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/CodeGen/CodeGenPrepare.h"
16#include "llvm/ADT/APInt.h"
17#include "llvm/ADT/ArrayRef.h"
18#include "llvm/ADT/DenseMap.h"
19#include "llvm/ADT/MapVector.h"
20#include "llvm/ADT/PointerIntPair.h"
21#include "llvm/ADT/STLExtras.h"
22#include "llvm/ADT/SmallPtrSet.h"
23#include "llvm/ADT/SmallVector.h"
24#include "llvm/ADT/Statistic.h"
25#include "llvm/Analysis/BlockFrequencyInfo.h"
26#include "llvm/Analysis/BranchProbabilityInfo.h"
27#include "llvm/Analysis/FloatingPointPredicateUtils.h"
28#include "llvm/Analysis/InstructionSimplify.h"
29#include "llvm/Analysis/LoopInfo.h"
30#include "llvm/Analysis/ProfileSummaryInfo.h"
31#include "llvm/Analysis/ScalarEvolutionExpressions.h"
32#include "llvm/Analysis/TargetLibraryInfo.h"
33#include "llvm/Analysis/TargetTransformInfo.h"
34#include "llvm/Analysis/ValueTracking.h"
35#include "llvm/Analysis/VectorUtils.h"
36#include "llvm/CodeGen/Analysis.h"
37#include "llvm/CodeGen/BasicBlockSectionsProfileReader.h"
38#include "llvm/CodeGen/ISDOpcodes.h"
39#include "llvm/CodeGen/SelectionDAGNodes.h"
40#include "llvm/CodeGen/TargetLowering.h"
41#include "llvm/CodeGen/TargetPassConfig.h"
42#include "llvm/CodeGen/TargetSubtargetInfo.h"
43#include "llvm/CodeGen/ValueTypes.h"
44#include "llvm/CodeGenTypes/MachineValueType.h"
45#include "llvm/Config/llvm-config.h"
46#include "llvm/IR/Argument.h"
47#include "llvm/IR/Attributes.h"
48#include "llvm/IR/BasicBlock.h"
49#include "llvm/IR/Constant.h"
50#include "llvm/IR/Constants.h"
51#include "llvm/IR/DataLayout.h"
52#include "llvm/IR/DebugInfo.h"
53#include "llvm/IR/DerivedTypes.h"
54#include "llvm/IR/Dominators.h"
55#include "llvm/IR/Function.h"
56#include "llvm/IR/GetElementPtrTypeIterator.h"
57#include "llvm/IR/GlobalValue.h"
58#include "llvm/IR/GlobalVariable.h"
59#include "llvm/IR/IRBuilder.h"
60#include "llvm/IR/InlineAsm.h"
61#include "llvm/IR/InstrTypes.h"
62#include "llvm/IR/Instruction.h"
63#include "llvm/IR/Instructions.h"
64#include "llvm/IR/IntrinsicInst.h"
65#include "llvm/IR/Intrinsics.h"
66#include "llvm/IR/IntrinsicsAArch64.h"
67#include "llvm/IR/LLVMContext.h"
68#include "llvm/IR/MDBuilder.h"
69#include "llvm/IR/Module.h"
70#include "llvm/IR/Operator.h"
71#include "llvm/IR/PatternMatch.h"
72#include "llvm/IR/ProfDataUtils.h"
73#include "llvm/IR/Statepoint.h"
74#include "llvm/IR/Type.h"
75#include "llvm/IR/Use.h"
76#include "llvm/IR/User.h"
77#include "llvm/IR/Value.h"
78#include "llvm/IR/ValueHandle.h"
79#include "llvm/IR/ValueMap.h"
80#include "llvm/InitializePasses.h"
81#include "llvm/Pass.h"
82#include "llvm/Support/BlockFrequency.h"
83#include "llvm/Support/BranchProbability.h"
84#include "llvm/Support/Casting.h"
85#include "llvm/Support/CommandLine.h"
86#include "llvm/Support/Compiler.h"
87#include "llvm/Support/Debug.h"
88#include "llvm/Support/ErrorHandling.h"
89#include "llvm/Support/raw_ostream.h"
90#include "llvm/Target/TargetMachine.h"
91#include "llvm/Target/TargetOptions.h"
92#include "llvm/Transforms/Utils/BasicBlockUtils.h"
93#include "llvm/Transforms/Utils/BypassSlowDivision.h"
94#include "llvm/Transforms/Utils/Local.h"
95#include "llvm/Transforms/Utils/SimplifyLibCalls.h"
96#include "llvm/Transforms/Utils/SizeOpts.h"
97#include <algorithm>
98#include <cassert>
99#include <cstdint>
100#include <iterator>
101#include <limits>
102#include <memory>
103#include <optional>
104#include <utility>
105#include <vector>
106
107using namespace llvm;
108using namespace llvm::PatternMatch;
109
110#define DEBUG_TYPE "codegenprepare"
111
112STATISTIC(NumBlocksElim, "Number of blocks eliminated");
113STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
114STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
115STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
116 "sunken Cmps");
117STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
118 "of sunken Casts");
119STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
120 "computations were sunk");
121STATISTIC(NumMemoryInstsPhiCreated,
122 "Number of phis created when address "
123 "computations were sunk to memory instructions");
124STATISTIC(NumMemoryInstsSelectCreated,
125 "Number of select created when address "
126 "computations were sunk to memory instructions");
127STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
128STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
129STATISTIC(NumAndsAdded,
130 "Number of and mask instructions added to form ext loads");
131STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
132STATISTIC(NumRetsDup, "Number of return instructions duplicated");
133STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
134STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
135STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
136
137static cl::opt<bool> DisableBranchOpts(
138 "disable-cgp-branch-opts", cl::Hidden, cl::init(Val: false),
139 cl::desc("Disable branch optimizations in CodeGenPrepare"));
140
141static cl::opt<bool>
142 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(Val: false),
143 cl::desc("Disable GC optimizations in CodeGenPrepare"));
144
145static cl::opt<bool>
146 DisableSelectToBranch("disable-cgp-select2branch", cl::Hidden,
147 cl::init(Val: false),
148 cl::desc("Disable select to branch conversion."));
149
150static cl::opt<bool>
151 AddrSinkUsingGEPs("addr-sink-using-gep", cl::Hidden, cl::init(Val: true),
152 cl::desc("Address sinking in CGP using GEPs."));
153
154static cl::opt<bool>
155 EnableAndCmpSinking("enable-andcmp-sinking", cl::Hidden, cl::init(Val: true),
156 cl::desc("Enable sinking and/cmp into branches."));
157
158static cl::opt<bool> DisableStoreExtract(
159 "disable-cgp-store-extract", cl::Hidden, cl::init(Val: false),
160 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
161
162static cl::opt<bool> StressStoreExtract(
163 "stress-cgp-store-extract", cl::Hidden, cl::init(Val: false),
164 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
165
166static cl::opt<bool> DisableExtLdPromotion(
167 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(Val: false),
168 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
169 "CodeGenPrepare"));
170
171static cl::opt<bool> StressExtLdPromotion(
172 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(Val: false),
173 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
174 "optimization in CodeGenPrepare"));
175
176static cl::opt<bool> DisablePreheaderProtect(
177 "disable-preheader-prot", cl::Hidden, cl::init(Val: false),
178 cl::desc("Disable protection against removing loop preheaders"));
179
180static cl::opt<bool> ProfileGuidedSectionPrefix(
181 "profile-guided-section-prefix", cl::Hidden, cl::init(Val: true),
182 cl::desc("Use profile info to add section prefix for hot/cold functions"));
183
184static cl::opt<bool> ProfileUnknownInSpecialSection(
185 "profile-unknown-in-special-section", cl::Hidden,
186 cl::desc("In profiling mode like sampleFDO, if a function doesn't have "
187 "profile, we cannot tell the function is cold for sure because "
188 "it may be a function newly added without ever being sampled. "
189 "With the flag enabled, compiler can put such profile unknown "
190 "functions into a special section, so runtime system can choose "
191 "to handle it in a different way than .text section, to save "
192 "RAM for example. "));
193
194static cl::opt<bool> BBSectionsGuidedSectionPrefix(
195 "bbsections-guided-section-prefix", cl::Hidden, cl::init(Val: true),
196 cl::desc("Use the basic-block-sections profile to determine the text "
197 "section prefix for hot functions. Functions with "
198 "basic-block-sections profile will be placed in `.text.hot` "
199 "regardless of their FDO profile info. Other functions won't be "
200 "impacted, i.e., their prefixes will be decided by FDO/sampleFDO "
201 "profiles."));
202
203static cl::opt<uint64_t> FreqRatioToSkipMerge(
204 "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(Val: 2),
205 cl::desc("Skip merging empty blocks if (frequency of empty block) / "
206 "(frequency of destination block) is greater than this ratio"));
207
208static cl::opt<bool> ForceSplitStore(
209 "force-split-store", cl::Hidden, cl::init(Val: false),
210 cl::desc("Force store splitting no matter what the target query says."));
211
212static cl::opt<bool> EnableTypePromotionMerge(
213 "cgp-type-promotion-merge", cl::Hidden,
214 cl::desc("Enable merging of redundant sexts when one is dominating"
215 " the other."),
216 cl::init(Val: true));
217
218static cl::opt<bool> DisableComplexAddrModes(
219 "disable-complex-addr-modes", cl::Hidden, cl::init(Val: false),
220 cl::desc("Disables combining addressing modes with different parts "
221 "in optimizeMemoryInst."));
222
223static cl::opt<bool>
224 AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(Val: false),
225 cl::desc("Allow creation of Phis in Address sinking."));
226
227static cl::opt<bool> AddrSinkNewSelects(
228 "addr-sink-new-select", cl::Hidden, cl::init(Val: true),
229 cl::desc("Allow creation of selects in Address sinking."));
230
231static cl::opt<bool> AddrSinkCombineBaseReg(
232 "addr-sink-combine-base-reg", cl::Hidden, cl::init(Val: true),
233 cl::desc("Allow combining of BaseReg field in Address sinking."));
234
235static cl::opt<bool> AddrSinkCombineBaseGV(
236 "addr-sink-combine-base-gv", cl::Hidden, cl::init(Val: true),
237 cl::desc("Allow combining of BaseGV field in Address sinking."));
238
239static cl::opt<bool> AddrSinkCombineBaseOffs(
240 "addr-sink-combine-base-offs", cl::Hidden, cl::init(Val: true),
241 cl::desc("Allow combining of BaseOffs field in Address sinking."));
242
243static cl::opt<bool> AddrSinkCombineScaledReg(
244 "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(Val: true),
245 cl::desc("Allow combining of ScaledReg field in Address sinking."));
246
247static cl::opt<bool>
248 EnableGEPOffsetSplit("cgp-split-large-offset-gep", cl::Hidden,
249 cl::init(Val: true),
250 cl::desc("Enable splitting large offset of GEP."));
251
252static cl::opt<bool> EnableICMP_EQToICMP_ST(
253 "cgp-icmp-eq2icmp-st", cl::Hidden, cl::init(Val: false),
254 cl::desc("Enable ICMP_EQ to ICMP_S(L|G)T conversion."));
255
256static cl::opt<bool>
257 VerifyBFIUpdates("cgp-verify-bfi-updates", cl::Hidden, cl::init(Val: false),
258 cl::desc("Enable BFI update verification for "
259 "CodeGenPrepare."));
260
261static cl::opt<bool>
262 OptimizePhiTypes("cgp-optimize-phi-types", cl::Hidden, cl::init(Val: true),
263 cl::desc("Enable converting phi types in CodeGenPrepare"));
264
265static cl::opt<unsigned>
266 HugeFuncThresholdInCGPP("cgpp-huge-func", cl::init(Val: 10000), cl::Hidden,
267 cl::desc("Least BB number of huge function."));
268
269static cl::opt<unsigned>
270 MaxAddressUsersToScan("cgp-max-address-users-to-scan", cl::init(Val: 100),
271 cl::Hidden,
272 cl::desc("Max number of address users to look at"));
273
274static cl::opt<bool>
275 DisableDeletePHIs("disable-cgp-delete-phis", cl::Hidden, cl::init(Val: false),
276 cl::desc("Disable elimination of dead PHI nodes."));
277
278namespace {
279
280enum ExtType {
281 ZeroExtension, // Zero extension has been seen.
282 SignExtension, // Sign extension has been seen.
283 BothExtension // This extension type is used if we saw sext after
284 // ZeroExtension had been set, or if we saw zext after
285 // SignExtension had been set. It makes the type
286 // information of a promoted instruction invalid.
287};
288
289enum ModifyDT {
290 NotModifyDT, // Not Modify any DT.
291 ModifyBBDT, // Modify the Basic Block Dominator Tree.
292 ModifyInstDT // Modify the Instruction Dominator in a Basic Block,
293 // This usually means we move/delete/insert instruction
294 // in a Basic Block. So we should re-iterate instructions
295 // in such Basic Block.
296};
297
298using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
299using TypeIsSExt = PointerIntPair<Type *, 2, ExtType>;
300using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
301using SExts = SmallVector<Instruction *, 16>;
302using ValueToSExts = MapVector<Value *, SExts>;
303
304class TypePromotionTransaction;
305
306class CodeGenPrepare {
307 friend class CodeGenPrepareLegacyPass;
308 const TargetMachine *TM = nullptr;
309 const TargetSubtargetInfo *SubtargetInfo = nullptr;
310 const TargetLowering *TLI = nullptr;
311 const TargetRegisterInfo *TRI = nullptr;
312 const TargetTransformInfo *TTI = nullptr;
313 const BasicBlockSectionsProfileReader *BBSectionsProfileReader = nullptr;
314 const TargetLibraryInfo *TLInfo = nullptr;
315 LoopInfo *LI = nullptr;
316 std::unique_ptr<BlockFrequencyInfo> BFI;
317 std::unique_ptr<BranchProbabilityInfo> BPI;
318 ProfileSummaryInfo *PSI = nullptr;
319
320 /// As we scan instructions optimizing them, this is the next instruction
321 /// to optimize. Transforms that can invalidate this should update it.
322 BasicBlock::iterator CurInstIterator;
323
324 /// Keeps track of non-local addresses that have been sunk into a block.
325 /// This allows us to avoid inserting duplicate code for blocks with
326 /// multiple load/stores of the same address. The usage of WeakTrackingVH
327 /// enables SunkAddrs to be treated as a cache whose entries can be
328 /// invalidated if a sunken address computation has been erased.
329 ValueMap<Value *, WeakTrackingVH> SunkAddrs;
330
331 /// Keeps track of all instructions inserted for the current function.
332 SetOfInstrs InsertedInsts;
333
334 /// Keeps track of the type of the related instruction before their
335 /// promotion for the current function.
336 InstrToOrigTy PromotedInsts;
337
338 /// Keep track of instructions removed during promotion.
339 SetOfInstrs RemovedInsts;
340
341 /// Keep track of sext chains based on their initial value.
342 DenseMap<Value *, Instruction *> SeenChainsForSExt;
343
344 /// Keep track of GEPs accessing the same data structures such as structs or
345 /// arrays that are candidates to be split later because of their large
346 /// size.
347 MapVector<AssertingVH<Value>,
348 SmallVector<std::pair<AssertingVH<GetElementPtrInst>, int64_t>, 32>>
349 LargeOffsetGEPMap;
350
351 /// Keep track of new GEP base after splitting the GEPs having large offset.
352 SmallSet<AssertingVH<Value>, 2> NewGEPBases;
353
354 /// Map serial numbers to Large offset GEPs.
355 DenseMap<AssertingVH<GetElementPtrInst>, int> LargeOffsetGEPID;
356
357 /// Keep track of SExt promoted.
358 ValueToSExts ValToSExtendedUses;
359
360 /// True if the function has the OptSize attribute.
361 bool OptSize;
362
363 /// DataLayout for the Function being processed.
364 const DataLayout *DL = nullptr;
365
366 /// Building the dominator tree can be expensive, so we only build it
367 /// lazily and update it when required.
368 std::unique_ptr<DominatorTree> DT;
369
370public:
371 CodeGenPrepare() = default;
372 CodeGenPrepare(const TargetMachine *TM) : TM(TM){};
373 /// If encounter huge function, we need to limit the build time.
374 bool IsHugeFunc = false;
375
376 /// FreshBBs is like worklist, it collected the updated BBs which need
377 /// to be optimized again.
378 /// Note: Consider building time in this pass, when a BB updated, we need
379 /// to insert such BB into FreshBBs for huge function.
380 SmallPtrSet<BasicBlock *, 32> FreshBBs;
381
382 void releaseMemory() {
383 // Clear per function information.
384 InsertedInsts.clear();
385 PromotedInsts.clear();
386 FreshBBs.clear();
387 BPI.reset();
388 BFI.reset();
389 }
390
391 bool run(Function &F, FunctionAnalysisManager &AM);
392
393private:
394 template <typename F>
395 void resetIteratorIfInvalidatedWhileCalling(BasicBlock *BB, F f) {
396 // Substituting can cause recursive simplifications, which can invalidate
397 // our iterator. Use a WeakTrackingVH to hold onto it in case this
398 // happens.
399 Value *CurValue = &*CurInstIterator;
400 WeakTrackingVH IterHandle(CurValue);
401
402 f();
403
404 // If the iterator instruction was recursively deleted, start over at the
405 // start of the block.
406 if (IterHandle != CurValue) {
407 CurInstIterator = BB->begin();
408 SunkAddrs.clear();
409 }
410 }
411
412 // Get the DominatorTree, building if necessary.
413 DominatorTree &getDT(Function &F) {
414 if (!DT)
415 DT = std::make_unique<DominatorTree>(args&: F);
416 return *DT;
417 }
418
419 void removeAllAssertingVHReferences(Value *V);
420 bool eliminateAssumptions(Function &F);
421 bool eliminateFallThrough(Function &F, DominatorTree *DT = nullptr);
422 bool eliminateMostlyEmptyBlocks(Function &F);
423 BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
424 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
425 void eliminateMostlyEmptyBlock(BasicBlock *BB);
426 bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
427 bool isPreheader);
428 bool makeBitReverse(Instruction &I);
429 bool optimizeBlock(BasicBlock &BB, ModifyDT &ModifiedDT);
430 bool optimizeInst(Instruction *I, ModifyDT &ModifiedDT);
431 bool optimizeMemoryInst(Instruction *MemoryInst, Value *Addr, Type *AccessTy,
432 unsigned AddrSpace);
433 bool optimizeGatherScatterInst(Instruction *MemoryInst, Value *Ptr);
434 bool optimizeMulWithOverflow(Instruction *I, bool IsSigned,
435 ModifyDT &ModifiedDT);
436 bool optimizeInlineAsmInst(CallInst *CS);
437 bool optimizeCallInst(CallInst *CI, ModifyDT &ModifiedDT);
438 bool optimizeExt(Instruction *&I);
439 bool optimizeExtUses(Instruction *I);
440 bool optimizeLoadExt(LoadInst *Load);
441 bool optimizeShiftInst(BinaryOperator *BO);
442 bool optimizeFunnelShift(IntrinsicInst *Fsh);
443 bool optimizeSelectInst(SelectInst *SI);
444 bool optimizeShuffleVectorInst(ShuffleVectorInst *SVI);
445 bool optimizeSwitchType(SwitchInst *SI);
446 bool optimizeSwitchPhiConstants(SwitchInst *SI);
447 bool optimizeSwitchInst(SwitchInst *SI);
448 bool optimizeExtractElementInst(Instruction *Inst);
449 bool dupRetToEnableTailCallOpts(BasicBlock *BB, ModifyDT &ModifiedDT);
450 bool fixupDbgVariableRecord(DbgVariableRecord &I);
451 bool fixupDbgVariableRecordsOnInst(Instruction &I);
452 bool placeDbgValues(Function &F);
453 bool placePseudoProbes(Function &F);
454 bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
455 LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
456 bool tryToPromoteExts(TypePromotionTransaction &TPT,
457 const SmallVectorImpl<Instruction *> &Exts,
458 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
459 unsigned CreatedInstsCost = 0);
460 bool mergeSExts(Function &F);
461 bool splitLargeGEPOffsets();
462 bool optimizePhiType(PHINode *Inst, SmallPtrSetImpl<PHINode *> &Visited,
463 SmallPtrSetImpl<Instruction *> &DeletedInstrs);
464 bool optimizePhiTypes(Function &F);
465 bool performAddressTypePromotion(
466 Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
467 bool HasPromoted, TypePromotionTransaction &TPT,
468 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
469 bool splitBranchCondition(Function &F, ModifyDT &ModifiedDT);
470 bool simplifyOffsetableRelocate(GCStatepointInst &I);
471
472 bool tryToSinkFreeOperands(Instruction *I);
473 bool replaceMathCmpWithIntrinsic(BinaryOperator *BO, Value *Arg0, Value *Arg1,
474 CmpInst *Cmp, Intrinsic::ID IID);
475 bool optimizeCmp(CmpInst *Cmp, ModifyDT &ModifiedDT);
476 bool optimizeURem(Instruction *Rem);
477 bool combineToUSubWithOverflow(CmpInst *Cmp, ModifyDT &ModifiedDT);
478 bool combineToUAddWithOverflow(CmpInst *Cmp, ModifyDT &ModifiedDT);
479 bool unfoldPowerOf2Test(CmpInst *Cmp);
480 void verifyBFIUpdates(Function &F);
481 bool _run(Function &F);
482};
483
484class CodeGenPrepareLegacyPass : public FunctionPass {
485public:
486 static char ID; // Pass identification, replacement for typeid
487
488 CodeGenPrepareLegacyPass() : FunctionPass(ID) {}
489
490 bool runOnFunction(Function &F) override;
491
492 StringRef getPassName() const override { return "CodeGen Prepare"; }
493
494 void getAnalysisUsage(AnalysisUsage &AU) const override {
495 // FIXME: When we can selectively preserve passes, preserve the domtree.
496 AU.addRequired<ProfileSummaryInfoWrapperPass>();
497 AU.addRequired<TargetLibraryInfoWrapperPass>();
498 AU.addRequired<TargetPassConfig>();
499 AU.addRequired<TargetTransformInfoWrapperPass>();
500 AU.addRequired<LoopInfoWrapperPass>();
501 AU.addUsedIfAvailable<BasicBlockSectionsProfileReaderWrapperPass>();
502 }
503};
504
505} // end anonymous namespace
506
507char CodeGenPrepareLegacyPass::ID = 0;
508
509bool CodeGenPrepareLegacyPass::runOnFunction(Function &F) {
510 if (skipFunction(F))
511 return false;
512 auto TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>();
513 CodeGenPrepare CGP(TM);
514 CGP.DL = &F.getDataLayout();
515 CGP.SubtargetInfo = TM->getSubtargetImpl(F);
516 CGP.TLI = CGP.SubtargetInfo->getTargetLowering();
517 CGP.TRI = CGP.SubtargetInfo->getRegisterInfo();
518 CGP.TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
519 CGP.TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
520 CGP.LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
521 CGP.BPI.reset(p: new BranchProbabilityInfo(F, *CGP.LI));
522 CGP.BFI.reset(p: new BlockFrequencyInfo(F, *CGP.BPI, *CGP.LI));
523 CGP.PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
524 auto BBSPRWP =
525 getAnalysisIfAvailable<BasicBlockSectionsProfileReaderWrapperPass>();
526 CGP.BBSectionsProfileReader = BBSPRWP ? &BBSPRWP->getBBSPR() : nullptr;
527
528 return CGP._run(F);
529}
530
531INITIALIZE_PASS_BEGIN(CodeGenPrepareLegacyPass, DEBUG_TYPE,
532 "Optimize for code generation", false, false)
533INITIALIZE_PASS_DEPENDENCY(BasicBlockSectionsProfileReaderWrapperPass)
534INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
535INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
536INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
537INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
538INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
539INITIALIZE_PASS_END(CodeGenPrepareLegacyPass, DEBUG_TYPE,
540 "Optimize for code generation", false, false)
541
542FunctionPass *llvm::createCodeGenPrepareLegacyPass() {
543 return new CodeGenPrepareLegacyPass();
544}
545
546PreservedAnalyses CodeGenPreparePass::run(Function &F,
547 FunctionAnalysisManager &AM) {
548 CodeGenPrepare CGP(TM);
549
550 bool Changed = CGP.run(F, AM);
551 if (!Changed)
552 return PreservedAnalyses::all();
553
554 PreservedAnalyses PA;
555 PA.preserve<TargetLibraryAnalysis>();
556 PA.preserve<TargetIRAnalysis>();
557 return PA;
558}
559
560bool CodeGenPrepare::run(Function &F, FunctionAnalysisManager &AM) {
561 DL = &F.getDataLayout();
562 SubtargetInfo = TM->getSubtargetImpl(F);
563 TLI = SubtargetInfo->getTargetLowering();
564 TRI = SubtargetInfo->getRegisterInfo();
565 TLInfo = &AM.getResult<TargetLibraryAnalysis>(IR&: F);
566 TTI = &AM.getResult<TargetIRAnalysis>(IR&: F);
567 LI = &AM.getResult<LoopAnalysis>(IR&: F);
568 BPI.reset(p: new BranchProbabilityInfo(F, *LI));
569 BFI.reset(p: new BlockFrequencyInfo(F, *BPI, *LI));
570 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(IR&: F);
571 PSI = MAMProxy.getCachedResult<ProfileSummaryAnalysis>(IR&: *F.getParent());
572 BBSectionsProfileReader =
573 AM.getCachedResult<BasicBlockSectionsProfileReaderAnalysis>(IR&: F);
574 return _run(F);
575}
576
577bool CodeGenPrepare::_run(Function &F) {
578 bool EverMadeChange = false;
579
580 OptSize = F.hasOptSize();
581 // Use the basic-block-sections profile to promote hot functions to .text.hot
582 // if requested.
583 if (BBSectionsGuidedSectionPrefix && BBSectionsProfileReader &&
584 BBSectionsProfileReader->isFunctionHot(FuncName: F.getName())) {
585 (void)F.setSectionPrefix("hot");
586 } else if (ProfileGuidedSectionPrefix) {
587 // The hot attribute overwrites profile count based hotness while profile
588 // counts based hotness overwrite the cold attribute.
589 // This is a conservative behabvior.
590 if (F.hasFnAttribute(Kind: Attribute::Hot) ||
591 PSI->isFunctionHotInCallGraph(F: &F, BFI&: *BFI))
592 (void)F.setSectionPrefix("hot");
593 // If PSI shows this function is not hot, we will placed the function
594 // into unlikely section if (1) PSI shows this is a cold function, or
595 // (2) the function has a attribute of cold.
596 else if (PSI->isFunctionColdInCallGraph(F: &F, BFI&: *BFI) ||
597 F.hasFnAttribute(Kind: Attribute::Cold))
598 (void)F.setSectionPrefix("unlikely");
599 else if (ProfileUnknownInSpecialSection && PSI->hasPartialSampleProfile() &&
600 PSI->isFunctionHotnessUnknown(F))
601 (void)F.setSectionPrefix("unknown");
602 }
603
604 /// This optimization identifies DIV instructions that can be
605 /// profitably bypassed and carried out with a shorter, faster divide.
606 if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI->isSlowDivBypassed()) {
607 const DenseMap<unsigned int, unsigned int> &BypassWidths =
608 TLI->getBypassSlowDivWidths();
609 BasicBlock *BB = &*F.begin();
610 while (BB != nullptr) {
611 // bypassSlowDivision may create new BBs, but we don't want to reapply the
612 // optimization to those blocks.
613 BasicBlock *Next = BB->getNextNode();
614 if (!llvm::shouldOptimizeForSize(BB, PSI, BFI: BFI.get()))
615 EverMadeChange |= bypassSlowDivision(BB, BypassWidth: BypassWidths);
616 BB = Next;
617 }
618 }
619
620 // Get rid of @llvm.assume builtins before attempting to eliminate empty
621 // blocks, since there might be blocks that only contain @llvm.assume calls
622 // (plus arguments that we can get rid of).
623 EverMadeChange |= eliminateAssumptions(F);
624
625 // Eliminate blocks that contain only PHI nodes and an
626 // unconditional branch.
627 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
628
629 ModifyDT ModifiedDT = ModifyDT::NotModifyDT;
630 if (!DisableBranchOpts)
631 EverMadeChange |= splitBranchCondition(F, ModifiedDT);
632
633 // Split some critical edges where one of the sources is an indirect branch,
634 // to help generate sane code for PHIs involving such edges.
635 EverMadeChange |=
636 SplitIndirectBrCriticalEdges(F, /*IgnoreBlocksWithoutPHI=*/true);
637
638 // If we are optimzing huge function, we need to consider the build time.
639 // Because the basic algorithm's complex is near O(N!).
640 IsHugeFunc = F.size() > HugeFuncThresholdInCGPP;
641
642 // Transformations above may invalidate dominator tree and/or loop info.
643 DT.reset();
644 LI->releaseMemory();
645 LI->analyze(DomTree: getDT(F));
646
647 bool MadeChange = true;
648 bool FuncIterated = false;
649 while (MadeChange) {
650 MadeChange = false;
651
652 for (BasicBlock &BB : llvm::make_early_inc_range(Range&: F)) {
653 if (FuncIterated && !FreshBBs.contains(Ptr: &BB))
654 continue;
655
656 ModifyDT ModifiedDTOnIteration = ModifyDT::NotModifyDT;
657 bool Changed = optimizeBlock(BB, ModifiedDT&: ModifiedDTOnIteration);
658
659 if (ModifiedDTOnIteration == ModifyDT::ModifyBBDT)
660 DT.reset();
661
662 MadeChange |= Changed;
663 if (IsHugeFunc) {
664 // If the BB is updated, it may still has chance to be optimized.
665 // This usually happen at sink optimization.
666 // For example:
667 //
668 // bb0：
669 // %and = and i32 %a, 4
670 // %cmp = icmp eq i32 %and, 0
671 //
672 // If the %cmp sink to other BB, the %and will has chance to sink.
673 if (Changed)
674 FreshBBs.insert(Ptr: &BB);
675 else if (FuncIterated)
676 FreshBBs.erase(Ptr: &BB);
677 } else {
678 // For small/normal functions, we restart BB iteration if the dominator
679 // tree of the Function was changed.
680 if (ModifiedDTOnIteration != ModifyDT::NotModifyDT)
681 break;
682 }
683 }
684 // We have iterated all the BB in the (only work for huge) function.
685 FuncIterated = IsHugeFunc;
686
687 if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
688 MadeChange |= mergeSExts(F);
689 if (!LargeOffsetGEPMap.empty())
690 MadeChange |= splitLargeGEPOffsets();
691 MadeChange |= optimizePhiTypes(F);
692
693 if (MadeChange)
694 eliminateFallThrough(F, DT: DT.get());
695
696#ifndef NDEBUG
697 if (MadeChange && VerifyLoopInfo)
698 LI->verify(getDT(F));
699#endif
700
701 // Really free removed instructions during promotion.
702 for (Instruction *I : RemovedInsts)
703 I->deleteValue();
704
705 EverMadeChange |= MadeChange;
706 SeenChainsForSExt.clear();
707 ValToSExtendedUses.clear();
708 RemovedInsts.clear();
709 LargeOffsetGEPMap.clear();
710 LargeOffsetGEPID.clear();
711 }
712
713 NewGEPBases.clear();
714 SunkAddrs.clear();
715
716 if (!DisableBranchOpts) {
717 MadeChange = false;
718 // Use a set vector to get deterministic iteration order. The order the
719 // blocks are removed may affect whether or not PHI nodes in successors
720 // are removed.
721 SmallSetVector<BasicBlock *, 8> WorkList;
722 for (BasicBlock &BB : F) {
723 SmallVector<BasicBlock *, 2> Successors(successors(BB: &BB));
724 MadeChange |= ConstantFoldTerminator(BB: &BB, DeleteDeadConditions: true);
725 if (!MadeChange)
726 continue;
727
728 for (BasicBlock *Succ : Successors)
729 if (pred_empty(BB: Succ))
730 WorkList.insert(X: Succ);
731 }
732
733 // Delete the dead blocks and any of their dead successors.
734 MadeChange |= !WorkList.empty();
735 while (!WorkList.empty()) {
736 BasicBlock *BB = WorkList.pop_back_val();
737 SmallVector<BasicBlock *, 2> Successors(successors(BB));
738
739 DeleteDeadBlock(BB);
740
741 for (BasicBlock *Succ : Successors)
742 if (pred_empty(BB: Succ))
743 WorkList.insert(X: Succ);
744 }
745
746 // Merge pairs of basic blocks with unconditional branches, connected by
747 // a single edge.
748 if (EverMadeChange || MadeChange)
749 MadeChange |= eliminateFallThrough(F);
750
751 EverMadeChange |= MadeChange;
752 }
753
754 if (!DisableGCOpts) {
755 SmallVector<GCStatepointInst *, 2> Statepoints;
756 for (BasicBlock &BB : F)
757 for (Instruction &I : BB)
758 if (auto *SP = dyn_cast<GCStatepointInst>(Val: &I))
759 Statepoints.push_back(Elt: SP);
760 for (auto &I : Statepoints)
761 EverMadeChange |= simplifyOffsetableRelocate(I&: *I);
762 }
763
764 // Do this last to clean up use-before-def scenarios introduced by other
765 // preparatory transforms.
766 EverMadeChange |= placeDbgValues(F);
767 EverMadeChange |= placePseudoProbes(F);
768
769#ifndef NDEBUG
770 if (VerifyBFIUpdates)
771 verifyBFIUpdates(F);
772#endif
773
774 return EverMadeChange;
775}
776
777bool CodeGenPrepare::eliminateAssumptions(Function &F) {
778 bool MadeChange = false;
779 for (BasicBlock &BB : F) {
780 CurInstIterator = BB.begin();
781 while (CurInstIterator != BB.end()) {
782 Instruction *I = &*(CurInstIterator++);
783 if (auto *Assume = dyn_cast<AssumeInst>(Val: I)) {
784 MadeChange = true;
785 Value *Operand = Assume->getOperand(i_nocapture: 0);
786 Assume->eraseFromParent();
787
788 resetIteratorIfInvalidatedWhileCalling(BB: &BB, f: [&]() {
789 RecursivelyDeleteTriviallyDeadInstructions(V: Operand, TLI: TLInfo, MSSAU: nullptr);
790 });
791 }
792 }
793 }
794 return MadeChange;
795}
796
797/// An instruction is about to be deleted, so remove all references to it in our
798/// GEP-tracking data strcutures.
799void CodeGenPrepare::removeAllAssertingVHReferences(Value *V) {
800 LargeOffsetGEPMap.erase(Key: V);
801 NewGEPBases.erase(V);
802
803 auto GEP = dyn_cast<GetElementPtrInst>(Val: V);
804 if (!GEP)
805 return;
806
807 LargeOffsetGEPID.erase(Val: GEP);
808
809 auto VecI = LargeOffsetGEPMap.find(Key: GEP->getPointerOperand());
810 if (VecI == LargeOffsetGEPMap.end())
811 return;
812
813 auto &GEPVector = VecI->second;
814 llvm::erase_if(C&: GEPVector, P: [=](auto &Elt) { return Elt.first == GEP; });
815
816 if (GEPVector.empty())
817 LargeOffsetGEPMap.erase(Iterator: VecI);
818}
819
820// Verify BFI has been updated correctly by recomputing BFI and comparing them.
821[[maybe_unused]] void CodeGenPrepare::verifyBFIUpdates(Function &F) {
822 DominatorTree NewDT(F);
823 LoopInfo NewLI(NewDT);
824 BranchProbabilityInfo NewBPI(F, NewLI, TLInfo);
825 BlockFrequencyInfo NewBFI(F, NewBPI, NewLI);
826 NewBFI.verifyMatch(Other&: *BFI);
827}
828
829/// Merge basic blocks which are connected by a single edge, where one of the
830/// basic blocks has a single successor pointing to the other basic block,
831/// which has a single predecessor.
832bool CodeGenPrepare::eliminateFallThrough(Function &F, DominatorTree *DT) {
833 bool Changed = false;
834 // Scan all of the blocks in the function, except for the entry block.
835 // Use a temporary array to avoid iterator being invalidated when
836 // deleting blocks.
837 SmallVector<WeakTrackingVH, 16> Blocks(
838 llvm::make_pointer_range(Range: llvm::drop_begin(RangeOrContainer&: F)));
839
840 SmallSet<WeakTrackingVH, 16> Preds;
841 for (auto &Block : Blocks) {
842 auto *BB = cast_or_null<BasicBlock>(Val&: Block);
843 if (!BB)
844 continue;
845 // If the destination block has a single pred, then this is a trivial
846 // edge, just collapse it.
847 BasicBlock *SinglePred = BB->getSinglePredecessor();
848
849 // Don't merge if BB's address is taken.
850 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken())
851 continue;
852
853 // Make an effort to skip unreachable blocks.
854 if (DT && !DT->isReachableFromEntry(A: BB))
855 continue;
856
857 BranchInst *Term = dyn_cast<BranchInst>(Val: SinglePred->getTerminator());
858 if (Term && !Term->isConditional()) {
859 Changed = true;
860 LLVM_DEBUG(dbgs() << "To merge:\n" << *BB << "\n\n\n");
861
862 // Merge BB into SinglePred and delete it.
863 MergeBlockIntoPredecessor(BB, /* DTU */ nullptr, LI, /* MSSAU */ nullptr,
864 /* MemDep */ nullptr,
865 /* PredecessorWithTwoSuccessors */ false, DT);
866 Preds.insert(V: SinglePred);
867
868 if (IsHugeFunc) {
869 // Update FreshBBs to optimize the merged BB.
870 FreshBBs.insert(Ptr: SinglePred);
871 FreshBBs.erase(Ptr: BB);
872 }
873 }
874 }
875
876 // (Repeatedly) merging blocks into their predecessors can create redundant
877 // debug intrinsics.
878 for (const auto &Pred : Preds)
879 if (auto *BB = cast_or_null<BasicBlock>(Val: Pred))
880 RemoveRedundantDbgInstrs(BB);
881
882 return Changed;
883}
884
885/// Find a destination block from BB if BB is mergeable empty block.
886BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
887 // If this block doesn't end with an uncond branch, ignore it.
888 BranchInst *BI = dyn_cast<BranchInst>(Val: BB->getTerminator());
889 if (!BI || !BI->isUnconditional())
890 return nullptr;
891
892 // If the instruction before the branch (skipping debug info) isn't a phi
893 // node, then other stuff is happening here.
894 BasicBlock::iterator BBI = BI->getIterator();
895 if (BBI != BB->begin()) {
896 --BBI;
897 if (!isa<PHINode>(Val: BBI))
898 return nullptr;
899 }
900
901 // Do not break infinite loops.
902 BasicBlock *DestBB = BI->getSuccessor(i: 0);
903 if (DestBB == BB)
904 return nullptr;
905
906 if (!canMergeBlocks(BB, DestBB))
907 DestBB = nullptr;
908
909 return DestBB;
910}
911
912/// Eliminate blocks that contain only PHI nodes, debug info directives, and an
913/// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
914/// edges in ways that are non-optimal for isel. Start by eliminating these
915/// blocks so we can split them the way we want them.
916bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
917 SmallPtrSet<BasicBlock *, 16> Preheaders;
918 SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
919 while (!LoopList.empty()) {
920 Loop *L = LoopList.pop_back_val();
921 llvm::append_range(C&: LoopList, R&: *L);
922 if (BasicBlock *Preheader = L->getLoopPreheader())
923 Preheaders.insert(Ptr: Preheader);
924 }
925
926 bool MadeChange = false;
927 // Copy blocks into a temporary array to avoid iterator invalidation issues
928 // as we remove them.
929 // Note that this intentionally skips the entry block.
930 SmallVector<WeakTrackingVH, 16> Blocks;
931 for (auto &Block : llvm::drop_begin(RangeOrContainer&: F)) {
932 // Delete phi nodes that could block deleting other empty blocks.
933 if (!DisableDeletePHIs)
934 MadeChange |= DeleteDeadPHIs(BB: &Block, TLI: TLInfo);
935 Blocks.push_back(Elt: &Block);
936 }
937
938 for (auto &Block : Blocks) {
939 BasicBlock *BB = cast_or_null<BasicBlock>(Val&: Block);
940 if (!BB)
941 continue;
942 BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
943 if (!DestBB ||
944 !isMergingEmptyBlockProfitable(BB, DestBB, isPreheader: Preheaders.count(Ptr: BB)))
945 continue;
946
947 eliminateMostlyEmptyBlock(BB);
948 MadeChange = true;
949 }
950 return MadeChange;
951}
952
953bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
954 BasicBlock *DestBB,
955 bool isPreheader) {
956 // Do not delete loop preheaders if doing so would create a critical edge.
957 // Loop preheaders can be good locations to spill registers. If the
958 // preheader is deleted and we create a critical edge, registers may be
959 // spilled in the loop body instead.
960 if (!DisablePreheaderProtect && isPreheader &&
961 !(BB->getSinglePredecessor() &&
962 BB->getSinglePredecessor()->getSingleSuccessor()))
963 return false;
964
965 // Skip merging if the block's successor is also a successor to any callbr
966 // that leads to this block.
967 // FIXME: Is this really needed? Is this a correctness issue?
968 for (BasicBlock *Pred : predecessors(BB)) {
969 if (isa<CallBrInst>(Val: Pred->getTerminator()) &&
970 llvm::is_contained(Range: successors(BB: Pred), Element: DestBB))
971 return false;
972 }
973
974 // Try to skip merging if the unique predecessor of BB is terminated by a
975 // switch or indirect branch instruction, and BB is used as an incoming block
976 // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
977 // add COPY instructions in the predecessor of BB instead of BB (if it is not
978 // merged). Note that the critical edge created by merging such blocks wont be
979 // split in MachineSink because the jump table is not analyzable. By keeping
980 // such empty block (BB), ISel will place COPY instructions in BB, not in the
981 // predecessor of BB.
982 BasicBlock *Pred = BB->getUniquePredecessor();
983 if (!Pred || !(isa<SwitchInst>(Val: Pred->getTerminator()) ||
984 isa<IndirectBrInst>(Val: Pred->getTerminator())))
985 return true;
986
987 if (BB->getTerminator() != &*BB->getFirstNonPHIOrDbg())
988 return true;
989
990 // We use a simple cost heuristic which determine skipping merging is
991 // profitable if the cost of skipping merging is less than the cost of
992 // merging : Cost(skipping merging) < Cost(merging BB), where the
993 // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
994 // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
995 // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
996 // Freq(Pred) / Freq(BB) > 2.
997 // Note that if there are multiple empty blocks sharing the same incoming
998 // value for the PHIs in the DestBB, we consider them together. In such
999 // case, Cost(merging BB) will be the sum of their frequencies.
1000
1001 if (!isa<PHINode>(Val: DestBB->begin()))
1002 return true;
1003
1004 SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
1005
1006 // Find all other incoming blocks from which incoming values of all PHIs in
1007 // DestBB are the same as the ones from BB.
1008 for (BasicBlock *DestBBPred : predecessors(BB: DestBB)) {
1009 if (DestBBPred == BB)
1010 continue;
1011
1012 if (llvm::all_of(Range: DestBB->phis(), P: [&](const PHINode &DestPN) {
1013 return DestPN.getIncomingValueForBlock(BB) ==
1014 DestPN.getIncomingValueForBlock(BB: DestBBPred);
1015 }))
1016 SameIncomingValueBBs.insert(Ptr: DestBBPred);
1017 }
1018
1019 // See if all BB's incoming values are same as the value from Pred. In this
1020 // case, no reason to skip merging because COPYs are expected to be place in
1021 // Pred already.
1022 if (SameIncomingValueBBs.count(Ptr: Pred))
1023 return true;
1024
1025 BlockFrequency PredFreq = BFI->getBlockFreq(BB: Pred);
1026 BlockFrequency BBFreq = BFI->getBlockFreq(BB);
1027
1028 for (auto *SameValueBB : SameIncomingValueBBs)
1029 if (SameValueBB->getUniquePredecessor() == Pred &&
1030 DestBB == findDestBlockOfMergeableEmptyBlock(BB: SameValueBB))
1031 BBFreq += BFI->getBlockFreq(BB: SameValueBB);
1032
1033 std::optional<BlockFrequency> Limit = BBFreq.mul(Factor: FreqRatioToSkipMerge);
1034 return !Limit || PredFreq <= *Limit;
1035}
1036
1037/// Return true if we can merge BB into DestBB if there is a single
1038/// unconditional branch between them, and BB contains no other non-phi
1039/// instructions.
1040bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
1041 const BasicBlock *DestBB) const {
1042 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
1043 // the successor. If there are more complex condition (e.g. preheaders),
1044 // don't mess around with them.
1045 for (const PHINode &PN : BB->phis()) {
1046 for (const User *U : PN.users()) {
1047 const Instruction *UI = cast<Instruction>(Val: U);
1048 if (UI->getParent() != DestBB || !isa<PHINode>(Val: UI))
1049 return false;
1050 // If User is inside DestBB block and it is a PHINode then check
1051 // incoming value. If incoming value is not from BB then this is
1052 // a complex condition (e.g. preheaders) we want to avoid here.
1053 if (UI->getParent() == DestBB) {
1054 if (const PHINode *UPN = dyn_cast<PHINode>(Val: UI))
1055 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
1056 Instruction *Insn = dyn_cast<Instruction>(Val: UPN->getIncomingValue(i: I));
1057 if (Insn && Insn->getParent() == BB &&
1058 Insn->getParent() != UPN->getIncomingBlock(i: I))
1059 return false;
1060 }
1061 }
1062 }
1063 }
1064
1065 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
1066 // and DestBB may have conflicting incoming values for the block. If so, we
1067 // can't merge the block.
1068 const PHINode *DestBBPN = dyn_cast<PHINode>(Val: DestBB->begin());
1069 if (!DestBBPN)
1070 return true; // no conflict.
1071
1072 // Collect the preds of BB.
1073 SmallPtrSet<const BasicBlock *, 16> BBPreds;
1074 if (const PHINode *BBPN = dyn_cast<PHINode>(Val: BB->begin())) {
1075 // It is faster to get preds from a PHI than with pred_iterator.
1076 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1077 BBPreds.insert(Ptr: BBPN->getIncomingBlock(i));
1078 } else {
1079 BBPreds.insert_range(R: predecessors(BB));
1080 }
1081
1082 // Walk the preds of DestBB.
1083 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
1084 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
1085 if (BBPreds.count(Ptr: Pred)) { // Common predecessor?
1086 for (const PHINode &PN : DestBB->phis()) {
1087 const Value *V1 = PN.getIncomingValueForBlock(BB: Pred);
1088 const Value *V2 = PN.getIncomingValueForBlock(BB);
1089
1090 // If V2 is a phi node in BB, look up what the mapped value will be.
1091 if (const PHINode *V2PN = dyn_cast<PHINode>(Val: V2))
1092 if (V2PN->getParent() == BB)
1093 V2 = V2PN->getIncomingValueForBlock(BB: Pred);
1094
1095 // If there is a conflict, bail out.
1096 if (V1 != V2)
1097 return false;
1098 }
1099 }
1100 }
1101
1102 return true;
1103}
1104
1105/// Replace all old uses with new ones, and push the updated BBs into FreshBBs.
1106static void replaceAllUsesWith(Value *Old, Value *New,
1107 SmallPtrSet<BasicBlock *, 32> &FreshBBs,
1108 bool IsHuge) {
1109 auto *OldI = dyn_cast<Instruction>(Val: Old);
1110 if (OldI) {
1111 for (Value::user_iterator UI = OldI->user_begin(), E = OldI->user_end();
1112 UI != E; ++UI) {
1113 Instruction *User = cast<Instruction>(Val: *UI);
1114 if (IsHuge)
1115 FreshBBs.insert(Ptr: User->getParent());
1116 }
1117 }
1118 Old->replaceAllUsesWith(V: New);
1119}
1120
1121/// Eliminate a basic block that has only phi's and an unconditional branch in
1122/// it.
1123void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
1124 BranchInst *BI = cast<BranchInst>(Val: BB->getTerminator());
1125 BasicBlock *DestBB = BI->getSuccessor(i: 0);
1126
1127 LLVM_DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"
1128 << *BB << *DestBB);
1129
1130 // If the destination block has a single pred, then this is a trivial edge,
1131 // just collapse it.
1132 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
1133 if (SinglePred != DestBB) {
1134 assert(SinglePred == BB &&
1135 "Single predecessor not the same as predecessor");
1136 // Merge DestBB into SinglePred/BB and delete it.
1137 MergeBlockIntoPredecessor(BB: DestBB);
1138 // Note: BB(=SinglePred) will not be deleted on this path.
1139 // DestBB(=its single successor) is the one that was deleted.
1140 LLVM_DEBUG(dbgs() << "AFTER:\n" << *SinglePred << "\n\n\n");
1141
1142 if (IsHugeFunc) {
1143 // Update FreshBBs to optimize the merged BB.
1144 FreshBBs.insert(Ptr: SinglePred);
1145 FreshBBs.erase(Ptr: DestBB);
1146 }
1147 return;
1148 }
1149 }
1150
1151 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
1152 // to handle the new incoming edges it is about to have.
1153 for (PHINode &PN : DestBB->phis()) {
1154 // Remove the incoming value for BB, and remember it.
1155 Value *InVal = PN.removeIncomingValue(BB, DeletePHIIfEmpty: false);
1156
1157 // Two options: either the InVal is a phi node defined in BB or it is some
1158 // value that dominates BB.
1159 PHINode *InValPhi = dyn_cast<PHINode>(Val: InVal);
1160 if (InValPhi && InValPhi->getParent() == BB) {
1161 // Add all of the input values of the input PHI as inputs of this phi.
1162 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
1163 PN.addIncoming(V: InValPhi->getIncomingValue(i),
1164 BB: InValPhi->getIncomingBlock(i));
1165 } else {
1166 // Otherwise, add one instance of the dominating value for each edge that
1167 // we will be adding.
1168 if (PHINode *BBPN = dyn_cast<PHINode>(Val: BB->begin())) {
1169 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1170 PN.addIncoming(V: InVal, BB: BBPN->getIncomingBlock(i));
1171 } else {
1172 for (BasicBlock *Pred : predecessors(BB))
1173 PN.addIncoming(V: InVal, BB: Pred);
1174 }
1175 }
1176 }
1177
1178 // Preserve loop Metadata.
1179 if (BI->hasMetadata(KindID: LLVMContext::MD_loop)) {
1180 for (auto *Pred : predecessors(BB))
1181 Pred->getTerminator()->copyMetadata(SrcInst: *BI, WL: LLVMContext::MD_loop);
1182 }
1183
1184 // The PHIs are now updated, change everything that refers to BB to use
1185 // DestBB and remove BB.
1186 BB->replaceAllUsesWith(V: DestBB);
1187 BB->eraseFromParent();
1188 ++NumBlocksElim;
1189
1190 LLVM_DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1191}
1192
1193// Computes a map of base pointer relocation instructions to corresponding
1194// derived pointer relocation instructions given a vector of all relocate calls
1195static void computeBaseDerivedRelocateMap(
1196 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
1197 MapVector<GCRelocateInst *, SmallVector<GCRelocateInst *, 0>>
1198 &RelocateInstMap) {
1199 // Collect information in two maps: one primarily for locating the base object
1200 // while filling the second map; the second map is the final structure holding
1201 // a mapping between Base and corresponding Derived relocate calls
1202 MapVector<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
1203 for (auto *ThisRelocate : AllRelocateCalls) {
1204 auto K = std::make_pair(x: ThisRelocate->getBasePtrIndex(),
1205 y: ThisRelocate->getDerivedPtrIndex());
1206 RelocateIdxMap.insert(KV: std::make_pair(x&: K, y&: ThisRelocate));
1207 }
1208 for (auto &Item : RelocateIdxMap) {
1209 std::pair<unsigned, unsigned> Key = Item.first;
1210 if (Key.first == Key.second)
1211 // Base relocation: nothing to insert
1212 continue;
1213
1214 GCRelocateInst *I = Item.second;
1215 auto BaseKey = std::make_pair(x&: Key.first, y&: Key.first);
1216
1217 // We're iterating over RelocateIdxMap so we cannot modify it.
1218 auto MaybeBase = RelocateIdxMap.find(Key: BaseKey);
1219 if (MaybeBase == RelocateIdxMap.end())
1220 // TODO: We might want to insert a new base object relocate and gep off
1221 // that, if there are enough derived object relocates.
1222 continue;
1223
1224 RelocateInstMap[MaybeBase->second].push_back(Elt: I);
1225 }
1226}
1227
1228// Accepts a GEP and extracts the operands into a vector provided they're all
1229// small integer constants
1230static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
1231 SmallVectorImpl<Value *> &OffsetV) {
1232 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
1233 // Only accept small constant integer operands
1234 auto *Op = dyn_cast<ConstantInt>(Val: GEP->getOperand(i_nocapture: i));
1235 if (!Op || Op->getZExtValue() > 20)
1236 return false;
1237 }
1238
1239 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
1240 OffsetV.push_back(Elt: GEP->getOperand(i_nocapture: i));
1241 return true;
1242}
1243
1244// Takes a RelocatedBase (base pointer relocation instruction) and Targets to
1245// replace, computes a replacement, and affects it.
