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/DomTreeUpdater.h"
28#include "llvm/Analysis/FloatingPointPredicateUtils.h"
29#include "llvm/Analysis/InstructionSimplify.h"
30#include "llvm/Analysis/LoopInfo.h"
31#include "llvm/Analysis/ProfileSummaryInfo.h"
32#include "llvm/Analysis/ScalarEvolutionExpressions.h"
33#include "llvm/Analysis/TargetLibraryInfo.h"
34#include "llvm/Analysis/TargetTransformInfo.h"
35#include "llvm/Analysis/ValueTracking.h"
36#include "llvm/Analysis/VectorUtils.h"
37#include "llvm/CodeGen/Analysis.h"
38#include "llvm/CodeGen/BasicBlockSectionsProfileReader.h"
39#include "llvm/CodeGen/ISDOpcodes.h"
40#include "llvm/CodeGen/SelectionDAGNodes.h"
41#include "llvm/CodeGen/TargetLowering.h"
42#include "llvm/CodeGen/TargetPassConfig.h"
43#include "llvm/CodeGen/TargetSubtargetInfo.h"
44#include "llvm/CodeGen/ValueTypes.h"
45#include "llvm/CodeGenTypes/MachineValueType.h"
46#include "llvm/Config/llvm-config.h"
47#include "llvm/IR/Argument.h"
48#include "llvm/IR/Attributes.h"
49#include "llvm/IR/BasicBlock.h"
50#include "llvm/IR/CFG.h"
51#include "llvm/IR/Constant.h"
52#include "llvm/IR/Constants.h"
53#include "llvm/IR/DataLayout.h"
54#include "llvm/IR/DebugInfo.h"
55#include "llvm/IR/DerivedTypes.h"
56#include "llvm/IR/Dominators.h"
57#include "llvm/IR/Function.h"
58#include "llvm/IR/GetElementPtrTypeIterator.h"
59#include "llvm/IR/GlobalValue.h"
60#include "llvm/IR/GlobalVariable.h"
61#include "llvm/IR/IRBuilder.h"
62#include "llvm/IR/InlineAsm.h"
63#include "llvm/IR/InstrTypes.h"
64#include "llvm/IR/Instruction.h"
65#include "llvm/IR/Instructions.h"
66#include "llvm/IR/IntrinsicInst.h"
67#include "llvm/IR/Intrinsics.h"
68#include "llvm/IR/IntrinsicsAArch64.h"
69#include "llvm/IR/LLVMContext.h"
70#include "llvm/IR/MDBuilder.h"
71#include "llvm/IR/Module.h"
72#include "llvm/IR/Operator.h"
73#include "llvm/IR/PatternMatch.h"
74#include "llvm/IR/ProfDataUtils.h"
75#include "llvm/IR/Statepoint.h"
76#include "llvm/IR/Type.h"
77#include "llvm/IR/Use.h"
78#include "llvm/IR/User.h"
79#include "llvm/IR/Value.h"
80#include "llvm/IR/ValueHandle.h"
81#include "llvm/IR/ValueMap.h"
82#include "llvm/InitializePasses.h"
83#include "llvm/Pass.h"
84#include "llvm/Support/BlockFrequency.h"
85#include "llvm/Support/BranchProbability.h"
86#include "llvm/Support/Casting.h"
87#include "llvm/Support/CommandLine.h"
88#include "llvm/Support/Compiler.h"
89#include "llvm/Support/Debug.h"
90#include "llvm/Support/ErrorHandling.h"
91#include "llvm/Support/raw_ostream.h"
92#include "llvm/Target/TargetMachine.h"
93#include "llvm/Target/TargetOptions.h"
94#include "llvm/Transforms/Utils/BasicBlockUtils.h"
95#include "llvm/Transforms/Utils/BypassSlowDivision.h"
96#include "llvm/Transforms/Utils/Local.h"
97#include "llvm/Transforms/Utils/SimplifyLibCalls.h"
98#include "llvm/Transforms/Utils/SizeOpts.h"
99#include <algorithm>
100#include <cassert>
101#include <cstdint>
102#include <iterator>
103#include <limits>
104#include <memory>
105#include <optional>
106#include <utility>
107#include <vector>
108
109using namespace llvm;
110using namespace llvm::PatternMatch;
111
112#define DEBUG_TYPE "codegenprepare"
113
114STATISTIC(NumBlocksElim, "Number of blocks eliminated");
115STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
116STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
117STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
118 "sunken Cmps");
119STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
120 "of sunken Casts");
121STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
122 "computations were sunk");
123STATISTIC(NumMemoryInstsPhiCreated,
124 "Number of phis created when address "
125 "computations were sunk to memory instructions");
126STATISTIC(NumMemoryInstsSelectCreated,
127 "Number of select created when address "
128 "computations were sunk to memory instructions");
129STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
130STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
131STATISTIC(NumAndsAdded,
132 "Number of and mask instructions added to form ext loads");
133STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
134STATISTIC(NumRetsDup, "Number of return instructions duplicated");
135STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
136STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
137STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
138
139static cl::opt<bool> DisableBranchOpts(
140 "disable-cgp-branch-opts", cl::Hidden, cl::init(Val: false),
141 cl::desc("Disable branch optimizations in CodeGenPrepare"));
142
143static cl::opt<bool>
144 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(Val: false),
145 cl::desc("Disable GC optimizations in CodeGenPrepare"));
146
147static cl::opt<bool>
148 DisableSelectToBranch("disable-cgp-select2branch", cl::Hidden,
149 cl::init(Val: false),
150 cl::desc("Disable select to branch conversion."));
151
152static cl::opt<bool>
153 AddrSinkUsingGEPs("addr-sink-using-gep", cl::Hidden, cl::init(Val: true),
154 cl::desc("Address sinking in CGP using GEPs."));
155
156static cl::opt<bool>
157 EnableAndCmpSinking("enable-andcmp-sinking", cl::Hidden, cl::init(Val: true),
158 cl::desc("Enable sinking and/cmp into branches."));
159
160static cl::opt<bool> DisableStoreExtract(
161 "disable-cgp-store-extract", cl::Hidden, cl::init(Val: false),
162 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
163
164static cl::opt<bool> StressStoreExtract(
165 "stress-cgp-store-extract", cl::Hidden, cl::init(Val: false),
166 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
167
168static cl::opt<bool> DisableExtLdPromotion(
169 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(Val: false),
170 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
171 "CodeGenPrepare"));
172
173static cl::opt<bool> StressExtLdPromotion(
174 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(Val: false),
175 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
176 "optimization in CodeGenPrepare"));
177
178static cl::opt<bool> DisablePreheaderProtect(
179 "disable-preheader-prot", cl::Hidden, cl::init(Val: false),
180 cl::desc("Disable protection against removing loop preheaders"));
181
182static cl::opt<bool> ProfileGuidedSectionPrefix(
183 "profile-guided-section-prefix", cl::Hidden, cl::init(Val: true),
184 cl::desc("Use profile info to add section prefix for hot/cold functions"));
185
186static cl::opt<bool> ProfileUnknownInSpecialSection(
187 "profile-unknown-in-special-section", cl::Hidden,
188 cl::desc("In profiling mode like sampleFDO, if a function doesn't have "
189 "profile, we cannot tell the function is cold for sure because "
190 "it may be a function newly added without ever being sampled. "
191 "With the flag enabled, compiler can put such profile unknown "
192 "functions into a special section, so runtime system can choose "
193 "to handle it in a different way than .text section, to save "
194 "RAM for example. "));
195
196static cl::opt<bool> BBSectionsGuidedSectionPrefix(
197 "bbsections-guided-section-prefix", cl::Hidden, cl::init(Val: true),
198 cl::desc("Use the basic-block-sections profile to determine the text "
199 "section prefix for hot functions. Functions with "
200 "basic-block-sections profile will be placed in `.text.hot` "
201 "regardless of their FDO profile info. Other functions won't be "
202 "impacted, i.e., their prefixes will be decided by FDO/sampleFDO "
203 "profiles."));
204
205static cl::opt<uint64_t> FreqRatioToSkipMerge(
206 "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(Val: 2),
207 cl::desc("Skip merging empty blocks if (frequency of empty block) / "
208 "(frequency of destination block) is greater than this ratio"));
209
210static cl::opt<bool> ForceSplitStore(
211 "force-split-store", cl::Hidden, cl::init(Val: false),
212 cl::desc("Force store splitting no matter what the target query says."));
213
214static cl::opt<bool> EnableTypePromotionMerge(
215 "cgp-type-promotion-merge", cl::Hidden,
216 cl::desc("Enable merging of redundant sexts when one is dominating"
217 " the other."),
218 cl::init(Val: true));
219
220static cl::opt<bool> DisableComplexAddrModes(
221 "disable-complex-addr-modes", cl::Hidden, cl::init(Val: false),
222 cl::desc("Disables combining addressing modes with different parts "
223 "in optimizeMemoryInst."));
224
225static cl::opt<bool>
226 AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(Val: false),
227 cl::desc("Allow creation of Phis in Address sinking."));
228
229static cl::opt<bool> AddrSinkNewSelects(
230 "addr-sink-new-select", cl::Hidden, cl::init(Val: true),
231 cl::desc("Allow creation of selects in Address sinking."));
232
233static cl::opt<bool> AddrSinkCombineBaseReg(
234 "addr-sink-combine-base-reg", cl::Hidden, cl::init(Val: true),
235 cl::desc("Allow combining of BaseReg field in Address sinking."));
236
237static cl::opt<bool> AddrSinkCombineBaseGV(
238 "addr-sink-combine-base-gv", cl::Hidden, cl::init(Val: true),
239 cl::desc("Allow combining of BaseGV field in Address sinking."));
240
241static cl::opt<bool> AddrSinkCombineBaseOffs(
242 "addr-sink-combine-base-offs", cl::Hidden, cl::init(Val: true),
243 cl::desc("Allow combining of BaseOffs field in Address sinking."));
244
245static cl::opt<bool> AddrSinkCombineScaledReg(
246 "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(Val: true),
247 cl::desc("Allow combining of ScaledReg field in Address sinking."));
248
249static cl::opt<bool>
250 EnableGEPOffsetSplit("cgp-split-large-offset-gep", cl::Hidden,
251 cl::init(Val: true),
252 cl::desc("Enable splitting large offset of GEP."));
253
254static cl::opt<bool> EnableICMP_EQToICMP_ST(
255 "cgp-icmp-eq2icmp-st", cl::Hidden, cl::init(Val: false),
256 cl::desc("Enable ICMP_EQ to ICMP_S(L|G)T conversion."));
257
258static cl::opt<bool>
259 VerifyBFIUpdates("cgp-verify-bfi-updates", cl::Hidden, cl::init(Val: false),
260 cl::desc("Enable BFI update verification for "
261 "CodeGenPrepare."));
262
263static cl::opt<bool>
264 OptimizePhiTypes("cgp-optimize-phi-types", cl::Hidden, cl::init(Val: true),
265 cl::desc("Enable converting phi types in CodeGenPrepare"));
266
267static cl::opt<unsigned>
268 HugeFuncThresholdInCGPP("cgpp-huge-func", cl::init(Val: 10000), cl::Hidden,
269 cl::desc("Least BB number of huge function."));
270
271static cl::opt<unsigned>
272 MaxAddressUsersToScan("cgp-max-address-users-to-scan", cl::init(Val: 100),
273 cl::Hidden,
274 cl::desc("Max number of address users to look at"));
275
276static cl::opt<bool>
277 DisableDeletePHIs("disable-cgp-delete-phis", cl::Hidden, cl::init(Val: false),
278 cl::desc("Disable elimination of dead PHI nodes."));
279
280namespace {
281
282enum ExtType {
283 ZeroExtension, // Zero extension has been seen.
284 SignExtension, // Sign extension has been seen.
285 BothExtension // This extension type is used if we saw sext after
286 // ZeroExtension had been set, or if we saw zext after
287 // SignExtension had been set. It makes the type
288 // information of a promoted instruction invalid.
289};
290
291enum ModifyDT {
292 NotModifyDT, // Not Modify any DT.
293 ModifyBBDT, // Modify the Basic Block Dominator Tree.
294 ModifyInstDT // Modify the Instruction Dominator in a Basic Block,
295 // This usually means we move/delete/insert instruction
296 // in a Basic Block. So we should re-iterate instructions
297 // in such Basic Block.
298};
299
300using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
301using TypeIsSExt = PointerIntPair<Type *, 2, ExtType>;
302using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
303using SExts = SmallVector<Instruction *, 16>;
304using ValueToSExts = MapVector<Value *, SExts>;
305
306class TypePromotionTransaction;
307
308class CodeGenPrepare {
309 friend class CodeGenPrepareLegacyPass;
310 const TargetMachine *TM = nullptr;
311 const TargetSubtargetInfo *SubtargetInfo = nullptr;
312 const TargetLowering *TLI = nullptr;
313 const TargetRegisterInfo *TRI = nullptr;
314 const TargetTransformInfo *TTI = nullptr;
315 const BasicBlockSectionsProfileReader *BBSectionsProfileReader = nullptr;
316 const TargetLibraryInfo *TLInfo = nullptr;
317 DomTreeUpdater *DTU = nullptr;
318 LoopInfo *LI = nullptr;
319 BlockFrequencyInfo *BFI;
320 BranchProbabilityInfo *BPI;
321 ProfileSummaryInfo *PSI = nullptr;
322
323 /// As we scan instructions optimizing them, this is the next instruction
324 /// to optimize. Transforms that can invalidate this should update it.
325 BasicBlock::iterator CurInstIterator;
326
327 /// Keeps track of non-local addresses that have been sunk into a block.
328 /// This allows us to avoid inserting duplicate code for blocks with
329 /// multiple load/stores of the same address. The usage of WeakTrackingVH
330 /// enables SunkAddrs to be treated as a cache whose entries can be
331 /// invalidated if a sunken address computation has been erased.
332 ValueMap<Value *, WeakTrackingVH> SunkAddrs;
333
334 /// Keeps track of all instructions inserted for the current function.
335 SetOfInstrs InsertedInsts;
336
337 /// Keeps track of the type of the related instruction before their
338 /// promotion for the current function.
339 InstrToOrigTy PromotedInsts;
340
341 /// Keep track of instructions removed during promotion.
342 SetOfInstrs RemovedInsts;
343
344 /// Keep track of sext chains based on their initial value.
345 DenseMap<Value *, Instruction *> SeenChainsForSExt;
346
347 /// Keep track of GEPs accessing the same data structures such as structs or
348 /// arrays that are candidates to be split later because of their large
349 /// size.
350 MapVector<AssertingVH<Value>,
351 SmallVector<std::pair<AssertingVH<GetElementPtrInst>, int64_t>, 32>>
352 LargeOffsetGEPMap;
353
354 /// Keep track of new GEP base after splitting the GEPs having large offset.
355 SmallSet<AssertingVH<Value>, 2> NewGEPBases;
356
357 /// Map serial numbers to Large offset GEPs.
358 DenseMap<AssertingVH<GetElementPtrInst>, int> LargeOffsetGEPID;
359
360 /// Keep track of SExt promoted.
361 ValueToSExts ValToSExtendedUses;
362
363 /// True if the function has the OptSize attribute.
364 bool OptSize;
365
366 /// DataLayout for the Function being processed.
367 const DataLayout *DL = nullptr;
368
369public:
370 CodeGenPrepare() = default;
371 CodeGenPrepare(const TargetMachine *TM) : TM(TM){};
372 /// If encounter huge function, we need to limit the build time.
373 bool IsHugeFunc = false;
374
375 /// FreshBBs is like worklist, it collected the updated BBs which need
376 /// to be optimized again.
377 /// Note: Consider building time in this pass, when a BB updated, we need
378 /// to insert such BB into FreshBBs for huge function.
379 SmallPtrSet<BasicBlock *, 32> FreshBBs;
380
381 void releaseMemory() {
382 // Clear per function information.
383 InsertedInsts.clear();
384 PromotedInsts.clear();
385 FreshBBs.clear();
386 }
387
388 bool run(Function &F, FunctionAnalysisManager &AM);
389
390private:
391 template <typename F>
392 void resetIteratorIfInvalidatedWhileCalling(BasicBlock *BB, F f) {
393 // Substituting can cause recursive simplifications, which can invalidate
394 // our iterator. Use a WeakTrackingVH to hold onto it in case this
395 // happens.
396 Value *CurValue = &*CurInstIterator;
397 WeakTrackingVH IterHandle(CurValue);
398
399 f();
400
401 // If the iterator instruction was recursively deleted, start over at the
402 // start of the block.
403 if (IterHandle != CurValue) {
404 CurInstIterator = BB->begin();
405 SunkAddrs.clear();
406 }
407 }
408
409 // Get the DominatorTree, updating it if necessary.
410 DominatorTree &getDT() { return DTU->getDomTree(); }
411
412 void removeAllAssertingVHReferences(Value *V);
413 bool eliminateAssumptions(Function &F);
414 bool eliminateFallThrough(Function &F);
415 bool eliminateMostlyEmptyBlocks(Function &F, bool &ResetLI);
416 BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
417 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
418 bool eliminateMostlyEmptyBlock(BasicBlock *BB);
419 bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
420 bool isPreheader);
421 bool makeBitReverse(Instruction &I);
422 bool optimizeBlock(BasicBlock &BB, ModifyDT &ModifiedDT);
423 bool optimizeInst(Instruction *I, ModifyDT &ModifiedDT);
424 bool optimizeMemoryInst(Instruction *MemoryInst, Value *Addr, Type *AccessTy,
425 unsigned AddrSpace);
426 bool optimizeGatherScatterInst(Instruction *MemoryInst, Value *Ptr);
427 bool optimizeMulWithOverflow(Instruction *I, bool IsSigned,
428 ModifyDT &ModifiedDT);
429 bool optimizeInlineAsmInst(CallInst *CS);
430 bool optimizeCallInst(CallInst *CI, ModifyDT &ModifiedDT);
431 bool optimizeExt(Instruction *&I);
432 bool optimizeExtUses(Instruction *I);
433 bool optimizeLoadExt(LoadInst *Load);
434 bool optimizeShiftInst(BinaryOperator *BO);
435 bool optimizeFunnelShift(IntrinsicInst *Fsh);
436 bool optimizeSelectInst(SelectInst *SI);
437 bool optimizeShuffleVectorInst(ShuffleVectorInst *SVI);
438 bool optimizeSwitchType(SwitchInst *SI);
439 bool optimizeSwitchPhiConstants(SwitchInst *SI);
440 bool optimizeSwitchInst(SwitchInst *SI);
441 bool optimizeExtractElementInst(Instruction *Inst);
442 bool dupRetToEnableTailCallOpts(BasicBlock *BB, ModifyDT &ModifiedDT);
443 bool fixupDbgVariableRecord(DbgVariableRecord &I);
444 bool fixupDbgVariableRecordsOnInst(Instruction &I);
445 bool placeDbgValues(Function &F);
446 bool placePseudoProbes(Function &F);
447 bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
448 LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
449 bool tryToPromoteExts(TypePromotionTransaction &TPT,
450 const SmallVectorImpl<Instruction *> &Exts,
451 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
452 unsigned CreatedInstsCost = 0);
453 bool mergeSExts(Function &F);
454 bool splitLargeGEPOffsets();
455 bool optimizePhiType(PHINode *Inst, SmallPtrSetImpl<PHINode *> &Visited,
456 SmallPtrSetImpl<Instruction *> &DeletedInstrs);
457 bool optimizePhiTypes(Function &F);
458 bool performAddressTypePromotion(
459 Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
460 bool HasPromoted, TypePromotionTransaction &TPT,
461 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
462 bool splitBranchCondition(Function &F);
463 bool simplifyOffsetableRelocate(GCStatepointInst &I);
464
465 bool tryToSinkFreeOperands(Instruction *I);
466 bool replaceMathCmpWithIntrinsic(BinaryOperator *BO, Value *Arg0, Value *Arg1,
467 CmpInst *Cmp, Intrinsic::ID IID);
468 bool optimizeCmp(CmpInst *Cmp, ModifyDT &ModifiedDT);
469 bool optimizeURem(Instruction *Rem);
470 bool combineToUSubWithOverflow(CmpInst *Cmp, ModifyDT &ModifiedDT);
471 bool combineToUAddWithOverflow(CmpInst *Cmp, ModifyDT &ModifiedDT);
472 bool unfoldPowerOf2Test(CmpInst *Cmp);
473 void verifyBFIUpdates(Function &F);
474 bool _run(Function &F);
475};
476
477class CodeGenPrepareLegacyPass : public FunctionPass {
478public:
479 static char ID; // Pass identification, replacement for typeid
480
481 CodeGenPrepareLegacyPass() : FunctionPass(ID) {}
482
483 bool runOnFunction(Function &F) override;
484
485 StringRef getPassName() const override { return "CodeGen Prepare"; }
486
487 void getAnalysisUsage(AnalysisUsage &AU) const override {
488 // FIXME: When we can selectively preserve passes, preserve the domtree.
489 AU.addRequired<ProfileSummaryInfoWrapperPass>();
490 AU.addRequired<TargetLibraryInfoWrapperPass>();
491 AU.addRequired<TargetPassConfig>();
492 AU.addRequired<TargetTransformInfoWrapperPass>();
493 AU.addRequired<DominatorTreeWrapperPass>();
494 AU.addRequired<LoopInfoWrapperPass>();
495 AU.addRequired<BranchProbabilityInfoWrapperPass>();
496 AU.addRequired<BlockFrequencyInfoWrapperPass>();
497 AU.addUsedIfAvailable<BasicBlockSectionsProfileReaderWrapperPass>();
498 }
499};
500
501} // end anonymous namespace
502
503char CodeGenPrepareLegacyPass::ID = 0;
504
505bool CodeGenPrepareLegacyPass::runOnFunction(Function &F) {
506 if (skipFunction(F))
507 return false;
508 auto TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>();
509 CodeGenPrepare CGP(TM);
510 CGP.DL = &F.getDataLayout();
511 CGP.SubtargetInfo = TM->getSubtargetImpl(F);
512 CGP.TLI = CGP.SubtargetInfo->getTargetLowering();
513 CGP.TRI = CGP.SubtargetInfo->getRegisterInfo();
514 CGP.TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
515 CGP.TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
516 CGP.LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
517 CGP.BPI = &getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI();
518 CGP.BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
519 CGP.PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
520 auto BBSPRWP =
521 getAnalysisIfAvailable<BasicBlockSectionsProfileReaderWrapperPass>();
522 CGP.BBSectionsProfileReader = BBSPRWP ? &BBSPRWP->getBBSPR() : nullptr;
523 DomTreeUpdater DTUpdater(
524 &getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
525 DomTreeUpdater::UpdateStrategy::Lazy);
526 CGP.DTU = &DTUpdater;
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(DominatorTreeWrapperPass)
535INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
536INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
537INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
538INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
539INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
540INITIALIZE_PASS_END(CodeGenPrepareLegacyPass, DEBUG_TYPE,
541 "Optimize for code generation", false, false)
542
543FunctionPass *llvm::createCodeGenPrepareLegacyPass() {
544 return new CodeGenPrepareLegacyPass();
545}
546
547PreservedAnalyses CodeGenPreparePass::run(Function &F,
548 FunctionAnalysisManager &AM) {
549 CodeGenPrepare CGP(TM);
550
551 bool Changed = CGP.run(F, AM);
552 if (!Changed)
553 return PreservedAnalyses::all();
554
555 PreservedAnalyses PA;
556 PA.preserve<TargetLibraryAnalysis>();
557 PA.preserve<TargetIRAnalysis>();
558 return PA;
559}
560
561bool CodeGenPrepare::run(Function &F, FunctionAnalysisManager &AM) {
562 DL = &F.getDataLayout();
563 SubtargetInfo = TM->getSubtargetImpl(F);
564 TLI = SubtargetInfo->getTargetLowering();
565 TRI = SubtargetInfo->getRegisterInfo();
566 TLInfo = &AM.getResult<TargetLibraryAnalysis>(IR&: F);
567 TTI = &AM.getResult<TargetIRAnalysis>(IR&: F);
568 LI = &AM.getResult<LoopAnalysis>(IR&: F);
569 BPI = &AM.getResult<BranchProbabilityAnalysis>(IR&: F);
570 BFI = &AM.getResult<BlockFrequencyAnalysis>(IR&: F);
571 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(IR&: F);
572 PSI = MAMProxy.getCachedResult<ProfileSummaryAnalysis>(IR&: *F.getParent());
573 BBSectionsProfileReader =
574 AM.getCachedResult<BasicBlockSectionsProfileReaderAnalysis>(IR&: F);
575 DomTreeUpdater DTUpdater(&AM.getResult<DominatorTreeAnalysis>(IR&: F),
576 DomTreeUpdater::UpdateStrategy::Lazy);
577 DTU = &DTUpdater;
578 return _run(F);
579}
580
581bool CodeGenPrepare::_run(Function &F) {
582 bool EverMadeChange = false;
583
584 OptSize = F.hasOptSize();
585 // Use the basic-block-sections profile to promote hot functions to .text.hot
586 // if requested.
587 if (BBSectionsGuidedSectionPrefix && BBSectionsProfileReader &&
588 BBSectionsProfileReader->isFunctionHot(FuncName: F.getName())) {
589 (void)F.setSectionPrefix("hot");
590 } else if (ProfileGuidedSectionPrefix) {
591 // The hot attribute overwrites profile count based hotness while profile
592 // counts based hotness overwrite the cold attribute.
593 // This is a conservative behabvior.
594 if (F.hasFnAttribute(Kind: Attribute::Hot) ||
595 PSI->isFunctionHotInCallGraph(F: &F, BFI&: *BFI))
596 (void)F.setSectionPrefix("hot");
597 // If PSI shows this function is not hot, we will placed the function
598 // into unlikely section if (1) PSI shows this is a cold function, or
599 // (2) the function has a attribute of cold.
600 else if (PSI->isFunctionColdInCallGraph(F: &F, BFI&: *BFI) ||
601 F.hasFnAttribute(Kind: Attribute::Cold))
602 (void)F.setSectionPrefix("unlikely");
603 else if (ProfileUnknownInSpecialSection && PSI->hasPartialSampleProfile() &&
604 PSI->isFunctionHotnessUnknown(F))
605 (void)F.setSectionPrefix("unknown");
606 }
607
608 /// This optimization identifies DIV instructions that can be
609 /// profitably bypassed and carried out with a shorter, faster divide.
610 if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI->isSlowDivBypassed()) {
611 const DenseMap<unsigned int, unsigned int> &BypassWidths =
612 TLI->getBypassSlowDivWidths();
613 BasicBlock *BB = &*F.begin();
614 while (BB != nullptr) {
615 // bypassSlowDivision may create new BBs, but we don't want to reapply the
616 // optimization to those blocks.
617 BasicBlock *Next = BB->getNextNode();
618 if (!llvm::shouldOptimizeForSize(BB, PSI, BFI))
619 EverMadeChange |= bypassSlowDivision(BB, BypassWidth: BypassWidths, DTU, LI);
620 BB = Next;
621 }
622 }
623
624 // Get rid of @llvm.assume builtins before attempting to eliminate empty
625 // blocks, since there might be blocks that only contain @llvm.assume calls
626 // (plus arguments that we can get rid of).
627 EverMadeChange |= eliminateAssumptions(F);
628
629 auto resetLoopInfo = [this]() {
630 LI->releaseMemory();
631 LI->analyze(DomTree: DTU->getDomTree());
632 };
633
634 // Eliminate blocks that contain only PHI nodes and an
635 // unconditional branch.
636 bool ResetLI = false;
637 EverMadeChange |= eliminateMostlyEmptyBlocks(F, ResetLI);
638 if (ResetLI)
639 resetLoopInfo();
640
641 if (!DisableBranchOpts)
642 EverMadeChange |= splitBranchCondition(F);
643
644 // Split some critical edges where one of the sources is an indirect branch,
645 // to help generate sane code for PHIs involving such edges.
646 bool Split = SplitIndirectBrCriticalEdges(F, /*IgnoreBlocksWithoutPHI=*/true,
647 BPI, BFI, DTU);
648 EverMadeChange |= Split;
649 if (Split)
650 resetLoopInfo();
651
652#ifndef NDEBUG
653 if (VerifyDomInfo)
654 assert(getDT().verify(DominatorTree::VerificationLevel::Fast) &&
655 "Incorrect DominatorTree updates in CGP");
656
657 if (VerifyLoopInfo)
658 LI->verify(getDT());
659#endif
660
661 // If we are optimzing huge function, we need to consider the build time.
662 // Because the basic algorithm's complex is near O(N!).
663 IsHugeFunc = F.size() > HugeFuncThresholdInCGPP;
664
665 bool MadeChange = true;
666 bool FuncIterated = false;
667 while (MadeChange) {
668 MadeChange = false;
669
670 // This is required because optimizeBlock() calls getDT() inside the loop
671 // below, which flushes pending updates and may delete dead blocks, leading
672 // to iterator invalidation.
673 DTU->flush();
674
675 for (BasicBlock &BB : llvm::make_early_inc_range(Range&: F)) {
676 if (FuncIterated && !FreshBBs.contains(Ptr: &BB))
677 continue;
678
679 ModifyDT ModifiedDTOnIteration = ModifyDT::NotModifyDT;
680 bool Changed = optimizeBlock(BB, ModifiedDT&: ModifiedDTOnIteration);
681
682 MadeChange |= Changed;
683 if (IsHugeFunc) {
684 // If the BB is updated, it may still has chance to be optimized.
685 // This usually happen at sink optimization.
686 // For example:
687 //
688 // bb0：
689 // %and = and i32 %a, 4
690 // %cmp = icmp eq i32 %and, 0
691 //
692 // If the %cmp sink to other BB, the %and will has chance to sink.
693 if (Changed)
694 FreshBBs.insert(Ptr: &BB);
695 else if (FuncIterated)
696 FreshBBs.erase(Ptr: &BB);
697 } else {
698 // For small/normal functions, we restart BB iteration if the dominator
699 // tree of the Function was changed.
700 if (ModifiedDTOnIteration != ModifyDT::NotModifyDT)
701 break;
702 }
703 }
704 // We have iterated all the BB in the (only work for huge) function.
705 FuncIterated = IsHugeFunc;
706
707 if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
708 MadeChange |= mergeSExts(F);
709 if (!LargeOffsetGEPMap.empty())
710 MadeChange |= splitLargeGEPOffsets();
711 MadeChange |= optimizePhiTypes(F);
712
713 if (MadeChange)
714 eliminateFallThrough(F);
715
716#ifndef NDEBUG
717 if (VerifyDomInfo)
718 assert(getDT().verify(DominatorTree::VerificationLevel::Fast) &&
719 "Incorrect DominatorTree updates in CGP");
720
721 if (VerifyLoopInfo)
722 LI->verify(getDT());
723#endif
724
725 // Really free removed instructions during promotion.
726 for (Instruction *I : RemovedInsts)
727 I->deleteValue();
728
729 EverMadeChange |= MadeChange;
730 SeenChainsForSExt.clear();
731 ValToSExtendedUses.clear();
732 RemovedInsts.clear();
733 LargeOffsetGEPMap.clear();
734 LargeOffsetGEPID.clear();
735 }
736
737 NewGEPBases.clear();
738 SunkAddrs.clear();
739
740 // LoopInfo is not needed anymore and ConstantFoldTerminator can break it.
741 LI = nullptr;
742
743 if (!DisableBranchOpts) {
744 MadeChange = false;
745 // Use a set vector to get deterministic iteration order. The order the
746 // blocks are removed may affect whether or not PHI nodes in successors
747 // are removed.
748 SmallSetVector<BasicBlock *, 8> WorkList;
749 for (BasicBlock &BB : F) {
750 SmallVector<BasicBlock *, 2> Successors(successors(BB: &BB));
751 MadeChange |= ConstantFoldTerminator(BB: &BB, DeleteDeadConditions: true, TLI: nullptr, DTU);
752 if (!MadeChange)
753 continue;
754
755 for (BasicBlock *Succ : Successors)
756 if (pred_empty(BB: Succ))
757 WorkList.insert(X: Succ);
758 }
759
760 // Delete the dead blocks and any of their dead successors.
761 MadeChange |= !WorkList.empty();
762 while (!WorkList.empty()) {
763 BasicBlock *BB = WorkList.pop_back_val();
764 SmallVector<BasicBlock *, 2> Successors(successors(BB));
765
766 DeleteDeadBlock(BB, DTU);
767
768 for (BasicBlock *Succ : Successors)
769 if (pred_empty(BB: Succ))
770 WorkList.insert(X: Succ);
771 }
772
773 // Flush pending DT updates in order to finalise deletion of dead blocks.
774 DTU->flush();
775
776 // Merge pairs of basic blocks with unconditional branches, connected by
777 // a single edge.
778 if (EverMadeChange || MadeChange)
779 MadeChange |= eliminateFallThrough(F);
780
781 EverMadeChange |= MadeChange;
782 }
783
784 if (!DisableGCOpts) {
785 SmallVector<GCStatepointInst *, 2> Statepoints;
786 for (BasicBlock &BB : F)
787 for (Instruction &I : BB)
788 if (auto *SP = dyn_cast<GCStatepointInst>(Val: &I))
789 Statepoints.push_back(Elt: SP);
790 for (auto &I : Statepoints)
791 EverMadeChange |= simplifyOffsetableRelocate(I&: *I);
792 }
793
794 // Do this last to clean up use-before-def scenarios introduced by other
795 // preparatory transforms.
796 EverMadeChange |= placeDbgValues(F);
797 EverMadeChange |= placePseudoProbes(F);
798
799#ifndef NDEBUG
800 if (VerifyBFIUpdates)
801 verifyBFIUpdates(F);
802#endif
803
804 return EverMadeChange;
805}
806
807bool CodeGenPrepare::eliminateAssumptions(Function &F) {
808 bool MadeChange = false;
809 for (BasicBlock &BB : F) {
810 CurInstIterator = BB.begin();
811 while (CurInstIterator != BB.end()) {
812 Instruction *I = &*(CurInstIterator++);
813 if (auto *Assume = dyn_cast<AssumeInst>(Val: I)) {
814 MadeChange = true;
815 Value *Operand = Assume->getOperand(i_nocapture: 0);
816 Assume->eraseFromParent();
817
818 resetIteratorIfInvalidatedWhileCalling(BB: &BB, f: [&]() {
819 RecursivelyDeleteTriviallyDeadInstructions(V: Operand, TLI: TLInfo, MSSAU: nullptr);
820 });
821 }
822 }
823 }
824 return MadeChange;
825}
826
827/// An instruction is about to be deleted, so remove all references to it in our
828/// GEP-tracking data strcutures.
829void CodeGenPrepare::removeAllAssertingVHReferences(Value *V) {
830 LargeOffsetGEPMap.erase(Key: V);
831 NewGEPBases.erase(V);
832
833 auto GEP = dyn_cast<GetElementPtrInst>(Val: V);
834 if (!GEP)
835 return;
836
837 LargeOffsetGEPID.erase(Val: GEP);
838
839 auto VecI = LargeOffsetGEPMap.find(Key: GEP->getPointerOperand());
840 if (VecI == LargeOffsetGEPMap.end())
841 return;
842
843 auto &GEPVector = VecI->second;
844 llvm::erase_if(C&: GEPVector, P: [=](auto &Elt) { return Elt.first == GEP; });
845
846 if (GEPVector.empty())
847 LargeOffsetGEPMap.erase(Iterator: VecI);
848}
849
850// Verify BFI has been updated correctly by recomputing BFI and comparing them.
851[[maybe_unused]] void CodeGenPrepare::verifyBFIUpdates(Function &F) {
852 DominatorTree NewDT(F);
853 LoopInfo NewLI(NewDT);
854 BranchProbabilityInfo NewBPI(F, NewLI, TLInfo);
855 BlockFrequencyInfo NewBFI(F, NewBPI, NewLI);
856 NewBFI.verifyMatch(Other&: *BFI);
857}
858
859/// Merge basic blocks which are connected by a single edge, where one of the
860/// basic blocks has a single successor pointing to the other basic block,
861/// which has a single predecessor.
862bool CodeGenPrepare::eliminateFallThrough(Function &F) {
863 bool Changed = false;
864 SmallPtrSet<BasicBlock *, 8> Preds;
865 // Scan all of the blocks in the function, except for the entry block.
866 for (auto &Block : llvm::drop_begin(RangeOrContainer&: F)) {
867 auto *BB = &Block;
868 if (DTU->isBBPendingDeletion(DelBB: BB))
869 continue;
870 // If the destination block has a single pred, then this is a trivial
871 // edge, just collapse it.
872 BasicBlock *SinglePred = BB->getSinglePredecessor();
873
874 // Don't merge if BB's address is taken.
875 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken())
876 continue;
877
878 if (isa<UncondBrInst>(Val: SinglePred->getTerminator())) {
879 Changed = true;
880 LLVM_DEBUG(dbgs() << "To merge:\n" << *BB << "\n\n\n");
881
882 // Merge BB into SinglePred and delete it.
883 MergeBlockIntoPredecessor(BB, DTU, LI);
884 Preds.insert(Ptr: SinglePred);
885
886 if (IsHugeFunc) {
887 // Update FreshBBs to optimize the merged BB.
888 FreshBBs.insert(Ptr: SinglePred);
889 FreshBBs.erase(Ptr: BB);
890 }
891 }
892 }
893
894 // (Repeatedly) merging blocks into their predecessors can create redundant
895 // debug intrinsics.
896 for (auto *Pred : Preds)
897 if (!DTU->isBBPendingDeletion(DelBB: Pred))
898 RemoveRedundantDbgInstrs(BB: Pred);
899
900 return Changed;
901}
902
903/// Find a destination block from BB if BB is mergeable empty block.
904BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
905 // If this block doesn't end with an uncond branch, ignore it.
906 UncondBrInst *BI = dyn_cast<UncondBrInst>(Val: BB->getTerminator());
907 if (!BI)
908 return nullptr;
909
910 // If the instruction before the branch (skipping debug info) isn't a phi
911 // node, then other stuff is happening here.
912 BasicBlock::iterator BBI = BI->getIterator();
913 if (BBI != BB->begin()) {
914 --BBI;
915 if (!isa<PHINode>(Val: BBI))
916 return nullptr;
917 }
918
919 // Do not break infinite loops.
920 BasicBlock *DestBB = BI->getSuccessor();
921 if (DestBB == BB)
922 return nullptr;
923
924 if (!canMergeBlocks(BB, DestBB))
925 DestBB = nullptr;
926
927 return DestBB;
928}
929
930/// Eliminate blocks that contain only PHI nodes, debug info directives, and an
931/// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
932/// edges in ways that are non-optimal for isel. Start by eliminating these
933/// blocks so we can split them the way we want them.
934bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F, bool &ResetLI) {
935 SmallPtrSet<BasicBlock *, 16> Preheaders;
936 SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
937 while (!LoopList.empty()) {
938 Loop *L = LoopList.pop_back_val();
939 llvm::append_range(C&: LoopList, R&: *L);
940 if (BasicBlock *Preheader = L->getLoopPreheader())
941 Preheaders.insert(Ptr: Preheader);
942 }
943
944 ResetLI = false;
945 bool MadeChange = false;
946 // Note that this intentionally skips the entry block.
947 for (auto &Block : llvm::drop_begin(RangeOrContainer&: F)) {
948 // Delete phi nodes that could block deleting other empty blocks.
949 if (!DisableDeletePHIs)
950 MadeChange |= DeleteDeadPHIs(BB: &Block, TLI: TLInfo);
951 }
952
953 for (auto &Block : llvm::drop_begin(RangeOrContainer&: F)) {
954 auto *BB = &Block;
955 if (DTU->isBBPendingDeletion(DelBB: BB))
956 continue;
957 BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
958 if (!DestBB ||
959 !isMergingEmptyBlockProfitable(BB, DestBB, isPreheader: Preheaders.count(Ptr: BB)))
960 continue;
961
962 ResetLI |= eliminateMostlyEmptyBlock(BB);
963 MadeChange = true;
964 }
965 return MadeChange;
966}
967
968bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
969 BasicBlock *DestBB,
970 bool isPreheader) {
971 // Do not delete loop preheaders if doing so would create a critical edge.
972 // Loop preheaders can be good locations to spill registers. If the
973 // preheader is deleted and we create a critical edge, registers may be
974 // spilled in the loop body instead.
975 if (!DisablePreheaderProtect && isPreheader &&
976 !(BB->getSinglePredecessor() &&
977 BB->getSinglePredecessor()->getSingleSuccessor()))
978 return false;
979
980 // Skip merging if the block's successor is also a successor to any callbr
981 // that leads to this block.
982 // FIXME: Is this really needed? Is this a correctness issue?
983 for (BasicBlock *Pred : predecessors(BB)) {
984 if (isa<CallBrInst>(Val: Pred->getTerminator()) &&
985 llvm::is_contained(Range: successors(BB: Pred), Element: DestBB))
986 return false;
987 }
988
989 // Try to skip merging if the unique predecessor of BB is terminated by a
990 // switch or indirect branch instruction, and BB is used as an incoming block
991 // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
992 // add COPY instructions in the predecessor of BB instead of BB (if it is not
993 // merged). Note that the critical edge created by merging such blocks wont be
994 // split in MachineSink because the jump table is not analyzable. By keeping
995 // such empty block (BB), ISel will place COPY instructions in BB, not in the
996 // predecessor of BB.
997 BasicBlock *Pred = BB->getUniquePredecessor();
998 if (!Pred || !(isa<SwitchInst>(Val: Pred->getTerminator()) ||
999 isa<IndirectBrInst>(Val: Pred->getTerminator())))
1000 return true;
1001
1002 if (BB->getTerminator() != &*BB->getFirstNonPHIOrDbg())
1003 return true;
1004
1005 // We use a simple cost heuristic which determine skipping merging is
1006 // profitable if the cost of skipping merging is less than the cost of
1007 // merging : Cost(skipping merging) < Cost(merging BB), where the
1008 // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
1009 // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
1010 // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
1011 // Freq(Pred) / Freq(BB) > 2.
1012 // Note that if there are multiple empty blocks sharing the same incoming
1013 // value for the PHIs in the DestBB, we consider them together. In such
1014 // case, Cost(merging BB) will be the sum of their frequencies.
1015
1016 if (!isa<PHINode>(Val: DestBB->begin()))
1017 return true;
1018
1019 SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
1020
1021 // Find all other incoming blocks from which incoming values of all PHIs in
1022 // DestBB are the same as the ones from BB.
1023 for (BasicBlock *DestBBPred : predecessors(BB: DestBB)) {
1024 if (DestBBPred == BB)
1025 continue;
1026
1027 if (llvm::all_of(Range: DestBB->phis(), P: [&](const PHINode &DestPN) {
1028 return DestPN.getIncomingValueForBlock(BB) ==
1029 DestPN.getIncomingValueForBlock(BB: DestBBPred);
1030 }))
1031 SameIncomingValueBBs.insert(Ptr: DestBBPred);
1032 }
1033
1034 // See if all BB's incoming values are same as the value from Pred. In this
1035 // case, no reason to skip merging because COPYs are expected to be place in
1036 // Pred already.
1037 if (SameIncomingValueBBs.count(Ptr: Pred))
1038 return true;
1039
1040 BlockFrequency PredFreq = BFI->getBlockFreq(BB: Pred);
1041 BlockFrequency BBFreq = BFI->getBlockFreq(BB);
1042
1043 for (auto *SameValueBB : SameIncomingValueBBs)
1044 if (SameValueBB->getUniquePredecessor() == Pred &&
1045 DestBB == findDestBlockOfMergeableEmptyBlock(BB: SameValueBB))
1046 BBFreq += BFI->getBlockFreq(BB: SameValueBB);
1047
1048 std::optional<BlockFrequency> Limit = BBFreq.mul(Factor: FreqRatioToSkipMerge);
1049 return !Limit || PredFreq <= *Limit;
1050}
1051
1052/// Return true if we can merge BB into DestBB if there is a single
1053/// unconditional branch between them, and BB contains no other non-phi
1054/// instructions.
1055bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
1056 const BasicBlock *DestBB) const {
1057 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
1058 // the successor. If there are more complex condition (e.g. preheaders),
1059 // don't mess around with them.
1060 for (const PHINode &PN : BB->phis()) {
1061 for (const User *U : PN.users()) {
1062 const Instruction *UI = cast<Instruction>(Val: U);
1063 if (UI->getParent() != DestBB || !isa<PHINode>(Val: UI))
1064 return false;
1065 // If User is inside DestBB block and it is a PHINode then check
1066 // incoming value. If incoming value is not from BB then this is
1067 // a complex condition (e.g. preheaders) we want to avoid here.
1068 if (UI->getParent() == DestBB) {
1069 if (const PHINode *UPN = dyn_cast<PHINode>(Val: UI))
1070 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
1071 Instruction *Insn = dyn_cast<Instruction>(Val: UPN->getIncomingValue(i: I));
1072 if (Insn && Insn->getParent() == BB &&
1073 Insn->getParent() != UPN->getIncomingBlock(i: I))
1074 return false;
1075 }
1076 }
1077 }
1078 }
1079
1080 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
1081 // and DestBB may have conflicting incoming values for the block. If so, we
1082 // can't merge the block.
1083 const PHINode *DestBBPN = dyn_cast<PHINode>(Val: DestBB->begin());
1084 if (!DestBBPN)
1085 return true; // no conflict.
1086
1087 // Collect the preds of BB.
1088 SmallPtrSet<const BasicBlock *, 16> BBPreds;
1089 if (const PHINode *BBPN = dyn_cast<PHINode>(Val: BB->begin())) {
1090 // It is faster to get preds from a PHI than with pred_iterator.
1091 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1092 BBPreds.insert(Ptr: BBPN->getIncomingBlock(i));
1093 } else {
1094 BBPreds.insert_range(R: predecessors(BB));
1095 }
1096
1097 // Walk the preds of DestBB.
1098 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
1099 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
1100 if (BBPreds.count(Ptr: Pred)) { // Common predecessor?
