File: | lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp |
Location: | line 1411, column 9 |
Description: | Value stored to 'Condition' is never read |
1 | //===-- SelectionDAGBuilder.cpp - Selection-DAG building ------------------===// |
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 implements routines for translating from LLVM IR into SelectionDAG IR. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "SelectionDAGBuilder.h" |
15 | #include "SDNodeDbgValue.h" |
16 | #include "llvm/ADT/BitVector.h" |
17 | #include "llvm/ADT/Optional.h" |
18 | #include "llvm/ADT/SmallSet.h" |
19 | #include "llvm/ADT/Statistic.h" |
20 | #include "llvm/Analysis/AliasAnalysis.h" |
21 | #include "llvm/Analysis/BranchProbabilityInfo.h" |
22 | #include "llvm/Analysis/ConstantFolding.h" |
23 | #include "llvm/Analysis/ValueTracking.h" |
24 | #include "llvm/CodeGen/Analysis.h" |
25 | #include "llvm/CodeGen/FastISel.h" |
26 | #include "llvm/CodeGen/FunctionLoweringInfo.h" |
27 | #include "llvm/CodeGen/GCMetadata.h" |
28 | #include "llvm/CodeGen/GCStrategy.h" |
29 | #include "llvm/CodeGen/MachineFrameInfo.h" |
30 | #include "llvm/CodeGen/MachineFunction.h" |
31 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
32 | #include "llvm/CodeGen/MachineJumpTableInfo.h" |
33 | #include "llvm/CodeGen/MachineModuleInfo.h" |
34 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
35 | #include "llvm/CodeGen/SelectionDAG.h" |
36 | #include "llvm/CodeGen/StackMaps.h" |
37 | #include "llvm/IR/CallingConv.h" |
38 | #include "llvm/IR/Constants.h" |
39 | #include "llvm/IR/DataLayout.h" |
40 | #include "llvm/IR/DebugInfo.h" |
41 | #include "llvm/IR/DerivedTypes.h" |
42 | #include "llvm/IR/Function.h" |
43 | #include "llvm/IR/GlobalVariable.h" |
44 | #include "llvm/IR/InlineAsm.h" |
45 | #include "llvm/IR/Instructions.h" |
46 | #include "llvm/IR/IntrinsicInst.h" |
47 | #include "llvm/IR/Intrinsics.h" |
48 | #include "llvm/IR/LLVMContext.h" |
49 | #include "llvm/IR/Module.h" |
50 | #include "llvm/IR/Statepoint.h" |
51 | #include "llvm/Support/CommandLine.h" |
52 | #include "llvm/Support/Debug.h" |
53 | #include "llvm/Support/ErrorHandling.h" |
54 | #include "llvm/Support/MathExtras.h" |
55 | #include "llvm/Support/raw_ostream.h" |
56 | #include "llvm/Target/TargetFrameLowering.h" |
57 | #include "llvm/Target/TargetInstrInfo.h" |
58 | #include "llvm/Target/TargetIntrinsicInfo.h" |
59 | #include "llvm/Target/TargetLibraryInfo.h" |
60 | #include "llvm/Target/TargetLowering.h" |
61 | #include "llvm/Target/TargetOptions.h" |
62 | #include "llvm/Target/TargetSelectionDAGInfo.h" |
63 | #include "llvm/Target/TargetSubtargetInfo.h" |
64 | #include <algorithm> |
65 | using namespace llvm; |
66 | |
67 | #define DEBUG_TYPE"isel" "isel" |
68 | |
69 | /// LimitFloatPrecision - Generate low-precision inline sequences for |
70 | /// some float libcalls (6, 8 or 12 bits). |
71 | static unsigned LimitFloatPrecision; |
72 | |
73 | static cl::opt<unsigned, true> |
74 | LimitFPPrecision("limit-float-precision", |
75 | cl::desc("Generate low-precision inline sequences " |
76 | "for some float libcalls"), |
77 | cl::location(LimitFloatPrecision), |
78 | cl::init(0)); |
79 | |
80 | // Limit the width of DAG chains. This is important in general to prevent |
81 | // prevent DAG-based analysis from blowing up. For example, alias analysis and |
82 | // load clustering may not complete in reasonable time. It is difficult to |
83 | // recognize and avoid this situation within each individual analysis, and |
84 | // future analyses are likely to have the same behavior. Limiting DAG width is |
85 | // the safe approach, and will be especially important with global DAGs. |
86 | // |
87 | // MaxParallelChains default is arbitrarily high to avoid affecting |
88 | // optimization, but could be lowered to improve compile time. Any ld-ld-st-st |
89 | // sequence over this should have been converted to llvm.memcpy by the |
90 | // frontend. It easy to induce this behavior with .ll code such as: |
91 | // %buffer = alloca [4096 x i8] |
92 | // %data = load [4096 x i8]* %argPtr |
93 | // store [4096 x i8] %data, [4096 x i8]* %buffer |
94 | static const unsigned MaxParallelChains = 64; |
95 | |
96 | static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL, |
97 | const SDValue *Parts, unsigned NumParts, |
98 | MVT PartVT, EVT ValueVT, const Value *V); |
99 | |
100 | /// getCopyFromParts - Create a value that contains the specified legal parts |
101 | /// combined into the value they represent. If the parts combine to a type |
102 | /// larger then ValueVT then AssertOp can be used to specify whether the extra |
103 | /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT |
104 | /// (ISD::AssertSext). |
105 | static SDValue getCopyFromParts(SelectionDAG &DAG, SDLoc DL, |
106 | const SDValue *Parts, |
107 | unsigned NumParts, MVT PartVT, EVT ValueVT, |
108 | const Value *V, |
109 | ISD::NodeType AssertOp = ISD::DELETED_NODE) { |
110 | if (ValueVT.isVector()) |
111 | return getCopyFromPartsVector(DAG, DL, Parts, NumParts, |
112 | PartVT, ValueVT, V); |
113 | |
114 | assert(NumParts > 0 && "No parts to assemble!")((NumParts > 0 && "No parts to assemble!") ? static_cast <void> (0) : __assert_fail ("NumParts > 0 && \"No parts to assemble!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 114, __PRETTY_FUNCTION__)); |
115 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
116 | SDValue Val = Parts[0]; |
117 | |
118 | if (NumParts > 1) { |
119 | // Assemble the value from multiple parts. |
120 | if (ValueVT.isInteger()) { |
121 | unsigned PartBits = PartVT.getSizeInBits(); |
122 | unsigned ValueBits = ValueVT.getSizeInBits(); |
123 | |
124 | // Assemble the power of 2 part. |
125 | unsigned RoundParts = NumParts & (NumParts - 1) ? |
126 | 1 << Log2_32(NumParts) : NumParts; |
127 | unsigned RoundBits = PartBits * RoundParts; |
128 | EVT RoundVT = RoundBits == ValueBits ? |
129 | ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits); |
130 | SDValue Lo, Hi; |
131 | |
132 | EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2); |
133 | |
134 | if (RoundParts > 2) { |
135 | Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2, |
136 | PartVT, HalfVT, V); |
137 | Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2, |
138 | RoundParts / 2, PartVT, HalfVT, V); |
139 | } else { |
140 | Lo = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[0]); |
141 | Hi = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[1]); |
142 | } |
143 | |
144 | if (TLI.isBigEndian()) |
145 | std::swap(Lo, Hi); |
146 | |
147 | Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi); |
148 | |
149 | if (RoundParts < NumParts) { |
150 | // Assemble the trailing non-power-of-2 part. |
151 | unsigned OddParts = NumParts - RoundParts; |
152 | EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits); |
153 | Hi = getCopyFromParts(DAG, DL, |
154 | Parts + RoundParts, OddParts, PartVT, OddVT, V); |
155 | |
156 | // Combine the round and odd parts. |
157 | Lo = Val; |
158 | if (TLI.isBigEndian()) |
159 | std::swap(Lo, Hi); |
160 | EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); |
161 | Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi); |
162 | Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi, |
163 | DAG.getConstant(Lo.getValueType().getSizeInBits(), |
164 | TLI.getPointerTy())); |
165 | Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo); |
166 | Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi); |
167 | } |
168 | } else if (PartVT.isFloatingPoint()) { |
169 | // FP split into multiple FP parts (for ppcf128) |
170 | assert(ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 &&((ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 && "Unexpected split") ? static_cast<void> (0) : __assert_fail ("ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 && \"Unexpected split\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 171, __PRETTY_FUNCTION__)) |
171 | "Unexpected split")((ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 && "Unexpected split") ? static_cast<void> (0) : __assert_fail ("ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 && \"Unexpected split\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 171, __PRETTY_FUNCTION__)); |
172 | SDValue Lo, Hi; |
173 | Lo = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[0]); |
174 | Hi = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[1]); |
175 | if (TLI.hasBigEndianPartOrdering(ValueVT)) |
176 | std::swap(Lo, Hi); |
177 | Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi); |
178 | } else { |
179 | // FP split into integer parts (soft fp) |
180 | assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&((ValueVT.isFloatingPoint() && PartVT.isInteger() && !PartVT.isVector() && "Unexpected split") ? static_cast <void> (0) : __assert_fail ("ValueVT.isFloatingPoint() && PartVT.isInteger() && !PartVT.isVector() && \"Unexpected split\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 181, __PRETTY_FUNCTION__)) |
181 | !PartVT.isVector() && "Unexpected split")((ValueVT.isFloatingPoint() && PartVT.isInteger() && !PartVT.isVector() && "Unexpected split") ? static_cast <void> (0) : __assert_fail ("ValueVT.isFloatingPoint() && PartVT.isInteger() && !PartVT.isVector() && \"Unexpected split\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 181, __PRETTY_FUNCTION__)); |
182 | EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits()); |
183 | Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT, V); |
184 | } |
185 | } |
186 | |
187 | // There is now one part, held in Val. Correct it to match ValueVT. |
188 | EVT PartEVT = Val.getValueType(); |
189 | |
190 | if (PartEVT == ValueVT) |
191 | return Val; |
192 | |
193 | if (PartEVT.isInteger() && ValueVT.isInteger()) { |
194 | if (ValueVT.bitsLT(PartEVT)) { |
195 | // For a truncate, see if we have any information to |
196 | // indicate whether the truncated bits will always be |
197 | // zero or sign-extension. |
198 | if (AssertOp != ISD::DELETED_NODE) |
199 | Val = DAG.getNode(AssertOp, DL, PartEVT, Val, |
200 | DAG.getValueType(ValueVT)); |
201 | return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); |
202 | } |
203 | return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val); |
204 | } |
205 | |
206 | if (PartEVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { |
207 | // FP_ROUND's are always exact here. |
208 | if (ValueVT.bitsLT(Val.getValueType())) |
209 | return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val, |
210 | DAG.getTargetConstant(1, TLI.getPointerTy())); |
211 | |
212 | return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val); |
213 | } |
214 | |
215 | if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits()) |
216 | return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val); |
217 | |
218 | llvm_unreachable("Unknown mismatch!")::llvm::llvm_unreachable_internal("Unknown mismatch!", "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 218); |
219 | } |
220 | |
221 | static void diagnosePossiblyInvalidConstraint(LLVMContext &Ctx, const Value *V, |
222 | const Twine &ErrMsg) { |
223 | const Instruction *I = dyn_cast_or_null<Instruction>(V); |
224 | if (!V) |
225 | return Ctx.emitError(ErrMsg); |
226 | |
227 | const char *AsmError = ", possible invalid constraint for vector type"; |
228 | if (const CallInst *CI = dyn_cast<CallInst>(I)) |
229 | if (isa<InlineAsm>(CI->getCalledValue())) |
230 | return Ctx.emitError(I, ErrMsg + AsmError); |
231 | |
232 | return Ctx.emitError(I, ErrMsg); |
233 | } |
234 | |
235 | /// getCopyFromPartsVector - Create a value that contains the specified legal |
236 | /// parts combined into the value they represent. If the parts combine to a |
237 | /// type larger then ValueVT then AssertOp can be used to specify whether the |
238 | /// extra bits are known to be zero (ISD::AssertZext) or sign extended from |
239 | /// ValueVT (ISD::AssertSext). |
240 | static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL, |
241 | const SDValue *Parts, unsigned NumParts, |
242 | MVT PartVT, EVT ValueVT, const Value *V) { |
243 | assert(ValueVT.isVector() && "Not a vector value")((ValueVT.isVector() && "Not a vector value") ? static_cast <void> (0) : __assert_fail ("ValueVT.isVector() && \"Not a vector value\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 243, __PRETTY_FUNCTION__)); |
244 | assert(NumParts > 0 && "No parts to assemble!")((NumParts > 0 && "No parts to assemble!") ? static_cast <void> (0) : __assert_fail ("NumParts > 0 && \"No parts to assemble!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 244, __PRETTY_FUNCTION__)); |
245 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
246 | SDValue Val = Parts[0]; |
247 | |
248 | // Handle a multi-element vector. |
249 | if (NumParts > 1) { |
250 | EVT IntermediateVT; |
251 | MVT RegisterVT; |
252 | unsigned NumIntermediates; |
253 | unsigned NumRegs = |
254 | TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT, |
255 | NumIntermediates, RegisterVT); |
256 | assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!")((NumRegs == NumParts && "Part count doesn't match vector breakdown!" ) ? static_cast<void> (0) : __assert_fail ("NumRegs == NumParts && \"Part count doesn't match vector breakdown!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 256, __PRETTY_FUNCTION__)); |
257 | NumParts = NumRegs; // Silence a compiler warning. |
258 | assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!")((RegisterVT == PartVT && "Part type doesn't match vector breakdown!" ) ? static_cast<void> (0) : __assert_fail ("RegisterVT == PartVT && \"Part type doesn't match vector breakdown!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 258, __PRETTY_FUNCTION__)); |
259 | assert(RegisterVT == Parts[0].getSimpleValueType() &&((RegisterVT == Parts[0].getSimpleValueType() && "Part type doesn't match part!" ) ? static_cast<void> (0) : __assert_fail ("RegisterVT == Parts[0].getSimpleValueType() && \"Part type doesn't match part!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 260, __PRETTY_FUNCTION__)) |
260 | "Part type doesn't match part!")((RegisterVT == Parts[0].getSimpleValueType() && "Part type doesn't match part!" ) ? static_cast<void> (0) : __assert_fail ("RegisterVT == Parts[0].getSimpleValueType() && \"Part type doesn't match part!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 260, __PRETTY_FUNCTION__)); |
261 | |
262 | // Assemble the parts into intermediate operands. |
263 | SmallVector<SDValue, 8> Ops(NumIntermediates); |
264 | if (NumIntermediates == NumParts) { |
265 | // If the register was not expanded, truncate or copy the value, |
266 | // as appropriate. |
267 | for (unsigned i = 0; i != NumParts; ++i) |
268 | Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1, |
269 | PartVT, IntermediateVT, V); |
270 | } else if (NumParts > 0) { |
271 | // If the intermediate type was expanded, build the intermediate |
272 | // operands from the parts. |
273 | assert(NumParts % NumIntermediates == 0 &&((NumParts % NumIntermediates == 0 && "Must expand into a divisible number of parts!" ) ? static_cast<void> (0) : __assert_fail ("NumParts % NumIntermediates == 0 && \"Must expand into a divisible number of parts!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 274, __PRETTY_FUNCTION__)) |
274 | "Must expand into a divisible number of parts!")((NumParts % NumIntermediates == 0 && "Must expand into a divisible number of parts!" ) ? static_cast<void> (0) : __assert_fail ("NumParts % NumIntermediates == 0 && \"Must expand into a divisible number of parts!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 274, __PRETTY_FUNCTION__)); |
275 | unsigned Factor = NumParts / NumIntermediates; |
276 | for (unsigned i = 0; i != NumIntermediates; ++i) |
277 | Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor, |
278 | PartVT, IntermediateVT, V); |
279 | } |
280 | |
281 | // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the |
282 | // intermediate operands. |
283 | Val = DAG.getNode(IntermediateVT.isVector() ? ISD::CONCAT_VECTORS |
284 | : ISD::BUILD_VECTOR, |
285 | DL, ValueVT, Ops); |
286 | } |
287 | |
288 | // There is now one part, held in Val. Correct it to match ValueVT. |
289 | EVT PartEVT = Val.getValueType(); |
290 | |
291 | if (PartEVT == ValueVT) |
292 | return Val; |
293 | |
294 | if (PartEVT.isVector()) { |
295 | // If the element type of the source/dest vectors are the same, but the |
296 | // parts vector has more elements than the value vector, then we have a |
297 | // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the |
298 | // elements we want. |
299 | if (PartEVT.getVectorElementType() == ValueVT.getVectorElementType()) { |
300 | assert(PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&((PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements () && "Cannot narrow, it would be a lossy transformation" ) ? static_cast<void> (0) : __assert_fail ("PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements() && \"Cannot narrow, it would be a lossy transformation\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 301, __PRETTY_FUNCTION__)) |
301 | "Cannot narrow, it would be a lossy transformation")((PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements () && "Cannot narrow, it would be a lossy transformation" ) ? static_cast<void> (0) : __assert_fail ("PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements() && \"Cannot narrow, it would be a lossy transformation\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 301, __PRETTY_FUNCTION__)); |
302 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val, |
303 | DAG.getConstant(0, TLI.getVectorIdxTy())); |
304 | } |
305 | |
306 | // Vector/Vector bitcast. |
307 | if (ValueVT.getSizeInBits() == PartEVT.getSizeInBits()) |
308 | return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val); |
309 | |
310 | assert(PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements() &&((PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements () && "Cannot handle this kind of promotion") ? static_cast <void> (0) : __assert_fail ("PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements() && \"Cannot handle this kind of promotion\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 311, __PRETTY_FUNCTION__)) |
311 | "Cannot handle this kind of promotion")((PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements () && "Cannot handle this kind of promotion") ? static_cast <void> (0) : __assert_fail ("PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements() && \"Cannot handle this kind of promotion\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 311, __PRETTY_FUNCTION__)); |
312 | // Promoted vector extract |
313 | bool Smaller = ValueVT.bitsLE(PartEVT); |
314 | return DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND), |
315 | DL, ValueVT, Val); |
316 | |
317 | } |
318 | |
319 | // Trivial bitcast if the types are the same size and the destination |
320 | // vector type is legal. |
321 | if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits() && |
322 | TLI.isTypeLegal(ValueVT)) |
323 | return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val); |
324 | |
325 | // Handle cases such as i8 -> <1 x i1> |
326 | if (ValueVT.getVectorNumElements() != 1) { |
327 | diagnosePossiblyInvalidConstraint(*DAG.getContext(), V, |
328 | "non-trivial scalar-to-vector conversion"); |
329 | return DAG.getUNDEF(ValueVT); |
330 | } |
331 | |
332 | if (ValueVT.getVectorNumElements() == 1 && |
333 | ValueVT.getVectorElementType() != PartEVT) { |
334 | bool Smaller = ValueVT.bitsLE(PartEVT); |
335 | Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND), |
336 | DL, ValueVT.getScalarType(), Val); |
337 | } |
338 | |
339 | return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val); |
340 | } |
341 | |
342 | static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc dl, |
343 | SDValue Val, SDValue *Parts, unsigned NumParts, |
344 | MVT PartVT, const Value *V); |
345 | |
346 | /// getCopyToParts - Create a series of nodes that contain the specified value |
347 | /// split into legal parts. If the parts contain more bits than Val, then, for |
348 | /// integers, ExtendKind can be used to specify how to generate the extra bits. |
349 | static void getCopyToParts(SelectionDAG &DAG, SDLoc DL, |
350 | SDValue Val, SDValue *Parts, unsigned NumParts, |
351 | MVT PartVT, const Value *V, |
352 | ISD::NodeType ExtendKind = ISD::ANY_EXTEND) { |
353 | EVT ValueVT = Val.getValueType(); |
354 | |
355 | // Handle the vector case separately. |
356 | if (ValueVT.isVector()) |
357 | return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT, V); |
358 | |
359 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
360 | unsigned PartBits = PartVT.getSizeInBits(); |
361 | unsigned OrigNumParts = NumParts; |
362 | assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!")((TLI.isTypeLegal(PartVT) && "Copying to an illegal type!" ) ? static_cast<void> (0) : __assert_fail ("TLI.isTypeLegal(PartVT) && \"Copying to an illegal type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 362, __PRETTY_FUNCTION__)); |
363 | |
364 | if (NumParts == 0) |
365 | return; |
366 | |
367 | assert(!ValueVT.isVector() && "Vector case handled elsewhere")((!ValueVT.isVector() && "Vector case handled elsewhere" ) ? static_cast<void> (0) : __assert_fail ("!ValueVT.isVector() && \"Vector case handled elsewhere\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 367, __PRETTY_FUNCTION__)); |
368 | EVT PartEVT = PartVT; |
369 | if (PartEVT == ValueVT) { |
370 | assert(NumParts == 1 && "No-op copy with multiple parts!")((NumParts == 1 && "No-op copy with multiple parts!") ? static_cast<void> (0) : __assert_fail ("NumParts == 1 && \"No-op copy with multiple parts!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 370, __PRETTY_FUNCTION__)); |
371 | Parts[0] = Val; |
372 | return; |
373 | } |
374 | |
375 | if (NumParts * PartBits > ValueVT.getSizeInBits()) { |
376 | // If the parts cover more bits than the value has, promote the value. |
377 | if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { |
378 | assert(NumParts == 1 && "Do not know what to promote to!")((NumParts == 1 && "Do not know what to promote to!") ? static_cast<void> (0) : __assert_fail ("NumParts == 1 && \"Do not know what to promote to!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 378, __PRETTY_FUNCTION__)); |
379 | Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val); |
380 | } else { |
381 | assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&(((PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT .isInteger() && "Unknown mismatch!") ? static_cast< void> (0) : __assert_fail ("(PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT.isInteger() && \"Unknown mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 383, __PRETTY_FUNCTION__)) |
382 | ValueVT.isInteger() &&(((PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT .isInteger() && "Unknown mismatch!") ? static_cast< void> (0) : __assert_fail ("(PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT.isInteger() && \"Unknown mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 383, __PRETTY_FUNCTION__)) |
383 | "Unknown mismatch!")(((PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT .isInteger() && "Unknown mismatch!") ? static_cast< void> (0) : __assert_fail ("(PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT.isInteger() && \"Unknown mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 383, __PRETTY_FUNCTION__)); |
384 | ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); |
385 | Val = DAG.getNode(ExtendKind, DL, ValueVT, Val); |
386 | if (PartVT == MVT::x86mmx) |
387 | Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val); |
388 | } |
389 | } else if (PartBits == ValueVT.getSizeInBits()) { |
390 | // Different types of the same size. |
391 | assert(NumParts == 1 && PartEVT != ValueVT)((NumParts == 1 && PartEVT != ValueVT) ? static_cast< void> (0) : __assert_fail ("NumParts == 1 && PartEVT != ValueVT" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 391, __PRETTY_FUNCTION__)); |
392 | Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val); |
393 | } else if (NumParts * PartBits < ValueVT.getSizeInBits()) { |
394 | // If the parts cover less bits than value has, truncate the value. |
395 | assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&(((PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT .isInteger() && "Unknown mismatch!") ? static_cast< void> (0) : __assert_fail ("(PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT.isInteger() && \"Unknown mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 397, __PRETTY_FUNCTION__)) |
396 | ValueVT.isInteger() &&(((PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT .isInteger() && "Unknown mismatch!") ? static_cast< void> (0) : __assert_fail ("(PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT.isInteger() && \"Unknown mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 397, __PRETTY_FUNCTION__)) |
397 | "Unknown mismatch!")(((PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT .isInteger() && "Unknown mismatch!") ? static_cast< void> (0) : __assert_fail ("(PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT.isInteger() && \"Unknown mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 397, __PRETTY_FUNCTION__)); |
398 | ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); |
399 | Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); |
400 | if (PartVT == MVT::x86mmx) |
401 | Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val); |
402 | } |
403 | |
404 | // The value may have changed - recompute ValueVT. |
405 | ValueVT = Val.getValueType(); |
406 | assert(NumParts * PartBits == ValueVT.getSizeInBits() &&((NumParts * PartBits == ValueVT.getSizeInBits() && "Failed to tile the value with PartVT!" ) ? static_cast<void> (0) : __assert_fail ("NumParts * PartBits == ValueVT.getSizeInBits() && \"Failed to tile the value with PartVT!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 407, __PRETTY_FUNCTION__)) |
407 | "Failed to tile the value with PartVT!")((NumParts * PartBits == ValueVT.getSizeInBits() && "Failed to tile the value with PartVT!" ) ? static_cast<void> (0) : __assert_fail ("NumParts * PartBits == ValueVT.getSizeInBits() && \"Failed to tile the value with PartVT!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 407, __PRETTY_FUNCTION__)); |
408 | |
409 | if (NumParts == 1) { |
410 | if (PartEVT != ValueVT) |
411 | diagnosePossiblyInvalidConstraint(*DAG.getContext(), V, |
412 | "scalar-to-vector conversion failed"); |
413 | |
414 | Parts[0] = Val; |
415 | return; |
416 | } |
417 | |
418 | // Expand the value into multiple parts. |
419 | if (NumParts & (NumParts - 1)) { |
420 | // The number of parts is not a power of 2. Split off and copy the tail. |
421 | assert(PartVT.isInteger() && ValueVT.isInteger() &&((PartVT.isInteger() && ValueVT.isInteger() && "Do not know what to expand to!") ? static_cast<void> ( 0) : __assert_fail ("PartVT.isInteger() && ValueVT.isInteger() && \"Do not know what to expand to!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 422, __PRETTY_FUNCTION__)) |
422 | "Do not know what to expand to!")((PartVT.isInteger() && ValueVT.isInteger() && "Do not know what to expand to!") ? static_cast<void> ( 0) : __assert_fail ("PartVT.isInteger() && ValueVT.isInteger() && \"Do not know what to expand to!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 422, __PRETTY_FUNCTION__)); |
423 | unsigned RoundParts = 1 << Log2_32(NumParts); |
424 | unsigned RoundBits = RoundParts * PartBits; |
425 | unsigned OddParts = NumParts - RoundParts; |
426 | SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val, |
427 | DAG.getIntPtrConstant(RoundBits)); |
428 | getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT, V); |
429 | |
430 | if (TLI.isBigEndian()) |
431 | // The odd parts were reversed by getCopyToParts - unreverse them. |
432 | std::reverse(Parts + RoundParts, Parts + NumParts); |
433 | |
434 | NumParts = RoundParts; |
435 | ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); |
436 | Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); |
437 | } |
438 | |
439 | // The number of parts is a power of 2. Repeatedly bisect the value using |
440 | // EXTRACT_ELEMENT. |
441 | Parts[0] = DAG.getNode(ISD::BITCAST, DL, |
442 | EVT::getIntegerVT(*DAG.getContext(), |
443 | ValueVT.getSizeInBits()), |
444 | Val); |
445 | |
446 | for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) { |
447 | for (unsigned i = 0; i < NumParts; i += StepSize) { |
448 | unsigned ThisBits = StepSize * PartBits / 2; |
449 | EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits); |
450 | SDValue &Part0 = Parts[i]; |
451 | SDValue &Part1 = Parts[i+StepSize/2]; |
452 | |
453 | Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, |
454 | ThisVT, Part0, DAG.getIntPtrConstant(1)); |
455 | Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, |
456 | ThisVT, Part0, DAG.getIntPtrConstant(0)); |
457 | |
458 | if (ThisBits == PartBits && ThisVT != PartVT) { |
459 | Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0); |
460 | Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1); |
461 | } |
462 | } |
463 | } |
464 | |
465 | if (TLI.isBigEndian()) |
466 | std::reverse(Parts, Parts + OrigNumParts); |
467 | } |
468 | |
469 | |
470 | /// getCopyToPartsVector - Create a series of nodes that contain the specified |
471 | /// value split into legal parts. |
472 | static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc DL, |
473 | SDValue Val, SDValue *Parts, unsigned NumParts, |
474 | MVT PartVT, const Value *V) { |
475 | EVT ValueVT = Val.getValueType(); |
476 | assert(ValueVT.isVector() && "Not a vector")((ValueVT.isVector() && "Not a vector") ? static_cast <void> (0) : __assert_fail ("ValueVT.isVector() && \"Not a vector\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 476, __PRETTY_FUNCTION__)); |
477 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
478 | |
479 | if (NumParts == 1) { |
480 | EVT PartEVT = PartVT; |
481 | if (PartEVT == ValueVT) { |
482 | // Nothing to do. |
483 | } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) { |
484 | // Bitconvert vector->vector case. |
485 | Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val); |
486 | } else if (PartVT.isVector() && |
487 | PartEVT.getVectorElementType() == ValueVT.getVectorElementType() && |
488 | PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements()) { |
489 | EVT ElementVT = PartVT.getVectorElementType(); |
490 | // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in |
491 | // undef elements. |
492 | SmallVector<SDValue, 16> Ops; |
493 | for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i) |
494 | Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, |
495 | ElementVT, Val, DAG.getConstant(i, |
496 | TLI.getVectorIdxTy()))); |
497 | |
498 | for (unsigned i = ValueVT.getVectorNumElements(), |
499 | e = PartVT.getVectorNumElements(); i != e; ++i) |
500 | Ops.push_back(DAG.getUNDEF(ElementVT)); |
501 | |
502 | Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, Ops); |
503 | |
504 | // FIXME: Use CONCAT for 2x -> 4x. |
505 | |
506 | //SDValue UndefElts = DAG.getUNDEF(VectorTy); |
507 | //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts); |
508 | } else if (PartVT.isVector() && |
509 | PartEVT.getVectorElementType().bitsGE( |
510 | ValueVT.getVectorElementType()) && |
511 | PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements()) { |
512 | |
513 | // Promoted vector extract |
514 | bool Smaller = PartEVT.bitsLE(ValueVT); |
515 | Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND), |
516 | DL, PartVT, Val); |
517 | } else{ |
518 | // Vector -> scalar conversion. |
519 | assert(ValueVT.getVectorNumElements() == 1 &&((ValueVT.getVectorNumElements() == 1 && "Only trivial vector-to-scalar conversions should get here!" ) ? static_cast<void> (0) : __assert_fail ("ValueVT.getVectorNumElements() == 1 && \"Only trivial vector-to-scalar conversions should get here!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 520, __PRETTY_FUNCTION__)) |
520 | "Only trivial vector-to-scalar conversions should get here!")((ValueVT.getVectorNumElements() == 1 && "Only trivial vector-to-scalar conversions should get here!" ) ? static_cast<void> (0) : __assert_fail ("ValueVT.getVectorNumElements() == 1 && \"Only trivial vector-to-scalar conversions should get here!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 520, __PRETTY_FUNCTION__)); |
521 | Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, |
522 | PartVT, Val, DAG.getConstant(0, TLI.getVectorIdxTy())); |
523 | |
524 | bool Smaller = ValueVT.bitsLE(PartVT); |
525 | Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND), |
526 | DL, PartVT, Val); |
527 | } |
528 | |
529 | Parts[0] = Val; |
530 | return; |
531 | } |
532 | |
533 | // Handle a multi-element vector. |
534 | EVT IntermediateVT; |
535 | MVT RegisterVT; |
536 | unsigned NumIntermediates; |
537 | unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, |
538 | IntermediateVT, |
539 | NumIntermediates, RegisterVT); |
540 | unsigned NumElements = ValueVT.getVectorNumElements(); |
541 | |
542 | assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!")((NumRegs == NumParts && "Part count doesn't match vector breakdown!" ) ? static_cast<void> (0) : __assert_fail ("NumRegs == NumParts && \"Part count doesn't match vector breakdown!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 542, __PRETTY_FUNCTION__)); |
543 | NumParts = NumRegs; // Silence a compiler warning. |
544 | assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!")((RegisterVT == PartVT && "Part type doesn't match vector breakdown!" ) ? static_cast<void> (0) : __assert_fail ("RegisterVT == PartVT && \"Part type doesn't match vector breakdown!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 544, __PRETTY_FUNCTION__)); |
545 | |
546 | // Split the vector into intermediate operands. |
547 | SmallVector<SDValue, 8> Ops(NumIntermediates); |
548 | for (unsigned i = 0; i != NumIntermediates; ++i) { |
549 | if (IntermediateVT.isVector()) |
550 | Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, |
551 | IntermediateVT, Val, |
552 | DAG.getConstant(i * (NumElements / NumIntermediates), |
553 | TLI.getVectorIdxTy())); |
554 | else |
555 | Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, |
556 | IntermediateVT, Val, |
557 | DAG.getConstant(i, TLI.getVectorIdxTy())); |
558 | } |
559 | |
560 | // Split the intermediate operands into legal parts. |
561 | if (NumParts == NumIntermediates) { |
562 | // If the register was not expanded, promote or copy the value, |
563 | // as appropriate. |
564 | for (unsigned i = 0; i != NumParts; ++i) |
565 | getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT, V); |
566 | } else if (NumParts > 0) { |
567 | // If the intermediate type was expanded, split each the value into |
568 | // legal parts. |
569 | assert(NumParts % NumIntermediates == 0 &&((NumParts % NumIntermediates == 0 && "Must expand into a divisible number of parts!" ) ? static_cast<void> (0) : __assert_fail ("NumParts % NumIntermediates == 0 && \"Must expand into a divisible number of parts!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 570, __PRETTY_FUNCTION__)) |
570 | "Must expand into a divisible number of parts!")((NumParts % NumIntermediates == 0 && "Must expand into a divisible number of parts!" ) ? static_cast<void> (0) : __assert_fail ("NumParts % NumIntermediates == 0 && \"Must expand into a divisible number of parts!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 570, __PRETTY_FUNCTION__)); |
571 | unsigned Factor = NumParts / NumIntermediates; |
572 | for (unsigned i = 0; i != NumIntermediates; ++i) |
573 | getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT, V); |
574 | } |
575 | } |
576 | |
577 | namespace { |
578 | /// RegsForValue - This struct represents the registers (physical or virtual) |
579 | /// that a particular set of values is assigned, and the type information |
580 | /// about the value. The most common situation is to represent one value at a |
581 | /// time, but struct or array values are handled element-wise as multiple |
582 | /// values. The splitting of aggregates is performed recursively, so that we |
583 | /// never have aggregate-typed registers. The values at this point do not |
584 | /// necessarily have legal types, so each value may require one or more |
585 | /// registers of some legal type. |
586 | /// |
587 | struct RegsForValue { |
588 | /// ValueVTs - The value types of the values, which may not be legal, and |
589 | /// may need be promoted or synthesized from one or more registers. |
590 | /// |
591 | SmallVector<EVT, 4> ValueVTs; |
592 | |
593 | /// RegVTs - The value types of the registers. This is the same size as |
594 | /// ValueVTs and it records, for each value, what the type of the assigned |
595 | /// register or registers are. (Individual values are never synthesized |
596 | /// from more than one type of register.) |
597 | /// |
598 | /// With virtual registers, the contents of RegVTs is redundant with TLI's |
599 | /// getRegisterType member function, however when with physical registers |
600 | /// it is necessary to have a separate record of the types. |
601 | /// |
602 | SmallVector<MVT, 4> RegVTs; |
603 | |
604 | /// Regs - This list holds the registers assigned to the values. |
605 | /// Each legal or promoted value requires one register, and each |
606 | /// expanded value requires multiple registers. |
607 | /// |
608 | SmallVector<unsigned, 4> Regs; |
609 | |
610 | RegsForValue() {} |
611 | |
612 | RegsForValue(const SmallVector<unsigned, 4> ®s, |
613 | MVT regvt, EVT valuevt) |
614 | : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {} |
615 | |
616 | RegsForValue(LLVMContext &Context, const TargetLowering &tli, |
617 | unsigned Reg, Type *Ty) { |
618 | ComputeValueVTs(tli, Ty, ValueVTs); |
619 | |
620 | for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) { |
621 | EVT ValueVT = ValueVTs[Value]; |
622 | unsigned NumRegs = tli.getNumRegisters(Context, ValueVT); |
623 | MVT RegisterVT = tli.getRegisterType(Context, ValueVT); |
624 | for (unsigned i = 0; i != NumRegs; ++i) |
625 | Regs.push_back(Reg + i); |
626 | RegVTs.push_back(RegisterVT); |
627 | Reg += NumRegs; |
628 | } |
629 | } |
630 | |
631 | /// append - Add the specified values to this one. |
632 | void append(const RegsForValue &RHS) { |
633 | ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end()); |
634 | RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end()); |
635 | Regs.append(RHS.Regs.begin(), RHS.Regs.end()); |
636 | } |
637 | |
638 | /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from |
639 | /// this value and returns the result as a ValueVTs value. This uses |
640 | /// Chain/Flag as the input and updates them for the output Chain/Flag. |
641 | /// If the Flag pointer is NULL, no flag is used. |
642 | SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo, |
643 | SDLoc dl, |
644 | SDValue &Chain, SDValue *Flag, |
645 | const Value *V = nullptr) const; |
646 | |
647 | /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the |
648 | /// specified value into the registers specified by this object. This uses |
649 | /// Chain/Flag as the input and updates them for the output Chain/Flag. |
650 | /// If the Flag pointer is NULL, no flag is used. |
651 | void |
652 | getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl, SDValue &Chain, |
653 | SDValue *Flag, const Value *V, |
654 | ISD::NodeType PreferredExtendType = ISD::ANY_EXTEND) const; |
655 | |
656 | /// AddInlineAsmOperands - Add this value to the specified inlineasm node |
657 | /// operand list. This adds the code marker, matching input operand index |
658 | /// (if applicable), and includes the number of values added into it. |
659 | void AddInlineAsmOperands(unsigned Kind, |
660 | bool HasMatching, unsigned MatchingIdx, |
661 | SelectionDAG &DAG, |
662 | std::vector<SDValue> &Ops) const; |
663 | }; |
664 | } |
665 | |
666 | /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from |
667 | /// this value and returns the result as a ValueVT value. This uses |
668 | /// Chain/Flag as the input and updates them for the output Chain/Flag. |
669 | /// If the Flag pointer is NULL, no flag is used. |
670 | SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG, |
671 | FunctionLoweringInfo &FuncInfo, |
672 | SDLoc dl, |
673 | SDValue &Chain, SDValue *Flag, |
674 | const Value *V) const { |
675 | // A Value with type {} or [0 x %t] needs no registers. |
676 | if (ValueVTs.empty()) |
677 | return SDValue(); |
678 | |
679 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
680 | |
681 | // Assemble the legal parts into the final values. |
682 | SmallVector<SDValue, 4> Values(ValueVTs.size()); |
683 | SmallVector<SDValue, 8> Parts; |
684 | for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) { |
685 | // Copy the legal parts from the registers. |
686 | EVT ValueVT = ValueVTs[Value]; |
687 | unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT); |
688 | MVT RegisterVT = RegVTs[Value]; |
689 | |
690 | Parts.resize(NumRegs); |
691 | for (unsigned i = 0; i != NumRegs; ++i) { |
692 | SDValue P; |
693 | if (!Flag) { |
694 | P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT); |
695 | } else { |
696 | P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag); |
697 | *Flag = P.getValue(2); |
698 | } |
699 | |
700 | Chain = P.getValue(1); |
701 | Parts[i] = P; |
702 | |
703 | // If the source register was virtual and if we know something about it, |
704 | // add an assert node. |
705 | if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) || |
706 | !RegisterVT.isInteger() || RegisterVT.isVector()) |
707 | continue; |
708 | |
709 | const FunctionLoweringInfo::LiveOutInfo *LOI = |
710 | FuncInfo.GetLiveOutRegInfo(Regs[Part+i]); |
711 | if (!LOI) |
712 | continue; |
713 | |
714 | unsigned RegSize = RegisterVT.getSizeInBits(); |
715 | unsigned NumSignBits = LOI->NumSignBits; |
716 | unsigned NumZeroBits = LOI->KnownZero.countLeadingOnes(); |
717 | |
718 | if (NumZeroBits == RegSize) { |
719 | // The current value is a zero. |
720 | // Explicitly express that as it would be easier for |
721 | // optimizations to kick in. |
722 | Parts[i] = DAG.getConstant(0, RegisterVT); |
723 | continue; |
724 | } |
725 | |
726 | // FIXME: We capture more information than the dag can represent. For |
727 | // now, just use the tightest assertzext/assertsext possible. |
728 | bool isSExt = true; |
729 | EVT FromVT(MVT::Other); |
730 | if (NumSignBits == RegSize) |
731 | isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1 |
732 | else if (NumZeroBits >= RegSize-1) |
733 | isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1 |
734 | else if (NumSignBits > RegSize-8) |
735 | isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8 |
736 | else if (NumZeroBits >= RegSize-8) |
737 | isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8 |
738 | else if (NumSignBits > RegSize-16) |
739 | isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16 |
740 | else if (NumZeroBits >= RegSize-16) |
741 | isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16 |
742 | else if (NumSignBits > RegSize-32) |
743 | isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32 |
744 | else if (NumZeroBits >= RegSize-32) |
745 | isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32 |
746 | else |
747 | continue; |
748 | |
749 | // Add an assertion node. |
750 | assert(FromVT != MVT::Other)((FromVT != MVT::Other) ? static_cast<void> (0) : __assert_fail ("FromVT != MVT::Other", "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 750, __PRETTY_FUNCTION__)); |
751 | Parts[i] = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl, |
752 | RegisterVT, P, DAG.getValueType(FromVT)); |
753 | } |
754 | |
755 | Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(), |
756 | NumRegs, RegisterVT, ValueVT, V); |
757 | Part += NumRegs; |
758 | Parts.clear(); |
759 | } |
760 | |
761 | return DAG.getNode(ISD::MERGE_VALUES, dl, DAG.getVTList(ValueVTs), Values); |
762 | } |
763 | |
764 | /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the |
765 | /// specified value into the registers specified by this object. This uses |
766 | /// Chain/Flag as the input and updates them for the output Chain/Flag. |
767 | /// If the Flag pointer is NULL, no flag is used. |
768 | void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl, |
769 | SDValue &Chain, SDValue *Flag, const Value *V, |
770 | ISD::NodeType PreferredExtendType) const { |
771 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
772 | ISD::NodeType ExtendKind = PreferredExtendType; |
773 | |
774 | // Get the list of the values's legal parts. |
775 | unsigned NumRegs = Regs.size(); |
776 | SmallVector<SDValue, 8> Parts(NumRegs); |
777 | for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) { |
778 | EVT ValueVT = ValueVTs[Value]; |
779 | unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT); |
780 | MVT RegisterVT = RegVTs[Value]; |
781 | |
782 | if (ExtendKind == ISD::ANY_EXTEND && TLI.isZExtFree(Val, RegisterVT)) |
783 | ExtendKind = ISD::ZERO_EXTEND; |
784 | |
785 | getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value), |
786 | &Parts[Part], NumParts, RegisterVT, V, ExtendKind); |
787 | Part += NumParts; |
788 | } |
789 | |
790 | // Copy the parts into the registers. |
791 | SmallVector<SDValue, 8> Chains(NumRegs); |
792 | for (unsigned i = 0; i != NumRegs; ++i) { |
793 | SDValue Part; |
794 | if (!Flag) { |
795 | Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]); |
796 | } else { |
797 | Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag); |
798 | *Flag = Part.getValue(1); |
799 | } |
800 | |
801 | Chains[i] = Part.getValue(0); |
802 | } |
803 | |
804 | if (NumRegs == 1 || Flag) |
805 | // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is |
806 | // flagged to it. That is the CopyToReg nodes and the user are considered |
807 | // a single scheduling unit. If we create a TokenFactor and return it as |
808 | // chain, then the TokenFactor is both a predecessor (operand) of the |
809 | // user as well as a successor (the TF operands are flagged to the user). |
810 | // c1, f1 = CopyToReg |
811 | // c2, f2 = CopyToReg |
812 | // c3 = TokenFactor c1, c2 |
813 | // ... |
814 | // = op c3, ..., f2 |
815 | Chain = Chains[NumRegs-1]; |
816 | else |
817 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains); |
818 | } |
819 | |
820 | /// AddInlineAsmOperands - Add this value to the specified inlineasm node |
821 | /// operand list. This adds the code marker and includes the number of |
822 | /// values added into it. |
823 | void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching, |
824 | unsigned MatchingIdx, |
825 | SelectionDAG &DAG, |
826 | std::vector<SDValue> &Ops) const { |
827 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
828 | |
829 | unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size()); |
830 | if (HasMatching) |
831 | Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx); |
832 | else if (!Regs.empty() && |
833 | TargetRegisterInfo::isVirtualRegister(Regs.front())) { |
834 | // Put the register class of the virtual registers in the flag word. That |
835 | // way, later passes can recompute register class constraints for inline |
836 | // assembly as well as normal instructions. |
837 | // Don't do this for tied operands that can use the regclass information |
838 | // from the def. |
839 | const MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo(); |
840 | const TargetRegisterClass *RC = MRI.getRegClass(Regs.front()); |
841 | Flag = InlineAsm::getFlagWordForRegClass(Flag, RC->getID()); |
842 | } |
843 | |
844 | SDValue Res = DAG.getTargetConstant(Flag, MVT::i32); |
845 | Ops.push_back(Res); |
846 | |
847 | unsigned SP = TLI.getStackPointerRegisterToSaveRestore(); |
