File: | build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/llvm/lib/CodeGen/CodeGenPrepare.cpp |
Warning: | line 1128, column 37 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===// | |||
2 | // | |||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
4 | // See https://llvm.org/LICENSE.txt for license information. | |||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
6 | // | |||
7 | //===----------------------------------------------------------------------===// | |||
8 | // | |||
9 | // This pass munges the code in the input function to better prepare it for | |||
10 | // SelectionDAG-based code generation. This works around limitations in it's | |||
11 | // basic-block-at-a-time approach. It should eventually be removed. | |||
12 | // | |||
13 | //===----------------------------------------------------------------------===// | |||
14 | ||||
15 | #include "llvm/ADT/APInt.h" | |||
16 | #include "llvm/ADT/ArrayRef.h" | |||
17 | #include "llvm/ADT/DenseMap.h" | |||
18 | #include "llvm/ADT/MapVector.h" | |||
19 | #include "llvm/ADT/PointerIntPair.h" | |||
20 | #include "llvm/ADT/STLExtras.h" | |||
21 | #include "llvm/ADT/SmallPtrSet.h" | |||
22 | #include "llvm/ADT/SmallVector.h" | |||
23 | #include "llvm/ADT/Statistic.h" | |||
24 | #include "llvm/Analysis/BlockFrequencyInfo.h" | |||
25 | #include "llvm/Analysis/BranchProbabilityInfo.h" | |||
26 | #include "llvm/Analysis/InstructionSimplify.h" | |||
27 | #include "llvm/Analysis/LoopInfo.h" | |||
28 | #include "llvm/Analysis/ProfileSummaryInfo.h" | |||
29 | #include "llvm/Analysis/TargetLibraryInfo.h" | |||
30 | #include "llvm/Analysis/TargetTransformInfo.h" | |||
31 | #include "llvm/Analysis/ValueTracking.h" | |||
32 | #include "llvm/Analysis/VectorUtils.h" | |||
33 | #include "llvm/CodeGen/Analysis.h" | |||
34 | #include "llvm/CodeGen/ISDOpcodes.h" | |||
35 | #include "llvm/CodeGen/SelectionDAGNodes.h" | |||
36 | #include "llvm/CodeGen/TargetLowering.h" | |||
37 | #include "llvm/CodeGen/TargetPassConfig.h" | |||
38 | #include "llvm/CodeGen/TargetSubtargetInfo.h" | |||
39 | #include "llvm/CodeGen/ValueTypes.h" | |||
40 | #include "llvm/Config/llvm-config.h" | |||
41 | #include "llvm/IR/Argument.h" | |||
42 | #include "llvm/IR/Attributes.h" | |||
43 | #include "llvm/IR/BasicBlock.h" | |||
44 | #include "llvm/IR/Constant.h" | |||
45 | #include "llvm/IR/Constants.h" | |||
46 | #include "llvm/IR/DataLayout.h" | |||
47 | #include "llvm/IR/DebugInfo.h" | |||
48 | #include "llvm/IR/DerivedTypes.h" | |||
49 | #include "llvm/IR/Dominators.h" | |||
50 | #include "llvm/IR/Function.h" | |||
51 | #include "llvm/IR/GetElementPtrTypeIterator.h" | |||
52 | #include "llvm/IR/GlobalValue.h" | |||
53 | #include "llvm/IR/GlobalVariable.h" | |||
54 | #include "llvm/IR/IRBuilder.h" | |||
55 | #include "llvm/IR/InlineAsm.h" | |||
56 | #include "llvm/IR/InstrTypes.h" | |||
57 | #include "llvm/IR/Instruction.h" | |||
58 | #include "llvm/IR/Instructions.h" | |||
59 | #include "llvm/IR/IntrinsicInst.h" | |||
60 | #include "llvm/IR/Intrinsics.h" | |||
61 | #include "llvm/IR/IntrinsicsAArch64.h" | |||
62 | #include "llvm/IR/LLVMContext.h" | |||
63 | #include "llvm/IR/MDBuilder.h" | |||
64 | #include "llvm/IR/Module.h" | |||
65 | #include "llvm/IR/Operator.h" | |||
66 | #include "llvm/IR/PatternMatch.h" | |||
67 | #include "llvm/IR/Statepoint.h" | |||
68 | #include "llvm/IR/Type.h" | |||
69 | #include "llvm/IR/Use.h" | |||
70 | #include "llvm/IR/User.h" | |||
71 | #include "llvm/IR/Value.h" | |||
72 | #include "llvm/IR/ValueHandle.h" | |||
73 | #include "llvm/IR/ValueMap.h" | |||
74 | #include "llvm/InitializePasses.h" | |||
75 | #include "llvm/Pass.h" | |||
76 | #include "llvm/Support/BlockFrequency.h" | |||
77 | #include "llvm/Support/BranchProbability.h" | |||
78 | #include "llvm/Support/Casting.h" | |||
79 | #include "llvm/Support/CommandLine.h" | |||
80 | #include "llvm/Support/Compiler.h" | |||
81 | #include "llvm/Support/Debug.h" | |||
82 | #include "llvm/Support/ErrorHandling.h" | |||
83 | #include "llvm/Support/MachineValueType.h" | |||
84 | #include "llvm/Support/MathExtras.h" | |||
85 | #include "llvm/Support/raw_ostream.h" | |||
86 | #include "llvm/Target/TargetMachine.h" | |||
87 | #include "llvm/Target/TargetOptions.h" | |||
88 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | |||
89 | #include "llvm/Transforms/Utils/BypassSlowDivision.h" | |||
90 | #include "llvm/Transforms/Utils/Local.h" | |||
91 | #include "llvm/Transforms/Utils/SimplifyLibCalls.h" | |||
92 | #include "llvm/Transforms/Utils/SizeOpts.h" | |||
93 | #include <algorithm> | |||
94 | #include <cassert> | |||
95 | #include <cstdint> | |||
96 | #include <iterator> | |||
97 | #include <limits> | |||
98 | #include <memory> | |||
99 | #include <utility> | |||
100 | #include <vector> | |||
101 | ||||
102 | using namespace llvm; | |||
103 | using namespace llvm::PatternMatch; | |||
104 | ||||
105 | #define DEBUG_TYPE"codegenprepare" "codegenprepare" | |||
106 | ||||
107 | STATISTIC(NumBlocksElim, "Number of blocks eliminated")static llvm::Statistic NumBlocksElim = {"codegenprepare", "NumBlocksElim" , "Number of blocks eliminated"}; | |||
108 | STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated")static llvm::Statistic NumPHIsElim = {"codegenprepare", "NumPHIsElim" , "Number of trivial PHIs eliminated"}; | |||
109 | STATISTIC(NumGEPsElim, "Number of GEPs converted to casts")static llvm::Statistic NumGEPsElim = {"codegenprepare", "NumGEPsElim" , "Number of GEPs converted to casts"}; | |||
110 | STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "static llvm::Statistic NumCmpUses = {"codegenprepare", "NumCmpUses" , "Number of uses of Cmp expressions replaced with uses of " "sunken Cmps" } | |||
111 | "sunken Cmps")static llvm::Statistic NumCmpUses = {"codegenprepare", "NumCmpUses" , "Number of uses of Cmp expressions replaced with uses of " "sunken Cmps" }; | |||
112 | STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "static llvm::Statistic NumCastUses = {"codegenprepare", "NumCastUses" , "Number of uses of Cast expressions replaced with uses " "of sunken Casts" } | |||
113 | "of sunken Casts")static llvm::Statistic NumCastUses = {"codegenprepare", "NumCastUses" , "Number of uses of Cast expressions replaced with uses " "of sunken Casts" }; | |||
114 | STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "static llvm::Statistic NumMemoryInsts = {"codegenprepare", "NumMemoryInsts" , "Number of memory instructions whose address " "computations were sunk" } | |||
115 | "computations were sunk")static llvm::Statistic NumMemoryInsts = {"codegenprepare", "NumMemoryInsts" , "Number of memory instructions whose address " "computations were sunk" }; | |||
116 | STATISTIC(NumMemoryInstsPhiCreated,static llvm::Statistic NumMemoryInstsPhiCreated = {"codegenprepare" , "NumMemoryInstsPhiCreated", "Number of phis created when address " "computations were sunk to memory instructions"} | |||
117 | "Number of phis created when address "static llvm::Statistic NumMemoryInstsPhiCreated = {"codegenprepare" , "NumMemoryInstsPhiCreated", "Number of phis created when address " "computations were sunk to memory instructions"} | |||
118 | "computations were sunk to memory instructions")static llvm::Statistic NumMemoryInstsPhiCreated = {"codegenprepare" , "NumMemoryInstsPhiCreated", "Number of phis created when address " "computations were sunk to memory instructions"}; | |||
119 | STATISTIC(NumMemoryInstsSelectCreated,static llvm::Statistic NumMemoryInstsSelectCreated = {"codegenprepare" , "NumMemoryInstsSelectCreated", "Number of select created when address " "computations were sunk to memory instructions"} | |||
120 | "Number of select created when address "static llvm::Statistic NumMemoryInstsSelectCreated = {"codegenprepare" , "NumMemoryInstsSelectCreated", "Number of select created when address " "computations were sunk to memory instructions"} | |||
121 | "computations were sunk to memory instructions")static llvm::Statistic NumMemoryInstsSelectCreated = {"codegenprepare" , "NumMemoryInstsSelectCreated", "Number of select created when address " "computations were sunk to memory instructions"}; | |||
122 | STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads")static llvm::Statistic NumExtsMoved = {"codegenprepare", "NumExtsMoved" , "Number of [s|z]ext instructions combined with loads"}; | |||
123 | STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized")static llvm::Statistic NumExtUses = {"codegenprepare", "NumExtUses" , "Number of uses of [s|z]ext instructions optimized"}; | |||
124 | STATISTIC(NumAndsAdded,static llvm::Statistic NumAndsAdded = {"codegenprepare", "NumAndsAdded" , "Number of and mask instructions added to form ext loads"} | |||
125 | "Number of and mask instructions added to form ext loads")static llvm::Statistic NumAndsAdded = {"codegenprepare", "NumAndsAdded" , "Number of and mask instructions added to form ext loads"}; | |||
126 | STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized")static llvm::Statistic NumAndUses = {"codegenprepare", "NumAndUses" , "Number of uses of and mask instructions optimized"}; | |||
127 | STATISTIC(NumRetsDup, "Number of return instructions duplicated")static llvm::Statistic NumRetsDup = {"codegenprepare", "NumRetsDup" , "Number of return instructions duplicated"}; | |||
128 | STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved")static llvm::Statistic NumDbgValueMoved = {"codegenprepare", "NumDbgValueMoved" , "Number of debug value instructions moved"}; | |||
129 | STATISTIC(NumSelectsExpanded, "Number of selects turned into branches")static llvm::Statistic NumSelectsExpanded = {"codegenprepare" , "NumSelectsExpanded", "Number of selects turned into branches" }; | |||
130 | STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed")static llvm::Statistic NumStoreExtractExposed = {"codegenprepare" , "NumStoreExtractExposed", "Number of store(extractelement) exposed" }; | |||
131 | ||||
132 | static cl::opt<bool> DisableBranchOpts( | |||
133 | "disable-cgp-branch-opts", cl::Hidden, cl::init(false), | |||
134 | cl::desc("Disable branch optimizations in CodeGenPrepare")); | |||
135 | ||||
136 | static cl::opt<bool> | |||
137 | DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false), | |||
138 | cl::desc("Disable GC optimizations in CodeGenPrepare")); | |||
139 | ||||
140 | static cl::opt<bool> DisableSelectToBranch( | |||
141 | "disable-cgp-select2branch", cl::Hidden, cl::init(false), | |||
142 | cl::desc("Disable select to branch conversion.")); | |||
143 | ||||
144 | static cl::opt<bool> AddrSinkUsingGEPs( | |||
145 | "addr-sink-using-gep", cl::Hidden, cl::init(true), | |||
146 | cl::desc("Address sinking in CGP using GEPs.")); | |||
147 | ||||
148 | static cl::opt<bool> EnableAndCmpSinking( | |||
149 | "enable-andcmp-sinking", cl::Hidden, cl::init(true), | |||
150 | cl::desc("Enable sinkinig and/cmp into branches.")); | |||
151 | ||||
152 | static cl::opt<bool> DisableStoreExtract( | |||
153 | "disable-cgp-store-extract", cl::Hidden, cl::init(false), | |||
154 | cl::desc("Disable store(extract) optimizations in CodeGenPrepare")); | |||
155 | ||||
156 | static cl::opt<bool> StressStoreExtract( | |||
157 | "stress-cgp-store-extract", cl::Hidden, cl::init(false), | |||
158 | cl::desc("Stress test store(extract) optimizations in CodeGenPrepare")); | |||
159 | ||||
160 | static cl::opt<bool> DisableExtLdPromotion( | |||
161 | "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false), | |||
162 | cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in " | |||
163 | "CodeGenPrepare")); | |||
164 | ||||
165 | static cl::opt<bool> StressExtLdPromotion( | |||
166 | "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false), | |||
167 | cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) " | |||
168 | "optimization in CodeGenPrepare")); | |||
169 | ||||
170 | static cl::opt<bool> DisablePreheaderProtect( | |||
171 | "disable-preheader-prot", cl::Hidden, cl::init(false), | |||
172 | cl::desc("Disable protection against removing loop preheaders")); | |||
173 | ||||
174 | static cl::opt<bool> ProfileGuidedSectionPrefix( | |||
175 | "profile-guided-section-prefix", cl::Hidden, cl::init(true), cl::ZeroOrMore, | |||
176 | cl::desc("Use profile info to add section prefix for hot/cold functions")); | |||
177 | ||||
178 | static cl::opt<bool> ProfileUnknownInSpecialSection( | |||
179 | "profile-unknown-in-special-section", cl::Hidden, cl::init(false), | |||
180 | cl::ZeroOrMore, | |||
181 | cl::desc("In profiling mode like sampleFDO, if a function doesn't have " | |||
182 | "profile, we cannot tell the function is cold for sure because " | |||
183 | "it may be a function newly added without ever being sampled. " | |||
184 | "With the flag enabled, compiler can put such profile unknown " | |||
185 | "functions into a special section, so runtime system can choose " | |||
186 | "to handle it in a different way than .text section, to save " | |||
187 | "RAM for example. ")); | |||
188 | ||||
189 | static cl::opt<unsigned> FreqRatioToSkipMerge( | |||
190 | "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2), | |||
191 | cl::desc("Skip merging empty blocks if (frequency of empty block) / " | |||
192 | "(frequency of destination block) is greater than this ratio")); | |||
193 | ||||
194 | static cl::opt<bool> ForceSplitStore( | |||
195 | "force-split-store", cl::Hidden, cl::init(false), | |||
196 | cl::desc("Force store splitting no matter what the target query says.")); | |||
197 | ||||
198 | static cl::opt<bool> | |||
199 | EnableTypePromotionMerge("cgp-type-promotion-merge", cl::Hidden, | |||
200 | cl::desc("Enable merging of redundant sexts when one is dominating" | |||
201 | " the other."), cl::init(true)); | |||
202 | ||||
203 | static cl::opt<bool> DisableComplexAddrModes( | |||
204 | "disable-complex-addr-modes", cl::Hidden, cl::init(false), | |||
205 | cl::desc("Disables combining addressing modes with different parts " | |||
206 | "in optimizeMemoryInst.")); | |||
207 | ||||
208 | static cl::opt<bool> | |||
209 | AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(false), | |||
210 | cl::desc("Allow creation of Phis in Address sinking.")); | |||
211 | ||||
212 | static cl::opt<bool> | |||
213 | AddrSinkNewSelects("addr-sink-new-select", cl::Hidden, cl::init(true), | |||
214 | cl::desc("Allow creation of selects in Address sinking.")); | |||
215 | ||||
216 | static cl::opt<bool> AddrSinkCombineBaseReg( | |||
217 | "addr-sink-combine-base-reg", cl::Hidden, cl::init(true), | |||
218 | cl::desc("Allow combining of BaseReg field in Address sinking.")); | |||
219 | ||||
220 | static cl::opt<bool> AddrSinkCombineBaseGV( | |||
221 | "addr-sink-combine-base-gv", cl::Hidden, cl::init(true), | |||
222 | cl::desc("Allow combining of BaseGV field in Address sinking.")); | |||
223 | ||||
224 | static cl::opt<bool> AddrSinkCombineBaseOffs( | |||
225 | "addr-sink-combine-base-offs", cl::Hidden, cl::init(true), | |||
226 | cl::desc("Allow combining of BaseOffs field in Address sinking.")); | |||
227 | ||||
228 | static cl::opt<bool> AddrSinkCombineScaledReg( | |||
229 | "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(true), | |||
230 | cl::desc("Allow combining of ScaledReg field in Address sinking.")); | |||
231 | ||||
232 | static cl::opt<bool> | |||
233 | EnableGEPOffsetSplit("cgp-split-large-offset-gep", cl::Hidden, | |||
234 | cl::init(true), | |||
235 | cl::desc("Enable splitting large offset of GEP.")); | |||
236 | ||||
237 | static cl::opt<bool> EnableICMP_EQToICMP_ST( | |||
238 | "cgp-icmp-eq2icmp-st", cl::Hidden, cl::init(false), | |||
239 | cl::desc("Enable ICMP_EQ to ICMP_S(L|G)T conversion.")); | |||
240 | ||||
241 | static cl::opt<bool> | |||
242 | VerifyBFIUpdates("cgp-verify-bfi-updates", cl::Hidden, cl::init(false), | |||
243 | cl::desc("Enable BFI update verification for " | |||
244 | "CodeGenPrepare.")); | |||
245 | ||||
246 | static cl::opt<bool> OptimizePhiTypes( | |||
247 | "cgp-optimize-phi-types", cl::Hidden, cl::init(false), | |||
248 | cl::desc("Enable converting phi types in CodeGenPrepare")); | |||
249 | ||||
250 | namespace { | |||
251 | ||||
252 | enum ExtType { | |||
253 | ZeroExtension, // Zero extension has been seen. | |||
254 | SignExtension, // Sign extension has been seen. | |||
255 | BothExtension // This extension type is used if we saw sext after | |||
256 | // ZeroExtension had been set, or if we saw zext after | |||
257 | // SignExtension had been set. It makes the type | |||
258 | // information of a promoted instruction invalid. | |||
259 | }; | |||
260 | ||||
261 | using SetOfInstrs = SmallPtrSet<Instruction *, 16>; | |||
262 | using TypeIsSExt = PointerIntPair<Type *, 2, ExtType>; | |||
263 | using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>; | |||
264 | using SExts = SmallVector<Instruction *, 16>; | |||
265 | using ValueToSExts = DenseMap<Value *, SExts>; | |||
266 | ||||
267 | class TypePromotionTransaction; | |||
268 | ||||
269 | class CodeGenPrepare : public FunctionPass { | |||
270 | const TargetMachine *TM = nullptr; | |||
271 | const TargetSubtargetInfo *SubtargetInfo; | |||
272 | const TargetLowering *TLI = nullptr; | |||
273 | const TargetRegisterInfo *TRI; | |||
274 | const TargetTransformInfo *TTI = nullptr; | |||
275 | const TargetLibraryInfo *TLInfo; | |||
276 | const LoopInfo *LI; | |||
277 | std::unique_ptr<BlockFrequencyInfo> BFI; | |||
278 | std::unique_ptr<BranchProbabilityInfo> BPI; | |||
279 | ProfileSummaryInfo *PSI; | |||
280 | ||||
281 | /// As we scan instructions optimizing them, this is the next instruction | |||
282 | /// to optimize. Transforms that can invalidate this should update it. | |||
283 | BasicBlock::iterator CurInstIterator; | |||
284 | ||||
285 | /// Keeps track of non-local addresses that have been sunk into a block. | |||
286 | /// This allows us to avoid inserting duplicate code for blocks with | |||
287 | /// multiple load/stores of the same address. The usage of WeakTrackingVH | |||
288 | /// enables SunkAddrs to be treated as a cache whose entries can be | |||
289 | /// invalidated if a sunken address computation has been erased. | |||
290 | ValueMap<Value*, WeakTrackingVH> SunkAddrs; | |||
291 | ||||
292 | /// Keeps track of all instructions inserted for the current function. | |||
293 | SetOfInstrs InsertedInsts; | |||
294 | ||||
295 | /// Keeps track of the type of the related instruction before their | |||
296 | /// promotion for the current function. | |||
297 | InstrToOrigTy PromotedInsts; | |||
298 | ||||
299 | /// Keep track of instructions removed during promotion. | |||
300 | SetOfInstrs RemovedInsts; | |||
301 | ||||
302 | /// Keep track of sext chains based on their initial value. | |||
303 | DenseMap<Value *, Instruction *> SeenChainsForSExt; | |||
304 | ||||
305 | /// Keep track of GEPs accessing the same data structures such as structs or | |||
306 | /// arrays that are candidates to be split later because of their large | |||
307 | /// size. | |||
308 | MapVector< | |||
309 | AssertingVH<Value>, | |||
310 | SmallVector<std::pair<AssertingVH<GetElementPtrInst>, int64_t>, 32>> | |||
311 | LargeOffsetGEPMap; | |||
312 | ||||
313 | /// Keep track of new GEP base after splitting the GEPs having large offset. | |||
314 | SmallSet<AssertingVH<Value>, 2> NewGEPBases; | |||
315 | ||||
316 | /// Map serial numbers to Large offset GEPs. | |||
317 | DenseMap<AssertingVH<GetElementPtrInst>, int> LargeOffsetGEPID; | |||
318 | ||||
319 | /// Keep track of SExt promoted. | |||
320 | ValueToSExts ValToSExtendedUses; | |||
321 | ||||
322 | /// True if the function has the OptSize attribute. | |||
323 | bool OptSize; | |||
324 | ||||
325 | /// DataLayout for the Function being processed. | |||
326 | const DataLayout *DL = nullptr; | |||
327 | ||||
328 | /// Building the dominator tree can be expensive, so we only build it | |||
329 | /// lazily and update it when required. | |||
330 | std::unique_ptr<DominatorTree> DT; | |||
331 | ||||
332 | public: | |||
333 | static char ID; // Pass identification, replacement for typeid | |||
334 | ||||
335 | CodeGenPrepare() : FunctionPass(ID) { | |||
336 | initializeCodeGenPreparePass(*PassRegistry::getPassRegistry()); | |||
337 | } | |||
338 | ||||
339 | bool runOnFunction(Function &F) override; | |||
340 | ||||
341 | StringRef getPassName() const override { return "CodeGen Prepare"; } | |||
342 | ||||
343 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
344 | // FIXME: When we can selectively preserve passes, preserve the domtree. | |||
345 | AU.addRequired<ProfileSummaryInfoWrapperPass>(); | |||
346 | AU.addRequired<TargetLibraryInfoWrapperPass>(); | |||
347 | AU.addRequired<TargetPassConfig>(); | |||
348 | AU.addRequired<TargetTransformInfoWrapperPass>(); | |||
349 | AU.addRequired<LoopInfoWrapperPass>(); | |||
350 | } | |||
351 | ||||
352 | private: | |||
353 | template <typename F> | |||
354 | void resetIteratorIfInvalidatedWhileCalling(BasicBlock *BB, F f) { | |||
355 | // Substituting can cause recursive simplifications, which can invalidate | |||
356 | // our iterator. Use a WeakTrackingVH to hold onto it in case this | |||
357 | // happens. | |||
358 | Value *CurValue = &*CurInstIterator; | |||
359 | WeakTrackingVH IterHandle(CurValue); | |||
360 | ||||
361 | f(); | |||
362 | ||||
363 | // If the iterator instruction was recursively deleted, start over at the | |||
364 | // start of the block. | |||
365 | if (IterHandle != CurValue) { | |||
366 | CurInstIterator = BB->begin(); | |||
367 | SunkAddrs.clear(); | |||
368 | } | |||
369 | } | |||
370 | ||||
371 | // Get the DominatorTree, building if necessary. | |||
372 | DominatorTree &getDT(Function &F) { | |||
373 | if (!DT) | |||
374 | DT = std::make_unique<DominatorTree>(F); | |||
375 | return *DT; | |||
376 | } | |||
377 | ||||
378 | void removeAllAssertingVHReferences(Value *V); | |||
379 | bool eliminateAssumptions(Function &F); | |||
380 | bool eliminateFallThrough(Function &F); | |||
381 | bool eliminateMostlyEmptyBlocks(Function &F); | |||
382 | BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB); | |||
383 | bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; | |||
384 | void eliminateMostlyEmptyBlock(BasicBlock *BB); | |||
385 | bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB, | |||
386 | bool isPreheader); | |||
387 | bool makeBitReverse(Instruction &I); | |||
388 | bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT); | |||
389 | bool optimizeInst(Instruction *I, bool &ModifiedDT); | |||
390 | bool optimizeMemoryInst(Instruction *MemoryInst, Value *Addr, | |||
391 | Type *AccessTy, unsigned AddrSpace); | |||
392 | bool optimizeGatherScatterInst(Instruction *MemoryInst, Value *Ptr); | |||
393 | bool optimizeInlineAsmInst(CallInst *CS); | |||
394 | bool optimizeCallInst(CallInst *CI, bool &ModifiedDT); | |||
395 | bool optimizeExt(Instruction *&I); | |||
396 | bool optimizeExtUses(Instruction *I); | |||
397 | bool optimizeLoadExt(LoadInst *Load); | |||
398 | bool optimizeShiftInst(BinaryOperator *BO); | |||
399 | bool optimizeFunnelShift(IntrinsicInst *Fsh); | |||
400 | bool optimizeSelectInst(SelectInst *SI); | |||
401 | bool optimizeShuffleVectorInst(ShuffleVectorInst *SVI); | |||
402 | bool optimizeSwitchInst(SwitchInst *SI); | |||
403 | bool optimizeExtractElementInst(Instruction *Inst); | |||
404 | bool dupRetToEnableTailCallOpts(BasicBlock *BB, bool &ModifiedDT); | |||
405 | bool fixupDbgValue(Instruction *I); | |||
406 | bool placeDbgValues(Function &F); | |||
407 | bool placePseudoProbes(Function &F); | |||
408 | bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts, | |||
409 | LoadInst *&LI, Instruction *&Inst, bool HasPromoted); | |||
410 | bool tryToPromoteExts(TypePromotionTransaction &TPT, | |||
411 | const SmallVectorImpl<Instruction *> &Exts, | |||
412 | SmallVectorImpl<Instruction *> &ProfitablyMovedExts, | |||
413 | unsigned CreatedInstsCost = 0); | |||
414 | bool mergeSExts(Function &F); | |||
415 | bool splitLargeGEPOffsets(); | |||
416 | bool optimizePhiType(PHINode *Inst, SmallPtrSetImpl<PHINode *> &Visited, | |||
417 | SmallPtrSetImpl<Instruction *> &DeletedInstrs); | |||
418 | bool optimizePhiTypes(Function &F); | |||
419 | bool performAddressTypePromotion( | |||
420 | Instruction *&Inst, | |||
421 | bool AllowPromotionWithoutCommonHeader, | |||
422 | bool HasPromoted, TypePromotionTransaction &TPT, | |||
423 | SmallVectorImpl<Instruction *> &SpeculativelyMovedExts); | |||
424 | bool splitBranchCondition(Function &F, bool &ModifiedDT); | |||
425 | bool simplifyOffsetableRelocate(GCStatepointInst &I); | |||
426 | ||||
427 | bool tryToSinkFreeOperands(Instruction *I); | |||
428 | bool replaceMathCmpWithIntrinsic(BinaryOperator *BO, Value *Arg0, | |||
429 | Value *Arg1, CmpInst *Cmp, | |||
430 | Intrinsic::ID IID); | |||
431 | bool optimizeCmp(CmpInst *Cmp, bool &ModifiedDT); | |||
432 | bool combineToUSubWithOverflow(CmpInst *Cmp, bool &ModifiedDT); | |||
433 | bool combineToUAddWithOverflow(CmpInst *Cmp, bool &ModifiedDT); | |||
434 | void verifyBFIUpdates(Function &F); | |||
435 | }; | |||
436 | ||||
437 | } // end anonymous namespace | |||
438 | ||||
439 | char CodeGenPrepare::ID = 0; | |||
440 | ||||
441 | INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,static void *initializeCodeGenPreparePassOnce(PassRegistry & Registry) { | |||
442 | "Optimize for code generation", false, false)static void *initializeCodeGenPreparePassOnce(PassRegistry & Registry) { | |||
443 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry); | |||
444 | INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)initializeProfileSummaryInfoWrapperPassPass(Registry); | |||
445 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); | |||
446 | INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)initializeTargetPassConfigPass(Registry); | |||
447 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | |||
448 | INITIALIZE_PASS_END(CodeGenPrepare, DEBUG_TYPE,PassInfo *PI = new PassInfo( "Optimize for code generation", "codegenprepare" , &CodeGenPrepare::ID, PassInfo::NormalCtor_t(callDefaultCtor <CodeGenPrepare>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeCodeGenPreparePassFlag ; void llvm::initializeCodeGenPreparePass(PassRegistry &Registry ) { llvm::call_once(InitializeCodeGenPreparePassFlag, initializeCodeGenPreparePassOnce , std::ref(Registry)); } | |||
449 | "Optimize for code generation", false, false)PassInfo *PI = new PassInfo( "Optimize for code generation", "codegenprepare" , &CodeGenPrepare::ID, PassInfo::NormalCtor_t(callDefaultCtor <CodeGenPrepare>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeCodeGenPreparePassFlag ; void llvm::initializeCodeGenPreparePass(PassRegistry &Registry ) { llvm::call_once(InitializeCodeGenPreparePassFlag, initializeCodeGenPreparePassOnce , std::ref(Registry)); } | |||
450 | ||||
451 | FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); } | |||
452 | ||||
453 | bool CodeGenPrepare::runOnFunction(Function &F) { | |||
454 | if (skipFunction(F)) | |||
455 | return false; | |||
456 | ||||
457 | DL = &F.getParent()->getDataLayout(); | |||
458 | ||||
459 | bool EverMadeChange = false; | |||
460 | // Clear per function information. | |||
461 | InsertedInsts.clear(); | |||
462 | PromotedInsts.clear(); | |||
463 | ||||
464 | TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>(); | |||
465 | SubtargetInfo = TM->getSubtargetImpl(F); | |||
466 | TLI = SubtargetInfo->getTargetLowering(); | |||
467 | TRI = SubtargetInfo->getRegisterInfo(); | |||
468 | TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); | |||
469 | TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | |||
470 | LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | |||
471 | BPI.reset(new BranchProbabilityInfo(F, *LI)); | |||
472 | BFI.reset(new BlockFrequencyInfo(F, *BPI, *LI)); | |||
473 | PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); | |||
474 | OptSize = F.hasOptSize(); | |||
475 | if (ProfileGuidedSectionPrefix) { | |||
476 | // The hot attribute overwrites profile count based hotness while profile | |||
477 | // counts based hotness overwrite the cold attribute. | |||
478 | // This is a conservative behabvior. | |||
479 | if (F.hasFnAttribute(Attribute::Hot) || | |||
480 | PSI->isFunctionHotInCallGraph(&F, *BFI)) | |||
481 | F.setSectionPrefix("hot"); | |||
482 | // If PSI shows this function is not hot, we will placed the function | |||
483 | // into unlikely section if (1) PSI shows this is a cold function, or | |||
484 | // (2) the function has a attribute of cold. | |||
485 | else if (PSI->isFunctionColdInCallGraph(&F, *BFI) || | |||
486 | F.hasFnAttribute(Attribute::Cold)) | |||
487 | F.setSectionPrefix("unlikely"); | |||
488 | else if (ProfileUnknownInSpecialSection && PSI->hasPartialSampleProfile() && | |||
489 | PSI->isFunctionHotnessUnknown(F)) | |||
490 | F.setSectionPrefix("unknown"); | |||
491 | } | |||
492 | ||||
493 | /// This optimization identifies DIV instructions that can be | |||
494 | /// profitably bypassed and carried out with a shorter, faster divide. | |||
495 | if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI->isSlowDivBypassed()) { | |||
496 | const DenseMap<unsigned int, unsigned int> &BypassWidths = | |||
497 | TLI->getBypassSlowDivWidths(); | |||
498 | BasicBlock* BB = &*F.begin(); | |||
499 | while (BB != nullptr) { | |||
500 | // bypassSlowDivision may create new BBs, but we don't want to reapply the | |||
501 | // optimization to those blocks. | |||
502 | BasicBlock* Next = BB->getNextNode(); | |||
503 | // F.hasOptSize is already checked in the outer if statement. | |||
504 | if (!llvm::shouldOptimizeForSize(BB, PSI, BFI.get())) | |||
505 | EverMadeChange |= bypassSlowDivision(BB, BypassWidths); | |||
506 | BB = Next; | |||
507 | } | |||
508 | } | |||
509 | ||||
510 | // Get rid of @llvm.assume builtins before attempting to eliminate empty | |||
511 | // blocks, since there might be blocks that only contain @llvm.assume calls | |||
512 | // (plus arguments that we can get rid of). | |||
513 | EverMadeChange |= eliminateAssumptions(F); | |||
514 | ||||
515 | // Eliminate blocks that contain only PHI nodes and an | |||
516 | // unconditional branch. | |||
517 | EverMadeChange |= eliminateMostlyEmptyBlocks(F); | |||
518 | ||||
519 | bool ModifiedDT = false; | |||
520 | if (!DisableBranchOpts) | |||
521 | EverMadeChange |= splitBranchCondition(F, ModifiedDT); | |||
522 | ||||
523 | // Split some critical edges where one of the sources is an indirect branch, | |||
524 | // to help generate sane code for PHIs involving such edges. | |||
525 | EverMadeChange |= | |||
526 | SplitIndirectBrCriticalEdges(F, /*IgnoreBlocksWithoutPHI=*/true); | |||
527 | ||||
528 | bool MadeChange = true; | |||
529 | while (MadeChange) { | |||
530 | MadeChange = false; | |||
531 | DT.reset(); | |||
532 | for (BasicBlock &BB : llvm::make_early_inc_range(F)) { | |||
533 | bool ModifiedDTOnIteration = false; | |||
534 | MadeChange |= optimizeBlock(BB, ModifiedDTOnIteration); | |||
535 | ||||
536 | // Restart BB iteration if the dominator tree of the Function was changed | |||
537 | if (ModifiedDTOnIteration) | |||
538 | break; | |||
539 | } | |||
540 | if (EnableTypePromotionMerge && !ValToSExtendedUses.empty()) | |||
541 | MadeChange |= mergeSExts(F); | |||
542 | if (!LargeOffsetGEPMap.empty()) | |||
543 | MadeChange |= splitLargeGEPOffsets(); | |||
544 | MadeChange |= optimizePhiTypes(F); | |||
545 | ||||
546 | if (MadeChange) | |||
547 | eliminateFallThrough(F); | |||
548 | ||||
549 | // Really free removed instructions during promotion. | |||
550 | for (Instruction *I : RemovedInsts) | |||
551 | I->deleteValue(); | |||
552 | ||||
553 | EverMadeChange |= MadeChange; | |||
554 | SeenChainsForSExt.clear(); | |||
555 | ValToSExtendedUses.clear(); | |||
556 | RemovedInsts.clear(); | |||
557 | LargeOffsetGEPMap.clear(); | |||
558 | LargeOffsetGEPID.clear(); | |||
559 | } | |||
560 | ||||
561 | NewGEPBases.clear(); | |||
562 | SunkAddrs.clear(); | |||
563 | ||||
564 | if (!DisableBranchOpts) { | |||
565 | MadeChange = false; | |||
566 | // Use a set vector to get deterministic iteration order. The order the | |||
567 | // blocks are removed may affect whether or not PHI nodes in successors | |||
568 | // are removed. | |||
569 | SmallSetVector<BasicBlock*, 8> WorkList; | |||
570 | for (BasicBlock &BB : F) { | |||
571 | SmallVector<BasicBlock *, 2> Successors(successors(&BB)); | |||
572 | MadeChange |= ConstantFoldTerminator(&BB, true); | |||
573 | if (!MadeChange) continue; | |||
574 | ||||
575 | for (BasicBlock *Succ : Successors) | |||
576 | if (pred_empty(Succ)) | |||
577 | WorkList.insert(Succ); | |||
578 | } | |||
579 | ||||
580 | // Delete the dead blocks and any of their dead successors. | |||
581 | MadeChange |= !WorkList.empty(); | |||
582 | while (!WorkList.empty()) { | |||
583 | BasicBlock *BB = WorkList.pop_back_val(); | |||
584 | SmallVector<BasicBlock*, 2> Successors(successors(BB)); | |||
585 | ||||
586 | DeleteDeadBlock(BB); | |||
587 | ||||
588 | for (BasicBlock *Succ : Successors) | |||
589 | if (pred_empty(Succ)) | |||
590 | WorkList.insert(Succ); | |||
591 | } | |||
592 | ||||
593 | // Merge pairs of basic blocks with unconditional branches, connected by | |||
594 | // a single edge. | |||
595 | if (EverMadeChange || MadeChange) | |||
596 | MadeChange |= eliminateFallThrough(F); | |||
597 | ||||
598 | EverMadeChange |= MadeChange; | |||
599 | } | |||
600 | ||||
601 | if (!DisableGCOpts) { | |||
602 | SmallVector<GCStatepointInst *, 2> Statepoints; | |||
603 | for (BasicBlock &BB : F) | |||
604 | for (Instruction &I : BB) | |||
605 | if (auto *SP = dyn_cast<GCStatepointInst>(&I)) | |||
606 | Statepoints.push_back(SP); | |||
607 | for (auto &I : Statepoints) | |||
608 | EverMadeChange |= simplifyOffsetableRelocate(*I); | |||
609 | } | |||
610 | ||||
611 | // Do this last to clean up use-before-def scenarios introduced by other | |||
612 | // preparatory transforms. | |||
613 | EverMadeChange |= placeDbgValues(F); | |||
614 | EverMadeChange |= placePseudoProbes(F); | |||
615 | ||||
616 | #ifndef NDEBUG | |||
617 | if (VerifyBFIUpdates) | |||
618 | verifyBFIUpdates(F); | |||
619 | #endif | |||
620 | ||||
621 | return EverMadeChange; | |||
622 | } | |||
623 | ||||
624 | bool CodeGenPrepare::eliminateAssumptions(Function &F) { | |||
625 | bool MadeChange = false; | |||
626 | for (BasicBlock &BB : F) { | |||
627 | CurInstIterator = BB.begin(); | |||
628 | while (CurInstIterator != BB.end()) { | |||
629 | Instruction *I = &*(CurInstIterator++); | |||
630 | if (auto *Assume = dyn_cast<AssumeInst>(I)) { | |||
631 | MadeChange = true; | |||
632 | Value *Operand = Assume->getOperand(0); | |||
633 | Assume->eraseFromParent(); | |||
634 | ||||
635 | resetIteratorIfInvalidatedWhileCalling(&BB, [&]() { | |||
636 | RecursivelyDeleteTriviallyDeadInstructions(Operand, TLInfo, nullptr); | |||
637 | }); | |||
638 | } | |||
639 | } | |||
640 | } | |||
641 | return MadeChange; | |||
642 | } | |||
643 | ||||
644 | /// An instruction is about to be deleted, so remove all references to it in our | |||
645 | /// GEP-tracking data strcutures. | |||
646 | void CodeGenPrepare::removeAllAssertingVHReferences(Value *V) { | |||
647 | LargeOffsetGEPMap.erase(V); | |||
648 | NewGEPBases.erase(V); | |||
649 | ||||
650 | auto GEP = dyn_cast<GetElementPtrInst>(V); | |||
651 | if (!GEP) | |||
652 | return; | |||
653 | ||||
654 | LargeOffsetGEPID.erase(GEP); | |||
655 | ||||
656 | auto VecI = LargeOffsetGEPMap.find(GEP->getPointerOperand()); | |||
657 | if (VecI == LargeOffsetGEPMap.end()) | |||
658 | return; | |||
659 | ||||
660 | auto &GEPVector = VecI->second; | |||
661 | llvm::erase_if(GEPVector, [=](auto &Elt) { return Elt.first == GEP; }); | |||
662 | ||||
663 | if (GEPVector.empty()) | |||
664 | LargeOffsetGEPMap.erase(VecI); | |||
665 | } | |||
666 | ||||
667 | // Verify BFI has been updated correctly by recomputing BFI and comparing them. | |||
668 | void LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) CodeGenPrepare::verifyBFIUpdates(Function &F) { | |||
669 | DominatorTree NewDT(F); | |||
670 | LoopInfo NewLI(NewDT); | |||
671 | BranchProbabilityInfo NewBPI(F, NewLI, TLInfo); | |||
672 | BlockFrequencyInfo NewBFI(F, NewBPI, NewLI); | |||
673 | NewBFI.verifyMatch(*BFI); | |||
674 | } | |||
675 | ||||
676 | /// Merge basic blocks which are connected by a single edge, where one of the | |||
677 | /// basic blocks has a single successor pointing to the other basic block, | |||
678 | /// which has a single predecessor. | |||
679 | bool CodeGenPrepare::eliminateFallThrough(Function &F) { | |||
680 | bool Changed = false; | |||
681 | // Scan all of the blocks in the function, except for the entry block. | |||
682 | // Use a temporary array to avoid iterator being invalidated when | |||
683 | // deleting blocks. | |||
684 | SmallVector<WeakTrackingVH, 16> Blocks; | |||
685 | for (auto &Block : llvm::drop_begin(F)) | |||
686 | Blocks.push_back(&Block); | |||
687 | ||||
688 | SmallSet<WeakTrackingVH, 16> Preds; | |||
689 | for (auto &Block : Blocks) { | |||
690 | auto *BB = cast_or_null<BasicBlock>(Block); | |||
691 | if (!BB) | |||
692 | continue; | |||
693 | // If the destination block has a single pred, then this is a trivial | |||
694 | // edge, just collapse it. | |||
695 | BasicBlock *SinglePred = BB->getSinglePredecessor(); | |||
696 | ||||
697 | // Don't merge if BB's address is taken. | |||
698 | if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue; | |||
699 | ||||
700 | BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator()); | |||
701 | if (Term && !Term->isConditional()) { | |||
702 | Changed = true; | |||
703 | LLVM_DEBUG(dbgs() << "To merge:\n" << *BB << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "To merge:\n" << * BB << "\n\n\n"; } } while (false); | |||
704 | ||||
705 | // Merge BB into SinglePred and delete it. | |||
706 | MergeBlockIntoPredecessor(BB); | |||
707 | Preds.insert(SinglePred); | |||
708 | } | |||
709 | } | |||
710 | ||||
711 | // (Repeatedly) merging blocks into their predecessors can create redundant | |||
712 | // debug intrinsics. | |||
713 | for (auto &Pred : Preds) | |||
714 | if (auto *BB = cast_or_null<BasicBlock>(Pred)) | |||
715 | RemoveRedundantDbgInstrs(BB); | |||
716 | ||||
717 | return Changed; | |||
718 | } | |||
719 | ||||
720 | /// Find a destination block from BB if BB is mergeable empty block. | |||
721 | BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) { | |||
722 | // If this block doesn't end with an uncond branch, ignore it. | |||
723 | BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); | |||
724 | if (!BI || !BI->isUnconditional()) | |||
725 | return nullptr; | |||
726 | ||||
727 | // If the instruction before the branch (skipping debug info) isn't a phi | |||
728 | // node, then other stuff is happening here. | |||
729 | BasicBlock::iterator BBI = BI->getIterator(); | |||
730 | if (BBI != BB->begin()) { | |||
731 | --BBI; | |||
732 | while (isa<DbgInfoIntrinsic>(BBI)) { | |||
733 | if (BBI == BB->begin()) | |||
734 | break; | |||
735 | --BBI; | |||
736 | } | |||
737 | if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI)) | |||
738 | return nullptr; | |||
739 | } | |||
740 | ||||
741 | // Do not break infinite loops. | |||
742 | BasicBlock *DestBB = BI->getSuccessor(0); | |||
743 | if (DestBB == BB) | |||
744 | return nullptr; | |||
745 | ||||
746 | if (!canMergeBlocks(BB, DestBB)) | |||
747 | DestBB = nullptr; | |||
748 | ||||
749 | return DestBB; | |||
750 | } | |||
751 | ||||
752 | /// Eliminate blocks that contain only PHI nodes, debug info directives, and an | |||
753 | /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split | |||
754 | /// edges in ways that are non-optimal for isel. Start by eliminating these | |||
755 | /// blocks so we can split them the way we want them. | |||
756 | bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) { | |||
757 | SmallPtrSet<BasicBlock *, 16> Preheaders; | |||
758 | SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end()); | |||
759 | while (!LoopList.empty()) { | |||
760 | Loop *L = LoopList.pop_back_val(); | |||
761 | llvm::append_range(LoopList, *L); | |||
762 | if (BasicBlock *Preheader = L->getLoopPreheader()) | |||
763 | Preheaders.insert(Preheader); | |||
764 | } | |||
765 | ||||
766 | bool MadeChange = false; | |||
767 | // Copy blocks into a temporary array to avoid iterator invalidation issues | |||
768 | // as we remove them. | |||
769 | // Note that this intentionally skips the entry block. | |||
770 | SmallVector<WeakTrackingVH, 16> Blocks; | |||
771 | for (auto &Block : llvm::drop_begin(F)) | |||
772 | Blocks.push_back(&Block); | |||
773 | ||||
774 | for (auto &Block : Blocks) { | |||
775 | BasicBlock *BB = cast_or_null<BasicBlock>(Block); | |||
776 | if (!BB) | |||
777 | continue; | |||
778 | BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB); | |||
779 | if (!DestBB || | |||
780 | !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB))) | |||
781 | continue; | |||
782 | ||||
783 | eliminateMostlyEmptyBlock(BB); | |||
784 | MadeChange = true; | |||
785 | } | |||
786 | return MadeChange; | |||
787 | } | |||
788 | ||||
789 | bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB, | |||
790 | BasicBlock *DestBB, | |||
791 | bool isPreheader) { | |||
792 | // Do not delete loop preheaders if doing so would create a critical edge. | |||
793 | // Loop preheaders can be good locations to spill registers. If the | |||
794 | // preheader is deleted and we create a critical edge, registers may be | |||
795 | // spilled in the loop body instead. | |||
796 | if (!DisablePreheaderProtect && isPreheader && | |||
797 | !(BB->getSinglePredecessor() && | |||
798 | BB->getSinglePredecessor()->getSingleSuccessor())) | |||
799 | return false; | |||
800 | ||||
801 | // Skip merging if the block's successor is also a successor to any callbr | |||
802 | // that leads to this block. | |||
803 | // FIXME: Is this really needed? Is this a correctness issue? | |||
804 | for (BasicBlock *Pred : predecessors(BB)) { | |||
805 | if (auto *CBI = dyn_cast<CallBrInst>((Pred)->getTerminator())) | |||
806 | for (unsigned i = 0, e = CBI->getNumSuccessors(); i != e; ++i) | |||
807 | if (DestBB == CBI->getSuccessor(i)) | |||
808 | return false; | |||
809 | } | |||
810 | ||||
811 | // Try to skip merging if the unique predecessor of BB is terminated by a | |||
812 | // switch or indirect branch instruction, and BB is used as an incoming block | |||
813 | // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to | |||
814 | // add COPY instructions in the predecessor of BB instead of BB (if it is not | |||
815 | // merged). Note that the critical edge created by merging such blocks wont be | |||
816 | // split in MachineSink because the jump table is not analyzable. By keeping | |||
817 | // such empty block (BB), ISel will place COPY instructions in BB, not in the | |||
818 | // predecessor of BB. | |||
819 | BasicBlock *Pred = BB->getUniquePredecessor(); | |||
820 | if (!Pred || | |||
821 | !(isa<SwitchInst>(Pred->getTerminator()) || | |||
822 | isa<IndirectBrInst>(Pred->getTerminator()))) | |||
823 | return true; | |||
824 | ||||
825 | if (BB->getTerminator() != BB->getFirstNonPHIOrDbg()) | |||
826 | return true; | |||
827 | ||||
828 | // We use a simple cost heuristic which determine skipping merging is | |||
829 | // profitable if the cost of skipping merging is less than the cost of | |||
830 | // merging : Cost(skipping merging) < Cost(merging BB), where the | |||
831 | // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and | |||
832 | // the Cost(merging BB) is Freq(Pred) * Cost(Copy). | |||
833 | // Assuming Cost(Copy) == Cost(Branch), we could simplify it to : | |||
834 | // Freq(Pred) / Freq(BB) > 2. | |||
835 | // Note that if there are multiple empty blocks sharing the same incoming | |||
836 | // value for the PHIs in the DestBB, we consider them together. In such | |||
837 | // case, Cost(merging BB) will be the sum of their frequencies. | |||
838 | ||||
839 | if (!isa<PHINode>(DestBB->begin())) | |||
840 | return true; | |||
841 | ||||
842 | SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs; | |||
843 | ||||
844 | // Find all other incoming blocks from which incoming values of all PHIs in | |||
845 | // DestBB are the same as the ones from BB. | |||
846 | for (BasicBlock *DestBBPred : predecessors(DestBB)) { | |||
847 | if (DestBBPred == BB) | |||
848 | continue; | |||
849 | ||||
850 | if (llvm::all_of(DestBB->phis(), [&](const PHINode &DestPN) { | |||
851 | return DestPN.getIncomingValueForBlock(BB) == | |||
852 | DestPN.getIncomingValueForBlock(DestBBPred); | |||
853 | })) | |||
854 | SameIncomingValueBBs.insert(DestBBPred); | |||
855 | } | |||
856 | ||||
857 | // See if all BB's incoming values are same as the value from Pred. In this | |||
858 | // case, no reason to skip merging because COPYs are expected to be place in | |||
859 | // Pred already. | |||
860 | if (SameIncomingValueBBs.count(Pred)) | |||
861 | return true; | |||
862 | ||||
863 | BlockFrequency PredFreq = BFI->getBlockFreq(Pred); | |||
864 | BlockFrequency BBFreq = BFI->getBlockFreq(BB); | |||
865 | ||||
866 | for (auto *SameValueBB : SameIncomingValueBBs) | |||
867 | if (SameValueBB->getUniquePredecessor() == Pred && | |||
868 | DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB)) | |||
869 | BBFreq += BFI->getBlockFreq(SameValueBB); | |||
870 | ||||
871 | return PredFreq.getFrequency() <= | |||
872 | BBFreq.getFrequency() * FreqRatioToSkipMerge; | |||
873 | } | |||
874 | ||||
875 | /// Return true if we can merge BB into DestBB if there is a single | |||
876 | /// unconditional branch between them, and BB contains no other non-phi | |||
877 | /// instructions. | |||
878 | bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB, | |||
879 | const BasicBlock *DestBB) const { | |||
880 | // We only want to eliminate blocks whose phi nodes are used by phi nodes in | |||
881 | // the successor. If there are more complex condition (e.g. preheaders), | |||
882 | // don't mess around with them. | |||
883 | for (const PHINode &PN : BB->phis()) { | |||
884 | for (const User *U : PN.users()) { | |||
885 | const Instruction *UI = cast<Instruction>(U); | |||
886 | if (UI->getParent() != DestBB || !isa<PHINode>(UI)) | |||
887 | return false; | |||
888 | // If User is inside DestBB block and it is a PHINode then check | |||
889 | // incoming value. If incoming value is not from BB then this is | |||
890 | // a complex condition (e.g. preheaders) we want to avoid here. | |||
891 | if (UI->getParent() == DestBB) { | |||
892 | if (const PHINode *UPN = dyn_cast<PHINode>(UI)) | |||
893 | for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) { | |||
894 | Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I)); | |||
895 | if (Insn && Insn->getParent() == BB && | |||
896 | Insn->getParent() != UPN->getIncomingBlock(I)) | |||
897 | return false; | |||
898 | } | |||
899 | } | |||
900 | } | |||
901 | } | |||
902 | ||||
903 | // If BB and DestBB contain any common predecessors, then the phi nodes in BB | |||
904 | // and DestBB may have conflicting incoming values for the block. If so, we | |||
905 | // can't merge the block. | |||
906 | const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin()); | |||
907 | if (!DestBBPN) return true; // no conflict. | |||
908 | ||||
909 | // Collect the preds of BB. | |||
910 | SmallPtrSet<const BasicBlock*, 16> BBPreds; | |||
911 | if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { | |||
912 | // It is faster to get preds from a PHI than with pred_iterator. | |||
913 | for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) | |||
914 | BBPreds.insert(BBPN->getIncomingBlock(i)); | |||
915 | } else { | |||
916 | BBPreds.insert(pred_begin(BB), pred_end(BB)); | |||
917 | } | |||
918 | ||||
919 | // Walk the preds of DestBB. | |||
920 | for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) { | |||
921 | BasicBlock *Pred = DestBBPN->getIncomingBlock(i); | |||
922 | if (BBPreds.count(Pred)) { // Common predecessor? | |||
923 | for (const PHINode &PN : DestBB->phis()) { | |||
924 | const Value *V1 = PN.getIncomingValueForBlock(Pred); | |||
925 | const Value *V2 = PN.getIncomingValueForBlock(BB); | |||
926 | ||||
927 | // If V2 is a phi node in BB, look up what the mapped value will be. | |||
928 | if (const PHINode *V2PN = dyn_cast<PHINode>(V2)) | |||
929 | if (V2PN->getParent() == BB) | |||
930 | V2 = V2PN->getIncomingValueForBlock(Pred); | |||
931 | ||||
932 | // If there is a conflict, bail out. | |||
933 | if (V1 != V2) return false; | |||
934 | } | |||
935 | } | |||
936 | } | |||
937 | ||||
938 | return true; | |||
939 | } | |||
940 | ||||
941 | /// Eliminate a basic block that has only phi's and an unconditional branch in | |||
942 | /// it. | |||
943 | void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) { | |||
944 | BranchInst *BI = cast<BranchInst>(BB->getTerminator()); | |||
945 | BasicBlock *DestBB = BI->getSuccessor(0); | |||
946 | ||||
947 | LLVM_DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB; } } while (false) | |||
948 | << *BB << *DestBB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB; } } while (false); | |||
949 | ||||
950 | // If the destination block has a single pred, then this is a trivial edge, | |||
951 | // just collapse it. | |||
952 | if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) { | |||
953 | if (SinglePred != DestBB) { | |||
954 | assert(SinglePred == BB &&(static_cast <bool> (SinglePred == BB && "Single predecessor not the same as predecessor" ) ? void (0) : __assert_fail ("SinglePred == BB && \"Single predecessor not the same as predecessor\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 955, __extension__ __PRETTY_FUNCTION__ )) | |||
955 | "Single predecessor not the same as predecessor")(static_cast <bool> (SinglePred == BB && "Single predecessor not the same as predecessor" ) ? void (0) : __assert_fail ("SinglePred == BB && \"Single predecessor not the same as predecessor\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 955, __extension__ __PRETTY_FUNCTION__ )); | |||
956 | // Merge DestBB into SinglePred/BB and delete it. | |||
957 | MergeBlockIntoPredecessor(DestBB); | |||
958 | // Note: BB(=SinglePred) will not be deleted on this path. | |||
959 | // DestBB(=its single successor) is the one that was deleted. | |||
960 | LLVM_DEBUG(dbgs() << "AFTER:\n" << *SinglePred << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "AFTER:\n" << *SinglePred << "\n\n\n"; } } while (false); | |||
961 | return; | |||
962 | } | |||
963 | } | |||
964 | ||||
965 | // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB | |||
966 | // to handle the new incoming edges it is about to have. | |||
967 | for (PHINode &PN : DestBB->phis()) { | |||
968 | // Remove the incoming value for BB, and remember it. | |||
969 | Value *InVal = PN.removeIncomingValue(BB, false); | |||
970 | ||||
971 | // Two options: either the InVal is a phi node defined in BB or it is some | |||
972 | // value that dominates BB. | |||
973 | PHINode *InValPhi = dyn_cast<PHINode>(InVal); | |||
974 | if (InValPhi && InValPhi->getParent() == BB) { | |||
975 | // Add all of the input values of the input PHI as inputs of this phi. | |||
976 | for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i) | |||
977 | PN.addIncoming(InValPhi->getIncomingValue(i), | |||
978 | InValPhi->getIncomingBlock(i)); | |||
979 | } else { | |||
980 | // Otherwise, add one instance of the dominating value for each edge that | |||
981 | // we will be adding. | |||
982 | if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { | |||
983 | for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) | |||
984 | PN.addIncoming(InVal, BBPN->getIncomingBlock(i)); | |||
985 | } else { | |||
986 | for (BasicBlock *Pred : predecessors(BB)) | |||
987 | PN.addIncoming(InVal, Pred); | |||
988 | } | |||
989 | } | |||
990 | } | |||
991 | ||||
992 | // The PHIs are now updated, change everything that refers to BB to use | |||
993 | // DestBB and remove BB. | |||
994 | BB->replaceAllUsesWith(DestBB); | |||
995 | BB->eraseFromParent(); | |||
996 | ++NumBlocksElim; | |||
997 | ||||
998 | LLVM_DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "AFTER:\n" << *DestBB << "\n\n\n"; } } while (false); | |||
999 | } | |||
1000 | ||||
1001 | // Computes a map of base pointer relocation instructions to corresponding | |||
1002 | // derived pointer relocation instructions given a vector of all relocate calls | |||