1101 for (const PHINode &PN : DestBB->phis()) {
1102 const Value *V1 = PN.getIncomingValueForBlock(BB: Pred);
1103 const Value *V2 = PN.getIncomingValueForBlock(BB);
1104
1105 // If V2 is a phi node in BB, look up what the mapped value will be.
1106 if (const PHINode *V2PN = dyn_cast<PHINode>(Val: V2))
1107 if (V2PN->getParent() == BB)
1108 V2 = V2PN->getIncomingValueForBlock(BB: Pred);
1109
1110 // If there is a conflict, bail out.
1111 if (V1 != V2)
1112 return false;
1113 }
1114 }
1115 }
1116
1117 return true;
1118}
1119
1120/// Replace all old uses with new ones, and push the updated BBs into FreshBBs.
1121static void replaceAllUsesWith(Value *Old, Value *New,
1122 SmallPtrSet<BasicBlock *, 32> &FreshBBs,
1123 bool IsHuge) {
1124 auto *OldI = dyn_cast<Instruction>(Val: Old);
1125 if (OldI) {
1126 for (Value::user_iterator UI = OldI->user_begin(), E = OldI->user_end();
1127 UI != E; ++UI) {
1128 Instruction *User = cast<Instruction>(Val: *UI);
1129 if (IsHuge)
1130 FreshBBs.insert(Ptr: User->getParent());
1131 }
1132 }
1133 Old->replaceAllUsesWith(V: New);
1134}
1135
1136/// Eliminate a basic block that has only phi's and an unconditional branch in
1137/// it.
1138/// Indicate that the LoopInfo was modified only if it wasn't updated.
1139bool CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
1140 UncondBrInst *BI = cast<UncondBrInst>(Val: BB->getTerminator());
1141 BasicBlock *DestBB = BI->getSuccessor();
1142
1143 LLVM_DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"
1144 << *BB << *DestBB);
1145
1146 // If the destination block has a single pred, then this is a trivial edge,
1147 // just collapse it.
1148 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
1149 if (SinglePred != DestBB) {
1150 assert(SinglePred == BB &&
1151 "Single predecessor not the same as predecessor");
1152 // Merge DestBB into SinglePred/BB and delete it.
1153 MergeBlockIntoPredecessor(BB: DestBB, DTU, LI);
1154 // Note: BB(=SinglePred) will not be deleted on this path.
1155 // DestBB(=its single successor) is the one that was deleted.
1156 LLVM_DEBUG(dbgs() << "AFTER:\n" << *SinglePred << "\n\n\n");
1157
1158 if (IsHugeFunc) {
1159 // Update FreshBBs to optimize the merged BB.
1160 FreshBBs.insert(Ptr: SinglePred);
1161 FreshBBs.erase(Ptr: DestBB);
1162 }
1163 return false;
1164 }
1165 }
1166
1167 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
1168 // to handle the new incoming edges it is about to have.
1169 for (PHINode &PN : DestBB->phis()) {
1170 // Remove the incoming value for BB, and remember it.
1171 Value *InVal = PN.removeIncomingValue(BB, DeletePHIIfEmpty: false);
1172
1173 // Two options: either the InVal is a phi node defined in BB or it is some
1174 // value that dominates BB.
1175 PHINode *InValPhi = dyn_cast<PHINode>(Val: InVal);
1176 if (InValPhi && InValPhi->getParent() == BB) {
1177 // Add all of the input values of the input PHI as inputs of this phi.
1178 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
1179 PN.addIncoming(V: InValPhi->getIncomingValue(i),
1180 BB: InValPhi->getIncomingBlock(i));
1181 } else {
1182 // Otherwise, add one instance of the dominating value for each edge that
1183 // we will be adding.
1184 if (PHINode *BBPN = dyn_cast<PHINode>(Val: BB->begin())) {
1185 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1186 PN.addIncoming(V: InVal, BB: BBPN->getIncomingBlock(i));
1187 } else {
1188 for (BasicBlock *Pred : predecessors(BB))
1189 PN.addIncoming(V: InVal, BB: Pred);
1190 }
1191 }
1192 }
1193
1194 // Preserve loop Metadata.
1195 if (BI->hasMetadata(KindID: LLVMContext::MD_loop)) {
1196 for (auto *Pred : predecessors(BB))
1197 Pred->getTerminator()->copyMetadata(SrcInst: *BI, WL: LLVMContext::MD_loop);
1198 }
1199
1200 // The PHIs are now updated, change everything that refers to BB to use
1201 // DestBB and remove BB.
1202 SmallVector<DominatorTree::UpdateType, 8> DTUpdates;
1203 SmallPtrSet<BasicBlock *, 8> SeenPreds;
1204 SmallPtrSet<BasicBlock *, 8> PredOfDestBB(llvm::from_range,
1205 predecessors(BB: DestBB));
1206 for (auto *Pred : predecessors(BB)) {
1207 if (!PredOfDestBB.contains(Ptr: Pred)) {
1208 if (SeenPreds.insert(Ptr: Pred).second)
1209 DTUpdates.push_back(Elt: {DominatorTree::Insert, Pred, DestBB});
1210 }
1211 }
1212 SeenPreds.clear();
1213 for (auto *Pred : predecessors(BB)) {
1214 if (SeenPreds.insert(Ptr: Pred).second)
1215 DTUpdates.push_back(Elt: {DominatorTree::Delete, Pred, BB});
1216 }
1217 DTUpdates.push_back(Elt: {DominatorTree::Delete, BB, DestBB});
1218 BB->replaceAllUsesWith(V: DestBB);
1219 DTU->applyUpdates(Updates: DTUpdates);
1220 DTU->deleteBB(DelBB: BB);
1221 ++NumBlocksElim;
1222
1223 LLVM_DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1224 return true;
1225}
1226
1227// Computes a map of base pointer relocation instructions to corresponding
1228// derived pointer relocation instructions given a vector of all relocate calls
1229static void computeBaseDerivedRelocateMap(
1230 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
1231 MapVector<GCRelocateInst *, SmallVector<GCRelocateInst *, 0>>
1232 &RelocateInstMap) {
1233 // Collect information in two maps: one primarily for locating the base object
1234 // while filling the second map; the second map is the final structure holding
1235 // a mapping between Base and corresponding Derived relocate calls
1236 MapVector<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
1237 for (auto *ThisRelocate : AllRelocateCalls) {
1238 auto K = std::make_pair(x: ThisRelocate->getBasePtrIndex(),
1239 y: ThisRelocate->getDerivedPtrIndex());
1240 RelocateIdxMap.insert(KV: std::make_pair(x&: K, y&: ThisRelocate));
1241 }
1242 for (auto &Item : RelocateIdxMap) {
1243 std::pair<unsigned, unsigned> Key = Item.first;
1244 if (Key.first == Key.second)
1245 // Base relocation: nothing to insert
1246 continue;
1247
1248 GCRelocateInst *I = Item.second;
1249 auto BaseKey = std::make_pair(x&: Key.first, y&: Key.first);
1250
1251 // We're iterating over RelocateIdxMap so we cannot modify it.
1252 auto MaybeBase = RelocateIdxMap.find(Key: BaseKey);
1253 if (MaybeBase == RelocateIdxMap.end())
1254 // TODO: We might want to insert a new base object relocate and gep off
1255 // that, if there are enough derived object relocates.
1256 continue;
1257
1258 RelocateInstMap[MaybeBase->second].push_back(Elt: I);
1259 }
1260}
1261
1262// Accepts a GEP and extracts the operands into a vector provided they're all
1263// small integer constants
1264static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
1265 SmallVectorImpl<Value *> &OffsetV) {
1266 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
1267 // Only accept small constant integer operands
1268 auto *Op = dyn_cast<ConstantInt>(Val: GEP->getOperand(i_nocapture: i));
1269 if (!Op || Op->getZExtValue() > 20)
1270 return false;
1271 }
1272
1273 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
1274 OffsetV.push_back(Elt: GEP->getOperand(i_nocapture: i));
1275 return true;
1276}
1277
1278// Takes a RelocatedBase (base pointer relocation instruction) and Targets to
1279// replace, computes a replacement, and affects it.
1280static bool
1281simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
1282 const SmallVectorImpl<GCRelocateInst *> &Targets) {
1283 bool MadeChange = false;
1284 // We must ensure the relocation of derived pointer is defined after
1285 // relocation of base pointer. If we find a relocation corresponding to base
1286 // defined earlier than relocation of base then we move relocation of base
1287 // right before found relocation. We consider only relocation in the same
1288 // basic block as relocation of base. Relocations from other basic block will
1289 // be skipped by optimization and we do not care about them.
1290 for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
1291 &*R != RelocatedBase; ++R)
1292 if (auto *RI = dyn_cast<GCRelocateInst>(Val&: R))
1293 if (RI->getStatepoint() == RelocatedBase->getStatepoint())
1294 if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
1295 RelocatedBase->moveBefore(InsertPos: RI->getIterator());
1296 MadeChange = true;
1297 break;
1298 }
1299
1300 for (GCRelocateInst *ToReplace : Targets) {
1301 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
1302 "Not relocating a derived object of the original base object");
1303 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
1304 // A duplicate relocate call. TODO: coalesce duplicates.
1305 continue;
1306 }
1307
1308 if (RelocatedBase->getParent() != ToReplace->getParent()) {
1309 // Base and derived relocates are in different basic blocks.
1310 // In this case transform is only valid when base dominates derived
1311 // relocate. However it would be too expensive to check dominance
1312 // for each such relocate, so we skip the whole transformation.
1313 continue;
1314 }
1315
1316 Value *Base = ToReplace->getBasePtr();
1317 auto *Derived = dyn_cast<GetElementPtrInst>(Val: ToReplace->getDerivedPtr());
1318 if (!Derived || Derived->getPointerOperand() != Base)
1319 continue;
1320
1321 SmallVector<Value *, 2> OffsetV;
1322 if (!getGEPSmallConstantIntOffsetV(GEP: Derived, OffsetV))
1323 continue;
1324
1325 // Create a Builder and replace the target callsite with a gep
1326 assert(RelocatedBase->getNextNode() &&
1327 "Should always have one since it's not a terminator");
1328
1329 // Insert after RelocatedBase
1330 IRBuilder<> Builder(RelocatedBase->getNextNode());
1331 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1332
1333 // If gc_relocate does not match the actual type, cast it to the right type.
1334 // In theory, there must be a bitcast after gc_relocate if the type does not
1335 // match, and we should reuse it to get the derived pointer. But it could be
1336 // cases like this:
1337 // bb1:
1338 // ...
1339 // %g1 = call coldcc i8 addrspace(1)*
1340 // @llvm.experimental.gc.relocate.p1i8(...) br label %merge
1341 //
1342 // bb2:
1343 // ...
1344 // %g2 = call coldcc i8 addrspace(1)*
1345 // @llvm.experimental.gc.relocate.p1i8(...) br label %merge
1346 //
1347 // merge:
1348 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
1349 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
1350 //
1351 // In this case, we can not find the bitcast any more. So we insert a new
1352 // bitcast no matter there is already one or not. In this way, we can handle
1353 // all cases, and the extra bitcast should be optimized away in later
1354 // passes.
1355 Value *ActualRelocatedBase = RelocatedBase;
1356 if (RelocatedBase->getType() != Base->getType()) {
1357 ActualRelocatedBase =
1358 Builder.CreateBitCast(V: RelocatedBase, DestTy: Base->getType());
1359 }
1360 Value *Replacement =
1361 Builder.CreateGEP(Ty: Derived->getSourceElementType(), Ptr: ActualRelocatedBase,
1362 IdxList: ArrayRef(OffsetV));
1363 Replacement->takeName(V: ToReplace);
1364 // If the newly generated derived pointer's type does not match the original
1365 // derived pointer's type, cast the new derived pointer to match it. Same
1366 // reasoning as above.
1367 Value *ActualReplacement = Replacement;
1368 if (Replacement->getType() != ToReplace->getType()) {
1369 ActualReplacement =
1370 Builder.CreateBitCast(V: Replacement, DestTy: ToReplace->getType());
1371 }
1372 ToReplace->replaceAllUsesWith(V: ActualReplacement);
1373 ToReplace->eraseFromParent();
1374
1375 MadeChange = true;
1376 }
1377 return MadeChange;
1378}
1379
1380// Turns this:
1381//
1382// %base = ...
1383// %ptr = gep %base + 15
1384// %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1385// %base' = relocate(%tok, i32 4, i32 4)
1386// %ptr' = relocate(%tok, i32 4, i32 5)
1387// %val = load %ptr'
1388//
1389// into this:
1390//
1391// %base = ...
1392// %ptr = gep %base + 15
1393// %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1394// %base' = gc.relocate(%tok, i32 4, i32 4)
1395// %ptr' = gep %base' + 15
1396// %val = load %ptr'
1397bool CodeGenPrepare::simplifyOffsetableRelocate(GCStatepointInst &I) {
1398 bool MadeChange = false;
1399 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1400 for (auto *U : I.users())
1401 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(Val: U))
1402 // Collect all the relocate calls associated with a statepoint
1403 AllRelocateCalls.push_back(Elt: Relocate);
1404
1405 // We need at least one base pointer relocation + one derived pointer
1406 // relocation to mangle
1407 if (AllRelocateCalls.size() < 2)
1408 return false;
1409
1410 // RelocateInstMap is a mapping from the base relocate instruction to the
1411 // corresponding derived relocate instructions
1412 MapVector<GCRelocateInst *, SmallVector<GCRelocateInst *, 0>> RelocateInstMap;
1413 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1414 if (RelocateInstMap.empty())
1415 return false;
1416
1417 for (auto &Item : RelocateInstMap)
1418 // Item.first is the RelocatedBase to offset against
1419 // Item.second is the vector of Targets to replace
1420 MadeChange = simplifyRelocatesOffABase(RelocatedBase: Item.first, Targets: Item.second);
1421 return MadeChange;
1422}
1423
1424/// Sink the specified cast instruction into its user blocks.
1425static bool SinkCast(CastInst *CI) {
1426 BasicBlock *DefBB = CI->getParent();
1427
1428 /// InsertedCasts - Only insert a cast in each block once.
1429 DenseMap<BasicBlock *, CastInst *> InsertedCasts;
1430
1431 bool MadeChange = false;
1432 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1433 UI != E;) {
1434 Use &TheUse = UI.getUse();
1435 Instruction *User = cast<Instruction>(Val: *UI);
1436
1437 // Figure out which BB this cast is used in. For PHI's this is the
1438 // appropriate predecessor block.
1439 BasicBlock *UserBB = User->getParent();
1440 if (PHINode *PN = dyn_cast<PHINode>(Val: User)) {
1441 UserBB = PN->getIncomingBlock(U: TheUse);
1442 }
1443
1444 // Preincrement use iterator so we don't invalidate it.
1445 ++UI;
1446
1447 // The first insertion point of a block containing an EH pad is after the
1448 // pad. If the pad is the user, we cannot sink the cast past the pad.
1449 if (User->isEHPad())
1450 continue;
1451
1452 // If the block selected to receive the cast is an EH pad that does not
1453 // allow non-PHI instructions before the terminator, we can't sink the
1454 // cast.
1455 if (UserBB->getTerminator()->isEHPad())
1456 continue;
1457
1458 // If this user is in the same block as the cast, don't change the cast.
1459 if (UserBB == DefBB)
1460 continue;
1461
1462 // If we have already inserted a cast into this block, use it.
1463 CastInst *&InsertedCast = InsertedCasts[UserBB];
1464
1465 if (!InsertedCast) {
1466 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1467 assert(InsertPt != UserBB->end());
1468 InsertedCast = cast<CastInst>(Val: CI->clone());
1469 InsertedCast->insertBefore(BB&: *UserBB, InsertPos: InsertPt);
1470 }
1471
1472 // Replace a use of the cast with a use of the new cast.
1473 TheUse = InsertedCast;
1474 MadeChange = true;
1475 ++NumCastUses;
1476 }
1477
1478 // If we removed all uses, nuke the cast.
1479 if (CI->use_empty()) {
1480 salvageDebugInfo(I&: *CI);
1481 CI->eraseFromParent();
1482 MadeChange = true;
1483 }
1484
1485 return MadeChange;
1486}
1487
1488/// If the specified cast instruction is a noop copy (e.g. it's casting from
1489/// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1490/// reduce the number of virtual registers that must be created and coalesced.
1491///
1492/// Return true if any changes are made.
1493static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1494 const DataLayout &DL) {
1495 // Sink only "cheap" (or nop) address-space casts. This is a weaker condition
1496 // than sinking only nop casts, but is helpful on some platforms.
1497 if (auto *ASC = dyn_cast<AddrSpaceCastInst>(Val: CI)) {
1498 if (!TLI.isFreeAddrSpaceCast(SrcAS: ASC->getSrcAddressSpace(),
1499 DestAS: ASC->getDestAddressSpace()))
1500 return false;
1501 }
1502
1503 // If this is a noop copy,
1504 EVT SrcVT = TLI.getValueType(DL, Ty: CI->getOperand(i_nocapture: 0)->getType());
1505 EVT DstVT = TLI.getValueType(DL, Ty: CI->getType());
1506
1507 // This is an fp<->int conversion?
1508 if (SrcVT.isInteger() != DstVT.isInteger())
1509 return false;
1510
1511 // If this is an extension, it will be a zero or sign extension, which
1512 // isn't a noop.
1513 if (SrcVT.bitsLT(VT: DstVT))
1514 return false;
1515
1516 // If these values will be promoted, find out what they will be promoted
1517 // to. This helps us consider truncates on PPC as noop copies when they
1518 // are.
1519 if (TLI.getTypeAction(Context&: CI->getContext(), VT: SrcVT) ==
1520 TargetLowering::TypePromoteInteger)
1521 SrcVT = TLI.getTypeToTransformTo(Context&: CI->getContext(), VT: SrcVT);
1522 if (TLI.getTypeAction(Context&: CI->getContext(), VT: DstVT) ==
1523 TargetLowering::TypePromoteInteger)
1524 DstVT = TLI.getTypeToTransformTo(Context&: CI->getContext(), VT: DstVT);
1525
1526 // If, after promotion, these are the same types, this is a noop copy.
1527 if (SrcVT != DstVT)
1528 return false;
1529
1530 return SinkCast(CI);
1531}
1532
1533// Match a simple increment by constant operation. Note that if a sub is
1534// matched, the step is negated (as if the step had been canonicalized to
1535// an add, even though we leave the instruction alone.)
1536static bool matchIncrement(const Instruction *IVInc, Instruction *&LHS,
1537 Constant *&Step) {
1538 if (match(V: IVInc, P: m_Add(L: m_Instruction(I&: LHS), R: m_Constant(C&: Step))) ||
1539 match(V: IVInc, P: m_ExtractValue<0>(V: m_Intrinsic<Intrinsic::uadd_with_overflow>(
1540 Op0: m_Instruction(I&: LHS), Op1: m_Constant(C&: Step)))))
1541 return true;
1542 if (match(V: IVInc, P: m_Sub(L: m_Instruction(I&: LHS), R: m_Constant(C&: Step))) ||
1543 match(V: IVInc, P: m_ExtractValue<0>(V: m_Intrinsic<Intrinsic::usub_with_overflow>(
1544 Op0: m_Instruction(I&: LHS), Op1: m_Constant(C&: Step))))) {
1545 Step = ConstantExpr::getNeg(C: Step);
1546 return true;
1547 }
1548 return false;
1549}
1550
1551/// If given \p PN is an inductive variable with value IVInc coming from the
1552/// backedge, and on each iteration it gets increased by Step, return pair
1553/// <IVInc, Step>. Otherwise, return std::nullopt.
1554static std::optional<std::pair<Instruction *, Constant *>>
1555getIVIncrement(const PHINode *PN, const LoopInfo *LI) {
1556 const Loop *L = LI->getLoopFor(BB: PN->getParent());
1557 if (!L || L->getHeader() != PN->getParent() || !L->getLoopLatch())
1558 return std::nullopt;
1559 auto *IVInc =
1560 dyn_cast<Instruction>(Val: PN->getIncomingValueForBlock(BB: L->getLoopLatch()));
1561 if (!IVInc || LI->getLoopFor(BB: IVInc->getParent()) != L)
1562 return std::nullopt;
1563 Instruction *LHS = nullptr;
1564 Constant *Step = nullptr;
1565 if (matchIncrement(IVInc, LHS, Step) && LHS == PN)
1566 return std::make_pair(x&: IVInc, y&: Step);
1567 return std::nullopt;
1568}
1569
1570static bool isIVIncrement(const Value *V, const LoopInfo *LI) {
1571 auto *I = dyn_cast<Instruction>(Val: V);
1572 if (!I)
1573 return false;
1574 Instruction *LHS = nullptr;
1575 Constant *Step = nullptr;
1576 if (!matchIncrement(IVInc: I, LHS, Step))
1577 return false;
1578 if (auto *PN = dyn_cast<PHINode>(Val: LHS))
1579 if (auto IVInc = getIVIncrement(PN, LI))
1580 return IVInc->first == I;
1581 return false;
1582}
1583
1584bool CodeGenPrepare::replaceMathCmpWithIntrinsic(BinaryOperator *BO,
1585 Value *Arg0, Value *Arg1,
1586 CmpInst *Cmp,
1587 Intrinsic::ID IID) {
1588 auto IsReplacableIVIncrement = [this, &Cmp](BinaryOperator *BO) {
1589 if (!isIVIncrement(V: BO, LI))
1590 return false;
1591 const Loop *L = LI->getLoopFor(BB: BO->getParent());
1592 assert(L && "L should not be null after isIVIncrement()");
1593 // Do not risk on moving increment into a child loop.
1594 if (LI->getLoopFor(BB: Cmp->getParent()) != L)
1595 return false;
1596
1597 // Finally, we need to ensure that the insert point will dominate all
1598 // existing uses of the increment.
1599
1600 auto &DT = getDT();
1601 if (DT.dominates(A: Cmp->getParent(), B: BO->getParent()))
1602 // If we're moving up the dom tree, all uses are trivially dominated.
1603 // (This is the common case for code produced by LSR.)
1604 return true;
1605
1606 // Otherwise, special case the single use in the phi recurrence.
1607 return BO->hasOneUse() && DT.dominates(A: Cmp->getParent(), B: L->getLoopLatch());
1608 };
1609 if (BO->getParent() != Cmp->getParent() && !IsReplacableIVIncrement(BO)) {
1610 // We used to use a dominator tree here to allow multi-block optimization.
1611 // But that was problematic because:
1612 // 1. It could cause a perf regression by hoisting the math op into the
1613 // critical path.
1614 // 2. It could cause a perf regression by creating a value that was live
1615 // across multiple blocks and increasing register pressure.
1616 // 3. Use of a dominator tree could cause large compile-time regression.
1617 // This is because we recompute the DT on every change in the main CGP
1618 // run-loop. The recomputing is probably unnecessary in many cases, so if
1619 // that was fixed, using a DT here would be ok.
1620 //
1621 // There is one important particular case we still want to handle: if BO is
1622 // the IV increment. Important properties that make it profitable:
1623 // - We can speculate IV increment anywhere in the loop (as long as the
1624 // indvar Phi is its only user);
1625 // - Upon computing Cmp, we effectively compute something equivalent to the
1626 // IV increment (despite it loops differently in the IR). So moving it up
1627 // to the cmp point does not really increase register pressure.
1628 return false;
1629 }
1630
1631 // We allow matching the canonical IR (add X, C) back to (usubo X, -C).
1632 if (BO->getOpcode() == Instruction::Add &&
1633 IID == Intrinsic::usub_with_overflow) {
1634 assert(isa<Constant>(Arg1) && "Unexpected input for usubo");
1635 Arg1 = ConstantExpr::getNeg(C: cast<Constant>(Val: Arg1));
1636 }
1637
1638 // Insert at the first instruction of the pair.
1639 Instruction *InsertPt = nullptr;
1640 for (Instruction &Iter : *Cmp->getParent()) {
1641 // If BO is an XOR, it is not guaranteed that it comes after both inputs to
1642 // the overflow intrinsic are defined.
1643 if ((BO->getOpcode() != Instruction::Xor && &Iter == BO) || &Iter == Cmp) {
1644 InsertPt = &Iter;
1645 break;
1646 }
1647 }
1648 assert(InsertPt != nullptr && "Parent block did not contain cmp or binop");
1649
1650 IRBuilder<> Builder(InsertPt);
1651 Value *MathOV = Builder.CreateBinaryIntrinsic(ID: IID, LHS: Arg0, RHS: Arg1);
1652 if (BO->getOpcode() != Instruction::Xor) {
1653 Value *Math = Builder.CreateExtractValue(Agg: MathOV, Idxs: 0, Name: "math");
1654 replaceAllUsesWith(Old: BO, New: Math, FreshBBs, IsHuge: IsHugeFunc);
1655 } else
1656 assert(BO->hasOneUse() &&
1657 "Patterns with XOr should use the BO only in the compare");
1658 Value *OV = Builder.CreateExtractValue(Agg: MathOV, Idxs: 1, Name: "ov");
1659 replaceAllUsesWith(Old: Cmp, New: OV, FreshBBs, IsHuge: IsHugeFunc);
1660 Cmp->eraseFromParent();
1661 BO->eraseFromParent();
1662 return true;
1663}
1664
1665/// Match special-case patterns that check for unsigned add overflow.
1666static bool matchUAddWithOverflowConstantEdgeCases(CmpInst *Cmp,
1667 BinaryOperator *&Add) {
1668 // Add = add A, 1; Cmp = icmp eq A,-1 (overflow if A is max val)
1669 // Add = add A,-1; Cmp = icmp ne A, 0 (overflow if A is non-zero)
1670 Value *A = Cmp->getOperand(i_nocapture: 0), *B = Cmp->getOperand(i_nocapture: 1);
1671
1672 // We are not expecting non-canonical/degenerate code. Just bail out.
1673 if (isa<Constant>(Val: A))
1674 return false;
1675
1676 ICmpInst::Predicate Pred = Cmp->getPredicate();
1677 if (Pred == ICmpInst::ICMP_EQ && match(V: B, P: m_AllOnes()))
1678 B = ConstantInt::get(Ty: B->getType(), V: 1);
1679 else if (Pred == ICmpInst::ICMP_NE && match(V: B, P: m_ZeroInt()))
1680 B = Constant::getAllOnesValue(Ty: B->getType());
1681 else
1682 return false;
1683
1684 // Check the users of the variable operand of the compare looking for an add
1685 // with the adjusted constant.
1686 for (User *U : A->users()) {
1687 if (match(V: U, P: m_Add(L: m_Specific(V: A), R: m_Specific(V: B)))) {
1688 Add = cast<BinaryOperator>(Val: U);
1689 return true;
1690 }
1691 }
1692 return false;
1693}
1694
1695/// Try to combine the compare into a call to the llvm.uadd.with.overflow
1696/// intrinsic. Return true if any changes were made.
1697bool CodeGenPrepare::combineToUAddWithOverflow(CmpInst *Cmp,
1698 ModifyDT &ModifiedDT) {
1699 bool EdgeCase = false;
1700 Value *A, *B;
1701 BinaryOperator *Add;
1702 if (!match(V: Cmp, P: m_UAddWithOverflow(L: m_Value(V&: A), R: m_Value(V&: B), S: m_BinOp(I&: Add)))) {
1703 if (!matchUAddWithOverflowConstantEdgeCases(Cmp, Add))
1704 return false;
1705 // Set A and B in case we match matchUAddWithOverflowConstantEdgeCases.
1706 A = Add->getOperand(i_nocapture: 0);
1707 B = Add->getOperand(i_nocapture: 1);
1708 EdgeCase = true;
1709 }
1710
1711 if (!TLI->shouldFormOverflowOp(Opcode: ISD::UADDO,
1712 VT: TLI->getValueType(DL: *DL, Ty: Add->getType()),
1713 MathUsed: Add->hasNUsesOrMore(N: EdgeCase ? 1 : 2)))
1714 return false;
1715
1716 // We don't want to move around uses of condition values this late, so we
1717 // check if it is legal to create the call to the intrinsic in the basic
1718 // block containing the icmp.
1719 if (Add->getParent() != Cmp->getParent() && !Add->hasOneUse())
1720 return false;
1721
1722 if (!replaceMathCmpWithIntrinsic(BO: Add, Arg0: A, Arg1: B, Cmp,
1723 IID: Intrinsic::uadd_with_overflow))
1724 return false;
1725
1726 // Reset callers - do not crash by iterating over a dead instruction.
1727 ModifiedDT = ModifyDT::ModifyInstDT;
1728 return true;
1729}
1730
1731bool CodeGenPrepare::combineToUSubWithOverflow(CmpInst *Cmp,
1732 ModifyDT &ModifiedDT) {
1733 // We are not expecting non-canonical/degenerate code. Just bail out.
1734 Value *A = Cmp->getOperand(i_nocapture: 0), *B = Cmp->getOperand(i_nocapture: 1);
1735 if (isa<Constant>(Val: A) && isa<Constant>(Val: B))
1736 return false;
1737
1738 // Convert (A u> B) to (A u< B) to simplify pattern matching.
1739 ICmpInst::Predicate Pred = Cmp->getPredicate();
1740 if (Pred == ICmpInst::ICMP_UGT) {
1741 std::swap(a&: A, b&: B);
1742 Pred = ICmpInst::ICMP_ULT;
1743 }
1744 // Convert special-case: (A == 0) is the same as (A u< 1).
1745 if (Pred == ICmpInst::ICMP_EQ && match(V: B, P: m_ZeroInt())) {
1746 B = ConstantInt::get(Ty: B->getType(), V: 1);
1747 Pred = ICmpInst::ICMP_ULT;
1748 }
1749 // Convert special-case: (A != 0) is the same as (0 u< A).
1750 if (Pred == ICmpInst::ICMP_NE && match(V: B, P: m_ZeroInt())) {
1751 std::swap(a&: A, b&: B);
1752 Pred = ICmpInst::ICMP_ULT;
1753 }
1754 if (Pred != ICmpInst::ICMP_ULT)
1755 return false;
1756
1757 // Walk the users of a variable operand of a compare looking for a subtract or
1758 // add with that same operand. Also match the 2nd operand of the compare to
1759 // the add/sub, but that may be a negated constant operand of an add.
1760 Value *CmpVariableOperand = isa<Constant>(Val: A) ? B : A;
1761 BinaryOperator *Sub = nullptr;
1762 for (User *U : CmpVariableOperand->users()) {
1763 // A - B, A u< B --> usubo(A, B)
1764 if (match(V: U, P: m_Sub(L: m_Specific(V: A), R: m_Specific(V: B)))) {
1765 Sub = cast<BinaryOperator>(Val: U);
1766 break;
1767 }
1768
1769 // A + (-C), A u< C (canonicalized form of (sub A, C))
1770 const APInt *CmpC, *AddC;
1771 if (match(V: U, P: m_Add(L: m_Specific(V: A), R: m_APInt(Res&: AddC))) &&
1772 match(V: B, P: m_APInt(Res&: CmpC)) && *AddC == -(*CmpC)) {
1773 Sub = cast<BinaryOperator>(Val: U);
1774 break;
1775 }
1776 }
1777 if (!Sub)
1778 return false;
1779
1780 if (!TLI->shouldFormOverflowOp(Opcode: ISD::USUBO,
1781 VT: TLI->getValueType(DL: *DL, Ty: Sub->getType()),
1782 MathUsed: Sub->hasNUsesOrMore(N: 1)))
1783 return false;
1784
1785 // We don't want to move around uses of condition values this late, so we
1786 // check if it is legal to create the call to the intrinsic in the basic
1787 // block containing the icmp.
1788 if (Sub->getParent() != Cmp->getParent() && !Sub->hasOneUse())
1789 return false;
1790
1791 if (!replaceMathCmpWithIntrinsic(BO: Sub, Arg0: Sub->getOperand(i_nocapture: 0), Arg1: Sub->getOperand(i_nocapture: 1),
1792 Cmp, IID: Intrinsic::usub_with_overflow))
1793 return false;
1794
1795 // Reset callers - do not crash by iterating over a dead instruction.
1796 ModifiedDT = ModifyDT::ModifyInstDT;
1797 return true;
1798}
1799
1800// Decanonicalizes icmp+ctpop power-of-two test if ctpop is slow.
1801// The same transformation exists in DAG combiner, but we repeat it here because
1802// DAG builder can break the pattern by moving icmp into a successor block.
1803bool CodeGenPrepare::unfoldPowerOf2Test(CmpInst *Cmp) {
1804 CmpPredicate Pred;
1805 Value *X;
1806 const APInt *C;
1807
1808 // (icmp (ctpop x), c)
1809 if (!match(V: Cmp, P: m_ICmp(Pred, L: m_Intrinsic<Intrinsic::ctpop>(Op0: m_Value(V&: X)),
1810 R: m_APIntAllowPoison(Res&: C))))
1811 return false;
1812
1813 // We're only interested in "is power of 2 [or zero]" patterns.
1814 bool IsStrictlyPowerOf2Test = ICmpInst::isEquality(P: Pred) && *C == 1;
1815 bool IsPowerOf2OrZeroTest = (Pred == CmpInst::ICMP_ULT && *C == 2) ||
1816 (Pred == CmpInst::ICMP_UGT && *C == 1);
1817 if (!IsStrictlyPowerOf2Test && !IsPowerOf2OrZeroTest)
1818 return false;
1819
1820 // Some targets have better codegen for `ctpop(x) u</u>= 2/1`than for
1821 // `ctpop(x) ==/!= 1`. If ctpop is fast, only try changing the comparison,
1822 // and otherwise expand ctpop into a few simple instructions.
1823 Type *OpTy = X->getType();
1824 if (TLI->isCtpopFast(VT: TLI->getValueType(DL: *DL, Ty: OpTy))) {
1825 // Look for `ctpop(x) ==/!= 1`, where `ctpop(x)` is known to be non-zero.
1826 if (!IsStrictlyPowerOf2Test || !isKnownNonZero(V: Cmp->getOperand(i_nocapture: 0), Q: *DL))
1827 return false;
1828
1829 // ctpop(x) == 1 -> ctpop(x) u< 2
1830 // ctpop(x) != 1 -> ctpop(x) u> 1
1831 if (Pred == ICmpInst::ICMP_EQ) {
1832 Cmp->setOperand(i_nocapture: 1, Val_nocapture: ConstantInt::get(Ty: OpTy, V: 2));
1833 Cmp->setPredicate(ICmpInst::ICMP_ULT);
1834 } else {
1835 Cmp->setPredicate(ICmpInst::ICMP_UGT);
1836 }
1837 return true;
1838 }
1839
1840 Value *NewCmp;
1841 if (IsPowerOf2OrZeroTest ||
1842 (IsStrictlyPowerOf2Test && isKnownNonZero(V: Cmp->getOperand(i_nocapture: 0), Q: *DL))) {
1843 // ctpop(x) u< 2 -> (x & (x - 1)) == 0
1844 // ctpop(x) u> 1 -> (x & (x - 1)) != 0
1845 IRBuilder<> Builder(Cmp);
1846 Value *Sub = Builder.CreateAdd(LHS: X, RHS: Constant::getAllOnesValue(Ty: OpTy));
1847 Value *And = Builder.CreateAnd(LHS: X, RHS: Sub);
1848 CmpInst::Predicate NewPred =
1849 (Pred == CmpInst::ICMP_ULT || Pred == CmpInst::ICMP_EQ)
1850 ? CmpInst::ICMP_EQ
1851 : CmpInst::ICMP_NE;
1852 NewCmp = Builder.CreateICmp(P: NewPred, LHS: And, RHS: ConstantInt::getNullValue(Ty: OpTy));
1853 } else {
1854 // ctpop(x) == 1 -> (x ^ (x - 1)) u> (x - 1)
1855 // ctpop(x) != 1 -> (x ^ (x - 1)) u<= (x - 1)
1856 IRBuilder<> Builder(Cmp);
1857 Value *Sub = Builder.CreateAdd(LHS: X, RHS: Constant::getAllOnesValue(Ty: OpTy));
1858 Value *Xor = Builder.CreateXor(LHS: X, RHS: Sub);
1859 CmpInst::Predicate NewPred =
1860 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT : CmpInst::ICMP_ULE;
1861 NewCmp = Builder.CreateICmp(P: NewPred, LHS: Xor, RHS: Sub);
1862 }
1863
1864 Cmp->replaceAllUsesWith(V: NewCmp);
1865 RecursivelyDeleteTriviallyDeadInstructions(V: Cmp);
1866 return true;
1867}
1868
1869/// Sink the given CmpInst into user blocks to reduce the number of virtual
1870/// registers that must be created and coalesced. This is a clear win except on
1871/// targets with multiple condition code registers (PowerPC), where it might
1872/// lose; some adjustment may be wanted there.
1873///
1874/// Return true if any changes are made.
1875static bool sinkCmpExpression(CmpInst *Cmp, const TargetLowering &TLI,
1876 const DataLayout &DL) {
1877 if (TLI.hasMultipleConditionRegisters(VT: EVT::getEVT(Ty: Cmp->getType())))
1878 return false;
1879
1880 // Avoid sinking soft-FP comparisons, since this can move them into a loop.
1881 if (TLI.useSoftFloat() && isa<FCmpInst>(Val: Cmp))
1882 return false;
1883
1884 bool UsedInPhiOrCurrentBlock = any_of(Range: Cmp->users(), P: [Cmp](User *U) {
1885 return isa<PHINode>(Val: U) ||
1886 cast<Instruction>(Val: U)->getParent() == Cmp->getParent();
1887 });
1888
1889 // Avoid sinking larger than legal integer comparisons unless its ONLY used in
1890 // another BB.
1891 if (UsedInPhiOrCurrentBlock && Cmp->getOperand(i_nocapture: 0)->getType()->isIntegerTy() &&
1892 Cmp->getOperand(i_nocapture: 0)->getType()->getScalarSizeInBits() >
1893 DL.getLargestLegalIntTypeSizeInBits())
1894 return false;
1895
1896 // Only insert a cmp in each block once.
1897 DenseMap<BasicBlock *, CmpInst *> InsertedCmps;
1898
1899 bool MadeChange = false;
1900 for (Value::user_iterator UI = Cmp->user_begin(), E = Cmp->user_end();
1901 UI != E;) {
1902 Use &TheUse = UI.getUse();
1903 Instruction *User = cast<Instruction>(Val: *UI);
1904
1905 // Preincrement use iterator so we don't invalidate it.
1906 ++UI;
1907
1908 // Don't bother for PHI nodes.
1909 if (isa<PHINode>(Val: User))
1910 continue;
1911
1912 // Figure out which BB this cmp is used in.
1913 BasicBlock *UserBB = User->getParent();
1914 BasicBlock *DefBB = Cmp->getParent();
1915
1916 // If this user is in the same block as the cmp, don't change the cmp.
1917 if (UserBB == DefBB)
1918 continue;
1919
1920 // If we have already inserted a cmp into this block, use it.
1921 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1922
1923 if (!InsertedCmp) {
1924 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1925 assert(InsertPt != UserBB->end());
1926 InsertedCmp = CmpInst::Create(Op: Cmp->getOpcode(), Pred: Cmp->getPredicate(),
1927 S1: Cmp->getOperand(i_nocapture: 0), S2: Cmp->getOperand(i_nocapture: 1), Name: "");
1928 InsertedCmp->insertBefore(BB&: *UserBB, InsertPos: InsertPt);
1929 // Propagate the debug info.
1930 InsertedCmp->setDebugLoc(Cmp->getDebugLoc());
1931 }
1932
1933 // Replace a use of the cmp with a use of the new cmp.
1934 TheUse = InsertedCmp;
1935 MadeChange = true;
1936 ++NumCmpUses;
1937 }
1938
1939 // If we removed all uses, nuke the cmp.
1940 if (Cmp->use_empty()) {
1941 Cmp->eraseFromParent();
1942 MadeChange = true;
1943 }
1944
1945 return MadeChange;
1946}
1947
1948/// For pattern like:
1949///
1950/// DomCond = icmp sgt/slt CmpOp0, CmpOp1 (might not be in DomBB)
1951/// ...
1952/// DomBB:
1953/// ...
1954/// br DomCond, TrueBB, CmpBB
1955/// CmpBB: (with DomBB being the single predecessor)
1956/// ...
1957/// Cmp = icmp eq CmpOp0, CmpOp1
1958/// ...
1959///
1960/// It would use two comparison on targets that lowering of icmp sgt/slt is
1961/// different from lowering of icmp eq (PowerPC). This function try to convert
1962/// 'Cmp = icmp eq CmpOp0, CmpOp1' to ' Cmp = icmp slt/sgt CmpOp0, CmpOp1'.
1963/// After that, DomCond and Cmp can use the same comparison so reduce one
1964/// comparison.
1965///
1966/// Return true if any changes are made.
1967static bool foldICmpWithDominatingICmp(CmpInst *Cmp,
1968 const TargetLowering &TLI) {
1969 if (!EnableICMP_EQToICMP_ST && TLI.isEqualityCmpFoldedWithSignedCmp())
1970 return false;
1971
1972 ICmpInst::Predicate Pred = Cmp->getPredicate();
1973 if (Pred != ICmpInst::ICMP_EQ)
1974 return false;
1975
1976 // If icmp eq has users other than CondBrInst and SelectInst, converting it to
1977 // icmp slt/sgt would introduce more redundant LLVM IR.
1978 for (User *U : Cmp->users()) {
1979 if (isa<CondBrInst>(Val: U))
1980 continue;
1981 if (isa<SelectInst>(Val: U) && cast<SelectInst>(Val: U)->getCondition() == Cmp)
1982 continue;
1983 return false;
1984 }
1985
1986 // This is a cheap/incomplete check for dominance - just match a single
1987 // predecessor with a conditional branch.
1988 BasicBlock *CmpBB = Cmp->getParent();
1989 BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1990 if (!DomBB)
1991 return false;
1992
1993 // We want to ensure that the only way control gets to the comparison of
1994 // interest is that a less/greater than comparison on the same operands is
1995 // false.
1996 Value *DomCond;
1997 BasicBlock *TrueBB, *FalseBB;
1998 if (!match(V: DomBB->getTerminator(), P: m_Br(C: m_Value(V&: DomCond), T&: TrueBB, F&: FalseBB)))
1999 return false;
2000 if (CmpBB != FalseBB)
2001 return false;
2002
2003 Value *CmpOp0 = Cmp->getOperand(i_nocapture: 0), *CmpOp1 = Cmp->getOperand(i_nocapture: 1);
2004 CmpPredicate DomPred;
2005 if (!match(V: DomCond, P: m_ICmp(Pred&: DomPred, L: m_Specific(V: CmpOp0), R: m_Specific(V: CmpOp1))))
2006 return false;
2007 if (DomPred != ICmpInst::ICMP_SGT && DomPred != ICmpInst::ICMP_SLT)
2008 return false;
2009
2010 // Convert the equality comparison to the opposite of the dominating
2011 // comparison and swap the direction for all branch/select users.
2012 // We have conceptually converted:
2013 // Res = (a < b) ? <LT_RES> : (a == b) ? <EQ_RES> : <GT_RES>;
2014 // to
2015 // Res = (a < b) ? <LT_RES> : (a > b) ? <GT_RES> : <EQ_RES>;
2016 // And similarly for branches.
2017 for (User *U : Cmp->users()) {
2018 if (auto *BI = dyn_cast<CondBrInst>(Val: U)) {
2019 BI->swapSuccessors();
2020 continue;
2021 }
2022 if (auto *SI = dyn_cast<SelectInst>(Val: U)) {
2023 // Swap operands
2024 SI->swapValues();
2025 SI->swapProfMetadata();
2026 continue;
2027 }
2028 llvm_unreachable("Must be a branch or a select");
2029 }
2030 Cmp->setPredicate(CmpInst::getSwappedPredicate(pred: DomPred));
2031 return true;
2032}
2033
2034/// Many architectures use the same instruction for both subtract and cmp. Try
2035/// to swap cmp operands to match subtract operations to allow for CSE.
2036static bool swapICmpOperandsToExposeCSEOpportunities(CmpInst *Cmp) {
2037 Value *Op0 = Cmp->getOperand(i_nocapture: 0);
2038 Value *Op1 = Cmp->getOperand(i_nocapture: 1);
2039 if (!Op0->getType()->isIntegerTy() || isa<Constant>(Val: Op0) ||
2040 isa<Constant>(Val: Op1) || Op0 == Op1)
2041 return false;
2042
2043 // If a subtract already has the same operands as a compare, swapping would be
2044 // bad. If a subtract has the same operands as a compare but in reverse order,
2045 // then swapping is good.
2046 int GoodToSwap = 0;
2047 unsigned NumInspected = 0;
2048 for (const User *U : Op0->users()) {
2049 // Avoid walking many users.
2050 if (++NumInspected > 128)
2051 return false;
2052 if (match(V: U, P: m_Sub(L: m_Specific(V: Op1), R: m_Specific(V: Op0))))
2053 GoodToSwap++;
2054 else if (match(V: U, P: m_Sub(L: m_Specific(V: Op0), R: m_Specific(V: Op1))))
2055 GoodToSwap--;
2056 }
2057
2058 if (GoodToSwap > 0) {
2059 Cmp->swapOperands();
2060 return true;
2061 }
2062 return false;
2063}
2064
2065static bool foldFCmpToFPClassTest(CmpInst *Cmp, const TargetLowering &TLI,
2066 const DataLayout &DL) {
2067 FCmpInst *FCmp = dyn_cast<FCmpInst>(Val: Cmp);
2068 if (!FCmp)
2069 return false;
2070
2071 // Don't fold if the target offers free fabs and the predicate is legal.