848 | for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) { |
849 | unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]); |
850 | MVT RegisterVT = RegVTs[Value]; |
851 | for (unsigned i = 0; i != NumRegs; ++i) { |
852 | assert(Reg < Regs.size() && "Mismatch in # registers expected")((Reg < Regs.size() && "Mismatch in # registers expected" ) ? static_cast<void> (0) : __assert_fail ("Reg < Regs.size() && \"Mismatch in # registers expected\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 852, __PRETTY_FUNCTION__)); |
853 | unsigned TheReg = Regs[Reg++]; |
854 | Ops.push_back(DAG.getRegister(TheReg, RegisterVT)); |
855 | |
856 | if (TheReg == SP && Code == InlineAsm::Kind_Clobber) { |
857 | // If we clobbered the stack pointer, MFI should know about it. |
858 | assert(DAG.getMachineFunction().getFrameInfo()->((DAG.getMachineFunction().getFrameInfo()-> hasInlineAsmWithSPAdjust ()) ? static_cast<void> (0) : __assert_fail ("DAG.getMachineFunction().getFrameInfo()-> hasInlineAsmWithSPAdjust()" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 859, __PRETTY_FUNCTION__)) |
859 | hasInlineAsmWithSPAdjust())((DAG.getMachineFunction().getFrameInfo()-> hasInlineAsmWithSPAdjust ()) ? static_cast<void> (0) : __assert_fail ("DAG.getMachineFunction().getFrameInfo()-> hasInlineAsmWithSPAdjust()" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 859, __PRETTY_FUNCTION__)); |
860 | } |
861 | } |
862 | } |
863 | } |
864 | |
865 | void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa, |
866 | const TargetLibraryInfo *li) { |
867 | AA = &aa; |
868 | GFI = gfi; |
869 | LibInfo = li; |
870 | DL = DAG.getSubtarget().getDataLayout(); |
871 | Context = DAG.getContext(); |
872 | LPadToCallSiteMap.clear(); |
873 | } |
874 | |
875 | /// clear - Clear out the current SelectionDAG and the associated |
876 | /// state and prepare this SelectionDAGBuilder object to be used |
877 | /// for a new block. This doesn't clear out information about |
878 | /// additional blocks that are needed to complete switch lowering |
879 | /// or PHI node updating; that information is cleared out as it is |
880 | /// consumed. |
881 | void SelectionDAGBuilder::clear() { |
882 | NodeMap.clear(); |
883 | UnusedArgNodeMap.clear(); |
884 | PendingLoads.clear(); |
885 | PendingExports.clear(); |
886 | CurInst = nullptr; |
887 | HasTailCall = false; |
888 | SDNodeOrder = LowestSDNodeOrder; |
889 | StatepointLowering.clear(); |
890 | } |
891 | |
892 | /// clearDanglingDebugInfo - Clear the dangling debug information |
893 | /// map. This function is separated from the clear so that debug |
894 | /// information that is dangling in a basic block can be properly |
895 | /// resolved in a different basic block. This allows the |
896 | /// SelectionDAG to resolve dangling debug information attached |
897 | /// to PHI nodes. |
898 | void SelectionDAGBuilder::clearDanglingDebugInfo() { |
899 | DanglingDebugInfoMap.clear(); |
900 | } |
901 | |
902 | /// getRoot - Return the current virtual root of the Selection DAG, |
903 | /// flushing any PendingLoad items. This must be done before emitting |
904 | /// a store or any other node that may need to be ordered after any |
905 | /// prior load instructions. |
906 | /// |
907 | SDValue SelectionDAGBuilder::getRoot() { |
908 | if (PendingLoads.empty()) |
909 | return DAG.getRoot(); |
910 | |
911 | if (PendingLoads.size() == 1) { |
912 | SDValue Root = PendingLoads[0]; |
913 | DAG.setRoot(Root); |
914 | PendingLoads.clear(); |
915 | return Root; |
916 | } |
917 | |
918 | // Otherwise, we have to make a token factor node. |
919 | SDValue Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, |
920 | PendingLoads); |
921 | PendingLoads.clear(); |
922 | DAG.setRoot(Root); |
923 | return Root; |
924 | } |
925 | |
926 | /// getControlRoot - Similar to getRoot, but instead of flushing all the |
927 | /// PendingLoad items, flush all the PendingExports items. It is necessary |
928 | /// to do this before emitting a terminator instruction. |
929 | /// |
930 | SDValue SelectionDAGBuilder::getControlRoot() { |
931 | SDValue Root = DAG.getRoot(); |
932 | |
933 | if (PendingExports.empty()) |
934 | return Root; |
935 | |
936 | // Turn all of the CopyToReg chains into one factored node. |
937 | if (Root.getOpcode() != ISD::EntryToken) { |
938 | unsigned i = 0, e = PendingExports.size(); |
939 | for (; i != e; ++i) { |
940 | assert(PendingExports[i].getNode()->getNumOperands() > 1)((PendingExports[i].getNode()->getNumOperands() > 1) ? static_cast <void> (0) : __assert_fail ("PendingExports[i].getNode()->getNumOperands() > 1" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 940, __PRETTY_FUNCTION__)); |
941 | if (PendingExports[i].getNode()->getOperand(0) == Root) |
942 | break; // Don't add the root if we already indirectly depend on it. |
943 | } |
944 | |
945 | if (i == e) |
946 | PendingExports.push_back(Root); |
947 | } |
948 | |
949 | Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, |
950 | PendingExports); |
951 | PendingExports.clear(); |
952 | DAG.setRoot(Root); |
953 | return Root; |
954 | } |
955 | |
956 | void SelectionDAGBuilder::visit(const Instruction &I) { |
957 | // Set up outgoing PHI node register values before emitting the terminator. |
958 | if (isa<TerminatorInst>(&I)) |
959 | HandlePHINodesInSuccessorBlocks(I.getParent()); |
960 | |
961 | ++SDNodeOrder; |
962 | |
963 | CurInst = &I; |
964 | |
965 | visit(I.getOpcode(), I); |
966 | |
967 | if (!isa<TerminatorInst>(&I) && !HasTailCall) |
968 | CopyToExportRegsIfNeeded(&I); |
969 | |
970 | CurInst = nullptr; |
971 | } |
972 | |
973 | void SelectionDAGBuilder::visitPHI(const PHINode &) { |
974 | llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!")::llvm::llvm_unreachable_internal("SelectionDAGBuilder shouldn't visit PHI nodes!" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 974); |
975 | } |
976 | |
977 | void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) { |
978 | // Note: this doesn't use InstVisitor, because it has to work with |
979 | // ConstantExpr's in addition to instructions. |
980 | switch (Opcode) { |
981 | default: llvm_unreachable("Unknown instruction type encountered!")::llvm::llvm_unreachable_internal("Unknown instruction type encountered!" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 981); |
982 | // Build the switch statement using the Instruction.def file. |
983 | #define HANDLE_INST(NUM, OPCODE, CLASS) \ |
984 | case Instruction::OPCODE: visit##OPCODE((const CLASS&)I); break; |
985 | #include "llvm/IR/Instruction.def" |
986 | } |
987 | } |
988 | |
989 | // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V, |
990 | // generate the debug data structures now that we've seen its definition. |
991 | void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V, |
992 | SDValue Val) { |
993 | DanglingDebugInfo &DDI = DanglingDebugInfoMap[V]; |
994 | if (DDI.getDI()) { |
995 | const DbgValueInst *DI = DDI.getDI(); |
996 | DebugLoc dl = DDI.getdl(); |
997 | unsigned DbgSDNodeOrder = DDI.getSDNodeOrder(); |
998 | MDNode *Variable = DI->getVariable(); |
999 | MDNode *Expr = DI->getExpression(); |
1000 | uint64_t Offset = DI->getOffset(); |
1001 | // A dbg.value for an alloca is always indirect. |
1002 | bool IsIndirect = isa<AllocaInst>(V) || Offset != 0; |
1003 | SDDbgValue *SDV; |
1004 | if (Val.getNode()) { |
1005 | if (!EmitFuncArgumentDbgValue(V, Variable, Expr, Offset, IsIndirect, |
1006 | Val)) { |
1007 | SDV = DAG.getDbgValue(Variable, Expr, Val.getNode(), Val.getResNo(), |
1008 | IsIndirect, Offset, dl, DbgSDNodeOrder); |
1009 | DAG.AddDbgValue(SDV, Val.getNode(), false); |
1010 | } |
1011 | } else |
1012 | DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Dropping debug info for " << *DI << "\n"; } } while (0); |
1013 | DanglingDebugInfoMap[V] = DanglingDebugInfo(); |
1014 | } |
1015 | } |
1016 | |
1017 | /// getValue - Return an SDValue for the given Value. |
1018 | SDValue SelectionDAGBuilder::getValue(const Value *V) { |
1019 | // If we already have an SDValue for this value, use it. It's important |
1020 | // to do this first, so that we don't create a CopyFromReg if we already |
1021 | // have a regular SDValue. |
1022 | SDValue &N = NodeMap[V]; |
1023 | if (N.getNode()) return N; |
1024 | |
1025 | // If there's a virtual register allocated and initialized for this |
1026 | // value, use it. |
1027 | DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V); |
1028 | if (It != FuncInfo.ValueMap.end()) { |
1029 | unsigned InReg = It->second; |
1030 | RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(), InReg, |
1031 | V->getType()); |
1032 | SDValue Chain = DAG.getEntryNode(); |
1033 | N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V); |
1034 | resolveDanglingDebugInfo(V, N); |
1035 | return N; |
1036 | } |
1037 | |
1038 | // Otherwise create a new SDValue and remember it. |
1039 | SDValue Val = getValueImpl(V); |
1040 | NodeMap[V] = Val; |
1041 | resolveDanglingDebugInfo(V, Val); |
1042 | return Val; |
1043 | } |
1044 | |
1045 | /// getNonRegisterValue - Return an SDValue for the given Value, but |
1046 | /// don't look in FuncInfo.ValueMap for a virtual register. |
1047 | SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) { |
1048 | // If we already have an SDValue for this value, use it. |
1049 | SDValue &N = NodeMap[V]; |
1050 | if (N.getNode()) return N; |
1051 | |
1052 | // Otherwise create a new SDValue and remember it. |
1053 | SDValue Val = getValueImpl(V); |
1054 | NodeMap[V] = Val; |
1055 | resolveDanglingDebugInfo(V, Val); |
1056 | return Val; |
1057 | } |
1058 | |
1059 | /// getValueImpl - Helper function for getValue and getNonRegisterValue. |
1060 | /// Create an SDValue for the given value. |
1061 | SDValue SelectionDAGBuilder::getValueImpl(const Value *V) { |
1062 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
1063 | |
1064 | if (const Constant *C = dyn_cast<Constant>(V)) { |
1065 | EVT VT = TLI.getValueType(V->getType(), true); |
1066 | |
1067 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(C)) |
1068 | return DAG.getConstant(*CI, VT); |
1069 | |
1070 | if (const GlobalValue *GV = dyn_cast<GlobalValue>(C)) |
1071 | return DAG.getGlobalAddress(GV, getCurSDLoc(), VT); |
1072 | |
1073 | if (isa<ConstantPointerNull>(C)) { |
1074 | unsigned AS = V->getType()->getPointerAddressSpace(); |
1075 | return DAG.getConstant(0, TLI.getPointerTy(AS)); |
1076 | } |
1077 | |
1078 | if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C)) |
1079 | return DAG.getConstantFP(*CFP, VT); |
1080 | |
1081 | if (isa<UndefValue>(C) && !V->getType()->isAggregateType()) |
1082 | return DAG.getUNDEF(VT); |
1083 | |
1084 | if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { |
1085 | visit(CE->getOpcode(), *CE); |
1086 | SDValue N1 = NodeMap[V]; |
1087 | assert(N1.getNode() && "visit didn't populate the NodeMap!")((N1.getNode() && "visit didn't populate the NodeMap!" ) ? static_cast<void> (0) : __assert_fail ("N1.getNode() && \"visit didn't populate the NodeMap!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1087, __PRETTY_FUNCTION__)); |
1088 | return N1; |
1089 | } |
1090 | |
1091 | if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) { |
1092 | SmallVector<SDValue, 4> Constants; |
1093 | for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end(); |
1094 | OI != OE; ++OI) { |
1095 | SDNode *Val = getValue(*OI).getNode(); |
1096 | // If the operand is an empty aggregate, there are no values. |
1097 | if (!Val) continue; |
1098 | // Add each leaf value from the operand to the Constants list |
1099 | // to form a flattened list of all the values. |
1100 | for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i) |
1101 | Constants.push_back(SDValue(Val, i)); |
1102 | } |
1103 | |
1104 | return DAG.getMergeValues(Constants, getCurSDLoc()); |
1105 | } |
1106 | |
1107 | if (const ConstantDataSequential *CDS = |
1108 | dyn_cast<ConstantDataSequential>(C)) { |
1109 | SmallVector<SDValue, 4> Ops; |
1110 | for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) { |
1111 | SDNode *Val = getValue(CDS->getElementAsConstant(i)).getNode(); |
1112 | // Add each leaf value from the operand to the Constants list |
1113 | // to form a flattened list of all the values. |
1114 | for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i) |
1115 | Ops.push_back(SDValue(Val, i)); |
1116 | } |
1117 | |
1118 | if (isa<ArrayType>(CDS->getType())) |
1119 | return DAG.getMergeValues(Ops, getCurSDLoc()); |
1120 | return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), |
1121 | VT, Ops); |
1122 | } |
1123 | |
1124 | if (C->getType()->isStructTy() || C->getType()->isArrayTy()) { |
1125 | assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&(((isa<ConstantAggregateZero>(C) || isa<UndefValue> (C)) && "Unknown struct or array constant!") ? static_cast <void> (0) : __assert_fail ("(isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) && \"Unknown struct or array constant!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1126, __PRETTY_FUNCTION__)) |
1126 | "Unknown struct or array constant!")(((isa<ConstantAggregateZero>(C) || isa<UndefValue> (C)) && "Unknown struct or array constant!") ? static_cast <void> (0) : __assert_fail ("(isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) && \"Unknown struct or array constant!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1126, __PRETTY_FUNCTION__)); |
1127 | |
1128 | SmallVector<EVT, 4> ValueVTs; |
1129 | ComputeValueVTs(TLI, C->getType(), ValueVTs); |
1130 | unsigned NumElts = ValueVTs.size(); |
1131 | if (NumElts == 0) |
1132 | return SDValue(); // empty struct |
1133 | SmallVector<SDValue, 4> Constants(NumElts); |
1134 | for (unsigned i = 0; i != NumElts; ++i) { |
1135 | EVT EltVT = ValueVTs[i]; |
1136 | if (isa<UndefValue>(C)) |
1137 | Constants[i] = DAG.getUNDEF(EltVT); |
1138 | else if (EltVT.isFloatingPoint()) |
1139 | Constants[i] = DAG.getConstantFP(0, EltVT); |
1140 | else |
1141 | Constants[i] = DAG.getConstant(0, EltVT); |
1142 | } |
1143 | |
1144 | return DAG.getMergeValues(Constants, getCurSDLoc()); |
1145 | } |
1146 | |
1147 | if (const BlockAddress *BA = dyn_cast<BlockAddress>(C)) |
1148 | return DAG.getBlockAddress(BA, VT); |
1149 | |
1150 | VectorType *VecTy = cast<VectorType>(V->getType()); |
1151 | unsigned NumElements = VecTy->getNumElements(); |
1152 | |
1153 | // Now that we know the number and type of the elements, get that number of |
1154 | // elements into the Ops array based on what kind of constant it is. |
1155 | SmallVector<SDValue, 16> Ops; |
1156 | if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) { |
1157 | for (unsigned i = 0; i != NumElements; ++i) |
1158 | Ops.push_back(getValue(CV->getOperand(i))); |
1159 | } else { |
1160 | assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!")((isa<ConstantAggregateZero>(C) && "Unknown vector constant!" ) ? static_cast<void> (0) : __assert_fail ("isa<ConstantAggregateZero>(C) && \"Unknown vector constant!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1160, __PRETTY_FUNCTION__)); |
1161 | EVT EltVT = TLI.getValueType(VecTy->getElementType()); |
1162 | |
1163 | SDValue Op; |
1164 | if (EltVT.isFloatingPoint()) |
1165 | Op = DAG.getConstantFP(0, EltVT); |
1166 | else |
1167 | Op = DAG.getConstant(0, EltVT); |
1168 | Ops.assign(NumElements, Op); |
1169 | } |
1170 | |
1171 | // Create a BUILD_VECTOR node. |
1172 | return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), VT, Ops); |
1173 | } |
1174 | |
1175 | // If this is a static alloca, generate it as the frameindex instead of |
1176 | // computation. |
1177 | if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) { |
1178 | DenseMap<const AllocaInst*, int>::iterator SI = |
1179 | FuncInfo.StaticAllocaMap.find(AI); |
1180 | if (SI != FuncInfo.StaticAllocaMap.end()) |
1181 | return DAG.getFrameIndex(SI->second, TLI.getPointerTy()); |
1182 | } |
1183 | |
1184 | // If this is an instruction which fast-isel has deferred, select it now. |
1185 | if (const Instruction *Inst = dyn_cast<Instruction>(V)) { |
1186 | unsigned InReg = FuncInfo.InitializeRegForValue(Inst); |
1187 | RegsForValue RFV(*DAG.getContext(), TLI, InReg, Inst->getType()); |
1188 | SDValue Chain = DAG.getEntryNode(); |
1189 | return RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V); |
1190 | } |
1191 | |
1192 | llvm_unreachable("Can't get register for value!")::llvm::llvm_unreachable_internal("Can't get register for value!" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1192); |
1193 | } |
1194 | |
1195 | void SelectionDAGBuilder::visitRet(const ReturnInst &I) { |
1196 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
1197 | SDValue Chain = getControlRoot(); |
1198 | SmallVector<ISD::OutputArg, 8> Outs; |
1199 | SmallVector<SDValue, 8> OutVals; |
1200 | |
1201 | if (!FuncInfo.CanLowerReturn) { |
1202 | unsigned DemoteReg = FuncInfo.DemoteRegister; |
1203 | const Function *F = I.getParent()->getParent(); |
1204 | |
1205 | // Emit a store of the return value through the virtual register. |
1206 | // Leave Outs empty so that LowerReturn won't try to load return |
1207 | // registers the usual way. |
1208 | SmallVector<EVT, 1> PtrValueVTs; |
1209 | ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()), |
1210 | PtrValueVTs); |
1211 | |
1212 | SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]); |
1213 | SDValue RetOp = getValue(I.getOperand(0)); |
1214 | |
1215 | SmallVector<EVT, 4> ValueVTs; |
1216 | SmallVector<uint64_t, 4> Offsets; |
1217 | ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets); |
1218 | unsigned NumValues = ValueVTs.size(); |
1219 | |
1220 | SmallVector<SDValue, 4> Chains(NumValues); |
1221 | for (unsigned i = 0; i != NumValues; ++i) { |
1222 | SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(), |
1223 | RetPtr.getValueType(), RetPtr, |
1224 | DAG.getIntPtrConstant(Offsets[i])); |
1225 | Chains[i] = |
1226 | DAG.getStore(Chain, getCurSDLoc(), |
1227 | SDValue(RetOp.getNode(), RetOp.getResNo() + i), |
1228 | // FIXME: better loc info would be nice. |
1229 | Add, MachinePointerInfo(), false, false, 0); |
1230 | } |
1231 | |
1232 | Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), |
1233 | MVT::Other, Chains); |
1234 | } else if (I.getNumOperands() != 0) { |
1235 | SmallVector<EVT, 4> ValueVTs; |
1236 | ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs); |
1237 | unsigned NumValues = ValueVTs.size(); |
1238 | if (NumValues) { |
1239 | SDValue RetOp = getValue(I.getOperand(0)); |
1240 | for (unsigned j = 0, f = NumValues; j != f; ++j) { |
1241 | EVT VT = ValueVTs[j]; |
1242 | |
1243 | ISD::NodeType ExtendKind = ISD::ANY_EXTEND; |
1244 | |
1245 | const Function *F = I.getParent()->getParent(); |
1246 | if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex, |
1247 | Attribute::SExt)) |
1248 | ExtendKind = ISD::SIGN_EXTEND; |
1249 | else if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex, |
1250 | Attribute::ZExt)) |
1251 | ExtendKind = ISD::ZERO_EXTEND; |
1252 | |
1253 | if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) |
1254 | VT = TLI.getTypeForExtArgOrReturn(*DAG.getContext(), VT, ExtendKind); |
1255 | |
1256 | unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), VT); |
1257 | MVT PartVT = TLI.getRegisterType(*DAG.getContext(), VT); |
1258 | SmallVector<SDValue, 4> Parts(NumParts); |
1259 | getCopyToParts(DAG, getCurSDLoc(), |
1260 | SDValue(RetOp.getNode(), RetOp.getResNo() + j), |
1261 | &Parts[0], NumParts, PartVT, &I, ExtendKind); |
1262 | |
1263 | // 'inreg' on function refers to return value |
1264 | ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); |
1265 | if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex, |
1266 | Attribute::InReg)) |
1267 | Flags.setInReg(); |
1268 | |
1269 | // Propagate extension type if any |
1270 | if (ExtendKind == ISD::SIGN_EXTEND) |
1271 | Flags.setSExt(); |
1272 | else if (ExtendKind == ISD::ZERO_EXTEND) |
1273 | Flags.setZExt(); |
1274 | |
1275 | for (unsigned i = 0; i < NumParts; ++i) { |
1276 | Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(), |
1277 | VT, /*isfixed=*/true, 0, 0)); |
1278 | OutVals.push_back(Parts[i]); |
1279 | } |
1280 | } |
1281 | } |
1282 | } |
1283 | |
1284 | bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg(); |
1285 | CallingConv::ID CallConv = |
1286 | DAG.getMachineFunction().getFunction()->getCallingConv(); |
1287 | Chain = DAG.getTargetLoweringInfo().LowerReturn( |
1288 | Chain, CallConv, isVarArg, Outs, OutVals, getCurSDLoc(), DAG); |
1289 | |
1290 | // Verify that the target's LowerReturn behaved as expected. |
1291 | assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&((Chain.getNode() && Chain.getValueType() == MVT::Other && "LowerReturn didn't return a valid chain!") ? static_cast <void> (0) : __assert_fail ("Chain.getNode() && Chain.getValueType() == MVT::Other && \"LowerReturn didn't return a valid chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1292, __PRETTY_FUNCTION__)) |
1292 | "LowerReturn didn't return a valid chain!")((Chain.getNode() && Chain.getValueType() == MVT::Other && "LowerReturn didn't return a valid chain!") ? static_cast <void> (0) : __assert_fail ("Chain.getNode() && Chain.getValueType() == MVT::Other && \"LowerReturn didn't return a valid chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1292, __PRETTY_FUNCTION__)); |
1293 | |
1294 | // Update the DAG with the new chain value resulting from return lowering. |
1295 | DAG.setRoot(Chain); |
1296 | } |
1297 | |
1298 | /// CopyToExportRegsIfNeeded - If the given value has virtual registers |
1299 | /// created for it, emit nodes to copy the value into the virtual |
1300 | /// registers. |
1301 | void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) { |
1302 | // Skip empty types |
1303 | if (V->getType()->isEmptyTy()) |
1304 | return; |
1305 | |
1306 | DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V); |
1307 | if (VMI != FuncInfo.ValueMap.end()) { |
1308 | assert(!V->use_empty() && "Unused value assigned virtual registers!")((!V->use_empty() && "Unused value assigned virtual registers!" ) ? static_cast<void> (0) : __assert_fail ("!V->use_empty() && \"Unused value assigned virtual registers!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1308, __PRETTY_FUNCTION__)); |
1309 | CopyValueToVirtualRegister(V, VMI->second); |
1310 | } |
1311 | } |
1312 | |
1313 | /// ExportFromCurrentBlock - If this condition isn't known to be exported from |
1314 | /// the current basic block, add it to ValueMap now so that we'll get a |
1315 | /// CopyTo/FromReg. |
1316 | void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) { |
1317 | // No need to export constants. |
1318 | if (!isa<Instruction>(V) && !isa<Argument>(V)) return; |
1319 | |
1320 | // Already exported? |
1321 | if (FuncInfo.isExportedInst(V)) return; |
1322 | |
1323 | unsigned Reg = FuncInfo.InitializeRegForValue(V); |
1324 | CopyValueToVirtualRegister(V, Reg); |
1325 | } |
1326 | |
1327 | bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V, |
1328 | const BasicBlock *FromBB) { |
1329 | // The operands of the setcc have to be in this block. We don't know |
1330 | // how to export them from some other block. |
1331 | if (const Instruction *VI = dyn_cast<Instruction>(V)) { |
1332 | // Can export from current BB. |
1333 | if (VI->getParent() == FromBB) |
1334 | return true; |
1335 | |
1336 | // Is already exported, noop. |
1337 | return FuncInfo.isExportedInst(V); |
1338 | } |
1339 | |
1340 | // If this is an argument, we can export it if the BB is the entry block or |
1341 | // if it is already exported. |
1342 | if (isa<Argument>(V)) { |
1343 | if (FromBB == &FromBB->getParent()->getEntryBlock()) |
1344 | return true; |
1345 | |
1346 | // Otherwise, can only export this if it is already exported. |
1347 | return FuncInfo.isExportedInst(V); |
1348 | } |
1349 | |
1350 | // Otherwise, constants can always be exported. |
1351 | return true; |
1352 | } |
1353 | |
1354 | /// Return branch probability calculated by BranchProbabilityInfo for IR blocks. |
1355 | uint32_t SelectionDAGBuilder::getEdgeWeight(const MachineBasicBlock *Src, |
1356 | const MachineBasicBlock *Dst) const { |
1357 | BranchProbabilityInfo *BPI = FuncInfo.BPI; |
1358 | if (!BPI) |
1359 | return 0; |
1360 | const BasicBlock *SrcBB = Src->getBasicBlock(); |
1361 | const BasicBlock *DstBB = Dst->getBasicBlock(); |
1362 | return BPI->getEdgeWeight(SrcBB, DstBB); |
1363 | } |
1364 | |
1365 | void SelectionDAGBuilder:: |
1366 | addSuccessorWithWeight(MachineBasicBlock *Src, MachineBasicBlock *Dst, |
1367 | uint32_t Weight /* = 0 */) { |
1368 | if (!Weight) |
1369 | Weight = getEdgeWeight(Src, Dst); |
1370 | Src->addSuccessor(Dst, Weight); |
1371 | } |
1372 | |
1373 | |
1374 | static bool InBlock(const Value *V, const BasicBlock *BB) { |
1375 | if (const Instruction *I = dyn_cast<Instruction>(V)) |
1376 | return I->getParent() == BB; |
1377 | return true; |
1378 | } |
1379 | |
1380 | /// EmitBranchForMergedCondition - Helper method for FindMergedConditions. |
1381 | /// This function emits a branch and is used at the leaves of an OR or an |
1382 | /// AND operator tree. |
1383 | /// |
1384 | void |
1385 | SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond, |
1386 | MachineBasicBlock *TBB, |
1387 | MachineBasicBlock *FBB, |
1388 | MachineBasicBlock *CurBB, |
1389 | MachineBasicBlock *SwitchBB, |
1390 | uint32_t TWeight, |
1391 | uint32_t FWeight) { |
1392 | const BasicBlock *BB = CurBB->getBasicBlock(); |
1393 | |
1394 | // If the leaf of the tree is a comparison, merge the condition into |
1395 | // the caseblock. |
1396 | if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) { |
1397 | // The operands of the cmp have to be in this block. We don't know |
1398 | // how to export them from some other block. If this is the first block |
1399 | // of the sequence, no exporting is needed. |
1400 | if (CurBB == SwitchBB || |
1401 | (isExportableFromCurrentBlock(BOp->getOperand(0), BB) && |
1402 | isExportableFromCurrentBlock(BOp->getOperand(1), BB))) { |
1403 | ISD::CondCode Condition; |
1404 | if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) { |
1405 | Condition = getICmpCondCode(IC->getPredicate()); |
1406 | } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) { |
1407 | Condition = getFCmpCondCode(FC->getPredicate()); |
1408 | if (TM.Options.NoNaNsFPMath) |
1409 | Condition = getFCmpCodeWithoutNaN(Condition); |
1410 | } else { |
1411 | Condition = ISD::SETEQ; // silence warning. |
Value stored to 'Condition' is never read | |
1412 | llvm_unreachable("Unknown compare instruction")::llvm::llvm_unreachable_internal("Unknown compare instruction" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1412); |
1413 | } |
1414 | |
1415 | CaseBlock CB(Condition, BOp->getOperand(0), BOp->getOperand(1), nullptr, |
1416 | TBB, FBB, CurBB, TWeight, FWeight); |
1417 | SwitchCases.push_back(CB); |
1418 | return; |
1419 | } |
1420 | } |
1421 | |
1422 | // Create a CaseBlock record representing this branch. |
1423 | CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()), |
1424 | nullptr, TBB, FBB, CurBB, TWeight, FWeight); |
1425 | SwitchCases.push_back(CB); |
1426 | } |
1427 | |
1428 | /// Scale down both weights to fit into uint32_t. |
1429 | static void ScaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) { |
1430 | uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse; |
1431 | uint32_t Scale = (NewMax / UINT32_MAX(4294967295U)) + 1; |
1432 | NewTrue = NewTrue / Scale; |
1433 | NewFalse = NewFalse / Scale; |
1434 | } |
1435 | |
1436 | /// FindMergedConditions - If Cond is an expression like |
1437 | void SelectionDAGBuilder::FindMergedConditions(const Value *Cond, |
1438 | MachineBasicBlock *TBB, |
1439 | MachineBasicBlock *FBB, |
1440 | MachineBasicBlock *CurBB, |
1441 | MachineBasicBlock *SwitchBB, |
1442 | unsigned Opc, uint32_t TWeight, |
1443 | uint32_t FWeight) { |
1444 | // If this node is not part of the or/and tree, emit it as a branch. |
1445 | const Instruction *BOp = dyn_cast<Instruction>(Cond); |
1446 | if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) || |
1447 | (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() || |
1448 | BOp->getParent() != CurBB->getBasicBlock() || |
1449 | !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) || |
1450 | !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) { |
1451 | EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB, |
1452 | TWeight, FWeight); |
1453 | return; |
1454 | } |
1455 | |
1456 | // Create TmpBB after CurBB. |
1457 | MachineFunction::iterator BBI = CurBB; |
1458 | MachineFunction &MF = DAG.getMachineFunction(); |
1459 | MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock()); |
1460 | CurBB->getParent()->insert(++BBI, TmpBB); |
1461 | |
1462 | if (Opc == Instruction::Or) { |
1463 | // Codegen X | Y as: |
1464 | // BB1: |
1465 | // jmp_if_X TBB |
1466 | // jmp TmpBB |
1467 | // TmpBB: |
1468 | // jmp_if_Y TBB |
1469 | // jmp FBB |
1470 | // |
1471 | |
1472 | // We have flexibility in setting Prob for BB1 and Prob for TmpBB. |
1473 | // The requirement is that |
1474 | // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB) |
1475 | // = TrueProb for orignal BB. |
1476 | // Assuming the orignal weights are A and B, one choice is to set BB1's |
1477 | // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice |
1478 | // assumes that |
1479 | // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB. |
1480 | // Another choice is to assume TrueProb for BB1 equals to TrueProb for |
1481 | // TmpBB, but the math is more complicated. |
1482 | |
1483 | uint64_t NewTrueWeight = TWeight; |
1484 | uint64_t NewFalseWeight = (uint64_t)TWeight + 2 * (uint64_t)FWeight; |
1485 | ScaleWeights(NewTrueWeight, NewFalseWeight); |
1486 | // Emit the LHS condition. |
1487 | FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc, |
1488 | NewTrueWeight, NewFalseWeight); |
1489 | |
1490 | NewTrueWeight = TWeight; |
1491 | NewFalseWeight = 2 * (uint64_t)FWeight; |
1492 | ScaleWeights(NewTrueWeight, NewFalseWeight); |
1493 | // Emit the RHS condition into TmpBB. |
1494 | FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc, |
1495 | NewTrueWeight, NewFalseWeight); |
1496 | } else { |
1497 | assert(Opc == Instruction::And && "Unknown merge op!")((Opc == Instruction::And && "Unknown merge op!") ? static_cast <void> (0) : __assert_fail ("Opc == Instruction::And && \"Unknown merge op!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1497, __PRETTY_FUNCTION__)); |
1498 | // Codegen X & Y as: |
1499 | // BB1: |
1500 | // jmp_if_X TmpBB |
1501 | // jmp FBB |
1502 | // TmpBB: |
1503 | // jmp_if_Y TBB |
1504 | // jmp FBB |
1505 | // |
1506 | // This requires creation of TmpBB after CurBB. |
1507 | |
1508 | // We have flexibility in setting Prob for BB1 and Prob for TmpBB. |
1509 | // The requirement is that |
1510 | // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB) |
1511 | // = FalseProb for orignal BB. |
1512 | // Assuming the orignal weights are A and B, one choice is to set BB1's |
1513 | // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice |
1514 | // assumes that |
1515 | // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB. |
1516 | |
1517 | uint64_t NewTrueWeight = 2 * (uint64_t)TWeight + (uint64_t)FWeight; |
1518 | uint64_t NewFalseWeight = FWeight; |
1519 | ScaleWeights(NewTrueWeight, NewFalseWeight); |
1520 | // Emit the LHS condition. |
1521 | FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc, |
1522 | NewTrueWeight, NewFalseWeight); |
1523 | |
1524 | NewTrueWeight = 2 * (uint64_t)TWeight; |
1525 | NewFalseWeight = FWeight; |
1526 | ScaleWeights(NewTrueWeight, NewFalseWeight); |
1527 | // Emit the RHS condition into TmpBB. |
1528 | FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc, |
1529 | NewTrueWeight, NewFalseWeight); |
1530 | } |
1531 | } |
1532 | |
1533 | /// If the set of cases should be emitted as a series of branches, return true. |
1534 | /// If we should emit this as a bunch of and/or'd together conditions, return |
1535 | /// false. |
1536 | bool |
1537 | SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases) { |
1538 | if (Cases.size() != 2) return true; |
1539 | |
1540 | // If this is two comparisons of the same values or'd or and'd together, they |
1541 | // will get folded into a single comparison, so don't emit two blocks. |
1542 | if ((Cases[0].CmpLHS == Cases[1].CmpLHS && |
1543 | Cases[0].CmpRHS == Cases[1].CmpRHS) || |
1544 | (Cases[0].CmpRHS == Cases[1].CmpLHS && |
1545 | Cases[0].CmpLHS == Cases[1].CmpRHS)) { |
1546 | return false; |
1547 | } |
1548 | |
1549 | // Handle: (X != null) | (Y != null) --> (X|Y) != 0 |
1550 | // Handle: (X == null) & (Y == null) --> (X|Y) == 0 |
1551 | if (Cases[0].CmpRHS == Cases[1].CmpRHS && |
1552 | Cases[0].CC == Cases[1].CC && |
1553 | isa<Constant>(Cases[0].CmpRHS) && |
1554 | cast<Constant>(Cases[0].CmpRHS)->isNullValue()) { |
1555 | if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB) |
1556 | return false; |
1557 | if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB) |
1558 | return false; |
1559 | } |
1560 | |
1561 | return true; |
1562 | } |
1563 | |
1564 | void SelectionDAGBuilder::visitBr(const BranchInst &I) { |
1565 | MachineBasicBlock *BrMBB = FuncInfo.MBB; |
1566 | |
1567 | // Update machine-CFG edges. |
1568 | MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)]; |
1569 | |
1570 | // Figure out which block is immediately after the current one. |
1571 | MachineBasicBlock *NextBlock = nullptr; |
1572 | MachineFunction::iterator BBI = BrMBB; |
1573 | if (++BBI != FuncInfo.MF->end()) |
1574 | NextBlock = BBI; |
1575 | |
1576 | if (I.isUnconditional()) { |
1577 | // Update machine-CFG edges. |
1578 | BrMBB->addSuccessor(Succ0MBB); |
1579 | |
1580 | // If this is not a fall-through branch or optimizations are switched off, |
1581 | // emit the branch. |
1582 | if (Succ0MBB != NextBlock || TM.getOptLevel() == CodeGenOpt::None) |
1583 | DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), |
1584 | MVT::Other, getControlRoot(), |
1585 | DAG.getBasicBlock(Succ0MBB))); |
1586 | |
1587 | return; |
1588 | } |
1589 | |
1590 | // If this condition is one of the special cases we handle, do special stuff |
1591 | // now. |
1592 | const Value *CondVal = I.getCondition(); |
1593 | MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)]; |
1594 | |
1595 | // If this is a series of conditions that are or'd or and'd together, emit |
1596 | // this as a sequence of branches instead of setcc's with and/or operations. |
1597 | // As long as jumps are not expensive, this should improve performance. |
1598 | // For example, instead of something like: |
1599 | // cmp A, B |
1600 | // C = seteq |
1601 | // cmp D, E |
1602 | // F = setle |
1603 | // or C, F |
1604 | // jnz foo |
1605 | // Emit: |
1606 | // cmp A, B |
1607 | // je foo |
1608 | // cmp D, E |
1609 | // jle foo |
1610 | // |
1611 | if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) { |
1612 | if (!DAG.getTargetLoweringInfo().isJumpExpensive() && |
1613 | BOp->hasOneUse() && (BOp->getOpcode() == Instruction::And || |
1614 | BOp->getOpcode() == Instruction::Or)) { |
1615 | FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB, |
1616 | BOp->getOpcode(), getEdgeWeight(BrMBB, Succ0MBB), |
1617 | getEdgeWeight(BrMBB, Succ1MBB)); |
1618 | // If the compares in later blocks need to use values not currently |
1619 | // exported from this block, export them now. This block should always |
1620 | // be the first entry. |
1621 | assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!")((SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!" ) ? static_cast<void> (0) : __assert_fail ("SwitchCases[0].ThisBB == BrMBB && \"Unexpected lowering!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1621, __PRETTY_FUNCTION__)); |
1622 | |
1623 | // Allow some cases to be rejected. |
1624 | if (ShouldEmitAsBranches(SwitchCases)) { |
1625 | for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) { |
1626 | ExportFromCurrentBlock(SwitchCases[i].CmpLHS); |
1627 | ExportFromCurrentBlock(SwitchCases[i].CmpRHS); |
1628 | } |
1629 | |
1630 | // Emit the branch for this block. |
1631 | visitSwitchCase(SwitchCases[0], BrMBB); |
1632 | SwitchCases.erase(SwitchCases.begin()); |
1633 | return; |
1634 | } |
1635 | |
1636 | // Okay, we decided not to do this, remove any inserted MBB's and clear |
1637 | // SwitchCases. |
1638 | for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) |
1639 | FuncInfo.MF->erase(SwitchCases[i].ThisBB); |
1640 | |
1641 | SwitchCases.clear(); |
1642 | } |
1643 | } |
1644 | |
1645 | // Create a CaseBlock record representing this branch. |
1646 | CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()), |
1647 | nullptr, Succ0MBB, Succ1MBB, BrMBB); |
1648 | |
1649 | // Use visitSwitchCase to actually insert the fast branch sequence for this |
1650 | // cond branch. |
1651 | visitSwitchCase(CB, BrMBB); |
1652 | } |
1653 | |
1654 | /// visitSwitchCase - Emits the necessary code to represent a single node in |
1655 | /// the binary search tree resulting from lowering a switch instruction. |
1656 | void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB, |
1657 | MachineBasicBlock *SwitchBB) { |
1658 | SDValue Cond; |
1659 | SDValue CondLHS = getValue(CB.CmpLHS); |
1660 | SDLoc dl = getCurSDLoc(); |
1661 | |
1662 | // Build the setcc now. |
1663 | if (!CB.CmpMHS) { |
1664 | // Fold "(X == true)" to X and "(X == false)" to !X to |
1665 | // handle common cases produced by branch lowering. |
1666 | if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) && |
1667 | CB.CC == ISD::SETEQ) |
1668 | Cond = CondLHS; |
1669 | else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) && |
1670 | CB.CC == ISD::SETEQ) { |
1671 | SDValue True = DAG.getConstant(1, CondLHS.getValueType()); |
1672 | Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True); |
1673 | } else |
1674 | Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC); |
1675 | } else { |
1676 | assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now")((CB.CC == ISD::SETLE && "Can handle only LE ranges now" ) ? static_cast<void> (0) : __assert_fail ("CB.CC == ISD::SETLE && \"Can handle only LE ranges now\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1676, __PRETTY_FUNCTION__)); |
1677 | |
1678 | const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue(); |
1679 | const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue(); |
1680 | |
1681 | SDValue CmpOp = getValue(CB.CmpMHS); |
1682 | EVT VT = CmpOp.getValueType(); |
1683 | |
1684 | if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) { |
1685 | Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT), |
1686 | ISD::SETLE); |
1687 | } else { |
1688 | SDValue SUB = DAG.getNode(ISD::SUB, dl, |
1689 | VT, CmpOp, DAG.getConstant(Low, VT)); |
1690 | Cond = DAG.getSetCC(dl, MVT::i1, SUB, |
1691 | DAG.getConstant(High-Low, VT), ISD::SETULE); |
1692 | } |
1693 | } |
1694 | |
1695 | // Update successor info |
1696 | addSuccessorWithWeight(SwitchBB, CB.TrueBB, CB.TrueWeight); |
1697 | // TrueBB and FalseBB are always different unless the incoming IR is |
1698 | // degenerate. This only happens when running llc on weird IR. |
1699 | if (CB.TrueBB != CB.FalseBB) |
1700 | addSuccessorWithWeight(SwitchBB, CB.FalseBB, CB.FalseWeight); |
1701 | |
1702 | // Set NextBlock to be the MBB immediately after the current one, if any. |
1703 | // This is used to avoid emitting unnecessary branches to the next block. |
1704 | MachineBasicBlock *NextBlock = nullptr; |
1705 | MachineFunction::iterator BBI = SwitchBB; |
1706 | if (++BBI != FuncInfo.MF->end()) |
1707 | NextBlock = BBI; |
1708 | |
1709 | // If the lhs block is the next block, invert the condition so that we can |
1710 | // fall through to the lhs instead of the rhs block. |
1711 | if (CB.TrueBB == NextBlock) { |
1712 | std::swap(CB.TrueBB, CB.FalseBB); |
1713 | SDValue True = DAG.getConstant(1, Cond.getValueType()); |
1714 | Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True); |
1715 | } |
1716 | |
1717 | SDValue BrCond = DAG.getNode(ISD::BRCOND, dl, |
1718 | MVT::Other, getControlRoot(), Cond, |
1719 | DAG.getBasicBlock(CB.TrueBB)); |
1720 | |
1721 | // Insert the false branch. Do this even if it's a fall through branch, |
1722 | // this makes it easier to do DAG optimizations which require inverting |
1723 | // the branch condition. |
1724 | BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond, |
1725 | DAG.getBasicBlock(CB.FalseBB)); |
1726 | |
1727 | DAG.setRoot(BrCond); |
1728 | } |
1729 | |
1730 | /// visitJumpTable - Emit JumpTable node in the current MBB |
1731 | void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) { |
1732 | // Emit the code for the jump table |
1733 | assert(JT.Reg != -1U && "Should lower JT Header first!")((JT.Reg != -1U && "Should lower JT Header first!") ? static_cast<void> (0) : __assert_fail ("JT.Reg != -1U && \"Should lower JT Header first!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1733, __PRETTY_FUNCTION__)); |
1734 | EVT PTy = DAG.getTargetLoweringInfo().getPointerTy(); |
1735 | SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(), |
1736 | JT.Reg, PTy); |
1737 | SDValue Table = DAG.getJumpTable(JT.JTI, PTy); |
1738 | SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurSDLoc(), |
1739 | MVT::Other, Index.getValue(1), |
1740 | Table, Index); |
1741 | DAG.setRoot(BrJumpTable); |
1742 | } |
1743 | |
1744 | /// visitJumpTableHeader - This function emits necessary code to produce index |
1745 | /// in the JumpTable from switch case. |
1746 | void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT, |
1747 | JumpTableHeader &JTH, |
1748 | MachineBasicBlock *SwitchBB) { |
1749 | // Subtract the lowest switch case value from the value being switched on and |
1750 | // conditional branch to default mbb if the result is greater than the |
1751 | // difference between smallest and largest cases. |
1752 | SDValue SwitchOp = getValue(JTH.SValue); |
1753 | EVT VT = SwitchOp.getValueType(); |
1754 | SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, SwitchOp, |
1755 | DAG.getConstant(JTH.First, VT)); |
1756 | |
1757 | // The SDNode we just created, which holds the value being switched on minus |
1758 | // the smallest case value, needs to be copied to a virtual register so it |
1759 | // can be used as an index into the jump table in a subsequent basic block. |
1760 | // This value may be smaller or larger than the target's pointer type, and |
1761 | // therefore require extension or truncating. |
1762 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
1763 | SwitchOp = DAG.getZExtOrTrunc(Sub, getCurSDLoc(), TLI.getPointerTy()); |
1764 | |
1765 | unsigned JumpTableReg = FuncInfo.CreateReg(TLI.getPointerTy()); |
1766 | SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurSDLoc(), |
1767 | JumpTableReg, SwitchOp); |
1768 | JT.Reg = JumpTableReg; |
1769 | |
1770 | // Emit the range check for the jump table, and branch to the default block |
1771 | // for the switch statement if the value being switched on exceeds the largest |
1772 | // case in the switch. |
1773 | SDValue CMP = |
1774 | DAG.getSetCC(getCurSDLoc(), TLI.getSetCCResultType(*DAG.getContext(), |
1775 | Sub.getValueType()), |
1776 | Sub, DAG.getConstant(JTH.Last - JTH.First, VT), ISD::SETUGT); |
1777 | |
1778 | // Set NextBlock to be the MBB immediately after the current one, if any. |
1779 | // This is used to avoid emitting unnecessary branches to the next block. |
1780 | MachineBasicBlock *NextBlock = nullptr; |
1781 | MachineFunction::iterator BBI = SwitchBB; |
1782 | |
1783 | if (++BBI != FuncInfo.MF->end()) |
1784 | NextBlock = BBI; |
1785 | |
1786 | SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurSDLoc(), |
1787 | MVT::Other, CopyTo, CMP, |
1788 | DAG.getBasicBlock(JT.Default)); |
1789 | |
1790 | if (JT.MBB != NextBlock) |
1791 | BrCond = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, BrCond, |
1792 | DAG.getBasicBlock(JT.MBB)); |
1793 | |
1794 | DAG.setRoot(BrCond); |
1795 | } |
1796 | |
1797 | /// Codegen a new tail for a stack protector check ParentMBB which has had its |
1798 | /// tail spliced into a stack protector check success bb. |
1799 | /// |
1800 | /// For a high level explanation of how this fits into the stack protector |
1801 | /// generation see the comment on the declaration of class |
1802 | /// StackProtectorDescriptor. |
1803 | void SelectionDAGBuilder::visitSPDescriptorParent(StackProtectorDescriptor &SPD, |
1804 | MachineBasicBlock *ParentBB) { |
1805 | |
1806 | // First create the loads to the guard/stack slot for the comparison. |
1807 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
1808 | EVT PtrTy = TLI.getPointerTy(); |
1809 | |
1810 | MachineFrameInfo *MFI = ParentBB->getParent()->getFrameInfo(); |
1811 | int FI = MFI->getStackProtectorIndex(); |
1812 | |
1813 | const Value *IRGuard = SPD.getGuard(); |
1814 | SDValue GuardPtr = getValue(IRGuard); |
1815 | SDValue StackSlotPtr = DAG.getFrameIndex(FI, PtrTy); |
1816 | |
1817 | unsigned Align = |
1818 | TLI.getDataLayout()->getPrefTypeAlignment(IRGuard->getType()); |
1819 | |
1820 | SDValue Guard; |
1821 | |
1822 | // If GuardReg is set and useLoadStackGuardNode returns true, retrieve the |
1823 | // guard value from the virtual register holding the value. Otherwise, emit a |
1824 | // volatile load to retrieve the stack guard value. |
1825 | unsigned GuardReg = SPD.getGuardReg(); |
1826 | |
1827 | if (GuardReg && TLI.useLoadStackGuardNode()) |
1828 | Guard = DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(), GuardReg, |
1829 | PtrTy); |
1830 | else |
1831 | Guard = DAG.getLoad(PtrTy, getCurSDLoc(), DAG.getEntryNode(), |
1832 | GuardPtr, MachinePointerInfo(IRGuard, 0), |
1833 | true, false, false, Align); |
1834 | |
1835 | SDValue StackSlot = DAG.getLoad(PtrTy, getCurSDLoc(), DAG.getEntryNode(), |
1836 | StackSlotPtr, |
1837 | MachinePointerInfo::getFixedStack(FI), |
1838 | true, false, false, Align); |
1839 | |
1840 | // Perform the comparison via a subtract/getsetcc. |
1841 | EVT VT = Guard.getValueType(); |
1842 | SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, Guard, StackSlot); |
1843 | |
1844 | SDValue Cmp = |
1845 | DAG.getSetCC(getCurSDLoc(), TLI.getSetCCResultType(*DAG.getContext(), |
1846 | Sub.getValueType()), |
1847 | Sub, DAG.getConstant(0, VT), ISD::SETNE); |
1848 | |
1849 | // If the sub is not 0, then we know the guard/stackslot do not equal, so |
1850 | // branch to failure MBB. |
1851 | SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurSDLoc(), |
1852 | MVT::Other, StackSlot.getOperand(0), |
1853 | Cmp, DAG.getBasicBlock(SPD.getFailureMBB())); |
1854 | // Otherwise branch to success MBB. |
1855 | SDValue Br = DAG.getNode(ISD::BR, getCurSDLoc(), |
1856 | MVT::Other, BrCond, |
1857 | DAG.getBasicBlock(SPD.getSuccessMBB())); |
1858 | |
1859 | DAG.setRoot(Br); |
1860 | } |
1861 | |
1862 | /// Codegen the failure basic block for a stack protector check. |
1863 | /// |
1864 | /// A failure stack protector machine basic block consists simply of a call to |
1865 | /// __stack_chk_fail(). |
1866 | /// |
1867 | /// For a high level explanation of how this fits into the stack protector |
1868 | /// generation see the comment on the declaration of class |
1869 | /// StackProtectorDescriptor. |
1870 | void |
1871 | SelectionDAGBuilder::visitSPDescriptorFailure(StackProtectorDescriptor &SPD) { |
1872 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
1873 | SDValue Chain = |
1874 | TLI.makeLibCall(DAG, RTLIB::STACKPROTECTOR_CHECK_FAIL, MVT::isVoid, |
1875 | nullptr, 0, false, getCurSDLoc(), false, false).second; |
1876 | DAG.setRoot(Chain); |
1877 | } |
1878 | |
1879 | /// visitBitTestHeader - This function emits necessary code to produce value |
1880 | /// suitable for "bit tests" |
1881 | void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B, |
1882 | MachineBasicBlock *SwitchBB) { |