1246static bool
1247simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
1248 const SmallVectorImpl<GCRelocateInst *> &Targets) {
1249 bool MadeChange = false;
1250 // We must ensure the relocation of derived pointer is defined after
1251 // relocation of base pointer. If we find a relocation corresponding to base
1252 // defined earlier than relocation of base then we move relocation of base
1253 // right before found relocation. We consider only relocation in the same
1254 // basic block as relocation of base. Relocations from other basic block will
1255 // be skipped by optimization and we do not care about them.
1256 for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
1257 &*R != RelocatedBase; ++R)
1258 if (auto *RI = dyn_cast<GCRelocateInst>(Val&: R))
1259 if (RI->getStatepoint() == RelocatedBase->getStatepoint())
1260 if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
1261 RelocatedBase->moveBefore(InsertPos: RI->getIterator());
1262 MadeChange = true;
1263 break;
1264 }
1265
1266 for (GCRelocateInst *ToReplace : Targets) {
1267 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
1268 "Not relocating a derived object of the original base object");
1269 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
1270 // A duplicate relocate call. TODO: coalesce duplicates.
1271 continue;
1272 }
1273
1274 if (RelocatedBase->getParent() != ToReplace->getParent()) {
1275 // Base and derived relocates are in different basic blocks.
1276 // In this case transform is only valid when base dominates derived
1277 // relocate. However it would be too expensive to check dominance
1278 // for each such relocate, so we skip the whole transformation.
1279 continue;
1280 }
1281
1282 Value *Base = ToReplace->getBasePtr();
1283 auto *Derived = dyn_cast<GetElementPtrInst>(Val: ToReplace->getDerivedPtr());
1284 if (!Derived || Derived->getPointerOperand() != Base)
1285 continue;
1286
1287 SmallVector<Value *, 2> OffsetV;
1288 if (!getGEPSmallConstantIntOffsetV(GEP: Derived, OffsetV))
1289 continue;
1290
1291 // Create a Builder and replace the target callsite with a gep
1292 assert(RelocatedBase->getNextNode() &&
1293 "Should always have one since it's not a terminator");
1294
1295 // Insert after RelocatedBase
1296 IRBuilder<> Builder(RelocatedBase->getNextNode());
1297 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1298
1299 // If gc_relocate does not match the actual type, cast it to the right type.
1300 // In theory, there must be a bitcast after gc_relocate if the type does not
1301 // match, and we should reuse it to get the derived pointer. But it could be
1302 // cases like this:
1303 // bb1:
1304 // ...
1305 // %g1 = call coldcc i8 addrspace(1)*
1306 // @llvm.experimental.gc.relocate.p1i8(...) br label %merge
1307 //
1308 // bb2:
1309 // ...
1310 // %g2 = call coldcc i8 addrspace(1)*
1311 // @llvm.experimental.gc.relocate.p1i8(...) br label %merge
1312 //
1313 // merge:
1314 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
1315 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
1316 //
1317 // In this case, we can not find the bitcast any more. So we insert a new
1318 // bitcast no matter there is already one or not. In this way, we can handle
1319 // all cases, and the extra bitcast should be optimized away in later
1320 // passes.
1321 Value *ActualRelocatedBase = RelocatedBase;
1322 if (RelocatedBase->getType() != Base->getType()) {
1323 ActualRelocatedBase =
1324 Builder.CreateBitCast(V: RelocatedBase, DestTy: Base->getType());
1325 }
1326 Value *Replacement =
1327 Builder.CreateGEP(Ty: Derived->getSourceElementType(), Ptr: ActualRelocatedBase,
1328 IdxList: ArrayRef(OffsetV));
1329 Replacement->takeName(V: ToReplace);
1330 // If the newly generated derived pointer's type does not match the original
1331 // derived pointer's type, cast the new derived pointer to match it. Same
1332 // reasoning as above.
1333 Value *ActualReplacement = Replacement;
1334 if (Replacement->getType() != ToReplace->getType()) {
1335 ActualReplacement =
1336 Builder.CreateBitCast(V: Replacement, DestTy: ToReplace->getType());
1337 }
1338 ToReplace->replaceAllUsesWith(V: ActualReplacement);
1339 ToReplace->eraseFromParent();
1340
1341 MadeChange = true;
1342 }
1343 return MadeChange;
1344}
1345
1346// Turns this:
1347//
1348// %base = ...
1349// %ptr = gep %base + 15
1350// %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1351// %base' = relocate(%tok, i32 4, i32 4)
1352// %ptr' = relocate(%tok, i32 4, i32 5)
1353// %val = load %ptr'
1354//
1355// into this:
1356//
1357// %base = ...
1358// %ptr = gep %base + 15
1359// %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1360// %base' = gc.relocate(%tok, i32 4, i32 4)
1361// %ptr' = gep %base' + 15
1362// %val = load %ptr'
1363bool CodeGenPrepare::simplifyOffsetableRelocate(GCStatepointInst &I) {
1364 bool MadeChange = false;
1365 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1366 for (auto *U : I.users())
1367 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(Val: U))
1368 // Collect all the relocate calls associated with a statepoint
1369 AllRelocateCalls.push_back(Elt: Relocate);
1370
1371 // We need at least one base pointer relocation + one derived pointer
1372 // relocation to mangle
1373 if (AllRelocateCalls.size() < 2)
1374 return false;
1375
1376 // RelocateInstMap is a mapping from the base relocate instruction to the
1377 // corresponding derived relocate instructions
1378 MapVector<GCRelocateInst *, SmallVector<GCRelocateInst *, 0>> RelocateInstMap;
1379 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1380 if (RelocateInstMap.empty())
1381 return false;
1382
1383 for (auto &Item : RelocateInstMap)
1384 // Item.first is the RelocatedBase to offset against
1385 // Item.second is the vector of Targets to replace
1386 MadeChange = simplifyRelocatesOffABase(RelocatedBase: Item.first, Targets: Item.second);
1387 return MadeChange;
1388}
1389
1390/// Sink the specified cast instruction into its user blocks.
1391static bool SinkCast(CastInst *CI) {
1392 BasicBlock *DefBB = CI->getParent();
1393
1394 /// InsertedCasts - Only insert a cast in each block once.
1395 DenseMap<BasicBlock *, CastInst *> InsertedCasts;
1396
1397 bool MadeChange = false;
1398 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1399 UI != E;) {
1400 Use &TheUse = UI.getUse();
1401 Instruction *User = cast<Instruction>(Val: *UI);
1402
1403 // Figure out which BB this cast is used in. For PHI's this is the
1404 // appropriate predecessor block.
1405 BasicBlock *UserBB = User->getParent();
1406 if (PHINode *PN = dyn_cast<PHINode>(Val: User)) {
1407 UserBB = PN->getIncomingBlock(U: TheUse);
1408 }
1409
1410 // Preincrement use iterator so we don't invalidate it.
1411 ++UI;
1412
1413 // The first insertion point of a block containing an EH pad is after the
1414 // pad. If the pad is the user, we cannot sink the cast past the pad.
1415 if (User->isEHPad())
1416 continue;
1417
1418 // If the block selected to receive the cast is an EH pad that does not
1419 // allow non-PHI instructions before the terminator, we can't sink the
1420 // cast.
1421 if (UserBB->getTerminator()->isEHPad())
1422 continue;
1423
1424 // If this user is in the same block as the cast, don't change the cast.
1425 if (UserBB == DefBB)
1426 continue;
1427
1428 // If we have already inserted a cast into this block, use it.
1429 CastInst *&InsertedCast = InsertedCasts[UserBB];
1430
1431 if (!InsertedCast) {
1432 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1433 assert(InsertPt != UserBB->end());
1434 InsertedCast = cast<CastInst>(Val: CI->clone());
1435 InsertedCast->insertBefore(BB&: *UserBB, InsertPos: InsertPt);
1436 }
1437
1438 // Replace a use of the cast with a use of the new cast.
1439 TheUse = InsertedCast;
1440 MadeChange = true;
1441 ++NumCastUses;
1442 }
1443
1444 // If we removed all uses, nuke the cast.
1445 if (CI->use_empty()) {
1446 salvageDebugInfo(I&: *CI);
1447 CI->eraseFromParent();
1448 MadeChange = true;
1449 }
1450
1451 return MadeChange;
1452}
1453
1454/// If the specified cast instruction is a noop copy (e.g. it's casting from
1455/// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1456/// reduce the number of virtual registers that must be created and coalesced.
1457///
1458/// Return true if any changes are made.
1459static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1460 const DataLayout &DL) {
1461 // Sink only "cheap" (or nop) address-space casts. This is a weaker condition
1462 // than sinking only nop casts, but is helpful on some platforms.
1463 if (auto *ASC = dyn_cast<AddrSpaceCastInst>(Val: CI)) {
1464 if (!TLI.isFreeAddrSpaceCast(SrcAS: ASC->getSrcAddressSpace(),
1465 DestAS: ASC->getDestAddressSpace()))
1466 return false;
1467 }
1468
1469 // If this is a noop copy,
1470 EVT SrcVT = TLI.getValueType(DL, Ty: CI->getOperand(i_nocapture: 0)->getType());
1471 EVT DstVT = TLI.getValueType(DL, Ty: CI->getType());
1472
1473 // This is an fp<->int conversion?
1474 if (SrcVT.isInteger() != DstVT.isInteger())
1475 return false;
1476
1477 // If this is an extension, it will be a zero or sign extension, which
1478 // isn't a noop.
1479 if (SrcVT.bitsLT(VT: DstVT))
1480 return false;
1481
1482 // If these values will be promoted, find out what they will be promoted
1483 // to. This helps us consider truncates on PPC as noop copies when they
1484 // are.
1485 if (TLI.getTypeAction(Context&: CI->getContext(), VT: SrcVT) ==
1486 TargetLowering::TypePromoteInteger)
1487 SrcVT = TLI.getTypeToTransformTo(Context&: CI->getContext(), VT: SrcVT);
1488 if (TLI.getTypeAction(Context&: CI->getContext(), VT: DstVT) ==
1489 TargetLowering::TypePromoteInteger)
1490 DstVT = TLI.getTypeToTransformTo(Context&: CI->getContext(), VT: DstVT);
1491
1492 // If, after promotion, these are the same types, this is a noop copy.
1493 if (SrcVT != DstVT)
1494 return false;
1495
1496 return SinkCast(CI);
1497}
1498
1499// Match a simple increment by constant operation. Note that if a sub is
1500// matched, the step is negated (as if the step had been canonicalized to
1501// an add, even though we leave the instruction alone.)
1502static bool matchIncrement(const Instruction *IVInc, Instruction *&LHS,
1503 Constant *&Step) {
1504 if (match(V: IVInc, P: m_Add(L: m_Instruction(I&: LHS), R: m_Constant(C&: Step))) ||
1505 match(V: IVInc, P: m_ExtractValue<0>(V: m_Intrinsic<Intrinsic::uadd_with_overflow>(
1506 Op0: m_Instruction(I&: LHS), Op1: m_Constant(C&: Step)))))
1507 return true;
1508 if (match(V: IVInc, P: m_Sub(L: m_Instruction(I&: LHS), R: m_Constant(C&: Step))) ||
1509 match(V: IVInc, P: m_ExtractValue<0>(V: m_Intrinsic<Intrinsic::usub_with_overflow>(
1510 Op0: m_Instruction(I&: LHS), Op1: m_Constant(C&: Step))))) {
1511 Step = ConstantExpr::getNeg(C: Step);
1512 return true;
1513 }
1514 return false;
1515}
1516
1517/// If given \p PN is an inductive variable with value IVInc coming from the
1518/// backedge, and on each iteration it gets increased by Step, return pair
1519/// <IVInc, Step>. Otherwise, return std::nullopt.
1520static std::optional<std::pair<Instruction *, Constant *>>
1521getIVIncrement(const PHINode *PN, const LoopInfo *LI) {
1522 const Loop *L = LI->getLoopFor(BB: PN->getParent());
1523 if (!L || L->getHeader() != PN->getParent() || !L->getLoopLatch())
1524 return std::nullopt;
1525 auto *IVInc =
1526 dyn_cast<Instruction>(Val: PN->getIncomingValueForBlock(BB: L->getLoopLatch()));
1527 if (!IVInc || LI->getLoopFor(BB: IVInc->getParent()) != L)
1528 return std::nullopt;
1529 Instruction *LHS = nullptr;
1530 Constant *Step = nullptr;
1531 if (matchIncrement(IVInc, LHS, Step) && LHS == PN)
1532 return std::make_pair(x&: IVInc, y&: Step);
1533 return std::nullopt;
1534}
1535
1536static bool isIVIncrement(const Value *V, const LoopInfo *LI) {
1537 auto *I = dyn_cast<Instruction>(Val: V);
1538 if (!I)
1539 return false;
1540 Instruction *LHS = nullptr;
1541 Constant *Step = nullptr;
1542 if (!matchIncrement(IVInc: I, LHS, Step))
1543 return false;
1544 if (auto *PN = dyn_cast<PHINode>(Val: LHS))
1545 if (auto IVInc = getIVIncrement(PN, LI))
1546 return IVInc->first == I;
1547 return false;
1548}
1549
1550bool CodeGenPrepare::replaceMathCmpWithIntrinsic(BinaryOperator *BO,
1551 Value *Arg0, Value *Arg1,
1552 CmpInst *Cmp,
1553 Intrinsic::ID IID) {
1554 auto IsReplacableIVIncrement = [this, &Cmp](BinaryOperator *BO) {
1555 if (!isIVIncrement(V: BO, LI))
1556 return false;
1557 const Loop *L = LI->getLoopFor(BB: BO->getParent());
1558 assert(L && "L should not be null after isIVIncrement()");
1559 // Do not risk on moving increment into a child loop.
1560 if (LI->getLoopFor(BB: Cmp->getParent()) != L)
1561 return false;
1562
1563 // Finally, we need to ensure that the insert point will dominate all
1564 // existing uses of the increment.
1565
1566 auto &DT = getDT(F&: *BO->getParent()->getParent());
1567 if (DT.dominates(A: Cmp->getParent(), B: BO->getParent()))
1568 // If we're moving up the dom tree, all uses are trivially dominated.
1569 // (This is the common case for code produced by LSR.)
1570 return true;
1571
1572 // Otherwise, special case the single use in the phi recurrence.
1573 return BO->hasOneUse() && DT.dominates(A: Cmp->getParent(), B: L->getLoopLatch());
1574 };
1575 if (BO->getParent() != Cmp->getParent() && !IsReplacableIVIncrement(BO)) {
1576 // We used to use a dominator tree here to allow multi-block optimization.
1577 // But that was problematic because:
1578 // 1. It could cause a perf regression by hoisting the math op into the
1579 // critical path.
1580 // 2. It could cause a perf regression by creating a value that was live
1581 // across multiple blocks and increasing register pressure.
1582 // 3. Use of a dominator tree could cause large compile-time regression.
1583 // This is because we recompute the DT on every change in the main CGP
1584 // run-loop. The recomputing is probably unnecessary in many cases, so if
1585 // that was fixed, using a DT here would be ok.
1586 //
1587 // There is one important particular case we still want to handle: if BO is
1588 // the IV increment. Important properties that make it profitable:
1589 // - We can speculate IV increment anywhere in the loop (as long as the
1590 // indvar Phi is its only user);
1591 // - Upon computing Cmp, we effectively compute something equivalent to the
1592 // IV increment (despite it loops differently in the IR). So moving it up
1593 // to the cmp point does not really increase register pressure.
1594 return false;
1595 }
1596
1597 // We allow matching the canonical IR (add X, C) back to (usubo X, -C).
1598 if (BO->getOpcode() == Instruction::Add &&
1599 IID == Intrinsic::usub_with_overflow) {
1600 assert(isa<Constant>(Arg1) && "Unexpected input for usubo");
1601 Arg1 = ConstantExpr::getNeg(C: cast<Constant>(Val: Arg1));
1602 }
1603
1604 // Insert at the first instruction of the pair.
1605 Instruction *InsertPt = nullptr;
1606 for (Instruction &Iter : *Cmp->getParent()) {
1607 // If BO is an XOR, it is not guaranteed that it comes after both inputs to
1608 // the overflow intrinsic are defined.
1609 if ((BO->getOpcode() != Instruction::Xor && &Iter == BO) || &Iter == Cmp) {
1610 InsertPt = &Iter;
1611 break;
1612 }
1613 }
1614 assert(InsertPt != nullptr && "Parent block did not contain cmp or binop");
1615
1616 IRBuilder<> Builder(InsertPt);
1617 Value *MathOV = Builder.CreateBinaryIntrinsic(ID: IID, LHS: Arg0, RHS: Arg1);
1618 if (BO->getOpcode() != Instruction::Xor) {
1619 Value *Math = Builder.CreateExtractValue(Agg: MathOV, Idxs: 0, Name: "math");
1620 replaceAllUsesWith(Old: BO, New: Math, FreshBBs, IsHuge: IsHugeFunc);
1621 } else
1622 assert(BO->hasOneUse() &&
1623 "Patterns with XOr should use the BO only in the compare");
1624 Value *OV = Builder.CreateExtractValue(Agg: MathOV, Idxs: 1, Name: "ov");
1625 replaceAllUsesWith(Old: Cmp, New: OV, FreshBBs, IsHuge: IsHugeFunc);
1626 Cmp->eraseFromParent();
1627 BO->eraseFromParent();
1628 return true;
1629}
1630
1631/// Match special-case patterns that check for unsigned add overflow.
1632static bool matchUAddWithOverflowConstantEdgeCases(CmpInst *Cmp,
1633 BinaryOperator *&Add) {
1634 // Add = add A, 1; Cmp = icmp eq A,-1 (overflow if A is max val)
1635 // Add = add A,-1; Cmp = icmp ne A, 0 (overflow if A is non-zero)
1636 Value *A = Cmp->getOperand(i_nocapture: 0), *B = Cmp->getOperand(i_nocapture: 1);
1637
1638 // We are not expecting non-canonical/degenerate code. Just bail out.
1639 if (isa<Constant>(Val: A))
1640 return false;
1641
1642 ICmpInst::Predicate Pred = Cmp->getPredicate();
1643 if (Pred == ICmpInst::ICMP_EQ && match(V: B, P: m_AllOnes()))
1644 B = ConstantInt::get(Ty: B->getType(), V: 1);
1645 else if (Pred == ICmpInst::ICMP_NE && match(V: B, P: m_ZeroInt()))
1646 B = Constant::getAllOnesValue(Ty: B->getType());
1647 else
1648 return false;
1649
1650 // Check the users of the variable operand of the compare looking for an add
1651 // with the adjusted constant.
1652 for (User *U : A->users()) {
1653 if (match(V: U, P: m_Add(L: m_Specific(V: A), R: m_Specific(V: B)))) {
1654 Add = cast<BinaryOperator>(Val: U);
1655 return true;
1656 }
1657 }
1658 return false;
1659}
1660
1661/// Try to combine the compare into a call to the llvm.uadd.with.overflow
1662/// intrinsic. Return true if any changes were made.
1663bool CodeGenPrepare::combineToUAddWithOverflow(CmpInst *Cmp,
1664 ModifyDT &ModifiedDT) {
1665 bool EdgeCase = false;
1666 Value *A, *B;
1667 BinaryOperator *Add;
1668 if (!match(V: Cmp, P: m_UAddWithOverflow(L: m_Value(V&: A), R: m_Value(V&: B), S: m_BinOp(I&: Add)))) {
1669 if (!matchUAddWithOverflowConstantEdgeCases(Cmp, Add))
1670 return false;
1671 // Set A and B in case we match matchUAddWithOverflowConstantEdgeCases.
1672 A = Add->getOperand(i_nocapture: 0);
1673 B = Add->getOperand(i_nocapture: 1);
1674 EdgeCase = true;
1675 }
1676
1677 if (!TLI->shouldFormOverflowOp(Opcode: ISD::UADDO,
1678 VT: TLI->getValueType(DL: *DL, Ty: Add->getType()),
1679 MathUsed: Add->hasNUsesOrMore(N: EdgeCase ? 1 : 2)))
1680 return false;
1681
1682 // We don't want to move around uses of condition values this late, so we
1683 // check if it is legal to create the call to the intrinsic in the basic
1684 // block containing the icmp.
1685 if (Add->getParent() != Cmp->getParent() && !Add->hasOneUse())
1686 return false;
1687
1688 if (!replaceMathCmpWithIntrinsic(BO: Add, Arg0: A, Arg1: B, Cmp,
1689 IID: Intrinsic::uadd_with_overflow))
1690 return false;
1691
1692 // Reset callers - do not crash by iterating over a dead instruction.
1693 ModifiedDT = ModifyDT::ModifyInstDT;
1694 return true;
1695}
1696
1697bool CodeGenPrepare::combineToUSubWithOverflow(CmpInst *Cmp,
1698 ModifyDT &ModifiedDT) {
1699 // We are not expecting non-canonical/degenerate code. Just bail out.
1700 Value *A = Cmp->getOperand(i_nocapture: 0), *B = Cmp->getOperand(i_nocapture: 1);
1701 if (isa<Constant>(Val: A) && isa<Constant>(Val: B))
1702 return false;
1703
1704 // Convert (A u> B) to (A u< B) to simplify pattern matching.
1705 ICmpInst::Predicate Pred = Cmp->getPredicate();
1706 if (Pred == ICmpInst::ICMP_UGT) {
1707 std::swap(a&: A, b&: B);
1708 Pred = ICmpInst::ICMP_ULT;
1709 }
1710 // Convert special-case: (A == 0) is the same as (A u< 1).
1711 if (Pred == ICmpInst::ICMP_EQ && match(V: B, P: m_ZeroInt())) {
1712 B = ConstantInt::get(Ty: B->getType(), V: 1);
1713 Pred = ICmpInst::ICMP_ULT;
1714 }
1715 // Convert special-case: (A != 0) is the same as (0 u< A).
1716 if (Pred == ICmpInst::ICMP_NE && match(V: B, P: m_ZeroInt())) {
1717 std::swap(a&: A, b&: B);
1718 Pred = ICmpInst::ICMP_ULT;
1719 }
1720 if (Pred != ICmpInst::ICMP_ULT)
1721 return false;
1722
1723 // Walk the users of a variable operand of a compare looking for a subtract or
1724 // add with that same operand. Also match the 2nd operand of the compare to
1725 // the add/sub, but that may be a negated constant operand of an add.
1726 Value *CmpVariableOperand = isa<Constant>(Val: A) ? B : A;
1727 BinaryOperator *Sub = nullptr;
1728 for (User *U : CmpVariableOperand->users()) {
1729 // A - B, A u< B --> usubo(A, B)
1730 if (match(V: U, P: m_Sub(L: m_Specific(V: A), R: m_Specific(V: B)))) {
1731 Sub = cast<BinaryOperator>(Val: U);
1732 break;
1733 }
1734
1735 // A + (-C), A u< C (canonicalized form of (sub A, C))
1736 const APInt *CmpC, *AddC;
1737 if (match(V: U, P: m_Add(L: m_Specific(V: A), R: m_APInt(Res&: AddC))) &&
1738 match(V: B, P: m_APInt(Res&: CmpC)) && *AddC == -(*CmpC)) {
1739 Sub = cast<BinaryOperator>(Val: U);
1740 break;
1741 }
1742 }
1743 if (!Sub)
1744 return false;
1745
1746 if (!TLI->shouldFormOverflowOp(Opcode: ISD::USUBO,
1747 VT: TLI->getValueType(DL: *DL, Ty: Sub->getType()),
1748 MathUsed: Sub->hasNUsesOrMore(N: 1)))
1749 return false;
1750
1751 // We don't want to move around uses of condition values this late, so we
1752 // check if it is legal to create the call to the intrinsic in the basic
1753 // block containing the icmp.
1754 if (Sub->getParent() != Cmp->getParent() && !Sub->hasOneUse())
1755 return false;
1756
1757 if (!replaceMathCmpWithIntrinsic(BO: Sub, Arg0: Sub->getOperand(i_nocapture: 0), Arg1: Sub->getOperand(i_nocapture: 1),
1758 Cmp, IID: Intrinsic::usub_with_overflow))
1759 return false;
1760
1761 // Reset callers - do not crash by iterating over a dead instruction.
1762 ModifiedDT = ModifyDT::ModifyInstDT;
1763 return true;
1764}
1765
1766// Decanonicalizes icmp+ctpop power-of-two test if ctpop is slow.
1767// The same transformation exists in DAG combiner, but we repeat it here because
1768// DAG builder can break the pattern by moving icmp into a successor block.
1769bool CodeGenPrepare::unfoldPowerOf2Test(CmpInst *Cmp) {
1770 CmpPredicate Pred;
1771 Value *X;
1772 const APInt *C;
1773
1774 // (icmp (ctpop x), c)
1775 if (!match(V: Cmp, P: m_ICmp(Pred, L: m_Intrinsic<Intrinsic::ctpop>(Op0: m_Value(V&: X)),
1776 R: m_APIntAllowPoison(Res&: C))))
1777 return false;
1778
1779 // We're only interested in "is power of 2 [or zero]" patterns.
1780 bool IsStrictlyPowerOf2Test = ICmpInst::isEquality(P: Pred) && *C == 1;
1781 bool IsPowerOf2OrZeroTest = (Pred == CmpInst::ICMP_ULT && *C == 2) ||
1782 (Pred == CmpInst::ICMP_UGT && *C == 1);
1783 if (!IsStrictlyPowerOf2Test && !IsPowerOf2OrZeroTest)
1784 return false;
1785
1786 // Some targets have better codegen for `ctpop(x) u</u>= 2/1`than for
1787 // `ctpop(x) ==/!= 1`. If ctpop is fast, only try changing the comparison,
1788 // and otherwise expand ctpop into a few simple instructions.
1789 Type *OpTy = X->getType();
1790 if (TLI->isCtpopFast(VT: TLI->getValueType(DL: *DL, Ty: OpTy))) {
1791 // Look for `ctpop(x) ==/!= 1`, where `ctpop(x)` is known to be non-zero.
1792 if (!IsStrictlyPowerOf2Test || !isKnownNonZero(V: Cmp->getOperand(i_nocapture: 0), Q: *DL))
1793 return false;
1794
1795 // ctpop(x) == 1 -> ctpop(x) u< 2
1796 // ctpop(x) != 1 -> ctpop(x) u> 1
1797 if (Pred == ICmpInst::ICMP_EQ) {
1798 Cmp->setOperand(i_nocapture: 1, Val_nocapture: ConstantInt::get(Ty: OpTy, V: 2));
1799 Cmp->setPredicate(ICmpInst::ICMP_ULT);
1800 } else {
1801 Cmp->setPredicate(ICmpInst::ICMP_UGT);
1802 }
1803 return true;
1804 }
1805
1806 Value *NewCmp;
1807 if (IsPowerOf2OrZeroTest ||
1808 (IsStrictlyPowerOf2Test && isKnownNonZero(V: Cmp->getOperand(i_nocapture: 0), Q: *DL))) {
1809 // ctpop(x) u< 2 -> (x & (x - 1)) == 0
1810 // ctpop(x) u> 1 -> (x & (x - 1)) != 0
1811 IRBuilder<> Builder(Cmp);
1812 Value *Sub = Builder.CreateAdd(LHS: X, RHS: Constant::getAllOnesValue(Ty: OpTy));
1813 Value *And = Builder.CreateAnd(LHS: X, RHS: Sub);
1814 CmpInst::Predicate NewPred =
1815 (Pred == CmpInst::ICMP_ULT || Pred == CmpInst::ICMP_EQ)
1816 ? CmpInst::ICMP_EQ
1817 : CmpInst::ICMP_NE;
1818 NewCmp = Builder.CreateICmp(P: NewPred, LHS: And, RHS: ConstantInt::getNullValue(Ty: OpTy));
1819 } else {
1820 // ctpop(x) == 1 -> (x ^ (x - 1)) u> (x - 1)
1821 // ctpop(x) != 1 -> (x ^ (x - 1)) u<= (x - 1)
1822 IRBuilder<> Builder(Cmp);
1823 Value *Sub = Builder.CreateAdd(LHS: X, RHS: Constant::getAllOnesValue(Ty: OpTy));
1824 Value *Xor = Builder.CreateXor(LHS: X, RHS: Sub);
1825 CmpInst::Predicate NewPred =
1826 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT : CmpInst::ICMP_ULE;
1827 NewCmp = Builder.CreateICmp(P: NewPred, LHS: Xor, RHS: Sub);
1828 }
1829
1830 Cmp->replaceAllUsesWith(V: NewCmp);
1831 RecursivelyDeleteTriviallyDeadInstructions(V: Cmp);
1832 return true;
1833}
1834
1835/// Sink the given CmpInst into user blocks to reduce the number of virtual
1836/// registers that must be created and coalesced. This is a clear win except on
1837/// targets with multiple condition code registers (PowerPC), where it might
1838/// lose; some adjustment may be wanted there.
1839///
1840/// Return true if any changes are made.
1841static bool sinkCmpExpression(CmpInst *Cmp, const TargetLowering &TLI,
1842 const DataLayout &DL) {
1843 if (TLI.hasMultipleConditionRegisters(VT: EVT::getEVT(Ty: Cmp->getType())))
1844 return false;
1845
1846 // Avoid sinking soft-FP comparisons, since this can move them into a loop.
1847 if (TLI.useSoftFloat() && isa<FCmpInst>(Val: Cmp))
1848 return false;
1849
1850 bool UsedInPhiOrCurrentBlock = any_of(Range: Cmp->users(), P: [Cmp](User *U) {
1851 return isa<PHINode>(Val: U) ||
1852 cast<Instruction>(Val: U)->getParent() == Cmp->getParent();
1853 });
1854
1855 // Avoid sinking larger than legal integer comparisons unless its ONLY used in
1856 // another BB.
1857 if (UsedInPhiOrCurrentBlock && Cmp->getOperand(i_nocapture: 0)->getType()->isIntegerTy() &&
1858 Cmp->getOperand(i_nocapture: 0)->getType()->getScalarSizeInBits() >
1859 DL.getLargestLegalIntTypeSizeInBits())
1860 return false;
1861
1862 // Only insert a cmp in each block once.
1863 DenseMap<BasicBlock *, CmpInst *> InsertedCmps;
1864
1865 bool MadeChange = false;
1866 for (Value::user_iterator UI = Cmp->user_begin(), E = Cmp->user_end();
1867 UI != E;) {
1868 Use &TheUse = UI.getUse();
1869 Instruction *User = cast<Instruction>(Val: *UI);
1870
1871 // Preincrement use iterator so we don't invalidate it.
1872 ++UI;
1873
1874 // Don't bother for PHI nodes.
1875 if (isa<PHINode>(Val: User))
1876 continue;
1877
1878 // Figure out which BB this cmp is used in.
1879 BasicBlock *UserBB = User->getParent();
1880 BasicBlock *DefBB = Cmp->getParent();
1881
1882 // If this user is in the same block as the cmp, don't change the cmp.
1883 if (UserBB == DefBB)
1884 continue;
1885
1886 // If we have already inserted a cmp into this block, use it.
1887 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1888
1889 if (!InsertedCmp) {
1890 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1891 assert(InsertPt != UserBB->end());
1892 InsertedCmp = CmpInst::Create(Op: Cmp->getOpcode(), Pred: Cmp->getPredicate(),
1893 S1: Cmp->getOperand(i_nocapture: 0), S2: Cmp->getOperand(i_nocapture: 1), Name: "");
1894 InsertedCmp->insertBefore(BB&: *UserBB, InsertPos: InsertPt);
1895 // Propagate the debug info.
1896 InsertedCmp->setDebugLoc(Cmp->getDebugLoc());
1897 }
1898
1899 // Replace a use of the cmp with a use of the new cmp.
1900 TheUse = InsertedCmp;
1901 MadeChange = true;
1902 ++NumCmpUses;
1903 }
1904
1905 // If we removed all uses, nuke the cmp.
1906 if (Cmp->use_empty()) {
1907 Cmp->eraseFromParent();
1908 MadeChange = true;
1909 }
1910
1911 return MadeChange;
1912}
1913
1914/// For pattern like:
1915///
1916/// DomCond = icmp sgt/slt CmpOp0, CmpOp1 (might not be in DomBB)
1917/// ...
1918/// DomBB:
1919/// ...
1920/// br DomCond, TrueBB, CmpBB
1921/// CmpBB: (with DomBB being the single predecessor)
1922/// ...
1923/// Cmp = icmp eq CmpOp0, CmpOp1
1924/// ...
1925///
1926/// It would use two comparison on targets that lowering of icmp sgt/slt is
1927/// different from lowering of icmp eq (PowerPC). This function try to convert
1928/// 'Cmp = icmp eq CmpOp0, CmpOp1' to ' Cmp = icmp slt/sgt CmpOp0, CmpOp1'.
1929/// After that, DomCond and Cmp can use the same comparison so reduce one
1930/// comparison.
1931///
1932/// Return true if any changes are made.
1933static bool foldICmpWithDominatingICmp(CmpInst *Cmp,
1934 const TargetLowering &TLI) {
1935 if (!EnableICMP_EQToICMP_ST && TLI.isEqualityCmpFoldedWithSignedCmp())
1936 return false;
1937
1938 ICmpInst::Predicate Pred = Cmp->getPredicate();
1939 if (Pred != ICmpInst::ICMP_EQ)
1940 return false;
1941
1942 // If icmp eq has users other than BranchInst and SelectInst, converting it to
1943 // icmp slt/sgt would introduce more redundant LLVM IR.
1944 for (User *U : Cmp->users()) {
1945 if (isa<BranchInst>(Val: U))
1946 continue;
1947 if (isa<SelectInst>(Val: U) && cast<SelectInst>(Val: U)->getCondition() == Cmp)
1948 continue;
1949 return false;
1950 }
1951
1952 // This is a cheap/incomplete check for dominance - just match a single
1953 // predecessor with a conditional branch.
1954 BasicBlock *CmpBB = Cmp->getParent();
1955 BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1956 if (!DomBB)
1957 return false;
1958
1959 // We want to ensure that the only way control gets to the comparison of
1960 // interest is that a less/greater than comparison on the same operands is
1961 // false.
1962 Value *DomCond;
1963 BasicBlock *TrueBB, *FalseBB;
1964 if (!match(V: DomBB->getTerminator(), P: m_Br(C: m_Value(V&: DomCond), T&: TrueBB, F&: FalseBB)))
1965 return false;
1966 if (CmpBB != FalseBB)
1967 return false;
1968
1969 Value *CmpOp0 = Cmp->getOperand(i_nocapture: 0), *CmpOp1 = Cmp->getOperand(i_nocapture: 1);
1970 CmpPredicate DomPred;
1971 if (!match(V: DomCond, P: m_ICmp(Pred&: DomPred, L: m_Specific(V: CmpOp0), R: m_Specific(V: CmpOp1))))
1972 return false;
1973 if (DomPred != ICmpInst::ICMP_SGT && DomPred != ICmpInst::ICMP_SLT)
1974 return false;
1975
1976 // Convert the equality comparison to the opposite of the dominating
1977 // comparison and swap the direction for all branch/select users.
1978 // We have conceptually converted:
1979 // Res = (a < b) ? <LT_RES> : (a == b) ? <EQ_RES> : <GT_RES>;
1980 // to
1981 // Res = (a < b) ? <LT_RES> : (a > b) ? <GT_RES> : <EQ_RES>;
1982 // And similarly for branches.
1983 for (User *U : Cmp->users()) {
1984 if (auto *BI = dyn_cast<BranchInst>(Val: U)) {
1985 assert(BI->isConditional() && "Must be conditional");
1986 BI->swapSuccessors();
1987 continue;
1988 }
1989 if (auto *SI = dyn_cast<SelectInst>(Val: U)) {
1990 // Swap operands
1991 SI->swapValues();
1992 SI->swapProfMetadata();
1993 continue;
1994 }
1995 llvm_unreachable("Must be a branch or a select");
1996 }
1997 Cmp->setPredicate(CmpInst::getSwappedPredicate(pred: DomPred));
1998 return true;
1999}
2000
2001/// Many architectures use the same instruction for both subtract and cmp. Try
2002/// to swap cmp operands to match subtract operations to allow for CSE.
2003static bool swapICmpOperandsToExposeCSEOpportunities(CmpInst *Cmp) {
2004 Value *Op0 = Cmp->getOperand(i_nocapture: 0);
2005 Value *Op1 = Cmp->getOperand(i_nocapture: 1);
2006 if (!Op0->getType()->isIntegerTy() || isa<Constant>(Val: Op0) ||
2007 isa<Constant>(Val: Op1) || Op0 == Op1)
2008 return false;
2009
2010 // If a subtract already has the same operands as a compare, swapping would be
2011 // bad. If a subtract has the same operands as a compare but in reverse order,
2012 // then swapping is good.
2013 int GoodToSwap = 0;
2014 unsigned NumInspected = 0;
2015 for (const User *U : Op0->users()) {
2016 // Avoid walking many users.
2017 if (++NumInspected > 128)
2018 return false;
2019 if (match(V: U, P: m_Sub(L: m_Specific(V: Op1), R: m_Specific(V: Op0))))
2020 GoodToSwap++;
2021 else if (match(V: U, P: m_Sub(L: m_Specific(V: Op0), R: m_Specific(V: Op1))))
2022 GoodToSwap--;
2023 }
2024
2025 if (GoodToSwap > 0) {
2026 Cmp->swapOperands();
2027 return true;
2028 }
2029 return false;
2030}
2031
2032static bool foldFCmpToFPClassTest(CmpInst *Cmp, const TargetLowering &TLI,
2033 const DataLayout &DL) {
2034 FCmpInst *FCmp = dyn_cast<FCmpInst>(Val: Cmp);
2035 if (!FCmp)
2036 return false;
2037
2038 // Don't fold if the target offers free fabs and the predicate is legal.
2039 EVT VT = TLI.getValueType(DL, Ty: Cmp->getOperand(i_nocapture: 0)->getType());
2040 if (TLI.isFAbsFree(VT) &&
2041 TLI.isCondCodeLegal(CC: getFCmpCondCode(Pred: FCmp->getPredicate()),
2042 VT: VT.getSimpleVT()))
2043 return false;
2044
2045 // Reverse the canonicalization if it is a FP class test
2046 auto ShouldReverseTransform = [](FPClassTest ClassTest) {
2047 return ClassTest == fcInf || ClassTest == (fcInf | fcNan);
2048 };
2049 auto [ClassVal, ClassTest] =
2050 fcmpToClassTest(Pred: FCmp->getPredicate(), F: *FCmp->getParent()->getParent(),
2051 LHS: FCmp->getOperand(i_nocapture: 0), RHS: FCmp->getOperand(i_nocapture: 1));
2052 if (!ClassVal)
2053 return false;
2054
2055 if (!ShouldReverseTransform(ClassTest) && !ShouldReverseTransform(~ClassTest))
2056 return false;
2057
2058 IRBuilder<> Builder(Cmp);
2059 Value *IsFPClass = Builder.createIsFPClass(FPNum: ClassVal, Test: ClassTest);
2060 Cmp->replaceAllUsesWith(V: IsFPClass);
2061 RecursivelyDeleteTriviallyDeadInstructions(V: Cmp);
2062 return true;
2063}
2064
2065static bool isRemOfLoopIncrementWithLoopInvariant(
2066 Instruction *Rem, const LoopInfo *LI, Value *&RemAmtOut, Value *&AddInstOut,
2067 Value *&AddOffsetOut, PHINode *&LoopIncrPNOut) {
2068 Value *Incr, *RemAmt;
2069 // NB: If RemAmt is a power of 2 it *should* have been transformed by now.
2070 if (!match(V: Rem, P: m_URem(L: m_Value(V&: Incr), R: m_Value(V&: RemAmt))))
2071 return false;
2072
2073 Value *AddInst, *AddOffset;
2074 // Find out loop increment PHI.
2075 auto *PN = dyn_cast<PHINode>(Val: Incr);
2076 if (PN != nullptr) {
2077 AddInst = nullptr;
2078 AddOffset = nullptr;
2079 } else {
2080 // Search through a NUW add on top of the loop increment.
2081 Value *V0, *V1;
2082 if (!match(V: Incr, P: m_NUWAdd(L: m_Value(V&: V0), R: m_Value(V&: V1))))
2083 return false;
2084
2085 AddInst = Incr;
2086 PN = dyn_cast<PHINode>(Val: V0);
2087 if (PN != nullptr) {
2088 AddOffset = V1;
2089 } else {
2090 PN = dyn_cast<PHINode>(Val: V1);
2091 AddOffset = V0;
2092 }
2093 }
2094
2095 if (!PN)
2096 return false;
2097
2098 // This isn't strictly necessary, what we really need is one increment and any
2099 // amount of initial values all being the same.
2100 if (PN->getNumIncomingValues() != 2)
2101 return false;
2102
2103 // Only trivially analyzable loops.
2104 Loop *L = LI->getLoopFor(BB: PN->getParent());
2105 if (!L || !L->getLoopPreheader() || !L->getLoopLatch())
2106 return false;
2107
2108 // Req that the remainder is in the loop
2109 if (!L->contains(Inst: Rem))
2110 return false;
2111
2112 // Only works if the remainder amount is a loop invaraint
2113 if (!L->isLoopInvariant(V: RemAmt))
2114 return false;
2115
2116 // Only works if the AddOffset is a loop invaraint
2117 if (AddOffset && !L->isLoopInvariant(V: AddOffset))
2118 return false;
2119
2120 // Is the PHI a loop increment?
2121 auto LoopIncrInfo = getIVIncrement(PN, LI);
2122 if (!LoopIncrInfo)
2123 return false;
2124
2125 // We need remainder_amount % increment_amount to be zero. Increment of one
2126 // satisfies that without any special logic and is overwhelmingly the common
2127 // case.
2128 if (!match(V: LoopIncrInfo->second, P: m_One()))
2129 return false;
2130
2131 // Need the increment to not overflow.
2132 if (!match(V: LoopIncrInfo->first, P: m_c_NUWAdd(L: m_Specific(V: PN), R: m_Value())))
2133 return false;
2134
2135 // Set output variables.
2136 RemAmtOut = RemAmt;
2137 LoopIncrPNOut = PN;
2138 AddInstOut = AddInst;
2139 AddOffsetOut = AddOffset;
2140
2141 return true;
2142}
2143
2144// Try to transform:
2145//
2146// for(i = Start; i < End; ++i)
2147// Rem = (i nuw+ IncrLoopInvariant) u% RemAmtLoopInvariant;
2148//
2149// ->
2150//
2151// Rem = (Start nuw+ IncrLoopInvariant) % RemAmtLoopInvariant;
2152// for(i = Start; i < End; ++i, ++rem)
2153// Rem = rem == RemAmtLoopInvariant ? 0 : Rem;
2154static bool foldURemOfLoopIncrement(Instruction *Rem, const DataLayout *DL,
2155 const LoopInfo *LI,
2156 SmallPtrSet<BasicBlock *, 32> &FreshBBs,
2157 bool IsHuge) {
2158 Value *AddOffset, *RemAmt, *AddInst;
2159 PHINode *LoopIncrPN;
2160 if (!isRemOfLoopIncrementWithLoopInvariant(Rem, LI, RemAmtOut&: RemAmt, AddInstOut&: AddInst,
2161 AddOffsetOut&: AddOffset, LoopIncrPNOut&: LoopIncrPN))
2162 return false;
2163
2164 // Only non-constant remainder as the extra IV is probably not profitable
2165 // in that case.
2166 //
2167 // Potential TODO(1): `urem` of a const ends up as `mul` + `shift` + `add`. If
2168 // we can rule out register pressure and ensure this `urem` is executed each
2169 // iteration, its probably profitable to handle the const case as well.
2170 //
2171 // Potential TODO(2): Should we have a check for how "nested" this remainder
2172 // operation is? The new code runs every iteration so if the remainder is
2173 // guarded behind unlikely conditions this might not be worth it.
2174 if (match(V: RemAmt, P: m_ImmConstant()))
2175 return false;
2176
2177 Loop *L = LI->getLoopFor(BB: LoopIncrPN->getParent());
2178 Value *Start = LoopIncrPN->getIncomingValueForBlock(BB: L->getLoopPreheader());
2179 // If we have add create initial value for remainder.
2180 // The logic here is:
2181 // (urem (add nuw Start, IncrLoopInvariant), RemAmtLoopInvariant
2182 //
2183 // Only proceed if the expression simplifies (otherwise we can't fully
2184 // optimize out the urem).
2185 if (AddInst) {
2186 assert(AddOffset && "We found an add but missing values");
2187 // Without dom-condition/assumption cache we aren't likely to get much out
2188 // of a context instruction.
2189 Start = simplifyAddInst(LHS: Start, RHS: AddOffset,
2190 IsNSW: match(V: AddInst, P: m_NSWAdd(L: m_Value(), R: m_Value())),
2191 /*IsNUW=*/true, Q: *DL);
2192 if (!Start)
2193 return false;
2194 }
2195
2196 // If we can't fully optimize out the `rem`, skip this transform.
2197 Start = simplifyURemInst(LHS: Start, RHS: RemAmt, Q: *DL);
2198 if (!Start)
2199 return false;
2200
2201 // Create new remainder with induction variable.
2202 Type *Ty = Rem->getType();
2203 IRBuilder<> Builder(Rem->getContext());
2204
2205 Builder.SetInsertPoint(LoopIncrPN);
2206 PHINode *NewRem = Builder.CreatePHI(Ty, NumReservedValues: 2);
2207
2208 Builder.SetInsertPoint(cast<Instruction>(
2209 Val: LoopIncrPN->getIncomingValueForBlock(BB: L->getLoopLatch())));
2210 // `(add (urem x, y), 1)` is always nuw.
2211 Value *RemAdd = Builder.CreateNUWAdd(LHS: NewRem, RHS: ConstantInt::get(Ty, V: 1));
2212 Value *RemCmp = Builder.CreateICmp(P: ICmpInst::ICMP_EQ, LHS: RemAdd, RHS: RemAmt);
2213 Value *RemSel =
2214 Builder.CreateSelect(C: RemCmp, True: Constant::getNullValue(Ty), False: RemAdd);
2215
2216 NewRem->addIncoming(V: Start, BB: L->getLoopPreheader());
2217 NewRem->addIncoming(V: RemSel, BB: L->getLoopLatch());
2218
2219 // Insert all touched BBs.
2220 FreshBBs.insert(Ptr: LoopIncrPN->getParent());
2221 FreshBBs.insert(Ptr: L->getLoopLatch());
2222 FreshBBs.insert(Ptr: Rem->getParent());
2223 if (AddInst)
2224 FreshBBs.insert(Ptr: cast<Instruction>(Val: AddInst)->getParent());
2225 replaceAllUsesWith(Old: Rem, New: NewRem, FreshBBs, IsHuge);
2226 Rem->eraseFromParent();
2227 if (AddInst && AddInst->use_empty())
2228 cast<Instruction>(Val: AddInst)->eraseFromParent();
2229 return true;
2230}
2231
2232bool CodeGenPrepare::optimizeURem(Instruction *Rem) {
2233 if (foldURemOfLoopIncrement(Rem, DL, LI, FreshBBs, IsHuge: IsHugeFunc))
2234 return true;
2235 return false;
2236}
2237
2238bool CodeGenPrepare::optimizeCmp(CmpInst *Cmp, ModifyDT &ModifiedDT) {
2239 if (sinkCmpExpression(Cmp, TLI: *TLI, DL: *DL))
2240 return true;
2241
2242 if (combineToUAddWithOverflow(Cmp, ModifiedDT))
2243 return true;
2244
2245 if (combineToUSubWithOverflow(Cmp, ModifiedDT))
2246 return true;
2247
2248 if (unfoldPowerOf2Test(Cmp))
2249 return true;
2250
2251 if (foldICmpWithDominatingICmp(Cmp, TLI: *TLI))
2252 return true;
2253
2254 if (swapICmpOperandsToExposeCSEOpportunities(Cmp))
2255 return true;
2256
2257 if (foldFCmpToFPClassTest(Cmp, TLI: *TLI, DL: *DL))
2258 return true;
2259
2260 return false;
2261}
2262
2263/// Duplicate and sink the given 'and' instruction into user blocks where it is
2264/// used in a compare to allow isel to generate better code for targets where
2265/// this operation can be combined.
2266///
2267/// Return true if any changes are made.
2268static bool sinkAndCmp0Expression(Instruction *AndI, const TargetLowering &TLI,
2269 SetOfInstrs &InsertedInsts) {
2270 // Double-check that we're not trying to optimize an instruction that was
2271 // already optimized by some other part of this pass.
2272 assert(!InsertedInsts.count(AndI) &&
2273 "Attempting to optimize already optimized and instruction");
2274 (void)InsertedInsts;
2275
2276 // Nothing to do for single use in same basic block.
2277 if (AndI->hasOneUse() &&
2278 AndI->getParent() == cast<Instruction>(Val: *AndI->user_begin())->getParent())
2279 return false;
2280
2281 // Try to avoid cases where sinking/duplicating is likely to increase register
2282 // pressure.
2283 if (!isa<ConstantInt>(Val: AndI->getOperand(i: 0)) &&
2284 !isa<ConstantInt>(Val: AndI->getOperand(i: 1)) &&
2285 AndI->getOperand(i: 0)->hasOneUse() && AndI->getOperand(i: 1)->hasOneUse())
2286 return false;
2287
2288 for (auto *U : AndI->users()) {
2289 Instruction *User = cast<Instruction>(Val: U);
2290
2291 // Only sink 'and' feeding icmp with 0.
2292 if (!isa<ICmpInst>(Val: User))
2293 return false;
2294
2295 auto *CmpC = dyn_cast<ConstantInt>(Val: User->getOperand(i: 1));
2296 if (!CmpC || !CmpC->isZero())
2297 return false;
2298 }
2299
2300 if (!TLI.isMaskAndCmp0FoldingBeneficial(AndI: *AndI))
2301 return false;
2302
2303 LLVM_DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n");
2304 LLVM_DEBUG(AndI->getParent()->dump());
2305
2306 // Push the 'and' into the same block as the icmp 0. There should only be
2307 // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
2308 // others, so we don't need to keep track of which BBs we insert into.
2309 for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
2310 UI != E;) {
2311 Use &TheUse = UI.getUse();
2312 Instruction *User = cast<Instruction>(Val: *UI);
2313
2314 // Preincrement use iterator so we don't invalidate it.
2315 ++UI;
2316
2317 LLVM_DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n");
2318
2319 // Keep the 'and' in the same place if the use is already in the same block.
2320 Instruction *InsertPt =
2321 User->getParent() == AndI->getParent() ? AndI : User;
2322 Instruction *InsertedAnd = BinaryOperator::Create(
2323 Op: Instruction::And, S1: AndI->getOperand(i: 0), S2: AndI->getOperand(i: 1), Name: "",
2324 InsertBefore: InsertPt->getIterator());
2325 // Propagate the debug info.
2326 InsertedAnd->setDebugLoc(AndI->getDebugLoc());
2327
2328 // Replace a use of the 'and' with a use of the new 'and'.
2329 TheUse = InsertedAnd;
2330 ++NumAndUses;
2331 LLVM_DEBUG(User->getParent()->dump());
2332 }
2333
2334 // We removed all uses, nuke the and.
2335 AndI->eraseFromParent();
2336 return true;
2337}
2338
2339/// Check if the candidates could be combined with a shift instruction, which
2340/// includes:
2341/// 1. Truncate instruction
2342/// 2. And instruction and the imm is a mask of the low bits:
2343/// imm & (imm+1) == 0
2344static bool isExtractBitsCandidateUse(Instruction *User) {
2345 if (!isa<TruncInst>(Val: User)) {
2346 if (User->getOpcode() != Instruction::And ||
2347 !isa<ConstantInt>(Val: User->getOperand(i: 1)))
2348 return false;
2349
2350 const APInt &Cimm = cast<ConstantInt>(Val: User->getOperand(i: 1))->getValue();
2351
2352 if ((Cimm & (Cimm + 1)).getBoolValue())
2353 return false;
2354 }
2355 return true;
2356}
2357
2358/// Sink both shift and truncate instruction to the use of truncate's BB.
2359static bool
2360SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
2361 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
2362 const TargetLowering &TLI, const DataLayout &DL) {
2363 BasicBlock *UserBB = User->getParent();
2364 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
2365 auto *TruncI = cast<TruncInst>(Val: User);
2366 bool MadeChange = false;
2367
2368 for (Value::user_iterator TruncUI = TruncI->user_begin(),
2369 TruncE = TruncI->user_end();
2370 TruncUI != TruncE;) {
2371
2372 Use &TruncTheUse = TruncUI.getUse();
2373 Instruction *TruncUser = cast<Instruction>(Val: *TruncUI);
2374 // Preincrement use iterator so we don't invalidate it.