1003 | static void computeBaseDerivedRelocateMap( | |||
1004 | const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls, | |||
1005 | DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> | |||
1006 | &RelocateInstMap) { | |||
1007 | // Collect information in two maps: one primarily for locating the base object | |||
1008 | // while filling the second map; the second map is the final structure holding | |||
1009 | // a mapping between Base and corresponding Derived relocate calls | |||
1010 | DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap; | |||
1011 | for (auto *ThisRelocate : AllRelocateCalls) { | |||
1012 | auto K = std::make_pair(ThisRelocate->getBasePtrIndex(), | |||
1013 | ThisRelocate->getDerivedPtrIndex()); | |||
1014 | RelocateIdxMap.insert(std::make_pair(K, ThisRelocate)); | |||
1015 | } | |||
1016 | for (auto &Item : RelocateIdxMap) { | |||
1017 | std::pair<unsigned, unsigned> Key = Item.first; | |||
1018 | if (Key.first == Key.second) | |||
1019 | // Base relocation: nothing to insert | |||
1020 | continue; | |||
1021 | ||||
1022 | GCRelocateInst *I = Item.second; | |||
1023 | auto BaseKey = std::make_pair(Key.first, Key.first); | |||
1024 | ||||
1025 | // We're iterating over RelocateIdxMap so we cannot modify it. | |||
1026 | auto MaybeBase = RelocateIdxMap.find(BaseKey); | |||
1027 | if (MaybeBase == RelocateIdxMap.end()) | |||
1028 | // TODO: We might want to insert a new base object relocate and gep off | |||
1029 | // that, if there are enough derived object relocates. | |||
1030 | continue; | |||
1031 | ||||
1032 | RelocateInstMap[MaybeBase->second].push_back(I); | |||
1033 | } | |||
1034 | } | |||
1035 | ||||
1036 | // Accepts a GEP and extracts the operands into a vector provided they're all | |||
1037 | // small integer constants | |||
1038 | static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP, | |||
1039 | SmallVectorImpl<Value *> &OffsetV) { | |||
1040 | for (unsigned i = 1; i < GEP->getNumOperands(); i++) { | |||
1041 | // Only accept small constant integer operands | |||
1042 | auto *Op = dyn_cast<ConstantInt>(GEP->getOperand(i)); | |||
1043 | if (!Op || Op->getZExtValue() > 20) | |||
1044 | return false; | |||
1045 | } | |||
1046 | ||||
1047 | for (unsigned i = 1; i < GEP->getNumOperands(); i++) | |||
1048 | OffsetV.push_back(GEP->getOperand(i)); | |||
1049 | return true; | |||
1050 | } | |||
1051 | ||||
1052 | // Takes a RelocatedBase (base pointer relocation instruction) and Targets to | |||
1053 | // replace, computes a replacement, and affects it. | |||
1054 | static bool | |||
1055 | simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase, | |||
1056 | const SmallVectorImpl<GCRelocateInst *> &Targets) { | |||
1057 | bool MadeChange = false; | |||
1058 | // We must ensure the relocation of derived pointer is defined after | |||
1059 | // relocation of base pointer. If we find a relocation corresponding to base | |||
1060 | // defined earlier than relocation of base then we move relocation of base | |||
1061 | // right before found relocation. We consider only relocation in the same | |||
1062 | // basic block as relocation of base. Relocations from other basic block will | |||
1063 | // be skipped by optimization and we do not care about them. | |||
1064 | for (auto R = RelocatedBase->getParent()->getFirstInsertionPt(); | |||
1065 | &*R != RelocatedBase; ++R) | |||
1066 | if (auto *RI = dyn_cast<GCRelocateInst>(R)) | |||
1067 | if (RI->getStatepoint() == RelocatedBase->getStatepoint()) | |||
1068 | if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) { | |||
1069 | RelocatedBase->moveBefore(RI); | |||
1070 | break; | |||
1071 | } | |||
1072 | ||||
1073 | for (GCRelocateInst *ToReplace : Targets) { | |||
1074 | assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&(static_cast <bool> (ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() && "Not relocating a derived object of the original base object" ) ? void (0) : __assert_fail ("ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() && \"Not relocating a derived object of the original base object\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1075, __extension__ __PRETTY_FUNCTION__ )) | |||
1075 | "Not relocating a derived object of the original base object")(static_cast <bool> (ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() && "Not relocating a derived object of the original base object" ) ? void (0) : __assert_fail ("ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() && \"Not relocating a derived object of the original base object\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1075, __extension__ __PRETTY_FUNCTION__ )); | |||
1076 | if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) { | |||
1077 | // A duplicate relocate call. TODO: coalesce duplicates. | |||
1078 | continue; | |||
1079 | } | |||
1080 | ||||
1081 | if (RelocatedBase->getParent() != ToReplace->getParent()) { | |||
1082 | // Base and derived relocates are in different basic blocks. | |||
1083 | // In this case transform is only valid when base dominates derived | |||
1084 | // relocate. However it would be too expensive to check dominance | |||
1085 | // for each such relocate, so we skip the whole transformation. | |||
1086 | continue; | |||
1087 | } | |||
1088 | ||||
1089 | Value *Base = ToReplace->getBasePtr(); | |||
1090 | auto *Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr()); | |||
1091 | if (!Derived
| |||
1092 | continue; | |||
1093 | ||||
1094 | SmallVector<Value *, 2> OffsetV; | |||
1095 | if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV)) | |||
1096 | continue; | |||
1097 | ||||
1098 | // Create a Builder and replace the target callsite with a gep | |||
1099 | assert(RelocatedBase->getNextNode() &&(static_cast <bool> (RelocatedBase->getNextNode() && "Should always have one since it's not a terminator") ? void (0) : __assert_fail ("RelocatedBase->getNextNode() && \"Should always have one since it's not a terminator\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1100, __extension__ __PRETTY_FUNCTION__ )) | |||
1100 | "Should always have one since it's not a terminator")(static_cast <bool> (RelocatedBase->getNextNode() && "Should always have one since it's not a terminator") ? void (0) : __assert_fail ("RelocatedBase->getNextNode() && \"Should always have one since it's not a terminator\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1100, __extension__ __PRETTY_FUNCTION__ )); | |||
1101 | ||||
1102 | // Insert after RelocatedBase | |||
1103 | IRBuilder<> Builder(RelocatedBase->getNextNode()); | |||
1104 | Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc()); | |||
1105 | ||||
1106 | // If gc_relocate does not match the actual type, cast it to the right type. | |||
1107 | // In theory, there must be a bitcast after gc_relocate if the type does not | |||
1108 | // match, and we should reuse it to get the derived pointer. But it could be | |||
1109 | // cases like this: | |||
1110 | // bb1: | |||
1111 | // ... | |||
1112 | // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...) | |||
1113 | // br label %merge | |||
1114 | // | |||
1115 | // bb2: | |||
1116 | // ... | |||
1117 | // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...) | |||
1118 | // br label %merge | |||
1119 | // | |||
1120 | // merge: | |||
1121 | // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ] | |||
1122 | // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)* | |||
1123 | // | |||
1124 | // In this case, we can not find the bitcast any more. So we insert a new bitcast | |||
1125 | // no matter there is already one or not. In this way, we can handle all cases, and | |||
1126 | // the extra bitcast should be optimized away in later passes. | |||
1127 | Value *ActualRelocatedBase = RelocatedBase; | |||
1128 | if (RelocatedBase->getType() != Base->getType()) { | |||
| ||||
1129 | ActualRelocatedBase = | |||
1130 | Builder.CreateBitCast(RelocatedBase, Base->getType()); | |||
1131 | } | |||
1132 | Value *Replacement = Builder.CreateGEP( | |||
1133 | Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV)); | |||
1134 | Replacement->takeName(ToReplace); | |||
1135 | // If the newly generated derived pointer's type does not match the original derived | |||
1136 | // pointer's type, cast the new derived pointer to match it. Same reasoning as above. | |||
1137 | Value *ActualReplacement = Replacement; | |||
1138 | if (Replacement->getType() != ToReplace->getType()) { | |||
1139 | ActualReplacement = | |||
1140 | Builder.CreateBitCast(Replacement, ToReplace->getType()); | |||
1141 | } | |||
1142 | ToReplace->replaceAllUsesWith(ActualReplacement); | |||
1143 | ToReplace->eraseFromParent(); | |||
1144 | ||||
1145 | MadeChange = true; | |||
1146 | } | |||
1147 | return MadeChange; | |||
1148 | } | |||
1149 | ||||
1150 | // Turns this: | |||
1151 | // | |||
1152 | // %base = ... | |||
1153 | // %ptr = gep %base + 15 | |||
1154 | // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr) | |||
1155 | // %base' = relocate(%tok, i32 4, i32 4) | |||
1156 | // %ptr' = relocate(%tok, i32 4, i32 5) | |||
1157 | // %val = load %ptr' | |||
1158 | // | |||
1159 | // into this: | |||
1160 | // | |||
1161 | // %base = ... | |||
1162 | // %ptr = gep %base + 15 | |||
1163 | // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr) | |||
1164 | // %base' = gc.relocate(%tok, i32 4, i32 4) | |||
1165 | // %ptr' = gep %base' + 15 | |||
1166 | // %val = load %ptr' | |||
1167 | bool CodeGenPrepare::simplifyOffsetableRelocate(GCStatepointInst &I) { | |||
1168 | bool MadeChange = false; | |||
1169 | SmallVector<GCRelocateInst *, 2> AllRelocateCalls; | |||
1170 | for (auto *U : I.users()) | |||
1171 | if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U)) | |||
1172 | // Collect all the relocate calls associated with a statepoint | |||
1173 | AllRelocateCalls.push_back(Relocate); | |||
1174 | ||||
1175 | // We need at least one base pointer relocation + one derived pointer | |||
1176 | // relocation to mangle | |||
1177 | if (AllRelocateCalls.size() < 2) | |||
| ||||
1178 | return false; | |||
1179 | ||||
1180 | // RelocateInstMap is a mapping from the base relocate instruction to the | |||
1181 | // corresponding derived relocate instructions | |||
1182 | DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap; | |||
1183 | computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap); | |||
1184 | if (RelocateInstMap.empty()) | |||
1185 | return false; | |||
1186 | ||||
1187 | for (auto &Item : RelocateInstMap) | |||
1188 | // Item.first is the RelocatedBase to offset against | |||
1189 | // Item.second is the vector of Targets to replace | |||
1190 | MadeChange = simplifyRelocatesOffABase(Item.first, Item.second); | |||
1191 | return MadeChange; | |||
1192 | } | |||
1193 | ||||
1194 | /// Sink the specified cast instruction into its user blocks. | |||
1195 | static bool SinkCast(CastInst *CI) { | |||
1196 | BasicBlock *DefBB = CI->getParent(); | |||
1197 | ||||
1198 | /// InsertedCasts - Only insert a cast in each block once. | |||
1199 | DenseMap<BasicBlock*, CastInst*> InsertedCasts; | |||
1200 | ||||
1201 | bool MadeChange = false; | |||
1202 | for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end(); | |||
1203 | UI != E; ) { | |||
1204 | Use &TheUse = UI.getUse(); | |||
1205 | Instruction *User = cast<Instruction>(*UI); | |||
1206 | ||||
1207 | // Figure out which BB this cast is used in. For PHI's this is the | |||
1208 | // appropriate predecessor block. | |||
1209 | BasicBlock *UserBB = User->getParent(); | |||
1210 | if (PHINode *PN = dyn_cast<PHINode>(User)) { | |||
1211 | UserBB = PN->getIncomingBlock(TheUse); | |||
1212 | } | |||
1213 | ||||
1214 | // Preincrement use iterator so we don't invalidate it. | |||
1215 | ++UI; | |||
1216 | ||||
1217 | // The first insertion point of a block containing an EH pad is after the | |||
1218 | // pad. If the pad is the user, we cannot sink the cast past the pad. | |||
1219 | if (User->isEHPad()) | |||
1220 | continue; | |||
1221 | ||||
1222 | // If the block selected to receive the cast is an EH pad that does not | |||
1223 | // allow non-PHI instructions before the terminator, we can't sink the | |||
1224 | // cast. | |||
1225 | if (UserBB->getTerminator()->isEHPad()) | |||
1226 | continue; | |||
1227 | ||||
1228 | // If this user is in the same block as the cast, don't change the cast. | |||
1229 | if (UserBB == DefBB) continue; | |||
1230 | ||||
1231 | // If we have already inserted a cast into this block, use it. | |||
1232 | CastInst *&InsertedCast = InsertedCasts[UserBB]; | |||
1233 | ||||
1234 | if (!InsertedCast) { | |||
1235 | BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); | |||
1236 | assert(InsertPt != UserBB->end())(static_cast <bool> (InsertPt != UserBB->end()) ? void (0) : __assert_fail ("InsertPt != UserBB->end()", "llvm/lib/CodeGen/CodeGenPrepare.cpp" , 1236, __extension__ __PRETTY_FUNCTION__)); | |||
1237 | InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0), | |||
1238 | CI->getType(), "", &*InsertPt); | |||
1239 | InsertedCast->setDebugLoc(CI->getDebugLoc()); | |||
1240 | } | |||
1241 | ||||
1242 | // Replace a use of the cast with a use of the new cast. | |||
1243 | TheUse = InsertedCast; | |||
1244 | MadeChange = true; | |||
1245 | ++NumCastUses; | |||
1246 | } | |||
1247 | ||||
1248 | // If we removed all uses, nuke the cast. | |||
1249 | if (CI->use_empty()) { | |||
1250 | salvageDebugInfo(*CI); | |||
1251 | CI->eraseFromParent(); | |||
1252 | MadeChange = true; | |||
1253 | } | |||
1254 | ||||
1255 | return MadeChange; | |||
1256 | } | |||
1257 | ||||
1258 | /// If the specified cast instruction is a noop copy (e.g. it's casting from | |||
1259 | /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to | |||
1260 | /// reduce the number of virtual registers that must be created and coalesced. | |||
1261 | /// | |||
1262 | /// Return true if any changes are made. | |||
1263 | static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI, | |||
1264 | const DataLayout &DL) { | |||
1265 | // Sink only "cheap" (or nop) address-space casts. This is a weaker condition | |||
1266 | // than sinking only nop casts, but is helpful on some platforms. | |||
1267 | if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) { | |||
1268 | if (!TLI.isFreeAddrSpaceCast(ASC->getSrcAddressSpace(), | |||
1269 | ASC->getDestAddressSpace())) | |||
1270 | return false; | |||
1271 | } | |||
1272 | ||||
1273 | // If this is a noop copy, | |||
1274 | EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType()); | |||
1275 | EVT DstVT = TLI.getValueType(DL, CI->getType()); | |||
1276 | ||||
1277 | // This is an fp<->int conversion? | |||
1278 | if (SrcVT.isInteger() != DstVT.isInteger()) | |||
1279 | return false; | |||
1280 | ||||
1281 | // If this is an extension, it will be a zero or sign extension, which | |||
1282 | // isn't a noop. | |||
1283 | if (SrcVT.bitsLT(DstVT)) return false; | |||
1284 | ||||
1285 | // If these values will be promoted, find out what they will be promoted | |||
1286 | // to. This helps us consider truncates on PPC as noop copies when they | |||
1287 | // are. | |||
1288 | if (TLI.getTypeAction(CI->getContext(), SrcVT) == | |||
1289 | TargetLowering::TypePromoteInteger) | |||
1290 | SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT); | |||
1291 | if (TLI.getTypeAction(CI->getContext(), DstVT) == | |||
1292 | TargetLowering::TypePromoteInteger) | |||
1293 | DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT); | |||
1294 | ||||
1295 | // If, after promotion, these are the same types, this is a noop copy. | |||
1296 | if (SrcVT != DstVT) | |||
1297 | return false; | |||
1298 | ||||
1299 | return SinkCast(CI); | |||
1300 | } | |||
1301 | ||||
1302 | // Match a simple increment by constant operation. Note that if a sub is | |||
1303 | // matched, the step is negated (as if the step had been canonicalized to | |||
1304 | // an add, even though we leave the instruction alone.) | |||
1305 | bool matchIncrement(const Instruction* IVInc, Instruction *&LHS, | |||
1306 | Constant *&Step) { | |||
1307 | if (match(IVInc, m_Add(m_Instruction(LHS), m_Constant(Step))) || | |||
1308 | match(IVInc, m_ExtractValue<0>(m_Intrinsic<Intrinsic::uadd_with_overflow>( | |||
1309 | m_Instruction(LHS), m_Constant(Step))))) | |||
1310 | return true; | |||
1311 | if (match(IVInc, m_Sub(m_Instruction(LHS), m_Constant(Step))) || | |||
1312 | match(IVInc, m_ExtractValue<0>(m_Intrinsic<Intrinsic::usub_with_overflow>( | |||
1313 | m_Instruction(LHS), m_Constant(Step))))) { | |||
1314 | Step = ConstantExpr::getNeg(Step); | |||
1315 | return true; | |||
1316 | } | |||
1317 | return false; | |||
1318 | } | |||
1319 | ||||
1320 | /// If given \p PN is an inductive variable with value IVInc coming from the | |||
1321 | /// backedge, and on each iteration it gets increased by Step, return pair | |||
1322 | /// <IVInc, Step>. Otherwise, return None. | |||
1323 | static Optional<std::pair<Instruction *, Constant *> > | |||
1324 | getIVIncrement(const PHINode *PN, const LoopInfo *LI) { | |||
1325 | const Loop *L = LI->getLoopFor(PN->getParent()); | |||
1326 | if (!L || L->getHeader() != PN->getParent() || !L->getLoopLatch()) | |||
1327 | return None; | |||
1328 | auto *IVInc = | |||
1329 | dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch())); | |||
1330 | if (!IVInc || LI->getLoopFor(IVInc->getParent()) != L) | |||
1331 | return None; | |||
1332 | Instruction *LHS = nullptr; | |||
1333 | Constant *Step = nullptr; | |||
1334 | if (matchIncrement(IVInc, LHS, Step) && LHS == PN) | |||
1335 | return std::make_pair(IVInc, Step); | |||
1336 | return None; | |||
1337 | } | |||
1338 | ||||
1339 | static bool isIVIncrement(const Value *V, const LoopInfo *LI) { | |||
1340 | auto *I = dyn_cast<Instruction>(V); | |||
1341 | if (!I) | |||
1342 | return false; | |||
1343 | Instruction *LHS = nullptr; | |||
1344 | Constant *Step = nullptr; | |||
1345 | if (!matchIncrement(I, LHS, Step)) | |||
1346 | return false; | |||
1347 | if (auto *PN = dyn_cast<PHINode>(LHS)) | |||
1348 | if (auto IVInc = getIVIncrement(PN, LI)) | |||
1349 | return IVInc->first == I; | |||
1350 | return false; | |||
1351 | } | |||
1352 | ||||
1353 | bool CodeGenPrepare::replaceMathCmpWithIntrinsic(BinaryOperator *BO, | |||
1354 | Value *Arg0, Value *Arg1, | |||
1355 | CmpInst *Cmp, | |||
1356 | Intrinsic::ID IID) { | |||
1357 | auto IsReplacableIVIncrement = [this, &Cmp](BinaryOperator *BO) { | |||
1358 | if (!isIVIncrement(BO, LI)) | |||
1359 | return false; | |||
1360 | const Loop *L = LI->getLoopFor(BO->getParent()); | |||
1361 | assert(L && "L should not be null after isIVIncrement()")(static_cast <bool> (L && "L should not be null after isIVIncrement()" ) ? void (0) : __assert_fail ("L && \"L should not be null after isIVIncrement()\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1361, __extension__ __PRETTY_FUNCTION__ )); | |||
1362 | // Do not risk on moving increment into a child loop. | |||
1363 | if (LI->getLoopFor(Cmp->getParent()) != L) | |||
1364 | return false; | |||
1365 | ||||
1366 | // Finally, we need to ensure that the insert point will dominate all | |||
1367 | // existing uses of the increment. | |||
1368 | ||||
1369 | auto &DT = getDT(*BO->getParent()->getParent()); | |||
1370 | if (DT.dominates(Cmp->getParent(), BO->getParent())) | |||
1371 | // If we're moving up the dom tree, all uses are trivially dominated. | |||
1372 | // (This is the common case for code produced by LSR.) | |||
1373 | return true; | |||
1374 | ||||
1375 | // Otherwise, special case the single use in the phi recurrence. | |||
1376 | return BO->hasOneUse() && DT.dominates(Cmp->getParent(), L->getLoopLatch()); | |||
1377 | }; | |||
1378 | if (BO->getParent() != Cmp->getParent() && !IsReplacableIVIncrement(BO)) { | |||
1379 | // We used to use a dominator tree here to allow multi-block optimization. | |||
1380 | // But that was problematic because: | |||
1381 | // 1. It could cause a perf regression by hoisting the math op into the | |||
1382 | // critical path. | |||
1383 | // 2. It could cause a perf regression by creating a value that was live | |||
1384 | // across multiple blocks and increasing register pressure. | |||
1385 | // 3. Use of a dominator tree could cause large compile-time regression. | |||
1386 | // This is because we recompute the DT on every change in the main CGP | |||
1387 | // run-loop. The recomputing is probably unnecessary in many cases, so if | |||
1388 | // that was fixed, using a DT here would be ok. | |||
1389 | // | |||
1390 | // There is one important particular case we still want to handle: if BO is | |||
1391 | // the IV increment. Important properties that make it profitable: | |||
1392 | // - We can speculate IV increment anywhere in the loop (as long as the | |||
1393 | // indvar Phi is its only user); | |||
1394 | // - Upon computing Cmp, we effectively compute something equivalent to the | |||
1395 | // IV increment (despite it loops differently in the IR). So moving it up | |||
1396 | // to the cmp point does not really increase register pressure. | |||
1397 | return false; | |||
1398 | } | |||
1399 | ||||
1400 | // We allow matching the canonical IR (add X, C) back to (usubo X, -C). | |||
1401 | if (BO->getOpcode() == Instruction::Add && | |||
1402 | IID == Intrinsic::usub_with_overflow) { | |||
1403 | assert(isa<Constant>(Arg1) && "Unexpected input for usubo")(static_cast <bool> (isa<Constant>(Arg1) && "Unexpected input for usubo") ? void (0) : __assert_fail ("isa<Constant>(Arg1) && \"Unexpected input for usubo\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1403, __extension__ __PRETTY_FUNCTION__ )); | |||
1404 | Arg1 = ConstantExpr::getNeg(cast<Constant>(Arg1)); | |||
1405 | } | |||
1406 | ||||
1407 | // Insert at the first instruction of the pair. | |||
1408 | Instruction *InsertPt = nullptr; | |||
1409 | for (Instruction &Iter : *Cmp->getParent()) { | |||
1410 | // If BO is an XOR, it is not guaranteed that it comes after both inputs to | |||
1411 | // the overflow intrinsic are defined. | |||
1412 | if ((BO->getOpcode() != Instruction::Xor && &Iter == BO) || &Iter == Cmp) { | |||
1413 | InsertPt = &Iter; | |||
1414 | break; | |||
1415 | } | |||
1416 | } | |||
1417 | assert(InsertPt != nullptr && "Parent block did not contain cmp or binop")(static_cast <bool> (InsertPt != nullptr && "Parent block did not contain cmp or binop" ) ? void (0) : __assert_fail ("InsertPt != nullptr && \"Parent block did not contain cmp or binop\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1417, __extension__ __PRETTY_FUNCTION__ )); | |||
1418 | ||||
1419 | IRBuilder<> Builder(InsertPt); | |||
1420 | Value *MathOV = Builder.CreateBinaryIntrinsic(IID, Arg0, Arg1); | |||
1421 | if (BO->getOpcode() != Instruction::Xor) { | |||
1422 | Value *Math = Builder.CreateExtractValue(MathOV, 0, "math"); | |||
1423 | BO->replaceAllUsesWith(Math); | |||
1424 | } else | |||
1425 | assert(BO->hasOneUse() &&(static_cast <bool> (BO->hasOneUse() && "Patterns with XOr should use the BO only in the compare" ) ? void (0) : __assert_fail ("BO->hasOneUse() && \"Patterns with XOr should use the BO only in the compare\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1426, __extension__ __PRETTY_FUNCTION__ )) | |||
1426 | "Patterns with XOr should use the BO only in the compare")(static_cast <bool> (BO->hasOneUse() && "Patterns with XOr should use the BO only in the compare" ) ? void (0) : __assert_fail ("BO->hasOneUse() && \"Patterns with XOr should use the BO only in the compare\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1426, __extension__ __PRETTY_FUNCTION__ )); | |||
1427 | Value *OV = Builder.CreateExtractValue(MathOV, 1, "ov"); | |||
1428 | Cmp->replaceAllUsesWith(OV); | |||
1429 | Cmp->eraseFromParent(); | |||
1430 | BO->eraseFromParent(); | |||
1431 | return true; | |||
1432 | } | |||
1433 | ||||
1434 | /// Match special-case patterns that check for unsigned add overflow. | |||
1435 | static bool matchUAddWithOverflowConstantEdgeCases(CmpInst *Cmp, | |||
1436 | BinaryOperator *&Add) { | |||
1437 | // Add = add A, 1; Cmp = icmp eq A,-1 (overflow if A is max val) | |||
1438 | // Add = add A,-1; Cmp = icmp ne A, 0 (overflow if A is non-zero) | |||
1439 | Value *A = Cmp->getOperand(0), *B = Cmp->getOperand(1); | |||
1440 | ||||
1441 | // We are not expecting non-canonical/degenerate code. Just bail out. | |||
1442 | if (isa<Constant>(A)) | |||
1443 | return false; | |||
1444 | ||||
1445 | ICmpInst::Predicate Pred = Cmp->getPredicate(); | |||
1446 | if (Pred == ICmpInst::ICMP_EQ && match(B, m_AllOnes())) | |||
1447 | B = ConstantInt::get(B->getType(), 1); | |||
1448 | else if (Pred == ICmpInst::ICMP_NE && match(B, m_ZeroInt())) | |||
1449 | B = ConstantInt::get(B->getType(), -1); | |||
1450 | else | |||
1451 | return false; | |||
1452 | ||||
1453 | // Check the users of the variable operand of the compare looking for an add | |||
1454 | // with the adjusted constant. | |||
1455 | for (User *U : A->users()) { | |||
1456 | if (match(U, m_Add(m_Specific(A), m_Specific(B)))) { | |||
1457 | Add = cast<BinaryOperator>(U); | |||
1458 | return true; | |||
1459 | } | |||
1460 | } | |||
1461 | return false; | |||
1462 | } | |||
1463 | ||||
1464 | /// Try to combine the compare into a call to the llvm.uadd.with.overflow | |||
1465 | /// intrinsic. Return true if any changes were made. | |||
1466 | bool CodeGenPrepare::combineToUAddWithOverflow(CmpInst *Cmp, | |||
1467 | bool &ModifiedDT) { | |||
1468 | Value *A, *B; | |||
1469 | BinaryOperator *Add; | |||
1470 | if (!match(Cmp, m_UAddWithOverflow(m_Value(A), m_Value(B), m_BinOp(Add)))) { | |||
1471 | if (!matchUAddWithOverflowConstantEdgeCases(Cmp, Add)) | |||
1472 | return false; | |||
1473 | // Set A and B in case we match matchUAddWithOverflowConstantEdgeCases. | |||
1474 | A = Add->getOperand(0); | |||
1475 | B = Add->getOperand(1); | |||
1476 | } | |||
1477 | ||||
1478 | if (!TLI->shouldFormOverflowOp(ISD::UADDO, | |||
1479 | TLI->getValueType(*DL, Add->getType()), | |||
1480 | Add->hasNUsesOrMore(2))) | |||
1481 | return false; | |||
1482 | ||||
1483 | // We don't want to move around uses of condition values this late, so we | |||
1484 | // check if it is legal to create the call to the intrinsic in the basic | |||
1485 | // block containing the icmp. | |||
1486 | if (Add->getParent() != Cmp->getParent() && !Add->hasOneUse()) | |||
1487 | return false; | |||
1488 | ||||
1489 | if (!replaceMathCmpWithIntrinsic(Add, A, B, Cmp, | |||
1490 | Intrinsic::uadd_with_overflow)) | |||
1491 | return false; | |||
1492 | ||||
1493 | // Reset callers - do not crash by iterating over a dead instruction. | |||
1494 | ModifiedDT = true; | |||
1495 | return true; | |||
1496 | } | |||
1497 | ||||
1498 | bool CodeGenPrepare::combineToUSubWithOverflow(CmpInst *Cmp, | |||
1499 | bool &ModifiedDT) { | |||
1500 | // We are not expecting non-canonical/degenerate code. Just bail out. | |||
1501 | Value *A = Cmp->getOperand(0), *B = Cmp->getOperand(1); | |||
1502 | if (isa<Constant>(A) && isa<Constant>(B)) | |||
1503 | return false; | |||
1504 | ||||
1505 | // Convert (A u> B) to (A u< B) to simplify pattern matching. | |||
1506 | ICmpInst::Predicate Pred = Cmp->getPredicate(); | |||
1507 | if (Pred == ICmpInst::ICMP_UGT) { | |||
1508 | std::swap(A, B); | |||
1509 | Pred = ICmpInst::ICMP_ULT; | |||
1510 | } | |||
1511 | // Convert special-case: (A == 0) is the same as (A u< 1). | |||
1512 | if (Pred == ICmpInst::ICMP_EQ && match(B, m_ZeroInt())) { | |||
1513 | B = ConstantInt::get(B->getType(), 1); | |||
1514 | Pred = ICmpInst::ICMP_ULT; | |||
1515 | } | |||
1516 | // Convert special-case: (A != 0) is the same as (0 u< A). | |||
1517 | if (Pred == ICmpInst::ICMP_NE && match(B, m_ZeroInt())) { | |||
1518 | std::swap(A, B); | |||
1519 | Pred = ICmpInst::ICMP_ULT; | |||
1520 | } | |||
1521 | if (Pred != ICmpInst::ICMP_ULT) | |||
1522 | return false; | |||
1523 | ||||
1524 | // Walk the users of a variable operand of a compare looking for a subtract or | |||
1525 | // add with that same operand. Also match the 2nd operand of the compare to | |||
1526 | // the add/sub, but that may be a negated constant operand of an add. | |||
1527 | Value *CmpVariableOperand = isa<Constant>(A) ? B : A; | |||
1528 | BinaryOperator *Sub = nullptr; | |||
1529 | for (User *U : CmpVariableOperand->users()) { | |||
1530 | // A - B, A u< B --> usubo(A, B) | |||
1531 | if (match(U, m_Sub(m_Specific(A), m_Specific(B)))) { | |||
1532 | Sub = cast<BinaryOperator>(U); | |||
1533 | break; | |||
1534 | } | |||
1535 | ||||
1536 | // A + (-C), A u< C (canonicalized form of (sub A, C)) | |||
1537 | const APInt *CmpC, *AddC; | |||
1538 | if (match(U, m_Add(m_Specific(A), m_APInt(AddC))) && | |||
1539 | match(B, m_APInt(CmpC)) && *AddC == -(*CmpC)) { | |||
1540 | Sub = cast<BinaryOperator>(U); | |||
1541 | break; | |||
1542 | } | |||
1543 | } | |||
1544 | if (!Sub) | |||
1545 | return false; | |||
1546 | ||||
1547 | if (!TLI->shouldFormOverflowOp(ISD::USUBO, | |||
1548 | TLI->getValueType(*DL, Sub->getType()), | |||
1549 | Sub->hasNUsesOrMore(2))) | |||
1550 | return false; | |||
1551 | ||||
1552 | if (!replaceMathCmpWithIntrinsic(Sub, Sub->getOperand(0), Sub->getOperand(1), | |||
1553 | Cmp, Intrinsic::usub_with_overflow)) | |||
1554 | return false; | |||
1555 | ||||
1556 | // Reset callers - do not crash by iterating over a dead instruction. | |||
1557 | ModifiedDT = true; | |||
1558 | return true; | |||
1559 | } | |||
1560 | ||||
1561 | /// Sink the given CmpInst into user blocks to reduce the number of virtual | |||
1562 | /// registers that must be created and coalesced. This is a clear win except on | |||
1563 | /// targets with multiple condition code registers (PowerPC), where it might | |||
1564 | /// lose; some adjustment may be wanted there. | |||
1565 | /// | |||
1566 | /// Return true if any changes are made. | |||
1567 | static bool sinkCmpExpression(CmpInst *Cmp, const TargetLowering &TLI) { | |||
1568 | if (TLI.hasMultipleConditionRegisters()) | |||
1569 | return false; | |||
1570 | ||||
1571 | // Avoid sinking soft-FP comparisons, since this can move them into a loop. | |||
1572 | if (TLI.useSoftFloat() && isa<FCmpInst>(Cmp)) | |||
1573 | return false; | |||
1574 | ||||
1575 | // Only insert a cmp in each block once. | |||
1576 | DenseMap<BasicBlock*, CmpInst*> InsertedCmps; | |||
1577 | ||||
1578 | bool MadeChange = false; | |||
1579 | for (Value::user_iterator UI = Cmp->user_begin(), E = Cmp->user_end(); | |||
1580 | UI != E; ) { | |||
1581 | Use &TheUse = UI.getUse(); | |||
1582 | Instruction *User = cast<Instruction>(*UI); | |||
1583 | ||||
1584 | // Preincrement use iterator so we don't invalidate it. | |||
1585 | ++UI; | |||
1586 | ||||
1587 | // Don't bother for PHI nodes. | |||
1588 | if (isa<PHINode>(User)) | |||
1589 | continue; | |||
1590 | ||||
1591 | // Figure out which BB this cmp is used in. | |||
1592 | BasicBlock *UserBB = User->getParent(); | |||
1593 | BasicBlock *DefBB = Cmp->getParent(); | |||
1594 | ||||
1595 | // If this user is in the same block as the cmp, don't change the cmp. | |||
1596 | if (UserBB == DefBB) continue; | |||
1597 | ||||
1598 | // If we have already inserted a cmp into this block, use it. | |||
1599 | CmpInst *&InsertedCmp = InsertedCmps[UserBB]; | |||
1600 | ||||
1601 | if (!InsertedCmp) { | |||
1602 | BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); | |||
1603 | assert(InsertPt != UserBB->end())(static_cast <bool> (InsertPt != UserBB->end()) ? void (0) : __assert_fail ("InsertPt != UserBB->end()", "llvm/lib/CodeGen/CodeGenPrepare.cpp" , 1603, __extension__ __PRETTY_FUNCTION__)); | |||
1604 | InsertedCmp = | |||
1605 | CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), | |||
1606 | Cmp->getOperand(0), Cmp->getOperand(1), "", | |||
1607 | &*InsertPt); | |||
1608 | // Propagate the debug info. | |||
1609 | InsertedCmp->setDebugLoc(Cmp->getDebugLoc()); | |||
1610 | } | |||
1611 | ||||
1612 | // Replace a use of the cmp with a use of the new cmp. | |||
1613 | TheUse = InsertedCmp; | |||
1614 | MadeChange = true; | |||
1615 | ++NumCmpUses; | |||
1616 | } | |||
1617 | ||||
1618 | // If we removed all uses, nuke the cmp. | |||
1619 | if (Cmp->use_empty()) { | |||
1620 | Cmp->eraseFromParent(); | |||
1621 | MadeChange = true; | |||
1622 | } | |||
1623 | ||||
1624 | return MadeChange; | |||
1625 | } | |||
1626 | ||||
1627 | /// For pattern like: | |||
1628 | /// | |||
1629 | /// DomCond = icmp sgt/slt CmpOp0, CmpOp1 (might not be in DomBB) | |||
1630 | /// ... | |||
1631 | /// DomBB: | |||
1632 | /// ... | |||
1633 | /// br DomCond, TrueBB, CmpBB | |||
1634 | /// CmpBB: (with DomBB being the single predecessor) | |||
1635 | /// ... | |||
1636 | /// Cmp = icmp eq CmpOp0, CmpOp1 | |||
1637 | /// ... | |||
1638 | /// | |||
1639 | /// It would use two comparison on targets that lowering of icmp sgt/slt is | |||
1640 | /// different from lowering of icmp eq (PowerPC). This function try to convert | |||
1641 | /// 'Cmp = icmp eq CmpOp0, CmpOp1' to ' Cmp = icmp slt/sgt CmpOp0, CmpOp1'. | |||
1642 | /// After that, DomCond and Cmp can use the same comparison so reduce one | |||
1643 | /// comparison. | |||
1644 | /// | |||
1645 | /// Return true if any changes are made. | |||
1646 | static bool foldICmpWithDominatingICmp(CmpInst *Cmp, | |||
1647 | const TargetLowering &TLI) { | |||
1648 | if (!EnableICMP_EQToICMP_ST && TLI.isEqualityCmpFoldedWithSignedCmp()) | |||
1649 | return false; | |||
1650 | ||||
1651 | ICmpInst::Predicate Pred = Cmp->getPredicate(); | |||
1652 | if (Pred != ICmpInst::ICMP_EQ) | |||
1653 | return false; | |||
1654 | ||||
1655 | // If icmp eq has users other than BranchInst and SelectInst, converting it to | |||
1656 | // icmp slt/sgt would introduce more redundant LLVM IR. | |||
1657 | for (User *U : Cmp->users()) { | |||
1658 | if (isa<BranchInst>(U)) | |||
1659 | continue; | |||
1660 | if (isa<SelectInst>(U) && cast<SelectInst>(U)->getCondition() == Cmp) | |||
1661 | continue; | |||
1662 | return false; | |||
1663 | } | |||
1664 | ||||
1665 | // This is a cheap/incomplete check for dominance - just match a single | |||
1666 | // predecessor with a conditional branch. | |||
1667 | BasicBlock *CmpBB = Cmp->getParent(); | |||
1668 | BasicBlock *DomBB = CmpBB->getSinglePredecessor(); | |||
1669 | if (!DomBB) | |||
1670 | return false; | |||
1671 | ||||
1672 | // We want to ensure that the only way control gets to the comparison of | |||
1673 | // interest is that a less/greater than comparison on the same operands is | |||
1674 | // false. | |||
1675 | Value *DomCond; | |||
1676 | BasicBlock *TrueBB, *FalseBB; | |||
1677 | if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB))) | |||
1678 | return false; | |||
1679 | if (CmpBB != FalseBB) | |||
1680 | return false; | |||
1681 | ||||
1682 | Value *CmpOp0 = Cmp->getOperand(0), *CmpOp1 = Cmp->getOperand(1); | |||
1683 | ICmpInst::Predicate DomPred; | |||
1684 | if (!match(DomCond, m_ICmp(DomPred, m_Specific(CmpOp0), m_Specific(CmpOp1)))) | |||
1685 | return false; | |||
1686 | if (DomPred != ICmpInst::ICMP_SGT && DomPred != ICmpInst::ICMP_SLT) | |||
1687 | return false; | |||
1688 | ||||
1689 | // Convert the equality comparison to the opposite of the dominating | |||
1690 | // comparison and swap the direction for all branch/select users. | |||
1691 | // We have conceptually converted: | |||
1692 | // Res = (a < b) ? <LT_RES> : (a == b) ? <EQ_RES> : <GT_RES>; | |||
1693 | // to | |||
1694 | // Res = (a < b) ? <LT_RES> : (a > b) ? <GT_RES> : <EQ_RES>; | |||
1695 | // And similarly for branches. | |||
1696 | for (User *U : Cmp->users()) { | |||
1697 | if (auto *BI = dyn_cast<BranchInst>(U)) { | |||
1698 | assert(BI->isConditional() && "Must be conditional")(static_cast <bool> (BI->isConditional() && "Must be conditional" ) ? void (0) : __assert_fail ("BI->isConditional() && \"Must be conditional\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1698, __extension__ __PRETTY_FUNCTION__ )); | |||
1699 | BI->swapSuccessors(); | |||
1700 | continue; | |||
1701 | } | |||
1702 | if (auto *SI = dyn_cast<SelectInst>(U)) { | |||
1703 | // Swap operands | |||
1704 | SI->swapValues(); | |||
1705 | SI->swapProfMetadata(); | |||
1706 | continue; | |||
1707 | } | |||
1708 | llvm_unreachable("Must be a branch or a select")::llvm::llvm_unreachable_internal("Must be a branch or a select" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1708); | |||
1709 | } | |||
1710 | Cmp->setPredicate(CmpInst::getSwappedPredicate(DomPred)); | |||
1711 | return true; | |||
1712 | } | |||
1713 | ||||
1714 | bool CodeGenPrepare::optimizeCmp(CmpInst *Cmp, bool &ModifiedDT) { | |||
1715 | if (sinkCmpExpression(Cmp, *TLI)) | |||
1716 | return true; | |||
1717 | ||||
1718 | if (combineToUAddWithOverflow(Cmp, ModifiedDT)) | |||
1719 | return true; | |||
1720 | ||||
1721 | if (combineToUSubWithOverflow(Cmp, ModifiedDT)) | |||
1722 | return true; | |||
1723 | ||||
1724 | if (foldICmpWithDominatingICmp(Cmp, *TLI)) | |||
1725 | return true; | |||
1726 | ||||
1727 | return false; | |||
1728 | } | |||
1729 | ||||
1730 | /// Duplicate and sink the given 'and' instruction into user blocks where it is | |||
1731 | /// used in a compare to allow isel to generate better code for targets where | |||
1732 | /// this operation can be combined. | |||
1733 | /// | |||
1734 | /// Return true if any changes are made. | |||
1735 | static bool sinkAndCmp0Expression(Instruction *AndI, | |||
1736 | const TargetLowering &TLI, | |||
1737 | SetOfInstrs &InsertedInsts) { | |||
1738 | // Double-check that we're not trying to optimize an instruction that was | |||
1739 | // already optimized by some other part of this pass. | |||
1740 | assert(!InsertedInsts.count(AndI) &&(static_cast <bool> (!InsertedInsts.count(AndI) && "Attempting to optimize already optimized and instruction") ? void (0) : __assert_fail ("!InsertedInsts.count(AndI) && \"Attempting to optimize already optimized and instruction\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1741, __extension__ __PRETTY_FUNCTION__ )) | |||
1741 | "Attempting to optimize already optimized and instruction")(static_cast <bool> (!InsertedInsts.count(AndI) && "Attempting to optimize already optimized and instruction") ? void (0) : __assert_fail ("!InsertedInsts.count(AndI) && \"Attempting to optimize already optimized and instruction\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1741, __extension__ __PRETTY_FUNCTION__ )); | |||
1742 | (void) InsertedInsts; | |||
1743 | ||||
1744 | // Nothing to do for single use in same basic block. | |||
1745 | if (AndI->hasOneUse() && | |||
1746 | AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent()) | |||
1747 | return false; | |||
1748 | ||||
1749 | // Try to avoid cases where sinking/duplicating is likely to increase register | |||
1750 | // pressure. | |||
1751 | if (!isa<ConstantInt>(AndI->getOperand(0)) && | |||
1752 | !isa<ConstantInt>(AndI->getOperand(1)) && | |||
1753 | AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse()) | |||
1754 | return false; | |||
1755 | ||||
1756 | for (auto *U : AndI->users()) { | |||
1757 | Instruction *User = cast<Instruction>(U); | |||
1758 | ||||
1759 | // Only sink 'and' feeding icmp with 0. | |||
1760 | if (!isa<ICmpInst>(User)) | |||
1761 | return false; | |||
1762 | ||||
1763 | auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1)); | |||
1764 | if (!CmpC || !CmpC->isZero()) | |||
1765 | return false; | |||
1766 | } | |||
1767 | ||||
1768 | if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI)) | |||
1769 | return false; | |||
1770 | ||||
1771 | LLVM_DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "found 'and' feeding only icmp 0;\n" ; } } while (false); | |||
1772 | LLVM_DEBUG(AndI->getParent()->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { AndI->getParent()->dump(); } } while (false); | |||
1773 | ||||
1774 | // Push the 'and' into the same block as the icmp 0. There should only be | |||
1775 | // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any | |||
1776 | // others, so we don't need to keep track of which BBs we insert into. | |||
1777 | for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end(); | |||
1778 | UI != E; ) { | |||
1779 | Use &TheUse = UI.getUse(); | |||
1780 | Instruction *User = cast<Instruction>(*UI); | |||
1781 | ||||
1782 | // Preincrement use iterator so we don't invalidate it. | |||
1783 | ++UI; | |||
1784 | ||||
1785 | LLVM_DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "sinking 'and' use: " << *User << "\n"; } } while (false); | |||
1786 | ||||
1787 | // Keep the 'and' in the same place if the use is already in the same block. | |||
1788 | Instruction *InsertPt = | |||
1789 | User->getParent() == AndI->getParent() ? AndI : User; | |||
1790 | Instruction *InsertedAnd = | |||
1791 | BinaryOperator::Create(Instruction::And, AndI->getOperand(0), | |||
1792 | AndI->getOperand(1), "", InsertPt); | |||
1793 | // Propagate the debug info. | |||
1794 | InsertedAnd->setDebugLoc(AndI->getDebugLoc()); | |||
1795 | ||||
1796 | // Replace a use of the 'and' with a use of the new 'and'. | |||
1797 | TheUse = InsertedAnd; | |||
1798 | ++NumAndUses; | |||
1799 | LLVM_DEBUG(User->getParent()->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { User->getParent()->dump(); } } while (false); | |||
1800 | } | |||
1801 | ||||
1802 | // We removed all uses, nuke the and. | |||
1803 | AndI->eraseFromParent(); | |||
1804 | return true; | |||
1805 | } | |||
1806 | ||||
1807 | /// Check if the candidates could be combined with a shift instruction, which | |||
1808 | /// includes: | |||
1809 | /// 1. Truncate instruction | |||
1810 | /// 2. And instruction and the imm is a mask of the low bits: | |||
1811 | /// imm & (imm+1) == 0 | |||
1812 | static bool isExtractBitsCandidateUse(Instruction *User) { | |||
1813 | if (!isa<TruncInst>(User)) { | |||
1814 | if (User->getOpcode() != Instruction::And || | |||
1815 | !isa<ConstantInt>(User->getOperand(1))) | |||
1816 | return false; | |||
1817 | ||||
1818 | const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue(); | |||
1819 | ||||
1820 | if ((Cimm & (Cimm + 1)).getBoolValue()) | |||
1821 | return false; | |||
1822 | } | |||
1823 | return true; | |||
1824 | } | |||
1825 | ||||
1826 | /// Sink both shift and truncate instruction to the use of truncate's BB. | |||
1827 | static bool | |||
1828 | SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI, | |||
1829 | DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts, | |||
1830 | const TargetLowering &TLI, const DataLayout &DL) { | |||
1831 | BasicBlock *UserBB = User->getParent(); | |||
1832 | DenseMap<BasicBlock *, CastInst *> InsertedTruncs; | |||
1833 | auto *TruncI = cast<TruncInst>(User); | |||
1834 | bool MadeChange = false; | |||
1835 | ||||
1836 | for (Value::user_iterator TruncUI = TruncI->user_begin(), | |||
1837 | TruncE = TruncI->user_end(); | |||
1838 | TruncUI != TruncE;) { | |||
1839 | ||||
1840 | Use &TruncTheUse = TruncUI.getUse(); | |||
1841 | Instruction *TruncUser = cast<Instruction>(*TruncUI); | |||
1842 | // Preincrement use iterator so we don't invalidate it. | |||
1843 | ||||
1844 | ++TruncUI; | |||
1845 | ||||
1846 | int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode()); | |||
1847 | if (!ISDOpcode) | |||
1848 | continue; | |||
1849 | ||||
1850 | // If the use is actually a legal node, there will not be an | |||
1851 | // implicit truncate. | |||
1852 | // FIXME: always querying the result type is just an | |||
1853 | // approximation; some nodes' legality is determined by the | |||
1854 | // operand or other means. There's no good way to find out though. | |||
1855 | if (TLI.isOperationLegalOrCustom( | |||
1856 | ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true))) | |||
1857 | continue; | |||
1858 | ||||
1859 | // Don't bother for PHI nodes. | |||
1860 | if (isa<PHINode>(TruncUser)) | |||
1861 | continue; | |||
1862 | ||||
1863 | BasicBlock *TruncUserBB = TruncUser->getParent(); | |||
1864 | ||||
1865 | if (UserBB == TruncUserBB) | |||
1866 | continue; | |||
1867 | ||||
1868 | BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB]; | |||
1869 | CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB]; | |||
1870 | ||||
1871 | if (!InsertedShift && !InsertedTrunc) { | |||
1872 | BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt(); | |||
1873 | assert(InsertPt != TruncUserBB->end())(static_cast <bool> (InsertPt != TruncUserBB->end()) ? void (0) : __assert_fail ("InsertPt != TruncUserBB->end()" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1873, __extension__ __PRETTY_FUNCTION__ )); | |||
1874 | // Sink the shift | |||
1875 | if (ShiftI->getOpcode() == Instruction::AShr) | |||
1876 | InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, | |||
1877 | "", &*InsertPt); | |||
1878 | else | |||
1879 | InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, | |||
1880 | "", &*InsertPt); | |||
1881 | InsertedShift->setDebugLoc(ShiftI->getDebugLoc()); | |||
1882 | ||||
1883 | // Sink the trunc | |||
1884 | BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt(); | |||
1885 | TruncInsertPt++; | |||
1886 | assert(TruncInsertPt != TruncUserBB->end())(static_cast <bool> (TruncInsertPt != TruncUserBB->end ()) ? void (0) : __assert_fail ("TruncInsertPt != TruncUserBB->end()" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1886, __extension__ __PRETTY_FUNCTION__ )); | |||
1887 | ||||
1888 | InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift, | |||
1889 | TruncI->getType(), "", &*TruncInsertPt); | |||
1890 | InsertedTrunc->setDebugLoc(TruncI->getDebugLoc()); | |||
1891 | ||||
1892 | MadeChange = true; | |||
1893 | ||||
1894 | TruncTheUse = InsertedTrunc; | |||
1895 | } | |||
1896 | } | |||
1897 | return MadeChange; | |||
1898 | } | |||
1899 | ||||
1900 | /// Sink the shift *right* instruction into user blocks if the uses could | |||
1901 | /// potentially be combined with this shift instruction and generate BitExtract | |||
1902 | /// instruction. It will only be applied if the architecture supports BitExtract | |||
1903 | /// instruction. Here is an example: | |||
1904 | /// BB1: | |||
1905 | /// %x.extract.shift = lshr i64 %arg1, 32 | |||
1906 | /// BB2: | |||
1907 | /// %x.extract.trunc = trunc i64 %x.extract.shift to i16 | |||
1908 | /// ==> | |||
1909 | /// | |||
1910 | /// BB2: | |||
1911 | /// %x.extract.shift.1 = lshr i64 %arg1, 32 | |||
1912 | /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16 | |||
1913 | /// | |||
1914 | /// CodeGen will recognize the pattern in BB2 and generate BitExtract | |||
1915 | /// instruction. | |||
1916 | /// Return true if any changes are made. | |||
1917 | static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI, | |||
1918 | const TargetLowering &TLI, | |||
1919 | const DataLayout &DL) { | |||
1920 | BasicBlock *DefBB = ShiftI->getParent(); | |||
1921 | ||||
1922 | /// Only insert instructions in each block once. | |||
1923 | DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts; | |||
1924 | ||||
1925 | bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType())); | |||
1926 | ||||
1927 | bool MadeChange = false; | |||
1928 | for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end(); | |||
1929 | UI != E;) { | |||
1930 | Use &TheUse = UI.getUse(); | |||
1931 | Instruction *User = cast<Instruction>(*UI); | |||
1932 | // Preincrement use iterator so we don't invalidate it. | |||
1933 | ++UI; | |||
1934 | ||||
1935 | // Don't bother for PHI nodes. | |||
1936 | if (isa<PHINode>(User)) | |||
1937 | continue; | |||
1938 | ||||
1939 | if (!isExtractBitsCandidateUse(User)) | |||
1940 | continue; | |||
1941 | ||||
1942 | BasicBlock *UserBB = User->getParent(); | |||
1943 | ||||
1944 | if (UserBB == DefBB) { | |||
1945 | // If the shift and truncate instruction are in the same BB. The use of | |||
1946 | // the truncate(TruncUse) may still introduce another truncate if not | |||
1947 | // legal. In this case, we would like to sink both shift and truncate | |||
1948 | // instruction to the BB of TruncUse. | |||
1949 | // for example: | |||
1950 | // BB1: | |||
1951 | // i64 shift.result = lshr i64 opnd, imm | |||
1952 | // trunc.result = trunc shift.result to i16 | |||
1953 | // | |||
1954 | // BB2: | |||
1955 | // ----> We will have an implicit truncate here if the architecture does | |||
1956 | // not have i16 compare. | |||
1957 | // cmp i16 trunc.result, opnd2 | |||
1958 | // | |||
1959 | if (isa<TruncInst>(User) && shiftIsLegal | |||
1960 | // If the type of the truncate is legal, no truncate will be | |||
1961 | // introduced in other basic blocks. | |||
1962 | && | |||
1963 | (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType())))) | |||
1964 | MadeChange = | |||
1965 | SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL); | |||
1966 | ||||
1967 | continue; | |||
1968 | } | |||
1969 | // If we have already inserted a shift into this block, use it. | |||
1970 | BinaryOperator *&InsertedShift = InsertedShifts[UserBB]; | |||
1971 | ||||
1972 | if (!InsertedShift) { | |||
1973 | BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); | |||
1974 | assert(InsertPt != UserBB->end())(static_cast <bool> (InsertPt != UserBB->end()) ? void (0) : __assert_fail ("InsertPt != UserBB->end()", "llvm/lib/CodeGen/CodeGenPrepare.cpp" , 1974, __extension__ __PRETTY_FUNCTION__)); | |||
1975 | ||||
1976 | if (ShiftI->getOpcode() == Instruction::AShr) | |||
1977 | InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, | |||
1978 | "", &*InsertPt); | |||
1979 | else | |||
1980 | InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, | |||
1981 | "", &*InsertPt); | |||
1982 | InsertedShift->setDebugLoc(ShiftI->getDebugLoc()); | |||
1983 | ||||
1984 | MadeChange = true; | |||
1985 | } | |||
1986 | ||||
1987 | // Replace a use of the shift with a use of the new shift. | |||
1988 | TheUse = InsertedShift; | |||
1989 | } | |||
1990 | ||||
1991 | // If we removed all uses, or there are none, nuke the shift. | |||
1992 | if (ShiftI->use_empty()) { | |||
1993 | salvageDebugInfo(*ShiftI); | |||
1994 | ShiftI->eraseFromParent(); | |||
1995 | MadeChange = true; | |||
1996 | } | |||
1997 | ||||
1998 | return MadeChange; | |||
1999 | } | |||
2000 | ||||
2001 | /// If counting leading or trailing zeros is an expensive operation and a zero | |||
2002 | /// input is defined, add a check for zero to avoid calling the intrinsic. | |||
2003 | /// | |||
2004 | /// We want to transform: | |||
2005 | /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false) | |||
2006 | /// | |||
2007 | /// into: | |||
2008 | /// entry: | |||
2009 | /// %cmpz = icmp eq i64 %A, 0 | |||
2010 | /// br i1 %cmpz, label %cond.end, label %cond.false | |||
2011 | /// cond.false: | |||
2012 | /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true) | |||
2013 | /// br label %cond.end | |||
2014 | /// cond.end: | |||
2015 | /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ] | |||
2016 | /// | |||
2017 | /// If the transform is performed, return true and set ModifiedDT to true. | |||
2018 | static bool despeculateCountZeros(IntrinsicInst *CountZeros, | |||
2019 | const TargetLowering *TLI, | |||
2020 | const DataLayout *DL, | |||
2021 | bool &ModifiedDT) { | |||
2022 | // If a zero input is undefined, it doesn't make sense to despeculate that. | |||
2023 | if (match(CountZeros->getOperand(1), m_One())) | |||
2024 | return false; | |||
2025 | ||||
2026 | // If it's cheap to speculate, there's nothing to do. | |||
2027 | auto IntrinsicID = CountZeros->getIntrinsicID(); | |||
2028 | if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) || | |||
2029 | (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz())) | |||
2030 | return false; | |||
2031 | ||||
2032 | // Only handle legal scalar cases. Anything else requires too much work. | |||
2033 | Type *Ty = CountZeros->getType(); | |||
2034 | unsigned SizeInBits = Ty->getScalarSizeInBits(); | |||
2035 | if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits()) | |||
2036 | return false; | |||
2037 | ||||
2038 | // Bail if the value is never zero. | |||
2039 | if (llvm::isKnownNonZero(CountZeros->getOperand(0), *DL)) | |||
2040 | return false; | |||
2041 | ||||
2042 | // The intrinsic will be sunk behind a compare against zero and branch. | |||
2043 | BasicBlock *StartBlock = CountZeros->getParent(); | |||
2044 | BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false"); | |||
2045 | ||||
2046 | // Create another block after the count zero intrinsic. A PHI will be added | |||