1883 | // Subtract the minimum value |
1884 | SDValue SwitchOp = getValue(B.SValue); |
1885 | EVT VT = SwitchOp.getValueType(); |
1886 | SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, SwitchOp, |
1887 | DAG.getConstant(B.First, VT)); |
1888 | |
1889 | // Check range |
1890 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
1891 | SDValue RangeCmp = |
1892 | DAG.getSetCC(getCurSDLoc(), TLI.getSetCCResultType(*DAG.getContext(), |
1893 | Sub.getValueType()), |
1894 | Sub, DAG.getConstant(B.Range, VT), ISD::SETUGT); |
1895 | |
1896 | // Determine the type of the test operands. |
1897 | bool UsePtrType = false; |
1898 | if (!TLI.isTypeLegal(VT)) |
1899 | UsePtrType = true; |
1900 | else { |
1901 | for (unsigned i = 0, e = B.Cases.size(); i != e; ++i) |
1902 | if (!isUIntN(VT.getSizeInBits(), B.Cases[i].Mask)) { |
1903 | // Switch table case range are encoded into series of masks. |
1904 | // Just use pointer type, it's guaranteed to fit. |
1905 | UsePtrType = true; |
1906 | break; |
1907 | } |
1908 | } |
1909 | if (UsePtrType) { |
1910 | VT = TLI.getPointerTy(); |
1911 | Sub = DAG.getZExtOrTrunc(Sub, getCurSDLoc(), VT); |
1912 | } |
1913 | |
1914 | B.RegVT = VT.getSimpleVT(); |
1915 | B.Reg = FuncInfo.CreateReg(B.RegVT); |
1916 | SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurSDLoc(), |
1917 | B.Reg, Sub); |
1918 | |
1919 | // Set NextBlock to be the MBB immediately after the current one, if any. |
1920 | // This is used to avoid emitting unnecessary branches to the next block. |
1921 | MachineBasicBlock *NextBlock = nullptr; |
1922 | MachineFunction::iterator BBI = SwitchBB; |
1923 | if (++BBI != FuncInfo.MF->end()) |
1924 | NextBlock = BBI; |
1925 | |
1926 | MachineBasicBlock* MBB = B.Cases[0].ThisBB; |
1927 | |
1928 | addSuccessorWithWeight(SwitchBB, B.Default); |
1929 | addSuccessorWithWeight(SwitchBB, MBB); |
1930 | |
1931 | SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurSDLoc(), |
1932 | MVT::Other, CopyTo, RangeCmp, |
1933 | DAG.getBasicBlock(B.Default)); |
1934 | |
1935 | if (MBB != NextBlock) |
1936 | BrRange = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, CopyTo, |
1937 | DAG.getBasicBlock(MBB)); |
1938 | |
1939 | DAG.setRoot(BrRange); |
1940 | } |
1941 | |
1942 | /// visitBitTestCase - this function produces one "bit test" |
1943 | void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB, |
1944 | MachineBasicBlock* NextMBB, |
1945 | uint32_t BranchWeightToNext, |
1946 | unsigned Reg, |
1947 | BitTestCase &B, |
1948 | MachineBasicBlock *SwitchBB) { |
1949 | MVT VT = BB.RegVT; |
1950 | SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(), |
1951 | Reg, VT); |
1952 | SDValue Cmp; |
1953 | unsigned PopCount = CountPopulation_64(B.Mask); |
1954 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
1955 | if (PopCount == 1) { |
1956 | // Testing for a single bit; just compare the shift count with what it |
1957 | // would need to be to shift a 1 bit in that position. |
1958 | Cmp = DAG.getSetCC( |
1959 | getCurSDLoc(), TLI.getSetCCResultType(*DAG.getContext(), VT), ShiftOp, |
1960 | DAG.getConstant(countTrailingZeros(B.Mask), VT), ISD::SETEQ); |
1961 | } else if (PopCount == BB.Range) { |
1962 | // There is only one zero bit in the range, test for it directly. |
1963 | Cmp = DAG.getSetCC( |
1964 | getCurSDLoc(), TLI.getSetCCResultType(*DAG.getContext(), VT), ShiftOp, |
1965 | DAG.getConstant(CountTrailingOnes_64(B.Mask), VT), ISD::SETNE); |
1966 | } else { |
1967 | // Make desired shift |
1968 | SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurSDLoc(), VT, |
1969 | DAG.getConstant(1, VT), ShiftOp); |
1970 | |
1971 | // Emit bit tests and jumps |
1972 | SDValue AndOp = DAG.getNode(ISD::AND, getCurSDLoc(), |
1973 | VT, SwitchVal, DAG.getConstant(B.Mask, VT)); |
1974 | Cmp = DAG.getSetCC(getCurSDLoc(), |
1975 | TLI.getSetCCResultType(*DAG.getContext(), VT), AndOp, |
1976 | DAG.getConstant(0, VT), ISD::SETNE); |
1977 | } |
1978 | |
1979 | // The branch weight from SwitchBB to B.TargetBB is B.ExtraWeight. |
1980 | addSuccessorWithWeight(SwitchBB, B.TargetBB, B.ExtraWeight); |
1981 | // The branch weight from SwitchBB to NextMBB is BranchWeightToNext. |
1982 | addSuccessorWithWeight(SwitchBB, NextMBB, BranchWeightToNext); |
1983 | |
1984 | SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurSDLoc(), |
1985 | MVT::Other, getControlRoot(), |
1986 | Cmp, DAG.getBasicBlock(B.TargetBB)); |
1987 | |
1988 | // Set NextBlock to be the MBB immediately after the current one, if any. |
1989 | // This is used to avoid emitting unnecessary branches to the next block. |
1990 | MachineBasicBlock *NextBlock = nullptr; |
1991 | MachineFunction::iterator BBI = SwitchBB; |
1992 | if (++BBI != FuncInfo.MF->end()) |
1993 | NextBlock = BBI; |
1994 | |
1995 | if (NextMBB != NextBlock) |
1996 | BrAnd = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, BrAnd, |
1997 | DAG.getBasicBlock(NextMBB)); |
1998 | |
1999 | DAG.setRoot(BrAnd); |
2000 | } |
2001 | |
2002 | void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) { |
2003 | MachineBasicBlock *InvokeMBB = FuncInfo.MBB; |
2004 | |
2005 | // Retrieve successors. |
2006 | MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)]; |
2007 | MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)]; |
2008 | |
2009 | const Value *Callee(I.getCalledValue()); |
2010 | const Function *Fn = dyn_cast<Function>(Callee); |
2011 | if (isa<InlineAsm>(Callee)) |
2012 | visitInlineAsm(&I); |
2013 | else if (Fn && Fn->isIntrinsic()) { |
2014 | switch (Fn->getIntrinsicID()) { |
2015 | default: |
2016 | llvm_unreachable("Cannot invoke this intrinsic")::llvm::llvm_unreachable_internal("Cannot invoke this intrinsic" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2016); |
2017 | case Intrinsic::donothing: |
2018 | // Ignore invokes to @llvm.donothing: jump directly to the next BB. |
2019 | break; |
2020 | case Intrinsic::experimental_patchpoint_void: |
2021 | case Intrinsic::experimental_patchpoint_i64: |
2022 | visitPatchpoint(&I, LandingPad); |
2023 | break; |
2024 | } |
2025 | } else |
2026 | LowerCallTo(&I, getValue(Callee), false, LandingPad); |
2027 | |
2028 | // If the value of the invoke is used outside of its defining block, make it |
2029 | // available as a virtual register. |
2030 | CopyToExportRegsIfNeeded(&I); |
2031 | |
2032 | // Update successor info |
2033 | addSuccessorWithWeight(InvokeMBB, Return); |
2034 | addSuccessorWithWeight(InvokeMBB, LandingPad); |
2035 | |
2036 | // Drop into normal successor. |
2037 | DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), |
2038 | MVT::Other, getControlRoot(), |
2039 | DAG.getBasicBlock(Return))); |
2040 | } |
2041 | |
2042 | void SelectionDAGBuilder::visitResume(const ResumeInst &RI) { |
2043 | llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!")::llvm::llvm_unreachable_internal("SelectionDAGBuilder shouldn't visit resume instructions!" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2043); |
2044 | } |
2045 | |
2046 | void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &LP) { |
2047 | assert(FuncInfo.MBB->isLandingPad() &&((FuncInfo.MBB->isLandingPad() && "Call to landingpad not in landing pad!" ) ? static_cast<void> (0) : __assert_fail ("FuncInfo.MBB->isLandingPad() && \"Call to landingpad not in landing pad!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2048, __PRETTY_FUNCTION__)) |
2048 | "Call to landingpad not in landing pad!")((FuncInfo.MBB->isLandingPad() && "Call to landingpad not in landing pad!" ) ? static_cast<void> (0) : __assert_fail ("FuncInfo.MBB->isLandingPad() && \"Call to landingpad not in landing pad!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2048, __PRETTY_FUNCTION__)); |
2049 | |
2050 | MachineBasicBlock *MBB = FuncInfo.MBB; |
2051 | MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); |
2052 | AddLandingPadInfo(LP, MMI, MBB); |
2053 | |
2054 | // If there aren't registers to copy the values into (e.g., during SjLj |
2055 | // exceptions), then don't bother to create these DAG nodes. |
2056 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
2057 | if (TLI.getExceptionPointerRegister() == 0 && |
2058 | TLI.getExceptionSelectorRegister() == 0) |
2059 | return; |
2060 | |
2061 | SmallVector<EVT, 2> ValueVTs; |
2062 | ComputeValueVTs(TLI, LP.getType(), ValueVTs); |
2063 | assert(ValueVTs.size() == 2 && "Only two-valued landingpads are supported")((ValueVTs.size() == 2 && "Only two-valued landingpads are supported" ) ? static_cast<void> (0) : __assert_fail ("ValueVTs.size() == 2 && \"Only two-valued landingpads are supported\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2063, __PRETTY_FUNCTION__)); |
2064 | |
2065 | // Get the two live-in registers as SDValues. The physregs have already been |
2066 | // copied into virtual registers. |
2067 | SDValue Ops[2]; |
2068 | Ops[0] = DAG.getZExtOrTrunc( |
2069 | DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(), |
2070 | FuncInfo.ExceptionPointerVirtReg, TLI.getPointerTy()), |
2071 | getCurSDLoc(), ValueVTs[0]); |
2072 | Ops[1] = DAG.getZExtOrTrunc( |
2073 | DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(), |
2074 | FuncInfo.ExceptionSelectorVirtReg, TLI.getPointerTy()), |
2075 | getCurSDLoc(), ValueVTs[1]); |
2076 | |
2077 | // Merge into one. |
2078 | SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), |
2079 | DAG.getVTList(ValueVTs), Ops); |
2080 | setValue(&LP, Res); |
2081 | } |
2082 | |
2083 | /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for |
2084 | /// small case ranges). |
2085 | bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR, |
2086 | CaseRecVector& WorkList, |
2087 | const Value* SV, |
2088 | MachineBasicBlock *Default, |
2089 | MachineBasicBlock *SwitchBB) { |
2090 | // Size is the number of Cases represented by this range. |
2091 | size_t Size = CR.Range.second - CR.Range.first; |
2092 | if (Size > 3) |
2093 | return false; |
2094 | |
2095 | // Get the MachineFunction which holds the current MBB. This is used when |
2096 | // inserting any additional MBBs necessary to represent the switch. |
2097 | MachineFunction *CurMF = FuncInfo.MF; |
2098 | |
2099 | // Figure out which block is immediately after the current one. |
2100 | MachineBasicBlock *NextBlock = nullptr; |
2101 | MachineFunction::iterator BBI = CR.CaseBB; |
2102 | |
2103 | if (++BBI != FuncInfo.MF->end()) |
2104 | NextBlock = BBI; |
2105 | |
2106 | BranchProbabilityInfo *BPI = FuncInfo.BPI; |
2107 | // If any two of the cases has the same destination, and if one value |
2108 | // is the same as the other, but has one bit unset that the other has set, |
2109 | // use bit manipulation to do two compares at once. For example: |
2110 | // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)" |
2111 | // TODO: This could be extended to merge any 2 cases in switches with 3 cases. |
2112 | // TODO: Handle cases where CR.CaseBB != SwitchBB. |
2113 | if (Size == 2 && CR.CaseBB == SwitchBB) { |
2114 | Case &Small = *CR.Range.first; |
2115 | Case &Big = *(CR.Range.second-1); |
2116 | |
2117 | if (Small.Low == Small.High && Big.Low == Big.High && Small.BB == Big.BB) { |
2118 | const APInt& SmallValue = cast<ConstantInt>(Small.Low)->getValue(); |
2119 | const APInt& BigValue = cast<ConstantInt>(Big.Low)->getValue(); |
2120 | |
2121 | // Check that there is only one bit different. |
2122 | if (BigValue.countPopulation() == SmallValue.countPopulation() + 1 && |
2123 | (SmallValue | BigValue) == BigValue) { |
2124 | // Isolate the common bit. |
2125 | APInt CommonBit = BigValue & ~SmallValue; |
2126 | assert((SmallValue | CommonBit) == BigValue &&(((SmallValue | CommonBit) == BigValue && CommonBit.countPopulation () == 1 && "Not a common bit?") ? static_cast<void > (0) : __assert_fail ("(SmallValue | CommonBit) == BigValue && CommonBit.countPopulation() == 1 && \"Not a common bit?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2127, __PRETTY_FUNCTION__)) |
2127 | CommonBit.countPopulation() == 1 && "Not a common bit?")(((SmallValue | CommonBit) == BigValue && CommonBit.countPopulation () == 1 && "Not a common bit?") ? static_cast<void > (0) : __assert_fail ("(SmallValue | CommonBit) == BigValue && CommonBit.countPopulation() == 1 && \"Not a common bit?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2127, __PRETTY_FUNCTION__)); |
2128 | |
2129 | SDValue CondLHS = getValue(SV); |
2130 | EVT VT = CondLHS.getValueType(); |
2131 | SDLoc DL = getCurSDLoc(); |
2132 | |
2133 | SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS, |
2134 | DAG.getConstant(CommonBit, VT)); |
2135 | SDValue Cond = DAG.getSetCC(DL, MVT::i1, |
2136 | Or, DAG.getConstant(BigValue, VT), |
2137 | ISD::SETEQ); |
2138 | |
2139 | // Update successor info. |
2140 | // Both Small and Big will jump to Small.BB, so we sum up the weights. |
2141 | addSuccessorWithWeight(SwitchBB, Small.BB, |
2142 | Small.ExtraWeight + Big.ExtraWeight); |
2143 | addSuccessorWithWeight(SwitchBB, Default, |
2144 | // The default destination is the first successor in IR. |
2145 | BPI ? BPI->getEdgeWeight(SwitchBB->getBasicBlock(), (unsigned)0) : 0); |
2146 | |
2147 | // Insert the true branch. |
2148 | SDValue BrCond = DAG.getNode(ISD::BRCOND, DL, MVT::Other, |
2149 | getControlRoot(), Cond, |
2150 | DAG.getBasicBlock(Small.BB)); |
2151 | |
2152 | // Insert the false branch. |
2153 | BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond, |
2154 | DAG.getBasicBlock(Default)); |
2155 | |
2156 | DAG.setRoot(BrCond); |
2157 | return true; |
2158 | } |
2159 | } |
2160 | } |
2161 | |
2162 | // Order cases by weight so the most likely case will be checked first. |
2163 | uint32_t UnhandledWeights = 0; |
2164 | if (BPI) { |
2165 | for (CaseItr I = CR.Range.first, IE = CR.Range.second; I != IE; ++I) { |
2166 | uint32_t IWeight = I->ExtraWeight; |
2167 | UnhandledWeights += IWeight; |
2168 | for (CaseItr J = CR.Range.first; J < I; ++J) { |
2169 | uint32_t JWeight = J->ExtraWeight; |
2170 | if (IWeight > JWeight) |
2171 | std::swap(*I, *J); |
2172 | } |
2173 | } |
2174 | } |
2175 | // Rearrange the case blocks so that the last one falls through if possible. |
2176 | Case &BackCase = *(CR.Range.second-1); |
2177 | if (Size > 1 && |
2178 | NextBlock && Default != NextBlock && BackCase.BB != NextBlock) { |
2179 | // The last case block won't fall through into 'NextBlock' if we emit the |
2180 | // branches in this order. See if rearranging a case value would help. |
2181 | // We start at the bottom as it's the case with the least weight. |
2182 | for (Case *I = &*(CR.Range.second-2), *E = &*CR.Range.first-1; I != E; --I) |
2183 | if (I->BB == NextBlock) { |
2184 | std::swap(*I, BackCase); |
2185 | break; |
2186 | } |
2187 | } |
2188 | |
2189 | // Create a CaseBlock record representing a conditional branch to |
2190 | // the Case's target mbb if the value being switched on SV is equal |
2191 | // to C. |
2192 | MachineBasicBlock *CurBlock = CR.CaseBB; |
2193 | for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) { |
2194 | MachineBasicBlock *FallThrough; |
2195 | if (I != E-1) { |
2196 | FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock()); |
2197 | CurMF->insert(BBI, FallThrough); |
2198 | |
2199 | // Put SV in a virtual register to make it available from the new blocks. |
2200 | ExportFromCurrentBlock(SV); |
2201 | } else { |
2202 | // If the last case doesn't match, go to the default block. |
2203 | FallThrough = Default; |
2204 | } |
2205 | |
2206 | const Value *RHS, *LHS, *MHS; |
2207 | ISD::CondCode CC; |
2208 | if (I->High == I->Low) { |
2209 | // This is just small small case range :) containing exactly 1 case |
2210 | CC = ISD::SETEQ; |
2211 | LHS = SV; RHS = I->High; MHS = nullptr; |
2212 | } else { |
2213 | CC = ISD::SETLE; |
2214 | LHS = I->Low; MHS = SV; RHS = I->High; |
2215 | } |
2216 | |
2217 | // The false weight should be sum of all un-handled cases. |
2218 | UnhandledWeights -= I->ExtraWeight; |
2219 | CaseBlock CB(CC, LHS, RHS, MHS, /* truebb */ I->BB, /* falsebb */ FallThrough, |
2220 | /* me */ CurBlock, |
2221 | /* trueweight */ I->ExtraWeight, |
2222 | /* falseweight */ UnhandledWeights); |
2223 | |
2224 | // If emitting the first comparison, just call visitSwitchCase to emit the |
2225 | // code into the current block. Otherwise, push the CaseBlock onto the |
2226 | // vector to be later processed by SDISel, and insert the node's MBB |
2227 | // before the next MBB. |
2228 | if (CurBlock == SwitchBB) |
2229 | visitSwitchCase(CB, SwitchBB); |
2230 | else |
2231 | SwitchCases.push_back(CB); |
2232 | |
2233 | CurBlock = FallThrough; |
2234 | } |
2235 | |
2236 | return true; |
2237 | } |
2238 | |
2239 | static inline bool areJTsAllowed(const TargetLowering &TLI) { |
2240 | return TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) || |
2241 | TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other); |
2242 | } |
2243 | |
2244 | static APInt ComputeRange(const APInt &First, const APInt &Last) { |
2245 | uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1; |
2246 | APInt LastExt = Last.sext(BitWidth), FirstExt = First.sext(BitWidth); |
2247 | return (LastExt - FirstExt + 1ULL); |
2248 | } |
2249 | |
2250 | /// handleJTSwitchCase - Emit jumptable for current switch case range |
2251 | bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec &CR, |
2252 | CaseRecVector &WorkList, |
2253 | const Value *SV, |
2254 | MachineBasicBlock *Default, |
2255 | MachineBasicBlock *SwitchBB) { |
2256 | Case& FrontCase = *CR.Range.first; |
2257 | Case& BackCase = *(CR.Range.second-1); |
2258 | |
2259 | const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue(); |
2260 | const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue(); |
2261 | |
2262 | APInt TSize(First.getBitWidth(), 0); |
2263 | for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) |
2264 | TSize += I->size(); |
2265 | |
2266 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
2267 | if (!areJTsAllowed(TLI) || TSize.ult(TLI.getMinimumJumpTableEntries())) |
2268 | return false; |
2269 | |
2270 | APInt Range = ComputeRange(First, Last); |
2271 | // The density is TSize / Range. Require at least 40%. |
2272 | // It should not be possible for IntTSize to saturate for sane code, but make |
2273 | // sure we handle Range saturation correctly. |
2274 | uint64_t IntRange = Range.getLimitedValue(UINT64_MAX(18446744073709551615UL)/10); |
2275 | uint64_t IntTSize = TSize.getLimitedValue(UINT64_MAX(18446744073709551615UL)/10); |
2276 | if (IntTSize * 10 < IntRange * 4) |
2277 | return false; |
2278 | |
2279 | DEBUG(dbgs() << "Lowering jump table\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Lowering jump table\n" << "First entry: " << First << ". Last entry: " << Last << '\n' << "Range: " << Range << ". Size: " << TSize << ".\n\n"; } } while (0) |
2280 | << "First entry: " << First << ". Last entry: " << Last << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Lowering jump table\n" << "First entry: " << First << ". Last entry: " << Last << '\n' << "Range: " << Range << ". Size: " << TSize << ".\n\n"; } } while (0) |
2281 | << "Range: " << Range << ". Size: " << TSize << ".\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Lowering jump table\n" << "First entry: " << First << ". Last entry: " << Last << '\n' << "Range: " << Range << ". Size: " << TSize << ".\n\n"; } } while (0); |
2282 | |
2283 | // Get the MachineFunction which holds the current MBB. This is used when |
2284 | // inserting any additional MBBs necessary to represent the switch. |
2285 | MachineFunction *CurMF = FuncInfo.MF; |
2286 | |
2287 | // Figure out which block is immediately after the current one. |
2288 | MachineFunction::iterator BBI = CR.CaseBB; |
2289 | ++BBI; |
2290 | |
2291 | const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); |
2292 | |
2293 | // Create a new basic block to hold the code for loading the address |
2294 | // of the jump table, and jumping to it. Update successor information; |
2295 | // we will either branch to the default case for the switch, or the jump |
2296 | // table. |
2297 | MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB); |
2298 | CurMF->insert(BBI, JumpTableBB); |
2299 | |
2300 | addSuccessorWithWeight(CR.CaseBB, Default); |
2301 | addSuccessorWithWeight(CR.CaseBB, JumpTableBB); |
2302 | |
2303 | // Build a vector of destination BBs, corresponding to each target |
2304 | // of the jump table. If the value of the jump table slot corresponds to |
2305 | // a case statement, push the case's BB onto the vector, otherwise, push |
2306 | // the default BB. |
2307 | std::vector<MachineBasicBlock*> DestBBs; |
2308 | APInt TEI = First; |
2309 | for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) { |
2310 | const APInt &Low = cast<ConstantInt>(I->Low)->getValue(); |
2311 | const APInt &High = cast<ConstantInt>(I->High)->getValue(); |
2312 | |
2313 | if (Low.sle(TEI) && TEI.sle(High)) { |
2314 | DestBBs.push_back(I->BB); |
2315 | if (TEI==High) |
2316 | ++I; |
2317 | } else { |
2318 | DestBBs.push_back(Default); |
2319 | } |
2320 | } |
2321 | |
2322 | // Calculate weight for each unique destination in CR. |
2323 | DenseMap<MachineBasicBlock*, uint32_t> DestWeights; |
2324 | if (FuncInfo.BPI) |
2325 | for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) { |
2326 | DenseMap<MachineBasicBlock*, uint32_t>::iterator Itr = |
2327 | DestWeights.find(I->BB); |
2328 | if (Itr != DestWeights.end()) |
2329 | Itr->second += I->ExtraWeight; |
2330 | else |
2331 | DestWeights[I->BB] = I->ExtraWeight; |
2332 | } |
2333 | |
2334 | // Update successor info. Add one edge to each unique successor. |
2335 | BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs()); |
2336 | for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(), |
2337 | E = DestBBs.end(); I != E; ++I) { |
2338 | if (!SuccsHandled[(*I)->getNumber()]) { |
2339 | SuccsHandled[(*I)->getNumber()] = true; |
2340 | DenseMap<MachineBasicBlock*, uint32_t>::iterator Itr = |
2341 | DestWeights.find(*I); |
2342 | addSuccessorWithWeight(JumpTableBB, *I, |
2343 | Itr != DestWeights.end() ? Itr->second : 0); |
2344 | } |
2345 | } |
2346 | |
2347 | // Create a jump table index for this jump table. |
2348 | unsigned JTEncoding = TLI.getJumpTableEncoding(); |
2349 | unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding) |
2350 | ->createJumpTableIndex(DestBBs); |
2351 | |
2352 | // Set the jump table information so that we can codegen it as a second |
2353 | // MachineBasicBlock |
2354 | JumpTable JT(-1U, JTI, JumpTableBB, Default); |
2355 | JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB)); |
2356 | if (CR.CaseBB == SwitchBB) |
2357 | visitJumpTableHeader(JT, JTH, SwitchBB); |
2358 | |
2359 | JTCases.push_back(JumpTableBlock(JTH, JT)); |
2360 | return true; |
2361 | } |
2362 | |
2363 | /// handleBTSplitSwitchCase - emit comparison and split binary search tree into |
2364 | /// 2 subtrees. |
2365 | bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR, |
2366 | CaseRecVector& WorkList, |
2367 | const Value* SV, |
2368 | MachineBasicBlock* SwitchBB) { |
2369 | // Get the MachineFunction which holds the current MBB. This is used when |
2370 | // inserting any additional MBBs necessary to represent the switch. |
2371 | MachineFunction *CurMF = FuncInfo.MF; |
2372 | |
2373 | // Figure out which block is immediately after the current one. |
2374 | MachineFunction::iterator BBI = CR.CaseBB; |
2375 | ++BBI; |
2376 | |
2377 | Case& FrontCase = *CR.Range.first; |
2378 | Case& BackCase = *(CR.Range.second-1); |
2379 | const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); |
2380 | |
2381 | // Size is the number of Cases represented by this range. |
2382 | unsigned Size = CR.Range.second - CR.Range.first; |
2383 | |
2384 | const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue(); |
2385 | const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue(); |
2386 | double FMetric = 0; |
2387 | CaseItr Pivot = CR.Range.first + Size/2; |
2388 | |
2389 | // Select optimal pivot, maximizing sum density of LHS and RHS. This will |
2390 | // (heuristically) allow us to emit JumpTable's later. |
2391 | APInt TSize(First.getBitWidth(), 0); |
2392 | for (CaseItr I = CR.Range.first, E = CR.Range.second; |
2393 | I!=E; ++I) |
2394 | TSize += I->size(); |
2395 | |
2396 | APInt LSize = FrontCase.size(); |
2397 | APInt RSize = TSize-LSize; |
2398 | DEBUG(dbgs() << "Selecting best pivot: \n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Selecting best pivot: \n" << "First: " << First << ", Last: " << Last << '\n' << "LSize: " << LSize << ", RSize: " << RSize << '\n'; } } while (0) |
2399 | << "First: " << First << ", Last: " << Last <<'\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Selecting best pivot: \n" << "First: " << First << ", Last: " << Last << '\n' << "LSize: " << LSize << ", RSize: " << RSize << '\n'; } } while (0) |
2400 | << "LSize: " << LSize << ", RSize: " << RSize << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Selecting best pivot: \n" << "First: " << First << ", Last: " << Last << '\n' << "LSize: " << LSize << ", RSize: " << RSize << '\n'; } } while (0); |
2401 | for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second; |
2402 | J!=E; ++I, ++J) { |
2403 | const APInt &LEnd = cast<ConstantInt>(I->High)->getValue(); |
2404 | const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue(); |
2405 | APInt Range = ComputeRange(LEnd, RBegin); |
2406 | assert((Range - 2ULL).isNonNegative() &&(((Range - 2ULL).isNonNegative() && "Invalid case distance" ) ? static_cast<void> (0) : __assert_fail ("(Range - 2ULL).isNonNegative() && \"Invalid case distance\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2407, __PRETTY_FUNCTION__)) |
2407 | "Invalid case distance")(((Range - 2ULL).isNonNegative() && "Invalid case distance" ) ? static_cast<void> (0) : __assert_fail ("(Range - 2ULL).isNonNegative() && \"Invalid case distance\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2407, __PRETTY_FUNCTION__)); |
2408 | // Use volatile double here to avoid excess precision issues on some hosts, |
2409 | // e.g. that use 80-bit X87 registers. |
2410 | volatile double LDensity = |
2411 | (double)LSize.roundToDouble() / |
2412 | (LEnd - First + 1ULL).roundToDouble(); |
2413 | volatile double RDensity = |
2414 | (double)RSize.roundToDouble() / |
2415 | (Last - RBegin + 1ULL).roundToDouble(); |
2416 | volatile double Metric = Range.logBase2()*(LDensity+RDensity); |
2417 | // Should always split in some non-trivial place |
2418 | DEBUG(dbgs() <<"=>Step\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() <<"=>Step\n" << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n' << "LDensity: " << LDensity << ", RDensity: " << RDensity << '\n' << "Metric: " << Metric << '\n'; } } while (0) |
2419 | << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() <<"=>Step\n" << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n' << "LDensity: " << LDensity << ", RDensity: " << RDensity << '\n' << "Metric: " << Metric << '\n'; } } while (0) |
2420 | << "LDensity: " << LDensitydo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() <<"=>Step\n" << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n' << "LDensity: " << LDensity << ", RDensity: " << RDensity << '\n' << "Metric: " << Metric << '\n'; } } while (0) |
2421 | << ", RDensity: " << RDensity << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() <<"=>Step\n" << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n' << "LDensity: " << LDensity << ", RDensity: " << RDensity << '\n' << "Metric: " << Metric << '\n'; } } while (0) |
2422 | << "Metric: " << Metric << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() <<"=>Step\n" << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n' << "LDensity: " << LDensity << ", RDensity: " << RDensity << '\n' << "Metric: " << Metric << '\n'; } } while (0); |
2423 | if (FMetric < Metric) { |
2424 | Pivot = J; |
2425 | FMetric = Metric; |
2426 | DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Current metric set to: " << FMetric << '\n'; } } while (0); |
2427 | } |
2428 | |
2429 | LSize += J->size(); |
2430 | RSize -= J->size(); |
2431 | } |
2432 | |
2433 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
2434 | if (areJTsAllowed(TLI)) { |
2435 | // If our case is dense we *really* should handle it earlier! |
2436 | assert((FMetric > 0) && "Should handle dense range earlier!")(((FMetric > 0) && "Should handle dense range earlier!" ) ? static_cast<void> (0) : __assert_fail ("(FMetric > 0) && \"Should handle dense range earlier!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2436, __PRETTY_FUNCTION__)); |
2437 | } else { |
2438 | Pivot = CR.Range.first + Size/2; |
2439 | } |
2440 | |
2441 | CaseRange LHSR(CR.Range.first, Pivot); |
2442 | CaseRange RHSR(Pivot, CR.Range.second); |
2443 | const Constant *C = Pivot->Low; |
2444 | MachineBasicBlock *FalseBB = nullptr, *TrueBB = nullptr; |
2445 | |
2446 | // We know that we branch to the LHS if the Value being switched on is |
2447 | // less than the Pivot value, C. We use this to optimize our binary |
2448 | // tree a bit, by recognizing that if SV is greater than or equal to the |
2449 | // LHS's Case Value, and that Case Value is exactly one less than the |
2450 | // Pivot's Value, then we can branch directly to the LHS's Target, |
2451 | // rather than creating a leaf node for it. |
2452 | if ((LHSR.second - LHSR.first) == 1 && |
2453 | LHSR.first->High == CR.GE && |
2454 | cast<ConstantInt>(C)->getValue() == |
2455 | (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) { |
2456 | TrueBB = LHSR.first->BB; |
2457 | } else { |
2458 | TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB); |
2459 | CurMF->insert(BBI, TrueBB); |
2460 | WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR)); |
2461 | |
2462 | // Put SV in a virtual register to make it available from the new blocks. |
2463 | ExportFromCurrentBlock(SV); |
2464 | } |
2465 | |
2466 | // Similar to the optimization above, if the Value being switched on is |
2467 | // known to be less than the Constant CR.LT, and the current Case Value |
2468 | // is CR.LT - 1, then we can branch directly to the target block for |
2469 | // the current Case Value, rather than emitting a RHS leaf node for it. |
2470 | if ((RHSR.second - RHSR.first) == 1 && CR.LT && |
2471 | cast<ConstantInt>(RHSR.first->Low)->getValue() == |
2472 | (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) { |
2473 | FalseBB = RHSR.first->BB; |
2474 | } else { |
2475 | FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB); |
2476 | CurMF->insert(BBI, FalseBB); |
2477 | WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR)); |
2478 | |
2479 | // Put SV in a virtual register to make it available from the new blocks. |
2480 | ExportFromCurrentBlock(SV); |
2481 | } |
2482 | |
2483 | // Create a CaseBlock record representing a conditional branch to |
2484 | // the LHS node if the value being switched on SV is less than C. |
2485 | // Otherwise, branch to LHS. |
2486 | CaseBlock CB(ISD::SETLT, SV, C, nullptr, TrueBB, FalseBB, CR.CaseBB); |
2487 | |
2488 | if (CR.CaseBB == SwitchBB) |
2489 | visitSwitchCase(CB, SwitchBB); |
2490 | else |
2491 | SwitchCases.push_back(CB); |
2492 | |
2493 | return true; |
2494 | } |
2495 | |
2496 | /// handleBitTestsSwitchCase - if current case range has few destination and |
2497 | /// range span less, than machine word bitwidth, encode case range into series |
2498 | /// of masks and emit bit tests with these masks. |
2499 | bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR, |
2500 | CaseRecVector& WorkList, |
2501 | const Value* SV, |
2502 | MachineBasicBlock* Default, |
2503 | MachineBasicBlock* SwitchBB) { |
2504 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
2505 | EVT PTy = TLI.getPointerTy(); |
2506 | unsigned IntPtrBits = PTy.getSizeInBits(); |
2507 | |
2508 | Case& FrontCase = *CR.Range.first; |
2509 | Case& BackCase = *(CR.Range.second-1); |
2510 | |
2511 | // Get the MachineFunction which holds the current MBB. This is used when |
2512 | // inserting any additional MBBs necessary to represent the switch. |
2513 | MachineFunction *CurMF = FuncInfo.MF; |
2514 | |
2515 | // If target does not have legal shift left, do not emit bit tests at all. |
2516 | if (!TLI.isOperationLegal(ISD::SHL, PTy)) |
2517 | return false; |
2518 | |
2519 | size_t numCmps = 0; |
2520 | for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) { |
2521 | // Single case counts one, case range - two. |
2522 | numCmps += (I->Low == I->High ? 1 : 2); |
2523 | } |
2524 | |
2525 | // Count unique destinations |
2526 | SmallSet<MachineBasicBlock*, 4> Dests; |
2527 | for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) { |
2528 | Dests.insert(I->BB); |
2529 | if (Dests.size() > 3) |
2530 | // Don't bother the code below, if there are too much unique destinations |
2531 | return false; |
2532 | } |
2533 | DEBUG(dbgs() << "Total number of unique destinations: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Total number of unique destinations: " << Dests.size() << '\n' << "Total number of comparisons: " << numCmps << '\n'; } } while (0) |
2534 | << Dests.size() << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Total number of unique destinations: " << Dests.size() << '\n' << "Total number of comparisons: " << numCmps << '\n'; } } while (0) |
2535 | << "Total number of comparisons: " << numCmps << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Total number of unique destinations: " << Dests.size() << '\n' << "Total number of comparisons: " << numCmps << '\n'; } } while (0); |
2536 | |
2537 | // Compute span of values. |
2538 | const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue(); |
2539 | const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue(); |
2540 | APInt cmpRange = maxValue - minValue; |
2541 | |
2542 | DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Compare range: " << cmpRange << '\n' << "Low bound: " << minValue << '\n' << "High bound: " << maxValue << '\n' ; } } while (0) |
2543 | << "Low bound: " << minValue << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Compare range: " << cmpRange << '\n' << "Low bound: " << minValue << '\n' << "High bound: " << maxValue << '\n' ; } } while (0) |
2544 | << "High bound: " << maxValue << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Compare range: " << cmpRange << '\n' << "Low bound: " << minValue << '\n' << "High bound: " << maxValue << '\n' ; } } while (0); |
2545 | |
2546 | if (cmpRange.uge(IntPtrBits) || |
2547 | (!(Dests.size() == 1 && numCmps >= 3) && |
2548 | !(Dests.size() == 2 && numCmps >= 5) && |
2549 | !(Dests.size() >= 3 && numCmps >= 6))) |
2550 | return false; |
2551 | |
2552 | DEBUG(dbgs() << "Emitting bit tests\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Emitting bit tests\n"; } } while (0); |
2553 | APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth()); |
2554 | |
2555 | // Optimize the case where all the case values fit in a |
2556 | // word without having to subtract minValue. In this case, |
2557 | // we can optimize away the subtraction. |
2558 | if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) { |
2559 | cmpRange = maxValue; |
2560 | } else { |
2561 | lowBound = minValue; |
2562 | } |
2563 | |
2564 | CaseBitsVector CasesBits; |
2565 | unsigned i, count = 0; |
2566 | |
2567 | for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) { |
2568 | MachineBasicBlock* Dest = I->BB; |
2569 | for (i = 0; i < count; ++i) |
2570 | if (Dest == CasesBits[i].BB) |
2571 | break; |
2572 | |
2573 | if (i == count) { |
2574 | assert((count < 3) && "Too much destinations to test!")(((count < 3) && "Too much destinations to test!") ? static_cast<void> (0) : __assert_fail ("(count < 3) && \"Too much destinations to test!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2574, __PRETTY_FUNCTION__)); |
2575 | CasesBits.push_back(CaseBits(0, Dest, 0, 0/*Weight*/)); |
2576 | count++; |
2577 | } |
2578 | |
2579 | const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue(); |
2580 | const APInt& highValue = cast<ConstantInt>(I->High)->getValue(); |
2581 | |
2582 | uint64_t lo = (lowValue - lowBound).getZExtValue(); |
2583 | uint64_t hi = (highValue - lowBound).getZExtValue(); |
2584 | CasesBits[i].ExtraWeight += I->ExtraWeight; |
2585 | |
2586 | for (uint64_t j = lo; j <= hi; j++) { |
2587 | CasesBits[i].Mask |= 1ULL << j; |
2588 | CasesBits[i].Bits++; |
2589 | } |
2590 | |
2591 | } |
2592 | std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp()); |
2593 | |
2594 | BitTestInfo BTC; |
2595 | |
2596 | // Figure out which block is immediately after the current one. |
2597 | MachineFunction::iterator BBI = CR.CaseBB; |
2598 | ++BBI; |
2599 | |
2600 | const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); |
2601 | |
2602 | DEBUG(dbgs() << "Cases:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Cases:\n"; } } while (0); |
2603 | for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) { |
2604 | DEBUG(dbgs() << "Mask: " << CasesBits[i].Maskdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Mask: " << CasesBits[i].Mask << ", Bits: " << CasesBits[i].Bits << ", BB: " << CasesBits[i].BB << '\n'; } } while (0) |
2605 | << ", Bits: " << CasesBits[i].Bitsdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Mask: " << CasesBits[i].Mask << ", Bits: " << CasesBits[i].Bits << ", BB: " << CasesBits[i].BB << '\n'; } } while (0) |
2606 | << ", BB: " << CasesBits[i].BB << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Mask: " << CasesBits[i].Mask << ", Bits: " << CasesBits[i].Bits << ", BB: " << CasesBits[i].BB << '\n'; } } while (0); |
2607 | |
2608 | MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB); |
2609 | CurMF->insert(BBI, CaseBB); |
2610 | BTC.push_back(BitTestCase(CasesBits[i].Mask, |
2611 | CaseBB, |
2612 | CasesBits[i].BB, CasesBits[i].ExtraWeight)); |
2613 | |
2614 | // Put SV in a virtual register to make it available from the new blocks. |
2615 | ExportFromCurrentBlock(SV); |
2616 | } |
2617 | |
2618 | BitTestBlock BTB(lowBound, cmpRange, SV, |
2619 | -1U, MVT::Other, (CR.CaseBB == SwitchBB), |
2620 | CR.CaseBB, Default, std::move(BTC)); |
2621 | |
2622 | if (CR.CaseBB == SwitchBB) |
2623 | visitBitTestHeader(BTB, SwitchBB); |
2624 | |
2625 | BitTestCases.push_back(std::move(BTB)); |
2626 | |
2627 | return true; |
2628 | } |
2629 | |
2630 | /// Clusterify - Transform simple list of Cases into list of CaseRange's |
2631 | void SelectionDAGBuilder::Clusterify(CaseVector& Cases, |
2632 | const SwitchInst& SI) { |
2633 | BranchProbabilityInfo *BPI = FuncInfo.BPI; |
2634 | // Start with "simple" cases. |
2635 | for (SwitchInst::ConstCaseIt i : SI.cases()) { |
2636 | const BasicBlock *SuccBB = i.getCaseSuccessor(); |
2637 | MachineBasicBlock *SMBB = FuncInfo.MBBMap[SuccBB]; |
2638 | |
2639 | uint32_t ExtraWeight = |
2640 | BPI ? BPI->getEdgeWeight(SI.getParent(), i.getSuccessorIndex()) : 0; |
2641 | |
2642 | Cases.push_back(Case(i.getCaseValue(), i.getCaseValue(), |
2643 | SMBB, ExtraWeight)); |
2644 | } |
2645 | std::sort(Cases.begin(), Cases.end(), CaseCmp()); |
2646 | |
2647 | // Merge case into clusters |
2648 | if (Cases.size() >= 2) |
2649 | // Must recompute end() each iteration because it may be |
2650 | // invalidated by erase if we hold on to it |
2651 | for (CaseItr I = Cases.begin(), J = std::next(Cases.begin()); |
2652 | J != Cases.end(); ) { |
2653 | const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue(); |
2654 | const APInt& currentValue = cast<ConstantInt>(I->High)->getValue(); |
2655 | MachineBasicBlock* nextBB = J->BB; |
2656 | MachineBasicBlock* currentBB = I->BB; |
2657 | |
2658 | // If the two neighboring cases go to the same destination, merge them |
2659 | // into a single case. |
2660 | if ((nextValue - currentValue == 1) && (currentBB == nextBB)) { |
2661 | I->High = J->High; |
2662 | I->ExtraWeight += J->ExtraWeight; |
2663 | J = Cases.erase(J); |
2664 | } else { |
2665 | I = J++; |
2666 | } |
2667 | } |
2668 | |
2669 | DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0) |
2670 | size_t numCmps = 0;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0) |
2671 | for (auto &I : Cases)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0) |
2672 | // A range counts double, since it requires two compares.do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0) |
2673 | numCmps += I.Low != I.High ? 2 : 1;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0) |
2674 | |
2675 | dbgs() << "Clusterify finished. Total clusters: " << Cases.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0) |
2676 | << ". Total compares: " << numCmps << '\n';do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0) |
2677 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0); |
2678 | } |
2679 | |
2680 | void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First, |
2681 | MachineBasicBlock *Last) { |
2682 | // Update JTCases. |
2683 | for (unsigned i = 0, e = JTCases.size(); i != e; ++i) |
2684 | if (JTCases[i].first.HeaderBB == First) |
2685 | JTCases[i].first.HeaderBB = Last; |
2686 | |
2687 | // Update BitTestCases. |
2688 | for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i) |
2689 | if (BitTestCases[i].Parent == First) |
2690 | BitTestCases[i].Parent = Last; |
2691 | } |
2692 | |
2693 | void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) { |
2694 | MachineBasicBlock *SwitchMBB = FuncInfo.MBB; |
2695 | |
2696 | // Figure out which block is immediately after the current one. |
2697 | MachineBasicBlock *NextBlock = nullptr; |
2698 | if (SwitchMBB + 1 != FuncInfo.MF->end()) |
2699 | NextBlock = SwitchMBB + 1; |
2700 | |
2701 | |
2702 | // Create a vector of Cases, sorted so that we can efficiently create a binary |
2703 | // search tree from them. |
2704 | CaseVector Cases; |
2705 | Clusterify(Cases, SI); |
2706 | |
2707 | // Get the default destination MBB. |
2708 | MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()]; |
2709 | |
2710 | if (isa<UnreachableInst>(SI.getDefaultDest()->getFirstNonPHIOrDbg()) && |
2711 | !Cases.empty()) { |
2712 | // Replace an unreachable default destination with the most popular case |
2713 | // destination. |
2714 | DenseMap<const BasicBlock *, uint64_t> Popularity; |
2715 | uint64_t MaxPop = 0; |
2716 | const BasicBlock *MaxBB = nullptr; |
2717 | for (auto I : SI.cases()) { |
2718 | const BasicBlock *BB = I.getCaseSuccessor(); |
2719 | if (++Popularity[BB] > MaxPop) { |
2720 | MaxPop = Popularity[BB]; |
2721 | MaxBB = BB; |
2722 | } |
2723 | } |
2724 | |
2725 | // Set new default. |
2726 | assert(MaxPop > 0)((MaxPop > 0) ? static_cast<void> (0) : __assert_fail ("MaxPop > 0", "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2726, __PRETTY_FUNCTION__)); |
2727 | assert(MaxBB)((MaxBB) ? static_cast<void> (0) : __assert_fail ("MaxBB" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2727, __PRETTY_FUNCTION__)); |
2728 | Default = FuncInfo.MBBMap[MaxBB]; |
2729 | |
2730 | // Remove cases that were pointing to the destination that is now the default. |
2731 | Cases.erase(std::remove_if(Cases.begin(), Cases.end(), |
2732 | [&](const Case &C) { return C.BB == Default; }), |
2733 | Cases.end()); |
2734 | } |
2735 | |
2736 | // If there is only the default destination, go there directly. |
2737 | if (Cases.empty()) { |
2738 | // Update machine-CFG edges. |
2739 | SwitchMBB->addSuccessor(Default); |
2740 | |
2741 | // If this is not a fall-through branch, emit the branch. |
2742 | if (Default != NextBlock) { |
2743 | DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, |
2744 | getControlRoot(), DAG.getBasicBlock(Default))); |
2745 | } |
2746 | return; |
2747 | } |
2748 | |
2749 | // Get the Value to be switched on. |
2750 | const Value *SV = SI.getCondition(); |
2751 | |
2752 | // Push the initial CaseRec onto the worklist |
2753 | CaseRecVector WorkList; |
2754 | WorkList.push_back(CaseRec(SwitchMBB,nullptr,nullptr, |
2755 | CaseRange(Cases.begin(),Cases.end()))); |
2756 | |
2757 | while (!WorkList.empty()) { |
2758 | // Grab a record representing a case range to process off the worklist |
2759 | CaseRec CR = WorkList.back(); |
2760 | WorkList.pop_back(); |
2761 | |
2762 | if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB)) |
2763 | continue; |
2764 | |
2765 | // If the range has few cases (two or less) emit a series of specific |
2766 | // tests. |
2767 | if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB)) |
2768 | continue; |
2769 | |
2770 | // If the switch has more than N blocks, and is at least 40% dense, and the |
2771 | // target supports indirect branches, then emit a jump table rather than |
2772 | // lowering the switch to a binary tree of conditional branches. |
2773 | // N defaults to 4 and is controlled via TLS.getMinimumJumpTableEntries(). |
2774 | if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB)) |
2775 | continue; |
2776 | |
2777 | // Emit binary tree. We need to pick a pivot, and push left and right ranges |
2778 | // onto the worklist. Leafs are handled via handleSmallSwitchRange() call. |
2779 | handleBTSplitSwitchCase(CR, WorkList, SV, SwitchMBB); |
2780 | } |
2781 | } |
2782 | |
2783 | void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) { |
2784 | MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB; |
2785 | |
2786 | // Update machine-CFG edges with unique successors. |
2787 | SmallSet<BasicBlock*, 32> Done; |
2788 | for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i) { |
2789 | BasicBlock *BB = I.getSuccessor(i); |
2790 | bool Inserted = Done.insert(BB).second; |
2791 | if (!Inserted) |
2792 | continue; |
2793 | |
2794 | MachineBasicBlock *Succ = FuncInfo.MBBMap[BB]; |
2795 | addSuccessorWithWeight(IndirectBrMBB, Succ); |
2796 | } |
2797 | |
2798 | DAG.setRoot(DAG.getNode(ISD::BRIND, getCurSDLoc(), |
2799 | MVT::Other, getControlRoot(), |
2800 | getValue(I.getAddress()))); |
2801 | } |
2802 | |
2803 | void SelectionDAGBuilder::visitUnreachable(const UnreachableInst &I) { |
2804 | if (DAG.getTarget().Options.TrapUnreachable) |
2805 | DAG.setRoot(DAG.getNode(ISD::TRAP, getCurSDLoc(), MVT::Other, DAG.getRoot())); |
2806 | } |
2807 | |
2808 | void SelectionDAGBuilder::visitFSub(const User &I) { |
2809 | // -0.0 - X --> fneg |
2810 | Type *Ty = I.getType(); |
2811 | if (isa<Constant>(I.getOperand(0)) && |
2812 | I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) { |
2813 | SDValue Op2 = getValue(I.getOperand(1)); |
2814 | setValue(&I, DAG.getNode(ISD::FNEG, getCurSDLoc(), |
2815 | Op2.getValueType(), Op2)); |
2816 | return; |
2817 | } |
2818 | |
2819 | visitBinary(I, ISD::FSUB); |
2820 | } |
2821 | |
2822 | void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) { |
2823 | SDValue Op1 = getValue(I.getOperand(0)); |
2824 | SDValue Op2 = getValue(I.getOperand(1)); |
2825 | |
2826 | bool nuw = false; |
2827 | bool nsw = false; |
2828 | bool exact = false; |
2829 | if (const OverflowingBinaryOperator *OFBinOp = |
2830 | dyn_cast<const OverflowingBinaryOperator>(&I)) { |
2831 | nuw = OFBinOp->hasNoUnsignedWrap(); |
2832 | nsw = OFBinOp->hasNoSignedWrap(); |
2833 | } |
2834 | if (const PossiblyExactOperator *ExactOp = |
2835 | dyn_cast<const PossiblyExactOperator>(&I)) |
2836 | exact = ExactOp->isExact(); |
2837 | |
2838 | SDValue BinNodeValue = DAG.getNode(OpCode, getCurSDLoc(), Op1.getValueType(), |
2839 | Op1, Op2, nuw, nsw, exact); |
2840 | setValue(&I, BinNodeValue); |
2841 | } |
2842 | |
2843 | void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) { |
2844 | SDValue Op1 = getValue(I.getOperand(0)); |
2845 | SDValue Op2 = getValue(I.getOperand(1)); |
2846 | |
2847 | EVT ShiftTy = |
2848 | DAG.getTargetLoweringInfo().getShiftAmountTy(Op2.getValueType()); |
2849 | |
2850 | // Coerce the shift amount to the right type if we can. |
2851 | if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) { |
2852 | unsigned ShiftSize = ShiftTy.getSizeInBits(); |
2853 | unsigned Op2Size = Op2.getValueType().getSizeInBits(); |