2375
2376 ++TruncUI;
2377
2378 int ISDOpcode = TLI.InstructionOpcodeToISD(Opcode: TruncUser->getOpcode());
2379 if (!ISDOpcode)
2380 continue;
2381
2382 // If the use is actually a legal node, there will not be an
2383 // implicit truncate.
2384 // FIXME: always querying the result type is just an
2385 // approximation; some nodes' legality is determined by the
2386 // operand or other means. There's no good way to find out though.
2387 if (TLI.isOperationLegalOrCustom(
2388 Op: ISDOpcode, VT: TLI.getValueType(DL, Ty: TruncUser->getType(), AllowUnknown: true)))
2389 continue;
2390
2391 // Don't bother for PHI nodes.
2392 if (isa<PHINode>(Val: TruncUser))
2393 continue;
2394
2395 BasicBlock *TruncUserBB = TruncUser->getParent();
2396
2397 if (UserBB == TruncUserBB)
2398 continue;
2399
2400 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
2401 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
2402
2403 if (!InsertedShift && !InsertedTrunc) {
2404 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
2405 assert(InsertPt != TruncUserBB->end());
2406 // Sink the shift
2407 if (ShiftI->getOpcode() == Instruction::AShr)
2408 InsertedShift =
2409 BinaryOperator::CreateAShr(V1: ShiftI->getOperand(i_nocapture: 0), V2: CI, Name: "");
2410 else
2411 InsertedShift =
2412 BinaryOperator::CreateLShr(V1: ShiftI->getOperand(i_nocapture: 0), V2: CI, Name: "");
2413 InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
2414 InsertedShift->insertBefore(BB&: *TruncUserBB, InsertPos: InsertPt);
2415
2416 // Sink the trunc
2417 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
2418 TruncInsertPt++;
2419 // It will go ahead of any debug-info.
2420 TruncInsertPt.setHeadBit(true);
2421 assert(TruncInsertPt != TruncUserBB->end());
2422
2423 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), S: InsertedShift,
2424 Ty: TruncI->getType(), Name: "");
2425 InsertedTrunc->insertBefore(BB&: *TruncUserBB, InsertPos: TruncInsertPt);
2426 InsertedTrunc->setDebugLoc(TruncI->getDebugLoc());
2427
2428 MadeChange = true;
2429
2430 TruncTheUse = InsertedTrunc;
2431 }
2432 }
2433 return MadeChange;
2434}
2435
2436/// Sink the shift *right* instruction into user blocks if the uses could
2437/// potentially be combined with this shift instruction and generate BitExtract
2438/// instruction. It will only be applied if the architecture supports BitExtract
2439/// instruction. Here is an example:
2440/// BB1:
2441/// %x.extract.shift = lshr i64 %arg1, 32
2442/// BB2:
2443/// %x.extract.trunc = trunc i64 %x.extract.shift to i16
2444/// ==>
2445///
2446/// BB2:
2447/// %x.extract.shift.1 = lshr i64 %arg1, 32
2448/// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
2449///
2450/// CodeGen will recognize the pattern in BB2 and generate BitExtract
2451/// instruction.
2452/// Return true if any changes are made.
2453static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
2454 const TargetLowering &TLI,
2455 const DataLayout &DL) {
2456 BasicBlock *DefBB = ShiftI->getParent();
2457
2458 /// Only insert instructions in each block once.
2459 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
2460
2461 bool shiftIsLegal = TLI.isTypeLegal(VT: TLI.getValueType(DL, Ty: ShiftI->getType()));
2462
2463 bool MadeChange = false;
2464 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
2465 UI != E;) {
2466 Use &TheUse = UI.getUse();
2467 Instruction *User = cast<Instruction>(Val: *UI);
2468 // Preincrement use iterator so we don't invalidate it.
2469 ++UI;
2470
2471 // Don't bother for PHI nodes.
2472 if (isa<PHINode>(Val: User))
2473 continue;
2474
2475 if (!isExtractBitsCandidateUse(User))
2476 continue;
2477
2478 BasicBlock *UserBB = User->getParent();
2479
2480 if (UserBB == DefBB) {
2481 // If the shift and truncate instruction are in the same BB. The use of
2482 // the truncate(TruncUse) may still introduce another truncate if not
2483 // legal. In this case, we would like to sink both shift and truncate
2484 // instruction to the BB of TruncUse.
2485 // for example:
2486 // BB1:
2487 // i64 shift.result = lshr i64 opnd, imm
2488 // trunc.result = trunc shift.result to i16
2489 //
2490 // BB2:
2491 // ----> We will have an implicit truncate here if the architecture does
2492 // not have i16 compare.
2493 // cmp i16 trunc.result, opnd2
2494 //
2495 if (isa<TruncInst>(Val: User) &&
2496 shiftIsLegal
2497 // If the type of the truncate is legal, no truncate will be
2498 // introduced in other basic blocks.
2499 && (!TLI.isTypeLegal(VT: TLI.getValueType(DL, Ty: User->getType()))))
2500 MadeChange =
2501 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
2502
2503 continue;
2504 }
2505 // If we have already inserted a shift into this block, use it.
2506 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
2507
2508 if (!InsertedShift) {
2509 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2510 assert(InsertPt != UserBB->end());
2511
2512 if (ShiftI->getOpcode() == Instruction::AShr)
2513 InsertedShift =
2514 BinaryOperator::CreateAShr(V1: ShiftI->getOperand(i_nocapture: 0), V2: CI, Name: "");
2515 else
2516 InsertedShift =
2517 BinaryOperator::CreateLShr(V1: ShiftI->getOperand(i_nocapture: 0), V2: CI, Name: "");
2518 InsertedShift->insertBefore(BB&: *UserBB, InsertPos: InsertPt);
2519 InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
2520
2521 MadeChange = true;
2522 }
2523
2524 // Replace a use of the shift with a use of the new shift.
2525 TheUse = InsertedShift;
2526 }
2527
2528 // If we removed all uses, or there are none, nuke the shift.
2529 if (ShiftI->use_empty()) {
2530 salvageDebugInfo(I&: *ShiftI);
2531 ShiftI->eraseFromParent();
2532 MadeChange = true;
2533 }
2534
2535 return MadeChange;
2536}
2537
2538/// If counting leading or trailing zeros is an expensive operation and a zero
2539/// input is defined, add a check for zero to avoid calling the intrinsic.
2540///
2541/// We want to transform:
2542/// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
2543///
2544/// into:
2545/// entry:
2546/// %cmpz = icmp eq i64 %A, 0
2547/// br i1 %cmpz, label %cond.end, label %cond.false
2548/// cond.false:
2549/// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
2550/// br label %cond.end
2551/// cond.end:
2552/// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
2553///
2554/// If the transform is performed, return true and set ModifiedDT to true.
2555static bool despeculateCountZeros(IntrinsicInst *CountZeros, LoopInfo &LI,
2556 const TargetLowering *TLI,
2557 const DataLayout *DL, ModifyDT &ModifiedDT,
2558 SmallPtrSet<BasicBlock *, 32> &FreshBBs,
2559 bool IsHugeFunc) {
2560 // If a zero input is undefined, it doesn't make sense to despeculate that.
2561 if (match(V: CountZeros->getOperand(i_nocapture: 1), P: m_One()))
2562 return false;
2563
2564 // If it's cheap to speculate, there's nothing to do.
2565 Type *Ty = CountZeros->getType();
2566 auto IntrinsicID = CountZeros->getIntrinsicID();
2567 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz(Ty)) ||
2568 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz(Ty)))
2569 return false;
2570
2571 // Only handle scalar cases. Anything else requires too much work.
2572 unsigned SizeInBits = Ty->getScalarSizeInBits();
2573 if (Ty->isVectorTy())
2574 return false;
2575
2576 // Bail if the value is never zero.
2577 Use &Op = CountZeros->getOperandUse(i: 0);
2578 if (isKnownNonZero(V: Op, Q: *DL))
2579 return false;
2580
2581 // The intrinsic will be sunk behind a compare against zero and branch.
2582 BasicBlock *StartBlock = CountZeros->getParent();
2583 BasicBlock *CallBlock = StartBlock->splitBasicBlock(I: CountZeros, BBName: "cond.false");
2584 if (IsHugeFunc)
2585 FreshBBs.insert(Ptr: CallBlock);
2586
2587 // Create another block after the count zero intrinsic. A PHI will be added
2588 // in this block to select the result of the intrinsic or the bit-width
2589 // constant if the input to the intrinsic is zero.
2590 BasicBlock::iterator SplitPt = std::next(x: BasicBlock::iterator(CountZeros));
2591 // Any debug-info after CountZeros should not be included.
2592 SplitPt.setHeadBit(true);
2593 BasicBlock *EndBlock = CallBlock->splitBasicBlock(I: SplitPt, BBName: "cond.end");
2594 if (IsHugeFunc)
2595 FreshBBs.insert(Ptr: EndBlock);
2596
2597 // Update the LoopInfo. The new blocks are in the same loop as the start
2598 // block.
2599 if (Loop *L = LI.getLoopFor(BB: StartBlock)) {
2600 L->addBasicBlockToLoop(NewBB: CallBlock, LI);
2601 L->addBasicBlockToLoop(NewBB: EndBlock, LI);
2602 }
2603
2604 // Set up a builder to create a compare, conditional branch, and PHI.
2605 IRBuilder<> Builder(CountZeros->getContext());
2606 Builder.SetInsertPoint(StartBlock->getTerminator());
2607 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
2608
2609 // Replace the unconditional branch that was created by the first split with
2610 // a compare against zero and a conditional branch.
2611 Value *Zero = Constant::getNullValue(Ty);
2612 // Avoid introducing branch on poison. This also replaces the ctz operand.
2613 if (!isGuaranteedNotToBeUndefOrPoison(V: Op))
2614 Op = Builder.CreateFreeze(V: Op, Name: Op->getName() + ".fr");
2615 Value *Cmp = Builder.CreateICmpEQ(LHS: Op, RHS: Zero, Name: "cmpz");
2616 Builder.CreateCondBr(Cond: Cmp, True: EndBlock, False: CallBlock);
2617 StartBlock->getTerminator()->eraseFromParent();
2618
2619 // Create a PHI in the end block to select either the output of the intrinsic
2620 // or the bit width of the operand.
2621 Builder.SetInsertPoint(TheBB: EndBlock, IP: EndBlock->begin());
2622 PHINode *PN = Builder.CreatePHI(Ty, NumReservedValues: 2, Name: "ctz");
2623 replaceAllUsesWith(Old: CountZeros, New: PN, FreshBBs, IsHuge: IsHugeFunc);
2624 Value *BitWidth = Builder.getInt(AI: APInt(SizeInBits, SizeInBits));
2625 PN->addIncoming(V: BitWidth, BB: StartBlock);
2626 PN->addIncoming(V: CountZeros, BB: CallBlock);
2627
2628 // We are explicitly handling the zero case, so we can set the intrinsic's
2629 // undefined zero argument to 'true'. This will also prevent reprocessing the
2630 // intrinsic; we only despeculate when a zero input is defined.
2631 CountZeros->setArgOperand(i: 1, v: Builder.getTrue());
2632 ModifiedDT = ModifyDT::ModifyBBDT;
2633 return true;
2634}
2635
2636bool CodeGenPrepare::optimizeCallInst(CallInst *CI, ModifyDT &ModifiedDT) {
2637 BasicBlock *BB = CI->getParent();
2638
2639 // Sink address computing for memory operands into the block.
2640 if (CI->isInlineAsm() && optimizeInlineAsmInst(CS: CI))
2641 return true;
2642
2643 // Align the pointer arguments to this call if the target thinks it's a good
2644 // idea
2645 unsigned MinSize;
2646 Align PrefAlign;
2647 if (TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
2648 for (auto &Arg : CI->args()) {
2649 // We want to align both objects whose address is used directly and
2650 // objects whose address is used in casts and GEPs, though it only makes
2651 // sense for GEPs if the offset is a multiple of the desired alignment and
2652 // if size - offset meets the size threshold.
2653 if (!Arg->getType()->isPointerTy())
2654 continue;
2655 APInt Offset(DL->getIndexSizeInBits(
2656 AS: cast<PointerType>(Val: Arg->getType())->getAddressSpace()),
2657 0);
2658 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(DL: *DL, Offset);
2659 uint64_t Offset2 = Offset.getLimitedValue();
2660 if (!isAligned(Lhs: PrefAlign, SizeInBytes: Offset2))
2661 continue;
2662 AllocaInst *AI;
2663 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlign() < PrefAlign) {
2664 std::optional<TypeSize> AllocaSize = AI->getAllocationSize(DL: *DL);
2665 if (AllocaSize && AllocaSize->getKnownMinValue() >= MinSize + Offset2)
2666 AI->setAlignment(PrefAlign);
2667 }
2668 // Global variables can only be aligned if they are defined in this
2669 // object (i.e. they are uniquely initialized in this object), and
2670 // over-aligning global variables that have an explicit section is
2671 // forbidden.
2672 GlobalVariable *GV;
2673 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
2674 GV->getPointerAlignment(DL: *DL) < PrefAlign &&
2675 GV->getGlobalSize(DL: *DL) >= MinSize + Offset2)
2676 GV->setAlignment(PrefAlign);
2677 }
2678 }
2679 // If this is a memcpy (or similar) then we may be able to improve the
2680 // alignment.
2681 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Val: CI)) {
2682 Align DestAlign = getKnownAlignment(V: MI->getDest(), DL: *DL);
2683 MaybeAlign MIDestAlign = MI->getDestAlign();
2684 if (!MIDestAlign || DestAlign > *MIDestAlign)
2685 MI->setDestAlignment(DestAlign);
2686 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(Val: MI)) {
2687 MaybeAlign MTISrcAlign = MTI->getSourceAlign();
2688 Align SrcAlign = getKnownAlignment(V: MTI->getSource(), DL: *DL);
2689 if (!MTISrcAlign || SrcAlign > *MTISrcAlign)
2690 MTI->setSourceAlignment(SrcAlign);
2691 }
2692 }
2693
2694 // If we have a cold call site, try to sink addressing computation into the
2695 // cold block. This interacts with our handling for loads and stores to
2696 // ensure that we can fold all uses of a potential addressing computation
2697 // into their uses. TODO: generalize this to work over profiling data
2698 if (CI->hasFnAttr(Kind: Attribute::Cold) &&
2699 !llvm::shouldOptimizeForSize(BB, PSI, BFI: BFI.get()))
2700 for (auto &Arg : CI->args()) {
2701 if (!Arg->getType()->isPointerTy())
2702 continue;
2703 unsigned AS = Arg->getType()->getPointerAddressSpace();
2704 if (optimizeMemoryInst(MemoryInst: CI, Addr: Arg, AccessTy: Arg->getType(), AddrSpace: AS))
2705 return true;
2706 }
2707
2708 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: CI);
2709 if (II) {
2710 switch (II->getIntrinsicID()) {
2711 default:
2712 break;
2713 case Intrinsic::assume:
2714 llvm_unreachable("llvm.assume should have been removed already");
2715 case Intrinsic::allow_runtime_check:
2716 case Intrinsic::allow_ubsan_check:
2717 case Intrinsic::experimental_widenable_condition: {
2718 // Give up on future widening opportunities so that we can fold away dead
2719 // paths and merge blocks before going into block-local instruction
2720 // selection.
2721 if (II->use_empty()) {
2722 II->eraseFromParent();
2723 return true;
2724 }
2725 Constant *RetVal = ConstantInt::getTrue(Context&: II->getContext());
2726 resetIteratorIfInvalidatedWhileCalling(BB, f: [&]() {
2727 replaceAndRecursivelySimplify(I: CI, SimpleV: RetVal, TLI: TLInfo, DT: nullptr);
2728 });
2729 return true;
2730 }
2731 case Intrinsic::objectsize:
2732 llvm_unreachable("llvm.objectsize.* should have been lowered already");
2733 case Intrinsic::is_constant:
2734 llvm_unreachable("llvm.is.constant.* should have been lowered already");
2735 case Intrinsic::aarch64_stlxr:
2736 case Intrinsic::aarch64_stxr: {
2737 ZExtInst *ExtVal = dyn_cast<ZExtInst>(Val: CI->getArgOperand(i: 0));
2738 if (!ExtVal || !ExtVal->hasOneUse() ||
2739 ExtVal->getParent() == CI->getParent())
2740 return false;
2741 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
2742 ExtVal->moveBefore(InsertPos: CI->getIterator());
2743 // Mark this instruction as "inserted by CGP", so that other
2744 // optimizations don't touch it.
2745 InsertedInsts.insert(Ptr: ExtVal);
2746 return true;
2747 }
2748
2749 case Intrinsic::launder_invariant_group:
2750 case Intrinsic::strip_invariant_group: {
2751 Value *ArgVal = II->getArgOperand(i: 0);
2752 auto it = LargeOffsetGEPMap.find(Key: II);
2753 if (it != LargeOffsetGEPMap.end()) {
2754 // Merge entries in LargeOffsetGEPMap to reflect the RAUW.
2755 // Make sure not to have to deal with iterator invalidation
2756 // after possibly adding ArgVal to LargeOffsetGEPMap.
2757 auto GEPs = std::move(it->second);
2758 LargeOffsetGEPMap[ArgVal].append(in_start: GEPs.begin(), in_end: GEPs.end());
2759 LargeOffsetGEPMap.erase(Key: II);
2760 }
2761
2762 replaceAllUsesWith(Old: II, New: ArgVal, FreshBBs, IsHuge: IsHugeFunc);
2763 II->eraseFromParent();
2764 return true;
2765 }
2766 case Intrinsic::cttz:
2767 case Intrinsic::ctlz:
2768 // If counting zeros is expensive, try to avoid it.
2769 return despeculateCountZeros(CountZeros: II, LI&: *LI, TLI, DL, ModifiedDT, FreshBBs,
2770 IsHugeFunc);
2771 case Intrinsic::fshl:
2772 case Intrinsic::fshr:
2773 return optimizeFunnelShift(Fsh: II);
2774 case Intrinsic::masked_gather:
2775 return optimizeGatherScatterInst(MemoryInst: II, Ptr: II->getArgOperand(i: 0));
2776 case Intrinsic::masked_scatter:
2777 return optimizeGatherScatterInst(MemoryInst: II, Ptr: II->getArgOperand(i: 1));
2778 case Intrinsic::masked_load:
2779 // Treat v1X masked load as load X type.
2780 if (auto *VT = dyn_cast<FixedVectorType>(Val: II->getType())) {
2781 if (VT->getNumElements() == 1) {
2782 Value *PtrVal = II->getArgOperand(i: 0);
2783 unsigned AS = PtrVal->getType()->getPointerAddressSpace();
2784 if (optimizeMemoryInst(MemoryInst: II, Addr: PtrVal, AccessTy: VT->getElementType(), AddrSpace: AS))
2785 return true;
2786 }
2787 }
2788 return false;
2789 case Intrinsic::masked_store:
2790 // Treat v1X masked store as store X type.
2791 if (auto *VT =
2792 dyn_cast<FixedVectorType>(Val: II->getArgOperand(i: 0)->getType())) {
2793 if (VT->getNumElements() == 1) {
2794 Value *PtrVal = II->getArgOperand(i: 1);
2795 unsigned AS = PtrVal->getType()->getPointerAddressSpace();
2796 if (optimizeMemoryInst(MemoryInst: II, Addr: PtrVal, AccessTy: VT->getElementType(), AddrSpace: AS))
2797 return true;
2798 }
2799 }
2800 return false;
2801 case Intrinsic::umul_with_overflow:
2802 return optimizeMulWithOverflow(I: II, /*IsSigned=*/false, ModifiedDT);
2803 case Intrinsic::smul_with_overflow:
2804 return optimizeMulWithOverflow(I: II, /*IsSigned=*/true, ModifiedDT);
2805 }
2806
2807 SmallVector<Value *, 2> PtrOps;
2808 Type *AccessTy;
2809 if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
2810 while (!PtrOps.empty()) {
2811 Value *PtrVal = PtrOps.pop_back_val();
2812 unsigned AS = PtrVal->getType()->getPointerAddressSpace();
2813 if (optimizeMemoryInst(MemoryInst: II, Addr: PtrVal, AccessTy, AddrSpace: AS))
2814 return true;
2815 }
2816 }
2817
2818 // From here on out we're working with named functions.
2819 auto *Callee = CI->getCalledFunction();
2820 if (!Callee)
2821 return false;
2822
2823 // Lower all default uses of _chk calls. This is very similar
2824 // to what InstCombineCalls does, but here we are only lowering calls
2825 // to fortified library functions (e.g. __memcpy_chk) that have the default
2826 // "don't know" as the objectsize. Anything else should be left alone.
2827 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
2828 IRBuilder<> Builder(CI);
2829 if (Value *V = Simplifier.optimizeCall(CI, B&: Builder)) {
2830 replaceAllUsesWith(Old: CI, New: V, FreshBBs, IsHuge: IsHugeFunc);
2831 CI->eraseFromParent();
2832 return true;
2833 }
2834
2835 // SCCP may have propagated, among other things, C++ static variables across
2836 // calls. If this happens to be the case, we may want to undo it in order to
2837 // avoid redundant pointer computation of the constant, as the function method
2838 // returning the constant needs to be executed anyways.
2839 auto GetUniformReturnValue = [](const Function *F) -> GlobalVariable * {
2840 if (!F->getReturnType()->isPointerTy())
2841 return nullptr;
2842
2843 GlobalVariable *UniformValue = nullptr;
2844 for (auto &BB : *F) {
2845 if (auto *RI = dyn_cast<ReturnInst>(Val: BB.getTerminator())) {
2846 if (auto *V = dyn_cast<GlobalVariable>(Val: RI->getReturnValue())) {
2847 if (!UniformValue)
2848 UniformValue = V;
2849 else if (V != UniformValue)
2850 return nullptr;
2851 } else {
2852 return nullptr;
2853 }
2854 }
2855 }
2856
2857 return UniformValue;
2858 };
2859
2860 if (Callee->hasExactDefinition()) {
2861 if (GlobalVariable *RV = GetUniformReturnValue(Callee)) {
2862 bool MadeChange = false;
2863 for (Use &U : make_early_inc_range(Range: RV->uses())) {
2864 auto *I = dyn_cast<Instruction>(Val: U.getUser());
2865 if (!I || I->getParent() != CI->getParent()) {
2866 // Limit to the same basic block to avoid extending the call-site live
2867 // range, which otherwise could increase register pressure.
2868 continue;
2869 }
2870 if (CI->comesBefore(Other: I)) {
2871 U.set(CI);
2872 MadeChange = true;
2873 }
2874 }
2875
2876 return MadeChange;
2877 }
2878 }
2879
2880 return false;
2881}
2882
2883static bool isIntrinsicOrLFToBeTailCalled(const TargetLibraryInfo *TLInfo,
2884 const CallInst *CI) {
2885 assert(CI && CI->use_empty());
2886
2887 if (const auto *II = dyn_cast<IntrinsicInst>(Val: CI))
2888 switch (II->getIntrinsicID()) {
2889 case Intrinsic::memset:
2890 case Intrinsic::memcpy:
2891 case Intrinsic::memmove:
2892 return true;
2893 default:
2894 return false;
2895 }
2896
2897 LibFunc LF;
2898 Function *Callee = CI->getCalledFunction();
2899 if (Callee && TLInfo && TLInfo->getLibFunc(FDecl: *Callee, F&: LF))
2900 switch (LF) {
2901 case LibFunc_strcpy:
2902 case LibFunc_strncpy:
2903 case LibFunc_strcat:
2904 case LibFunc_strncat:
2905 return true;
2906 default:
2907 return false;
2908 }
2909
2910 return false;
2911}
2912
2913/// Look for opportunities to duplicate return instructions to the predecessor
2914/// to enable tail call optimizations. The case it is currently looking for is
2915/// the following one. Known intrinsics or library function that may be tail
2916/// called are taken into account as well.
2917/// @code
2918/// bb0:
2919/// %tmp0 = tail call i32 @f0()
2920/// br label %return
2921/// bb1:
2922/// %tmp1 = tail call i32 @f1()
2923/// br label %return
2924/// bb2:
2925/// %tmp2 = tail call i32 @f2()
2926/// br label %return
2927/// return:
2928/// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
2929/// ret i32 %retval
2930/// @endcode
2931///
2932/// =>
2933///
2934/// @code
2935/// bb0:
2936/// %tmp0 = tail call i32 @f0()
2937/// ret i32 %tmp0
2938/// bb1:
2939/// %tmp1 = tail call i32 @f1()
2940/// ret i32 %tmp1
2941/// bb2:
2942/// %tmp2 = tail call i32 @f2()
2943/// ret i32 %tmp2
2944/// @endcode
2945bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB,
2946 ModifyDT &ModifiedDT) {
2947 if (!BB->getTerminator())
2948 return false;
2949
2950 ReturnInst *RetI = dyn_cast<ReturnInst>(Val: BB->getTerminator());
2951 if (!RetI)
2952 return false;
2953
2954 assert(LI->getLoopFor(BB) == nullptr && "A return block cannot be in a loop");
2955
2956 PHINode *PN = nullptr;
2957 ExtractValueInst *EVI = nullptr;
2958 BitCastInst *BCI = nullptr;
2959 Value *V = RetI->getReturnValue();
2960 if (V) {
2961 BCI = dyn_cast<BitCastInst>(Val: V);
2962 if (BCI)
2963 V = BCI->getOperand(i_nocapture: 0);
2964
2965 EVI = dyn_cast<ExtractValueInst>(Val: V);
2966 if (EVI) {
2967 V = EVI->getOperand(i_nocapture: 0);
2968 if (!llvm::all_of(Range: EVI->indices(), P: equal_to(Arg: 0)))
2969 return false;
2970 }
2971
2972 PN = dyn_cast<PHINode>(Val: V);
2973 }
2974
2975 if (PN && PN->getParent() != BB)
2976 return false;
2977
2978 auto isLifetimeEndOrBitCastFor = [](const Instruction *Inst) {
2979 const BitCastInst *BC = dyn_cast<BitCastInst>(Val: Inst);
2980 if (BC && BC->hasOneUse())
2981 Inst = BC->user_back();
2982
2983 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: Inst))
2984 return II->getIntrinsicID() == Intrinsic::lifetime_end;
2985 return false;
2986 };
2987
2988 SmallVector<const IntrinsicInst *, 4> FakeUses;
2989
2990 auto isFakeUse = [&FakeUses](const Instruction *Inst) {
2991 if (auto *II = dyn_cast<IntrinsicInst>(Val: Inst);
2992 II && II->getIntrinsicID() == Intrinsic::fake_use) {
2993 // Record the instruction so it can be preserved when the exit block is
2994 // removed. Do not preserve the fake use that uses the result of the
2995 // PHI instruction.
2996 // Do not copy fake uses that use the result of a PHI node.
2997 // FIXME: If we do want to copy the fake use into the return blocks, we
2998 // have to figure out which of the PHI node operands to use for each
2999 // copy.
3000 if (!isa<PHINode>(Val: II->getOperand(i_nocapture: 0))) {
3001 FakeUses.push_back(Elt: II);
3002 }
3003 return true;
3004 }
3005
3006 return false;
3007 };
3008
3009 // Make sure there are no instructions between the first instruction
3010 // and return.
3011 BasicBlock::const_iterator BI = BB->getFirstNonPHIIt();
3012 // Skip over pseudo-probes and the bitcast.
3013 while (&*BI == BCI || &*BI == EVI || isa<PseudoProbeInst>(Val: BI) ||
3014 isLifetimeEndOrBitCastFor(&*BI) || isFakeUse(&*BI))
3015 BI = std::next(x: BI);
3016 if (&*BI != RetI)
3017 return false;
3018
3019 // Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
3020 // call.
3021 auto MayBePermittedAsTailCall = [&](const auto *CI) {
3022 return TLI->mayBeEmittedAsTailCall(CI) &&
3023 attributesPermitTailCall(BB->getParent(), CI, RetI, *TLI);
3024 };
3025
3026 SmallVector<BasicBlock *, 4> TailCallBBs;
3027 // Record the call instructions so we can insert any fake uses
3028 // that need to be preserved before them.
3029 SmallVector<CallInst *, 4> CallInsts;
3030 if (PN) {
3031 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
3032 // Look through bitcasts.
3033 Value *IncomingVal = PN->getIncomingValue(i: I)->stripPointerCasts();
3034 CallInst *CI = dyn_cast<CallInst>(Val: IncomingVal);
3035 BasicBlock *PredBB = PN->getIncomingBlock(i: I);
3036 // Make sure the phi value is indeed produced by the tail call.
3037 if (CI && CI->hasOneUse() && CI->getParent() == PredBB &&
3038 MayBePermittedAsTailCall(CI)) {
3039 TailCallBBs.push_back(Elt: PredBB);
3040 CallInsts.push_back(Elt: CI);
3041 } else {
3042 // Consider the cases in which the phi value is indirectly produced by
3043 // the tail call, for example when encountering memset(), memmove(),
3044 // strcpy(), whose return value may have been optimized out. In such
3045 // cases, the value needs to be the first function argument.
3046 //
3047 // bb0:
3048 // tail call void @llvm.memset.p0.i64(ptr %0, i8 0, i64 %1)
3049 // br label %return
3050 // return:
3051 // %phi = phi ptr [ %0, %bb0 ], [ %2, %entry ]
3052 if (PredBB && PredBB->getSingleSuccessor() == BB)
3053 CI = dyn_cast_or_null<CallInst>(
3054 Val: PredBB->getTerminator()->getPrevNode());
3055
3056 if (CI && CI->use_empty() &&
3057 isIntrinsicOrLFToBeTailCalled(TLInfo, CI) &&
3058 IncomingVal == CI->getArgOperand(i: 0) &&
3059 MayBePermittedAsTailCall(CI)) {
3060 TailCallBBs.push_back(Elt: PredBB);
3061 CallInsts.push_back(Elt: CI);
3062 }
3063 }
3064 }
3065 } else {
3066 SmallPtrSet<BasicBlock *, 4> VisitedBBs;
3067 for (BasicBlock *Pred : predecessors(BB)) {
3068 if (!VisitedBBs.insert(Ptr: Pred).second)
3069 continue;
3070 if (Instruction *I = Pred->rbegin()->getPrevNode()) {
3071 CallInst *CI = dyn_cast<CallInst>(Val: I);
3072 if (CI && CI->use_empty() && MayBePermittedAsTailCall(CI)) {
3073 // Either we return void or the return value must be the first
3074 // argument of a known intrinsic or library function.
3075 if (!V || isa<UndefValue>(Val: V) ||
3076 (isIntrinsicOrLFToBeTailCalled(TLInfo, CI) &&
3077 V == CI->getArgOperand(i: 0))) {
3078 TailCallBBs.push_back(Elt: Pred);
3079 CallInsts.push_back(Elt: CI);
3080 }
3081 }
3082 }
3083 }
3084 }
3085
3086 bool Changed = false;
3087 for (auto const &TailCallBB : TailCallBBs) {
3088 // Make sure the call instruction is followed by an unconditional branch to
3089 // the return block.
3090 BranchInst *BI = dyn_cast<BranchInst>(Val: TailCallBB->getTerminator());
3091 if (!BI || !BI->isUnconditional() || BI->getSuccessor(i: 0) != BB)
3092 continue;
3093
3094 // Duplicate the return into TailCallBB.
3095 (void)FoldReturnIntoUncondBranch(RI: RetI, BB, Pred: TailCallBB);
3096 assert(!VerifyBFIUpdates ||
3097 BFI->getBlockFreq(BB) >= BFI->getBlockFreq(TailCallBB));
3098 BFI->setBlockFreq(BB,
3099 Freq: (BFI->getBlockFreq(BB) - BFI->getBlockFreq(BB: TailCallBB)));
3100 ModifiedDT = ModifyDT::ModifyBBDT;
3101 Changed = true;
3102 ++NumRetsDup;
3103 }
3104
3105 // If we eliminated all predecessors of the block, delete the block now.
3106 if (Changed && !BB->hasAddressTaken() && pred_empty(BB)) {
3107 // Copy the fake uses found in the original return block to all blocks
3108 // that contain tail calls.
3109 for (auto *CI : CallInsts) {
3110 for (auto const *FakeUse : FakeUses) {
3111 auto *ClonedInst = FakeUse->clone();
3112 ClonedInst->insertBefore(InsertPos: CI->getIterator());
3113 }
3114 }
3115 BB->eraseFromParent();
3116 }
3117
3118 return Changed;
3119}
3120
3121//===----------------------------------------------------------------------===//
3122// Memory Optimization
3123//===----------------------------------------------------------------------===//
3124
3125namespace {
3126
3127/// This is an extended version of TargetLowering::AddrMode
3128/// which holds actual Value*'s for register values.
3129struct ExtAddrMode : public TargetLowering::AddrMode {
3130 Value *BaseReg = nullptr;
3131 Value *ScaledReg = nullptr;
3132 Value *OriginalValue = nullptr;
3133 bool InBounds = true;
3134
3135 enum FieldName {
3136 NoField = 0x00,
3137 BaseRegField = 0x01,
3138 BaseGVField = 0x02,
3139 BaseOffsField = 0x04,
3140 ScaledRegField = 0x08,
3141 ScaleField = 0x10,
3142 MultipleFields = 0xff
3143 };
3144
3145 ExtAddrMode() = default;
3146
3147 void print(raw_ostream &OS) const;
3148 void dump() const;
3149
3150 // Replace From in ExtAddrMode with To.
3151 // E.g., SExt insts may be promoted and deleted. We should replace them with
3152 // the promoted values.
3153 void replaceWith(Value *From, Value *To) {
3154 if (ScaledReg == From)
3155 ScaledReg = To;
3156 }
3157
3158 FieldName compare(const ExtAddrMode &other) {
3159 // First check that the types are the same on each field, as differing types
3160 // is something we can't cope with later on.
3161 if (BaseReg && other.BaseReg &&
3162 BaseReg->getType() != other.BaseReg->getType())
3163 return MultipleFields;
3164 if (BaseGV && other.BaseGV && BaseGV->getType() != other.BaseGV->getType())
3165 return MultipleFields;
3166 if (ScaledReg && other.ScaledReg &&
3167 ScaledReg->getType() != other.ScaledReg->getType())
3168 return MultipleFields;
3169
3170 // Conservatively reject 'inbounds' mismatches.
3171 if (InBounds != other.InBounds)
3172 return MultipleFields;
3173
3174 // Check each field to see if it differs.
3175 unsigned Result = NoField;
3176 if (BaseReg != other.BaseReg)
3177 Result |= BaseRegField;
3178 if (BaseGV != other.BaseGV)
3179 Result |= BaseGVField;
3180 if (BaseOffs != other.BaseOffs)
3181 Result |= BaseOffsField;
3182 if (ScaledReg != other.ScaledReg)
3183 Result |= ScaledRegField;
3184 // Don't count 0 as being a different scale, because that actually means
3185 // unscaled (which will already be counted by having no ScaledReg).
3186 if (Scale && other.Scale && Scale != other.Scale)
3187 Result |= ScaleField;
3188
3189 if (llvm::popcount(Value: Result) > 1)
3190 return MultipleFields;
3191 else
3192 return static_cast<FieldName>(Result);
3193 }
3194
3195 // An AddrMode is trivial if it involves no calculation i.e. it is just a base
3196 // with no offset.
3197 bool isTrivial() {
3198 // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
3199 // trivial if at most one of these terms is nonzero, except that BaseGV and
3200 // BaseReg both being zero actually means a null pointer value, which we
3201 // consider to be 'non-zero' here.
3202 return !BaseOffs && !Scale && !(BaseGV && BaseReg);
3203 }
3204
3205 Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
3206 switch (Field) {
3207 default:
3208 return nullptr;
3209 case BaseRegField:
3210 return BaseReg;
3211 case BaseGVField:
3212 return BaseGV;
3213 case ScaledRegField:
3214 return ScaledReg;
3215 case BaseOffsField:
3216 return ConstantInt::getSigned(Ty: IntPtrTy, V: BaseOffs);
3217 }
3218 }
3219
3220 void SetCombinedField(FieldName Field, Value *V,
3221 const SmallVectorImpl<ExtAddrMode> &AddrModes) {
3222 switch (Field) {
3223 default:
3224 llvm_unreachable("Unhandled fields are expected to be rejected earlier");
3225 break;
3226 case ExtAddrMode::BaseRegField:
3227 BaseReg = V;
3228 break;
3229 case ExtAddrMode::BaseGVField:
3230 // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
3231 // in the BaseReg field.
3232 assert(BaseReg == nullptr);
3233 BaseReg = V;
3234 BaseGV = nullptr;
3235 break;
3236 case ExtAddrMode::ScaledRegField:
3237 ScaledReg = V;
3238 // If we have a mix of scaled and unscaled addrmodes then we want scale
3239 // to be the scale and not zero.
3240 if (!Scale)
3241 for (const ExtAddrMode &AM : AddrModes)
3242 if (AM.Scale) {
3243 Scale = AM.Scale;
3244 break;
3245 }
3246 break;
3247 case ExtAddrMode::BaseOffsField:
3248 // The offset is no longer a constant, so it goes in ScaledReg with a
3249 // scale of 1.
3250 assert(ScaledReg == nullptr);
3251 ScaledReg = V;
3252 Scale = 1;
3253 BaseOffs = 0;
3254 break;
3255 }
3256 }
3257};
3258
3259#ifndef NDEBUG
3260static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
3261 AM.print(OS);
3262 return OS;
3263}
3264#endif
3265
3266#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3267void ExtAddrMode::print(raw_ostream &OS) const {
3268 bool NeedPlus = false;
3269 OS << "[";
3270 if (InBounds)
3271 OS << "inbounds ";
3272 if (BaseGV) {
3273 OS << "GV:";
3274 BaseGV->printAsOperand(OS, /*PrintType=*/false);
3275 NeedPlus = true;
3276 }
3277
3278 if (BaseOffs) {
3279 OS << (NeedPlus ? " + " : "") << BaseOffs;
3280 NeedPlus = true;
3281 }
3282
3283 if (BaseReg) {
3284 OS << (NeedPlus ? " + " : "") << "Base:";
3285 BaseReg->printAsOperand(OS, /*PrintType=*/false);
3286 NeedPlus = true;
3287 }
3288 if (Scale) {
3289 OS << (NeedPlus ? " + " : "") << Scale << "*";
3290 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
3291 }
3292
3293 OS << ']';
3294}
3295
3296LLVM_DUMP_METHOD void ExtAddrMode::dump() const {
3297 print(dbgs());
3298 dbgs() << '\n';
3299}
3300#endif
3301
3302} // end anonymous namespace
3303
3304namespace {
3305
3306/// This class provides transaction based operation on the IR.
3307/// Every change made through this class is recorded in the internal state and
3308/// can be undone (rollback) until commit is called.
3309/// CGP does not check if instructions could be speculatively executed when
3310/// moved. Preserving the original location would pessimize the debugging
3311/// experience, as well as negatively impact the quality of sample PGO.
3312class TypePromotionTransaction {
3313 /// This represents the common interface of the individual transaction.
3314 /// Each class implements the logic for doing one specific modification on
3315 /// the IR via the TypePromotionTransaction.
3316 class TypePromotionAction {
3317 protected:
3318 /// The Instruction modified.
3319 Instruction *Inst;
3320
3321 public:
3322 /// Constructor of the action.
3323 /// The constructor performs the related action on the IR.
3324 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
3325
3326 virtual ~TypePromotionAction() = default;
3327
3328 /// Undo the modification done by this action.
3329 /// When this method is called, the IR must be in the same state as it was
3330 /// before this action was applied.
3331 /// \pre Undoing the action works if and only if the IR is in the exact same
3332 /// state as it was directly after this action was applied.
3333 virtual void undo() = 0;
3334
3335 /// Advocate every change made by this action.
3336 /// When the results on the IR of the action are to be kept, it is important
3337 /// to call this function, otherwise hidden information may be kept forever.
3338 virtual void commit() {
3339 // Nothing to be done, this action is not doing anything.
3340 }
3341 };
3342
3343 /// Utility to remember the position of an instruction.
3344 class InsertionHandler {
3345 /// Position of an instruction.
3346 /// Either an instruction:
3347 /// - Is the first in a basic block: BB is used.
3348 /// - Has a previous instruction: PrevInst is used.
3349 struct {
3350 BasicBlock::iterator PrevInst;
3351 BasicBlock *BB;
3352 } Point;
3353 std::optional<DbgRecord::self_iterator> BeforeDbgRecord = std::nullopt;
3354
3355 /// Remember whether or not the instruction had a previous instruction.
3356 bool HasPrevInstruction;
3357
3358 public:
3359 /// Record the position of \p Inst.
3360 InsertionHandler(Instruction *Inst) {
3361 HasPrevInstruction = (Inst != &*(Inst->getParent()->begin()));
3362 BasicBlock *BB = Inst->getParent();
3363
3364 // Record where we would have to re-insert the instruction in the sequence
3365 // of DbgRecords, if we ended up reinserting.
3366 BeforeDbgRecord = Inst->getDbgReinsertionPosition();
3367
3368 if (HasPrevInstruction) {
3369 Point.PrevInst = std::prev(x: Inst->getIterator());
3370 } else {
3371 Point.BB = BB;
3372 }
3373 }
3374
3375 /// Insert \p Inst at the recorded position.
3376 void insert(Instruction *Inst) {
3377 if (HasPrevInstruction) {
3378 if (Inst->getParent())
3379 Inst->removeFromParent();
3380 Inst->insertAfter(InsertPos: Point.PrevInst);
3381 } else {
3382 BasicBlock::iterator Position = Point.BB->getFirstInsertionPt();
3383 if (Inst->getParent())
3384 Inst->moveBefore(BB&: *Point.BB, I: Position);
3385 else
3386 Inst->insertBefore(BB&: *Point.BB, InsertPos: Position);
3387 }
3388
3389 Inst->getParent()->reinsertInstInDbgRecords(I: Inst, Pos: BeforeDbgRecord);
3390 }
3391 };
3392
3393 /// Move an instruction before another.
3394 class InstructionMoveBefore : public TypePromotionAction {
3395 /// Original position of the instruction.
3396 InsertionHandler Position;
3397
3398 public:
3399 /// Move \p Inst before \p Before.
3400 InstructionMoveBefore(Instruction *Inst, BasicBlock::iterator Before)
3401 : TypePromotionAction(Inst), Position(Inst) {
3402 LLVM_DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before
3403 << "\n");
3404 Inst->moveBefore(InsertPos: Before);
3405 }
3406
3407 /// Move the instruction back to its original position.
3408 void undo() override {
3409 LLVM_DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
3410 Position.insert(Inst);
3411 }
3412 };
3413
3414 /// Set the operand of an instruction with a new value.
3415 class OperandSetter : public TypePromotionAction {
3416 /// Original operand of the instruction.
3417 Value *Origin;
3418
3419 /// Index of the modified instruction.
3420 unsigned Idx;
3421
3422 public:
3423 /// Set \p Idx operand of \p Inst with \p NewVal.
3424 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
3425 : TypePromotionAction(Inst), Idx(Idx) {
3426 LLVM_DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
3427 << "for:" << *Inst << "\n"
3428 << "with:" << *NewVal << "\n");
3429 Origin = Inst->getOperand(i: Idx);
3430 Inst->setOperand(i: Idx, Val: NewVal);
3431 }
3432
3433 /// Restore the original value of the instruction.
3434 void undo() override {
3435 LLVM_DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
3436 << "for: " << *Inst << "\n"
3437 << "with: " << *Origin << "\n");
3438 Inst->setOperand(i: Idx, Val: Origin);
3439 }
3440 };
3441
3442 /// Hide the operands of an instruction.
3443 /// Do as if this instruction was not using any of its operands.
3444 class OperandsHider : public TypePromotionAction {
3445 /// The list of original operands.
3446 SmallVector<Value *, 4> OriginalValues;
3447
3448 public:
3449 /// Remove \p Inst from the uses of the operands of \p Inst.
3450 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
3451 LLVM_DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
3452 unsigned NumOpnds = Inst->getNumOperands();
3453 OriginalValues.reserve(N: NumOpnds);
3454 for (unsigned It = 0; It < NumOpnds; ++It) {
3455 // Save the current operand.
3456 Value *Val = Inst->getOperand(i: It);
3457 OriginalValues.push_back(Elt: Val);
3458 // Set a dummy one.
3459 // We could use OperandSetter here, but that would imply an overhead
3460 // that we are not willing to pay.
3461 Inst->setOperand(i: It, Val: PoisonValue::get(T: Val->getType()));
3462 }
3463 }
3464
3465 /// Restore the original list of uses.
3466 void undo() override {
3467 LLVM_DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
3468 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
3469 Inst->setOperand(i: It, Val: OriginalValues[It]);
3470 }
3471 };
3472
3473 /// Build a truncate instruction.
3474 class TruncBuilder : public TypePromotionAction {
3475 Value *Val;
3476
3477 public:
3478 /// Build a truncate instruction of \p Opnd producing a \p Ty
3479 /// result.
3480 /// trunc Opnd to Ty.
3481 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
3482 IRBuilder<> Builder(Opnd);
3483 Builder.SetCurrentDebugLocation(DebugLoc());
3484 Val = Builder.CreateTrunc(V: Opnd, DestTy: Ty, Name: "promoted");
3485 LLVM_DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
3486 }
3487
3488 /// Get the built value.
3489 Value *getBuiltValue() { return Val; }
3490
3491 /// Remove the built instruction.
3492 void undo() override {
3493 LLVM_DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
3494 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3495 IVal->eraseFromParent();
3496 }
3497 };
3498
3499 /// Build a sign extension instruction.
3500 class SExtBuilder : public TypePromotionAction {
3501 Value *Val;
3502
3503 public:
3504 /// Build a sign extension instruction of \p Opnd producing a \p Ty
3505 /// result.
3506 /// sext Opnd to Ty.
3507 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3508 : TypePromotionAction(InsertPt) {
3509 IRBuilder<> Builder(InsertPt);
3510 Val = Builder.CreateSExt(V: Opnd, DestTy: Ty, Name: "promoted");
3511 LLVM_DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
3512 }
3513
3514 /// Get the built value.
3515 Value *getBuiltValue() { return Val; }
3516
3517 /// Remove the built instruction.
3518 void undo() override {
3519 LLVM_DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
3520 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3521 IVal->eraseFromParent();
3522 }
3523 };
3524
3525 /// Build a zero extension instruction.
3526 class ZExtBuilder : public TypePromotionAction {
3527 Value *Val;
3528
3529 public:
3530 /// Build a zero extension instruction of \p Opnd producing a \p Ty
3531 /// result.
3532 /// zext Opnd to Ty.
3533 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3534 : TypePromotionAction(InsertPt) {
3535 IRBuilder<> Builder(InsertPt);
3536 Builder.SetCurrentDebugLocation(DebugLoc());
3537 Val = Builder.CreateZExt(V: Opnd, DestTy: Ty, Name: "promoted");
3538 LLVM_DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
3539 }
3540
3541 /// Get the built value.
3542 Value *getBuiltValue() { return Val; }
3543
3544 /// Remove the built instruction.
3545 void undo() override {
3546 LLVM_DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
3547 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3548 IVal->eraseFromParent();
3549 }
3550 };
3551
3552 /// Mutate an instruction to another type.
3553 class TypeMutator : public TypePromotionAction {
3554 /// Record the original type.
3555 Type *OrigTy;
3556
3557 public:
3558 /// Mutate the type of \p Inst into \p NewTy.
3559 TypeMutator(Instruction *Inst, Type *NewTy)
3560 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
3561 LLVM_DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
3562 << "\n");
3563 Inst->mutateType(Ty: NewTy);
3564 }
3565
3566 /// Mutate the instruction back to its original type.
3567 void undo() override {
3568 LLVM_DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
3569 << "\n");
3570 Inst->mutateType(Ty: OrigTy);
3571 }
3572 };
3573
3574 /// Replace the uses of an instruction by another instruction.
3575 class UsesReplacer : public TypePromotionAction {
3576 /// Helper structure to keep track of the replaced uses.
3577 struct InstructionAndIdx {
3578 /// The instruction using the instruction.
3579 Instruction *Inst;
3580
3581 /// The index where this instruction is used for Inst.
3582 unsigned Idx;
3583
3584 InstructionAndIdx(Instruction *Inst, unsigned Idx)
3585 : Inst(Inst), Idx(Idx) {}
3586 };
3587
3588 /// Keep track of the original uses (pair Instruction, Index).
3589 SmallVector<InstructionAndIdx, 4> OriginalUses;
3590 /// Keep track of the debug users.
3591 SmallVector<DbgVariableRecord *, 1> DbgVariableRecords;
3592
3593 /// Keep track of the new value so that we can undo it by replacing
3594 /// instances of the new value with the original value.
3595 Value *New;
3596
3597 using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
3598
3599 public:
3600 /// Replace all the use of \p Inst by \p New.
3601 UsesReplacer(Instruction *Inst, Value *New)
3602 : TypePromotionAction(Inst), New(New) {
3603 LLVM_DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
3604 << "\n");
3605 // Record the original uses.
3606 for (Use &U : Inst->uses()) {
3607 Instruction *UserI = cast<Instruction>(Val: U.getUser());
3608 OriginalUses.push_back(Elt: InstructionAndIdx(UserI, U.getOperandNo()));
3609 }
3610 // Record the debug uses separately. They are not in the instruction's
3611 // use list, but they are replaced by RAUW.
3612 findDbgValues(V: Inst, DbgVariableRecords);
3613
3614 // Now, we can replace the uses.
3615 Inst->replaceAllUsesWith(V: New);
3616 }
3617
3618 /// Reassign the original uses of Inst to Inst.
3619 void undo() override {
3620 LLVM_DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
3621 for (InstructionAndIdx &Use : OriginalUses)
3622 Use.Inst->setOperand(i: Use.Idx, Val: Inst);
3623 // RAUW has replaced all original uses with references to the new value,
3624 // including the debug uses. Since we are undoing the replacements,
3625 // the original debug uses must also be reinstated to maintain the
3626 // correctness and utility of debug value records.
3627 for (DbgVariableRecord *DVR : DbgVariableRecords)
3628 DVR->replaceVariableLocationOp(OldValue: New, NewValue: Inst);
3629 }
3630 };
3631
3632 /// Remove an instruction from the IR.
3633 class InstructionRemover : public TypePromotionAction {
3634 /// Original position of the instruction.
3635 InsertionHandler Inserter;
3636
3637 /// Helper structure to hide all the link to the instruction. In other
3638 /// words, this helps to do as if the instruction was removed.
3639 OperandsHider Hider;
3640
3641 /// Keep track of the uses replaced, if any.
3642 UsesReplacer *Replacer = nullptr;
3643
3644 /// Keep track of instructions removed.
3645 SetOfInstrs &RemovedInsts;
3646
3647 public:
3648 /// Remove all reference of \p Inst and optionally replace all its
3649 /// uses with New.
3650 /// \p RemovedInsts Keep track of the instructions removed by this Action.
3651 /// \pre If !Inst->use_empty(), then New != nullptr
3652 InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
3653 Value *New = nullptr)
3654 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
3655 RemovedInsts(RemovedInsts) {
3656 if (New)
3657 Replacer = new UsesReplacer(Inst, New);
3658 LLVM_DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
3659 RemovedInsts.insert(Ptr: Inst);
3660 /// The instructions removed here will be freed after completing
3661 /// optimizeBlock() for all blocks as we need to keep track of the
3662 /// removed instructions during promotion.
3663 Inst->removeFromParent();
3664 }
3665
3666 ~InstructionRemover() override { delete Replacer; }
3667
3668 InstructionRemover &operator=(const InstructionRemover &other) = delete;
3669 InstructionRemover(const InstructionRemover &other) = delete;
3670
3671 /// Resurrect the instruction and reassign it to the proper uses if
3672 /// new value was provided when build this action.
3673 void undo() override {
3674 LLVM_DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
3675 Inserter.insert(Inst);
3676 if (Replacer)
3677 Replacer->undo();
3678 Hider.undo();
3679 RemovedInsts.erase(Ptr: Inst);
3680 }
3681 };
3682
3683public:
3684 /// Restoration point.
3685 /// The restoration point is a pointer to an action instead of an iterator
3686 /// because the iterator may be invalidated but not the pointer.
3687 using ConstRestorationPt = const TypePromotionAction *;
3688
3689 TypePromotionTransaction(SetOfInstrs &RemovedInsts)
3690 : RemovedInsts(RemovedInsts) {}
3691
3692 /// Advocate every changes made in that transaction. Return true if any change
3693 /// happen.
