File: | lib/CodeGen/SelectionDAG/DAGCombiner.cpp |
Warning: | line 391, column 32 The result of the right shift is undefined due to shifting by '64', which is greater or equal to the width of type 'llvm::APInt::WordType' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- DAGCombiner.cpp - Implement a DAG node combiner --------------------===// | |||
2 | // | |||
3 | // The LLVM Compiler Infrastructure | |||
4 | // | |||
5 | // This file is distributed under the University of Illinois Open Source | |||
6 | // License. See LICENSE.TXT for details. | |||
7 | // | |||
8 | //===----------------------------------------------------------------------===// | |||
9 | // | |||
10 | // This pass combines dag nodes to form fewer, simpler DAG nodes. It can be run | |||
11 | // both before and after the DAG is legalized. | |||
12 | // | |||
13 | // This pass is not a substitute for the LLVM IR instcombine pass. This pass is | |||
14 | // primarily intended to handle simplification opportunities that are implicit | |||
15 | // in the LLVM IR and exposed by the various codegen lowering phases. | |||
16 | // | |||
17 | //===----------------------------------------------------------------------===// | |||
18 | ||||
19 | #include "llvm/ADT/APFloat.h" | |||
20 | #include "llvm/ADT/APInt.h" | |||
21 | #include "llvm/ADT/ArrayRef.h" | |||
22 | #include "llvm/ADT/DenseMap.h" | |||
23 | #include "llvm/ADT/None.h" | |||
24 | #include "llvm/ADT/Optional.h" | |||
25 | #include "llvm/ADT/STLExtras.h" | |||
26 | #include "llvm/ADT/SetVector.h" | |||
27 | #include "llvm/ADT/SmallBitVector.h" | |||
28 | #include "llvm/ADT/SmallPtrSet.h" | |||
29 | #include "llvm/ADT/SmallSet.h" | |||
30 | #include "llvm/ADT/SmallVector.h" | |||
31 | #include "llvm/ADT/Statistic.h" | |||
32 | #include "llvm/Analysis/AliasAnalysis.h" | |||
33 | #include "llvm/Analysis/MemoryLocation.h" | |||
34 | #include "llvm/CodeGen/DAGCombine.h" | |||
35 | #include "llvm/CodeGen/ISDOpcodes.h" | |||
36 | #include "llvm/CodeGen/MachineFrameInfo.h" | |||
37 | #include "llvm/CodeGen/MachineFunction.h" | |||
38 | #include "llvm/CodeGen/MachineMemOperand.h" | |||
39 | #include "llvm/CodeGen/MachineValueType.h" | |||
40 | #include "llvm/CodeGen/RuntimeLibcalls.h" | |||
41 | #include "llvm/CodeGen/SelectionDAG.h" | |||
42 | #include "llvm/CodeGen/SelectionDAGAddressAnalysis.h" | |||
43 | #include "llvm/CodeGen/SelectionDAGNodes.h" | |||
44 | #include "llvm/CodeGen/SelectionDAGTargetInfo.h" | |||
45 | #include "llvm/CodeGen/TargetLowering.h" | |||
46 | #include "llvm/CodeGen/TargetRegisterInfo.h" | |||
47 | #include "llvm/CodeGen/TargetSubtargetInfo.h" | |||
48 | #include "llvm/CodeGen/ValueTypes.h" | |||
49 | #include "llvm/IR/Attributes.h" | |||
50 | #include "llvm/IR/Constant.h" | |||
51 | #include "llvm/IR/DataLayout.h" | |||
52 | #include "llvm/IR/DerivedTypes.h" | |||
53 | #include "llvm/IR/Function.h" | |||
54 | #include "llvm/IR/LLVMContext.h" | |||
55 | #include "llvm/IR/Metadata.h" | |||
56 | #include "llvm/Support/Casting.h" | |||
57 | #include "llvm/Support/CodeGen.h" | |||
58 | #include "llvm/Support/CommandLine.h" | |||
59 | #include "llvm/Support/Compiler.h" | |||
60 | #include "llvm/Support/Debug.h" | |||
61 | #include "llvm/Support/ErrorHandling.h" | |||
62 | #include "llvm/Support/KnownBits.h" | |||
63 | #include "llvm/Support/MathExtras.h" | |||
64 | #include "llvm/Support/raw_ostream.h" | |||
65 | #include "llvm/Target/TargetMachine.h" | |||
66 | #include "llvm/Target/TargetOptions.h" | |||
67 | #include <algorithm> | |||
68 | #include <cassert> | |||
69 | #include <cstdint> | |||
70 | #include <functional> | |||
71 | #include <iterator> | |||
72 | #include <string> | |||
73 | #include <tuple> | |||
74 | #include <utility> | |||
75 | #include <vector> | |||
76 | ||||
77 | using namespace llvm; | |||
78 | ||||
79 | #define DEBUG_TYPE"dagcombine" "dagcombine" | |||
80 | ||||
81 | STATISTIC(NodesCombined , "Number of dag nodes combined")static llvm::Statistic NodesCombined = {"dagcombine", "NodesCombined" , "Number of dag nodes combined", {0}, {false}}; | |||
82 | STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created")static llvm::Statistic PreIndexedNodes = {"dagcombine", "PreIndexedNodes" , "Number of pre-indexed nodes created", {0}, {false}}; | |||
83 | STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created")static llvm::Statistic PostIndexedNodes = {"dagcombine", "PostIndexedNodes" , "Number of post-indexed nodes created", {0}, {false}}; | |||
84 | STATISTIC(OpsNarrowed , "Number of load/op/store narrowed")static llvm::Statistic OpsNarrowed = {"dagcombine", "OpsNarrowed" , "Number of load/op/store narrowed", {0}, {false}}; | |||
85 | STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int")static llvm::Statistic LdStFP2Int = {"dagcombine", "LdStFP2Int" , "Number of fp load/store pairs transformed to int", {0}, {false }}; | |||
86 | STATISTIC(SlicedLoads, "Number of load sliced")static llvm::Statistic SlicedLoads = {"dagcombine", "SlicedLoads" , "Number of load sliced", {0}, {false}}; | |||
87 | ||||
88 | static cl::opt<bool> | |||
89 | CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden, | |||
90 | cl::desc("Enable DAG combiner's use of IR alias analysis")); | |||
91 | ||||
92 | static cl::opt<bool> | |||
93 | UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(true), | |||
94 | cl::desc("Enable DAG combiner's use of TBAA")); | |||
95 | ||||
96 | #ifndef NDEBUG | |||
97 | static cl::opt<std::string> | |||
98 | CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden, | |||
99 | cl::desc("Only use DAG-combiner alias analysis in this" | |||
100 | " function")); | |||
101 | #endif | |||
102 | ||||
103 | /// Hidden option to stress test load slicing, i.e., when this option | |||
104 | /// is enabled, load slicing bypasses most of its profitability guards. | |||
105 | static cl::opt<bool> | |||
106 | StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden, | |||
107 | cl::desc("Bypass the profitability model of load slicing"), | |||
108 | cl::init(false)); | |||
109 | ||||
110 | static cl::opt<bool> | |||
111 | MaySplitLoadIndex("combiner-split-load-index", cl::Hidden, cl::init(true), | |||
112 | cl::desc("DAG combiner may split indexing from loads")); | |||
113 | ||||
114 | namespace { | |||
115 | ||||
116 | class DAGCombiner { | |||
117 | SelectionDAG &DAG; | |||
118 | const TargetLowering &TLI; | |||
119 | CombineLevel Level; | |||
120 | CodeGenOpt::Level OptLevel; | |||
121 | bool LegalOperations = false; | |||
122 | bool LegalTypes = false; | |||
123 | bool ForCodeSize; | |||
124 | ||||
125 | /// \brief Worklist of all of the nodes that need to be simplified. | |||
126 | /// | |||
127 | /// This must behave as a stack -- new nodes to process are pushed onto the | |||
128 | /// back and when processing we pop off of the back. | |||
129 | /// | |||
130 | /// The worklist will not contain duplicates but may contain null entries | |||
131 | /// due to nodes being deleted from the underlying DAG. | |||
132 | SmallVector<SDNode *, 64> Worklist; | |||
133 | ||||
134 | /// \brief Mapping from an SDNode to its position on the worklist. | |||
135 | /// | |||
136 | /// This is used to find and remove nodes from the worklist (by nulling | |||
137 | /// them) when they are deleted from the underlying DAG. It relies on | |||
138 | /// stable indices of nodes within the worklist. | |||
139 | DenseMap<SDNode *, unsigned> WorklistMap; | |||
140 | ||||
141 | /// \brief Set of nodes which have been combined (at least once). | |||
142 | /// | |||
143 | /// This is used to allow us to reliably add any operands of a DAG node | |||
144 | /// which have not yet been combined to the worklist. | |||
145 | SmallPtrSet<SDNode *, 32> CombinedNodes; | |||
146 | ||||
147 | // AA - Used for DAG load/store alias analysis. | |||
148 | AliasAnalysis *AA; | |||
149 | ||||
150 | /// When an instruction is simplified, add all users of the instruction to | |||
151 | /// the work lists because they might get more simplified now. | |||
152 | void AddUsersToWorklist(SDNode *N) { | |||
153 | for (SDNode *Node : N->uses()) | |||
154 | AddToWorklist(Node); | |||
155 | } | |||
156 | ||||
157 | /// Call the node-specific routine that folds each particular type of node. | |||
158 | SDValue visit(SDNode *N); | |||
159 | ||||
160 | public: | |||
161 | DAGCombiner(SelectionDAG &D, AliasAnalysis *AA, CodeGenOpt::Level OL) | |||
162 | : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes), | |||
163 | OptLevel(OL), AA(AA) { | |||
164 | ForCodeSize = DAG.getMachineFunction().getFunction().optForSize(); | |||
165 | ||||
166 | MaximumLegalStoreInBits = 0; | |||
167 | for (MVT VT : MVT::all_valuetypes()) | |||
168 | if (EVT(VT).isSimple() && VT != MVT::Other && | |||
169 | TLI.isTypeLegal(EVT(VT)) && | |||
170 | VT.getSizeInBits() >= MaximumLegalStoreInBits) | |||
171 | MaximumLegalStoreInBits = VT.getSizeInBits(); | |||
172 | } | |||
173 | ||||
174 | /// Add to the worklist making sure its instance is at the back (next to be | |||
175 | /// processed.) | |||
176 | void AddToWorklist(SDNode *N) { | |||
177 | assert(N->getOpcode() != ISD::DELETED_NODE &&(static_cast <bool> (N->getOpcode() != ISD::DELETED_NODE && "Deleted Node added to Worklist") ? void (0) : __assert_fail ("N->getOpcode() != ISD::DELETED_NODE && \"Deleted Node added to Worklist\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 178, __extension__ __PRETTY_FUNCTION__)) | |||
178 | "Deleted Node added to Worklist")(static_cast <bool> (N->getOpcode() != ISD::DELETED_NODE && "Deleted Node added to Worklist") ? void (0) : __assert_fail ("N->getOpcode() != ISD::DELETED_NODE && \"Deleted Node added to Worklist\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 178, __extension__ __PRETTY_FUNCTION__)); | |||
179 | ||||
180 | // Skip handle nodes as they can't usefully be combined and confuse the | |||
181 | // zero-use deletion strategy. | |||
182 | if (N->getOpcode() == ISD::HANDLENODE) | |||
183 | return; | |||
184 | ||||
185 | if (WorklistMap.insert(std::make_pair(N, Worklist.size())).second) | |||
186 | Worklist.push_back(N); | |||
187 | } | |||
188 | ||||
189 | /// Remove all instances of N from the worklist. | |||
190 | void removeFromWorklist(SDNode *N) { | |||
191 | CombinedNodes.erase(N); | |||
192 | ||||
193 | auto It = WorklistMap.find(N); | |||
194 | if (It == WorklistMap.end()) | |||
195 | return; // Not in the worklist. | |||
196 | ||||
197 | // Null out the entry rather than erasing it to avoid a linear operation. | |||
198 | Worklist[It->second] = nullptr; | |||
199 | WorklistMap.erase(It); | |||
200 | } | |||
201 | ||||
202 | void deleteAndRecombine(SDNode *N); | |||
203 | bool recursivelyDeleteUnusedNodes(SDNode *N); | |||
204 | ||||
205 | /// Replaces all uses of the results of one DAG node with new values. | |||
206 | SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo, | |||
207 | bool AddTo = true); | |||
208 | ||||
209 | /// Replaces all uses of the results of one DAG node with new values. | |||
210 | SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) { | |||
211 | return CombineTo(N, &Res, 1, AddTo); | |||
212 | } | |||
213 | ||||
214 | /// Replaces all uses of the results of one DAG node with new values. | |||
215 | SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, | |||
216 | bool AddTo = true) { | |||
217 | SDValue To[] = { Res0, Res1 }; | |||
218 | return CombineTo(N, To, 2, AddTo); | |||
219 | } | |||
220 | ||||
221 | void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO); | |||
222 | ||||
223 | private: | |||
224 | unsigned MaximumLegalStoreInBits; | |||
225 | ||||
226 | /// Check the specified integer node value to see if it can be simplified or | |||
227 | /// if things it uses can be simplified by bit propagation. | |||
228 | /// If so, return true. | |||
229 | bool SimplifyDemandedBits(SDValue Op) { | |||
230 | unsigned BitWidth = Op.getScalarValueSizeInBits(); | |||
231 | APInt Demanded = APInt::getAllOnesValue(BitWidth); | |||
232 | return SimplifyDemandedBits(Op, Demanded); | |||
233 | } | |||
234 | ||||
235 | /// Check the specified vector node value to see if it can be simplified or | |||
236 | /// if things it uses can be simplified as it only uses some of the | |||
237 | /// elements. If so, return true. | |||
238 | bool SimplifyDemandedVectorElts(SDValue Op) { | |||
239 | unsigned NumElts = Op.getValueType().getVectorNumElements(); | |||
240 | APInt Demanded = APInt::getAllOnesValue(NumElts); | |||
241 | return SimplifyDemandedVectorElts(Op, Demanded); | |||
242 | } | |||
243 | ||||
244 | bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded); | |||
245 | bool SimplifyDemandedVectorElts(SDValue Op, const APInt &Demanded); | |||
246 | ||||
247 | bool CombineToPreIndexedLoadStore(SDNode *N); | |||
248 | bool CombineToPostIndexedLoadStore(SDNode *N); | |||
249 | SDValue SplitIndexingFromLoad(LoadSDNode *LD); | |||
250 | bool SliceUpLoad(SDNode *N); | |||
251 | ||||
252 | /// \brief Replace an ISD::EXTRACT_VECTOR_ELT of a load with a narrowed | |||
253 | /// load. | |||
254 | /// | |||
255 | /// \param EVE ISD::EXTRACT_VECTOR_ELT to be replaced. | |||
256 | /// \param InVecVT type of the input vector to EVE with bitcasts resolved. | |||
257 | /// \param EltNo index of the vector element to load. | |||
258 | /// \param OriginalLoad load that EVE came from to be replaced. | |||
259 | /// \returns EVE on success SDValue() on failure. | |||
260 | SDValue ReplaceExtractVectorEltOfLoadWithNarrowedLoad( | |||
261 | SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad); | |||
262 | void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad); | |||
263 | SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace); | |||
264 | SDValue SExtPromoteOperand(SDValue Op, EVT PVT); | |||
265 | SDValue ZExtPromoteOperand(SDValue Op, EVT PVT); | |||
266 | SDValue PromoteIntBinOp(SDValue Op); | |||
267 | SDValue PromoteIntShiftOp(SDValue Op); | |||
268 | SDValue PromoteExtend(SDValue Op); | |||
269 | bool PromoteLoad(SDValue Op); | |||
270 | ||||
271 | void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs, | |||
272 | SDValue OrigLoad, SDValue ExtLoad, | |||
273 | const SDLoc &DL, | |||
274 | ISD::NodeType ExtType); | |||
275 | ||||
276 | /// Call the node-specific routine that knows how to fold each | |||
277 | /// particular type of node. If that doesn't do anything, try the | |||
278 | /// target-specific DAG combines. | |||
279 | SDValue combine(SDNode *N); | |||
280 | ||||
281 | // Visitation implementation - Implement dag node combining for different | |||
282 | // node types. The semantics are as follows: | |||
283 | // Return Value: | |||
284 | // SDValue.getNode() == 0 - No change was made | |||
285 | // SDValue.getNode() == N - N was replaced, is dead and has been handled. | |||
286 | // otherwise - N should be replaced by the returned Operand. | |||
287 | // | |||
288 | SDValue visitTokenFactor(SDNode *N); | |||
289 | SDValue visitMERGE_VALUES(SDNode *N); | |||
290 | SDValue visitADD(SDNode *N); | |||
291 | SDValue visitADDLike(SDValue N0, SDValue N1, SDNode *LocReference); | |||
292 | SDValue visitSUB(SDNode *N); | |||
293 | SDValue visitADDC(SDNode *N); | |||
294 | SDValue visitUADDO(SDNode *N); | |||
295 | SDValue visitUADDOLike(SDValue N0, SDValue N1, SDNode *N); | |||
296 | SDValue visitSUBC(SDNode *N); | |||
297 | SDValue visitUSUBO(SDNode *N); | |||
298 | SDValue visitADDE(SDNode *N); | |||
299 | SDValue visitADDCARRY(SDNode *N); | |||
300 | SDValue visitADDCARRYLike(SDValue N0, SDValue N1, SDValue CarryIn, SDNode *N); | |||
301 | SDValue visitSUBE(SDNode *N); | |||
302 | SDValue visitSUBCARRY(SDNode *N); | |||
303 | SDValue visitMUL(SDNode *N); | |||
304 | SDValue useDivRem(SDNode *N); | |||
305 | SDValue visitSDIV(SDNode *N); | |||
306 | SDValue visitUDIV(SDNode *N); | |||
307 | SDValue visitREM(SDNode *N); | |||
308 | SDValue visitMULHU(SDNode *N); | |||
309 | SDValue visitMULHS(SDNode *N); | |||
310 | SDValue visitSMUL_LOHI(SDNode *N); | |||
311 | SDValue visitUMUL_LOHI(SDNode *N); | |||
312 | SDValue visitSMULO(SDNode *N); | |||
313 | SDValue visitUMULO(SDNode *N); | |||
314 | SDValue visitIMINMAX(SDNode *N); | |||
315 | SDValue visitAND(SDNode *N); | |||
316 | SDValue visitANDLike(SDValue N0, SDValue N1, SDNode *LocReference); | |||
317 | SDValue visitOR(SDNode *N); | |||
318 | SDValue visitORLike(SDValue N0, SDValue N1, SDNode *LocReference); | |||
319 | SDValue visitXOR(SDNode *N); | |||
320 | SDValue SimplifyVBinOp(SDNode *N); | |||
321 | SDValue visitSHL(SDNode *N); | |||
322 | SDValue visitSRA(SDNode *N); | |||
323 | SDValue visitSRL(SDNode *N); | |||
324 | SDValue visitRotate(SDNode *N); | |||
325 | SDValue visitABS(SDNode *N); | |||
326 | SDValue visitBSWAP(SDNode *N); | |||
327 | SDValue visitBITREVERSE(SDNode *N); | |||
328 | SDValue visitCTLZ(SDNode *N); | |||
329 | SDValue visitCTLZ_ZERO_UNDEF(SDNode *N); | |||
330 | SDValue visitCTTZ(SDNode *N); | |||
331 | SDValue visitCTTZ_ZERO_UNDEF(SDNode *N); | |||
332 | SDValue visitCTPOP(SDNode *N); | |||
333 | SDValue visitSELECT(SDNode *N); | |||
334 | SDValue visitVSELECT(SDNode *N); | |||
335 | SDValue visitSELECT_CC(SDNode *N); | |||
336 | SDValue visitSETCC(SDNode *N); | |||
337 | SDValue visitSETCCE(SDNode *N); | |||
338 | SDValue visitSETCCCARRY(SDNode *N); | |||
339 | SDValue visitSIGN_EXTEND(SDNode *N); | |||
340 | SDValue visitZERO_EXTEND(SDNode *N); | |||
341 | SDValue visitANY_EXTEND(SDNode *N); | |||
342 | SDValue visitAssertExt(SDNode *N); | |||
343 | SDValue visitSIGN_EXTEND_INREG(SDNode *N); | |||
344 | SDValue visitSIGN_EXTEND_VECTOR_INREG(SDNode *N); | |||
345 | SDValue visitZERO_EXTEND_VECTOR_INREG(SDNode *N); | |||
346 | SDValue visitTRUNCATE(SDNode *N); | |||
347 | SDValue visitBITCAST(SDNode *N); | |||
348 | SDValue visitBUILD_PAIR(SDNode *N); | |||
349 | SDValue visitFADD(SDNode *N); | |||
350 | SDValue visitFSUB(SDNode *N); | |||
351 | SDValue visitFMUL(SDNode *N); | |||
352 | SDValue visitFMA(SDNode *N); | |||
353 | SDValue visitFDIV(SDNode *N); | |||
354 | SDValue visitFREM(SDNode *N); | |||
355 | SDValue visitFSQRT(SDNode *N); | |||
356 | SDValue visitFCOPYSIGN(SDNode *N); | |||
357 | SDValue visitSINT_TO_FP(SDNode *N); | |||
358 | SDValue visitUINT_TO_FP(SDNode *N); | |||
359 | SDValue visitFP_TO_SINT(SDNode *N); | |||
360 | SDValue visitFP_TO_UINT(SDNode *N); | |||
361 | SDValue visitFP_ROUND(SDNode *N); | |||
362 | SDValue visitFP_ROUND_INREG(SDNode *N); | |||
363 | SDValue visitFP_EXTEND(SDNode *N); | |||
364 | SDValue visitFNEG(SDNode *N); | |||
365 | SDValue visitFABS(SDNode *N); | |||
366 | SDValue visitFCEIL(SDNode *N); | |||
367 | SDValue visitFTRUNC(SDNode *N); | |||
368 | SDValue visitFFLOOR(SDNode *N); | |||
369 | SDValue visitFMINNUM(SDNode *N); | |||
370 | SDValue visitFMAXNUM(SDNode *N); | |||
371 | SDValue visitBRCOND(SDNode *N); | |||
372 | SDValue visitBR_CC(SDNode *N); | |||
373 | SDValue visitLOAD(SDNode *N); | |||
374 | ||||
375 | SDValue replaceStoreChain(StoreSDNode *ST, SDValue BetterChain); | |||
376 | SDValue replaceStoreOfFPConstant(StoreSDNode *ST); | |||
377 | ||||
378 | SDValue visitSTORE(SDNode *N); | |||
379 | SDValue visitINSERT_VECTOR_ELT(SDNode *N); | |||
380 | SDValue visitEXTRACT_VECTOR_ELT(SDNode *N); | |||
381 | SDValue visitBUILD_VECTOR(SDNode *N); | |||
382 | SDValue visitCONCAT_VECTORS(SDNode *N); | |||
383 | SDValue visitEXTRACT_SUBVECTOR(SDNode *N); | |||
384 | SDValue visitVECTOR_SHUFFLE(SDNode *N); | |||
385 | SDValue visitSCALAR_TO_VECTOR(SDNode *N); | |||
386 | SDValue visitINSERT_SUBVECTOR(SDNode *N); | |||
387 | SDValue visitMLOAD(SDNode *N); | |||
388 | SDValue visitMSTORE(SDNode *N); | |||
389 | SDValue visitMGATHER(SDNode *N); | |||
390 | SDValue visitMSCATTER(SDNode *N); | |||
391 | SDValue visitFP_TO_FP16(SDNode *N); | |||
392 | SDValue visitFP16_TO_FP(SDNode *N); | |||
393 | ||||
394 | SDValue visitFADDForFMACombine(SDNode *N); | |||
395 | SDValue visitFSUBForFMACombine(SDNode *N); | |||
396 | SDValue visitFMULForFMADistributiveCombine(SDNode *N); | |||
397 | ||||
398 | SDValue XformToShuffleWithZero(SDNode *N); | |||
399 | SDValue ReassociateOps(unsigned Opc, const SDLoc &DL, SDValue LHS, | |||
400 | SDValue RHS); | |||
401 | ||||
402 | SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt); | |||
403 | ||||
404 | SDValue foldSelectOfConstants(SDNode *N); | |||
405 | SDValue foldVSelectOfConstants(SDNode *N); | |||
406 | SDValue foldBinOpIntoSelect(SDNode *BO); | |||
407 | bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS); | |||
408 | SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N); | |||
409 | SDValue SimplifySelect(const SDLoc &DL, SDValue N0, SDValue N1, SDValue N2); | |||
410 | SDValue SimplifySelectCC(const SDLoc &DL, SDValue N0, SDValue N1, | |||
411 | SDValue N2, SDValue N3, ISD::CondCode CC, | |||
412 | bool NotExtCompare = false); | |||
413 | SDValue foldSelectCCToShiftAnd(const SDLoc &DL, SDValue N0, SDValue N1, | |||
414 | SDValue N2, SDValue N3, ISD::CondCode CC); | |||
415 | SDValue foldLogicOfSetCCs(bool IsAnd, SDValue N0, SDValue N1, | |||
416 | const SDLoc &DL); | |||
417 | SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond, | |||
418 | const SDLoc &DL, bool foldBooleans); | |||
419 | SDValue rebuildSetCC(SDValue N); | |||
420 | ||||
421 | bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS, | |||
422 | SDValue &CC) const; | |||
423 | bool isOneUseSetCC(SDValue N) const; | |||
424 | ||||
425 | SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp, | |||
426 | unsigned HiOp); | |||
427 | SDValue CombineConsecutiveLoads(SDNode *N, EVT VT); | |||
428 | SDValue CombineExtLoad(SDNode *N); | |||
429 | SDValue combineRepeatedFPDivisors(SDNode *N); | |||
430 | SDValue combineInsertEltToShuffle(SDNode *N, unsigned InsIndex); | |||
431 | SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT); | |||
432 | SDValue BuildSDIV(SDNode *N); | |||
433 | SDValue BuildSDIVPow2(SDNode *N); | |||
434 | SDValue BuildUDIV(SDNode *N); | |||
435 | SDValue BuildLogBase2(SDValue Op, const SDLoc &DL); | |||
436 | SDValue BuildReciprocalEstimate(SDValue Op, SDNodeFlags Flags); | |||
437 | SDValue buildRsqrtEstimate(SDValue Op, SDNodeFlags Flags); | |||
438 | SDValue buildSqrtEstimate(SDValue Op, SDNodeFlags Flags); | |||
439 | SDValue buildSqrtEstimateImpl(SDValue Op, SDNodeFlags Flags, bool Recip); | |||
440 | SDValue buildSqrtNROneConst(SDValue Op, SDValue Est, unsigned Iterations, | |||
441 | SDNodeFlags Flags, bool Reciprocal); | |||
442 | SDValue buildSqrtNRTwoConst(SDValue Op, SDValue Est, unsigned Iterations, | |||
443 | SDNodeFlags Flags, bool Reciprocal); | |||
444 | SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1, | |||
445 | bool DemandHighBits = true); | |||
446 | SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1); | |||
447 | SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg, | |||
448 | SDValue InnerPos, SDValue InnerNeg, | |||
449 | unsigned PosOpcode, unsigned NegOpcode, | |||
450 | const SDLoc &DL); | |||
451 | SDNode *MatchRotate(SDValue LHS, SDValue RHS, const SDLoc &DL); | |||
452 | SDValue MatchLoadCombine(SDNode *N); | |||
453 | SDValue ReduceLoadWidth(SDNode *N); | |||
454 | SDValue ReduceLoadOpStoreWidth(SDNode *N); | |||
455 | SDValue splitMergedValStore(StoreSDNode *ST); | |||
456 | SDValue TransformFPLoadStorePair(SDNode *N); | |||
457 | SDValue reduceBuildVecExtToExtBuildVec(SDNode *N); | |||
458 | SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N); | |||
459 | SDValue reduceBuildVecToShuffle(SDNode *N); | |||
460 | SDValue createBuildVecShuffle(const SDLoc &DL, SDNode *N, | |||
461 | ArrayRef<int> VectorMask, SDValue VecIn1, | |||
462 | SDValue VecIn2, unsigned LeftIdx); | |||
463 | SDValue matchVSelectOpSizesWithSetCC(SDNode *N); | |||
464 | ||||
465 | /// Walk up chain skipping non-aliasing memory nodes, | |||
466 | /// looking for aliasing nodes and adding them to the Aliases vector. | |||
467 | void GatherAllAliases(SDNode *N, SDValue OriginalChain, | |||
468 | SmallVectorImpl<SDValue> &Aliases); | |||
469 | ||||
470 | /// Return true if there is any possibility that the two addresses overlap. | |||
471 | bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const; | |||
472 | ||||
473 | /// Walk up chain skipping non-aliasing memory nodes, looking for a better | |||
474 | /// chain (aliasing node.) | |||
475 | SDValue FindBetterChain(SDNode *N, SDValue Chain); | |||
476 | ||||
477 | /// Try to replace a store and any possibly adjacent stores on | |||
478 | /// consecutive chains with better chains. Return true only if St is | |||
479 | /// replaced. | |||
480 | /// | |||
481 | /// Notice that other chains may still be replaced even if the function | |||
482 | /// returns false. | |||
483 | bool findBetterNeighborChains(StoreSDNode *St); | |||
484 | ||||
485 | /// Match "(X shl/srl V1) & V2" where V2 may not be present. | |||
486 | bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask); | |||
487 | ||||
488 | /// Holds a pointer to an LSBaseSDNode as well as information on where it | |||
489 | /// is located in a sequence of memory operations connected by a chain. | |||
490 | struct MemOpLink { | |||
491 | // Ptr to the mem node. | |||
492 | LSBaseSDNode *MemNode; | |||
493 | ||||
494 | // Offset from the base ptr. | |||
495 | int64_t OffsetFromBase; | |||
496 | ||||
497 | MemOpLink(LSBaseSDNode *N, int64_t Offset) | |||
498 | : MemNode(N), OffsetFromBase(Offset) {} | |||
499 | }; | |||
500 | ||||
501 | /// This is a helper function for visitMUL to check the profitability | |||
502 | /// of folding (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2). | |||
503 | /// MulNode is the original multiply, AddNode is (add x, c1), | |||
504 | /// and ConstNode is c2. | |||
505 | bool isMulAddWithConstProfitable(SDNode *MulNode, | |||
506 | SDValue &AddNode, | |||
507 | SDValue &ConstNode); | |||
508 | ||||
509 | /// This is a helper function for visitAND and visitZERO_EXTEND. Returns | |||
510 | /// true if the (and (load x) c) pattern matches an extload. ExtVT returns | |||
511 | /// the type of the loaded value to be extended. | |||
512 | bool isAndLoadExtLoad(ConstantSDNode *AndC, LoadSDNode *LoadN, | |||
513 | EVT LoadResultTy, EVT &ExtVT); | |||
514 | ||||
515 | /// Helper function to calculate whether the given Load can have its | |||
516 | /// width reduced to ExtVT. | |||
517 | bool isLegalNarrowLoad(LoadSDNode *LoadN, ISD::LoadExtType ExtType, | |||
518 | EVT &ExtVT, unsigned ShAmt = 0); | |||
519 | ||||
520 | /// Used by BackwardsPropagateMask to find suitable loads. | |||
521 | bool SearchForAndLoads(SDNode *N, SmallPtrSetImpl<LoadSDNode*> &Loads, | |||
522 | SmallPtrSetImpl<SDNode*> &NodeWithConsts, | |||
523 | ConstantSDNode *Mask, SDNode *&UncombinedNode); | |||
524 | /// Attempt to propagate a given AND node back to load leaves so that they | |||
525 | /// can be combined into narrow loads. | |||
526 | bool BackwardsPropagateMask(SDNode *N, SelectionDAG &DAG); | |||
527 | ||||
528 | /// Helper function for MergeConsecutiveStores which merges the | |||
529 | /// component store chains. | |||
530 | SDValue getMergeStoreChains(SmallVectorImpl<MemOpLink> &StoreNodes, | |||
531 | unsigned NumStores); | |||
532 | ||||
533 | /// This is a helper function for MergeConsecutiveStores. When the | |||
534 | /// source elements of the consecutive stores are all constants or | |||
535 | /// all extracted vector elements, try to merge them into one | |||
536 | /// larger store introducing bitcasts if necessary. \return True | |||
537 | /// if a merged store was created. | |||
538 | bool MergeStoresOfConstantsOrVecElts(SmallVectorImpl<MemOpLink> &StoreNodes, | |||
539 | EVT MemVT, unsigned NumStores, | |||
540 | bool IsConstantSrc, bool UseVector, | |||
541 | bool UseTrunc); | |||
542 | ||||
543 | /// This is a helper function for MergeConsecutiveStores. Stores | |||
544 | /// that potentially may be merged with St are placed in | |||
545 | /// StoreNodes. | |||
546 | void getStoreMergeCandidates(StoreSDNode *St, | |||
547 | SmallVectorImpl<MemOpLink> &StoreNodes); | |||
548 | ||||
549 | /// Helper function for MergeConsecutiveStores. Checks if | |||
550 | /// candidate stores have indirect dependency through their | |||
551 | /// operands. \return True if safe to merge. | |||
552 | bool checkMergeStoreCandidatesForDependencies( | |||
553 | SmallVectorImpl<MemOpLink> &StoreNodes, unsigned NumStores); | |||
554 | ||||
555 | /// Merge consecutive store operations into a wide store. | |||
556 | /// This optimization uses wide integers or vectors when possible. | |||
557 | /// \return number of stores that were merged into a merged store (the | |||
558 | /// affected nodes are stored as a prefix in \p StoreNodes). | |||
559 | bool MergeConsecutiveStores(StoreSDNode *N); | |||
560 | ||||
561 | /// \brief Try to transform a truncation where C is a constant: | |||
562 | /// (trunc (and X, C)) -> (and (trunc X), (trunc C)) | |||
563 | /// | |||
564 | /// \p N needs to be a truncation and its first operand an AND. Other | |||
565 | /// requirements are checked by the function (e.g. that trunc is | |||
566 | /// single-use) and if missed an empty SDValue is returned. | |||
567 | SDValue distributeTruncateThroughAnd(SDNode *N); | |||
568 | ||||
569 | public: | |||
570 | /// Runs the dag combiner on all nodes in the work list | |||
571 | void Run(CombineLevel AtLevel); | |||
572 | ||||
573 | SelectionDAG &getDAG() const { return DAG; } | |||
574 | ||||
575 | /// Returns a type large enough to hold any valid shift amount - before type | |||
576 | /// legalization these can be huge. | |||
577 | EVT getShiftAmountTy(EVT LHSTy) { | |||
578 | assert(LHSTy.isInteger() && "Shift amount is not an integer type!")(static_cast <bool> (LHSTy.isInteger() && "Shift amount is not an integer type!" ) ? void (0) : __assert_fail ("LHSTy.isInteger() && \"Shift amount is not an integer type!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 578, __extension__ __PRETTY_FUNCTION__)); | |||
579 | return TLI.getShiftAmountTy(LHSTy, DAG.getDataLayout(), LegalTypes); | |||
580 | } | |||
581 | ||||
582 | /// This method returns true if we are running before type legalization or | |||
583 | /// if the specified VT is legal. | |||
584 | bool isTypeLegal(const EVT &VT) { | |||
585 | if (!LegalTypes) return true; | |||
586 | return TLI.isTypeLegal(VT); | |||
587 | } | |||
588 | ||||
589 | /// Convenience wrapper around TargetLowering::getSetCCResultType | |||
590 | EVT getSetCCResultType(EVT VT) const { | |||
591 | return TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); | |||
592 | } | |||
593 | }; | |||
594 | ||||
595 | /// This class is a DAGUpdateListener that removes any deleted | |||
596 | /// nodes from the worklist. | |||
597 | class WorklistRemover : public SelectionDAG::DAGUpdateListener { | |||
598 | DAGCombiner &DC; | |||
599 | ||||
600 | public: | |||
601 | explicit WorklistRemover(DAGCombiner &dc) | |||
602 | : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {} | |||
603 | ||||
604 | void NodeDeleted(SDNode *N, SDNode *E) override { | |||
605 | DC.removeFromWorklist(N); | |||
606 | } | |||
607 | }; | |||
608 | ||||
609 | } // end anonymous namespace | |||
610 | ||||
611 | //===----------------------------------------------------------------------===// | |||
612 | // TargetLowering::DAGCombinerInfo implementation | |||
613 | //===----------------------------------------------------------------------===// | |||
614 | ||||
615 | void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) { | |||
616 | ((DAGCombiner*)DC)->AddToWorklist(N); | |||
617 | } | |||
618 | ||||
619 | SDValue TargetLowering::DAGCombinerInfo:: | |||
620 | CombineTo(SDNode *N, ArrayRef<SDValue> To, bool AddTo) { | |||
621 | return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo); | |||
622 | } | |||
623 | ||||
624 | SDValue TargetLowering::DAGCombinerInfo:: | |||
625 | CombineTo(SDNode *N, SDValue Res, bool AddTo) { | |||
626 | return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo); | |||
627 | } | |||
628 | ||||
629 | SDValue TargetLowering::DAGCombinerInfo:: | |||
630 | CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) { | |||
631 | return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo); | |||
632 | } | |||
633 | ||||
634 | void TargetLowering::DAGCombinerInfo:: | |||
635 | CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) { | |||
636 | return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO); | |||
637 | } | |||
638 | ||||
639 | //===----------------------------------------------------------------------===// | |||
640 | // Helper Functions | |||
641 | //===----------------------------------------------------------------------===// | |||
642 | ||||
643 | void DAGCombiner::deleteAndRecombine(SDNode *N) { | |||
644 | removeFromWorklist(N); | |||
645 | ||||
646 | // If the operands of this node are only used by the node, they will now be | |||
647 | // dead. Make sure to re-visit them and recursively delete dead nodes. | |||
648 | for (const SDValue &Op : N->ops()) | |||
649 | // For an operand generating multiple values, one of the values may | |||
650 | // become dead allowing further simplification (e.g. split index | |||
651 | // arithmetic from an indexed load). | |||
652 | if (Op->hasOneUse() || Op->getNumValues() > 1) | |||
653 | AddToWorklist(Op.getNode()); | |||
654 | ||||
655 | DAG.DeleteNode(N); | |||
656 | } | |||
657 | ||||
658 | /// Return 1 if we can compute the negated form of the specified expression for | |||
659 | /// the same cost as the expression itself, or 2 if we can compute the negated | |||
660 | /// form more cheaply than the expression itself. | |||
661 | static char isNegatibleForFree(SDValue Op, bool LegalOperations, | |||
662 | const TargetLowering &TLI, | |||
663 | const TargetOptions *Options, | |||
664 | unsigned Depth = 0) { | |||
665 | // fneg is removable even if it has multiple uses. | |||
666 | if (Op.getOpcode() == ISD::FNEG) return 2; | |||
667 | ||||
668 | // Don't allow anything with multiple uses. | |||
669 | if (!Op.hasOneUse()) return 0; | |||
670 | ||||
671 | // Don't recurse exponentially. | |||
672 | if (Depth > 6) return 0; | |||
673 | ||||
674 | switch (Op.getOpcode()) { | |||
675 | default: return false; | |||
676 | case ISD::ConstantFP: { | |||
677 | if (!LegalOperations) | |||
678 | return 1; | |||
679 | ||||
680 | // Don't invert constant FP values after legalization unless the target says | |||
681 | // the negated constant is legal. | |||
682 | EVT VT = Op.getValueType(); | |||
683 | return TLI.isOperationLegal(ISD::ConstantFP, VT) || | |||
684 | TLI.isFPImmLegal(neg(cast<ConstantFPSDNode>(Op)->getValueAPF()), VT); | |||
685 | } | |||
686 | case ISD::FADD: | |||
687 | // FIXME: determine better conditions for this xform. | |||
688 | if (!Options->UnsafeFPMath) return 0; | |||
689 | ||||
690 | // After operation legalization, it might not be legal to create new FSUBs. | |||
691 | if (LegalOperations && | |||
692 | !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType())) | |||
693 | return 0; | |||
694 | ||||
695 | // fold (fneg (fadd A, B)) -> (fsub (fneg A), B) | |||
696 | if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, | |||
697 | Options, Depth + 1)) | |||
698 | return V; | |||
699 | // fold (fneg (fadd A, B)) -> (fsub (fneg B), A) | |||
700 | return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options, | |||
701 | Depth + 1); | |||
702 | case ISD::FSUB: | |||
703 | // We can't turn -(A-B) into B-A when we honor signed zeros. | |||
704 | if (!Options->NoSignedZerosFPMath && | |||
705 | !Op.getNode()->getFlags().hasNoSignedZeros()) | |||
706 | return 0; | |||
707 | ||||
708 | // fold (fneg (fsub A, B)) -> (fsub B, A) | |||
709 | return 1; | |||
710 | ||||
711 | case ISD::FMUL: | |||
712 | case ISD::FDIV: | |||
713 | if (Options->HonorSignDependentRoundingFPMath()) return 0; | |||
714 | ||||
715 | // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y)) | |||
716 | if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, | |||
717 | Options, Depth + 1)) | |||
718 | return V; | |||
719 | ||||
720 | return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options, | |||
721 | Depth + 1); | |||
722 | ||||
723 | case ISD::FP_EXTEND: | |||
724 | case ISD::FP_ROUND: | |||
725 | case ISD::FSIN: | |||
726 | return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options, | |||
727 | Depth + 1); | |||
728 | } | |||
729 | } | |||
730 | ||||
731 | /// If isNegatibleForFree returns true, return the newly negated expression. | |||
732 | static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG, | |||
733 | bool LegalOperations, unsigned Depth = 0) { | |||
734 | const TargetOptions &Options = DAG.getTarget().Options; | |||
735 | // fneg is removable even if it has multiple uses. | |||
736 | if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0); | |||
737 | ||||
738 | // Don't allow anything with multiple uses. | |||
739 | assert(Op.hasOneUse() && "Unknown reuse!")(static_cast <bool> (Op.hasOneUse() && "Unknown reuse!" ) ? void (0) : __assert_fail ("Op.hasOneUse() && \"Unknown reuse!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 739, __extension__ __PRETTY_FUNCTION__)); | |||
740 | ||||
741 | assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree")(static_cast <bool> (Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree" ) ? void (0) : __assert_fail ("Depth <= 6 && \"GetNegatedExpression doesn't match isNegatibleForFree\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 741, __extension__ __PRETTY_FUNCTION__)); | |||
742 | ||||
743 | const SDNodeFlags Flags = Op.getNode()->getFlags(); | |||
744 | ||||
745 | switch (Op.getOpcode()) { | |||
746 | default: llvm_unreachable("Unknown code")::llvm::llvm_unreachable_internal("Unknown code", "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 746); | |||
747 | case ISD::ConstantFP: { | |||
748 | APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF(); | |||
749 | V.changeSign(); | |||
750 | return DAG.getConstantFP(V, SDLoc(Op), Op.getValueType()); | |||
751 | } | |||
752 | case ISD::FADD: | |||
753 | // FIXME: determine better conditions for this xform. | |||
754 | assert(Options.UnsafeFPMath)(static_cast <bool> (Options.UnsafeFPMath) ? void (0) : __assert_fail ("Options.UnsafeFPMath", "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 754, __extension__ __PRETTY_FUNCTION__)); | |||
755 | ||||
756 | // fold (fneg (fadd A, B)) -> (fsub (fneg A), B) | |||
757 | if (isNegatibleForFree(Op.getOperand(0), LegalOperations, | |||
758 | DAG.getTargetLoweringInfo(), &Options, Depth+1)) | |||
759 | return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(), | |||
760 | GetNegatedExpression(Op.getOperand(0), DAG, | |||
761 | LegalOperations, Depth+1), | |||
762 | Op.getOperand(1), Flags); | |||
763 | // fold (fneg (fadd A, B)) -> (fsub (fneg B), A) | |||
764 | return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(), | |||
765 | GetNegatedExpression(Op.getOperand(1), DAG, | |||
766 | LegalOperations, Depth+1), | |||
767 | Op.getOperand(0), Flags); | |||
768 | case ISD::FSUB: | |||
769 | // fold (fneg (fsub 0, B)) -> B | |||
770 | if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0))) | |||
771 | if (N0CFP->isZero()) | |||
772 | return Op.getOperand(1); | |||
773 | ||||
774 | // fold (fneg (fsub A, B)) -> (fsub B, A) | |||
775 | return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(), | |||
776 | Op.getOperand(1), Op.getOperand(0), Flags); | |||
777 | ||||
778 | case ISD::FMUL: | |||
779 | case ISD::FDIV: | |||
780 | assert(!Options.HonorSignDependentRoundingFPMath())(static_cast <bool> (!Options.HonorSignDependentRoundingFPMath ()) ? void (0) : __assert_fail ("!Options.HonorSignDependentRoundingFPMath()" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 780, __extension__ __PRETTY_FUNCTION__)); | |||
781 | ||||
782 | // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) | |||
783 | if (isNegatibleForFree(Op.getOperand(0), LegalOperations, | |||
784 | DAG.getTargetLoweringInfo(), &Options, Depth+1)) | |||
785 | return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(), | |||
786 | GetNegatedExpression(Op.getOperand(0), DAG, | |||
787 | LegalOperations, Depth+1), | |||
788 | Op.getOperand(1), Flags); | |||
789 | ||||
790 | // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y)) | |||
791 | return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(), | |||
792 | Op.getOperand(0), | |||
793 | GetNegatedExpression(Op.getOperand(1), DAG, | |||
794 | LegalOperations, Depth+1), Flags); | |||
795 | ||||
796 | case ISD::FP_EXTEND: | |||
797 | case ISD::FSIN: | |||
798 | return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(), | |||
799 | GetNegatedExpression(Op.getOperand(0), DAG, | |||
800 | LegalOperations, Depth+1)); | |||
801 | case ISD::FP_ROUND: | |||
802 | return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(), | |||
803 | GetNegatedExpression(Op.getOperand(0), DAG, | |||
804 | LegalOperations, Depth+1), | |||
805 | Op.getOperand(1)); | |||
806 | } | |||
807 | } | |||
808 | ||||
809 | // APInts must be the same size for most operations, this helper | |||
810 | // function zero extends the shorter of the pair so that they match. | |||
811 | // We provide an Offset so that we can create bitwidths that won't overflow. | |||
812 | static void zeroExtendToMatch(APInt &LHS, APInt &RHS, unsigned Offset = 0) { | |||
813 | unsigned Bits = Offset + std::max(LHS.getBitWidth(), RHS.getBitWidth()); | |||
814 | LHS = LHS.zextOrSelf(Bits); | |||
815 | RHS = RHS.zextOrSelf(Bits); | |||
816 | } | |||
817 | ||||
818 | // Return true if this node is a setcc, or is a select_cc | |||
819 | // that selects between the target values used for true and false, making it | |||
820 | // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to | |||
821 | // the appropriate nodes based on the type of node we are checking. This | |||
822 | // simplifies life a bit for the callers. | |||
823 | bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS, | |||
824 | SDValue &CC) const { | |||
825 | if (N.getOpcode() == ISD::SETCC) { | |||
826 | LHS = N.getOperand(0); | |||
827 | RHS = N.getOperand(1); | |||
828 | CC = N.getOperand(2); | |||
829 | return true; | |||
830 | } | |||
831 | ||||
832 | if (N.getOpcode() != ISD::SELECT_CC || | |||
833 | !TLI.isConstTrueVal(N.getOperand(2).getNode()) || | |||
834 | !TLI.isConstFalseVal(N.getOperand(3).getNode())) | |||
835 | return false; | |||
836 | ||||
837 | if (TLI.getBooleanContents(N.getValueType()) == | |||
838 | TargetLowering::UndefinedBooleanContent) | |||
839 | return false; | |||
840 | ||||
841 | LHS = N.getOperand(0); | |||
842 | RHS = N.getOperand(1); | |||
843 | CC = N.getOperand(4); | |||
844 | return true; | |||
845 | } | |||
846 | ||||
847 | /// Return true if this is a SetCC-equivalent operation with only one use. | |||
848 | /// If this is true, it allows the users to invert the operation for free when | |||
849 | /// it is profitable to do so. | |||
850 | bool DAGCombiner::isOneUseSetCC(SDValue N) const { | |||
851 | SDValue N0, N1, N2; | |||
852 | if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse()) | |||
853 | return true; | |||
854 | return false; | |||
855 | } | |||
856 | ||||
857 | // \brief Returns the SDNode if it is a constant float BuildVector | |||
858 | // or constant float. | |||
859 | static SDNode *isConstantFPBuildVectorOrConstantFP(SDValue N) { | |||
860 | if (isa<ConstantFPSDNode>(N)) | |||
861 | return N.getNode(); | |||
862 | if (ISD::isBuildVectorOfConstantFPSDNodes(N.getNode())) | |||
863 | return N.getNode(); | |||
864 | return nullptr; | |||
865 | } | |||
866 | ||||
867 | // Determines if it is a constant integer or a build vector of constant | |||
868 | // integers (and undefs). | |||
869 | // Do not permit build vector implicit truncation. | |||
870 | static bool isConstantOrConstantVector(SDValue N, bool NoOpaques = false) { | |||
871 | if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N)) | |||
872 | return !(Const->isOpaque() && NoOpaques); | |||
873 | if (N.getOpcode() != ISD::BUILD_VECTOR) | |||
874 | return false; | |||
875 | unsigned BitWidth = N.getScalarValueSizeInBits(); | |||
876 | for (const SDValue &Op : N->op_values()) { | |||
877 | if (Op.isUndef()) | |||
878 | continue; | |||
879 | ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Op); | |||
880 | if (!Const || Const->getAPIntValue().getBitWidth() != BitWidth || | |||
881 | (Const->isOpaque() && NoOpaques)) | |||
882 | return false; | |||
883 | } | |||
884 | return true; | |||
885 | } | |||
886 | ||||
887 | // Determines if it is a constant null integer or a splatted vector of a | |||
888 | // constant null integer (with no undefs). | |||
889 | // Build vector implicit truncation is not an issue for null values. | |||
890 | static bool isNullConstantOrNullSplatConstant(SDValue N) { | |||
891 | if (ConstantSDNode *Splat = isConstOrConstSplat(N)) | |||
892 | return Splat->isNullValue(); | |||
893 | return false; | |||
894 | } | |||
895 | ||||
896 | // Determines if it is a constant integer of one or a splatted vector of a | |||
897 | // constant integer of one (with no undefs). | |||
898 | // Do not permit build vector implicit truncation. | |||
899 | static bool isOneConstantOrOneSplatConstant(SDValue N) { | |||
900 | unsigned BitWidth = N.getScalarValueSizeInBits(); | |||
901 | if (ConstantSDNode *Splat = isConstOrConstSplat(N)) | |||
902 | return Splat->isOne() && Splat->getAPIntValue().getBitWidth() == BitWidth; | |||
903 | return false; | |||
904 | } | |||
905 | ||||
906 | // Determines if it is a constant integer of all ones or a splatted vector of a | |||
907 | // constant integer of all ones (with no undefs). | |||
908 | // Do not permit build vector implicit truncation. | |||
909 | static bool isAllOnesConstantOrAllOnesSplatConstant(SDValue N) { | |||
910 | unsigned BitWidth = N.getScalarValueSizeInBits(); | |||
911 | if (ConstantSDNode *Splat = isConstOrConstSplat(N)) | |||
912 | return Splat->isAllOnesValue() && | |||
913 | Splat->getAPIntValue().getBitWidth() == BitWidth; | |||
914 | return false; | |||
915 | } | |||
916 | ||||
917 | // Determines if a BUILD_VECTOR is composed of all-constants possibly mixed with | |||
918 | // undef's. | |||
919 | static bool isAnyConstantBuildVector(const SDNode *N) { | |||
920 | return ISD::isBuildVectorOfConstantSDNodes(N) || | |||
921 | ISD::isBuildVectorOfConstantFPSDNodes(N); | |||
922 | } | |||
923 | ||||
924 | SDValue DAGCombiner::ReassociateOps(unsigned Opc, const SDLoc &DL, SDValue N0, | |||
925 | SDValue N1) { | |||
926 | EVT VT = N0.getValueType(); | |||
927 | if (N0.getOpcode() == Opc) { | |||
928 | if (SDNode *L = DAG.isConstantIntBuildVectorOrConstantInt(N0.getOperand(1))) { | |||
929 | if (SDNode *R = DAG.isConstantIntBuildVectorOrConstantInt(N1)) { | |||
930 | // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2)) | |||
931 | if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, DL, VT, L, R)) | |||
932 | return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode); | |||
933 | return SDValue(); | |||
934 | } | |||
935 | if (N0.hasOneUse()) { | |||
936 | // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one | |||
937 | // use | |||
938 | SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1); | |||
939 | if (!OpNode.getNode()) | |||
940 | return SDValue(); | |||
941 | AddToWorklist(OpNode.getNode()); | |||
942 | return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1)); | |||
943 | } | |||
944 | } | |||
945 | } | |||
946 | ||||
947 | if (N1.getOpcode() == Opc) { | |||
948 | if (SDNode *R = DAG.isConstantIntBuildVectorOrConstantInt(N1.getOperand(1))) { | |||
949 | if (SDNode *L = DAG.isConstantIntBuildVectorOrConstantInt(N0)) { | |||
950 | // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2)) | |||
951 | if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, DL, VT, R, L)) | |||
952 | return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode); | |||
953 | return SDValue(); | |||
954 | } | |||
955 | if (N1.hasOneUse()) { | |||
956 | // reassoc. (op x, (op y, c1)) -> (op (op x, y), c1) iff x+c1 has one | |||
957 | // use | |||
958 | SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0, N1.getOperand(0)); | |||
959 | if (!OpNode.getNode()) | |||
960 | return SDValue(); | |||
961 | AddToWorklist(OpNode.getNode()); | |||
962 | return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1)); | |||
963 | } | |||
964 | } | |||
965 | } | |||
966 | ||||
967 | return SDValue(); | |||
968 | } | |||
969 | ||||
970 | SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo, | |||
971 | bool AddTo) { | |||
972 | assert(N->getNumValues() == NumTo && "Broken CombineTo call!")(static_cast <bool> (N->getNumValues() == NumTo && "Broken CombineTo call!") ? void (0) : __assert_fail ("N->getNumValues() == NumTo && \"Broken CombineTo call!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 972, __extension__ __PRETTY_FUNCTION__)); | |||
973 | ++NodesCombined; | |||
974 | DEBUG(dbgs() << "\nReplacing.1 ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.1 "; N->dump (&DAG); dbgs() << "\nWith: "; To[0].getNode()->dump (&DAG); dbgs() << " and " << NumTo-1 << " other values\n"; } } while (false) | |||
975 | N->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.1 "; N->dump (&DAG); dbgs() << "\nWith: "; To[0].getNode()->dump (&DAG); dbgs() << " and " << NumTo-1 << " other values\n"; } } while (false) | |||
976 | dbgs() << "\nWith: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.1 "; N->dump (&DAG); dbgs() << "\nWith: "; To[0].getNode()->dump (&DAG); dbgs() << " and " << NumTo-1 << " other values\n"; } } while (false) | |||
977 | To[0].getNode()->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.1 "; N->dump (&DAG); dbgs() << "\nWith: "; To[0].getNode()->dump (&DAG); dbgs() << " and " << NumTo-1 << " other values\n"; } } while (false) | |||
978 | dbgs() << " and " << NumTo-1 << " other values\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.1 "; N->dump (&DAG); dbgs() << "\nWith: "; To[0].getNode()->dump (&DAG); dbgs() << " and " << NumTo-1 << " other values\n"; } } while (false); | |||
979 | for (unsigned i = 0, e = NumTo; i != e; ++i) | |||
980 | assert((!To[i].getNode() ||(static_cast <bool> ((!To[i].getNode() || N->getValueType (i) == To[i].getValueType()) && "Cannot combine value to value of different type!" ) ? void (0) : __assert_fail ("(!To[i].getNode() || N->getValueType(i) == To[i].getValueType()) && \"Cannot combine value to value of different type!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 982, __extension__ __PRETTY_FUNCTION__)) | |||
981 | N->getValueType(i) == To[i].getValueType()) &&(static_cast <bool> ((!To[i].getNode() || N->getValueType (i) == To[i].getValueType()) && "Cannot combine value to value of different type!" ) ? void (0) : __assert_fail ("(!To[i].getNode() || N->getValueType(i) == To[i].getValueType()) && \"Cannot combine value to value of different type!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 982, __extension__ __PRETTY_FUNCTION__)) | |||
982 | "Cannot combine value to value of different type!")(static_cast <bool> ((!To[i].getNode() || N->getValueType (i) == To[i].getValueType()) && "Cannot combine value to value of different type!" ) ? void (0) : __assert_fail ("(!To[i].getNode() || N->getValueType(i) == To[i].getValueType()) && \"Cannot combine value to value of different type!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 982, __extension__ __PRETTY_FUNCTION__)); | |||
983 | ||||
984 | WorklistRemover DeadNodes(*this); | |||
985 | DAG.ReplaceAllUsesWith(N, To); | |||
986 | if (AddTo) { | |||
987 | // Push the new nodes and any users onto the worklist | |||
988 | for (unsigned i = 0, e = NumTo; i != e; ++i) { | |||
989 | if (To[i].getNode()) { | |||
990 | AddToWorklist(To[i].getNode()); | |||
991 | AddUsersToWorklist(To[i].getNode()); | |||
992 | } | |||
993 | } | |||
994 | } | |||
995 | ||||
996 | // Finally, if the node is now dead, remove it from the graph. The node | |||
997 | // may not be dead if the replacement process recursively simplified to | |||
998 | // something else needing this node. | |||
999 | if (N->use_empty()) | |||
1000 | deleteAndRecombine(N); | |||
1001 | return SDValue(N, 0); | |||
1002 | } | |||
1003 | ||||
1004 | void DAGCombiner:: | |||
1005 | CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) { | |||
1006 | // Replace all uses. If any nodes become isomorphic to other nodes and | |||
1007 | // are deleted, make sure to remove them from our worklist. | |||
1008 | WorklistRemover DeadNodes(*this); | |||
1009 | DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New); | |||
1010 | ||||
1011 | // Push the new node and any (possibly new) users onto the worklist. | |||
1012 | AddToWorklist(TLO.New.getNode()); | |||
1013 | AddUsersToWorklist(TLO.New.getNode()); | |||
1014 | ||||
1015 | // Finally, if the node is now dead, remove it from the graph. The node | |||
1016 | // may not be dead if the replacement process recursively simplified to | |||
1017 | // something else needing this node. | |||
1018 | if (TLO.Old.getNode()->use_empty()) | |||
1019 | deleteAndRecombine(TLO.Old.getNode()); | |||
1020 | } | |||
1021 | ||||
1022 | /// Check the specified integer node value to see if it can be simplified or if | |||
1023 | /// things it uses can be simplified by bit propagation. If so, return true. | |||
1024 | bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) { | |||
1025 | TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations); | |||
1026 | KnownBits Known; | |||
1027 | if (!TLI.SimplifyDemandedBits(Op, Demanded, Known, TLO)) | |||
1028 | return false; | |||
1029 | ||||
1030 | // Revisit the node. | |||
1031 | AddToWorklist(Op.getNode()); | |||
1032 | ||||
1033 | // Replace the old value with the new one. | |||
1034 | ++NodesCombined; | |||
1035 | DEBUG(dbgs() << "\nReplacing.2 ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.2 "; TLO.Old.getNode ()->dump(&DAG); dbgs() << "\nWith: "; TLO.New.getNode ()->dump(&DAG); dbgs() << '\n'; } } while (false ) | |||
1036 | TLO.Old.getNode()->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.2 "; TLO.Old.getNode ()->dump(&DAG); dbgs() << "\nWith: "; TLO.New.getNode ()->dump(&DAG); dbgs() << '\n'; } } while (false ) | |||
1037 | dbgs() << "\nWith: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.2 "; TLO.Old.getNode ()->dump(&DAG); dbgs() << "\nWith: "; TLO.New.getNode ()->dump(&DAG); dbgs() << '\n'; } } while (false ) | |||
1038 | TLO.New.getNode()->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.2 "; TLO.Old.getNode ()->dump(&DAG); dbgs() << "\nWith: "; TLO.New.getNode ()->dump(&DAG); dbgs() << '\n'; } } while (false ) | |||
1039 | dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.2 "; TLO.Old.getNode ()->dump(&DAG); dbgs() << "\nWith: "; TLO.New.getNode ()->dump(&DAG); dbgs() << '\n'; } } while (false ); | |||
1040 | ||||
1041 | CommitTargetLoweringOpt(TLO); | |||
1042 | return true; | |||
1043 | } | |||
1044 | ||||
1045 | /// Check the specified vector node value to see if it can be simplified or | |||
1046 | /// if things it uses can be simplified as it only uses some of the elements. | |||
1047 | /// If so, return true. | |||
1048 | bool DAGCombiner::SimplifyDemandedVectorElts(SDValue Op, | |||
1049 | const APInt &Demanded) { | |||
1050 | TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations); | |||
1051 | APInt KnownUndef, KnownZero; | |||
1052 | if (!TLI.SimplifyDemandedVectorElts(Op, Demanded, KnownUndef, KnownZero, TLO)) | |||
1053 | return false; | |||
1054 | ||||
1055 | // Revisit the node. | |||
1056 | AddToWorklist(Op.getNode()); | |||
1057 | ||||
1058 | // Replace the old value with the new one. | |||
1059 | ++NodesCombined; | |||
1060 | DEBUG(dbgs() << "\nReplacing.2 "; TLO.Old.getNode()->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.2 "; TLO.Old.getNode ()->dump(&DAG); dbgs() << "\nWith: "; TLO.New.getNode ()->dump(&DAG); dbgs() << '\n'; } } while (false ) | |||
1061 | dbgs() << "\nWith: "; TLO.New.getNode()->dump(&DAG); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.2 "; TLO.Old.getNode ()->dump(&DAG); dbgs() << "\nWith: "; TLO.New.getNode ()->dump(&DAG); dbgs() << '\n'; } } while (false ); | |||
1062 | ||||
1063 | CommitTargetLoweringOpt(TLO); | |||
1064 | return true; | |||
1065 | } | |||
1066 | ||||
1067 | void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) { | |||
1068 | SDLoc DL(Load); | |||
1069 | EVT VT = Load->getValueType(0); | |||
1070 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, VT, SDValue(ExtLoad, 0)); | |||
1071 | ||||
1072 | DEBUG(dbgs() << "\nReplacing.9 ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.9 "; Load-> dump(&DAG); dbgs() << "\nWith: "; Trunc.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false) | |||
1073 | Load->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.9 "; Load-> dump(&DAG); dbgs() << "\nWith: "; Trunc.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false) | |||
1074 | dbgs() << "\nWith: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.9 "; Load-> dump(&DAG); dbgs() << "\nWith: "; Trunc.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false) | |||
1075 | Trunc.getNode()->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.9 "; Load-> dump(&DAG); dbgs() << "\nWith: "; Trunc.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false) | |||
1076 | dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.9 "; Load-> dump(&DAG); dbgs() << "\nWith: "; Trunc.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false); | |||
1077 | WorklistRemover DeadNodes(*this); | |||
1078 | DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc); | |||
1079 | DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1)); | |||
1080 | deleteAndRecombine(Load); | |||
1081 | AddToWorklist(Trunc.getNode()); | |||
1082 | } | |||
1083 | ||||
1084 | SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) { | |||
1085 | Replace = false; | |||
1086 | SDLoc DL(Op); | |||
1087 | if (ISD::isUNINDEXEDLoad(Op.getNode())) { | |||
1088 | LoadSDNode *LD = cast<LoadSDNode>(Op); | |||
1089 | EVT MemVT = LD->getMemoryVT(); | |||
1090 | ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD) | |||
1091 | ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD | |||
1092 | : ISD::EXTLOAD) | |||
1093 | : LD->getExtensionType(); | |||
1094 | Replace = true; | |||
1095 | return DAG.getExtLoad(ExtType, DL, PVT, | |||
1096 | LD->getChain(), LD->getBasePtr(), | |||
1097 | MemVT, LD->getMemOperand()); | |||
1098 | } | |||
1099 | ||||
1100 | unsigned Opc = Op.getOpcode(); | |||
1101 | switch (Opc) { | |||
1102 | default: break; | |||
1103 | case ISD::AssertSext: | |||
1104 | if (SDValue Op0 = SExtPromoteOperand(Op.getOperand(0), PVT)) | |||
1105 | return DAG.getNode(ISD::AssertSext, DL, PVT, Op0, Op.getOperand(1)); | |||
1106 | break; | |||
1107 | case ISD::AssertZext: | |||
1108 | if (SDValue Op0 = ZExtPromoteOperand(Op.getOperand(0), PVT)) | |||
1109 | return DAG.getNode(ISD::AssertZext, DL, PVT, Op0, Op.getOperand(1)); | |||
1110 | break; | |||
1111 | case ISD::Constant: { | |||
1112 | unsigned ExtOpc = | |||
1113 | Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; | |||
1114 | return DAG.getNode(ExtOpc, DL, PVT, Op); | |||
1115 | } | |||
1116 | } | |||
1117 | ||||
1118 | if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT)) | |||
1119 | return SDValue(); | |||
1120 | return DAG.getNode(ISD::ANY_EXTEND, DL, PVT, Op); | |||
1121 | } | |||
1122 | ||||
1123 | SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) { | |||
1124 | if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT)) | |||
1125 | return SDValue(); | |||
1126 | EVT OldVT = Op.getValueType(); | |||
1127 | SDLoc DL(Op); | |||
1128 | bool Replace = false; | |||
1129 | SDValue NewOp = PromoteOperand(Op, PVT, Replace); | |||
1130 | if (!NewOp.getNode()) | |||
1131 | return SDValue(); | |||
1132 | AddToWorklist(NewOp.getNode()); | |||
1133 | ||||
1134 | if (Replace) | |||
1135 | ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode()); | |||
1136 | return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, NewOp.getValueType(), NewOp, | |||
1137 | DAG.getValueType(OldVT)); | |||
1138 | } | |||
1139 | ||||
1140 | SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) { | |||
1141 | EVT OldVT = Op.getValueType(); | |||
1142 | SDLoc DL(Op); | |||
1143 | bool Replace = false; | |||
1144 | SDValue NewOp = PromoteOperand(Op, PVT, Replace); | |||
1145 | if (!NewOp.getNode()) | |||
1146 | return SDValue(); | |||
1147 | AddToWorklist(NewOp.getNode()); | |||
1148 | ||||
1149 | if (Replace) | |||
1150 | ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode()); | |||
1151 | return DAG.getZeroExtendInReg(NewOp, DL, OldVT); | |||
1152 | } | |||
1153 | ||||
1154 | /// Promote the specified integer binary operation if the target indicates it is | |||
1155 | /// beneficial. e.g. On x86, it's usually better to promote i16 operations to | |||
1156 | /// i32 since i16 instructions are longer. | |||
1157 | SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) { | |||
1158 | if (!LegalOperations) | |||
1159 | return SDValue(); | |||
1160 | ||||
1161 | EVT VT = Op.getValueType(); | |||
1162 | if (VT.isVector() || !VT.isInteger()) | |||
1163 | return SDValue(); | |||
1164 | ||||
1165 | // If operation type is 'undesirable', e.g. i16 on x86, consider | |||
1166 | // promoting it. | |||
1167 | unsigned Opc = Op.getOpcode(); | |||
1168 | if (TLI.isTypeDesirableForOp(Opc, VT)) | |||
1169 | return SDValue(); | |||
1170 | ||||
1171 | EVT PVT = VT; | |||
1172 | // Consult target whether it is a good idea to promote this operation and | |||
1173 | // what's the right type to promote it to. | |||
1174 | if (TLI.IsDesirableToPromoteOp(Op, PVT)) { | |||
1175 | assert(PVT != VT && "Don't know what type to promote to!")(static_cast <bool> (PVT != VT && "Don't know what type to promote to!" ) ? void (0) : __assert_fail ("PVT != VT && \"Don't know what type to promote to!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1175, __extension__ __PRETTY_FUNCTION__)); | |||
1176 | ||||
1177 | DEBUG(dbgs() << "\nPromoting "; Op.getNode()->dump(&DAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; Op.getNode( )->dump(&DAG); } } while (false); | |||
1178 | ||||
1179 | bool Replace0 = false; | |||
1180 | SDValue N0 = Op.getOperand(0); | |||
1181 | SDValue NN0 = PromoteOperand(N0, PVT, Replace0); | |||
1182 | ||||
1183 | bool Replace1 = false; | |||
1184 | SDValue N1 = Op.getOperand(1); | |||
1185 | SDValue NN1 = PromoteOperand(N1, PVT, Replace1); | |||
1186 | SDLoc DL(Op); | |||
1187 | ||||
1188 | SDValue RV = | |||
1189 | DAG.getNode(ISD::TRUNCATE, DL, VT, DAG.getNode(Opc, DL, PVT, NN0, NN1)); | |||
1190 | ||||
1191 | // We are always replacing N0/N1's use in N and only need | |||
1192 | // additional replacements if there are additional uses. | |||
1193 | Replace0 &= !N0->hasOneUse(); | |||
1194 | Replace1 &= (N0 != N1) && !N1->hasOneUse(); | |||
1195 | ||||
1196 | // Combine Op here so it is preserved past replacements. | |||
1197 | CombineTo(Op.getNode(), RV); | |||
1198 | ||||
1199 | // If operands have a use ordering, make sure we deal with | |||
1200 | // predecessor first. | |||
1201 | if (Replace0 && Replace1 && N0.getNode()->isPredecessorOf(N1.getNode())) { | |||
1202 | std::swap(N0, N1); | |||
1203 | std::swap(NN0, NN1); | |||
1204 | } | |||
1205 | ||||
1206 | if (Replace0) { | |||
1207 | AddToWorklist(NN0.getNode()); | |||
1208 | ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode()); | |||
1209 | } | |||
1210 | if (Replace1) { | |||
1211 | AddToWorklist(NN1.getNode()); | |||
1212 | ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode()); | |||
1213 | } | |||
1214 | return Op; | |||
1215 | } | |||
1216 | return SDValue(); | |||
1217 | } | |||
1218 | ||||
1219 | /// Promote the specified integer shift operation if the target indicates it is | |||
1220 | /// beneficial. e.g. On x86, it's usually better to promote i16 operations to | |||
1221 | /// i32 since i16 instructions are longer. | |||
1222 | SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) { | |||
1223 | if (!LegalOperations) | |||
1224 | return SDValue(); | |||
1225 | ||||
1226 | EVT VT = Op.getValueType(); | |||
1227 | if (VT.isVector() || !VT.isInteger()) | |||
1228 | return SDValue(); | |||
1229 | ||||
1230 | // If operation type is 'undesirable', e.g. i16 on x86, consider | |||
1231 | // promoting it. | |||
1232 | unsigned Opc = Op.getOpcode(); | |||
1233 | if (TLI.isTypeDesirableForOp(Opc, VT)) | |||
1234 | return SDValue(); | |||
1235 | ||||
1236 | EVT PVT = VT; | |||
1237 | // Consult target whether it is a good idea to promote this operation and | |||
1238 | // what's the right type to promote it to. | |||
1239 | if (TLI.IsDesirableToPromoteOp(Op, PVT)) { | |||
1240 | assert(PVT != VT && "Don't know what type to promote to!")(static_cast <bool> (PVT != VT && "Don't know what type to promote to!" ) ? void (0) : __assert_fail ("PVT != VT && \"Don't know what type to promote to!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1240, __extension__ __PRETTY_FUNCTION__)); | |||
1241 | ||||
1242 | DEBUG(dbgs() << "\nPromoting "; Op.getNode()->dump(&DAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; Op.getNode( )->dump(&DAG); } } while (false); | |||
1243 | ||||
1244 | bool Replace = false; | |||
1245 | SDValue N0 = Op.getOperand(0); | |||
1246 | SDValue N1 = Op.getOperand(1); | |||
1247 | if (Opc == ISD::SRA) | |||
1248 | N0 = SExtPromoteOperand(N0, PVT); | |||
1249 | else if (Opc == ISD::SRL) | |||
1250 | N0 = ZExtPromoteOperand(N0, PVT); | |||
1251 | else | |||
1252 | N0 = PromoteOperand(N0, PVT, Replace); | |||
1253 | ||||
1254 | if (!N0.getNode()) | |||
1255 | return SDValue(); | |||
1256 | ||||
1257 | SDLoc DL(Op); | |||
1258 | SDValue RV = | |||
1259 | DAG.getNode(ISD::TRUNCATE, DL, VT, DAG.getNode(Opc, DL, PVT, N0, N1)); | |||
1260 | ||||
1261 | AddToWorklist(N0.getNode()); | |||
1262 | if (Replace) | |||
1263 | ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode()); | |||
1264 | ||||
1265 | // Deal with Op being deleted. | |||
1266 | if (Op && Op.getOpcode() != ISD::DELETED_NODE) | |||
1267 | return RV; | |||
1268 | } | |||
1269 | return SDValue(); | |||
1270 | } | |||
1271 | ||||
1272 | SDValue DAGCombiner::PromoteExtend(SDValue Op) { | |||
1273 | if (!LegalOperations) | |||
1274 | return SDValue(); | |||
1275 | ||||
1276 | EVT VT = Op.getValueType(); | |||
1277 | if (VT.isVector() || !VT.isInteger()) | |||
1278 | return SDValue(); | |||
1279 | ||||
1280 | // If operation type is 'undesirable', e.g. i16 on x86, consider | |||
1281 | // promoting it. | |||
1282 | unsigned Opc = Op.getOpcode(); | |||
1283 | if (TLI.isTypeDesirableForOp(Opc, VT)) | |||
1284 | return SDValue(); | |||
1285 | ||||
1286 | EVT PVT = VT; | |||
1287 | // Consult target whether it is a good idea to promote this operation and | |||
1288 | // what's the right type to promote it to. | |||
1289 | if (TLI.IsDesirableToPromoteOp(Op, PVT)) { | |||
1290 | assert(PVT != VT && "Don't know what type to promote to!")(static_cast <bool> (PVT != VT && "Don't know what type to promote to!" ) ? void (0) : __assert_fail ("PVT != VT && \"Don't know what type to promote to!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1290, __extension__ __PRETTY_FUNCTION__)); | |||
1291 | // fold (aext (aext x)) -> (aext x) | |||
1292 | // fold (aext (zext x)) -> (zext x) | |||
1293 | // fold (aext (sext x)) -> (sext x) | |||
1294 | DEBUG(dbgs() << "\nPromoting ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; Op.getNode( )->dump(&DAG); } } while (false) | |||
1295 | Op.getNode()->dump(&DAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; Op.getNode( )->dump(&DAG); } } while (false); | |||
1296 | return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0)); | |||
1297 | } | |||
1298 | return SDValue(); | |||
1299 | } | |||
1300 | ||||
1301 | bool DAGCombiner::PromoteLoad(SDValue Op) { | |||
1302 | if (!LegalOperations) | |||
1303 | return false; | |||
1304 | ||||
1305 | if (!ISD::isUNINDEXEDLoad(Op.getNode())) | |||
1306 | return false; | |||
1307 | ||||
1308 | EVT VT = Op.getValueType(); | |||
1309 | if (VT.isVector() || !VT.isInteger()) | |||
1310 | return false; | |||
1311 | ||||
1312 | // If operation type is 'undesirable', e.g. i16 on x86, consider | |||
1313 | // promoting it. | |||
1314 | unsigned Opc = Op.getOpcode(); | |||
1315 | if (TLI.isTypeDesirableForOp(Opc, VT)) | |||
1316 | return false; | |||
1317 | ||||
1318 | EVT PVT = VT; | |||
1319 | // Consult target whether it is a good idea to promote this operation and | |||
1320 | // what's the right type to promote it to. | |||
1321 | if (TLI.IsDesirableToPromoteOp(Op, PVT)) { | |||
1322 | assert(PVT != VT && "Don't know what type to promote to!")(static_cast <bool> (PVT != VT && "Don't know what type to promote to!" ) ? void (0) : __assert_fail ("PVT != VT && \"Don't know what type to promote to!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1322, __extension__ __PRETTY_FUNCTION__)); | |||
1323 | ||||
1324 | SDLoc DL(Op); | |||
1325 | SDNode *N = Op.getNode(); | |||
1326 | LoadSDNode *LD = cast<LoadSDNode>(N); | |||
1327 | EVT MemVT = LD->getMemoryVT(); | |||
1328 | ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD) | |||
1329 | ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD | |||
1330 | : ISD::EXTLOAD) | |||
1331 | : LD->getExtensionType(); | |||
1332 | SDValue NewLD = DAG.getExtLoad(ExtType, DL, PVT, | |||
1333 | LD->getChain(), LD->getBasePtr(), | |||
1334 | MemVT, LD->getMemOperand()); | |||
1335 | SDValue Result = DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD); | |||
1336 | ||||
1337 | DEBUG(dbgs() << "\nPromoting ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; N->dump( &DAG); dbgs() << "\nTo: "; Result.getNode()->dump (&DAG); dbgs() << '\n'; } } while (false) | |||
1338 | N->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; N->dump( &DAG); dbgs() << "\nTo: "; Result.getNode()->dump (&DAG); dbgs() << '\n'; } } while (false) | |||
1339 | dbgs() << "\nTo: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; N->dump( &DAG); dbgs() << "\nTo: "; Result.getNode()->dump (&DAG); dbgs() << '\n'; } } while (false) | |||
1340 | Result.getNode()->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; N->dump( &DAG); dbgs() << "\nTo: "; Result.getNode()->dump (&DAG); dbgs() << '\n'; } } while (false) | |||
1341 | dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nPromoting "; N->dump( &DAG); dbgs() << "\nTo: "; Result.getNode()->dump (&DAG); dbgs() << '\n'; } } while (false); | |||
1342 | WorklistRemover DeadNodes(*this); | |||
1343 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result); | |||
1344 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1)); | |||
1345 | deleteAndRecombine(N); | |||
1346 | AddToWorklist(Result.getNode()); | |||
1347 | return true; | |||
1348 | } | |||
1349 | return false; | |||
1350 | } | |||
1351 | ||||
1352 | /// \brief Recursively delete a node which has no uses and any operands for | |||
1353 | /// which it is the only use. | |||
1354 | /// | |||
1355 | /// Note that this both deletes the nodes and removes them from the worklist. | |||
1356 | /// It also adds any nodes who have had a user deleted to the worklist as they | |||
1357 | /// may now have only one use and subject to other combines. | |||
1358 | bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) { | |||
1359 | if (!N->use_empty()) | |||
1360 | return false; | |||
1361 | ||||
1362 | SmallSetVector<SDNode *, 16> Nodes; | |||
1363 | Nodes.insert(N); | |||
1364 | do { | |||
1365 | N = Nodes.pop_back_val(); | |||
1366 | if (!N) | |||
1367 | continue; | |||
1368 | ||||
1369 | if (N->use_empty()) { | |||
1370 | for (const SDValue &ChildN : N->op_values()) | |||
1371 | Nodes.insert(ChildN.getNode()); | |||
1372 | ||||
1373 | removeFromWorklist(N); | |||
1374 | DAG.DeleteNode(N); | |||
1375 | } else { | |||
1376 | AddToWorklist(N); | |||
1377 | } | |||
1378 | } while (!Nodes.empty()); | |||
1379 | return true; | |||
1380 | } | |||
1381 | ||||
1382 | //===----------------------------------------------------------------------===// | |||
1383 | // Main DAG Combiner implementation | |||
1384 | //===----------------------------------------------------------------------===// | |||
1385 | ||||
1386 | void DAGCombiner::Run(CombineLevel AtLevel) { | |||
1387 | // set the instance variables, so that the various visit routines may use it. | |||
1388 | Level = AtLevel; | |||
1389 | LegalOperations = Level >= AfterLegalizeVectorOps; | |||
1390 | LegalTypes = Level >= AfterLegalizeTypes; | |||
1391 | ||||
1392 | // Add all the dag nodes to the worklist. | |||
1393 | for (SDNode &Node : DAG.allnodes()) | |||
1394 | AddToWorklist(&Node); | |||
1395 | ||||
1396 | // Create a dummy node (which is not added to allnodes), that adds a reference | |||
1397 | // to the root node, preventing it from being deleted, and tracking any | |||
1398 | // changes of the root. | |||
1399 | HandleSDNode Dummy(DAG.getRoot()); | |||
1400 | ||||
1401 | // While the worklist isn't empty, find a node and try to combine it. | |||
1402 | while (!WorklistMap.empty()) { | |||
1403 | SDNode *N; | |||
1404 | // The Worklist holds the SDNodes in order, but it may contain null entries. | |||
1405 | do { | |||
1406 | N = Worklist.pop_back_val(); | |||
1407 | } while (!N); | |||
1408 | ||||
1409 | bool GoodWorklistEntry = WorklistMap.erase(N); | |||
1410 | (void)GoodWorklistEntry; | |||
1411 | assert(GoodWorklistEntry &&(static_cast <bool> (GoodWorklistEntry && "Found a worklist entry without a corresponding map entry!" ) ? void (0) : __assert_fail ("GoodWorklistEntry && \"Found a worklist entry without a corresponding map entry!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1412, __extension__ __PRETTY_FUNCTION__)) | |||
1412 | "Found a worklist entry without a corresponding map entry!")(static_cast <bool> (GoodWorklistEntry && "Found a worklist entry without a corresponding map entry!" ) ? void (0) : __assert_fail ("GoodWorklistEntry && \"Found a worklist entry without a corresponding map entry!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1412, __extension__ __PRETTY_FUNCTION__)); | |||
1413 | ||||
1414 | // If N has no uses, it is dead. Make sure to revisit all N's operands once | |||
1415 | // N is deleted from the DAG, since they too may now be dead or may have a | |||
1416 | // reduced number of uses, allowing other xforms. | |||
1417 | if (recursivelyDeleteUnusedNodes(N)) | |||
1418 | continue; | |||
1419 | ||||
1420 | WorklistRemover DeadNodes(*this); | |||
1421 | ||||
1422 | // If this combine is running after legalizing the DAG, re-legalize any | |||
1423 | // nodes pulled off the worklist. | |||
1424 | if (Level == AfterLegalizeDAG) { | |||
1425 | SmallSetVector<SDNode *, 16> UpdatedNodes; | |||
1426 | bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes); | |||
1427 | ||||
1428 | for (SDNode *LN : UpdatedNodes) { | |||
1429 | AddToWorklist(LN); | |||
1430 | AddUsersToWorklist(LN); | |||
1431 | } | |||
1432 | if (!NIsValid) | |||
1433 | continue; | |||
1434 | } | |||
1435 | ||||
1436 | DEBUG(dbgs() << "\nCombining: "; N->dump(&DAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nCombining: "; N->dump (&DAG); } } while (false); | |||
1437 | ||||
1438 | // Add any operands of the new node which have not yet been combined to the | |||
1439 | // worklist as well. Because the worklist uniques things already, this | |||
1440 | // won't repeatedly process the same operand. | |||
1441 | CombinedNodes.insert(N); | |||
1442 | for (const SDValue &ChildN : N->op_values()) | |||
1443 | if (!CombinedNodes.count(ChildN.getNode())) | |||
1444 | AddToWorklist(ChildN.getNode()); | |||
1445 | ||||
1446 | SDValue RV = combine(N); | |||
1447 | ||||
1448 | if (!RV.getNode()) | |||
1449 | continue; | |||
1450 | ||||
1451 | ++NodesCombined; | |||
1452 | ||||
1453 | // If we get back the same node we passed in, rather than a new node or | |||
1454 | // zero, we know that the node must have defined multiple values and | |||
1455 | // CombineTo was used. Since CombineTo takes care of the worklist | |||
1456 | // mechanics for us, we have no work to do in this case. | |||
1457 | if (RV.getNode() == N) | |||
1458 | continue; | |||
1459 | ||||
1460 | assert(N->getOpcode() != ISD::DELETED_NODE &&(static_cast <bool> (N->getOpcode() != ISD::DELETED_NODE && RV.getOpcode() != ISD::DELETED_NODE && "Node was deleted but visit returned new node!" ) ? void (0) : __assert_fail ("N->getOpcode() != ISD::DELETED_NODE && RV.getOpcode() != ISD::DELETED_NODE && \"Node was deleted but visit returned new node!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1462, __extension__ __PRETTY_FUNCTION__)) | |||
1461 | RV.getOpcode() != ISD::DELETED_NODE &&(static_cast <bool> (N->getOpcode() != ISD::DELETED_NODE && RV.getOpcode() != ISD::DELETED_NODE && "Node was deleted but visit returned new node!" ) ? void (0) : __assert_fail ("N->getOpcode() != ISD::DELETED_NODE && RV.getOpcode() != ISD::DELETED_NODE && \"Node was deleted but visit returned new node!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1462, __extension__ __PRETTY_FUNCTION__)) | |||
1462 | "Node was deleted but visit returned new node!")(static_cast <bool> (N->getOpcode() != ISD::DELETED_NODE && RV.getOpcode() != ISD::DELETED_NODE && "Node was deleted but visit returned new node!" ) ? void (0) : __assert_fail ("N->getOpcode() != ISD::DELETED_NODE && RV.getOpcode() != ISD::DELETED_NODE && \"Node was deleted but visit returned new node!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1462, __extension__ __PRETTY_FUNCTION__)); | |||
1463 | ||||
1464 | DEBUG(dbgs() << " ... into: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << " ... into: "; RV.getNode() ->dump(&DAG); } } while (false) | |||
1465 | RV.getNode()->dump(&DAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << " ... into: "; RV.getNode() ->dump(&DAG); } } while (false); | |||
1466 | ||||
1467 | if (N->getNumValues() == RV.getNode()->getNumValues()) | |||
1468 | DAG.ReplaceAllUsesWith(N, RV.getNode()); | |||
1469 | else { | |||
1470 | assert(N->getValueType(0) == RV.getValueType() &&(static_cast <bool> (N->getValueType(0) == RV.getValueType () && N->getNumValues() == 1 && "Type mismatch" ) ? void (0) : __assert_fail ("N->getValueType(0) == RV.getValueType() && N->getNumValues() == 1 && \"Type mismatch\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1471, __extension__ __PRETTY_FUNCTION__)) | |||
1471 | N->getNumValues() == 1 && "Type mismatch")(static_cast <bool> (N->getValueType(0) == RV.getValueType () && N->getNumValues() == 1 && "Type mismatch" ) ? void (0) : __assert_fail ("N->getValueType(0) == RV.getValueType() && N->getNumValues() == 1 && \"Type mismatch\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1471, __extension__ __PRETTY_FUNCTION__)); | |||
1472 | DAG.ReplaceAllUsesWith(N, &RV); | |||
1473 | } | |||
1474 | ||||
1475 | // Push the new node and any users onto the worklist | |||
1476 | AddToWorklist(RV.getNode()); | |||
1477 | AddUsersToWorklist(RV.getNode()); | |||
1478 | ||||
1479 | // Finally, if the node is now dead, remove it from the graph. The node | |||
1480 | // may not be dead if the replacement process recursively simplified to | |||
1481 | // something else needing this node. This will also take care of adding any | |||
1482 | // operands which have lost a user to the worklist. | |||
1483 | recursivelyDeleteUnusedNodes(N); | |||
1484 | } | |||
1485 | ||||
1486 | // If the root changed (e.g. it was a dead load, update the root). | |||
1487 | DAG.setRoot(Dummy.getValue()); | |||
1488 | DAG.RemoveDeadNodes(); | |||
1489 | } | |||
1490 | ||||
1491 | SDValue DAGCombiner::visit(SDNode *N) { | |||
1492 | switch (N->getOpcode()) { | |||
1493 | default: break; | |||
1494 | case ISD::TokenFactor: return visitTokenFactor(N); | |||
1495 | case ISD::MERGE_VALUES: return visitMERGE_VALUES(N); | |||
1496 | case ISD::ADD: return visitADD(N); | |||
1497 | case ISD::SUB: return visitSUB(N); | |||
1498 | case ISD::ADDC: return visitADDC(N); | |||
1499 | case ISD::UADDO: return visitUADDO(N); | |||
1500 | case ISD::SUBC: return visitSUBC(N); | |||
1501 | case ISD::USUBO: return visitUSUBO(N); | |||
1502 | case ISD::ADDE: return visitADDE(N); | |||
1503 | case ISD::ADDCARRY: return visitADDCARRY(N); | |||
1504 | case ISD::SUBE: return visitSUBE(N); | |||
1505 | case ISD::SUBCARRY: return visitSUBCARRY(N); | |||
1506 | case ISD::MUL: return visitMUL(N); | |||
1507 | case ISD::SDIV: return visitSDIV(N); | |||
1508 | case ISD::UDIV: return visitUDIV(N); | |||
1509 | case ISD::SREM: | |||
1510 | case ISD::UREM: return visitREM(N); | |||
1511 | case ISD::MULHU: return visitMULHU(N); | |||
1512 | case ISD::MULHS: return visitMULHS(N); | |||
1513 | case ISD::SMUL_LOHI: return visitSMUL_LOHI(N); | |||
1514 | case ISD::UMUL_LOHI: return visitUMUL_LOHI(N); | |||
1515 | case ISD::SMULO: return visitSMULO(N); | |||
1516 | case ISD::UMULO: return visitUMULO(N); | |||
1517 | case ISD::SMIN: | |||
1518 | case ISD::SMAX: | |||
1519 | case ISD::UMIN: | |||
1520 | case ISD::UMAX: return visitIMINMAX(N); | |||
1521 | case ISD::AND: return visitAND(N); | |||
1522 | case ISD::OR: return visitOR(N); | |||
1523 | case ISD::XOR: return visitXOR(N); | |||
1524 | case ISD::SHL: return visitSHL(N); | |||
1525 | case ISD::SRA: return visitSRA(N); | |||
1526 | case ISD::SRL: return visitSRL(N); | |||
1527 | case ISD::ROTR: | |||
1528 | case ISD::ROTL: return visitRotate(N); | |||
1529 | case ISD::ABS: return visitABS(N); | |||
1530 | case ISD::BSWAP: return visitBSWAP(N); | |||
1531 | case ISD::BITREVERSE: return visitBITREVERSE(N); | |||
1532 | case ISD::CTLZ: return visitCTLZ(N); | |||
1533 | case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N); | |||
1534 | case ISD::CTTZ: return visitCTTZ(N); | |||
1535 | case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N); | |||
1536 | case ISD::CTPOP: return visitCTPOP(N); | |||
1537 | case ISD::SELECT: return visitSELECT(N); | |||
1538 | case ISD::VSELECT: return visitVSELECT(N); | |||
1539 | case ISD::SELECT_CC: return visitSELECT_CC(N); | |||
1540 | case ISD::SETCC: return visitSETCC(N); | |||
1541 | case ISD::SETCCE: return visitSETCCE(N); | |||
1542 | case ISD::SETCCCARRY: return visitSETCCCARRY(N); | |||
1543 | case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N); | |||
1544 | case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N); | |||
1545 | case ISD::ANY_EXTEND: return visitANY_EXTEND(N); | |||
1546 | case ISD::AssertSext: | |||
1547 | case ISD::AssertZext: return visitAssertExt(N); | |||
1548 | case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N); | |||
1549 | case ISD::SIGN_EXTEND_VECTOR_INREG: return visitSIGN_EXTEND_VECTOR_INREG(N); | |||
1550 | case ISD::ZERO_EXTEND_VECTOR_INREG: return visitZERO_EXTEND_VECTOR_INREG(N); | |||
1551 | case ISD::TRUNCATE: return visitTRUNCATE(N); | |||
1552 | case ISD::BITCAST: return visitBITCAST(N); | |||
1553 | case ISD::BUILD_PAIR: return visitBUILD_PAIR(N); | |||
1554 | case ISD::FADD: return visitFADD(N); | |||
1555 | case ISD::FSUB: return visitFSUB(N); | |||
1556 | case ISD::FMUL: return visitFMUL(N); | |||
1557 | case ISD::FMA: return visitFMA(N); | |||
1558 | case ISD::FDIV: return visitFDIV(N); | |||
1559 | case ISD::FREM: return visitFREM(N); | |||
1560 | case ISD::FSQRT: return visitFSQRT(N); | |||
1561 | case ISD::FCOPYSIGN: return visitFCOPYSIGN(N); | |||
1562 | case ISD::SINT_TO_FP: return visitSINT_TO_FP(N); | |||
1563 | case ISD::UINT_TO_FP: return visitUINT_TO_FP(N); | |||
1564 | case ISD::FP_TO_SINT: return visitFP_TO_SINT(N); | |||
1565 | case ISD::FP_TO_UINT: return visitFP_TO_UINT(N); | |||
1566 | case ISD::FP_ROUND: return visitFP_ROUND(N); | |||
1567 | case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N); | |||
1568 | case ISD::FP_EXTEND: return visitFP_EXTEND(N); | |||
1569 | case ISD::FNEG: return visitFNEG(N); | |||
1570 | case ISD::FABS: return visitFABS(N); | |||
1571 | case ISD::FFLOOR: return visitFFLOOR(N); | |||
1572 | case ISD::FMINNUM: return visitFMINNUM(N); | |||
1573 | case ISD::FMAXNUM: return visitFMAXNUM(N); | |||
1574 | case ISD::FCEIL: return visitFCEIL(N); | |||
1575 | case ISD::FTRUNC: return visitFTRUNC(N); | |||
1576 | case ISD::BRCOND: return visitBRCOND(N); | |||
1577 | case ISD::BR_CC: return visitBR_CC(N); | |||
1578 | case ISD::LOAD: return visitLOAD(N); | |||
1579 | case ISD::STORE: return visitSTORE(N); | |||
1580 | case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N); | |||
1581 | case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N); | |||
1582 | case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N); | |||
1583 | case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N); | |||
1584 | case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N); | |||
1585 | case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N); | |||
1586 | case ISD::SCALAR_TO_VECTOR: return visitSCALAR_TO_VECTOR(N); | |||
1587 | case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N); | |||
1588 | case ISD::MGATHER: return visitMGATHER(N); | |||
1589 | case ISD::MLOAD: return visitMLOAD(N); | |||
1590 | case ISD::MSCATTER: return visitMSCATTER(N); | |||
1591 | case ISD::MSTORE: return visitMSTORE(N); | |||
1592 | case ISD::FP_TO_FP16: return visitFP_TO_FP16(N); | |||
1593 | case ISD::FP16_TO_FP: return visitFP16_TO_FP(N); | |||
1594 | } | |||
1595 | return SDValue(); | |||
1596 | } | |||
1597 | ||||
1598 | SDValue DAGCombiner::combine(SDNode *N) { | |||
1599 | SDValue RV = visit(N); | |||
1600 | ||||
1601 | // If nothing happened, try a target-specific DAG combine. | |||
1602 | if (!RV.getNode()) { | |||
1603 | assert(N->getOpcode() != ISD::DELETED_NODE &&(static_cast <bool> (N->getOpcode() != ISD::DELETED_NODE && "Node was deleted but visit returned NULL!") ? void (0) : __assert_fail ("N->getOpcode() != ISD::DELETED_NODE && \"Node was deleted but visit returned NULL!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1604, __extension__ __PRETTY_FUNCTION__)) | |||
1604 | "Node was deleted but visit returned NULL!")(static_cast <bool> (N->getOpcode() != ISD::DELETED_NODE && "Node was deleted but visit returned NULL!") ? void (0) : __assert_fail ("N->getOpcode() != ISD::DELETED_NODE && \"Node was deleted but visit returned NULL!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1604, __extension__ __PRETTY_FUNCTION__)); | |||
1605 | ||||
1606 | if (N->getOpcode() >= ISD::BUILTIN_OP_END || | |||
1607 | TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) { | |||
1608 | ||||
1609 | // Expose the DAG combiner to the target combiner impls. | |||
1610 | TargetLowering::DAGCombinerInfo | |||
1611 | DagCombineInfo(DAG, Level, false, this); | |||
1612 | ||||
1613 | RV = TLI.PerformDAGCombine(N, DagCombineInfo); | |||
1614 | } | |||
1615 | } | |||
1616 | ||||
1617 | // If nothing happened still, try promoting the operation. | |||
1618 | if (!RV.getNode()) { | |||
1619 | switch (N->getOpcode()) { | |||
1620 | default: break; | |||
1621 | case ISD::ADD: | |||
1622 | case ISD::SUB: | |||
1623 | case ISD::MUL: | |||
1624 | case ISD::AND: | |||
1625 | case ISD::OR: | |||
1626 | case ISD::XOR: | |||
1627 | RV = PromoteIntBinOp(SDValue(N, 0)); | |||
1628 | break; | |||
1629 | case ISD::SHL: | |||
1630 | case ISD::SRA: | |||
1631 | case ISD::SRL: | |||
1632 | RV = PromoteIntShiftOp(SDValue(N, 0)); | |||
1633 | break; | |||
1634 | case ISD::SIGN_EXTEND: | |||
1635 | case ISD::ZERO_EXTEND: | |||
1636 | case ISD::ANY_EXTEND: | |||
1637 | RV = PromoteExtend(SDValue(N, 0)); | |||
1638 | break; | |||
1639 | case ISD::LOAD: | |||
1640 | if (PromoteLoad(SDValue(N, 0))) | |||
1641 | RV = SDValue(N, 0); | |||
1642 | break; | |||
1643 | } | |||
1644 | } | |||
1645 | ||||
1646 | // If N is a commutative binary node, try eliminate it if the commuted | |||
1647 | // version is already present in the DAG. | |||
1648 | if (!RV.getNode() && TLI.isCommutativeBinOp(N->getOpcode()) && | |||
1649 | N->getNumValues() == 1) { | |||
1650 | SDValue N0 = N->getOperand(0); | |||
1651 | SDValue N1 = N->getOperand(1); | |||
1652 | ||||
1653 | // Constant operands are canonicalized to RHS. | |||
1654 | if (N0 != N1 && (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1))) { | |||
1655 | SDValue Ops[] = {N1, N0}; | |||
1656 | SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops, | |||
1657 | N->getFlags()); | |||
1658 | if (CSENode) | |||
1659 | return SDValue(CSENode, 0); | |||
1660 | } | |||
1661 | } | |||
1662 | ||||
1663 | return RV; | |||
1664 | } | |||
1665 | ||||
1666 | /// Given a node, return its input chain if it has one, otherwise return a null | |||
1667 | /// sd operand. | |||
1668 | static SDValue getInputChainForNode(SDNode *N) { | |||
1669 | if (unsigned NumOps = N->getNumOperands()) { | |||
1670 | if (N->getOperand(0).getValueType() == MVT::Other) | |||
1671 | return N->getOperand(0); | |||
1672 | if (N->getOperand(NumOps-1).getValueType() == MVT::Other) | |||
1673 | return N->getOperand(NumOps-1); | |||
1674 | for (unsigned i = 1; i < NumOps-1; ++i) | |||
1675 | if (N->getOperand(i).getValueType() == MVT::Other) | |||
1676 | return N->getOperand(i); | |||
1677 | } | |||
1678 | return SDValue(); | |||
1679 | } | |||
1680 | ||||
1681 | SDValue DAGCombiner::visitTokenFactor(SDNode *N) { | |||
1682 | // If N has two operands, where one has an input chain equal to the other, | |||
1683 | // the 'other' chain is redundant. | |||
1684 | if (N->getNumOperands() == 2) { | |||
1685 | if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1)) | |||
1686 | return N->getOperand(0); | |||
1687 | if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0)) | |||
1688 | return N->getOperand(1); | |||
1689 | } | |||
1690 | ||||
1691 | SmallVector<SDNode *, 8> TFs; // List of token factors to visit. | |||
1692 | SmallVector<SDValue, 8> Ops; // Ops for replacing token factor. | |||
1693 | SmallPtrSet<SDNode*, 16> SeenOps; | |||
1694 | bool Changed = false; // If we should replace this token factor. | |||
1695 | ||||
1696 | // Start out with this token factor. | |||
1697 | TFs.push_back(N); | |||
1698 | ||||
1699 | // Iterate through token factors. The TFs grows when new token factors are | |||
1700 | // encountered. | |||
1701 | for (unsigned i = 0; i < TFs.size(); ++i) { | |||
1702 | SDNode *TF = TFs[i]; | |||
1703 | ||||
1704 | // Check each of the operands. | |||
1705 | for (const SDValue &Op : TF->op_values()) { | |||
1706 | switch (Op.getOpcode()) { | |||
1707 | case ISD::EntryToken: | |||
1708 | // Entry tokens don't need to be added to the list. They are | |||
1709 | // redundant. | |||
1710 | Changed = true; | |||
1711 | break; | |||
1712 | ||||
1713 | case ISD::TokenFactor: | |||
1714 | if (Op.hasOneUse() && !is_contained(TFs, Op.getNode())) { | |||
1715 | // Queue up for processing. | |||
1716 | TFs.push_back(Op.getNode()); | |||
1717 | // Clean up in case the token factor is removed. | |||
1718 | AddToWorklist(Op.getNode()); | |||
1719 | Changed = true; | |||
1720 | break; | |||
1721 | } | |||
1722 | LLVM_FALLTHROUGH[[clang::fallthrough]]; | |||
1723 | ||||
1724 | default: | |||
1725 | // Only add if it isn't already in the list. | |||
1726 | if (SeenOps.insert(Op.getNode()).second) | |||
1727 | Ops.push_back(Op); | |||
1728 | else | |||
1729 | Changed = true; | |||
1730 | break; | |||
1731 | } | |||
1732 | } | |||
1733 | } | |||
1734 | ||||
1735 | // Remove Nodes that are chained to another node in the list. Do so | |||
1736 | // by walking up chains breath-first stopping when we've seen | |||
1737 | // another operand. In general we must climb to the EntryNode, but we can exit | |||
1738 | // early if we find all remaining work is associated with just one operand as | |||
1739 | // no further pruning is possible. | |||
1740 | ||||
1741 | // List of nodes to search through and original Ops from which they originate. | |||
1742 | SmallVector<std::pair<SDNode *, unsigned>, 8> Worklist; | |||
1743 | SmallVector<unsigned, 8> OpWorkCount; // Count of work for each Op. | |||
1744 | SmallPtrSet<SDNode *, 16> SeenChains; | |||
1745 | bool DidPruneOps = false; | |||
1746 | ||||
1747 | unsigned NumLeftToConsider = 0; | |||
1748 | for (const SDValue &Op : Ops) { | |||
1749 | Worklist.push_back(std::make_pair(Op.getNode(), NumLeftToConsider++)); | |||
1750 | OpWorkCount.push_back(1); | |||
1751 | } | |||
1752 | ||||
1753 | auto AddToWorklist = [&](unsigned CurIdx, SDNode *Op, unsigned OpNumber) { | |||
1754 | // If this is an Op, we can remove the op from the list. Remark any | |||
1755 | // search associated with it as from the current OpNumber. | |||
1756 | if (SeenOps.count(Op) != 0) { | |||
1757 | Changed = true; | |||
1758 | DidPruneOps = true; | |||
1759 | unsigned OrigOpNumber = 0; | |||
1760 | while (OrigOpNumber < Ops.size() && Ops[OrigOpNumber].getNode() != Op) | |||
1761 | OrigOpNumber++; | |||
1762 | assert((OrigOpNumber != Ops.size()) &&(static_cast <bool> ((OrigOpNumber != Ops.size()) && "expected to find TokenFactor Operand") ? void (0) : __assert_fail ("(OrigOpNumber != Ops.size()) && \"expected to find TokenFactor Operand\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1763, __extension__ __PRETTY_FUNCTION__)) | |||
1763 | "expected to find TokenFactor Operand")(static_cast <bool> ((OrigOpNumber != Ops.size()) && "expected to find TokenFactor Operand") ? void (0) : __assert_fail ("(OrigOpNumber != Ops.size()) && \"expected to find TokenFactor Operand\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1763, __extension__ __PRETTY_FUNCTION__)); | |||
1764 | // Re-mark worklist from OrigOpNumber to OpNumber | |||
1765 | for (unsigned i = CurIdx + 1; i < Worklist.size(); ++i) { | |||
1766 | if (Worklist[i].second == OrigOpNumber) { | |||
1767 | Worklist[i].second = OpNumber; | |||
1768 | } | |||
1769 | } | |||
1770 | OpWorkCount[OpNumber] += OpWorkCount[OrigOpNumber]; | |||
1771 | OpWorkCount[OrigOpNumber] = 0; | |||
1772 | NumLeftToConsider--; | |||
1773 | } | |||
1774 | // Add if it's a new chain | |||
1775 | if (SeenChains.insert(Op).second) { | |||
1776 | OpWorkCount[OpNumber]++; | |||
1777 | Worklist.push_back(std::make_pair(Op, OpNumber)); | |||
1778 | } | |||
1779 | }; | |||
1780 | ||||
1781 | for (unsigned i = 0; i < Worklist.size() && i < 1024; ++i) { | |||
1782 | // We need at least be consider at least 2 Ops to prune. | |||
1783 | if (NumLeftToConsider <= 1) | |||
1784 | break; | |||
1785 | auto CurNode = Worklist[i].first; | |||
1786 | auto CurOpNumber = Worklist[i].second; | |||
1787 | assert((OpWorkCount[CurOpNumber] > 0) &&(static_cast <bool> ((OpWorkCount[CurOpNumber] > 0) && "Node should not appear in worklist") ? void (0) : __assert_fail ("(OpWorkCount[CurOpNumber] > 0) && \"Node should not appear in worklist\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1788, __extension__ __PRETTY_FUNCTION__)) | |||
1788 | "Node should not appear in worklist")(static_cast <bool> ((OpWorkCount[CurOpNumber] > 0) && "Node should not appear in worklist") ? void (0) : __assert_fail ("(OpWorkCount[CurOpNumber] > 0) && \"Node should not appear in worklist\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1788, __extension__ __PRETTY_FUNCTION__)); | |||
1789 | switch (CurNode->getOpcode()) { | |||
1790 | case ISD::EntryToken: | |||
1791 | // Hitting EntryToken is the only way for the search to terminate without | |||
1792 | // hitting | |||
1793 | // another operand's search. Prevent us from marking this operand | |||
1794 | // considered. | |||
1795 | NumLeftToConsider++; | |||
1796 | break; | |||
1797 | case ISD::TokenFactor: | |||
1798 | for (const SDValue &Op : CurNode->op_values()) | |||
1799 | AddToWorklist(i, Op.getNode(), CurOpNumber); | |||
1800 | break; | |||
1801 | case ISD::CopyFromReg: | |||
1802 | case ISD::CopyToReg: | |||
1803 | AddToWorklist(i, CurNode->getOperand(0).getNode(), CurOpNumber); | |||
1804 | break; | |||
1805 | default: | |||
1806 | if (auto *MemNode = dyn_cast<MemSDNode>(CurNode)) | |||
1807 | AddToWorklist(i, MemNode->getChain().getNode(), CurOpNumber); | |||
1808 | break; | |||
1809 | } | |||
1810 | OpWorkCount[CurOpNumber]--; | |||
1811 | if (OpWorkCount[CurOpNumber] == 0) | |||
1812 | NumLeftToConsider--; | |||
1813 | } | |||
1814 | ||||
1815 | // If we've changed things around then replace token factor. | |||
1816 | if (Changed) { | |||
1817 | SDValue Result; | |||
1818 | if (Ops.empty()) { | |||
1819 | // The entry token is the only possible outcome. | |||
1820 | Result = DAG.getEntryNode(); | |||
1821 | } else { | |||
1822 | if (DidPruneOps) { | |||
1823 | SmallVector<SDValue, 8> PrunedOps; | |||
1824 | // | |||
1825 | for (const SDValue &Op : Ops) { | |||
1826 | if (SeenChains.count(Op.getNode()) == 0) | |||
1827 | PrunedOps.push_back(Op); | |||
1828 | } | |||
1829 | Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, PrunedOps); | |||
1830 | } else { | |||
1831 | Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops); | |||
1832 | } | |||
1833 | } | |||
1834 | return Result; | |||
1835 | } | |||
1836 | return SDValue(); | |||
1837 | } | |||
1838 | ||||
1839 | /// MERGE_VALUES can always be eliminated. | |||
1840 | SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) { | |||
1841 | WorklistRemover DeadNodes(*this); | |||
1842 | // Replacing results may cause a different MERGE_VALUES to suddenly | |||
1843 | // be CSE'd with N, and carry its uses with it. Iterate until no | |||
1844 | // uses remain, to ensure that the node can be safely deleted. | |||
1845 | // First add the users of this node to the work list so that they | |||
1846 | // can be tried again once they have new operands. | |||
1847 | AddUsersToWorklist(N); | |||
1848 | do { | |||
1849 | for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) | |||
1850 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i)); | |||
1851 | } while (!N->use_empty()); | |||
1852 | deleteAndRecombine(N); | |||
1853 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
1854 | } | |||
1855 | ||||
1856 | /// If \p N is a ConstantSDNode with isOpaque() == false return it casted to a | |||
1857 | /// ConstantSDNode pointer else nullptr. | |||
1858 | static ConstantSDNode *getAsNonOpaqueConstant(SDValue N) { | |||
1859 | ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N); | |||
1860 | return Const != nullptr && !Const->isOpaque() ? Const : nullptr; | |||
1861 | } | |||
1862 | ||||
1863 | SDValue DAGCombiner::foldBinOpIntoSelect(SDNode *BO) { | |||
1864 | auto BinOpcode = BO->getOpcode(); | |||
1865 | assert((BinOpcode == ISD::ADD || BinOpcode == ISD::SUB ||(static_cast <bool> ((BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && "Unexpected binary operator" ) ? void (0) : __assert_fail ("(BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && \"Unexpected binary operator\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1874, __extension__ __PRETTY_FUNCTION__)) | |||
1866 | BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV ||(static_cast <bool> ((BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && "Unexpected binary operator" ) ? void (0) : __assert_fail ("(BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && \"Unexpected binary operator\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1874, __extension__ __PRETTY_FUNCTION__)) | |||
1867 | BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM ||(static_cast <bool> ((BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && "Unexpected binary operator" ) ? void (0) : __assert_fail ("(BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && \"Unexpected binary operator\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1874, __extension__ __PRETTY_FUNCTION__)) | |||
1868 | BinOpcode == ISD::UREM || BinOpcode == ISD::AND ||(static_cast <bool> ((BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && "Unexpected binary operator" ) ? void (0) : __assert_fail ("(BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && \"Unexpected binary operator\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1874, __extension__ __PRETTY_FUNCTION__)) | |||
1869 | BinOpcode == ISD::OR || BinOpcode == ISD::XOR ||(static_cast <bool> ((BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && "Unexpected binary operator" ) ? void (0) : __assert_fail ("(BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && \"Unexpected binary operator\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1874, __extension__ __PRETTY_FUNCTION__)) | |||
1870 | BinOpcode == ISD::SHL || BinOpcode == ISD::SRL ||(static_cast <bool> ((BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && "Unexpected binary operator" ) ? void (0) : __assert_fail ("(BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && \"Unexpected binary operator\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1874, __extension__ __PRETTY_FUNCTION__)) | |||
1871 | BinOpcode == ISD::SRA || BinOpcode == ISD::FADD ||(static_cast <bool> ((BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && "Unexpected binary operator" ) ? void (0) : __assert_fail ("(BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && \"Unexpected binary operator\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1874, __extension__ __PRETTY_FUNCTION__)) | |||
1872 | BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL ||(static_cast <bool> ((BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && "Unexpected binary operator" ) ? void (0) : __assert_fail ("(BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && \"Unexpected binary operator\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1874, __extension__ __PRETTY_FUNCTION__)) | |||
1873 | BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) &&(static_cast <bool> ((BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && "Unexpected binary operator" ) ? void (0) : __assert_fail ("(BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && \"Unexpected binary operator\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1874, __extension__ __PRETTY_FUNCTION__)) | |||
1874 | "Unexpected binary operator")(static_cast <bool> ((BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && "Unexpected binary operator" ) ? void (0) : __assert_fail ("(BinOpcode == ISD::ADD || BinOpcode == ISD::SUB || BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV || BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM || BinOpcode == ISD::UREM || BinOpcode == ISD::AND || BinOpcode == ISD::OR || BinOpcode == ISD::XOR || BinOpcode == ISD::SHL || BinOpcode == ISD::SRL || BinOpcode == ISD::SRA || BinOpcode == ISD::FADD || BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL || BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) && \"Unexpected binary operator\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 1874, __extension__ __PRETTY_FUNCTION__)); | |||
1875 | ||||
1876 | // Bail out if any constants are opaque because we can't constant fold those. | |||
1877 | SDValue C1 = BO->getOperand(1); | |||
1878 | if (!isConstantOrConstantVector(C1, true) && | |||
1879 | !isConstantFPBuildVectorOrConstantFP(C1)) | |||
1880 | return SDValue(); | |||
1881 | ||||
1882 | // Don't do this unless the old select is going away. We want to eliminate the | |||
1883 | // binary operator, not replace a binop with a select. | |||
1884 | // TODO: Handle ISD::SELECT_CC. | |||
1885 | SDValue Sel = BO->getOperand(0); | |||
1886 | if (Sel.getOpcode() != ISD::SELECT || !Sel.hasOneUse()) | |||
1887 | return SDValue(); | |||
1888 | ||||
1889 | SDValue CT = Sel.getOperand(1); | |||
1890 | if (!isConstantOrConstantVector(CT, true) && | |||
1891 | !isConstantFPBuildVectorOrConstantFP(CT)) | |||
1892 | return SDValue(); | |||
1893 | ||||
1894 | SDValue CF = Sel.getOperand(2); | |||
1895 | if (!isConstantOrConstantVector(CF, true) && | |||
1896 | !isConstantFPBuildVectorOrConstantFP(CF)) | |||
1897 | return SDValue(); | |||
1898 | ||||
1899 | // We have a select-of-constants followed by a binary operator with a | |||
1900 | // constant. Eliminate the binop by pulling the constant math into the select. | |||
1901 | // Example: add (select Cond, CT, CF), C1 --> select Cond, CT + C1, CF + C1 | |||
1902 | EVT VT = Sel.getValueType(); | |||
1903 | SDLoc DL(Sel); | |||
1904 | SDValue NewCT = DAG.getNode(BinOpcode, DL, VT, CT, C1); | |||
1905 | if (!NewCT.isUndef() && | |||
1906 | !isConstantOrConstantVector(NewCT, true) && | |||
1907 | !isConstantFPBuildVectorOrConstantFP(NewCT)) | |||
1908 | return SDValue(); | |||
1909 | ||||
1910 | SDValue NewCF = DAG.getNode(BinOpcode, DL, VT, CF, C1); | |||
1911 | if (!NewCF.isUndef() && | |||
1912 | !isConstantOrConstantVector(NewCF, true) && | |||
1913 | !isConstantFPBuildVectorOrConstantFP(NewCF)) | |||
1914 | return SDValue(); | |||
1915 | ||||
1916 | return DAG.getSelect(DL, VT, Sel.getOperand(0), NewCT, NewCF); | |||
1917 | } | |||
1918 | ||||
1919 | SDValue DAGCombiner::visitADD(SDNode *N) { | |||
1920 | SDValue N0 = N->getOperand(0); | |||
1921 | SDValue N1 = N->getOperand(1); | |||
1922 | EVT VT = N0.getValueType(); | |||
1923 | SDLoc DL(N); | |||
1924 | ||||
1925 | // fold vector ops | |||
1926 | if (VT.isVector()) { | |||
1927 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
1928 | return FoldedVOp; | |||
1929 | ||||
1930 | // fold (add x, 0) -> x, vector edition | |||
1931 | if (ISD::isBuildVectorAllZeros(N1.getNode())) | |||
1932 | return N0; | |||
1933 | if (ISD::isBuildVectorAllZeros(N0.getNode())) | |||
1934 | return N1; | |||
1935 | } | |||
1936 | ||||
1937 | // fold (add x, undef) -> undef | |||
1938 | if (N0.isUndef()) | |||
1939 | return N0; | |||
1940 | ||||
1941 | if (N1.isUndef()) | |||
1942 | return N1; | |||
1943 | ||||
1944 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) { | |||
1945 | // canonicalize constant to RHS | |||
1946 | if (!DAG.isConstantIntBuildVectorOrConstantInt(N1)) | |||
1947 | return DAG.getNode(ISD::ADD, DL, VT, N1, N0); | |||
1948 | // fold (add c1, c2) -> c1+c2 | |||
1949 | return DAG.FoldConstantArithmetic(ISD::ADD, DL, VT, N0.getNode(), | |||
1950 | N1.getNode()); | |||
1951 | } | |||
1952 | ||||
1953 | // fold (add x, 0) -> x | |||
1954 | if (isNullConstant(N1)) | |||
1955 | return N0; | |||
1956 | ||||
1957 | if (isConstantOrConstantVector(N1, /* NoOpaque */ true)) { | |||
1958 | // fold ((c1-A)+c2) -> (c1+c2)-A | |||
1959 | if (N0.getOpcode() == ISD::SUB && | |||
1960 | isConstantOrConstantVector(N0.getOperand(0), /* NoOpaque */ true)) { | |||
1961 | // FIXME: Adding 2 constants should be handled by FoldConstantArithmetic. | |||
1962 | return DAG.getNode(ISD::SUB, DL, VT, | |||
1963 | DAG.getNode(ISD::ADD, DL, VT, N1, N0.getOperand(0)), | |||
1964 | N0.getOperand(1)); | |||
1965 | } | |||
1966 | ||||
1967 | // add (sext i1 X), 1 -> zext (not i1 X) | |||
1968 | // We don't transform this pattern: | |||
1969 | // add (zext i1 X), -1 -> sext (not i1 X) | |||
1970 | // because most (?) targets generate better code for the zext form. | |||
1971 | if (N0.getOpcode() == ISD::SIGN_EXTEND && N0.hasOneUse() && | |||
1972 | isOneConstantOrOneSplatConstant(N1)) { | |||
1973 | SDValue X = N0.getOperand(0); | |||
1974 | if ((!LegalOperations || | |||
1975 | (TLI.isOperationLegal(ISD::XOR, X.getValueType()) && | |||
1976 | TLI.isOperationLegal(ISD::ZERO_EXTEND, VT))) && | |||
1977 | X.getScalarValueSizeInBits() == 1) { | |||
1978 | SDValue Not = DAG.getNOT(DL, X, X.getValueType()); | |||
1979 | return DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Not); | |||
1980 | } | |||
1981 | } | |||
1982 | ||||
1983 | // Undo the add -> or combine to merge constant offsets from a frame index. | |||
1984 | if (N0.getOpcode() == ISD::OR && | |||
1985 | isa<FrameIndexSDNode>(N0.getOperand(0)) && | |||
1986 | isa<ConstantSDNode>(N0.getOperand(1)) && | |||
1987 | DAG.haveNoCommonBitsSet(N0.getOperand(0), N0.getOperand(1))) { | |||
1988 | SDValue Add0 = DAG.getNode(ISD::ADD, DL, VT, N1, N0.getOperand(1)); | |||
1989 | return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), Add0); | |||
1990 | } | |||
1991 | } | |||
1992 | ||||
1993 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
1994 | return NewSel; | |||
1995 | ||||
1996 | // reassociate add | |||
1997 | if (SDValue RADD = ReassociateOps(ISD::ADD, DL, N0, N1)) | |||
1998 | return RADD; | |||
1999 | ||||
2000 | // fold ((0-A) + B) -> B-A | |||
2001 | if (N0.getOpcode() == ISD::SUB && | |||
2002 | isNullConstantOrNullSplatConstant(N0.getOperand(0))) | |||
2003 | return DAG.getNode(ISD::SUB, DL, VT, N1, N0.getOperand(1)); | |||
2004 | ||||
2005 | // fold (A + (0-B)) -> A-B | |||
2006 | if (N1.getOpcode() == ISD::SUB && | |||
2007 | isNullConstantOrNullSplatConstant(N1.getOperand(0))) | |||
2008 | return DAG.getNode(ISD::SUB, DL, VT, N0, N1.getOperand(1)); | |||
2009 | ||||
2010 | // fold (A+(B-A)) -> B | |||
2011 | if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1)) | |||
2012 | return N1.getOperand(0); | |||
2013 | ||||
2014 | // fold ((B-A)+A) -> B | |||
2015 | if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1)) | |||
2016 | return N0.getOperand(0); | |||
2017 | ||||
2018 | // fold (A+(B-(A+C))) to (B-C) | |||
2019 | if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD && | |||
2020 | N0 == N1.getOperand(1).getOperand(0)) | |||
2021 | return DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(0), | |||
2022 | N1.getOperand(1).getOperand(1)); | |||
2023 | ||||
2024 | // fold (A+(B-(C+A))) to (B-C) | |||
2025 | if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD && | |||
2026 | N0 == N1.getOperand(1).getOperand(1)) | |||
2027 | return DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(0), | |||
2028 | N1.getOperand(1).getOperand(0)); | |||
2029 | ||||
2030 | // fold (A+((B-A)+or-C)) to (B+or-C) | |||
2031 | if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) && | |||
2032 | N1.getOperand(0).getOpcode() == ISD::SUB && | |||
2033 | N0 == N1.getOperand(0).getOperand(1)) | |||
2034 | return DAG.getNode(N1.getOpcode(), DL, VT, N1.getOperand(0).getOperand(0), | |||
2035 | N1.getOperand(1)); | |||
2036 | ||||
2037 | // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant | |||
2038 | if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) { | |||
2039 | SDValue N00 = N0.getOperand(0); | |||
2040 | SDValue N01 = N0.getOperand(1); | |||
2041 | SDValue N10 = N1.getOperand(0); | |||
2042 | SDValue N11 = N1.getOperand(1); | |||
2043 | ||||
2044 | if (isConstantOrConstantVector(N00) || isConstantOrConstantVector(N10)) | |||
2045 | return DAG.getNode(ISD::SUB, DL, VT, | |||
2046 | DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10), | |||
2047 | DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11)); | |||
2048 | } | |||
2049 | ||||
2050 | if (SimplifyDemandedBits(SDValue(N, 0))) | |||
2051 | return SDValue(N, 0); | |||
2052 | ||||
2053 | // fold (a+b) -> (a|b) iff a and b share no bits. | |||
2054 | if ((!LegalOperations || TLI.isOperationLegal(ISD::OR, VT)) && | |||
2055 | DAG.haveNoCommonBitsSet(N0, N1)) | |||
2056 | return DAG.getNode(ISD::OR, DL, VT, N0, N1); | |||
2057 | ||||
2058 | if (SDValue Combined = visitADDLike(N0, N1, N)) | |||
2059 | return Combined; | |||
2060 | ||||
2061 | if (SDValue Combined = visitADDLike(N1, N0, N)) | |||
2062 | return Combined; | |||
2063 | ||||
2064 | return SDValue(); | |||
2065 | } | |||
2066 | ||||
2067 | static SDValue getAsCarry(const TargetLowering &TLI, SDValue V) { | |||
2068 | bool Masked = false; | |||
2069 | ||||
2070 | // First, peel away TRUNCATE/ZERO_EXTEND/AND nodes due to legalization. | |||
2071 | while (true) { | |||
2072 | if (V.getOpcode() == ISD::TRUNCATE || V.getOpcode() == ISD::ZERO_EXTEND) { | |||
2073 | V = V.getOperand(0); | |||
2074 | continue; | |||
2075 | } | |||
2076 | ||||
2077 | if (V.getOpcode() == ISD::AND && isOneConstant(V.getOperand(1))) { | |||
2078 | Masked = true; | |||
2079 | V = V.getOperand(0); | |||
2080 | continue; | |||
2081 | } | |||
2082 | ||||
2083 | break; | |||
2084 | } | |||
2085 | ||||
2086 | // If this is not a carry, return. | |||
2087 | if (V.getResNo() != 1) | |||
2088 | return SDValue(); | |||
2089 | ||||
2090 | if (V.getOpcode() != ISD::ADDCARRY && V.getOpcode() != ISD::SUBCARRY && | |||
2091 | V.getOpcode() != ISD::UADDO && V.getOpcode() != ISD::USUBO) | |||
2092 | return SDValue(); | |||
2093 | ||||
2094 | // If the result is masked, then no matter what kind of bool it is we can | |||
2095 | // return. If it isn't, then we need to make sure the bool type is either 0 or | |||
2096 | // 1 and not other values. | |||
2097 | if (Masked || | |||
2098 | TLI.getBooleanContents(V.getValueType()) == | |||
2099 | TargetLoweringBase::ZeroOrOneBooleanContent) | |||
2100 | return V; | |||
2101 | ||||
2102 | return SDValue(); | |||
2103 | } | |||
2104 | ||||
2105 | SDValue DAGCombiner::visitADDLike(SDValue N0, SDValue N1, SDNode *LocReference) { | |||
2106 | EVT VT = N0.getValueType(); | |||
2107 | SDLoc DL(LocReference); | |||
2108 | ||||
2109 | // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n)) | |||
2110 | if (N1.getOpcode() == ISD::SHL && N1.getOperand(0).getOpcode() == ISD::SUB && | |||
2111 | isNullConstantOrNullSplatConstant(N1.getOperand(0).getOperand(0))) | |||
2112 | return DAG.getNode(ISD::SUB, DL, VT, N0, | |||
2113 | DAG.getNode(ISD::SHL, DL, VT, | |||
2114 | N1.getOperand(0).getOperand(1), | |||
2115 | N1.getOperand(1))); | |||
2116 | ||||
2117 | if (N1.getOpcode() == ISD::AND) { | |||
2118 | SDValue AndOp0 = N1.getOperand(0); | |||
2119 | unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0); | |||
2120 | unsigned DestBits = VT.getScalarSizeInBits(); | |||
2121 | ||||
2122 | // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x)) | |||
2123 | // and similar xforms where the inner op is either ~0 or 0. | |||
2124 | if (NumSignBits == DestBits && | |||
2125 | isOneConstantOrOneSplatConstant(N1->getOperand(1))) | |||
2126 | return DAG.getNode(ISD::SUB, DL, VT, N0, AndOp0); | |||
2127 | } | |||
2128 | ||||
2129 | // add (sext i1), X -> sub X, (zext i1) | |||
2130 | if (N0.getOpcode() == ISD::SIGN_EXTEND && | |||
2131 | N0.getOperand(0).getValueType() == MVT::i1 && | |||
2132 | !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) { | |||
2133 | SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)); | |||
2134 | return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt); | |||
2135 | } | |||
2136 | ||||
2137 | // add X, (sextinreg Y i1) -> sub X, (and Y 1) | |||
2138 | if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) { | |||
2139 | VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1)); | |||
2140 | if (TN->getVT() == MVT::i1) { | |||
2141 | SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0), | |||
2142 | DAG.getConstant(1, DL, VT)); | |||
2143 | return DAG.getNode(ISD::SUB, DL, VT, N0, ZExt); | |||
2144 | } | |||
2145 | } | |||
2146 | ||||
2147 | // (add X, (addcarry Y, 0, Carry)) -> (addcarry X, Y, Carry) | |||
2148 | if (N1.getOpcode() == ISD::ADDCARRY && isNullConstant(N1.getOperand(1)) && | |||
2149 | N1.getResNo() == 0) | |||
2150 | return DAG.getNode(ISD::ADDCARRY, DL, N1->getVTList(), | |||
2151 | N0, N1.getOperand(0), N1.getOperand(2)); | |||
2152 | ||||
2153 | // (add X, Carry) -> (addcarry X, 0, Carry) | |||
2154 | if (TLI.isOperationLegalOrCustom(ISD::ADDCARRY, VT)) | |||
2155 | if (SDValue Carry = getAsCarry(TLI, N1)) | |||
2156 | return DAG.getNode(ISD::ADDCARRY, DL, | |||
2157 | DAG.getVTList(VT, Carry.getValueType()), N0, | |||
2158 | DAG.getConstant(0, DL, VT), Carry); | |||
2159 | ||||
2160 | return SDValue(); | |||
2161 | } | |||
2162 | ||||
2163 | SDValue DAGCombiner::visitADDC(SDNode *N) { | |||
2164 | SDValue N0 = N->getOperand(0); | |||
2165 | SDValue N1 = N->getOperand(1); | |||
2166 | EVT VT = N0.getValueType(); | |||
2167 | SDLoc DL(N); | |||
2168 | ||||
2169 | // If the flag result is dead, turn this into an ADD. | |||
2170 | if (!N->hasAnyUseOfValue(1)) | |||
2171 | return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), | |||
2172 | DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); | |||
2173 | ||||
2174 | // canonicalize constant to RHS. | |||
2175 | ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); | |||
2176 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); | |||
2177 | if (N0C && !N1C) | |||
2178 | return DAG.getNode(ISD::ADDC, DL, N->getVTList(), N1, N0); | |||
2179 | ||||
2180 | // fold (addc x, 0) -> x + no carry out | |||
2181 | if (isNullConstant(N1)) | |||
2182 | return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, | |||
2183 | DL, MVT::Glue)); | |||
2184 | ||||
2185 | // If it cannot overflow, transform into an add. | |||
2186 | if (DAG.computeOverflowKind(N0, N1) == SelectionDAG::OFK_Never) | |||
2187 | return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), | |||
2188 | DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); | |||
2189 | ||||
2190 | return SDValue(); | |||
2191 | } | |||
2192 | ||||
2193 | SDValue DAGCombiner::visitUADDO(SDNode *N) { | |||
2194 | SDValue N0 = N->getOperand(0); | |||
2195 | SDValue N1 = N->getOperand(1); | |||
2196 | EVT VT = N0.getValueType(); | |||
2197 | if (VT.isVector()) | |||
2198 | return SDValue(); | |||
2199 | ||||
2200 | EVT CarryVT = N->getValueType(1); | |||
2201 | SDLoc DL(N); | |||
2202 | ||||
2203 | // If the flag result is dead, turn this into an ADD. | |||
2204 | if (!N->hasAnyUseOfValue(1)) | |||
2205 | return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), | |||
2206 | DAG.getUNDEF(CarryVT)); | |||
2207 | ||||
2208 | // canonicalize constant to RHS. | |||
2209 | ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); | |||
2210 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); | |||
2211 | if (N0C && !N1C) | |||
2212 | return DAG.getNode(ISD::UADDO, DL, N->getVTList(), N1, N0); | |||
2213 | ||||
2214 | // fold (uaddo x, 0) -> x + no carry out | |||
2215 | if (isNullConstant(N1)) | |||
2216 | return CombineTo(N, N0, DAG.getConstant(0, DL, CarryVT)); | |||
2217 | ||||
2218 | // If it cannot overflow, transform into an add. | |||
2219 | if (DAG.computeOverflowKind(N0, N1) == SelectionDAG::OFK_Never) | |||
2220 | return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), | |||
2221 | DAG.getConstant(0, DL, CarryVT)); | |||
2222 | ||||
2223 | if (SDValue Combined = visitUADDOLike(N0, N1, N)) | |||
2224 | return Combined; | |||
2225 | ||||
2226 | if (SDValue Combined = visitUADDOLike(N1, N0, N)) | |||
2227 | return Combined; | |||
2228 | ||||
2229 | return SDValue(); | |||
2230 | } | |||
2231 | ||||
2232 | SDValue DAGCombiner::visitUADDOLike(SDValue N0, SDValue N1, SDNode *N) { | |||
2233 | auto VT = N0.getValueType(); | |||
2234 | ||||
2235 | // (uaddo X, (addcarry Y, 0, Carry)) -> (addcarry X, Y, Carry) | |||
2236 | // If Y + 1 cannot overflow. | |||
2237 | if (N1.getOpcode() == ISD::ADDCARRY && isNullConstant(N1.getOperand(1))) { | |||
2238 | SDValue Y = N1.getOperand(0); | |||
2239 | SDValue One = DAG.getConstant(1, SDLoc(N), Y.getValueType()); | |||
2240 | if (DAG.computeOverflowKind(Y, One) == SelectionDAG::OFK_Never) | |||
2241 | return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), N0, Y, | |||
2242 | N1.getOperand(2)); | |||
2243 | } | |||
2244 | ||||
2245 | // (uaddo X, Carry) -> (addcarry X, 0, Carry) | |||
2246 | if (TLI.isOperationLegalOrCustom(ISD::ADDCARRY, VT)) | |||
2247 | if (SDValue Carry = getAsCarry(TLI, N1)) | |||
2248 | return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), N0, | |||
2249 | DAG.getConstant(0, SDLoc(N), VT), Carry); | |||
2250 | ||||
2251 | return SDValue(); | |||
2252 | } | |||
2253 | ||||
2254 | SDValue DAGCombiner::visitADDE(SDNode *N) { | |||
2255 | SDValue N0 = N->getOperand(0); | |||
2256 | SDValue N1 = N->getOperand(1); | |||
2257 | SDValue CarryIn = N->getOperand(2); | |||
2258 | ||||
2259 | // canonicalize constant to RHS | |||
2260 | ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); | |||
2261 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); | |||
2262 | if (N0C && !N1C) | |||
2263 | return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(), | |||
2264 | N1, N0, CarryIn); | |||
2265 | ||||
2266 | // fold (adde x, y, false) -> (addc x, y) | |||
2267 | if (CarryIn.getOpcode() == ISD::CARRY_FALSE) | |||
2268 | return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1); | |||
2269 | ||||
2270 | return SDValue(); | |||
2271 | } | |||
2272 | ||||
2273 | SDValue DAGCombiner::visitADDCARRY(SDNode *N) { | |||
2274 | SDValue N0 = N->getOperand(0); | |||
2275 | SDValue N1 = N->getOperand(1); | |||
2276 | SDValue CarryIn = N->getOperand(2); | |||
2277 | SDLoc DL(N); | |||
2278 | ||||
2279 | // canonicalize constant to RHS | |||
2280 | ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); | |||
2281 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); | |||
2282 | if (N0C && !N1C) | |||
2283 | return DAG.getNode(ISD::ADDCARRY, DL, N->getVTList(), N1, N0, CarryIn); | |||
2284 | ||||
2285 | // fold (addcarry x, y, false) -> (uaddo x, y) | |||
2286 | if (isNullConstant(CarryIn)) | |||
2287 | return DAG.getNode(ISD::UADDO, DL, N->getVTList(), N0, N1); | |||
2288 | ||||
2289 | // fold (addcarry 0, 0, X) -> (and (ext/trunc X), 1) and no carry. | |||
2290 | if (isNullConstant(N0) && isNullConstant(N1)) { | |||
2291 | EVT VT = N0.getValueType(); | |||
2292 | EVT CarryVT = CarryIn.getValueType(); | |||
2293 | SDValue CarryExt = DAG.getBoolExtOrTrunc(CarryIn, DL, VT, CarryVT); | |||
2294 | AddToWorklist(CarryExt.getNode()); | |||
2295 | return CombineTo(N, DAG.getNode(ISD::AND, DL, VT, CarryExt, | |||
2296 | DAG.getConstant(1, DL, VT)), | |||
2297 | DAG.getConstant(0, DL, CarryVT)); | |||
2298 | } | |||
2299 | ||||
2300 | if (SDValue Combined = visitADDCARRYLike(N0, N1, CarryIn, N)) | |||
2301 | return Combined; | |||
2302 | ||||
2303 | if (SDValue Combined = visitADDCARRYLike(N1, N0, CarryIn, N)) | |||
2304 | return Combined; | |||
2305 | ||||
2306 | return SDValue(); | |||
2307 | } | |||
2308 | ||||
2309 | SDValue DAGCombiner::visitADDCARRYLike(SDValue N0, SDValue N1, SDValue CarryIn, | |||
2310 | SDNode *N) { | |||
2311 | // Iff the flag result is dead: | |||
2312 | // (addcarry (add|uaddo X, Y), 0, Carry) -> (addcarry X, Y, Carry) | |||
2313 | if ((N0.getOpcode() == ISD::ADD || | |||
2314 | (N0.getOpcode() == ISD::UADDO && N0.getResNo() == 0)) && | |||
2315 | isNullConstant(N1) && !N->hasAnyUseOfValue(1)) | |||
2316 | return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), | |||
2317 | N0.getOperand(0), N0.getOperand(1), CarryIn); | |||
2318 | ||||
2319 | /** | |||
2320 | * When one of the addcarry argument is itself a carry, we may be facing | |||
2321 | * a diamond carry propagation. In which case we try to transform the DAG | |||
2322 | * to ensure linear carry propagation if that is possible. | |||
2323 | * | |||
2324 | * We are trying to get: | |||
2325 | * (addcarry X, 0, (addcarry A, B, Z):Carry) | |||
2326 | */ | |||
2327 | if (auto Y = getAsCarry(TLI, N1)) { | |||
2328 | /** | |||
2329 | * (uaddo A, B) | |||
2330 | * / \ | |||
2331 | * Carry Sum | |||
2332 | * | \ | |||
2333 | * | (addcarry *, 0, Z) | |||
2334 | * | / | |||
2335 | * \ Carry | |||
2336 | * | / | |||
2337 | * (addcarry X, *, *) | |||
2338 | */ | |||
2339 | if (Y.getOpcode() == ISD::UADDO && | |||
2340 | CarryIn.getResNo() == 1 && | |||
2341 | CarryIn.getOpcode() == ISD::ADDCARRY && | |||
2342 | isNullConstant(CarryIn.getOperand(1)) && | |||
2343 | CarryIn.getOperand(0) == Y.getValue(0)) { | |||
2344 | auto NewY = DAG.getNode(ISD::ADDCARRY, SDLoc(N), Y->getVTList(), | |||
2345 | Y.getOperand(0), Y.getOperand(1), | |||
2346 | CarryIn.getOperand(2)); | |||
2347 | AddToWorklist(NewY.getNode()); | |||
2348 | return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), N0, | |||
2349 | DAG.getConstant(0, SDLoc(N), N0.getValueType()), | |||
2350 | NewY.getValue(1)); | |||
2351 | } | |||
2352 | } | |||
2353 | ||||
2354 | return SDValue(); | |||
2355 | } | |||
2356 | ||||
2357 | // Since it may not be valid to emit a fold to zero for vector initializers | |||
2358 | // check if we can before folding. | |||
2359 | static SDValue tryFoldToZero(const SDLoc &DL, const TargetLowering &TLI, EVT VT, | |||
2360 | SelectionDAG &DAG, bool LegalOperations, | |||
2361 | bool LegalTypes) { | |||
2362 | if (!VT.isVector()) | |||
2363 | return DAG.getConstant(0, DL, VT); | |||
2364 | if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) | |||
2365 | return DAG.getConstant(0, DL, VT); | |||
2366 | return SDValue(); | |||
2367 | } | |||
2368 | ||||
2369 | SDValue DAGCombiner::visitSUB(SDNode *N) { | |||
2370 | SDValue N0 = N->getOperand(0); | |||