2047 | // in this block to select the result of the intrinsic or the bit-width | |||
2048 | // constant if the input to the intrinsic is zero. | |||
2049 | BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros)); | |||
2050 | BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end"); | |||
2051 | ||||
2052 | // Set up a builder to create a compare, conditional branch, and PHI. | |||
2053 | IRBuilder<> Builder(CountZeros->getContext()); | |||
2054 | Builder.SetInsertPoint(StartBlock->getTerminator()); | |||
2055 | Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc()); | |||
2056 | ||||
2057 | // Replace the unconditional branch that was created by the first split with | |||
2058 | // a compare against zero and a conditional branch. | |||
2059 | Value *Zero = Constant::getNullValue(Ty); | |||
2060 | Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz"); | |||
2061 | Builder.CreateCondBr(Cmp, EndBlock, CallBlock); | |||
2062 | StartBlock->getTerminator()->eraseFromParent(); | |||
2063 | ||||
2064 | // Create a PHI in the end block to select either the output of the intrinsic | |||
2065 | // or the bit width of the operand. | |||
2066 | Builder.SetInsertPoint(&EndBlock->front()); | |||
2067 | PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz"); | |||
2068 | CountZeros->replaceAllUsesWith(PN); | |||
2069 | Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits)); | |||
2070 | PN->addIncoming(BitWidth, StartBlock); | |||
2071 | PN->addIncoming(CountZeros, CallBlock); | |||
2072 | ||||
2073 | // We are explicitly handling the zero case, so we can set the intrinsic's | |||
2074 | // undefined zero argument to 'true'. This will also prevent reprocessing the | |||
2075 | // intrinsic; we only despeculate when a zero input is defined. | |||
2076 | CountZeros->setArgOperand(1, Builder.getTrue()); | |||
2077 | ModifiedDT = true; | |||
2078 | return true; | |||
2079 | } | |||
2080 | ||||
2081 | bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) { | |||
2082 | BasicBlock *BB = CI->getParent(); | |||
2083 | ||||
2084 | // Lower inline assembly if we can. | |||
2085 | // If we found an inline asm expession, and if the target knows how to | |||
2086 | // lower it to normal LLVM code, do so now. | |||
2087 | if (CI->isInlineAsm()) { | |||
2088 | if (TLI->ExpandInlineAsm(CI)) { | |||
2089 | // Avoid invalidating the iterator. | |||
2090 | CurInstIterator = BB->begin(); | |||
2091 | // Avoid processing instructions out of order, which could cause | |||
2092 | // reuse before a value is defined. | |||
2093 | SunkAddrs.clear(); | |||
2094 | return true; | |||
2095 | } | |||
2096 | // Sink address computing for memory operands into the block. | |||
2097 | if (optimizeInlineAsmInst(CI)) | |||
2098 | return true; | |||
2099 | } | |||
2100 | ||||
2101 | // Align the pointer arguments to this call if the target thinks it's a good | |||
2102 | // idea | |||
2103 | unsigned MinSize, PrefAlign; | |||
2104 | if (TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) { | |||
2105 | for (auto &Arg : CI->args()) { | |||
2106 | // We want to align both objects whose address is used directly and | |||
2107 | // objects whose address is used in casts and GEPs, though it only makes | |||
2108 | // sense for GEPs if the offset is a multiple of the desired alignment and | |||
2109 | // if size - offset meets the size threshold. | |||
2110 | if (!Arg->getType()->isPointerTy()) | |||
2111 | continue; | |||
2112 | APInt Offset(DL->getIndexSizeInBits( | |||
2113 | cast<PointerType>(Arg->getType())->getAddressSpace()), | |||
2114 | 0); | |||
2115 | Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset); | |||
2116 | uint64_t Offset2 = Offset.getLimitedValue(); | |||
2117 | if ((Offset2 & (PrefAlign-1)) != 0) | |||
2118 | continue; | |||
2119 | AllocaInst *AI; | |||
2120 | if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign && | |||
2121 | DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2) | |||
2122 | AI->setAlignment(Align(PrefAlign)); | |||
2123 | // Global variables can only be aligned if they are defined in this | |||
2124 | // object (i.e. they are uniquely initialized in this object), and | |||
2125 | // over-aligning global variables that have an explicit section is | |||
2126 | // forbidden. | |||
2127 | GlobalVariable *GV; | |||
2128 | if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() && | |||
2129 | GV->getPointerAlignment(*DL) < PrefAlign && | |||
2130 | DL->getTypeAllocSize(GV->getValueType()) >= | |||
2131 | MinSize + Offset2) | |||
2132 | GV->setAlignment(MaybeAlign(PrefAlign)); | |||
2133 | } | |||
2134 | // If this is a memcpy (or similar) then we may be able to improve the | |||
2135 | // alignment | |||
2136 | if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) { | |||
2137 | Align DestAlign = getKnownAlignment(MI->getDest(), *DL); | |||
2138 | MaybeAlign MIDestAlign = MI->getDestAlign(); | |||
2139 | if (!MIDestAlign || DestAlign > *MIDestAlign) | |||
2140 | MI->setDestAlignment(DestAlign); | |||
2141 | if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { | |||
2142 | MaybeAlign MTISrcAlign = MTI->getSourceAlign(); | |||
2143 | Align SrcAlign = getKnownAlignment(MTI->getSource(), *DL); | |||
2144 | if (!MTISrcAlign || SrcAlign > *MTISrcAlign) | |||
2145 | MTI->setSourceAlignment(SrcAlign); | |||
2146 | } | |||
2147 | } | |||
2148 | } | |||
2149 | ||||
2150 | // If we have a cold call site, try to sink addressing computation into the | |||
2151 | // cold block. This interacts with our handling for loads and stores to | |||
2152 | // ensure that we can fold all uses of a potential addressing computation | |||
2153 | // into their uses. TODO: generalize this to work over profiling data | |||
2154 | if (CI->hasFnAttr(Attribute::Cold) && | |||
2155 | !OptSize && !llvm::shouldOptimizeForSize(BB, PSI, BFI.get())) | |||
2156 | for (auto &Arg : CI->args()) { | |||
2157 | if (!Arg->getType()->isPointerTy()) | |||
2158 | continue; | |||
2159 | unsigned AS = Arg->getType()->getPointerAddressSpace(); | |||
2160 | return optimizeMemoryInst(CI, Arg, Arg->getType(), AS); | |||
2161 | } | |||
2162 | ||||
2163 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI); | |||
2164 | if (II) { | |||
2165 | switch (II->getIntrinsicID()) { | |||
2166 | default: break; | |||
2167 | case Intrinsic::assume: | |||
2168 | llvm_unreachable("llvm.assume should have been removed already")::llvm::llvm_unreachable_internal("llvm.assume should have been removed already" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 2168); | |||
2169 | case Intrinsic::experimental_widenable_condition: { | |||
2170 | // Give up on future widening oppurtunties so that we can fold away dead | |||
2171 | // paths and merge blocks before going into block-local instruction | |||
2172 | // selection. | |||
2173 | if (II->use_empty()) { | |||
2174 | II->eraseFromParent(); | |||
2175 | return true; | |||
2176 | } | |||
2177 | Constant *RetVal = ConstantInt::getTrue(II->getContext()); | |||
2178 | resetIteratorIfInvalidatedWhileCalling(BB, [&]() { | |||
2179 | replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr); | |||
2180 | }); | |||
2181 | return true; | |||
2182 | } | |||
2183 | case Intrinsic::objectsize: | |||
2184 | llvm_unreachable("llvm.objectsize.* should have been lowered already")::llvm::llvm_unreachable_internal("llvm.objectsize.* should have been lowered already" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 2184); | |||
2185 | case Intrinsic::is_constant: | |||
2186 | llvm_unreachable("llvm.is.constant.* should have been lowered already")::llvm::llvm_unreachable_internal("llvm.is.constant.* should have been lowered already" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 2186); | |||
2187 | case Intrinsic::aarch64_stlxr: | |||
2188 | case Intrinsic::aarch64_stxr: { | |||
2189 | ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0)); | |||
2190 | if (!ExtVal || !ExtVal->hasOneUse() || | |||
2191 | ExtVal->getParent() == CI->getParent()) | |||
2192 | return false; | |||
2193 | // Sink a zext feeding stlxr/stxr before it, so it can be folded into it. | |||
2194 | ExtVal->moveBefore(CI); | |||
2195 | // Mark this instruction as "inserted by CGP", so that other | |||
2196 | // optimizations don't touch it. | |||
2197 | InsertedInsts.insert(ExtVal); | |||
2198 | return true; | |||
2199 | } | |||
2200 | ||||
2201 | case Intrinsic::launder_invariant_group: | |||
2202 | case Intrinsic::strip_invariant_group: { | |||
2203 | Value *ArgVal = II->getArgOperand(0); | |||
2204 | auto it = LargeOffsetGEPMap.find(II); | |||
2205 | if (it != LargeOffsetGEPMap.end()) { | |||
2206 | // Merge entries in LargeOffsetGEPMap to reflect the RAUW. | |||
2207 | // Make sure not to have to deal with iterator invalidation | |||
2208 | // after possibly adding ArgVal to LargeOffsetGEPMap. | |||
2209 | auto GEPs = std::move(it->second); | |||
2210 | LargeOffsetGEPMap[ArgVal].append(GEPs.begin(), GEPs.end()); | |||
2211 | LargeOffsetGEPMap.erase(II); | |||
2212 | } | |||
2213 | ||||
2214 | II->replaceAllUsesWith(ArgVal); | |||
2215 | II->eraseFromParent(); | |||
2216 | return true; | |||
2217 | } | |||
2218 | case Intrinsic::cttz: | |||
2219 | case Intrinsic::ctlz: | |||
2220 | // If counting zeros is expensive, try to avoid it. | |||
2221 | return despeculateCountZeros(II, TLI, DL, ModifiedDT); | |||
2222 | case Intrinsic::fshl: | |||
2223 | case Intrinsic::fshr: | |||
2224 | return optimizeFunnelShift(II); | |||
2225 | case Intrinsic::dbg_value: | |||
2226 | return fixupDbgValue(II); | |||
2227 | case Intrinsic::vscale: { | |||
2228 | // If datalayout has no special restrictions on vector data layout, | |||
2229 | // replace `llvm.vscale` by an equivalent constant expression | |||
2230 | // to benefit from cheap constant propagation. | |||
2231 | Type *ScalableVectorTy = | |||
2232 | VectorType::get(Type::getInt8Ty(II->getContext()), 1, true); | |||
2233 | if (DL->getTypeAllocSize(ScalableVectorTy).getKnownMinSize() == 8) { | |||
2234 | auto *Null = Constant::getNullValue(ScalableVectorTy->getPointerTo()); | |||
2235 | auto *One = ConstantInt::getSigned(II->getType(), 1); | |||
2236 | auto *CGep = | |||
2237 | ConstantExpr::getGetElementPtr(ScalableVectorTy, Null, One); | |||
2238 | II->replaceAllUsesWith(ConstantExpr::getPtrToInt(CGep, II->getType())); | |||
2239 | II->eraseFromParent(); | |||
2240 | return true; | |||
2241 | } | |||
2242 | break; | |||
2243 | } | |||
2244 | case Intrinsic::masked_gather: | |||
2245 | return optimizeGatherScatterInst(II, II->getArgOperand(0)); | |||
2246 | case Intrinsic::masked_scatter: | |||
2247 | return optimizeGatherScatterInst(II, II->getArgOperand(1)); | |||
2248 | } | |||
2249 | ||||
2250 | SmallVector<Value *, 2> PtrOps; | |||
2251 | Type *AccessTy; | |||
2252 | if (TLI->getAddrModeArguments(II, PtrOps, AccessTy)) | |||
2253 | while (!PtrOps.empty()) { | |||
2254 | Value *PtrVal = PtrOps.pop_back_val(); | |||
2255 | unsigned AS = PtrVal->getType()->getPointerAddressSpace(); | |||
2256 | if (optimizeMemoryInst(II, PtrVal, AccessTy, AS)) | |||
2257 | return true; | |||
2258 | } | |||
2259 | } | |||
2260 | ||||
2261 | // From here on out we're working with named functions. | |||
2262 | if (!CI->getCalledFunction()) return false; | |||
2263 | ||||
2264 | // Lower all default uses of _chk calls. This is very similar | |||
2265 | // to what InstCombineCalls does, but here we are only lowering calls | |||
2266 | // to fortified library functions (e.g. __memcpy_chk) that have the default | |||
2267 | // "don't know" as the objectsize. Anything else should be left alone. | |||
2268 | FortifiedLibCallSimplifier Simplifier(TLInfo, true); | |||
2269 | IRBuilder<> Builder(CI); | |||
2270 | if (Value *V = Simplifier.optimizeCall(CI, Builder)) { | |||
2271 | CI->replaceAllUsesWith(V); | |||
2272 | CI->eraseFromParent(); | |||
2273 | return true; | |||
2274 | } | |||
2275 | ||||
2276 | return false; | |||
2277 | } | |||
2278 | ||||
2279 | /// Look for opportunities to duplicate return instructions to the predecessor | |||
2280 | /// to enable tail call optimizations. The case it is currently looking for is: | |||
2281 | /// @code | |||
2282 | /// bb0: | |||
2283 | /// %tmp0 = tail call i32 @f0() | |||
2284 | /// br label %return | |||
2285 | /// bb1: | |||
2286 | /// %tmp1 = tail call i32 @f1() | |||
2287 | /// br label %return | |||
2288 | /// bb2: | |||
2289 | /// %tmp2 = tail call i32 @f2() | |||
2290 | /// br label %return | |||
2291 | /// return: | |||
2292 | /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ] | |||
2293 | /// ret i32 %retval | |||
2294 | /// @endcode | |||
2295 | /// | |||
2296 | /// => | |||
2297 | /// | |||
2298 | /// @code | |||
2299 | /// bb0: | |||
2300 | /// %tmp0 = tail call i32 @f0() | |||
2301 | /// ret i32 %tmp0 | |||
2302 | /// bb1: | |||
2303 | /// %tmp1 = tail call i32 @f1() | |||
2304 | /// ret i32 %tmp1 | |||
2305 | /// bb2: | |||
2306 | /// %tmp2 = tail call i32 @f2() | |||
2307 | /// ret i32 %tmp2 | |||
2308 | /// @endcode | |||
2309 | bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB, bool &ModifiedDT) { | |||
2310 | ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator()); | |||
2311 | if (!RetI) | |||
2312 | return false; | |||
2313 | ||||
2314 | PHINode *PN = nullptr; | |||
2315 | ExtractValueInst *EVI = nullptr; | |||
2316 | BitCastInst *BCI = nullptr; | |||
2317 | Value *V = RetI->getReturnValue(); | |||
2318 | if (V) { | |||
2319 | BCI = dyn_cast<BitCastInst>(V); | |||
2320 | if (BCI) | |||
2321 | V = BCI->getOperand(0); | |||
2322 | ||||
2323 | EVI = dyn_cast<ExtractValueInst>(V); | |||
2324 | if (EVI) { | |||
2325 | V = EVI->getOperand(0); | |||
2326 | if (!llvm::all_of(EVI->indices(), [](unsigned idx) { return idx == 0; })) | |||
2327 | return false; | |||
2328 | } | |||
2329 | ||||
2330 | PN = dyn_cast<PHINode>(V); | |||
2331 | if (!PN) | |||
2332 | return false; | |||
2333 | } | |||
2334 | ||||
2335 | if (PN && PN->getParent() != BB) | |||
2336 | return false; | |||
2337 | ||||
2338 | auto isLifetimeEndOrBitCastFor = [](const Instruction *Inst) { | |||
2339 | const BitCastInst *BC = dyn_cast<BitCastInst>(Inst); | |||
2340 | if (BC && BC->hasOneUse()) | |||
2341 | Inst = BC->user_back(); | |||
2342 | ||||
2343 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) | |||
2344 | return II->getIntrinsicID() == Intrinsic::lifetime_end; | |||
2345 | return false; | |||
2346 | }; | |||
2347 | ||||
2348 | // Make sure there are no instructions between the first instruction | |||
2349 | // and return. | |||
2350 | const Instruction *BI = BB->getFirstNonPHI(); | |||
2351 | // Skip over debug and the bitcast. | |||
2352 | while (isa<DbgInfoIntrinsic>(BI) || BI == BCI || BI == EVI || | |||
2353 | isa<PseudoProbeInst>(BI) || isLifetimeEndOrBitCastFor(BI)) | |||
2354 | BI = BI->getNextNode(); | |||
2355 | if (BI != RetI) | |||
2356 | return false; | |||
2357 | ||||
2358 | /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail | |||
2359 | /// call. | |||
2360 | const Function *F = BB->getParent(); | |||
2361 | SmallVector<BasicBlock*, 4> TailCallBBs; | |||
2362 | if (PN) { | |||
2363 | for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) { | |||
2364 | // Look through bitcasts. | |||
2365 | Value *IncomingVal = PN->getIncomingValue(I)->stripPointerCasts(); | |||
2366 | CallInst *CI = dyn_cast<CallInst>(IncomingVal); | |||
2367 | BasicBlock *PredBB = PN->getIncomingBlock(I); | |||
2368 | // Make sure the phi value is indeed produced by the tail call. | |||
2369 | if (CI && CI->hasOneUse() && CI->getParent() == PredBB && | |||
2370 | TLI->mayBeEmittedAsTailCall(CI) && | |||
2371 | attributesPermitTailCall(F, CI, RetI, *TLI)) | |||
2372 | TailCallBBs.push_back(PredBB); | |||
2373 | } | |||
2374 | } else { | |||
2375 | SmallPtrSet<BasicBlock*, 4> VisitedBBs; | |||
2376 | for (BasicBlock *Pred : predecessors(BB)) { | |||
2377 | if (!VisitedBBs.insert(Pred).second) | |||
2378 | continue; | |||
2379 | if (Instruction *I = Pred->rbegin()->getPrevNonDebugInstruction(true)) { | |||
2380 | CallInst *CI = dyn_cast<CallInst>(I); | |||
2381 | if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) && | |||
2382 | attributesPermitTailCall(F, CI, RetI, *TLI)) | |||
2383 | TailCallBBs.push_back(Pred); | |||
2384 | } | |||
2385 | } | |||
2386 | } | |||
2387 | ||||
2388 | bool Changed = false; | |||
2389 | for (auto const &TailCallBB : TailCallBBs) { | |||
2390 | // Make sure the call instruction is followed by an unconditional branch to | |||
2391 | // the return block. | |||
2392 | BranchInst *BI = dyn_cast<BranchInst>(TailCallBB->getTerminator()); | |||
2393 | if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB) | |||
2394 | continue; | |||
2395 | ||||
2396 | // Duplicate the return into TailCallBB. | |||
2397 | (void)FoldReturnIntoUncondBranch(RetI, BB, TailCallBB); | |||
2398 | assert(!VerifyBFIUpdates ||(static_cast <bool> (!VerifyBFIUpdates || BFI->getBlockFreq (BB) >= BFI->getBlockFreq(TailCallBB)) ? void (0) : __assert_fail ("!VerifyBFIUpdates || BFI->getBlockFreq(BB) >= BFI->getBlockFreq(TailCallBB)" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 2399, __extension__ __PRETTY_FUNCTION__ )) | |||
2399 | BFI->getBlockFreq(BB) >= BFI->getBlockFreq(TailCallBB))(static_cast <bool> (!VerifyBFIUpdates || BFI->getBlockFreq (BB) >= BFI->getBlockFreq(TailCallBB)) ? void (0) : __assert_fail ("!VerifyBFIUpdates || BFI->getBlockFreq(BB) >= BFI->getBlockFreq(TailCallBB)" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 2399, __extension__ __PRETTY_FUNCTION__ )); | |||
2400 | BFI->setBlockFreq( | |||
2401 | BB, | |||
2402 | (BFI->getBlockFreq(BB) - BFI->getBlockFreq(TailCallBB)).getFrequency()); | |||
2403 | ModifiedDT = Changed = true; | |||
2404 | ++NumRetsDup; | |||
2405 | } | |||
2406 | ||||
2407 | // If we eliminated all predecessors of the block, delete the block now. | |||
2408 | if (Changed && !BB->hasAddressTaken() && pred_empty(BB)) | |||
2409 | BB->eraseFromParent(); | |||
2410 | ||||
2411 | return Changed; | |||
2412 | } | |||
2413 | ||||
2414 | //===----------------------------------------------------------------------===// | |||
2415 | // Memory Optimization | |||
2416 | //===----------------------------------------------------------------------===// | |||
2417 | ||||
2418 | namespace { | |||
2419 | ||||
2420 | /// This is an extended version of TargetLowering::AddrMode | |||
2421 | /// which holds actual Value*'s for register values. | |||
2422 | struct ExtAddrMode : public TargetLowering::AddrMode { | |||
2423 | Value *BaseReg = nullptr; | |||
2424 | Value *ScaledReg = nullptr; | |||
2425 | Value *OriginalValue = nullptr; | |||
2426 | bool InBounds = true; | |||
2427 | ||||
2428 | enum FieldName { | |||
2429 | NoField = 0x00, | |||
2430 | BaseRegField = 0x01, | |||
2431 | BaseGVField = 0x02, | |||
2432 | BaseOffsField = 0x04, | |||
2433 | ScaledRegField = 0x08, | |||
2434 | ScaleField = 0x10, | |||
2435 | MultipleFields = 0xff | |||
2436 | }; | |||
2437 | ||||
2438 | ||||
2439 | ExtAddrMode() = default; | |||
2440 | ||||
2441 | void print(raw_ostream &OS) const; | |||
2442 | void dump() const; | |||
2443 | ||||
2444 | FieldName compare(const ExtAddrMode &other) { | |||
2445 | // First check that the types are the same on each field, as differing types | |||
2446 | // is something we can't cope with later on. | |||
2447 | if (BaseReg && other.BaseReg && | |||
2448 | BaseReg->getType() != other.BaseReg->getType()) | |||
2449 | return MultipleFields; | |||
2450 | if (BaseGV && other.BaseGV && | |||
2451 | BaseGV->getType() != other.BaseGV->getType()) | |||
2452 | return MultipleFields; | |||
2453 | if (ScaledReg && other.ScaledReg && | |||
2454 | ScaledReg->getType() != other.ScaledReg->getType()) | |||
2455 | return MultipleFields; | |||
2456 | ||||
2457 | // Conservatively reject 'inbounds' mismatches. | |||
2458 | if (InBounds != other.InBounds) | |||
2459 | return MultipleFields; | |||
2460 | ||||
2461 | // Check each field to see if it differs. | |||
2462 | unsigned Result = NoField; | |||
2463 | if (BaseReg != other.BaseReg) | |||
2464 | Result |= BaseRegField; | |||
2465 | if (BaseGV != other.BaseGV) | |||
2466 | Result |= BaseGVField; | |||
2467 | if (BaseOffs != other.BaseOffs) | |||
2468 | Result |= BaseOffsField; | |||
2469 | if (ScaledReg != other.ScaledReg) | |||
2470 | Result |= ScaledRegField; | |||
2471 | // Don't count 0 as being a different scale, because that actually means | |||
2472 | // unscaled (which will already be counted by having no ScaledReg). | |||
2473 | if (Scale && other.Scale && Scale != other.Scale) | |||
2474 | Result |= ScaleField; | |||
2475 | ||||
2476 | if (countPopulation(Result) > 1) | |||
2477 | return MultipleFields; | |||
2478 | else | |||
2479 | return static_cast<FieldName>(Result); | |||
2480 | } | |||
2481 | ||||
2482 | // An AddrMode is trivial if it involves no calculation i.e. it is just a base | |||
2483 | // with no offset. | |||
2484 | bool isTrivial() { | |||
2485 | // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is | |||
2486 | // trivial if at most one of these terms is nonzero, except that BaseGV and | |||
2487 | // BaseReg both being zero actually means a null pointer value, which we | |||
2488 | // consider to be 'non-zero' here. | |||
2489 | return !BaseOffs && !Scale && !(BaseGV && BaseReg); | |||
2490 | } | |||
2491 | ||||
2492 | Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) { | |||
2493 | switch (Field) { | |||
2494 | default: | |||
2495 | return nullptr; | |||
2496 | case BaseRegField: | |||
2497 | return BaseReg; | |||
2498 | case BaseGVField: | |||
2499 | return BaseGV; | |||
2500 | case ScaledRegField: | |||
2501 | return ScaledReg; | |||
2502 | case BaseOffsField: | |||
2503 | return ConstantInt::get(IntPtrTy, BaseOffs); | |||
2504 | } | |||
2505 | } | |||
2506 | ||||
2507 | void SetCombinedField(FieldName Field, Value *V, | |||
2508 | const SmallVectorImpl<ExtAddrMode> &AddrModes) { | |||
2509 | switch (Field) { | |||
2510 | default: | |||
2511 | llvm_unreachable("Unhandled fields are expected to be rejected earlier")::llvm::llvm_unreachable_internal("Unhandled fields are expected to be rejected earlier" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 2511); | |||
2512 | break; | |||
2513 | case ExtAddrMode::BaseRegField: | |||
2514 | BaseReg = V; | |||
2515 | break; | |||
2516 | case ExtAddrMode::BaseGVField: | |||
2517 | // A combined BaseGV is an Instruction, not a GlobalValue, so it goes | |||
2518 | // in the BaseReg field. | |||
2519 | assert(BaseReg == nullptr)(static_cast <bool> (BaseReg == nullptr) ? void (0) : __assert_fail ("BaseReg == nullptr", "llvm/lib/CodeGen/CodeGenPrepare.cpp" , 2519, __extension__ __PRETTY_FUNCTION__)); | |||
2520 | BaseReg = V; | |||
2521 | BaseGV = nullptr; | |||
2522 | break; | |||
2523 | case ExtAddrMode::ScaledRegField: | |||
2524 | ScaledReg = V; | |||
2525 | // If we have a mix of scaled and unscaled addrmodes then we want scale | |||
2526 | // to be the scale and not zero. | |||
2527 | if (!Scale) | |||
2528 | for (const ExtAddrMode &AM : AddrModes) | |||
2529 | if (AM.Scale) { | |||
2530 | Scale = AM.Scale; | |||
2531 | break; | |||
2532 | } | |||
2533 | break; | |||
2534 | case ExtAddrMode::BaseOffsField: | |||
2535 | // The offset is no longer a constant, so it goes in ScaledReg with a | |||
2536 | // scale of 1. | |||
2537 | assert(ScaledReg == nullptr)(static_cast <bool> (ScaledReg == nullptr) ? void (0) : __assert_fail ("ScaledReg == nullptr", "llvm/lib/CodeGen/CodeGenPrepare.cpp" , 2537, __extension__ __PRETTY_FUNCTION__)); | |||
2538 | ScaledReg = V; | |||
2539 | Scale = 1; | |||
2540 | BaseOffs = 0; | |||
2541 | break; | |||
2542 | } | |||
2543 | } | |||
2544 | }; | |||
2545 | ||||
2546 | } // end anonymous namespace | |||
2547 | ||||
2548 | #ifndef NDEBUG | |||
2549 | static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) { | |||
2550 | AM.print(OS); | |||
2551 | return OS; | |||
2552 | } | |||
2553 | #endif | |||
2554 | ||||
2555 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | |||
2556 | void ExtAddrMode::print(raw_ostream &OS) const { | |||
2557 | bool NeedPlus = false; | |||
2558 | OS << "["; | |||
2559 | if (InBounds) | |||
2560 | OS << "inbounds "; | |||
2561 | if (BaseGV) { | |||
2562 | OS << (NeedPlus ? " + " : "") | |||
2563 | << "GV:"; | |||
2564 | BaseGV->printAsOperand(OS, /*PrintType=*/false); | |||
2565 | NeedPlus = true; | |||
2566 | } | |||
2567 | ||||
2568 | if (BaseOffs) { | |||
2569 | OS << (NeedPlus ? " + " : "") | |||
2570 | << BaseOffs; | |||
2571 | NeedPlus = true; | |||
2572 | } | |||
2573 | ||||
2574 | if (BaseReg) { | |||
2575 | OS << (NeedPlus ? " + " : "") | |||
2576 | << "Base:"; | |||
2577 | BaseReg->printAsOperand(OS, /*PrintType=*/false); | |||
2578 | NeedPlus = true; | |||
2579 | } | |||
2580 | if (Scale) { | |||
2581 | OS << (NeedPlus ? " + " : "") | |||
2582 | << Scale << "*"; | |||
2583 | ScaledReg->printAsOperand(OS, /*PrintType=*/false); | |||
2584 | } | |||
2585 | ||||
2586 | OS << ']'; | |||
2587 | } | |||
2588 | ||||
2589 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void ExtAddrMode::dump() const { | |||
2590 | print(dbgs()); | |||
2591 | dbgs() << '\n'; | |||
2592 | } | |||
2593 | #endif | |||
2594 | ||||
2595 | namespace { | |||
2596 | ||||
2597 | /// This class provides transaction based operation on the IR. | |||
2598 | /// Every change made through this class is recorded in the internal state and | |||
2599 | /// can be undone (rollback) until commit is called. | |||
2600 | /// CGP does not check if instructions could be speculatively executed when | |||
2601 | /// moved. Preserving the original location would pessimize the debugging | |||
2602 | /// experience, as well as negatively impact the quality of sample PGO. | |||
2603 | class TypePromotionTransaction { | |||
2604 | /// This represents the common interface of the individual transaction. | |||
2605 | /// Each class implements the logic for doing one specific modification on | |||
2606 | /// the IR via the TypePromotionTransaction. | |||
2607 | class TypePromotionAction { | |||
2608 | protected: | |||
2609 | /// The Instruction modified. | |||
2610 | Instruction *Inst; | |||
2611 | ||||
2612 | public: | |||
2613 | /// Constructor of the action. | |||
2614 | /// The constructor performs the related action on the IR. | |||
2615 | TypePromotionAction(Instruction *Inst) : Inst(Inst) {} | |||
2616 | ||||
2617 | virtual ~TypePromotionAction() = default; | |||
2618 | ||||
2619 | /// Undo the modification done by this action. | |||
2620 | /// When this method is called, the IR must be in the same state as it was | |||
2621 | /// before this action was applied. | |||
2622 | /// \pre Undoing the action works if and only if the IR is in the exact same | |||
2623 | /// state as it was directly after this action was applied. | |||
2624 | virtual void undo() = 0; | |||
2625 | ||||
2626 | /// Advocate every change made by this action. | |||
2627 | /// When the results on the IR of the action are to be kept, it is important | |||
2628 | /// to call this function, otherwise hidden information may be kept forever. | |||
2629 | virtual void commit() { | |||
2630 | // Nothing to be done, this action is not doing anything. | |||
2631 | } | |||
2632 | }; | |||
2633 | ||||
2634 | /// Utility to remember the position of an instruction. | |||
2635 | class InsertionHandler { | |||
2636 | /// Position of an instruction. | |||
2637 | /// Either an instruction: | |||
2638 | /// - Is the first in a basic block: BB is used. | |||
2639 | /// - Has a previous instruction: PrevInst is used. | |||
2640 | union { | |||
2641 | Instruction *PrevInst; | |||
2642 | BasicBlock *BB; | |||
2643 | } Point; | |||
2644 | ||||
2645 | /// Remember whether or not the instruction had a previous instruction. | |||
2646 | bool HasPrevInstruction; | |||
2647 | ||||
2648 | public: | |||
2649 | /// Record the position of \p Inst. | |||
2650 | InsertionHandler(Instruction *Inst) { | |||
2651 | BasicBlock::iterator It = Inst->getIterator(); | |||
2652 | HasPrevInstruction = (It != (Inst->getParent()->begin())); | |||
2653 | if (HasPrevInstruction) | |||
2654 | Point.PrevInst = &*--It; | |||
2655 | else | |||
2656 | Point.BB = Inst->getParent(); | |||
2657 | } | |||
2658 | ||||
2659 | /// Insert \p Inst at the recorded position. | |||
2660 | void insert(Instruction *Inst) { | |||
2661 | if (HasPrevInstruction) { | |||
2662 | if (Inst->getParent()) | |||
2663 | Inst->removeFromParent(); | |||
2664 | Inst->insertAfter(Point.PrevInst); | |||
2665 | } else { | |||
2666 | Instruction *Position = &*Point.BB->getFirstInsertionPt(); | |||
2667 | if (Inst->getParent()) | |||
2668 | Inst->moveBefore(Position); | |||
2669 | else | |||
2670 | Inst->insertBefore(Position); | |||
2671 | } | |||
2672 | } | |||
2673 | }; | |||
2674 | ||||
2675 | /// Move an instruction before another. | |||
2676 | class InstructionMoveBefore : public TypePromotionAction { | |||
2677 | /// Original position of the instruction. | |||
2678 | InsertionHandler Position; | |||
2679 | ||||
2680 | public: | |||
2681 | /// Move \p Inst before \p Before. | |||
2682 | InstructionMoveBefore(Instruction *Inst, Instruction *Before) | |||
2683 | : TypePromotionAction(Inst), Position(Inst) { | |||
2684 | LLVM_DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Beforedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: move: " << * Inst << "\nbefore: " << *Before << "\n"; } } while (false) | |||
2685 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: move: " << * Inst << "\nbefore: " << *Before << "\n"; } } while (false); | |||
2686 | Inst->moveBefore(Before); | |||
2687 | } | |||
2688 | ||||
2689 | /// Move the instruction back to its original position. | |||
2690 | void undo() override { | |||
2691 | LLVM_DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: moveBefore: " << *Inst << "\n"; } } while (false); | |||
2692 | Position.insert(Inst); | |||
2693 | } | |||
2694 | }; | |||
2695 | ||||
2696 | /// Set the operand of an instruction with a new value. | |||
2697 | class OperandSetter : public TypePromotionAction { | |||
2698 | /// Original operand of the instruction. | |||
2699 | Value *Origin; | |||
2700 | ||||
2701 | /// Index of the modified instruction. | |||
2702 | unsigned Idx; | |||
2703 | ||||
2704 | public: | |||
2705 | /// Set \p Idx operand of \p Inst with \p NewVal. | |||
2706 | OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal) | |||
2707 | : TypePromotionAction(Inst), Idx(Idx) { | |||
2708 | LLVM_DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: setOperand: " << Idx << "\n" << "for:" << *Inst << "\n" << "with:" << *NewVal << "\n"; } } while ( false) | |||
2709 | << "for:" << *Inst << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: setOperand: " << Idx << "\n" << "for:" << *Inst << "\n" << "with:" << *NewVal << "\n"; } } while ( false) | |||
2710 | << "with:" << *NewVal << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: setOperand: " << Idx << "\n" << "for:" << *Inst << "\n" << "with:" << *NewVal << "\n"; } } while ( false); | |||
2711 | Origin = Inst->getOperand(Idx); | |||
2712 | Inst->setOperand(Idx, NewVal); | |||
2713 | } | |||
2714 | ||||
2715 | /// Restore the original value of the instruction. | |||
2716 | void undo() override { | |||
2717 | LLVM_DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: setOperand:" << Idx << "\n" << "for: " << *Inst << "\n" << "with: " << *Origin << "\n"; } } while ( false) | |||
2718 | << "for: " << *Inst << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: setOperand:" << Idx << "\n" << "for: " << *Inst << "\n" << "with: " << *Origin << "\n"; } } while ( false) | |||
2719 | << "with: " << *Origin << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: setOperand:" << Idx << "\n" << "for: " << *Inst << "\n" << "with: " << *Origin << "\n"; } } while ( false); | |||
2720 | Inst->setOperand(Idx, Origin); | |||
2721 | } | |||
2722 | }; | |||
2723 | ||||
2724 | /// Hide the operands of an instruction. | |||
2725 | /// Do as if this instruction was not using any of its operands. | |||
2726 | class OperandsHider : public TypePromotionAction { | |||
2727 | /// The list of original operands. | |||
2728 | SmallVector<Value *, 4> OriginalValues; | |||
2729 | ||||
2730 | public: | |||
2731 | /// Remove \p Inst from the uses of the operands of \p Inst. | |||
2732 | OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) { | |||
2733 | LLVM_DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: OperandsHider: " << *Inst << "\n"; } } while (false); | |||
2734 | unsigned NumOpnds = Inst->getNumOperands(); | |||
2735 | OriginalValues.reserve(NumOpnds); | |||
2736 | for (unsigned It = 0; It < NumOpnds; ++It) { | |||
2737 | // Save the current operand. | |||
2738 | Value *Val = Inst->getOperand(It); | |||
2739 | OriginalValues.push_back(Val); | |||
2740 | // Set a dummy one. | |||
2741 | // We could use OperandSetter here, but that would imply an overhead | |||
2742 | // that we are not willing to pay. | |||
2743 | Inst->setOperand(It, UndefValue::get(Val->getType())); | |||
2744 | } | |||
2745 | } | |||
2746 | ||||
2747 | /// Restore the original list of uses. | |||
2748 | void undo() override { | |||
2749 | LLVM_DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: OperandsHider: " << *Inst << "\n"; } } while (false); | |||
2750 | for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It) | |||
2751 | Inst->setOperand(It, OriginalValues[It]); | |||
2752 | } | |||
2753 | }; | |||
2754 | ||||
2755 | /// Build a truncate instruction. | |||
2756 | class TruncBuilder : public TypePromotionAction { | |||
2757 | Value *Val; | |||
2758 | ||||
2759 | public: | |||
2760 | /// Build a truncate instruction of \p Opnd producing a \p Ty | |||
2761 | /// result. | |||
2762 | /// trunc Opnd to Ty. | |||
2763 | TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) { | |||
2764 | IRBuilder<> Builder(Opnd); | |||
2765 | Builder.SetCurrentDebugLocation(DebugLoc()); | |||
2766 | Val = Builder.CreateTrunc(Opnd, Ty, "promoted"); | |||
2767 | LLVM_DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: TruncBuilder: " << *Val << "\n"; } } while (false); | |||
2768 | } | |||
2769 | ||||
2770 | /// Get the built value. | |||
2771 | Value *getBuiltValue() { return Val; } | |||
2772 | ||||
2773 | /// Remove the built instruction. | |||
2774 | void undo() override { | |||
2775 | LLVM_DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: TruncBuilder: " << *Val << "\n"; } } while (false); | |||
2776 | if (Instruction *IVal = dyn_cast<Instruction>(Val)) | |||
2777 | IVal->eraseFromParent(); | |||
2778 | } | |||
2779 | }; | |||
2780 | ||||
2781 | /// Build a sign extension instruction. | |||
2782 | class SExtBuilder : public TypePromotionAction { | |||
2783 | Value *Val; | |||
2784 | ||||
2785 | public: | |||
2786 | /// Build a sign extension instruction of \p Opnd producing a \p Ty | |||
2787 | /// result. | |||
2788 | /// sext Opnd to Ty. | |||
2789 | SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty) | |||
2790 | : TypePromotionAction(InsertPt) { | |||
2791 | IRBuilder<> Builder(InsertPt); | |||
2792 | Val = Builder.CreateSExt(Opnd, Ty, "promoted"); | |||
2793 | LLVM_DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: SExtBuilder: " << *Val << "\n"; } } while (false); | |||
2794 | } | |||
2795 | ||||
2796 | /// Get the built value. | |||
2797 | Value *getBuiltValue() { return Val; } | |||
2798 | ||||
2799 | /// Remove the built instruction. | |||
2800 | void undo() override { | |||
2801 | LLVM_DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: SExtBuilder: " << *Val << "\n"; } } while (false); | |||
2802 | if (Instruction *IVal = dyn_cast<Instruction>(Val)) | |||
2803 | IVal->eraseFromParent(); | |||
2804 | } | |||
2805 | }; | |||
2806 | ||||
2807 | /// Build a zero extension instruction. | |||
2808 | class ZExtBuilder : public TypePromotionAction { | |||
2809 | Value *Val; | |||
2810 | ||||
2811 | public: | |||
2812 | /// Build a zero extension instruction of \p Opnd producing a \p Ty | |||
2813 | /// result. | |||
2814 | /// zext Opnd to Ty. | |||
2815 | ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty) | |||
2816 | : TypePromotionAction(InsertPt) { | |||
2817 | IRBuilder<> Builder(InsertPt); | |||
2818 | Builder.SetCurrentDebugLocation(DebugLoc()); | |||
2819 | Val = Builder.CreateZExt(Opnd, Ty, "promoted"); | |||
2820 | LLVM_DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: ZExtBuilder: " << *Val << "\n"; } } while (false); | |||
2821 | } | |||
2822 | ||||
2823 | /// Get the built value. | |||
2824 | Value *getBuiltValue() { return Val; } | |||
2825 | ||||
2826 | /// Remove the built instruction. | |||
2827 | void undo() override { | |||
2828 | LLVM_DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: ZExtBuilder: " << *Val << "\n"; } } while (false); | |||
2829 | if (Instruction *IVal = dyn_cast<Instruction>(Val)) | |||
2830 | IVal->eraseFromParent(); | |||
2831 | } | |||
2832 | }; | |||
2833 | ||||
2834 | /// Mutate an instruction to another type. | |||
2835 | class TypeMutator : public TypePromotionAction { | |||
2836 | /// Record the original type. | |||
2837 | Type *OrigTy; | |||
2838 | ||||
2839 | public: | |||
2840 | /// Mutate the type of \p Inst into \p NewTy. | |||
2841 | TypeMutator(Instruction *Inst, Type *NewTy) | |||
2842 | : TypePromotionAction(Inst), OrigTy(Inst->getType()) { | |||
2843 | LLVM_DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTydo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy << "\n"; } } while (false) | |||
2844 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy << "\n"; } } while (false); | |||
2845 | Inst->mutateType(NewTy); | |||
2846 | } | |||
2847 | ||||
2848 | /// Mutate the instruction back to its original type. | |||
2849 | void undo() override { | |||
2850 | LLVM_DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTydo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy << "\n"; } } while (false) | |||
2851 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy << "\n"; } } while (false); | |||
2852 | Inst->mutateType(OrigTy); | |||
2853 | } | |||
2854 | }; | |||
2855 | ||||
2856 | /// Replace the uses of an instruction by another instruction. | |||
2857 | class UsesReplacer : public TypePromotionAction { | |||
2858 | /// Helper structure to keep track of the replaced uses. | |||
2859 | struct InstructionAndIdx { | |||
2860 | /// The instruction using the instruction. | |||
2861 | Instruction *Inst; | |||
2862 | ||||
2863 | /// The index where this instruction is used for Inst. | |||
2864 | unsigned Idx; | |||
2865 | ||||
2866 | InstructionAndIdx(Instruction *Inst, unsigned Idx) | |||
2867 | : Inst(Inst), Idx(Idx) {} | |||
2868 | }; | |||
2869 | ||||
2870 | /// Keep track of the original uses (pair Instruction, Index). | |||
2871 | SmallVector<InstructionAndIdx, 4> OriginalUses; | |||
2872 | /// Keep track of the debug users. | |||
2873 | SmallVector<DbgValueInst *, 1> DbgValues; | |||
2874 | ||||
2875 | /// Keep track of the new value so that we can undo it by replacing | |||
2876 | /// instances of the new value with the original value. | |||
2877 | Value *New; | |||
2878 | ||||
2879 | using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator; | |||
2880 | ||||
2881 | public: | |||
2882 | /// Replace all the use of \p Inst by \p New. | |||
2883 | UsesReplacer(Instruction *Inst, Value *New) | |||
2884 | : TypePromotionAction(Inst), New(New) { | |||
2885 | LLVM_DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *Newdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New << "\n"; } } while (false) | |||
2886 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New << "\n"; } } while (false); | |||
2887 | // Record the original uses. | |||
2888 | for (Use &U : Inst->uses()) { | |||
2889 | Instruction *UserI = cast<Instruction>(U.getUser()); | |||
2890 | OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo())); | |||
2891 | } | |||
2892 | // Record the debug uses separately. They are not in the instruction's | |||
2893 | // use list, but they are replaced by RAUW. | |||
2894 | findDbgValues(DbgValues, Inst); | |||
2895 | ||||
2896 | // Now, we can replace the uses. | |||
2897 | Inst->replaceAllUsesWith(New); | |||
2898 | } | |||
2899 | ||||
2900 | /// Reassign the original uses of Inst to Inst. | |||
2901 | void undo() override { | |||
2902 | LLVM_DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: UsersReplacer: " << *Inst << "\n"; } } while (false); | |||
2903 | for (InstructionAndIdx &Use : OriginalUses) | |||
2904 | Use.Inst->setOperand(Use.Idx, Inst); | |||
2905 | // RAUW has replaced all original uses with references to the new value, | |||
2906 | // including the debug uses. Since we are undoing the replacements, | |||
2907 | // the original debug uses must also be reinstated to maintain the | |||
2908 | // correctness and utility of debug value instructions. | |||
2909 | for (auto *DVI : DbgValues) | |||
2910 | DVI->replaceVariableLocationOp(New, Inst); | |||
2911 | } | |||
2912 | }; | |||
2913 | ||||
2914 | /// Remove an instruction from the IR. | |||
2915 | class InstructionRemover : public TypePromotionAction { | |||
2916 | /// Original position of the instruction. | |||
2917 | InsertionHandler Inserter; | |||
2918 | ||||
2919 | /// Helper structure to hide all the link to the instruction. In other | |||
2920 | /// words, this helps to do as if the instruction was removed. | |||
2921 | OperandsHider Hider; | |||
2922 | ||||
2923 | /// Keep track of the uses replaced, if any. | |||
2924 | UsesReplacer *Replacer = nullptr; | |||
2925 | ||||
2926 | /// Keep track of instructions removed. | |||
2927 | SetOfInstrs &RemovedInsts; | |||
2928 | ||||
2929 | public: | |||
2930 | /// Remove all reference of \p Inst and optionally replace all its | |||
2931 | /// uses with New. | |||
2932 | /// \p RemovedInsts Keep track of the instructions removed by this Action. | |||
2933 | /// \pre If !Inst->use_empty(), then New != nullptr | |||
2934 | InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts, | |||
2935 | Value *New = nullptr) | |||
2936 | : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst), | |||
2937 | RemovedInsts(RemovedInsts) { | |||
2938 | if (New) | |||
2939 | Replacer = new UsesReplacer(Inst, New); | |||
2940 | LLVM_DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: InstructionRemover: " << *Inst << "\n"; } } while (false); | |||
2941 | RemovedInsts.insert(Inst); | |||
2942 | /// The instructions removed here will be freed after completing | |||
2943 | /// optimizeBlock() for all blocks as we need to keep track of the | |||
2944 | /// removed instructions during promotion. | |||
2945 | Inst->removeFromParent(); | |||
2946 | } | |||
2947 | ||||
2948 | ~InstructionRemover() override { delete Replacer; } | |||
2949 | ||||
2950 | /// Resurrect the instruction and reassign it to the proper uses if | |||
2951 | /// new value was provided when build this action. | |||
2952 | void undo() override { | |||
2953 | LLVM_DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: InstructionRemover: " << *Inst << "\n"; } } while (false); | |||
2954 | Inserter.insert(Inst); | |||
2955 | if (Replacer) | |||
2956 | Replacer->undo(); | |||
2957 | Hider.undo(); | |||
2958 | RemovedInsts.erase(Inst); | |||
2959 | } | |||
2960 | }; | |||
2961 | ||||
2962 | public: | |||
2963 | /// Restoration point. | |||
2964 | /// The restoration point is a pointer to an action instead of an iterator | |||
2965 | /// because the iterator may be invalidated but not the pointer. | |||
2966 | using ConstRestorationPt = const TypePromotionAction *; | |||
2967 | ||||
2968 | TypePromotionTransaction(SetOfInstrs &RemovedInsts) | |||
2969 | : RemovedInsts(RemovedInsts) {} | |||
2970 | ||||
2971 | /// Advocate every changes made in that transaction. Return true if any change | |||
2972 | /// happen. | |||
2973 | bool commit(); | |||
2974 | ||||
2975 | /// Undo all the changes made after the given point. | |||
2976 | void rollback(ConstRestorationPt Point); | |||
2977 | ||||
2978 | /// Get the current restoration point. | |||
2979 | ConstRestorationPt getRestorationPoint() const; | |||
2980 | ||||
2981 | /// \name API for IR modification with state keeping to support rollback. | |||
2982 | /// @{ | |||
2983 | /// Same as Instruction::setOperand. | |||
2984 | void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal); | |||
2985 | ||||
2986 | /// Same as Instruction::eraseFromParent. | |||
2987 | void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr); | |||
2988 | ||||
2989 | /// Same as Value::replaceAllUsesWith. | |||
2990 | void replaceAllUsesWith(Instruction *Inst, Value *New); | |||
2991 | ||||
2992 | /// Same as Value::mutateType. | |||
2993 | void mutateType(Instruction *Inst, Type *NewTy); | |||
2994 | ||||
2995 | /// Same as IRBuilder::createTrunc. | |||
2996 | Value *createTrunc(Instruction *Opnd, Type *Ty); | |||
2997 | ||||
2998 | /// Same as IRBuilder::createSExt. | |||
2999 | Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty); | |||
3000 | ||||
3001 | /// Same as IRBuilder::createZExt. | |||
3002 | Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty); | |||
3003 | ||||
3004 | /// Same as Instruction::moveBefore. | |||
3005 | void moveBefore(Instruction *Inst, Instruction *Before); | |||
3006 | /// @} | |||
3007 | ||||
3008 | private: | |||
3009 | /// The ordered list of actions made so far. | |||
3010 | SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions; | |||
3011 | ||||
3012 | using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator; | |||
3013 | ||||
3014 | SetOfInstrs &RemovedInsts; | |||
3015 | }; | |||
3016 | ||||
3017 | } // end anonymous namespace | |||
3018 | ||||
3019 | void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx, | |||
3020 | Value *NewVal) { | |||
3021 | Actions.push_back(std::make_unique<TypePromotionTransaction::OperandSetter>( | |||
3022 | Inst, Idx, NewVal)); | |||
3023 | } | |||
3024 | ||||
3025 | void TypePromotionTransaction::eraseInstruction(Instruction *Inst, | |||
3026 | Value *NewVal) { | |||
3027 | Actions.push_back( | |||
3028 | std::make_unique<TypePromotionTransaction::InstructionRemover>( | |||
3029 | Inst, RemovedInsts, NewVal)); | |||
3030 | } | |||
3031 | ||||
3032 | void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst, | |||
3033 | Value *New) { | |||
3034 | Actions.push_back( | |||
3035 | std::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New)); | |||
3036 | } | |||
3037 | ||||
3038 | void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) { | |||
3039 | Actions.push_back( | |||
3040 | std::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy)); | |||
3041 | } | |||
3042 | ||||
3043 | Value *TypePromotionTransaction::createTrunc(Instruction *Opnd, | |||
3044 | Type *Ty) { | |||
3045 | std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty)); | |||
3046 | Value *Val = Ptr->getBuiltValue(); | |||
3047 | Actions.push_back(std::move(Ptr)); | |||
3048 | return Val; | |||
3049 | } | |||
3050 | ||||
3051 | Value *TypePromotionTransaction::createSExt(Instruction *Inst, | |||
3052 | Value *Opnd, Type *Ty) { | |||
3053 | std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty)); | |||
3054 | Value *Val = Ptr->getBuiltValue(); | |||
3055 | Actions.push_back(std::move(Ptr)); | |||
3056 | return Val; | |||
3057 | } | |||
3058 | ||||
3059 | Value *TypePromotionTransaction::createZExt(Instruction *Inst, | |||
3060 | Value *Opnd, Type *Ty) { | |||
3061 | std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty)); | |||
3062 | Value *Val = Ptr->getBuiltValue(); | |||
3063 | Actions.push_back(std::move(Ptr)); | |||
3064 | return Val; | |||
3065 | } | |||
3066 | ||||
3067 | void TypePromotionTransaction::moveBefore(Instruction *Inst, | |||
3068 | Instruction *Before) { | |||
3069 | Actions.push_back( | |||
3070 | std::make_unique<TypePromotionTransaction::InstructionMoveBefore>( | |||
3071 | Inst, Before)); | |||
3072 | } | |||
3073 | ||||
3074 | TypePromotionTransaction::ConstRestorationPt | |||
3075 | TypePromotionTransaction::getRestorationPoint() const { | |||
3076 | return !Actions.empty() ? Actions.back().get() : nullptr; | |||
3077 | } | |||
3078 | ||||
3079 | bool TypePromotionTransaction::commit() { | |||
3080 | for (std::unique_ptr<TypePromotionAction> &Action : Actions) | |||
3081 | Action->commit(); | |||
3082 | bool Modified = !Actions.empty(); | |||
3083 | Actions.clear(); | |||
3084 | return Modified; | |||
3085 | } | |||
3086 | ||||
3087 | void TypePromotionTransaction::rollback( | |||
3088 | TypePromotionTransaction::ConstRestorationPt Point) { | |||
3089 | while (!Actions.empty() && Point != Actions.back().get()) { | |||
3090 | std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val(); | |||
3091 | Curr->undo(); | |||
3092 | } | |||
3093 | } | |||
3094 | ||||
3095 | namespace { | |||
3096 | ||||
3097 | /// A helper class for matching addressing modes. | |||
3098 | /// | |||
3099 | /// This encapsulates the logic for matching the target-legal addressing modes. | |||
3100 | class AddressingModeMatcher { | |||
3101 | SmallVectorImpl<Instruction*> &AddrModeInsts; | |||
3102 | const TargetLowering &TLI; | |||
3103 | const TargetRegisterInfo &TRI; | |||
3104 | const DataLayout &DL; | |||
3105 | const LoopInfo &LI; | |||
3106 | const std::function<const DominatorTree &()> getDTFn; | |||
3107 | ||||
3108 | /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and | |||
3109 | /// the memory instruction that we're computing this address for. | |||
3110 | Type *AccessTy; | |||
3111 | unsigned AddrSpace; | |||
3112 | Instruction *MemoryInst; | |||
3113 | ||||
3114 | /// This is the addressing mode that we're building up. This is | |||
3115 | /// part of the return value of this addressing mode matching stuff. | |||
3116 | ExtAddrMode &AddrMode; | |||
3117 | ||||
3118 | /// The instructions inserted by other CodeGenPrepare optimizations. | |||
3119 | const SetOfInstrs &InsertedInsts; | |||
3120 | ||||
3121 | /// A map from the instructions to their type before promotion. | |||
3122 | InstrToOrigTy &PromotedInsts; | |||
3123 | ||||
3124 | /// The ongoing transaction where every action should be registered. | |||
3125 | TypePromotionTransaction &TPT; | |||
3126 | ||||
3127 | // A GEP which has too large offset to be folded into the addressing mode. | |||
3128 | std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP; | |||
3129 | ||||
3130 | /// This is set to true when we should not do profitability checks. | |||
3131 | /// When true, IsProfitableToFoldIntoAddressingMode always returns true. | |||
3132 | bool IgnoreProfitability; | |||
3133 | ||||
3134 | /// True if we are optimizing for size. | |||
3135 | bool OptSize; | |||
3136 | ||||
3137 | ProfileSummaryInfo *PSI; | |||
3138 | BlockFrequencyInfo *BFI; | |||
3139 | ||||
3140 | AddressingModeMatcher( | |||
3141 | SmallVectorImpl<Instruction *> &AMI, const TargetLowering &TLI, | |||
3142 | const TargetRegisterInfo &TRI, const LoopInfo &LI, | |||
3143 | const std::function<const DominatorTree &()> getDTFn, | |||
3144 | Type *AT, unsigned AS, Instruction *MI, ExtAddrMode &AM, | |||
3145 | const SetOfInstrs &InsertedInsts, InstrToOrigTy &PromotedInsts, | |||
3146 | TypePromotionTransaction &TPT, | |||
3147 | std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP, | |||
3148 | bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) | |||
3149 | : AddrModeInsts(AMI), TLI(TLI), TRI(TRI), | |||
3150 | DL(MI->getModule()->getDataLayout()), LI(LI), getDTFn(getDTFn), | |||
3151 | AccessTy(AT), AddrSpace(AS), MemoryInst(MI), AddrMode(AM), | |||
3152 | InsertedInsts(InsertedInsts), PromotedInsts(PromotedInsts), TPT(TPT), | |||
3153 | LargeOffsetGEP(LargeOffsetGEP), OptSize(OptSize), PSI(PSI), BFI(BFI) { | |||
3154 | IgnoreProfitability = false; | |||
3155 | } | |||
3156 | ||||
3157 | public: | |||
3158 | /// Find the maximal addressing mode that a load/store of V can fold, | |||
3159 | /// give an access type of AccessTy. This returns a list of involved | |||
3160 | /// instructions in AddrModeInsts. | |||
3161 | /// \p InsertedInsts The instructions inserted by other CodeGenPrepare | |||
3162 | /// optimizations. | |||
3163 | /// \p PromotedInsts maps the instructions to their type before promotion. | |||
3164 | /// \p The ongoing transaction where every action should be registered. | |||
3165 | static ExtAddrMode | |||
3166 | Match(Value *V, Type *AccessTy, unsigned AS, Instruction *MemoryInst, | |||
3167 | SmallVectorImpl<Instruction *> &AddrModeInsts, | |||