2854 | SDLoc DL = getCurSDLoc(); |
2855 | |
2856 | // If the operand is smaller than the shift count type, promote it. |
2857 | if (ShiftSize > Op2Size) |
2858 | Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2); |
2859 | |
2860 | // If the operand is larger than the shift count type but the shift |
2861 | // count type has enough bits to represent any shift value, truncate |
2862 | // it now. This is a common case and it exposes the truncate to |
2863 | // optimization early. |
2864 | else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits())) |
2865 | Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2); |
2866 | // Otherwise we'll need to temporarily settle for some other convenient |
2867 | // type. Type legalization will make adjustments once the shiftee is split. |
2868 | else |
2869 | Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32); |
2870 | } |
2871 | |
2872 | bool nuw = false; |
2873 | bool nsw = false; |
2874 | bool exact = false; |
2875 | |
2876 | if (Opcode == ISD::SRL || Opcode == ISD::SRA || Opcode == ISD::SHL) { |
2877 | |
2878 | if (const OverflowingBinaryOperator *OFBinOp = |
2879 | dyn_cast<const OverflowingBinaryOperator>(&I)) { |
2880 | nuw = OFBinOp->hasNoUnsignedWrap(); |
2881 | nsw = OFBinOp->hasNoSignedWrap(); |
2882 | } |
2883 | if (const PossiblyExactOperator *ExactOp = |
2884 | dyn_cast<const PossiblyExactOperator>(&I)) |
2885 | exact = ExactOp->isExact(); |
2886 | } |
2887 | |
2888 | SDValue Res = DAG.getNode(Opcode, getCurSDLoc(), Op1.getValueType(), Op1, Op2, |
2889 | nuw, nsw, exact); |
2890 | setValue(&I, Res); |
2891 | } |
2892 | |
2893 | void SelectionDAGBuilder::visitSDiv(const User &I) { |
2894 | SDValue Op1 = getValue(I.getOperand(0)); |
2895 | SDValue Op2 = getValue(I.getOperand(1)); |
2896 | |
2897 | // Turn exact SDivs into multiplications. |
2898 | // FIXME: This should be in DAGCombiner, but it doesn't have access to the |
2899 | // exact bit. |
2900 | if (isa<BinaryOperator>(&I) && cast<BinaryOperator>(&I)->isExact() && |
2901 | !isa<ConstantSDNode>(Op1) && |
2902 | isa<ConstantSDNode>(Op2) && !cast<ConstantSDNode>(Op2)->isNullValue()) |
2903 | setValue(&I, DAG.getTargetLoweringInfo() |
2904 | .BuildExactSDIV(Op1, Op2, getCurSDLoc(), DAG)); |
2905 | else |
2906 | setValue(&I, DAG.getNode(ISD::SDIV, getCurSDLoc(), Op1.getValueType(), |
2907 | Op1, Op2)); |
2908 | } |
2909 | |
2910 | void SelectionDAGBuilder::visitICmp(const User &I) { |
2911 | ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE; |
2912 | if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I)) |
2913 | predicate = IC->getPredicate(); |
2914 | else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I)) |
2915 | predicate = ICmpInst::Predicate(IC->getPredicate()); |
2916 | SDValue Op1 = getValue(I.getOperand(0)); |
2917 | SDValue Op2 = getValue(I.getOperand(1)); |
2918 | ISD::CondCode Opcode = getICmpCondCode(predicate); |
2919 | |
2920 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); |
2921 | setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Opcode)); |
2922 | } |
2923 | |
2924 | void SelectionDAGBuilder::visitFCmp(const User &I) { |
2925 | FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE; |
2926 | if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I)) |
2927 | predicate = FC->getPredicate(); |
2928 | else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I)) |
2929 | predicate = FCmpInst::Predicate(FC->getPredicate()); |
2930 | SDValue Op1 = getValue(I.getOperand(0)); |
2931 | SDValue Op2 = getValue(I.getOperand(1)); |
2932 | ISD::CondCode Condition = getFCmpCondCode(predicate); |
2933 | if (TM.Options.NoNaNsFPMath) |
2934 | Condition = getFCmpCodeWithoutNaN(Condition); |
2935 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); |
2936 | setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Condition)); |
2937 | } |
2938 | |
2939 | void SelectionDAGBuilder::visitSelect(const User &I) { |
2940 | SmallVector<EVT, 4> ValueVTs; |
2941 | ComputeValueVTs(DAG.getTargetLoweringInfo(), I.getType(), ValueVTs); |
2942 | unsigned NumValues = ValueVTs.size(); |
2943 | if (NumValues == 0) return; |
2944 | |
2945 | SmallVector<SDValue, 4> Values(NumValues); |
2946 | SDValue Cond = getValue(I.getOperand(0)); |
2947 | SDValue TrueVal = getValue(I.getOperand(1)); |
2948 | SDValue FalseVal = getValue(I.getOperand(2)); |
2949 | ISD::NodeType OpCode = Cond.getValueType().isVector() ? |
2950 | ISD::VSELECT : ISD::SELECT; |
2951 | |
2952 | for (unsigned i = 0; i != NumValues; ++i) |
2953 | Values[i] = DAG.getNode(OpCode, getCurSDLoc(), |
2954 | TrueVal.getNode()->getValueType(TrueVal.getResNo()+i), |
2955 | Cond, |
2956 | SDValue(TrueVal.getNode(), |
2957 | TrueVal.getResNo() + i), |
2958 | SDValue(FalseVal.getNode(), |
2959 | FalseVal.getResNo() + i)); |
2960 | |
2961 | setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), |
2962 | DAG.getVTList(ValueVTs), Values)); |
2963 | } |
2964 | |
2965 | void SelectionDAGBuilder::visitTrunc(const User &I) { |
2966 | // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest). |
2967 | SDValue N = getValue(I.getOperand(0)); |
2968 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); |
2969 | setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), DestVT, N)); |
2970 | } |
2971 | |
2972 | void SelectionDAGBuilder::visitZExt(const User &I) { |
2973 | // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest). |
2974 | // ZExt also can't be a cast to bool for same reason. So, nothing much to do |
2975 | SDValue N = getValue(I.getOperand(0)); |
2976 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); |
2977 | setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurSDLoc(), DestVT, N)); |
2978 | } |
2979 | |
2980 | void SelectionDAGBuilder::visitSExt(const User &I) { |
2981 | // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest). |
2982 | // SExt also can't be a cast to bool for same reason. So, nothing much to do |
2983 | SDValue N = getValue(I.getOperand(0)); |
2984 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); |
2985 | setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurSDLoc(), DestVT, N)); |
2986 | } |
2987 | |
2988 | void SelectionDAGBuilder::visitFPTrunc(const User &I) { |
2989 | // FPTrunc is never a no-op cast, no need to check |
2990 | SDValue N = getValue(I.getOperand(0)); |
2991 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
2992 | EVT DestVT = TLI.getValueType(I.getType()); |
2993 | setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurSDLoc(), DestVT, N, |
2994 | DAG.getTargetConstant(0, TLI.getPointerTy()))); |
2995 | } |
2996 | |
2997 | void SelectionDAGBuilder::visitFPExt(const User &I) { |
2998 | // FPExt is never a no-op cast, no need to check |
2999 | SDValue N = getValue(I.getOperand(0)); |
3000 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); |
3001 | setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurSDLoc(), DestVT, N)); |
3002 | } |
3003 | |
3004 | void SelectionDAGBuilder::visitFPToUI(const User &I) { |
3005 | // FPToUI is never a no-op cast, no need to check |
3006 | SDValue N = getValue(I.getOperand(0)); |
3007 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); |
3008 | setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurSDLoc(), DestVT, N)); |
3009 | } |
3010 | |
3011 | void SelectionDAGBuilder::visitFPToSI(const User &I) { |
3012 | // FPToSI is never a no-op cast, no need to check |
3013 | SDValue N = getValue(I.getOperand(0)); |
3014 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); |
3015 | setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurSDLoc(), DestVT, N)); |
3016 | } |
3017 | |
3018 | void SelectionDAGBuilder::visitUIToFP(const User &I) { |
3019 | // UIToFP is never a no-op cast, no need to check |
3020 | SDValue N = getValue(I.getOperand(0)); |
3021 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); |
3022 | setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurSDLoc(), DestVT, N)); |
3023 | } |
3024 | |
3025 | void SelectionDAGBuilder::visitSIToFP(const User &I) { |
3026 | // SIToFP is never a no-op cast, no need to check |
3027 | SDValue N = getValue(I.getOperand(0)); |
3028 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); |
3029 | setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurSDLoc(), DestVT, N)); |
3030 | } |
3031 | |
3032 | void SelectionDAGBuilder::visitPtrToInt(const User &I) { |
3033 | // What to do depends on the size of the integer and the size of the pointer. |
3034 | // We can either truncate, zero extend, or no-op, accordingly. |
3035 | SDValue N = getValue(I.getOperand(0)); |
3036 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); |
3037 | setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT)); |
3038 | } |
3039 | |
3040 | void SelectionDAGBuilder::visitIntToPtr(const User &I) { |
3041 | // What to do depends on the size of the integer and the size of the pointer. |
3042 | // We can either truncate, zero extend, or no-op, accordingly. |
3043 | SDValue N = getValue(I.getOperand(0)); |
3044 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); |
3045 | setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT)); |
3046 | } |
3047 | |
3048 | void SelectionDAGBuilder::visitBitCast(const User &I) { |
3049 | SDValue N = getValue(I.getOperand(0)); |
3050 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); |
3051 | |
3052 | // BitCast assures us that source and destination are the same size so this is |
3053 | // either a BITCAST or a no-op. |
3054 | if (DestVT != N.getValueType()) |
3055 | setValue(&I, DAG.getNode(ISD::BITCAST, getCurSDLoc(), |
3056 | DestVT, N)); // convert types. |
3057 | // Check if the original LLVM IR Operand was a ConstantInt, because getValue() |
3058 | // might fold any kind of constant expression to an integer constant and that |
3059 | // is not what we are looking for. Only regcognize a bitcast of a genuine |
3060 | // constant integer as an opaque constant. |
3061 | else if(ConstantInt *C = dyn_cast<ConstantInt>(I.getOperand(0))) |
3062 | setValue(&I, DAG.getConstant(C->getValue(), DestVT, /*isTarget=*/false, |
3063 | /*isOpaque*/true)); |
3064 | else |
3065 | setValue(&I, N); // noop cast. |
3066 | } |
3067 | |
3068 | void SelectionDAGBuilder::visitAddrSpaceCast(const User &I) { |
3069 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
3070 | const Value *SV = I.getOperand(0); |
3071 | SDValue N = getValue(SV); |
3072 | EVT DestVT = TLI.getValueType(I.getType()); |
3073 | |
3074 | unsigned SrcAS = SV->getType()->getPointerAddressSpace(); |
3075 | unsigned DestAS = I.getType()->getPointerAddressSpace(); |
3076 | |
3077 | if (!TLI.isNoopAddrSpaceCast(SrcAS, DestAS)) |
3078 | N = DAG.getAddrSpaceCast(getCurSDLoc(), DestVT, N, SrcAS, DestAS); |
3079 | |
3080 | setValue(&I, N); |
3081 | } |
3082 | |
3083 | void SelectionDAGBuilder::visitInsertElement(const User &I) { |
3084 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
3085 | SDValue InVec = getValue(I.getOperand(0)); |
3086 | SDValue InVal = getValue(I.getOperand(1)); |
3087 | SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(2)), |
3088 | getCurSDLoc(), TLI.getVectorIdxTy()); |
3089 | setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurSDLoc(), |
3090 | TLI.getValueType(I.getType()), InVec, InVal, InIdx)); |
3091 | } |
3092 | |
3093 | void SelectionDAGBuilder::visitExtractElement(const User &I) { |
3094 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
3095 | SDValue InVec = getValue(I.getOperand(0)); |
3096 | SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(1)), |
3097 | getCurSDLoc(), TLI.getVectorIdxTy()); |
3098 | setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(), |
3099 | TLI.getValueType(I.getType()), InVec, InIdx)); |
3100 | } |
3101 | |
3102 | // Utility for visitShuffleVector - Return true if every element in Mask, |
3103 | // beginning from position Pos and ending in Pos+Size, falls within the |
3104 | // specified sequential range [L, L+Pos). or is undef. |
3105 | static bool isSequentialInRange(const SmallVectorImpl<int> &Mask, |
3106 | unsigned Pos, unsigned Size, int Low) { |
3107 | for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low) |
3108 | if (Mask[i] >= 0 && Mask[i] != Low) |
3109 | return false; |
3110 | return true; |
3111 | } |
3112 | |
3113 | void SelectionDAGBuilder::visitShuffleVector(const User &I) { |
3114 | SDValue Src1 = getValue(I.getOperand(0)); |
3115 | SDValue Src2 = getValue(I.getOperand(1)); |
3116 | |
3117 | SmallVector<int, 8> Mask; |
3118 | ShuffleVectorInst::getShuffleMask(cast<Constant>(I.getOperand(2)), Mask); |
3119 | unsigned MaskNumElts = Mask.size(); |
3120 | |
3121 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
3122 | EVT VT = TLI.getValueType(I.getType()); |
3123 | EVT SrcVT = Src1.getValueType(); |
3124 | unsigned SrcNumElts = SrcVT.getVectorNumElements(); |
3125 | |
3126 | if (SrcNumElts == MaskNumElts) { |
3127 | setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2, |
3128 | &Mask[0])); |
3129 | return; |
3130 | } |
3131 | |
3132 | // Normalize the shuffle vector since mask and vector length don't match. |
3133 | if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) { |
3134 | // Mask is longer than the source vectors and is a multiple of the source |
3135 | // vectors. We can use concatenate vector to make the mask and vectors |
3136 | // lengths match. |
3137 | if (SrcNumElts*2 == MaskNumElts) { |
3138 | // First check for Src1 in low and Src2 in high |
3139 | if (isSequentialInRange(Mask, 0, SrcNumElts, 0) && |
3140 | isSequentialInRange(Mask, SrcNumElts, SrcNumElts, SrcNumElts)) { |
3141 | // The shuffle is concatenating two vectors together. |
3142 | setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(), |
3143 | VT, Src1, Src2)); |
3144 | return; |
3145 | } |
3146 | // Then check for Src2 in low and Src1 in high |
3147 | if (isSequentialInRange(Mask, 0, SrcNumElts, SrcNumElts) && |
3148 | isSequentialInRange(Mask, SrcNumElts, SrcNumElts, 0)) { |
3149 | // The shuffle is concatenating two vectors together. |
3150 | setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(), |
3151 | VT, Src2, Src1)); |
3152 | return; |
3153 | } |
3154 | } |
3155 | |
3156 | // Pad both vectors with undefs to make them the same length as the mask. |
3157 | unsigned NumConcat = MaskNumElts / SrcNumElts; |
3158 | bool Src1U = Src1.getOpcode() == ISD::UNDEF; |
3159 | bool Src2U = Src2.getOpcode() == ISD::UNDEF; |
3160 | SDValue UndefVal = DAG.getUNDEF(SrcVT); |
3161 | |
3162 | SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal); |
3163 | SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal); |
3164 | MOps1[0] = Src1; |
3165 | MOps2[0] = Src2; |
3166 | |
3167 | Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS, |
3168 | getCurSDLoc(), VT, MOps1); |
3169 | Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS, |
3170 | getCurSDLoc(), VT, MOps2); |
3171 | |
3172 | // Readjust mask for new input vector length. |
3173 | SmallVector<int, 8> MappedOps; |
3174 | for (unsigned i = 0; i != MaskNumElts; ++i) { |
3175 | int Idx = Mask[i]; |
3176 | if (Idx >= (int)SrcNumElts) |
3177 | Idx -= SrcNumElts - MaskNumElts; |
3178 | MappedOps.push_back(Idx); |
3179 | } |
3180 | |
3181 | setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2, |
3182 | &MappedOps[0])); |
3183 | return; |
3184 | } |
3185 | |
3186 | if (SrcNumElts > MaskNumElts) { |
3187 | // Analyze the access pattern of the vector to see if we can extract |
3188 | // two subvectors and do the shuffle. The analysis is done by calculating |
3189 | // the range of elements the mask access on both vectors. |
3190 | int MinRange[2] = { static_cast<int>(SrcNumElts), |
3191 | static_cast<int>(SrcNumElts)}; |
3192 | int MaxRange[2] = {-1, -1}; |
3193 | |
3194 | for (unsigned i = 0; i != MaskNumElts; ++i) { |
3195 | int Idx = Mask[i]; |
3196 | unsigned Input = 0; |
3197 | if (Idx < 0) |
3198 | continue; |
3199 | |
3200 | if (Idx >= (int)SrcNumElts) { |
3201 | Input = 1; |
3202 | Idx -= SrcNumElts; |
3203 | } |
3204 | if (Idx > MaxRange[Input]) |
3205 | MaxRange[Input] = Idx; |
3206 | if (Idx < MinRange[Input]) |
3207 | MinRange[Input] = Idx; |
3208 | } |
3209 | |
3210 | // Check if the access is smaller than the vector size and can we find |
3211 | // a reasonable extract index. |
3212 | int RangeUse[2] = { -1, -1 }; // 0 = Unused, 1 = Extract, -1 = Can not |
3213 | // Extract. |
3214 | int StartIdx[2]; // StartIdx to extract from |
3215 | for (unsigned Input = 0; Input < 2; ++Input) { |
3216 | if (MinRange[Input] >= (int)SrcNumElts && MaxRange[Input] < 0) { |
3217 | RangeUse[Input] = 0; // Unused |
3218 | StartIdx[Input] = 0; |
3219 | continue; |
3220 | } |
3221 | |
3222 | // Find a good start index that is a multiple of the mask length. Then |
3223 | // see if the rest of the elements are in range. |
3224 | StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts; |
3225 | if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts && |
3226 | StartIdx[Input] + MaskNumElts <= SrcNumElts) |
3227 | RangeUse[Input] = 1; // Extract from a multiple of the mask length. |
3228 | } |
3229 | |
3230 | if (RangeUse[0] == 0 && RangeUse[1] == 0) { |
3231 | setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used. |
3232 | return; |
3233 | } |
3234 | if (RangeUse[0] >= 0 && RangeUse[1] >= 0) { |
3235 | // Extract appropriate subvector and generate a vector shuffle |
3236 | for (unsigned Input = 0; Input < 2; ++Input) { |
3237 | SDValue &Src = Input == 0 ? Src1 : Src2; |
3238 | if (RangeUse[Input] == 0) |
3239 | Src = DAG.getUNDEF(VT); |
3240 | else |
3241 | Src = DAG.getNode( |
3242 | ISD::EXTRACT_SUBVECTOR, getCurSDLoc(), VT, Src, |
3243 | DAG.getConstant(StartIdx[Input], TLI.getVectorIdxTy())); |
3244 | } |
3245 | |
3246 | // Calculate new mask. |
3247 | SmallVector<int, 8> MappedOps; |
3248 | for (unsigned i = 0; i != MaskNumElts; ++i) { |
3249 | int Idx = Mask[i]; |
3250 | if (Idx >= 0) { |
3251 | if (Idx < (int)SrcNumElts) |
3252 | Idx -= StartIdx[0]; |
3253 | else |
3254 | Idx -= SrcNumElts + StartIdx[1] - MaskNumElts; |
3255 | } |
3256 | MappedOps.push_back(Idx); |
3257 | } |
3258 | |
3259 | setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2, |
3260 | &MappedOps[0])); |
3261 | return; |
3262 | } |
3263 | } |
3264 | |
3265 | // We can't use either concat vectors or extract subvectors so fall back to |
3266 | // replacing the shuffle with extract and build vector. |
3267 | // to insert and build vector. |
3268 | EVT EltVT = VT.getVectorElementType(); |
3269 | EVT IdxVT = TLI.getVectorIdxTy(); |
3270 | SmallVector<SDValue,8> Ops; |
3271 | for (unsigned i = 0; i != MaskNumElts; ++i) { |
3272 | int Idx = Mask[i]; |
3273 | SDValue Res; |
3274 | |
3275 | if (Idx < 0) { |
3276 | Res = DAG.getUNDEF(EltVT); |
3277 | } else { |
3278 | SDValue &Src = Idx < (int)SrcNumElts ? Src1 : Src2; |
3279 | if (Idx >= (int)SrcNumElts) Idx -= SrcNumElts; |
3280 | |
3281 | Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(), |
3282 | EltVT, Src, DAG.getConstant(Idx, IdxVT)); |
3283 | } |
3284 | |
3285 | Ops.push_back(Res); |
3286 | } |
3287 | |
3288 | setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), VT, Ops)); |
3289 | } |
3290 | |
3291 | void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) { |
3292 | const Value *Op0 = I.getOperand(0); |
3293 | const Value *Op1 = I.getOperand(1); |
3294 | Type *AggTy = I.getType(); |
3295 | Type *ValTy = Op1->getType(); |
3296 | bool IntoUndef = isa<UndefValue>(Op0); |
3297 | bool FromUndef = isa<UndefValue>(Op1); |
3298 | |
3299 | unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices()); |
3300 | |
3301 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
3302 | SmallVector<EVT, 4> AggValueVTs; |
3303 | ComputeValueVTs(TLI, AggTy, AggValueVTs); |
3304 | SmallVector<EVT, 4> ValValueVTs; |
3305 | ComputeValueVTs(TLI, ValTy, ValValueVTs); |
3306 | |
3307 | unsigned NumAggValues = AggValueVTs.size(); |
3308 | unsigned NumValValues = ValValueVTs.size(); |
3309 | SmallVector<SDValue, 4> Values(NumAggValues); |
3310 | |
3311 | // Ignore an insertvalue that produces an empty object |
3312 | if (!NumAggValues) { |
3313 | setValue(&I, DAG.getUNDEF(MVT(MVT::Other))); |
3314 | return; |
3315 | } |
3316 | |
3317 | SDValue Agg = getValue(Op0); |
3318 | unsigned i = 0; |
3319 | // Copy the beginning value(s) from the original aggregate. |
3320 | for (; i != LinearIndex; ++i) |
3321 | Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) : |
3322 | SDValue(Agg.getNode(), Agg.getResNo() + i); |
3323 | // Copy values from the inserted value(s). |
3324 | if (NumValValues) { |
3325 | SDValue Val = getValue(Op1); |
3326 | for (; i != LinearIndex + NumValValues; ++i) |
3327 | Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) : |
3328 | SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex); |
3329 | } |
3330 | // Copy remaining value(s) from the original aggregate. |
3331 | for (; i != NumAggValues; ++i) |
3332 | Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) : |
3333 | SDValue(Agg.getNode(), Agg.getResNo() + i); |
3334 | |
3335 | setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), |
3336 | DAG.getVTList(AggValueVTs), Values)); |
3337 | } |
3338 | |
3339 | void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) { |
3340 | const Value *Op0 = I.getOperand(0); |
3341 | Type *AggTy = Op0->getType(); |
3342 | Type *ValTy = I.getType(); |
3343 | bool OutOfUndef = isa<UndefValue>(Op0); |
3344 | |
3345 | unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices()); |
3346 | |
3347 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
3348 | SmallVector<EVT, 4> ValValueVTs; |
3349 | ComputeValueVTs(TLI, ValTy, ValValueVTs); |
3350 | |
3351 | unsigned NumValValues = ValValueVTs.size(); |
3352 | |
3353 | // Ignore a extractvalue that produces an empty object |
3354 | if (!NumValValues) { |
3355 | setValue(&I, DAG.getUNDEF(MVT(MVT::Other))); |
3356 | return; |
3357 | } |
3358 | |
3359 | SmallVector<SDValue, 4> Values(NumValValues); |
3360 | |
3361 | SDValue Agg = getValue(Op0); |
3362 | // Copy out the selected value(s). |
3363 | for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i) |
3364 | Values[i - LinearIndex] = |
3365 | OutOfUndef ? |
3366 | DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) : |
3367 | SDValue(Agg.getNode(), Agg.getResNo() + i); |
3368 | |
3369 | setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), |
3370 | DAG.getVTList(ValValueVTs), Values)); |
3371 | } |
3372 | |
3373 | void SelectionDAGBuilder::visitGetElementPtr(const User &I) { |
3374 | Value *Op0 = I.getOperand(0); |
3375 | // Note that the pointer operand may be a vector of pointers. Take the scalar |
3376 | // element which holds a pointer. |
3377 | Type *Ty = Op0->getType()->getScalarType(); |
3378 | unsigned AS = Ty->getPointerAddressSpace(); |
3379 | SDValue N = getValue(Op0); |
3380 | |
3381 | for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end(); |
3382 | OI != E; ++OI) { |
3383 | const Value *Idx = *OI; |
3384 | if (StructType *StTy = dyn_cast<StructType>(Ty)) { |
3385 | unsigned Field = cast<Constant>(Idx)->getUniqueInteger().getZExtValue(); |
3386 | if (Field) { |
3387 | // N = N + Offset |
3388 | uint64_t Offset = DL->getStructLayout(StTy)->getElementOffset(Field); |
3389 | N = DAG.getNode(ISD::ADD, getCurSDLoc(), N.getValueType(), N, |
3390 | DAG.getConstant(Offset, N.getValueType())); |
3391 | } |
3392 | |
3393 | Ty = StTy->getElementType(Field); |
3394 | } else { |
3395 | Ty = cast<SequentialType>(Ty)->getElementType(); |
3396 | |
3397 | // If this is a constant subscript, handle it quickly. |
3398 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
3399 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) { |
3400 | if (CI->isZero()) continue; |
3401 | uint64_t Offs = |
3402 | DL->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue(); |
3403 | SDValue OffsVal; |
3404 | EVT PTy = TLI.getPointerTy(AS); |
3405 | unsigned PtrBits = PTy.getSizeInBits(); |
3406 | if (PtrBits < 64) |
3407 | OffsVal = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), PTy, |
3408 | DAG.getConstant(Offs, MVT::i64)); |
3409 | else |
3410 | OffsVal = DAG.getConstant(Offs, PTy); |
3411 | |
3412 | N = DAG.getNode(ISD::ADD, getCurSDLoc(), N.getValueType(), N, |
3413 | OffsVal); |
3414 | continue; |
3415 | } |
3416 | |
3417 | // N = N + Idx * ElementSize; |
3418 | APInt ElementSize = |
3419 | APInt(TLI.getPointerSizeInBits(AS), DL->getTypeAllocSize(Ty)); |
3420 | SDValue IdxN = getValue(Idx); |
3421 | |
3422 | // If the index is smaller or larger than intptr_t, truncate or extend |
3423 | // it. |
3424 | IdxN = DAG.getSExtOrTrunc(IdxN, getCurSDLoc(), N.getValueType()); |
3425 | |
3426 | // If this is a multiply by a power of two, turn it into a shl |
3427 | // immediately. This is a very common case. |
3428 | if (ElementSize != 1) { |
3429 | if (ElementSize.isPowerOf2()) { |
3430 | unsigned Amt = ElementSize.logBase2(); |
3431 | IdxN = DAG.getNode(ISD::SHL, getCurSDLoc(), |
3432 | N.getValueType(), IdxN, |
3433 | DAG.getConstant(Amt, IdxN.getValueType())); |
3434 | } else { |
3435 | SDValue Scale = DAG.getConstant(ElementSize, IdxN.getValueType()); |
3436 | IdxN = DAG.getNode(ISD::MUL, getCurSDLoc(), |
3437 | N.getValueType(), IdxN, Scale); |
3438 | } |
3439 | } |
3440 | |
3441 | N = DAG.getNode(ISD::ADD, getCurSDLoc(), |
3442 | N.getValueType(), N, IdxN); |
3443 | } |
3444 | } |
3445 | |
3446 | setValue(&I, N); |
3447 | } |
3448 | |
3449 | void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) { |
3450 | // If this is a fixed sized alloca in the entry block of the function, |
3451 | // allocate it statically on the stack. |
3452 | if (FuncInfo.StaticAllocaMap.count(&I)) |
3453 | return; // getValue will auto-populate this. |
3454 | |
3455 | Type *Ty = I.getAllocatedType(); |
3456 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
3457 | uint64_t TySize = TLI.getDataLayout()->getTypeAllocSize(Ty); |
3458 | unsigned Align = |
3459 | std::max((unsigned)TLI.getDataLayout()->getPrefTypeAlignment(Ty), |
3460 | I.getAlignment()); |
3461 | |
3462 | SDValue AllocSize = getValue(I.getArraySize()); |
3463 | |
3464 | EVT IntPtr = TLI.getPointerTy(); |
3465 | if (AllocSize.getValueType() != IntPtr) |
3466 | AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurSDLoc(), IntPtr); |
3467 | |
3468 | AllocSize = DAG.getNode(ISD::MUL, getCurSDLoc(), IntPtr, |
3469 | AllocSize, |
3470 | DAG.getConstant(TySize, IntPtr)); |
3471 | |
3472 | // Handle alignment. If the requested alignment is less than or equal to |
3473 | // the stack alignment, ignore it. If the size is greater than or equal to |
3474 | // the stack alignment, we note this in the DYNAMIC_STACKALLOC node. |
3475 | unsigned StackAlign = |
3476 | DAG.getSubtarget().getFrameLowering()->getStackAlignment(); |
3477 | if (Align <= StackAlign) |
3478 | Align = 0; |
3479 | |
3480 | // Round the size of the allocation up to the stack alignment size |
3481 | // by add SA-1 to the size. |
3482 | AllocSize = DAG.getNode(ISD::ADD, getCurSDLoc(), |
3483 | AllocSize.getValueType(), AllocSize, |
3484 | DAG.getIntPtrConstant(StackAlign-1)); |
3485 | |
3486 | // Mask out the low bits for alignment purposes. |
3487 | AllocSize = DAG.getNode(ISD::AND, getCurSDLoc(), |
3488 | AllocSize.getValueType(), AllocSize, |
3489 | DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1))); |
3490 | |
3491 | SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) }; |
3492 | SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other); |
3493 | SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurSDLoc(), VTs, Ops); |
3494 | setValue(&I, DSA); |
3495 | DAG.setRoot(DSA.getValue(1)); |
3496 | |
3497 | assert(FuncInfo.MF->getFrameInfo()->hasVarSizedObjects())((FuncInfo.MF->getFrameInfo()->hasVarSizedObjects()) ? static_cast <void> (0) : __assert_fail ("FuncInfo.MF->getFrameInfo()->hasVarSizedObjects()" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 3497, __PRETTY_FUNCTION__)); |
3498 | } |
3499 | |
3500 | void SelectionDAGBuilder::visitLoad(const LoadInst &I) { |
3501 | if (I.isAtomic()) |
3502 | return visitAtomicLoad(I); |
3503 | |
3504 | const Value *SV = I.getOperand(0); |
3505 | SDValue Ptr = getValue(SV); |
3506 | |
3507 | Type *Ty = I.getType(); |
3508 | |
3509 | bool isVolatile = I.isVolatile(); |
3510 | bool isNonTemporal = I.getMetadata(LLVMContext::MD_nontemporal) != nullptr; |
3511 | bool isInvariant = I.getMetadata(LLVMContext::MD_invariant_load) != nullptr; |
3512 | unsigned Alignment = I.getAlignment(); |
3513 | |
3514 | AAMDNodes AAInfo; |
3515 | I.getAAMetadata(AAInfo); |
3516 | const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range); |
3517 | |
3518 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
3519 | SmallVector<EVT, 4> ValueVTs; |
3520 | SmallVector<uint64_t, 4> Offsets; |
3521 | ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets); |
3522 | unsigned NumValues = ValueVTs.size(); |
3523 | if (NumValues == 0) |
3524 | return; |
3525 | |
3526 | SDValue Root; |
3527 | bool ConstantMemory = false; |
3528 | if (isVolatile || NumValues > MaxParallelChains) |
3529 | // Serialize volatile loads with other side effects. |
3530 | Root = getRoot(); |
3531 | else if (AA->pointsToConstantMemory( |
3532 | AliasAnalysis::Location(SV, AA->getTypeStoreSize(Ty), AAInfo))) { |
3533 | // Do not serialize (non-volatile) loads of constant memory with anything. |
3534 | Root = DAG.getEntryNode(); |
3535 | ConstantMemory = true; |
3536 | } else { |
3537 | // Do not serialize non-volatile loads against each other. |
3538 | Root = DAG.getRoot(); |
3539 | } |
3540 | |
3541 | if (isVolatile) |
3542 | Root = TLI.prepareVolatileOrAtomicLoad(Root, getCurSDLoc(), DAG); |
3543 | |
3544 | SmallVector<SDValue, 4> Values(NumValues); |
3545 | SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains), |
3546 | NumValues)); |
3547 | EVT PtrVT = Ptr.getValueType(); |
3548 | unsigned ChainI = 0; |
3549 | for (unsigned i = 0; i != NumValues; ++i, ++ChainI) { |
3550 | // Serializing loads here may result in excessive register pressure, and |
3551 | // TokenFactor places arbitrary choke points on the scheduler. SD scheduling |
3552 | // could recover a bit by hoisting nodes upward in the chain by recognizing |
3553 | // they are side-effect free or do not alias. The optimizer should really |
3554 | // avoid this case by converting large object/array copies to llvm.memcpy |
3555 | // (MaxParallelChains should always remain as failsafe). |
3556 | if (ChainI == MaxParallelChains) { |
3557 | assert(PendingLoads.empty() && "PendingLoads must be serialized first")((PendingLoads.empty() && "PendingLoads must be serialized first" ) ? static_cast<void> (0) : __assert_fail ("PendingLoads.empty() && \"PendingLoads must be serialized first\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 3557, __PRETTY_FUNCTION__)); |
3558 | SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, |
3559 | makeArrayRef(Chains.data(), ChainI)); |
3560 | Root = Chain; |
3561 | ChainI = 0; |
3562 | } |
3563 | SDValue A = DAG.getNode(ISD::ADD, getCurSDLoc(), |
3564 | PtrVT, Ptr, |
3565 | DAG.getConstant(Offsets[i], PtrVT)); |
3566 | SDValue L = DAG.getLoad(ValueVTs[i], getCurSDLoc(), Root, |
3567 | A, MachinePointerInfo(SV, Offsets[i]), isVolatile, |
3568 | isNonTemporal, isInvariant, Alignment, AAInfo, |
3569 | Ranges); |
3570 | |
3571 | Values[i] = L; |
3572 | Chains[ChainI] = L.getValue(1); |
3573 | } |
3574 | |
3575 | if (!ConstantMemory) { |
3576 | SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, |
3577 | makeArrayRef(Chains.data(), ChainI)); |
3578 | if (isVolatile) |
3579 | DAG.setRoot(Chain); |
3580 | else |
3581 | PendingLoads.push_back(Chain); |
3582 | } |
3583 | |
3584 | setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), |
3585 | DAG.getVTList(ValueVTs), Values)); |
3586 | } |
3587 | |
3588 | void SelectionDAGBuilder::visitStore(const StoreInst &I) { |
3589 | if (I.isAtomic()) |
3590 | return visitAtomicStore(I); |
3591 | |
3592 | const Value *SrcV = I.getOperand(0); |
3593 | const Value *PtrV = I.getOperand(1); |
3594 | |
3595 | SmallVector<EVT, 4> ValueVTs; |
3596 | SmallVector<uint64_t, 4> Offsets; |
3597 | ComputeValueVTs(DAG.getTargetLoweringInfo(), SrcV->getType(), |
3598 | ValueVTs, &Offsets); |
3599 | unsigned NumValues = ValueVTs.size(); |
3600 | if (NumValues == 0) |
3601 | return; |
3602 | |
3603 | // Get the lowered operands. Note that we do this after |
3604 | // checking if NumResults is zero, because with zero results |
3605 | // the operands won't have values in the map. |
3606 | SDValue Src = getValue(SrcV); |
3607 | SDValue Ptr = getValue(PtrV); |
3608 | |
3609 | SDValue Root = getRoot(); |
3610 | SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains), |
3611 | NumValues)); |
3612 | EVT PtrVT = Ptr.getValueType(); |
3613 | bool isVolatile = I.isVolatile(); |
3614 | bool isNonTemporal = I.getMetadata(LLVMContext::MD_nontemporal) != nullptr; |
3615 | unsigned Alignment = I.getAlignment(); |
3616 | |
3617 | AAMDNodes AAInfo; |
3618 | I.getAAMetadata(AAInfo); |
3619 | |
3620 | unsigned ChainI = 0; |
3621 | for (unsigned i = 0; i != NumValues; ++i, ++ChainI) { |
3622 | // See visitLoad comments. |
3623 | if (ChainI == MaxParallelChains) { |
3624 | SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, |
3625 | makeArrayRef(Chains.data(), ChainI)); |
3626 | Root = Chain; |
3627 | ChainI = 0; |
3628 | } |
3629 | SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(), PtrVT, Ptr, |
3630 | DAG.getConstant(Offsets[i], PtrVT)); |
3631 | SDValue St = DAG.getStore(Root, getCurSDLoc(), |
3632 | SDValue(Src.getNode(), Src.getResNo() + i), |
3633 | Add, MachinePointerInfo(PtrV, Offsets[i]), |
3634 | isVolatile, isNonTemporal, Alignment, AAInfo); |
3635 | Chains[ChainI] = St; |
3636 | } |
3637 | |
3638 | SDValue StoreNode = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, |
3639 | makeArrayRef(Chains.data(), ChainI)); |
3640 | DAG.setRoot(StoreNode); |
3641 | } |
3642 | |
3643 | void SelectionDAGBuilder::visitMaskedStore(const CallInst &I) { |
3644 | SDLoc sdl = getCurSDLoc(); |
3645 | |
3646 | Value *PtrOperand = I.getArgOperand(0); |
3647 | SDValue Ptr = getValue(PtrOperand); |
3648 | SDValue Src0 = getValue(I.getArgOperand(1)); |
3649 | SDValue Mask = getValue(I.getArgOperand(3)); |
3650 | EVT VT = Src0.getValueType(); |
3651 | unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(2)))->getZExtValue(); |
3652 | if (!Alignment) |
3653 | Alignment = DAG.getEVTAlignment(VT); |
3654 | |
3655 | AAMDNodes AAInfo; |
3656 | I.getAAMetadata(AAInfo); |
3657 | |
3658 | MachineMemOperand *MMO = |
3659 | DAG.getMachineFunction(). |
3660 | getMachineMemOperand(MachinePointerInfo(PtrOperand), |
3661 | MachineMemOperand::MOStore, VT.getStoreSize(), |
3662 | Alignment, AAInfo); |
3663 | SDValue StoreNode = DAG.getMaskedStore(getRoot(), sdl, Src0, Ptr, Mask, MMO); |
3664 | DAG.setRoot(StoreNode); |
3665 | setValue(&I, StoreNode); |
3666 | } |
3667 | |
3668 | void SelectionDAGBuilder::visitMaskedLoad(const CallInst &I) { |
3669 | SDLoc sdl = getCurSDLoc(); |
3670 | |
3671 | Value *PtrOperand = I.getArgOperand(0); |
3672 | SDValue Ptr = getValue(PtrOperand); |
3673 | SDValue Src0 = getValue(I.getArgOperand(1)); |
3674 | SDValue Mask = getValue(I.getArgOperand(3)); |
3675 | |
3676 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
3677 | EVT VT = TLI.getValueType(I.getType()); |
3678 | unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(2)))->getZExtValue(); |
3679 | if (!Alignment) |
3680 | Alignment = DAG.getEVTAlignment(VT); |
3681 | |
3682 | AAMDNodes AAInfo; |
3683 | I.getAAMetadata(AAInfo); |
3684 | const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range); |
3685 | |
3686 | SDValue InChain = DAG.getRoot(); |
3687 | if (AA->pointsToConstantMemory( |
3688 | AliasAnalysis::Location(PtrOperand, |
3689 | AA->getTypeStoreSize(I.getType()), |
3690 | AAInfo))) { |
3691 | // Do not serialize (non-volatile) loads of constant memory with anything. |
3692 | InChain = DAG.getEntryNode(); |
3693 | } |
3694 | |
3695 | MachineMemOperand *MMO = |
3696 | DAG.getMachineFunction(). |
3697 | getMachineMemOperand(MachinePointerInfo(PtrOperand), |
3698 | MachineMemOperand::MOLoad, VT.getStoreSize(), |
3699 | Alignment, AAInfo, Ranges); |
3700 | |
3701 | SDValue Load = DAG.getMaskedLoad(VT, sdl, InChain, Ptr, Mask, Src0, MMO); |
3702 | SDValue OutChain = Load.getValue(1); |
3703 | DAG.setRoot(OutChain); |
3704 | setValue(&I, Load); |
3705 | } |
3706 | |
3707 | void SelectionDAGBuilder::visitAtomicCmpXchg(const AtomicCmpXchgInst &I) { |
3708 | SDLoc dl = getCurSDLoc(); |
3709 | AtomicOrdering SuccessOrder = I.getSuccessOrdering(); |
3710 | AtomicOrdering FailureOrder = I.getFailureOrdering(); |
3711 | SynchronizationScope Scope = I.getSynchScope(); |
3712 | |
3713 | SDValue InChain = getRoot(); |
3714 | |
3715 | MVT MemVT = getValue(I.getCompareOperand()).getSimpleValueType(); |
3716 | SDVTList VTs = DAG.getVTList(MemVT, MVT::i1, MVT::Other); |
3717 | SDValue L = DAG.getAtomicCmpSwap( |
3718 | ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, dl, MemVT, VTs, InChain, |
3719 | getValue(I.getPointerOperand()), getValue(I.getCompareOperand()), |
3720 | getValue(I.getNewValOperand()), MachinePointerInfo(I.getPointerOperand()), |
3721 | /*Alignment=*/ 0, SuccessOrder, FailureOrder, Scope); |
3722 | |
3723 | SDValue OutChain = L.getValue(2); |
3724 | |
3725 | setValue(&I, L); |
3726 | DAG.setRoot(OutChain); |
3727 | } |
3728 | |
3729 | void SelectionDAGBuilder::visitAtomicRMW(const AtomicRMWInst &I) { |
3730 | SDLoc dl = getCurSDLoc(); |
3731 | ISD::NodeType NT; |
3732 | switch (I.getOperation()) { |
3733 | default: llvm_unreachable("Unknown atomicrmw operation")::llvm::llvm_unreachable_internal("Unknown atomicrmw operation" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 3733); |
3734 | case AtomicRMWInst::Xchg: NT = ISD::ATOMIC_SWAP; break; |
3735 | case AtomicRMWInst::Add: NT = ISD::ATOMIC_LOAD_ADD; break; |
3736 | case AtomicRMWInst::Sub: NT = ISD::ATOMIC_LOAD_SUB; break; |
3737 | case AtomicRMWInst::And: NT = ISD::ATOMIC_LOAD_AND; break; |
3738 | case AtomicRMWInst::Nand: NT = ISD::ATOMIC_LOAD_NAND; break; |
3739 | case AtomicRMWInst::Or: NT = ISD::ATOMIC_LOAD_OR; break; |
3740 | case AtomicRMWInst::Xor: NT = ISD::ATOMIC_LOAD_XOR; break; |
3741 | case AtomicRMWInst::Max: NT = ISD::ATOMIC_LOAD_MAX; break; |
3742 | case AtomicRMWInst::Min: NT = ISD::ATOMIC_LOAD_MIN; break; |
3743 | case AtomicRMWInst::UMax: NT = ISD::ATOMIC_LOAD_UMAX; break; |
3744 | case AtomicRMWInst::UMin: NT = ISD::ATOMIC_LOAD_UMIN; break; |
3745 | } |
3746 | AtomicOrdering Order = I.getOrdering(); |
3747 | SynchronizationScope Scope = I.getSynchScope(); |
3748 | |
3749 | SDValue InChain = getRoot(); |
3750 | |
3751 | SDValue L = |
3752 | DAG.getAtomic(NT, dl, |
3753 | getValue(I.getValOperand()).getSimpleValueType(), |
3754 | InChain, |
3755 | getValue(I.getPointerOperand()), |
3756 | getValue(I.getValOperand()), |
3757 | I.getPointerOperand(), |
3758 | /* Alignment=*/ 0, Order, Scope); |
3759 | |
3760 | SDValue OutChain = L.getValue(1); |
3761 | |
3762 | setValue(&I, L); |
3763 | DAG.setRoot(OutChain); |
3764 | } |
3765 | |
3766 | void SelectionDAGBuilder::visitFence(const FenceInst &I) { |
3767 | SDLoc dl = getCurSDLoc(); |
3768 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
3769 | SDValue Ops[3]; |
3770 | Ops[0] = getRoot(); |
3771 | Ops[1] = DAG.getConstant(I.getOrdering(), TLI.getPointerTy()); |
3772 | Ops[2] = DAG.getConstant(I.getSynchScope(), TLI.getPointerTy()); |
3773 | DAG.setRoot(DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops)); |
3774 | } |
3775 | |
3776 | void SelectionDAGBuilder::visitAtomicLoad(const LoadInst &I) { |
3777 | SDLoc dl = getCurSDLoc(); |
3778 | AtomicOrdering Order = I.getOrdering(); |
3779 | SynchronizationScope Scope = I.getSynchScope(); |
3780 | |
3781 | SDValue InChain = getRoot(); |
3782 | |
3783 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
3784 | EVT VT = TLI.getValueType(I.getType()); |
3785 | |
3786 | if (I.getAlignment() < VT.getSizeInBits() / 8) |
3787 | report_fatal_error("Cannot generate unaligned atomic load"); |
3788 | |
3789 | MachineMemOperand *MMO = |
3790 | DAG.getMachineFunction(). |
3791 | getMachineMemOperand(MachinePointerInfo(I.getPointerOperand()), |
3792 | MachineMemOperand::MOVolatile | |
3793 | MachineMemOperand::MOLoad, |
3794 | VT.getStoreSize(), |
3795 | I.getAlignment() ? I.getAlignment() : |
3796 | DAG.getEVTAlignment(VT)); |
3797 | |
3798 | InChain = TLI.prepareVolatileOrAtomicLoad(InChain, dl, DAG); |
3799 | SDValue L = |
3800 | DAG.getAtomic(ISD::ATOMIC_LOAD, dl, VT, VT, InChain, |
3801 | getValue(I.getPointerOperand()), MMO, |
3802 | Order, Scope); |
3803 | |
3804 | SDValue OutChain = L.getValue(1); |
3805 | |
3806 | setValue(&I, L); |
3807 | DAG.setRoot(OutChain); |
3808 | } |
3809 | |
3810 | void SelectionDAGBuilder::visitAtomicStore(const StoreInst &I) { |
3811 | SDLoc dl = getCurSDLoc(); |
3812 | |
3813 | AtomicOrdering Order = I.getOrdering(); |
3814 | SynchronizationScope Scope = I.getSynchScope(); |
3815 | |
3816 | SDValue InChain = getRoot(); |
3817 | |
3818 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
3819 | EVT VT = TLI.getValueType(I.getValueOperand()->getType()); |
3820 | |
3821 | if (I.getAlignment() < VT.getSizeInBits() / 8) |
3822 | report_fatal_error("Cannot generate unaligned atomic store"); |
3823 | |
3824 | SDValue OutChain = |
3825 | DAG.getAtomic(ISD::ATOMIC_STORE, dl, VT, |
3826 | InChain, |
3827 | getValue(I.getPointerOperand()), |
3828 | getValue(I.getValueOperand()), |
3829 | I.getPointerOperand(), I.getAlignment(), |
3830 | Order, Scope); |
3831 | |
3832 | DAG.setRoot(OutChain); |
3833 | } |
3834 | |
3835 | /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC |
3836 | /// node. |
3837 | void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I, |
3838 | unsigned Intrinsic) { |
3839 | bool HasChain = !I.doesNotAccessMemory(); |
3840 | bool OnlyLoad = HasChain && I.onlyReadsMemory(); |
3841 | |
3842 | // Build the operand list. |
3843 | SmallVector<SDValue, 8> Ops; |
3844 | if (HasChain) { // If this intrinsic has side-effects, chainify it. |
3845 | if (OnlyLoad) { |
3846 | // We don't need to serialize loads against other loads. |
3847 | Ops.push_back(DAG.getRoot()); |
3848 | } else { |
3849 | Ops.push_back(getRoot()); |
3850 | } |
3851 | } |
3852 | |
3853 | // Info is set by getTgtMemInstrinsic |
3854 | TargetLowering::IntrinsicInfo Info; |
3855 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
3856 | bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic); |
3857 | |
3858 | // Add the intrinsic ID as an integer operand if it's not a target intrinsic. |
3859 | if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID || |
3860 | Info.opc == ISD::INTRINSIC_W_CHAIN) |
3861 | Ops.push_back(DAG.getTargetConstant(Intrinsic, TLI.getPointerTy())); |
3862 | |
3863 | // Add all operands of the call to the operand list. |
3864 | for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) { |
3865 | SDValue Op = getValue(I.getArgOperand(i)); |
3866 | Ops.push_back(Op); |
3867 | } |
3868 | |
3869 | SmallVector<EVT, 4> ValueVTs; |
3870 | ComputeValueVTs(TLI, I.getType(), ValueVTs); |
3871 | |
3872 | if (HasChain) |
3873 | ValueVTs.push_back(MVT::Other); |
3874 | |
3875 | SDVTList VTs = DAG.getVTList(ValueVTs); |
3876 | |
3877 | // Create the node. |
3878 | SDValue Result; |
3879 | if (IsTgtIntrinsic) { |
3880 | // This is target intrinsic that touches memory |
3881 | Result = DAG.getMemIntrinsicNode(Info.opc, getCurSDLoc(), |
3882 | VTs, Ops, Info.memVT, |
3883 | MachinePointerInfo(Info.ptrVal, Info.offset), |
3884 | Info.align, Info.vol, |
3885 | Info.readMem, Info.writeMem, Info.size); |
3886 | } else if (!HasChain) { |
3887 | Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurSDLoc(), VTs, Ops); |
3888 | } else if (!I.getType()->isVoidTy()) { |
3889 | Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurSDLoc(), VTs, Ops); |
3890 | } else { |
3891 | Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurSDLoc(), VTs, Ops); |
3892 | } |
3893 | |
3894 | if (HasChain) { |
3895 | SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1); |
3896 | if (OnlyLoad) |
3897 | PendingLoads.push_back(Chain); |
3898 | else |
3899 | DAG.setRoot(Chain); |
3900 | } |
3901 | |
3902 | if (!I.getType()->isVoidTy()) { |
3903 | if (VectorType *PTy = dyn_cast<VectorType>(I.getType())) { |
3904 | EVT VT = TLI.getValueType(PTy); |
3905 | Result = DAG.getNode(ISD::BITCAST, getCurSDLoc(), VT, Result); |
3906 | } |
3907 | |
3908 | setValue(&I, Result); |
3909 | } |
3910 | } |
3911 | |
3912 | /// GetSignificand - Get the significand and build it into a floating-point |
3913 | /// number with exponent of 1: |
3914 | /// |
3915 | /// Op = (Op & 0x007fffff) | 0x3f800000; |
3916 | /// |
3917 | /// where Op is the hexadecimal representation of floating point value. |
3918 | static SDValue |
3919 | GetSignificand(SelectionDAG &DAG, SDValue Op, SDLoc dl) { |
3920 | SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op, |
3921 | DAG.getConstant(0x007fffff, MVT::i32)); |
3922 | SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1, |
3923 | DAG.getConstant(0x3f800000, MVT::i32)); |
3924 | return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2); |
3925 | } |
3926 | |
3927 | /// GetExponent - Get the exponent: |
3928 | /// |
3929 | /// (float)(int)(((Op & 0x7f800000) >> 23) - 127); |
3930 | /// |
3931 | /// where Op is the hexadecimal representation of floating point value. |
3932 | static SDValue |
3933 | GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI, |
3934 | SDLoc dl) { |
3935 | SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op, |
3936 | DAG.getConstant(0x7f800000, MVT::i32)); |
3937 | SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0, |
3938 | DAG.getConstant(23, TLI.getPointerTy())); |
3939 | SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1, |
3940 | DAG.getConstant(127, MVT::i32)); |
3941 | return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2); |
3942 | } |
3943 | |
3944 | /// getF32Constant - Get 32-bit floating point constant. |
3945 | static SDValue |
3946 | getF32Constant(SelectionDAG &DAG, unsigned Flt) { |
3947 | return DAG.getConstantFP(APFloat(APFloat::IEEEsingle, APInt(32, Flt)), |
3948 | MVT::f32); |
3949 | } |
3950 | |
3951 | /// expandExp - Lower an exp intrinsic. Handles the special sequences for |
3952 | /// limited-precision mode. |
3953 | static SDValue expandExp(SDLoc dl, SDValue Op, SelectionDAG &DAG, |
3954 | const TargetLowering &TLI) { |
3955 | if (Op.getValueType() == MVT::f32 && |
3956 | LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { |
3957 | |
3958 | // Put the exponent in the right bit position for later addition to the |
3959 | // final result: |
3960 | // |
3961 | // #define LOG2OFe 1.4426950f |
3962 | // IntegerPartOfX = ((int32_t)(X * LOG2OFe)); |
3963 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op, |
3964 | getF32Constant(DAG, 0x3fb8aa3b)); |
3965 | SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0); |
3966 | |
3967 | // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX; |
3968 | SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX); |
3969 | SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1); |
3970 | |
3971 | // IntegerPartOfX <<= 23; |
3972 | IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX, |
3973 | DAG.getConstant(23, TLI.getPointerTy())); |
3974 | |
3975 | SDValue TwoToFracPartOfX; |
3976 | if (LimitFloatPrecision <= 6) { |
3977 | // For floating-point precision of 6: |
3978 | // |
3979 | // TwoToFractionalPartOfX = |
3980 | // 0.997535578f + |
3981 | // (0.735607626f + 0.252464424f * x) * x; |
3982 | // |
3983 | // error 0.0144103317, which is 6 bits |
3984 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
3985 | getF32Constant(DAG, 0x3e814304)); |