2072 EVT VT = TLI.getValueType(DL, Ty: Cmp->getOperand(i_nocapture: 0)->getType());
2073 if (TLI.isFAbsFree(VT) &&
2074 TLI.isCondCodeLegal(CC: getFCmpCondCode(Pred: FCmp->getPredicate()),
2075 VT: VT.getSimpleVT()))
2076 return false;
2077
2078 // Reverse the canonicalization if it is a FP class test
2079 auto ShouldReverseTransform = [](FPClassTest ClassTest) {
2080 return ClassTest == fcInf || ClassTest == (fcInf | fcNan);
2081 };
2082 auto [ClassVal, ClassTest] =
2083 fcmpToClassTest(Pred: FCmp->getPredicate(), F: *FCmp->getParent()->getParent(),
2084 LHS: FCmp->getOperand(i_nocapture: 0), RHS: FCmp->getOperand(i_nocapture: 1));
2085 if (!ClassVal)
2086 return false;
2087
2088 if (!ShouldReverseTransform(ClassTest) && !ShouldReverseTransform(~ClassTest))
2089 return false;
2090
2091 IRBuilder<> Builder(Cmp);
2092 Value *IsFPClass = Builder.createIsFPClass(FPNum: ClassVal, Test: ClassTest);
2093 Cmp->replaceAllUsesWith(V: IsFPClass);
2094 RecursivelyDeleteTriviallyDeadInstructions(V: Cmp);
2095 return true;
2096}
2097
2098static bool isRemOfLoopIncrementWithLoopInvariant(
2099 Instruction *Rem, const LoopInfo *LI, Value *&RemAmtOut, Value *&AddInstOut,
2100 Value *&AddOffsetOut, PHINode *&LoopIncrPNOut) {
2101 Value *Incr, *RemAmt;
2102 // NB: If RemAmt is a power of 2 it *should* have been transformed by now.
2103 if (!match(V: Rem, P: m_URem(L: m_Value(V&: Incr), R: m_Value(V&: RemAmt))))
2104 return false;
2105
2106 Value *AddInst, *AddOffset;
2107 // Find out loop increment PHI.
2108 auto *PN = dyn_cast<PHINode>(Val: Incr);
2109 if (PN != nullptr) {
2110 AddInst = nullptr;
2111 AddOffset = nullptr;
2112 } else {
2113 // Search through a NUW add on top of the loop increment.
2114 Value *V0, *V1;
2115 if (!match(V: Incr, P: m_NUWAdd(L: m_Value(V&: V0), R: m_Value(V&: V1))))
2116 return false;
2117
2118 AddInst = Incr;
2119 PN = dyn_cast<PHINode>(Val: V0);
2120 if (PN != nullptr) {
2121 AddOffset = V1;
2122 } else {
2123 PN = dyn_cast<PHINode>(Val: V1);
2124 AddOffset = V0;
2125 }
2126 }
2127
2128 if (!PN)
2129 return false;
2130
2131 // This isn't strictly necessary, what we really need is one increment and any
2132 // amount of initial values all being the same.
2133 if (PN->getNumIncomingValues() != 2)
2134 return false;
2135
2136 // Only trivially analyzable loops.
2137 Loop *L = LI->getLoopFor(BB: PN->getParent());
2138 if (!L || !L->getLoopPreheader() || !L->getLoopLatch())
2139 return false;
2140
2141 // Req that the remainder is in the loop
2142 if (!L->contains(Inst: Rem))
2143 return false;
2144
2145 // Only works if the remainder amount is a loop invaraint
2146 if (!L->isLoopInvariant(V: RemAmt))
2147 return false;
2148
2149 // Only works if the AddOffset is a loop invaraint
2150 if (AddOffset && !L->isLoopInvariant(V: AddOffset))
2151 return false;
2152
2153 // Is the PHI a loop increment?
2154 auto LoopIncrInfo = getIVIncrement(PN, LI);
2155 if (!LoopIncrInfo)
2156 return false;
2157
2158 // We need remainder_amount % increment_amount to be zero. Increment of one
2159 // satisfies that without any special logic and is overwhelmingly the common
2160 // case.
2161 if (!match(V: LoopIncrInfo->second, P: m_One()))
2162 return false;
2163
2164 // Need the increment to not overflow.
2165 if (!match(V: LoopIncrInfo->first, P: m_c_NUWAdd(L: m_Specific(V: PN), R: m_Value())))
2166 return false;
2167
2168 // Set output variables.
2169 RemAmtOut = RemAmt;
2170 LoopIncrPNOut = PN;
2171 AddInstOut = AddInst;
2172 AddOffsetOut = AddOffset;
2173
2174 return true;
2175}
2176
2177// Try to transform:
2178//
2179// for(i = Start; i < End; ++i)
2180// Rem = (i nuw+ IncrLoopInvariant) u% RemAmtLoopInvariant;
2181//
2182// ->
2183//
2184// Rem = (Start nuw+ IncrLoopInvariant) % RemAmtLoopInvariant;
2185// for(i = Start; i < End; ++i, ++rem)
2186// Rem = rem == RemAmtLoopInvariant ? 0 : Rem;
2187static bool foldURemOfLoopIncrement(Instruction *Rem, const DataLayout *DL,
2188 const LoopInfo *LI,
2189 SmallPtrSet<BasicBlock *, 32> &FreshBBs,
2190 bool IsHuge) {
2191 Value *AddOffset, *RemAmt, *AddInst;
2192 PHINode *LoopIncrPN;
2193 if (!isRemOfLoopIncrementWithLoopInvariant(Rem, LI, RemAmtOut&: RemAmt, AddInstOut&: AddInst,
2194 AddOffsetOut&: AddOffset, LoopIncrPNOut&: LoopIncrPN))
2195 return false;
2196
2197 // Only non-constant remainder as the extra IV is probably not profitable
2198 // in that case.
2199 //
2200 // Potential TODO(1): `urem` of a const ends up as `mul` + `shift` + `add`. If
2201 // we can rule out register pressure and ensure this `urem` is executed each
2202 // iteration, its probably profitable to handle the const case as well.
2203 //
2204 // Potential TODO(2): Should we have a check for how "nested" this remainder
2205 // operation is? The new code runs every iteration so if the remainder is
2206 // guarded behind unlikely conditions this might not be worth it.
2207 if (match(V: RemAmt, P: m_ImmConstant()))
2208 return false;
2209
2210 Loop *L = LI->getLoopFor(BB: LoopIncrPN->getParent());
2211 Value *Start = LoopIncrPN->getIncomingValueForBlock(BB: L->getLoopPreheader());
2212 // If we have add create initial value for remainder.
2213 // The logic here is:
2214 // (urem (add nuw Start, IncrLoopInvariant), RemAmtLoopInvariant
2215 //
2216 // Only proceed if the expression simplifies (otherwise we can't fully
2217 // optimize out the urem).
2218 if (AddInst) {
2219 assert(AddOffset && "We found an add but missing values");
2220 // Without dom-condition/assumption cache we aren't likely to get much out
2221 // of a context instruction.
2222 Start = simplifyAddInst(LHS: Start, RHS: AddOffset,
2223 IsNSW: match(V: AddInst, P: m_NSWAdd(L: m_Value(), R: m_Value())),
2224 /*IsNUW=*/true, Q: *DL);
2225 if (!Start)
2226 return false;
2227 }
2228
2229 // If we can't fully optimize out the `rem`, skip this transform.
2230 Start = simplifyURemInst(LHS: Start, RHS: RemAmt, Q: *DL);
2231 if (!Start)
2232 return false;
2233
2234 // Create new remainder with induction variable.
2235 Type *Ty = Rem->getType();
2236 IRBuilder<> Builder(Rem->getContext());
2237
2238 Builder.SetInsertPoint(LoopIncrPN);
2239 PHINode *NewRem = Builder.CreatePHI(Ty, NumReservedValues: 2);
2240
2241 Builder.SetInsertPoint(cast<Instruction>(
2242 Val: LoopIncrPN->getIncomingValueForBlock(BB: L->getLoopLatch())));
2243 // `(add (urem x, y), 1)` is always nuw.
2244 Value *RemAdd = Builder.CreateNUWAdd(LHS: NewRem, RHS: ConstantInt::get(Ty, V: 1));
2245 Value *RemCmp = Builder.CreateICmp(P: ICmpInst::ICMP_EQ, LHS: RemAdd, RHS: RemAmt);
2246 Value *RemSel =
2247 Builder.CreateSelect(C: RemCmp, True: Constant::getNullValue(Ty), False: RemAdd);
2248
2249 NewRem->addIncoming(V: Start, BB: L->getLoopPreheader());
2250 NewRem->addIncoming(V: RemSel, BB: L->getLoopLatch());
2251
2252 // Insert all touched BBs.
2253 FreshBBs.insert(Ptr: LoopIncrPN->getParent());
2254 FreshBBs.insert(Ptr: L->getLoopLatch());
2255 FreshBBs.insert(Ptr: Rem->getParent());
2256 if (AddInst)
2257 FreshBBs.insert(Ptr: cast<Instruction>(Val: AddInst)->getParent());
2258 replaceAllUsesWith(Old: Rem, New: NewRem, FreshBBs, IsHuge);
2259 Rem->eraseFromParent();
2260 if (AddInst && AddInst->use_empty())
2261 cast<Instruction>(Val: AddInst)->eraseFromParent();
2262 return true;
2263}
2264
2265bool CodeGenPrepare::optimizeURem(Instruction *Rem) {
2266 if (foldURemOfLoopIncrement(Rem, DL, LI, FreshBBs, IsHuge: IsHugeFunc))
2267 return true;
2268 return false;
2269}
2270
2271bool CodeGenPrepare::optimizeCmp(CmpInst *Cmp, ModifyDT &ModifiedDT) {
2272 if (sinkCmpExpression(Cmp, TLI: *TLI, DL: *DL))
2273 return true;
2274
2275 if (combineToUAddWithOverflow(Cmp, ModifiedDT))
2276 return true;
2277
2278 if (combineToUSubWithOverflow(Cmp, ModifiedDT))
2279 return true;
2280
2281 if (unfoldPowerOf2Test(Cmp))
2282 return true;
2283
2284 if (foldICmpWithDominatingICmp(Cmp, TLI: *TLI))
2285 return true;
2286
2287 if (swapICmpOperandsToExposeCSEOpportunities(Cmp))
2288 return true;
2289
2290 if (foldFCmpToFPClassTest(Cmp, TLI: *TLI, DL: *DL))
2291 return true;
2292
2293 return false;
2294}
2295
2296/// Duplicate and sink the given 'and' instruction into user blocks where it is
2297/// used in a compare to allow isel to generate better code for targets where
2298/// this operation can be combined.
2299///
2300/// Return true if any changes are made.
2301static bool sinkAndCmp0Expression(Instruction *AndI, const TargetLowering &TLI,
2302 SetOfInstrs &InsertedInsts) {
2303 // Double-check that we're not trying to optimize an instruction that was
2304 // already optimized by some other part of this pass.
2305 assert(!InsertedInsts.count(AndI) &&
2306 "Attempting to optimize already optimized and instruction");
2307 (void)InsertedInsts;
2308
2309 // Nothing to do for single use in same basic block.
2310 if (AndI->hasOneUse() &&
2311 AndI->getParent() == cast<Instruction>(Val: *AndI->user_begin())->getParent())
2312 return false;
2313
2314 // Try to avoid cases where sinking/duplicating is likely to increase register
2315 // pressure.
2316 if (!isa<ConstantInt>(Val: AndI->getOperand(i: 0)) &&
2317 !isa<ConstantInt>(Val: AndI->getOperand(i: 1)) &&
2318 AndI->getOperand(i: 0)->hasOneUse() && AndI->getOperand(i: 1)->hasOneUse())
2319 return false;
2320
2321 for (auto *U : AndI->users()) {
2322 Instruction *User = cast<Instruction>(Val: U);
2323
2324 // Only sink 'and' feeding icmp with 0.
2325 if (!isa<ICmpInst>(Val: User))
2326 return false;
2327
2328 auto *CmpC = dyn_cast<ConstantInt>(Val: User->getOperand(i: 1));
2329 if (!CmpC || !CmpC->isZero())
2330 return false;
2331 }
2332
2333 if (!TLI.isMaskAndCmp0FoldingBeneficial(AndI: *AndI))
2334 return false;
2335
2336 LLVM_DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n");
2337 LLVM_DEBUG(AndI->getParent()->dump());
2338
2339 // Push the 'and' into the same block as the icmp 0. There should only be
2340 // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
2341 // others, so we don't need to keep track of which BBs we insert into.
2342 for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
2343 UI != E;) {
2344 Use &TheUse = UI.getUse();
2345 Instruction *User = cast<Instruction>(Val: *UI);
2346
2347 // Preincrement use iterator so we don't invalidate it.
2348 ++UI;
2349
2350 LLVM_DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n");
2351
2352 // Keep the 'and' in the same place if the use is already in the same block.
2353 Instruction *InsertPt =
2354 User->getParent() == AndI->getParent() ? AndI : User;
2355 Instruction *InsertedAnd = BinaryOperator::Create(
2356 Op: Instruction::And, S1: AndI->getOperand(i: 0), S2: AndI->getOperand(i: 1), Name: "",
2357 InsertBefore: InsertPt->getIterator());
2358 // Propagate the debug info.
2359 InsertedAnd->setDebugLoc(AndI->getDebugLoc());
2360
2361 // Replace a use of the 'and' with a use of the new 'and'.
2362 TheUse = InsertedAnd;
2363 ++NumAndUses;
2364 LLVM_DEBUG(User->getParent()->dump());
2365 }
2366
2367 // We removed all uses, nuke the and.
2368 AndI->eraseFromParent();
2369 return true;
2370}
2371
2372/// Check if the candidates could be combined with a shift instruction, which
2373/// includes:
2374/// 1. Truncate instruction
2375/// 2. And instruction and the imm is a mask of the low bits:
2376/// imm & (imm+1) == 0
2377static bool isExtractBitsCandidateUse(Instruction *User) {
2378 if (!isa<TruncInst>(Val: User)) {
2379 if (User->getOpcode() != Instruction::And ||
2380 !isa<ConstantInt>(Val: User->getOperand(i: 1)))
2381 return false;
2382
2383 const APInt &Cimm = cast<ConstantInt>(Val: User->getOperand(i: 1))->getValue();
2384
2385 if ((Cimm & (Cimm + 1)).getBoolValue())
2386 return false;
2387 }
2388 return true;
2389}
2390
2391/// Sink both shift and truncate instruction to the use of truncate's BB.
2392static bool
2393SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
2394 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
2395 const TargetLowering &TLI, const DataLayout &DL) {
2396 BasicBlock *UserBB = User->getParent();
2397 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
2398 auto *TruncI = cast<TruncInst>(Val: User);
2399 bool MadeChange = false;
2400
2401 for (Value::user_iterator TruncUI = TruncI->user_begin(),
2402 TruncE = TruncI->user_end();
2403 TruncUI != TruncE;) {
2404
2405 Use &TruncTheUse = TruncUI.getUse();
2406 Instruction *TruncUser = cast<Instruction>(Val: *TruncUI);
2407 // Preincrement use iterator so we don't invalidate it.
2408
2409 ++TruncUI;
2410
2411 int ISDOpcode = TLI.InstructionOpcodeToISD(Opcode: TruncUser->getOpcode());
2412 if (!ISDOpcode)
2413 continue;
2414
2415 // If the use is actually a legal node, there will not be an
2416 // implicit truncate.
2417 // FIXME: always querying the result type is just an
2418 // approximation; some nodes' legality is determined by the
2419 // operand or other means. There's no good way to find out though.
2420 if (TLI.isOperationLegalOrCustom(
2421 Op: ISDOpcode, VT: TLI.getValueType(DL, Ty: TruncUser->getType(), AllowUnknown: true)))
2422 continue;
2423
2424 // Don't bother for PHI nodes.
2425 if (isa<PHINode>(Val: TruncUser))
2426 continue;
2427
2428 BasicBlock *TruncUserBB = TruncUser->getParent();
2429
2430 if (UserBB == TruncUserBB)
2431 continue;
2432
2433 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
2434 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
2435
2436 if (!InsertedShift && !InsertedTrunc) {
2437 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
2438 assert(InsertPt != TruncUserBB->end());
2439 // Sink the shift
2440 if (ShiftI->getOpcode() == Instruction::AShr)
2441 InsertedShift =
2442 BinaryOperator::CreateAShr(V1: ShiftI->getOperand(i_nocapture: 0), V2: CI, Name: "");
2443 else
2444 InsertedShift =
2445 BinaryOperator::CreateLShr(V1: ShiftI->getOperand(i_nocapture: 0), V2: CI, Name: "");
2446 InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
2447 InsertedShift->insertBefore(BB&: *TruncUserBB, InsertPos: InsertPt);
2448
2449 // Sink the trunc
2450 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
2451 TruncInsertPt++;
2452 // It will go ahead of any debug-info.
2453 TruncInsertPt.setHeadBit(true);
2454 assert(TruncInsertPt != TruncUserBB->end());
2455
2456 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), S: InsertedShift,
2457 Ty: TruncI->getType(), Name: "");
2458 InsertedTrunc->insertBefore(BB&: *TruncUserBB, InsertPos: TruncInsertPt);
2459 InsertedTrunc->setDebugLoc(TruncI->getDebugLoc());
2460
2461 MadeChange = true;
2462
2463 TruncTheUse = InsertedTrunc;
2464 }
2465 }
2466 return MadeChange;
2467}
2468
2469/// Sink the shift *right* instruction into user blocks if the uses could
2470/// potentially be combined with this shift instruction and generate BitExtract
2471/// instruction. It will only be applied if the architecture supports BitExtract
2472/// instruction. Here is an example:
2473/// BB1:
2474/// %x.extract.shift = lshr i64 %arg1, 32
2475/// BB2:
2476/// %x.extract.trunc = trunc i64 %x.extract.shift to i16
2477/// ==>
2478///
2479/// BB2:
2480/// %x.extract.shift.1 = lshr i64 %arg1, 32
2481/// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
2482///
2483/// CodeGen will recognize the pattern in BB2 and generate BitExtract
2484/// instruction.
2485/// Return true if any changes are made.
2486static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
2487 const TargetLowering &TLI,
2488 const DataLayout &DL) {
2489 BasicBlock *DefBB = ShiftI->getParent();
2490
2491 /// Only insert instructions in each block once.
2492 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
2493
2494 bool shiftIsLegal = TLI.isTypeLegal(VT: TLI.getValueType(DL, Ty: ShiftI->getType()));
2495
2496 bool MadeChange = false;
2497 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
2498 UI != E;) {
2499 Use &TheUse = UI.getUse();
2500 Instruction *User = cast<Instruction>(Val: *UI);
2501 // Preincrement use iterator so we don't invalidate it.
2502 ++UI;
2503
2504 // Don't bother for PHI nodes.
2505 if (isa<PHINode>(Val: User))
2506 continue;
2507
2508 if (!isExtractBitsCandidateUse(User))
2509 continue;
2510
2511 BasicBlock *UserBB = User->getParent();
2512
2513 if (UserBB == DefBB) {
2514 // If the shift and truncate instruction are in the same BB. The use of
2515 // the truncate(TruncUse) may still introduce another truncate if not
2516 // legal. In this case, we would like to sink both shift and truncate
2517 // instruction to the BB of TruncUse.
2518 // for example:
2519 // BB1:
2520 // i64 shift.result = lshr i64 opnd, imm
2521 // trunc.result = trunc shift.result to i16
2522 //
2523 // BB2:
2524 // ----> We will have an implicit truncate here if the architecture does
2525 // not have i16 compare.
2526 // cmp i16 trunc.result, opnd2
2527 //
2528 if (isa<TruncInst>(Val: User) &&
2529 shiftIsLegal
2530 // If the type of the truncate is legal, no truncate will be
2531 // introduced in other basic blocks.
2532 && (!TLI.isTypeLegal(VT: TLI.getValueType(DL, Ty: User->getType()))))
2533 MadeChange =
2534 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
2535
2536 continue;
2537 }
2538 // If we have already inserted a shift into this block, use it.
2539 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
2540
2541 if (!InsertedShift) {
2542 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2543 assert(InsertPt != UserBB->end());
2544
2545 if (ShiftI->getOpcode() == Instruction::AShr)
2546 InsertedShift =
2547 BinaryOperator::CreateAShr(V1: ShiftI->getOperand(i_nocapture: 0), V2: CI, Name: "");
2548 else
2549 InsertedShift =
2550 BinaryOperator::CreateLShr(V1: ShiftI->getOperand(i_nocapture: 0), V2: CI, Name: "");
2551 InsertedShift->insertBefore(BB&: *UserBB, InsertPos: InsertPt);
2552 InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
2553
2554 MadeChange = true;
2555 }
2556
2557 // Replace a use of the shift with a use of the new shift.
2558 TheUse = InsertedShift;
2559 }
2560
2561 // If we removed all uses, or there are none, nuke the shift.
2562 if (ShiftI->use_empty()) {
2563 salvageDebugInfo(I&: *ShiftI);
2564 ShiftI->eraseFromParent();
2565 MadeChange = true;
2566 }
2567
2568 return MadeChange;
2569}
2570
2571/// If counting leading or trailing zeros is an expensive operation and a zero
2572/// input is defined, add a check for zero to avoid calling the intrinsic.
2573///
2574/// We want to transform:
2575/// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
2576///
2577/// into:
2578/// entry:
2579/// %cmpz = icmp eq i64 %A, 0
2580/// br i1 %cmpz, label %cond.end, label %cond.false
2581/// cond.false:
2582/// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
2583/// br label %cond.end
2584/// cond.end:
2585/// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
2586///
2587/// If the transform is performed, return true and set ModifiedDT to true.
2588static bool despeculateCountZeros(IntrinsicInst *CountZeros,
2589 DomTreeUpdater *DTU, LoopInfo *LI,
2590 const TargetLowering *TLI,
2591 const DataLayout *DL, ModifyDT &ModifiedDT,
2592 SmallPtrSet<BasicBlock *, 32> &FreshBBs,
2593 bool IsHugeFunc) {
2594 // If a zero input is undefined, it doesn't make sense to despeculate that.
2595 if (match(V: CountZeros->getOperand(i_nocapture: 1), P: m_One()))
2596 return false;
2597
2598 // If it's cheap to speculate, there's nothing to do.
2599 Type *Ty = CountZeros->getType();
2600 auto IntrinsicID = CountZeros->getIntrinsicID();
2601 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz(Ty)) ||
2602 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz(Ty)))
2603 return false;
2604
2605 // Only handle scalar cases. Anything else requires too much work.
2606 unsigned SizeInBits = Ty->getScalarSizeInBits();
2607 if (Ty->isVectorTy())
2608 return false;
2609
2610 // Bail if the value is never zero.
2611 Use &Op = CountZeros->getOperandUse(i: 0);
2612 if (isKnownNonZero(V: Op, Q: *DL))
2613 return false;
2614
2615 // The intrinsic will be sunk behind a compare against zero and branch.
2616 BasicBlock *StartBlock = CountZeros->getParent();
2617 BasicBlock *CallBlock = SplitBlock(Old: StartBlock, SplitPt: CountZeros, DTU, LI,
2618 /* MSSAU */ nullptr, BBName: "cond.false");
2619 if (IsHugeFunc)
2620 FreshBBs.insert(Ptr: CallBlock);
2621
2622 // Create another block after the count zero intrinsic. A PHI will be added
2623 // in this block to select the result of the intrinsic or the bit-width
2624 // constant if the input to the intrinsic is zero.
2625 BasicBlock::iterator SplitPt = std::next(x: BasicBlock::iterator(CountZeros));
2626 // Any debug-info after CountZeros should not be included.
2627 SplitPt.setHeadBit(true);
2628 BasicBlock *EndBlock = SplitBlock(Old: CallBlock, SplitPt: &*SplitPt, DTU, LI,
2629 /* MSSAU */ nullptr, BBName: "cond.end");
2630 if (IsHugeFunc)
2631 FreshBBs.insert(Ptr: EndBlock);
2632
2633 // Set up a builder to create a compare, conditional branch, and PHI.
2634 IRBuilder<> Builder(CountZeros->getContext());
2635 Builder.SetInsertPoint(StartBlock->getTerminator());
2636 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
2637
2638 // Replace the unconditional branch that was created by the first split with
2639 // a compare against zero and a conditional branch.
2640 Value *Zero = Constant::getNullValue(Ty);
2641 // Avoid introducing branch on poison. This also replaces the ctz operand.
2642 if (!isGuaranteedNotToBeUndefOrPoison(V: Op))
2643 Op = Builder.CreateFreeze(V: Op, Name: Op->getName() + ".fr");
2644 Value *Cmp = Builder.CreateICmpEQ(LHS: Op, RHS: Zero, Name: "cmpz");
2645 Builder.CreateCondBr(Cond: Cmp, True: EndBlock, False: CallBlock);
2646 StartBlock->getTerminator()->eraseFromParent();
2647 DTU->applyUpdates(Updates: {{DominatorTree::Insert, StartBlock, EndBlock}});
2648
2649 // Create a PHI in the end block to select either the output of the intrinsic
2650 // or the bit width of the operand.
2651 Builder.SetInsertPoint(TheBB: EndBlock, IP: EndBlock->begin());
2652 PHINode *PN = Builder.CreatePHI(Ty, NumReservedValues: 2, Name: "ctz");
2653 replaceAllUsesWith(Old: CountZeros, New: PN, FreshBBs, IsHuge: IsHugeFunc);
2654 Value *BitWidth = Builder.getInt(AI: APInt(SizeInBits, SizeInBits));
2655 PN->addIncoming(V: BitWidth, BB: StartBlock);
2656 PN->addIncoming(V: CountZeros, BB: CallBlock);
2657
2658 // We are explicitly handling the zero case, so we can set the intrinsic's
2659 // undefined zero argument to 'true'. This will also prevent reprocessing the
2660 // intrinsic; we only despeculate when a zero input is defined.
2661 CountZeros->setArgOperand(i: 1, v: Builder.getTrue());
2662 ModifiedDT = ModifyDT::ModifyBBDT;
2663 return true;
2664}
2665
2666bool CodeGenPrepare::optimizeCallInst(CallInst *CI, ModifyDT &ModifiedDT) {
2667 BasicBlock *BB = CI->getParent();
2668
2669 // Sink address computing for memory operands into the block.
2670 if (CI->isInlineAsm() && optimizeInlineAsmInst(CS: CI))
2671 return true;
2672
2673 // Align the pointer arguments to this call if the target thinks it's a good
2674 // idea
2675 unsigned MinSize;
2676 Align PrefAlign;
2677 if (TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
2678 for (auto &Arg : CI->args()) {
2679 // We want to align both objects whose address is used directly and
2680 // objects whose address is used in casts and GEPs, though it only makes
2681 // sense for GEPs if the offset is a multiple of the desired alignment and
2682 // if size - offset meets the size threshold.
2683 if (!Arg->getType()->isPointerTy())
2684 continue;
2685 APInt Offset(DL->getIndexSizeInBits(
2686 AS: cast<PointerType>(Val: Arg->getType())->getAddressSpace()),
2687 0);
2688 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(DL: *DL, Offset);
2689 uint64_t Offset2 = Offset.getLimitedValue();
2690 if (!isAligned(Lhs: PrefAlign, SizeInBytes: Offset2))
2691 continue;
2692 AllocaInst *AI;
2693 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlign() < PrefAlign) {
2694 std::optional<TypeSize> AllocaSize = AI->getAllocationSize(DL: *DL);
2695 if (AllocaSize && AllocaSize->getKnownMinValue() >= MinSize + Offset2)
2696 AI->setAlignment(PrefAlign);
2697 }
2698 // Global variables can only be aligned if they are defined in this
2699 // object (i.e. they are uniquely initialized in this object), and
2700 // over-aligning global variables that have an explicit section is
2701 // forbidden.
2702 GlobalVariable *GV;
2703 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
2704 GV->getPointerAlignment(DL: *DL) < PrefAlign &&
2705 GV->getGlobalSize(DL: *DL) >= MinSize + Offset2)
2706 GV->setAlignment(PrefAlign);
2707 }
2708 }
2709 // If this is a memcpy (or similar) then we may be able to improve the
2710 // alignment.
2711 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Val: CI)) {
2712 Align DestAlign = getKnownAlignment(V: MI->getDest(), DL: *DL);
2713 MaybeAlign MIDestAlign = MI->getDestAlign();
2714 if (!MIDestAlign || DestAlign > *MIDestAlign)
2715 MI->setDestAlignment(DestAlign);
2716 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(Val: MI)) {
2717 MaybeAlign MTISrcAlign = MTI->getSourceAlign();
2718 Align SrcAlign = getKnownAlignment(V: MTI->getSource(), DL: *DL);
2719 if (!MTISrcAlign || SrcAlign > *MTISrcAlign)
2720 MTI->setSourceAlignment(SrcAlign);
2721 }
2722 }
2723
2724 // If we have a cold call site, try to sink addressing computation into the
2725 // cold block. This interacts with our handling for loads and stores to
2726 // ensure that we can fold all uses of a potential addressing computation
2727 // into their uses. TODO: generalize this to work over profiling data
2728 if (CI->hasFnAttr(Kind: Attribute::Cold) &&
2729 !llvm::shouldOptimizeForSize(BB, PSI, BFI))
2730 for (auto &Arg : CI->args()) {
2731 if (!Arg->getType()->isPointerTy())
2732 continue;
2733 unsigned AS = Arg->getType()->getPointerAddressSpace();
2734 if (optimizeMemoryInst(MemoryInst: CI, Addr: Arg, AccessTy: Arg->getType(), AddrSpace: AS))
2735 return true;
2736 }
2737
2738 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: CI);
2739 if (II) {
2740 switch (II->getIntrinsicID()) {
2741 default:
2742 break;
2743 case Intrinsic::assume:
2744 llvm_unreachable("llvm.assume should have been removed already");
2745 case Intrinsic::allow_runtime_check:
2746 case Intrinsic::allow_ubsan_check:
2747 case Intrinsic::experimental_widenable_condition: {
2748 // Give up on future widening opportunities so that we can fold away dead
2749 // paths and merge blocks before going into block-local instruction
2750 // selection.
2751 if (II->use_empty()) {
2752 II->eraseFromParent();
2753 return true;
2754 }
2755 Constant *RetVal = ConstantInt::getTrue(Context&: II->getContext());
2756 resetIteratorIfInvalidatedWhileCalling(BB, f: [&]() {
2757 replaceAndRecursivelySimplify(I: CI, SimpleV: RetVal, TLI: TLInfo, DT: nullptr);
2758 });
2759 return true;
2760 }
2761 case Intrinsic::objectsize:
2762 llvm_unreachable("llvm.objectsize.* should have been lowered already");
2763 case Intrinsic::is_constant:
2764 llvm_unreachable("llvm.is.constant.* should have been lowered already");
2765 case Intrinsic::aarch64_stlxr:
2766 case Intrinsic::aarch64_stxr: {
2767 ZExtInst *ExtVal = dyn_cast<ZExtInst>(Val: CI->getArgOperand(i: 0));
2768 if (!ExtVal || !ExtVal->hasOneUse() ||
2769 ExtVal->getParent() == CI->getParent())
2770 return false;
2771 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
2772 ExtVal->moveBefore(InsertPos: CI->getIterator());
2773 // Mark this instruction as "inserted by CGP", so that other
2774 // optimizations don't touch it.
2775 InsertedInsts.insert(Ptr: ExtVal);
2776 return true;
2777 }
2778
2779 case Intrinsic::launder_invariant_group:
2780 case Intrinsic::strip_invariant_group: {
2781 Value *ArgVal = II->getArgOperand(i: 0);
2782 auto it = LargeOffsetGEPMap.find(Key: II);
2783 if (it != LargeOffsetGEPMap.end()) {
2784 // Merge entries in LargeOffsetGEPMap to reflect the RAUW.
2785 // Make sure not to have to deal with iterator invalidation
2786 // after possibly adding ArgVal to LargeOffsetGEPMap.
2787 auto GEPs = std::move(it->second);
2788 LargeOffsetGEPMap[ArgVal].append(in_start: GEPs.begin(), in_end: GEPs.end());
2789 LargeOffsetGEPMap.erase(Key: II);
2790 }
2791
2792 replaceAllUsesWith(Old: II, New: ArgVal, FreshBBs, IsHuge: IsHugeFunc);
2793 II->eraseFromParent();
2794 return true;
2795 }
2796 case Intrinsic::cttz:
2797 case Intrinsic::ctlz:
2798 // If counting zeros is expensive, try to avoid it.
2799 return despeculateCountZeros(CountZeros: II, DTU, LI, TLI, DL, ModifiedDT, FreshBBs,
2800 IsHugeFunc);
2801 case Intrinsic::fshl:
2802 case Intrinsic::fshr:
2803 return optimizeFunnelShift(Fsh: II);
2804 case Intrinsic::masked_gather:
2805 return optimizeGatherScatterInst(MemoryInst: II, Ptr: II->getArgOperand(i: 0));
2806 case Intrinsic::masked_scatter:
2807 return optimizeGatherScatterInst(MemoryInst: II, Ptr: II->getArgOperand(i: 1));
2808 case Intrinsic::masked_load:
2809 // Treat v1X masked load as load X type.
2810 if (auto *VT = dyn_cast<FixedVectorType>(Val: II->getType())) {
2811 if (VT->getNumElements() == 1) {
2812 Value *PtrVal = II->getArgOperand(i: 0);
2813 unsigned AS = PtrVal->getType()->getPointerAddressSpace();
2814 if (optimizeMemoryInst(MemoryInst: II, Addr: PtrVal, AccessTy: VT->getElementType(), AddrSpace: AS))
2815 return true;
2816 }
2817 }
2818 return false;
2819 case Intrinsic::masked_store:
2820 // Treat v1X masked store as store X type.
2821 if (auto *VT =
2822 dyn_cast<FixedVectorType>(Val: II->getArgOperand(i: 0)->getType())) {
2823 if (VT->getNumElements() == 1) {
2824 Value *PtrVal = II->getArgOperand(i: 1);
2825 unsigned AS = PtrVal->getType()->getPointerAddressSpace();
2826 if (optimizeMemoryInst(MemoryInst: II, Addr: PtrVal, AccessTy: VT->getElementType(), AddrSpace: AS))
2827 return true;
2828 }
2829 }
2830 return false;
2831 case Intrinsic::umul_with_overflow:
2832 return optimizeMulWithOverflow(I: II, /*IsSigned=*/false, ModifiedDT);
2833 case Intrinsic::smul_with_overflow:
2834 return optimizeMulWithOverflow(I: II, /*IsSigned=*/true, ModifiedDT);
2835 }
2836
2837 SmallVector<Value *, 2> PtrOps;
2838 Type *AccessTy;
2839 if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
2840 while (!PtrOps.empty()) {
2841 Value *PtrVal = PtrOps.pop_back_val();
2842 unsigned AS = PtrVal->getType()->getPointerAddressSpace();
2843 if (optimizeMemoryInst(MemoryInst: II, Addr: PtrVal, AccessTy, AddrSpace: AS))
2844 return true;
2845 }
2846 }
2847
2848 // From here on out we're working with named functions.
2849 auto *Callee = CI->getCalledFunction();
2850 if (!Callee)
2851 return false;
2852
2853 // Lower all default uses of _chk calls. This is very similar
2854 // to what InstCombineCalls does, but here we are only lowering calls
2855 // to fortified library functions (e.g. __memcpy_chk) that have the default
2856 // "don't know" as the objectsize. Anything else should be left alone.
2857 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
2858 IRBuilder<> Builder(CI);
2859 if (Value *V = Simplifier.optimizeCall(CI, B&: Builder)) {
2860 replaceAllUsesWith(Old: CI, New: V, FreshBBs, IsHuge: IsHugeFunc);
2861 CI->eraseFromParent();
2862 return true;
2863 }
2864
2865 // SCCP may have propagated, among other things, C++ static variables across
2866 // calls. If this happens to be the case, we may want to undo it in order to
2867 // avoid redundant pointer computation of the constant, as the function method
2868 // returning the constant needs to be executed anyways.
2869 auto GetUniformReturnValue = [](const Function *F) -> GlobalVariable * {
2870 if (!F->getReturnType()->isPointerTy())
2871 return nullptr;
2872
2873 GlobalVariable *UniformValue = nullptr;
2874 for (auto &BB : *F) {
2875 if (auto *RI = dyn_cast<ReturnInst>(Val: BB.getTerminator())) {
2876 if (auto *V = dyn_cast<GlobalVariable>(Val: RI->getReturnValue())) {
2877 if (!UniformValue)
2878 UniformValue = V;
2879 else if (V != UniformValue)
2880 return nullptr;
2881 } else {
2882 return nullptr;
2883 }
2884 }
2885 }
2886
2887 return UniformValue;
2888 };
2889
2890 if (Callee->hasExactDefinition()) {
2891 if (GlobalVariable *RV = GetUniformReturnValue(Callee)) {
2892 bool MadeChange = false;
2893 for (Use &U : make_early_inc_range(Range: RV->uses())) {
2894 auto *I = dyn_cast<Instruction>(Val: U.getUser());
2895 if (!I || I->getParent() != CI->getParent()) {
2896 // Limit to the same basic block to avoid extending the call-site live
2897 // range, which otherwise could increase register pressure.
2898 continue;
2899 }
2900 if (CI->comesBefore(Other: I)) {
2901 U.set(CI);
2902 MadeChange = true;
2903 }
2904 }
2905
2906 return MadeChange;
2907 }
2908 }
2909
2910 return false;
2911}
2912
2913static bool isIntrinsicOrLFToBeTailCalled(const TargetLibraryInfo *TLInfo,
2914 const CallInst *CI) {
2915 assert(CI && CI->use_empty());
2916
2917 if (const auto *II = dyn_cast<IntrinsicInst>(Val: CI))
2918 switch (II->getIntrinsicID()) {
2919 case Intrinsic::memset:
2920 case Intrinsic::memcpy:
2921 case Intrinsic::memmove:
2922 return true;
2923 default:
2924 return false;
2925 }
2926
2927 LibFunc LF;
2928 Function *Callee = CI->getCalledFunction();
2929 if (Callee && TLInfo && TLInfo->getLibFunc(FDecl: *Callee, F&: LF))
2930 switch (LF) {
2931 case LibFunc_strcpy:
2932 case LibFunc_strncpy:
2933 case LibFunc_strcat:
2934 case LibFunc_strncat:
2935 return true;
2936 default:
2937 return false;
2938 }
2939
2940 return false;
2941}
2942
2943/// Look for opportunities to duplicate return instructions to the predecessor
2944/// to enable tail call optimizations. The case it is currently looking for is
2945/// the following one. Known intrinsics or library function that may be tail
2946/// called are taken into account as well.
2947/// @code
2948/// bb0:
2949/// %tmp0 = tail call i32 @f0()
2950/// br label %return
2951/// bb1:
2952/// %tmp1 = tail call i32 @f1()
2953/// br label %return
2954/// bb2:
2955/// %tmp2 = tail call i32 @f2()
2956/// br label %return
2957/// return:
2958/// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
2959/// ret i32 %retval
2960/// @endcode
2961///
2962/// =>
2963///
2964/// @code
2965/// bb0:
2966/// %tmp0 = tail call i32 @f0()
2967/// ret i32 %tmp0
2968/// bb1:
2969/// %tmp1 = tail call i32 @f1()
2970/// ret i32 %tmp1
2971/// bb2:
2972/// %tmp2 = tail call i32 @f2()
2973/// ret i32 %tmp2
2974/// @endcode
2975bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB,
2976 ModifyDT &ModifiedDT) {
2977 if (!BB->getTerminator())
2978 return false;
2979
2980 ReturnInst *RetI = dyn_cast<ReturnInst>(Val: BB->getTerminator());
2981 if (!RetI)
2982 return false;
2983
2984 assert(LI->getLoopFor(BB) == nullptr && "A return block cannot be in a loop");
2985
2986 PHINode *PN = nullptr;
2987 ExtractValueInst *EVI = nullptr;
2988 BitCastInst *BCI = nullptr;
2989 Value *V = RetI->getReturnValue();
2990 if (V) {
2991 BCI = dyn_cast<BitCastInst>(Val: V);
2992 if (BCI)
2993 V = BCI->getOperand(i_nocapture: 0);
2994
2995 EVI = dyn_cast<ExtractValueInst>(Val: V);
2996 if (EVI) {
2997 V = EVI->getOperand(i_nocapture: 0);
2998 if (!llvm::all_of(Range: EVI->indices(), P: equal_to(Arg: 0)))
2999 return false;
3000 }
3001
3002 PN = dyn_cast<PHINode>(Val: V);
3003 }
3004
3005 if (PN && PN->getParent() != BB)
3006 return false;
3007
3008 auto isLifetimeEndOrBitCastFor = [](const Instruction *Inst) {
3009 const BitCastInst *BC = dyn_cast<BitCastInst>(Val: Inst);
3010 if (BC && BC->hasOneUse())
3011 Inst = BC->user_back();
3012
3013 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: Inst))
3014 return II->getIntrinsicID() == Intrinsic::lifetime_end;
3015 return false;
3016 };
3017
3018 SmallVector<const IntrinsicInst *, 4> FakeUses;
3019
3020 auto isFakeUse = [&FakeUses](const Instruction *Inst) {
3021 if (auto *II = dyn_cast<IntrinsicInst>(Val: Inst);
3022 II && II->getIntrinsicID() == Intrinsic::fake_use) {
3023 // Record the instruction so it can be preserved when the exit block is
3024 // removed. Do not preserve the fake use that uses the result of the
3025 // PHI instruction.
3026 // Do not copy fake uses that use the result of a PHI node.
3027 // FIXME: If we do want to copy the fake use into the return blocks, we
3028 // have to figure out which of the PHI node operands to use for each
3029 // copy.
3030 if (!isa<PHINode>(Val: II->getOperand(i_nocapture: 0))) {
3031 FakeUses.push_back(Elt: II);
3032 }
3033 return true;
3034 }
3035
3036 return false;
3037 };
3038
3039 // Make sure there are no instructions between the first instruction
3040 // and return.
3041 BasicBlock::const_iterator BI = BB->getFirstNonPHIIt();
3042 // Skip over pseudo-probes and the bitcast.
3043 while (&*BI == BCI || &*BI == EVI || isa<PseudoProbeInst>(Val: BI) ||
3044 isLifetimeEndOrBitCastFor(&*BI) || isFakeUse(&*BI))
3045 BI = std::next(x: BI);
3046 if (&*BI != RetI)
3047 return false;
3048
3049 // Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
3050 // call.
3051 auto MayBePermittedAsTailCall = [&](const auto *CI) {
3052 return TLI->mayBeEmittedAsTailCall(CI) &&
3053 attributesPermitTailCall(BB->getParent(), CI, RetI, *TLI);
3054 };
3055
3056 SmallVector<BasicBlock *, 4> TailCallBBs;
3057 // Record the call instructions so we can insert any fake uses
3058 // that need to be preserved before them.
3059 SmallVector<CallInst *, 4> CallInsts;
3060 if (PN) {
3061 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
3062 // Look through bitcasts.
3063 Value *IncomingVal = PN->getIncomingValue(i: I)->stripPointerCasts();
3064 CallInst *CI = dyn_cast<CallInst>(Val: IncomingVal);
3065 BasicBlock *PredBB = PN->getIncomingBlock(i: I);
3066 // Make sure the phi value is indeed produced by the tail call.
3067 if (CI && CI->hasOneUse() && CI->getParent() == PredBB &&
3068 MayBePermittedAsTailCall(CI)) {
3069 TailCallBBs.push_back(Elt: PredBB);
3070 CallInsts.push_back(Elt: CI);
3071 } else {
3072 // Consider the cases in which the phi value is indirectly produced by
3073 // the tail call, for example when encountering memset(), memmove(),
3074 // strcpy(), whose return value may have been optimized out. In such
3075 // cases, the value needs to be the first function argument.
3076 //
3077 // bb0:
3078 // tail call void @llvm.memset.p0.i64(ptr %0, i8 0, i64 %1)
3079 // br label %return
3080 // return:
3081 // %phi = phi ptr [ %0, %bb0 ], [ %2, %entry ]
3082 if (PredBB && PredBB->getSingleSuccessor() == BB)
3083 CI = dyn_cast_or_null<CallInst>(
3084 Val: PredBB->getTerminator()->getPrevNode());
3085
3086 if (CI && CI->use_empty() &&
3087 isIntrinsicOrLFToBeTailCalled(TLInfo, CI) &&
3088 IncomingVal == CI->getArgOperand(i: 0) &&
3089 MayBePermittedAsTailCall(CI)) {
3090 TailCallBBs.push_back(Elt: PredBB);
3091 CallInsts.push_back(Elt: CI);
3092 }
3093 }
3094 }
3095 } else {
3096 SmallPtrSet<BasicBlock *, 4> VisitedBBs;
3097 for (BasicBlock *Pred : predecessors(BB)) {
3098 if (!VisitedBBs.insert(Ptr: Pred).second)
3099 continue;
3100 if (Instruction *I = Pred->rbegin()->getPrevNode()) {
3101 CallInst *CI = dyn_cast<CallInst>(Val: I);
3102 if (CI && CI->use_empty() && MayBePermittedAsTailCall(CI)) {
3103 // Either we return void or the return value must be the first
3104 // argument of a known intrinsic or library function.
3105 if (!V || isa<UndefValue>(Val: V) ||
3106 (isIntrinsicOrLFToBeTailCalled(TLInfo, CI) &&
3107 V == CI->getArgOperand(i: 0))) {
3108 TailCallBBs.push_back(Elt: Pred);
3109 CallInsts.push_back(Elt: CI);
3110 }
3111 }
3112 }
3113 }
3114 }
3115
3116 bool Changed = false;
3117 for (auto const &TailCallBB : TailCallBBs) {
3118 // Make sure the call instruction is followed by an unconditional branch to
3119 // the return block.
3120 UncondBrInst *BI = dyn_cast<UncondBrInst>(Val: TailCallBB->getTerminator());
3121 if (!BI || BI->getSuccessor() != BB)
3122 continue;
3123
3124 // Duplicate the return into TailCallBB.
3125 (void)FoldReturnIntoUncondBranch(RI: RetI, BB, Pred: TailCallBB, DTU);
3126 assert(!VerifyBFIUpdates ||
3127 BFI->getBlockFreq(BB) >= BFI->getBlockFreq(TailCallBB));
3128 BFI->setBlockFreq(BB,
3129 Freq: (BFI->getBlockFreq(BB) - BFI->getBlockFreq(BB: TailCallBB)));
3130 ModifiedDT = ModifyDT::ModifyBBDT;
3131 Changed = true;
3132 ++NumRetsDup;
3133 }
3134
3135 // If we eliminated all predecessors of the block, delete the block now.