2371 | SDValue N1 = N->getOperand(1); | |||
2372 | EVT VT = N0.getValueType(); | |||
2373 | SDLoc DL(N); | |||
2374 | ||||
2375 | // fold vector ops | |||
2376 | if (VT.isVector()) { | |||
2377 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
2378 | return FoldedVOp; | |||
2379 | ||||
2380 | // fold (sub x, 0) -> x, vector edition | |||
2381 | if (ISD::isBuildVectorAllZeros(N1.getNode())) | |||
2382 | return N0; | |||
2383 | } | |||
2384 | ||||
2385 | // fold (sub x, x) -> 0 | |||
2386 | // FIXME: Refactor this and xor and other similar operations together. | |||
2387 | if (N0 == N1) | |||
2388 | return tryFoldToZero(DL, TLI, VT, DAG, LegalOperations, LegalTypes); | |||
2389 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
2390 | DAG.isConstantIntBuildVectorOrConstantInt(N1)) { | |||
2391 | // fold (sub c1, c2) -> c1-c2 | |||
2392 | return DAG.FoldConstantArithmetic(ISD::SUB, DL, VT, N0.getNode(), | |||
2393 | N1.getNode()); | |||
2394 | } | |||
2395 | ||||
2396 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
2397 | return NewSel; | |||
2398 | ||||
2399 | ConstantSDNode *N1C = getAsNonOpaqueConstant(N1); | |||
2400 | ||||
2401 | // fold (sub x, c) -> (add x, -c) | |||
2402 | if (N1C) { | |||
2403 | return DAG.getNode(ISD::ADD, DL, VT, N0, | |||
2404 | DAG.getConstant(-N1C->getAPIntValue(), DL, VT)); | |||
2405 | } | |||
2406 | ||||
2407 | if (isNullConstantOrNullSplatConstant(N0)) { | |||
2408 | unsigned BitWidth = VT.getScalarSizeInBits(); | |||
2409 | // Right-shifting everything out but the sign bit followed by negation is | |||
2410 | // the same as flipping arithmetic/logical shift type without the negation: | |||
2411 | // -(X >>u 31) -> (X >>s 31) | |||
2412 | // -(X >>s 31) -> (X >>u 31) | |||
2413 | if (N1->getOpcode() == ISD::SRA || N1->getOpcode() == ISD::SRL) { | |||
2414 | ConstantSDNode *ShiftAmt = isConstOrConstSplat(N1.getOperand(1)); | |||
2415 | if (ShiftAmt && ShiftAmt->getZExtValue() == BitWidth - 1) { | |||
2416 | auto NewSh = N1->getOpcode() == ISD::SRA ? ISD::SRL : ISD::SRA; | |||
2417 | if (!LegalOperations || TLI.isOperationLegal(NewSh, VT)) | |||
2418 | return DAG.getNode(NewSh, DL, VT, N1.getOperand(0), N1.getOperand(1)); | |||
2419 | } | |||
2420 | } | |||
2421 | ||||
2422 | // 0 - X --> 0 if the sub is NUW. | |||
2423 | if (N->getFlags().hasNoUnsignedWrap()) | |||
2424 | return N0; | |||
2425 | ||||
2426 | if (DAG.MaskedValueIsZero(N1, ~APInt::getSignMask(BitWidth))) { | |||
2427 | // N1 is either 0 or the minimum signed value. If the sub is NSW, then | |||
2428 | // N1 must be 0 because negating the minimum signed value is undefined. | |||
2429 | if (N->getFlags().hasNoSignedWrap()) | |||
2430 | return N0; | |||
2431 | ||||
2432 | // 0 - X --> X if X is 0 or the minimum signed value. | |||
2433 | return N1; | |||
2434 | } | |||
2435 | } | |||
2436 | ||||
2437 | // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) | |||
2438 | if (isAllOnesConstantOrAllOnesSplatConstant(N0)) | |||
2439 | return DAG.getNode(ISD::XOR, DL, VT, N1, N0); | |||
2440 | ||||
2441 | // fold A-(A-B) -> B | |||
2442 | if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0)) | |||
2443 | return N1.getOperand(1); | |||
2444 | ||||
2445 | // fold (A+B)-A -> B | |||
2446 | if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1) | |||
2447 | return N0.getOperand(1); | |||
2448 | ||||
2449 | // fold (A+B)-B -> A | |||
2450 | if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1) | |||
2451 | return N0.getOperand(0); | |||
2452 | ||||
2453 | // fold C2-(A+C1) -> (C2-C1)-A | |||
2454 | if (N1.getOpcode() == ISD::ADD) { | |||
2455 | SDValue N11 = N1.getOperand(1); | |||
2456 | if (isConstantOrConstantVector(N0, /* NoOpaques */ true) && | |||
2457 | isConstantOrConstantVector(N11, /* NoOpaques */ true)) { | |||
2458 | SDValue NewC = DAG.getNode(ISD::SUB, DL, VT, N0, N11); | |||
2459 | return DAG.getNode(ISD::SUB, DL, VT, NewC, N1.getOperand(0)); | |||
2460 | } | |||
2461 | } | |||
2462 | ||||
2463 | // fold ((A+(B+or-C))-B) -> A+or-C | |||
2464 | if (N0.getOpcode() == ISD::ADD && | |||
2465 | (N0.getOperand(1).getOpcode() == ISD::SUB || | |||
2466 | N0.getOperand(1).getOpcode() == ISD::ADD) && | |||
2467 | N0.getOperand(1).getOperand(0) == N1) | |||
2468 | return DAG.getNode(N0.getOperand(1).getOpcode(), DL, VT, N0.getOperand(0), | |||
2469 | N0.getOperand(1).getOperand(1)); | |||
2470 | ||||
2471 | // fold ((A+(C+B))-B) -> A+C | |||
2472 | if (N0.getOpcode() == ISD::ADD && N0.getOperand(1).getOpcode() == ISD::ADD && | |||
2473 | N0.getOperand(1).getOperand(1) == N1) | |||
2474 | return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), | |||
2475 | N0.getOperand(1).getOperand(0)); | |||
2476 | ||||
2477 | // fold ((A-(B-C))-C) -> A-B | |||
2478 | if (N0.getOpcode() == ISD::SUB && N0.getOperand(1).getOpcode() == ISD::SUB && | |||
2479 | N0.getOperand(1).getOperand(1) == N1) | |||
2480 | return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), | |||
2481 | N0.getOperand(1).getOperand(0)); | |||
2482 | ||||
2483 | // If either operand of a sub is undef, the result is undef | |||
2484 | if (N0.isUndef()) | |||
2485 | return N0; | |||
2486 | if (N1.isUndef()) | |||
2487 | return N1; | |||
2488 | ||||
2489 | // If the relocation model supports it, consider symbol offsets. | |||
2490 | if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0)) | |||
2491 | if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) { | |||
2492 | // fold (sub Sym, c) -> Sym-c | |||
2493 | if (N1C && GA->getOpcode() == ISD::GlobalAddress) | |||
2494 | return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT, | |||
2495 | GA->getOffset() - | |||
2496 | (uint64_t)N1C->getSExtValue()); | |||
2497 | // fold (sub Sym+c1, Sym+c2) -> c1-c2 | |||
2498 | if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1)) | |||
2499 | if (GA->getGlobal() == GB->getGlobal()) | |||
2500 | return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(), | |||
2501 | DL, VT); | |||
2502 | } | |||
2503 | ||||
2504 | // sub X, (sextinreg Y i1) -> add X, (and Y 1) | |||
2505 | if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) { | |||
2506 | VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1)); | |||
2507 | if (TN->getVT() == MVT::i1) { | |||
2508 | SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0), | |||
2509 | DAG.getConstant(1, DL, VT)); | |||
2510 | return DAG.getNode(ISD::ADD, DL, VT, N0, ZExt); | |||
2511 | } | |||
2512 | } | |||
2513 | ||||
2514 | return SDValue(); | |||
2515 | } | |||
2516 | ||||
2517 | SDValue DAGCombiner::visitSUBC(SDNode *N) { | |||
2518 | SDValue N0 = N->getOperand(0); | |||
2519 | SDValue N1 = N->getOperand(1); | |||
2520 | EVT VT = N0.getValueType(); | |||
2521 | SDLoc DL(N); | |||
2522 | ||||
2523 | // If the flag result is dead, turn this into an SUB. | |||
2524 | if (!N->hasAnyUseOfValue(1)) | |||
2525 | return CombineTo(N, DAG.getNode(ISD::SUB, DL, VT, N0, N1), | |||
2526 | DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); | |||
2527 | ||||
2528 | // fold (subc x, x) -> 0 + no borrow | |||
2529 | if (N0 == N1) | |||
2530 | return CombineTo(N, DAG.getConstant(0, DL, VT), | |||
2531 | DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); | |||
2532 | ||||
2533 | // fold (subc x, 0) -> x + no borrow | |||
2534 | if (isNullConstant(N1)) | |||
2535 | return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); | |||
2536 | ||||
2537 | // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow | |||
2538 | if (isAllOnesConstant(N0)) | |||
2539 | return CombineTo(N, DAG.getNode(ISD::XOR, DL, VT, N1, N0), | |||
2540 | DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); | |||
2541 | ||||
2542 | return SDValue(); | |||
2543 | } | |||
2544 | ||||
2545 | SDValue DAGCombiner::visitUSUBO(SDNode *N) { | |||
2546 | SDValue N0 = N->getOperand(0); | |||
2547 | SDValue N1 = N->getOperand(1); | |||
2548 | EVT VT = N0.getValueType(); | |||
2549 | if (VT.isVector()) | |||
2550 | return SDValue(); | |||
2551 | ||||
2552 | EVT CarryVT = N->getValueType(1); | |||
2553 | SDLoc DL(N); | |||
2554 | ||||
2555 | // If the flag result is dead, turn this into an SUB. | |||
2556 | if (!N->hasAnyUseOfValue(1)) | |||
2557 | return CombineTo(N, DAG.getNode(ISD::SUB, DL, VT, N0, N1), | |||
2558 | DAG.getUNDEF(CarryVT)); | |||
2559 | ||||
2560 | // fold (usubo x, x) -> 0 + no borrow | |||
2561 | if (N0 == N1) | |||
2562 | return CombineTo(N, DAG.getConstant(0, DL, VT), | |||
2563 | DAG.getConstant(0, DL, CarryVT)); | |||
2564 | ||||
2565 | // fold (usubo x, 0) -> x + no borrow | |||
2566 | if (isNullConstant(N1)) | |||
2567 | return CombineTo(N, N0, DAG.getConstant(0, DL, CarryVT)); | |||
2568 | ||||
2569 | // Canonicalize (usubo -1, x) -> ~x, i.e. (xor x, -1) + no borrow | |||
2570 | if (isAllOnesConstant(N0)) | |||
2571 | return CombineTo(N, DAG.getNode(ISD::XOR, DL, VT, N1, N0), | |||
2572 | DAG.getConstant(0, DL, CarryVT)); | |||
2573 | ||||
2574 | return SDValue(); | |||
2575 | } | |||
2576 | ||||
2577 | SDValue DAGCombiner::visitSUBE(SDNode *N) { | |||
2578 | SDValue N0 = N->getOperand(0); | |||
2579 | SDValue N1 = N->getOperand(1); | |||
2580 | SDValue CarryIn = N->getOperand(2); | |||
2581 | ||||
2582 | // fold (sube x, y, false) -> (subc x, y) | |||
2583 | if (CarryIn.getOpcode() == ISD::CARRY_FALSE) | |||
2584 | return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1); | |||
2585 | ||||
2586 | return SDValue(); | |||
2587 | } | |||
2588 | ||||
2589 | SDValue DAGCombiner::visitSUBCARRY(SDNode *N) { | |||
2590 | SDValue N0 = N->getOperand(0); | |||
2591 | SDValue N1 = N->getOperand(1); | |||
2592 | SDValue CarryIn = N->getOperand(2); | |||
2593 | ||||
2594 | // fold (subcarry x, y, false) -> (usubo x, y) | |||
2595 | if (isNullConstant(CarryIn)) | |||
2596 | return DAG.getNode(ISD::USUBO, SDLoc(N), N->getVTList(), N0, N1); | |||
2597 | ||||
2598 | return SDValue(); | |||
2599 | } | |||
2600 | ||||
2601 | SDValue DAGCombiner::visitMUL(SDNode *N) { | |||
2602 | SDValue N0 = N->getOperand(0); | |||
2603 | SDValue N1 = N->getOperand(1); | |||
2604 | EVT VT = N0.getValueType(); | |||
2605 | ||||
2606 | // fold (mul x, undef) -> 0 | |||
2607 | if (N0.isUndef() || N1.isUndef()) | |||
2608 | return DAG.getConstant(0, SDLoc(N), VT); | |||
2609 | ||||
2610 | bool N0IsConst = false; | |||
2611 | bool N1IsConst = false; | |||
2612 | bool N1IsOpaqueConst = false; | |||
2613 | bool N0IsOpaqueConst = false; | |||
2614 | APInt ConstValue0, ConstValue1; | |||
2615 | // fold vector ops | |||
2616 | if (VT.isVector()) { | |||
2617 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
2618 | return FoldedVOp; | |||
2619 | ||||
2620 | N0IsConst = ISD::isConstantSplatVector(N0.getNode(), ConstValue0); | |||
2621 | N1IsConst = ISD::isConstantSplatVector(N1.getNode(), ConstValue1); | |||
2622 | assert((!N0IsConst ||(static_cast <bool> ((!N0IsConst || ConstValue0.getBitWidth () == VT.getScalarSizeInBits()) && "Splat APInt should be element width" ) ? void (0) : __assert_fail ("(!N0IsConst || ConstValue0.getBitWidth() == VT.getScalarSizeInBits()) && \"Splat APInt should be element width\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2624, __extension__ __PRETTY_FUNCTION__)) | |||
2623 | ConstValue0.getBitWidth() == VT.getScalarSizeInBits()) &&(static_cast <bool> ((!N0IsConst || ConstValue0.getBitWidth () == VT.getScalarSizeInBits()) && "Splat APInt should be element width" ) ? void (0) : __assert_fail ("(!N0IsConst || ConstValue0.getBitWidth() == VT.getScalarSizeInBits()) && \"Splat APInt should be element width\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2624, __extension__ __PRETTY_FUNCTION__)) | |||
2624 | "Splat APInt should be element width")(static_cast <bool> ((!N0IsConst || ConstValue0.getBitWidth () == VT.getScalarSizeInBits()) && "Splat APInt should be element width" ) ? void (0) : __assert_fail ("(!N0IsConst || ConstValue0.getBitWidth() == VT.getScalarSizeInBits()) && \"Splat APInt should be element width\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2624, __extension__ __PRETTY_FUNCTION__)); | |||
2625 | assert((!N1IsConst ||(static_cast <bool> ((!N1IsConst || ConstValue1.getBitWidth () == VT.getScalarSizeInBits()) && "Splat APInt should be element width" ) ? void (0) : __assert_fail ("(!N1IsConst || ConstValue1.getBitWidth() == VT.getScalarSizeInBits()) && \"Splat APInt should be element width\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2627, __extension__ __PRETTY_FUNCTION__)) | |||
2626 | ConstValue1.getBitWidth() == VT.getScalarSizeInBits()) &&(static_cast <bool> ((!N1IsConst || ConstValue1.getBitWidth () == VT.getScalarSizeInBits()) && "Splat APInt should be element width" ) ? void (0) : __assert_fail ("(!N1IsConst || ConstValue1.getBitWidth() == VT.getScalarSizeInBits()) && \"Splat APInt should be element width\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2627, __extension__ __PRETTY_FUNCTION__)) | |||
2627 | "Splat APInt should be element width")(static_cast <bool> ((!N1IsConst || ConstValue1.getBitWidth () == VT.getScalarSizeInBits()) && "Splat APInt should be element width" ) ? void (0) : __assert_fail ("(!N1IsConst || ConstValue1.getBitWidth() == VT.getScalarSizeInBits()) && \"Splat APInt should be element width\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2627, __extension__ __PRETTY_FUNCTION__)); | |||
2628 | } else { | |||
2629 | N0IsConst = isa<ConstantSDNode>(N0); | |||
2630 | if (N0IsConst) { | |||
2631 | ConstValue0 = cast<ConstantSDNode>(N0)->getAPIntValue(); | |||
2632 | N0IsOpaqueConst = cast<ConstantSDNode>(N0)->isOpaque(); | |||
2633 | } | |||
2634 | N1IsConst = isa<ConstantSDNode>(N1); | |||
2635 | if (N1IsConst) { | |||
2636 | ConstValue1 = cast<ConstantSDNode>(N1)->getAPIntValue(); | |||
2637 | N1IsOpaqueConst = cast<ConstantSDNode>(N1)->isOpaque(); | |||
2638 | } | |||
2639 | } | |||
2640 | ||||
2641 | // fold (mul c1, c2) -> c1*c2 | |||
2642 | if (N0IsConst && N1IsConst && !N0IsOpaqueConst && !N1IsOpaqueConst) | |||
2643 | return DAG.FoldConstantArithmetic(ISD::MUL, SDLoc(N), VT, | |||
2644 | N0.getNode(), N1.getNode()); | |||
2645 | ||||
2646 | // canonicalize constant to RHS (vector doesn't have to splat) | |||
2647 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
2648 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) | |||
2649 | return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0); | |||
2650 | // fold (mul x, 0) -> 0 | |||
2651 | if (N1IsConst && ConstValue1.isNullValue()) | |||
2652 | return N1; | |||
2653 | // fold (mul x, 1) -> x | |||
2654 | if (N1IsConst && ConstValue1.isOneValue()) | |||
2655 | return N0; | |||
2656 | ||||
2657 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
2658 | return NewSel; | |||
2659 | ||||
2660 | // fold (mul x, -1) -> 0-x | |||
2661 | if (N1IsConst && ConstValue1.isAllOnesValue()) { | |||
2662 | SDLoc DL(N); | |||
2663 | return DAG.getNode(ISD::SUB, DL, VT, | |||
2664 | DAG.getConstant(0, DL, VT), N0); | |||
2665 | } | |||
2666 | // fold (mul x, (1 << c)) -> x << c | |||
2667 | if (isConstantOrConstantVector(N1, /*NoOpaques*/ true) && | |||
2668 | DAG.isKnownToBeAPowerOfTwo(N1) && | |||
2669 | (!VT.isVector() || Level <= AfterLegalizeVectorOps)) { | |||
2670 | SDLoc DL(N); | |||
2671 | SDValue LogBase2 = BuildLogBase2(N1, DL); | |||
2672 | AddToWorklist(LogBase2.getNode()); | |||
2673 | ||||
2674 | EVT ShiftVT = getShiftAmountTy(N0.getValueType()); | |||
2675 | SDValue Trunc = DAG.getZExtOrTrunc(LogBase2, DL, ShiftVT); | |||
2676 | AddToWorklist(Trunc.getNode()); | |||
2677 | return DAG.getNode(ISD::SHL, DL, VT, N0, Trunc); | |||
2678 | } | |||
2679 | // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c | |||
2680 | if (N1IsConst && !N1IsOpaqueConst && (-ConstValue1).isPowerOf2()) { | |||
2681 | unsigned Log2Val = (-ConstValue1).logBase2(); | |||
2682 | SDLoc DL(N); | |||
2683 | // FIXME: If the input is something that is easily negated (e.g. a | |||
2684 | // single-use add), we should put the negate there. | |||
2685 | return DAG.getNode(ISD::SUB, DL, VT, | |||
2686 | DAG.getConstant(0, DL, VT), | |||
2687 | DAG.getNode(ISD::SHL, DL, VT, N0, | |||
2688 | DAG.getConstant(Log2Val, DL, | |||
2689 | getShiftAmountTy(N0.getValueType())))); | |||
2690 | } | |||
2691 | ||||
2692 | // (mul (shl X, c1), c2) -> (mul X, c2 << c1) | |||
2693 | if (N0.getOpcode() == ISD::SHL && | |||
2694 | isConstantOrConstantVector(N1, /* NoOpaques */ true) && | |||
2695 | isConstantOrConstantVector(N0.getOperand(1), /* NoOpaques */ true)) { | |||
2696 | SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT, N1, N0.getOperand(1)); | |||
2697 | if (isConstantOrConstantVector(C3)) | |||
2698 | return DAG.getNode(ISD::MUL, SDLoc(N), VT, N0.getOperand(0), C3); | |||
2699 | } | |||
2700 | ||||
2701 | // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one | |||
2702 | // use. | |||
2703 | { | |||
2704 | SDValue Sh(nullptr, 0), Y(nullptr, 0); | |||
2705 | ||||
2706 | // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)). | |||
2707 | if (N0.getOpcode() == ISD::SHL && | |||
2708 | isConstantOrConstantVector(N0.getOperand(1)) && | |||
2709 | N0.getNode()->hasOneUse()) { | |||
2710 | Sh = N0; Y = N1; | |||
2711 | } else if (N1.getOpcode() == ISD::SHL && | |||
2712 | isConstantOrConstantVector(N1.getOperand(1)) && | |||
2713 | N1.getNode()->hasOneUse()) { | |||
2714 | Sh = N1; Y = N0; | |||
2715 | } | |||
2716 | ||||
2717 | if (Sh.getNode()) { | |||
2718 | SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT, Sh.getOperand(0), Y); | |||
2719 | return DAG.getNode(ISD::SHL, SDLoc(N), VT, Mul, Sh.getOperand(1)); | |||
2720 | } | |||
2721 | } | |||
2722 | ||||
2723 | // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2) | |||
2724 | if (DAG.isConstantIntBuildVectorOrConstantInt(N1) && | |||
2725 | N0.getOpcode() == ISD::ADD && | |||
2726 | DAG.isConstantIntBuildVectorOrConstantInt(N0.getOperand(1)) && | |||
2727 | isMulAddWithConstProfitable(N, N0, N1)) | |||
2728 | return DAG.getNode(ISD::ADD, SDLoc(N), VT, | |||
2729 | DAG.getNode(ISD::MUL, SDLoc(N0), VT, | |||
2730 | N0.getOperand(0), N1), | |||
2731 | DAG.getNode(ISD::MUL, SDLoc(N1), VT, | |||
2732 | N0.getOperand(1), N1)); | |||
2733 | ||||
2734 | // reassociate mul | |||
2735 | if (SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1)) | |||
2736 | return RMUL; | |||
2737 | ||||
2738 | return SDValue(); | |||
2739 | } | |||
2740 | ||||
2741 | /// Return true if divmod libcall is available. | |||
2742 | static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned, | |||
2743 | const TargetLowering &TLI) { | |||
2744 | RTLIB::Libcall LC; | |||
2745 | EVT NodeType = Node->getValueType(0); | |||
2746 | if (!NodeType.isSimple()) | |||
2747 | return false; | |||
2748 | switch (NodeType.getSimpleVT().SimpleTy) { | |||
2749 | default: return false; // No libcall for vector types. | |||
2750 | case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break; | |||
2751 | case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break; | |||
2752 | case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break; | |||
2753 | case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break; | |||
2754 | case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break; | |||
2755 | } | |||
2756 | ||||
2757 | return TLI.getLibcallName(LC) != nullptr; | |||
2758 | } | |||
2759 | ||||
2760 | /// Issue divrem if both quotient and remainder are needed. | |||
2761 | SDValue DAGCombiner::useDivRem(SDNode *Node) { | |||
2762 | if (Node->use_empty()) | |||
2763 | return SDValue(); // This is a dead node, leave it alone. | |||
2764 | ||||
2765 | unsigned Opcode = Node->getOpcode(); | |||
2766 | bool isSigned = (Opcode == ISD::SDIV) || (Opcode == ISD::SREM); | |||
2767 | unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM; | |||
2768 | ||||
2769 | // DivMod lib calls can still work on non-legal types if using lib-calls. | |||
2770 | EVT VT = Node->getValueType(0); | |||
2771 | if (VT.isVector() || !VT.isInteger()) | |||
2772 | return SDValue(); | |||
2773 | ||||
2774 | if (!TLI.isTypeLegal(VT) && !TLI.isOperationCustom(DivRemOpc, VT)) | |||
2775 | return SDValue(); | |||
2776 | ||||
2777 | // If DIVREM is going to get expanded into a libcall, | |||
2778 | // but there is no libcall available, then don't combine. | |||
2779 | if (!TLI.isOperationLegalOrCustom(DivRemOpc, VT) && | |||
2780 | !isDivRemLibcallAvailable(Node, isSigned, TLI)) | |||
2781 | return SDValue(); | |||
2782 | ||||
2783 | // If div is legal, it's better to do the normal expansion | |||
2784 | unsigned OtherOpcode = 0; | |||
2785 | if ((Opcode == ISD::SDIV) || (Opcode == ISD::UDIV)) { | |||
2786 | OtherOpcode = isSigned ? ISD::SREM : ISD::UREM; | |||
2787 | if (TLI.isOperationLegalOrCustom(Opcode, VT)) | |||
2788 | return SDValue(); | |||
2789 | } else { | |||
2790 | OtherOpcode = isSigned ? ISD::SDIV : ISD::UDIV; | |||
2791 | if (TLI.isOperationLegalOrCustom(OtherOpcode, VT)) | |||
2792 | return SDValue(); | |||
2793 | } | |||
2794 | ||||
2795 | SDValue Op0 = Node->getOperand(0); | |||
2796 | SDValue Op1 = Node->getOperand(1); | |||
2797 | SDValue combined; | |||
2798 | for (SDNode::use_iterator UI = Op0.getNode()->use_begin(), | |||
2799 | UE = Op0.getNode()->use_end(); UI != UE; ++UI) { | |||
2800 | SDNode *User = *UI; | |||
2801 | if (User == Node || User->use_empty()) | |||
2802 | continue; | |||
2803 | // Convert the other matching node(s), too; | |||
2804 | // otherwise, the DIVREM may get target-legalized into something | |||
2805 | // target-specific that we won't be able to recognize. | |||
2806 | unsigned UserOpc = User->getOpcode(); | |||
2807 | if ((UserOpc == Opcode || UserOpc == OtherOpcode || UserOpc == DivRemOpc) && | |||
2808 | User->getOperand(0) == Op0 && | |||
2809 | User->getOperand(1) == Op1) { | |||
2810 | if (!combined) { | |||
2811 | if (UserOpc == OtherOpcode) { | |||
2812 | SDVTList VTs = DAG.getVTList(VT, VT); | |||
2813 | combined = DAG.getNode(DivRemOpc, SDLoc(Node), VTs, Op0, Op1); | |||
2814 | } else if (UserOpc == DivRemOpc) { | |||
2815 | combined = SDValue(User, 0); | |||
2816 | } else { | |||
2817 | assert(UserOpc == Opcode)(static_cast <bool> (UserOpc == Opcode) ? void (0) : __assert_fail ("UserOpc == Opcode", "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 2817, __extension__ __PRETTY_FUNCTION__)); | |||
2818 | continue; | |||
2819 | } | |||
2820 | } | |||
2821 | if (UserOpc == ISD::SDIV || UserOpc == ISD::UDIV) | |||
2822 | CombineTo(User, combined); | |||
2823 | else if (UserOpc == ISD::SREM || UserOpc == ISD::UREM) | |||
2824 | CombineTo(User, combined.getValue(1)); | |||
2825 | } | |||
2826 | } | |||
2827 | return combined; | |||
2828 | } | |||
2829 | ||||
2830 | static SDValue simplifyDivRem(SDNode *N, SelectionDAG &DAG) { | |||
2831 | SDValue N0 = N->getOperand(0); | |||
2832 | SDValue N1 = N->getOperand(1); | |||
2833 | EVT VT = N->getValueType(0); | |||
2834 | SDLoc DL(N); | |||
2835 | ||||
2836 | if (DAG.isUndef(N->getOpcode(), {N0, N1})) | |||
2837 | return DAG.getUNDEF(VT); | |||
2838 | ||||
2839 | // undef / X -> 0 | |||
2840 | // undef % X -> 0 | |||
2841 | if (N0.isUndef()) | |||
2842 | return DAG.getConstant(0, DL, VT); | |||
2843 | ||||
2844 | return SDValue(); | |||
2845 | } | |||
2846 | ||||
2847 | SDValue DAGCombiner::visitSDIV(SDNode *N) { | |||
2848 | SDValue N0 = N->getOperand(0); | |||
2849 | SDValue N1 = N->getOperand(1); | |||
2850 | EVT VT = N->getValueType(0); | |||
2851 | ||||
2852 | // fold vector ops | |||
2853 | if (VT.isVector()) | |||
2854 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
2855 | return FoldedVOp; | |||
2856 | ||||
2857 | SDLoc DL(N); | |||
2858 | ||||
2859 | // fold (sdiv c1, c2) -> c1/c2 | |||
2860 | ConstantSDNode *N0C = isConstOrConstSplat(N0); | |||
2861 | ConstantSDNode *N1C = isConstOrConstSplat(N1); | |||
2862 | if (N0C && N1C && !N0C->isOpaque() && !N1C->isOpaque()) | |||
2863 | return DAG.FoldConstantArithmetic(ISD::SDIV, DL, VT, N0C, N1C); | |||
2864 | // fold (sdiv X, 1) -> X | |||
2865 | if (N1C && N1C->isOne()) | |||
2866 | return N0; | |||
2867 | // fold (sdiv X, -1) -> 0-X | |||
2868 | if (N1C && N1C->isAllOnesValue()) | |||
2869 | return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), N0); | |||
2870 | ||||
2871 | if (SDValue V = simplifyDivRem(N, DAG)) | |||
2872 | return V; | |||
2873 | ||||
2874 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
2875 | return NewSel; | |||
2876 | ||||
2877 | // If we know the sign bits of both operands are zero, strength reduce to a | |||
2878 | // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2 | |||
2879 | if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0)) | |||
2880 | return DAG.getNode(ISD::UDIV, DL, N1.getValueType(), N0, N1); | |||
2881 | ||||
2882 | // fold (sdiv X, pow2) -> simple ops after legalize | |||
2883 | // FIXME: We check for the exact bit here because the generic lowering gives | |||
2884 | // better results in that case. The target-specific lowering should learn how | |||
2885 | // to handle exact sdivs efficiently. | |||
2886 | if (N1C && !N1C->isNullValue() && !N1C->isOpaque() && | |||
2887 | !N->getFlags().hasExact() && (N1C->getAPIntValue().isPowerOf2() || | |||
2888 | (-N1C->getAPIntValue()).isPowerOf2())) { | |||
2889 | // Target-specific implementation of sdiv x, pow2. | |||
2890 | if (SDValue Res = BuildSDIVPow2(N)) | |||
2891 | return Res; | |||
2892 | ||||
2893 | unsigned lg2 = N1C->getAPIntValue().countTrailingZeros(); | |||
2894 | ||||
2895 | // Splat the sign bit into the register | |||
2896 | SDValue SGN = | |||
2897 | DAG.getNode(ISD::SRA, DL, VT, N0, | |||
2898 | DAG.getConstant(VT.getScalarSizeInBits() - 1, DL, | |||
2899 | getShiftAmountTy(N0.getValueType()))); | |||
2900 | AddToWorklist(SGN.getNode()); | |||
2901 | ||||
2902 | // Add (N0 < 0) ? abs2 - 1 : 0; | |||
2903 | SDValue SRL = | |||
2904 | DAG.getNode(ISD::SRL, DL, VT, SGN, | |||
2905 | DAG.getConstant(VT.getScalarSizeInBits() - lg2, DL, | |||
2906 | getShiftAmountTy(SGN.getValueType()))); | |||
2907 | SDValue ADD = DAG.getNode(ISD::ADD, DL, VT, N0, SRL); | |||
2908 | AddToWorklist(SRL.getNode()); | |||
2909 | AddToWorklist(ADD.getNode()); // Divide by pow2 | |||
2910 | SDValue SRA = DAG.getNode(ISD::SRA, DL, VT, ADD, | |||
2911 | DAG.getConstant(lg2, DL, | |||
2912 | getShiftAmountTy(ADD.getValueType()))); | |||
2913 | ||||
2914 | // If we're dividing by a positive value, we're done. Otherwise, we must | |||
2915 | // negate the result. | |||
2916 | if (N1C->getAPIntValue().isNonNegative()) | |||
2917 | return SRA; | |||
2918 | ||||
2919 | AddToWorklist(SRA.getNode()); | |||
2920 | return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), SRA); | |||
2921 | } | |||
2922 | ||||
2923 | // If integer divide is expensive and we satisfy the requirements, emit an | |||
2924 | // alternate sequence. Targets may check function attributes for size/speed | |||
2925 | // trade-offs. | |||
2926 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); | |||
2927 | if (N1C && !TLI.isIntDivCheap(N->getValueType(0), Attr)) | |||
2928 | if (SDValue Op = BuildSDIV(N)) | |||
2929 | return Op; | |||
2930 | ||||
2931 | // sdiv, srem -> sdivrem | |||
2932 | // If the divisor is constant, then return DIVREM only if isIntDivCheap() is | |||
2933 | // true. Otherwise, we break the simplification logic in visitREM(). | |||
2934 | if (!N1C || TLI.isIntDivCheap(N->getValueType(0), Attr)) | |||
2935 | if (SDValue DivRem = useDivRem(N)) | |||
2936 | return DivRem; | |||
2937 | ||||
2938 | return SDValue(); | |||
2939 | } | |||
2940 | ||||
2941 | SDValue DAGCombiner::visitUDIV(SDNode *N) { | |||
2942 | SDValue N0 = N->getOperand(0); | |||
2943 | SDValue N1 = N->getOperand(1); | |||
2944 | EVT VT = N->getValueType(0); | |||
2945 | ||||
2946 | // fold vector ops | |||
2947 | if (VT.isVector()) | |||
2948 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
2949 | return FoldedVOp; | |||
2950 | ||||
2951 | SDLoc DL(N); | |||
2952 | ||||
2953 | // fold (udiv c1, c2) -> c1/c2 | |||
2954 | ConstantSDNode *N0C = isConstOrConstSplat(N0); | |||
2955 | ConstantSDNode *N1C = isConstOrConstSplat(N1); | |||
2956 | if (N0C && N1C) | |||
2957 | if (SDValue Folded = DAG.FoldConstantArithmetic(ISD::UDIV, DL, VT, | |||
2958 | N0C, N1C)) | |||
2959 | return Folded; | |||
2960 | ||||
2961 | if (SDValue V = simplifyDivRem(N, DAG)) | |||
2962 | return V; | |||
2963 | ||||
2964 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
2965 | return NewSel; | |||
2966 | ||||
2967 | // fold (udiv x, (1 << c)) -> x >>u c | |||
2968 | if (isConstantOrConstantVector(N1, /*NoOpaques*/ true) && | |||
2969 | DAG.isKnownToBeAPowerOfTwo(N1)) { | |||
2970 | SDValue LogBase2 = BuildLogBase2(N1, DL); | |||
2971 | AddToWorklist(LogBase2.getNode()); | |||
2972 | ||||
2973 | EVT ShiftVT = getShiftAmountTy(N0.getValueType()); | |||
2974 | SDValue Trunc = DAG.getZExtOrTrunc(LogBase2, DL, ShiftVT); | |||
2975 | AddToWorklist(Trunc.getNode()); | |||
2976 | return DAG.getNode(ISD::SRL, DL, VT, N0, Trunc); | |||
2977 | } | |||
2978 | ||||
2979 | // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2 | |||
2980 | if (N1.getOpcode() == ISD::SHL) { | |||
2981 | SDValue N10 = N1.getOperand(0); | |||
2982 | if (isConstantOrConstantVector(N10, /*NoOpaques*/ true) && | |||
2983 | DAG.isKnownToBeAPowerOfTwo(N10)) { | |||
2984 | SDValue LogBase2 = BuildLogBase2(N10, DL); | |||
2985 | AddToWorklist(LogBase2.getNode()); | |||
2986 | ||||
2987 | EVT ADDVT = N1.getOperand(1).getValueType(); | |||
2988 | SDValue Trunc = DAG.getZExtOrTrunc(LogBase2, DL, ADDVT); | |||
2989 | AddToWorklist(Trunc.getNode()); | |||
2990 | SDValue Add = DAG.getNode(ISD::ADD, DL, ADDVT, N1.getOperand(1), Trunc); | |||
2991 | AddToWorklist(Add.getNode()); | |||
2992 | return DAG.getNode(ISD::SRL, DL, VT, N0, Add); | |||
2993 | } | |||
2994 | } | |||
2995 | ||||
2996 | // fold (udiv x, c) -> alternate | |||
2997 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); | |||
2998 | if (N1C && !TLI.isIntDivCheap(N->getValueType(0), Attr)) | |||
2999 | if (SDValue Op = BuildUDIV(N)) | |||
3000 | return Op; | |||
3001 | ||||
3002 | // sdiv, srem -> sdivrem | |||
3003 | // If the divisor is constant, then return DIVREM only if isIntDivCheap() is | |||
3004 | // true. Otherwise, we break the simplification logic in visitREM(). | |||
3005 | if (!N1C || TLI.isIntDivCheap(N->getValueType(0), Attr)) | |||
3006 | if (SDValue DivRem = useDivRem(N)) | |||
3007 | return DivRem; | |||
3008 | ||||
3009 | return SDValue(); | |||
3010 | } | |||
3011 | ||||
3012 | // handles ISD::SREM and ISD::UREM | |||
3013 | SDValue DAGCombiner::visitREM(SDNode *N) { | |||
3014 | unsigned Opcode = N->getOpcode(); | |||
3015 | SDValue N0 = N->getOperand(0); | |||
3016 | SDValue N1 = N->getOperand(1); | |||
3017 | EVT VT = N->getValueType(0); | |||
3018 | bool isSigned = (Opcode == ISD::SREM); | |||
3019 | SDLoc DL(N); | |||
3020 | ||||
3021 | // fold (rem c1, c2) -> c1%c2 | |||
3022 | ConstantSDNode *N0C = isConstOrConstSplat(N0); | |||
3023 | ConstantSDNode *N1C = isConstOrConstSplat(N1); | |||
3024 | if (N0C && N1C) | |||
3025 | if (SDValue Folded = DAG.FoldConstantArithmetic(Opcode, DL, VT, N0C, N1C)) | |||
3026 | return Folded; | |||
3027 | ||||
3028 | if (SDValue V = simplifyDivRem(N, DAG)) | |||
3029 | return V; | |||
3030 | ||||
3031 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
3032 | return NewSel; | |||
3033 | ||||
3034 | if (isSigned) { | |||
3035 | // If we know the sign bits of both operands are zero, strength reduce to a | |||
3036 | // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15 | |||
3037 | if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0)) | |||
3038 | return DAG.getNode(ISD::UREM, DL, VT, N0, N1); | |||
3039 | } else { | |||
3040 | SDValue NegOne = DAG.getAllOnesConstant(DL, VT); | |||
3041 | if (DAG.isKnownToBeAPowerOfTwo(N1)) { | |||
3042 | // fold (urem x, pow2) -> (and x, pow2-1) | |||
3043 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N1, NegOne); | |||
3044 | AddToWorklist(Add.getNode()); | |||
3045 | return DAG.getNode(ISD::AND, DL, VT, N0, Add); | |||
3046 | } | |||
3047 | if (N1.getOpcode() == ISD::SHL && | |||
3048 | DAG.isKnownToBeAPowerOfTwo(N1.getOperand(0))) { | |||
3049 | // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1)) | |||
3050 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N1, NegOne); | |||
3051 | AddToWorklist(Add.getNode()); | |||
3052 | return DAG.getNode(ISD::AND, DL, VT, N0, Add); | |||
3053 | } | |||
3054 | } | |||
3055 | ||||
3056 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); | |||
3057 | ||||
3058 | // If X/C can be simplified by the division-by-constant logic, lower | |||
3059 | // X%C to the equivalent of X-X/C*C. | |||
3060 | // To avoid mangling nodes, this simplification requires that the combine() | |||
3061 | // call for the speculative DIV must not cause a DIVREM conversion. We guard | |||
3062 | // against this by skipping the simplification if isIntDivCheap(). When | |||
3063 | // div is not cheap, combine will not return a DIVREM. Regardless, | |||
3064 | // checking cheapness here makes sense since the simplification results in | |||
3065 | // fatter code. | |||
3066 | if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap(VT, Attr)) { | |||
3067 | unsigned DivOpcode = isSigned ? ISD::SDIV : ISD::UDIV; | |||
3068 | SDValue Div = DAG.getNode(DivOpcode, DL, VT, N0, N1); | |||
3069 | AddToWorklist(Div.getNode()); | |||
3070 | SDValue OptimizedDiv = combine(Div.getNode()); | |||
3071 | if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) { | |||
3072 | assert((OptimizedDiv.getOpcode() != ISD::UDIVREM) &&(static_cast <bool> ((OptimizedDiv.getOpcode() != ISD:: UDIVREM) && (OptimizedDiv.getOpcode() != ISD::SDIVREM )) ? void (0) : __assert_fail ("(OptimizedDiv.getOpcode() != ISD::UDIVREM) && (OptimizedDiv.getOpcode() != ISD::SDIVREM)" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3073, __extension__ __PRETTY_FUNCTION__)) | |||
3073 | (OptimizedDiv.getOpcode() != ISD::SDIVREM))(static_cast <bool> ((OptimizedDiv.getOpcode() != ISD:: UDIVREM) && (OptimizedDiv.getOpcode() != ISD::SDIVREM )) ? void (0) : __assert_fail ("(OptimizedDiv.getOpcode() != ISD::UDIVREM) && (OptimizedDiv.getOpcode() != ISD::SDIVREM)" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3073, __extension__ __PRETTY_FUNCTION__)); | |||
3074 | SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, OptimizedDiv, N1); | |||
3075 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0, Mul); | |||
3076 | AddToWorklist(Mul.getNode()); | |||
3077 | return Sub; | |||
3078 | } | |||
3079 | } | |||
3080 | ||||
3081 | // sdiv, srem -> sdivrem | |||
3082 | if (SDValue DivRem = useDivRem(N)) | |||
3083 | return DivRem.getValue(1); | |||
3084 | ||||
3085 | return SDValue(); | |||
3086 | } | |||
3087 | ||||
3088 | SDValue DAGCombiner::visitMULHS(SDNode *N) { | |||
3089 | SDValue N0 = N->getOperand(0); | |||
3090 | SDValue N1 = N->getOperand(1); | |||
3091 | EVT VT = N->getValueType(0); | |||
3092 | SDLoc DL(N); | |||
3093 | ||||
3094 | if (VT.isVector()) { | |||
3095 | // fold (mulhs x, 0) -> 0 | |||
3096 | if (ISD::isBuildVectorAllZeros(N1.getNode())) | |||
3097 | return N1; | |||
3098 | if (ISD::isBuildVectorAllZeros(N0.getNode())) | |||
3099 | return N0; | |||
3100 | } | |||
3101 | ||||
3102 | // fold (mulhs x, 0) -> 0 | |||
3103 | if (isNullConstant(N1)) | |||
3104 | return N1; | |||
3105 | // fold (mulhs x, 1) -> (sra x, size(x)-1) | |||
3106 | if (isOneConstant(N1)) | |||
3107 | return DAG.getNode(ISD::SRA, DL, N0.getValueType(), N0, | |||
3108 | DAG.getConstant(N0.getValueSizeInBits() - 1, DL, | |||
3109 | getShiftAmountTy(N0.getValueType()))); | |||
3110 | ||||
3111 | // fold (mulhs x, undef) -> 0 | |||
3112 | if (N0.isUndef() || N1.isUndef()) | |||
3113 | return DAG.getConstant(0, DL, VT); | |||
3114 | ||||
3115 | // If the type twice as wide is legal, transform the mulhs to a wider multiply | |||
3116 | // plus a shift. | |||
3117 | if (VT.isSimple() && !VT.isVector()) { | |||
3118 | MVT Simple = VT.getSimpleVT(); | |||
3119 | unsigned SimpleSize = Simple.getSizeInBits(); | |||
3120 | EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); | |||
3121 | if (TLI.isOperationLegal(ISD::MUL, NewVT)) { | |||
3122 | N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0); | |||
3123 | N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1); | |||
3124 | N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1); | |||
3125 | N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1, | |||
3126 | DAG.getConstant(SimpleSize, DL, | |||
3127 | getShiftAmountTy(N1.getValueType()))); | |||
3128 | return DAG.getNode(ISD::TRUNCATE, DL, VT, N1); | |||
3129 | } | |||
3130 | } | |||
3131 | ||||
3132 | return SDValue(); | |||
3133 | } | |||
3134 | ||||
3135 | SDValue DAGCombiner::visitMULHU(SDNode *N) { | |||
3136 | SDValue N0 = N->getOperand(0); | |||
3137 | SDValue N1 = N->getOperand(1); | |||
3138 | EVT VT = N->getValueType(0); | |||
3139 | SDLoc DL(N); | |||
3140 | ||||
3141 | if (VT.isVector()) { | |||
3142 | // fold (mulhu x, 0) -> 0 | |||
3143 | if (ISD::isBuildVectorAllZeros(N1.getNode())) | |||
3144 | return N1; | |||
3145 | if (ISD::isBuildVectorAllZeros(N0.getNode())) | |||
3146 | return N0; | |||
3147 | } | |||
3148 | ||||
3149 | // fold (mulhu x, 0) -> 0 | |||
3150 | if (isNullConstant(N1)) | |||
3151 | return N1; | |||
3152 | // fold (mulhu x, 1) -> 0 | |||
3153 | if (isOneConstant(N1)) | |||
3154 | return DAG.getConstant(0, DL, N0.getValueType()); | |||
3155 | // fold (mulhu x, undef) -> 0 | |||
3156 | if (N0.isUndef() || N1.isUndef()) | |||
3157 | return DAG.getConstant(0, DL, VT); | |||
3158 | ||||
3159 | // If the type twice as wide is legal, transform the mulhu to a wider multiply | |||
3160 | // plus a shift. | |||
3161 | if (VT.isSimple() && !VT.isVector()) { | |||
3162 | MVT Simple = VT.getSimpleVT(); | |||
3163 | unsigned SimpleSize = Simple.getSizeInBits(); | |||
3164 | EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); | |||
3165 | if (TLI.isOperationLegal(ISD::MUL, NewVT)) { | |||
3166 | N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0); | |||
3167 | N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1); | |||
3168 | N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1); | |||
3169 | N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1, | |||
3170 | DAG.getConstant(SimpleSize, DL, | |||
3171 | getShiftAmountTy(N1.getValueType()))); | |||
3172 | return DAG.getNode(ISD::TRUNCATE, DL, VT, N1); | |||
3173 | } | |||
3174 | } | |||
3175 | ||||
3176 | return SDValue(); | |||
3177 | } | |||
3178 | ||||
3179 | /// Perform optimizations common to nodes that compute two values. LoOp and HiOp | |||
3180 | /// give the opcodes for the two computations that are being performed. Return | |||
3181 | /// true if a simplification was made. | |||
3182 | SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp, | |||
3183 | unsigned HiOp) { | |||
3184 | // If the high half is not needed, just compute the low half. | |||
3185 | bool HiExists = N->hasAnyUseOfValue(1); | |||
3186 | if (!HiExists && | |||
3187 | (!LegalOperations || | |||
3188 | TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) { | |||
3189 | SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops()); | |||
3190 | return CombineTo(N, Res, Res); | |||
3191 | } | |||
3192 | ||||
3193 | // If the low half is not needed, just compute the high half. | |||
3194 | bool LoExists = N->hasAnyUseOfValue(0); | |||
3195 | if (!LoExists && | |||
3196 | (!LegalOperations || | |||
3197 | TLI.isOperationLegal(HiOp, N->getValueType(1)))) { | |||
3198 | SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops()); | |||
3199 | return CombineTo(N, Res, Res); | |||
3200 | } | |||
3201 | ||||
3202 | // If both halves are used, return as it is. | |||
3203 | if (LoExists && HiExists) | |||
3204 | return SDValue(); | |||
3205 | ||||
3206 | // If the two computed results can be simplified separately, separate them. | |||
3207 | if (LoExists) { | |||
3208 | SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops()); | |||
3209 | AddToWorklist(Lo.getNode()); | |||
3210 | SDValue LoOpt = combine(Lo.getNode()); | |||
3211 | if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() && | |||
3212 | (!LegalOperations || | |||
3213 | TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType()))) | |||
3214 | return CombineTo(N, LoOpt, LoOpt); | |||
3215 | } | |||
3216 | ||||
3217 | if (HiExists) { | |||
3218 | SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops()); | |||
3219 | AddToWorklist(Hi.getNode()); | |||
3220 | SDValue HiOpt = combine(Hi.getNode()); | |||
3221 | if (HiOpt.getNode() && HiOpt != Hi && | |||
3222 | (!LegalOperations || | |||
3223 | TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType()))) | |||
3224 | return CombineTo(N, HiOpt, HiOpt); | |||
3225 | } | |||
3226 | ||||
3227 | return SDValue(); | |||
3228 | } | |||
3229 | ||||
3230 | SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) { | |||
3231 | if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS)) | |||
3232 | return Res; | |||
3233 | ||||
3234 | EVT VT = N->getValueType(0); | |||
3235 | SDLoc DL(N); | |||
3236 | ||||
3237 | // If the type is twice as wide is legal, transform the mulhu to a wider | |||
3238 | // multiply plus a shift. | |||
3239 | if (VT.isSimple() && !VT.isVector()) { | |||
3240 | MVT Simple = VT.getSimpleVT(); | |||
3241 | unsigned SimpleSize = Simple.getSizeInBits(); | |||
3242 | EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); | |||
3243 | if (TLI.isOperationLegal(ISD::MUL, NewVT)) { | |||
3244 | SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0)); | |||
3245 | SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1)); | |||
3246 | Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi); | |||
3247 | // Compute the high part as N1. | |||
3248 | Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo, | |||
3249 | DAG.getConstant(SimpleSize, DL, | |||
3250 | getShiftAmountTy(Lo.getValueType()))); | |||
3251 | Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi); | |||
3252 | // Compute the low part as N0. | |||
3253 | Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo); | |||
3254 | return CombineTo(N, Lo, Hi); | |||
3255 | } | |||
3256 | } | |||
3257 | ||||
3258 | return SDValue(); | |||
3259 | } | |||
3260 | ||||
3261 | SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) { | |||
3262 | if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU)) | |||
3263 | return Res; | |||
3264 | ||||
3265 | EVT VT = N->getValueType(0); | |||
3266 | SDLoc DL(N); | |||
3267 | ||||
3268 | // If the type is twice as wide is legal, transform the mulhu to a wider | |||
3269 | // multiply plus a shift. | |||
3270 | if (VT.isSimple() && !VT.isVector()) { | |||
3271 | MVT Simple = VT.getSimpleVT(); | |||
3272 | unsigned SimpleSize = Simple.getSizeInBits(); | |||
3273 | EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); | |||
3274 | if (TLI.isOperationLegal(ISD::MUL, NewVT)) { | |||
3275 | SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0)); | |||
3276 | SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1)); | |||
3277 | Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi); | |||
3278 | // Compute the high part as N1. | |||
3279 | Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo, | |||
3280 | DAG.getConstant(SimpleSize, DL, | |||
3281 | getShiftAmountTy(Lo.getValueType()))); | |||
3282 | Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi); | |||
3283 | // Compute the low part as N0. | |||
3284 | Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo); | |||
3285 | return CombineTo(N, Lo, Hi); | |||
3286 | } | |||
3287 | } | |||
3288 | ||||
3289 | return SDValue(); | |||
3290 | } | |||
3291 | ||||
3292 | SDValue DAGCombiner::visitSMULO(SDNode *N) { | |||
3293 | // (smulo x, 2) -> (saddo x, x) | |||
3294 | if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1))) | |||
3295 | if (C2->getAPIntValue() == 2) | |||
3296 | return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(), | |||
3297 | N->getOperand(0), N->getOperand(0)); | |||
3298 | ||||
3299 | return SDValue(); | |||
3300 | } | |||
3301 | ||||
3302 | SDValue DAGCombiner::visitUMULO(SDNode *N) { | |||
3303 | // (umulo x, 2) -> (uaddo x, x) | |||
3304 | if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1))) | |||
3305 | if (C2->getAPIntValue() == 2) | |||
3306 | return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(), | |||
3307 | N->getOperand(0), N->getOperand(0)); | |||
3308 | ||||
3309 | return SDValue(); | |||
3310 | } | |||
3311 | ||||
3312 | SDValue DAGCombiner::visitIMINMAX(SDNode *N) { | |||
3313 | SDValue N0 = N->getOperand(0); | |||
3314 | SDValue N1 = N->getOperand(1); | |||
3315 | EVT VT = N0.getValueType(); | |||
3316 | ||||
3317 | // fold vector ops | |||
3318 | if (VT.isVector()) | |||
3319 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
3320 | return FoldedVOp; | |||
3321 | ||||
3322 | // fold operation with constant operands. | |||
3323 | ConstantSDNode *N0C = getAsNonOpaqueConstant(N0); | |||
3324 | ConstantSDNode *N1C = getAsNonOpaqueConstant(N1); | |||
3325 | if (N0C && N1C) | |||
3326 | return DAG.FoldConstantArithmetic(N->getOpcode(), SDLoc(N), VT, N0C, N1C); | |||
3327 | ||||
3328 | // canonicalize constant to RHS | |||
3329 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
3330 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) | |||
3331 | return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N1, N0); | |||
3332 | ||||
3333 | // Is sign bits are zero, flip between UMIN/UMAX and SMIN/SMAX. | |||
3334 | // Only do this if the current op isn't legal and the flipped is. | |||
3335 | unsigned Opcode = N->getOpcode(); | |||
3336 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3337 | if (!TLI.isOperationLegal(Opcode, VT) && | |||
3338 | (N0.isUndef() || DAG.SignBitIsZero(N0)) && | |||
3339 | (N1.isUndef() || DAG.SignBitIsZero(N1))) { | |||
3340 | unsigned AltOpcode; | |||
3341 | switch (Opcode) { | |||
3342 | case ISD::SMIN: AltOpcode = ISD::UMIN; break; | |||
3343 | case ISD::SMAX: AltOpcode = ISD::UMAX; break; | |||
3344 | case ISD::UMIN: AltOpcode = ISD::SMIN; break; | |||
3345 | case ISD::UMAX: AltOpcode = ISD::SMAX; break; | |||
3346 | default: llvm_unreachable("Unknown MINMAX opcode")::llvm::llvm_unreachable_internal("Unknown MINMAX opcode", "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3346); | |||
3347 | } | |||
3348 | if (TLI.isOperationLegal(AltOpcode, VT)) | |||
3349 | return DAG.getNode(AltOpcode, SDLoc(N), VT, N0, N1); | |||
3350 | } | |||
3351 | ||||
3352 | return SDValue(); | |||
3353 | } | |||
3354 | ||||
3355 | /// If this is a binary operator with two operands of the same opcode, try to | |||
3356 | /// simplify it. | |||
3357 | SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) { | |||
3358 | SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); | |||
3359 | EVT VT = N0.getValueType(); | |||
3360 | assert(N0.getOpcode() == N1.getOpcode() && "Bad input!")(static_cast <bool> (N0.getOpcode() == N1.getOpcode() && "Bad input!") ? void (0) : __assert_fail ("N0.getOpcode() == N1.getOpcode() && \"Bad input!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3360, __extension__ __PRETTY_FUNCTION__)); | |||
3361 | ||||
3362 | // Bail early if none of these transforms apply. | |||
3363 | if (N0.getNumOperands() == 0) return SDValue(); | |||
3364 | ||||
3365 | // For each of OP in AND/OR/XOR: | |||
3366 | // fold (OP (zext x), (zext y)) -> (zext (OP x, y)) | |||
3367 | // fold (OP (sext x), (sext y)) -> (sext (OP x, y)) | |||
3368 | // fold (OP (aext x), (aext y)) -> (aext (OP x, y)) | |||
3369 | // fold (OP (bswap x), (bswap y)) -> (bswap (OP x, y)) | |||
3370 | // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free) | |||
3371 | // | |||
3372 | // do not sink logical op inside of a vector extend, since it may combine | |||
3373 | // into a vsetcc. | |||
3374 | EVT Op0VT = N0.getOperand(0).getValueType(); | |||
3375 | if ((N0.getOpcode() == ISD::ZERO_EXTEND || | |||
3376 | N0.getOpcode() == ISD::SIGN_EXTEND || | |||
3377 | N0.getOpcode() == ISD::BSWAP || | |||
3378 | // Avoid infinite looping with PromoteIntBinOp. | |||
3379 | (N0.getOpcode() == ISD::ANY_EXTEND && | |||
3380 | (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) || | |||
3381 | (N0.getOpcode() == ISD::TRUNCATE && | |||
3382 | (!TLI.isZExtFree(VT, Op0VT) || | |||
3383 | !TLI.isTruncateFree(Op0VT, VT)) && | |||
3384 | TLI.isTypeLegal(Op0VT))) && | |||
3385 | !VT.isVector() && | |||
3386 | Op0VT == N1.getOperand(0).getValueType() && | |||
3387 | (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) { | |||
3388 | SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0), | |||
3389 | N0.getOperand(0).getValueType(), | |||
3390 | N0.getOperand(0), N1.getOperand(0)); | |||
3391 | AddToWorklist(ORNode.getNode()); | |||
3392 | return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode); | |||
3393 | } | |||
3394 | ||||
3395 | // For each of OP in SHL/SRL/SRA/AND... | |||
3396 | // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z) | |||
3397 | // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z) | |||
3398 | // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z) | |||
3399 | if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL || | |||
3400 | N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) && | |||
3401 | N0.getOperand(1) == N1.getOperand(1)) { | |||
3402 | SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0), | |||
3403 | N0.getOperand(0).getValueType(), | |||
3404 | N0.getOperand(0), N1.getOperand(0)); | |||
3405 | AddToWorklist(ORNode.getNode()); | |||
3406 | return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, | |||
3407 | ORNode, N0.getOperand(1)); | |||
3408 | } | |||
3409 | ||||
3410 | // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B)) | |||
3411 | // Only perform this optimization up until type legalization, before | |||
3412 | // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by | |||
3413 | // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and | |||
3414 | // we don't want to undo this promotion. | |||
3415 | // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper | |||
3416 | // on scalars. | |||
3417 | if ((N0.getOpcode() == ISD::BITCAST || | |||
3418 | N0.getOpcode() == ISD::SCALAR_TO_VECTOR) && | |||
3419 | Level <= AfterLegalizeTypes) { | |||
3420 | SDValue In0 = N0.getOperand(0); | |||
3421 | SDValue In1 = N1.getOperand(0); | |||
3422 | EVT In0Ty = In0.getValueType(); | |||
3423 | EVT In1Ty = In1.getValueType(); | |||
3424 | SDLoc DL(N); | |||
3425 | // If both incoming values are integers, and the original types are the | |||
3426 | // same. | |||
3427 | if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) { | |||
3428 | SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1); | |||
3429 | SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op); | |||
3430 | AddToWorklist(Op.getNode()); | |||
3431 | return BC; | |||
3432 | } | |||
3433 | } | |||
3434 | ||||
3435 | // Xor/and/or are indifferent to the swizzle operation (shuffle of one value). | |||
3436 | // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B)) | |||
3437 | // If both shuffles use the same mask, and both shuffle within a single | |||
3438 | // vector, then it is worthwhile to move the swizzle after the operation. | |||
3439 | // The type-legalizer generates this pattern when loading illegal | |||
3440 | // vector types from memory. In many cases this allows additional shuffle | |||
3441 | // optimizations. | |||
3442 | // There are other cases where moving the shuffle after the xor/and/or | |||
3443 | // is profitable even if shuffles don't perform a swizzle. | |||
3444 | // If both shuffles use the same mask, and both shuffles have the same first | |||
3445 | // or second operand, then it might still be profitable to move the shuffle | |||
3446 | // after the xor/and/or operation. | |||
3447 | if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) { | |||
3448 | ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0); | |||
3449 | ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1); | |||
3450 | ||||
3451 | assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&(static_cast <bool> (N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() && "Inputs to shuffles are not the same type" ) ? void (0) : __assert_fail ("N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() && \"Inputs to shuffles are not the same type\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3452, __extension__ __PRETTY_FUNCTION__)) | |||
3452 | "Inputs to shuffles are not the same type")(static_cast <bool> (N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() && "Inputs to shuffles are not the same type" ) ? void (0) : __assert_fail ("N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() && \"Inputs to shuffles are not the same type\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3452, __extension__ __PRETTY_FUNCTION__)); | |||
3453 | ||||
3454 | // Check that both shuffles use the same mask. The masks are known to be of | |||
3455 | // the same length because the result vector type is the same. | |||
3456 | // Check also that shuffles have only one use to avoid introducing extra | |||
3457 | // instructions. | |||
3458 | if (SVN0->hasOneUse() && SVN1->hasOneUse() && | |||
3459 | SVN0->getMask().equals(SVN1->getMask())) { | |||
3460 | SDValue ShOp = N0->getOperand(1); | |||
3461 | ||||
3462 | // Don't try to fold this node if it requires introducing a | |||
3463 | // build vector of all zeros that might be illegal at this stage. | |||
3464 | if (N->getOpcode() == ISD::XOR && !ShOp.isUndef()) { | |||
3465 | if (!LegalTypes) | |||
3466 | ShOp = DAG.getConstant(0, SDLoc(N), VT); | |||
3467 | else | |||
3468 | ShOp = SDValue(); | |||
3469 | } | |||
3470 | ||||
3471 | // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C) | |||
3472 | // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C) | |||
3473 | // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0) | |||
3474 | if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) { | |||
3475 | SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT, | |||
3476 | N0->getOperand(0), N1->getOperand(0)); | |||
3477 | AddToWorklist(NewNode.getNode()); | |||
3478 | return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp, | |||
3479 | SVN0->getMask()); | |||
3480 | } | |||
3481 | ||||
3482 | // Don't try to fold this node if it requires introducing a | |||
3483 | // build vector of all zeros that might be illegal at this stage. | |||
3484 | ShOp = N0->getOperand(0); | |||
3485 | if (N->getOpcode() == ISD::XOR && !ShOp.isUndef()) { | |||
3486 | if (!LegalTypes) | |||
3487 | ShOp = DAG.getConstant(0, SDLoc(N), VT); | |||
3488 | else | |||
3489 | ShOp = SDValue(); | |||
3490 | } | |||
3491 | ||||
3492 | // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B)) | |||
3493 | // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B)) | |||
3494 | // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B)) | |||
3495 | if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) { | |||
3496 | SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT, | |||
3497 | N0->getOperand(1), N1->getOperand(1)); | |||
3498 | AddToWorklist(NewNode.getNode()); | |||
3499 | return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode, | |||
3500 | SVN0->getMask()); | |||
3501 | } | |||
3502 | } | |||
3503 | } | |||
3504 | ||||
3505 | return SDValue(); | |||
3506 | } | |||
3507 | ||||
3508 | /// Try to make (and/or setcc (LL, LR), setcc (RL, RR)) more efficient. | |||
3509 | SDValue DAGCombiner::foldLogicOfSetCCs(bool IsAnd, SDValue N0, SDValue N1, | |||
3510 | const SDLoc &DL) { | |||
3511 | SDValue LL, LR, RL, RR, N0CC, N1CC; | |||
3512 | if (!isSetCCEquivalent(N0, LL, LR, N0CC) || | |||
3513 | !isSetCCEquivalent(N1, RL, RR, N1CC)) | |||
3514 | return SDValue(); | |||
3515 | ||||
3516 | assert(N0.getValueType() == N1.getValueType() &&(static_cast <bool> (N0.getValueType() == N1.getValueType () && "Unexpected operand types for bitwise logic op" ) ? void (0) : __assert_fail ("N0.getValueType() == N1.getValueType() && \"Unexpected operand types for bitwise logic op\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3517, __extension__ __PRETTY_FUNCTION__)) | |||
3517 | "Unexpected operand types for bitwise logic op")(static_cast <bool> (N0.getValueType() == N1.getValueType () && "Unexpected operand types for bitwise logic op" ) ? void (0) : __assert_fail ("N0.getValueType() == N1.getValueType() && \"Unexpected operand types for bitwise logic op\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3517, __extension__ __PRETTY_FUNCTION__)); | |||
3518 | assert(LL.getValueType() == LR.getValueType() &&(static_cast <bool> (LL.getValueType() == LR.getValueType () && RL.getValueType() == RR.getValueType() && "Unexpected operand types for setcc") ? void (0) : __assert_fail ("LL.getValueType() == LR.getValueType() && RL.getValueType() == RR.getValueType() && \"Unexpected operand types for setcc\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3520, __extension__ __PRETTY_FUNCTION__)) | |||
3519 | RL.getValueType() == RR.getValueType() &&(static_cast <bool> (LL.getValueType() == LR.getValueType () && RL.getValueType() == RR.getValueType() && "Unexpected operand types for setcc") ? void (0) : __assert_fail ("LL.getValueType() == LR.getValueType() && RL.getValueType() == RR.getValueType() && \"Unexpected operand types for setcc\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3520, __extension__ __PRETTY_FUNCTION__)) | |||
3520 | "Unexpected operand types for setcc")(static_cast <bool> (LL.getValueType() == LR.getValueType () && RL.getValueType() == RR.getValueType() && "Unexpected operand types for setcc") ? void (0) : __assert_fail ("LL.getValueType() == LR.getValueType() && RL.getValueType() == RR.getValueType() && \"Unexpected operand types for setcc\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3520, __extension__ __PRETTY_FUNCTION__)); | |||
3521 | ||||
3522 | // If we're here post-legalization or the logic op type is not i1, the logic | |||
3523 | // op type must match a setcc result type. Also, all folds require new | |||
3524 | // operations on the left and right operands, so those types must match. | |||
3525 | EVT VT = N0.getValueType(); | |||
3526 | EVT OpVT = LL.getValueType(); | |||
3527 | if (LegalOperations || VT.getScalarType() != MVT::i1) | |||
3528 | if (VT != getSetCCResultType(OpVT)) | |||
3529 | return SDValue(); | |||
3530 | if (OpVT != RL.getValueType()) | |||
3531 | return SDValue(); | |||
3532 | ||||
3533 | ISD::CondCode CC0 = cast<CondCodeSDNode>(N0CC)->get(); | |||
3534 | ISD::CondCode CC1 = cast<CondCodeSDNode>(N1CC)->get(); | |||
3535 | bool IsInteger = OpVT.isInteger(); | |||
3536 | if (LR == RR && CC0 == CC1 && IsInteger) { | |||
3537 | bool IsZero = isNullConstantOrNullSplatConstant(LR); | |||
3538 | bool IsNeg1 = isAllOnesConstantOrAllOnesSplatConstant(LR); | |||
3539 | ||||
3540 | // All bits clear? | |||
3541 | bool AndEqZero = IsAnd && CC1 == ISD::SETEQ && IsZero; | |||
3542 | // All sign bits clear? | |||
3543 | bool AndGtNeg1 = IsAnd && CC1 == ISD::SETGT && IsNeg1; | |||
3544 | // Any bits set? | |||
3545 | bool OrNeZero = !IsAnd && CC1 == ISD::SETNE && IsZero; | |||
3546 | // Any sign bits set? | |||
3547 | bool OrLtZero = !IsAnd && CC1 == ISD::SETLT && IsZero; | |||
3548 | ||||
3549 | // (and (seteq X, 0), (seteq Y, 0)) --> (seteq (or X, Y), 0) | |||
3550 | // (and (setgt X, -1), (setgt Y, -1)) --> (setgt (or X, Y), -1) | |||
3551 | // (or (setne X, 0), (setne Y, 0)) --> (setne (or X, Y), 0) | |||
3552 | // (or (setlt X, 0), (setlt Y, 0)) --> (setlt (or X, Y), 0) | |||
3553 | if (AndEqZero || AndGtNeg1 || OrNeZero || OrLtZero) { | |||
3554 | SDValue Or = DAG.getNode(ISD::OR, SDLoc(N0), OpVT, LL, RL); | |||
3555 | AddToWorklist(Or.getNode()); | |||
3556 | return DAG.getSetCC(DL, VT, Or, LR, CC1); | |||
3557 | } | |||
3558 | ||||
3559 | // All bits set? | |||
3560 | bool AndEqNeg1 = IsAnd && CC1 == ISD::SETEQ && IsNeg1; | |||
3561 | // All sign bits set? | |||
3562 | bool AndLtZero = IsAnd && CC1 == ISD::SETLT && IsZero; | |||
3563 | // Any bits clear? | |||
3564 | bool OrNeNeg1 = !IsAnd && CC1 == ISD::SETNE && IsNeg1; | |||
3565 | // Any sign bits clear? | |||
3566 | bool OrGtNeg1 = !IsAnd && CC1 == ISD::SETGT && IsNeg1; | |||
3567 | ||||
3568 | // (and (seteq X, -1), (seteq Y, -1)) --> (seteq (and X, Y), -1) | |||
3569 | // (and (setlt X, 0), (setlt Y, 0)) --> (setlt (and X, Y), 0) | |||
3570 | // (or (setne X, -1), (setne Y, -1)) --> (setne (and X, Y), -1) | |||
3571 | // (or (setgt X, -1), (setgt Y -1)) --> (setgt (and X, Y), -1) | |||
3572 | if (AndEqNeg1 || AndLtZero || OrNeNeg1 || OrGtNeg1) { | |||
3573 | SDValue And = DAG.getNode(ISD::AND, SDLoc(N0), OpVT, LL, RL); | |||
3574 | AddToWorklist(And.getNode()); | |||
3575 | return DAG.getSetCC(DL, VT, And, LR, CC1); | |||
3576 | } | |||
3577 | } | |||
3578 | ||||
3579 | // TODO: What is the 'or' equivalent of this fold? | |||
3580 | // (and (setne X, 0), (setne X, -1)) --> (setuge (add X, 1), 2) | |||
3581 | if (IsAnd && LL == RL && CC0 == CC1 && OpVT.getScalarSizeInBits() > 1 && | |||
3582 | IsInteger && CC0 == ISD::SETNE && | |||
3583 | ((isNullConstant(LR) && isAllOnesConstant(RR)) || | |||
3584 | (isAllOnesConstant(LR) && isNullConstant(RR)))) { | |||
3585 | SDValue One = DAG.getConstant(1, DL, OpVT); | |||
3586 | SDValue Two = DAG.getConstant(2, DL, OpVT); | |||
3587 | SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0), OpVT, LL, One); | |||
3588 | AddToWorklist(Add.getNode()); | |||
3589 | return DAG.getSetCC(DL, VT, Add, Two, ISD::SETUGE); | |||
3590 | } | |||
3591 | ||||
3592 | // Try more general transforms if the predicates match and the only user of | |||
3593 | // the compares is the 'and' or 'or'. | |||
3594 | if (IsInteger && TLI.convertSetCCLogicToBitwiseLogic(OpVT) && CC0 == CC1 && | |||
3595 | N0.hasOneUse() && N1.hasOneUse()) { | |||
3596 | // and (seteq A, B), (seteq C, D) --> seteq (or (xor A, B), (xor C, D)), 0 | |||
3597 | // or (setne A, B), (setne C, D) --> setne (or (xor A, B), (xor C, D)), 0 | |||
3598 | if ((IsAnd && CC1 == ISD::SETEQ) || (!IsAnd && CC1 == ISD::SETNE)) { | |||
3599 | SDValue XorL = DAG.getNode(ISD::XOR, SDLoc(N0), OpVT, LL, LR); | |||
3600 | SDValue XorR = DAG.getNode(ISD::XOR, SDLoc(N1), OpVT, RL, RR); | |||
3601 | SDValue Or = DAG.getNode(ISD::OR, DL, OpVT, XorL, XorR); | |||
3602 | SDValue Zero = DAG.getConstant(0, DL, OpVT); | |||
3603 | return DAG.getSetCC(DL, VT, Or, Zero, CC1); | |||
3604 | } | |||
3605 | } | |||
3606 | ||||
3607 | // Canonicalize equivalent operands to LL == RL. | |||
3608 | if (LL == RR && LR == RL) { | |||
3609 | CC1 = ISD::getSetCCSwappedOperands(CC1); | |||
3610 | std::swap(RL, RR); | |||
3611 | } | |||
3612 | ||||
3613 | // (and (setcc X, Y, CC0), (setcc X, Y, CC1)) --> (setcc X, Y, NewCC) | |||
3614 | // (or (setcc X, Y, CC0), (setcc X, Y, CC1)) --> (setcc X, Y, NewCC) | |||
3615 | if (LL == RL && LR == RR) { | |||
3616 | ISD::CondCode NewCC = IsAnd ? ISD::getSetCCAndOperation(CC0, CC1, IsInteger) | |||
3617 | : ISD::getSetCCOrOperation(CC0, CC1, IsInteger); | |||
3618 | if (NewCC != ISD::SETCC_INVALID && | |||
3619 | (!LegalOperations || | |||
3620 | (TLI.isCondCodeLegal(NewCC, LL.getSimpleValueType()) && | |||
3621 | TLI.isOperationLegal(ISD::SETCC, OpVT)))) | |||
3622 | return DAG.getSetCC(DL, VT, LL, LR, NewCC); | |||
3623 | } | |||
3624 | ||||
3625 | return SDValue(); | |||
3626 | } | |||
3627 | ||||
3628 | /// This contains all DAGCombine rules which reduce two values combined by | |||
3629 | /// an And operation to a single value. This makes them reusable in the context | |||
3630 | /// of visitSELECT(). Rules involving constants are not included as | |||
3631 | /// visitSELECT() already handles those cases. | |||
3632 | SDValue DAGCombiner::visitANDLike(SDValue N0, SDValue N1, SDNode *N) { | |||
3633 | EVT VT = N1.getValueType(); | |||
3634 | SDLoc DL(N); | |||
3635 | ||||
3636 | // fold (and x, undef) -> 0 | |||
3637 | if (N0.isUndef() || N1.isUndef()) | |||
3638 | return DAG.getConstant(0, DL, VT); | |||
3639 | ||||
3640 | if (SDValue V = foldLogicOfSetCCs(true, N0, N1, DL)) | |||
3641 | return V; | |||
3642 | ||||
3643 | if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL && | |||
3644 | VT.getSizeInBits() <= 64) { | |||
3645 | if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { | |||
3646 | if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) { | |||
3647 | // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal | |||
3648 | // immediate for an add, but it is legal if its top c2 bits are set, | |||
3649 | // transform the ADD so the immediate doesn't need to be materialized | |||
3650 | // in a register. | |||
3651 | APInt ADDC = ADDI->getAPIntValue(); | |||
3652 | APInt SRLC = SRLI->getAPIntValue(); | |||
3653 | if (ADDC.getMinSignedBits() <= 64 && | |||
3654 | SRLC.ult(VT.getSizeInBits()) && | |||
3655 | !TLI.isLegalAddImmediate(ADDC.getSExtValue())) { | |||
3656 | APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(), | |||
3657 | SRLC.getZExtValue()); | |||
3658 | if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) { | |||
3659 | ADDC |= Mask; | |||
3660 | if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) { | |||
3661 | SDLoc DL0(N0); | |||
3662 | SDValue NewAdd = | |||
3663 | DAG.getNode(ISD::ADD, DL0, VT, | |||
3664 | N0.getOperand(0), DAG.getConstant(ADDC, DL, VT)); | |||
3665 | CombineTo(N0.getNode(), NewAdd); | |||
3666 | // Return N so it doesn't get rechecked! | |||
3667 | return SDValue(N, 0); | |||
3668 | } | |||
3669 | } | |||
3670 | } | |||
3671 | } | |||
3672 | } | |||
3673 | } | |||
3674 | ||||
3675 | // Reduce bit extract of low half of an integer to the narrower type. | |||
3676 | // (and (srl i64:x, K), KMask) -> | |||
3677 | // (i64 zero_extend (and (srl (i32 (trunc i64:x)), K)), KMask) | |||
3678 | if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) { | |||
3679 | if (ConstantSDNode *CAnd = dyn_cast<ConstantSDNode>(N1)) { | |||
3680 | if (ConstantSDNode *CShift = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { | |||
3681 | unsigned Size = VT.getSizeInBits(); | |||
3682 | const APInt &AndMask = CAnd->getAPIntValue(); | |||
3683 | unsigned ShiftBits = CShift->getZExtValue(); | |||
3684 | ||||
3685 | // Bail out, this node will probably disappear anyway. | |||
3686 | if (ShiftBits == 0) | |||
3687 | return SDValue(); | |||
3688 | ||||
3689 | unsigned MaskBits = AndMask.countTrailingOnes(); | |||
3690 | EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), Size / 2); | |||
3691 | ||||
3692 | if (AndMask.isMask() && | |||
3693 | // Required bits must not span the two halves of the integer and | |||
3694 | // must fit in the half size type. | |||
3695 | (ShiftBits + MaskBits <= Size / 2) && | |||
3696 | TLI.isNarrowingProfitable(VT, HalfVT) && | |||
3697 | TLI.isTypeDesirableForOp(ISD::AND, HalfVT) && | |||
3698 | TLI.isTypeDesirableForOp(ISD::SRL, HalfVT) && | |||
3699 | TLI.isTruncateFree(VT, HalfVT) && | |||
3700 | TLI.isZExtFree(HalfVT, VT)) { | |||
3701 | // The isNarrowingProfitable is to avoid regressions on PPC and | |||
3702 | // AArch64 which match a few 64-bit bit insert / bit extract patterns | |||
3703 | // on downstream users of this. Those patterns could probably be | |||
3704 | // extended to handle extensions mixed in. | |||
3705 | ||||
3706 | SDValue SL(N0); | |||
3707 | assert(MaskBits <= Size)(static_cast <bool> (MaskBits <= Size) ? void (0) : __assert_fail ("MaskBits <= Size", "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3707, __extension__ __PRETTY_FUNCTION__)); | |||
3708 | ||||
3709 | // Extracting the highest bit of the low half. | |||
3710 | EVT ShiftVT = TLI.getShiftAmountTy(HalfVT, DAG.getDataLayout()); | |||
3711 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SL, HalfVT, | |||
3712 | N0.getOperand(0)); | |||
3713 | ||||
3714 | SDValue NewMask = DAG.getConstant(AndMask.trunc(Size / 2), SL, HalfVT); | |||
3715 | SDValue ShiftK = DAG.getConstant(ShiftBits, SL, ShiftVT); | |||
3716 | SDValue Shift = DAG.getNode(ISD::SRL, SL, HalfVT, Trunc, ShiftK); | |||
3717 | SDValue And = DAG.getNode(ISD::AND, SL, HalfVT, Shift, NewMask); | |||
3718 | return DAG.getNode(ISD::ZERO_EXTEND, SL, VT, And); | |||
3719 | } | |||
3720 | } | |||
3721 | } | |||
3722 | } | |||
3723 | ||||
3724 | return SDValue(); | |||
3725 | } | |||
3726 | ||||
3727 | bool DAGCombiner::isAndLoadExtLoad(ConstantSDNode *AndC, LoadSDNode *LoadN, | |||
3728 | EVT LoadResultTy, EVT &ExtVT) { | |||
3729 | if (!AndC->getAPIntValue().isMask()) | |||
3730 | return false; | |||
3731 | ||||
3732 | unsigned ActiveBits = AndC->getAPIntValue().countTrailingOnes(); | |||
3733 | ||||
3734 | ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits); | |||
3735 | EVT LoadedVT = LoadN->getMemoryVT(); | |||
3736 | ||||
3737 | if (ExtVT == LoadedVT && | |||
3738 | (!LegalOperations || | |||
3739 | TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy, ExtVT))) { | |||
3740 | // ZEXTLOAD will match without needing to change the size of the value being | |||
3741 | // loaded. | |||
3742 | return true; | |||
3743 | } | |||
3744 | ||||
3745 | // Do not change the width of a volatile load. | |||
3746 | if (LoadN->isVolatile()) | |||
3747 | return false; | |||
3748 | ||||
3749 | // Do not generate loads of non-round integer types since these can | |||
3750 | // be expensive (and would be wrong if the type is not byte sized). | |||
3751 | if (!LoadedVT.bitsGT(ExtVT) || !ExtVT.isRound()) | |||
3752 | return false; | |||
3753 | ||||
3754 | if (LegalOperations && | |||
3755 | !TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy, ExtVT)) | |||
3756 | return false; | |||
3757 | ||||
3758 | if (!TLI.shouldReduceLoadWidth(LoadN, ISD::ZEXTLOAD, ExtVT)) | |||
3759 | return false; | |||
3760 | ||||
3761 | return true; | |||
3762 | } | |||
3763 | ||||
3764 | bool DAGCombiner::isLegalNarrowLoad(LoadSDNode *LoadN, ISD::LoadExtType ExtType, | |||
3765 | EVT &ExtVT, unsigned ShAmt) { | |||
3766 | // Don't transform one with multiple uses, this would require adding a new | |||
3767 | // load. | |||
3768 | if (!SDValue(LoadN, 0).hasOneUse()) | |||
3769 | return false; | |||
3770 | ||||
3771 | if (LegalOperations && | |||
3772 | !TLI.isLoadExtLegal(ExtType, LoadN->getValueType(0), ExtVT)) | |||
3773 | return false; | |||
3774 | ||||
3775 | // Do not generate loads of non-round integer types since these can | |||
3776 | // be expensive (and would be wrong if the type is not byte sized). | |||
3777 | if (!ExtVT.isRound()) | |||
3778 | return false; | |||
3779 | ||||
3780 | // Don't change the width of a volatile load. | |||
3781 | if (LoadN->isVolatile()) | |||
3782 | return false; | |||
3783 | ||||
3784 | // Verify that we are actually reducing a load width here. | |||
3785 | if (LoadN->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits()) | |||
3786 | return false; | |||
3787 | ||||
3788 | // For the transform to be legal, the load must produce only two values | |||
3789 | // (the value loaded and the chain). Don't transform a pre-increment | |||
3790 | // load, for example, which produces an extra value. Otherwise the | |||
3791 | // transformation is not equivalent, and the downstream logic to replace | |||
3792 | // uses gets things wrong. | |||
3793 | if (LoadN->getNumValues() > 2) | |||
3794 | return false; | |||
3795 | ||||
3796 | // If the load that we're shrinking is an extload and we're not just | |||
3797 | // discarding the extension we can't simply shrink the load. Bail. | |||
3798 | // TODO: It would be possible to merge the extensions in some cases. | |||
3799 | if (LoadN->getExtensionType() != ISD::NON_EXTLOAD && | |||
3800 | LoadN->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt) | |||
3801 | return false; | |||
3802 | ||||
3803 | if (!TLI.shouldReduceLoadWidth(LoadN, ExtType, ExtVT)) | |||
3804 | return false; | |||
3805 | ||||
3806 | // It's not possible to generate a constant of extended or untyped type. | |||
3807 | EVT PtrType = LoadN->getOperand(1).getValueType(); | |||
3808 | if (PtrType == MVT::Untyped || PtrType.isExtended()) | |||
3809 | return false; | |||
3810 | ||||
3811 | return true; | |||
3812 | } | |||
3813 | ||||
3814 | bool DAGCombiner::SearchForAndLoads(SDNode *N, | |||
3815 | SmallPtrSetImpl<LoadSDNode*> &Loads, | |||
3816 | SmallPtrSetImpl<SDNode*> &NodesWithConsts, | |||
3817 | ConstantSDNode *Mask, | |||
3818 | SDNode *&NodeToMask) { | |||
3819 | // Recursively search for the operands, looking for loads which can be | |||
3820 | // narrowed. | |||
3821 | for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i) { | |||
3822 | SDValue Op = N->getOperand(i); | |||
3823 | ||||
3824 | if (Op.getValueType().isVector()) | |||
3825 | return false; | |||
3826 | ||||
3827 | // Some constants may need fixing up later if they are too large. | |||
3828 | if (auto *C = dyn_cast<ConstantSDNode>(Op)) { | |||
3829 | if ((N->getOpcode() == ISD::OR || N->getOpcode() == ISD::XOR) && | |||
3830 | (Mask->getAPIntValue() & C->getAPIntValue()) != C->getAPIntValue()) | |||
3831 | NodesWithConsts.insert(N); | |||
3832 | continue; | |||
3833 | } | |||
3834 | ||||
3835 | if (!Op.hasOneUse()) | |||
3836 | return false; | |||
3837 | ||||
3838 | switch(Op.getOpcode()) { | |||
3839 | case ISD::LOAD: { | |||
3840 | auto *Load = cast<LoadSDNode>(Op); | |||
3841 | EVT ExtVT; | |||
3842 | if (isAndLoadExtLoad(Mask, Load, Load->getValueType(0), ExtVT) && | |||
3843 | isLegalNarrowLoad(Load, ISD::ZEXTLOAD, ExtVT)) { | |||
3844 | ||||
3845 | // ZEXTLOAD is already small enough. | |||
3846 | if (Load->getExtensionType() == ISD::ZEXTLOAD && | |||
3847 | ExtVT.bitsGE(Load->getMemoryVT())) | |||
3848 | continue; | |||
3849 | ||||
3850 | // Use LE to convert equal sized loads to zext. | |||
3851 | if (ExtVT.bitsLE(Load->getMemoryVT())) | |||
3852 | Loads.insert(Load); | |||
3853 | ||||
3854 | continue; | |||
3855 | } | |||
3856 | return false; | |||
3857 | } | |||
3858 | case ISD::ZERO_EXTEND: | |||
3859 | case ISD::AssertZext: { | |||
3860 | unsigned ActiveBits = Mask->getAPIntValue().countTrailingOnes(); | |||
3861 | EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits); | |||
3862 | EVT VT = Op.getOpcode() == ISD::AssertZext ? | |||
3863 | cast<VTSDNode>(Op.getOperand(1))->getVT() : | |||
3864 | Op.getOperand(0).getValueType(); | |||
3865 | ||||
3866 | // We can accept extending nodes if the mask is wider or an equal | |||
3867 | // width to the original type. | |||
3868 | if (ExtVT.bitsGE(VT)) | |||
3869 | continue; | |||
3870 | break; | |||
3871 | } | |||
3872 | case ISD::OR: | |||
3873 | case ISD::XOR: | |||
3874 | case ISD::AND: | |||
3875 | if (!SearchForAndLoads(Op.getNode(), Loads, NodesWithConsts, Mask, | |||
3876 | NodeToMask)) | |||
3877 | return false; | |||
3878 | continue; | |||
3879 | } | |||
3880 | ||||
3881 | // Allow one node which will masked along with any loads found. | |||
3882 | if (NodeToMask) | |||
3883 | return false; | |||
3884 | NodeToMask = Op.getNode(); | |||
3885 | } | |||
3886 | return true; | |||
3887 | } | |||
3888 | ||||
3889 | bool DAGCombiner::BackwardsPropagateMask(SDNode *N, SelectionDAG &DAG) { | |||
3890 | auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(1)); | |||
3891 | if (!Mask) | |||
3892 | return false; | |||
3893 | ||||
3894 | if (!Mask->getAPIntValue().isMask()) | |||
3895 | return false; | |||
3896 | ||||
3897 | // No need to do anything if the and directly uses a load. | |||
3898 | if (isa<LoadSDNode>(N->getOperand(0))) | |||
3899 | return false; | |||
3900 | ||||
3901 | SmallPtrSet<LoadSDNode*, 8> Loads; | |||
3902 | SmallPtrSet<SDNode*, 2> NodesWithConsts; | |||
3903 | SDNode *FixupNode = nullptr; | |||
3904 | if (SearchForAndLoads(N, Loads, NodesWithConsts, Mask, FixupNode)) { | |||
3905 | if (Loads.size() == 0) | |||
3906 | return false; | |||
3907 | ||||
3908 | DEBUG(dbgs() << "Backwards propagate AND: "; N->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "Backwards propagate AND: " ; N->dump(); } } while (false); | |||
3909 | SDValue MaskOp = N->getOperand(1); | |||
3910 | ||||
3911 | // If it exists, fixup the single node we allow in the tree that needs | |||
3912 | // masking. | |||
3913 | if (FixupNode) { | |||
3914 | DEBUG(dbgs() << "First, need to fix up: "; FixupNode->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "First, need to fix up: "; FixupNode ->dump(); } } while (false); | |||
3915 | SDValue And = DAG.getNode(ISD::AND, SDLoc(FixupNode), | |||
3916 | FixupNode->getValueType(0), | |||
3917 | SDValue(FixupNode, 0), MaskOp); | |||
3918 | DAG.ReplaceAllUsesOfValueWith(SDValue(FixupNode, 0), And); | |||
3919 | DAG.UpdateNodeOperands(And.getNode(), SDValue(FixupNode, 0), | |||
3920 | MaskOp); | |||
3921 | } | |||
3922 | ||||
3923 | // Narrow any constants that need it. | |||
3924 | for (auto *LogicN : NodesWithConsts) { | |||
3925 | SDValue Op0 = LogicN->getOperand(0); | |||
3926 | SDValue Op1 = LogicN->getOperand(1); | |||
3927 | ||||
3928 | if (isa<ConstantSDNode>(Op0)) | |||
3929 | std::swap(Op0, Op1); | |||
3930 | ||||
3931 | SDValue And = DAG.getNode(ISD::AND, SDLoc(Op1), Op1.getValueType(), | |||
3932 | Op1, MaskOp); | |||
3933 | ||||
3934 | DAG.UpdateNodeOperands(LogicN, Op0, And); | |||
3935 | } | |||
3936 | ||||
3937 | // Create narrow loads. | |||
3938 | for (auto *Load : Loads) { | |||
3939 | DEBUG(dbgs() << "Propagate AND back to: "; Load->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "Propagate AND back to: "; Load ->dump(); } } while (false); | |||
3940 | SDValue And = DAG.getNode(ISD::AND, SDLoc(Load), Load->getValueType(0), | |||
3941 | SDValue(Load, 0), MaskOp); | |||
3942 | DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), And); | |||
3943 | DAG.UpdateNodeOperands(And.getNode(), SDValue(Load, 0), MaskOp); | |||
3944 | SDValue NewLoad = ReduceLoadWidth(And.getNode()); | |||
3945 | assert(NewLoad &&(static_cast <bool> (NewLoad && "Shouldn't be masking the load if it can't be narrowed" ) ? void (0) : __assert_fail ("NewLoad && \"Shouldn't be masking the load if it can't be narrowed\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3946, __extension__ __PRETTY_FUNCTION__)) | |||
3946 | "Shouldn't be masking the load if it can't be narrowed")(static_cast <bool> (NewLoad && "Shouldn't be masking the load if it can't be narrowed" ) ? void (0) : __assert_fail ("NewLoad && \"Shouldn't be masking the load if it can't be narrowed\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 3946, __extension__ __PRETTY_FUNCTION__)); | |||
3947 | CombineTo(Load, NewLoad, NewLoad.getValue(1)); | |||
3948 | } | |||
3949 | DAG.ReplaceAllUsesWith(N, N->getOperand(0).getNode()); | |||
3950 | return true; | |||
3951 | } | |||
3952 | return false; | |||
3953 | } | |||
3954 | ||||
3955 | SDValue DAGCombiner::visitAND(SDNode *N) { | |||
3956 | SDValue N0 = N->getOperand(0); | |||
3957 | SDValue N1 = N->getOperand(1); | |||
3958 | EVT VT = N1.getValueType(); | |||
3959 | ||||
3960 | // x & x --> x | |||
3961 | if (N0 == N1) | |||
3962 | return N0; | |||
3963 | ||||
3964 | // fold vector ops | |||
3965 | if (VT.isVector()) { | |||
3966 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
3967 | return FoldedVOp; | |||
3968 | ||||
3969 | // fold (and x, 0) -> 0, vector edition | |||
3970 | if (ISD::isBuildVectorAllZeros(N0.getNode())) | |||
3971 | // do not return N0, because undef node may exist in N0 | |||
3972 | return DAG.getConstant(APInt::getNullValue(N0.getScalarValueSizeInBits()), | |||
3973 | SDLoc(N), N0.getValueType()); | |||
3974 | if (ISD::isBuildVectorAllZeros(N1.getNode())) | |||
3975 | // do not return N1, because undef node may exist in N1 | |||
3976 | return DAG.getConstant(APInt::getNullValue(N1.getScalarValueSizeInBits()), | |||
3977 | SDLoc(N), N1.getValueType()); | |||
3978 | ||||
3979 | // fold (and x, -1) -> x, vector edition | |||
3980 | if (ISD::isBuildVectorAllOnes(N0.getNode())) | |||
3981 | return N1; | |||
3982 | if (ISD::isBuildVectorAllOnes(N1.getNode())) | |||
3983 | return N0; | |||
3984 | } | |||
3985 | ||||
3986 | // fold (and c1, c2) -> c1&c2 | |||
3987 | ConstantSDNode *N0C = getAsNonOpaqueConstant(N0); | |||
3988 | ConstantSDNode *N1C = isConstOrConstSplat(N1); | |||
3989 | if (N0C && N1C && !N1C->isOpaque()) | |||
3990 | return DAG.FoldConstantArithmetic(ISD::AND, SDLoc(N), VT, N0C, N1C); | |||
3991 | // canonicalize constant to RHS | |||
3992 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
3993 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) | |||
3994 | return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0); | |||
3995 | // fold (and x, -1) -> x | |||
3996 | if (isAllOnesConstant(N1)) | |||
3997 | return N0; | |||
3998 | // if (and x, c) is known to be zero, return 0 | |||
3999 | unsigned BitWidth = VT.getScalarSizeInBits(); | |||
4000 | if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0), | |||
4001 | APInt::getAllOnesValue(BitWidth))) | |||
4002 | return DAG.getConstant(0, SDLoc(N), VT); | |||
4003 | ||||
4004 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
4005 | return NewSel; | |||
4006 | ||||
4007 | // reassociate and | |||
4008 | if (SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1)) | |||
4009 | return RAND; | |||
4010 | ||||
4011 | // Try to convert a constant mask AND into a shuffle clear mask. | |||
4012 | if (VT.isVector()) | |||
4013 | if (SDValue Shuffle = XformToShuffleWithZero(N)) | |||
4014 | return Shuffle; | |||
4015 | ||||
4016 | // fold (and (or x, C), D) -> D if (C & D) == D | |||
4017 | auto MatchSubset = [](ConstantSDNode *LHS, ConstantSDNode *RHS) { | |||
4018 | return RHS->getAPIntValue().isSubsetOf(LHS->getAPIntValue()); | |||
4019 | }; | |||
4020 | if (N0.getOpcode() == ISD::OR && | |||
4021 | ISD::matchBinaryPredicate(N0.getOperand(1), N1, MatchSubset)) | |||
4022 | return N1; | |||
4023 | // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits. | |||
4024 | if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) { | |||
4025 | SDValue N0Op0 = N0.getOperand(0); | |||
4026 | APInt Mask = ~N1C->getAPIntValue(); | |||
4027 | Mask = Mask.trunc(N0Op0.getScalarValueSizeInBits()); | |||
4028 | if (DAG.MaskedValueIsZero(N0Op0, Mask)) { | |||
4029 | SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), | |||
4030 | N0.getValueType(), N0Op0); | |||
4031 | ||||
4032 | // Replace uses of the AND with uses of the Zero extend node. | |||
4033 | CombineTo(N, Zext); | |||
4034 | ||||
4035 | // We actually want to replace all uses of the any_extend with the | |||
4036 | // zero_extend, to avoid duplicating things. This will later cause this | |||
4037 | // AND to be folded. | |||
4038 | CombineTo(N0.getNode(), Zext); | |||
4039 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
4040 | } | |||
4041 | } | |||
4042 | // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) -> | |||
4043 | // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must | |||
4044 | // already be zero by virtue of the width of the base type of the load. | |||
4045 | // | |||
4046 | // the 'X' node here can either be nothing or an extract_vector_elt to catch | |||
4047 | // more cases. | |||
4048 | if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT && | |||
4049 | N0.getValueSizeInBits() == N0.getOperand(0).getScalarValueSizeInBits() && | |||
4050 | N0.getOperand(0).getOpcode() == ISD::LOAD && | |||
4051 | N0.getOperand(0).getResNo() == 0) || | |||
4052 | (N0.getOpcode() == ISD::LOAD && N0.getResNo() == 0)) { | |||
4053 | LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ? | |||
4054 | N0 : N0.getOperand(0) ); | |||
4055 | ||||
4056 | // Get the constant (if applicable) the zero'th operand is being ANDed with. | |||
4057 | // This can be a pure constant or a vector splat, in which case we treat the | |||
4058 | // vector as a scalar and use the splat value. | |||
4059 | APInt Constant = APInt::getNullValue(1); | |||
4060 | if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) { | |||
4061 | Constant = C->getAPIntValue(); | |||
4062 | } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) { | |||
4063 | APInt SplatValue, SplatUndef; | |||
4064 | unsigned SplatBitSize; | |||
4065 | bool HasAnyUndefs; | |||
4066 | bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef, | |||
4067 | SplatBitSize, HasAnyUndefs); | |||
4068 | if (IsSplat) { | |||
4069 | // Undef bits can contribute to a possible optimisation if set, so | |||
4070 | // set them. | |||
4071 | SplatValue |= SplatUndef; | |||
4072 | ||||
4073 | // The splat value may be something like "0x00FFFFFF", which means 0 for | |||
4074 | // the first vector value and FF for the rest, repeating. We need a mask | |||
4075 | // that will apply equally to all members of the vector, so AND all the | |||
4076 | // lanes of the constant together. | |||
4077 | EVT VT = Vector->getValueType(0); | |||
4078 | unsigned BitWidth = VT.getScalarSizeInBits(); | |||
4079 | ||||
4080 | // If the splat value has been compressed to a bitlength lower | |||
4081 | // than the size of the vector lane, we need to re-expand it to | |||
4082 | // the lane size. | |||
4083 | if (BitWidth > SplatBitSize) | |||
4084 | for (SplatValue = SplatValue.zextOrTrunc(BitWidth); | |||
4085 | SplatBitSize < BitWidth; | |||
4086 | SplatBitSize = SplatBitSize * 2) | |||
4087 | SplatValue |= SplatValue.shl(SplatBitSize); | |||
4088 | ||||
4089 | // Make sure that variable 'Constant' is only set if 'SplatBitSize' is a | |||
4090 | // multiple of 'BitWidth'. Otherwise, we could propagate a wrong value. | |||
4091 | if (SplatBitSize % BitWidth == 0) { | |||
4092 | Constant = APInt::getAllOnesValue(BitWidth); | |||
4093 | for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i) | |||
4094 | Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth); | |||
4095 | } | |||
4096 | } | |||
4097 | } | |||
4098 | ||||
4099 | // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is | |||
4100 | // actually legal and isn't going to get expanded, else this is a false | |||
4101 | // optimisation. | |||
4102 | bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD, | |||
4103 | Load->getValueType(0), | |||
4104 | Load->getMemoryVT()); | |||
4105 | ||||
4106 | // Resize the constant to the same size as the original memory access before | |||
4107 | // extension. If it is still the AllOnesValue then this AND is completely | |||
4108 | // unneeded. | |||
4109 | Constant = Constant.zextOrTrunc(Load->getMemoryVT().getScalarSizeInBits()); | |||
4110 | ||||
4111 | bool B; | |||
4112 | switch (Load->getExtensionType()) { | |||
4113 | default: B = false; break; | |||
4114 | case ISD::EXTLOAD: B = CanZextLoadProfitably; break; | |||
4115 | case ISD::ZEXTLOAD: | |||
4116 | case ISD::NON_EXTLOAD: B = true; break; | |||
4117 | } | |||
4118 | ||||
4119 | if (B && Constant.isAllOnesValue()) { | |||
4120 | // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to | |||
4121 | // preserve semantics once we get rid of the AND. | |||
4122 | SDValue NewLoad(Load, 0); | |||
4123 | ||||
4124 | // Fold the AND away. NewLoad may get replaced immediately. | |||
4125 | CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0); | |||
4126 | ||||
4127 | if (Load->getExtensionType() == ISD::EXTLOAD) { | |||
4128 | NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD, | |||
4129 | Load->getValueType(0), SDLoc(Load), | |||
4130 | Load->getChain(), Load->getBasePtr(), | |||
4131 | Load->getOffset(), Load->getMemoryVT(), | |||
4132 | Load->getMemOperand()); | |||
4133 | // Replace uses of the EXTLOAD with the new ZEXTLOAD. | |||
4134 | if (Load->getNumValues() == 3) { | |||
4135 | // PRE/POST_INC loads have 3 values. | |||
4136 | SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1), | |||
4137 | NewLoad.getValue(2) }; | |||
4138 | CombineTo(Load, To, 3, true); | |||
4139 | } else { | |||
4140 | CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1)); | |||
4141 | } | |||
4142 | } | |||
4143 | ||||
4144 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
4145 | } | |||
4146 | } | |||
4147 | ||||
4148 | // fold (and (load x), 255) -> (zextload x, i8) | |||
4149 | // fold (and (extload x, i16), 255) -> (zextload x, i8) | |||
4150 | // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8) | |||
4151 | if (!VT.isVector() && N1C && (N0.getOpcode() == ISD::LOAD || | |||
4152 | (N0.getOpcode() == ISD::ANY_EXTEND && | |||
4153 | N0.getOperand(0).getOpcode() == ISD::LOAD))) { | |||
4154 | if (SDValue Res = ReduceLoadWidth(N)) { | |||
4155 | LoadSDNode *LN0 = N0->getOpcode() == ISD::ANY_EXTEND | |||
4156 | ? cast<LoadSDNode>(N0.getOperand(0)) : cast<LoadSDNode>(N0); | |||
4157 | ||||
4158 | AddToWorklist(N); | |||
4159 | CombineTo(LN0, Res, Res.getValue(1)); | |||
4160 | return SDValue(N, 0); | |||
4161 | } | |||
4162 | } | |||
4163 | ||||
4164 | if (Level >= AfterLegalizeTypes) { | |||
4165 | // Attempt to propagate the AND back up to the leaves which, if they're | |||
4166 | // loads, can be combined to narrow loads and the AND node can be removed. | |||
4167 | // Perform after legalization so that extend nodes will already be | |||
4168 | // combined into the loads. | |||
4169 | if (BackwardsPropagateMask(N, DAG)) { | |||
4170 | return SDValue(N, 0); | |||
4171 | } | |||
4172 | } | |||
4173 | ||||
4174 | if (SDValue Combined = visitANDLike(N0, N1, N)) | |||
4175 | return Combined; | |||
4176 | ||||
4177 | // Simplify: (and (op x...), (op y...)) -> (op (and x, y)) | |||
4178 | if (N0.getOpcode() == N1.getOpcode()) | |||
4179 | if (SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N)) | |||
4180 | return Tmp; | |||
4181 | ||||
4182 | // Masking the negated extension of a boolean is just the zero-extended | |||
4183 | // boolean: | |||
4184 | // and (sub 0, zext(bool X)), 1 --> zext(bool X) | |||
4185 | // and (sub 0, sext(bool X)), 1 --> zext(bool X) | |||
4186 | // | |||
4187 | // Note: the SimplifyDemandedBits fold below can make an information-losing | |||
4188 | // transform, and then we have no way to find this better fold. | |||
4189 | if (N1C && N1C->isOne() && N0.getOpcode() == ISD::SUB) { | |||
4190 | if (isNullConstantOrNullSplatConstant(N0.getOperand(0))) { | |||
4191 | SDValue SubRHS = N0.getOperand(1); | |||
4192 | if (SubRHS.getOpcode() == ISD::ZERO_EXTEND && | |||
4193 | SubRHS.getOperand(0).getScalarValueSizeInBits() == 1) | |||
4194 | return SubRHS; | |||
4195 | if (SubRHS.getOpcode() == ISD::SIGN_EXTEND && | |||
4196 | SubRHS.getOperand(0).getScalarValueSizeInBits() == 1) | |||
4197 | return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, SubRHS.getOperand(0)); | |||
4198 | } | |||
4199 | } | |||
4200 | ||||
4201 | // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1) | |||
4202 | // fold (and (sra)) -> (and (srl)) when possible. | |||
4203 | if (SimplifyDemandedBits(SDValue(N, 0))) | |||
4204 | return SDValue(N, 0); | |||
4205 | ||||
4206 | // fold (zext_inreg (extload x)) -> (zextload x) | |||
4207 | if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) { | |||
4208 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
4209 | EVT MemVT = LN0->getMemoryVT(); | |||
4210 | // If we zero all the possible extended bits, then we can turn this into | |||
4211 | // a zextload if we are running before legalize or the operation is legal. | |||
4212 | unsigned BitWidth = N1.getScalarValueSizeInBits(); | |||
4213 | if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth, | |||
4214 | BitWidth - MemVT.getScalarSizeInBits())) && | |||
4215 | ((!LegalOperations && !LN0->isVolatile()) || | |||
4216 | TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) { | |||
4217 | SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT, | |||
4218 | LN0->getChain(), LN0->getBasePtr(), | |||
4219 | MemVT, LN0->getMemOperand()); | |||
4220 | AddToWorklist(N); | |||
4221 | CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); | |||
4222 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
4223 | } | |||
4224 | } | |||
4225 | // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use | |||
4226 | if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) && | |||
4227 | N0.hasOneUse()) { | |||
4228 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
4229 | EVT MemVT = LN0->getMemoryVT(); | |||
4230 | // If we zero all the possible extended bits, then we can turn this into | |||
4231 | // a zextload if we are running before legalize or the operation is legal. | |||
4232 | unsigned BitWidth = N1.getScalarValueSizeInBits(); | |||
4233 | if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth, | |||
4234 | BitWidth - MemVT.getScalarSizeInBits())) && | |||
4235 | ((!LegalOperations && !LN0->isVolatile()) || | |||
4236 | TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) { | |||
4237 | SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT, | |||
4238 | LN0->getChain(), LN0->getBasePtr(), | |||
4239 | MemVT, LN0->getMemOperand()); | |||
4240 | AddToWorklist(N); | |||
4241 | CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); | |||
4242 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
4243 | } | |||
4244 | } | |||
4245 | // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const) | |||
4246 | if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) { | |||
4247 | if (SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0), | |||
4248 | N0.getOperand(1), false)) | |||
4249 | return BSwap; | |||
4250 | } | |||
4251 | ||||
4252 | return SDValue(); | |||
4253 | } | |||
4254 | ||||
4255 | /// Match (a >> 8) | (a << 8) as (bswap a) >> 16. | |||
4256 | SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1, | |||
4257 | bool DemandHighBits) { | |||
4258 | if (!LegalOperations) | |||
4259 | return SDValue(); | |||
4260 | ||||
4261 | EVT VT = N->getValueType(0); | |||
4262 | if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16) | |||
4263 | return SDValue(); | |||
4264 | if (!TLI.isOperationLegalOrCustom(ISD::BSWAP, VT)) | |||
4265 | return SDValue(); | |||
4266 | ||||
4267 | // Recognize (and (shl a, 8), 0xff00), (and (srl a, 8), 0xff) | |||
4268 | bool LookPassAnd0 = false; | |||
4269 | bool LookPassAnd1 = false; | |||
4270 | if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL) | |||
4271 | std::swap(N0, N1); | |||
4272 | if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL) | |||
4273 | std::swap(N0, N1); | |||
4274 | if (N0.getOpcode() == ISD::AND) { | |||
4275 | if (!N0.getNode()->hasOneUse()) | |||
4276 | return SDValue(); | |||
4277 | ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); | |||
4278 | if (!N01C || N01C->getZExtValue() != 0xFF00) | |||
4279 | return SDValue(); | |||
4280 | N0 = N0.getOperand(0); | |||
4281 | LookPassAnd0 = true; | |||
4282 | } | |||
4283 | ||||
4284 | if (N1.getOpcode() == ISD::AND) { | |||
4285 | if (!N1.getNode()->hasOneUse()) | |||
4286 | return SDValue(); | |||
4287 | ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); | |||
4288 | if (!N11C || N11C->getZExtValue() != 0xFF) | |||
4289 | return SDValue(); | |||
4290 | N1 = N1.getOperand(0); | |||
4291 | LookPassAnd1 = true; | |||
4292 | } | |||
4293 | ||||
4294 | if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL) | |||
4295 | std::swap(N0, N1); | |||
4296 | if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL) | |||
4297 | return SDValue(); | |||
4298 | if (!N0.getNode()->hasOneUse() || !N1.getNode()->hasOneUse()) | |||
4299 | return SDValue(); | |||
4300 | ||||
4301 | ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); | |||
4302 | ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); | |||
4303 | if (!N01C || !N11C) | |||
4304 | return SDValue(); | |||
4305 | if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8) | |||
4306 | return SDValue(); | |||
4307 | ||||
4308 | // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8) | |||
4309 | SDValue N00 = N0->getOperand(0); | |||
4310 | if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) { | |||
4311 | if (!N00.getNode()->hasOneUse()) | |||
4312 | return SDValue(); | |||
4313 | ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1)); | |||
4314 | if (!N001C || N001C->getZExtValue() != 0xFF) | |||
4315 | return SDValue(); | |||
4316 | N00 = N00.getOperand(0); | |||
4317 | LookPassAnd0 = true; | |||
4318 | } | |||
4319 | ||||
4320 | SDValue N10 = N1->getOperand(0); | |||
4321 | if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) { | |||
4322 | if (!N10.getNode()->hasOneUse()) | |||
4323 | return SDValue(); | |||
4324 | ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1)); | |||
4325 | if (!N101C || N101C->getZExtValue() != 0xFF00) | |||
4326 | return SDValue(); | |||
4327 | N10 = N10.getOperand(0); | |||
4328 | LookPassAnd1 = true; | |||
4329 | } | |||
4330 | ||||
4331 | if (N00 != N10) | |||
4332 | return SDValue(); | |||
4333 | ||||
4334 | // Make sure everything beyond the low halfword gets set to zero since the SRL | |||
4335 | // 16 will clear the top bits. | |||
4336 | unsigned OpSizeInBits = VT.getSizeInBits(); | |||
4337 | if (DemandHighBits && OpSizeInBits > 16) { | |||
4338 | // If the left-shift isn't masked out then the only way this is a bswap is | |||
4339 | // if all bits beyond the low 8 are 0. In that case the entire pattern | |||
4340 | // reduces to a left shift anyway: leave it for other parts of the combiner. | |||
4341 | if (!LookPassAnd0) | |||
4342 | return SDValue(); | |||
4343 | ||||
4344 | // However, if the right shift isn't masked out then it might be because | |||
4345 | // it's not needed. See if we can spot that too. | |||
4346 | if (!LookPassAnd1 && | |||
4347 | !DAG.MaskedValueIsZero( | |||
4348 | N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16))) | |||
4349 | return SDValue(); | |||
4350 | } | |||
4351 | ||||
4352 | SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00); | |||
4353 | if (OpSizeInBits > 16) { | |||
4354 | SDLoc DL(N); | |||
4355 | Res = DAG.getNode(ISD::SRL, DL, VT, Res, | |||
4356 | DAG.getConstant(OpSizeInBits - 16, DL, | |||
4357 | getShiftAmountTy(VT))); | |||
4358 | } | |||
4359 | return Res; | |||
4360 | } | |||
4361 | ||||
4362 | /// Return true if the specified node is an element that makes up a 32-bit | |||
4363 | /// packed halfword byteswap. | |||
4364 | /// ((x & 0x000000ff) << 8) | | |||
4365 | /// ((x & 0x0000ff00) >> 8) | | |||
4366 | /// ((x & 0x00ff0000) << 8) | | |||
4367 | /// ((x & 0xff000000) >> 8) | |||
4368 | static bool isBSwapHWordElement(SDValue N, MutableArrayRef<SDNode *> Parts) { | |||
4369 | if (!N.getNode()->hasOneUse()) | |||
4370 | return false; | |||
4371 | ||||
4372 | unsigned Opc = N.getOpcode(); | |||
4373 | if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL) | |||
4374 | return false; | |||
4375 | ||||
4376 | SDValue N0 = N.getOperand(0); | |||
4377 | unsigned Opc0 = N0.getOpcode(); | |||
4378 | if (Opc0 != ISD::AND && Opc0 != ISD::SHL && Opc0 != ISD::SRL) | |||
4379 | return false; | |||
4380 | ||||
4381 | ConstantSDNode *N1C = nullptr; | |||
4382 | // SHL or SRL: look upstream for AND mask operand | |||
4383 | if (Opc == ISD::AND) | |||
4384 | N1C = dyn_cast<ConstantSDNode>(N.getOperand(1)); | |||
4385 | else if (Opc0 == ISD::AND) | |||
4386 | N1C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); | |||
4387 | if (!N1C) | |||
4388 | return false; | |||
4389 | ||||
4390 | unsigned MaskByteOffset; | |||
4391 | switch (N1C->getZExtValue()) { | |||
4392 | default: | |||
4393 | return false; | |||
4394 | case 0xFF: MaskByteOffset = 0; break; | |||
4395 | case 0xFF00: MaskByteOffset = 1; break; | |||
4396 | case 0xFF0000: MaskByteOffset = 2; break; | |||
4397 | case 0xFF000000: MaskByteOffset = 3; break; | |||
4398 | } | |||
4399 | ||||
4400 | // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00). | |||
4401 | if (Opc == ISD::AND) { | |||
4402 | if (MaskByteOffset == 0 || MaskByteOffset == 2) { | |||
4403 | // (x >> 8) & 0xff | |||
4404 | // (x >> 8) & 0xff0000 | |||
4405 | if (Opc0 != ISD::SRL) | |||
4406 | return false; | |||
4407 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); | |||
4408 | if (!C || C->getZExtValue() != 8) | |||
4409 | return false; | |||
4410 | } else { | |||
4411 | // (x << 8) & 0xff00 | |||
4412 | // (x << 8) & 0xff000000 | |||
4413 | if (Opc0 != ISD::SHL) | |||
4414 | return false; | |||
4415 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); | |||
4416 | if (!C || C->getZExtValue() != 8) | |||
4417 | return false; | |||
4418 | } | |||
4419 | } else if (Opc == ISD::SHL) { | |||
4420 | // (x & 0xff) << 8 | |||
4421 | // (x & 0xff0000) << 8 | |||
4422 | if (MaskByteOffset != 0 && MaskByteOffset != 2) | |||
4423 | return false; | |||
4424 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1)); | |||
4425 | if (!C || C->getZExtValue() != 8) | |||
4426 | return false; | |||
4427 | } else { // Opc == ISD::SRL | |||
4428 | // (x & 0xff00) >> 8 | |||
4429 | // (x & 0xff000000) >> 8 | |||
4430 | if (MaskByteOffset != 1 && MaskByteOffset != 3) | |||
4431 | return false; | |||
4432 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1)); | |||
4433 | if (!C || C->getZExtValue() != 8) | |||
4434 | return false; | |||
4435 | } | |||
4436 | ||||
4437 | if (Parts[MaskByteOffset]) | |||
4438 | return false; | |||
4439 | ||||
4440 | Parts[MaskByteOffset] = N0.getOperand(0).getNode(); | |||
4441 | return true; | |||
4442 | } | |||
4443 | ||||
4444 | /// Match a 32-bit packed halfword bswap. That is | |||
4445 | /// ((x & 0x000000ff) << 8) | | |||
4446 | /// ((x & 0x0000ff00) >> 8) | | |||
4447 | /// ((x & 0x00ff0000) << 8) | | |||
4448 | /// ((x & 0xff000000) >> 8) | |||
4449 | /// => (rotl (bswap x), 16) | |||
4450 | SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) { | |||
4451 | if (!LegalOperations) | |||
4452 | return SDValue(); | |||
4453 | ||||
4454 | EVT VT = N->getValueType(0); | |||
4455 | if (VT != MVT::i32) | |||
4456 | return SDValue(); | |||
4457 | if (!TLI.isOperationLegalOrCustom(ISD::BSWAP, VT)) | |||
4458 | return SDValue(); | |||
4459 | ||||
4460 | // Look for either | |||
4461 | // (or (or (and), (and)), (or (and), (and))) | |||
4462 | // (or (or (or (and), (and)), (and)), (and)) | |||
4463 | if (N0.getOpcode() != ISD::OR) | |||
4464 | return SDValue(); | |||
4465 | SDValue N00 = N0.getOperand(0); | |||
4466 | SDValue N01 = N0.getOperand(1); | |||
4467 | SDNode *Parts[4] = {}; | |||
4468 | ||||
4469 | if (N1.getOpcode() == ISD::OR && | |||
4470 | N00.getNumOperands() == 2 && N01.getNumOperands() == 2) { | |||
4471 | // (or (or (and), (and)), (or (and), (and))) | |||
4472 | if (!isBSwapHWordElement(N00, Parts)) | |||
4473 | return SDValue(); | |||
4474 | ||||
4475 | if (!isBSwapHWordElement(N01, Parts)) | |||
4476 | return SDValue(); | |||
4477 | SDValue N10 = N1.getOperand(0); | |||
4478 | if (!isBSwapHWordElement(N10, Parts)) | |||
4479 | return SDValue(); | |||
4480 | SDValue N11 = N1.getOperand(1); | |||
4481 | if (!isBSwapHWordElement(N11, Parts)) | |||
4482 | return SDValue(); | |||
4483 | } else { | |||
4484 | // (or (or (or (and), (and)), (and)), (and)) | |||
4485 | if (!isBSwapHWordElement(N1, Parts)) | |||
4486 | return SDValue(); | |||
4487 | if (!isBSwapHWordElement(N01, Parts)) | |||
4488 | return SDValue(); | |||
4489 | if (N00.getOpcode() != ISD::OR) | |||
4490 | return SDValue(); | |||
4491 | SDValue N000 = N00.getOperand(0); | |||
4492 | if (!isBSwapHWordElement(N000, Parts)) | |||
4493 | return SDValue(); | |||
4494 | SDValue N001 = N00.getOperand(1); | |||
4495 | if (!isBSwapHWordElement(N001, Parts)) | |||
4496 | return SDValue(); | |||
4497 | } | |||
4498 | ||||
4499 | // Make sure the parts are all coming from the same node. | |||
4500 | if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3]) | |||
4501 | return SDValue(); | |||
4502 | ||||
4503 | SDLoc DL(N); | |||
4504 | SDValue BSwap = DAG.getNode(ISD::BSWAP, DL, VT, | |||
4505 | SDValue(Parts[0], 0)); | |||
4506 | ||||
4507 | // Result of the bswap should be rotated by 16. If it's not legal, then | |||
4508 | // do (x << 16) | (x >> 16). | |||
4509 | SDValue ShAmt = DAG.getConstant(16, DL, getShiftAmountTy(VT)); | |||
4510 | if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT)) | |||
4511 | return DAG.getNode(ISD::ROTL, DL, VT, BSwap, ShAmt); | |||
4512 | if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT)) | |||
4513 | return DAG.getNode(ISD::ROTR, DL, VT, BSwap, ShAmt); | |||
4514 | return DAG.getNode(ISD::OR, DL, VT, | |||
4515 | DAG.getNode(ISD::SHL, DL, VT, BSwap, ShAmt), | |||
4516 | DAG.getNode(ISD::SRL, DL, VT, BSwap, ShAmt)); | |||
4517 | } | |||
4518 | ||||
4519 | /// This contains all DAGCombine rules which reduce two values combined by | |||
4520 | /// an Or operation to a single value \see visitANDLike(). | |||
4521 | SDValue DAGCombiner::visitORLike(SDValue N0, SDValue N1, SDNode *N) { | |||
4522 | EVT VT = N1.getValueType(); | |||
4523 | SDLoc DL(N); | |||
4524 | ||||
4525 | // fold (or x, undef) -> -1 | |||
4526 | if (!LegalOperations && (N0.isUndef() || N1.isUndef())) | |||
4527 | return DAG.getAllOnesConstant(DL, VT); | |||
4528 | ||||
4529 | if (SDValue V = foldLogicOfSetCCs(false, N0, N1, DL)) | |||
4530 | return V; | |||
4531 | ||||
4532 | // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible. | |||
4533 | if (N0.getOpcode() == ISD::AND && N1.getOpcode() == ISD::AND && | |||
4534 | // Don't increase # computations. | |||
4535 | (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) { | |||
4536 | // We can only do this xform if we know that bits from X that are set in C2 | |||
4537 | // but not in C1 are already zero. Likewise for Y. | |||
4538 | if (const ConstantSDNode *N0O1C = | |||
4539 | getAsNonOpaqueConstant(N0.getOperand(1))) { | |||
4540 | if (const ConstantSDNode *N1O1C = | |||
4541 | getAsNonOpaqueConstant(N1.getOperand(1))) { | |||
4542 | // We can only do this xform if we know that bits from X that are set in | |||
4543 | // C2 but not in C1 are already zero. Likewise for Y. | |||
4544 | const APInt &LHSMask = N0O1C->getAPIntValue(); | |||
4545 | const APInt &RHSMask = N1O1C->getAPIntValue(); | |||
4546 | ||||
4547 | if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) && | |||
4548 | DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) { | |||
4549 | SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT, | |||
4550 | N0.getOperand(0), N1.getOperand(0)); | |||
4551 | return DAG.getNode(ISD::AND, DL, VT, X, | |||
4552 | DAG.getConstant(LHSMask | RHSMask, DL, VT)); | |||
4553 | } | |||
4554 | } | |||
4555 | } | |||
4556 | } | |||
4557 | ||||
4558 | // (or (and X, M), (and X, N)) -> (and X, (or M, N)) | |||
4559 | if (N0.getOpcode() == ISD::AND && | |||
4560 | N1.getOpcode() == ISD::AND && | |||
4561 | N0.getOperand(0) == N1.getOperand(0) && | |||
4562 | // Don't increase # computations. | |||
4563 | (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) { | |||
4564 | SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT, | |||
4565 | N0.getOperand(1), N1.getOperand(1)); | |||
4566 | return DAG.getNode(ISD::AND, DL, VT, N0.getOperand(0), X); | |||
4567 | } | |||
4568 | ||||
4569 | return SDValue(); | |||
4570 | } | |||
4571 | ||||
4572 | SDValue DAGCombiner::visitOR(SDNode *N) { | |||
4573 | SDValue N0 = N->getOperand(0); | |||
4574 | SDValue N1 = N->getOperand(1); | |||
4575 | EVT VT = N1.getValueType(); | |||
4576 | ||||
4577 | // x | x --> x | |||
4578 | if (N0 == N1) | |||
4579 | return N0; | |||
4580 | ||||
4581 | // fold vector ops | |||
4582 | if (VT.isVector()) { | |||
4583 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
4584 | return FoldedVOp; | |||
4585 | ||||
4586 | // fold (or x, 0) -> x, vector edition | |||
4587 | if (ISD::isBuildVectorAllZeros(N0.getNode())) | |||
4588 | return N1; | |||
4589 | if (ISD::isBuildVectorAllZeros(N1.getNode())) | |||
4590 | return N0; | |||
4591 | ||||
4592 | // fold (or x, -1) -> -1, vector edition | |||
4593 | if (ISD::isBuildVectorAllOnes(N0.getNode())) | |||
4594 | // do not return N0, because undef node may exist in N0 | |||
4595 | return DAG.getAllOnesConstant(SDLoc(N), N0.getValueType()); | |||
4596 | if (ISD::isBuildVectorAllOnes(N1.getNode())) | |||
4597 | // do not return N1, because undef node may exist in N1 | |||
4598 | return DAG.getAllOnesConstant(SDLoc(N), N1.getValueType()); | |||
4599 | ||||
4600 | // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask) | |||
4601 | // Do this only if the resulting shuffle is legal. | |||
4602 | if (isa<ShuffleVectorSDNode>(N0) && | |||
4603 | isa<ShuffleVectorSDNode>(N1) && | |||
4604 | // Avoid folding a node with illegal type. | |||
4605 | TLI.isTypeLegal(VT)) { | |||
4606 | bool ZeroN00 = ISD::isBuildVectorAllZeros(N0.getOperand(0).getNode()); | |||
4607 | bool ZeroN01 = ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode()); | |||
4608 | bool ZeroN10 = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode()); | |||
4609 | bool ZeroN11 = ISD::isBuildVectorAllZeros(N1.getOperand(1).getNode()); | |||
4610 | // Ensure both shuffles have a zero input. | |||
4611 | if ((ZeroN00 != ZeroN01) && (ZeroN10 != ZeroN11)) { | |||
4612 | assert((!ZeroN00 || !ZeroN01) && "Both inputs zero!")(static_cast <bool> ((!ZeroN00 || !ZeroN01) && "Both inputs zero!" ) ? void (0) : __assert_fail ("(!ZeroN00 || !ZeroN01) && \"Both inputs zero!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4612, __extension__ __PRETTY_FUNCTION__)); | |||
4613 | assert((!ZeroN10 || !ZeroN11) && "Both inputs zero!")(static_cast <bool> ((!ZeroN10 || !ZeroN11) && "Both inputs zero!" ) ? void (0) : __assert_fail ("(!ZeroN10 || !ZeroN11) && \"Both inputs zero!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4613, __extension__ __PRETTY_FUNCTION__)); | |||
4614 | const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0); | |||
4615 | const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1); | |||
4616 | bool CanFold = true; | |||
4617 | int NumElts = VT.getVectorNumElements(); | |||
4618 | SmallVector<int, 4> Mask(NumElts); | |||
4619 | ||||
4620 | for (int i = 0; i != NumElts; ++i) { | |||
4621 | int M0 = SV0->getMaskElt(i); | |||
4622 | int M1 = SV1->getMaskElt(i); | |||
4623 | ||||
4624 | // Determine if either index is pointing to a zero vector. | |||
4625 | bool M0Zero = M0 < 0 || (ZeroN00 == (M0 < NumElts)); | |||
4626 | bool M1Zero = M1 < 0 || (ZeroN10 == (M1 < NumElts)); | |||
4627 | ||||
4628 | // If one element is zero and the otherside is undef, keep undef. | |||
4629 | // This also handles the case that both are undef. | |||
4630 | if ((M0Zero && M1 < 0) || (M1Zero && M0 < 0)) { | |||
4631 | Mask[i] = -1; | |||
4632 | continue; | |||
4633 | } | |||
4634 | ||||
4635 | // Make sure only one of the elements is zero. | |||
4636 | if (M0Zero == M1Zero) { | |||
4637 | CanFold = false; | |||
4638 | break; | |||
4639 | } | |||
4640 | ||||
4641 | assert((M0 >= 0 || M1 >= 0) && "Undef index!")(static_cast <bool> ((M0 >= 0 || M1 >= 0) && "Undef index!") ? void (0) : __assert_fail ("(M0 >= 0 || M1 >= 0) && \"Undef index!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4641, __extension__ __PRETTY_FUNCTION__)); | |||
4642 | ||||
4643 | // We have a zero and non-zero element. If the non-zero came from | |||
4644 | // SV0 make the index a LHS index. If it came from SV1, make it | |||
4645 | // a RHS index. We need to mod by NumElts because we don't care | |||
4646 | // which operand it came from in the original shuffles. | |||
4647 | Mask[i] = M1Zero ? M0 % NumElts : (M1 % NumElts) + NumElts; | |||
4648 | } | |||
4649 | ||||
4650 | if (CanFold) { | |||
4651 | SDValue NewLHS = ZeroN00 ? N0.getOperand(1) : N0.getOperand(0); | |||
4652 | SDValue NewRHS = ZeroN10 ? N1.getOperand(1) : N1.getOperand(0); | |||
4653 | ||||
4654 | bool LegalMask = TLI.isShuffleMaskLegal(Mask, VT); | |||
4655 | if (!LegalMask) { | |||
4656 | std::swap(NewLHS, NewRHS); | |||
4657 | ShuffleVectorSDNode::commuteMask(Mask); | |||
4658 | LegalMask = TLI.isShuffleMaskLegal(Mask, VT); | |||
4659 | } | |||
4660 | ||||
4661 | if (LegalMask) | |||
4662 | return DAG.getVectorShuffle(VT, SDLoc(N), NewLHS, NewRHS, Mask); | |||
4663 | } | |||
4664 | } | |||
4665 | } | |||
4666 | } | |||
4667 | ||||
4668 | // fold (or c1, c2) -> c1|c2 | |||
4669 | ConstantSDNode *N0C = getAsNonOpaqueConstant(N0); | |||
4670 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); | |||
4671 | if (N0C && N1C && !N1C->isOpaque()) | |||
4672 | return DAG.FoldConstantArithmetic(ISD::OR, SDLoc(N), VT, N0C, N1C); | |||
4673 | // canonicalize constant to RHS | |||
4674 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
4675 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) | |||
4676 | return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0); | |||
4677 | // fold (or x, 0) -> x | |||
4678 | if (isNullConstant(N1)) | |||
4679 | return N0; | |||
4680 | // fold (or x, -1) -> -1 | |||
4681 | if (isAllOnesConstant(N1)) | |||
4682 | return N1; | |||
4683 | ||||
4684 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
4685 | return NewSel; | |||
4686 | ||||
4687 | // fold (or x, c) -> c iff (x & ~c) == 0 | |||
4688 | if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue())) | |||
4689 | return N1; | |||
4690 | ||||
4691 | if (SDValue Combined = visitORLike(N0, N1, N)) | |||
4692 | return Combined; | |||
4693 | ||||
4694 | // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16) | |||
4695 | if (SDValue BSwap = MatchBSwapHWord(N, N0, N1)) | |||
4696 | return BSwap; | |||
4697 | if (SDValue BSwap = MatchBSwapHWordLow(N, N0, N1)) | |||
4698 | return BSwap; | |||
4699 | ||||
4700 | // reassociate or | |||
4701 | if (SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1)) | |||
4702 | return ROR; | |||
4703 | ||||
4704 | // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2) | |||
4705 | // iff (c1 & c2) != 0. | |||
4706 | auto MatchIntersect = [](ConstantSDNode *LHS, ConstantSDNode *RHS) { | |||
4707 | return LHS->getAPIntValue().intersects(RHS->getAPIntValue()); | |||
4708 | }; | |||
4709 | if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() && | |||
4710 | ISD::matchBinaryPredicate(N0.getOperand(1), N1, MatchIntersect)) { | |||
4711 | if (SDValue COR = DAG.FoldConstantArithmetic( | |||
4712 | ISD::OR, SDLoc(N1), VT, N1.getNode(), N0.getOperand(1).getNode())) { | |||
4713 | SDValue IOR = DAG.getNode(ISD::OR, SDLoc(N0), VT, N0.getOperand(0), N1); | |||
4714 | AddToWorklist(IOR.getNode()); | |||
4715 | return DAG.getNode(ISD::AND, SDLoc(N), VT, COR, IOR); | |||
4716 | } | |||
4717 | } | |||
4718 | ||||
4719 | // Simplify: (or (op x...), (op y...)) -> (op (or x, y)) | |||
4720 | if (N0.getOpcode() == N1.getOpcode()) | |||
4721 | if (SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N)) | |||
4722 | return Tmp; | |||
4723 | ||||
4724 | // See if this is some rotate idiom. | |||
4725 | if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N))) | |||
4726 | return SDValue(Rot, 0); | |||
4727 | ||||
4728 | if (SDValue Load = MatchLoadCombine(N)) | |||
4729 | return Load; | |||
4730 | ||||
4731 | // Simplify the operands using demanded-bits information. | |||
4732 | if (SimplifyDemandedBits(SDValue(N, 0))) | |||
4733 | return SDValue(N, 0); | |||
4734 | ||||
4735 | return SDValue(); | |||
4736 | } | |||
4737 | ||||
4738 | /// Match "(X shl/srl V1) & V2" where V2 may not be present. | |||
4739 | bool DAGCombiner::MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) { | |||
4740 | if (Op.getOpcode() == ISD::AND) { | |||
4741 | if (DAG.isConstantIntBuildVectorOrConstantInt(Op.getOperand(1))) { | |||
4742 | Mask = Op.getOperand(1); | |||
4743 | Op = Op.getOperand(0); | |||
4744 | } else { | |||
4745 | return false; | |||
4746 | } | |||
4747 | } | |||
4748 | ||||
4749 | if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) { | |||
4750 | Shift = Op; | |||
4751 | return true; | |||
4752 | } | |||
4753 | ||||
4754 | return false; | |||
4755 | } | |||
4756 | ||||
4757 | // Return true if we can prove that, whenever Neg and Pos are both in the | |||
4758 | // range [0, EltSize), Neg == (Pos == 0 ? 0 : EltSize - Pos). This means that | |||
4759 | // for two opposing shifts shift1 and shift2 and a value X with OpBits bits: | |||
4760 | // | |||
4761 | // (or (shift1 X, Neg), (shift2 X, Pos)) | |||
4762 | // | |||
4763 | // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate | |||
4764 | // in direction shift1 by Neg. The range [0, EltSize) means that we only need | |||
4765 | // to consider shift amounts with defined behavior. | |||
4766 | static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned EltSize) { | |||
4767 | // If EltSize is a power of 2 then: | |||
4768 | // | |||
4769 | // (a) (Pos == 0 ? 0 : EltSize - Pos) == (EltSize - Pos) & (EltSize - 1) | |||
4770 | // (b) Neg == Neg & (EltSize - 1) whenever Neg is in [0, EltSize). | |||
4771 | // | |||
4772 | // So if EltSize is a power of 2 and Neg is (and Neg', EltSize-1), we check | |||
4773 | // for the stronger condition: | |||
4774 | // | |||
4775 | // Neg & (EltSize - 1) == (EltSize - Pos) & (EltSize - 1) [A] | |||
4776 | // | |||
4777 | // for all Neg and Pos. Since Neg & (EltSize - 1) == Neg' & (EltSize - 1) | |||
4778 | // we can just replace Neg with Neg' for the rest of the function. | |||
4779 | // | |||
4780 | // In other cases we check for the even stronger condition: | |||
4781 | // | |||
4782 | // Neg == EltSize - Pos [B] | |||
4783 | // | |||
4784 | // for all Neg and Pos. Note that the (or ...) then invokes undefined | |||
4785 | // behavior if Pos == 0 (and consequently Neg == EltSize). | |||
4786 | // | |||
4787 | // We could actually use [A] whenever EltSize is a power of 2, but the | |||
4788 | // only extra cases that it would match are those uninteresting ones | |||
4789 | // where Neg and Pos are never in range at the same time. E.g. for | |||
4790 | // EltSize == 32, using [A] would allow a Neg of the form (sub 64, Pos) | |||
4791 | // as well as (sub 32, Pos), but: | |||
4792 | // | |||
4793 | // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos)) | |||
4794 | // | |||
4795 | // always invokes undefined behavior for 32-bit X. | |||
4796 | // | |||
4797 | // Below, Mask == EltSize - 1 when using [A] and is all-ones otherwise. | |||
4798 | unsigned MaskLoBits = 0; | |||
4799 | if (Neg.getOpcode() == ISD::AND && isPowerOf2_64(EltSize)) { | |||
4800 | if (ConstantSDNode *NegC = isConstOrConstSplat(Neg.getOperand(1))) { | |||
4801 | if (NegC->getAPIntValue() == EltSize - 1) { | |||
4802 | Neg = Neg.getOperand(0); | |||
4803 | MaskLoBits = Log2_64(EltSize); | |||
4804 | } | |||
4805 | } | |||
4806 | } | |||
4807 | ||||
4808 | // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1. | |||
4809 | if (Neg.getOpcode() != ISD::SUB) | |||
4810 | return false; | |||
4811 | ConstantSDNode *NegC = isConstOrConstSplat(Neg.getOperand(0)); | |||
4812 | if (!NegC) | |||
4813 | return false; | |||
4814 | SDValue NegOp1 = Neg.getOperand(1); | |||
4815 | ||||
4816 | // On the RHS of [A], if Pos is Pos' & (EltSize - 1), just replace Pos with | |||
4817 | // Pos'. The truncation is redundant for the purpose of the equality. | |||
4818 | if (MaskLoBits && Pos.getOpcode() == ISD::AND) | |||
4819 | if (ConstantSDNode *PosC = isConstOrConstSplat(Pos.getOperand(1))) | |||
4820 | if (PosC->getAPIntValue() == EltSize - 1) | |||
4821 | Pos = Pos.getOperand(0); | |||
4822 | ||||
4823 | // The condition we need is now: | |||
4824 | // | |||
4825 | // (NegC - NegOp1) & Mask == (EltSize - Pos) & Mask | |||
4826 | // | |||
4827 | // If NegOp1 == Pos then we need: | |||
4828 | // | |||
4829 | // EltSize & Mask == NegC & Mask | |||
4830 | // | |||
4831 | // (because "x & Mask" is a truncation and distributes through subtraction). | |||
4832 | APInt Width; | |||
4833 | if (Pos == NegOp1) | |||
4834 | Width = NegC->getAPIntValue(); | |||
4835 | ||||
4836 | // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC. | |||
4837 | // Then the condition we want to prove becomes: | |||
4838 | // | |||
4839 | // (NegC - NegOp1) & Mask == (EltSize - (NegOp1 + PosC)) & Mask | |||
4840 | // | |||
4841 | // which, again because "x & Mask" is a truncation, becomes: | |||
4842 | // | |||
4843 | // NegC & Mask == (EltSize - PosC) & Mask | |||
4844 | // EltSize & Mask == (NegC + PosC) & Mask | |||
4845 | else if (Pos.getOpcode() == ISD::ADD && Pos.getOperand(0) == NegOp1) { | |||
4846 | if (ConstantSDNode *PosC = isConstOrConstSplat(Pos.getOperand(1))) | |||
4847 | Width = PosC->getAPIntValue() + NegC->getAPIntValue(); | |||
4848 | else | |||
4849 | return false; | |||
4850 | } else | |||
4851 | return false; | |||
4852 | ||||
4853 | // Now we just need to check that EltSize & Mask == Width & Mask. | |||
4854 | if (MaskLoBits) | |||
4855 | // EltSize & Mask is 0 since Mask is EltSize - 1. | |||
4856 | return Width.getLoBits(MaskLoBits) == 0; | |||
4857 | return Width == EltSize; | |||
4858 | } | |||
4859 | ||||
4860 | // A subroutine of MatchRotate used once we have found an OR of two opposite | |||
4861 | // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces | |||
4862 | // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the | |||
4863 | // former being preferred if supported. InnerPos and InnerNeg are Pos and | |||
4864 | // Neg with outer conversions stripped away. | |||
4865 | SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos, | |||
4866 | SDValue Neg, SDValue InnerPos, | |||
4867 | SDValue InnerNeg, unsigned PosOpcode, | |||
4868 | unsigned NegOpcode, const SDLoc &DL) { | |||
4869 | // fold (or (shl x, (*ext y)), | |||
4870 | // (srl x, (*ext (sub 32, y)))) -> | |||
4871 | // (rotl x, y) or (rotr x, (sub 32, y)) | |||
4872 | // | |||
4873 | // fold (or (shl x, (*ext (sub 32, y))), | |||
4874 | // (srl x, (*ext y))) -> | |||
4875 | // (rotr x, y) or (rotl x, (sub 32, y)) | |||
4876 | EVT VT = Shifted.getValueType(); | |||
4877 | if (matchRotateSub(InnerPos, InnerNeg, VT.getScalarSizeInBits())) { | |||
4878 | bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT); | |||
4879 | return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted, | |||
4880 | HasPos ? Pos : Neg).getNode(); | |||
4881 | } | |||
4882 | ||||
4883 | return nullptr; | |||
4884 | } | |||
4885 | ||||
4886 | // MatchRotate - Handle an 'or' of two operands. If this is one of the many | |||
4887 | // idioms for rotate, and if the target supports rotation instructions, generate | |||
4888 | // a rot[lr]. | |||
4889 | SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, const SDLoc &DL) { | |||
4890 | // Must be a legal type. Expanded 'n promoted things won't work with rotates. | |||
4891 | EVT VT = LHS.getValueType(); | |||
4892 | if (!TLI.isTypeLegal(VT)) return nullptr; | |||
4893 | ||||
4894 | // The target must have at least one rotate flavor. | |||
4895 | bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT); | |||
4896 | bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT); | |||
4897 | if (!HasROTL && !HasROTR) return nullptr; | |||
4898 | ||||
4899 | // Check for truncated rotate. | |||
4900 | if (LHS.getOpcode() == ISD::TRUNCATE && RHS.getOpcode() == ISD::TRUNCATE && | |||
4901 | LHS.getOperand(0).getValueType() == RHS.getOperand(0).getValueType()) { | |||
4902 | assert(LHS.getValueType() == RHS.getValueType())(static_cast <bool> (LHS.getValueType() == RHS.getValueType ()) ? void (0) : __assert_fail ("LHS.getValueType() == RHS.getValueType()" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 4902, __extension__ __PRETTY_FUNCTION__)); | |||
4903 | if (SDNode *Rot = MatchRotate(LHS.getOperand(0), RHS.getOperand(0), DL)) { | |||
4904 | return DAG.getNode(ISD::TRUNCATE, SDLoc(LHS), LHS.getValueType(), | |||
4905 | SDValue(Rot, 0)).getNode(); | |||
4906 | } | |||
4907 | } | |||
4908 | ||||
4909 | // Match "(X shl/srl V1) & V2" where V2 may not be present. | |||
4910 | SDValue LHSShift; // The shift. | |||
4911 | SDValue LHSMask; // AND value if any. | |||
4912 | if (!MatchRotateHalf(LHS, LHSShift, LHSMask)) | |||
4913 | return nullptr; // Not part of a rotate. | |||
4914 | ||||
4915 | SDValue RHSShift; // The shift. | |||
4916 | SDValue RHSMask; // AND value if any. | |||
4917 | if (!MatchRotateHalf(RHS, RHSShift, RHSMask)) | |||
4918 | return nullptr; // Not part of a rotate. | |||
4919 | ||||
4920 | if (LHSShift.getOperand(0) != RHSShift.getOperand(0)) | |||
4921 | return nullptr; // Not shifting the same value. | |||
4922 | ||||
4923 | if (LHSShift.getOpcode() == RHSShift.getOpcode()) | |||
4924 | return nullptr; // Shifts must disagree. | |||
4925 | ||||
4926 | // Canonicalize shl to left side in a shl/srl pair. | |||
4927 | if (RHSShift.getOpcode() == ISD::SHL) { | |||
4928 | std::swap(LHS, RHS); | |||
4929 | std::swap(LHSShift, RHSShift); | |||
4930 | std::swap(LHSMask, RHSMask); | |||