3986 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, |
3987 | getF32Constant(DAG, 0x3f3c50c8)); |
3988 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); |
3989 | TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, |
3990 | getF32Constant(DAG, 0x3f7f5e7e)); |
3991 | } else if (LimitFloatPrecision <= 12) { |
3992 | // For floating-point precision of 12: |
3993 | // |
3994 | // TwoToFractionalPartOfX = |
3995 | // 0.999892986f + |
3996 | // (0.696457318f + |
3997 | // (0.224338339f + 0.792043434e-1f * x) * x) * x; |
3998 | // |
3999 | // 0.000107046256 error, which is 13 to 14 bits |
4000 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4001 | getF32Constant(DAG, 0x3da235e3)); |
4002 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, |
4003 | getF32Constant(DAG, 0x3e65b8f3)); |
4004 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); |
4005 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, |
4006 | getF32Constant(DAG, 0x3f324b07)); |
4007 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); |
4008 | TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, |
4009 | getF32Constant(DAG, 0x3f7ff8fd)); |
4010 | } else { // LimitFloatPrecision <= 18 |
4011 | // For floating-point precision of 18: |
4012 | // |
4013 | // TwoToFractionalPartOfX = |
4014 | // 0.999999982f + |
4015 | // (0.693148872f + |
4016 | // (0.240227044f + |
4017 | // (0.554906021e-1f + |
4018 | // (0.961591928e-2f + |
4019 | // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; |
4020 | // |
4021 | // error 2.47208000*10^(-7), which is better than 18 bits |
4022 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4023 | getF32Constant(DAG, 0x3924b03e)); |
4024 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, |
4025 | getF32Constant(DAG, 0x3ab24b87)); |
4026 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); |
4027 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, |
4028 | getF32Constant(DAG, 0x3c1d8c17)); |
4029 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); |
4030 | SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, |
4031 | getF32Constant(DAG, 0x3d634a1d)); |
4032 | SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); |
4033 | SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, |
4034 | getF32Constant(DAG, 0x3e75fe14)); |
4035 | SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); |
4036 | SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10, |
4037 | getF32Constant(DAG, 0x3f317234)); |
4038 | SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X); |
4039 | TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12, |
4040 | getF32Constant(DAG, 0x3f800000)); |
4041 | } |
4042 | |
4043 | // Add the exponent into the result in integer domain. |
4044 | SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, TwoToFracPartOfX); |
4045 | return DAG.getNode(ISD::BITCAST, dl, MVT::f32, |
4046 | DAG.getNode(ISD::ADD, dl, MVT::i32, |
4047 | t13, IntegerPartOfX)); |
4048 | } |
4049 | |
4050 | // No special expansion. |
4051 | return DAG.getNode(ISD::FEXP, dl, Op.getValueType(), Op); |
4052 | } |
4053 | |
4054 | /// expandLog - Lower a log intrinsic. Handles the special sequences for |
4055 | /// limited-precision mode. |
4056 | static SDValue expandLog(SDLoc dl, SDValue Op, SelectionDAG &DAG, |
4057 | const TargetLowering &TLI) { |
4058 | if (Op.getValueType() == MVT::f32 && |
4059 | LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { |
4060 | SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op); |
4061 | |
4062 | // Scale the exponent by log(2) [0.69314718f]. |
4063 | SDValue Exp = GetExponent(DAG, Op1, TLI, dl); |
4064 | SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp, |
4065 | getF32Constant(DAG, 0x3f317218)); |
4066 | |
4067 | // Get the significand and build it into a floating-point number with |
4068 | // exponent of 1. |
4069 | SDValue X = GetSignificand(DAG, Op1, dl); |
4070 | |
4071 | SDValue LogOfMantissa; |
4072 | if (LimitFloatPrecision <= 6) { |
4073 | // For floating-point precision of 6: |
4074 | // |
4075 | // LogofMantissa = |
4076 | // -1.1609546f + |
4077 | // (1.4034025f - 0.23903021f * x) * x; |
4078 | // |
4079 | // error 0.0034276066, which is better than 8 bits |
4080 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4081 | getF32Constant(DAG, 0xbe74c456)); |
4082 | SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, |
4083 | getF32Constant(DAG, 0x3fb3a2b1)); |
4084 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); |
4085 | LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, |
4086 | getF32Constant(DAG, 0x3f949a29)); |
4087 | } else if (LimitFloatPrecision <= 12) { |
4088 | // For floating-point precision of 12: |
4089 | // |
4090 | // LogOfMantissa = |
4091 | // -1.7417939f + |
4092 | // (2.8212026f + |
4093 | // (-1.4699568f + |
4094 | // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x; |
4095 | // |
4096 | // error 0.000061011436, which is 14 bits |
4097 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4098 | getF32Constant(DAG, 0xbd67b6d6)); |
4099 | SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, |
4100 | getF32Constant(DAG, 0x3ee4f4b8)); |
4101 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); |
4102 | SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, |
4103 | getF32Constant(DAG, 0x3fbc278b)); |
4104 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); |
4105 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, |
4106 | getF32Constant(DAG, 0x40348e95)); |
4107 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); |
4108 | LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, |
4109 | getF32Constant(DAG, 0x3fdef31a)); |
4110 | } else { // LimitFloatPrecision <= 18 |
4111 | // For floating-point precision of 18: |
4112 | // |
4113 | // LogOfMantissa = |
4114 | // -2.1072184f + |
4115 | // (4.2372794f + |
4116 | // (-3.7029485f + |
4117 | // (2.2781945f + |
4118 | // (-0.87823314f + |
4119 | // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x; |
4120 | // |
4121 | // error 0.0000023660568, which is better than 18 bits |
4122 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4123 | getF32Constant(DAG, 0xbc91e5ac)); |
4124 | SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, |
4125 | getF32Constant(DAG, 0x3e4350aa)); |
4126 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); |
4127 | SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, |
4128 | getF32Constant(DAG, 0x3f60d3e3)); |
4129 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); |
4130 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, |
4131 | getF32Constant(DAG, 0x4011cdf0)); |
4132 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); |
4133 | SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, |
4134 | getF32Constant(DAG, 0x406cfd1c)); |
4135 | SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); |
4136 | SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, |
4137 | getF32Constant(DAG, 0x408797cb)); |
4138 | SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); |
4139 | LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10, |
4140 | getF32Constant(DAG, 0x4006dcab)); |
4141 | } |
4142 | |
4143 | return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, LogOfMantissa); |
4144 | } |
4145 | |
4146 | // No special expansion. |
4147 | return DAG.getNode(ISD::FLOG, dl, Op.getValueType(), Op); |
4148 | } |
4149 | |
4150 | /// expandLog2 - Lower a log2 intrinsic. Handles the special sequences for |
4151 | /// limited-precision mode. |
4152 | static SDValue expandLog2(SDLoc dl, SDValue Op, SelectionDAG &DAG, |
4153 | const TargetLowering &TLI) { |
4154 | if (Op.getValueType() == MVT::f32 && |
4155 | LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { |
4156 | SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op); |
4157 | |
4158 | // Get the exponent. |
4159 | SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl); |
4160 | |
4161 | // Get the significand and build it into a floating-point number with |
4162 | // exponent of 1. |
4163 | SDValue X = GetSignificand(DAG, Op1, dl); |
4164 | |
4165 | // Different possible minimax approximations of significand in |
4166 | // floating-point for various degrees of accuracy over [1,2]. |
4167 | SDValue Log2ofMantissa; |
4168 | if (LimitFloatPrecision <= 6) { |
4169 | // For floating-point precision of 6: |
4170 | // |
4171 | // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x; |
4172 | // |
4173 | // error 0.0049451742, which is more than 7 bits |
4174 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4175 | getF32Constant(DAG, 0xbeb08fe0)); |
4176 | SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, |
4177 | getF32Constant(DAG, 0x40019463)); |
4178 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); |
4179 | Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, |
4180 | getF32Constant(DAG, 0x3fd6633d)); |
4181 | } else if (LimitFloatPrecision <= 12) { |
4182 | // For floating-point precision of 12: |
4183 | // |
4184 | // Log2ofMantissa = |
4185 | // -2.51285454f + |
4186 | // (4.07009056f + |
4187 | // (-2.12067489f + |
4188 | // (.645142248f - 0.816157886e-1f * x) * x) * x) * x; |
4189 | // |
4190 | // error 0.0000876136000, which is better than 13 bits |
4191 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4192 | getF32Constant(DAG, 0xbda7262e)); |
4193 | SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, |
4194 | getF32Constant(DAG, 0x3f25280b)); |
4195 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); |
4196 | SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, |
4197 | getF32Constant(DAG, 0x4007b923)); |
4198 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); |
4199 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, |
4200 | getF32Constant(DAG, 0x40823e2f)); |
4201 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); |
4202 | Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, |
4203 | getF32Constant(DAG, 0x4020d29c)); |
4204 | } else { // LimitFloatPrecision <= 18 |
4205 | // For floating-point precision of 18: |
4206 | // |
4207 | // Log2ofMantissa = |
4208 | // -3.0400495f + |
4209 | // (6.1129976f + |
4210 | // (-5.3420409f + |
4211 | // (3.2865683f + |
4212 | // (-1.2669343f + |
4213 | // (0.27515199f - |
4214 | // 0.25691327e-1f * x) * x) * x) * x) * x) * x; |
4215 | // |
4216 | // error 0.0000018516, which is better than 18 bits |
4217 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4218 | getF32Constant(DAG, 0xbcd2769e)); |
4219 | SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, |
4220 | getF32Constant(DAG, 0x3e8ce0b9)); |
4221 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); |
4222 | SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, |
4223 | getF32Constant(DAG, 0x3fa22ae7)); |
4224 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); |
4225 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, |
4226 | getF32Constant(DAG, 0x40525723)); |
4227 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); |
4228 | SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, |
4229 | getF32Constant(DAG, 0x40aaf200)); |
4230 | SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); |
4231 | SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, |
4232 | getF32Constant(DAG, 0x40c39dad)); |
4233 | SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); |
4234 | Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10, |
4235 | getF32Constant(DAG, 0x4042902c)); |
4236 | } |
4237 | |
4238 | return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log2ofMantissa); |
4239 | } |
4240 | |
4241 | // No special expansion. |
4242 | return DAG.getNode(ISD::FLOG2, dl, Op.getValueType(), Op); |
4243 | } |
4244 | |
4245 | /// expandLog10 - Lower a log10 intrinsic. Handles the special sequences for |
4246 | /// limited-precision mode. |
4247 | static SDValue expandLog10(SDLoc dl, SDValue Op, SelectionDAG &DAG, |
4248 | const TargetLowering &TLI) { |
4249 | if (Op.getValueType() == MVT::f32 && |
4250 | LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { |
4251 | SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op); |
4252 | |
4253 | // Scale the exponent by log10(2) [0.30102999f]. |
4254 | SDValue Exp = GetExponent(DAG, Op1, TLI, dl); |
4255 | SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp, |
4256 | getF32Constant(DAG, 0x3e9a209a)); |
4257 | |
4258 | // Get the significand and build it into a floating-point number with |
4259 | // exponent of 1. |
4260 | SDValue X = GetSignificand(DAG, Op1, dl); |
4261 | |
4262 | SDValue Log10ofMantissa; |
4263 | if (LimitFloatPrecision <= 6) { |
4264 | // For floating-point precision of 6: |
4265 | // |
4266 | // Log10ofMantissa = |
4267 | // -0.50419619f + |
4268 | // (0.60948995f - 0.10380950f * x) * x; |
4269 | // |
4270 | // error 0.0014886165, which is 6 bits |
4271 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4272 | getF32Constant(DAG, 0xbdd49a13)); |
4273 | SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, |
4274 | getF32Constant(DAG, 0x3f1c0789)); |
4275 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); |
4276 | Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, |
4277 | getF32Constant(DAG, 0x3f011300)); |
4278 | } else if (LimitFloatPrecision <= 12) { |
4279 | // For floating-point precision of 12: |
4280 | // |
4281 | // Log10ofMantissa = |
4282 | // -0.64831180f + |
4283 | // (0.91751397f + |
4284 | // (-0.31664806f + 0.47637168e-1f * x) * x) * x; |
4285 | // |
4286 | // error 0.00019228036, which is better than 12 bits |
4287 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4288 | getF32Constant(DAG, 0x3d431f31)); |
4289 | SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, |
4290 | getF32Constant(DAG, 0x3ea21fb2)); |
4291 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); |
4292 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, |
4293 | getF32Constant(DAG, 0x3f6ae232)); |
4294 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); |
4295 | Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4, |
4296 | getF32Constant(DAG, 0x3f25f7c3)); |
4297 | } else { // LimitFloatPrecision <= 18 |
4298 | // For floating-point precision of 18: |
4299 | // |
4300 | // Log10ofMantissa = |
4301 | // -0.84299375f + |
4302 | // (1.5327582f + |
4303 | // (-1.0688956f + |
4304 | // (0.49102474f + |
4305 | // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x; |
4306 | // |
4307 | // error 0.0000037995730, which is better than 18 bits |
4308 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4309 | getF32Constant(DAG, 0x3c5d51ce)); |
4310 | SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, |
4311 | getF32Constant(DAG, 0x3e00685a)); |
4312 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); |
4313 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, |
4314 | getF32Constant(DAG, 0x3efb6798)); |
4315 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); |
4316 | SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4, |
4317 | getF32Constant(DAG, 0x3f88d192)); |
4318 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); |
4319 | SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, |
4320 | getF32Constant(DAG, 0x3fc4316c)); |
4321 | SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); |
4322 | Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8, |
4323 | getF32Constant(DAG, 0x3f57ce70)); |
4324 | } |
4325 | |
4326 | return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log10ofMantissa); |
4327 | } |
4328 | |
4329 | // No special expansion. |
4330 | return DAG.getNode(ISD::FLOG10, dl, Op.getValueType(), Op); |
4331 | } |
4332 | |
4333 | /// expandExp2 - Lower an exp2 intrinsic. Handles the special sequences for |
4334 | /// limited-precision mode. |
4335 | static SDValue expandExp2(SDLoc dl, SDValue Op, SelectionDAG &DAG, |
4336 | const TargetLowering &TLI) { |
4337 | if (Op.getValueType() == MVT::f32 && |
4338 | LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { |
4339 | SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op); |
4340 | |
4341 | // FractionalPartOfX = x - (float)IntegerPartOfX; |
4342 | SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX); |
4343 | SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1); |
4344 | |
4345 | // IntegerPartOfX <<= 23; |
4346 | IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX, |
4347 | DAG.getConstant(23, TLI.getPointerTy())); |
4348 | |
4349 | SDValue TwoToFractionalPartOfX; |
4350 | if (LimitFloatPrecision <= 6) { |
4351 | // For floating-point precision of 6: |
4352 | // |
4353 | // TwoToFractionalPartOfX = |
4354 | // 0.997535578f + |
4355 | // (0.735607626f + 0.252464424f * x) * x; |
4356 | // |
4357 | // error 0.0144103317, which is 6 bits |
4358 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4359 | getF32Constant(DAG, 0x3e814304)); |
4360 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, |
4361 | getF32Constant(DAG, 0x3f3c50c8)); |
4362 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); |
4363 | TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, |
4364 | getF32Constant(DAG, 0x3f7f5e7e)); |
4365 | } else if (LimitFloatPrecision <= 12) { |
4366 | // For floating-point precision of 12: |
4367 | // |
4368 | // TwoToFractionalPartOfX = |
4369 | // 0.999892986f + |
4370 | // (0.696457318f + |
4371 | // (0.224338339f + 0.792043434e-1f * x) * x) * x; |
4372 | // |
4373 | // error 0.000107046256, which is 13 to 14 bits |
4374 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4375 | getF32Constant(DAG, 0x3da235e3)); |
4376 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, |
4377 | getF32Constant(DAG, 0x3e65b8f3)); |
4378 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); |
4379 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, |
4380 | getF32Constant(DAG, 0x3f324b07)); |
4381 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); |
4382 | TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, |
4383 | getF32Constant(DAG, 0x3f7ff8fd)); |
4384 | } else { // LimitFloatPrecision <= 18 |
4385 | // For floating-point precision of 18: |
4386 | // |
4387 | // TwoToFractionalPartOfX = |
4388 | // 0.999999982f + |
4389 | // (0.693148872f + |
4390 | // (0.240227044f + |
4391 | // (0.554906021e-1f + |
4392 | // (0.961591928e-2f + |
4393 | // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; |
4394 | // error 2.47208000*10^(-7), which is better than 18 bits |
4395 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4396 | getF32Constant(DAG, 0x3924b03e)); |
4397 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, |
4398 | getF32Constant(DAG, 0x3ab24b87)); |
4399 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); |
4400 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, |
4401 | getF32Constant(DAG, 0x3c1d8c17)); |
4402 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); |
4403 | SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, |
4404 | getF32Constant(DAG, 0x3d634a1d)); |
4405 | SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); |
4406 | SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, |
4407 | getF32Constant(DAG, 0x3e75fe14)); |
4408 | SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); |
4409 | SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10, |
4410 | getF32Constant(DAG, 0x3f317234)); |
4411 | SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X); |
4412 | TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12, |
4413 | getF32Constant(DAG, 0x3f800000)); |
4414 | } |
4415 | |
4416 | // Add the exponent into the result in integer domain. |
4417 | SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, |
4418 | TwoToFractionalPartOfX); |
4419 | return DAG.getNode(ISD::BITCAST, dl, MVT::f32, |
4420 | DAG.getNode(ISD::ADD, dl, MVT::i32, |
4421 | t13, IntegerPartOfX)); |
4422 | } |
4423 | |
4424 | // No special expansion. |
4425 | return DAG.getNode(ISD::FEXP2, dl, Op.getValueType(), Op); |
4426 | } |
4427 | |
4428 | /// visitPow - Lower a pow intrinsic. Handles the special sequences for |
4429 | /// limited-precision mode with x == 10.0f. |
4430 | static SDValue expandPow(SDLoc dl, SDValue LHS, SDValue RHS, |
4431 | SelectionDAG &DAG, const TargetLowering &TLI) { |
4432 | bool IsExp10 = false; |
4433 | if (LHS.getValueType() == MVT::f32 && RHS.getValueType() == MVT::f32 && |
4434 | LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { |
4435 | if (ConstantFPSDNode *LHSC = dyn_cast<ConstantFPSDNode>(LHS)) { |
4436 | APFloat Ten(10.0f); |
4437 | IsExp10 = LHSC->isExactlyValue(Ten); |
4438 | } |
4439 | } |
4440 | |
4441 | if (IsExp10) { |
4442 | // Put the exponent in the right bit position for later addition to the |
4443 | // final result: |
4444 | // |
4445 | // #define LOG2OF10 3.3219281f |
4446 | // IntegerPartOfX = (int32_t)(x * LOG2OF10); |
4447 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, RHS, |
4448 | getF32Constant(DAG, 0x40549a78)); |
4449 | SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0); |
4450 | |
4451 | // FractionalPartOfX = x - (float)IntegerPartOfX; |
4452 | SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX); |
4453 | SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1); |
4454 | |
4455 | // IntegerPartOfX <<= 23; |
4456 | IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX, |
4457 | DAG.getConstant(23, TLI.getPointerTy())); |
4458 | |
4459 | SDValue TwoToFractionalPartOfX; |
4460 | if (LimitFloatPrecision <= 6) { |
4461 | // For floating-point precision of 6: |
4462 | // |
4463 | // twoToFractionalPartOfX = |
4464 | // 0.997535578f + |
4465 | // (0.735607626f + 0.252464424f * x) * x; |
4466 | // |
4467 | // error 0.0144103317, which is 6 bits |
4468 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4469 | getF32Constant(DAG, 0x3e814304)); |
4470 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, |
4471 | getF32Constant(DAG, 0x3f3c50c8)); |
4472 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); |
4473 | TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, |
4474 | getF32Constant(DAG, 0x3f7f5e7e)); |
4475 | } else if (LimitFloatPrecision <= 12) { |
4476 | // For floating-point precision of 12: |
4477 | // |
4478 | // TwoToFractionalPartOfX = |
4479 | // 0.999892986f + |
4480 | // (0.696457318f + |
4481 | // (0.224338339f + 0.792043434e-1f * x) * x) * x; |
4482 | // |
4483 | // error 0.000107046256, which is 13 to 14 bits |
4484 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4485 | getF32Constant(DAG, 0x3da235e3)); |
4486 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, |
4487 | getF32Constant(DAG, 0x3e65b8f3)); |
4488 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); |
4489 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, |
4490 | getF32Constant(DAG, 0x3f324b07)); |
4491 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); |
4492 | TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, |
4493 | getF32Constant(DAG, 0x3f7ff8fd)); |
4494 | } else { // LimitFloatPrecision <= 18 |
4495 | // For floating-point precision of 18: |
4496 | // |
4497 | // TwoToFractionalPartOfX = |
4498 | // 0.999999982f + |
4499 | // (0.693148872f + |
4500 | // (0.240227044f + |
4501 | // (0.554906021e-1f + |
4502 | // (0.961591928e-2f + |
4503 | // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; |
4504 | // error 2.47208000*10^(-7), which is better than 18 bits |
4505 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, |
4506 | getF32Constant(DAG, 0x3924b03e)); |
4507 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, |
4508 | getF32Constant(DAG, 0x3ab24b87)); |
4509 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); |
4510 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, |
4511 | getF32Constant(DAG, 0x3c1d8c17)); |
4512 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); |
4513 | SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, |
4514 | getF32Constant(DAG, 0x3d634a1d)); |
4515 | SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); |
4516 | SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, |
4517 | getF32Constant(DAG, 0x3e75fe14)); |
4518 | SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); |
4519 | SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10, |
4520 | getF32Constant(DAG, 0x3f317234)); |
4521 | SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X); |
4522 | TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12, |
4523 | getF32Constant(DAG, 0x3f800000)); |
4524 | } |
4525 | |
4526 | SDValue t13 = DAG.getNode(ISD::BITCAST, dl,MVT::i32,TwoToFractionalPartOfX); |
4527 | return DAG.getNode(ISD::BITCAST, dl, MVT::f32, |
4528 | DAG.getNode(ISD::ADD, dl, MVT::i32, |
4529 | t13, IntegerPartOfX)); |
4530 | } |
4531 | |
4532 | // No special expansion. |
4533 | return DAG.getNode(ISD::FPOW, dl, LHS.getValueType(), LHS, RHS); |
4534 | } |
4535 | |
4536 | |
4537 | /// ExpandPowI - Expand a llvm.powi intrinsic. |
4538 | static SDValue ExpandPowI(SDLoc DL, SDValue LHS, SDValue RHS, |
4539 | SelectionDAG &DAG) { |
4540 | // If RHS is a constant, we can expand this out to a multiplication tree, |
4541 | // otherwise we end up lowering to a call to __powidf2 (for example). When |
4542 | // optimizing for size, we only want to do this if the expansion would produce |
4543 | // a small number of multiplies, otherwise we do the full expansion. |
4544 | if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) { |
4545 | // Get the exponent as a positive value. |
4546 | unsigned Val = RHSC->getSExtValue(); |
4547 | if ((int)Val < 0) Val = -Val; |
4548 | |
4549 | // powi(x, 0) -> 1.0 |
4550 | if (Val == 0) |
4551 | return DAG.getConstantFP(1.0, LHS.getValueType()); |
4552 | |
4553 | const Function *F = DAG.getMachineFunction().getFunction(); |
4554 | if (!F->getAttributes().hasAttribute(AttributeSet::FunctionIndex, |
4555 | Attribute::OptimizeForSize) || |
4556 | // If optimizing for size, don't insert too many multiplies. This |
4557 | // inserts up to 5 multiplies. |
4558 | CountPopulation_32(Val)+Log2_32(Val) < 7) { |
4559 | // We use the simple binary decomposition method to generate the multiply |
4560 | // sequence. There are more optimal ways to do this (for example, |
4561 | // powi(x,15) generates one more multiply than it should), but this has |
4562 | // the benefit of being both really simple and much better than a libcall. |
4563 | SDValue Res; // Logically starts equal to 1.0 |
4564 | SDValue CurSquare = LHS; |
4565 | while (Val) { |
4566 | if (Val & 1) { |
4567 | if (Res.getNode()) |
4568 | Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare); |
4569 | else |
4570 | Res = CurSquare; // 1.0*CurSquare. |
4571 | } |
4572 | |
4573 | CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(), |
4574 | CurSquare, CurSquare); |
4575 | Val >>= 1; |
4576 | } |
4577 | |
4578 | // If the original was negative, invert the result, producing 1/(x*x*x). |
4579 | if (RHSC->getSExtValue() < 0) |
4580 | Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(), |
4581 | DAG.getConstantFP(1.0, LHS.getValueType()), Res); |
4582 | return Res; |
4583 | } |
4584 | } |
4585 | |
4586 | // Otherwise, expand to a libcall. |
4587 | return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS); |
4588 | } |
4589 | |
4590 | // getTruncatedArgReg - Find underlying register used for an truncated |
4591 | // argument. |
4592 | static unsigned getTruncatedArgReg(const SDValue &N) { |
4593 | if (N.getOpcode() != ISD::TRUNCATE) |
4594 | return 0; |
4595 | |
4596 | const SDValue &Ext = N.getOperand(0); |
4597 | if (Ext.getOpcode() == ISD::AssertZext || |
4598 | Ext.getOpcode() == ISD::AssertSext) { |
4599 | const SDValue &CFR = Ext.getOperand(0); |
4600 | if (CFR.getOpcode() == ISD::CopyFromReg) |
4601 | return cast<RegisterSDNode>(CFR.getOperand(1))->getReg(); |
4602 | if (CFR.getOpcode() == ISD::TRUNCATE) |
4603 | return getTruncatedArgReg(CFR); |
4604 | } |
4605 | return 0; |
4606 | } |
4607 | |
4608 | /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function |
4609 | /// argument, create the corresponding DBG_VALUE machine instruction for it now. |
4610 | /// At the end of instruction selection, they will be inserted to the entry BB. |
4611 | bool SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, |
4612 | MDNode *Variable, |
4613 | MDNode *Expr, int64_t Offset, |
4614 | bool IsIndirect, |
4615 | const SDValue &N) { |
4616 | const Argument *Arg = dyn_cast<Argument>(V); |
4617 | if (!Arg) |
4618 | return false; |
4619 | |
4620 | MachineFunction &MF = DAG.getMachineFunction(); |
4621 | const TargetInstrInfo *TII = DAG.getSubtarget().getInstrInfo(); |
4622 | |
4623 | // Ignore inlined function arguments here. |
4624 | DIVariable DV(Variable); |
4625 | if (DV.isInlinedFnArgument(MF.getFunction())) |
4626 | return false; |
4627 | |
4628 | Optional<MachineOperand> Op; |
4629 | // Some arguments' frame index is recorded during argument lowering. |
4630 | if (int FI = FuncInfo.getArgumentFrameIndex(Arg)) |
4631 | Op = MachineOperand::CreateFI(FI); |
4632 | |
4633 | if (!Op && N.getNode()) { |
4634 | unsigned Reg; |
4635 | if (N.getOpcode() == ISD::CopyFromReg) |
4636 | Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg(); |
4637 | else |
4638 | Reg = getTruncatedArgReg(N); |
4639 | if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) { |
4640 | MachineRegisterInfo &RegInfo = MF.getRegInfo(); |
4641 | unsigned PR = RegInfo.getLiveInPhysReg(Reg); |
4642 | if (PR) |
4643 | Reg = PR; |
4644 | } |
4645 | if (Reg) |
4646 | Op = MachineOperand::CreateReg(Reg, false); |
4647 | } |
4648 | |
4649 | if (!Op) { |
4650 | // Check if ValueMap has reg number. |
4651 | DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V); |
4652 | if (VMI != FuncInfo.ValueMap.end()) |
4653 | Op = MachineOperand::CreateReg(VMI->second, false); |
4654 | } |
4655 | |
4656 | if (!Op && N.getNode()) |
4657 | // Check if frame index is available. |
4658 | if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode())) |
4659 | if (FrameIndexSDNode *FINode = |
4660 | dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode())) |
4661 | Op = MachineOperand::CreateFI(FINode->getIndex()); |
4662 | |
4663 | if (!Op) |
4664 | return false; |
4665 | |
4666 | if (Op->isReg()) |
4667 | FuncInfo.ArgDbgValues.push_back( |
4668 | BuildMI(MF, getCurDebugLoc(), TII->get(TargetOpcode::DBG_VALUE), |
4669 | IsIndirect, Op->getReg(), Offset, Variable, Expr)); |
4670 | else |
4671 | FuncInfo.ArgDbgValues.push_back( |
4672 | BuildMI(MF, getCurDebugLoc(), TII->get(TargetOpcode::DBG_VALUE)) |
4673 | .addOperand(*Op) |
4674 | .addImm(Offset) |
4675 | .addMetadata(Variable) |
4676 | .addMetadata(Expr)); |
4677 | |
4678 | return true; |
4679 | } |
4680 | |
4681 | // VisualStudio defines setjmp as _setjmp |
4682 | #if defined(_MSC_VER) && defined(setjmp) && \ |
4683 | !defined(setjmp_undefined_for_msvc) |
4684 | # pragma push_macro("setjmp") |
4685 | # undef setjmp |
4686 | # define setjmp_undefined_for_msvc |
4687 | #endif |
4688 | |
4689 | /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If |
4690 | /// we want to emit this as a call to a named external function, return the name |
4691 | /// otherwise lower it and return null. |
4692 | const char * |
4693 | SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) { |
4694 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
4695 | SDLoc sdl = getCurSDLoc(); |
4696 | DebugLoc dl = getCurDebugLoc(); |
4697 | SDValue Res; |
4698 | |
4699 | switch (Intrinsic) { |
4700 | default: |
4701 | // By default, turn this into a target intrinsic node. |
4702 | visitTargetIntrinsic(I, Intrinsic); |
4703 | return nullptr; |
4704 | case Intrinsic::vastart: visitVAStart(I); return nullptr; |
4705 | case Intrinsic::vaend: visitVAEnd(I); return nullptr; |
4706 | case Intrinsic::vacopy: visitVACopy(I); return nullptr; |
4707 | case Intrinsic::returnaddress: |
4708 | setValue(&I, DAG.getNode(ISD::RETURNADDR, sdl, TLI.getPointerTy(), |
4709 | getValue(I.getArgOperand(0)))); |
4710 | return nullptr; |
4711 | case Intrinsic::frameaddress: |
4712 | setValue(&I, DAG.getNode(ISD::FRAMEADDR, sdl, TLI.getPointerTy(), |
4713 | getValue(I.getArgOperand(0)))); |
4714 | return nullptr; |
4715 | case Intrinsic::read_register: { |
4716 | Value *Reg = I.getArgOperand(0); |
4717 | SDValue RegName = |
4718 | DAG.getMDNode(cast<MDNode>(cast<MetadataAsValue>(Reg)->getMetadata())); |
4719 | EVT VT = TLI.getValueType(I.getType()); |
4720 | setValue(&I, DAG.getNode(ISD::READ_REGISTER, sdl, VT, RegName)); |
4721 | return nullptr; |
4722 | } |
4723 | case Intrinsic::write_register: { |
4724 | Value *Reg = I.getArgOperand(0); |
4725 | Value *RegValue = I.getArgOperand(1); |
4726 | SDValue Chain = getValue(RegValue).getOperand(0); |
4727 | SDValue RegName = |
4728 | DAG.getMDNode(cast<MDNode>(cast<MetadataAsValue>(Reg)->getMetadata())); |
4729 | DAG.setRoot(DAG.getNode(ISD::WRITE_REGISTER, sdl, MVT::Other, Chain, |
4730 | RegName, getValue(RegValue))); |
4731 | return nullptr; |
4732 | } |
4733 | case Intrinsic::setjmp: |
4734 | return &"_setjmp"[!TLI.usesUnderscoreSetJmp()]; |
4735 | case Intrinsic::longjmp: |
4736 | return &"_longjmp"[!TLI.usesUnderscoreLongJmp()]; |
4737 | case Intrinsic::memcpy: { |
4738 | // Assert for address < 256 since we support only user defined address |
4739 | // spaces. |
4740 | assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4744, __PRETTY_FUNCTION__)) |
4741 | < 256 &&((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4744, __PRETTY_FUNCTION__)) |
4742 | cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4744, __PRETTY_FUNCTION__)) |
4743 | < 256 &&((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4744, __PRETTY_FUNCTION__)) |
4744 | "Unknown address space")((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4744, __PRETTY_FUNCTION__)); |
4745 | SDValue Op1 = getValue(I.getArgOperand(0)); |
4746 | SDValue Op2 = getValue(I.getArgOperand(1)); |
4747 | SDValue Op3 = getValue(I.getArgOperand(2)); |
4748 | unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue(); |
4749 | if (!Align) |
4750 | Align = 1; // @llvm.memcpy defines 0 and 1 to both mean no alignment. |
4751 | bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue(); |
4752 | DAG.setRoot(DAG.getMemcpy(getRoot(), sdl, Op1, Op2, Op3, Align, isVol, false, |
4753 | MachinePointerInfo(I.getArgOperand(0)), |
4754 | MachinePointerInfo(I.getArgOperand(1)))); |
4755 | return nullptr; |
4756 | } |
4757 | case Intrinsic::memset: { |
4758 | // Assert for address < 256 since we support only user defined address |
4759 | // spaces. |
4760 | assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && "Unknown address space" ) ? static_cast<void> (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4762, __PRETTY_FUNCTION__)) |
4761 | < 256 &&((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && "Unknown address space" ) ? static_cast<void> (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4762, __PRETTY_FUNCTION__)) |
4762 | "Unknown address space")((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && "Unknown address space" ) ? static_cast<void> (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4762, __PRETTY_FUNCTION__)); |
4763 | SDValue Op1 = getValue(I.getArgOperand(0)); |
4764 | SDValue Op2 = getValue(I.getArgOperand(1)); |
4765 | SDValue Op3 = getValue(I.getArgOperand(2)); |
4766 | unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue(); |
4767 | if (!Align) |
4768 | Align = 1; // @llvm.memset defines 0 and 1 to both mean no alignment. |
4769 | bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue(); |
4770 | DAG.setRoot(DAG.getMemset(getRoot(), sdl, Op1, Op2, Op3, Align, isVol, |
4771 | MachinePointerInfo(I.getArgOperand(0)))); |
4772 | return nullptr; |
4773 | } |
4774 | case Intrinsic::memmove: { |
4775 | // Assert for address < 256 since we support only user defined address |
4776 | // spaces. |
4777 | assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4781, __PRETTY_FUNCTION__)) |
4778 | < 256 &&((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4781, __PRETTY_FUNCTION__)) |
4779 | cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4781, __PRETTY_FUNCTION__)) |
4780 | < 256 &&((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4781, __PRETTY_FUNCTION__)) |
4781 | "Unknown address space")((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4781, __PRETTY_FUNCTION__)); |
4782 | SDValue Op1 = getValue(I.getArgOperand(0)); |
4783 | SDValue Op2 = getValue(I.getArgOperand(1)); |
4784 | SDValue Op3 = getValue(I.getArgOperand(2)); |
4785 | unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue(); |
4786 | if (!Align) |
4787 | Align = 1; // @llvm.memmove defines 0 and 1 to both mean no alignment. |
4788 | bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue(); |
4789 | DAG.setRoot(DAG.getMemmove(getRoot(), sdl, Op1, Op2, Op3, Align, isVol, |
4790 | MachinePointerInfo(I.getArgOperand(0)), |
4791 | MachinePointerInfo(I.getArgOperand(1)))); |
4792 | return nullptr; |
4793 | } |
4794 | case Intrinsic::dbg_declare: { |
4795 | const DbgDeclareInst &DI = cast<DbgDeclareInst>(I); |
4796 | MDNode *Variable = DI.getVariable(); |
4797 | MDNode *Expression = DI.getExpression(); |
4798 | const Value *Address = DI.getAddress(); |
4799 | DIVariable DIVar(Variable); |
4800 | assert((!DIVar || DIVar.isVariable()) &&(((!DIVar || DIVar.isVariable()) && "Variable in DbgDeclareInst should be either null or a DIVariable." ) ? static_cast<void> (0) : __assert_fail ("(!DIVar || DIVar.isVariable()) && \"Variable in DbgDeclareInst should be either null or a DIVariable.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4801, __PRETTY_FUNCTION__)) |
4801 | "Variable in DbgDeclareInst should be either null or a DIVariable.")(((!DIVar || DIVar.isVariable()) && "Variable in DbgDeclareInst should be either null or a DIVariable." ) ? static_cast<void> (0) : __assert_fail ("(!DIVar || DIVar.isVariable()) && \"Variable in DbgDeclareInst should be either null or a DIVariable.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4801, __PRETTY_FUNCTION__)); |
4802 | if (!Address || !DIVar) { |
4803 | DEBUG(dbgs() << "Dropping debug info for " << DI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Dropping debug info for " << DI << "\n"; } } while (0); |
4804 | return nullptr; |
4805 | } |
4806 | |
4807 | // Check if address has undef value. |
4808 | if (isa<UndefValue>(Address) || |
4809 | (Address->use_empty() && !isa<Argument>(Address))) { |
4810 | DEBUG(dbgs() << "Dropping debug info for " << DI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Dropping debug info for " << DI << "\n"; } } while (0); |
4811 | return nullptr; |
4812 | } |
4813 | |
4814 | SDValue &N = NodeMap[Address]; |
4815 | if (!N.getNode() && isa<Argument>(Address)) |
4816 | // Check unused arguments map. |
4817 | N = UnusedArgNodeMap[Address]; |
4818 | SDDbgValue *SDV; |
4819 | if (N.getNode()) { |
4820 | if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address)) |
4821 | Address = BCI->getOperand(0); |
4822 | // Parameters are handled specially. |
4823 | bool isParameter = |
4824 | (DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable || |
4825 | isa<Argument>(Address)); |
4826 | |
4827 | const AllocaInst *AI = dyn_cast<AllocaInst>(Address); |
4828 | |
4829 | if (isParameter && !AI) { |
4830 | FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode()); |
4831 | if (FINode) |
4832 | // Byval parameter. We have a frame index at this point. |
4833 | SDV = DAG.getFrameIndexDbgValue( |
4834 | Variable, Expression, FINode->getIndex(), 0, dl, SDNodeOrder); |
4835 | else { |
4836 | // Address is an argument, so try to emit its dbg value using |
4837 | // virtual register info from the FuncInfo.ValueMap. |
4838 | EmitFuncArgumentDbgValue(Address, Variable, Expression, 0, false, N); |
4839 | return nullptr; |
4840 | } |
4841 | } else if (AI) |
4842 | SDV = DAG.getDbgValue(Variable, Expression, N.getNode(), N.getResNo(), |
4843 | true, 0, dl, SDNodeOrder); |
4844 | else { |
4845 | // Can't do anything with other non-AI cases yet. |
4846 | DEBUG(dbgs() << "Dropping debug info for " << DI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Dropping debug info for " << DI << "\n"; } } while (0); |
4847 | DEBUG(dbgs() << "non-AllocaInst issue for Address: \n\t")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "non-AllocaInst issue for Address: \n\t" ; } } while (0); |
4848 | DEBUG(Address->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { Address->dump(); } } while (0); |
4849 | return nullptr; |
4850 | } |
4851 | DAG.AddDbgValue(SDV, N.getNode(), isParameter); |
4852 | } else { |
4853 | // If Address is an argument then try to emit its dbg value using |
4854 | // virtual register info from the FuncInfo.ValueMap. |
4855 | if (!EmitFuncArgumentDbgValue(Address, Variable, Expression, 0, false, |
4856 | N)) { |
4857 | // If variable is pinned by a alloca in dominating bb then |
4858 | // use StaticAllocaMap. |
4859 | if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) { |
4860 | if (AI->getParent() != DI.getParent()) { |
4861 | DenseMap<const AllocaInst*, int>::iterator SI = |
4862 | FuncInfo.StaticAllocaMap.find(AI); |
4863 | if (SI != FuncInfo.StaticAllocaMap.end()) { |
4864 | SDV = DAG.getFrameIndexDbgValue(Variable, Expression, SI->second, |
4865 | 0, dl, SDNodeOrder); |
4866 | DAG.AddDbgValue(SDV, nullptr, false); |
4867 | return nullptr; |
4868 | } |
4869 | } |
4870 | } |
4871 | DEBUG(dbgs() << "Dropping debug info for " << DI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Dropping debug info for " << DI << "\n"; } } while (0); |
4872 | } |
4873 | } |
4874 | return nullptr; |
4875 | } |
4876 | case Intrinsic::dbg_value: { |
4877 | const DbgValueInst &DI = cast<DbgValueInst>(I); |
4878 | DIVariable DIVar(DI.getVariable()); |
4879 | assert((!DIVar || DIVar.isVariable()) &&(((!DIVar || DIVar.isVariable()) && "Variable in DbgValueInst should be either null or a DIVariable." ) ? static_cast<void> (0) : __assert_fail ("(!DIVar || DIVar.isVariable()) && \"Variable in DbgValueInst should be either null or a DIVariable.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4880, __PRETTY_FUNCTION__)) |
4880 | "Variable in DbgValueInst should be either null or a DIVariable.")(((!DIVar || DIVar.isVariable()) && "Variable in DbgValueInst should be either null or a DIVariable." ) ? static_cast<void> (0) : __assert_fail ("(!DIVar || DIVar.isVariable()) && \"Variable in DbgValueInst should be either null or a DIVariable.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4880, __PRETTY_FUNCTION__)); |
4881 | if (!DIVar) |
4882 | return nullptr; |
4883 | |
4884 | MDNode *Variable = DI.getVariable(); |
4885 | MDNode *Expression = DI.getExpression(); |
4886 | uint64_t Offset = DI.getOffset(); |
4887 | const Value *V = DI.getValue(); |
4888 | if (!V) |
4889 | return nullptr; |
4890 | |
4891 | SDDbgValue *SDV; |
4892 | if (isa<ConstantInt>(V) || isa<ConstantFP>(V) || isa<UndefValue>(V)) { |
4893 | SDV = DAG.getConstantDbgValue(Variable, Expression, V, Offset, dl, |
4894 | SDNodeOrder); |
4895 | DAG.AddDbgValue(SDV, nullptr, false); |
4896 | } else { |
4897 | // Do not use getValue() in here; we don't want to generate code at |
4898 | // this point if it hasn't been done yet. |
4899 | SDValue N = NodeMap[V]; |
4900 | if (!N.getNode() && isa<Argument>(V)) |
4901 | // Check unused arguments map. |
4902 | N = UnusedArgNodeMap[V]; |
4903 | if (N.getNode()) { |
4904 | // A dbg.value for an alloca is always indirect. |
4905 | bool IsIndirect = isa<AllocaInst>(V) || Offset != 0; |
4906 | if (!EmitFuncArgumentDbgValue(V, Variable, Expression, Offset, |
4907 | IsIndirect, N)) { |
4908 | SDV = DAG.getDbgValue(Variable, Expression, N.getNode(), N.getResNo(), |
4909 | IsIndirect, Offset, dl, SDNodeOrder); |
4910 | DAG.AddDbgValue(SDV, N.getNode(), false); |
4911 | } |
4912 | } else if (!V->use_empty() ) { |
4913 | // Do not call getValue(V) yet, as we don't want to generate code. |
4914 | // Remember it for later. |
4915 | DanglingDebugInfo DDI(&DI, dl, SDNodeOrder); |
4916 | DanglingDebugInfoMap[V] = DDI; |
4917 | } else { |
4918 | // We may expand this to cover more cases. One case where we have no |
4919 | // data available is an unreferenced parameter. |
4920 | DEBUG(dbgs() << "Dropping debug info for " << DI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Dropping debug info for " << DI << "\n"; } } while (0); |
4921 | } |
4922 | } |
4923 | |
4924 | // Build a debug info table entry. |
4925 | if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V)) |
4926 | V = BCI->getOperand(0); |
4927 | const AllocaInst *AI = dyn_cast<AllocaInst>(V); |
4928 | // Don't handle byval struct arguments or VLAs, for example. |
4929 | if (!AI) { |
4930 | DEBUG(dbgs() << "Dropping debug location info for:\n " << DI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Dropping debug location info for:\n " << DI << "\n"; } } while (0); |
4931 | DEBUG(dbgs() << " Last seen at:\n " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << " Last seen at:\n " << * V << "\n"; } } while (0); |
4932 | return nullptr; |
4933 | } |
4934 | DenseMap<const AllocaInst*, int>::iterator SI = |
4935 | FuncInfo.StaticAllocaMap.find(AI); |
4936 | if (SI == FuncInfo.StaticAllocaMap.end()) |
4937 | return nullptr; // VLAs. |
4938 | return nullptr; |
4939 | } |
4940 | |
4941 | case Intrinsic::eh_typeid_for: { |
4942 | // Find the type id for the given typeinfo. |
4943 | GlobalValue *GV = ExtractTypeInfo(I.getArgOperand(0)); |
4944 | unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV); |
4945 | Res = DAG.getConstant(TypeID, MVT::i32); |
4946 | setValue(&I, Res); |
4947 | return nullptr; |
4948 | } |
4949 | |
4950 | case Intrinsic::eh_return_i32: |
4951 | case Intrinsic::eh_return_i64: |
4952 | DAG.getMachineFunction().getMMI().setCallsEHReturn(true); |
4953 | DAG.setRoot(DAG.getNode(ISD::EH_RETURN, sdl, |
4954 | MVT::Other, |
4955 | getControlRoot(), |
4956 | getValue(I.getArgOperand(0)), |
4957 | getValue(I.getArgOperand(1)))); |
4958 | return nullptr; |
4959 | case Intrinsic::eh_unwind_init: |
4960 | DAG.getMachineFunction().getMMI().setCallsUnwindInit(true); |
4961 | return nullptr; |
4962 | case Intrinsic::eh_dwarf_cfa: { |
4963 | SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), sdl, |
4964 | TLI.getPointerTy()); |
4965 | SDValue Offset = DAG.getNode(ISD::ADD, sdl, |
4966 | CfaArg.getValueType(), |
4967 | DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, sdl, |
4968 | CfaArg.getValueType()), |
4969 | CfaArg); |
4970 | SDValue FA = DAG.getNode(ISD::FRAMEADDR, sdl, TLI.getPointerTy(), |
4971 | DAG.getConstant(0, TLI.getPointerTy())); |
4972 | setValue(&I, DAG.getNode(ISD::ADD, sdl, FA.getValueType(), |
4973 | FA, Offset)); |
4974 | return nullptr; |
4975 | } |
4976 | case Intrinsic::eh_sjlj_callsite: { |
4977 | MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); |
4978 | ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0)); |
4979 | assert(CI && "Non-constant call site value in eh.sjlj.callsite!")((CI && "Non-constant call site value in eh.sjlj.callsite!" ) ? static_cast<void> (0) : __assert_fail ("CI && \"Non-constant call site value in eh.sjlj.callsite!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4979, __PRETTY_FUNCTION__)); |
4980 | assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!")((MMI.getCurrentCallSite() == 0 && "Overlapping call sites!" ) ? static_cast<void> (0) : __assert_fail ("MMI.getCurrentCallSite() == 0 && \"Overlapping call sites!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4980, __PRETTY_FUNCTION__)); |
4981 | |
4982 | MMI.setCurrentCallSite(CI->getZExtValue()); |
4983 | return nullptr; |
4984 | } |
4985 | case Intrinsic::eh_sjlj_functioncontext: { |
4986 | // Get and store the index of the function context. |
4987 | MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); |
4988 | AllocaInst *FnCtx = |
4989 | cast<AllocaInst>(I.getArgOperand(0)->stripPointerCasts()); |
4990 | int FI = FuncInfo.StaticAllocaMap[FnCtx]; |
4991 | MFI->setFunctionContextIndex(FI); |
4992 | return nullptr; |
4993 | } |