3136 if (Changed && !BB->hasAddressTaken() && pred_empty(BB)) {
3137 // Copy the fake uses found in the original return block to all blocks
3138 // that contain tail calls.
3139 for (auto *CI : CallInsts) {
3140 for (auto const *FakeUse : FakeUses) {
3141 auto *ClonedInst = FakeUse->clone();
3142 ClonedInst->insertBefore(InsertPos: CI->getIterator());
3143 }
3144 }
3145 DTU->deleteBB(DelBB: BB);
3146 }
3147
3148 return Changed;
3149}
3150
3151//===----------------------------------------------------------------------===//
3152// Memory Optimization
3153//===----------------------------------------------------------------------===//
3154
3155namespace {
3156
3157/// This is an extended version of TargetLowering::AddrMode
3158/// which holds actual Value*'s for register values.
3159struct ExtAddrMode : public TargetLowering::AddrMode {
3160 Value *BaseReg = nullptr;
3161 Value *ScaledReg = nullptr;
3162 Value *OriginalValue = nullptr;
3163 bool InBounds = true;
3164
3165 enum FieldName {
3166 NoField = 0x00,
3167 BaseRegField = 0x01,
3168 BaseGVField = 0x02,
3169 BaseOffsField = 0x04,
3170 ScaledRegField = 0x08,
3171 ScaleField = 0x10,
3172 MultipleFields = 0xff
3173 };
3174
3175 ExtAddrMode() = default;
3176
3177 void print(raw_ostream &OS) const;
3178 void dump() const;
3179
3180 // Replace From in ExtAddrMode with To.
3181 // E.g., SExt insts may be promoted and deleted. We should replace them with
3182 // the promoted values.
3183 void replaceWith(Value *From, Value *To) {
3184 if (ScaledReg == From)
3185 ScaledReg = To;
3186 }
3187
3188 FieldName compare(const ExtAddrMode &other) {
3189 // First check that the types are the same on each field, as differing types
3190 // is something we can't cope with later on.
3191 if (BaseReg && other.BaseReg &&
3192 BaseReg->getType() != other.BaseReg->getType())
3193 return MultipleFields;
3194 if (BaseGV && other.BaseGV && BaseGV->getType() != other.BaseGV->getType())
3195 return MultipleFields;
3196 if (ScaledReg && other.ScaledReg &&
3197 ScaledReg->getType() != other.ScaledReg->getType())
3198 return MultipleFields;
3199
3200 // Conservatively reject 'inbounds' mismatches.
3201 if (InBounds != other.InBounds)
3202 return MultipleFields;
3203
3204 // Check each field to see if it differs.
3205 unsigned Result = NoField;
3206 if (BaseReg != other.BaseReg)
3207 Result |= BaseRegField;
3208 if (BaseGV != other.BaseGV)
3209 Result |= BaseGVField;
3210 if (BaseOffs != other.BaseOffs)
3211 Result |= BaseOffsField;
3212 if (ScaledReg != other.ScaledReg)
3213 Result |= ScaledRegField;
3214 // Don't count 0 as being a different scale, because that actually means
3215 // unscaled (which will already be counted by having no ScaledReg).
3216 if (Scale && other.Scale && Scale != other.Scale)
3217 Result |= ScaleField;
3218
3219 if (llvm::popcount(Value: Result) > 1)
3220 return MultipleFields;
3221 else
3222 return static_cast<FieldName>(Result);
3223 }
3224
3225 // An AddrMode is trivial if it involves no calculation i.e. it is just a base
3226 // with no offset.
3227 bool isTrivial() {
3228 // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
3229 // trivial if at most one of these terms is nonzero, except that BaseGV and
3230 // BaseReg both being zero actually means a null pointer value, which we
3231 // consider to be 'non-zero' here.
3232 return !BaseOffs && !Scale && !(BaseGV && BaseReg);
3233 }
3234
3235 Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
3236 switch (Field) {
3237 default:
3238 return nullptr;
3239 case BaseRegField:
3240 return BaseReg;
3241 case BaseGVField:
3242 return BaseGV;
3243 case ScaledRegField:
3244 return ScaledReg;
3245 case BaseOffsField:
3246 return ConstantInt::getSigned(Ty: IntPtrTy, V: BaseOffs);
3247 }
3248 }
3249
3250 void SetCombinedField(FieldName Field, Value *V,
3251 const SmallVectorImpl<ExtAddrMode> &AddrModes) {
3252 switch (Field) {
3253 default:
3254 llvm_unreachable("Unhandled fields are expected to be rejected earlier");
3255 break;
3256 case ExtAddrMode::BaseRegField:
3257 BaseReg = V;
3258 break;
3259 case ExtAddrMode::BaseGVField:
3260 // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
3261 // in the BaseReg field.
3262 assert(BaseReg == nullptr);
3263 BaseReg = V;
3264 BaseGV = nullptr;
3265 break;
3266 case ExtAddrMode::ScaledRegField:
3267 ScaledReg = V;
3268 // If we have a mix of scaled and unscaled addrmodes then we want scale
3269 // to be the scale and not zero.
3270 if (!Scale)
3271 for (const ExtAddrMode &AM : AddrModes)
3272 if (AM.Scale) {
3273 Scale = AM.Scale;
3274 break;
3275 }
3276 break;
3277 case ExtAddrMode::BaseOffsField:
3278 // The offset is no longer a constant, so it goes in ScaledReg with a
3279 // scale of 1.
3280 assert(ScaledReg == nullptr);
3281 ScaledReg = V;
3282 Scale = 1;
3283 BaseOffs = 0;
3284 break;
3285 }
3286 }
3287};
3288
3289#ifndef NDEBUG
3290static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
3291 AM.print(OS);
3292 return OS;
3293}
3294#endif
3295
3296#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3297void ExtAddrMode::print(raw_ostream &OS) const {
3298 bool NeedPlus = false;
3299 OS << "[";
3300 if (InBounds)
3301 OS << "inbounds ";
3302 if (BaseGV) {
3303 OS << "GV:";
3304 BaseGV->printAsOperand(OS, /*PrintType=*/false);
3305 NeedPlus = true;
3306 }
3307
3308 if (BaseOffs) {
3309 OS << (NeedPlus ? " + " : "") << BaseOffs;
3310 NeedPlus = true;
3311 }
3312
3313 if (BaseReg) {
3314 OS << (NeedPlus ? " + " : "") << "Base:";
3315 BaseReg->printAsOperand(OS, /*PrintType=*/false);
3316 NeedPlus = true;
3317 }
3318 if (Scale) {
3319 OS << (NeedPlus ? " + " : "") << Scale << "*";
3320 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
3321 }
3322
3323 OS << ']';
3324}
3325
3326LLVM_DUMP_METHOD void ExtAddrMode::dump() const {
3327 print(dbgs());
3328 dbgs() << '\n';
3329}
3330#endif
3331
3332} // end anonymous namespace
3333
3334namespace {
3335
3336/// This class provides transaction based operation on the IR.
3337/// Every change made through this class is recorded in the internal state and
3338/// can be undone (rollback) until commit is called.
3339/// CGP does not check if instructions could be speculatively executed when
3340/// moved. Preserving the original location would pessimize the debugging
3341/// experience, as well as negatively impact the quality of sample PGO.
3342class TypePromotionTransaction {
3343 /// This represents the common interface of the individual transaction.
3344 /// Each class implements the logic for doing one specific modification on
3345 /// the IR via the TypePromotionTransaction.
3346 class TypePromotionAction {
3347 protected:
3348 /// The Instruction modified.
3349 Instruction *Inst;
3350
3351 public:
3352 /// Constructor of the action.
3353 /// The constructor performs the related action on the IR.
3354 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
3355
3356 virtual ~TypePromotionAction() = default;
3357
3358 /// Undo the modification done by this action.
3359 /// When this method is called, the IR must be in the same state as it was
3360 /// before this action was applied.
3361 /// \pre Undoing the action works if and only if the IR is in the exact same
3362 /// state as it was directly after this action was applied.
3363 virtual void undo() = 0;
3364
3365 /// Advocate every change made by this action.
3366 /// When the results on the IR of the action are to be kept, it is important
3367 /// to call this function, otherwise hidden information may be kept forever.
3368 virtual void commit() {
3369 // Nothing to be done, this action is not doing anything.
3370 }
3371 };
3372
3373 /// Utility to remember the position of an instruction.
3374 class InsertionHandler {
3375 /// Position of an instruction.
3376 /// Either an instruction:
3377 /// - Is the first in a basic block: BB is used.
3378 /// - Has a previous instruction: PrevInst is used.
3379 struct {
3380 BasicBlock::iterator PrevInst;
3381 BasicBlock *BB;
3382 } Point;
3383 std::optional<DbgRecord::self_iterator> BeforeDbgRecord = std::nullopt;
3384
3385 /// Remember whether or not the instruction had a previous instruction.
3386 bool HasPrevInstruction;
3387
3388 public:
3389 /// Record the position of \p Inst.
3390 InsertionHandler(Instruction *Inst) {
3391 HasPrevInstruction = (Inst != &*(Inst->getParent()->begin()));
3392 BasicBlock *BB = Inst->getParent();
3393
3394 // Record where we would have to re-insert the instruction in the sequence
3395 // of DbgRecords, if we ended up reinserting.
3396 BeforeDbgRecord = Inst->getDbgReinsertionPosition();
3397
3398 if (HasPrevInstruction) {
3399 Point.PrevInst = std::prev(x: Inst->getIterator());
3400 } else {
3401 Point.BB = BB;
3402 }
3403 }
3404
3405 /// Insert \p Inst at the recorded position.
3406 void insert(Instruction *Inst) {
3407 if (HasPrevInstruction) {
3408 if (Inst->getParent())
3409 Inst->removeFromParent();
3410 Inst->insertAfter(InsertPos: Point.PrevInst);
3411 } else {
3412 BasicBlock::iterator Position = Point.BB->getFirstInsertionPt();
3413 if (Inst->getParent())
3414 Inst->moveBefore(BB&: *Point.BB, I: Position);
3415 else
3416 Inst->insertBefore(BB&: *Point.BB, InsertPos: Position);
3417 }
3418
3419 Inst->getParent()->reinsertInstInDbgRecords(I: Inst, Pos: BeforeDbgRecord);
3420 }
3421 };
3422
3423 /// Move an instruction before another.
3424 class InstructionMoveBefore : public TypePromotionAction {
3425 /// Original position of the instruction.
3426 InsertionHandler Position;
3427
3428 public:
3429 /// Move \p Inst before \p Before.
3430 InstructionMoveBefore(Instruction *Inst, BasicBlock::iterator Before)
3431 : TypePromotionAction(Inst), Position(Inst) {
3432 LLVM_DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before
3433 << "\n");
3434 Inst->moveBefore(InsertPos: Before);
3435 }
3436
3437 /// Move the instruction back to its original position.
3438 void undo() override {
3439 LLVM_DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
3440 Position.insert(Inst);
3441 }
3442 };
3443
3444 /// Set the operand of an instruction with a new value.
3445 class OperandSetter : public TypePromotionAction {
3446 /// Original operand of the instruction.
3447 Value *Origin;
3448
3449 /// Index of the modified instruction.
3450 unsigned Idx;
3451
3452 public:
3453 /// Set \p Idx operand of \p Inst with \p NewVal.
3454 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
3455 : TypePromotionAction(Inst), Idx(Idx) {
3456 LLVM_DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
3457 << "for:" << *Inst << "\n"
3458 << "with:" << *NewVal << "\n");
3459 Origin = Inst->getOperand(i: Idx);
3460 Inst->setOperand(i: Idx, Val: NewVal);
3461 }
3462
3463 /// Restore the original value of the instruction.
3464 void undo() override {
3465 LLVM_DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
3466 << "for: " << *Inst << "\n"
3467 << "with: " << *Origin << "\n");
3468 Inst->setOperand(i: Idx, Val: Origin);
3469 }
3470 };
3471
3472 /// Hide the operands of an instruction.
3473 /// Do as if this instruction was not using any of its operands.
3474 class OperandsHider : public TypePromotionAction {
3475 /// The list of original operands.
3476 SmallVector<Value *, 4> OriginalValues;
3477
3478 public:
3479 /// Remove \p Inst from the uses of the operands of \p Inst.
3480 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
3481 LLVM_DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
3482 unsigned NumOpnds = Inst->getNumOperands();
3483 OriginalValues.reserve(N: NumOpnds);
3484 for (unsigned It = 0; It < NumOpnds; ++It) {
3485 // Save the current operand.
3486 Value *Val = Inst->getOperand(i: It);
3487 OriginalValues.push_back(Elt: Val);
3488 // Set a dummy one.
3489 // We could use OperandSetter here, but that would imply an overhead
3490 // that we are not willing to pay.
3491 Inst->setOperand(i: It, Val: PoisonValue::get(T: Val->getType()));
3492 }
3493 }
3494
3495 /// Restore the original list of uses.
3496 void undo() override {
3497 LLVM_DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
3498 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
3499 Inst->setOperand(i: It, Val: OriginalValues[It]);
3500 }
3501 };
3502
3503 /// Build a truncate instruction.
3504 class TruncBuilder : public TypePromotionAction {
3505 Value *Val;
3506
3507 public:
3508 /// Build a truncate instruction of \p Opnd producing a \p Ty
3509 /// result.
3510 /// trunc Opnd to Ty.
3511 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
3512 IRBuilder<> Builder(Opnd);
3513 Builder.SetCurrentDebugLocation(DebugLoc());
3514 Val = Builder.CreateTrunc(V: Opnd, DestTy: Ty, Name: "promoted");
3515 LLVM_DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
3516 }
3517
3518 /// Get the built value.
3519 Value *getBuiltValue() { return Val; }
3520
3521 /// Remove the built instruction.
3522 void undo() override {
3523 LLVM_DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
3524 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3525 IVal->eraseFromParent();
3526 }
3527 };
3528
3529 /// Build a sign extension instruction.
3530 class SExtBuilder : public TypePromotionAction {
3531 Value *Val;
3532
3533 public:
3534 /// Build a sign extension instruction of \p Opnd producing a \p Ty
3535 /// result.
3536 /// sext Opnd to Ty.
3537 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3538 : TypePromotionAction(InsertPt) {
3539 IRBuilder<> Builder(InsertPt);
3540 Val = Builder.CreateSExt(V: Opnd, DestTy: Ty, Name: "promoted");
3541 LLVM_DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
3542 }
3543
3544 /// Get the built value.
3545 Value *getBuiltValue() { return Val; }
3546
3547 /// Remove the built instruction.
3548 void undo() override {
3549 LLVM_DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
3550 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3551 IVal->eraseFromParent();
3552 }
3553 };
3554
3555 /// Build a zero extension instruction.
3556 class ZExtBuilder : public TypePromotionAction {
3557 Value *Val;
3558
3559 public:
3560 /// Build a zero extension instruction of \p Opnd producing a \p Ty
3561 /// result.
3562 /// zext Opnd to Ty.
3563 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3564 : TypePromotionAction(InsertPt) {
3565 IRBuilder<> Builder(InsertPt);
3566 Builder.SetCurrentDebugLocation(DebugLoc());
3567 Val = Builder.CreateZExt(V: Opnd, DestTy: Ty, Name: "promoted");
3568 LLVM_DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
3569 }
3570
3571 /// Get the built value.
3572 Value *getBuiltValue() { return Val; }
3573
3574 /// Remove the built instruction.
3575 void undo() override {
3576 LLVM_DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
3577 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3578 IVal->eraseFromParent();
3579 }
3580 };
3581
3582 /// Mutate an instruction to another type.
3583 class TypeMutator : public TypePromotionAction {
3584 /// Record the original type.
3585 Type *OrigTy;
3586
3587 public:
3588 /// Mutate the type of \p Inst into \p NewTy.
3589 TypeMutator(Instruction *Inst, Type *NewTy)
3590 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
3591 LLVM_DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
3592 << "\n");
3593 Inst->mutateType(Ty: NewTy);
3594 }
3595
3596 /// Mutate the instruction back to its original type.
3597 void undo() override {
3598 LLVM_DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
3599 << "\n");
3600 Inst->mutateType(Ty: OrigTy);
3601 }
3602 };
3603
3604 /// Replace the uses of an instruction by another instruction.
3605 class UsesReplacer : public TypePromotionAction {
3606 /// Helper structure to keep track of the replaced uses.
3607 struct InstructionAndIdx {
3608 /// The instruction using the instruction.
3609 Instruction *Inst;
3610
3611 /// The index where this instruction is used for Inst.
3612 unsigned Idx;
3613
3614 InstructionAndIdx(Instruction *Inst, unsigned Idx)
3615 : Inst(Inst), Idx(Idx) {}
3616 };
3617
3618 /// Keep track of the original uses (pair Instruction, Index).
3619 SmallVector<InstructionAndIdx, 4> OriginalUses;
3620 /// Keep track of the debug users.
3621 SmallVector<DbgVariableRecord *, 1> DbgVariableRecords;
3622
3623 /// Keep track of the new value so that we can undo it by replacing
3624 /// instances of the new value with the original value.
3625 Value *New;
3626
3627 using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
3628
3629 public:
3630 /// Replace all the use of \p Inst by \p New.
3631 UsesReplacer(Instruction *Inst, Value *New)
3632 : TypePromotionAction(Inst), New(New) {
3633 LLVM_DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
3634 << "\n");
3635 // Record the original uses.
3636 for (Use &U : Inst->uses()) {
3637 Instruction *UserI = cast<Instruction>(Val: U.getUser());
3638 OriginalUses.push_back(Elt: InstructionAndIdx(UserI, U.getOperandNo()));
3639 }
3640 // Record the debug uses separately. They are not in the instruction's
3641 // use list, but they are replaced by RAUW.
3642 findDbgValues(V: Inst, DbgVariableRecords);
3643
3644 // Now, we can replace the uses.
3645 Inst->replaceAllUsesWith(V: New);
3646 }
3647
3648 /// Reassign the original uses of Inst to Inst.
3649 void undo() override {
3650 LLVM_DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
3651 for (InstructionAndIdx &Use : OriginalUses)
3652 Use.Inst->setOperand(i: Use.Idx, Val: Inst);
3653 // RAUW has replaced all original uses with references to the new value,
3654 // including the debug uses. Since we are undoing the replacements,
3655 // the original debug uses must also be reinstated to maintain the
3656 // correctness and utility of debug value records.
3657 for (DbgVariableRecord *DVR : DbgVariableRecords)
3658 DVR->replaceVariableLocationOp(OldValue: New, NewValue: Inst);
3659 }
3660 };
3661
3662 /// Remove an instruction from the IR.
3663 class InstructionRemover : public TypePromotionAction {
3664 /// Original position of the instruction.
3665 InsertionHandler Inserter;
3666
3667 /// Helper structure to hide all the link to the instruction. In other
3668 /// words, this helps to do as if the instruction was removed.
3669 OperandsHider Hider;
3670
3671 /// Keep track of the uses replaced, if any.
3672 UsesReplacer *Replacer = nullptr;
3673
3674 /// Keep track of instructions removed.
3675 SetOfInstrs &RemovedInsts;
3676
3677 public:
3678 /// Remove all reference of \p Inst and optionally replace all its
3679 /// uses with New.
3680 /// \p RemovedInsts Keep track of the instructions removed by this Action.
3681 /// \pre If !Inst->use_empty(), then New != nullptr
3682 InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
3683 Value *New = nullptr)
3684 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
3685 RemovedInsts(RemovedInsts) {
3686 if (New)
3687 Replacer = new UsesReplacer(Inst, New);
3688 LLVM_DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
3689 RemovedInsts.insert(Ptr: Inst);
3690 /// The instructions removed here will be freed after completing
3691 /// optimizeBlock() for all blocks as we need to keep track of the
3692 /// removed instructions during promotion.
3693 Inst->removeFromParent();
3694 }
3695
3696 ~InstructionRemover() override { delete Replacer; }
3697
3698 InstructionRemover &operator=(const InstructionRemover &other) = delete;
3699 InstructionRemover(const InstructionRemover &other) = delete;
3700
3701 /// Resurrect the instruction and reassign it to the proper uses if
3702 /// new value was provided when build this action.
3703 void undo() override {
3704 LLVM_DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
3705 Inserter.insert(Inst);
3706 if (Replacer)
3707 Replacer->undo();
3708 Hider.undo();
3709 RemovedInsts.erase(Ptr: Inst);
3710 }
3711 };
3712
3713public:
3714 /// Restoration point.
3715 /// The restoration point is a pointer to an action instead of an iterator
3716 /// because the iterator may be invalidated but not the pointer.
3717 using ConstRestorationPt = const TypePromotionAction *;
3718
3719 TypePromotionTransaction(SetOfInstrs &RemovedInsts)
3720 : RemovedInsts(RemovedInsts) {}
3721
3722 /// Advocate every changes made in that transaction. Return true if any change
3723 /// happen.
3724 bool commit();
3725
3726 /// Undo all the changes made after the given point.
3727 void rollback(ConstRestorationPt Point);
3728
3729 /// Get the current restoration point.
3730 ConstRestorationPt getRestorationPoint() const;
3731
3732 /// \name API for IR modification with state keeping to support rollback.
3733 /// @{
3734 /// Same as Instruction::setOperand.
3735 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
3736
3737 /// Same as Instruction::eraseFromParent.
3738 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
3739
3740 /// Same as Value::replaceAllUsesWith.
3741 void replaceAllUsesWith(Instruction *Inst, Value *New);
3742
3743 /// Same as Value::mutateType.
3744 void mutateType(Instruction *Inst, Type *NewTy);
3745
3746 /// Same as IRBuilder::createTrunc.
3747 Value *createTrunc(Instruction *Opnd, Type *Ty);
3748
3749 /// Same as IRBuilder::createSExt.
3750 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
3751
3752 /// Same as IRBuilder::createZExt.
3753 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
3754
3755private:
3756 /// The ordered list of actions made so far.
3757 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
3758
3759 using CommitPt =
3760 SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
3761
3762 SetOfInstrs &RemovedInsts;
3763};
3764
3765} // end anonymous namespace
3766
3767void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
3768 Value *NewVal) {
3769 Actions.push_back(Elt: std::make_unique<TypePromotionTransaction::OperandSetter>(
3770 args&: Inst, args&: Idx, args&: NewVal));
3771}
3772
3773void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
3774 Value *NewVal) {
3775 Actions.push_back(
3776 Elt: std::make_unique<TypePromotionTransaction::InstructionRemover>(
3777 args&: Inst, args&: RemovedInsts, args&: NewVal));
3778}
3779
3780void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
3781 Value *New) {
3782 Actions.push_back(
3783 Elt: std::make_unique<TypePromotionTransaction::UsesReplacer>(args&: Inst, args&: New));
3784}
3785
3786void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
3787 Actions.push_back(
3788 Elt: std::make_unique<TypePromotionTransaction::TypeMutator>(args&: Inst, args&: NewTy));
3789}
3790
3791Value *TypePromotionTransaction::createTrunc(Instruction *Opnd, Type *Ty) {
3792 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
3793 Value *Val = Ptr->getBuiltValue();
3794 Actions.push_back(Elt: std::move(Ptr));
3795 return Val;
3796}
3797
3798Value *TypePromotionTransaction::createSExt(Instruction *Inst, Value *Opnd,
3799 Type *Ty) {
3800 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
3801 Value *Val = Ptr->getBuiltValue();
3802 Actions.push_back(Elt: std::move(Ptr));
3803 return Val;
3804}
3805
3806Value *TypePromotionTransaction::createZExt(Instruction *Inst, Value *Opnd,
3807 Type *Ty) {
3808 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
3809 Value *Val = Ptr->getBuiltValue();
3810 Actions.push_back(Elt: std::move(Ptr));
3811 return Val;
3812}
3813
3814TypePromotionTransaction::ConstRestorationPt
3815TypePromotionTransaction::getRestorationPoint() const {
3816 return !Actions.empty() ? Actions.back().get() : nullptr;
3817}
3818
3819bool TypePromotionTransaction::commit() {
3820 for (std::unique_ptr<TypePromotionAction> &Action : Actions)
3821 Action->commit();
3822 bool Modified = !Actions.empty();
3823 Actions.clear();
3824 return Modified;
3825}
3826
3827void TypePromotionTransaction::rollback(
3828 TypePromotionTransaction::ConstRestorationPt Point) {
3829 while (!Actions.empty() && Point != Actions.back().get()) {
3830 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
3831 Curr->undo();
3832 }
3833}
3834
3835namespace {
3836
3837/// A helper class for matching addressing modes.
3838///
3839/// This encapsulates the logic for matching the target-legal addressing modes.
3840class AddressingModeMatcher {
3841 SmallVectorImpl<Instruction *> &AddrModeInsts;
3842 const TargetLowering &TLI;
3843 const TargetRegisterInfo &TRI;
3844 const DataLayout &DL;
3845 const LoopInfo &LI;
3846 const std::function<const DominatorTree &()> getDTFn;
3847
3848 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
3849 /// the memory instruction that we're computing this address for.
3850 Type *AccessTy;
3851 unsigned AddrSpace;
3852 Instruction *MemoryInst;
3853
3854 /// This is the addressing mode that we're building up. This is
3855 /// part of the return value of this addressing mode matching stuff.
3856 ExtAddrMode &AddrMode;
3857
3858 /// The instructions inserted by other CodeGenPrepare optimizations.
3859 const SetOfInstrs &InsertedInsts;
3860
3861 /// A map from the instructions to their type before promotion.
3862 InstrToOrigTy &PromotedInsts;
3863
3864 /// The ongoing transaction where every action should be registered.
3865 TypePromotionTransaction &TPT;
3866
3867 // A GEP which has too large offset to be folded into the addressing mode.
3868 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP;
3869
3870 /// This is set to true when we should not do profitability checks.
3871 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
3872 bool IgnoreProfitability;
3873
3874 /// True if we are optimizing for size.
3875 bool OptSize = false;
3876
3877 ProfileSummaryInfo *PSI;
3878 BlockFrequencyInfo *BFI;
3879
3880 AddressingModeMatcher(
3881 SmallVectorImpl<Instruction *> &AMI, const TargetLowering &TLI,
3882 const TargetRegisterInfo &TRI, const LoopInfo &LI,
3883 const std::function<const DominatorTree &()> getDTFn, Type *AT,
3884 unsigned AS, Instruction *MI, ExtAddrMode &AM,
3885 const SetOfInstrs &InsertedInsts, InstrToOrigTy &PromotedInsts,
3886 TypePromotionTransaction &TPT,
3887 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP,
3888 bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI)
3889 : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
3890 DL(MI->getDataLayout()), LI(LI), getDTFn(getDTFn),
3891 AccessTy(AT), AddrSpace(AS), MemoryInst(MI), AddrMode(AM),
3892 InsertedInsts(InsertedInsts), PromotedInsts(PromotedInsts), TPT(TPT),
3893 LargeOffsetGEP(LargeOffsetGEP), OptSize(OptSize), PSI(PSI), BFI(BFI) {
3894 IgnoreProfitability = false;
3895 }
3896
3897public:
3898 /// Find the maximal addressing mode that a load/store of V can fold,
3899 /// give an access type of AccessTy. This returns a list of involved
3900 /// instructions in AddrModeInsts.
3901 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
3902 /// optimizations.
3903 /// \p PromotedInsts maps the instructions to their type before promotion.
3904 /// \p The ongoing transaction where every action should be registered.
3905 static ExtAddrMode
3906 Match(Value *V, Type *AccessTy, unsigned AS, Instruction *MemoryInst,
3907 SmallVectorImpl<Instruction *> &AddrModeInsts,
3908 const TargetLowering &TLI, const LoopInfo &LI,
3909 const std::function<const DominatorTree &()> getDTFn,
3910 const TargetRegisterInfo &TRI, const SetOfInstrs &InsertedInsts,
3911 InstrToOrigTy &PromotedInsts, TypePromotionTransaction &TPT,
3912 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP,
3913 bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) {
3914 ExtAddrMode Result;
3915
3916 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI, LI, getDTFn,
3917 AccessTy, AS, MemoryInst, Result,
3918 InsertedInsts, PromotedInsts, TPT,
3919 LargeOffsetGEP, OptSize, PSI, BFI)
3920 .matchAddr(Addr: V, Depth: 0);
3921 (void)Success;
3922 assert(Success && "Couldn't select *anything*?");
3923 return Result;
3924 }
3925
3926private:
3927 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
3928 bool matchAddr(Value *Addr, unsigned Depth);
3929 bool matchOperationAddr(User *AddrInst, unsigned Opcode, unsigned Depth,
3930 bool *MovedAway = nullptr);
3931 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
3932 ExtAddrMode &AMBefore,
3933 ExtAddrMode &AMAfter);
3934 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
3935 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
3936 Value *PromotedOperand) const;
3937};
3938
3939class PhiNodeSet;
3940
3941/// An iterator for PhiNodeSet.
3942class PhiNodeSetIterator {
3943 PhiNodeSet *const Set;
3944 size_t CurrentIndex = 0;
3945
3946public:
3947 /// The constructor. Start should point to either a valid element, or be equal
3948 /// to the size of the underlying SmallVector of the PhiNodeSet.
3949 PhiNodeSetIterator(PhiNodeSet *const Set, size_t Start);
3950 PHINode *operator*() const;
3951 PhiNodeSetIterator &operator++();
3952 bool operator==(const PhiNodeSetIterator &RHS) const;
3953 bool operator!=(const PhiNodeSetIterator &RHS) const;
3954};
3955
3956/// Keeps a set of PHINodes.
3957///
3958/// This is a minimal set implementation for a specific use case:
3959/// It is very fast when there are very few elements, but also provides good
3960/// performance when there are many. It is similar to SmallPtrSet, but also
3961/// provides iteration by insertion order, which is deterministic and stable
3962/// across runs. It is also similar to SmallSetVector, but provides removing
3963/// elements in O(1) time. This is achieved by not actually removing the element
3964/// from the underlying vector, so comes at the cost of using more memory, but
3965/// that is fine, since PhiNodeSets are used as short lived objects.
3966class PhiNodeSet {
3967 friend class PhiNodeSetIterator;
3968
3969 using MapType = SmallDenseMap<PHINode *, size_t, 32>;
3970 using iterator = PhiNodeSetIterator;
3971
3972 /// Keeps the elements in the order of their insertion in the underlying
3973 /// vector. To achieve constant time removal, it never deletes any element.
3974 SmallVector<PHINode *, 32> NodeList;
3975
3976 /// Keeps the elements in the underlying set implementation. This (and not the
3977 /// NodeList defined above) is the source of truth on whether an element
3978 /// is actually in the collection.
3979 MapType NodeMap;
3980
3981 /// Points to the first valid (not deleted) element when the set is not empty
3982 /// and the value is not zero. Equals to the size of the underlying vector
3983 /// when the set is empty. When the value is 0, as in the beginning, the
3984 /// first element may or may not be valid.
3985 size_t FirstValidElement = 0;
3986
3987public:
3988 /// Inserts a new element to the collection.
3989 /// \returns true if the element is actually added, i.e. was not in the
3990 /// collection before the operation.
3991 bool insert(PHINode *Ptr) {
3992 if (NodeMap.insert(KV: std::make_pair(x&: Ptr, y: NodeList.size())).second) {
3993 NodeList.push_back(Elt: Ptr);
3994 return true;
3995 }
3996 return false;
3997 }
3998
3999 /// Removes the element from the collection.
4000 /// \returns whether the element is actually removed, i.e. was in the
4001 /// collection before the operation.
4002 bool erase(PHINode *Ptr) {
4003 if (NodeMap.erase(Val: Ptr)) {
4004 SkipRemovedElements(CurrentIndex&: FirstValidElement);
4005 return true;
4006 }
4007 return false;
4008 }
4009
4010 /// Removes all elements and clears the collection.
4011 void clear() {
4012 NodeMap.clear();
4013 NodeList.clear();
4014 FirstValidElement = 0;
4015 }
4016
4017 /// \returns an iterator that will iterate the elements in the order of
4018 /// insertion.
4019 iterator begin() {
4020 if (FirstValidElement == 0)
4021 SkipRemovedElements(CurrentIndex&: FirstValidElement);
4022 return PhiNodeSetIterator(this, FirstValidElement);
4023 }
4024
4025 /// \returns an iterator that points to the end of the collection.
4026 iterator end() { return PhiNodeSetIterator(this, NodeList.size()); }
4027
4028 /// Returns the number of elements in the collection.
4029 size_t size() const { return NodeMap.size(); }
4030
4031 /// \returns 1 if the given element is in the collection, and 0 if otherwise.
4032 size_t count(PHINode *Ptr) const { return NodeMap.count(Val: Ptr); }
4033
4034private:
4035 /// Updates the CurrentIndex so that it will point to a valid element.
4036 ///
4037 /// If the element of NodeList at CurrentIndex is valid, it does not
4038 /// change it. If there are no more valid elements, it updates CurrentIndex
4039 /// to point to the end of the NodeList.
4040 void SkipRemovedElements(size_t &CurrentIndex) {
4041 while (CurrentIndex < NodeList.size()) {
4042 auto it = NodeMap.find(Val: NodeList[CurrentIndex]);
4043 // If the element has been deleted and added again later, NodeMap will
4044 // point to a different index, so CurrentIndex will still be invalid.
4045 if (it != NodeMap.end() && it->second == CurrentIndex)
4046 break;
4047 ++CurrentIndex;
4048 }
4049 }
4050};
4051
4052PhiNodeSetIterator::PhiNodeSetIterator(PhiNodeSet *const Set, size_t Start)
4053 : Set(Set), CurrentIndex(Start) {}
4054
4055PHINode *PhiNodeSetIterator::operator*() const {
4056 assert(CurrentIndex < Set->NodeList.size() &&
4057 "PhiNodeSet access out of range");
4058 return Set->NodeList[CurrentIndex];
4059}
4060
4061PhiNodeSetIterator &PhiNodeSetIterator::operator++() {
4062 assert(CurrentIndex < Set->NodeList.size() &&
4063 "PhiNodeSet access out of range");
4064 ++CurrentIndex;
4065 Set->SkipRemovedElements(CurrentIndex);
4066 return *this;
4067}
4068
4069bool PhiNodeSetIterator::operator==(const PhiNodeSetIterator &RHS) const {
4070 return CurrentIndex == RHS.CurrentIndex;
4071}
4072
4073bool PhiNodeSetIterator::operator!=(const PhiNodeSetIterator &RHS) const {
4074 return !((*this) == RHS);
4075}
4076
4077/// Keep track of simplification of Phi nodes.
4078/// Accept the set of all phi nodes and erase phi node from this set
4079/// if it is simplified.
4080class SimplificationTracker {
4081 DenseMap<Value *, Value *> Storage;
4082 // Tracks newly created Phi nodes. The elements are iterated by insertion
4083 // order.
4084 PhiNodeSet AllPhiNodes;
4085 // Tracks newly created Select nodes.
4086 SmallPtrSet<SelectInst *, 32> AllSelectNodes;
4087
4088public:
4089 Value *Get(Value *V) {
4090 do {
4091 auto SV = Storage.find(Val: V);
4092 if (SV == Storage.end())
4093 return V;
4094 V = SV->second;
4095 } while (true);
4096 }
4097
4098 void Put(Value *From, Value *To) { Storage.insert(KV: {From, To}); }
4099
4100 void ReplacePhi(PHINode *From, PHINode *To) {
4101 Value *OldReplacement = Get(V: From);
4102 while (OldReplacement != From) {
4103 From = To;
4104 To = dyn_cast<PHINode>(Val: OldReplacement);
4105 OldReplacement = Get(V: From);
4106 }
4107 assert(To && Get(To) == To && "Replacement PHI node is already replaced.");
4108 Put(From, To);
4109 From->replaceAllUsesWith(V: To);
4110 AllPhiNodes.erase(Ptr: From);
4111 From->eraseFromParent();
4112 }
4113
4114 PhiNodeSet &newPhiNodes() { return AllPhiNodes; }
4115
4116 void insertNewPhi(PHINode *PN) { AllPhiNodes.insert(Ptr: PN); }
4117
4118 void insertNewSelect(SelectInst *SI) { AllSelectNodes.insert(Ptr: SI); }
4119
4120 unsigned countNewPhiNodes() const { return AllPhiNodes.size(); }
4121
4122 unsigned countNewSelectNodes() const { return AllSelectNodes.size(); }
4123
4124 void destroyNewNodes(Type *CommonType) {
4125 // For safe erasing, replace the uses with dummy value first.
4126 auto *Dummy = PoisonValue::get(T: CommonType);
4127 for (auto *I : AllPhiNodes) {
4128 I->replaceAllUsesWith(V: Dummy);
4129 I->eraseFromParent();
4130 }
4131 AllPhiNodes.clear();
4132 for (auto *I : AllSelectNodes) {
4133 I->replaceAllUsesWith(V: Dummy);
4134 I->eraseFromParent();
4135 }
4136 AllSelectNodes.clear();
4137 }
4138};
4139
4140/// A helper class for combining addressing modes.
4141class AddressingModeCombiner {
4142 typedef DenseMap<Value *, Value *> FoldAddrToValueMapping;
4143 typedef std::pair<PHINode *, PHINode *> PHIPair;
4144
4145private:
4146 /// The addressing modes we've collected.
4147 SmallVector<ExtAddrMode, 16> AddrModes;
4148
4149 /// The field in which the AddrModes differ, when we have more than one.
4150 ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
4151
4152 /// Are the AddrModes that we have all just equal to their original values?
4153 bool AllAddrModesTrivial = true;
4154
4155 /// Common Type for all different fields in addressing modes.
4156 Type *CommonType = nullptr;
4157
4158 const DataLayout &DL;
4159
4160 /// Original Address.
4161 Value *Original;
4162
4163 /// Common value among addresses
4164 Value *CommonValue = nullptr;
4165
4166public:
4167 AddressingModeCombiner(const DataLayout &DL, Value *OriginalValue)
4168 : DL(DL), Original(OriginalValue) {}
4169
4170 ~AddressingModeCombiner() { eraseCommonValueIfDead(); }
4171
4172 /// Get the combined AddrMode
4173 const ExtAddrMode &getAddrMode() const { return AddrModes[0]; }
4174
4175 /// Add a new AddrMode if it's compatible with the AddrModes we already
4176 /// have.
4177 /// \return True iff we succeeded in doing so.
4178 bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
4179 // Take note of if we have any non-trivial AddrModes, as we need to detect
4180 // when all AddrModes are trivial as then we would introduce a phi or select
4181 // which just duplicates what's already there.
4182 AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
4183
4184 // If this is the first addrmode then everything is fine.
4185 if (AddrModes.empty()) {
4186 AddrModes.emplace_back(Args&: NewAddrMode);
4187 return true;
4188 }
4189
4190 // Figure out how different this is from the other address modes, which we
4191 // can do just by comparing against the first one given that we only care
4192 // about the cumulative difference.
4193 ExtAddrMode::FieldName ThisDifferentField =
4194 AddrModes[0].compare(other: NewAddrMode);
4195 if (DifferentField == ExtAddrMode::NoField)
4196 DifferentField = ThisDifferentField;
4197 else if (DifferentField != ThisDifferentField)
4198 DifferentField = ExtAddrMode::MultipleFields;
4199
4200 // If NewAddrMode differs in more than one dimension we cannot handle it.
4201 bool CanHandle = DifferentField != ExtAddrMode::MultipleFields;
4202
4203 // If Scale Field is different then we reject.
4204 CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField;
4205
4206 // We also must reject the case when base offset is different and
4207 // scale reg is not null, we cannot handle this case due to merge of
4208 // different offsets will be used as ScaleReg.
4209 CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField ||
4210 !NewAddrMode.ScaledReg);
4211
4212 // We also must reject the case when GV is different and BaseReg installed
4213 // due to we want to use base reg as a merge of GV values.
4214 CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField ||
4215 !NewAddrMode.HasBaseReg);
4216
4217 // Even if NewAddMode is the same we still need to collect it due to
4218 // original value is different. And later we will need all original values
4219 // as anchors during finding the common Phi node.
4220 if (CanHandle)
4221 AddrModes.emplace_back(Args&: NewAddrMode);
4222 else
4223 AddrModes.clear();
4224
4225 return CanHandle;
4226 }
4227
4228 /// Combine the addressing modes we've collected into a single
4229 /// addressing mode.
4230 /// \return True iff we successfully combined them or we only had one so
4231 /// didn't need to combine them anyway.
4232 bool combineAddrModes() {
4233 // If we have no AddrModes then they can't be combined.
4234 if (AddrModes.size() == 0)
4235 return false;
4236
4237 // A single AddrMode can trivially be combined.
4238 if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
4239 return true;
4240
4241 // If the AddrModes we collected are all just equal to the value they are
4242 // derived from then combining them wouldn't do anything useful.
4243 if (AllAddrModesTrivial)
4244 return false;
4245
4246 if (!addrModeCombiningAllowed())
4247 return false;
4248
4249 // Build a map between <original value, basic block where we saw it> to
4250 // value of base register.
4251 // Bail out if there is no common type.
4252 FoldAddrToValueMapping Map;
4253 if (!initializeMap(Map))
4254 return false;
4255
4256 CommonValue = findCommon(Map);
4257 if (CommonValue)
4258 AddrModes[0].SetCombinedField(Field: DifferentField, V: CommonValue, AddrModes);
4259 return CommonValue != nullptr;
4260 }
4261
4262private:
4263 /// `CommonValue` may be a placeholder inserted by us.
4264 /// If the placeholder is not used, we should remove this dead instruction.
4265 void eraseCommonValueIfDead() {
4266 if (CommonValue && CommonValue->use_empty())
4267 if (Instruction *CommonInst = dyn_cast<Instruction>(Val: CommonValue))
4268 CommonInst->eraseFromParent();
4269 }
4270
4271 /// Initialize Map with anchor values. For address seen
4272 /// we set the value of different field saw in this address.
4273 /// At the same time we find a common type for different field we will
4274 /// use to create new Phi/Select nodes. Keep it in CommonType field.
4275 /// Return false if there is no common type found.
4276 bool initializeMap(FoldAddrToValueMapping &Map) {
4277 // Keep track of keys where the value is null. We will need to replace it
4278 // with constant null when we know the common type.
4279 SmallVector<Value *, 2> NullValue;
4280 Type *IntPtrTy = DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
4281 for (auto &AM : AddrModes) {
4282 Value *DV = AM.GetFieldAsValue(Field: DifferentField, IntPtrTy);
4283 if (DV) {
4284 auto *Type = DV->getType();
4285 if (CommonType && CommonType != Type)
4286 return false;
4287 CommonType = Type;
4288 Map[AM.OriginalValue] = DV;
4289 } else {
4290 NullValue.push_back(Elt: AM.OriginalValue);
4291 }
4292 }
4293 assert(CommonType && "At least one non-null value must be!");
4294 for (auto *V : NullValue)
4295 Map[V] = Constant::getNullValue(Ty: CommonType);
4296 return true;
4297 }
4298
4299 /// We have mapping between value A and other value B where B was a field in
4300 /// addressing mode represented by A. Also we have an original value C
4301 /// representing an address we start with. Traversing from C through phi and
4302 /// selects we ended up with A's in a map. This utility function tries to find
4303 /// a value V which is a field in addressing mode C and traversing through phi
4304 /// nodes and selects we will end up in corresponded values B in a map.
4305 /// The utility will create a new Phi/Selects if needed.
4306 // The simple example looks as follows:
4307 // BB1:
4308 // p1 = b1 + 40
4309 // br cond BB2, BB3
4310 // BB2:
4311 // p2 = b2 + 40
4312 // br BB3
4313 // BB3:
4314 // p = phi [p1, BB1], [p2, BB2]
4315 // v = load p
4316 // Map is
4317 // p1 -> b1
4318 // p2 -> b2
4319 // Request is
4320 // p -> ?
4321 // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3.
4322 Value *findCommon(FoldAddrToValueMapping &Map) {
4323 // Tracks the simplification of newly created phi nodes. The reason we use
4324 // this mapping is because we will add new created Phi nodes in AddrToBase.
4325 // Simplification of Phi nodes is recursive, so some Phi node may
4326 // be simplified after we added it to AddrToBase. In reality this
4327 // simplification is possible only if original phi/selects were not
4328 // simplified yet.
4329 // Using this mapping we can find the current value in AddrToBase.
4330 SimplificationTracker ST;
4331
4332 // First step, DFS to create PHI nodes for all intermediate blocks.
4333 // Also fill traverse order for the second step.
4334 SmallVector<Value *, 32> TraverseOrder;
4335 InsertPlaceholders(Map, TraverseOrder, ST);
4336
4337 // Second Step, fill new nodes by merged values and simplify if possible.
4338 FillPlaceholders(Map, TraverseOrder, ST);
4339
4340 if (!AddrSinkNewSelects && ST.countNewSelectNodes() > 0) {
4341 ST.destroyNewNodes(CommonType);
4342 return nullptr;
4343 }
4344
4345 // Now we'd like to match New Phi nodes to existed ones.
4346 unsigned PhiNotMatchedCount = 0;
4347 if (!MatchPhiSet(ST, AllowNewPhiNodes: AddrSinkNewPhis, PhiNotMatchedCount)) {
4348 ST.destroyNewNodes(CommonType);
4349 return nullptr;
4350 }
4351
4352 auto *Result = ST.Get(V: Map.find(Val: Original)->second);
4353 if (Result) {
4354 NumMemoryInstsPhiCreated += ST.countNewPhiNodes() + PhiNotMatchedCount;
4355 NumMemoryInstsSelectCreated += ST.countNewSelectNodes();
4356 }
4357 return Result;
4358 }
4359
4360 /// Try to match PHI node to Candidate.
4361 /// Matcher tracks the matched Phi nodes.
4362 bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
4363 SmallSetVector<PHIPair, 8> &Matcher,
4364 PhiNodeSet &PhiNodesToMatch) {
4365 SmallVector<PHIPair, 8> WorkList;
4366 Matcher.insert(X: {PHI, Candidate});
4367 SmallPtrSet<PHINode *, 8> MatchedPHIs;
4368 MatchedPHIs.insert(Ptr: PHI);
4369 WorkList.push_back(Elt: {PHI, Candidate});
4370 SmallSet<PHIPair, 8> Visited;
4371 while (!WorkList.empty()) {
4372 auto Item = WorkList.pop_back_val();
4373 if (!Visited.insert(V: Item).second)
4374 continue;
4375 // We iterate over all incoming values to Phi to compare them.