4931 | } | |||
4932 | ||||
4933 | unsigned EltSizeInBits = VT.getScalarSizeInBits(); | |||
4934 | SDValue LHSShiftArg = LHSShift.getOperand(0); | |||
4935 | SDValue LHSShiftAmt = LHSShift.getOperand(1); | |||
4936 | SDValue RHSShiftArg = RHSShift.getOperand(0); | |||
4937 | SDValue RHSShiftAmt = RHSShift.getOperand(1); | |||
4938 | ||||
4939 | // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1) | |||
4940 | // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2) | |||
4941 | auto MatchRotateSum = [EltSizeInBits](ConstantSDNode *LHS, | |||
4942 | ConstantSDNode *RHS) { | |||
4943 | return (LHS->getAPIntValue() + RHS->getAPIntValue()) == EltSizeInBits; | |||
4944 | }; | |||
4945 | if (ISD::matchBinaryPredicate(LHSShiftAmt, RHSShiftAmt, MatchRotateSum)) { | |||
4946 | SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT, | |||
4947 | LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt); | |||
4948 | ||||
4949 | // If there is an AND of either shifted operand, apply it to the result. | |||
4950 | if (LHSMask.getNode() || RHSMask.getNode()) { | |||
4951 | SDValue AllOnes = DAG.getAllOnesConstant(DL, VT); | |||
4952 | SDValue Mask = AllOnes; | |||
4953 | ||||
4954 | if (LHSMask.getNode()) { | |||
4955 | SDValue RHSBits = DAG.getNode(ISD::SRL, DL, VT, AllOnes, RHSShiftAmt); | |||
4956 | Mask = DAG.getNode(ISD::AND, DL, VT, Mask, | |||
4957 | DAG.getNode(ISD::OR, DL, VT, LHSMask, RHSBits)); | |||
4958 | } | |||
4959 | if (RHSMask.getNode()) { | |||
4960 | SDValue LHSBits = DAG.getNode(ISD::SHL, DL, VT, AllOnes, LHSShiftAmt); | |||
4961 | Mask = DAG.getNode(ISD::AND, DL, VT, Mask, | |||
4962 | DAG.getNode(ISD::OR, DL, VT, RHSMask, LHSBits)); | |||
4963 | } | |||
4964 | ||||
4965 | Rot = DAG.getNode(ISD::AND, DL, VT, Rot, Mask); | |||
4966 | } | |||
4967 | ||||
4968 | return Rot.getNode(); | |||
4969 | } | |||
4970 | ||||
4971 | // If there is a mask here, and we have a variable shift, we can't be sure | |||
4972 | // that we're masking out the right stuff. | |||
4973 | if (LHSMask.getNode() || RHSMask.getNode()) | |||
4974 | return nullptr; | |||
4975 | ||||
4976 | // If the shift amount is sign/zext/any-extended just peel it off. | |||
4977 | SDValue LExtOp0 = LHSShiftAmt; | |||
4978 | SDValue RExtOp0 = RHSShiftAmt; | |||
4979 | if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND || | |||
4980 | LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND || | |||
4981 | LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND || | |||
4982 | LHSShiftAmt.getOpcode() == ISD::TRUNCATE) && | |||
4983 | (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND || | |||
4984 | RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND || | |||
4985 | RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND || | |||
4986 | RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) { | |||
4987 | LExtOp0 = LHSShiftAmt.getOperand(0); | |||
4988 | RExtOp0 = RHSShiftAmt.getOperand(0); | |||
4989 | } | |||
4990 | ||||
4991 | SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt, | |||
4992 | LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL); | |||
4993 | if (TryL) | |||
4994 | return TryL; | |||
4995 | ||||
4996 | SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt, | |||
4997 | RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL); | |||
4998 | if (TryR) | |||
4999 | return TryR; | |||
5000 | ||||
5001 | return nullptr; | |||
5002 | } | |||
5003 | ||||
5004 | namespace { | |||
5005 | ||||
5006 | /// Represents known origin of an individual byte in load combine pattern. The | |||
5007 | /// value of the byte is either constant zero or comes from memory. | |||
5008 | struct ByteProvider { | |||
5009 | // For constant zero providers Load is set to nullptr. For memory providers | |||
5010 | // Load represents the node which loads the byte from memory. | |||
5011 | // ByteOffset is the offset of the byte in the value produced by the load. | |||
5012 | LoadSDNode *Load = nullptr; | |||
5013 | unsigned ByteOffset = 0; | |||
5014 | ||||
5015 | ByteProvider() = default; | |||
5016 | ||||
5017 | static ByteProvider getMemory(LoadSDNode *Load, unsigned ByteOffset) { | |||
5018 | return ByteProvider(Load, ByteOffset); | |||
5019 | } | |||
5020 | ||||
5021 | static ByteProvider getConstantZero() { return ByteProvider(nullptr, 0); } | |||
5022 | ||||
5023 | bool isConstantZero() const { return !Load; } | |||
5024 | bool isMemory() const { return Load; } | |||
5025 | ||||
5026 | bool operator==(const ByteProvider &Other) const { | |||
5027 | return Other.Load == Load && Other.ByteOffset == ByteOffset; | |||
5028 | } | |||
5029 | ||||
5030 | private: | |||
5031 | ByteProvider(LoadSDNode *Load, unsigned ByteOffset) | |||
5032 | : Load(Load), ByteOffset(ByteOffset) {} | |||
5033 | }; | |||
5034 | ||||
5035 | } // end anonymous namespace | |||
5036 | ||||
5037 | /// Recursively traverses the expression calculating the origin of the requested | |||
5038 | /// byte of the given value. Returns None if the provider can't be calculated. | |||
5039 | /// | |||
5040 | /// For all the values except the root of the expression verifies that the value | |||
5041 | /// has exactly one use and if it's not true return None. This way if the origin | |||
5042 | /// of the byte is returned it's guaranteed that the values which contribute to | |||
5043 | /// the byte are not used outside of this expression. | |||
5044 | /// | |||
5045 | /// Because the parts of the expression are not allowed to have more than one | |||
5046 | /// use this function iterates over trees, not DAGs. So it never visits the same | |||
5047 | /// node more than once. | |||
5048 | static const Optional<ByteProvider> | |||
5049 | calculateByteProvider(SDValue Op, unsigned Index, unsigned Depth, | |||
5050 | bool Root = false) { | |||
5051 | // Typical i64 by i8 pattern requires recursion up to 8 calls depth | |||
5052 | if (Depth == 10) | |||
5053 | return None; | |||
5054 | ||||
5055 | if (!Root && !Op.hasOneUse()) | |||
5056 | return None; | |||
5057 | ||||
5058 | assert(Op.getValueType().isScalarInteger() && "can't handle other types")(static_cast <bool> (Op.getValueType().isScalarInteger( ) && "can't handle other types") ? void (0) : __assert_fail ("Op.getValueType().isScalarInteger() && \"can't handle other types\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5058, __extension__ __PRETTY_FUNCTION__)); | |||
5059 | unsigned BitWidth = Op.getValueSizeInBits(); | |||
5060 | if (BitWidth % 8 != 0) | |||
5061 | return None; | |||
5062 | unsigned ByteWidth = BitWidth / 8; | |||
5063 | assert(Index < ByteWidth && "invalid index requested")(static_cast <bool> (Index < ByteWidth && "invalid index requested" ) ? void (0) : __assert_fail ("Index < ByteWidth && \"invalid index requested\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5063, __extension__ __PRETTY_FUNCTION__)); | |||
5064 | (void) ByteWidth; | |||
5065 | ||||
5066 | switch (Op.getOpcode()) { | |||
5067 | case ISD::OR: { | |||
5068 | auto LHS = calculateByteProvider(Op->getOperand(0), Index, Depth + 1); | |||
5069 | if (!LHS) | |||
5070 | return None; | |||
5071 | auto RHS = calculateByteProvider(Op->getOperand(1), Index, Depth + 1); | |||
5072 | if (!RHS) | |||
5073 | return None; | |||
5074 | ||||
5075 | if (LHS->isConstantZero()) | |||
5076 | return RHS; | |||
5077 | if (RHS->isConstantZero()) | |||
5078 | return LHS; | |||
5079 | return None; | |||
5080 | } | |||
5081 | case ISD::SHL: { | |||
5082 | auto ShiftOp = dyn_cast<ConstantSDNode>(Op->getOperand(1)); | |||
5083 | if (!ShiftOp) | |||
5084 | return None; | |||
5085 | ||||
5086 | uint64_t BitShift = ShiftOp->getZExtValue(); | |||
5087 | if (BitShift % 8 != 0) | |||
5088 | return None; | |||
5089 | uint64_t ByteShift = BitShift / 8; | |||
5090 | ||||
5091 | return Index < ByteShift | |||
5092 | ? ByteProvider::getConstantZero() | |||
5093 | : calculateByteProvider(Op->getOperand(0), Index - ByteShift, | |||
5094 | Depth + 1); | |||
5095 | } | |||
5096 | case ISD::ANY_EXTEND: | |||
5097 | case ISD::SIGN_EXTEND: | |||
5098 | case ISD::ZERO_EXTEND: { | |||
5099 | SDValue NarrowOp = Op->getOperand(0); | |||
5100 | unsigned NarrowBitWidth = NarrowOp.getScalarValueSizeInBits(); | |||
5101 | if (NarrowBitWidth % 8 != 0) | |||
5102 | return None; | |||
5103 | uint64_t NarrowByteWidth = NarrowBitWidth / 8; | |||
5104 | ||||
5105 | if (Index >= NarrowByteWidth) | |||
5106 | return Op.getOpcode() == ISD::ZERO_EXTEND | |||
5107 | ? Optional<ByteProvider>(ByteProvider::getConstantZero()) | |||
5108 | : None; | |||
5109 | return calculateByteProvider(NarrowOp, Index, Depth + 1); | |||
5110 | } | |||
5111 | case ISD::BSWAP: | |||
5112 | return calculateByteProvider(Op->getOperand(0), ByteWidth - Index - 1, | |||
5113 | Depth + 1); | |||
5114 | case ISD::LOAD: { | |||
5115 | auto L = cast<LoadSDNode>(Op.getNode()); | |||
5116 | if (L->isVolatile() || L->isIndexed()) | |||
5117 | return None; | |||
5118 | ||||
5119 | unsigned NarrowBitWidth = L->getMemoryVT().getSizeInBits(); | |||
5120 | if (NarrowBitWidth % 8 != 0) | |||
5121 | return None; | |||
5122 | uint64_t NarrowByteWidth = NarrowBitWidth / 8; | |||
5123 | ||||
5124 | if (Index >= NarrowByteWidth) | |||
5125 | return L->getExtensionType() == ISD::ZEXTLOAD | |||
5126 | ? Optional<ByteProvider>(ByteProvider::getConstantZero()) | |||
5127 | : None; | |||
5128 | return ByteProvider::getMemory(L, Index); | |||
5129 | } | |||
5130 | } | |||
5131 | ||||
5132 | return None; | |||
5133 | } | |||
5134 | ||||
5135 | /// Match a pattern where a wide type scalar value is loaded by several narrow | |||
5136 | /// loads and combined by shifts and ors. Fold it into a single load or a load | |||
5137 | /// and a BSWAP if the targets supports it. | |||
5138 | /// | |||
5139 | /// Assuming little endian target: | |||
5140 | /// i8 *a = ... | |||
5141 | /// i32 val = a[0] | (a[1] << 8) | (a[2] << 16) | (a[3] << 24) | |||
5142 | /// => | |||
5143 | /// i32 val = *((i32)a) | |||
5144 | /// | |||
5145 | /// i8 *a = ... | |||
5146 | /// i32 val = (a[0] << 24) | (a[1] << 16) | (a[2] << 8) | a[3] | |||
5147 | /// => | |||
5148 | /// i32 val = BSWAP(*((i32)a)) | |||
5149 | /// | |||
5150 | /// TODO: This rule matches complex patterns with OR node roots and doesn't | |||
5151 | /// interact well with the worklist mechanism. When a part of the pattern is | |||
5152 | /// updated (e.g. one of the loads) its direct users are put into the worklist, | |||
5153 | /// but the root node of the pattern which triggers the load combine is not | |||
5154 | /// necessarily a direct user of the changed node. For example, once the address | |||
5155 | /// of t28 load is reassociated load combine won't be triggered: | |||
5156 | /// t25: i32 = add t4, Constant:i32<2> | |||
5157 | /// t26: i64 = sign_extend t25 | |||
5158 | /// t27: i64 = add t2, t26 | |||
5159 | /// t28: i8,ch = load<LD1[%tmp9]> t0, t27, undef:i64 | |||
5160 | /// t29: i32 = zero_extend t28 | |||
5161 | /// t32: i32 = shl t29, Constant:i8<8> | |||
5162 | /// t33: i32 = or t23, t32 | |||
5163 | /// As a possible fix visitLoad can check if the load can be a part of a load | |||
5164 | /// combine pattern and add corresponding OR roots to the worklist. | |||
5165 | SDValue DAGCombiner::MatchLoadCombine(SDNode *N) { | |||
5166 | assert(N->getOpcode() == ISD::OR &&(static_cast <bool> (N->getOpcode() == ISD::OR && "Can only match load combining against OR nodes") ? void (0) : __assert_fail ("N->getOpcode() == ISD::OR && \"Can only match load combining against OR nodes\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5167, __extension__ __PRETTY_FUNCTION__)) | |||
5167 | "Can only match load combining against OR nodes")(static_cast <bool> (N->getOpcode() == ISD::OR && "Can only match load combining against OR nodes") ? void (0) : __assert_fail ("N->getOpcode() == ISD::OR && \"Can only match load combining against OR nodes\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5167, __extension__ __PRETTY_FUNCTION__)); | |||
5168 | ||||
5169 | // Handles simple types only | |||
5170 | EVT VT = N->getValueType(0); | |||
5171 | if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64) | |||
5172 | return SDValue(); | |||
5173 | unsigned ByteWidth = VT.getSizeInBits() / 8; | |||
5174 | ||||
5175 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
5176 | // Before legalize we can introduce too wide illegal loads which will be later | |||
5177 | // split into legal sized loads. This enables us to combine i64 load by i8 | |||
5178 | // patterns to a couple of i32 loads on 32 bit targets. | |||
5179 | if (LegalOperations && !TLI.isOperationLegal(ISD::LOAD, VT)) | |||
5180 | return SDValue(); | |||
5181 | ||||
5182 | std::function<unsigned(unsigned, unsigned)> LittleEndianByteAt = []( | |||
5183 | unsigned BW, unsigned i) { return i; }; | |||
5184 | std::function<unsigned(unsigned, unsigned)> BigEndianByteAt = []( | |||
5185 | unsigned BW, unsigned i) { return BW - i - 1; }; | |||
5186 | ||||
5187 | bool IsBigEndianTarget = DAG.getDataLayout().isBigEndian(); | |||
5188 | auto MemoryByteOffset = [&] (ByteProvider P) { | |||
5189 | assert(P.isMemory() && "Must be a memory byte provider")(static_cast <bool> (P.isMemory() && "Must be a memory byte provider" ) ? void (0) : __assert_fail ("P.isMemory() && \"Must be a memory byte provider\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5189, __extension__ __PRETTY_FUNCTION__)); | |||
5190 | unsigned LoadBitWidth = P.Load->getMemoryVT().getSizeInBits(); | |||
5191 | assert(LoadBitWidth % 8 == 0 &&(static_cast <bool> (LoadBitWidth % 8 == 0 && "can only analyze providers for individual bytes not bit" ) ? void (0) : __assert_fail ("LoadBitWidth % 8 == 0 && \"can only analyze providers for individual bytes not bit\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5192, __extension__ __PRETTY_FUNCTION__)) | |||
5192 | "can only analyze providers for individual bytes not bit")(static_cast <bool> (LoadBitWidth % 8 == 0 && "can only analyze providers for individual bytes not bit" ) ? void (0) : __assert_fail ("LoadBitWidth % 8 == 0 && \"can only analyze providers for individual bytes not bit\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5192, __extension__ __PRETTY_FUNCTION__)); | |||
5193 | unsigned LoadByteWidth = LoadBitWidth / 8; | |||
5194 | return IsBigEndianTarget | |||
5195 | ? BigEndianByteAt(LoadByteWidth, P.ByteOffset) | |||
5196 | : LittleEndianByteAt(LoadByteWidth, P.ByteOffset); | |||
5197 | }; | |||
5198 | ||||
5199 | Optional<BaseIndexOffset> Base; | |||
5200 | SDValue Chain; | |||
5201 | ||||
5202 | SmallSet<LoadSDNode *, 8> Loads; | |||
5203 | Optional<ByteProvider> FirstByteProvider; | |||
5204 | int64_t FirstOffset = INT64_MAX(9223372036854775807L); | |||
5205 | ||||
5206 | // Check if all the bytes of the OR we are looking at are loaded from the same | |||
5207 | // base address. Collect bytes offsets from Base address in ByteOffsets. | |||
5208 | SmallVector<int64_t, 4> ByteOffsets(ByteWidth); | |||
5209 | for (unsigned i = 0; i < ByteWidth; i++) { | |||
5210 | auto P = calculateByteProvider(SDValue(N, 0), i, 0, /*Root=*/true); | |||
5211 | if (!P || !P->isMemory()) // All the bytes must be loaded from memory | |||
5212 | return SDValue(); | |||
5213 | ||||
5214 | LoadSDNode *L = P->Load; | |||
5215 | assert(L->hasNUsesOfValue(1, 0) && !L->isVolatile() && !L->isIndexed() &&(static_cast <bool> (L->hasNUsesOfValue(1, 0) && !L->isVolatile() && !L->isIndexed() && "Must be enforced by calculateByteProvider") ? void (0) : __assert_fail ("L->hasNUsesOfValue(1, 0) && !L->isVolatile() && !L->isIndexed() && \"Must be enforced by calculateByteProvider\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5216, __extension__ __PRETTY_FUNCTION__)) | |||
5216 | "Must be enforced by calculateByteProvider")(static_cast <bool> (L->hasNUsesOfValue(1, 0) && !L->isVolatile() && !L->isIndexed() && "Must be enforced by calculateByteProvider") ? void (0) : __assert_fail ("L->hasNUsesOfValue(1, 0) && !L->isVolatile() && !L->isIndexed() && \"Must be enforced by calculateByteProvider\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5216, __extension__ __PRETTY_FUNCTION__)); | |||
5217 | assert(L->getOffset().isUndef() && "Unindexed load must have undef offset")(static_cast <bool> (L->getOffset().isUndef() && "Unindexed load must have undef offset") ? void (0) : __assert_fail ("L->getOffset().isUndef() && \"Unindexed load must have undef offset\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5217, __extension__ __PRETTY_FUNCTION__)); | |||
5218 | ||||
5219 | // All loads must share the same chain | |||
5220 | SDValue LChain = L->getChain(); | |||
5221 | if (!Chain) | |||
5222 | Chain = LChain; | |||
5223 | else if (Chain != LChain) | |||
5224 | return SDValue(); | |||
5225 | ||||
5226 | // Loads must share the same base address | |||
5227 | BaseIndexOffset Ptr = BaseIndexOffset::match(L, DAG); | |||
5228 | int64_t ByteOffsetFromBase = 0; | |||
5229 | if (!Base) | |||
5230 | Base = Ptr; | |||
5231 | else if (!Base->equalBaseIndex(Ptr, DAG, ByteOffsetFromBase)) | |||
5232 | return SDValue(); | |||
5233 | ||||
5234 | // Calculate the offset of the current byte from the base address | |||
5235 | ByteOffsetFromBase += MemoryByteOffset(*P); | |||
5236 | ByteOffsets[i] = ByteOffsetFromBase; | |||
5237 | ||||
5238 | // Remember the first byte load | |||
5239 | if (ByteOffsetFromBase < FirstOffset) { | |||
5240 | FirstByteProvider = P; | |||
5241 | FirstOffset = ByteOffsetFromBase; | |||
5242 | } | |||
5243 | ||||
5244 | Loads.insert(L); | |||
5245 | } | |||
5246 | assert(!Loads.empty() && "All the bytes of the value must be loaded from "(static_cast <bool> (!Loads.empty() && "All the bytes of the value must be loaded from " "memory, so there must be at least one load which produces the value" ) ? void (0) : __assert_fail ("!Loads.empty() && \"All the bytes of the value must be loaded from \" \"memory, so there must be at least one load which produces the value\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5247, __extension__ __PRETTY_FUNCTION__)) | |||
5247 | "memory, so there must be at least one load which produces the value")(static_cast <bool> (!Loads.empty() && "All the bytes of the value must be loaded from " "memory, so there must be at least one load which produces the value" ) ? void (0) : __assert_fail ("!Loads.empty() && \"All the bytes of the value must be loaded from \" \"memory, so there must be at least one load which produces the value\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5247, __extension__ __PRETTY_FUNCTION__)); | |||
5248 | assert(Base && "Base address of the accessed memory location must be set")(static_cast <bool> (Base && "Base address of the accessed memory location must be set" ) ? void (0) : __assert_fail ("Base && \"Base address of the accessed memory location must be set\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5248, __extension__ __PRETTY_FUNCTION__)); | |||
5249 | assert(FirstOffset != INT64_MAX && "First byte offset must be set")(static_cast <bool> (FirstOffset != (9223372036854775807L ) && "First byte offset must be set") ? void (0) : __assert_fail ("FirstOffset != INT64_MAX && \"First byte offset must be set\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5249, __extension__ __PRETTY_FUNCTION__)); | |||
5250 | ||||
5251 | // Check if the bytes of the OR we are looking at match with either big or | |||
5252 | // little endian value load | |||
5253 | bool BigEndian = true, LittleEndian = true; | |||
5254 | for (unsigned i = 0; i < ByteWidth; i++) { | |||
5255 | int64_t CurrentByteOffset = ByteOffsets[i] - FirstOffset; | |||
5256 | LittleEndian &= CurrentByteOffset == LittleEndianByteAt(ByteWidth, i); | |||
5257 | BigEndian &= CurrentByteOffset == BigEndianByteAt(ByteWidth, i); | |||
5258 | if (!BigEndian && !LittleEndian) | |||
5259 | return SDValue(); | |||
5260 | } | |||
5261 | assert((BigEndian != LittleEndian) && "should be either or")(static_cast <bool> ((BigEndian != LittleEndian) && "should be either or") ? void (0) : __assert_fail ("(BigEndian != LittleEndian) && \"should be either or\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5261, __extension__ __PRETTY_FUNCTION__)); | |||
5262 | assert(FirstByteProvider && "must be set")(static_cast <bool> (FirstByteProvider && "must be set" ) ? void (0) : __assert_fail ("FirstByteProvider && \"must be set\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5262, __extension__ __PRETTY_FUNCTION__)); | |||
5263 | ||||
5264 | // Ensure that the first byte is loaded from zero offset of the first load. | |||
5265 | // So the combined value can be loaded from the first load address. | |||
5266 | if (MemoryByteOffset(*FirstByteProvider) != 0) | |||
5267 | return SDValue(); | |||
5268 | LoadSDNode *FirstLoad = FirstByteProvider->Load; | |||
5269 | ||||
5270 | // The node we are looking at matches with the pattern, check if we can | |||
5271 | // replace it with a single load and bswap if needed. | |||
5272 | ||||
5273 | // If the load needs byte swap check if the target supports it | |||
5274 | bool NeedsBswap = IsBigEndianTarget != BigEndian; | |||
5275 | ||||
5276 | // Before legalize we can introduce illegal bswaps which will be later | |||
5277 | // converted to an explicit bswap sequence. This way we end up with a single | |||
5278 | // load and byte shuffling instead of several loads and byte shuffling. | |||
5279 | if (NeedsBswap && LegalOperations && !TLI.isOperationLegal(ISD::BSWAP, VT)) | |||
5280 | return SDValue(); | |||
5281 | ||||
5282 | // Check that a load of the wide type is both allowed and fast on the target | |||
5283 | bool Fast = false; | |||
5284 | bool Allowed = TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), | |||
5285 | VT, FirstLoad->getAddressSpace(), | |||
5286 | FirstLoad->getAlignment(), &Fast); | |||
5287 | if (!Allowed || !Fast) | |||
5288 | return SDValue(); | |||
5289 | ||||
5290 | SDValue NewLoad = | |||
5291 | DAG.getLoad(VT, SDLoc(N), Chain, FirstLoad->getBasePtr(), | |||
5292 | FirstLoad->getPointerInfo(), FirstLoad->getAlignment()); | |||
5293 | ||||
5294 | // Transfer chain users from old loads to the new load. | |||
5295 | for (LoadSDNode *L : Loads) | |||
5296 | DAG.ReplaceAllUsesOfValueWith(SDValue(L, 1), SDValue(NewLoad.getNode(), 1)); | |||
5297 | ||||
5298 | return NeedsBswap ? DAG.getNode(ISD::BSWAP, SDLoc(N), VT, NewLoad) : NewLoad; | |||
5299 | } | |||
5300 | ||||
5301 | SDValue DAGCombiner::visitXOR(SDNode *N) { | |||
5302 | SDValue N0 = N->getOperand(0); | |||
5303 | SDValue N1 = N->getOperand(1); | |||
5304 | EVT VT = N0.getValueType(); | |||
5305 | ||||
5306 | // fold vector ops | |||
5307 | if (VT.isVector()) { | |||
5308 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
5309 | return FoldedVOp; | |||
5310 | ||||
5311 | // fold (xor x, 0) -> x, vector edition | |||
5312 | if (ISD::isBuildVectorAllZeros(N0.getNode())) | |||
5313 | return N1; | |||
5314 | if (ISD::isBuildVectorAllZeros(N1.getNode())) | |||
5315 | return N0; | |||
5316 | } | |||
5317 | ||||
5318 | // fold (xor undef, undef) -> 0. This is a common idiom (misuse). | |||
5319 | if (N0.isUndef() && N1.isUndef()) | |||
5320 | return DAG.getConstant(0, SDLoc(N), VT); | |||
5321 | // fold (xor x, undef) -> undef | |||
5322 | if (N0.isUndef()) | |||
5323 | return N0; | |||
5324 | if (N1.isUndef()) | |||
5325 | return N1; | |||
5326 | // fold (xor c1, c2) -> c1^c2 | |||
5327 | ConstantSDNode *N0C = getAsNonOpaqueConstant(N0); | |||
5328 | ConstantSDNode *N1C = getAsNonOpaqueConstant(N1); | |||
5329 | if (N0C && N1C) | |||
5330 | return DAG.FoldConstantArithmetic(ISD::XOR, SDLoc(N), VT, N0C, N1C); | |||
5331 | // canonicalize constant to RHS | |||
5332 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
5333 | !DAG.isConstantIntBuildVectorOrConstantInt(N1)) | |||
5334 | return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0); | |||
5335 | // fold (xor x, 0) -> x | |||
5336 | if (isNullConstant(N1)) | |||
5337 | return N0; | |||
5338 | ||||
5339 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
5340 | return NewSel; | |||
5341 | ||||
5342 | // reassociate xor | |||
5343 | if (SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1)) | |||
5344 | return RXOR; | |||
5345 | ||||
5346 | // fold !(x cc y) -> (x !cc y) | |||
5347 | SDValue LHS, RHS, CC; | |||
5348 | if (TLI.isConstTrueVal(N1.getNode()) && isSetCCEquivalent(N0, LHS, RHS, CC)) { | |||
5349 | bool isInt = LHS.getValueType().isInteger(); | |||
5350 | ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(), | |||
5351 | isInt); | |||
5352 | ||||
5353 | if (!LegalOperations || | |||
5354 | TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) { | |||
5355 | switch (N0.getOpcode()) { | |||
5356 | default: | |||
5357 | llvm_unreachable("Unhandled SetCC Equivalent!")::llvm::llvm_unreachable_internal("Unhandled SetCC Equivalent!" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5357); | |||
5358 | case ISD::SETCC: | |||
5359 | return DAG.getSetCC(SDLoc(N0), VT, LHS, RHS, NotCC); | |||
5360 | case ISD::SELECT_CC: | |||
5361 | return DAG.getSelectCC(SDLoc(N0), LHS, RHS, N0.getOperand(2), | |||
5362 | N0.getOperand(3), NotCC); | |||
5363 | } | |||
5364 | } | |||
5365 | } | |||
5366 | ||||
5367 | // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y))) | |||
5368 | if (isOneConstant(N1) && N0.getOpcode() == ISD::ZERO_EXTEND && | |||
5369 | N0.getNode()->hasOneUse() && | |||
5370 | isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){ | |||
5371 | SDValue V = N0.getOperand(0); | |||
5372 | SDLoc DL(N0); | |||
5373 | V = DAG.getNode(ISD::XOR, DL, V.getValueType(), V, | |||
5374 | DAG.getConstant(1, DL, V.getValueType())); | |||
5375 | AddToWorklist(V.getNode()); | |||
5376 | return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V); | |||
5377 | } | |||
5378 | ||||
5379 | // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc | |||
5380 | if (isOneConstant(N1) && VT == MVT::i1 && N0.hasOneUse() && | |||
5381 | (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) { | |||
5382 | SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1); | |||
5383 | if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) { | |||
5384 | unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND; | |||
5385 | LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS | |||
5386 | RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS | |||
5387 | AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode()); | |||
5388 | return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS); | |||
5389 | } | |||
5390 | } | |||
5391 | // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants | |||
5392 | if (isAllOnesConstant(N1) && N0.hasOneUse() && | |||
5393 | (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) { | |||
5394 | SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1); | |||
5395 | if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) { | |||
5396 | unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND; | |||
5397 | LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS | |||
5398 | RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS | |||
5399 | AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode()); | |||
5400 | return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS); | |||
5401 | } | |||
5402 | } | |||
5403 | // fold (xor (and x, y), y) -> (and (not x), y) | |||
5404 | if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() && | |||
5405 | N0->getOperand(1) == N1) { | |||
5406 | SDValue X = N0->getOperand(0); | |||
5407 | SDValue NotX = DAG.getNOT(SDLoc(X), X, VT); | |||
5408 | AddToWorklist(NotX.getNode()); | |||
5409 | return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1); | |||
5410 | } | |||
5411 | ||||
5412 | // fold Y = sra (X, size(X)-1); xor (add (X, Y), Y) -> (abs X) | |||
5413 | unsigned OpSizeInBits = VT.getScalarSizeInBits(); | |||
5414 | if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 && | |||
5415 | N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0) && | |||
5416 | TLI.isOperationLegalOrCustom(ISD::ABS, VT)) { | |||
5417 | if (ConstantSDNode *C = isConstOrConstSplat(N1.getOperand(1))) | |||
5418 | if (C->getAPIntValue() == (OpSizeInBits - 1)) | |||
5419 | return DAG.getNode(ISD::ABS, SDLoc(N), VT, N0.getOperand(0)); | |||
5420 | } | |||
5421 | ||||
5422 | // fold (xor x, x) -> 0 | |||
5423 | if (N0 == N1) | |||
5424 | return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes); | |||
5425 | ||||
5426 | // fold (xor (shl 1, x), -1) -> (rotl ~1, x) | |||
5427 | // Here is a concrete example of this equivalence: | |||
5428 | // i16 x == 14 | |||
5429 | // i16 shl == 1 << 14 == 16384 == 0b0100000000000000 | |||
5430 | // i16 xor == ~(1 << 14) == 49151 == 0b1011111111111111 | |||
5431 | // | |||
5432 | // => | |||
5433 | // | |||
5434 | // i16 ~1 == 0b1111111111111110 | |||
5435 | // i16 rol(~1, 14) == 0b1011111111111111 | |||
5436 | // | |||
5437 | // Some additional tips to help conceptualize this transform: | |||
5438 | // - Try to see the operation as placing a single zero in a value of all ones. | |||
5439 | // - There exists no value for x which would allow the result to contain zero. | |||
5440 | // - Values of x larger than the bitwidth are undefined and do not require a | |||
5441 | // consistent result. | |||
5442 | // - Pushing the zero left requires shifting one bits in from the right. | |||
5443 | // A rotate left of ~1 is a nice way of achieving the desired result. | |||
5444 | if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT) && N0.getOpcode() == ISD::SHL | |||
5445 | && isAllOnesConstant(N1) && isOneConstant(N0.getOperand(0))) { | |||
5446 | SDLoc DL(N); | |||
5447 | return DAG.getNode(ISD::ROTL, DL, VT, DAG.getConstant(~1, DL, VT), | |||
5448 | N0.getOperand(1)); | |||
5449 | } | |||
5450 | ||||
5451 | // Simplify: xor (op x...), (op y...) -> (op (xor x, y)) | |||
5452 | if (N0.getOpcode() == N1.getOpcode()) | |||
5453 | if (SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N)) | |||
5454 | return Tmp; | |||
5455 | ||||
5456 | // Simplify the expression using non-local knowledge. | |||
5457 | if (SimplifyDemandedBits(SDValue(N, 0))) | |||
5458 | return SDValue(N, 0); | |||
5459 | ||||
5460 | return SDValue(); | |||
5461 | } | |||
5462 | ||||
5463 | /// Handle transforms common to the three shifts, when the shift amount is a | |||
5464 | /// constant. | |||
5465 | SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) { | |||
5466 | SDNode *LHS = N->getOperand(0).getNode(); | |||
5467 | if (!LHS->hasOneUse()) return SDValue(); | |||
5468 | ||||
5469 | // We want to pull some binops through shifts, so that we have (and (shift)) | |||
5470 | // instead of (shift (and)), likewise for add, or, xor, etc. This sort of | |||
5471 | // thing happens with address calculations, so it's important to canonicalize | |||
5472 | // it. | |||
5473 | bool HighBitSet = false; // Can we transform this if the high bit is set? | |||
5474 | ||||
5475 | switch (LHS->getOpcode()) { | |||
5476 | default: return SDValue(); | |||
5477 | case ISD::OR: | |||
5478 | case ISD::XOR: | |||
5479 | HighBitSet = false; // We can only transform sra if the high bit is clear. | |||
5480 | break; | |||
5481 | case ISD::AND: | |||
5482 | HighBitSet = true; // We can only transform sra if the high bit is set. | |||
5483 | break; | |||
5484 | case ISD::ADD: | |||
5485 | if (N->getOpcode() != ISD::SHL) | |||
5486 | return SDValue(); // only shl(add) not sr[al](add). | |||
5487 | HighBitSet = false; // We can only transform sra if the high bit is clear. | |||
5488 | break; | |||
5489 | } | |||
5490 | ||||
5491 | // We require the RHS of the binop to be a constant and not opaque as well. | |||
5492 | ConstantSDNode *BinOpCst = getAsNonOpaqueConstant(LHS->getOperand(1)); | |||
5493 | if (!BinOpCst) return SDValue(); | |||
5494 | ||||
5495 | // FIXME: disable this unless the input to the binop is a shift by a constant | |||
5496 | // or is copy/select.Enable this in other cases when figure out it's exactly profitable. | |||
5497 | SDNode *BinOpLHSVal = LHS->getOperand(0).getNode(); | |||
5498 | bool isShift = BinOpLHSVal->getOpcode() == ISD::SHL || | |||
5499 | BinOpLHSVal->getOpcode() == ISD::SRA || | |||
5500 | BinOpLHSVal->getOpcode() == ISD::SRL; | |||
5501 | bool isCopyOrSelect = BinOpLHSVal->getOpcode() == ISD::CopyFromReg || | |||
5502 | BinOpLHSVal->getOpcode() == ISD::SELECT; | |||
5503 | ||||
5504 | if ((!isShift || !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1))) && | |||
5505 | !isCopyOrSelect) | |||
5506 | return SDValue(); | |||
5507 | ||||
5508 | if (isCopyOrSelect && N->hasOneUse()) | |||
5509 | return SDValue(); | |||
5510 | ||||
5511 | EVT VT = N->getValueType(0); | |||
5512 | ||||
5513 | // If this is a signed shift right, and the high bit is modified by the | |||
5514 | // logical operation, do not perform the transformation. The highBitSet | |||
5515 | // boolean indicates the value of the high bit of the constant which would | |||
5516 | // cause it to be modified for this operation. | |||
5517 | if (N->getOpcode() == ISD::SRA) { | |||
5518 | bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative(); | |||
5519 | if (BinOpRHSSignSet != HighBitSet) | |||
5520 | return SDValue(); | |||
5521 | } | |||
5522 | ||||
5523 | if (!TLI.isDesirableToCommuteWithShift(LHS)) | |||
5524 | return SDValue(); | |||
5525 | ||||
5526 | // Fold the constants, shifting the binop RHS by the shift amount. | |||
5527 | SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)), | |||
5528 | N->getValueType(0), | |||
5529 | LHS->getOperand(1), N->getOperand(1)); | |||
5530 | assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!")(static_cast <bool> (isa<ConstantSDNode>(NewRHS) && "Folding was not successful!") ? void (0) : __assert_fail ("isa<ConstantSDNode>(NewRHS) && \"Folding was not successful!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5530, __extension__ __PRETTY_FUNCTION__)); | |||
5531 | ||||
5532 | // Create the new shift. | |||
5533 | SDValue NewShift = DAG.getNode(N->getOpcode(), | |||
5534 | SDLoc(LHS->getOperand(0)), | |||
5535 | VT, LHS->getOperand(0), N->getOperand(1)); | |||
5536 | ||||
5537 | // Create the new binop. | |||
5538 | return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS); | |||
5539 | } | |||
5540 | ||||
5541 | SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) { | |||
5542 | assert(N->getOpcode() == ISD::TRUNCATE)(static_cast <bool> (N->getOpcode() == ISD::TRUNCATE ) ? void (0) : __assert_fail ("N->getOpcode() == ISD::TRUNCATE" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5542, __extension__ __PRETTY_FUNCTION__)); | |||
5543 | assert(N->getOperand(0).getOpcode() == ISD::AND)(static_cast <bool> (N->getOperand(0).getOpcode() == ISD::AND) ? void (0) : __assert_fail ("N->getOperand(0).getOpcode() == ISD::AND" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 5543, __extension__ __PRETTY_FUNCTION__)); | |||
5544 | ||||
5545 | // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC) | |||
5546 | if (N->hasOneUse() && N->getOperand(0).hasOneUse()) { | |||
5547 | SDValue N01 = N->getOperand(0).getOperand(1); | |||
5548 | if (isConstantOrConstantVector(N01, /* NoOpaques */ true)) { | |||
5549 | SDLoc DL(N); | |||
5550 | EVT TruncVT = N->getValueType(0); | |||
5551 | SDValue N00 = N->getOperand(0).getOperand(0); | |||
5552 | SDValue Trunc00 = DAG.getNode(ISD::TRUNCATE, DL, TruncVT, N00); | |||
5553 | SDValue Trunc01 = DAG.getNode(ISD::TRUNCATE, DL, TruncVT, N01); | |||
5554 | AddToWorklist(Trunc00.getNode()); | |||
5555 | AddToWorklist(Trunc01.getNode()); | |||
5556 | return DAG.getNode(ISD::AND, DL, TruncVT, Trunc00, Trunc01); | |||
5557 | } | |||
5558 | } | |||
5559 | ||||
5560 | return SDValue(); | |||
5561 | } | |||
5562 | ||||
5563 | SDValue DAGCombiner::visitRotate(SDNode *N) { | |||
5564 | SDLoc dl(N); | |||
5565 | SDValue N0 = N->getOperand(0); | |||
5566 | SDValue N1 = N->getOperand(1); | |||
5567 | EVT VT = N->getValueType(0); | |||
5568 | unsigned Bitsize = VT.getScalarSizeInBits(); | |||
5569 | ||||
5570 | // fold (rot x, 0) -> x | |||
5571 | if (isNullConstantOrNullSplatConstant(N1)) | |||
5572 | return N0; | |||
5573 | ||||
5574 | // fold (rot x, c) -> (rot x, c % BitSize) | |||
5575 | if (ConstantSDNode *Cst = isConstOrConstSplat(N1)) { | |||
5576 | if (Cst->getAPIntValue().uge(Bitsize)) { | |||
5577 | uint64_t RotAmt = Cst->getAPIntValue().urem(Bitsize); | |||
5578 | return DAG.getNode(N->getOpcode(), dl, VT, N0, | |||
5579 | DAG.getConstant(RotAmt, dl, N1.getValueType())); | |||
5580 | } | |||
5581 | } | |||
5582 | ||||
5583 | // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))). | |||
5584 | if (N1.getOpcode() == ISD::TRUNCATE && | |||
5585 | N1.getOperand(0).getOpcode() == ISD::AND) { | |||
5586 | if (SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode())) | |||
5587 | return DAG.getNode(N->getOpcode(), dl, VT, N0, NewOp1); | |||
5588 | } | |||
5589 | ||||
5590 | unsigned NextOp = N0.getOpcode(); | |||
5591 | // fold (rot* (rot* x, c2), c1) -> (rot* x, c1 +- c2 % bitsize) | |||
5592 | if (NextOp == ISD::ROTL || NextOp == ISD::ROTR) { | |||
5593 | SDNode *C1 = DAG.isConstantIntBuildVectorOrConstantInt(N1); | |||
5594 | SDNode *C2 = DAG.isConstantIntBuildVectorOrConstantInt(N0.getOperand(1)); | |||
5595 | if (C1 && C2 && C1->getValueType(0) == C2->getValueType(0)) { | |||
5596 | EVT ShiftVT = C1->getValueType(0); | |||
5597 | bool SameSide = (N->getOpcode() == NextOp); | |||
5598 | unsigned CombineOp = SameSide ? ISD::ADD : ISD::SUB; | |||
5599 | if (SDValue CombinedShift = | |||
5600 | DAG.FoldConstantArithmetic(CombineOp, dl, ShiftVT, C1, C2)) { | |||
5601 | SDValue BitsizeC = DAG.getConstant(Bitsize, dl, ShiftVT); | |||
5602 | SDValue CombinedShiftNorm = DAG.FoldConstantArithmetic( | |||
5603 | ISD::SREM, dl, ShiftVT, CombinedShift.getNode(), | |||
5604 | BitsizeC.getNode()); | |||
5605 | return DAG.getNode(N->getOpcode(), dl, VT, N0->getOperand(0), | |||
5606 | CombinedShiftNorm); | |||
5607 | } | |||
5608 | } | |||
5609 | } | |||
5610 | return SDValue(); | |||
5611 | } | |||
5612 | ||||
5613 | SDValue DAGCombiner::visitSHL(SDNode *N) { | |||
5614 | SDValue N0 = N->getOperand(0); | |||
5615 | SDValue N1 = N->getOperand(1); | |||
5616 | EVT VT = N0.getValueType(); | |||
5617 | unsigned OpSizeInBits = VT.getScalarSizeInBits(); | |||
5618 | ||||
5619 | // fold vector ops | |||
5620 | if (VT.isVector()) { | |||
5621 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
5622 | return FoldedVOp; | |||
5623 | ||||
5624 | BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1); | |||
5625 | // If setcc produces all-one true value then: | |||
5626 | // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV) | |||
5627 | if (N1CV && N1CV->isConstant()) { | |||
5628 | if (N0.getOpcode() == ISD::AND) { | |||
5629 | SDValue N00 = N0->getOperand(0); | |||
5630 | SDValue N01 = N0->getOperand(1); | |||
5631 | BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01); | |||
5632 | ||||
5633 | if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC && | |||
5634 | TLI.getBooleanContents(N00.getOperand(0).getValueType()) == | |||
5635 | TargetLowering::ZeroOrNegativeOneBooleanContent) { | |||
5636 | if (SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, SDLoc(N), VT, | |||
5637 | N01CV, N1CV)) | |||
5638 | return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C); | |||
5639 | } | |||
5640 | } | |||
5641 | } | |||
5642 | } | |||
5643 | ||||
5644 | ConstantSDNode *N1C = isConstOrConstSplat(N1); | |||
5645 | ||||
5646 | // fold (shl c1, c2) -> c1<<c2 | |||
5647 | ConstantSDNode *N0C = getAsNonOpaqueConstant(N0); | |||
5648 | if (N0C && N1C && !N1C->isOpaque()) | |||
5649 | return DAG.FoldConstantArithmetic(ISD::SHL, SDLoc(N), VT, N0C, N1C); | |||
5650 | // fold (shl 0, x) -> 0 | |||
5651 | if (isNullConstantOrNullSplatConstant(N0)) | |||
5652 | return N0; | |||
5653 | // fold (shl x, c >= size(x)) -> undef | |||
5654 | // NOTE: ALL vector elements must be too big to avoid partial UNDEFs. | |||
5655 | auto MatchShiftTooBig = [OpSizeInBits](ConstantSDNode *Val) { | |||
5656 | return Val->getAPIntValue().uge(OpSizeInBits); | |||
5657 | }; | |||
5658 | if (ISD::matchUnaryPredicate(N1, MatchShiftTooBig)) | |||
5659 | return DAG.getUNDEF(VT); | |||
5660 | // fold (shl x, 0) -> x | |||
5661 | if (N1C && N1C->isNullValue()) | |||
5662 | return N0; | |||
5663 | // fold (shl undef, x) -> 0 | |||
5664 | if (N0.isUndef()) | |||
5665 | return DAG.getConstant(0, SDLoc(N), VT); | |||
5666 | ||||
5667 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
5668 | return NewSel; | |||
5669 | ||||
5670 | // if (shl x, c) is known to be zero, return 0 | |||
5671 | if (DAG.MaskedValueIsZero(SDValue(N, 0), | |||
5672 | APInt::getAllOnesValue(OpSizeInBits))) | |||
5673 | return DAG.getConstant(0, SDLoc(N), VT); | |||
5674 | // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))). | |||
5675 | if (N1.getOpcode() == ISD::TRUNCATE && | |||
5676 | N1.getOperand(0).getOpcode() == ISD::AND) { | |||
5677 | if (SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode())) | |||
5678 | return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1); | |||
5679 | } | |||
5680 | ||||
5681 | if (N1C && SimplifyDemandedBits(SDValue(N, 0))) | |||
5682 | return SDValue(N, 0); | |||
5683 | ||||
5684 | // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2)) | |||
5685 | if (N0.getOpcode() == ISD::SHL) { | |||
5686 | auto MatchOutOfRange = [OpSizeInBits](ConstantSDNode *LHS, | |||
5687 | ConstantSDNode *RHS) { | |||
5688 | APInt c1 = LHS->getAPIntValue(); | |||
5689 | APInt c2 = RHS->getAPIntValue(); | |||
5690 | zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */); | |||
5691 | return (c1 + c2).uge(OpSizeInBits); | |||
5692 | }; | |||
5693 | if (ISD::matchBinaryPredicate(N1, N0.getOperand(1), MatchOutOfRange)) | |||
5694 | return DAG.getConstant(0, SDLoc(N), VT); | |||
5695 | ||||
5696 | auto MatchInRange = [OpSizeInBits](ConstantSDNode *LHS, | |||
5697 | ConstantSDNode *RHS) { | |||
5698 | APInt c1 = LHS->getAPIntValue(); | |||
5699 | APInt c2 = RHS->getAPIntValue(); | |||
5700 | zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */); | |||
5701 | return (c1 + c2).ult(OpSizeInBits); | |||
5702 | }; | |||
5703 | if (ISD::matchBinaryPredicate(N1, N0.getOperand(1), MatchInRange)) { | |||
5704 | SDLoc DL(N); | |||
5705 | EVT ShiftVT = N1.getValueType(); | |||
5706 | SDValue Sum = DAG.getNode(ISD::ADD, DL, ShiftVT, N1, N0.getOperand(1)); | |||
5707 | return DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0), Sum); | |||
5708 | } | |||
5709 | } | |||
5710 | ||||
5711 | // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2))) | |||
5712 | // For this to be valid, the second form must not preserve any of the bits | |||
5713 | // that are shifted out by the inner shift in the first form. This means | |||
5714 | // the outer shift size must be >= the number of bits added by the ext. | |||
5715 | // As a corollary, we don't care what kind of ext it is. | |||
5716 | if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND || | |||
5717 | N0.getOpcode() == ISD::ANY_EXTEND || | |||
5718 | N0.getOpcode() == ISD::SIGN_EXTEND) && | |||
5719 | N0.getOperand(0).getOpcode() == ISD::SHL) { | |||
5720 | SDValue N0Op0 = N0.getOperand(0); | |||
5721 | if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) { | |||
5722 | APInt c1 = N0Op0C1->getAPIntValue(); | |||
5723 | APInt c2 = N1C->getAPIntValue(); | |||
5724 | zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */); | |||
5725 | ||||
5726 | EVT InnerShiftVT = N0Op0.getValueType(); | |||
5727 | uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits(); | |||
5728 | if (c2.uge(OpSizeInBits - InnerShiftSize)) { | |||
5729 | SDLoc DL(N0); | |||
5730 | APInt Sum = c1 + c2; | |||
5731 | if (Sum.uge(OpSizeInBits)) | |||
5732 | return DAG.getConstant(0, DL, VT); | |||
5733 | ||||
5734 | return DAG.getNode( | |||
5735 | ISD::SHL, DL, VT, | |||
5736 | DAG.getNode(N0.getOpcode(), DL, VT, N0Op0->getOperand(0)), | |||
5737 | DAG.getConstant(Sum.getZExtValue(), DL, N1.getValueType())); | |||
5738 | } | |||
5739 | } | |||
5740 | } | |||
5741 | ||||
5742 | // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C)) | |||
5743 | // Only fold this if the inner zext has no other uses to avoid increasing | |||
5744 | // the total number of instructions. | |||
5745 | if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() && | |||
5746 | N0.getOperand(0).getOpcode() == ISD::SRL) { | |||
5747 | SDValue N0Op0 = N0.getOperand(0); | |||
5748 | if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) { | |||
5749 | if (N0Op0C1->getAPIntValue().ult(VT.getScalarSizeInBits())) { | |||
5750 | uint64_t c1 = N0Op0C1->getZExtValue(); | |||
5751 | uint64_t c2 = N1C->getZExtValue(); | |||
5752 | if (c1 == c2) { | |||
5753 | SDValue NewOp0 = N0.getOperand(0); | |||
5754 | EVT CountVT = NewOp0.getOperand(1).getValueType(); | |||
5755 | SDLoc DL(N); | |||
5756 | SDValue NewSHL = DAG.getNode(ISD::SHL, DL, NewOp0.getValueType(), | |||
5757 | NewOp0, | |||
5758 | DAG.getConstant(c2, DL, CountVT)); | |||
5759 | AddToWorklist(NewSHL.getNode()); | |||
5760 | return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL); | |||
5761 | } | |||
5762 | } | |||
5763 | } | |||
5764 | } | |||
5765 | ||||
5766 | // fold (shl (sr[la] exact X, C1), C2) -> (shl X, (C2-C1)) if C1 <= C2 | |||
5767 | // fold (shl (sr[la] exact X, C1), C2) -> (sr[la] X, (C2-C1)) if C1 > C2 | |||
5768 | if (N1C && (N0.getOpcode() == ISD::SRL || N0.getOpcode() == ISD::SRA) && | |||
5769 | N0->getFlags().hasExact()) { | |||
5770 | if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) { | |||
5771 | uint64_t C1 = N0C1->getZExtValue(); | |||
5772 | uint64_t C2 = N1C->getZExtValue(); | |||
5773 | SDLoc DL(N); | |||
5774 | if (C1 <= C2) | |||
5775 | return DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0), | |||
5776 | DAG.getConstant(C2 - C1, DL, N1.getValueType())); | |||
5777 | return DAG.getNode(N0.getOpcode(), DL, VT, N0.getOperand(0), | |||
5778 | DAG.getConstant(C1 - C2, DL, N1.getValueType())); | |||
5779 | } | |||
5780 | } | |||
5781 | ||||
5782 | // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or | |||
5783 | // (and (srl x, (sub c1, c2), MASK) | |||
5784 | // Only fold this if the inner shift has no other uses -- if it does, folding | |||
5785 | // this will increase the total number of instructions. | |||
5786 | if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) { | |||
5787 | if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) { | |||
5788 | uint64_t c1 = N0C1->getZExtValue(); | |||
5789 | if (c1 < OpSizeInBits) { | |||
5790 | uint64_t c2 = N1C->getZExtValue(); | |||
5791 | APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1); | |||
5792 | SDValue Shift; | |||
5793 | if (c2 > c1) { | |||
5794 | Mask <<= c2 - c1; | |||
5795 | SDLoc DL(N); | |||
5796 | Shift = DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0), | |||
5797 | DAG.getConstant(c2 - c1, DL, N1.getValueType())); | |||
5798 | } else { | |||
5799 | Mask.lshrInPlace(c1 - c2); | |||
5800 | SDLoc DL(N); | |||
5801 | Shift = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0), | |||
5802 | DAG.getConstant(c1 - c2, DL, N1.getValueType())); | |||
5803 | } | |||
5804 | SDLoc DL(N0); | |||
5805 | return DAG.getNode(ISD::AND, DL, VT, Shift, | |||
5806 | DAG.getConstant(Mask, DL, VT)); | |||
5807 | } | |||
5808 | } | |||
5809 | } | |||
5810 | ||||
5811 | // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1)) | |||
5812 | if (N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1) && | |||
5813 | isConstantOrConstantVector(N1, /* No Opaques */ true)) { | |||
5814 | SDLoc DL(N); | |||
5815 | SDValue AllBits = DAG.getAllOnesConstant(DL, VT); | |||
5816 | SDValue HiBitsMask = DAG.getNode(ISD::SHL, DL, VT, AllBits, N1); | |||
5817 | return DAG.getNode(ISD::AND, DL, VT, N0.getOperand(0), HiBitsMask); | |||
5818 | } | |||
5819 | ||||
5820 | // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) | |||
5821 | // fold (shl (or x, c1), c2) -> (or (shl x, c2), c1 << c2) | |||
5822 | // Variant of version done on multiply, except mul by a power of 2 is turned | |||
5823 | // into a shift. | |||
5824 | if ((N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::OR) && | |||
5825 | N0.getNode()->hasOneUse() && | |||
5826 | isConstantOrConstantVector(N1, /* No Opaques */ true) && | |||
5827 | isConstantOrConstantVector(N0.getOperand(1), /* No Opaques */ true)) { | |||
5828 | SDValue Shl0 = DAG.getNode(ISD::SHL, SDLoc(N0), VT, N0.getOperand(0), N1); | |||
5829 | SDValue Shl1 = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1); | |||
5830 | AddToWorklist(Shl0.getNode()); | |||
5831 | AddToWorklist(Shl1.getNode()); | |||
5832 | return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, Shl0, Shl1); | |||
5833 | } | |||
5834 | ||||
5835 | // fold (shl (mul x, c1), c2) -> (mul x, c1 << c2) | |||
5836 | if (N0.getOpcode() == ISD::MUL && N0.getNode()->hasOneUse() && | |||
5837 | isConstantOrConstantVector(N1, /* No Opaques */ true) && | |||
5838 | isConstantOrConstantVector(N0.getOperand(1), /* No Opaques */ true)) { | |||
5839 | SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1); | |||
5840 | if (isConstantOrConstantVector(Shl)) | |||
5841 | return DAG.getNode(ISD::MUL, SDLoc(N), VT, N0.getOperand(0), Shl); | |||
5842 | } | |||
5843 | ||||
5844 | if (N1C && !N1C->isOpaque()) | |||
5845 | if (SDValue NewSHL = visitShiftByConstant(N, N1C)) | |||
5846 | return NewSHL; | |||
5847 | ||||
5848 | return SDValue(); | |||
5849 | } | |||
5850 | ||||
5851 | SDValue DAGCombiner::visitSRA(SDNode *N) { | |||
5852 | SDValue N0 = N->getOperand(0); | |||
5853 | SDValue N1 = N->getOperand(1); | |||
5854 | EVT VT = N0.getValueType(); | |||
5855 | unsigned OpSizeInBits = VT.getScalarSizeInBits(); | |||
5856 | ||||
5857 | // Arithmetic shifting an all-sign-bit value is a no-op. | |||
5858 | // fold (sra 0, x) -> 0 | |||
5859 | // fold (sra -1, x) -> -1 | |||
5860 | if (DAG.ComputeNumSignBits(N0) == OpSizeInBits) | |||
5861 | return N0; | |||
5862 | ||||
5863 | // fold vector ops | |||
5864 | if (VT.isVector()) | |||
5865 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
5866 | return FoldedVOp; | |||
5867 | ||||
5868 | ConstantSDNode *N1C = isConstOrConstSplat(N1); | |||
5869 | ||||
5870 | // fold (sra c1, c2) -> (sra c1, c2) | |||
5871 | ConstantSDNode *N0C = getAsNonOpaqueConstant(N0); | |||
5872 | if (N0C && N1C && !N1C->isOpaque()) | |||
5873 | return DAG.FoldConstantArithmetic(ISD::SRA, SDLoc(N), VT, N0C, N1C); | |||
5874 | // fold (sra x, c >= size(x)) -> undef | |||
5875 | // NOTE: ALL vector elements must be too big to avoid partial UNDEFs. | |||
5876 | auto MatchShiftTooBig = [OpSizeInBits](ConstantSDNode *Val) { | |||
5877 | return Val->getAPIntValue().uge(OpSizeInBits); | |||
5878 | }; | |||
5879 | if (ISD::matchUnaryPredicate(N1, MatchShiftTooBig)) | |||
5880 | return DAG.getUNDEF(VT); | |||
5881 | // fold (sra x, 0) -> x | |||
5882 | if (N1C && N1C->isNullValue()) | |||
5883 | return N0; | |||
5884 | ||||
5885 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
5886 | return NewSel; | |||
5887 | ||||
5888 | // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports | |||
5889 | // sext_inreg. | |||
5890 | if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) { | |||
5891 | unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue(); | |||
5892 | EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits); | |||
5893 | if (VT.isVector()) | |||
5894 | ExtVT = EVT::getVectorVT(*DAG.getContext(), | |||
5895 | ExtVT, VT.getVectorNumElements()); | |||
5896 | if ((!LegalOperations || | |||
5897 | TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT))) | |||
5898 | return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, | |||
5899 | N0.getOperand(0), DAG.getValueType(ExtVT)); | |||
5900 | } | |||
5901 | ||||
5902 | // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2)) | |||
5903 | if (N0.getOpcode() == ISD::SRA) { | |||
5904 | SDLoc DL(N); | |||
5905 | EVT ShiftVT = N1.getValueType(); | |||
5906 | ||||
5907 | auto MatchOutOfRange = [OpSizeInBits](ConstantSDNode *LHS, | |||
5908 | ConstantSDNode *RHS) { | |||
5909 | APInt c1 = LHS->getAPIntValue(); | |||
5910 | APInt c2 = RHS->getAPIntValue(); | |||
5911 | zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */); | |||
5912 | return (c1 + c2).uge(OpSizeInBits); | |||
5913 | }; | |||
5914 | if (ISD::matchBinaryPredicate(N1, N0.getOperand(1), MatchOutOfRange)) | |||
5915 | return DAG.getNode(ISD::SRA, DL, VT, N0.getOperand(0), | |||
5916 | DAG.getConstant(OpSizeInBits - 1, DL, ShiftVT)); | |||
5917 | ||||
5918 | auto MatchInRange = [OpSizeInBits](ConstantSDNode *LHS, | |||
5919 | ConstantSDNode *RHS) { | |||
5920 | APInt c1 = LHS->getAPIntValue(); | |||
5921 | APInt c2 = RHS->getAPIntValue(); | |||
5922 | zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */); | |||
5923 | return (c1 + c2).ult(OpSizeInBits); | |||
5924 | }; | |||
5925 | if (ISD::matchBinaryPredicate(N1, N0.getOperand(1), MatchInRange)) { | |||
5926 | SDValue Sum = DAG.getNode(ISD::ADD, DL, ShiftVT, N1, N0.getOperand(1)); | |||
5927 | return DAG.getNode(ISD::SRA, DL, VT, N0.getOperand(0), Sum); | |||
5928 | } | |||
5929 | } | |||
5930 | ||||
5931 | // fold (sra (shl X, m), (sub result_size, n)) | |||
5932 | // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for | |||
5933 | // result_size - n != m. | |||
5934 | // If truncate is free for the target sext(shl) is likely to result in better | |||
5935 | // code. | |||
5936 | if (N0.getOpcode() == ISD::SHL && N1C) { | |||
5937 | // Get the two constanst of the shifts, CN0 = m, CN = n. | |||
5938 | const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1)); | |||
5939 | if (N01C) { | |||
5940 | LLVMContext &Ctx = *DAG.getContext(); | |||
5941 | // Determine what the truncate's result bitsize and type would be. | |||
5942 | EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue()); | |||
5943 | ||||
5944 | if (VT.isVector()) | |||
5945 | TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements()); | |||
5946 | ||||
5947 | // Determine the residual right-shift amount. | |||
5948 | int ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue(); | |||
5949 | ||||
5950 | // If the shift is not a no-op (in which case this should be just a sign | |||
5951 | // extend already), the truncated to type is legal, sign_extend is legal | |||
5952 | // on that type, and the truncate to that type is both legal and free, | |||
5953 | // perform the transform. | |||
5954 | if ((ShiftAmt > 0) && | |||
5955 | TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) && | |||
5956 | TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) && | |||
5957 | TLI.isTruncateFree(VT, TruncVT)) { | |||
5958 | SDLoc DL(N); | |||
5959 | SDValue Amt = DAG.getConstant(ShiftAmt, DL, | |||
5960 | getShiftAmountTy(N0.getOperand(0).getValueType())); | |||
5961 | SDValue Shift = DAG.getNode(ISD::SRL, DL, VT, | |||
5962 | N0.getOperand(0), Amt); | |||
5963 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, TruncVT, | |||
5964 | Shift); | |||
5965 | return DAG.getNode(ISD::SIGN_EXTEND, DL, | |||
5966 | N->getValueType(0), Trunc); | |||
5967 | } | |||
5968 | } | |||
5969 | } | |||
5970 | ||||
5971 | // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))). | |||
5972 | if (N1.getOpcode() == ISD::TRUNCATE && | |||
5973 | N1.getOperand(0).getOpcode() == ISD::AND) { | |||
5974 | if (SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode())) | |||
5975 | return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1); | |||
5976 | } | |||
5977 | ||||
5978 | // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2)) | |||
5979 | // if c1 is equal to the number of bits the trunc removes | |||
5980 | if (N0.getOpcode() == ISD::TRUNCATE && | |||
5981 | (N0.getOperand(0).getOpcode() == ISD::SRL || | |||
5982 | N0.getOperand(0).getOpcode() == ISD::SRA) && | |||
5983 | N0.getOperand(0).hasOneUse() && | |||
5984 | N0.getOperand(0).getOperand(1).hasOneUse() && | |||
5985 | N1C) { | |||
5986 | SDValue N0Op0 = N0.getOperand(0); | |||
5987 | if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) { | |||
5988 | unsigned LargeShiftVal = LargeShift->getZExtValue(); | |||
5989 | EVT LargeVT = N0Op0.getValueType(); | |||
5990 | ||||
5991 | if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) { | |||
5992 | SDLoc DL(N); | |||
5993 | SDValue Amt = | |||
5994 | DAG.getConstant(LargeShiftVal + N1C->getZExtValue(), DL, | |||
5995 | getShiftAmountTy(N0Op0.getOperand(0).getValueType())); | |||
5996 | SDValue SRA = DAG.getNode(ISD::SRA, DL, LargeVT, | |||
5997 | N0Op0.getOperand(0), Amt); | |||
5998 | return DAG.getNode(ISD::TRUNCATE, DL, VT, SRA); | |||
5999 | } | |||
6000 | } | |||
6001 | } | |||
6002 | ||||
6003 | // Simplify, based on bits shifted out of the LHS. | |||
6004 | if (N1C && SimplifyDemandedBits(SDValue(N, 0))) | |||
6005 | return SDValue(N, 0); | |||
6006 | ||||
6007 | // If the sign bit is known to be zero, switch this to a SRL. | |||
6008 | if (DAG.SignBitIsZero(N0)) | |||
6009 | return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1); | |||
6010 | ||||
6011 | if (N1C && !N1C->isOpaque()) | |||
6012 | if (SDValue NewSRA = visitShiftByConstant(N, N1C)) | |||
6013 | return NewSRA; | |||
6014 | ||||
6015 | return SDValue(); | |||
6016 | } | |||
6017 | ||||
6018 | SDValue DAGCombiner::visitSRL(SDNode *N) { | |||
6019 | SDValue N0 = N->getOperand(0); | |||
6020 | SDValue N1 = N->getOperand(1); | |||
6021 | EVT VT = N0.getValueType(); | |||
6022 | unsigned OpSizeInBits = VT.getScalarSizeInBits(); | |||
6023 | ||||
6024 | // fold vector ops | |||
6025 | if (VT.isVector()) | |||
6026 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
6027 | return FoldedVOp; | |||
6028 | ||||
6029 | ConstantSDNode *N1C = isConstOrConstSplat(N1); | |||
6030 | ||||
6031 | // fold (srl c1, c2) -> c1 >>u c2 | |||
6032 | ConstantSDNode *N0C = getAsNonOpaqueConstant(N0); | |||
6033 | if (N0C && N1C && !N1C->isOpaque()) | |||
6034 | return DAG.FoldConstantArithmetic(ISD::SRL, SDLoc(N), VT, N0C, N1C); | |||
6035 | // fold (srl 0, x) -> 0 | |||
6036 | if (isNullConstantOrNullSplatConstant(N0)) | |||
6037 | return N0; | |||
6038 | // fold (srl x, c >= size(x)) -> undef | |||
6039 | // NOTE: ALL vector elements must be too big to avoid partial UNDEFs. | |||
6040 | auto MatchShiftTooBig = [OpSizeInBits](ConstantSDNode *Val) { | |||
6041 | return Val->getAPIntValue().uge(OpSizeInBits); | |||
6042 | }; | |||
6043 | if (ISD::matchUnaryPredicate(N1, MatchShiftTooBig)) | |||
6044 | return DAG.getUNDEF(VT); | |||
6045 | // fold (srl x, 0) -> x | |||
6046 | if (N1C && N1C->isNullValue()) | |||
6047 | return N0; | |||
6048 | ||||
6049 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
6050 | return NewSel; | |||
6051 | ||||
6052 | // if (srl x, c) is known to be zero, return 0 | |||
6053 | if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0), | |||
6054 | APInt::getAllOnesValue(OpSizeInBits))) | |||
6055 | return DAG.getConstant(0, SDLoc(N), VT); | |||
6056 | ||||
6057 | // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2)) | |||
6058 | if (N0.getOpcode() == ISD::SRL) { | |||
6059 | auto MatchOutOfRange = [OpSizeInBits](ConstantSDNode *LHS, | |||
6060 | ConstantSDNode *RHS) { | |||
6061 | APInt c1 = LHS->getAPIntValue(); | |||
6062 | APInt c2 = RHS->getAPIntValue(); | |||
6063 | zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */); | |||
6064 | return (c1 + c2).uge(OpSizeInBits); | |||
6065 | }; | |||
6066 | if (ISD::matchBinaryPredicate(N1, N0.getOperand(1), MatchOutOfRange)) | |||
6067 | return DAG.getConstant(0, SDLoc(N), VT); | |||
6068 | ||||
6069 | auto MatchInRange = [OpSizeInBits](ConstantSDNode *LHS, | |||
6070 | ConstantSDNode *RHS) { | |||
6071 | APInt c1 = LHS->getAPIntValue(); | |||
6072 | APInt c2 = RHS->getAPIntValue(); | |||
6073 | zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */); | |||
6074 | return (c1 + c2).ult(OpSizeInBits); | |||
6075 | }; | |||
6076 | if (ISD::matchBinaryPredicate(N1, N0.getOperand(1), MatchInRange)) { | |||
6077 | SDLoc DL(N); | |||
6078 | EVT ShiftVT = N1.getValueType(); | |||
6079 | SDValue Sum = DAG.getNode(ISD::ADD, DL, ShiftVT, N1, N0.getOperand(1)); | |||
6080 | return DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0), Sum); | |||
6081 | } | |||
6082 | } | |||
6083 | ||||
6084 | // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2))) | |||
6085 | if (N1C && N0.getOpcode() == ISD::TRUNCATE && | |||
6086 | N0.getOperand(0).getOpcode() == ISD::SRL) { | |||
6087 | if (auto N001C = isConstOrConstSplat(N0.getOperand(0).getOperand(1))) { | |||
6088 | uint64_t c1 = N001C->getZExtValue(); | |||
6089 | uint64_t c2 = N1C->getZExtValue(); | |||
6090 | EVT InnerShiftVT = N0.getOperand(0).getValueType(); | |||
6091 | EVT ShiftCountVT = N0.getOperand(0).getOperand(1).getValueType(); | |||
6092 | uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits(); | |||
6093 | // This is only valid if the OpSizeInBits + c1 = size of inner shift. | |||
6094 | if (c1 + OpSizeInBits == InnerShiftSize) { | |||
6095 | SDLoc DL(N0); | |||
6096 | if (c1 + c2 >= InnerShiftSize) | |||
6097 | return DAG.getConstant(0, DL, VT); | |||
6098 | return DAG.getNode(ISD::TRUNCATE, DL, VT, | |||
6099 | DAG.getNode(ISD::SRL, DL, InnerShiftVT, | |||
6100 | N0.getOperand(0).getOperand(0), | |||
6101 | DAG.getConstant(c1 + c2, DL, | |||
6102 | ShiftCountVT))); | |||
6103 | } | |||
6104 | } | |||
6105 | } | |||
6106 | ||||
6107 | // fold (srl (shl x, c), c) -> (and x, cst2) | |||
6108 | if (N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 && | |||
6109 | isConstantOrConstantVector(N1, /* NoOpaques */ true)) { | |||
6110 | SDLoc DL(N); | |||
6111 | SDValue Mask = | |||
6112 | DAG.getNode(ISD::SRL, DL, VT, DAG.getAllOnesConstant(DL, VT), N1); | |||
6113 | AddToWorklist(Mask.getNode()); | |||
6114 | return DAG.getNode(ISD::AND, DL, VT, N0.getOperand(0), Mask); | |||
6115 | } | |||
6116 | ||||
6117 | // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask) | |||
6118 | if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) { | |||
6119 | // Shifting in all undef bits? | |||
6120 | EVT SmallVT = N0.getOperand(0).getValueType(); | |||
6121 | unsigned BitSize = SmallVT.getScalarSizeInBits(); | |||
6122 | if (N1C->getZExtValue() >= BitSize) | |||
6123 | return DAG.getUNDEF(VT); | |||
6124 | ||||
6125 | if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) { | |||
6126 | uint64_t ShiftAmt = N1C->getZExtValue(); | |||
6127 | SDLoc DL0(N0); | |||
6128 | SDValue SmallShift = DAG.getNode(ISD::SRL, DL0, SmallVT, | |||
6129 | N0.getOperand(0), | |||
6130 | DAG.getConstant(ShiftAmt, DL0, | |||
6131 | getShiftAmountTy(SmallVT))); | |||
6132 | AddToWorklist(SmallShift.getNode()); | |||
6133 | APInt Mask = APInt::getLowBitsSet(OpSizeInBits, OpSizeInBits - ShiftAmt); | |||
6134 | SDLoc DL(N); | |||
6135 | return DAG.getNode(ISD::AND, DL, VT, | |||
6136 | DAG.getNode(ISD::ANY_EXTEND, DL, VT, SmallShift), | |||
6137 | DAG.getConstant(Mask, DL, VT)); | |||
6138 | } | |||
6139 | } | |||
6140 | ||||
6141 | // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign | |||
6142 | // bit, which is unmodified by sra. | |||
6143 | if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) { | |||
6144 | if (N0.getOpcode() == ISD::SRA) | |||
6145 | return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1); | |||
6146 | } | |||
6147 | ||||
6148 | // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit). | |||
6149 | if (N1C && N0.getOpcode() == ISD::CTLZ && | |||
6150 | N1C->getAPIntValue() == Log2_32(OpSizeInBits)) { | |||
6151 | KnownBits Known; | |||
6152 | DAG.computeKnownBits(N0.getOperand(0), Known); | |||
6153 | ||||
6154 | // If any of the input bits are KnownOne, then the input couldn't be all | |||
6155 | // zeros, thus the result of the srl will always be zero. | |||
6156 | if (Known.One.getBoolValue()) return DAG.getConstant(0, SDLoc(N0), VT); | |||
6157 | ||||
6158 | // If all of the bits input the to ctlz node are known to be zero, then | |||
6159 | // the result of the ctlz is "32" and the result of the shift is one. | |||
6160 | APInt UnknownBits = ~Known.Zero; | |||
6161 | if (UnknownBits == 0) return DAG.getConstant(1, SDLoc(N0), VT); | |||
6162 | ||||
6163 | // Otherwise, check to see if there is exactly one bit input to the ctlz. | |||
6164 | if (UnknownBits.isPowerOf2()) { | |||
6165 | // Okay, we know that only that the single bit specified by UnknownBits | |||
6166 | // could be set on input to the CTLZ node. If this bit is set, the SRL | |||
6167 | // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair | |||
6168 | // to an SRL/XOR pair, which is likely to simplify more. | |||
6169 | unsigned ShAmt = UnknownBits.countTrailingZeros(); | |||
6170 | SDValue Op = N0.getOperand(0); | |||
6171 | ||||
6172 | if (ShAmt) { | |||
6173 | SDLoc DL(N0); | |||
6174 | Op = DAG.getNode(ISD::SRL, DL, VT, Op, | |||
6175 | DAG.getConstant(ShAmt, DL, | |||
6176 | getShiftAmountTy(Op.getValueType()))); | |||
6177 | AddToWorklist(Op.getNode()); | |||
6178 | } | |||
6179 | ||||
6180 | SDLoc DL(N); | |||
6181 | return DAG.getNode(ISD::XOR, DL, VT, | |||
6182 | Op, DAG.getConstant(1, DL, VT)); | |||
6183 | } | |||
6184 | } | |||
6185 | ||||
6186 | // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))). | |||
6187 | if (N1.getOpcode() == ISD::TRUNCATE && | |||
6188 | N1.getOperand(0).getOpcode() == ISD::AND) { | |||
6189 | if (SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode())) | |||
6190 | return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1); | |||
6191 | } | |||
6192 | ||||
6193 | // fold operands of srl based on knowledge that the low bits are not | |||
6194 | // demanded. | |||
6195 | if (N1C && SimplifyDemandedBits(SDValue(N, 0))) | |||
6196 | return SDValue(N, 0); | |||
6197 | ||||
6198 | if (N1C && !N1C->isOpaque()) | |||
6199 | if (SDValue NewSRL = visitShiftByConstant(N, N1C)) | |||
6200 | return NewSRL; | |||
6201 | ||||
6202 | // Attempt to convert a srl of a load into a narrower zero-extending load. | |||
6203 | if (SDValue NarrowLoad = ReduceLoadWidth(N)) | |||
6204 | return NarrowLoad; | |||
6205 | ||||
6206 | // Here is a common situation. We want to optimize: | |||
6207 | // | |||
6208 | // %a = ... | |||
6209 | // %b = and i32 %a, 2 | |||
6210 | // %c = srl i32 %b, 1 | |||
6211 | // brcond i32 %c ... | |||
6212 | // | |||
6213 | // into | |||
6214 | // | |||
6215 | // %a = ... | |||
6216 | // %b = and %a, 2 | |||
6217 | // %c = setcc eq %b, 0 | |||
6218 | // brcond %c ... | |||
6219 | // | |||
6220 | // However when after the source operand of SRL is optimized into AND, the SRL | |||
6221 | // itself may not be optimized further. Look for it and add the BRCOND into | |||
6222 | // the worklist. | |||
6223 | if (N->hasOneUse()) { | |||
6224 | SDNode *Use = *N->use_begin(); | |||
6225 | if (Use->getOpcode() == ISD::BRCOND) | |||
6226 | AddToWorklist(Use); | |||
6227 | else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) { | |||
6228 | // Also look pass the truncate. | |||
6229 | Use = *Use->use_begin(); | |||
6230 | if (Use->getOpcode() == ISD::BRCOND) | |||
6231 | AddToWorklist(Use); | |||
6232 | } | |||
6233 | } | |||
6234 | ||||
6235 | return SDValue(); | |||
6236 | } | |||
6237 | ||||
6238 | SDValue DAGCombiner::visitABS(SDNode *N) { | |||
6239 | SDValue N0 = N->getOperand(0); | |||
6240 | EVT VT = N->getValueType(0); | |||
6241 | ||||
6242 | // fold (abs c1) -> c2 | |||
6243 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) | |||
6244 | return DAG.getNode(ISD::ABS, SDLoc(N), VT, N0); | |||
6245 | // fold (abs (abs x)) -> (abs x) | |||
6246 | if (N0.getOpcode() == ISD::ABS) | |||
6247 | return N0; | |||
6248 | // fold (abs x) -> x iff not-negative | |||
6249 | if (DAG.SignBitIsZero(N0)) | |||
6250 | return N0; | |||
6251 | return SDValue(); | |||
6252 | } | |||
6253 | ||||
6254 | SDValue DAGCombiner::visitBSWAP(SDNode *N) { | |||
6255 | SDValue N0 = N->getOperand(0); | |||
6256 | EVT VT = N->getValueType(0); | |||
6257 | ||||
6258 | // fold (bswap c1) -> c2 | |||
6259 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) | |||
6260 | return DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N0); | |||
6261 | // fold (bswap (bswap x)) -> x | |||
6262 | if (N0.getOpcode() == ISD::BSWAP) | |||
6263 | return N0->getOperand(0); | |||
6264 | return SDValue(); | |||
6265 | } | |||
6266 | ||||
6267 | SDValue DAGCombiner::visitBITREVERSE(SDNode *N) { | |||
6268 | SDValue N0 = N->getOperand(0); | |||
6269 | EVT VT = N->getValueType(0); | |||
6270 | ||||
6271 | // fold (bitreverse c1) -> c2 | |||
6272 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) | |||
6273 | return DAG.getNode(ISD::BITREVERSE, SDLoc(N), VT, N0); | |||
6274 | // fold (bitreverse (bitreverse x)) -> x | |||
6275 | if (N0.getOpcode() == ISD::BITREVERSE) | |||
6276 | return N0.getOperand(0); | |||
6277 | return SDValue(); | |||
6278 | } | |||
6279 | ||||
6280 | SDValue DAGCombiner::visitCTLZ(SDNode *N) { | |||
6281 | SDValue N0 = N->getOperand(0); | |||
6282 | EVT VT = N->getValueType(0); | |||
6283 | ||||
6284 | // fold (ctlz c1) -> c2 | |||
6285 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) | |||
6286 | return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0); | |||
6287 | ||||
6288 | // If the value is known never to be zero, switch to the undef version. | |||
6289 | if (!LegalOperations || TLI.isOperationLegal(ISD::CTLZ_ZERO_UNDEF, VT)) { | |||
6290 | if (DAG.isKnownNeverZero(N0)) | |||
6291 | return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0); | |||
6292 | } | |||
6293 | ||||
6294 | return SDValue(); | |||
6295 | } | |||
6296 | ||||
6297 | SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) { | |||
6298 | SDValue N0 = N->getOperand(0); | |||
6299 | EVT VT = N->getValueType(0); | |||
6300 | ||||
6301 | // fold (ctlz_zero_undef c1) -> c2 | |||
6302 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) | |||
6303 | return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0); | |||
6304 | return SDValue(); | |||
6305 | } | |||
6306 | ||||
6307 | SDValue DAGCombiner::visitCTTZ(SDNode *N) { | |||
6308 | SDValue N0 = N->getOperand(0); | |||
6309 | EVT VT = N->getValueType(0); | |||
6310 | ||||
6311 | // fold (cttz c1) -> c2 | |||
6312 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) | |||
6313 | return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0); | |||
6314 | ||||
6315 | // If the value is known never to be zero, switch to the undef version. | |||
6316 | if (!LegalOperations || TLI.isOperationLegal(ISD::CTTZ_ZERO_UNDEF, VT)) { | |||
6317 | if (DAG.isKnownNeverZero(N0)) | |||
6318 | return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0); | |||
6319 | } | |||
6320 | ||||
6321 | return SDValue(); | |||
6322 | } | |||
6323 | ||||
6324 | SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) { | |||
6325 | SDValue N0 = N->getOperand(0); | |||
6326 | EVT VT = N->getValueType(0); | |||
6327 | ||||
6328 | // fold (cttz_zero_undef c1) -> c2 | |||
6329 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) | |||
6330 | return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0); | |||
6331 | return SDValue(); | |||
6332 | } | |||
6333 | ||||
6334 | SDValue DAGCombiner::visitCTPOP(SDNode *N) { | |||
6335 | SDValue N0 = N->getOperand(0); | |||
6336 | EVT VT = N->getValueType(0); | |||
6337 | ||||
6338 | // fold (ctpop c1) -> c2 | |||
6339 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) | |||
6340 | return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0); | |||
6341 | return SDValue(); | |||
6342 | } | |||
6343 | ||||
6344 | /// \brief Generate Min/Max node | |||
6345 | static SDValue combineMinNumMaxNum(const SDLoc &DL, EVT VT, SDValue LHS, | |||
6346 | SDValue RHS, SDValue True, SDValue False, | |||
6347 | ISD::CondCode CC, const TargetLowering &TLI, | |||
6348 | SelectionDAG &DAG) { | |||
6349 | if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True)) | |||
6350 | return SDValue(); | |||
6351 | ||||
6352 | switch (CC) { | |||
6353 | case ISD::SETOLT: | |||
6354 | case ISD::SETOLE: | |||
6355 | case ISD::SETLT: | |||
6356 | case ISD::SETLE: | |||
6357 | case ISD::SETULT: | |||
6358 | case ISD::SETULE: { | |||
6359 | unsigned Opcode = (LHS == True) ? ISD::FMINNUM : ISD::FMAXNUM; | |||
6360 | if (TLI.isOperationLegal(Opcode, VT)) | |||
6361 | return DAG.getNode(Opcode, DL, VT, LHS, RHS); | |||
6362 | return SDValue(); | |||
6363 | } | |||
6364 | case ISD::SETOGT: | |||
6365 | case ISD::SETOGE: | |||
6366 | case ISD::SETGT: | |||
6367 | case ISD::SETGE: | |||
6368 | case ISD::SETUGT: | |||
6369 | case ISD::SETUGE: { | |||
6370 | unsigned Opcode = (LHS == True) ? ISD::FMAXNUM : ISD::FMINNUM; | |||
6371 | if (TLI.isOperationLegal(Opcode, VT)) | |||
6372 | return DAG.getNode(Opcode, DL, VT, LHS, RHS); | |||
6373 | return SDValue(); | |||
6374 | } | |||
6375 | default: | |||
6376 | return SDValue(); | |||
6377 | } | |||
6378 | } | |||
6379 | ||||
6380 | SDValue DAGCombiner::foldSelectOfConstants(SDNode *N) { | |||
6381 | SDValue Cond = N->getOperand(0); | |||
6382 | SDValue N1 = N->getOperand(1); | |||
6383 | SDValue N2 = N->getOperand(2); | |||
6384 | EVT VT = N->getValueType(0); | |||
6385 | EVT CondVT = Cond.getValueType(); | |||
6386 | SDLoc DL(N); | |||
6387 | ||||
6388 | if (!VT.isInteger()) | |||
6389 | return SDValue(); | |||
6390 | ||||
6391 | auto *C1 = dyn_cast<ConstantSDNode>(N1); | |||
6392 | auto *C2 = dyn_cast<ConstantSDNode>(N2); | |||
6393 | if (!C1 || !C2) | |||
6394 | return SDValue(); | |||
6395 | ||||
6396 | // Only do this before legalization to avoid conflicting with target-specific | |||
6397 | // transforms in the other direction (create a select from a zext/sext). There | |||
6398 | // is also a target-independent combine here in DAGCombiner in the other | |||
6399 | // direction for (select Cond, -1, 0) when the condition is not i1. | |||
6400 | if (CondVT == MVT::i1 && !LegalOperations) { | |||
6401 | if (C1->isNullValue() && C2->isOne()) { | |||
6402 | // select Cond, 0, 1 --> zext (!Cond) | |||
6403 | SDValue NotCond = DAG.getNOT(DL, Cond, MVT::i1); | |||
6404 | if (VT != MVT::i1) | |||
6405 | NotCond = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, NotCond); | |||
6406 | return NotCond; | |||
6407 | } | |||
6408 | if (C1->isNullValue() && C2->isAllOnesValue()) { | |||
6409 | // select Cond, 0, -1 --> sext (!Cond) | |||
6410 | SDValue NotCond = DAG.getNOT(DL, Cond, MVT::i1); | |||
6411 | if (VT != MVT::i1) | |||
6412 | NotCond = DAG.getNode(ISD::SIGN_EXTEND, DL, VT, NotCond); | |||
6413 | return NotCond; | |||
6414 | } | |||
6415 | if (C1->isOne() && C2->isNullValue()) { | |||
6416 | // select Cond, 1, 0 --> zext (Cond) | |||
6417 | if (VT != MVT::i1) | |||
6418 | Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Cond); | |||
6419 | return Cond; | |||
6420 | } | |||
6421 | if (C1->isAllOnesValue() && C2->isNullValue()) { | |||
6422 | // select Cond, -1, 0 --> sext (Cond) | |||
6423 | if (VT != MVT::i1) | |||
6424 | Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, VT, Cond); | |||
6425 | return Cond; | |||
6426 | } | |||
6427 | ||||
6428 | // For any constants that differ by 1, we can transform the select into an | |||
6429 | // extend and add. Use a target hook because some targets may prefer to | |||
6430 | // transform in the other direction. | |||
6431 | if (TLI.convertSelectOfConstantsToMath(VT)) { | |||
6432 | if (C1->getAPIntValue() - 1 == C2->getAPIntValue()) { | |||
6433 | // select Cond, C1, C1-1 --> add (zext Cond), C1-1 | |||
6434 | if (VT != MVT::i1) | |||
6435 | Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Cond); | |||
6436 | return DAG.getNode(ISD::ADD, DL, VT, Cond, N2); | |||
6437 | } | |||
6438 | if (C1->getAPIntValue() + 1 == C2->getAPIntValue()) { | |||
6439 | // select Cond, C1, C1+1 --> add (sext Cond), C1+1 | |||
6440 | if (VT != MVT::i1) | |||
6441 | Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, VT, Cond); | |||
6442 | return DAG.getNode(ISD::ADD, DL, VT, Cond, N2); | |||
6443 | } | |||
6444 | } | |||
6445 | ||||
6446 | return SDValue(); | |||
6447 | } | |||
6448 | ||||
6449 | // fold (select Cond, 0, 1) -> (xor Cond, 1) | |||
6450 | // We can't do this reliably if integer based booleans have different contents | |||
6451 | // to floating point based booleans. This is because we can't tell whether we | |||
6452 | // have an integer-based boolean or a floating-point-based boolean unless we | |||
6453 | // can find the SETCC that produced it and inspect its operands. This is | |||
6454 | // fairly easy if C is the SETCC node, but it can potentially be | |||
6455 | // undiscoverable (or not reasonably discoverable). For example, it could be | |||
6456 | // in another basic block or it could require searching a complicated | |||
6457 | // expression. | |||
6458 | if (CondVT.isInteger() && | |||
6459 | TLI.getBooleanContents(false, true) == | |||
6460 | TargetLowering::ZeroOrOneBooleanContent && | |||
6461 | TLI.getBooleanContents(false, false) == | |||
6462 | TargetLowering::ZeroOrOneBooleanContent && | |||
6463 | C1->isNullValue() && C2->isOne()) { | |||
6464 | SDValue NotCond = | |||
6465 | DAG.getNode(ISD::XOR, DL, CondVT, Cond, DAG.getConstant(1, DL, CondVT)); | |||
6466 | if (VT.bitsEq(CondVT)) | |||
6467 | return NotCond; | |||
6468 | return DAG.getZExtOrTrunc(NotCond, DL, VT); | |||
6469 | } | |||
6470 | ||||
6471 | return SDValue(); | |||
6472 | } | |||
6473 | ||||
6474 | SDValue DAGCombiner::visitSELECT(SDNode *N) { | |||
6475 | SDValue N0 = N->getOperand(0); | |||
6476 | SDValue N1 = N->getOperand(1); | |||
6477 | SDValue N2 = N->getOperand(2); | |||
6478 | EVT VT = N->getValueType(0); | |||
6479 | EVT VT0 = N0.getValueType(); | |||
6480 | SDLoc DL(N); | |||
6481 | ||||
6482 | // fold (select C, X, X) -> X | |||
6483 | if (N1 == N2) | |||
6484 | return N1; | |||
6485 | ||||
6486 | if (const ConstantSDNode *N0C = dyn_cast<const ConstantSDNode>(N0)) { | |||
6487 | // fold (select true, X, Y) -> X | |||
6488 | // fold (select false, X, Y) -> Y | |||
6489 | return !N0C->isNullValue() ? N1 : N2; | |||
6490 | } | |||
6491 | ||||
6492 | // fold (select X, X, Y) -> (or X, Y) | |||
6493 | // fold (select X, 1, Y) -> (or C, Y) | |||
6494 | if (VT == VT0 && VT == MVT::i1 && (N0 == N1 || isOneConstant(N1))) | |||
6495 | return DAG.getNode(ISD::OR, DL, VT, N0, N2); | |||
6496 | ||||
6497 | if (SDValue V = foldSelectOfConstants(N)) | |||
6498 | return V; | |||
6499 | ||||
6500 | // fold (select C, 0, X) -> (and (not C), X) | |||
6501 | if (VT == VT0 && VT == MVT::i1 && isNullConstant(N1)) { | |||
6502 | SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT); | |||
6503 | AddToWorklist(NOTNode.getNode()); | |||
6504 | return DAG.getNode(ISD::AND, DL, VT, NOTNode, N2); | |||
6505 | } | |||
6506 | // fold (select C, X, 1) -> (or (not C), X) | |||
6507 | if (VT == VT0 && VT == MVT::i1 && isOneConstant(N2)) { | |||
6508 | SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT); | |||
6509 | AddToWorklist(NOTNode.getNode()); | |||
6510 | return DAG.getNode(ISD::OR, DL, VT, NOTNode, N1); | |||
6511 | } | |||
6512 | // fold (select X, Y, X) -> (and X, Y) | |||
6513 | // fold (select X, Y, 0) -> (and X, Y) | |||
6514 | if (VT == VT0 && VT == MVT::i1 && (N0 == N2 || isNullConstant(N2))) | |||
6515 | return DAG.getNode(ISD::AND, DL, VT, N0, N1); | |||
6516 | ||||
6517 | // If we can fold this based on the true/false value, do so. | |||
6518 | if (SimplifySelectOps(N, N1, N2)) | |||
6519 | return SDValue(N, 0); // Don't revisit N. | |||
6520 | ||||
6521 | if (VT0 == MVT::i1) { | |||
6522 | // The code in this block deals with the following 2 equivalences: | |||
6523 | // select(C0|C1, x, y) <=> select(C0, x, select(C1, x, y)) | |||
6524 | // select(C0&C1, x, y) <=> select(C0, select(C1, x, y), y) | |||
6525 | // The target can specify its preferred form with the | |||
6526 | // shouldNormalizeToSelectSequence() callback. However we always transform | |||
6527 | // to the right anyway if we find the inner select exists in the DAG anyway | |||
6528 | // and we always transform to the left side if we know that we can further | |||
6529 | // optimize the combination of the conditions. | |||
6530 | bool normalizeToSequence = | |||
6531 | TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT); | |||
6532 | // select (and Cond0, Cond1), X, Y | |||
6533 | // -> select Cond0, (select Cond1, X, Y), Y | |||
6534 | if (N0->getOpcode() == ISD::AND && N0->hasOneUse()) { | |||
6535 | SDValue Cond0 = N0->getOperand(0); | |||
6536 | SDValue Cond1 = N0->getOperand(1); | |||
6537 | SDValue InnerSelect = | |||
6538 | DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Cond1, N1, N2); | |||
6539 | if (normalizeToSequence || !InnerSelect.use_empty()) | |||
6540 | return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Cond0, | |||
6541 | InnerSelect, N2); | |||
6542 | } | |||
6543 | // select (or Cond0, Cond1), X, Y -> select Cond0, X, (select Cond1, X, Y) | |||
6544 | if (N0->getOpcode() == ISD::OR && N0->hasOneUse()) { | |||
6545 | SDValue Cond0 = N0->getOperand(0); | |||
6546 | SDValue Cond1 = N0->getOperand(1); | |||
6547 | SDValue InnerSelect = | |||
6548 | DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Cond1, N1, N2); | |||
6549 | if (normalizeToSequence || !InnerSelect.use_empty()) | |||
6550 | return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Cond0, N1, | |||
6551 | InnerSelect); | |||
6552 | } | |||
6553 | ||||
6554 | // select Cond0, (select Cond1, X, Y), Y -> select (and Cond0, Cond1), X, Y | |||
6555 | if (N1->getOpcode() == ISD::SELECT && N1->hasOneUse()) { | |||
6556 | SDValue N1_0 = N1->getOperand(0); | |||
6557 | SDValue N1_1 = N1->getOperand(1); | |||
6558 | SDValue N1_2 = N1->getOperand(2); | |||
6559 | if (N1_2 == N2 && N0.getValueType() == N1_0.getValueType()) { | |||
6560 | // Create the actual and node if we can generate good code for it. | |||
6561 | if (!normalizeToSequence) { | |||
6562 | SDValue And = DAG.getNode(ISD::AND, DL, N0.getValueType(), N0, N1_0); | |||
6563 | return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), And, N1_1, N2); | |||
6564 | } | |||
6565 | // Otherwise see if we can optimize the "and" to a better pattern. | |||
6566 | if (SDValue Combined = visitANDLike(N0, N1_0, N)) | |||
6567 | return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Combined, N1_1, | |||
6568 | N2); | |||
6569 | } | |||
6570 | } | |||
6571 | // select Cond0, X, (select Cond1, X, Y) -> select (or Cond0, Cond1), X, Y | |||
6572 | if (N2->getOpcode() == ISD::SELECT && N2->hasOneUse()) { | |||
6573 | SDValue N2_0 = N2->getOperand(0); | |||
6574 | SDValue N2_1 = N2->getOperand(1); | |||
6575 | SDValue N2_2 = N2->getOperand(2); | |||
6576 | if (N2_1 == N1 && N0.getValueType() == N2_0.getValueType()) { | |||
6577 | // Create the actual or node if we can generate good code for it. | |||
6578 | if (!normalizeToSequence) { | |||
6579 | SDValue Or = DAG.getNode(ISD::OR, DL, N0.getValueType(), N0, N2_0); | |||
6580 | return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Or, N1, N2_2); | |||
6581 | } | |||
6582 | // Otherwise see if we can optimize to a better pattern. | |||
6583 | if (SDValue Combined = visitORLike(N0, N2_0, N)) | |||
6584 | return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Combined, N1, | |||
6585 | N2_2); | |||
6586 | } | |||
6587 | } | |||
6588 | } | |||
6589 | ||||
6590 | // select (xor Cond, 1), X, Y -> select Cond, Y, X | |||
6591 | if (VT0 == MVT::i1) { | |||
6592 | if (N0->getOpcode() == ISD::XOR) { | |||
6593 | if (auto *C = dyn_cast<ConstantSDNode>(N0->getOperand(1))) { | |||
6594 | SDValue Cond0 = N0->getOperand(0); | |||
6595 | if (C->isOne()) | |||
6596 | return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Cond0, N2, N1); | |||
6597 | } | |||
6598 | } | |||
6599 | } | |||
6600 | ||||
6601 | // fold selects based on a setcc into other things, such as min/max/abs | |||
6602 | if (N0.getOpcode() == ISD::SETCC) { | |||
6603 | // select x, y (fcmp lt x, y) -> fminnum x, y | |||
6604 | // select x, y (fcmp gt x, y) -> fmaxnum x, y | |||
6605 | // | |||
6606 | // This is OK if we don't care about what happens if either operand is a | |||
6607 | // NaN. | |||
6608 | // | |||
6609 | ||||
6610 | // FIXME: Instead of testing for UnsafeFPMath, this should be checking for | |||
6611 | // no signed zeros as well as no nans. | |||
6612 | const TargetOptions &Options = DAG.getTarget().Options; | |||
6613 | if (Options.UnsafeFPMath && VT.isFloatingPoint() && N0.hasOneUse() && | |||
6614 | DAG.isKnownNeverNaN(N1) && DAG.isKnownNeverNaN(N2)) { | |||
6615 | ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); | |||
6616 | ||||
6617 | if (SDValue FMinMax = combineMinNumMaxNum( | |||
6618 | DL, VT, N0.getOperand(0), N0.getOperand(1), N1, N2, CC, TLI, DAG)) | |||
6619 | return FMinMax; | |||
6620 | } | |||
6621 | ||||
6622 | if ((!LegalOperations && | |||
6623 | TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) || | |||
6624 | TLI.isOperationLegal(ISD::SELECT_CC, VT)) | |||
6625 | return DAG.getNode(ISD::SELECT_CC, DL, VT, N0.getOperand(0), | |||
6626 | N0.getOperand(1), N1, N2, N0.getOperand(2)); | |||
6627 | return SimplifySelect(DL, N0, N1, N2); | |||
6628 | } | |||
6629 | ||||
6630 | return SDValue(); | |||
6631 | } | |||
6632 | ||||
6633 | static | |||
6634 | std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) { | |||
6635 | SDLoc DL(N); | |||
6636 | EVT LoVT, HiVT; | |||
6637 | std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0)); | |||
6638 | ||||
6639 | // Split the inputs. | |||
6640 | SDValue Lo, Hi, LL, LH, RL, RH; | |||
6641 | std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0); | |||
6642 | std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1); | |||
6643 | ||||
6644 | Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2)); | |||
6645 | Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2)); | |||
6646 | ||||
6647 | return std::make_pair(Lo, Hi); | |||
6648 | } | |||
6649 | ||||
6650 | // This function assumes all the vselect's arguments are CONCAT_VECTOR | |||
6651 | // nodes and that the condition is a BV of ConstantSDNodes (or undefs). | |||
6652 | static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) { | |||
6653 | SDLoc DL(N); | |||
6654 | SDValue Cond = N->getOperand(0); | |||
6655 | SDValue LHS = N->getOperand(1); | |||
6656 | SDValue RHS = N->getOperand(2); | |||
6657 | EVT VT = N->getValueType(0); | |||
6658 | int NumElems = VT.getVectorNumElements(); | |||
6659 | assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&(static_cast <bool> (LHS.getOpcode() == ISD::CONCAT_VECTORS && RHS.getOpcode() == ISD::CONCAT_VECTORS && Cond.getOpcode() == ISD::BUILD_VECTOR) ? void (0) : __assert_fail ("LHS.getOpcode() == ISD::CONCAT_VECTORS && RHS.getOpcode() == ISD::CONCAT_VECTORS && Cond.getOpcode() == ISD::BUILD_VECTOR" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 6661, __extension__ __PRETTY_FUNCTION__)) | |||
6660 | RHS.getOpcode() == ISD::CONCAT_VECTORS &&(static_cast <bool> (LHS.getOpcode() == ISD::CONCAT_VECTORS && RHS.getOpcode() == ISD::CONCAT_VECTORS && Cond.getOpcode() == ISD::BUILD_VECTOR) ? void (0) : __assert_fail ("LHS.getOpcode() == ISD::CONCAT_VECTORS && RHS.getOpcode() == ISD::CONCAT_VECTORS && Cond.getOpcode() == ISD::BUILD_VECTOR" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 6661, __extension__ __PRETTY_FUNCTION__)) | |||
6661 | Cond.getOpcode() == ISD::BUILD_VECTOR)(static_cast <bool> (LHS.getOpcode() == ISD::CONCAT_VECTORS && RHS.getOpcode() == ISD::CONCAT_VECTORS && Cond.getOpcode() == ISD::BUILD_VECTOR) ? void (0) : __assert_fail ("LHS.getOpcode() == ISD::CONCAT_VECTORS && RHS.getOpcode() == ISD::CONCAT_VECTORS && Cond.getOpcode() == ISD::BUILD_VECTOR" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 6661, __extension__ __PRETTY_FUNCTION__)); | |||
6662 | ||||
6663 | // CONCAT_VECTOR can take an arbitrary number of arguments. We only care about | |||
6664 | // binary ones here. | |||
6665 | if (LHS->getNumOperands() != 2 || RHS->getNumOperands() != 2) | |||
6666 | return SDValue(); | |||
6667 | ||||
6668 | // We're sure we have an even number of elements due to the | |||
6669 | // concat_vectors we have as arguments to vselect. | |||
6670 | // Skip BV elements until we find one that's not an UNDEF | |||
6671 | // After we find an UNDEF element, keep looping until we get to half the | |||
6672 | // length of the BV and see if all the non-undef nodes are the same. | |||
6673 | ConstantSDNode *BottomHalf = nullptr; | |||
6674 | for (int i = 0; i < NumElems / 2; ++i) { | |||
6675 | if (Cond->getOperand(i)->isUndef()) | |||
6676 | continue; | |||
6677 | ||||
6678 | if (BottomHalf == nullptr) | |||
6679 | BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i)); | |||
6680 | else if (Cond->getOperand(i).getNode() != BottomHalf) | |||
6681 | return SDValue(); | |||
6682 | } | |||
6683 | ||||
6684 | // Do the same for the second half of the BuildVector | |||
6685 | ConstantSDNode *TopHalf = nullptr; | |||
6686 | for (int i = NumElems / 2; i < NumElems; ++i) { | |||
6687 | if (Cond->getOperand(i)->isUndef()) | |||
6688 | continue; | |||
6689 | ||||
6690 | if (TopHalf == nullptr) | |||
6691 | TopHalf = cast<ConstantSDNode>(Cond.getOperand(i)); | |||
6692 | else if (Cond->getOperand(i).getNode() != TopHalf) | |||
6693 | return SDValue(); | |||
6694 | } | |||
6695 | ||||
6696 | assert(TopHalf && BottomHalf &&(static_cast <bool> (TopHalf && BottomHalf && "One half of the selector was all UNDEFs and the other was all the " "same value. This should have been addressed before this function." ) ? void (0) : __assert_fail ("TopHalf && BottomHalf && \"One half of the selector was all UNDEFs and the other was all the \" \"same value. This should have been addressed before this function.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 6698, __extension__ __PRETTY_FUNCTION__)) | |||
6697 | "One half of the selector was all UNDEFs and the other was all the "(static_cast <bool> (TopHalf && BottomHalf && "One half of the selector was all UNDEFs and the other was all the " "same value. This should have been addressed before this function." ) ? void (0) : __assert_fail ("TopHalf && BottomHalf && \"One half of the selector was all UNDEFs and the other was all the \" \"same value. This should have been addressed before this function.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 6698, __extension__ __PRETTY_FUNCTION__)) | |||
6698 | "same value. This should have been addressed before this function.")(static_cast <bool> (TopHalf && BottomHalf && "One half of the selector was all UNDEFs and the other was all the " "same value. This should have been addressed before this function." ) ? void (0) : __assert_fail ("TopHalf && BottomHalf && \"One half of the selector was all UNDEFs and the other was all the \" \"same value. This should have been addressed before this function.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 6698, __extension__ __PRETTY_FUNCTION__)); | |||
6699 | return DAG.getNode( | |||
6700 | ISD::CONCAT_VECTORS, DL, VT, | |||
6701 | BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0), | |||
6702 | TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1)); | |||
6703 | } | |||
6704 | ||||
6705 | SDValue DAGCombiner::visitMSCATTER(SDNode *N) { | |||
6706 | if (Level >= AfterLegalizeTypes) | |||
6707 | return SDValue(); | |||
6708 | ||||
6709 | MaskedScatterSDNode *MSC = cast<MaskedScatterSDNode>(N); | |||
6710 | SDValue Mask = MSC->getMask(); | |||
6711 | SDValue Data = MSC->getValue(); | |||
6712 | SDLoc DL(N); | |||
6713 | ||||
6714 | // If the MSCATTER data type requires splitting and the mask is provided by a | |||
6715 | // SETCC, then split both nodes and its operands before legalization. This | |||
6716 | // prevents the type legalizer from unrolling SETCC into scalar comparisons | |||
6717 | // and enables future optimizations (e.g. min/max pattern matching on X86). | |||
6718 | if (Mask.getOpcode() != ISD::SETCC) | |||
6719 | return SDValue(); | |||
6720 | ||||
6721 | // Check if any splitting is required. | |||
6722 | if (TLI.getTypeAction(*DAG.getContext(), Data.getValueType()) != | |||
6723 | TargetLowering::TypeSplitVector) | |||
6724 | return SDValue(); | |||
6725 | SDValue MaskLo, MaskHi, Lo, Hi; | |||
6726 | std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG); | |||
6727 | ||||
6728 | EVT LoVT, HiVT; | |||
6729 | std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MSC->getValueType(0)); | |||
6730 | ||||
6731 | SDValue Chain = MSC->getChain(); | |||
6732 | ||||
6733 | EVT MemoryVT = MSC->getMemoryVT(); | |||
6734 | unsigned Alignment = MSC->getOriginalAlignment(); | |||
6735 | ||||
6736 | EVT LoMemVT, HiMemVT; | |||
6737 | std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT); | |||
6738 | ||||
6739 | SDValue DataLo, DataHi; | |||
6740 | std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL); | |||
6741 | ||||
6742 | SDValue Scale = MSC->getScale(); | |||
6743 | SDValue BasePtr = MSC->getBasePtr(); | |||
6744 | SDValue IndexLo, IndexHi; | |||
6745 | std::tie(IndexLo, IndexHi) = DAG.SplitVector(MSC->getIndex(), DL); | |||
6746 | ||||
6747 | MachineMemOperand *MMO = DAG.getMachineFunction(). | |||
6748 | getMachineMemOperand(MSC->getPointerInfo(), | |||
6749 | MachineMemOperand::MOStore, LoMemVT.getStoreSize(), | |||
6750 | Alignment, MSC->getAAInfo(), MSC->getRanges()); | |||
6751 | ||||
6752 | SDValue OpsLo[] = { Chain, DataLo, MaskLo, BasePtr, IndexLo, Scale }; | |||
6753 | Lo = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataLo.getValueType(), | |||
6754 | DL, OpsLo, MMO); | |||
6755 | ||||
6756 | SDValue OpsHi[] = { Chain, DataHi, MaskHi, BasePtr, IndexHi, Scale }; | |||
6757 | Hi = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataHi.getValueType(), | |||
6758 | DL, OpsHi, MMO); | |||
6759 | ||||
6760 | AddToWorklist(Lo.getNode()); | |||
6761 | AddToWorklist(Hi.getNode()); | |||
6762 | ||||
6763 | return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi); | |||
6764 | } | |||
6765 | ||||
6766 | SDValue DAGCombiner::visitMSTORE(SDNode *N) { | |||
6767 | if (Level >= AfterLegalizeTypes) | |||
6768 | return SDValue(); | |||
6769 | ||||
6770 | MaskedStoreSDNode *MST = dyn_cast<MaskedStoreSDNode>(N); | |||
6771 | SDValue Mask = MST->getMask(); | |||
6772 | SDValue Data = MST->getValue(); | |||
6773 | EVT VT = Data.getValueType(); | |||
6774 | SDLoc DL(N); | |||
6775 | ||||
6776 | // If the MSTORE data type requires splitting and the mask is provided by a | |||
6777 | // SETCC, then split both nodes and its operands before legalization. This | |||
6778 | // prevents the type legalizer from unrolling SETCC into scalar comparisons | |||
6779 | // and enables future optimizations (e.g. min/max pattern matching on X86). | |||
6780 | if (Mask.getOpcode() == ISD::SETCC) { | |||
6781 | // Check if any splitting is required. | |||
6782 | if (TLI.getTypeAction(*DAG.getContext(), VT) != | |||
6783 | TargetLowering::TypeSplitVector) | |||
6784 | return SDValue(); | |||
6785 | ||||
6786 | SDValue MaskLo, MaskHi, Lo, Hi; | |||
6787 | std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG); | |||
6788 | ||||
6789 | SDValue Chain = MST->getChain(); | |||
6790 | SDValue Ptr = MST->getBasePtr(); | |||
6791 | ||||
6792 | EVT MemoryVT = MST->getMemoryVT(); | |||
6793 | unsigned Alignment = MST->getOriginalAlignment(); | |||
6794 | ||||
6795 | // if Alignment is equal to the vector size, | |||
6796 | // take the half of it for the second part | |||
6797 | unsigned SecondHalfAlignment = | |||
6798 | (Alignment == VT.getSizeInBits() / 8) ? Alignment / 2 : Alignment; | |||
6799 | ||||
6800 | EVT LoMemVT, HiMemVT; | |||
6801 | std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT); | |||
6802 | ||||
6803 | SDValue DataLo, DataHi; | |||
6804 | std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL); | |||
6805 | ||||
6806 | MachineMemOperand *MMO = DAG.getMachineFunction(). | |||
6807 | getMachineMemOperand(MST->getPointerInfo(), | |||
6808 | MachineMemOperand::MOStore, LoMemVT.getStoreSize(), | |||
6809 | Alignment, MST->getAAInfo(), MST->getRanges()); | |||
6810 | ||||
6811 | Lo = DAG.getMaskedStore(Chain, DL, DataLo, Ptr, MaskLo, LoMemVT, MMO, | |||
6812 | MST->isTruncatingStore(), | |||
6813 | MST->isCompressingStore()); | |||
6814 | ||||
6815 | Ptr = TLI.IncrementMemoryAddress(Ptr, MaskLo, DL, LoMemVT, DAG, | |||
6816 | MST->isCompressingStore()); | |||
6817 | unsigned HiOffset = LoMemVT.getStoreSize(); | |||
6818 | ||||
6819 | MMO = DAG.getMachineFunction().getMachineMemOperand( | |||
6820 | MST->getPointerInfo().getWithOffset(HiOffset), | |||
6821 | MachineMemOperand::MOStore, HiMemVT.getStoreSize(), SecondHalfAlignment, | |||
6822 | MST->getAAInfo(), MST->getRanges()); | |||
6823 | ||||
6824 | Hi = DAG.getMaskedStore(Chain, DL, DataHi, Ptr, MaskHi, HiMemVT, MMO, | |||
6825 | MST->isTruncatingStore(), | |||
6826 | MST->isCompressingStore()); | |||
6827 | ||||
6828 | AddToWorklist(Lo.getNode()); | |||
6829 | AddToWorklist(Hi.getNode()); | |||
6830 | ||||
6831 | return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi); | |||
6832 | } | |||
6833 | return SDValue(); | |||
6834 | } | |||
6835 | ||||
6836 | SDValue DAGCombiner::visitMGATHER(SDNode *N) { | |||
6837 | if (Level >= AfterLegalizeTypes) | |||
6838 | return SDValue(); | |||
6839 | ||||
6840 | MaskedGatherSDNode *MGT = cast<MaskedGatherSDNode>(N); | |||
6841 | SDValue Mask = MGT->getMask(); | |||
6842 | SDLoc DL(N); | |||
6843 | ||||
6844 | // If the MGATHER result requires splitting and the mask is provided by a | |||
6845 | // SETCC, then split both nodes and its operands before legalization. This | |||
6846 | // prevents the type legalizer from unrolling SETCC into scalar comparisons | |||
6847 | // and enables future optimizations (e.g. min/max pattern matching on X86). | |||
6848 | ||||
6849 | if (Mask.getOpcode() != ISD::SETCC) | |||
6850 | return SDValue(); | |||
6851 | ||||
6852 | EVT VT = N->getValueType(0); | |||
6853 | ||||
6854 | // Check if any splitting is required. | |||
6855 | if (TLI.getTypeAction(*DAG.getContext(), VT) != | |||
6856 | TargetLowering::TypeSplitVector) | |||
6857 | return SDValue(); | |||
6858 | ||||
6859 | SDValue MaskLo, MaskHi, Lo, Hi; | |||
6860 | std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG); | |||
6861 | ||||
6862 | SDValue Src0 = MGT->getValue(); | |||
6863 | SDValue Src0Lo, Src0Hi; | |||
6864 | std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL); | |||
6865 | ||||
6866 | EVT LoVT, HiVT; | |||
6867 | std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT); | |||
6868 | ||||
6869 | SDValue Chain = MGT->getChain(); | |||
6870 | EVT MemoryVT = MGT->getMemoryVT(); | |||
6871 | unsigned Alignment = MGT->getOriginalAlignment(); | |||
6872 | ||||
6873 | EVT LoMemVT, HiMemVT; | |||
6874 | std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT); | |||
6875 | ||||
6876 | SDValue Scale = MGT->getScale(); | |||
6877 | SDValue BasePtr = MGT->getBasePtr(); | |||
6878 | SDValue Index = MGT->getIndex(); | |||
6879 | SDValue IndexLo, IndexHi; | |||
6880 | std::tie(IndexLo, IndexHi) = DAG.SplitVector(Index, DL); | |||
6881 | ||||
6882 | MachineMemOperand *MMO = DAG.getMachineFunction(). | |||
6883 | getMachineMemOperand(MGT->getPointerInfo(), | |||
6884 | MachineMemOperand::MOLoad, LoMemVT.getStoreSize(), | |||
6885 | Alignment, MGT->getAAInfo(), MGT->getRanges()); | |||
6886 | ||||
6887 | SDValue OpsLo[] = { Chain, Src0Lo, MaskLo, BasePtr, IndexLo, Scale }; | |||
6888 | Lo = DAG.getMaskedGather(DAG.getVTList(LoVT, MVT::Other), LoVT, DL, OpsLo, | |||
6889 | MMO); | |||
6890 | ||||
6891 | SDValue OpsHi[] = { Chain, Src0Hi, MaskHi, BasePtr, IndexHi, Scale }; | |||
6892 | Hi = DAG.getMaskedGather(DAG.getVTList(HiVT, MVT::Other), HiVT, DL, OpsHi, | |||
6893 | MMO); | |||
6894 | ||||
6895 | AddToWorklist(Lo.getNode()); | |||
6896 | AddToWorklist(Hi.getNode()); | |||
6897 | ||||
6898 | // Build a factor node to remember that this load is independent of the | |||
6899 | // other one. | |||
6900 | Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1), | |||
6901 | Hi.getValue(1)); | |||
6902 | ||||
6903 | // Legalized the chain result - switch anything that used the old chain to | |||
6904 | // use the new one. | |||
6905 | DAG.ReplaceAllUsesOfValueWith(SDValue(MGT, 1), Chain); | |||
6906 | ||||
6907 | SDValue GatherRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi); | |||
6908 | ||||
6909 | SDValue RetOps[] = { GatherRes, Chain }; | |||
6910 | return DAG.getMergeValues(RetOps, DL); | |||
6911 | } | |||
6912 | ||||
6913 | SDValue DAGCombiner::visitMLOAD(SDNode *N) { | |||
6914 | if (Level >= AfterLegalizeTypes) | |||
6915 | return SDValue(); | |||
6916 | ||||
6917 | MaskedLoadSDNode *MLD = dyn_cast<MaskedLoadSDNode>(N); | |||
6918 | SDValue Mask = MLD->getMask(); | |||
6919 | SDLoc DL(N); | |||
6920 | ||||
6921 | // If the MLOAD result requires splitting and the mask is provided by a | |||
6922 | // SETCC, then split both nodes and its operands before legalization. This | |||
6923 | // prevents the type legalizer from unrolling SETCC into scalar comparisons | |||
6924 | // and enables future optimizations (e.g. min/max pattern matching on X86). | |||
6925 | if (Mask.getOpcode() == ISD::SETCC) { | |||
6926 | EVT VT = N->getValueType(0); | |||
6927 | ||||
6928 | // Check if any splitting is required. | |||
6929 | if (TLI.getTypeAction(*DAG.getContext(), VT) != | |||
6930 | TargetLowering::TypeSplitVector) | |||
6931 | return SDValue(); | |||
6932 | ||||
6933 | SDValue MaskLo, MaskHi, Lo, Hi; | |||
6934 | std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG); | |||
6935 | ||||
6936 | SDValue Src0 = MLD->getSrc0(); | |||
6937 | SDValue Src0Lo, Src0Hi; | |||
6938 | std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL); | |||
6939 | ||||
6940 | EVT LoVT, HiVT; | |||
6941 | std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MLD->getValueType(0)); | |||
6942 | ||||
6943 | SDValue Chain = MLD->getChain(); | |||
6944 | SDValue Ptr = MLD->getBasePtr(); | |||
6945 | EVT MemoryVT = MLD->getMemoryVT(); | |||
6946 | unsigned Alignment = MLD->getOriginalAlignment(); | |||
6947 | ||||
6948 | // if Alignment is equal to the vector size, | |||
6949 | // take the half of it for the second part | |||
6950 | unsigned SecondHalfAlignment = | |||
6951 | (Alignment == MLD->getValueType(0).getSizeInBits()/8) ? | |||
6952 | Alignment/2 : Alignment; | |||
6953 | ||||
6954 | EVT LoMemVT, HiMemVT; | |||
6955 | std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT); | |||
6956 | ||||
6957 | MachineMemOperand *MMO = DAG.getMachineFunction(). | |||
6958 | getMachineMemOperand(MLD->getPointerInfo(), | |||
6959 | MachineMemOperand::MOLoad, LoMemVT.getStoreSize(), | |||
6960 | Alignment, MLD->getAAInfo(), MLD->getRanges()); | |||
6961 | ||||
6962 | Lo = DAG.getMaskedLoad(LoVT, DL, Chain, Ptr, MaskLo, Src0Lo, LoMemVT, MMO, | |||
6963 | ISD::NON_EXTLOAD, MLD->isExpandingLoad()); | |||
6964 | ||||
6965 | Ptr = TLI.IncrementMemoryAddress(Ptr, MaskLo, DL, LoMemVT, DAG, | |||
6966 | MLD->isExpandingLoad()); | |||
6967 | unsigned HiOffset = LoMemVT.getStoreSize(); | |||
6968 | ||||
6969 | MMO = DAG.getMachineFunction().getMachineMemOperand( | |||
6970 | MLD->getPointerInfo().getWithOffset(HiOffset), | |||
6971 | MachineMemOperand::MOLoad, HiMemVT.getStoreSize(), SecondHalfAlignment, | |||
6972 | MLD->getAAInfo(), MLD->getRanges()); | |||
6973 | ||||
6974 | Hi = DAG.getMaskedLoad(HiVT, DL, Chain, Ptr, MaskHi, Src0Hi, HiMemVT, MMO, | |||
6975 | ISD::NON_EXTLOAD, MLD->isExpandingLoad()); | |||
6976 | ||||
6977 | AddToWorklist(Lo.getNode()); | |||
6978 | AddToWorklist(Hi.getNode()); | |||
6979 | ||||
6980 | // Build a factor node to remember that this load is independent of the | |||
6981 | // other one. | |||
6982 | Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1), | |||
6983 | Hi.getValue(1)); | |||
6984 | ||||
6985 | // Legalized the chain result - switch anything that used the old chain to | |||
6986 | // use the new one. | |||
6987 | DAG.ReplaceAllUsesOfValueWith(SDValue(MLD, 1), Chain); | |||
6988 | ||||
6989 | SDValue LoadRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi); | |||
6990 | ||||
6991 | SDValue RetOps[] = { LoadRes, Chain }; | |||
6992 | return DAG.getMergeValues(RetOps, DL); | |||
6993 | } | |||
6994 | return SDValue(); | |||
6995 | } | |||
6996 | ||||
6997 | /// A vector select of 2 constant vectors can be simplified to math/logic to | |||
6998 | /// avoid a variable select instruction and possibly avoid constant loads. | |||
6999 | SDValue DAGCombiner::foldVSelectOfConstants(SDNode *N) { | |||
7000 | SDValue Cond = N->getOperand(0); | |||
7001 | SDValue N1 = N->getOperand(1); | |||
7002 | SDValue N2 = N->getOperand(2); | |||
7003 | EVT VT = N->getValueType(0); | |||
7004 | if (!Cond.hasOneUse() || Cond.getScalarValueSizeInBits() != 1 || | |||
7005 | !TLI.convertSelectOfConstantsToMath(VT) || | |||
7006 | !ISD::isBuildVectorOfConstantSDNodes(N1.getNode()) || | |||
7007 | !ISD::isBuildVectorOfConstantSDNodes(N2.getNode())) | |||
7008 | return SDValue(); | |||
7009 | ||||
7010 | // Check if we can use the condition value to increment/decrement a single | |||
7011 | // constant value. This simplifies a select to an add and removes a constant | |||
7012 | // load/materialization from the general case. | |||
7013 | bool AllAddOne = true; | |||
7014 | bool AllSubOne = true; | |||
7015 | unsigned Elts = VT.getVectorNumElements(); | |||
7016 | for (unsigned i = 0; i != Elts; ++i) { | |||
7017 | SDValue N1Elt = N1.getOperand(i); | |||
7018 | SDValue N2Elt = N2.getOperand(i); | |||
7019 | if (N1Elt.isUndef() || N2Elt.isUndef()) | |||
7020 | continue; | |||
7021 | ||||
7022 | const APInt &C1 = cast<ConstantSDNode>(N1Elt)->getAPIntValue(); | |||
7023 | const APInt &C2 = cast<ConstantSDNode>(N2Elt)->getAPIntValue(); | |||
7024 | if (C1 != C2 + 1) | |||
7025 | AllAddOne = false; | |||
7026 | if (C1 != C2 - 1) | |||
7027 | AllSubOne = false; | |||
7028 | } | |||
7029 | ||||
7030 | // Further simplifications for the extra-special cases where the constants are | |||
7031 | // all 0 or all -1 should be implemented as folds of these patterns. | |||
7032 | SDLoc DL(N); | |||
7033 | if (AllAddOne || AllSubOne) { | |||
7034 | // vselect <N x i1> Cond, C+1, C --> add (zext Cond), C | |||
7035 | // vselect <N x i1> Cond, C-1, C --> add (sext Cond), C | |||
7036 | auto ExtendOpcode = AllAddOne ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND; | |||
7037 | SDValue ExtendedCond = DAG.getNode(ExtendOpcode, DL, VT, Cond); | |||
7038 | return DAG.getNode(ISD::ADD, DL, VT, ExtendedCond, N2); | |||
7039 | } | |||
7040 | ||||
7041 | // The general case for select-of-constants: | |||
7042 | // vselect <N x i1> Cond, C1, C2 --> xor (and (sext Cond), (C1^C2)), C2 | |||
7043 | // ...but that only makes sense if a vselect is slower than 2 logic ops, so | |||
7044 | // leave that to a machine-specific pass. | |||
7045 | return SDValue(); | |||
7046 | } | |||
7047 | ||||
7048 | SDValue DAGCombiner::visitVSELECT(SDNode *N) { | |||
7049 | SDValue N0 = N->getOperand(0); | |||
7050 | SDValue N1 = N->getOperand(1); | |||
7051 | SDValue N2 = N->getOperand(2); | |||
7052 | SDLoc DL(N); | |||
7053 | ||||
7054 | // fold (vselect C, X, X) -> X | |||
7055 | if (N1 == N2) | |||
7056 | return N1; | |||
7057 | ||||
7058 | // Canonicalize integer abs. | |||
7059 | // vselect (setg[te] X, 0), X, -X -> | |||
7060 | // vselect (setgt X, -1), X, -X -> | |||
7061 | // vselect (setl[te] X, 0), -X, X -> | |||
7062 | // Y = sra (X, size(X)-1); xor (add (X, Y), Y) | |||
7063 | if (N0.getOpcode() == ISD::SETCC) { | |||
7064 | SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1); | |||
7065 | ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); | |||
7066 | bool isAbs = false; | |||
7067 | bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode()); | |||
7068 | ||||
7069 | if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) || | |||
7070 | (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) && | |||
7071 | N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1)) | |||
7072 | isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode()); | |||
7073 | else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) && | |||
7074 | N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1)) | |||
7075 | isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode()); | |||
7076 | ||||
7077 | if (isAbs) { | |||
7078 | EVT VT = LHS.getValueType(); | |||
7079 | if (TLI.isOperationLegalOrCustom(ISD::ABS, VT)) | |||
7080 | return DAG.getNode(ISD::ABS, DL, VT, LHS); | |||
7081 | ||||
7082 | SDValue Shift = DAG.getNode( | |||
7083 | ISD::SRA, DL, VT, LHS, | |||
7084 | DAG.getConstant(VT.getScalarSizeInBits() - 1, DL, VT)); | |||
7085 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift); | |||
7086 | AddToWorklist(Shift.getNode()); | |||
7087 | AddToWorklist(Add.getNode()); | |||
7088 | return DAG.getNode(ISD::XOR, DL, VT, Add, Shift); | |||
7089 | } | |||
7090 | } | |||
7091 | ||||
7092 | if (SimplifySelectOps(N, N1, N2)) | |||
7093 | return SDValue(N, 0); // Don't revisit N. | |||
7094 | ||||
7095 | // Fold (vselect (build_vector all_ones), N1, N2) -> N1 | |||
7096 | if (ISD::isBuildVectorAllOnes(N0.getNode())) | |||
7097 | return N1; | |||
7098 | // Fold (vselect (build_vector all_zeros), N1, N2) -> N2 | |||
7099 | if (ISD::isBuildVectorAllZeros(N0.getNode())) | |||
7100 | return N2; | |||
7101 | ||||
7102 | // The ConvertSelectToConcatVector function is assuming both the above | |||
7103 | // checks for (vselect (build_vector all{ones,zeros) ...) have been made | |||
7104 | // and addressed. | |||
7105 | if (N1.getOpcode() == ISD::CONCAT_VECTORS && | |||
7106 | N2.getOpcode() == ISD::CONCAT_VECTORS && | |||
7107 | ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) { | |||
7108 | if (SDValue CV = ConvertSelectToConcatVector(N, DAG)) | |||
7109 | return CV; | |||
7110 | } | |||
7111 | ||||
7112 | if (SDValue V = foldVSelectOfConstants(N)) | |||
7113 | return V; | |||
7114 | ||||
7115 | return SDValue(); | |||
7116 | } | |||
7117 | ||||
7118 | SDValue DAGCombiner::visitSELECT_CC(SDNode *N) { | |||
7119 | SDValue N0 = N->getOperand(0); | |||
7120 | SDValue N1 = N->getOperand(1); | |||
7121 | SDValue N2 = N->getOperand(2); | |||
7122 | SDValue N3 = N->getOperand(3); | |||
7123 | SDValue N4 = N->getOperand(4); | |||
7124 | ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get(); | |||
7125 | ||||
7126 | // fold select_cc lhs, rhs, x, x, cc -> x | |||
7127 | if (N2 == N3) | |||
7128 | return N2; | |||
7129 | ||||
7130 | // Determine if the condition we're dealing with is constant | |||
7131 | if (SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()), N0, N1, | |||
7132 | CC, SDLoc(N), false)) { | |||
7133 | AddToWorklist(SCC.getNode()); | |||
7134 | ||||
7135 | if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) { | |||
7136 | if (!SCCC->isNullValue()) | |||
7137 | return N2; // cond always true -> true val | |||
7138 | else | |||
7139 | return N3; // cond always false -> false val | |||
7140 | } else if (SCC->isUndef()) { | |||
7141 | // When the condition is UNDEF, just return the first operand. This is | |||
7142 | // coherent the DAG creation, no setcc node is created in this case | |||
7143 | return N2; | |||
7144 | } else if (SCC.getOpcode() == ISD::SETCC) { | |||
7145 | // Fold to a simpler select_cc | |||
7146 | return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(), | |||
7147 | SCC.getOperand(0), SCC.getOperand(1), N2, N3, | |||
7148 | SCC.getOperand(2)); | |||
7149 | } | |||
7150 | } | |||
7151 | ||||
7152 | // If we can fold this based on the true/false value, do so. | |||
7153 | if (SimplifySelectOps(N, N2, N3)) | |||
7154 | return SDValue(N, 0); // Don't revisit N. | |||
7155 | ||||
7156 | // fold select_cc into other things, such as min/max/abs | |||
7157 | return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC); | |||
7158 | } | |||
7159 | ||||
7160 | SDValue DAGCombiner::visitSETCC(SDNode *N) { | |||
7161 | // setcc is very commonly used as an argument to brcond. This pattern | |||
7162 | // also lend itself to numerous combines and, as a result, it is desired | |||
7163 | // we keep the argument to a brcond as a setcc as much as possible. | |||
7164 | bool PreferSetCC = | |||
7165 | N->hasOneUse() && N->use_begin()->getOpcode() == ISD::BRCOND; | |||
7166 | ||||
7167 | SDValue Combined = SimplifySetCC( | |||
7168 | N->getValueType(0), N->getOperand(0), N->getOperand(1), | |||
7169 | cast<CondCodeSDNode>(N->getOperand(2))->get(), SDLoc(N), !PreferSetCC); | |||
7170 | ||||
7171 | if (!Combined) | |||
7172 | return SDValue(); | |||
7173 | ||||
7174 | // If we prefer to have a setcc, and we don't, we'll try our best to | |||
7175 | // recreate one using rebuildSetCC. | |||
7176 | if (PreferSetCC && Combined.getOpcode() != ISD::SETCC) { | |||
7177 | SDValue NewSetCC = rebuildSetCC(Combined); | |||
7178 | ||||
7179 | // We don't have anything interesting to combine to. | |||
7180 | if (NewSetCC.getNode() == N) | |||
7181 | return SDValue(); | |||
7182 | ||||
7183 | if (NewSetCC) | |||
7184 | return NewSetCC; | |||
7185 | } | |||
7186 | ||||
7187 | return Combined; | |||
7188 | } | |||
7189 | ||||
7190 | SDValue DAGCombiner::visitSETCCE(SDNode *N) { | |||
7191 | SDValue LHS = N->getOperand(0); | |||
7192 | SDValue RHS = N->getOperand(1); | |||
7193 | SDValue Carry = N->getOperand(2); | |||
7194 | SDValue Cond = N->getOperand(3); | |||
7195 | ||||
7196 | // If Carry is false, fold to a regular SETCC. | |||
7197 | if (Carry.getOpcode() == ISD::CARRY_FALSE) | |||
7198 | return DAG.getNode(ISD::SETCC, SDLoc(N), N->getVTList(), LHS, RHS, Cond); | |||
7199 | ||||
7200 | return SDValue(); | |||
7201 | } | |||
7202 | ||||
7203 | SDValue DAGCombiner::visitSETCCCARRY(SDNode *N) { | |||
7204 | SDValue LHS = N->getOperand(0); | |||
7205 | SDValue RHS = N->getOperand(1); | |||
7206 | SDValue Carry = N->getOperand(2); | |||
7207 | SDValue Cond = N->getOperand(3); | |||
7208 | ||||
7209 | // If Carry is false, fold to a regular SETCC. | |||
7210 | if (isNullConstant(Carry)) | |||
7211 | return DAG.getNode(ISD::SETCC, SDLoc(N), N->getVTList(), LHS, RHS, Cond); | |||
7212 | ||||
7213 | return SDValue(); | |||
7214 | } | |||
7215 | ||||
7216 | /// Try to fold a sext/zext/aext dag node into a ConstantSDNode or | |||
7217 | /// a build_vector of constants. | |||
7218 | /// This function is called by the DAGCombiner when visiting sext/zext/aext | |||
7219 | /// dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND). | |||
7220 | /// Vector extends are not folded if operations are legal; this is to | |||
7221 | /// avoid introducing illegal build_vector dag nodes. | |||
7222 | static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI, | |||
7223 | SelectionDAG &DAG, bool LegalTypes, | |||
7224 | bool LegalOperations) { | |||
7225 | unsigned Opcode = N->getOpcode(); | |||
7226 | SDValue N0 = N->getOperand(0); | |||
7227 | EVT VT = N->getValueType(0); | |||
7228 | ||||
7229 | assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||(static_cast <bool> ((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND || Opcode == ISD::ANY_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG || Opcode == ISD::ZERO_EXTEND_VECTOR_INREG ) && "Expected EXTEND dag node in input!") ? void (0) : __assert_fail ("(Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND || Opcode == ISD::ANY_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG || Opcode == ISD::ZERO_EXTEND_VECTOR_INREG) && \"Expected EXTEND dag node in input!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 7232, __extension__ __PRETTY_FUNCTION__)) | |||
7230 | Opcode == ISD::ANY_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG ||(static_cast <bool> ((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND || Opcode == ISD::ANY_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG || Opcode == ISD::ZERO_EXTEND_VECTOR_INREG ) && "Expected EXTEND dag node in input!") ? void (0) : __assert_fail ("(Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND || Opcode == ISD::ANY_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG || Opcode == ISD::ZERO_EXTEND_VECTOR_INREG) && \"Expected EXTEND dag node in input!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 7232, __extension__ __PRETTY_FUNCTION__)) | |||
7231 | Opcode == ISD::ZERO_EXTEND_VECTOR_INREG)(static_cast <bool> ((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND || Opcode == ISD::ANY_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG || Opcode == ISD::ZERO_EXTEND_VECTOR_INREG ) && "Expected EXTEND dag node in input!") ? void (0) : __assert_fail ("(Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND || Opcode == ISD::ANY_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG || Opcode == ISD::ZERO_EXTEND_VECTOR_INREG) && \"Expected EXTEND dag node in input!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 7232, __extension__ __PRETTY_FUNCTION__)) | |||
7232 | && "Expected EXTEND dag node in input!")(static_cast <bool> ((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND || Opcode == ISD::ANY_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG || Opcode == ISD::ZERO_EXTEND_VECTOR_INREG ) && "Expected EXTEND dag node in input!") ? void (0) : __assert_fail ("(Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND || Opcode == ISD::ANY_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG || Opcode == ISD::ZERO_EXTEND_VECTOR_INREG) && \"Expected EXTEND dag node in input!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 7232, __extension__ __PRETTY_FUNCTION__)); | |||
7233 | ||||
7234 | // fold (sext c1) -> c1 | |||
7235 | // fold (zext c1) -> c1 | |||
7236 | // fold (aext c1) -> c1 | |||
7237 | if (isa<ConstantSDNode>(N0)) | |||
7238 | return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode(); | |||
7239 | ||||
7240 | // fold (sext (build_vector AllConstants) -> (build_vector AllConstants) | |||
7241 | // fold (zext (build_vector AllConstants) -> (build_vector AllConstants) | |||
7242 | // fold (aext (build_vector AllConstants) -> (build_vector AllConstants) | |||
7243 | EVT SVT = VT.getScalarType(); | |||
7244 | if (!(VT.isVector() && | |||
7245 | (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) && | |||
7246 | ISD::isBuildVectorOfConstantSDNodes(N0.getNode()))) | |||
7247 | return nullptr; | |||
7248 | ||||
7249 | // We can fold this node into a build_vector. | |||
7250 | unsigned VTBits = SVT.getSizeInBits(); | |||
7251 | unsigned EVTBits = N0->getValueType(0).getScalarSizeInBits(); | |||
7252 | SmallVector<SDValue, 8> Elts; | |||
7253 | unsigned NumElts = VT.getVectorNumElements(); | |||
7254 | SDLoc DL(N); | |||
7255 | ||||
7256 | for (unsigned i=0; i != NumElts; ++i) { | |||
7257 | SDValue Op = N0->getOperand(i); | |||
7258 | if (Op->isUndef()) { | |||
7259 | Elts.push_back(DAG.getUNDEF(SVT)); | |||
7260 | continue; | |||
7261 | } | |||
7262 | ||||
7263 | SDLoc DL(Op); | |||
7264 | // Get the constant value and if needed trunc it to the size of the type. | |||
7265 | // Nodes like build_vector might have constants wider than the scalar type. | |||
7266 | APInt C = cast<ConstantSDNode>(Op)->getAPIntValue().zextOrTrunc(EVTBits); | |||
7267 | if (Opcode == ISD::SIGN_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG) | |||
7268 | Elts.push_back(DAG.getConstant(C.sext(VTBits), DL, SVT)); | |||
7269 | else | |||
7270 | Elts.push_back(DAG.getConstant(C.zext(VTBits), DL, SVT)); | |||
7271 | } | |||
7272 | ||||
7273 | return DAG.getBuildVector(VT, DL, Elts).getNode(); | |||
7274 | } | |||
7275 | ||||
7276 | // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this: | |||
7277 | // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))" | |||
7278 | // transformation. Returns true if extension are possible and the above | |||
7279 | // mentioned transformation is profitable. | |||
7280 | static bool ExtendUsesToFormExtLoad(EVT VT, SDNode *N, SDValue N0, | |||
7281 | unsigned ExtOpc, | |||
7282 | SmallVectorImpl<SDNode *> &ExtendNodes, | |||
7283 | const TargetLowering &TLI) { | |||
7284 | bool HasCopyToRegUses = false; | |||
7285 | bool isTruncFree = TLI.isTruncateFree(VT, N0.getValueType()); | |||
7286 | for (SDNode::use_iterator UI = N0.getNode()->use_begin(), | |||
7287 | UE = N0.getNode()->use_end(); | |||
7288 | UI != UE; ++UI) { | |||
7289 | SDNode *User = *UI; | |||
7290 | if (User == N) | |||
7291 | continue; | |||
7292 | if (UI.getUse().getResNo() != N0.getResNo()) | |||
7293 | continue; | |||
7294 | // FIXME: Only extend SETCC N, N and SETCC N, c for now. | |||
7295 | if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) { | |||
7296 | ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get(); | |||
7297 | if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC)) | |||
7298 | // Sign bits will be lost after a zext. | |||
7299 | return false; | |||
7300 | bool Add = false; | |||
7301 | for (unsigned i = 0; i != 2; ++i) { | |||
7302 | SDValue UseOp = User->getOperand(i); | |||
7303 | if (UseOp == N0) | |||
7304 | continue; | |||
7305 | if (!isa<ConstantSDNode>(UseOp)) | |||
7306 | return false; | |||
7307 | Add = true; | |||
7308 | } | |||
7309 | if (Add) | |||
7310 | ExtendNodes.push_back(User); | |||
7311 | continue; | |||
7312 | } | |||
7313 | // If truncates aren't free and there are users we can't | |||
7314 | // extend, it isn't worthwhile. | |||
7315 | if (!isTruncFree) | |||
7316 | return false; | |||
7317 | // Remember if this value is live-out. | |||
7318 | if (User->getOpcode() == ISD::CopyToReg) | |||
7319 | HasCopyToRegUses = true; | |||
7320 | } | |||
7321 | ||||
7322 | if (HasCopyToRegUses) { | |||
7323 | bool BothLiveOut = false; | |||
7324 | for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); | |||
7325 | UI != UE; ++UI) { | |||
7326 | SDUse &Use = UI.getUse(); | |||
7327 | if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) { | |||
7328 | BothLiveOut = true; | |||
7329 | break; | |||
7330 | } | |||
7331 | } | |||
7332 | if (BothLiveOut) | |||
7333 | // Both unextended and extended values are live out. There had better be | |||
7334 | // a good reason for the transformation. | |||
7335 | return ExtendNodes.size(); | |||
7336 | } | |||
7337 | return true; | |||
7338 | } | |||
7339 | ||||
7340 | void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs, | |||
7341 | SDValue OrigLoad, SDValue ExtLoad, | |||
7342 | const SDLoc &DL, ISD::NodeType ExtType) { | |||
7343 | // Extend SetCC uses if necessary. | |||
7344 | for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) { | |||
7345 | SDNode *SetCC = SetCCs[i]; | |||
7346 | SmallVector<SDValue, 4> Ops; | |||
7347 | ||||
7348 | for (unsigned j = 0; j != 2; ++j) { | |||
7349 | SDValue SOp = SetCC->getOperand(j); | |||
7350 | if (SOp == OrigLoad) | |||
7351 | Ops.push_back(ExtLoad); | |||
7352 | else | |||
7353 | Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp)); | |||
7354 | } | |||
7355 | ||||
7356 | Ops.push_back(SetCC->getOperand(2)); | |||
7357 | CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops)); | |||
7358 | } | |||
7359 | } | |||
7360 | ||||
7361 | // FIXME: Bring more similar combines here, common to sext/zext (maybe aext?). | |||
7362 | SDValue DAGCombiner::CombineExtLoad(SDNode *N) { | |||
7363 | SDValue N0 = N->getOperand(0); | |||
7364 | EVT DstVT = N->getValueType(0); | |||
7365 | EVT SrcVT = N0.getValueType(); | |||
7366 | ||||
7367 | assert((N->getOpcode() == ISD::SIGN_EXTEND ||(static_cast <bool> ((N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND) && "Unexpected node type (not an extend)!" ) ? void (0) : __assert_fail ("(N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND) && \"Unexpected node type (not an extend)!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 7369, __extension__ __PRETTY_FUNCTION__)) | |||
7368 | N->getOpcode() == ISD::ZERO_EXTEND) &&(static_cast <bool> ((N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND) && "Unexpected node type (not an extend)!" ) ? void (0) : __assert_fail ("(N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND) && \"Unexpected node type (not an extend)!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 7369, __extension__ __PRETTY_FUNCTION__)) | |||
7369 | "Unexpected node type (not an extend)!")(static_cast <bool> ((N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND) && "Unexpected node type (not an extend)!" ) ? void (0) : __assert_fail ("(N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND) && \"Unexpected node type (not an extend)!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 7369, __extension__ __PRETTY_FUNCTION__)); | |||