3168 | const TargetLowering &TLI, const LoopInfo &LI, | |||
3169 | const std::function<const DominatorTree &()> getDTFn, | |||
3170 | const TargetRegisterInfo &TRI, const SetOfInstrs &InsertedInsts, | |||
3171 | InstrToOrigTy &PromotedInsts, TypePromotionTransaction &TPT, | |||
3172 | std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP, | |||
3173 | bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) { | |||
3174 | ExtAddrMode Result; | |||
3175 | ||||
3176 | bool Success = AddressingModeMatcher( | |||
3177 | AddrModeInsts, TLI, TRI, LI, getDTFn, AccessTy, AS, MemoryInst, Result, | |||
3178 | InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP, OptSize, PSI, | |||
3179 | BFI).matchAddr(V, 0); | |||
3180 | (void)Success; assert(Success && "Couldn't select *anything*?")(static_cast <bool> (Success && "Couldn't select *anything*?" ) ? void (0) : __assert_fail ("Success && \"Couldn't select *anything*?\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3180, __extension__ __PRETTY_FUNCTION__ )); | |||
3181 | return Result; | |||
3182 | } | |||
3183 | ||||
3184 | private: | |||
3185 | bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth); | |||
3186 | bool matchAddr(Value *Addr, unsigned Depth); | |||
3187 | bool matchOperationAddr(User *AddrInst, unsigned Opcode, unsigned Depth, | |||
3188 | bool *MovedAway = nullptr); | |||
3189 | bool isProfitableToFoldIntoAddressingMode(Instruction *I, | |||
3190 | ExtAddrMode &AMBefore, | |||
3191 | ExtAddrMode &AMAfter); | |||
3192 | bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2); | |||
3193 | bool isPromotionProfitable(unsigned NewCost, unsigned OldCost, | |||
3194 | Value *PromotedOperand) const; | |||
3195 | }; | |||
3196 | ||||
3197 | class PhiNodeSet; | |||
3198 | ||||
3199 | /// An iterator for PhiNodeSet. | |||
3200 | class PhiNodeSetIterator { | |||
3201 | PhiNodeSet * const Set; | |||
3202 | size_t CurrentIndex = 0; | |||
3203 | ||||
3204 | public: | |||
3205 | /// The constructor. Start should point to either a valid element, or be equal | |||
3206 | /// to the size of the underlying SmallVector of the PhiNodeSet. | |||
3207 | PhiNodeSetIterator(PhiNodeSet * const Set, size_t Start); | |||
3208 | PHINode * operator*() const; | |||
3209 | PhiNodeSetIterator& operator++(); | |||
3210 | bool operator==(const PhiNodeSetIterator &RHS) const; | |||
3211 | bool operator!=(const PhiNodeSetIterator &RHS) const; | |||
3212 | }; | |||
3213 | ||||
3214 | /// Keeps a set of PHINodes. | |||
3215 | /// | |||
3216 | /// This is a minimal set implementation for a specific use case: | |||
3217 | /// It is very fast when there are very few elements, but also provides good | |||
3218 | /// performance when there are many. It is similar to SmallPtrSet, but also | |||
3219 | /// provides iteration by insertion order, which is deterministic and stable | |||
3220 | /// across runs. It is also similar to SmallSetVector, but provides removing | |||
3221 | /// elements in O(1) time. This is achieved by not actually removing the element | |||
3222 | /// from the underlying vector, so comes at the cost of using more memory, but | |||
3223 | /// that is fine, since PhiNodeSets are used as short lived objects. | |||
3224 | class PhiNodeSet { | |||
3225 | friend class PhiNodeSetIterator; | |||
3226 | ||||
3227 | using MapType = SmallDenseMap<PHINode *, size_t, 32>; | |||
3228 | using iterator = PhiNodeSetIterator; | |||
3229 | ||||
3230 | /// Keeps the elements in the order of their insertion in the underlying | |||
3231 | /// vector. To achieve constant time removal, it never deletes any element. | |||
3232 | SmallVector<PHINode *, 32> NodeList; | |||
3233 | ||||
3234 | /// Keeps the elements in the underlying set implementation. This (and not the | |||
3235 | /// NodeList defined above) is the source of truth on whether an element | |||
3236 | /// is actually in the collection. | |||
3237 | MapType NodeMap; | |||
3238 | ||||
3239 | /// Points to the first valid (not deleted) element when the set is not empty | |||
3240 | /// and the value is not zero. Equals to the size of the underlying vector | |||
3241 | /// when the set is empty. When the value is 0, as in the beginning, the | |||
3242 | /// first element may or may not be valid. | |||
3243 | size_t FirstValidElement = 0; | |||
3244 | ||||
3245 | public: | |||
3246 | /// Inserts a new element to the collection. | |||
3247 | /// \returns true if the element is actually added, i.e. was not in the | |||
3248 | /// collection before the operation. | |||
3249 | bool insert(PHINode *Ptr) { | |||
3250 | if (NodeMap.insert(std::make_pair(Ptr, NodeList.size())).second) { | |||
3251 | NodeList.push_back(Ptr); | |||
3252 | return true; | |||
3253 | } | |||
3254 | return false; | |||
3255 | } | |||
3256 | ||||
3257 | /// Removes the element from the collection. | |||
3258 | /// \returns whether the element is actually removed, i.e. was in the | |||
3259 | /// collection before the operation. | |||
3260 | bool erase(PHINode *Ptr) { | |||
3261 | if (NodeMap.erase(Ptr)) { | |||
3262 | SkipRemovedElements(FirstValidElement); | |||
3263 | return true; | |||
3264 | } | |||
3265 | return false; | |||
3266 | } | |||
3267 | ||||
3268 | /// Removes all elements and clears the collection. | |||
3269 | void clear() { | |||
3270 | NodeMap.clear(); | |||
3271 | NodeList.clear(); | |||
3272 | FirstValidElement = 0; | |||
3273 | } | |||
3274 | ||||
3275 | /// \returns an iterator that will iterate the elements in the order of | |||
3276 | /// insertion. | |||
3277 | iterator begin() { | |||
3278 | if (FirstValidElement == 0) | |||
3279 | SkipRemovedElements(FirstValidElement); | |||
3280 | return PhiNodeSetIterator(this, FirstValidElement); | |||
3281 | } | |||
3282 | ||||
3283 | /// \returns an iterator that points to the end of the collection. | |||
3284 | iterator end() { return PhiNodeSetIterator(this, NodeList.size()); } | |||
3285 | ||||
3286 | /// Returns the number of elements in the collection. | |||
3287 | size_t size() const { | |||
3288 | return NodeMap.size(); | |||
3289 | } | |||
3290 | ||||
3291 | /// \returns 1 if the given element is in the collection, and 0 if otherwise. | |||
3292 | size_t count(PHINode *Ptr) const { | |||
3293 | return NodeMap.count(Ptr); | |||
3294 | } | |||
3295 | ||||
3296 | private: | |||
3297 | /// Updates the CurrentIndex so that it will point to a valid element. | |||
3298 | /// | |||
3299 | /// If the element of NodeList at CurrentIndex is valid, it does not | |||
3300 | /// change it. If there are no more valid elements, it updates CurrentIndex | |||
3301 | /// to point to the end of the NodeList. | |||
3302 | void SkipRemovedElements(size_t &CurrentIndex) { | |||
3303 | while (CurrentIndex < NodeList.size()) { | |||
3304 | auto it = NodeMap.find(NodeList[CurrentIndex]); | |||
3305 | // If the element has been deleted and added again later, NodeMap will | |||
3306 | // point to a different index, so CurrentIndex will still be invalid. | |||
3307 | if (it != NodeMap.end() && it->second == CurrentIndex) | |||
3308 | break; | |||
3309 | ++CurrentIndex; | |||
3310 | } | |||
3311 | } | |||
3312 | }; | |||
3313 | ||||
3314 | PhiNodeSetIterator::PhiNodeSetIterator(PhiNodeSet *const Set, size_t Start) | |||
3315 | : Set(Set), CurrentIndex(Start) {} | |||
3316 | ||||
3317 | PHINode * PhiNodeSetIterator::operator*() const { | |||
3318 | assert(CurrentIndex < Set->NodeList.size() &&(static_cast <bool> (CurrentIndex < Set->NodeList .size() && "PhiNodeSet access out of range") ? void ( 0) : __assert_fail ("CurrentIndex < Set->NodeList.size() && \"PhiNodeSet access out of range\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3319, __extension__ __PRETTY_FUNCTION__ )) | |||
3319 | "PhiNodeSet access out of range")(static_cast <bool> (CurrentIndex < Set->NodeList .size() && "PhiNodeSet access out of range") ? void ( 0) : __assert_fail ("CurrentIndex < Set->NodeList.size() && \"PhiNodeSet access out of range\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3319, __extension__ __PRETTY_FUNCTION__ )); | |||
3320 | return Set->NodeList[CurrentIndex]; | |||
3321 | } | |||
3322 | ||||
3323 | PhiNodeSetIterator& PhiNodeSetIterator::operator++() { | |||
3324 | assert(CurrentIndex < Set->NodeList.size() &&(static_cast <bool> (CurrentIndex < Set->NodeList .size() && "PhiNodeSet access out of range") ? void ( 0) : __assert_fail ("CurrentIndex < Set->NodeList.size() && \"PhiNodeSet access out of range\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3325, __extension__ __PRETTY_FUNCTION__ )) | |||
3325 | "PhiNodeSet access out of range")(static_cast <bool> (CurrentIndex < Set->NodeList .size() && "PhiNodeSet access out of range") ? void ( 0) : __assert_fail ("CurrentIndex < Set->NodeList.size() && \"PhiNodeSet access out of range\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3325, __extension__ __PRETTY_FUNCTION__ )); | |||
3326 | ++CurrentIndex; | |||
3327 | Set->SkipRemovedElements(CurrentIndex); | |||
3328 | return *this; | |||
3329 | } | |||
3330 | ||||
3331 | bool PhiNodeSetIterator::operator==(const PhiNodeSetIterator &RHS) const { | |||
3332 | return CurrentIndex == RHS.CurrentIndex; | |||
3333 | } | |||
3334 | ||||
3335 | bool PhiNodeSetIterator::operator!=(const PhiNodeSetIterator &RHS) const { | |||
3336 | return !((*this) == RHS); | |||
3337 | } | |||
3338 | ||||
3339 | /// Keep track of simplification of Phi nodes. | |||
3340 | /// Accept the set of all phi nodes and erase phi node from this set | |||
3341 | /// if it is simplified. | |||
3342 | class SimplificationTracker { | |||
3343 | DenseMap<Value *, Value *> Storage; | |||
3344 | const SimplifyQuery &SQ; | |||
3345 | // Tracks newly created Phi nodes. The elements are iterated by insertion | |||
3346 | // order. | |||
3347 | PhiNodeSet AllPhiNodes; | |||
3348 | // Tracks newly created Select nodes. | |||
3349 | SmallPtrSet<SelectInst *, 32> AllSelectNodes; | |||
3350 | ||||
3351 | public: | |||
3352 | SimplificationTracker(const SimplifyQuery &sq) | |||
3353 | : SQ(sq) {} | |||
3354 | ||||
3355 | Value *Get(Value *V) { | |||
3356 | do { | |||
3357 | auto SV = Storage.find(V); | |||
3358 | if (SV == Storage.end()) | |||
3359 | return V; | |||
3360 | V = SV->second; | |||
3361 | } while (true); | |||
3362 | } | |||
3363 | ||||
3364 | Value *Simplify(Value *Val) { | |||
3365 | SmallVector<Value *, 32> WorkList; | |||
3366 | SmallPtrSet<Value *, 32> Visited; | |||
3367 | WorkList.push_back(Val); | |||
3368 | while (!WorkList.empty()) { | |||
3369 | auto *P = WorkList.pop_back_val(); | |||
3370 | if (!Visited.insert(P).second) | |||
3371 | continue; | |||
3372 | if (auto *PI = dyn_cast<Instruction>(P)) | |||
3373 | if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) { | |||
3374 | for (auto *U : PI->users()) | |||
3375 | WorkList.push_back(cast<Value>(U)); | |||
3376 | Put(PI, V); | |||
3377 | PI->replaceAllUsesWith(V); | |||
3378 | if (auto *PHI = dyn_cast<PHINode>(PI)) | |||
3379 | AllPhiNodes.erase(PHI); | |||
3380 | if (auto *Select = dyn_cast<SelectInst>(PI)) | |||
3381 | AllSelectNodes.erase(Select); | |||
3382 | PI->eraseFromParent(); | |||
3383 | } | |||
3384 | } | |||
3385 | return Get(Val); | |||
3386 | } | |||
3387 | ||||
3388 | void Put(Value *From, Value *To) { | |||
3389 | Storage.insert({ From, To }); | |||
3390 | } | |||
3391 | ||||
3392 | void ReplacePhi(PHINode *From, PHINode *To) { | |||
3393 | Value* OldReplacement = Get(From); | |||
3394 | while (OldReplacement != From) { | |||
3395 | From = To; | |||
3396 | To = dyn_cast<PHINode>(OldReplacement); | |||
3397 | OldReplacement = Get(From); | |||
3398 | } | |||
3399 | assert(To && Get(To) == To && "Replacement PHI node is already replaced.")(static_cast <bool> (To && Get(To) == To && "Replacement PHI node is already replaced.") ? void (0) : __assert_fail ("To && Get(To) == To && \"Replacement PHI node is already replaced.\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3399, __extension__ __PRETTY_FUNCTION__ )); | |||
3400 | Put(From, To); | |||
3401 | From->replaceAllUsesWith(To); | |||
3402 | AllPhiNodes.erase(From); | |||
3403 | From->eraseFromParent(); | |||
3404 | } | |||
3405 | ||||
3406 | PhiNodeSet& newPhiNodes() { return AllPhiNodes; } | |||
3407 | ||||
3408 | void insertNewPhi(PHINode *PN) { AllPhiNodes.insert(PN); } | |||
3409 | ||||
3410 | void insertNewSelect(SelectInst *SI) { AllSelectNodes.insert(SI); } | |||
3411 | ||||
3412 | unsigned countNewPhiNodes() const { return AllPhiNodes.size(); } | |||
3413 | ||||
3414 | unsigned countNewSelectNodes() const { return AllSelectNodes.size(); } | |||
3415 | ||||
3416 | void destroyNewNodes(Type *CommonType) { | |||
3417 | // For safe erasing, replace the uses with dummy value first. | |||
3418 | auto *Dummy = UndefValue::get(CommonType); | |||
3419 | for (auto *I : AllPhiNodes) { | |||
3420 | I->replaceAllUsesWith(Dummy); | |||
3421 | I->eraseFromParent(); | |||
3422 | } | |||
3423 | AllPhiNodes.clear(); | |||
3424 | for (auto *I : AllSelectNodes) { | |||
3425 | I->replaceAllUsesWith(Dummy); | |||
3426 | I->eraseFromParent(); | |||
3427 | } | |||
3428 | AllSelectNodes.clear(); | |||
3429 | } | |||
3430 | }; | |||
3431 | ||||
3432 | /// A helper class for combining addressing modes. | |||
3433 | class AddressingModeCombiner { | |||
3434 | typedef DenseMap<Value *, Value *> FoldAddrToValueMapping; | |||
3435 | typedef std::pair<PHINode *, PHINode *> PHIPair; | |||
3436 | ||||
3437 | private: | |||
3438 | /// The addressing modes we've collected. | |||
3439 | SmallVector<ExtAddrMode, 16> AddrModes; | |||
3440 | ||||
3441 | /// The field in which the AddrModes differ, when we have more than one. | |||
3442 | ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField; | |||
3443 | ||||
3444 | /// Are the AddrModes that we have all just equal to their original values? | |||
3445 | bool AllAddrModesTrivial = true; | |||
3446 | ||||
3447 | /// Common Type for all different fields in addressing modes. | |||
3448 | Type *CommonType = nullptr; | |||
3449 | ||||
3450 | /// SimplifyQuery for simplifyInstruction utility. | |||
3451 | const SimplifyQuery &SQ; | |||
3452 | ||||
3453 | /// Original Address. | |||
3454 | Value *Original; | |||
3455 | ||||
3456 | public: | |||
3457 | AddressingModeCombiner(const SimplifyQuery &_SQ, Value *OriginalValue) | |||
3458 | : SQ(_SQ), Original(OriginalValue) {} | |||
3459 | ||||
3460 | /// Get the combined AddrMode | |||
3461 | const ExtAddrMode &getAddrMode() const { | |||
3462 | return AddrModes[0]; | |||
3463 | } | |||
3464 | ||||
3465 | /// Add a new AddrMode if it's compatible with the AddrModes we already | |||
3466 | /// have. | |||
3467 | /// \return True iff we succeeded in doing so. | |||
3468 | bool addNewAddrMode(ExtAddrMode &NewAddrMode) { | |||
3469 | // Take note of if we have any non-trivial AddrModes, as we need to detect | |||
3470 | // when all AddrModes are trivial as then we would introduce a phi or select | |||
3471 | // which just duplicates what's already there. | |||
3472 | AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial(); | |||
3473 | ||||
3474 | // If this is the first addrmode then everything is fine. | |||
3475 | if (AddrModes.empty()) { | |||
3476 | AddrModes.emplace_back(NewAddrMode); | |||
3477 | return true; | |||
3478 | } | |||
3479 | ||||
3480 | // Figure out how different this is from the other address modes, which we | |||
3481 | // can do just by comparing against the first one given that we only care | |||
3482 | // about the cumulative difference. | |||
3483 | ExtAddrMode::FieldName ThisDifferentField = | |||
3484 | AddrModes[0].compare(NewAddrMode); | |||
3485 | if (DifferentField == ExtAddrMode::NoField) | |||
3486 | DifferentField = ThisDifferentField; | |||
3487 | else if (DifferentField != ThisDifferentField) | |||
3488 | DifferentField = ExtAddrMode::MultipleFields; | |||
3489 | ||||
3490 | // If NewAddrMode differs in more than one dimension we cannot handle it. | |||
3491 | bool CanHandle = DifferentField != ExtAddrMode::MultipleFields; | |||
3492 | ||||
3493 | // If Scale Field is different then we reject. | |||
3494 | CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField; | |||
3495 | ||||
3496 | // We also must reject the case when base offset is different and | |||
3497 | // scale reg is not null, we cannot handle this case due to merge of | |||
3498 | // different offsets will be used as ScaleReg. | |||
3499 | CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField || | |||
3500 | !NewAddrMode.ScaledReg); | |||
3501 | ||||
3502 | // We also must reject the case when GV is different and BaseReg installed | |||
3503 | // due to we want to use base reg as a merge of GV values. | |||
3504 | CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField || | |||
3505 | !NewAddrMode.HasBaseReg); | |||
3506 | ||||
3507 | // Even if NewAddMode is the same we still need to collect it due to | |||
3508 | // original value is different. And later we will need all original values | |||
3509 | // as anchors during finding the common Phi node. | |||
3510 | if (CanHandle) | |||
3511 | AddrModes.emplace_back(NewAddrMode); | |||
3512 | else | |||
3513 | AddrModes.clear(); | |||
3514 | ||||
3515 | return CanHandle; | |||
3516 | } | |||
3517 | ||||
3518 | /// Combine the addressing modes we've collected into a single | |||
3519 | /// addressing mode. | |||
3520 | /// \return True iff we successfully combined them or we only had one so | |||
3521 | /// didn't need to combine them anyway. | |||
3522 | bool combineAddrModes() { | |||
3523 | // If we have no AddrModes then they can't be combined. | |||
3524 | if (AddrModes.size() == 0) | |||
3525 | return false; | |||
3526 | ||||
3527 | // A single AddrMode can trivially be combined. | |||
3528 | if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField) | |||
3529 | return true; | |||
3530 | ||||
3531 | // If the AddrModes we collected are all just equal to the value they are | |||
3532 | // derived from then combining them wouldn't do anything useful. | |||
3533 | if (AllAddrModesTrivial) | |||
3534 | return false; | |||
3535 | ||||
3536 | if (!addrModeCombiningAllowed()) | |||
3537 | return false; | |||
3538 | ||||
3539 | // Build a map between <original value, basic block where we saw it> to | |||
3540 | // value of base register. | |||
3541 | // Bail out if there is no common type. | |||
3542 | FoldAddrToValueMapping Map; | |||
3543 | if (!initializeMap(Map)) | |||
3544 | return false; | |||
3545 | ||||
3546 | Value *CommonValue = findCommon(Map); | |||
3547 | if (CommonValue) | |||
3548 | AddrModes[0].SetCombinedField(DifferentField, CommonValue, AddrModes); | |||
3549 | return CommonValue != nullptr; | |||
3550 | } | |||
3551 | ||||
3552 | private: | |||
3553 | /// Initialize Map with anchor values. For address seen | |||
3554 | /// we set the value of different field saw in this address. | |||
3555 | /// At the same time we find a common type for different field we will | |||
3556 | /// use to create new Phi/Select nodes. Keep it in CommonType field. | |||
3557 | /// Return false if there is no common type found. | |||
3558 | bool initializeMap(FoldAddrToValueMapping &Map) { | |||
3559 | // Keep track of keys where the value is null. We will need to replace it | |||
3560 | // with constant null when we know the common type. | |||
3561 | SmallVector<Value *, 2> NullValue; | |||
3562 | Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType()); | |||
3563 | for (auto &AM : AddrModes) { | |||
3564 | Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy); | |||
3565 | if (DV) { | |||
3566 | auto *Type = DV->getType(); | |||
3567 | if (CommonType && CommonType != Type) | |||
3568 | return false; | |||
3569 | CommonType = Type; | |||
3570 | Map[AM.OriginalValue] = DV; | |||
3571 | } else { | |||
3572 | NullValue.push_back(AM.OriginalValue); | |||
3573 | } | |||
3574 | } | |||
3575 | assert(CommonType && "At least one non-null value must be!")(static_cast <bool> (CommonType && "At least one non-null value must be!" ) ? void (0) : __assert_fail ("CommonType && \"At least one non-null value must be!\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3575, __extension__ __PRETTY_FUNCTION__ )); | |||
3576 | for (auto *V : NullValue) | |||
3577 | Map[V] = Constant::getNullValue(CommonType); | |||
3578 | return true; | |||
3579 | } | |||
3580 | ||||
3581 | /// We have mapping between value A and other value B where B was a field in | |||
3582 | /// addressing mode represented by A. Also we have an original value C | |||
3583 | /// representing an address we start with. Traversing from C through phi and | |||
3584 | /// selects we ended up with A's in a map. This utility function tries to find | |||
3585 | /// a value V which is a field in addressing mode C and traversing through phi | |||
3586 | /// nodes and selects we will end up in corresponded values B in a map. | |||
3587 | /// The utility will create a new Phi/Selects if needed. | |||
3588 | // The simple example looks as follows: | |||
3589 | // BB1: | |||
3590 | // p1 = b1 + 40 | |||
3591 | // br cond BB2, BB3 | |||
3592 | // BB2: | |||
3593 | // p2 = b2 + 40 | |||
3594 | // br BB3 | |||
3595 | // BB3: | |||
3596 | // p = phi [p1, BB1], [p2, BB2] | |||
3597 | // v = load p | |||
3598 | // Map is | |||
3599 | // p1 -> b1 | |||
3600 | // p2 -> b2 | |||
3601 | // Request is | |||
3602 | // p -> ? | |||
3603 | // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3. | |||
3604 | Value *findCommon(FoldAddrToValueMapping &Map) { | |||
3605 | // Tracks the simplification of newly created phi nodes. The reason we use | |||
3606 | // this mapping is because we will add new created Phi nodes in AddrToBase. | |||
3607 | // Simplification of Phi nodes is recursive, so some Phi node may | |||
3608 | // be simplified after we added it to AddrToBase. In reality this | |||
3609 | // simplification is possible only if original phi/selects were not | |||
3610 | // simplified yet. | |||
3611 | // Using this mapping we can find the current value in AddrToBase. | |||
3612 | SimplificationTracker ST(SQ); | |||
3613 | ||||
3614 | // First step, DFS to create PHI nodes for all intermediate blocks. | |||
3615 | // Also fill traverse order for the second step. | |||
3616 | SmallVector<Value *, 32> TraverseOrder; | |||
3617 | InsertPlaceholders(Map, TraverseOrder, ST); | |||
3618 | ||||
3619 | // Second Step, fill new nodes by merged values and simplify if possible. | |||
3620 | FillPlaceholders(Map, TraverseOrder, ST); | |||
3621 | ||||
3622 | if (!AddrSinkNewSelects && ST.countNewSelectNodes() > 0) { | |||
3623 | ST.destroyNewNodes(CommonType); | |||
3624 | return nullptr; | |||
3625 | } | |||
3626 | ||||
3627 | // Now we'd like to match New Phi nodes to existed ones. | |||
3628 | unsigned PhiNotMatchedCount = 0; | |||
3629 | if (!MatchPhiSet(ST, AddrSinkNewPhis, PhiNotMatchedCount)) { | |||
3630 | ST.destroyNewNodes(CommonType); | |||
3631 | return nullptr; | |||
3632 | } | |||
3633 | ||||
3634 | auto *Result = ST.Get(Map.find(Original)->second); | |||
3635 | if (Result) { | |||
3636 | NumMemoryInstsPhiCreated += ST.countNewPhiNodes() + PhiNotMatchedCount; | |||
3637 | NumMemoryInstsSelectCreated += ST.countNewSelectNodes(); | |||
3638 | } | |||
3639 | return Result; | |||
3640 | } | |||
3641 | ||||
3642 | /// Try to match PHI node to Candidate. | |||
3643 | /// Matcher tracks the matched Phi nodes. | |||
3644 | bool MatchPhiNode(PHINode *PHI, PHINode *Candidate, | |||
3645 | SmallSetVector<PHIPair, 8> &Matcher, | |||
3646 | PhiNodeSet &PhiNodesToMatch) { | |||
3647 | SmallVector<PHIPair, 8> WorkList; | |||
3648 | Matcher.insert({ PHI, Candidate }); | |||
3649 | SmallSet<PHINode *, 8> MatchedPHIs; | |||
3650 | MatchedPHIs.insert(PHI); | |||
3651 | WorkList.push_back({ PHI, Candidate }); | |||
3652 | SmallSet<PHIPair, 8> Visited; | |||
3653 | while (!WorkList.empty()) { | |||
3654 | auto Item = WorkList.pop_back_val(); | |||
3655 | if (!Visited.insert(Item).second) | |||
3656 | continue; | |||
3657 | // We iterate over all incoming values to Phi to compare them. | |||
3658 | // If values are different and both of them Phi and the first one is a | |||
3659 | // Phi we added (subject to match) and both of them is in the same basic | |||
3660 | // block then we can match our pair if values match. So we state that | |||
3661 | // these values match and add it to work list to verify that. | |||
3662 | for (auto B : Item.first->blocks()) { | |||
3663 | Value *FirstValue = Item.first->getIncomingValueForBlock(B); | |||
3664 | Value *SecondValue = Item.second->getIncomingValueForBlock(B); | |||
3665 | if (FirstValue == SecondValue) | |||
3666 | continue; | |||
3667 | ||||
3668 | PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue); | |||
3669 | PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue); | |||
3670 | ||||
3671 | // One of them is not Phi or | |||
3672 | // The first one is not Phi node from the set we'd like to match or | |||
3673 | // Phi nodes from different basic blocks then | |||
3674 | // we will not be able to match. | |||
3675 | if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) || | |||
3676 | FirstPhi->getParent() != SecondPhi->getParent()) | |||
3677 | return false; | |||
3678 | ||||
3679 | // If we already matched them then continue. | |||
3680 | if (Matcher.count({ FirstPhi, SecondPhi })) | |||
3681 | continue; | |||
3682 | // So the values are different and does not match. So we need them to | |||
3683 | // match. (But we register no more than one match per PHI node, so that | |||
3684 | // we won't later try to replace them twice.) | |||
3685 | if (MatchedPHIs.insert(FirstPhi).second) | |||
3686 | Matcher.insert({ FirstPhi, SecondPhi }); | |||
3687 | // But me must check it. | |||
3688 | WorkList.push_back({ FirstPhi, SecondPhi }); | |||
3689 | } | |||
3690 | } | |||
3691 | return true; | |||
3692 | } | |||
3693 | ||||
3694 | /// For the given set of PHI nodes (in the SimplificationTracker) try | |||
3695 | /// to find their equivalents. | |||
3696 | /// Returns false if this matching fails and creation of new Phi is disabled. | |||
3697 | bool MatchPhiSet(SimplificationTracker &ST, bool AllowNewPhiNodes, | |||
3698 | unsigned &PhiNotMatchedCount) { | |||
3699 | // Matched and PhiNodesToMatch iterate their elements in a deterministic | |||
3700 | // order, so the replacements (ReplacePhi) are also done in a deterministic | |||
3701 | // order. | |||
3702 | SmallSetVector<PHIPair, 8> Matched; | |||
3703 | SmallPtrSet<PHINode *, 8> WillNotMatch; | |||
3704 | PhiNodeSet &PhiNodesToMatch = ST.newPhiNodes(); | |||
3705 | while (PhiNodesToMatch.size()) { | |||
3706 | PHINode *PHI = *PhiNodesToMatch.begin(); | |||
3707 | ||||
3708 | // Add us, if no Phi nodes in the basic block we do not match. | |||
3709 | WillNotMatch.clear(); | |||
3710 | WillNotMatch.insert(PHI); | |||
3711 | ||||
3712 | // Traverse all Phis until we found equivalent or fail to do that. | |||
3713 | bool IsMatched = false; | |||
3714 | for (auto &P : PHI->getParent()->phis()) { | |||
3715 | // Skip new Phi nodes. | |||
3716 | if (PhiNodesToMatch.count(&P)) | |||
3717 | continue; | |||
3718 | if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch))) | |||
3719 | break; | |||
3720 | // If it does not match, collect all Phi nodes from matcher. | |||
3721 | // if we end up with no match, them all these Phi nodes will not match | |||
3722 | // later. | |||
3723 | for (auto M : Matched) | |||
3724 | WillNotMatch.insert(M.first); | |||
3725 | Matched.clear(); | |||
3726 | } | |||
3727 | if (IsMatched) { | |||
3728 | // Replace all matched values and erase them. | |||
3729 | for (auto MV : Matched) | |||
3730 | ST.ReplacePhi(MV.first, MV.second); | |||
3731 | Matched.clear(); | |||
3732 | continue; | |||
3733 | } | |||
3734 | // If we are not allowed to create new nodes then bail out. | |||
3735 | if (!AllowNewPhiNodes) | |||
3736 | return false; | |||
3737 | // Just remove all seen values in matcher. They will not match anything. | |||
3738 | PhiNotMatchedCount += WillNotMatch.size(); | |||
3739 | for (auto *P : WillNotMatch) | |||
3740 | PhiNodesToMatch.erase(P); | |||
3741 | } | |||
3742 | return true; | |||
3743 | } | |||
3744 | /// Fill the placeholders with values from predecessors and simplify them. | |||
3745 | void FillPlaceholders(FoldAddrToValueMapping &Map, | |||
3746 | SmallVectorImpl<Value *> &TraverseOrder, | |||
3747 | SimplificationTracker &ST) { | |||
3748 | while (!TraverseOrder.empty()) { | |||
3749 | Value *Current = TraverseOrder.pop_back_val(); | |||
3750 | assert(Map.find(Current) != Map.end() && "No node to fill!!!")(static_cast <bool> (Map.find(Current) != Map.end() && "No node to fill!!!") ? void (0) : __assert_fail ("Map.find(Current) != Map.end() && \"No node to fill!!!\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3750, __extension__ __PRETTY_FUNCTION__ )); | |||
3751 | Value *V = Map[Current]; | |||
3752 | ||||
3753 | if (SelectInst *Select = dyn_cast<SelectInst>(V)) { | |||
3754 | // CurrentValue also must be Select. | |||
3755 | auto *CurrentSelect = cast<SelectInst>(Current); | |||
3756 | auto *TrueValue = CurrentSelect->getTrueValue(); | |||
3757 | assert(Map.find(TrueValue) != Map.end() && "No True Value!")(static_cast <bool> (Map.find(TrueValue) != Map.end() && "No True Value!") ? void (0) : __assert_fail ("Map.find(TrueValue) != Map.end() && \"No True Value!\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3757, __extension__ __PRETTY_FUNCTION__ )); | |||
3758 | Select->setTrueValue(ST.Get(Map[TrueValue])); | |||
3759 | auto *FalseValue = CurrentSelect->getFalseValue(); | |||
3760 | assert(Map.find(FalseValue) != Map.end() && "No False Value!")(static_cast <bool> (Map.find(FalseValue) != Map.end() && "No False Value!") ? void (0) : __assert_fail ("Map.find(FalseValue) != Map.end() && \"No False Value!\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3760, __extension__ __PRETTY_FUNCTION__ )); | |||
3761 | Select->setFalseValue(ST.Get(Map[FalseValue])); | |||
3762 | } else { | |||
3763 | // Must be a Phi node then. | |||
3764 | auto *PHI = cast<PHINode>(V); | |||
3765 | // Fill the Phi node with values from predecessors. | |||
3766 | for (auto *B : predecessors(PHI->getParent())) { | |||
3767 | Value *PV = cast<PHINode>(Current)->getIncomingValueForBlock(B); | |||
3768 | assert(Map.find(PV) != Map.end() && "No predecessor Value!")(static_cast <bool> (Map.find(PV) != Map.end() && "No predecessor Value!") ? void (0) : __assert_fail ("Map.find(PV) != Map.end() && \"No predecessor Value!\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3768, __extension__ __PRETTY_FUNCTION__ )); | |||
3769 | PHI->addIncoming(ST.Get(Map[PV]), B); | |||
3770 | } | |||
3771 | } | |||
3772 | Map[Current] = ST.Simplify(V); | |||
3773 | } | |||
3774 | } | |||
3775 | ||||
3776 | /// Starting from original value recursively iterates over def-use chain up to | |||
3777 | /// known ending values represented in a map. For each traversed phi/select | |||
3778 | /// inserts a placeholder Phi or Select. | |||
3779 | /// Reports all new created Phi/Select nodes by adding them to set. | |||
3780 | /// Also reports and order in what values have been traversed. | |||
3781 | void InsertPlaceholders(FoldAddrToValueMapping &Map, | |||
3782 | SmallVectorImpl<Value *> &TraverseOrder, | |||
3783 | SimplificationTracker &ST) { | |||
3784 | SmallVector<Value *, 32> Worklist; | |||
3785 | assert((isa<PHINode>(Original) || isa<SelectInst>(Original)) &&(static_cast <bool> ((isa<PHINode>(Original) || isa <SelectInst>(Original)) && "Address must be a Phi or Select node" ) ? void (0) : __assert_fail ("(isa<PHINode>(Original) || isa<SelectInst>(Original)) && \"Address must be a Phi or Select node\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3786, __extension__ __PRETTY_FUNCTION__ )) | |||
3786 | "Address must be a Phi or Select node")(static_cast <bool> ((isa<PHINode>(Original) || isa <SelectInst>(Original)) && "Address must be a Phi or Select node" ) ? void (0) : __assert_fail ("(isa<PHINode>(Original) || isa<SelectInst>(Original)) && \"Address must be a Phi or Select node\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3786, __extension__ __PRETTY_FUNCTION__ )); | |||
3787 | auto *Dummy = UndefValue::get(CommonType); | |||
3788 | Worklist.push_back(Original); | |||
3789 | while (!Worklist.empty()) { | |||
3790 | Value *Current = Worklist.pop_back_val(); | |||
3791 | // if it is already visited or it is an ending value then skip it. | |||
3792 | if (Map.find(Current) != Map.end()) | |||
3793 | continue; | |||
3794 | TraverseOrder.push_back(Current); | |||
3795 | ||||
3796 | // CurrentValue must be a Phi node or select. All others must be covered | |||
3797 | // by anchors. | |||
3798 | if (SelectInst *CurrentSelect = dyn_cast<SelectInst>(Current)) { | |||
3799 | // Is it OK to get metadata from OrigSelect?! | |||
3800 | // Create a Select placeholder with dummy value. | |||
3801 | SelectInst *Select = SelectInst::Create( | |||
3802 | CurrentSelect->getCondition(), Dummy, Dummy, | |||
3803 | CurrentSelect->getName(), CurrentSelect, CurrentSelect); | |||
3804 | Map[Current] = Select; | |||
3805 | ST.insertNewSelect(Select); | |||
3806 | // We are interested in True and False values. | |||
3807 | Worklist.push_back(CurrentSelect->getTrueValue()); | |||
3808 | Worklist.push_back(CurrentSelect->getFalseValue()); | |||
3809 | } else { | |||
3810 | // It must be a Phi node then. | |||
3811 | PHINode *CurrentPhi = cast<PHINode>(Current); | |||
3812 | unsigned PredCount = CurrentPhi->getNumIncomingValues(); | |||
3813 | PHINode *PHI = | |||
3814 | PHINode::Create(CommonType, PredCount, "sunk_phi", CurrentPhi); | |||
3815 | Map[Current] = PHI; | |||
3816 | ST.insertNewPhi(PHI); | |||
3817 | append_range(Worklist, CurrentPhi->incoming_values()); | |||
3818 | } | |||
3819 | } | |||
3820 | } | |||
3821 | ||||
3822 | bool addrModeCombiningAllowed() { | |||
3823 | if (DisableComplexAddrModes) | |||
3824 | return false; | |||
3825 | switch (DifferentField) { | |||
3826 | default: | |||
3827 | return false; | |||
3828 | case ExtAddrMode::BaseRegField: | |||
3829 | return AddrSinkCombineBaseReg; | |||
3830 | case ExtAddrMode::BaseGVField: | |||
3831 | return AddrSinkCombineBaseGV; | |||
3832 | case ExtAddrMode::BaseOffsField: | |||
3833 | return AddrSinkCombineBaseOffs; | |||
3834 | case ExtAddrMode::ScaledRegField: | |||
3835 | return AddrSinkCombineScaledReg; | |||
3836 | } | |||
3837 | } | |||
3838 | }; | |||
3839 | } // end anonymous namespace | |||
3840 | ||||
3841 | /// Try adding ScaleReg*Scale to the current addressing mode. | |||
3842 | /// Return true and update AddrMode if this addr mode is legal for the target, | |||
3843 | /// false if not. | |||
3844 | bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale, | |||
3845 | unsigned Depth) { | |||
3846 | // If Scale is 1, then this is the same as adding ScaleReg to the addressing | |||
3847 | // mode. Just process that directly. | |||
3848 | if (Scale == 1) | |||
3849 | return matchAddr(ScaleReg, Depth); | |||
3850 | ||||
3851 | // If the scale is 0, it takes nothing to add this. | |||
3852 | if (Scale == 0) | |||
3853 | return true; | |||
3854 | ||||
3855 | // If we already have a scale of this value, we can add to it, otherwise, we | |||
3856 | // need an available scale field. | |||
3857 | if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg) | |||
3858 | return false; | |||
3859 | ||||
3860 | ExtAddrMode TestAddrMode = AddrMode; | |||
3861 | ||||
3862 | // Add scale to turn X*4+X*3 -> X*7. This could also do things like | |||
3863 | // [A+B + A*7] -> [B+A*8]. | |||
3864 | TestAddrMode.Scale += Scale; | |||
3865 | TestAddrMode.ScaledReg = ScaleReg; | |||
3866 | ||||
3867 | // If the new address isn't legal, bail out. | |||
3868 | if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) | |||
3869 | return false; | |||
3870 | ||||
3871 | // It was legal, so commit it. | |||
3872 | AddrMode = TestAddrMode; | |||
3873 | ||||
3874 | // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now | |||
3875 | // to see if ScaleReg is actually X+C. If so, we can turn this into adding | |||
3876 | // X*Scale + C*Scale to addr mode. If we found available IV increment, do not | |||
3877 | // go any further: we can reuse it and cannot eliminate it. | |||
3878 | ConstantInt *CI = nullptr; Value *AddLHS = nullptr; | |||
3879 | if (isa<Instruction>(ScaleReg) && // not a constant expr. | |||
3880 | match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI))) && | |||
3881 | !isIVIncrement(ScaleReg, &LI) && CI->getValue().isSignedIntN(64)) { | |||
3882 | TestAddrMode.InBounds = false; | |||
3883 | TestAddrMode.ScaledReg = AddLHS; | |||
3884 | TestAddrMode.BaseOffs += CI->getSExtValue() * TestAddrMode.Scale; | |||
3885 | ||||
3886 | // If this addressing mode is legal, commit it and remember that we folded | |||
3887 | // this instruction. | |||
3888 | if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) { | |||
3889 | AddrModeInsts.push_back(cast<Instruction>(ScaleReg)); | |||
3890 | AddrMode = TestAddrMode; | |||
3891 | return true; | |||
3892 | } | |||
3893 | // Restore status quo. | |||
3894 | TestAddrMode = AddrMode; | |||
3895 | } | |||
3896 | ||||
3897 | // If this is an add recurrence with a constant step, return the increment | |||
3898 | // instruction and the canonicalized step. | |||
3899 | auto GetConstantStep = [this](const Value * V) | |||
3900 | ->Optional<std::pair<Instruction *, APInt> > { | |||
3901 | auto *PN = dyn_cast<PHINode>(V); | |||
3902 | if (!PN) | |||
3903 | return None; | |||
3904 | auto IVInc = getIVIncrement(PN, &LI); | |||
3905 | if (!IVInc) | |||
3906 | return None; | |||
3907 | // TODO: The result of the intrinsics above is two-compliment. However when | |||
3908 | // IV inc is expressed as add or sub, iv.next is potentially a poison value. | |||
3909 | // If it has nuw or nsw flags, we need to make sure that these flags are | |||
3910 | // inferrable at the point of memory instruction. Otherwise we are replacing | |||
3911 | // well-defined two-compliment computation with poison. Currently, to avoid | |||
3912 | // potentially complex analysis needed to prove this, we reject such cases. | |||
3913 | if (auto *OIVInc = dyn_cast<OverflowingBinaryOperator>(IVInc->first)) | |||
3914 | if (OIVInc->hasNoSignedWrap() || OIVInc->hasNoUnsignedWrap()) | |||
3915 | return None; | |||
3916 | if (auto *ConstantStep = dyn_cast<ConstantInt>(IVInc->second)) | |||
3917 | return std::make_pair(IVInc->first, ConstantStep->getValue()); | |||
3918 | return None; | |||
3919 | }; | |||
3920 | ||||
3921 | // Try to account for the following special case: | |||
3922 | // 1. ScaleReg is an inductive variable; | |||
3923 | // 2. We use it with non-zero offset; | |||
3924 | // 3. IV's increment is available at the point of memory instruction. | |||
3925 | // | |||
3926 | // In this case, we may reuse the IV increment instead of the IV Phi to | |||
3927 | // achieve the following advantages: | |||
3928 | // 1. If IV step matches the offset, we will have no need in the offset; | |||
3929 | // 2. Even if they don't match, we will reduce the overlap of living IV | |||
3930 | // and IV increment, that will potentially lead to better register | |||
3931 | // assignment. | |||
3932 | if (AddrMode.BaseOffs) { | |||
3933 | if (auto IVStep = GetConstantStep(ScaleReg)) { | |||
3934 | Instruction *IVInc = IVStep->first; | |||
3935 | // The following assert is important to ensure a lack of infinite loops. | |||
3936 | // This transforms is (intentionally) the inverse of the one just above. | |||
3937 | // If they don't agree on the definition of an increment, we'd alternate | |||
3938 | // back and forth indefinitely. | |||
3939 | assert(isIVIncrement(IVInc, &LI) && "implied by GetConstantStep")(static_cast <bool> (isIVIncrement(IVInc, &LI) && "implied by GetConstantStep") ? void (0) : __assert_fail ("isIVIncrement(IVInc, &LI) && \"implied by GetConstantStep\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3939, __extension__ __PRETTY_FUNCTION__ )); | |||
3940 | APInt Step = IVStep->second; | |||
3941 | APInt Offset = Step * AddrMode.Scale; | |||
3942 | if (Offset.isSignedIntN(64)) { | |||
3943 | TestAddrMode.InBounds = false; | |||
3944 | TestAddrMode.ScaledReg = IVInc; | |||
3945 | TestAddrMode.BaseOffs -= Offset.getLimitedValue(); | |||
3946 | // If this addressing mode is legal, commit it.. | |||
3947 | // (Note that we defer the (expensive) domtree base legality check | |||
3948 | // to the very last possible point.) | |||
3949 | if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace) && | |||
3950 | getDTFn().dominates(IVInc, MemoryInst)) { | |||
3951 | AddrModeInsts.push_back(cast<Instruction>(IVInc)); | |||
3952 | AddrMode = TestAddrMode; | |||
3953 | return true; | |||
3954 | } | |||
3955 | // Restore status quo. | |||
3956 | TestAddrMode = AddrMode; | |||
3957 | } | |||
3958 | } | |||
3959 | } | |||
3960 | ||||
3961 | // Otherwise, just return what we have. | |||
3962 | return true; | |||
3963 | } | |||
3964 | ||||
3965 | /// This is a little filter, which returns true if an addressing computation | |||
3966 | /// involving I might be folded into a load/store accessing it. | |||
3967 | /// This doesn't need to be perfect, but needs to accept at least | |||
3968 | /// the set of instructions that MatchOperationAddr can. | |||
3969 | static bool MightBeFoldableInst(Instruction *I) { | |||
3970 | switch (I->getOpcode()) { | |||
3971 | case Instruction::BitCast: | |||
3972 | case Instruction::AddrSpaceCast: | |||
3973 | // Don't touch identity bitcasts. | |||
3974 | if (I->getType() == I->getOperand(0)->getType()) | |||
3975 | return false; | |||
3976 | return I->getType()->isIntOrPtrTy(); | |||
3977 | case Instruction::PtrToInt: | |||
3978 | // PtrToInt is always a noop, as we know that the int type is pointer sized. | |||
3979 | return true; | |||
3980 | case Instruction::IntToPtr: | |||
3981 | // We know the input is intptr_t, so this is foldable. | |||
3982 | return true; | |||
3983 | case Instruction::Add: | |||
3984 | return true; | |||
3985 | case Instruction::Mul: | |||
3986 | case Instruction::Shl: | |||
3987 | // Can only handle X*C and X << C. | |||
3988 | return isa<ConstantInt>(I->getOperand(1)); | |||
3989 | case Instruction::GetElementPtr: | |||
3990 | return true; | |||
3991 | default: | |||
3992 | return false; | |||
3993 | } | |||
3994 | } | |||
3995 | ||||
3996 | /// Check whether or not \p Val is a legal instruction for \p TLI. | |||
3997 | /// \note \p Val is assumed to be the product of some type promotion. | |||
3998 | /// Therefore if \p Val has an undefined state in \p TLI, this is assumed | |||
3999 | /// to be legal, as the non-promoted value would have had the same state. | |||
4000 | static bool isPromotedInstructionLegal(const TargetLowering &TLI, | |||
4001 | const DataLayout &DL, Value *Val) { | |||
4002 | Instruction *PromotedInst = dyn_cast<Instruction>(Val); | |||
4003 | if (!PromotedInst) | |||
4004 | return false; | |||
4005 | int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode()); | |||
4006 | // If the ISDOpcode is undefined, it was undefined before the promotion. | |||
4007 | if (!ISDOpcode) | |||
4008 | return true; | |||
4009 | // Otherwise, check if the promoted instruction is legal or not. | |||
4010 | return TLI.isOperationLegalOrCustom( | |||
4011 | ISDOpcode, TLI.getValueType(DL, PromotedInst->getType())); | |||
4012 | } | |||
4013 | ||||
4014 | namespace { | |||
4015 | ||||
4016 | /// Hepler class to perform type promotion. | |||
4017 | class TypePromotionHelper { | |||
4018 | /// Utility function to add a promoted instruction \p ExtOpnd to | |||
4019 | /// \p PromotedInsts and record the type of extension we have seen. | |||
4020 | static void addPromotedInst(InstrToOrigTy &PromotedInsts, | |||
4021 | Instruction *ExtOpnd, | |||
4022 | bool IsSExt) { | |||
4023 | ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension; | |||
4024 | InstrToOrigTy::iterator It = PromotedInsts.find(ExtOpnd); | |||
4025 | if (It != PromotedInsts.end()) { | |||
4026 | // If the new extension is same as original, the information in | |||
4027 | // PromotedInsts[ExtOpnd] is still correct. | |||
4028 | if (It->second.getInt() == ExtTy) | |||
4029 | return; | |||
4030 | ||||
4031 | // Now the new extension is different from old extension, we make | |||
4032 | // the type information invalid by setting extension type to | |||
4033 | // BothExtension. | |||
4034 | ExtTy = BothExtension; | |||
4035 | } | |||
4036 | PromotedInsts[ExtOpnd] = TypeIsSExt(ExtOpnd->getType(), ExtTy); | |||
4037 | } | |||
4038 | ||||
4039 | /// Utility function to query the original type of instruction \p Opnd | |||
4040 | /// with a matched extension type. If the extension doesn't match, we | |||
4041 | /// cannot use the information we had on the original type. | |||
4042 | /// BothExtension doesn't match any extension type. | |||
4043 | static const Type *getOrigType(const InstrToOrigTy &PromotedInsts, | |||
4044 | Instruction *Opnd, | |||
4045 | bool IsSExt) { | |||
4046 | ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension; | |||
4047 | InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd); | |||
4048 | if (It != PromotedInsts.end() && It->second.getInt() == ExtTy) | |||
4049 | return It->second.getPointer(); | |||
4050 | return nullptr; | |||
4051 | } | |||
4052 | ||||
4053 | /// Utility function to check whether or not a sign or zero extension | |||
4054 | /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by | |||
4055 | /// either using the operands of \p Inst or promoting \p Inst. | |||
4056 | /// The type of the extension is defined by \p IsSExt. | |||
4057 | /// In other words, check if: | |||
4058 | /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType. | |||
4059 | /// #1 Promotion applies: | |||
4060 | /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...). | |||
4061 | /// #2 Operand reuses: | |||
4062 | /// ext opnd1 to ConsideredExtType. | |||
4063 | /// \p PromotedInsts maps the instructions to their type before promotion. | |||
4064 | static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType, | |||
4065 | const InstrToOrigTy &PromotedInsts, bool IsSExt); | |||
4066 | ||||
4067 | /// Utility function to determine if \p OpIdx should be promoted when | |||
4068 | /// promoting \p Inst. | |||
4069 | static bool shouldExtOperand(const Instruction *Inst, int OpIdx) { | |||
4070 | return !(isa<SelectInst>(Inst) && OpIdx == 0); | |||
4071 | } | |||
4072 | ||||
4073 | /// Utility function to promote the operand of \p Ext when this | |||
4074 | /// operand is a promotable trunc or sext or zext. | |||
4075 | /// \p PromotedInsts maps the instructions to their type before promotion. | |||
4076 | /// \p CreatedInstsCost[out] contains the cost of all instructions | |||
4077 | /// created to promote the operand of Ext. | |||
4078 | /// Newly added extensions are inserted in \p Exts. | |||
4079 | /// Newly added truncates are inserted in \p Truncs. | |||
4080 | /// Should never be called directly. | |||
4081 | /// \return The promoted value which is used instead of Ext. | |||
4082 | static Value *promoteOperandForTruncAndAnyExt( | |||
4083 | Instruction *Ext, TypePromotionTransaction &TPT, | |||
4084 | InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, | |||
4085 | SmallVectorImpl<Instruction *> *Exts, | |||
4086 | SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI); | |||
4087 | ||||
4088 | /// Utility function to promote the operand of \p Ext when this | |||
4089 | /// operand is promotable and is not a supported trunc or sext. | |||
4090 | /// \p PromotedInsts maps the instructions to their type before promotion. | |||
4091 | /// \p CreatedInstsCost[out] contains the cost of all the instructions | |||
4092 | /// created to promote the operand of Ext. | |||
4093 | /// Newly added extensions are inserted in \p Exts. | |||
4094 | /// Newly added truncates are inserted in \p Truncs. | |||
4095 | /// Should never be called directly. | |||
4096 | /// \return The promoted value which is used instead of Ext. | |||
4097 | static Value *promoteOperandForOther(Instruction *Ext, | |||
4098 | TypePromotionTransaction &TPT, | |||
4099 | InstrToOrigTy &PromotedInsts, | |||
4100 | unsigned &CreatedInstsCost, | |||
4101 | SmallVectorImpl<Instruction *> *Exts, | |||
4102 | SmallVectorImpl<Instruction *> *Truncs, | |||
4103 | const TargetLowering &TLI, bool IsSExt); | |||
4104 | ||||
4105 | /// \see promoteOperandForOther. | |||
4106 | static Value *signExtendOperandForOther( | |||
4107 | Instruction *Ext, TypePromotionTransaction &TPT, | |||
4108 | InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, | |||
4109 | SmallVectorImpl<Instruction *> *Exts, | |||
4110 | SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) { | |||
4111 | return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost, | |||
4112 | Exts, Truncs, TLI, true); | |||
4113 | } | |||
4114 | ||||
4115 | /// \see promoteOperandForOther. | |||
4116 | static Value *zeroExtendOperandForOther( | |||
4117 | Instruction *Ext, TypePromotionTransaction &TPT, | |||
4118 | InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, | |||
4119 | SmallVectorImpl<Instruction *> *Exts, | |||
4120 | SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) { | |||
4121 | return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost, | |||
4122 | Exts, Truncs, TLI, false); | |||
4123 | } | |||
4124 | ||||
4125 | public: | |||
4126 | /// Type for the utility function that promotes the operand of Ext. | |||
4127 | using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT, | |||
4128 | InstrToOrigTy &PromotedInsts, | |||
4129 | unsigned &CreatedInstsCost, | |||
4130 | SmallVectorImpl<Instruction *> *Exts, | |||
4131 | SmallVectorImpl<Instruction *> *Truncs, | |||
4132 | const TargetLowering &TLI); | |||
4133 | ||||
4134 | /// Given a sign/zero extend instruction \p Ext, return the appropriate | |||
4135 | /// action to promote the operand of \p Ext instead of using Ext. | |||
4136 | /// \return NULL if no promotable action is possible with the current | |||
4137 | /// sign extension. | |||
4138 | /// \p InsertedInsts keeps track of all the instructions inserted by the | |||
4139 | /// other CodeGenPrepare optimizations. This information is important | |||
4140 | /// because we do not want to promote these instructions as CodeGenPrepare | |||
4141 | /// will reinsert them later. Thus creating an infinite loop: create/remove. | |||
4142 | /// \p PromotedInsts maps the instructions to their type before promotion. | |||
4143 | static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts, | |||
4144 | const TargetLowering &TLI, | |||
4145 | const InstrToOrigTy &PromotedInsts); | |||
4146 | }; | |||
4147 | ||||
4148 | } // end anonymous namespace | |||
4149 | ||||
4150 | bool TypePromotionHelper::canGetThrough(const Instruction *Inst, | |||
4151 | Type *ConsideredExtType, | |||
4152 | const InstrToOrigTy &PromotedInsts, | |||
4153 | bool IsSExt) { | |||
4154 | // The promotion helper does not know how to deal with vector types yet. | |||
4155 | // To be able to fix that, we would need to fix the places where we | |||
4156 | // statically extend, e.g., constants and such. | |||
4157 | if (Inst->getType()->isVectorTy()) | |||
4158 | return false; | |||
4159 | ||||
4160 | // We can always get through zext. | |||
4161 | if (isa<ZExtInst>(Inst)) | |||
4162 | return true; | |||
4163 | ||||
4164 | // sext(sext) is ok too. | |||
4165 | if (IsSExt && isa<SExtInst>(Inst)) | |||
4166 | return true; | |||
4167 | ||||
4168 | // We can get through binary operator, if it is legal. In other words, the | |||
4169 | // binary operator must have a nuw or nsw flag. | |||
4170 | if (const auto *BinOp = dyn_cast<BinaryOperator>(Inst)) | |||
4171 | if (isa<OverflowingBinaryOperator>(BinOp) && | |||
4172 | ((!IsSExt && BinOp->hasNoUnsignedWrap()) || | |||
4173 | (IsSExt && BinOp->hasNoSignedWrap()))) | |||
4174 | return true; | |||
4175 | ||||
4176 | // ext(and(opnd, cst)) --> and(ext(opnd), ext(cst)) | |||
4177 | if ((Inst->getOpcode() == Instruction::And || | |||
4178 | Inst->getOpcode() == Instruction::Or)) | |||
4179 | return true; | |||
4180 | ||||
4181 | // ext(xor(opnd, cst)) --> xor(ext(opnd), ext(cst)) | |||
4182 | if (Inst->getOpcode() == Instruction::Xor) { | |||
4183 | // Make sure it is not a NOT. | |||
4184 | if (const auto *Cst = dyn_cast<ConstantInt>(Inst->getOperand(1))) | |||
4185 | if (!Cst->getValue().isAllOnes()) | |||