4376 // If values are different and both of them Phi and the first one is a
4377 // Phi we added (subject to match) and both of them is in the same basic
4378 // block then we can match our pair if values match. So we state that
4379 // these values match and add it to work list to verify that.
4380 for (auto *B : Item.first->blocks()) {
4381 Value *FirstValue = Item.first->getIncomingValueForBlock(BB: B);
4382 Value *SecondValue = Item.second->getIncomingValueForBlock(BB: B);
4383 if (FirstValue == SecondValue)
4384 continue;
4385
4386 PHINode *FirstPhi = dyn_cast<PHINode>(Val: FirstValue);
4387 PHINode *SecondPhi = dyn_cast<PHINode>(Val: SecondValue);
4388
4389 // One of them is not Phi or
4390 // The first one is not Phi node from the set we'd like to match or
4391 // Phi nodes from different basic blocks then
4392 // we will not be able to match.
4393 if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(Ptr: FirstPhi) ||
4394 FirstPhi->getParent() != SecondPhi->getParent())
4395 return false;
4396
4397 // If we already matched them then continue.
4398 if (Matcher.count(key: {FirstPhi, SecondPhi}))
4399 continue;
4400 // So the values are different and does not match. So we need them to
4401 // match. (But we register no more than one match per PHI node, so that
4402 // we won't later try to replace them twice.)
4403 if (MatchedPHIs.insert(Ptr: FirstPhi).second)
4404 Matcher.insert(X: {FirstPhi, SecondPhi});
4405 // But me must check it.
4406 WorkList.push_back(Elt: {FirstPhi, SecondPhi});
4407 }
4408 }
4409 return true;
4410 }
4411
4412 /// For the given set of PHI nodes (in the SimplificationTracker) try
4413 /// to find their equivalents.
4414 /// Returns false if this matching fails and creation of new Phi is disabled.
4415 bool MatchPhiSet(SimplificationTracker &ST, bool AllowNewPhiNodes,
4416 unsigned &PhiNotMatchedCount) {
4417 // Matched and PhiNodesToMatch iterate their elements in a deterministic
4418 // order, so the replacements (ReplacePhi) are also done in a deterministic
4419 // order.
4420 SmallSetVector<PHIPair, 8> Matched;
4421 SmallPtrSet<PHINode *, 8> WillNotMatch;
4422 PhiNodeSet &PhiNodesToMatch = ST.newPhiNodes();
4423 while (PhiNodesToMatch.size()) {
4424 PHINode *PHI = *PhiNodesToMatch.begin();
4425
4426 // Add us, if no Phi nodes in the basic block we do not match.
4427 WillNotMatch.clear();
4428 WillNotMatch.insert(Ptr: PHI);
4429
4430 // Traverse all Phis until we found equivalent or fail to do that.
4431 bool IsMatched = false;
4432 for (auto &P : PHI->getParent()->phis()) {
4433 // Skip new Phi nodes.
4434 if (PhiNodesToMatch.count(Ptr: &P))
4435 continue;
4436 if ((IsMatched = MatchPhiNode(PHI, Candidate: &P, Matcher&: Matched, PhiNodesToMatch)))
4437 break;
4438 // If it does not match, collect all Phi nodes from matcher.
4439 // if we end up with no match, them all these Phi nodes will not match
4440 // later.
4441 WillNotMatch.insert_range(R: llvm::make_first_range(c&: Matched));
4442 Matched.clear();
4443 }
4444 if (IsMatched) {
4445 // Replace all matched values and erase them.
4446 for (auto MV : Matched)
4447 ST.ReplacePhi(From: MV.first, To: MV.second);
4448 Matched.clear();
4449 continue;
4450 }
4451 // If we are not allowed to create new nodes then bail out.
4452 if (!AllowNewPhiNodes)
4453 return false;
4454 // Just remove all seen values in matcher. They will not match anything.
4455 PhiNotMatchedCount += WillNotMatch.size();
4456 for (auto *P : WillNotMatch)
4457 PhiNodesToMatch.erase(Ptr: P);
4458 }
4459 return true;
4460 }
4461 /// Fill the placeholders with values from predecessors and simplify them.
4462 void FillPlaceholders(FoldAddrToValueMapping &Map,
4463 SmallVectorImpl<Value *> &TraverseOrder,
4464 SimplificationTracker &ST) {
4465 while (!TraverseOrder.empty()) {
4466 Value *Current = TraverseOrder.pop_back_val();
4467 assert(Map.contains(Current) && "No node to fill!!!");
4468 Value *V = Map[Current];
4469
4470 if (SelectInst *Select = dyn_cast<SelectInst>(Val: V)) {
4471 // CurrentValue also must be Select.
4472 auto *CurrentSelect = cast<SelectInst>(Val: Current);
4473 auto *TrueValue = CurrentSelect->getTrueValue();
4474 assert(Map.contains(TrueValue) && "No True Value!");
4475 Select->setTrueValue(ST.Get(V: Map[TrueValue]));
4476 auto *FalseValue = CurrentSelect->getFalseValue();
4477 assert(Map.contains(FalseValue) && "No False Value!");
4478 Select->setFalseValue(ST.Get(V: Map[FalseValue]));
4479 } else {
4480 // Must be a Phi node then.
4481 auto *PHI = cast<PHINode>(Val: V);
4482 // Fill the Phi node with values from predecessors.
4483 for (auto *B : predecessors(BB: PHI->getParent())) {
4484 Value *PV = cast<PHINode>(Val: Current)->getIncomingValueForBlock(BB: B);
4485 assert(Map.contains(PV) && "No predecessor Value!");
4486 PHI->addIncoming(V: ST.Get(V: Map[PV]), BB: B);
4487 }
4488 }
4489 }
4490 }
4491
4492 /// Starting from original value recursively iterates over def-use chain up to
4493 /// known ending values represented in a map. For each traversed phi/select
4494 /// inserts a placeholder Phi or Select.
4495 /// Reports all new created Phi/Select nodes by adding them to set.
4496 /// Also reports and order in what values have been traversed.
4497 void InsertPlaceholders(FoldAddrToValueMapping &Map,
4498 SmallVectorImpl<Value *> &TraverseOrder,
4499 SimplificationTracker &ST) {
4500 SmallVector<Value *, 32> Worklist;
4501 assert((isa<PHINode>(Original) || isa<SelectInst>(Original)) &&
4502 "Address must be a Phi or Select node");
4503 auto *Dummy = PoisonValue::get(T: CommonType);
4504 Worklist.push_back(Elt: Original);
4505 while (!Worklist.empty()) {
4506 Value *Current = Worklist.pop_back_val();
4507 // if it is already visited or it is an ending value then skip it.
4508 if (Map.contains(Val: Current))
4509 continue;
4510 TraverseOrder.push_back(Elt: Current);
4511
4512 // CurrentValue must be a Phi node or select. All others must be covered
4513 // by anchors.
4514 if (SelectInst *CurrentSelect = dyn_cast<SelectInst>(Val: Current)) {
4515 // Is it OK to get metadata from OrigSelect?!
4516 // Create a Select placeholder with dummy value.
4517 SelectInst *Select =
4518 SelectInst::Create(C: CurrentSelect->getCondition(), S1: Dummy, S2: Dummy,
4519 NameStr: CurrentSelect->getName(),
4520 InsertBefore: CurrentSelect->getIterator(), MDFrom: CurrentSelect);
4521 Map[Current] = Select;
4522 ST.insertNewSelect(SI: Select);
4523 // We are interested in True and False values.
4524 Worklist.push_back(Elt: CurrentSelect->getTrueValue());
4525 Worklist.push_back(Elt: CurrentSelect->getFalseValue());
4526 } else {
4527 // It must be a Phi node then.
4528 PHINode *CurrentPhi = cast<PHINode>(Val: Current);
4529 unsigned PredCount = CurrentPhi->getNumIncomingValues();
4530 PHINode *PHI =
4531 PHINode::Create(Ty: CommonType, NumReservedValues: PredCount, NameStr: "sunk_phi", InsertBefore: CurrentPhi->getIterator());
4532 Map[Current] = PHI;
4533 ST.insertNewPhi(PN: PHI);
4534 append_range(C&: Worklist, R: CurrentPhi->incoming_values());
4535 }
4536 }
4537 }
4538
4539 bool addrModeCombiningAllowed() {
4540 if (DisableComplexAddrModes)
4541 return false;
4542 switch (DifferentField) {
4543 default:
4544 return false;
4545 case ExtAddrMode::BaseRegField:
4546 return AddrSinkCombineBaseReg;
4547 case ExtAddrMode::BaseGVField:
4548 return AddrSinkCombineBaseGV;
4549 case ExtAddrMode::BaseOffsField:
4550 return AddrSinkCombineBaseOffs;
4551 case ExtAddrMode::ScaledRegField:
4552 return AddrSinkCombineScaledReg;
4553 }
4554 }
4555};
4556} // end anonymous namespace
4557
4558/// Try adding ScaleReg*Scale to the current addressing mode.
4559/// Return true and update AddrMode if this addr mode is legal for the target,
4560/// false if not.
4561bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
4562 unsigned Depth) {
4563 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
4564 // mode. Just process that directly.
4565 if (Scale == 1)
4566 return matchAddr(Addr: ScaleReg, Depth);
4567
4568 // If the scale is 0, it takes nothing to add this.
4569 if (Scale == 0)
4570 return true;
4571
4572 // If we already have a scale of this value, we can add to it, otherwise, we
4573 // need an available scale field.
4574 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
4575 return false;
4576
4577 ExtAddrMode TestAddrMode = AddrMode;
4578
4579 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
4580 // [A+B + A*7] -> [B+A*8].
4581 TestAddrMode.Scale += Scale;
4582 TestAddrMode.ScaledReg = ScaleReg;
4583
4584 // If the new address isn't legal, bail out.
4585 if (!TLI.isLegalAddressingMode(DL, AM: TestAddrMode, Ty: AccessTy, AddrSpace))
4586 return false;
4587
4588 // It was legal, so commit it.
4589 AddrMode = TestAddrMode;
4590
4591 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
4592 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
4593 // X*Scale + C*Scale to addr mode. If we found available IV increment, do not
4594 // go any further: we can reuse it and cannot eliminate it.
4595 ConstantInt *CI = nullptr;
4596 Value *AddLHS = nullptr;
4597 if (isa<Instruction>(Val: ScaleReg) && // not a constant expr.
4598 match(V: ScaleReg, P: m_Add(L: m_Value(V&: AddLHS), R: m_ConstantInt(CI))) &&
4599 !isIVIncrement(V: ScaleReg, LI: &LI) && CI->getValue().isSignedIntN(N: 64)) {
4600 TestAddrMode.InBounds = false;
4601 TestAddrMode.ScaledReg = AddLHS;
4602 TestAddrMode.BaseOffs += CI->getSExtValue() * TestAddrMode.Scale;
4603
4604 // If this addressing mode is legal, commit it and remember that we folded
4605 // this instruction.
4606 if (TLI.isLegalAddressingMode(DL, AM: TestAddrMode, Ty: AccessTy, AddrSpace)) {
4607 AddrModeInsts.push_back(Elt: cast<Instruction>(Val: ScaleReg));
4608 AddrMode = TestAddrMode;
4609 return true;
4610 }
4611 // Restore status quo.
4612 TestAddrMode = AddrMode;
4613 }
4614
4615 // If this is an add recurrence with a constant step, return the increment
4616 // instruction and the canonicalized step.
4617 auto GetConstantStep =
4618 [this](const Value *V) -> std::optional<std::pair<Instruction *, APInt>> {
4619 auto *PN = dyn_cast<PHINode>(Val: V);
4620 if (!PN)
4621 return std::nullopt;
4622 auto IVInc = getIVIncrement(PN, LI: &LI);
4623 if (!IVInc)
4624 return std::nullopt;
4625 // TODO: The result of the intrinsics above is two-complement. However when
4626 // IV inc is expressed as add or sub, iv.next is potentially a poison value.
4627 // If it has nuw or nsw flags, we need to make sure that these flags are
4628 // inferrable at the point of memory instruction. Otherwise we are replacing
4629 // well-defined two-complement computation with poison. Currently, to avoid
4630 // potentially complex analysis needed to prove this, we reject such cases.
4631 if (auto *OIVInc = dyn_cast<OverflowingBinaryOperator>(Val: IVInc->first))
4632 if (OIVInc->hasNoSignedWrap() || OIVInc->hasNoUnsignedWrap())
4633 return std::nullopt;
4634 if (auto *ConstantStep = dyn_cast<ConstantInt>(Val: IVInc->second))
4635 return std::make_pair(x&: IVInc->first, y: ConstantStep->getValue());
4636 return std::nullopt;
4637 };
4638
4639 // Try to account for the following special case:
4640 // 1. ScaleReg is an inductive variable;
4641 // 2. We use it with non-zero offset;
4642 // 3. IV's increment is available at the point of memory instruction.
4643 //
4644 // In this case, we may reuse the IV increment instead of the IV Phi to
4645 // achieve the following advantages:
4646 // 1. If IV step matches the offset, we will have no need in the offset;
4647 // 2. Even if they don't match, we will reduce the overlap of living IV
4648 // and IV increment, that will potentially lead to better register
4649 // assignment.
4650 if (AddrMode.BaseOffs) {
4651 if (auto IVStep = GetConstantStep(ScaleReg)) {
4652 Instruction *IVInc = IVStep->first;
4653 // The following assert is important to ensure a lack of infinite loops.
4654 // This transforms is (intentionally) the inverse of the one just above.
4655 // If they don't agree on the definition of an increment, we'd alternate
4656 // back and forth indefinitely.
4657 assert(isIVIncrement(IVInc, &LI) && "implied by GetConstantStep");
4658 APInt Step = IVStep->second;
4659 APInt Offset = Step * AddrMode.Scale;
4660 if (Offset.isSignedIntN(N: 64)) {
4661 TestAddrMode.InBounds = false;
4662 TestAddrMode.ScaledReg = IVInc;
4663 TestAddrMode.BaseOffs -= Offset.getLimitedValue();
4664 // If this addressing mode is legal, commit it..
4665 // (Note that we defer the (expensive) domtree base legality check
4666 // to the very last possible point.)
4667 if (TLI.isLegalAddressingMode(DL, AM: TestAddrMode, Ty: AccessTy, AddrSpace) &&
4668 getDTFn().dominates(Def: IVInc, User: MemoryInst)) {
4669 AddrModeInsts.push_back(Elt: cast<Instruction>(Val: IVInc));
4670 AddrMode = TestAddrMode;
4671 return true;
4672 }
4673 // Restore status quo.
4674 TestAddrMode = AddrMode;
4675 }
4676 }
4677 }
4678
4679 // Otherwise, just return what we have.
4680 return true;
4681}
4682
4683/// This is a little filter, which returns true if an addressing computation
4684/// involving I might be folded into a load/store accessing it.
4685/// This doesn't need to be perfect, but needs to accept at least
4686/// the set of instructions that MatchOperationAddr can.
4687static bool MightBeFoldableInst(Instruction *I) {
4688 switch (I->getOpcode()) {
4689 case Instruction::BitCast:
4690 case Instruction::AddrSpaceCast:
4691 // Don't touch identity bitcasts.
4692 if (I->getType() == I->getOperand(i: 0)->getType())
4693 return false;
4694 return I->getType()->isIntOrPtrTy();
4695 case Instruction::PtrToInt:
4696 // PtrToInt is always a noop, as we know that the int type is pointer sized.
4697 return true;
4698 case Instruction::IntToPtr:
4699 // We know the input is intptr_t, so this is foldable.
4700 return true;
4701 case Instruction::Add:
4702 return true;
4703 case Instruction::Mul:
4704 case Instruction::Shl:
4705 // Can only handle X*C and X << C.
4706 return isa<ConstantInt>(Val: I->getOperand(i: 1));
4707 case Instruction::GetElementPtr:
4708 return true;
4709 default:
4710 return false;
4711 }
4712}
4713
4714/// Check whether or not \p Val is a legal instruction for \p TLI.
4715/// \note \p Val is assumed to be the product of some type promotion.
4716/// Therefore if \p Val has an undefined state in \p TLI, this is assumed
4717/// to be legal, as the non-promoted value would have had the same state.
4718static bool isPromotedInstructionLegal(const TargetLowering &TLI,
4719 const DataLayout &DL, Value *Val) {
4720 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
4721 if (!PromotedInst)
4722 return false;
4723 int ISDOpcode = TLI.InstructionOpcodeToISD(Opcode: PromotedInst->getOpcode());
4724 // If the ISDOpcode is undefined, it was undefined before the promotion.
4725 if (!ISDOpcode)
4726 return true;
4727 // Otherwise, check if the promoted instruction is legal or not.
4728 return TLI.isOperationLegalOrCustom(
4729 Op: ISDOpcode, VT: TLI.getValueType(DL, Ty: PromotedInst->getType()));
4730}
4731
4732namespace {
4733
4734/// Hepler class to perform type promotion.
4735class TypePromotionHelper {
4736 /// Utility function to add a promoted instruction \p ExtOpnd to
4737 /// \p PromotedInsts and record the type of extension we have seen.
4738 static void addPromotedInst(InstrToOrigTy &PromotedInsts,
4739 Instruction *ExtOpnd, bool IsSExt) {
4740 ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
4741 auto [It, Inserted] = PromotedInsts.try_emplace(Key: ExtOpnd);
4742 if (!Inserted) {
4743 // If the new extension is same as original, the information in
4744 // PromotedInsts[ExtOpnd] is still correct.
4745 if (It->second.getInt() == ExtTy)
4746 return;
4747
4748 // Now the new extension is different from old extension, we make
4749 // the type information invalid by setting extension type to
4750 // BothExtension.
4751 ExtTy = BothExtension;
4752 }
4753 It->second = TypeIsSExt(ExtOpnd->getType(), ExtTy);
4754 }
4755
4756 /// Utility function to query the original type of instruction \p Opnd
4757 /// with a matched extension type. If the extension doesn't match, we
4758 /// cannot use the information we had on the original type.
4759 /// BothExtension doesn't match any extension type.
4760 static const Type *getOrigType(const InstrToOrigTy &PromotedInsts,
4761 Instruction *Opnd, bool IsSExt) {
4762 ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
4763 InstrToOrigTy::const_iterator It = PromotedInsts.find(Val: Opnd);
4764 if (It != PromotedInsts.end() && It->second.getInt() == ExtTy)
4765 return It->second.getPointer();
4766 return nullptr;
4767 }
4768
4769 /// Utility function to check whether or not a sign or zero extension
4770 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
4771 /// either using the operands of \p Inst or promoting \p Inst.
4772 /// The type of the extension is defined by \p IsSExt.
4773 /// In other words, check if:
4774 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
4775 /// #1 Promotion applies:
4776 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
4777 /// #2 Operand reuses:
4778 /// ext opnd1 to ConsideredExtType.
4779 /// \p PromotedInsts maps the instructions to their type before promotion.
4780 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
4781 const InstrToOrigTy &PromotedInsts, bool IsSExt);
4782
4783 /// Utility function to determine if \p OpIdx should be promoted when
4784 /// promoting \p Inst.
4785 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
4786 return !(isa<SelectInst>(Val: Inst) && OpIdx == 0);
4787 }
4788
4789 /// Utility function to promote the operand of \p Ext when this
4790 /// operand is a promotable trunc or sext or zext.
4791 /// \p PromotedInsts maps the instructions to their type before promotion.
4792 /// \p CreatedInstsCost[out] contains the cost of all instructions
4793 /// created to promote the operand of Ext.
4794 /// Newly added extensions are inserted in \p Exts.
4795 /// Newly added truncates are inserted in \p Truncs.
4796 /// Should never be called directly.
4797 /// \return The promoted value which is used instead of Ext.
4798 static Value *promoteOperandForTruncAndAnyExt(
4799 Instruction *Ext, TypePromotionTransaction &TPT,
4800 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4801 SmallVectorImpl<Instruction *> *Exts,
4802 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
4803
4804 /// Utility function to promote the operand of \p Ext when this
4805 /// operand is promotable and is not a supported trunc or sext.
4806 /// \p PromotedInsts maps the instructions to their type before promotion.
4807 /// \p CreatedInstsCost[out] contains the cost of all the instructions
4808 /// created to promote the operand of Ext.
4809 /// Newly added extensions are inserted in \p Exts.
4810 /// Newly added truncates are inserted in \p Truncs.
4811 /// Should never be called directly.
4812 /// \return The promoted value which is used instead of Ext.
4813 static Value *promoteOperandForOther(Instruction *Ext,
4814 TypePromotionTransaction &TPT,
4815 InstrToOrigTy &PromotedInsts,
4816 unsigned &CreatedInstsCost,
4817 SmallVectorImpl<Instruction *> *Exts,
4818 SmallVectorImpl<Instruction *> *Truncs,
4819 const TargetLowering &TLI, bool IsSExt);
4820
4821 /// \see promoteOperandForOther.
4822 static Value *signExtendOperandForOther(
4823 Instruction *Ext, TypePromotionTransaction &TPT,
4824 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4825 SmallVectorImpl<Instruction *> *Exts,
4826 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4827 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
4828 Exts, Truncs, TLI, IsSExt: true);
4829 }
4830
4831 /// \see promoteOperandForOther.
4832 static Value *zeroExtendOperandForOther(
4833 Instruction *Ext, TypePromotionTransaction &TPT,
4834 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4835 SmallVectorImpl<Instruction *> *Exts,
4836 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4837 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
4838 Exts, Truncs, TLI, IsSExt: false);
4839 }
4840
4841public:
4842 /// Type for the utility function that promotes the operand of Ext.
4843 using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
4844 InstrToOrigTy &PromotedInsts,
4845 unsigned &CreatedInstsCost,
4846 SmallVectorImpl<Instruction *> *Exts,
4847 SmallVectorImpl<Instruction *> *Truncs,
4848 const TargetLowering &TLI);
4849
4850 /// Given a sign/zero extend instruction \p Ext, return the appropriate
4851 /// action to promote the operand of \p Ext instead of using Ext.
4852 /// \return NULL if no promotable action is possible with the current
4853 /// sign extension.
4854 /// \p InsertedInsts keeps track of all the instructions inserted by the
4855 /// other CodeGenPrepare optimizations. This information is important
4856 /// because we do not want to promote these instructions as CodeGenPrepare
4857 /// will reinsert them later. Thus creating an infinite loop: create/remove.
4858 /// \p PromotedInsts maps the instructions to their type before promotion.
4859 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
4860 const TargetLowering &TLI,
4861 const InstrToOrigTy &PromotedInsts);
4862};
4863
4864} // end anonymous namespace
4865
4866bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
4867 Type *ConsideredExtType,
4868 const InstrToOrigTy &PromotedInsts,
4869 bool IsSExt) {
4870 // The promotion helper does not know how to deal with vector types yet.
4871 // To be able to fix that, we would need to fix the places where we
4872 // statically extend, e.g., constants and such.
4873 if (Inst->getType()->isVectorTy())
4874 return false;
4875
4876 // We can always get through zext.
4877 if (isa<ZExtInst>(Val: Inst))
4878 return true;
4879
4880 // sext(sext) is ok too.
4881 if (IsSExt && isa<SExtInst>(Val: Inst))
4882 return true;
4883
4884 // We can get through binary operator, if it is legal. In other words, the
4885 // binary operator must have a nuw or nsw flag.
4886 if (const auto *BinOp = dyn_cast<BinaryOperator>(Val: Inst))
4887 if (isa<OverflowingBinaryOperator>(Val: BinOp) &&
4888 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
4889 (IsSExt && BinOp->hasNoSignedWrap())))
4890 return true;
4891
4892 // ext(and(opnd, cst)) --> and(ext(opnd), ext(cst))
4893 if ((Inst->getOpcode() == Instruction::And ||
4894 Inst->getOpcode() == Instruction::Or))
4895 return true;
4896
4897 // ext(xor(opnd, cst)) --> xor(ext(opnd), ext(cst))
4898 if (Inst->getOpcode() == Instruction::Xor) {
4899 // Make sure it is not a NOT.
4900 if (const auto *Cst = dyn_cast<ConstantInt>(Val: Inst->getOperand(i: 1)))
4901 if (!Cst->getValue().isAllOnes())
4902 return true;
4903 }
4904
4905 // zext(shrl(opnd, cst)) --> shrl(zext(opnd), zext(cst))
4906 // It may change a poisoned value into a regular value, like
4907 // zext i32 (shrl i8 %val, 12) --> shrl i32 (zext i8 %val), 12
4908 // poisoned value regular value
4909 // It should be OK since undef covers valid value.
4910 if (Inst->getOpcode() == Instruction::LShr && !IsSExt)
4911 return true;
4912
4913 // and(ext(shl(opnd, cst)), cst) --> and(shl(ext(opnd), ext(cst)), cst)
4914 // It may change a poisoned value into a regular value, like
4915 // zext i32 (shl i8 %val, 12) --> shl i32 (zext i8 %val), 12
4916 // poisoned value regular value
4917 // It should be OK since undef covers valid value.
4918 if (Inst->getOpcode() == Instruction::Shl && Inst->hasOneUse()) {
4919 const auto *ExtInst = cast<const Instruction>(Val: *Inst->user_begin());
4920 if (ExtInst->hasOneUse()) {
4921 const auto *AndInst = dyn_cast<const Instruction>(Val: *ExtInst->user_begin());
4922 if (AndInst && AndInst->getOpcode() == Instruction::And) {
4923 const auto *Cst = dyn_cast<ConstantInt>(Val: AndInst->getOperand(i: 1));
4924 if (Cst &&
4925 Cst->getValue().isIntN(N: Inst->getType()->getIntegerBitWidth()))
4926 return true;
4927 }
4928 }
4929 }
4930
4931 // Check if we can do the following simplification.
4932 // ext(trunc(opnd)) --> ext(opnd)
4933 if (!isa<TruncInst>(Val: Inst))
4934 return false;
4935
4936 Value *OpndVal = Inst->getOperand(i: 0);
4937 // Check if we can use this operand in the extension.
4938 // If the type is larger than the result type of the extension, we cannot.
4939 if (!OpndVal->getType()->isIntegerTy() ||
4940 OpndVal->getType()->getIntegerBitWidth() >
4941 ConsideredExtType->getIntegerBitWidth())
4942 return false;
4943
4944 // If the operand of the truncate is not an instruction, we will not have
4945 // any information on the dropped bits.
4946 // (Actually we could for constant but it is not worth the extra logic).
4947 Instruction *Opnd = dyn_cast<Instruction>(Val: OpndVal);
4948 if (!Opnd)
4949 return false;
4950
4951 // Check if the source of the type is narrow enough.
4952 // I.e., check that trunc just drops extended bits of the same kind of
4953 // the extension.
4954 // #1 get the type of the operand and check the kind of the extended bits.
4955 const Type *OpndType = getOrigType(PromotedInsts, Opnd, IsSExt);
4956 if (OpndType)
4957 ;
4958 else if ((IsSExt && isa<SExtInst>(Val: Opnd)) || (!IsSExt && isa<ZExtInst>(Val: Opnd)))
4959 OpndType = Opnd->getOperand(i: 0)->getType();
4960 else
4961 return false;
4962
4963 // #2 check that the truncate just drops extended bits.
4964 return Inst->getType()->getIntegerBitWidth() >=
4965 OpndType->getIntegerBitWidth();
4966}
4967
4968TypePromotionHelper::Action TypePromotionHelper::getAction(
4969 Instruction *Ext, const SetOfInstrs &InsertedInsts,
4970 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
4971 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
4972 "Unexpected instruction type");
4973 Instruction *ExtOpnd = dyn_cast<Instruction>(Val: Ext->getOperand(i: 0));
4974 Type *ExtTy = Ext->getType();
4975 bool IsSExt = isa<SExtInst>(Val: Ext);
4976 // If the operand of the extension is not an instruction, we cannot
4977 // get through.
4978 // If it, check we can get through.
4979 if (!ExtOpnd || !canGetThrough(Inst: ExtOpnd, ConsideredExtType: ExtTy, PromotedInsts, IsSExt))
4980 return nullptr;
4981
4982 // Do not promote if the operand has been added by codegenprepare.
4983 // Otherwise, it means we are undoing an optimization that is likely to be
4984 // redone, thus causing potential infinite loop.
4985 if (isa<TruncInst>(Val: ExtOpnd) && InsertedInsts.count(Ptr: ExtOpnd))
4986 return nullptr;
4987
4988 // SExt or Trunc instructions.
4989 // Return the related handler.
4990 if (isa<SExtInst>(Val: ExtOpnd) || isa<TruncInst>(Val: ExtOpnd) ||
4991 isa<ZExtInst>(Val: ExtOpnd))
4992 return promoteOperandForTruncAndAnyExt;
4993
4994 // Regular instruction.
4995 // Abort early if we will have to insert non-free instructions.
4996 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(FromTy: ExtTy, ToTy: ExtOpnd->getType()))
4997 return nullptr;
4998 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
4999}
5000
5001Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
5002 Instruction *SExt, TypePromotionTransaction &TPT,
5003 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
5004 SmallVectorImpl<Instruction *> *Exts,
5005 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
5006 // By construction, the operand of SExt is an instruction. Otherwise we cannot
5007 // get through it and this method should not be called.
5008 Instruction *SExtOpnd = cast<Instruction>(Val: SExt->getOperand(i: 0));
5009 Value *ExtVal = SExt;
5010 bool HasMergedNonFreeExt = false;
5011 if (isa<ZExtInst>(Val: SExtOpnd)) {
5012 // Replace s|zext(zext(opnd))
5013 // => zext(opnd).
5014 HasMergedNonFreeExt = !TLI.isExtFree(I: SExtOpnd);
5015 Value *ZExt =
5016 TPT.createZExt(Inst: SExt, Opnd: SExtOpnd->getOperand(i: 0), Ty: SExt->getType());
5017 TPT.replaceAllUsesWith(Inst: SExt, New: ZExt);
5018 TPT.eraseInstruction(Inst: SExt);
5019 ExtVal = ZExt;
5020 } else {
5021 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
5022 // => z|sext(opnd).
5023 TPT.setOperand(Inst: SExt, Idx: 0, NewVal: SExtOpnd->getOperand(i: 0));
5024 }
5025 CreatedInstsCost = 0;
5026
5027 // Remove dead code.
5028 if (SExtOpnd->use_empty())
5029 TPT.eraseInstruction(Inst: SExtOpnd);
5030
5031 // Check if the extension is still needed.
5032 Instruction *ExtInst = dyn_cast<Instruction>(Val: ExtVal);
5033 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(i: 0)->getType()) {
5034 if (ExtInst) {
5035 if (Exts)
5036 Exts->push_back(Elt: ExtInst);
5037 CreatedInstsCost = !TLI.isExtFree(I: ExtInst) && !HasMergedNonFreeExt;
5038 }
5039 return ExtVal;
5040 }
5041
5042 // At this point we have: ext ty opnd to ty.
5043 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
5044 Value *NextVal = ExtInst->getOperand(i: 0);
5045 TPT.eraseInstruction(Inst: ExtInst, NewVal: NextVal);
5046 return NextVal;
5047}
5048
5049Value *TypePromotionHelper::promoteOperandForOther(
5050 Instruction *Ext, TypePromotionTransaction &TPT,
5051 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
5052 SmallVectorImpl<Instruction *> *Exts,
5053 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
5054 bool IsSExt) {
5055 // By construction, the operand of Ext is an instruction. Otherwise we cannot
5056 // get through it and this method should not be called.
5057 Instruction *ExtOpnd = cast<Instruction>(Val: Ext->getOperand(i: 0));
5058 CreatedInstsCost = 0;
5059 if (!ExtOpnd->hasOneUse()) {
5060 // ExtOpnd will be promoted.
5061 // All its uses, but Ext, will need to use a truncated value of the
5062 // promoted version.
5063 // Create the truncate now.
5064 Value *Trunc = TPT.createTrunc(Opnd: Ext, Ty: ExtOpnd->getType());
5065 if (Instruction *ITrunc = dyn_cast<Instruction>(Val: Trunc)) {
5066 // Insert it just after the definition.
5067 ITrunc->moveAfter(MovePos: ExtOpnd);
5068 if (Truncs)
5069 Truncs->push_back(Elt: ITrunc);
5070 }
5071
5072 TPT.replaceAllUsesWith(Inst: ExtOpnd, New: Trunc);
5073 // Restore the operand of Ext (which has been replaced by the previous call
5074 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
5075 TPT.setOperand(Inst: Ext, Idx: 0, NewVal: ExtOpnd);
5076 }
5077
5078 // Get through the Instruction:
5079 // 1. Update its type.
5080 // 2. Replace the uses of Ext by Inst.
5081 // 3. Extend each operand that needs to be extended.
5082
5083 // Remember the original type of the instruction before promotion.
5084 // This is useful to know that the high bits are sign extended bits.
5085 addPromotedInst(PromotedInsts, ExtOpnd, IsSExt);
5086 // Step #1.
5087 TPT.mutateType(Inst: ExtOpnd, NewTy: Ext->getType());
5088 // Step #2.
5089 TPT.replaceAllUsesWith(Inst: Ext, New: ExtOpnd);
5090 // Step #3.
5091 LLVM_DEBUG(dbgs() << "Propagate Ext to operands\n");
5092 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
5093 ++OpIdx) {
5094 LLVM_DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
5095 if (ExtOpnd->getOperand(i: OpIdx)->getType() == Ext->getType() ||
5096 !shouldExtOperand(Inst: ExtOpnd, OpIdx)) {
5097 LLVM_DEBUG(dbgs() << "No need to propagate\n");
5098 continue;
5099 }
5100 // Check if we can statically extend the operand.
5101 Value *Opnd = ExtOpnd->getOperand(i: OpIdx);
5102 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Val: Opnd)) {
5103 LLVM_DEBUG(dbgs() << "Statically extend\n");
5104 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
5105 APInt CstVal = IsSExt ? Cst->getValue().sext(width: BitWidth)
5106 : Cst->getValue().zext(width: BitWidth);
5107 TPT.setOperand(Inst: ExtOpnd, Idx: OpIdx, NewVal: ConstantInt::get(Ty: Ext->getType(), V: CstVal));
5108 continue;
5109 }
5110 // UndefValue are typed, so we have to statically sign extend them.
5111 if (isa<UndefValue>(Val: Opnd)) {
5112 LLVM_DEBUG(dbgs() << "Statically extend\n");
5113 TPT.setOperand(Inst: ExtOpnd, Idx: OpIdx, NewVal: UndefValue::get(T: Ext->getType()));
5114 continue;
5115 }
5116
5117 // Otherwise we have to explicitly sign extend the operand.
5118 Value *ValForExtOpnd = IsSExt
5119 ? TPT.createSExt(Inst: ExtOpnd, Opnd, Ty: Ext->getType())
5120 : TPT.createZExt(Inst: ExtOpnd, Opnd, Ty: Ext->getType());
5121 TPT.setOperand(Inst: ExtOpnd, Idx: OpIdx, NewVal: ValForExtOpnd);
5122 Instruction *InstForExtOpnd = dyn_cast<Instruction>(Val: ValForExtOpnd);
5123 if (!InstForExtOpnd)
5124 continue;
5125
5126 if (Exts)
5127 Exts->push_back(Elt: InstForExtOpnd);
5128
5129 CreatedInstsCost += !TLI.isExtFree(I: InstForExtOpnd);
5130 }
5131 LLVM_DEBUG(dbgs() << "Extension is useless now\n");
5132 TPT.eraseInstruction(Inst: Ext);
5133 return ExtOpnd;
5134}
5135
5136/// Check whether or not promoting an instruction to a wider type is profitable.
5137/// \p NewCost gives the cost of extension instructions created by the
5138/// promotion.
5139/// \p OldCost gives the cost of extension instructions before the promotion
5140/// plus the number of instructions that have been
5141/// matched in the addressing mode the promotion.
5142/// \p PromotedOperand is the value that has been promoted.
5143/// \return True if the promotion is profitable, false otherwise.
5144bool AddressingModeMatcher::isPromotionProfitable(
5145 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
5146 LLVM_DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost
5147 << '\n');
5148 // The cost of the new extensions is greater than the cost of the
5149 // old extension plus what we folded.
5150 // This is not profitable.
5151 if (NewCost > OldCost)
5152 return false;
5153 if (NewCost < OldCost)
5154 return true;
5155 // The promotion is neutral but it may help folding the sign extension in
5156 // loads for instance.
5157 // Check that we did not create an illegal instruction.
5158 return isPromotedInstructionLegal(TLI, DL, Val: PromotedOperand);
5159}
5160
5161/// Given an instruction or constant expr, see if we can fold the operation
5162/// into the addressing mode. If so, update the addressing mode and return
5163/// true, otherwise return false without modifying AddrMode.
5164/// If \p MovedAway is not NULL, it contains the information of whether or
5165/// not AddrInst has to be folded into the addressing mode on success.
5166/// If \p MovedAway == true, \p AddrInst will not be part of the addressing
5167/// because it has been moved away.
5168/// Thus AddrInst must not be added in the matched instructions.
5169/// This state can happen when AddrInst is a sext, since it may be moved away.
5170/// Therefore, AddrInst may not be valid when MovedAway is true and it must
5171/// not be referenced anymore.
5172bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
5173 unsigned Depth,
5174 bool *MovedAway) {
5175 // Avoid exponential behavior on extremely deep expression trees.
5176 if (Depth >= 5)
5177 return false;
5178
5179 // By default, all matched instructions stay in place.
5180 if (MovedAway)
5181 *MovedAway = false;
5182
5183 switch (Opcode) {
5184 case Instruction::PtrToInt:
5185 // PtrToInt is always a noop, as we know that the int type is pointer sized.
5186 return matchAddr(Addr: AddrInst->getOperand(i: 0), Depth);
5187 case Instruction::IntToPtr: {
5188 auto AS = AddrInst->getType()->getPointerAddressSpace();
5189 auto PtrTy = MVT::getIntegerVT(BitWidth: DL.getPointerSizeInBits(AS));
5190 // This inttoptr is a no-op if the integer type is pointer sized.
5191 if (TLI.getValueType(DL, Ty: AddrInst->getOperand(i: 0)->getType()) == PtrTy)
5192 return matchAddr(Addr: AddrInst->getOperand(i: 0), Depth);
5193 return false;
5194 }
5195 case Instruction::BitCast:
5196 // BitCast is always a noop, and we can handle it as long as it is
5197 // int->int or pointer->pointer (we don't want int<->fp or something).
5198 if (AddrInst->getOperand(i: 0)->getType()->isIntOrPtrTy() &&
5199 // Don't touch identity bitcasts. These were probably put here by LSR,
5200 // and we don't want to mess around with them. Assume it knows what it
5201 // is doing.
5202 AddrInst->getOperand(i: 0)->getType() != AddrInst->getType())
5203 return matchAddr(Addr: AddrInst->getOperand(i: 0), Depth);
5204 return false;
5205 case Instruction::AddrSpaceCast: {
5206 unsigned SrcAS =
5207 AddrInst->getOperand(i: 0)->getType()->getPointerAddressSpace();
5208 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
5209 if (TLI.getTargetMachine().isNoopAddrSpaceCast(SrcAS, DestAS))
5210 return matchAddr(Addr: AddrInst->getOperand(i: 0), Depth);
5211 return false;
5212 }
5213 case Instruction::Add: {
5214 // Check to see if we can merge in one operand, then the other. If so, we
5215 // win.
5216 ExtAddrMode BackupAddrMode = AddrMode;
5217 unsigned OldSize = AddrModeInsts.size();
5218 // Start a transaction at this point.
5219 // The LHS may match but not the RHS.
5220 // Therefore, we need a higher level restoration point to undo partially
5221 // matched operation.
5222 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5223 TPT.getRestorationPoint();
5224
5225 // Try to match an integer constant second to increase its chance of ending
5226 // up in `BaseOffs`, resp. decrease its chance of ending up in `BaseReg`.
5227 int First = 0, Second = 1;
5228 if (isa<ConstantInt>(Val: AddrInst->getOperand(i: First))
5229 && !isa<ConstantInt>(Val: AddrInst->getOperand(i: Second)))
5230 std::swap(a&: First, b&: Second);
5231 AddrMode.InBounds = false;
5232 if (matchAddr(Addr: AddrInst->getOperand(i: First), Depth: Depth + 1) &&
5233 matchAddr(Addr: AddrInst->getOperand(i: Second), Depth: Depth + 1))
5234 return true;
5235
5236 // Restore the old addr mode info.
5237 AddrMode = BackupAddrMode;
5238 AddrModeInsts.resize(N: OldSize);
5239 TPT.rollback(Point: LastKnownGood);
5240
5241 // Otherwise this was over-aggressive. Try merging operands in the opposite
5242 // order.
5243 if (matchAddr(Addr: AddrInst->getOperand(i: Second), Depth: Depth + 1) &&
5244 matchAddr(Addr: AddrInst->getOperand(i: First), Depth: Depth + 1))
5245 return true;
5246
5247 // Otherwise we definitely can't merge the ADD in.
5248 AddrMode = BackupAddrMode;
5249 AddrModeInsts.resize(N: OldSize);
5250 TPT.rollback(Point: LastKnownGood);
5251 break;
5252 }
5253 // case Instruction::Or:
5254 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
5255 // break;
5256 case Instruction::Mul:
5257 case Instruction::Shl: {
5258 // Can only handle X*C and X << C.
5259 AddrMode.InBounds = false;
5260 ConstantInt *RHS = dyn_cast<ConstantInt>(Val: AddrInst->getOperand(i: 1));
5261 if (!RHS || RHS->getBitWidth() > 64)
5262 return false;
5263 int64_t Scale = Opcode == Instruction::Shl
5264 ? 1LL << RHS->getLimitedValue(Limit: RHS->getBitWidth() - 1)
5265 : RHS->getSExtValue();
5266
5267 return matchScaledValue(ScaleReg: AddrInst->getOperand(i: 0), Scale, Depth);
5268 }
5269 case Instruction::GetElementPtr: {
5270 // Scan the GEP. We check it if it contains constant offsets and at most
5271 // one variable offset.
5272 int VariableOperand = -1;
5273 unsigned VariableScale = 0;
5274
5275 int64_t ConstantOffset = 0;
5276 gep_type_iterator GTI = gep_type_begin(GEP: AddrInst);
5277 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
5278 if (StructType *STy = GTI.getStructTypeOrNull()) {
5279 const StructLayout *SL = DL.getStructLayout(Ty: STy);
5280 unsigned Idx =
5281 cast<ConstantInt>(Val: AddrInst->getOperand(i))->getZExtValue();
5282 ConstantOffset += SL->getElementOffset(Idx);
5283 } else {
5284 TypeSize TS = GTI.getSequentialElementStride(DL);
5285 if (TS.isNonZero()) {
5286 // The optimisations below currently only work for fixed offsets.
5287 if (TS.isScalable())
5288 return false;
5289 int64_t TypeSize = TS.getFixedValue();
5290 if (ConstantInt *CI =
5291 dyn_cast<ConstantInt>(Val: AddrInst->getOperand(i))) {
5292 const APInt &CVal = CI->getValue();
5293 if (CVal.getSignificantBits() <= 64) {
5294 ConstantOffset += CVal.getSExtValue() * TypeSize;
5295 continue;
5296 }
5297 }
5298 // We only allow one variable index at the moment.
5299 if (VariableOperand != -1)
5300 return false;
5301
5302 // Remember the variable index.
5303 VariableOperand = i;
5304 VariableScale = TypeSize;
5305 }
5306 }
5307 }
5308
5309 // A common case is for the GEP to only do a constant offset. In this case,
5310 // just add it to the disp field and check validity.
5311 if (VariableOperand == -1) {
5312 AddrMode.BaseOffs += ConstantOffset;
5313 if (matchAddr(Addr: AddrInst->getOperand(i: 0), Depth: Depth + 1)) {
5314 if (!cast<GEPOperator>(Val: AddrInst)->isInBounds())
5315 AddrMode.InBounds = false;
5316 return true;
5317 }
5318 AddrMode.BaseOffs -= ConstantOffset;
5319
5320 if (EnableGEPOffsetSplit && isa<GetElementPtrInst>(Val: AddrInst) &&
5321 TLI.shouldConsiderGEPOffsetSplit() && Depth == 0 &&
5322 ConstantOffset > 0) {
5323 // Record GEPs with non-zero offsets as candidates for splitting in
5324 // the event that the offset cannot fit into the r+i addressing mode.
5325 // Simple and common case that only one GEP is used in calculating the
5326 // address for the memory access.
5327 Value *Base = AddrInst->getOperand(i: 0);
5328 auto *BaseI = dyn_cast<Instruction>(Val: Base);
5329 auto *GEP = cast<GetElementPtrInst>(Val: AddrInst);
5330 if (isa<Argument>(Val: Base) || isa<GlobalValue>(Val: Base) ||
5331 (BaseI && !isa<CastInst>(Val: BaseI) &&
5332 !isa<GetElementPtrInst>(Val: BaseI))) {
5333 // Make sure the parent block allows inserting non-PHI instructions
5334 // before the terminator.
5335 BasicBlock *Parent = BaseI ? BaseI->getParent()
5336 : &GEP->getFunction()->getEntryBlock();
5337 if (!Parent->getTerminator()->isEHPad())
5338 LargeOffsetGEP = std::make_pair(x&: GEP, y&: ConstantOffset);
5339 }
5340 }
5341
5342 return false;
5343 }
5344
5345 // Save the valid addressing mode in case we can't match.
5346 ExtAddrMode BackupAddrMode = AddrMode;
5347 unsigned OldSize = AddrModeInsts.size();
5348
5349 // See if the scale and offset amount is valid for this target.
5350 AddrMode.BaseOffs += ConstantOffset;
5351 if (!cast<GEPOperator>(Val: AddrInst)->isInBounds())
5352 AddrMode.InBounds = false;
5353
5354 // Match the base operand of the GEP.