7370 | ||||
7371 | // fold (sext (load x)) to multiple smaller sextloads; same for zext. | |||
7372 | // For example, on a target with legal v4i32, but illegal v8i32, turn: | |||
7373 | // (v8i32 (sext (v8i16 (load x)))) | |||
7374 | // into: | |||
7375 | // (v8i32 (concat_vectors (v4i32 (sextload x)), | |||
7376 | // (v4i32 (sextload (x + 16))))) | |||
7377 | // Where uses of the original load, i.e.: | |||
7378 | // (v8i16 (load x)) | |||
7379 | // are replaced with: | |||
7380 | // (v8i16 (truncate | |||
7381 | // (v8i32 (concat_vectors (v4i32 (sextload x)), | |||
7382 | // (v4i32 (sextload (x + 16))))))) | |||
7383 | // | |||
7384 | // This combine is only applicable to illegal, but splittable, vectors. | |||
7385 | // All legal types, and illegal non-vector types, are handled elsewhere. | |||
7386 | // This combine is controlled by TargetLowering::isVectorLoadExtDesirable. | |||
7387 | // | |||
7388 | if (N0->getOpcode() != ISD::LOAD) | |||
7389 | return SDValue(); | |||
7390 | ||||
7391 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
7392 | ||||
7393 | if (!ISD::isNON_EXTLoad(LN0) || !ISD::isUNINDEXEDLoad(LN0) || | |||
7394 | !N0.hasOneUse() || LN0->isVolatile() || !DstVT.isVector() || | |||
7395 | !DstVT.isPow2VectorType() || !TLI.isVectorLoadExtDesirable(SDValue(N, 0))) | |||
7396 | return SDValue(); | |||
7397 | ||||
7398 | SmallVector<SDNode *, 4> SetCCs; | |||
7399 | if (!ExtendUsesToFormExtLoad(DstVT, N, N0, N->getOpcode(), SetCCs, TLI)) | |||
7400 | return SDValue(); | |||
7401 | ||||
7402 | ISD::LoadExtType ExtType = | |||
7403 | N->getOpcode() == ISD::SIGN_EXTEND ? ISD::SEXTLOAD : ISD::ZEXTLOAD; | |||
7404 | ||||
7405 | // Try to split the vector types to get down to legal types. | |||
7406 | EVT SplitSrcVT = SrcVT; | |||
7407 | EVT SplitDstVT = DstVT; | |||
7408 | while (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT) && | |||
7409 | SplitSrcVT.getVectorNumElements() > 1) { | |||
7410 | SplitDstVT = DAG.GetSplitDestVTs(SplitDstVT).first; | |||
7411 | SplitSrcVT = DAG.GetSplitDestVTs(SplitSrcVT).first; | |||
7412 | } | |||
7413 | ||||
7414 | if (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT)) | |||
7415 | return SDValue(); | |||
7416 | ||||
7417 | SDLoc DL(N); | |||
7418 | const unsigned NumSplits = | |||
7419 | DstVT.getVectorNumElements() / SplitDstVT.getVectorNumElements(); | |||
7420 | const unsigned Stride = SplitSrcVT.getStoreSize(); | |||
7421 | SmallVector<SDValue, 4> Loads; | |||
7422 | SmallVector<SDValue, 4> Chains; | |||
7423 | ||||
7424 | SDValue BasePtr = LN0->getBasePtr(); | |||
7425 | for (unsigned Idx = 0; Idx < NumSplits; Idx++) { | |||
7426 | const unsigned Offset = Idx * Stride; | |||
7427 | const unsigned Align = MinAlign(LN0->getAlignment(), Offset); | |||
7428 | ||||
7429 | SDValue SplitLoad = DAG.getExtLoad( | |||
7430 | ExtType, DL, SplitDstVT, LN0->getChain(), BasePtr, | |||
7431 | LN0->getPointerInfo().getWithOffset(Offset), SplitSrcVT, Align, | |||
7432 | LN0->getMemOperand()->getFlags(), LN0->getAAInfo()); | |||
7433 | ||||
7434 | BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr, | |||
7435 | DAG.getConstant(Stride, DL, BasePtr.getValueType())); | |||
7436 | ||||
7437 | Loads.push_back(SplitLoad.getValue(0)); | |||
7438 | Chains.push_back(SplitLoad.getValue(1)); | |||
7439 | } | |||
7440 | ||||
7441 | SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains); | |||
7442 | SDValue NewValue = DAG.getNode(ISD::CONCAT_VECTORS, DL, DstVT, Loads); | |||
7443 | ||||
7444 | // Simplify TF. | |||
7445 | AddToWorklist(NewChain.getNode()); | |||
7446 | ||||
7447 | CombineTo(N, NewValue); | |||
7448 | ||||
7449 | // Replace uses of the original load (before extension) | |||
7450 | // with a truncate of the concatenated sextloaded vectors. | |||
7451 | SDValue Trunc = | |||
7452 | DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(), NewValue); | |||
7453 | ExtendSetCCUses(SetCCs, N0, NewValue, DL, | |||
7454 | (ISD::NodeType)N->getOpcode()); | |||
7455 | CombineTo(N0.getNode(), Trunc, NewChain); | |||
7456 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
7457 | } | |||
7458 | ||||
7459 | /// If we're narrowing or widening the result of a vector select and the final | |||
7460 | /// size is the same size as a setcc (compare) feeding the select, then try to | |||
7461 | /// apply the cast operation to the select's operands because matching vector | |||
7462 | /// sizes for a select condition and other operands should be more efficient. | |||
7463 | SDValue DAGCombiner::matchVSelectOpSizesWithSetCC(SDNode *Cast) { | |||
7464 | unsigned CastOpcode = Cast->getOpcode(); | |||
7465 | assert((CastOpcode == ISD::SIGN_EXTEND || CastOpcode == ISD::ZERO_EXTEND ||(static_cast <bool> ((CastOpcode == ISD::SIGN_EXTEND || CastOpcode == ISD::ZERO_EXTEND || CastOpcode == ISD::TRUNCATE || CastOpcode == ISD::FP_EXTEND || CastOpcode == ISD::FP_ROUND ) && "Unexpected opcode for vector select narrowing/widening" ) ? void (0) : __assert_fail ("(CastOpcode == ISD::SIGN_EXTEND || CastOpcode == ISD::ZERO_EXTEND || CastOpcode == ISD::TRUNCATE || CastOpcode == ISD::FP_EXTEND || CastOpcode == ISD::FP_ROUND) && \"Unexpected opcode for vector select narrowing/widening\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 7468, __extension__ __PRETTY_FUNCTION__)) | |||
7466 | CastOpcode == ISD::TRUNCATE || CastOpcode == ISD::FP_EXTEND ||(static_cast <bool> ((CastOpcode == ISD::SIGN_EXTEND || CastOpcode == ISD::ZERO_EXTEND || CastOpcode == ISD::TRUNCATE || CastOpcode == ISD::FP_EXTEND || CastOpcode == ISD::FP_ROUND ) && "Unexpected opcode for vector select narrowing/widening" ) ? void (0) : __assert_fail ("(CastOpcode == ISD::SIGN_EXTEND || CastOpcode == ISD::ZERO_EXTEND || CastOpcode == ISD::TRUNCATE || CastOpcode == ISD::FP_EXTEND || CastOpcode == ISD::FP_ROUND) && \"Unexpected opcode for vector select narrowing/widening\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 7468, __extension__ __PRETTY_FUNCTION__)) | |||
7467 | CastOpcode == ISD::FP_ROUND) &&(static_cast <bool> ((CastOpcode == ISD::SIGN_EXTEND || CastOpcode == ISD::ZERO_EXTEND || CastOpcode == ISD::TRUNCATE || CastOpcode == ISD::FP_EXTEND || CastOpcode == ISD::FP_ROUND ) && "Unexpected opcode for vector select narrowing/widening" ) ? void (0) : __assert_fail ("(CastOpcode == ISD::SIGN_EXTEND || CastOpcode == ISD::ZERO_EXTEND || CastOpcode == ISD::TRUNCATE || CastOpcode == ISD::FP_EXTEND || CastOpcode == ISD::FP_ROUND) && \"Unexpected opcode for vector select narrowing/widening\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 7468, __extension__ __PRETTY_FUNCTION__)) | |||
7468 | "Unexpected opcode for vector select narrowing/widening")(static_cast <bool> ((CastOpcode == ISD::SIGN_EXTEND || CastOpcode == ISD::ZERO_EXTEND || CastOpcode == ISD::TRUNCATE || CastOpcode == ISD::FP_EXTEND || CastOpcode == ISD::FP_ROUND ) && "Unexpected opcode for vector select narrowing/widening" ) ? void (0) : __assert_fail ("(CastOpcode == ISD::SIGN_EXTEND || CastOpcode == ISD::ZERO_EXTEND || CastOpcode == ISD::TRUNCATE || CastOpcode == ISD::FP_EXTEND || CastOpcode == ISD::FP_ROUND) && \"Unexpected opcode for vector select narrowing/widening\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 7468, __extension__ __PRETTY_FUNCTION__)); | |||
7469 | ||||
7470 | // We only do this transform before legal ops because the pattern may be | |||
7471 | // obfuscated by target-specific operations after legalization. Do not create | |||
7472 | // an illegal select op, however, because that may be difficult to lower. | |||
7473 | EVT VT = Cast->getValueType(0); | |||
7474 | if (LegalOperations || !TLI.isOperationLegalOrCustom(ISD::VSELECT, VT)) | |||
7475 | return SDValue(); | |||
7476 | ||||
7477 | SDValue VSel = Cast->getOperand(0); | |||
7478 | if (VSel.getOpcode() != ISD::VSELECT || !VSel.hasOneUse() || | |||
7479 | VSel.getOperand(0).getOpcode() != ISD::SETCC) | |||
7480 | return SDValue(); | |||
7481 | ||||
7482 | // Does the setcc have the same vector size as the casted select? | |||
7483 | SDValue SetCC = VSel.getOperand(0); | |||
7484 | EVT SetCCVT = getSetCCResultType(SetCC.getOperand(0).getValueType()); | |||
7485 | if (SetCCVT.getSizeInBits() != VT.getSizeInBits()) | |||
7486 | return SDValue(); | |||
7487 | ||||
7488 | // cast (vsel (setcc X), A, B) --> vsel (setcc X), (cast A), (cast B) | |||
7489 | SDValue A = VSel.getOperand(1); | |||
7490 | SDValue B = VSel.getOperand(2); | |||
7491 | SDValue CastA, CastB; | |||
7492 | SDLoc DL(Cast); | |||
7493 | if (CastOpcode == ISD::FP_ROUND) { | |||
7494 | // FP_ROUND (fptrunc) has an extra flag operand to pass along. | |||
7495 | CastA = DAG.getNode(CastOpcode, DL, VT, A, Cast->getOperand(1)); | |||
7496 | CastB = DAG.getNode(CastOpcode, DL, VT, B, Cast->getOperand(1)); | |||
7497 | } else { | |||
7498 | CastA = DAG.getNode(CastOpcode, DL, VT, A); | |||
7499 | CastB = DAG.getNode(CastOpcode, DL, VT, B); | |||
7500 | } | |||
7501 | return DAG.getNode(ISD::VSELECT, DL, VT, SetCC, CastA, CastB); | |||
7502 | } | |||
7503 | ||||
7504 | SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) { | |||
7505 | SDValue N0 = N->getOperand(0); | |||
7506 | EVT VT = N->getValueType(0); | |||
7507 | SDLoc DL(N); | |||
7508 | ||||
7509 | if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes, | |||
7510 | LegalOperations)) | |||
7511 | return SDValue(Res, 0); | |||
7512 | ||||
7513 | // fold (sext (sext x)) -> (sext x) | |||
7514 | // fold (sext (aext x)) -> (sext x) | |||
7515 | if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) | |||
7516 | return DAG.getNode(ISD::SIGN_EXTEND, DL, VT, N0.getOperand(0)); | |||
7517 | ||||
7518 | if (N0.getOpcode() == ISD::TRUNCATE) { | |||
7519 | // fold (sext (truncate (load x))) -> (sext (smaller load x)) | |||
7520 | // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n))) | |||
7521 | if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) { | |||
7522 | SDNode *oye = N0.getOperand(0).getNode(); | |||
7523 | if (NarrowLoad.getNode() != N0.getNode()) { | |||
7524 | CombineTo(N0.getNode(), NarrowLoad); | |||
7525 | // CombineTo deleted the truncate, if needed, but not what's under it. | |||
7526 | AddToWorklist(oye); | |||
7527 | } | |||
7528 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
7529 | } | |||
7530 | ||||
7531 | // See if the value being truncated is already sign extended. If so, just | |||
7532 | // eliminate the trunc/sext pair. | |||
7533 | SDValue Op = N0.getOperand(0); | |||
7534 | unsigned OpBits = Op.getScalarValueSizeInBits(); | |||
7535 | unsigned MidBits = N0.getScalarValueSizeInBits(); | |||
7536 | unsigned DestBits = VT.getScalarSizeInBits(); | |||
7537 | unsigned NumSignBits = DAG.ComputeNumSignBits(Op); | |||
7538 | ||||
7539 | if (OpBits == DestBits) { | |||
7540 | // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign | |||
7541 | // bits, it is already ready. | |||
7542 | if (NumSignBits > DestBits-MidBits) | |||
7543 | return Op; | |||
7544 | } else if (OpBits < DestBits) { | |||
7545 | // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign | |||
7546 | // bits, just sext from i32. | |||
7547 | if (NumSignBits > OpBits-MidBits) | |||
7548 | return DAG.getNode(ISD::SIGN_EXTEND, DL, VT, Op); | |||
7549 | } else { | |||
7550 | // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign | |||
7551 | // bits, just truncate to i32. | |||
7552 | if (NumSignBits > OpBits-MidBits) | |||
7553 | return DAG.getNode(ISD::TRUNCATE, DL, VT, Op); | |||
7554 | } | |||
7555 | ||||
7556 | // fold (sext (truncate x)) -> (sextinreg x). | |||
7557 | if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, | |||
7558 | N0.getValueType())) { | |||
7559 | if (OpBits < DestBits) | |||
7560 | Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op); | |||
7561 | else if (OpBits > DestBits) | |||
7562 | Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op); | |||
7563 | return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Op, | |||
7564 | DAG.getValueType(N0.getValueType())); | |||
7565 | } | |||
7566 | } | |||
7567 | ||||
7568 | // fold (sext (load x)) -> (sext (truncate (sextload x))) | |||
7569 | // Only generate vector extloads when 1) they're legal, and 2) they are | |||
7570 | // deemed desirable by the target. | |||
7571 | if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) && | |||
7572 | ((!LegalOperations && !VT.isVector() && | |||
7573 | !cast<LoadSDNode>(N0)->isVolatile()) || | |||
7574 | TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()))) { | |||
7575 | bool DoXform = true; | |||
7576 | SmallVector<SDNode*, 4> SetCCs; | |||
7577 | if (!N0.hasOneUse()) | |||
7578 | DoXform = ExtendUsesToFormExtLoad(VT, N, N0, ISD::SIGN_EXTEND, SetCCs, | |||
7579 | TLI); | |||
7580 | if (VT.isVector()) | |||
7581 | DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0)); | |||
7582 | if (DoXform) { | |||
7583 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
7584 | SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, DL, VT, LN0->getChain(), | |||
7585 | LN0->getBasePtr(), N0.getValueType(), | |||
7586 | LN0->getMemOperand()); | |||
7587 | ExtendSetCCUses(SetCCs, N0, ExtLoad, DL, ISD::SIGN_EXTEND); | |||
7588 | // If the load value is used only by N, replace it via CombineTo N. | |||
7589 | bool NoReplaceTrunc = SDValue(LN0, 0).hasOneUse(); | |||
7590 | CombineTo(N, ExtLoad); | |||
7591 | if (NoReplaceTrunc) { | |||
7592 | DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), ExtLoad.getValue(1)); | |||
7593 | } else { | |||
7594 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), | |||
7595 | N0.getValueType(), ExtLoad); | |||
7596 | CombineTo(LN0, Trunc, ExtLoad.getValue(1)); | |||
7597 | } | |||
7598 | return SDValue(N, 0); | |||
7599 | } | |||
7600 | } | |||
7601 | ||||
7602 | // fold (sext (load x)) to multiple smaller sextloads. | |||
7603 | // Only on illegal but splittable vectors. | |||
7604 | if (SDValue ExtLoad = CombineExtLoad(N)) | |||
7605 | return ExtLoad; | |||
7606 | ||||
7607 | // fold (sext (sextload x)) -> (sext (truncate (sextload x))) | |||
7608 | // fold (sext ( extload x)) -> (sext (truncate (sextload x))) | |||
7609 | if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) && | |||
7610 | ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) { | |||
7611 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
7612 | EVT MemVT = LN0->getMemoryVT(); | |||
7613 | if ((!LegalOperations && !LN0->isVolatile()) || | |||
7614 | TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, MemVT)) { | |||
7615 | SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, DL, VT, LN0->getChain(), | |||
7616 | LN0->getBasePtr(), MemVT, | |||
7617 | LN0->getMemOperand()); | |||
7618 | CombineTo(N, ExtLoad); | |||
7619 | DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), ExtLoad.getValue(1)); | |||
7620 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
7621 | } | |||
7622 | } | |||
7623 | ||||
7624 | // fold (sext (and/or/xor (load x), cst)) -> | |||
7625 | // (and/or/xor (sextload x), (sext cst)) | |||
7626 | if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR || | |||
7627 | N0.getOpcode() == ISD::XOR) && | |||
7628 | isa<LoadSDNode>(N0.getOperand(0)) && | |||
7629 | N0.getOperand(1).getOpcode() == ISD::Constant && | |||
7630 | (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) { | |||
7631 | LoadSDNode *LN00 = cast<LoadSDNode>(N0.getOperand(0)); | |||
7632 | EVT MemVT = LN00->getMemoryVT(); | |||
7633 | if (TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, MemVT) && | |||
7634 | LN00->getExtensionType() != ISD::ZEXTLOAD && LN00->isUnindexed()) { | |||
7635 | SmallVector<SDNode*, 4> SetCCs; | |||
7636 | bool DoXform = ExtendUsesToFormExtLoad(VT, N0.getNode(), N0.getOperand(0), | |||
7637 | ISD::SIGN_EXTEND, SetCCs, TLI); | |||
7638 | if (DoXform) { | |||
7639 | SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN00), VT, | |||
7640 | LN00->getChain(), LN00->getBasePtr(), | |||
7641 | LN00->getMemoryVT(), | |||
7642 | LN00->getMemOperand()); | |||
7643 | APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); | |||
7644 | Mask = Mask.sext(VT.getSizeInBits()); | |||
7645 | SDValue And = DAG.getNode(N0.getOpcode(), DL, VT, | |||
7646 | ExtLoad, DAG.getConstant(Mask, DL, VT)); | |||
7647 | ExtendSetCCUses(SetCCs, N0.getOperand(0), ExtLoad, DL, | |||
7648 | ISD::SIGN_EXTEND); | |||
7649 | bool NoReplaceTruncAnd = !N0.hasOneUse(); | |||
7650 | bool NoReplaceTrunc = SDValue(LN00, 0).hasOneUse(); | |||
7651 | CombineTo(N, And); | |||
7652 | // If N0 has multiple uses, change other uses as well. | |||
7653 | if (NoReplaceTruncAnd) { | |||
7654 | SDValue TruncAnd = | |||
7655 | DAG.getNode(ISD::TRUNCATE, DL, N0.getValueType(), And); | |||
7656 | CombineTo(N0.getNode(), TruncAnd); | |||
7657 | } | |||
7658 | if (NoReplaceTrunc) { | |||
7659 | DAG.ReplaceAllUsesOfValueWith(SDValue(LN00, 1), ExtLoad.getValue(1)); | |||
7660 | } else { | |||
7661 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(LN00), | |||
7662 | LN00->getValueType(0), ExtLoad); | |||
7663 | CombineTo(LN00, Trunc, ExtLoad.getValue(1)); | |||
7664 | } | |||
7665 | return SDValue(N,0); // Return N so it doesn't get rechecked! | |||
7666 | } | |||
7667 | } | |||
7668 | } | |||
7669 | ||||
7670 | if (N0.getOpcode() == ISD::SETCC) { | |||
7671 | SDValue N00 = N0.getOperand(0); | |||
7672 | SDValue N01 = N0.getOperand(1); | |||
7673 | ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); | |||
7674 | EVT N00VT = N0.getOperand(0).getValueType(); | |||
7675 | ||||
7676 | // sext(setcc) -> sext_in_reg(vsetcc) for vectors. | |||
7677 | // Only do this before legalize for now. | |||
7678 | if (VT.isVector() && !LegalOperations && | |||
7679 | TLI.getBooleanContents(N00VT) == | |||
7680 | TargetLowering::ZeroOrNegativeOneBooleanContent) { | |||
7681 | // On some architectures (such as SSE/NEON/etc) the SETCC result type is | |||
7682 | // of the same size as the compared operands. Only optimize sext(setcc()) | |||
7683 | // if this is the case. | |||
7684 | EVT SVT = getSetCCResultType(N00VT); | |||
7685 | ||||
7686 | // We know that the # elements of the results is the same as the | |||
7687 | // # elements of the compare (and the # elements of the compare result | |||
7688 | // for that matter). Check to see that they are the same size. If so, | |||
7689 | // we know that the element size of the sext'd result matches the | |||
7690 | // element size of the compare operands. | |||
7691 | if (VT.getSizeInBits() == SVT.getSizeInBits()) | |||
7692 | return DAG.getSetCC(DL, VT, N00, N01, CC); | |||
7693 | ||||
7694 | // If the desired elements are smaller or larger than the source | |||
7695 | // elements, we can use a matching integer vector type and then | |||
7696 | // truncate/sign extend. | |||
7697 | EVT MatchingVecType = N00VT.changeVectorElementTypeToInteger(); | |||
7698 | if (SVT == MatchingVecType) { | |||
7699 | SDValue VsetCC = DAG.getSetCC(DL, MatchingVecType, N00, N01, CC); | |||
7700 | return DAG.getSExtOrTrunc(VsetCC, DL, VT); | |||
7701 | } | |||
7702 | } | |||
7703 | ||||
7704 | // sext(setcc x, y, cc) -> (select (setcc x, y, cc), T, 0) | |||
7705 | // Here, T can be 1 or -1, depending on the type of the setcc and | |||
7706 | // getBooleanContents(). | |||
7707 | unsigned SetCCWidth = N0.getScalarValueSizeInBits(); | |||
7708 | ||||
7709 | // To determine the "true" side of the select, we need to know the high bit | |||
7710 | // of the value returned by the setcc if it evaluates to true. | |||
7711 | // If the type of the setcc is i1, then the true case of the select is just | |||
7712 | // sext(i1 1), that is, -1. | |||
7713 | // If the type of the setcc is larger (say, i8) then the value of the high | |||
7714 | // bit depends on getBooleanContents(), so ask TLI for a real "true" value | |||
7715 | // of the appropriate width. | |||
7716 | SDValue ExtTrueVal = (SetCCWidth == 1) | |||
7717 | ? DAG.getAllOnesConstant(DL, VT) | |||
7718 | : DAG.getBoolConstant(true, DL, VT, N00VT); | |||
7719 | SDValue Zero = DAG.getConstant(0, DL, VT); | |||
7720 | if (SDValue SCC = | |||
7721 | SimplifySelectCC(DL, N00, N01, ExtTrueVal, Zero, CC, true)) | |||
7722 | return SCC; | |||
7723 | ||||
7724 | if (!VT.isVector() && !TLI.convertSelectOfConstantsToMath(VT)) { | |||
7725 | EVT SetCCVT = getSetCCResultType(N00VT); | |||
7726 | // Don't do this transform for i1 because there's a select transform | |||
7727 | // that would reverse it. | |||
7728 | // TODO: We should not do this transform at all without a target hook | |||
7729 | // because a sext is likely cheaper than a select? | |||
7730 | if (SetCCVT.getScalarSizeInBits() != 1 && | |||
7731 | (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, N00VT))) { | |||
7732 | SDValue SetCC = DAG.getSetCC(DL, SetCCVT, N00, N01, CC); | |||
7733 | return DAG.getSelect(DL, VT, SetCC, ExtTrueVal, Zero); | |||
7734 | } | |||
7735 | } | |||
7736 | } | |||
7737 | ||||
7738 | // fold (sext x) -> (zext x) if the sign bit is known zero. | |||
7739 | if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) && | |||
7740 | DAG.SignBitIsZero(N0)) | |||
7741 | return DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0); | |||
7742 | ||||
7743 | if (SDValue NewVSel = matchVSelectOpSizesWithSetCC(N)) | |||
7744 | return NewVSel; | |||
7745 | ||||
7746 | return SDValue(); | |||
7747 | } | |||
7748 | ||||
7749 | // isTruncateOf - If N is a truncate of some other value, return true, record | |||
7750 | // the value being truncated in Op and which of Op's bits are zero/one in Known. | |||
7751 | // This function computes KnownBits to avoid a duplicated call to | |||
7752 | // computeKnownBits in the caller. | |||
7753 | static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op, | |||
7754 | KnownBits &Known) { | |||
7755 | if (N->getOpcode() == ISD::TRUNCATE) { | |||
7756 | Op = N->getOperand(0); | |||
7757 | DAG.computeKnownBits(Op, Known); | |||
7758 | return true; | |||
7759 | } | |||
7760 | ||||
7761 | if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 || | |||
7762 | cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE) | |||
7763 | return false; | |||
7764 | ||||
7765 | SDValue Op0 = N->getOperand(0); | |||
7766 | SDValue Op1 = N->getOperand(1); | |||
7767 | assert(Op0.getValueType() == Op1.getValueType())(static_cast <bool> (Op0.getValueType() == Op1.getValueType ()) ? void (0) : __assert_fail ("Op0.getValueType() == Op1.getValueType()" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 7767, __extension__ __PRETTY_FUNCTION__)); | |||
7768 | ||||
7769 | if (isNullConstant(Op0)) | |||
7770 | Op = Op1; | |||
7771 | else if (isNullConstant(Op1)) | |||
7772 | Op = Op0; | |||
7773 | else | |||
7774 | return false; | |||
7775 | ||||
7776 | DAG.computeKnownBits(Op, Known); | |||
7777 | ||||
7778 | if (!(Known.Zero | 1).isAllOnesValue()) | |||
7779 | return false; | |||
7780 | ||||
7781 | return true; | |||
7782 | } | |||
7783 | ||||
7784 | SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) { | |||
7785 | SDValue N0 = N->getOperand(0); | |||
7786 | EVT VT = N->getValueType(0); | |||
7787 | ||||
7788 | if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes, | |||
7789 | LegalOperations)) | |||
7790 | return SDValue(Res, 0); | |||
7791 | ||||
7792 | // fold (zext (zext x)) -> (zext x) | |||
7793 | // fold (zext (aext x)) -> (zext x) | |||
7794 | if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) | |||
7795 | return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, | |||
7796 | N0.getOperand(0)); | |||
7797 | ||||
7798 | // fold (zext (truncate x)) -> (zext x) or | |||
7799 | // (zext (truncate x)) -> (truncate x) | |||
7800 | // This is valid when the truncated bits of x are already zero. | |||
7801 | // FIXME: We should extend this to work for vectors too. | |||
7802 | SDValue Op; | |||
7803 | KnownBits Known; | |||
7804 | if (!VT.isVector() && isTruncateOf(DAG, N0, Op, Known)) { | |||
7805 | APInt TruncatedBits = | |||
7806 | (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ? | |||
7807 | APInt(Op.getValueSizeInBits(), 0) : | |||
7808 | APInt::getBitsSet(Op.getValueSizeInBits(), | |||
7809 | N0.getValueSizeInBits(), | |||
7810 | std::min(Op.getValueSizeInBits(), | |||
7811 | VT.getSizeInBits())); | |||
7812 | if (TruncatedBits.isSubsetOf(Known.Zero)) | |||
7813 | return DAG.getZExtOrTrunc(Op, SDLoc(N), VT); | |||
7814 | } | |||
7815 | ||||
7816 | // fold (zext (truncate x)) -> (and x, mask) | |||
7817 | if (N0.getOpcode() == ISD::TRUNCATE) { | |||
7818 | // fold (zext (truncate (load x))) -> (zext (smaller load x)) | |||
7819 | // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n))) | |||
7820 | if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) { | |||
7821 | SDNode *oye = N0.getOperand(0).getNode(); | |||
7822 | if (NarrowLoad.getNode() != N0.getNode()) { | |||
7823 | CombineTo(N0.getNode(), NarrowLoad); | |||
7824 | // CombineTo deleted the truncate, if needed, but not what's under it. | |||
7825 | AddToWorklist(oye); | |||
7826 | } | |||
7827 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
7828 | } | |||
7829 | ||||
7830 | EVT SrcVT = N0.getOperand(0).getValueType(); | |||
7831 | EVT MinVT = N0.getValueType(); | |||
7832 | ||||
7833 | // Try to mask before the extension to avoid having to generate a larger mask, | |||
7834 | // possibly over several sub-vectors. | |||
7835 | if (SrcVT.bitsLT(VT)) { | |||
7836 | if (!LegalOperations || (TLI.isOperationLegal(ISD::AND, SrcVT) && | |||
7837 | TLI.isOperationLegal(ISD::ZERO_EXTEND, VT))) { | |||
7838 | SDValue Op = N0.getOperand(0); | |||
7839 | Op = DAG.getZeroExtendInReg(Op, SDLoc(N), MinVT.getScalarType()); | |||
7840 | AddToWorklist(Op.getNode()); | |||
7841 | SDValue ZExtOrTrunc = DAG.getZExtOrTrunc(Op, SDLoc(N), VT); | |||
7842 | // Transfer the debug info; the new node is equivalent to N0. | |||
7843 | DAG.transferDbgValues(N0, ZExtOrTrunc); | |||
7844 | return ZExtOrTrunc; | |||
7845 | } | |||
7846 | } | |||
7847 | ||||
7848 | if (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT)) { | |||
7849 | SDValue Op = DAG.getAnyExtOrTrunc(N0.getOperand(0), SDLoc(N), VT); | |||
7850 | AddToWorklist(Op.getNode()); | |||
7851 | SDValue And = DAG.getZeroExtendInReg(Op, SDLoc(N), MinVT.getScalarType()); | |||
7852 | // We may safely transfer the debug info describing the truncate node over | |||
7853 | // to the equivalent and operation. | |||
7854 | DAG.transferDbgValues(N0, And); | |||
7855 | return And; | |||
7856 | } | |||
7857 | } | |||
7858 | ||||
7859 | // Fold (zext (and (trunc x), cst)) -> (and x, cst), | |||
7860 | // if either of the casts is not free. | |||
7861 | if (N0.getOpcode() == ISD::AND && | |||
7862 | N0.getOperand(0).getOpcode() == ISD::TRUNCATE && | |||
7863 | N0.getOperand(1).getOpcode() == ISD::Constant && | |||
7864 | (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(), | |||
7865 | N0.getValueType()) || | |||
7866 | !TLI.isZExtFree(N0.getValueType(), VT))) { | |||
7867 | SDValue X = N0.getOperand(0).getOperand(0); | |||
7868 | X = DAG.getAnyExtOrTrunc(X, SDLoc(X), VT); | |||
7869 | APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); | |||
7870 | Mask = Mask.zext(VT.getSizeInBits()); | |||
7871 | SDLoc DL(N); | |||
7872 | return DAG.getNode(ISD::AND, DL, VT, | |||
7873 | X, DAG.getConstant(Mask, DL, VT)); | |||
7874 | } | |||
7875 | ||||
7876 | // fold (zext (load x)) -> (zext (truncate (zextload x))) | |||
7877 | // Only generate vector extloads when 1) they're legal, and 2) they are | |||
7878 | // deemed desirable by the target. | |||
7879 | if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) && | |||
7880 | ((!LegalOperations && !VT.isVector() && | |||
7881 | !cast<LoadSDNode>(N0)->isVolatile()) || | |||
7882 | TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()))) { | |||
7883 | bool DoXform = true; | |||
7884 | SmallVector<SDNode*, 4> SetCCs; | |||
7885 | if (!N0.hasOneUse()) | |||
7886 | DoXform = ExtendUsesToFormExtLoad(VT, N, N0, ISD::ZERO_EXTEND, SetCCs, | |||
7887 | TLI); | |||
7888 | if (VT.isVector()) | |||
7889 | DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0)); | |||
7890 | if (DoXform) { | |||
7891 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
7892 | SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT, | |||
7893 | LN0->getChain(), | |||
7894 | LN0->getBasePtr(), N0.getValueType(), | |||
7895 | LN0->getMemOperand()); | |||
7896 | ||||
7897 | ExtendSetCCUses(SetCCs, N0, ExtLoad, SDLoc(N), ISD::ZERO_EXTEND); | |||
7898 | // If the load value is used only by N, replace it via CombineTo N. | |||
7899 | bool NoReplaceTrunc = SDValue(LN0, 0).hasOneUse(); | |||
7900 | CombineTo(N, ExtLoad); | |||
7901 | if (NoReplaceTrunc) { | |||
7902 | DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), ExtLoad.getValue(1)); | |||
7903 | } else { | |||
7904 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), | |||
7905 | N0.getValueType(), ExtLoad); | |||
7906 | CombineTo(LN0, Trunc, ExtLoad.getValue(1)); | |||
7907 | } | |||
7908 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
7909 | } | |||
7910 | } | |||
7911 | ||||
7912 | // fold (zext (load x)) to multiple smaller zextloads. | |||
7913 | // Only on illegal but splittable vectors. | |||
7914 | if (SDValue ExtLoad = CombineExtLoad(N)) | |||
7915 | return ExtLoad; | |||
7916 | ||||
7917 | // fold (zext (and/or/xor (load x), cst)) -> | |||
7918 | // (and/or/xor (zextload x), (zext cst)) | |||
7919 | // Unless (and (load x) cst) will match as a zextload already and has | |||
7920 | // additional users. | |||
7921 | if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR || | |||
7922 | N0.getOpcode() == ISD::XOR) && | |||
7923 | isa<LoadSDNode>(N0.getOperand(0)) && | |||
7924 | N0.getOperand(1).getOpcode() == ISD::Constant && | |||
7925 | (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) { | |||
7926 | LoadSDNode *LN00 = cast<LoadSDNode>(N0.getOperand(0)); | |||
7927 | EVT MemVT = LN00->getMemoryVT(); | |||
7928 | if (TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT) && | |||
7929 | LN00->getExtensionType() != ISD::SEXTLOAD && LN00->isUnindexed()) { | |||
7930 | bool DoXform = true; | |||
7931 | SmallVector<SDNode*, 4> SetCCs; | |||
7932 | if (!N0.hasOneUse()) { | |||
7933 | if (N0.getOpcode() == ISD::AND) { | |||
7934 | auto *AndC = cast<ConstantSDNode>(N0.getOperand(1)); | |||
7935 | EVT LoadResultTy = AndC->getValueType(0); | |||
7936 | EVT ExtVT; | |||
7937 | if (isAndLoadExtLoad(AndC, LN00, LoadResultTy, ExtVT)) | |||
7938 | DoXform = false; | |||
7939 | } | |||
7940 | } | |||
7941 | if (DoXform) | |||
7942 | DoXform = ExtendUsesToFormExtLoad(VT, N0.getNode(), N0.getOperand(0), | |||
7943 | ISD::ZERO_EXTEND, SetCCs, TLI); | |||
7944 | if (DoXform) { | |||
7945 | SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN00), VT, | |||
7946 | LN00->getChain(), LN00->getBasePtr(), | |||
7947 | LN00->getMemoryVT(), | |||
7948 | LN00->getMemOperand()); | |||
7949 | APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); | |||
7950 | Mask = Mask.zext(VT.getSizeInBits()); | |||
7951 | SDLoc DL(N); | |||
7952 | SDValue And = DAG.getNode(N0.getOpcode(), DL, VT, | |||
7953 | ExtLoad, DAG.getConstant(Mask, DL, VT)); | |||
7954 | ExtendSetCCUses(SetCCs, N0.getOperand(0), ExtLoad, DL, | |||
7955 | ISD::ZERO_EXTEND); | |||
7956 | bool NoReplaceTruncAnd = !N0.hasOneUse(); | |||
7957 | bool NoReplaceTrunc = SDValue(LN00, 0).hasOneUse(); | |||
7958 | CombineTo(N, And); | |||
7959 | // If N0 has multiple uses, change other uses as well. | |||
7960 | if (NoReplaceTruncAnd) { | |||
7961 | SDValue TruncAnd = | |||
7962 | DAG.getNode(ISD::TRUNCATE, DL, N0.getValueType(), And); | |||
7963 | CombineTo(N0.getNode(), TruncAnd); | |||
7964 | } | |||
7965 | if (NoReplaceTrunc) { | |||
7966 | DAG.ReplaceAllUsesOfValueWith(SDValue(LN00, 1), ExtLoad.getValue(1)); | |||
7967 | } else { | |||
7968 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(LN00), | |||
7969 | LN00->getValueType(0), ExtLoad); | |||
7970 | CombineTo(LN00, Trunc, ExtLoad.getValue(1)); | |||
7971 | } | |||
7972 | return SDValue(N,0); // Return N so it doesn't get rechecked! | |||
7973 | } | |||
7974 | } | |||
7975 | } | |||
7976 | ||||
7977 | // fold (zext (zextload x)) -> (zext (truncate (zextload x))) | |||
7978 | // fold (zext ( extload x)) -> (zext (truncate (zextload x))) | |||
7979 | if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) && | |||
7980 | ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) { | |||
7981 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
7982 | EVT MemVT = LN0->getMemoryVT(); | |||
7983 | if ((!LegalOperations && !LN0->isVolatile()) || | |||
7984 | TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT)) { | |||
7985 | SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT, | |||
7986 | LN0->getChain(), | |||
7987 | LN0->getBasePtr(), MemVT, | |||
7988 | LN0->getMemOperand()); | |||
7989 | CombineTo(N, ExtLoad); | |||
7990 | DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), ExtLoad.getValue(1)); | |||
7991 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
7992 | } | |||
7993 | } | |||
7994 | ||||
7995 | if (N0.getOpcode() == ISD::SETCC) { | |||
7996 | // Only do this before legalize for now. | |||
7997 | if (!LegalOperations && VT.isVector() && | |||
7998 | N0.getValueType().getVectorElementType() == MVT::i1) { | |||
7999 | EVT N00VT = N0.getOperand(0).getValueType(); | |||
8000 | if (getSetCCResultType(N00VT) == N0.getValueType()) | |||
8001 | return SDValue(); | |||
8002 | ||||
8003 | // We know that the # elements of the results is the same as the # | |||
8004 | // elements of the compare (and the # elements of the compare result for | |||
8005 | // that matter). Check to see that they are the same size. If so, we know | |||
8006 | // that the element size of the sext'd result matches the element size of | |||
8007 | // the compare operands. | |||
8008 | SDLoc DL(N); | |||
8009 | SDValue VecOnes = DAG.getConstant(1, DL, VT); | |||
8010 | if (VT.getSizeInBits() == N00VT.getSizeInBits()) { | |||
8011 | // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors. | |||
8012 | SDValue VSetCC = DAG.getNode(ISD::SETCC, DL, VT, N0.getOperand(0), | |||
8013 | N0.getOperand(1), N0.getOperand(2)); | |||
8014 | return DAG.getNode(ISD::AND, DL, VT, VSetCC, VecOnes); | |||
8015 | } | |||
8016 | ||||
8017 | // If the desired elements are smaller or larger than the source | |||
8018 | // elements we can use a matching integer vector type and then | |||
8019 | // truncate/sign extend. | |||
8020 | EVT MatchingVectorType = N00VT.changeVectorElementTypeToInteger(); | |||
8021 | SDValue VsetCC = | |||
8022 | DAG.getNode(ISD::SETCC, DL, MatchingVectorType, N0.getOperand(0), | |||
8023 | N0.getOperand(1), N0.getOperand(2)); | |||
8024 | return DAG.getNode(ISD::AND, DL, VT, DAG.getSExtOrTrunc(VsetCC, DL, VT), | |||
8025 | VecOnes); | |||
8026 | } | |||
8027 | ||||
8028 | // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc | |||
8029 | SDLoc DL(N); | |||
8030 | if (SDValue SCC = SimplifySelectCC( | |||
8031 | DL, N0.getOperand(0), N0.getOperand(1), DAG.getConstant(1, DL, VT), | |||
8032 | DAG.getConstant(0, DL, VT), | |||
8033 | cast<CondCodeSDNode>(N0.getOperand(2))->get(), true)) | |||
8034 | return SCC; | |||
8035 | } | |||
8036 | ||||
8037 | // (zext (shl (zext x), cst)) -> (shl (zext x), cst) | |||
8038 | if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) && | |||
8039 | isa<ConstantSDNode>(N0.getOperand(1)) && | |||
8040 | N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND && | |||
8041 | N0.hasOneUse()) { | |||
8042 | SDValue ShAmt = N0.getOperand(1); | |||
8043 | unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue(); | |||
8044 | if (N0.getOpcode() == ISD::SHL) { | |||
8045 | SDValue InnerZExt = N0.getOperand(0); | |||
8046 | // If the original shl may be shifting out bits, do not perform this | |||
8047 | // transformation. | |||
8048 | unsigned KnownZeroBits = InnerZExt.getValueSizeInBits() - | |||
8049 | InnerZExt.getOperand(0).getValueSizeInBits(); | |||
8050 | if (ShAmtVal > KnownZeroBits) | |||
8051 | return SDValue(); | |||
8052 | } | |||
8053 | ||||
8054 | SDLoc DL(N); | |||
8055 | ||||
8056 | // Ensure that the shift amount is wide enough for the shifted value. | |||
8057 | if (VT.getSizeInBits() >= 256) | |||
8058 | ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt); | |||
8059 | ||||
8060 | return DAG.getNode(N0.getOpcode(), DL, VT, | |||
8061 | DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)), | |||
8062 | ShAmt); | |||
8063 | } | |||
8064 | ||||
8065 | if (SDValue NewVSel = matchVSelectOpSizesWithSetCC(N)) | |||
8066 | return NewVSel; | |||
8067 | ||||
8068 | return SDValue(); | |||
8069 | } | |||
8070 | ||||
8071 | SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) { | |||
8072 | SDValue N0 = N->getOperand(0); | |||
8073 | EVT VT = N->getValueType(0); | |||
8074 | ||||
8075 | if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes, | |||
8076 | LegalOperations)) | |||
8077 | return SDValue(Res, 0); | |||
8078 | ||||
8079 | // fold (aext (aext x)) -> (aext x) | |||
8080 | // fold (aext (zext x)) -> (zext x) | |||
8081 | // fold (aext (sext x)) -> (sext x) | |||
8082 | if (N0.getOpcode() == ISD::ANY_EXTEND || | |||
8083 | N0.getOpcode() == ISD::ZERO_EXTEND || | |||
8084 | N0.getOpcode() == ISD::SIGN_EXTEND) | |||
8085 | return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0)); | |||
8086 | ||||
8087 | // fold (aext (truncate (load x))) -> (aext (smaller load x)) | |||
8088 | // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n))) | |||
8089 | if (N0.getOpcode() == ISD::TRUNCATE) { | |||
8090 | if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) { | |||
8091 | SDNode *oye = N0.getOperand(0).getNode(); | |||
8092 | if (NarrowLoad.getNode() != N0.getNode()) { | |||
8093 | CombineTo(N0.getNode(), NarrowLoad); | |||
8094 | // CombineTo deleted the truncate, if needed, but not what's under it. | |||
8095 | AddToWorklist(oye); | |||
8096 | } | |||
8097 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
8098 | } | |||
8099 | } | |||
8100 | ||||
8101 | // fold (aext (truncate x)) | |||
8102 | if (N0.getOpcode() == ISD::TRUNCATE) | |||
8103 | return DAG.getAnyExtOrTrunc(N0.getOperand(0), SDLoc(N), VT); | |||
8104 | ||||
8105 | // Fold (aext (and (trunc x), cst)) -> (and x, cst) | |||
8106 | // if the trunc is not free. | |||
8107 | if (N0.getOpcode() == ISD::AND && | |||
8108 | N0.getOperand(0).getOpcode() == ISD::TRUNCATE && | |||
8109 | N0.getOperand(1).getOpcode() == ISD::Constant && | |||
8110 | !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(), | |||
8111 | N0.getValueType())) { | |||
8112 | SDLoc DL(N); | |||
8113 | SDValue X = N0.getOperand(0).getOperand(0); | |||
8114 | X = DAG.getAnyExtOrTrunc(X, DL, VT); | |||
8115 | APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); | |||
8116 | Mask = Mask.zext(VT.getSizeInBits()); | |||
8117 | return DAG.getNode(ISD::AND, DL, VT, | |||
8118 | X, DAG.getConstant(Mask, DL, VT)); | |||
8119 | } | |||
8120 | ||||
8121 | // fold (aext (load x)) -> (aext (truncate (extload x))) | |||
8122 | // None of the supported targets knows how to perform load and any_ext | |||
8123 | // on vectors in one instruction. We only perform this transformation on | |||
8124 | // scalars. | |||
8125 | if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() && | |||
8126 | ISD::isUNINDEXEDLoad(N0.getNode()) && | |||
8127 | TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) { | |||
8128 | bool DoXform = true; | |||
8129 | SmallVector<SDNode*, 4> SetCCs; | |||
8130 | if (!N0.hasOneUse()) | |||
8131 | DoXform = ExtendUsesToFormExtLoad(VT, N, N0, ISD::ANY_EXTEND, SetCCs, | |||
8132 | TLI); | |||
8133 | if (DoXform) { | |||
8134 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
8135 | SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT, | |||
8136 | LN0->getChain(), | |||
8137 | LN0->getBasePtr(), N0.getValueType(), | |||
8138 | LN0->getMemOperand()); | |||
8139 | ExtendSetCCUses(SetCCs, N0, ExtLoad, SDLoc(N), | |||
8140 | ISD::ANY_EXTEND); | |||
8141 | // If the load value is used only by N, replace it via CombineTo N. | |||
8142 | bool NoReplaceTrunc = N0.hasOneUse(); | |||
8143 | CombineTo(N, ExtLoad); | |||
8144 | if (NoReplaceTrunc) { | |||
8145 | DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), ExtLoad.getValue(1)); | |||
8146 | } else { | |||
8147 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), | |||
8148 | N0.getValueType(), ExtLoad); | |||
8149 | CombineTo(LN0, Trunc, ExtLoad.getValue(1)); | |||
8150 | } | |||
8151 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
8152 | } | |||
8153 | } | |||
8154 | ||||
8155 | // fold (aext (zextload x)) -> (aext (truncate (zextload x))) | |||
8156 | // fold (aext (sextload x)) -> (aext (truncate (sextload x))) | |||
8157 | // fold (aext ( extload x)) -> (aext (truncate (extload x))) | |||
8158 | if (N0.getOpcode() == ISD::LOAD && !ISD::isNON_EXTLoad(N0.getNode()) && | |||
8159 | ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) { | |||
8160 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
8161 | ISD::LoadExtType ExtType = LN0->getExtensionType(); | |||
8162 | EVT MemVT = LN0->getMemoryVT(); | |||
8163 | if (!LegalOperations || TLI.isLoadExtLegal(ExtType, VT, MemVT)) { | |||
8164 | SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N), | |||
8165 | VT, LN0->getChain(), LN0->getBasePtr(), | |||
8166 | MemVT, LN0->getMemOperand()); | |||
8167 | CombineTo(N, ExtLoad); | |||
8168 | DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), ExtLoad.getValue(1)); | |||
8169 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
8170 | } | |||
8171 | } | |||
8172 | ||||
8173 | if (N0.getOpcode() == ISD::SETCC) { | |||
8174 | // For vectors: | |||
8175 | // aext(setcc) -> vsetcc | |||
8176 | // aext(setcc) -> truncate(vsetcc) | |||
8177 | // aext(setcc) -> aext(vsetcc) | |||
8178 | // Only do this before legalize for now. | |||
8179 | if (VT.isVector() && !LegalOperations) { | |||
8180 | EVT N00VT = N0.getOperand(0).getValueType(); | |||
8181 | if (getSetCCResultType(N00VT) == N0.getValueType()) | |||
8182 | return SDValue(); | |||
8183 | ||||
8184 | // We know that the # elements of the results is the same as the | |||
8185 | // # elements of the compare (and the # elements of the compare result | |||
8186 | // for that matter). Check to see that they are the same size. If so, | |||
8187 | // we know that the element size of the sext'd result matches the | |||
8188 | // element size of the compare operands. | |||
8189 | if (VT.getSizeInBits() == N00VT.getSizeInBits()) | |||
8190 | return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0), | |||
8191 | N0.getOperand(1), | |||
8192 | cast<CondCodeSDNode>(N0.getOperand(2))->get()); | |||
8193 | // If the desired elements are smaller or larger than the source | |||
8194 | // elements we can use a matching integer vector type and then | |||
8195 | // truncate/any extend | |||
8196 | else { | |||
8197 | EVT MatchingVectorType = N00VT.changeVectorElementTypeToInteger(); | |||
8198 | SDValue VsetCC = | |||
8199 | DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0), | |||
8200 | N0.getOperand(1), | |||
8201 | cast<CondCodeSDNode>(N0.getOperand(2))->get()); | |||
8202 | return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT); | |||
8203 | } | |||
8204 | } | |||
8205 | ||||
8206 | // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc | |||
8207 | SDLoc DL(N); | |||
8208 | if (SDValue SCC = SimplifySelectCC( | |||
8209 | DL, N0.getOperand(0), N0.getOperand(1), DAG.getConstant(1, DL, VT), | |||
8210 | DAG.getConstant(0, DL, VT), | |||
8211 | cast<CondCodeSDNode>(N0.getOperand(2))->get(), true)) | |||
8212 | return SCC; | |||
8213 | } | |||
8214 | ||||
8215 | return SDValue(); | |||
8216 | } | |||
8217 | ||||
8218 | SDValue DAGCombiner::visitAssertExt(SDNode *N) { | |||
8219 | unsigned Opcode = N->getOpcode(); | |||
8220 | SDValue N0 = N->getOperand(0); | |||
8221 | SDValue N1 = N->getOperand(1); | |||
8222 | EVT AssertVT = cast<VTSDNode>(N1)->getVT(); | |||
8223 | ||||
8224 | // fold (assert?ext (assert?ext x, vt), vt) -> (assert?ext x, vt) | |||
8225 | if (N0.getOpcode() == Opcode && | |||
8226 | AssertVT == cast<VTSDNode>(N0.getOperand(1))->getVT()) | |||
8227 | return N0; | |||
8228 | ||||
8229 | if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() && | |||
8230 | N0.getOperand(0).getOpcode() == Opcode) { | |||
8231 | // We have an assert, truncate, assert sandwich. Make one stronger assert | |||
8232 | // by asserting on the smallest asserted type to the larger source type. | |||
8233 | // This eliminates the later assert: | |||
8234 | // assert (trunc (assert X, i8) to iN), i1 --> trunc (assert X, i1) to iN | |||
8235 | // assert (trunc (assert X, i1) to iN), i8 --> trunc (assert X, i1) to iN | |||
8236 | SDValue BigA = N0.getOperand(0); | |||
8237 | EVT BigA_AssertVT = cast<VTSDNode>(BigA.getOperand(1))->getVT(); | |||
8238 | assert(BigA_AssertVT.bitsLE(N0.getValueType()) &&(static_cast <bool> (BigA_AssertVT.bitsLE(N0.getValueType ()) && "Asserting zero/sign-extended bits to a type larger than the " "truncated destination does not provide information") ? void (0) : __assert_fail ("BigA_AssertVT.bitsLE(N0.getValueType()) && \"Asserting zero/sign-extended bits to a type larger than the \" \"truncated destination does not provide information\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 8240, __extension__ __PRETTY_FUNCTION__)) | |||
8239 | "Asserting zero/sign-extended bits to a type larger than the "(static_cast <bool> (BigA_AssertVT.bitsLE(N0.getValueType ()) && "Asserting zero/sign-extended bits to a type larger than the " "truncated destination does not provide information") ? void (0) : __assert_fail ("BigA_AssertVT.bitsLE(N0.getValueType()) && \"Asserting zero/sign-extended bits to a type larger than the \" \"truncated destination does not provide information\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 8240, __extension__ __PRETTY_FUNCTION__)) | |||
8240 | "truncated destination does not provide information")(static_cast <bool> (BigA_AssertVT.bitsLE(N0.getValueType ()) && "Asserting zero/sign-extended bits to a type larger than the " "truncated destination does not provide information") ? void (0) : __assert_fail ("BigA_AssertVT.bitsLE(N0.getValueType()) && \"Asserting zero/sign-extended bits to a type larger than the \" \"truncated destination does not provide information\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 8240, __extension__ __PRETTY_FUNCTION__)); | |||
8241 | ||||
8242 | SDLoc DL(N); | |||
8243 | EVT MinAssertVT = AssertVT.bitsLT(BigA_AssertVT) ? AssertVT : BigA_AssertVT; | |||
8244 | SDValue MinAssertVTVal = DAG.getValueType(MinAssertVT); | |||
8245 | SDValue NewAssert = DAG.getNode(Opcode, DL, BigA.getValueType(), | |||
8246 | BigA.getOperand(0), MinAssertVTVal); | |||
8247 | return DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), NewAssert); | |||
8248 | } | |||
8249 | ||||
8250 | return SDValue(); | |||
8251 | } | |||
8252 | ||||
8253 | /// If the result of a wider load is shifted to right of N bits and then | |||
8254 | /// truncated to a narrower type and where N is a multiple of number of bits of | |||
8255 | /// the narrower type, transform it to a narrower load from address + N / num of | |||
8256 | /// bits of new type. Also narrow the load if the result is masked with an AND | |||
8257 | /// to effectively produce a smaller type. If the result is to be extended, also | |||
8258 | /// fold the extension to form a extending load. | |||
8259 | SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) { | |||
8260 | unsigned Opc = N->getOpcode(); | |||
8261 | ||||
8262 | ISD::LoadExtType ExtType = ISD::NON_EXTLOAD; | |||
8263 | SDValue N0 = N->getOperand(0); | |||
8264 | EVT VT = N->getValueType(0); | |||
8265 | EVT ExtVT = VT; | |||
8266 | ||||
8267 | // This transformation isn't valid for vector loads. | |||
8268 | if (VT.isVector()) | |||
8269 | return SDValue(); | |||
8270 | ||||
8271 | // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then | |||
8272 | // extended to VT. | |||
8273 | if (Opc == ISD::SIGN_EXTEND_INREG) { | |||
8274 | ExtType = ISD::SEXTLOAD; | |||
8275 | ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT(); | |||
8276 | } else if (Opc == ISD::SRL) { | |||
8277 | // Another special-case: SRL is basically zero-extending a narrower value, | |||
8278 | // or it maybe shifting a higher subword, half or byte into the lowest | |||
8279 | // bits. | |||
8280 | ExtType = ISD::ZEXTLOAD; | |||
8281 | N0 = SDValue(N, 0); | |||
8282 | ||||
8283 | auto *LN0 = dyn_cast<LoadSDNode>(N0.getOperand(0)); | |||
8284 | auto *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1)); | |||
8285 | if (!N01 || !LN0) | |||
8286 | return SDValue(); | |||
8287 | ||||
8288 | uint64_t ShiftAmt = N01->getZExtValue(); | |||
8289 | uint64_t MemoryWidth = LN0->getMemoryVT().getSizeInBits(); | |||
8290 | if (LN0->getExtensionType() != ISD::SEXTLOAD && MemoryWidth > ShiftAmt) | |||
8291 | ExtVT = EVT::getIntegerVT(*DAG.getContext(), MemoryWidth - ShiftAmt); | |||
8292 | else | |||
8293 | ExtVT = EVT::getIntegerVT(*DAG.getContext(), | |||
8294 | VT.getSizeInBits() - ShiftAmt); | |||
8295 | } else if (Opc == ISD::AND) { | |||
8296 | // An AND with a constant mask is the same as a truncate + zero-extend. | |||
8297 | auto AndC = dyn_cast<ConstantSDNode>(N->getOperand(1)); | |||
8298 | if (!AndC || !AndC->getAPIntValue().isMask()) | |||
8299 | return SDValue(); | |||
8300 | ||||
8301 | unsigned ActiveBits = AndC->getAPIntValue().countTrailingOnes(); | |||
8302 | ExtType = ISD::ZEXTLOAD; | |||
8303 | ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits); | |||
8304 | } | |||
8305 | ||||
8306 | unsigned ShAmt = 0; | |||
8307 | if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) { | |||
8308 | if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { | |||
8309 | ShAmt = N01->getZExtValue(); | |||
8310 | unsigned EVTBits = ExtVT.getSizeInBits(); | |||
8311 | // Is the shift amount a multiple of size of VT? | |||
8312 | if ((ShAmt & (EVTBits-1)) == 0) { | |||
8313 | N0 = N0.getOperand(0); | |||
8314 | // Is the load width a multiple of size of VT? | |||
8315 | if ((N0.getValueSizeInBits() & (EVTBits-1)) != 0) | |||
8316 | return SDValue(); | |||
8317 | } | |||
8318 | ||||
8319 | // At this point, we must have a load or else we can't do the transform. | |||
8320 | if (!isa<LoadSDNode>(N0)) return SDValue(); | |||
8321 | ||||
8322 | // Because a SRL must be assumed to *need* to zero-extend the high bits | |||
8323 | // (as opposed to anyext the high bits), we can't combine the zextload | |||
8324 | // lowering of SRL and an sextload. | |||
8325 | if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD) | |||
8326 | return SDValue(); | |||
8327 | ||||
8328 | // If the shift amount is larger than the input type then we're not | |||
8329 | // accessing any of the loaded bytes. If the load was a zextload/extload | |||
8330 | // then the result of the shift+trunc is zero/undef (handled elsewhere). | |||
8331 | if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits()) | |||
8332 | return SDValue(); | |||
8333 | } | |||
8334 | } | |||
8335 | ||||
8336 | // If the load is shifted left (and the result isn't shifted back right), | |||
8337 | // we can fold the truncate through the shift. | |||
8338 | unsigned ShLeftAmt = 0; | |||
8339 | if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() && | |||
8340 | ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) { | |||
8341 | if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { | |||
8342 | ShLeftAmt = N01->getZExtValue(); | |||
8343 | N0 = N0.getOperand(0); | |||
8344 | } | |||
8345 | } | |||
8346 | ||||
8347 | // If we haven't found a load, we can't narrow it. | |||
8348 | if (!isa<LoadSDNode>(N0)) | |||
8349 | return SDValue(); | |||
8350 | ||||
8351 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
8352 | if (!isLegalNarrowLoad(LN0, ExtType, ExtVT, ShAmt)) | |||
8353 | return SDValue(); | |||
8354 | ||||
8355 | // For big endian targets, we need to adjust the offset to the pointer to | |||
8356 | // load the correct bytes. | |||
8357 | if (DAG.getDataLayout().isBigEndian()) { | |||
8358 | unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits(); | |||
8359 | unsigned EVTStoreBits = ExtVT.getStoreSizeInBits(); | |||
8360 | ShAmt = LVTStoreBits - EVTStoreBits - ShAmt; | |||
8361 | } | |||
8362 | ||||
8363 | EVT PtrType = N0.getOperand(1).getValueType(); | |||
8364 | uint64_t PtrOff = ShAmt / 8; | |||
8365 | unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff); | |||
8366 | SDLoc DL(LN0); | |||
8367 | // The original load itself didn't wrap, so an offset within it doesn't. | |||
8368 | SDNodeFlags Flags; | |||
8369 | Flags.setNoUnsignedWrap(true); | |||
8370 | SDValue NewPtr = DAG.getNode(ISD::ADD, DL, | |||
8371 | PtrType, LN0->getBasePtr(), | |||
8372 | DAG.getConstant(PtrOff, DL, PtrType), | |||
8373 | Flags); | |||
8374 | AddToWorklist(NewPtr.getNode()); | |||
8375 | ||||
8376 | SDValue Load; | |||
8377 | if (ExtType == ISD::NON_EXTLOAD) | |||
8378 | Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr, | |||
8379 | LN0->getPointerInfo().getWithOffset(PtrOff), NewAlign, | |||
8380 | LN0->getMemOperand()->getFlags(), LN0->getAAInfo()); | |||
8381 | else | |||
8382 | Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(), NewPtr, | |||
8383 | LN0->getPointerInfo().getWithOffset(PtrOff), ExtVT, | |||
8384 | NewAlign, LN0->getMemOperand()->getFlags(), | |||
8385 | LN0->getAAInfo()); | |||
8386 | ||||
8387 | // Replace the old load's chain with the new load's chain. | |||
8388 | WorklistRemover DeadNodes(*this); | |||
8389 | DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1)); | |||
8390 | ||||
8391 | // Shift the result left, if we've swallowed a left shift. | |||
8392 | SDValue Result = Load; | |||
8393 | if (ShLeftAmt != 0) { | |||
8394 | EVT ShImmTy = getShiftAmountTy(Result.getValueType()); | |||
8395 | if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt)) | |||
8396 | ShImmTy = VT; | |||
8397 | // If the shift amount is as large as the result size (but, presumably, | |||
8398 | // no larger than the source) then the useful bits of the result are | |||
8399 | // zero; we can't simply return the shortened shift, because the result | |||
8400 | // of that operation is undefined. | |||
8401 | SDLoc DL(N0); | |||
8402 | if (ShLeftAmt >= VT.getSizeInBits()) | |||
8403 | Result = DAG.getConstant(0, DL, VT); | |||
8404 | else | |||
8405 | Result = DAG.getNode(ISD::SHL, DL, VT, | |||
8406 | Result, DAG.getConstant(ShLeftAmt, DL, ShImmTy)); | |||
8407 | } | |||
8408 | ||||
8409 | // Return the new loaded value. | |||
8410 | return Result; | |||
8411 | } | |||
8412 | ||||
8413 | SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) { | |||
8414 | SDValue N0 = N->getOperand(0); | |||
8415 | SDValue N1 = N->getOperand(1); | |||
8416 | EVT VT = N->getValueType(0); | |||
8417 | EVT EVT = cast<VTSDNode>(N1)->getVT(); | |||
8418 | unsigned VTBits = VT.getScalarSizeInBits(); | |||
8419 | unsigned EVTBits = EVT.getScalarSizeInBits(); | |||
8420 | ||||
8421 | if (N0.isUndef()) | |||
8422 | return DAG.getUNDEF(VT); | |||
8423 | ||||
8424 | // fold (sext_in_reg c1) -> c1 | |||
8425 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) | |||
8426 | return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1); | |||
8427 | ||||
8428 | // If the input is already sign extended, just drop the extension. | |||
8429 | if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1) | |||
8430 | return N0; | |||
8431 | ||||
8432 | // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2 | |||
8433 | if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && | |||
8434 | EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT())) | |||
8435 | return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, | |||
8436 | N0.getOperand(0), N1); | |||
8437 | ||||
8438 | // fold (sext_in_reg (sext x)) -> (sext x) | |||
8439 | // fold (sext_in_reg (aext x)) -> (sext x) | |||
8440 | // if x is small enough. | |||
8441 | if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) { | |||
8442 | SDValue N00 = N0.getOperand(0); | |||
8443 | if (N00.getScalarValueSizeInBits() <= EVTBits && | |||
8444 | (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT))) | |||
8445 | return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1); | |||
8446 | } | |||
8447 | ||||
8448 | // fold (sext_in_reg (*_extend_vector_inreg x)) -> (sext_vector_in_reg x) | |||
8449 | if ((N0.getOpcode() == ISD::ANY_EXTEND_VECTOR_INREG || | |||
8450 | N0.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG || | |||
8451 | N0.getOpcode() == ISD::ZERO_EXTEND_VECTOR_INREG) && | |||
8452 | N0.getOperand(0).getScalarValueSizeInBits() == EVTBits) { | |||
8453 | if (!LegalOperations || | |||
8454 | TLI.isOperationLegal(ISD::SIGN_EXTEND_VECTOR_INREG, VT)) | |||
8455 | return DAG.getSignExtendVectorInReg(N0.getOperand(0), SDLoc(N), VT); | |||
8456 | } | |||
8457 | ||||
8458 | // fold (sext_in_reg (zext x)) -> (sext x) | |||
8459 | // iff we are extending the source sign bit. | |||
8460 | if (N0.getOpcode() == ISD::ZERO_EXTEND) { | |||
8461 | SDValue N00 = N0.getOperand(0); | |||
8462 | if (N00.getScalarValueSizeInBits() == EVTBits && | |||
8463 | (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT))) | |||
8464 | return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1); | |||
8465 | } | |||
8466 | ||||
8467 | // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero. | |||
8468 | if (DAG.MaskedValueIsZero(N0, APInt::getOneBitSet(VTBits, EVTBits - 1))) | |||
8469 | return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT.getScalarType()); | |||
8470 | ||||
8471 | // fold operands of sext_in_reg based on knowledge that the top bits are not | |||
8472 | // demanded. | |||
8473 | if (SimplifyDemandedBits(SDValue(N, 0))) | |||
8474 | return SDValue(N, 0); | |||
8475 | ||||
8476 | // fold (sext_in_reg (load x)) -> (smaller sextload x) | |||
8477 | // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits)) | |||
8478 | if (SDValue NarrowLoad = ReduceLoadWidth(N)) | |||
8479 | return NarrowLoad; | |||
8480 | ||||
8481 | // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24) | |||
8482 | // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible. | |||
8483 | // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above. | |||
8484 | if (N0.getOpcode() == ISD::SRL) { | |||
8485 | if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1))) | |||
8486 | if (ShAmt->getZExtValue()+EVTBits <= VTBits) { | |||
8487 | // We can turn this into an SRA iff the input to the SRL is already sign | |||
8488 | // extended enough. | |||
8489 | unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0)); | |||
8490 | if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits) | |||
8491 | return DAG.getNode(ISD::SRA, SDLoc(N), VT, | |||
8492 | N0.getOperand(0), N0.getOperand(1)); | |||
8493 | } | |||
8494 | } | |||
8495 | ||||
8496 | // fold (sext_inreg (extload x)) -> (sextload x) | |||
8497 | // If sextload is not supported by target, we can only do the combine when | |||
8498 | // load has one use. Doing otherwise can block folding the extload with other | |||
8499 | // extends that the target does support. | |||
8500 | if (ISD::isEXTLoad(N0.getNode()) && | |||
8501 | ISD::isUNINDEXEDLoad(N0.getNode()) && | |||
8502 | EVT == cast<LoadSDNode>(N0)->getMemoryVT() && | |||
8503 | ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile() && | |||
8504 | N0.hasOneUse()) || | |||
8505 | TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) { | |||
8506 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
8507 | SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT, | |||
8508 | LN0->getChain(), | |||
8509 | LN0->getBasePtr(), EVT, | |||
8510 | LN0->getMemOperand()); | |||
8511 | CombineTo(N, ExtLoad); | |||
8512 | CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); | |||
8513 | AddToWorklist(ExtLoad.getNode()); | |||
8514 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
8515 | } | |||
8516 | // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use | |||
8517 | if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) && | |||
8518 | N0.hasOneUse() && | |||
8519 | EVT == cast<LoadSDNode>(N0)->getMemoryVT() && | |||
8520 | ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) || | |||
8521 | TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) { | |||
8522 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
8523 | SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT, | |||
8524 | LN0->getChain(), | |||
8525 | LN0->getBasePtr(), EVT, | |||
8526 | LN0->getMemOperand()); | |||
8527 | CombineTo(N, ExtLoad); | |||
8528 | CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); | |||
8529 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
8530 | } | |||
8531 | ||||
8532 | // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16)) | |||
8533 | if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) { | |||
8534 | if (SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0), | |||
8535 | N0.getOperand(1), false)) | |||
8536 | return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, | |||
8537 | BSwap, N1); | |||
8538 | } | |||
8539 | ||||
8540 | return SDValue(); | |||
8541 | } | |||
8542 | ||||
8543 | SDValue DAGCombiner::visitSIGN_EXTEND_VECTOR_INREG(SDNode *N) { | |||
8544 | SDValue N0 = N->getOperand(0); | |||
8545 | EVT VT = N->getValueType(0); | |||
8546 | ||||
8547 | if (N0.isUndef()) | |||
8548 | return DAG.getUNDEF(VT); | |||
8549 | ||||
8550 | if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes, | |||
8551 | LegalOperations)) | |||
8552 | return SDValue(Res, 0); | |||
8553 | ||||
8554 | return SDValue(); | |||
8555 | } | |||
8556 | ||||
8557 | SDValue DAGCombiner::visitZERO_EXTEND_VECTOR_INREG(SDNode *N) { | |||
8558 | SDValue N0 = N->getOperand(0); | |||
8559 | EVT VT = N->getValueType(0); | |||
8560 | ||||
8561 | if (N0.isUndef()) | |||
8562 | return DAG.getUNDEF(VT); | |||
8563 | ||||
8564 | if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes, | |||
8565 | LegalOperations)) | |||
8566 | return SDValue(Res, 0); | |||
8567 | ||||
8568 | return SDValue(); | |||
8569 | } | |||
8570 | ||||
8571 | SDValue DAGCombiner::visitTRUNCATE(SDNode *N) { | |||
8572 | SDValue N0 = N->getOperand(0); | |||
8573 | EVT VT = N->getValueType(0); | |||
8574 | bool isLE = DAG.getDataLayout().isLittleEndian(); | |||
8575 | ||||
8576 | // noop truncate | |||
8577 | if (N0.getValueType() == N->getValueType(0)) | |||
8578 | return N0; | |||
8579 | ||||
8580 | // fold (truncate (truncate x)) -> (truncate x) | |||
8581 | if (N0.getOpcode() == ISD::TRUNCATE) | |||
8582 | return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0)); | |||
8583 | ||||
8584 | // fold (truncate c1) -> c1 | |||
8585 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) { | |||
8586 | SDValue C = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0); | |||
8587 | if (C.getNode() != N) | |||
8588 | return C; | |||
8589 | } | |||
8590 | ||||
8591 | // fold (truncate (ext x)) -> (ext x) or (truncate x) or x | |||
8592 | if (N0.getOpcode() == ISD::ZERO_EXTEND || | |||
8593 | N0.getOpcode() == ISD::SIGN_EXTEND || | |||
8594 | N0.getOpcode() == ISD::ANY_EXTEND) { | |||
8595 | // if the source is smaller than the dest, we still need an extend. | |||
8596 | if (N0.getOperand(0).getValueType().bitsLT(VT)) | |||
8597 | return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0)); | |||
8598 | // if the source is larger than the dest, than we just need the truncate. | |||
8599 | if (N0.getOperand(0).getValueType().bitsGT(VT)) | |||
8600 | return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0)); | |||
8601 | // if the source and dest are the same type, we can drop both the extend | |||
8602 | // and the truncate. | |||
8603 | return N0.getOperand(0); | |||
8604 | } | |||
8605 | ||||
8606 | // If this is anyext(trunc), don't fold it, allow ourselves to be folded. | |||
8607 | if (N->hasOneUse() && (N->use_begin()->getOpcode() == ISD::ANY_EXTEND)) | |||
8608 | return SDValue(); | |||
8609 | ||||
8610 | // Fold extract-and-trunc into a narrow extract. For example: | |||
8611 | // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1) | |||
8612 | // i32 y = TRUNCATE(i64 x) | |||
8613 | // -- becomes -- | |||
8614 | // v16i8 b = BITCAST (v2i64 val) | |||
8615 | // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8) | |||
8616 | // | |||
8617 | // Note: We only run this optimization after type legalization (which often | |||
8618 | // creates this pattern) and before operation legalization after which | |||
8619 | // we need to be more careful about the vector instructions that we generate. | |||
8620 | if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT && | |||
8621 | LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) { | |||
8622 | EVT VecTy = N0.getOperand(0).getValueType(); | |||
8623 | EVT ExTy = N0.getValueType(); | |||
8624 | EVT TrTy = N->getValueType(0); | |||
8625 | ||||
8626 | unsigned NumElem = VecTy.getVectorNumElements(); | |||
8627 | unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits(); | |||
8628 | ||||
8629 | EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem); | |||
8630 | assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size")(static_cast <bool> (NVT.getSizeInBits() == VecTy.getSizeInBits () && "Invalid Size") ? void (0) : __assert_fail ("NVT.getSizeInBits() == VecTy.getSizeInBits() && \"Invalid Size\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 8630, __extension__ __PRETTY_FUNCTION__)); | |||
8631 | ||||
8632 | SDValue EltNo = N0->getOperand(1); | |||
8633 | if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) { | |||
8634 | int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue(); | |||
8635 | EVT IndexTy = TLI.getVectorIdxTy(DAG.getDataLayout()); | |||
8636 | int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1)); | |||
8637 | ||||
8638 | SDLoc DL(N); | |||
8639 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, TrTy, | |||
8640 | DAG.getBitcast(NVT, N0.getOperand(0)), | |||
8641 | DAG.getConstant(Index, DL, IndexTy)); | |||
8642 | } | |||
8643 | } | |||
8644 | ||||
8645 | // trunc (select c, a, b) -> select c, (trunc a), (trunc b) | |||
8646 | if (N0.getOpcode() == ISD::SELECT && N0.hasOneUse()) { | |||
8647 | EVT SrcVT = N0.getValueType(); | |||
8648 | if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) && | |||
8649 | TLI.isTruncateFree(SrcVT, VT)) { | |||
8650 | SDLoc SL(N0); | |||
8651 | SDValue Cond = N0.getOperand(0); | |||
8652 | SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1)); | |||
8653 | SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2)); | |||
8654 | return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1); | |||
8655 | } | |||
8656 | } | |||
8657 | ||||
8658 | // trunc (shl x, K) -> shl (trunc x), K => K < VT.getScalarSizeInBits() | |||
8659 | if (N0.getOpcode() == ISD::SHL && N0.hasOneUse() && | |||
8660 | (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::SHL, VT)) && | |||
8661 | TLI.isTypeDesirableForOp(ISD::SHL, VT)) { | |||
8662 | SDValue Amt = N0.getOperand(1); | |||
8663 | KnownBits Known; | |||
8664 | DAG.computeKnownBits(Amt, Known); | |||
8665 | unsigned Size = VT.getScalarSizeInBits(); | |||
8666 | if (Known.getBitWidth() - Known.countMinLeadingZeros() <= Log2_32(Size)) { | |||
8667 | SDLoc SL(N); | |||
8668 | EVT AmtVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout()); | |||
8669 | ||||
8670 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(0)); | |||
8671 | if (AmtVT != Amt.getValueType()) { | |||
8672 | Amt = DAG.getZExtOrTrunc(Amt, SL, AmtVT); | |||
8673 | AddToWorklist(Amt.getNode()); | |||
8674 | } | |||
8675 | return DAG.getNode(ISD::SHL, SL, VT, Trunc, Amt); | |||
8676 | } | |||
8677 | } | |||
8678 | ||||
8679 | // Fold a series of buildvector, bitcast, and truncate if possible. | |||
8680 | // For example fold | |||
8681 | // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to | |||
8682 | // (2xi32 (buildvector x, y)). | |||
8683 | if (Level == AfterLegalizeVectorOps && VT.isVector() && | |||
8684 | N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() && | |||
8685 | N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR && | |||
8686 | N0.getOperand(0).hasOneUse()) { | |||
8687 | SDValue BuildVect = N0.getOperand(0); | |||
8688 | EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType(); | |||
8689 | EVT TruncVecEltTy = VT.getVectorElementType(); | |||
8690 | ||||
8691 | // Check that the element types match. | |||
8692 | if (BuildVectEltTy == TruncVecEltTy) { | |||
8693 | // Now we only need to compute the offset of the truncated elements. | |||
8694 | unsigned BuildVecNumElts = BuildVect.getNumOperands(); | |||
8695 | unsigned TruncVecNumElts = VT.getVectorNumElements(); | |||
8696 | unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts; | |||
8697 | ||||
8698 | assert((BuildVecNumElts % TruncVecNumElts) == 0 &&(static_cast <bool> ((BuildVecNumElts % TruncVecNumElts ) == 0 && "Invalid number of elements") ? void (0) : __assert_fail ("(BuildVecNumElts % TruncVecNumElts) == 0 && \"Invalid number of elements\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 8699, __extension__ __PRETTY_FUNCTION__)) | |||
8699 | "Invalid number of elements")(static_cast <bool> ((BuildVecNumElts % TruncVecNumElts ) == 0 && "Invalid number of elements") ? void (0) : __assert_fail ("(BuildVecNumElts % TruncVecNumElts) == 0 && \"Invalid number of elements\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 8699, __extension__ __PRETTY_FUNCTION__)); | |||
8700 | ||||
8701 | SmallVector<SDValue, 8> Opnds; | |||
8702 | for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset) | |||
8703 | Opnds.push_back(BuildVect.getOperand(i)); | |||
8704 | ||||
8705 | return DAG.getBuildVector(VT, SDLoc(N), Opnds); | |||
8706 | } | |||
8707 | } | |||
8708 | ||||
8709 | // See if we can simplify the input to this truncate through knowledge that | |||
8710 | // only the low bits are being used. | |||
8711 | // For example "trunc (or (shl x, 8), y)" // -> trunc y | |||
8712 | // Currently we only perform this optimization on scalars because vectors | |||
8713 | // may have different active low bits. | |||
8714 | if (!VT.isVector()) { | |||
8715 | APInt Mask = | |||
8716 | APInt::getLowBitsSet(N0.getValueSizeInBits(), VT.getSizeInBits()); | |||
8717 | if (SDValue Shorter = DAG.GetDemandedBits(N0, Mask)) | |||
8718 | return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter); | |||
8719 | } | |||
8720 | ||||
8721 | // fold (truncate (load x)) -> (smaller load x) | |||
8722 | // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits)) | |||
8723 | if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) { | |||
8724 | if (SDValue Reduced = ReduceLoadWidth(N)) | |||
8725 | return Reduced; | |||
8726 | ||||
8727 | // Handle the case where the load remains an extending load even | |||
8728 | // after truncation. | |||
8729 | if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) { | |||
8730 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
8731 | if (!LN0->isVolatile() && | |||
8732 | LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) { | |||
8733 | SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0), | |||
8734 | VT, LN0->getChain(), LN0->getBasePtr(), | |||
8735 | LN0->getMemoryVT(), | |||
8736 | LN0->getMemOperand()); | |||
8737 | DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1)); | |||
8738 | return NewLoad; | |||
8739 | } | |||
8740 | } | |||
8741 | } | |||
8742 | ||||
8743 | // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)), | |||
8744 | // where ... are all 'undef'. | |||
8745 | if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) { | |||
8746 | SmallVector<EVT, 8> VTs; | |||
8747 | SDValue V; | |||
8748 | unsigned Idx = 0; | |||
8749 | unsigned NumDefs = 0; | |||
8750 | ||||
8751 | for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) { | |||
8752 | SDValue X = N0.getOperand(i); | |||
8753 | if (!X.isUndef()) { | |||
8754 | V = X; | |||
8755 | Idx = i; | |||
8756 | NumDefs++; | |||
8757 | } | |||
8758 | // Stop if more than one members are non-undef. | |||
8759 | if (NumDefs > 1) | |||
8760 | break; | |||
8761 | VTs.push_back(EVT::getVectorVT(*DAG.getContext(), | |||
8762 | VT.getVectorElementType(), | |||
8763 | X.getValueType().getVectorNumElements())); | |||
8764 | } | |||
8765 | ||||
8766 | if (NumDefs == 0) | |||
8767 | return DAG.getUNDEF(VT); | |||
8768 | ||||
8769 | if (NumDefs == 1) { | |||
8770 | assert(V.getNode() && "The single defined operand is empty!")(static_cast <bool> (V.getNode() && "The single defined operand is empty!" ) ? void (0) : __assert_fail ("V.getNode() && \"The single defined operand is empty!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 8770, __extension__ __PRETTY_FUNCTION__)); | |||
8771 | SmallVector<SDValue, 8> Opnds; | |||
8772 | for (unsigned i = 0, e = VTs.size(); i != e; ++i) { | |||
8773 | if (i != Idx) { | |||
8774 | Opnds.push_back(DAG.getUNDEF(VTs[i])); | |||
8775 | continue; | |||
8776 | } | |||
8777 | SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V); | |||
8778 | AddToWorklist(NV.getNode()); | |||
8779 | Opnds.push_back(NV); | |||
8780 | } | |||
8781 | return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds); | |||
8782 | } | |||
8783 | } | |||
8784 | ||||
8785 | // Fold truncate of a bitcast of a vector to an extract of the low vector | |||
8786 | // element. | |||
8787 | // | |||
8788 | // e.g. trunc (i64 (bitcast v2i32:x)) -> extract_vector_elt v2i32:x, idx | |||
8789 | if (N0.getOpcode() == ISD::BITCAST && !VT.isVector()) { | |||
8790 | SDValue VecSrc = N0.getOperand(0); | |||
8791 | EVT SrcVT = VecSrc.getValueType(); | |||
8792 | if (SrcVT.isVector() && SrcVT.getScalarType() == VT && | |||
8793 | (!LegalOperations || | |||
8794 | TLI.isOperationLegal(ISD::EXTRACT_VECTOR_ELT, SrcVT))) { | |||
8795 | SDLoc SL(N); | |||
8796 | ||||
8797 | EVT IdxVT = TLI.getVectorIdxTy(DAG.getDataLayout()); | |||
8798 | unsigned Idx = isLE ? 0 : SrcVT.getVectorNumElements() - 1; | |||
8799 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, VT, | |||
8800 | VecSrc, DAG.getConstant(Idx, SL, IdxVT)); | |||
8801 | } | |||
8802 | } | |||
8803 | ||||
8804 | // Simplify the operands using demanded-bits information. | |||
8805 | if (!VT.isVector() && | |||
8806 | SimplifyDemandedBits(SDValue(N, 0))) | |||
8807 | return SDValue(N, 0); | |||
8808 | ||||
8809 | // (trunc adde(X, Y, Carry)) -> (adde trunc(X), trunc(Y), Carry) | |||
8810 | // (trunc addcarry(X, Y, Carry)) -> (addcarry trunc(X), trunc(Y), Carry) | |||
8811 | // When the adde's carry is not used. | |||
8812 | if ((N0.getOpcode() == ISD::ADDE || N0.getOpcode() == ISD::ADDCARRY) && | |||
8813 | N0.hasOneUse() && !N0.getNode()->hasAnyUseOfValue(1) && | |||
8814 | (!LegalOperations || TLI.isOperationLegal(N0.getOpcode(), VT))) { | |||
8815 | SDLoc SL(N); | |||
8816 | auto X = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(0)); | |||
8817 | auto Y = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1)); | |||
8818 | auto VTs = DAG.getVTList(VT, N0->getValueType(1)); | |||
8819 | return DAG.getNode(N0.getOpcode(), SL, VTs, X, Y, N0.getOperand(2)); | |||
8820 | } | |||
8821 | ||||
8822 | // fold (truncate (extract_subvector(ext x))) -> | |||
8823 | // (extract_subvector x) | |||
8824 | // TODO: This can be generalized to cover cases where the truncate and extract | |||
8825 | // do not fully cancel each other out. | |||
8826 | if (!LegalTypes && N0.getOpcode() == ISD::EXTRACT_SUBVECTOR) { | |||
8827 | SDValue N00 = N0.getOperand(0); | |||
8828 | if (N00.getOpcode() == ISD::SIGN_EXTEND || | |||
8829 | N00.getOpcode() == ISD::ZERO_EXTEND || | |||
8830 | N00.getOpcode() == ISD::ANY_EXTEND) { | |||
8831 | if (N00.getOperand(0)->getValueType(0).getVectorElementType() == | |||
8832 | VT.getVectorElementType()) | |||
8833 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N0->getOperand(0)), VT, | |||
8834 | N00.getOperand(0), N0.getOperand(1)); | |||
8835 | } | |||
8836 | } | |||
8837 | ||||
8838 | if (SDValue NewVSel = matchVSelectOpSizesWithSetCC(N)) | |||
8839 | return NewVSel; | |||
8840 | ||||
8841 | return SDValue(); | |||
8842 | } | |||
8843 | ||||
8844 | static SDNode *getBuildPairElt(SDNode *N, unsigned i) { | |||
8845 | SDValue Elt = N->getOperand(i); | |||
8846 | if (Elt.getOpcode() != ISD::MERGE_VALUES) | |||
8847 | return Elt.getNode(); | |||
8848 | return Elt.getOperand(Elt.getResNo()).getNode(); | |||
8849 | } | |||
8850 | ||||
8851 | /// build_pair (load, load) -> load | |||
8852 | /// if load locations are consecutive. | |||
8853 | SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) { | |||
8854 | assert(N->getOpcode() == ISD::BUILD_PAIR)(static_cast <bool> (N->getOpcode() == ISD::BUILD_PAIR ) ? void (0) : __assert_fail ("N->getOpcode() == ISD::BUILD_PAIR" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 8854, __extension__ __PRETTY_FUNCTION__)); | |||
8855 | ||||
8856 | LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0)); | |||
8857 | LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1)); | |||
8858 | ||||
8859 | // A BUILD_PAIR is always having the least significant part in elt 0 and the | |||
8860 | // most significant part in elt 1. So when combining into one large load, we | |||
8861 | // need to consider the endianness. | |||
8862 | if (DAG.getDataLayout().isBigEndian()) | |||
8863 | std::swap(LD1, LD2); | |||
8864 | ||||
8865 | if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() || | |||
8866 | LD1->getAddressSpace() != LD2->getAddressSpace()) | |||
8867 | return SDValue(); | |||
8868 | EVT LD1VT = LD1->getValueType(0); | |||
8869 | unsigned LD1Bytes = LD1VT.getStoreSize(); | |||
8870 | if (ISD::isNON_EXTLoad(LD2) && LD2->hasOneUse() && | |||
8871 | DAG.areNonVolatileConsecutiveLoads(LD2, LD1, LD1Bytes, 1)) { | |||
8872 | unsigned Align = LD1->getAlignment(); | |||
8873 | unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment( | |||
8874 | VT.getTypeForEVT(*DAG.getContext())); | |||
8875 | ||||
8876 | if (NewAlign <= Align && | |||
8877 | (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) | |||
8878 | return DAG.getLoad(VT, SDLoc(N), LD1->getChain(), LD1->getBasePtr(), | |||
8879 | LD1->getPointerInfo(), Align); | |||
8880 | } | |||
8881 | ||||
8882 | return SDValue(); | |||
8883 | } | |||
8884 | ||||
8885 | static unsigned getPPCf128HiElementSelector(const SelectionDAG &DAG) { | |||
8886 | // On little-endian machines, bitcasting from ppcf128 to i128 does swap the Hi | |||
8887 | // and Lo parts; on big-endian machines it doesn't. | |||
8888 | return DAG.getDataLayout().isBigEndian() ? 1 : 0; | |||
8889 | } | |||
8890 | ||||
8891 | static SDValue foldBitcastedFPLogic(SDNode *N, SelectionDAG &DAG, | |||
8892 | const TargetLowering &TLI) { | |||
8893 | // If this is not a bitcast to an FP type or if the target doesn't have | |||
8894 | // IEEE754-compliant FP logic, we're done. | |||
8895 | EVT VT = N->getValueType(0); | |||
8896 | if (!VT.isFloatingPoint() || !TLI.hasBitPreservingFPLogic(VT)) | |||
8897 | return SDValue(); | |||
8898 | ||||
8899 | // TODO: Use splat values for the constant-checking below and remove this | |||
8900 | // restriction. | |||
8901 | SDValue N0 = N->getOperand(0); | |||
8902 | EVT SourceVT = N0.getValueType(); | |||
8903 | if (SourceVT.isVector()) | |||
8904 | return SDValue(); | |||
8905 | ||||
8906 | unsigned FPOpcode; | |||
8907 | APInt SignMask; | |||
8908 | switch (N0.getOpcode()) { | |||
8909 | case ISD::AND: | |||
8910 | FPOpcode = ISD::FABS; | |||
8911 | SignMask = ~APInt::getSignMask(SourceVT.getSizeInBits()); | |||
8912 | break; | |||
8913 | case ISD::XOR: | |||
8914 | FPOpcode = ISD::FNEG; | |||
8915 | SignMask = APInt::getSignMask(SourceVT.getSizeInBits()); | |||
8916 | break; | |||
8917 | // TODO: ISD::OR --> ISD::FNABS? | |||
8918 | default: | |||
8919 | return SDValue(); | |||
8920 | } | |||
8921 | ||||
8922 | // Fold (bitcast int (and (bitcast fp X to int), 0x7fff...) to fp) -> fabs X | |||
8923 | // Fold (bitcast int (xor (bitcast fp X to int), 0x8000...) to fp) -> fneg X | |||
8924 | SDValue LogicOp0 = N0.getOperand(0); | |||
8925 | ConstantSDNode *LogicOp1 = dyn_cast<ConstantSDNode>(N0.getOperand(1)); | |||
8926 | if (LogicOp1 && LogicOp1->getAPIntValue() == SignMask && | |||
8927 | LogicOp0.getOpcode() == ISD::BITCAST && | |||
8928 | LogicOp0->getOperand(0).getValueType() == VT) | |||
8929 | return DAG.getNode(FPOpcode, SDLoc(N), VT, LogicOp0->getOperand(0)); | |||
8930 | ||||
8931 | return SDValue(); | |||
8932 | } | |||
8933 | ||||
8934 | SDValue DAGCombiner::visitBITCAST(SDNode *N) { | |||
8935 | SDValue N0 = N->getOperand(0); | |||
8936 | EVT VT = N->getValueType(0); | |||
8937 | ||||
8938 | if (N0.isUndef()) | |||
8939 | return DAG.getUNDEF(VT); | |||
8940 | ||||
8941 | // If the input is a BUILD_VECTOR with all constant elements, fold this now. | |||
8942 | // Only do this before legalize, since afterward the target may be depending | |||
8943 | // on the bitconvert. | |||
8944 | // First check to see if this is all constant. | |||
8945 | if (!LegalTypes && | |||
8946 | N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() && | |||
8947 | VT.isVector()) { | |||
8948 | bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant(); | |||
8949 | ||||
8950 | EVT DestEltVT = N->getValueType(0).getVectorElementType(); | |||
8951 | assert(!DestEltVT.isVector() &&(static_cast <bool> (!DestEltVT.isVector() && "Element type of vector ValueType must not be vector!" ) ? void (0) : __assert_fail ("!DestEltVT.isVector() && \"Element type of vector ValueType must not be vector!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 8952, __extension__ __PRETTY_FUNCTION__)) | |||
8952 | "Element type of vector ValueType must not be vector!")(static_cast <bool> (!DestEltVT.isVector() && "Element type of vector ValueType must not be vector!" ) ? void (0) : __assert_fail ("!DestEltVT.isVector() && \"Element type of vector ValueType must not be vector!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 8952, __extension__ __PRETTY_FUNCTION__)); | |||
8953 | if (isSimple) | |||
8954 | return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT); | |||
8955 | } | |||
8956 | ||||
8957 | // If the input is a constant, let getNode fold it. | |||
8958 | if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) { | |||
8959 | // If we can't allow illegal operations, we need to check that this is just | |||
8960 | // a fp -> int or int -> conversion and that the resulting operation will | |||
8961 | // be legal. | |||
8962 | if (!LegalOperations || | |||
8963 | (isa<ConstantSDNode>(N0) && VT.isFloatingPoint() && !VT.isVector() && | |||
8964 | TLI.isOperationLegal(ISD::ConstantFP, VT)) || | |||
8965 | (isa<ConstantFPSDNode>(N0) && VT.isInteger() && !VT.isVector() && | |||
8966 | TLI.isOperationLegal(ISD::Constant, VT))) | |||
8967 | return DAG.getBitcast(VT, N0); | |||
8968 | } | |||
8969 | ||||
8970 | // (conv (conv x, t1), t2) -> (conv x, t2) | |||
8971 | if (N0.getOpcode() == ISD::BITCAST) | |||
8972 | return DAG.getBitcast(VT, N0.getOperand(0)); | |||
8973 | ||||
8974 | // fold (conv (load x)) -> (load (conv*)x) | |||
8975 | // If the resultant load doesn't need a higher alignment than the original! | |||
8976 | if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() && | |||
8977 | // Do not change the width of a volatile load. | |||
8978 | !cast<LoadSDNode>(N0)->isVolatile() && | |||
8979 | // Do not remove the cast if the types differ in endian layout. | |||
8980 | TLI.hasBigEndianPartOrdering(N0.getValueType(), DAG.getDataLayout()) == | |||
8981 | TLI.hasBigEndianPartOrdering(VT, DAG.getDataLayout()) && | |||
8982 | (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) && | |||
8983 | TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) { | |||
8984 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
8985 | unsigned OrigAlign = LN0->getAlignment(); | |||
8986 | ||||
8987 | bool Fast = false; | |||
8988 | if (TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT, | |||
8989 | LN0->getAddressSpace(), OrigAlign, &Fast) && | |||
8990 | Fast) { | |||
8991 | SDValue Load = | |||
8992 | DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(), | |||
8993 | LN0->getPointerInfo(), OrigAlign, | |||
8994 | LN0->getMemOperand()->getFlags(), LN0->getAAInfo()); | |||
8995 | DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1)); | |||
8996 | return Load; | |||
8997 | } | |||
8998 | } | |||
8999 | ||||
9000 | if (SDValue V = foldBitcastedFPLogic(N, DAG, TLI)) | |||
9001 | return V; | |||
9002 | ||||
9003 | // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit) | |||
9004 | // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit)) | |||
9005 | // | |||
9006 | // For ppc_fp128: | |||
9007 | // fold (bitcast (fneg x)) -> | |||
9008 | // flipbit = signbit | |||
9009 | // (xor (bitcast x) (build_pair flipbit, flipbit)) | |||
9010 | // | |||
9011 | // fold (bitcast (fabs x)) -> | |||
9012 | // flipbit = (and (extract_element (bitcast x), 0), signbit) | |||
9013 | // (xor (bitcast x) (build_pair flipbit, flipbit)) | |||
9014 | // This often reduces constant pool loads. | |||
9015 | if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) || | |||
9016 | (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) && | |||
9017 | N0.getNode()->hasOneUse() && VT.isInteger() && | |||
9018 | !VT.isVector() && !N0.getValueType().isVector()) { | |||
9019 | SDValue NewConv = DAG.getBitcast(VT, N0.getOperand(0)); | |||
9020 | AddToWorklist(NewConv.getNode()); | |||
9021 | ||||
9022 | SDLoc DL(N); | |||
9023 | if (N0.getValueType() == MVT::ppcf128 && !LegalTypes) { | |||
9024 | assert(VT.getSizeInBits() == 128)(static_cast <bool> (VT.getSizeInBits() == 128) ? void ( 0) : __assert_fail ("VT.getSizeInBits() == 128", "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 9024, __extension__ __PRETTY_FUNCTION__)); | |||
9025 | SDValue SignBit = DAG.getConstant( | |||
9026 | APInt::getSignMask(VT.getSizeInBits() / 2), SDLoc(N0), MVT::i64); | |||
9027 | SDValue FlipBit; | |||
9028 | if (N0.getOpcode() == ISD::FNEG) { | |||
9029 | FlipBit = SignBit; | |||
9030 | AddToWorklist(FlipBit.getNode()); | |||
9031 | } else { | |||
9032 | assert(N0.getOpcode() == ISD::FABS)(static_cast <bool> (N0.getOpcode() == ISD::FABS) ? void (0) : __assert_fail ("N0.getOpcode() == ISD::FABS", "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 9032, __extension__ __PRETTY_FUNCTION__)); | |||
9033 | SDValue Hi = | |||
9034 | DAG.getNode(ISD::EXTRACT_ELEMENT, SDLoc(NewConv), MVT::i64, NewConv, | |||
9035 | DAG.getIntPtrConstant(getPPCf128HiElementSelector(DAG), | |||
9036 | SDLoc(NewConv))); | |||
9037 | AddToWorklist(Hi.getNode()); | |||
9038 | FlipBit = DAG.getNode(ISD::AND, SDLoc(N0), MVT::i64, Hi, SignBit); | |||
9039 | AddToWorklist(FlipBit.getNode()); | |||
9040 | } | |||
9041 | SDValue FlipBits = | |||
9042 | DAG.getNode(ISD::BUILD_PAIR, SDLoc(N0), VT, FlipBit, FlipBit); | |||
9043 | AddToWorklist(FlipBits.getNode()); | |||
9044 | return DAG.getNode(ISD::XOR, DL, VT, NewConv, FlipBits); | |||
9045 | } | |||
9046 | APInt SignBit = APInt::getSignMask(VT.getSizeInBits()); | |||
9047 | if (N0.getOpcode() == ISD::FNEG) | |||
9048 | return DAG.getNode(ISD::XOR, DL, VT, | |||
9049 | NewConv, DAG.getConstant(SignBit, DL, VT)); | |||
9050 | assert(N0.getOpcode() == ISD::FABS)(static_cast <bool> (N0.getOpcode() == ISD::FABS) ? void (0) : __assert_fail ("N0.getOpcode() == ISD::FABS", "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 9050, __extension__ __PRETTY_FUNCTION__)); | |||
9051 | return DAG.getNode(ISD::AND, DL, VT, | |||
9052 | NewConv, DAG.getConstant(~SignBit, DL, VT)); | |||
9053 | } | |||
9054 | ||||
9055 | // fold (bitconvert (fcopysign cst, x)) -> | |||
9056 | // (or (and (bitconvert x), sign), (and cst, (not sign))) | |||
9057 | // Note that we don't handle (copysign x, cst) because this can always be | |||
9058 | // folded to an fneg or fabs. | |||
9059 | // | |||
9060 | // For ppc_fp128: | |||
9061 | // fold (bitcast (fcopysign cst, x)) -> | |||
9062 | // flipbit = (and (extract_element | |||
9063 | // (xor (bitcast cst), (bitcast x)), 0), | |||
9064 | // signbit) | |||
9065 | // (xor (bitcast cst) (build_pair flipbit, flipbit)) | |||
9066 | if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() && | |||
9067 | isa<ConstantFPSDNode>(N0.getOperand(0)) && | |||
9068 | VT.isInteger() && !VT.isVector()) { | |||
9069 | unsigned OrigXWidth = N0.getOperand(1).getValueSizeInBits(); | |||
9070 | EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth); | |||
9071 | if (isTypeLegal(IntXVT)) { | |||
9072 | SDValue X = DAG.getBitcast(IntXVT, N0.getOperand(1)); | |||
9073 | AddToWorklist(X.getNode()); | |||
9074 | ||||
9075 | // If X has a different width than the result/lhs, sext it or truncate it. | |||
9076 | unsigned VTWidth = VT.getSizeInBits(); | |||
9077 | if (OrigXWidth < VTWidth) { | |||
9078 | X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X); | |||
9079 | AddToWorklist(X.getNode()); | |||
9080 | } else if (OrigXWidth > VTWidth) { | |||
9081 | // To get the sign bit in the right place, we have to shift it right | |||
9082 | // before truncating. | |||
9083 | SDLoc DL(X); | |||
9084 | X = DAG.getNode(ISD::SRL, DL, | |||
9085 | X.getValueType(), X, | |||
9086 | DAG.getConstant(OrigXWidth-VTWidth, DL, | |||
9087 | X.getValueType())); | |||
9088 | AddToWorklist(X.getNode()); | |||
9089 | X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X); | |||
9090 | AddToWorklist(X.getNode()); | |||
9091 | } | |||
9092 | ||||
9093 | if (N0.getValueType() == MVT::ppcf128 && !LegalTypes) { | |||
9094 | APInt SignBit = APInt::getSignMask(VT.getSizeInBits() / 2); | |||
9095 | SDValue Cst = DAG.getBitcast(VT, N0.getOperand(0)); | |||
9096 | AddToWorklist(Cst.getNode()); | |||
9097 | SDValue X = DAG.getBitcast(VT, N0.getOperand(1)); | |||
9098 | AddToWorklist(X.getNode()); | |||
9099 | SDValue XorResult = DAG.getNode(ISD::XOR, SDLoc(N0), VT, Cst, X); | |||
9100 | AddToWorklist(XorResult.getNode()); | |||
9101 | SDValue XorResult64 = DAG.getNode( | |||
9102 | ISD::EXTRACT_ELEMENT, SDLoc(XorResult), MVT::i64, XorResult, | |||
9103 | DAG.getIntPtrConstant(getPPCf128HiElementSelector(DAG), | |||
9104 | SDLoc(XorResult))); | |||
9105 | AddToWorklist(XorResult64.getNode()); | |||
9106 | SDValue FlipBit = | |||
9107 | DAG.getNode(ISD::AND, SDLoc(XorResult64), MVT::i64, XorResult64, | |||
9108 | DAG.getConstant(SignBit, SDLoc(XorResult64), MVT::i64)); | |||
9109 | AddToWorklist(FlipBit.getNode()); | |||
9110 | SDValue FlipBits = | |||
9111 | DAG.getNode(ISD::BUILD_PAIR, SDLoc(N0), VT, FlipBit, FlipBit); | |||
9112 | AddToWorklist(FlipBits.getNode()); | |||
9113 | return DAG.getNode(ISD::XOR, SDLoc(N), VT, Cst, FlipBits); | |||
9114 | } | |||
9115 | APInt SignBit = APInt::getSignMask(VT.getSizeInBits()); | |||
9116 | X = DAG.getNode(ISD::AND, SDLoc(X), VT, | |||
9117 | X, DAG.getConstant(SignBit, SDLoc(X), VT)); | |||
9118 | AddToWorklist(X.getNode()); | |||
9119 | ||||
9120 | SDValue Cst = DAG.getBitcast(VT, N0.getOperand(0)); | |||
9121 | Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT, | |||
9122 | Cst, DAG.getConstant(~SignBit, SDLoc(Cst), VT)); | |||
9123 | AddToWorklist(Cst.getNode()); | |||
9124 | ||||
9125 | return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst); | |||
9126 | } | |||
9127 | } | |||
9128 | ||||
9129 | // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive. | |||
9130 | if (N0.getOpcode() == ISD::BUILD_PAIR) | |||
9131 | if (SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT)) | |||
9132 | return CombineLD; | |||
9133 | ||||
9134 | // Remove double bitcasts from shuffles - this is often a legacy of | |||
9135 | // XformToShuffleWithZero being used to combine bitmaskings (of | |||
9136 | // float vectors bitcast to integer vectors) into shuffles. | |||
9137 | // bitcast(shuffle(bitcast(s0),bitcast(s1))) -> shuffle(s0,s1) | |||
9138 | if (Level < AfterLegalizeDAG && TLI.isTypeLegal(VT) && VT.isVector() && | |||
9139 | N0->getOpcode() == ISD::VECTOR_SHUFFLE && | |||
9140 | VT.getVectorNumElements() >= N0.getValueType().getVectorNumElements() && | |||
9141 | !(VT.getVectorNumElements() % N0.getValueType().getVectorNumElements())) { | |||
9142 | ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N0); | |||
9143 | ||||
9144 | // If operands are a bitcast, peek through if it casts the original VT. | |||
9145 | // If operands are a constant, just bitcast back to original VT. | |||
9146 | auto PeekThroughBitcast = [&](SDValue Op) { | |||
9147 | if (Op.getOpcode() == ISD::BITCAST && | |||
9148 | Op.getOperand(0).getValueType() == VT) | |||
9149 | return SDValue(Op.getOperand(0)); | |||
9150 | if (Op.isUndef() || ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) || | |||
9151 | ISD::isBuildVectorOfConstantFPSDNodes(Op.getNode())) | |||
9152 | return DAG.getBitcast(VT, Op); | |||
9153 | return SDValue(); | |||
9154 | }; | |||
9155 | ||||
9156 | // FIXME: If either input vector is bitcast, try to convert the shuffle to | |||
9157 | // the result type of this bitcast. This would eliminate at least one | |||
9158 | // bitcast. See the transform in InstCombine. | |||
9159 | SDValue SV0 = PeekThroughBitcast(N0->getOperand(0)); | |||
9160 | SDValue SV1 = PeekThroughBitcast(N0->getOperand(1)); | |||
9161 | if (!(SV0 && SV1)) | |||
9162 | return SDValue(); | |||
9163 | ||||
9164 | int MaskScale = | |||
9165 | VT.getVectorNumElements() / N0.getValueType().getVectorNumElements(); | |||
9166 | SmallVector<int, 8> NewMask; | |||
9167 | for (int M : SVN->getMask()) | |||
9168 | for (int i = 0; i != MaskScale; ++i) | |||
9169 | NewMask.push_back(M < 0 ? -1 : M * MaskScale + i); | |||
9170 | ||||
9171 | bool LegalMask = TLI.isShuffleMaskLegal(NewMask, VT); | |||
9172 | if (!LegalMask) { | |||
9173 | std::swap(SV0, SV1); | |||
9174 | ShuffleVectorSDNode::commuteMask(NewMask); | |||
9175 | LegalMask = TLI.isShuffleMaskLegal(NewMask, VT); | |||
9176 | } | |||
9177 | ||||
9178 | if (LegalMask) | |||
9179 | return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, NewMask); | |||
9180 | } | |||
9181 | ||||
9182 | return SDValue(); | |||
9183 | } | |||
9184 | ||||
9185 | SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) { | |||
9186 | EVT VT = N->getValueType(0); | |||
9187 | return CombineConsecutiveLoads(N, VT); | |||
9188 | } | |||
9189 | ||||
9190 | /// We know that BV is a build_vector node with Constant, ConstantFP or Undef | |||
9191 | /// operands. DstEltVT indicates the destination element value type. | |||
9192 | SDValue DAGCombiner:: | |||
9193 | ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) { | |||
9194 | EVT SrcEltVT = BV->getValueType(0).getVectorElementType(); | |||
9195 | ||||
9196 | // If this is already the right type, we're done. | |||
9197 | if (SrcEltVT == DstEltVT) return SDValue(BV, 0); | |||
9198 | ||||
9199 | unsigned SrcBitSize = SrcEltVT.getSizeInBits(); | |||
9200 | unsigned DstBitSize = DstEltVT.getSizeInBits(); | |||
9201 | ||||
9202 | // If this is a conversion of N elements of one type to N elements of another | |||
9203 | // type, convert each element. This handles FP<->INT cases. | |||
9204 | if (SrcBitSize == DstBitSize) { | |||
9205 | EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, | |||
9206 | BV->getValueType(0).getVectorNumElements()); | |||
9207 | ||||
9208 | // Due to the FP element handling below calling this routine recursively, | |||
9209 | // we can end up with a scalar-to-vector node here. | |||
9210 | if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR) | |||
9211 | return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT, | |||
9212 | DAG.getBitcast(DstEltVT, BV->getOperand(0))); | |||
9213 | ||||
9214 | SmallVector<SDValue, 8> Ops; | |||
9215 | for (SDValue Op : BV->op_values()) { | |||
9216 | // If the vector element type is not legal, the BUILD_VECTOR operands | |||
9217 | // are promoted and implicitly truncated. Make that explicit here. | |||
9218 | if (Op.getValueType() != SrcEltVT) | |||
9219 | Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op); | |||
9220 | Ops.push_back(DAG.getBitcast(DstEltVT, Op)); | |||
9221 | AddToWorklist(Ops.back().getNode()); | |||
9222 | } | |||
9223 | return DAG.getBuildVector(VT, SDLoc(BV), Ops); | |||
9224 | } | |||
9225 | ||||
9226 | // Otherwise, we're growing or shrinking the elements. To avoid having to | |||
9227 | // handle annoying details of growing/shrinking FP values, we convert them to | |||
9228 | // int first. | |||
9229 | if (SrcEltVT.isFloatingPoint()) { | |||
9230 | // Convert the input float vector to a int vector where the elements are the | |||
9231 | // same sizes. | |||
9232 | EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits()); | |||
9233 | BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode(); | |||
9234 | SrcEltVT = IntVT; | |||
9235 | } | |||
9236 | ||||
9237 | // Now we know the input is an integer vector. If the output is a FP type, | |||
9238 | // convert to integer first, then to FP of the right size. | |||
9239 | if (DstEltVT.isFloatingPoint()) { | |||
9240 | EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits()); | |||
9241 | SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode(); | |||
9242 | ||||
9243 | // Next, convert to FP elements of the same size. | |||
9244 | return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT); | |||
9245 | } | |||
9246 | ||||
9247 | SDLoc DL(BV); | |||
9248 | ||||
9249 | // Okay, we know the src/dst types are both integers of differing types. | |||
9250 | // Handling growing first. | |||
9251 | assert(SrcEltVT.isInteger() && DstEltVT.isInteger())(static_cast <bool> (SrcEltVT.isInteger() && DstEltVT .isInteger()) ? void (0) : __assert_fail ("SrcEltVT.isInteger() && DstEltVT.isInteger()" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 9251, __extension__ __PRETTY_FUNCTION__)); | |||
9252 | if (SrcBitSize < DstBitSize) { | |||
9253 | unsigned NumInputsPerOutput = DstBitSize/SrcBitSize; | |||
9254 | ||||
9255 | SmallVector<SDValue, 8> Ops; | |||
9256 | for (unsigned i = 0, e = BV->getNumOperands(); i != e; | |||
9257 | i += NumInputsPerOutput) { | |||
9258 | bool isLE = DAG.getDataLayout().isLittleEndian(); | |||
9259 | APInt NewBits = APInt(DstBitSize, 0); | |||
9260 | bool EltIsUndef = true; | |||
9261 | for (unsigned j = 0; j != NumInputsPerOutput; ++j) { | |||
9262 | // Shift the previously computed bits over. | |||
9263 | NewBits <<= SrcBitSize; | |||
9264 | SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j)); | |||
9265 | if (Op.isUndef()) continue; | |||
9266 | EltIsUndef = false; | |||
9267 | ||||
9268 | NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue(). | |||
9269 | zextOrTrunc(SrcBitSize).zext(DstBitSize); | |||
9270 | } | |||
9271 | ||||
9272 | if (EltIsUndef) | |||
9273 | Ops.push_back(DAG.getUNDEF(DstEltVT)); | |||
9274 | else | |||
9275 | Ops.push_back(DAG.getConstant(NewBits, DL, DstEltVT)); | |||
9276 | } | |||
9277 | ||||
9278 | EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size()); | |||
9279 | return DAG.getBuildVector(VT, DL, Ops); | |||
9280 | } | |||
9281 | ||||
9282 | // Finally, this must be the case where we are shrinking elements: each input | |||
9283 | // turns into multiple outputs. | |||
9284 | unsigned NumOutputsPerInput = SrcBitSize/DstBitSize; | |||
9285 | EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, | |||
9286 | NumOutputsPerInput*BV->getNumOperands()); | |||
9287 | SmallVector<SDValue, 8> Ops; | |||
9288 | ||||
9289 | for (const SDValue &Op : BV->op_values()) { | |||
9290 | if (Op.isUndef()) { | |||
9291 | Ops.append(NumOutputsPerInput, DAG.getUNDEF(DstEltVT)); | |||
9292 | continue; | |||
9293 | } | |||
9294 | ||||
9295 | APInt OpVal = cast<ConstantSDNode>(Op)-> | |||
9296 | getAPIntValue().zextOrTrunc(SrcBitSize); | |||
9297 | ||||
9298 | for (unsigned j = 0; j != NumOutputsPerInput; ++j) { | |||
9299 | APInt ThisVal = OpVal.trunc(DstBitSize); | |||
9300 | Ops.push_back(DAG.getConstant(ThisVal, DL, DstEltVT)); | |||
9301 | OpVal.lshrInPlace(DstBitSize); | |||
9302 | } | |||
9303 | ||||
9304 | // For big endian targets, swap the order of the pieces of each element. | |||
9305 | if (DAG.getDataLayout().isBigEndian()) | |||
9306 | std::reverse(Ops.end()-NumOutputsPerInput, Ops.end()); | |||
9307 | } | |||
9308 | ||||
9309 | return DAG.getBuildVector(VT, DL, Ops); | |||
9310 | } | |||
9311 | ||||
9312 | static bool isContractable(SDNode *N) { | |||
9313 | SDNodeFlags F = N->getFlags(); | |||
9314 | return F.hasAllowContract() || F.hasUnsafeAlgebra(); | |||
9315 | } | |||
9316 | ||||
9317 | /// Try to perform FMA combining on a given FADD node. | |||
9318 | SDValue DAGCombiner::visitFADDForFMACombine(SDNode *N) { | |||
9319 | SDValue N0 = N->getOperand(0); | |||
9320 | SDValue N1 = N->getOperand(1); | |||
9321 | EVT VT = N->getValueType(0); | |||
9322 | SDLoc SL(N); | |||
9323 | ||||
9324 | const TargetOptions &Options = DAG.getTarget().Options; | |||
9325 | ||||
9326 | // Floating-point multiply-add with intermediate rounding. | |||
9327 | bool HasFMAD = (LegalOperations && TLI.isOperationLegal(ISD::FMAD, VT)); | |||
9328 | ||||
9329 | // Floating-point multiply-add without intermediate rounding. | |||
9330 | bool HasFMA = | |||
9331 | TLI.isFMAFasterThanFMulAndFAdd(VT) && | |||
9332 | (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT)); | |||
9333 | ||||
9334 | // No valid opcode, do not combine. | |||
9335 | if (!HasFMAD && !HasFMA) | |||
9336 | return SDValue(); | |||
9337 | ||||
9338 | bool AllowFusionGlobally = (Options.AllowFPOpFusion == FPOpFusion::Fast || | |||
9339 | Options.UnsafeFPMath || HasFMAD); | |||
9340 | // If the addition is not contractable, do not combine. | |||
9341 | if (!AllowFusionGlobally && !isContractable(N)) | |||
9342 | return SDValue(); | |||
9343 | ||||
9344 | const SelectionDAGTargetInfo *STI = DAG.getSubtarget().getSelectionDAGInfo(); | |||
9345 | if (STI && STI->generateFMAsInMachineCombiner(OptLevel)) | |||
9346 | return SDValue(); | |||
9347 | ||||
9348 | // Always prefer FMAD to FMA for precision. | |||
9349 | unsigned PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA; | |||
9350 | bool Aggressive = TLI.enableAggressiveFMAFusion(VT); | |||
9351 | ||||
9352 | // Is the node an FMUL and contractable either due to global flags or | |||
9353 | // SDNodeFlags. | |||
9354 | auto isContractableFMUL = [AllowFusionGlobally](SDValue N) { | |||
9355 | if (N.getOpcode() != ISD::FMUL) | |||
9356 | return false; | |||
9357 | return AllowFusionGlobally || isContractable(N.getNode()); | |||
9358 | }; | |||
9359 | // If we have two choices trying to fold (fadd (fmul u, v), (fmul x, y)), | |||
9360 | // prefer to fold the multiply with fewer uses. | |||
9361 | if (Aggressive && isContractableFMUL(N0) && isContractableFMUL(N1)) { | |||
9362 | if (N0.getNode()->use_size() > N1.getNode()->use_size()) | |||
9363 | std::swap(N0, N1); | |||
9364 | } | |||
9365 | ||||
9366 | // fold (fadd (fmul x, y), z) -> (fma x, y, z) | |||
9367 | if (isContractableFMUL(N0) && (Aggressive || N0->hasOneUse())) { | |||
9368 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9369 | N0.getOperand(0), N0.getOperand(1), N1); | |||
9370 | } | |||
9371 | ||||
9372 | // fold (fadd x, (fmul y, z)) -> (fma y, z, x) | |||
9373 | // Note: Commutes FADD operands. | |||
9374 | if (isContractableFMUL(N1) && (Aggressive || N1->hasOneUse())) { | |||
9375 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9376 | N1.getOperand(0), N1.getOperand(1), N0); | |||
9377 | } | |||
9378 | ||||
9379 | // Look through FP_EXTEND nodes to do more combining. | |||
9380 | ||||
9381 | // fold (fadd (fpext (fmul x, y)), z) -> (fma (fpext x), (fpext y), z) | |||
9382 | if (N0.getOpcode() == ISD::FP_EXTEND) { | |||
9383 | SDValue N00 = N0.getOperand(0); | |||
9384 | if (isContractableFMUL(N00) && | |||
9385 | TLI.isFPExtFoldable(PreferredFusedOpcode, VT, N00.getValueType())) { | |||
9386 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9387 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9388 | N00.getOperand(0)), | |||
9389 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9390 | N00.getOperand(1)), N1); | |||
9391 | } | |||
9392 | } | |||
9393 | ||||
9394 | // fold (fadd x, (fpext (fmul y, z))) -> (fma (fpext y), (fpext z), x) | |||
9395 | // Note: Commutes FADD operands. | |||
9396 | if (N1.getOpcode() == ISD::FP_EXTEND) { | |||
9397 | SDValue N10 = N1.getOperand(0); | |||
9398 | if (isContractableFMUL(N10) && | |||
9399 | TLI.isFPExtFoldable(PreferredFusedOpcode, VT, N10.getValueType())) { | |||
9400 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9401 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9402 | N10.getOperand(0)), | |||
9403 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9404 | N10.getOperand(1)), N0); | |||
9405 | } | |||
9406 | } | |||
9407 | ||||
9408 | // More folding opportunities when target permits. | |||
9409 | if (Aggressive) { | |||
9410 | // fold (fadd (fma x, y, (fmul u, v)), z) -> (fma x, y (fma u, v, z)) | |||
9411 | // FIXME: The UnsafeAlgebra flag should be propagated to FMA/FMAD, but FMF | |||
9412 | // are currently only supported on binary nodes. | |||
9413 | if (Options.UnsafeFPMath && | |||
9414 | N0.getOpcode() == PreferredFusedOpcode && | |||
9415 | N0.getOperand(2).getOpcode() == ISD::FMUL && | |||
9416 | N0->hasOneUse() && N0.getOperand(2)->hasOneUse()) { | |||
9417 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9418 | N0.getOperand(0), N0.getOperand(1), | |||
9419 | DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9420 | N0.getOperand(2).getOperand(0), | |||
9421 | N0.getOperand(2).getOperand(1), | |||
9422 | N1)); | |||
9423 | } | |||
9424 | ||||
9425 | // fold (fadd x, (fma y, z, (fmul u, v)) -> (fma y, z (fma u, v, x)) | |||
9426 | // FIXME: The UnsafeAlgebra flag should be propagated to FMA/FMAD, but FMF | |||
9427 | // are currently only supported on binary nodes. | |||
9428 | if (Options.UnsafeFPMath && | |||
9429 | N1->getOpcode() == PreferredFusedOpcode && | |||
9430 | N1.getOperand(2).getOpcode() == ISD::FMUL && | |||
9431 | N1->hasOneUse() && N1.getOperand(2)->hasOneUse()) { | |||
9432 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9433 | N1.getOperand(0), N1.getOperand(1), | |||
9434 | DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9435 | N1.getOperand(2).getOperand(0), | |||
9436 | N1.getOperand(2).getOperand(1), | |||
9437 | N0)); | |||
9438 | } | |||
9439 | ||||
9440 | ||||
9441 | // fold (fadd (fma x, y, (fpext (fmul u, v))), z) | |||
9442 | // -> (fma x, y, (fma (fpext u), (fpext v), z)) | |||
9443 | auto FoldFAddFMAFPExtFMul = [&] ( | |||
9444 | SDValue X, SDValue Y, SDValue U, SDValue V, SDValue Z) { | |||
9445 | return DAG.getNode(PreferredFusedOpcode, SL, VT, X, Y, | |||
9446 | DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9447 | DAG.getNode(ISD::FP_EXTEND, SL, VT, U), | |||
9448 | DAG.getNode(ISD::FP_EXTEND, SL, VT, V), | |||
9449 | Z)); | |||
9450 | }; | |||
9451 | if (N0.getOpcode() == PreferredFusedOpcode) { | |||
9452 | SDValue N02 = N0.getOperand(2); | |||
9453 | if (N02.getOpcode() == ISD::FP_EXTEND) { | |||
9454 | SDValue N020 = N02.getOperand(0); | |||
9455 | if (isContractableFMUL(N020) && | |||
9456 | TLI.isFPExtFoldable(PreferredFusedOpcode, VT, N020.getValueType())) { | |||
9457 | return FoldFAddFMAFPExtFMul(N0.getOperand(0), N0.getOperand(1), | |||
9458 | N020.getOperand(0), N020.getOperand(1), | |||
9459 | N1); | |||
9460 | } | |||
9461 | } | |||
9462 | } | |||
9463 | ||||
9464 | // fold (fadd (fpext (fma x, y, (fmul u, v))), z) | |||
9465 | // -> (fma (fpext x), (fpext y), (fma (fpext u), (fpext v), z)) | |||
9466 | // FIXME: This turns two single-precision and one double-precision | |||
9467 | // operation into two double-precision operations, which might not be | |||
9468 | // interesting for all targets, especially GPUs. | |||
9469 | auto FoldFAddFPExtFMAFMul = [&] ( | |||
9470 | SDValue X, SDValue Y, SDValue U, SDValue V, SDValue Z) { | |||
9471 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9472 | DAG.getNode(ISD::FP_EXTEND, SL, VT, X), | |||
9473 | DAG.getNode(ISD::FP_EXTEND, SL, VT, Y), | |||
9474 | DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9475 | DAG.getNode(ISD::FP_EXTEND, SL, VT, U), | |||
9476 | DAG.getNode(ISD::FP_EXTEND, SL, VT, V), | |||
9477 | Z)); | |||
9478 | }; | |||
9479 | if (N0.getOpcode() == ISD::FP_EXTEND) { | |||
9480 | SDValue N00 = N0.getOperand(0); | |||
9481 | if (N00.getOpcode() == PreferredFusedOpcode) { | |||
9482 | SDValue N002 = N00.getOperand(2); | |||
9483 | if (isContractableFMUL(N002) && | |||
9484 | TLI.isFPExtFoldable(PreferredFusedOpcode, VT, N00.getValueType())) { | |||
9485 | return FoldFAddFPExtFMAFMul(N00.getOperand(0), N00.getOperand(1), | |||
9486 | N002.getOperand(0), N002.getOperand(1), | |||
9487 | N1); | |||
9488 | } | |||
9489 | } | |||
9490 | } | |||
9491 | ||||
9492 | // fold (fadd x, (fma y, z, (fpext (fmul u, v))) | |||
9493 | // -> (fma y, z, (fma (fpext u), (fpext v), x)) | |||
9494 | if (N1.getOpcode() == PreferredFusedOpcode) { | |||
9495 | SDValue N12 = N1.getOperand(2); | |||
9496 | if (N12.getOpcode() == ISD::FP_EXTEND) { | |||
9497 | SDValue N120 = N12.getOperand(0); | |||
9498 | if (isContractableFMUL(N120) && | |||
9499 | TLI.isFPExtFoldable(PreferredFusedOpcode, VT, N120.getValueType())) { | |||
9500 | return FoldFAddFMAFPExtFMul(N1.getOperand(0), N1.getOperand(1), | |||
9501 | N120.getOperand(0), N120.getOperand(1), | |||
9502 | N0); | |||
9503 | } | |||
9504 | } | |||
9505 | } | |||
9506 | ||||
9507 | // fold (fadd x, (fpext (fma y, z, (fmul u, v))) | |||
9508 | // -> (fma (fpext y), (fpext z), (fma (fpext u), (fpext v), x)) | |||
9509 | // FIXME: This turns two single-precision and one double-precision | |||
9510 | // operation into two double-precision operations, which might not be | |||
9511 | // interesting for all targets, especially GPUs. | |||
9512 | if (N1.getOpcode() == ISD::FP_EXTEND) { | |||
9513 | SDValue N10 = N1.getOperand(0); | |||
9514 | if (N10.getOpcode() == PreferredFusedOpcode) { | |||
9515 | SDValue N102 = N10.getOperand(2); | |||
9516 | if (isContractableFMUL(N102) && | |||
9517 | TLI.isFPExtFoldable(PreferredFusedOpcode, VT, N10.getValueType())) { | |||
9518 | return FoldFAddFPExtFMAFMul(N10.getOperand(0), N10.getOperand(1), | |||
9519 | N102.getOperand(0), N102.getOperand(1), | |||
9520 | N0); | |||
9521 | } | |||
9522 | } | |||
9523 | } | |||
9524 | } | |||
9525 | ||||
9526 | return SDValue(); | |||
9527 | } | |||
9528 | ||||
9529 | /// Try to perform FMA combining on a given FSUB node. | |||
9530 | SDValue DAGCombiner::visitFSUBForFMACombine(SDNode *N) { | |||
9531 | SDValue N0 = N->getOperand(0); | |||
9532 | SDValue N1 = N->getOperand(1); | |||
9533 | EVT VT = N->getValueType(0); | |||
9534 | SDLoc SL(N); | |||
9535 | ||||
9536 | const TargetOptions &Options = DAG.getTarget().Options; | |||
9537 | // Floating-point multiply-add with intermediate rounding. | |||
9538 | bool HasFMAD = (LegalOperations && TLI.isOperationLegal(ISD::FMAD, VT)); | |||
9539 | ||||
9540 | // Floating-point multiply-add without intermediate rounding. | |||
9541 | bool HasFMA = | |||
9542 | TLI.isFMAFasterThanFMulAndFAdd(VT) && | |||
9543 | (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT)); | |||
9544 | ||||
9545 | // No valid opcode, do not combine. | |||
9546 | if (!HasFMAD && !HasFMA) | |||
9547 | return SDValue(); | |||
9548 | ||||
9549 | bool AllowFusionGlobally = (Options.AllowFPOpFusion == FPOpFusion::Fast || | |||
9550 | Options.UnsafeFPMath || HasFMAD); | |||
9551 | // If the subtraction is not contractable, do not combine. | |||
9552 | if (!AllowFusionGlobally && !isContractable(N)) | |||
9553 | return SDValue(); | |||
9554 | ||||
9555 | const SelectionDAGTargetInfo *STI = DAG.getSubtarget().getSelectionDAGInfo(); | |||
9556 | if (STI && STI->generateFMAsInMachineCombiner(OptLevel)) | |||
9557 | return SDValue(); | |||
9558 | ||||
9559 | // Always prefer FMAD to FMA for precision. | |||
9560 | unsigned PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA; | |||
9561 | bool Aggressive = TLI.enableAggressiveFMAFusion(VT); | |||
9562 | ||||
9563 | // Is the node an FMUL and contractable either due to global flags or | |||
9564 | // SDNodeFlags. | |||
9565 | auto isContractableFMUL = [AllowFusionGlobally](SDValue N) { | |||
9566 | if (N.getOpcode() != ISD::FMUL) | |||
9567 | return false; | |||
9568 | return AllowFusionGlobally || isContractable(N.getNode()); | |||
9569 | }; | |||
9570 | ||||
9571 | // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z)) | |||
9572 | if (isContractableFMUL(N0) && (Aggressive || N0->hasOneUse())) { | |||
9573 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9574 | N0.getOperand(0), N0.getOperand(1), | |||
9575 | DAG.getNode(ISD::FNEG, SL, VT, N1)); | |||
9576 | } | |||
9577 | ||||
9578 | // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x) | |||
9579 | // Note: Commutes FSUB operands. | |||
9580 | if (isContractableFMUL(N1) && (Aggressive || N1->hasOneUse())) | |||
9581 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9582 | DAG.getNode(ISD::FNEG, SL, VT, | |||
9583 | N1.getOperand(0)), | |||
9584 | N1.getOperand(1), N0); | |||
9585 | ||||
9586 | // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z)) | |||
9587 | if (N0.getOpcode() == ISD::FNEG && isContractableFMUL(N0.getOperand(0)) && | |||
9588 | (Aggressive || (N0->hasOneUse() && N0.getOperand(0).hasOneUse()))) { | |||
9589 | SDValue N00 = N0.getOperand(0).getOperand(0); | |||
9590 | SDValue N01 = N0.getOperand(0).getOperand(1); | |||
9591 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9592 | DAG.getNode(ISD::FNEG, SL, VT, N00), N01, | |||
9593 | DAG.getNode(ISD::FNEG, SL, VT, N1)); | |||
9594 | } | |||
9595 | ||||
9596 | // Look through FP_EXTEND nodes to do more combining. | |||
9597 | ||||
9598 | // fold (fsub (fpext (fmul x, y)), z) | |||
9599 | // -> (fma (fpext x), (fpext y), (fneg z)) | |||
9600 | if (N0.getOpcode() == ISD::FP_EXTEND) { | |||
9601 | SDValue N00 = N0.getOperand(0); | |||
9602 | if (isContractableFMUL(N00) && | |||
9603 | TLI.isFPExtFoldable(PreferredFusedOpcode, VT, N00.getValueType())) { | |||
9604 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9605 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9606 | N00.getOperand(0)), | |||
9607 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9608 | N00.getOperand(1)), | |||
9609 | DAG.getNode(ISD::FNEG, SL, VT, N1)); | |||
9610 | } | |||
9611 | } | |||
9612 | ||||
9613 | // fold (fsub x, (fpext (fmul y, z))) | |||
9614 | // -> (fma (fneg (fpext y)), (fpext z), x) | |||
9615 | // Note: Commutes FSUB operands. | |||
9616 | if (N1.getOpcode() == ISD::FP_EXTEND) { | |||
9617 | SDValue N10 = N1.getOperand(0); | |||
9618 | if (isContractableFMUL(N10) && | |||
9619 | TLI.isFPExtFoldable(PreferredFusedOpcode, VT, N10.getValueType())) { | |||
9620 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9621 | DAG.getNode(ISD::FNEG, SL, VT, | |||
9622 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9623 | N10.getOperand(0))), | |||
9624 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9625 | N10.getOperand(1)), | |||
9626 | N0); | |||
9627 | } | |||
9628 | } | |||
9629 | ||||
9630 | // fold (fsub (fpext (fneg (fmul, x, y))), z) | |||
9631 | // -> (fneg (fma (fpext x), (fpext y), z)) | |||
9632 | // Note: This could be removed with appropriate canonicalization of the | |||
9633 | // input expression into (fneg (fadd (fpext (fmul, x, y)), z). However, the | |||
9634 | // orthogonal flags -fp-contract=fast and -enable-unsafe-fp-math prevent | |||
9635 | // from implementing the canonicalization in visitFSUB. | |||
9636 | if (N0.getOpcode() == ISD::FP_EXTEND) { | |||
9637 | SDValue N00 = N0.getOperand(0); | |||
9638 | if (N00.getOpcode() == ISD::FNEG) { | |||
9639 | SDValue N000 = N00.getOperand(0); | |||
9640 | if (isContractableFMUL(N000) && | |||
9641 | TLI.isFPExtFoldable(PreferredFusedOpcode, VT, N00.getValueType())) { | |||
9642 | return DAG.getNode(ISD::FNEG, SL, VT, | |||
9643 | DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9644 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9645 | N000.getOperand(0)), | |||
9646 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9647 | N000.getOperand(1)), | |||
9648 | N1)); | |||
9649 | } | |||
9650 | } | |||
9651 | } | |||
9652 | ||||
9653 | // fold (fsub (fneg (fpext (fmul, x, y))), z) | |||
9654 | // -> (fneg (fma (fpext x)), (fpext y), z) | |||
9655 | // Note: This could be removed with appropriate canonicalization of the | |||
9656 | // input expression into (fneg (fadd (fpext (fmul, x, y)), z). However, the | |||
9657 | // orthogonal flags -fp-contract=fast and -enable-unsafe-fp-math prevent | |||
9658 | // from implementing the canonicalization in visitFSUB. | |||
9659 | if (N0.getOpcode() == ISD::FNEG) { | |||
9660 | SDValue N00 = N0.getOperand(0); | |||
9661 | if (N00.getOpcode() == ISD::FP_EXTEND) { | |||
9662 | SDValue N000 = N00.getOperand(0); | |||
9663 | if (isContractableFMUL(N000) && | |||
9664 | TLI.isFPExtFoldable(PreferredFusedOpcode, VT, N000.getValueType())) { | |||
9665 | return DAG.getNode(ISD::FNEG, SL, VT, | |||
9666 | DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9667 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9668 | N000.getOperand(0)), | |||
9669 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9670 | N000.getOperand(1)), | |||
9671 | N1)); | |||
9672 | } | |||
9673 | } | |||
9674 | } | |||
9675 | ||||
9676 | // More folding opportunities when target permits. | |||
9677 | if (Aggressive) { | |||
9678 | // fold (fsub (fma x, y, (fmul u, v)), z) | |||
9679 | // -> (fma x, y (fma u, v, (fneg z))) | |||
9680 | // FIXME: The UnsafeAlgebra flag should be propagated to FMA/FMAD, but FMF | |||
9681 | // are currently only supported on binary nodes. | |||
9682 | if (Options.UnsafeFPMath && N0.getOpcode() == PreferredFusedOpcode && | |||
9683 | isContractableFMUL(N0.getOperand(2)) && N0->hasOneUse() && | |||
9684 | N0.getOperand(2)->hasOneUse()) { | |||
9685 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9686 | N0.getOperand(0), N0.getOperand(1), | |||
9687 | DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9688 | N0.getOperand(2).getOperand(0), | |||
9689 | N0.getOperand(2).getOperand(1), | |||
9690 | DAG.getNode(ISD::FNEG, SL, VT, | |||
9691 | N1))); | |||
9692 | } | |||
9693 | ||||
9694 | // fold (fsub x, (fma y, z, (fmul u, v))) | |||
9695 | // -> (fma (fneg y), z, (fma (fneg u), v, x)) | |||
9696 | // FIXME: The UnsafeAlgebra flag should be propagated to FMA/FMAD, but FMF | |||
9697 | // are currently only supported on binary nodes. | |||
9698 | if (Options.UnsafeFPMath && N1.getOpcode() == PreferredFusedOpcode && | |||
9699 | isContractableFMUL(N1.getOperand(2))) { | |||
9700 | SDValue N20 = N1.getOperand(2).getOperand(0); | |||
9701 | SDValue N21 = N1.getOperand(2).getOperand(1); | |||
9702 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9703 | DAG.getNode(ISD::FNEG, SL, VT, | |||
9704 | N1.getOperand(0)), | |||
9705 | N1.getOperand(1), | |||
9706 | DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9707 | DAG.getNode(ISD::FNEG, SL, VT, N20), | |||
9708 | ||||
9709 | N21, N0)); | |||
9710 | } | |||
9711 | ||||
9712 | ||||
9713 | // fold (fsub (fma x, y, (fpext (fmul u, v))), z) | |||
9714 | // -> (fma x, y (fma (fpext u), (fpext v), (fneg z))) | |||
9715 | if (N0.getOpcode() == PreferredFusedOpcode) { | |||
9716 | SDValue N02 = N0.getOperand(2); | |||
9717 | if (N02.getOpcode() == ISD::FP_EXTEND) { | |||
9718 | SDValue N020 = N02.getOperand(0); | |||
9719 | if (isContractableFMUL(N020) && | |||
9720 | TLI.isFPExtFoldable(PreferredFusedOpcode, VT, N020.getValueType())) { | |||
9721 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9722 | N0.getOperand(0), N0.getOperand(1), | |||
9723 | DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9724 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9725 | N020.getOperand(0)), | |||
9726 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9727 | N020.getOperand(1)), | |||
9728 | DAG.getNode(ISD::FNEG, SL, VT, | |||
9729 | N1))); | |||
9730 | } | |||
9731 | } | |||
9732 | } | |||
9733 | ||||
9734 | // fold (fsub (fpext (fma x, y, (fmul u, v))), z) | |||
9735 | // -> (fma (fpext x), (fpext y), | |||
9736 | // (fma (fpext u), (fpext v), (fneg z))) | |||
9737 | // FIXME: This turns two single-precision and one double-precision | |||
9738 | // operation into two double-precision operations, which might not be | |||
9739 | // interesting for all targets, especially GPUs. | |||
9740 | if (N0.getOpcode() == ISD::FP_EXTEND) { | |||
9741 | SDValue N00 = N0.getOperand(0); | |||
9742 | if (N00.getOpcode() == PreferredFusedOpcode) { | |||
9743 | SDValue N002 = N00.getOperand(2); | |||
9744 | if (isContractableFMUL(N002) && | |||
9745 | TLI.isFPExtFoldable(PreferredFusedOpcode, VT, N00.getValueType())) { | |||
9746 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9747 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9748 | N00.getOperand(0)), | |||
9749 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9750 | N00.getOperand(1)), | |||
9751 | DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9752 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9753 | N002.getOperand(0)), | |||
9754 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9755 | N002.getOperand(1)), | |||
9756 | DAG.getNode(ISD::FNEG, SL, VT, | |||
9757 | N1))); | |||
9758 | } | |||
9759 | } | |||
9760 | } | |||
9761 | ||||
9762 | // fold (fsub x, (fma y, z, (fpext (fmul u, v)))) | |||
9763 | // -> (fma (fneg y), z, (fma (fneg (fpext u)), (fpext v), x)) | |||
9764 | if (N1.getOpcode() == PreferredFusedOpcode && | |||
9765 | N1.getOperand(2).getOpcode() == ISD::FP_EXTEND) { | |||
9766 | SDValue N120 = N1.getOperand(2).getOperand(0); | |||
9767 | if (isContractableFMUL(N120) && | |||
9768 | TLI.isFPExtFoldable(PreferredFusedOpcode, VT, N120.getValueType())) { | |||
9769 | SDValue N1200 = N120.getOperand(0); | |||
9770 | SDValue N1201 = N120.getOperand(1); | |||
9771 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9772 | DAG.getNode(ISD::FNEG, SL, VT, N1.getOperand(0)), | |||
9773 | N1.getOperand(1), | |||
9774 | DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9775 | DAG.getNode(ISD::FNEG, SL, VT, | |||
9776 | DAG.getNode(ISD::FP_EXTEND, SL, | |||
9777 | VT, N1200)), | |||
9778 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9779 | N1201), | |||
9780 | N0)); | |||
9781 | } | |||
9782 | } | |||
9783 | ||||
9784 | // fold (fsub x, (fpext (fma y, z, (fmul u, v)))) | |||
9785 | // -> (fma (fneg (fpext y)), (fpext z), | |||
9786 | // (fma (fneg (fpext u)), (fpext v), x)) | |||
9787 | // FIXME: This turns two single-precision and one double-precision | |||
9788 | // operation into two double-precision operations, which might not be | |||
9789 | // interesting for all targets, especially GPUs. | |||
9790 | if (N1.getOpcode() == ISD::FP_EXTEND && | |||
9791 | N1.getOperand(0).getOpcode() == PreferredFusedOpcode) { | |||
9792 | SDValue CvtSrc = N1.getOperand(0); | |||
9793 | SDValue N100 = CvtSrc.getOperand(0); | |||
9794 | SDValue N101 = CvtSrc.getOperand(1); | |||
9795 | SDValue N102 = CvtSrc.getOperand(2); | |||
9796 | if (isContractableFMUL(N102) && | |||
9797 | TLI.isFPExtFoldable(PreferredFusedOpcode, VT, CvtSrc.getValueType())) { | |||
9798 | SDValue N1020 = N102.getOperand(0); | |||
9799 | SDValue N1021 = N102.getOperand(1); | |||
9800 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9801 | DAG.getNode(ISD::FNEG, SL, VT, | |||
9802 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9803 | N100)), | |||
9804 | DAG.getNode(ISD::FP_EXTEND, SL, VT, N101), | |||
9805 | DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9806 | DAG.getNode(ISD::FNEG, SL, VT, | |||
9807 | DAG.getNode(ISD::FP_EXTEND, SL, | |||
9808 | VT, N1020)), | |||
9809 | DAG.getNode(ISD::FP_EXTEND, SL, VT, | |||
9810 | N1021), | |||
9811 | N0)); | |||
9812 | } | |||
9813 | } | |||
9814 | } | |||
9815 | ||||
9816 | return SDValue(); | |||
9817 | } | |||
9818 | ||||
9819 | /// Try to perform FMA combining on a given FMUL node based on the distributive | |||
9820 | /// law x * (y + 1) = x * y + x and variants thereof (commuted versions, | |||