3694 bool commit();
3695
3696 /// Undo all the changes made after the given point.
3697 void rollback(ConstRestorationPt Point);
3698
3699 /// Get the current restoration point.
3700 ConstRestorationPt getRestorationPoint() const;
3701
3702 /// \name API for IR modification with state keeping to support rollback.
3703 /// @{
3704 /// Same as Instruction::setOperand.
3705 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
3706
3707 /// Same as Instruction::eraseFromParent.
3708 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
3709
3710 /// Same as Value::replaceAllUsesWith.
3711 void replaceAllUsesWith(Instruction *Inst, Value *New);
3712
3713 /// Same as Value::mutateType.
3714 void mutateType(Instruction *Inst, Type *NewTy);
3715
3716 /// Same as IRBuilder::createTrunc.
3717 Value *createTrunc(Instruction *Opnd, Type *Ty);
3718
3719 /// Same as IRBuilder::createSExt.
3720 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
3721
3722 /// Same as IRBuilder::createZExt.
3723 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
3724
3725private:
3726 /// The ordered list of actions made so far.
3727 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
3728
3729 using CommitPt =
3730 SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
3731
3732 SetOfInstrs &RemovedInsts;
3733};
3734
3735} // end anonymous namespace
3736
3737void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
3738 Value *NewVal) {
3739 Actions.push_back(Elt: std::make_unique<TypePromotionTransaction::OperandSetter>(
3740 args&: Inst, args&: Idx, args&: NewVal));
3741}
3742
3743void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
3744 Value *NewVal) {
3745 Actions.push_back(
3746 Elt: std::make_unique<TypePromotionTransaction::InstructionRemover>(
3747 args&: Inst, args&: RemovedInsts, args&: NewVal));
3748}
3749
3750void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
3751 Value *New) {
3752 Actions.push_back(
3753 Elt: std::make_unique<TypePromotionTransaction::UsesReplacer>(args&: Inst, args&: New));
3754}
3755
3756void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
3757 Actions.push_back(
3758 Elt: std::make_unique<TypePromotionTransaction::TypeMutator>(args&: Inst, args&: NewTy));
3759}
3760
3761Value *TypePromotionTransaction::createTrunc(Instruction *Opnd, Type *Ty) {
3762 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
3763 Value *Val = Ptr->getBuiltValue();
3764 Actions.push_back(Elt: std::move(Ptr));
3765 return Val;
3766}
3767
3768Value *TypePromotionTransaction::createSExt(Instruction *Inst, Value *Opnd,
3769 Type *Ty) {
3770 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
3771 Value *Val = Ptr->getBuiltValue();
3772 Actions.push_back(Elt: std::move(Ptr));
3773 return Val;
3774}
3775
3776Value *TypePromotionTransaction::createZExt(Instruction *Inst, Value *Opnd,
3777 Type *Ty) {
3778 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
3779 Value *Val = Ptr->getBuiltValue();
3780 Actions.push_back(Elt: std::move(Ptr));
3781 return Val;
3782}
3783
3784TypePromotionTransaction::ConstRestorationPt
3785TypePromotionTransaction::getRestorationPoint() const {
3786 return !Actions.empty() ? Actions.back().get() : nullptr;
3787}
3788
3789bool TypePromotionTransaction::commit() {
3790 for (std::unique_ptr<TypePromotionAction> &Action : Actions)
3791 Action->commit();
3792 bool Modified = !Actions.empty();
3793 Actions.clear();
3794 return Modified;
3795}
3796
3797void TypePromotionTransaction::rollback(
3798 TypePromotionTransaction::ConstRestorationPt Point) {
3799 while (!Actions.empty() && Point != Actions.back().get()) {
3800 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
3801 Curr->undo();
3802 }
3803}
3804
3805namespace {
3806
3807/// A helper class for matching addressing modes.
3808///
3809/// This encapsulates the logic for matching the target-legal addressing modes.
3810class AddressingModeMatcher {
3811 SmallVectorImpl<Instruction *> &AddrModeInsts;
3812 const TargetLowering &TLI;
3813 const TargetRegisterInfo &TRI;
3814 const DataLayout &DL;
3815 const LoopInfo &LI;
3816 const std::function<const DominatorTree &()> getDTFn;
3817
3818 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
3819 /// the memory instruction that we're computing this address for.
3820 Type *AccessTy;
3821 unsigned AddrSpace;
3822 Instruction *MemoryInst;
3823
3824 /// This is the addressing mode that we're building up. This is
3825 /// part of the return value of this addressing mode matching stuff.
3826 ExtAddrMode &AddrMode;
3827
3828 /// The instructions inserted by other CodeGenPrepare optimizations.
3829 const SetOfInstrs &InsertedInsts;
3830
3831 /// A map from the instructions to their type before promotion.
3832 InstrToOrigTy &PromotedInsts;
3833
3834 /// The ongoing transaction where every action should be registered.
3835 TypePromotionTransaction &TPT;
3836
3837 // A GEP which has too large offset to be folded into the addressing mode.
3838 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP;
3839
3840 /// This is set to true when we should not do profitability checks.
3841 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
3842 bool IgnoreProfitability;
3843
3844 /// True if we are optimizing for size.
3845 bool OptSize = false;
3846
3847 ProfileSummaryInfo *PSI;
3848 BlockFrequencyInfo *BFI;
3849
3850 AddressingModeMatcher(
3851 SmallVectorImpl<Instruction *> &AMI, const TargetLowering &TLI,
3852 const TargetRegisterInfo &TRI, const LoopInfo &LI,
3853 const std::function<const DominatorTree &()> getDTFn, Type *AT,
3854 unsigned AS, Instruction *MI, ExtAddrMode &AM,
3855 const SetOfInstrs &InsertedInsts, InstrToOrigTy &PromotedInsts,
3856 TypePromotionTransaction &TPT,
3857 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP,
3858 bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI)
3859 : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
3860 DL(MI->getDataLayout()), LI(LI), getDTFn(getDTFn),
3861 AccessTy(AT), AddrSpace(AS), MemoryInst(MI), AddrMode(AM),
3862 InsertedInsts(InsertedInsts), PromotedInsts(PromotedInsts), TPT(TPT),
3863 LargeOffsetGEP(LargeOffsetGEP), OptSize(OptSize), PSI(PSI), BFI(BFI) {
3864 IgnoreProfitability = false;
3865 }
3866
3867public:
3868 /// Find the maximal addressing mode that a load/store of V can fold,
3869 /// give an access type of AccessTy. This returns a list of involved
3870 /// instructions in AddrModeInsts.
3871 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
3872 /// optimizations.
3873 /// \p PromotedInsts maps the instructions to their type before promotion.
3874 /// \p The ongoing transaction where every action should be registered.
3875 static ExtAddrMode
3876 Match(Value *V, Type *AccessTy, unsigned AS, Instruction *MemoryInst,
3877 SmallVectorImpl<Instruction *> &AddrModeInsts,
3878 const TargetLowering &TLI, const LoopInfo &LI,
3879 const std::function<const DominatorTree &()> getDTFn,
3880 const TargetRegisterInfo &TRI, const SetOfInstrs &InsertedInsts,
3881 InstrToOrigTy &PromotedInsts, TypePromotionTransaction &TPT,
3882 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP,
3883 bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) {
3884 ExtAddrMode Result;
3885
3886 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI, LI, getDTFn,
3887 AccessTy, AS, MemoryInst, Result,
3888 InsertedInsts, PromotedInsts, TPT,
3889 LargeOffsetGEP, OptSize, PSI, BFI)
3890 .matchAddr(Addr: V, Depth: 0);
3891 (void)Success;
3892 assert(Success && "Couldn't select *anything*?");
3893 return Result;
3894 }
3895
3896private:
3897 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
3898 bool matchAddr(Value *Addr, unsigned Depth);
3899 bool matchOperationAddr(User *AddrInst, unsigned Opcode, unsigned Depth,
3900 bool *MovedAway = nullptr);
3901 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
3902 ExtAddrMode &AMBefore,
3903 ExtAddrMode &AMAfter);
3904 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
3905 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
3906 Value *PromotedOperand) const;
3907};
3908
3909class PhiNodeSet;
3910
3911/// An iterator for PhiNodeSet.
3912class PhiNodeSetIterator {
3913 PhiNodeSet *const Set;
3914 size_t CurrentIndex = 0;
3915
3916public:
3917 /// The constructor. Start should point to either a valid element, or be equal
3918 /// to the size of the underlying SmallVector of the PhiNodeSet.
3919 PhiNodeSetIterator(PhiNodeSet *const Set, size_t Start);
3920 PHINode *operator*() const;
3921 PhiNodeSetIterator &operator++();
3922 bool operator==(const PhiNodeSetIterator &RHS) const;
3923 bool operator!=(const PhiNodeSetIterator &RHS) const;
3924};
3925
3926/// Keeps a set of PHINodes.
3927///
3928/// This is a minimal set implementation for a specific use case:
3929/// It is very fast when there are very few elements, but also provides good
3930/// performance when there are many. It is similar to SmallPtrSet, but also
3931/// provides iteration by insertion order, which is deterministic and stable
3932/// across runs. It is also similar to SmallSetVector, but provides removing
3933/// elements in O(1) time. This is achieved by not actually removing the element
3934/// from the underlying vector, so comes at the cost of using more memory, but
3935/// that is fine, since PhiNodeSets are used as short lived objects.
3936class PhiNodeSet {
3937 friend class PhiNodeSetIterator;
3938
3939 using MapType = SmallDenseMap<PHINode *, size_t, 32>;
3940 using iterator = PhiNodeSetIterator;
3941
3942 /// Keeps the elements in the order of their insertion in the underlying
3943 /// vector. To achieve constant time removal, it never deletes any element.
3944 SmallVector<PHINode *, 32> NodeList;
3945
3946 /// Keeps the elements in the underlying set implementation. This (and not the
3947 /// NodeList defined above) is the source of truth on whether an element
3948 /// is actually in the collection.
3949 MapType NodeMap;
3950
3951 /// Points to the first valid (not deleted) element when the set is not empty
3952 /// and the value is not zero. Equals to the size of the underlying vector
3953 /// when the set is empty. When the value is 0, as in the beginning, the
3954 /// first element may or may not be valid.
3955 size_t FirstValidElement = 0;
3956
3957public:
3958 /// Inserts a new element to the collection.
3959 /// \returns true if the element is actually added, i.e. was not in the
3960 /// collection before the operation.
3961 bool insert(PHINode *Ptr) {
3962 if (NodeMap.insert(KV: std::make_pair(x&: Ptr, y: NodeList.size())).second) {
3963 NodeList.push_back(Elt: Ptr);
3964 return true;
3965 }
3966 return false;
3967 }
3968
3969 /// Removes the element from the collection.
3970 /// \returns whether the element is actually removed, i.e. was in the
3971 /// collection before the operation.
3972 bool erase(PHINode *Ptr) {
3973 if (NodeMap.erase(Val: Ptr)) {
3974 SkipRemovedElements(CurrentIndex&: FirstValidElement);
3975 return true;
3976 }
3977 return false;
3978 }
3979
3980 /// Removes all elements and clears the collection.
3981 void clear() {
3982 NodeMap.clear();
3983 NodeList.clear();
3984 FirstValidElement = 0;
3985 }
3986
3987 /// \returns an iterator that will iterate the elements in the order of
3988 /// insertion.
3989 iterator begin() {
3990 if (FirstValidElement == 0)
3991 SkipRemovedElements(CurrentIndex&: FirstValidElement);
3992 return PhiNodeSetIterator(this, FirstValidElement);
3993 }
3994
3995 /// \returns an iterator that points to the end of the collection.
3996 iterator end() { return PhiNodeSetIterator(this, NodeList.size()); }
3997
3998 /// Returns the number of elements in the collection.
3999 size_t size() const { return NodeMap.size(); }
4000
4001 /// \returns 1 if the given element is in the collection, and 0 if otherwise.
4002 size_t count(PHINode *Ptr) const { return NodeMap.count(Val: Ptr); }
4003
4004private:
4005 /// Updates the CurrentIndex so that it will point to a valid element.
4006 ///
4007 /// If the element of NodeList at CurrentIndex is valid, it does not
4008 /// change it. If there are no more valid elements, it updates CurrentIndex
4009 /// to point to the end of the NodeList.
4010 void SkipRemovedElements(size_t &CurrentIndex) {
4011 while (CurrentIndex < NodeList.size()) {
4012 auto it = NodeMap.find(Val: NodeList[CurrentIndex]);
4013 // If the element has been deleted and added again later, NodeMap will
4014 // point to a different index, so CurrentIndex will still be invalid.
4015 if (it != NodeMap.end() && it->second == CurrentIndex)
4016 break;
4017 ++CurrentIndex;
4018 }
4019 }
4020};
4021
4022PhiNodeSetIterator::PhiNodeSetIterator(PhiNodeSet *const Set, size_t Start)
4023 : Set(Set), CurrentIndex(Start) {}
4024
4025PHINode *PhiNodeSetIterator::operator*() const {
4026 assert(CurrentIndex < Set->NodeList.size() &&
4027 "PhiNodeSet access out of range");
4028 return Set->NodeList[CurrentIndex];
4029}
4030
4031PhiNodeSetIterator &PhiNodeSetIterator::operator++() {
4032 assert(CurrentIndex < Set->NodeList.size() &&
4033 "PhiNodeSet access out of range");
4034 ++CurrentIndex;
4035 Set->SkipRemovedElements(CurrentIndex);
4036 return *this;
4037}
4038
4039bool PhiNodeSetIterator::operator==(const PhiNodeSetIterator &RHS) const {
4040 return CurrentIndex == RHS.CurrentIndex;
4041}
4042
4043bool PhiNodeSetIterator::operator!=(const PhiNodeSetIterator &RHS) const {
4044 return !((*this) == RHS);
4045}
4046
4047/// Keep track of simplification of Phi nodes.
4048/// Accept the set of all phi nodes and erase phi node from this set
4049/// if it is simplified.
4050class SimplificationTracker {
4051 DenseMap<Value *, Value *> Storage;
4052 // Tracks newly created Phi nodes. The elements are iterated by insertion
4053 // order.
4054 PhiNodeSet AllPhiNodes;
4055 // Tracks newly created Select nodes.
4056 SmallPtrSet<SelectInst *, 32> AllSelectNodes;
4057
4058public:
4059 Value *Get(Value *V) {
4060 do {
4061 auto SV = Storage.find(Val: V);
4062 if (SV == Storage.end())
4063 return V;
4064 V = SV->second;
4065 } while (true);
4066 }
4067
4068 void Put(Value *From, Value *To) { Storage.insert(KV: {From, To}); }
4069
4070 void ReplacePhi(PHINode *From, PHINode *To) {
4071 Value *OldReplacement = Get(V: From);
4072 while (OldReplacement != From) {
4073 From = To;
4074 To = dyn_cast<PHINode>(Val: OldReplacement);
4075 OldReplacement = Get(V: From);
4076 }
4077 assert(To && Get(To) == To && "Replacement PHI node is already replaced.");
4078 Put(From, To);
4079 From->replaceAllUsesWith(V: To);
4080 AllPhiNodes.erase(Ptr: From);
4081 From->eraseFromParent();
4082 }
4083
4084 PhiNodeSet &newPhiNodes() { return AllPhiNodes; }
4085
4086 void insertNewPhi(PHINode *PN) { AllPhiNodes.insert(Ptr: PN); }
4087
4088 void insertNewSelect(SelectInst *SI) { AllSelectNodes.insert(Ptr: SI); }
4089
4090 unsigned countNewPhiNodes() const { return AllPhiNodes.size(); }
4091
4092 unsigned countNewSelectNodes() const { return AllSelectNodes.size(); }
4093
4094 void destroyNewNodes(Type *CommonType) {
4095 // For safe erasing, replace the uses with dummy value first.
4096 auto *Dummy = PoisonValue::get(T: CommonType);
4097 for (auto *I : AllPhiNodes) {
4098 I->replaceAllUsesWith(V: Dummy);
4099 I->eraseFromParent();
4100 }
4101 AllPhiNodes.clear();
4102 for (auto *I : AllSelectNodes) {
4103 I->replaceAllUsesWith(V: Dummy);
4104 I->eraseFromParent();
4105 }
4106 AllSelectNodes.clear();
4107 }
4108};
4109
4110/// A helper class for combining addressing modes.
4111class AddressingModeCombiner {
4112 typedef DenseMap<Value *, Value *> FoldAddrToValueMapping;
4113 typedef std::pair<PHINode *, PHINode *> PHIPair;
4114
4115private:
4116 /// The addressing modes we've collected.
4117 SmallVector<ExtAddrMode, 16> AddrModes;
4118
4119 /// The field in which the AddrModes differ, when we have more than one.
4120 ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
4121
4122 /// Are the AddrModes that we have all just equal to their original values?
4123 bool AllAddrModesTrivial = true;
4124
4125 /// Common Type for all different fields in addressing modes.
4126 Type *CommonType = nullptr;
4127
4128 const DataLayout &DL;
4129
4130 /// Original Address.
4131 Value *Original;
4132
4133 /// Common value among addresses
4134 Value *CommonValue = nullptr;
4135
4136public:
4137 AddressingModeCombiner(const DataLayout &DL, Value *OriginalValue)
4138 : DL(DL), Original(OriginalValue) {}
4139
4140 ~AddressingModeCombiner() { eraseCommonValueIfDead(); }
4141
4142 /// Get the combined AddrMode
4143 const ExtAddrMode &getAddrMode() const { return AddrModes[0]; }
4144
4145 /// Add a new AddrMode if it's compatible with the AddrModes we already
4146 /// have.
4147 /// \return True iff we succeeded in doing so.
4148 bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
4149 // Take note of if we have any non-trivial AddrModes, as we need to detect
4150 // when all AddrModes are trivial as then we would introduce a phi or select
4151 // which just duplicates what's already there.
4152 AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
4153
4154 // If this is the first addrmode then everything is fine.
4155 if (AddrModes.empty()) {
4156 AddrModes.emplace_back(Args&: NewAddrMode);
4157 return true;
4158 }
4159
4160 // Figure out how different this is from the other address modes, which we
4161 // can do just by comparing against the first one given that we only care
4162 // about the cumulative difference.
4163 ExtAddrMode::FieldName ThisDifferentField =
4164 AddrModes[0].compare(other: NewAddrMode);
4165 if (DifferentField == ExtAddrMode::NoField)
4166 DifferentField = ThisDifferentField;
4167 else if (DifferentField != ThisDifferentField)
4168 DifferentField = ExtAddrMode::MultipleFields;
4169
4170 // If NewAddrMode differs in more than one dimension we cannot handle it.
4171 bool CanHandle = DifferentField != ExtAddrMode::MultipleFields;
4172
4173 // If Scale Field is different then we reject.
4174 CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField;
4175
4176 // We also must reject the case when base offset is different and
4177 // scale reg is not null, we cannot handle this case due to merge of
4178 // different offsets will be used as ScaleReg.
4179 CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField ||
4180 !NewAddrMode.ScaledReg);
4181
4182 // We also must reject the case when GV is different and BaseReg installed
4183 // due to we want to use base reg as a merge of GV values.
4184 CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField ||
4185 !NewAddrMode.HasBaseReg);
4186
4187 // Even if NewAddMode is the same we still need to collect it due to
4188 // original value is different. And later we will need all original values
4189 // as anchors during finding the common Phi node.
4190 if (CanHandle)
4191 AddrModes.emplace_back(Args&: NewAddrMode);
4192 else
4193 AddrModes.clear();
4194
4195 return CanHandle;
4196 }
4197
4198 /// Combine the addressing modes we've collected into a single
4199 /// addressing mode.
4200 /// \return True iff we successfully combined them or we only had one so
4201 /// didn't need to combine them anyway.
4202 bool combineAddrModes() {
4203 // If we have no AddrModes then they can't be combined.
4204 if (AddrModes.size() == 0)
4205 return false;
4206
4207 // A single AddrMode can trivially be combined.
4208 if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
4209 return true;
4210
4211 // If the AddrModes we collected are all just equal to the value they are
4212 // derived from then combining them wouldn't do anything useful.
4213 if (AllAddrModesTrivial)
4214 return false;
4215
4216 if (!addrModeCombiningAllowed())
4217 return false;
4218
4219 // Build a map between <original value, basic block where we saw it> to
4220 // value of base register.
4221 // Bail out if there is no common type.
4222 FoldAddrToValueMapping Map;
4223 if (!initializeMap(Map))
4224 return false;
4225
4226 CommonValue = findCommon(Map);
4227 if (CommonValue)
4228 AddrModes[0].SetCombinedField(Field: DifferentField, V: CommonValue, AddrModes);
4229 return CommonValue != nullptr;
4230 }
4231
4232private:
4233 /// `CommonValue` may be a placeholder inserted by us.
4234 /// If the placeholder is not used, we should remove this dead instruction.
4235 void eraseCommonValueIfDead() {
4236 if (CommonValue && CommonValue->use_empty())
4237 if (Instruction *CommonInst = dyn_cast<Instruction>(Val: CommonValue))
4238 CommonInst->eraseFromParent();
4239 }
4240
4241 /// Initialize Map with anchor values. For address seen
4242 /// we set the value of different field saw in this address.
4243 /// At the same time we find a common type for different field we will
4244 /// use to create new Phi/Select nodes. Keep it in CommonType field.
4245 /// Return false if there is no common type found.
4246 bool initializeMap(FoldAddrToValueMapping &Map) {
4247 // Keep track of keys where the value is null. We will need to replace it
4248 // with constant null when we know the common type.
4249 SmallVector<Value *, 2> NullValue;
4250 Type *IntPtrTy = DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
4251 for (auto &AM : AddrModes) {
4252 Value *DV = AM.GetFieldAsValue(Field: DifferentField, IntPtrTy);
4253 if (DV) {
4254 auto *Type = DV->getType();
4255 if (CommonType && CommonType != Type)
4256 return false;
4257 CommonType = Type;
4258 Map[AM.OriginalValue] = DV;
4259 } else {
4260 NullValue.push_back(Elt: AM.OriginalValue);
4261 }
4262 }
4263 assert(CommonType && "At least one non-null value must be!");
4264 for (auto *V : NullValue)
4265 Map[V] = Constant::getNullValue(Ty: CommonType);
4266 return true;
4267 }
4268
4269 /// We have mapping between value A and other value B where B was a field in
4270 /// addressing mode represented by A. Also we have an original value C
4271 /// representing an address we start with. Traversing from C through phi and
4272 /// selects we ended up with A's in a map. This utility function tries to find
4273 /// a value V which is a field in addressing mode C and traversing through phi
4274 /// nodes and selects we will end up in corresponded values B in a map.
4275 /// The utility will create a new Phi/Selects if needed.
4276 // The simple example looks as follows:
4277 // BB1:
4278 // p1 = b1 + 40
4279 // br cond BB2, BB3
4280 // BB2:
4281 // p2 = b2 + 40
4282 // br BB3
4283 // BB3:
4284 // p = phi [p1, BB1], [p2, BB2]
4285 // v = load p
4286 // Map is
4287 // p1 -> b1
4288 // p2 -> b2
4289 // Request is
4290 // p -> ?
4291 // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3.
4292 Value *findCommon(FoldAddrToValueMapping &Map) {
4293 // Tracks the simplification of newly created phi nodes. The reason we use
4294 // this mapping is because we will add new created Phi nodes in AddrToBase.
4295 // Simplification of Phi nodes is recursive, so some Phi node may
4296 // be simplified after we added it to AddrToBase. In reality this
4297 // simplification is possible only if original phi/selects were not
4298 // simplified yet.
4299 // Using this mapping we can find the current value in AddrToBase.
4300 SimplificationTracker ST;
4301
4302 // First step, DFS to create PHI nodes for all intermediate blocks.
4303 // Also fill traverse order for the second step.
4304 SmallVector<Value *, 32> TraverseOrder;
4305 InsertPlaceholders(Map, TraverseOrder, ST);
4306
4307 // Second Step, fill new nodes by merged values and simplify if possible.
4308 FillPlaceholders(Map, TraverseOrder, ST);
4309
4310 if (!AddrSinkNewSelects && ST.countNewSelectNodes() > 0) {
4311 ST.destroyNewNodes(CommonType);
4312 return nullptr;
4313 }
4314
4315 // Now we'd like to match New Phi nodes to existed ones.
4316 unsigned PhiNotMatchedCount = 0;
4317 if (!MatchPhiSet(ST, AllowNewPhiNodes: AddrSinkNewPhis, PhiNotMatchedCount)) {
4318 ST.destroyNewNodes(CommonType);
4319 return nullptr;
4320 }
4321
4322 auto *Result = ST.Get(V: Map.find(Val: Original)->second);
4323 if (Result) {
4324 NumMemoryInstsPhiCreated += ST.countNewPhiNodes() + PhiNotMatchedCount;
4325 NumMemoryInstsSelectCreated += ST.countNewSelectNodes();
4326 }
4327 return Result;
4328 }
4329
4330 /// Try to match PHI node to Candidate.
4331 /// Matcher tracks the matched Phi nodes.
4332 bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
4333 SmallSetVector<PHIPair, 8> &Matcher,
4334 PhiNodeSet &PhiNodesToMatch) {
4335 SmallVector<PHIPair, 8> WorkList;
4336 Matcher.insert(X: {PHI, Candidate});
4337 SmallPtrSet<PHINode *, 8> MatchedPHIs;
4338 MatchedPHIs.insert(Ptr: PHI);
4339 WorkList.push_back(Elt: {PHI, Candidate});
4340 SmallSet<PHIPair, 8> Visited;
4341 while (!WorkList.empty()) {
4342 auto Item = WorkList.pop_back_val();
4343 if (!Visited.insert(V: Item).second)
4344 continue;
4345 // We iterate over all incoming values to Phi to compare them.
4346 // If values are different and both of them Phi and the first one is a
4347 // Phi we added (subject to match) and both of them is in the same basic
4348 // block then we can match our pair if values match. So we state that
4349 // these values match and add it to work list to verify that.
4350 for (auto *B : Item.first->blocks()) {
4351 Value *FirstValue = Item.first->getIncomingValueForBlock(BB: B);
4352 Value *SecondValue = Item.second->getIncomingValueForBlock(BB: B);
4353 if (FirstValue == SecondValue)
4354 continue;
4355
4356 PHINode *FirstPhi = dyn_cast<PHINode>(Val: FirstValue);
4357 PHINode *SecondPhi = dyn_cast<PHINode>(Val: SecondValue);
4358
4359 // One of them is not Phi or
4360 // The first one is not Phi node from the set we'd like to match or
4361 // Phi nodes from different basic blocks then
4362 // we will not be able to match.
4363 if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(Ptr: FirstPhi) ||
4364 FirstPhi->getParent() != SecondPhi->getParent())
4365 return false;
4366
4367 // If we already matched them then continue.
4368 if (Matcher.count(key: {FirstPhi, SecondPhi}))
4369 continue;
4370 // So the values are different and does not match. So we need them to
4371 // match. (But we register no more than one match per PHI node, so that
4372 // we won't later try to replace them twice.)
4373 if (MatchedPHIs.insert(Ptr: FirstPhi).second)
4374 Matcher.insert(X: {FirstPhi, SecondPhi});
4375 // But me must check it.
4376 WorkList.push_back(Elt: {FirstPhi, SecondPhi});
4377 }
4378 }
4379 return true;
4380 }
4381
4382 /// For the given set of PHI nodes (in the SimplificationTracker) try
4383 /// to find their equivalents.
4384 /// Returns false if this matching fails and creation of new Phi is disabled.
4385 bool MatchPhiSet(SimplificationTracker &ST, bool AllowNewPhiNodes,
4386 unsigned &PhiNotMatchedCount) {
4387 // Matched and PhiNodesToMatch iterate their elements in a deterministic
4388 // order, so the replacements (ReplacePhi) are also done in a deterministic
4389 // order.
4390 SmallSetVector<PHIPair, 8> Matched;
4391 SmallPtrSet<PHINode *, 8> WillNotMatch;
4392 PhiNodeSet &PhiNodesToMatch = ST.newPhiNodes();
4393 while (PhiNodesToMatch.size()) {
4394 PHINode *PHI = *PhiNodesToMatch.begin();
4395
4396 // Add us, if no Phi nodes in the basic block we do not match.
4397 WillNotMatch.clear();
4398 WillNotMatch.insert(Ptr: PHI);
4399
4400 // Traverse all Phis until we found equivalent or fail to do that.
4401 bool IsMatched = false;
4402 for (auto &P : PHI->getParent()->phis()) {
4403 // Skip new Phi nodes.
4404 if (PhiNodesToMatch.count(Ptr: &P))
4405 continue;
4406 if ((IsMatched = MatchPhiNode(PHI, Candidate: &P, Matcher&: Matched, PhiNodesToMatch)))
4407 break;
4408 // If it does not match, collect all Phi nodes from matcher.
4409 // if we end up with no match, them all these Phi nodes will not match
4410 // later.
4411 WillNotMatch.insert_range(R: llvm::make_first_range(c&: Matched));
4412 Matched.clear();
4413 }
4414 if (IsMatched) {
4415 // Replace all matched values and erase them.
4416 for (auto MV : Matched)
4417 ST.ReplacePhi(From: MV.first, To: MV.second);
4418 Matched.clear();
4419 continue;
4420 }
4421 // If we are not allowed to create new nodes then bail out.
4422 if (!AllowNewPhiNodes)
4423 return false;
4424 // Just remove all seen values in matcher. They will not match anything.
4425 PhiNotMatchedCount += WillNotMatch.size();
4426 for (auto *P : WillNotMatch)
4427 PhiNodesToMatch.erase(Ptr: P);
4428 }
4429 return true;
4430 }
4431 /// Fill the placeholders with values from predecessors and simplify them.
4432 void FillPlaceholders(FoldAddrToValueMapping &Map,
4433 SmallVectorImpl<Value *> &TraverseOrder,
4434 SimplificationTracker &ST) {
4435 while (!TraverseOrder.empty()) {
4436 Value *Current = TraverseOrder.pop_back_val();
4437 assert(Map.contains(Current) && "No node to fill!!!");
4438 Value *V = Map[Current];
4439
4440 if (SelectInst *Select = dyn_cast<SelectInst>(Val: V)) {
4441 // CurrentValue also must be Select.
4442 auto *CurrentSelect = cast<SelectInst>(Val: Current);
4443 auto *TrueValue = CurrentSelect->getTrueValue();
4444 assert(Map.contains(TrueValue) && "No True Value!");
4445 Select->setTrueValue(ST.Get(V: Map[TrueValue]));
4446 auto *FalseValue = CurrentSelect->getFalseValue();
4447 assert(Map.contains(FalseValue) && "No False Value!");
4448 Select->setFalseValue(ST.Get(V: Map[FalseValue]));
4449 } else {
4450 // Must be a Phi node then.
4451 auto *PHI = cast<PHINode>(Val: V);
4452 // Fill the Phi node with values from predecessors.
4453 for (auto *B : predecessors(BB: PHI->getParent())) {
4454 Value *PV = cast<PHINode>(Val: Current)->getIncomingValueForBlock(BB: B);
4455 assert(Map.contains(PV) && "No predecessor Value!");
4456 PHI->addIncoming(V: ST.Get(V: Map[PV]), BB: B);
4457 }
4458 }
4459 }
4460 }
4461
4462 /// Starting from original value recursively iterates over def-use chain up to
4463 /// known ending values represented in a map. For each traversed phi/select
4464 /// inserts a placeholder Phi or Select.
4465 /// Reports all new created Phi/Select nodes by adding them to set.
4466 /// Also reports and order in what values have been traversed.
4467 void InsertPlaceholders(FoldAddrToValueMapping &Map,
4468 SmallVectorImpl<Value *> &TraverseOrder,
4469 SimplificationTracker &ST) {
4470 SmallVector<Value *, 32> Worklist;
4471 assert((isa<PHINode>(Original) || isa<SelectInst>(Original)) &&
4472 "Address must be a Phi or Select node");
4473 auto *Dummy = PoisonValue::get(T: CommonType);
4474 Worklist.push_back(Elt: Original);
4475 while (!Worklist.empty()) {
4476 Value *Current = Worklist.pop_back_val();
4477 // if it is already visited or it is an ending value then skip it.
4478 if (Map.contains(Val: Current))
4479 continue;
4480 TraverseOrder.push_back(Elt: Current);
4481
4482 // CurrentValue must be a Phi node or select. All others must be covered
4483 // by anchors.
4484 if (SelectInst *CurrentSelect = dyn_cast<SelectInst>(Val: Current)) {
4485 // Is it OK to get metadata from OrigSelect?!
4486 // Create a Select placeholder with dummy value.
4487 SelectInst *Select =
4488 SelectInst::Create(C: CurrentSelect->getCondition(), S1: Dummy, S2: Dummy,
4489 NameStr: CurrentSelect->getName(),
4490 InsertBefore: CurrentSelect->getIterator(), MDFrom: CurrentSelect);
4491 Map[Current] = Select;
4492 ST.insertNewSelect(SI: Select);
4493 // We are interested in True and False values.
4494 Worklist.push_back(Elt: CurrentSelect->getTrueValue());
4495 Worklist.push_back(Elt: CurrentSelect->getFalseValue());
4496 } else {
4497 // It must be a Phi node then.
4498 PHINode *CurrentPhi = cast<PHINode>(Val: Current);
4499 unsigned PredCount = CurrentPhi->getNumIncomingValues();
4500 PHINode *PHI =
4501 PHINode::Create(Ty: CommonType, NumReservedValues: PredCount, NameStr: "sunk_phi", InsertBefore: CurrentPhi->getIterator());
4502 Map[Current] = PHI;
4503 ST.insertNewPhi(PN: PHI);
4504 append_range(C&: Worklist, R: CurrentPhi->incoming_values());
4505 }
4506 }
4507 }
4508
4509 bool addrModeCombiningAllowed() {
4510 if (DisableComplexAddrModes)
4511 return false;
4512 switch (DifferentField) {
4513 default:
4514 return false;
4515 case ExtAddrMode::BaseRegField:
4516 return AddrSinkCombineBaseReg;
4517 case ExtAddrMode::BaseGVField:
4518 return AddrSinkCombineBaseGV;
4519 case ExtAddrMode::BaseOffsField:
4520 return AddrSinkCombineBaseOffs;
4521 case ExtAddrMode::ScaledRegField:
4522 return AddrSinkCombineScaledReg;
4523 }
4524 }
4525};
4526} // end anonymous namespace
4527
4528/// Try adding ScaleReg*Scale to the current addressing mode.
4529/// Return true and update AddrMode if this addr mode is legal for the target,
4530/// false if not.
4531bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
4532 unsigned Depth) {
4533 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
4534 // mode. Just process that directly.
4535 if (Scale == 1)
4536 return matchAddr(Addr: ScaleReg, Depth);
4537
4538 // If the scale is 0, it takes nothing to add this.
4539 if (Scale == 0)
4540 return true;
4541
4542 // If we already have a scale of this value, we can add to it, otherwise, we
4543 // need an available scale field.
4544 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
4545 return false;
4546
4547 ExtAddrMode TestAddrMode = AddrMode;
4548
4549 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
4550 // [A+B + A*7] -> [B+A*8].
4551 TestAddrMode.Scale += Scale;
4552 TestAddrMode.ScaledReg = ScaleReg;
4553
4554 // If the new address isn't legal, bail out.
4555 if (!TLI.isLegalAddressingMode(DL, AM: TestAddrMode, Ty: AccessTy, AddrSpace))
4556 return false;
4557
4558 // It was legal, so commit it.
4559 AddrMode = TestAddrMode;
4560
4561 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
4562 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
4563 // X*Scale + C*Scale to addr mode. If we found available IV increment, do not
4564 // go any further: we can reuse it and cannot eliminate it.
4565 ConstantInt *CI = nullptr;
4566 Value *AddLHS = nullptr;
4567 if (isa<Instruction>(Val: ScaleReg) && // not a constant expr.
4568 match(V: ScaleReg, P: m_Add(L: m_Value(V&: AddLHS), R: m_ConstantInt(CI))) &&
4569 !isIVIncrement(V: ScaleReg, LI: &LI) && CI->getValue().isSignedIntN(N: 64)) {
4570 TestAddrMode.InBounds = false;
4571 TestAddrMode.ScaledReg = AddLHS;
4572 TestAddrMode.BaseOffs += CI->getSExtValue() * TestAddrMode.Scale;
4573
4574 // If this addressing mode is legal, commit it and remember that we folded
4575 // this instruction.
4576 if (TLI.isLegalAddressingMode(DL, AM: TestAddrMode, Ty: AccessTy, AddrSpace)) {
4577 AddrModeInsts.push_back(Elt: cast<Instruction>(Val: ScaleReg));
4578 AddrMode = TestAddrMode;
4579 return true;
4580 }
4581 // Restore status quo.
4582 TestAddrMode = AddrMode;
4583 }
4584
4585 // If this is an add recurrence with a constant step, return the increment
4586 // instruction and the canonicalized step.
4587 auto GetConstantStep =
4588 [this](const Value *V) -> std::optional<std::pair<Instruction *, APInt>> {
4589 auto *PN = dyn_cast<PHINode>(Val: V);
4590 if (!PN)
4591 return std::nullopt;
4592 auto IVInc = getIVIncrement(PN, LI: &LI);
4593 if (!IVInc)
4594 return std::nullopt;
4595 // TODO: The result of the intrinsics above is two-complement. However when
4596 // IV inc is expressed as add or sub, iv.next is potentially a poison value.
4597 // If it has nuw or nsw flags, we need to make sure that these flags are
4598 // inferrable at the point of memory instruction. Otherwise we are replacing
4599 // well-defined two-complement computation with poison. Currently, to avoid
4600 // potentially complex analysis needed to prove this, we reject such cases.
4601 if (auto *OIVInc = dyn_cast<OverflowingBinaryOperator>(Val: IVInc->first))
4602 if (OIVInc->hasNoSignedWrap() || OIVInc->hasNoUnsignedWrap())
4603 return std::nullopt;
4604 if (auto *ConstantStep = dyn_cast<ConstantInt>(Val: IVInc->second))
4605 return std::make_pair(x&: IVInc->first, y: ConstantStep->getValue());
4606 return std::nullopt;
4607 };
4608
4609 // Try to account for the following special case:
4610 // 1. ScaleReg is an inductive variable;
4611 // 2. We use it with non-zero offset;
4612 // 3. IV's increment is available at the point of memory instruction.
4613 //
4614 // In this case, we may reuse the IV increment instead of the IV Phi to
4615 // achieve the following advantages:
4616 // 1. If IV step matches the offset, we will have no need in the offset;
4617 // 2. Even if they don't match, we will reduce the overlap of living IV
4618 // and IV increment, that will potentially lead to better register
4619 // assignment.
4620 if (AddrMode.BaseOffs) {
4621 if (auto IVStep = GetConstantStep(ScaleReg)) {
4622 Instruction *IVInc = IVStep->first;
4623 // The following assert is important to ensure a lack of infinite loops.
4624 // This transforms is (intentionally) the inverse of the one just above.
4625 // If they don't agree on the definition of an increment, we'd alternate
4626 // back and forth indefinitely.
4627 assert(isIVIncrement(IVInc, &LI) && "implied by GetConstantStep");
4628 APInt Step = IVStep->second;
4629 APInt Offset = Step * AddrMode.Scale;
4630 if (Offset.isSignedIntN(N: 64)) {
4631 TestAddrMode.InBounds = false;
4632 TestAddrMode.ScaledReg = IVInc;
4633 TestAddrMode.BaseOffs -= Offset.getLimitedValue();
4634 // If this addressing mode is legal, commit it..
4635 // (Note that we defer the (expensive) domtree base legality check
4636 // to the very last possible point.)
4637 if (TLI.isLegalAddressingMode(DL, AM: TestAddrMode, Ty: AccessTy, AddrSpace) &&
4638 getDTFn().dominates(Def: IVInc, User: MemoryInst)) {
4639 AddrModeInsts.push_back(Elt: cast<Instruction>(Val: IVInc));
4640 AddrMode = TestAddrMode;
4641 return true;
4642 }
4643 // Restore status quo.
4644 TestAddrMode = AddrMode;
4645 }
4646 }
4647 }
4648
4649 // Otherwise, just return what we have.
4650 return true;
4651}
4652
4653/// This is a little filter, which returns true if an addressing computation
4654/// involving I might be folded into a load/store accessing it.
4655/// This doesn't need to be perfect, but needs to accept at least
4656/// the set of instructions that MatchOperationAddr can.
4657static bool MightBeFoldableInst(Instruction *I) {
4658 switch (I->getOpcode()) {
4659 case Instruction::BitCast:
4660 case Instruction::AddrSpaceCast:
4661 // Don't touch identity bitcasts.
4662 if (I->getType() == I->getOperand(i: 0)->getType())
4663 return false;
4664 return I->getType()->isIntOrPtrTy();
4665 case Instruction::PtrToInt:
4666 // PtrToInt is always a noop, as we know that the int type is pointer sized.
4667 return true;
4668 case Instruction::IntToPtr:
4669 // We know the input is intptr_t, so this is foldable.
4670 return true;
4671 case Instruction::Add:
4672 return true;
4673 case Instruction::Mul:
4674 case Instruction::Shl:
4675 // Can only handle X*C and X << C.
4676 return isa<ConstantInt>(Val: I->getOperand(i: 1));
4677 case Instruction::GetElementPtr:
4678 return true;
4679 default:
4680 return false;
4681 }
4682}
4683
4684/// Check whether or not \p Val is a legal instruction for \p TLI.
4685/// \note \p Val is assumed to be the product of some type promotion.
4686/// Therefore if \p Val has an undefined state in \p TLI, this is assumed
4687/// to be legal, as the non-promoted value would have had the same state.
4688static bool isPromotedInstructionLegal(const TargetLowering &TLI,
4689 const DataLayout &DL, Value *Val) {
4690 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
4691 if (!PromotedInst)
4692 return false;
4693 int ISDOpcode = TLI.InstructionOpcodeToISD(Opcode: PromotedInst->getOpcode());
4694 // If the ISDOpcode is undefined, it was undefined before the promotion.
4695 if (!ISDOpcode)
4696 return true;
4697 // Otherwise, check if the promoted instruction is legal or not.
4698 return TLI.isOperationLegalOrCustom(
4699 Op: ISDOpcode, VT: TLI.getValueType(DL, Ty: PromotedInst->getType()));
4700}
4701
4702namespace {
4703
4704/// Hepler class to perform type promotion.
4705class TypePromotionHelper {
4706 /// Utility function to add a promoted instruction \p ExtOpnd to
4707 /// \p PromotedInsts and record the type of extension we have seen.
4708 static void addPromotedInst(InstrToOrigTy &PromotedInsts,
4709 Instruction *ExtOpnd, bool IsSExt) {
4710 ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
4711 auto [It, Inserted] = PromotedInsts.try_emplace(Key: ExtOpnd);
4712 if (!Inserted) {
4713 // If the new extension is same as original, the information in
4714 // PromotedInsts[ExtOpnd] is still correct.
4715 if (It->second.getInt() == ExtTy)
4716 return;
4717
4718 // Now the new extension is different from old extension, we make
4719 // the type information invalid by setting extension type to
4720 // BothExtension.
4721 ExtTy = BothExtension;
4722 }
4723 It->second = TypeIsSExt(ExtOpnd->getType(), ExtTy);
4724 }
4725
4726 /// Utility function to query the original type of instruction \p Opnd
4727 /// with a matched extension type. If the extension doesn't match, we
4728 /// cannot use the information we had on the original type.
4729 /// BothExtension doesn't match any extension type.
4730 static const Type *getOrigType(const InstrToOrigTy &PromotedInsts,
4731 Instruction *Opnd, bool IsSExt) {
4732 ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
4733 InstrToOrigTy::const_iterator It = PromotedInsts.find(Val: Opnd);
4734 if (It != PromotedInsts.end() && It->second.getInt() == ExtTy)
4735 return It->second.getPointer();
4736 return nullptr;
4737 }
4738
4739 /// Utility function to check whether or not a sign or zero extension
4740 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
4741 /// either using the operands of \p Inst or promoting \p Inst.
4742 /// The type of the extension is defined by \p IsSExt.
4743 /// In other words, check if:
4744 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
4745 /// #1 Promotion applies:
4746 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
4747 /// #2 Operand reuses:
4748 /// ext opnd1 to ConsideredExtType.
4749 /// \p PromotedInsts maps the instructions to their type before promotion.
4750 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
4751 const InstrToOrigTy &PromotedInsts, bool IsSExt);
4752
4753 /// Utility function to determine if \p OpIdx should be promoted when
4754 /// promoting \p Inst.
4755 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
4756 return !(isa<SelectInst>(Val: Inst) && OpIdx == 0);
4757 }
4758
4759 /// Utility function to promote the operand of \p Ext when this
4760 /// operand is a promotable trunc or sext or zext.
4761 /// \p PromotedInsts maps the instructions to their type before promotion.
4762 /// \p CreatedInstsCost[out] contains the cost of all instructions
4763 /// created to promote the operand of Ext.
4764 /// Newly added extensions are inserted in \p Exts.
4765 /// Newly added truncates are inserted in \p Truncs.
4766 /// Should never be called directly.
4767 /// \return The promoted value which is used instead of Ext.
4768 static Value *promoteOperandForTruncAndAnyExt(
4769 Instruction *Ext, TypePromotionTransaction &TPT,
4770 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4771 SmallVectorImpl<Instruction *> *Exts,
4772 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
4773
4774 /// Utility function to promote the operand of \p Ext when this
4775 /// operand is promotable and is not a supported trunc or sext.
4776 /// \p PromotedInsts maps the instructions to their type before promotion.
4777 /// \p CreatedInstsCost[out] contains the cost of all the instructions
4778 /// created to promote the operand of Ext.
4779 /// Newly added extensions are inserted in \p Exts.
4780 /// Newly added truncates are inserted in \p Truncs.
4781 /// Should never be called directly.
4782 /// \return The promoted value which is used instead of Ext.
4783 static Value *promoteOperandForOther(Instruction *Ext,
4784 TypePromotionTransaction &TPT,
4785 InstrToOrigTy &PromotedInsts,
4786 unsigned &CreatedInstsCost,
4787 SmallVectorImpl<Instruction *> *Exts,
4788 SmallVectorImpl<Instruction *> *Truncs,
4789 const TargetLowering &TLI, bool IsSExt);
4790
4791 /// \see promoteOperandForOther.
4792 static Value *signExtendOperandForOther(
4793 Instruction *Ext, TypePromotionTransaction &TPT,
4794 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4795 SmallVectorImpl<Instruction *> *Exts,
4796 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4797 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
4798 Exts, Truncs, TLI, IsSExt: true);
4799 }
4800
4801 /// \see promoteOperandForOther.
4802 static Value *zeroExtendOperandForOther(
4803 Instruction *Ext, TypePromotionTransaction &TPT,
4804 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4805 SmallVectorImpl<Instruction *> *Exts,
4806 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4807 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
4808 Exts, Truncs, TLI, IsSExt: false);
4809 }
4810
4811public:
4812 /// Type for the utility function that promotes the operand of Ext.
4813 using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
4814 InstrToOrigTy &PromotedInsts,
4815 unsigned &CreatedInstsCost,
4816 SmallVectorImpl<Instruction *> *Exts,
4817 SmallVectorImpl<Instruction *> *Truncs,
4818 const TargetLowering &TLI);
4819
4820 /// Given a sign/zero extend instruction \p Ext, return the appropriate
4821 /// action to promote the operand of \p Ext instead of using Ext.
4822 /// \return NULL if no promotable action is possible with the current
4823 /// sign extension.
4824 /// \p InsertedInsts keeps track of all the instructions inserted by the
4825 /// other CodeGenPrepare optimizations. This information is important
4826 /// because we do not want to promote these instructions as CodeGenPrepare
4827 /// will reinsert them later. Thus creating an infinite loop: create/remove.
4828 /// \p PromotedInsts maps the instructions to their type before promotion.
4829 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
4830 const TargetLowering &TLI,
4831 const InstrToOrigTy &PromotedInsts);
4832};
4833
4834} // end anonymous namespace
4835
4836bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
4837 Type *ConsideredExtType,
4838 const InstrToOrigTy &PromotedInsts,
4839 bool IsSExt) {
4840 // The promotion helper does not know how to deal with vector types yet.
4841 // To be able to fix that, we would need to fix the places where we
4842 // statically extend, e.g., constants and such.
4843 if (Inst->getType()->isVectorTy())
4844 return false;
4845
4846 // We can always get through zext.
4847 if (isa<ZExtInst>(Val: Inst))
4848 return true;
4849
4850 // sext(sext) is ok too.
4851 if (IsSExt && isa<SExtInst>(Val: Inst))
4852 return true;
4853
4854 // We can get through binary operator, if it is legal. In other words, the
4855 // binary operator must have a nuw or nsw flag.
4856 if (const auto *BinOp = dyn_cast<BinaryOperator>(Val: Inst))
4857 if (isa<OverflowingBinaryOperator>(Val: BinOp) &&
4858 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
4859 (IsSExt && BinOp->hasNoSignedWrap())))
4860 return true;
4861
4862 // ext(and(opnd, cst)) --> and(ext(opnd), ext(cst))
4863 if ((Inst->getOpcode() == Instruction::And ||
4864 Inst->getOpcode() == Instruction::Or))
4865 return true;
4866
4867 // ext(xor(opnd, cst)) --> xor(ext(opnd), ext(cst))
4868 if (Inst->getOpcode() == Instruction::Xor) {
4869 // Make sure it is not a NOT.
4870 if (const auto *Cst = dyn_cast<ConstantInt>(Val: Inst->getOperand(i: 1)))
4871 if (!Cst->getValue().isAllOnes())
4872 return true;
4873 }
4874
4875 // zext(shrl(opnd, cst)) --> shrl(zext(opnd), zext(cst))
4876 // It may change a poisoned value into a regular value, like
4877 // zext i32 (shrl i8 %val, 12) --> shrl i32 (zext i8 %val), 12
4878 // poisoned value regular value
4879 // It should be OK since undef covers valid value.
4880 if (Inst->getOpcode() == Instruction::LShr && !IsSExt)
4881 return true;
4882
4883 // and(ext(shl(opnd, cst)), cst) --> and(shl(ext(opnd), ext(cst)), cst)
4884 // It may change a poisoned value into a regular value, like
4885 // zext i32 (shl i8 %val, 12) --> shl i32 (zext i8 %val), 12
4886 // poisoned value regular value
4887 // It should be OK since undef covers valid value.
4888 if (Inst->getOpcode() == Instruction::Shl && Inst->hasOneUse()) {
4889 const auto *ExtInst = cast<const Instruction>(Val: *Inst->user_begin());
4890 if (ExtInst->hasOneUse()) {
4891 const auto *AndInst = dyn_cast<const Instruction>(Val: *ExtInst->user_begin());
4892 if (AndInst && AndInst->getOpcode() == Instruction::And) {
4893 const auto *Cst = dyn_cast<ConstantInt>(Val: AndInst->getOperand(i: 1));
4894 if (Cst &&
4895 Cst->getValue().isIntN(N: Inst->getType()->getIntegerBitWidth()))
4896 return true;
4897 }
4898 }
4899 }
4900
4901 // Check if we can do the following simplification.
4902 // ext(trunc(opnd)) --> ext(opnd)
4903 if (!isa<TruncInst>(Val: Inst))
4904 return false;
4905
4906 Value *OpndVal = Inst->getOperand(i: 0);
4907 // Check if we can use this operand in the extension.
4908 // If the type is larger than the result type of the extension, we cannot.
4909 if (!OpndVal->getType()->isIntegerTy() ||
4910 OpndVal->getType()->getIntegerBitWidth() >
4911 ConsideredExtType->getIntegerBitWidth())
4912 return false;
4913
4914 // If the operand of the truncate is not an instruction, we will not have
4915 // any information on the dropped bits.
4916 // (Actually we could for constant but it is not worth the extra logic).
4917 Instruction *Opnd = dyn_cast<Instruction>(Val: OpndVal);
4918 if (!Opnd)
4919 return false;
4920
4921 // Check if the source of the type is narrow enough.
4922 // I.e., check that trunc just drops extended bits of the same kind of
4923 // the extension.
4924 // #1 get the type of the operand and check the kind of the extended bits.
4925 const Type *OpndType = getOrigType(PromotedInsts, Opnd, IsSExt);
4926 if (OpndType)
4927 ;
4928 else if ((IsSExt && isa<SExtInst>(Val: Opnd)) || (!IsSExt && isa<ZExtInst>(Val: Opnd)))
4929 OpndType = Opnd->getOperand(i: 0)->getType();
4930 else
4931 return false;
4932
4933 // #2 check that the truncate just drops extended bits.