5355 if (!matchAddr(Addr: AddrInst->getOperand(i: 0), Depth: Depth + 1)) {
5356 // If it couldn't be matched, just stuff the value in a register.
5357 if (AddrMode.HasBaseReg) {
5358 AddrMode = BackupAddrMode;
5359 AddrModeInsts.resize(N: OldSize);
5360 return false;
5361 }
5362 AddrMode.HasBaseReg = true;
5363 AddrMode.BaseReg = AddrInst->getOperand(i: 0);
5364 }
5365
5366 // Match the remaining variable portion of the GEP.
5367 if (!matchScaledValue(ScaleReg: AddrInst->getOperand(i: VariableOperand), Scale: VariableScale,
5368 Depth)) {
5369 // If it couldn't be matched, try stuffing the base into a register
5370 // instead of matching it, and retrying the match of the scale.
5371 AddrMode = BackupAddrMode;
5372 AddrModeInsts.resize(N: OldSize);
5373 if (AddrMode.HasBaseReg)
5374 return false;
5375 AddrMode.HasBaseReg = true;
5376 AddrMode.BaseReg = AddrInst->getOperand(i: 0);
5377 AddrMode.BaseOffs += ConstantOffset;
5378 if (!matchScaledValue(ScaleReg: AddrInst->getOperand(i: VariableOperand),
5379 Scale: VariableScale, Depth)) {
5380 // If even that didn't work, bail.
5381 AddrMode = BackupAddrMode;
5382 AddrModeInsts.resize(N: OldSize);
5383 return false;
5384 }
5385 }
5386
5387 return true;
5388 }
5389 case Instruction::SExt:
5390 case Instruction::ZExt: {
5391 Instruction *Ext = dyn_cast<Instruction>(Val: AddrInst);
5392 if (!Ext)
5393 return false;
5394
5395 // Try to move this ext out of the way of the addressing mode.
5396 // Ask for a method for doing so.
5397 TypePromotionHelper::Action TPH =
5398 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
5399 if (!TPH)
5400 return false;
5401
5402 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5403 TPT.getRestorationPoint();
5404 unsigned CreatedInstsCost = 0;
5405 unsigned ExtCost = !TLI.isExtFree(I: Ext);
5406 Value *PromotedOperand =
5407 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
5408 // SExt has been moved away.
5409 // Thus either it will be rematched later in the recursive calls or it is
5410 // gone. Anyway, we must not fold it into the addressing mode at this point.
5411 // E.g.,
5412 // op = add opnd, 1
5413 // idx = ext op
5414 // addr = gep base, idx
5415 // is now:
5416 // promotedOpnd = ext opnd <- no match here
5417 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
5418 // addr = gep base, op <- match
5419 if (MovedAway)
5420 *MovedAway = true;
5421
5422 assert(PromotedOperand &&
5423 "TypePromotionHelper should have filtered out those cases");
5424
5425 ExtAddrMode BackupAddrMode = AddrMode;
5426 unsigned OldSize = AddrModeInsts.size();
5427
5428 if (!matchAddr(Addr: PromotedOperand, Depth) ||
5429 // The total of the new cost is equal to the cost of the created
5430 // instructions.
5431 // The total of the old cost is equal to the cost of the extension plus
5432 // what we have saved in the addressing mode.
5433 !isPromotionProfitable(NewCost: CreatedInstsCost,
5434 OldCost: ExtCost + (AddrModeInsts.size() - OldSize),
5435 PromotedOperand)) {
5436 AddrMode = BackupAddrMode;
5437 AddrModeInsts.resize(N: OldSize);
5438 LLVM_DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
5439 TPT.rollback(Point: LastKnownGood);
5440 return false;
5441 }
5442
5443 // SExt has been deleted. Make sure it is not referenced by the AddrMode.
5444 AddrMode.replaceWith(From: Ext, To: PromotedOperand);
5445 return true;
5446 }
5447 case Instruction::Call:
5448 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: AddrInst)) {
5449 if (II->getIntrinsicID() == Intrinsic::threadlocal_address) {
5450 GlobalValue &GV = cast<GlobalValue>(Val&: *II->getArgOperand(i: 0));
5451 if (TLI.addressingModeSupportsTLS(GV))
5452 return matchAddr(Addr: AddrInst->getOperand(i: 0), Depth);
5453 }
5454 }
5455 break;
5456 }
5457 return false;
5458}
5459
5460/// If we can, try to add the value of 'Addr' into the current addressing mode.
5461/// If Addr can't be added to AddrMode this returns false and leaves AddrMode
5462/// unmodified. This assumes that Addr is either a pointer type or intptr_t
5463/// for the target.
5464///
5465bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
5466 // Start a transaction at this point that we will rollback if the matching
5467 // fails.
5468 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5469 TPT.getRestorationPoint();
5470 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: Addr)) {
5471 if (CI->getValue().isSignedIntN(N: 64)) {
5472 // Check if the addition would result in a signed overflow.
5473 int64_t Result;
5474 bool Overflow =
5475 AddOverflow(X: AddrMode.BaseOffs, Y: CI->getSExtValue(), Result);
5476 if (!Overflow) {
5477 // Fold in immediates if legal for the target.
5478 AddrMode.BaseOffs = Result;
5479 if (TLI.isLegalAddressingMode(DL, AM: AddrMode, Ty: AccessTy, AddrSpace))
5480 return true;
5481 AddrMode.BaseOffs -= CI->getSExtValue();
5482 }
5483 }
5484 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Val: Addr)) {
5485 // If this is a global variable, try to fold it into the addressing mode.
5486 if (!AddrMode.BaseGV) {
5487 AddrMode.BaseGV = GV;
5488 if (TLI.isLegalAddressingMode(DL, AM: AddrMode, Ty: AccessTy, AddrSpace))
5489 return true;
5490 AddrMode.BaseGV = nullptr;
5491 }
5492 } else if (Instruction *I = dyn_cast<Instruction>(Val: Addr)) {
5493 ExtAddrMode BackupAddrMode = AddrMode;
5494 unsigned OldSize = AddrModeInsts.size();
5495
5496 // Check to see if it is possible to fold this operation.
5497 bool MovedAway = false;
5498 if (matchOperationAddr(AddrInst: I, Opcode: I->getOpcode(), Depth, MovedAway: &MovedAway)) {
5499 // This instruction may have been moved away. If so, there is nothing
5500 // to check here.
5501 if (MovedAway)
5502 return true;
5503 // Okay, it's possible to fold this. Check to see if it is actually
5504 // *profitable* to do so. We use a simple cost model to avoid increasing
5505 // register pressure too much.
5506 if (I->hasOneUse() ||
5507 isProfitableToFoldIntoAddressingMode(I, AMBefore&: BackupAddrMode, AMAfter&: AddrMode)) {
5508 AddrModeInsts.push_back(Elt: I);
5509 return true;
5510 }
5511
5512 // It isn't profitable to do this, roll back.
5513 AddrMode = BackupAddrMode;
5514 AddrModeInsts.resize(N: OldSize);
5515 TPT.rollback(Point: LastKnownGood);
5516 }
5517 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Val: Addr)) {
5518 if (matchOperationAddr(AddrInst: CE, Opcode: CE->getOpcode(), Depth))
5519 return true;
5520 TPT.rollback(Point: LastKnownGood);
5521 } else if (isa<ConstantPointerNull>(Val: Addr)) {
5522 // Null pointer gets folded without affecting the addressing mode.
5523 return true;
5524 }
5525
5526 // Worse case, the target should support [reg] addressing modes. :)
5527 if (!AddrMode.HasBaseReg) {
5528 AddrMode.HasBaseReg = true;
5529 AddrMode.BaseReg = Addr;
5530 // Still check for legality in case the target supports [imm] but not [i+r].
5531 if (TLI.isLegalAddressingMode(DL, AM: AddrMode, Ty: AccessTy, AddrSpace))
5532 return true;
5533 AddrMode.HasBaseReg = false;
5534 AddrMode.BaseReg = nullptr;
5535 }
5536
5537 // If the base register is already taken, see if we can do [r+r].
5538 if (AddrMode.Scale == 0) {
5539 AddrMode.Scale = 1;
5540 AddrMode.ScaledReg = Addr;
5541 if (TLI.isLegalAddressingMode(DL, AM: AddrMode, Ty: AccessTy, AddrSpace))
5542 return true;
5543 AddrMode.Scale = 0;
5544 AddrMode.ScaledReg = nullptr;
5545 }
5546 // Couldn't match.
5547 TPT.rollback(Point: LastKnownGood);
5548 return false;
5549}
5550
5551/// Check to see if all uses of OpVal by the specified inline asm call are due
5552/// to memory operands. If so, return true, otherwise return false.
5553static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
5554 const TargetLowering &TLI,
5555 const TargetRegisterInfo &TRI) {
5556 const Function *F = CI->getFunction();
5557 TargetLowering::AsmOperandInfoVector TargetConstraints =
5558 TLI.ParseConstraints(DL: F->getDataLayout(), TRI: &TRI, Call: *CI);
5559
5560 for (TargetLowering::AsmOperandInfo &OpInfo : TargetConstraints) {
5561 // Compute the constraint code and ConstraintType to use.
5562 TLI.ComputeConstraintToUse(OpInfo, Op: SDValue());
5563
5564 // If this asm operand is our Value*, and if it isn't an indirect memory
5565 // operand, we can't fold it! TODO: Also handle C_Address?
5566 if (OpInfo.CallOperandVal == OpVal &&
5567 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
5568 !OpInfo.isIndirect))
5569 return false;
5570 }
5571
5572 return true;
5573}
5574
5575/// Recursively walk all the uses of I until we find a memory use.
5576/// If we find an obviously non-foldable instruction, return true.
5577/// Add accessed addresses and types to MemoryUses.
5578static bool FindAllMemoryUses(
5579 Instruction *I, SmallVectorImpl<std::pair<Use *, Type *>> &MemoryUses,
5580 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
5581 const TargetRegisterInfo &TRI, bool OptSize, ProfileSummaryInfo *PSI,
5582 BlockFrequencyInfo *BFI, unsigned &SeenInsts) {
5583 // If we already considered this instruction, we're done.
5584 if (!ConsideredInsts.insert(Ptr: I).second)
5585 return false;
5586
5587 // If this is an obviously unfoldable instruction, bail out.
5588 if (!MightBeFoldableInst(I))
5589 return true;
5590
5591 // Loop over all the uses, recursively processing them.
5592 for (Use &U : I->uses()) {
5593 // Conservatively return true if we're seeing a large number or a deep chain
5594 // of users. This avoids excessive compilation times in pathological cases.
5595 if (SeenInsts++ >= MaxAddressUsersToScan)
5596 return true;
5597
5598 Instruction *UserI = cast<Instruction>(Val: U.getUser());
5599 if (LoadInst *LI = dyn_cast<LoadInst>(Val: UserI)) {
5600 MemoryUses.push_back(Elt: {&U, LI->getType()});
5601 continue;
5602 }
5603
5604 if (StoreInst *SI = dyn_cast<StoreInst>(Val: UserI)) {
5605 if (U.getOperandNo() != StoreInst::getPointerOperandIndex())
5606 return true; // Storing addr, not into addr.
5607 MemoryUses.push_back(Elt: {&U, SI->getValueOperand()->getType()});
5608 continue;
5609 }
5610
5611 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: UserI)) {
5612 if (U.getOperandNo() != AtomicRMWInst::getPointerOperandIndex())
5613 return true; // Storing addr, not into addr.
5614 MemoryUses.push_back(Elt: {&U, RMW->getValOperand()->getType()});
5615 continue;
5616 }
5617
5618 if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Val: UserI)) {
5619 if (U.getOperandNo() != AtomicCmpXchgInst::getPointerOperandIndex())
5620 return true; // Storing addr, not into addr.
5621 MemoryUses.push_back(Elt: {&U, CmpX->getCompareOperand()->getType()});
5622 continue;
5623 }
5624
5625 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: UserI)) {
5626 SmallVector<Value *, 2> PtrOps;
5627 Type *AccessTy;
5628 if (!TLI.getAddrModeArguments(II, PtrOps, AccessTy))
5629 return true;
5630
5631 if (!find(Range&: PtrOps, Val: U.get()))
5632 return true;
5633
5634 MemoryUses.push_back(Elt: {&U, AccessTy});
5635 continue;
5636 }
5637
5638 if (CallInst *CI = dyn_cast<CallInst>(Val: UserI)) {
5639 if (CI->hasFnAttr(Kind: Attribute::Cold)) {
5640 // If this is a cold call, we can sink the addressing calculation into
5641 // the cold path. See optimizeCallInst
5642 if (!llvm::shouldOptimizeForSize(BB: CI->getParent(), PSI, BFI))
5643 continue;
5644 }
5645
5646 InlineAsm *IA = dyn_cast<InlineAsm>(Val: CI->getCalledOperand());
5647 if (!IA)
5648 return true;
5649
5650 // If this is a memory operand, we're cool, otherwise bail out.
5651 if (!IsOperandAMemoryOperand(CI, IA, OpVal: I, TLI, TRI))
5652 return true;
5653 continue;
5654 }
5655
5656 if (FindAllMemoryUses(I: UserI, MemoryUses, ConsideredInsts, TLI, TRI, OptSize,
5657 PSI, BFI, SeenInsts))
5658 return true;
5659 }
5660
5661 return false;
5662}
5663
5664static bool FindAllMemoryUses(
5665 Instruction *I, SmallVectorImpl<std::pair<Use *, Type *>> &MemoryUses,
5666 const TargetLowering &TLI, const TargetRegisterInfo &TRI, bool OptSize,
5667 ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) {
5668 unsigned SeenInsts = 0;
5669 SmallPtrSet<Instruction *, 16> ConsideredInsts;
5670 return FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI, OptSize,
5671 PSI, BFI, SeenInsts);
5672}
5673
5674
5675/// Return true if Val is already known to be live at the use site that we're
5676/// folding it into. If so, there is no cost to include it in the addressing
5677/// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
5678/// instruction already.
5679bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,
5680 Value *KnownLive1,
5681 Value *KnownLive2) {
5682 // If Val is either of the known-live values, we know it is live!
5683 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
5684 return true;
5685
5686 // All values other than instructions and arguments (e.g. constants) are live.
5687 if (!isa<Instruction>(Val) && !isa<Argument>(Val))
5688 return true;
5689
5690 // If Val is a constant sized alloca in the entry block, it is live, this is
5691 // true because it is just a reference to the stack/frame pointer, which is
5692 // live for the whole function.
5693 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
5694 if (AI->isStaticAlloca())
5695 return true;
5696
5697 // Check to see if this value is already used in the memory instruction's
5698 // block. If so, it's already live into the block at the very least, so we
5699 // can reasonably fold it.
5700 return Val->isUsedInBasicBlock(BB: MemoryInst->getParent());
5701}
5702
5703/// It is possible for the addressing mode of the machine to fold the specified
5704/// instruction into a load or store that ultimately uses it.
5705/// However, the specified instruction has multiple uses.
5706/// Given this, it may actually increase register pressure to fold it
5707/// into the load. For example, consider this code:
5708///
5709/// X = ...
5710/// Y = X+1
5711/// use(Y) -> nonload/store
5712/// Z = Y+1
5713/// load Z
5714///
5715/// In this case, Y has multiple uses, and can be folded into the load of Z
5716/// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
5717/// be live at the use(Y) line. If we don't fold Y into load Z, we use one
5718/// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
5719/// number of computations either.
5720///
5721/// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
5722/// X was live across 'load Z' for other reasons, we actually *would* want to
5723/// fold the addressing mode in the Z case. This would make Y die earlier.
5724bool AddressingModeMatcher::isProfitableToFoldIntoAddressingMode(
5725 Instruction *I, ExtAddrMode &AMBefore, ExtAddrMode &AMAfter) {
5726 if (IgnoreProfitability)
5727 return true;
5728
5729 // AMBefore is the addressing mode before this instruction was folded into it,
5730 // and AMAfter is the addressing mode after the instruction was folded. Get
5731 // the set of registers referenced by AMAfter and subtract out those
5732 // referenced by AMBefore: this is the set of values which folding in this
5733 // address extends the lifetime of.
5734 //
5735 // Note that there are only two potential values being referenced here,
5736 // BaseReg and ScaleReg (global addresses are always available, as are any
5737 // folded immediates).
5738 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
5739
5740 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
5741 // lifetime wasn't extended by adding this instruction.
5742 if (valueAlreadyLiveAtInst(Val: BaseReg, KnownLive1: AMBefore.BaseReg, KnownLive2: AMBefore.ScaledReg))
5743 BaseReg = nullptr;
5744 if (valueAlreadyLiveAtInst(Val: ScaledReg, KnownLive1: AMBefore.BaseReg, KnownLive2: AMBefore.ScaledReg))
5745 ScaledReg = nullptr;
5746
5747 // If folding this instruction (and it's subexprs) didn't extend any live
5748 // ranges, we're ok with it.
5749 if (!BaseReg && !ScaledReg)
5750 return true;
5751
5752 // If all uses of this instruction can have the address mode sunk into them,
5753 // we can remove the addressing mode and effectively trade one live register
5754 // for another (at worst.) In this context, folding an addressing mode into
5755 // the use is just a particularly nice way of sinking it.
5756 SmallVector<std::pair<Use *, Type *>, 16> MemoryUses;
5757 if (FindAllMemoryUses(I, MemoryUses, TLI, TRI, OptSize, PSI, BFI))
5758 return false; // Has a non-memory, non-foldable use!
5759
5760 // Now that we know that all uses of this instruction are part of a chain of
5761 // computation involving only operations that could theoretically be folded
5762 // into a memory use, loop over each of these memory operation uses and see
5763 // if they could *actually* fold the instruction. The assumption is that
5764 // addressing modes are cheap and that duplicating the computation involved
5765 // many times is worthwhile, even on a fastpath. For sinking candidates
5766 // (i.e. cold call sites), this serves as a way to prevent excessive code
5767 // growth since most architectures have some reasonable small and fast way to
5768 // compute an effective address. (i.e LEA on x86)
5769 SmallVector<Instruction *, 32> MatchedAddrModeInsts;
5770 for (const std::pair<Use *, Type *> &Pair : MemoryUses) {
5771 Value *Address = Pair.first->get();
5772 Instruction *UserI = cast<Instruction>(Val: Pair.first->getUser());
5773 Type *AddressAccessTy = Pair.second;
5774 unsigned AS = Address->getType()->getPointerAddressSpace();
5775
5776 // Do a match against the root of this address, ignoring profitability. This
5777 // will tell us if the addressing mode for the memory operation will
5778 // *actually* cover the shared instruction.
5779 ExtAddrMode Result;
5780 std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
5781 0);
5782 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5783 TPT.getRestorationPoint();
5784 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, TRI, LI, getDTFn,
5785 AddressAccessTy, AS, UserI, Result,
5786 InsertedInsts, PromotedInsts, TPT,
5787 LargeOffsetGEP, OptSize, PSI, BFI);
5788 Matcher.IgnoreProfitability = true;
5789 bool Success = Matcher.matchAddr(Addr: Address, Depth: 0);
5790 (void)Success;
5791 assert(Success && "Couldn't select *anything*?");
5792
5793 // The match was to check the profitability, the changes made are not
5794 // part of the original matcher. Therefore, they should be dropped
5795 // otherwise the original matcher will not present the right state.
5796 TPT.rollback(Point: LastKnownGood);
5797
5798 // If the match didn't cover I, then it won't be shared by it.
5799 if (!is_contained(Range&: MatchedAddrModeInsts, Element: I))
5800 return false;
5801
5802 MatchedAddrModeInsts.clear();
5803 }
5804
5805 return true;
5806}
5807
5808/// Return true if the specified values are defined in a
5809/// different basic block than BB.
5810static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
5811 if (Instruction *I = dyn_cast<Instruction>(Val: V))
5812 return I->getParent() != BB;
5813 return false;
5814}
5815
5816// Find an insert position of Addr for MemoryInst. We can't guarantee MemoryInst
5817// is the first instruction that will use Addr. So we need to find the first
5818// user of Addr in current BB.
5819static BasicBlock::iterator findInsertPos(Value *Addr, Instruction *MemoryInst,
5820 Value *SunkAddr) {
5821 if (Addr->hasOneUse())
5822 return MemoryInst->getIterator();
5823
5824 // We already have a SunkAddr in current BB, but we may need to insert cast
5825 // instruction after it.
5826 if (SunkAddr) {
5827 if (Instruction *AddrInst = dyn_cast<Instruction>(Val: SunkAddr))
5828 return std::next(x: AddrInst->getIterator());
5829 }
5830
5831 // Find the first user of Addr in current BB.
5832 Instruction *Earliest = MemoryInst;
5833 for (User *U : Addr->users()) {
5834 Instruction *UserInst = dyn_cast<Instruction>(Val: U);
5835 if (UserInst && UserInst->getParent() == MemoryInst->getParent()) {
5836 if (isa<PHINode>(Val: UserInst) || UserInst->isDebugOrPseudoInst())
5837 continue;
5838 if (UserInst->comesBefore(Other: Earliest))
5839 Earliest = UserInst;
5840 }
5841 }
5842 return Earliest->getIterator();
5843}
5844
5845/// Sink addressing mode computation immediate before MemoryInst if doing so
5846/// can be done without increasing register pressure. The need for the
5847/// register pressure constraint means this can end up being an all or nothing
5848/// decision for all uses of the same addressing computation.
5849///
5850/// Load and Store Instructions often have addressing modes that can do
5851/// significant amounts of computation. As such, instruction selection will try
5852/// to get the load or store to do as much computation as possible for the
5853/// program. The problem is that isel can only see within a single block. As
5854/// such, we sink as much legal addressing mode work into the block as possible.
5855///
5856/// This method is used to optimize both load/store and inline asms with memory
5857/// operands. It's also used to sink addressing computations feeding into cold
5858/// call sites into their (cold) basic block.
5859///
5860/// The motivation for handling sinking into cold blocks is that doing so can
5861/// both enable other address mode sinking (by satisfying the register pressure
5862/// constraint above), and reduce register pressure globally (by removing the
5863/// addressing mode computation from the fast path entirely.).
5864bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
5865 Type *AccessTy, unsigned AddrSpace) {
5866 Value *Repl = Addr;
5867
5868 // Try to collapse single-value PHI nodes. This is necessary to undo
5869 // unprofitable PRE transformations.
5870 SmallVector<Value *, 8> worklist;
5871 SmallPtrSet<Value *, 16> Visited;
5872 worklist.push_back(Elt: Addr);
5873
5874 // Use a worklist to iteratively look through PHI and select nodes, and
5875 // ensure that the addressing mode obtained from the non-PHI/select roots of
5876 // the graph are compatible.
5877 bool PhiOrSelectSeen = false;
5878 SmallVector<Instruction *, 16> AddrModeInsts;
5879 AddressingModeCombiner AddrModes(*DL, Addr);
5880 TypePromotionTransaction TPT(RemovedInsts);
5881 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5882 TPT.getRestorationPoint();
5883 while (!worklist.empty()) {
5884 Value *V = worklist.pop_back_val();
5885
5886 // We allow traversing cyclic Phi nodes.
5887 // In case of success after this loop we ensure that traversing through
5888 // Phi nodes ends up with all cases to compute address of the form
5889 // BaseGV + Base + Scale * Index + Offset
5890 // where Scale and Offset are constans and BaseGV, Base and Index
5891 // are exactly the same Values in all cases.
5892 // It means that BaseGV, Scale and Offset dominate our memory instruction
5893 // and have the same value as they had in address computation represented
5894 // as Phi. So we can safely sink address computation to memory instruction.
5895 if (!Visited.insert(Ptr: V).second)
5896 continue;
5897
5898 // For a PHI node, push all of its incoming values.
5899 if (PHINode *P = dyn_cast<PHINode>(Val: V)) {
5900 append_range(C&: worklist, R: P->incoming_values());
5901 PhiOrSelectSeen = true;
5902 continue;
5903 }
5904 // Similar for select.
5905 if (SelectInst *SI = dyn_cast<SelectInst>(Val: V)) {
5906 worklist.push_back(Elt: SI->getFalseValue());
5907 worklist.push_back(Elt: SI->getTrueValue());
5908 PhiOrSelectSeen = true;
5909 continue;
5910 }
5911
5912 // For non-PHIs, determine the addressing mode being computed. Note that
5913 // the result may differ depending on what other uses our candidate
5914 // addressing instructions might have.
5915 AddrModeInsts.clear();
5916 std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
5917 0);
5918 // Defer the query (and possible computation of) the dom tree to point of
5919 // actual use. It's expected that most address matches don't actually need
5920 // the domtree.
5921 auto getDTFn = [this]() -> const DominatorTree & { return getDT(); };
5922 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
5923 V, AccessTy, AS: AddrSpace, MemoryInst, AddrModeInsts, TLI: *TLI, LI: *LI, getDTFn,
5924 TRI: *TRI, InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP, OptSize, PSI,
5925 BFI);
5926
5927 GetElementPtrInst *GEP = LargeOffsetGEP.first;
5928 if (GEP && !NewGEPBases.count(V: GEP)) {
5929 // If splitting the underlying data structure can reduce the offset of a
5930 // GEP, collect the GEP. Skip the GEPs that are the new bases of
5931 // previously split data structures.
5932 LargeOffsetGEPMap[GEP->getPointerOperand()].push_back(Elt: LargeOffsetGEP);
5933 LargeOffsetGEPID.insert(KV: std::make_pair(x&: GEP, y: LargeOffsetGEPID.size()));
5934 }
5935
5936 NewAddrMode.OriginalValue = V;
5937 if (!AddrModes.addNewAddrMode(NewAddrMode))
5938 break;
5939 }
5940
5941 // Try to combine the AddrModes we've collected. If we couldn't collect any,
5942 // or we have multiple but either couldn't combine them or combining them
5943 // wouldn't do anything useful, bail out now.
5944 if (!AddrModes.combineAddrModes()) {
5945 TPT.rollback(Point: LastKnownGood);
5946 return false;
5947 }
5948 bool Modified = TPT.commit();
5949
5950 // Get the combined AddrMode (or the only AddrMode, if we only had one).
5951 ExtAddrMode AddrMode = AddrModes.getAddrMode();
5952
5953 // If all the instructions matched are already in this BB, don't do anything.
5954 // If we saw a Phi node then it is not local definitely, and if we saw a
5955 // select then we want to push the address calculation past it even if it's
5956 // already in this BB.
5957 if (!PhiOrSelectSeen && none_of(Range&: AddrModeInsts, P: [&](Value *V) {
5958 return IsNonLocalValue(V, BB: MemoryInst->getParent());
5959 })) {
5960 LLVM_DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode
5961 << "\n");
5962 return Modified;
5963 }
5964
5965 // Now that we determined the addressing expression we want to use and know
5966 // that we have to sink it into this block. Check to see if we have already
5967 // done this for some other load/store instr in this block. If so, reuse
5968 // the computation. Before attempting reuse, check if the address is valid
5969 // as it may have been erased.
5970
5971 WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
5972
5973 Value *SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
5974 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
5975
5976 // The current BB may be optimized multiple times, we can't guarantee the
5977 // reuse of Addr happens later, call findInsertPos to find an appropriate
5978 // insert position.
5979 auto InsertPos = findInsertPos(Addr, MemoryInst, SunkAddr);
5980
5981 // TODO: Adjust insert point considering (Base|Scaled)Reg if possible.
5982 if (!SunkAddr) {
5983 auto &DT = getDT();
5984 if ((AddrMode.BaseReg && !DT.dominates(Def: AddrMode.BaseReg, User: &*InsertPos)) ||
5985 (AddrMode.ScaledReg && !DT.dominates(Def: AddrMode.ScaledReg, User: &*InsertPos)))
5986 return Modified;
5987 }
5988
5989 IRBuilder<> Builder(MemoryInst->getParent(), InsertPos);
5990
5991 if (SunkAddr) {
5992 LLVM_DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode
5993 << " for " << *MemoryInst << "\n");
5994 if (SunkAddr->getType() != Addr->getType()) {
5995 if (SunkAddr->getType()->getPointerAddressSpace() !=
5996 Addr->getType()->getPointerAddressSpace() &&
5997 !DL->isNonIntegralPointerType(Ty: Addr->getType())) {
5998 // There are two reasons the address spaces might not match: a no-op
5999 // addrspacecast, or a ptrtoint/inttoptr pair. Either way, we emit a
6000 // ptrtoint/inttoptr pair to ensure we match the original semantics.
6001 // TODO: allow bitcast between different address space pointers with the
6002 // same size.
6003 SunkAddr = Builder.CreatePtrToInt(V: SunkAddr, DestTy: IntPtrTy, Name: "sunkaddr");
6004 SunkAddr =
6005 Builder.CreateIntToPtr(V: SunkAddr, DestTy: Addr->getType(), Name: "sunkaddr");
6006 } else
6007 SunkAddr = Builder.CreatePointerCast(V: SunkAddr, DestTy: Addr->getType());
6008 }
6009 } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() &&
6010 SubtargetInfo->addrSinkUsingGEPs())) {
6011 // By default, we use the GEP-based method when AA is used later. This
6012 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
6013 LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
6014 << " for " << *MemoryInst << "\n");
6015 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
6016
6017 // First, find the pointer.
6018 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
6019 ResultPtr = AddrMode.BaseReg;
6020 AddrMode.BaseReg = nullptr;
6021 }
6022
6023 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
6024 // We can't add more than one pointer together, nor can we scale a
6025 // pointer (both of which seem meaningless).
6026 if (ResultPtr || AddrMode.Scale != 1)
6027 return Modified;
6028
6029 ResultPtr = AddrMode.ScaledReg;
6030 AddrMode.Scale = 0;
6031 }
6032
6033 // It is only safe to sign extend the BaseReg if we know that the math
6034 // required to create it did not overflow before we extend it. Since
6035 // the original IR value was tossed in favor of a constant back when
6036 // the AddrMode was created we need to bail out gracefully if widths
6037 // do not match instead of extending it.
6038 //
6039 // (See below for code to add the scale.)
6040 if (AddrMode.Scale) {
6041 Type *ScaledRegTy = AddrMode.ScaledReg->getType();
6042 if (cast<IntegerType>(Val: IntPtrTy)->getBitWidth() >
6043 cast<IntegerType>(Val: ScaledRegTy)->getBitWidth())
6044 return Modified;
6045 }
6046
6047 GlobalValue *BaseGV = AddrMode.BaseGV;
6048 if (BaseGV != nullptr) {
6049 if (ResultPtr)
6050 return Modified;
6051
6052 if (BaseGV->isThreadLocal()) {
6053 ResultPtr = Builder.CreateThreadLocalAddress(Ptr: BaseGV);
6054 } else {
6055 ResultPtr = BaseGV;
6056 }
6057 }
6058
6059 // If the real base value actually came from an inttoptr, then the matcher
6060 // will look through it and provide only the integer value. In that case,
6061 // use it here.
6062 if (!DL->isNonIntegralPointerType(Ty: Addr->getType())) {
6063 if (!ResultPtr && AddrMode.BaseReg) {
6064 ResultPtr = Builder.CreateIntToPtr(V: AddrMode.BaseReg, DestTy: Addr->getType(),
6065 Name: "sunkaddr");
6066 AddrMode.BaseReg = nullptr;
6067 } else if (!ResultPtr && AddrMode.Scale == 1) {
6068 ResultPtr = Builder.CreateIntToPtr(V: AddrMode.ScaledReg, DestTy: Addr->getType(),
6069 Name: "sunkaddr");
6070 AddrMode.Scale = 0;
6071 }
6072 }
6073
6074 if (!ResultPtr && !AddrMode.BaseReg && !AddrMode.Scale &&
6075 !AddrMode.BaseOffs) {
6076 SunkAddr = Constant::getNullValue(Ty: Addr->getType());
6077 } else if (!ResultPtr) {
6078 return Modified;
6079 } else {
6080 Type *I8PtrTy =
6081 Builder.getPtrTy(AddrSpace: Addr->getType()->getPointerAddressSpace());
6082
6083 // Start with the base register. Do this first so that subsequent address
6084 // matching finds it last, which will prevent it from trying to match it
6085 // as the scaled value in case it happens to be a mul. That would be
6086 // problematic if we've sunk a different mul for the scale, because then
6087 // we'd end up sinking both muls.
6088 if (AddrMode.BaseReg) {
6089 Value *V = AddrMode.BaseReg;
6090 if (V->getType() != IntPtrTy)
6091 V = Builder.CreateIntCast(V, DestTy: IntPtrTy, /*isSigned=*/true, Name: "sunkaddr");
6092
6093 ResultIndex = V;
6094 }
6095
6096 // Add the scale value.
6097 if (AddrMode.Scale) {
6098 Value *V = AddrMode.ScaledReg;
6099 if (V->getType() == IntPtrTy) {
6100 // done.
6101 } else {
6102 assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <
6103 cast<IntegerType>(V->getType())->getBitWidth() &&
6104 "We can't transform if ScaledReg is too narrow");
6105 V = Builder.CreateTrunc(V, DestTy: IntPtrTy, Name: "sunkaddr");
6106 }
6107
6108 if (AddrMode.Scale != 1)
6109 V = Builder.CreateMul(
6110 LHS: V, RHS: ConstantInt::getSigned(Ty: IntPtrTy, V: AddrMode.Scale), Name: "sunkaddr");
6111 if (ResultIndex)
6112 ResultIndex = Builder.CreateAdd(LHS: ResultIndex, RHS: V, Name: "sunkaddr");
6113 else
6114 ResultIndex = V;
6115 }
6116
6117 // Add in the Base Offset if present.
6118 if (AddrMode.BaseOffs) {
6119 Value *V = ConstantInt::getSigned(Ty: IntPtrTy, V: AddrMode.BaseOffs);
6120 if (ResultIndex) {
6121 // We need to add this separately from the scale above to help with
6122 // SDAG consecutive load/store merging.
6123 if (ResultPtr->getType() != I8PtrTy)
6124 ResultPtr = Builder.CreatePointerCast(V: ResultPtr, DestTy: I8PtrTy);
6125 ResultPtr = Builder.CreatePtrAdd(Ptr: ResultPtr, Offset: ResultIndex, Name: "sunkaddr",
6126 NW: AddrMode.InBounds);
6127 }
6128
6129 ResultIndex = V;
6130 }
6131
6132 if (!ResultIndex) {
6133 auto PtrInst = dyn_cast<Instruction>(Val: ResultPtr);
6134 // We know that we have a pointer without any offsets. If this pointer
6135 // originates from a different basic block than the current one, we
6136 // must be able to recreate it in the current basic block.
6137 // We do not support the recreation of any instructions yet.
6138 if (PtrInst && PtrInst->getParent() != MemoryInst->getParent())
6139 return Modified;
6140 SunkAddr = ResultPtr;
6141 } else {
6142 if (ResultPtr->getType() != I8PtrTy)
6143 ResultPtr = Builder.CreatePointerCast(V: ResultPtr, DestTy: I8PtrTy);
6144 SunkAddr = Builder.CreatePtrAdd(Ptr: ResultPtr, Offset: ResultIndex, Name: "sunkaddr",
6145 NW: AddrMode.InBounds);
6146 }
6147
6148 if (SunkAddr->getType() != Addr->getType()) {
6149 if (SunkAddr->getType()->getPointerAddressSpace() !=
6150 Addr->getType()->getPointerAddressSpace() &&
6151 !DL->isNonIntegralPointerType(Ty: Addr->getType())) {
6152 // There are two reasons the address spaces might not match: a no-op
6153 // addrspacecast, or a ptrtoint/inttoptr pair. Either way, we emit a
6154 // ptrtoint/inttoptr pair to ensure we match the original semantics.
6155 // TODO: allow bitcast between different address space pointers with
6156 // the same size.
6157 SunkAddr = Builder.CreatePtrToInt(V: SunkAddr, DestTy: IntPtrTy, Name: "sunkaddr");
6158 SunkAddr =
6159 Builder.CreateIntToPtr(V: SunkAddr, DestTy: Addr->getType(), Name: "sunkaddr");
6160 } else
6161 SunkAddr = Builder.CreatePointerCast(V: SunkAddr, DestTy: Addr->getType());
6162 }
6163 }
6164 } else {
6165 // We'd require a ptrtoint/inttoptr down the line, which we can't do for
6166 // non-integral pointers, so in that case bail out now.
6167 Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
6168 Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
6169 PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(Val: BaseTy);
6170 PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(Val: ScaleTy);
6171 if (DL->isNonIntegralPointerType(Ty: Addr->getType()) ||
6172 (BasePtrTy && DL->isNonIntegralPointerType(PT: BasePtrTy)) ||
6173 (ScalePtrTy && DL->isNonIntegralPointerType(PT: ScalePtrTy)) ||
6174 (AddrMode.BaseGV &&
6175 DL->isNonIntegralPointerType(PT: AddrMode.BaseGV->getType())))
6176 return Modified;
6177
6178 LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
6179 << " for " << *MemoryInst << "\n");
6180 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
6181 Value *Result = nullptr;
6182
6183 // Start with the base register. Do this first so that subsequent address
6184 // matching finds it last, which will prevent it from trying to match it
6185 // as the scaled value in case it happens to be a mul. That would be
6186 // problematic if we've sunk a different mul for the scale, because then
6187 // we'd end up sinking both muls.
6188 if (AddrMode.BaseReg) {
6189 Value *V = AddrMode.BaseReg;
6190 if (V->getType()->isPointerTy())
6191 V = Builder.CreatePtrToInt(V, DestTy: IntPtrTy, Name: "sunkaddr");
6192 if (V->getType() != IntPtrTy)
6193 V = Builder.CreateIntCast(V, DestTy: IntPtrTy, /*isSigned=*/true, Name: "sunkaddr");
6194 Result = V;
6195 }
6196
6197 // Add the scale value.
6198 if (AddrMode.Scale) {
6199 Value *V = AddrMode.ScaledReg;
6200 if (V->getType() == IntPtrTy) {
6201 // done.
6202 } else if (V->getType()->isPointerTy()) {
6203 V = Builder.CreatePtrToInt(V, DestTy: IntPtrTy, Name: "sunkaddr");
6204 } else if (cast<IntegerType>(Val: IntPtrTy)->getBitWidth() <
6205 cast<IntegerType>(Val: V->getType())->getBitWidth()) {
6206 V = Builder.CreateTrunc(V, DestTy: IntPtrTy, Name: "sunkaddr");
6207 } else {
6208 // It is only safe to sign extend the BaseReg if we know that the math
6209 // required to create it did not overflow before we extend it. Since
6210 // the original IR value was tossed in favor of a constant back when
6211 // the AddrMode was created we need to bail out gracefully if widths
6212 // do not match instead of extending it.
6213 Instruction *I = dyn_cast_or_null<Instruction>(Val: Result);
6214 if (I && (Result != AddrMode.BaseReg))
6215 I->eraseFromParent();
6216 return Modified;
6217 }
6218 if (AddrMode.Scale != 1)
6219 V = Builder.CreateMul(
6220 LHS: V, RHS: ConstantInt::getSigned(Ty: IntPtrTy, V: AddrMode.Scale), Name: "sunkaddr");
6221 if (Result)
6222 Result = Builder.CreateAdd(LHS: Result, RHS: V, Name: "sunkaddr");
6223 else
6224 Result = V;
6225 }
6226
6227 // Add in the BaseGV if present.
6228 GlobalValue *BaseGV = AddrMode.BaseGV;
6229 if (BaseGV != nullptr) {
6230 Value *BaseGVPtr;
6231 if (BaseGV->isThreadLocal()) {
6232 BaseGVPtr = Builder.CreateThreadLocalAddress(Ptr: BaseGV);
6233 } else {
6234 BaseGVPtr = BaseGV;
6235 }
6236 Value *V = Builder.CreatePtrToInt(V: BaseGVPtr, DestTy: IntPtrTy, Name: "sunkaddr");
6237 if (Result)
6238 Result = Builder.CreateAdd(LHS: Result, RHS: V, Name: "sunkaddr");
6239 else
6240 Result = V;
6241 }
6242
6243 // Add in the Base Offset if present.
6244 if (AddrMode.BaseOffs) {
6245 Value *V = ConstantInt::getSigned(Ty: IntPtrTy, V: AddrMode.BaseOffs);
6246 if (Result)
6247 Result = Builder.CreateAdd(LHS: Result, RHS: V, Name: "sunkaddr");
6248 else
6249 Result = V;
6250 }
6251
6252 if (!Result)
6253 SunkAddr = Constant::getNullValue(Ty: Addr->getType());
6254 else
6255 SunkAddr = Builder.CreateIntToPtr(V: Result, DestTy: Addr->getType(), Name: "sunkaddr");
6256 }
6257
6258 MemoryInst->replaceUsesOfWith(From: Repl, To: SunkAddr);
6259 // Store the newly computed address into the cache. In the case we reused a
6260 // value, this should be idempotent.
6261 SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
6262
6263 // If we have no uses, recursively delete the value and all dead instructions
6264 // using it.
6265 if (Repl->use_empty()) {
6266 resetIteratorIfInvalidatedWhileCalling(BB: CurInstIterator->getParent(), f: [&]() {
6267 RecursivelyDeleteTriviallyDeadInstructions(
6268 V: Repl, TLI: TLInfo, MSSAU: nullptr,
6269 AboutToDeleteCallback: [&](Value *V) { removeAllAssertingVHReferences(V); });
6270 });
6271 }
6272 ++NumMemoryInsts;
6273 return true;
6274}
6275
6276/// Rewrite GEP input to gather/scatter to enable SelectionDAGBuilder to find
6277/// a uniform base to use for ISD::MGATHER/MSCATTER. SelectionDAGBuilder can
6278/// only handle a 2 operand GEP in the same basic block or a splat constant
6279/// vector. The 2 operands to the GEP must have a scalar pointer and a vector
6280/// index.
6281///
6282/// If the existing GEP has a vector base pointer that is splat, we can look
6283/// through the splat to find the scalar pointer. If we can't find a scalar
6284/// pointer there's nothing we can do.
6285///
6286/// If we have a GEP with more than 2 indices where the middle indices are all
6287/// zeroes, we can replace it with 2 GEPs where the second has 2 operands.
6288///
6289/// If the final index isn't a vector or is a splat, we can emit a scalar GEP
6290/// followed by a GEP with an all zeroes vector index. This will enable
6291/// SelectionDAGBuilder to use the scalar GEP as the uniform base and have a
6292/// zero index.
6293bool CodeGenPrepare::optimizeGatherScatterInst(Instruction *MemoryInst,
6294 Value *Ptr) {
6295 Value *NewAddr;
6296
6297 if (const auto *GEP = dyn_cast<GetElementPtrInst>(Val: Ptr)) {
6298 // Don't optimize GEPs that don't have indices.
6299 if (!GEP->hasIndices())
6300 return false;
6301
6302 // If the GEP and the gather/scatter aren't in the same BB, don't optimize.
6303 // FIXME: We should support this by sinking the GEP.
6304 if (MemoryInst->getParent() != GEP->getParent())
6305 return false;
6306
6307 SmallVector<Value *, 2> Ops(GEP->operands());
6308
6309 bool RewriteGEP = false;
6310
6311 if (Ops[0]->getType()->isVectorTy()) {
6312 Ops[0] = getSplatValue(V: Ops[0]);
6313 if (!Ops[0])
6314 return false;
6315 RewriteGEP = true;
6316 }
6317
6318 unsigned FinalIndex = Ops.size() - 1;
6319
6320 // Ensure all but the last index is 0.
6321 // FIXME: This isn't strictly required. All that's required is that they are
6322 // all scalars or splats.
6323 for (unsigned i = 1; i < FinalIndex; ++i) {
6324 auto *C = dyn_cast<Constant>(Val: Ops[i]);
6325 if (!C)
6326 return false;
6327 if (isa<VectorType>(Val: C->getType()))
6328 C = C->getSplatValue();
6329 auto *CI = dyn_cast_or_null<ConstantInt>(Val: C);
6330 if (!CI || !CI->isZero())
6331 return false;
6332 // Scalarize the index if needed.
6333 Ops[i] = CI;
6334 }
6335
6336 // Try to scalarize the final index.
6337 if (Ops[FinalIndex]->getType()->isVectorTy()) {
6338 if (Value *V = getSplatValue(V: Ops[FinalIndex])) {
6339 auto *C = dyn_cast<ConstantInt>(Val: V);
6340 // Don't scalarize all zeros vector.
6341 if (!C || !C->isZero()) {
6342 Ops[FinalIndex] = V;
6343 RewriteGEP = true;
6344 }
6345 }
6346 }
6347
6348 // If we made any changes or the we have extra operands, we need to generate
6349 // new instructions.
6350 if (!RewriteGEP && Ops.size() == 2)
6351 return false;
6352
6353 auto NumElts = cast<VectorType>(Val: Ptr->getType())->getElementCount();
6354
6355 IRBuilder<> Builder(MemoryInst);
6356
6357 Type *SourceTy = GEP->getSourceElementType();
6358 Type *ScalarIndexTy = DL->getIndexType(PtrTy: Ops[0]->getType()->getScalarType());
6359
6360 // If the final index isn't a vector, emit a scalar GEP containing all ops
6361 // and a vector GEP with all zeroes final index.
6362 if (!Ops[FinalIndex]->getType()->isVectorTy()) {
6363 NewAddr = Builder.CreateGEP(Ty: SourceTy, Ptr: Ops[0], IdxList: ArrayRef(Ops).drop_front());
6364 auto *IndexTy = VectorType::get(ElementType: ScalarIndexTy, EC: NumElts);
6365 auto *SecondTy = GetElementPtrInst::getIndexedType(
6366 Ty: SourceTy, IdxList: ArrayRef(Ops).drop_front());
6367 NewAddr =
6368 Builder.CreateGEP(Ty: SecondTy, Ptr: NewAddr, IdxList: Constant::getNullValue(Ty: IndexTy));
6369 } else {
6370 Value *Base = Ops[0];
6371 Value *Index = Ops[FinalIndex];
6372
6373 // Create a scalar GEP if there are more than 2 operands.
6374 if (Ops.size() != 2) {
6375 // Replace the last index with 0.