4994 | case Intrinsic::eh_sjlj_setjmp: { |
4995 | SDValue Ops[2]; |
4996 | Ops[0] = getRoot(); |
4997 | Ops[1] = getValue(I.getArgOperand(0)); |
4998 | SDValue Op = DAG.getNode(ISD::EH_SJLJ_SETJMP, sdl, |
4999 | DAG.getVTList(MVT::i32, MVT::Other), Ops); |
5000 | setValue(&I, Op.getValue(0)); |
5001 | DAG.setRoot(Op.getValue(1)); |
5002 | return nullptr; |
5003 | } |
5004 | case Intrinsic::eh_sjlj_longjmp: { |
5005 | DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, sdl, MVT::Other, |
5006 | getRoot(), getValue(I.getArgOperand(0)))); |
5007 | return nullptr; |
5008 | } |
5009 | |
5010 | case Intrinsic::masked_load: |
5011 | visitMaskedLoad(I); |
5012 | return nullptr; |
5013 | case Intrinsic::masked_store: |
5014 | visitMaskedStore(I); |
5015 | return nullptr; |
5016 | case Intrinsic::x86_mmx_pslli_w: |
5017 | case Intrinsic::x86_mmx_pslli_d: |
5018 | case Intrinsic::x86_mmx_pslli_q: |
5019 | case Intrinsic::x86_mmx_psrli_w: |
5020 | case Intrinsic::x86_mmx_psrli_d: |
5021 | case Intrinsic::x86_mmx_psrli_q: |
5022 | case Intrinsic::x86_mmx_psrai_w: |
5023 | case Intrinsic::x86_mmx_psrai_d: { |
5024 | SDValue ShAmt = getValue(I.getArgOperand(1)); |
5025 | if (isa<ConstantSDNode>(ShAmt)) { |
5026 | visitTargetIntrinsic(I, Intrinsic); |
5027 | return nullptr; |
5028 | } |
5029 | unsigned NewIntrinsic = 0; |
5030 | EVT ShAmtVT = MVT::v2i32; |
5031 | switch (Intrinsic) { |
5032 | case Intrinsic::x86_mmx_pslli_w: |
5033 | NewIntrinsic = Intrinsic::x86_mmx_psll_w; |
5034 | break; |
5035 | case Intrinsic::x86_mmx_pslli_d: |
5036 | NewIntrinsic = Intrinsic::x86_mmx_psll_d; |
5037 | break; |
5038 | case Intrinsic::x86_mmx_pslli_q: |
5039 | NewIntrinsic = Intrinsic::x86_mmx_psll_q; |
5040 | break; |
5041 | case Intrinsic::x86_mmx_psrli_w: |
5042 | NewIntrinsic = Intrinsic::x86_mmx_psrl_w; |
5043 | break; |
5044 | case Intrinsic::x86_mmx_psrli_d: |
5045 | NewIntrinsic = Intrinsic::x86_mmx_psrl_d; |
5046 | break; |
5047 | case Intrinsic::x86_mmx_psrli_q: |
5048 | NewIntrinsic = Intrinsic::x86_mmx_psrl_q; |
5049 | break; |
5050 | case Intrinsic::x86_mmx_psrai_w: |
5051 | NewIntrinsic = Intrinsic::x86_mmx_psra_w; |
5052 | break; |
5053 | case Intrinsic::x86_mmx_psrai_d: |
5054 | NewIntrinsic = Intrinsic::x86_mmx_psra_d; |
5055 | break; |
5056 | default: llvm_unreachable("Impossible intrinsic")::llvm::llvm_unreachable_internal("Impossible intrinsic", "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5056); // Can't reach here. |
5057 | } |
5058 | |
5059 | // The vector shift intrinsics with scalars uses 32b shift amounts but |
5060 | // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits |
5061 | // to be zero. |
5062 | // We must do this early because v2i32 is not a legal type. |
5063 | SDValue ShOps[2]; |
5064 | ShOps[0] = ShAmt; |
5065 | ShOps[1] = DAG.getConstant(0, MVT::i32); |
5066 | ShAmt = DAG.getNode(ISD::BUILD_VECTOR, sdl, ShAmtVT, ShOps); |
5067 | EVT DestVT = TLI.getValueType(I.getType()); |
5068 | ShAmt = DAG.getNode(ISD::BITCAST, sdl, DestVT, ShAmt); |
5069 | Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, sdl, DestVT, |
5070 | DAG.getConstant(NewIntrinsic, MVT::i32), |
5071 | getValue(I.getArgOperand(0)), ShAmt); |
5072 | setValue(&I, Res); |
5073 | return nullptr; |
5074 | } |
5075 | case Intrinsic::x86_avx_vinsertf128_pd_256: |
5076 | case Intrinsic::x86_avx_vinsertf128_ps_256: |
5077 | case Intrinsic::x86_avx_vinsertf128_si_256: |
5078 | case Intrinsic::x86_avx2_vinserti128: { |
5079 | EVT DestVT = TLI.getValueType(I.getType()); |
5080 | EVT ElVT = TLI.getValueType(I.getArgOperand(1)->getType()); |
5081 | uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(2))->getZExtValue() & 1) * |
5082 | ElVT.getVectorNumElements(); |
5083 | Res = |
5084 | DAG.getNode(ISD::INSERT_SUBVECTOR, sdl, DestVT, |
5085 | getValue(I.getArgOperand(0)), getValue(I.getArgOperand(1)), |
5086 | DAG.getConstant(Idx, TLI.getVectorIdxTy())); |
5087 | setValue(&I, Res); |
5088 | return nullptr; |
5089 | } |
5090 | case Intrinsic::x86_avx_vextractf128_pd_256: |
5091 | case Intrinsic::x86_avx_vextractf128_ps_256: |
5092 | case Intrinsic::x86_avx_vextractf128_si_256: |
5093 | case Intrinsic::x86_avx2_vextracti128: { |
5094 | EVT DestVT = TLI.getValueType(I.getType()); |
5095 | uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(1))->getZExtValue() & 1) * |
5096 | DestVT.getVectorNumElements(); |
5097 | Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, sdl, DestVT, |
5098 | getValue(I.getArgOperand(0)), |
5099 | DAG.getConstant(Idx, TLI.getVectorIdxTy())); |
5100 | setValue(&I, Res); |
5101 | return nullptr; |
5102 | } |
5103 | case Intrinsic::convertff: |
5104 | case Intrinsic::convertfsi: |
5105 | case Intrinsic::convertfui: |
5106 | case Intrinsic::convertsif: |
5107 | case Intrinsic::convertuif: |
5108 | case Intrinsic::convertss: |
5109 | case Intrinsic::convertsu: |
5110 | case Intrinsic::convertus: |
5111 | case Intrinsic::convertuu: { |
5112 | ISD::CvtCode Code = ISD::CVT_INVALID; |
5113 | switch (Intrinsic) { |
5114 | default: llvm_unreachable("Impossible intrinsic")::llvm::llvm_unreachable_internal("Impossible intrinsic", "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5114); // Can't reach here. |
5115 | case Intrinsic::convertff: Code = ISD::CVT_FF; break; |
5116 | case Intrinsic::convertfsi: Code = ISD::CVT_FS; break; |
5117 | case Intrinsic::convertfui: Code = ISD::CVT_FU; break; |
5118 | case Intrinsic::convertsif: Code = ISD::CVT_SF; break; |
5119 | case Intrinsic::convertuif: Code = ISD::CVT_UF; break; |
5120 | case Intrinsic::convertss: Code = ISD::CVT_SS; break; |
5121 | case Intrinsic::convertsu: Code = ISD::CVT_SU; break; |
5122 | case Intrinsic::convertus: Code = ISD::CVT_US; break; |
5123 | case Intrinsic::convertuu: Code = ISD::CVT_UU; break; |
5124 | } |
5125 | EVT DestVT = TLI.getValueType(I.getType()); |
5126 | const Value *Op1 = I.getArgOperand(0); |
5127 | Res = DAG.getConvertRndSat(DestVT, sdl, getValue(Op1), |
5128 | DAG.getValueType(DestVT), |
5129 | DAG.getValueType(getValue(Op1).getValueType()), |
5130 | getValue(I.getArgOperand(1)), |
5131 | getValue(I.getArgOperand(2)), |
5132 | Code); |
5133 | setValue(&I, Res); |
5134 | return nullptr; |
5135 | } |
5136 | case Intrinsic::powi: |
5137 | setValue(&I, ExpandPowI(sdl, getValue(I.getArgOperand(0)), |
5138 | getValue(I.getArgOperand(1)), DAG)); |
5139 | return nullptr; |
5140 | case Intrinsic::log: |
5141 | setValue(&I, expandLog(sdl, getValue(I.getArgOperand(0)), DAG, TLI)); |
5142 | return nullptr; |
5143 | case Intrinsic::log2: |
5144 | setValue(&I, expandLog2(sdl, getValue(I.getArgOperand(0)), DAG, TLI)); |
5145 | return nullptr; |
5146 | case Intrinsic::log10: |
5147 | setValue(&I, expandLog10(sdl, getValue(I.getArgOperand(0)), DAG, TLI)); |
5148 | return nullptr; |
5149 | case Intrinsic::exp: |
5150 | setValue(&I, expandExp(sdl, getValue(I.getArgOperand(0)), DAG, TLI)); |
5151 | return nullptr; |
5152 | case Intrinsic::exp2: |
5153 | setValue(&I, expandExp2(sdl, getValue(I.getArgOperand(0)), DAG, TLI)); |
5154 | return nullptr; |
5155 | case Intrinsic::pow: |
5156 | setValue(&I, expandPow(sdl, getValue(I.getArgOperand(0)), |
5157 | getValue(I.getArgOperand(1)), DAG, TLI)); |
5158 | return nullptr; |
5159 | case Intrinsic::sqrt: |
5160 | case Intrinsic::fabs: |
5161 | case Intrinsic::sin: |
5162 | case Intrinsic::cos: |
5163 | case Intrinsic::floor: |
5164 | case Intrinsic::ceil: |
5165 | case Intrinsic::trunc: |
5166 | case Intrinsic::rint: |
5167 | case Intrinsic::nearbyint: |
5168 | case Intrinsic::round: { |
5169 | unsigned Opcode; |
5170 | switch (Intrinsic) { |
5171 | default: llvm_unreachable("Impossible intrinsic")::llvm::llvm_unreachable_internal("Impossible intrinsic", "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5171); // Can't reach here. |
5172 | case Intrinsic::sqrt: Opcode = ISD::FSQRT; break; |
5173 | case Intrinsic::fabs: Opcode = ISD::FABS; break; |
5174 | case Intrinsic::sin: Opcode = ISD::FSIN; break; |
5175 | case Intrinsic::cos: Opcode = ISD::FCOS; break; |
5176 | case Intrinsic::floor: Opcode = ISD::FFLOOR; break; |
5177 | case Intrinsic::ceil: Opcode = ISD::FCEIL; break; |
5178 | case Intrinsic::trunc: Opcode = ISD::FTRUNC; break; |
5179 | case Intrinsic::rint: Opcode = ISD::FRINT; break; |
5180 | case Intrinsic::nearbyint: Opcode = ISD::FNEARBYINT; break; |
5181 | case Intrinsic::round: Opcode = ISD::FROUND; break; |
5182 | } |
5183 | |
5184 | setValue(&I, DAG.getNode(Opcode, sdl, |
5185 | getValue(I.getArgOperand(0)).getValueType(), |
5186 | getValue(I.getArgOperand(0)))); |
5187 | return nullptr; |
5188 | } |
5189 | case Intrinsic::minnum: |
5190 | setValue(&I, DAG.getNode(ISD::FMINNUM, sdl, |
5191 | getValue(I.getArgOperand(0)).getValueType(), |
5192 | getValue(I.getArgOperand(0)), |
5193 | getValue(I.getArgOperand(1)))); |
5194 | return nullptr; |
5195 | case Intrinsic::maxnum: |
5196 | setValue(&I, DAG.getNode(ISD::FMAXNUM, sdl, |
5197 | getValue(I.getArgOperand(0)).getValueType(), |
5198 | getValue(I.getArgOperand(0)), |
5199 | getValue(I.getArgOperand(1)))); |
5200 | return nullptr; |
5201 | case Intrinsic::copysign: |
5202 | setValue(&I, DAG.getNode(ISD::FCOPYSIGN, sdl, |
5203 | getValue(I.getArgOperand(0)).getValueType(), |
5204 | getValue(I.getArgOperand(0)), |
5205 | getValue(I.getArgOperand(1)))); |
5206 | return nullptr; |
5207 | case Intrinsic::fma: |
5208 | setValue(&I, DAG.getNode(ISD::FMA, sdl, |
5209 | getValue(I.getArgOperand(0)).getValueType(), |
5210 | getValue(I.getArgOperand(0)), |
5211 | getValue(I.getArgOperand(1)), |
5212 | getValue(I.getArgOperand(2)))); |
5213 | return nullptr; |
5214 | case Intrinsic::fmuladd: { |
5215 | EVT VT = TLI.getValueType(I.getType()); |
5216 | if (TM.Options.AllowFPOpFusion != FPOpFusion::Strict && |
5217 | TLI.isFMAFasterThanFMulAndFAdd(VT)) { |
5218 | setValue(&I, DAG.getNode(ISD::FMA, sdl, |
5219 | getValue(I.getArgOperand(0)).getValueType(), |
5220 | getValue(I.getArgOperand(0)), |
5221 | getValue(I.getArgOperand(1)), |
5222 | getValue(I.getArgOperand(2)))); |
5223 | } else { |
5224 | SDValue Mul = DAG.getNode(ISD::FMUL, sdl, |
5225 | getValue(I.getArgOperand(0)).getValueType(), |
5226 | getValue(I.getArgOperand(0)), |
5227 | getValue(I.getArgOperand(1))); |
5228 | SDValue Add = DAG.getNode(ISD::FADD, sdl, |
5229 | getValue(I.getArgOperand(0)).getValueType(), |
5230 | Mul, |
5231 | getValue(I.getArgOperand(2))); |
5232 | setValue(&I, Add); |
5233 | } |
5234 | return nullptr; |
5235 | } |
5236 | case Intrinsic::convert_to_fp16: |
5237 | setValue(&I, DAG.getNode(ISD::BITCAST, sdl, MVT::i16, |
5238 | DAG.getNode(ISD::FP_ROUND, sdl, MVT::f16, |
5239 | getValue(I.getArgOperand(0)), |
5240 | DAG.getTargetConstant(0, MVT::i32)))); |
5241 | return nullptr; |
5242 | case Intrinsic::convert_from_fp16: |
5243 | setValue(&I, |
5244 | DAG.getNode(ISD::FP_EXTEND, sdl, TLI.getValueType(I.getType()), |
5245 | DAG.getNode(ISD::BITCAST, sdl, MVT::f16, |
5246 | getValue(I.getArgOperand(0))))); |
5247 | return nullptr; |
5248 | case Intrinsic::pcmarker: { |
5249 | SDValue Tmp = getValue(I.getArgOperand(0)); |
5250 | DAG.setRoot(DAG.getNode(ISD::PCMARKER, sdl, MVT::Other, getRoot(), Tmp)); |
5251 | return nullptr; |
5252 | } |
5253 | case Intrinsic::readcyclecounter: { |
5254 | SDValue Op = getRoot(); |
5255 | Res = DAG.getNode(ISD::READCYCLECOUNTER, sdl, |
5256 | DAG.getVTList(MVT::i64, MVT::Other), Op); |
5257 | setValue(&I, Res); |
5258 | DAG.setRoot(Res.getValue(1)); |
5259 | return nullptr; |
5260 | } |
5261 | case Intrinsic::bswap: |
5262 | setValue(&I, DAG.getNode(ISD::BSWAP, sdl, |
5263 | getValue(I.getArgOperand(0)).getValueType(), |
5264 | getValue(I.getArgOperand(0)))); |
5265 | return nullptr; |
5266 | case Intrinsic::cttz: { |
5267 | SDValue Arg = getValue(I.getArgOperand(0)); |
5268 | ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1)); |
5269 | EVT Ty = Arg.getValueType(); |
5270 | setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTTZ : ISD::CTTZ_ZERO_UNDEF, |
5271 | sdl, Ty, Arg)); |
5272 | return nullptr; |
5273 | } |
5274 | case Intrinsic::ctlz: { |
5275 | SDValue Arg = getValue(I.getArgOperand(0)); |
5276 | ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1)); |
5277 | EVT Ty = Arg.getValueType(); |
5278 | setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTLZ : ISD::CTLZ_ZERO_UNDEF, |
5279 | sdl, Ty, Arg)); |
5280 | return nullptr; |
5281 | } |
5282 | case Intrinsic::ctpop: { |
5283 | SDValue Arg = getValue(I.getArgOperand(0)); |
5284 | EVT Ty = Arg.getValueType(); |
5285 | setValue(&I, DAG.getNode(ISD::CTPOP, sdl, Ty, Arg)); |
5286 | return nullptr; |
5287 | } |
5288 | case Intrinsic::stacksave: { |
5289 | SDValue Op = getRoot(); |
5290 | Res = DAG.getNode(ISD::STACKSAVE, sdl, |
5291 | DAG.getVTList(TLI.getPointerTy(), MVT::Other), Op); |
5292 | setValue(&I, Res); |
5293 | DAG.setRoot(Res.getValue(1)); |
5294 | return nullptr; |
5295 | } |
5296 | case Intrinsic::stackrestore: { |
5297 | Res = getValue(I.getArgOperand(0)); |
5298 | DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, sdl, MVT::Other, getRoot(), Res)); |
5299 | return nullptr; |
5300 | } |
5301 | case Intrinsic::stackprotector: { |
5302 | // Emit code into the DAG to store the stack guard onto the stack. |
5303 | MachineFunction &MF = DAG.getMachineFunction(); |
5304 | MachineFrameInfo *MFI = MF.getFrameInfo(); |
5305 | EVT PtrTy = TLI.getPointerTy(); |
5306 | SDValue Src, Chain = getRoot(); |
5307 | const Value *Ptr = cast<LoadInst>(I.getArgOperand(0))->getPointerOperand(); |
5308 | const GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr); |
5309 | |
5310 | // See if Ptr is a bitcast. If it is, look through it and see if we can get |
5311 | // global variable __stack_chk_guard. |
5312 | if (!GV) |
5313 | if (const Operator *BC = dyn_cast<Operator>(Ptr)) |
5314 | if (BC->getOpcode() == Instruction::BitCast) |
5315 | GV = dyn_cast<GlobalVariable>(BC->getOperand(0)); |
5316 | |
5317 | if (GV && TLI.useLoadStackGuardNode()) { |
5318 | // Emit a LOAD_STACK_GUARD node. |
5319 | MachineSDNode *Node = DAG.getMachineNode(TargetOpcode::LOAD_STACK_GUARD, |
5320 | sdl, PtrTy, Chain); |
5321 | MachinePointerInfo MPInfo(GV); |
5322 | MachineInstr::mmo_iterator MemRefs = MF.allocateMemRefsArray(1); |
5323 | unsigned Flags = MachineMemOperand::MOLoad | |
5324 | MachineMemOperand::MOInvariant; |
5325 | *MemRefs = MF.getMachineMemOperand(MPInfo, Flags, |
5326 | PtrTy.getSizeInBits() / 8, |
5327 | DAG.getEVTAlignment(PtrTy)); |
5328 | Node->setMemRefs(MemRefs, MemRefs + 1); |
5329 | |
5330 | // Copy the guard value to a virtual register so that it can be |
5331 | // retrieved in the epilogue. |
5332 | Src = SDValue(Node, 0); |
5333 | const TargetRegisterClass *RC = |
5334 | TLI.getRegClassFor(Src.getSimpleValueType()); |
5335 | unsigned Reg = MF.getRegInfo().createVirtualRegister(RC); |
5336 | |
5337 | SPDescriptor.setGuardReg(Reg); |
5338 | Chain = DAG.getCopyToReg(Chain, sdl, Reg, Src); |
5339 | } else { |
5340 | Src = getValue(I.getArgOperand(0)); // The guard's value. |
5341 | } |
5342 | |
5343 | AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1)); |
5344 | |
5345 | int FI = FuncInfo.StaticAllocaMap[Slot]; |
5346 | MFI->setStackProtectorIndex(FI); |
5347 | |
5348 | SDValue FIN = DAG.getFrameIndex(FI, PtrTy); |
5349 | |
5350 | // Store the stack protector onto the stack. |
5351 | Res = DAG.getStore(Chain, sdl, Src, FIN, |
5352 | MachinePointerInfo::getFixedStack(FI), |
5353 | true, false, 0); |
5354 | setValue(&I, Res); |
5355 | DAG.setRoot(Res); |
5356 | return nullptr; |
5357 | } |
5358 | case Intrinsic::objectsize: { |
5359 | // If we don't know by now, we're never going to know. |
5360 | ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1)); |
5361 | |
5362 | assert(CI && "Non-constant type in __builtin_object_size?")((CI && "Non-constant type in __builtin_object_size?" ) ? static_cast<void> (0) : __assert_fail ("CI && \"Non-constant type in __builtin_object_size?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5362, __PRETTY_FUNCTION__)); |
5363 | |
5364 | SDValue Arg = getValue(I.getCalledValue()); |
5365 | EVT Ty = Arg.getValueType(); |
5366 | |
5367 | if (CI->isZero()) |
5368 | Res = DAG.getConstant(-1ULL, Ty); |
5369 | else |
5370 | Res = DAG.getConstant(0, Ty); |
5371 | |
5372 | setValue(&I, Res); |
5373 | return nullptr; |
5374 | } |
5375 | case Intrinsic::annotation: |
5376 | case Intrinsic::ptr_annotation: |
5377 | // Drop the intrinsic, but forward the value |
5378 | setValue(&I, getValue(I.getOperand(0))); |
5379 | return nullptr; |
5380 | case Intrinsic::assume: |
5381 | case Intrinsic::var_annotation: |
5382 | // Discard annotate attributes and assumptions |
5383 | return nullptr; |
5384 | |
5385 | case Intrinsic::init_trampoline: { |
5386 | const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts()); |
5387 | |
5388 | SDValue Ops[6]; |
5389 | Ops[0] = getRoot(); |
5390 | Ops[1] = getValue(I.getArgOperand(0)); |
5391 | Ops[2] = getValue(I.getArgOperand(1)); |
5392 | Ops[3] = getValue(I.getArgOperand(2)); |
5393 | Ops[4] = DAG.getSrcValue(I.getArgOperand(0)); |
5394 | Ops[5] = DAG.getSrcValue(F); |
5395 | |
5396 | Res = DAG.getNode(ISD::INIT_TRAMPOLINE, sdl, MVT::Other, Ops); |
5397 | |
5398 | DAG.setRoot(Res); |
5399 | return nullptr; |
5400 | } |
5401 | case Intrinsic::adjust_trampoline: { |
5402 | setValue(&I, DAG.getNode(ISD::ADJUST_TRAMPOLINE, sdl, |
5403 | TLI.getPointerTy(), |
5404 | getValue(I.getArgOperand(0)))); |
5405 | return nullptr; |
5406 | } |
5407 | case Intrinsic::gcroot: |
5408 | if (GFI) { |
5409 | const Value *Alloca = I.getArgOperand(0)->stripPointerCasts(); |
5410 | const Constant *TypeMap = cast<Constant>(I.getArgOperand(1)); |
5411 | |
5412 | FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode()); |
5413 | GFI->addStackRoot(FI->getIndex(), TypeMap); |
5414 | } |
5415 | return nullptr; |
5416 | case Intrinsic::gcread: |
5417 | case Intrinsic::gcwrite: |
5418 | llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!")::llvm::llvm_unreachable_internal("GC failed to lower gcread/gcwrite intrinsics!" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5418); |
5419 | case Intrinsic::flt_rounds: |
5420 | setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, sdl, MVT::i32)); |
5421 | return nullptr; |
5422 | |
5423 | case Intrinsic::expect: { |
5424 | // Just replace __builtin_expect(exp, c) with EXP. |
5425 | setValue(&I, getValue(I.getArgOperand(0))); |
5426 | return nullptr; |
5427 | } |
5428 | |
5429 | case Intrinsic::debugtrap: |
5430 | case Intrinsic::trap: { |
5431 | StringRef TrapFuncName = TM.Options.getTrapFunctionName(); |
5432 | if (TrapFuncName.empty()) { |
5433 | ISD::NodeType Op = (Intrinsic == Intrinsic::trap) ? |
5434 | ISD::TRAP : ISD::DEBUGTRAP; |
5435 | DAG.setRoot(DAG.getNode(Op, sdl,MVT::Other, getRoot())); |
5436 | return nullptr; |
5437 | } |
5438 | TargetLowering::ArgListTy Args; |
5439 | |
5440 | TargetLowering::CallLoweringInfo CLI(DAG); |
5441 | CLI.setDebugLoc(sdl).setChain(getRoot()) |
5442 | .setCallee(CallingConv::C, I.getType(), |
5443 | DAG.getExternalSymbol(TrapFuncName.data(), TLI.getPointerTy()), |
5444 | std::move(Args), 0); |
5445 | |
5446 | std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI); |
5447 | DAG.setRoot(Result.second); |
5448 | return nullptr; |
5449 | } |
5450 | |
5451 | case Intrinsic::uadd_with_overflow: |
5452 | case Intrinsic::sadd_with_overflow: |
5453 | case Intrinsic::usub_with_overflow: |
5454 | case Intrinsic::ssub_with_overflow: |
5455 | case Intrinsic::umul_with_overflow: |
5456 | case Intrinsic::smul_with_overflow: { |
5457 | ISD::NodeType Op; |
5458 | switch (Intrinsic) { |
5459 | default: llvm_unreachable("Impossible intrinsic")::llvm::llvm_unreachable_internal("Impossible intrinsic", "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5459); // Can't reach here. |
5460 | case Intrinsic::uadd_with_overflow: Op = ISD::UADDO; break; |
5461 | case Intrinsic::sadd_with_overflow: Op = ISD::SADDO; break; |
5462 | case Intrinsic::usub_with_overflow: Op = ISD::USUBO; break; |
5463 | case Intrinsic::ssub_with_overflow: Op = ISD::SSUBO; break; |
5464 | case Intrinsic::umul_with_overflow: Op = ISD::UMULO; break; |
5465 | case Intrinsic::smul_with_overflow: Op = ISD::SMULO; break; |
5466 | } |
5467 | SDValue Op1 = getValue(I.getArgOperand(0)); |
5468 | SDValue Op2 = getValue(I.getArgOperand(1)); |
5469 | |
5470 | SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1); |
5471 | setValue(&I, DAG.getNode(Op, sdl, VTs, Op1, Op2)); |
5472 | return nullptr; |
5473 | } |
5474 | case Intrinsic::prefetch: { |
5475 | SDValue Ops[5]; |
5476 | unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue(); |
5477 | Ops[0] = getRoot(); |
5478 | Ops[1] = getValue(I.getArgOperand(0)); |
5479 | Ops[2] = getValue(I.getArgOperand(1)); |
5480 | Ops[3] = getValue(I.getArgOperand(2)); |
5481 | Ops[4] = getValue(I.getArgOperand(3)); |
5482 | DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, sdl, |
5483 | DAG.getVTList(MVT::Other), Ops, |
5484 | EVT::getIntegerVT(*Context, 8), |
5485 | MachinePointerInfo(I.getArgOperand(0)), |
5486 | 0, /* align */ |
5487 | false, /* volatile */ |
5488 | rw==0, /* read */ |
5489 | rw==1)); /* write */ |
5490 | return nullptr; |
5491 | } |
5492 | case Intrinsic::lifetime_start: |
5493 | case Intrinsic::lifetime_end: { |
5494 | bool IsStart = (Intrinsic == Intrinsic::lifetime_start); |
5495 | // Stack coloring is not enabled in O0, discard region information. |
5496 | if (TM.getOptLevel() == CodeGenOpt::None) |
5497 | return nullptr; |
5498 | |
5499 | SmallVector<Value *, 4> Allocas; |
5500 | GetUnderlyingObjects(I.getArgOperand(1), Allocas, DL); |
5501 | |
5502 | for (SmallVectorImpl<Value*>::iterator Object = Allocas.begin(), |
5503 | E = Allocas.end(); Object != E; ++Object) { |
5504 | AllocaInst *LifetimeObject = dyn_cast_or_null<AllocaInst>(*Object); |
5505 | |
5506 | // Could not find an Alloca. |
5507 | if (!LifetimeObject) |
5508 | continue; |
5509 | |
5510 | // First check that the Alloca is static, otherwise it won't have a |
5511 | // valid frame index. |
5512 | auto SI = FuncInfo.StaticAllocaMap.find(LifetimeObject); |
5513 | if (SI == FuncInfo.StaticAllocaMap.end()) |
5514 | return nullptr; |
5515 | |
5516 | int FI = SI->second; |
5517 | |
5518 | SDValue Ops[2]; |
5519 | Ops[0] = getRoot(); |
5520 | Ops[1] = DAG.getFrameIndex(FI, TLI.getPointerTy(), true); |
5521 | unsigned Opcode = (IsStart ? ISD::LIFETIME_START : ISD::LIFETIME_END); |
5522 | |
5523 | Res = DAG.getNode(Opcode, sdl, MVT::Other, Ops); |
5524 | DAG.setRoot(Res); |
5525 | } |
5526 | return nullptr; |
5527 | } |
5528 | case Intrinsic::invariant_start: |
5529 | // Discard region information. |
5530 | setValue(&I, DAG.getUNDEF(TLI.getPointerTy())); |
5531 | return nullptr; |
5532 | case Intrinsic::invariant_end: |
5533 | // Discard region information. |
5534 | return nullptr; |
5535 | case Intrinsic::stackprotectorcheck: { |
5536 | // Do not actually emit anything for this basic block. Instead we initialize |
5537 | // the stack protector descriptor and export the guard variable so we can |
5538 | // access it in FinishBasicBlock. |
5539 | const BasicBlock *BB = I.getParent(); |
5540 | SPDescriptor.initialize(BB, FuncInfo.MBBMap[BB], I); |
5541 | ExportFromCurrentBlock(SPDescriptor.getGuard()); |
5542 | |
5543 | // Flush our exports since we are going to process a terminator. |
5544 | (void)getControlRoot(); |
5545 | return nullptr; |
5546 | } |
5547 | case Intrinsic::clear_cache: |
5548 | return TLI.getClearCacheBuiltinName(); |
5549 | case Intrinsic::donothing: |
5550 | // ignore |
5551 | return nullptr; |
5552 | case Intrinsic::experimental_stackmap: { |
5553 | visitStackmap(I); |
5554 | return nullptr; |
5555 | } |
5556 | case Intrinsic::experimental_patchpoint_void: |
5557 | case Intrinsic::experimental_patchpoint_i64: { |
5558 | visitPatchpoint(&I); |
5559 | return nullptr; |
5560 | } |
5561 | case Intrinsic::experimental_gc_statepoint: { |
5562 | visitStatepoint(I); |
5563 | return nullptr; |
5564 | } |
5565 | case Intrinsic::experimental_gc_result_int: |
5566 | case Intrinsic::experimental_gc_result_float: |
5567 | case Intrinsic::experimental_gc_result_ptr: { |
5568 | visitGCResult(I); |
5569 | return nullptr; |
5570 | } |
5571 | case Intrinsic::experimental_gc_relocate: { |
5572 | visitGCRelocate(I); |
5573 | return nullptr; |
5574 | } |
5575 | case Intrinsic::instrprof_increment: |
5576 | llvm_unreachable("instrprof failed to lower an increment")::llvm::llvm_unreachable_internal("instrprof failed to lower an increment" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5576); |
5577 | } |
5578 | } |
5579 | |
5580 | std::pair<SDValue, SDValue> |
5581 | SelectionDAGBuilder::lowerInvokable(TargetLowering::CallLoweringInfo &CLI, |
5582 | MachineBasicBlock *LandingPad) { |
5583 | MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); |
5584 | MCSymbol *BeginLabel = nullptr; |
5585 | |
5586 | if (LandingPad) { |
5587 | // Insert a label before the invoke call to mark the try range. This can be |
5588 | // used to detect deletion of the invoke via the MachineModuleInfo. |
5589 | BeginLabel = MMI.getContext().CreateTempSymbol(); |
5590 | |
5591 | // For SjLj, keep track of which landing pads go with which invokes |
5592 | // so as to maintain the ordering of pads in the LSDA. |
5593 | unsigned CallSiteIndex = MMI.getCurrentCallSite(); |
5594 | if (CallSiteIndex) { |
5595 | MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex); |
5596 | LPadToCallSiteMap[LandingPad].push_back(CallSiteIndex); |
5597 | |
5598 | // Now that the call site is handled, stop tracking it. |
5599 | MMI.setCurrentCallSite(0); |
5600 | } |
5601 | |
5602 | // Both PendingLoads and PendingExports must be flushed here; |
5603 | // this call might not return. |
5604 | (void)getRoot(); |
5605 | DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getControlRoot(), BeginLabel)); |
5606 | |
5607 | CLI.setChain(getRoot()); |
5608 | } |
5609 | |
5610 | const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering(); |
5611 | std::pair<SDValue, SDValue> Result = TLI->LowerCallTo(CLI); |
5612 | |
5613 | assert((CLI.IsTailCall || Result.second.getNode()) &&(((CLI.IsTailCall || Result.second.getNode()) && "Non-null chain expected with non-tail call!" ) ? static_cast<void> (0) : __assert_fail ("(CLI.IsTailCall || Result.second.getNode()) && \"Non-null chain expected with non-tail call!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5614, __PRETTY_FUNCTION__)) |
5614 | "Non-null chain expected with non-tail call!")(((CLI.IsTailCall || Result.second.getNode()) && "Non-null chain expected with non-tail call!" ) ? static_cast<void> (0) : __assert_fail ("(CLI.IsTailCall || Result.second.getNode()) && \"Non-null chain expected with non-tail call!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5614, __PRETTY_FUNCTION__)); |
5615 | assert((Result.second.getNode() || !Result.first.getNode()) &&(((Result.second.getNode() || !Result.first.getNode()) && "Null value expected with tail call!") ? static_cast<void > (0) : __assert_fail ("(Result.second.getNode() || !Result.first.getNode()) && \"Null value expected with tail call!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5616, __PRETTY_FUNCTION__)) |
5616 | "Null value expected with tail call!")(((Result.second.getNode() || !Result.first.getNode()) && "Null value expected with tail call!") ? static_cast<void > (0) : __assert_fail ("(Result.second.getNode() || !Result.first.getNode()) && \"Null value expected with tail call!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5616, __PRETTY_FUNCTION__)); |
5617 | |
5618 | if (!Result.second.getNode()) { |
5619 | // As a special case, a null chain means that a tail call has been emitted |
5620 | // and the DAG root is already updated. |
5621 | HasTailCall = true; |
5622 | |
5623 | // Since there's no actual continuation from this block, nothing can be |
5624 | // relying on us setting vregs for them. |
5625 | PendingExports.clear(); |
5626 | } else { |
5627 | DAG.setRoot(Result.second); |
5628 | } |
5629 | |
5630 | if (LandingPad) { |
5631 | // Insert a label at the end of the invoke call to mark the try range. This |
5632 | // can be used to detect deletion of the invoke via the MachineModuleInfo. |
5633 | MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol(); |
5634 | DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getRoot(), EndLabel)); |
5635 | |
5636 | // Inform MachineModuleInfo of range. |
5637 | MMI.addInvoke(LandingPad, BeginLabel, EndLabel); |
5638 | } |
5639 | |
5640 | return Result; |
5641 | } |
5642 | |
5643 | void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee, |
5644 | bool isTailCall, |
5645 | MachineBasicBlock *LandingPad) { |
5646 | PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType()); |
5647 | FunctionType *FTy = cast<FunctionType>(PT->getElementType()); |
5648 | Type *RetTy = FTy->getReturnType(); |
5649 | |
5650 | TargetLowering::ArgListTy Args; |
5651 | TargetLowering::ArgListEntry Entry; |
5652 | Args.reserve(CS.arg_size()); |
5653 | |
5654 | for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end(); |
5655 | i != e; ++i) { |
5656 | const Value *V = *i; |
5657 | |
5658 | // Skip empty types |
5659 | if (V->getType()->isEmptyTy()) |
5660 | continue; |
5661 | |
5662 | SDValue ArgNode = getValue(V); |
5663 | Entry.Node = ArgNode; Entry.Ty = V->getType(); |
5664 | |
5665 | // Skip the first return-type Attribute to get to params. |
5666 | Entry.setAttributes(&CS, i - CS.arg_begin() + 1); |
5667 | Args.push_back(Entry); |
5668 | } |
5669 | |
5670 | // Check if target-independent constraints permit a tail call here. |
5671 | // Target-dependent constraints are checked within TLI->LowerCallTo. |
5672 | if (isTailCall && !isInTailCallPosition(CS, DAG.getTarget())) |
5673 | isTailCall = false; |
5674 | |
5675 | TargetLowering::CallLoweringInfo CLI(DAG); |
5676 | CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot()) |
5677 | .setCallee(RetTy, FTy, Callee, std::move(Args), CS) |
5678 | .setTailCall(isTailCall); |
5679 | std::pair<SDValue,SDValue> Result = lowerInvokable(CLI, LandingPad); |
5680 | |
5681 | if (Result.first.getNode()) |
5682 | setValue(CS.getInstruction(), Result.first); |
5683 | } |
5684 | |
5685 | /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the |
5686 | /// value is equal or not-equal to zero. |
5687 | static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) { |
5688 | for (const User *U : V->users()) { |
5689 | if (const ICmpInst *IC = dyn_cast<ICmpInst>(U)) |
5690 | if (IC->isEquality()) |
5691 | if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1))) |
5692 | if (C->isNullValue()) |
5693 | continue; |
5694 | // Unknown instruction. |
5695 | return false; |
5696 | } |
5697 | return true; |
5698 | } |
5699 | |
5700 | static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT, |
5701 | Type *LoadTy, |
5702 | SelectionDAGBuilder &Builder) { |
5703 | |
5704 | // Check to see if this load can be trivially constant folded, e.g. if the |
5705 | // input is from a string literal. |
5706 | if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) { |
5707 | // Cast pointer to the type we really want to load. |
5708 | LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput), |
5709 | PointerType::getUnqual(LoadTy)); |
5710 | |
5711 | if (const Constant *LoadCst = |
5712 | ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput), |
5713 | Builder.DL)) |
5714 | return Builder.getValue(LoadCst); |
5715 | } |
5716 | |
5717 | // Otherwise, we have to emit the load. If the pointer is to unfoldable but |
5718 | // still constant memory, the input chain can be the entry node. |
5719 | SDValue Root; |
5720 | bool ConstantMemory = false; |
5721 | |
5722 | // Do not serialize (non-volatile) loads of constant memory with anything. |
5723 | if (Builder.AA->pointsToConstantMemory(PtrVal)) { |
5724 | Root = Builder.DAG.getEntryNode(); |
5725 | ConstantMemory = true; |
5726 | } else { |
5727 | // Do not serialize non-volatile loads against each other. |
5728 | Root = Builder.DAG.getRoot(); |
5729 | } |
5730 | |
5731 | SDValue Ptr = Builder.getValue(PtrVal); |
5732 | SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurSDLoc(), Root, |
5733 | Ptr, MachinePointerInfo(PtrVal), |
5734 | false /*volatile*/, |
5735 | false /*nontemporal*/, |
5736 | false /*isinvariant*/, 1 /* align=1 */); |
5737 | |
5738 | if (!ConstantMemory) |
5739 | Builder.PendingLoads.push_back(LoadVal.getValue(1)); |
5740 | return LoadVal; |
5741 | } |
5742 | |
5743 | /// processIntegerCallValue - Record the value for an instruction that |
5744 | /// produces an integer result, converting the type where necessary. |
5745 | void SelectionDAGBuilder::processIntegerCallValue(const Instruction &I, |
5746 | SDValue Value, |
5747 | bool IsSigned) { |
5748 | EVT VT = DAG.getTargetLoweringInfo().getValueType(I.getType(), true); |
5749 | if (IsSigned) |
5750 | Value = DAG.getSExtOrTrunc(Value, getCurSDLoc(), VT); |
5751 | else |
5752 | Value = DAG.getZExtOrTrunc(Value, getCurSDLoc(), VT); |
5753 | setValue(&I, Value); |
5754 | } |
5755 | |
5756 | /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form. |
5757 | /// If so, return true and lower it, otherwise return false and it will be |
5758 | /// lowered like a normal call. |
5759 | bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) { |
5760 | // Verify that the prototype makes sense. int memcmp(void*,void*,size_t) |
5761 | if (I.getNumArgOperands() != 3) |
5762 | return false; |
5763 | |
5764 | const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1); |
5765 | if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() || |
5766 | !I.getArgOperand(2)->getType()->isIntegerTy() || |
5767 | !I.getType()->isIntegerTy()) |
5768 | return false; |
5769 | |
5770 | const Value *Size = I.getArgOperand(2); |
5771 | const ConstantInt *CSize = dyn_cast<ConstantInt>(Size); |
5772 | if (CSize && CSize->getZExtValue() == 0) { |
5773 | EVT CallVT = DAG.getTargetLoweringInfo().getValueType(I.getType(), true); |
5774 | setValue(&I, DAG.getConstant(0, CallVT)); |
5775 | return true; |
5776 | } |
5777 | |
5778 | const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); |
5779 | std::pair<SDValue, SDValue> Res = |
5780 | TSI.EmitTargetCodeForMemcmp(DAG, getCurSDLoc(), DAG.getRoot(), |
5781 | getValue(LHS), getValue(RHS), getValue(Size), |
5782 | MachinePointerInfo(LHS), |
5783 | MachinePointerInfo(RHS)); |
5784 | if (Res.first.getNode()) { |
5785 | processIntegerCallValue(I, Res.first, true); |
5786 | PendingLoads.push_back(Res.second); |
5787 | return true; |
5788 | } |
5789 | |
5790 | // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0 |
5791 | // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0 |
5792 | if (CSize && IsOnlyUsedInZeroEqualityComparison(&I)) { |
5793 | bool ActuallyDoIt = true; |
5794 | MVT LoadVT; |
5795 | Type *LoadTy; |
5796 | switch (CSize->getZExtValue()) { |
5797 | default: |
5798 | LoadVT = MVT::Other; |
5799 | LoadTy = nullptr; |
5800 | ActuallyDoIt = false; |
5801 | break; |
5802 | case 2: |
5803 | LoadVT = MVT::i16; |
5804 | LoadTy = Type::getInt16Ty(CSize->getContext()); |
5805 | break; |
5806 | case 4: |
5807 | LoadVT = MVT::i32; |
5808 | LoadTy = Type::getInt32Ty(CSize->getContext()); |
5809 | break; |
5810 | case 8: |
5811 | LoadVT = MVT::i64; |
5812 | LoadTy = Type::getInt64Ty(CSize->getContext()); |
5813 | break; |
5814 | /* |
5815 | case 16: |
5816 | LoadVT = MVT::v4i32; |
5817 | LoadTy = Type::getInt32Ty(CSize->getContext()); |
5818 | LoadTy = VectorType::get(LoadTy, 4); |
5819 | break; |
5820 | */ |
5821 | } |
5822 | |
5823 | // This turns into unaligned loads. We only do this if the target natively |
5824 | // supports the MVT we'll be loading or if it is small enough (<= 4) that |
5825 | // we'll only produce a small number of byte loads. |
5826 | |
5827 | // Require that we can find a legal MVT, and only do this if the target |
5828 | // supports unaligned loads of that type. Expanding into byte loads would |
5829 | // bloat the code. |
5830 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
5831 | if (ActuallyDoIt && CSize->getZExtValue() > 4) { |
5832 | unsigned DstAS = LHS->getType()->getPointerAddressSpace(); |
5833 | unsigned SrcAS = RHS->getType()->getPointerAddressSpace(); |
5834 | // TODO: Handle 5 byte compare as 4-byte + 1 byte. |
5835 | // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads. |
5836 | // TODO: Check alignment of src and dest ptrs. |
5837 | if (!TLI.isTypeLegal(LoadVT) || |
5838 | !TLI.allowsMisalignedMemoryAccesses(LoadVT, SrcAS) || |
5839 | !TLI.allowsMisalignedMemoryAccesses(LoadVT, DstAS)) |
5840 | ActuallyDoIt = false; |
5841 | } |
5842 | |
5843 | if (ActuallyDoIt) { |
5844 | SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this); |
5845 | SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this); |
5846 | |
5847 | SDValue Res = DAG.getSetCC(getCurSDLoc(), MVT::i1, LHSVal, RHSVal, |
5848 | ISD::SETNE); |
5849 | processIntegerCallValue(I, Res, false); |
5850 | return true; |
5851 | } |
5852 | } |
5853 | |
5854 | |
5855 | return false; |
5856 | } |
5857 | |
5858 | /// visitMemChrCall -- See if we can lower a memchr call into an optimized |
5859 | /// form. If so, return true and lower it, otherwise return false and it |
5860 | /// will be lowered like a normal call. |
5861 | bool SelectionDAGBuilder::visitMemChrCall(const CallInst &I) { |
5862 | // Verify that the prototype makes sense. void *memchr(void *, int, size_t) |
5863 | if (I.getNumArgOperands() != 3) |
5864 | return false; |
5865 | |
5866 | const Value *Src = I.getArgOperand(0); |
5867 | const Value *Char = I.getArgOperand(1); |
5868 | const Value *Length = I.getArgOperand(2); |
5869 | if (!Src->getType()->isPointerTy() || |
5870 | !Char->getType()->isIntegerTy() || |
5871 | !Length->getType()->isIntegerTy() || |
5872 | !I.getType()->isPointerTy()) |
5873 | return false; |
5874 | |
5875 | const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); |
5876 | std::pair<SDValue, SDValue> Res = |
5877 | TSI.EmitTargetCodeForMemchr(DAG, getCurSDLoc(), DAG.getRoot(), |
5878 | getValue(Src), getValue(Char), getValue(Length), |
5879 | MachinePointerInfo(Src)); |
5880 | if (Res.first.getNode()) { |
5881 | setValue(&I, Res.first); |
5882 | PendingLoads.push_back(Res.second); |
5883 | return true; |
5884 | } |
5885 | |
5886 | return false; |
5887 | } |
5888 | |
5889 | /// visitStrCpyCall -- See if we can lower a strcpy or stpcpy call into an |
5890 | /// optimized form. If so, return true and lower it, otherwise return false |
5891 | /// and it will be lowered like a normal call. |
5892 | bool SelectionDAGBuilder::visitStrCpyCall(const CallInst &I, bool isStpcpy) { |
5893 | // Verify that the prototype makes sense. char *strcpy(char *, char *) |
5894 | if (I.getNumArgOperands() != 2) |
5895 | return false; |
5896 | |
5897 | const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1); |
5898 | if (!Arg0->getType()->isPointerTy() || |
5899 | !Arg1->getType()->isPointerTy() || |
5900 | !I.getType()->isPointerTy()) |
5901 | return false; |
5902 | |
5903 | const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); |
5904 | std::pair<SDValue, SDValue> Res = |
5905 | TSI.EmitTargetCodeForStrcpy(DAG, getCurSDLoc(), getRoot(), |
5906 | getValue(Arg0), getValue(Arg1), |
5907 | MachinePointerInfo(Arg0), |
5908 | MachinePointerInfo(Arg1), isStpcpy); |
5909 | if (Res.first.getNode()) { |
5910 | setValue(&I, Res.first); |
5911 | DAG.setRoot(Res.second); |
5912 | return true; |
5913 | } |
5914 | |
5915 | return false; |
5916 | } |
5917 | |
5918 | /// visitStrCmpCall - See if we can lower a call to strcmp in an optimized form. |
5919 | /// If so, return true and lower it, otherwise return false and it will be |
5920 | /// lowered like a normal call. |
5921 | bool SelectionDAGBuilder::visitStrCmpCall(const CallInst &I) { |
5922 | // Verify that the prototype makes sense. int strcmp(void*,void*) |
5923 | if (I.getNumArgOperands() != 2) |
5924 | return false; |
5925 | |
5926 | const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1); |
5927 | if (!Arg0->getType()->isPointerTy() || |
5928 | !Arg1->getType()->isPointerTy() || |
5929 | !I.getType()->isIntegerTy()) |
5930 | return false; |
5931 | |
5932 | const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); |
5933 | std::pair<SDValue, SDValue> Res = |
5934 | TSI.EmitTargetCodeForStrcmp(DAG, getCurSDLoc(), DAG.getRoot(), |
5935 | getValue(Arg0), getValue(Arg1), |
5936 | MachinePointerInfo(Arg0), |
5937 | MachinePointerInfo(Arg1)); |
5938 | if (Res.first.getNode()) { |
5939 | processIntegerCallValue(I, Res.first, true); |
5940 | PendingLoads.push_back(Res.second); |
5941 | return true; |
5942 | } |
5943 | |
5944 | return false; |
5945 | } |
5946 | |
5947 | /// visitStrLenCall -- See if we can lower a strlen call into an optimized |
5948 | /// form. If so, return true and lower it, otherwise return false and it |
5949 | /// will be lowered like a normal call. |
5950 | bool SelectionDAGBuilder::visitStrLenCall(const CallInst &I) { |
5951 | // Verify that the prototype makes sense. size_t strlen(char *) |
5952 | if (I.getNumArgOperands() != 1) |
5953 | return false; |
5954 | |
5955 | const Value *Arg0 = I.getArgOperand(0); |
5956 | if (!Arg0->getType()->isPointerTy() || !I.getType()->isIntegerTy()) |
5957 | return false; |
5958 | |
5959 | const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); |
5960 | std::pair<SDValue, SDValue> Res = |
5961 | TSI.EmitTargetCodeForStrlen(DAG, getCurSDLoc(), DAG.getRoot(), |
5962 | getValue(Arg0), MachinePointerInfo(Arg0)); |
5963 | if (Res.first.getNode()) { |
5964 | processIntegerCallValue(I, Res.first, false); |
5965 | PendingLoads.push_back(Res.second); |
5966 | return true; |
5967 | } |
5968 | |
5969 | return false; |
5970 | } |
5971 | |
5972 | /// visitStrNLenCall -- See if we can lower a strnlen call into an optimized |
5973 | /// form. If so, return true and lower it, otherwise return false and it |
5974 | /// will be lowered like a normal call. |
5975 | bool SelectionDAGBuilder::visitStrNLenCall(const CallInst &I) { |
5976 | // Verify that the prototype makes sense. size_t strnlen(char *, size_t) |
5977 | if (I.getNumArgOperands() != 2) |
5978 | return false; |
5979 | |
5980 | const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1); |
5981 | if (!Arg0->getType()->isPointerTy() || |
5982 | !Arg1->getType()->isIntegerTy() || |
5983 | !I.getType()->isIntegerTy()) |
5984 | return false; |
5985 | |
5986 | const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); |
5987 | std::pair<SDValue, SDValue> Res = |
5988 | TSI.EmitTargetCodeForStrnlen(DAG, getCurSDLoc(), DAG.getRoot(), |
5989 | getValue(Arg0), getValue(Arg1), |
5990 | MachinePointerInfo(Arg0)); |
5991 | if (Res.first.getNode()) { |
5992 | processIntegerCallValue(I, Res.first, false); |
5993 | PendingLoads.push_back(Res.second); |
5994 | return true; |
5995 | } |
5996 | |
5997 | return false; |
5998 | } |
5999 | |
6000 | /// visitUnaryFloatCall - If a call instruction is a unary floating-point |
6001 | /// operation (as expected), translate it to an SDNode with the specified opcode |
6002 | /// and return true. |
6003 | bool SelectionDAGBuilder::visitUnaryFloatCall(const CallInst &I, |
6004 | unsigned Opcode) { |
6005 | // Sanity check that it really is a unary floating-point call. |
6006 | if (I.getNumArgOperands() != 1 || |
6007 | !I.getArgOperand(0)->getType()->isFloatingPointTy() || |
6008 | I.getType() != I.getArgOperand(0)->getType() || |
6009 | !I.onlyReadsMemory()) |
6010 | return false; |
6011 | |
6012 | SDValue Tmp = getValue(I.getArgOperand(0)); |
6013 | setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), Tmp.getValueType(), Tmp)); |
6014 | return true; |
6015 | } |
6016 | |
6017 | /// visitBinaryFloatCall - If a call instruction is a binary floating-point |
6018 | /// operation (as expected), translate it to an SDNode with the specified opcode |
6019 | /// and return true. |
6020 | bool SelectionDAGBuilder::visitBinaryFloatCall(const CallInst &I, |
6021 | unsigned Opcode) { |
6022 | // Sanity check that it really is a binary floating-point call. |
6023 | if (I.getNumArgOperands() != 2 || |
6024 | !I.getArgOperand(0)->getType()->isFloatingPointTy() || |
6025 | I.getType() != I.getArgOperand(0)->getType() || |
6026 | I.getType() != I.getArgOperand(1)->getType() || |
6027 | !I.onlyReadsMemory()) |
6028 | return false; |
6029 | |
6030 | SDValue Tmp0 = getValue(I.getArgOperand(0)); |
6031 | SDValue Tmp1 = getValue(I.getArgOperand(1)); |
6032 | EVT VT = Tmp0.getValueType(); |
6033 | setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), VT, Tmp0, Tmp1)); |
6034 | return true; |
6035 | } |
6036 | |
6037 | void SelectionDAGBuilder::visitCall(const CallInst &I) { |
6038 | // Handle inline assembly differently. |
6039 | if (isa<InlineAsm>(I.getCalledValue())) { |
6040 | visitInlineAsm(&I); |
6041 | return; |
6042 | } |
6043 | |
6044 | MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); |
6045 | ComputeUsesVAFloatArgument(I, &MMI); |
6046 | |
6047 | const char *RenameFn = nullptr; |
6048 | if (Function *F = I.getCalledFunction()) { |
6049 | if (F->isDeclaration()) { |
6050 | if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) { |
6051 | if (unsigned IID = II->getIntrinsicID(F)) { |
6052 | RenameFn = visitIntrinsicCall(I, IID); |
6053 | if (!RenameFn) |
6054 | return; |
6055 | } |
6056 | } |
6057 | if (unsigned IID = F->getIntrinsicID()) { |
6058 | RenameFn = visitIntrinsicCall(I, IID); |
6059 | if (!RenameFn) |
6060 | return; |
6061 | } |
6062 | } |
6063 | |
6064 | // Check for well-known libc/libm calls. If the function is internal, it |
6065 | // can't be a library call. |
6066 | LibFunc::Func Func; |
6067 | if (!F->hasLocalLinkage() && F->hasName() && |
6068 | LibInfo->getLibFunc(F->getName(), Func) && |
6069 | LibInfo->hasOptimizedCodeGen(Func)) { |
6070 | switch (Func) { |
6071 | default: break; |
6072 | case LibFunc::copysign: |
6073 | case LibFunc::copysignf: |
6074 | case LibFunc::copysignl: |
6075 | if (I.getNumArgOperands() == 2 && // Basic sanity checks. |
6076 | I.getArgOperand(0)->getType()->isFloatingPointTy() && |
6077 | I.getType() == I.getArgOperand(0)->getType() && |
6078 | I.getType() == I.getArgOperand(1)->getType() && |
6079 | I.onlyReadsMemory()) { |
6080 | SDValue LHS = getValue(I.getArgOperand(0)); |
6081 | SDValue RHS = getValue(I.getArgOperand(1)); |
6082 | setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurSDLoc(), |
6083 | LHS.getValueType(), LHS, RHS)); |
6084 | return; |
6085 | } |
6086 | break; |
6087 | case LibFunc::fabs: |
6088 | case LibFunc::fabsf: |
6089 | case LibFunc::fabsl: |
6090 | if (visitUnaryFloatCall(I, ISD::FABS)) |
6091 | return; |
6092 | break; |
6093 | case LibFunc::fmin: |
6094 | case LibFunc::fminf: |
6095 | case LibFunc::fminl: |
6096 | if (visitBinaryFloatCall(I, ISD::FMINNUM)) |
6097 | return; |
6098 | break; |
6099 | case LibFunc::fmax: |
6100 | case LibFunc::fmaxf: |
6101 | case LibFunc::fmaxl: |
6102 | if (visitBinaryFloatCall(I, ISD::FMAXNUM)) |
6103 | return; |
6104 | break; |
6105 | case LibFunc::sin: |
6106 | case LibFunc::sinf: |
6107 | case LibFunc::sinl: |
6108 | if (visitUnaryFloatCall(I, ISD::FSIN)) |
6109 | return; |
6110 | break; |
6111 | case LibFunc::cos: |
6112 | case LibFunc::cosf: |
6113 | case LibFunc::cosl: |
6114 | if (visitUnaryFloatCall(I, ISD::FCOS)) |
6115 | return; |
6116 | break; |
6117 | case LibFunc::sqrt: |
6118 | case LibFunc::sqrtf: |
6119 | case LibFunc::sqrtl: |
6120 | case LibFunc::sqrt_finite: |
6121 | case LibFunc::sqrtf_finite: |
6122 | case LibFunc::sqrtl_finite: |
6123 | if (visitUnaryFloatCall(I, ISD::FSQRT)) |
6124 | return; |
6125 | break; |
6126 | case LibFunc::floor: |
6127 | case LibFunc::floorf: |
6128 | case LibFunc::floorl: |
6129 | if (visitUnaryFloatCall(I, ISD::FFLOOR)) |
6130 | return; |
6131 | break; |