4934 return Inst->getType()->getIntegerBitWidth() >=
4935 OpndType->getIntegerBitWidth();
4936}
4937
4938TypePromotionHelper::Action TypePromotionHelper::getAction(
4939 Instruction *Ext, const SetOfInstrs &InsertedInsts,
4940 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
4941 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
4942 "Unexpected instruction type");
4943 Instruction *ExtOpnd = dyn_cast<Instruction>(Val: Ext->getOperand(i: 0));
4944 Type *ExtTy = Ext->getType();
4945 bool IsSExt = isa<SExtInst>(Val: Ext);
4946 // If the operand of the extension is not an instruction, we cannot
4947 // get through.
4948 // If it, check we can get through.
4949 if (!ExtOpnd || !canGetThrough(Inst: ExtOpnd, ConsideredExtType: ExtTy, PromotedInsts, IsSExt))
4950 return nullptr;
4951
4952 // Do not promote if the operand has been added by codegenprepare.
4953 // Otherwise, it means we are undoing an optimization that is likely to be
4954 // redone, thus causing potential infinite loop.
4955 if (isa<TruncInst>(Val: ExtOpnd) && InsertedInsts.count(Ptr: ExtOpnd))
4956 return nullptr;
4957
4958 // SExt or Trunc instructions.
4959 // Return the related handler.
4960 if (isa<SExtInst>(Val: ExtOpnd) || isa<TruncInst>(Val: ExtOpnd) ||
4961 isa<ZExtInst>(Val: ExtOpnd))
4962 return promoteOperandForTruncAndAnyExt;
4963
4964 // Regular instruction.
4965 // Abort early if we will have to insert non-free instructions.
4966 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(FromTy: ExtTy, ToTy: ExtOpnd->getType()))
4967 return nullptr;
4968 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
4969}
4970
4971Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
4972 Instruction *SExt, TypePromotionTransaction &TPT,
4973 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4974 SmallVectorImpl<Instruction *> *Exts,
4975 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4976 // By construction, the operand of SExt is an instruction. Otherwise we cannot
4977 // get through it and this method should not be called.
4978 Instruction *SExtOpnd = cast<Instruction>(Val: SExt->getOperand(i: 0));
4979 Value *ExtVal = SExt;
4980 bool HasMergedNonFreeExt = false;
4981 if (isa<ZExtInst>(Val: SExtOpnd)) {
4982 // Replace s|zext(zext(opnd))
4983 // => zext(opnd).
4984 HasMergedNonFreeExt = !TLI.isExtFree(I: SExtOpnd);
4985 Value *ZExt =
4986 TPT.createZExt(Inst: SExt, Opnd: SExtOpnd->getOperand(i: 0), Ty: SExt->getType());
4987 TPT.replaceAllUsesWith(Inst: SExt, New: ZExt);
4988 TPT.eraseInstruction(Inst: SExt);
4989 ExtVal = ZExt;
4990 } else {
4991 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
4992 // => z|sext(opnd).
4993 TPT.setOperand(Inst: SExt, Idx: 0, NewVal: SExtOpnd->getOperand(i: 0));
4994 }
4995 CreatedInstsCost = 0;
4996
4997 // Remove dead code.
4998 if (SExtOpnd->use_empty())
4999 TPT.eraseInstruction(Inst: SExtOpnd);
5000
5001 // Check if the extension is still needed.
5002 Instruction *ExtInst = dyn_cast<Instruction>(Val: ExtVal);
5003 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(i: 0)->getType()) {
5004 if (ExtInst) {
5005 if (Exts)
5006 Exts->push_back(Elt: ExtInst);
5007 CreatedInstsCost = !TLI.isExtFree(I: ExtInst) && !HasMergedNonFreeExt;
5008 }
5009 return ExtVal;
5010 }
5011
5012 // At this point we have: ext ty opnd to ty.
5013 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
5014 Value *NextVal = ExtInst->getOperand(i: 0);
5015 TPT.eraseInstruction(Inst: ExtInst, NewVal: NextVal);
5016 return NextVal;
5017}
5018
5019Value *TypePromotionHelper::promoteOperandForOther(
5020 Instruction *Ext, TypePromotionTransaction &TPT,
5021 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
5022 SmallVectorImpl<Instruction *> *Exts,
5023 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
5024 bool IsSExt) {
5025 // By construction, the operand of Ext is an instruction. Otherwise we cannot
5026 // get through it and this method should not be called.
5027 Instruction *ExtOpnd = cast<Instruction>(Val: Ext->getOperand(i: 0));
5028 CreatedInstsCost = 0;
5029 if (!ExtOpnd->hasOneUse()) {
5030 // ExtOpnd will be promoted.
5031 // All its uses, but Ext, will need to use a truncated value of the
5032 // promoted version.
5033 // Create the truncate now.
5034 Value *Trunc = TPT.createTrunc(Opnd: Ext, Ty: ExtOpnd->getType());
5035 if (Instruction *ITrunc = dyn_cast<Instruction>(Val: Trunc)) {
5036 // Insert it just after the definition.
5037 ITrunc->moveAfter(MovePos: ExtOpnd);
5038 if (Truncs)
5039 Truncs->push_back(Elt: ITrunc);
5040 }
5041
5042 TPT.replaceAllUsesWith(Inst: ExtOpnd, New: Trunc);
5043 // Restore the operand of Ext (which has been replaced by the previous call
5044 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
5045 TPT.setOperand(Inst: Ext, Idx: 0, NewVal: ExtOpnd);
5046 }
5047
5048 // Get through the Instruction:
5049 // 1. Update its type.
5050 // 2. Replace the uses of Ext by Inst.
5051 // 3. Extend each operand that needs to be extended.
5052
5053 // Remember the original type of the instruction before promotion.
5054 // This is useful to know that the high bits are sign extended bits.
5055 addPromotedInst(PromotedInsts, ExtOpnd, IsSExt);
5056 // Step #1.
5057 TPT.mutateType(Inst: ExtOpnd, NewTy: Ext->getType());
5058 // Step #2.
5059 TPT.replaceAllUsesWith(Inst: Ext, New: ExtOpnd);
5060 // Step #3.
5061 LLVM_DEBUG(dbgs() << "Propagate Ext to operands\n");
5062 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
5063 ++OpIdx) {
5064 LLVM_DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
5065 if (ExtOpnd->getOperand(i: OpIdx)->getType() == Ext->getType() ||
5066 !shouldExtOperand(Inst: ExtOpnd, OpIdx)) {
5067 LLVM_DEBUG(dbgs() << "No need to propagate\n");
5068 continue;
5069 }
5070 // Check if we can statically extend the operand.
5071 Value *Opnd = ExtOpnd->getOperand(i: OpIdx);
5072 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Val: Opnd)) {
5073 LLVM_DEBUG(dbgs() << "Statically extend\n");
5074 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
5075 APInt CstVal = IsSExt ? Cst->getValue().sext(width: BitWidth)
5076 : Cst->getValue().zext(width: BitWidth);
5077 TPT.setOperand(Inst: ExtOpnd, Idx: OpIdx, NewVal: ConstantInt::get(Ty: Ext->getType(), V: CstVal));
5078 continue;
5079 }
5080 // UndefValue are typed, so we have to statically sign extend them.
5081 if (isa<UndefValue>(Val: Opnd)) {
5082 LLVM_DEBUG(dbgs() << "Statically extend\n");
5083 TPT.setOperand(Inst: ExtOpnd, Idx: OpIdx, NewVal: UndefValue::get(T: Ext->getType()));
5084 continue;
5085 }
5086
5087 // Otherwise we have to explicitly sign extend the operand.
5088 Value *ValForExtOpnd = IsSExt
5089 ? TPT.createSExt(Inst: ExtOpnd, Opnd, Ty: Ext->getType())
5090 : TPT.createZExt(Inst: ExtOpnd, Opnd, Ty: Ext->getType());
5091 TPT.setOperand(Inst: ExtOpnd, Idx: OpIdx, NewVal: ValForExtOpnd);
5092 Instruction *InstForExtOpnd = dyn_cast<Instruction>(Val: ValForExtOpnd);
5093 if (!InstForExtOpnd)
5094 continue;
5095
5096 if (Exts)
5097 Exts->push_back(Elt: InstForExtOpnd);
5098
5099 CreatedInstsCost += !TLI.isExtFree(I: InstForExtOpnd);
5100 }
5101 LLVM_DEBUG(dbgs() << "Extension is useless now\n");
5102 TPT.eraseInstruction(Inst: Ext);
5103 return ExtOpnd;
5104}
5105
5106/// Check whether or not promoting an instruction to a wider type is profitable.
5107/// \p NewCost gives the cost of extension instructions created by the
5108/// promotion.
5109/// \p OldCost gives the cost of extension instructions before the promotion
5110/// plus the number of instructions that have been
5111/// matched in the addressing mode the promotion.
5112/// \p PromotedOperand is the value that has been promoted.
5113/// \return True if the promotion is profitable, false otherwise.
5114bool AddressingModeMatcher::isPromotionProfitable(
5115 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
5116 LLVM_DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost
5117 << '\n');
5118 // The cost of the new extensions is greater than the cost of the
5119 // old extension plus what we folded.
5120 // This is not profitable.
5121 if (NewCost > OldCost)
5122 return false;
5123 if (NewCost < OldCost)
5124 return true;
5125 // The promotion is neutral but it may help folding the sign extension in
5126 // loads for instance.
5127 // Check that we did not create an illegal instruction.
5128 return isPromotedInstructionLegal(TLI, DL, Val: PromotedOperand);
5129}
5130
5131/// Given an instruction or constant expr, see if we can fold the operation
5132/// into the addressing mode. If so, update the addressing mode and return
5133/// true, otherwise return false without modifying AddrMode.
5134/// If \p MovedAway is not NULL, it contains the information of whether or
5135/// not AddrInst has to be folded into the addressing mode on success.
5136/// If \p MovedAway == true, \p AddrInst will not be part of the addressing
5137/// because it has been moved away.
5138/// Thus AddrInst must not be added in the matched instructions.
5139/// This state can happen when AddrInst is a sext, since it may be moved away.
5140/// Therefore, AddrInst may not be valid when MovedAway is true and it must
5141/// not be referenced anymore.
5142bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
5143 unsigned Depth,
5144 bool *MovedAway) {
5145 // Avoid exponential behavior on extremely deep expression trees.
5146 if (Depth >= 5)
5147 return false;
5148
5149 // By default, all matched instructions stay in place.
5150 if (MovedAway)
5151 *MovedAway = false;
5152
5153 switch (Opcode) {
5154 case Instruction::PtrToInt:
5155 // PtrToInt is always a noop, as we know that the int type is pointer sized.
5156 return matchAddr(Addr: AddrInst->getOperand(i: 0), Depth);
5157 case Instruction::IntToPtr: {
5158 auto AS = AddrInst->getType()->getPointerAddressSpace();
5159 auto PtrTy = MVT::getIntegerVT(BitWidth: DL.getPointerSizeInBits(AS));
5160 // This inttoptr is a no-op if the integer type is pointer sized.
5161 if (TLI.getValueType(DL, Ty: AddrInst->getOperand(i: 0)->getType()) == PtrTy)
5162 return matchAddr(Addr: AddrInst->getOperand(i: 0), Depth);
5163 return false;
5164 }
5165 case Instruction::BitCast:
5166 // BitCast is always a noop, and we can handle it as long as it is
5167 // int->int or pointer->pointer (we don't want int<->fp or something).
5168 if (AddrInst->getOperand(i: 0)->getType()->isIntOrPtrTy() &&
5169 // Don't touch identity bitcasts. These were probably put here by LSR,
5170 // and we don't want to mess around with them. Assume it knows what it
5171 // is doing.
5172 AddrInst->getOperand(i: 0)->getType() != AddrInst->getType())
5173 return matchAddr(Addr: AddrInst->getOperand(i: 0), Depth);
5174 return false;
5175 case Instruction::AddrSpaceCast: {
5176 unsigned SrcAS =
5177 AddrInst->getOperand(i: 0)->getType()->getPointerAddressSpace();
5178 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
5179 if (TLI.getTargetMachine().isNoopAddrSpaceCast(SrcAS, DestAS))
5180 return matchAddr(Addr: AddrInst->getOperand(i: 0), Depth);
5181 return false;
5182 }
5183 case Instruction::Add: {
5184 // Check to see if we can merge in one operand, then the other. If so, we
5185 // win.
5186 ExtAddrMode BackupAddrMode = AddrMode;
5187 unsigned OldSize = AddrModeInsts.size();
5188 // Start a transaction at this point.
5189 // The LHS may match but not the RHS.
5190 // Therefore, we need a higher level restoration point to undo partially
5191 // matched operation.
5192 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5193 TPT.getRestorationPoint();
5194
5195 // Try to match an integer constant second to increase its chance of ending
5196 // up in `BaseOffs`, resp. decrease its chance of ending up in `BaseReg`.
5197 int First = 0, Second = 1;
5198 if (isa<ConstantInt>(Val: AddrInst->getOperand(i: First))
5199 && !isa<ConstantInt>(Val: AddrInst->getOperand(i: Second)))
5200 std::swap(a&: First, b&: Second);
5201 AddrMode.InBounds = false;
5202 if (matchAddr(Addr: AddrInst->getOperand(i: First), Depth: Depth + 1) &&
5203 matchAddr(Addr: AddrInst->getOperand(i: Second), Depth: Depth + 1))
5204 return true;
5205
5206 // Restore the old addr mode info.
5207 AddrMode = BackupAddrMode;
5208 AddrModeInsts.resize(N: OldSize);
5209 TPT.rollback(Point: LastKnownGood);
5210
5211 // Otherwise this was over-aggressive. Try merging operands in the opposite
5212 // order.
5213 if (matchAddr(Addr: AddrInst->getOperand(i: Second), Depth: Depth + 1) &&
5214 matchAddr(Addr: AddrInst->getOperand(i: First), Depth: Depth + 1))
5215 return true;
5216
5217 // Otherwise we definitely can't merge the ADD in.
5218 AddrMode = BackupAddrMode;
5219 AddrModeInsts.resize(N: OldSize);
5220 TPT.rollback(Point: LastKnownGood);
5221 break;
5222 }
5223 // case Instruction::Or:
5224 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
5225 // break;
5226 case Instruction::Mul:
5227 case Instruction::Shl: {
5228 // Can only handle X*C and X << C.
5229 AddrMode.InBounds = false;
5230 ConstantInt *RHS = dyn_cast<ConstantInt>(Val: AddrInst->getOperand(i: 1));
5231 if (!RHS || RHS->getBitWidth() > 64)
5232 return false;
5233 int64_t Scale = Opcode == Instruction::Shl
5234 ? 1LL << RHS->getLimitedValue(Limit: RHS->getBitWidth() - 1)
5235 : RHS->getSExtValue();
5236
5237 return matchScaledValue(ScaleReg: AddrInst->getOperand(i: 0), Scale, Depth);
5238 }
5239 case Instruction::GetElementPtr: {
5240 // Scan the GEP. We check it if it contains constant offsets and at most
5241 // one variable offset.
5242 int VariableOperand = -1;
5243 unsigned VariableScale = 0;
5244
5245 int64_t ConstantOffset = 0;
5246 gep_type_iterator GTI = gep_type_begin(GEP: AddrInst);
5247 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
5248 if (StructType *STy = GTI.getStructTypeOrNull()) {
5249 const StructLayout *SL = DL.getStructLayout(Ty: STy);
5250 unsigned Idx =
5251 cast<ConstantInt>(Val: AddrInst->getOperand(i))->getZExtValue();
5252 ConstantOffset += SL->getElementOffset(Idx);
5253 } else {
5254 TypeSize TS = GTI.getSequentialElementStride(DL);
5255 if (TS.isNonZero()) {
5256 // The optimisations below currently only work for fixed offsets.
5257 if (TS.isScalable())
5258 return false;
5259 int64_t TypeSize = TS.getFixedValue();
5260 if (ConstantInt *CI =
5261 dyn_cast<ConstantInt>(Val: AddrInst->getOperand(i))) {
5262 const APInt &CVal = CI->getValue();
5263 if (CVal.getSignificantBits() <= 64) {
5264 ConstantOffset += CVal.getSExtValue() * TypeSize;
5265 continue;
5266 }
5267 }
5268 // We only allow one variable index at the moment.
5269 if (VariableOperand != -1)
5270 return false;
5271
5272 // Remember the variable index.
5273 VariableOperand = i;
5274 VariableScale = TypeSize;
5275 }
5276 }
5277 }
5278
5279 // A common case is for the GEP to only do a constant offset. In this case,
5280 // just add it to the disp field and check validity.
5281 if (VariableOperand == -1) {
5282 AddrMode.BaseOffs += ConstantOffset;
5283 if (matchAddr(Addr: AddrInst->getOperand(i: 0), Depth: Depth + 1)) {
5284 if (!cast<GEPOperator>(Val: AddrInst)->isInBounds())
5285 AddrMode.InBounds = false;
5286 return true;
5287 }
5288 AddrMode.BaseOffs -= ConstantOffset;
5289
5290 if (EnableGEPOffsetSplit && isa<GetElementPtrInst>(Val: AddrInst) &&
5291 TLI.shouldConsiderGEPOffsetSplit() && Depth == 0 &&
5292 ConstantOffset > 0) {
5293 // Record GEPs with non-zero offsets as candidates for splitting in
5294 // the event that the offset cannot fit into the r+i addressing mode.
5295 // Simple and common case that only one GEP is used in calculating the
5296 // address for the memory access.
5297 Value *Base = AddrInst->getOperand(i: 0);
5298 auto *BaseI = dyn_cast<Instruction>(Val: Base);
5299 auto *GEP = cast<GetElementPtrInst>(Val: AddrInst);
5300 if (isa<Argument>(Val: Base) || isa<GlobalValue>(Val: Base) ||
5301 (BaseI && !isa<CastInst>(Val: BaseI) &&
5302 !isa<GetElementPtrInst>(Val: BaseI))) {
5303 // Make sure the parent block allows inserting non-PHI instructions
5304 // before the terminator.
5305 BasicBlock *Parent = BaseI ? BaseI->getParent()
5306 : &GEP->getFunction()->getEntryBlock();
5307 if (!Parent->getTerminator()->isEHPad())
5308 LargeOffsetGEP = std::make_pair(x&: GEP, y&: ConstantOffset);
5309 }
5310 }
5311
5312 return false;
5313 }
5314
5315 // Save the valid addressing mode in case we can't match.
5316 ExtAddrMode BackupAddrMode = AddrMode;
5317 unsigned OldSize = AddrModeInsts.size();
5318
5319 // See if the scale and offset amount is valid for this target.
5320 AddrMode.BaseOffs += ConstantOffset;
5321 if (!cast<GEPOperator>(Val: AddrInst)->isInBounds())
5322 AddrMode.InBounds = false;
5323
5324 // Match the base operand of the GEP.
5325 if (!matchAddr(Addr: AddrInst->getOperand(i: 0), Depth: Depth + 1)) {
5326 // If it couldn't be matched, just stuff the value in a register.
5327 if (AddrMode.HasBaseReg) {
5328 AddrMode = BackupAddrMode;
5329 AddrModeInsts.resize(N: OldSize);
5330 return false;
5331 }
5332 AddrMode.HasBaseReg = true;
5333 AddrMode.BaseReg = AddrInst->getOperand(i: 0);
5334 }
5335
5336 // Match the remaining variable portion of the GEP.
5337 if (!matchScaledValue(ScaleReg: AddrInst->getOperand(i: VariableOperand), Scale: VariableScale,
5338 Depth)) {
5339 // If it couldn't be matched, try stuffing the base into a register
5340 // instead of matching it, and retrying the match of the scale.
5341 AddrMode = BackupAddrMode;
5342 AddrModeInsts.resize(N: OldSize);
5343 if (AddrMode.HasBaseReg)
5344 return false;
5345 AddrMode.HasBaseReg = true;
5346 AddrMode.BaseReg = AddrInst->getOperand(i: 0);
5347 AddrMode.BaseOffs += ConstantOffset;
5348 if (!matchScaledValue(ScaleReg: AddrInst->getOperand(i: VariableOperand),
5349 Scale: VariableScale, Depth)) {
5350 // If even that didn't work, bail.
5351 AddrMode = BackupAddrMode;
5352 AddrModeInsts.resize(N: OldSize);
5353 return false;
5354 }
5355 }
5356
5357 return true;
5358 }
5359 case Instruction::SExt:
5360 case Instruction::ZExt: {
5361 Instruction *Ext = dyn_cast<Instruction>(Val: AddrInst);
5362 if (!Ext)
5363 return false;
5364
5365 // Try to move this ext out of the way of the addressing mode.
5366 // Ask for a method for doing so.
5367 TypePromotionHelper::Action TPH =
5368 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
5369 if (!TPH)
5370 return false;
5371
5372 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5373 TPT.getRestorationPoint();
5374 unsigned CreatedInstsCost = 0;
5375 unsigned ExtCost = !TLI.isExtFree(I: Ext);
5376 Value *PromotedOperand =
5377 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
5378 // SExt has been moved away.
5379 // Thus either it will be rematched later in the recursive calls or it is
5380 // gone. Anyway, we must not fold it into the addressing mode at this point.
5381 // E.g.,
5382 // op = add opnd, 1
5383 // idx = ext op
5384 // addr = gep base, idx
5385 // is now:
5386 // promotedOpnd = ext opnd <- no match here
5387 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
5388 // addr = gep base, op <- match
5389 if (MovedAway)
5390 *MovedAway = true;
5391
5392 assert(PromotedOperand &&
5393 "TypePromotionHelper should have filtered out those cases");
5394
5395 ExtAddrMode BackupAddrMode = AddrMode;
5396 unsigned OldSize = AddrModeInsts.size();
5397
5398 if (!matchAddr(Addr: PromotedOperand, Depth) ||
5399 // The total of the new cost is equal to the cost of the created
5400 // instructions.
5401 // The total of the old cost is equal to the cost of the extension plus
5402 // what we have saved in the addressing mode.
5403 !isPromotionProfitable(NewCost: CreatedInstsCost,
5404 OldCost: ExtCost + (AddrModeInsts.size() - OldSize),
5405 PromotedOperand)) {
5406 AddrMode = BackupAddrMode;
5407 AddrModeInsts.resize(N: OldSize);
5408 LLVM_DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
5409 TPT.rollback(Point: LastKnownGood);
5410 return false;
5411 }
5412
5413 // SExt has been deleted. Make sure it is not referenced by the AddrMode.
5414 AddrMode.replaceWith(From: Ext, To: PromotedOperand);
5415 return true;
5416 }
5417 case Instruction::Call:
5418 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: AddrInst)) {
5419 if (II->getIntrinsicID() == Intrinsic::threadlocal_address) {
5420 GlobalValue &GV = cast<GlobalValue>(Val&: *II->getArgOperand(i: 0));
5421 if (TLI.addressingModeSupportsTLS(GV))
5422 return matchAddr(Addr: AddrInst->getOperand(i: 0), Depth);
5423 }
5424 }
5425 break;
5426 }
5427 return false;
5428}
5429
5430/// If we can, try to add the value of 'Addr' into the current addressing mode.
5431/// If Addr can't be added to AddrMode this returns false and leaves AddrMode
5432/// unmodified. This assumes that Addr is either a pointer type or intptr_t
5433/// for the target.
5434///
5435bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
5436 // Start a transaction at this point that we will rollback if the matching
5437 // fails.
5438 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5439 TPT.getRestorationPoint();
5440 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: Addr)) {
5441 if (CI->getValue().isSignedIntN(N: 64)) {
5442 // Check if the addition would result in a signed overflow.
5443 int64_t Result;
5444 bool Overflow =
5445 AddOverflow(X: AddrMode.BaseOffs, Y: CI->getSExtValue(), Result);
5446 if (!Overflow) {
5447 // Fold in immediates if legal for the target.
5448 AddrMode.BaseOffs = Result;
5449 if (TLI.isLegalAddressingMode(DL, AM: AddrMode, Ty: AccessTy, AddrSpace))
5450 return true;
5451 AddrMode.BaseOffs -= CI->getSExtValue();
5452 }
5453 }
5454 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Val: Addr)) {
5455 // If this is a global variable, try to fold it into the addressing mode.
5456 if (!AddrMode.BaseGV) {
5457 AddrMode.BaseGV = GV;
5458 if (TLI.isLegalAddressingMode(DL, AM: AddrMode, Ty: AccessTy, AddrSpace))
5459 return true;
5460 AddrMode.BaseGV = nullptr;
5461 }
5462 } else if (Instruction *I = dyn_cast<Instruction>(Val: Addr)) {
5463 ExtAddrMode BackupAddrMode = AddrMode;
5464 unsigned OldSize = AddrModeInsts.size();
5465
5466 // Check to see if it is possible to fold this operation.
5467 bool MovedAway = false;
5468 if (matchOperationAddr(AddrInst: I, Opcode: I->getOpcode(), Depth, MovedAway: &MovedAway)) {
5469 // This instruction may have been moved away. If so, there is nothing
5470 // to check here.
5471 if (MovedAway)
5472 return true;
5473 // Okay, it's possible to fold this. Check to see if it is actually
5474 // *profitable* to do so. We use a simple cost model to avoid increasing
5475 // register pressure too much.
5476 if (I->hasOneUse() ||
5477 isProfitableToFoldIntoAddressingMode(I, AMBefore&: BackupAddrMode, AMAfter&: AddrMode)) {
5478 AddrModeInsts.push_back(Elt: I);
5479 return true;
5480 }
5481
5482 // It isn't profitable to do this, roll back.
5483 AddrMode = BackupAddrMode;
5484 AddrModeInsts.resize(N: OldSize);
5485 TPT.rollback(Point: LastKnownGood);
5486 }
5487 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Val: Addr)) {
5488 if (matchOperationAddr(AddrInst: CE, Opcode: CE->getOpcode(), Depth))
5489 return true;
5490 TPT.rollback(Point: LastKnownGood);
5491 } else if (isa<ConstantPointerNull>(Val: Addr)) {
5492 // Null pointer gets folded without affecting the addressing mode.
5493 return true;
5494 }
5495
5496 // Worse case, the target should support [reg] addressing modes. :)
5497 if (!AddrMode.HasBaseReg) {
5498 AddrMode.HasBaseReg = true;
5499 AddrMode.BaseReg = Addr;
5500 // Still check for legality in case the target supports [imm] but not [i+r].
5501 if (TLI.isLegalAddressingMode(DL, AM: AddrMode, Ty: AccessTy, AddrSpace))
5502 return true;
5503 AddrMode.HasBaseReg = false;
5504 AddrMode.BaseReg = nullptr;
5505 }
5506
5507 // If the base register is already taken, see if we can do [r+r].
5508 if (AddrMode.Scale == 0) {
5509 AddrMode.Scale = 1;
5510 AddrMode.ScaledReg = Addr;
5511 if (TLI.isLegalAddressingMode(DL, AM: AddrMode, Ty: AccessTy, AddrSpace))
5512 return true;
5513 AddrMode.Scale = 0;
5514 AddrMode.ScaledReg = nullptr;
5515 }
5516 // Couldn't match.
5517 TPT.rollback(Point: LastKnownGood);
5518 return false;
5519}
5520
5521/// Check to see if all uses of OpVal by the specified inline asm call are due
5522/// to memory operands. If so, return true, otherwise return false.
5523static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
5524 const TargetLowering &TLI,
5525 const TargetRegisterInfo &TRI) {
5526 const Function *F = CI->getFunction();
5527 TargetLowering::AsmOperandInfoVector TargetConstraints =
5528 TLI.ParseConstraints(DL: F->getDataLayout(), TRI: &TRI, Call: *CI);
5529
5530 for (TargetLowering::AsmOperandInfo &OpInfo : TargetConstraints) {
5531 // Compute the constraint code and ConstraintType to use.
5532 TLI.ComputeConstraintToUse(OpInfo, Op: SDValue());
5533
5534 // If this asm operand is our Value*, and if it isn't an indirect memory
5535 // operand, we can't fold it! TODO: Also handle C_Address?
5536 if (OpInfo.CallOperandVal == OpVal &&
5537 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
5538 !OpInfo.isIndirect))
5539 return false;
5540 }
5541
5542 return true;
5543}
5544
5545/// Recursively walk all the uses of I until we find a memory use.
5546/// If we find an obviously non-foldable instruction, return true.
5547/// Add accessed addresses and types to MemoryUses.
5548static bool FindAllMemoryUses(
5549 Instruction *I, SmallVectorImpl<std::pair<Use *, Type *>> &MemoryUses,
5550 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
5551 const TargetRegisterInfo &TRI, bool OptSize, ProfileSummaryInfo *PSI,
5552 BlockFrequencyInfo *BFI, unsigned &SeenInsts) {
5553 // If we already considered this instruction, we're done.
5554 if (!ConsideredInsts.insert(Ptr: I).second)
5555 return false;
5556
5557 // If this is an obviously unfoldable instruction, bail out.
5558 if (!MightBeFoldableInst(I))
5559 return true;
5560
5561 // Loop over all the uses, recursively processing them.
5562 for (Use &U : I->uses()) {
5563 // Conservatively return true if we're seeing a large number or a deep chain
5564 // of users. This avoids excessive compilation times in pathological cases.
5565 if (SeenInsts++ >= MaxAddressUsersToScan)
5566 return true;
5567
5568 Instruction *UserI = cast<Instruction>(Val: U.getUser());
5569 if (LoadInst *LI = dyn_cast<LoadInst>(Val: UserI)) {
5570 MemoryUses.push_back(Elt: {&U, LI->getType()});
5571 continue;
5572 }
5573
5574 if (StoreInst *SI = dyn_cast<StoreInst>(Val: UserI)) {
5575 if (U.getOperandNo() != StoreInst::getPointerOperandIndex())
5576 return true; // Storing addr, not into addr.
5577 MemoryUses.push_back(Elt: {&U, SI->getValueOperand()->getType()});
5578 continue;
5579 }
5580
5581 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: UserI)) {
5582 if (U.getOperandNo() != AtomicRMWInst::getPointerOperandIndex())
5583 return true; // Storing addr, not into addr.
5584 MemoryUses.push_back(Elt: {&U, RMW->getValOperand()->getType()});
5585 continue;
5586 }
5587
5588 if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Val: UserI)) {
5589 if (U.getOperandNo() != AtomicCmpXchgInst::getPointerOperandIndex())
5590 return true; // Storing addr, not into addr.
5591 MemoryUses.push_back(Elt: {&U, CmpX->getCompareOperand()->getType()});
5592 continue;
5593 }
5594
5595 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: UserI)) {
5596 SmallVector<Value *, 2> PtrOps;
5597 Type *AccessTy;
5598 if (!TLI.getAddrModeArguments(II, PtrOps, AccessTy))
5599 return true;
5600
5601 if (!find(Range&: PtrOps, Val: U.get()))
5602 return true;
5603
5604 MemoryUses.push_back(Elt: {&U, AccessTy});
5605 continue;
5606 }
5607
5608 if (CallInst *CI = dyn_cast<CallInst>(Val: UserI)) {
5609 if (CI->hasFnAttr(Kind: Attribute::Cold)) {
5610 // If this is a cold call, we can sink the addressing calculation into
5611 // the cold path. See optimizeCallInst
5612 if (!llvm::shouldOptimizeForSize(BB: CI->getParent(), PSI, BFI))
5613 continue;
5614 }
5615
5616 InlineAsm *IA = dyn_cast<InlineAsm>(Val: CI->getCalledOperand());
5617 if (!IA)
5618 return true;
5619
5620 // If this is a memory operand, we're cool, otherwise bail out.
5621 if (!IsOperandAMemoryOperand(CI, IA, OpVal: I, TLI, TRI))
5622 return true;
5623 continue;
5624 }
5625
5626 if (FindAllMemoryUses(I: UserI, MemoryUses, ConsideredInsts, TLI, TRI, OptSize,
5627 PSI, BFI, SeenInsts))
5628 return true;
5629 }
5630
5631 return false;
5632}
5633
5634static bool FindAllMemoryUses(
5635 Instruction *I, SmallVectorImpl<std::pair<Use *, Type *>> &MemoryUses,
5636 const TargetLowering &TLI, const TargetRegisterInfo &TRI, bool OptSize,
5637 ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) {
5638 unsigned SeenInsts = 0;
5639 SmallPtrSet<Instruction *, 16> ConsideredInsts;
5640 return FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI, OptSize,
5641 PSI, BFI, SeenInsts);
5642}
5643
5644
5645/// Return true if Val is already known to be live at the use site that we're
5646/// folding it into. If so, there is no cost to include it in the addressing
5647/// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
5648/// instruction already.
5649bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,
5650 Value *KnownLive1,
5651 Value *KnownLive2) {
5652 // If Val is either of the known-live values, we know it is live!
5653 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
5654 return true;
5655
5656 // All values other than instructions and arguments (e.g. constants) are live.
5657 if (!isa<Instruction>(Val) && !isa<Argument>(Val))
5658 return true;
5659
5660 // If Val is a constant sized alloca in the entry block, it is live, this is
5661 // true because it is just a reference to the stack/frame pointer, which is
5662 // live for the whole function.
5663 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
5664 if (AI->isStaticAlloca())
5665 return true;
5666
5667 // Check to see if this value is already used in the memory instruction's
5668 // block. If so, it's already live into the block at the very least, so we
5669 // can reasonably fold it.
5670 return Val->isUsedInBasicBlock(BB: MemoryInst->getParent());
5671}
5672
5673/// It is possible for the addressing mode of the machine to fold the specified
5674/// instruction into a load or store that ultimately uses it.
5675/// However, the specified instruction has multiple uses.
5676/// Given this, it may actually increase register pressure to fold it
5677/// into the load. For example, consider this code:
5678///
5679/// X = ...
5680/// Y = X+1
5681/// use(Y) -> nonload/store
5682/// Z = Y+1
5683/// load Z
5684///
5685/// In this case, Y has multiple uses, and can be folded into the load of Z
5686/// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
5687/// be live at the use(Y) line. If we don't fold Y into load Z, we use one
5688/// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
5689/// number of computations either.
5690///
5691/// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
5692/// X was live across 'load Z' for other reasons, we actually *would* want to
5693/// fold the addressing mode in the Z case. This would make Y die earlier.
5694bool AddressingModeMatcher::isProfitableToFoldIntoAddressingMode(
5695 Instruction *I, ExtAddrMode &AMBefore, ExtAddrMode &AMAfter) {
5696 if (IgnoreProfitability)
5697 return true;
5698
5699 // AMBefore is the addressing mode before this instruction was folded into it,
5700 // and AMAfter is the addressing mode after the instruction was folded. Get
5701 // the set of registers referenced by AMAfter and subtract out those
5702 // referenced by AMBefore: this is the set of values which folding in this
5703 // address extends the lifetime of.
5704 //
5705 // Note that there are only two potential values being referenced here,
5706 // BaseReg and ScaleReg (global addresses are always available, as are any
5707 // folded immediates).
5708 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
5709
5710 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
5711 // lifetime wasn't extended by adding this instruction.
5712 if (valueAlreadyLiveAtInst(Val: BaseReg, KnownLive1: AMBefore.BaseReg, KnownLive2: AMBefore.ScaledReg))
5713 BaseReg = nullptr;
5714 if (valueAlreadyLiveAtInst(Val: ScaledReg, KnownLive1: AMBefore.BaseReg, KnownLive2: AMBefore.ScaledReg))
5715 ScaledReg = nullptr;
5716
5717 // If folding this instruction (and it's subexprs) didn't extend any live
5718 // ranges, we're ok with it.
5719 if (!BaseReg && !ScaledReg)
5720 return true;
5721
5722 // If all uses of this instruction can have the address mode sunk into them,
5723 // we can remove the addressing mode and effectively trade one live register
5724 // for another (at worst.) In this context, folding an addressing mode into
5725 // the use is just a particularly nice way of sinking it.
5726 SmallVector<std::pair<Use *, Type *>, 16> MemoryUses;
5727 if (FindAllMemoryUses(I, MemoryUses, TLI, TRI, OptSize, PSI, BFI))
5728 return false; // Has a non-memory, non-foldable use!
5729
5730 // Now that we know that all uses of this instruction are part of a chain of
5731 // computation involving only operations that could theoretically be folded
5732 // into a memory use, loop over each of these memory operation uses and see
5733 // if they could *actually* fold the instruction. The assumption is that
5734 // addressing modes are cheap and that duplicating the computation involved
5735 // many times is worthwhile, even on a fastpath. For sinking candidates
5736 // (i.e. cold call sites), this serves as a way to prevent excessive code
5737 // growth since most architectures have some reasonable small and fast way to
5738 // compute an effective address. (i.e LEA on x86)
5739 SmallVector<Instruction *, 32> MatchedAddrModeInsts;
5740 for (const std::pair<Use *, Type *> &Pair : MemoryUses) {
5741 Value *Address = Pair.first->get();
5742 Instruction *UserI = cast<Instruction>(Val: Pair.first->getUser());
5743 Type *AddressAccessTy = Pair.second;
5744 unsigned AS = Address->getType()->getPointerAddressSpace();
5745
5746 // Do a match against the root of this address, ignoring profitability. This
5747 // will tell us if the addressing mode for the memory operation will
5748 // *actually* cover the shared instruction.
5749 ExtAddrMode Result;
5750 std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
5751 0);
5752 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5753 TPT.getRestorationPoint();
5754 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, TRI, LI, getDTFn,
5755 AddressAccessTy, AS, UserI, Result,
5756 InsertedInsts, PromotedInsts, TPT,
5757 LargeOffsetGEP, OptSize, PSI, BFI);
5758 Matcher.IgnoreProfitability = true;
5759 bool Success = Matcher.matchAddr(Addr: Address, Depth: 0);
5760 (void)Success;
5761 assert(Success && "Couldn't select *anything*?");
5762
5763 // The match was to check the profitability, the changes made are not
5764 // part of the original matcher. Therefore, they should be dropped
5765 // otherwise the original matcher will not present the right state.
5766 TPT.rollback(Point: LastKnownGood);
5767
5768 // If the match didn't cover I, then it won't be shared by it.
5769 if (!is_contained(Range&: MatchedAddrModeInsts, Element: I))
5770 return false;
5771
5772 MatchedAddrModeInsts.clear();
5773 }
5774
5775 return true;
5776}
5777
5778/// Return true if the specified values are defined in a
5779/// different basic block than BB.
5780static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
5781 if (Instruction *I = dyn_cast<Instruction>(Val: V))
5782 return I->getParent() != BB;
5783 return false;
5784}
5785
5786// Find an insert position of Addr for MemoryInst. We can't guarantee MemoryInst
5787// is the first instruction that will use Addr. So we need to find the first
5788// user of Addr in current BB.
5789static BasicBlock::iterator findInsertPos(Value *Addr, Instruction *MemoryInst,
5790 Value *SunkAddr) {
5791 if (Addr->hasOneUse())
5792 return MemoryInst->getIterator();
5793
5794 // We already have a SunkAddr in current BB, but we may need to insert cast
5795 // instruction after it.
5796 if (SunkAddr) {
5797 if (Instruction *AddrInst = dyn_cast<Instruction>(Val: SunkAddr))
5798 return std::next(x: AddrInst->getIterator());
5799 }
5800
5801 // Find the first user of Addr in current BB.
5802 Instruction *Earliest = MemoryInst;
5803 for (User *U : Addr->users()) {
5804 Instruction *UserInst = dyn_cast<Instruction>(Val: U);
5805 if (UserInst && UserInst->getParent() == MemoryInst->getParent()) {
5806 if (isa<PHINode>(Val: UserInst) || UserInst->isDebugOrPseudoInst())
5807 continue;
5808 if (UserInst->comesBefore(Other: Earliest))
5809 Earliest = UserInst;
5810 }
5811 }
5812 return Earliest->getIterator();
5813}
5814
5815/// Sink addressing mode computation immediate before MemoryInst if doing so
5816/// can be done without increasing register pressure. The need for the
5817/// register pressure constraint means this can end up being an all or nothing
5818/// decision for all uses of the same addressing computation.
5819///
5820/// Load and Store Instructions often have addressing modes that can do
5821/// significant amounts of computation. As such, instruction selection will try
5822/// to get the load or store to do as much computation as possible for the
5823/// program. The problem is that isel can only see within a single block. As
5824/// such, we sink as much legal addressing mode work into the block as possible.
5825///
5826/// This method is used to optimize both load/store and inline asms with memory
5827/// operands. It's also used to sink addressing computations feeding into cold
5828/// call sites into their (cold) basic block.
5829///
5830/// The motivation for handling sinking into cold blocks is that doing so can
5831/// both enable other address mode sinking (by satisfying the register pressure
5832/// constraint above), and reduce register pressure globally (by removing the
5833/// addressing mode computation from the fast path entirely.).
5834bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
5835 Type *AccessTy, unsigned AddrSpace) {
5836 Value *Repl = Addr;
5837
5838 // Try to collapse single-value PHI nodes. This is necessary to undo
5839 // unprofitable PRE transformations.
5840 SmallVector<Value *, 8> worklist;
5841 SmallPtrSet<Value *, 16> Visited;
5842 worklist.push_back(Elt: Addr);
5843
5844 // Use a worklist to iteratively look through PHI and select nodes, and
5845 // ensure that the addressing mode obtained from the non-PHI/select roots of
5846 // the graph are compatible.
5847 bool PhiOrSelectSeen = false;
5848 SmallVector<Instruction *, 16> AddrModeInsts;
5849 AddressingModeCombiner AddrModes(*DL, Addr);
5850 TypePromotionTransaction TPT(RemovedInsts);
5851 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5852 TPT.getRestorationPoint();
5853 while (!worklist.empty()) {
5854 Value *V = worklist.pop_back_val();
5855
5856 // We allow traversing cyclic Phi nodes.
5857 // In case of success after this loop we ensure that traversing through
5858 // Phi nodes ends up with all cases to compute address of the form
5859 // BaseGV + Base + Scale * Index + Offset
5860 // where Scale and Offset are constans and BaseGV, Base and Index
5861 // are exactly the same Values in all cases.
5862 // It means that BaseGV, Scale and Offset dominate our memory instruction
5863 // and have the same value as they had in address computation represented
5864 // as Phi. So we can safely sink address computation to memory instruction.
5865 if (!Visited.insert(Ptr: V).second)
5866 continue;
5867
5868 // For a PHI node, push all of its incoming values.
5869 if (PHINode *P = dyn_cast<PHINode>(Val: V)) {
5870 append_range(C&: worklist, R: P->incoming_values());
5871 PhiOrSelectSeen = true;
5872 continue;
5873 }
5874 // Similar for select.
5875 if (SelectInst *SI = dyn_cast<SelectInst>(Val: V)) {
5876 worklist.push_back(Elt: SI->getFalseValue());
5877 worklist.push_back(Elt: SI->getTrueValue());
5878 PhiOrSelectSeen = true;
5879 continue;
5880 }
5881
5882 // For non-PHIs, determine the addressing mode being computed. Note that
5883 // the result may differ depending on what other uses our candidate
5884 // addressing instructions might have.
5885 AddrModeInsts.clear();
5886 std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
5887 0);
5888 // Defer the query (and possible computation of) the dom tree to point of
5889 // actual use. It's expected that most address matches don't actually need
5890 // the domtree.
5891 auto getDTFn = [MemoryInst, this]() -> const DominatorTree & {
5892 Function *F = MemoryInst->getParent()->getParent();
5893 return this->getDT(F&: *F);
5894 };
5895 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
5896 V, AccessTy, AS: AddrSpace, MemoryInst, AddrModeInsts, TLI: *TLI, LI: *LI, getDTFn,
5897 TRI: *TRI, InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP, OptSize, PSI,
5898 BFI: BFI.get());
5899
5900 GetElementPtrInst *GEP = LargeOffsetGEP.first;
5901 if (GEP && !NewGEPBases.count(V: GEP)) {
5902 // If splitting the underlying data structure can reduce the offset of a
5903 // GEP, collect the GEP. Skip the GEPs that are the new bases of
5904 // previously split data structures.
5905 LargeOffsetGEPMap[GEP->getPointerOperand()].push_back(Elt: LargeOffsetGEP);
5906 LargeOffsetGEPID.insert(KV: std::make_pair(x&: GEP, y: LargeOffsetGEPID.size()));
5907 }
5908
5909 NewAddrMode.OriginalValue = V;
5910 if (!AddrModes.addNewAddrMode(NewAddrMode))
5911 break;
5912 }
5913
5914 // Try to combine the AddrModes we've collected. If we couldn't collect any,
5915 // or we have multiple but either couldn't combine them or combining them
5916 // wouldn't do anything useful, bail out now.
5917 if (!AddrModes.combineAddrModes()) {
5918 TPT.rollback(Point: LastKnownGood);
5919 return false;
5920 }
5921 bool Modified = TPT.commit();
5922
5923 // Get the combined AddrMode (or the only AddrMode, if we only had one).
5924 ExtAddrMode AddrMode = AddrModes.getAddrMode();
5925
5926 // If all the instructions matched are already in this BB, don't do anything.
5927 // If we saw a Phi node then it is not local definitely, and if we saw a
5928 // select then we want to push the address calculation past it even if it's
5929 // already in this BB.
5930 if (!PhiOrSelectSeen && none_of(Range&: AddrModeInsts, P: [&](Value *V) {
5931 return IsNonLocalValue(V, BB: MemoryInst->getParent());
5932 })) {
5933 LLVM_DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode
5934 << "\n");
5935 return Modified;
5936 }
5937
5938 // Now that we determined the addressing expression we want to use and know
5939 // that we have to sink it into this block. Check to see if we have already
5940 // done this for some other load/store instr in this block. If so, reuse
5941 // the computation. Before attempting reuse, check if the address is valid
5942 // as it may have been erased.
5943
5944 WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
5945
5946 Value *SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
5947 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
5948
5949 // The current BB may be optimized multiple times, we can't guarantee the
5950 // reuse of Addr happens later, call findInsertPos to find an appropriate
5951 // insert position.
5952 auto InsertPos = findInsertPos(Addr, MemoryInst, SunkAddr);
5953
5954 // TODO: Adjust insert point considering (Base|Scaled)Reg if possible.
5955 if (!SunkAddr) {
5956 auto &DT = getDT(F&: *MemoryInst->getFunction());
5957 if ((AddrMode.BaseReg && !DT.dominates(Def: AddrMode.BaseReg, User: &*InsertPos)) ||
5958 (AddrMode.ScaledReg && !DT.dominates(Def: AddrMode.ScaledReg, User: &*InsertPos)))
5959 return Modified;
5960 }
5961
5962 IRBuilder<> Builder(MemoryInst->getParent(), InsertPos);
5963
5964 if (SunkAddr) {
5965 LLVM_DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode
5966 << " for " << *MemoryInst << "\n");
5967 if (SunkAddr->getType() != Addr->getType()) {
5968 if (SunkAddr->getType()->getPointerAddressSpace() !=
5969 Addr->getType()->getPointerAddressSpace() &&
5970 !DL->isNonIntegralPointerType(Ty: Addr->getType())) {
5971 // There are two reasons the address spaces might not match: a no-op
5972 // addrspacecast, or a ptrtoint/inttoptr pair. Either way, we emit a
5973 // ptrtoint/inttoptr pair to ensure we match the original semantics.
5974 // TODO: allow bitcast between different address space pointers with the
5975 // same size.
5976 SunkAddr = Builder.CreatePtrToInt(V: SunkAddr, DestTy: IntPtrTy, Name: "sunkaddr");
5977 SunkAddr =
5978 Builder.CreateIntToPtr(V: SunkAddr, DestTy: Addr->getType(), Name: "sunkaddr");
5979 } else
5980 SunkAddr = Builder.CreatePointerCast(V: SunkAddr, DestTy: Addr->getType());
5981 }
5982 } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() &&
5983 SubtargetInfo->addrSinkUsingGEPs())) {
5984 // By default, we use the GEP-based method when AA is used later. This
5985 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
5986 LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
5987 << " for " << *MemoryInst << "\n");
5988 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
5989
5990 // First, find the pointer.
5991 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
5992 ResultPtr = AddrMode.BaseReg;
5993 AddrMode.BaseReg = nullptr;
5994 }
5995
5996 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
5997 // We can't add more than one pointer together, nor can we scale a
5998 // pointer (both of which seem meaningless).
5999 if (ResultPtr || AddrMode.Scale != 1)
6000 return Modified;
6001
6002 ResultPtr = AddrMode.ScaledReg;
6003 AddrMode.Scale = 0;
6004 }
6005
6006 // It is only safe to sign extend the BaseReg if we know that the math
6007 // required to create it did not overflow before we extend it. Since
6008 // the original IR value was tossed in favor of a constant back when
6009 // the AddrMode was created we need to bail out gracefully if widths
6010 // do not match instead of extending it.
6011 //
6012 // (See below for code to add the scale.)
6013 if (AddrMode.Scale) {
6014 Type *ScaledRegTy = AddrMode.ScaledReg->getType();
6015 if (cast<IntegerType>(Val: IntPtrTy)->getBitWidth() >
6016 cast<IntegerType>(Val: ScaledRegTy)->getBitWidth())
6017 return Modified;
6018 }
6019
6020 GlobalValue *BaseGV = AddrMode.BaseGV;
6021 if (BaseGV != nullptr) {
6022 if (ResultPtr)
6023 return Modified;
6024
6025 if (BaseGV->isThreadLocal()) {
6026 ResultPtr = Builder.CreateThreadLocalAddress(Ptr: BaseGV);
6027 } else {
6028 ResultPtr = BaseGV;
6029 }
6030 }
6031
6032 // If the real base value actually came from an inttoptr, then the matcher
6033 // will look through it and provide only the integer value. In that case,
6034 // use it here.
6035 if (!DL->isNonIntegralPointerType(Ty: Addr->getType())) {
6036 if (!ResultPtr && AddrMode.BaseReg) {
6037 ResultPtr = Builder.CreateIntToPtr(V: AddrMode.BaseReg, DestTy: Addr->getType(),
6038 Name: "sunkaddr");
6039 AddrMode.BaseReg = nullptr;
6040 } else if (!ResultPtr && AddrMode.Scale == 1) {
6041 ResultPtr = Builder.CreateIntToPtr(V: AddrMode.ScaledReg, DestTy: Addr->getType(),
6042 Name: "sunkaddr");
6043 AddrMode.Scale = 0;
6044 }
6045 }
6046
6047 if (!ResultPtr && !AddrMode.BaseReg && !AddrMode.Scale &&
6048 !AddrMode.BaseOffs) {
6049 SunkAddr = Constant::getNullValue(Ty: Addr->getType());
6050 } else if (!ResultPtr) {
6051 return Modified;
6052 } else {
6053 Type *I8PtrTy =
6054 Builder.getPtrTy(AddrSpace: Addr->getType()->getPointerAddressSpace());
6055
6056 // Start with the base register. Do this first so that subsequent address
6057 // matching finds it last, which will prevent it from trying to match it
6058 // as the scaled value in case it happens to be a mul. That would be
6059 // problematic if we've sunk a different mul for the scale, because then
6060 // we'd end up sinking both muls.
6061 if (AddrMode.BaseReg) {
6062 Value *V = AddrMode.BaseReg;
6063 if (V->getType() != IntPtrTy)
6064 V = Builder.CreateIntCast(V, DestTy: IntPtrTy, /*isSigned=*/true, Name: "sunkaddr");
6065
6066 ResultIndex = V;
6067 }
6068
6069 // Add the scale value.
6070 if (AddrMode.Scale) {
6071 Value *V = AddrMode.ScaledReg;
6072 if (V->getType() == IntPtrTy) {
6073 // done.
6074 } else {
6075 assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <
6076 cast<IntegerType>(V->getType())->getBitWidth() &&
6077 "We can't transform if ScaledReg is too narrow");
6078 V = Builder.CreateTrunc(V, DestTy: IntPtrTy, Name: "sunkaddr");
6079 }
6080
6081 if (AddrMode.Scale != 1)
6082 V = Builder.CreateMul(
6083 LHS: V, RHS: ConstantInt::getSigned(Ty: IntPtrTy, V: AddrMode.Scale), Name: "sunkaddr");
6084 if (ResultIndex)
6085 ResultIndex = Builder.CreateAdd(LHS: ResultIndex, RHS: V, Name: "sunkaddr");
6086 else
6087 ResultIndex = V;
6088 }
6089
6090 // Add in the Base Offset if present.
6091 if (AddrMode.BaseOffs) {
6092 Value *V = ConstantInt::getSigned(Ty: IntPtrTy, V: AddrMode.BaseOffs);
6093 if (ResultIndex) {
6094 // We need to add this separately from the scale above to help with
6095 // SDAG consecutive load/store merging.