9821 | /// subtraction instead of addition). | |||
9822 | SDValue DAGCombiner::visitFMULForFMADistributiveCombine(SDNode *N) { | |||
9823 | SDValue N0 = N->getOperand(0); | |||
9824 | SDValue N1 = N->getOperand(1); | |||
9825 | EVT VT = N->getValueType(0); | |||
9826 | SDLoc SL(N); | |||
9827 | ||||
9828 | assert(N->getOpcode() == ISD::FMUL && "Expected FMUL Operation")(static_cast <bool> (N->getOpcode() == ISD::FMUL && "Expected FMUL Operation") ? void (0) : __assert_fail ("N->getOpcode() == ISD::FMUL && \"Expected FMUL Operation\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 9828, __extension__ __PRETTY_FUNCTION__)); | |||
9829 | ||||
9830 | const TargetOptions &Options = DAG.getTarget().Options; | |||
9831 | ||||
9832 | // The transforms below are incorrect when x == 0 and y == inf, because the | |||
9833 | // intermediate multiplication produces a nan. | |||
9834 | if (!Options.NoInfsFPMath) | |||
9835 | return SDValue(); | |||
9836 | ||||
9837 | // Floating-point multiply-add without intermediate rounding. | |||
9838 | bool HasFMA = | |||
9839 | (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) && | |||
9840 | TLI.isFMAFasterThanFMulAndFAdd(VT) && | |||
9841 | (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT)); | |||
9842 | ||||
9843 | // Floating-point multiply-add with intermediate rounding. This can result | |||
9844 | // in a less precise result due to the changed rounding order. | |||
9845 | bool HasFMAD = Options.UnsafeFPMath && | |||
9846 | (LegalOperations && TLI.isOperationLegal(ISD::FMAD, VT)); | |||
9847 | ||||
9848 | // No valid opcode, do not combine. | |||
9849 | if (!HasFMAD && !HasFMA) | |||
9850 | return SDValue(); | |||
9851 | ||||
9852 | // Always prefer FMAD to FMA for precision. | |||
9853 | unsigned PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA; | |||
9854 | bool Aggressive = TLI.enableAggressiveFMAFusion(VT); | |||
9855 | ||||
9856 | // fold (fmul (fadd x, +1.0), y) -> (fma x, y, y) | |||
9857 | // fold (fmul (fadd x, -1.0), y) -> (fma x, y, (fneg y)) | |||
9858 | auto FuseFADD = [&](SDValue X, SDValue Y) { | |||
9859 | if (X.getOpcode() == ISD::FADD && (Aggressive || X->hasOneUse())) { | |||
9860 | auto XC1 = isConstOrConstSplatFP(X.getOperand(1)); | |||
9861 | if (XC1 && XC1->isExactlyValue(+1.0)) | |||
9862 | return DAG.getNode(PreferredFusedOpcode, SL, VT, X.getOperand(0), Y, Y); | |||
9863 | if (XC1 && XC1->isExactlyValue(-1.0)) | |||
9864 | return DAG.getNode(PreferredFusedOpcode, SL, VT, X.getOperand(0), Y, | |||
9865 | DAG.getNode(ISD::FNEG, SL, VT, Y)); | |||
9866 | } | |||
9867 | return SDValue(); | |||
9868 | }; | |||
9869 | ||||
9870 | if (SDValue FMA = FuseFADD(N0, N1)) | |||
9871 | return FMA; | |||
9872 | if (SDValue FMA = FuseFADD(N1, N0)) | |||
9873 | return FMA; | |||
9874 | ||||
9875 | // fold (fmul (fsub +1.0, x), y) -> (fma (fneg x), y, y) | |||
9876 | // fold (fmul (fsub -1.0, x), y) -> (fma (fneg x), y, (fneg y)) | |||
9877 | // fold (fmul (fsub x, +1.0), y) -> (fma x, y, (fneg y)) | |||
9878 | // fold (fmul (fsub x, -1.0), y) -> (fma x, y, y) | |||
9879 | auto FuseFSUB = [&](SDValue X, SDValue Y) { | |||
9880 | if (X.getOpcode() == ISD::FSUB && (Aggressive || X->hasOneUse())) { | |||
9881 | auto XC0 = isConstOrConstSplatFP(X.getOperand(0)); | |||
9882 | if (XC0 && XC0->isExactlyValue(+1.0)) | |||
9883 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9884 | DAG.getNode(ISD::FNEG, SL, VT, X.getOperand(1)), Y, | |||
9885 | Y); | |||
9886 | if (XC0 && XC0->isExactlyValue(-1.0)) | |||
9887 | return DAG.getNode(PreferredFusedOpcode, SL, VT, | |||
9888 | DAG.getNode(ISD::FNEG, SL, VT, X.getOperand(1)), Y, | |||
9889 | DAG.getNode(ISD::FNEG, SL, VT, Y)); | |||
9890 | ||||
9891 | auto XC1 = isConstOrConstSplatFP(X.getOperand(1)); | |||
9892 | if (XC1 && XC1->isExactlyValue(+1.0)) | |||
9893 | return DAG.getNode(PreferredFusedOpcode, SL, VT, X.getOperand(0), Y, | |||
9894 | DAG.getNode(ISD::FNEG, SL, VT, Y)); | |||
9895 | if (XC1 && XC1->isExactlyValue(-1.0)) | |||
9896 | return DAG.getNode(PreferredFusedOpcode, SL, VT, X.getOperand(0), Y, Y); | |||
9897 | } | |||
9898 | return SDValue(); | |||
9899 | }; | |||
9900 | ||||
9901 | if (SDValue FMA = FuseFSUB(N0, N1)) | |||
9902 | return FMA; | |||
9903 | if (SDValue FMA = FuseFSUB(N1, N0)) | |||
9904 | return FMA; | |||
9905 | ||||
9906 | return SDValue(); | |||
9907 | } | |||
9908 | ||||
9909 | static bool isFMulNegTwo(SDValue &N) { | |||
9910 | if (N.getOpcode() != ISD::FMUL) | |||
9911 | return false; | |||
9912 | if (ConstantFPSDNode *CFP = isConstOrConstSplatFP(N.getOperand(1))) | |||
9913 | return CFP->isExactlyValue(-2.0); | |||
9914 | return false; | |||
9915 | } | |||
9916 | ||||
9917 | SDValue DAGCombiner::visitFADD(SDNode *N) { | |||
9918 | SDValue N0 = N->getOperand(0); | |||
9919 | SDValue N1 = N->getOperand(1); | |||
9920 | bool N0CFP = isConstantFPBuildVectorOrConstantFP(N0); | |||
9921 | bool N1CFP = isConstantFPBuildVectorOrConstantFP(N1); | |||
9922 | EVT VT = N->getValueType(0); | |||
9923 | SDLoc DL(N); | |||
9924 | const TargetOptions &Options = DAG.getTarget().Options; | |||
9925 | const SDNodeFlags Flags = N->getFlags(); | |||
9926 | ||||
9927 | // fold vector ops | |||
9928 | if (VT.isVector()) | |||
9929 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
9930 | return FoldedVOp; | |||
9931 | ||||
9932 | // fold (fadd c1, c2) -> c1 + c2 | |||
9933 | if (N0CFP && N1CFP) | |||
9934 | return DAG.getNode(ISD::FADD, DL, VT, N0, N1, Flags); | |||
9935 | ||||
9936 | // canonicalize constant to RHS | |||
9937 | if (N0CFP && !N1CFP) | |||
9938 | return DAG.getNode(ISD::FADD, DL, VT, N1, N0, Flags); | |||
9939 | ||||
9940 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
9941 | return NewSel; | |||
9942 | ||||
9943 | // fold (fadd A, (fneg B)) -> (fsub A, B) | |||
9944 | if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) && | |||
9945 | isNegatibleForFree(N1, LegalOperations, TLI, &Options) == 2) | |||
9946 | return DAG.getNode(ISD::FSUB, DL, VT, N0, | |||
9947 | GetNegatedExpression(N1, DAG, LegalOperations), Flags); | |||
9948 | ||||
9949 | // fold (fadd (fneg A), B) -> (fsub B, A) | |||
9950 | if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) && | |||
9951 | isNegatibleForFree(N0, LegalOperations, TLI, &Options) == 2) | |||
9952 | return DAG.getNode(ISD::FSUB, DL, VT, N1, | |||
9953 | GetNegatedExpression(N0, DAG, LegalOperations), Flags); | |||
9954 | ||||
9955 | // fold (fadd A, (fmul B, -2.0)) -> (fsub A, (fadd B, B)) | |||
9956 | // fold (fadd (fmul B, -2.0), A) -> (fsub A, (fadd B, B)) | |||
9957 | if ((isFMulNegTwo(N0) && N0.hasOneUse()) || | |||
9958 | (isFMulNegTwo(N1) && N1.hasOneUse())) { | |||
9959 | bool N1IsFMul = isFMulNegTwo(N1); | |||
9960 | SDValue AddOp = N1IsFMul ? N1.getOperand(0) : N0.getOperand(0); | |||
9961 | SDValue Add = DAG.getNode(ISD::FADD, DL, VT, AddOp, AddOp, Flags); | |||
9962 | return DAG.getNode(ISD::FSUB, DL, VT, N1IsFMul ? N0 : N1, Add, Flags); | |||
9963 | } | |||
9964 | ||||
9965 | // FIXME: Auto-upgrade the target/function-level option. | |||
9966 | if (Options.NoSignedZerosFPMath || N->getFlags().hasNoSignedZeros()) { | |||
9967 | // fold (fadd A, 0) -> A | |||
9968 | if (ConstantFPSDNode *N1C = isConstOrConstSplatFP(N1)) | |||
9969 | if (N1C->isZero()) | |||
9970 | return N0; | |||
9971 | } | |||
9972 | ||||
9973 | // If 'unsafe math' is enabled, fold lots of things. | |||
9974 | if (Options.UnsafeFPMath) { | |||
9975 | // No FP constant should be created after legalization as Instruction | |||
9976 | // Selection pass has a hard time dealing with FP constants. | |||
9977 | bool AllowNewConst = (Level < AfterLegalizeDAG); | |||
9978 | ||||
9979 | // fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2)) | |||
9980 | if (N1CFP && N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() && | |||
9981 | isConstantFPBuildVectorOrConstantFP(N0.getOperand(1))) | |||
9982 | return DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(0), | |||
9983 | DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(1), N1, | |||
9984 | Flags), | |||
9985 | Flags); | |||
9986 | ||||
9987 | // If allowed, fold (fadd (fneg x), x) -> 0.0 | |||
9988 | if (AllowNewConst && N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1) | |||
9989 | return DAG.getConstantFP(0.0, DL, VT); | |||
9990 | ||||
9991 | // If allowed, fold (fadd x, (fneg x)) -> 0.0 | |||
9992 | if (AllowNewConst && N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0) | |||
9993 | return DAG.getConstantFP(0.0, DL, VT); | |||
9994 | ||||
9995 | // We can fold chains of FADD's of the same value into multiplications. | |||
9996 | // This transform is not safe in general because we are reducing the number | |||
9997 | // of rounding steps. | |||
9998 | if (TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && !N0CFP && !N1CFP) { | |||
9999 | if (N0.getOpcode() == ISD::FMUL) { | |||
10000 | bool CFP00 = isConstantFPBuildVectorOrConstantFP(N0.getOperand(0)); | |||
10001 | bool CFP01 = isConstantFPBuildVectorOrConstantFP(N0.getOperand(1)); | |||
10002 | ||||
10003 | // (fadd (fmul x, c), x) -> (fmul x, c+1) | |||
10004 | if (CFP01 && !CFP00 && N0.getOperand(0) == N1) { | |||
10005 | SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(1), | |||
10006 | DAG.getConstantFP(1.0, DL, VT), Flags); | |||
10007 | return DAG.getNode(ISD::FMUL, DL, VT, N1, NewCFP, Flags); | |||
10008 | } | |||
10009 | ||||
10010 | // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2) | |||
10011 | if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD && | |||
10012 | N1.getOperand(0) == N1.getOperand(1) && | |||
10013 | N0.getOperand(0) == N1.getOperand(0)) { | |||
10014 | SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(1), | |||
10015 | DAG.getConstantFP(2.0, DL, VT), Flags); | |||
10016 | return DAG.getNode(ISD::FMUL, DL, VT, N0.getOperand(0), NewCFP, Flags); | |||
10017 | } | |||
10018 | } | |||
10019 | ||||
10020 | if (N1.getOpcode() == ISD::FMUL) { | |||
10021 | bool CFP10 = isConstantFPBuildVectorOrConstantFP(N1.getOperand(0)); | |||
10022 | bool CFP11 = isConstantFPBuildVectorOrConstantFP(N1.getOperand(1)); | |||
10023 | ||||
10024 | // (fadd x, (fmul x, c)) -> (fmul x, c+1) | |||
10025 | if (CFP11 && !CFP10 && N1.getOperand(0) == N0) { | |||
10026 | SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, N1.getOperand(1), | |||
10027 | DAG.getConstantFP(1.0, DL, VT), Flags); | |||
10028 | return DAG.getNode(ISD::FMUL, DL, VT, N0, NewCFP, Flags); | |||
10029 | } | |||
10030 | ||||
10031 | // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2) | |||
10032 | if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD && | |||
10033 | N0.getOperand(0) == N0.getOperand(1) && | |||
10034 | N1.getOperand(0) == N0.getOperand(0)) { | |||
10035 | SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, N1.getOperand(1), | |||
10036 | DAG.getConstantFP(2.0, DL, VT), Flags); | |||
10037 | return DAG.getNode(ISD::FMUL, DL, VT, N1.getOperand(0), NewCFP, Flags); | |||
10038 | } | |||
10039 | } | |||
10040 | ||||
10041 | if (N0.getOpcode() == ISD::FADD && AllowNewConst) { | |||
10042 | bool CFP00 = isConstantFPBuildVectorOrConstantFP(N0.getOperand(0)); | |||
10043 | // (fadd (fadd x, x), x) -> (fmul x, 3.0) | |||
10044 | if (!CFP00 && N0.getOperand(0) == N0.getOperand(1) && | |||
10045 | (N0.getOperand(0) == N1)) { | |||
10046 | return DAG.getNode(ISD::FMUL, DL, VT, | |||
10047 | N1, DAG.getConstantFP(3.0, DL, VT), Flags); | |||
10048 | } | |||
10049 | } | |||
10050 | ||||
10051 | if (N1.getOpcode() == ISD::FADD && AllowNewConst) { | |||
10052 | bool CFP10 = isConstantFPBuildVectorOrConstantFP(N1.getOperand(0)); | |||
10053 | // (fadd x, (fadd x, x)) -> (fmul x, 3.0) | |||
10054 | if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) && | |||
10055 | N1.getOperand(0) == N0) { | |||
10056 | return DAG.getNode(ISD::FMUL, DL, VT, | |||
10057 | N0, DAG.getConstantFP(3.0, DL, VT), Flags); | |||
10058 | } | |||
10059 | } | |||
10060 | ||||
10061 | // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0) | |||
10062 | if (AllowNewConst && | |||
10063 | N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD && | |||
10064 | N0.getOperand(0) == N0.getOperand(1) && | |||
10065 | N1.getOperand(0) == N1.getOperand(1) && | |||
10066 | N0.getOperand(0) == N1.getOperand(0)) { | |||
10067 | return DAG.getNode(ISD::FMUL, DL, VT, N0.getOperand(0), | |||
10068 | DAG.getConstantFP(4.0, DL, VT), Flags); | |||
10069 | } | |||
10070 | } | |||
10071 | } // enable-unsafe-fp-math | |||
10072 | ||||
10073 | // FADD -> FMA combines: | |||
10074 | if (SDValue Fused = visitFADDForFMACombine(N)) { | |||
10075 | AddToWorklist(Fused.getNode()); | |||
10076 | return Fused; | |||
10077 | } | |||
10078 | return SDValue(); | |||
10079 | } | |||
10080 | ||||
10081 | SDValue DAGCombiner::visitFSUB(SDNode *N) { | |||
10082 | SDValue N0 = N->getOperand(0); | |||
10083 | SDValue N1 = N->getOperand(1); | |||
10084 | ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0); | |||
10085 | ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1); | |||
10086 | EVT VT = N->getValueType(0); | |||
10087 | SDLoc DL(N); | |||
10088 | const TargetOptions &Options = DAG.getTarget().Options; | |||
10089 | const SDNodeFlags Flags = N->getFlags(); | |||
10090 | ||||
10091 | // fold vector ops | |||
10092 | if (VT.isVector()) | |||
10093 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
10094 | return FoldedVOp; | |||
10095 | ||||
10096 | // fold (fsub c1, c2) -> c1-c2 | |||
10097 | if (N0CFP && N1CFP) | |||
10098 | return DAG.getNode(ISD::FSUB, DL, VT, N0, N1, Flags); | |||
10099 | ||||
10100 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
10101 | return NewSel; | |||
10102 | ||||
10103 | // fold (fsub A, (fneg B)) -> (fadd A, B) | |||
10104 | if (isNegatibleForFree(N1, LegalOperations, TLI, &Options)) | |||
10105 | return DAG.getNode(ISD::FADD, DL, VT, N0, | |||
10106 | GetNegatedExpression(N1, DAG, LegalOperations), Flags); | |||
10107 | ||||
10108 | // FIXME: Auto-upgrade the target/function-level option. | |||
10109 | if (Options.NoSignedZerosFPMath || N->getFlags().hasNoSignedZeros()) { | |||
10110 | // (fsub 0, B) -> -B | |||
10111 | if (N0CFP && N0CFP->isZero()) { | |||
10112 | if (isNegatibleForFree(N1, LegalOperations, TLI, &Options)) | |||
10113 | return GetNegatedExpression(N1, DAG, LegalOperations); | |||
10114 | if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT)) | |||
10115 | return DAG.getNode(ISD::FNEG, DL, VT, N1, Flags); | |||
10116 | } | |||
10117 | } | |||
10118 | ||||
10119 | // If 'unsafe math' is enabled, fold lots of things. | |||
10120 | if (Options.UnsafeFPMath) { | |||
10121 | // (fsub A, 0) -> A | |||
10122 | if (N1CFP && N1CFP->isZero()) | |||
10123 | return N0; | |||
10124 | ||||
10125 | // (fsub x, x) -> 0.0 | |||
10126 | if (N0 == N1) | |||
10127 | return DAG.getConstantFP(0.0f, DL, VT); | |||
10128 | ||||
10129 | // (fsub x, (fadd x, y)) -> (fneg y) | |||
10130 | // (fsub x, (fadd y, x)) -> (fneg y) | |||
10131 | if (N1.getOpcode() == ISD::FADD) { | |||
10132 | SDValue N10 = N1->getOperand(0); | |||
10133 | SDValue N11 = N1->getOperand(1); | |||
10134 | ||||
10135 | if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI, &Options)) | |||
10136 | return GetNegatedExpression(N11, DAG, LegalOperations); | |||
10137 | ||||
10138 | if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI, &Options)) | |||
10139 | return GetNegatedExpression(N10, DAG, LegalOperations); | |||
10140 | } | |||
10141 | } | |||
10142 | ||||
10143 | // FSUB -> FMA combines: | |||
10144 | if (SDValue Fused = visitFSUBForFMACombine(N)) { | |||
10145 | AddToWorklist(Fused.getNode()); | |||
10146 | return Fused; | |||
10147 | } | |||
10148 | ||||
10149 | return SDValue(); | |||
10150 | } | |||
10151 | ||||
10152 | SDValue DAGCombiner::visitFMUL(SDNode *N) { | |||
10153 | SDValue N0 = N->getOperand(0); | |||
10154 | SDValue N1 = N->getOperand(1); | |||
10155 | ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0); | |||
10156 | ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1); | |||
10157 | EVT VT = N->getValueType(0); | |||
10158 | SDLoc DL(N); | |||
10159 | const TargetOptions &Options = DAG.getTarget().Options; | |||
10160 | const SDNodeFlags Flags = N->getFlags(); | |||
10161 | ||||
10162 | // fold vector ops | |||
10163 | if (VT.isVector()) { | |||
10164 | // This just handles C1 * C2 for vectors. Other vector folds are below. | |||
10165 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
10166 | return FoldedVOp; | |||
10167 | } | |||
10168 | ||||
10169 | // fold (fmul c1, c2) -> c1*c2 | |||
10170 | if (N0CFP && N1CFP) | |||
10171 | return DAG.getNode(ISD::FMUL, DL, VT, N0, N1, Flags); | |||
10172 | ||||
10173 | // canonicalize constant to RHS | |||
10174 | if (isConstantFPBuildVectorOrConstantFP(N0) && | |||
10175 | !isConstantFPBuildVectorOrConstantFP(N1)) | |||
10176 | return DAG.getNode(ISD::FMUL, DL, VT, N1, N0, Flags); | |||
10177 | ||||
10178 | // fold (fmul A, 1.0) -> A | |||
10179 | if (N1CFP && N1CFP->isExactlyValue(1.0)) | |||
10180 | return N0; | |||
10181 | ||||
10182 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
10183 | return NewSel; | |||
10184 | ||||
10185 | if (Options.UnsafeFPMath) { | |||
10186 | // fold (fmul A, 0) -> 0 | |||
10187 | if (N1CFP && N1CFP->isZero()) | |||
10188 | return N1; | |||
10189 | ||||
10190 | // fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2)) | |||
10191 | if (N0.getOpcode() == ISD::FMUL) { | |||
10192 | // Fold scalars or any vector constants (not just splats). | |||
10193 | // This fold is done in general by InstCombine, but extra fmul insts | |||
10194 | // may have been generated during lowering. | |||
10195 | SDValue N00 = N0.getOperand(0); | |||
10196 | SDValue N01 = N0.getOperand(1); | |||
10197 | auto *BV1 = dyn_cast<BuildVectorSDNode>(N1); | |||
10198 | auto *BV00 = dyn_cast<BuildVectorSDNode>(N00); | |||
10199 | auto *BV01 = dyn_cast<BuildVectorSDNode>(N01); | |||
10200 | ||||
10201 | // Check 1: Make sure that the first operand of the inner multiply is NOT | |||
10202 | // a constant. Otherwise, we may induce infinite looping. | |||
10203 | if (!(isConstOrConstSplatFP(N00) || (BV00 && BV00->isConstant()))) { | |||
10204 | // Check 2: Make sure that the second operand of the inner multiply and | |||
10205 | // the second operand of the outer multiply are constants. | |||
10206 | if ((N1CFP && isConstOrConstSplatFP(N01)) || | |||
10207 | (BV1 && BV01 && BV1->isConstant() && BV01->isConstant())) { | |||
10208 | SDValue MulConsts = DAG.getNode(ISD::FMUL, DL, VT, N01, N1, Flags); | |||
10209 | return DAG.getNode(ISD::FMUL, DL, VT, N00, MulConsts, Flags); | |||
10210 | } | |||
10211 | } | |||
10212 | } | |||
10213 | ||||
10214 | // fold (fmul (fadd x, x), c) -> (fmul x, (fmul 2.0, c)) | |||
10215 | // Undo the fmul 2.0, x -> fadd x, x transformation, since if it occurs | |||
10216 | // during an early run of DAGCombiner can prevent folding with fmuls | |||
10217 | // inserted during lowering. | |||
10218 | if (N0.getOpcode() == ISD::FADD && | |||
10219 | (N0.getOperand(0) == N0.getOperand(1)) && | |||
10220 | N0.hasOneUse()) { | |||
10221 | const SDValue Two = DAG.getConstantFP(2.0, DL, VT); | |||
10222 | SDValue MulConsts = DAG.getNode(ISD::FMUL, DL, VT, Two, N1, Flags); | |||
10223 | return DAG.getNode(ISD::FMUL, DL, VT, N0.getOperand(0), MulConsts, Flags); | |||
10224 | } | |||
10225 | } | |||
10226 | ||||
10227 | // fold (fmul X, 2.0) -> (fadd X, X) | |||
10228 | if (N1CFP && N1CFP->isExactlyValue(+2.0)) | |||
10229 | return DAG.getNode(ISD::FADD, DL, VT, N0, N0, Flags); | |||
10230 | ||||
10231 | // fold (fmul X, -1.0) -> (fneg X) | |||
10232 | if (N1CFP && N1CFP->isExactlyValue(-1.0)) | |||
10233 | if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT)) | |||
10234 | return DAG.getNode(ISD::FNEG, DL, VT, N0); | |||
10235 | ||||
10236 | // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y) | |||
10237 | if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) { | |||
10238 | if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) { | |||
10239 | // Both can be negated for free, check to see if at least one is cheaper | |||
10240 | // negated. | |||
10241 | if (LHSNeg == 2 || RHSNeg == 2) | |||
10242 | return DAG.getNode(ISD::FMUL, DL, VT, | |||
10243 | GetNegatedExpression(N0, DAG, LegalOperations), | |||
10244 | GetNegatedExpression(N1, DAG, LegalOperations), | |||
10245 | Flags); | |||
10246 | } | |||
10247 | } | |||
10248 | ||||
10249 | // fold (fmul X, (select (fcmp X > 0.0), -1.0, 1.0)) -> (fneg (fabs X)) | |||
10250 | // fold (fmul X, (select (fcmp X > 0.0), 1.0, -1.0)) -> (fabs X) | |||
10251 | if (Flags.hasNoNaNs() && Flags.hasNoSignedZeros() && | |||
10252 | (N0.getOpcode() == ISD::SELECT || N1.getOpcode() == ISD::SELECT) && | |||
10253 | TLI.isOperationLegal(ISD::FABS, VT)) { | |||
10254 | SDValue Select = N0, X = N1; | |||
10255 | if (Select.getOpcode() != ISD::SELECT) | |||
10256 | std::swap(Select, X); | |||
10257 | ||||
10258 | SDValue Cond = Select.getOperand(0); | |||
10259 | auto TrueOpnd = dyn_cast<ConstantFPSDNode>(Select.getOperand(1)); | |||
10260 | auto FalseOpnd = dyn_cast<ConstantFPSDNode>(Select.getOperand(2)); | |||
10261 | ||||
10262 | if (TrueOpnd && FalseOpnd && | |||
10263 | Cond.getOpcode() == ISD::SETCC && Cond.getOperand(0) == X && | |||
10264 | isa<ConstantFPSDNode>(Cond.getOperand(1)) && | |||
10265 | cast<ConstantFPSDNode>(Cond.getOperand(1))->isExactlyValue(0.0)) { | |||
10266 | ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get(); | |||
10267 | switch (CC) { | |||
10268 | default: break; | |||
10269 | case ISD::SETOLT: | |||
10270 | case ISD::SETULT: | |||
10271 | case ISD::SETOLE: | |||
10272 | case ISD::SETULE: | |||
10273 | case ISD::SETLT: | |||
10274 | case ISD::SETLE: | |||
10275 | std::swap(TrueOpnd, FalseOpnd); | |||
10276 | LLVM_FALLTHROUGH[[clang::fallthrough]]; | |||
10277 | case ISD::SETOGT: | |||
10278 | case ISD::SETUGT: | |||
10279 | case ISD::SETOGE: | |||
10280 | case ISD::SETUGE: | |||
10281 | case ISD::SETGT: | |||
10282 | case ISD::SETGE: | |||
10283 | if (TrueOpnd->isExactlyValue(-1.0) && FalseOpnd->isExactlyValue(1.0) && | |||
10284 | TLI.isOperationLegal(ISD::FNEG, VT)) | |||
10285 | return DAG.getNode(ISD::FNEG, DL, VT, | |||
10286 | DAG.getNode(ISD::FABS, DL, VT, X)); | |||
10287 | if (TrueOpnd->isExactlyValue(1.0) && FalseOpnd->isExactlyValue(-1.0)) | |||
10288 | return DAG.getNode(ISD::FABS, DL, VT, X); | |||
10289 | ||||
10290 | break; | |||
10291 | } | |||
10292 | } | |||
10293 | } | |||
10294 | ||||
10295 | // FMUL -> FMA combines: | |||
10296 | if (SDValue Fused = visitFMULForFMADistributiveCombine(N)) { | |||
10297 | AddToWorklist(Fused.getNode()); | |||
10298 | return Fused; | |||
10299 | } | |||
10300 | ||||
10301 | return SDValue(); | |||
10302 | } | |||
10303 | ||||
10304 | SDValue DAGCombiner::visitFMA(SDNode *N) { | |||
10305 | SDValue N0 = N->getOperand(0); | |||
10306 | SDValue N1 = N->getOperand(1); | |||
10307 | SDValue N2 = N->getOperand(2); | |||
10308 | ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); | |||
10309 | ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); | |||
10310 | EVT VT = N->getValueType(0); | |||
10311 | SDLoc DL(N); | |||
10312 | const TargetOptions &Options = DAG.getTarget().Options; | |||
10313 | ||||
10314 | // Constant fold FMA. | |||
10315 | if (isa<ConstantFPSDNode>(N0) && | |||
10316 | isa<ConstantFPSDNode>(N1) && | |||
10317 | isa<ConstantFPSDNode>(N2)) { | |||
10318 | return DAG.getNode(ISD::FMA, DL, VT, N0, N1, N2); | |||
10319 | } | |||
10320 | ||||
10321 | if (Options.UnsafeFPMath) { | |||
10322 | if (N0CFP && N0CFP->isZero()) | |||
10323 | return N2; | |||
10324 | if (N1CFP && N1CFP->isZero()) | |||
10325 | return N2; | |||
10326 | } | |||
10327 | // TODO: The FMA node should have flags that propagate to these nodes. | |||
10328 | if (N0CFP && N0CFP->isExactlyValue(1.0)) | |||
10329 | return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2); | |||
10330 | if (N1CFP && N1CFP->isExactlyValue(1.0)) | |||
10331 | return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2); | |||
10332 | ||||
10333 | // Canonicalize (fma c, x, y) -> (fma x, c, y) | |||
10334 | if (isConstantFPBuildVectorOrConstantFP(N0) && | |||
10335 | !isConstantFPBuildVectorOrConstantFP(N1)) | |||
10336 | return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2); | |||
10337 | ||||
10338 | // TODO: FMA nodes should have flags that propagate to the created nodes. | |||
10339 | // For now, create a Flags object for use with all unsafe math transforms. | |||
10340 | SDNodeFlags Flags; | |||
10341 | Flags.setUnsafeAlgebra(true); | |||
10342 | ||||
10343 | if (Options.UnsafeFPMath) { | |||
10344 | // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2) | |||
10345 | if (N2.getOpcode() == ISD::FMUL && N0 == N2.getOperand(0) && | |||
10346 | isConstantFPBuildVectorOrConstantFP(N1) && | |||
10347 | isConstantFPBuildVectorOrConstantFP(N2.getOperand(1))) { | |||
10348 | return DAG.getNode(ISD::FMUL, DL, VT, N0, | |||
10349 | DAG.getNode(ISD::FADD, DL, VT, N1, N2.getOperand(1), | |||
10350 | Flags), Flags); | |||
10351 | } | |||
10352 | ||||
10353 | // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y) | |||
10354 | if (N0.getOpcode() == ISD::FMUL && | |||
10355 | isConstantFPBuildVectorOrConstantFP(N1) && | |||
10356 | isConstantFPBuildVectorOrConstantFP(N0.getOperand(1))) { | |||
10357 | return DAG.getNode(ISD::FMA, DL, VT, | |||
10358 | N0.getOperand(0), | |||
10359 | DAG.getNode(ISD::FMUL, DL, VT, N1, N0.getOperand(1), | |||
10360 | Flags), | |||
10361 | N2); | |||
10362 | } | |||
10363 | } | |||
10364 | ||||
10365 | // (fma x, 1, y) -> (fadd x, y) | |||
10366 | // (fma x, -1, y) -> (fadd (fneg x), y) | |||
10367 | if (N1CFP) { | |||
10368 | if (N1CFP->isExactlyValue(1.0)) | |||
10369 | // TODO: The FMA node should have flags that propagate to this node. | |||
10370 | return DAG.getNode(ISD::FADD, DL, VT, N0, N2); | |||
10371 | ||||
10372 | if (N1CFP->isExactlyValue(-1.0) && | |||
10373 | (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) { | |||
10374 | SDValue RHSNeg = DAG.getNode(ISD::FNEG, DL, VT, N0); | |||
10375 | AddToWorklist(RHSNeg.getNode()); | |||
10376 | // TODO: The FMA node should have flags that propagate to this node. | |||
10377 | return DAG.getNode(ISD::FADD, DL, VT, N2, RHSNeg); | |||
10378 | } | |||
10379 | ||||
10380 | // fma (fneg x), K, y -> fma x -K, y | |||
10381 | if (N0.getOpcode() == ISD::FNEG && | |||
10382 | (TLI.isOperationLegal(ISD::ConstantFP, VT) || | |||
10383 | (N1.hasOneUse() && !TLI.isFPImmLegal(N1CFP->getValueAPF(), VT)))) { | |||
10384 | return DAG.getNode(ISD::FMA, DL, VT, N0.getOperand(0), | |||
10385 | DAG.getNode(ISD::FNEG, DL, VT, N1, Flags), N2); | |||
10386 | } | |||
10387 | } | |||
10388 | ||||
10389 | if (Options.UnsafeFPMath) { | |||
10390 | // (fma x, c, x) -> (fmul x, (c+1)) | |||
10391 | if (N1CFP && N0 == N2) { | |||
10392 | return DAG.getNode(ISD::FMUL, DL, VT, N0, | |||
10393 | DAG.getNode(ISD::FADD, DL, VT, N1, | |||
10394 | DAG.getConstantFP(1.0, DL, VT), Flags), | |||
10395 | Flags); | |||
10396 | } | |||
10397 | ||||
10398 | // (fma x, c, (fneg x)) -> (fmul x, (c-1)) | |||
10399 | if (N1CFP && N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0) { | |||
10400 | return DAG.getNode(ISD::FMUL, DL, VT, N0, | |||
10401 | DAG.getNode(ISD::FADD, DL, VT, N1, | |||
10402 | DAG.getConstantFP(-1.0, DL, VT), Flags), | |||
10403 | Flags); | |||
10404 | } | |||
10405 | } | |||
10406 | ||||
10407 | return SDValue(); | |||
10408 | } | |||
10409 | ||||
10410 | // Combine multiple FDIVs with the same divisor into multiple FMULs by the | |||
10411 | // reciprocal. | |||
10412 | // E.g., (a / D; b / D;) -> (recip = 1.0 / D; a * recip; b * recip) | |||
10413 | // Notice that this is not always beneficial. One reason is different targets | |||
10414 | // may have different costs for FDIV and FMUL, so sometimes the cost of two | |||
10415 | // FDIVs may be lower than the cost of one FDIV and two FMULs. Another reason | |||
10416 | // is the critical path is increased from "one FDIV" to "one FDIV + one FMUL". | |||
10417 | SDValue DAGCombiner::combineRepeatedFPDivisors(SDNode *N) { | |||
10418 | bool UnsafeMath = DAG.getTarget().Options.UnsafeFPMath; | |||
10419 | const SDNodeFlags Flags = N->getFlags(); | |||
10420 | if (!UnsafeMath && !Flags.hasAllowReciprocal()) | |||
10421 | return SDValue(); | |||
10422 | ||||
10423 | // Skip if current node is a reciprocal. | |||
10424 | SDValue N0 = N->getOperand(0); | |||
10425 | ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); | |||
10426 | if (N0CFP && N0CFP->isExactlyValue(1.0)) | |||
10427 | return SDValue(); | |||
10428 | ||||
10429 | // Exit early if the target does not want this transform or if there can't | |||
10430 | // possibly be enough uses of the divisor to make the transform worthwhile. | |||
10431 | SDValue N1 = N->getOperand(1); | |||
10432 | unsigned MinUses = TLI.combineRepeatedFPDivisors(); | |||
10433 | if (!MinUses || N1->use_size() < MinUses) | |||
10434 | return SDValue(); | |||
10435 | ||||
10436 | // Find all FDIV users of the same divisor. | |||
10437 | // Use a set because duplicates may be present in the user list. | |||
10438 | SetVector<SDNode *> Users; | |||
10439 | for (auto *U : N1->uses()) { | |||
10440 | if (U->getOpcode() == ISD::FDIV && U->getOperand(1) == N1) { | |||
10441 | // This division is eligible for optimization only if global unsafe math | |||
10442 | // is enabled or if this division allows reciprocal formation. | |||
10443 | if (UnsafeMath || U->getFlags().hasAllowReciprocal()) | |||
10444 | Users.insert(U); | |||
10445 | } | |||
10446 | } | |||
10447 | ||||
10448 | // Now that we have the actual number of divisor uses, make sure it meets | |||
10449 | // the minimum threshold specified by the target. | |||
10450 | if (Users.size() < MinUses) | |||
10451 | return SDValue(); | |||
10452 | ||||
10453 | EVT VT = N->getValueType(0); | |||
10454 | SDLoc DL(N); | |||
10455 | SDValue FPOne = DAG.getConstantFP(1.0, DL, VT); | |||
10456 | SDValue Reciprocal = DAG.getNode(ISD::FDIV, DL, VT, FPOne, N1, Flags); | |||
10457 | ||||
10458 | // Dividend / Divisor -> Dividend * Reciprocal | |||
10459 | for (auto *U : Users) { | |||
10460 | SDValue Dividend = U->getOperand(0); | |||
10461 | if (Dividend != FPOne) { | |||
10462 | SDValue NewNode = DAG.getNode(ISD::FMUL, SDLoc(U), VT, Dividend, | |||
10463 | Reciprocal, Flags); | |||
10464 | CombineTo(U, NewNode); | |||
10465 | } else if (U != Reciprocal.getNode()) { | |||
10466 | // In the absence of fast-math-flags, this user node is always the | |||
10467 | // same node as Reciprocal, but with FMF they may be different nodes. | |||
10468 | CombineTo(U, Reciprocal); | |||
10469 | } | |||
10470 | } | |||
10471 | return SDValue(N, 0); // N was replaced. | |||
10472 | } | |||
10473 | ||||
10474 | SDValue DAGCombiner::visitFDIV(SDNode *N) { | |||
10475 | SDValue N0 = N->getOperand(0); | |||
10476 | SDValue N1 = N->getOperand(1); | |||
10477 | ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); | |||
10478 | ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); | |||
10479 | EVT VT = N->getValueType(0); | |||
10480 | SDLoc DL(N); | |||
10481 | const TargetOptions &Options = DAG.getTarget().Options; | |||
10482 | SDNodeFlags Flags = N->getFlags(); | |||
10483 | ||||
10484 | // fold vector ops | |||
10485 | if (VT.isVector()) | |||
10486 | if (SDValue FoldedVOp = SimplifyVBinOp(N)) | |||
10487 | return FoldedVOp; | |||
10488 | ||||
10489 | // fold (fdiv c1, c2) -> c1/c2 | |||
10490 | if (N0CFP && N1CFP) | |||
10491 | return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1, Flags); | |||
10492 | ||||
10493 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
10494 | return NewSel; | |||
10495 | ||||
10496 | if (Options.UnsafeFPMath) { | |||
10497 | // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable. | |||
10498 | if (N1CFP) { | |||
10499 | // Compute the reciprocal 1.0 / c2. | |||
10500 | const APFloat &N1APF = N1CFP->getValueAPF(); | |||
10501 | APFloat Recip(N1APF.getSemantics(), 1); // 1.0 | |||
10502 | APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven); | |||
10503 | // Only do the transform if the reciprocal is a legal fp immediate that | |||
10504 | // isn't too nasty (eg NaN, denormal, ...). | |||
10505 | if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty | |||
10506 | (!LegalOperations || | |||
10507 | // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM | |||
10508 | // backend)... we should handle this gracefully after Legalize. | |||
10509 | // TLI.isOperationLegalOrCustom(ISD::ConstantFP, VT) || | |||
10510 | TLI.isOperationLegal(ISD::ConstantFP, VT) || | |||
10511 | TLI.isFPImmLegal(Recip, VT))) | |||
10512 | return DAG.getNode(ISD::FMUL, DL, VT, N0, | |||
10513 | DAG.getConstantFP(Recip, DL, VT), Flags); | |||
10514 | } | |||
10515 | ||||
10516 | // If this FDIV is part of a reciprocal square root, it may be folded | |||
10517 | // into a target-specific square root estimate instruction. | |||
10518 | if (N1.getOpcode() == ISD::FSQRT) { | |||
10519 | if (SDValue RV = buildRsqrtEstimate(N1.getOperand(0), Flags)) { | |||
10520 | return DAG.getNode(ISD::FMUL, DL, VT, N0, RV, Flags); | |||
10521 | } | |||
10522 | } else if (N1.getOpcode() == ISD::FP_EXTEND && | |||
10523 | N1.getOperand(0).getOpcode() == ISD::FSQRT) { | |||
10524 | if (SDValue RV = buildRsqrtEstimate(N1.getOperand(0).getOperand(0), | |||
10525 | Flags)) { | |||
10526 | RV = DAG.getNode(ISD::FP_EXTEND, SDLoc(N1), VT, RV); | |||
10527 | AddToWorklist(RV.getNode()); | |||
10528 | return DAG.getNode(ISD::FMUL, DL, VT, N0, RV, Flags); | |||
10529 | } | |||
10530 | } else if (N1.getOpcode() == ISD::FP_ROUND && | |||
10531 | N1.getOperand(0).getOpcode() == ISD::FSQRT) { | |||
10532 | if (SDValue RV = buildRsqrtEstimate(N1.getOperand(0).getOperand(0), | |||
10533 | Flags)) { | |||
10534 | RV = DAG.getNode(ISD::FP_ROUND, SDLoc(N1), VT, RV, N1.getOperand(1)); | |||
10535 | AddToWorklist(RV.getNode()); | |||
10536 | return DAG.getNode(ISD::FMUL, DL, VT, N0, RV, Flags); | |||
10537 | } | |||
10538 | } else if (N1.getOpcode() == ISD::FMUL) { | |||
10539 | // Look through an FMUL. Even though this won't remove the FDIV directly, | |||
10540 | // it's still worthwhile to get rid of the FSQRT if possible. | |||
10541 | SDValue SqrtOp; | |||
10542 | SDValue OtherOp; | |||
10543 | if (N1.getOperand(0).getOpcode() == ISD::FSQRT) { | |||
10544 | SqrtOp = N1.getOperand(0); | |||
10545 | OtherOp = N1.getOperand(1); | |||
10546 | } else if (N1.getOperand(1).getOpcode() == ISD::FSQRT) { | |||
10547 | SqrtOp = N1.getOperand(1); | |||
10548 | OtherOp = N1.getOperand(0); | |||
10549 | } | |||
10550 | if (SqrtOp.getNode()) { | |||
10551 | // We found a FSQRT, so try to make this fold: | |||
10552 | // x / (y * sqrt(z)) -> x * (rsqrt(z) / y) | |||
10553 | if (SDValue RV = buildRsqrtEstimate(SqrtOp.getOperand(0), Flags)) { | |||
10554 | RV = DAG.getNode(ISD::FDIV, SDLoc(N1), VT, RV, OtherOp, Flags); | |||
10555 | AddToWorklist(RV.getNode()); | |||
10556 | return DAG.getNode(ISD::FMUL, DL, VT, N0, RV, Flags); | |||
10557 | } | |||
10558 | } | |||
10559 | } | |||
10560 | ||||
10561 | // Fold into a reciprocal estimate and multiply instead of a real divide. | |||
10562 | if (SDValue RV = BuildReciprocalEstimate(N1, Flags)) { | |||
10563 | AddToWorklist(RV.getNode()); | |||
10564 | return DAG.getNode(ISD::FMUL, DL, VT, N0, RV, Flags); | |||
10565 | } | |||
10566 | } | |||
10567 | ||||
10568 | // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y) | |||
10569 | if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) { | |||
10570 | if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) { | |||
10571 | // Both can be negated for free, check to see if at least one is cheaper | |||
10572 | // negated. | |||
10573 | if (LHSNeg == 2 || RHSNeg == 2) | |||
10574 | return DAG.getNode(ISD::FDIV, SDLoc(N), VT, | |||
10575 | GetNegatedExpression(N0, DAG, LegalOperations), | |||
10576 | GetNegatedExpression(N1, DAG, LegalOperations), | |||
10577 | Flags); | |||
10578 | } | |||
10579 | } | |||
10580 | ||||
10581 | if (SDValue CombineRepeatedDivisors = combineRepeatedFPDivisors(N)) | |||
10582 | return CombineRepeatedDivisors; | |||
10583 | ||||
10584 | return SDValue(); | |||
10585 | } | |||
10586 | ||||
10587 | SDValue DAGCombiner::visitFREM(SDNode *N) { | |||
10588 | SDValue N0 = N->getOperand(0); | |||
10589 | SDValue N1 = N->getOperand(1); | |||
10590 | ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); | |||
10591 | ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); | |||
10592 | EVT VT = N->getValueType(0); | |||
10593 | ||||
10594 | // fold (frem c1, c2) -> fmod(c1,c2) | |||
10595 | if (N0CFP && N1CFP) | |||
10596 | return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1, N->getFlags()); | |||
10597 | ||||
10598 | if (SDValue NewSel = foldBinOpIntoSelect(N)) | |||
10599 | return NewSel; | |||
10600 | ||||
10601 | return SDValue(); | |||
10602 | } | |||
10603 | ||||
10604 | SDValue DAGCombiner::visitFSQRT(SDNode *N) { | |||
10605 | if (!DAG.getTarget().Options.UnsafeFPMath) | |||
10606 | return SDValue(); | |||
10607 | ||||
10608 | SDValue N0 = N->getOperand(0); | |||
10609 | if (TLI.isFsqrtCheap(N0, DAG)) | |||
10610 | return SDValue(); | |||
10611 | ||||
10612 | // TODO: FSQRT nodes should have flags that propagate to the created nodes. | |||
10613 | // For now, create a Flags object for use with all unsafe math transforms. | |||
10614 | SDNodeFlags Flags; | |||
10615 | Flags.setUnsafeAlgebra(true); | |||
10616 | return buildSqrtEstimate(N0, Flags); | |||
10617 | } | |||
10618 | ||||
10619 | /// copysign(x, fp_extend(y)) -> copysign(x, y) | |||
10620 | /// copysign(x, fp_round(y)) -> copysign(x, y) | |||
10621 | static inline bool CanCombineFCOPYSIGN_EXTEND_ROUND(SDNode *N) { | |||
10622 | SDValue N1 = N->getOperand(1); | |||
10623 | if ((N1.getOpcode() == ISD::FP_EXTEND || | |||
10624 | N1.getOpcode() == ISD::FP_ROUND)) { | |||
10625 | // Do not optimize out type conversion of f128 type yet. | |||
10626 | // For some targets like x86_64, configuration is changed to keep one f128 | |||
10627 | // value in one SSE register, but instruction selection cannot handle | |||
10628 | // FCOPYSIGN on SSE registers yet. | |||
10629 | EVT N1VT = N1->getValueType(0); | |||
10630 | EVT N1Op0VT = N1->getOperand(0).getValueType(); | |||
10631 | return (N1VT == N1Op0VT || N1Op0VT != MVT::f128); | |||
10632 | } | |||
10633 | return false; | |||
10634 | } | |||
10635 | ||||
10636 | SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) { | |||
10637 | SDValue N0 = N->getOperand(0); | |||
10638 | SDValue N1 = N->getOperand(1); | |||
10639 | ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); | |||
10640 | ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); | |||
10641 | EVT VT = N->getValueType(0); | |||
10642 | ||||
10643 | if (N0CFP && N1CFP) // Constant fold | |||
10644 | return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1); | |||
10645 | ||||
10646 | if (N1CFP) { | |||
10647 | const APFloat &V = N1CFP->getValueAPF(); | |||
10648 | // copysign(x, c1) -> fabs(x) iff ispos(c1) | |||
10649 | // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1) | |||
10650 | if (!V.isNegative()) { | |||
10651 | if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT)) | |||
10652 | return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0); | |||
10653 | } else { | |||
10654 | if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT)) | |||
10655 | return DAG.getNode(ISD::FNEG, SDLoc(N), VT, | |||
10656 | DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0)); | |||
10657 | } | |||
10658 | } | |||
10659 | ||||
10660 | // copysign(fabs(x), y) -> copysign(x, y) | |||
10661 | // copysign(fneg(x), y) -> copysign(x, y) | |||
10662 | // copysign(copysign(x,z), y) -> copysign(x, y) | |||
10663 | if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG || | |||
10664 | N0.getOpcode() == ISD::FCOPYSIGN) | |||
10665 | return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0.getOperand(0), N1); | |||
10666 | ||||
10667 | // copysign(x, abs(y)) -> abs(x) | |||
10668 | if (N1.getOpcode() == ISD::FABS) | |||
10669 | return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0); | |||
10670 | ||||
10671 | // copysign(x, copysign(y,z)) -> copysign(x, z) | |||
10672 | if (N1.getOpcode() == ISD::FCOPYSIGN) | |||
10673 | return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1.getOperand(1)); | |||
10674 | ||||
10675 | // copysign(x, fp_extend(y)) -> copysign(x, y) | |||
10676 | // copysign(x, fp_round(y)) -> copysign(x, y) | |||
10677 | if (CanCombineFCOPYSIGN_EXTEND_ROUND(N)) | |||
10678 | return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1.getOperand(0)); | |||
10679 | ||||
10680 | return SDValue(); | |||
10681 | } | |||
10682 | ||||
10683 | SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) { | |||
10684 | SDValue N0 = N->getOperand(0); | |||
10685 | EVT VT = N->getValueType(0); | |||
10686 | EVT OpVT = N0.getValueType(); | |||
10687 | ||||
10688 | // fold (sint_to_fp c1) -> c1fp | |||
10689 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
10690 | // ...but only if the target supports immediate floating-point values | |||
10691 | (!LegalOperations || | |||
10692 | TLI.isOperationLegalOrCustom(ISD::ConstantFP, VT))) | |||
10693 | return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0); | |||
10694 | ||||
10695 | // If the input is a legal type, and SINT_TO_FP is not legal on this target, | |||
10696 | // but UINT_TO_FP is legal on this target, try to convert. | |||
10697 | if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) && | |||
10698 | TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) { | |||
10699 | // If the sign bit is known to be zero, we can change this to UINT_TO_FP. | |||
10700 | if (DAG.SignBitIsZero(N0)) | |||
10701 | return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0); | |||
10702 | } | |||
10703 | ||||
10704 | // The next optimizations are desirable only if SELECT_CC can be lowered. | |||
10705 | if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) { | |||
10706 | // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc) | |||
10707 | if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 && | |||
10708 | !VT.isVector() && | |||
10709 | (!LegalOperations || | |||
10710 | TLI.isOperationLegalOrCustom(ISD::ConstantFP, VT))) { | |||
10711 | SDLoc DL(N); | |||
10712 | SDValue Ops[] = | |||
10713 | { N0.getOperand(0), N0.getOperand(1), | |||
10714 | DAG.getConstantFP(-1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT), | |||
10715 | N0.getOperand(2) }; | |||
10716 | return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops); | |||
10717 | } | |||
10718 | ||||
10719 | // fold (sint_to_fp (zext (setcc x, y, cc))) -> | |||
10720 | // (select_cc x, y, 1.0, 0.0,, cc) | |||
10721 | if (N0.getOpcode() == ISD::ZERO_EXTEND && | |||
10722 | N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() && | |||
10723 | (!LegalOperations || | |||
10724 | TLI.isOperationLegalOrCustom(ISD::ConstantFP, VT))) { | |||
10725 | SDLoc DL(N); | |||
10726 | SDValue Ops[] = | |||
10727 | { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1), | |||
10728 | DAG.getConstantFP(1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT), | |||
10729 | N0.getOperand(0).getOperand(2) }; | |||
10730 | return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops); | |||
10731 | } | |||
10732 | } | |||
10733 | ||||
10734 | return SDValue(); | |||
10735 | } | |||
10736 | ||||
10737 | SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) { | |||
10738 | SDValue N0 = N->getOperand(0); | |||
10739 | EVT VT = N->getValueType(0); | |||
10740 | EVT OpVT = N0.getValueType(); | |||
10741 | ||||
10742 | // fold (uint_to_fp c1) -> c1fp | |||
10743 | if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && | |||
10744 | // ...but only if the target supports immediate floating-point values | |||
10745 | (!LegalOperations || | |||
10746 | TLI.isOperationLegalOrCustom(ISD::ConstantFP, VT))) | |||
10747 | return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0); | |||
10748 | ||||
10749 | // If the input is a legal type, and UINT_TO_FP is not legal on this target, | |||
10750 | // but SINT_TO_FP is legal on this target, try to convert. | |||
10751 | if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) && | |||
10752 | TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) { | |||
10753 | // If the sign bit is known to be zero, we can change this to SINT_TO_FP. | |||
10754 | if (DAG.SignBitIsZero(N0)) | |||
10755 | return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0); | |||
10756 | } | |||
10757 | ||||
10758 | // The next optimizations are desirable only if SELECT_CC can be lowered. | |||
10759 | if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) { | |||
10760 | // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc) | |||
10761 | if (N0.getOpcode() == ISD::SETCC && !VT.isVector() && | |||
10762 | (!LegalOperations || | |||
10763 | TLI.isOperationLegalOrCustom(ISD::ConstantFP, VT))) { | |||
10764 | SDLoc DL(N); | |||
10765 | SDValue Ops[] = | |||
10766 | { N0.getOperand(0), N0.getOperand(1), | |||
10767 | DAG.getConstantFP(1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT), | |||
10768 | N0.getOperand(2) }; | |||
10769 | return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops); | |||
10770 | } | |||
10771 | } | |||
10772 | ||||
10773 | return SDValue(); | |||
10774 | } | |||
10775 | ||||
10776 | // Fold (fp_to_{s/u}int ({s/u}int_to_fpx)) -> zext x, sext x, trunc x, or x | |||
10777 | static SDValue FoldIntToFPToInt(SDNode *N, SelectionDAG &DAG) { | |||
10778 | SDValue N0 = N->getOperand(0); | |||
10779 | EVT VT = N->getValueType(0); | |||
10780 | ||||
10781 | if (N0.getOpcode() != ISD::UINT_TO_FP && N0.getOpcode() != ISD::SINT_TO_FP) | |||
10782 | return SDValue(); | |||
10783 | ||||
10784 | SDValue Src = N0.getOperand(0); | |||
10785 | EVT SrcVT = Src.getValueType(); | |||
10786 | bool IsInputSigned = N0.getOpcode() == ISD::SINT_TO_FP; | |||
10787 | bool IsOutputSigned = N->getOpcode() == ISD::FP_TO_SINT; | |||
10788 | ||||
10789 | // We can safely assume the conversion won't overflow the output range, | |||
10790 | // because (for example) (uint8_t)18293.f is undefined behavior. | |||
10791 | ||||
10792 | // Since we can assume the conversion won't overflow, our decision as to | |||
10793 | // whether the input will fit in the float should depend on the minimum | |||
10794 | // of the input range and output range. | |||
10795 | ||||
10796 | // This means this is also safe for a signed input and unsigned output, since | |||
10797 | // a negative input would lead to undefined behavior. | |||
10798 | unsigned InputSize = (int)SrcVT.getScalarSizeInBits() - IsInputSigned; | |||
10799 | unsigned OutputSize = (int)VT.getScalarSizeInBits() - IsOutputSigned; | |||
10800 | unsigned ActualSize = std::min(InputSize, OutputSize); | |||
10801 | const fltSemantics &sem = DAG.EVTToAPFloatSemantics(N0.getValueType()); | |||
10802 | ||||
10803 | // We can only fold away the float conversion if the input range can be | |||
10804 | // represented exactly in the float range. | |||
10805 | if (APFloat::semanticsPrecision(sem) >= ActualSize) { | |||
10806 | if (VT.getScalarSizeInBits() > SrcVT.getScalarSizeInBits()) { | |||
10807 | unsigned ExtOp = IsInputSigned && IsOutputSigned ? ISD::SIGN_EXTEND | |||
10808 | : ISD::ZERO_EXTEND; | |||
10809 | return DAG.getNode(ExtOp, SDLoc(N), VT, Src); | |||
10810 | } | |||
10811 | if (VT.getScalarSizeInBits() < SrcVT.getScalarSizeInBits()) | |||
10812 | return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Src); | |||
10813 | return DAG.getBitcast(VT, Src); | |||
10814 | } | |||
10815 | return SDValue(); | |||
10816 | } | |||
10817 | ||||
10818 | SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) { | |||
10819 | SDValue N0 = N->getOperand(0); | |||
10820 | EVT VT = N->getValueType(0); | |||
10821 | ||||
10822 | // fold (fp_to_sint c1fp) -> c1 | |||
10823 | if (isConstantFPBuildVectorOrConstantFP(N0)) | |||
10824 | return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0); | |||
10825 | ||||
10826 | return FoldIntToFPToInt(N, DAG); | |||
10827 | } | |||
10828 | ||||
10829 | SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) { | |||
10830 | SDValue N0 = N->getOperand(0); | |||
10831 | EVT VT = N->getValueType(0); | |||
10832 | ||||
10833 | // fold (fp_to_uint c1fp) -> c1 | |||
10834 | if (isConstantFPBuildVectorOrConstantFP(N0)) | |||
10835 | return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0); | |||
10836 | ||||
10837 | return FoldIntToFPToInt(N, DAG); | |||
10838 | } | |||
10839 | ||||
10840 | SDValue DAGCombiner::visitFP_ROUND(SDNode *N) { | |||
10841 | SDValue N0 = N->getOperand(0); | |||
10842 | SDValue N1 = N->getOperand(1); | |||
10843 | ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); | |||
10844 | EVT VT = N->getValueType(0); | |||
10845 | ||||
10846 | // fold (fp_round c1fp) -> c1fp | |||
10847 | if (N0CFP) | |||
10848 | return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1); | |||
10849 | ||||
10850 | // fold (fp_round (fp_extend x)) -> x | |||
10851 | if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType()) | |||
10852 | return N0.getOperand(0); | |||
10853 | ||||
10854 | // fold (fp_round (fp_round x)) -> (fp_round x) | |||
10855 | if (N0.getOpcode() == ISD::FP_ROUND) { | |||
10856 | const bool NIsTrunc = N->getConstantOperandVal(1) == 1; | |||
10857 | const bool N0IsTrunc = N0.getConstantOperandVal(1) == 1; | |||
10858 | ||||
10859 | // Skip this folding if it results in an fp_round from f80 to f16. | |||
10860 | // | |||
10861 | // f80 to f16 always generates an expensive (and as yet, unimplemented) | |||
10862 | // libcall to __truncxfhf2 instead of selecting native f16 conversion | |||
10863 | // instructions from f32 or f64. Moreover, the first (value-preserving) | |||
10864 | // fp_round from f80 to either f32 or f64 may become a NOP in platforms like | |||
10865 | // x86. | |||
10866 | if (N0.getOperand(0).getValueType() == MVT::f80 && VT == MVT::f16) | |||
10867 | return SDValue(); | |||
10868 | ||||
10869 | // If the first fp_round isn't a value preserving truncation, it might | |||
10870 | // introduce a tie in the second fp_round, that wouldn't occur in the | |||
10871 | // single-step fp_round we want to fold to. | |||
10872 | // In other words, double rounding isn't the same as rounding. | |||
10873 | // Also, this is a value preserving truncation iff both fp_round's are. | |||
10874 | if (DAG.getTarget().Options.UnsafeFPMath || N0IsTrunc) { | |||
10875 | SDLoc DL(N); | |||
10876 | return DAG.getNode(ISD::FP_ROUND, DL, VT, N0.getOperand(0), | |||
10877 | DAG.getIntPtrConstant(NIsTrunc && N0IsTrunc, DL)); | |||
10878 | } | |||
10879 | } | |||
10880 | ||||
10881 | // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y) | |||
10882 | if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) { | |||
10883 | SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT, | |||
10884 | N0.getOperand(0), N1); | |||
10885 | AddToWorklist(Tmp.getNode()); | |||
10886 | return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, | |||
10887 | Tmp, N0.getOperand(1)); | |||
10888 | } | |||
10889 | ||||
10890 | if (SDValue NewVSel = matchVSelectOpSizesWithSetCC(N)) | |||
10891 | return NewVSel; | |||
10892 | ||||
10893 | return SDValue(); | |||
10894 | } | |||
10895 | ||||
10896 | SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) { | |||
10897 | SDValue N0 = N->getOperand(0); | |||
10898 | EVT VT = N->getValueType(0); | |||
10899 | EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT(); | |||
10900 | ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); | |||
10901 | ||||
10902 | // fold (fp_round_inreg c1fp) -> c1fp | |||
10903 | if (N0CFP && isTypeLegal(EVT)) { | |||
10904 | SDLoc DL(N); | |||
10905 | SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), DL, EVT); | |||
10906 | return DAG.getNode(ISD::FP_EXTEND, DL, VT, Round); | |||
10907 | } | |||
10908 | ||||
10909 | return SDValue(); | |||
10910 | } | |||
10911 | ||||
10912 | SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) { | |||
10913 | SDValue N0 = N->getOperand(0); | |||
10914 | EVT VT = N->getValueType(0); | |||
10915 | ||||
10916 | // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded. | |||
10917 | if (N->hasOneUse() && | |||
10918 | N->use_begin()->getOpcode() == ISD::FP_ROUND) | |||
10919 | return SDValue(); | |||
10920 | ||||
10921 | // fold (fp_extend c1fp) -> c1fp | |||
10922 | if (isConstantFPBuildVectorOrConstantFP(N0)) | |||
10923 | return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0); | |||
10924 | ||||
10925 | // fold (fp_extend (fp16_to_fp op)) -> (fp16_to_fp op) | |||
10926 | if (N0.getOpcode() == ISD::FP16_TO_FP && | |||
10927 | TLI.getOperationAction(ISD::FP16_TO_FP, VT) == TargetLowering::Legal) | |||
10928 | return DAG.getNode(ISD::FP16_TO_FP, SDLoc(N), VT, N0.getOperand(0)); | |||
10929 | ||||
10930 | // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the | |||
10931 | // value of X. | |||
10932 | if (N0.getOpcode() == ISD::FP_ROUND | |||
10933 | && N0.getConstantOperandVal(1) == 1) { | |||
10934 | SDValue In = N0.getOperand(0); | |||
10935 | if (In.getValueType() == VT) return In; | |||
10936 | if (VT.bitsLT(In.getValueType())) | |||
10937 | return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, | |||
10938 | In, N0.getOperand(1)); | |||
10939 | return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In); | |||
10940 | } | |||
10941 | ||||
10942 | // fold (fpext (load x)) -> (fpext (fptrunc (extload x))) | |||
10943 | if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() && | |||
10944 | TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) { | |||
10945 | LoadSDNode *LN0 = cast<LoadSDNode>(N0); | |||
10946 | SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT, | |||
10947 | LN0->getChain(), | |||
10948 | LN0->getBasePtr(), N0.getValueType(), | |||
10949 | LN0->getMemOperand()); | |||
10950 | CombineTo(N, ExtLoad); | |||
10951 | CombineTo(N0.getNode(), | |||
10952 | DAG.getNode(ISD::FP_ROUND, SDLoc(N0), | |||
10953 | N0.getValueType(), ExtLoad, | |||
10954 | DAG.getIntPtrConstant(1, SDLoc(N0))), | |||
10955 | ExtLoad.getValue(1)); | |||
10956 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
10957 | } | |||
10958 | ||||
10959 | if (SDValue NewVSel = matchVSelectOpSizesWithSetCC(N)) | |||
10960 | return NewVSel; | |||
10961 | ||||
10962 | return SDValue(); | |||
10963 | } | |||
10964 | ||||
10965 | SDValue DAGCombiner::visitFCEIL(SDNode *N) { | |||
10966 | SDValue N0 = N->getOperand(0); | |||
10967 | EVT VT = N->getValueType(0); | |||
10968 | ||||
10969 | // fold (fceil c1) -> fceil(c1) | |||
10970 | if (isConstantFPBuildVectorOrConstantFP(N0)) | |||
10971 | return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0); | |||
10972 | ||||
10973 | return SDValue(); | |||
10974 | } | |||
10975 | ||||
10976 | SDValue DAGCombiner::visitFTRUNC(SDNode *N) { | |||
10977 | SDValue N0 = N->getOperand(0); | |||
10978 | EVT VT = N->getValueType(0); | |||
10979 | ||||
10980 | // fold (ftrunc c1) -> ftrunc(c1) | |||
10981 | if (isConstantFPBuildVectorOrConstantFP(N0)) | |||
10982 | return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0); | |||
10983 | ||||
10984 | // fold ftrunc (known rounded int x) -> x | |||
10985 | // ftrunc is a part of fptosi/fptoui expansion on some targets, so this is | |||
10986 | // likely to be generated to extract integer from a rounded floating value. | |||
10987 | switch (N0.getOpcode()) { | |||
10988 | default: break; | |||
10989 | case ISD::FRINT: | |||
10990 | case ISD::FTRUNC: | |||
10991 | case ISD::FNEARBYINT: | |||
10992 | case ISD::FFLOOR: | |||
10993 | case ISD::FCEIL: | |||
10994 | return N0; | |||
10995 | } | |||
10996 | ||||
10997 | return SDValue(); | |||
10998 | } | |||
10999 | ||||
11000 | SDValue DAGCombiner::visitFFLOOR(SDNode *N) { | |||
11001 | SDValue N0 = N->getOperand(0); | |||
11002 | EVT VT = N->getValueType(0); | |||
11003 | ||||
11004 | // fold (ffloor c1) -> ffloor(c1) | |||
11005 | if (isConstantFPBuildVectorOrConstantFP(N0)) | |||
11006 | return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0); | |||
11007 | ||||
11008 | return SDValue(); | |||
11009 | } | |||
11010 | ||||
11011 | // FIXME: FNEG and FABS have a lot in common; refactor. | |||
11012 | SDValue DAGCombiner::visitFNEG(SDNode *N) { | |||
11013 | SDValue N0 = N->getOperand(0); | |||
11014 | EVT VT = N->getValueType(0); | |||
11015 | ||||
11016 | // Constant fold FNEG. | |||
11017 | if (isConstantFPBuildVectorOrConstantFP(N0)) | |||
11018 | return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0); | |||
11019 | ||||
11020 | if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(), | |||
11021 | &DAG.getTarget().Options)) | |||
11022 | return GetNegatedExpression(N0, DAG, LegalOperations); | |||
11023 | ||||
11024 | // Transform fneg(bitconvert(x)) -> bitconvert(x ^ sign) to avoid loading | |||
11025 | // constant pool values. | |||
11026 | if (!TLI.isFNegFree(VT) && | |||
11027 | N0.getOpcode() == ISD::BITCAST && | |||
11028 | N0.getNode()->hasOneUse()) { | |||
11029 | SDValue Int = N0.getOperand(0); | |||
11030 | EVT IntVT = Int.getValueType(); | |||
11031 | if (IntVT.isInteger() && !IntVT.isVector()) { | |||
11032 | APInt SignMask; | |||
11033 | if (N0.getValueType().isVector()) { | |||
11034 | // For a vector, get a mask such as 0x80... per scalar element | |||
11035 | // and splat it. | |||
11036 | SignMask = APInt::getSignMask(N0.getScalarValueSizeInBits()); | |||
11037 | SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask); | |||
11038 | } else { | |||
11039 | // For a scalar, just generate 0x80... | |||
11040 | SignMask = APInt::getSignMask(IntVT.getSizeInBits()); | |||
11041 | } | |||
11042 | SDLoc DL0(N0); | |||
11043 | Int = DAG.getNode(ISD::XOR, DL0, IntVT, Int, | |||
11044 | DAG.getConstant(SignMask, DL0, IntVT)); | |||
11045 | AddToWorklist(Int.getNode()); | |||
11046 | return DAG.getBitcast(VT, Int); | |||
11047 | } | |||
11048 | } | |||
11049 | ||||
11050 | // (fneg (fmul c, x)) -> (fmul -c, x) | |||
11051 | if (N0.getOpcode() == ISD::FMUL && | |||
11052 | (N0.getNode()->hasOneUse() || !TLI.isFNegFree(VT))) { | |||
11053 | ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1)); | |||
11054 | if (CFP1) { | |||
11055 | APFloat CVal = CFP1->getValueAPF(); | |||
11056 | CVal.changeSign(); | |||
11057 | if (Level >= AfterLegalizeDAG && | |||
11058 | (TLI.isFPImmLegal(CVal, VT) || | |||
11059 | TLI.isOperationLegal(ISD::ConstantFP, VT))) | |||
11060 | return DAG.getNode( | |||
11061 | ISD::FMUL, SDLoc(N), VT, N0.getOperand(0), | |||
11062 | DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)), | |||
11063 | N0->getFlags()); | |||
11064 | } | |||
11065 | } | |||
11066 | ||||
11067 | return SDValue(); | |||
11068 | } | |||
11069 | ||||
11070 | SDValue DAGCombiner::visitFMINNUM(SDNode *N) { | |||
11071 | SDValue N0 = N->getOperand(0); | |||
11072 | SDValue N1 = N->getOperand(1); | |||
11073 | EVT VT = N->getValueType(0); | |||
11074 | const ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0); | |||
11075 | const ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1); | |||
11076 | ||||
11077 | if (N0CFP && N1CFP) { | |||
11078 | const APFloat &C0 = N0CFP->getValueAPF(); | |||
11079 | const APFloat &C1 = N1CFP->getValueAPF(); | |||
11080 | return DAG.getConstantFP(minnum(C0, C1), SDLoc(N), VT); | |||
11081 | } | |||
11082 | ||||
11083 | // Canonicalize to constant on RHS. | |||
11084 | if (isConstantFPBuildVectorOrConstantFP(N0) && | |||
11085 | !isConstantFPBuildVectorOrConstantFP(N1)) | |||
11086 | return DAG.getNode(ISD::FMINNUM, SDLoc(N), VT, N1, N0); | |||
11087 | ||||
11088 | return SDValue(); | |||
11089 | } | |||
11090 | ||||
11091 | SDValue DAGCombiner::visitFMAXNUM(SDNode *N) { | |||
11092 | SDValue N0 = N->getOperand(0); | |||
11093 | SDValue N1 = N->getOperand(1); | |||
11094 | EVT VT = N->getValueType(0); | |||
11095 | const ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0); | |||
11096 | const ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1); | |||
11097 | ||||
11098 | if (N0CFP && N1CFP) { | |||
11099 | const APFloat &C0 = N0CFP->getValueAPF(); | |||
11100 | const APFloat &C1 = N1CFP->getValueAPF(); | |||
11101 | return DAG.getConstantFP(maxnum(C0, C1), SDLoc(N), VT); | |||
11102 | } | |||
11103 | ||||
11104 | // Canonicalize to constant on RHS. | |||
11105 | if (isConstantFPBuildVectorOrConstantFP(N0) && | |||
11106 | !isConstantFPBuildVectorOrConstantFP(N1)) | |||
11107 | return DAG.getNode(ISD::FMAXNUM, SDLoc(N), VT, N1, N0); | |||
11108 | ||||
11109 | return SDValue(); | |||
11110 | } | |||
11111 | ||||
11112 | SDValue DAGCombiner::visitFABS(SDNode *N) { | |||
11113 | SDValue N0 = N->getOperand(0); | |||
11114 | EVT VT = N->getValueType(0); | |||
11115 | ||||
11116 | // fold (fabs c1) -> fabs(c1) | |||
11117 | if (isConstantFPBuildVectorOrConstantFP(N0)) | |||
11118 | return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0); | |||
11119 | ||||
11120 | // fold (fabs (fabs x)) -> (fabs x) | |||
11121 | if (N0.getOpcode() == ISD::FABS) | |||
11122 | return N->getOperand(0); | |||
11123 | ||||
11124 | // fold (fabs (fneg x)) -> (fabs x) | |||
11125 | // fold (fabs (fcopysign x, y)) -> (fabs x) | |||
11126 | if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN) | |||
11127 | return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0)); | |||
11128 | ||||
11129 | // Transform fabs(bitconvert(x)) -> bitconvert(x & ~sign) to avoid loading | |||
11130 | // constant pool values. | |||
11131 | if (!TLI.isFAbsFree(VT) && | |||
11132 | N0.getOpcode() == ISD::BITCAST && | |||
11133 | N0.getNode()->hasOneUse()) { | |||
11134 | SDValue Int = N0.getOperand(0); | |||
11135 | EVT IntVT = Int.getValueType(); | |||
11136 | if (IntVT.isInteger() && !IntVT.isVector()) { | |||
11137 | APInt SignMask; | |||
11138 | if (N0.getValueType().isVector()) { | |||
11139 | // For a vector, get a mask such as 0x7f... per scalar element | |||
11140 | // and splat it. | |||
11141 | SignMask = ~APInt::getSignMask(N0.getScalarValueSizeInBits()); | |||
11142 | SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask); | |||
11143 | } else { | |||
11144 | // For a scalar, just generate 0x7f... | |||
11145 | SignMask = ~APInt::getSignMask(IntVT.getSizeInBits()); | |||
11146 | } | |||
11147 | SDLoc DL(N0); | |||
11148 | Int = DAG.getNode(ISD::AND, DL, IntVT, Int, | |||
11149 | DAG.getConstant(SignMask, DL, IntVT)); | |||
11150 | AddToWorklist(Int.getNode()); | |||
11151 | return DAG.getBitcast(N->getValueType(0), Int); | |||
11152 | } | |||
11153 | } | |||
11154 | ||||
11155 | return SDValue(); | |||
11156 | } | |||
11157 | ||||
11158 | SDValue DAGCombiner::visitBRCOND(SDNode *N) { | |||
11159 | SDValue Chain = N->getOperand(0); | |||
11160 | SDValue N1 = N->getOperand(1); | |||
11161 | SDValue N2 = N->getOperand(2); | |||
11162 | ||||
11163 | // If N is a constant we could fold this into a fallthrough or unconditional | |||
11164 | // branch. However that doesn't happen very often in normal code, because | |||
11165 | // Instcombine/SimplifyCFG should have handled the available opportunities. | |||
11166 | // If we did this folding here, it would be necessary to update the | |||
11167 | // MachineBasicBlock CFG, which is awkward. | |||
11168 | ||||
11169 | // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal | |||
11170 | // on the target. | |||
11171 | if (N1.getOpcode() == ISD::SETCC && | |||
11172 | TLI.isOperationLegalOrCustom(ISD::BR_CC, | |||
11173 | N1.getOperand(0).getValueType())) { | |||
11174 | return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other, | |||
11175 | Chain, N1.getOperand(2), | |||
11176 | N1.getOperand(0), N1.getOperand(1), N2); | |||
11177 | } | |||
11178 | ||||
11179 | if (N1.hasOneUse()) { | |||
11180 | if (SDValue NewN1 = rebuildSetCC(N1)) | |||
11181 | return DAG.getNode(ISD::BRCOND, SDLoc(N), MVT::Other, Chain, NewN1, N2); | |||
11182 | } | |||
11183 | ||||
11184 | return SDValue(); | |||
11185 | } | |||
11186 | ||||
11187 | SDValue DAGCombiner::rebuildSetCC(SDValue N) { | |||
11188 | if (N.getOpcode() == ISD::SRL || | |||
11189 | (N.getOpcode() == ISD::TRUNCATE && | |||
11190 | (N.getOperand(0).hasOneUse() && | |||
11191 | N.getOperand(0).getOpcode() == ISD::SRL))) { | |||
11192 | // Look pass the truncate. | |||
11193 | if (N.getOpcode() == ISD::TRUNCATE) | |||
11194 | N = N.getOperand(0); | |||
11195 | ||||
11196 | // Match this pattern so that we can generate simpler code: | |||
11197 | // | |||
11198 | // %a = ... | |||
11199 | // %b = and i32 %a, 2 | |||
11200 | // %c = srl i32 %b, 1 | |||
11201 | // brcond i32 %c ... | |||
11202 | // | |||
11203 | // into | |||
11204 | // | |||
11205 | // %a = ... | |||
11206 | // %b = and i32 %a, 2 | |||
11207 | // %c = setcc eq %b, 0 | |||
11208 | // brcond %c ... | |||
11209 | // | |||
11210 | // This applies only when the AND constant value has one bit set and the | |||
11211 | // SRL constant is equal to the log2 of the AND constant. The back-end is | |||
11212 | // smart enough to convert the result into a TEST/JMP sequence. | |||
11213 | SDValue Op0 = N.getOperand(0); | |||
11214 | SDValue Op1 = N.getOperand(1); | |||
11215 | ||||
11216 | if (Op0.getOpcode() == ISD::AND && Op1.getOpcode() == ISD::Constant) { | |||
11217 | SDValue AndOp1 = Op0.getOperand(1); | |||
11218 | ||||
11219 | if (AndOp1.getOpcode() == ISD::Constant) { | |||
11220 | const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue(); | |||
11221 | ||||
11222 | if (AndConst.isPowerOf2() && | |||
11223 | cast<ConstantSDNode>(Op1)->getAPIntValue() == AndConst.logBase2()) { | |||
11224 | SDLoc DL(N); | |||
11225 | return DAG.getSetCC(DL, getSetCCResultType(Op0.getValueType()), | |||
11226 | Op0, DAG.getConstant(0, DL, Op0.getValueType()), | |||
11227 | ISD::SETNE); | |||
11228 | } | |||
11229 | } | |||
11230 | } | |||
11231 | } | |||
11232 | ||||
11233 | // Transform br(xor(x, y)) -> br(x != y) | |||
11234 | // Transform br(xor(xor(x,y), 1)) -> br (x == y) | |||
11235 | if (N.getOpcode() == ISD::XOR) { | |||
11236 | SDNode *TheXor = N.getNode(); | |||
11237 | ||||
11238 | // Avoid missing important xor optimizations. | |||
11239 | while (SDValue Tmp = visitXOR(TheXor)) { | |||
11240 | // We don't have a XOR anymore, bail. | |||
11241 | if (Tmp.getOpcode() != ISD::XOR) | |||
11242 | return Tmp; | |||
11243 | ||||
11244 | TheXor = Tmp.getNode(); | |||
11245 | } | |||
11246 | ||||
11247 | SDValue Op0 = TheXor->getOperand(0); | |||
11248 | SDValue Op1 = TheXor->getOperand(1); | |||
11249 | ||||
11250 | if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) { | |||
11251 | bool Equal = false; | |||
11252 | if (isOneConstant(Op0) && Op0.hasOneUse() && | |||
11253 | Op0.getOpcode() == ISD::XOR) { | |||
11254 | TheXor = Op0.getNode(); | |||
11255 | Equal = true; | |||
11256 | } | |||
11257 | ||||
11258 | EVT SetCCVT = N.getValueType(); | |||
11259 | if (LegalTypes) | |||
11260 | SetCCVT = getSetCCResultType(SetCCVT); | |||
11261 | // Replace the uses of XOR with SETCC | |||
11262 | return DAG.getSetCC(SDLoc(TheXor), SetCCVT, Op0, Op1, | |||
11263 | Equal ? ISD::SETEQ : ISD::SETNE); | |||
11264 | } | |||
11265 | } | |||
11266 | ||||
11267 | return SDValue(); | |||
11268 | } | |||
11269 | ||||
11270 | // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB. | |||
11271 | // | |||
11272 | SDValue DAGCombiner::visitBR_CC(SDNode *N) { | |||
11273 | CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1)); | |||
11274 | SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3); | |||
11275 | ||||
11276 | // If N is a constant we could fold this into a fallthrough or unconditional | |||
11277 | // branch. However that doesn't happen very often in normal code, because | |||
11278 | // Instcombine/SimplifyCFG should have handled the available opportunities. | |||
11279 | // If we did this folding here, it would be necessary to update the | |||
11280 | // MachineBasicBlock CFG, which is awkward. | |||
11281 | ||||
11282 | // Use SimplifySetCC to simplify SETCC's. | |||
11283 | SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()), | |||
11284 | CondLHS, CondRHS, CC->get(), SDLoc(N), | |||
11285 | false); | |||
11286 | if (Simp.getNode()) AddToWorklist(Simp.getNode()); | |||
11287 | ||||
11288 | // fold to a simpler setcc | |||
11289 | if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC) | |||
11290 | return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other, | |||
11291 | N->getOperand(0), Simp.getOperand(2), | |||
11292 | Simp.getOperand(0), Simp.getOperand(1), | |||
11293 | N->getOperand(4)); | |||
11294 | ||||
11295 | return SDValue(); | |||
11296 | } | |||
11297 | ||||
11298 | /// Return true if 'Use' is a load or a store that uses N as its base pointer | |||
11299 | /// and that N may be folded in the load / store addressing mode. | |||
11300 | static bool canFoldInAddressingMode(SDNode *N, SDNode *Use, | |||
11301 | SelectionDAG &DAG, | |||
11302 | const TargetLowering &TLI) { | |||
11303 | EVT VT; | |||
11304 | unsigned AS; | |||
11305 | ||||
11306 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) { | |||
11307 | if (LD->isIndexed() || LD->getBasePtr().getNode() != N) | |||
11308 | return false; | |||
11309 | VT = LD->getMemoryVT(); | |||
11310 | AS = LD->getAddressSpace(); | |||
11311 | } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) { | |||
11312 | if (ST->isIndexed() || ST->getBasePtr().getNode() != N) | |||
11313 | return false; | |||
11314 | VT = ST->getMemoryVT(); | |||
11315 | AS = ST->getAddressSpace(); | |||
11316 | } else | |||
11317 | return false; | |||
11318 | ||||
11319 | TargetLowering::AddrMode AM; | |||
11320 | if (N->getOpcode() == ISD::ADD) { | |||
11321 | ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1)); | |||
11322 | if (Offset) | |||
11323 | // [reg +/- imm] | |||
11324 | AM.BaseOffs = Offset->getSExtValue(); | |||
11325 | else | |||
11326 | // [reg +/- reg] | |||
11327 | AM.Scale = 1; | |||
11328 | } else if (N->getOpcode() == ISD::SUB) { | |||
11329 | ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1)); | |||
11330 | if (Offset) | |||
11331 | // [reg +/- imm] | |||
11332 | AM.BaseOffs = -Offset->getSExtValue(); | |||
11333 | else | |||
11334 | // [reg +/- reg] | |||
11335 | AM.Scale = 1; | |||
11336 | } else | |||
11337 | return false; | |||
11338 | ||||
11339 | return TLI.isLegalAddressingMode(DAG.getDataLayout(), AM, | |||
11340 | VT.getTypeForEVT(*DAG.getContext()), AS); | |||
11341 | } | |||
11342 | ||||
11343 | /// Try turning a load/store into a pre-indexed load/store when the base | |||
11344 | /// pointer is an add or subtract and it has other uses besides the load/store. | |||
11345 | /// After the transformation, the new indexed load/store has effectively folded | |||
11346 | /// the add/subtract in and all of its other uses are redirected to the | |||
11347 | /// new load/store. | |||
11348 | bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) { | |||
11349 | if (Level < AfterLegalizeDAG) | |||
11350 | return false; | |||
11351 | ||||
11352 | bool isLoad = true; | |||
11353 | SDValue Ptr; | |||
11354 | EVT VT; | |||
11355 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { | |||
11356 | if (LD->isIndexed()) | |||
11357 | return false; | |||
11358 | VT = LD->getMemoryVT(); | |||
11359 | if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) && | |||
11360 | !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT)) | |||
11361 | return false; | |||
11362 | Ptr = LD->getBasePtr(); | |||
11363 | } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { | |||
11364 | if (ST->isIndexed()) | |||
11365 | return false; | |||
11366 | VT = ST->getMemoryVT(); | |||
11367 | if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) && | |||
11368 | !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT)) | |||
11369 | return false; | |||
11370 | Ptr = ST->getBasePtr(); | |||
11371 | isLoad = false; | |||
11372 | } else { | |||
11373 | return false; | |||
11374 | } | |||
11375 | ||||
11376 | // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail | |||
11377 | // out. There is no reason to make this a preinc/predec. | |||
11378 | if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) || | |||
11379 | Ptr.getNode()->hasOneUse()) | |||
11380 | return false; | |||
11381 | ||||
11382 | // Ask the target to do addressing mode selection. | |||
11383 | SDValue BasePtr; | |||
11384 | SDValue Offset; | |||
11385 | ISD::MemIndexedMode AM = ISD::UNINDEXED; | |||
11386 | if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG)) | |||
11387 | return false; | |||
11388 | ||||
11389 | // Backends without true r+i pre-indexed forms may need to pass a | |||
11390 | // constant base with a variable offset so that constant coercion | |||
11391 | // will work with the patterns in canonical form. | |||
11392 | bool Swapped = false; | |||
11393 | if (isa<ConstantSDNode>(BasePtr)) { | |||
11394 | std::swap(BasePtr, Offset); | |||
11395 | Swapped = true; | |||
11396 | } | |||
11397 | ||||
11398 | // Don't create a indexed load / store with zero offset. | |||
11399 | if (isNullConstant(Offset)) | |||
11400 | return false; | |||
11401 | ||||
11402 | // Try turning it into a pre-indexed load / store except when: | |||
11403 | // 1) The new base ptr is a frame index. | |||
11404 | // 2) If N is a store and the new base ptr is either the same as or is a | |||
11405 | // predecessor of the value being stored. | |||
11406 | // 3) Another use of old base ptr is a predecessor of N. If ptr is folded | |||
11407 | // that would create a cycle. | |||
11408 | // 4) All uses are load / store ops that use it as old base ptr. | |||
11409 | ||||
11410 | // Check #1. Preinc'ing a frame index would require copying the stack pointer | |||
11411 | // (plus the implicit offset) to a register to preinc anyway. | |||
11412 | if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr)) | |||
11413 | return false; | |||
11414 | ||||
11415 | // Check #2. | |||
11416 | if (!isLoad) { | |||
11417 | SDValue Val = cast<StoreSDNode>(N)->getValue(); | |||
11418 | if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode())) | |||
11419 | return false; | |||
11420 | } | |||
11421 | ||||
11422 | // Caches for hasPredecessorHelper. | |||
11423 | SmallPtrSet<const SDNode *, 32> Visited; | |||
11424 | SmallVector<const SDNode *, 16> Worklist; | |||
11425 | Worklist.push_back(N); | |||
11426 | ||||
11427 | // If the offset is a constant, there may be other adds of constants that | |||
11428 | // can be folded with this one. We should do this to avoid having to keep | |||
11429 | // a copy of the original base pointer. | |||
11430 | SmallVector<SDNode *, 16> OtherUses; | |||
11431 | if (isa<ConstantSDNode>(Offset)) | |||
11432 | for (SDNode::use_iterator UI = BasePtr.getNode()->use_begin(), | |||
11433 | UE = BasePtr.getNode()->use_end(); | |||
11434 | UI != UE; ++UI) { | |||
11435 | SDUse &Use = UI.getUse(); | |||
11436 | // Skip the use that is Ptr and uses of other results from BasePtr's | |||
11437 | // node (important for nodes that return multiple results). | |||
11438 | if (Use.getUser() == Ptr.getNode() || Use != BasePtr) | |||
11439 | continue; | |||
11440 | ||||
11441 | if (SDNode::hasPredecessorHelper(Use.getUser(), Visited, Worklist)) | |||
11442 | continue; | |||
11443 | ||||
11444 | if (Use.getUser()->getOpcode() != ISD::ADD && | |||
11445 | Use.getUser()->getOpcode() != ISD::SUB) { | |||
11446 | OtherUses.clear(); | |||
11447 | break; | |||
11448 | } | |||
11449 | ||||
11450 | SDValue Op1 = Use.getUser()->getOperand((UI.getOperandNo() + 1) & 1); | |||
11451 | if (!isa<ConstantSDNode>(Op1)) { | |||
11452 | OtherUses.clear(); | |||
11453 | break; | |||
11454 | } | |||
11455 | ||||
11456 | // FIXME: In some cases, we can be smarter about this. | |||
11457 | if (Op1.getValueType() != Offset.getValueType()) { | |||
11458 | OtherUses.clear(); | |||
11459 | break; | |||
11460 | } | |||
11461 | ||||
11462 | OtherUses.push_back(Use.getUser()); | |||
11463 | } | |||
11464 | ||||
11465 | if (Swapped) | |||
11466 | std::swap(BasePtr, Offset); | |||
11467 | ||||
11468 | // Now check for #3 and #4. | |||
11469 | bool RealUse = false; | |||
11470 | ||||
11471 | for (SDNode *Use : Ptr.getNode()->uses()) { | |||
11472 | if (Use == N) | |||
11473 | continue; | |||
11474 | if (SDNode::hasPredecessorHelper(Use, Visited, Worklist)) | |||
11475 | return false; | |||
11476 | ||||
11477 | // If Ptr may be folded in addressing mode of other use, then it's | |||
11478 | // not profitable to do this transformation. | |||
11479 | if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI)) | |||
11480 | RealUse = true; | |||
11481 | } | |||
11482 | ||||
11483 | if (!RealUse) | |||
11484 | return false; | |||
11485 | ||||
11486 | SDValue Result; | |||
11487 | if (isLoad) | |||
11488 | Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N), | |||
11489 | BasePtr, Offset, AM); | |||
11490 | else | |||
11491 | Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N), | |||
11492 | BasePtr, Offset, AM); | |||
11493 | ++PreIndexedNodes; | |||
11494 | ++NodesCombined; | |||
11495 | DEBUG(dbgs() << "\nReplacing.4 ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.4 "; N->dump (&DAG); dbgs() << "\nWith: "; Result.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false) | |||
11496 | N->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.4 "; N->dump (&DAG); dbgs() << "\nWith: "; Result.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false) | |||
11497 | dbgs() << "\nWith: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.4 "; N->dump (&DAG); dbgs() << "\nWith: "; Result.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false) | |||
11498 | Result.getNode()->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.4 "; N->dump (&DAG); dbgs() << "\nWith: "; Result.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false) | |||
11499 | dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.4 "; N->dump (&DAG); dbgs() << "\nWith: "; Result.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false); | |||
11500 | WorklistRemover DeadNodes(*this); | |||
11501 | if (isLoad) { | |||
11502 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0)); | |||
11503 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2)); | |||
11504 | } else { | |||
11505 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1)); | |||
11506 | } | |||
11507 | ||||
11508 | // Finally, since the node is now dead, remove it from the graph. | |||
11509 | deleteAndRecombine(N); | |||
11510 | ||||
11511 | if (Swapped) | |||
11512 | std::swap(BasePtr, Offset); | |||
11513 | ||||
11514 | // Replace other uses of BasePtr that can be updated to use Ptr | |||
11515 | for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) { | |||
11516 | unsigned OffsetIdx = 1; | |||
11517 | if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode()) | |||
11518 | OffsetIdx = 0; | |||
11519 | assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==(static_cast <bool> (OtherUses[i]->getOperand(!OffsetIdx ).getNode() == BasePtr.getNode() && "Expected BasePtr operand" ) ? void (0) : __assert_fail ("OtherUses[i]->getOperand(!OffsetIdx).getNode() == BasePtr.getNode() && \"Expected BasePtr operand\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 11520, __extension__ __PRETTY_FUNCTION__)) | |||
11520 | BasePtr.getNode() && "Expected BasePtr operand")(static_cast <bool> (OtherUses[i]->getOperand(!OffsetIdx ).getNode() == BasePtr.getNode() && "Expected BasePtr operand" ) ? void (0) : __assert_fail ("OtherUses[i]->getOperand(!OffsetIdx).getNode() == BasePtr.getNode() && \"Expected BasePtr operand\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 11520, __extension__ __PRETTY_FUNCTION__)); | |||
11521 | ||||
11522 | // We need to replace ptr0 in the following expression: | |||
11523 | // x0 * offset0 + y0 * ptr0 = t0 | |||
11524 | // knowing that | |||
11525 | // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store) | |||
11526 | // | |||
11527 | // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the | |||
11528 | // indexed load/store and the expression that needs to be re-written. | |||
11529 | // | |||
11530 | // Therefore, we have: | |||
11531 | // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1 | |||
11532 | ||||
11533 | ConstantSDNode *CN = | |||
11534 | cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx)); | |||
11535 | int X0, X1, Y0, Y1; | |||
11536 | const APInt &Offset0 = CN->getAPIntValue(); | |||
11537 | APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue(); | |||
11538 | ||||
11539 | X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1; | |||
11540 | Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1; | |||
11541 | X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1; | |||
11542 | Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1; | |||
11543 | ||||
11544 | unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD; | |||
11545 | ||||
11546 | APInt CNV = Offset0; | |||
11547 | if (X0 < 0) CNV = -CNV; | |||
11548 | if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1; | |||
11549 | else CNV = CNV - Offset1; | |||
11550 | ||||
11551 | SDLoc DL(OtherUses[i]); | |||
11552 | ||||
11553 | // We can now generate the new expression. | |||
11554 | SDValue NewOp1 = DAG.getConstant(CNV, DL, CN->getValueType(0)); | |||
11555 | SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0); | |||
11556 | ||||
11557 | SDValue NewUse = DAG.getNode(Opcode, | |||
11558 | DL, | |||
11559 | OtherUses[i]->getValueType(0), NewOp1, NewOp2); | |||
11560 | DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse); | |||
11561 | deleteAndRecombine(OtherUses[i]); | |||
11562 | } | |||
11563 | ||||
11564 | // Replace the uses of Ptr with uses of the updated base value. | |||
11565 | DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0)); | |||
11566 | deleteAndRecombine(Ptr.getNode()); | |||
11567 | AddToWorklist(Result.getNode()); | |||
11568 | ||||
11569 | return true; | |||
11570 | } | |||
11571 | ||||
11572 | /// Try to combine a load/store with a add/sub of the base pointer node into a | |||
11573 | /// post-indexed load/store. The transformation folded the add/subtract into the | |||
11574 | /// new indexed load/store effectively and all of its uses are redirected to the | |||
11575 | /// new load/store. | |||
11576 | bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) { | |||
11577 | if (Level < AfterLegalizeDAG) | |||
11578 | return false; | |||
11579 | ||||
11580 | bool isLoad = true; | |||
11581 | SDValue Ptr; | |||
11582 | EVT VT; | |||
11583 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { | |||
11584 | if (LD->isIndexed()) | |||
11585 | return false; | |||
11586 | VT = LD->getMemoryVT(); | |||
11587 | if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) && | |||
11588 | !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT)) | |||
11589 | return false; | |||
11590 | Ptr = LD->getBasePtr(); | |||
11591 | } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { | |||
11592 | if (ST->isIndexed()) | |||
11593 | return false; | |||
11594 | VT = ST->getMemoryVT(); | |||
11595 | if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) && | |||
11596 | !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT)) | |||
11597 | return false; | |||
11598 | Ptr = ST->getBasePtr(); | |||
11599 | isLoad = false; | |||
11600 | } else { | |||
11601 | return false; | |||
11602 | } | |||
11603 | ||||
11604 | if (Ptr.getNode()->hasOneUse()) | |||
11605 | return false; | |||
11606 | ||||
11607 | for (SDNode *Op : Ptr.getNode()->uses()) { | |||
11608 | if (Op == N || | |||
11609 | (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)) | |||
11610 | continue; | |||
11611 | ||||
11612 | SDValue BasePtr; | |||
11613 | SDValue Offset; | |||
11614 | ISD::MemIndexedMode AM = ISD::UNINDEXED; | |||
11615 | if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) { | |||
11616 | // Don't create a indexed load / store with zero offset. | |||
11617 | if (isNullConstant(Offset)) | |||
11618 | continue; | |||
11619 | ||||
11620 | // Try turning it into a post-indexed load / store except when | |||
11621 | // 1) All uses are load / store ops that use it as base ptr (and | |||
11622 | // it may be folded as addressing mmode). | |||
11623 | // 2) Op must be independent of N, i.e. Op is neither a predecessor | |||
11624 | // nor a successor of N. Otherwise, if Op is folded that would | |||
11625 | // create a cycle. | |||
11626 | ||||
11627 | if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr)) | |||
11628 | continue; | |||
11629 | ||||
11630 | // Check for #1. | |||
11631 | bool TryNext = false; | |||
11632 | for (SDNode *Use : BasePtr.getNode()->uses()) { | |||
11633 | if (Use == Ptr.getNode()) | |||
11634 | continue; | |||
11635 | ||||
11636 | // If all the uses are load / store addresses, then don't do the | |||
11637 | // transformation. | |||
11638 | if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){ | |||
11639 | bool RealUse = false; | |||
11640 | for (SDNode *UseUse : Use->uses()) { | |||
11641 | if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI)) | |||
11642 | RealUse = true; | |||
11643 | } | |||
11644 | ||||
11645 | if (!RealUse) { | |||
11646 | TryNext = true; | |||
11647 | break; | |||
11648 | } | |||
11649 | } | |||
11650 | } | |||
11651 | ||||
11652 | if (TryNext) | |||
11653 | continue; | |||
11654 | ||||
11655 | // Check for #2 | |||
11656 | if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) { | |||
11657 | SDValue Result = isLoad | |||
11658 | ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N), | |||
11659 | BasePtr, Offset, AM) | |||
11660 | : DAG.getIndexedStore(SDValue(N,0), SDLoc(N), | |||
11661 | BasePtr, Offset, AM); | |||
11662 | ++PostIndexedNodes; | |||
11663 | ++NodesCombined; | |||
11664 | DEBUG(dbgs() << "\nReplacing.5 ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.5 "; N->dump (&DAG); dbgs() << "\nWith: "; Result.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false) | |||
11665 | N->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.5 "; N->dump (&DAG); dbgs() << "\nWith: "; Result.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false) | |||
11666 | dbgs() << "\nWith: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.5 "; N->dump (&DAG); dbgs() << "\nWith: "; Result.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false) | |||
11667 | Result.getNode()->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.5 "; N->dump (&DAG); dbgs() << "\nWith: "; Result.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false) | |||
11668 | dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.5 "; N->dump (&DAG); dbgs() << "\nWith: "; Result.getNode()-> dump(&DAG); dbgs() << '\n'; } } while (false); | |||
11669 | WorklistRemover DeadNodes(*this); | |||
11670 | if (isLoad) { | |||
11671 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0)); | |||
11672 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2)); | |||
11673 | } else { | |||
11674 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1)); | |||
11675 | } | |||
11676 | ||||
11677 | // Finally, since the node is now dead, remove it from the graph. | |||
11678 | deleteAndRecombine(N); | |||
11679 | ||||
11680 | // Replace the uses of Use with uses of the updated base value. | |||
11681 | DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0), | |||
11682 | Result.getValue(isLoad ? 1 : 0)); | |||
11683 | deleteAndRecombine(Op); | |||
11684 | return true; | |||
11685 | } | |||
11686 | } | |||
11687 | } | |||
11688 | ||||
11689 | return false; | |||
11690 | } | |||
11691 | ||||
11692 | /// \brief Return the base-pointer arithmetic from an indexed \p LD. | |||
11693 | SDValue DAGCombiner::SplitIndexingFromLoad(LoadSDNode *LD) { | |||
11694 | ISD::MemIndexedMode AM = LD->getAddressingMode(); | |||
11695 | assert(AM != ISD::UNINDEXED)(static_cast <bool> (AM != ISD::UNINDEXED) ? void (0) : __assert_fail ("AM != ISD::UNINDEXED", "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 11695, __extension__ __PRETTY_FUNCTION__)); | |||
11696 | SDValue BP = LD->getOperand(1); | |||
11697 | SDValue Inc = LD->getOperand(2); | |||
11698 | ||||
11699 | // Some backends use TargetConstants for load offsets, but don't expect | |||
11700 | // TargetConstants in general ADD nodes. We can convert these constants into | |||
11701 | // regular Constants (if the constant is not opaque). | |||
11702 | assert((Inc.getOpcode() != ISD::TargetConstant ||(static_cast <bool> ((Inc.getOpcode() != ISD::TargetConstant || !cast<ConstantSDNode>(Inc)->isOpaque()) && "Cannot split out indexing using opaque target constants") ? void (0) : __assert_fail ("(Inc.getOpcode() != ISD::TargetConstant || !cast<ConstantSDNode>(Inc)->isOpaque()) && \"Cannot split out indexing using opaque target constants\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 11704, __extension__ __PRETTY_FUNCTION__)) | |||
11703 | !cast<ConstantSDNode>(Inc)->isOpaque()) &&(static_cast <bool> ((Inc.getOpcode() != ISD::TargetConstant || !cast<ConstantSDNode>(Inc)->isOpaque()) && "Cannot split out indexing using opaque target constants") ? void (0) : __assert_fail ("(Inc.getOpcode() != ISD::TargetConstant || !cast<ConstantSDNode>(Inc)->isOpaque()) && \"Cannot split out indexing using opaque target constants\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 11704, __extension__ __PRETTY_FUNCTION__)) | |||
11704 | "Cannot split out indexing using opaque target constants")(static_cast <bool> ((Inc.getOpcode() != ISD::TargetConstant || !cast<ConstantSDNode>(Inc)->isOpaque()) && "Cannot split out indexing using opaque target constants") ? void (0) : __assert_fail ("(Inc.getOpcode() != ISD::TargetConstant || !cast<ConstantSDNode>(Inc)->isOpaque()) && \"Cannot split out indexing using opaque target constants\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 11704, __extension__ __PRETTY_FUNCTION__)); | |||
11705 | if (Inc.getOpcode() == ISD::TargetConstant) { | |||
11706 | ConstantSDNode *ConstInc = cast<ConstantSDNode>(Inc); | |||
11707 | Inc = DAG.getConstant(*ConstInc->getConstantIntValue(), SDLoc(Inc), | |||
11708 | ConstInc->getValueType(0)); | |||
11709 | } | |||
11710 | ||||
11711 | unsigned Opc = | |||
11712 | (AM == ISD::PRE_INC || AM == ISD::POST_INC ? ISD::ADD : ISD::SUB); | |||
11713 | return DAG.getNode(Opc, SDLoc(LD), BP.getSimpleValueType(), BP, Inc); | |||
11714 | } | |||
11715 | ||||
11716 | SDValue DAGCombiner::visitLOAD(SDNode *N) { | |||
11717 | LoadSDNode *LD = cast<LoadSDNode>(N); | |||
11718 | SDValue Chain = LD->getChain(); | |||
11719 | SDValue Ptr = LD->getBasePtr(); | |||
11720 | ||||
11721 | // If load is not volatile and there are no uses of the loaded value (and | |||
11722 | // the updated indexed value in case of indexed loads), change uses of the | |||
11723 | // chain value into uses of the chain input (i.e. delete the dead load). | |||
11724 | if (!LD->isVolatile()) { | |||
11725 | if (N->getValueType(1) == MVT::Other) { | |||
11726 | // Unindexed loads. | |||
11727 | if (!N->hasAnyUseOfValue(0)) { | |||
11728 | // It's not safe to use the two value CombineTo variant here. e.g. | |||
11729 | // v1, chain2 = load chain1, loc | |||
11730 | // v2, chain3 = load chain2, loc | |||
11731 | // v3 = add v2, c | |||
11732 | // Now we replace use of chain2 with chain1. This makes the second load | |||
11733 | // isomorphic to the one we are deleting, and thus makes this load live. | |||
11734 | DEBUG(dbgs() << "\nReplacing.6 ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.6 "; N->dump (&DAG); dbgs() << "\nWith chain: "; Chain.getNode() ->dump(&DAG); dbgs() << "\n"; } } while (false) | |||
11735 | N->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.6 "; N->dump (&DAG); dbgs() << "\nWith chain: "; Chain.getNode() ->dump(&DAG); dbgs() << "\n"; } } while (false) | |||
11736 | dbgs() << "\nWith chain: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.6 "; N->dump (&DAG); dbgs() << "\nWith chain: "; Chain.getNode() ->dump(&DAG); dbgs() << "\n"; } } while (false) | |||
11737 | Chain.getNode()->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.6 "; N->dump (&DAG); dbgs() << "\nWith chain: "; Chain.getNode() ->dump(&DAG); dbgs() << "\n"; } } while (false) | |||
11738 | dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.6 "; N->dump (&DAG); dbgs() << "\nWith chain: "; Chain.getNode() ->dump(&DAG); dbgs() << "\n"; } } while (false); | |||
11739 | WorklistRemover DeadNodes(*this); | |||
11740 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain); | |||
11741 | AddUsersToWorklist(Chain.getNode()); | |||
11742 | if (N->use_empty()) | |||
11743 | deleteAndRecombine(N); | |||
11744 | ||||
11745 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
11746 | } | |||
11747 | } else { | |||
11748 | // Indexed loads. | |||
11749 | assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?")(static_cast <bool> (N->getValueType(2) == MVT::Other && "Malformed indexed loads?") ? void (0) : __assert_fail ("N->getValueType(2) == MVT::Other && \"Malformed indexed loads?\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 11749, __extension__ __PRETTY_FUNCTION__)); | |||
11750 | ||||
11751 | // If this load has an opaque TargetConstant offset, then we cannot split | |||
11752 | // the indexing into an add/sub directly (that TargetConstant may not be | |||
11753 | // valid for a different type of node, and we cannot convert an opaque | |||
11754 | // target constant into a regular constant). | |||
11755 | bool HasOTCInc = LD->getOperand(2).getOpcode() == ISD::TargetConstant && | |||
11756 | cast<ConstantSDNode>(LD->getOperand(2))->isOpaque(); | |||
11757 | ||||
11758 | if (!N->hasAnyUseOfValue(0) && | |||
11759 | ((MaySplitLoadIndex && !HasOTCInc) || !N->hasAnyUseOfValue(1))) { | |||
11760 | SDValue Undef = DAG.getUNDEF(N->getValueType(0)); | |||
11761 | SDValue Index; | |||
11762 | if (N->hasAnyUseOfValue(1) && MaySplitLoadIndex && !HasOTCInc) { | |||
11763 | Index = SplitIndexingFromLoad(LD); | |||
11764 | // Try to fold the base pointer arithmetic into subsequent loads and | |||
11765 | // stores. | |||
11766 | AddUsersToWorklist(N); | |||
11767 | } else | |||
11768 | Index = DAG.getUNDEF(N->getValueType(1)); | |||
11769 | DEBUG(dbgs() << "\nReplacing.7 ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.7 "; N->dump (&DAG); dbgs() << "\nWith: "; Undef.getNode()->dump (&DAG); dbgs() << " and 2 other values\n"; } } while (false) | |||
11770 | N->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.7 "; N->dump (&DAG); dbgs() << "\nWith: "; Undef.getNode()->dump (&DAG); dbgs() << " and 2 other values\n"; } } while (false) | |||
11771 | dbgs() << "\nWith: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.7 "; N->dump (&DAG); dbgs() << "\nWith: "; Undef.getNode()->dump (&DAG); dbgs() << " and 2 other values\n"; } } while (false) | |||
11772 | Undef.getNode()->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.7 "; N->dump (&DAG); dbgs() << "\nWith: "; Undef.getNode()->dump (&DAG); dbgs() << " and 2 other values\n"; } } while (false) | |||
11773 | dbgs() << " and 2 other values\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dagcombine")) { dbgs() << "\nReplacing.7 "; N->dump (&DAG); dbgs() << "\nWith: "; Undef.getNode()->dump (&DAG); dbgs() << " and 2 other values\n"; } } while (false); | |||
11774 | WorklistRemover DeadNodes(*this); | |||
11775 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef); | |||
11776 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Index); | |||
11777 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain); | |||
11778 | deleteAndRecombine(N); | |||
11779 | return SDValue(N, 0); // Return N so it doesn't get rechecked! | |||
11780 | } | |||
11781 | } | |||
11782 | } | |||
11783 | ||||
11784 | // If this load is directly stored, replace the load value with the stored | |||
11785 | // value. | |||
11786 | // TODO: Handle store large -> read small portion. | |||
11787 | // TODO: Handle TRUNCSTORE/LOADEXT | |||
11788 | if (OptLevel != CodeGenOpt::None && | |||
11789 | ISD::isNormalLoad(N) && !LD->isVolatile()) { | |||
11790 | if (ISD::isNON_TRUNCStore(Chain.getNode())) { | |||
11791 | StoreSDNode *PrevST = cast<StoreSDNode>(Chain); | |||
11792 | if (PrevST->getBasePtr() == Ptr && | |||
11793 | PrevST->getValue().getValueType() == N->getValueType(0)) | |||
11794 | return CombineTo(N, PrevST->getOperand(1), Chain); | |||
11795 | } | |||
11796 | } | |||
11797 | ||||
11798 | // Try to infer better alignment information than the load already has. | |||
11799 | if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) { | |||
11800 | if (unsigned Align = DAG.InferPtrAlignment(Ptr)) { | |||
11801 | if (Align > LD->getMemOperand()->getBaseAlignment()) { | |||
11802 | SDValue NewLoad = DAG.getExtLoad( | |||
11803 | LD->getExtensionType(), SDLoc(N), LD->getValueType(0), Chain, Ptr, | |||
11804 | LD->getPointerInfo(), LD->getMemoryVT(), Align, | |||
11805 | LD->getMemOperand()->getFlags(), LD->getAAInfo()); | |||
11806 | if (NewLoad.getNode() != N) | |||
11807 | return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true); | |||
11808 | } | |||
11809 | } | |||
11810 | } | |||
11811 | ||||
11812 | if (LD->isUnindexed()) { | |||
11813 | // Walk up chain skipping non-aliasing memory nodes. | |||
11814 | SDValue BetterChain = FindBetterChain(N, Chain); | |||
11815 | ||||
11816 | // If there is a better chain. | |||
11817 | if (Chain != BetterChain) { | |||
11818 | SDValue ReplLoad; | |||
11819 | ||||
11820 | // Replace the chain to void dependency. | |||
11821 | if (LD->getExtensionType() == ISD::NON_EXTLOAD) { | |||
11822 | ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD), | |||
11823 | BetterChain, Ptr, LD->getMemOperand()); | |||
11824 | } else { | |||
11825 | ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), | |||
11826 | LD->getValueType(0), | |||
11827 | BetterChain, Ptr, LD->getMemoryVT(), | |||
11828 | LD->getMemOperand()); | |||
11829 | } | |||
11830 | ||||
11831 | // Create token factor to keep old chain connected. | |||
11832 | SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N), | |||
11833 | MVT::Other, Chain, ReplLoad.getValue(1)); | |||
11834 | ||||
11835 | // Replace uses with load result and token factor | |||
11836 | return CombineTo(N, ReplLoad.getValue(0), Token); | |||
11837 | } | |||
11838 | } | |||
11839 | ||||
11840 | // Try transforming N to an indexed load. | |||
11841 | if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N)) | |||
11842 | return SDValue(N, 0); | |||
11843 | ||||
11844 | // Try to slice up N to more direct loads if the slices are mapped to | |||
11845 | // different register banks or pairing can take place. | |||
11846 | if (SliceUpLoad(N)) | |||
11847 | return SDValue(N, 0); | |||
11848 | ||||
11849 | return SDValue(); | |||
11850 | } | |||
11851 | ||||
11852 | namespace { | |||
11853 | ||||
11854 | /// \brief Helper structure used to slice a load in smaller loads. | |||
11855 | /// Basically a slice is obtained from the following sequence: | |||
11856 | /// Origin = load Ty1, Base | |||
11857 | /// Shift = srl Ty1 Origin, CstTy Amount | |||
11858 | /// Inst = trunc Shift to Ty2 | |||
11859 | /// | |||
11860 | /// Then, it will be rewritten into: | |||
11861 | /// Slice = load SliceTy, Base + SliceOffset | |||
11862 | /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2 | |||
11863 | /// | |||
11864 | /// SliceTy is deduced from the number of bits that are actually used to | |||
11865 | /// build Inst. | |||
11866 | struct LoadedSlice { | |||
11867 | /// \brief Helper structure used to compute the cost of a slice. | |||
11868 | struct Cost { | |||
11869 | /// Are we optimizing for code size. | |||
11870 | bool ForCodeSize; | |||
11871 | ||||
11872 | /// Various cost. | |||
11873 | unsigned Loads = 0; | |||
11874 | unsigned Truncates = 0; | |||
11875 | unsigned CrossRegisterBanksCopies = 0; | |||
11876 | unsigned ZExts = 0; | |||
11877 | unsigned Shift = 0; | |||
11878 | ||||
11879 | Cost(bool ForCodeSize = false) : ForCodeSize(ForCodeSize) {} | |||
11880 | ||||
11881 | /// \brief Get the cost of one isolated slice. | |||
11882 | Cost(const LoadedSlice &LS, bool ForCodeSize = false) | |||
11883 | : ForCodeSize(ForCodeSize), Loads(1) { | |||
11884 | EVT TruncType = LS.Inst->getValueType(0); | |||
11885 | EVT LoadedType = LS.getLoadedType(); | |||
11886 | if (TruncType != LoadedType && | |||
11887 | !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType)) | |||
11888 | ZExts = 1; | |||
11889 | } | |||
11890 | ||||
11891 | /// \brief Account for slicing gain in the current cost. | |||
11892 | /// Slicing provide a few gains like removing a shift or a | |||
11893 | /// truncate. This method allows to grow the cost of the original | |||
11894 | /// load with the gain from this slice. | |||
11895 | void addSliceGain(const LoadedSlice &LS) { | |||
11896 | // Each slice saves a truncate. | |||
11897 | const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo(); | |||
11898 | if (!TLI.isTruncateFree(LS.Inst->getOperand(0).getValueType(), | |||
11899 | LS.Inst->getValueType(0))) | |||
11900 | ++Truncates; | |||
11901 | // If there is a shift amount, this slice gets rid of it. | |||
11902 | if (LS.Shift) | |||
11903 | ++Shift; | |||
11904 | // If this slice can merge a cross register bank copy, account for it. | |||
11905 | if (LS.canMergeExpensiveCrossRegisterBankCopy()) | |||
11906 | ++CrossRegisterBanksCopies; | |||
11907 | } | |||
11908 | ||||
11909 | Cost &operator+=(const Cost &RHS) { | |||
11910 | Loads += RHS.Loads; | |||
11911 | Truncates += RHS.Truncates; | |||
11912 | CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies; | |||
11913 | ZExts += RHS.ZExts; | |||
11914 | Shift += RHS.Shift; | |||
11915 | return *this; | |||
11916 | } | |||
11917 | ||||
11918 | bool operator==(const Cost &RHS) const { | |||
11919 | return Loads == RHS.Loads && Truncates == RHS.Truncates && | |||
11920 | CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies && | |||
11921 | ZExts == RHS.ZExts && Shift == RHS.Shift; | |||
11922 | } | |||
11923 | ||||
11924 | bool operator!=(const Cost &RHS) const { return !(*this == RHS); } | |||
11925 | ||||
11926 | bool operator<(const Cost &RHS) const { | |||
11927 | // Assume cross register banks copies are as expensive as loads. | |||
11928 | // FIXME: Do we want some more target hooks? | |||
11929 | unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies; | |||
11930 | unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies; | |||
11931 | // Unless we are optimizing for code size, consider the | |||
11932 | // expensive operation first. | |||
11933 | if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS) | |||
11934 | return ExpensiveOpsLHS < ExpensiveOpsRHS; | |||
11935 | return (Truncates + ZExts + Shift + ExpensiveOpsLHS) < | |||
11936 | (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS); | |||
11937 | } | |||
11938 | ||||
11939 | bool operator>(const Cost &RHS) const { return RHS < *this; } | |||
11940 | ||||
11941 | bool operator<=(const Cost &RHS) const { return !(RHS < *this); } | |||
11942 | ||||
11943 | bool operator>=(const Cost &RHS) const { return !(*this < RHS); } | |||
11944 | }; | |||
11945 | ||||
11946 | // The last instruction that represent the slice. This should be a | |||
11947 | // truncate instruction. | |||
11948 | SDNode *Inst; | |||
11949 | ||||
11950 | // The original load instruction. | |||
11951 | LoadSDNode *Origin; | |||
11952 | ||||
11953 | // The right shift amount in bits from the original load. | |||
11954 | unsigned Shift; | |||
11955 | ||||
11956 | // The DAG from which Origin came from. | |||
11957 | // This is used to get some contextual information about legal types, etc. | |||
11958 | SelectionDAG *DAG; | |||
11959 | ||||
11960 | LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr, | |||
11961 | unsigned Shift = 0, SelectionDAG *DAG = nullptr) | |||
11962 | : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {} | |||
11963 | ||||
11964 | /// \brief Get the bits used in a chunk of bits \p BitWidth large. | |||
11965 | /// \return Result is \p BitWidth and has used bits set to 1 and | |||
11966 | /// not used bits set to 0. | |||
11967 | APInt getUsedBits() const { | |||
11968 | // Reproduce the trunc(lshr) sequence: | |||
11969 | // - Start from the truncated value. | |||
11970 | // - Zero extend to the desired bit width. | |||
11971 | // - Shift left. | |||
11972 | assert(Origin && "No original load to compare against.")(static_cast <bool> (Origin && "No original load to compare against." ) ? void (0) : __assert_fail ("Origin && \"No original load to compare against.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 11972, __extension__ __PRETTY_FUNCTION__)); | |||
11973 | unsigned BitWidth = Origin->getValueSizeInBits(0); | |||
11974 | assert(Inst && "This slice is not bound to an instruction")(static_cast <bool> (Inst && "This slice is not bound to an instruction" ) ? void (0) : __assert_fail ("Inst && \"This slice is not bound to an instruction\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 11974, __extension__ __PRETTY_FUNCTION__)); | |||
11975 | assert(Inst->getValueSizeInBits(0) <= BitWidth &&(static_cast <bool> (Inst->getValueSizeInBits(0) <= BitWidth && "Extracted slice is bigger than the whole type!" ) ? void (0) : __assert_fail ("Inst->getValueSizeInBits(0) <= BitWidth && \"Extracted slice is bigger than the whole type!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 11976, __extension__ __PRETTY_FUNCTION__)) | |||
11976 | "Extracted slice is bigger than the whole type!")(static_cast <bool> (Inst->getValueSizeInBits(0) <= BitWidth && "Extracted slice is bigger than the whole type!" ) ? void (0) : __assert_fail ("Inst->getValueSizeInBits(0) <= BitWidth && \"Extracted slice is bigger than the whole type!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 11976, __extension__ __PRETTY_FUNCTION__)); | |||
11977 | APInt UsedBits(Inst->getValueSizeInBits(0), 0); | |||
11978 | UsedBits.setAllBits(); | |||
11979 | UsedBits = UsedBits.zext(BitWidth); | |||
11980 | UsedBits <<= Shift; | |||
11981 | return UsedBits; | |||
11982 | } | |||
11983 | ||||
11984 | /// \brief Get the size of the slice to be loaded in bytes. | |||
11985 | unsigned getLoadedSize() const { | |||
11986 | unsigned SliceSize = getUsedBits().countPopulation(); | |||
11987 | assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.")(static_cast <bool> (!(SliceSize & 0x7) && "Size is not a multiple of a byte." ) ? void (0) : __assert_fail ("!(SliceSize & 0x7) && \"Size is not a multiple of a byte.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 11987, __extension__ __PRETTY_FUNCTION__)); | |||
11988 | return SliceSize / 8; | |||
11989 | } | |||
11990 | ||||
11991 | /// \brief Get the type that will be loaded for this slice. | |||
11992 | /// Note: This may not be the final type for the slice. | |||
11993 | EVT getLoadedType() const { | |||
11994 | assert(DAG && "Missing context")(static_cast <bool> (DAG && "Missing context") ? void (0) : __assert_fail ("DAG && \"Missing context\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 11994, __extension__ __PRETTY_FUNCTION__)); | |||
11995 | LLVMContext &Ctxt = *DAG->getContext(); | |||
11996 | return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8); | |||
11997 | } | |||
11998 | ||||
11999 | /// \brief Get the alignment of the load used for this slice. | |||
12000 | unsigned getAlignment() const { | |||
12001 | unsigned Alignment = Origin->getAlignment(); | |||
12002 | unsigned Offset = getOffsetFromBase(); | |||
12003 | if (Offset != 0) | |||
12004 | Alignment = MinAlign(Alignment, Alignment + Offset); | |||
12005 | return Alignment; | |||
12006 | } | |||
12007 | ||||
12008 | /// \brief Check if this slice can be rewritten with legal operations. | |||
12009 | bool isLegal() const { | |||
12010 | // An invalid slice is not legal. | |||
12011 | if (!Origin || !Inst || !DAG) | |||
12012 | return false; | |||
12013 | ||||
12014 | // Offsets are for indexed load only, we do not handle that. | |||
12015 | if (!Origin->getOffset().isUndef()) | |||
12016 | return false; | |||
12017 | ||||
12018 | const TargetLowering &TLI = DAG->getTargetLoweringInfo(); | |||
12019 | ||||
12020 | // Check that the type is legal. | |||
12021 | EVT SliceType = getLoadedType(); | |||
12022 | if (!TLI.isTypeLegal(SliceType)) | |||
12023 | return false; | |||
12024 | ||||
12025 | // Check that the load is legal for this type. | |||
12026 | if (!TLI.isOperationLegal(ISD::LOAD, SliceType)) | |||
12027 | return false; | |||
12028 | ||||
12029 | // Check that the offset can be computed. | |||
12030 | // 1. Check its type. | |||
12031 | EVT PtrType = Origin->getBasePtr().getValueType(); | |||
12032 | if (PtrType == MVT::Untyped || PtrType.isExtended()) | |||
12033 | return false; | |||
12034 | ||||
12035 | // 2. Check that it fits in the immediate. | |||
12036 | if (!TLI.isLegalAddImmediate(getOffsetFromBase())) | |||
12037 | return false; | |||
12038 | ||||
12039 | // 3. Check that the computation is legal. | |||
12040 | if (!TLI.isOperationLegal(ISD::ADD, PtrType)) | |||
12041 | return false; | |||
12042 | ||||
12043 | // Check that the zext is legal if it needs one. | |||
12044 | EVT TruncateType = Inst->getValueType(0); | |||
12045 | if (TruncateType != SliceType && | |||
12046 | !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType)) | |||
12047 | return false; | |||
12048 | ||||
12049 | return true; | |||
12050 | } | |||
12051 | ||||
12052 | /// \brief Get the offset in bytes of this slice in the original chunk of | |||
12053 | /// bits. | |||
12054 | /// \pre DAG != nullptr. | |||
12055 | uint64_t getOffsetFromBase() const { | |||
12056 | assert(DAG && "Missing context.")(static_cast <bool> (DAG && "Missing context.") ? void (0) : __assert_fail ("DAG && \"Missing context.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12056, __extension__ __PRETTY_FUNCTION__)); | |||
12057 | bool IsBigEndian = DAG->getDataLayout().isBigEndian(); | |||
12058 | assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.")(static_cast <bool> (!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported." ) ? void (0) : __assert_fail ("!(Shift & 0x7) && \"Shifts not aligned on Bytes are not supported.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12058, __extension__ __PRETTY_FUNCTION__)); | |||
12059 | uint64_t Offset = Shift / 8; | |||
12060 | unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8; | |||
12061 | assert(!(Origin->getValueSizeInBits(0) & 0x7) &&(static_cast <bool> (!(Origin->getValueSizeInBits(0) & 0x7) && "The size of the original loaded type is not a multiple of a" " byte.") ? void (0) : __assert_fail ("!(Origin->getValueSizeInBits(0) & 0x7) && \"The size of the original loaded type is not a multiple of a\" \" byte.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12063, __extension__ __PRETTY_FUNCTION__)) | |||
12062 | "The size of the original loaded type is not a multiple of a"(static_cast <bool> (!(Origin->getValueSizeInBits(0) & 0x7) && "The size of the original loaded type is not a multiple of a" " byte.") ? void (0) : __assert_fail ("!(Origin->getValueSizeInBits(0) & 0x7) && \"The size of the original loaded type is not a multiple of a\" \" byte.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12063, __extension__ __PRETTY_FUNCTION__)) | |||
12063 | " byte.")(static_cast <bool> (!(Origin->getValueSizeInBits(0) & 0x7) && "The size of the original loaded type is not a multiple of a" " byte.") ? void (0) : __assert_fail ("!(Origin->getValueSizeInBits(0) & 0x7) && \"The size of the original loaded type is not a multiple of a\" \" byte.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12063, __extension__ __PRETTY_FUNCTION__)); | |||
12064 | // If Offset is bigger than TySizeInBytes, it means we are loading all | |||
12065 | // zeros. This should have been optimized before in the process. | |||
12066 | assert(TySizeInBytes > Offset &&(static_cast <bool> (TySizeInBytes > Offset && "Invalid shift amount for given loaded size") ? void (0) : __assert_fail ("TySizeInBytes > Offset && \"Invalid shift amount for given loaded size\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12067, __extension__ __PRETTY_FUNCTION__)) | |||
12067 | "Invalid shift amount for given loaded size")(static_cast <bool> (TySizeInBytes > Offset && "Invalid shift amount for given loaded size") ? void (0) : __assert_fail ("TySizeInBytes > Offset && \"Invalid shift amount for given loaded size\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12067, __extension__ __PRETTY_FUNCTION__)); | |||
12068 | if (IsBigEndian) | |||
12069 | Offset = TySizeInBytes - Offset - getLoadedSize(); | |||
12070 | return Offset; | |||
12071 | } | |||
12072 | ||||
12073 | /// \brief Generate the sequence of instructions to load the slice | |||
12074 | /// represented by this object and redirect the uses of this slice to | |||
12075 | /// this new sequence of instructions. | |||
12076 | /// \pre this->Inst && this->Origin are valid Instructions and this | |||
12077 | /// object passed the legal check: LoadedSlice::isLegal returned true. | |||
12078 | /// \return The last instruction of the sequence used to load the slice. | |||
12079 | SDValue loadSlice() const { | |||
12080 | assert(Inst && Origin && "Unable to replace a non-existing slice.")(static_cast <bool> (Inst && Origin && "Unable to replace a non-existing slice." ) ? void (0) : __assert_fail ("Inst && Origin && \"Unable to replace a non-existing slice.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12080, __extension__ __PRETTY_FUNCTION__)); | |||
12081 | const SDValue &OldBaseAddr = Origin->getBasePtr(); | |||
12082 | SDValue BaseAddr = OldBaseAddr; | |||
12083 | // Get the offset in that chunk of bytes w.r.t. the endianness. | |||
12084 | int64_t Offset = static_cast<int64_t>(getOffsetFromBase()); | |||
12085 | assert(Offset >= 0 && "Offset too big to fit in int64_t!")(static_cast <bool> (Offset >= 0 && "Offset too big to fit in int64_t!" ) ? void (0) : __assert_fail ("Offset >= 0 && \"Offset too big to fit in int64_t!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12085, __extension__ __PRETTY_FUNCTION__)); | |||
12086 | if (Offset) { | |||
12087 | // BaseAddr = BaseAddr + Offset. | |||
12088 | EVT ArithType = BaseAddr.getValueType(); | |||
12089 | SDLoc DL(Origin); | |||
12090 | BaseAddr = DAG->getNode(ISD::ADD, DL, ArithType, BaseAddr, | |||
12091 | DAG->getConstant(Offset, DL, ArithType)); | |||
12092 | } | |||
12093 | ||||
12094 | // Create the type of the loaded slice according to its size. | |||
12095 | EVT SliceType = getLoadedType(); | |||
12096 | ||||
12097 | // Create the load for the slice. | |||
12098 | SDValue LastInst = | |||
12099 | DAG->getLoad(SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr, | |||
12100 | Origin->getPointerInfo().getWithOffset(Offset), | |||
12101 | getAlignment(), Origin->getMemOperand()->getFlags()); | |||
12102 | // If the final type is not the same as the loaded type, this means that | |||
12103 | // we have to pad with zero. Create a zero extend for that. | |||
12104 | EVT FinalType = Inst->getValueType(0); | |||
12105 | if (SliceType != FinalType) | |||
12106 | LastInst = | |||
12107 | DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst); | |||
12108 | return LastInst; | |||
12109 | } | |||
12110 | ||||
12111 | /// \brief Check if this slice can be merged with an expensive cross register | |||
12112 | /// bank copy. E.g., | |||
12113 | /// i = load i32 | |||
12114 | /// f = bitcast i32 i to float | |||
12115 | bool canMergeExpensiveCrossRegisterBankCopy() const { | |||
12116 | if (!Inst || !Inst->hasOneUse()) | |||
12117 | return false; | |||
12118 | SDNode *Use = *Inst->use_begin(); | |||
12119 | if (Use->getOpcode() != ISD::BITCAST) | |||
12120 | return false; | |||
12121 | assert(DAG && "Missing context")(static_cast <bool> (DAG && "Missing context") ? void (0) : __assert_fail ("DAG && \"Missing context\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12121, __extension__ __PRETTY_FUNCTION__)); | |||
12122 | const TargetLowering &TLI = DAG->getTargetLoweringInfo(); | |||
12123 | EVT ResVT = Use->getValueType(0); | |||
12124 | const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT()); | |||
12125 | const TargetRegisterClass *ArgRC = | |||
12126 | TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT()); | |||
12127 | if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT)) | |||
12128 | return false; | |||
12129 | ||||
12130 | // At this point, we know that we perform a cross-register-bank copy. | |||
12131 | // Check if it is expensive. | |||
12132 | const TargetRegisterInfo *TRI = DAG->getSubtarget().getRegisterInfo(); | |||
12133 | // Assume bitcasts are cheap, unless both register classes do not | |||
12134 | // explicitly share a common sub class. | |||
12135 | if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC)) | |||
12136 | return false; | |||
12137 | ||||
12138 | // Check if it will be merged with the load. | |||
12139 | // 1. Check the alignment constraint. | |||
12140 | unsigned RequiredAlignment = DAG->getDataLayout().getABITypeAlignment( | |||
12141 | ResVT.getTypeForEVT(*DAG->getContext())); | |||
12142 | ||||
12143 | if (RequiredAlignment > getAlignment()) | |||
12144 | return false; | |||
12145 | ||||
12146 | // 2. Check that the load is a legal operation for that type. | |||
12147 | if (!TLI.isOperationLegal(ISD::LOAD, ResVT)) | |||
12148 | return false; | |||
12149 | ||||
12150 | // 3. Check that we do not have a zext in the way. | |||
12151 | if (Inst->getValueType(0) != getLoadedType()) | |||
12152 | return false; | |||
12153 | ||||
12154 | return true; | |||
12155 | } | |||
12156 | }; | |||
12157 | ||||
12158 | } // end anonymous namespace | |||
12159 | ||||
12160 | /// \brief Check that all bits set in \p UsedBits form a dense region, i.e., | |||
12161 | /// \p UsedBits looks like 0..0 1..1 0..0. | |||
12162 | static bool areUsedBitsDense(const APInt &UsedBits) { | |||
12163 | // If all the bits are one, this is dense! | |||
12164 | if (UsedBits.isAllOnesValue()) | |||
12165 | return true; | |||
12166 | ||||
12167 | // Get rid of the unused bits on the right. | |||
12168 | APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros()); | |||
12169 | // Get rid of the unused bits on the left. | |||
12170 | if (NarrowedUsedBits.countLeadingZeros()) | |||
12171 | NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits()); | |||
12172 | // Check that the chunk of bits is completely used. | |||
12173 | return NarrowedUsedBits.isAllOnesValue(); | |||
12174 | } | |||
12175 | ||||
12176 | /// \brief Check whether or not \p First and \p Second are next to each other | |||
12177 | /// in memory. This means that there is no hole between the bits loaded | |||
12178 | /// by \p First and the bits loaded by \p Second. | |||
12179 | static bool areSlicesNextToEachOther(const LoadedSlice &First, | |||
12180 | const LoadedSlice &Second) { | |||
12181 | assert(First.Origin == Second.Origin && First.Origin &&(static_cast <bool> (First.Origin == Second.Origin && First.Origin && "Unable to match different memory origins." ) ? void (0) : __assert_fail ("First.Origin == Second.Origin && First.Origin && \"Unable to match different memory origins.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12182, __extension__ __PRETTY_FUNCTION__)) | |||
12182 | "Unable to match different memory origins.")(static_cast <bool> (First.Origin == Second.Origin && First.Origin && "Unable to match different memory origins." ) ? void (0) : __assert_fail ("First.Origin == Second.Origin && First.Origin && \"Unable to match different memory origins.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12182, __extension__ __PRETTY_FUNCTION__)); | |||
12183 | APInt UsedBits = First.getUsedBits(); | |||
12184 | assert((UsedBits & Second.getUsedBits()) == 0 &&(static_cast <bool> ((UsedBits & Second.getUsedBits ()) == 0 && "Slices are not supposed to overlap.") ? void (0) : __assert_fail ("(UsedBits & Second.getUsedBits()) == 0 && \"Slices are not supposed to overlap.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12185, __extension__ __PRETTY_FUNCTION__)) | |||
12185 | "Slices are not supposed to overlap.")(static_cast <bool> ((UsedBits & Second.getUsedBits ()) == 0 && "Slices are not supposed to overlap.") ? void (0) : __assert_fail ("(UsedBits & Second.getUsedBits()) == 0 && \"Slices are not supposed to overlap.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12185, __extension__ __PRETTY_FUNCTION__)); | |||
12186 | UsedBits |= Second.getUsedBits(); | |||
12187 | return areUsedBitsDense(UsedBits); | |||
12188 | } | |||
12189 | ||||
12190 | /// \brief Adjust the \p GlobalLSCost according to the target | |||
12191 | /// paring capabilities and the layout of the slices. | |||
12192 | /// \pre \p GlobalLSCost should account for at least as many loads as | |||
12193 | /// there is in the slices in \p LoadedSlices. | |||
12194 | static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices, | |||
12195 | LoadedSlice::Cost &GlobalLSCost) { | |||
12196 | unsigned NumberOfSlices = LoadedSlices.size(); | |||
12197 | // If there is less than 2 elements, no pairing is possible. | |||
12198 | if (NumberOfSlices < 2) | |||
12199 | return; | |||
12200 | ||||
12201 | // Sort the slices so that elements that are likely to be next to each | |||
12202 | // other in memory are next to each other in the list. | |||
12203 | std::sort(LoadedSlices.begin(), LoadedSlices.end(), | |||
12204 | [](const LoadedSlice &LHS, const LoadedSlice &RHS) { | |||
12205 | assert(LHS.Origin == RHS.Origin && "Different bases not implemented.")(static_cast <bool> (LHS.Origin == RHS.Origin && "Different bases not implemented.") ? void (0) : __assert_fail ("LHS.Origin == RHS.Origin && \"Different bases not implemented.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12205, __extension__ __PRETTY_FUNCTION__)); | |||
12206 | return LHS.getOffsetFromBase() < RHS.getOffsetFromBase(); | |||
12207 | }); | |||
12208 | const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo(); | |||
12209 | // First (resp. Second) is the first (resp. Second) potentially candidate | |||
12210 | // to be placed in a paired load. | |||
12211 | const LoadedSlice *First = nullptr; | |||
12212 | const LoadedSlice *Second = nullptr; | |||
12213 | for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice, | |||
12214 | // Set the beginning of the pair. | |||
12215 | First = Second) { | |||
12216 | Second = &LoadedSlices[CurrSlice]; | |||
12217 | ||||
12218 | // If First is NULL, it means we start a new pair. | |||
12219 | // Get to the next slice. | |||
12220 | if (!First) | |||
12221 | continue; | |||
12222 | ||||
12223 | EVT LoadedType = First->getLoadedType(); | |||
12224 | ||||
12225 | // If the types of the slices are different, we cannot pair them. | |||
12226 | if (LoadedType != Second->getLoadedType()) | |||
12227 | continue; | |||
12228 | ||||
12229 | // Check if the target supplies paired loads for this type. | |||
12230 | unsigned RequiredAlignment = 0; | |||
12231 | if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) { | |||
12232 | // move to the next pair, this type is hopeless. | |||
12233 | Second = nullptr; | |||
12234 | continue; | |||
12235 | } | |||
12236 | // Check if we meet the alignment requirement. | |||
12237 | if (RequiredAlignment > First->getAlignment()) | |||
12238 | continue; | |||
12239 | ||||
12240 | // Check that both loads are next to each other in memory. | |||
12241 | if (!areSlicesNextToEachOther(*First, *Second)) | |||
12242 | continue; | |||
12243 | ||||
12244 | assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!")(static_cast <bool> (GlobalLSCost.Loads > 0 && "We save more loads than we created!") ? void (0) : __assert_fail ("GlobalLSCost.Loads > 0 && \"We save more loads than we created!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12244, __extension__ __PRETTY_FUNCTION__)); | |||
12245 | --GlobalLSCost.Loads; | |||
12246 | // Move to the next pair. | |||
12247 | Second = nullptr; | |||
12248 | } | |||
12249 | } | |||
12250 | ||||
12251 | /// \brief Check the profitability of all involved LoadedSlice. | |||