6132 | case LibFunc::nearbyint: |
6133 | case LibFunc::nearbyintf: |
6134 | case LibFunc::nearbyintl: |
6135 | if (visitUnaryFloatCall(I, ISD::FNEARBYINT)) |
6136 | return; |
6137 | break; |
6138 | case LibFunc::ceil: |
6139 | case LibFunc::ceilf: |
6140 | case LibFunc::ceill: |
6141 | if (visitUnaryFloatCall(I, ISD::FCEIL)) |
6142 | return; |
6143 | break; |
6144 | case LibFunc::rint: |
6145 | case LibFunc::rintf: |
6146 | case LibFunc::rintl: |
6147 | if (visitUnaryFloatCall(I, ISD::FRINT)) |
6148 | return; |
6149 | break; |
6150 | case LibFunc::round: |
6151 | case LibFunc::roundf: |
6152 | case LibFunc::roundl: |
6153 | if (visitUnaryFloatCall(I, ISD::FROUND)) |
6154 | return; |
6155 | break; |
6156 | case LibFunc::trunc: |
6157 | case LibFunc::truncf: |
6158 | case LibFunc::truncl: |
6159 | if (visitUnaryFloatCall(I, ISD::FTRUNC)) |
6160 | return; |
6161 | break; |
6162 | case LibFunc::log2: |
6163 | case LibFunc::log2f: |
6164 | case LibFunc::log2l: |
6165 | if (visitUnaryFloatCall(I, ISD::FLOG2)) |
6166 | return; |
6167 | break; |
6168 | case LibFunc::exp2: |
6169 | case LibFunc::exp2f: |
6170 | case LibFunc::exp2l: |
6171 | if (visitUnaryFloatCall(I, ISD::FEXP2)) |
6172 | return; |
6173 | break; |
6174 | case LibFunc::memcmp: |
6175 | if (visitMemCmpCall(I)) |
6176 | return; |
6177 | break; |
6178 | case LibFunc::memchr: |
6179 | if (visitMemChrCall(I)) |
6180 | return; |
6181 | break; |
6182 | case LibFunc::strcpy: |
6183 | if (visitStrCpyCall(I, false)) |
6184 | return; |
6185 | break; |
6186 | case LibFunc::stpcpy: |
6187 | if (visitStrCpyCall(I, true)) |
6188 | return; |
6189 | break; |
6190 | case LibFunc::strcmp: |
6191 | if (visitStrCmpCall(I)) |
6192 | return; |
6193 | break; |
6194 | case LibFunc::strlen: |
6195 | if (visitStrLenCall(I)) |
6196 | return; |
6197 | break; |
6198 | case LibFunc::strnlen: |
6199 | if (visitStrNLenCall(I)) |
6200 | return; |
6201 | break; |
6202 | } |
6203 | } |
6204 | } |
6205 | |
6206 | SDValue Callee; |
6207 | if (!RenameFn) |
6208 | Callee = getValue(I.getCalledValue()); |
6209 | else |
6210 | Callee = DAG.getExternalSymbol(RenameFn, |
6211 | DAG.getTargetLoweringInfo().getPointerTy()); |
6212 | |
6213 | // Check if we can potentially perform a tail call. More detailed checking is |
6214 | // be done within LowerCallTo, after more information about the call is known. |
6215 | LowerCallTo(&I, Callee, I.isTailCall()); |
6216 | } |
6217 | |
6218 | namespace { |
6219 | |
6220 | /// AsmOperandInfo - This contains information for each constraint that we are |
6221 | /// lowering. |
6222 | class SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo { |
6223 | public: |
6224 | /// CallOperand - If this is the result output operand or a clobber |
6225 | /// this is null, otherwise it is the incoming operand to the CallInst. |
6226 | /// This gets modified as the asm is processed. |
6227 | SDValue CallOperand; |
6228 | |
6229 | /// AssignedRegs - If this is a register or register class operand, this |
6230 | /// contains the set of register corresponding to the operand. |
6231 | RegsForValue AssignedRegs; |
6232 | |
6233 | explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info) |
6234 | : TargetLowering::AsmOperandInfo(info), CallOperand(nullptr,0) { |
6235 | } |
6236 | |
6237 | /// getCallOperandValEVT - Return the EVT of the Value* that this operand |
6238 | /// corresponds to. If there is no Value* for this operand, it returns |
6239 | /// MVT::Other. |
6240 | EVT getCallOperandValEVT(LLVMContext &Context, |
6241 | const TargetLowering &TLI, |
6242 | const DataLayout *DL) const { |
6243 | if (!CallOperandVal) return MVT::Other; |
6244 | |
6245 | if (isa<BasicBlock>(CallOperandVal)) |
6246 | return TLI.getPointerTy(); |
6247 | |
6248 | llvm::Type *OpTy = CallOperandVal->getType(); |
6249 | |
6250 | // FIXME: code duplicated from TargetLowering::ParseConstraints(). |
6251 | // If this is an indirect operand, the operand is a pointer to the |
6252 | // accessed type. |
6253 | if (isIndirect) { |
6254 | llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy); |
6255 | if (!PtrTy) |
6256 | report_fatal_error("Indirect operand for inline asm not a pointer!"); |
6257 | OpTy = PtrTy->getElementType(); |
6258 | } |
6259 | |
6260 | // Look for vector wrapped in a struct. e.g. { <16 x i8> }. |
6261 | if (StructType *STy = dyn_cast<StructType>(OpTy)) |
6262 | if (STy->getNumElements() == 1) |
6263 | OpTy = STy->getElementType(0); |
6264 | |
6265 | // If OpTy is not a single value, it may be a struct/union that we |
6266 | // can tile with integers. |
6267 | if (!OpTy->isSingleValueType() && OpTy->isSized()) { |
6268 | unsigned BitSize = DL->getTypeSizeInBits(OpTy); |
6269 | switch (BitSize) { |
6270 | default: break; |
6271 | case 1: |
6272 | case 8: |
6273 | case 16: |
6274 | case 32: |
6275 | case 64: |
6276 | case 128: |
6277 | OpTy = IntegerType::get(Context, BitSize); |
6278 | break; |
6279 | } |
6280 | } |
6281 | |
6282 | return TLI.getValueType(OpTy, true); |
6283 | } |
6284 | }; |
6285 | |
6286 | typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector; |
6287 | |
6288 | } // end anonymous namespace |
6289 | |
6290 | /// GetRegistersForValue - Assign registers (virtual or physical) for the |
6291 | /// specified operand. We prefer to assign virtual registers, to allow the |
6292 | /// register allocator to handle the assignment process. However, if the asm |
6293 | /// uses features that we can't model on machineinstrs, we have SDISel do the |
6294 | /// allocation. This produces generally horrible, but correct, code. |
6295 | /// |
6296 | /// OpInfo describes the operand. |
6297 | /// |
6298 | static void GetRegistersForValue(SelectionDAG &DAG, |
6299 | const TargetLowering &TLI, |
6300 | SDLoc DL, |
6301 | SDISelAsmOperandInfo &OpInfo) { |
6302 | LLVMContext &Context = *DAG.getContext(); |
6303 | |
6304 | MachineFunction &MF = DAG.getMachineFunction(); |
6305 | SmallVector<unsigned, 4> Regs; |
6306 | |
6307 | // If this is a constraint for a single physreg, or a constraint for a |
6308 | // register class, find it. |
6309 | std::pair<unsigned, const TargetRegisterClass*> PhysReg = |
6310 | TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode, |
6311 | OpInfo.ConstraintVT); |
6312 | |
6313 | unsigned NumRegs = 1; |
6314 | if (OpInfo.ConstraintVT != MVT::Other) { |
6315 | // If this is a FP input in an integer register (or visa versa) insert a bit |
6316 | // cast of the input value. More generally, handle any case where the input |
6317 | // value disagrees with the register class we plan to stick this in. |
6318 | if (OpInfo.Type == InlineAsm::isInput && |
6319 | PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) { |
6320 | // Try to convert to the first EVT that the reg class contains. If the |
6321 | // types are identical size, use a bitcast to convert (e.g. two differing |
6322 | // vector types). |
6323 | MVT RegVT = *PhysReg.second->vt_begin(); |
6324 | if (RegVT.getSizeInBits() == OpInfo.CallOperand.getValueSizeInBits()) { |
6325 | OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL, |
6326 | RegVT, OpInfo.CallOperand); |
6327 | OpInfo.ConstraintVT = RegVT; |
6328 | } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) { |
6329 | // If the input is a FP value and we want it in FP registers, do a |
6330 | // bitcast to the corresponding integer type. This turns an f64 value |
6331 | // into i64, which can be passed with two i32 values on a 32-bit |
6332 | // machine. |
6333 | RegVT = MVT::getIntegerVT(OpInfo.ConstraintVT.getSizeInBits()); |
6334 | OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL, |
6335 | RegVT, OpInfo.CallOperand); |
6336 | OpInfo.ConstraintVT = RegVT; |
6337 | } |
6338 | } |
6339 | |
6340 | NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT); |
6341 | } |
6342 | |
6343 | MVT RegVT; |
6344 | EVT ValueVT = OpInfo.ConstraintVT; |
6345 | |
6346 | // If this is a constraint for a specific physical register, like {r17}, |
6347 | // assign it now. |
6348 | if (unsigned AssignedReg = PhysReg.first) { |
6349 | const TargetRegisterClass *RC = PhysReg.second; |
6350 | if (OpInfo.ConstraintVT == MVT::Other) |
6351 | ValueVT = *RC->vt_begin(); |
6352 | |
6353 | // Get the actual register value type. This is important, because the user |
6354 | // may have asked for (e.g.) the AX register in i32 type. We need to |
6355 | // remember that AX is actually i16 to get the right extension. |
6356 | RegVT = *RC->vt_begin(); |
6357 | |
6358 | // This is a explicit reference to a physical register. |
6359 | Regs.push_back(AssignedReg); |
6360 | |
6361 | // If this is an expanded reference, add the rest of the regs to Regs. |
6362 | if (NumRegs != 1) { |
6363 | TargetRegisterClass::iterator I = RC->begin(); |
6364 | for (; *I != AssignedReg; ++I) |
6365 | assert(I != RC->end() && "Didn't find reg!")((I != RC->end() && "Didn't find reg!") ? static_cast <void> (0) : __assert_fail ("I != RC->end() && \"Didn't find reg!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6365, __PRETTY_FUNCTION__)); |
6366 | |
6367 | // Already added the first reg. |
6368 | --NumRegs; ++I; |
6369 | for (; NumRegs; --NumRegs, ++I) { |
6370 | assert(I != RC->end() && "Ran out of registers to allocate!")((I != RC->end() && "Ran out of registers to allocate!" ) ? static_cast<void> (0) : __assert_fail ("I != RC->end() && \"Ran out of registers to allocate!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6370, __PRETTY_FUNCTION__)); |
6371 | Regs.push_back(*I); |
6372 | } |
6373 | } |
6374 | |
6375 | OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT); |
6376 | return; |
6377 | } |
6378 | |
6379 | // Otherwise, if this was a reference to an LLVM register class, create vregs |
6380 | // for this reference. |
6381 | if (const TargetRegisterClass *RC = PhysReg.second) { |
6382 | RegVT = *RC->vt_begin(); |
6383 | if (OpInfo.ConstraintVT == MVT::Other) |
6384 | ValueVT = RegVT; |
6385 | |
6386 | // Create the appropriate number of virtual registers. |
6387 | MachineRegisterInfo &RegInfo = MF.getRegInfo(); |
6388 | for (; NumRegs; --NumRegs) |
6389 | Regs.push_back(RegInfo.createVirtualRegister(RC)); |
6390 | |
6391 | OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT); |
6392 | return; |
6393 | } |
6394 | |
6395 | // Otherwise, we couldn't allocate enough registers for this. |
6396 | } |
6397 | |
6398 | /// visitInlineAsm - Handle a call to an InlineAsm object. |
6399 | /// |
6400 | void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) { |
6401 | const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); |
6402 | |
6403 | /// ConstraintOperands - Information about all of the constraints. |
6404 | SDISelAsmOperandInfoVector ConstraintOperands; |
6405 | |
6406 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
6407 | TargetLowering::AsmOperandInfoVector |
6408 | TargetConstraints = TLI.ParseConstraints(CS); |
6409 | |
6410 | bool hasMemory = false; |
6411 | |
6412 | unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. |
6413 | unsigned ResNo = 0; // ResNo - The result number of the next output. |
6414 | for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) { |
6415 | ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i])); |
6416 | SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back(); |
6417 | |
6418 | MVT OpVT = MVT::Other; |
6419 | |
6420 | // Compute the value type for each operand. |
6421 | switch (OpInfo.Type) { |
6422 | case InlineAsm::isOutput: |
6423 | // Indirect outputs just consume an argument. |
6424 | if (OpInfo.isIndirect) { |
6425 | OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); |
6426 | break; |
6427 | } |
6428 | |
6429 | // The return value of the call is this value. As such, there is no |
6430 | // corresponding argument. |
6431 | assert(!CS.getType()->isVoidTy() && "Bad inline asm!")((!CS.getType()->isVoidTy() && "Bad inline asm!") ? static_cast<void> (0) : __assert_fail ("!CS.getType()->isVoidTy() && \"Bad inline asm!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6431, __PRETTY_FUNCTION__)); |
6432 | if (StructType *STy = dyn_cast<StructType>(CS.getType())) { |
6433 | OpVT = TLI.getSimpleValueType(STy->getElementType(ResNo)); |
6434 | } else { |
6435 | assert(ResNo == 0 && "Asm only has one result!")((ResNo == 0 && "Asm only has one result!") ? static_cast <void> (0) : __assert_fail ("ResNo == 0 && \"Asm only has one result!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6435, __PRETTY_FUNCTION__)); |
6436 | OpVT = TLI.getSimpleValueType(CS.getType()); |
6437 | } |
6438 | ++ResNo; |
6439 | break; |
6440 | case InlineAsm::isInput: |
6441 | OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); |
6442 | break; |
6443 | case InlineAsm::isClobber: |
6444 | // Nothing to do. |
6445 | break; |
6446 | } |
6447 | |
6448 | // If this is an input or an indirect output, process the call argument. |
6449 | // BasicBlocks are labels, currently appearing only in asm's. |
6450 | if (OpInfo.CallOperandVal) { |
6451 | if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) { |
6452 | OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]); |
6453 | } else { |
6454 | OpInfo.CallOperand = getValue(OpInfo.CallOperandVal); |
6455 | } |
6456 | |
6457 | OpVT = |
6458 | OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, DL).getSimpleVT(); |
6459 | } |
6460 | |
6461 | OpInfo.ConstraintVT = OpVT; |
6462 | |
6463 | // Indirect operand accesses access memory. |
6464 | if (OpInfo.isIndirect) |
6465 | hasMemory = true; |
6466 | else { |
6467 | for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) { |
6468 | TargetLowering::ConstraintType |
6469 | CType = TLI.getConstraintType(OpInfo.Codes[j]); |
6470 | if (CType == TargetLowering::C_Memory) { |
6471 | hasMemory = true; |
6472 | break; |
6473 | } |
6474 | } |
6475 | } |
6476 | } |
6477 | |
6478 | SDValue Chain, Flag; |
6479 | |
6480 | // We won't need to flush pending loads if this asm doesn't touch |
6481 | // memory and is nonvolatile. |
6482 | if (hasMemory || IA->hasSideEffects()) |
6483 | Chain = getRoot(); |
6484 | else |
6485 | Chain = DAG.getRoot(); |
6486 | |
6487 | // Second pass over the constraints: compute which constraint option to use |
6488 | // and assign registers to constraints that want a specific physreg. |
6489 | for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) { |
6490 | SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i]; |
6491 | |
6492 | // If this is an output operand with a matching input operand, look up the |
6493 | // matching input. If their types mismatch, e.g. one is an integer, the |
6494 | // other is floating point, or their sizes are different, flag it as an |
6495 | // error. |
6496 | if (OpInfo.hasMatchingInput()) { |
6497 | SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; |
6498 | |
6499 | if (OpInfo.ConstraintVT != Input.ConstraintVT) { |
6500 | std::pair<unsigned, const TargetRegisterClass*> MatchRC = |
6501 | TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode, |
6502 | OpInfo.ConstraintVT); |
6503 | std::pair<unsigned, const TargetRegisterClass*> InputRC = |
6504 | TLI.getRegForInlineAsmConstraint(Input.ConstraintCode, |
6505 | Input.ConstraintVT); |
6506 | if ((OpInfo.ConstraintVT.isInteger() != |
6507 | Input.ConstraintVT.isInteger()) || |
6508 | (MatchRC.second != InputRC.second)) { |
6509 | report_fatal_error("Unsupported asm: input constraint" |
6510 | " with a matching output constraint of" |
6511 | " incompatible type!"); |
6512 | } |
6513 | Input.ConstraintVT = OpInfo.ConstraintVT; |
6514 | } |
6515 | } |
6516 | |
6517 | // Compute the constraint code and ConstraintType to use. |
6518 | TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG); |
6519 | |
6520 | if (OpInfo.ConstraintType == TargetLowering::C_Memory && |
6521 | OpInfo.Type == InlineAsm::isClobber) |
6522 | continue; |
6523 | |
6524 | // If this is a memory input, and if the operand is not indirect, do what we |
6525 | // need to to provide an address for the memory input. |
6526 | if (OpInfo.ConstraintType == TargetLowering::C_Memory && |
6527 | !OpInfo.isIndirect) { |
6528 | assert((OpInfo.isMultipleAlternative ||(((OpInfo.isMultipleAlternative || (OpInfo.Type == InlineAsm:: isInput)) && "Can only indirectify direct input operands!" ) ? static_cast<void> (0) : __assert_fail ("(OpInfo.isMultipleAlternative || (OpInfo.Type == InlineAsm::isInput)) && \"Can only indirectify direct input operands!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6530, __PRETTY_FUNCTION__)) |
6529 | (OpInfo.Type == InlineAsm::isInput)) &&(((OpInfo.isMultipleAlternative || (OpInfo.Type == InlineAsm:: isInput)) && "Can only indirectify direct input operands!" ) ? static_cast<void> (0) : __assert_fail ("(OpInfo.isMultipleAlternative || (OpInfo.Type == InlineAsm::isInput)) && \"Can only indirectify direct input operands!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6530, __PRETTY_FUNCTION__)) |
6530 | "Can only indirectify direct input operands!")(((OpInfo.isMultipleAlternative || (OpInfo.Type == InlineAsm:: isInput)) && "Can only indirectify direct input operands!" ) ? static_cast<void> (0) : __assert_fail ("(OpInfo.isMultipleAlternative || (OpInfo.Type == InlineAsm::isInput)) && \"Can only indirectify direct input operands!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6530, __PRETTY_FUNCTION__)); |
6531 | |
6532 | // Memory operands really want the address of the value. If we don't have |
6533 | // an indirect input, put it in the constpool if we can, otherwise spill |
6534 | // it to a stack slot. |
6535 | // TODO: This isn't quite right. We need to handle these according to |
6536 | // the addressing mode that the constraint wants. Also, this may take |
6537 | // an additional register for the computation and we don't want that |
6538 | // either. |
6539 | |
6540 | // If the operand is a float, integer, or vector constant, spill to a |
6541 | // constant pool entry to get its address. |
6542 | const Value *OpVal = OpInfo.CallOperandVal; |
6543 | if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) || |
6544 | isa<ConstantVector>(OpVal) || isa<ConstantDataVector>(OpVal)) { |
6545 | OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal), |
6546 | TLI.getPointerTy()); |
6547 | } else { |
6548 | // Otherwise, create a stack slot and emit a store to it before the |
6549 | // asm. |
6550 | Type *Ty = OpVal->getType(); |
6551 | uint64_t TySize = TLI.getDataLayout()->getTypeAllocSize(Ty); |
6552 | unsigned Align = TLI.getDataLayout()->getPrefTypeAlignment(Ty); |
6553 | MachineFunction &MF = DAG.getMachineFunction(); |
6554 | int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false); |
6555 | SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy()); |
6556 | Chain = DAG.getStore(Chain, getCurSDLoc(), |
6557 | OpInfo.CallOperand, StackSlot, |
6558 | MachinePointerInfo::getFixedStack(SSFI), |
6559 | false, false, 0); |
6560 | OpInfo.CallOperand = StackSlot; |
6561 | } |
6562 | |
6563 | // There is no longer a Value* corresponding to this operand. |
6564 | OpInfo.CallOperandVal = nullptr; |
6565 | |
6566 | // It is now an indirect operand. |
6567 | OpInfo.isIndirect = true; |
6568 | } |
6569 | |
6570 | // If this constraint is for a specific register, allocate it before |
6571 | // anything else. |
6572 | if (OpInfo.ConstraintType == TargetLowering::C_Register) |
6573 | GetRegistersForValue(DAG, TLI, getCurSDLoc(), OpInfo); |
6574 | } |
6575 | |
6576 | // Second pass - Loop over all of the operands, assigning virtual or physregs |
6577 | // to register class operands. |
6578 | for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) { |
6579 | SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i]; |
6580 | |
6581 | // C_Register operands have already been allocated, Other/Memory don't need |
6582 | // to be. |
6583 | if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass) |
6584 | GetRegistersForValue(DAG, TLI, getCurSDLoc(), OpInfo); |
6585 | } |
6586 | |
6587 | // AsmNodeOperands - The operands for the ISD::INLINEASM node. |
6588 | std::vector<SDValue> AsmNodeOperands; |
6589 | AsmNodeOperands.push_back(SDValue()); // reserve space for input chain |
6590 | AsmNodeOperands.push_back( |
6591 | DAG.getTargetExternalSymbol(IA->getAsmString().c_str(), |
6592 | TLI.getPointerTy())); |
6593 | |
6594 | // If we have a !srcloc metadata node associated with it, we want to attach |
6595 | // this to the ultimately generated inline asm machineinstr. To do this, we |
6596 | // pass in the third operand as this (potentially null) inline asm MDNode. |
6597 | const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc"); |
6598 | AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc)); |
6599 | |
6600 | // Remember the HasSideEffect, AlignStack, AsmDialect, MayLoad and MayStore |
6601 | // bits as operand 3. |
6602 | unsigned ExtraInfo = 0; |
6603 | if (IA->hasSideEffects()) |
6604 | ExtraInfo |= InlineAsm::Extra_HasSideEffects; |
6605 | if (IA->isAlignStack()) |
6606 | ExtraInfo |= InlineAsm::Extra_IsAlignStack; |
6607 | // Set the asm dialect. |
6608 | ExtraInfo |= IA->getDialect() * InlineAsm::Extra_AsmDialect; |
6609 | |
6610 | // Determine if this InlineAsm MayLoad or MayStore based on the constraints. |
6611 | for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) { |
6612 | TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i]; |
6613 | |
6614 | // Compute the constraint code and ConstraintType to use. |
6615 | TLI.ComputeConstraintToUse(OpInfo, SDValue()); |
6616 | |
6617 | // Ideally, we would only check against memory constraints. However, the |
6618 | // meaning of an other constraint can be target-specific and we can't easily |
6619 | // reason about it. Therefore, be conservative and set MayLoad/MayStore |
6620 | // for other constriants as well. |
6621 | if (OpInfo.ConstraintType == TargetLowering::C_Memory || |
6622 | OpInfo.ConstraintType == TargetLowering::C_Other) { |
6623 | if (OpInfo.Type == InlineAsm::isInput) |
6624 | ExtraInfo |= InlineAsm::Extra_MayLoad; |
6625 | else if (OpInfo.Type == InlineAsm::isOutput) |
6626 | ExtraInfo |= InlineAsm::Extra_MayStore; |
6627 | else if (OpInfo.Type == InlineAsm::isClobber) |
6628 | ExtraInfo |= (InlineAsm::Extra_MayLoad | InlineAsm::Extra_MayStore); |
6629 | } |
6630 | } |
6631 | |
6632 | AsmNodeOperands.push_back(DAG.getTargetConstant(ExtraInfo, |
6633 | TLI.getPointerTy())); |
6634 | |
6635 | // Loop over all of the inputs, copying the operand values into the |
6636 | // appropriate registers and processing the output regs. |
6637 | RegsForValue RetValRegs; |
6638 | |
6639 | // IndirectStoresToEmit - The set of stores to emit after the inline asm node. |
6640 | std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit; |
6641 | |
6642 | for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) { |
6643 | SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i]; |
6644 | |
6645 | switch (OpInfo.Type) { |
6646 | case InlineAsm::isOutput: { |
6647 | if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass && |
6648 | OpInfo.ConstraintType != TargetLowering::C_Register) { |
6649 | // Memory output, or 'other' output (e.g. 'X' constraint). |
6650 | assert(OpInfo.isIndirect && "Memory output must be indirect operand")((OpInfo.isIndirect && "Memory output must be indirect operand" ) ? static_cast<void> (0) : __assert_fail ("OpInfo.isIndirect && \"Memory output must be indirect operand\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6650, __PRETTY_FUNCTION__)); |
6651 | |
6652 | // Add information to the INLINEASM node to know about this output. |
6653 | unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1); |
6654 | AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags, |
6655 | TLI.getPointerTy())); |
6656 | AsmNodeOperands.push_back(OpInfo.CallOperand); |
6657 | break; |
6658 | } |
6659 | |
6660 | // Otherwise, this is a register or register class output. |
6661 | |
6662 | // Copy the output from the appropriate register. Find a register that |
6663 | // we can use. |
6664 | if (OpInfo.AssignedRegs.Regs.empty()) { |
6665 | LLVMContext &Ctx = *DAG.getContext(); |
6666 | Ctx.emitError(CS.getInstruction(), |
6667 | "couldn't allocate output register for constraint '" + |
6668 | Twine(OpInfo.ConstraintCode) + "'"); |
6669 | return; |
6670 | } |
6671 | |
6672 | // If this is an indirect operand, store through the pointer after the |
6673 | // asm. |
6674 | if (OpInfo.isIndirect) { |
6675 | IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs, |
6676 | OpInfo.CallOperandVal)); |
6677 | } else { |
6678 | // This is the result value of the call. |
6679 | assert(!CS.getType()->isVoidTy() && "Bad inline asm!")((!CS.getType()->isVoidTy() && "Bad inline asm!") ? static_cast<void> (0) : __assert_fail ("!CS.getType()->isVoidTy() && \"Bad inline asm!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6679, __PRETTY_FUNCTION__)); |
6680 | // Concatenate this output onto the outputs list. |
6681 | RetValRegs.append(OpInfo.AssignedRegs); |
6682 | } |
6683 | |
6684 | // Add information to the INLINEASM node to know that this register is |
6685 | // set. |
6686 | OpInfo.AssignedRegs |
6687 | .AddInlineAsmOperands(OpInfo.isEarlyClobber |
6688 | ? InlineAsm::Kind_RegDefEarlyClobber |
6689 | : InlineAsm::Kind_RegDef, |
6690 | false, 0, DAG, AsmNodeOperands); |
6691 | break; |
6692 | } |
6693 | case InlineAsm::isInput: { |
6694 | SDValue InOperandVal = OpInfo.CallOperand; |
6695 | |
6696 | if (OpInfo.isMatchingInputConstraint()) { // Matching constraint? |
6697 | // If this is required to match an output register we have already set, |
6698 | // just use its register. |
6699 | unsigned OperandNo = OpInfo.getMatchedOperand(); |
6700 | |
6701 | // Scan until we find the definition we already emitted of this operand. |
6702 | // When we find it, create a RegsForValue operand. |
6703 | unsigned CurOp = InlineAsm::Op_FirstOperand; |
6704 | for (; OperandNo; --OperandNo) { |
6705 | // Advance to the next operand. |
6706 | unsigned OpFlag = |
6707 | cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue(); |
6708 | assert((InlineAsm::isRegDefKind(OpFlag) ||(((InlineAsm::isRegDefKind(OpFlag) || InlineAsm::isRegDefEarlyClobberKind (OpFlag) || InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?" ) ? static_cast<void> (0) : __assert_fail ("(InlineAsm::isRegDefKind(OpFlag) || InlineAsm::isRegDefEarlyClobberKind(OpFlag) || InlineAsm::isMemKind(OpFlag)) && \"Skipped past definitions?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6710, __PRETTY_FUNCTION__)) |
6709 | InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||(((InlineAsm::isRegDefKind(OpFlag) || InlineAsm::isRegDefEarlyClobberKind (OpFlag) || InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?" ) ? static_cast<void> (0) : __assert_fail ("(InlineAsm::isRegDefKind(OpFlag) || InlineAsm::isRegDefEarlyClobberKind(OpFlag) || InlineAsm::isMemKind(OpFlag)) && \"Skipped past definitions?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6710, __PRETTY_FUNCTION__)) |
6710 | InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?")(((InlineAsm::isRegDefKind(OpFlag) || InlineAsm::isRegDefEarlyClobberKind (OpFlag) || InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?" ) ? static_cast<void> (0) : __assert_fail ("(InlineAsm::isRegDefKind(OpFlag) || InlineAsm::isRegDefEarlyClobberKind(OpFlag) || InlineAsm::isMemKind(OpFlag)) && \"Skipped past definitions?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6710, __PRETTY_FUNCTION__)); |
6711 | CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1; |
6712 | } |
6713 | |
6714 | unsigned OpFlag = |
6715 | cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue(); |
6716 | if (InlineAsm::isRegDefKind(OpFlag) || |
6717 | InlineAsm::isRegDefEarlyClobberKind(OpFlag)) { |
6718 | // Add (OpFlag&0xffff)>>3 registers to MatchedRegs. |
6719 | if (OpInfo.isIndirect) { |
6720 | // This happens on gcc/testsuite/gcc.dg/pr8788-1.c |
6721 | LLVMContext &Ctx = *DAG.getContext(); |
6722 | Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:" |
6723 | " don't know how to handle tied " |
6724 | "indirect register inputs"); |
6725 | return; |
6726 | } |
6727 | |
6728 | RegsForValue MatchedRegs; |
6729 | MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType()); |
6730 | MVT RegVT = AsmNodeOperands[CurOp+1].getSimpleValueType(); |
6731 | MatchedRegs.RegVTs.push_back(RegVT); |
6732 | MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo(); |
6733 | for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag); |
6734 | i != e; ++i) { |
6735 | if (const TargetRegisterClass *RC = TLI.getRegClassFor(RegVT)) |
6736 | MatchedRegs.Regs.push_back(RegInfo.createVirtualRegister(RC)); |
6737 | else { |
6738 | LLVMContext &Ctx = *DAG.getContext(); |
6739 | Ctx.emitError(CS.getInstruction(), |
6740 | "inline asm error: This value" |
6741 | " type register class is not natively supported!"); |
6742 | return; |
6743 | } |
6744 | } |
6745 | // Use the produced MatchedRegs object to |
6746 | MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurSDLoc(), |
6747 | Chain, &Flag, CS.getInstruction()); |
6748 | MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, |
6749 | true, OpInfo.getMatchedOperand(), |
6750 | DAG, AsmNodeOperands); |
6751 | break; |
6752 | } |
6753 | |
6754 | assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!")((InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!" ) ? static_cast<void> (0) : __assert_fail ("InlineAsm::isMemKind(OpFlag) && \"Unknown matching constraint!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6754, __PRETTY_FUNCTION__)); |
6755 | assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&((InlineAsm::getNumOperandRegisters(OpFlag) == 1 && "Unexpected number of operands" ) ? static_cast<void> (0) : __assert_fail ("InlineAsm::getNumOperandRegisters(OpFlag) == 1 && \"Unexpected number of operands\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6756, __PRETTY_FUNCTION__)) |
6756 | "Unexpected number of operands")((InlineAsm::getNumOperandRegisters(OpFlag) == 1 && "Unexpected number of operands" ) ? static_cast<void> (0) : __assert_fail ("InlineAsm::getNumOperandRegisters(OpFlag) == 1 && \"Unexpected number of operands\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6756, __PRETTY_FUNCTION__)); |
6757 | // Add information to the INLINEASM node to know about this input. |
6758 | // See InlineAsm.h isUseOperandTiedToDef. |
6759 | OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag, |
6760 | OpInfo.getMatchedOperand()); |
6761 | AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag, |
6762 | TLI.getPointerTy())); |
6763 | AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]); |
6764 | break; |
6765 | } |
6766 | |
6767 | // Treat indirect 'X' constraint as memory. |
6768 | if (OpInfo.ConstraintType == TargetLowering::C_Other && |
6769 | OpInfo.isIndirect) |
6770 | OpInfo.ConstraintType = TargetLowering::C_Memory; |
6771 | |
6772 | if (OpInfo.ConstraintType == TargetLowering::C_Other) { |
6773 | std::vector<SDValue> Ops; |
6774 | TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode, |
6775 | Ops, DAG); |
6776 | if (Ops.empty()) { |
6777 | LLVMContext &Ctx = *DAG.getContext(); |
6778 | Ctx.emitError(CS.getInstruction(), |
6779 | "invalid operand for inline asm constraint '" + |
6780 | Twine(OpInfo.ConstraintCode) + "'"); |
6781 | return; |
6782 | } |
6783 | |
6784 | // Add information to the INLINEASM node to know about this input. |
6785 | unsigned ResOpType = |
6786 | InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size()); |
6787 | AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType, |
6788 | TLI.getPointerTy())); |
6789 | AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end()); |
6790 | break; |
6791 | } |
6792 | |
6793 | if (OpInfo.ConstraintType == TargetLowering::C_Memory) { |
6794 | assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!")((OpInfo.isIndirect && "Operand must be indirect to be a mem!" ) ? static_cast<void> (0) : __assert_fail ("OpInfo.isIndirect && \"Operand must be indirect to be a mem!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6794, __PRETTY_FUNCTION__)); |
6795 | assert(InOperandVal.getValueType() == TLI.getPointerTy() &&((InOperandVal.getValueType() == TLI.getPointerTy() && "Memory operands expect pointer values") ? static_cast<void > (0) : __assert_fail ("InOperandVal.getValueType() == TLI.getPointerTy() && \"Memory operands expect pointer values\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6796, __PRETTY_FUNCTION__)) |
6796 | "Memory operands expect pointer values")((InOperandVal.getValueType() == TLI.getPointerTy() && "Memory operands expect pointer values") ? static_cast<void > (0) : __assert_fail ("InOperandVal.getValueType() == TLI.getPointerTy() && \"Memory operands expect pointer values\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6796, __PRETTY_FUNCTION__)); |
6797 | |
6798 | // Add information to the INLINEASM node to know about this input. |
6799 | unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1); |
6800 | AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType, |
6801 | TLI.getPointerTy())); |
6802 | AsmNodeOperands.push_back(InOperandVal); |
6803 | break; |
6804 | } |
6805 | |
6806 | assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||(((OpInfo.ConstraintType == TargetLowering::C_RegisterClass || OpInfo.ConstraintType == TargetLowering::C_Register) && "Unknown constraint type!") ? static_cast<void> (0) : __assert_fail ("(OpInfo.ConstraintType == TargetLowering::C_RegisterClass || OpInfo.ConstraintType == TargetLowering::C_Register) && \"Unknown constraint type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6808, __PRETTY_FUNCTION__)) |
6807 | OpInfo.ConstraintType == TargetLowering::C_Register) &&(((OpInfo.ConstraintType == TargetLowering::C_RegisterClass || OpInfo.ConstraintType == TargetLowering::C_Register) && "Unknown constraint type!") ? static_cast<void> (0) : __assert_fail ("(OpInfo.ConstraintType == TargetLowering::C_RegisterClass || OpInfo.ConstraintType == TargetLowering::C_Register) && \"Unknown constraint type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6808, __PRETTY_FUNCTION__)) |
6808 | "Unknown constraint type!")(((OpInfo.ConstraintType == TargetLowering::C_RegisterClass || OpInfo.ConstraintType == TargetLowering::C_Register) && "Unknown constraint type!") ? static_cast<void> (0) : __assert_fail ("(OpInfo.ConstraintType == TargetLowering::C_RegisterClass || OpInfo.ConstraintType == TargetLowering::C_Register) && \"Unknown constraint type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6808, __PRETTY_FUNCTION__)); |
6809 | |
6810 | // TODO: Support this. |
6811 | if (OpInfo.isIndirect) { |
6812 | LLVMContext &Ctx = *DAG.getContext(); |
6813 | Ctx.emitError(CS.getInstruction(), |
6814 | "Don't know how to handle indirect register inputs yet " |
6815 | "for constraint '" + |
6816 | Twine(OpInfo.ConstraintCode) + "'"); |
6817 | return; |
6818 | } |
6819 | |
6820 | // Copy the input into the appropriate registers. |
6821 | if (OpInfo.AssignedRegs.Regs.empty()) { |
6822 | LLVMContext &Ctx = *DAG.getContext(); |
6823 | Ctx.emitError(CS.getInstruction(), |
6824 | "couldn't allocate input reg for constraint '" + |
6825 | Twine(OpInfo.ConstraintCode) + "'"); |
6826 | return; |
6827 | } |
6828 | |
6829 | OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurSDLoc(), |
6830 | Chain, &Flag, CS.getInstruction()); |
6831 | |
6832 | OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0, |
6833 | DAG, AsmNodeOperands); |
6834 | break; |
6835 | } |
6836 | case InlineAsm::isClobber: { |
6837 | // Add the clobbered value to the operand list, so that the register |
6838 | // allocator is aware that the physreg got clobbered. |
6839 | if (!OpInfo.AssignedRegs.Regs.empty()) |
6840 | OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_Clobber, |
6841 | false, 0, DAG, |
6842 | AsmNodeOperands); |
6843 | break; |
6844 | } |
6845 | } |
6846 | } |
6847 | |
6848 | // Finish up input operands. Set the input chain and add the flag last. |
6849 | AsmNodeOperands[InlineAsm::Op_InputChain] = Chain; |
6850 | if (Flag.getNode()) AsmNodeOperands.push_back(Flag); |
6851 | |
6852 | Chain = DAG.getNode(ISD::INLINEASM, getCurSDLoc(), |
6853 | DAG.getVTList(MVT::Other, MVT::Glue), AsmNodeOperands); |
6854 | Flag = Chain.getValue(1); |
6855 | |
6856 | // If this asm returns a register value, copy the result from that register |
6857 | // and set it as the value of the call. |
6858 | if (!RetValRegs.Regs.empty()) { |
6859 | SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), |
6860 | Chain, &Flag, CS.getInstruction()); |
6861 | |
6862 | // FIXME: Why don't we do this for inline asms with MRVs? |
6863 | if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) { |
6864 | EVT ResultType = TLI.getValueType(CS.getType()); |
6865 | |
6866 | // If any of the results of the inline asm is a vector, it may have the |
6867 | // wrong width/num elts. This can happen for register classes that can |
6868 | // contain multiple different value types. The preg or vreg allocated may |
6869 | // not have the same VT as was expected. Convert it to the right type |
6870 | // with bit_convert. |
6871 | if (ResultType != Val.getValueType() && Val.getValueType().isVector()) { |
6872 | Val = DAG.getNode(ISD::BITCAST, getCurSDLoc(), |
6873 | ResultType, Val); |
6874 | |
6875 | } else if (ResultType != Val.getValueType() && |
6876 | ResultType.isInteger() && Val.getValueType().isInteger()) { |
6877 | // If a result value was tied to an input value, the computed result may |
6878 | // have a wider width than the expected result. Extract the relevant |
6879 | // portion. |
6880 | Val = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), ResultType, Val); |
6881 | } |
6882 | |
6883 | assert(ResultType == Val.getValueType() && "Asm result value mismatch!")((ResultType == Val.getValueType() && "Asm result value mismatch!" ) ? static_cast<void> (0) : __assert_fail ("ResultType == Val.getValueType() && \"Asm result value mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6883, __PRETTY_FUNCTION__)); |
6884 | } |
6885 | |
6886 | setValue(CS.getInstruction(), Val); |
6887 | // Don't need to use this as a chain in this case. |
6888 | if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty()) |
6889 | return; |
6890 | } |
6891 | |
6892 | std::vector<std::pair<SDValue, const Value *> > StoresToEmit; |
6893 | |
6894 | // Process indirect outputs, first output all of the flagged copies out of |
6895 | // physregs. |
6896 | for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) { |
6897 | RegsForValue &OutRegs = IndirectStoresToEmit[i].first; |
6898 | const Value *Ptr = IndirectStoresToEmit[i].second; |
6899 | SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), |
6900 | Chain, &Flag, IA); |
6901 | StoresToEmit.push_back(std::make_pair(OutVal, Ptr)); |
6902 | } |
6903 | |
6904 | // Emit the non-flagged stores from the physregs. |
6905 | SmallVector<SDValue, 8> OutChains; |
6906 | for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) { |
6907 | SDValue Val = DAG.getStore(Chain, getCurSDLoc(), |
6908 | StoresToEmit[i].first, |
6909 | getValue(StoresToEmit[i].second), |
6910 | MachinePointerInfo(StoresToEmit[i].second), |
6911 | false, false, 0); |
6912 | OutChains.push_back(Val); |
6913 | } |
6914 | |
6915 | if (!OutChains.empty()) |
6916 | Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, OutChains); |
6917 | |
6918 | DAG.setRoot(Chain); |
6919 | } |
6920 | |
6921 | void SelectionDAGBuilder::visitVAStart(const CallInst &I) { |
6922 | DAG.setRoot(DAG.getNode(ISD::VASTART, getCurSDLoc(), |
6923 | MVT::Other, getRoot(), |
6924 | getValue(I.getArgOperand(0)), |
6925 | DAG.getSrcValue(I.getArgOperand(0)))); |
6926 | } |
6927 | |
6928 | void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) { |
6929 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
6930 | const DataLayout &DL = *TLI.getDataLayout(); |
6931 | SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurSDLoc(), |
6932 | getRoot(), getValue(I.getOperand(0)), |
6933 | DAG.getSrcValue(I.getOperand(0)), |
6934 | DL.getABITypeAlignment(I.getType())); |
6935 | setValue(&I, V); |
6936 | DAG.setRoot(V.getValue(1)); |
6937 | } |
6938 | |
6939 | void SelectionDAGBuilder::visitVAEnd(const CallInst &I) { |
6940 | DAG.setRoot(DAG.getNode(ISD::VAEND, getCurSDLoc(), |
6941 | MVT::Other, getRoot(), |
6942 | getValue(I.getArgOperand(0)), |
6943 | DAG.getSrcValue(I.getArgOperand(0)))); |
6944 | } |
6945 | |
6946 | void SelectionDAGBuilder::visitVACopy(const CallInst &I) { |
6947 | DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurSDLoc(), |
6948 | MVT::Other, getRoot(), |
6949 | getValue(I.getArgOperand(0)), |
6950 | getValue(I.getArgOperand(1)), |
6951 | DAG.getSrcValue(I.getArgOperand(0)), |
6952 | DAG.getSrcValue(I.getArgOperand(1)))); |
6953 | } |
6954 | |
6955 | /// \brief Lower an argument list according to the target calling convention. |
6956 | /// |
6957 | /// \return A tuple of <return-value, token-chain> |
6958 | /// |
6959 | /// This is a helper for lowering intrinsics that follow a target calling |
6960 | /// convention or require stack pointer adjustment. Only a subset of the |
6961 | /// intrinsic's operands need to participate in the calling convention. |
6962 | std::pair<SDValue, SDValue> |
6963 | SelectionDAGBuilder::lowerCallOperands(ImmutableCallSite CS, unsigned ArgIdx, |
6964 | unsigned NumArgs, SDValue Callee, |
6965 | bool UseVoidTy, |
6966 | MachineBasicBlock *LandingPad) { |
6967 | TargetLowering::ArgListTy Args; |
6968 | Args.reserve(NumArgs); |
6969 | |
6970 | // Populate the argument list. |
6971 | // Attributes for args start at offset 1, after the return attribute. |
6972 | for (unsigned ArgI = ArgIdx, ArgE = ArgIdx + NumArgs, AttrI = ArgIdx + 1; |
6973 | ArgI != ArgE; ++ArgI) { |
6974 | const Value *V = CS->getOperand(ArgI); |
6975 | |
6976 | assert(!V->getType()->isEmptyTy() && "Empty type passed to intrinsic.")((!V->getType()->isEmptyTy() && "Empty type passed to intrinsic." ) ? static_cast<void> (0) : __assert_fail ("!V->getType()->isEmptyTy() && \"Empty type passed to intrinsic.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6976, __PRETTY_FUNCTION__)); |
6977 | |
6978 | TargetLowering::ArgListEntry Entry; |
6979 | Entry.Node = getValue(V); |
6980 | Entry.Ty = V->getType(); |
6981 | Entry.setAttributes(&CS, AttrI); |
6982 | Args.push_back(Entry); |
6983 | } |
6984 | |
6985 | Type *retTy = UseVoidTy ? Type::getVoidTy(*DAG.getContext()) : CS->getType(); |
6986 | TargetLowering::CallLoweringInfo CLI(DAG); |
6987 | CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot()) |
6988 | .setCallee(CS.getCallingConv(), retTy, Callee, std::move(Args), NumArgs) |
6989 | .setDiscardResult(CS->use_empty()); |
6990 | |
6991 | return lowerInvokable(CLI, LandingPad); |
6992 | } |
6993 | |
6994 | /// \brief Add a stack map intrinsic call's live variable operands to a stackmap |
6995 | /// or patchpoint target node's operand list. |
6996 | /// |
6997 | /// Constants are converted to TargetConstants purely as an optimization to |
6998 | /// avoid constant materialization and register allocation. |
6999 | /// |
7000 | /// FrameIndex operands are converted to TargetFrameIndex so that ISEL does not |
7001 | /// generate addess computation nodes, and so ExpandISelPseudo can convert the |
7002 | /// TargetFrameIndex into a DirectMemRefOp StackMap location. This avoids |
7003 | /// address materialization and register allocation, but may also be required |
7004 | /// for correctness. If a StackMap (or PatchPoint) intrinsic directly uses an |
7005 | /// alloca in the entry block, then the runtime may assume that the alloca's |
7006 | /// StackMap location can be read immediately after compilation and that the |
7007 | /// location is valid at any point during execution (this is similar to the |
7008 | /// assumption made by the llvm.gcroot intrinsic). If the alloca's location were |
7009 | /// only available in a register, then the runtime would need to trap when |
7010 | /// execution reaches the StackMap in order to read the alloca's location. |
7011 | static void addStackMapLiveVars(ImmutableCallSite CS, unsigned StartIdx, |
7012 | SmallVectorImpl<SDValue> &Ops, |
7013 | SelectionDAGBuilder &Builder) { |
7014 | for (unsigned i = StartIdx, e = CS.arg_size(); i != e; ++i) { |
7015 | SDValue OpVal = Builder.getValue(CS.getArgument(i)); |
7016 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(OpVal)) { |
7017 | Ops.push_back( |
7018 | Builder.DAG.getTargetConstant(StackMaps::ConstantOp, MVT::i64)); |
7019 | Ops.push_back( |
7020 | Builder.DAG.getTargetConstant(C->getSExtValue(), MVT::i64)); |
7021 | } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(OpVal)) { |
7022 | const TargetLowering &TLI = Builder.DAG.getTargetLoweringInfo(); |
7023 | Ops.push_back( |
7024 | Builder.DAG.getTargetFrameIndex(FI->getIndex(), TLI.getPointerTy())); |
7025 | } else |
7026 | Ops.push_back(OpVal); |
7027 | } |
7028 | } |
7029 | |
7030 | /// \brief Lower llvm.experimental.stackmap directly to its target opcode. |
7031 | void SelectionDAGBuilder::visitStackmap(const CallInst &CI) { |
7032 | // void @llvm.experimental.stackmap(i32 <id>, i32 <numShadowBytes>, |
7033 | // [live variables...]) |
7034 | |
7035 | assert(CI.getType()->isVoidTy() && "Stackmap cannot return a value.")((CI.getType()->isVoidTy() && "Stackmap cannot return a value." ) ? static_cast<void> (0) : __assert_fail ("CI.getType()->isVoidTy() && \"Stackmap cannot return a value.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7035, __PRETTY_FUNCTION__)); |
7036 | |
7037 | SDValue Chain, InFlag, Callee, NullPtr; |
7038 | SmallVector<SDValue, 32> Ops; |
7039 | |
7040 | SDLoc DL = getCurSDLoc(); |
7041 | Callee = getValue(CI.getCalledValue()); |
7042 | NullPtr = DAG.getIntPtrConstant(0, true); |
7043 | |
7044 | // The stackmap intrinsic only records the live variables (the arguemnts |
7045 | // passed to it) and emits NOPS (if requested). Unlike the patchpoint |
7046 | // intrinsic, this won't be lowered to a function call. This means we don't |
7047 | // have to worry about calling conventions and target specific lowering code. |
7048 | // Instead we perform the call lowering right here. |
7049 | // |
7050 | // chain, flag = CALLSEQ_START(chain, 0) |
7051 | // chain, flag = STACKMAP(id, nbytes, ..., chain, flag) |
7052 | // chain, flag = CALLSEQ_END(chain, 0, 0, flag) |
7053 | // |
7054 | Chain = DAG.getCALLSEQ_START(getRoot(), NullPtr, DL); |
7055 | InFlag = Chain.getValue(1); |
7056 | |
7057 | // Add the <id> and <numBytes> constants. |
7058 | SDValue IDVal = getValue(CI.getOperand(PatchPointOpers::IDPos)); |
7059 | Ops.push_back(DAG.getTargetConstant( |
7060 | cast<ConstantSDNode>(IDVal)->getZExtValue(), MVT::i64)); |
7061 | SDValue NBytesVal = getValue(CI.getOperand(PatchPointOpers::NBytesPos)); |
7062 | Ops.push_back(DAG.getTargetConstant( |
7063 | cast<ConstantSDNode>(NBytesVal)->getZExtValue(), MVT::i32)); |
7064 | |
7065 | // Push live variables for the stack map. |
7066 | addStackMapLiveVars(&CI, 2, Ops, *this); |
7067 | |
7068 | // We are not pushing any register mask info here on the operands list, |
7069 | // because the stackmap doesn't clobber anything. |
7070 | |
7071 | // Push the chain and the glue flag. |
7072 | Ops.push_back(Chain); |
7073 | Ops.push_back(InFlag); |
7074 | |
7075 | // Create the STACKMAP node. |
7076 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); |
7077 | SDNode *SM = DAG.getMachineNode(TargetOpcode::STACKMAP, DL, NodeTys, Ops); |
7078 | Chain = SDValue(SM, 0); |
7079 | InFlag = Chain.getValue(1); |
7080 | |
7081 | Chain = DAG.getCALLSEQ_END(Chain, NullPtr, NullPtr, InFlag, DL); |
7082 | |
7083 | // Stackmaps don't generate values, so nothing goes into the NodeMap. |
7084 | |
7085 | // Set the root to the target-lowered call chain. |
7086 | DAG.setRoot(Chain); |
7087 | |