6096 if (ResultPtr->getType() != I8PtrTy)
6097 ResultPtr = Builder.CreatePointerCast(V: ResultPtr, DestTy: I8PtrTy);
6098 ResultPtr = Builder.CreatePtrAdd(Ptr: ResultPtr, Offset: ResultIndex, Name: "sunkaddr",
6099 NW: AddrMode.InBounds);
6100 }
6101
6102 ResultIndex = V;
6103 }
6104
6105 if (!ResultIndex) {
6106 auto PtrInst = dyn_cast<Instruction>(Val: ResultPtr);
6107 // We know that we have a pointer without any offsets. If this pointer
6108 // originates from a different basic block than the current one, we
6109 // must be able to recreate it in the current basic block.
6110 // We do not support the recreation of any instructions yet.
6111 if (PtrInst && PtrInst->getParent() != MemoryInst->getParent())
6112 return Modified;
6113 SunkAddr = ResultPtr;
6114 } else {
6115 if (ResultPtr->getType() != I8PtrTy)
6116 ResultPtr = Builder.CreatePointerCast(V: ResultPtr, DestTy: I8PtrTy);
6117 SunkAddr = Builder.CreatePtrAdd(Ptr: ResultPtr, Offset: ResultIndex, Name: "sunkaddr",
6118 NW: AddrMode.InBounds);
6119 }
6120
6121 if (SunkAddr->getType() != Addr->getType()) {
6122 if (SunkAddr->getType()->getPointerAddressSpace() !=
6123 Addr->getType()->getPointerAddressSpace() &&
6124 !DL->isNonIntegralPointerType(Ty: Addr->getType())) {
6125 // There are two reasons the address spaces might not match: a no-op
6126 // addrspacecast, or a ptrtoint/inttoptr pair. Either way, we emit a
6127 // ptrtoint/inttoptr pair to ensure we match the original semantics.
6128 // TODO: allow bitcast between different address space pointers with
6129 // the same size.
6130 SunkAddr = Builder.CreatePtrToInt(V: SunkAddr, DestTy: IntPtrTy, Name: "sunkaddr");
6131 SunkAddr =
6132 Builder.CreateIntToPtr(V: SunkAddr, DestTy: Addr->getType(), Name: "sunkaddr");
6133 } else
6134 SunkAddr = Builder.CreatePointerCast(V: SunkAddr, DestTy: Addr->getType());
6135 }
6136 }
6137 } else {
6138 // We'd require a ptrtoint/inttoptr down the line, which we can't do for
6139 // non-integral pointers, so in that case bail out now.
6140 Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
6141 Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
6142 PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(Val: BaseTy);
6143 PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(Val: ScaleTy);
6144 if (DL->isNonIntegralPointerType(Ty: Addr->getType()) ||
6145 (BasePtrTy && DL->isNonIntegralPointerType(PT: BasePtrTy)) ||
6146 (ScalePtrTy && DL->isNonIntegralPointerType(PT: ScalePtrTy)) ||
6147 (AddrMode.BaseGV &&
6148 DL->isNonIntegralPointerType(PT: AddrMode.BaseGV->getType())))
6149 return Modified;
6150
6151 LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
6152 << " for " << *MemoryInst << "\n");
6153 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
6154 Value *Result = nullptr;
6155
6156 // Start with the base register. Do this first so that subsequent address
6157 // matching finds it last, which will prevent it from trying to match it
6158 // as the scaled value in case it happens to be a mul. That would be
6159 // problematic if we've sunk a different mul for the scale, because then
6160 // we'd end up sinking both muls.
6161 if (AddrMode.BaseReg) {
6162 Value *V = AddrMode.BaseReg;
6163 if (V->getType()->isPointerTy())
6164 V = Builder.CreatePtrToInt(V, DestTy: IntPtrTy, Name: "sunkaddr");
6165 if (V->getType() != IntPtrTy)
6166 V = Builder.CreateIntCast(V, DestTy: IntPtrTy, /*isSigned=*/true, Name: "sunkaddr");
6167 Result = V;
6168 }
6169
6170 // Add the scale value.
6171 if (AddrMode.Scale) {
6172 Value *V = AddrMode.ScaledReg;
6173 if (V->getType() == IntPtrTy) {
6174 // done.
6175 } else if (V->getType()->isPointerTy()) {
6176 V = Builder.CreatePtrToInt(V, DestTy: IntPtrTy, Name: "sunkaddr");
6177 } else if (cast<IntegerType>(Val: IntPtrTy)->getBitWidth() <
6178 cast<IntegerType>(Val: V->getType())->getBitWidth()) {
6179 V = Builder.CreateTrunc(V, DestTy: IntPtrTy, Name: "sunkaddr");
6180 } else {
6181 // It is only safe to sign extend the BaseReg if we know that the math
6182 // required to create it did not overflow before we extend it. Since
6183 // the original IR value was tossed in favor of a constant back when
6184 // the AddrMode was created we need to bail out gracefully if widths
6185 // do not match instead of extending it.
6186 Instruction *I = dyn_cast_or_null<Instruction>(Val: Result);
6187 if (I && (Result != AddrMode.BaseReg))
6188 I->eraseFromParent();
6189 return Modified;
6190 }
6191 if (AddrMode.Scale != 1)
6192 V = Builder.CreateMul(
6193 LHS: V, RHS: ConstantInt::getSigned(Ty: IntPtrTy, V: AddrMode.Scale), Name: "sunkaddr");
6194 if (Result)
6195 Result = Builder.CreateAdd(LHS: Result, RHS: V, Name: "sunkaddr");
6196 else
6197 Result = V;
6198 }
6199
6200 // Add in the BaseGV if present.
6201 GlobalValue *BaseGV = AddrMode.BaseGV;
6202 if (BaseGV != nullptr) {
6203 Value *BaseGVPtr;
6204 if (BaseGV->isThreadLocal()) {
6205 BaseGVPtr = Builder.CreateThreadLocalAddress(Ptr: BaseGV);
6206 } else {
6207 BaseGVPtr = BaseGV;
6208 }
6209 Value *V = Builder.CreatePtrToInt(V: BaseGVPtr, DestTy: IntPtrTy, Name: "sunkaddr");
6210 if (Result)
6211 Result = Builder.CreateAdd(LHS: Result, RHS: V, Name: "sunkaddr");
6212 else
6213 Result = V;
6214 }
6215
6216 // Add in the Base Offset if present.
6217 if (AddrMode.BaseOffs) {
6218 Value *V = ConstantInt::getSigned(Ty: IntPtrTy, V: AddrMode.BaseOffs);
6219 if (Result)
6220 Result = Builder.CreateAdd(LHS: Result, RHS: V, Name: "sunkaddr");
6221 else
6222 Result = V;
6223 }
6224
6225 if (!Result)
6226 SunkAddr = Constant::getNullValue(Ty: Addr->getType());
6227 else
6228 SunkAddr = Builder.CreateIntToPtr(V: Result, DestTy: Addr->getType(), Name: "sunkaddr");
6229 }
6230
6231 MemoryInst->replaceUsesOfWith(From: Repl, To: SunkAddr);
6232 // Store the newly computed address into the cache. In the case we reused a
6233 // value, this should be idempotent.
6234 SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
6235
6236 // If we have no uses, recursively delete the value and all dead instructions
6237 // using it.
6238 if (Repl->use_empty()) {
6239 resetIteratorIfInvalidatedWhileCalling(BB: CurInstIterator->getParent(), f: [&]() {
6240 RecursivelyDeleteTriviallyDeadInstructions(
6241 V: Repl, TLI: TLInfo, MSSAU: nullptr,
6242 AboutToDeleteCallback: [&](Value *V) { removeAllAssertingVHReferences(V); });
6243 });
6244 }
6245 ++NumMemoryInsts;
6246 return true;
6247}
6248
6249/// Rewrite GEP input to gather/scatter to enable SelectionDAGBuilder to find
6250/// a uniform base to use for ISD::MGATHER/MSCATTER. SelectionDAGBuilder can
6251/// only handle a 2 operand GEP in the same basic block or a splat constant
6252/// vector. The 2 operands to the GEP must have a scalar pointer and a vector
6253/// index.
6254///
6255/// If the existing GEP has a vector base pointer that is splat, we can look
6256/// through the splat to find the scalar pointer. If we can't find a scalar
6257/// pointer there's nothing we can do.
6258///
6259/// If we have a GEP with more than 2 indices where the middle indices are all
6260/// zeroes, we can replace it with 2 GEPs where the second has 2 operands.
6261///
6262/// If the final index isn't a vector or is a splat, we can emit a scalar GEP
6263/// followed by a GEP with an all zeroes vector index. This will enable
6264/// SelectionDAGBuilder to use the scalar GEP as the uniform base and have a
6265/// zero index.
6266bool CodeGenPrepare::optimizeGatherScatterInst(Instruction *MemoryInst,
6267 Value *Ptr) {
6268 Value *NewAddr;
6269
6270 if (const auto *GEP = dyn_cast<GetElementPtrInst>(Val: Ptr)) {
6271 // Don't optimize GEPs that don't have indices.
6272 if (!GEP->hasIndices())
6273 return false;
6274
6275 // If the GEP and the gather/scatter aren't in the same BB, don't optimize.
6276 // FIXME: We should support this by sinking the GEP.
6277 if (MemoryInst->getParent() != GEP->getParent())
6278 return false;
6279
6280 SmallVector<Value *, 2> Ops(GEP->operands());
6281
6282 bool RewriteGEP = false;
6283
6284 if (Ops[0]->getType()->isVectorTy()) {
6285 Ops[0] = getSplatValue(V: Ops[0]);
6286 if (!Ops[0])
6287 return false;
6288 RewriteGEP = true;
6289 }
6290
6291 unsigned FinalIndex = Ops.size() - 1;
6292
6293 // Ensure all but the last index is 0.
6294 // FIXME: This isn't strictly required. All that's required is that they are
6295 // all scalars or splats.
6296 for (unsigned i = 1; i < FinalIndex; ++i) {
6297 auto *C = dyn_cast<Constant>(Val: Ops[i]);
6298 if (!C)
6299 return false;
6300 if (isa<VectorType>(Val: C->getType()))
6301 C = C->getSplatValue();
6302 auto *CI = dyn_cast_or_null<ConstantInt>(Val: C);
6303 if (!CI || !CI->isZero())
6304 return false;
6305 // Scalarize the index if needed.
6306 Ops[i] = CI;
6307 }
6308
6309 // Try to scalarize the final index.
6310 if (Ops[FinalIndex]->getType()->isVectorTy()) {
6311 if (Value *V = getSplatValue(V: Ops[FinalIndex])) {
6312 auto *C = dyn_cast<ConstantInt>(Val: V);
6313 // Don't scalarize all zeros vector.
6314 if (!C || !C->isZero()) {
6315 Ops[FinalIndex] = V;
6316 RewriteGEP = true;
6317 }
6318 }
6319 }
6320
6321 // If we made any changes or the we have extra operands, we need to generate
6322 // new instructions.
6323 if (!RewriteGEP && Ops.size() == 2)
6324 return false;
6325
6326 auto NumElts = cast<VectorType>(Val: Ptr->getType())->getElementCount();
6327
6328 IRBuilder<> Builder(MemoryInst);
6329
6330 Type *SourceTy = GEP->getSourceElementType();
6331 Type *ScalarIndexTy = DL->getIndexType(PtrTy: Ops[0]->getType()->getScalarType());
6332
6333 // If the final index isn't a vector, emit a scalar GEP containing all ops
6334 // and a vector GEP with all zeroes final index.
6335 if (!Ops[FinalIndex]->getType()->isVectorTy()) {
6336 NewAddr = Builder.CreateGEP(Ty: SourceTy, Ptr: Ops[0], IdxList: ArrayRef(Ops).drop_front());
6337 auto *IndexTy = VectorType::get(ElementType: ScalarIndexTy, EC: NumElts);
6338 auto *SecondTy = GetElementPtrInst::getIndexedType(
6339 Ty: SourceTy, IdxList: ArrayRef(Ops).drop_front());
6340 NewAddr =
6341 Builder.CreateGEP(Ty: SecondTy, Ptr: NewAddr, IdxList: Constant::getNullValue(Ty: IndexTy));
6342 } else {
6343 Value *Base = Ops[0];
6344 Value *Index = Ops[FinalIndex];
6345
6346 // Create a scalar GEP if there are more than 2 operands.
6347 if (Ops.size() != 2) {
6348 // Replace the last index with 0.
6349 Ops[FinalIndex] =
6350 Constant::getNullValue(Ty: Ops[FinalIndex]->getType()->getScalarType());
6351 Base = Builder.CreateGEP(Ty: SourceTy, Ptr: Base, IdxList: ArrayRef(Ops).drop_front());
6352 SourceTy = GetElementPtrInst::getIndexedType(
6353 Ty: SourceTy, IdxList: ArrayRef(Ops).drop_front());
6354 }
6355
6356 // Now create the GEP with scalar pointer and vector index.
6357 NewAddr = Builder.CreateGEP(Ty: SourceTy, Ptr: Base, IdxList: Index);
6358 }
6359 } else if (!isa<Constant>(Val: Ptr)) {
6360 // Not a GEP, maybe its a splat and we can create a GEP to enable
6361 // SelectionDAGBuilder to use it as a uniform base.
6362 Value *V = getSplatValue(V: Ptr);
6363 if (!V)
6364 return false;
6365
6366 auto NumElts = cast<VectorType>(Val: Ptr->getType())->getElementCount();
6367
6368 IRBuilder<> Builder(MemoryInst);
6369
6370 // Emit a vector GEP with a scalar pointer and all 0s vector index.
6371 Type *ScalarIndexTy = DL->getIndexType(PtrTy: V->getType()->getScalarType());
6372 auto *IndexTy = VectorType::get(ElementType: ScalarIndexTy, EC: NumElts);
6373 Type *ScalarTy;
6374 if (cast<IntrinsicInst>(Val: MemoryInst)->getIntrinsicID() ==
6375 Intrinsic::masked_gather) {
6376 ScalarTy = MemoryInst->getType()->getScalarType();
6377 } else {
6378 assert(cast<IntrinsicInst>(MemoryInst)->getIntrinsicID() ==
6379 Intrinsic::masked_scatter);
6380 ScalarTy = MemoryInst->getOperand(i: 0)->getType()->getScalarType();
6381 }
6382 NewAddr = Builder.CreateGEP(Ty: ScalarTy, Ptr: V, IdxList: Constant::getNullValue(Ty: IndexTy));
6383 } else {
6384 // Constant, SelectionDAGBuilder knows to check if its a splat.
6385 return false;
6386 }
6387
6388 MemoryInst->replaceUsesOfWith(From: Ptr, To: NewAddr);
6389
6390 // If we have no uses, recursively delete the value and all dead instructions
6391 // using it.
6392 if (Ptr->use_empty())
6393 RecursivelyDeleteTriviallyDeadInstructions(
6394 V: Ptr, TLI: TLInfo, MSSAU: nullptr,
6395 AboutToDeleteCallback: [&](Value *V) { removeAllAssertingVHReferences(V); });
6396
6397 return true;
6398}
6399
6400// This is a helper for CodeGenPrepare::optimizeMulWithOverflow.
6401// Check the pattern we are interested in where there are maximum 2 uses
6402// of the intrinsic which are the extract instructions.
6403static bool matchOverflowPattern(Instruction *&I, ExtractValueInst *&MulExtract,
6404 ExtractValueInst *&OverflowExtract) {
6405 // Bail out if it's more than 2 users:
6406 if (I->hasNUsesOrMore(N: 3))
6407 return false;
6408
6409 for (User *U : I->users()) {
6410 auto *Extract = dyn_cast<ExtractValueInst>(Val: U);
6411 if (!Extract || Extract->getNumIndices() != 1)
6412 return false;
6413
6414 unsigned Index = Extract->getIndices()[0];
6415 if (Index == 0)
6416 MulExtract = Extract;
6417 else if (Index == 1)
6418 OverflowExtract = Extract;
6419 else
6420 return false;
6421 }
6422 return true;
6423}
6424
6425// Rewrite the mul_with_overflow intrinsic by checking if both of the
6426// operands' value ranges are within the legal type. If so, we can optimize the
6427// multiplication algorithm. This code is supposed to be written during the step
6428// of type legalization, but given that we need to reconstruct the IR which is
6429// not doable there, we do it here.
6430// The IR after the optimization will look like:
6431// entry:
6432// if signed:
6433// ( (lhs_lo>>BW-1) ^ lhs_hi) || ( (rhs_lo>>BW-1) ^ rhs_hi) ? overflow,
6434// overflow_no
6435// else:
6436// (lhs_hi != 0) || (rhs_hi != 0) ? overflow, overflow_no
6437// overflow_no:
6438// overflow:
6439// overflow.res:
6440// \returns true if optimization was applied
6441// TODO: This optimization can be further improved to optimize branching on
6442// overflow where the 'overflow_no' BB can branch directly to the false
6443// successor of overflow, but that would add additional complexity so we leave
6444// it for future work.
6445bool CodeGenPrepare::optimizeMulWithOverflow(Instruction *I, bool IsSigned,
6446 ModifyDT &ModifiedDT) {
6447 // Check if target supports this optimization.
6448 if (!TLI->shouldOptimizeMulOverflowWithZeroHighBits(
6449 Context&: I->getContext(),
6450 VT: TLI->getValueType(DL: *DL, Ty: I->getType()->getContainedType(i: 0))))
6451 return false;
6452
6453 ExtractValueInst *MulExtract = nullptr, *OverflowExtract = nullptr;
6454 if (!matchOverflowPattern(I, MulExtract, OverflowExtract))
6455 return false;
6456
6457 // Keep track of the instruction to stop reoptimizing it again.
6458 InsertedInsts.insert(Ptr: I);
6459
6460 Value *LHS = I->getOperand(i: 0);
6461 Value *RHS = I->getOperand(i: 1);
6462 Type *Ty = LHS->getType();
6463 unsigned VTHalfBitWidth = Ty->getScalarSizeInBits() / 2;
6464 Type *LegalTy = Ty->getWithNewBitWidth(NewBitWidth: VTHalfBitWidth);
6465
6466 // New BBs:
6467 BasicBlock *OverflowEntryBB = I->getParent()->splitBasicBlockBefore(I, BBName: "");
6468 OverflowEntryBB->takeName(V: I->getParent());
6469 // Keep the 'br' instruction that is generated as a result of the split to be
6470 // erased/replaced later.
6471 Instruction *OldTerminator = OverflowEntryBB->getTerminator();
6472 BasicBlock *NoOverflowBB =
6473 BasicBlock::Create(Context&: I->getContext(), Name: "overflow.no", Parent: I->getFunction());
6474 NoOverflowBB->moveAfter(MovePos: OverflowEntryBB);
6475 BasicBlock *OverflowBB =
6476 BasicBlock::Create(Context&: I->getContext(), Name: "overflow", Parent: I->getFunction());
6477 OverflowBB->moveAfter(MovePos: NoOverflowBB);
6478
6479 // BB overflow.entry:
6480 IRBuilder<> Builder(OverflowEntryBB);
6481 // Extract low and high halves of LHS:
6482 Value *LoLHS = Builder.CreateTrunc(V: LHS, DestTy: LegalTy, Name: "lo.lhs");
6483 Value *HiLHS = Builder.CreateLShr(LHS, RHS: VTHalfBitWidth, Name: "lhs.lsr");
6484 HiLHS = Builder.CreateTrunc(V: HiLHS, DestTy: LegalTy, Name: "hi.lhs");
6485
6486 // Extract low and high halves of RHS:
6487 Value *LoRHS = Builder.CreateTrunc(V: RHS, DestTy: LegalTy, Name: "lo.rhs");
6488 Value *HiRHS = Builder.CreateLShr(LHS: RHS, RHS: VTHalfBitWidth, Name: "rhs.lsr");
6489 HiRHS = Builder.CreateTrunc(V: HiRHS, DestTy: LegalTy, Name: "hi.rhs");
6490
6491 Value *IsAnyBitTrue;
6492 if (IsSigned) {
6493 Value *SignLoLHS =
6494 Builder.CreateAShr(LHS: LoLHS, RHS: VTHalfBitWidth - 1, Name: "sign.lo.lhs");
6495 Value *SignLoRHS =
6496 Builder.CreateAShr(LHS: LoRHS, RHS: VTHalfBitWidth - 1, Name: "sign.lo.rhs");
6497 Value *XorLHS = Builder.CreateXor(LHS: HiLHS, RHS: SignLoLHS);
6498 Value *XorRHS = Builder.CreateXor(LHS: HiRHS, RHS: SignLoRHS);
6499 Value *Or = Builder.CreateOr(LHS: XorLHS, RHS: XorRHS, Name: "or.lhs.rhs");
6500 IsAnyBitTrue = Builder.CreateCmp(Pred: ICmpInst::ICMP_NE, LHS: Or,
6501 RHS: ConstantInt::getNullValue(Ty: Or->getType()));
6502 } else {
6503 Value *CmpLHS = Builder.CreateCmp(Pred: ICmpInst::ICMP_NE, LHS: HiLHS,
6504 RHS: ConstantInt::getNullValue(Ty: LegalTy));
6505 Value *CmpRHS = Builder.CreateCmp(Pred: ICmpInst::ICMP_NE, LHS: HiRHS,
6506 RHS: ConstantInt::getNullValue(Ty: LegalTy));
6507 IsAnyBitTrue = Builder.CreateOr(LHS: CmpLHS, RHS: CmpRHS, Name: "or.lhs.rhs");
6508 }
6509 Builder.CreateCondBr(Cond: IsAnyBitTrue, True: OverflowBB, False: NoOverflowBB);
6510
6511 // BB overflow.no:
6512 Builder.SetInsertPoint(NoOverflowBB);
6513 Value *ExtLoLHS, *ExtLoRHS;
6514 if (IsSigned) {
6515 ExtLoLHS = Builder.CreateSExt(V: LoLHS, DestTy: Ty, Name: "lo.lhs.ext");
6516 ExtLoRHS = Builder.CreateSExt(V: LoRHS, DestTy: Ty, Name: "lo.rhs.ext");
6517 } else {
6518 ExtLoLHS = Builder.CreateZExt(V: LoLHS, DestTy: Ty, Name: "lo.lhs.ext");
6519 ExtLoRHS = Builder.CreateZExt(V: LoRHS, DestTy: Ty, Name: "lo.rhs.ext");
6520 }
6521
6522 Value *Mul = Builder.CreateMul(LHS: ExtLoLHS, RHS: ExtLoRHS, Name: "mul.overflow.no");
6523
6524 // Create the 'overflow.res' BB to merge the results of
6525 // the two paths:
6526 BasicBlock *OverflowResBB = I->getParent();
6527 OverflowResBB->setName("overflow.res");
6528
6529 // BB overflow.no: jump to overflow.res BB
6530 Builder.CreateBr(Dest: OverflowResBB);
6531 // No we don't need the old terminator in overflow.entry BB, erase it:
6532 OldTerminator->eraseFromParent();
6533
6534 // BB overflow.res:
6535 Builder.SetInsertPoint(TheBB: OverflowResBB, IP: OverflowResBB->getFirstInsertionPt());
6536 // Create PHI nodes to merge results from no.overflow BB and overflow BB to
6537 // replace the extract instructions.
6538 PHINode *OverflowResPHI = Builder.CreatePHI(Ty, NumReservedValues: 2),
6539 *OverflowFlagPHI =
6540 Builder.CreatePHI(Ty: IntegerType::getInt1Ty(C&: I->getContext()), NumReservedValues: 2);
6541
6542 // Add the incoming values from no.overflow BB and later from overflow BB.
6543 OverflowResPHI->addIncoming(V: Mul, BB: NoOverflowBB);
6544 OverflowFlagPHI->addIncoming(V: ConstantInt::getFalse(Context&: I->getContext()),
6545 BB: NoOverflowBB);
6546
6547 // Replace all users of MulExtract and OverflowExtract to use the PHI nodes.
6548 if (MulExtract) {
6549 MulExtract->replaceAllUsesWith(V: OverflowResPHI);
6550 MulExtract->eraseFromParent();
6551 }
6552 if (OverflowExtract) {
6553 OverflowExtract->replaceAllUsesWith(V: OverflowFlagPHI);
6554 OverflowExtract->eraseFromParent();
6555 }
6556
6557 // Remove the intrinsic from parent (overflow.res BB) as it will be part of
6558 // overflow BB
6559 I->removeFromParent();
6560 // BB overflow:
6561 I->insertInto(ParentBB: OverflowBB, It: OverflowBB->end());
6562 Builder.SetInsertPoint(TheBB: OverflowBB, IP: OverflowBB->end());
6563 Value *MulOverflow = Builder.CreateExtractValue(Agg: I, Idxs: {0}, Name: "mul.overflow");
6564 Value *OverflowFlag = Builder.CreateExtractValue(Agg: I, Idxs: {1}, Name: "overflow.flag");
6565 Builder.CreateBr(Dest: OverflowResBB);
6566
6567 // Add The Extracted values to the PHINodes in the overflow.res BB.
6568 OverflowResPHI->addIncoming(V: MulOverflow, BB: OverflowBB);
6569 OverflowFlagPHI->addIncoming(V: OverflowFlag, BB: OverflowBB);
6570
6571 ModifiedDT = ModifyDT::ModifyBBDT;
6572 return true;
6573}
6574
6575/// If there are any memory operands, use OptimizeMemoryInst to sink their
6576/// address computing into the block when possible / profitable.
6577bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
6578 bool MadeChange = false;
6579
6580 const TargetRegisterInfo *TRI =
6581 TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
6582 TargetLowering::AsmOperandInfoVector TargetConstraints =
6583 TLI->ParseConstraints(DL: *DL, TRI, Call: *CS);
6584 unsigned ArgNo = 0;
6585 for (TargetLowering::AsmOperandInfo &OpInfo : TargetConstraints) {
6586 // Compute the constraint code and ConstraintType to use.
6587 TLI->ComputeConstraintToUse(OpInfo, Op: SDValue());
6588
6589 // TODO: Also handle C_Address?
6590 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6591 OpInfo.isIndirect) {
6592 Value *OpVal = CS->getArgOperand(i: ArgNo++);
6593 MadeChange |= optimizeMemoryInst(MemoryInst: CS, Addr: OpVal, AccessTy: OpVal->getType(), AddrSpace: ~0u);
6594 } else if (OpInfo.Type == InlineAsm::isInput)
6595 ArgNo++;
6596 }
6597
6598 return MadeChange;
6599}
6600
6601/// Check if all the uses of \p Val are equivalent (or free) zero or
6602/// sign extensions.
6603static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
6604 assert(!Val->use_empty() && "Input must have at least one use");
6605 const Instruction *FirstUser = cast<Instruction>(Val: *Val->user_begin());
6606 bool IsSExt = isa<SExtInst>(Val: FirstUser);
6607 Type *ExtTy = FirstUser->getType();
6608 for (const User *U : Val->users()) {
6609 const Instruction *UI = cast<Instruction>(Val: U);
6610 if ((IsSExt && !isa<SExtInst>(Val: UI)) || (!IsSExt && !isa<ZExtInst>(Val: UI)))
6611 return false;
6612 Type *CurTy = UI->getType();
6613 // Same input and output types: Same instruction after CSE.
6614 if (CurTy == ExtTy)
6615 continue;
6616
6617 // If IsSExt is true, we are in this situation:
6618 // a = Val
6619 // b = sext ty1 a to ty2
6620 // c = sext ty1 a to ty3
6621 // Assuming ty2 is shorter than ty3, this could be turned into:
6622 // a = Val
6623 // b = sext ty1 a to ty2
6624 // c = sext ty2 b to ty3
6625 // However, the last sext is not free.
6626 if (IsSExt)
6627 return false;
6628
6629 // This is a ZExt, maybe this is free to extend from one type to another.
6630 // In that case, we would not account for a different use.
6631 Type *NarrowTy;
6632 Type *LargeTy;
6633 if (ExtTy->getScalarType()->getIntegerBitWidth() >
6634 CurTy->getScalarType()->getIntegerBitWidth()) {
6635 NarrowTy = CurTy;
6636 LargeTy = ExtTy;
6637 } else {
6638 NarrowTy = ExtTy;
6639 LargeTy = CurTy;
6640 }
6641
6642 if (!TLI.isZExtFree(FromTy: NarrowTy, ToTy: LargeTy))
6643 return false;
6644 }
6645 // All uses are the same or can be derived from one another for free.
6646 return true;
6647}
6648
6649/// Try to speculatively promote extensions in \p Exts and continue
6650/// promoting through newly promoted operands recursively as far as doing so is
6651/// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
6652/// When some promotion happened, \p TPT contains the proper state to revert
6653/// them.
6654///
6655/// \return true if some promotion happened, false otherwise.
6656bool CodeGenPrepare::tryToPromoteExts(
6657 TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
6658 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
6659 unsigned CreatedInstsCost) {
6660 bool Promoted = false;
6661
6662 // Iterate over all the extensions to try to promote them.
6663 for (auto *I : Exts) {
6664 // Early check if we directly have ext(load).
6665 if (isa<LoadInst>(Val: I->getOperand(i: 0))) {
6666 ProfitablyMovedExts.push_back(Elt: I);
6667 continue;
6668 }
6669
6670 // Check whether or not we want to do any promotion. The reason we have
6671 // this check inside the for loop is to catch the case where an extension
6672 // is directly fed by a load because in such case the extension can be moved
6673 // up without any promotion on its operands.
6674 if (!TLI->enableExtLdPromotion() || DisableExtLdPromotion)
6675 return false;
6676
6677 // Get the action to perform the promotion.
6678 TypePromotionHelper::Action TPH =
6679 TypePromotionHelper::getAction(Ext: I, InsertedInsts, TLI: *TLI, PromotedInsts);
6680 // Check if we can promote.
6681 if (!TPH) {
6682 // Save the current extension as we cannot move up through its operand.
6683 ProfitablyMovedExts.push_back(Elt: I);
6684 continue;
6685 }
6686
6687 // Save the current state.
6688 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
6689 TPT.getRestorationPoint();
6690 SmallVector<Instruction *, 4> NewExts;
6691 unsigned NewCreatedInstsCost = 0;
6692 unsigned ExtCost = !TLI->isExtFree(I);
6693 // Promote.
6694 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
6695 &NewExts, nullptr, *TLI);
6696 assert(PromotedVal &&
6697 "TypePromotionHelper should have filtered out those cases");
6698
6699 // We would be able to merge only one extension in a load.
6700 // Therefore, if we have more than 1 new extension we heuristically
6701 // cut this search path, because it means we degrade the code quality.
6702 // With exactly 2, the transformation is neutral, because we will merge
6703 // one extension but leave one. However, we optimistically keep going,
6704 // because the new extension may be removed too. Also avoid replacing a
6705 // single free extension with multiple extensions, as this increases the
6706 // number of IR instructions while not providing any savings.
6707 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
6708 // FIXME: It would be possible to propagate a negative value instead of
6709 // conservatively ceiling it to 0.
6710 TotalCreatedInstsCost =
6711 std::max(a: (long long)0, b: (TotalCreatedInstsCost - ExtCost));
6712 if (!StressExtLdPromotion &&
6713 (TotalCreatedInstsCost > 1 ||
6714 !isPromotedInstructionLegal(TLI: *TLI, DL: *DL, Val: PromotedVal) ||
6715 (ExtCost == 0 && NewExts.size() > 1))) {
6716 // This promotion is not profitable, rollback to the previous state, and
6717 // save the current extension in ProfitablyMovedExts as the latest
6718 // speculative promotion turned out to be unprofitable.
6719 TPT.rollback(Point: LastKnownGood);
6720 ProfitablyMovedExts.push_back(Elt: I);
6721 continue;
6722 }
6723 // Continue promoting NewExts as far as doing so is profitable.
6724 SmallVector<Instruction *, 2> NewlyMovedExts;
6725 (void)tryToPromoteExts(TPT, Exts: NewExts, ProfitablyMovedExts&: NewlyMovedExts, CreatedInstsCost: TotalCreatedInstsCost);
6726 bool NewPromoted = false;
6727 for (auto *ExtInst : NewlyMovedExts) {
6728 Instruction *MovedExt = cast<Instruction>(Val: ExtInst);
6729 Value *ExtOperand = MovedExt->getOperand(i: 0);
6730 // If we have reached to a load, we need this extra profitability check
6731 // as it could potentially be merged into an ext(load).
6732 if (isa<LoadInst>(Val: ExtOperand) &&
6733 !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
6734 (ExtOperand->hasOneUse() || hasSameExtUse(Val: ExtOperand, TLI: *TLI))))
6735 continue;
6736
6737 ProfitablyMovedExts.push_back(Elt: MovedExt);
6738 NewPromoted = true;
6739 }
6740
6741 // If none of speculative promotions for NewExts is profitable, rollback
6742 // and save the current extension (I) as the last profitable extension.
6743 if (!NewPromoted) {
6744 TPT.rollback(Point: LastKnownGood);
6745 ProfitablyMovedExts.push_back(Elt: I);
6746 continue;
6747 }
6748 // The promotion is profitable.
6749 Promoted = true;
6750 }
6751 return Promoted;
6752}
6753
6754/// Merging redundant sexts when one is dominating the other.
6755bool CodeGenPrepare::mergeSExts(Function &F) {
6756 bool Changed = false;
6757 for (auto &Entry : ValToSExtendedUses) {
6758 SExts &Insts = Entry.second;
6759 SExts CurPts;
6760 for (Instruction *Inst : Insts) {
6761 if (RemovedInsts.count(Ptr: Inst) || !isa<SExtInst>(Val: Inst) ||
6762 Inst->getOperand(i: 0) != Entry.first)
6763 continue;
6764 bool inserted = false;
6765 for (auto &Pt : CurPts) {
6766 if (getDT(F).dominates(Def: Inst, User: Pt)) {
6767 replaceAllUsesWith(Old: Pt, New: Inst, FreshBBs, IsHuge: IsHugeFunc);
6768 RemovedInsts.insert(Ptr: Pt);
6769 Pt->removeFromParent();
6770 Pt = Inst;
6771 inserted = true;
6772 Changed = true;
6773 break;
6774 }
6775 if (!getDT(F).dominates(Def: Pt, User: Inst))
6776 // Give up if we need to merge in a common dominator as the
6777 // experiments show it is not profitable.
6778 continue;
6779 replaceAllUsesWith(Old: Inst, New: Pt, FreshBBs, IsHuge: IsHugeFunc);
6780 RemovedInsts.insert(Ptr: Inst);
6781 Inst->removeFromParent();
6782 inserted = true;
6783 Changed = true;
6784 break;
6785 }
6786 if (!inserted)
6787 CurPts.push_back(Elt: Inst);
6788 }
6789 }
6790 return Changed;
6791}
6792
6793// Splitting large data structures so that the GEPs accessing them can have
6794// smaller offsets so that they can be sunk to the same blocks as their users.
6795// For example, a large struct starting from %base is split into two parts
6796// where the second part starts from %new_base.
6797//
6798// Before:
6799// BB0:
6800// %base =
6801//
6802// BB1:
6803// %gep0 = gep %base, off0
6804// %gep1 = gep %base, off1
6805// %gep2 = gep %base, off2
6806//
6807// BB2:
6808// %load1 = load %gep0
6809// %load2 = load %gep1
6810// %load3 = load %gep2
6811//
6812// After:
6813// BB0:
6814// %base =
6815// %new_base = gep %base, off0
6816//
6817// BB1:
6818// %new_gep0 = %new_base
6819// %new_gep1 = gep %new_base, off1 - off0
6820// %new_gep2 = gep %new_base, off2 - off0
6821//
6822// BB2:
6823// %load1 = load i32, i32* %new_gep0
6824// %load2 = load i32, i32* %new_gep1
6825// %load3 = load i32, i32* %new_gep2
6826//
6827// %new_gep1 and %new_gep2 can be sunk to BB2 now after the splitting because
6828// their offsets are smaller enough to fit into the addressing mode.
6829bool CodeGenPrepare::splitLargeGEPOffsets() {
6830 bool Changed = false;
6831 for (auto &Entry : LargeOffsetGEPMap) {
6832 Value *OldBase = Entry.first;
6833 SmallVectorImpl<std::pair<AssertingVH<GetElementPtrInst>, int64_t>>
6834 &LargeOffsetGEPs = Entry.second;
6835 auto compareGEPOffset =
6836 [&](const std::pair<GetElementPtrInst *, int64_t> &LHS,
6837 const std::pair<GetElementPtrInst *, int64_t> &RHS) {
6838 if (LHS.first == RHS.first)
6839 return false;
6840 if (LHS.second != RHS.second)
6841 return LHS.second < RHS.second;
6842 return LargeOffsetGEPID[LHS.first] < LargeOffsetGEPID[RHS.first];
6843 };
6844 // Sorting all the GEPs of the same data structures based on the offsets.
6845 llvm::sort(C&: LargeOffsetGEPs, Comp: compareGEPOffset);
6846 LargeOffsetGEPs.erase(CS: llvm::unique(R&: LargeOffsetGEPs), CE: LargeOffsetGEPs.end());
6847 // Skip if all the GEPs have the same offsets.
6848 if (LargeOffsetGEPs.front().second == LargeOffsetGEPs.back().second)
6849 continue;
6850 GetElementPtrInst *BaseGEP = LargeOffsetGEPs.begin()->first;
6851 int64_t BaseOffset = LargeOffsetGEPs.begin()->second;
6852 Value *NewBaseGEP = nullptr;
6853
6854 auto createNewBase = [&](int64_t BaseOffset, Value *OldBase,
6855 GetElementPtrInst *GEP) {
6856 LLVMContext &Ctx = GEP->getContext();
6857 Type *PtrIdxTy = DL->getIndexType(PtrTy: GEP->getType());
6858 Type *I8PtrTy =
6859 PointerType::get(C&: Ctx, AddressSpace: GEP->getType()->getPointerAddressSpace());
6860
6861 BasicBlock::iterator NewBaseInsertPt;
6862 BasicBlock *NewBaseInsertBB;
6863 if (auto *BaseI = dyn_cast<Instruction>(Val: OldBase)) {
6864 // If the base of the struct is an instruction, the new base will be
6865 // inserted close to it.
6866 NewBaseInsertBB = BaseI->getParent();
6867 if (isa<PHINode>(Val: BaseI))
6868 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
6869 else if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Val: BaseI)) {
6870 NewBaseInsertBB =
6871 SplitEdge(From: NewBaseInsertBB, To: Invoke->getNormalDest(),
6872 DT: &getDT(F&: *NewBaseInsertBB->getParent()), LI);
6873 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
6874 } else
6875 NewBaseInsertPt = std::next(x: BaseI->getIterator());
6876 } else {
6877 // If the current base is an argument or global value, the new base
6878 // will be inserted to the entry block.
6879 NewBaseInsertBB = &BaseGEP->getFunction()->getEntryBlock();
6880 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
6881 }
6882 IRBuilder<> NewBaseBuilder(NewBaseInsertBB, NewBaseInsertPt);
6883 // Create a new base.
6884 // TODO: Avoid implicit trunc?
6885 // See https://github.com/llvm/llvm-project/issues/112510.
6886 Value *BaseIndex =
6887 ConstantInt::getSigned(Ty: PtrIdxTy, V: BaseOffset, /*ImplicitTrunc=*/true);
6888 NewBaseGEP = OldBase;
6889 if (NewBaseGEP->getType() != I8PtrTy)
6890 NewBaseGEP = NewBaseBuilder.CreatePointerCast(V: NewBaseGEP, DestTy: I8PtrTy);
6891 NewBaseGEP =
6892 NewBaseBuilder.CreatePtrAdd(Ptr: NewBaseGEP, Offset: BaseIndex, Name: "splitgep");
6893 NewGEPBases.insert(V: NewBaseGEP);
6894 return;
6895 };
6896
6897 // Check whether all the offsets can be encoded with prefered common base.
6898 if (int64_t PreferBase = TLI->getPreferredLargeGEPBaseOffset(
6899 MinOffset: LargeOffsetGEPs.front().second, MaxOffset: LargeOffsetGEPs.back().second)) {
6900 BaseOffset = PreferBase;
6901 // Create a new base if the offset of the BaseGEP can be decoded with one
6902 // instruction.
6903 createNewBase(BaseOffset, OldBase, BaseGEP);
6904 }
6905
6906 auto *LargeOffsetGEP = LargeOffsetGEPs.begin();
6907 while (LargeOffsetGEP != LargeOffsetGEPs.end()) {
6908 GetElementPtrInst *GEP = LargeOffsetGEP->first;
6909 int64_t Offset = LargeOffsetGEP->second;
6910 if (Offset != BaseOffset) {
6911 TargetLowering::AddrMode AddrMode;
6912 AddrMode.HasBaseReg = true;
6913 AddrMode.BaseOffs = Offset - BaseOffset;
6914 // The result type of the GEP might not be the type of the memory
6915 // access.
6916 if (!TLI->isLegalAddressingMode(DL: *DL, AM: AddrMode,
6917 Ty: GEP->getResultElementType(),
6918 AddrSpace: GEP->getAddressSpace())) {
6919 // We need to create a new base if the offset to the current base is
6920 // too large to fit into the addressing mode. So, a very large struct
6921 // may be split into several parts.
6922 BaseGEP = GEP;
6923 BaseOffset = Offset;
6924 NewBaseGEP = nullptr;
6925 }
6926 }
6927
6928 // Generate a new GEP to replace the current one.
6929 Type *PtrIdxTy = DL->getIndexType(PtrTy: GEP->getType());
6930
6931 if (!NewBaseGEP) {
6932 // Create a new base if we don't have one yet. Find the insertion
6933 // pointer for the new base first.
6934 createNewBase(BaseOffset, OldBase, GEP);
6935 }
6936
6937 IRBuilder<> Builder(GEP);
6938 Value *NewGEP = NewBaseGEP;
6939 if (Offset != BaseOffset) {
6940 // Calculate the new offset for the new GEP.
6941 Value *Index = ConstantInt::get(Ty: PtrIdxTy, V: Offset - BaseOffset);
6942 NewGEP = Builder.CreatePtrAdd(Ptr: NewBaseGEP, Offset: Index);
6943 }
6944 replaceAllUsesWith(Old: GEP, New: NewGEP, FreshBBs, IsHuge: IsHugeFunc);
6945 LargeOffsetGEPID.erase(Val: GEP);
6946 LargeOffsetGEP = LargeOffsetGEPs.erase(CI: LargeOffsetGEP);
6947 GEP->eraseFromParent();
6948 Changed = true;
6949 }
6950 }
6951 return Changed;
6952}
6953
6954bool CodeGenPrepare::optimizePhiType(
6955 PHINode *I, SmallPtrSetImpl<PHINode *> &Visited,
6956 SmallPtrSetImpl<Instruction *> &DeletedInstrs) {
6957 // We are looking for a collection on interconnected phi nodes that together
6958 // only use loads/bitcasts and are used by stores/bitcasts, and the bitcasts
6959 // are of the same type. Convert the whole set of nodes to the type of the
6960 // bitcast.
6961 Type *PhiTy = I->getType();
6962 Type *ConvertTy = nullptr;
6963 if (Visited.count(Ptr: I) ||
6964 (!I->getType()->isIntegerTy() && !I->getType()->isFloatingPointTy()))
6965 return false;
6966
6967 SmallVector<Instruction *, 4> Worklist;
6968 Worklist.push_back(Elt: cast<Instruction>(Val: I));
6969 SmallPtrSet<PHINode *, 4> PhiNodes;
6970 SmallPtrSet<ConstantData *, 4> Constants;
6971 PhiNodes.insert(Ptr: I);
6972 Visited.insert(Ptr: I);
6973 SmallPtrSet<Instruction *, 4> Defs;
6974 SmallPtrSet<Instruction *, 4> Uses;
6975 // This works by adding extra bitcasts between load/stores and removing
6976 // existing bicasts. If we have a phi(bitcast(load)) or a store(bitcast(phi))
6977 // we can get in the situation where we remove a bitcast in one iteration
6978 // just to add it again in the next. We need to ensure that at least one
6979 // bitcast we remove are anchored to something that will not change back.
6980 bool AnyAnchored = false;
6981
6982 while (!Worklist.empty()) {
6983 Instruction *II = Worklist.pop_back_val();
6984
6985 if (auto *Phi = dyn_cast<PHINode>(Val: II)) {
6986 // Handle Defs, which might also be PHI's
6987 for (Value *V : Phi->incoming_values()) {
6988 if (auto *OpPhi = dyn_cast<PHINode>(Val: V)) {
6989 if (!PhiNodes.count(Ptr: OpPhi)) {
6990 if (!Visited.insert(Ptr: OpPhi).second)
6991 return false;
6992 PhiNodes.insert(Ptr: OpPhi);
6993 Worklist.push_back(Elt: OpPhi);
6994 }
6995 } else if (auto *OpLoad = dyn_cast<LoadInst>(Val: V)) {
6996 if (!OpLoad->isSimple())
6997 return false;
6998 if (Defs.insert(Ptr: OpLoad).second)
6999 Worklist.push_back(Elt: OpLoad);
7000 } else if (auto *OpEx = dyn_cast<ExtractElementInst>(Val: V)) {
7001 if (Defs.insert(Ptr: OpEx).second)
7002 Worklist.push_back(Elt: OpEx);
7003 } else if (auto *OpBC = dyn_cast<BitCastInst>(Val: V)) {
7004 if (!ConvertTy)
7005 ConvertTy = OpBC->getOperand(i_nocapture: 0)->getType();
7006 if (OpBC->getOperand(i_nocapture: 0)->getType() != ConvertTy)
7007 return false;
7008 if (Defs.insert(Ptr: OpBC).second) {
7009 Worklist.push_back(Elt: OpBC);
7010 AnyAnchored |= !isa<LoadInst>(Val: OpBC->getOperand(i_nocapture: 0)) &&
7011 !isa<ExtractElementInst>(Val: OpBC->getOperand(i_nocapture: 0));
7012 }
7013 } else if (auto *OpC = dyn_cast<ConstantData>(Val: V))
7014 Constants.insert(Ptr: OpC);
7015 else
7016 return false;
7017 }
7018 }
7019
7020 // Handle uses which might also be phi's
7021 for (User *V : II->users()) {
7022 if (auto *OpPhi = dyn_cast<PHINode>(Val: V)) {
7023 if (!PhiNodes.count(Ptr: OpPhi)) {
7024 if (Visited.count(Ptr: OpPhi))
7025 return false;
7026 PhiNodes.insert(Ptr: OpPhi);
7027 Visited.insert(Ptr: OpPhi);
7028 Worklist.push_back(Elt: OpPhi);
7029 }
7030 } else if (auto *OpStore = dyn_cast<StoreInst>(Val: V)) {
7031 if (!OpStore->isSimple() || OpStore->getOperand(i_nocapture: 0) != II)
7032 return false;
7033 Uses.insert(Ptr: OpStore);
7034 } else if (auto *OpBC = dyn_cast<BitCastInst>(Val: V)) {
7035 if (!ConvertTy)
7036 ConvertTy = OpBC->getType();
7037 if (OpBC->getType() != ConvertTy)
7038 return false;
7039 Uses.insert(Ptr: OpBC);
7040 AnyAnchored |=
7041 any_of(Range: OpBC->users(), P: [](User *U) { return !isa<StoreInst>(Val: U); });
7042 } else {
7043 return false;
7044 }
7045 }
7046 }
7047
7048 if (!ConvertTy || !AnyAnchored || PhiTy == ConvertTy ||
7049 !TLI->shouldConvertPhiType(From: PhiTy, To: ConvertTy))
7050 return false;
7051
7052 LLVM_DEBUG(dbgs() << "Converting " << *I << "\n and connected nodes to "
7053 << *ConvertTy << "\n");
7054
7055 // Create all the new phi nodes of the new type, and bitcast any loads to the
7056 // correct type.
7057 ValueToValueMap ValMap;
7058 for (ConstantData *C : Constants)
7059 ValMap[C] = ConstantExpr::getBitCast(C, Ty: ConvertTy);
7060 for (Instruction *D : Defs) {
7061 if (isa<BitCastInst>(Val: D)) {
7062 ValMap[D] = D->getOperand(i: 0);
7063 DeletedInstrs.insert(Ptr: D);
7064 } else {
7065 BasicBlock::iterator insertPt = std::next(x: D->getIterator());
7066 ValMap[D] = new BitCastInst(D, ConvertTy, D->getName() + ".bc", insertPt);
7067 }
7068 }
7069 for (PHINode *Phi : PhiNodes)
7070 ValMap[Phi] = PHINode::Create(Ty: ConvertTy, NumReservedValues: Phi->getNumIncomingValues(),
7071 NameStr: Phi->getName() + ".tc", InsertBefore: Phi->getIterator());
7072 // Pipe together all the PhiNodes.
7073 for (PHINode *Phi : PhiNodes) {
7074 PHINode *NewPhi = cast<PHINode>(Val: ValMap[Phi]);
7075 for (int i = 0, e = Phi->getNumIncomingValues(); i < e; i++)
7076 NewPhi->addIncoming(V: ValMap[Phi->getIncomingValue(i)],
7077 BB: Phi->getIncomingBlock(i));
7078 Visited.insert(Ptr: NewPhi);
7079 }
7080 // And finally pipe up the stores and bitcasts
7081 for (Instruction *U : Uses) {
7082 if (isa<BitCastInst>(Val: U)) {
7083 DeletedInstrs.insert(Ptr: U);
7084 replaceAllUsesWith(Old: U, New: ValMap[U->getOperand(i: 0)], FreshBBs, IsHuge: IsHugeFunc);
7085 } else {
7086 U->setOperand(i: 0, Val: new BitCastInst(ValMap[U->getOperand(i: 0)], PhiTy, "bc",
7087 U->getIterator()));
7088 }
7089 }
7090
7091 // Save the removed phis to be deleted later.
7092 DeletedInstrs.insert_range(R&: PhiNodes);
7093 return true;
7094}
7095
7096bool CodeGenPrepare::optimizePhiTypes(Function &F) {
7097 if (!OptimizePhiTypes)
7098 return false;
7099
7100 bool Changed = false;
7101 SmallPtrSet<PHINode *, 4> Visited;
7102 SmallPtrSet<Instruction *, 4> DeletedInstrs;
7103
7104 // Attempt to optimize all the phis in the functions to the correct type.
7105 for (auto &BB : F)
7106 for (auto &Phi : BB.phis())
7107 Changed |= optimizePhiType(I: &Phi, Visited, DeletedInstrs);
7108
7109 // Remove any old phi's that have been converted.
7110 for (auto *I : DeletedInstrs) {
7111 replaceAllUsesWith(Old: I, New: PoisonValue::get(T: I->getType()), FreshBBs, IsHuge: IsHugeFunc);
7112 I->eraseFromParent();
7113 }
7114
7115 return Changed;
7116}
7117
7118/// Return true, if an ext(load) can be formed from an extension in
7119/// \p MovedExts.
7120bool CodeGenPrepare::canFormExtLd(
7121 const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
7122 Instruction *&Inst, bool HasPromoted) {
7123 for (auto *MovedExtInst : MovedExts) {
7124 if (isa<LoadInst>(Val: MovedExtInst->getOperand(i: 0))) {
7125 LI = cast<LoadInst>(Val: MovedExtInst->getOperand(i: 0));
7126 Inst = MovedExtInst;
7127 break;
7128 }
7129 }
7130 if (!LI)
7131 return false;
7132
7133 // If they're already in the same block, there's nothing to do.
7134 // Make the cheap checks first if we did not promote.
7135 // If we promoted, we need to check if it is indeed profitable.
7136 if (!HasPromoted && LI->getParent() == Inst->getParent())
7137 return false;
7138
7139 return TLI->isExtLoad(Load: LI, Ext: Inst, DL: *DL);
7140}
7141
7142/// Move a zext or sext fed by a load into the same basic block as the load,
7143/// unless conditions are unfavorable. This allows SelectionDAG to fold the
7144/// extend into the load.
7145///
7146/// E.g.,
7147/// \code
7148/// %ld = load i32* %addr
7149/// %add = add nuw i32 %ld, 4
7150/// %zext = zext i32 %add to i64
7151// \endcode
7152/// =>
7153/// \code
7154/// %ld = load i32* %addr
7155/// %zext = zext i32 %ld to i64
7156/// %add = add nuw i64 %zext, 4
7157/// \encode
7158/// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
7159/// allow us to match zext(load i32*) to i64.
7160///
7161/// Also, try to promote the computations used to obtain a sign extended
7162/// value used into memory accesses.
7163/// E.g.,
7164/// \code
7165/// a = add nsw i32 b, 3
7166/// d = sext i32 a to i64
7167/// e = getelementptr ..., i64 d
7168/// \endcode
7169/// =>
7170/// \code
7171/// f = sext i32 b to i64
7172/// a = add nsw i64 f, 3
7173/// e = getelementptr ..., i64 a
7174/// \endcode
7175///
7176/// \p Inst[in/out] the extension may be modified during the process if some
7177/// promotions apply.
7178bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
7179 bool AllowPromotionWithoutCommonHeader = false;
7180 /// See if it is an interesting sext operations for the address type
7181 /// promotion before trying to promote it, e.g., the ones with the right
7182 /// type and used in memory accesses.