12252 | /// Currently, it is considered profitable if there is exactly two | |||
12253 | /// involved slices (1) which are (2) next to each other in memory, and | |||
12254 | /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3). | |||
12255 | /// | |||
12256 | /// Note: The order of the elements in \p LoadedSlices may be modified, but not | |||
12257 | /// the elements themselves. | |||
12258 | /// | |||
12259 | /// FIXME: When the cost model will be mature enough, we can relax | |||
12260 | /// constraints (1) and (2). | |||
12261 | static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices, | |||
12262 | const APInt &UsedBits, bool ForCodeSize) { | |||
12263 | unsigned NumberOfSlices = LoadedSlices.size(); | |||
12264 | if (StressLoadSlicing) | |||
12265 | return NumberOfSlices > 1; | |||
12266 | ||||
12267 | // Check (1). | |||
12268 | if (NumberOfSlices != 2) | |||
12269 | return false; | |||
12270 | ||||
12271 | // Check (2). | |||
12272 | if (!areUsedBitsDense(UsedBits)) | |||
12273 | return false; | |||
12274 | ||||
12275 | // Check (3). | |||
12276 | LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize); | |||
12277 | // The original code has one big load. | |||
12278 | OrigCost.Loads = 1; | |||
12279 | for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) { | |||
12280 | const LoadedSlice &LS = LoadedSlices[CurrSlice]; | |||
12281 | // Accumulate the cost of all the slices. | |||
12282 | LoadedSlice::Cost SliceCost(LS, ForCodeSize); | |||
12283 | GlobalSlicingCost += SliceCost; | |||
12284 | ||||
12285 | // Account as cost in the original configuration the gain obtained | |||
12286 | // with the current slices. | |||
12287 | OrigCost.addSliceGain(LS); | |||
12288 | } | |||
12289 | ||||
12290 | // If the target supports paired load, adjust the cost accordingly. | |||
12291 | adjustCostForPairing(LoadedSlices, GlobalSlicingCost); | |||
12292 | return OrigCost > GlobalSlicingCost; | |||
12293 | } | |||
12294 | ||||
12295 | /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr) | |||
12296 | /// operations, split it in the various pieces being extracted. | |||
12297 | /// | |||
12298 | /// This sort of thing is introduced by SROA. | |||
12299 | /// This slicing takes care not to insert overlapping loads. | |||
12300 | /// \pre LI is a simple load (i.e., not an atomic or volatile load). | |||
12301 | bool DAGCombiner::SliceUpLoad(SDNode *N) { | |||
12302 | if (Level < AfterLegalizeDAG) | |||
12303 | return false; | |||
12304 | ||||
12305 | LoadSDNode *LD = cast<LoadSDNode>(N); | |||
12306 | if (LD->isVolatile() || !ISD::isNormalLoad(LD) || | |||
12307 | !LD->getValueType(0).isInteger()) | |||
12308 | return false; | |||
12309 | ||||
12310 | // Keep track of already used bits to detect overlapping values. | |||
12311 | // In that case, we will just abort the transformation. | |||
12312 | APInt UsedBits(LD->getValueSizeInBits(0), 0); | |||
12313 | ||||
12314 | SmallVector<LoadedSlice, 4> LoadedSlices; | |||
12315 | ||||
12316 | // Check if this load is used as several smaller chunks of bits. | |||
12317 | // Basically, look for uses in trunc or trunc(lshr) and record a new chain | |||
12318 | // of computation for each trunc. | |||
12319 | for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end(); | |||
12320 | UI != UIEnd; ++UI) { | |||
12321 | // Skip the uses of the chain. | |||
12322 | if (UI.getUse().getResNo() != 0) | |||
12323 | continue; | |||
12324 | ||||
12325 | SDNode *User = *UI; | |||
12326 | unsigned Shift = 0; | |||
12327 | ||||
12328 | // Check if this is a trunc(lshr). | |||
12329 | if (User->getOpcode() == ISD::SRL && User->hasOneUse() && | |||
12330 | isa<ConstantSDNode>(User->getOperand(1))) { | |||
12331 | Shift = User->getConstantOperandVal(1); | |||
12332 | User = *User->use_begin(); | |||
12333 | } | |||
12334 | ||||
12335 | // At this point, User is a Truncate, iff we encountered, trunc or | |||
12336 | // trunc(lshr). | |||
12337 | if (User->getOpcode() != ISD::TRUNCATE) | |||
12338 | return false; | |||
12339 | ||||
12340 | // The width of the type must be a power of 2 and greater than 8-bits. | |||
12341 | // Otherwise the load cannot be represented in LLVM IR. | |||
12342 | // Moreover, if we shifted with a non-8-bits multiple, the slice | |||
12343 | // will be across several bytes. We do not support that. | |||
12344 | unsigned Width = User->getValueSizeInBits(0); | |||
12345 | if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7)) | |||
12346 | return false; | |||
12347 | ||||
12348 | // Build the slice for this chain of computations. | |||
12349 | LoadedSlice LS(User, LD, Shift, &DAG); | |||
12350 | APInt CurrentUsedBits = LS.getUsedBits(); | |||
12351 | ||||
12352 | // Check if this slice overlaps with another. | |||
12353 | if ((CurrentUsedBits & UsedBits) != 0) | |||
12354 | return false; | |||
12355 | // Update the bits used globally. | |||
12356 | UsedBits |= CurrentUsedBits; | |||
12357 | ||||
12358 | // Check if the new slice would be legal. | |||
12359 | if (!LS.isLegal()) | |||
12360 | return false; | |||
12361 | ||||
12362 | // Record the slice. | |||
12363 | LoadedSlices.push_back(LS); | |||
12364 | } | |||
12365 | ||||
12366 | // Abort slicing if it does not seem to be profitable. | |||
12367 | if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize)) | |||
12368 | return false; | |||
12369 | ||||
12370 | ++SlicedLoads; | |||
12371 | ||||
12372 | // Rewrite each chain to use an independent load. | |||
12373 | // By construction, each chain can be represented by a unique load. | |||
12374 | ||||
12375 | // Prepare the argument for the new token factor for all the slices. | |||
12376 | SmallVector<SDValue, 8> ArgChains; | |||
12377 | for (SmallVectorImpl<LoadedSlice>::const_iterator | |||
12378 | LSIt = LoadedSlices.begin(), | |||
12379 | LSItEnd = LoadedSlices.end(); | |||
12380 | LSIt != LSItEnd; ++LSIt) { | |||
12381 | SDValue SliceInst = LSIt->loadSlice(); | |||
12382 | CombineTo(LSIt->Inst, SliceInst, true); | |||
12383 | if (SliceInst.getOpcode() != ISD::LOAD) | |||
12384 | SliceInst = SliceInst.getOperand(0); | |||
12385 | assert(SliceInst->getOpcode() == ISD::LOAD &&(static_cast <bool> (SliceInst->getOpcode() == ISD:: LOAD && "It takes more than a zext to get to the loaded slice!!" ) ? void (0) : __assert_fail ("SliceInst->getOpcode() == ISD::LOAD && \"It takes more than a zext to get to the loaded slice!!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12386, __extension__ __PRETTY_FUNCTION__)) | |||
12386 | "It takes more than a zext to get to the loaded slice!!")(static_cast <bool> (SliceInst->getOpcode() == ISD:: LOAD && "It takes more than a zext to get to the loaded slice!!" ) ? void (0) : __assert_fail ("SliceInst->getOpcode() == ISD::LOAD && \"It takes more than a zext to get to the loaded slice!!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12386, __extension__ __PRETTY_FUNCTION__)); | |||
12387 | ArgChains.push_back(SliceInst.getValue(1)); | |||
12388 | } | |||
12389 | ||||
12390 | SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other, | |||
12391 | ArgChains); | |||
12392 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain); | |||
12393 | AddToWorklist(Chain.getNode()); | |||
12394 | return true; | |||
12395 | } | |||
12396 | ||||
12397 | /// Check to see if V is (and load (ptr), imm), where the load is having | |||
12398 | /// specific bytes cleared out. If so, return the byte size being masked out | |||
12399 | /// and the shift amount. | |||
12400 | static std::pair<unsigned, unsigned> | |||
12401 | CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) { | |||
12402 | std::pair<unsigned, unsigned> Result(0, 0); | |||
12403 | ||||
12404 | // Check for the structure we're looking for. | |||
12405 | if (V->getOpcode() != ISD::AND || | |||
12406 | !isa<ConstantSDNode>(V->getOperand(1)) || | |||
12407 | !ISD::isNormalLoad(V->getOperand(0).getNode())) | |||
12408 | return Result; | |||
12409 | ||||
12410 | // Check the chain and pointer. | |||
12411 | LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0)); | |||
12412 | if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer. | |||
12413 | ||||
12414 | // The store should be chained directly to the load or be an operand of a | |||
12415 | // tokenfactor. | |||
12416 | if (LD == Chain.getNode()) | |||
12417 | ; // ok. | |||
12418 | else if (Chain->getOpcode() != ISD::TokenFactor) | |||
12419 | return Result; // Fail. | |||
12420 | else { | |||
12421 | bool isOk = false; | |||
12422 | for (const SDValue &ChainOp : Chain->op_values()) | |||
12423 | if (ChainOp.getNode() == LD) { | |||
12424 | isOk = true; | |||
12425 | break; | |||
12426 | } | |||
12427 | if (!isOk) return Result; | |||
12428 | } | |||
12429 | ||||
12430 | // This only handles simple types. | |||
12431 | if (V.getValueType() != MVT::i16 && | |||
12432 | V.getValueType() != MVT::i32 && | |||
12433 | V.getValueType() != MVT::i64) | |||
12434 | return Result; | |||
12435 | ||||
12436 | // Check the constant mask. Invert it so that the bits being masked out are | |||
12437 | // 0 and the bits being kept are 1. Use getSExtValue so that leading bits | |||
12438 | // follow the sign bit for uniformity. | |||
12439 | uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue(); | |||
12440 | unsigned NotMaskLZ = countLeadingZeros(NotMask); | |||
12441 | if (NotMaskLZ & 7) return Result; // Must be multiple of a byte. | |||
12442 | unsigned NotMaskTZ = countTrailingZeros(NotMask); | |||
12443 | if (NotMaskTZ & 7) return Result; // Must be multiple of a byte. | |||
12444 | if (NotMaskLZ == 64) return Result; // All zero mask. | |||
12445 | ||||
12446 | // See if we have a continuous run of bits. If so, we have 0*1+0* | |||
12447 | if (countTrailingOnes(NotMask >> NotMaskTZ) + NotMaskTZ + NotMaskLZ != 64) | |||
12448 | return Result; | |||
12449 | ||||
12450 | // Adjust NotMaskLZ down to be from the actual size of the int instead of i64. | |||
12451 | if (V.getValueType() != MVT::i64 && NotMaskLZ) | |||
12452 | NotMaskLZ -= 64-V.getValueSizeInBits(); | |||
12453 | ||||
12454 | unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8; | |||
12455 | switch (MaskedBytes) { | |||
12456 | case 1: | |||
12457 | case 2: | |||
12458 | case 4: break; | |||
12459 | default: return Result; // All one mask, or 5-byte mask. | |||
12460 | } | |||
12461 | ||||
12462 | // Verify that the first bit starts at a multiple of mask so that the access | |||
12463 | // is aligned the same as the access width. | |||
12464 | if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result; | |||
12465 | ||||
12466 | Result.first = MaskedBytes; | |||
12467 | Result.second = NotMaskTZ/8; | |||
12468 | return Result; | |||
12469 | } | |||
12470 | ||||
12471 | /// Check to see if IVal is something that provides a value as specified by | |||
12472 | /// MaskInfo. If so, replace the specified store with a narrower store of | |||
12473 | /// truncated IVal. | |||
12474 | static SDNode * | |||
12475 | ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo, | |||
12476 | SDValue IVal, StoreSDNode *St, | |||
12477 | DAGCombiner *DC) { | |||
12478 | unsigned NumBytes = MaskInfo.first; | |||
12479 | unsigned ByteShift = MaskInfo.second; | |||
12480 | SelectionDAG &DAG = DC->getDAG(); | |||
12481 | ||||
12482 | // Check to see if IVal is all zeros in the part being masked in by the 'or' | |||
12483 | // that uses this. If not, this is not a replacement. | |||
12484 | APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(), | |||
12485 | ByteShift*8, (ByteShift+NumBytes)*8); | |||
12486 | if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr; | |||
12487 | ||||
12488 | // Check that it is legal on the target to do this. It is legal if the new | |||
12489 | // VT we're shrinking to (i8/i16/i32) is legal or we're still before type | |||
12490 | // legalization. | |||
12491 | MVT VT = MVT::getIntegerVT(NumBytes*8); | |||
12492 | if (!DC->isTypeLegal(VT)) | |||
12493 | return nullptr; | |||
12494 | ||||
12495 | // Okay, we can do this! Replace the 'St' store with a store of IVal that is | |||
12496 | // shifted by ByteShift and truncated down to NumBytes. | |||
12497 | if (ByteShift) { | |||
12498 | SDLoc DL(IVal); | |||
12499 | IVal = DAG.getNode(ISD::SRL, DL, IVal.getValueType(), IVal, | |||
12500 | DAG.getConstant(ByteShift*8, DL, | |||
12501 | DC->getShiftAmountTy(IVal.getValueType()))); | |||
12502 | } | |||
12503 | ||||
12504 | // Figure out the offset for the store and the alignment of the access. | |||
12505 | unsigned StOffset; | |||
12506 | unsigned NewAlign = St->getAlignment(); | |||
12507 | ||||
12508 | if (DAG.getDataLayout().isLittleEndian()) | |||
12509 | StOffset = ByteShift; | |||
12510 | else | |||
12511 | StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes; | |||
12512 | ||||
12513 | SDValue Ptr = St->getBasePtr(); | |||
12514 | if (StOffset) { | |||
12515 | SDLoc DL(IVal); | |||
12516 | Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), | |||
12517 | Ptr, DAG.getConstant(StOffset, DL, Ptr.getValueType())); | |||
12518 | NewAlign = MinAlign(NewAlign, StOffset); | |||
12519 | } | |||
12520 | ||||
12521 | // Truncate down to the new size. | |||
12522 | IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal); | |||
12523 | ||||
12524 | ++OpsNarrowed; | |||
12525 | return DAG | |||
12526 | .getStore(St->getChain(), SDLoc(St), IVal, Ptr, | |||
12527 | St->getPointerInfo().getWithOffset(StOffset), NewAlign) | |||
12528 | .getNode(); | |||
12529 | } | |||
12530 | ||||
12531 | /// Look for sequence of load / op / store where op is one of 'or', 'xor', and | |||
12532 | /// 'and' of immediates. If 'op' is only touching some of the loaded bits, try | |||
12533 | /// narrowing the load and store if it would end up being a win for performance | |||
12534 | /// or code size. | |||
12535 | SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) { | |||
12536 | StoreSDNode *ST = cast<StoreSDNode>(N); | |||
12537 | if (ST->isVolatile()) | |||
12538 | return SDValue(); | |||
12539 | ||||
12540 | SDValue Chain = ST->getChain(); | |||
12541 | SDValue Value = ST->getValue(); | |||
12542 | SDValue Ptr = ST->getBasePtr(); | |||
12543 | EVT VT = Value.getValueType(); | |||
12544 | ||||
12545 | if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse()) | |||
12546 | return SDValue(); | |||
12547 | ||||
12548 | unsigned Opc = Value.getOpcode(); | |||
12549 | ||||
12550 | // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst | |||
12551 | // is a byte mask indicating a consecutive number of bytes, check to see if | |||
12552 | // Y is known to provide just those bytes. If so, we try to replace the | |||
12553 | // load + replace + store sequence with a single (narrower) store, which makes | |||
12554 | // the load dead. | |||
12555 | if (Opc == ISD::OR) { | |||
12556 | std::pair<unsigned, unsigned> MaskedLoad; | |||
12557 | MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain); | |||
12558 | if (MaskedLoad.first) | |||
12559 | if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad, | |||
12560 | Value.getOperand(1), ST,this)) | |||
12561 | return SDValue(NewST, 0); | |||
12562 | ||||
12563 | // Or is commutative, so try swapping X and Y. | |||
12564 | MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain); | |||
12565 | if (MaskedLoad.first) | |||
12566 | if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad, | |||
12567 | Value.getOperand(0), ST,this)) | |||
12568 | return SDValue(NewST, 0); | |||
12569 | } | |||
12570 | ||||
12571 | if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) || | |||
12572 | Value.getOperand(1).getOpcode() != ISD::Constant) | |||
12573 | return SDValue(); | |||
12574 | ||||
12575 | SDValue N0 = Value.getOperand(0); | |||
12576 | if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() && | |||
12577 | Chain == SDValue(N0.getNode(), 1)) { | |||
12578 | LoadSDNode *LD = cast<LoadSDNode>(N0); | |||
12579 | if (LD->getBasePtr() != Ptr || | |||
12580 | LD->getPointerInfo().getAddrSpace() != | |||
12581 | ST->getPointerInfo().getAddrSpace()) | |||
12582 | return SDValue(); | |||
12583 | ||||
12584 | // Find the type to narrow it the load / op / store to. | |||
12585 | SDValue N1 = Value.getOperand(1); | |||
12586 | unsigned BitWidth = N1.getValueSizeInBits(); | |||
12587 | APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue(); | |||
12588 | if (Opc == ISD::AND) | |||
12589 | Imm ^= APInt::getAllOnesValue(BitWidth); | |||
12590 | if (Imm == 0 || Imm.isAllOnesValue()) | |||
12591 | return SDValue(); | |||
12592 | unsigned ShAmt = Imm.countTrailingZeros(); | |||
12593 | unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1; | |||
12594 | unsigned NewBW = NextPowerOf2(MSB - ShAmt); | |||
12595 | EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW); | |||
12596 | // The narrowing should be profitable, the load/store operation should be | |||
12597 | // legal (or custom) and the store size should be equal to the NewVT width. | |||
12598 | while (NewBW < BitWidth && | |||
12599 | (NewVT.getStoreSizeInBits() != NewBW || | |||
12600 | !TLI.isOperationLegalOrCustom(Opc, NewVT) || | |||
12601 | !TLI.isNarrowingProfitable(VT, NewVT))) { | |||
12602 | NewBW = NextPowerOf2(NewBW); | |||
12603 | NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW); | |||
12604 | } | |||
12605 | if (NewBW >= BitWidth) | |||
12606 | return SDValue(); | |||
12607 | ||||
12608 | // If the lsb changed does not start at the type bitwidth boundary, | |||
12609 | // start at the previous one. | |||
12610 | if (ShAmt % NewBW) | |||
12611 | ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW; | |||
12612 | APInt Mask = APInt::getBitsSet(BitWidth, ShAmt, | |||
12613 | std::min(BitWidth, ShAmt + NewBW)); | |||
12614 | if ((Imm & Mask) == Imm) { | |||
12615 | APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW); | |||
12616 | if (Opc == ISD::AND) | |||
12617 | NewImm ^= APInt::getAllOnesValue(NewBW); | |||
12618 | uint64_t PtrOff = ShAmt / 8; | |||
12619 | // For big endian targets, we need to adjust the offset to the pointer to | |||
12620 | // load the correct bytes. | |||
12621 | if (DAG.getDataLayout().isBigEndian()) | |||
12622 | PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff; | |||
12623 | ||||
12624 | unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff); | |||
12625 | Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext()); | |||
12626 | if (NewAlign < DAG.getDataLayout().getABITypeAlignment(NewVTTy)) | |||
12627 | return SDValue(); | |||
12628 | ||||
12629 | SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD), | |||
12630 | Ptr.getValueType(), Ptr, | |||
12631 | DAG.getConstant(PtrOff, SDLoc(LD), | |||
12632 | Ptr.getValueType())); | |||
12633 | SDValue NewLD = | |||
12634 | DAG.getLoad(NewVT, SDLoc(N0), LD->getChain(), NewPtr, | |||
12635 | LD->getPointerInfo().getWithOffset(PtrOff), NewAlign, | |||
12636 | LD->getMemOperand()->getFlags(), LD->getAAInfo()); | |||
12637 | SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD, | |||
12638 | DAG.getConstant(NewImm, SDLoc(Value), | |||
12639 | NewVT)); | |||
12640 | SDValue NewST = | |||
12641 | DAG.getStore(Chain, SDLoc(N), NewVal, NewPtr, | |||
12642 | ST->getPointerInfo().getWithOffset(PtrOff), NewAlign); | |||
12643 | ||||
12644 | AddToWorklist(NewPtr.getNode()); | |||
12645 | AddToWorklist(NewLD.getNode()); | |||
12646 | AddToWorklist(NewVal.getNode()); | |||
12647 | WorklistRemover DeadNodes(*this); | |||
12648 | DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1)); | |||
12649 | ++OpsNarrowed; | |||
12650 | return NewST; | |||
12651 | } | |||
12652 | } | |||
12653 | ||||
12654 | return SDValue(); | |||
12655 | } | |||
12656 | ||||
12657 | /// For a given floating point load / store pair, if the load value isn't used | |||
12658 | /// by any other operations, then consider transforming the pair to integer | |||
12659 | /// load / store operations if the target deems the transformation profitable. | |||
12660 | SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) { | |||
12661 | StoreSDNode *ST = cast<StoreSDNode>(N); | |||
12662 | SDValue Chain = ST->getChain(); | |||
12663 | SDValue Value = ST->getValue(); | |||
12664 | if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) && | |||
12665 | Value.hasOneUse() && | |||
12666 | Chain == SDValue(Value.getNode(), 1)) { | |||
12667 | LoadSDNode *LD = cast<LoadSDNode>(Value); | |||
12668 | EVT VT = LD->getMemoryVT(); | |||
12669 | if (!VT.isFloatingPoint() || | |||
12670 | VT != ST->getMemoryVT() || | |||
12671 | LD->isNonTemporal() || | |||
12672 | ST->isNonTemporal() || | |||
12673 | LD->getPointerInfo().getAddrSpace() != 0 || | |||
12674 | ST->getPointerInfo().getAddrSpace() != 0) | |||
12675 | return SDValue(); | |||
12676 | ||||
12677 | EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits()); | |||
12678 | if (!TLI.isOperationLegal(ISD::LOAD, IntVT) || | |||
12679 | !TLI.isOperationLegal(ISD::STORE, IntVT) || | |||
12680 | !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) || | |||
12681 | !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT)) | |||
12682 | return SDValue(); | |||
12683 | ||||
12684 | unsigned LDAlign = LD->getAlignment(); | |||
12685 | unsigned STAlign = ST->getAlignment(); | |||
12686 | Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext()); | |||
12687 | unsigned ABIAlign = DAG.getDataLayout().getABITypeAlignment(IntVTTy); | |||
12688 | if (LDAlign < ABIAlign || STAlign < ABIAlign) | |||
12689 | return SDValue(); | |||
12690 | ||||
12691 | SDValue NewLD = | |||
12692 | DAG.getLoad(IntVT, SDLoc(Value), LD->getChain(), LD->getBasePtr(), | |||
12693 | LD->getPointerInfo(), LDAlign); | |||
12694 | ||||
12695 | SDValue NewST = | |||
12696 | DAG.getStore(NewLD.getValue(1), SDLoc(N), NewLD, ST->getBasePtr(), | |||
12697 | ST->getPointerInfo(), STAlign); | |||
12698 | ||||
12699 | AddToWorklist(NewLD.getNode()); | |||
12700 | AddToWorklist(NewST.getNode()); | |||
12701 | WorklistRemover DeadNodes(*this); | |||
12702 | DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1)); | |||
12703 | ++LdStFP2Int; | |||
12704 | return NewST; | |||
12705 | } | |||
12706 | ||||
12707 | return SDValue(); | |||
12708 | } | |||
12709 | ||||
12710 | // This is a helper function for visitMUL to check the profitability | |||
12711 | // of folding (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2). | |||
12712 | // MulNode is the original multiply, AddNode is (add x, c1), | |||
12713 | // and ConstNode is c2. | |||
12714 | // | |||
12715 | // If the (add x, c1) has multiple uses, we could increase | |||
12716 | // the number of adds if we make this transformation. | |||
12717 | // It would only be worth doing this if we can remove a | |||
12718 | // multiply in the process. Check for that here. | |||
12719 | // To illustrate: | |||
12720 | // (A + c1) * c3 | |||
12721 | // (A + c2) * c3 | |||
12722 | // We're checking for cases where we have common "c3 * A" expressions. | |||
12723 | bool DAGCombiner::isMulAddWithConstProfitable(SDNode *MulNode, | |||
12724 | SDValue &AddNode, | |||
12725 | SDValue &ConstNode) { | |||
12726 | APInt Val; | |||
12727 | ||||
12728 | // If the add only has one use, this would be OK to do. | |||
12729 | if (AddNode.getNode()->hasOneUse()) | |||
12730 | return true; | |||
12731 | ||||
12732 | // Walk all the users of the constant with which we're multiplying. | |||
12733 | for (SDNode *Use : ConstNode->uses()) { | |||
12734 | if (Use == MulNode) // This use is the one we're on right now. Skip it. | |||
12735 | continue; | |||
12736 | ||||
12737 | if (Use->getOpcode() == ISD::MUL) { // We have another multiply use. | |||
12738 | SDNode *OtherOp; | |||
12739 | SDNode *MulVar = AddNode.getOperand(0).getNode(); | |||
12740 | ||||
12741 | // OtherOp is what we're multiplying against the constant. | |||
12742 | if (Use->getOperand(0) == ConstNode) | |||
12743 | OtherOp = Use->getOperand(1).getNode(); | |||
12744 | else | |||
12745 | OtherOp = Use->getOperand(0).getNode(); | |||
12746 | ||||
12747 | // Check to see if multiply is with the same operand of our "add". | |||
12748 | // | |||
12749 | // ConstNode = CONST | |||
12750 | // Use = ConstNode * A <-- visiting Use. OtherOp is A. | |||
12751 | // ... | |||
12752 | // AddNode = (A + c1) <-- MulVar is A. | |||
12753 | // = AddNode * ConstNode <-- current visiting instruction. | |||
12754 | // | |||
12755 | // If we make this transformation, we will have a common | |||
12756 | // multiply (ConstNode * A) that we can save. | |||
12757 | if (OtherOp == MulVar) | |||
12758 | return true; | |||
12759 | ||||
12760 | // Now check to see if a future expansion will give us a common | |||
12761 | // multiply. | |||
12762 | // | |||
12763 | // ConstNode = CONST | |||
12764 | // AddNode = (A + c1) | |||
12765 | // ... = AddNode * ConstNode <-- current visiting instruction. | |||
12766 | // ... | |||
12767 | // OtherOp = (A + c2) | |||
12768 | // Use = OtherOp * ConstNode <-- visiting Use. | |||
12769 | // | |||
12770 | // If we make this transformation, we will have a common | |||
12771 | // multiply (CONST * A) after we also do the same transformation | |||
12772 | // to the "t2" instruction. | |||
12773 | if (OtherOp->getOpcode() == ISD::ADD && | |||
12774 | DAG.isConstantIntBuildVectorOrConstantInt(OtherOp->getOperand(1)) && | |||
12775 | OtherOp->getOperand(0).getNode() == MulVar) | |||
12776 | return true; | |||
12777 | } | |||
12778 | } | |||
12779 | ||||
12780 | // Didn't find a case where this would be profitable. | |||
12781 | return false; | |||
12782 | } | |||
12783 | ||||
12784 | static SDValue peekThroughBitcast(SDValue V) { | |||
12785 | while (V.getOpcode() == ISD::BITCAST) | |||
12786 | V = V.getOperand(0); | |||
12787 | return V; | |||
12788 | } | |||
12789 | ||||
12790 | SDValue DAGCombiner::getMergeStoreChains(SmallVectorImpl<MemOpLink> &StoreNodes, | |||
12791 | unsigned NumStores) { | |||
12792 | SmallVector<SDValue, 8> Chains; | |||
12793 | SmallPtrSet<const SDNode *, 8> Visited; | |||
12794 | SDLoc StoreDL(StoreNodes[0].MemNode); | |||
12795 | ||||
12796 | for (unsigned i = 0; i < NumStores; ++i) { | |||
12797 | Visited.insert(StoreNodes[i].MemNode); | |||
12798 | } | |||
12799 | ||||
12800 | // don't include nodes that are children | |||
12801 | for (unsigned i = 0; i < NumStores; ++i) { | |||
12802 | if (Visited.count(StoreNodes[i].MemNode->getChain().getNode()) == 0) | |||
12803 | Chains.push_back(StoreNodes[i].MemNode->getChain()); | |||
12804 | } | |||
12805 | ||||
12806 | assert(Chains.size() > 0 && "Chain should have generated a chain")(static_cast <bool> (Chains.size() > 0 && "Chain should have generated a chain" ) ? void (0) : __assert_fail ("Chains.size() > 0 && \"Chain should have generated a chain\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12806, __extension__ __PRETTY_FUNCTION__)); | |||
12807 | return DAG.getNode(ISD::TokenFactor, StoreDL, MVT::Other, Chains); | |||
12808 | } | |||
12809 | ||||
12810 | bool DAGCombiner::MergeStoresOfConstantsOrVecElts( | |||
12811 | SmallVectorImpl<MemOpLink> &StoreNodes, EVT MemVT, unsigned NumStores, | |||
12812 | bool IsConstantSrc, bool UseVector, bool UseTrunc) { | |||
12813 | // Make sure we have something to merge. | |||
12814 | if (NumStores < 2) | |||
12815 | return false; | |||
12816 | ||||
12817 | // The latest Node in the DAG. | |||
12818 | SDLoc DL(StoreNodes[0].MemNode); | |||
12819 | ||||
12820 | int64_t ElementSizeBits = MemVT.getStoreSizeInBits(); | |||
12821 | unsigned SizeInBits = NumStores * ElementSizeBits; | |||
12822 | unsigned NumMemElts = MemVT.isVector() ? MemVT.getVectorNumElements() : 1; | |||
12823 | ||||
12824 | EVT StoreTy; | |||
12825 | if (UseVector) { | |||
12826 | unsigned Elts = NumStores * NumMemElts; | |||
12827 | // Get the type for the merged vector store. | |||
12828 | StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(), Elts); | |||
12829 | } else | |||
12830 | StoreTy = EVT::getIntegerVT(*DAG.getContext(), SizeInBits); | |||
12831 | ||||
12832 | SDValue StoredVal; | |||
12833 | if (UseVector) { | |||
12834 | if (IsConstantSrc) { | |||
12835 | SmallVector<SDValue, 8> BuildVector; | |||
12836 | for (unsigned I = 0; I != NumStores; ++I) { | |||
12837 | StoreSDNode *St = cast<StoreSDNode>(StoreNodes[I].MemNode); | |||
12838 | SDValue Val = St->getValue(); | |||
12839 | // If constant is of the wrong type, convert it now. | |||
12840 | if (MemVT != Val.getValueType()) { | |||
12841 | Val = peekThroughBitcast(Val); | |||
12842 | // Deal with constants of wrong size. | |||
12843 | if (ElementSizeBits != Val.getValueSizeInBits()) { | |||
12844 | EVT IntMemVT = | |||
12845 | EVT::getIntegerVT(*DAG.getContext(), MemVT.getSizeInBits()); | |||
12846 | if (isa<ConstantFPSDNode>(Val)) { | |||
12847 | // Not clear how to truncate FP values. | |||
12848 | return false; | |||
12849 | } else if (auto *C = dyn_cast<ConstantSDNode>(Val)) | |||
12850 | Val = DAG.getConstant(C->getAPIntValue() | |||
12851 | .zextOrTrunc(Val.getValueSizeInBits()) | |||
12852 | .zextOrTrunc(ElementSizeBits), | |||
12853 | SDLoc(C), IntMemVT); | |||
12854 | } | |||
12855 | // Make sure correctly size type is the correct type. | |||
12856 | Val = DAG.getBitcast(MemVT, Val); | |||
12857 | } | |||
12858 | BuildVector.push_back(Val); | |||
12859 | } | |||
12860 | StoredVal = DAG.getNode(MemVT.isVector() ? ISD::CONCAT_VECTORS | |||
12861 | : ISD::BUILD_VECTOR, | |||
12862 | DL, StoreTy, BuildVector); | |||
12863 | } else { | |||
12864 | SmallVector<SDValue, 8> Ops; | |||
12865 | for (unsigned i = 0; i < NumStores; ++i) { | |||
12866 | StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode); | |||
12867 | SDValue Val = peekThroughBitcast(St->getValue()); | |||
12868 | // All operands of BUILD_VECTOR / CONCAT_VECTOR must be of | |||
12869 | // type MemVT. If the underlying value is not the correct | |||
12870 | // type, but it is an extraction of an appropriate vector we | |||
12871 | // can recast Val to be of the correct type. This may require | |||
12872 | // converting between EXTRACT_VECTOR_ELT and | |||
12873 | // EXTRACT_SUBVECTOR. | |||
12874 | if ((MemVT != Val.getValueType()) && | |||
12875 | (Val.getOpcode() == ISD::EXTRACT_VECTOR_ELT || | |||
12876 | Val.getOpcode() == ISD::EXTRACT_SUBVECTOR)) { | |||
12877 | SDValue Vec = Val.getOperand(0); | |||
12878 | EVT MemVTScalarTy = MemVT.getScalarType(); | |||
12879 | // We may need to add a bitcast here to get types to line up. | |||
12880 | if (MemVTScalarTy != Vec.getValueType()) { | |||
12881 | unsigned Elts = Vec.getValueType().getSizeInBits() / | |||
12882 | MemVTScalarTy.getSizeInBits(); | |||
12883 | EVT NewVecTy = | |||
12884 | EVT::getVectorVT(*DAG.getContext(), MemVTScalarTy, Elts); | |||
12885 | Vec = DAG.getBitcast(NewVecTy, Vec); | |||
12886 | } | |||
12887 | auto OpC = (MemVT.isVector()) ? ISD::EXTRACT_SUBVECTOR | |||
12888 | : ISD::EXTRACT_VECTOR_ELT; | |||
12889 | Val = DAG.getNode(OpC, SDLoc(Val), MemVT, Vec, Val.getOperand(1)); | |||
12890 | } | |||
12891 | Ops.push_back(Val); | |||
12892 | } | |||
12893 | ||||
12894 | // Build the extracted vector elements back into a vector. | |||
12895 | StoredVal = DAG.getNode(MemVT.isVector() ? ISD::CONCAT_VECTORS | |||
12896 | : ISD::BUILD_VECTOR, | |||
12897 | DL, StoreTy, Ops); | |||
12898 | } | |||
12899 | } else { | |||
12900 | // We should always use a vector store when merging extracted vector | |||
12901 | // elements, so this path implies a store of constants. | |||
12902 | assert(IsConstantSrc && "Merged vector elements should use vector store")(static_cast <bool> (IsConstantSrc && "Merged vector elements should use vector store" ) ? void (0) : __assert_fail ("IsConstantSrc && \"Merged vector elements should use vector store\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12902, __extension__ __PRETTY_FUNCTION__)); | |||
12903 | ||||
12904 | APInt StoreInt(SizeInBits, 0); | |||
12905 | ||||
12906 | // Construct a single integer constant which is made of the smaller | |||
12907 | // constant inputs. | |||
12908 | bool IsLE = DAG.getDataLayout().isLittleEndian(); | |||
12909 | for (unsigned i = 0; i < NumStores; ++i) { | |||
12910 | unsigned Idx = IsLE ? (NumStores - 1 - i) : i; | |||
12911 | StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode); | |||
12912 | ||||
12913 | SDValue Val = St->getValue(); | |||
12914 | StoreInt <<= ElementSizeBits; | |||
12915 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) { | |||
12916 | StoreInt |= C->getAPIntValue() | |||
12917 | .zextOrTrunc(ElementSizeBits) | |||
12918 | .zextOrTrunc(SizeInBits); | |||
12919 | } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) { | |||
12920 | StoreInt |= C->getValueAPF() | |||
12921 | .bitcastToAPInt() | |||
12922 | .zextOrTrunc(ElementSizeBits) | |||
12923 | .zextOrTrunc(SizeInBits); | |||
12924 | // If fp truncation is necessary give up for now. | |||
12925 | if (MemVT.getSizeInBits() != ElementSizeBits) | |||
12926 | return false; | |||
12927 | } else { | |||
12928 | llvm_unreachable("Invalid constant element type")::llvm::llvm_unreachable_internal("Invalid constant element type" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 12928); | |||
12929 | } | |||
12930 | } | |||
12931 | ||||
12932 | // Create the new Load and Store operations. | |||
12933 | StoredVal = DAG.getConstant(StoreInt, DL, StoreTy); | |||
12934 | } | |||
12935 | ||||
12936 | LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode; | |||
12937 | SDValue NewChain = getMergeStoreChains(StoreNodes, NumStores); | |||
12938 | ||||
12939 | // make sure we use trunc store if it's necessary to be legal. | |||
12940 | SDValue NewStore; | |||
12941 | if (!UseTrunc) { | |||
12942 | NewStore = DAG.getStore(NewChain, DL, StoredVal, FirstInChain->getBasePtr(), | |||
12943 | FirstInChain->getPointerInfo(), | |||
12944 | FirstInChain->getAlignment()); | |||
12945 | } else { // Must be realized as a trunc store | |||
12946 | EVT LegalizedStoredValueTy = | |||
12947 | TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType()); | |||
12948 | unsigned LegalizedStoreSize = LegalizedStoredValueTy.getSizeInBits(); | |||
12949 | ConstantSDNode *C = cast<ConstantSDNode>(StoredVal); | |||
12950 | SDValue ExtendedStoreVal = | |||
12951 | DAG.getConstant(C->getAPIntValue().zextOrTrunc(LegalizedStoreSize), DL, | |||
12952 | LegalizedStoredValueTy); | |||
12953 | NewStore = DAG.getTruncStore( | |||
12954 | NewChain, DL, ExtendedStoreVal, FirstInChain->getBasePtr(), | |||
12955 | FirstInChain->getPointerInfo(), StoredVal.getValueType() /*TVT*/, | |||
12956 | FirstInChain->getAlignment(), | |||
12957 | FirstInChain->getMemOperand()->getFlags()); | |||
12958 | } | |||
12959 | ||||
12960 | // Replace all merged stores with the new store. | |||
12961 | for (unsigned i = 0; i < NumStores; ++i) | |||
12962 | CombineTo(StoreNodes[i].MemNode, NewStore); | |||
12963 | ||||
12964 | AddToWorklist(NewChain.getNode()); | |||
12965 | return true; | |||
12966 | } | |||
12967 | ||||
12968 | void DAGCombiner::getStoreMergeCandidates( | |||
12969 | StoreSDNode *St, SmallVectorImpl<MemOpLink> &StoreNodes) { | |||
12970 | // This holds the base pointer, index, and the offset in bytes from the base | |||
12971 | // pointer. | |||
12972 | BaseIndexOffset BasePtr = BaseIndexOffset::match(St, DAG); | |||
12973 | EVT MemVT = St->getMemoryVT(); | |||
12974 | ||||
12975 | SDValue Val = peekThroughBitcast(St->getValue()); | |||
12976 | // We must have a base and an offset. | |||
12977 | if (!BasePtr.getBase().getNode()) | |||
12978 | return; | |||
12979 | ||||
12980 | // Do not handle stores to undef base pointers. | |||
12981 | if (BasePtr.getBase().isUndef()) | |||
12982 | return; | |||
12983 | ||||
12984 | bool IsConstantSrc = isa<ConstantSDNode>(Val) || isa<ConstantFPSDNode>(Val); | |||
12985 | bool IsExtractVecSrc = (Val.getOpcode() == ISD::EXTRACT_VECTOR_ELT || | |||
12986 | Val.getOpcode() == ISD::EXTRACT_SUBVECTOR); | |||
12987 | bool IsLoadSrc = isa<LoadSDNode>(Val); | |||
12988 | BaseIndexOffset LBasePtr; | |||
12989 | // Match on loadbaseptr if relevant. | |||
12990 | EVT LoadVT; | |||
12991 | if (IsLoadSrc) { | |||
12992 | auto *Ld = cast<LoadSDNode>(Val); | |||
12993 | LBasePtr = BaseIndexOffset::match(Ld, DAG); | |||
12994 | LoadVT = Ld->getMemoryVT(); | |||
12995 | // Load and store should be the same type. | |||
12996 | if (MemVT != LoadVT) | |||
12997 | return; | |||
12998 | } | |||
12999 | auto CandidateMatch = [&](StoreSDNode *Other, BaseIndexOffset &Ptr, | |||
13000 | int64_t &Offset) -> bool { | |||
13001 | if (Other->isVolatile() || Other->isIndexed()) | |||
13002 | return false; | |||
13003 | SDValue Val = peekThroughBitcast(Other->getValue()); | |||
13004 | // Allow merging constants of different types as integers. | |||
13005 | bool NoTypeMatch = (MemVT.isInteger()) ? !MemVT.bitsEq(Other->getMemoryVT()) | |||
13006 | : Other->getMemoryVT() != MemVT; | |||
13007 | if (IsLoadSrc) { | |||
13008 | if (NoTypeMatch) | |||
13009 | return false; | |||
13010 | // The Load's Base Ptr must also match | |||
13011 | if (LoadSDNode *OtherLd = dyn_cast<LoadSDNode>(Val)) { | |||
13012 | auto LPtr = BaseIndexOffset::match(OtherLd, DAG); | |||
13013 | if (LoadVT != OtherLd->getMemoryVT()) | |||
13014 | return false; | |||
13015 | if (!(LBasePtr.equalBaseIndex(LPtr, DAG))) | |||
13016 | return false; | |||
13017 | } else | |||
13018 | return false; | |||
13019 | } | |||
13020 | if (IsConstantSrc) { | |||
13021 | if (NoTypeMatch) | |||
13022 | return false; | |||
13023 | if (!(isa<ConstantSDNode>(Val) || isa<ConstantFPSDNode>(Val))) | |||
13024 | return false; | |||
13025 | } | |||
13026 | if (IsExtractVecSrc) { | |||
13027 | // Do not merge truncated stores here. | |||
13028 | if (Other->isTruncatingStore()) | |||
13029 | return false; | |||
13030 | if (!MemVT.bitsEq(Val.getValueType())) | |||
13031 | return false; | |||
13032 | if (Val.getOpcode() != ISD::EXTRACT_VECTOR_ELT && | |||
13033 | Val.getOpcode() != ISD::EXTRACT_SUBVECTOR) | |||
13034 | return false; | |||
13035 | } | |||
13036 | Ptr = BaseIndexOffset::match(Other, DAG); | |||
13037 | return (BasePtr.equalBaseIndex(Ptr, DAG, Offset)); | |||
13038 | }; | |||
13039 | ||||
13040 | // We looking for a root node which is an ancestor to all mergable | |||
13041 | // stores. We search up through a load, to our root and then down | |||
13042 | // through all children. For instance we will find Store{1,2,3} if | |||
13043 | // St is Store1, Store2. or Store3 where the root is not a load | |||
13044 | // which always true for nonvolatile ops. TODO: Expand | |||
13045 | // the search to find all valid candidates through multiple layers of loads. | |||
13046 | // | |||
13047 | // Root | |||
13048 | // |-------|-------| | |||
13049 | // Load Load Store3 | |||
13050 | // | | | |||
13051 | // Store1 Store2 | |||
13052 | // | |||
13053 | // FIXME: We should be able to climb and | |||
13054 | // descend TokenFactors to find candidates as well. | |||
13055 | ||||
13056 | SDNode *RootNode = (St->getChain()).getNode(); | |||
13057 | ||||
13058 | if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(RootNode)) { | |||
13059 | RootNode = Ldn->getChain().getNode(); | |||
13060 | for (auto I = RootNode->use_begin(), E = RootNode->use_end(); I != E; ++I) | |||
13061 | if (I.getOperandNo() == 0 && isa<LoadSDNode>(*I)) // walk down chain | |||
13062 | for (auto I2 = (*I)->use_begin(), E2 = (*I)->use_end(); I2 != E2; ++I2) | |||
13063 | if (I2.getOperandNo() == 0) | |||
13064 | if (StoreSDNode *OtherST = dyn_cast<StoreSDNode>(*I2)) { | |||
13065 | BaseIndexOffset Ptr; | |||
13066 | int64_t PtrDiff; | |||
13067 | if (CandidateMatch(OtherST, Ptr, PtrDiff)) | |||
13068 | StoreNodes.push_back(MemOpLink(OtherST, PtrDiff)); | |||
13069 | } | |||
13070 | } else | |||
13071 | for (auto I = RootNode->use_begin(), E = RootNode->use_end(); I != E; ++I) | |||
13072 | if (I.getOperandNo() == 0) | |||
13073 | if (StoreSDNode *OtherST = dyn_cast<StoreSDNode>(*I)) { | |||
13074 | BaseIndexOffset Ptr; | |||
13075 | int64_t PtrDiff; | |||
13076 | if (CandidateMatch(OtherST, Ptr, PtrDiff)) | |||
13077 | StoreNodes.push_back(MemOpLink(OtherST, PtrDiff)); | |||
13078 | } | |||
13079 | } | |||
13080 | ||||
13081 | // We need to check that merging these stores does not cause a loop in | |||
13082 | // the DAG. Any store candidate may depend on another candidate | |||
13083 | // indirectly through its operand (we already consider dependencies | |||
13084 | // through the chain). Check in parallel by searching up from | |||
13085 | // non-chain operands of candidates. | |||
13086 | bool DAGCombiner::checkMergeStoreCandidatesForDependencies( | |||
13087 | SmallVectorImpl<MemOpLink> &StoreNodes, unsigned NumStores) { | |||
13088 | // FIXME: We should be able to truncate a full search of | |||
13089 | // predecessors by doing a BFS and keeping tabs the originating | |||
13090 | // stores from which worklist nodes come from in a similar way to | |||
13091 | // TokenFactor simplfication. | |||
13092 | ||||
13093 | SmallPtrSet<const SDNode *, 16> Visited; | |||
13094 | SmallVector<const SDNode *, 8> Worklist; | |||
13095 | unsigned int Max = 8192; | |||
13096 | // Search Ops of store candidates. | |||
13097 | for (unsigned i = 0; i < NumStores; ++i) { | |||
13098 | SDNode *n = StoreNodes[i].MemNode; | |||
13099 | // Potential loops may happen only through non-chain operands | |||
13100 | for (unsigned j = 1; j < n->getNumOperands(); ++j) | |||
13101 | Worklist.push_back(n->getOperand(j).getNode()); | |||
13102 | } | |||
13103 | // Search through DAG. We can stop early if we find a store node. | |||
13104 | for (unsigned i = 0; i < NumStores; ++i) { | |||
13105 | if (SDNode::hasPredecessorHelper(StoreNodes[i].MemNode, Visited, Worklist, | |||
13106 | Max)) | |||
13107 | return false; | |||
13108 | // Check if we ended early, failing conservatively if so. | |||
13109 | if (Visited.size() >= Max) | |||
13110 | return false; | |||
13111 | } | |||
13112 | return true; | |||
13113 | } | |||
13114 | ||||
13115 | bool DAGCombiner::MergeConsecutiveStores(StoreSDNode *St) { | |||
13116 | if (OptLevel == CodeGenOpt::None) | |||
13117 | return false; | |||
13118 | ||||
13119 | EVT MemVT = St->getMemoryVT(); | |||
13120 | int64_t ElementSizeBytes = MemVT.getStoreSize(); | |||
13121 | unsigned NumMemElts = MemVT.isVector() ? MemVT.getVectorNumElements() : 1; | |||
13122 | ||||
13123 | if (MemVT.getSizeInBits() * 2 > MaximumLegalStoreInBits) | |||
13124 | return false; | |||
13125 | ||||
13126 | bool NoVectors = DAG.getMachineFunction().getFunction().hasFnAttribute( | |||
13127 | Attribute::NoImplicitFloat); | |||
13128 | ||||
13129 | // This function cannot currently deal with non-byte-sized memory sizes. | |||
13130 | if (ElementSizeBytes * 8 != MemVT.getSizeInBits()) | |||
13131 | return false; | |||
13132 | ||||
13133 | if (!MemVT.isSimple()) | |||
13134 | return false; | |||
13135 | ||||
13136 | // Perform an early exit check. Do not bother looking at stored values that | |||
13137 | // are not constants, loads, or extracted vector elements. | |||
13138 | SDValue StoredVal = peekThroughBitcast(St->getValue()); | |||
13139 | bool IsLoadSrc = isa<LoadSDNode>(StoredVal); | |||
13140 | bool IsConstantSrc = isa<ConstantSDNode>(StoredVal) || | |||
13141 | isa<ConstantFPSDNode>(StoredVal); | |||
13142 | bool IsExtractVecSrc = (StoredVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT || | |||
13143 | StoredVal.getOpcode() == ISD::EXTRACT_SUBVECTOR); | |||
13144 | ||||
13145 | if (!IsConstantSrc && !IsLoadSrc && !IsExtractVecSrc) | |||
13146 | return false; | |||
13147 | ||||
13148 | SmallVector<MemOpLink, 8> StoreNodes; | |||
13149 | // Find potential store merge candidates by searching through chain sub-DAG | |||
13150 | getStoreMergeCandidates(St, StoreNodes); | |||
13151 | ||||
13152 | // Check if there is anything to merge. | |||
13153 | if (StoreNodes.size() < 2) | |||
13154 | return false; | |||
13155 | ||||
13156 | // Sort the memory operands according to their distance from the | |||
13157 | // base pointer. | |||
13158 | std::sort(StoreNodes.begin(), StoreNodes.end(), | |||
13159 | [](MemOpLink LHS, MemOpLink RHS) { | |||
13160 | return LHS.OffsetFromBase < RHS.OffsetFromBase; | |||
13161 | }); | |||
13162 | ||||
13163 | // Store Merge attempts to merge the lowest stores. This generally | |||
13164 | // works out as if successful, as the remaining stores are checked | |||
13165 | // after the first collection of stores is merged. However, in the | |||
13166 | // case that a non-mergeable store is found first, e.g., {p[-2], | |||
13167 | // p[0], p[1], p[2], p[3]}, we would fail and miss the subsequent | |||
13168 | // mergeable cases. To prevent this, we prune such stores from the | |||
13169 | // front of StoreNodes here. | |||
13170 | ||||
13171 | bool RV = false; | |||
13172 | while (StoreNodes.size() > 1) { | |||
13173 | unsigned StartIdx = 0; | |||
13174 | while ((StartIdx + 1 < StoreNodes.size()) && | |||
13175 | StoreNodes[StartIdx].OffsetFromBase + ElementSizeBytes != | |||
13176 | StoreNodes[StartIdx + 1].OffsetFromBase) | |||
13177 | ++StartIdx; | |||
13178 | ||||
13179 | // Bail if we don't have enough candidates to merge. | |||
13180 | if (StartIdx + 1 >= StoreNodes.size()) | |||
13181 | return RV; | |||
13182 | ||||
13183 | if (StartIdx) | |||
13184 | StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + StartIdx); | |||
13185 | ||||
13186 | // Scan the memory operations on the chain and find the first | |||
13187 | // non-consecutive store memory address. | |||
13188 | unsigned NumConsecutiveStores = 1; | |||
13189 | int64_t StartAddress = StoreNodes[0].OffsetFromBase; | |||
13190 | // Check that the addresses are consecutive starting from the second | |||
13191 | // element in the list of stores. | |||
13192 | for (unsigned i = 1, e = StoreNodes.size(); i < e; ++i) { | |||
13193 | int64_t CurrAddress = StoreNodes[i].OffsetFromBase; | |||
13194 | if (CurrAddress - StartAddress != (ElementSizeBytes * i)) | |||
13195 | break; | |||
13196 | NumConsecutiveStores = i + 1; | |||
13197 | } | |||
13198 | ||||
13199 | if (NumConsecutiveStores < 2) { | |||
13200 | StoreNodes.erase(StoreNodes.begin(), | |||
13201 | StoreNodes.begin() + NumConsecutiveStores); | |||
13202 | continue; | |||
13203 | } | |||
13204 | ||||
13205 | // Check that we can merge these candidates without causing a cycle | |||
13206 | if (!checkMergeStoreCandidatesForDependencies(StoreNodes, | |||
13207 | NumConsecutiveStores)) { | |||
13208 | StoreNodes.erase(StoreNodes.begin(), | |||
13209 | StoreNodes.begin() + NumConsecutiveStores); | |||
13210 | continue; | |||
13211 | } | |||
13212 | ||||
13213 | // The node with the lowest store address. | |||
13214 | LLVMContext &Context = *DAG.getContext(); | |||
13215 | const DataLayout &DL = DAG.getDataLayout(); | |||
13216 | ||||
13217 | // Store the constants into memory as one consecutive store. | |||
13218 | if (IsConstantSrc) { | |||
13219 | LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode; | |||
13220 | unsigned FirstStoreAS = FirstInChain->getAddressSpace(); | |||
13221 | unsigned FirstStoreAlign = FirstInChain->getAlignment(); | |||
13222 | unsigned LastLegalType = 1; | |||
13223 | unsigned LastLegalVectorType = 1; | |||
13224 | bool LastIntegerTrunc = false; | |||
13225 | bool NonZero = false; | |||
13226 | unsigned FirstZeroAfterNonZero = NumConsecutiveStores; | |||
13227 | for (unsigned i = 0; i < NumConsecutiveStores; ++i) { | |||
13228 | StoreSDNode *ST = cast<StoreSDNode>(StoreNodes[i].MemNode); | |||
13229 | SDValue StoredVal = ST->getValue(); | |||
13230 | bool IsElementZero = false; | |||
13231 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) | |||
13232 | IsElementZero = C->isNullValue(); | |||
13233 | else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) | |||
13234 | IsElementZero = C->getConstantFPValue()->isNullValue(); | |||
13235 | if (IsElementZero) { | |||
13236 | if (NonZero && FirstZeroAfterNonZero == NumConsecutiveStores) | |||
13237 | FirstZeroAfterNonZero = i; | |||
13238 | } | |||
13239 | NonZero |= !IsElementZero; | |||
13240 | ||||
13241 | // Find a legal type for the constant store. | |||
13242 | unsigned SizeInBits = (i + 1) * ElementSizeBytes * 8; | |||
13243 | EVT StoreTy = EVT::getIntegerVT(Context, SizeInBits); | |||
13244 | bool IsFast = false; | |||
13245 | if (TLI.isTypeLegal(StoreTy) && | |||
13246 | TLI.canMergeStoresTo(FirstStoreAS, StoreTy, DAG) && | |||
13247 | TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstStoreAS, | |||
13248 | FirstStoreAlign, &IsFast) && | |||
13249 | IsFast) { | |||
13250 | LastIntegerTrunc = false; | |||
13251 | LastLegalType = i + 1; | |||
13252 | // Or check whether a truncstore is legal. | |||
13253 | } else if (TLI.getTypeAction(Context, StoreTy) == | |||
13254 | TargetLowering::TypePromoteInteger) { | |||
13255 | EVT LegalizedStoredValueTy = | |||
13256 | TLI.getTypeToTransformTo(Context, StoredVal.getValueType()); | |||
13257 | if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) && | |||
13258 | TLI.canMergeStoresTo(FirstStoreAS, LegalizedStoredValueTy, DAG) && | |||
13259 | TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstStoreAS, | |||
13260 | FirstStoreAlign, &IsFast) && | |||
13261 | IsFast) { | |||
13262 | LastIntegerTrunc = true; | |||
13263 | LastLegalType = i + 1; | |||
13264 | } | |||
13265 | } | |||
13266 | ||||
13267 | // We only use vectors if the constant is known to be zero or the target | |||
13268 | // allows it and the function is not marked with the noimplicitfloat | |||
13269 | // attribute. | |||
13270 | if ((!NonZero || | |||
13271 | TLI.storeOfVectorConstantIsCheap(MemVT, i + 1, FirstStoreAS)) && | |||
13272 | !NoVectors) { | |||
13273 | // Find a legal type for the vector store. | |||
13274 | unsigned Elts = (i + 1) * NumMemElts; | |||
13275 | EVT Ty = EVT::getVectorVT(Context, MemVT.getScalarType(), Elts); | |||
13276 | if (TLI.isTypeLegal(Ty) && TLI.isTypeLegal(MemVT) && | |||
13277 | TLI.canMergeStoresTo(FirstStoreAS, Ty, DAG) && | |||
13278 | TLI.allowsMemoryAccess(Context, DL, Ty, FirstStoreAS, | |||
13279 | FirstStoreAlign, &IsFast) && | |||
13280 | IsFast) | |||
13281 | LastLegalVectorType = i + 1; | |||
13282 | } | |||
13283 | } | |||
13284 | ||||
13285 | bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors; | |||
13286 | unsigned NumElem = (UseVector) ? LastLegalVectorType : LastLegalType; | |||
13287 | ||||
13288 | // Check if we found a legal integer type that creates a meaningful merge. | |||
13289 | if (NumElem < 2) { | |||
13290 | // We know that candidate stores are in order and of correct | |||
13291 | // shape. While there is no mergeable sequence from the | |||
13292 | // beginning one may start later in the sequence. The only | |||
13293 | // reason a merge of size N could have failed where another of | |||
13294 | // the same size would not have, is if the alignment has | |||
13295 | // improved or we've dropped a non-zero value. Drop as many | |||
13296 | // candidates as we can here. | |||
13297 | unsigned NumSkip = 1; | |||
13298 | while ( | |||
13299 | (NumSkip < NumConsecutiveStores) && | |||
13300 | (NumSkip < FirstZeroAfterNonZero) && | |||
13301 | (StoreNodes[NumSkip].MemNode->getAlignment() <= FirstStoreAlign)) { | |||
13302 | NumSkip++; | |||
13303 | } | |||
13304 | StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumSkip); | |||
13305 | continue; | |||
13306 | } | |||
13307 | ||||
13308 | bool Merged = MergeStoresOfConstantsOrVecElts( | |||
13309 | StoreNodes, MemVT, NumElem, true, UseVector, LastIntegerTrunc); | |||
13310 | RV |= Merged; | |||
13311 | ||||
13312 | // Remove merged stores for next iteration. | |||
13313 | StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumElem); | |||
13314 | continue; | |||
13315 | } | |||
13316 | ||||
13317 | // When extracting multiple vector elements, try to store them | |||
13318 | // in one vector store rather than a sequence of scalar stores. | |||
13319 | if (IsExtractVecSrc) { | |||
13320 | LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode; | |||
13321 | unsigned FirstStoreAS = FirstInChain->getAddressSpace(); | |||
13322 | unsigned FirstStoreAlign = FirstInChain->getAlignment(); | |||
13323 | unsigned NumStoresToMerge = 1; | |||
13324 | for (unsigned i = 0; i < NumConsecutiveStores; ++i) { | |||
13325 | StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode); | |||
13326 | SDValue StVal = peekThroughBitcast(St->getValue()); | |||
13327 | // This restriction could be loosened. | |||
13328 | // Bail out if any stored values are not elements extracted from a | |||
13329 | // vector. It should be possible to handle mixed sources, but load | |||
13330 | // sources need more careful handling (see the block of code below that | |||
13331 | // handles consecutive loads). | |||
13332 | if (StVal.getOpcode() != ISD::EXTRACT_VECTOR_ELT && | |||
13333 | StVal.getOpcode() != ISD::EXTRACT_SUBVECTOR) | |||
13334 | return RV; | |||
13335 | ||||
13336 | // Find a legal type for the vector store. | |||
13337 | unsigned Elts = (i + 1) * NumMemElts; | |||
13338 | EVT Ty = | |||
13339 | EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(), Elts); | |||
13340 | bool IsFast; | |||
13341 | if (TLI.isTypeLegal(Ty) && | |||
13342 | TLI.canMergeStoresTo(FirstStoreAS, Ty, DAG) && | |||
13343 | TLI.allowsMemoryAccess(Context, DL, Ty, FirstStoreAS, | |||
13344 | FirstStoreAlign, &IsFast) && | |||
13345 | IsFast) | |||
13346 | NumStoresToMerge = i + 1; | |||
13347 | } | |||
13348 | ||||
13349 | // Check if we found a legal integer type that creates a meaningful merge. | |||
13350 | if (NumStoresToMerge < 2) { | |||
13351 | // We know that candidate stores are in order and of correct | |||
13352 | // shape. While there is no mergeable sequence from the | |||
13353 | // beginning one may start later in the sequence. The only | |||
13354 | // reason a merge of size N could have failed where another of | |||
13355 | // the same size would not have, is if the alignment has | |||
13356 | // improved. Drop as many candidates as we can here. | |||
13357 | unsigned NumSkip = 1; | |||
13358 | while ((NumSkip < NumConsecutiveStores) && | |||
13359 | (StoreNodes[NumSkip].MemNode->getAlignment() <= FirstStoreAlign)) | |||
13360 | NumSkip++; | |||
13361 | ||||
13362 | StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumSkip); | |||
13363 | continue; | |||
13364 | } | |||
13365 | ||||
13366 | bool Merged = MergeStoresOfConstantsOrVecElts( | |||
13367 | StoreNodes, MemVT, NumStoresToMerge, false, true, false); | |||
13368 | if (!Merged) { | |||
13369 | StoreNodes.erase(StoreNodes.begin(), | |||
13370 | StoreNodes.begin() + NumStoresToMerge); | |||
13371 | continue; | |||
13372 | } | |||
13373 | // Remove merged stores for next iteration. | |||
13374 | StoreNodes.erase(StoreNodes.begin(), | |||
13375 | StoreNodes.begin() + NumStoresToMerge); | |||
13376 | RV = true; | |||
13377 | continue; | |||
13378 | } | |||
13379 | ||||
13380 | // Below we handle the case of multiple consecutive stores that | |||
13381 | // come from multiple consecutive loads. We merge them into a single | |||
13382 | // wide load and a single wide store. | |||
13383 | ||||
13384 | // Look for load nodes which are used by the stored values. | |||
13385 | SmallVector<MemOpLink, 8> LoadNodes; | |||
13386 | ||||
13387 | // Find acceptable loads. Loads need to have the same chain (token factor), | |||
13388 | // must not be zext, volatile, indexed, and they must be consecutive. | |||
13389 | BaseIndexOffset LdBasePtr; | |||
13390 | for (unsigned i = 0; i < NumConsecutiveStores; ++i) { | |||
13391 | StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode); | |||
13392 | SDValue Val = peekThroughBitcast(St->getValue()); | |||
13393 | LoadSDNode *Ld = dyn_cast<LoadSDNode>(Val); | |||
13394 | if (!Ld) | |||
13395 | break; | |||
13396 | ||||
13397 | // Loads must only have one use. | |||
13398 | if (!Ld->hasNUsesOfValue(1, 0)) | |||
13399 | break; | |||
13400 | ||||
13401 | // The memory operands must not be volatile. | |||
13402 | if (Ld->isVolatile() || Ld->isIndexed()) | |||
13403 | break; | |||
13404 | ||||
13405 | // The stored memory type must be the same. | |||
13406 | if (Ld->getMemoryVT() != MemVT) | |||
13407 | break; | |||
13408 | ||||
13409 | BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld, DAG); | |||
13410 | // If this is not the first ptr that we check. | |||
13411 | int64_t LdOffset = 0; | |||
13412 | if (LdBasePtr.getBase().getNode()) { | |||
13413 | // The base ptr must be the same. | |||
13414 | if (!LdBasePtr.equalBaseIndex(LdPtr, DAG, LdOffset)) | |||
13415 | break; | |||
13416 | } else { | |||
13417 | // Check that all other base pointers are the same as this one. | |||
13418 | LdBasePtr = LdPtr; | |||
13419 | } | |||
13420 | ||||
13421 | // We found a potential memory operand to merge. | |||
13422 | LoadNodes.push_back(MemOpLink(Ld, LdOffset)); | |||
13423 | } | |||
13424 | ||||
13425 | if (LoadNodes.size() < 2) { | |||
13426 | StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + 1); | |||
13427 | continue; | |||
13428 | } | |||
13429 | ||||
13430 | // If we have load/store pair instructions and we only have two values, | |||
13431 | // don't bother merging. | |||
13432 | unsigned RequiredAlignment; | |||
13433 | if (LoadNodes.size() == 2 && TLI.hasPairedLoad(MemVT, RequiredAlignment) && | |||
13434 | StoreNodes[0].MemNode->getAlignment() >= RequiredAlignment) { | |||
13435 | StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + 2); | |||
13436 | continue; | |||
13437 | } | |||
13438 | LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode; | |||
13439 | unsigned FirstStoreAS = FirstInChain->getAddressSpace(); | |||
13440 | unsigned FirstStoreAlign = FirstInChain->getAlignment(); | |||
13441 | LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode); | |||
13442 | unsigned FirstLoadAS = FirstLoad->getAddressSpace(); | |||
13443 | unsigned FirstLoadAlign = FirstLoad->getAlignment(); | |||
13444 | ||||
13445 | // Scan the memory operations on the chain and find the first | |||
13446 | // non-consecutive load memory address. These variables hold the index in | |||
13447 | // the store node array. | |||
13448 | unsigned LastConsecutiveLoad = 1; | |||
13449 | // This variable refers to the size and not index in the array. | |||
13450 | unsigned LastLegalVectorType = 1; | |||
13451 | unsigned LastLegalIntegerType = 1; | |||
13452 | bool isDereferenceable = true; | |||
13453 | bool DoIntegerTruncate = false; | |||
13454 | StartAddress = LoadNodes[0].OffsetFromBase; | |||
13455 | SDValue FirstChain = FirstLoad->getChain(); | |||
13456 | for (unsigned i = 1; i < LoadNodes.size(); ++i) { | |||
13457 | // All loads must share the same chain. | |||
13458 | if (LoadNodes[i].MemNode->getChain() != FirstChain) | |||
13459 | break; | |||
13460 | ||||
13461 | int64_t CurrAddress = LoadNodes[i].OffsetFromBase; | |||
13462 | if (CurrAddress - StartAddress != (ElementSizeBytes * i)) | |||
13463 | break; | |||
13464 | LastConsecutiveLoad = i; | |||
13465 | ||||
13466 | if (isDereferenceable && !LoadNodes[i].MemNode->isDereferenceable()) | |||
13467 | isDereferenceable = false; | |||
13468 | ||||
13469 | // Find a legal type for the vector store. | |||
13470 | unsigned Elts = (i + 1) * NumMemElts; | |||
13471 | EVT StoreTy = EVT::getVectorVT(Context, MemVT.getScalarType(), Elts); | |||
13472 | ||||
13473 | bool IsFastSt, IsFastLd; | |||
13474 | if (TLI.isTypeLegal(StoreTy) && | |||
13475 | TLI.canMergeStoresTo(FirstStoreAS, StoreTy, DAG) && | |||
13476 | TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstStoreAS, | |||
13477 | FirstStoreAlign, &IsFastSt) && | |||
13478 | IsFastSt && | |||
13479 | TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstLoadAS, | |||
13480 | FirstLoadAlign, &IsFastLd) && | |||
13481 | IsFastLd) { | |||
13482 | LastLegalVectorType = i + 1; | |||
13483 | } | |||
13484 | ||||
13485 | // Find a legal type for the integer store. | |||
13486 | unsigned SizeInBits = (i + 1) * ElementSizeBytes * 8; | |||
13487 | StoreTy = EVT::getIntegerVT(Context, SizeInBits); | |||
13488 | if (TLI.isTypeLegal(StoreTy) && | |||
13489 | TLI.canMergeStoresTo(FirstStoreAS, StoreTy, DAG) && | |||
13490 | TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstStoreAS, | |||
13491 | FirstStoreAlign, &IsFastSt) && | |||
13492 | IsFastSt && | |||
13493 | TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstLoadAS, | |||
13494 | FirstLoadAlign, &IsFastLd) && | |||
13495 | IsFastLd) { | |||
13496 | LastLegalIntegerType = i + 1; | |||
13497 | DoIntegerTruncate = false; | |||
13498 | // Or check whether a truncstore and extload is legal. | |||
13499 | } else if (TLI.getTypeAction(Context, StoreTy) == | |||
13500 | TargetLowering::TypePromoteInteger) { | |||
13501 | EVT LegalizedStoredValueTy = TLI.getTypeToTransformTo(Context, StoreTy); | |||
13502 | if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) && | |||
13503 | TLI.canMergeStoresTo(FirstStoreAS, LegalizedStoredValueTy, DAG) && | |||
13504 | TLI.isLoadExtLegal(ISD::ZEXTLOAD, LegalizedStoredValueTy, | |||
13505 | StoreTy) && | |||
13506 | TLI.isLoadExtLegal(ISD::SEXTLOAD, LegalizedStoredValueTy, | |||
13507 | StoreTy) && | |||
13508 | TLI.isLoadExtLegal(ISD::EXTLOAD, LegalizedStoredValueTy, StoreTy) && | |||
13509 | TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstStoreAS, | |||
13510 | FirstStoreAlign, &IsFastSt) && | |||
13511 | IsFastSt && | |||
13512 | TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstLoadAS, | |||
13513 | FirstLoadAlign, &IsFastLd) && | |||
13514 | IsFastLd) { | |||
13515 | LastLegalIntegerType = i + 1; | |||
13516 | DoIntegerTruncate = true; | |||
13517 | } | |||
13518 | } | |||
13519 | } | |||
13520 | ||||
13521 | // Only use vector types if the vector type is larger than the integer type. | |||
13522 | // If they are the same, use integers. | |||
13523 | bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors; | |||
13524 | unsigned LastLegalType = | |||
13525 | std::max(LastLegalVectorType, LastLegalIntegerType); | |||
13526 | ||||
13527 | // We add +1 here because the LastXXX variables refer to location while | |||
13528 | // the NumElem refers to array/index size. | |||
13529 | unsigned NumElem = std::min(NumConsecutiveStores, LastConsecutiveLoad + 1); | |||
13530 | NumElem = std::min(LastLegalType, NumElem); | |||
13531 | ||||
13532 | if (NumElem < 2) { | |||
13533 | // We know that candidate stores are in order and of correct | |||
13534 | // shape. While there is no mergeable sequence from the | |||
13535 | // beginning one may start later in the sequence. The only | |||
13536 | // reason a merge of size N could have failed where another of | |||
13537 | // the same size would not have is if the alignment or either | |||
13538 | // the load or store has improved. Drop as many candidates as we | |||
13539 | // can here. | |||
13540 | unsigned NumSkip = 1; | |||
13541 | while ((NumSkip < LoadNodes.size()) && | |||
13542 | (LoadNodes[NumSkip].MemNode->getAlignment() <= FirstLoadAlign) && | |||
13543 | (StoreNodes[NumSkip].MemNode->getAlignment() <= FirstStoreAlign)) | |||
13544 | NumSkip++; | |||
13545 | StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumSkip); | |||
13546 | continue; | |||
13547 | } | |||
13548 | ||||
13549 | // Find if it is better to use vectors or integers to load and store | |||
13550 | // to memory. | |||
13551 | EVT JointMemOpVT; | |||
13552 | if (UseVectorTy) { | |||
13553 | // Find a legal type for the vector store. | |||
13554 | unsigned Elts = NumElem * NumMemElts; | |||
13555 | JointMemOpVT = EVT::getVectorVT(Context, MemVT.getScalarType(), Elts); | |||
13556 | } else { | |||
13557 | unsigned SizeInBits = NumElem * ElementSizeBytes * 8; | |||
13558 | JointMemOpVT = EVT::getIntegerVT(Context, SizeInBits); | |||
13559 | } | |||
13560 | ||||
13561 | SDLoc LoadDL(LoadNodes[0].MemNode); | |||
13562 | SDLoc StoreDL(StoreNodes[0].MemNode); | |||
13563 | ||||
13564 | // The merged loads are required to have the same incoming chain, so | |||
13565 | // using the first's chain is acceptable. | |||
13566 | ||||
13567 | SDValue NewStoreChain = getMergeStoreChains(StoreNodes, NumElem); | |||
13568 | AddToWorklist(NewStoreChain.getNode()); | |||
13569 | ||||
13570 | MachineMemOperand::Flags MMOFlags = isDereferenceable ? | |||
13571 | MachineMemOperand::MODereferenceable: | |||
13572 | MachineMemOperand::MONone; | |||
13573 | ||||
13574 | SDValue NewLoad, NewStore; | |||
13575 | if (UseVectorTy || !DoIntegerTruncate) { | |||
13576 | NewLoad = DAG.getLoad(JointMemOpVT, LoadDL, FirstLoad->getChain(), | |||
13577 | FirstLoad->getBasePtr(), | |||
13578 | FirstLoad->getPointerInfo(), FirstLoadAlign, | |||
13579 | MMOFlags); | |||
13580 | NewStore = DAG.getStore(NewStoreChain, StoreDL, NewLoad, | |||
13581 | FirstInChain->getBasePtr(), | |||
13582 | FirstInChain->getPointerInfo(), FirstStoreAlign); | |||
13583 | } else { // This must be the truncstore/extload case | |||
13584 | EVT ExtendedTy = | |||
13585 | TLI.getTypeToTransformTo(*DAG.getContext(), JointMemOpVT); | |||
13586 | NewLoad = | |||
13587 | DAG.getExtLoad(ISD::EXTLOAD, LoadDL, ExtendedTy, FirstLoad->getChain(), | |||
13588 | FirstLoad->getBasePtr(), FirstLoad->getPointerInfo(), | |||
13589 | JointMemOpVT, FirstLoadAlign, MMOFlags); | |||
13590 | NewStore = DAG.getTruncStore(NewStoreChain, StoreDL, NewLoad, | |||
13591 | FirstInChain->getBasePtr(), | |||
13592 | FirstInChain->getPointerInfo(), JointMemOpVT, | |||
13593 | FirstInChain->getAlignment(), | |||
13594 | FirstInChain->getMemOperand()->getFlags()); | |||
13595 | } | |||
13596 | ||||
13597 | // Transfer chain users from old loads to the new load. | |||
13598 | for (unsigned i = 0; i < NumElem; ++i) { | |||
13599 | LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode); | |||
13600 | DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), | |||
13601 | SDValue(NewLoad.getNode(), 1)); | |||
13602 | } | |||
13603 | ||||
13604 | // Replace the all stores with the new store. Recursively remove | |||
13605 | // corresponding value if its no longer used. | |||
13606 | for (unsigned i = 0; i < NumElem; ++i) { | |||
13607 | SDValue Val = StoreNodes[i].MemNode->getOperand(1); | |||
13608 | CombineTo(StoreNodes[i].MemNode, NewStore); | |||
13609 | if (Val.getNode()->use_empty()) | |||
13610 | recursivelyDeleteUnusedNodes(Val.getNode()); | |||
13611 | } | |||
13612 | ||||
13613 | RV = true; | |||
13614 | StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumElem); | |||
13615 | } | |||
13616 | return RV; | |||
13617 | } | |||
13618 | ||||
13619 | SDValue DAGCombiner::replaceStoreChain(StoreSDNode *ST, SDValue BetterChain) { | |||
13620 | SDLoc SL(ST); | |||
13621 | SDValue ReplStore; | |||
13622 | ||||
13623 | // Replace the chain to avoid dependency. | |||
13624 | if (ST->isTruncatingStore()) { | |||
13625 | ReplStore = DAG.getTruncStore(BetterChain, SL, ST->getValue(), | |||
13626 | ST->getBasePtr(), ST->getMemoryVT(), | |||
13627 | ST->getMemOperand()); | |||
13628 | } else { | |||
13629 | ReplStore = DAG.getStore(BetterChain, SL, ST->getValue(), ST->getBasePtr(), | |||
13630 | ST->getMemOperand()); | |||
13631 | } | |||
13632 | ||||
13633 | // Create token to keep both nodes around. | |||
13634 | SDValue Token = DAG.getNode(ISD::TokenFactor, SL, | |||
13635 | MVT::Other, ST->getChain(), ReplStore); | |||
13636 | ||||
13637 | // Make sure the new and old chains are cleaned up. | |||
13638 | AddToWorklist(Token.getNode()); | |||
13639 | ||||
13640 | // Don't add users to work list. | |||
13641 | return CombineTo(ST, Token, false); | |||
13642 | } | |||
13643 | ||||
13644 | SDValue DAGCombiner::replaceStoreOfFPConstant(StoreSDNode *ST) { | |||
13645 | SDValue Value = ST->getValue(); | |||
13646 | if (Value.getOpcode() == ISD::TargetConstantFP) | |||
13647 | return SDValue(); | |||
13648 | ||||
13649 | SDLoc DL(ST); | |||
13650 | ||||
13651 | SDValue Chain = ST->getChain(); | |||
13652 | SDValue Ptr = ST->getBasePtr(); | |||
13653 | ||||
13654 | const ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Value); | |||
13655 | ||||
13656 | // NOTE: If the original store is volatile, this transform must not increase | |||
13657 | // the number of stores. For example, on x86-32 an f64 can be stored in one | |||
13658 | // processor operation but an i64 (which is not legal) requires two. So the | |||
13659 | // transform should not be done in this case. | |||
13660 | ||||
13661 | SDValue Tmp; | |||
13662 | switch (CFP->getSimpleValueType(0).SimpleTy) { | |||
13663 | default: | |||
13664 | llvm_unreachable("Unknown FP type")::llvm::llvm_unreachable_internal("Unknown FP type", "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 13664); | |||
13665 | case MVT::f16: // We don't do this for these yet. | |||
13666 | case MVT::f80: | |||
13667 | case MVT::f128: | |||
13668 | case MVT::ppcf128: | |||
13669 | return SDValue(); | |||
13670 | case MVT::f32: | |||
13671 | if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) || | |||
13672 | TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) { | |||
13673 | ; | |||
13674 | Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF(). | |||
13675 | bitcastToAPInt().getZExtValue(), SDLoc(CFP), | |||
13676 | MVT::i32); | |||
13677 | return DAG.getStore(Chain, DL, Tmp, Ptr, ST->getMemOperand()); | |||
13678 | } | |||
13679 | ||||
13680 | return SDValue(); | |||
13681 | case MVT::f64: | |||
13682 | if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations && | |||
13683 | !ST->isVolatile()) || | |||
13684 | TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) { | |||
13685 | ; | |||
13686 | Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt(). | |||
13687 | getZExtValue(), SDLoc(CFP), MVT::i64); | |||
13688 | return DAG.getStore(Chain, DL, Tmp, | |||
13689 | Ptr, ST->getMemOperand()); | |||
13690 | } | |||
13691 | ||||
13692 | if (!ST->isVolatile() && | |||
13693 | TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) { | |||
13694 | // Many FP stores are not made apparent until after legalize, e.g. for | |||
13695 | // argument passing. Since this is so common, custom legalize the | |||
13696 | // 64-bit integer store into two 32-bit stores. | |||
13697 | uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue(); | |||
13698 | SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, SDLoc(CFP), MVT::i32); | |||
13699 | SDValue Hi = DAG.getConstant(Val >> 32, SDLoc(CFP), MVT::i32); | |||
13700 | if (DAG.getDataLayout().isBigEndian()) | |||
13701 | std::swap(Lo, Hi); | |||
13702 | ||||
13703 | unsigned Alignment = ST->getAlignment(); | |||
13704 | MachineMemOperand::Flags MMOFlags = ST->getMemOperand()->getFlags(); | |||
13705 | AAMDNodes AAInfo = ST->getAAInfo(); | |||
13706 | ||||
13707 | SDValue St0 = DAG.getStore(Chain, DL, Lo, Ptr, ST->getPointerInfo(), | |||
13708 | ST->getAlignment(), MMOFlags, AAInfo); | |||
13709 | Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr, | |||
13710 | DAG.getConstant(4, DL, Ptr.getValueType())); | |||
13711 | Alignment = MinAlign(Alignment, 4U); | |||
13712 | SDValue St1 = DAG.getStore(Chain, DL, Hi, Ptr, | |||
13713 | ST->getPointerInfo().getWithOffset(4), | |||
13714 | Alignment, MMOFlags, AAInfo); | |||
13715 | return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, | |||
13716 | St0, St1); | |||
13717 | } | |||
13718 | ||||
13719 | return SDValue(); | |||
13720 | } | |||
13721 | } | |||
13722 | ||||
13723 | SDValue DAGCombiner::visitSTORE(SDNode *N) { | |||
13724 | StoreSDNode *ST = cast<StoreSDNode>(N); | |||
13725 | SDValue Chain = ST->getChain(); | |||
13726 | SDValue Value = ST->getValue(); | |||
13727 | SDValue Ptr = ST->getBasePtr(); | |||
13728 | ||||
13729 | // If this is a store of a bit convert, store the input value if the | |||
13730 | // resultant store does not need a higher alignment than the original. | |||
13731 | if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() && | |||
13732 | ST->isUnindexed()) { | |||
13733 | EVT SVT = Value.getOperand(0).getValueType(); | |||
13734 | if (((!LegalOperations && !ST->isVolatile()) || | |||
13735 | TLI.isOperationLegalOrCustom(ISD::STORE, SVT)) && | |||
13736 | TLI.isStoreBitCastBeneficial(Value.getValueType(), SVT)) { | |||
13737 | unsigned OrigAlign = ST->getAlignment(); | |||
13738 | bool Fast = false; | |||
13739 | if (TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), SVT, | |||
13740 | ST->getAddressSpace(), OrigAlign, &Fast) && | |||
13741 | Fast) { | |||
13742 | return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0), Ptr, | |||
13743 | ST->getPointerInfo(), OrigAlign, | |||
13744 | ST->getMemOperand()->getFlags(), ST->getAAInfo()); | |||
13745 | } | |||
13746 | } | |||
13747 | } | |||
13748 | ||||
13749 | // Turn 'store undef, Ptr' -> nothing. | |||
13750 | if (Value.isUndef() && ST->isUnindexed()) | |||
13751 | return Chain; | |||
13752 | ||||
13753 | // Try to infer better alignment information than the store already has. | |||
13754 | if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) { | |||
13755 | if (unsigned Align = DAG.InferPtrAlignment(Ptr)) { | |||
13756 | if (Align > ST->getAlignment()) { | |||
13757 | SDValue NewStore = | |||
13758 | DAG.getTruncStore(Chain, SDLoc(N), Value, Ptr, ST->getPointerInfo(), | |||
13759 | ST->getMemoryVT(), Align, | |||
13760 | ST->getMemOperand()->getFlags(), ST->getAAInfo()); | |||
13761 | if (NewStore.getNode() != N) | |||
13762 | return CombineTo(ST, NewStore, true); | |||
13763 | } | |||
13764 | } | |||
13765 | } | |||
13766 | ||||
13767 | // Try transforming a pair floating point load / store ops to integer | |||
13768 | // load / store ops. | |||
13769 | if (SDValue NewST = TransformFPLoadStorePair(N)) | |||
13770 | return NewST; | |||
13771 | ||||
13772 | if (ST->isUnindexed()) { | |||
13773 | // Walk up chain skipping non-aliasing memory nodes, on this store and any | |||
13774 | // adjacent stores. | |||
13775 | if (findBetterNeighborChains(ST)) { | |||
13776 | // replaceStoreChain uses CombineTo, which handled all of the worklist | |||
13777 | // manipulation. Return the original node to not do anything else. | |||
13778 | return SDValue(ST, 0); | |||
13779 | } | |||
13780 | Chain = ST->getChain(); | |||
13781 | } | |||
13782 | ||||
13783 | // FIXME: is there such a thing as a truncating indexed store? | |||
13784 | if (ST->isTruncatingStore() && ST->isUnindexed() && | |||
13785 | Value.getValueType().isInteger()) { | |||
13786 | // See if we can simplify the input to this truncstore with knowledge that | |||
13787 | // only the low bits are being used. For example: | |||
13788 | // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8" | |||
13789 | SDValue Shorter = DAG.GetDemandedBits( | |||
13790 | Value, APInt::getLowBitsSet(Value.getScalarValueSizeInBits(), | |||
13791 | ST->getMemoryVT().getScalarSizeInBits())); | |||
13792 | AddToWorklist(Value.getNode()); | |||
13793 | if (Shorter.getNode()) | |||
13794 | return DAG.getTruncStore(Chain, SDLoc(N), Shorter, | |||
13795 | Ptr, ST->getMemoryVT(), ST->getMemOperand()); | |||
13796 | ||||
13797 | // Otherwise, see if we can simplify the operation with | |||
13798 | // SimplifyDemandedBits, which only works if the value has a single use. | |||
13799 | if (SimplifyDemandedBits( | |||
13800 | Value, | |||
13801 | APInt::getLowBitsSet(Value.getScalarValueSizeInBits(), | |||
13802 | ST->getMemoryVT().getScalarSizeInBits()))) { | |||
13803 | // Re-visit the store if anything changed and the store hasn't been merged | |||
13804 | // with another node (N is deleted) SimplifyDemandedBits will add Value's | |||
13805 | // node back to the worklist if necessary, but we also need to re-visit | |||
13806 | // the Store node itself. | |||
13807 | if (N->getOpcode() != ISD::DELETED_NODE) | |||
13808 | AddToWorklist(N); | |||
13809 | return SDValue(N, 0); | |||
13810 | } | |||
13811 | } | |||
13812 | ||||
13813 | // If this is a load followed by a store to the same location, then the store | |||
13814 | // is dead/noop. | |||
13815 | if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) { | |||
13816 | if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() && | |||
13817 | ST->isUnindexed() && !ST->isVolatile() && | |||
13818 | // There can't be any side effects between the load and store, such as | |||
13819 | // a call or store. | |||
13820 | Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) { | |||
13821 | // The store is dead, remove it. | |||
13822 | return Chain; | |||
13823 | } | |||
13824 | } | |||
13825 | ||||
13826 | if (StoreSDNode *ST1 = dyn_cast<StoreSDNode>(Chain)) { | |||
13827 | if (ST->isUnindexed() && !ST->isVolatile() && ST1->isUnindexed() && | |||
13828 | !ST1->isVolatile() && ST1->getBasePtr() == Ptr && | |||
13829 | ST->getMemoryVT() == ST1->getMemoryVT()) { | |||
13830 | // If this is a store followed by a store with the same value to the same | |||
13831 | // location, then the store is dead/noop. | |||
13832 | if (ST1->getValue() == Value) { | |||
13833 | // The store is dead, remove it. | |||
13834 | return Chain; | |||
13835 | } | |||
13836 | ||||
13837 | // If this is a store who's preceeding store to the same location | |||
13838 | // and no one other node is chained to that store we can effectively | |||
13839 | // drop the store. Do not remove stores to undef as they may be used as | |||
13840 | // data sinks. | |||
13841 | if (OptLevel != CodeGenOpt::None && ST1->hasOneUse() && | |||
13842 | !ST1->getBasePtr().isUndef()) { | |||
13843 | // ST1 is fully overwritten and can be elided. Combine with it's chain | |||
13844 | // value. | |||
13845 | CombineTo(ST1, ST1->getChain()); | |||
13846 | return SDValue(); | |||
13847 | } | |||
13848 | } | |||
13849 | } | |||
13850 | ||||
13851 | // If this is an FP_ROUND or TRUNC followed by a store, fold this into a | |||
13852 | // truncating store. We can do this even if this is already a truncstore. | |||
13853 | if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE) | |||
13854 | && Value.getNode()->hasOneUse() && ST->isUnindexed() && | |||
13855 | TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(), | |||
13856 | ST->getMemoryVT())) { | |||
13857 | return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0), | |||
13858 | Ptr, ST->getMemoryVT(), ST->getMemOperand()); | |||
13859 | } | |||
13860 | ||||
13861 | // Always perform this optimization before types are legal. If the target | |||
13862 | // prefers, also try this after legalization to catch stores that were created | |||
13863 | // by intrinsics or other nodes. | |||
13864 | if (!LegalTypes || (TLI.mergeStoresAfterLegalization())) { | |||
13865 | while (true) { | |||
13866 | // There can be multiple store sequences on the same chain. | |||
13867 | // Keep trying to merge store sequences until we are unable to do so | |||
13868 | // or until we merge the last store on the chain. | |||
13869 | bool Changed = MergeConsecutiveStores(ST); | |||
13870 | if (!Changed) break; | |||
13871 | // Return N as merge only uses CombineTo and no worklist clean | |||
13872 | // up is necessary. | |||
13873 | if (N->getOpcode() == ISD::DELETED_NODE || !isa<StoreSDNode>(N)) | |||
13874 | return SDValue(N, 0); | |||
13875 | } | |||
13876 | } | |||
13877 | ||||
13878 | // Try transforming N to an indexed store. | |||
13879 | if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N)) | |||
13880 | return SDValue(N, 0); | |||
13881 | ||||
13882 | // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr' | |||
13883 | // | |||
13884 | // Make sure to do this only after attempting to merge stores in order to | |||
13885 | // avoid changing the types of some subset of stores due to visit order, | |||
13886 | // preventing their merging. | |||
13887 | if (isa<ConstantFPSDNode>(ST->getValue())) { | |||
13888 | if (SDValue NewSt = replaceStoreOfFPConstant(ST)) | |||
13889 | return NewSt; | |||
13890 | } | |||
13891 | ||||
13892 | if (SDValue NewSt = splitMergedValStore(ST)) | |||
13893 | return NewSt; | |||
13894 | ||||
13895 | return ReduceLoadOpStoreWidth(N); | |||
13896 | } | |||
13897 | ||||
13898 | /// For the instruction sequence of store below, F and I values | |||
13899 | /// are bundled together as an i64 value before being stored into memory. | |||
13900 | /// Sometimes it is more efficent to generate separate stores for F and I, | |||
13901 | /// which can remove the bitwise instructions or sink them to colder places. | |||
13902 | /// | |||
13903 | /// (store (or (zext (bitcast F to i32) to i64), | |||
13904 | /// (shl (zext I to i64), 32)), addr) --> | |||
13905 | /// (store F, addr) and (store I, addr+4) | |||
13906 | /// | |||
13907 | /// Similarly, splitting for other merged store can also be beneficial, like: | |||
13908 | /// For pair of {i32, i32}, i64 store --> two i32 stores. | |||
13909 | /// For pair of {i32, i16}, i64 store --> two i32 stores. | |||
13910 | /// For pair of {i16, i16}, i32 store --> two i16 stores. | |||
13911 | /// For pair of {i16, i8}, i32 store --> two i16 stores. | |||
13912 | /// For pair of {i8, i8}, i16 store --> two i8 stores. | |||
13913 | /// | |||
13914 | /// We allow each target to determine specifically which kind of splitting is | |||
13915 | /// supported. | |||
13916 | /// | |||
13917 | /// The store patterns are commonly seen from the simple code snippet below | |||
13918 | /// if only std::make_pair(...) is sroa transformed before inlined into hoo. | |||
13919 | /// void goo(const std::pair<int, float> &); | |||
13920 | /// hoo() { | |||
13921 | /// ... | |||
13922 | /// goo(std::make_pair(tmp, ftmp)); | |||
13923 | /// ... | |||
13924 | /// } | |||
13925 | /// | |||
13926 | SDValue DAGCombiner::splitMergedValStore(StoreSDNode *ST) { | |||
13927 | if (OptLevel == CodeGenOpt::None) | |||
13928 | return SDValue(); | |||
13929 | ||||
13930 | SDValue Val = ST->getValue(); | |||
13931 | SDLoc DL(ST); | |||
13932 | ||||
13933 | // Match OR operand. | |||
13934 | if (!Val.getValueType().isScalarInteger() || Val.getOpcode() != ISD::OR) | |||
13935 | return SDValue(); | |||
13936 | ||||
13937 | // Match SHL operand and get Lower and Higher parts of Val. | |||
13938 | SDValue Op1 = Val.getOperand(0); | |||
13939 | SDValue Op2 = Val.getOperand(1); | |||
13940 | SDValue Lo, Hi; | |||
13941 | if (Op1.getOpcode() != ISD::SHL) { | |||
13942 | std::swap(Op1, Op2); | |||
13943 | if (Op1.getOpcode() != ISD::SHL) | |||
13944 | return SDValue(); | |||
13945 | } | |||
13946 | Lo = Op2; | |||
13947 | Hi = Op1.getOperand(0); | |||
13948 | if (!Op1.hasOneUse()) | |||
13949 | return SDValue(); | |||
13950 | ||||
13951 | // Match shift amount to HalfValBitSize. | |||
13952 | unsigned HalfValBitSize = Val.getValueSizeInBits() / 2; | |||
13953 | ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(Op1.getOperand(1)); | |||
13954 | if (!ShAmt || ShAmt->getAPIntValue() != HalfValBitSize) | |||
13955 | return SDValue(); | |||
13956 | ||||
13957 | // Lo and Hi are zero-extended from int with size less equal than 32 | |||
13958 | // to i64. | |||
13959 | if (Lo.getOpcode() != ISD::ZERO_EXTEND || !Lo.hasOneUse() || | |||
13960 | !Lo.getOperand(0).getValueType().isScalarInteger() || | |||
13961 | Lo.getOperand(0).getValueSizeInBits() > HalfValBitSize || | |||
13962 | Hi.getOpcode() != ISD::ZERO_EXTEND || !Hi.hasOneUse() || | |||
13963 | !Hi.getOperand(0).getValueType().isScalarInteger() || | |||
13964 | Hi.getOperand(0).getValueSizeInBits() > HalfValBitSize) | |||
13965 | return SDValue(); | |||
13966 | ||||
13967 | // Use the EVT of low and high parts before bitcast as the input | |||
13968 | // of target query. | |||
13969 | EVT LowTy = (Lo.getOperand(0).getOpcode() == ISD::BITCAST) | |||
13970 | ? Lo.getOperand(0).getValueType() | |||
13971 | : Lo.getValueType(); | |||
13972 | EVT HighTy = (Hi.getOperand(0).getOpcode() == ISD::BITCAST) | |||
13973 | ? Hi.getOperand(0).getValueType() | |||
13974 | : Hi.getValueType(); | |||
13975 | if (!TLI.isMultiStoresCheaperThanBitsMerge(LowTy, HighTy)) | |||
13976 | return SDValue(); | |||
13977 | ||||
13978 | // Start to split store. | |||
13979 | unsigned Alignment = ST->getAlignment(); | |||
13980 | MachineMemOperand::Flags MMOFlags = ST->getMemOperand()->getFlags(); | |||
13981 | AAMDNodes AAInfo = ST->getAAInfo(); | |||
13982 | ||||
13983 | // Change the sizes of Lo and Hi's value types to HalfValBitSize. | |||
13984 | EVT VT = EVT::getIntegerVT(*DAG.getContext(), HalfValBitSize); | |||
13985 | Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Lo.getOperand(0)); | |||
13986 | Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Hi.getOperand(0)); | |||
13987 | ||||
13988 | SDValue Chain = ST->getChain(); | |||
13989 | SDValue Ptr = ST->getBasePtr(); | |||
13990 | // Lower value store. | |||
13991 | SDValue St0 = DAG.getStore(Chain, DL, Lo, Ptr, ST->getPointerInfo(), | |||
13992 | ST->getAlignment(), MMOFlags, AAInfo); | |||
13993 | Ptr = | |||
13994 | DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr, | |||
13995 | DAG.getConstant(HalfValBitSize / 8, DL, Ptr.getValueType())); | |||
13996 | // Higher value store. | |||
13997 | SDValue St1 = | |||
13998 | DAG.getStore(St0, DL, Hi, Ptr, | |||
13999 | ST->getPointerInfo().getWithOffset(HalfValBitSize / 8), | |||
14000 | Alignment / 2, MMOFlags, AAInfo); | |||
14001 | return St1; | |||
14002 | } | |||
14003 | ||||
14004 | /// Convert a disguised subvector insertion into a shuffle: | |||
14005 | /// insert_vector_elt V, (bitcast X from vector type), IdxC --> | |||
14006 | /// bitcast(shuffle (bitcast V), (extended X), Mask) | |||
14007 | /// Note: We do not use an insert_subvector node because that requires a legal | |||
14008 | /// subvector type. | |||
14009 | SDValue DAGCombiner::combineInsertEltToShuffle(SDNode *N, unsigned InsIndex) { | |||
14010 | SDValue InsertVal = N->getOperand(1); | |||
14011 | if (InsertVal.getOpcode() != ISD::BITCAST || !InsertVal.hasOneUse() || | |||
14012 | !InsertVal.getOperand(0).getValueType().isVector()) | |||
14013 | return SDValue(); | |||
14014 | ||||
14015 | SDValue SubVec = InsertVal.getOperand(0); | |||
14016 | SDValue DestVec = N->getOperand(0); | |||
14017 | EVT SubVecVT = SubVec.getValueType(); | |||
14018 | EVT VT = DestVec.getValueType(); | |||
14019 | unsigned NumSrcElts = SubVecVT.getVectorNumElements(); | |||
14020 | unsigned ExtendRatio = VT.getSizeInBits() / SubVecVT.getSizeInBits(); | |||
14021 | unsigned NumMaskVals = ExtendRatio * NumSrcElts; | |||
14022 | ||||
14023 | // Step 1: Create a shuffle mask that implements this insert operation. The | |||
14024 | // vector that we are inserting into will be operand 0 of the shuffle, so | |||
14025 | // those elements are just 'i'. The inserted subvector is in the first | |||
14026 | // positions of operand 1 of the shuffle. Example: | |||
14027 | // insert v4i32 V, (v2i16 X), 2 --> shuffle v8i16 V', X', {0,1,2,3,8,9,6,7} | |||
14028 | SmallVector<int, 16> Mask(NumMaskVals); | |||
14029 | for (unsigned i = 0; i != NumMaskVals; ++i) { | |||
14030 | if (i / NumSrcElts == InsIndex) | |||
14031 | Mask[i] = (i % NumSrcElts) + NumMaskVals; | |||
14032 | else | |||
14033 | Mask[i] = i; | |||
14034 | } | |||
14035 | ||||
14036 | // Bail out if the target can not handle the shuffle we want to create. | |||
14037 | EVT SubVecEltVT = SubVecVT.getVectorElementType(); | |||
14038 | EVT ShufVT = EVT::getVectorVT(*DAG.getContext(), SubVecEltVT, NumMaskVals); | |||
14039 | if (!TLI.isShuffleMaskLegal(Mask, ShufVT)) | |||
14040 | return SDValue(); | |||
14041 | ||||
14042 | // Step 2: Create a wide vector from the inserted source vector by appending | |||
14043 | // undefined elements. This is the same size as our destination vector. | |||
14044 | SDLoc DL(N); | |||
14045 | SmallVector<SDValue, 8> ConcatOps(ExtendRatio, DAG.getUNDEF(SubVecVT)); | |||
14046 | ConcatOps[0] = SubVec; | |||
14047 | SDValue PaddedSubV = DAG.getNode(ISD::CONCAT_VECTORS, DL, ShufVT, ConcatOps); | |||
14048 | ||||
14049 | // Step 3: Shuffle in the padded subvector. | |||
14050 | SDValue DestVecBC = DAG.getBitcast(ShufVT, DestVec); | |||
14051 | SDValue Shuf = DAG.getVectorShuffle(ShufVT, DL, DestVecBC, PaddedSubV, Mask); | |||
14052 | AddToWorklist(PaddedSubV.getNode()); | |||
14053 | AddToWorklist(DestVecBC.getNode()); | |||
14054 | AddToWorklist(Shuf.getNode()); | |||
14055 | return DAG.getBitcast(VT, Shuf); | |||
14056 | } | |||
14057 | ||||
14058 | SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) { | |||
14059 | SDValue InVec = N->getOperand(0); | |||
14060 | SDValue InVal = N->getOperand(1); | |||
14061 | SDValue EltNo = N->getOperand(2); | |||
14062 | SDLoc DL(N); | |||
14063 | ||||
14064 | // If the inserted element is an UNDEF, just use the input vector. | |||
14065 | if (InVal.isUndef()) | |||
14066 | return InVec; | |||
14067 | ||||
14068 | EVT VT = InVec.getValueType(); | |||
14069 | ||||
14070 | // Remove redundant insertions: | |||
14071 | // (insert_vector_elt x (extract_vector_elt x idx) idx) -> x | |||
14072 | if (InVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT && | |||
14073 | InVec == InVal.getOperand(0) && EltNo == InVal.getOperand(1)) | |||
14074 | return InVec; | |||
14075 | ||||
14076 | // We must know which element is being inserted for folds below here. | |||
14077 | auto *IndexC = dyn_cast<ConstantSDNode>(EltNo); | |||
14078 | if (!IndexC) | |||
14079 | return SDValue(); | |||
14080 | unsigned Elt = IndexC->getZExtValue(); | |||
14081 | ||||
14082 | if (SDValue Shuf = combineInsertEltToShuffle(N, Elt)) | |||
14083 | return Shuf; | |||
14084 | ||||
14085 | // Canonicalize insert_vector_elt dag nodes. | |||
14086 | // Example: | |||
14087 | // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1) | |||
14088 | // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0) | |||
14089 | // | |||
14090 | // Do this only if the child insert_vector node has one use; also | |||
14091 | // do this only if indices are both constants and Idx1 < Idx0. | |||
14092 | if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse() | |||
14093 | && isa<ConstantSDNode>(InVec.getOperand(2))) { | |||
14094 | unsigned OtherElt = InVec.getConstantOperandVal(2); | |||
14095 | if (Elt < OtherElt) { | |||
14096 | // Swap nodes. | |||
14097 | SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, | |||
14098 | InVec.getOperand(0), InVal, EltNo); | |||
14099 | AddToWorklist(NewOp.getNode()); | |||
14100 | return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()), | |||
14101 | VT, NewOp, InVec.getOperand(1), InVec.getOperand(2)); | |||
14102 | } | |||
14103 | } | |||
14104 | ||||
14105 | // If we can't generate a legal BUILD_VECTOR, exit | |||
14106 | if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) | |||
14107 | return SDValue(); | |||
14108 | ||||
14109 | // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially | |||
14110 | // be converted to a BUILD_VECTOR). Fill in the Ops vector with the | |||
14111 | // vector elements. | |||
14112 | SmallVector<SDValue, 8> Ops; | |||
14113 | // Do not combine these two vectors if the output vector will not replace | |||
14114 | // the input vector. | |||
14115 | if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) { | |||
14116 | Ops.append(InVec.getNode()->op_begin(), | |||
14117 | InVec.getNode()->op_end()); | |||
14118 | } else if (InVec.isUndef()) { | |||
14119 | unsigned NElts = VT.getVectorNumElements(); | |||
14120 | Ops.append(NElts, DAG.getUNDEF(InVal.getValueType())); | |||
14121 | } else { | |||
14122 | return SDValue(); | |||
14123 | } | |||
14124 | ||||
14125 | // Insert the element | |||
14126 | if (Elt < Ops.size()) { | |||
14127 | // All the operands of BUILD_VECTOR must have the same type; | |||
14128 | // we enforce that here. | |||
14129 | EVT OpVT = Ops[0].getValueType(); | |||
14130 | Ops[Elt] = OpVT.isInteger() ? DAG.getAnyExtOrTrunc(InVal, DL, OpVT) : InVal; | |||
14131 | } | |||
14132 | ||||
14133 | // Return the new vector | |||
14134 | return DAG.getBuildVector(VT, DL, Ops); | |||
14135 | } | |||
14136 | ||||
14137 | SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad( | |||
14138 | SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) { | |||
14139 | assert(!OriginalLoad->isVolatile())(static_cast <bool> (!OriginalLoad->isVolatile()) ? void (0) : __assert_fail ("!OriginalLoad->isVolatile()", "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 14139, __extension__ __PRETTY_FUNCTION__)); | |||
14140 | ||||
14141 | EVT ResultVT = EVE->getValueType(0); | |||
14142 | EVT VecEltVT = InVecVT.getVectorElementType(); | |||
14143 | unsigned Align = OriginalLoad->getAlignment(); | |||
14144 | unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment( | |||
14145 | VecEltVT.getTypeForEVT(*DAG.getContext())); | |||
14146 | ||||
14147 | if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT)) | |||
14148 | return SDValue(); | |||
14149 | ||||
14150 | ISD::LoadExtType ExtTy = ResultVT.bitsGT(VecEltVT) ? | |||
14151 | ISD::NON_EXTLOAD : ISD::EXTLOAD; | |||
14152 | if (!TLI.shouldReduceLoadWidth(OriginalLoad, ExtTy, VecEltVT)) | |||
14153 | return SDValue(); | |||
14154 | ||||
14155 | Align = NewAlign; | |||
14156 | ||||
14157 | SDValue NewPtr = OriginalLoad->getBasePtr(); | |||
14158 | SDValue Offset; | |||
14159 | EVT PtrType = NewPtr.getValueType(); | |||
14160 | MachinePointerInfo MPI; | |||
14161 | SDLoc DL(EVE); | |||
14162 | if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) { | |||
14163 | int Elt = ConstEltNo->getZExtValue(); | |||
14164 | unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8; | |||
14165 | Offset = DAG.getConstant(PtrOff, DL, PtrType); | |||
14166 | MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff); | |||
14167 | } else { | |||
14168 | Offset = DAG.getZExtOrTrunc(EltNo, DL, PtrType); | |||
14169 | Offset = DAG.getNode( | |||
14170 | ISD::MUL, DL, PtrType, Offset, | |||
14171 | DAG.getConstant(VecEltVT.getStoreSize(), DL, PtrType)); | |||
14172 | MPI = OriginalLoad->getPointerInfo(); | |||
14173 | } | |||
14174 | NewPtr = DAG.getNode(ISD::ADD, DL, PtrType, NewPtr, Offset); | |||
14175 | ||||
14176 | // The replacement we need to do here is a little tricky: we need to | |||
14177 | // replace an extractelement of a load with a load. | |||
14178 | // Use ReplaceAllUsesOfValuesWith to do the replacement. | |||
14179 | // Note that this replacement assumes that the extractvalue is the only | |||
14180 | // use of the load; that's okay because we don't want to perform this | |||
14181 | // transformation in other cases anyway. | |||
14182 | SDValue Load; | |||
14183 | SDValue Chain; | |||
14184 | if (ResultVT.bitsGT(VecEltVT)) { | |||
14185 | // If the result type of vextract is wider than the load, then issue an | |||
14186 | // extending load instead. | |||
14187 | ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, ResultVT, | |||
14188 | VecEltVT) | |||
14189 | ? ISD::ZEXTLOAD | |||
14190 | : ISD::EXTLOAD; | |||
14191 | Load = DAG.getExtLoad(ExtType, SDLoc(EVE), ResultVT, | |||
14192 | OriginalLoad->getChain(), NewPtr, MPI, VecEltVT, | |||
14193 | Align, OriginalLoad->getMemOperand()->getFlags(), | |||
14194 | OriginalLoad->getAAInfo()); | |||
14195 | Chain = Load.getValue(1); | |||
14196 | } else { | |||
14197 | Load = DAG.getLoad(VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr, | |||
14198 | MPI, Align, OriginalLoad->getMemOperand()->getFlags(), | |||
14199 | OriginalLoad->getAAInfo()); | |||
14200 | Chain = Load.getValue(1); | |||
14201 | if (ResultVT.bitsLT(VecEltVT)) | |||
14202 | Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load); | |||
14203 | else | |||
14204 | Load = DAG.getBitcast(ResultVT, Load); | |||
14205 | } | |||
14206 | WorklistRemover DeadNodes(*this); | |||
14207 | SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) }; | |||
14208 | SDValue To[] = { Load, Chain }; | |||
14209 | DAG.ReplaceAllUsesOfValuesWith(From, To, 2); | |||
14210 | // Since we're explicitly calling ReplaceAllUses, add the new node to the | |||
14211 | // worklist explicitly as well. | |||
14212 | AddToWorklist(Load.getNode()); | |||
14213 | AddUsersToWorklist(Load.getNode()); // Add users too | |||
14214 | // Make sure to revisit this node to clean it up; it will usually be dead. | |||
14215 | AddToWorklist(EVE); | |||
14216 | ++OpsNarrowed; | |||
14217 | return SDValue(EVE, 0); | |||
14218 | } | |||
14219 | ||||
14220 | SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) { | |||
14221 | // (vextract (scalar_to_vector val, 0) -> val | |||
14222 | SDValue InVec = N->getOperand(0); | |||
14223 | EVT VT = InVec.getValueType(); | |||
14224 | EVT NVT = N->getValueType(0); | |||
14225 | ||||
14226 | if (InVec.isUndef()) | |||
14227 | return DAG.getUNDEF(NVT); | |||
14228 | ||||
14229 | if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) { | |||
14230 | // Check if the result type doesn't match the inserted element type. A | |||
14231 | // SCALAR_TO_VECTOR may truncate the inserted element and the | |||
14232 | // EXTRACT_VECTOR_ELT may widen the extracted vector. | |||
14233 | SDValue InOp = InVec.getOperand(0); | |||
14234 | if (InOp.getValueType() != NVT) { | |||
14235 | assert(InOp.getValueType().isInteger() && NVT.isInteger())(static_cast <bool> (InOp.getValueType().isInteger() && NVT.isInteger()) ? void (0) : __assert_fail ("InOp.getValueType().isInteger() && NVT.isInteger()" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 14235, __extension__ __PRETTY_FUNCTION__)); | |||
14236 | return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT); | |||
14237 | } | |||
14238 | return InOp; | |||
14239 | } | |||
14240 | ||||
14241 | SDValue EltNo = N->getOperand(1); | |||
14242 | ConstantSDNode *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo); | |||
14243 | ||||
14244 | // extract_vector_elt of out-of-bounds element -> UNDEF | |||
14245 | if (ConstEltNo && ConstEltNo->getAPIntValue().uge(VT.getVectorNumElements())) | |||
14246 | return DAG.getUNDEF(NVT); | |||
14247 | ||||
14248 | // extract_vector_elt (build_vector x, y), 1 -> y | |||
14249 | if (ConstEltNo && | |||
14250 | InVec.getOpcode() == ISD::BUILD_VECTOR && | |||
14251 | TLI.isTypeLegal(VT) && | |||
14252 | (InVec.hasOneUse() || | |||
14253 | TLI.aggressivelyPreferBuildVectorSources(VT))) { | |||
14254 | SDValue Elt = InVec.getOperand(ConstEltNo->getZExtValue()); | |||
14255 | EVT InEltVT = Elt.getValueType(); | |||
14256 | ||||
14257 | // Sometimes build_vector's scalar input types do not match result type. | |||
14258 | if (NVT == InEltVT) | |||
14259 | return Elt; | |||
14260 | ||||
14261 | // TODO: It may be useful to truncate if free if the build_vector implicitly | |||
14262 | // converts. | |||
14263 | } | |||
14264 | ||||
14265 | // extract_vector_elt (v2i32 (bitcast i64:x)), EltTrunc -> i32 (trunc i64:x) | |||
14266 | bool isLE = DAG.getDataLayout().isLittleEndian(); | |||
14267 | unsigned EltTrunc = isLE ? 0 : VT.getVectorNumElements() - 1; | |||
14268 | if (ConstEltNo && InVec.getOpcode() == ISD::BITCAST && InVec.hasOneUse() && | |||
14269 | ConstEltNo->getZExtValue() == EltTrunc && VT.isInteger()) { | |||
14270 | SDValue BCSrc = InVec.getOperand(0); | |||
14271 | if (BCSrc.getValueType().isScalarInteger()) | |||
14272 | return DAG.getNode(ISD::TRUNCATE, SDLoc(N), NVT, BCSrc); | |||
14273 | } | |||
14274 | ||||
14275 | // extract_vector_elt (insert_vector_elt vec, val, idx), idx) -> val | |||
14276 | // | |||
14277 | // This only really matters if the index is non-constant since other combines | |||
14278 | // on the constant elements already work. | |||
14279 | if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && | |||
14280 | EltNo == InVec.getOperand(2)) { | |||
14281 | SDValue Elt = InVec.getOperand(1); | |||
14282 | return VT.isInteger() ? DAG.getAnyExtOrTrunc(Elt, SDLoc(N), NVT) : Elt; | |||
14283 | } | |||
14284 | ||||
14285 | // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT. | |||
14286 | // We only perform this optimization before the op legalization phase because | |||
14287 | // we may introduce new vector instructions which are not backed by TD | |||
14288 | // patterns. For example on AVX, extracting elements from a wide vector | |||
14289 | // without using extract_subvector. However, if we can find an underlying | |||
14290 | // scalar value, then we can always use that. | |||
14291 | if (ConstEltNo && InVec.getOpcode() == ISD::VECTOR_SHUFFLE) { | |||
14292 | int NumElem = VT.getVectorNumElements(); | |||
14293 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec); | |||
14294 | // Find the new index to extract from. | |||
14295 | int OrigElt = SVOp->getMaskElt(ConstEltNo->getZExtValue()); | |||
14296 | ||||
14297 | // Extracting an undef index is undef. | |||
14298 | if (OrigElt == -1) | |||
14299 | return DAG.getUNDEF(NVT); | |||
14300 | ||||
14301 | // Select the right vector half to extract from. | |||
14302 | SDValue SVInVec; | |||
14303 | if (OrigElt < NumElem) { | |||
14304 | SVInVec = InVec->getOperand(0); | |||
14305 | } else { | |||
14306 | SVInVec = InVec->getOperand(1); | |||
14307 | OrigElt -= NumElem; | |||
14308 | } | |||
14309 | ||||
14310 | if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) { | |||
14311 | SDValue InOp = SVInVec.getOperand(OrigElt); | |||
14312 | if (InOp.getValueType() != NVT) { | |||
14313 | assert(InOp.getValueType().isInteger() && NVT.isInteger())(static_cast <bool> (InOp.getValueType().isInteger() && NVT.isInteger()) ? void (0) : __assert_fail ("InOp.getValueType().isInteger() && NVT.isInteger()" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 14313, __extension__ __PRETTY_FUNCTION__)); | |||
14314 | InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT); | |||
14315 | } | |||
14316 | ||||
14317 | return InOp; | |||
14318 | } | |||
14319 | ||||
14320 | // FIXME: We should handle recursing on other vector shuffles and | |||
14321 | // scalar_to_vector here as well. | |||
14322 | ||||
14323 | if (!LegalOperations || | |||
14324 | // FIXME: Should really be just isOperationLegalOrCustom. | |||
14325 | TLI.isOperationLegal(ISD::EXTRACT_VECTOR_ELT, VT) || | |||
14326 | TLI.isOperationExpand(ISD::VECTOR_SHUFFLE, VT)) { | |||
14327 | EVT IndexTy = TLI.getVectorIdxTy(DAG.getDataLayout()); | |||
14328 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT, SVInVec, | |||
14329 | DAG.getConstant(OrigElt, SDLoc(SVOp), IndexTy)); | |||
14330 | } | |||
14331 | } | |||
14332 | ||||
14333 | bool BCNumEltsChanged = false; | |||
14334 | EVT ExtVT = VT.getVectorElementType(); | |||
14335 | EVT LVT = ExtVT; | |||
14336 | ||||
14337 | // If the result of load has to be truncated, then it's not necessarily | |||
14338 | // profitable. | |||
14339 | if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT)) | |||
14340 | return SDValue(); | |||
14341 | ||||
14342 | if (InVec.getOpcode() == ISD::BITCAST) { | |||
14343 | // Don't duplicate a load with other uses. | |||
14344 | if (!InVec.hasOneUse()) | |||
14345 | return SDValue(); | |||
14346 | ||||
14347 | EVT BCVT = InVec.getOperand(0).getValueType(); | |||
14348 | if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType())) | |||
14349 | return SDValue(); | |||
14350 | if (VT.getVectorNumElements() != BCVT.getVectorNumElements()) | |||
14351 | BCNumEltsChanged = true; | |||
14352 | InVec = InVec.getOperand(0); | |||
14353 | ExtVT = BCVT.getVectorElementType(); | |||
14354 | } | |||
14355 | ||||
14356 | // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size) | |||
14357 | if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() && | |||
14358 | ISD::isNormalLoad(InVec.getNode()) && | |||
14359 | !N->getOperand(1)->hasPredecessor(InVec.getNode())) { | |||
14360 | SDValue Index = N->getOperand(1); | |||
14361 | if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec)) { | |||
14362 | if (!OrigLoad->isVolatile()) { | |||
14363 | return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index, | |||
14364 | OrigLoad); | |||
14365 | } | |||
14366 | } | |||
14367 | } | |||
14368 | ||||
14369 | // Perform only after legalization to ensure build_vector / vector_shuffle | |||
14370 | // optimizations have already been done. | |||
14371 | if (!LegalOperations) return SDValue(); | |||
14372 | ||||
14373 | // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size) | |||
14374 | // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size) | |||
14375 | // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr) | |||
14376 | ||||
14377 | if (ConstEltNo) { | |||
14378 | int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue(); | |||
14379 | ||||
14380 | LoadSDNode *LN0 = nullptr; | |||
14381 | const ShuffleVectorSDNode *SVN = nullptr; | |||
14382 | if (ISD::isNormalLoad(InVec.getNode())) { | |||
14383 | LN0 = cast<LoadSDNode>(InVec); | |||
14384 | } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR && | |||
14385 | InVec.getOperand(0).getValueType() == ExtVT && | |||
14386 | ISD::isNormalLoad(InVec.getOperand(0).getNode())) { | |||
14387 | // Don't duplicate a load with other uses. | |||
14388 | if (!InVec.hasOneUse()) | |||
14389 | return SDValue(); | |||
14390 | ||||
14391 | LN0 = cast<LoadSDNode>(InVec.getOperand(0)); | |||
14392 | } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) { | |||
14393 | // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1) | |||
14394 | // => | |||
14395 | // (load $addr+1*size) | |||
14396 | ||||
14397 | // Don't duplicate a load with other uses. | |||
14398 | if (!InVec.hasOneUse()) | |||
14399 | return SDValue(); | |||
14400 | ||||
14401 | // If the bit convert changed the number of elements, it is unsafe | |||
14402 | // to examine the mask. | |||
14403 | if (BCNumEltsChanged) | |||
14404 | return SDValue(); | |||
14405 | ||||
14406 | // Select the input vector, guarding against out of range extract vector. | |||
14407 | unsigned NumElems = VT.getVectorNumElements(); | |||
14408 | int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt); | |||
14409 | InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1); | |||
14410 | ||||
14411 | if (InVec.getOpcode() == ISD::BITCAST) { | |||
14412 | // Don't duplicate a load with other uses. | |||
14413 | if (!InVec.hasOneUse()) | |||
14414 | return SDValue(); | |||
14415 | ||||
14416 | InVec = InVec.getOperand(0); | |||
14417 | } | |||
14418 | if (ISD::isNormalLoad(InVec.getNode())) { | |||
14419 | LN0 = cast<LoadSDNode>(InVec); | |||
14420 | Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems; | |||
14421 | EltNo = DAG.getConstant(Elt, SDLoc(EltNo), EltNo.getValueType()); | |||
14422 | } | |||
14423 | } | |||
14424 | ||||
14425 | // Make sure we found a non-volatile load and the extractelement is | |||
14426 | // the only use. | |||
14427 | if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile()) | |||
14428 | return SDValue(); | |||
14429 | ||||
14430 | // If Idx was -1 above, Elt is going to be -1, so just return undef. | |||
14431 | if (Elt == -1) | |||
14432 | return DAG.getUNDEF(LVT); | |||
14433 | ||||
14434 | return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0); | |||
14435 | } | |||
14436 | ||||
14437 | return SDValue(); | |||
14438 | } | |||
14439 | ||||
14440 | // Simplify (build_vec (ext )) to (bitcast (build_vec )) | |||
14441 | SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) { | |||
14442 | // We perform this optimization post type-legalization because | |||
14443 | // the type-legalizer often scalarizes integer-promoted vectors. | |||
14444 | // Performing this optimization before may create bit-casts which | |||
14445 | // will be type-legalized to complex code sequences. | |||
14446 | // We perform this optimization only before the operation legalizer because we | |||
14447 | // may introduce illegal operations. | |||
14448 | if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes) | |||
14449 | return SDValue(); | |||
14450 | ||||
14451 | unsigned NumInScalars = N->getNumOperands(); | |||
14452 | SDLoc DL(N); | |||
14453 | EVT VT = N->getValueType(0); | |||
14454 | ||||
14455 | // Check to see if this is a BUILD_VECTOR of a bunch of values | |||
14456 | // which come from any_extend or zero_extend nodes. If so, we can create | |||
14457 | // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR | |||
14458 | // optimizations. We do not handle sign-extend because we can't fill the sign | |||
14459 | // using shuffles. | |||
14460 | EVT SourceType = MVT::Other; | |||
14461 | bool AllAnyExt = true; | |||
14462 | ||||
14463 | for (unsigned i = 0; i != NumInScalars; ++i) { | |||
14464 | SDValue In = N->getOperand(i); | |||
14465 | // Ignore undef inputs. | |||
14466 | if (In.isUndef()) continue; | |||
14467 | ||||
14468 | bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND; | |||
14469 | bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND; | |||
14470 | ||||
14471 | // Abort if the element is not an extension. | |||
14472 | if (!ZeroExt && !AnyExt) { | |||
14473 | SourceType = MVT::Other; | |||
14474 | break; | |||
14475 | } | |||
14476 | ||||
14477 | // The input is a ZeroExt or AnyExt. Check the original type. | |||
14478 | EVT InTy = In.getOperand(0).getValueType(); | |||
14479 | ||||
14480 | // Check that all of the widened source types are the same. | |||
14481 | if (SourceType == MVT::Other) | |||
14482 | // First time. | |||
14483 | SourceType = InTy; | |||
14484 | else if (InTy != SourceType) { | |||
14485 | // Multiple income types. Abort. | |||
14486 | SourceType = MVT::Other; | |||
14487 | break; | |||
14488 | } | |||
14489 | ||||
14490 | // Check if all of the extends are ANY_EXTENDs. | |||
14491 | AllAnyExt &= AnyExt; | |||
14492 | } | |||
14493 | ||||
14494 | // In order to have valid types, all of the inputs must be extended from the | |||
14495 | // same source type and all of the inputs must be any or zero extend. | |||
14496 | // Scalar sizes must be a power of two. | |||
14497 | EVT OutScalarTy = VT.getScalarType(); | |||
14498 | bool ValidTypes = SourceType != MVT::Other && | |||
14499 | isPowerOf2_32(OutScalarTy.getSizeInBits()) && | |||
14500 | isPowerOf2_32(SourceType.getSizeInBits()); | |||
14501 | ||||
14502 | // Create a new simpler BUILD_VECTOR sequence which other optimizations can | |||
14503 | // turn into a single shuffle instruction. | |||
14504 | if (!ValidTypes) | |||
14505 | return SDValue(); | |||
14506 | ||||
14507 | bool isLE = DAG.getDataLayout().isLittleEndian(); | |||
14508 | unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits(); | |||
14509 | assert(ElemRatio > 1 && "Invalid element size ratio")(static_cast <bool> (ElemRatio > 1 && "Invalid element size ratio" ) ? void (0) : __assert_fail ("ElemRatio > 1 && \"Invalid element size ratio\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 14509, __extension__ __PRETTY_FUNCTION__)); | |||
14510 | SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType): | |||
14511 | DAG.getConstant(0, DL, SourceType); | |||
14512 | ||||
14513 | unsigned NewBVElems = ElemRatio * VT.getVectorNumElements(); | |||
14514 | SmallVector<SDValue, 8> Ops(NewBVElems, Filler); | |||
14515 | ||||
14516 | // Populate the new build_vector | |||
14517 | for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { | |||
14518 | SDValue Cast = N->getOperand(i); | |||
14519 | assert((Cast.getOpcode() == ISD::ANY_EXTEND ||(static_cast <bool> ((Cast.getOpcode() == ISD::ANY_EXTEND || Cast.getOpcode() == ISD::ZERO_EXTEND || Cast.isUndef()) && "Invalid cast opcode") ? void (0) : __assert_fail ("(Cast.getOpcode() == ISD::ANY_EXTEND || Cast.getOpcode() == ISD::ZERO_EXTEND || Cast.isUndef()) && \"Invalid cast opcode\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 14521, __extension__ __PRETTY_FUNCTION__)) | |||
14520 | Cast.getOpcode() == ISD::ZERO_EXTEND ||(static_cast <bool> ((Cast.getOpcode() == ISD::ANY_EXTEND || Cast.getOpcode() == ISD::ZERO_EXTEND || Cast.isUndef()) && "Invalid cast opcode") ? void (0) : __assert_fail ("(Cast.getOpcode() == ISD::ANY_EXTEND || Cast.getOpcode() == ISD::ZERO_EXTEND || Cast.isUndef()) && \"Invalid cast opcode\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 14521, __extension__ __PRETTY_FUNCTION__)) | |||
14521 | Cast.isUndef()) && "Invalid cast opcode")(static_cast <bool> ((Cast.getOpcode() == ISD::ANY_EXTEND || Cast.getOpcode() == ISD::ZERO_EXTEND || Cast.isUndef()) && "Invalid cast opcode") ? void (0) : __assert_fail ("(Cast.getOpcode() == ISD::ANY_EXTEND || Cast.getOpcode() == ISD::ZERO_EXTEND || Cast.isUndef()) && \"Invalid cast opcode\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 14521, __extension__ __PRETTY_FUNCTION__)); | |||
14522 | SDValue In; | |||
14523 | if (Cast.isUndef()) | |||
14524 | In = DAG.getUNDEF(SourceType); | |||
14525 | else | |||
14526 | In = Cast->getOperand(0); | |||
14527 | unsigned Index = isLE ? (i * ElemRatio) : | |||
14528 | (i * ElemRatio + (ElemRatio - 1)); | |||
14529 | ||||
14530 | assert(Index < Ops.size() && "Invalid index")(static_cast <bool> (Index < Ops.size() && "Invalid index" ) ? void (0) : __assert_fail ("Index < Ops.size() && \"Invalid index\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 14530, __extension__ __PRETTY_FUNCTION__)); | |||
14531 | Ops[Index] = In; | |||
14532 | } | |||
14533 | ||||
14534 | // The type of the new BUILD_VECTOR node. | |||
14535 | EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems); | |||
14536 | assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&(static_cast <bool> (VecVT.getSizeInBits() == VT.getSizeInBits () && "Invalid vector size") ? void (0) : __assert_fail ("VecVT.getSizeInBits() == VT.getSizeInBits() && \"Invalid vector size\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 14537, __extension__ __PRETTY_FUNCTION__)) | |||
14537 | "Invalid vector size")(static_cast <bool> (VecVT.getSizeInBits() == VT.getSizeInBits () && "Invalid vector size") ? void (0) : __assert_fail ("VecVT.getSizeInBits() == VT.getSizeInBits() && \"Invalid vector size\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 14537, __extension__ __PRETTY_FUNCTION__)); | |||
14538 | // Check if the new vector type is legal. | |||
14539 | if (!isTypeLegal(VecVT)) return SDValue(); | |||
14540 | ||||
14541 | // Make the new BUILD_VECTOR. | |||
14542 | SDValue BV = DAG.getBuildVector(VecVT, DL, Ops); | |||
14543 | ||||
14544 | // The new BUILD_VECTOR node has the potential to be further optimized. | |||
14545 | AddToWorklist(BV.getNode()); | |||
14546 | // Bitcast to the desired type. | |||
14547 | return DAG.getBitcast(VT, BV); | |||
14548 | } | |||
14549 | ||||
14550 | SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) { | |||
14551 | EVT VT = N->getValueType(0); | |||
14552 | ||||
14553 | unsigned NumInScalars = N->getNumOperands(); | |||
14554 | SDLoc DL(N); | |||
14555 | ||||
14556 | EVT SrcVT = MVT::Other; | |||
14557 | unsigned Opcode = ISD::DELETED_NODE; | |||
14558 | unsigned NumDefs = 0; | |||
14559 | ||||
14560 | for (unsigned i = 0; i != NumInScalars; ++i) { | |||
14561 | SDValue In = N->getOperand(i); | |||
14562 | unsigned Opc = In.getOpcode(); | |||
14563 | ||||
14564 | if (Opc == ISD::UNDEF) | |||
14565 | continue; | |||
14566 | ||||
14567 | // If all scalar values are floats and converted from integers. | |||
14568 | if (Opcode == ISD::DELETED_NODE && | |||
14569 | (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) { | |||
14570 | Opcode = Opc; | |||
14571 | } | |||
14572 | ||||
14573 | if (Opc != Opcode) | |||
14574 | return SDValue(); | |||
14575 | ||||
14576 | EVT InVT = In.getOperand(0).getValueType(); | |||
14577 | ||||
14578 | // If all scalar values are typed differently, bail out. It's chosen to | |||
14579 | // simplify BUILD_VECTOR of integer types. | |||
14580 | if (SrcVT == MVT::Other) | |||
14581 | SrcVT = InVT; | |||
14582 | if (SrcVT != InVT) | |||
14583 | return SDValue(); | |||
14584 | NumDefs++; | |||
14585 | } | |||
14586 | ||||
14587 | // If the vector has just one element defined, it's not worth to fold it into | |||
14588 | // a vectorized one. | |||
14589 | if (NumDefs < 2) | |||
14590 | return SDValue(); | |||
14591 | ||||
14592 | assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)(static_cast <bool> ((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP) && "Should only handle conversion from integer to float." ) ? void (0) : __assert_fail ("(Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP) && \"Should only handle conversion from integer to float.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 14593, __extension__ __PRETTY_FUNCTION__)) | |||
14593 | && "Should only handle conversion from integer to float.")(static_cast <bool> ((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP) && "Should only handle conversion from integer to float." ) ? void (0) : __assert_fail ("(Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP) && \"Should only handle conversion from integer to float.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 14593, __extension__ __PRETTY_FUNCTION__)); | |||
14594 | assert(SrcVT != MVT::Other && "Cannot determine source type!")(static_cast <bool> (SrcVT != MVT::Other && "Cannot determine source type!" ) ? void (0) : __assert_fail ("SrcVT != MVT::Other && \"Cannot determine source type!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 14594, __extension__ __PRETTY_FUNCTION__)); | |||
14595 | ||||
14596 | EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars); | |||
14597 | ||||
14598 | if (!TLI.isOperationLegalOrCustom(Opcode, NVT)) | |||
14599 | return SDValue(); | |||
14600 | ||||
14601 | // Just because the floating-point vector type is legal does not necessarily | |||
14602 | // mean that the corresponding integer vector type is. | |||
14603 | if (!isTypeLegal(NVT)) | |||
14604 | return SDValue(); | |||
14605 | ||||
14606 | SmallVector<SDValue, 8> Opnds; | |||
14607 | for (unsigned i = 0; i != NumInScalars; ++i) { | |||
14608 | SDValue In = N->getOperand(i); | |||
14609 | ||||
14610 | if (In.isUndef()) | |||
14611 | Opnds.push_back(DAG.getUNDEF(SrcVT)); | |||
14612 | else | |||
14613 | Opnds.push_back(In.getOperand(0)); | |||
14614 | } | |||
14615 | SDValue BV = DAG.getBuildVector(NVT, DL, Opnds); | |||
14616 | AddToWorklist(BV.getNode()); | |||
14617 | ||||
14618 | return DAG.getNode(Opcode, DL, VT, BV); | |||
14619 | } | |||
14620 | ||||
14621 | SDValue DAGCombiner::createBuildVecShuffle(const SDLoc &DL, SDNode *N, | |||
14622 | ArrayRef<int> VectorMask, | |||
14623 | SDValue VecIn1, SDValue VecIn2, | |||
14624 | unsigned LeftIdx) { | |||
14625 | MVT IdxTy = TLI.getVectorIdxTy(DAG.getDataLayout()); | |||
14626 | SDValue ZeroIdx = DAG.getConstant(0, DL, IdxTy); | |||
14627 | ||||
14628 | EVT VT = N->getValueType(0); | |||
14629 | EVT InVT1 = VecIn1.getValueType(); | |||
14630 | EVT InVT2 = VecIn2.getNode() ? VecIn2.getValueType() : InVT1; | |||
14631 | ||||
14632 | unsigned Vec2Offset = 0; | |||
14633 | unsigned NumElems = VT.getVectorNumElements(); | |||
14634 | unsigned ShuffleNumElems = NumElems; | |||
14635 | ||||
14636 | // In case both the input vectors are extracted from same base | |||
14637 | // vector we do not need extra addend (Vec2Offset) while | |||
14638 | // computing shuffle mask. | |||
14639 | if (!VecIn2 || !(VecIn1.getOpcode() == ISD::EXTRACT_SUBVECTOR) || | |||
14640 | !(VecIn2.getOpcode() == ISD::EXTRACT_SUBVECTOR) || | |||
14641 | !(VecIn1.getOperand(0) == VecIn2.getOperand(0))) | |||
14642 | Vec2Offset = InVT1.getVectorNumElements(); | |||
14643 | ||||
14644 | // We can't generate a shuffle node with mismatched input and output types. | |||
14645 | // Try to make the types match the type of the output. | |||
14646 | if (InVT1 != VT || InVT2 != VT) { | |||
14647 | if ((VT.getSizeInBits() % InVT1.getSizeInBits() == 0) && InVT1 == InVT2) { | |||
14648 | // If the output vector length is a multiple of both input lengths, | |||
14649 | // we can concatenate them and pad the rest with undefs. | |||
14650 | unsigned NumConcats = VT.getSizeInBits() / InVT1.getSizeInBits(); | |||
14651 | assert(NumConcats >= 2 && "Concat needs at least two inputs!")(static_cast <bool> (NumConcats >= 2 && "Concat needs at least two inputs!" ) ? void (0) : __assert_fail ("NumConcats >= 2 && \"Concat needs at least two inputs!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 14651, __extension__ __PRETTY_FUNCTION__)); | |||
14652 | SmallVector<SDValue, 2> ConcatOps(NumConcats, DAG.getUNDEF(InVT1)); | |||
14653 | ConcatOps[0] = VecIn1; | |||
14654 | ConcatOps[1] = VecIn2 ? VecIn2 : DAG.getUNDEF(InVT1); | |||
14655 | VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, ConcatOps); | |||
14656 | VecIn2 = SDValue(); | |||
14657 | } else if (InVT1.getSizeInBits() == VT.getSizeInBits() * 2) { | |||
14658 | if (!TLI.isExtractSubvectorCheap(VT, InVT1, NumElems)) | |||
14659 | return SDValue(); | |||
14660 | ||||
14661 | if (!VecIn2.getNode()) { | |||
14662 | // If we only have one input vector, and it's twice the size of the | |||
14663 | // output, split it in two. | |||
14664 | VecIn2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, VecIn1, | |||
14665 | DAG.getConstant(NumElems, DL, IdxTy)); | |||
14666 | VecIn1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, VecIn1, ZeroIdx); | |||
14667 | // Since we now have shorter input vectors, adjust the offset of the | |||
14668 | // second vector's start. | |||
14669 | Vec2Offset = NumElems; | |||
14670 | } else if (InVT2.getSizeInBits() <= InVT1.getSizeInBits()) { | |||
14671 | // VecIn1 is wider than the output, and we have another, possibly | |||
14672 | // smaller input. Pad the smaller input with undefs, shuffle at the | |||
14673 | // input vector width, and extract the output. | |||
14674 | // The shuffle type is different than VT, so check legality again. | |||
14675 | if (LegalOperations && | |||
14676 | !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, InVT1)) | |||
14677 | return SDValue(); | |||
14678 | ||||
14679 | // Legalizing INSERT_SUBVECTOR is tricky - you basically have to | |||
14680 | // lower it back into a BUILD_VECTOR. So if the inserted type is | |||
14681 | // illegal, don't even try. | |||
14682 | if (InVT1 != InVT2) { | |||
14683 | if (!TLI.isTypeLegal(InVT2)) | |||
14684 | return SDValue(); | |||
14685 | VecIn2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, InVT1, | |||
14686 | DAG.getUNDEF(InVT1), VecIn2, ZeroIdx); | |||
14687 | } | |||
14688 | ShuffleNumElems = NumElems * 2; | |||
14689 | } else { | |||
14690 | // Both VecIn1 and VecIn2 are wider than the output, and VecIn2 is wider | |||
14691 | // than VecIn1. We can't handle this for now - this case will disappear | |||
14692 | // when we start sorting the vectors by type. | |||
14693 | return SDValue(); | |||
14694 | } | |||
14695 | } else if (InVT2.getSizeInBits() * 2 == VT.getSizeInBits() && | |||
14696 | InVT1.getSizeInBits() == VT.getSizeInBits()) { | |||
14697 | SmallVector<SDValue, 2> ConcatOps(2, DAG.getUNDEF(InVT2)); | |||
14698 | ConcatOps[0] = VecIn2; | |||
14699 | VecIn2 = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, ConcatOps); | |||
14700 | } else { | |||
14701 | // TODO: Support cases where the length mismatch isn't exactly by a | |||
14702 | // factor of 2. | |||
14703 | // TODO: Move this check upwards, so that if we have bad type | |||
14704 | // mismatches, we don't create any DAG nodes. | |||
14705 | return SDValue(); | |||
14706 | } | |||
14707 | } | |||
14708 | ||||
14709 | // Initialize mask to undef. | |||
14710 | SmallVector<int, 8> Mask(ShuffleNumElems, -1); | |||
14711 | ||||
14712 | // Only need to run up to the number of elements actually used, not the | |||
14713 | // total number of elements in the shuffle - if we are shuffling a wider | |||
14714 | // vector, the high lanes should be set to undef. | |||
14715 | for (unsigned i = 0; i != NumElems; ++i) { | |||
14716 | if (VectorMask[i] <= 0) | |||
14717 | continue; | |||
14718 | ||||
14719 | unsigned ExtIndex = N->getOperand(i).getConstantOperandVal(1); | |||
14720 | if (VectorMask[i] == (int)LeftIdx) { | |||
14721 | Mask[i] = ExtIndex; | |||
14722 | } else if (VectorMask[i] == (int)LeftIdx + 1) { | |||
14723 | Mask[i] = Vec2Offset + ExtIndex; | |||
14724 | } | |||
14725 | } | |||
14726 | ||||
14727 | // The type the input vectors may have changed above. | |||
14728 | InVT1 = VecIn1.getValueType(); | |||
14729 | ||||
14730 | // If we already have a VecIn2, it should have the same type as VecIn1. | |||
14731 | // If we don't, get an undef/zero vector of the appropriate type. | |||
14732 | VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(InVT1); | |||