6376 Ops[FinalIndex] =
6377 Constant::getNullValue(Ty: Ops[FinalIndex]->getType()->getScalarType());
6378 Base = Builder.CreateGEP(Ty: SourceTy, Ptr: Base, IdxList: ArrayRef(Ops).drop_front());
6379 SourceTy = GetElementPtrInst::getIndexedType(
6380 Ty: SourceTy, IdxList: ArrayRef(Ops).drop_front());
6381 }
6382
6383 // Now create the GEP with scalar pointer and vector index.
6384 NewAddr = Builder.CreateGEP(Ty: SourceTy, Ptr: Base, IdxList: Index);
6385 }
6386 } else if (!isa<Constant>(Val: Ptr)) {
6387 // Not a GEP, maybe its a splat and we can create a GEP to enable
6388 // SelectionDAGBuilder to use it as a uniform base.
6389 Value *V = getSplatValue(V: Ptr);
6390 if (!V)
6391 return false;
6392
6393 auto NumElts = cast<VectorType>(Val: Ptr->getType())->getElementCount();
6394
6395 IRBuilder<> Builder(MemoryInst);
6396
6397 // Emit a vector GEP with a scalar pointer and all 0s vector index.
6398 Type *ScalarIndexTy = DL->getIndexType(PtrTy: V->getType()->getScalarType());
6399 auto *IndexTy = VectorType::get(ElementType: ScalarIndexTy, EC: NumElts);
6400 Type *ScalarTy;
6401 if (cast<IntrinsicInst>(Val: MemoryInst)->getIntrinsicID() ==
6402 Intrinsic::masked_gather) {
6403 ScalarTy = MemoryInst->getType()->getScalarType();
6404 } else {
6405 assert(cast<IntrinsicInst>(MemoryInst)->getIntrinsicID() ==
6406 Intrinsic::masked_scatter);
6407 ScalarTy = MemoryInst->getOperand(i: 0)->getType()->getScalarType();
6408 }
6409 NewAddr = Builder.CreateGEP(Ty: ScalarTy, Ptr: V, IdxList: Constant::getNullValue(Ty: IndexTy));
6410 } else {
6411 // Constant, SelectionDAGBuilder knows to check if its a splat.
6412 return false;
6413 }
6414
6415 MemoryInst->replaceUsesOfWith(From: Ptr, To: NewAddr);
6416
6417 // If we have no uses, recursively delete the value and all dead instructions
6418 // using it.
6419 if (Ptr->use_empty())
6420 RecursivelyDeleteTriviallyDeadInstructions(
6421 V: Ptr, TLI: TLInfo, MSSAU: nullptr,
6422 AboutToDeleteCallback: [&](Value *V) { removeAllAssertingVHReferences(V); });
6423
6424 return true;
6425}
6426
6427// This is a helper for CodeGenPrepare::optimizeMulWithOverflow.
6428// Check the pattern we are interested in where there are maximum 2 uses
6429// of the intrinsic which are the extract instructions.
6430static bool matchOverflowPattern(Instruction *&I, ExtractValueInst *&MulExtract,
6431 ExtractValueInst *&OverflowExtract) {
6432 // Bail out if it's more than 2 users:
6433 if (I->hasNUsesOrMore(N: 3))
6434 return false;
6435
6436 for (User *U : I->users()) {
6437 auto *Extract = dyn_cast<ExtractValueInst>(Val: U);
6438 if (!Extract || Extract->getNumIndices() != 1)
6439 return false;
6440
6441 unsigned Index = Extract->getIndices()[0];
6442 if (Index == 0)
6443 MulExtract = Extract;
6444 else if (Index == 1)
6445 OverflowExtract = Extract;
6446 else
6447 return false;
6448 }
6449 return true;
6450}
6451
6452// Rewrite the mul_with_overflow intrinsic by checking if both of the
6453// operands' value ranges are within the legal type. If so, we can optimize the
6454// multiplication algorithm. This code is supposed to be written during the step
6455// of type legalization, but given that we need to reconstruct the IR which is
6456// not doable there, we do it here.
6457// The IR after the optimization will look like:
6458// entry:
6459// if signed:
6460// ( (lhs_lo>>BW-1) ^ lhs_hi) || ( (rhs_lo>>BW-1) ^ rhs_hi) ? overflow,
6461// overflow_no
6462// else:
6463// (lhs_hi != 0) || (rhs_hi != 0) ? overflow, overflow_no
6464// overflow_no:
6465// overflow:
6466// overflow.res:
6467// \returns true if optimization was applied
6468// TODO: This optimization can be further improved to optimize branching on
6469// overflow where the 'overflow_no' BB can branch directly to the false
6470// successor of overflow, but that would add additional complexity so we leave
6471// it for future work.
6472bool CodeGenPrepare::optimizeMulWithOverflow(Instruction *I, bool IsSigned,
6473 ModifyDT &ModifiedDT) {
6474 // Check if target supports this optimization.
6475 if (!TLI->shouldOptimizeMulOverflowWithZeroHighBits(
6476 Context&: I->getContext(),
6477 VT: TLI->getValueType(DL: *DL, Ty: I->getType()->getContainedType(i: 0))))
6478 return false;
6479
6480 ExtractValueInst *MulExtract = nullptr, *OverflowExtract = nullptr;
6481 if (!matchOverflowPattern(I, MulExtract, OverflowExtract))
6482 return false;
6483
6484 // Keep track of the instruction to stop reoptimizing it again.
6485 InsertedInsts.insert(Ptr: I);
6486
6487 Value *LHS = I->getOperand(i: 0);
6488 Value *RHS = I->getOperand(i: 1);
6489 Type *Ty = LHS->getType();
6490 unsigned VTHalfBitWidth = Ty->getScalarSizeInBits() / 2;
6491 Type *LegalTy = Ty->getWithNewBitWidth(NewBitWidth: VTHalfBitWidth);
6492
6493 // New BBs:
6494 BasicBlock *OverflowEntryBB =
6495 splitBlockBefore(Old: I->getParent(), SplitPt: I, DTU, LI, MSSAU: nullptr, BBName: "");
6496 OverflowEntryBB->takeName(V: I->getParent());
6497 // Keep the 'br' instruction that is generated as a result of the split to be
6498 // erased/replaced later.
6499 Instruction *OldTerminator = OverflowEntryBB->getTerminator();
6500 BasicBlock *NoOverflowBB =
6501 BasicBlock::Create(Context&: I->getContext(), Name: "overflow.no", Parent: I->getFunction());
6502 NoOverflowBB->moveAfter(MovePos: OverflowEntryBB);
6503 BasicBlock *OverflowBB =
6504 BasicBlock::Create(Context&: I->getContext(), Name: "overflow", Parent: I->getFunction());
6505 OverflowBB->moveAfter(MovePos: NoOverflowBB);
6506
6507 // BB overflow.entry:
6508 IRBuilder<> Builder(OverflowEntryBB);
6509 // Extract low and high halves of LHS:
6510 Value *LoLHS = Builder.CreateTrunc(V: LHS, DestTy: LegalTy, Name: "lo.lhs");
6511 Value *HiLHS = Builder.CreateLShr(LHS, RHS: VTHalfBitWidth, Name: "lhs.lsr");
6512 HiLHS = Builder.CreateTrunc(V: HiLHS, DestTy: LegalTy, Name: "hi.lhs");
6513
6514 // Extract low and high halves of RHS:
6515 Value *LoRHS = Builder.CreateTrunc(V: RHS, DestTy: LegalTy, Name: "lo.rhs");
6516 Value *HiRHS = Builder.CreateLShr(LHS: RHS, RHS: VTHalfBitWidth, Name: "rhs.lsr");
6517 HiRHS = Builder.CreateTrunc(V: HiRHS, DestTy: LegalTy, Name: "hi.rhs");
6518
6519 Value *IsAnyBitTrue;
6520 if (IsSigned) {
6521 Value *SignLoLHS =
6522 Builder.CreateAShr(LHS: LoLHS, RHS: VTHalfBitWidth - 1, Name: "sign.lo.lhs");
6523 Value *SignLoRHS =
6524 Builder.CreateAShr(LHS: LoRHS, RHS: VTHalfBitWidth - 1, Name: "sign.lo.rhs");
6525 Value *XorLHS = Builder.CreateXor(LHS: HiLHS, RHS: SignLoLHS);
6526 Value *XorRHS = Builder.CreateXor(LHS: HiRHS, RHS: SignLoRHS);
6527 Value *Or = Builder.CreateOr(LHS: XorLHS, RHS: XorRHS, Name: "or.lhs.rhs");
6528 IsAnyBitTrue = Builder.CreateCmp(Pred: ICmpInst::ICMP_NE, LHS: Or,
6529 RHS: ConstantInt::getNullValue(Ty: Or->getType()));
6530 } else {
6531 Value *CmpLHS = Builder.CreateCmp(Pred: ICmpInst::ICMP_NE, LHS: HiLHS,
6532 RHS: ConstantInt::getNullValue(Ty: LegalTy));
6533 Value *CmpRHS = Builder.CreateCmp(Pred: ICmpInst::ICMP_NE, LHS: HiRHS,
6534 RHS: ConstantInt::getNullValue(Ty: LegalTy));
6535 IsAnyBitTrue = Builder.CreateOr(LHS: CmpLHS, RHS: CmpRHS, Name: "or.lhs.rhs");
6536 }
6537 Builder.CreateCondBr(Cond: IsAnyBitTrue, True: OverflowBB, False: NoOverflowBB);
6538
6539 // BB overflow.no:
6540 Builder.SetInsertPoint(NoOverflowBB);
6541 Value *ExtLoLHS, *ExtLoRHS;
6542 if (IsSigned) {
6543 ExtLoLHS = Builder.CreateSExt(V: LoLHS, DestTy: Ty, Name: "lo.lhs.ext");
6544 ExtLoRHS = Builder.CreateSExt(V: LoRHS, DestTy: Ty, Name: "lo.rhs.ext");
6545 } else {
6546 ExtLoLHS = Builder.CreateZExt(V: LoLHS, DestTy: Ty, Name: "lo.lhs.ext");
6547 ExtLoRHS = Builder.CreateZExt(V: LoRHS, DestTy: Ty, Name: "lo.rhs.ext");
6548 }
6549
6550 Value *Mul = Builder.CreateMul(LHS: ExtLoLHS, RHS: ExtLoRHS, Name: "mul.overflow.no");
6551
6552 // Create the 'overflow.res' BB to merge the results of
6553 // the two paths:
6554 BasicBlock *OverflowResBB = I->getParent();
6555 OverflowResBB->setName("overflow.res");
6556
6557 // BB overflow.no: jump to overflow.res BB
6558 Builder.CreateBr(Dest: OverflowResBB);
6559 // No we don't need the old terminator in overflow.entry BB, erase it:
6560 OldTerminator->eraseFromParent();
6561
6562 // BB overflow.res:
6563 Builder.SetInsertPoint(TheBB: OverflowResBB, IP: OverflowResBB->getFirstInsertionPt());
6564 // Create PHI nodes to merge results from no.overflow BB and overflow BB to
6565 // replace the extract instructions.
6566 PHINode *OverflowResPHI = Builder.CreatePHI(Ty, NumReservedValues: 2),
6567 *OverflowFlagPHI =
6568 Builder.CreatePHI(Ty: IntegerType::getInt1Ty(C&: I->getContext()), NumReservedValues: 2);
6569
6570 // Add the incoming values from no.overflow BB and later from overflow BB.
6571 OverflowResPHI->addIncoming(V: Mul, BB: NoOverflowBB);
6572 OverflowFlagPHI->addIncoming(V: ConstantInt::getFalse(Context&: I->getContext()),
6573 BB: NoOverflowBB);
6574
6575 // Replace all users of MulExtract and OverflowExtract to use the PHI nodes.
6576 if (MulExtract) {
6577 MulExtract->replaceAllUsesWith(V: OverflowResPHI);
6578 MulExtract->eraseFromParent();
6579 }
6580 if (OverflowExtract) {
6581 OverflowExtract->replaceAllUsesWith(V: OverflowFlagPHI);
6582 OverflowExtract->eraseFromParent();
6583 }
6584
6585 // Remove the intrinsic from parent (overflow.res BB) as it will be part of
6586 // overflow BB
6587 I->removeFromParent();
6588 // BB overflow:
6589 I->insertInto(ParentBB: OverflowBB, It: OverflowBB->end());
6590 Builder.SetInsertPoint(TheBB: OverflowBB, IP: OverflowBB->end());
6591 Value *MulOverflow = Builder.CreateExtractValue(Agg: I, Idxs: {0}, Name: "mul.overflow");
6592 Value *OverflowFlag = Builder.CreateExtractValue(Agg: I, Idxs: {1}, Name: "overflow.flag");
6593 Builder.CreateBr(Dest: OverflowResBB);
6594
6595 // Add The Extracted values to the PHINodes in the overflow.res BB.
6596 OverflowResPHI->addIncoming(V: MulOverflow, BB: OverflowBB);
6597 OverflowFlagPHI->addIncoming(V: OverflowFlag, BB: OverflowBB);
6598
6599 DTU->applyUpdates(Updates: {{DominatorTree::Insert, OverflowEntryBB, OverflowBB},
6600 {DominatorTree::Insert, OverflowEntryBB, NoOverflowBB},
6601 {DominatorTree::Insert, NoOverflowBB, OverflowResBB},
6602 {DominatorTree::Delete, OverflowEntryBB, OverflowResBB},
6603 {DominatorTree::Insert, OverflowBB, OverflowResBB}});
6604
6605 ModifiedDT = ModifyDT::ModifyBBDT;
6606 return true;
6607}
6608
6609/// If there are any memory operands, use OptimizeMemoryInst to sink their
6610/// address computing into the block when possible / profitable.
6611bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
6612 bool MadeChange = false;
6613
6614 const TargetRegisterInfo *TRI =
6615 TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
6616 TargetLowering::AsmOperandInfoVector TargetConstraints =
6617 TLI->ParseConstraints(DL: *DL, TRI, Call: *CS);
6618 unsigned ArgNo = 0;
6619 for (TargetLowering::AsmOperandInfo &OpInfo : TargetConstraints) {
6620 // Compute the constraint code and ConstraintType to use.
6621 TLI->ComputeConstraintToUse(OpInfo, Op: SDValue());
6622
6623 // TODO: Also handle C_Address?
6624 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6625 OpInfo.isIndirect) {
6626 Value *OpVal = CS->getArgOperand(i: ArgNo++);
6627 MadeChange |= optimizeMemoryInst(MemoryInst: CS, Addr: OpVal, AccessTy: OpVal->getType(), AddrSpace: ~0u);
6628 } else if (OpInfo.Type == InlineAsm::isInput)
6629 ArgNo++;
6630 }
6631
6632 return MadeChange;
6633}
6634
6635/// Check if all the uses of \p Val are equivalent (or free) zero or
6636/// sign extensions.
6637static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
6638 assert(!Val->use_empty() && "Input must have at least one use");
6639 const Instruction *FirstUser = cast<Instruction>(Val: *Val->user_begin());
6640 bool IsSExt = isa<SExtInst>(Val: FirstUser);
6641 Type *ExtTy = FirstUser->getType();
6642 for (const User *U : Val->users()) {
6643 const Instruction *UI = cast<Instruction>(Val: U);
6644 if ((IsSExt && !isa<SExtInst>(Val: UI)) || (!IsSExt && !isa<ZExtInst>(Val: UI)))
6645 return false;
6646 Type *CurTy = UI->getType();
6647 // Same input and output types: Same instruction after CSE.
6648 if (CurTy == ExtTy)
6649 continue;
6650
6651 // If IsSExt is true, we are in this situation:
6652 // a = Val
6653 // b = sext ty1 a to ty2
6654 // c = sext ty1 a to ty3
6655 // Assuming ty2 is shorter than ty3, this could be turned into:
6656 // a = Val
6657 // b = sext ty1 a to ty2
6658 // c = sext ty2 b to ty3
6659 // However, the last sext is not free.
6660 if (IsSExt)
6661 return false;
6662
6663 // This is a ZExt, maybe this is free to extend from one type to another.
6664 // In that case, we would not account for a different use.
6665 Type *NarrowTy;
6666 Type *LargeTy;
6667 if (ExtTy->getScalarType()->getIntegerBitWidth() >
6668 CurTy->getScalarType()->getIntegerBitWidth()) {
6669 NarrowTy = CurTy;
6670 LargeTy = ExtTy;
6671 } else {
6672 NarrowTy = ExtTy;
6673 LargeTy = CurTy;
6674 }
6675
6676 if (!TLI.isZExtFree(FromTy: NarrowTy, ToTy: LargeTy))
6677 return false;
6678 }
6679 // All uses are the same or can be derived from one another for free.
6680 return true;
6681}
6682
6683/// Try to speculatively promote extensions in \p Exts and continue
6684/// promoting through newly promoted operands recursively as far as doing so is
6685/// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
6686/// When some promotion happened, \p TPT contains the proper state to revert
6687/// them.
6688///
6689/// \return true if some promotion happened, false otherwise.
6690bool CodeGenPrepare::tryToPromoteExts(
6691 TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
6692 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
6693 unsigned CreatedInstsCost) {
6694 bool Promoted = false;
6695
6696 // Iterate over all the extensions to try to promote them.
6697 for (auto *I : Exts) {
6698 // Early check if we directly have ext(load).
6699 if (isa<LoadInst>(Val: I->getOperand(i: 0))) {
6700 ProfitablyMovedExts.push_back(Elt: I);
6701 continue;
6702 }
6703
6704 // Check whether or not we want to do any promotion. The reason we have
6705 // this check inside the for loop is to catch the case where an extension
6706 // is directly fed by a load because in such case the extension can be moved
6707 // up without any promotion on its operands.
6708 if (!TLI->enableExtLdPromotion() || DisableExtLdPromotion)
6709 return false;
6710
6711 // Get the action to perform the promotion.
6712 TypePromotionHelper::Action TPH =
6713 TypePromotionHelper::getAction(Ext: I, InsertedInsts, TLI: *TLI, PromotedInsts);
6714 // Check if we can promote.
6715 if (!TPH) {
6716 // Save the current extension as we cannot move up through its operand.
6717 ProfitablyMovedExts.push_back(Elt: I);
6718 continue;
6719 }
6720
6721 // Save the current state.
6722 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
6723 TPT.getRestorationPoint();
6724 SmallVector<Instruction *, 4> NewExts;
6725 unsigned NewCreatedInstsCost = 0;
6726 unsigned ExtCost = !TLI->isExtFree(I);
6727 // Promote.
6728 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
6729 &NewExts, nullptr, *TLI);
6730 assert(PromotedVal &&
6731 "TypePromotionHelper should have filtered out those cases");
6732
6733 // We would be able to merge only one extension in a load.
6734 // Therefore, if we have more than 1 new extension we heuristically
6735 // cut this search path, because it means we degrade the code quality.
6736 // With exactly 2, the transformation is neutral, because we will merge
6737 // one extension but leave one. However, we optimistically keep going,
6738 // because the new extension may be removed too. Also avoid replacing a
6739 // single free extension with multiple extensions, as this increases the
6740 // number of IR instructions while not providing any savings.
6741 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
6742 // FIXME: It would be possible to propagate a negative value instead of
6743 // conservatively ceiling it to 0.
6744 TotalCreatedInstsCost =
6745 std::max(a: (long long)0, b: (TotalCreatedInstsCost - ExtCost));
6746 if (!StressExtLdPromotion &&
6747 (TotalCreatedInstsCost > 1 ||
6748 !isPromotedInstructionLegal(TLI: *TLI, DL: *DL, Val: PromotedVal) ||
6749 (ExtCost == 0 && NewExts.size() > 1))) {
6750 // This promotion is not profitable, rollback to the previous state, and
6751 // save the current extension in ProfitablyMovedExts as the latest
6752 // speculative promotion turned out to be unprofitable.
6753 TPT.rollback(Point: LastKnownGood);
6754 ProfitablyMovedExts.push_back(Elt: I);
6755 continue;
6756 }
6757 // Continue promoting NewExts as far as doing so is profitable.
6758 SmallVector<Instruction *, 2> NewlyMovedExts;
6759 (void)tryToPromoteExts(TPT, Exts: NewExts, ProfitablyMovedExts&: NewlyMovedExts, CreatedInstsCost: TotalCreatedInstsCost);
6760 bool NewPromoted = false;
6761 for (auto *ExtInst : NewlyMovedExts) {
6762 Instruction *MovedExt = cast<Instruction>(Val: ExtInst);
6763 Value *ExtOperand = MovedExt->getOperand(i: 0);
6764 // If we have reached to a load, we need this extra profitability check
6765 // as it could potentially be merged into an ext(load).
6766 if (isa<LoadInst>(Val: ExtOperand) &&
6767 !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
6768 (ExtOperand->hasOneUse() || hasSameExtUse(Val: ExtOperand, TLI: *TLI))))
6769 continue;
6770
6771 ProfitablyMovedExts.push_back(Elt: MovedExt);
6772 NewPromoted = true;
6773 }
6774
6775 // If none of speculative promotions for NewExts is profitable, rollback
6776 // and save the current extension (I) as the last profitable extension.
6777 if (!NewPromoted) {
6778 TPT.rollback(Point: LastKnownGood);
6779 ProfitablyMovedExts.push_back(Elt: I);
6780 continue;
6781 }
6782 // The promotion is profitable.
6783 Promoted = true;
6784 }
6785 return Promoted;
6786}
6787
6788/// Merging redundant sexts when one is dominating the other.
6789bool CodeGenPrepare::mergeSExts(Function &F) {
6790 bool Changed = false;
6791 for (auto &Entry : ValToSExtendedUses) {
6792 SExts &Insts = Entry.second;
6793 SExts CurPts;
6794 for (Instruction *Inst : Insts) {
6795 if (RemovedInsts.count(Ptr: Inst) || !isa<SExtInst>(Val: Inst) ||
6796 Inst->getOperand(i: 0) != Entry.first)
6797 continue;
6798 bool inserted = false;
6799 for (auto &Pt : CurPts) {
6800 if (getDT().dominates(Def: Inst, User: Pt)) {
6801 replaceAllUsesWith(Old: Pt, New: Inst, FreshBBs, IsHuge: IsHugeFunc);
6802 RemovedInsts.insert(Ptr: Pt);
6803 Pt->removeFromParent();
6804 Pt = Inst;
6805 inserted = true;
6806 Changed = true;
6807 break;
6808 }
6809 if (!getDT().dominates(Def: Pt, User: Inst))
6810 // Give up if we need to merge in a common dominator as the
6811 // experiments show it is not profitable.
6812 continue;
6813 replaceAllUsesWith(Old: Inst, New: Pt, FreshBBs, IsHuge: IsHugeFunc);
6814 RemovedInsts.insert(Ptr: Inst);
6815 Inst->removeFromParent();
6816 inserted = true;
6817 Changed = true;
6818 break;
6819 }
6820 if (!inserted)
6821 CurPts.push_back(Elt: Inst);
6822 }
6823 }
6824 return Changed;
6825}
6826
6827// Splitting large data structures so that the GEPs accessing them can have
6828// smaller offsets so that they can be sunk to the same blocks as their users.
6829// For example, a large struct starting from %base is split into two parts
6830// where the second part starts from %new_base.
6831//
6832// Before:
6833// BB0:
6834// %base =
6835//
6836// BB1:
6837// %gep0 = gep %base, off0
6838// %gep1 = gep %base, off1
6839// %gep2 = gep %base, off2
6840//
6841// BB2:
6842// %load1 = load %gep0
6843// %load2 = load %gep1
6844// %load3 = load %gep2
6845//
6846// After:
6847// BB0:
6848// %base =
6849// %new_base = gep %base, off0
6850//
6851// BB1:
6852// %new_gep0 = %new_base
6853// %new_gep1 = gep %new_base, off1 - off0
6854// %new_gep2 = gep %new_base, off2 - off0
6855//
6856// BB2:
6857// %load1 = load i32, i32* %new_gep0
6858// %load2 = load i32, i32* %new_gep1
6859// %load3 = load i32, i32* %new_gep2
6860//
6861// %new_gep1 and %new_gep2 can be sunk to BB2 now after the splitting because
6862// their offsets are smaller enough to fit into the addressing mode.
6863bool CodeGenPrepare::splitLargeGEPOffsets() {
6864 bool Changed = false;
6865 for (auto &Entry : LargeOffsetGEPMap) {
6866 Value *OldBase = Entry.first;
6867 SmallVectorImpl<std::pair<AssertingVH<GetElementPtrInst>, int64_t>>
6868 &LargeOffsetGEPs = Entry.second;
6869 auto compareGEPOffset =
6870 [&](const std::pair<GetElementPtrInst *, int64_t> &LHS,
6871 const std::pair<GetElementPtrInst *, int64_t> &RHS) {
6872 if (LHS.first == RHS.first)
6873 return false;
6874 if (LHS.second != RHS.second)
6875 return LHS.second < RHS.second;
6876 return LargeOffsetGEPID[LHS.first] < LargeOffsetGEPID[RHS.first];
6877 };
6878 // Sorting all the GEPs of the same data structures based on the offsets.
6879 llvm::sort(C&: LargeOffsetGEPs, Comp: compareGEPOffset);
6880 LargeOffsetGEPs.erase(CS: llvm::unique(R&: LargeOffsetGEPs), CE: LargeOffsetGEPs.end());
6881 // Skip if all the GEPs have the same offsets.
6882 if (LargeOffsetGEPs.front().second == LargeOffsetGEPs.back().second)
6883 continue;
6884 GetElementPtrInst *BaseGEP = LargeOffsetGEPs.begin()->first;
6885 int64_t BaseOffset = LargeOffsetGEPs.begin()->second;
6886 Value *NewBaseGEP = nullptr;
6887
6888 auto createNewBase = [&](int64_t BaseOffset, Value *OldBase,
6889 GetElementPtrInst *GEP) {
6890 LLVMContext &Ctx = GEP->getContext();
6891 Type *PtrIdxTy = DL->getIndexType(PtrTy: GEP->getType());
6892 Type *I8PtrTy =
6893 PointerType::get(C&: Ctx, AddressSpace: GEP->getType()->getPointerAddressSpace());
6894
6895 BasicBlock::iterator NewBaseInsertPt;
6896 BasicBlock *NewBaseInsertBB;
6897 if (auto *BaseI = dyn_cast<Instruction>(Val: OldBase)) {
6898 // If the base of the struct is an instruction, the new base will be
6899 // inserted close to it.
6900 NewBaseInsertBB = BaseI->getParent();
6901 if (isa<PHINode>(Val: BaseI))
6902 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
6903 else if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Val: BaseI)) {
6904 NewBaseInsertBB =
6905 SplitEdge(From: NewBaseInsertBB, To: Invoke->getNormalDest(), DT: &getDT(), LI);
6906 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
6907 } else
6908 NewBaseInsertPt = std::next(x: BaseI->getIterator());
6909 } else {
6910 // If the current base is an argument or global value, the new base
6911 // will be inserted to the entry block.
6912 NewBaseInsertBB = &BaseGEP->getFunction()->getEntryBlock();
6913 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
6914 }
6915 IRBuilder<> NewBaseBuilder(NewBaseInsertBB, NewBaseInsertPt);
6916 // Create a new base.
6917 // TODO: Avoid implicit trunc?
6918 // See https://github.com/llvm/llvm-project/issues/112510.
6919 Value *BaseIndex =
6920 ConstantInt::getSigned(Ty: PtrIdxTy, V: BaseOffset, /*ImplicitTrunc=*/true);
6921 NewBaseGEP = OldBase;
6922 if (NewBaseGEP->getType() != I8PtrTy)
6923 NewBaseGEP = NewBaseBuilder.CreatePointerCast(V: NewBaseGEP, DestTy: I8PtrTy);
6924 NewBaseGEP =
6925 NewBaseBuilder.CreatePtrAdd(Ptr: NewBaseGEP, Offset: BaseIndex, Name: "splitgep");
6926 NewGEPBases.insert(V: NewBaseGEP);
6927 return;
6928 };
6929
6930 // Check whether all the offsets can be encoded with prefered common base.
6931 if (int64_t PreferBase = TLI->getPreferredLargeGEPBaseOffset(
6932 MinOffset: LargeOffsetGEPs.front().second, MaxOffset: LargeOffsetGEPs.back().second)) {
6933 BaseOffset = PreferBase;
6934 // Create a new base if the offset of the BaseGEP can be decoded with one
6935 // instruction.
6936 createNewBase(BaseOffset, OldBase, BaseGEP);
6937 }
6938
6939 auto *LargeOffsetGEP = LargeOffsetGEPs.begin();
6940 while (LargeOffsetGEP != LargeOffsetGEPs.end()) {
6941 GetElementPtrInst *GEP = LargeOffsetGEP->first;
6942 int64_t Offset = LargeOffsetGEP->second;
6943 if (Offset != BaseOffset) {
6944 TargetLowering::AddrMode AddrMode;
6945 AddrMode.HasBaseReg = true;
6946 AddrMode.BaseOffs = Offset - BaseOffset;
6947 // The result type of the GEP might not be the type of the memory
6948 // access.
6949 if (!TLI->isLegalAddressingMode(DL: *DL, AM: AddrMode,
6950 Ty: GEP->getResultElementType(),
6951 AddrSpace: GEP->getAddressSpace())) {
6952 // We need to create a new base if the offset to the current base is
6953 // too large to fit into the addressing mode. So, a very large struct
6954 // may be split into several parts.
6955 BaseGEP = GEP;
6956 BaseOffset = Offset;
6957 NewBaseGEP = nullptr;
6958 }
6959 }
6960
6961 // Generate a new GEP to replace the current one.
6962 Type *PtrIdxTy = DL->getIndexType(PtrTy: GEP->getType());
6963
6964 if (!NewBaseGEP) {
6965 // Create a new base if we don't have one yet. Find the insertion
6966 // pointer for the new base first.
6967 createNewBase(BaseOffset, OldBase, GEP);
6968 }
6969
6970 IRBuilder<> Builder(GEP);
6971 Value *NewGEP = NewBaseGEP;
6972 if (Offset != BaseOffset) {
6973 // Calculate the new offset for the new GEP.
6974 Value *Index = ConstantInt::get(Ty: PtrIdxTy, V: Offset - BaseOffset);
6975 NewGEP = Builder.CreatePtrAdd(Ptr: NewBaseGEP, Offset: Index);
6976 }
6977 replaceAllUsesWith(Old: GEP, New: NewGEP, FreshBBs, IsHuge: IsHugeFunc);
6978 LargeOffsetGEPID.erase(Val: GEP);
6979 LargeOffsetGEP = LargeOffsetGEPs.erase(CI: LargeOffsetGEP);
6980 GEP->eraseFromParent();
6981 Changed = true;
6982 }
6983 }
6984 return Changed;
6985}
6986
6987bool CodeGenPrepare::optimizePhiType(
6988 PHINode *I, SmallPtrSetImpl<PHINode *> &Visited,
6989 SmallPtrSetImpl<Instruction *> &DeletedInstrs) {
6990 // We are looking for a collection on interconnected phi nodes that together
6991 // only use loads/bitcasts and are used by stores/bitcasts, and the bitcasts
6992 // are of the same type. Convert the whole set of nodes to the type of the
6993 // bitcast.
6994 Type *PhiTy = I->getType();
6995 Type *ConvertTy = nullptr;
6996 if (Visited.count(Ptr: I) ||
6997 (!I->getType()->isIntegerTy() && !I->getType()->isFloatingPointTy()))
6998 return false;
6999
7000 SmallVector<Instruction *, 4> Worklist;
7001 Worklist.push_back(Elt: cast<Instruction>(Val: I));
7002 SmallPtrSet<PHINode *, 4> PhiNodes;
7003 SmallPtrSet<ConstantData *, 4> Constants;
7004 PhiNodes.insert(Ptr: I);
7005 Visited.insert(Ptr: I);
7006 SmallPtrSet<Instruction *, 4> Defs;
7007 SmallPtrSet<Instruction *, 4> Uses;
7008 // This works by adding extra bitcasts between load/stores and removing
7009 // existing bitcasts. If we have a phi(bitcast(load)) or a store(bitcast(phi))
7010 // we can get in the situation where we remove a bitcast in one iteration
7011 // just to add it again in the next. We need to ensure that at least one
7012 // bitcast we remove are anchored to something that will not change back.
7013 bool AnyAnchored = false;
7014
7015 while (!Worklist.empty()) {
7016 Instruction *II = Worklist.pop_back_val();
7017
7018 if (auto *Phi = dyn_cast<PHINode>(Val: II)) {
7019 // Handle Defs, which might also be PHI's
7020 for (Value *V : Phi->incoming_values()) {
7021 if (auto *OpPhi = dyn_cast<PHINode>(Val: V)) {
7022 if (!PhiNodes.count(Ptr: OpPhi)) {
7023 if (!Visited.insert(Ptr: OpPhi).second)
7024 return false;
7025 PhiNodes.insert(Ptr: OpPhi);
7026 Worklist.push_back(Elt: OpPhi);
7027 }
7028 } else if (auto *OpLoad = dyn_cast<LoadInst>(Val: V)) {
7029 if (!OpLoad->isSimple())
7030 return false;
7031 if (Defs.insert(Ptr: OpLoad).second)
7032 Worklist.push_back(Elt: OpLoad);
7033 } else if (auto *OpEx = dyn_cast<ExtractElementInst>(Val: V)) {
7034 if (Defs.insert(Ptr: OpEx).second)
7035 Worklist.push_back(Elt: OpEx);
7036 } else if (auto *OpBC = dyn_cast<BitCastInst>(Val: V)) {
7037 if (!ConvertTy)
7038 ConvertTy = OpBC->getOperand(i_nocapture: 0)->getType();
7039 if (OpBC->getOperand(i_nocapture: 0)->getType() != ConvertTy)
7040 return false;
7041 if (Defs.insert(Ptr: OpBC).second) {
7042 Worklist.push_back(Elt: OpBC);
7043 AnyAnchored |= !isa<LoadInst>(Val: OpBC->getOperand(i_nocapture: 0)) &&
7044 !isa<ExtractElementInst>(Val: OpBC->getOperand(i_nocapture: 0));
7045 }
7046 } else if (auto *OpC = dyn_cast<ConstantData>(Val: V))
7047 Constants.insert(Ptr: OpC);
7048 else
7049 return false;
7050 }
7051 }
7052
7053 // Handle uses which might also be phi's
7054 for (User *V : II->users()) {
7055 if (auto *OpPhi = dyn_cast<PHINode>(Val: V)) {
7056 if (!PhiNodes.count(Ptr: OpPhi)) {
7057 if (Visited.count(Ptr: OpPhi))
7058 return false;
7059 PhiNodes.insert(Ptr: OpPhi);
7060 Visited.insert(Ptr: OpPhi);
7061 Worklist.push_back(Elt: OpPhi);
7062 }
7063 } else if (auto *OpStore = dyn_cast<StoreInst>(Val: V)) {
7064 if (!OpStore->isSimple() || OpStore->getOperand(i_nocapture: 0) != II)
7065 return false;
7066 Uses.insert(Ptr: OpStore);
7067 } else if (auto *OpBC = dyn_cast<BitCastInst>(Val: V)) {
7068 if (!ConvertTy)
7069 ConvertTy = OpBC->getType();
7070 if (OpBC->getType() != ConvertTy)
7071 return false;
7072 Uses.insert(Ptr: OpBC);
7073 AnyAnchored |=
7074 any_of(Range: OpBC->users(), P: [](User *U) { return !isa<StoreInst>(Val: U); });
7075 } else {
7076 return false;
7077 }
7078 }
7079 }
7080
7081 if (!ConvertTy || !AnyAnchored || PhiTy == ConvertTy ||
7082 !TLI->shouldConvertPhiType(From: PhiTy, To: ConvertTy))
7083 return false;
7084
7085 LLVM_DEBUG(dbgs() << "Converting " << *I << "\n and connected nodes to "
7086 << *ConvertTy << "\n");
7087
7088 // Create all the new phi nodes of the new type, and bitcast any loads to the
7089 // correct type.
7090 ValueToValueMap ValMap;
7091 for (ConstantData *C : Constants)
7092 ValMap[C] = ConstantExpr::getBitCast(C, Ty: ConvertTy);
7093 for (Instruction *D : Defs) {
7094 if (isa<BitCastInst>(Val: D)) {
7095 ValMap[D] = D->getOperand(i: 0);
7096 DeletedInstrs.insert(Ptr: D);
7097 } else {
7098 BasicBlock::iterator insertPt = std::next(x: D->getIterator());
7099 ValMap[D] = new BitCastInst(D, ConvertTy, D->getName() + ".bc", insertPt);
7100 }
7101 }
7102 for (PHINode *Phi : PhiNodes)
7103 ValMap[Phi] = PHINode::Create(Ty: ConvertTy, NumReservedValues: Phi->getNumIncomingValues(),
7104 NameStr: Phi->getName() + ".tc", InsertBefore: Phi->getIterator());
7105 // Pipe together all the PhiNodes.
7106 for (PHINode *Phi : PhiNodes) {
7107 PHINode *NewPhi = cast<PHINode>(Val: ValMap[Phi]);
7108 for (int i = 0, e = Phi->getNumIncomingValues(); i < e; i++)
7109 NewPhi->addIncoming(V: ValMap[Phi->getIncomingValue(i)],
7110 BB: Phi->getIncomingBlock(i));
7111 Visited.insert(Ptr: NewPhi);
7112 }
7113 // And finally pipe up the stores and bitcasts
7114 for (Instruction *U : Uses) {
7115 if (isa<BitCastInst>(Val: U)) {
7116 DeletedInstrs.insert(Ptr: U);
7117 replaceAllUsesWith(Old: U, New: ValMap[U->getOperand(i: 0)], FreshBBs, IsHuge: IsHugeFunc);
7118 } else {
7119 U->setOperand(i: 0, Val: new BitCastInst(ValMap[U->getOperand(i: 0)], PhiTy, "bc",
7120 U->getIterator()));
7121 }
7122 }
7123
7124 // Save the removed phis to be deleted later.
7125 DeletedInstrs.insert_range(R&: PhiNodes);
7126 return true;
7127}
7128
7129bool CodeGenPrepare::optimizePhiTypes(Function &F) {
7130 if (!OptimizePhiTypes)
7131 return false;
7132
7133 bool Changed = false;
7134 SmallPtrSet<PHINode *, 4> Visited;
7135 SmallPtrSet<Instruction *, 4> DeletedInstrs;
7136
7137 // Attempt to optimize all the phis in the functions to the correct type.
7138 for (auto &BB : F)
7139 for (auto &Phi : BB.phis())
7140 Changed |= optimizePhiType(I: &Phi, Visited, DeletedInstrs);
7141
7142 // Remove any old phi's that have been converted.
7143 for (auto *I : DeletedInstrs) {
7144 replaceAllUsesWith(Old: I, New: PoisonValue::get(T: I->getType()), FreshBBs, IsHuge: IsHugeFunc);
7145 I->eraseFromParent();
7146 }
7147
7148 return Changed;
7149}
7150
7151/// Return true, if an ext(load) can be formed from an extension in
7152/// \p MovedExts.
7153bool CodeGenPrepare::canFormExtLd(
7154 const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
7155 Instruction *&Inst, bool HasPromoted) {
7156 for (auto *MovedExtInst : MovedExts) {
7157 if (isa<LoadInst>(Val: MovedExtInst->getOperand(i: 0))) {
7158 LI = cast<LoadInst>(Val: MovedExtInst->getOperand(i: 0));
7159 Inst = MovedExtInst;
7160 break;
7161 }
7162 }
7163 if (!LI)
7164 return false;
7165
7166 // If they're already in the same block, there's nothing to do.
7167 // Make the cheap checks first if we did not promote.
7168 // If we promoted, we need to check if it is indeed profitable.
7169 if (!HasPromoted && LI->getParent() == Inst->getParent())
7170 return false;
7171
7172 return TLI->isExtLoad(Load: LI, Ext: Inst, DL: *DL);
7173}
7174
7175/// Move a zext or sext fed by a load into the same basic block as the load,
7176/// unless conditions are unfavorable. This allows SelectionDAG to fold the
7177/// extend into the load.
7178///
7179/// E.g.,
7180/// \code
7181/// %ld = load i32* %addr
7182/// %add = add nuw i32 %ld, 4
7183/// %zext = zext i32 %add to i64
7184// \endcode
7185/// =>
7186/// \code
7187/// %ld = load i32* %addr
7188/// %zext = zext i32 %ld to i64
7189/// %add = add nuw i64 %zext, 4
7190/// \encode
7191/// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
7192/// allow us to match zext(load i32*) to i64.
7193///
7194/// Also, try to promote the computations used to obtain a sign extended
7195/// value used into memory accesses.
7196/// E.g.,
7197/// \code
7198/// a = add nsw i32 b, 3
7199/// d = sext i32 a to i64
7200/// e = getelementptr ..., i64 d
7201/// \endcode
7202/// =>
7203/// \code
7204/// f = sext i32 b to i64
7205/// a = add nsw i64 f, 3
7206/// e = getelementptr ..., i64 a
7207/// \endcode
7208///
7209/// \p Inst[in/out] the extension may be modified during the process if some
7210/// promotions apply.
7211bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
7212 bool AllowPromotionWithoutCommonHeader = false;
7213 /// See if it is an interesting sext operations for the address type
7214 /// promotion before trying to promote it, e.g., the ones with the right
7215 /// type and used in memory accesses.
7216 bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
7217 I: *Inst, AllowPromotionWithoutCommonHeader);
7218 TypePromotionTransaction TPT(RemovedInsts);
7219 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
7220 TPT.getRestorationPoint();
7221 SmallVector<Instruction *, 1> Exts;
7222 SmallVector<Instruction *, 2> SpeculativelyMovedExts;
7223 Exts.push_back(Elt: Inst);
7224
7225 bool HasPromoted = tryToPromoteExts(TPT, Exts, ProfitablyMovedExts&: SpeculativelyMovedExts);
7226
7227 // Look for a load being extended.
7228 LoadInst *LI = nullptr;
7229 Instruction *ExtFedByLoad;
7230
7231 // Try to promote a chain of computation if it allows to form an extended
7232 // load.
7233 if (canFormExtLd(MovedExts: SpeculativelyMovedExts, LI, Inst&: ExtFedByLoad, HasPromoted)) {
7234 assert(LI && ExtFedByLoad && "Expect a valid load and extension");
7235 TPT.commit();
7236 // Move the extend into the same block as the load.
7237 ExtFedByLoad->moveAfter(MovePos: LI);
7238 ++NumExtsMoved;
7239 Inst = ExtFedByLoad;
7240 return true;
7241 }
7242
7243 // Continue promoting SExts if known as considerable depending on targets.
7244 if (ATPConsiderable &&
7245 performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
7246 HasPromoted, TPT, SpeculativelyMovedExts))
7247 return true;
7248
7249 TPT.rollback(Point: LastKnownGood);
7250 return false;
7251}
7252
7253// Perform address type promotion if doing so is profitable.
7254// If AllowPromotionWithoutCommonHeader == false, we should find other sext
7255// instructions that sign extended the same initial value. However, if
7256// AllowPromotionWithoutCommonHeader == true, we expect promoting the
7257// extension is just profitable.
7258bool CodeGenPrepare::performAddressTypePromotion(
7259 Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
7260 bool HasPromoted, TypePromotionTransaction &TPT,
7261 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
7262 bool Promoted = false;
7263 SmallPtrSet<Instruction *, 1> UnhandledExts;
7264 bool AllSeenFirst = true;
7265 for (auto *I : SpeculativelyMovedExts) {
7266 Value *HeadOfChain = I->getOperand(i: 0);
7267 DenseMap<Value *, Instruction *>::iterator AlreadySeen =
7268 SeenChainsForSExt.find(Val: HeadOfChain);
7269 // If there is an unhandled SExt which has the same header, try to promote
7270 // it as well.
7271 if (AlreadySeen != SeenChainsForSExt.end()) {
7272 if (AlreadySeen->second != nullptr)
7273 UnhandledExts.insert(Ptr: AlreadySeen->second);
7274 AllSeenFirst = false;
7275 }
7276 }
7277
7278 if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
7279 SpeculativelyMovedExts.size() == 1)) {
7280 TPT.commit();
7281 if (HasPromoted)
7282 Promoted = true;
7283 for (auto *I : SpeculativelyMovedExts) {
7284 Value *HeadOfChain = I->getOperand(i: 0);
7285 SeenChainsForSExt[HeadOfChain] = nullptr;
7286 ValToSExtendedUses[HeadOfChain].push_back(Elt: I);
7287 }
7288 // Update Inst as promotion happen.
7289 Inst = SpeculativelyMovedExts.pop_back_val();
7290 } else {
7291 // This is the first chain visited from the header, keep the current chain
7292 // as unhandled. Defer to promote this until we encounter another SExt
7293 // chain derived from the same header.
7294 for (auto *I : SpeculativelyMovedExts) {
7295 Value *HeadOfChain = I->getOperand(i: 0);
7296 SeenChainsForSExt[HeadOfChain] = Inst;
7297 }
7298 return false;
7299 }
7300
7301 if (!AllSeenFirst && !UnhandledExts.empty())
7302 for (auto *VisitedSExt : UnhandledExts) {
7303 if (RemovedInsts.count(Ptr: VisitedSExt))
7304 continue;
7305 TypePromotionTransaction TPT(RemovedInsts);
7306 SmallVector<Instruction *, 1> Exts;
7307 SmallVector<Instruction *, 2> Chains;
7308 Exts.push_back(Elt: VisitedSExt);
7309 bool HasPromoted = tryToPromoteExts(TPT, Exts, ProfitablyMovedExts&: Chains);
7310 TPT.commit();
7311 if (HasPromoted)
7312 Promoted = true;
7313 for (auto *I : Chains) {
7314 Value *HeadOfChain = I->getOperand(i: 0);
7315 // Mark this as handled.
7316 SeenChainsForSExt[HeadOfChain] = nullptr;
7317 ValToSExtendedUses[HeadOfChain].push_back(Elt: I);
7318 }
7319 }
7320 return Promoted;
7321}
7322
7323bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
7324 BasicBlock *DefBB = I->getParent();
7325
7326 // If the result of a {s|z}ext and its source are both live out, rewrite all
7327 // other uses of the source with result of extension.