7183 bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
7184 I: *Inst, AllowPromotionWithoutCommonHeader);
7185 TypePromotionTransaction TPT(RemovedInsts);
7186 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
7187 TPT.getRestorationPoint();
7188 SmallVector<Instruction *, 1> Exts;
7189 SmallVector<Instruction *, 2> SpeculativelyMovedExts;
7190 Exts.push_back(Elt: Inst);
7191
7192 bool HasPromoted = tryToPromoteExts(TPT, Exts, ProfitablyMovedExts&: SpeculativelyMovedExts);
7193
7194 // Look for a load being extended.
7195 LoadInst *LI = nullptr;
7196 Instruction *ExtFedByLoad;
7197
7198 // Try to promote a chain of computation if it allows to form an extended
7199 // load.
7200 if (canFormExtLd(MovedExts: SpeculativelyMovedExts, LI, Inst&: ExtFedByLoad, HasPromoted)) {
7201 assert(LI && ExtFedByLoad && "Expect a valid load and extension");
7202 TPT.commit();
7203 // Move the extend into the same block as the load.
7204 ExtFedByLoad->moveAfter(MovePos: LI);
7205 ++NumExtsMoved;
7206 Inst = ExtFedByLoad;
7207 return true;
7208 }
7209
7210 // Continue promoting SExts if known as considerable depending on targets.
7211 if (ATPConsiderable &&
7212 performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
7213 HasPromoted, TPT, SpeculativelyMovedExts))
7214 return true;
7215
7216 TPT.rollback(Point: LastKnownGood);
7217 return false;
7218}
7219
7220// Perform address type promotion if doing so is profitable.
7221// If AllowPromotionWithoutCommonHeader == false, we should find other sext
7222// instructions that sign extended the same initial value. However, if
7223// AllowPromotionWithoutCommonHeader == true, we expect promoting the
7224// extension is just profitable.
7225bool CodeGenPrepare::performAddressTypePromotion(
7226 Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
7227 bool HasPromoted, TypePromotionTransaction &TPT,
7228 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
7229 bool Promoted = false;
7230 SmallPtrSet<Instruction *, 1> UnhandledExts;
7231 bool AllSeenFirst = true;
7232 for (auto *I : SpeculativelyMovedExts) {
7233 Value *HeadOfChain = I->getOperand(i: 0);
7234 DenseMap<Value *, Instruction *>::iterator AlreadySeen =
7235 SeenChainsForSExt.find(Val: HeadOfChain);
7236 // If there is an unhandled SExt which has the same header, try to promote
7237 // it as well.
7238 if (AlreadySeen != SeenChainsForSExt.end()) {
7239 if (AlreadySeen->second != nullptr)
7240 UnhandledExts.insert(Ptr: AlreadySeen->second);
7241 AllSeenFirst = false;
7242 }
7243 }
7244
7245 if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
7246 SpeculativelyMovedExts.size() == 1)) {
7247 TPT.commit();
7248 if (HasPromoted)
7249 Promoted = true;
7250 for (auto *I : SpeculativelyMovedExts) {
7251 Value *HeadOfChain = I->getOperand(i: 0);
7252 SeenChainsForSExt[HeadOfChain] = nullptr;
7253 ValToSExtendedUses[HeadOfChain].push_back(Elt: I);
7254 }
7255 // Update Inst as promotion happen.
7256 Inst = SpeculativelyMovedExts.pop_back_val();
7257 } else {
7258 // This is the first chain visited from the header, keep the current chain
7259 // as unhandled. Defer to promote this until we encounter another SExt
7260 // chain derived from the same header.
7261 for (auto *I : SpeculativelyMovedExts) {
7262 Value *HeadOfChain = I->getOperand(i: 0);
7263 SeenChainsForSExt[HeadOfChain] = Inst;
7264 }
7265 return false;
7266 }
7267
7268 if (!AllSeenFirst && !UnhandledExts.empty())
7269 for (auto *VisitedSExt : UnhandledExts) {
7270 if (RemovedInsts.count(Ptr: VisitedSExt))
7271 continue;
7272 TypePromotionTransaction TPT(RemovedInsts);
7273 SmallVector<Instruction *, 1> Exts;
7274 SmallVector<Instruction *, 2> Chains;
7275 Exts.push_back(Elt: VisitedSExt);
7276 bool HasPromoted = tryToPromoteExts(TPT, Exts, ProfitablyMovedExts&: Chains);
7277 TPT.commit();
7278 if (HasPromoted)
7279 Promoted = true;
7280 for (auto *I : Chains) {
7281 Value *HeadOfChain = I->getOperand(i: 0);
7282 // Mark this as handled.
7283 SeenChainsForSExt[HeadOfChain] = nullptr;
7284 ValToSExtendedUses[HeadOfChain].push_back(Elt: I);
7285 }
7286 }
7287 return Promoted;
7288}
7289
7290bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
7291 BasicBlock *DefBB = I->getParent();
7292
7293 // If the result of a {s|z}ext and its source are both live out, rewrite all
7294 // other uses of the source with result of extension.
7295 Value *Src = I->getOperand(i: 0);
7296 if (Src->hasOneUse())
7297 return false;
7298
7299 // Only do this xform if truncating is free.
7300 if (!TLI->isTruncateFree(FromTy: I->getType(), ToTy: Src->getType()))
7301 return false;
7302
7303 // Only safe to perform the optimization if the source is also defined in
7304 // this block.
7305 if (!isa<Instruction>(Val: Src) || DefBB != cast<Instruction>(Val: Src)->getParent())
7306 return false;
7307
7308 bool DefIsLiveOut = false;
7309 for (User *U : I->users()) {
7310 Instruction *UI = cast<Instruction>(Val: U);
7311
7312 // Figure out which BB this ext is used in.
7313 BasicBlock *UserBB = UI->getParent();
7314 if (UserBB == DefBB)
7315 continue;
7316 DefIsLiveOut = true;
7317 break;
7318 }
7319 if (!DefIsLiveOut)
7320 return false;
7321
7322 // Make sure none of the uses are PHI nodes.
7323 for (User *U : Src->users()) {
7324 Instruction *UI = cast<Instruction>(Val: U);
7325 BasicBlock *UserBB = UI->getParent();
7326 if (UserBB == DefBB)
7327 continue;
7328 // Be conservative. We don't want this xform to end up introducing
7329 // reloads just before load / store instructions.
7330 if (isa<PHINode>(Val: UI) || isa<LoadInst>(Val: UI) || isa<StoreInst>(Val: UI))
7331 return false;
7332 }
7333
7334 // InsertedTruncs - Only insert one trunc in each block once.
7335 DenseMap<BasicBlock *, Instruction *> InsertedTruncs;
7336
7337 bool MadeChange = false;
7338 for (Use &U : Src->uses()) {
7339 Instruction *User = cast<Instruction>(Val: U.getUser());
7340
7341 // Figure out which BB this ext is used in.
7342 BasicBlock *UserBB = User->getParent();
7343 if (UserBB == DefBB)
7344 continue;
7345
7346 // Both src and def are live in this block. Rewrite the use.
7347 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
7348
7349 if (!InsertedTrunc) {
7350 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
7351 assert(InsertPt != UserBB->end());
7352 InsertedTrunc = new TruncInst(I, Src->getType(), "");
7353 InsertedTrunc->insertBefore(BB&: *UserBB, InsertPos: InsertPt);
7354 InsertedInsts.insert(Ptr: InsertedTrunc);
7355 }
7356
7357 // Replace a use of the {s|z}ext source with a use of the result.
7358 U = InsertedTrunc;
7359 ++NumExtUses;
7360 MadeChange = true;
7361 }
7362
7363 return MadeChange;
7364}
7365
7366// Find loads whose uses only use some of the loaded value's bits. Add an "and"
7367// just after the load if the target can fold this into one extload instruction,
7368// with the hope of eliminating some of the other later "and" instructions using
7369// the loaded value. "and"s that are made trivially redundant by the insertion
7370// of the new "and" are removed by this function, while others (e.g. those whose
7371// path from the load goes through a phi) are left for isel to potentially
7372// remove.
7373//
7374// For example:
7375//
7376// b0:
7377// x = load i32
7378// ...
7379// b1:
7380// y = and x, 0xff
7381// z = use y
7382//
7383// becomes:
7384//
7385// b0:
7386// x = load i32
7387// x' = and x, 0xff
7388// ...
7389// b1:
7390// z = use x'
7391//
7392// whereas:
7393//
7394// b0:
7395// x1 = load i32
7396// ...
7397// b1:
7398// x2 = load i32
7399// ...
7400// b2:
7401// x = phi x1, x2
7402// y = and x, 0xff
7403//
7404// becomes (after a call to optimizeLoadExt for each load):
7405//
7406// b0:
7407// x1 = load i32
7408// x1' = and x1, 0xff
7409// ...
7410// b1:
7411// x2 = load i32
7412// x2' = and x2, 0xff
7413// ...
7414// b2:
7415// x = phi x1', x2'
7416// y = and x, 0xff
7417bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
7418 if (!Load->isSimple() || !Load->getType()->isIntOrPtrTy())
7419 return false;
7420
7421 // Skip loads we've already transformed.
7422 if (Load->hasOneUse() &&
7423 InsertedInsts.count(Ptr: cast<Instruction>(Val: *Load->user_begin())))
7424 return false;
7425
7426 // Look at all uses of Load, looking through phis, to determine how many bits
7427 // of the loaded value are needed.
7428 SmallVector<Instruction *, 8> WorkList;
7429 SmallPtrSet<Instruction *, 16> Visited;
7430 SmallVector<Instruction *, 8> AndsToMaybeRemove;
7431 SmallVector<Instruction *, 8> DropFlags;
7432 for (auto *U : Load->users())
7433 WorkList.push_back(Elt: cast<Instruction>(Val: U));
7434
7435 EVT LoadResultVT = TLI->getValueType(DL: *DL, Ty: Load->getType());
7436 unsigned BitWidth = LoadResultVT.getSizeInBits();
7437 // If the BitWidth is 0, do not try to optimize the type
7438 if (BitWidth == 0)
7439 return false;
7440
7441 APInt DemandBits(BitWidth, 0);
7442 APInt WidestAndBits(BitWidth, 0);
7443
7444 while (!WorkList.empty()) {
7445 Instruction *I = WorkList.pop_back_val();
7446
7447 // Break use-def graph loops.
7448 if (!Visited.insert(Ptr: I).second)
7449 continue;
7450
7451 // For a PHI node, push all of its users.
7452 if (auto *Phi = dyn_cast<PHINode>(Val: I)) {
7453 for (auto *U : Phi->users())
7454 WorkList.push_back(Elt: cast<Instruction>(Val: U));
7455 continue;
7456 }
7457
7458 switch (I->getOpcode()) {
7459 case Instruction::And: {
7460 auto *AndC = dyn_cast<ConstantInt>(Val: I->getOperand(i: 1));
7461 if (!AndC)
7462 return false;
7463 APInt AndBits = AndC->getValue();
7464 DemandBits |= AndBits;
7465 // Keep track of the widest and mask we see.
7466 if (AndBits.ugt(RHS: WidestAndBits))
7467 WidestAndBits = AndBits;
7468 if (AndBits == WidestAndBits && I->getOperand(i: 0) == Load)
7469 AndsToMaybeRemove.push_back(Elt: I);
7470 break;
7471 }
7472
7473 case Instruction::Shl: {
7474 auto *ShlC = dyn_cast<ConstantInt>(Val: I->getOperand(i: 1));
7475 if (!ShlC)
7476 return false;
7477 uint64_t ShiftAmt = ShlC->getLimitedValue(Limit: BitWidth - 1);
7478 DemandBits.setLowBits(BitWidth - ShiftAmt);
7479 DropFlags.push_back(Elt: I);
7480 break;
7481 }
7482
7483 case Instruction::Trunc: {
7484 EVT TruncVT = TLI->getValueType(DL: *DL, Ty: I->getType());
7485 unsigned TruncBitWidth = TruncVT.getSizeInBits();
7486 DemandBits.setLowBits(TruncBitWidth);
7487 DropFlags.push_back(Elt: I);
7488 break;
7489 }
7490
7491 default:
7492 return false;
7493 }
7494 }
7495
7496 uint32_t ActiveBits = DemandBits.getActiveBits();
7497 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
7498 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
7499 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
7500 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
7501 // followed by an AND.
7502 // TODO: Look into removing this restriction by fixing backends to either
7503 // return false for isLoadExtLegal for i1 or have them select this pattern to
7504 // a single instruction.
7505 //
7506 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
7507 // mask, since these are the only ands that will be removed by isel.
7508 if (ActiveBits <= 1 || !DemandBits.isMask(numBits: ActiveBits) ||
7509 WidestAndBits != DemandBits)
7510 return false;
7511
7512 LLVMContext &Ctx = Load->getType()->getContext();
7513 Type *TruncTy = Type::getIntNTy(C&: Ctx, N: ActiveBits);
7514 EVT TruncVT = TLI->getValueType(DL: *DL, Ty: TruncTy);
7515
7516 // Reject cases that won't be matched as extloads.
7517 if (!LoadResultVT.bitsGT(VT: TruncVT) || !TruncVT.isRound() ||
7518 !TLI->isLoadExtLegal(ExtType: ISD::ZEXTLOAD, ValVT: LoadResultVT, MemVT: TruncVT))
7519 return false;
7520
7521 IRBuilder<> Builder(Load->getNextNode());
7522 auto *NewAnd = cast<Instruction>(
7523 Val: Builder.CreateAnd(LHS: Load, RHS: ConstantInt::get(Context&: Ctx, V: DemandBits)));
7524 // Mark this instruction as "inserted by CGP", so that other
7525 // optimizations don't touch it.
7526 InsertedInsts.insert(Ptr: NewAnd);
7527
7528 // Replace all uses of load with new and (except for the use of load in the
7529 // new and itself).
7530 replaceAllUsesWith(Old: Load, New: NewAnd, FreshBBs, IsHuge: IsHugeFunc);
7531 NewAnd->setOperand(i: 0, Val: Load);
7532
7533 // Remove any and instructions that are now redundant.
7534 for (auto *And : AndsToMaybeRemove)
7535 // Check that the and mask is the same as the one we decided to put on the
7536 // new and.
7537 if (cast<ConstantInt>(Val: And->getOperand(i: 1))->getValue() == DemandBits) {
7538 replaceAllUsesWith(Old: And, New: NewAnd, FreshBBs, IsHuge: IsHugeFunc);
7539 if (&*CurInstIterator == And)
7540 CurInstIterator = std::next(x: And->getIterator());
7541 And->eraseFromParent();
7542 ++NumAndUses;
7543 }
7544
7545 // NSW flags may not longer hold.
7546 for (auto *Inst : DropFlags)
7547 Inst->setHasNoSignedWrap(false);
7548
7549 ++NumAndsAdded;
7550 return true;
7551}
7552
7553/// Check if V (an operand of a select instruction) is an expensive instruction
7554/// that is only used once.
7555static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
7556 auto *I = dyn_cast<Instruction>(Val: V);
7557 // If it's safe to speculatively execute, then it should not have side
7558 // effects; therefore, it's safe to sink and possibly *not* execute.
7559 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
7560 TTI->isExpensiveToSpeculativelyExecute(I);
7561}
7562
7563/// Returns true if a SelectInst should be turned into an explicit branch.
7564static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
7565 const TargetLowering *TLI,
7566 SelectInst *SI) {
7567 // If even a predictable select is cheap, then a branch can't be cheaper.
7568 if (!TLI->isPredictableSelectExpensive())
7569 return false;
7570
7571 // FIXME: This should use the same heuristics as IfConversion to determine
7572 // whether a select is better represented as a branch.
7573
7574 // If metadata tells us that the select condition is obviously predictable,
7575 // then we want to replace the select with a branch.
7576 uint64_t TrueWeight, FalseWeight;
7577 if (extractBranchWeights(I: *SI, TrueVal&: TrueWeight, FalseVal&: FalseWeight)) {
7578 uint64_t Max = std::max(a: TrueWeight, b: FalseWeight);
7579 uint64_t Sum = TrueWeight + FalseWeight;
7580 if (Sum != 0) {
7581 auto Probability = BranchProbability::getBranchProbability(Numerator: Max, Denominator: Sum);
7582 if (Probability > TTI->getPredictableBranchThreshold())
7583 return true;
7584 }
7585 }
7586
7587 CmpInst *Cmp = dyn_cast<CmpInst>(Val: SI->getCondition());
7588
7589 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
7590 // comparison condition. If the compare has more than one use, there's
7591 // probably another cmov or setcc around, so it's not worth emitting a branch.
7592 if (!Cmp || !Cmp->hasOneUse())
7593 return false;
7594
7595 // If either operand of the select is expensive and only needed on one side
7596 // of the select, we should form a branch.
7597 if (sinkSelectOperand(TTI, V: SI->getTrueValue()) ||
7598 sinkSelectOperand(TTI, V: SI->getFalseValue()))
7599 return true;
7600
7601 return false;
7602}
7603
7604/// If \p isTrue is true, return the true value of \p SI, otherwise return
7605/// false value of \p SI. If the true/false value of \p SI is defined by any
7606/// select instructions in \p Selects, look through the defining select
7607/// instruction until the true/false value is not defined in \p Selects.
7608static Value *
7609getTrueOrFalseValue(SelectInst *SI, bool isTrue,
7610 const SmallPtrSet<const Instruction *, 2> &Selects) {
7611 Value *V = nullptr;
7612
7613 for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(Ptr: DefSI);
7614 DefSI = dyn_cast<SelectInst>(Val: V)) {
7615 assert(DefSI->getCondition() == SI->getCondition() &&
7616 "The condition of DefSI does not match with SI");
7617 V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
7618 }
7619
7620 assert(V && "Failed to get select true/false value");
7621 return V;
7622}
7623
7624bool CodeGenPrepare::optimizeShiftInst(BinaryOperator *Shift) {
7625 assert(Shift->isShift() && "Expected a shift");
7626
7627 // If this is (1) a vector shift, (2) shifts by scalars are cheaper than
7628 // general vector shifts, and (3) the shift amount is a select-of-splatted
7629 // values, hoist the shifts before the select:
7630 // shift Op0, (select Cond, TVal, FVal) -->
7631 // select Cond, (shift Op0, TVal), (shift Op0, FVal)
7632 //
7633 // This is inverting a generic IR transform when we know that the cost of a
7634 // general vector shift is more than the cost of 2 shift-by-scalars.
7635 // We can't do this effectively in SDAG because we may not be able to
7636 // determine if the select operands are splats from within a basic block.
7637 Type *Ty = Shift->getType();
7638 if (!Ty->isVectorTy() || !TTI->isVectorShiftByScalarCheap(Ty))
7639 return false;
7640 Value *Cond, *TVal, *FVal;
7641 if (!match(V: Shift->getOperand(i_nocapture: 1),
7642 P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: TVal), R: m_Value(V&: FVal)))))
7643 return false;
7644 if (!isSplatValue(V: TVal) || !isSplatValue(V: FVal))
7645 return false;
7646
7647 IRBuilder<> Builder(Shift);
7648 BinaryOperator::BinaryOps Opcode = Shift->getOpcode();
7649 Value *NewTVal = Builder.CreateBinOp(Opc: Opcode, LHS: Shift->getOperand(i_nocapture: 0), RHS: TVal);
7650 Value *NewFVal = Builder.CreateBinOp(Opc: Opcode, LHS: Shift->getOperand(i_nocapture: 0), RHS: FVal);
7651 Value *NewSel = Builder.CreateSelect(C: Cond, True: NewTVal, False: NewFVal);
7652 replaceAllUsesWith(Old: Shift, New: NewSel, FreshBBs, IsHuge: IsHugeFunc);
7653 Shift->eraseFromParent();
7654 return true;
7655}
7656
7657bool CodeGenPrepare::optimizeFunnelShift(IntrinsicInst *Fsh) {
7658 Intrinsic::ID Opcode = Fsh->getIntrinsicID();
7659 assert((Opcode == Intrinsic::fshl || Opcode == Intrinsic::fshr) &&
7660 "Expected a funnel shift");
7661
7662 // If this is (1) a vector funnel shift, (2) shifts by scalars are cheaper
7663 // than general vector shifts, and (3) the shift amount is select-of-splatted
7664 // values, hoist the funnel shifts before the select:
7665 // fsh Op0, Op1, (select Cond, TVal, FVal) -->
7666 // select Cond, (fsh Op0, Op1, TVal), (fsh Op0, Op1, FVal)
7667 //
7668 // This is inverting a generic IR transform when we know that the cost of a
7669 // general vector shift is more than the cost of 2 shift-by-scalars.
7670 // We can't do this effectively in SDAG because we may not be able to
7671 // determine if the select operands are splats from within a basic block.
7672 Type *Ty = Fsh->getType();
7673 if (!Ty->isVectorTy() || !TTI->isVectorShiftByScalarCheap(Ty))
7674 return false;
7675 Value *Cond, *TVal, *FVal;
7676 if (!match(V: Fsh->getOperand(i_nocapture: 2),
7677 P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: TVal), R: m_Value(V&: FVal)))))
7678 return false;
7679 if (!isSplatValue(V: TVal) || !isSplatValue(V: FVal))
7680 return false;
7681
7682 IRBuilder<> Builder(Fsh);
7683 Value *X = Fsh->getOperand(i_nocapture: 0), *Y = Fsh->getOperand(i_nocapture: 1);
7684 Value *NewTVal = Builder.CreateIntrinsic(ID: Opcode, Types: Ty, Args: {X, Y, TVal});
7685 Value *NewFVal = Builder.CreateIntrinsic(ID: Opcode, Types: Ty, Args: {X, Y, FVal});
7686 Value *NewSel = Builder.CreateSelect(C: Cond, True: NewTVal, False: NewFVal);
7687 replaceAllUsesWith(Old: Fsh, New: NewSel, FreshBBs, IsHuge: IsHugeFunc);
7688 Fsh->eraseFromParent();
7689 return true;
7690}
7691
7692/// If we have a SelectInst that will likely profit from branch prediction,
7693/// turn it into a branch.
7694bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
7695 if (DisableSelectToBranch)
7696 return false;
7697
7698 // If the SelectOptimize pass is enabled, selects have already been optimized.
7699 if (!getCGPassBuilderOption().DisableSelectOptimize)
7700 return false;
7701
7702 // Find all consecutive select instructions that share the same condition.
7703 SmallVector<SelectInst *, 2> ASI;
7704 ASI.push_back(Elt: SI);
7705 for (BasicBlock::iterator It = ++BasicBlock::iterator(SI);
7706 It != SI->getParent()->end(); ++It) {
7707 SelectInst *I = dyn_cast<SelectInst>(Val: &*It);
7708 if (I && SI->getCondition() == I->getCondition()) {
7709 ASI.push_back(Elt: I);
7710 } else {
7711 break;
7712 }
7713 }
7714
7715 SelectInst *LastSI = ASI.back();
7716 // Increment the current iterator to skip all the rest of select instructions
7717 // because they will be either "not lowered" or "all lowered" to branch.
7718 CurInstIterator = std::next(x: LastSI->getIterator());
7719 // Examine debug-info attached to the consecutive select instructions. They
7720 // won't be individually optimised by optimizeInst, so we need to perform
7721 // DbgVariableRecord maintenence here instead.
7722 for (SelectInst *SI : ArrayRef(ASI).drop_front())
7723 fixupDbgVariableRecordsOnInst(I&: *SI);
7724
7725 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(Bitwidth: 1);
7726
7727 // Can we convert the 'select' to CF ?
7728 if (VectorCond || SI->getMetadata(KindID: LLVMContext::MD_unpredictable))
7729 return false;
7730
7731 TargetLowering::SelectSupportKind SelectKind;
7732 if (SI->getType()->isVectorTy())
7733 SelectKind = TargetLowering::ScalarCondVectorVal;
7734 else
7735 SelectKind = TargetLowering::ScalarValSelect;
7736
7737 if (TLI->isSelectSupported(SelectKind) &&
7738 (!isFormingBranchFromSelectProfitable(TTI, TLI, SI) ||
7739 llvm::shouldOptimizeForSize(BB: SI->getParent(), PSI, BFI: BFI.get())))
7740 return false;
7741
7742 // The DominatorTree needs to be rebuilt by any consumers after this
7743 // transformation. We simply reset here rather than setting the ModifiedDT
7744 // flag to avoid restarting the function walk in runOnFunction for each
7745 // select optimized.
7746 DT.reset();
7747
7748 // Transform a sequence like this:
7749 // start:
7750 // %cmp = cmp uge i32 %a, %b
7751 // %sel = select i1 %cmp, i32 %c, i32 %d
7752 //
7753 // Into:
7754 // start:
7755 // %cmp = cmp uge i32 %a, %b
7756 // %cmp.frozen = freeze %cmp
7757 // br i1 %cmp.frozen, label %select.true, label %select.false
7758 // select.true:
7759 // br label %select.end
7760 // select.false:
7761 // br label %select.end
7762 // select.end:
7763 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
7764 //
7765 // %cmp should be frozen, otherwise it may introduce undefined behavior.
7766 // In addition, we may sink instructions that produce %c or %d from
7767 // the entry block into the destination(s) of the new branch.
7768 // If the true or false blocks do not contain a sunken instruction, that
7769 // block and its branch may be optimized away. In that case, one side of the
7770 // first branch will point directly to select.end, and the corresponding PHI
7771 // predecessor block will be the start block.
7772
7773 // Collect values that go on the true side and the values that go on the false
7774 // side.
7775 SmallVector<Instruction *> TrueInstrs, FalseInstrs;
7776 for (SelectInst *SI : ASI) {
7777 if (Value *V = SI->getTrueValue(); sinkSelectOperand(TTI, V))
7778 TrueInstrs.push_back(Elt: cast<Instruction>(Val: V));
7779 if (Value *V = SI->getFalseValue(); sinkSelectOperand(TTI, V))
7780 FalseInstrs.push_back(Elt: cast<Instruction>(Val: V));
7781 }
7782
7783 // Split the select block, according to how many (if any) values go on each
7784 // side.
7785 BasicBlock *StartBlock = SI->getParent();
7786 BasicBlock::iterator SplitPt = std::next(x: BasicBlock::iterator(LastSI));
7787 // We should split before any debug-info.
7788 SplitPt.setHeadBit(true);
7789
7790 IRBuilder<> IB(SI);
7791 auto *CondFr = IB.CreateFreeze(V: SI->getCondition(), Name: SI->getName() + ".frozen");
7792
7793 BasicBlock *TrueBlock = nullptr;
7794 BasicBlock *FalseBlock = nullptr;
7795 BasicBlock *EndBlock = nullptr;
7796 BranchInst *TrueBranch = nullptr;
7797 BranchInst *FalseBranch = nullptr;
7798 if (TrueInstrs.size() == 0) {
7799 FalseBranch = cast<BranchInst>(Val: SplitBlockAndInsertIfElse(
7800 Cond: CondFr, SplitBefore: SplitPt, Unreachable: false, BranchWeights: nullptr, DTU: nullptr, LI));
7801 FalseBlock = FalseBranch->getParent();
7802 EndBlock = cast<BasicBlock>(Val: FalseBranch->getOperand(i_nocapture: 0));
7803 } else if (FalseInstrs.size() == 0) {
7804 TrueBranch = cast<BranchInst>(Val: SplitBlockAndInsertIfThen(
7805 Cond: CondFr, SplitBefore: SplitPt, Unreachable: false, BranchWeights: nullptr, DTU: nullptr, LI));
7806 TrueBlock = TrueBranch->getParent();
7807 EndBlock = cast<BasicBlock>(Val: TrueBranch->getOperand(i_nocapture: 0));
7808 } else {
7809 Instruction *ThenTerm = nullptr;
7810 Instruction *ElseTerm = nullptr;
7811 SplitBlockAndInsertIfThenElse(Cond: CondFr, SplitBefore: SplitPt, ThenTerm: &ThenTerm, ElseTerm: &ElseTerm,
7812 BranchWeights: nullptr, DTU: nullptr, LI);
7813 TrueBranch = cast<BranchInst>(Val: ThenTerm);
7814 FalseBranch = cast<BranchInst>(Val: ElseTerm);
7815 TrueBlock = TrueBranch->getParent();
7816 FalseBlock = FalseBranch->getParent();
7817 EndBlock = cast<BasicBlock>(Val: TrueBranch->getOperand(i_nocapture: 0));
7818 }
7819
7820 EndBlock->setName("select.end");
7821 if (TrueBlock)
7822 TrueBlock->setName("select.true.sink");
7823 if (FalseBlock)
7824 FalseBlock->setName(FalseInstrs.size() == 0 ? "select.false"
7825 : "select.false.sink");
7826
7827 if (IsHugeFunc) {
7828 if (TrueBlock)
7829 FreshBBs.insert(Ptr: TrueBlock);
7830 if (FalseBlock)
7831 FreshBBs.insert(Ptr: FalseBlock);
7832 FreshBBs.insert(Ptr: EndBlock);
7833 }
7834
7835 BFI->setBlockFreq(BB: EndBlock, Freq: BFI->getBlockFreq(BB: StartBlock));
7836
7837 static const unsigned MD[] = {
7838 LLVMContext::MD_prof, LLVMContext::MD_unpredictable,
7839 LLVMContext::MD_make_implicit, LLVMContext::MD_dbg};
7840 StartBlock->getTerminator()->copyMetadata(SrcInst: *SI, WL: MD);
7841
7842 // Sink expensive instructions into the conditional blocks to avoid executing
7843 // them speculatively.
7844 for (Instruction *I : TrueInstrs)
7845 I->moveBefore(InsertPos: TrueBranch->getIterator());
7846 for (Instruction *I : FalseInstrs)
7847 I->moveBefore(InsertPos: FalseBranch->getIterator());
7848
7849 // If we did not create a new block for one of the 'true' or 'false' paths
7850 // of the condition, it means that side of the branch goes to the end block
7851 // directly and the path originates from the start block from the point of
7852 // view of the new PHI.
7853 if (TrueBlock == nullptr)
7854 TrueBlock = StartBlock;
7855 else if (FalseBlock == nullptr)
7856 FalseBlock = StartBlock;
7857
7858 SmallPtrSet<const Instruction *, 2> INS(llvm::from_range, ASI);
7859 // Use reverse iterator because later select may use the value of the
7860 // earlier select, and we need to propagate value through earlier select
7861 // to get the PHI operand.
7862 for (SelectInst *SI : llvm::reverse(C&: ASI)) {
7863 // The select itself is replaced with a PHI Node.
7864 PHINode *PN = PHINode::Create(Ty: SI->getType(), NumReservedValues: 2, NameStr: "");
7865 PN->insertBefore(InsertPos: EndBlock->begin());
7866 PN->takeName(V: SI);
7867 PN->addIncoming(V: getTrueOrFalseValue(SI, isTrue: true, Selects: INS), BB: TrueBlock);
7868 PN->addIncoming(V: getTrueOrFalseValue(SI, isTrue: false, Selects: INS), BB: FalseBlock);
7869 PN->setDebugLoc(SI->getDebugLoc());
7870
7871 replaceAllUsesWith(Old: SI, New: PN, FreshBBs, IsHuge: IsHugeFunc);
7872 SI->eraseFromParent();
7873 INS.erase(Ptr: SI);
7874 ++NumSelectsExpanded;
7875 }
7876
7877 // Instruct OptimizeBlock to skip to the next block.
7878 CurInstIterator = StartBlock->end();
7879 return true;
7880}
7881
7882/// Some targets only accept certain types for splat inputs. For example a VDUP
7883/// in MVE takes a GPR (integer) register, and the instruction that incorporate
7884/// a VDUP (such as a VADD qd, qm, rm) also require a gpr register.
7885bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
7886 // Accept shuf(insertelem(undef/poison, val, 0), undef/poison, <0,0,..>) only
7887 if (!match(V: SVI, P: m_Shuffle(v1: m_InsertElt(Val: m_Undef(), Elt: m_Value(), Idx: m_ZeroInt()),
7888 v2: m_Undef(), mask: m_ZeroMask())))
7889 return false;
7890 Type *NewType = TLI->shouldConvertSplatType(SVI);
7891 if (!NewType)
7892 return false;
7893
7894 auto *SVIVecType = cast<FixedVectorType>(Val: SVI->getType());
7895 assert(!NewType->isVectorTy() && "Expected a scalar type!");
7896 assert(NewType->getScalarSizeInBits() == SVIVecType->getScalarSizeInBits() &&
7897 "Expected a type of the same size!");
7898 auto *NewVecType =
7899 FixedVectorType::get(ElementType: NewType, NumElts: SVIVecType->getNumElements());
7900
7901 // Create a bitcast (shuffle (insert (bitcast(..))))
7902 IRBuilder<> Builder(SVI->getContext());
7903 Builder.SetInsertPoint(SVI);
7904 Value *BC1 = Builder.CreateBitCast(
7905 V: cast<Instruction>(Val: SVI->getOperand(i_nocapture: 0))->getOperand(i: 1), DestTy: NewType);
7906 Value *Shuffle = Builder.CreateVectorSplat(NumElts: NewVecType->getNumElements(), V: BC1);
7907 Value *BC2 = Builder.CreateBitCast(V: Shuffle, DestTy: SVIVecType);
7908
7909 replaceAllUsesWith(Old: SVI, New: BC2, FreshBBs, IsHuge: IsHugeFunc);
7910 RecursivelyDeleteTriviallyDeadInstructions(
7911 V: SVI, TLI: TLInfo, MSSAU: nullptr,
7912 AboutToDeleteCallback: [&](Value *V) { removeAllAssertingVHReferences(V); });
7913
7914 // Also hoist the bitcast up to its operand if it they are not in the same
7915 // block.
7916 if (auto *BCI = dyn_cast<Instruction>(Val: BC1))
7917 if (auto *Op = dyn_cast<Instruction>(Val: BCI->getOperand(i: 0)))
7918 if (BCI->getParent() != Op->getParent() && !isa<PHINode>(Val: Op) &&
7919 !Op->isTerminator() && !Op->isEHPad())
7920 BCI->moveAfter(MovePos: Op);
7921
7922 return true;
7923}
7924
7925bool CodeGenPrepare::tryToSinkFreeOperands(Instruction *I) {
7926 // If the operands of I can be folded into a target instruction together with
7927 // I, duplicate and sink them.
7928 SmallVector<Use *, 4> OpsToSink;
7929 if (!TTI->isProfitableToSinkOperands(I, Ops&: OpsToSink))
7930 return false;
7931
7932 // OpsToSink can contain multiple uses in a use chain (e.g.
7933 // (%u1 with %u1 = shufflevector), (%u2 with %u2 = zext %u1)). The dominating
7934 // uses must come first, so we process the ops in reverse order so as to not
7935 // create invalid IR.
7936 BasicBlock *TargetBB = I->getParent();
7937 bool Changed = false;
7938 SmallVector<Use *, 4> ToReplace;
7939 Instruction *InsertPoint = I;
7940 DenseMap<const Instruction *, unsigned long> InstOrdering;
7941 unsigned long InstNumber = 0;
7942 for (const auto &I : *TargetBB)
7943 InstOrdering[&I] = InstNumber++;
7944
7945 for (Use *U : reverse(C&: OpsToSink)) {
7946 auto *UI = cast<Instruction>(Val: U->get());
7947 if (isa<PHINode>(Val: UI) || UI->mayHaveSideEffects() || UI->mayReadFromMemory())
7948 continue;
7949 if (UI->getParent() == TargetBB) {
7950 if (InstOrdering[UI] < InstOrdering[InsertPoint])
7951 InsertPoint = UI;
7952 continue;
7953 }
7954 ToReplace.push_back(Elt: U);
7955 }
7956
7957 SetVector<Instruction *> MaybeDead;
7958 DenseMap<Instruction *, Instruction *> NewInstructions;
7959 for (Use *U : ToReplace) {
7960 auto *UI = cast<Instruction>(Val: U->get());
7961 Instruction *NI = UI->clone();
7962
7963 if (IsHugeFunc) {
7964 // Now we clone an instruction, its operands' defs may sink to this BB
7965 // now. So we put the operands defs' BBs into FreshBBs to do optimization.
7966 for (Value *Op : NI->operands())
7967 if (auto *OpDef = dyn_cast<Instruction>(Val: Op))
7968 FreshBBs.insert(Ptr: OpDef->getParent());
7969 }
7970
7971 NewInstructions[UI] = NI;
7972 MaybeDead.insert(X: UI);
7973 LLVM_DEBUG(dbgs() << "Sinking " << *UI << " to user " << *I << "\n");
7974 NI->insertBefore(InsertPos: InsertPoint->getIterator());
7975 InsertPoint = NI;
7976 InsertedInsts.insert(Ptr: NI);
7977
7978 // Update the use for the new instruction, making sure that we update the
7979 // sunk instruction uses, if it is part of a chain that has already been
7980 // sunk.
7981 Instruction *OldI = cast<Instruction>(Val: U->getUser());
7982 if (auto It = NewInstructions.find(Val: OldI); It != NewInstructions.end())
7983 It->second->setOperand(i: U->getOperandNo(), Val: NI);
7984 else
7985 U->set(NI);
7986 Changed = true;
7987 }
7988
7989 // Remove instructions that are dead after sinking.
7990 for (auto *I : MaybeDead) {
7991 if (!I->hasNUsesOrMore(N: 1)) {
7992 LLVM_DEBUG(dbgs() << "Removing dead instruction: " << *I << "\n");
7993 I->eraseFromParent();
7994 }
7995 }
7996
7997 return Changed;
7998}
7999
8000bool CodeGenPrepare::optimizeSwitchType(SwitchInst *SI) {
8001 Value *Cond = SI->getCondition();
8002 Type *OldType = Cond->getType();
8003 LLVMContext &Context = Cond->getContext();
8004 EVT OldVT = TLI->getValueType(DL: *DL, Ty: OldType);
8005 MVT RegType = TLI->getPreferredSwitchConditionType(Context, ConditionVT: OldVT);
8006 unsigned RegWidth = RegType.getSizeInBits();
8007
8008 if (RegWidth <= cast<IntegerType>(Val: OldType)->getBitWidth())
8009 return false;
8010
8011 // If the register width is greater than the type width, expand the condition
8012 // of the switch instruction and each case constant to the width of the
8013 // register. By widening the type of the switch condition, subsequent
8014 // comparisons (for case comparisons) will not need to be extended to the
8015 // preferred register width, so we will potentially eliminate N-1 extends,
8016 // where N is the number of cases in the switch.
8017 auto *NewType = Type::getIntNTy(C&: Context, N: RegWidth);
8018
8019 // Extend the switch condition and case constants using the target preferred
8020 // extend unless the switch condition is a function argument with an extend
8021 // attribute. In that case, we can avoid an unnecessary mask/extension by
8022 // matching the argument extension instead.
8023 Instruction::CastOps ExtType = Instruction::ZExt;
8024 // Some targets prefer SExt over ZExt.
8025 if (TLI->isSExtCheaperThanZExt(FromTy: OldVT, ToTy: RegType))
8026 ExtType = Instruction::SExt;
8027
8028 if (auto *Arg = dyn_cast<Argument>(Val: Cond)) {
8029 if (Arg->hasSExtAttr())
8030 ExtType = Instruction::SExt;
8031 if (Arg->hasZExtAttr())
8032 ExtType = Instruction::ZExt;
8033 }
8034
8035 auto *ExtInst = CastInst::Create(ExtType, S: Cond, Ty: NewType);
8036 ExtInst->insertBefore(InsertPos: SI->getIterator());
8037 ExtInst->setDebugLoc(SI->getDebugLoc());
8038 SI->setCondition(ExtInst);
8039 for (auto Case : SI->cases()) {
8040 const APInt &NarrowConst = Case.getCaseValue()->getValue();
8041 APInt WideConst = (ExtType == Instruction::ZExt)
8042 ? NarrowConst.zext(width: RegWidth)
8043 : NarrowConst.sext(width: RegWidth);
8044 Case.setValue(ConstantInt::get(Context, V: WideConst));
8045 }
8046
8047 return true;
8048}
8049
8050bool CodeGenPrepare::optimizeSwitchPhiConstants(SwitchInst *SI) {
8051 // The SCCP optimization tends to produce code like this:
8052 // switch(x) { case 42: phi(42, ...) }
8053 // Materializing the constant for the phi-argument needs instructions; So we
8054 // change the code to:
8055 // switch(x) { case 42: phi(x, ...) }
8056
8057 Value *Condition = SI->getCondition();
8058 // Avoid endless loop in degenerate case.
8059 if (isa<ConstantInt>(Val: *Condition))
8060 return false;
8061
8062 bool Changed = false;
8063 BasicBlock *SwitchBB = SI->getParent();
8064 Type *ConditionType = Condition->getType();
8065
8066 for (const SwitchInst::CaseHandle &Case : SI->cases()) {
8067 ConstantInt *CaseValue = Case.getCaseValue();
8068 BasicBlock *CaseBB = Case.getCaseSuccessor();
8069 // Set to true if we previously checked that `CaseBB` is only reached by
8070 // a single case from this switch.
8071 bool CheckedForSinglePred = false;
8072 for (PHINode &PHI : CaseBB->phis()) {
8073 Type *PHIType = PHI.getType();
8074 // If ZExt is free then we can also catch patterns like this:
8075 // switch((i32)x) { case 42: phi((i64)42, ...); }
8076 // and replace `(i64)42` with `zext i32 %x to i64`.
8077 bool TryZExt =
8078 PHIType->isIntegerTy() &&
8079 PHIType->getIntegerBitWidth() > ConditionType->getIntegerBitWidth() &&
8080 TLI->isZExtFree(FromTy: ConditionType, ToTy: PHIType);
8081 if (PHIType == ConditionType || TryZExt) {
8082 // Set to true to skip this case because of multiple preds.
8083 bool SkipCase = false;
8084 Value *Replacement = nullptr;
8085 for (unsigned I = 0, E = PHI.getNumIncomingValues(); I != E; I++) {
8086 Value *PHIValue = PHI.getIncomingValue(i: I);
8087 if (PHIValue != CaseValue) {
8088 if (!TryZExt)
8089 continue;
8090 ConstantInt *PHIValueInt = dyn_cast<ConstantInt>(Val: PHIValue);
8091 if (!PHIValueInt ||
8092 PHIValueInt->getValue() !=
8093 CaseValue->getValue().zext(width: PHIType->getIntegerBitWidth()))
8094 continue;
8095 }
8096 if (PHI.getIncomingBlock(i: I) != SwitchBB)
8097 continue;
8098 // We cannot optimize if there are multiple case labels jumping to
8099 // this block. This check may get expensive when there are many
8100 // case labels so we test for it last.
8101 if (!CheckedForSinglePred) {
8102 CheckedForSinglePred = true;
8103 if (SI->findCaseDest(BB: CaseBB) == nullptr) {
8104 SkipCase = true;
8105 break;
8106 }
8107 }
8108
8109 if (Replacement == nullptr) {
8110 if (PHIValue == CaseValue) {
8111 Replacement = Condition;
8112 } else {
8113 IRBuilder<> Builder(SI);
8114 Replacement = Builder.CreateZExt(V: Condition, DestTy: PHIType);
8115 }
8116 }
8117 PHI.setIncomingValue(i: I, V: Replacement);
8118 Changed = true;
8119 }
8120 if (SkipCase)
8121 break;
8122 }
8123 }
8124 }
8125 return Changed;
8126}
8127
8128bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
8129 bool Changed = optimizeSwitchType(SI);
8130 Changed |= optimizeSwitchPhiConstants(SI);
8131 return Changed;
8132}
8133
8134namespace {
8135
8136/// Helper class to promote a scalar operation to a vector one.
8137/// This class is used to move downward extractelement transition.
8138/// E.g.,
8139/// a = vector_op <2 x i32>
8140/// b = extractelement <2 x i32> a, i32 0
8141/// c = scalar_op b
8142/// store c
8143///
8144/// =>
8145/// a = vector_op <2 x i32>
8146/// c = vector_op a (equivalent to scalar_op on the related lane)
8147/// * d = extractelement <2 x i32> c, i32 0
8148/// * store d
8149/// Assuming both extractelement and store can be combine, we get rid of the
8150/// transition.
8151class VectorPromoteHelper {
8152 /// DataLayout associated with the current module.
8153 const DataLayout &DL;
8154
8155 /// Used to perform some checks on the legality of vector operations.
8156 const TargetLowering &TLI;
8157
8158 /// Used to estimated the cost of the promoted chain.
8159 const TargetTransformInfo &TTI;
8160
8161 /// The transition being moved downwards.
8162 Instruction *Transition;
8163
8164 /// The sequence of instructions to be promoted.
8165 SmallVector<Instruction *, 4> InstsToBePromoted;
8166
8167 /// Cost of combining a store and an extract.
8168 unsigned StoreExtractCombineCost;
8169
8170 /// Instruction that will be combined with the transition.
8171 Instruction *CombineInst = nullptr;
8172
8173 /// The instruction that represents the current end of the transition.
8174 /// Since we are faking the promotion until we reach the end of the chain
8175 /// of computation, we need a way to get the current end of the transition.
8176 Instruction *getEndOfTransition() const {
8177 if (InstsToBePromoted.empty())
8178 return Transition;
8179 return InstsToBePromoted.back();
8180 }
8181
8182 /// Return the index of the original value in the transition.
8183 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
8184 /// c, is at index 0.
8185 unsigned getTransitionOriginalValueIdx() const {
8186 assert(isa<ExtractElementInst>(Transition) &&
8187 "Other kind of transitions are not supported yet");
8188 return 0;
8189 }
8190
8191 /// Return the index of the index in the transition.
8192 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
8193 /// is at index 1.
8194 unsigned getTransitionIdx() const {
8195 assert(isa<ExtractElementInst>(Transition) &&
8196 "Other kind of transitions are not supported yet");
8197 return 1;
8198 }
8199
8200 /// Get the type of the transition.
8201 /// This is the type of the original value.
8202 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
8203 /// transition is <2 x i32>.
8204 Type *getTransitionType() const {
8205 return Transition->getOperand(i: getTransitionOriginalValueIdx())->getType();
8206 }
8207
8208 /// Promote \p ToBePromoted by moving \p Def downward through.
8209 /// I.e., we have the following sequence:
8210 /// Def = Transition <ty1> a to <ty2>
8211 /// b = ToBePromoted <ty2> Def, ...
8212 /// =>
8213 /// b = ToBePromoted <ty1> a, ...
8214 /// Def = Transition <ty1> ToBePromoted to <ty2>
8215 void promoteImpl(Instruction *ToBePromoted);
8216
8217 /// Check whether or not it is profitable to promote all the
8218 /// instructions enqueued to be promoted.
8219 bool isProfitableToPromote() {
8220 Value *ValIdx = Transition->getOperand(i: getTransitionOriginalValueIdx());
8221 unsigned Index = isa<ConstantInt>(Val: ValIdx)
8222 ? cast<ConstantInt>(Val: ValIdx)->getZExtValue()
8223 : -1;
8224 Type *PromotedType = getTransitionType();
8225
8226 StoreInst *ST = cast<StoreInst>(Val: CombineInst);
8227 unsigned AS = ST->getPointerAddressSpace();
8228 // Check if this store is supported.
8229 if (!TLI.allowsMisalignedMemoryAccesses(
8230 TLI.getValueType(DL, Ty: ST->getValueOperand()->getType()), AddrSpace: AS,
8231 Alignment: ST->getAlign())) {
8232 // If this is not supported, there is no way we can combine
8233 // the extract with the store.
8234 return false;
8235 }
8236
8237 // The scalar chain of computation has to pay for the transition
8238 // scalar to vector.
8239 // The vector chain has to account for the combining cost.
8240 enum TargetTransformInfo::TargetCostKind CostKind =
8241 TargetTransformInfo::TCK_RecipThroughput;
8242 InstructionCost ScalarCost =
8243 TTI.getVectorInstrCost(I: *Transition, Val: PromotedType, CostKind, Index);
8244 InstructionCost VectorCost = StoreExtractCombineCost;
8245 for (const auto &Inst : InstsToBePromoted) {
8246 // Compute the cost.
8247 // By construction, all instructions being promoted are arithmetic ones.
8248 // Moreover, one argument is a constant that can be viewed as a splat
8249 // constant.
8250 Value *Arg0 = Inst->getOperand(i: 0);
8251 bool IsArg0Constant = isa<UndefValue>(Val: Arg0) || isa<ConstantInt>(Val: Arg0) ||
8252 isa<ConstantFP>(Val: Arg0);
8253 TargetTransformInfo::OperandValueInfo Arg0Info, Arg1Info;
8254 if (IsArg0Constant)
8255 Arg0Info.Kind = TargetTransformInfo::OK_UniformConstantValue;
8256 else
8257 Arg1Info.Kind = TargetTransformInfo::OK_UniformConstantValue;
8258
8259 ScalarCost += TTI.getArithmeticInstrCost(
8260 Opcode: Inst->getOpcode(), Ty: Inst->getType(), CostKind, Opd1Info: Arg0Info, Opd2Info: Arg1Info);
8261 VectorCost += TTI.getArithmeticInstrCost(Opcode: Inst->getOpcode(), Ty: PromotedType,
8262 CostKind, Opd1Info: Arg0Info, Opd2Info: Arg1Info);
8263 }
8264 LLVM_DEBUG(
8265 dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
8266 << ScalarCost << "\nVector: " << VectorCost << '\n');
8267 return ScalarCost > VectorCost;
8268 }
8269
8270 /// Generate a constant vector with \p Val with the same
8271 /// number of elements as the transition.
8272 /// \p UseSplat defines whether or not \p Val should be replicated
8273 /// across the whole vector.
8274 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
8275 /// otherwise we generate a vector with as many poison as possible:
8276 /// <poison, ..., poison, Val, poison, ..., poison> where \p Val is only
8277 /// used at the index of the extract.
8278 Value *getConstantVector(Constant *Val, bool UseSplat) const {
8279 unsigned ExtractIdx = std::numeric_limits<unsigned>::max();
8280 if (!UseSplat) {
8281 // If we cannot determine where the constant must be, we have to
8282 // use a splat constant.
8283 Value *ValExtractIdx = Transition->getOperand(i: getTransitionIdx());
8284 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Val: ValExtractIdx))
8285 ExtractIdx = CstVal->getSExtValue();
8286 else
8287 UseSplat = true;
8288 }
8289
8290 ElementCount EC = cast<VectorType>(Val: getTransitionType())->getElementCount();
8291 if (UseSplat)
8292 return ConstantVector::getSplat(EC, Elt: Val);
8293
8294 if (!EC.isScalable()) {
8295 SmallVector<Constant *, 4> ConstVec;
8296 PoisonValue *PoisonVal = PoisonValue::get(T: Val->getType());
8297 for (unsigned Idx = 0; Idx != EC.getKnownMinValue(); ++Idx) {
8298 if (Idx == ExtractIdx)
8299 ConstVec.push_back(Elt: Val);
8300 else
8301 ConstVec.push_back(Elt: PoisonVal);
8302 }
8303 return ConstantVector::get(V: ConstVec);
8304 } else
8305 llvm_unreachable(
8306 "Generate scalable vector for non-splat is unimplemented");
8307 }
8308
8309 /// Check if promoting to a vector type an operand at \p OperandIdx
8310 /// in \p Use can trigger undefined behavior.
8311 static bool canCauseUndefinedBehavior(const Instruction *Use,
8312 unsigned OperandIdx) {
8313 // This is not safe to introduce undef when the operand is on
8314 // the right hand side of a division-like instruction.
8315 if (OperandIdx != 1)
8316 return false;
8317 switch (Use->getOpcode()) {
8318 default:
8319 return false;
8320 case Instruction::SDiv:
8321 case Instruction::UDiv:
8322 case Instruction::SRem:
8323 case Instruction::URem:
8324 return true;
8325 case Instruction::FDiv:
8326 case Instruction::FRem:
8327 return !Use->hasNoNaNs();
8328 }
8329 llvm_unreachable(nullptr);
8330 }
8331
8332public:
8333 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
8334 const TargetTransformInfo &TTI, Instruction *Transition,
8335 unsigned CombineCost)
8336 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
8337 StoreExtractCombineCost(CombineCost) {
8338 assert(Transition && "Do not know how to promote null");
8339 }
8340
8341 /// Check if we can promote \p ToBePromoted to \p Type.
8342 bool canPromote(const Instruction *ToBePromoted) const {
8343 // We could support CastInst too.
8344 return isa<BinaryOperator>(Val: ToBePromoted);
8345 }
8346
8347 /// Check if it is profitable to promote \p ToBePromoted
8348 /// by moving downward the transition through.
8349 bool shouldPromote(const Instruction *ToBePromoted) const {
8350 // Promote only if all the operands can be statically expanded.
8351 // Indeed, we do not want to introduce any new kind of transitions.