14733 | assert(InVT1 == VecIn2.getValueType() && "Unexpected second input type.")(static_cast <bool> (InVT1 == VecIn2.getValueType() && "Unexpected second input type.") ? void (0) : __assert_fail ( "InVT1 == VecIn2.getValueType() && \"Unexpected second input type.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 14733, __extension__ __PRETTY_FUNCTION__)); | |||
14734 | ||||
14735 | SDValue Shuffle = DAG.getVectorShuffle(InVT1, DL, VecIn1, VecIn2, Mask); | |||
14736 | if (ShuffleNumElems > NumElems) | |||
14737 | Shuffle = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Shuffle, ZeroIdx); | |||
14738 | ||||
14739 | return Shuffle; | |||
14740 | } | |||
14741 | ||||
14742 | // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT | |||
14743 | // operations. If the types of the vectors we're extracting from allow it, | |||
14744 | // turn this into a vector_shuffle node. | |||
14745 | SDValue DAGCombiner::reduceBuildVecToShuffle(SDNode *N) { | |||
14746 | SDLoc DL(N); | |||
14747 | EVT VT = N->getValueType(0); | |||
14748 | ||||
14749 | // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes. | |||
14750 | if (!isTypeLegal(VT)) | |||
14751 | return SDValue(); | |||
14752 | ||||
14753 | // May only combine to shuffle after legalize if shuffle is legal. | |||
14754 | if (LegalOperations && !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, VT)) | |||
14755 | return SDValue(); | |||
14756 | ||||
14757 | bool UsesZeroVector = false; | |||
14758 | unsigned NumElems = N->getNumOperands(); | |||
14759 | ||||
14760 | // Record, for each element of the newly built vector, which input vector | |||
14761 | // that element comes from. -1 stands for undef, 0 for the zero vector, | |||
14762 | // and positive values for the input vectors. | |||
14763 | // VectorMask maps each element to its vector number, and VecIn maps vector | |||
14764 | // numbers to their initial SDValues. | |||
14765 | ||||
14766 | SmallVector<int, 8> VectorMask(NumElems, -1); | |||
14767 | SmallVector<SDValue, 8> VecIn; | |||
14768 | VecIn.push_back(SDValue()); | |||
14769 | ||||
14770 | for (unsigned i = 0; i != NumElems; ++i) { | |||
14771 | SDValue Op = N->getOperand(i); | |||
14772 | ||||
14773 | if (Op.isUndef()) | |||
14774 | continue; | |||
14775 | ||||
14776 | // See if we can use a blend with a zero vector. | |||
14777 | // TODO: Should we generalize this to a blend with an arbitrary constant | |||
14778 | // vector? | |||
14779 | if (isNullConstant(Op) || isNullFPConstant(Op)) { | |||
14780 | UsesZeroVector = true; | |||
14781 | VectorMask[i] = 0; | |||
14782 | continue; | |||
14783 | } | |||
14784 | ||||
14785 | // Not an undef or zero. If the input is something other than an | |||
14786 | // EXTRACT_VECTOR_ELT with an in-range constant index, bail out. | |||
14787 | if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT || | |||
14788 | !isa<ConstantSDNode>(Op.getOperand(1))) | |||
14789 | return SDValue(); | |||
14790 | SDValue ExtractedFromVec = Op.getOperand(0); | |||
14791 | ||||
14792 | APInt ExtractIdx = cast<ConstantSDNode>(Op.getOperand(1))->getAPIntValue(); | |||
14793 | if (ExtractIdx.uge(ExtractedFromVec.getValueType().getVectorNumElements())) | |||
14794 | return SDValue(); | |||
14795 | ||||
14796 | // All inputs must have the same element type as the output. | |||
14797 | if (VT.getVectorElementType() != | |||
14798 | ExtractedFromVec.getValueType().getVectorElementType()) | |||
14799 | return SDValue(); | |||
14800 | ||||
14801 | // Have we seen this input vector before? | |||
14802 | // The vectors are expected to be tiny (usually 1 or 2 elements), so using | |||
14803 | // a map back from SDValues to numbers isn't worth it. | |||
14804 | unsigned Idx = std::distance( | |||
14805 | VecIn.begin(), std::find(VecIn.begin(), VecIn.end(), ExtractedFromVec)); | |||
14806 | if (Idx == VecIn.size()) | |||
14807 | VecIn.push_back(ExtractedFromVec); | |||
14808 | ||||
14809 | VectorMask[i] = Idx; | |||
14810 | } | |||
14811 | ||||
14812 | // If we didn't find at least one input vector, bail out. | |||
14813 | if (VecIn.size() < 2) | |||
14814 | return SDValue(); | |||
14815 | ||||
14816 | // If all the Operands of BUILD_VECTOR extract from same | |||
14817 | // vector, then split the vector efficiently based on the maximum | |||
14818 | // vector access index and adjust the VectorMask and | |||
14819 | // VecIn accordingly. | |||
14820 | if (VecIn.size() == 2) { | |||
14821 | unsigned MaxIndex = 0; | |||
14822 | unsigned NearestPow2 = 0; | |||
14823 | SDValue Vec = VecIn.back(); | |||
14824 | EVT InVT = Vec.getValueType(); | |||
14825 | MVT IdxTy = TLI.getVectorIdxTy(DAG.getDataLayout()); | |||
14826 | SmallVector<unsigned, 8> IndexVec(NumElems, 0); | |||
14827 | ||||
14828 | for (unsigned i = 0; i < NumElems; i++) { | |||
14829 | if (VectorMask[i] <= 0) | |||
14830 | continue; | |||
14831 | unsigned Index = N->getOperand(i).getConstantOperandVal(1); | |||
14832 | IndexVec[i] = Index; | |||
14833 | MaxIndex = std::max(MaxIndex, Index); | |||
14834 | } | |||
14835 | ||||
14836 | NearestPow2 = PowerOf2Ceil(MaxIndex); | |||
14837 | if (InVT.isSimple() && NearestPow2 > 2 && MaxIndex < NearestPow2 && | |||
14838 | NumElems * 2 < NearestPow2) { | |||
14839 | unsigned SplitSize = NearestPow2 / 2; | |||
14840 | EVT SplitVT = EVT::getVectorVT(*DAG.getContext(), | |||
14841 | InVT.getVectorElementType(), SplitSize); | |||
14842 | if (TLI.isTypeLegal(SplitVT)) { | |||
14843 | SDValue VecIn2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, Vec, | |||
14844 | DAG.getConstant(SplitSize, DL, IdxTy)); | |||
14845 | SDValue VecIn1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, Vec, | |||
14846 | DAG.getConstant(0, DL, IdxTy)); | |||
14847 | VecIn.pop_back(); | |||
14848 | VecIn.push_back(VecIn1); | |||
14849 | VecIn.push_back(VecIn2); | |||
14850 | ||||
14851 | for (unsigned i = 0; i < NumElems; i++) { | |||
14852 | if (VectorMask[i] <= 0) | |||
14853 | continue; | |||
14854 | VectorMask[i] = (IndexVec[i] < SplitSize) ? 1 : 2; | |||
14855 | } | |||
14856 | } | |||
14857 | } | |||
14858 | } | |||
14859 | ||||
14860 | // TODO: We want to sort the vectors by descending length, so that adjacent | |||
14861 | // pairs have similar length, and the longer vector is always first in the | |||
14862 | // pair. | |||
14863 | ||||
14864 | // TODO: Should this fire if some of the input vectors has illegal type (like | |||
14865 | // it does now), or should we let legalization run its course first? | |||
14866 | ||||
14867 | // Shuffle phase: | |||
14868 | // Take pairs of vectors, and shuffle them so that the result has elements | |||
14869 | // from these vectors in the correct places. | |||
14870 | // For example, given: | |||
14871 | // t10: i32 = extract_vector_elt t1, Constant:i64<0> | |||
14872 | // t11: i32 = extract_vector_elt t2, Constant:i64<0> | |||
14873 | // t12: i32 = extract_vector_elt t3, Constant:i64<0> | |||
14874 | // t13: i32 = extract_vector_elt t1, Constant:i64<1> | |||
14875 | // t14: v4i32 = BUILD_VECTOR t10, t11, t12, t13 | |||
14876 | // We will generate: | |||
14877 | // t20: v4i32 = vector_shuffle<0,4,u,1> t1, t2 | |||
14878 | // t21: v4i32 = vector_shuffle<u,u,0,u> t3, undef | |||
14879 | SmallVector<SDValue, 4> Shuffles; | |||
14880 | for (unsigned In = 0, Len = (VecIn.size() / 2); In < Len; ++In) { | |||
14881 | unsigned LeftIdx = 2 * In + 1; | |||
14882 | SDValue VecLeft = VecIn[LeftIdx]; | |||
14883 | SDValue VecRight = | |||
14884 | (LeftIdx + 1) < VecIn.size() ? VecIn[LeftIdx + 1] : SDValue(); | |||
14885 | ||||
14886 | if (SDValue Shuffle = createBuildVecShuffle(DL, N, VectorMask, VecLeft, | |||
14887 | VecRight, LeftIdx)) | |||
14888 | Shuffles.push_back(Shuffle); | |||
14889 | else | |||
14890 | return SDValue(); | |||
14891 | } | |||
14892 | ||||
14893 | // If we need the zero vector as an "ingredient" in the blend tree, add it | |||
14894 | // to the list of shuffles. | |||
14895 | if (UsesZeroVector) | |||
14896 | Shuffles.push_back(VT.isInteger() ? DAG.getConstant(0, DL, VT) | |||
14897 | : DAG.getConstantFP(0.0, DL, VT)); | |||
14898 | ||||
14899 | // If we only have one shuffle, we're done. | |||
14900 | if (Shuffles.size() == 1) | |||
14901 | return Shuffles[0]; | |||
14902 | ||||
14903 | // Update the vector mask to point to the post-shuffle vectors. | |||
14904 | for (int &Vec : VectorMask) | |||
14905 | if (Vec == 0) | |||
14906 | Vec = Shuffles.size() - 1; | |||
14907 | else | |||
14908 | Vec = (Vec - 1) / 2; | |||
14909 | ||||
14910 | // More than one shuffle. Generate a binary tree of blends, e.g. if from | |||
14911 | // the previous step we got the set of shuffles t10, t11, t12, t13, we will | |||
14912 | // generate: | |||
14913 | // t10: v8i32 = vector_shuffle<0,8,u,u,u,u,u,u> t1, t2 | |||
14914 | // t11: v8i32 = vector_shuffle<u,u,0,8,u,u,u,u> t3, t4 | |||
14915 | // t12: v8i32 = vector_shuffle<u,u,u,u,0,8,u,u> t5, t6 | |||
14916 | // t13: v8i32 = vector_shuffle<u,u,u,u,u,u,0,8> t7, t8 | |||
14917 | // t20: v8i32 = vector_shuffle<0,1,10,11,u,u,u,u> t10, t11 | |||
14918 | // t21: v8i32 = vector_shuffle<u,u,u,u,4,5,14,15> t12, t13 | |||
14919 | // t30: v8i32 = vector_shuffle<0,1,2,3,12,13,14,15> t20, t21 | |||
14920 | ||||
14921 | // Make sure the initial size of the shuffle list is even. | |||
14922 | if (Shuffles.size() % 2) | |||
14923 | Shuffles.push_back(DAG.getUNDEF(VT)); | |||
14924 | ||||
14925 | for (unsigned CurSize = Shuffles.size(); CurSize > 1; CurSize /= 2) { | |||
14926 | if (CurSize % 2) { | |||
14927 | Shuffles[CurSize] = DAG.getUNDEF(VT); | |||
14928 | CurSize++; | |||
14929 | } | |||
14930 | for (unsigned In = 0, Len = CurSize / 2; In < Len; ++In) { | |||
14931 | int Left = 2 * In; | |||
14932 | int Right = 2 * In + 1; | |||
14933 | SmallVector<int, 8> Mask(NumElems, -1); | |||
14934 | for (unsigned i = 0; i != NumElems; ++i) { | |||
14935 | if (VectorMask[i] == Left) { | |||
14936 | Mask[i] = i; | |||
14937 | VectorMask[i] = In; | |||
14938 | } else if (VectorMask[i] == Right) { | |||
14939 | Mask[i] = i + NumElems; | |||
14940 | VectorMask[i] = In; | |||
14941 | } | |||
14942 | } | |||
14943 | ||||
14944 | Shuffles[In] = | |||
14945 | DAG.getVectorShuffle(VT, DL, Shuffles[Left], Shuffles[Right], Mask); | |||
14946 | } | |||
14947 | } | |||
14948 | return Shuffles[0]; | |||
14949 | } | |||
14950 | ||||
14951 | SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) { | |||
14952 | EVT VT = N->getValueType(0); | |||
14953 | ||||
14954 | // A vector built entirely of undefs is undef. | |||
14955 | if (ISD::allOperandsUndef(N)) | |||
14956 | return DAG.getUNDEF(VT); | |||
14957 | ||||
14958 | // If this is a splat of a bitcast from another vector, change to a | |||
14959 | // concat_vector. | |||
14960 | // For example: | |||
14961 | // (build_vector (i64 (bitcast (v2i32 X))), (i64 (bitcast (v2i32 X)))) -> | |||
14962 | // (v2i64 (bitcast (concat_vectors (v2i32 X), (v2i32 X)))) | |||
14963 | // | |||
14964 | // If X is a build_vector itself, the concat can become a larger build_vector. | |||
14965 | // TODO: Maybe this is useful for non-splat too? | |||
14966 | if (!LegalOperations) { | |||
14967 | if (SDValue Splat = cast<BuildVectorSDNode>(N)->getSplatValue()) { | |||
14968 | Splat = peekThroughBitcast(Splat); | |||
14969 | EVT SrcVT = Splat.getValueType(); | |||
14970 | if (SrcVT.isVector()) { | |||
14971 | unsigned NumElts = N->getNumOperands() * SrcVT.getVectorNumElements(); | |||
14972 | EVT NewVT = EVT::getVectorVT(*DAG.getContext(), | |||
14973 | SrcVT.getVectorElementType(), NumElts); | |||
14974 | SmallVector<SDValue, 8> Ops(N->getNumOperands(), Splat); | |||
14975 | SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), NewVT, Ops); | |||
14976 | return DAG.getBitcast(VT, Concat); | |||
14977 | } | |||
14978 | } | |||
14979 | } | |||
14980 | ||||
14981 | // Check if we can express BUILD VECTOR via subvector extract. | |||
14982 | if (!LegalTypes && (N->getNumOperands() > 1)) { | |||
14983 | SDValue Op0 = N->getOperand(0); | |||
14984 | auto checkElem = [&](SDValue Op) -> uint64_t { | |||
14985 | if ((Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT) && | |||
14986 | (Op0.getOperand(0) == Op.getOperand(0))) | |||
14987 | if (auto CNode = dyn_cast<ConstantSDNode>(Op.getOperand(1))) | |||
14988 | return CNode->getZExtValue(); | |||
14989 | return -1; | |||
14990 | }; | |||
14991 | ||||
14992 | int Offset = checkElem(Op0); | |||
14993 | for (unsigned i = 0; i < N->getNumOperands(); ++i) { | |||
14994 | if (Offset + i != checkElem(N->getOperand(i))) { | |||
14995 | Offset = -1; | |||
14996 | break; | |||
14997 | } | |||
14998 | } | |||
14999 | ||||
15000 | if ((Offset == 0) && | |||
15001 | (Op0.getOperand(0).getValueType() == N->getValueType(0))) | |||
15002 | return Op0.getOperand(0); | |||
15003 | if ((Offset != -1) && | |||
15004 | ((Offset % N->getValueType(0).getVectorNumElements()) == | |||
15005 | 0)) // IDX must be multiple of output size. | |||
15006 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N), N->getValueType(0), | |||
15007 | Op0.getOperand(0), Op0.getOperand(1)); | |||
15008 | } | |||
15009 | ||||
15010 | if (SDValue V = reduceBuildVecExtToExtBuildVec(N)) | |||
15011 | return V; | |||
15012 | ||||
15013 | if (SDValue V = reduceBuildVecConvertToConvertBuildVec(N)) | |||
15014 | return V; | |||
15015 | ||||
15016 | if (SDValue V = reduceBuildVecToShuffle(N)) | |||
15017 | return V; | |||
15018 | ||||
15019 | return SDValue(); | |||
15020 | } | |||
15021 | ||||
15022 | static SDValue combineConcatVectorOfScalars(SDNode *N, SelectionDAG &DAG) { | |||
15023 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
15024 | EVT OpVT = N->getOperand(0).getValueType(); | |||
15025 | ||||
15026 | // If the operands are legal vectors, leave them alone. | |||
15027 | if (TLI.isTypeLegal(OpVT)) | |||
15028 | return SDValue(); | |||
15029 | ||||
15030 | SDLoc DL(N); | |||
15031 | EVT VT = N->getValueType(0); | |||
15032 | SmallVector<SDValue, 8> Ops; | |||
15033 | ||||
15034 | EVT SVT = EVT::getIntegerVT(*DAG.getContext(), OpVT.getSizeInBits()); | |||
15035 | SDValue ScalarUndef = DAG.getNode(ISD::UNDEF, DL, SVT); | |||
15036 | ||||
15037 | // Keep track of what we encounter. | |||
15038 | bool AnyInteger = false; | |||
15039 | bool AnyFP = false; | |||
15040 | for (const SDValue &Op : N->ops()) { | |||
15041 | if (ISD::BITCAST == Op.getOpcode() && | |||
15042 | !Op.getOperand(0).getValueType().isVector()) | |||
15043 | Ops.push_back(Op.getOperand(0)); | |||
15044 | else if (ISD::UNDEF == Op.getOpcode()) | |||
15045 | Ops.push_back(ScalarUndef); | |||
15046 | else | |||
15047 | return SDValue(); | |||
15048 | ||||
15049 | // Note whether we encounter an integer or floating point scalar. | |||
15050 | // If it's neither, bail out, it could be something weird like x86mmx. | |||
15051 | EVT LastOpVT = Ops.back().getValueType(); | |||
15052 | if (LastOpVT.isFloatingPoint()) | |||
15053 | AnyFP = true; | |||
15054 | else if (LastOpVT.isInteger()) | |||
15055 | AnyInteger = true; | |||
15056 | else | |||
15057 | return SDValue(); | |||
15058 | } | |||
15059 | ||||
15060 | // If any of the operands is a floating point scalar bitcast to a vector, | |||
15061 | // use floating point types throughout, and bitcast everything. | |||
15062 | // Replace UNDEFs by another scalar UNDEF node, of the final desired type. | |||
15063 | if (AnyFP) { | |||
15064 | SVT = EVT::getFloatingPointVT(OpVT.getSizeInBits()); | |||
15065 | ScalarUndef = DAG.getNode(ISD::UNDEF, DL, SVT); | |||
15066 | if (AnyInteger) { | |||
15067 | for (SDValue &Op : Ops) { | |||
15068 | if (Op.getValueType() == SVT) | |||
15069 | continue; | |||
15070 | if (Op.isUndef()) | |||
15071 | Op = ScalarUndef; | |||
15072 | else | |||
15073 | Op = DAG.getBitcast(SVT, Op); | |||
15074 | } | |||
15075 | } | |||
15076 | } | |||
15077 | ||||
15078 | EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SVT, | |||
15079 | VT.getSizeInBits() / SVT.getSizeInBits()); | |||
15080 | return DAG.getBitcast(VT, DAG.getBuildVector(VecVT, DL, Ops)); | |||
15081 | } | |||
15082 | ||||
15083 | // Check to see if this is a CONCAT_VECTORS of a bunch of EXTRACT_SUBVECTOR | |||
15084 | // operations. If so, and if the EXTRACT_SUBVECTOR vector inputs come from at | |||
15085 | // most two distinct vectors the same size as the result, attempt to turn this | |||
15086 | // into a legal shuffle. | |||
15087 | static SDValue combineConcatVectorOfExtracts(SDNode *N, SelectionDAG &DAG) { | |||
15088 | EVT VT = N->getValueType(0); | |||
15089 | EVT OpVT = N->getOperand(0).getValueType(); | |||
15090 | int NumElts = VT.getVectorNumElements(); | |||
15091 | int NumOpElts = OpVT.getVectorNumElements(); | |||
15092 | ||||
15093 | SDValue SV0 = DAG.getUNDEF(VT), SV1 = DAG.getUNDEF(VT); | |||
15094 | SmallVector<int, 8> Mask; | |||
15095 | ||||
15096 | for (SDValue Op : N->ops()) { | |||
15097 | // Peek through any bitcast. | |||
15098 | Op = peekThroughBitcast(Op); | |||
15099 | ||||
15100 | // UNDEF nodes convert to UNDEF shuffle mask values. | |||
15101 | if (Op.isUndef()) { | |||
15102 | Mask.append((unsigned)NumOpElts, -1); | |||
15103 | continue; | |||
15104 | } | |||
15105 | ||||
15106 | if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR) | |||
15107 | return SDValue(); | |||
15108 | ||||
15109 | // What vector are we extracting the subvector from and at what index? | |||
15110 | SDValue ExtVec = Op.getOperand(0); | |||
15111 | ||||
15112 | // We want the EVT of the original extraction to correctly scale the | |||
15113 | // extraction index. | |||
15114 | EVT ExtVT = ExtVec.getValueType(); | |||
15115 | ||||
15116 | // Peek through any bitcast. | |||
15117 | ExtVec = peekThroughBitcast(ExtVec); | |||
15118 | ||||
15119 | // UNDEF nodes convert to UNDEF shuffle mask values. | |||
15120 | if (ExtVec.isUndef()) { | |||
15121 | Mask.append((unsigned)NumOpElts, -1); | |||
15122 | continue; | |||
15123 | } | |||
15124 | ||||
15125 | if (!isa<ConstantSDNode>(Op.getOperand(1))) | |||
15126 | return SDValue(); | |||
15127 | int ExtIdx = Op.getConstantOperandVal(1); | |||
15128 | ||||
15129 | // Ensure that we are extracting a subvector from a vector the same | |||
15130 | // size as the result. | |||
15131 | if (ExtVT.getSizeInBits() != VT.getSizeInBits()) | |||
15132 | return SDValue(); | |||
15133 | ||||
15134 | // Scale the subvector index to account for any bitcast. | |||
15135 | int NumExtElts = ExtVT.getVectorNumElements(); | |||
15136 | if (0 == (NumExtElts % NumElts)) | |||
15137 | ExtIdx /= (NumExtElts / NumElts); | |||
15138 | else if (0 == (NumElts % NumExtElts)) | |||
15139 | ExtIdx *= (NumElts / NumExtElts); | |||
15140 | else | |||
15141 | return SDValue(); | |||
15142 | ||||
15143 | // At most we can reference 2 inputs in the final shuffle. | |||
15144 | if (SV0.isUndef() || SV0 == ExtVec) { | |||
15145 | SV0 = ExtVec; | |||
15146 | for (int i = 0; i != NumOpElts; ++i) | |||
15147 | Mask.push_back(i + ExtIdx); | |||
15148 | } else if (SV1.isUndef() || SV1 == ExtVec) { | |||
15149 | SV1 = ExtVec; | |||
15150 | for (int i = 0; i != NumOpElts; ++i) | |||
15151 | Mask.push_back(i + ExtIdx + NumElts); | |||
15152 | } else { | |||
15153 | return SDValue(); | |||
15154 | } | |||
15155 | } | |||
15156 | ||||
15157 | if (!DAG.getTargetLoweringInfo().isShuffleMaskLegal(Mask, VT)) | |||
15158 | return SDValue(); | |||
15159 | ||||
15160 | return DAG.getVectorShuffle(VT, SDLoc(N), DAG.getBitcast(VT, SV0), | |||
15161 | DAG.getBitcast(VT, SV1), Mask); | |||
15162 | } | |||
15163 | ||||
15164 | SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) { | |||
15165 | // If we only have one input vector, we don't need to do any concatenation. | |||
15166 | if (N->getNumOperands() == 1) | |||
15167 | return N->getOperand(0); | |||
15168 | ||||
15169 | // Check if all of the operands are undefs. | |||
15170 | EVT VT = N->getValueType(0); | |||
15171 | if (ISD::allOperandsUndef(N)) | |||
15172 | return DAG.getUNDEF(VT); | |||
15173 | ||||
15174 | // Optimize concat_vectors where all but the first of the vectors are undef. | |||
15175 | if (std::all_of(std::next(N->op_begin()), N->op_end(), [](const SDValue &Op) { | |||
15176 | return Op.isUndef(); | |||
15177 | })) { | |||
15178 | SDValue In = N->getOperand(0); | |||
15179 | assert(In.getValueType().isVector() && "Must concat vectors")(static_cast <bool> (In.getValueType().isVector() && "Must concat vectors") ? void (0) : __assert_fail ("In.getValueType().isVector() && \"Must concat vectors\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15179, __extension__ __PRETTY_FUNCTION__)); | |||
15180 | ||||
15181 | // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr). | |||
15182 | if (In->getOpcode() == ISD::BITCAST && | |||
15183 | !In->getOperand(0).getValueType().isVector()) { | |||
15184 | SDValue Scalar = In->getOperand(0); | |||
15185 | ||||
15186 | // If the bitcast type isn't legal, it might be a trunc of a legal type; | |||
15187 | // look through the trunc so we can still do the transform: | |||
15188 | // concat_vectors(trunc(scalar), undef) -> scalar_to_vector(scalar) | |||
15189 | if (Scalar->getOpcode() == ISD::TRUNCATE && | |||
15190 | !TLI.isTypeLegal(Scalar.getValueType()) && | |||
15191 | TLI.isTypeLegal(Scalar->getOperand(0).getValueType())) | |||
15192 | Scalar = Scalar->getOperand(0); | |||
15193 | ||||
15194 | EVT SclTy = Scalar->getValueType(0); | |||
15195 | ||||
15196 | if (!SclTy.isFloatingPoint() && !SclTy.isInteger()) | |||
15197 | return SDValue(); | |||
15198 | ||||
15199 | // Bail out if the vector size is not a multiple of the scalar size. | |||
15200 | if (VT.getSizeInBits() % SclTy.getSizeInBits()) | |||
15201 | return SDValue(); | |||
15202 | ||||
15203 | unsigned VNTNumElms = VT.getSizeInBits() / SclTy.getSizeInBits(); | |||
15204 | if (VNTNumElms < 2) | |||
15205 | return SDValue(); | |||
15206 | ||||
15207 | EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy, VNTNumElms); | |||
15208 | if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType())) | |||
15209 | return SDValue(); | |||
15210 | ||||
15211 | SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), NVT, Scalar); | |||
15212 | return DAG.getBitcast(VT, Res); | |||
15213 | } | |||
15214 | } | |||
15215 | ||||
15216 | // Fold any combination of BUILD_VECTOR or UNDEF nodes into one BUILD_VECTOR. | |||
15217 | // We have already tested above for an UNDEF only concatenation. | |||
15218 | // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...)) | |||
15219 | // -> (BUILD_VECTOR A, B, ..., C, D, ...) | |||
15220 | auto IsBuildVectorOrUndef = [](const SDValue &Op) { | |||
15221 | return ISD::UNDEF == Op.getOpcode() || ISD::BUILD_VECTOR == Op.getOpcode(); | |||
15222 | }; | |||
15223 | if (llvm::all_of(N->ops(), IsBuildVectorOrUndef)) { | |||
15224 | SmallVector<SDValue, 8> Opnds; | |||
15225 | EVT SVT = VT.getScalarType(); | |||
15226 | ||||
15227 | EVT MinVT = SVT; | |||
15228 | if (!SVT.isFloatingPoint()) { | |||
15229 | // If BUILD_VECTOR are from built from integer, they may have different | |||
15230 | // operand types. Get the smallest type and truncate all operands to it. | |||
15231 | bool FoundMinVT = false; | |||
15232 | for (const SDValue &Op : N->ops()) | |||
15233 | if (ISD::BUILD_VECTOR == Op.getOpcode()) { | |||
15234 | EVT OpSVT = Op.getOperand(0).getValueType(); | |||
15235 | MinVT = (!FoundMinVT || OpSVT.bitsLE(MinVT)) ? OpSVT : MinVT; | |||
15236 | FoundMinVT = true; | |||
15237 | } | |||
15238 | assert(FoundMinVT && "Concat vector type mismatch")(static_cast <bool> (FoundMinVT && "Concat vector type mismatch" ) ? void (0) : __assert_fail ("FoundMinVT && \"Concat vector type mismatch\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15238, __extension__ __PRETTY_FUNCTION__)); | |||
15239 | } | |||
15240 | ||||
15241 | for (const SDValue &Op : N->ops()) { | |||
15242 | EVT OpVT = Op.getValueType(); | |||
15243 | unsigned NumElts = OpVT.getVectorNumElements(); | |||
15244 | ||||
15245 | if (ISD::UNDEF == Op.getOpcode()) | |||
15246 | Opnds.append(NumElts, DAG.getUNDEF(MinVT)); | |||
15247 | ||||
15248 | if (ISD::BUILD_VECTOR == Op.getOpcode()) { | |||
15249 | if (SVT.isFloatingPoint()) { | |||
15250 | assert(SVT == OpVT.getScalarType() && "Concat vector type mismatch")(static_cast <bool> (SVT == OpVT.getScalarType() && "Concat vector type mismatch") ? void (0) : __assert_fail ("SVT == OpVT.getScalarType() && \"Concat vector type mismatch\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15250, __extension__ __PRETTY_FUNCTION__)); | |||
15251 | Opnds.append(Op->op_begin(), Op->op_begin() + NumElts); | |||
15252 | } else { | |||
15253 | for (unsigned i = 0; i != NumElts; ++i) | |||
15254 | Opnds.push_back( | |||
15255 | DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinVT, Op.getOperand(i))); | |||
15256 | } | |||
15257 | } | |||
15258 | } | |||
15259 | ||||
15260 | assert(VT.getVectorNumElements() == Opnds.size() &&(static_cast <bool> (VT.getVectorNumElements() == Opnds .size() && "Concat vector type mismatch") ? void (0) : __assert_fail ("VT.getVectorNumElements() == Opnds.size() && \"Concat vector type mismatch\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15261, __extension__ __PRETTY_FUNCTION__)) | |||
15261 | "Concat vector type mismatch")(static_cast <bool> (VT.getVectorNumElements() == Opnds .size() && "Concat vector type mismatch") ? void (0) : __assert_fail ("VT.getVectorNumElements() == Opnds.size() && \"Concat vector type mismatch\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15261, __extension__ __PRETTY_FUNCTION__)); | |||
15262 | return DAG.getBuildVector(VT, SDLoc(N), Opnds); | |||
15263 | } | |||
15264 | ||||
15265 | // Fold CONCAT_VECTORS of only bitcast scalars (or undef) to BUILD_VECTOR. | |||
15266 | if (SDValue V = combineConcatVectorOfScalars(N, DAG)) | |||
15267 | return V; | |||
15268 | ||||
15269 | // Fold CONCAT_VECTORS of EXTRACT_SUBVECTOR (or undef) to VECTOR_SHUFFLE. | |||
15270 | if (Level < AfterLegalizeVectorOps && TLI.isTypeLegal(VT)) | |||
15271 | if (SDValue V = combineConcatVectorOfExtracts(N, DAG)) | |||
15272 | return V; | |||
15273 | ||||
15274 | // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR | |||
15275 | // nodes often generate nop CONCAT_VECTOR nodes. | |||
15276 | // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that | |||
15277 | // place the incoming vectors at the exact same location. | |||
15278 | SDValue SingleSource = SDValue(); | |||
15279 | unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements(); | |||
15280 | ||||
15281 | for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { | |||
15282 | SDValue Op = N->getOperand(i); | |||
15283 | ||||
15284 | if (Op.isUndef()) | |||
15285 | continue; | |||
15286 | ||||
15287 | // Check if this is the identity extract: | |||
15288 | if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR) | |||
15289 | return SDValue(); | |||
15290 | ||||
15291 | // Find the single incoming vector for the extract_subvector. | |||
15292 | if (SingleSource.getNode()) { | |||
15293 | if (Op.getOperand(0) != SingleSource) | |||
15294 | return SDValue(); | |||
15295 | } else { | |||
15296 | SingleSource = Op.getOperand(0); | |||
15297 | ||||
15298 | // Check the source type is the same as the type of the result. | |||
15299 | // If not, this concat may extend the vector, so we can not | |||
15300 | // optimize it away. | |||
15301 | if (SingleSource.getValueType() != N->getValueType(0)) | |||
15302 | return SDValue(); | |||
15303 | } | |||
15304 | ||||
15305 | unsigned IdentityIndex = i * PartNumElem; | |||
15306 | ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1)); | |||
15307 | // The extract index must be constant. | |||
15308 | if (!CS) | |||
15309 | return SDValue(); | |||
15310 | ||||
15311 | // Check that we are reading from the identity index. | |||
15312 | if (CS->getZExtValue() != IdentityIndex) | |||
15313 | return SDValue(); | |||
15314 | } | |||
15315 | ||||
15316 | if (SingleSource.getNode()) | |||
15317 | return SingleSource; | |||
15318 | ||||
15319 | return SDValue(); | |||
15320 | } | |||
15321 | ||||
15322 | /// If we are extracting a subvector produced by a wide binary operator with at | |||
15323 | /// at least one operand that was the result of a vector concatenation, then try | |||
15324 | /// to use the narrow vector operands directly to avoid the concatenation and | |||
15325 | /// extraction. | |||
15326 | static SDValue narrowExtractedVectorBinOp(SDNode *Extract, SelectionDAG &DAG) { | |||
15327 | // TODO: Refactor with the caller (visitEXTRACT_SUBVECTOR), so we can share | |||
15328 | // some of these bailouts with other transforms. | |||
15329 | ||||
15330 | // The extract index must be a constant, so we can map it to a concat operand. | |||
15331 | auto *ExtractIndex = dyn_cast<ConstantSDNode>(Extract->getOperand(1)); | |||
15332 | if (!ExtractIndex) | |||
15333 | return SDValue(); | |||
15334 | ||||
15335 | // Only handle the case where we are doubling and then halving. A larger ratio | |||
15336 | // may require more than two narrow binops to replace the wide binop. | |||
15337 | EVT VT = Extract->getValueType(0); | |||
15338 | unsigned NumElems = VT.getVectorNumElements(); | |||
15339 | assert((ExtractIndex->getZExtValue() % NumElems) == 0 &&(static_cast <bool> ((ExtractIndex->getZExtValue() % NumElems) == 0 && "Extract index is not a multiple of the vector length." ) ? void (0) : __assert_fail ("(ExtractIndex->getZExtValue() % NumElems) == 0 && \"Extract index is not a multiple of the vector length.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15340, __extension__ __PRETTY_FUNCTION__)) | |||
15340 | "Extract index is not a multiple of the vector length.")(static_cast <bool> ((ExtractIndex->getZExtValue() % NumElems) == 0 && "Extract index is not a multiple of the vector length." ) ? void (0) : __assert_fail ("(ExtractIndex->getZExtValue() % NumElems) == 0 && \"Extract index is not a multiple of the vector length.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15340, __extension__ __PRETTY_FUNCTION__)); | |||
15341 | if (Extract->getOperand(0).getValueSizeInBits() != VT.getSizeInBits() * 2) | |||
15342 | return SDValue(); | |||
15343 | ||||
15344 | // We are looking for an optionally bitcasted wide vector binary operator | |||
15345 | // feeding an extract subvector. | |||
15346 | SDValue BinOp = peekThroughBitcast(Extract->getOperand(0)); | |||
15347 | ||||
15348 | // TODO: The motivating case for this transform is an x86 AVX1 target. That | |||
15349 | // target has temptingly almost legal versions of bitwise logic ops in 256-bit | |||
15350 | // flavors, but no other 256-bit integer support. This could be extended to | |||
15351 | // handle any binop, but that may require fixing/adding other folds to avoid | |||
15352 | // codegen regressions. | |||
15353 | unsigned BOpcode = BinOp.getOpcode(); | |||
15354 | if (BOpcode != ISD::AND && BOpcode != ISD::OR && BOpcode != ISD::XOR) | |||
15355 | return SDValue(); | |||
15356 | ||||
15357 | // The binop must be a vector type, so we can chop it in half. | |||
15358 | EVT WideBVT = BinOp.getValueType(); | |||
15359 | if (!WideBVT.isVector()) | |||
15360 | return SDValue(); | |||
15361 | ||||
15362 | // Bail out if the target does not support a narrower version of the binop. | |||
15363 | EVT NarrowBVT = EVT::getVectorVT(*DAG.getContext(), WideBVT.getScalarType(), | |||
15364 | WideBVT.getVectorNumElements() / 2); | |||
15365 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
15366 | if (!TLI.isOperationLegalOrCustomOrPromote(BOpcode, NarrowBVT)) | |||
15367 | return SDValue(); | |||
15368 | ||||
15369 | // Peek through bitcasts of the binary operator operands if needed. | |||
15370 | SDValue LHS = peekThroughBitcast(BinOp.getOperand(0)); | |||
15371 | SDValue RHS = peekThroughBitcast(BinOp.getOperand(1)); | |||
15372 | ||||
15373 | // We need at least one concatenation operation of a binop operand to make | |||
15374 | // this transform worthwhile. The concat must double the input vector sizes. | |||
15375 | // TODO: Should we also handle INSERT_SUBVECTOR patterns? | |||
15376 | bool ConcatL = | |||
15377 | LHS.getOpcode() == ISD::CONCAT_VECTORS && LHS.getNumOperands() == 2; | |||
15378 | bool ConcatR = | |||
15379 | RHS.getOpcode() == ISD::CONCAT_VECTORS && RHS.getNumOperands() == 2; | |||
15380 | if (!ConcatL && !ConcatR) | |||
15381 | return SDValue(); | |||
15382 | ||||
15383 | // If one of the binop operands was not the result of a concat, we must | |||
15384 | // extract a half-sized operand for our new narrow binop. We can't just reuse | |||
15385 | // the original extract index operand because we may have bitcasted. | |||
15386 | unsigned ConcatOpNum = ExtractIndex->getZExtValue() / NumElems; | |||
15387 | unsigned ExtBOIdx = ConcatOpNum * NarrowBVT.getVectorNumElements(); | |||
15388 | EVT ExtBOIdxVT = Extract->getOperand(1).getValueType(); | |||
15389 | SDLoc DL(Extract); | |||
15390 | ||||
15391 | // extract (binop (concat X1, X2), (concat Y1, Y2)), N --> binop XN, YN | |||
15392 | // extract (binop (concat X1, X2), Y), N --> binop XN, (extract Y, N) | |||
15393 | // extract (binop X, (concat Y1, Y2)), N --> binop (extract X, N), YN | |||
15394 | SDValue X = ConcatL ? DAG.getBitcast(NarrowBVT, LHS.getOperand(ConcatOpNum)) | |||
15395 | : DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, NarrowBVT, | |||
15396 | BinOp.getOperand(0), | |||
15397 | DAG.getConstant(ExtBOIdx, DL, ExtBOIdxVT)); | |||
15398 | ||||
15399 | SDValue Y = ConcatR ? DAG.getBitcast(NarrowBVT, RHS.getOperand(ConcatOpNum)) | |||
15400 | : DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, NarrowBVT, | |||
15401 | BinOp.getOperand(1), | |||
15402 | DAG.getConstant(ExtBOIdx, DL, ExtBOIdxVT)); | |||
15403 | ||||
15404 | SDValue NarrowBinOp = DAG.getNode(BOpcode, DL, NarrowBVT, X, Y); | |||
15405 | return DAG.getBitcast(VT, NarrowBinOp); | |||
15406 | } | |||
15407 | ||||
15408 | /// If we are extracting a subvector from a wide vector load, convert to a | |||
15409 | /// narrow load to eliminate the extraction: | |||
15410 | /// (extract_subvector (load wide vector)) --> (load narrow vector) | |||
15411 | static SDValue narrowExtractedVectorLoad(SDNode *Extract, SelectionDAG &DAG) { | |||
15412 | // TODO: Add support for big-endian. The offset calculation must be adjusted. | |||
15413 | if (DAG.getDataLayout().isBigEndian()) | |||
15414 | return SDValue(); | |||
15415 | ||||
15416 | // TODO: The one-use check is overly conservative. Check the cost of the | |||
15417 | // extract instead or remove that condition entirely. | |||
15418 | auto *Ld = dyn_cast<LoadSDNode>(Extract->getOperand(0)); | |||
15419 | auto *ExtIdx = dyn_cast<ConstantSDNode>(Extract->getOperand(1)); | |||
15420 | if (!Ld || !Ld->hasOneUse() || Ld->getExtensionType() || Ld->isVolatile() || | |||
15421 | !ExtIdx) | |||
15422 | return SDValue(); | |||
15423 | ||||
15424 | // The narrow load will be offset from the base address of the old load if | |||
15425 | // we are extracting from something besides index 0 (little-endian). | |||
15426 | EVT VT = Extract->getValueType(0); | |||
15427 | SDLoc DL(Extract); | |||
15428 | SDValue BaseAddr = Ld->getOperand(1); | |||
15429 | unsigned Offset = ExtIdx->getZExtValue() * VT.getScalarType().getStoreSize(); | |||
15430 | ||||
15431 | // TODO: Use "BaseIndexOffset" to make this more effective. | |||
15432 | SDValue NewAddr = DAG.getMemBasePlusOffset(BaseAddr, Offset, DL); | |||
15433 | MachineFunction &MF = DAG.getMachineFunction(); | |||
15434 | MachineMemOperand *MMO = MF.getMachineMemOperand(Ld->getMemOperand(), Offset, | |||
15435 | VT.getStoreSize()); | |||
15436 | SDValue NewLd = DAG.getLoad(VT, DL, Ld->getChain(), NewAddr, MMO); | |||
15437 | DAG.makeEquivalentMemoryOrdering(Ld, NewLd); | |||
15438 | return NewLd; | |||
15439 | } | |||
15440 | ||||
15441 | SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) { | |||
15442 | EVT NVT = N->getValueType(0); | |||
15443 | SDValue V = N->getOperand(0); | |||
15444 | ||||
15445 | // Extract from UNDEF is UNDEF. | |||
15446 | if (V.isUndef()) | |||
15447 | return DAG.getUNDEF(NVT); | |||
15448 | ||||
15449 | if (TLI.isOperationLegalOrCustomOrPromote(ISD::LOAD, NVT)) | |||
15450 | if (SDValue NarrowLoad = narrowExtractedVectorLoad(N, DAG)) | |||
15451 | return NarrowLoad; | |||
15452 | ||||
15453 | // Combine: | |||
15454 | // (extract_subvec (concat V1, V2, ...), i) | |||
15455 | // Into: | |||
15456 | // Vi if possible | |||
15457 | // Only operand 0 is checked as 'concat' assumes all inputs of the same | |||
15458 | // type. | |||
15459 | if (V->getOpcode() == ISD::CONCAT_VECTORS && | |||
15460 | isa<ConstantSDNode>(N->getOperand(1)) && | |||
15461 | V->getOperand(0).getValueType() == NVT) { | |||
15462 | unsigned Idx = N->getConstantOperandVal(1); | |||
15463 | unsigned NumElems = NVT.getVectorNumElements(); | |||
15464 | assert((Idx % NumElems) == 0 &&(static_cast <bool> ((Idx % NumElems) == 0 && "IDX in concat is not a multiple of the result vector length." ) ? void (0) : __assert_fail ("(Idx % NumElems) == 0 && \"IDX in concat is not a multiple of the result vector length.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15465, __extension__ __PRETTY_FUNCTION__)) | |||
15465 | "IDX in concat is not a multiple of the result vector length.")(static_cast <bool> ((Idx % NumElems) == 0 && "IDX in concat is not a multiple of the result vector length." ) ? void (0) : __assert_fail ("(Idx % NumElems) == 0 && \"IDX in concat is not a multiple of the result vector length.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15465, __extension__ __PRETTY_FUNCTION__)); | |||
15466 | return V->getOperand(Idx / NumElems); | |||
15467 | } | |||
15468 | ||||
15469 | // Skip bitcasting | |||
15470 | V = peekThroughBitcast(V); | |||
15471 | ||||
15472 | // If the input is a build vector. Try to make a smaller build vector. | |||
15473 | if (V->getOpcode() == ISD::BUILD_VECTOR) { | |||
15474 | if (auto *Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))) { | |||
15475 | EVT InVT = V->getValueType(0); | |||
15476 | unsigned ExtractSize = NVT.getSizeInBits(); | |||
15477 | unsigned EltSize = InVT.getScalarSizeInBits(); | |||
15478 | // Only do this if we won't split any elements. | |||
15479 | if (ExtractSize % EltSize == 0) { | |||
15480 | unsigned NumElems = ExtractSize / EltSize; | |||
15481 | EVT ExtractVT = EVT::getVectorVT(*DAG.getContext(), | |||
15482 | InVT.getVectorElementType(), NumElems); | |||
15483 | if ((!LegalOperations || | |||
15484 | TLI.isOperationLegal(ISD::BUILD_VECTOR, ExtractVT)) && | |||
15485 | (!LegalTypes || TLI.isTypeLegal(ExtractVT))) { | |||
15486 | unsigned IdxVal = (Idx->getZExtValue() * NVT.getScalarSizeInBits()) / | |||
15487 | EltSize; | |||
15488 | ||||
15489 | // Extract the pieces from the original build_vector. | |||
15490 | SDValue BuildVec = DAG.getBuildVector(ExtractVT, SDLoc(N), | |||
15491 | makeArrayRef(V->op_begin() + IdxVal, | |||
15492 | NumElems)); | |||
15493 | return DAG.getBitcast(NVT, BuildVec); | |||
15494 | } | |||
15495 | } | |||
15496 | } | |||
15497 | } | |||
15498 | ||||
15499 | if (V->getOpcode() == ISD::INSERT_SUBVECTOR) { | |||
15500 | // Handle only simple case where vector being inserted and vector | |||
15501 | // being extracted are of same size. | |||
15502 | EVT SmallVT = V->getOperand(1).getValueType(); | |||
15503 | if (!NVT.bitsEq(SmallVT)) | |||
15504 | return SDValue(); | |||
15505 | ||||
15506 | // Only handle cases where both indexes are constants. | |||
15507 | ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1)); | |||
15508 | ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2)); | |||
15509 | ||||
15510 | if (InsIdx && ExtIdx) { | |||
15511 | // Combine: | |||
15512 | // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx) | |||
15513 | // Into: | |||
15514 | // indices are equal or bit offsets are equal => V1 | |||
15515 | // otherwise => (extract_subvec V1, ExtIdx) | |||
15516 | if (InsIdx->getZExtValue() * SmallVT.getScalarSizeInBits() == | |||
15517 | ExtIdx->getZExtValue() * NVT.getScalarSizeInBits()) | |||
15518 | return DAG.getBitcast(NVT, V->getOperand(1)); | |||
15519 | return DAG.getNode( | |||
15520 | ISD::EXTRACT_SUBVECTOR, SDLoc(N), NVT, | |||
15521 | DAG.getBitcast(N->getOperand(0).getValueType(), V->getOperand(0)), | |||
15522 | N->getOperand(1)); | |||
15523 | } | |||
15524 | } | |||
15525 | ||||
15526 | if (SDValue NarrowBOp = narrowExtractedVectorBinOp(N, DAG)) | |||
15527 | return NarrowBOp; | |||
15528 | ||||
15529 | return SDValue(); | |||
15530 | } | |||
15531 | ||||
15532 | // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat, | |||
15533 | // or turn a shuffle of a single concat into simpler shuffle then concat. | |||
15534 | static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) { | |||
15535 | EVT VT = N->getValueType(0); | |||
15536 | unsigned NumElts = VT.getVectorNumElements(); | |||
15537 | ||||
15538 | SDValue N0 = N->getOperand(0); | |||
15539 | SDValue N1 = N->getOperand(1); | |||
15540 | ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); | |||
15541 | ||||
15542 | SmallVector<SDValue, 4> Ops; | |||
15543 | EVT ConcatVT = N0.getOperand(0).getValueType(); | |||
15544 | unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements(); | |||
15545 | unsigned NumConcats = NumElts / NumElemsPerConcat; | |||
15546 | ||||
15547 | // Special case: shuffle(concat(A,B)) can be more efficiently represented | |||
15548 | // as concat(shuffle(A,B),UNDEF) if the shuffle doesn't set any of the high | |||
15549 | // half vector elements. | |||
15550 | if (NumElemsPerConcat * 2 == NumElts && N1.isUndef() && | |||
15551 | std::all_of(SVN->getMask().begin() + NumElemsPerConcat, | |||
15552 | SVN->getMask().end(), [](int i) { return i == -1; })) { | |||
15553 | N0 = DAG.getVectorShuffle(ConcatVT, SDLoc(N), N0.getOperand(0), N0.getOperand(1), | |||
15554 | makeArrayRef(SVN->getMask().begin(), NumElemsPerConcat)); | |||
15555 | N1 = DAG.getUNDEF(ConcatVT); | |||
15556 | return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, N0, N1); | |||
15557 | } | |||
15558 | ||||
15559 | // Look at every vector that's inserted. We're looking for exact | |||
15560 | // subvector-sized copies from a concatenated vector | |||
15561 | for (unsigned I = 0; I != NumConcats; ++I) { | |||
15562 | // Make sure we're dealing with a copy. | |||
15563 | unsigned Begin = I * NumElemsPerConcat; | |||
15564 | bool AllUndef = true, NoUndef = true; | |||
15565 | for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) { | |||
15566 | if (SVN->getMaskElt(J) >= 0) | |||
15567 | AllUndef = false; | |||
15568 | else | |||
15569 | NoUndef = false; | |||
15570 | } | |||
15571 | ||||
15572 | if (NoUndef) { | |||
15573 | if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0) | |||
15574 | return SDValue(); | |||
15575 | ||||
15576 | for (unsigned J = 1; J != NumElemsPerConcat; ++J) | |||
15577 | if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J)) | |||
15578 | return SDValue(); | |||
15579 | ||||
15580 | unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat; | |||
15581 | if (FirstElt < N0.getNumOperands()) | |||
15582 | Ops.push_back(N0.getOperand(FirstElt)); | |||
15583 | else | |||
15584 | Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands())); | |||
15585 | ||||
15586 | } else if (AllUndef) { | |||
15587 | Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType())); | |||
15588 | } else { // Mixed with general masks and undefs, can't do optimization. | |||
15589 | return SDValue(); | |||
15590 | } | |||
15591 | } | |||
15592 | ||||
15593 | return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops); | |||
15594 | } | |||
15595 | ||||
15596 | // Attempt to combine a shuffle of 2 inputs of 'scalar sources' - | |||
15597 | // BUILD_VECTOR or SCALAR_TO_VECTOR into a single BUILD_VECTOR. | |||
15598 | // | |||
15599 | // SHUFFLE(BUILD_VECTOR(), BUILD_VECTOR()) -> BUILD_VECTOR() is always | |||
15600 | // a simplification in some sense, but it isn't appropriate in general: some | |||
15601 | // BUILD_VECTORs are substantially cheaper than others. The general case | |||
15602 | // of a BUILD_VECTOR requires inserting each element individually (or | |||
15603 | // performing the equivalent in a temporary stack variable). A BUILD_VECTOR of | |||
15604 | // all constants is a single constant pool load. A BUILD_VECTOR where each | |||
15605 | // element is identical is a splat. A BUILD_VECTOR where most of the operands | |||
15606 | // are undef lowers to a small number of element insertions. | |||
15607 | // | |||
15608 | // To deal with this, we currently use a bunch of mostly arbitrary heuristics. | |||
15609 | // We don't fold shuffles where one side is a non-zero constant, and we don't | |||
15610 | // fold shuffles if the resulting (non-splat) BUILD_VECTOR would have duplicate | |||
15611 | // non-constant operands. This seems to work out reasonably well in practice. | |||
15612 | static SDValue combineShuffleOfScalars(ShuffleVectorSDNode *SVN, | |||
15613 | SelectionDAG &DAG, | |||
15614 | const TargetLowering &TLI) { | |||
15615 | EVT VT = SVN->getValueType(0); | |||
15616 | unsigned NumElts = VT.getVectorNumElements(); | |||
15617 | SDValue N0 = SVN->getOperand(0); | |||
15618 | SDValue N1 = SVN->getOperand(1); | |||
15619 | ||||
15620 | if (!N0->hasOneUse() || !N1->hasOneUse()) | |||
15621 | return SDValue(); | |||
15622 | ||||
15623 | // If only one of N1,N2 is constant, bail out if it is not ALL_ZEROS as | |||
15624 | // discussed above. | |||
15625 | if (!N1.isUndef()) { | |||
15626 | bool N0AnyConst = isAnyConstantBuildVector(N0.getNode()); | |||
15627 | bool N1AnyConst = isAnyConstantBuildVector(N1.getNode()); | |||
15628 | if (N0AnyConst && !N1AnyConst && !ISD::isBuildVectorAllZeros(N0.getNode())) | |||
15629 | return SDValue(); | |||
15630 | if (!N0AnyConst && N1AnyConst && !ISD::isBuildVectorAllZeros(N1.getNode())) | |||
15631 | return SDValue(); | |||
15632 | } | |||
15633 | ||||
15634 | // If both inputs are splats of the same value then we can safely merge this | |||
15635 | // to a single BUILD_VECTOR with undef elements based on the shuffle mask. | |||
15636 | bool IsSplat = false; | |||
15637 | auto *BV0 = dyn_cast<BuildVectorSDNode>(N0); | |||
15638 | auto *BV1 = dyn_cast<BuildVectorSDNode>(N1); | |||
15639 | if (BV0 && BV1) | |||
15640 | if (SDValue Splat0 = BV0->getSplatValue()) | |||
15641 | IsSplat = (Splat0 == BV1->getSplatValue()); | |||
15642 | ||||
15643 | SmallVector<SDValue, 8> Ops; | |||
15644 | SmallSet<SDValue, 16> DuplicateOps; | |||
15645 | for (int M : SVN->getMask()) { | |||
15646 | SDValue Op = DAG.getUNDEF(VT.getScalarType()); | |||
15647 | if (M >= 0) { | |||
15648 | int Idx = M < (int)NumElts ? M : M - NumElts; | |||
15649 | SDValue &S = (M < (int)NumElts ? N0 : N1); | |||
15650 | if (S.getOpcode() == ISD::BUILD_VECTOR) { | |||
15651 | Op = S.getOperand(Idx); | |||
15652 | } else if (S.getOpcode() == ISD::SCALAR_TO_VECTOR) { | |||
15653 | assert(Idx == 0 && "Unexpected SCALAR_TO_VECTOR operand index.")(static_cast <bool> (Idx == 0 && "Unexpected SCALAR_TO_VECTOR operand index." ) ? void (0) : __assert_fail ("Idx == 0 && \"Unexpected SCALAR_TO_VECTOR operand index.\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15653, __extension__ __PRETTY_FUNCTION__)); | |||
15654 | Op = S.getOperand(0); | |||
15655 | } else { | |||
15656 | // Operand can't be combined - bail out. | |||
15657 | return SDValue(); | |||
15658 | } | |||
15659 | } | |||
15660 | ||||
15661 | // Don't duplicate a non-constant BUILD_VECTOR operand unless we're | |||
15662 | // generating a splat; semantically, this is fine, but it's likely to | |||
15663 | // generate low-quality code if the target can't reconstruct an appropriate | |||
15664 | // shuffle. | |||
15665 | if (!Op.isUndef() && !isa<ConstantSDNode>(Op) && !isa<ConstantFPSDNode>(Op)) | |||
15666 | if (!IsSplat && !DuplicateOps.insert(Op).second) | |||
15667 | return SDValue(); | |||
15668 | ||||
15669 | Ops.push_back(Op); | |||
15670 | } | |||
15671 | ||||
15672 | // BUILD_VECTOR requires all inputs to be of the same type, find the | |||
15673 | // maximum type and extend them all. | |||
15674 | EVT SVT = VT.getScalarType(); | |||
15675 | if (SVT.isInteger()) | |||
15676 | for (SDValue &Op : Ops) | |||
15677 | SVT = (SVT.bitsLT(Op.getValueType()) ? Op.getValueType() : SVT); | |||
15678 | if (SVT != VT.getScalarType()) | |||
15679 | for (SDValue &Op : Ops) | |||
15680 | Op = TLI.isZExtFree(Op.getValueType(), SVT) | |||
15681 | ? DAG.getZExtOrTrunc(Op, SDLoc(SVN), SVT) | |||
15682 | : DAG.getSExtOrTrunc(Op, SDLoc(SVN), SVT); | |||
15683 | return DAG.getBuildVector(VT, SDLoc(SVN), Ops); | |||
15684 | } | |||
15685 | ||||
15686 | // Match shuffles that can be converted to any_vector_extend_in_reg. | |||
15687 | // This is often generated during legalization. | |||
15688 | // e.g. v4i32 <0,u,1,u> -> (v2i64 any_vector_extend_in_reg(v4i32 src)) | |||
15689 | // TODO Add support for ZERO_EXTEND_VECTOR_INREG when we have a test case. | |||
15690 | static SDValue combineShuffleToVectorExtend(ShuffleVectorSDNode *SVN, | |||
15691 | SelectionDAG &DAG, | |||
15692 | const TargetLowering &TLI, | |||
15693 | bool LegalOperations, | |||
15694 | bool LegalTypes) { | |||
15695 | EVT VT = SVN->getValueType(0); | |||
15696 | bool IsBigEndian = DAG.getDataLayout().isBigEndian(); | |||
15697 | ||||
15698 | // TODO Add support for big-endian when we have a test case. | |||
15699 | if (!VT.isInteger() || IsBigEndian) | |||
15700 | return SDValue(); | |||
15701 | ||||
15702 | unsigned NumElts = VT.getVectorNumElements(); | |||
15703 | unsigned EltSizeInBits = VT.getScalarSizeInBits(); | |||
15704 | ArrayRef<int> Mask = SVN->getMask(); | |||
15705 | SDValue N0 = SVN->getOperand(0); | |||
15706 | ||||
15707 | // shuffle<0,-1,1,-1> == (v2i64 anyextend_vector_inreg(v4i32)) | |||
15708 | auto isAnyExtend = [&Mask, &NumElts](unsigned Scale) { | |||
15709 | for (unsigned i = 0; i != NumElts; ++i) { | |||
15710 | if (Mask[i] < 0) | |||
15711 | continue; | |||
15712 | if ((i % Scale) == 0 && Mask[i] == (int)(i / Scale)) | |||
15713 | continue; | |||
15714 | return false; | |||
15715 | } | |||
15716 | return true; | |||
15717 | }; | |||
15718 | ||||
15719 | // Attempt to match a '*_extend_vector_inreg' shuffle, we just search for | |||
15720 | // power-of-2 extensions as they are the most likely. | |||
15721 | for (unsigned Scale = 2; Scale < NumElts; Scale *= 2) { | |||
15722 | // Check for non power of 2 vector sizes | |||
15723 | if (NumElts % Scale != 0) | |||
15724 | continue; | |||
15725 | if (!isAnyExtend(Scale)) | |||
15726 | continue; | |||
15727 | ||||
15728 | EVT OutSVT = EVT::getIntegerVT(*DAG.getContext(), EltSizeInBits * Scale); | |||
15729 | EVT OutVT = EVT::getVectorVT(*DAG.getContext(), OutSVT, NumElts / Scale); | |||
15730 | if (!LegalTypes || TLI.isTypeLegal(OutVT)) | |||
15731 | if (!LegalOperations || | |||
15732 | TLI.isOperationLegalOrCustom(ISD::ANY_EXTEND_VECTOR_INREG, OutVT)) | |||
15733 | return DAG.getBitcast(VT, | |||
15734 | DAG.getAnyExtendVectorInReg(N0, SDLoc(SVN), OutVT)); | |||
15735 | } | |||
15736 | ||||
15737 | return SDValue(); | |||
15738 | } | |||
15739 | ||||
15740 | // Detect 'truncate_vector_inreg' style shuffles that pack the lower parts of | |||
15741 | // each source element of a large type into the lowest elements of a smaller | |||
15742 | // destination type. This is often generated during legalization. | |||
15743 | // If the source node itself was a '*_extend_vector_inreg' node then we should | |||
15744 | // then be able to remove it. | |||
15745 | static SDValue combineTruncationShuffle(ShuffleVectorSDNode *SVN, | |||
15746 | SelectionDAG &DAG) { | |||
15747 | EVT VT = SVN->getValueType(0); | |||
15748 | bool IsBigEndian = DAG.getDataLayout().isBigEndian(); | |||
15749 | ||||
15750 | // TODO Add support for big-endian when we have a test case. | |||
15751 | if (!VT.isInteger() || IsBigEndian) | |||
15752 | return SDValue(); | |||
15753 | ||||
15754 | SDValue N0 = peekThroughBitcast(SVN->getOperand(0)); | |||
15755 | ||||
15756 | unsigned Opcode = N0.getOpcode(); | |||
15757 | if (Opcode != ISD::ANY_EXTEND_VECTOR_INREG && | |||
15758 | Opcode != ISD::SIGN_EXTEND_VECTOR_INREG && | |||
15759 | Opcode != ISD::ZERO_EXTEND_VECTOR_INREG) | |||
15760 | return SDValue(); | |||
15761 | ||||
15762 | SDValue N00 = N0.getOperand(0); | |||
15763 | ArrayRef<int> Mask = SVN->getMask(); | |||
15764 | unsigned NumElts = VT.getVectorNumElements(); | |||
15765 | unsigned EltSizeInBits = VT.getScalarSizeInBits(); | |||
15766 | unsigned ExtSrcSizeInBits = N00.getScalarValueSizeInBits(); | |||
15767 | unsigned ExtDstSizeInBits = N0.getScalarValueSizeInBits(); | |||
15768 | ||||
15769 | if (ExtDstSizeInBits % ExtSrcSizeInBits != 0) | |||
15770 | return SDValue(); | |||
15771 | unsigned ExtScale = ExtDstSizeInBits / ExtSrcSizeInBits; | |||
15772 | ||||
15773 | // (v4i32 truncate_vector_inreg(v2i64)) == shuffle<0,2-1,-1> | |||
15774 | // (v8i16 truncate_vector_inreg(v4i32)) == shuffle<0,2,4,6,-1,-1,-1,-1> | |||
15775 | // (v8i16 truncate_vector_inreg(v2i64)) == shuffle<0,4,-1,-1,-1,-1,-1,-1> | |||
15776 | auto isTruncate = [&Mask, &NumElts](unsigned Scale) { | |||
15777 | for (unsigned i = 0; i != NumElts; ++i) { | |||
15778 | if (Mask[i] < 0) | |||
15779 | continue; | |||
15780 | if ((i * Scale) < NumElts && Mask[i] == (int)(i * Scale)) | |||
15781 | continue; | |||
15782 | return false; | |||
15783 | } | |||
15784 | return true; | |||
15785 | }; | |||
15786 | ||||
15787 | // At the moment we just handle the case where we've truncated back to the | |||
15788 | // same size as before the extension. | |||
15789 | // TODO: handle more extension/truncation cases as cases arise. | |||
15790 | if (EltSizeInBits != ExtSrcSizeInBits) | |||
15791 | return SDValue(); | |||
15792 | ||||
15793 | // We can remove *extend_vector_inreg only if the truncation happens at | |||
15794 | // the same scale as the extension. | |||
15795 | if (isTruncate(ExtScale)) | |||
15796 | return DAG.getBitcast(VT, N00); | |||
15797 | ||||
15798 | return SDValue(); | |||
15799 | } | |||
15800 | ||||
15801 | // Combine shuffles of splat-shuffles of the form: | |||
15802 | // shuffle (shuffle V, undef, splat-mask), undef, M | |||
15803 | // If splat-mask contains undef elements, we need to be careful about | |||
15804 | // introducing undef's in the folded mask which are not the result of composing | |||
15805 | // the masks of the shuffles. | |||
15806 | static SDValue combineShuffleOfSplat(ArrayRef<int> UserMask, | |||
15807 | ShuffleVectorSDNode *Splat, | |||
15808 | SelectionDAG &DAG) { | |||
15809 | ArrayRef<int> SplatMask = Splat->getMask(); | |||
15810 | assert(UserMask.size() == SplatMask.size() && "Mask length mismatch")(static_cast <bool> (UserMask.size() == SplatMask.size( ) && "Mask length mismatch") ? void (0) : __assert_fail ("UserMask.size() == SplatMask.size() && \"Mask length mismatch\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15810, __extension__ __PRETTY_FUNCTION__)); | |||
15811 | ||||
15812 | // Prefer simplifying to the splat-shuffle, if possible. This is legal if | |||
15813 | // every undef mask element in the splat-shuffle has a corresponding undef | |||
15814 | // element in the user-shuffle's mask or if the composition of mask elements | |||
15815 | // would result in undef. | |||
15816 | // Examples for (shuffle (shuffle v, undef, SplatMask), undef, UserMask): | |||
15817 | // * UserMask=[0,2,u,u], SplatMask=[2,u,2,u] -> [2,2,u,u] | |||
15818 | // In this case it is not legal to simplify to the splat-shuffle because we | |||
15819 | // may be exposing the users of the shuffle an undef element at index 1 | |||
15820 | // which was not there before the combine. | |||
15821 | // * UserMask=[0,u,2,u], SplatMask=[2,u,2,u] -> [2,u,2,u] | |||
15822 | // In this case the composition of masks yields SplatMask, so it's ok to | |||
15823 | // simplify to the splat-shuffle. | |||
15824 | // * UserMask=[3,u,2,u], SplatMask=[2,u,2,u] -> [u,u,2,u] | |||
15825 | // In this case the composed mask includes all undef elements of SplatMask | |||
15826 | // and in addition sets element zero to undef. It is safe to simplify to | |||
15827 | // the splat-shuffle. | |||
15828 | auto CanSimplifyToExistingSplat = [](ArrayRef<int> UserMask, | |||
15829 | ArrayRef<int> SplatMask) { | |||
15830 | for (unsigned i = 0, e = UserMask.size(); i != e; ++i) | |||
15831 | if (UserMask[i] != -1 && SplatMask[i] == -1 && | |||
15832 | SplatMask[UserMask[i]] != -1) | |||
15833 | return false; | |||
15834 | return true; | |||
15835 | }; | |||
15836 | if (CanSimplifyToExistingSplat(UserMask, SplatMask)) | |||
15837 | return SDValue(Splat, 0); | |||
15838 | ||||
15839 | // Create a new shuffle with a mask that is composed of the two shuffles' | |||
15840 | // masks. | |||
15841 | SmallVector<int, 32> NewMask; | |||
15842 | for (int Idx : UserMask) | |||
15843 | NewMask.push_back(Idx == -1 ? -1 : SplatMask[Idx]); | |||
15844 | ||||
15845 | return DAG.getVectorShuffle(Splat->getValueType(0), SDLoc(Splat), | |||
15846 | Splat->getOperand(0), Splat->getOperand(1), | |||
15847 | NewMask); | |||
15848 | } | |||
15849 | ||||
15850 | /// If the shuffle mask is taking exactly one element from the first vector | |||
15851 | /// operand and passing through all other elements from the second vector | |||
15852 | /// operand, return the index of the mask element that is choosing an element | |||
15853 | /// from the first operand. Otherwise, return -1. | |||
15854 | static int getShuffleMaskIndexOfOneElementFromOp0IntoOp1(ArrayRef<int> Mask) { | |||
15855 | int MaskSize = Mask.size(); | |||
15856 | int EltFromOp0 = -1; | |||
15857 | // TODO: This does not match if there are undef elements in the shuffle mask. | |||
15858 | // Should we ignore undefs in the shuffle mask instead? The trade-off is | |||
15859 | // removing an instruction (a shuffle), but losing the knowledge that some | |||
15860 | // vector lanes are not needed. | |||
15861 | for (int i = 0; i != MaskSize; ++i) { | |||
15862 | if (Mask[i] >= 0 && Mask[i] < MaskSize) { | |||
15863 | // We're looking for a shuffle of exactly one element from operand 0. | |||
15864 | if (EltFromOp0 != -1) | |||
15865 | return -1; | |||
15866 | EltFromOp0 = i; | |||
15867 | } else if (Mask[i] != i + MaskSize) { | |||
15868 | // Nothing from operand 1 can change lanes. | |||
15869 | return -1; | |||
15870 | } | |||
15871 | } | |||
15872 | return EltFromOp0; | |||
15873 | } | |||
15874 | ||||
15875 | /// If a shuffle inserts exactly one element from a source vector operand into | |||
15876 | /// another vector operand and we can access the specified element as a scalar, | |||
15877 | /// then we can eliminate the shuffle. | |||
15878 | static SDValue replaceShuffleOfInsert(ShuffleVectorSDNode *Shuf, | |||
15879 | SelectionDAG &DAG) { | |||
15880 | // First, check if we are taking one element of a vector and shuffling that | |||
15881 | // element into another vector. | |||
15882 | ArrayRef<int> Mask = Shuf->getMask(); | |||
15883 | SmallVector<int, 16> CommutedMask(Mask.begin(), Mask.end()); | |||
15884 | SDValue Op0 = Shuf->getOperand(0); | |||
15885 | SDValue Op1 = Shuf->getOperand(1); | |||
15886 | int ShufOp0Index = getShuffleMaskIndexOfOneElementFromOp0IntoOp1(Mask); | |||
15887 | if (ShufOp0Index == -1) { | |||
15888 | // Commute mask and check again. | |||
15889 | ShuffleVectorSDNode::commuteMask(CommutedMask); | |||
15890 | ShufOp0Index = getShuffleMaskIndexOfOneElementFromOp0IntoOp1(CommutedMask); | |||
15891 | if (ShufOp0Index == -1) | |||
15892 | return SDValue(); | |||
15893 | // Commute operands to match the commuted shuffle mask. | |||
15894 | std::swap(Op0, Op1); | |||
15895 | Mask = CommutedMask; | |||
15896 | } | |||
15897 | ||||
15898 | // The shuffle inserts exactly one element from operand 0 into operand 1. | |||
15899 | // Now see if we can access that element as a scalar via a real insert element | |||
15900 | // instruction. | |||
15901 | // TODO: We can try harder to locate the element as a scalar. Examples: it | |||
15902 | // could be an operand of SCALAR_TO_VECTOR, BUILD_VECTOR, or a constant. | |||
15903 | assert(Mask[ShufOp0Index] >= 0 && Mask[ShufOp0Index] < (int)Mask.size() &&(static_cast <bool> (Mask[ShufOp0Index] >= 0 && Mask[ShufOp0Index] < (int)Mask.size() && "Shuffle mask value must be from operand 0" ) ? void (0) : __assert_fail ("Mask[ShufOp0Index] >= 0 && Mask[ShufOp0Index] < (int)Mask.size() && \"Shuffle mask value must be from operand 0\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15904, __extension__ __PRETTY_FUNCTION__)) | |||
15904 | "Shuffle mask value must be from operand 0")(static_cast <bool> (Mask[ShufOp0Index] >= 0 && Mask[ShufOp0Index] < (int)Mask.size() && "Shuffle mask value must be from operand 0" ) ? void (0) : __assert_fail ("Mask[ShufOp0Index] >= 0 && Mask[ShufOp0Index] < (int)Mask.size() && \"Shuffle mask value must be from operand 0\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15904, __extension__ __PRETTY_FUNCTION__)); | |||
15905 | if (Op0.getOpcode() != ISD::INSERT_VECTOR_ELT) | |||
15906 | return SDValue(); | |||
15907 | ||||
15908 | auto *InsIndexC = dyn_cast<ConstantSDNode>(Op0.getOperand(2)); | |||
15909 | if (!InsIndexC || InsIndexC->getSExtValue() != Mask[ShufOp0Index]) | |||
15910 | return SDValue(); | |||
15911 | ||||
15912 | // There's an existing insertelement with constant insertion index, so we | |||
15913 | // don't need to check the legality/profitability of a replacement operation | |||
15914 | // that differs at most in the constant value. The target should be able to | |||
15915 | // lower any of those in a similar way. If not, legalization will expand this | |||
15916 | // to a scalar-to-vector plus shuffle. | |||
15917 | // | |||
15918 | // Note that the shuffle may move the scalar from the position that the insert | |||
15919 | // element used. Therefore, our new insert element occurs at the shuffle's | |||
15920 | // mask index value, not the insert's index value. | |||
15921 | // shuffle (insertelt v1, x, C), v2, mask --> insertelt v2, x, C' | |||
15922 | SDValue NewInsIndex = DAG.getConstant(ShufOp0Index, SDLoc(Shuf), | |||
15923 | Op0.getOperand(2).getValueType()); | |||
15924 | return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(Shuf), Op0.getValueType(), | |||
15925 | Op1, Op0.getOperand(1), NewInsIndex); | |||
15926 | } | |||
15927 | ||||
15928 | SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) { | |||
15929 | EVT VT = N->getValueType(0); | |||
15930 | unsigned NumElts = VT.getVectorNumElements(); | |||
15931 | ||||
15932 | SDValue N0 = N->getOperand(0); | |||
15933 | SDValue N1 = N->getOperand(1); | |||
15934 | ||||
15935 | assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG")(static_cast <bool> (N0.getValueType() == VT && "Vector shuffle must be normalized in DAG") ? void (0) : __assert_fail ("N0.getValueType() == VT && \"Vector shuffle must be normalized in DAG\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15935, __extension__ __PRETTY_FUNCTION__)); | |||
15936 | ||||
15937 | // Canonicalize shuffle undef, undef -> undef | |||
15938 | if (N0.isUndef() && N1.isUndef()) | |||
15939 | return DAG.getUNDEF(VT); | |||
15940 | ||||
15941 | ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); | |||
15942 | ||||
15943 | // Canonicalize shuffle v, v -> v, undef | |||
15944 | if (N0 == N1) { | |||
15945 | SmallVector<int, 8> NewMask; | |||
15946 | for (unsigned i = 0; i != NumElts; ++i) { | |||
15947 | int Idx = SVN->getMaskElt(i); | |||
15948 | if (Idx >= (int)NumElts) Idx -= NumElts; | |||
15949 | NewMask.push_back(Idx); | |||
15950 | } | |||
15951 | return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT), NewMask); | |||
15952 | } | |||
15953 | ||||
15954 | // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask. | |||
15955 | if (N0.isUndef()) | |||
15956 | return DAG.getCommutedVectorShuffle(*SVN); | |||
15957 | ||||
15958 | // Remove references to rhs if it is undef | |||
15959 | if (N1.isUndef()) { | |||
15960 | bool Changed = false; | |||
15961 | SmallVector<int, 8> NewMask; | |||
15962 | for (unsigned i = 0; i != NumElts; ++i) { | |||
15963 | int Idx = SVN->getMaskElt(i); | |||
15964 | if (Idx >= (int)NumElts) { | |||
15965 | Idx = -1; | |||
15966 | Changed = true; | |||
15967 | } | |||
15968 | NewMask.push_back(Idx); | |||
15969 | } | |||
15970 | if (Changed) | |||
15971 | return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, NewMask); | |||
15972 | } | |||
15973 | ||||
15974 | if (SDValue InsElt = replaceShuffleOfInsert(SVN, DAG)) | |||
15975 | return InsElt; | |||
15976 | ||||
15977 | // A shuffle of a single vector that is a splat can always be folded. | |||
15978 | if (auto *N0Shuf = dyn_cast<ShuffleVectorSDNode>(N0)) | |||
15979 | if (N1->isUndef() && N0Shuf->isSplat()) | |||
15980 | return combineShuffleOfSplat(SVN->getMask(), N0Shuf, DAG); | |||
15981 | ||||
15982 | // If it is a splat, check if the argument vector is another splat or a | |||
15983 | // build_vector. | |||
15984 | if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) { | |||
15985 | SDNode *V = N0.getNode(); | |||
15986 | ||||
15987 | // If this is a bit convert that changes the element type of the vector but | |||
15988 | // not the number of vector elements, look through it. Be careful not to | |||
15989 | // look though conversions that change things like v4f32 to v2f64. | |||
15990 | if (V->getOpcode() == ISD::BITCAST) { | |||
15991 | SDValue ConvInput = V->getOperand(0); | |||
15992 | if (ConvInput.getValueType().isVector() && | |||
15993 | ConvInput.getValueType().getVectorNumElements() == NumElts) | |||
15994 | V = ConvInput.getNode(); | |||
15995 | } | |||
15996 | ||||
15997 | if (V->getOpcode() == ISD::BUILD_VECTOR) { | |||
15998 | assert(V->getNumOperands() == NumElts &&(static_cast <bool> (V->getNumOperands() == NumElts && "BUILD_VECTOR has wrong number of operands") ? void (0) : __assert_fail ("V->getNumOperands() == NumElts && \"BUILD_VECTOR has wrong number of operands\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15999, __extension__ __PRETTY_FUNCTION__)) | |||
15999 | "BUILD_VECTOR has wrong number of operands")(static_cast <bool> (V->getNumOperands() == NumElts && "BUILD_VECTOR has wrong number of operands") ? void (0) : __assert_fail ("V->getNumOperands() == NumElts && \"BUILD_VECTOR has wrong number of operands\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 15999, __extension__ __PRETTY_FUNCTION__)); | |||
16000 | SDValue Base; | |||
16001 | bool AllSame = true; | |||
16002 | for (unsigned i = 0; i != NumElts; ++i) { | |||
16003 | if (!V->getOperand(i).isUndef()) { | |||
16004 | Base = V->getOperand(i); | |||
16005 | break; | |||
16006 | } | |||
16007 | } | |||
16008 | // Splat of <u, u, u, u>, return <u, u, u, u> | |||
16009 | if (!Base.getNode()) | |||
16010 | return N0; | |||
16011 | for (unsigned i = 0; i != NumElts; ++i) { | |||
16012 | if (V->getOperand(i) != Base) { | |||
16013 | AllSame = false; | |||
16014 | break; | |||
16015 | } | |||
16016 | } | |||
16017 | // Splat of <x, x, x, x>, return <x, x, x, x> | |||
16018 | if (AllSame) | |||
16019 | return N0; | |||
16020 | ||||
16021 | // Canonicalize any other splat as a build_vector. | |||
16022 | const SDValue &Splatted = V->getOperand(SVN->getSplatIndex()); | |||
16023 | SmallVector<SDValue, 8> Ops(NumElts, Splatted); | |||
16024 | SDValue NewBV = DAG.getBuildVector(V->getValueType(0), SDLoc(N), Ops); | |||
16025 | ||||
16026 | // We may have jumped through bitcasts, so the type of the | |||
16027 | // BUILD_VECTOR may not match the type of the shuffle. | |||
16028 | if (V->getValueType(0) != VT) | |||
16029 | NewBV = DAG.getBitcast(VT, NewBV); | |||
16030 | return NewBV; | |||
16031 | } | |||
16032 | } | |||
16033 | ||||
16034 | // Simplify source operands based on shuffle mask. | |||
16035 | if (SimplifyDemandedVectorElts(SDValue(N, 0))) | |||
16036 | return SDValue(N, 0); | |||
16037 | ||||
16038 | // Match shuffles that can be converted to any_vector_extend_in_reg. | |||
16039 | if (SDValue V = combineShuffleToVectorExtend(SVN, DAG, TLI, LegalOperations, LegalTypes)) | |||
16040 | return V; | |||
16041 | ||||
16042 | // Combine "truncate_vector_in_reg" style shuffles. | |||
16043 | if (SDValue V = combineTruncationShuffle(SVN, DAG)) | |||
16044 | return V; | |||
16045 | ||||
16046 | if (N0.getOpcode() == ISD::CONCAT_VECTORS && | |||
16047 | Level < AfterLegalizeVectorOps && | |||
16048 | (N1.isUndef() || | |||
16049 | (N1.getOpcode() == ISD::CONCAT_VECTORS && | |||
16050 | N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) { | |||
16051 | if (SDValue V = partitionShuffleOfConcats(N, DAG)) | |||
16052 | return V; | |||
16053 | } | |||
16054 | ||||
16055 | // Attempt to combine a shuffle of 2 inputs of 'scalar sources' - | |||
16056 | // BUILD_VECTOR or SCALAR_TO_VECTOR into a single BUILD_VECTOR. | |||
16057 | if (Level < AfterLegalizeVectorOps && TLI.isTypeLegal(VT)) | |||
16058 | if (SDValue Res = combineShuffleOfScalars(SVN, DAG, TLI)) | |||
16059 | return Res; | |||
16060 | ||||
16061 | // If this shuffle only has a single input that is a bitcasted shuffle, | |||
16062 | // attempt to merge the 2 shuffles and suitably bitcast the inputs/output | |||
16063 | // back to their original types. | |||
16064 | if (N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() && | |||
16065 | N1.isUndef() && Level < AfterLegalizeVectorOps && | |||
16066 | TLI.isTypeLegal(VT)) { | |||
16067 | ||||
16068 | // Peek through the bitcast only if there is one user. | |||
16069 | SDValue BC0 = N0; | |||
16070 | while (BC0.getOpcode() == ISD::BITCAST) { | |||
16071 | if (!BC0.hasOneUse()) | |||
16072 | break; | |||
16073 | BC0 = BC0.getOperand(0); | |||
16074 | } | |||
16075 | ||||
16076 | auto ScaleShuffleMask = [](ArrayRef<int> Mask, int Scale) { | |||
16077 | if (Scale == 1) | |||
16078 | return SmallVector<int, 8>(Mask.begin(), Mask.end()); | |||
16079 | ||||
16080 | SmallVector<int, 8> NewMask; | |||
16081 | for (int M : Mask) | |||
16082 | for (int s = 0; s != Scale; ++s) | |||
16083 | NewMask.push_back(M < 0 ? -1 : Scale * M + s); | |||
16084 | return NewMask; | |||
16085 | }; | |||
16086 | ||||
16087 | if (BC0.getOpcode() == ISD::VECTOR_SHUFFLE && BC0.hasOneUse()) { | |||
16088 | EVT SVT = VT.getScalarType(); | |||
16089 | EVT InnerVT = BC0->getValueType(0); | |||
16090 | EVT InnerSVT = InnerVT.getScalarType(); | |||
16091 | ||||
16092 | // Determine which shuffle works with the smaller scalar type. | |||
16093 | EVT ScaleVT = SVT.bitsLT(InnerSVT) ? VT : InnerVT; | |||
16094 | EVT ScaleSVT = ScaleVT.getScalarType(); | |||
16095 | ||||
16096 | if (TLI.isTypeLegal(ScaleVT) && | |||
16097 | 0 == (InnerSVT.getSizeInBits() % ScaleSVT.getSizeInBits()) && | |||
16098 | 0 == (SVT.getSizeInBits() % ScaleSVT.getSizeInBits())) { | |||
16099 | int InnerScale = InnerSVT.getSizeInBits() / ScaleSVT.getSizeInBits(); | |||
16100 | int OuterScale = SVT.getSizeInBits() / ScaleSVT.getSizeInBits(); | |||
16101 | ||||
16102 | // Scale the shuffle masks to the smaller scalar type. | |||
16103 | ShuffleVectorSDNode *InnerSVN = cast<ShuffleVectorSDNode>(BC0); | |||
16104 | SmallVector<int, 8> InnerMask = | |||
16105 | ScaleShuffleMask(InnerSVN->getMask(), InnerScale); | |||
16106 | SmallVector<int, 8> OuterMask = | |||
16107 | ScaleShuffleMask(SVN->getMask(), OuterScale); | |||
16108 | ||||
16109 | // Merge the shuffle masks. | |||
16110 | SmallVector<int, 8> NewMask; | |||
16111 | for (int M : OuterMask) | |||
16112 | NewMask.push_back(M < 0 ? -1 : InnerMask[M]); | |||
16113 | ||||
16114 | // Test for shuffle mask legality over both commutations. | |||
16115 | SDValue SV0 = BC0->getOperand(0); | |||
16116 | SDValue SV1 = BC0->getOperand(1); | |||
16117 | bool LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT); | |||
16118 | if (!LegalMask) { | |||
16119 | std::swap(SV0, SV1); | |||
16120 | ShuffleVectorSDNode::commuteMask(NewMask); | |||
16121 | LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT); | |||
16122 | } | |||
16123 | ||||
16124 | if (LegalMask) { | |||
16125 | SV0 = DAG.getBitcast(ScaleVT, SV0); | |||
16126 | SV1 = DAG.getBitcast(ScaleVT, SV1); | |||
16127 | return DAG.getBitcast( | |||
16128 | VT, DAG.getVectorShuffle(ScaleVT, SDLoc(N), SV0, SV1, NewMask)); | |||
16129 | } | |||
16130 | } | |||
16131 | } | |||
16132 | } | |||
16133 | ||||
16134 | // Canonicalize shuffles according to rules: | |||
16135 | // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A) | |||
16136 | // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B) | |||
16137 | // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B) | |||
16138 | if (N1.getOpcode() == ISD::VECTOR_SHUFFLE && | |||
16139 | N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG && | |||
16140 | TLI.isTypeLegal(VT)) { | |||
16141 | // The incoming shuffle must be of the same type as the result of the | |||
16142 | // current shuffle. | |||
16143 | assert(N1->getOperand(0).getValueType() == VT &&(static_cast <bool> (N1->getOperand(0).getValueType( ) == VT && "Shuffle types don't match") ? void (0) : __assert_fail ("N1->getOperand(0).getValueType() == VT && \"Shuffle types don't match\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 16144, __extension__ __PRETTY_FUNCTION__)) | |||
16144 | "Shuffle types don't match")(static_cast <bool> (N1->getOperand(0).getValueType( ) == VT && "Shuffle types don't match") ? void (0) : __assert_fail ("N1->getOperand(0).getValueType() == VT && \"Shuffle types don't match\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 16144, __extension__ __PRETTY_FUNCTION__)); | |||
16145 | ||||
16146 | SDValue SV0 = N1->getOperand(0); | |||
16147 | SDValue SV1 = N1->getOperand(1); | |||
16148 | bool HasSameOp0 = N0 == SV0; | |||
16149 | bool IsSV1Undef = SV1.isUndef(); | |||
16150 | if (HasSameOp0 || IsSV1Undef || N0 == SV1) | |||
16151 | // Commute the operands of this shuffle so that next rule | |||
16152 | // will trigger. | |||
16153 | return DAG.getCommutedVectorShuffle(*SVN); | |||
16154 | } | |||
16155 | ||||
16156 | // Try to fold according to rules: | |||
16157 | // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2) | |||
16158 | // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2) | |||
16159 | // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2) | |||
16160 | // Don't try to fold shuffles with illegal type. | |||
16161 | // Only fold if this shuffle is the only user of the other shuffle. | |||
16162 | if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && N->isOnlyUserOf(N0.getNode()) && | |||
16163 | Level < AfterLegalizeDAG && TLI.isTypeLegal(VT)) { | |||
16164 | ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0); | |||
16165 | ||||
16166 | // Don't try to fold splats; they're likely to simplify somehow, or they | |||
16167 | // might be free. | |||
16168 | if (OtherSV->isSplat()) | |||
16169 | return SDValue(); | |||
16170 | ||||
16171 | // The incoming shuffle must be of the same type as the result of the | |||
16172 | // current shuffle. | |||
16173 | assert(OtherSV->getOperand(0).getValueType() == VT &&(static_cast <bool> (OtherSV->getOperand(0).getValueType () == VT && "Shuffle types don't match") ? void (0) : __assert_fail ("OtherSV->getOperand(0).getValueType() == VT && \"Shuffle types don't match\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 16174, __extension__ __PRETTY_FUNCTION__)) | |||
16174 | "Shuffle types don't match")(static_cast <bool> (OtherSV->getOperand(0).getValueType () == VT && "Shuffle types don't match") ? void (0) : __assert_fail ("OtherSV->getOperand(0).getValueType() == VT && \"Shuffle types don't match\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 16174, __extension__ __PRETTY_FUNCTION__)); | |||
16175 | ||||
16176 | SDValue SV0, SV1; | |||
16177 | SmallVector<int, 4> Mask; | |||
16178 | // Compute the combined shuffle mask for a shuffle with SV0 as the first | |||
16179 | // operand, and SV1 as the second operand. | |||
16180 | for (unsigned i = 0; i != NumElts; ++i) { | |||
16181 | int Idx = SVN->getMaskElt(i); | |||
16182 | if (Idx < 0) { | |||
16183 | // Propagate Undef. | |||
16184 | Mask.push_back(Idx); | |||
16185 | continue; | |||
16186 | } | |||
16187 | ||||
16188 | SDValue CurrentVec; | |||
16189 | if (Idx < (int)NumElts) { | |||
16190 | // This shuffle index refers to the inner shuffle N0. Lookup the inner | |||
16191 | // shuffle mask to identify which vector is actually referenced. | |||
16192 | Idx = OtherSV->getMaskElt(Idx); | |||
16193 | if (Idx < 0) { | |||
16194 | // Propagate Undef. | |||
16195 | Mask.push_back(Idx); | |||
16196 | continue; | |||
16197 | } | |||
16198 | ||||
16199 | CurrentVec = (Idx < (int) NumElts) ? OtherSV->getOperand(0) | |||
16200 | : OtherSV->getOperand(1); | |||
16201 | } else { | |||
16202 | // This shuffle index references an element within N1. | |||
16203 | CurrentVec = N1; | |||
16204 | } | |||
16205 | ||||
16206 | // Simple case where 'CurrentVec' is UNDEF. | |||
16207 | if (CurrentVec.isUndef()) { | |||
16208 | Mask.push_back(-1); | |||
16209 | continue; | |||
16210 | } | |||
16211 | ||||
16212 | // Canonicalize the shuffle index. We don't know yet if CurrentVec | |||
16213 | // will be the first or second operand of the combined shuffle. | |||
16214 | Idx = Idx % NumElts; | |||
16215 | if (!SV0.getNode() || SV0 == CurrentVec) { | |||
16216 | // Ok. CurrentVec is the left hand side. | |||
16217 | // Update the mask accordingly. | |||
16218 | SV0 = CurrentVec; | |||
16219 | Mask.push_back(Idx); | |||
16220 | continue; | |||
16221 | } | |||
16222 | ||||
16223 | // Bail out if we cannot convert the shuffle pair into a single shuffle. | |||
16224 | if (SV1.getNode() && SV1 != CurrentVec) | |||
16225 | return SDValue(); | |||
16226 | ||||
16227 | // Ok. CurrentVec is the right hand side. | |||
16228 | // Update the mask accordingly. | |||
16229 | SV1 = CurrentVec; | |||
16230 | Mask.push_back(Idx + NumElts); | |||
16231 | } | |||
16232 | ||||
16233 | // Check if all indices in Mask are Undef. In case, propagate Undef. | |||
16234 | bool isUndefMask = true; | |||
16235 | for (unsigned i = 0; i != NumElts && isUndefMask; ++i) | |||
16236 | isUndefMask &= Mask[i] < 0; | |||
16237 | ||||
16238 | if (isUndefMask) | |||
16239 | return DAG.getUNDEF(VT); | |||
16240 | ||||
16241 | if (!SV0.getNode()) | |||
16242 | SV0 = DAG.getUNDEF(VT); | |||
16243 | if (!SV1.getNode()) | |||
16244 | SV1 = DAG.getUNDEF(VT); | |||
16245 | ||||
16246 | // Avoid introducing shuffles with illegal mask. | |||
16247 | if (!TLI.isShuffleMaskLegal(Mask, VT)) { | |||
16248 | ShuffleVectorSDNode::commuteMask(Mask); | |||
16249 | ||||
16250 | if (!TLI.isShuffleMaskLegal(Mask, VT)) | |||
16251 | return SDValue(); | |||
16252 | ||||
16253 | // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, A, M2) | |||
16254 | // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, A, M2) | |||
16255 | // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, B, M2) | |||
16256 | std::swap(SV0, SV1); | |||
16257 | } | |||
16258 | ||||
16259 | // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2) | |||
16260 | // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2) | |||
16261 | // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2) | |||
16262 | return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, Mask); | |||
16263 | } | |||
16264 | ||||
16265 | return SDValue(); | |||
16266 | } | |||
16267 | ||||
16268 | SDValue DAGCombiner::visitSCALAR_TO_VECTOR(SDNode *N) { | |||
16269 | SDValue InVal = N->getOperand(0); | |||
16270 | EVT VT = N->getValueType(0); | |||
16271 | ||||
16272 | // Replace a SCALAR_TO_VECTOR(EXTRACT_VECTOR_ELT(V,C0)) pattern | |||
16273 | // with a VECTOR_SHUFFLE and possible truncate. | |||
16274 | if (InVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT) { | |||
16275 | SDValue InVec = InVal->getOperand(0); | |||
16276 | SDValue EltNo = InVal->getOperand(1); | |||
16277 | auto InVecT = InVec.getValueType(); | |||
16278 | if (ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(EltNo)) { | |||
16279 | SmallVector<int, 8> NewMask(InVecT.getVectorNumElements(), -1); | |||
16280 | int Elt = C0->getZExtValue(); | |||
16281 | NewMask[0] = Elt; | |||
16282 | SDValue Val; | |||
16283 | // If we have an implict truncate do truncate here as long as it's legal. | |||
16284 | // if it's not legal, this should | |||
16285 | if (VT.getScalarType() != InVal.getValueType() && | |||
16286 | InVal.getValueType().isScalarInteger() && | |||
16287 | isTypeLegal(VT.getScalarType())) { | |||
16288 | Val = | |||
16289 | DAG.getNode(ISD::TRUNCATE, SDLoc(InVal), VT.getScalarType(), InVal); | |||
16290 | return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), VT, Val); | |||
16291 | } | |||
16292 | if (VT.getScalarType() == InVecT.getScalarType() && | |||
16293 | VT.getVectorNumElements() <= InVecT.getVectorNumElements() && | |||
16294 | TLI.isShuffleMaskLegal(NewMask, VT)) { | |||
16295 | Val = DAG.getVectorShuffle(InVecT, SDLoc(N), InVec, | |||
16296 | DAG.getUNDEF(InVecT), NewMask); | |||
16297 | // If the initial vector is the correct size this shuffle is a | |||
16298 | // valid result. | |||
16299 | if (VT == InVecT) | |||
16300 | return Val; | |||
16301 | // If not we must truncate the vector. | |||
16302 | if (VT.getVectorNumElements() != InVecT.getVectorNumElements()) { | |||
16303 | MVT IdxTy = TLI.getVectorIdxTy(DAG.getDataLayout()); | |||
16304 | SDValue ZeroIdx = DAG.getConstant(0, SDLoc(N), IdxTy); | |||
16305 | EVT SubVT = | |||
16306 | EVT::getVectorVT(*DAG.getContext(), InVecT.getVectorElementType(), | |||
16307 | VT.getVectorNumElements()); | |||
16308 | Val = DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N), SubVT, Val, | |||
16309 | ZeroIdx); | |||
16310 | return Val; | |||
16311 | } | |||
16312 | } | |||
16313 | } | |||
16314 | } | |||
16315 | ||||
16316 | return SDValue(); | |||
16317 | } | |||
16318 | ||||
16319 | SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) { | |||
16320 | EVT VT = N->getValueType(0); | |||
16321 | SDValue N0 = N->getOperand(0); | |||
16322 | SDValue N1 = N->getOperand(1); | |||
16323 | SDValue N2 = N->getOperand(2); | |||
16324 | ||||
16325 | // If inserting an UNDEF, just return the original vector. | |||
16326 | if (N1.isUndef()) | |||
16327 | return N0; | |||
16328 | ||||
16329 | // For nested INSERT_SUBVECTORs, attempt to combine inner node first to allow | |||
16330 | // us to pull BITCASTs from input to output. | |||
16331 | if (N0.hasOneUse() && N0->getOpcode() == ISD::INSERT_SUBVECTOR) | |||
16332 | if (SDValue NN0 = visitINSERT_SUBVECTOR(N0.getNode())) | |||
16333 | return DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), VT, NN0, N1, N2); | |||
16334 | ||||
16335 | // If this is an insert of an extracted vector into an undef vector, we can | |||
16336 | // just use the input to the extract. | |||
16337 | if (N0.isUndef() && N1.getOpcode() == ISD::EXTRACT_SUBVECTOR && | |||
16338 | N1.getOperand(1) == N2 && N1.getOperand(0).getValueType() == VT) | |||
16339 | return N1.getOperand(0); | |||
16340 | ||||
16341 | // If we are inserting a bitcast value into an undef, with the same | |||
16342 | // number of elements, just use the bitcast input of the extract. | |||
16343 | // i.e. INSERT_SUBVECTOR UNDEF (BITCAST N1) N2 -> | |||
16344 | // BITCAST (INSERT_SUBVECTOR UNDEF N1 N2) | |||
16345 | if (N0.isUndef() && N1.getOpcode() == ISD::BITCAST && | |||
16346 | N1.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR && | |||
16347 | N1.getOperand(0).getOperand(1) == N2 && | |||
16348 | N1.getOperand(0).getOperand(0).getValueType().getVectorNumElements() == | |||
16349 | VT.getVectorNumElements() && | |||
16350 | N1.getOperand(0).getOperand(0).getValueType().getSizeInBits() == | |||
16351 | VT.getSizeInBits()) { | |||
16352 | return DAG.getBitcast(VT, N1.getOperand(0).getOperand(0)); | |||
16353 | } | |||
16354 | ||||
16355 | // If both N1 and N2 are bitcast values on which insert_subvector | |||
16356 | // would makes sense, pull the bitcast through. | |||
16357 | // i.e. INSERT_SUBVECTOR (BITCAST N0) (BITCAST N1) N2 -> | |||
16358 | // BITCAST (INSERT_SUBVECTOR N0 N1 N2) | |||
16359 | if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST) { | |||
16360 | SDValue CN0 = N0.getOperand(0); | |||
16361 | SDValue CN1 = N1.getOperand(0); | |||
16362 | EVT CN0VT = CN0.getValueType(); | |||
16363 | EVT CN1VT = CN1.getValueType(); | |||
16364 | if (CN0VT.isVector() && CN1VT.isVector() && | |||
16365 | CN0VT.getVectorElementType() == CN1VT.getVectorElementType() && | |||
16366 | CN0VT.getVectorNumElements() == VT.getVectorNumElements()) { | |||
16367 | SDValue NewINSERT = DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), | |||
16368 | CN0.getValueType(), CN0, CN1, N2); | |||
16369 | return DAG.getBitcast(VT, NewINSERT); | |||
16370 | } | |||
16371 | } | |||
16372 | ||||
16373 | // Combine INSERT_SUBVECTORs where we are inserting to the same index. | |||
16374 | // INSERT_SUBVECTOR( INSERT_SUBVECTOR( Vec, SubOld, Idx ), SubNew, Idx ) | |||
16375 | // --> INSERT_SUBVECTOR( Vec, SubNew, Idx ) | |||
16376 | if (N0.getOpcode() == ISD::INSERT_SUBVECTOR && | |||
16377 | N0.getOperand(1).getValueType() == N1.getValueType() && | |||
16378 | N0.getOperand(2) == N2) | |||
16379 | return DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), VT, N0.getOperand(0), | |||
16380 | N1, N2); | |||
16381 | ||||
16382 | if (!isa<ConstantSDNode>(N2)) | |||
16383 | return SDValue(); | |||
16384 | ||||
16385 | unsigned InsIdx = cast<ConstantSDNode>(N2)->getZExtValue(); | |||
16386 | ||||
16387 | // Canonicalize insert_subvector dag nodes. | |||
16388 | // Example: | |||
16389 | // (insert_subvector (insert_subvector A, Idx0), Idx1) | |||
16390 | // -> (insert_subvector (insert_subvector A, Idx1), Idx0) | |||
16391 | if (N0.getOpcode() == ISD::INSERT_SUBVECTOR && N0.hasOneUse() && | |||
16392 | N1.getValueType() == N0.getOperand(1).getValueType() && | |||
16393 | isa<ConstantSDNode>(N0.getOperand(2))) { | |||
16394 | unsigned OtherIdx = N0.getConstantOperandVal(2); | |||
16395 | if (InsIdx < OtherIdx) { | |||
16396 | // Swap nodes. | |||
16397 | SDValue NewOp = DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), VT, | |||
16398 | N0.getOperand(0), N1, N2); | |||
16399 | AddToWorklist(NewOp.getNode()); | |||
16400 | return DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N0.getNode()), | |||
16401 | VT, NewOp, N0.getOperand(1), N0.getOperand(2)); | |||
16402 | } | |||
16403 | } | |||
16404 | ||||
16405 | // If the input vector is a concatenation, and the insert replaces | |||
16406 | // one of the pieces, we can optimize into a single concat_vectors. | |||
16407 | if (N0.getOpcode() == ISD::CONCAT_VECTORS && N0.hasOneUse() && | |||
16408 | N0.getOperand(0).getValueType() == N1.getValueType()) { | |||
16409 | unsigned Factor = N1.getValueType().getVectorNumElements(); | |||
16410 | ||||
16411 | SmallVector<SDValue, 8> Ops(N0->op_begin(), N0->op_end()); | |||
16412 | Ops[cast<ConstantSDNode>(N2)->getZExtValue() / Factor] = N1; | |||
16413 | ||||
16414 | return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops); | |||
16415 | } | |||
16416 | ||||
16417 | return SDValue(); | |||
16418 | } | |||
16419 | ||||
16420 | SDValue DAGCombiner::visitFP_TO_FP16(SDNode *N) { | |||
16421 | SDValue N0 = N->getOperand(0); | |||
16422 | ||||
16423 | // fold (fp_to_fp16 (fp16_to_fp op)) -> op | |||
16424 | if (N0->getOpcode() == ISD::FP16_TO_FP) | |||
16425 | return N0->getOperand(0); | |||
16426 | ||||
16427 | return SDValue(); | |||
16428 | } | |||
16429 | ||||
16430 | SDValue DAGCombiner::visitFP16_TO_FP(SDNode *N) { | |||
16431 | SDValue N0 = N->getOperand(0); | |||
16432 | ||||
16433 | // fold fp16_to_fp(op & 0xffff) -> fp16_to_fp(op) | |||
16434 | if (N0->getOpcode() == ISD::AND) { | |||
16435 | ConstantSDNode *AndConst = getAsNonOpaqueConstant(N0.getOperand(1)); | |||
16436 | if (AndConst && AndConst->getAPIntValue() == 0xffff) { | |||
16437 | return DAG.getNode(ISD::FP16_TO_FP, SDLoc(N), N->getValueType(0), | |||
16438 | N0.getOperand(0)); | |||
16439 | } | |||
16440 | } | |||
16441 | ||||
16442 | return SDValue(); | |||
16443 | } | |||
16444 | ||||
16445 | /// Returns a vector_shuffle if it able to transform an AND to a vector_shuffle | |||
16446 | /// with the destination vector and a zero vector. | |||
16447 | /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==> | |||
16448 | /// vector_shuffle V, Zero, <0, 4, 2, 4> | |||
16449 | SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) { | |||
16450 | assert(N->getOpcode() == ISD::AND && "Unexpected opcode!")(static_cast <bool> (N->getOpcode() == ISD::AND && "Unexpected opcode!") ? void (0) : __assert_fail ("N->getOpcode() == ISD::AND && \"Unexpected opcode!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 16450, __extension__ __PRETTY_FUNCTION__)); | |||
16451 | ||||
16452 | EVT VT = N->getValueType(0); | |||
16453 | SDValue LHS = N->getOperand(0); | |||
16454 | SDValue RHS = peekThroughBitcast(N->getOperand(1)); | |||
16455 | SDLoc DL(N); | |||
16456 | ||||
16457 | // Make sure we're not running after operation legalization where it | |||
16458 | // may have custom lowered the vector shuffles. | |||
16459 | if (LegalOperations) | |||
| ||||
16460 | return SDValue(); | |||
16461 | ||||
16462 | if (RHS.getOpcode() != ISD::BUILD_VECTOR) | |||
16463 | return SDValue(); | |||
16464 | ||||
16465 | EVT RVT = RHS.getValueType(); | |||
16466 | unsigned NumElts = RHS.getNumOperands(); | |||
16467 | ||||
16468 | // Attempt to create a valid clear mask, splitting the mask into | |||
16469 | // sub elements and checking to see if each is | |||
16470 | // all zeros or all ones - suitable for shuffle masking. | |||
16471 | auto BuildClearMask = [&](int Split) { | |||
16472 | int NumSubElts = NumElts * Split; | |||
16473 | int NumSubBits = RVT.getScalarSizeInBits() / Split; | |||
16474 | ||||
16475 | SmallVector<int, 8> Indices; | |||
16476 | for (int i = 0; i != NumSubElts; ++i) { | |||
16477 | int EltIdx = i / Split; | |||
16478 | int SubIdx = i % Split; | |||
16479 | SDValue Elt = RHS.getOperand(EltIdx); | |||
16480 | if (Elt.isUndef()) { | |||
16481 | Indices.push_back(-1); | |||
16482 | continue; | |||
16483 | } | |||
16484 | ||||
16485 | APInt Bits; | |||
16486 | if (isa<ConstantSDNode>(Elt)) | |||
16487 | Bits = cast<ConstantSDNode>(Elt)->getAPIntValue(); | |||
16488 | else if (isa<ConstantFPSDNode>(Elt)) | |||
16489 | Bits = cast<ConstantFPSDNode>(Elt)->getValueAPF().bitcastToAPInt(); | |||
16490 | else | |||
16491 | return SDValue(); | |||
16492 | ||||
16493 | // Extract the sub element from the constant bit mask. | |||
16494 | if (DAG.getDataLayout().isBigEndian()) { | |||
16495 | Bits.lshrInPlace((Split - SubIdx - 1) * NumSubBits); | |||
16496 | } else { | |||
16497 | Bits.lshrInPlace(SubIdx * NumSubBits); | |||
16498 | } | |||
16499 | ||||
16500 | if (Split > 1) | |||
16501 | Bits = Bits.trunc(NumSubBits); | |||
16502 | ||||
16503 | if (Bits.isAllOnesValue()) | |||
16504 | Indices.push_back(i); | |||
16505 | else if (Bits == 0) | |||
16506 | Indices.push_back(i + NumSubElts); | |||
16507 | else | |||
16508 | return SDValue(); | |||
16509 | } | |||
16510 | ||||
16511 | // Let's see if the target supports this vector_shuffle. | |||
16512 | EVT ClearSVT = EVT::getIntegerVT(*DAG.getContext(), NumSubBits); | |||
16513 | EVT ClearVT = EVT::getVectorVT(*DAG.getContext(), ClearSVT, NumSubElts); | |||
16514 | if (!TLI.isVectorClearMaskLegal(Indices, ClearVT)) | |||
16515 | return SDValue(); | |||
16516 | ||||
16517 | SDValue Zero = DAG.getConstant(0, DL, ClearVT); | |||
16518 | return DAG.getBitcast(VT, DAG.getVectorShuffle(ClearVT, DL, | |||
16519 | DAG.getBitcast(ClearVT, LHS), | |||
16520 | Zero, Indices)); | |||
16521 | }; | |||
16522 | ||||
16523 | // Determine maximum split level (byte level masking). | |||
16524 | int MaxSplit = 1; | |||
16525 | if (RVT.getScalarSizeInBits() % 8 == 0) | |||
16526 | MaxSplit = RVT.getScalarSizeInBits() / 8; | |||
16527 | ||||
16528 | for (int Split = 1; Split <= MaxSplit; ++Split) | |||
16529 | if (RVT.getScalarSizeInBits() % Split == 0) | |||
16530 | if (SDValue S = BuildClearMask(Split)) | |||
16531 | return S; | |||
16532 | ||||
16533 | return SDValue(); | |||
16534 | } | |||
16535 | ||||
16536 | /// Visit a binary vector operation, like ADD. | |||
16537 | SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) { | |||
16538 | assert(N->getValueType(0).isVector() &&(static_cast <bool> (N->getValueType(0).isVector() && "SimplifyVBinOp only works on vectors!") ? void (0) : __assert_fail ("N->getValueType(0).isVector() && \"SimplifyVBinOp only works on vectors!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 16539, __extension__ __PRETTY_FUNCTION__)) | |||
16539 | "SimplifyVBinOp only works on vectors!")(static_cast <bool> (N->getValueType(0).isVector() && "SimplifyVBinOp only works on vectors!") ? void (0) : __assert_fail ("N->getValueType(0).isVector() && \"SimplifyVBinOp only works on vectors!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 16539, __extension__ __PRETTY_FUNCTION__)); | |||
16540 | ||||
16541 | SDValue LHS = N->getOperand(0); | |||
16542 | SDValue RHS = N->getOperand(1); | |||
16543 | SDValue Ops[] = {LHS, RHS}; | |||
16544 | ||||
16545 | // See if we can constant fold the vector operation. | |||
16546 | if (SDValue Fold = DAG.FoldConstantVectorArithmetic( | |||
16547 | N->getOpcode(), SDLoc(LHS), LHS.getValueType(), Ops, N->getFlags())) | |||
16548 | return Fold; | |||
16549 | ||||
16550 | // Type legalization might introduce new shuffles in the DAG. | |||
16551 | // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask))) | |||
16552 | // -> (shuffle (VBinOp (A, B)), Undef, Mask). | |||
16553 | if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) && | |||
16554 | isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() && | |||
16555 | LHS.getOperand(1).isUndef() && | |||
16556 | RHS.getOperand(1).isUndef()) { | |||
16557 | ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS); | |||
16558 | ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS); | |||
16559 | ||||
16560 | if (SVN0->getMask().equals(SVN1->getMask())) { | |||
16561 | EVT VT = N->getValueType(0); | |||
16562 | SDValue UndefVector = LHS.getOperand(1); | |||
16563 | SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT, | |||
16564 | LHS.getOperand(0), RHS.getOperand(0), | |||
16565 | N->getFlags()); | |||
16566 | AddUsersToWorklist(N); | |||
16567 | return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector, | |||
16568 | SVN0->getMask()); | |||
16569 | } | |||
16570 | } | |||
16571 | ||||
16572 | return SDValue(); | |||
16573 | } | |||
16574 | ||||
16575 | SDValue DAGCombiner::SimplifySelect(const SDLoc &DL, SDValue N0, SDValue N1, | |||
16576 | SDValue N2) { | |||
16577 | assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!")(static_cast <bool> (N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!") ? void (0) : __assert_fail ("N0.getOpcode() ==ISD::SETCC && \"First argument must be a SetCC node!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 16577, __extension__ __PRETTY_FUNCTION__)); | |||
16578 | ||||
16579 | SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2, | |||
16580 | cast<CondCodeSDNode>(N0.getOperand(2))->get()); | |||
16581 | ||||
16582 | // If we got a simplified select_cc node back from SimplifySelectCC, then | |||
16583 | // break it down into a new SETCC node, and a new SELECT node, and then return | |||
16584 | // the SELECT node, since we were called with a SELECT node. | |||
16585 | if (SCC.getNode()) { | |||
16586 | // Check to see if we got a select_cc back (to turn into setcc/select). | |||
16587 | // Otherwise, just return whatever node we got back, like fabs. | |||
16588 | if (SCC.getOpcode() == ISD::SELECT_CC) { | |||
16589 | SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0), | |||
16590 | N0.getValueType(), | |||
16591 | SCC.getOperand(0), SCC.getOperand(1), | |||
16592 | SCC.getOperand(4)); | |||
16593 | AddToWorklist(SETCC.getNode()); | |||
16594 | return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC, | |||
16595 | SCC.getOperand(2), SCC.getOperand(3)); | |||
16596 | } | |||
16597 | ||||
16598 | return SCC; | |||
16599 | } | |||
16600 | return SDValue(); | |||
16601 | } | |||
16602 | ||||
16603 | /// Given a SELECT or a SELECT_CC node, where LHS and RHS are the two values | |||
16604 | /// being selected between, see if we can simplify the select. Callers of this | |||
16605 | /// should assume that TheSelect is deleted if this returns true. As such, they | |||
16606 | /// should return the appropriate thing (e.g. the node) back to the top-level of | |||
16607 | /// the DAG combiner loop to avoid it being looked at. | |||
16608 | bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS, | |||
16609 | SDValue RHS) { | |||
16610 | // fold (select (setcc x, [+-]0.0, *lt), NaN, (fsqrt x)) | |||
16611 | // The select + setcc is redundant, because fsqrt returns NaN for X < 0. | |||
16612 | if (const ConstantFPSDNode *NaN = isConstOrConstSplatFP(LHS)) { | |||
16613 | if (NaN->isNaN() && RHS.getOpcode() == ISD::FSQRT) { | |||
16614 | // We have: (select (setcc ?, ?, ?), NaN, (fsqrt ?)) | |||
16615 | SDValue Sqrt = RHS; | |||
16616 | ISD::CondCode CC; | |||
16617 | SDValue CmpLHS; | |||
16618 | const ConstantFPSDNode *Zero = nullptr; | |||
16619 | ||||
16620 | if (TheSelect->getOpcode() == ISD::SELECT_CC) { | |||
16621 | CC = dyn_cast<CondCodeSDNode>(TheSelect->getOperand(4))->get(); | |||
16622 | CmpLHS = TheSelect->getOperand(0); | |||
16623 | Zero = isConstOrConstSplatFP(TheSelect->getOperand(1)); | |||
16624 | } else { | |||
16625 | // SELECT or VSELECT | |||
16626 | SDValue Cmp = TheSelect->getOperand(0); | |||
16627 | if (Cmp.getOpcode() == ISD::SETCC) { | |||
16628 | CC = dyn_cast<CondCodeSDNode>(Cmp.getOperand(2))->get(); | |||
16629 | CmpLHS = Cmp.getOperand(0); | |||
16630 | Zero = isConstOrConstSplatFP(Cmp.getOperand(1)); | |||
16631 | } | |||
16632 | } | |||
16633 | if (Zero && Zero->isZero() && | |||
16634 | Sqrt.getOperand(0) == CmpLHS && (CC == ISD::SETOLT || | |||
16635 | CC == ISD::SETULT || CC == ISD::SETLT)) { | |||
16636 | // We have: (select (setcc x, [+-]0.0, *lt), NaN, (fsqrt x)) | |||
16637 | CombineTo(TheSelect, Sqrt); | |||
16638 | return true; | |||
16639 | } | |||
16640 | } | |||
16641 | } | |||
16642 | // Cannot simplify select with vector condition | |||
16643 | if (TheSelect->getOperand(0).getValueType().isVector()) return false; | |||
16644 | ||||
16645 | // If this is a select from two identical things, try to pull the operation | |||
16646 | // through the select. | |||
16647 | if (LHS.getOpcode() != RHS.getOpcode() || | |||
16648 | !LHS.hasOneUse() || !RHS.hasOneUse()) | |||
16649 | return false; | |||
16650 | ||||
16651 | // If this is a load and the token chain is identical, replace the select | |||
16652 | // of two loads with a load through a select of the address to load from. | |||
16653 | // This triggers in things like "select bool X, 10.0, 123.0" after the FP | |||
16654 | // constants have been dropped into the constant pool. | |||
16655 | if (LHS.getOpcode() == ISD::LOAD) { | |||
16656 | LoadSDNode *LLD = cast<LoadSDNode>(LHS); | |||
16657 | LoadSDNode *RLD = cast<LoadSDNode>(RHS); | |||
16658 | ||||
16659 | // Token chains must be identical. | |||
16660 | if (LHS.getOperand(0) != RHS.getOperand(0) || | |||
16661 | // Do not let this transformation reduce the number of volatile loads. | |||
16662 | LLD->isVolatile() || RLD->isVolatile() || | |||
16663 | // FIXME: If either is a pre/post inc/dec load, | |||
16664 | // we'd need to split out the address adjustment. | |||
16665 | LLD->isIndexed() || RLD->isIndexed() || | |||
16666 | // If this is an EXTLOAD, the VT's must match. | |||
16667 | LLD->getMemoryVT() != RLD->getMemoryVT() || | |||
16668 | // If this is an EXTLOAD, the kind of extension must match. | |||
16669 | (LLD->getExtensionType() != RLD->getExtensionType() && | |||
16670 | // The only exception is if one of the extensions is anyext. | |||
16671 | LLD->getExtensionType() != ISD::EXTLOAD && | |||
16672 | RLD->getExtensionType() != ISD::EXTLOAD) || | |||
16673 | // FIXME: this discards src value information. This is | |||
16674 | // over-conservative. It would be beneficial to be able to remember | |||
16675 | // both potential memory locations. Since we are discarding | |||
16676 | // src value info, don't do the transformation if the memory | |||
16677 | // locations are not in the default address space. | |||
16678 | LLD->getPointerInfo().getAddrSpace() != 0 || | |||
16679 | RLD->getPointerInfo().getAddrSpace() != 0 || | |||
16680 | !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(), | |||
16681 | LLD->getBasePtr().getValueType())) | |||
16682 | return false; | |||
16683 | ||||
16684 | // Check that the select condition doesn't reach either load. If so, | |||
16685 | // folding this will induce a cycle into the DAG. If not, this is safe to | |||
16686 | // xform, so create a select of the addresses. | |||
16687 | SDValue Addr; | |||
16688 | if (TheSelect->getOpcode() == ISD::SELECT) { | |||
16689 | SDNode *CondNode = TheSelect->getOperand(0).getNode(); | |||
16690 | if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) || | |||
16691 | (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode))) | |||
16692 | return false; | |||
16693 | // The loads must not depend on one another. | |||
16694 | if (LLD->isPredecessorOf(RLD) || | |||
16695 | RLD->isPredecessorOf(LLD)) | |||
16696 | return false; | |||
16697 | Addr = DAG.getSelect(SDLoc(TheSelect), | |||
16698 | LLD->getBasePtr().getValueType(), | |||
16699 | TheSelect->getOperand(0), LLD->getBasePtr(), | |||
16700 | RLD->getBasePtr()); | |||
16701 | } else { // Otherwise SELECT_CC | |||
16702 | SDNode *CondLHS = TheSelect->getOperand(0).getNode(); | |||
16703 | SDNode *CondRHS = TheSelect->getOperand(1).getNode(); | |||
16704 | ||||
16705 | if ((LLD->hasAnyUseOfValue(1) && | |||
16706 | (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) || | |||
16707 | (RLD->hasAnyUseOfValue(1) && | |||
16708 | (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS)))) | |||
16709 | return false; | |||
16710 | ||||
16711 | Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect), | |||
16712 | LLD->getBasePtr().getValueType(), | |||
16713 | TheSelect->getOperand(0), | |||
16714 | TheSelect->getOperand(1), | |||
16715 | LLD->getBasePtr(), RLD->getBasePtr(), | |||
16716 | TheSelect->getOperand(4)); | |||
16717 | } | |||
16718 | ||||
16719 | SDValue Load; | |||
16720 | // It is safe to replace the two loads if they have different alignments, | |||
16721 | // but the new load must be the minimum (most restrictive) alignment of the | |||
16722 | // inputs. | |||
16723 | unsigned Alignment = std::min(LLD->getAlignment(), RLD->getAlignment()); | |||
16724 | MachineMemOperand::Flags MMOFlags = LLD->getMemOperand()->getFlags(); | |||
16725 | if (!RLD->isInvariant()) | |||
16726 | MMOFlags &= ~MachineMemOperand::MOInvariant; | |||
16727 | if (!RLD->isDereferenceable()) | |||
16728 | MMOFlags &= ~MachineMemOperand::MODereferenceable; | |||
16729 | if (LLD->getExtensionType() == ISD::NON_EXTLOAD) { | |||
16730 | // FIXME: Discards pointer and AA info. | |||
16731 | Load = DAG.getLoad(TheSelect->getValueType(0), SDLoc(TheSelect), | |||
16732 | LLD->getChain(), Addr, MachinePointerInfo(), Alignment, | |||
16733 | MMOFlags); | |||
16734 | } else { | |||
16735 | // FIXME: Discards pointer and AA info. | |||
16736 | Load = DAG.getExtLoad( | |||
16737 | LLD->getExtensionType() == ISD::EXTLOAD ? RLD->getExtensionType() | |||
16738 | : LLD->getExtensionType(), | |||
16739 | SDLoc(TheSelect), TheSelect->getValueType(0), LLD->getChain(), Addr, | |||
16740 | MachinePointerInfo(), LLD->getMemoryVT(), Alignment, MMOFlags); | |||
16741 | } | |||
16742 | ||||
16743 | // Users of the select now use the result of the load. | |||
16744 | CombineTo(TheSelect, Load); | |||
16745 | ||||
16746 | // Users of the old loads now use the new load's chain. We know the | |||
16747 | // old-load value is dead now. | |||
16748 | CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1)); | |||
16749 | CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1)); | |||
16750 | return true; | |||
16751 | } | |||
16752 | ||||
16753 | return false; | |||
16754 | } | |||
16755 | ||||
16756 | /// Try to fold an expression of the form (N0 cond N1) ? N2 : N3 to a shift and | |||
16757 | /// bitwise 'and'. | |||
16758 | SDValue DAGCombiner::foldSelectCCToShiftAnd(const SDLoc &DL, SDValue N0, | |||
16759 | SDValue N1, SDValue N2, SDValue N3, | |||
16760 | ISD::CondCode CC) { | |||
16761 | // If this is a select where the false operand is zero and the compare is a | |||
16762 | // check of the sign bit, see if we can perform the "gzip trick": | |||
16763 | // select_cc setlt X, 0, A, 0 -> and (sra X, size(X)-1), A | |||
16764 | // select_cc setgt X, 0, A, 0 -> and (not (sra X, size(X)-1)), A | |||
16765 | EVT XType = N0.getValueType(); | |||
16766 | EVT AType = N2.getValueType(); | |||
16767 | if (!isNullConstant(N3) || !XType.bitsGE(AType)) | |||
16768 | return SDValue(); | |||
16769 | ||||
16770 | // If the comparison is testing for a positive value, we have to invert | |||
16771 | // the sign bit mask, so only do that transform if the target has a bitwise | |||
16772 | // 'and not' instruction (the invert is free). | |||
16773 | if (CC == ISD::SETGT && TLI.hasAndNot(N2)) { | |||
16774 | // (X > -1) ? A : 0 | |||
16775 | // (X > 0) ? X : 0 <-- This is canonical signed max. | |||
16776 | if (!(isAllOnesConstant(N1) || (isNullConstant(N1) && N0 == N2))) | |||
16777 | return SDValue(); | |||
16778 | } else if (CC == ISD::SETLT) { | |||
16779 | // (X < 0) ? A : 0 | |||
16780 | // (X < 1) ? X : 0 <-- This is un-canonicalized signed min. | |||
16781 | if (!(isNullConstant(N1) || (isOneConstant(N1) && N0 == N2))) | |||
16782 | return SDValue(); | |||
16783 | } else { | |||
16784 | return SDValue(); | |||
16785 | } | |||
16786 | ||||
16787 | // and (sra X, size(X)-1), A -> "and (srl X, C2), A" iff A is a single-bit | |||
16788 | // constant. | |||
16789 | EVT ShiftAmtTy = getShiftAmountTy(N0.getValueType()); | |||
16790 | auto *N2C = dyn_cast<ConstantSDNode>(N2.getNode()); | |||
16791 | if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue() - 1)) == 0)) { | |||
16792 | unsigned ShCt = XType.getSizeInBits() - N2C->getAPIntValue().logBase2() - 1; | |||
16793 | SDValue ShiftAmt = DAG.getConstant(ShCt, DL, ShiftAmtTy); | |||
16794 | SDValue Shift = DAG.getNode(ISD::SRL, DL, XType, N0, ShiftAmt); | |||
16795 | AddToWorklist(Shift.getNode()); | |||
16796 | ||||
16797 | if (XType.bitsGT(AType)) { | |||
16798 | Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift); | |||
16799 | AddToWorklist(Shift.getNode()); | |||
16800 | } | |||
16801 | ||||
16802 | if (CC == ISD::SETGT) | |||
16803 | Shift = DAG.getNOT(DL, Shift, AType); | |||
16804 | ||||
16805 | return DAG.getNode(ISD::AND, DL, AType, Shift, N2); | |||
16806 | } | |||
16807 | ||||
16808 | SDValue ShiftAmt = DAG.getConstant(XType.getSizeInBits() - 1, DL, ShiftAmtTy); | |||
16809 | SDValue Shift = DAG.getNode(ISD::SRA, DL, XType, N0, ShiftAmt); | |||
16810 | AddToWorklist(Shift.getNode()); | |||
16811 | ||||
16812 | if (XType.bitsGT(AType)) { | |||
16813 | Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift); | |||
16814 | AddToWorklist(Shift.getNode()); | |||
16815 | } | |||
16816 | ||||
16817 | if (CC == ISD::SETGT) | |||
16818 | Shift = DAG.getNOT(DL, Shift, AType); | |||
16819 | ||||
16820 | return DAG.getNode(ISD::AND, DL, AType, Shift, N2); | |||
16821 | } | |||
16822 | ||||
16823 | /// Simplify an expression of the form (N0 cond N1) ? N2 : N3 | |||
16824 | /// where 'cond' is the comparison specified by CC. | |||
16825 | SDValue DAGCombiner::SimplifySelectCC(const SDLoc &DL, SDValue N0, SDValue N1, | |||
16826 | SDValue N2, SDValue N3, ISD::CondCode CC, | |||
16827 | bool NotExtCompare) { | |||
16828 | // (x ? y : y) -> y. | |||
16829 | if (N2 == N3) return N2; | |||
16830 | ||||
16831 | EVT VT = N2.getValueType(); | |||
16832 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode()); | |||
16833 | ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode()); | |||
16834 | ||||
16835 | // Determine if the condition we're dealing with is constant | |||
16836 | SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()), | |||
16837 | N0, N1, CC, DL, false); | |||
16838 | if (SCC.getNode()) AddToWorklist(SCC.getNode()); | |||
16839 | ||||
16840 | if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode())) { | |||
16841 | // fold select_cc true, x, y -> x | |||
16842 | // fold select_cc false, x, y -> y | |||
16843 | return !SCCC->isNullValue() ? N2 : N3; | |||
16844 | } | |||
16845 | ||||
16846 | // Check to see if we can simplify the select into an fabs node | |||
16847 | if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) { | |||
16848 | // Allow either -0.0 or 0.0 | |||
16849 | if (CFP->isZero()) { | |||
16850 | // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs | |||
16851 | if ((CC == ISD::SETGE || CC == ISD::SETGT) && | |||
16852 | N0 == N2 && N3.getOpcode() == ISD::FNEG && | |||
16853 | N2 == N3.getOperand(0)) | |||
16854 | return DAG.getNode(ISD::FABS, DL, VT, N0); | |||
16855 | ||||
16856 | // select (setl[te] X, +/-0.0), fneg(X), X -> fabs | |||
16857 | if ((CC == ISD::SETLT || CC == ISD::SETLE) && | |||
16858 | N0 == N3 && N2.getOpcode() == ISD::FNEG && | |||
16859 | N2.getOperand(0) == N3) | |||
16860 | return DAG.getNode(ISD::FABS, DL, VT, N3); | |||
16861 | } | |||
16862 | } | |||
16863 | ||||
16864 | // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)" | |||
16865 | // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0 | |||
16866 | // in it. This is a win when the constant is not otherwise available because | |||
16867 | // it replaces two constant pool loads with one. We only do this if the FP | |||
16868 | // type is known to be legal, because if it isn't, then we are before legalize | |||
16869 | // types an we want the other legalization to happen first (e.g. to avoid | |||
16870 | // messing with soft float) and if the ConstantFP is not legal, because if | |||
16871 | // it is legal, we may not need to store the FP constant in a constant pool. | |||
16872 | if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2)) | |||
16873 | if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) { | |||
16874 | if (TLI.isTypeLegal(N2.getValueType()) && | |||
16875 | (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) != | |||
16876 | TargetLowering::Legal && | |||
16877 | !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) && | |||
16878 | !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) && | |||
16879 | // If both constants have multiple uses, then we won't need to do an | |||
16880 | // extra load, they are likely around in registers for other users. | |||
16881 | (TV->hasOneUse() || FV->hasOneUse())) { | |||
16882 | Constant *Elts[] = { | |||
16883 | const_cast<ConstantFP*>(FV->getConstantFPValue()), | |||
16884 | const_cast<ConstantFP*>(TV->getConstantFPValue()) | |||
16885 | }; | |||
16886 | Type *FPTy = Elts[0]->getType(); | |||
16887 | const DataLayout &TD = DAG.getDataLayout(); | |||
16888 | ||||
16889 | // Create a ConstantArray of the two constants. | |||
16890 | Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts); | |||
16891 | SDValue CPIdx = | |||
16892 | DAG.getConstantPool(CA, TLI.getPointerTy(DAG.getDataLayout()), | |||
16893 | TD.getPrefTypeAlignment(FPTy)); | |||
16894 | unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment(); | |||
16895 | ||||
16896 | // Get the offsets to the 0 and 1 element of the array so that we can | |||
16897 | // select between them. | |||
16898 | SDValue Zero = DAG.getIntPtrConstant(0, DL); | |||
16899 | unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType()); | |||
16900 | SDValue One = DAG.getIntPtrConstant(EltSize, SDLoc(FV)); | |||
16901 | ||||
16902 | SDValue Cond = DAG.getSetCC(DL, | |||
16903 | getSetCCResultType(N0.getValueType()), | |||
16904 | N0, N1, CC); | |||
16905 | AddToWorklist(Cond.getNode()); | |||
16906 | SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(), | |||
16907 | Cond, One, Zero); | |||
16908 | AddToWorklist(CstOffset.getNode()); | |||
16909 | CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx, | |||
16910 | CstOffset); | |||
16911 | AddToWorklist(CPIdx.getNode()); | |||
16912 | return DAG.getLoad( | |||
16913 | TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx, | |||
16914 | MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), | |||
16915 | Alignment); | |||
16916 | } | |||
16917 | } | |||
16918 | ||||
16919 | if (SDValue V = foldSelectCCToShiftAnd(DL, N0, N1, N2, N3, CC)) | |||
16920 | return V; | |||
16921 | ||||
16922 | // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A) | |||
16923 | // where y is has a single bit set. | |||
16924 | // A plaintext description would be, we can turn the SELECT_CC into an AND | |||
16925 | // when the condition can be materialized as an all-ones register. Any | |||
16926 | // single bit-test can be materialized as an all-ones register with | |||
16927 | // shift-left and shift-right-arith. | |||
16928 | if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND && | |||
16929 | N0->getValueType(0) == VT && isNullConstant(N1) && isNullConstant(N2)) { | |||
16930 | SDValue AndLHS = N0->getOperand(0); | |||
16931 | ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1)); | |||
16932 | if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) { | |||
16933 | // Shift the tested bit over the sign bit. | |||
16934 | const APInt &AndMask = ConstAndRHS->getAPIntValue(); | |||
16935 | SDValue ShlAmt = | |||
16936 | DAG.getConstant(AndMask.countLeadingZeros(), SDLoc(AndLHS), | |||
16937 | getShiftAmountTy(AndLHS.getValueType())); | |||
16938 | SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt); | |||
16939 | ||||
16940 | // Now arithmetic right shift it all the way over, so the result is either | |||
16941 | // all-ones, or zero. | |||
16942 | SDValue ShrAmt = | |||
16943 | DAG.getConstant(AndMask.getBitWidth() - 1, SDLoc(Shl), | |||
16944 | getShiftAmountTy(Shl.getValueType())); | |||
16945 | SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt); | |||
16946 | ||||
16947 | return DAG.getNode(ISD::AND, DL, VT, Shr, N3); | |||
16948 | } | |||
16949 | } | |||
16950 | ||||
16951 | // fold select C, 16, 0 -> shl C, 4 | |||
16952 | if (N2C && isNullConstant(N3) && N2C->getAPIntValue().isPowerOf2() && | |||
16953 | TLI.getBooleanContents(N0.getValueType()) == | |||
16954 | TargetLowering::ZeroOrOneBooleanContent) { | |||
16955 | ||||
16956 | // If the caller doesn't want us to simplify this into a zext of a compare, | |||
16957 | // don't do it. | |||
16958 | if (NotExtCompare && N2C->isOne()) | |||
16959 | return SDValue(); | |||
16960 | ||||
16961 | // Get a SetCC of the condition | |||
16962 | // NOTE: Don't create a SETCC if it's not legal on this target. | |||
16963 | if (!LegalOperations || | |||
16964 | TLI.isOperationLegal(ISD::SETCC, N0.getValueType())) { | |||
16965 | SDValue Temp, SCC; | |||
16966 | // cast from setcc result type to select result type | |||
16967 | if (LegalTypes) { | |||
16968 | SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()), | |||
16969 | N0, N1, CC); | |||
16970 | if (N2.getValueType().bitsLT(SCC.getValueType())) | |||
16971 | Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2), | |||
16972 | N2.getValueType()); | |||
16973 | else | |||
16974 | Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2), | |||
16975 | N2.getValueType(), SCC); | |||
16976 | } else { | |||
16977 | SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC); | |||
16978 | Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2), | |||
16979 | N2.getValueType(), SCC); | |||
16980 | } | |||
16981 | ||||
16982 | AddToWorklist(SCC.getNode()); | |||
16983 | AddToWorklist(Temp.getNode()); | |||
16984 | ||||
16985 | if (N2C->isOne()) | |||
16986 | return Temp; | |||
16987 | ||||
16988 | // shl setcc result by log2 n2c | |||
16989 | return DAG.getNode( | |||
16990 | ISD::SHL, DL, N2.getValueType(), Temp, | |||
16991 | DAG.getConstant(N2C->getAPIntValue().logBase2(), SDLoc(Temp), | |||
16992 | getShiftAmountTy(Temp.getValueType()))); | |||
16993 | } | |||
16994 | } | |||
16995 | ||||
16996 | // Check to see if this is an integer abs. | |||
16997 | // select_cc setg[te] X, 0, X, -X -> | |||
16998 | // select_cc setgt X, -1, X, -X -> | |||
16999 | // select_cc setl[te] X, 0, -X, X -> | |||
17000 | // select_cc setlt X, 1, -X, X -> | |||
17001 | // Y = sra (X, size(X)-1); xor (add (X, Y), Y) | |||
17002 | if (N1C) { | |||
17003 | ConstantSDNode *SubC = nullptr; | |||
17004 | if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) || | |||
17005 | (N1C->isAllOnesValue() && CC == ISD::SETGT)) && | |||
17006 | N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1)) | |||
17007 | SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0)); | |||
17008 | else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) || | |||
17009 | (N1C->isOne() && CC == ISD::SETLT)) && | |||
17010 | N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1)) | |||
17011 | SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0)); | |||
17012 | ||||
17013 | EVT XType = N0.getValueType(); | |||
17014 | if (SubC && SubC->isNullValue() && XType.isInteger()) { | |||
17015 | SDLoc DL(N0); | |||
17016 | SDValue Shift = DAG.getNode(ISD::SRA, DL, XType, | |||
17017 | N0, | |||
17018 | DAG.getConstant(XType.getSizeInBits() - 1, DL, | |||
17019 | getShiftAmountTy(N0.getValueType()))); | |||
17020 | SDValue Add = DAG.getNode(ISD::ADD, DL, | |||
17021 | XType, N0, Shift); | |||
17022 | AddToWorklist(Shift.getNode()); | |||
17023 | AddToWorklist(Add.getNode()); | |||
17024 | return DAG.getNode(ISD::XOR, DL, XType, Add, Shift); | |||
17025 | } | |||
17026 | } | |||
17027 | ||||
17028 | // select_cc seteq X, 0, sizeof(X), ctlz(X) -> ctlz(X) | |||
17029 | // select_cc seteq X, 0, sizeof(X), ctlz_zero_undef(X) -> ctlz(X) | |||
17030 | // select_cc seteq X, 0, sizeof(X), cttz(X) -> cttz(X) | |||
17031 | // select_cc seteq X, 0, sizeof(X), cttz_zero_undef(X) -> cttz(X) | |||
17032 | // select_cc setne X, 0, ctlz(X), sizeof(X) -> ctlz(X) | |||
17033 | // select_cc setne X, 0, ctlz_zero_undef(X), sizeof(X) -> ctlz(X) | |||
17034 | // select_cc setne X, 0, cttz(X), sizeof(X) -> cttz(X) | |||
17035 | // select_cc setne X, 0, cttz_zero_undef(X), sizeof(X) -> cttz(X) | |||
17036 | if (N1C && N1C->isNullValue() && (CC == ISD::SETEQ || CC == ISD::SETNE)) { | |||
17037 | SDValue ValueOnZero = N2; | |||
17038 | SDValue Count = N3; | |||
17039 | // If the condition is NE instead of E, swap the operands. | |||
17040 | if (CC == ISD::SETNE) | |||
17041 | std::swap(ValueOnZero, Count); | |||
17042 | // Check if the value on zero is a constant equal to the bits in the type. | |||
17043 | if (auto *ValueOnZeroC = dyn_cast<ConstantSDNode>(ValueOnZero)) { | |||
17044 | if (ValueOnZeroC->getAPIntValue() == VT.getSizeInBits()) { | |||
17045 | // If the other operand is cttz/cttz_zero_undef of N0, and cttz is | |||
17046 | // legal, combine to just cttz. | |||
17047 | if ((Count.getOpcode() == ISD::CTTZ || | |||
17048 | Count.getOpcode() == ISD::CTTZ_ZERO_UNDEF) && | |||
17049 | N0 == Count.getOperand(0) && | |||
17050 | (!LegalOperations || TLI.isOperationLegal(ISD::CTTZ, VT))) | |||
17051 | return DAG.getNode(ISD::CTTZ, DL, VT, N0); | |||
17052 | // If the other operand is ctlz/ctlz_zero_undef of N0, and ctlz is | |||
17053 | // legal, combine to just ctlz. | |||
17054 | if ((Count.getOpcode() == ISD::CTLZ || | |||
17055 | Count.getOpcode() == ISD::CTLZ_ZERO_UNDEF) && | |||
17056 | N0 == Count.getOperand(0) && | |||
17057 | (!LegalOperations || TLI.isOperationLegal(ISD::CTLZ, VT))) | |||
17058 | return DAG.getNode(ISD::CTLZ, DL, VT, N0); | |||
17059 | } | |||
17060 | } | |||
17061 | } | |||
17062 | ||||
17063 | return SDValue(); | |||
17064 | } | |||
17065 | ||||
17066 | /// This is a stub for TargetLowering::SimplifySetCC. | |||
17067 | SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, | |||
17068 | ISD::CondCode Cond, const SDLoc &DL, | |||
17069 | bool foldBooleans) { | |||
17070 | TargetLowering::DAGCombinerInfo | |||
17071 | DagCombineInfo(DAG, Level, false, this); | |||
17072 | return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL); | |||
17073 | } | |||
17074 | ||||
17075 | /// Given an ISD::SDIV node expressing a divide by constant, return | |||
17076 | /// a DAG expression to select that will generate the same value by multiplying | |||
17077 | /// by a magic number. | |||
17078 | /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide". | |||
17079 | SDValue DAGCombiner::BuildSDIV(SDNode *N) { | |||
17080 | // when optimising for minimum size, we don't want to expand a div to a mul | |||
17081 | // and a shift. | |||
17082 | if (DAG.getMachineFunction().getFunction().optForMinSize()) | |||
17083 | return SDValue(); | |||
17084 | ||||
17085 | ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1)); | |||
17086 | if (!C) | |||
17087 | return SDValue(); | |||
17088 | ||||
17089 | // Avoid division by zero. | |||
17090 | if (C->isNullValue()) | |||
17091 | return SDValue(); | |||
17092 | ||||
17093 | std::vector<SDNode *> Built; | |||
17094 | SDValue S = | |||
17095 | TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built); | |||
17096 | ||||
17097 | for (SDNode *N : Built) | |||
17098 | AddToWorklist(N); | |||
17099 | return S; | |||
17100 | } | |||
17101 | ||||
17102 | /// Given an ISD::SDIV node expressing a divide by constant power of 2, return a | |||
17103 | /// DAG expression that will generate the same value by right shifting. | |||
17104 | SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) { | |||
17105 | ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1)); | |||
17106 | if (!C) | |||
17107 | return SDValue(); | |||
17108 | ||||
17109 | // Avoid division by zero. | |||
17110 | if (C->isNullValue()) | |||
17111 | return SDValue(); | |||
17112 | ||||
17113 | std::vector<SDNode *> Built; | |||
17114 | SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, &Built); | |||
17115 | ||||
17116 | for (SDNode *N : Built) | |||
17117 | AddToWorklist(N); | |||
17118 | return S; | |||
17119 | } | |||
17120 | ||||
17121 | /// Given an ISD::UDIV node expressing a divide by constant, return a DAG | |||
17122 | /// expression that will generate the same value by multiplying by a magic | |||
17123 | /// number. | |||
17124 | /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide". | |||
17125 | SDValue DAGCombiner::BuildUDIV(SDNode *N) { | |||
17126 | // when optimising for minimum size, we don't want to expand a div to a mul | |||
17127 | // and a shift. | |||
17128 | if (DAG.getMachineFunction().getFunction().optForMinSize()) | |||
17129 | return SDValue(); | |||
17130 | ||||
17131 | ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1)); | |||
17132 | if (!C) | |||
17133 | return SDValue(); | |||
17134 | ||||
17135 | // Avoid division by zero. | |||
17136 | if (C->isNullValue()) | |||
17137 | return SDValue(); | |||
17138 | ||||
17139 | std::vector<SDNode *> Built; | |||
17140 | SDValue S = | |||
17141 | TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built); | |||
17142 | ||||
17143 | for (SDNode *N : Built) | |||
17144 | AddToWorklist(N); | |||
17145 | return S; | |||
17146 | } | |||
17147 | ||||
17148 | /// Determines the LogBase2 value for a non-null input value using the | |||
17149 | /// transform: LogBase2(V) = (EltBits - 1) - ctlz(V). | |||
17150 | SDValue DAGCombiner::BuildLogBase2(SDValue V, const SDLoc &DL) { | |||
17151 | EVT VT = V.getValueType(); | |||
17152 | unsigned EltBits = VT.getScalarSizeInBits(); | |||
17153 | SDValue Ctlz = DAG.getNode(ISD::CTLZ, DL, VT, V); | |||
17154 | SDValue Base = DAG.getConstant(EltBits - 1, DL, VT); | |||
17155 | SDValue LogBase2 = DAG.getNode(ISD::SUB, DL, VT, Base, Ctlz); | |||
17156 | return LogBase2; | |||
17157 | } | |||
17158 | ||||
17159 | /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i) | |||
17160 | /// For the reciprocal, we need to find the zero of the function: | |||
17161 | /// F(X) = A X - 1 [which has a zero at X = 1/A] | |||
17162 | /// => | |||
17163 | /// X_{i+1} = X_i (2 - A X_i) = X_i + X_i (1 - A X_i) [this second form | |||