4186 | return true; | |||
4187 | } | |||
4188 | ||||
4189 | // zext(shrl(opnd, cst)) --> shrl(zext(opnd), zext(cst)) | |||
4190 | // It may change a poisoned value into a regular value, like | |||
4191 | // zext i32 (shrl i8 %val, 12) --> shrl i32 (zext i8 %val), 12 | |||
4192 | // poisoned value regular value | |||
4193 | // It should be OK since undef covers valid value. | |||
4194 | if (Inst->getOpcode() == Instruction::LShr && !IsSExt) | |||
4195 | return true; | |||
4196 | ||||
4197 | // and(ext(shl(opnd, cst)), cst) --> and(shl(ext(opnd), ext(cst)), cst) | |||
4198 | // It may change a poisoned value into a regular value, like | |||
4199 | // zext i32 (shl i8 %val, 12) --> shl i32 (zext i8 %val), 12 | |||
4200 | // poisoned value regular value | |||
4201 | // It should be OK since undef covers valid value. | |||
4202 | if (Inst->getOpcode() == Instruction::Shl && Inst->hasOneUse()) { | |||
4203 | const auto *ExtInst = cast<const Instruction>(*Inst->user_begin()); | |||
4204 | if (ExtInst->hasOneUse()) { | |||
4205 | const auto *AndInst = dyn_cast<const Instruction>(*ExtInst->user_begin()); | |||
4206 | if (AndInst && AndInst->getOpcode() == Instruction::And) { | |||
4207 | const auto *Cst = dyn_cast<ConstantInt>(AndInst->getOperand(1)); | |||
4208 | if (Cst && | |||
4209 | Cst->getValue().isIntN(Inst->getType()->getIntegerBitWidth())) | |||
4210 | return true; | |||
4211 | } | |||
4212 | } | |||
4213 | } | |||
4214 | ||||
4215 | // Check if we can do the following simplification. | |||
4216 | // ext(trunc(opnd)) --> ext(opnd) | |||
4217 | if (!isa<TruncInst>(Inst)) | |||
4218 | return false; | |||
4219 | ||||
4220 | Value *OpndVal = Inst->getOperand(0); | |||
4221 | // Check if we can use this operand in the extension. | |||
4222 | // If the type is larger than the result type of the extension, we cannot. | |||
4223 | if (!OpndVal->getType()->isIntegerTy() || | |||
4224 | OpndVal->getType()->getIntegerBitWidth() > | |||
4225 | ConsideredExtType->getIntegerBitWidth()) | |||
4226 | return false; | |||
4227 | ||||
4228 | // If the operand of the truncate is not an instruction, we will not have | |||
4229 | // any information on the dropped bits. | |||
4230 | // (Actually we could for constant but it is not worth the extra logic). | |||
4231 | Instruction *Opnd = dyn_cast<Instruction>(OpndVal); | |||
4232 | if (!Opnd) | |||
4233 | return false; | |||
4234 | ||||
4235 | // Check if the source of the type is narrow enough. | |||
4236 | // I.e., check that trunc just drops extended bits of the same kind of | |||
4237 | // the extension. | |||
4238 | // #1 get the type of the operand and check the kind of the extended bits. | |||
4239 | const Type *OpndType = getOrigType(PromotedInsts, Opnd, IsSExt); | |||
4240 | if (OpndType) | |||
4241 | ; | |||
4242 | else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd))) | |||
4243 | OpndType = Opnd->getOperand(0)->getType(); | |||
4244 | else | |||
4245 | return false; | |||
4246 | ||||
4247 | // #2 check that the truncate just drops extended bits. | |||
4248 | return Inst->getType()->getIntegerBitWidth() >= | |||
4249 | OpndType->getIntegerBitWidth(); | |||
4250 | } | |||
4251 | ||||
4252 | TypePromotionHelper::Action TypePromotionHelper::getAction( | |||
4253 | Instruction *Ext, const SetOfInstrs &InsertedInsts, | |||
4254 | const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) { | |||
4255 | assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&(static_cast <bool> ((isa<SExtInst>(Ext) || isa< ZExtInst>(Ext)) && "Unexpected instruction type") ? void (0) : __assert_fail ("(isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && \"Unexpected instruction type\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 4256, __extension__ __PRETTY_FUNCTION__ )) | |||
4256 | "Unexpected instruction type")(static_cast <bool> ((isa<SExtInst>(Ext) || isa< ZExtInst>(Ext)) && "Unexpected instruction type") ? void (0) : __assert_fail ("(isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && \"Unexpected instruction type\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 4256, __extension__ __PRETTY_FUNCTION__ )); | |||
4257 | Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0)); | |||
4258 | Type *ExtTy = Ext->getType(); | |||
4259 | bool IsSExt = isa<SExtInst>(Ext); | |||
4260 | // If the operand of the extension is not an instruction, we cannot | |||
4261 | // get through. | |||
4262 | // If it, check we can get through. | |||
4263 | if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt)) | |||
4264 | return nullptr; | |||
4265 | ||||
4266 | // Do not promote if the operand has been added by codegenprepare. | |||
4267 | // Otherwise, it means we are undoing an optimization that is likely to be | |||
4268 | // redone, thus causing potential infinite loop. | |||
4269 | if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd)) | |||
4270 | return nullptr; | |||
4271 | ||||
4272 | // SExt or Trunc instructions. | |||
4273 | // Return the related handler. | |||
4274 | if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) || | |||
4275 | isa<ZExtInst>(ExtOpnd)) | |||
4276 | return promoteOperandForTruncAndAnyExt; | |||
4277 | ||||
4278 | // Regular instruction. | |||
4279 | // Abort early if we will have to insert non-free instructions. | |||
4280 | if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType())) | |||
4281 | return nullptr; | |||
4282 | return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther; | |||
4283 | } | |||
4284 | ||||
4285 | Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt( | |||
4286 | Instruction *SExt, TypePromotionTransaction &TPT, | |||
4287 | InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, | |||
4288 | SmallVectorImpl<Instruction *> *Exts, | |||
4289 | SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) { | |||
4290 | // By construction, the operand of SExt is an instruction. Otherwise we cannot | |||
4291 | // get through it and this method should not be called. | |||
4292 | Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0)); | |||
4293 | Value *ExtVal = SExt; | |||
4294 | bool HasMergedNonFreeExt = false; | |||
4295 | if (isa<ZExtInst>(SExtOpnd)) { | |||
4296 | // Replace s|zext(zext(opnd)) | |||
4297 | // => zext(opnd). | |||
4298 | HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd); | |||
4299 | Value *ZExt = | |||
4300 | TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType()); | |||
4301 | TPT.replaceAllUsesWith(SExt, ZExt); | |||
4302 | TPT.eraseInstruction(SExt); | |||
4303 | ExtVal = ZExt; | |||
4304 | } else { | |||
4305 | // Replace z|sext(trunc(opnd)) or sext(sext(opnd)) | |||
4306 | // => z|sext(opnd). | |||
4307 | TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0)); | |||
4308 | } | |||
4309 | CreatedInstsCost = 0; | |||
4310 | ||||
4311 | // Remove dead code. | |||
4312 | if (SExtOpnd->use_empty()) | |||
4313 | TPT.eraseInstruction(SExtOpnd); | |||
4314 | ||||
4315 | // Check if the extension is still needed. | |||
4316 | Instruction *ExtInst = dyn_cast<Instruction>(ExtVal); | |||
4317 | if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) { | |||
4318 | if (ExtInst) { | |||
4319 | if (Exts) | |||
4320 | Exts->push_back(ExtInst); | |||
4321 | CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt; | |||
4322 | } | |||
4323 | return ExtVal; | |||
4324 | } | |||
4325 | ||||
4326 | // At this point we have: ext ty opnd to ty. | |||
4327 | // Reassign the uses of ExtInst to the opnd and remove ExtInst. | |||
4328 | Value *NextVal = ExtInst->getOperand(0); | |||
4329 | TPT.eraseInstruction(ExtInst, NextVal); | |||
4330 | return NextVal; | |||
4331 | } | |||
4332 | ||||
4333 | Value *TypePromotionHelper::promoteOperandForOther( | |||
4334 | Instruction *Ext, TypePromotionTransaction &TPT, | |||
4335 | InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, | |||
4336 | SmallVectorImpl<Instruction *> *Exts, | |||
4337 | SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI, | |||
4338 | bool IsSExt) { | |||
4339 | // By construction, the operand of Ext is an instruction. Otherwise we cannot | |||
4340 | // get through it and this method should not be called. | |||
4341 | Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0)); | |||
4342 | CreatedInstsCost = 0; | |||
4343 | if (!ExtOpnd->hasOneUse()) { | |||
4344 | // ExtOpnd will be promoted. | |||
4345 | // All its uses, but Ext, will need to use a truncated value of the | |||
4346 | // promoted version. | |||
4347 | // Create the truncate now. | |||
4348 | Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType()); | |||
4349 | if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) { | |||
4350 | // Insert it just after the definition. | |||
4351 | ITrunc->moveAfter(ExtOpnd); | |||
4352 | if (Truncs) | |||
4353 | Truncs->push_back(ITrunc); | |||
4354 | } | |||
4355 | ||||
4356 | TPT.replaceAllUsesWith(ExtOpnd, Trunc); | |||
4357 | // Restore the operand of Ext (which has been replaced by the previous call | |||
4358 | // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext. | |||
4359 | TPT.setOperand(Ext, 0, ExtOpnd); | |||
4360 | } | |||
4361 | ||||
4362 | // Get through the Instruction: | |||
4363 | // 1. Update its type. | |||
4364 | // 2. Replace the uses of Ext by Inst. | |||
4365 | // 3. Extend each operand that needs to be extended. | |||
4366 | ||||
4367 | // Remember the original type of the instruction before promotion. | |||
4368 | // This is useful to know that the high bits are sign extended bits. | |||
4369 | addPromotedInst(PromotedInsts, ExtOpnd, IsSExt); | |||
4370 | // Step #1. | |||
4371 | TPT.mutateType(ExtOpnd, Ext->getType()); | |||
4372 | // Step #2. | |||
4373 | TPT.replaceAllUsesWith(Ext, ExtOpnd); | |||
4374 | // Step #3. | |||
4375 | Instruction *ExtForOpnd = Ext; | |||
4376 | ||||
4377 | LLVM_DEBUG(dbgs() << "Propagate Ext to operands\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Propagate Ext to operands\n" ; } } while (false); | |||
4378 | for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx; | |||
4379 | ++OpIdx) { | |||
4380 | LLVM_DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Operand:\n" << * (ExtOpnd->getOperand(OpIdx)) << '\n'; } } while (false ); | |||
4381 | if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() || | |||
4382 | !shouldExtOperand(ExtOpnd, OpIdx)) { | |||
4383 | LLVM_DEBUG(dbgs() << "No need to propagate\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "No need to propagate\n" ; } } while (false); | |||
4384 | continue; | |||
4385 | } | |||
4386 | // Check if we can statically extend the operand. | |||
4387 | Value *Opnd = ExtOpnd->getOperand(OpIdx); | |||
4388 | if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) { | |||
4389 | LLVM_DEBUG(dbgs() << "Statically extend\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Statically extend\n"; } } while (false); | |||
4390 | unsigned BitWidth = Ext->getType()->getIntegerBitWidth(); | |||
4391 | APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth) | |||
4392 | : Cst->getValue().zext(BitWidth); | |||
4393 | TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal)); | |||
4394 | continue; | |||
4395 | } | |||
4396 | // UndefValue are typed, so we have to statically sign extend them. | |||
4397 | if (isa<UndefValue>(Opnd)) { | |||
4398 | LLVM_DEBUG(dbgs() << "Statically extend\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Statically extend\n"; } } while (false); | |||
4399 | TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType())); | |||
4400 | continue; | |||
4401 | } | |||
4402 | ||||
4403 | // Otherwise we have to explicitly sign extend the operand. | |||
4404 | // Check if Ext was reused to extend an operand. | |||
4405 | if (!ExtForOpnd) { | |||
4406 | // If yes, create a new one. | |||
4407 | LLVM_DEBUG(dbgs() << "More operands to ext\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "More operands to ext\n" ; } } while (false); | |||
4408 | Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType()) | |||
4409 | : TPT.createZExt(Ext, Opnd, Ext->getType()); | |||
4410 | if (!isa<Instruction>(ValForExtOpnd)) { | |||
4411 | TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd); | |||
4412 | continue; | |||
4413 | } | |||
4414 | ExtForOpnd = cast<Instruction>(ValForExtOpnd); | |||
4415 | } | |||
4416 | if (Exts) | |||
4417 | Exts->push_back(ExtForOpnd); | |||
4418 | TPT.setOperand(ExtForOpnd, 0, Opnd); | |||
4419 | ||||
4420 | // Move the sign extension before the insertion point. | |||
4421 | TPT.moveBefore(ExtForOpnd, ExtOpnd); | |||
4422 | TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd); | |||
4423 | CreatedInstsCost += !TLI.isExtFree(ExtForOpnd); | |||
4424 | // If more sext are required, new instructions will have to be created. | |||
4425 | ExtForOpnd = nullptr; | |||
4426 | } | |||
4427 | if (ExtForOpnd == Ext) { | |||
4428 | LLVM_DEBUG(dbgs() << "Extension is useless now\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Extension is useless now\n" ; } } while (false); | |||
4429 | TPT.eraseInstruction(Ext); | |||
4430 | } | |||
4431 | return ExtOpnd; | |||
4432 | } | |||
4433 | ||||
4434 | /// Check whether or not promoting an instruction to a wider type is profitable. | |||
4435 | /// \p NewCost gives the cost of extension instructions created by the | |||
4436 | /// promotion. | |||
4437 | /// \p OldCost gives the cost of extension instructions before the promotion | |||
4438 | /// plus the number of instructions that have been | |||
4439 | /// matched in the addressing mode the promotion. | |||
4440 | /// \p PromotedOperand is the value that has been promoted. | |||
4441 | /// \return True if the promotion is profitable, false otherwise. | |||
4442 | bool AddressingModeMatcher::isPromotionProfitable( | |||
4443 | unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const { | |||
4444 | LLVM_DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n'; } } while (false) | |||
4445 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n'; } } while (false); | |||
4446 | // The cost of the new extensions is greater than the cost of the | |||
4447 | // old extension plus what we folded. | |||
4448 | // This is not profitable. | |||
4449 | if (NewCost > OldCost) | |||
4450 | return false; | |||
4451 | if (NewCost < OldCost) | |||
4452 | return true; | |||
4453 | // The promotion is neutral but it may help folding the sign extension in | |||
4454 | // loads for instance. | |||
4455 | // Check that we did not create an illegal instruction. | |||
4456 | return isPromotedInstructionLegal(TLI, DL, PromotedOperand); | |||
4457 | } | |||
4458 | ||||
4459 | /// Given an instruction or constant expr, see if we can fold the operation | |||
4460 | /// into the addressing mode. If so, update the addressing mode and return | |||
4461 | /// true, otherwise return false without modifying AddrMode. | |||
4462 | /// If \p MovedAway is not NULL, it contains the information of whether or | |||
4463 | /// not AddrInst has to be folded into the addressing mode on success. | |||
4464 | /// If \p MovedAway == true, \p AddrInst will not be part of the addressing | |||
4465 | /// because it has been moved away. | |||
4466 | /// Thus AddrInst must not be added in the matched instructions. | |||
4467 | /// This state can happen when AddrInst is a sext, since it may be moved away. | |||
4468 | /// Therefore, AddrInst may not be valid when MovedAway is true and it must | |||
4469 | /// not be referenced anymore. | |||
4470 | bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode, | |||
4471 | unsigned Depth, | |||
4472 | bool *MovedAway) { | |||
4473 | // Avoid exponential behavior on extremely deep expression trees. | |||
4474 | if (Depth >= 5) return false; | |||
4475 | ||||
4476 | // By default, all matched instructions stay in place. | |||
4477 | if (MovedAway) | |||
4478 | *MovedAway = false; | |||
4479 | ||||
4480 | switch (Opcode) { | |||
4481 | case Instruction::PtrToInt: | |||
4482 | // PtrToInt is always a noop, as we know that the int type is pointer sized. | |||
4483 | return matchAddr(AddrInst->getOperand(0), Depth); | |||
4484 | case Instruction::IntToPtr: { | |||
4485 | auto AS = AddrInst->getType()->getPointerAddressSpace(); | |||
4486 | auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS)); | |||
4487 | // This inttoptr is a no-op if the integer type is pointer sized. | |||
4488 | if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy) | |||
4489 | return matchAddr(AddrInst->getOperand(0), Depth); | |||
4490 | return false; | |||
4491 | } | |||
4492 | case Instruction::BitCast: | |||
4493 | // BitCast is always a noop, and we can handle it as long as it is | |||
4494 | // int->int or pointer->pointer (we don't want int<->fp or something). | |||
4495 | if (AddrInst->getOperand(0)->getType()->isIntOrPtrTy() && | |||
4496 | // Don't touch identity bitcasts. These were probably put here by LSR, | |||
4497 | // and we don't want to mess around with them. Assume it knows what it | |||
4498 | // is doing. | |||
4499 | AddrInst->getOperand(0)->getType() != AddrInst->getType()) | |||
4500 | return matchAddr(AddrInst->getOperand(0), Depth); | |||
4501 | return false; | |||
4502 | case Instruction::AddrSpaceCast: { | |||
4503 | unsigned SrcAS | |||
4504 | = AddrInst->getOperand(0)->getType()->getPointerAddressSpace(); | |||
4505 | unsigned DestAS = AddrInst->getType()->getPointerAddressSpace(); | |||
4506 | if (TLI.getTargetMachine().isNoopAddrSpaceCast(SrcAS, DestAS)) | |||
4507 | return matchAddr(AddrInst->getOperand(0), Depth); | |||
4508 | return false; | |||
4509 | } | |||
4510 | case Instruction::Add: { | |||
4511 | // Check to see if we can merge in the RHS then the LHS. If so, we win. | |||
4512 | ExtAddrMode BackupAddrMode = AddrMode; | |||
4513 | unsigned OldSize = AddrModeInsts.size(); | |||
4514 | // Start a transaction at this point. | |||
4515 | // The LHS may match but not the RHS. | |||
4516 | // Therefore, we need a higher level restoration point to undo partially | |||
4517 | // matched operation. | |||
4518 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | |||
4519 | TPT.getRestorationPoint(); | |||
4520 | ||||
4521 | AddrMode.InBounds = false; | |||
4522 | if (matchAddr(AddrInst->getOperand(1), Depth+1) && | |||
4523 | matchAddr(AddrInst->getOperand(0), Depth+1)) | |||
4524 | return true; | |||
4525 | ||||
4526 | // Restore the old addr mode info. | |||
4527 | AddrMode = BackupAddrMode; | |||
4528 | AddrModeInsts.resize(OldSize); | |||
4529 | TPT.rollback(LastKnownGood); | |||
4530 | ||||
4531 | // Otherwise this was over-aggressive. Try merging in the LHS then the RHS. | |||
4532 | if (matchAddr(AddrInst->getOperand(0), Depth+1) && | |||
4533 | matchAddr(AddrInst->getOperand(1), Depth+1)) | |||
4534 | return true; | |||
4535 | ||||
4536 | // Otherwise we definitely can't merge the ADD in. | |||
4537 | AddrMode = BackupAddrMode; | |||
4538 | AddrModeInsts.resize(OldSize); | |||
4539 | TPT.rollback(LastKnownGood); | |||
4540 | break; | |||
4541 | } | |||
4542 | //case Instruction::Or: | |||
4543 | // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD. | |||
4544 | //break; | |||
4545 | case Instruction::Mul: | |||
4546 | case Instruction::Shl: { | |||
4547 | // Can only handle X*C and X << C. | |||
4548 | AddrMode.InBounds = false; | |||
4549 | ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1)); | |||
4550 | if (!RHS || RHS->getBitWidth() > 64) | |||
4551 | return false; | |||
4552 | int64_t Scale = Opcode == Instruction::Shl | |||
4553 | ? 1LL << RHS->getLimitedValue(RHS->getBitWidth() - 1) | |||
4554 | : RHS->getSExtValue(); | |||
4555 | ||||
4556 | return matchScaledValue(AddrInst->getOperand(0), Scale, Depth); | |||
4557 | } | |||
4558 | case Instruction::GetElementPtr: { | |||
4559 | // Scan the GEP. We check it if it contains constant offsets and at most | |||
4560 | // one variable offset. | |||
4561 | int VariableOperand = -1; | |||
4562 | unsigned VariableScale = 0; | |||
4563 | ||||
4564 | int64_t ConstantOffset = 0; | |||
4565 | gep_type_iterator GTI = gep_type_begin(AddrInst); | |||
4566 | for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) { | |||
4567 | if (StructType *STy = GTI.getStructTypeOrNull()) { | |||
4568 | const StructLayout *SL = DL.getStructLayout(STy); | |||
4569 | unsigned Idx = | |||
4570 | cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue(); | |||
4571 | ConstantOffset += SL->getElementOffset(Idx); | |||
4572 | } else { | |||
4573 | TypeSize TS = DL.getTypeAllocSize(GTI.getIndexedType()); | |||
4574 | if (TS.isNonZero()) { | |||
4575 | // The optimisations below currently only work for fixed offsets. | |||
4576 | if (TS.isScalable()) | |||
4577 | return false; | |||
4578 | int64_t TypeSize = TS.getFixedSize(); | |||
4579 | if (ConstantInt *CI = | |||
4580 | dyn_cast<ConstantInt>(AddrInst->getOperand(i))) { | |||
4581 | const APInt &CVal = CI->getValue(); | |||
4582 | if (CVal.getMinSignedBits() <= 64) { | |||
4583 | ConstantOffset += CVal.getSExtValue() * TypeSize; | |||
4584 | continue; | |||
4585 | } | |||
4586 | } | |||
4587 | // We only allow one variable index at the moment. | |||
4588 | if (VariableOperand != -1) | |||
4589 | return false; | |||
4590 | ||||
4591 | // Remember the variable index. | |||
4592 | VariableOperand = i; | |||
4593 | VariableScale = TypeSize; | |||
4594 | } | |||
4595 | } | |||
4596 | } | |||
4597 | ||||
4598 | // A common case is for the GEP to only do a constant offset. In this case, | |||
4599 | // just add it to the disp field and check validity. | |||
4600 | if (VariableOperand == -1) { | |||
4601 | AddrMode.BaseOffs += ConstantOffset; | |||
4602 | if (ConstantOffset == 0 || | |||
4603 | TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) { | |||
4604 | // Check to see if we can fold the base pointer in too. | |||
4605 | if (matchAddr(AddrInst->getOperand(0), Depth+1)) { | |||
4606 | if (!cast<GEPOperator>(AddrInst)->isInBounds()) | |||
4607 | AddrMode.InBounds = false; | |||
4608 | return true; | |||
4609 | } | |||
4610 | } else if (EnableGEPOffsetSplit && isa<GetElementPtrInst>(AddrInst) && | |||
4611 | TLI.shouldConsiderGEPOffsetSplit() && Depth == 0 && | |||
4612 | ConstantOffset > 0) { | |||
4613 | // Record GEPs with non-zero offsets as candidates for splitting in the | |||
4614 | // event that the offset cannot fit into the r+i addressing mode. | |||
4615 | // Simple and common case that only one GEP is used in calculating the | |||
4616 | // address for the memory access. | |||
4617 | Value *Base = AddrInst->getOperand(0); | |||
4618 | auto *BaseI = dyn_cast<Instruction>(Base); | |||
4619 | auto *GEP = cast<GetElementPtrInst>(AddrInst); | |||
4620 | if (isa<Argument>(Base) || isa<GlobalValue>(Base) || | |||
4621 | (BaseI && !isa<CastInst>(BaseI) && | |||
4622 | !isa<GetElementPtrInst>(BaseI))) { | |||
4623 | // Make sure the parent block allows inserting non-PHI instructions | |||
4624 | // before the terminator. | |||
4625 | BasicBlock *Parent = | |||
4626 | BaseI ? BaseI->getParent() : &GEP->getFunction()->getEntryBlock(); | |||
4627 | if (!Parent->getTerminator()->isEHPad()) | |||
4628 | LargeOffsetGEP = std::make_pair(GEP, ConstantOffset); | |||
4629 | } | |||
4630 | } | |||
4631 | AddrMode.BaseOffs -= ConstantOffset; | |||
4632 | return false; | |||
4633 | } | |||
4634 | ||||
4635 | // Save the valid addressing mode in case we can't match. | |||
4636 | ExtAddrMode BackupAddrMode = AddrMode; | |||
4637 | unsigned OldSize = AddrModeInsts.size(); | |||
4638 | ||||
4639 | // See if the scale and offset amount is valid for this target. | |||
4640 | AddrMode.BaseOffs += ConstantOffset; | |||
4641 | if (!cast<GEPOperator>(AddrInst)->isInBounds()) | |||
4642 | AddrMode.InBounds = false; | |||
4643 | ||||
4644 | // Match the base operand of the GEP. | |||
4645 | if (!matchAddr(AddrInst->getOperand(0), Depth+1)) { | |||
4646 | // If it couldn't be matched, just stuff the value in a register. | |||
4647 | if (AddrMode.HasBaseReg) { | |||
4648 | AddrMode = BackupAddrMode; | |||
4649 | AddrModeInsts.resize(OldSize); | |||
4650 | return false; | |||
4651 | } | |||
4652 | AddrMode.HasBaseReg = true; | |||
4653 | AddrMode.BaseReg = AddrInst->getOperand(0); | |||
4654 | } | |||
4655 | ||||
4656 | // Match the remaining variable portion of the GEP. | |||
4657 | if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale, | |||
4658 | Depth)) { | |||
4659 | // If it couldn't be matched, try stuffing the base into a register | |||
4660 | // instead of matching it, and retrying the match of the scale. | |||
4661 | AddrMode = BackupAddrMode; | |||
4662 | AddrModeInsts.resize(OldSize); | |||
4663 | if (AddrMode.HasBaseReg) | |||
4664 | return false; | |||
4665 | AddrMode.HasBaseReg = true; | |||
4666 | AddrMode.BaseReg = AddrInst->getOperand(0); | |||
4667 | AddrMode.BaseOffs += ConstantOffset; | |||
4668 | if (!matchScaledValue(AddrInst->getOperand(VariableOperand), | |||
4669 | VariableScale, Depth)) { | |||
4670 | // If even that didn't work, bail. | |||
4671 | AddrMode = BackupAddrMode; | |||
4672 | AddrModeInsts.resize(OldSize); | |||
4673 | return false; | |||
4674 | } | |||
4675 | } | |||
4676 | ||||
4677 | return true; | |||
4678 | } | |||
4679 | case Instruction::SExt: | |||
4680 | case Instruction::ZExt: { | |||
4681 | Instruction *Ext = dyn_cast<Instruction>(AddrInst); | |||
4682 | if (!Ext) | |||
4683 | return false; | |||
4684 | ||||
4685 | // Try to move this ext out of the way of the addressing mode. | |||
4686 | // Ask for a method for doing so. | |||
4687 | TypePromotionHelper::Action TPH = | |||
4688 | TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts); | |||
4689 | if (!TPH) | |||
4690 | return false; | |||
4691 | ||||
4692 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | |||
4693 | TPT.getRestorationPoint(); | |||
4694 | unsigned CreatedInstsCost = 0; | |||
4695 | unsigned ExtCost = !TLI.isExtFree(Ext); | |||
4696 | Value *PromotedOperand = | |||
4697 | TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI); | |||
4698 | // SExt has been moved away. | |||
4699 | // Thus either it will be rematched later in the recursive calls or it is | |||
4700 | // gone. Anyway, we must not fold it into the addressing mode at this point. | |||
4701 | // E.g., | |||
4702 | // op = add opnd, 1 | |||
4703 | // idx = ext op | |||
4704 | // addr = gep base, idx | |||
4705 | // is now: | |||
4706 | // promotedOpnd = ext opnd <- no match here | |||
4707 | // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls) | |||
4708 | // addr = gep base, op <- match | |||
4709 | if (MovedAway) | |||
4710 | *MovedAway = true; | |||
4711 | ||||
4712 | assert(PromotedOperand &&(static_cast <bool> (PromotedOperand && "TypePromotionHelper should have filtered out those cases" ) ? void (0) : __assert_fail ("PromotedOperand && \"TypePromotionHelper should have filtered out those cases\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 4713, __extension__ __PRETTY_FUNCTION__ )) | |||
4713 | "TypePromotionHelper should have filtered out those cases")(static_cast <bool> (PromotedOperand && "TypePromotionHelper should have filtered out those cases" ) ? void (0) : __assert_fail ("PromotedOperand && \"TypePromotionHelper should have filtered out those cases\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 4713, __extension__ __PRETTY_FUNCTION__ )); | |||
4714 | ||||
4715 | ExtAddrMode BackupAddrMode = AddrMode; | |||
4716 | unsigned OldSize = AddrModeInsts.size(); | |||
4717 | ||||
4718 | if (!matchAddr(PromotedOperand, Depth) || | |||
4719 | // The total of the new cost is equal to the cost of the created | |||
4720 | // instructions. | |||
4721 | // The total of the old cost is equal to the cost of the extension plus | |||
4722 | // what we have saved in the addressing mode. | |||
4723 | !isPromotionProfitable(CreatedInstsCost, | |||
4724 | ExtCost + (AddrModeInsts.size() - OldSize), | |||
4725 | PromotedOperand)) { | |||
4726 | AddrMode = BackupAddrMode; | |||
4727 | AddrModeInsts.resize(OldSize); | |||
4728 | LLVM_DEBUG(dbgs() << "Sign extension does not pay off: rollback\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Sign extension does not pay off: rollback\n" ; } } while (false); | |||
4729 | TPT.rollback(LastKnownGood); | |||
4730 | return false; | |||
4731 | } | |||
4732 | return true; | |||
4733 | } | |||
4734 | } | |||
4735 | return false; | |||
4736 | } | |||
4737 | ||||
4738 | /// If we can, try to add the value of 'Addr' into the current addressing mode. | |||
4739 | /// If Addr can't be added to AddrMode this returns false and leaves AddrMode | |||
4740 | /// unmodified. This assumes that Addr is either a pointer type or intptr_t | |||
4741 | /// for the target. | |||
4742 | /// | |||
4743 | bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) { | |||
4744 | // Start a transaction at this point that we will rollback if the matching | |||
4745 | // fails. | |||
4746 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | |||
4747 | TPT.getRestorationPoint(); | |||
4748 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) { | |||
4749 | if (CI->getValue().isSignedIntN(64)) { | |||
4750 | // Fold in immediates if legal for the target. | |||
4751 | AddrMode.BaseOffs += CI->getSExtValue(); | |||
4752 | if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) | |||
4753 | return true; | |||
4754 | AddrMode.BaseOffs -= CI->getSExtValue(); | |||
4755 | } | |||
4756 | } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) { | |||
4757 | // If this is a global variable, try to fold it into the addressing mode. | |||
4758 | if (!AddrMode.BaseGV) { | |||
4759 | AddrMode.BaseGV = GV; | |||
4760 | if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) | |||
4761 | return true; | |||
4762 | AddrMode.BaseGV = nullptr; | |||
4763 | } | |||
4764 | } else if (Instruction *I = dyn_cast<Instruction>(Addr)) { | |||
4765 | ExtAddrMode BackupAddrMode = AddrMode; | |||
4766 | unsigned OldSize = AddrModeInsts.size(); | |||
4767 | ||||
4768 | // Check to see if it is possible to fold this operation. | |||
4769 | bool MovedAway = false; | |||
4770 | if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) { | |||
4771 | // This instruction may have been moved away. If so, there is nothing | |||
4772 | // to check here. | |||
4773 | if (MovedAway) | |||
4774 | return true; | |||
4775 | // Okay, it's possible to fold this. Check to see if it is actually | |||
4776 | // *profitable* to do so. We use a simple cost model to avoid increasing | |||
4777 | // register pressure too much. | |||
4778 | if (I->hasOneUse() || | |||
4779 | isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) { | |||
4780 | AddrModeInsts.push_back(I); | |||
4781 | return true; | |||
4782 | } | |||
4783 | ||||
4784 | // It isn't profitable to do this, roll back. | |||
4785 | //cerr << "NOT FOLDING: " << *I; | |||
4786 | AddrMode = BackupAddrMode; | |||
4787 | AddrModeInsts.resize(OldSize); | |||
4788 | TPT.rollback(LastKnownGood); | |||
4789 | } | |||
4790 | } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) { | |||
4791 | if (matchOperationAddr(CE, CE->getOpcode(), Depth)) | |||
4792 | return true; | |||
4793 | TPT.rollback(LastKnownGood); | |||
4794 | } else if (isa<ConstantPointerNull>(Addr)) { | |||
4795 | // Null pointer gets folded without affecting the addressing mode. | |||
4796 | return true; | |||
4797 | } | |||
4798 | ||||
4799 | // Worse case, the target should support [reg] addressing modes. :) | |||
4800 | if (!AddrMode.HasBaseReg) { | |||
4801 | AddrMode.HasBaseReg = true; | |||
4802 | AddrMode.BaseReg = Addr; | |||
4803 | // Still check for legality in case the target supports [imm] but not [i+r]. | |||
4804 | if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) | |||
4805 | return true; | |||
4806 | AddrMode.HasBaseReg = false; | |||
4807 | AddrMode.BaseReg = nullptr; | |||
4808 | } | |||
4809 | ||||
4810 | // If the base register is already taken, see if we can do [r+r]. | |||
4811 | if (AddrMode.Scale == 0) { | |||
4812 | AddrMode.Scale = 1; | |||
4813 | AddrMode.ScaledReg = Addr; | |||
4814 | if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) | |||
4815 | return true; | |||
4816 | AddrMode.Scale = 0; | |||
4817 | AddrMode.ScaledReg = nullptr; | |||
4818 | } | |||
4819 | // Couldn't match. | |||
4820 | TPT.rollback(LastKnownGood); | |||
4821 | return false; | |||
4822 | } | |||
4823 | ||||
4824 | /// Check to see if all uses of OpVal by the specified inline asm call are due | |||
4825 | /// to memory operands. If so, return true, otherwise return false. | |||
4826 | static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal, | |||
4827 | const TargetLowering &TLI, | |||
4828 | const TargetRegisterInfo &TRI) { | |||
4829 | const Function *F = CI->getFunction(); | |||
4830 | TargetLowering::AsmOperandInfoVector TargetConstraints = | |||
4831 | TLI.ParseConstraints(F->getParent()->getDataLayout(), &TRI, *CI); | |||
4832 | ||||
4833 | for (TargetLowering::AsmOperandInfo &OpInfo : TargetConstraints) { | |||
4834 | // Compute the constraint code and ConstraintType to use. | |||
4835 | TLI.ComputeConstraintToUse(OpInfo, SDValue()); | |||
4836 | ||||
4837 | // If this asm operand is our Value*, and if it isn't an indirect memory | |||
4838 | // operand, we can't fold it! TODO: Also handle C_Address? | |||
4839 | if (OpInfo.CallOperandVal == OpVal && | |||
4840 | (OpInfo.ConstraintType != TargetLowering::C_Memory || | |||
4841 | !OpInfo.isIndirect)) | |||
4842 | return false; | |||
4843 | } | |||
4844 | ||||
4845 | return true; | |||
4846 | } | |||
4847 | ||||
4848 | // Max number of memory uses to look at before aborting the search to conserve | |||
4849 | // compile time. | |||
4850 | static constexpr int MaxMemoryUsesToScan = 20; | |||
4851 | ||||
4852 | /// Recursively walk all the uses of I until we find a memory use. | |||
4853 | /// If we find an obviously non-foldable instruction, return true. | |||
4854 | /// Add accessed addresses and types to MemoryUses. | |||
4855 | static bool FindAllMemoryUses( | |||
4856 | Instruction *I, SmallVectorImpl<std::pair<Value *, Type *>> &MemoryUses, | |||
4857 | SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI, | |||
4858 | const TargetRegisterInfo &TRI, bool OptSize, ProfileSummaryInfo *PSI, | |||
4859 | BlockFrequencyInfo *BFI, int SeenInsts = 0) { | |||
4860 | // If we already considered this instruction, we're done. | |||
4861 | if (!ConsideredInsts.insert(I).second) | |||
4862 | return false; | |||
4863 | ||||
4864 | // If this is an obviously unfoldable instruction, bail out. | |||
4865 | if (!MightBeFoldableInst(I)) | |||
4866 | return true; | |||
4867 | ||||
4868 | // Loop over all the uses, recursively processing them. | |||
4869 | for (Use &U : I->uses()) { | |||
4870 | // Conservatively return true if we're seeing a large number or a deep chain | |||
4871 | // of users. This avoids excessive compilation times in pathological cases. | |||
4872 | if (SeenInsts++ >= MaxMemoryUsesToScan) | |||
4873 | return true; | |||
4874 | ||||
4875 | Instruction *UserI = cast<Instruction>(U.getUser()); | |||
4876 | if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) { | |||
4877 | MemoryUses.push_back({U.get(), LI->getType()}); | |||
4878 | continue; | |||
4879 | } | |||
4880 | ||||
4881 | if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) { | |||
4882 | if (U.getOperandNo() != StoreInst::getPointerOperandIndex()) | |||
4883 | return true; // Storing addr, not into addr. | |||
4884 | MemoryUses.push_back({U.get(), SI->getValueOperand()->getType()}); | |||
4885 | continue; | |||
4886 | } | |||
4887 | ||||
4888 | if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) { | |||
4889 | if (U.getOperandNo() != AtomicRMWInst::getPointerOperandIndex()) | |||
4890 | return true; // Storing addr, not into addr. | |||
4891 | MemoryUses.push_back({U.get(), RMW->getValOperand()->getType()}); | |||
4892 | continue; | |||
4893 | } | |||
4894 | ||||
4895 | if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) { | |||
4896 | if (U.getOperandNo() != AtomicCmpXchgInst::getPointerOperandIndex()) | |||
4897 | return true; // Storing addr, not into addr. | |||
4898 | MemoryUses.push_back({U.get(), CmpX->getCompareOperand()->getType()}); | |||
4899 | continue; | |||
4900 | } | |||
4901 | ||||
4902 | if (CallInst *CI = dyn_cast<CallInst>(UserI)) { | |||
4903 | if (CI->hasFnAttr(Attribute::Cold)) { | |||
4904 | // If this is a cold call, we can sink the addressing calculation into | |||
4905 | // the cold path. See optimizeCallInst | |||
4906 | bool OptForSize = OptSize || | |||
4907 | llvm::shouldOptimizeForSize(CI->getParent(), PSI, BFI); | |||
4908 | if (!OptForSize) | |||
4909 | continue; | |||
4910 | } | |||
4911 | ||||
4912 | InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledOperand()); | |||
4913 | if (!IA) return true; | |||
4914 | ||||
4915 | // If this is a memory operand, we're cool, otherwise bail out. | |||
4916 | if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI)) | |||
4917 | return true; | |||
4918 | continue; | |||
4919 | } | |||
4920 | ||||
4921 | if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI, OptSize, | |||
4922 | PSI, BFI, SeenInsts)) | |||
4923 | return true; | |||
4924 | } | |||
4925 | ||||
4926 | return false; | |||
4927 | } | |||
4928 | ||||
4929 | /// Return true if Val is already known to be live at the use site that we're | |||
4930 | /// folding it into. If so, there is no cost to include it in the addressing | |||
4931 | /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the | |||
4932 | /// instruction already. | |||
4933 | bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1, | |||
4934 | Value *KnownLive2) { | |||
4935 | // If Val is either of the known-live values, we know it is live! | |||
4936 | if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2) | |||
4937 | return true; | |||
4938 | ||||
4939 | // All values other than instructions and arguments (e.g. constants) are live. | |||
4940 | if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true; | |||
4941 | ||||
4942 | // If Val is a constant sized alloca in the entry block, it is live, this is | |||
4943 | // true because it is just a reference to the stack/frame pointer, which is | |||
4944 | // live for the whole function. | |||
4945 | if (AllocaInst *AI = dyn_cast<AllocaInst>(Val)) | |||
4946 | if (AI->isStaticAlloca()) | |||
4947 | return true; | |||
4948 | ||||
4949 | // Check to see if this value is already used in the memory instruction's | |||
4950 | // block. If so, it's already live into the block at the very least, so we | |||
4951 | // can reasonably fold it. | |||
4952 | return Val->isUsedInBasicBlock(MemoryInst->getParent()); | |||
4953 | } | |||
4954 | ||||
4955 | /// It is possible for the addressing mode of the machine to fold the specified | |||
4956 | /// instruction into a load or store that ultimately uses it. | |||
4957 | /// However, the specified instruction has multiple uses. | |||
4958 | /// Given this, it may actually increase register pressure to fold it | |||
4959 | /// into the load. For example, consider this code: | |||
4960 | /// | |||
4961 | /// X = ... | |||
4962 | /// Y = X+1 | |||
4963 | /// use(Y) -> nonload/store | |||
4964 | /// Z = Y+1 | |||
4965 | /// load Z | |||
4966 | /// | |||
4967 | /// In this case, Y has multiple uses, and can be folded into the load of Z | |||
4968 | /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to | |||
4969 | /// be live at the use(Y) line. If we don't fold Y into load Z, we use one | |||
4970 | /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the | |||
4971 | /// number of computations either. | |||
4972 | /// | |||
4973 | /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If | |||
4974 | /// X was live across 'load Z' for other reasons, we actually *would* want to | |||
4975 | /// fold the addressing mode in the Z case. This would make Y die earlier. | |||
4976 | bool AddressingModeMatcher:: | |||
4977 | isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore, | |||
4978 | ExtAddrMode &AMAfter) { | |||
4979 | if (IgnoreProfitability) return true; | |||
4980 | ||||
4981 | // AMBefore is the addressing mode before this instruction was folded into it, | |||
4982 | // and AMAfter is the addressing mode after the instruction was folded. Get | |||
4983 | // the set of registers referenced by AMAfter and subtract out those | |||
4984 | // referenced by AMBefore: this is the set of values which folding in this | |||
4985 | // address extends the lifetime of. | |||
4986 | // | |||
4987 | // Note that there are only two potential values being referenced here, | |||
4988 | // BaseReg and ScaleReg (global addresses are always available, as are any | |||
4989 | // folded immediates). | |||
4990 | Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg; | |||
4991 | ||||
4992 | // If the BaseReg or ScaledReg was referenced by the previous addrmode, their | |||
4993 | // lifetime wasn't extended by adding this instruction. | |||
4994 | if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg)) | |||
4995 | BaseReg = nullptr; | |||
4996 | if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg)) | |||
4997 | ScaledReg = nullptr; | |||
4998 | ||||
4999 | // If folding this instruction (and it's subexprs) didn't extend any live | |||
5000 | // ranges, we're ok with it. | |||
5001 | if (!BaseReg && !ScaledReg) | |||
5002 | return true; | |||
5003 | ||||
5004 | // If all uses of this instruction can have the address mode sunk into them, | |||
5005 | // we can remove the addressing mode and effectively trade one live register | |||
5006 | // for another (at worst.) In this context, folding an addressing mode into | |||
5007 | // the use is just a particularly nice way of sinking it. | |||
5008 | SmallVector<std::pair<Value *, Type *>, 16> MemoryUses; | |||
5009 | SmallPtrSet<Instruction*, 16> ConsideredInsts; | |||
5010 | if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI, OptSize, | |||
5011 | PSI, BFI)) | |||
5012 | return false; // Has a non-memory, non-foldable use! | |||
5013 | ||||
5014 | // Now that we know that all uses of this instruction are part of a chain of | |||
5015 | // computation involving only operations that could theoretically be folded | |||
5016 | // into a memory use, loop over each of these memory operation uses and see | |||
5017 | // if they could *actually* fold the instruction. The assumption is that | |||
5018 | // addressing modes are cheap and that duplicating the computation involved | |||
5019 | // many times is worthwhile, even on a fastpath. For sinking candidates | |||
5020 | // (i.e. cold call sites), this serves as a way to prevent excessive code | |||
5021 | // growth since most architectures have some reasonable small and fast way to | |||
5022 | // compute an effective address. (i.e LEA on x86) | |||
5023 | SmallVector<Instruction*, 32> MatchedAddrModeInsts; | |||
5024 | for (const std::pair<Value *, Type *> &Pair : MemoryUses) { | |||
5025 | Value *Address = Pair.first; | |||
5026 | Type *AddressAccessTy = Pair.second; | |||
5027 | unsigned AS = Address->getType()->getPointerAddressSpace(); | |||
5028 | ||||
5029 | // Do a match against the root of this address, ignoring profitability. This | |||
5030 | // will tell us if the addressing mode for the memory operation will | |||
5031 | // *actually* cover the shared instruction. | |||
5032 | ExtAddrMode Result; | |||
5033 | std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr, | |||
5034 | 0); | |||
5035 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | |||
5036 | TPT.getRestorationPoint(); | |||
5037 | AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, TRI, LI, getDTFn, | |||
5038 | AddressAccessTy, AS, MemoryInst, Result, | |||
5039 | InsertedInsts, PromotedInsts, TPT, | |||
5040 | LargeOffsetGEP, OptSize, PSI, BFI); | |||
5041 | Matcher.IgnoreProfitability = true; | |||
5042 | bool Success = Matcher.matchAddr(Address, 0); | |||
5043 | (void)Success; assert(Success && "Couldn't select *anything*?")(static_cast <bool> (Success && "Couldn't select *anything*?" ) ? void (0) : __assert_fail ("Success && \"Couldn't select *anything*?\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5043, __extension__ __PRETTY_FUNCTION__ )); | |||
5044 | ||||
5045 | // The match was to check the profitability, the changes made are not | |||
5046 | // part of the original matcher. Therefore, they should be dropped | |||
5047 | // otherwise the original matcher will not present the right state. | |||
5048 | TPT.rollback(LastKnownGood); | |||
5049 | ||||
5050 | // If the match didn't cover I, then it won't be shared by it. | |||
5051 | if (!is_contained(MatchedAddrModeInsts, I)) | |||
5052 | return false; | |||
5053 | ||||
5054 | MatchedAddrModeInsts.clear(); | |||
5055 | } | |||
5056 | ||||
5057 | return true; | |||
5058 | } | |||
5059 | ||||
5060 | /// Return true if the specified values are defined in a | |||
5061 | /// different basic block than BB. | |||
5062 | static bool IsNonLocalValue(Value *V, BasicBlock *BB) { | |||
5063 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
5064 | return I->getParent() != BB; | |||
5065 | return false; | |||
5066 | } | |||
5067 | ||||
5068 | /// Sink addressing mode computation immediate before MemoryInst if doing so | |||
5069 | /// can be done without increasing register pressure. The need for the | |||
5070 | /// register pressure constraint means this can end up being an all or nothing | |||
5071 | /// decision for all uses of the same addressing computation. | |||
5072 | /// | |||
5073 | /// Load and Store Instructions often have addressing modes that can do | |||
5074 | /// significant amounts of computation. As such, instruction selection will try | |||
5075 | /// to get the load or store to do as much computation as possible for the | |||
5076 | /// program. The problem is that isel can only see within a single block. As | |||
5077 | /// such, we sink as much legal addressing mode work into the block as possible. | |||
5078 | /// | |||
5079 | /// This method is used to optimize both load/store and inline asms with memory | |||
5080 | /// operands. It's also used to sink addressing computations feeding into cold | |||
5081 | /// call sites into their (cold) basic block. | |||
5082 | /// | |||
5083 | /// The motivation for handling sinking into cold blocks is that doing so can | |||
5084 | /// both enable other address mode sinking (by satisfying the register pressure | |||
5085 | /// constraint above), and reduce register pressure globally (by removing the | |||
5086 | /// addressing mode computation from the fast path entirely.). | |||
5087 | bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr, | |||
5088 | Type *AccessTy, unsigned AddrSpace) { | |||
5089 | Value *Repl = Addr; | |||
5090 | ||||
5091 | // Try to collapse single-value PHI nodes. This is necessary to undo | |||
5092 | // unprofitable PRE transformations. | |||
5093 | SmallVector<Value*, 8> worklist; | |||
5094 | SmallPtrSet<Value*, 16> Visited; | |||
5095 | worklist.push_back(Addr); | |||
5096 | ||||
5097 | // Use a worklist to iteratively look through PHI and select nodes, and | |||
5098 | // ensure that the addressing mode obtained from the non-PHI/select roots of | |||
5099 | // the graph are compatible. | |||
5100 | bool PhiOrSelectSeen = false; | |||
5101 | SmallVector<Instruction*, 16> AddrModeInsts; | |||
5102 | const SimplifyQuery SQ(*DL, TLInfo); | |||
5103 | AddressingModeCombiner AddrModes(SQ, Addr); | |||
5104 | TypePromotionTransaction TPT(RemovedInsts); | |||
5105 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | |||
5106 | TPT.getRestorationPoint(); | |||
5107 | while (!worklist.empty()) { | |||
5108 | Value *V = worklist.pop_back_val(); | |||
5109 | ||||
5110 | // We allow traversing cyclic Phi nodes. | |||
5111 | // In case of success after this loop we ensure that traversing through | |||
5112 | // Phi nodes ends up with all cases to compute address of the form | |||
5113 | // BaseGV + Base + Scale * Index + Offset | |||
5114 | // where Scale and Offset are constans and BaseGV, Base and Index | |||
5115 | // are exactly the same Values in all cases. | |||
5116 | // It means that BaseGV, Scale and Offset dominate our memory instruction | |||
5117 | // and have the same value as they had in address computation represented | |||
5118 | // as Phi. So we can safely sink address computation to memory instruction. | |||
5119 | if (!Visited.insert(V).second) | |||
5120 | continue; | |||
5121 | ||||
5122 | // For a PHI node, push all of its incoming values. | |||
5123 | if (PHINode *P = dyn_cast<PHINode>(V)) { | |||
5124 | append_range(worklist, P->incoming_values()); | |||
5125 | PhiOrSelectSeen = true; | |||
5126 | continue; | |||
5127 | } | |||
5128 | // Similar for select. | |||
5129 | if (SelectInst *SI = dyn_cast<SelectInst>(V)) { | |||
5130 | worklist.push_back(SI->getFalseValue()); | |||
5131 | worklist.push_back(SI->getTrueValue()); | |||
5132 | PhiOrSelectSeen = true; | |||
5133 | continue; | |||
5134 | } | |||
5135 | ||||
5136 | // For non-PHIs, determine the addressing mode being computed. Note that | |||
5137 | // the result may differ depending on what other uses our candidate | |||
5138 | // addressing instructions might have. | |||
5139 | AddrModeInsts.clear(); | |||
5140 | std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr, | |||
5141 | 0); | |||
5142 | // Defer the query (and possible computation of) the dom tree to point of | |||
5143 | // actual use. It's expected that most address matches don't actually need | |||
5144 | // the domtree. | |||
5145 | auto getDTFn = [MemoryInst, this]() -> const DominatorTree & { | |||
5146 | Function *F = MemoryInst->getParent()->getParent(); | |||
5147 | return this->getDT(*F); | |||
5148 | }; | |||
5149 | ExtAddrMode NewAddrMode = AddressingModeMatcher::Match( | |||
5150 | V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *LI, getDTFn, | |||
5151 | *TRI, InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP, OptSize, PSI, | |||
5152 | BFI.get()); | |||
5153 | ||||
5154 | GetElementPtrInst *GEP = LargeOffsetGEP.first; | |||
5155 | if (GEP && !NewGEPBases.count(GEP)) { | |||
5156 | // If splitting the underlying data structure can reduce the offset of a | |||
5157 | // GEP, collect the GEP. Skip the GEPs that are the new bases of | |||
5158 | // previously split data structures. | |||
5159 | LargeOffsetGEPMap[GEP->getPointerOperand()].push_back(LargeOffsetGEP); | |||
5160 | if (LargeOffsetGEPID.find(GEP) == LargeOffsetGEPID.end()) | |||
5161 | LargeOffsetGEPID[GEP] = LargeOffsetGEPID.size(); | |||
5162 | } | |||
5163 | ||||
5164 | NewAddrMode.OriginalValue = V; | |||
5165 | if (!AddrModes.addNewAddrMode(NewAddrMode)) | |||
5166 | break; | |||
5167 | } | |||
5168 | ||||
5169 | // Try to combine the AddrModes we've collected. If we couldn't collect any, | |||
5170 | // or we have multiple but either couldn't combine them or combining them | |||
5171 | // wouldn't do anything useful, bail out now. | |||
5172 | if (!AddrModes.combineAddrModes()) { | |||
5173 | TPT.rollback(LastKnownGood); | |||
5174 | return false; | |||
5175 | } | |||
5176 | bool Modified = TPT.commit(); | |||
5177 | ||||
5178 | // Get the combined AddrMode (or the only AddrMode, if we only had one). | |||
5179 | ExtAddrMode AddrMode = AddrModes.getAddrMode(); | |||
5180 | ||||
5181 | // If all the instructions matched are already in this BB, don't do anything. | |||
5182 | // If we saw a Phi node then it is not local definitely, and if we saw a select | |||
5183 | // then we want to push the address calculation past it even if it's already | |||
5184 | // in this BB. | |||
5185 | if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) { | |||
5186 | return IsNonLocalValue(V, MemoryInst->getParent()); | |||
5187 | })) { | |||
5188 | LLVM_DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrModedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n"; } } while (false) | |||
5189 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n"; } } while (false); | |||
5190 | return Modified; | |||
5191 | } | |||
5192 | ||||
5193 | // Insert this computation right after this user. Since our caller is | |||
5194 | // scanning from the top of the BB to the bottom, reuse of the expr are | |||
5195 | // guaranteed to happen later. | |||
5196 | IRBuilder<> Builder(MemoryInst); | |||
5197 | ||||
5198 | // Now that we determined the addressing expression we want to use and know | |||
5199 | // that we have to sink it into this block. Check to see if we have already | |||