7088 | // Inform the Frame Information that we have a stackmap in this function. |
7089 | FuncInfo.MF->getFrameInfo()->setHasStackMap(); |
7090 | } |
7091 | |
7092 | /// \brief Lower llvm.experimental.patchpoint directly to its target opcode. |
7093 | void SelectionDAGBuilder::visitPatchpoint(ImmutableCallSite CS, |
7094 | MachineBasicBlock *LandingPad) { |
7095 | // void|i64 @llvm.experimental.patchpoint.void|i64(i64 <id>, |
7096 | // i32 <numBytes>, |
7097 | // i8* <target>, |
7098 | // i32 <numArgs>, |
7099 | // [Args...], |
7100 | // [live variables...]) |
7101 | |
7102 | CallingConv::ID CC = CS.getCallingConv(); |
7103 | bool IsAnyRegCC = CC == CallingConv::AnyReg; |
7104 | bool HasDef = !CS->getType()->isVoidTy(); |
7105 | SDValue Callee = getValue(CS->getOperand(2)); // <target> |
7106 | |
7107 | // Get the real number of arguments participating in the call <numArgs> |
7108 | SDValue NArgVal = getValue(CS.getArgument(PatchPointOpers::NArgPos)); |
7109 | unsigned NumArgs = cast<ConstantSDNode>(NArgVal)->getZExtValue(); |
7110 | |
7111 | // Skip the four meta args: <id>, <numNopBytes>, <target>, <numArgs> |
7112 | // Intrinsics include all meta-operands up to but not including CC. |
7113 | unsigned NumMetaOpers = PatchPointOpers::CCPos; |
7114 | assert(CS.arg_size() >= NumMetaOpers + NumArgs &&((CS.arg_size() >= NumMetaOpers + NumArgs && "Not enough arguments provided to the patchpoint intrinsic" ) ? static_cast<void> (0) : __assert_fail ("CS.arg_size() >= NumMetaOpers + NumArgs && \"Not enough arguments provided to the patchpoint intrinsic\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7115, __PRETTY_FUNCTION__)) |
7115 | "Not enough arguments provided to the patchpoint intrinsic")((CS.arg_size() >= NumMetaOpers + NumArgs && "Not enough arguments provided to the patchpoint intrinsic" ) ? static_cast<void> (0) : __assert_fail ("CS.arg_size() >= NumMetaOpers + NumArgs && \"Not enough arguments provided to the patchpoint intrinsic\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7115, __PRETTY_FUNCTION__)); |
7116 | |
7117 | // For AnyRegCC the arguments are lowered later on manually. |
7118 | unsigned NumCallArgs = IsAnyRegCC ? 0 : NumArgs; |
7119 | std::pair<SDValue, SDValue> Result = |
7120 | lowerCallOperands(CS, NumMetaOpers, NumCallArgs, Callee, IsAnyRegCC, |
7121 | LandingPad); |
7122 | |
7123 | SDNode *CallEnd = Result.second.getNode(); |
7124 | if (HasDef && (CallEnd->getOpcode() == ISD::CopyFromReg)) |
7125 | CallEnd = CallEnd->getOperand(0).getNode(); |
7126 | |
7127 | /// Get a call instruction from the call sequence chain. |
7128 | /// Tail calls are not allowed. |
7129 | assert(CallEnd->getOpcode() == ISD::CALLSEQ_END &&((CallEnd->getOpcode() == ISD::CALLSEQ_END && "Expected a callseq node." ) ? static_cast<void> (0) : __assert_fail ("CallEnd->getOpcode() == ISD::CALLSEQ_END && \"Expected a callseq node.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7130, __PRETTY_FUNCTION__)) |
7130 | "Expected a callseq node.")((CallEnd->getOpcode() == ISD::CALLSEQ_END && "Expected a callseq node." ) ? static_cast<void> (0) : __assert_fail ("CallEnd->getOpcode() == ISD::CALLSEQ_END && \"Expected a callseq node.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7130, __PRETTY_FUNCTION__)); |
7131 | SDNode *Call = CallEnd->getOperand(0).getNode(); |
7132 | bool HasGlue = Call->getGluedNode(); |
7133 | |
7134 | // Replace the target specific call node with the patchable intrinsic. |
7135 | SmallVector<SDValue, 8> Ops; |
7136 | |
7137 | // Add the <id> and <numBytes> constants. |
7138 | SDValue IDVal = getValue(CS->getOperand(PatchPointOpers::IDPos)); |
7139 | Ops.push_back(DAG.getTargetConstant( |
7140 | cast<ConstantSDNode>(IDVal)->getZExtValue(), MVT::i64)); |
7141 | SDValue NBytesVal = getValue(CS->getOperand(PatchPointOpers::NBytesPos)); |
7142 | Ops.push_back(DAG.getTargetConstant( |
7143 | cast<ConstantSDNode>(NBytesVal)->getZExtValue(), MVT::i32)); |
7144 | |
7145 | // Assume that the Callee is a constant address. |
7146 | // FIXME: handle function symbols in the future. |
7147 | Ops.push_back( |
7148 | DAG.getIntPtrConstant(cast<ConstantSDNode>(Callee)->getZExtValue(), |
7149 | /*isTarget=*/true)); |
7150 | |
7151 | // Adjust <numArgs> to account for any arguments that have been passed on the |
7152 | // stack instead. |
7153 | // Call Node: Chain, Target, {Args}, RegMask, [Glue] |
7154 | unsigned NumCallRegArgs = Call->getNumOperands() - (HasGlue ? 4 : 3); |
7155 | NumCallRegArgs = IsAnyRegCC ? NumArgs : NumCallRegArgs; |
7156 | Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, MVT::i32)); |
7157 | |
7158 | // Add the calling convention |
7159 | Ops.push_back(DAG.getTargetConstant((unsigned)CC, MVT::i32)); |
7160 | |
7161 | // Add the arguments we omitted previously. The register allocator should |
7162 | // place these in any free register. |
7163 | if (IsAnyRegCC) |
7164 | for (unsigned i = NumMetaOpers, e = NumMetaOpers + NumArgs; i != e; ++i) |
7165 | Ops.push_back(getValue(CS.getArgument(i))); |
7166 | |
7167 | // Push the arguments from the call instruction up to the register mask. |
7168 | SDNode::op_iterator e = HasGlue ? Call->op_end()-2 : Call->op_end()-1; |
7169 | for (SDNode::op_iterator i = Call->op_begin()+2; i != e; ++i) |
7170 | Ops.push_back(*i); |
7171 | |
7172 | // Push live variables for the stack map. |
7173 | addStackMapLiveVars(CS, NumMetaOpers + NumArgs, Ops, *this); |
7174 | |
7175 | // Push the register mask info. |
7176 | if (HasGlue) |
7177 | Ops.push_back(*(Call->op_end()-2)); |
7178 | else |
7179 | Ops.push_back(*(Call->op_end()-1)); |
7180 | |
7181 | // Push the chain (this is originally the first operand of the call, but |
7182 | // becomes now the last or second to last operand). |
7183 | Ops.push_back(*(Call->op_begin())); |
7184 | |
7185 | // Push the glue flag (last operand). |
7186 | if (HasGlue) |
7187 | Ops.push_back(*(Call->op_end()-1)); |
7188 | |
7189 | SDVTList NodeTys; |
7190 | if (IsAnyRegCC && HasDef) { |
7191 | // Create the return types based on the intrinsic definition |
7192 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
7193 | SmallVector<EVT, 3> ValueVTs; |
7194 | ComputeValueVTs(TLI, CS->getType(), ValueVTs); |
7195 | assert(ValueVTs.size() == 1 && "Expected only one return value type.")((ValueVTs.size() == 1 && "Expected only one return value type." ) ? static_cast<void> (0) : __assert_fail ("ValueVTs.size() == 1 && \"Expected only one return value type.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7195, __PRETTY_FUNCTION__)); |
7196 | |
7197 | // There is always a chain and a glue type at the end |
7198 | ValueVTs.push_back(MVT::Other); |
7199 | ValueVTs.push_back(MVT::Glue); |
7200 | NodeTys = DAG.getVTList(ValueVTs); |
7201 | } else |
7202 | NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); |
7203 | |
7204 | // Replace the target specific call node with a PATCHPOINT node. |
7205 | MachineSDNode *MN = DAG.getMachineNode(TargetOpcode::PATCHPOINT, |
7206 | getCurSDLoc(), NodeTys, Ops); |
7207 | |
7208 | // Update the NodeMap. |
7209 | if (HasDef) { |
7210 | if (IsAnyRegCC) |
7211 | setValue(CS.getInstruction(), SDValue(MN, 0)); |
7212 | else |
7213 | setValue(CS.getInstruction(), Result.first); |
7214 | } |
7215 | |
7216 | // Fixup the consumers of the intrinsic. The chain and glue may be used in the |
7217 | // call sequence. Furthermore the location of the chain and glue can change |
7218 | // when the AnyReg calling convention is used and the intrinsic returns a |
7219 | // value. |
7220 | if (IsAnyRegCC && HasDef) { |
7221 | SDValue From[] = {SDValue(Call, 0), SDValue(Call, 1)}; |
7222 | SDValue To[] = {SDValue(MN, 1), SDValue(MN, 2)}; |
7223 | DAG.ReplaceAllUsesOfValuesWith(From, To, 2); |
7224 | } else |
7225 | DAG.ReplaceAllUsesWith(Call, MN); |
7226 | DAG.DeleteNode(Call); |
7227 | |
7228 | // Inform the Frame Information that we have a patchpoint in this function. |
7229 | FuncInfo.MF->getFrameInfo()->setHasPatchPoint(); |
7230 | } |
7231 | |
7232 | /// Returns an AttributeSet representing the attributes applied to the return |
7233 | /// value of the given call. |
7234 | static AttributeSet getReturnAttrs(TargetLowering::CallLoweringInfo &CLI) { |
7235 | SmallVector<Attribute::AttrKind, 2> Attrs; |
7236 | if (CLI.RetSExt) |
7237 | Attrs.push_back(Attribute::SExt); |
7238 | if (CLI.RetZExt) |
7239 | Attrs.push_back(Attribute::ZExt); |
7240 | if (CLI.IsInReg) |
7241 | Attrs.push_back(Attribute::InReg); |
7242 | |
7243 | return AttributeSet::get(CLI.RetTy->getContext(), AttributeSet::ReturnIndex, |
7244 | Attrs); |
7245 | } |
7246 | |
7247 | /// TargetLowering::LowerCallTo - This is the default LowerCallTo |
7248 | /// implementation, which just calls LowerCall. |
7249 | /// FIXME: When all targets are |
7250 | /// migrated to using LowerCall, this hook should be integrated into SDISel. |
7251 | std::pair<SDValue, SDValue> |
7252 | TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const { |
7253 | // Handle the incoming return values from the call. |
7254 | CLI.Ins.clear(); |
7255 | Type *OrigRetTy = CLI.RetTy; |
7256 | SmallVector<EVT, 4> RetTys; |
7257 | SmallVector<uint64_t, 4> Offsets; |
7258 | ComputeValueVTs(*this, CLI.RetTy, RetTys, &Offsets); |
7259 | |
7260 | SmallVector<ISD::OutputArg, 4> Outs; |
7261 | GetReturnInfo(CLI.RetTy, getReturnAttrs(CLI), Outs, *this); |
7262 | |
7263 | bool CanLowerReturn = |
7264 | this->CanLowerReturn(CLI.CallConv, CLI.DAG.getMachineFunction(), |
7265 | CLI.IsVarArg, Outs, CLI.RetTy->getContext()); |
7266 | |
7267 | SDValue DemoteStackSlot; |
7268 | int DemoteStackIdx = -100; |
7269 | if (!CanLowerReturn) { |
7270 | // FIXME: equivalent assert? |
7271 | // assert(!CS.hasInAllocaArgument() && |
7272 | // "sret demotion is incompatible with inalloca"); |
7273 | uint64_t TySize = getDataLayout()->getTypeAllocSize(CLI.RetTy); |
7274 | unsigned Align = getDataLayout()->getPrefTypeAlignment(CLI.RetTy); |
7275 | MachineFunction &MF = CLI.DAG.getMachineFunction(); |
7276 | DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false); |
7277 | Type *StackSlotPtrType = PointerType::getUnqual(CLI.RetTy); |
7278 | |
7279 | DemoteStackSlot = CLI.DAG.getFrameIndex(DemoteStackIdx, getPointerTy()); |
7280 | ArgListEntry Entry; |
7281 | Entry.Node = DemoteStackSlot; |
7282 | Entry.Ty = StackSlotPtrType; |
7283 | Entry.isSExt = false; |
7284 | Entry.isZExt = false; |
7285 | Entry.isInReg = false; |
7286 | Entry.isSRet = true; |
7287 | Entry.isNest = false; |
7288 | Entry.isByVal = false; |
7289 | Entry.isReturned = false; |
7290 | Entry.Alignment = Align; |
7291 | CLI.getArgs().insert(CLI.getArgs().begin(), Entry); |
7292 | CLI.RetTy = Type::getVoidTy(CLI.RetTy->getContext()); |
7293 | } else { |
7294 | for (unsigned I = 0, E = RetTys.size(); I != E; ++I) { |
7295 | EVT VT = RetTys[I]; |
7296 | MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT); |
7297 | unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT); |
7298 | for (unsigned i = 0; i != NumRegs; ++i) { |
7299 | ISD::InputArg MyFlags; |
7300 | MyFlags.VT = RegisterVT; |
7301 | MyFlags.ArgVT = VT; |
7302 | MyFlags.Used = CLI.IsReturnValueUsed; |
7303 | if (CLI.RetSExt) |
7304 | MyFlags.Flags.setSExt(); |
7305 | if (CLI.RetZExt) |
7306 | MyFlags.Flags.setZExt(); |
7307 | if (CLI.IsInReg) |
7308 | MyFlags.Flags.setInReg(); |
7309 | CLI.Ins.push_back(MyFlags); |
7310 | } |
7311 | } |
7312 | } |
7313 | |
7314 | // Handle all of the outgoing arguments. |
7315 | CLI.Outs.clear(); |
7316 | CLI.OutVals.clear(); |
7317 | ArgListTy &Args = CLI.getArgs(); |
7318 | for (unsigned i = 0, e = Args.size(); i != e; ++i) { |
7319 | SmallVector<EVT, 4> ValueVTs; |
7320 | ComputeValueVTs(*this, Args[i].Ty, ValueVTs); |
7321 | Type *FinalType = Args[i].Ty; |
7322 | if (Args[i].isByVal) |
7323 | FinalType = cast<PointerType>(Args[i].Ty)->getElementType(); |
7324 | bool NeedsRegBlock = functionArgumentNeedsConsecutiveRegisters( |
7325 | FinalType, CLI.CallConv, CLI.IsVarArg); |
7326 | for (unsigned Value = 0, NumValues = ValueVTs.size(); Value != NumValues; |
7327 | ++Value) { |
7328 | EVT VT = ValueVTs[Value]; |
7329 | Type *ArgTy = VT.getTypeForEVT(CLI.RetTy->getContext()); |
7330 | SDValue Op = SDValue(Args[i].Node.getNode(), |
7331 | Args[i].Node.getResNo() + Value); |
7332 | ISD::ArgFlagsTy Flags; |
7333 | unsigned OriginalAlignment = getDataLayout()->getABITypeAlignment(ArgTy); |
7334 | |
7335 | if (Args[i].isZExt) |
7336 | Flags.setZExt(); |
7337 | if (Args[i].isSExt) |
7338 | Flags.setSExt(); |
7339 | if (Args[i].isInReg) |
7340 | Flags.setInReg(); |
7341 | if (Args[i].isSRet) |
7342 | Flags.setSRet(); |
7343 | if (Args[i].isByVal) |
7344 | Flags.setByVal(); |
7345 | if (Args[i].isInAlloca) { |
7346 | Flags.setInAlloca(); |
7347 | // Set the byval flag for CCAssignFn callbacks that don't know about |
7348 | // inalloca. This way we can know how many bytes we should've allocated |
7349 | // and how many bytes a callee cleanup function will pop. If we port |
7350 | // inalloca to more targets, we'll have to add custom inalloca handling |
7351 | // in the various CC lowering callbacks. |
7352 | Flags.setByVal(); |
7353 | } |
7354 | if (Args[i].isByVal || Args[i].isInAlloca) { |
7355 | PointerType *Ty = cast<PointerType>(Args[i].Ty); |
7356 | Type *ElementTy = Ty->getElementType(); |
7357 | Flags.setByValSize(getDataLayout()->getTypeAllocSize(ElementTy)); |
7358 | // For ByVal, alignment should come from FE. BE will guess if this |
7359 | // info is not there but there are cases it cannot get right. |
7360 | unsigned FrameAlign; |
7361 | if (Args[i].Alignment) |
7362 | FrameAlign = Args[i].Alignment; |
7363 | else |
7364 | FrameAlign = getByValTypeAlignment(ElementTy); |
7365 | Flags.setByValAlign(FrameAlign); |
7366 | } |
7367 | if (Args[i].isNest) |
7368 | Flags.setNest(); |
7369 | if (NeedsRegBlock) { |
7370 | Flags.setInConsecutiveRegs(); |
7371 | if (Value == NumValues - 1) |
7372 | Flags.setInConsecutiveRegsLast(); |
7373 | } |
7374 | Flags.setOrigAlign(OriginalAlignment); |
7375 | |
7376 | MVT PartVT = getRegisterType(CLI.RetTy->getContext(), VT); |
7377 | unsigned NumParts = getNumRegisters(CLI.RetTy->getContext(), VT); |
7378 | SmallVector<SDValue, 4> Parts(NumParts); |
7379 | ISD::NodeType ExtendKind = ISD::ANY_EXTEND; |
7380 | |
7381 | if (Args[i].isSExt) |
7382 | ExtendKind = ISD::SIGN_EXTEND; |
7383 | else if (Args[i].isZExt) |
7384 | ExtendKind = ISD::ZERO_EXTEND; |
7385 | |
7386 | // Conservatively only handle 'returned' on non-vectors for now |
7387 | if (Args[i].isReturned && !Op.getValueType().isVector()) { |
7388 | assert(CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues &&((CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues && "unexpected use of 'returned'") ? static_cast< void> (0) : __assert_fail ("CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues && \"unexpected use of 'returned'\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7389, __PRETTY_FUNCTION__)) |
7389 | "unexpected use of 'returned'")((CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues && "unexpected use of 'returned'") ? static_cast< void> (0) : __assert_fail ("CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues && \"unexpected use of 'returned'\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7389, __PRETTY_FUNCTION__)); |
7390 | // Before passing 'returned' to the target lowering code, ensure that |
7391 | // either the register MVT and the actual EVT are the same size or that |
7392 | // the return value and argument are extended in the same way; in these |
7393 | // cases it's safe to pass the argument register value unchanged as the |
7394 | // return register value (although it's at the target's option whether |
7395 | // to do so) |
7396 | // TODO: allow code generation to take advantage of partially preserved |
7397 | // registers rather than clobbering the entire register when the |
7398 | // parameter extension method is not compatible with the return |
7399 | // extension method |
7400 | if ((NumParts * PartVT.getSizeInBits() == VT.getSizeInBits()) || |
7401 | (ExtendKind != ISD::ANY_EXTEND && |
7402 | CLI.RetSExt == Args[i].isSExt && CLI.RetZExt == Args[i].isZExt)) |
7403 | Flags.setReturned(); |
7404 | } |
7405 | |
7406 | getCopyToParts(CLI.DAG, CLI.DL, Op, &Parts[0], NumParts, PartVT, |
7407 | CLI.CS ? CLI.CS->getInstruction() : nullptr, ExtendKind); |
7408 | |
7409 | for (unsigned j = 0; j != NumParts; ++j) { |
7410 | // if it isn't first piece, alignment must be 1 |
7411 | ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(), VT, |
7412 | i < CLI.NumFixedArgs, |
7413 | i, j*Parts[j].getValueType().getStoreSize()); |
7414 | if (NumParts > 1 && j == 0) |
7415 | MyFlags.Flags.setSplit(); |
7416 | else if (j != 0) |
7417 | MyFlags.Flags.setOrigAlign(1); |
7418 | |
7419 | CLI.Outs.push_back(MyFlags); |
7420 | CLI.OutVals.push_back(Parts[j]); |
7421 | } |
7422 | } |
7423 | } |
7424 | |
7425 | SmallVector<SDValue, 4> InVals; |
7426 | CLI.Chain = LowerCall(CLI, InVals); |
7427 | |
7428 | // Verify that the target's LowerCall behaved as expected. |
7429 | assert(CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other &&((CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT ::Other && "LowerCall didn't return a valid chain!") ? static_cast<void> (0) : __assert_fail ("CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other && \"LowerCall didn't return a valid chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7430, __PRETTY_FUNCTION__)) |
7430 | "LowerCall didn't return a valid chain!")((CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT ::Other && "LowerCall didn't return a valid chain!") ? static_cast<void> (0) : __assert_fail ("CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other && \"LowerCall didn't return a valid chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7430, __PRETTY_FUNCTION__)); |
7431 | assert((!CLI.IsTailCall || InVals.empty()) &&(((!CLI.IsTailCall || InVals.empty()) && "LowerCall emitted a return value for a tail call!" ) ? static_cast<void> (0) : __assert_fail ("(!CLI.IsTailCall || InVals.empty()) && \"LowerCall emitted a return value for a tail call!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7432, __PRETTY_FUNCTION__)) |
7432 | "LowerCall emitted a return value for a tail call!")(((!CLI.IsTailCall || InVals.empty()) && "LowerCall emitted a return value for a tail call!" ) ? static_cast<void> (0) : __assert_fail ("(!CLI.IsTailCall || InVals.empty()) && \"LowerCall emitted a return value for a tail call!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7432, __PRETTY_FUNCTION__)); |
7433 | assert((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) &&(((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) && "LowerCall didn't emit the correct number of values!") ? static_cast <void> (0) : __assert_fail ("(CLI.IsTailCall || InVals.size() == CLI.Ins.size()) && \"LowerCall didn't emit the correct number of values!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7434, __PRETTY_FUNCTION__)) |
7434 | "LowerCall didn't emit the correct number of values!")(((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) && "LowerCall didn't emit the correct number of values!") ? static_cast <void> (0) : __assert_fail ("(CLI.IsTailCall || InVals.size() == CLI.Ins.size()) && \"LowerCall didn't emit the correct number of values!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7434, __PRETTY_FUNCTION__)); |
7435 | |
7436 | // For a tail call, the return value is merely live-out and there aren't |
7437 | // any nodes in the DAG representing it. Return a special value to |
7438 | // indicate that a tail call has been emitted and no more Instructions |
7439 | // should be processed in the current block. |
7440 | if (CLI.IsTailCall) { |
7441 | CLI.DAG.setRoot(CLI.Chain); |
7442 | return std::make_pair(SDValue(), SDValue()); |
7443 | } |
7444 | |
7445 | DEBUG(for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++i) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerCall emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerCall emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7447, __PRETTY_FUNCTION__)); ((EVT(CLI.Ins[i].VT) == InVals [i].getValueType() && "LowerCall emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(CLI.Ins[i].VT) == InVals[i].getValueType() && \"LowerCall emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7449, __PRETTY_FUNCTION__)); }; } } while (0) |
7446 | assert(InVals[i].getNode() &&do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerCall emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerCall emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7447, __PRETTY_FUNCTION__)); ((EVT(CLI.Ins[i].VT) == InVals [i].getValueType() && "LowerCall emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(CLI.Ins[i].VT) == InVals[i].getValueType() && \"LowerCall emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7449, __PRETTY_FUNCTION__)); }; } } while (0) |
7447 | "LowerCall emitted a null value!");do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerCall emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerCall emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7447, __PRETTY_FUNCTION__)); ((EVT(CLI.Ins[i].VT) == InVals [i].getValueType() && "LowerCall emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(CLI.Ins[i].VT) == InVals[i].getValueType() && \"LowerCall emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7449, __PRETTY_FUNCTION__)); }; } } while (0) |
7448 | assert(EVT(CLI.Ins[i].VT) == InVals[i].getValueType() &&do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerCall emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerCall emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7447, __PRETTY_FUNCTION__)); ((EVT(CLI.Ins[i].VT) == InVals [i].getValueType() && "LowerCall emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(CLI.Ins[i].VT) == InVals[i].getValueType() && \"LowerCall emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7449, __PRETTY_FUNCTION__)); }; } } while (0) |
7449 | "LowerCall emitted a value with the wrong type!");do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerCall emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerCall emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7447, __PRETTY_FUNCTION__)); ((EVT(CLI.Ins[i].VT) == InVals [i].getValueType() && "LowerCall emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(CLI.Ins[i].VT) == InVals[i].getValueType() && \"LowerCall emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7449, __PRETTY_FUNCTION__)); }; } } while (0) |
7450 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerCall emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerCall emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7447, __PRETTY_FUNCTION__)); ((EVT(CLI.Ins[i].VT) == InVals [i].getValueType() && "LowerCall emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(CLI.Ins[i].VT) == InVals[i].getValueType() && \"LowerCall emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7449, __PRETTY_FUNCTION__)); }; } } while (0); |
7451 | |
7452 | SmallVector<SDValue, 4> ReturnValues; |
7453 | if (!CanLowerReturn) { |
7454 | // The instruction result is the result of loading from the |
7455 | // hidden sret parameter. |
7456 | SmallVector<EVT, 1> PVTs; |
7457 | Type *PtrRetTy = PointerType::getUnqual(OrigRetTy); |
7458 | |
7459 | ComputeValueVTs(*this, PtrRetTy, PVTs); |
7460 | assert(PVTs.size() == 1 && "Pointers should fit in one register")((PVTs.size() == 1 && "Pointers should fit in one register" ) ? static_cast<void> (0) : __assert_fail ("PVTs.size() == 1 && \"Pointers should fit in one register\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7460, __PRETTY_FUNCTION__)); |
7461 | EVT PtrVT = PVTs[0]; |
7462 | |
7463 | unsigned NumValues = RetTys.size(); |
7464 | ReturnValues.resize(NumValues); |
7465 | SmallVector<SDValue, 4> Chains(NumValues); |
7466 | |
7467 | for (unsigned i = 0; i < NumValues; ++i) { |
7468 | SDValue Add = CLI.DAG.getNode(ISD::ADD, CLI.DL, PtrVT, DemoteStackSlot, |
7469 | CLI.DAG.getConstant(Offsets[i], PtrVT)); |
7470 | SDValue L = CLI.DAG.getLoad( |
7471 | RetTys[i], CLI.DL, CLI.Chain, Add, |
7472 | MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]), false, |
7473 | false, false, 1); |
7474 | ReturnValues[i] = L; |
7475 | Chains[i] = L.getValue(1); |
7476 | } |
7477 | |
7478 | CLI.Chain = CLI.DAG.getNode(ISD::TokenFactor, CLI.DL, MVT::Other, Chains); |
7479 | } else { |
7480 | // Collect the legal value parts into potentially illegal values |
7481 | // that correspond to the original function's return values. |
7482 | ISD::NodeType AssertOp = ISD::DELETED_NODE; |
7483 | if (CLI.RetSExt) |
7484 | AssertOp = ISD::AssertSext; |
7485 | else if (CLI.RetZExt) |
7486 | AssertOp = ISD::AssertZext; |
7487 | unsigned CurReg = 0; |
7488 | for (unsigned I = 0, E = RetTys.size(); I != E; ++I) { |
7489 | EVT VT = RetTys[I]; |
7490 | MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT); |
7491 | unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT); |
7492 | |
7493 | ReturnValues.push_back(getCopyFromParts(CLI.DAG, CLI.DL, &InVals[CurReg], |
7494 | NumRegs, RegisterVT, VT, nullptr, |
7495 | AssertOp)); |
7496 | CurReg += NumRegs; |
7497 | } |
7498 | |
7499 | // For a function returning void, there is no return value. We can't create |
7500 | // such a node, so we just return a null return value in that case. In |
7501 | // that case, nothing will actually look at the value. |
7502 | if (ReturnValues.empty()) |
7503 | return std::make_pair(SDValue(), CLI.Chain); |
7504 | } |
7505 | |
7506 | SDValue Res = CLI.DAG.getNode(ISD::MERGE_VALUES, CLI.DL, |
7507 | CLI.DAG.getVTList(RetTys), ReturnValues); |
7508 | return std::make_pair(Res, CLI.Chain); |
7509 | } |
7510 | |
7511 | void TargetLowering::LowerOperationWrapper(SDNode *N, |
7512 | SmallVectorImpl<SDValue> &Results, |
7513 | SelectionDAG &DAG) const { |
7514 | SDValue Res = LowerOperation(SDValue(N, 0), DAG); |
7515 | if (Res.getNode()) |
7516 | Results.push_back(Res); |
7517 | } |
7518 | |
7519 | SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { |
7520 | llvm_unreachable("LowerOperation not implemented for this target!")::llvm::llvm_unreachable_internal("LowerOperation not implemented for this target!" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7520); |
7521 | } |
7522 | |
7523 | void |
7524 | SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) { |
7525 | SDValue Op = getNonRegisterValue(V); |
7526 | assert((Op.getOpcode() != ISD::CopyFromReg ||(((Op.getOpcode() != ISD::CopyFromReg || cast<RegisterSDNode >(Op.getOperand(1))->getReg() != Reg) && "Copy from a reg to the same reg!" ) ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() != ISD::CopyFromReg || cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) && \"Copy from a reg to the same reg!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7528, __PRETTY_FUNCTION__)) |
7527 | cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&(((Op.getOpcode() != ISD::CopyFromReg || cast<RegisterSDNode >(Op.getOperand(1))->getReg() != Reg) && "Copy from a reg to the same reg!" ) ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() != ISD::CopyFromReg || cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) && \"Copy from a reg to the same reg!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7528, __PRETTY_FUNCTION__)) |
7528 | "Copy from a reg to the same reg!")(((Op.getOpcode() != ISD::CopyFromReg || cast<RegisterSDNode >(Op.getOperand(1))->getReg() != Reg) && "Copy from a reg to the same reg!" ) ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() != ISD::CopyFromReg || cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) && \"Copy from a reg to the same reg!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7528, __PRETTY_FUNCTION__)); |
7529 | assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg")((!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg" ) ? static_cast<void> (0) : __assert_fail ("!TargetRegisterInfo::isPhysicalRegister(Reg) && \"Is a physreg\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7529, __PRETTY_FUNCTION__)); |
7530 | |
7531 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
7532 | RegsForValue RFV(V->getContext(), TLI, Reg, V->getType()); |
7533 | SDValue Chain = DAG.getEntryNode(); |
7534 | |
7535 | ISD::NodeType ExtendType = (FuncInfo.PreferredExtendType.find(V) == |
7536 | FuncInfo.PreferredExtendType.end()) |
7537 | ? ISD::ANY_EXTEND |
7538 | : FuncInfo.PreferredExtendType[V]; |
7539 | RFV.getCopyToRegs(Op, DAG, getCurSDLoc(), Chain, nullptr, V, ExtendType); |
7540 | PendingExports.push_back(Chain); |
7541 | } |
7542 | |
7543 | #include "llvm/CodeGen/SelectionDAGISel.h" |
7544 | |
7545 | /// isOnlyUsedInEntryBlock - If the specified argument is only used in the |
7546 | /// entry block, return true. This includes arguments used by switches, since |
7547 | /// the switch may expand into multiple basic blocks. |
7548 | static bool isOnlyUsedInEntryBlock(const Argument *A, bool FastISel) { |
7549 | // With FastISel active, we may be splitting blocks, so force creation |
7550 | // of virtual registers for all non-dead arguments. |
7551 | if (FastISel) |
7552 | return A->use_empty(); |
7553 | |
7554 | const BasicBlock *Entry = A->getParent()->begin(); |
7555 | for (const User *U : A->users()) |
7556 | if (cast<Instruction>(U)->getParent() != Entry || isa<SwitchInst>(U)) |
7557 | return false; // Use not in entry block. |
7558 | |
7559 | return true; |
7560 | } |
7561 | |
7562 | void SelectionDAGISel::LowerArguments(const Function &F) { |
7563 | SelectionDAG &DAG = SDB->DAG; |
7564 | SDLoc dl = SDB->getCurSDLoc(); |
7565 | const DataLayout *DL = TLI->getDataLayout(); |
7566 | SmallVector<ISD::InputArg, 16> Ins; |
7567 | |
7568 | if (!FuncInfo->CanLowerReturn) { |
7569 | // Put in an sret pointer parameter before all the other parameters. |
7570 | SmallVector<EVT, 1> ValueVTs; |
7571 | ComputeValueVTs(*TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs); |
7572 | |
7573 | // NOTE: Assuming that a pointer will never break down to more than one VT |
7574 | // or one register. |
7575 | ISD::ArgFlagsTy Flags; |
7576 | Flags.setSRet(); |
7577 | MVT RegisterVT = TLI->getRegisterType(*DAG.getContext(), ValueVTs[0]); |
7578 | ISD::InputArg RetArg(Flags, RegisterVT, ValueVTs[0], true, 0, 0); |
7579 | Ins.push_back(RetArg); |
7580 | } |
7581 | |
7582 | // Set up the incoming argument description vector. |
7583 | unsigned Idx = 1; |
7584 | for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); |
7585 | I != E; ++I, ++Idx) { |
7586 | SmallVector<EVT, 4> ValueVTs; |
7587 | ComputeValueVTs(*TLI, I->getType(), ValueVTs); |
7588 | bool isArgValueUsed = !I->use_empty(); |
7589 | unsigned PartBase = 0; |
7590 | Type *FinalType = I->getType(); |
7591 | if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal)) |
7592 | FinalType = cast<PointerType>(FinalType)->getElementType(); |
7593 | bool NeedsRegBlock = TLI->functionArgumentNeedsConsecutiveRegisters( |
7594 | FinalType, F.getCallingConv(), F.isVarArg()); |
7595 | for (unsigned Value = 0, NumValues = ValueVTs.size(); |
7596 | Value != NumValues; ++Value) { |
7597 | EVT VT = ValueVTs[Value]; |
7598 | Type *ArgTy = VT.getTypeForEVT(*DAG.getContext()); |
7599 | ISD::ArgFlagsTy Flags; |
7600 | unsigned OriginalAlignment = DL->getABITypeAlignment(ArgTy); |
7601 | |
7602 | if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt)) |
7603 | Flags.setZExt(); |
7604 | if (F.getAttributes().hasAttribute(Idx, Attribute::SExt)) |
7605 | Flags.setSExt(); |
7606 | if (F.getAttributes().hasAttribute(Idx, Attribute::InReg)) |
7607 | Flags.setInReg(); |
7608 | if (F.getAttributes().hasAttribute(Idx, Attribute::StructRet)) |
7609 | Flags.setSRet(); |
7610 | if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal)) |
7611 | Flags.setByVal(); |
7612 | if (F.getAttributes().hasAttribute(Idx, Attribute::InAlloca)) { |
7613 | Flags.setInAlloca(); |
7614 | // Set the byval flag for CCAssignFn callbacks that don't know about |
7615 | // inalloca. This way we can know how many bytes we should've allocated |
7616 | // and how many bytes a callee cleanup function will pop. If we port |
7617 | // inalloca to more targets, we'll have to add custom inalloca handling |
7618 | // in the various CC lowering callbacks. |
7619 | Flags.setByVal(); |
7620 | } |
7621 | if (Flags.isByVal() || Flags.isInAlloca()) { |
7622 | PointerType *Ty = cast<PointerType>(I->getType()); |
7623 | Type *ElementTy = Ty->getElementType(); |
7624 | Flags.setByValSize(DL->getTypeAllocSize(ElementTy)); |
7625 | // For ByVal, alignment should be passed from FE. BE will guess if |
7626 | // this info is not there but there are cases it cannot get right. |
7627 | unsigned FrameAlign; |
7628 | if (F.getParamAlignment(Idx)) |
7629 | FrameAlign = F.getParamAlignment(Idx); |
7630 | else |
7631 | FrameAlign = TLI->getByValTypeAlignment(ElementTy); |
7632 | Flags.setByValAlign(FrameAlign); |
7633 | } |
7634 | if (F.getAttributes().hasAttribute(Idx, Attribute::Nest)) |
7635 | Flags.setNest(); |
7636 | if (NeedsRegBlock) { |
7637 | Flags.setInConsecutiveRegs(); |
7638 | if (Value == NumValues - 1) |
7639 | Flags.setInConsecutiveRegsLast(); |
7640 | } |
7641 | Flags.setOrigAlign(OriginalAlignment); |
7642 | |
7643 | MVT RegisterVT = TLI->getRegisterType(*CurDAG->getContext(), VT); |
7644 | unsigned NumRegs = TLI->getNumRegisters(*CurDAG->getContext(), VT); |
7645 | for (unsigned i = 0; i != NumRegs; ++i) { |
7646 | ISD::InputArg MyFlags(Flags, RegisterVT, VT, isArgValueUsed, |
7647 | Idx-1, PartBase+i*RegisterVT.getStoreSize()); |
7648 | if (NumRegs > 1 && i == 0) |
7649 | MyFlags.Flags.setSplit(); |
7650 | // if it isn't first piece, alignment must be 1 |
7651 | else if (i > 0) |
7652 | MyFlags.Flags.setOrigAlign(1); |
7653 | Ins.push_back(MyFlags); |
7654 | } |
7655 | PartBase += VT.getStoreSize(); |
7656 | } |
7657 | } |
7658 | |
7659 | // Call the target to set up the argument values. |
7660 | SmallVector<SDValue, 8> InVals; |
7661 | SDValue NewRoot = TLI->LowerFormalArguments( |
7662 | DAG.getRoot(), F.getCallingConv(), F.isVarArg(), Ins, dl, DAG, InVals); |
7663 | |
7664 | // Verify that the target's LowerFormalArguments behaved as expected. |
7665 | assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&((NewRoot.getNode() && NewRoot.getValueType() == MVT:: Other && "LowerFormalArguments didn't return a valid chain!" ) ? static_cast<void> (0) : __assert_fail ("NewRoot.getNode() && NewRoot.getValueType() == MVT::Other && \"LowerFormalArguments didn't return a valid chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7666, __PRETTY_FUNCTION__)) |
7666 | "LowerFormalArguments didn't return a valid chain!")((NewRoot.getNode() && NewRoot.getValueType() == MVT:: Other && "LowerFormalArguments didn't return a valid chain!" ) ? static_cast<void> (0) : __assert_fail ("NewRoot.getNode() && NewRoot.getValueType() == MVT::Other && \"LowerFormalArguments didn't return a valid chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7666, __PRETTY_FUNCTION__)); |
7667 | assert(InVals.size() == Ins.size() &&((InVals.size() == Ins.size() && "LowerFormalArguments didn't emit the correct number of values!" ) ? static_cast<void> (0) : __assert_fail ("InVals.size() == Ins.size() && \"LowerFormalArguments didn't emit the correct number of values!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7668, __PRETTY_FUNCTION__)) |
7668 | "LowerFormalArguments didn't emit the correct number of values!")((InVals.size() == Ins.size() && "LowerFormalArguments didn't emit the correct number of values!" ) ? static_cast<void> (0) : __assert_fail ("InVals.size() == Ins.size() && \"LowerFormalArguments didn't emit the correct number of values!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7668, __PRETTY_FUNCTION__)); |
7669 | DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7672, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7674, __PRETTY_FUNCTION__)); } }; } } while (0) |
7670 | for (unsigned i = 0, e = Ins.size(); i != e; ++i) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7672, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7674, __PRETTY_FUNCTION__)); } }; } } while (0) |
7671 | assert(InVals[i].getNode() &&do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7672, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7674, __PRETTY_FUNCTION__)); } }; } } while (0) |
7672 | "LowerFormalArguments emitted a null value!");do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7672, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7674, __PRETTY_FUNCTION__)); } }; } } while (0) |
7673 | assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7672, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7674, __PRETTY_FUNCTION__)); } }; } } while (0) |
7674 | "LowerFormalArguments emitted a value with the wrong type!");do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7672, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7674, __PRETTY_FUNCTION__)); } }; } } while (0) |
7675 | }do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7672, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7674, __PRETTY_FUNCTION__)); } }; } } while (0) |
7676 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7672, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7674, __PRETTY_FUNCTION__)); } }; } } while (0); |
7677 | |
7678 | // Update the DAG with the new chain value resulting from argument lowering. |
7679 | DAG.setRoot(NewRoot); |
7680 | |
7681 | // Set up the argument values. |
7682 | unsigned i = 0; |
7683 | Idx = 1; |
7684 | if (!FuncInfo->CanLowerReturn) { |
7685 | // Create a virtual register for the sret pointer, and put in a copy |
7686 | // from the sret argument into it. |
7687 | SmallVector<EVT, 1> ValueVTs; |
7688 | ComputeValueVTs(*TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs); |
7689 | MVT VT = ValueVTs[0].getSimpleVT(); |
7690 | MVT RegVT = TLI->getRegisterType(*CurDAG->getContext(), VT); |
7691 | ISD::NodeType AssertOp = ISD::DELETED_NODE; |
7692 | SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1, |
7693 | RegVT, VT, nullptr, AssertOp); |
7694 | |
7695 | MachineFunction& MF = SDB->DAG.getMachineFunction(); |
7696 | MachineRegisterInfo& RegInfo = MF.getRegInfo(); |
7697 | unsigned SRetReg = RegInfo.createVirtualRegister(TLI->getRegClassFor(RegVT)); |
7698 | FuncInfo->DemoteRegister = SRetReg; |
7699 | NewRoot = |
7700 | SDB->DAG.getCopyToReg(NewRoot, SDB->getCurSDLoc(), SRetReg, ArgValue); |
7701 | DAG.setRoot(NewRoot); |
7702 | |
7703 | // i indexes lowered arguments. Bump it past the hidden sret argument. |
7704 | // Idx indexes LLVM arguments. Don't touch it. |
7705 | ++i; |
7706 | } |
7707 | |
7708 | for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; |
7709 | ++I, ++Idx) { |
7710 | SmallVector<SDValue, 4> ArgValues; |
7711 | SmallVector<EVT, 4> ValueVTs; |
7712 | ComputeValueVTs(*TLI, I->getType(), ValueVTs); |
7713 | unsigned NumValues = ValueVTs.size(); |
7714 | |
7715 | // If this argument is unused then remember its value. It is used to generate |
7716 | // debugging information. |
7717 | if (I->use_empty() && NumValues) { |
7718 | SDB->setUnusedArgValue(I, InVals[i]); |
7719 | |
7720 | // Also remember any frame index for use in FastISel. |
7721 | if (FrameIndexSDNode *FI = |
7722 | dyn_cast<FrameIndexSDNode>(InVals[i].getNode())) |
7723 | FuncInfo->setArgumentFrameIndex(I, FI->getIndex()); |
7724 | } |
7725 | |
7726 | for (unsigned Val = 0; Val != NumValues; ++Val) { |
7727 | EVT VT = ValueVTs[Val]; |
7728 | MVT PartVT = TLI->getRegisterType(*CurDAG->getContext(), VT); |
7729 | unsigned NumParts = TLI->getNumRegisters(*CurDAG->getContext(), VT); |
7730 | |
7731 | if (!I->use_empty()) { |
7732 | ISD::NodeType AssertOp = ISD::DELETED_NODE; |
7733 | if (F.getAttributes().hasAttribute(Idx, Attribute::SExt)) |
7734 | AssertOp = ISD::AssertSext; |
7735 | else if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt)) |
7736 | AssertOp = ISD::AssertZext; |
7737 | |
7738 | ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i], |
7739 | NumParts, PartVT, VT, |
7740 | nullptr, AssertOp)); |
7741 | } |
7742 | |
7743 | i += NumParts; |
7744 | } |
7745 | |
7746 | // We don't need to do anything else for unused arguments. |
7747 | if (ArgValues.empty()) |
7748 | continue; |
7749 | |
7750 | // Note down frame index. |
7751 | if (FrameIndexSDNode *FI = |
7752 | dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode())) |
7753 | FuncInfo->setArgumentFrameIndex(I, FI->getIndex()); |
7754 | |
7755 | SDValue Res = DAG.getMergeValues(makeArrayRef(ArgValues.data(), NumValues), |
7756 | SDB->getCurSDLoc()); |
7757 | |
7758 | SDB->setValue(I, Res); |
7759 | if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::BUILD_PAIR) { |
7760 | if (LoadSDNode *LNode = |
7761 | dyn_cast<LoadSDNode>(Res.getOperand(0).getNode())) |
7762 | if (FrameIndexSDNode *FI = |
7763 | dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode())) |
7764 | FuncInfo->setArgumentFrameIndex(I, FI->getIndex()); |
7765 | } |
7766 | |
7767 | // If this argument is live outside of the entry block, insert a copy from |
7768 | // wherever we got it to the vreg that other BB's will reference it as. |
7769 | if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::CopyFromReg) { |
7770 | // If we can, though, try to skip creating an unnecessary vreg. |
7771 | // FIXME: This isn't very clean... it would be nice to make this more |
7772 | // general. It's also subtly incompatible with the hacks FastISel |
7773 | // uses with vregs. |
7774 | unsigned Reg = cast<RegisterSDNode>(Res.getOperand(1))->getReg(); |
7775 | if (TargetRegisterInfo::isVirtualRegister(Reg)) { |
7776 | FuncInfo->ValueMap[I] = Reg; |
7777 | continue; |
7778 | } |
7779 | } |
7780 | if (!isOnlyUsedInEntryBlock(I, TM.Options.EnableFastISel)) { |
7781 | FuncInfo->InitializeRegForValue(I); |
7782 | SDB->CopyToExportRegsIfNeeded(I); |
7783 | } |
7784 | } |
7785 | |
7786 | assert(i == InVals.size() && "Argument register count mismatch!")((i == InVals.size() && "Argument register count mismatch!" ) ? static_cast<void> (0) : __assert_fail ("i == InVals.size() && \"Argument register count mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7786, __PRETTY_FUNCTION__)); |
7787 | |
7788 | // Finally, if the target has anything special to do, allow it to do so. |
7789 | // FIXME: this should insert code into the DAG! |
7790 | EmitFunctionEntryCode(); |
7791 | } |
7792 | |
7793 | /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to |
7794 | /// ensure constants are generated when needed. Remember the virtual registers |
7795 | /// that need to be added to the Machine PHI nodes as input. We cannot just |
7796 | /// directly add them, because expansion might result in multiple MBB's for one |
7797 | /// BB. As such, the start of the BB might correspond to a different MBB than |
7798 | /// the end. |
7799 | /// |
7800 | void |
7801 | SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) { |
7802 | const TerminatorInst *TI = LLVMBB->getTerminator(); |
7803 | |
7804 | SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled; |
7805 | |
7806 | // Check successor nodes' PHI nodes that expect a constant to be available |
7807 | // from this block. |
7808 | for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) { |
7809 | const BasicBlock *SuccBB = TI->getSuccessor(succ); |
7810 | if (!isa<PHINode>(SuccBB->begin())) continue; |
7811 | MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB]; |
7812 | |
7813 | // If this terminator has multiple identical successors (common for |
7814 | // switches), only handle each succ once. |
7815 | if (!SuccsHandled.insert(SuccMBB).second) |
7816 | continue; |
7817 | |
7818 | MachineBasicBlock::iterator MBBI = SuccMBB->begin(); |
7819 | |
7820 | // At this point we know that there is a 1-1 correspondence between LLVM PHI |
7821 | // nodes and Machine PHI nodes, but the incoming operands have not been |
7822 | // emitted yet. |
7823 | for (BasicBlock::const_iterator I = SuccBB->begin(); |
7824 | const PHINode *PN = dyn_cast<PHINode>(I); ++I) { |
7825 | // Ignore dead phi's. |
7826 | if (PN->use_empty()) continue; |
7827 | |
7828 | // Skip empty types |
7829 | if (PN->getType()->isEmptyTy()) |
7830 | continue; |
7831 | |
7832 | unsigned Reg; |
7833 | const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB); |
7834 | |
7835 | if (const Constant *C = dyn_cast<Constant>(PHIOp)) { |
7836 | unsigned &RegOut = ConstantsOut[C]; |
7837 | if (RegOut == 0) { |
7838 | RegOut = FuncInfo.CreateRegs(C->getType()); |
7839 | CopyValueToVirtualRegister(C, RegOut); |
7840 | } |
7841 | Reg = RegOut; |
7842 | } else { |
7843 | DenseMap<const Value *, unsigned>::iterator I = |
7844 | FuncInfo.ValueMap.find(PHIOp); |
7845 | if (I != FuncInfo.ValueMap.end()) |
7846 | Reg = I->second; |
7847 | else { |
7848 | assert(isa<AllocaInst>(PHIOp) &&((isa<AllocaInst>(PHIOp) && FuncInfo.StaticAllocaMap .count(cast<AllocaInst>(PHIOp)) && "Didn't codegen value into a register!??" ) ? static_cast<void> (0) : __assert_fail ("isa<AllocaInst>(PHIOp) && FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) && \"Didn't codegen value into a register!??\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7850, __PRETTY_FUNCTION__)) |
7849 | FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&((isa<AllocaInst>(PHIOp) && FuncInfo.StaticAllocaMap .count(cast<AllocaInst>(PHIOp)) && "Didn't codegen value into a register!??" ) ? static_cast<void> (0) : __assert_fail ("isa<AllocaInst>(PHIOp) && FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) && \"Didn't codegen value into a register!??\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7850, __PRETTY_FUNCTION__)) |
7850 | "Didn't codegen value into a register!??")((isa<AllocaInst>(PHIOp) && FuncInfo.StaticAllocaMap .count(cast<AllocaInst>(PHIOp)) && "Didn't codegen value into a register!??" ) ? static_cast<void> (0) : __assert_fail ("isa<AllocaInst>(PHIOp) && FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) && \"Didn't codegen value into a register!??\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224240/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7850, __PRETTY_FUNCTION__)); |
7851 | Reg = FuncInfo.CreateRegs(PHIOp->getType()); |
7852 | CopyValueToVirtualRegister(PHIOp, Reg); |
7853 | } |
7854 | } |
7855 | |
7856 | // Remember that this register needs to added to the machine PHI node as |
7857 | // the input for this MBB. |
7858 | SmallVector<EVT, 4> ValueVTs; |
7859 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
7860 | ComputeValueVTs(TLI, PN->getType(), ValueVTs); |
7861 | for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) { |
7862 | EVT VT = ValueVTs[vti]; |
7863 | unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT); |
7864 | for (unsigned i = 0, e = NumRegisters; i != e; ++i) |
7865 | FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i)); |
7866 | Reg += NumRegisters; |
7867 | } |
7868 | } |
7869 | } |
7870 | |
7871 | ConstantsOut.clear(); |
7872 | } |
7873 | |
7874 | /// Add a successor MBB to ParentMBB< creating a new MachineBB for BB if SuccMBB |
7875 | /// is 0. |
7876 | MachineBasicBlock * |
7877 | SelectionDAGBuilder::StackProtectorDescriptor:: |
7878 | AddSuccessorMBB(const BasicBlock *BB, |
7879 | MachineBasicBlock *ParentMBB, |
7880 | bool IsLikely, |
7881 | MachineBasicBlock *SuccMBB) { |
7882 | // If SuccBB has not been created yet, create it. |
7883 | if (!SuccMBB) { |
7884 | MachineFunction *MF = ParentMBB->getParent(); |
7885 | MachineFunction::iterator BBI = ParentMBB; |
7886 | SuccMBB = MF->CreateMachineBasicBlock(BB); |
7887 | MF->insert(++BBI, SuccMBB); |
7888 | } |
7889 | // Add it as a successor of ParentMBB. |
7890 | ParentMBB->addSuccessor( |
7891 | SuccMBB, BranchProbabilityInfo::getBranchWeightStackProtector(IsLikely)); |
7892 | return SuccMBB; |
7893 | } |