17164 | /// does not require additional intermediate precision] | |||
17165 | SDValue DAGCombiner::BuildReciprocalEstimate(SDValue Op, SDNodeFlags Flags) { | |||
17166 | if (Level >= AfterLegalizeDAG) | |||
17167 | return SDValue(); | |||
17168 | ||||
17169 | // TODO: Handle half and/or extended types? | |||
17170 | EVT VT = Op.getValueType(); | |||
17171 | if (VT.getScalarType() != MVT::f32 && VT.getScalarType() != MVT::f64) | |||
17172 | return SDValue(); | |||
17173 | ||||
17174 | // If estimates are explicitly disabled for this function, we're done. | |||
17175 | MachineFunction &MF = DAG.getMachineFunction(); | |||
17176 | int Enabled = TLI.getRecipEstimateDivEnabled(VT, MF); | |||
17177 | if (Enabled == TLI.ReciprocalEstimate::Disabled) | |||
17178 | return SDValue(); | |||
17179 | ||||
17180 | // Estimates may be explicitly enabled for this type with a custom number of | |||
17181 | // refinement steps. | |||
17182 | int Iterations = TLI.getDivRefinementSteps(VT, MF); | |||
17183 | if (SDValue Est = TLI.getRecipEstimate(Op, DAG, Enabled, Iterations)) { | |||
17184 | AddToWorklist(Est.getNode()); | |||
17185 | ||||
17186 | if (Iterations) { | |||
17187 | EVT VT = Op.getValueType(); | |||
17188 | SDLoc DL(Op); | |||
17189 | SDValue FPOne = DAG.getConstantFP(1.0, DL, VT); | |||
17190 | ||||
17191 | // Newton iterations: Est = Est + Est (1 - Arg * Est) | |||
17192 | for (int i = 0; i < Iterations; ++i) { | |||
17193 | SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Op, Est, Flags); | |||
17194 | AddToWorklist(NewEst.getNode()); | |||
17195 | ||||
17196 | NewEst = DAG.getNode(ISD::FSUB, DL, VT, FPOne, NewEst, Flags); | |||
17197 | AddToWorklist(NewEst.getNode()); | |||
17198 | ||||
17199 | NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst, Flags); | |||
17200 | AddToWorklist(NewEst.getNode()); | |||
17201 | ||||
17202 | Est = DAG.getNode(ISD::FADD, DL, VT, Est, NewEst, Flags); | |||
17203 | AddToWorklist(Est.getNode()); | |||
17204 | } | |||
17205 | } | |||
17206 | return Est; | |||
17207 | } | |||
17208 | ||||
17209 | return SDValue(); | |||
17210 | } | |||
17211 | ||||
17212 | /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i) | |||
17213 | /// For the reciprocal sqrt, we need to find the zero of the function: | |||
17214 | /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)] | |||
17215 | /// => | |||
17216 | /// X_{i+1} = X_i (1.5 - A X_i^2 / 2) | |||
17217 | /// As a result, we precompute A/2 prior to the iteration loop. | |||
17218 | SDValue DAGCombiner::buildSqrtNROneConst(SDValue Arg, SDValue Est, | |||
17219 | unsigned Iterations, | |||
17220 | SDNodeFlags Flags, bool Reciprocal) { | |||
17221 | EVT VT = Arg.getValueType(); | |||
17222 | SDLoc DL(Arg); | |||
17223 | SDValue ThreeHalves = DAG.getConstantFP(1.5, DL, VT); | |||
17224 | ||||
17225 | // We now need 0.5 * Arg which we can write as (1.5 * Arg - Arg) so that | |||
17226 | // this entire sequence requires only one FP constant. | |||
17227 | SDValue HalfArg = DAG.getNode(ISD::FMUL, DL, VT, ThreeHalves, Arg, Flags); | |||
17228 | AddToWorklist(HalfArg.getNode()); | |||
17229 | ||||
17230 | HalfArg = DAG.getNode(ISD::FSUB, DL, VT, HalfArg, Arg, Flags); | |||
17231 | AddToWorklist(HalfArg.getNode()); | |||
17232 | ||||
17233 | // Newton iterations: Est = Est * (1.5 - HalfArg * Est * Est) | |||
17234 | for (unsigned i = 0; i < Iterations; ++i) { | |||
17235 | SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, Est, Flags); | |||
17236 | AddToWorklist(NewEst.getNode()); | |||
17237 | ||||
17238 | NewEst = DAG.getNode(ISD::FMUL, DL, VT, HalfArg, NewEst, Flags); | |||
17239 | AddToWorklist(NewEst.getNode()); | |||
17240 | ||||
17241 | NewEst = DAG.getNode(ISD::FSUB, DL, VT, ThreeHalves, NewEst, Flags); | |||
17242 | AddToWorklist(NewEst.getNode()); | |||
17243 | ||||
17244 | Est = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst, Flags); | |||
17245 | AddToWorklist(Est.getNode()); | |||
17246 | } | |||
17247 | ||||
17248 | // If non-reciprocal square root is requested, multiply the result by Arg. | |||
17249 | if (!Reciprocal) { | |||
17250 | Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Arg, Flags); | |||
17251 | AddToWorklist(Est.getNode()); | |||
17252 | } | |||
17253 | ||||
17254 | return Est; | |||
17255 | } | |||
17256 | ||||
17257 | /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i) | |||
17258 | /// For the reciprocal sqrt, we need to find the zero of the function: | |||
17259 | /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)] | |||
17260 | /// => | |||
17261 | /// X_{i+1} = (-0.5 * X_i) * (A * X_i * X_i + (-3.0)) | |||
17262 | SDValue DAGCombiner::buildSqrtNRTwoConst(SDValue Arg, SDValue Est, | |||
17263 | unsigned Iterations, | |||
17264 | SDNodeFlags Flags, bool Reciprocal) { | |||
17265 | EVT VT = Arg.getValueType(); | |||
17266 | SDLoc DL(Arg); | |||
17267 | SDValue MinusThree = DAG.getConstantFP(-3.0, DL, VT); | |||
17268 | SDValue MinusHalf = DAG.getConstantFP(-0.5, DL, VT); | |||
17269 | ||||
17270 | // This routine must enter the loop below to work correctly | |||
17271 | // when (Reciprocal == false). | |||
17272 | assert(Iterations > 0)(static_cast <bool> (Iterations > 0) ? void (0) : __assert_fail ("Iterations > 0", "/build/llvm-toolchain-snapshot-7~svn326246/lib/CodeGen/SelectionDAG/DAGCombiner.cpp" , 17272, __extension__ __PRETTY_FUNCTION__)); | |||
17273 | ||||
17274 | // Newton iterations for reciprocal square root: | |||
17275 | // E = (E * -0.5) * ((A * E) * E + -3.0) | |||
17276 | for (unsigned i = 0; i < Iterations; ++i) { | |||
17277 | SDValue AE = DAG.getNode(ISD::FMUL, DL, VT, Arg, Est, Flags); | |||
17278 | AddToWorklist(AE.getNode()); | |||
17279 | ||||
17280 | SDValue AEE = DAG.getNode(ISD::FMUL, DL, VT, AE, Est, Flags); | |||
17281 | AddToWorklist(AEE.getNode()); | |||
17282 | ||||
17283 | SDValue RHS = DAG.getNode(ISD::FADD, DL, VT, AEE, MinusThree, Flags); | |||
17284 | AddToWorklist(RHS.getNode()); | |||
17285 | ||||
17286 | // When calculating a square root at the last iteration build: | |||
17287 | // S = ((A * E) * -0.5) * ((A * E) * E + -3.0) | |||
17288 | // (notice a common subexpression) | |||
17289 | SDValue LHS; | |||
17290 | if (Reciprocal || (i + 1) < Iterations) { | |||
17291 | // RSQRT: LHS = (E * -0.5) | |||
17292 | LHS = DAG.getNode(ISD::FMUL, DL, VT, Est, MinusHalf, Flags); | |||
17293 | } else { | |||
17294 | // SQRT: LHS = (A * E) * -0.5 | |||
17295 | LHS = DAG.getNode(ISD::FMUL, DL, VT, AE, MinusHalf, Flags); | |||
17296 | } | |||
17297 | AddToWorklist(LHS.getNode()); | |||
17298 | ||||
17299 | Est = DAG.getNode(ISD::FMUL, DL, VT, LHS, RHS, Flags); | |||
17300 | AddToWorklist(Est.getNode()); | |||
17301 | } | |||
17302 | ||||
17303 | return Est; | |||
17304 | } | |||
17305 | ||||
17306 | /// Build code to calculate either rsqrt(Op) or sqrt(Op). In the latter case | |||
17307 | /// Op*rsqrt(Op) is actually computed, so additional postprocessing is needed if | |||
17308 | /// Op can be zero. | |||
17309 | SDValue DAGCombiner::buildSqrtEstimateImpl(SDValue Op, SDNodeFlags Flags, | |||
17310 | bool Reciprocal) { | |||
17311 | if (Level >= AfterLegalizeDAG) | |||
17312 | return SDValue(); | |||
17313 | ||||
17314 | // TODO: Handle half and/or extended types? | |||
17315 | EVT VT = Op.getValueType(); | |||
17316 | if (VT.getScalarType() != MVT::f32 && VT.getScalarType() != MVT::f64) | |||
17317 | return SDValue(); | |||
17318 | ||||
17319 | // If estimates are explicitly disabled for this function, we're done. | |||
17320 | MachineFunction &MF = DAG.getMachineFunction(); | |||
17321 | int Enabled = TLI.getRecipEstimateSqrtEnabled(VT, MF); | |||
17322 | if (Enabled == TLI.ReciprocalEstimate::Disabled) | |||
17323 | return SDValue(); | |||
17324 | ||||
17325 | // Estimates may be explicitly enabled for this type with a custom number of | |||
17326 | // refinement steps. | |||
17327 | int Iterations = TLI.getSqrtRefinementSteps(VT, MF); | |||
17328 | ||||
17329 | bool UseOneConstNR = false; | |||
17330 | if (SDValue Est = | |||
17331 | TLI.getSqrtEstimate(Op, DAG, Enabled, Iterations, UseOneConstNR, | |||
17332 | Reciprocal)) { | |||
17333 | AddToWorklist(Est.getNode()); | |||
17334 | ||||
17335 | if (Iterations) { | |||
17336 | Est = UseOneConstNR | |||
17337 | ? buildSqrtNROneConst(Op, Est, Iterations, Flags, Reciprocal) | |||
17338 | : buildSqrtNRTwoConst(Op, Est, Iterations, Flags, Reciprocal); | |||
17339 | ||||
17340 | if (!Reciprocal) { | |||
17341 | // The estimate is now completely wrong if the input was exactly 0.0 or | |||
17342 | // possibly a denormal. Force the answer to 0.0 for those cases. | |||
17343 | EVT VT = Op.getValueType(); | |||
17344 | SDLoc DL(Op); | |||
17345 | EVT CCVT = getSetCCResultType(VT); | |||
17346 | ISD::NodeType SelOpcode = VT.isVector() ? ISD::VSELECT : ISD::SELECT; | |||
17347 | const Function &F = DAG.getMachineFunction().getFunction(); | |||
17348 | Attribute Denorms = F.getFnAttribute("denormal-fp-math"); | |||
17349 | if (Denorms.getValueAsString().equals("ieee")) { | |||
17350 | // fabs(X) < SmallestNormal ? 0.0 : Est | |||
17351 | const fltSemantics &FltSem = DAG.EVTToAPFloatSemantics(VT); | |||
17352 | APFloat SmallestNorm = APFloat::getSmallestNormalized(FltSem); | |||
17353 | SDValue NormC = DAG.getConstantFP(SmallestNorm, DL, VT); | |||
17354 | SDValue FPZero = DAG.getConstantFP(0.0, DL, VT); | |||
17355 | SDValue Fabs = DAG.getNode(ISD::FABS, DL, VT, Op); | |||
17356 | SDValue IsDenorm = DAG.getSetCC(DL, CCVT, Fabs, NormC, ISD::SETLT); | |||
17357 | Est = DAG.getNode(SelOpcode, DL, VT, IsDenorm, FPZero, Est); | |||
17358 | AddToWorklist(Fabs.getNode()); | |||
17359 | AddToWorklist(IsDenorm.getNode()); | |||
17360 | AddToWorklist(Est.getNode()); | |||
17361 | } else { | |||
17362 | // X == 0.0 ? 0.0 : Est | |||
17363 | SDValue FPZero = DAG.getConstantFP(0.0, DL, VT); | |||
17364 | SDValue IsZero = DAG.getSetCC(DL, CCVT, Op, FPZero, ISD::SETEQ); | |||
17365 | Est = DAG.getNode(SelOpcode, DL, VT, IsZero, FPZero, Est); | |||
17366 | AddToWorklist(IsZero.getNode()); | |||
17367 | AddToWorklist(Est.getNode()); | |||
17368 | } | |||
17369 | } | |||
17370 | } | |||
17371 | return Est; | |||
17372 | } | |||
17373 | ||||
17374 | return SDValue(); | |||
17375 | } | |||
17376 | ||||
17377 | SDValue DAGCombiner::buildRsqrtEstimate(SDValue Op, SDNodeFlags Flags) { | |||
17378 | return buildSqrtEstimateImpl(Op, Flags, true); | |||
17379 | } | |||
17380 | ||||
17381 | SDValue DAGCombiner::buildSqrtEstimate(SDValue Op, SDNodeFlags Flags) { | |||
17382 | return buildSqrtEstimateImpl(Op, Flags, false); | |||
17383 | } | |||
17384 | ||||
17385 | /// Return true if there is any possibility that the two addresses overlap. | |||
17386 | bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const { | |||
17387 | // If they are the same then they must be aliases. | |||
17388 | if (Op0->getBasePtr() == Op1->getBasePtr()) return true; | |||
17389 | ||||
17390 | // If they are both volatile then they cannot be reordered. | |||
17391 | if (Op0->isVolatile() && Op1->isVolatile()) return true; | |||
17392 | ||||
17393 | // If one operation reads from invariant memory, and the other may store, they | |||
17394 | // cannot alias. These should really be checking the equivalent of mayWrite, | |||
17395 | // but it only matters for memory nodes other than load /store. | |||
17396 | if (Op0->isInvariant() && Op1->writeMem()) | |||
17397 | return false; | |||
17398 | ||||
17399 | if (Op1->isInvariant() && Op0->writeMem()) | |||
17400 | return false; | |||
17401 | ||||
17402 | unsigned NumBytes0 = Op0->getMemoryVT().getStoreSize(); | |||
17403 | unsigned NumBytes1 = Op1->getMemoryVT().getStoreSize(); | |||
17404 | ||||
17405 | // Check for BaseIndexOffset matching. | |||
17406 | BaseIndexOffset BasePtr0 = BaseIndexOffset::match(Op0, DAG); | |||
17407 | BaseIndexOffset BasePtr1 = BaseIndexOffset::match(Op1, DAG); | |||
17408 | int64_t PtrDiff; | |||
17409 | if (BasePtr0.getBase().getNode() && BasePtr1.getBase().getNode()) { | |||
17410 | if (BasePtr0.equalBaseIndex(BasePtr1, DAG, PtrDiff)) | |||
17411 | return !((NumBytes0 <= PtrDiff) || (PtrDiff + NumBytes1 <= 0)); | |||
17412 | ||||
17413 | // If both BasePtr0 and BasePtr1 are FrameIndexes, we will not be | |||
17414 | // able to calculate their relative offset if at least one arises | |||
17415 | // from an alloca. However, these allocas cannot overlap and we | |||
17416 | // can infer there is no alias. | |||
17417 | if (auto *A = dyn_cast<FrameIndexSDNode>(BasePtr0.getBase())) | |||
17418 | if (auto *B = dyn_cast<FrameIndexSDNode>(BasePtr1.getBase())) { | |||
17419 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); | |||
17420 | // If the base are the same frame index but the we couldn't find a | |||
17421 | // constant offset, (indices are different) be conservative. | |||
17422 | if (A != B && (!MFI.isFixedObjectIndex(A->getIndex()) || | |||
17423 | !MFI.isFixedObjectIndex(B->getIndex()))) | |||
17424 | return false; | |||
17425 | } | |||
17426 | ||||
17427 | bool IsFI0 = isa<FrameIndexSDNode>(BasePtr0.getBase()); | |||
17428 | bool IsFI1 = isa<FrameIndexSDNode>(BasePtr1.getBase()); | |||
17429 | bool IsGV0 = isa<GlobalAddressSDNode>(BasePtr0.getBase()); | |||
17430 | bool IsGV1 = isa<GlobalAddressSDNode>(BasePtr1.getBase()); | |||
17431 | bool IsCV0 = isa<ConstantPoolSDNode>(BasePtr0.getBase()); | |||
17432 | bool IsCV1 = isa<ConstantPoolSDNode>(BasePtr1.getBase()); | |||
17433 | ||||
17434 | // If of mismatched base types or checkable indices we can check | |||
17435 | // they do not alias. | |||
17436 | if ((BasePtr0.getIndex() == BasePtr1.getIndex() || (IsFI0 != IsFI1) || | |||
17437 | (IsGV0 != IsGV1) || (IsCV0 != IsCV1)) && | |||
17438 | (IsFI0 || IsGV0 || IsCV0) && (IsFI1 || IsGV1 || IsCV1)) | |||
17439 | return false; | |||
17440 | } | |||
17441 | ||||
17442 | // If we know required SrcValue1 and SrcValue2 have relatively large | |||
17443 | // alignment compared to the size and offset of the access, we may be able | |||
17444 | // to prove they do not alias. This check is conservative for now to catch | |||
17445 | // cases created by splitting vector types. | |||
17446 | int64_t SrcValOffset0 = Op0->getSrcValueOffset(); | |||
17447 | int64_t SrcValOffset1 = Op1->getSrcValueOffset(); | |||
17448 | unsigned OrigAlignment0 = Op0->getOriginalAlignment(); | |||
17449 | unsigned OrigAlignment1 = Op1->getOriginalAlignment(); | |||
17450 | if (OrigAlignment0 == OrigAlignment1 && SrcValOffset0 != SrcValOffset1 && | |||
17451 | NumBytes0 == NumBytes1 && OrigAlignment0 > NumBytes0) { | |||
17452 | int64_t OffAlign0 = SrcValOffset0 % OrigAlignment0; | |||
17453 | int64_t OffAlign1 = SrcValOffset1 % OrigAlignment1; | |||
17454 | ||||
17455 | // There is no overlap between these relatively aligned accesses of | |||
17456 | // similar size. Return no alias. | |||
17457 | if ((OffAlign0 + NumBytes0) <= OffAlign1 || | |||
17458 | (OffAlign1 + NumBytes1) <= OffAlign0) | |||
17459 | return false; | |||
17460 | } | |||
17461 | ||||
17462 | bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0 | |||
17463 | ? CombinerGlobalAA | |||
17464 | : DAG.getSubtarget().useAA(); | |||
17465 | #ifndef NDEBUG | |||
17466 | if (CombinerAAOnlyFunc.getNumOccurrences() && | |||
17467 | CombinerAAOnlyFunc != DAG.getMachineFunction().getName()) | |||
17468 | UseAA = false; | |||
17469 | #endif | |||
17470 | ||||
17471 | if (UseAA && AA && | |||
17472 | Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) { | |||
17473 | // Use alias analysis information. | |||
17474 | int64_t MinOffset = std::min(SrcValOffset0, SrcValOffset1); | |||
17475 | int64_t Overlap0 = NumBytes0 + SrcValOffset0 - MinOffset; | |||
17476 | int64_t Overlap1 = NumBytes1 + SrcValOffset1 - MinOffset; | |||
17477 | AliasResult AAResult = | |||
17478 | AA->alias(MemoryLocation(Op0->getMemOperand()->getValue(), Overlap0, | |||
17479 | UseTBAA ? Op0->getAAInfo() : AAMDNodes()), | |||
17480 | MemoryLocation(Op1->getMemOperand()->getValue(), Overlap1, | |||
17481 | UseTBAA ? Op1->getAAInfo() : AAMDNodes()) ); | |||
17482 | if (AAResult == NoAlias) | |||
17483 | return false; | |||
17484 | } | |||
17485 | ||||
17486 | // Otherwise we have to assume they alias. | |||
17487 | return true; | |||
17488 | } | |||
17489 | ||||
17490 | /// Walk up chain skipping non-aliasing memory nodes, | |||
17491 | /// looking for aliasing nodes and adding them to the Aliases vector. | |||
17492 | void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain, | |||
17493 | SmallVectorImpl<SDValue> &Aliases) { | |||
17494 | SmallVector<SDValue, 8> Chains; // List of chains to visit. | |||
17495 | SmallPtrSet<SDNode *, 16> Visited; // Visited node set. | |||
17496 | ||||
17497 | // Get alias information for node. | |||
17498 | bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile(); | |||
17499 | ||||
17500 | // Starting off. | |||
17501 | Chains.push_back(OriginalChain); | |||
17502 | unsigned Depth = 0; | |||
17503 | ||||
17504 | // Look at each chain and determine if it is an alias. If so, add it to the | |||
17505 | // aliases list. If not, then continue up the chain looking for the next | |||
17506 | // candidate. | |||
17507 | while (!Chains.empty()) { | |||
17508 | SDValue Chain = Chains.pop_back_val(); | |||
17509 | ||||
17510 | // For TokenFactor nodes, look at each operand and only continue up the | |||
17511 | // chain until we reach the depth limit. | |||
17512 | // | |||
17513 | // FIXME: The depth check could be made to return the last non-aliasing | |||
17514 | // chain we found before we hit a tokenfactor rather than the original | |||
17515 | // chain. | |||
17516 | if (Depth > TLI.getGatherAllAliasesMaxDepth()) { | |||
17517 | Aliases.clear(); | |||
17518 | Aliases.push_back(OriginalChain); | |||
17519 | return; | |||
17520 | } | |||
17521 | ||||
17522 | // Don't bother if we've been before. | |||
17523 | if (!Visited.insert(Chain.getNode()).second) | |||
17524 | continue; | |||
17525 | ||||
17526 | switch (Chain.getOpcode()) { | |||
17527 | case ISD::EntryToken: | |||
17528 | // Entry token is ideal chain operand, but handled in FindBetterChain. | |||
17529 | break; | |||
17530 | ||||
17531 | case ISD::LOAD: | |||
17532 | case ISD::STORE: { | |||
17533 | // Get alias information for Chain. | |||
17534 | bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) && | |||
17535 | !cast<LSBaseSDNode>(Chain.getNode())->isVolatile(); | |||
17536 | ||||
17537 | // If chain is alias then stop here. | |||
17538 | if (!(IsLoad && IsOpLoad) && | |||
17539 | isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) { | |||
17540 | Aliases.push_back(Chain); | |||
17541 | } else { | |||
17542 | // Look further up the chain. | |||
17543 | Chains.push_back(Chain.getOperand(0)); | |||
17544 | ++Depth; | |||
17545 | } | |||
17546 | break; | |||
17547 | } | |||
17548 | ||||
17549 | case ISD::TokenFactor: | |||
17550 | // We have to check each of the operands of the token factor for "small" | |||
17551 | // token factors, so we queue them up. Adding the operands to the queue | |||
17552 | // (stack) in reverse order maintains the original order and increases the | |||
17553 | // likelihood that getNode will find a matching token factor (CSE.) | |||
17554 | if (Chain.getNumOperands() > 16) { | |||
17555 | Aliases.push_back(Chain); | |||
17556 | break; | |||
17557 | } | |||
17558 | for (unsigned n = Chain.getNumOperands(); n;) | |||
17559 | Chains.push_back(Chain.getOperand(--n)); | |||
17560 | ++Depth; | |||
17561 | break; | |||
17562 | ||||
17563 | case ISD::CopyFromReg: | |||
17564 | // Forward past CopyFromReg. | |||
17565 | Chains.push_back(Chain.getOperand(0)); | |||
17566 | ++Depth; | |||
17567 | break; | |||
17568 | ||||
17569 | default: | |||
17570 | // For all other instructions we will just have to take what we can get. | |||
17571 | Aliases.push_back(Chain); | |||
17572 | break; | |||
17573 | } | |||
17574 | } | |||
17575 | } | |||
17576 | ||||
17577 | /// Walk up chain skipping non-aliasing memory nodes, looking for a better chain | |||
17578 | /// (aliasing node.) | |||
17579 | SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) { | |||
17580 | if (OptLevel == CodeGenOpt::None) | |||
17581 | return OldChain; | |||
17582 | ||||
17583 | // Ops for replacing token factor. | |||
17584 | SmallVector<SDValue, 8> Aliases; | |||
17585 | ||||
17586 | // Accumulate all the aliases to this node. | |||
17587 | GatherAllAliases(N, OldChain, Aliases); | |||
17588 | ||||
17589 | // If no operands then chain to entry token. | |||
17590 | if (Aliases.size() == 0) | |||
17591 | return DAG.getEntryNode(); | |||
17592 | ||||
17593 | // If a single operand then chain to it. We don't need to revisit it. | |||
17594 | if (Aliases.size() == 1) | |||
17595 | return Aliases[0]; | |||
17596 | ||||
17597 | // Construct a custom tailored token factor. | |||
17598 | return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases); | |||
17599 | } | |||
17600 | ||||
17601 | // This function tries to collect a bunch of potentially interesting | |||
17602 | // nodes to improve the chains of, all at once. This might seem | |||
17603 | // redundant, as this function gets called when visiting every store | |||
17604 | // node, so why not let the work be done on each store as it's visited? | |||
17605 | // | |||
17606 | // I believe this is mainly important because MergeConsecutiveStores | |||
17607 | // is unable to deal with merging stores of different sizes, so unless | |||
17608 | // we improve the chains of all the potential candidates up-front | |||
17609 | // before running MergeConsecutiveStores, it might only see some of | |||
17610 | // the nodes that will eventually be candidates, and then not be able | |||
17611 | // to go from a partially-merged state to the desired final | |||
17612 | // fully-merged state. | |||
17613 | bool DAGCombiner::findBetterNeighborChains(StoreSDNode *St) { | |||
17614 | if (OptLevel == CodeGenOpt::None) | |||
17615 | return false; | |||
17616 | ||||
17617 | // This holds the base pointer, index, and the offset in bytes from the base | |||
17618 | // pointer. | |||
17619 | BaseIndexOffset BasePtr = BaseIndexOffset::match(St, DAG); | |||
17620 | ||||
17621 | // We must have a base and an offset. | |||
17622 | if (!BasePtr.getBase().getNode()) | |||
17623 | return false; | |||
17624 | ||||
17625 | // Do not handle stores to undef base pointers. | |||
17626 | if (BasePtr.getBase().isUndef()) | |||
17627 | return false; | |||
17628 | ||||
17629 | SmallVector<StoreSDNode *, 8> ChainedStores; | |||
17630 | ChainedStores.push_back(St); | |||
17631 | ||||
17632 | // Walk up the chain and look for nodes with offsets from the same | |||
17633 | // base pointer. Stop when reaching an instruction with a different kind | |||
17634 | // or instruction which has a different base pointer. | |||
17635 | StoreSDNode *Index = St; | |||
17636 | while (Index) { | |||
17637 | // If the chain has more than one use, then we can't reorder the mem ops. | |||
17638 | if (Index != St && !SDValue(Index, 0)->hasOneUse()) | |||
17639 | break; | |||
17640 | ||||
17641 | if (Index->isVolatile() || Index->isIndexed()) | |||
17642 | break; | |||
17643 | ||||
17644 | // Find the base pointer and offset for this memory node. | |||
17645 | BaseIndexOffset Ptr = BaseIndexOffset::match(Index, DAG); | |||
17646 | ||||
17647 | // Check that the base pointer is the same as the original one. | |||
17648 | if (!BasePtr.equalBaseIndex(Ptr, DAG)) | |||
17649 | break; | |||
17650 | ||||
17651 | // Walk up the chain to find the next store node, ignoring any | |||
17652 | // intermediate loads. Any other kind of node will halt the loop. | |||
17653 | SDNode *NextInChain = Index->getChain().getNode(); | |||
17654 | while (true) { | |||
17655 | if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) { | |||
17656 | // We found a store node. Use it for the next iteration. | |||
17657 | if (STn->isVolatile() || STn->isIndexed()) { | |||
17658 | Index = nullptr; | |||
17659 | break; | |||
17660 | } | |||
17661 | ChainedStores.push_back(STn); | |||
17662 | Index = STn; | |||
17663 | break; | |||
17664 | } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) { | |||
17665 | NextInChain = Ldn->getChain().getNode(); | |||
17666 | continue; | |||
17667 | } else { | |||
17668 | Index = nullptr; | |||
17669 | break; | |||
17670 | } | |||
17671 | } // end while | |||
17672 | } | |||
17673 | ||||
17674 | // At this point, ChainedStores lists all of the Store nodes | |||
17675 | // reachable by iterating up through chain nodes matching the above | |||
17676 | // conditions. For each such store identified, try to find an | |||
17677 | // earlier chain to attach the store to which won't violate the | |||
17678 | // required ordering. | |||
17679 | bool MadeChangeToSt = false; | |||
17680 | SmallVector<std::pair<StoreSDNode *, SDValue>, 8> BetterChains; | |||
17681 | ||||
17682 | for (StoreSDNode *ChainedStore : ChainedStores) { | |||
17683 | SDValue Chain = ChainedStore->getChain(); | |||
17684 | SDValue BetterChain = FindBetterChain(ChainedStore, Chain); | |||
17685 | ||||
17686 | if (Chain != BetterChain) { | |||
17687 | if (ChainedStore == St) | |||
17688 | MadeChangeToSt = true; | |||
17689 | BetterChains.push_back(std::make_pair(ChainedStore, BetterChain)); | |||
17690 | } | |||
17691 | } | |||
17692 | ||||
17693 | // Do all replacements after finding the replacements to make to avoid making | |||
17694 | // the chains more complicated by introducing new TokenFactors. | |||
17695 | for (auto Replacement : BetterChains) | |||
17696 | replaceStoreChain(Replacement.first, Replacement.second); | |||
17697 | ||||
17698 | return MadeChangeToSt; | |||
17699 | } | |||
17700 | ||||
17701 | /// This is the entry point for the file. | |||
17702 | void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis *AA, | |||
17703 | CodeGenOpt::Level OptLevel) { | |||
17704 | /// This is the main entry point to this class. | |||
17705 | DAGCombiner(*this, AA, OptLevel).Run(Level); | |||
17706 | } |
1 | //===-- llvm/ADT/APInt.h - For Arbitrary Precision Integer -----*- C++ -*--===// | |||
2 | // | |||
3 | // The LLVM Compiler Infrastructure | |||
4 | // | |||
5 | // This file is distributed under the University of Illinois Open Source | |||
6 | // License. See LICENSE.TXT for details. | |||
7 | // | |||
8 | //===----------------------------------------------------------------------===// | |||
9 | /// | |||
10 | /// \file | |||
11 | /// \brief This file implements a class to represent arbitrary precision | |||
12 | /// integral constant values and operations on them. | |||
13 | /// | |||
14 | //===----------------------------------------------------------------------===// | |||
15 | ||||
16 | #ifndef LLVM_ADT_APINT_H | |||
17 | #define LLVM_ADT_APINT_H | |||
18 | ||||
19 | #include "llvm/Support/Compiler.h" | |||
20 | #include "llvm/Support/MathExtras.h" | |||
21 | #include <cassert> | |||
22 | #include <climits> | |||
23 | #include <cstring> | |||
24 | #include <string> | |||
25 | ||||
26 | namespace llvm { | |||
27 | class FoldingSetNodeID; | |||
28 | class StringRef; | |||
29 | class hash_code; | |||
30 | class raw_ostream; | |||
31 | ||||
32 | template <typename T> class SmallVectorImpl; | |||
33 | template <typename T> class ArrayRef; | |||
34 | ||||
35 | class APInt; | |||
36 | ||||
37 | inline APInt operator-(APInt); | |||
38 | ||||
39 | //===----------------------------------------------------------------------===// | |||
40 | // APInt Class | |||
41 | //===----------------------------------------------------------------------===// | |||
42 | ||||
43 | /// \brief Class for arbitrary precision integers. | |||
44 | /// | |||
45 | /// APInt is a functional replacement for common case unsigned integer type like | |||
46 | /// "unsigned", "unsigned long" or "uint64_t", but also allows non-byte-width | |||
47 | /// integer sizes and large integer value types such as 3-bits, 15-bits, or more | |||
48 | /// than 64-bits of precision. APInt provides a variety of arithmetic operators | |||
49 | /// and methods to manipulate integer values of any bit-width. It supports both | |||
50 | /// the typical integer arithmetic and comparison operations as well as bitwise | |||
51 | /// manipulation. | |||
52 | /// | |||
53 | /// The class has several invariants worth noting: | |||
54 | /// * All bit, byte, and word positions are zero-based. | |||
55 | /// * Once the bit width is set, it doesn't change except by the Truncate, | |||
56 | /// SignExtend, or ZeroExtend operations. | |||
57 | /// * All binary operators must be on APInt instances of the same bit width. | |||
58 | /// Attempting to use these operators on instances with different bit | |||
59 | /// widths will yield an assertion. | |||
60 | /// * The value is stored canonically as an unsigned value. For operations | |||
61 | /// where it makes a difference, there are both signed and unsigned variants | |||
62 | /// of the operation. For example, sdiv and udiv. However, because the bit | |||
63 | /// widths must be the same, operations such as Mul and Add produce the same | |||
64 | /// results regardless of whether the values are interpreted as signed or | |||
65 | /// not. | |||
66 | /// * In general, the class tries to follow the style of computation that LLVM | |||
67 | /// uses in its IR. This simplifies its use for LLVM. | |||
68 | /// | |||
69 | class LLVM_NODISCARD[[clang::warn_unused_result]] APInt { | |||
70 | public: | |||
71 | typedef uint64_t WordType; | |||
72 | ||||
73 | /// This enum is used to hold the constants we needed for APInt. | |||
74 | enum : unsigned { | |||
75 | /// Byte size of a word. | |||
76 | APINT_WORD_SIZE = sizeof(WordType), | |||
77 | /// Bits in a word. | |||
78 | APINT_BITS_PER_WORD = APINT_WORD_SIZE * CHAR_BIT8 | |||
79 | }; | |||
80 | ||||
81 | static const WordType WORD_MAX = ~WordType(0); | |||
82 | ||||
83 | private: | |||
84 | /// This union is used to store the integer value. When the | |||
85 | /// integer bit-width <= 64, it uses VAL, otherwise it uses pVal. | |||
86 | union { | |||
87 | uint64_t VAL; ///< Used to store the <= 64 bits integer value. | |||
88 | uint64_t *pVal; ///< Used to store the >64 bits integer value. | |||
89 | } U; | |||
90 | ||||
91 | unsigned BitWidth; ///< The number of bits in this APInt. | |||
92 | ||||
93 | friend struct DenseMapAPIntKeyInfo; | |||
94 | ||||
95 | friend class APSInt; | |||
96 | ||||
97 | /// \brief Fast internal constructor | |||
98 | /// | |||
99 | /// This constructor is used only internally for speed of construction of | |||
100 | /// temporaries. It is unsafe for general use so it is not public. | |||
101 | APInt(uint64_t *val, unsigned bits) : BitWidth(bits) { | |||
102 | U.pVal = val; | |||
103 | } | |||
104 | ||||
105 | /// \brief Determine if this APInt just has one word to store value. | |||
106 | /// | |||
107 | /// \returns true if the number of bits <= 64, false otherwise. | |||
108 | bool isSingleWord() const { return BitWidth <= APINT_BITS_PER_WORD; } | |||
109 | ||||
110 | /// \brief Determine which word a bit is in. | |||
111 | /// | |||
112 | /// \returns the word position for the specified bit position. | |||
113 | static unsigned whichWord(unsigned bitPosition) { | |||
114 | return bitPosition / APINT_BITS_PER_WORD; | |||
115 | } | |||
116 | ||||
117 | /// \brief Determine which bit in a word a bit is in. | |||
118 | /// | |||
119 | /// \returns the bit position in a word for the specified bit position | |||
120 | /// in the APInt. | |||
121 | static unsigned whichBit(unsigned bitPosition) { | |||
122 | return bitPosition % APINT_BITS_PER_WORD; | |||
123 | } | |||
124 | ||||
125 | /// \brief Get a single bit mask. | |||
126 | /// | |||
127 | /// \returns a uint64_t with only bit at "whichBit(bitPosition)" set | |||
128 | /// This method generates and returns a uint64_t (word) mask for a single | |||
129 | /// bit at a specific bit position. This is used to mask the bit in the | |||
130 | /// corresponding word. | |||
131 | static uint64_t maskBit(unsigned bitPosition) { | |||
132 | return 1ULL << whichBit(bitPosition); | |||
133 | } | |||
134 | ||||
135 | /// \brief Clear unused high order bits | |||
136 | /// | |||
137 | /// This method is used internally to clear the top "N" bits in the high order | |||
138 | /// word that are not used by the APInt. This is needed after the most | |||
139 | /// significant word is assigned a value to ensure that those bits are | |||
140 | /// zero'd out. | |||
141 | APInt &clearUnusedBits() { | |||
142 | // Compute how many bits are used in the final word | |||
143 | unsigned WordBits = ((BitWidth-1) % APINT_BITS_PER_WORD) + 1; | |||
144 | ||||
145 | // Mask out the high bits. | |||
146 | uint64_t mask = WORD_MAX >> (APINT_BITS_PER_WORD - WordBits); | |||
147 | if (isSingleWord()) | |||
148 | U.VAL &= mask; | |||
149 | else | |||
150 | U.pVal[getNumWords() - 1] &= mask; | |||
151 | return *this; | |||
152 | } | |||
153 | ||||
154 | /// \brief Get the word corresponding to a bit position | |||
155 | /// \returns the corresponding word for the specified bit position. | |||
156 | uint64_t getWord(unsigned bitPosition) const { | |||
157 | return isSingleWord() ? U.VAL : U.pVal[whichWord(bitPosition)]; | |||
158 | } | |||
159 | ||||
160 | /// Utility method to change the bit width of this APInt to new bit width, | |||
161 | /// allocating and/or deallocating as necessary. There is no guarantee on the | |||
162 | /// value of any bits upon return. Caller should populate the bits after. | |||
163 | void reallocate(unsigned NewBitWidth); | |||
164 | ||||
165 | /// \brief Convert a char array into an APInt | |||
166 | /// | |||
167 | /// \param radix 2, 8, 10, 16, or 36 | |||
168 | /// Converts a string into a number. The string must be non-empty | |||
169 | /// and well-formed as a number of the given base. The bit-width | |||
170 | /// must be sufficient to hold the result. | |||
171 | /// | |||
172 | /// This is used by the constructors that take string arguments. | |||
173 | /// | |||
174 | /// StringRef::getAsInteger is superficially similar but (1) does | |||
175 | /// not assume that the string is well-formed and (2) grows the | |||
176 | /// result to hold the input. | |||
177 | void fromString(unsigned numBits, StringRef str, uint8_t radix); | |||
178 | ||||
179 | /// \brief An internal division function for dividing APInts. | |||
180 | /// | |||
181 | /// This is used by the toString method to divide by the radix. It simply | |||
182 | /// provides a more convenient form of divide for internal use since KnuthDiv | |||
183 | /// has specific constraints on its inputs. If those constraints are not met | |||
184 | /// then it provides a simpler form of divide. | |||
185 | static void divide(const WordType *LHS, unsigned lhsWords, | |||
186 | const WordType *RHS, unsigned rhsWords, WordType *Quotient, | |||
187 | WordType *Remainder); | |||
188 | ||||
189 | /// out-of-line slow case for inline constructor | |||
190 | void initSlowCase(uint64_t val, bool isSigned); | |||
191 | ||||
192 | /// shared code between two array constructors | |||
193 | void initFromArray(ArrayRef<uint64_t> array); | |||
194 | ||||
195 | /// out-of-line slow case for inline copy constructor | |||
196 | void initSlowCase(const APInt &that); | |||
197 | ||||
198 | /// out-of-line slow case for shl | |||
199 | void shlSlowCase(unsigned ShiftAmt); | |||
200 | ||||
201 | /// out-of-line slow case for lshr. | |||
202 | void lshrSlowCase(unsigned ShiftAmt); | |||
203 | ||||
204 | /// out-of-line slow case for ashr. | |||
205 | void ashrSlowCase(unsigned ShiftAmt); | |||
206 | ||||
207 | /// out-of-line slow case for operator= | |||
208 | void AssignSlowCase(const APInt &RHS); | |||
209 | ||||
210 | /// out-of-line slow case for operator== | |||
211 | bool EqualSlowCase(const APInt &RHS) const LLVM_READONLY__attribute__((__pure__)); | |||
212 | ||||
213 | /// out-of-line slow case for countLeadingZeros | |||
214 | unsigned countLeadingZerosSlowCase() const LLVM_READONLY__attribute__((__pure__)); | |||
215 | ||||
216 | /// out-of-line slow case for countLeadingOnes. | |||
217 | unsigned countLeadingOnesSlowCase() const LLVM_READONLY__attribute__((__pure__)); | |||
218 | ||||
219 | /// out-of-line slow case for countTrailingZeros. | |||
220 | unsigned countTrailingZerosSlowCase() const LLVM_READONLY__attribute__((__pure__)); | |||
221 | ||||
222 | /// out-of-line slow case for countTrailingOnes | |||
223 | unsigned countTrailingOnesSlowCase() const LLVM_READONLY__attribute__((__pure__)); | |||
224 | ||||
225 | /// out-of-line slow case for countPopulation | |||
226 | unsigned countPopulationSlowCase() const LLVM_READONLY__attribute__((__pure__)); | |||
227 | ||||
228 | /// out-of-line slow case for intersects. | |||
229 | bool intersectsSlowCase(const APInt &RHS) const LLVM_READONLY__attribute__((__pure__)); | |||
230 | ||||
231 | /// out-of-line slow case for isSubsetOf. | |||
232 | bool isSubsetOfSlowCase(const APInt &RHS) const LLVM_READONLY__attribute__((__pure__)); | |||
233 | ||||
234 | /// out-of-line slow case for setBits. | |||
235 | void setBitsSlowCase(unsigned loBit, unsigned hiBit); | |||
236 | ||||
237 | /// out-of-line slow case for flipAllBits. | |||
238 | void flipAllBitsSlowCase(); | |||
239 | ||||
240 | /// out-of-line slow case for operator&=. | |||
241 | void AndAssignSlowCase(const APInt& RHS); | |||
242 | ||||
243 | /// out-of-line slow case for operator|=. | |||
244 | void OrAssignSlowCase(const APInt& RHS); | |||
245 | ||||
246 | /// out-of-line slow case for operator^=. | |||
247 | void XorAssignSlowCase(const APInt& RHS); | |||
248 | ||||
249 | /// Unsigned comparison. Returns -1, 0, or 1 if this APInt is less than, equal | |||
250 | /// to, or greater than RHS. | |||
251 | int compare(const APInt &RHS) const LLVM_READONLY__attribute__((__pure__)); | |||
252 | ||||
253 | /// Signed comparison. Returns -1, 0, or 1 if this APInt is less than, equal | |||
254 | /// to, or greater than RHS. | |||
255 | int compareSigned(const APInt &RHS) const LLVM_READONLY__attribute__((__pure__)); | |||
256 | ||||
257 | public: | |||
258 | /// \name Constructors | |||
259 | /// @{ | |||
260 | ||||
261 | /// \brief Create a new APInt of numBits width, initialized as val. | |||
262 | /// | |||
263 | /// If isSigned is true then val is treated as if it were a signed value | |||
264 | /// (i.e. as an int64_t) and the appropriate sign extension to the bit width | |||
265 | /// will be done. Otherwise, no sign extension occurs (high order bits beyond | |||
266 | /// the range of val are zero filled). | |||
267 | /// | |||
268 | /// \param numBits the bit width of the constructed APInt | |||
269 | /// \param val the initial value of the APInt | |||
270 | /// \param isSigned how to treat signedness of val | |||
271 | APInt(unsigned numBits, uint64_t val, bool isSigned = false) | |||
272 | : BitWidth(numBits) { | |||
273 | assert(BitWidth && "bitwidth too small")(static_cast <bool> (BitWidth && "bitwidth too small" ) ? void (0) : __assert_fail ("BitWidth && \"bitwidth too small\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 273, __extension__ __PRETTY_FUNCTION__)); | |||
274 | if (isSingleWord()) { | |||
275 | U.VAL = val; | |||
276 | clearUnusedBits(); | |||
277 | } else { | |||
278 | initSlowCase(val, isSigned); | |||
279 | } | |||
280 | } | |||
281 | ||||
282 | /// \brief Construct an APInt of numBits width, initialized as bigVal[]. | |||
283 | /// | |||
284 | /// Note that bigVal.size() can be smaller or larger than the corresponding | |||
285 | /// bit width but any extraneous bits will be dropped. | |||
286 | /// | |||
287 | /// \param numBits the bit width of the constructed APInt | |||
288 | /// \param bigVal a sequence of words to form the initial value of the APInt | |||
289 | APInt(unsigned numBits, ArrayRef<uint64_t> bigVal); | |||
290 | ||||
291 | /// Equivalent to APInt(numBits, ArrayRef<uint64_t>(bigVal, numWords)), but | |||
292 | /// deprecated because this constructor is prone to ambiguity with the | |||
293 | /// APInt(unsigned, uint64_t, bool) constructor. | |||
294 | /// | |||
295 | /// If this overload is ever deleted, care should be taken to prevent calls | |||
296 | /// from being incorrectly captured by the APInt(unsigned, uint64_t, bool) | |||
297 | /// constructor. | |||
298 | APInt(unsigned numBits, unsigned numWords, const uint64_t bigVal[]); | |||
299 | ||||
300 | /// \brief Construct an APInt from a string representation. | |||
301 | /// | |||
302 | /// This constructor interprets the string \p str in the given radix. The | |||
303 | /// interpretation stops when the first character that is not suitable for the | |||
304 | /// radix is encountered, or the end of the string. Acceptable radix values | |||
305 | /// are 2, 8, 10, 16, and 36. It is an error for the value implied by the | |||
306 | /// string to require more bits than numBits. | |||
307 | /// | |||
308 | /// \param numBits the bit width of the constructed APInt | |||
309 | /// \param str the string to be interpreted | |||
310 | /// \param radix the radix to use for the conversion | |||
311 | APInt(unsigned numBits, StringRef str, uint8_t radix); | |||
312 | ||||
313 | /// Simply makes *this a copy of that. | |||
314 | /// @brief Copy Constructor. | |||
315 | APInt(const APInt &that) : BitWidth(that.BitWidth) { | |||
316 | if (isSingleWord()) | |||
317 | U.VAL = that.U.VAL; | |||
318 | else | |||
319 | initSlowCase(that); | |||
320 | } | |||
321 | ||||
322 | /// \brief Move Constructor. | |||
323 | APInt(APInt &&that) : BitWidth(that.BitWidth) { | |||
324 | memcpy(&U, &that.U, sizeof(U)); | |||
325 | that.BitWidth = 0; | |||
326 | } | |||
327 | ||||
328 | /// \brief Destructor. | |||
329 | ~APInt() { | |||
330 | if (needsCleanup()) | |||
331 | delete[] U.pVal; | |||
332 | } | |||
333 | ||||
334 | /// \brief Default constructor that creates an uninteresting APInt | |||
335 | /// representing a 1-bit zero value. | |||
336 | /// | |||
337 | /// This is useful for object deserialization (pair this with the static | |||
338 | /// method Read). | |||
339 | explicit APInt() : BitWidth(1) { U.VAL = 0; } | |||
340 | ||||
341 | /// \brief Returns whether this instance allocated memory. | |||
342 | bool needsCleanup() const { return !isSingleWord(); } | |||
343 | ||||
344 | /// Used to insert APInt objects, or objects that contain APInt objects, into | |||
345 | /// FoldingSets. | |||
346 | void Profile(FoldingSetNodeID &id) const; | |||
347 | ||||
348 | /// @} | |||
349 | /// \name Value Tests | |||
350 | /// @{ | |||
351 | ||||
352 | /// \brief Determine sign of this APInt. | |||
353 | /// | |||
354 | /// This tests the high bit of this APInt to determine if it is set. | |||
355 | /// | |||
356 | /// \returns true if this APInt is negative, false otherwise | |||
357 | bool isNegative() const { return (*this)[BitWidth - 1]; } | |||
358 | ||||
359 | /// \brief Determine if this APInt Value is non-negative (>= 0) | |||
360 | /// | |||
361 | /// This tests the high bit of the APInt to determine if it is unset. | |||
362 | bool isNonNegative() const { return !isNegative(); } | |||
363 | ||||
364 | /// \brief Determine if sign bit of this APInt is set. | |||
365 | /// | |||
366 | /// This tests the high bit of this APInt to determine if it is set. | |||
367 | /// | |||
368 | /// \returns true if this APInt has its sign bit set, false otherwise. | |||
369 | bool isSignBitSet() const { return (*this)[BitWidth-1]; } | |||
370 | ||||
371 | /// \brief Determine if sign bit of this APInt is clear. | |||
372 | /// | |||
373 | /// This tests the high bit of this APInt to determine if it is clear. | |||
374 | /// | |||
375 | /// \returns true if this APInt has its sign bit clear, false otherwise. | |||
376 | bool isSignBitClear() const { return !isSignBitSet(); } | |||
377 | ||||
378 | /// \brief Determine if this APInt Value is positive. | |||
379 | /// | |||
380 | /// This tests if the value of this APInt is positive (> 0). Note | |||
381 | /// that 0 is not a positive value. | |||
382 | /// | |||
383 | /// \returns true if this APInt is positive. | |||
384 | bool isStrictlyPositive() const { return isNonNegative() && !isNullValue(); } | |||
385 | ||||
386 | /// \brief Determine if all bits are set | |||
387 | /// | |||
388 | /// This checks to see if the value has all bits of the APInt are set or not. | |||
389 | bool isAllOnesValue() const { | |||
390 | if (isSingleWord()) | |||
391 | return U.VAL == WORD_MAX >> (APINT_BITS_PER_WORD - BitWidth); | |||
| ||||
392 | return countTrailingOnesSlowCase() == BitWidth; | |||
393 | } | |||
394 | ||||
395 | /// \brief Determine if all bits are clear | |||
396 | /// | |||
397 | /// This checks to see if the value has all bits of the APInt are clear or | |||
398 | /// not. | |||
399 | bool isNullValue() const { return !*this; } | |||
400 | ||||
401 | /// \brief Determine if this is a value of 1. | |||
402 | /// | |||
403 | /// This checks to see if the value of this APInt is one. | |||
404 | bool isOneValue() const { | |||
405 | if (isSingleWord()) | |||
406 | return U.VAL == 1; | |||
407 | return countLeadingZerosSlowCase() == BitWidth - 1; | |||
408 | } | |||
409 | ||||
410 | /// \brief Determine if this is the largest unsigned value. | |||
411 | /// | |||
412 | /// This checks to see if the value of this APInt is the maximum unsigned | |||
413 | /// value for the APInt's bit width. | |||
414 | bool isMaxValue() const { return isAllOnesValue(); } | |||
415 | ||||
416 | /// \brief Determine if this is the largest signed value. | |||
417 | /// | |||
418 | /// This checks to see if the value of this APInt is the maximum signed | |||
419 | /// value for the APInt's bit width. | |||
420 | bool isMaxSignedValue() const { | |||
421 | if (isSingleWord()) | |||
422 | return U.VAL == ((WordType(1) << (BitWidth - 1)) - 1); | |||
423 | return !isNegative() && countTrailingOnesSlowCase() == BitWidth - 1; | |||
424 | } | |||
425 | ||||
426 | /// \brief Determine if this is the smallest unsigned value. | |||
427 | /// | |||
428 | /// This checks to see if the value of this APInt is the minimum unsigned | |||
429 | /// value for the APInt's bit width. | |||
430 | bool isMinValue() const { return isNullValue(); } | |||
431 | ||||
432 | /// \brief Determine if this is the smallest signed value. | |||
433 | /// | |||
434 | /// This checks to see if the value of this APInt is the minimum signed | |||
435 | /// value for the APInt's bit width. | |||
436 | bool isMinSignedValue() const { | |||
437 | if (isSingleWord()) | |||
438 | return U.VAL == (WordType(1) << (BitWidth - 1)); | |||
439 | return isNegative() && countTrailingZerosSlowCase() == BitWidth - 1; | |||
440 | } | |||
441 | ||||
442 | /// \brief Check if this APInt has an N-bits unsigned integer value. | |||
443 | bool isIntN(unsigned N) const { | |||
444 | assert(N && "N == 0 ???")(static_cast <bool> (N && "N == 0 ???") ? void ( 0) : __assert_fail ("N && \"N == 0 ???\"", "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 444, __extension__ __PRETTY_FUNCTION__)); | |||
445 | return getActiveBits() <= N; | |||
446 | } | |||
447 | ||||
448 | /// \brief Check if this APInt has an N-bits signed integer value. | |||
449 | bool isSignedIntN(unsigned N) const { | |||
450 | assert(N && "N == 0 ???")(static_cast <bool> (N && "N == 0 ???") ? void ( 0) : __assert_fail ("N && \"N == 0 ???\"", "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 450, __extension__ __PRETTY_FUNCTION__)); | |||
451 | return getMinSignedBits() <= N; | |||
452 | } | |||
453 | ||||
454 | /// \brief Check if this APInt's value is a power of two greater than zero. | |||
455 | /// | |||
456 | /// \returns true if the argument APInt value is a power of two > 0. | |||
457 | bool isPowerOf2() const { | |||
458 | if (isSingleWord()) | |||
459 | return isPowerOf2_64(U.VAL); | |||
460 | return countPopulationSlowCase() == 1; | |||
461 | } | |||
462 | ||||
463 | /// \brief Check if the APInt's value is returned by getSignMask. | |||
464 | /// | |||
465 | /// \returns true if this is the value returned by getSignMask. | |||
466 | bool isSignMask() const { return isMinSignedValue(); } | |||
467 | ||||
468 | /// \brief Convert APInt to a boolean value. | |||
469 | /// | |||
470 | /// This converts the APInt to a boolean value as a test against zero. | |||
471 | bool getBoolValue() const { return !!*this; } | |||
472 | ||||
473 | /// If this value is smaller than the specified limit, return it, otherwise | |||
474 | /// return the limit value. This causes the value to saturate to the limit. | |||
475 | uint64_t getLimitedValue(uint64_t Limit = UINT64_MAX(18446744073709551615UL)) const { | |||
476 | return ugt(Limit) ? Limit : getZExtValue(); | |||
477 | } | |||
478 | ||||
479 | /// \brief Check if the APInt consists of a repeated bit pattern. | |||
480 | /// | |||
481 | /// e.g. 0x01010101 satisfies isSplat(8). | |||
482 | /// \param SplatSizeInBits The size of the pattern in bits. Must divide bit | |||
483 | /// width without remainder. | |||
484 | bool isSplat(unsigned SplatSizeInBits) const; | |||
485 | ||||
486 | /// \returns true if this APInt value is a sequence of \param numBits ones | |||
487 | /// starting at the least significant bit with the remainder zero. | |||
488 | bool isMask(unsigned numBits) const { | |||
489 | assert(numBits != 0 && "numBits must be non-zero")(static_cast <bool> (numBits != 0 && "numBits must be non-zero" ) ? void (0) : __assert_fail ("numBits != 0 && \"numBits must be non-zero\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 489, __extension__ __PRETTY_FUNCTION__)); | |||
490 | assert(numBits <= BitWidth && "numBits out of range")(static_cast <bool> (numBits <= BitWidth && "numBits out of range" ) ? void (0) : __assert_fail ("numBits <= BitWidth && \"numBits out of range\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 490, __extension__ __PRETTY_FUNCTION__)); | |||
491 | if (isSingleWord()) | |||
492 | return U.VAL == (WORD_MAX >> (APINT_BITS_PER_WORD - numBits)); | |||
493 | unsigned Ones = countTrailingOnesSlowCase(); | |||
494 | return (numBits == Ones) && | |||
495 | ((Ones + countLeadingZerosSlowCase()) == BitWidth); | |||
496 | } | |||
497 | ||||
498 | /// \returns true if this APInt is a non-empty sequence of ones starting at | |||
499 | /// the least significant bit with the remainder zero. | |||
500 | /// Ex. isMask(0x0000FFFFU) == true. | |||
501 | bool isMask() const { | |||
502 | if (isSingleWord()) | |||
503 | return isMask_64(U.VAL); | |||
504 | unsigned Ones = countTrailingOnesSlowCase(); | |||
505 | return (Ones > 0) && ((Ones + countLeadingZerosSlowCase()) == BitWidth); | |||
506 | } | |||
507 | ||||
508 | /// \brief Return true if this APInt value contains a sequence of ones with | |||
509 | /// the remainder zero. | |||
510 | bool isShiftedMask() const { | |||
511 | if (isSingleWord()) | |||
512 | return isShiftedMask_64(U.VAL); | |||
513 | unsigned Ones = countPopulationSlowCase(); | |||
514 | unsigned LeadZ = countLeadingZerosSlowCase(); | |||
515 | return (Ones + LeadZ + countTrailingZeros()) == BitWidth; | |||
516 | } | |||
517 | ||||
518 | /// @} | |||
519 | /// \name Value Generators | |||
520 | /// @{ | |||
521 | ||||
522 | /// \brief Gets maximum unsigned value of APInt for specific bit width. | |||
523 | static APInt getMaxValue(unsigned numBits) { | |||
524 | return getAllOnesValue(numBits); | |||
525 | } | |||
526 | ||||
527 | /// \brief Gets maximum signed value of APInt for a specific bit width. | |||
528 | static APInt getSignedMaxValue(unsigned numBits) { | |||
529 | APInt API = getAllOnesValue(numBits); | |||
530 | API.clearBit(numBits - 1); | |||
531 | return API; | |||
532 | } | |||
533 | ||||
534 | /// \brief Gets minimum unsigned value of APInt for a specific bit width. | |||
535 | static APInt getMinValue(unsigned numBits) { return APInt(numBits, 0); } | |||
536 | ||||
537 | /// \brief Gets minimum signed value of APInt for a specific bit width. | |||
538 | static APInt getSignedMinValue(unsigned numBits) { | |||
539 | APInt API(numBits, 0); | |||
540 | API.setBit(numBits - 1); | |||
541 | return API; | |||
542 | } | |||
543 | ||||
544 | /// \brief Get the SignMask for a specific bit width. | |||
545 | /// | |||
546 | /// This is just a wrapper function of getSignedMinValue(), and it helps code | |||
547 | /// readability when we want to get a SignMask. | |||
548 | static APInt getSignMask(unsigned BitWidth) { | |||
549 | return getSignedMinValue(BitWidth); | |||
550 | } | |||
551 | ||||
552 | /// \brief Get the all-ones value. | |||
553 | /// | |||
554 | /// \returns the all-ones value for an APInt of the specified bit-width. | |||
555 | static APInt getAllOnesValue(unsigned numBits) { | |||
556 | return APInt(numBits, WORD_MAX, true); | |||
557 | } | |||
558 | ||||
559 | /// \brief Get the '0' value. | |||
560 | /// | |||
561 | /// \returns the '0' value for an APInt of the specified bit-width. | |||
562 | static APInt getNullValue(unsigned numBits) { return APInt(numBits, 0); } | |||
563 | ||||
564 | /// \brief Compute an APInt containing numBits highbits from this APInt. | |||
565 | /// | |||
566 | /// Get an APInt with the same BitWidth as this APInt, just zero mask | |||
567 | /// the low bits and right shift to the least significant bit. | |||
568 | /// | |||
569 | /// \returns the high "numBits" bits of this APInt. | |||
570 | APInt getHiBits(unsigned numBits) const; | |||
571 | ||||
572 | /// \brief Compute an APInt containing numBits lowbits from this APInt. | |||
573 | /// | |||
574 | /// Get an APInt with the same BitWidth as this APInt, just zero mask | |||
575 | /// the high bits. | |||
576 | /// | |||
577 | /// \returns the low "numBits" bits of this APInt. | |||
578 | APInt getLoBits(unsigned numBits) const; | |||
579 | ||||
580 | /// \brief Return an APInt with exactly one bit set in the result. | |||
581 | static APInt getOneBitSet(unsigned numBits, unsigned BitNo) { | |||
582 | APInt Res(numBits, 0); | |||
583 | Res.setBit(BitNo); | |||
584 | return Res; | |||
585 | } | |||
586 | ||||
587 | /// \brief Get a value with a block of bits set. | |||
588 | /// | |||
589 | /// Constructs an APInt value that has a contiguous range of bits set. The | |||
590 | /// bits from loBit (inclusive) to hiBit (exclusive) will be set. All other | |||
591 | /// bits will be zero. For example, with parameters(32, 0, 16) you would get | |||
592 | /// 0x0000FFFF. If hiBit is less than loBit then the set bits "wrap". For | |||
593 | /// example, with parameters (32, 28, 4), you would get 0xF000000F. | |||
594 | /// | |||
595 | /// \param numBits the intended bit width of the result | |||
596 | /// \param loBit the index of the lowest bit set. | |||
597 | /// \param hiBit the index of the highest bit set. | |||
598 | /// | |||
599 | /// \returns An APInt value with the requested bits set. | |||
600 | static APInt getBitsSet(unsigned numBits, unsigned loBit, unsigned hiBit) { | |||
601 | APInt Res(numBits, 0); | |||
602 | Res.setBits(loBit, hiBit); | |||
603 | return Res; | |||
604 | } | |||
605 | ||||
606 | /// \brief Get a value with upper bits starting at loBit set. | |||
607 | /// | |||
608 | /// Constructs an APInt value that has a contiguous range of bits set. The | |||
609 | /// bits from loBit (inclusive) to numBits (exclusive) will be set. All other | |||
610 | /// bits will be zero. For example, with parameters(32, 12) you would get | |||
611 | /// 0xFFFFF000. | |||
612 | /// | |||
613 | /// \param numBits the intended bit width of the result | |||
614 | /// \param loBit the index of the lowest bit to set. | |||
615 | /// | |||
616 | /// \returns An APInt value with the requested bits set. | |||
617 | static APInt getBitsSetFrom(unsigned numBits, unsigned loBit) { | |||
618 | APInt Res(numBits, 0); | |||
619 | Res.setBitsFrom(loBit); | |||
620 | return Res; | |||
621 | } | |||
622 | ||||
623 | /// \brief Get a value with high bits set | |||
624 | /// | |||
625 | /// Constructs an APInt value that has the top hiBitsSet bits set. | |||
626 | /// | |||
627 | /// \param numBits the bitwidth of the result | |||
628 | /// \param hiBitsSet the number of high-order bits set in the result. | |||
629 | static APInt getHighBitsSet(unsigned numBits, unsigned hiBitsSet) { | |||
630 | APInt Res(numBits, 0); | |||
631 | Res.setHighBits(hiBitsSet); | |||
632 | return Res; | |||
633 | } | |||
634 | ||||
635 | /// \brief Get a value with low bits set | |||
636 | /// | |||
637 | /// Constructs an APInt value that has the bottom loBitsSet bits set. | |||
638 | /// | |||
639 | /// \param numBits the bitwidth of the result | |||
640 | /// \param loBitsSet the number of low-order bits set in the result. | |||
641 | static APInt getLowBitsSet(unsigned numBits, unsigned loBitsSet) { | |||
642 | APInt Res(numBits, 0); | |||
643 | Res.setLowBits(loBitsSet); | |||
644 | return Res; | |||
645 | } | |||
646 | ||||
647 | /// \brief Return a value containing V broadcasted over NewLen bits. | |||
648 | static APInt getSplat(unsigned NewLen, const APInt &V); | |||
649 | ||||
650 | /// \brief Determine if two APInts have the same value, after zero-extending | |||
651 | /// one of them (if needed!) to ensure that the bit-widths match. | |||
652 | static bool isSameValue(const APInt &I1, const APInt &I2) { | |||
653 | if (I1.getBitWidth() == I2.getBitWidth()) | |||
654 | return I1 == I2; | |||
655 | ||||
656 | if (I1.getBitWidth() > I2.getBitWidth()) | |||
657 | return I1 == I2.zext(I1.getBitWidth()); | |||
658 | ||||
659 | return I1.zext(I2.getBitWidth()) == I2; | |||
660 | } | |||
661 | ||||
662 | /// \brief Overload to compute a hash_code for an APInt value. | |||
663 | friend hash_code hash_value(const APInt &Arg); | |||
664 | ||||
665 | /// This function returns a pointer to the internal storage of the APInt. | |||
666 | /// This is useful for writing out the APInt in binary form without any | |||
667 | /// conversions. | |||
668 | const uint64_t *getRawData() const { | |||
669 | if (isSingleWord()) | |||
670 | return &U.VAL; | |||
671 | return &U.pVal[0]; | |||
672 | } | |||
673 | ||||
674 | /// @} | |||
675 | /// \name Unary Operators | |||
676 | /// @{ | |||
677 | ||||
678 | /// \brief Postfix increment operator. | |||
679 | /// | |||
680 | /// Increments *this by 1. | |||
681 | /// | |||
682 | /// \returns a new APInt value representing the original value of *this. | |||
683 | const APInt operator++(int) { | |||
684 | APInt API(*this); | |||
685 | ++(*this); | |||
686 | return API; | |||
687 | } | |||
688 | ||||
689 | /// \brief Prefix increment operator. | |||
690 | /// | |||
691 | /// \returns *this incremented by one | |||
692 | APInt &operator++(); | |||
693 | ||||
694 | /// \brief Postfix decrement operator. | |||
695 | /// | |||
696 | /// Decrements *this by 1. | |||
697 | /// | |||
698 | /// \returns a new APInt value representing the original value of *this. | |||
699 | const APInt operator--(int) { | |||
700 | APInt API(*this); | |||
701 | --(*this); | |||
702 | return API; | |||
703 | } | |||
704 | ||||
705 | /// \brief Prefix decrement operator. | |||
706 | /// | |||
707 | /// \returns *this decremented by one. | |||
708 | APInt &operator--(); | |||
709 | ||||
710 | /// \brief Logical negation operator. | |||
711 | /// | |||
712 | /// Performs logical negation operation on this APInt. | |||
713 | /// | |||
714 | /// \returns true if *this is zero, false otherwise. | |||
715 | bool operator!() const { | |||
716 | if (isSingleWord()) | |||
717 | return U.VAL == 0; | |||
718 | return countLeadingZerosSlowCase() == BitWidth; | |||
719 | } | |||
720 | ||||
721 | /// @} | |||
722 | /// \name Assignment Operators | |||
723 | /// @{ | |||
724 | ||||
725 | /// \brief Copy assignment operator. | |||
726 | /// | |||
727 | /// \returns *this after assignment of RHS. | |||
728 | APInt &operator=(const APInt &RHS) { | |||
729 | // If the bitwidths are the same, we can avoid mucking with memory | |||
730 | if (isSingleWord() && RHS.isSingleWord()) { | |||
731 | U.VAL = RHS.U.VAL; | |||
732 | BitWidth = RHS.BitWidth; | |||
733 | return clearUnusedBits(); | |||
734 | } | |||
735 | ||||
736 | AssignSlowCase(RHS); | |||
737 | return *this; | |||
738 | } | |||
739 | ||||
740 | /// @brief Move assignment operator. | |||
741 | APInt &operator=(APInt &&that) { | |||
742 | assert(this != &that && "Self-move not supported")(static_cast <bool> (this != &that && "Self-move not supported" ) ? void (0) : __assert_fail ("this != &that && \"Self-move not supported\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 742, __extension__ __PRETTY_FUNCTION__)); | |||
743 | if (!isSingleWord()) | |||
744 | delete[] U.pVal; | |||
745 | ||||
746 | // Use memcpy so that type based alias analysis sees both VAL and pVal | |||
747 | // as modified. | |||
748 | memcpy(&U, &that.U, sizeof(U)); | |||
749 | ||||
750 | BitWidth = that.BitWidth; | |||
751 | that.BitWidth = 0; | |||
752 | ||||
753 | return *this; | |||
754 | } | |||
755 | ||||
756 | /// \brief Assignment operator. | |||
757 | /// | |||
758 | /// The RHS value is assigned to *this. If the significant bits in RHS exceed | |||
759 | /// the bit width, the excess bits are truncated. If the bit width is larger | |||
760 | /// than 64, the value is zero filled in the unspecified high order bits. | |||
761 | /// | |||
762 | /// \returns *this after assignment of RHS value. | |||
763 | APInt &operator=(uint64_t RHS) { | |||
764 | if (isSingleWord()) { | |||
765 | U.VAL = RHS; | |||
766 | clearUnusedBits(); | |||
767 | } else { | |||
768 | U.pVal[0] = RHS; | |||
769 | memset(U.pVal+1, 0, (getNumWords() - 1) * APINT_WORD_SIZE); | |||
770 | } | |||
771 | return *this; | |||
772 | } | |||
773 | ||||
774 | /// \brief Bitwise AND assignment operator. | |||
775 | /// | |||
776 | /// Performs a bitwise AND operation on this APInt and RHS. The result is | |||
777 | /// assigned to *this. | |||
778 | /// | |||
779 | /// \returns *this after ANDing with RHS. | |||
780 | APInt &operator&=(const APInt &RHS) { | |||
781 | assert(BitWidth == RHS.BitWidth && "Bit widths must be the same")(static_cast <bool> (BitWidth == RHS.BitWidth && "Bit widths must be the same") ? void (0) : __assert_fail ("BitWidth == RHS.BitWidth && \"Bit widths must be the same\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 781, __extension__ __PRETTY_FUNCTION__)); | |||
782 | if (isSingleWord()) | |||
783 | U.VAL &= RHS.U.VAL; | |||
784 | else | |||
785 | AndAssignSlowCase(RHS); | |||
786 | return *this; | |||
787 | } | |||
788 | ||||
789 | /// \brief Bitwise AND assignment operator. | |||
790 | /// | |||
791 | /// Performs a bitwise AND operation on this APInt and RHS. RHS is | |||
792 | /// logically zero-extended or truncated to match the bit-width of | |||
793 | /// the LHS. | |||
794 | APInt &operator&=(uint64_t RHS) { | |||
795 | if (isSingleWord()) { | |||
796 | U.VAL &= RHS; | |||
797 | return *this; | |||
798 | } | |||
799 | U.pVal[0] &= RHS; | |||
800 | memset(U.pVal+1, 0, (getNumWords() - 1) * APINT_WORD_SIZE); | |||
801 | return *this; | |||
802 | } | |||
803 | ||||
804 | /// \brief Bitwise OR assignment operator. | |||
805 | /// | |||
806 | /// Performs a bitwise OR operation on this APInt and RHS. The result is | |||
807 | /// assigned *this; | |||
808 | /// | |||
809 | /// \returns *this after ORing with RHS. | |||
810 | APInt &operator|=(const APInt &RHS) { | |||
811 | assert(BitWidth == RHS.BitWidth && "Bit widths must be the same")(static_cast <bool> (BitWidth == RHS.BitWidth && "Bit widths must be the same") ? void (0) : __assert_fail ("BitWidth == RHS.BitWidth && \"Bit widths must be the same\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 811, __extension__ __PRETTY_FUNCTION__)); | |||
812 | if (isSingleWord()) | |||
813 | U.VAL |= RHS.U.VAL; | |||
814 | else | |||
815 | OrAssignSlowCase(RHS); | |||
816 | return *this; | |||
817 | } | |||
818 | ||||
819 | /// \brief Bitwise OR assignment operator. | |||
820 | /// | |||
821 | /// Performs a bitwise OR operation on this APInt and RHS. RHS is | |||
822 | /// logically zero-extended or truncated to match the bit-width of | |||
823 | /// the LHS. | |||
824 | APInt &operator|=(uint64_t RHS) { | |||
825 | if (isSingleWord()) { | |||
826 | U.VAL |= RHS; | |||
827 | clearUnusedBits(); | |||
828 | } else { | |||
829 | U.pVal[0] |= RHS; | |||
830 | } | |||
831 | return *this; | |||
832 | } | |||
833 | ||||
834 | /// \brief Bitwise XOR assignment operator. | |||
835 | /// | |||
836 | /// Performs a bitwise XOR operation on this APInt and RHS. The result is | |||
837 | /// assigned to *this. | |||
838 | /// | |||
839 | /// \returns *this after XORing with RHS. | |||
840 | APInt &operator^=(const APInt &RHS) { | |||
841 | assert(BitWidth == RHS.BitWidth && "Bit widths must be the same")(static_cast <bool> (BitWidth == RHS.BitWidth && "Bit widths must be the same") ? void (0) : __assert_fail ("BitWidth == RHS.BitWidth && \"Bit widths must be the same\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 841, __extension__ __PRETTY_FUNCTION__)); | |||
842 | if (isSingleWord()) | |||
843 | U.VAL ^= RHS.U.VAL; | |||
844 | else | |||
845 | XorAssignSlowCase(RHS); | |||
846 | return *this; | |||
847 | } | |||
848 | ||||
849 | /// \brief Bitwise XOR assignment operator. | |||
850 | /// | |||
851 | /// Performs a bitwise XOR operation on this APInt and RHS. RHS is | |||
852 | /// logically zero-extended or truncated to match the bit-width of | |||
853 | /// the LHS. | |||
854 | APInt &operator^=(uint64_t RHS) { | |||
855 | if (isSingleWord()) { | |||
856 | U.VAL ^= RHS; | |||
857 | clearUnusedBits(); | |||
858 | } else { | |||
859 | U.pVal[0] ^= RHS; | |||
860 | } | |||
861 | return *this; | |||
862 | } | |||
863 | ||||
864 | /// \brief Multiplication assignment operator. | |||
865 | /// | |||
866 | /// Multiplies this APInt by RHS and assigns the result to *this. | |||
867 | /// | |||
868 | /// \returns *this | |||
869 | APInt &operator*=(const APInt &RHS); | |||
870 | APInt &operator*=(uint64_t RHS); | |||
871 | ||||
872 | /// \brief Addition assignment operator. | |||
873 | /// | |||
874 | /// Adds RHS to *this and assigns the result to *this. | |||
875 | /// | |||
876 | /// \returns *this | |||
877 | APInt &operator+=(const APInt &RHS); | |||
878 | APInt &operator+=(uint64_t RHS); | |||
879 | ||||
880 | /// \brief Subtraction assignment operator. | |||
881 | /// | |||
882 | /// Subtracts RHS from *this and assigns the result to *this. | |||
883 | /// | |||
884 | /// \returns *this | |||
885 | APInt &operator-=(const APInt &RHS); | |||
886 | APInt &operator-=(uint64_t RHS); | |||
887 | ||||
888 | /// \brief Left-shift assignment function. | |||
889 | /// | |||
890 | /// Shifts *this left by shiftAmt and assigns the result to *this. | |||
891 | /// | |||
892 | /// \returns *this after shifting left by ShiftAmt | |||
893 | APInt &operator<<=(unsigned ShiftAmt) { | |||
894 | assert(ShiftAmt <= BitWidth && "Invalid shift amount")(static_cast <bool> (ShiftAmt <= BitWidth && "Invalid shift amount") ? void (0) : __assert_fail ("ShiftAmt <= BitWidth && \"Invalid shift amount\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 894, __extension__ __PRETTY_FUNCTION__)); | |||
895 | if (isSingleWord()) { | |||
896 | if (ShiftAmt == BitWidth) | |||
897 | U.VAL = 0; | |||
898 | else | |||
899 | U.VAL <<= ShiftAmt; | |||
900 | return clearUnusedBits(); | |||
901 | } | |||
902 | shlSlowCase(ShiftAmt); | |||
903 | return *this; | |||
904 | } | |||
905 | ||||
906 | /// \brief Left-shift assignment function. | |||
907 | /// | |||
908 | /// Shifts *this left by shiftAmt and assigns the result to *this. | |||
909 | /// | |||
910 | /// \returns *this after shifting left by ShiftAmt | |||
911 | APInt &operator<<=(const APInt &ShiftAmt); | |||
912 | ||||
913 | /// @} | |||
914 | /// \name Binary Operators | |||
915 | /// @{ | |||
916 | ||||
917 | /// \brief Multiplication operator. | |||
918 | /// | |||
919 | /// Multiplies this APInt by RHS and returns the result. | |||
920 | APInt operator*(const APInt &RHS) const; | |||
921 | ||||
922 | /// \brief Left logical shift operator. | |||
923 | /// | |||
924 | /// Shifts this APInt left by \p Bits and returns the result. | |||
925 | APInt operator<<(unsigned Bits) const { return shl(Bits); } | |||
926 | ||||
927 | /// \brief Left logical shift operator. | |||
928 | /// | |||
929 | /// Shifts this APInt left by \p Bits and returns the result. | |||
930 | APInt operator<<(const APInt &Bits) const { return shl(Bits); } | |||
931 | ||||
932 | /// \brief Arithmetic right-shift function. | |||
933 | /// | |||
934 | /// Arithmetic right-shift this APInt by shiftAmt. | |||
935 | APInt ashr(unsigned ShiftAmt) const { | |||
936 | APInt R(*this); | |||
937 | R.ashrInPlace(ShiftAmt); | |||
938 | return R; | |||
939 | } | |||
940 | ||||
941 | /// Arithmetic right-shift this APInt by ShiftAmt in place. | |||
942 | void ashrInPlace(unsigned ShiftAmt) { | |||
943 | assert(ShiftAmt <= BitWidth && "Invalid shift amount")(static_cast <bool> (ShiftAmt <= BitWidth && "Invalid shift amount") ? void (0) : __assert_fail ("ShiftAmt <= BitWidth && \"Invalid shift amount\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 943, __extension__ __PRETTY_FUNCTION__)); | |||
944 | if (isSingleWord()) { | |||
945 | int64_t SExtVAL = SignExtend64(U.VAL, BitWidth); | |||
946 | if (ShiftAmt == BitWidth) | |||
947 | U.VAL = SExtVAL >> (APINT_BITS_PER_WORD - 1); // Fill with sign bit. | |||
948 | else | |||
949 | U.VAL = SExtVAL >> ShiftAmt; | |||
950 | clearUnusedBits(); | |||
951 | return; | |||
952 | } | |||
953 | ashrSlowCase(ShiftAmt); | |||
954 | } | |||
955 | ||||
956 | /// \brief Logical right-shift function. | |||
957 | /// | |||
958 | /// Logical right-shift this APInt by shiftAmt. | |||
959 | APInt lshr(unsigned shiftAmt) const { | |||
960 | APInt R(*this); | |||
961 | R.lshrInPlace(shiftAmt); | |||
962 | return R; | |||
963 | } | |||
964 | ||||
965 | /// Logical right-shift this APInt by ShiftAmt in place. | |||
966 | void lshrInPlace(unsigned ShiftAmt) { | |||
967 | assert(ShiftAmt <= BitWidth && "Invalid shift amount")(static_cast <bool> (ShiftAmt <= BitWidth && "Invalid shift amount") ? void (0) : __assert_fail ("ShiftAmt <= BitWidth && \"Invalid shift amount\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 967, __extension__ __PRETTY_FUNCTION__)); | |||
968 | if (isSingleWord()) { | |||
969 | if (ShiftAmt == BitWidth) | |||
970 | U.VAL = 0; | |||
971 | else | |||
972 | U.VAL >>= ShiftAmt; | |||
973 | return; | |||
974 | } | |||
975 | lshrSlowCase(ShiftAmt); | |||
976 | } | |||
977 | ||||
978 | /// \brief Left-shift function. | |||
979 | /// | |||
980 | /// Left-shift this APInt by shiftAmt. | |||
981 | APInt shl(unsigned shiftAmt) const { | |||
982 | APInt R(*this); | |||
983 | R <<= shiftAmt; | |||
984 | return R; | |||
985 | } | |||
986 | ||||
987 | /// \brief Rotate left by rotateAmt. | |||
988 | APInt rotl(unsigned rotateAmt) const; | |||
989 | ||||
990 | /// \brief Rotate right by rotateAmt. | |||
991 | APInt rotr(unsigned rotateAmt) const; | |||
992 | ||||
993 | /// \brief Arithmetic right-shift function. | |||
994 | /// | |||
995 | /// Arithmetic right-shift this APInt by shiftAmt. | |||
996 | APInt ashr(const APInt &ShiftAmt) const { | |||
997 | APInt R(*this); | |||
998 | R.ashrInPlace(ShiftAmt); | |||
999 | return R; | |||
1000 | } | |||
1001 | ||||
1002 | /// Arithmetic right-shift this APInt by shiftAmt in place. | |||
1003 | void ashrInPlace(const APInt &shiftAmt); | |||
1004 | ||||
1005 | /// \brief Logical right-shift function. | |||
1006 | /// | |||
1007 | /// Logical right-shift this APInt by shiftAmt. | |||
1008 | APInt lshr(const APInt &ShiftAmt) const { | |||
1009 | APInt R(*this); | |||
1010 | R.lshrInPlace(ShiftAmt); | |||
1011 | return R; | |||
1012 | } | |||
1013 | ||||
1014 | /// Logical right-shift this APInt by ShiftAmt in place. | |||
1015 | void lshrInPlace(const APInt &ShiftAmt); | |||
1016 | ||||
1017 | /// \brief Left-shift function. | |||
1018 | /// | |||
1019 | /// Left-shift this APInt by shiftAmt. | |||
1020 | APInt shl(const APInt &ShiftAmt) const { | |||
1021 | APInt R(*this); | |||
1022 | R <<= ShiftAmt; | |||
1023 | return R; | |||
1024 | } | |||
1025 | ||||
1026 | /// \brief Rotate left by rotateAmt. | |||
1027 | APInt rotl(const APInt &rotateAmt) const; | |||
1028 | ||||
1029 | /// \brief Rotate right by rotateAmt. | |||
1030 | APInt rotr(const APInt &rotateAmt) const; | |||
1031 | ||||
1032 | /// \brief Unsigned division operation. | |||
1033 | /// | |||
1034 | /// Perform an unsigned divide operation on this APInt by RHS. Both this and | |||
1035 | /// RHS are treated as unsigned quantities for purposes of this division. | |||
1036 | /// | |||
1037 | /// \returns a new APInt value containing the division result | |||
1038 | APInt udiv(const APInt &RHS) const; | |||
1039 | APInt udiv(uint64_t RHS) const; | |||
1040 | ||||
1041 | /// \brief Signed division function for APInt. | |||
1042 | /// | |||
1043 | /// Signed divide this APInt by APInt RHS. | |||
1044 | APInt sdiv(const APInt &RHS) const; | |||
1045 | APInt sdiv(int64_t RHS) const; | |||
1046 | ||||
1047 | /// \brief Unsigned remainder operation. | |||
1048 | /// | |||
1049 | /// Perform an unsigned remainder operation on this APInt with RHS being the | |||
1050 | /// divisor. Both this and RHS are treated as unsigned quantities for purposes | |||
1051 | /// of this operation. Note that this is a true remainder operation and not a | |||
1052 | /// modulo operation because the sign follows the sign of the dividend which | |||
1053 | /// is *this. | |||
1054 | /// | |||
1055 | /// \returns a new APInt value containing the remainder result | |||
1056 | APInt urem(const APInt &RHS) const; | |||
1057 | uint64_t urem(uint64_t RHS) const; | |||
1058 | ||||
1059 | /// \brief Function for signed remainder operation. | |||
1060 | /// | |||
1061 | /// Signed remainder operation on APInt. | |||
1062 | APInt srem(const APInt &RHS) const; | |||
1063 | int64_t srem(int64_t RHS) const; | |||
1064 | ||||
1065 | /// \brief Dual division/remainder interface. | |||
1066 | /// | |||
1067 | /// Sometimes it is convenient to divide two APInt values and obtain both the | |||
1068 | /// quotient and remainder. This function does both operations in the same | |||
1069 | /// computation making it a little more efficient. The pair of input arguments | |||
1070 | /// may overlap with the pair of output arguments. It is safe to call | |||
1071 | /// udivrem(X, Y, X, Y), for example. | |||
1072 | static void udivrem(const APInt &LHS, const APInt &RHS, APInt &Quotient, | |||
1073 | APInt &Remainder); | |||
1074 | static void udivrem(const APInt &LHS, uint64_t RHS, APInt &Quotient, | |||
1075 | uint64_t &Remainder); | |||
1076 | ||||
1077 | static void sdivrem(const APInt &LHS, const APInt &RHS, APInt &Quotient, | |||
1078 | APInt &Remainder); | |||
1079 | static void sdivrem(const APInt &LHS, int64_t RHS, APInt &Quotient, | |||
1080 | int64_t &Remainder); | |||
1081 | ||||
1082 | // Operations that return overflow indicators. | |||
1083 | APInt sadd_ov(const APInt &RHS, bool &Overflow) const; | |||
1084 | APInt uadd_ov(const APInt &RHS, bool &Overflow) const; | |||
1085 | APInt ssub_ov(const APInt &RHS, bool &Overflow) const; | |||
1086 | APInt usub_ov(const APInt &RHS, bool &Overflow) const; | |||
1087 | APInt sdiv_ov(const APInt &RHS, bool &Overflow) const; | |||
1088 | APInt smul_ov(const APInt &RHS, bool &Overflow) const; | |||
1089 | APInt umul_ov(const APInt &RHS, bool &Overflow) const; | |||
1090 | APInt sshl_ov(const APInt &Amt, bool &Overflow) const; | |||
1091 | APInt ushl_ov(const APInt &Amt, bool &Overflow) const; | |||
1092 | ||||
1093 | /// \brief Array-indexing support. | |||
1094 | /// | |||
1095 | /// \returns the bit value at bitPosition | |||
1096 | bool operator[](unsigned bitPosition) const { | |||
1097 | assert(bitPosition < getBitWidth() && "Bit position out of bounds!")(static_cast <bool> (bitPosition < getBitWidth() && "Bit position out of bounds!") ? void (0) : __assert_fail ("bitPosition < getBitWidth() && \"Bit position out of bounds!\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 1097, __extension__ __PRETTY_FUNCTION__)); | |||
1098 | return (maskBit(bitPosition) & getWord(bitPosition)) != 0; | |||
1099 | } | |||
1100 | ||||
1101 | /// @} | |||
1102 | /// \name Comparison Operators | |||
1103 | /// @{ | |||
1104 | ||||
1105 | /// \brief Equality operator. | |||
1106 | /// | |||
1107 | /// Compares this APInt with RHS for the validity of the equality | |||
1108 | /// relationship. | |||
1109 | bool operator==(const APInt &RHS) const { | |||
1110 | assert(BitWidth == RHS.BitWidth && "Comparison requires equal bit widths")(static_cast <bool> (BitWidth == RHS.BitWidth && "Comparison requires equal bit widths") ? void (0) : __assert_fail ("BitWidth == RHS.BitWidth && \"Comparison requires equal bit widths\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 1110, __extension__ __PRETTY_FUNCTION__)); | |||
1111 | if (isSingleWord()) | |||
1112 | return U.VAL == RHS.U.VAL; | |||
1113 | return EqualSlowCase(RHS); | |||
1114 | } | |||
1115 | ||||
1116 | /// \brief Equality operator. | |||
1117 | /// | |||
1118 | /// Compares this APInt with a uint64_t for the validity of the equality | |||
1119 | /// relationship. | |||
1120 | /// | |||
1121 | /// \returns true if *this == Val | |||
1122 | bool operator==(uint64_t Val) const { | |||
1123 | return (isSingleWord() || getActiveBits() <= 64) && getZExtValue() == Val; | |||
1124 | } | |||
1125 | ||||
1126 | /// \brief Equality comparison. | |||
1127 | /// | |||
1128 | /// Compares this APInt with RHS for the validity of the equality | |||
1129 | /// relationship. | |||
1130 | /// | |||
1131 | /// \returns true if *this == Val | |||
1132 | bool eq(const APInt &RHS) const { return (*this) == RHS; } | |||
1133 | ||||
1134 | /// \brief Inequality operator. | |||
1135 | /// | |||
1136 | /// Compares this APInt with RHS for the validity of the inequality | |||
1137 | /// relationship. | |||
1138 | /// | |||
1139 | /// \returns true if *this != Val | |||
1140 | bool operator!=(const APInt &RHS) const { return !((*this) == RHS); } | |||
1141 | ||||
1142 | /// \brief Inequality operator. | |||
1143 | /// | |||
1144 | /// Compares this APInt with a uint64_t for the validity of the inequality | |||
1145 | /// relationship. | |||
1146 | /// | |||
1147 | /// \returns true if *this != Val | |||
1148 | bool operator!=(uint64_t Val) const { return !((*this) == Val); } | |||
1149 | ||||
1150 | /// \brief Inequality comparison | |||
1151 | /// | |||
1152 | /// Compares this APInt with RHS for the validity of the inequality | |||
1153 | /// relationship. | |||
1154 | /// | |||
1155 | /// \returns true if *this != Val | |||
1156 | bool ne(const APInt &RHS) const { return !((*this) == RHS); } | |||
1157 | ||||
1158 | /// \brief Unsigned less than comparison | |||
1159 | /// | |||
1160 | /// Regards both *this and RHS as unsigned quantities and compares them for | |||
1161 | /// the validity of the less-than relationship. | |||
1162 | /// | |||
1163 | /// \returns true if *this < RHS when both are considered unsigned. | |||
1164 | bool ult(const APInt &RHS) const { return compare(RHS) < 0; } | |||
1165 | ||||
1166 | /// \brief Unsigned less than comparison | |||
1167 | /// | |||
1168 | /// Regards both *this as an unsigned quantity and compares it with RHS for | |||
1169 | /// the validity of the less-than relationship. | |||
1170 | /// | |||
1171 | /// \returns true if *this < RHS when considered unsigned. | |||
1172 | bool ult(uint64_t RHS) const { | |||
1173 | // Only need to check active bits if not a single word. | |||
1174 | return (isSingleWord() || getActiveBits() <= 64) && getZExtValue() < RHS; | |||
1175 | } | |||
1176 | ||||
1177 | /// \brief Signed less than comparison | |||
1178 | /// | |||
1179 | /// Regards both *this and RHS as signed quantities and compares them for | |||
1180 | /// validity of the less-than relationship. | |||
1181 | /// | |||
1182 | /// \returns true if *this < RHS when both are considered signed. | |||
1183 | bool slt(const APInt &RHS) const { return compareSigned(RHS) < 0; } | |||
1184 | ||||
1185 | /// \brief Signed less than comparison | |||
1186 | /// | |||
1187 | /// Regards both *this as a signed quantity and compares it with RHS for | |||
1188 | /// the validity of the less-than relationship. | |||
1189 | /// | |||
1190 | /// \returns true if *this < RHS when considered signed. | |||
1191 | bool slt(int64_t RHS) const { | |||
1192 | return (!isSingleWord() && getMinSignedBits() > 64) ? isNegative() | |||
1193 | : getSExtValue() < RHS; | |||
1194 | } | |||
1195 | ||||
1196 | /// \brief Unsigned less or equal comparison | |||
1197 | /// | |||
1198 | /// Regards both *this and RHS as unsigned quantities and compares them for | |||
1199 | /// validity of the less-or-equal relationship. | |||
1200 | /// | |||
1201 | /// \returns true if *this <= RHS when both are considered unsigned. | |||
1202 | bool ule(const APInt &RHS) const { return compare(RHS) <= 0; } | |||
1203 | ||||
1204 | /// \brief Unsigned less or equal comparison | |||
1205 | /// | |||
1206 | /// Regards both *this as an unsigned quantity and compares it with RHS for | |||
1207 | /// the validity of the less-or-equal relationship. | |||
1208 | /// | |||
1209 | /// \returns true if *this <= RHS when considered unsigned. | |||
1210 | bool ule(uint64_t RHS) const { return !ugt(RHS); } | |||
1211 | ||||
1212 | /// \brief Signed less or equal comparison | |||
1213 | /// | |||
1214 | /// Regards both *this and RHS as signed quantities and compares them for | |||
1215 | /// validity of the less-or-equal relationship. | |||
1216 | /// | |||
1217 | /// \returns true if *this <= RHS when both are considered signed. | |||
1218 | bool sle(const APInt &RHS) const { return compareSigned(RHS) <= 0; } | |||
1219 | ||||
1220 | /// \brief Signed less or equal comparison | |||
1221 | /// | |||
1222 | /// Regards both *this as a signed quantity and compares it with RHS for the | |||
1223 | /// validity of the less-or-equal relationship. | |||
1224 | /// | |||
1225 | /// \returns true if *this <= RHS when considered signed. | |||
1226 | bool sle(uint64_t RHS) const { return !sgt(RHS); } | |||
1227 | ||||
1228 | /// \brief Unsigned greather than comparison | |||
1229 | /// | |||
1230 | /// Regards both *this and RHS as unsigned quantities and compares them for | |||
1231 | /// the validity of the greater-than relationship. | |||
1232 | /// | |||
1233 | /// \returns true if *this > RHS when both are considered unsigned. | |||
1234 | bool ugt(const APInt &RHS) const { return !ule(RHS); } | |||
1235 | ||||
1236 | /// \brief Unsigned greater than comparison | |||
1237 | /// | |||
1238 | /// Regards both *this as an unsigned quantity and compares it with RHS for | |||
1239 | /// the validity of the greater-than relationship. | |||
1240 | /// | |||
1241 | /// \returns true if *this > RHS when considered unsigned. | |||
1242 | bool ugt(uint64_t RHS) const { | |||
1243 | // Only need to check active bits if not a single word. | |||
1244 | return (!isSingleWord() && getActiveBits() > 64) || getZExtValue() > RHS; | |||
1245 | } | |||
1246 | ||||
1247 | /// \brief Signed greather than comparison | |||
1248 | /// | |||
1249 | /// Regards both *this and RHS as signed quantities and compares them for the | |||
1250 | /// validity of the greater-than relationship. | |||
1251 | /// | |||
1252 | /// \returns true if *this > RHS when both are considered signed. | |||
1253 | bool sgt(const APInt &RHS) const { return !sle(RHS); } | |||
1254 | ||||
1255 | /// \brief Signed greater than comparison | |||
1256 | /// | |||
1257 | /// Regards both *this as a signed quantity and compares it with RHS for | |||
1258 | /// the validity of the greater-than relationship. | |||
1259 | /// | |||
1260 | /// \returns true if *this > RHS when considered signed. | |||
1261 | bool sgt(int64_t RHS) const { | |||
1262 | return (!isSingleWord() && getMinSignedBits() > 64) ? !isNegative() | |||
1263 | : getSExtValue() > RHS; | |||
1264 | } | |||
1265 | ||||
1266 | /// \brief Unsigned greater or equal comparison | |||
1267 | /// | |||
1268 | /// Regards both *this and RHS as unsigned quantities and compares them for | |||
1269 | /// validity of the greater-or-equal relationship. | |||
1270 | /// | |||
1271 | /// \returns true if *this >= RHS when both are considered unsigned. | |||
1272 | bool uge(const APInt &RHS) const { return !ult(RHS); } | |||
1273 | ||||
1274 | /// \brief Unsigned greater or equal comparison | |||
1275 | /// | |||
1276 | /// Regards both *this as an unsigned quantity and compares it with RHS for | |||
1277 | /// the validity of the greater-or-equal relationship. | |||
1278 | /// | |||
1279 | /// \returns true if *this >= RHS when considered unsigned. | |||
1280 | bool uge(uint64_t RHS) const { return !ult(RHS); } | |||
1281 | ||||
1282 | /// \brief Signed greater or equal comparison | |||
1283 | /// | |||
1284 | /// Regards both *this and RHS as signed quantities and compares them for | |||
1285 | /// validity of the greater-or-equal relationship. | |||
1286 | /// | |||
1287 | /// \returns true if *this >= RHS when both are considered signed. | |||
1288 | bool sge(const APInt &RHS) const { return !slt(RHS); } | |||
1289 | ||||
1290 | /// \brief Signed greater or equal comparison | |||
1291 | /// | |||
1292 | /// Regards both *this as a signed quantity and compares it with RHS for | |||
1293 | /// the validity of the greater-or-equal relationship. | |||
1294 | /// | |||
1295 | /// \returns true if *this >= RHS when considered signed. | |||
1296 | bool sge(int64_t RHS) const { return !slt(RHS); } | |||
1297 | ||||
1298 | /// This operation tests if there are any pairs of corresponding bits | |||
1299 | /// between this APInt and RHS that are both set. | |||
1300 | bool intersects(const APInt &RHS) const { | |||
1301 | assert(BitWidth == RHS.BitWidth && "Bit widths must be the same")(static_cast <bool> (BitWidth == RHS.BitWidth && "Bit widths must be the same") ? void (0) : __assert_fail ("BitWidth == RHS.BitWidth && \"Bit widths must be the same\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 1301, __extension__ __PRETTY_FUNCTION__)); | |||
1302 | if (isSingleWord()) | |||
1303 | return (U.VAL & RHS.U.VAL) != 0; | |||
1304 | return intersectsSlowCase(RHS); | |||
1305 | } | |||
1306 | ||||
1307 | /// This operation checks that all bits set in this APInt are also set in RHS. | |||
1308 | bool isSubsetOf(const APInt &RHS) const { | |||
1309 | assert(BitWidth == RHS.BitWidth && "Bit widths must be the same")(static_cast <bool> (BitWidth == RHS.BitWidth && "Bit widths must be the same") ? void (0) : __assert_fail ("BitWidth == RHS.BitWidth && \"Bit widths must be the same\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 1309, __extension__ __PRETTY_FUNCTION__)); | |||
1310 | if (isSingleWord()) | |||
1311 | return (U.VAL & ~RHS.U.VAL) == 0; | |||
1312 | return isSubsetOfSlowCase(RHS); | |||
1313 | } | |||
1314 | ||||
1315 | /// @} | |||
1316 | /// \name Resizing Operators | |||
1317 | /// @{ | |||
1318 | ||||
1319 | /// \brief Truncate to new width. | |||
1320 | /// | |||
1321 | /// Truncate the APInt to a specified width. It is an error to specify a width | |||
1322 | /// that is greater than or equal to the current width. | |||
1323 | APInt trunc(unsigned width) const; | |||
1324 | ||||
1325 | /// \brief Sign extend to a new width. | |||
1326 | /// | |||
1327 | /// This operation sign extends the APInt to a new width. If the high order | |||
1328 | /// bit is set, the fill on the left will be done with 1 bits, otherwise zero. | |||
1329 | /// It is an error to specify a width that is less than or equal to the | |||
1330 | /// current width. | |||
1331 | APInt sext(unsigned width) const; | |||
1332 | ||||
1333 | /// \brief Zero extend to a new width. | |||
1334 | /// | |||
1335 | /// This operation zero extends the APInt to a new width. The high order bits | |||
1336 | /// are filled with 0 bits. It is an error to specify a width that is less | |||
1337 | /// than or equal to the current width. | |||
1338 | APInt zext(unsigned width) const; | |||
1339 | ||||
1340 | /// \brief Sign extend or truncate to width | |||
1341 | /// | |||
1342 | /// Make this APInt have the bit width given by \p width. The value is sign | |||
1343 | /// extended, truncated, or left alone to make it that width. | |||
1344 | APInt sextOrTrunc(unsigned width) const; | |||
1345 | ||||
1346 | /// \brief Zero extend or truncate to width | |||
1347 | /// | |||
1348 | /// Make this APInt have the bit width given by \p width. The value is zero | |||
1349 | /// extended, truncated, or left alone to make it that width. | |||
1350 | APInt zextOrTrunc(unsigned width) const; | |||
1351 | ||||
1352 | /// \brief Sign extend or truncate to width | |||
1353 | /// | |||
1354 | /// Make this APInt have the bit width given by \p width. The value is sign | |||
1355 | /// extended, or left alone to make it that width. | |||
1356 | APInt sextOrSelf(unsigned width) const; | |||
1357 | ||||
1358 | /// \brief Zero extend or truncate to width | |||
1359 | /// | |||
1360 | /// Make this APInt have the bit width given by \p width. The value is zero | |||
1361 | /// extended, or left alone to make it that width. | |||
1362 | APInt zextOrSelf(unsigned width) const; | |||
1363 | ||||
1364 | /// @} | |||
1365 | /// \name Bit Manipulation Operators | |||
1366 | /// @{ | |||
1367 | ||||
1368 | /// \brief Set every bit to 1. | |||
1369 | void setAllBits() { | |||
1370 | if (isSingleWord()) | |||
1371 | U.VAL = WORD_MAX; | |||
1372 | else | |||
1373 | // Set all the bits in all the words. | |||
1374 | memset(U.pVal, -1, getNumWords() * APINT_WORD_SIZE); | |||
1375 | // Clear the unused ones | |||
1376 | clearUnusedBits(); | |||
1377 | } | |||
1378 | ||||
1379 | /// \brief Set a given bit to 1. | |||
1380 | /// | |||
1381 | /// Set the given bit to 1 whose position is given as "bitPosition". | |||
1382 | void setBit(unsigned BitPosition) { | |||
1383 | assert(BitPosition <= BitWidth && "BitPosition out of range")(static_cast <bool> (BitPosition <= BitWidth && "BitPosition out of range") ? void (0) : __assert_fail ("BitPosition <= BitWidth && \"BitPosition out of range\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 1383, __extension__ __PRETTY_FUNCTION__)); | |||
1384 | WordType Mask = maskBit(BitPosition); | |||
1385 | if (isSingleWord()) | |||
1386 | U.VAL |= Mask; | |||
1387 | else | |||
1388 | U.pVal[whichWord(BitPosition)] |= Mask; | |||
1389 | } | |||
1390 | ||||
1391 | /// Set the sign bit to 1. | |||
1392 | void setSignBit() { | |||
1393 | setBit(BitWidth - 1); | |||
1394 | } | |||
1395 | ||||
1396 | /// Set the bits from loBit (inclusive) to hiBit (exclusive) to 1. | |||
1397 | void setBits(unsigned loBit, unsigned hiBit) { | |||
1398 | assert(hiBit <= BitWidth && "hiBit out of range")(static_cast <bool> (hiBit <= BitWidth && "hiBit out of range" ) ? void (0) : __assert_fail ("hiBit <= BitWidth && \"hiBit out of range\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 1398, __extension__ __PRETTY_FUNCTION__)); | |||
1399 | assert(loBit <= BitWidth && "loBit out of range")(static_cast <bool> (loBit <= BitWidth && "loBit out of range" ) ? void (0) : __assert_fail ("loBit <= BitWidth && \"loBit out of range\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 1399, __extension__ __PRETTY_FUNCTION__)); | |||
1400 | assert(loBit <= hiBit && "loBit greater than hiBit")(static_cast <bool> (loBit <= hiBit && "loBit greater than hiBit" ) ? void (0) : __assert_fail ("loBit <= hiBit && \"loBit greater than hiBit\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 1400, __extension__ __PRETTY_FUNCTION__)); | |||
1401 | if (loBit == hiBit) | |||
1402 | return; | |||
1403 | if (loBit < APINT_BITS_PER_WORD && hiBit <= APINT_BITS_PER_WORD) { | |||
1404 | uint64_t mask = WORD_MAX >> (APINT_BITS_PER_WORD - (hiBit - loBit)); | |||
1405 | mask <<= loBit; | |||
1406 | if (isSingleWord()) | |||
1407 | U.VAL |= mask; | |||
1408 | else | |||
1409 | U.pVal[0] |= mask; | |||
1410 | } else { | |||
1411 | setBitsSlowCase(loBit, hiBit); | |||
1412 | } | |||
1413 | } | |||
1414 | ||||
1415 | /// Set the top bits starting from loBit. | |||
1416 | void setBitsFrom(unsigned loBit) { | |||
1417 | return setBits(loBit, BitWidth); | |||
1418 | } | |||
1419 | ||||
1420 | /// Set the bottom loBits bits. | |||
1421 | void setLowBits(unsigned loBits) { | |||
1422 | return setBits(0, loBits); | |||
1423 | } | |||
1424 | ||||
1425 | /// Set the top hiBits bits. | |||
1426 | void setHighBits(unsigned hiBits) { | |||
1427 | return setBits(BitWidth - hiBits, BitWidth); | |||
1428 | } | |||
1429 | ||||
1430 | /// \brief Set every bit to 0. | |||
1431 | void clearAllBits() { | |||
1432 | if (isSingleWord()) | |||
1433 | U.VAL = 0; | |||
1434 | else | |||
1435 | memset(U.pVal, 0, getNumWords() * APINT_WORD_SIZE); | |||
1436 | } | |||
1437 | ||||
1438 | /// \brief Set a given bit to 0. | |||
1439 | /// | |||
1440 | /// Set the given bit to 0 whose position is given as "bitPosition". | |||
1441 | void clearBit(unsigned BitPosition) { | |||
1442 | assert(BitPosition <= BitWidth && "BitPosition out of range")(static_cast <bool> (BitPosition <= BitWidth && "BitPosition out of range") ? void (0) : __assert_fail ("BitPosition <= BitWidth && \"BitPosition out of range\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 1442, __extension__ __PRETTY_FUNCTION__)); | |||
1443 | WordType Mask = ~maskBit(BitPosition); | |||
1444 | if (isSingleWord()) | |||
1445 | U.VAL &= Mask; | |||
1446 | else | |||
1447 | U.pVal[whichWord(BitPosition)] &= Mask; | |||
1448 | } | |||
1449 | ||||
1450 | /// Set the sign bit to 0. | |||
1451 | void clearSignBit() { | |||
1452 | clearBit(BitWidth - 1); | |||
1453 | } | |||
1454 | ||||
1455 | /// \brief Toggle every bit to its opposite value. | |||
1456 | void flipAllBits() { | |||
1457 | if (isSingleWord()) { | |||
1458 | U.VAL ^= WORD_MAX; | |||
1459 | clearUnusedBits(); | |||
1460 | } else { | |||
1461 | flipAllBitsSlowCase(); | |||
1462 | } | |||
1463 | } | |||
1464 | ||||
1465 | /// \brief Toggles a given bit to its opposite value. | |||
1466 | /// | |||
1467 | /// Toggle a given bit to its opposite value whose position is given | |||
1468 | /// as "bitPosition". | |||
1469 | void flipBit(unsigned bitPosition); | |||
1470 | ||||
1471 | /// Negate this APInt in place. | |||
1472 | void negate() { | |||
1473 | flipAllBits(); | |||
1474 | ++(*this); | |||
1475 | } | |||
1476 | ||||
1477 | /// Insert the bits from a smaller APInt starting at bitPosition. | |||
1478 | void insertBits(const APInt &SubBits, unsigned bitPosition); | |||
1479 | ||||
1480 | /// Return an APInt with the extracted bits [bitPosition,bitPosition+numBits). | |||
1481 | APInt extractBits(unsigned numBits, unsigned bitPosition) const; | |||
1482 | ||||
1483 | /// @} | |||
1484 | /// \name Value Characterization Functions | |||
1485 | /// @{ | |||
1486 | ||||
1487 | /// \brief Return the number of bits in the APInt. | |||
1488 | unsigned getBitWidth() const { return BitWidth; } | |||
1489 | ||||
1490 | /// \brief Get the number of words. | |||
1491 | /// | |||
1492 | /// Here one word's bitwidth equals to that of uint64_t. | |||
1493 | /// | |||
1494 | /// \returns the number of words to hold the integer value of this APInt. | |||
1495 | unsigned getNumWords() const { return getNumWords(BitWidth); } | |||
1496 | ||||
1497 | /// \brief Get the number of words. | |||
1498 | /// | |||
1499 | /// *NOTE* Here one word's bitwidth equals to that of uint64_t. | |||
1500 | /// | |||
1501 | /// \returns the number of words to hold the integer value with a given bit | |||
1502 | /// width. | |||
1503 | static unsigned getNumWords(unsigned BitWidth) { | |||
1504 | return ((uint64_t)BitWidth + APINT_BITS_PER_WORD - 1) / APINT_BITS_PER_WORD; | |||
1505 | } | |||
1506 | ||||
1507 | /// \brief Compute the number of active bits in the value | |||
1508 | /// | |||
1509 | /// This function returns the number of active bits which is defined as the | |||
1510 | /// bit width minus the number of leading zeros. This is used in several | |||
1511 | /// computations to see how "wide" the value is. | |||
1512 | unsigned getActiveBits() const { return BitWidth - countLeadingZeros(); } | |||
1513 | ||||
1514 | /// \brief Compute the number of active words in the value of this APInt. | |||
1515 | /// | |||
1516 | /// This is used in conjunction with getActiveData to extract the raw value of | |||
1517 | /// the APInt. | |||
1518 | unsigned getActiveWords() const { | |||
1519 | unsigned numActiveBits = getActiveBits(); | |||
1520 | return numActiveBits ? whichWord(numActiveBits - 1) + 1 : 1; | |||
1521 | } | |||
1522 | ||||
1523 | /// \brief Get the minimum bit size for this signed APInt | |||
1524 | /// | |||
1525 | /// Computes the minimum bit width for this APInt while considering it to be a | |||
1526 | /// signed (and probably negative) value. If the value is not negative, this | |||
1527 | /// function returns the same value as getActiveBits()+1. Otherwise, it | |||
1528 | /// returns the smallest bit width that will retain the negative value. For | |||
1529 | /// example, -1 can be written as 0b1 or 0xFFFFFFFFFF. 0b1 is shorter and so | |||
1530 | /// for -1, this function will always return 1. | |||
1531 | unsigned getMinSignedBits() const { | |||
1532 | if (isNegative()) | |||
1533 | return BitWidth - countLeadingOnes() + 1; | |||
1534 | return getActiveBits() + 1; | |||
1535 | } | |||
1536 | ||||
1537 | /// \brief Get zero extended value | |||
1538 | /// | |||
1539 | /// This method attempts to return the value of this APInt as a zero extended | |||
1540 | /// uint64_t. The bitwidth must be <= 64 or the value must fit within a | |||
1541 | /// uint64_t. Otherwise an assertion will result. | |||
1542 | uint64_t getZExtValue() const { | |||
1543 | if (isSingleWord()) | |||
1544 | return U.VAL; | |||
1545 | assert(getActiveBits() <= 64 && "Too many bits for uint64_t")(static_cast <bool> (getActiveBits() <= 64 && "Too many bits for uint64_t") ? void (0) : __assert_fail ("getActiveBits() <= 64 && \"Too many bits for uint64_t\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 1545, __extension__ __PRETTY_FUNCTION__)); | |||
1546 | return U.pVal[0]; | |||
1547 | } | |||
1548 | ||||
1549 | /// \brief Get sign extended value | |||
1550 | /// | |||
1551 | /// This method attempts to return the value of this APInt as a sign extended | |||
1552 | /// int64_t. The bit width must be <= 64 or the value must fit within an | |||
1553 | /// int64_t. Otherwise an assertion will result. | |||
1554 | int64_t getSExtValue() const { | |||
1555 | if (isSingleWord()) | |||
1556 | return SignExtend64(U.VAL, BitWidth); | |||
1557 | assert(getMinSignedBits() <= 64 && "Too many bits for int64_t")(static_cast <bool> (getMinSignedBits() <= 64 && "Too many bits for int64_t") ? void (0) : __assert_fail ("getMinSignedBits() <= 64 && \"Too many bits for int64_t\"" , "/build/llvm-toolchain-snapshot-7~svn326246/include/llvm/ADT/APInt.h" , 1557, __extension__ __PRETTY_FUNCTION__)); | |||
1558 | return int64_t(U.pVal[0]); | |||
1559 | } | |||
1560 | ||||
1561 | /// \brief Get bits required for string value. | |||
1562 | /// | |||
1563 | /// This method determines how many bits are required to hold the APInt | |||
1564 | /// equivalent of the string given by \p str. | |||
1565 | static unsigned getBitsNeeded(StringRef str, uint8_t radix); | |||
1566 | ||||
1567 | /// \brief The APInt version of the countLeadingZeros functions in | |||
1568 | /// MathExtras.h. | |||
1569 | /// | |||
1570 | /// It counts the number of zeros from the most significant bit to the first | |||
1571 | /// one bit. | |||
1572 | /// | |||
1573 | /// \returns BitWidth if the value is zero, otherwise returns the number of | |||
1574 | /// zeros from the most significant bit to the first one bits. | |||
1575 | unsigned countLeadingZeros() const { | |||
1576 | if (isSingleWord()) { | |||
1577 | unsigned unusedBits = APINT_BITS_PER_WORD - BitWidth; | |||
1578 | return llvm::countLeadingZeros(U.VAL) - unusedBits; | |||
1579 | } | |||
1580 | return countLeadingZerosSlowCase(); | |||
1581 | } | |||
1582 | ||||
1583 | /// \brief Count the number of leading one bits. | |||
1584 | /// | |||
1585 | /// This function is an APInt version of the countLeadingOnes | |||
1586 | /// functions in MathExtras.h. It counts the number of ones from the most | |||
1587 | /// significant bit to the first zero bit. | |||
1588 | /// | |||
1589 | /// \returns 0 if the high order bit is not set, otherwise returns the number | |||
1590 | /// of 1 bits from the most significant to the least | |||
1591 | unsigned countLeadingOnes() const { | |||
1592 | if (isSingleWord()) | |||
1593 | return llvm::countLeadingOnes(U.VAL << (APINT_BITS_PER_WORD - BitWidth)); | |||
1594 | return countLeadingOnesSlowCase(); | |||
1595 | } | |||
1596 | ||||
1597 | /// Computes the number of leading bits of this APInt that are equal to its | |||
1598 | /// sign bit. | |||
1599 | unsigned getNumSignBits() const { | |||
1600 | return isNegative() ? countLeadingOnes() : countLeadingZeros(); | |||
1601 | } | |||
1602 | ||||
1603 | /// \brief Count the number of trailing zero bits. | |||
1604 | /// | |||
1605 | /// This function is an APInt version of the countTrailingZeros | |||
1606 | /// functions in MathExtras.h. It counts the number of zeros from the least | |||
1607 | /// significant bit to the first set bit. | |||
1608 | /// | |||
1609 | /// \returns BitWidth if the value is zero, otherwise returns the number of | |||
1610 | /// zeros from the least significant bit to the first one bit. | |||
1611 | unsigned countTrailingZeros() const { | |||
1612 | if (isSingleWord()) | |||
1613 | return std::min(unsigned(llvm::countTrailingZeros(U.VAL)), BitWidth); | |||
1614 | return countTrailingZerosSlowCase(); | |||
1615 | } | |||
1616 | ||||
1617 | /// \brief Count the number of trailing one bits. | |||
1618 | /// | |||
1619 | /// This function is an APInt version of the countTrailingOnes | |||
1620 | /// functions in MathExtras.h. It counts the number of ones from the least | |||
1621 | /// significant bit to the first zero bit. | |||
1622 | /// | |||
1623 | /// \returns BitWidth if the value is all ones, otherwise returns the number | |||
1624 | /// of ones from the least significant bit to the first zero bit. | |||
1625 | unsigned countTrailingOnes() const { | |||
1626 | if (isSingleWord()) | |||
1627 | return llvm::countTrailingOnes(U.VAL); | |||
1628 | return countTrailingOnesSlowCase(); | |||
1629 | } | |||
1630 | ||||
1631 | /// \brief Count the number of bits set. | |||
1632 | /// | |||
1633 | /// This function is an APInt version of the countPopulation functions | |||
1634 | /// in MathExtras.h. It counts the number of 1 bits in the APInt value. | |||
1635 | /// | |||
1636 | /// \returns 0 if the value is zero, otherwise returns the number of set bits. | |||
1637 | unsigned countPopulation() const { | |||
1638 | if (isSingleWord()) | |||
1639 | return llvm::countPopulation(U.VAL); | |||
1640 | return countPopulationSlowCase(); | |||
1641 | } | |||
1642 | ||||
1643 | /// @} | |||
1644 | /// \name Conversion Functions | |||
1645 | /// @{ | |||
1646 | void print(raw_ostream &OS, bool isSigned) const; | |||
1647 | ||||
1648 | /// Converts an APInt to a string and append it to Str. Str is commonly a | |||
1649 | /// SmallString. | |||
1650 | void toString(SmallVectorImpl<char> &Str, unsigned Radix, bool Signed, | |||
1651 | bool formatAsCLiteral = false) const; | |||
1652 | ||||
1653 | /// Considers the APInt to be unsigned and converts it into a string in the | |||
1654 | /// radix given. The radix can be 2, 8, 10 16, or 36. | |||
1655 | void toStringUnsigned(SmallVectorImpl<char> &Str, unsigned Radix = 10) const { | |||
1656 | toString(Str, Radix, false, false); | |||
1657 | } | |||
1658 | ||||
1659 | /// Considers the APInt to be signed and converts it into a string in the | |||
1660 | /// radix given. The radix can be 2, 8, 10, 16, or 36. | |||
1661 | void toStringSigned(SmallVectorImpl<char> &Str, unsigned Radix = 10) const { | |||
1662 | toString(Str, Radix, true, false); | |||
1663 | } | |||
1664 | ||||
1665 | /// \brief Return the APInt as a std::string. | |||
1666 | /// | |||
1667 | /// Note that this is an inefficient method. It is better to pass in a | |||
1668 | /// SmallVector/SmallString to the methods above to avoid thrashing the heap | |||
1669 | /// for the string. | |||
1670 | std::string toString(unsigned Radix, bool Signed) const; | |||
1671 | ||||
1672 | /// \returns a byte-swapped representation of this APInt Value. | |||
1673 | APInt byteSwap() const; | |||
1674 | ||||
1675 | /// \returns the value with the bit representation reversed of this APInt | |||
1676 | /// Value. | |||
1677 | APInt reverseBits() const; | |||
1678 | ||||
1679 | /// \brief Converts this APInt to a double value. | |||
1680 | double roundToDouble(bool isSigned) const; | |||
1681 | ||||
1682 | /// \brief Converts this unsigned APInt to a double value. | |||
1683 | double roundToDouble() const { return roundToDouble(false); } | |||
1684 | ||||
1685 | /// \brief Converts this signed APInt to a double value. | |||
1686 | double signedRoundToDouble() const { return roundToDouble(true); } | |||
1687 | ||||
1688 | /// \brief Converts APInt bits to a double | |||
1689 | /// | |||
1690 | /// The conversion does not do a translation from integer to double, it just | |||
1691 | /// re-interprets the bits as a double. Note that it is valid to do this on | |||
1692 | /// any bit width. Exactly 64 bits will be translated. | |||
1693 | double bitsToDouble() const { | |||
1694 | return BitsToDouble(getWord(0)); | |||
1695 | } | |||
1696 | ||||
1697 | /// \brief Converts APInt bits to a double | |||
1698 | /// | |||
1699 | /// The conversion does not do a translation from integer to float, it just | |||
1700 | /// re-interprets the bits as a float. Note that it is valid to do this on | |||
1701 | /// any bit width. Exactly 32 bits will be translated. | |||
1702 | float bitsToFloat() const { | |||
1703 | return BitsToFloat(getWord(0)); | |||
1704 | } | |||
1705 | ||||
1706 | /// \brief Converts a double to APInt bits. | |||
1707 | /// | |||
1708 | /// The conversion does not do a translation from double to integer, it just | |||
1709 | /// re-interprets the bits of the double. | |||
1710 | static APInt doubleToBits(double V) { | |||
1711 | return APInt(sizeof(double) * CHAR_BIT8, DoubleToBits(V)); | |||
1712 | } | |||
1713 | ||||
1714 | /// \brief Converts a float to APInt bits. | |||
1715 | /// | |||
1716 | /// The conversion does not do a translation from float to integer, it just | |||
1717 | /// re-interprets the bits of the float. | |||
1718 | static APInt floatToBits(float V) { | |||
1719 | return APInt(sizeof(float) * CHAR_BIT8, FloatToBits(V)); | |||
1720 | } | |||
1721 | ||||
1722 | /// @} | |||
1723 | /// \name Mathematics Operations | |||
1724 | /// @{ | |||
1725 | ||||
1726 | /// \returns the floor log base 2 of this APInt. | |||
1727 | unsigned logBase2() const { return getActiveBits() - 1; } | |||
1728 | ||||
1729 | /// \returns the ceil log base 2 of this APInt. | |||
1730 | unsigned ceilLogBase2() const { | |||
1731 | APInt temp(*this); | |||
1732 | --temp; | |||
1733 | return temp.getActiveBits(); | |||
1734 | } | |||
1735 | ||||
1736 | /// \returns the nearest log base 2 of this APInt. Ties round up. | |||
1737 | /// | |||
1738 | /// NOTE: When we have a BitWidth of 1, we define: | |||
1739 | /// | |||
1740 | /// log2(0) = UINT32_MAX | |||
1741 | /// log2(1) = 0 | |||
1742 | /// | |||
1743 | /// to get around any mathematical concerns resulting from | |||
1744 | /// referencing 2 in a space where 2 does no exist. | |||
1745 | unsigned nearestLogBase2() const { | |||
1746 | // Special case when we have a bitwidth of 1. If VAL is 1, then we | |||
1747 | // get 0. If VAL is 0, we get WORD_MAX which gets truncated to | |||
1748 | // UINT32_MAX. | |||
1749 | if (BitWidth == 1) | |||
1750 | return U.VAL - 1; | |||
1751 | ||||
1752 | // Handle the zero case. | |||
1753 | if (isNullValue()) | |||
1754 | return UINT32_MAX(4294967295U); | |||
1755 | ||||
1756 | // The non-zero case is handled by computing: | |||
1757 | // | |||
1758 | // nearestLogBase2(x) = logBase2(x) + x[logBase2(x)-1]. | |||
1759 | // | |||
1760 | // where x[i] is referring to the value of the ith bit of x. | |||
1761 | unsigned lg = logBase2(); | |||
1762 | return lg + unsigned((*this)[lg - 1]); | |||
1763 | } | |||
1764 | ||||
1765 | /// \returns the log base 2 of this APInt if its an exact power of two, -1 | |||
1766 | /// otherwise | |||
1767 | int32_t exactLogBase2() const { | |||
1768 | if (!isPowerOf2()) | |||
1769 | return -1; | |||
1770 | return logBase2(); | |||
1771 | } | |||
1772 | ||||
1773 | /// \brief Compute the square root | |||
1774 | APInt sqrt() const; | |||
1775 | ||||
1776 | /// \brief Get the absolute value; | |||
1777 | /// | |||
1778 | /// If *this is < 0 then return -(*this), otherwise *this; | |||
1779 | APInt abs() const { | |||
1780 | if (isNegative()) | |||
1781 | return -(*this); | |||
1782 | return *this; | |||
1783 | } | |||
1784 | ||||
1785 | /// \returns the multiplicative inverse for a given modulo. | |||
1786 | APInt multiplicativeInverse(const APInt &modulo) const; | |||
1787 | ||||
1788 | /// @} | |||
1789 | /// \name Support for division by constant | |||
1790 | /// @{ | |||
1791 | ||||
1792 | /// Calculate the magic number for signed division by a constant. | |||
1793 | struct ms; | |||
1794 | ms magic() const; | |||
1795 | ||||
1796 | /// Calculate the magic number for unsigned division by a constant. | |||
1797 | struct mu; | |||
1798 | mu magicu(unsigned LeadingZeros = 0) const; | |||
1799 | ||||
1800 | /// @} | |||
1801 | /// \name Building-block Operations for APInt and APFloat | |||
1802 | /// @{ | |||
1803 | ||||
1804 | // These building block operations operate on a representation of arbitrary | |||
1805 | // precision, two's-complement, bignum integer values. They should be | |||
1806 | // sufficient to implement APInt and APFloat bignum requirements. Inputs are | |||
1807 | // generally a pointer to the base of an array of integer parts, representing | |||
1808 | // an unsigned bignum, and a count of how many parts there are. | |||
1809 | ||||
1810 | /// Sets the least significant part of a bignum to the input value, and zeroes | |||
1811 | /// out higher parts. | |||
1812 | static void tcSet(WordType *, WordType, unsigned); | |||
1813 | ||||
1814 | /// Assign one bignum to another. | |||
1815 | static void tcAssign(WordType *, const WordType *, unsigned); | |||
1816 | ||||
1817 | /// Returns true if a bignum is zero, false otherwise. | |||
1818 | static bool tcIsZero(const WordType *, unsigned); | |||
1819 | ||||
1820 | /// Extract the given bit of a bignum; returns 0 or 1. Zero-based. | |||
1821 | static int tcExtractBit(const WordType *, unsigned bit); | |||
1822 | ||||
1823 | /// Copy the bit vector of width srcBITS from SRC, starting at bit srcLSB, to | |||
1824 | /// DST, of dstCOUNT parts, such that the bit srcLSB becomes the least | |||
1825 | /// significant bit of DST. All high bits above srcBITS in DST are | |||
1826 | /// zero-filled. | |||
1827 | static void tcExtract(WordType *, unsigned dstCount, | |||
1828 | const WordType *, unsigned srcBits, | |||
1829 | unsigned srcLSB); | |||
1830 | ||||
1831 | /// Set the given bit of a bignum. Zero-based. | |||
1832 | static void tcSetBit(WordType *, unsigned bit); | |||
1833 | ||||
1834 | /// Clear the given bit of a bignum. Zero-based. | |||
1835 | static void tcClearBit(WordType *, unsigned bit); | |||
1836 | ||||
1837 | /// Returns the bit number of the least or most significant set bit of a | |||
1838 | /// number. If the input number has no bits set -1U is returned. | |||
1839 | static unsigned tcLSB(const WordType *, unsigned n); | |||
1840 | static unsigned tcMSB(const WordType *parts, unsigned n); | |||
1841 | ||||
1842 | /// Negate a bignum in-place. | |||
1843 | static void tcNegate(WordType *, unsigned); | |||
1844 | ||||
1845 | /// DST += RHS + CARRY where CARRY is zero or one. Returns the carry flag. | |||
1846 | static WordType tcAdd(WordType *, const WordType *, | |||
1847 | WordType carry, unsigned); | |||
1848 | /// DST += RHS. Returns the carry flag. | |||
1849 | static WordType tcAddPart(WordType *, WordType, unsigned); | |||
1850 | ||||
1851 | /// DST -= RHS + CARRY where CARRY is zero or one. Returns the carry flag. | |||
1852 | static WordType tcSubtract(WordType *, const WordType *, | |||
1853 | WordType carry, unsigned); | |||
1854 | /// DST -= RHS. Returns the carry flag. | |||
1855 | static WordType tcSubtractPart(WordType *, WordType, unsigned); | |||
1856 | ||||
1857 | /// DST += SRC * MULTIPLIER + PART if add is true | |||
1858 | /// DST = SRC * MULTIPLIER + PART if add is false | |||
1859 | /// | |||
1860 | /// Requires 0 <= DSTPARTS <= SRCPARTS + 1. If DST overlaps SRC they must | |||
1861 | /// start at the same point, i.e. DST == SRC. | |||
1862 | /// | |||
1863 | /// If DSTPARTS == SRC_PARTS + 1 no overflow occurs and zero is returned. | |||
1864 | /// Otherwise DST is filled with the least significant DSTPARTS parts of the | |||
1865 | /// result, and if all of the omitted higher parts were zero return zero, | |||
1866 | /// otherwise overflow occurred and return one. | |||
1867 | static int tcMultiplyPart(WordType *dst, const WordType *src, | |||
1868 | WordType multiplier, WordType carry, | |||
1869 | unsigned srcParts, unsigned dstParts, | |||
1870 | bool add); | |||
1871 | ||||
1872 | /// DST = LHS * RHS, where DST has the same width as the operands and is | |||
1873 | /// filled with the least significant parts of the result. Returns one if | |||
1874 | /// overflow occurred, otherwise zero. DST must be disjoint from both | |||
1875 | /// operands. | |||
1876 | static int tcMultiply(WordType *, const WordType *, const WordType *, | |||
1877 | unsigned); | |||
1878 | ||||
1879 | /// DST = LHS * RHS, where DST has width the sum of the widths of the | |||
1880 | /// operands. No overflow occurs. DST must be disjoint from both operands. | |||
1881 | static void tcFullMultiply(WordType *, const WordType *, | |||
1882 | const WordType *, unsigned, unsigned); | |||
1883 | ||||
1884 | /// If RHS is zero LHS and REMAINDER are left unchanged, return one. | |||
1885 | /// Otherwise set LHS to LHS / RHS with the fractional part discarded, set | |||
1886 | /// REMAINDER to the remainder, return zero. i.e. | |||
1887 | /// | |||
1888 | /// OLD_LHS = RHS * LHS + REMAINDER | |||
1889 | /// | |||
1890 | /// SCRATCH is a bignum of the same size as the operands and result for use by | |||
1891 | /// the routine; its contents need not be initialized and are destroyed. LHS, | |||
1892 | /// REMAINDER and SCRATCH must be distinct. | |||
1893 | static int tcDivide(WordType *lhs, const WordType *rhs, | |||
1894 | WordType *remainder, WordType *scratch, | |||
1895 | unsigned parts); | |||
1896 | ||||
1897 | /// Shift a bignum left Count bits. Shifted in bits are zero. There are no | |||
1898 | /// restrictions on Count. | |||
1899 | static void tcShiftLeft(WordType *, unsigned Words, unsigned Count); | |||
1900 | ||||
1901 | /// Shift a bignum right Count bits. Shifted in bits are zero. There are no | |||
1902 | /// restrictions on Count. | |||
1903 | static void tcShiftRight(WordType *, unsigned Words, unsigned Count); | |||
1904 | ||||
1905 | /// The obvious AND, OR and XOR and complement operations. | |||
1906 | static void tcAnd(WordType *, const WordType *, unsigned); | |||
1907 | static void tcOr(WordType *, const WordType *, unsigned); | |||
1908 | static void tcXor(WordType *, const WordType *, unsigned); | |||
1909 | static void tcComplement(WordType *, unsigned); | |||
1910 | ||||
1911 | /// Comparison (unsigned) of two bignums. | |||
1912 | static int tcCompare(const WordType *, const WordType *, unsigned); | |||
1913 | ||||
1914 | /// Increment a bignum in-place. Return the carry flag. | |||
1915 | static WordType tcIncrement(WordType *dst, unsigned parts) { | |||
1916 | return tcAddPart(dst, 1, parts); | |||
1917 | } | |||
1918 | ||||
1919 | /// Decrement a bignum in-place. Return the borrow flag. | |||
1920 | static WordType tcDecrement(WordType *dst, unsigned parts) { | |||
1921 | return tcSubtractPart(dst, 1, parts); | |||
1922 | } | |||
1923 | ||||
1924 | /// Set the least significant BITS and clear the rest. | |||
1925 | static void tcSetLeastSignificantBits(WordType *, unsigned, unsigned bits); | |||
1926 | ||||
1927 | /// \brief debug method | |||
1928 | void dump() const; | |||
1929 | ||||
1930 | /// @} | |||
1931 | }; | |||
1932 | ||||
1933 | /// Magic data for optimising signed division by a constant. | |||
1934 | struct APInt::ms { | |||
1935 | APInt m; ///< magic number | |||
1936 | unsigned s; ///< shift amount | |||
1937 | }; | |||
1938 | ||||
1939 | /// Magic data for optimising unsigned division by a constant. | |||
1940 | struct APInt::mu { | |||
1941 | APInt m; ///< magic number | |||
1942 | bool a; ///< add indicator | |||
1943 | unsigned s; ///< shift amount | |||
1944 | }; | |||
1945 | ||||
1946 | inline bool operator==(uint64_t V1, const APInt &V2) { return V2 == V1; } | |||
1947 | ||||
1948 | inline bool operator!=(uint64_t V1, const APInt &V2) { return V2 != V1; } | |||
1949 | ||||
1950 | /// \brief Unary bitwise complement operator. | |||
1951 | /// | |||
1952 | /// \returns an APInt that is the bitwise complement of \p v. | |||
1953 | inline APInt operator~(APInt v) { | |||
1954 | v.flipAllBits(); | |||
1955 | return v; | |||
1956 | } | |||
1957 | ||||
1958 | inline APInt operator&(APInt a, const APInt &b) { | |||
1959 | a &= b; | |||
1960 | return a; | |||
1961 | } | |||
1962 | ||||
1963 | inline APInt operator&(const APInt &a, APInt &&b) { | |||
1964 | b &= a; | |||
1965 | return std::move(b); | |||
1966 | } | |||
1967 | ||||
1968 | inline APInt operator&(APInt a, uint64_t RHS) { | |||
1969 | a &= RHS; | |||
1970 | return a; | |||
1971 | } | |||
1972 | ||||
1973 | inline APInt operator&(uint64_t LHS, APInt b) { | |||
1974 | b &= LHS; | |||
1975 | return b; | |||
1976 | } | |||
1977 | ||||
1978 | inline APInt operator|(APInt a, const APInt &b) { | |||
1979 | a |= b; | |||
1980 | return a; | |||
1981 | } | |||
1982 | ||||
1983 | inline APInt operator|(const APInt &a, APInt &&b) { | |||
1984 | b |= a; | |||
1985 | return std::move(b); | |||
1986 | } | |||
1987 | ||||
1988 | inline APInt operator|(APInt a, uint64_t RHS) { | |||
1989 | a |= RHS; | |||
1990 | return a; | |||
1991 | } | |||
1992 | ||||
1993 | inline APInt operator|(uint64_t LHS, APInt b) { | |||
1994 | b |= LHS; | |||
1995 | return b; | |||
1996 | } | |||
1997 | ||||
1998 | inline APInt operator^(APInt a, const APInt &b) { | |||
1999 | a ^= b; | |||
2000 | return a; | |||
2001 | } | |||
2002 | ||||
2003 | inline APInt operator^(const APInt &a, APInt &&b) { | |||
2004 | b ^= a; | |||
2005 | return std::move(b); | |||
2006 | } | |||
2007 | ||||
2008 | inline APInt operator^(APInt a, uint64_t RHS) { | |||
2009 | a ^= RHS; | |||
2010 | return a; | |||
2011 | } | |||
2012 | ||||
2013 | inline APInt operator^(uint64_t LHS, APInt b) { | |||
2014 | b ^= LHS; | |||
2015 | return b; | |||
2016 | } | |||
2017 | ||||
2018 | inline raw_ostream &operator<<(raw_ostream &OS, const APInt &I) { | |||
2019 | I.print(OS, true); | |||
2020 | return OS; | |||
2021 | } | |||
2022 | ||||
2023 | inline APInt operator-(APInt v) { | |||
2024 | v.negate(); | |||
2025 | return v; | |||
2026 | } | |||
2027 | ||||
2028 | inline APInt operator+(APInt a, const APInt &b) { | |||
2029 | a += b; | |||
2030 | return a; | |||
2031 | } | |||
2032 | ||||
2033 | inline APInt operator+(const APInt &a, APInt &&b) { | |||
2034 | b += a; | |||
2035 | return std::move(b); | |||
2036 | } | |||
2037 | ||||
2038 | inline APInt operator+(APInt a, uint64_t RHS) { | |||
2039 | a += RHS; | |||
2040 | return a; | |||
2041 | } | |||
2042 | ||||
2043 | inline APInt operator+(uint64_t LHS, APInt b) { | |||
2044 | b += LHS; | |||
2045 | return b; | |||
2046 | } | |||
2047 | ||||
2048 | inline APInt operator-(APInt a, const APInt &b) { | |||
2049 | a -= b; | |||
2050 | return a; | |||
2051 | } | |||
2052 | ||||
2053 | inline APInt operator-(const APInt &a, APInt &&b) { | |||
2054 | b.negate(); | |||
2055 | b += a; | |||
2056 | return std::move(b); | |||
2057 | } | |||
2058 | ||||
2059 | inline APInt operator-(APInt a, uint64_t RHS) { | |||
2060 | a -= RHS; | |||
2061 | return a; | |||
2062 | } | |||
2063 | ||||
2064 | inline APInt operator-(uint64_t LHS, APInt b) { | |||
2065 | b.negate(); | |||
2066 | b += LHS; | |||
2067 | return b; | |||
2068 | } | |||
2069 | ||||
2070 | inline APInt operator*(APInt a, uint64_t RHS) { | |||
2071 | a *= RHS; | |||
2072 | return a; | |||
2073 | } | |||
2074 | ||||
2075 | inline APInt operator*(uint64_t LHS, APInt b) { | |||
2076 | b *= LHS; | |||
2077 | return b; | |||
2078 | } | |||
2079 | ||||
2080 | ||||
2081 | namespace APIntOps { | |||
2082 | ||||
2083 | /// \brief Determine the smaller of two APInts considered to be signed. | |||
2084 | inline const APInt &smin(const APInt &A, const APInt &B) { | |||
2085 | return A.slt(B) ? A : B; | |||
2086 | } | |||
2087 | ||||
2088 | /// \brief Determine the larger of two APInts considered to be signed. | |||
2089 | inline const APInt &smax(const APInt &A, const APInt &B) { | |||
2090 | return A.sgt(B) ? A : B; | |||
2091 | } | |||
2092 | ||||
2093 | /// \brief Determine the smaller of two APInts considered to be signed. | |||
2094 | inline const APInt &umin(const APInt &A, const APInt &B) { | |||
2095 | return A.ult(B) ? A : B; | |||
2096 | } | |||
2097 | ||||
2098 | /// \brief Determine the larger of two APInts considered to be unsigned. | |||
2099 | inline const APInt &umax(const APInt &A, const APInt &B) { | |||
2100 | return A.ugt(B) ? A : B; | |||
2101 | } | |||
2102 | ||||
2103 | /// \brief Compute GCD of two unsigned APInt values. | |||
2104 | /// | |||
2105 | /// This function returns the greatest common divisor of the two APInt values | |||
2106 | /// using Stein's algorithm. | |||
2107 | /// | |||
2108 | /// \returns the greatest common divisor of A and B. | |||
2109 | APInt GreatestCommonDivisor(APInt A, APInt B); | |||
2110 | ||||
2111 | /// \brief Converts the given APInt to a double value. | |||
2112 | /// | |||
2113 | /// Treats the APInt as an unsigned value for conversion purposes. | |||
2114 | inline double RoundAPIntToDouble(const APInt &APIVal) { | |||
2115 | return APIVal.roundToDouble(); | |||
2116 | } | |||
2117 | ||||
2118 | /// \brief Converts the given APInt to a double value. | |||
2119 | /// | |||
2120 | /// Treats the APInt as a signed value for conversion purposes. | |||
2121 | inline double RoundSignedAPIntToDouble(const APInt &APIVal) { | |||
2122 | return APIVal.signedRoundToDouble(); | |||
2123 | } | |||
2124 | ||||
2125 | /// \brief Converts the given APInt to a float vlalue. | |||
2126 | inline float RoundAPIntToFloat(const APInt &APIVal) { | |||
2127 | return float(RoundAPIntToDouble(APIVal)); | |||
2128 | } | |||
2129 | ||||
2130 | /// \brief Converts the given APInt to a float value. | |||
2131 | /// | |||
2132 | /// Treast the APInt as a signed value for conversion purposes. | |||
2133 | inline float RoundSignedAPIntToFloat(const APInt &APIVal) { | |||
2134 | return float(APIVal.signedRoundToDouble()); | |||
2135 | } | |||
2136 | ||||
2137 | /// \brief Converts the given double value into a APInt. | |||
2138 | /// | |||
2139 | /// This function convert a double value to an APInt value. | |||
2140 | APInt RoundDoubleToAPInt(double Double, unsigned width); | |||
2141 | ||||
2142 | /// \brief Converts a float value into a APInt. | |||
2143 | /// | |||
2144 | /// Converts a float value into an APInt value. | |||
2145 | inline APInt RoundFloatToAPInt(float Float, unsigned width) { | |||
2146 | return RoundDoubleToAPInt(double(Float), width); | |||
2147 | } | |||
2148 | ||||
2149 | } // End of APIntOps namespace | |||
2150 | ||||
2151 | // See friend declaration above. This additional declaration is required in | |||
2152 | // order to compile LLVM with IBM xlC compiler. | |||
2153 | hash_code hash_value(const APInt &Arg); | |||
2154 | } // End of llvm namespace | |||
2155 | ||||
2156 | #endif |