5200 | // done this for some other load/store instr in this block. If so, reuse | |||
5201 | // the computation. Before attempting reuse, check if the address is valid | |||
5202 | // as it may have been erased. | |||
5203 | ||||
5204 | WeakTrackingVH SunkAddrVH = SunkAddrs[Addr]; | |||
5205 | ||||
5206 | Value * SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr; | |||
5207 | if (SunkAddr) { | |||
5208 | LLVM_DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrModedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for " << *MemoryInst << "\n"; } } while (false) | |||
5209 | << " for " << *MemoryInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for " << *MemoryInst << "\n"; } } while (false); | |||
5210 | if (SunkAddr->getType() != Addr->getType()) | |||
5211 | SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType()); | |||
5212 | } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() && | |||
5213 | SubtargetInfo->addrSinkUsingGEPs())) { | |||
5214 | // By default, we use the GEP-based method when AA is used later. This | |||
5215 | // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities. | |||
5216 | LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrModedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " << *MemoryInst << "\n"; } } while (false) | |||
5217 | << " for " << *MemoryInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " << *MemoryInst << "\n"; } } while (false); | |||
5218 | Type *IntPtrTy = DL->getIntPtrType(Addr->getType()); | |||
5219 | Value *ResultPtr = nullptr, *ResultIndex = nullptr; | |||
5220 | ||||
5221 | // First, find the pointer. | |||
5222 | if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) { | |||
5223 | ResultPtr = AddrMode.BaseReg; | |||
5224 | AddrMode.BaseReg = nullptr; | |||
5225 | } | |||
5226 | ||||
5227 | if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) { | |||
5228 | // We can't add more than one pointer together, nor can we scale a | |||
5229 | // pointer (both of which seem meaningless). | |||
5230 | if (ResultPtr || AddrMode.Scale != 1) | |||
5231 | return Modified; | |||
5232 | ||||
5233 | ResultPtr = AddrMode.ScaledReg; | |||
5234 | AddrMode.Scale = 0; | |||
5235 | } | |||
5236 | ||||
5237 | // It is only safe to sign extend the BaseReg if we know that the math | |||
5238 | // required to create it did not overflow before we extend it. Since | |||
5239 | // the original IR value was tossed in favor of a constant back when | |||
5240 | // the AddrMode was created we need to bail out gracefully if widths | |||
5241 | // do not match instead of extending it. | |||
5242 | // | |||
5243 | // (See below for code to add the scale.) | |||
5244 | if (AddrMode.Scale) { | |||
5245 | Type *ScaledRegTy = AddrMode.ScaledReg->getType(); | |||
5246 | if (cast<IntegerType>(IntPtrTy)->getBitWidth() > | |||
5247 | cast<IntegerType>(ScaledRegTy)->getBitWidth()) | |||
5248 | return Modified; | |||
5249 | } | |||
5250 | ||||
5251 | if (AddrMode.BaseGV) { | |||
5252 | if (ResultPtr) | |||
5253 | return Modified; | |||
5254 | ||||
5255 | ResultPtr = AddrMode.BaseGV; | |||
5256 | } | |||
5257 | ||||
5258 | // If the real base value actually came from an inttoptr, then the matcher | |||
5259 | // will look through it and provide only the integer value. In that case, | |||
5260 | // use it here. | |||
5261 | if (!DL->isNonIntegralPointerType(Addr->getType())) { | |||
5262 | if (!ResultPtr && AddrMode.BaseReg) { | |||
5263 | ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), | |||
5264 | "sunkaddr"); | |||
5265 | AddrMode.BaseReg = nullptr; | |||
5266 | } else if (!ResultPtr && AddrMode.Scale == 1) { | |||
5267 | ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), | |||
5268 | "sunkaddr"); | |||
5269 | AddrMode.Scale = 0; | |||
5270 | } | |||
5271 | } | |||
5272 | ||||
5273 | if (!ResultPtr && | |||
5274 | !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) { | |||
5275 | SunkAddr = Constant::getNullValue(Addr->getType()); | |||
5276 | } else if (!ResultPtr) { | |||
5277 | return Modified; | |||
5278 | } else { | |||
5279 | Type *I8PtrTy = | |||
5280 | Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace()); | |||
5281 | Type *I8Ty = Builder.getInt8Ty(); | |||
5282 | ||||
5283 | // Start with the base register. Do this first so that subsequent address | |||
5284 | // matching finds it last, which will prevent it from trying to match it | |||
5285 | // as the scaled value in case it happens to be a mul. That would be | |||
5286 | // problematic if we've sunk a different mul for the scale, because then | |||
5287 | // we'd end up sinking both muls. | |||
5288 | if (AddrMode.BaseReg) { | |||
5289 | Value *V = AddrMode.BaseReg; | |||
5290 | if (V->getType() != IntPtrTy) | |||
5291 | V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr"); | |||
5292 | ||||
5293 | ResultIndex = V; | |||
5294 | } | |||
5295 | ||||
5296 | // Add the scale value. | |||
5297 | if (AddrMode.Scale) { | |||
5298 | Value *V = AddrMode.ScaledReg; | |||
5299 | if (V->getType() == IntPtrTy) { | |||
5300 | // done. | |||
5301 | } else { | |||
5302 | assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <(static_cast <bool> (cast<IntegerType>(IntPtrTy)-> getBitWidth() < cast<IntegerType>(V->getType())-> getBitWidth() && "We can't transform if ScaledReg is too narrow" ) ? void (0) : __assert_fail ("cast<IntegerType>(IntPtrTy)->getBitWidth() < cast<IntegerType>(V->getType())->getBitWidth() && \"We can't transform if ScaledReg is too narrow\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5304, __extension__ __PRETTY_FUNCTION__ )) | |||
5303 | cast<IntegerType>(V->getType())->getBitWidth() &&(static_cast <bool> (cast<IntegerType>(IntPtrTy)-> getBitWidth() < cast<IntegerType>(V->getType())-> getBitWidth() && "We can't transform if ScaledReg is too narrow" ) ? void (0) : __assert_fail ("cast<IntegerType>(IntPtrTy)->getBitWidth() < cast<IntegerType>(V->getType())->getBitWidth() && \"We can't transform if ScaledReg is too narrow\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5304, __extension__ __PRETTY_FUNCTION__ )) | |||
5304 | "We can't transform if ScaledReg is too narrow")(static_cast <bool> (cast<IntegerType>(IntPtrTy)-> getBitWidth() < cast<IntegerType>(V->getType())-> getBitWidth() && "We can't transform if ScaledReg is too narrow" ) ? void (0) : __assert_fail ("cast<IntegerType>(IntPtrTy)->getBitWidth() < cast<IntegerType>(V->getType())->getBitWidth() && \"We can't transform if ScaledReg is too narrow\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5304, __extension__ __PRETTY_FUNCTION__ )); | |||
5305 | V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr"); | |||
5306 | } | |||
5307 | ||||
5308 | if (AddrMode.Scale != 1) | |||
5309 | V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale), | |||
5310 | "sunkaddr"); | |||
5311 | if (ResultIndex) | |||
5312 | ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr"); | |||
5313 | else | |||
5314 | ResultIndex = V; | |||
5315 | } | |||
5316 | ||||
5317 | // Add in the Base Offset if present. | |||
5318 | if (AddrMode.BaseOffs) { | |||
5319 | Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); | |||
5320 | if (ResultIndex) { | |||
5321 | // We need to add this separately from the scale above to help with | |||
5322 | // SDAG consecutive load/store merging. | |||
5323 | if (ResultPtr->getType() != I8PtrTy) | |||
5324 | ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy); | |||
5325 | ResultPtr = | |||
5326 | AddrMode.InBounds | |||
5327 | ? Builder.CreateInBoundsGEP(I8Ty, ResultPtr, ResultIndex, | |||
5328 | "sunkaddr") | |||
5329 | : Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr"); | |||
5330 | } | |||
5331 | ||||
5332 | ResultIndex = V; | |||
5333 | } | |||
5334 | ||||
5335 | if (!ResultIndex) { | |||
5336 | SunkAddr = ResultPtr; | |||
5337 | } else { | |||
5338 | if (ResultPtr->getType() != I8PtrTy) | |||
5339 | ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy); | |||
5340 | SunkAddr = | |||
5341 | AddrMode.InBounds | |||
5342 | ? Builder.CreateInBoundsGEP(I8Ty, ResultPtr, ResultIndex, | |||
5343 | "sunkaddr") | |||
5344 | : Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr"); | |||
5345 | } | |||
5346 | ||||
5347 | if (SunkAddr->getType() != Addr->getType()) | |||
5348 | SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType()); | |||
5349 | } | |||
5350 | } else { | |||
5351 | // We'd require a ptrtoint/inttoptr down the line, which we can't do for | |||
5352 | // non-integral pointers, so in that case bail out now. | |||
5353 | Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr; | |||
5354 | Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr; | |||
5355 | PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy); | |||
5356 | PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy); | |||
5357 | if (DL->isNonIntegralPointerType(Addr->getType()) || | |||
5358 | (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) || | |||
5359 | (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) || | |||
5360 | (AddrMode.BaseGV && | |||
5361 | DL->isNonIntegralPointerType(AddrMode.BaseGV->getType()))) | |||
5362 | return Modified; | |||
5363 | ||||
5364 | LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrModedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " << *MemoryInst << "\n"; } } while (false) | |||
5365 | << " for " << *MemoryInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " << *MemoryInst << "\n"; } } while (false); | |||
5366 | Type *IntPtrTy = DL->getIntPtrType(Addr->getType()); | |||
5367 | Value *Result = nullptr; | |||
5368 | ||||
5369 | // Start with the base register. Do this first so that subsequent address | |||
5370 | // matching finds it last, which will prevent it from trying to match it | |||
5371 | // as the scaled value in case it happens to be a mul. That would be | |||
5372 | // problematic if we've sunk a different mul for the scale, because then | |||
5373 | // we'd end up sinking both muls. | |||
5374 | if (AddrMode.BaseReg) { | |||
5375 | Value *V = AddrMode.BaseReg; | |||
5376 | if (V->getType()->isPointerTy()) | |||
5377 | V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr"); | |||
5378 | if (V->getType() != IntPtrTy) | |||
5379 | V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr"); | |||
5380 | Result = V; | |||
5381 | } | |||
5382 | ||||
5383 | // Add the scale value. | |||
5384 | if (AddrMode.Scale) { | |||
5385 | Value *V = AddrMode.ScaledReg; | |||
5386 | if (V->getType() == IntPtrTy) { | |||
5387 | // done. | |||
5388 | } else if (V->getType()->isPointerTy()) { | |||
5389 | V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr"); | |||
5390 | } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < | |||
5391 | cast<IntegerType>(V->getType())->getBitWidth()) { | |||
5392 | V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr"); | |||
5393 | } else { | |||
5394 | // It is only safe to sign extend the BaseReg if we know that the math | |||
5395 | // required to create it did not overflow before we extend it. Since | |||
5396 | // the original IR value was tossed in favor of a constant back when | |||
5397 | // the AddrMode was created we need to bail out gracefully if widths | |||
5398 | // do not match instead of extending it. | |||
5399 | Instruction *I = dyn_cast_or_null<Instruction>(Result); | |||
5400 | if (I && (Result != AddrMode.BaseReg)) | |||
5401 | I->eraseFromParent(); | |||
5402 | return Modified; | |||
5403 | } | |||
5404 | if (AddrMode.Scale != 1) | |||
5405 | V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale), | |||
5406 | "sunkaddr"); | |||
5407 | if (Result) | |||
5408 | Result = Builder.CreateAdd(Result, V, "sunkaddr"); | |||
5409 | else | |||
5410 | Result = V; | |||
5411 | } | |||
5412 | ||||
5413 | // Add in the BaseGV if present. | |||
5414 | if (AddrMode.BaseGV) { | |||
5415 | Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr"); | |||
5416 | if (Result) | |||
5417 | Result = Builder.CreateAdd(Result, V, "sunkaddr"); | |||
5418 | else | |||
5419 | Result = V; | |||
5420 | } | |||
5421 | ||||
5422 | // Add in the Base Offset if present. | |||
5423 | if (AddrMode.BaseOffs) { | |||
5424 | Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); | |||
5425 | if (Result) | |||
5426 | Result = Builder.CreateAdd(Result, V, "sunkaddr"); | |||
5427 | else | |||
5428 | Result = V; | |||
5429 | } | |||
5430 | ||||
5431 | if (!Result) | |||
5432 | SunkAddr = Constant::getNullValue(Addr->getType()); | |||
5433 | else | |||
5434 | SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr"); | |||
5435 | } | |||
5436 | ||||
5437 | MemoryInst->replaceUsesOfWith(Repl, SunkAddr); | |||
5438 | // Store the newly computed address into the cache. In the case we reused a | |||
5439 | // value, this should be idempotent. | |||
5440 | SunkAddrs[Addr] = WeakTrackingVH(SunkAddr); | |||
5441 | ||||
5442 | // If we have no uses, recursively delete the value and all dead instructions | |||
5443 | // using it. | |||
5444 | if (Repl->use_empty()) { | |||
5445 | resetIteratorIfInvalidatedWhileCalling(CurInstIterator->getParent(), [&]() { | |||
5446 | RecursivelyDeleteTriviallyDeadInstructions( | |||
5447 | Repl, TLInfo, nullptr, | |||
5448 | [&](Value *V) { removeAllAssertingVHReferences(V); }); | |||
5449 | }); | |||
5450 | } | |||
5451 | ++NumMemoryInsts; | |||
5452 | return true; | |||
5453 | } | |||
5454 | ||||
5455 | /// Rewrite GEP input to gather/scatter to enable SelectionDAGBuilder to find | |||
5456 | /// a uniform base to use for ISD::MGATHER/MSCATTER. SelectionDAGBuilder can | |||
5457 | /// only handle a 2 operand GEP in the same basic block or a splat constant | |||
5458 | /// vector. The 2 operands to the GEP must have a scalar pointer and a vector | |||
5459 | /// index. | |||
5460 | /// | |||
5461 | /// If the existing GEP has a vector base pointer that is splat, we can look | |||
5462 | /// through the splat to find the scalar pointer. If we can't find a scalar | |||
5463 | /// pointer there's nothing we can do. | |||
5464 | /// | |||
5465 | /// If we have a GEP with more than 2 indices where the middle indices are all | |||
5466 | /// zeroes, we can replace it with 2 GEPs where the second has 2 operands. | |||
5467 | /// | |||
5468 | /// If the final index isn't a vector or is a splat, we can emit a scalar GEP | |||
5469 | /// followed by a GEP with an all zeroes vector index. This will enable | |||
5470 | /// SelectionDAGBuilder to use the scalar GEP as the uniform base and have a | |||
5471 | /// zero index. | |||
5472 | bool CodeGenPrepare::optimizeGatherScatterInst(Instruction *MemoryInst, | |||
5473 | Value *Ptr) { | |||
5474 | Value *NewAddr; | |||
5475 | ||||
5476 | if (const auto *GEP = dyn_cast<GetElementPtrInst>(Ptr)) { | |||
5477 | // Don't optimize GEPs that don't have indices. | |||
5478 | if (!GEP->hasIndices()) | |||
5479 | return false; | |||
5480 | ||||
5481 | // If the GEP and the gather/scatter aren't in the same BB, don't optimize. | |||
5482 | // FIXME: We should support this by sinking the GEP. | |||
5483 | if (MemoryInst->getParent() != GEP->getParent()) | |||
5484 | return false; | |||
5485 | ||||
5486 | SmallVector<Value *, 2> Ops(GEP->operands()); | |||
5487 | ||||
5488 | bool RewriteGEP = false; | |||
5489 | ||||
5490 | if (Ops[0]->getType()->isVectorTy()) { | |||
5491 | Ops[0] = getSplatValue(Ops[0]); | |||
5492 | if (!Ops[0]) | |||
5493 | return false; | |||
5494 | RewriteGEP = true; | |||
5495 | } | |||
5496 | ||||
5497 | unsigned FinalIndex = Ops.size() - 1; | |||
5498 | ||||
5499 | // Ensure all but the last index is 0. | |||
5500 | // FIXME: This isn't strictly required. All that's required is that they are | |||
5501 | // all scalars or splats. | |||
5502 | for (unsigned i = 1; i < FinalIndex; ++i) { | |||
5503 | auto *C = dyn_cast<Constant>(Ops[i]); | |||
5504 | if (!C) | |||
5505 | return false; | |||
5506 | if (isa<VectorType>(C->getType())) | |||
5507 | C = C->getSplatValue(); | |||
5508 | auto *CI = dyn_cast_or_null<ConstantInt>(C); | |||
5509 | if (!CI || !CI->isZero()) | |||
5510 | return false; | |||
5511 | // Scalarize the index if needed. | |||
5512 | Ops[i] = CI; | |||
5513 | } | |||
5514 | ||||
5515 | // Try to scalarize the final index. | |||
5516 | if (Ops[FinalIndex]->getType()->isVectorTy()) { | |||
5517 | if (Value *V = getSplatValue(Ops[FinalIndex])) { | |||
5518 | auto *C = dyn_cast<ConstantInt>(V); | |||
5519 | // Don't scalarize all zeros vector. | |||
5520 | if (!C || !C->isZero()) { | |||
5521 | Ops[FinalIndex] = V; | |||
5522 | RewriteGEP = true; | |||
5523 | } | |||
5524 | } | |||
5525 | } | |||
5526 | ||||
5527 | // If we made any changes or the we have extra operands, we need to generate | |||
5528 | // new instructions. | |||
5529 | if (!RewriteGEP && Ops.size() == 2) | |||
5530 | return false; | |||
5531 | ||||
5532 | auto NumElts = cast<VectorType>(Ptr->getType())->getElementCount(); | |||
5533 | ||||
5534 | IRBuilder<> Builder(MemoryInst); | |||
5535 | ||||
5536 | Type *SourceTy = GEP->getSourceElementType(); | |||
5537 | Type *ScalarIndexTy = DL->getIndexType(Ops[0]->getType()->getScalarType()); | |||
5538 | ||||
5539 | // If the final index isn't a vector, emit a scalar GEP containing all ops | |||
5540 | // and a vector GEP with all zeroes final index. | |||
5541 | if (!Ops[FinalIndex]->getType()->isVectorTy()) { | |||
5542 | NewAddr = Builder.CreateGEP(SourceTy, Ops[0], | |||
5543 | makeArrayRef(Ops).drop_front()); | |||
5544 | auto *IndexTy = VectorType::get(ScalarIndexTy, NumElts); | |||
5545 | auto *SecondTy = GetElementPtrInst::getIndexedType( | |||
5546 | SourceTy, makeArrayRef(Ops).drop_front()); | |||
5547 | NewAddr = | |||
5548 | Builder.CreateGEP(SecondTy, NewAddr, Constant::getNullValue(IndexTy)); | |||
5549 | } else { | |||
5550 | Value *Base = Ops[0]; | |||
5551 | Value *Index = Ops[FinalIndex]; | |||
5552 | ||||
5553 | // Create a scalar GEP if there are more than 2 operands. | |||
5554 | if (Ops.size() != 2) { | |||
5555 | // Replace the last index with 0. | |||
5556 | Ops[FinalIndex] = Constant::getNullValue(ScalarIndexTy); | |||
5557 | Base = Builder.CreateGEP(SourceTy, Base, | |||
5558 | makeArrayRef(Ops).drop_front()); | |||
5559 | SourceTy = GetElementPtrInst::getIndexedType( | |||
5560 | SourceTy, makeArrayRef(Ops).drop_front()); | |||
5561 | } | |||
5562 | ||||
5563 | // Now create the GEP with scalar pointer and vector index. | |||
5564 | NewAddr = Builder.CreateGEP(SourceTy, Base, Index); | |||
5565 | } | |||
5566 | } else if (!isa<Constant>(Ptr)) { | |||
5567 | // Not a GEP, maybe its a splat and we can create a GEP to enable | |||
5568 | // SelectionDAGBuilder to use it as a uniform base. | |||
5569 | Value *V = getSplatValue(Ptr); | |||
5570 | if (!V) | |||
5571 | return false; | |||
5572 | ||||
5573 | auto NumElts = cast<VectorType>(Ptr->getType())->getElementCount(); | |||
5574 | ||||
5575 | IRBuilder<> Builder(MemoryInst); | |||
5576 | ||||
5577 | // Emit a vector GEP with a scalar pointer and all 0s vector index. | |||
5578 | Type *ScalarIndexTy = DL->getIndexType(V->getType()->getScalarType()); | |||
5579 | auto *IndexTy = VectorType::get(ScalarIndexTy, NumElts); | |||
5580 | Type *ScalarTy; | |||
5581 | if (cast<IntrinsicInst>(MemoryInst)->getIntrinsicID() == | |||
5582 | Intrinsic::masked_gather) { | |||
5583 | ScalarTy = MemoryInst->getType()->getScalarType(); | |||
5584 | } else { | |||
5585 | assert(cast<IntrinsicInst>(MemoryInst)->getIntrinsicID() ==(static_cast <bool> (cast<IntrinsicInst>(MemoryInst )->getIntrinsicID() == Intrinsic::masked_scatter) ? void ( 0) : __assert_fail ("cast<IntrinsicInst>(MemoryInst)->getIntrinsicID() == Intrinsic::masked_scatter" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5586, __extension__ __PRETTY_FUNCTION__ )) | |||
5586 | Intrinsic::masked_scatter)(static_cast <bool> (cast<IntrinsicInst>(MemoryInst )->getIntrinsicID() == Intrinsic::masked_scatter) ? void ( 0) : __assert_fail ("cast<IntrinsicInst>(MemoryInst)->getIntrinsicID() == Intrinsic::masked_scatter" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5586, __extension__ __PRETTY_FUNCTION__ )); | |||
5587 | ScalarTy = MemoryInst->getOperand(0)->getType()->getScalarType(); | |||
5588 | } | |||
5589 | NewAddr = Builder.CreateGEP(ScalarTy, V, Constant::getNullValue(IndexTy)); | |||
5590 | } else { | |||
5591 | // Constant, SelectionDAGBuilder knows to check if its a splat. | |||
5592 | return false; | |||
5593 | } | |||
5594 | ||||
5595 | MemoryInst->replaceUsesOfWith(Ptr, NewAddr); | |||
5596 | ||||
5597 | // If we have no uses, recursively delete the value and all dead instructions | |||
5598 | // using it. | |||
5599 | if (Ptr->use_empty()) | |||
5600 | RecursivelyDeleteTriviallyDeadInstructions( | |||
5601 | Ptr, TLInfo, nullptr, | |||
5602 | [&](Value *V) { removeAllAssertingVHReferences(V); }); | |||
5603 | ||||
5604 | return true; | |||
5605 | } | |||
5606 | ||||
5607 | /// If there are any memory operands, use OptimizeMemoryInst to sink their | |||
5608 | /// address computing into the block when possible / profitable. | |||
5609 | bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) { | |||
5610 | bool MadeChange = false; | |||
5611 | ||||
5612 | const TargetRegisterInfo *TRI = | |||
5613 | TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo(); | |||
5614 | TargetLowering::AsmOperandInfoVector TargetConstraints = | |||
5615 | TLI->ParseConstraints(*DL, TRI, *CS); | |||
5616 | unsigned ArgNo = 0; | |||
5617 | for (TargetLowering::AsmOperandInfo &OpInfo : TargetConstraints) { | |||
5618 | // Compute the constraint code and ConstraintType to use. | |||
5619 | TLI->ComputeConstraintToUse(OpInfo, SDValue()); | |||
5620 | ||||
5621 | // TODO: Also handle C_Address? | |||
5622 | if (OpInfo.ConstraintType == TargetLowering::C_Memory && | |||
5623 | OpInfo.isIndirect) { | |||
5624 | Value *OpVal = CS->getArgOperand(ArgNo++); | |||
5625 | MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u); | |||
5626 | } else if (OpInfo.Type == InlineAsm::isInput) | |||
5627 | ArgNo++; | |||
5628 | } | |||
5629 | ||||
5630 | return MadeChange; | |||
5631 | } | |||
5632 | ||||
5633 | /// Check if all the uses of \p Val are equivalent (or free) zero or | |||
5634 | /// sign extensions. | |||
5635 | static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) { | |||
5636 | assert(!Val->use_empty() && "Input must have at least one use")(static_cast <bool> (!Val->use_empty() && "Input must have at least one use" ) ? void (0) : __assert_fail ("!Val->use_empty() && \"Input must have at least one use\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5636, __extension__ __PRETTY_FUNCTION__ )); | |||
5637 | const Instruction *FirstUser = cast<Instruction>(*Val->user_begin()); | |||
5638 | bool IsSExt = isa<SExtInst>(FirstUser); | |||
5639 | Type *ExtTy = FirstUser->getType(); | |||
5640 | for (const User *U : Val->users()) { | |||
5641 | const Instruction *UI = cast<Instruction>(U); | |||
5642 | if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI))) | |||
5643 | return false; | |||
5644 | Type *CurTy = UI->getType(); | |||
5645 | // Same input and output types: Same instruction after CSE. | |||
5646 | if (CurTy == ExtTy) | |||
5647 | continue; | |||
5648 | ||||
5649 | // If IsSExt is true, we are in this situation: | |||
5650 | // a = Val | |||
5651 | // b = sext ty1 a to ty2 | |||
5652 | // c = sext ty1 a to ty3 | |||
5653 | // Assuming ty2 is shorter than ty3, this could be turned into: | |||
5654 | // a = Val | |||
5655 | // b = sext ty1 a to ty2 | |||
5656 | // c = sext ty2 b to ty3 | |||
5657 | // However, the last sext is not free. | |||
5658 | if (IsSExt) | |||
5659 | return false; | |||
5660 | ||||
5661 | // This is a ZExt, maybe this is free to extend from one type to another. | |||
5662 | // In that case, we would not account for a different use. | |||
5663 | Type *NarrowTy; | |||
5664 | Type *LargeTy; | |||
5665 | if (ExtTy->getScalarType()->getIntegerBitWidth() > | |||
5666 | CurTy->getScalarType()->getIntegerBitWidth()) { | |||
5667 | NarrowTy = CurTy; | |||
5668 | LargeTy = ExtTy; | |||
5669 | } else { | |||
5670 | NarrowTy = ExtTy; | |||
5671 | LargeTy = CurTy; | |||
5672 | } | |||
5673 | ||||
5674 | if (!TLI.isZExtFree(NarrowTy, LargeTy)) | |||
5675 | return false; | |||
5676 | } | |||
5677 | // All uses are the same or can be derived from one another for free. | |||
5678 | return true; | |||
5679 | } | |||
5680 | ||||
5681 | /// Try to speculatively promote extensions in \p Exts and continue | |||
5682 | /// promoting through newly promoted operands recursively as far as doing so is | |||
5683 | /// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts. | |||
5684 | /// When some promotion happened, \p TPT contains the proper state to revert | |||
5685 | /// them. | |||
5686 | /// | |||
5687 | /// \return true if some promotion happened, false otherwise. | |||
5688 | bool CodeGenPrepare::tryToPromoteExts( | |||
5689 | TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts, | |||
5690 | SmallVectorImpl<Instruction *> &ProfitablyMovedExts, | |||
5691 | unsigned CreatedInstsCost) { | |||
5692 | bool Promoted = false; | |||
5693 | ||||
5694 | // Iterate over all the extensions to try to promote them. | |||
5695 | for (auto *I : Exts) { | |||
5696 | // Early check if we directly have ext(load). | |||
5697 | if (isa<LoadInst>(I->getOperand(0))) { | |||
5698 | ProfitablyMovedExts.push_back(I); | |||
5699 | continue; | |||
5700 | } | |||
5701 | ||||
5702 | // Check whether or not we want to do any promotion. The reason we have | |||
5703 | // this check inside the for loop is to catch the case where an extension | |||
5704 | // is directly fed by a load because in such case the extension can be moved | |||
5705 | // up without any promotion on its operands. | |||
5706 | if (!TLI->enableExtLdPromotion() || DisableExtLdPromotion) | |||
5707 | return false; | |||
5708 | ||||
5709 | // Get the action to perform the promotion. | |||
5710 | TypePromotionHelper::Action TPH = | |||
5711 | TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts); | |||
5712 | // Check if we can promote. | |||
5713 | if (!TPH) { | |||
5714 | // Save the current extension as we cannot move up through its operand. | |||
5715 | ProfitablyMovedExts.push_back(I); | |||
5716 | continue; | |||
5717 | } | |||
5718 | ||||
5719 | // Save the current state. | |||
5720 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | |||
5721 | TPT.getRestorationPoint(); | |||
5722 | SmallVector<Instruction *, 4> NewExts; | |||
5723 | unsigned NewCreatedInstsCost = 0; | |||
5724 | unsigned ExtCost = !TLI->isExtFree(I); | |||
5725 | // Promote. | |||
5726 | Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost, | |||
5727 | &NewExts, nullptr, *TLI); | |||
5728 | assert(PromotedVal &&(static_cast <bool> (PromotedVal && "TypePromotionHelper should have filtered out those cases" ) ? void (0) : __assert_fail ("PromotedVal && \"TypePromotionHelper should have filtered out those cases\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5729, __extension__ __PRETTY_FUNCTION__ )) | |||
5729 | "TypePromotionHelper should have filtered out those cases")(static_cast <bool> (PromotedVal && "TypePromotionHelper should have filtered out those cases" ) ? void (0) : __assert_fail ("PromotedVal && \"TypePromotionHelper should have filtered out those cases\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5729, __extension__ __PRETTY_FUNCTION__ )); | |||
5730 | ||||
5731 | // We would be able to merge only one extension in a load. | |||
5732 | // Therefore, if we have more than 1 new extension we heuristically | |||
5733 | // cut this search path, because it means we degrade the code quality. | |||
5734 | // With exactly 2, the transformation is neutral, because we will merge | |||
5735 | // one extension but leave one. However, we optimistically keep going, | |||
5736 | // because the new extension may be removed too. | |||
5737 | long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost; | |||
5738 | // FIXME: It would be possible to propagate a negative value instead of | |||
5739 | // conservatively ceiling it to 0. | |||
5740 | TotalCreatedInstsCost = | |||
5741 | std::max((long long)0, (TotalCreatedInstsCost - ExtCost)); | |||
5742 | if (!StressExtLdPromotion && | |||
5743 | (TotalCreatedInstsCost > 1 || | |||
5744 | !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) { | |||
5745 | // This promotion is not profitable, rollback to the previous state, and | |||
5746 | // save the current extension in ProfitablyMovedExts as the latest | |||
5747 | // speculative promotion turned out to be unprofitable. | |||
5748 | TPT.rollback(LastKnownGood); | |||
5749 | ProfitablyMovedExts.push_back(I); | |||
5750 | continue; | |||
5751 | } | |||
5752 | // Continue promoting NewExts as far as doing so is profitable. | |||
5753 | SmallVector<Instruction *, 2> NewlyMovedExts; | |||
5754 | (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost); | |||
5755 | bool NewPromoted = false; | |||
5756 | for (auto *ExtInst : NewlyMovedExts) { | |||
5757 | Instruction *MovedExt = cast<Instruction>(ExtInst); | |||
5758 | Value *ExtOperand = MovedExt->getOperand(0); | |||
5759 | // If we have reached to a load, we need this extra profitability check | |||
5760 | // as it could potentially be merged into an ext(load). | |||
5761 | if (isa<LoadInst>(ExtOperand) && | |||
5762 | !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost || | |||
5763 | (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI)))) | |||
5764 | continue; | |||
5765 | ||||
5766 | ProfitablyMovedExts.push_back(MovedExt); | |||
5767 | NewPromoted = true; | |||
5768 | } | |||
5769 | ||||
5770 | // If none of speculative promotions for NewExts is profitable, rollback | |||
5771 | // and save the current extension (I) as the last profitable extension. | |||
5772 | if (!NewPromoted) { | |||
5773 | TPT.rollback(LastKnownGood); | |||
5774 | ProfitablyMovedExts.push_back(I); | |||
5775 | continue; | |||
5776 | } | |||
5777 | // The promotion is profitable. | |||
5778 | Promoted = true; | |||
5779 | } | |||
5780 | return Promoted; | |||
5781 | } | |||
5782 | ||||
5783 | /// Merging redundant sexts when one is dominating the other. | |||
5784 | bool CodeGenPrepare::mergeSExts(Function &F) { | |||
5785 | bool Changed = false; | |||
5786 | for (auto &Entry : ValToSExtendedUses) { | |||
5787 | SExts &Insts = Entry.second; | |||
5788 | SExts CurPts; | |||
5789 | for (Instruction *Inst : Insts) { | |||
5790 | if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) || | |||
5791 | Inst->getOperand(0) != Entry.first) | |||
5792 | continue; | |||
5793 | bool inserted = false; | |||
5794 | for (auto &Pt : CurPts) { | |||
5795 | if (getDT(F).dominates(Inst, Pt)) { | |||
5796 | Pt->replaceAllUsesWith(Inst); | |||
5797 | RemovedInsts.insert(Pt); | |||
5798 | Pt->removeFromParent(); | |||
5799 | Pt = Inst; | |||
5800 | inserted = true; | |||
5801 | Changed = true; | |||
5802 | break; | |||
5803 | } | |||
5804 | if (!getDT(F).dominates(Pt, Inst)) | |||
5805 | // Give up if we need to merge in a common dominator as the | |||
5806 | // experiments show it is not profitable. | |||
5807 | continue; | |||
5808 | Inst->replaceAllUsesWith(Pt); | |||
5809 | RemovedInsts.insert(Inst); | |||
5810 | Inst->removeFromParent(); | |||
5811 | inserted = true; | |||
5812 | Changed = true; | |||
5813 | break; | |||
5814 | } | |||
5815 | if (!inserted) | |||
5816 | CurPts.push_back(Inst); | |||
5817 | } | |||
5818 | } | |||
5819 | return Changed; | |||
5820 | } | |||
5821 | ||||
5822 | // Splitting large data structures so that the GEPs accessing them can have | |||
5823 | // smaller offsets so that they can be sunk to the same blocks as their users. | |||
5824 | // For example, a large struct starting from %base is split into two parts | |||
5825 | // where the second part starts from %new_base. | |||
5826 | // | |||
5827 | // Before: | |||
5828 | // BB0: | |||
5829 | // %base = | |||
5830 | // | |||
5831 | // BB1: | |||
5832 | // %gep0 = gep %base, off0 | |||
5833 | // %gep1 = gep %base, off1 | |||
5834 | // %gep2 = gep %base, off2 | |||
5835 | // | |||
5836 | // BB2: | |||
5837 | // %load1 = load %gep0 | |||
5838 | // %load2 = load %gep1 | |||
5839 | // %load3 = load %gep2 | |||
5840 | // | |||
5841 | // After: | |||
5842 | // BB0: | |||
5843 | // %base = | |||
5844 | // %new_base = gep %base, off0 | |||
5845 | // | |||
5846 | // BB1: | |||
5847 | // %new_gep0 = %new_base | |||
5848 | // %new_gep1 = gep %new_base, off1 - off0 | |||
5849 | // %new_gep2 = gep %new_base, off2 - off0 | |||
5850 | // | |||
5851 | // BB2: | |||
5852 | // %load1 = load i32, i32* %new_gep0 | |||
5853 | // %load2 = load i32, i32* %new_gep1 | |||
5854 | // %load3 = load i32, i32* %new_gep2 | |||
5855 | // | |||
5856 | // %new_gep1 and %new_gep2 can be sunk to BB2 now after the splitting because | |||
5857 | // their offsets are smaller enough to fit into the addressing mode. | |||
5858 | bool CodeGenPrepare::splitLargeGEPOffsets() { | |||
5859 | bool Changed = false; | |||
5860 | for (auto &Entry : LargeOffsetGEPMap) { | |||
5861 | Value *OldBase = Entry.first; | |||
5862 | SmallVectorImpl<std::pair<AssertingVH<GetElementPtrInst>, int64_t>> | |||
5863 | &LargeOffsetGEPs = Entry.second; | |||
5864 | auto compareGEPOffset = | |||
5865 | [&](const std::pair<GetElementPtrInst *, int64_t> &LHS, | |||
5866 | const std::pair<GetElementPtrInst *, int64_t> &RHS) { | |||
5867 | if (LHS.first == RHS.first) | |||
5868 | return false; | |||
5869 | if (LHS.second != RHS.second) | |||
5870 | return LHS.second < RHS.second; | |||
5871 | return LargeOffsetGEPID[LHS.first] < LargeOffsetGEPID[RHS.first]; | |||
5872 | }; | |||
5873 | // Sorting all the GEPs of the same data structures based on the offsets. | |||
5874 | llvm::sort(LargeOffsetGEPs, compareGEPOffset); | |||
5875 | LargeOffsetGEPs.erase( | |||
5876 | std::unique(LargeOffsetGEPs.begin(), LargeOffsetGEPs.end()), | |||
5877 | LargeOffsetGEPs.end()); | |||
5878 | // Skip if all the GEPs have the same offsets. | |||
5879 | if (LargeOffsetGEPs.front().second == LargeOffsetGEPs.back().second) | |||
5880 | continue; | |||
5881 | GetElementPtrInst *BaseGEP = LargeOffsetGEPs.begin()->first; | |||
5882 | int64_t BaseOffset = LargeOffsetGEPs.begin()->second; | |||
5883 | Value *NewBaseGEP = nullptr; | |||
5884 | ||||
5885 | auto *LargeOffsetGEP = LargeOffsetGEPs.begin(); | |||
5886 | while (LargeOffsetGEP != LargeOffsetGEPs.end()) { | |||
5887 | GetElementPtrInst *GEP = LargeOffsetGEP->first; | |||
5888 | int64_t Offset = LargeOffsetGEP->second; | |||
5889 | if (Offset != BaseOffset) { | |||
5890 | TargetLowering::AddrMode AddrMode; | |||
5891 | AddrMode.BaseOffs = Offset - BaseOffset; | |||
5892 | // The result type of the GEP might not be the type of the memory | |||
5893 | // access. | |||
5894 | if (!TLI->isLegalAddressingMode(*DL, AddrMode, | |||
5895 | GEP->getResultElementType(), | |||
5896 | GEP->getAddressSpace())) { | |||
5897 | // We need to create a new base if the offset to the current base is | |||
5898 | // too large to fit into the addressing mode. So, a very large struct | |||
5899 | // may be split into several parts. | |||
5900 | BaseGEP = GEP; | |||
5901 | BaseOffset = Offset; | |||
5902 | NewBaseGEP = nullptr; | |||
5903 | } | |||
5904 | } | |||
5905 | ||||
5906 | // Generate a new GEP to replace the current one. | |||
5907 | LLVMContext &Ctx = GEP->getContext(); | |||
5908 | Type *IntPtrTy = DL->getIntPtrType(GEP->getType()); | |||
5909 | Type *I8PtrTy = | |||
5910 | Type::getInt8PtrTy(Ctx, GEP->getType()->getPointerAddressSpace()); | |||
5911 | Type *I8Ty = Type::getInt8Ty(Ctx); | |||
5912 | ||||
5913 | if (!NewBaseGEP) { | |||
5914 | // Create a new base if we don't have one yet. Find the insertion | |||
5915 | // pointer for the new base first. | |||
5916 | BasicBlock::iterator NewBaseInsertPt; | |||
5917 | BasicBlock *NewBaseInsertBB; | |||
5918 | if (auto *BaseI = dyn_cast<Instruction>(OldBase)) { | |||
5919 | // If the base of the struct is an instruction, the new base will be | |||
5920 | // inserted close to it. | |||
5921 | NewBaseInsertBB = BaseI->getParent(); | |||
5922 | if (isa<PHINode>(BaseI)) | |||
5923 | NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt(); | |||
5924 | else if (InvokeInst *Invoke = dyn_cast<InvokeInst>(BaseI)) { | |||
5925 | NewBaseInsertBB = | |||
5926 | SplitEdge(NewBaseInsertBB, Invoke->getNormalDest()); | |||
5927 | NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt(); | |||
5928 | } else | |||
5929 | NewBaseInsertPt = std::next(BaseI->getIterator()); | |||
5930 | } else { | |||
5931 | // If the current base is an argument or global value, the new base | |||
5932 | // will be inserted to the entry block. | |||
5933 | NewBaseInsertBB = &BaseGEP->getFunction()->getEntryBlock(); | |||
5934 | NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt(); | |||
5935 | } | |||
5936 | IRBuilder<> NewBaseBuilder(NewBaseInsertBB, NewBaseInsertPt); | |||
5937 | // Create a new base. | |||
5938 | Value *BaseIndex = ConstantInt::get(IntPtrTy, BaseOffset); | |||
5939 | NewBaseGEP = OldBase; | |||
5940 | if (NewBaseGEP->getType() != I8PtrTy) | |||
5941 | NewBaseGEP = NewBaseBuilder.CreatePointerCast(NewBaseGEP, I8PtrTy); | |||
5942 | NewBaseGEP = | |||
5943 | NewBaseBuilder.CreateGEP(I8Ty, NewBaseGEP, BaseIndex, "splitgep"); | |||
5944 | NewGEPBases.insert(NewBaseGEP); | |||
5945 | } | |||
5946 | ||||
5947 | IRBuilder<> Builder(GEP); | |||
5948 | Value *NewGEP = NewBaseGEP; | |||
5949 | if (Offset == BaseOffset) { | |||
5950 | if (GEP->getType() != I8PtrTy) | |||
5951 | NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType()); | |||
5952 | } else { | |||
5953 | // Calculate the new offset for the new GEP. | |||
5954 | Value *Index = ConstantInt::get(IntPtrTy, Offset - BaseOffset); | |||
5955 | NewGEP = Builder.CreateGEP(I8Ty, NewBaseGEP, Index); | |||
5956 | ||||
5957 | if (GEP->getType() != I8PtrTy) | |||
5958 | NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType()); | |||
5959 | } | |||
5960 | GEP->replaceAllUsesWith(NewGEP); | |||
5961 | LargeOffsetGEPID.erase(GEP); | |||
5962 | LargeOffsetGEP = LargeOffsetGEPs.erase(LargeOffsetGEP); | |||
5963 | GEP->eraseFromParent(); | |||
5964 | Changed = true; | |||
5965 | } | |||
5966 | } | |||
5967 | return Changed; | |||
5968 | } | |||
5969 | ||||
5970 | bool CodeGenPrepare::optimizePhiType( | |||
5971 | PHINode *I, SmallPtrSetImpl<PHINode *> &Visited, | |||
5972 | SmallPtrSetImpl<Instruction *> &DeletedInstrs) { | |||
5973 | // We are looking for a collection on interconnected phi nodes that together | |||
5974 | // only use loads/bitcasts and are used by stores/bitcasts, and the bitcasts | |||
5975 | // are of the same type. Convert the whole set of nodes to the type of the | |||
5976 | // bitcast. | |||
5977 | Type *PhiTy = I->getType(); | |||
5978 | Type *ConvertTy = nullptr; | |||
5979 | if (Visited.count(I) || | |||
5980 | (!I->getType()->isIntegerTy() && !I->getType()->isFloatingPointTy())) | |||
5981 | return false; | |||
5982 | ||||
5983 | SmallVector<Instruction *, 4> Worklist; | |||
5984 | Worklist.push_back(cast<Instruction>(I)); | |||
5985 | SmallPtrSet<PHINode *, 4> PhiNodes; | |||
5986 | PhiNodes.insert(I); | |||
5987 | Visited.insert(I); | |||
5988 | SmallPtrSet<Instruction *, 4> Defs; | |||
5989 | SmallPtrSet<Instruction *, 4> Uses; | |||
5990 | // This works by adding extra bitcasts between load/stores and removing | |||
5991 | // existing bicasts. If we have a phi(bitcast(load)) or a store(bitcast(phi)) | |||
5992 | // we can get in the situation where we remove a bitcast in one iteration | |||
5993 | // just to add it again in the next. We need to ensure that at least one | |||
5994 | // bitcast we remove are anchored to something that will not change back. | |||
5995 | bool AnyAnchored = false; | |||
5996 | ||||
5997 | while (!Worklist.empty()) { | |||
5998 | Instruction *II = Worklist.pop_back_val(); | |||
5999 | ||||
6000 | if (auto *Phi = dyn_cast<PHINode>(II)) { | |||
6001 | // Handle Defs, which might also be PHI's | |||
6002 | for (Value *V : Phi->incoming_values()) { | |||
6003 | if (auto *OpPhi = dyn_cast<PHINode>(V)) { | |||
6004 | if (!PhiNodes.count(OpPhi)) { | |||
6005 | if (Visited.count(OpPhi)) | |||
6006 | return false; | |||
6007 | PhiNodes.insert(OpPhi); | |||
6008 | Visited.insert(OpPhi); | |||
6009 | Worklist.push_back(OpPhi); | |||
6010 | } | |||
6011 | } else if (auto *OpLoad = dyn_cast<LoadInst>(V)) { | |||
6012 | if (!OpLoad->isSimple()) | |||
6013 | return false; | |||
6014 | if (!Defs.count(OpLoad)) { | |||
6015 | Defs.insert(OpLoad); | |||
6016 | Worklist.push_back(OpLoad); | |||
6017 | } | |||
6018 | } else if (auto *OpEx = dyn_cast<ExtractElementInst>(V)) { | |||
6019 | if (!Defs.count(OpEx)) { | |||
6020 | Defs.insert(OpEx); | |||
6021 | Worklist.push_back(OpEx); | |||
6022 | } | |||
6023 | } else if (auto *OpBC = dyn_cast<BitCastInst>(V)) { | |||
6024 | if (!ConvertTy) | |||
6025 | ConvertTy = OpBC->getOperand(0)->getType(); | |||
6026 | if (OpBC->getOperand(0)->getType() != ConvertTy) | |||
6027 | return false; | |||
6028 | if (!Defs.count(OpBC)) { | |||
6029 | Defs.insert(OpBC); | |||
6030 | Worklist.push_back(OpBC); | |||
6031 | AnyAnchored |= !isa<LoadInst>(OpBC->getOperand(0)) && | |||
6032 | !isa<ExtractElementInst>(OpBC->getOperand(0)); | |||
6033 | } | |||
6034 | } else if (!isa<UndefValue>(V)) { | |||
6035 | return false; | |||
6036 | } | |||
6037 | } | |||
6038 | } | |||
6039 | ||||
6040 | // Handle uses which might also be phi's | |||
6041 | for (User *V : II->users()) { | |||
6042 | if (auto *OpPhi = dyn_cast<PHINode>(V)) { | |||
6043 | if (!PhiNodes.count(OpPhi)) { | |||
6044 | if (Visited.count(OpPhi)) | |||
6045 | return false; | |||
6046 | PhiNodes.insert(OpPhi); | |||
6047 | Visited.insert(OpPhi); | |||
6048 | Worklist.push_back(OpPhi); | |||
6049 | } | |||
6050 | } else if (auto *OpStore = dyn_cast<StoreInst>(V)) { | |||
6051 | if (!OpStore->isSimple() || OpStore->getOperand(0) != II) | |||
6052 | return false; | |||
6053 | Uses.insert(OpStore); | |||
6054 | } else if (auto *OpBC = dyn_cast<BitCastInst>(V)) { | |||
6055 | if (!ConvertTy) | |||
6056 | ConvertTy = OpBC->getType(); | |||
6057 | if (OpBC->getType() != ConvertTy) | |||
6058 | return false; | |||
6059 | Uses.insert(OpBC); | |||
6060 | AnyAnchored |= | |||
6061 | any_of(OpBC->users(), [](User *U) { return !isa<StoreInst>(U); }); | |||
6062 | } else { | |||
6063 | return false; | |||
6064 | } | |||
6065 | } | |||
6066 | } | |||
6067 | ||||
6068 | if (!ConvertTy || !AnyAnchored || !TLI->shouldConvertPhiType(PhiTy, ConvertTy)) | |||
6069 | return false; | |||
6070 | ||||
6071 | LLVM_DEBUG(dbgs() << "Converting " << *I << "\n and connected nodes to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Converting " << * I << "\n and connected nodes to " << *ConvertTy << "\n"; } } while (false) | |||
6072 | << *ConvertTy << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Converting " << * I << "\n and connected nodes to " << *ConvertTy << "\n"; } } while (false); | |||
6073 | ||||
6074 | // Create all the new phi nodes of the new type, and bitcast any loads to the | |||
6075 | // correct type. | |||
6076 | ValueToValueMap ValMap; | |||
6077 | ValMap[UndefValue::get(PhiTy)] = UndefValue::get(ConvertTy); | |||
6078 | for (Instruction *D : Defs) { | |||
6079 | if (isa<BitCastInst>(D)) { | |||
6080 | ValMap[D] = D->getOperand(0); | |||
6081 | DeletedInstrs.insert(D); | |||
6082 | } else { | |||
6083 | ValMap[D] = | |||
6084 | new BitCastInst(D, ConvertTy, D->getName() + ".bc", D->getNextNode()); | |||
6085 | } | |||
6086 | } | |||
6087 | for (PHINode *Phi : PhiNodes) | |||
6088 | ValMap[Phi] = PHINode::Create(ConvertTy, Phi->getNumIncomingValues(), | |||
6089 | Phi->getName() + ".tc", Phi); | |||
6090 | // Pipe together all the PhiNodes. | |||
6091 | for (PHINode *Phi : PhiNodes) { | |||
6092 | PHINode *NewPhi = cast<PHINode>(ValMap[Phi]); | |||
6093 | for (int i = 0, e = Phi->getNumIncomingValues(); i < e; i++) | |||
6094 | NewPhi->addIncoming(ValMap[Phi->getIncomingValue(i)], | |||
6095 | Phi->getIncomingBlock(i)); | |||
6096 | Visited.insert(NewPhi); | |||
6097 | } | |||
6098 | // And finally pipe up the stores and bitcasts | |||
6099 | for (Instruction *U : Uses) { | |||
6100 | if (isa<BitCastInst>(U)) { | |||
6101 | DeletedInstrs.insert(U); | |||
6102 | U->replaceAllUsesWith(ValMap[U->getOperand(0)]); | |||
6103 | } else { | |||
6104 | U->setOperand(0, | |||
6105 | new BitCastInst(ValMap[U->getOperand(0)], PhiTy, "bc", U)); | |||
6106 | } | |||
6107 | } | |||
6108 | ||||
6109 | // Save the removed phis to be deleted later. | |||
6110 | for (PHINode *Phi : PhiNodes) | |||
6111 | DeletedInstrs.insert(Phi); | |||
6112 | return true; | |||
6113 | } | |||
6114 | ||||
6115 | bool CodeGenPrepare::optimizePhiTypes(Function &F) { | |||
6116 | if (!OptimizePhiTypes) | |||
6117 | return false; | |||
6118 | ||||
6119 | bool Changed = false; | |||
6120 | SmallPtrSet<PHINode *, 4> Visited; | |||
6121 | SmallPtrSet<Instruction *, 4> DeletedInstrs; | |||
6122 | ||||
6123 | // Attempt to optimize all the phis in the functions to the correct type. | |||
6124 | for (auto &BB : F) | |||
6125 | for (auto &Phi : BB.phis()) | |||
6126 | Changed |= optimizePhiType(&Phi, Visited, DeletedInstrs); | |||
6127 | ||||
6128 | // Remove any old phi's that have been converted. | |||
6129 | for (auto *I : DeletedInstrs) { | |||
6130 | I->replaceAllUsesWith(UndefValue::get(I->getType())); | |||
6131 | I->eraseFromParent(); | |||
6132 | } | |||
6133 | ||||
6134 | return Changed; | |||
6135 | } | |||
6136 | ||||
6137 | /// Return true, if an ext(load) can be formed from an extension in | |||
6138 | /// \p MovedExts. | |||
6139 | bool CodeGenPrepare::canFormExtLd( | |||
6140 | const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI, | |||
6141 | Instruction *&Inst, bool HasPromoted) { | |||
6142 | for (auto *MovedExtInst : MovedExts) { | |||
6143 | if (isa<LoadInst>(MovedExtInst->getOperand(0))) { | |||
6144 | LI = cast<LoadInst>(MovedExtInst->getOperand(0)); | |||
6145 | Inst = MovedExtInst; | |||
6146 | break; | |||
6147 | } | |||
6148 | } | |||
6149 | if (!LI) | |||
6150 | return false; | |||
6151 | ||||
6152 | // If they're already in the same block, there's nothing to do. | |||
6153 | // Make the cheap checks first if we did not promote. | |||
6154 | // If we promoted, we need to check if it is indeed profitable. | |||
6155 | if (!HasPromoted && LI->getParent() == Inst->getParent()) | |||
6156 | return false; | |||
6157 | ||||
6158 | return TLI->isExtLoad(LI, Inst, *DL); | |||
6159 | } | |||
6160 | ||||
6161 | /// Move a zext or sext fed by a load into the same basic block as the load, | |||
6162 | /// unless conditions are unfavorable. This allows SelectionDAG to fold the | |||
6163 | /// extend into the load. | |||
6164 | /// | |||
6165 | /// E.g., | |||
6166 | /// \code | |||
6167 | /// %ld = load i32* %addr | |||
6168 | /// %add = add nuw i32 %ld, 4 | |||
6169 | /// %zext = zext i32 %add to i64 | |||
6170 | // \endcode | |||
6171 | /// => | |||
6172 | /// \code | |||
6173 | /// %ld = load i32* %addr | |||
6174 | /// %zext = zext i32 %ld to i64 | |||
6175 | /// %add = add nuw i64 %zext, 4 | |||
6176 | /// \encode | |||
6177 | /// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which | |||
6178 | /// allow us to match zext(load i32*) to i64. | |||
6179 | /// | |||
6180 | /// Also, try to promote the computations used to obtain a sign extended | |||
6181 | /// value used into memory accesses. | |||
6182 | /// E.g., | |||
6183 | /// \code | |||
6184 | /// a = add nsw i32 b, 3 | |||
6185 | /// d = sext i32 a to i64 | |||
6186 | /// e = getelementptr ..., i64 d | |||
6187 | /// \endcode | |||
6188 | /// => | |||
6189 | /// \code | |||
6190 | /// f = sext i32 b to i64 | |||
6191 | /// a = add nsw i64 f, 3 | |||
6192 | /// e = getelementptr ..., i64 a | |||
6193 | /// \endcode | |||
6194 | /// | |||
6195 | /// \p Inst[in/out] the extension may be modified during the process if some | |||
6196 | /// promotions apply. | |||
6197 | bool CodeGenPrepare::optimizeExt(Instruction *&Inst) { | |||
6198 | bool AllowPromotionWithoutCommonHeader = false; | |||
6199 | /// See if it is an interesting sext operations for the address type | |||
6200 | /// promotion before trying to promote it, e.g., the ones with the right | |||
6201 | /// type and used in memory accesses. | |||
6202 | bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion( | |||
6203 | *Inst, AllowPromotionWithoutCommonHeader); | |||
6204 | TypePromotionTransaction TPT(RemovedInsts); | |||
6205 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | |||
6206 | TPT.getRestorationPoint(); | |||
6207 | SmallVector<Instruction *, 1> Exts; | |||
6208 | SmallVector<Instruction *, 2> SpeculativelyMovedExts; | |||
6209 | Exts.push_back(Inst); | |||
6210 | ||||
6211 | bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts); | |||
6212 | ||||
6213 | // Look for a load being extended. | |||
6214 | LoadInst *LI = nullptr; | |||
6215 | Instruction *ExtFedByLoad; | |||
6216 | ||||
6217 | // Try to promote a chain of computation if it allows to form an extended | |||
6218 | // load. | |||
6219 | if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) { | |||
6220 | assert(LI && ExtFedByLoad && "Expect a valid load and extension")(static_cast <bool> (LI && ExtFedByLoad && "Expect a valid load and extension") ? void (0) : __assert_fail ("LI && ExtFedByLoad && \"Expect a valid load and extension\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6220, __extension__ __PRETTY_FUNCTION__ )); | |||
6221 | TPT.commit(); | |||
6222 | // Move the extend into the same block as the load. | |||
6223 | ExtFedByLoad->moveAfter(LI); | |||
6224 | ++NumExtsMoved; | |||
6225 | Inst = ExtFedByLoad; | |||
6226 | return true; | |||
6227 | } | |||
6228 | ||||
6229 | // Continue promoting SExts if known as considerable depending on targets. | |||
6230 | if (ATPConsiderable && | |||
6231 | performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader, | |||
6232 | HasPromoted, TPT, SpeculativelyMovedExts)) | |||
6233 | return true; | |||
6234 | ||||
6235 | TPT.rollback(LastKnownGood); | |||
6236 | return false; | |||
6237 | } | |||
6238 | ||||
6239 | // Perform address type promotion if doing so is profitable. | |||
6240 | // If AllowPromotionWithoutCommonHeader == false, we should find other sext | |||
6241 | // instructions that sign extended the same initial value. However, if | |||
6242 | // AllowPromotionWithoutCommonHeader == true, we expect promoting the | |||
6243 | // extension is just profitable. | |||
6244 | bool CodeGenPrepare::performAddressTypePromotion( | |||
6245 | Instruction *&Inst, bool AllowPromotionWithoutCommonHeader, | |||
6246 | bool HasPromoted, TypePromotionTransaction &TPT, | |||
6247 | SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) { | |||
6248 | bool Promoted = false; | |||
6249 | SmallPtrSet<Instruction *, 1> UnhandledExts; | |||
6250 | bool AllSeenFirst = true; | |||
6251 | for (auto *I : SpeculativelyMovedExts) { | |||
6252 | Value *HeadOfChain = I->getOperand(0); | |||
6253 | DenseMap<Value *, Instruction *>::iterator AlreadySeen = | |||
6254 | SeenChainsForSExt.find(HeadOfChain); | |||
6255 | // If there is an unhandled SExt which has the same header, try to promote | |||
6256 | // it as well. | |||
6257 | if (AlreadySeen != SeenChainsForSExt.end()) { | |||
6258 | if (AlreadySeen->second != nullptr) | |||
6259 | UnhandledExts.insert(AlreadySeen->second); | |||
6260 | AllSeenFirst = false; | |||
6261 | } | |||
6262 | } | |||
6263 | ||||
6264 | if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader && | |||
6265 | SpeculativelyMovedExts.size() == 1)) { | |||
6266 | TPT.commit(); | |||
6267 | if (HasPromoted) | |||
6268 | Promoted = true; | |||
6269 | for (auto *I : SpeculativelyMovedExts) { | |||
6270 | Value *HeadOfChain = I->getOperand(0); | |||
6271 | SeenChainsForSExt[HeadOfChain] = nullptr; | |||
6272 | ValToSExtendedUses[HeadOfChain].push_back(I); | |||
6273 | } | |||
6274 | // Update Inst as promotion happen. | |||
6275 | Inst = SpeculativelyMovedExts.pop_back_val(); | |||
6276 | } else { | |||
6277 | // This is the first chain visited from the header, keep the current chain | |||
6278 | // as unhandled. Defer to promote this until we encounter another SExt | |||
6279 | // chain derived from the same header. | |||
6280 | for (auto *I : SpeculativelyMovedExts) { | |||
6281 | Value *HeadOfChain = I->getOperand(0); | |||