8352 for (const Use &U : ToBePromoted->operands()) {
8353 const Value *Val = U.get();
8354 if (Val == getEndOfTransition()) {
8355 // If the use is a division and the transition is on the rhs,
8356 // we cannot promote the operation, otherwise we may create a
8357 // division by zero.
8358 if (canCauseUndefinedBehavior(Use: ToBePromoted, OperandIdx: U.getOperandNo()))
8359 return false;
8360 continue;
8361 }
8362 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
8363 !isa<ConstantFP>(Val))
8364 return false;
8365 }
8366 // Check that the resulting operation is legal.
8367 int ISDOpcode = TLI.InstructionOpcodeToISD(Opcode: ToBePromoted->getOpcode());
8368 if (!ISDOpcode)
8369 return false;
8370 return StressStoreExtract ||
8371 TLI.isOperationLegalOrCustom(
8372 Op: ISDOpcode, VT: TLI.getValueType(DL, Ty: getTransitionType(), AllowUnknown: true));
8373 }
8374
8375 /// Check whether or not \p Use can be combined
8376 /// with the transition.
8377 /// I.e., is it possible to do Use(Transition) => AnotherUse?
8378 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Val: Use); }
8379
8380 /// Record \p ToBePromoted as part of the chain to be promoted.
8381 void enqueueForPromotion(Instruction *ToBePromoted) {
8382 InstsToBePromoted.push_back(Elt: ToBePromoted);
8383 }
8384
8385 /// Set the instruction that will be combined with the transition.
8386 void recordCombineInstruction(Instruction *ToBeCombined) {
8387 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
8388 CombineInst = ToBeCombined;
8389 }
8390
8391 /// Promote all the instructions enqueued for promotion if it is
8392 /// is profitable.
8393 /// \return True if the promotion happened, false otherwise.
8394 bool promote() {
8395 // Check if there is something to promote.
8396 // Right now, if we do not have anything to combine with,
8397 // we assume the promotion is not profitable.
8398 if (InstsToBePromoted.empty() || !CombineInst)
8399 return false;
8400
8401 // Check cost.
8402 if (!StressStoreExtract && !isProfitableToPromote())
8403 return false;
8404
8405 // Promote.
8406 for (auto &ToBePromoted : InstsToBePromoted)
8407 promoteImpl(ToBePromoted);
8408 InstsToBePromoted.clear();
8409 return true;
8410 }
8411};
8412
8413} // end anonymous namespace
8414
8415void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
8416 // At this point, we know that all the operands of ToBePromoted but Def
8417 // can be statically promoted.
8418 // For Def, we need to use its parameter in ToBePromoted:
8419 // b = ToBePromoted ty1 a
8420 // Def = Transition ty1 b to ty2
8421 // Move the transition down.
8422 // 1. Replace all uses of the promoted operation by the transition.
8423 // = ... b => = ... Def.
8424 assert(ToBePromoted->getType() == Transition->getType() &&
8425 "The type of the result of the transition does not match "
8426 "the final type");
8427 ToBePromoted->replaceAllUsesWith(V: Transition);
8428 // 2. Update the type of the uses.
8429 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
8430 Type *TransitionTy = getTransitionType();
8431 ToBePromoted->mutateType(Ty: TransitionTy);
8432 // 3. Update all the operands of the promoted operation with promoted
8433 // operands.
8434 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
8435 for (Use &U : ToBePromoted->operands()) {
8436 Value *Val = U.get();
8437 Value *NewVal = nullptr;
8438 if (Val == Transition)
8439 NewVal = Transition->getOperand(i: getTransitionOriginalValueIdx());
8440 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
8441 isa<ConstantFP>(Val)) {
8442 // Use a splat constant if it is not safe to use undef.
8443 NewVal = getConstantVector(
8444 Val: cast<Constant>(Val),
8445 UseSplat: isa<UndefValue>(Val) ||
8446 canCauseUndefinedBehavior(Use: ToBePromoted, OperandIdx: U.getOperandNo()));
8447 } else
8448 llvm_unreachable("Did you modified shouldPromote and forgot to update "
8449 "this?");
8450 ToBePromoted->setOperand(i: U.getOperandNo(), Val: NewVal);
8451 }
8452 Transition->moveAfter(MovePos: ToBePromoted);
8453 Transition->setOperand(i: getTransitionOriginalValueIdx(), Val: ToBePromoted);
8454}
8455
8456/// Some targets can do store(extractelement) with one instruction.
8457/// Try to push the extractelement towards the stores when the target
8458/// has this feature and this is profitable.
8459bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
8460 unsigned CombineCost = std::numeric_limits<unsigned>::max();
8461 if (DisableStoreExtract ||
8462 (!StressStoreExtract &&
8463 !TLI->canCombineStoreAndExtract(VectorTy: Inst->getOperand(i: 0)->getType(),
8464 Idx: Inst->getOperand(i: 1), Cost&: CombineCost)))
8465 return false;
8466
8467 // At this point we know that Inst is a vector to scalar transition.
8468 // Try to move it down the def-use chain, until:
8469 // - We can combine the transition with its single use
8470 // => we got rid of the transition.
8471 // - We escape the current basic block
8472 // => we would need to check that we are moving it at a cheaper place and
8473 // we do not do that for now.
8474 BasicBlock *Parent = Inst->getParent();
8475 LLVM_DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
8476 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
8477 // If the transition has more than one use, assume this is not going to be
8478 // beneficial.
8479 while (Inst->hasOneUse()) {
8480 Instruction *ToBePromoted = cast<Instruction>(Val: *Inst->user_begin());
8481 LLVM_DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
8482
8483 if (ToBePromoted->getParent() != Parent) {
8484 LLVM_DEBUG(dbgs() << "Instruction to promote is in a different block ("
8485 << ToBePromoted->getParent()->getName()
8486 << ") than the transition (" << Parent->getName()
8487 << ").\n");
8488 return false;
8489 }
8490
8491 if (VPH.canCombine(Use: ToBePromoted)) {
8492 LLVM_DEBUG(dbgs() << "Assume " << *Inst << '\n'
8493 << "will be combined with: " << *ToBePromoted << '\n');
8494 VPH.recordCombineInstruction(ToBeCombined: ToBePromoted);
8495 bool Changed = VPH.promote();
8496 NumStoreExtractExposed += Changed;
8497 return Changed;
8498 }
8499
8500 LLVM_DEBUG(dbgs() << "Try promoting.\n");
8501 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
8502 return false;
8503
8504 LLVM_DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
8505
8506 VPH.enqueueForPromotion(ToBePromoted);
8507 Inst = ToBePromoted;
8508 }
8509 return false;
8510}
8511
8512/// For the instruction sequence of store below, F and I values
8513/// are bundled together as an i64 value before being stored into memory.
8514/// Sometimes it is more efficient to generate separate stores for F and I,
8515/// which can remove the bitwise instructions or sink them to colder places.
8516///
8517/// (store (or (zext (bitcast F to i32) to i64),
8518/// (shl (zext I to i64), 32)), addr) -->
8519/// (store F, addr) and (store I, addr+4)
8520///
8521/// Similarly, splitting for other merged store can also be beneficial, like:
8522/// For pair of {i32, i32}, i64 store --> two i32 stores.
8523/// For pair of {i32, i16}, i64 store --> two i32 stores.
8524/// For pair of {i16, i16}, i32 store --> two i16 stores.
8525/// For pair of {i16, i8}, i32 store --> two i16 stores.
8526/// For pair of {i8, i8}, i16 store --> two i8 stores.
8527///
8528/// We allow each target to determine specifically which kind of splitting is
8529/// supported.
8530///
8531/// The store patterns are commonly seen from the simple code snippet below
8532/// if only std::make_pair(...) is sroa transformed before inlined into hoo.
8533/// void goo(const std::pair<int, float> &);
8534/// hoo() {
8535/// ...
8536/// goo(std::make_pair(tmp, ftmp));
8537/// ...
8538/// }
8539///
8540/// Although we already have similar splitting in DAG Combine, we duplicate
8541/// it in CodeGenPrepare to catch the case in which pattern is across
8542/// multiple BBs. The logic in DAG Combine is kept to catch case generated
8543/// during code expansion.
8544static bool splitMergedValStore(StoreInst &SI, const DataLayout &DL,
8545 const TargetLowering &TLI) {
8546 // Handle simple but common cases only.
8547 Type *StoreType = SI.getValueOperand()->getType();
8548
8549 // The code below assumes shifting a value by <number of bits>,
8550 // whereas scalable vectors would have to be shifted by
8551 // <2log(vscale) + number of bits> in order to store the
8552 // low/high parts. Bailing out for now.
8553 if (StoreType->isScalableTy())
8554 return false;
8555
8556 if (!DL.typeSizeEqualsStoreSize(Ty: StoreType) ||
8557 DL.getTypeSizeInBits(Ty: StoreType) == 0)
8558 return false;
8559
8560 unsigned HalfValBitSize = DL.getTypeSizeInBits(Ty: StoreType) / 2;
8561 Type *SplitStoreType = Type::getIntNTy(C&: SI.getContext(), N: HalfValBitSize);
8562 if (!DL.typeSizeEqualsStoreSize(Ty: SplitStoreType))
8563 return false;
8564
8565 // Don't split the store if it is volatile.
8566 if (SI.isVolatile())
8567 return false;
8568
8569 // Match the following patterns:
8570 // (store (or (zext LValue to i64),
8571 // (shl (zext HValue to i64), 32)), HalfValBitSize)
8572 // or
8573 // (store (or (shl (zext HValue to i64), 32)), HalfValBitSize)
8574 // (zext LValue to i64),
8575 // Expect both operands of OR and the first operand of SHL have only
8576 // one use.
8577 Value *LValue, *HValue;
8578 if (!match(V: SI.getValueOperand(),
8579 P: m_c_Or(L: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: LValue))),
8580 R: m_OneUse(SubPattern: m_Shl(L: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: HValue))),
8581 R: m_SpecificInt(V: HalfValBitSize))))))
8582 return false;
8583
8584 // Check LValue and HValue are int with size less or equal than 32.
8585 if (!LValue->getType()->isIntegerTy() ||
8586 DL.getTypeSizeInBits(Ty: LValue->getType()) > HalfValBitSize ||
8587 !HValue->getType()->isIntegerTy() ||
8588 DL.getTypeSizeInBits(Ty: HValue->getType()) > HalfValBitSize)
8589 return false;
8590
8591 // If LValue/HValue is a bitcast instruction, use the EVT before bitcast
8592 // as the input of target query.
8593 auto *LBC = dyn_cast<BitCastInst>(Val: LValue);
8594 auto *HBC = dyn_cast<BitCastInst>(Val: HValue);
8595 EVT LowTy = LBC ? EVT::getEVT(Ty: LBC->getOperand(i_nocapture: 0)->getType())
8596 : EVT::getEVT(Ty: LValue->getType());
8597 EVT HighTy = HBC ? EVT::getEVT(Ty: HBC->getOperand(i_nocapture: 0)->getType())
8598 : EVT::getEVT(Ty: HValue->getType());
8599 if (!ForceSplitStore && !TLI.isMultiStoresCheaperThanBitsMerge(LTy: LowTy, HTy: HighTy))
8600 return false;
8601
8602 // Start to split store.
8603 IRBuilder<> Builder(SI.getContext());
8604 Builder.SetInsertPoint(&SI);
8605
8606 // If LValue/HValue is a bitcast in another BB, create a new one in current
8607 // BB so it may be merged with the splitted stores by dag combiner.
8608 if (LBC && LBC->getParent() != SI.getParent())
8609 LValue = Builder.CreateBitCast(V: LBC->getOperand(i_nocapture: 0), DestTy: LBC->getType());
8610 if (HBC && HBC->getParent() != SI.getParent())
8611 HValue = Builder.CreateBitCast(V: HBC->getOperand(i_nocapture: 0), DestTy: HBC->getType());
8612
8613 bool IsLE = SI.getDataLayout().isLittleEndian();
8614 auto CreateSplitStore = [&](Value *V, bool Upper) {
8615 V = Builder.CreateZExtOrBitCast(V, DestTy: SplitStoreType);
8616 Value *Addr = SI.getPointerOperand();
8617 Align Alignment = SI.getAlign();
8618 const bool IsOffsetStore = (IsLE && Upper) || (!IsLE && !Upper);
8619 if (IsOffsetStore) {
8620 Addr = Builder.CreateGEP(
8621 Ty: SplitStoreType, Ptr: Addr,
8622 IdxList: ConstantInt::get(Ty: Type::getInt32Ty(C&: SI.getContext()), V: 1));
8623
8624 // When splitting the store in half, naturally one half will retain the
8625 // alignment of the original wider store, regardless of whether it was
8626 // over-aligned or not, while the other will require adjustment.
8627 Alignment = commonAlignment(A: Alignment, Offset: HalfValBitSize / 8);
8628 }
8629 Builder.CreateAlignedStore(Val: V, Ptr: Addr, Align: Alignment);
8630 };
8631
8632 CreateSplitStore(LValue, false);
8633 CreateSplitStore(HValue, true);
8634
8635 // Delete the old store.
8636 SI.eraseFromParent();
8637 return true;
8638}
8639
8640// Return true if the GEP has two operands, the first operand is of a sequential
8641// type, and the second operand is a constant.
8642static bool GEPSequentialConstIndexed(GetElementPtrInst *GEP) {
8643 gep_type_iterator I = gep_type_begin(GEP: *GEP);
8644 return GEP->getNumOperands() == 2 && I.isSequential() &&
8645 isa<ConstantInt>(Val: GEP->getOperand(i_nocapture: 1));
8646}
8647
8648// Try unmerging GEPs to reduce liveness interference (register pressure) across
8649// IndirectBr edges. Since IndirectBr edges tend to touch on many blocks,
8650// reducing liveness interference across those edges benefits global register
8651// allocation. Currently handles only certain cases.
8652//
8653// For example, unmerge %GEPI and %UGEPI as below.
8654//
8655// ---------- BEFORE ----------
8656// SrcBlock:
8657// ...
8658// %GEPIOp = ...
8659// ...
8660// %GEPI = gep %GEPIOp, Idx
8661// ...
8662// indirectbr ... [ label %DstB0, label %DstB1, ... label %DstBi ... ]
8663// (* %GEPI is alive on the indirectbr edges due to other uses ahead)
8664// (* %GEPIOp is alive on the indirectbr edges only because of it's used by
8665// %UGEPI)
8666//
8667// DstB0: ... (there may be a gep similar to %UGEPI to be unmerged)
8668// DstB1: ... (there may be a gep similar to %UGEPI to be unmerged)
8669// ...
8670//
8671// DstBi:
8672// ...
8673// %UGEPI = gep %GEPIOp, UIdx
8674// ...
8675// ---------------------------
8676//
8677// ---------- AFTER ----------
8678// SrcBlock:
8679// ... (same as above)
8680// (* %GEPI is still alive on the indirectbr edges)
8681// (* %GEPIOp is no longer alive on the indirectbr edges as a result of the
8682// unmerging)
8683// ...
8684//
8685// DstBi:
8686// ...
8687// %UGEPI = gep %GEPI, (UIdx-Idx)
8688// ...
8689// ---------------------------
8690//
8691// The register pressure on the IndirectBr edges is reduced because %GEPIOp is
8692// no longer alive on them.
8693//
8694// We try to unmerge GEPs here in CodGenPrepare, as opposed to limiting merging
8695// of GEPs in the first place in InstCombiner::visitGetElementPtrInst() so as
8696// not to disable further simplications and optimizations as a result of GEP
8697// merging.
8698//
8699// Note this unmerging may increase the length of the data flow critical path
8700// (the path from %GEPIOp to %UGEPI would go through %GEPI), which is a tradeoff
8701// between the register pressure and the length of data-flow critical
8702// path. Restricting this to the uncommon IndirectBr case would minimize the
8703// impact of potentially longer critical path, if any, and the impact on compile
8704// time.
8705static bool tryUnmergingGEPsAcrossIndirectBr(GetElementPtrInst *GEPI,
8706 const TargetTransformInfo *TTI) {
8707 BasicBlock *SrcBlock = GEPI->getParent();
8708 // Check that SrcBlock ends with an IndirectBr. If not, give up. The common
8709 // (non-IndirectBr) cases exit early here.
8710 if (!isa<IndirectBrInst>(Val: SrcBlock->getTerminator()))
8711 return false;
8712 // Check that GEPI is a simple gep with a single constant index.
8713 if (!GEPSequentialConstIndexed(GEP: GEPI))
8714 return false;
8715 ConstantInt *GEPIIdx = cast<ConstantInt>(Val: GEPI->getOperand(i_nocapture: 1));
8716 // Check that GEPI is a cheap one.
8717 if (TTI->getIntImmCost(Imm: GEPIIdx->getValue(), Ty: GEPIIdx->getType(),
8718 CostKind: TargetTransformInfo::TCK_SizeAndLatency) >
8719 TargetTransformInfo::TCC_Basic)
8720 return false;
8721 Value *GEPIOp = GEPI->getOperand(i_nocapture: 0);
8722 // Check that GEPIOp is an instruction that's also defined in SrcBlock.
8723 if (!isa<Instruction>(Val: GEPIOp))
8724 return false;
8725 auto *GEPIOpI = cast<Instruction>(Val: GEPIOp);
8726 if (GEPIOpI->getParent() != SrcBlock)
8727 return false;
8728 // Check that GEP is used outside the block, meaning it's alive on the
8729 // IndirectBr edge(s).
8730 if (llvm::none_of(Range: GEPI->users(), P: [&](User *Usr) {
8731 if (auto *I = dyn_cast<Instruction>(Val: Usr)) {
8732 if (I->getParent() != SrcBlock) {
8733 return true;
8734 }
8735 }
8736 return false;
8737 }))
8738 return false;
8739 // The second elements of the GEP chains to be unmerged.
8740 std::vector<GetElementPtrInst *> UGEPIs;
8741 // Check each user of GEPIOp to check if unmerging would make GEPIOp not alive
8742 // on IndirectBr edges.
8743 for (User *Usr : GEPIOp->users()) {
8744 if (Usr == GEPI)
8745 continue;
8746 // Check if Usr is an Instruction. If not, give up.
8747 if (!isa<Instruction>(Val: Usr))
8748 return false;
8749 auto *UI = cast<Instruction>(Val: Usr);
8750 // Check if Usr in the same block as GEPIOp, which is fine, skip.
8751 if (UI->getParent() == SrcBlock)
8752 continue;
8753 // Check if Usr is a GEP. If not, give up.
8754 if (!isa<GetElementPtrInst>(Val: Usr))
8755 return false;
8756 auto *UGEPI = cast<GetElementPtrInst>(Val: Usr);
8757 // Check if UGEPI is a simple gep with a single constant index and GEPIOp is
8758 // the pointer operand to it. If so, record it in the vector. If not, give
8759 // up.
8760 if (!GEPSequentialConstIndexed(GEP: UGEPI))
8761 return false;
8762 if (UGEPI->getOperand(i_nocapture: 0) != GEPIOp)
8763 return false;
8764 if (UGEPI->getSourceElementType() != GEPI->getSourceElementType())
8765 return false;
8766 if (GEPIIdx->getType() !=
8767 cast<ConstantInt>(Val: UGEPI->getOperand(i_nocapture: 1))->getType())
8768 return false;
8769 ConstantInt *UGEPIIdx = cast<ConstantInt>(Val: UGEPI->getOperand(i_nocapture: 1));
8770 if (TTI->getIntImmCost(Imm: UGEPIIdx->getValue(), Ty: UGEPIIdx->getType(),
8771 CostKind: TargetTransformInfo::TCK_SizeAndLatency) >
8772 TargetTransformInfo::TCC_Basic)
8773 return false;
8774 UGEPIs.push_back(x: UGEPI);
8775 }
8776 if (UGEPIs.size() == 0)
8777 return false;
8778 // Check the materializing cost of (Uidx-Idx).
8779 for (GetElementPtrInst *UGEPI : UGEPIs) {
8780 ConstantInt *UGEPIIdx = cast<ConstantInt>(Val: UGEPI->getOperand(i_nocapture: 1));
8781 APInt NewIdx = UGEPIIdx->getValue() - GEPIIdx->getValue();
8782 InstructionCost ImmCost = TTI->getIntImmCost(
8783 Imm: NewIdx, Ty: GEPIIdx->getType(), CostKind: TargetTransformInfo::TCK_SizeAndLatency);
8784 if (ImmCost > TargetTransformInfo::TCC_Basic)
8785 return false;
8786 }
8787 // Now unmerge between GEPI and UGEPIs.
8788 for (GetElementPtrInst *UGEPI : UGEPIs) {
8789 UGEPI->setOperand(i_nocapture: 0, Val_nocapture: GEPI);
8790 ConstantInt *UGEPIIdx = cast<ConstantInt>(Val: UGEPI->getOperand(i_nocapture: 1));
8791 Constant *NewUGEPIIdx = ConstantInt::get(
8792 Ty: GEPIIdx->getType(), V: UGEPIIdx->getValue() - GEPIIdx->getValue());
8793 UGEPI->setOperand(i_nocapture: 1, Val_nocapture: NewUGEPIIdx);
8794 // If GEPI is not inbounds but UGEPI is inbounds, change UGEPI to not
8795 // inbounds to avoid UB.
8796 if (!GEPI->isInBounds()) {
8797 UGEPI->setIsInBounds(false);
8798 }
8799 }
8800 // After unmerging, verify that GEPIOp is actually only used in SrcBlock (not
8801 // alive on IndirectBr edges).
8802 assert(llvm::none_of(GEPIOp->users(),
8803 [&](User *Usr) {
8804 return cast<Instruction>(Usr)->getParent() != SrcBlock;
8805 }) &&
8806 "GEPIOp is used outside SrcBlock");
8807 return true;
8808}
8809
8810static bool optimizeBranch(BranchInst *Branch, const TargetLowering &TLI,
8811 SmallPtrSet<BasicBlock *, 32> &FreshBBs,
8812 bool IsHugeFunc) {
8813 // Try and convert
8814 // %c = icmp ult %x, 8
8815 // br %c, bla, blb
8816 // %tc = lshr %x, 3
8817 // to
8818 // %tc = lshr %x, 3
8819 // %c = icmp eq %tc, 0
8820 // br %c, bla, blb
8821 // Creating the cmp to zero can be better for the backend, especially if the
8822 // lshr produces flags that can be used automatically.
8823 if (!TLI.preferZeroCompareBranch() || !Branch->isConditional())
8824 return false;
8825
8826 ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: Branch->getCondition());
8827 if (!Cmp || !isa<ConstantInt>(Val: Cmp->getOperand(i_nocapture: 1)) || !Cmp->hasOneUse())
8828 return false;
8829
8830 Value *X = Cmp->getOperand(i_nocapture: 0);
8831 if (!X->hasUseList())
8832 return false;
8833
8834 APInt CmpC = cast<ConstantInt>(Val: Cmp->getOperand(i_nocapture: 1))->getValue();
8835
8836 for (auto *U : X->users()) {
8837 Instruction *UI = dyn_cast<Instruction>(Val: U);
8838 // A quick dominance check
8839 if (!UI ||
8840 (UI->getParent() != Branch->getParent() &&
8841 UI->getParent() != Branch->getSuccessor(i: 0) &&
8842 UI->getParent() != Branch->getSuccessor(i: 1)) ||
8843 (UI->getParent() != Branch->getParent() &&
8844 !UI->getParent()->getSinglePredecessor()))
8845 continue;
8846
8847 if (CmpC.isPowerOf2() && Cmp->getPredicate() == ICmpInst::ICMP_ULT &&
8848 match(V: UI, P: m_Shr(L: m_Specific(V: X), R: m_SpecificInt(V: CmpC.logBase2())))) {
8849 IRBuilder<> Builder(Branch);
8850 if (UI->getParent() != Branch->getParent())
8851 UI->moveBefore(InsertPos: Branch->getIterator());
8852 UI->dropPoisonGeneratingFlags();
8853 Value *NewCmp = Builder.CreateCmp(Pred: ICmpInst::ICMP_EQ, LHS: UI,
8854 RHS: ConstantInt::get(Ty: UI->getType(), V: 0));
8855 LLVM_DEBUG(dbgs() << "Converting " << *Cmp << "\n");
8856 LLVM_DEBUG(dbgs() << " to compare on zero: " << *NewCmp << "\n");
8857 replaceAllUsesWith(Old: Cmp, New: NewCmp, FreshBBs, IsHuge: IsHugeFunc);
8858 return true;
8859 }
8860 if (Cmp->isEquality() &&
8861 (match(V: UI, P: m_Add(L: m_Specific(V: X), R: m_SpecificInt(V: -CmpC))) ||
8862 match(V: UI, P: m_Sub(L: m_Specific(V: X), R: m_SpecificInt(V: CmpC))) ||
8863 match(V: UI, P: m_Xor(L: m_Specific(V: X), R: m_SpecificInt(V: CmpC))))) {
8864 IRBuilder<> Builder(Branch);
8865 if (UI->getParent() != Branch->getParent())
8866 UI->moveBefore(InsertPos: Branch->getIterator());
8867 UI->dropPoisonGeneratingFlags();
8868 Value *NewCmp = Builder.CreateCmp(Pred: Cmp->getPredicate(), LHS: UI,
8869 RHS: ConstantInt::get(Ty: UI->getType(), V: 0));
8870 LLVM_DEBUG(dbgs() << "Converting " << *Cmp << "\n");
8871 LLVM_DEBUG(dbgs() << " to compare on zero: " << *NewCmp << "\n");
8872 replaceAllUsesWith(Old: Cmp, New: NewCmp, FreshBBs, IsHuge: IsHugeFunc);
8873 return true;
8874 }
8875 }
8876 return false;
8877}
8878
8879bool CodeGenPrepare::optimizeInst(Instruction *I, ModifyDT &ModifiedDT) {
8880 bool AnyChange = false;
8881 AnyChange = fixupDbgVariableRecordsOnInst(I&: *I);
8882
8883 // Bail out if we inserted the instruction to prevent optimizations from
8884 // stepping on each other's toes.
8885 if (InsertedInsts.count(Ptr: I))
8886 return AnyChange;
8887
8888 // TODO: Move into the switch on opcode below here.
8889 if (PHINode *P = dyn_cast<PHINode>(Val: I)) {
8890 // It is possible for very late stage optimizations (such as SimplifyCFG)
8891 // to introduce PHI nodes too late to be cleaned up. If we detect such a
8892 // trivial PHI, go ahead and zap it here.
8893 if (Value *V = simplifyInstruction(I: P, Q: {*DL, TLInfo})) {
8894 LargeOffsetGEPMap.erase(Key: P);
8895 replaceAllUsesWith(Old: P, New: V, FreshBBs, IsHuge: IsHugeFunc);
8896 P->eraseFromParent();
8897 ++NumPHIsElim;
8898 return true;
8899 }
8900 return AnyChange;
8901 }
8902
8903 if (CastInst *CI = dyn_cast<CastInst>(Val: I)) {
8904 // If the source of the cast is a constant, then this should have
8905 // already been constant folded. The only reason NOT to constant fold
8906 // it is if something (e.g. LSR) was careful to place the constant
8907 // evaluation in a block other than then one that uses it (e.g. to hoist
8908 // the address of globals out of a loop). If this is the case, we don't
8909 // want to forward-subst the cast.
8910 if (isa<Constant>(Val: CI->getOperand(i_nocapture: 0)))
8911 return AnyChange;
8912
8913 if (OptimizeNoopCopyExpression(CI, TLI: *TLI, DL: *DL))
8914 return true;
8915
8916 if ((isa<UIToFPInst>(Val: I) || isa<SIToFPInst>(Val: I) || isa<FPToUIInst>(Val: I) ||
8917 isa<TruncInst>(Val: I)) &&
8918 TLI->optimizeExtendOrTruncateConversion(
8919 I, L: LI->getLoopFor(BB: I->getParent()), TTI: *TTI))
8920 return true;
8921
8922 if (isa<ZExtInst>(Val: I) || isa<SExtInst>(Val: I)) {
8923 /// Sink a zext or sext into its user blocks if the target type doesn't
8924 /// fit in one register
8925 if (TLI->getTypeAction(Context&: CI->getContext(),
8926 VT: TLI->getValueType(DL: *DL, Ty: CI->getType())) ==
8927 TargetLowering::TypeExpandInteger) {
8928 return SinkCast(CI);
8929 } else {
8930 if (TLI->optimizeExtendOrTruncateConversion(
8931 I, L: LI->getLoopFor(BB: I->getParent()), TTI: *TTI))
8932 return true;
8933
8934 bool MadeChange = optimizeExt(Inst&: I);
8935 return MadeChange | optimizeExtUses(I);
8936 }
8937 }
8938 return AnyChange;
8939 }
8940
8941 if (auto *Cmp = dyn_cast<CmpInst>(Val: I))
8942 if (optimizeCmp(Cmp, ModifiedDT))
8943 return true;
8944
8945 if (match(V: I, P: m_URem(L: m_Value(), R: m_Value())))
8946 if (optimizeURem(Rem: I))
8947 return true;
8948
8949 if (LoadInst *LI = dyn_cast<LoadInst>(Val: I)) {
8950 LI->setMetadata(KindID: LLVMContext::MD_invariant_group, Node: nullptr);
8951 bool Modified = optimizeLoadExt(Load: LI);
8952 unsigned AS = LI->getPointerAddressSpace();
8953 Modified |= optimizeMemoryInst(MemoryInst: I, Addr: I->getOperand(i: 0), AccessTy: LI->getType(), AddrSpace: AS);
8954 return Modified;
8955 }
8956
8957 if (StoreInst *SI = dyn_cast<StoreInst>(Val: I)) {
8958 if (splitMergedValStore(SI&: *SI, DL: *DL, TLI: *TLI))
8959 return true;
8960 SI->setMetadata(KindID: LLVMContext::MD_invariant_group, Node: nullptr);
8961 unsigned AS = SI->getPointerAddressSpace();
8962 return optimizeMemoryInst(MemoryInst: I, Addr: SI->getOperand(i_nocapture: 1),
8963 AccessTy: SI->getOperand(i_nocapture: 0)->getType(), AddrSpace: AS);
8964 }
8965
8966 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: I)) {
8967 unsigned AS = RMW->getPointerAddressSpace();
8968 return optimizeMemoryInst(MemoryInst: I, Addr: RMW->getPointerOperand(), AccessTy: RMW->getType(), AddrSpace: AS);
8969 }
8970
8971 if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Val: I)) {
8972 unsigned AS = CmpX->getPointerAddressSpace();
8973 return optimizeMemoryInst(MemoryInst: I, Addr: CmpX->getPointerOperand(),
8974 AccessTy: CmpX->getCompareOperand()->getType(), AddrSpace: AS);
8975 }
8976
8977 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Val: I);
8978
8979 if (BinOp && BinOp->getOpcode() == Instruction::And && EnableAndCmpSinking &&
8980 sinkAndCmp0Expression(AndI: BinOp, TLI: *TLI, InsertedInsts))
8981 return true;
8982
8983 // TODO: Move this into the switch on opcode - it handles shifts already.
8984 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
8985 BinOp->getOpcode() == Instruction::LShr)) {
8986 ConstantInt *CI = dyn_cast<ConstantInt>(Val: BinOp->getOperand(i_nocapture: 1));
8987 if (CI && TLI->hasExtractBitsInsn())
8988 if (OptimizeExtractBits(ShiftI: BinOp, CI, TLI: *TLI, DL: *DL))
8989 return true;
8990 }
8991
8992 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Val: I)) {
8993 if (GEPI->hasAllZeroIndices()) {
8994 /// The GEP operand must be a pointer, so must its result -> BitCast
8995 Instruction *NC = new BitCastInst(GEPI->getOperand(i_nocapture: 0), GEPI->getType(),
8996 GEPI->getName(), GEPI->getIterator());
8997 NC->setDebugLoc(GEPI->getDebugLoc());
8998 replaceAllUsesWith(Old: GEPI, New: NC, FreshBBs, IsHuge: IsHugeFunc);
8999 RecursivelyDeleteTriviallyDeadInstructions(
9000 V: GEPI, TLI: TLInfo, MSSAU: nullptr,
9001 AboutToDeleteCallback: [&](Value *V) { removeAllAssertingVHReferences(V); });
9002 ++NumGEPsElim;
9003 optimizeInst(I: NC, ModifiedDT);
9004 return true;
9005 }
9006 if (tryUnmergingGEPsAcrossIndirectBr(GEPI, TTI)) {
9007 return true;
9008 }
9009 }
9010
9011 if (FreezeInst *FI = dyn_cast<FreezeInst>(Val: I)) {
9012 // freeze(icmp a, const)) -> icmp (freeze a), const
9013 // This helps generate efficient conditional jumps.
9014 Instruction *CmpI = nullptr;
9015 if (ICmpInst *II = dyn_cast<ICmpInst>(Val: FI->getOperand(i_nocapture: 0)))
9016 CmpI = II;
9017 else if (FCmpInst *F = dyn_cast<FCmpInst>(Val: FI->getOperand(i_nocapture: 0)))
9018 CmpI = F->getFastMathFlags().none() ? F : nullptr;
9019
9020 if (CmpI && CmpI->hasOneUse()) {
9021 auto Op0 = CmpI->getOperand(i: 0), Op1 = CmpI->getOperand(i: 1);
9022 bool Const0 = isa<ConstantInt>(Val: Op0) || isa<ConstantFP>(Val: Op0) ||
9023 isa<ConstantPointerNull>(Val: Op0);
9024 bool Const1 = isa<ConstantInt>(Val: Op1) || isa<ConstantFP>(Val: Op1) ||
9025 isa<ConstantPointerNull>(Val: Op1);
9026 if (Const0 || Const1) {
9027 if (!Const0 || !Const1) {
9028 auto *F = new FreezeInst(Const0 ? Op1 : Op0, "", CmpI->getIterator());
9029 F->takeName(V: FI);
9030 CmpI->setOperand(i: Const0 ? 1 : 0, Val: F);
9031 }
9032 replaceAllUsesWith(Old: FI, New: CmpI, FreshBBs, IsHuge: IsHugeFunc);
9033 FI->eraseFromParent();
9034 return true;
9035 }
9036 }
9037 return AnyChange;
9038 }
9039
9040 if (tryToSinkFreeOperands(I))
9041 return true;
9042
9043 switch (I->getOpcode()) {
9044 case Instruction::Shl:
9045 case Instruction::LShr:
9046 case Instruction::AShr:
9047 return optimizeShiftInst(Shift: cast<BinaryOperator>(Val: I));
9048 case Instruction::Call:
9049 return optimizeCallInst(CI: cast<CallInst>(Val: I), ModifiedDT);
9050 case Instruction::Select:
9051 return optimizeSelectInst(SI: cast<SelectInst>(Val: I));
9052 case Instruction::ShuffleVector:
9053 return optimizeShuffleVectorInst(SVI: cast<ShuffleVectorInst>(Val: I));
9054 case Instruction::Switch:
9055 return optimizeSwitchInst(SI: cast<SwitchInst>(Val: I));
9056 case Instruction::ExtractElement:
9057 return optimizeExtractElementInst(Inst: cast<ExtractElementInst>(Val: I));
9058 case Instruction::Br:
9059 return optimizeBranch(Branch: cast<BranchInst>(Val: I), TLI: *TLI, FreshBBs, IsHugeFunc);
9060 }
9061
9062 return AnyChange;
9063}
9064
9065/// Given an OR instruction, check to see if this is a bitreverse
9066/// idiom. If so, insert the new intrinsic and return true.
9067bool CodeGenPrepare::makeBitReverse(Instruction &I) {
9068 if (!I.getType()->isIntegerTy() ||
9069 !TLI->isOperationLegalOrCustom(Op: ISD::BITREVERSE,
9070 VT: TLI->getValueType(DL: *DL, Ty: I.getType(), AllowUnknown: true)))
9071 return false;
9072
9073 SmallVector<Instruction *, 4> Insts;
9074 if (!recognizeBSwapOrBitReverseIdiom(I: &I, MatchBSwaps: false, MatchBitReversals: true, InsertedInsts&: Insts))
9075 return false;
9076 Instruction *LastInst = Insts.back();
9077 replaceAllUsesWith(Old: &I, New: LastInst, FreshBBs, IsHuge: IsHugeFunc);
9078 RecursivelyDeleteTriviallyDeadInstructions(
9079 V: &I, TLI: TLInfo, MSSAU: nullptr,
9080 AboutToDeleteCallback: [&](Value *V) { removeAllAssertingVHReferences(V); });
9081 return true;
9082}
9083
9084// In this pass we look for GEP and cast instructions that are used
9085// across basic blocks and rewrite them to improve basic-block-at-a-time
9086// selection.
9087bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, ModifyDT &ModifiedDT) {
9088 SunkAddrs.clear();
9089 bool MadeChange = false;
9090
9091 do {
9092 CurInstIterator = BB.begin();
9093 ModifiedDT = ModifyDT::NotModifyDT;
9094 while (CurInstIterator != BB.end()) {
9095 MadeChange |= optimizeInst(I: &*CurInstIterator++, ModifiedDT);
9096 if (ModifiedDT != ModifyDT::NotModifyDT) {
9097 // For huge function we tend to quickly go though the inner optmization
9098 // opportunities in the BB. So we go back to the BB head to re-optimize
9099 // each instruction instead of go back to the function head.
9100 if (IsHugeFunc) {
9101 DT.reset();
9102 getDT(F&: *BB.getParent());
9103 break;
9104 } else {
9105 return true;
9106 }
9107 }
9108 }
9109 } while (ModifiedDT == ModifyDT::ModifyInstDT);
9110
9111 bool MadeBitReverse = true;
9112 while (MadeBitReverse) {
9113 MadeBitReverse = false;
9114 for (auto &I : reverse(C&: BB)) {
9115 if (makeBitReverse(I)) {
9116 MadeBitReverse = MadeChange = true;
9117 break;
9118 }
9119 }
9120 }
9121 MadeChange |= dupRetToEnableTailCallOpts(BB: &BB, ModifiedDT);
9122
9123 return MadeChange;
9124}
9125
9126bool CodeGenPrepare::fixupDbgVariableRecordsOnInst(Instruction &I) {
9127 bool AnyChange = false;
9128 for (DbgVariableRecord &DVR : filterDbgVars(R: I.getDbgRecordRange()))
9129 AnyChange |= fixupDbgVariableRecord(I&: DVR);
9130 return AnyChange;
9131}
9132
9133// FIXME: should updating debug-info really cause the "changed" flag to fire,
9134// which can cause a function to be reprocessed?
9135bool CodeGenPrepare::fixupDbgVariableRecord(DbgVariableRecord &DVR) {
9136 if (DVR.Type != DbgVariableRecord::LocationType::Value &&
9137 DVR.Type != DbgVariableRecord::LocationType::Assign)
9138 return false;
9139
9140 // Does this DbgVariableRecord refer to a sunk address calculation?
9141 bool AnyChange = false;
9142 SmallDenseSet<Value *> LocationOps(DVR.location_ops().begin(),
9143 DVR.location_ops().end());
9144 for (Value *Location : LocationOps) {
9145 WeakTrackingVH SunkAddrVH = SunkAddrs[Location];
9146 Value *SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
9147 if (SunkAddr) {
9148 // Point dbg.value at locally computed address, which should give the best
9149 // opportunity to be accurately lowered. This update may change the type
9150 // of pointer being referred to; however this makes no difference to
9151 // debugging information, and we can't generate bitcasts that may affect
9152 // codegen.
9153 DVR.replaceVariableLocationOp(OldValue: Location, NewValue: SunkAddr);
9154 AnyChange = true;
9155 }
9156 }
9157 return AnyChange;
9158}
9159
9160static void DbgInserterHelper(DbgVariableRecord *DVR, BasicBlock::iterator VI) {
9161 DVR->removeFromParent();
9162 BasicBlock *VIBB = VI->getParent();
9163 if (isa<PHINode>(Val: VI))
9164 VIBB->insertDbgRecordBefore(DR: DVR, Here: VIBB->getFirstInsertionPt());
9165 else
9166 VIBB->insertDbgRecordAfter(DR: DVR, I: &*VI);
9167}
9168
9169// A llvm.dbg.value may be using a value before its definition, due to
9170// optimizations in this pass and others. Scan for such dbg.values, and rescue
9171// them by moving the dbg.value to immediately after the value definition.
9172// FIXME: Ideally this should never be necessary, and this has the potential
9173// to re-order dbg.value intrinsics.
9174bool CodeGenPrepare::placeDbgValues(Function &F) {
9175 bool MadeChange = false;
9176 DominatorTree DT(F);
9177
9178 auto DbgProcessor = [&](auto *DbgItem, Instruction *Position) {
9179 SmallVector<Instruction *, 4> VIs;
9180 for (Value *V : DbgItem->location_ops())
9181 if (Instruction *VI = dyn_cast_or_null<Instruction>(Val: V))
9182 VIs.push_back(Elt: VI);
9183
9184 // This item may depend on multiple instructions, complicating any
9185 // potential sink. This block takes the defensive approach, opting to
9186 // "undef" the item if it has more than one instruction and any of them do
9187 // not dominate iem.
9188 for (Instruction *VI : VIs) {
9189 if (VI->isTerminator())
9190 continue;
9191
9192 // If VI is a phi in a block with an EHPad terminator, we can't insert
9193 // after it.
9194 if (isa<PHINode>(Val: VI) && VI->getParent()->getTerminator()->isEHPad())
9195 continue;
9196
9197 // If the defining instruction dominates the dbg.value, we do not need
9198 // to move the dbg.value.
9199 if (DT.dominates(Def: VI, User: Position))
9200 continue;
9201
9202 // If we depend on multiple instructions and any of them doesn't
9203 // dominate this DVI, we probably can't salvage it: moving it to
9204 // after any of the instructions could cause us to lose the others.
9205 if (VIs.size() > 1) {
9206 LLVM_DEBUG(
9207 dbgs()
9208 << "Unable to find valid location for Debug Value, undefing:\n"
9209 << *DbgItem);
9210 DbgItem->setKillLocation();
9211 break;
9212 }
9213
9214 LLVM_DEBUG(dbgs() << "Moving Debug Value before :\n"
9215 << *DbgItem << ' ' << *VI);
9216 DbgInserterHelper(DbgItem, VI->getIterator());
9217 MadeChange = true;
9218 ++NumDbgValueMoved;
9219 }
9220 };
9221
9222 for (BasicBlock &BB : F) {
9223 for (Instruction &Insn : llvm::make_early_inc_range(Range&: BB)) {
9224 // Process any DbgVariableRecord records attached to this
9225 // instruction.
9226 for (DbgVariableRecord &DVR : llvm::make_early_inc_range(
9227 Range: filterDbgVars(R: Insn.getDbgRecordRange()))) {
9228 if (DVR.Type != DbgVariableRecord::LocationType::Value)
9229 continue;
9230 DbgProcessor(&DVR, &Insn);
9231 }
9232 }
9233 }
9234
9235 return MadeChange;
9236}
9237
9238// Group scattered pseudo probes in a block to favor SelectionDAG. Scattered
9239// probes can be chained dependencies of other regular DAG nodes and block DAG
9240// combine optimizations.
9241bool CodeGenPrepare::placePseudoProbes(Function &F) {
9242 bool MadeChange = false;
9243 for (auto &Block : F) {
9244 // Move the rest probes to the beginning of the block.
9245 auto FirstInst = Block.getFirstInsertionPt();
9246 while (FirstInst != Block.end() && FirstInst->isDebugOrPseudoInst())
9247 ++FirstInst;
9248 BasicBlock::iterator I(FirstInst);
9249 I++;
9250 while (I != Block.end()) {
9251 if (auto *II = dyn_cast<PseudoProbeInst>(Val: I++)) {
9252 II->moveBefore(InsertPos: FirstInst);
9253 MadeChange = true;
9254 }
9255 }
9256 }
9257 return MadeChange;
9258}
9259
9260/// Scale down both weights to fit into uint32_t.
9261static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
9262 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
9263 uint32_t Scale = (NewMax / std::numeric_limits<uint32_t>::max()) + 1;
9264 NewTrue = NewTrue / Scale;
9265 NewFalse = NewFalse / Scale;
9266}
9267
9268/// Some targets prefer to split a conditional branch like:
9269/// \code
9270/// %0 = icmp ne i32 %a, 0
9271/// %1 = icmp ne i32 %b, 0
9272/// %or.cond = or i1 %0, %1
9273/// br i1 %or.cond, label %TrueBB, label %FalseBB
9274/// \endcode
9275/// into multiple branch instructions like:
9276/// \code
9277/// bb1:
9278/// %0 = icmp ne i32 %a, 0
9279/// br i1 %0, label %TrueBB, label %bb2
9280/// bb2:
9281/// %1 = icmp ne i32 %b, 0
9282/// br i1 %1, label %TrueBB, label %FalseBB
9283/// \endcode
9284/// This usually allows instruction selection to do even further optimizations
9285/// and combine the compare with the branch instruction. Currently this is
9286/// applied for targets which have "cheap" jump instructions.
9287///
9288/// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
9289///
9290bool CodeGenPrepare::splitBranchCondition(Function &F, ModifyDT &ModifiedDT) {
9291 if (!TM->Options.EnableFastISel || TLI->isJumpExpensive())
9292 return false;
9293
9294 bool MadeChange = false;
9295 for (auto &BB : F) {
9296 // Does this BB end with the following?
9297 // %cond1 = icmp|fcmp|binary instruction ...
9298 // %cond2 = icmp|fcmp|binary instruction ...
9299 // %cond.or = or|and i1 %cond1, cond2
9300 // br i1 %cond.or label %dest1, label %dest2"
9301 Instruction *LogicOp;
9302 BasicBlock *TBB, *FBB;
9303 if (!match(V: BB.getTerminator(),
9304 P: m_Br(C: m_OneUse(SubPattern: m_Instruction(I&: LogicOp)), T&: TBB, F&: FBB)))
9305 continue;
9306
9307 auto *Br1 = cast<BranchInst>(Val: BB.getTerminator());
9308 if (Br1->getMetadata(KindID: LLVMContext::MD_unpredictable))
9309 continue;
9310
9311 // The merging of mostly empty BB can cause a degenerate branch.
9312 if (TBB == FBB)
9313 continue;
9314
9315 unsigned Opc;
9316 Value *Cond1, *Cond2;
9317 if (match(V: LogicOp,
9318 P: m_LogicalAnd(L: m_OneUse(SubPattern: m_Value(V&: Cond1)), R: m_OneUse(SubPattern: m_Value(V&: Cond2)))))
9319 Opc = Instruction::And;
9320 else if (match(V: LogicOp, P: m_LogicalOr(L: m_OneUse(SubPattern: m_Value(V&: Cond1)),
9321 R: m_OneUse(SubPattern: m_Value(V&: Cond2)))))
9322 Opc = Instruction::Or;
9323 else
9324 continue;
9325
9326 auto IsGoodCond = [](Value *Cond) {
9327 return match(
9328 V: Cond,
9329 P: m_CombineOr(L: m_Cmp(), R: m_CombineOr(L: m_LogicalAnd(L: m_Value(), R: m_Value()),
9330 R: m_LogicalOr(L: m_Value(), R: m_Value()))));
9331 };
9332 if (!IsGoodCond(Cond1) || !IsGoodCond(Cond2))
9333 continue;
9334
9335 LLVM_DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
9336
9337 // Create a new BB.
9338 auto *TmpBB =
9339 BasicBlock::Create(Context&: BB.getContext(), Name: BB.getName() + ".cond.split",
9340 Parent: BB.getParent(), InsertBefore: BB.getNextNode());
9341 if (IsHugeFunc)
9342 FreshBBs.insert(Ptr: TmpBB);
9343
9344 // Update original basic block by using the first condition directly by the
9345 // branch instruction and removing the no longer needed and/or instruction.
9346 Br1->setCondition(Cond1);
9347 LogicOp->eraseFromParent();
9348
9349 // Depending on the condition we have to either replace the true or the
9350 // false successor of the original branch instruction.
9351 if (Opc == Instruction::And)
9352 Br1->setSuccessor(idx: 0, NewSucc: TmpBB);
9353 else
9354 Br1->setSuccessor(idx: 1, NewSucc: TmpBB);
9355
9356 // Fill in the new basic block.
9357 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond: Cond2, True: TBB, False: FBB);
9358 if (auto *I = dyn_cast<Instruction>(Val: Cond2)) {
9359 I->removeFromParent();
9360 I->insertBefore(InsertPos: Br2->getIterator());
9361 }
9362
9363 // Update PHI nodes in both successors. The original BB needs to be
9364 // replaced in one successor's PHI nodes, because the branch comes now from
9365 // the newly generated BB (NewBB). In the other successor we need to add one
9366 // incoming edge to the PHI nodes, because both branch instructions target
9367 // now the same successor. Depending on the original branch condition
9368 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
9369 // we perform the correct update for the PHI nodes.
9370 // This doesn't change the successor order of the just created branch
9371 // instruction (or any other instruction).
9372 if (Opc == Instruction::Or)
9373 std::swap(a&: TBB, b&: FBB);
9374
9375 // Replace the old BB with the new BB.
9376 TBB->replacePhiUsesWith(Old: &BB, New: TmpBB);
9377
9378 // Add another incoming edge from the new BB.
9379 for (PHINode &PN : FBB->phis()) {
9380 auto *Val = PN.getIncomingValueForBlock(BB: &BB);
9381 PN.addIncoming(V: Val, BB: TmpBB);
9382 }
9383
9384 // Update the branch weights (from SelectionDAGBuilder::
9385 // FindMergedConditions).
9386 if (Opc == Instruction::Or) {
9387 // Codegen X | Y as:
9388 // BB1:
9389 // jmp_if_X TBB
9390 // jmp TmpBB
9391 // TmpBB:
9392 // jmp_if_Y TBB
9393 // jmp FBB
9394 //
9395
9396 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
9397 // The requirement is that
9398 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
9399 // = TrueProb for original BB.
9400 // Assuming the original weights are A and B, one choice is to set BB1's
9401 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
9402 // assumes that
9403 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
9404 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
9405 // TmpBB, but the math is more complicated.
9406 uint64_t TrueWeight, FalseWeight;
9407 if (extractBranchWeights(I: *Br1, TrueVal&: TrueWeight, FalseVal&: FalseWeight)) {
9408 uint64_t NewTrueWeight = TrueWeight;
9409 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
9410 scaleWeights(NewTrue&: NewTrueWeight, NewFalse&: NewFalseWeight);
9411 Br1->setMetadata(KindID: LLVMContext::MD_prof,
9412 Node: MDBuilder(Br1->getContext())
9413 .createBranchWeights(TrueWeight, FalseWeight,
9414 IsExpected: hasBranchWeightOrigin(I: *Br1)));
9415
9416 NewTrueWeight = TrueWeight;
9417 NewFalseWeight = 2 * FalseWeight;
9418 scaleWeights(NewTrue&: NewTrueWeight, NewFalse&: NewFalseWeight);
9419 Br2->setMetadata(KindID: LLVMContext::MD_prof,
9420 Node: MDBuilder(Br2->getContext())
9421 .createBranchWeights(TrueWeight, FalseWeight));
9422 }
9423 } else {
9424 // Codegen X & Y as:
9425 // BB1:
9426 // jmp_if_X TmpBB
9427 // jmp FBB
9428 // TmpBB:
9429 // jmp_if_Y TBB
9430 // jmp FBB
9431 //
9432 // This requires creation of TmpBB after CurBB.
9433
9434 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
9435 // The requirement is that
9436 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
9437 // = FalseProb for original BB.
9438 // Assuming the original weights are A and B, one choice is to set BB1's
9439 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
9440 // assumes that
9441 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
9442 uint64_t TrueWeight, FalseWeight;
9443 if (extractBranchWeights(I: *Br1, TrueVal&: TrueWeight, FalseVal&: FalseWeight)) {
9444 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
9445 uint64_t NewFalseWeight = FalseWeight;
9446 scaleWeights(NewTrue&: NewTrueWeight, NewFalse&: NewFalseWeight);
9447 Br1->setMetadata(KindID: LLVMContext::MD_prof,
9448 Node: MDBuilder(Br1->getContext())
9449 .createBranchWeights(TrueWeight, FalseWeight));
9450
9451 NewTrueWeight = 2 * TrueWeight;
9452 NewFalseWeight = FalseWeight;
9453 scaleWeights(NewTrue&: NewTrueWeight, NewFalse&: NewFalseWeight);
9454 Br2->setMetadata(KindID: LLVMContext::MD_prof,
9455 Node: MDBuilder(Br2->getContext())
9456 .createBranchWeights(TrueWeight, FalseWeight));
9457 }
9458 }
9459
9460 ModifiedDT = ModifyDT::ModifyBBDT;
9461 MadeChange = true;
9462
9463 LLVM_DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
9464 TmpBB->dump());
9465 }
9466 return MadeChange;
9467}
9468