6282 | SeenChainsForSExt[HeadOfChain] = Inst; | |||
6283 | } | |||
6284 | return false; | |||
6285 | } | |||
6286 | ||||
6287 | if (!AllSeenFirst && !UnhandledExts.empty()) | |||
6288 | for (auto *VisitedSExt : UnhandledExts) { | |||
6289 | if (RemovedInsts.count(VisitedSExt)) | |||
6290 | continue; | |||
6291 | TypePromotionTransaction TPT(RemovedInsts); | |||
6292 | SmallVector<Instruction *, 1> Exts; | |||
6293 | SmallVector<Instruction *, 2> Chains; | |||
6294 | Exts.push_back(VisitedSExt); | |||
6295 | bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains); | |||
6296 | TPT.commit(); | |||
6297 | if (HasPromoted) | |||
6298 | Promoted = true; | |||
6299 | for (auto *I : Chains) { | |||
6300 | Value *HeadOfChain = I->getOperand(0); | |||
6301 | // Mark this as handled. | |||
6302 | SeenChainsForSExt[HeadOfChain] = nullptr; | |||
6303 | ValToSExtendedUses[HeadOfChain].push_back(I); | |||
6304 | } | |||
6305 | } | |||
6306 | return Promoted; | |||
6307 | } | |||
6308 | ||||
6309 | bool CodeGenPrepare::optimizeExtUses(Instruction *I) { | |||
6310 | BasicBlock *DefBB = I->getParent(); | |||
6311 | ||||
6312 | // If the result of a {s|z}ext and its source are both live out, rewrite all | |||
6313 | // other uses of the source with result of extension. | |||
6314 | Value *Src = I->getOperand(0); | |||
6315 | if (Src->hasOneUse()) | |||
6316 | return false; | |||
6317 | ||||
6318 | // Only do this xform if truncating is free. | |||
6319 | if (!TLI->isTruncateFree(I->getType(), Src->getType())) | |||
6320 | return false; | |||
6321 | ||||
6322 | // Only safe to perform the optimization if the source is also defined in | |||
6323 | // this block. | |||
6324 | if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent()) | |||
6325 | return false; | |||
6326 | ||||
6327 | bool DefIsLiveOut = false; | |||
6328 | for (User *U : I->users()) { | |||
6329 | Instruction *UI = cast<Instruction>(U); | |||
6330 | ||||
6331 | // Figure out which BB this ext is used in. | |||
6332 | BasicBlock *UserBB = UI->getParent(); | |||
6333 | if (UserBB == DefBB) continue; | |||
6334 | DefIsLiveOut = true; | |||
6335 | break; | |||
6336 | } | |||
6337 | if (!DefIsLiveOut) | |||
6338 | return false; | |||
6339 | ||||
6340 | // Make sure none of the uses are PHI nodes. | |||
6341 | for (User *U : Src->users()) { | |||
6342 | Instruction *UI = cast<Instruction>(U); | |||
6343 | BasicBlock *UserBB = UI->getParent(); | |||
6344 | if (UserBB == DefBB) continue; | |||
6345 | // Be conservative. We don't want this xform to end up introducing | |||
6346 | // reloads just before load / store instructions. | |||
6347 | if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI)) | |||
6348 | return false; | |||
6349 | } | |||
6350 | ||||
6351 | // InsertedTruncs - Only insert one trunc in each block once. | |||
6352 | DenseMap<BasicBlock*, Instruction*> InsertedTruncs; | |||
6353 | ||||
6354 | bool MadeChange = false; | |||
6355 | for (Use &U : Src->uses()) { | |||
6356 | Instruction *User = cast<Instruction>(U.getUser()); | |||
6357 | ||||
6358 | // Figure out which BB this ext is used in. | |||
6359 | BasicBlock *UserBB = User->getParent(); | |||
6360 | if (UserBB == DefBB) continue; | |||
6361 | ||||
6362 | // Both src and def are live in this block. Rewrite the use. | |||
6363 | Instruction *&InsertedTrunc = InsertedTruncs[UserBB]; | |||
6364 | ||||
6365 | if (!InsertedTrunc) { | |||
6366 | BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); | |||
6367 | assert(InsertPt != UserBB->end())(static_cast <bool> (InsertPt != UserBB->end()) ? void (0) : __assert_fail ("InsertPt != UserBB->end()", "llvm/lib/CodeGen/CodeGenPrepare.cpp" , 6367, __extension__ __PRETTY_FUNCTION__)); | |||
6368 | InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt); | |||
6369 | InsertedInsts.insert(InsertedTrunc); | |||
6370 | } | |||
6371 | ||||
6372 | // Replace a use of the {s|z}ext source with a use of the result. | |||
6373 | U = InsertedTrunc; | |||
6374 | ++NumExtUses; | |||
6375 | MadeChange = true; | |||
6376 | } | |||
6377 | ||||
6378 | return MadeChange; | |||
6379 | } | |||
6380 | ||||
6381 | // Find loads whose uses only use some of the loaded value's bits. Add an "and" | |||
6382 | // just after the load if the target can fold this into one extload instruction, | |||
6383 | // with the hope of eliminating some of the other later "and" instructions using | |||
6384 | // the loaded value. "and"s that are made trivially redundant by the insertion | |||
6385 | // of the new "and" are removed by this function, while others (e.g. those whose | |||
6386 | // path from the load goes through a phi) are left for isel to potentially | |||
6387 | // remove. | |||
6388 | // | |||
6389 | // For example: | |||
6390 | // | |||
6391 | // b0: | |||
6392 | // x = load i32 | |||
6393 | // ... | |||
6394 | // b1: | |||
6395 | // y = and x, 0xff | |||
6396 | // z = use y | |||
6397 | // | |||
6398 | // becomes: | |||
6399 | // | |||
6400 | // b0: | |||
6401 | // x = load i32 | |||
6402 | // x' = and x, 0xff | |||
6403 | // ... | |||
6404 | // b1: | |||
6405 | // z = use x' | |||
6406 | // | |||
6407 | // whereas: | |||
6408 | // | |||
6409 | // b0: | |||
6410 | // x1 = load i32 | |||
6411 | // ... | |||
6412 | // b1: | |||
6413 | // x2 = load i32 | |||
6414 | // ... | |||
6415 | // b2: | |||
6416 | // x = phi x1, x2 | |||
6417 | // y = and x, 0xff | |||
6418 | // | |||
6419 | // becomes (after a call to optimizeLoadExt for each load): | |||
6420 | // | |||
6421 | // b0: | |||
6422 | // x1 = load i32 | |||
6423 | // x1' = and x1, 0xff | |||
6424 | // ... | |||
6425 | // b1: | |||
6426 | // x2 = load i32 | |||
6427 | // x2' = and x2, 0xff | |||
6428 | // ... | |||
6429 | // b2: | |||
6430 | // x = phi x1', x2' | |||
6431 | // y = and x, 0xff | |||
6432 | bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) { | |||
6433 | if (!Load->isSimple() || !Load->getType()->isIntOrPtrTy()) | |||
6434 | return false; | |||
6435 | ||||
6436 | // Skip loads we've already transformed. | |||
6437 | if (Load->hasOneUse() && | |||
6438 | InsertedInsts.count(cast<Instruction>(*Load->user_begin()))) | |||
6439 | return false; | |||
6440 | ||||
6441 | // Look at all uses of Load, looking through phis, to determine how many bits | |||
6442 | // of the loaded value are needed. | |||
6443 | SmallVector<Instruction *, 8> WorkList; | |||
6444 | SmallPtrSet<Instruction *, 16> Visited; | |||
6445 | SmallVector<Instruction *, 8> AndsToMaybeRemove; | |||
6446 | for (auto *U : Load->users()) | |||
6447 | WorkList.push_back(cast<Instruction>(U)); | |||
6448 | ||||
6449 | EVT LoadResultVT = TLI->getValueType(*DL, Load->getType()); | |||
6450 | unsigned BitWidth = LoadResultVT.getSizeInBits(); | |||
6451 | // If the BitWidth is 0, do not try to optimize the type | |||
6452 | if (BitWidth == 0) | |||
6453 | return false; | |||
6454 | ||||
6455 | APInt DemandBits(BitWidth, 0); | |||
6456 | APInt WidestAndBits(BitWidth, 0); | |||
6457 | ||||
6458 | while (!WorkList.empty()) { | |||
6459 | Instruction *I = WorkList.pop_back_val(); | |||
6460 | ||||
6461 | // Break use-def graph loops. | |||
6462 | if (!Visited.insert(I).second) | |||
6463 | continue; | |||
6464 | ||||
6465 | // For a PHI node, push all of its users. | |||
6466 | if (auto *Phi = dyn_cast<PHINode>(I)) { | |||
6467 | for (auto *U : Phi->users()) | |||
6468 | WorkList.push_back(cast<Instruction>(U)); | |||
6469 | continue; | |||
6470 | } | |||
6471 | ||||
6472 | switch (I->getOpcode()) { | |||
6473 | case Instruction::And: { | |||
6474 | auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1)); | |||
6475 | if (!AndC) | |||
6476 | return false; | |||
6477 | APInt AndBits = AndC->getValue(); | |||
6478 | DemandBits |= AndBits; | |||
6479 | // Keep track of the widest and mask we see. | |||
6480 | if (AndBits.ugt(WidestAndBits)) | |||
6481 | WidestAndBits = AndBits; | |||
6482 | if (AndBits == WidestAndBits && I->getOperand(0) == Load) | |||
6483 | AndsToMaybeRemove.push_back(I); | |||
6484 | break; | |||
6485 | } | |||
6486 | ||||
6487 | case Instruction::Shl: { | |||
6488 | auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1)); | |||
6489 | if (!ShlC) | |||
6490 | return false; | |||
6491 | uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1); | |||
6492 | DemandBits.setLowBits(BitWidth - ShiftAmt); | |||
6493 | break; | |||
6494 | } | |||
6495 | ||||
6496 | case Instruction::Trunc: { | |||
6497 | EVT TruncVT = TLI->getValueType(*DL, I->getType()); | |||
6498 | unsigned TruncBitWidth = TruncVT.getSizeInBits(); | |||
6499 | DemandBits.setLowBits(TruncBitWidth); | |||
6500 | break; | |||
6501 | } | |||
6502 | ||||
6503 | default: | |||
6504 | return false; | |||
6505 | } | |||
6506 | } | |||
6507 | ||||
6508 | uint32_t ActiveBits = DemandBits.getActiveBits(); | |||
6509 | // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the | |||
6510 | // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example, | |||
6511 | // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but | |||
6512 | // (and (load x) 1) is not matched as a single instruction, rather as a LDR | |||
6513 | // followed by an AND. | |||
6514 | // TODO: Look into removing this restriction by fixing backends to either | |||
6515 | // return false for isLoadExtLegal for i1 or have them select this pattern to | |||
6516 | // a single instruction. | |||
6517 | // | |||
6518 | // Also avoid hoisting if we didn't see any ands with the exact DemandBits | |||
6519 | // mask, since these are the only ands that will be removed by isel. | |||
6520 | if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) || | |||
6521 | WidestAndBits != DemandBits) | |||
6522 | return false; | |||
6523 | ||||
6524 | LLVMContext &Ctx = Load->getType()->getContext(); | |||
6525 | Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits); | |||
6526 | EVT TruncVT = TLI->getValueType(*DL, TruncTy); | |||
6527 | ||||
6528 | // Reject cases that won't be matched as extloads. | |||
6529 | if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() || | |||
6530 | !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT)) | |||
6531 | return false; | |||
6532 | ||||
6533 | IRBuilder<> Builder(Load->getNextNode()); | |||
6534 | auto *NewAnd = cast<Instruction>( | |||
6535 | Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits))); | |||
6536 | // Mark this instruction as "inserted by CGP", so that other | |||
6537 | // optimizations don't touch it. | |||
6538 | InsertedInsts.insert(NewAnd); | |||
6539 | ||||
6540 | // Replace all uses of load with new and (except for the use of load in the | |||
6541 | // new and itself). | |||
6542 | Load->replaceAllUsesWith(NewAnd); | |||
6543 | NewAnd->setOperand(0, Load); | |||
6544 | ||||
6545 | // Remove any and instructions that are now redundant. | |||
6546 | for (auto *And : AndsToMaybeRemove) | |||
6547 | // Check that the and mask is the same as the one we decided to put on the | |||
6548 | // new and. | |||
6549 | if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) { | |||
6550 | And->replaceAllUsesWith(NewAnd); | |||
6551 | if (&*CurInstIterator == And) | |||
6552 | CurInstIterator = std::next(And->getIterator()); | |||
6553 | And->eraseFromParent(); | |||
6554 | ++NumAndUses; | |||
6555 | } | |||
6556 | ||||
6557 | ++NumAndsAdded; | |||
6558 | return true; | |||
6559 | } | |||
6560 | ||||
6561 | /// Check if V (an operand of a select instruction) is an expensive instruction | |||
6562 | /// that is only used once. | |||
6563 | static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) { | |||
6564 | auto *I = dyn_cast<Instruction>(V); | |||
6565 | // If it's safe to speculatively execute, then it should not have side | |||
6566 | // effects; therefore, it's safe to sink and possibly *not* execute. | |||
6567 | return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) && | |||
6568 | TTI->getUserCost(I, TargetTransformInfo::TCK_SizeAndLatency) >= | |||
6569 | TargetTransformInfo::TCC_Expensive; | |||
6570 | } | |||
6571 | ||||
6572 | /// Returns true if a SelectInst should be turned into an explicit branch. | |||
6573 | static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI, | |||
6574 | const TargetLowering *TLI, | |||
6575 | SelectInst *SI) { | |||
6576 | // If even a predictable select is cheap, then a branch can't be cheaper. | |||
6577 | if (!TLI->isPredictableSelectExpensive()) | |||
6578 | return false; | |||
6579 | ||||
6580 | // FIXME: This should use the same heuristics as IfConversion to determine | |||
6581 | // whether a select is better represented as a branch. | |||
6582 | ||||
6583 | // If metadata tells us that the select condition is obviously predictable, | |||
6584 | // then we want to replace the select with a branch. | |||
6585 | uint64_t TrueWeight, FalseWeight; | |||
6586 | if (SI->extractProfMetadata(TrueWeight, FalseWeight)) { | |||
6587 | uint64_t Max = std::max(TrueWeight, FalseWeight); | |||
6588 | uint64_t Sum = TrueWeight + FalseWeight; | |||
6589 | if (Sum != 0) { | |||
6590 | auto Probability = BranchProbability::getBranchProbability(Max, Sum); | |||
6591 | if (Probability > TTI->getPredictableBranchThreshold()) | |||
6592 | return true; | |||
6593 | } | |||
6594 | } | |||
6595 | ||||
6596 | CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition()); | |||
6597 | ||||
6598 | // If a branch is predictable, an out-of-order CPU can avoid blocking on its | |||
6599 | // comparison condition. If the compare has more than one use, there's | |||
6600 | // probably another cmov or setcc around, so it's not worth emitting a branch. | |||
6601 | if (!Cmp || !Cmp->hasOneUse()) | |||
6602 | return false; | |||
6603 | ||||
6604 | // If either operand of the select is expensive and only needed on one side | |||
6605 | // of the select, we should form a branch. | |||
6606 | if (sinkSelectOperand(TTI, SI->getTrueValue()) || | |||
6607 | sinkSelectOperand(TTI, SI->getFalseValue())) | |||
6608 | return true; | |||
6609 | ||||
6610 | return false; | |||
6611 | } | |||
6612 | ||||
6613 | /// If \p isTrue is true, return the true value of \p SI, otherwise return | |||
6614 | /// false value of \p SI. If the true/false value of \p SI is defined by any | |||
6615 | /// select instructions in \p Selects, look through the defining select | |||
6616 | /// instruction until the true/false value is not defined in \p Selects. | |||
6617 | static Value *getTrueOrFalseValue( | |||
6618 | SelectInst *SI, bool isTrue, | |||
6619 | const SmallPtrSet<const Instruction *, 2> &Selects) { | |||
6620 | Value *V = nullptr; | |||
6621 | ||||
6622 | for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI); | |||
6623 | DefSI = dyn_cast<SelectInst>(V)) { | |||
6624 | assert(DefSI->getCondition() == SI->getCondition() &&(static_cast <bool> (DefSI->getCondition() == SI-> getCondition() && "The condition of DefSI does not match with SI" ) ? void (0) : __assert_fail ("DefSI->getCondition() == SI->getCondition() && \"The condition of DefSI does not match with SI\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6625, __extension__ __PRETTY_FUNCTION__ )) | |||
6625 | "The condition of DefSI does not match with SI")(static_cast <bool> (DefSI->getCondition() == SI-> getCondition() && "The condition of DefSI does not match with SI" ) ? void (0) : __assert_fail ("DefSI->getCondition() == SI->getCondition() && \"The condition of DefSI does not match with SI\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6625, __extension__ __PRETTY_FUNCTION__ )); | |||
6626 | V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue()); | |||
6627 | } | |||
6628 | ||||
6629 | assert(V && "Failed to get select true/false value")(static_cast <bool> (V && "Failed to get select true/false value" ) ? void (0) : __assert_fail ("V && \"Failed to get select true/false value\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6629, __extension__ __PRETTY_FUNCTION__ )); | |||
6630 | return V; | |||
6631 | } | |||
6632 | ||||
6633 | bool CodeGenPrepare::optimizeShiftInst(BinaryOperator *Shift) { | |||
6634 | assert(Shift->isShift() && "Expected a shift")(static_cast <bool> (Shift->isShift() && "Expected a shift" ) ? void (0) : __assert_fail ("Shift->isShift() && \"Expected a shift\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6634, __extension__ __PRETTY_FUNCTION__ )); | |||
6635 | ||||
6636 | // If this is (1) a vector shift, (2) shifts by scalars are cheaper than | |||
6637 | // general vector shifts, and (3) the shift amount is a select-of-splatted | |||
6638 | // values, hoist the shifts before the select: | |||
6639 | // shift Op0, (select Cond, TVal, FVal) --> | |||
6640 | // select Cond, (shift Op0, TVal), (shift Op0, FVal) | |||
6641 | // | |||
6642 | // This is inverting a generic IR transform when we know that the cost of a | |||
6643 | // general vector shift is more than the cost of 2 shift-by-scalars. | |||
6644 | // We can't do this effectively in SDAG because we may not be able to | |||
6645 | // determine if the select operands are splats from within a basic block. | |||
6646 | Type *Ty = Shift->getType(); | |||
6647 | if (!Ty->isVectorTy() || !TLI->isVectorShiftByScalarCheap(Ty)) | |||
6648 | return false; | |||
6649 | Value *Cond, *TVal, *FVal; | |||
6650 | if (!match(Shift->getOperand(1), | |||
6651 | m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal))))) | |||
6652 | return false; | |||
6653 | if (!isSplatValue(TVal) || !isSplatValue(FVal)) | |||
6654 | return false; | |||
6655 | ||||
6656 | IRBuilder<> Builder(Shift); | |||
6657 | BinaryOperator::BinaryOps Opcode = Shift->getOpcode(); | |||
6658 | Value *NewTVal = Builder.CreateBinOp(Opcode, Shift->getOperand(0), TVal); | |||
6659 | Value *NewFVal = Builder.CreateBinOp(Opcode, Shift->getOperand(0), FVal); | |||
6660 | Value *NewSel = Builder.CreateSelect(Cond, NewTVal, NewFVal); | |||
6661 | Shift->replaceAllUsesWith(NewSel); | |||
6662 | Shift->eraseFromParent(); | |||
6663 | return true; | |||
6664 | } | |||
6665 | ||||
6666 | bool CodeGenPrepare::optimizeFunnelShift(IntrinsicInst *Fsh) { | |||
6667 | Intrinsic::ID Opcode = Fsh->getIntrinsicID(); | |||
6668 | assert((Opcode == Intrinsic::fshl || Opcode == Intrinsic::fshr) &&(static_cast <bool> ((Opcode == Intrinsic::fshl || Opcode == Intrinsic::fshr) && "Expected a funnel shift") ? void (0) : __assert_fail ("(Opcode == Intrinsic::fshl || Opcode == Intrinsic::fshr) && \"Expected a funnel shift\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6669, __extension__ __PRETTY_FUNCTION__ )) | |||
6669 | "Expected a funnel shift")(static_cast <bool> ((Opcode == Intrinsic::fshl || Opcode == Intrinsic::fshr) && "Expected a funnel shift") ? void (0) : __assert_fail ("(Opcode == Intrinsic::fshl || Opcode == Intrinsic::fshr) && \"Expected a funnel shift\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6669, __extension__ __PRETTY_FUNCTION__ )); | |||
6670 | ||||
6671 | // If this is (1) a vector funnel shift, (2) shifts by scalars are cheaper | |||
6672 | // than general vector shifts, and (3) the shift amount is select-of-splatted | |||
6673 | // values, hoist the funnel shifts before the select: | |||
6674 | // fsh Op0, Op1, (select Cond, TVal, FVal) --> | |||
6675 | // select Cond, (fsh Op0, Op1, TVal), (fsh Op0, Op1, FVal) | |||
6676 | // | |||
6677 | // This is inverting a generic IR transform when we know that the cost of a | |||
6678 | // general vector shift is more than the cost of 2 shift-by-scalars. | |||
6679 | // We can't do this effectively in SDAG because we may not be able to | |||
6680 | // determine if the select operands are splats from within a basic block. | |||
6681 | Type *Ty = Fsh->getType(); | |||
6682 | if (!Ty->isVectorTy() || !TLI->isVectorShiftByScalarCheap(Ty)) | |||
6683 | return false; | |||
6684 | Value *Cond, *TVal, *FVal; | |||
6685 | if (!match(Fsh->getOperand(2), | |||
6686 | m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal))))) | |||
6687 | return false; | |||
6688 | if (!isSplatValue(TVal) || !isSplatValue(FVal)) | |||
6689 | return false; | |||
6690 | ||||
6691 | IRBuilder<> Builder(Fsh); | |||
6692 | Value *X = Fsh->getOperand(0), *Y = Fsh->getOperand(1); | |||
6693 | Value *NewTVal = Builder.CreateIntrinsic(Opcode, Ty, { X, Y, TVal }); | |||
6694 | Value *NewFVal = Builder.CreateIntrinsic(Opcode, Ty, { X, Y, FVal }); | |||
6695 | Value *NewSel = Builder.CreateSelect(Cond, NewTVal, NewFVal); | |||
6696 | Fsh->replaceAllUsesWith(NewSel); | |||
6697 | Fsh->eraseFromParent(); | |||
6698 | return true; | |||
6699 | } | |||
6700 | ||||
6701 | /// If we have a SelectInst that will likely profit from branch prediction, | |||
6702 | /// turn it into a branch. | |||
6703 | bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) { | |||
6704 | if (DisableSelectToBranch) | |||
6705 | return false; | |||
6706 | ||||
6707 | // Find all consecutive select instructions that share the same condition. | |||
6708 | SmallVector<SelectInst *, 2> ASI; | |||
6709 | ASI.push_back(SI); | |||
6710 | for (BasicBlock::iterator It = ++BasicBlock::iterator(SI); | |||
6711 | It != SI->getParent()->end(); ++It) { | |||
6712 | SelectInst *I = dyn_cast<SelectInst>(&*It); | |||
6713 | if (I && SI->getCondition() == I->getCondition()) { | |||
6714 | ASI.push_back(I); | |||
6715 | } else { | |||
6716 | break; | |||
6717 | } | |||
6718 | } | |||
6719 | ||||
6720 | SelectInst *LastSI = ASI.back(); | |||
6721 | // Increment the current iterator to skip all the rest of select instructions | |||
6722 | // because they will be either "not lowered" or "all lowered" to branch. | |||
6723 | CurInstIterator = std::next(LastSI->getIterator()); | |||
6724 | ||||
6725 | bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1); | |||
6726 | ||||
6727 | // Can we convert the 'select' to CF ? | |||
6728 | if (VectorCond || SI->getMetadata(LLVMContext::MD_unpredictable)) | |||
6729 | return false; | |||
6730 | ||||
6731 | TargetLowering::SelectSupportKind SelectKind; | |||
6732 | if (VectorCond) | |||
6733 | SelectKind = TargetLowering::VectorMaskSelect; | |||
6734 | else if (SI->getType()->isVectorTy()) | |||
6735 | SelectKind = TargetLowering::ScalarCondVectorVal; | |||
6736 | else | |||
6737 | SelectKind = TargetLowering::ScalarValSelect; | |||
6738 | ||||
6739 | if (TLI->isSelectSupported(SelectKind) && | |||
6740 | (!isFormingBranchFromSelectProfitable(TTI, TLI, SI) || OptSize || | |||
6741 | llvm::shouldOptimizeForSize(SI->getParent(), PSI, BFI.get()))) | |||
6742 | return false; | |||
6743 | ||||
6744 | // The DominatorTree needs to be rebuilt by any consumers after this | |||
6745 | // transformation. We simply reset here rather than setting the ModifiedDT | |||
6746 | // flag to avoid restarting the function walk in runOnFunction for each | |||
6747 | // select optimized. | |||
6748 | DT.reset(); | |||
6749 | ||||
6750 | // Transform a sequence like this: | |||
6751 | // start: | |||
6752 | // %cmp = cmp uge i32 %a, %b | |||
6753 | // %sel = select i1 %cmp, i32 %c, i32 %d | |||
6754 | // | |||
6755 | // Into: | |||
6756 | // start: | |||
6757 | // %cmp = cmp uge i32 %a, %b | |||
6758 | // %cmp.frozen = freeze %cmp | |||
6759 | // br i1 %cmp.frozen, label %select.true, label %select.false | |||
6760 | // select.true: | |||
6761 | // br label %select.end | |||
6762 | // select.false: | |||
6763 | // br label %select.end | |||
6764 | // select.end: | |||
6765 | // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ] | |||
6766 | // | |||
6767 | // %cmp should be frozen, otherwise it may introduce undefined behavior. | |||
6768 | // In addition, we may sink instructions that produce %c or %d from | |||
6769 | // the entry block into the destination(s) of the new branch. | |||
6770 | // If the true or false blocks do not contain a sunken instruction, that | |||
6771 | // block and its branch may be optimized away. In that case, one side of the | |||
6772 | // first branch will point directly to select.end, and the corresponding PHI | |||
6773 | // predecessor block will be the start block. | |||
6774 | ||||
6775 | // First, we split the block containing the select into 2 blocks. | |||
6776 | BasicBlock *StartBlock = SI->getParent(); | |||
6777 | BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI)); | |||
6778 | BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end"); | |||
6779 | BFI->setBlockFreq(EndBlock, BFI->getBlockFreq(StartBlock).getFrequency()); | |||
6780 | ||||
6781 | // Delete the unconditional branch that was just created by the split. | |||
6782 | StartBlock->getTerminator()->eraseFromParent(); | |||
6783 | ||||
6784 | // These are the new basic blocks for the conditional branch. | |||
6785 | // At least one will become an actual new basic block. | |||
6786 | BasicBlock *TrueBlock = nullptr; | |||
6787 | BasicBlock *FalseBlock = nullptr; | |||
6788 | BranchInst *TrueBranch = nullptr; | |||
6789 | BranchInst *FalseBranch = nullptr; | |||
6790 | ||||
6791 | // Sink expensive instructions into the conditional blocks to avoid executing | |||
6792 | // them speculatively. | |||
6793 | for (SelectInst *SI : ASI) { | |||
6794 | if (sinkSelectOperand(TTI, SI->getTrueValue())) { | |||
6795 | if (TrueBlock == nullptr) { | |||
6796 | TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink", | |||
6797 | EndBlock->getParent(), EndBlock); | |||
6798 | TrueBranch = BranchInst::Create(EndBlock, TrueBlock); | |||
6799 | TrueBranch->setDebugLoc(SI->getDebugLoc()); | |||
6800 | } | |||
6801 | auto *TrueInst = cast<Instruction>(SI->getTrueValue()); | |||
6802 | TrueInst->moveBefore(TrueBranch); | |||
6803 | } | |||
6804 | if (sinkSelectOperand(TTI, SI->getFalseValue())) { | |||
6805 | if (FalseBlock == nullptr) { | |||
6806 | FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink", | |||
6807 | EndBlock->getParent(), EndBlock); | |||
6808 | FalseBranch = BranchInst::Create(EndBlock, FalseBlock); | |||
6809 | FalseBranch->setDebugLoc(SI->getDebugLoc()); | |||
6810 | } | |||
6811 | auto *FalseInst = cast<Instruction>(SI->getFalseValue()); | |||
6812 | FalseInst->moveBefore(FalseBranch); | |||
6813 | } | |||
6814 | } | |||
6815 | ||||
6816 | // If there was nothing to sink, then arbitrarily choose the 'false' side | |||
6817 | // for a new input value to the PHI. | |||
6818 | if (TrueBlock == FalseBlock) { | |||
6819 | assert(TrueBlock == nullptr &&(static_cast <bool> (TrueBlock == nullptr && "Unexpected basic block transform while optimizing select" ) ? void (0) : __assert_fail ("TrueBlock == nullptr && \"Unexpected basic block transform while optimizing select\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6820, __extension__ __PRETTY_FUNCTION__ )) | |||
6820 | "Unexpected basic block transform while optimizing select")(static_cast <bool> (TrueBlock == nullptr && "Unexpected basic block transform while optimizing select" ) ? void (0) : __assert_fail ("TrueBlock == nullptr && \"Unexpected basic block transform while optimizing select\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6820, __extension__ __PRETTY_FUNCTION__ )); | |||
6821 | ||||
6822 | FalseBlock = BasicBlock::Create(SI->getContext(), "select.false", | |||
6823 | EndBlock->getParent(), EndBlock); | |||
6824 | auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock); | |||
6825 | FalseBranch->setDebugLoc(SI->getDebugLoc()); | |||
6826 | } | |||
6827 | ||||
6828 | // Insert the real conditional branch based on the original condition. | |||
6829 | // If we did not create a new block for one of the 'true' or 'false' paths | |||
6830 | // of the condition, it means that side of the branch goes to the end block | |||
6831 | // directly and the path originates from the start block from the point of | |||
6832 | // view of the new PHI. | |||
6833 | BasicBlock *TT, *FT; | |||
6834 | if (TrueBlock == nullptr) { | |||
6835 | TT = EndBlock; | |||
6836 | FT = FalseBlock; | |||
6837 | TrueBlock = StartBlock; | |||
6838 | } else if (FalseBlock == nullptr) { | |||
6839 | TT = TrueBlock; | |||
6840 | FT = EndBlock; | |||
6841 | FalseBlock = StartBlock; | |||
6842 | } else { | |||
6843 | TT = TrueBlock; | |||
6844 | FT = FalseBlock; | |||
6845 | } | |||
6846 | IRBuilder<> IB(SI); | |||
6847 | auto *CondFr = IB.CreateFreeze(SI->getCondition(), SI->getName() + ".frozen"); | |||
6848 | IB.CreateCondBr(CondFr, TT, FT, SI); | |||
6849 | ||||
6850 | SmallPtrSet<const Instruction *, 2> INS; | |||
6851 | INS.insert(ASI.begin(), ASI.end()); | |||
6852 | // Use reverse iterator because later select may use the value of the | |||
6853 | // earlier select, and we need to propagate value through earlier select | |||
6854 | // to get the PHI operand. | |||
6855 | for (SelectInst *SI : llvm::reverse(ASI)) { | |||
6856 | // The select itself is replaced with a PHI Node. | |||
6857 | PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front()); | |||
6858 | PN->takeName(SI); | |||
6859 | PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock); | |||
6860 | PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock); | |||
6861 | PN->setDebugLoc(SI->getDebugLoc()); | |||
6862 | ||||
6863 | SI->replaceAllUsesWith(PN); | |||
6864 | SI->eraseFromParent(); | |||
6865 | INS.erase(SI); | |||
6866 | ++NumSelectsExpanded; | |||
6867 | } | |||
6868 | ||||
6869 | // Instruct OptimizeBlock to skip to the next block. | |||
6870 | CurInstIterator = StartBlock->end(); | |||
6871 | return true; | |||
6872 | } | |||
6873 | ||||
6874 | /// Some targets only accept certain types for splat inputs. For example a VDUP | |||
6875 | /// in MVE takes a GPR (integer) register, and the instruction that incorporate | |||
6876 | /// a VDUP (such as a VADD qd, qm, rm) also require a gpr register. | |||
6877 | bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) { | |||
6878 | // Accept shuf(insertelem(undef/poison, val, 0), undef/poison, <0,0,..>) only | |||
6879 | if (!match(SVI, m_Shuffle(m_InsertElt(m_Undef(), m_Value(), m_ZeroInt()), | |||
6880 | m_Undef(), m_ZeroMask()))) | |||
6881 | return false; | |||
6882 | Type *NewType = TLI->shouldConvertSplatType(SVI); | |||
6883 | if (!NewType) | |||
6884 | return false; | |||
6885 | ||||
6886 | auto *SVIVecType = cast<FixedVectorType>(SVI->getType()); | |||
6887 | assert(!NewType->isVectorTy() && "Expected a scalar type!")(static_cast <bool> (!NewType->isVectorTy() && "Expected a scalar type!") ? void (0) : __assert_fail ("!NewType->isVectorTy() && \"Expected a scalar type!\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6887, __extension__ __PRETTY_FUNCTION__ )); | |||
6888 | assert(NewType->getScalarSizeInBits() == SVIVecType->getScalarSizeInBits() &&(static_cast <bool> (NewType->getScalarSizeInBits() == SVIVecType->getScalarSizeInBits() && "Expected a type of the same size!" ) ? void (0) : __assert_fail ("NewType->getScalarSizeInBits() == SVIVecType->getScalarSizeInBits() && \"Expected a type of the same size!\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6889, __extension__ __PRETTY_FUNCTION__ )) | |||
6889 | "Expected a type of the same size!")(static_cast <bool> (NewType->getScalarSizeInBits() == SVIVecType->getScalarSizeInBits() && "Expected a type of the same size!" ) ? void (0) : __assert_fail ("NewType->getScalarSizeInBits() == SVIVecType->getScalarSizeInBits() && \"Expected a type of the same size!\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6889, __extension__ __PRETTY_FUNCTION__ )); | |||
6890 | auto *NewVecType = | |||
6891 | FixedVectorType::get(NewType, SVIVecType->getNumElements()); | |||
6892 | ||||
6893 | // Create a bitcast (shuffle (insert (bitcast(..)))) | |||
6894 | IRBuilder<> Builder(SVI->getContext()); | |||
6895 | Builder.SetInsertPoint(SVI); | |||
6896 | Value *BC1 = Builder.CreateBitCast( | |||
6897 | cast<Instruction>(SVI->getOperand(0))->getOperand(1), NewType); | |||
6898 | Value *Shuffle = Builder.CreateVectorSplat(NewVecType->getNumElements(), BC1); | |||
6899 | Value *BC2 = Builder.CreateBitCast(Shuffle, SVIVecType); | |||
6900 | ||||
6901 | SVI->replaceAllUsesWith(BC2); | |||
6902 | RecursivelyDeleteTriviallyDeadInstructions( | |||
6903 | SVI, TLInfo, nullptr, [&](Value *V) { removeAllAssertingVHReferences(V); }); | |||
6904 | ||||
6905 | // Also hoist the bitcast up to its operand if it they are not in the same | |||
6906 | // block. | |||
6907 | if (auto *BCI = dyn_cast<Instruction>(BC1)) | |||
6908 | if (auto *Op = dyn_cast<Instruction>(BCI->getOperand(0))) | |||
6909 | if (BCI->getParent() != Op->getParent() && !isa<PHINode>(Op) && | |||
6910 | !Op->isTerminator() && !Op->isEHPad()) | |||
6911 | BCI->moveAfter(Op); | |||
6912 | ||||
6913 | return true; | |||
6914 | } | |||
6915 | ||||
6916 | bool CodeGenPrepare::tryToSinkFreeOperands(Instruction *I) { | |||
6917 | // If the operands of I can be folded into a target instruction together with | |||
6918 | // I, duplicate and sink them. | |||
6919 | SmallVector<Use *, 4> OpsToSink; | |||
6920 | if (!TLI->shouldSinkOperands(I, OpsToSink)) | |||
6921 | return false; | |||
6922 | ||||
6923 | // OpsToSink can contain multiple uses in a use chain (e.g. | |||
6924 | // (%u1 with %u1 = shufflevector), (%u2 with %u2 = zext %u1)). The dominating | |||
6925 | // uses must come first, so we process the ops in reverse order so as to not | |||
6926 | // create invalid IR. | |||
6927 | BasicBlock *TargetBB = I->getParent(); | |||
6928 | bool Changed = false; | |||
6929 | SmallVector<Use *, 4> ToReplace; | |||
6930 | Instruction *InsertPoint = I; | |||
6931 | DenseMap<const Instruction *, unsigned long> InstOrdering; | |||
6932 | unsigned long InstNumber = 0; | |||
6933 | for (const auto &I : *TargetBB) | |||
6934 | InstOrdering[&I] = InstNumber++; | |||
6935 | ||||
6936 | for (Use *U : reverse(OpsToSink)) { | |||
6937 | auto *UI = cast<Instruction>(U->get()); | |||
6938 | if (isa<PHINode>(UI)) | |||
6939 | continue; | |||
6940 | if (UI->getParent() == TargetBB) { | |||
6941 | if (InstOrdering[UI] < InstOrdering[InsertPoint]) | |||
6942 | InsertPoint = UI; | |||
6943 | continue; | |||
6944 | } | |||
6945 | ToReplace.push_back(U); | |||
6946 | } | |||
6947 | ||||
6948 | SetVector<Instruction *> MaybeDead; | |||
6949 | DenseMap<Instruction *, Instruction *> NewInstructions; | |||
6950 | for (Use *U : ToReplace) { | |||
6951 | auto *UI = cast<Instruction>(U->get()); | |||
6952 | Instruction *NI = UI->clone(); | |||
6953 | NewInstructions[UI] = NI; | |||
6954 | MaybeDead.insert(UI); | |||
6955 | LLVM_DEBUG(dbgs() << "Sinking " << *UI << " to user " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Sinking " << *UI << " to user " << *I << "\n"; } } while (false ); | |||
6956 | NI->insertBefore(InsertPoint); | |||
6957 | InsertPoint = NI; | |||
6958 | InsertedInsts.insert(NI); | |||
6959 | ||||
6960 | // Update the use for the new instruction, making sure that we update the | |||
6961 | // sunk instruction uses, if it is part of a chain that has already been | |||
6962 | // sunk. | |||
6963 | Instruction *OldI = cast<Instruction>(U->getUser()); | |||
6964 | if (NewInstructions.count(OldI)) | |||
6965 | NewInstructions[OldI]->setOperand(U->getOperandNo(), NI); | |||
6966 | else | |||
6967 | U->set(NI); | |||
6968 | Changed = true; | |||
6969 | } | |||
6970 | ||||
6971 | // Remove instructions that are dead after sinking. | |||
6972 | for (auto *I : MaybeDead) { | |||
6973 | if (!I->hasNUsesOrMore(1)) { | |||
6974 | LLVM_DEBUG(dbgs() << "Removing dead instruction: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Removing dead instruction: " << *I << "\n"; } } while (false); | |||
6975 | I->eraseFromParent(); | |||
6976 | } | |||
6977 | } | |||
6978 | ||||
6979 | return Changed; | |||
6980 | } | |||
6981 | ||||
6982 | bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) { | |||
6983 | Value *Cond = SI->getCondition(); | |||
6984 | Type *OldType = Cond->getType(); | |||
6985 | LLVMContext &Context = Cond->getContext(); | |||
6986 | EVT OldVT = TLI->getValueType(*DL, OldType); | |||
6987 | MVT RegType = TLI->getRegisterType(Context, OldVT); | |||
6988 | unsigned RegWidth = RegType.getSizeInBits(); | |||
6989 | ||||
6990 | if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth()) | |||
6991 | return false; | |||
6992 | ||||
6993 | // If the register width is greater than the type width, expand the condition | |||
6994 | // of the switch instruction and each case constant to the width of the | |||
6995 | // register. By widening the type of the switch condition, subsequent | |||
6996 | // comparisons (for case comparisons) will not need to be extended to the | |||
6997 | // preferred register width, so we will potentially eliminate N-1 extends, | |||
6998 | // where N is the number of cases in the switch. | |||
6999 | auto *NewType = Type::getIntNTy(Context, RegWidth); | |||
7000 | ||||
7001 | // Extend the switch condition and case constants using the target preferred | |||
7002 | // extend unless the switch condition is a function argument with an extend | |||
7003 | // attribute. In that case, we can avoid an unnecessary mask/extension by | |||
7004 | // matching the argument extension instead. | |||
7005 | Instruction::CastOps ExtType = Instruction::ZExt; | |||
7006 | // Some targets prefer SExt over ZExt. | |||
7007 | if (TLI->isSExtCheaperThanZExt(OldVT, RegType)) | |||
7008 | ExtType = Instruction::SExt; | |||
7009 | ||||
7010 | if (auto *Arg = dyn_cast<Argument>(Cond)) { | |||
7011 | if (Arg->hasSExtAttr()) | |||
7012 | ExtType = Instruction::SExt; | |||
7013 | if (Arg->hasZExtAttr()) | |||
7014 | ExtType = Instruction::ZExt; | |||
7015 | } | |||
7016 | ||||
7017 | auto *ExtInst = CastInst::Create(ExtType, Cond, NewType); | |||
7018 | ExtInst->insertBefore(SI); | |||
7019 | ExtInst->setDebugLoc(SI->getDebugLoc()); | |||
7020 | SI->setCondition(ExtInst); | |||
7021 | for (auto Case : SI->cases()) { | |||
7022 | APInt NarrowConst = Case.getCaseValue()->getValue(); | |||
7023 | APInt WideConst = (ExtType == Instruction::ZExt) ? | |||
7024 | NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth); | |||
7025 | Case.setValue(ConstantInt::get(Context, WideConst)); | |||
7026 | } | |||
7027 | ||||
7028 | return true; | |||
7029 | } | |||
7030 | ||||
7031 | ||||
7032 | namespace { | |||
7033 | ||||
7034 | /// Helper class to promote a scalar operation to a vector one. | |||
7035 | /// This class is used to move downward extractelement transition. | |||
7036 | /// E.g., | |||
7037 | /// a = vector_op <2 x i32> | |||
7038 | /// b = extractelement <2 x i32> a, i32 0 | |||
7039 | /// c = scalar_op b | |||
7040 | /// store c | |||
7041 | /// | |||
7042 | /// => | |||
7043 | /// a = vector_op <2 x i32> | |||
7044 | /// c = vector_op a (equivalent to scalar_op on the related lane) | |||
7045 | /// * d = extractelement <2 x i32> c, i32 0 | |||
7046 | /// * store d | |||
7047 | /// Assuming both extractelement and store can be combine, we get rid of the | |||
7048 | /// transition. | |||
7049 | class VectorPromoteHelper { | |||
7050 | /// DataLayout associated with the current module. | |||
7051 | const DataLayout &DL; | |||
7052 | ||||
7053 | /// Used to perform some checks on the legality of vector operations. | |||
7054 | const TargetLowering &TLI; | |||
7055 | ||||
7056 | /// Used to estimated the cost of the promoted chain. | |||
7057 | const TargetTransformInfo &TTI; | |||
7058 | ||||
7059 | /// The transition being moved downwards. | |||
7060 | Instruction *Transition; | |||
7061 | ||||
7062 | /// The sequence of instructions to be promoted. | |||
7063 | SmallVector<Instruction *, 4> InstsToBePromoted; | |||
7064 | ||||
7065 | /// Cost of combining a store and an extract. | |||
7066 | unsigned StoreExtractCombineCost; | |||
7067 | ||||
7068 | /// Instruction that will be combined with the transition. | |||
7069 | Instruction *CombineInst = nullptr; | |||
7070 | ||||
7071 | /// The instruction that represents the current end of the transition. | |||
7072 | /// Since we are faking the promotion until we reach the end of the chain | |||
7073 | /// of computation, we need a way to get the current end of the transition. | |||
7074 | Instruction *getEndOfTransition() const { | |||
7075 | if (InstsToBePromoted.empty()) | |||
7076 | return Transition; | |||
7077 | return InstsToBePromoted.back(); | |||
7078 | } | |||
7079 | ||||
7080 | /// Return the index of the original value in the transition. | |||
7081 | /// E.g., for "extractelement <2 x i32> c, i32 1" the original value, | |||
7082 | /// c, is at index 0. | |||
7083 | unsigned getTransitionOriginalValueIdx() const { | |||
7084 | assert(isa<ExtractElementInst>(Transition) &&(static_cast <bool> (isa<ExtractElementInst>(Transition ) && "Other kind of transitions are not supported yet" ) ? void (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7085, __extension__ __PRETTY_FUNCTION__ )) | |||
7085 | "Other kind of transitions are not supported yet")(static_cast <bool> (isa<ExtractElementInst>(Transition ) && "Other kind of transitions are not supported yet" ) ? void (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7085, __extension__ __PRETTY_FUNCTION__ )); | |||
7086 | return 0; | |||
7087 | } | |||
7088 | ||||
7089 | /// Return the index of the index in the transition. | |||
7090 | /// E.g., for "extractelement <2 x i32> c, i32 0" the index | |||
7091 | /// is at index 1. | |||
7092 | unsigned getTransitionIdx() const { | |||
7093 | assert(isa<ExtractElementInst>(Transition) &&(static_cast <bool> (isa<ExtractElementInst>(Transition ) && "Other kind of transitions are not supported yet" ) ? void (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7094, __extension__ __PRETTY_FUNCTION__ )) | |||
7094 | "Other kind of transitions are not supported yet")(static_cast <bool> (isa<ExtractElementInst>(Transition ) && "Other kind of transitions are not supported yet" ) ? void (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7094, __extension__ __PRETTY_FUNCTION__ )); | |||
7095 | return 1; | |||
7096 | } | |||
7097 | ||||
7098 | /// Get the type of the transition. | |||
7099 | /// This is the type of the original value. | |||
7100 | /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the | |||
7101 | /// transition is <2 x i32>. | |||
7102 | Type *getTransitionType() const { | |||
7103 | return Transition->getOperand(getTransitionOriginalValueIdx())->getType(); | |||
7104 | } | |||
7105 | ||||
7106 | /// Promote \p ToBePromoted by moving \p Def downward through. | |||
7107 | /// I.e., we have the following sequence: | |||
7108 | /// Def = Transition <ty1> a to <ty2> | |||
7109 | /// b = ToBePromoted <ty2> Def, ... | |||
7110 | /// => | |||
7111 | /// b = ToBePromoted <ty1> a, ... | |||
7112 | /// Def = Transition <ty1> ToBePromoted to <ty2> | |||
7113 | void promoteImpl(Instruction *ToBePromoted); | |||
7114 | ||||
7115 | /// Check whether or not it is profitable to promote all the | |||
7116 | /// instructions enqueued to be promoted. | |||
7117 | bool isProfitableToPromote() { | |||
7118 | Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx()); | |||
7119 | unsigned Index = isa<ConstantInt>(ValIdx) | |||
7120 | ? cast<ConstantInt>(ValIdx)->getZExtValue() | |||
7121 | : -1; | |||
7122 | Type *PromotedType = getTransitionType(); | |||
7123 | ||||
7124 | StoreInst *ST = cast<StoreInst>(CombineInst); | |||
7125 | unsigned AS = ST->getPointerAddressSpace(); | |||
7126 | // Check if this store is supported. | |||
7127 | if (!TLI.allowsMisalignedMemoryAccesses( | |||
7128 | TLI.getValueType(DL, ST->getValueOperand()->getType()), AS, | |||
7129 | ST->getAlign())) { | |||
7130 | // If this is not supported, there is no way we can combine | |||
7131 | // the extract with the store. | |||
7132 | return false; | |||
7133 | } | |||
7134 | ||||
7135 | // The scalar chain of computation has to pay for the transition | |||
7136 | // scalar to vector. | |||
7137 | // The vector chain has to account for the combining cost. | |||
7138 | InstructionCost ScalarCost = | |||
7139 | TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index); | |||
7140 | InstructionCost VectorCost = StoreExtractCombineCost; | |||
7141 | enum TargetTransformInfo::TargetCostKind CostKind = | |||
7142 | TargetTransformInfo::TCK_RecipThroughput; | |||
7143 | for (const auto &Inst : InstsToBePromoted) { | |||
7144 | // Compute the cost. | |||
7145 | // By construction, all instructions being promoted are arithmetic ones. | |||
7146 | // Moreover, one argument is a constant that can be viewed as a splat | |||
7147 | // constant. | |||
7148 | Value *Arg0 = Inst->getOperand(0); | |||
7149 | bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) || | |||
7150 | isa<ConstantFP>(Arg0); | |||
7151 | TargetTransformInfo::OperandValueKind Arg0OVK = | |||
7152 | IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue | |||
7153 | : TargetTransformInfo::OK_AnyValue; | |||
7154 | TargetTransformInfo::OperandValueKind Arg1OVK = | |||
7155 | !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue | |||
7156 | : TargetTransformInfo::OK_AnyValue; | |||
7157 | ScalarCost += TTI.getArithmeticInstrCost( | |||
7158 | Inst->getOpcode(), Inst->getType(), CostKind, Arg0OVK, Arg1OVK); | |||
7159 | VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType, | |||
7160 | CostKind, | |||
7161 | Arg0OVK, Arg1OVK); | |||
7162 | } | |||
7163 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Estimated cost of computation to be promoted:\nScalar: " << ScalarCost << "\nVector: " << VectorCost << '\n'; } } while (false) | |||
7164 | dbgs() << "Estimated cost of computation to be promoted:\nScalar: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Estimated cost of computation to be promoted:\nScalar: " << ScalarCost << "\nVector: " << VectorCost << '\n'; } } while (false) | |||
7165 | << ScalarCost << "\nVector: " << VectorCost << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Estimated cost of computation to be promoted:\nScalar: " << ScalarCost << "\nVector: " << VectorCost << '\n'; } } while (false); | |||
7166 | return ScalarCost > VectorCost; | |||
7167 | } | |||
7168 | ||||
7169 | /// Generate a constant vector with \p Val with the same | |||
7170 | /// number of elements as the transition. | |||
7171 | /// \p UseSplat defines whether or not \p Val should be replicated | |||
7172 | /// across the whole vector. | |||
7173 | /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>, | |||
7174 | /// otherwise we generate a vector with as many undef as possible: | |||
7175 | /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only | |||
7176 | /// used at the index of the extract. | |||
7177 | Value *getConstantVector(Constant *Val, bool UseSplat) const { | |||
7178 | unsigned ExtractIdx = std::numeric_limits<unsigned>::max(); | |||
7179 | if (!UseSplat) { | |||
7180 | // If we cannot determine where the constant must be, we have to | |||
7181 | // use a splat constant. | |||
7182 | Value *ValExtractIdx = Transition->getOperand(getTransitionIdx()); | |||
7183 | if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx)) | |||
7184 | ExtractIdx = CstVal->getSExtValue(); | |||
7185 | else | |||
7186 | UseSplat = true; | |||
7187 | } | |||
7188 | ||||
7189 | ElementCount EC = cast<VectorType>(getTransitionType())->getElementCount(); | |||
7190 | if (UseSplat) | |||
7191 | return ConstantVector::getSplat(EC, Val); | |||
7192 | ||||
7193 | if (!EC.isScalable()) { | |||
7194 | SmallVector<Constant *, 4> ConstVec; | |||
7195 | UndefValue *UndefVal = UndefValue::get(Val->getType()); | |||
7196 | for (unsigned Idx = 0; Idx != EC.getKnownMinValue(); ++Idx) { | |||
7197 | if (Idx == ExtractIdx) | |||
7198 | ConstVec.push_back(Val); | |||
7199 | else | |||
7200 | ConstVec.push_back(UndefVal); | |||
7201 | } | |||
7202 | return ConstantVector::get(ConstVec); | |||
7203 | } else | |||
7204 | llvm_unreachable(::llvm::llvm_unreachable_internal("Generate scalable vector for non-splat is unimplemented" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7205) | |||
7205 | "Generate scalable vector for non-splat is unimplemented")::llvm::llvm_unreachable_internal("Generate scalable vector for non-splat is unimplemented" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7205); | |||
7206 | } | |||
7207 | ||||
7208 | /// Check if promoting to a vector type an operand at \p OperandIdx | |||
7209 | /// in \p Use can trigger undefined behavior. | |||
7210 | static bool canCauseUndefinedBehavior(const Instruction *Use, | |||
7211 | unsigned OperandIdx) { | |||
7212 | // This is not safe to introduce undef when the operand is on | |||
7213 | // the right hand side of a division-like instruction. | |||
7214 | if (OperandIdx != 1) | |||
7215 | return false; | |||
7216 | switch (Use->getOpcode()) { | |||
7217 | default: | |||
7218 | return false; | |||
7219 | case Instruction::SDiv: | |||
7220 | case Instruction::UDiv: | |||
7221 | case Instruction::SRem: | |||
7222 | case Instruction::URem: | |||
7223 | return true; | |||
7224 | case Instruction::FDiv: | |||
7225 | case Instruction::FRem: | |||
7226 | return !Use->hasNoNaNs(); | |||
7227 | } | |||
7228 | llvm_unreachable(nullptr)::llvm::llvm_unreachable_internal(nullptr, "llvm/lib/CodeGen/CodeGenPrepare.cpp" , 7228); | |||
7229 | } | |||
7230 | ||||
7231 | public: | |||
7232 | VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI, | |||
7233 | const TargetTransformInfo &TTI, Instruction *Transition, | |||
7234 | unsigned CombineCost) | |||
7235 | : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition), | |||
7236 | StoreExtractCombineCost(CombineCost) { | |||
7237 | assert(Transition && "Do not know how to promote null")(static_cast <bool> (Transition && "Do not know how to promote null" ) ? void (0) : __assert_fail ("Transition && \"Do not know how to promote null\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7237, __extension__ __PRETTY_FUNCTION__ )); | |||
7238 | } | |||
7239 | ||||
7240 | /// Check if we can promote \p ToBePromoted to \p Type. | |||
7241 | bool canPromote(const Instruction *ToBePromoted) const { | |||
7242 | // We could support CastInst too. | |||
7243 | return isa<BinaryOperator>(ToBePromoted); | |||
7244 | } | |||
7245 | ||||
7246 | /// Check if it is profitable to promote \p ToBePromoted | |||
7247 | /// by moving downward the transition through. | |||
7248 | bool shouldPromote(const Instruction *ToBePromoted) const { | |||
7249 | // Promote only if all the operands can be statically expanded. | |||
7250 | // Indeed, we do not want to introduce any new kind of transitions. | |||
7251 | for (const Use &U : ToBePromoted->operands()) { | |||
7252 | const Value *Val = U.get(); | |||
7253 | if (Val == getEndOfTransition()) { | |||
7254 | // If the use is a division and the transition is on the rhs, | |||
7255 | // we cannot promote the operation, otherwise we may create a | |||
7256 | // division by zero. | |||
7257 | if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo())) | |||
7258 | return false; | |||
7259 | continue; | |||
7260 | } | |||
7261 | if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) && | |||
7262 | !isa<ConstantFP>(Val)) | |||
7263 | return false; | |||
7264 | } | |||
7265 | // Check that the resulting operation is legal. | |||
7266 | int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode()); | |||
7267 | if (!ISDOpcode) | |||
7268 | return false; | |||
7269 | return StressStoreExtract || | |||
7270 | TLI.isOperationLegalOrCustom( | |||
7271 | ISDOpcode, TLI.getValueType(DL, getTransitionType(), true)); | |||
7272 | } | |||
7273 | ||||
7274 | /// Check whether or not \p Use can be combined | |||
7275 | /// with the transition. | |||
7276 | /// I.e., is it possible to do Use(Transition) => AnotherUse? | |||
7277 | bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); } | |||
7278 | ||||
7279 | /// Record \p ToBePromoted as part of the chain to be promoted. | |||
7280 | void enqueueForPromotion(Instruction *ToBePromoted) { | |||
7281 | InstsToBePromoted.push_back(ToBePromoted); | |||
7282 | } | |||
7283 | ||||
7284 | /// Set the instruction that will be combined with the transition. | |||
7285 | void recordCombineInstruction(Instruction *ToBeCombined) { | |||
7286 | assert(canCombine(ToBeCombined) && "Unsupported instruction to combine")(static_cast <bool> (canCombine(ToBeCombined) && "Unsupported instruction to combine") ? void (0) : __assert_fail ("canCombine(ToBeCombined) && \"Unsupported instruction to combine\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7286, __extension__ __PRETTY_FUNCTION__ )); | |||
7287 | CombineInst = ToBeCombined; | |||
7288 | } | |||
7289 | ||||
7290 | /// Promote all the instructions enqueued for promotion if it is | |||
7291 | /// is profitable. | |||
7292 | /// \return True if the promotion happened, false otherwise. | |||
7293 | bool promote() { | |||
7294 | // Check if there is something to promote. | |||
7295 | // Right now, if we do not have anything to combine with, | |||
7296 | // we assume the promotion is not profitable. | |||
7297 | if (InstsToBePromoted.empty() || !CombineInst) | |||
7298 | return false; | |||
7299 | ||||
7300 | // Check cost. | |||
7301 | if (!StressStoreExtract && !isProfitableToPromote()) | |||
7302 | return false; | |||
7303 | ||||
7304 | // Promote. | |||
7305 | for (auto &ToBePromoted : InstsToBePromoted) | |||
7306 | promoteImpl(ToBePromoted); | |||
7307 | InstsToBePromoted.clear(); | |||
7308 | return true; | |||
7309 | } | |||
7310 | }; | |||
7311 | ||||
7312 | } // end anonymous namespace | |||
7313 | ||||
7314 | void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) { | |||
7315 | // At this point, we know that all the operands of ToBePromoted but Def | |||
7316 | // can be statically promoted. | |||
7317 | // For Def, we need to use its parameter in ToBePromoted: | |||
7318 | // b = ToBePromoted ty1 a | |||
7319 | // Def = Transition ty1 b to ty2 | |||
7320 | // Move the transition down. | |||
7321 | // 1. Replace all uses of the promoted operation by the transition. | |||
7322 | // = ... b => = ... Def. | |||
7323 | assert(ToBePromoted->getType() == Transition->getType() &&(static_cast <bool> (ToBePromoted->getType() == Transition ->getType() && "The type of the result of the transition does not match " "the final type") ? void (0) : __assert_fail ("ToBePromoted->getType() == Transition->getType() && \"The type of the result of the transition does not match \" \"the final type\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7325, __extension__ __PRETTY_FUNCTION__ )) | |||
7324 | "The type of the result of the transition does not match "(static_cast <bool> (ToBePromoted->getType() == Transition ->getType() && "The type of the result of the transition does not match " "the final type") ? void (0) : __assert_fail ("ToBePromoted->getType() == Transition->getType() && \"The type of the result of the transition does not match \" \"the final type\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7325, __extension__ __PRETTY_FUNCTION__ )) | |||
7325 | "the final type")(static_cast <bool> (ToBePromoted->getType() == Transition ->getType() && "The type of the result of the transition does not match " "the final type") ? void (0) : __assert_fail ("ToBePromoted->getType() == Transition->getType() && \"The type of the result of the transition does not match \" \"the final type\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7325, __extension__ __PRETTY_FUNCTION__ )); | |||
7326 | ToBePromoted->replaceAllUsesWith(Transition); | |||
7327 | // 2. Update the type of the uses. | |||
7328 | // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def. | |||
7329 | Type *TransitionTy = getTransitionType(); | |||
7330 | ToBePromoted->mutateType(TransitionTy); | |||
7331 | // 3. Update all the operands of the promoted operation with promoted | |||
7332 | // operands. | |||
7333 | // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a. | |||
7334 | for (Use &U : ToBePromoted->operands()) { | |||
7335 | Value *Val = U.get(); | |||
7336 | Value *NewVal = nullptr; | |||
7337 | if (Val == Transition) | |||
7338 | NewVal = Transition->getOperand(getTransitionOriginalValueIdx()); | |||
7339 | else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) || | |||
7340 | isa<ConstantFP>(Val)) { | |||
7341 | // Use a splat constant if it is not safe to use undef. | |||
7342 | NewVal = getConstantVector( | |||
7343 | cast<Constant>(Val), | |||
7344 | isa<UndefValue>(Val) || | |||
7345 | canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo())); | |||
7346 | } else | |||
7347 | llvm_unreachable("Did you modified shouldPromote and forgot to update "::llvm::llvm_unreachable_internal("Did you modified shouldPromote and forgot to update " "this?", "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7348) | |||
7348 | "this?")::llvm::llvm_unreachable_internal("Did you modified shouldPromote and forgot to update " "this?", "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7348); | |||
7349 | ToBePromoted->setOperand(U.getOperandNo(), NewVal); | |||
7350 | } | |||
7351 | Transition->moveAfter(ToBePromoted); | |||
7352 | Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted); | |||
7353 | } | |||
7354 | ||||
7355 | /// Some targets can do store(extractelement) with one instruction. | |||
7356 | /// Try to push the extractelement towards the stores when the target | |||
7357 | /// has this feature and this is profitable. | |||
7358 | bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) { | |||
7359 | unsigned CombineCost = std::numeric_limits<unsigned>::max(); | |||
7360 | if (DisableStoreExtract || | |||
7361 | (!StressStoreExtract && | |||
7362 | !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(), | |||
7363 | Inst->getOperand(1), CombineCost))) | |||
7364 | return false; | |||
7365 | ||||
7366 | // At this point we know that Inst is a vector to scalar transition. | |||
7367 | // Try to move it down the def-use chain, until: | |||
7368 | // - We can combine the transition with its single use | |||
7369 | // => we got rid of the transition. | |||
7370 | // - We escape the current basic block | |||
7371 | // => we would need to check that we are moving it at a cheaper place and | |||
7372 | // we do not do that for now. | |||
7373 | BasicBlock *Parent = Inst->getParent(); | |||
7374 | LLVM_DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Found an interesting transition: " << *Inst << '\n'; } } while (false); | |||
7375 | VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost); | |||
7376 | // If the transition has more than one use, assume this is not going to be | |||
7377 | // beneficial. | |||
7378 | while (Inst->hasOneUse()) { | |||
7379 | Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin()); | |||
7380 | LLVM_DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Use: " << *ToBePromoted << '\n'; } } while (false); | |||
7381 | ||||
7382 | if (ToBePromoted->getParent() != Parent) { | |||
7383 | LLVM_DEBUG(dbgs() << "Instruction to promote is in a different block ("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Instruction to promote is in a different block (" << ToBePromoted->getParent()->getName() << ") than the transition (" << Parent->getName() << ").\n"; } } while (false) | |||
7384 | << ToBePromoted->getParent()->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Instruction to promote is in a different block (" << ToBePromoted->getParent()->getName() << ") than the transition (" << Parent->getName() << ").\n"; } } while (false) | |||
7385 | << ") than the transition (" << Parent->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Instruction to promote is in a different block (" << ToBePromoted->getParent()->getName() << ") than the transition (" << Parent->getName() << ").\n"; } } while (false) | |||
7386 | << ").\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Instruction to promote is in a different block (" << ToBePromoted->getParent()->getName() << ") than the transition (" << Parent->getName() << ").\n"; } } while (false); | |||
7387 | return false; | |||
7388 | } | |||
7389 | ||||
7390 | if (VPH.canCombine(ToBePromoted)) { | |||
7391 | LLVM_DEBUG(dbgs() << "Assume " << *Inst << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Assume " << *Inst << '\n' << "will be combined with: " << *ToBePromoted << '\n'; } } while (false) | |||
7392 | << "will be combined with: " << *ToBePromoted << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Assume " << *Inst << '\n' << "will be combined with: " << *ToBePromoted << '\n'; } } while (false); | |||
7393 | VPH.recordCombineInstruction(ToBePromoted); | |||
7394 | bool Changed = VPH.promote(); | |||
7395 | NumStoreExtractExposed += Changed; | |||
7396 | return Changed; | |||
7397 | } | |||
7398 | ||||
7399 | LLVM_DEBUG(dbgs() << "Try promoting.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Try promoting.\n"; } } while (false); | |||
7400 | if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted)) | |||
7401 | return false; | |||
7402 | ||||
7403 | LLVM_DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Promoting is possible... Enqueue for promotion!\n" ; } } while (false); | |||
7404 | ||||
7405 | VPH.enqueueForPromotion(ToBePromoted); | |||
7406 | Inst = ToBePromoted; | |||
7407 | } | |||
7408 | return false; | |||
7409 | } | |||
7410 | ||||
7411 | /// For the instruction sequence of store below, F and I values | |||
7412 | /// are bundled together as an i64 value before being stored into memory. | |||
7413 | /// Sometimes it is more efficient to generate separate stores for F and I, | |||
7414 | /// which can remove the bitwise instructions or sink them to colder places. | |||
7415 | /// | |||
7416 | /// (store (or (zext (bitcast F to i32) to i64), | |||
7417 | /// (shl (zext I to i64), 32)), addr) --> | |||
7418 | /// (store F, addr) and (store I, addr+4) | |||
7419 | /// | |||
7420 | /// Similarly, splitting for other merged store can also be beneficial, like: | |||
7421 | /// For pair of {i32, i32}, i64 store --> two i32 stores. | |||
7422 | /// For pair of {i32, i16}, i64 store --> two i32 stores. | |||
7423 | /// For pair of {i16, i16}, i32 store --> two i16 stores. | |||
7424 | /// For pair of {i16, i8}, i32 store --> two i16 stores. | |||
7425 | /// For pair of {i8, i8}, i16 store --> two i8 stores. | |||
7426 | /// | |||
7427 | /// We allow each target to determine specifically which kind of splitting is | |||
7428 | /// supported. | |||
7429 | /// | |||
7430 | /// The store patterns are commonly seen from the simple code snippet below | |||
7431 | /// if only std::make_pair(...) is sroa transformed before inlined into hoo. | |||
7432 | /// void goo(const std::pair<int, float> &); | |||
7433 | /// hoo() { | |||
7434 | /// ... | |||
7435 | /// goo(std::make_pair(tmp, ftmp)); | |||
7436 | /// ... | |||
7437 | /// } | |||
7438 | /// | |||
7439 | /// Although we already have similar splitting in DAG Combine, we duplicate | |||
7440 | /// it in CodeGenPrepare to catch the case in which pattern is across | |||
7441 | /// multiple BBs. The logic in DAG Combine is kept to catch case generated | |||
7442 | /// during code expansion. | |||
7443 | static bool splitMergedValStore(StoreInst &SI, const DataLayout &DL, | |||
7444 | const TargetLowering &TLI) { | |||
7445 | // Handle simple but common cases only. | |||
7446 | Type *StoreType = SI.getValueOperand()->getType(); | |||
7447 | ||||
7448 | // The code below assumes shifting a value by <number of bits>, | |||
7449 | // whereas scalable vectors would have to be shifted by | |||
7450 | // <2log(vscale) + number of bits> in order to store the | |||
7451 | // low/high parts. Bailing out for now. | |||
7452 | if (isa<ScalableVectorType>(StoreType)) | |||
7453 | return false; | |||
7454 | ||||
7455 | if (!DL.typeSizeEqualsStoreSize(StoreType) || | |||
7456 | DL.getTypeSizeInBits(StoreType) == 0) | |||
7457 | return false; | |||
7458 | ||||
7459 | unsigned HalfValBitSize = DL.getTypeSizeInBits(StoreType) / 2; | |||
7460 | Type *SplitStoreType = Type::getIntNTy(SI.getContext(), HalfValBitSize); | |||
7461 | if (!DL.typeSizeEqualsStoreSize(SplitStoreType)) | |||
7462 | return false; | |||
7463 | ||||
7464 | // Don't split the store if it is volatile. | |||
7465 | if (SI.isVolatile()) | |||
7466 | return false; | |||
7467 | ||||
7468 | // Match the following patterns: | |||
7469 | // (store (or (zext LValue to i64), | |||
7470 | // (shl (zext HValue to i64), 32)), HalfValBitSize) | |||
7471 | // or | |||
7472 | // (store (or (shl (zext HValue to i64), 32)), HalfValBitSize) | |||
7473 | // (zext LValue to i64), | |||
7474 | // Expect both operands of OR and the first operand of SHL have only | |||
7475 | // one use. | |||
7476 | Value *LValue, *HValue; | |||
7477 | if (!match(SI.getValueOperand(), | |||
7478 | m_c_Or(m_OneUse(m_ZExt(m_Value(LValue))), | |||
7479 | m_OneUse(m_Shl(m_OneUse(m_ZExt(m_Value(HValue))), | |||
7480 | m_SpecificInt(HalfValBitSize)))))) | |||
7481 | return false; | |||
7482 | ||||
7483 | // Check LValue and HValue are int with size less or equal than 32. | |||
7484 | if (!LValue->getType()->isIntegerTy() || | |||
7485 | DL.getTypeSizeInBits(LValue->getType()) > HalfValBitSize || | |||
7486 | !HValue->getType()->isIntegerTy() || | |||
7487 | DL.getTypeSizeInBits(HValue->getType()) > HalfValBitSize) | |||
7488 | return false; | |||
7489 | ||||
7490 | // If LValue/HValue is a bitcast instruction, use the EVT before bitcast | |||
7491 | // as the input of target query. | |||
7492 | auto *LBC = dyn_cast<BitCastInst>(LValue); | |||
7493 | auto *HBC = dyn_cast<BitCastInst>(HValue); | |||
7494 | EVT LowTy = LBC ? EVT::getEVT(LBC->getOperand(0)->getType()) | |||
7495 | : EVT::getEVT(LValue->getType()); | |||
7496 | EVT HighTy = HBC ? EVT::getEVT(HBC->getOperand(0)->getType()) | |||
7497 | : EVT::getEVT(HValue->getType()); | |||
7498 | if (!ForceSplitStore && !TLI.isMultiStoresCheaperThanBitsMerge(LowTy, HighTy)) | |||
7499 | return false; | |||
7500 | ||||
7501 | // Start to split store. | |||
7502 | IRBuilder<> Builder(SI.getContext()); | |||
7503 | Builder.SetInsertPoint(&SI); | |||
7504 | ||||
7505 | // If LValue/HValue is a bitcast in another BB, create a new one in current | |||
7506 | // BB so it may be merged with the splitted stores by dag combiner. | |||
7507 | if (LBC && LBC->getParent() != SI.getParent()) | |||
7508 | LValue = Builder.CreateBitCast(LBC->getOperand(0), LBC->getType()); | |||
7509 | if (HBC && HBC->getParent() != SI.getParent()) | |||
7510 | HValue = Builder.CreateBitCast(HBC->getOperand(0), HBC->getType()); | |||
7511 | ||||
7512 | bool IsLE = SI.getModule()->getDataLayout().isLittleEndian(); | |||
7513 | auto CreateSplitStore = [&](Value *V, bool Upper) { | |||
7514 | V = Builder.CreateZExtOrBitCast(V, SplitStoreType); | |||
7515 | Value *Addr = Builder.CreateBitCast( | |||
7516 | SI.getOperand(1), | |||
7517 | SplitStoreType->getPointerTo(SI.getPointerAddressSpace())); | |||
7518 | Align Alignment = SI.getAlign(); | |||
7519 | const bool IsOffsetStore = (IsLE && Upper) || (!IsLE && !Upper); | |||
7520 | if (IsOffsetStore) { | |||
7521 | Addr = Builder.CreateGEP( | |||
7522 | SplitStoreType, Addr, | |||
7523 | ConstantInt::get(Type::getInt32Ty(SI.getContext()), 1)); | |||
7524 | ||||
7525 | // When splitting the store in half, naturally one half will retain the | |||
7526 | // alignment of the original wider store, regardless of whether it was | |||
7527 | // over-aligned or not, while the other will require adjustment. | |||
7528 | Alignment = commonAlignment(Alignment, HalfValBitSize / 8); | |||
7529 | } | |||
7530 | Builder.CreateAlignedStore(V, Addr, Alignment); | |||
7531 | }; | |||
7532 | ||||
7533 | CreateSplitStore(LValue, false); | |||
7534 | CreateSplitStore(HValue, true); | |||
7535 | ||||
7536 | // Delete the old store. | |||
7537 | SI.eraseFromParent(); | |||
7538 | return true; | |||
7539 | } | |||
7540 | ||||
7541 | // Return true if the GEP has two operands, the first operand is of a sequential | |||
7542 | // type, and the second operand is a constant. | |||
7543 | static bool GEPSequentialConstIndexed(GetElementPtrInst *GEP) { | |||
7544 | gep_type_iterator I = gep_type_begin(*GEP); | |||
7545 | return GEP->getNumOperands() == 2 && | |||
7546 | I.isSequential() && | |||
7547 | isa<ConstantInt>(GEP->getOperand(1)); | |||
7548 | } | |||
7549 | ||||
7550 | // Try unmerging GEPs to reduce liveness interference (register pressure) across | |||
7551 | // IndirectBr edges. Since IndirectBr edges tend to touch on many blocks, | |||
7552 | // reducing liveness interference across those edges benefits global register | |||
7553 | // allocation. Currently handles only certain cases. | |||
7554 | // | |||
7555 | // For example, unmerge %GEPI and %UGEPI as below. | |||
7556 | // | |||
7557 | // ---------- BEFORE ---------- | |||
7558 | // SrcBlock: | |||
7559 | // ... | |||
7560 | // %GEPIOp = ... | |||
7561 | // ... | |||
7562 | // %GEPI = gep %GEPIOp, Idx | |||
7563 | // ... | |||
7564 | // indirectbr ... [ label %DstB0, label %DstB1, ... label %DstBi ... ] | |||
7565 | // (* %GEPI is alive on the indirectbr edges due to other uses ahead) | |||
7566 | // (* %GEPIOp is alive on the indirectbr edges only because of it's used by | |||
7567 | // %UGEPI) | |||
7568 | // | |||
7569 | // DstB0: ... (there may be a gep similar to %UGEPI to be unmerged) | |||
7570 | // DstB1: ... (there may be a gep similar to %UGEPI to be unmerged) | |||
7571 | // ... | |||
7572 | // | |||
7573 | // DstBi: | |||
7574 | // ... | |||
7575 | // %UGEPI = gep %GEPIOp, UIdx | |||
7576 | // ... | |||
7577 | // --------------------------- | |||
7578 | // | |||
7579 | // ---------- AFTER ---------- | |||
7580 | // SrcBlock: | |||
7581 | // ... (same as above) | |||
7582 | // (* %GEPI is still alive on the indirectbr edges) | |||
7583 | // (* %GEPIOp is no longer alive on the indirectbr edges as a result of the | |||
7584 | // unmerging) | |||
7585 | // ... | |||
7586 | // | |||
7587 | // DstBi: | |||
7588 | // ... | |||
7589 | // %UGEPI = gep %GEPI, (UIdx-Idx) | |||
7590 | // ... | |||
7591 | // --------------------------- | |||
7592 | // | |||
7593 | // The register pressure on the IndirectBr edges is reduced because %GEPIOp is | |||
7594 | // no longer alive on them. | |||
7595 | // | |||
7596 | // We try to unmerge GEPs here in CodGenPrepare, as opposed to limiting merging | |||
7597 | // of GEPs in the first place in InstCombiner::visitGetElementPtrInst() so as | |||
7598 | // not to disable further simplications and optimizations as a result of GEP | |||
7599 | // merging. | |||
7600 | // | |||
7601 | // Note this unmerging may increase the length of the data flow critical path | |||
7602 | // (the path from %GEPIOp to %UGEPI would go through %GEPI), which is a tradeoff | |||
7603 | // between the register pressure and the length of data-flow critical | |||
7604 | // path. Restricting this to the uncommon IndirectBr case would minimize the | |||
7605 | // impact of potentially longer critical path, if any, and the impact on compile | |||
7606 | // time. | |||
7607 | static bool tryUnmergingGEPsAcrossIndirectBr(GetElementPtrInst *GEPI, | |||
7608 | const TargetTransformInfo *TTI) { | |||
7609 | BasicBlock *SrcBlock = GEPI->getParent(); | |||
7610 | // Check that SrcBlock ends with an IndirectBr. If not, give up. The common | |||
7611 | // (non-IndirectBr) cases exit early here. | |||
7612 | if (!isa<IndirectBrInst>(SrcBlock->getTerminator())) | |||
7613 | return false; | |||
7614 | // Check that GEPI is a simple gep with a single constant index. | |||
7615 | if (!GEPSequentialConstIndexed(GEPI)) | |||
7616 | return false; | |||
7617 | ConstantInt *GEPIIdx = cast<ConstantInt>(GEPI->getOperand(1)); | |||
7618 | // Check that GEPI is a cheap one. | |||
7619 | if (TTI->getIntImmCost(GEPIIdx->getValue(), GEPIIdx->getType(), | |||
7620 | TargetTransformInfo::TCK_SizeAndLatency) | |||
7621 | > TargetTransformInfo::TCC_Basic) | |||
7622 | return false; | |||
7623 | Value *GEPIOp = GEPI->getOperand(0); | |||
7624 | // Check that GEPIOp is an instruction that's also defined in SrcBlock. | |||
7625 | if (!isa<Instruction>(GEPIOp)) | |||
7626 | return false; | |||
7627 | auto *GEPIOpI = cast<Instruction>(GEPIOp); | |||
7628 | if (GEPIOpI->getParent() != SrcBlock) | |||
7629 | return false; | |||
7630 | // Check that GEP is used outside the block, meaning it's alive on the | |||
7631 | // IndirectBr edge(s). | |||
7632 | if (find_if(GEPI->users(), [&](User *Usr) { | |||
7633 | if (auto *I = dyn_cast<Instruction>(Usr)) { | |||
7634 | if (I->getParent() != SrcBlock) { | |||
7635 | return true; | |||
7636 | } | |||
7637 | } | |||
7638 | return false; | |||
7639 | }) == GEPI->users().end()) | |||
7640 | return false; | |||
7641 | // The second elements of the GEP chains to be unmerged. | |||
7642 | std::vector<GetElementPtrInst *> UGEPIs; | |||
7643 | // Check each user of GEPIOp to check if unmerging would make GEPIOp not alive | |||
7644 | // on IndirectBr edges. | |||
7645 | for (User *Usr : GEPIOp->users()) { | |||
7646 | if (Usr == GEPI) continue; | |||
7647 | // Check if Usr is an Instruction. If not, give up. | |||
7648 | if (!isa<Instruction>(Usr)) | |||
7649 | return false; | |||
7650 | auto *UI = cast<Instruction>(Usr); | |||
7651 | // Check if Usr in the same block as GEPIOp, which is fine, skip. | |||
7652 | if (UI->getParent() == SrcBlock) | |||
7653 | continue; | |||
7654 | // Check if Usr is a GEP. If not, give up. | |||
7655 | if (!isa<GetElementPtrInst>(Usr)) | |||
7656 | return false; | |||
7657 | auto *UGEPI = cast<GetElementPtrInst>(Usr); | |||
7658 | // Check if UGEPI is a simple gep with a single constant index and GEPIOp is | |||
7659 | // the pointer operand to it. If so, record it in the vector. If not, give | |||
7660 | // up. | |||
7661 | if (!GEPSequentialConstIndexed(UGEPI)) | |||
7662 | return false; | |||
7663 | if (UGEPI->getOperand(0) != GEPIOp) | |||
7664 | return false; | |||
7665 | if (GEPIIdx->getType() != | |||
7666 | cast<ConstantInt>(UGEPI->getOperand(1))->getType()) | |||
7667 | return false; | |||
7668 | ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1)); | |||
7669 | if (TTI->getIntImmCost(UGEPIIdx->getValue(), UGEPIIdx->getType(), | |||
7670 | TargetTransformInfo::TCK_SizeAndLatency) | |||
7671 | > TargetTransformInfo::TCC_Basic) | |||
7672 | return false; | |||
7673 | UGEPIs.push_back(UGEPI); | |||
7674 | } | |||
7675 | if (UGEPIs.size() == 0) | |||
7676 | return false; | |||
7677 | // Check the materializing cost of (Uidx-Idx). | |||
7678 | for (GetElementPtrInst *UGEPI : UGEPIs) { | |||
7679 | ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1)); | |||
7680 | APInt NewIdx = UGEPIIdx->getValue() - GEPIIdx->getValue(); | |||
7681 | InstructionCost ImmCost = TTI->getIntImmCost( | |||
7682 | NewIdx, GEPIIdx->getType(), TargetTransformInfo::TCK_SizeAndLatency); | |||
7683 | if (ImmCost > TargetTransformInfo::TCC_Basic) | |||
7684 | return false; | |||
7685 | } | |||
7686 | // Now unmerge between GEPI and UGEPIs. | |||
7687 | for (GetElementPtrInst *UGEPI : UGEPIs) { | |||
7688 | UGEPI->setOperand(0, GEPI); | |||
7689 | ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1)); | |||
7690 | Constant *NewUGEPIIdx = | |||
7691 | ConstantInt::get(GEPIIdx->getType(), | |||
7692 | UGEPIIdx->getValue() - GEPIIdx->getValue()); | |||
7693 | UGEPI->setOperand(1, NewUGEPIIdx); | |||
7694 | // If GEPI is not inbounds but UGEPI is inbounds, change UGEPI to not | |||
7695 | // inbounds to avoid UB. | |||
7696 | if (!GEPI->isInBounds()) { | |||
7697 | UGEPI->setIsInBounds(false); | |||
7698 | } | |||
7699 | } | |||
7700 | // After unmerging, verify that GEPIOp is actually only used in SrcBlock (not | |||
7701 | // alive on IndirectBr edges). | |||
7702 | assert(find_if(GEPIOp->users(), [&](User *Usr) {(static_cast <bool> (find_if(GEPIOp->users(), [& ](User *Usr) { return cast<Instruction>(Usr)->getParent () != SrcBlock; }) == GEPIOp->users().end() && "GEPIOp is used outside SrcBlock" ) ? void (0) : __assert_fail ("find_if(GEPIOp->users(), [&](User *Usr) { return cast<Instruction>(Usr)->getParent() != SrcBlock; }) == GEPIOp->users().end() && \"GEPIOp is used outside SrcBlock\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7704, __extension__ __PRETTY_FUNCTION__ )) | |||
7703 | return cast<Instruction>(Usr)->getParent() != SrcBlock;(static_cast <bool> (find_if(GEPIOp->users(), [& ](User *Usr) { return cast<Instruction>(Usr)->getParent () != SrcBlock; }) == GEPIOp->users().end() && "GEPIOp is used outside SrcBlock" ) ? void (0) : __assert_fail ("find_if(GEPIOp->users(), [&](User *Usr) { return cast<Instruction>(Usr)->getParent() != SrcBlock; }) == GEPIOp->users().end() && \"GEPIOp is used outside SrcBlock\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7704, __extension__ __PRETTY_FUNCTION__ )) | |||
7704 | }) == GEPIOp->users().end() && "GEPIOp is used outside SrcBlock")(static_cast <bool> (find_if(GEPIOp->users(), [& ](User *Usr) { return cast<Instruction>(Usr)->getParent () != SrcBlock; }) == GEPIOp->users().end() && "GEPIOp is used outside SrcBlock" ) ? void (0) : __assert_fail ("find_if(GEPIOp->users(), [&](User *Usr) { return cast<Instruction>(Usr)->getParent() != SrcBlock; }) == GEPIOp->users().end() && \"GEPIOp is used outside SrcBlock\"" , "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7704, __extension__ __PRETTY_FUNCTION__ )); | |||
7705 | return true; | |||
7706 | } | |||
7707 | ||||
7708 | static bool optimizeBranch(BranchInst *Branch, const TargetLowering &TLI) { | |||
7709 | // Try and convert | |||
7710 | // %c = icmp ult %x, 8 | |||
7711 | // br %c, bla, blb | |||
7712 | // %tc = lshr %x, 3 | |||
7713 | // to | |||
7714 | // %tc = lshr %x, 3 | |||
7715 | // %c = icmp eq %tc, 0 | |||
7716 | // br %c, bla, blb | |||
7717 | // Creating the cmp to zero can be better for the backend, especially if the | |||
7718 | // lshr produces flags that can be used automatically. | |||
7719 | if (!TLI.preferZeroCompareBranch() || !Branch->isConditional()) | |||
7720 | return false; | |||
7721 | ||||
7722 | ICmpInst *Cmp = dyn_cast<ICmpInst>(Branch->getCondition()); | |||
7723 | if (!Cmp || !isa<ConstantInt>(Cmp->getOperand(1)) || !Cmp->hasOneUse()) | |||
7724 | return false; | |||
7725 | ||||
7726 | Value *X = Cmp->getOperand(0); | |||
7727 | APInt CmpC = cast<ConstantInt>(Cmp->getOperand(1))->getValue(); | |||
7728 | ||||
7729 | for (auto *U : X->users()) { | |||
7730 | Instruction *UI = dyn_cast<Instruction>(U); | |||
7731 | // A quick dominance check | |||
7732 | if (!UI || | |||
7733 | (UI->getParent() != Branch->getParent() && | |||
7734 | UI->getParent() != Branch->getSuccessor(0) && | |||
7735 | UI->getParent() != Branch->getSuccessor(1)) || | |||
7736 | (UI->getParent() != Branch->getParent() && | |||
7737 | !UI->getParent()->getSinglePredecessor())) | |||
7738 | continue; | |||
7739 | ||||
7740 | if (CmpC.isPowerOf2() && Cmp->getPredicate() == ICmpInst::ICMP_ULT && | |||
7741 | match(UI, m_Shr(m_Specific(X), m_SpecificInt(CmpC.logBase2())))) { | |||
7742 | IRBuilder<> Builder(Branch); | |||
7743 | if (UI->getParent() != Branch->getParent()) | |||
7744 | UI->moveBefore(Branch); | |||
7745 | Value *NewCmp = Builder.CreateCmp(ICmpInst::ICMP_EQ, UI, | |||
7746 | ConstantInt::get(UI->getType(), 0)); | |||
7747 | LLVM_DEBUG(dbgs() << "Converting " << *Cmp << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Converting " << * Cmp << "\n"; } } while (false); | |||
7748 | LLVM_DEBUG(dbgs() << " to compare on zero: " << *NewCmp << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << " to compare on zero: " << *NewCmp << "\n"; } } while (false); | |||
7749 | Cmp->replaceAllUsesWith(NewCmp); | |||
7750 | return true; | |||
7751 | } | |||
7752 | if (Cmp->isEquality() && | |||
7753 | (match(UI, m_Add(m_Specific(X), m_SpecificInt(-CmpC))) || | |||
7754 | match(UI, m_Sub(m_Specific(X), m_SpecificInt(CmpC))))) { | |||
7755 | IRBuilder<> Builder(Branch); | |||
7756 | if (UI->getParent() != Branch->getParent()) | |||
7757 | UI->moveBefore(Branch); | |||
7758 | Value *NewCmp = Builder.CreateCmp(Cmp->getPredicate(), UI, | |||
7759 | ConstantInt::get(UI->getType(), 0)); | |||
7760 | LLVM_DEBUG(dbgs() << "Converting " << *Cmp << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Converting " << * Cmp << "\n"; } } while (false); | |||
7761 | LLVM_DEBUG(dbgs() << " to compare on zero: " << *NewCmp << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << " to compare on zero: " << *NewCmp << "\n"; } } while (false); | |||
7762 | Cmp->replaceAllUsesWith(NewCmp); | |||
7763 | return true; | |||
7764 | } | |||
7765 | } | |||
7766 | return false; | |||
7767 | } | |||
7768 | ||||
7769 | bool CodeGenPrepare::optimizeInst(Instruction *I, bool &ModifiedDT) { | |||
7770 | // Bail out if we inserted the instruction to prevent optimizations from | |||
7771 | // stepping on each other's toes. | |||
7772 | if (InsertedInsts.count(I)) | |||
7773 | return false; | |||
7774 | ||||
7775 | // TODO: Move into the switch on opcode below here. | |||
7776 | if (PHINode *P = dyn_cast<PHINode>(I)) { | |||
7777 | // It is possible for very late stage optimizations (such as SimplifyCFG) | |||
7778 | // to introduce PHI nodes too late to be cleaned up. If we detect such a | |||
7779 | // trivial PHI, go ahead and zap it here. | |||
7780 | if (Value *V = SimplifyInstruction(P, {*DL, TLInfo})) { | |||
7781 | LargeOffsetGEPMap.erase(P); | |||
7782 | P->replaceAllUsesWith(V); | |||
7783 | P->eraseFromParent(); | |||
7784 | ++NumPHIsElim; | |||
7785 | return true; | |||
7786 | } | |||
7787 | return false; | |||
7788 | } | |||
7789 | ||||
7790 | if (CastInst *CI = dyn_cast<CastInst>(I)) { | |||
7791 | // If the source of the cast is a constant, then this should have | |||
7792 | // already been constant folded. The only reason NOT to constant fold | |||
7793 | // it is if something (e.g. LSR) was careful to place the constant | |||
7794 | // evaluation in a block other than then one that uses it (e.g. to hoist | |||
7795 | // the address of globals out of a loop). If this is the case, we don't | |||
7796 | // want to forward-subst the cast. | |||
7797 | if (isa<Constant>(CI->getOperand(0))) | |||
7798 | return false; | |||
7799 | ||||
7800 | if (OptimizeNoopCopyExpression(CI, *TLI, *DL)) | |||
7801 | return true; | |||
7802 | ||||
7803 | if (isa<ZExtInst>(I) || isa<SExtInst>(I)) { | |||
7804 | /// Sink a zext or sext into its user blocks if the target type doesn't | |||
7805 | /// fit in one register | |||
7806 | if (TLI->getTypeAction(CI->getContext(), | |||
7807 | TLI->getValueType(*DL, CI->getType())) == | |||
7808 | TargetLowering::TypeExpandInteger) { | |||
7809 | return SinkCast(CI); | |||
7810 | } else { | |||
7811 | bool MadeChange = optimizeExt(I); | |||
7812 | return MadeChange | optimizeExtUses(I); | |||
7813 | } | |||
7814 | } | |||
7815 | return false; | |||
7816 | } | |||
7817 | ||||
7818 | if (auto *Cmp = dyn_cast<CmpInst>(I)) | |||
7819 | if (optimizeCmp(Cmp, ModifiedDT)) | |||
7820 | return true; | |||
7821 | ||||
7822 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) { | |||
7823 | LI->setMetadata(LLVMContext::MD_invariant_group, nullptr); | |||
7824 | bool Modified = optimizeLoadExt(LI); | |||
7825 | unsigned AS = LI->getPointerAddressSpace(); | |||
7826 | Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS); | |||
7827 | return Modified; | |||
7828 | } | |||
7829 | ||||
7830 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) { | |||
7831 | if (splitMergedValStore(*SI, *DL, *TLI)) | |||
7832 | return true; | |||
7833 | SI->setMetadata(LLVMContext::MD_invariant_group, nullptr); | |||
7834 | unsigned AS = SI->getPointerAddressSpace(); | |||
7835 | return optimizeMemoryInst(I, SI->getOperand(1), | |||
7836 | SI->getOperand(0)->getType(), AS); | |||
7837 | } | |||
7838 | ||||
7839 | if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) { | |||
7840 | unsigned AS = RMW->getPointerAddressSpace(); | |||
7841 | return optimizeMemoryInst(I, RMW->getPointerOperand(), | |||
7842 | RMW->getType(), AS); | |||
7843 | } | |||
7844 | ||||
7845 | if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(I)) { | |||
7846 | unsigned AS = CmpX->getPointerAddressSpace(); | |||
7847 | return optimizeMemoryInst(I, CmpX->getPointerOperand(), | |||
7848 | CmpX->getCompareOperand()->getType(), AS); | |||
7849 | } | |||
7850 | ||||
7851 | BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I); | |||
7852 | ||||
7853 | if (BinOp && BinOp->getOpcode() == Instruction::And && EnableAndCmpSinking && | |||
7854 | sinkAndCmp0Expression(BinOp, *TLI, InsertedInsts)) | |||
7855 | return true; | |||
7856 | ||||
7857 | // TODO: Move this into the switch on opcode - it handles shifts already. | |||
7858 | if (BinOp && (BinOp->getOpcode() == Instruction::AShr || | |||
7859 | BinOp->getOpcode() == Instruction::LShr)) { | |||
7860 | ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1)); | |||
7861 | if (CI && TLI->hasExtractBitsInsn()) | |||
7862 | if (OptimizeExtractBits(BinOp, CI, *TLI, *DL)) | |||
7863 | return true; | |||
7864 | } | |||
7865 | ||||
7866 | if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { | |||
7867 | if (GEPI->hasAllZeroIndices()) { | |||
7868 | /// The GEP operand must be a pointer, so must its result -> BitCast | |||
7869 | Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), | |||
7870 | GEPI->getName(), GEPI); | |||
7871 | NC->setDebugLoc(GEPI->getDebugLoc()); | |||
7872 | GEPI->replaceAllUsesWith(NC); | |||
7873 | GEPI->eraseFromParent(); | |||
7874 | ++NumGEPsElim; | |||
7875 | optimizeInst(NC, ModifiedDT); | |||
7876 | return true; | |||
7877 | } | |||
7878 | if (tryUnmergingGEPsAcrossIndirectBr(GEPI, TTI)) { | |||
7879 | return true; | |||
7880 | } | |||
7881 | return false; | |||
7882 | } | |||
7883 | ||||
7884 | if (FreezeInst *FI = dyn_cast<FreezeInst>(I)) { | |||
7885 | // freeze(icmp a, const)) -> icmp (freeze a), const | |||
7886 | // This helps generate efficient conditional jumps. | |||
7887 | Instruction *CmpI = nullptr; | |||
7888 | if (ICmpInst *II = dyn_cast<ICmpInst>(FI->getOperand(0))) | |||
7889 | CmpI = II; | |||
7890 | else if (FCmpInst *F = dyn_cast<FCmpInst>(FI->getOperand(0))) | |||
7891 | CmpI = F->getFastMathFlags().none() ? F : nullptr; | |||
7892 | ||||
7893 | if (CmpI && CmpI->hasOneUse()) { | |||
7894 | auto Op0 = CmpI->getOperand(0), Op1 = CmpI->getOperand(1); | |||
7895 | bool Const0 = isa<ConstantInt>(Op0) || isa<ConstantFP>(Op0) || | |||
7896 | isa<ConstantPointerNull>(Op0); | |||
7897 | bool Const1 = isa<ConstantInt>(Op1) || isa<ConstantFP>(Op1) || | |||
7898 | isa<ConstantPointerNull>(Op1); | |||
7899 | if (Const0 || Const1) { | |||
7900 | if (!Const0 || !Const1) { | |||
7901 | auto *F = new FreezeInst(Const0 ? Op1 : Op0, "", CmpI); | |||
7902 | F->takeName(FI); | |||
7903 | CmpI->setOperand(Const0 ? 1 : 0, F); | |||
7904 | } | |||
7905 | FI->replaceAllUsesWith(CmpI); | |||
7906 | FI->eraseFromParent(); | |||
7907 | return true; | |||
7908 | } | |||
7909 | } | |||
7910 | return false; | |||
7911 | } | |||
7912 | ||||
7913 | if (tryToSinkFreeOperands(I)) | |||
7914 | return true; | |||
7915 | ||||
7916 | switch (I->getOpcode()) { | |||
7917 | case Instruction::Shl: | |||
7918 | case Instruction::LShr: | |||
7919 | case Instruction::AShr: | |||
7920 | return optimizeShiftInst(cast<BinaryOperator>(I)); | |||
7921 | case Instruction::Call: | |||
7922 | return optimizeCallInst(cast<CallInst>(I), ModifiedDT); | |||
7923 | case Instruction::Select: | |||
7924 | return optimizeSelectInst(cast<SelectInst>(I)); | |||
7925 | case Instruction::ShuffleVector: | |||
7926 | return optimizeShuffleVectorInst(cast<ShuffleVectorInst>(I)); | |||
7927 | case Instruction::Switch: | |||
7928 | return optimizeSwitchInst(cast<SwitchInst>(I)); | |||
7929 | case Instruction::ExtractElement: | |||
7930 | return optimizeExtractElementInst(cast<ExtractElementInst>(I)); | |||
7931 | case Instruction::Br: | |||
7932 | return optimizeBranch(cast<BranchInst>(I), *TLI); | |||
7933 | } | |||
7934 | ||||
7935 | return false; | |||
7936 | } | |||
7937 | ||||
7938 | /// Given an OR instruction, check to see if this is a bitreverse | |||
7939 | /// idiom. If so, insert the new intrinsic and return true. | |||
7940 | bool CodeGenPrepare::makeBitReverse(Instruction &I) { | |||
7941 | if (!I.getType()->isIntegerTy() || | |||
7942 | !TLI->isOperationLegalOrCustom(ISD::BITREVERSE, | |||
7943 | TLI->getValueType(*DL, I.getType(), true))) | |||
7944 | return false; | |||
7945 | ||||
7946 | SmallVector<Instruction*, 4> Insts; | |||
7947 | if (!recognizeBSwapOrBitReverseIdiom(&I, false, true, Insts)) | |||
7948 | return false; | |||
7949 | Instruction *LastInst = Insts.back(); | |||
7950 | I.replaceAllUsesWith(LastInst); | |||
7951 | RecursivelyDeleteTriviallyDeadInstructions( | |||
7952 | &I, TLInfo, nullptr, [&](Value *V) { removeAllAssertingVHReferences(V); }); | |||
7953 | return true; | |||
7954 | } | |||
7955 | ||||
7956 | // In this pass we look for GEP and cast instructions that are used | |||
7957 | // across basic blocks and rewrite them to improve basic-block-at-a-time | |||
7958 | // selection. | |||
7959 | bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool &ModifiedDT) { | |||
7960 | SunkAddrs.clear(); | |||
7961 | bool MadeChange = false; | |||
7962 | ||||
7963 | CurInstIterator = BB.begin(); | |||
7964 | while (CurInstIterator != BB.end()) { | |||
7965 | MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT); | |||
7966 | if (ModifiedDT) | |||
7967 | return true; | |||
7968 | } | |||
7969 | ||||
7970 | bool MadeBitReverse = true; | |||
7971 | while (MadeBitReverse) { | |||
7972 | MadeBitReverse = false; | |||
7973 | for (auto &I : reverse(BB)) { | |||
7974 | if (makeBitReverse(I)) { | |||
7975 | MadeBitReverse = MadeChange = true; | |||
7976 | break; | |||
7977 | } | |||
7978 | } | |||
7979 | } | |||
7980 | MadeChange |= dupRetToEnableTailCallOpts(&BB, ModifiedDT); | |||
7981 | ||||
7982 | return MadeChange; | |||
7983 | } | |||
7984 | ||||
7985 | // Some CGP optimizations may move or alter what's computed in a block. Check | |||
7986 | // whether a dbg.value intrinsic could be pointed at a more appropriate operand. | |||
7987 | bool CodeGenPrepare::fixupDbgValue(Instruction *I) { | |||
7988 | assert(isa<DbgValueInst>(I))(static_cast <bool> (isa<DbgValueInst>(I)) ? void (0) : __assert_fail ("isa<DbgValueInst>(I)", "llvm/lib/CodeGen/CodeGenPrepare.cpp" , 7988, __extension__ __PRETTY_FUNCTION__)); | |||
7989 | DbgValueInst &DVI = *cast<DbgValueInst>(I); | |||
7990 | ||||
7991 | // Does this dbg.value refer to a sunk address calculation? | |||
7992 | bool AnyChange = false; | |||
7993 | SmallDenseSet<Value *> LocationOps(DVI.location_ops().begin(), | |||
7994 | DVI.location_ops().end()); | |||
7995 | for (Value *Location : LocationOps) { | |||
7996 | WeakTrackingVH SunkAddrVH = SunkAddrs[Location]; | |||
7997 | Value *SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr; | |||
7998 | if (SunkAddr) { | |||
7999 | // Point dbg.value at locally computed address, which should give the best | |||
8000 | // opportunity to be accurately lowered. This update may change the type | |||
8001 | // of pointer being referred to; however this makes no difference to | |||
8002 | // debugging information, and we can't generate bitcasts that may affect | |||
8003 | // codegen. | |||
8004 | DVI.replaceVariableLocationOp(Location, SunkAddr); | |||
8005 | AnyChange = true; | |||
8006 | } | |||
8007 | } | |||
8008 | return AnyChange; | |||
8009 | } | |||
8010 | ||||
8011 | // A llvm.dbg.value may be using a value before its definition, due to | |||
8012 | // optimizations in this pass and others. Scan for such dbg.values, and rescue | |||
8013 | // them by moving the dbg.value to immediately after the value definition. | |||
8014 | // FIXME: Ideally this should never be necessary, and this has the potential | |||
8015 | // to re-order dbg.value intrinsics. | |||
8016 | bool CodeGenPrepare::placeDbgValues(Function &F) { | |||
8017 | bool MadeChange = false; | |||
8018 | DominatorTree DT(F); | |||
8019 | ||||
8020 | for (BasicBlock &BB : F) { | |||
8021 | for (Instruction &Insn : llvm::make_early_inc_range(BB)) { | |||
8022 | DbgValueInst *DVI = dyn_cast<DbgValueInst>(&Insn); | |||
8023 | if (!DVI) | |||
8024 | continue; | |||
8025 | ||||
8026 | SmallVector<Instruction *, 4> VIs; | |||
8027 | for (Value *V : DVI->getValues()) | |||
8028 | if (Instruction *VI = dyn_cast_or_null<Instruction>(V)) | |||
8029 | VIs.push_back(VI); | |||
8030 | ||||
8031 | // This DVI may depend on multiple instructions, complicating any | |||
8032 | // potential sink. This block takes the defensive approach, opting to | |||
8033 | // "undef" the DVI if it has more than one instruction and any of them do | |||
8034 | // not dominate DVI. | |||
8035 | for (Instruction *VI : VIs) { | |||
8036 | if (VI->isTerminator()) | |||
8037 | continue; | |||
8038 | ||||
8039 | // If VI is a phi in a block with an EHPad terminator, we can't insert | |||
8040 | // after it. | |||
8041 | if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad()) | |||
8042 | continue; | |||
8043 | ||||
8044 | // If the defining instruction dominates the dbg.value, we do not need | |||
8045 | // to move the dbg.value. | |||
8046 | if (DT.dominates(VI, DVI)) | |||
8047 | continue; | |||
8048 | ||||
8049 | // If we depend on multiple instructions and any of them doesn't | |||
8050 | // dominate this DVI, we probably can't salvage it: moving it to | |||
8051 | // after any of the instructions could cause us to lose the others. | |||
8052 | if (VIs.size() > 1) { | |||
8053 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Unable to find valid location for Debug Value, undefing:\n" << *DVI; } } while (false) | |||
8054 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Unable to find valid location for Debug Value, undefing:\n" << *DVI; } } while (false) | |||
8055 | << "Unable to find valid location for Debug Value, undefing:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Unable to find valid location for Debug Value, undefing:\n" << *DVI; } } while (false) | |||
8056 | << *DVI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Unable to find valid location for Debug Value, undefing:\n" << *DVI; } } while (false); | |||
8057 | DVI->setUndef(); | |||
8058 | break; | |||
8059 | } | |||
8060 | ||||
8061 | LLVM_DEBUG(dbgs() << "Moving Debug Value before :\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI; } } while (false) | |||
8062 | << *DVI << ' ' << *VI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI; } } while (false); | |||
8063 | DVI->removeFromParent(); | |||
8064 | if (isa<PHINode>(VI)) | |||
8065 | DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt()); | |||
8066 | else | |||
8067 | DVI->insertAfter(VI); | |||
8068 | MadeChange = true; | |||
8069 | ++NumDbgValueMoved; | |||
8070 | } | |||
8071 | } | |||
8072 | } | |||
8073 | return MadeChange; | |||
8074 | } | |||
8075 | ||||
8076 | // Group scattered pseudo probes in a block to favor SelectionDAG. Scattered | |||
8077 | // probes can be chained dependencies of other regular DAG nodes and block DAG | |||
8078 | // combine optimizations. | |||
8079 | bool CodeGenPrepare::placePseudoProbes(Function &F) { | |||
8080 | bool MadeChange = false; | |||
8081 | for (auto &Block : F) { | |||
8082 | // Move the rest probes to the beginning of the block. | |||
8083 | auto FirstInst = Block.getFirstInsertionPt(); | |||
8084 | while (FirstInst != Block.end() && FirstInst->isDebugOrPseudoInst()) | |||
8085 | ++FirstInst; | |||
8086 | BasicBlock::iterator I(FirstInst); | |||
8087 | I++; | |||
8088 | while (I != Block.end()) { | |||
8089 | if (auto *II = dyn_cast<PseudoProbeInst>(I++)) { | |||
8090 | II->moveBefore(&*FirstInst); | |||
8091 | MadeChange = true; | |||
8092 | } | |||
8093 | } | |||
8094 | } | |||
8095 | return MadeChange; | |||
8096 | } | |||
8097 | ||||
8098 | /// Scale down both weights to fit into uint32_t. | |||
8099 | static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) { | |||
8100 | uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse; | |||
8101 | uint32_t Scale = (NewMax / std::numeric_limits<uint32_t>::max()) + 1; | |||
8102 | NewTrue = NewTrue / Scale; | |||
8103 | NewFalse = NewFalse / Scale; | |||
8104 | } | |||
8105 | ||||
8106 | /// Some targets prefer to split a conditional branch like: | |||
8107 | /// \code | |||
8108 | /// %0 = icmp ne i32 %a, 0 | |||
8109 | /// %1 = icmp ne i32 %b, 0 | |||
8110 | /// %or.cond = or i1 %0, %1 | |||
8111 | /// br i1 %or.cond, label %TrueBB, label %FalseBB | |||
8112 | /// \endcode | |||
8113 | /// into multiple branch instructions like: | |||
8114 | /// \code | |||
8115 | /// bb1: | |||
8116 | /// %0 = icmp ne i32 %a, 0 | |||
8117 | /// br i1 %0, label %TrueBB, label %bb2 | |||
8118 | /// bb2: | |||
8119 | /// %1 = icmp ne i32 %b, 0 | |||
8120 | /// br i1 %1, label %TrueBB, label %FalseBB | |||
8121 | /// \endcode | |||
8122 | /// This usually allows instruction selection to do even further optimizations | |||
8123 | /// and combine the compare with the branch instruction. Currently this is | |||
8124 | /// applied for targets which have "cheap" jump instructions. | |||
8125 | /// | |||
8126 | /// FIXME: Remove the (equivalent?) implementation in SelectionDAG. | |||
8127 | /// | |||
8128 | bool CodeGenPrepare::splitBranchCondition(Function &F, bool &ModifiedDT) { | |||
8129 | if (!TM->Options.EnableFastISel || TLI->isJumpExpensive()) | |||
8130 | return false; | |||
8131 | ||||
8132 | bool MadeChange = false; | |||
8133 | for (auto &BB : F) { | |||
8134 | // Does this BB end with the following? | |||
8135 | // %cond1 = icmp|fcmp|binary instruction ... | |||
8136 | // %cond2 = icmp|fcmp|binary instruction ... | |||
8137 | // %cond.or = or|and i1 %cond1, cond2 | |||
8138 | // br i1 %cond.or label %dest1, label %dest2" | |||
8139 | Instruction *LogicOp; | |||
8140 | BasicBlock *TBB, *FBB; | |||
8141 | if (!match(BB.getTerminator(), | |||
8142 | m_Br(m_OneUse(m_Instruction(LogicOp)), TBB, FBB))) | |||
8143 | continue; | |||
8144 | ||||
8145 | auto *Br1 = cast<BranchInst>(BB.getTerminator()); | |||
8146 | if (Br1->getMetadata(LLVMContext::MD_unpredictable)) | |||
8147 | continue; | |||
8148 | ||||
8149 | // The merging of mostly empty BB can cause a degenerate branch. | |||
8150 | if (TBB == FBB) | |||
8151 | continue; | |||
8152 | ||||
8153 | unsigned Opc; | |||
8154 | Value *Cond1, *Cond2; | |||
8155 | if (match(LogicOp, | |||
8156 | m_LogicalAnd(m_OneUse(m_Value(Cond1)), m_OneUse(m_Value(Cond2))))) | |||
8157 | Opc = Instruction::And; | |||
8158 | else if (match(LogicOp, m_LogicalOr(m_OneUse(m_Value(Cond1)), | |||
8159 | m_OneUse(m_Value(Cond2))))) | |||
8160 | Opc = Instruction::Or; | |||
8161 | else | |||
8162 | continue; | |||
8163 | ||||
8164 | auto IsGoodCond = [](Value *Cond) { | |||
8165 | return match( | |||
8166 | Cond, | |||
8167 | m_CombineOr(m_Cmp(), m_CombineOr(m_LogicalAnd(m_Value(), m_Value()), | |||
8168 | m_LogicalOr(m_Value(), m_Value())))); | |||
8169 | }; | |||
8170 | if (!IsGoodCond(Cond1) || !IsGoodCond(Cond2)) | |||
8171 | continue; | |||
8172 | ||||
8173 | LLVM_DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Before branch condition splitting\n" ; BB.dump(); } } while (false); | |||
8174 | ||||
8175 | // Create a new BB. | |||
8176 | auto *TmpBB = | |||
8177 | BasicBlock::Create(BB.getContext(), BB.getName() + ".cond.split", | |||
8178 | BB.getParent(), BB.getNextNode()); | |||
8179 | ||||
8180 | // Update original basic block by using the first condition directly by the | |||
8181 | // branch instruction and removing the no longer needed and/or instruction. | |||
8182 | Br1->setCondition(Cond1); | |||
8183 | LogicOp->eraseFromParent(); | |||
8184 | ||||
8185 | // Depending on the condition we have to either replace the true or the | |||
8186 | // false successor of the original branch instruction. | |||
8187 | if (Opc == Instruction::And) | |||
8188 | Br1->setSuccessor(0, TmpBB); | |||
8189 | else | |||
8190 | Br1->setSuccessor(1, TmpBB); | |||
8191 | ||||
8192 | // Fill in the new basic block. | |||
8193 | auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB); | |||
8194 | if (auto *I = dyn_cast<Instruction>(Cond2)) { | |||
8195 | I->removeFromParent(); | |||
8196 | I->insertBefore(Br2); | |||
8197 | } | |||
8198 | ||||
8199 | // Update PHI nodes in both successors. The original BB needs to be | |||
8200 | // replaced in one successor's PHI nodes, because the branch comes now from | |||
8201 | // the newly generated BB (NewBB). In the other successor we need to add one | |||
8202 | // incoming edge to the PHI nodes, because both branch instructions target | |||
8203 | // now the same successor. Depending on the original branch condition | |||
8204 | // (and/or) we have to swap the successors (TrueDest, FalseDest), so that | |||
8205 | // we perform the correct update for the PHI nodes. | |||
8206 | // This doesn't change the successor order of the just created branch | |||
8207 | // instruction (or any other instruction). | |||
8208 | if (Opc == Instruction::Or) | |||
8209 | std::swap(TBB, FBB); | |||
8210 | ||||
8211 | // Replace the old BB with the new BB. | |||
8212 | TBB->replacePhiUsesWith(&BB, TmpBB); | |||
8213 | ||||
8214 | // Add another incoming edge form the new BB. | |||
8215 | for (PHINode &PN : FBB->phis()) { | |||
8216 | auto *Val = PN.getIncomingValueForBlock(&BB); | |||
8217 | PN.addIncoming(Val, TmpBB); | |||
8218 | } | |||
8219 | ||||
8220 | // Update the branch weights (from SelectionDAGBuilder:: | |||
8221 | // FindMergedConditions). | |||
8222 | if (Opc == Instruction::Or) { | |||
8223 | // Codegen X | Y as: | |||
8224 | // BB1: | |||
8225 | // jmp_if_X TBB | |||
8226 | // jmp TmpBB | |||
8227 | // TmpBB: | |||
8228 | // jmp_if_Y TBB | |||
8229 | // jmp FBB | |||
8230 | // | |||
8231 | ||||
8232 | // We have flexibility in setting Prob for BB1 and Prob for NewBB. | |||
8233 | // The requirement is that | |||
8234 | // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB) | |||
8235 | // = TrueProb for original BB. | |||
8236 | // Assuming the original weights are A and B, one choice is to set BB1's | |||
8237 | // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice | |||
8238 | // assumes that | |||
8239 | // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB. | |||
8240 | // Another choice is to assume TrueProb for BB1 equals to TrueProb for | |||
8241 | // TmpBB, but the math is more complicated. | |||
8242 | uint64_t TrueWeight, FalseWeight; | |||
8243 | if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) { | |||
8244 | uint64_t NewTrueWeight = TrueWeight; | |||
8245 | uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight; | |||
8246 | scaleWeights(NewTrueWeight, NewFalseWeight); | |||
8247 | Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext()) | |||
8248 | .createBranchWeights(TrueWeight, FalseWeight)); | |||
8249 | ||||
8250 | NewTrueWeight = TrueWeight; | |||
8251 | NewFalseWeight = 2 * FalseWeight; | |||
8252 | scaleWeights(NewTrueWeight, NewFalseWeight); | |||
8253 | Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext()) | |||
8254 | .createBranchWeights(TrueWeight, FalseWeight)); | |||
8255 | } | |||
8256 | } else { | |||
8257 | // Codegen X & Y as: | |||
8258 | // BB1: | |||
8259 | // jmp_if_X TmpBB | |||
8260 | // jmp FBB | |||
8261 | // TmpBB: | |||
8262 | // jmp_if_Y TBB | |||
8263 | // jmp FBB | |||
8264 | // | |||
8265 | // This requires creation of TmpBB after CurBB. | |||
8266 | ||||
8267 | // We have flexibility in setting Prob for BB1 and Prob for TmpBB. | |||
8268 | // The requirement is that | |||
8269 | // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB) | |||
8270 | // = FalseProb for original BB. | |||
8271 | // Assuming the original weights are A and B, one choice is to set BB1's | |||
8272 | // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice | |||
8273 | // assumes that | |||
8274 | // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB. | |||
8275 | uint64_t TrueWeight, FalseWeight; | |||
8276 | if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) { | |||
8277 | uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight; | |||
8278 | uint64_t NewFalseWeight = FalseWeight; | |||
8279 | scaleWeights(NewTrueWeight, NewFalseWeight); | |||
8280 | Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext()) | |||
8281 | .createBranchWeights(TrueWeight, FalseWeight)); | |||
8282 | ||||
8283 | NewTrueWeight = 2 * TrueWeight; | |||
8284 | NewFalseWeight = FalseWeight; | |||
8285 | scaleWeights(NewTrueWeight, NewFalseWeight); | |||
8286 | Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext()) | |||
8287 | .createBranchWeights(TrueWeight, FalseWeight)); | |||
8288 | } | |||
8289 | } | |||
8290 | ||||
8291 | ModifiedDT = true; | |||
8292 | MadeChange = true; | |||
8293 | ||||
8294 | LLVM_DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "After branch condition splitting\n" ; BB.dump(); TmpBB->dump(); } } while (false) | |||
8295 | TmpBB->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "After branch condition splitting\n" ; BB.dump(); TmpBB->dump(); } } while (false); | |||
8296 | } | |||
8297 | return MadeChange; | |||
8298 | } |