7328 Value *Src = I->getOperand(i: 0);
7329 if (Src->hasOneUse())
7330 return false;
7331
7332 // Only do this xform if truncating is free.
7333 if (!TLI->isTruncateFree(FromTy: I->getType(), ToTy: Src->getType()))
7334 return false;
7335
7336 // Only safe to perform the optimization if the source is also defined in
7337 // this block.
7338 if (!isa<Instruction>(Val: Src) || DefBB != cast<Instruction>(Val: Src)->getParent())
7339 return false;
7340
7341 bool DefIsLiveOut = false;
7342 for (User *U : I->users()) {
7343 Instruction *UI = cast<Instruction>(Val: U);
7344
7345 // Figure out which BB this ext is used in.
7346 BasicBlock *UserBB = UI->getParent();
7347 if (UserBB == DefBB)
7348 continue;
7349 DefIsLiveOut = true;
7350 break;
7351 }
7352 if (!DefIsLiveOut)
7353 return false;
7354
7355 // Make sure none of the uses are PHI nodes.
7356 for (User *U : Src->users()) {
7357 Instruction *UI = cast<Instruction>(Val: U);
7358 BasicBlock *UserBB = UI->getParent();
7359 if (UserBB == DefBB)
7360 continue;
7361 // Be conservative. We don't want this xform to end up introducing
7362 // reloads just before load / store instructions.
7363 if (isa<PHINode>(Val: UI) || isa<LoadInst>(Val: UI) || isa<StoreInst>(Val: UI))
7364 return false;
7365 }
7366
7367 // InsertedTruncs - Only insert one trunc in each block once.
7368 DenseMap<BasicBlock *, Instruction *> InsertedTruncs;
7369
7370 bool MadeChange = false;
7371 for (Use &U : Src->uses()) {
7372 Instruction *User = cast<Instruction>(Val: U.getUser());
7373
7374 // Figure out which BB this ext is used in.
7375 BasicBlock *UserBB = User->getParent();
7376 if (UserBB == DefBB)
7377 continue;
7378
7379 // Both src and def are live in this block. Rewrite the use.
7380 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
7381
7382 if (!InsertedTrunc) {
7383 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
7384 assert(InsertPt != UserBB->end());
7385 InsertedTrunc = new TruncInst(I, Src->getType(), "");
7386 InsertedTrunc->insertBefore(BB&: *UserBB, InsertPos: InsertPt);
7387 InsertedInsts.insert(Ptr: InsertedTrunc);
7388 }
7389
7390 // Replace a use of the {s|z}ext source with a use of the result.
7391 U = InsertedTrunc;
7392 ++NumExtUses;
7393 MadeChange = true;
7394 }
7395
7396 return MadeChange;
7397}
7398
7399// Find loads whose uses only use some of the loaded value's bits. Add an "and"
7400// just after the load if the target can fold this into one extload instruction,
7401// with the hope of eliminating some of the other later "and" instructions using
7402// the loaded value. "and"s that are made trivially redundant by the insertion
7403// of the new "and" are removed by this function, while others (e.g. those whose
7404// path from the load goes through a phi) are left for isel to potentially
7405// remove.
7406//
7407// For example:
7408//
7409// b0:
7410// x = load i32
7411// ...
7412// b1:
7413// y = and x, 0xff
7414// z = use y
7415//
7416// becomes:
7417//
7418// b0:
7419// x = load i32
7420// x' = and x, 0xff
7421// ...
7422// b1:
7423// z = use x'
7424//
7425// whereas:
7426//
7427// b0:
7428// x1 = load i32
7429// ...
7430// b1:
7431// x2 = load i32
7432// ...
7433// b2:
7434// x = phi x1, x2
7435// y = and x, 0xff
7436//
7437// becomes (after a call to optimizeLoadExt for each load):
7438//
7439// b0:
7440// x1 = load i32
7441// x1' = and x1, 0xff
7442// ...
7443// b1:
7444// x2 = load i32
7445// x2' = and x2, 0xff
7446// ...
7447// b2:
7448// x = phi x1', x2'
7449// y = and x, 0xff
7450bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
7451 if (!Load->isSimple() || !Load->getType()->isIntOrPtrTy())
7452 return false;
7453
7454 // Skip loads we've already transformed.
7455 if (Load->hasOneUse() &&
7456 InsertedInsts.count(Ptr: cast<Instruction>(Val: *Load->user_begin())))
7457 return false;
7458
7459 // Look at all uses of Load, looking through phis, to determine how many bits
7460 // of the loaded value are needed.
7461 SmallVector<Instruction *, 8> WorkList;
7462 SmallPtrSet<Instruction *, 16> Visited;
7463 SmallVector<Instruction *, 8> AndsToMaybeRemove;
7464 SmallVector<Instruction *, 8> DropFlags;
7465 for (auto *U : Load->users())
7466 WorkList.push_back(Elt: cast<Instruction>(Val: U));
7467
7468 EVT LoadResultVT = TLI->getValueType(DL: *DL, Ty: Load->getType());
7469 unsigned BitWidth = LoadResultVT.getSizeInBits();
7470 // If the BitWidth is 0, do not try to optimize the type
7471 if (BitWidth == 0)
7472 return false;
7473
7474 APInt DemandBits(BitWidth, 0);
7475 APInt WidestAndBits(BitWidth, 0);
7476
7477 while (!WorkList.empty()) {
7478 Instruction *I = WorkList.pop_back_val();
7479
7480 // Break use-def graph loops.
7481 if (!Visited.insert(Ptr: I).second)
7482 continue;
7483
7484 // For a PHI node, push all of its users.
7485 if (auto *Phi = dyn_cast<PHINode>(Val: I)) {
7486 for (auto *U : Phi->users())
7487 WorkList.push_back(Elt: cast<Instruction>(Val: U));
7488 continue;
7489 }
7490
7491 switch (I->getOpcode()) {
7492 case Instruction::And: {
7493 auto *AndC = dyn_cast<ConstantInt>(Val: I->getOperand(i: 1));
7494 if (!AndC)
7495 return false;
7496 APInt AndBits = AndC->getValue();
7497 DemandBits |= AndBits;
7498 // Keep track of the widest and mask we see.
7499 if (AndBits.ugt(RHS: WidestAndBits))
7500 WidestAndBits = AndBits;
7501 if (AndBits == WidestAndBits && I->getOperand(i: 0) == Load)
7502 AndsToMaybeRemove.push_back(Elt: I);
7503 break;
7504 }
7505
7506 case Instruction::Shl: {
7507 auto *ShlC = dyn_cast<ConstantInt>(Val: I->getOperand(i: 1));
7508 if (!ShlC)
7509 return false;
7510 uint64_t ShiftAmt = ShlC->getLimitedValue(Limit: BitWidth - 1);
7511 DemandBits.setLowBits(BitWidth - ShiftAmt);
7512 DropFlags.push_back(Elt: I);
7513 break;
7514 }
7515
7516 case Instruction::Trunc: {
7517 EVT TruncVT = TLI->getValueType(DL: *DL, Ty: I->getType());
7518 unsigned TruncBitWidth = TruncVT.getSizeInBits();
7519 DemandBits.setLowBits(TruncBitWidth);
7520 DropFlags.push_back(Elt: I);
7521 break;
7522 }
7523
7524 default:
7525 return false;
7526 }
7527 }
7528
7529 uint32_t ActiveBits = DemandBits.getActiveBits();
7530 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
7531 // target even if isLoadLegal says an i1 EXTLOAD is valid. For example,
7532 // for the AArch64 target isLoadLegal(i32, i1, ..., ZEXTLOAD, false) returns
7533 // true, but (and (load x) 1) is not matched as a single instruction, rather
7534 // as a LDR followed by an AND.
7535 // TODO: Look into removing this restriction by fixing backends to either
7536 // return false for isLoadLegal for i1 or have them select this pattern to
7537 // a single instruction.
7538 //
7539 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
7540 // mask, since these are the only ands that will be removed by isel.
7541 if (ActiveBits <= 1 || !DemandBits.isMask(numBits: ActiveBits) ||
7542 WidestAndBits != DemandBits)
7543 return false;
7544
7545 LLVMContext &Ctx = Load->getType()->getContext();
7546 Type *TruncTy = Type::getIntNTy(C&: Ctx, N: ActiveBits);
7547 EVT TruncVT = TLI->getValueType(DL: *DL, Ty: TruncTy);
7548
7549 // Reject cases that won't be matched as extloads.
7550 if (!LoadResultVT.bitsGT(VT: TruncVT) || !TruncVT.isRound() ||
7551 !TLI->isLoadLegal(ValVT: LoadResultVT, MemVT: TruncVT, Alignment: Load->getAlign(),
7552 AddrSpace: Load->getPointerAddressSpace(), ExtType: ISD::ZEXTLOAD, Atomic: false))
7553 return false;
7554
7555 IRBuilder<> Builder(Load->getNextNode());
7556 auto *NewAnd = cast<Instruction>(
7557 Val: Builder.CreateAnd(LHS: Load, RHS: ConstantInt::get(Context&: Ctx, V: DemandBits)));
7558 // Mark this instruction as "inserted by CGP", so that other
7559 // optimizations don't touch it.
7560 InsertedInsts.insert(Ptr: NewAnd);
7561
7562 // Replace all uses of load with new and (except for the use of load in the
7563 // new and itself).
7564 replaceAllUsesWith(Old: Load, New: NewAnd, FreshBBs, IsHuge: IsHugeFunc);
7565 NewAnd->setOperand(i: 0, Val: Load);
7566
7567 // Remove any and instructions that are now redundant.
7568 for (auto *And : AndsToMaybeRemove)
7569 // Check that the and mask is the same as the one we decided to put on the
7570 // new and.
7571 if (cast<ConstantInt>(Val: And->getOperand(i: 1))->getValue() == DemandBits) {
7572 replaceAllUsesWith(Old: And, New: NewAnd, FreshBBs, IsHuge: IsHugeFunc);
7573 if (&*CurInstIterator == And)
7574 CurInstIterator = std::next(x: And->getIterator());
7575 And->eraseFromParent();
7576 ++NumAndUses;
7577 }
7578
7579 // NSW flags may not longer hold.
7580 for (auto *Inst : DropFlags)
7581 Inst->setHasNoSignedWrap(false);
7582
7583 ++NumAndsAdded;
7584 return true;
7585}
7586
7587/// Check if V (an operand of a select instruction) is an expensive instruction
7588/// that is only used once.
7589static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
7590 auto *I = dyn_cast<Instruction>(Val: V);
7591 // If it's safe to speculatively execute, then it should not have side
7592 // effects; therefore, it's safe to sink and possibly *not* execute.
7593 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
7594 TTI->isExpensiveToSpeculativelyExecute(I);
7595}
7596
7597/// Returns true if a SelectInst should be turned into an explicit branch.
7598static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
7599 const TargetLowering *TLI,
7600 SelectInst *SI) {
7601 // If even a predictable select is cheap, then a branch can't be cheaper.
7602 if (!TLI->isPredictableSelectExpensive())
7603 return false;
7604
7605 // FIXME: This should use the same heuristics as IfConversion to determine
7606 // whether a select is better represented as a branch.
7607
7608 // If metadata tells us that the select condition is obviously predictable,
7609 // then we want to replace the select with a branch.
7610 uint64_t TrueWeight, FalseWeight;
7611 if (extractBranchWeights(I: *SI, TrueVal&: TrueWeight, FalseVal&: FalseWeight)) {
7612 uint64_t Max = std::max(a: TrueWeight, b: FalseWeight);
7613 uint64_t Sum = TrueWeight + FalseWeight;
7614 if (Sum != 0) {
7615 auto Probability = BranchProbability::getBranchProbability(Numerator: Max, Denominator: Sum);
7616 if (Probability > TTI->getPredictableBranchThreshold())
7617 return true;
7618 }
7619 }
7620
7621 CmpInst *Cmp = dyn_cast<CmpInst>(Val: SI->getCondition());
7622
7623 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
7624 // comparison condition. If the compare has more than one use, there's
7625 // probably another cmov or setcc around, so it's not worth emitting a branch.
7626 if (!Cmp || !Cmp->hasOneUse())
7627 return false;
7628
7629 // If either operand of the select is expensive and only needed on one side
7630 // of the select, we should form a branch.
7631 if (sinkSelectOperand(TTI, V: SI->getTrueValue()) ||
7632 sinkSelectOperand(TTI, V: SI->getFalseValue()))
7633 return true;
7634
7635 return false;
7636}
7637
7638/// If \p isTrue is true, return the true value of \p SI, otherwise return
7639/// false value of \p SI. If the true/false value of \p SI is defined by any
7640/// select instructions in \p Selects, look through the defining select
7641/// instruction until the true/false value is not defined in \p Selects.
7642static Value *
7643getTrueOrFalseValue(SelectInst *SI, bool isTrue,
7644 const SmallPtrSet<const Instruction *, 2> &Selects) {
7645 Value *V = nullptr;
7646
7647 for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(Ptr: DefSI);
7648 DefSI = dyn_cast<SelectInst>(Val: V)) {
7649 assert(DefSI->getCondition() == SI->getCondition() &&
7650 "The condition of DefSI does not match with SI");
7651 V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
7652 }
7653
7654 assert(V && "Failed to get select true/false value");
7655 return V;
7656}
7657
7658bool CodeGenPrepare::optimizeShiftInst(BinaryOperator *Shift) {
7659 assert(Shift->isShift() && "Expected a shift");
7660
7661 // If this is (1) a vector shift, (2) shifts by scalars are cheaper than
7662 // general vector shifts, and (3) the shift amount is a select-of-splatted
7663 // values, hoist the shifts before the select:
7664 // shift Op0, (select Cond, TVal, FVal) -->
7665 // select Cond, (shift Op0, TVal), (shift Op0, FVal)
7666 //
7667 // This is inverting a generic IR transform when we know that the cost of a
7668 // general vector shift is more than the cost of 2 shift-by-scalars.
7669 // We can't do this effectively in SDAG because we may not be able to
7670 // determine if the select operands are splats from within a basic block.
7671 Type *Ty = Shift->getType();
7672 if (!Ty->isVectorTy() || !TTI->isVectorShiftByScalarCheap(Ty))
7673 return false;
7674 Value *Cond, *TVal, *FVal;
7675 if (!match(V: Shift->getOperand(i_nocapture: 1),
7676 P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: TVal), R: m_Value(V&: FVal)))))
7677 return false;
7678 if (!isSplatValue(V: TVal) || !isSplatValue(V: FVal))
7679 return false;
7680
7681 IRBuilder<> Builder(Shift);
7682 BinaryOperator::BinaryOps Opcode = Shift->getOpcode();
7683 Value *NewTVal = Builder.CreateBinOp(Opc: Opcode, LHS: Shift->getOperand(i_nocapture: 0), RHS: TVal);
7684 Value *NewFVal = Builder.CreateBinOp(Opc: Opcode, LHS: Shift->getOperand(i_nocapture: 0), RHS: FVal);
7685 Value *NewSel = Builder.CreateSelect(C: Cond, True: NewTVal, False: NewFVal);
7686 replaceAllUsesWith(Old: Shift, New: NewSel, FreshBBs, IsHuge: IsHugeFunc);
7687 Shift->eraseFromParent();
7688 return true;
7689}
7690
7691bool CodeGenPrepare::optimizeFunnelShift(IntrinsicInst *Fsh) {
7692 Intrinsic::ID Opcode = Fsh->getIntrinsicID();
7693 assert((Opcode == Intrinsic::fshl || Opcode == Intrinsic::fshr) &&
7694 "Expected a funnel shift");
7695
7696 // If this is (1) a vector funnel shift, (2) shifts by scalars are cheaper
7697 // than general vector shifts, and (3) the shift amount is select-of-splatted
7698 // values, hoist the funnel shifts before the select:
7699 // fsh Op0, Op1, (select Cond, TVal, FVal) -->
7700 // select Cond, (fsh Op0, Op1, TVal), (fsh Op0, Op1, FVal)
7701 //
7702 // This is inverting a generic IR transform when we know that the cost of a
7703 // general vector shift is more than the cost of 2 shift-by-scalars.
7704 // We can't do this effectively in SDAG because we may not be able to
7705 // determine if the select operands are splats from within a basic block.
7706 Type *Ty = Fsh->getType();
7707 if (!Ty->isVectorTy() || !TTI->isVectorShiftByScalarCheap(Ty))
7708 return false;
7709 Value *Cond, *TVal, *FVal;
7710 if (!match(V: Fsh->getOperand(i_nocapture: 2),
7711 P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: TVal), R: m_Value(V&: FVal)))))
7712 return false;
7713 if (!isSplatValue(V: TVal) || !isSplatValue(V: FVal))
7714 return false;
7715
7716 IRBuilder<> Builder(Fsh);
7717 Value *X = Fsh->getOperand(i_nocapture: 0), *Y = Fsh->getOperand(i_nocapture: 1);
7718 Value *NewTVal = Builder.CreateIntrinsic(ID: Opcode, Types: Ty, Args: {X, Y, TVal});
7719 Value *NewFVal = Builder.CreateIntrinsic(ID: Opcode, Types: Ty, Args: {X, Y, FVal});
7720 Value *NewSel = Builder.CreateSelect(C: Cond, True: NewTVal, False: NewFVal);
7721 replaceAllUsesWith(Old: Fsh, New: NewSel, FreshBBs, IsHuge: IsHugeFunc);
7722 Fsh->eraseFromParent();
7723 return true;
7724}
7725
7726/// If we have a SelectInst that will likely profit from branch prediction,
7727/// turn it into a branch.
7728bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
7729 if (DisableSelectToBranch)
7730 return false;
7731
7732 // If the SelectOptimize pass is enabled, selects have already been optimized.
7733 if (!getCGPassBuilderOption().DisableSelectOptimize)
7734 return false;
7735
7736 // Find all consecutive select instructions that share the same condition.
7737 SmallVector<SelectInst *, 2> ASI;
7738 ASI.push_back(Elt: SI);
7739 for (BasicBlock::iterator It = ++BasicBlock::iterator(SI);
7740 It != SI->getParent()->end(); ++It) {
7741 SelectInst *I = dyn_cast<SelectInst>(Val: &*It);
7742 if (I && SI->getCondition() == I->getCondition()) {
7743 ASI.push_back(Elt: I);
7744 } else {
7745 break;
7746 }
7747 }
7748
7749 SelectInst *LastSI = ASI.back();
7750 // Increment the current iterator to skip all the rest of select instructions
7751 // because they will be either "not lowered" or "all lowered" to branch.
7752 CurInstIterator = std::next(x: LastSI->getIterator());
7753 // Examine debug-info attached to the consecutive select instructions. They
7754 // won't be individually optimised by optimizeInst, so we need to perform
7755 // DbgVariableRecord maintenence here instead.
7756 for (SelectInst *SI : ArrayRef(ASI).drop_front())
7757 fixupDbgVariableRecordsOnInst(I&: *SI);
7758
7759 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(Bitwidth: 1);
7760
7761 // Can we convert the 'select' to CF ?
7762 if (VectorCond || SI->getMetadata(KindID: LLVMContext::MD_unpredictable))
7763 return false;
7764
7765 TargetLowering::SelectSupportKind SelectKind;
7766 if (SI->getType()->isVectorTy())
7767 SelectKind = TargetLowering::ScalarCondVectorVal;
7768 else
7769 SelectKind = TargetLowering::ScalarValSelect;
7770
7771 if (TLI->isSelectSupported(SelectKind) &&
7772 (!isFormingBranchFromSelectProfitable(TTI, TLI, SI) ||
7773 llvm::shouldOptimizeForSize(BB: SI->getParent(), PSI, BFI)))
7774 return false;
7775
7776 // Transform a sequence like this:
7777 // start:
7778 // %cmp = cmp uge i32 %a, %b
7779 // %sel = select i1 %cmp, i32 %c, i32 %d
7780 //
7781 // Into:
7782 // start:
7783 // %cmp = cmp uge i32 %a, %b
7784 // %cmp.frozen = freeze %cmp
7785 // br i1 %cmp.frozen, label %select.true, label %select.false
7786 // select.true:
7787 // br label %select.end
7788 // select.false:
7789 // br label %select.end
7790 // select.end:
7791 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
7792 //
7793 // %cmp should be frozen, otherwise it may introduce undefined behavior.
7794 // In addition, we may sink instructions that produce %c or %d from
7795 // the entry block into the destination(s) of the new branch.
7796 // If the true or false blocks do not contain a sunken instruction, that
7797 // block and its branch may be optimized away. In that case, one side of the
7798 // first branch will point directly to select.end, and the corresponding PHI
7799 // predecessor block will be the start block.
7800 // The CFG is altered here and we update the DominatorTree and the LoopInfo,
7801 // but we don't set a ModifiedDT flag to avoid restarting the function walk in
7802 // runOnFunction for each select optimized.
7803
7804 // Collect values that go on the true side and the values that go on the false
7805 // side.
7806 SmallVector<Instruction *> TrueInstrs, FalseInstrs;
7807 for (SelectInst *SI : ASI) {
7808 if (Value *V = SI->getTrueValue(); sinkSelectOperand(TTI, V))
7809 TrueInstrs.push_back(Elt: cast<Instruction>(Val: V));
7810 if (Value *V = SI->getFalseValue(); sinkSelectOperand(TTI, V))
7811 FalseInstrs.push_back(Elt: cast<Instruction>(Val: V));
7812 }
7813
7814 // Split the select block, according to how many (if any) values go on each
7815 // side.
7816 BasicBlock *StartBlock = SI->getParent();
7817 BasicBlock::iterator SplitPt = std::next(x: BasicBlock::iterator(LastSI));
7818 // We should split before any debug-info.
7819 SplitPt.setHeadBit(true);
7820
7821 IRBuilder<> IB(SI);
7822 auto *CondFr = IB.CreateFreeze(V: SI->getCondition(), Name: SI->getName() + ".frozen");
7823
7824 BasicBlock *TrueBlock = nullptr;
7825 BasicBlock *FalseBlock = nullptr;
7826 BasicBlock *EndBlock = nullptr;
7827 UncondBrInst *TrueBranch = nullptr;
7828 UncondBrInst *FalseBranch = nullptr;
7829 if (TrueInstrs.size() == 0) {
7830 FalseBranch = cast<UncondBrInst>(
7831 Val: SplitBlockAndInsertIfElse(Cond: CondFr, SplitBefore: SplitPt, Unreachable: false, BranchWeights: nullptr, DTU, LI));
7832 FalseBlock = FalseBranch->getParent();
7833 EndBlock = cast<BasicBlock>(Val: FalseBranch->getOperand(i_nocapture: 0));
7834 } else if (FalseInstrs.size() == 0) {
7835 TrueBranch = cast<UncondBrInst>(
7836 Val: SplitBlockAndInsertIfThen(Cond: CondFr, SplitBefore: SplitPt, Unreachable: false, BranchWeights: nullptr, DTU, LI));
7837 TrueBlock = TrueBranch->getParent();
7838 EndBlock = TrueBranch->getSuccessor();
7839 } else {
7840 Instruction *ThenTerm = nullptr;
7841 Instruction *ElseTerm = nullptr;
7842 SplitBlockAndInsertIfThenElse(Cond: CondFr, SplitBefore: SplitPt, ThenTerm: &ThenTerm, ElseTerm: &ElseTerm,
7843 BranchWeights: nullptr, DTU, LI);
7844 TrueBranch = cast<UncondBrInst>(Val: ThenTerm);
7845 FalseBranch = cast<UncondBrInst>(Val: ElseTerm);
7846 TrueBlock = TrueBranch->getParent();
7847 FalseBlock = FalseBranch->getParent();
7848 EndBlock = TrueBranch->getSuccessor();
7849 }
7850
7851 EndBlock->setName("select.end");
7852 if (TrueBlock)
7853 TrueBlock->setName("select.true.sink");
7854 if (FalseBlock)
7855 FalseBlock->setName(FalseInstrs.size() == 0 ? "select.false"
7856 : "select.false.sink");
7857
7858 if (IsHugeFunc) {
7859 if (TrueBlock)
7860 FreshBBs.insert(Ptr: TrueBlock);
7861 if (FalseBlock)
7862 FreshBBs.insert(Ptr: FalseBlock);
7863 FreshBBs.insert(Ptr: EndBlock);
7864 }
7865
7866 BFI->setBlockFreq(BB: EndBlock, Freq: BFI->getBlockFreq(BB: StartBlock));
7867
7868 static const unsigned MD[] = {
7869 LLVMContext::MD_prof, LLVMContext::MD_unpredictable,
7870 LLVMContext::MD_make_implicit, LLVMContext::MD_dbg};
7871 StartBlock->getTerminator()->copyMetadata(SrcInst: *SI, WL: MD);
7872
7873 // Sink expensive instructions into the conditional blocks to avoid executing
7874 // them speculatively.
7875 for (Instruction *I : TrueInstrs)
7876 I->moveBefore(InsertPos: TrueBranch->getIterator());
7877 for (Instruction *I : FalseInstrs)
7878 I->moveBefore(InsertPos: FalseBranch->getIterator());
7879
7880 // If we did not create a new block for one of the 'true' or 'false' paths
7881 // of the condition, it means that side of the branch goes to the end block
7882 // directly and the path originates from the start block from the point of
7883 // view of the new PHI.
7884 if (TrueBlock == nullptr)
7885 TrueBlock = StartBlock;
7886 else if (FalseBlock == nullptr)
7887 FalseBlock = StartBlock;
7888
7889 SmallPtrSet<const Instruction *, 2> INS(llvm::from_range, ASI);
7890 // Use reverse iterator because later select may use the value of the
7891 // earlier select, and we need to propagate value through earlier select
7892 // to get the PHI operand.
7893 for (SelectInst *SI : llvm::reverse(C&: ASI)) {
7894 // The select itself is replaced with a PHI Node.
7895 PHINode *PN = PHINode::Create(Ty: SI->getType(), NumReservedValues: 2, NameStr: "");
7896 PN->insertBefore(InsertPos: EndBlock->begin());
7897 PN->takeName(V: SI);
7898 PN->addIncoming(V: getTrueOrFalseValue(SI, isTrue: true, Selects: INS), BB: TrueBlock);
7899 PN->addIncoming(V: getTrueOrFalseValue(SI, isTrue: false, Selects: INS), BB: FalseBlock);
7900 PN->setDebugLoc(SI->getDebugLoc());
7901
7902 replaceAllUsesWith(Old: SI, New: PN, FreshBBs, IsHuge: IsHugeFunc);
7903 SI->eraseFromParent();
7904 INS.erase(Ptr: SI);
7905 ++NumSelectsExpanded;
7906 }
7907
7908 // Instruct OptimizeBlock to skip to the next block.
7909 CurInstIterator = StartBlock->end();
7910 return true;
7911}
7912
7913/// Some targets only accept certain types for splat inputs. For example a VDUP
7914/// in MVE takes a GPR (integer) register, and the instruction that incorporate
7915/// a VDUP (such as a VADD qd, qm, rm) also require a gpr register.
7916bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
7917 // Accept shuf(insertelem(undef/poison, val, 0), undef/poison, <0,0,..>) only
7918 if (!match(V: SVI, P: m_Shuffle(v1: m_InsertElt(Val: m_Undef(), Elt: m_Value(), Idx: m_ZeroInt()),
7919 v2: m_Undef(), mask: m_ZeroMask())))
7920 return false;
7921 Type *NewType = TLI->shouldConvertSplatType(SVI);
7922 if (!NewType)
7923 return false;
7924
7925 auto *SVIVecType = cast<FixedVectorType>(Val: SVI->getType());
7926 assert(!NewType->isVectorTy() && "Expected a scalar type!");
7927 assert(NewType->getScalarSizeInBits() == SVIVecType->getScalarSizeInBits() &&
7928 "Expected a type of the same size!");
7929 auto *NewVecType =
7930 FixedVectorType::get(ElementType: NewType, NumElts: SVIVecType->getNumElements());
7931
7932 // Create a bitcast (shuffle (insert (bitcast(..))))
7933 IRBuilder<> Builder(SVI->getContext());
7934 Builder.SetInsertPoint(SVI);
7935 Value *BC1 = Builder.CreateBitCast(
7936 V: cast<Instruction>(Val: SVI->getOperand(i_nocapture: 0))->getOperand(i: 1), DestTy: NewType);
7937 Value *Shuffle = Builder.CreateVectorSplat(NumElts: NewVecType->getNumElements(), V: BC1);
7938 Value *BC2 = Builder.CreateBitCast(V: Shuffle, DestTy: SVIVecType);
7939
7940 replaceAllUsesWith(Old: SVI, New: BC2, FreshBBs, IsHuge: IsHugeFunc);
7941 RecursivelyDeleteTriviallyDeadInstructions(
7942 V: SVI, TLI: TLInfo, MSSAU: nullptr,
7943 AboutToDeleteCallback: [&](Value *V) { removeAllAssertingVHReferences(V); });
7944
7945 // Also hoist the bitcast up to its operand if it they are not in the same
7946 // block.
7947 if (auto *BCI = dyn_cast<Instruction>(Val: BC1))
7948 if (auto *Op = dyn_cast<Instruction>(Val: BCI->getOperand(i: 0)))
7949 if (BCI->getParent() != Op->getParent() && !isa<PHINode>(Val: Op) &&
7950 !Op->isTerminator() && !Op->isEHPad())
7951 BCI->moveAfter(MovePos: Op);
7952
7953 return true;
7954}
7955
7956bool CodeGenPrepare::tryToSinkFreeOperands(Instruction *I) {
7957 // If the operands of I can be folded into a target instruction together with
7958 // I, duplicate and sink them.
7959 SmallVector<Use *, 4> OpsToSink;
7960 if (!TTI->isProfitableToSinkOperands(I, Ops&: OpsToSink))
7961 return false;
7962
7963 // OpsToSink can contain multiple uses in a use chain (e.g.
7964 // (%u1 with %u1 = shufflevector), (%u2 with %u2 = zext %u1)). The dominating
7965 // uses must come first, so we process the ops in reverse order so as to not
7966 // create invalid IR.
7967 BasicBlock *TargetBB = I->getParent();
7968 bool Changed = false;
7969 SmallVector<Use *, 4> ToReplace;
7970 Instruction *InsertPoint = I;
7971 DenseMap<const Instruction *, unsigned long> InstOrdering;
7972 unsigned long InstNumber = 0;
7973 for (const auto &I : *TargetBB)
7974 InstOrdering[&I] = InstNumber++;
7975
7976 for (Use *U : reverse(C&: OpsToSink)) {
7977 auto *UI = cast<Instruction>(Val: U->get());
7978 if (isa<PHINode>(Val: UI) || UI->mayHaveSideEffects() || UI->mayReadFromMemory())
7979 continue;
7980 if (UI->getParent() == TargetBB) {
7981 if (InstOrdering[UI] < InstOrdering[InsertPoint])
7982 InsertPoint = UI;
7983 continue;
7984 }
7985 ToReplace.push_back(Elt: U);
7986 }
7987
7988 SetVector<Instruction *> MaybeDead;
7989 DenseMap<Instruction *, Instruction *> NewInstructions;
7990 for (Use *U : ToReplace) {
7991 auto *UI = cast<Instruction>(Val: U->get());
7992 Instruction *NI = UI->clone();
7993
7994 if (IsHugeFunc) {
7995 // Now we clone an instruction, its operands' defs may sink to this BB
7996 // now. So we put the operands defs' BBs into FreshBBs to do optimization.
7997 for (Value *Op : NI->operands())
7998 if (auto *OpDef = dyn_cast<Instruction>(Val: Op))
7999 FreshBBs.insert(Ptr: OpDef->getParent());
8000 }
8001
8002 NewInstructions[UI] = NI;
8003 MaybeDead.insert(X: UI);
8004 LLVM_DEBUG(dbgs() << "Sinking " << *UI << " to user " << *I << "\n");
8005 NI->insertBefore(InsertPos: InsertPoint->getIterator());
8006 InsertPoint = NI;
8007 InsertedInsts.insert(Ptr: NI);
8008
8009 // Update the use for the new instruction, making sure that we update the
8010 // sunk instruction uses, if it is part of a chain that has already been
8011 // sunk.
8012 Instruction *OldI = cast<Instruction>(Val: U->getUser());
8013 if (auto It = NewInstructions.find(Val: OldI); It != NewInstructions.end())
8014 It->second->setOperand(i: U->getOperandNo(), Val: NI);
8015 else
8016 U->set(NI);
8017 Changed = true;
8018 }
8019
8020 // Remove instructions that are dead after sinking.
8021 for (auto *I : MaybeDead) {
8022 if (!I->hasNUsesOrMore(N: 1)) {
8023 LLVM_DEBUG(dbgs() << "Removing dead instruction: " << *I << "\n");
8024 I->eraseFromParent();
8025 }
8026 }
8027
8028 return Changed;
8029}
8030
8031bool CodeGenPrepare::optimizeSwitchType(SwitchInst *SI) {
8032 Value *Cond = SI->getCondition();
8033 Type *OldType = Cond->getType();
8034 LLVMContext &Context = Cond->getContext();
8035 EVT OldVT = TLI->getValueType(DL: *DL, Ty: OldType);
8036 MVT RegType = TLI->getPreferredSwitchConditionType(Context, ConditionVT: OldVT);
8037 unsigned RegWidth = RegType.getSizeInBits();
8038
8039 if (RegWidth <= cast<IntegerType>(Val: OldType)->getBitWidth())
8040 return false;
8041
8042 // If the register width is greater than the type width, expand the condition
8043 // of the switch instruction and each case constant to the width of the
8044 // register. By widening the type of the switch condition, subsequent
8045 // comparisons (for case comparisons) will not need to be extended to the
8046 // preferred register width, so we will potentially eliminate N-1 extends,
8047 // where N is the number of cases in the switch.
8048 auto *NewType = Type::getIntNTy(C&: Context, N: RegWidth);
8049
8050 // Extend the switch condition and case constants using the target preferred
8051 // extend unless the switch condition is a function argument with an extend
8052 // attribute. In that case, we can avoid an unnecessary mask/extension by
8053 // matching the argument extension instead.
8054 Instruction::CastOps ExtType = Instruction::ZExt;
8055 // Some targets prefer SExt over ZExt.
8056 if (TLI->isSExtCheaperThanZExt(FromTy: OldVT, ToTy: RegType))
8057 ExtType = Instruction::SExt;
8058
8059 if (auto *Arg = dyn_cast<Argument>(Val: Cond)) {
8060 if (Arg->hasSExtAttr())
8061 ExtType = Instruction::SExt;
8062 if (Arg->hasZExtAttr())
8063 ExtType = Instruction::ZExt;
8064 }
8065
8066 auto *ExtInst = CastInst::Create(ExtType, S: Cond, Ty: NewType);
8067 ExtInst->insertBefore(InsertPos: SI->getIterator());
8068 ExtInst->setDebugLoc(SI->getDebugLoc());
8069 SI->setCondition(ExtInst);
8070 for (auto Case : SI->cases()) {
8071 const APInt &NarrowConst = Case.getCaseValue()->getValue();
8072 APInt WideConst = (ExtType == Instruction::ZExt)
8073 ? NarrowConst.zext(width: RegWidth)
8074 : NarrowConst.sext(width: RegWidth);
8075 Case.setValue(ConstantInt::get(Context, V: WideConst));
8076 }
8077
8078 return true;
8079}
8080
8081bool CodeGenPrepare::optimizeSwitchPhiConstants(SwitchInst *SI) {
8082 // The SCCP optimization tends to produce code like this:
8083 // switch(x) { case 42: phi(42, ...) }
8084 // Materializing the constant for the phi-argument needs instructions; So we
8085 // change the code to:
8086 // switch(x) { case 42: phi(x, ...) }
8087
8088 Value *Condition = SI->getCondition();
8089 // Avoid endless loop in degenerate case.
8090 if (isa<ConstantInt>(Val: *Condition))
8091 return false;
8092
8093 bool Changed = false;
8094 BasicBlock *SwitchBB = SI->getParent();
8095 Type *ConditionType = Condition->getType();
8096
8097 for (const SwitchInst::CaseHandle &Case : SI->cases()) {
8098 ConstantInt *CaseValue = Case.getCaseValue();
8099 BasicBlock *CaseBB = Case.getCaseSuccessor();
8100 // Set to true if we previously checked that `CaseBB` is only reached by
8101 // a single case from this switch.
8102 bool CheckedForSinglePred = false;
8103 for (PHINode &PHI : CaseBB->phis()) {
8104 Type *PHIType = PHI.getType();
8105 // If ZExt is free then we can also catch patterns like this:
8106 // switch((i32)x) { case 42: phi((i64)42, ...); }
8107 // and replace `(i64)42` with `zext i32 %x to i64`.
8108 bool TryZExt =
8109 PHIType->isIntegerTy() &&
8110 PHIType->getIntegerBitWidth() > ConditionType->getIntegerBitWidth() &&
8111 TLI->isZExtFree(FromTy: ConditionType, ToTy: PHIType);
8112 if (PHIType == ConditionType || TryZExt) {
8113 // Set to true to skip this case because of multiple preds.
8114 bool SkipCase = false;
8115 Value *Replacement = nullptr;
8116 for (unsigned I = 0, E = PHI.getNumIncomingValues(); I != E; I++) {
8117 Value *PHIValue = PHI.getIncomingValue(i: I);
8118 if (PHIValue != CaseValue) {
8119 if (!TryZExt)
8120 continue;
8121 ConstantInt *PHIValueInt = dyn_cast<ConstantInt>(Val: PHIValue);
8122 if (!PHIValueInt ||
8123 PHIValueInt->getValue() !=
8124 CaseValue->getValue().zext(width: PHIType->getIntegerBitWidth()))
8125 continue;
8126 }
8127 if (PHI.getIncomingBlock(i: I) != SwitchBB)
8128 continue;
8129 // We cannot optimize if there are multiple case labels jumping to
8130 // this block. This check may get expensive when there are many
8131 // case labels so we test for it last.
8132 if (!CheckedForSinglePred) {
8133 CheckedForSinglePred = true;
8134 if (SI->findCaseDest(BB: CaseBB) == nullptr) {
8135 SkipCase = true;
8136 break;
8137 }
8138 }
8139
8140 if (Replacement == nullptr) {
8141 if (PHIValue == CaseValue) {
8142 Replacement = Condition;
8143 } else {
8144 IRBuilder<> Builder(SI);
8145 Replacement = Builder.CreateZExt(V: Condition, DestTy: PHIType);
8146 }
8147 }
8148 PHI.setIncomingValue(i: I, V: Replacement);
8149 Changed = true;
8150 }
8151 if (SkipCase)
8152 break;
8153 }
8154 }
8155 }
8156 return Changed;
8157}
8158
8159bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
8160 bool Changed = optimizeSwitchType(SI);
8161 Changed |= optimizeSwitchPhiConstants(SI);
8162 return Changed;
8163}
8164
8165namespace {
8166
8167/// Helper class to promote a scalar operation to a vector one.
8168/// This class is used to move downward extractelement transition.
8169/// E.g.,
8170/// a = vector_op <2 x i32>
8171/// b = extractelement <2 x i32> a, i32 0
8172/// c = scalar_op b
8173/// store c
8174///
8175/// =>
8176/// a = vector_op <2 x i32>
8177/// c = vector_op a (equivalent to scalar_op on the related lane)
8178/// * d = extractelement <2 x i32> c, i32 0
8179/// * store d
8180/// Assuming both extractelement and store can be combine, we get rid of the
8181/// transition.
8182class VectorPromoteHelper {
8183 /// DataLayout associated with the current module.
8184 const DataLayout &DL;
8185
8186 /// Used to perform some checks on the legality of vector operations.
8187 const TargetLowering &TLI;
8188
8189 /// Used to estimated the cost of the promoted chain.
8190 const TargetTransformInfo &TTI;
8191
8192 /// The transition being moved downwards.
8193 Instruction *Transition;
8194
8195 /// The sequence of instructions to be promoted.
8196 SmallVector<Instruction *, 4> InstsToBePromoted;
8197
8198 /// Cost of combining a store and an extract.
8199 unsigned StoreExtractCombineCost;
8200
8201 /// Instruction that will be combined with the transition.
8202 Instruction *CombineInst = nullptr;
8203
8204 /// The instruction that represents the current end of the transition.
8205 /// Since we are faking the promotion until we reach the end of the chain
8206 /// of computation, we need a way to get the current end of the transition.
8207 Instruction *getEndOfTransition() const {
8208 if (InstsToBePromoted.empty())
8209 return Transition;
8210 return InstsToBePromoted.back();
8211 }
8212
8213 /// Return the index of the original value in the transition.
8214 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
8215 /// c, is at index 0.
8216 unsigned getTransitionOriginalValueIdx() const {
8217 assert(isa<ExtractElementInst>(Transition) &&
8218 "Other kind of transitions are not supported yet");
8219 return 0;
8220 }
8221
8222 /// Return the index of the index in the transition.
8223 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
8224 /// is at index 1.
8225 unsigned getTransitionIdx() const {
8226 assert(isa<ExtractElementInst>(Transition) &&
8227 "Other kind of transitions are not supported yet");
8228 return 1;
8229 }
8230
8231 /// Get the type of the transition.
8232 /// This is the type of the original value.
8233 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
8234 /// transition is <2 x i32>.
8235 Type *getTransitionType() const {
8236 return Transition->getOperand(i: getTransitionOriginalValueIdx())->getType();
8237 }
8238
8239 /// Promote \p ToBePromoted by moving \p Def downward through.
8240 /// I.e., we have the following sequence:
8241 /// Def = Transition <ty1> a to <ty2>
8242 /// b = ToBePromoted <ty2> Def, ...
8243 /// =>
8244 /// b = ToBePromoted <ty1> a, ...
8245 /// Def = Transition <ty1> ToBePromoted to <ty2>
8246 void promoteImpl(Instruction *ToBePromoted);
8247
8248 /// Check whether or not it is profitable to promote all the
8249 /// instructions enqueued to be promoted.
8250 bool isProfitableToPromote() {
8251 Value *ValIdx = Transition->getOperand(i: getTransitionOriginalValueIdx());
8252 unsigned Index = isa<ConstantInt>(Val: ValIdx)
8253 ? cast<ConstantInt>(Val: ValIdx)->getZExtValue()
8254 : -1;
8255 Type *PromotedType = getTransitionType();
8256
8257 StoreInst *ST = cast<StoreInst>(Val: CombineInst);
8258 unsigned AS = ST->getPointerAddressSpace();
8259 // Check if this store is supported.
8260 if (!TLI.allowsMisalignedMemoryAccesses(
8261 TLI.getValueType(DL, Ty: ST->getValueOperand()->getType()), AddrSpace: AS,
8262 Alignment: ST->getAlign())) {
8263 // If this is not supported, there is no way we can combine
8264 // the extract with the store.
8265 return false;
8266 }
8267
8268 // The scalar chain of computation has to pay for the transition
8269 // scalar to vector.
8270 // The vector chain has to account for the combining cost.
8271 enum TargetTransformInfo::TargetCostKind CostKind =
8272 TargetTransformInfo::TCK_RecipThroughput;
8273 InstructionCost ScalarCost =
8274 TTI.getVectorInstrCost(I: *Transition, Val: PromotedType, CostKind, Index);
8275 InstructionCost VectorCost = StoreExtractCombineCost;
8276 for (const auto &Inst : InstsToBePromoted) {
8277 // Compute the cost.
8278 // By construction, all instructions being promoted are arithmetic ones.
8279 // Moreover, one argument is a constant that can be viewed as a splat
8280 // constant.
8281 Value *Arg0 = Inst->getOperand(i: 0);
8282 bool IsArg0Constant = isa<UndefValue>(Val: Arg0) || isa<ConstantInt>(Val: Arg0) ||
8283 isa<ConstantFP>(Val: Arg0);
8284 TargetTransformInfo::OperandValueInfo Arg0Info, Arg1Info;
8285 if (IsArg0Constant)
8286 Arg0Info.Kind = TargetTransformInfo::OK_UniformConstantValue;
8287 else
8288 Arg1Info.Kind = TargetTransformInfo::OK_UniformConstantValue;
8289
8290 ScalarCost += TTI.getArithmeticInstrCost(
8291 Opcode: Inst->getOpcode(), Ty: Inst->getType(), CostKind, Opd1Info: Arg0Info, Opd2Info: Arg1Info);
8292 VectorCost += TTI.getArithmeticInstrCost(Opcode: Inst->getOpcode(), Ty: PromotedType,
8293 CostKind, Opd1Info: Arg0Info, Opd2Info: Arg1Info);
8294 }
8295 LLVM_DEBUG(
8296 dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
8297 << ScalarCost << "\nVector: " << VectorCost << '\n');
8298 return ScalarCost > VectorCost;
8299 }
8300
8301 /// Generate a constant vector with \p Val with the same
8302 /// number of elements as the transition.
8303 /// \p UseSplat defines whether or not \p Val should be replicated
8304 /// across the whole vector.
8305 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
8306 /// otherwise we generate a vector with as many poison as possible:
8307 /// <poison, ..., poison, Val, poison, ..., poison> where \p Val is only
8308 /// used at the index of the extract.
8309 Value *getConstantVector(Constant *Val, bool UseSplat) const {
8310 unsigned ExtractIdx = std::numeric_limits<unsigned>::max();
8311 if (!UseSplat) {
8312 // If we cannot determine where the constant must be, we have to
8313 // use a splat constant.
8314 Value *ValExtractIdx = Transition->getOperand(i: getTransitionIdx());
8315 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Val: ValExtractIdx))
8316 ExtractIdx = CstVal->getSExtValue();
8317 else
8318 UseSplat = true;
8319 }
8320
8321 ElementCount EC = cast<VectorType>(Val: getTransitionType())->getElementCount();
8322 if (UseSplat)
8323 return ConstantVector::getSplat(EC, Elt: Val);
8324
8325 if (!EC.isScalable()) {
8326 SmallVector<Constant *, 4> ConstVec;
8327 PoisonValue *PoisonVal = PoisonValue::get(T: Val->getType());
8328 for (unsigned Idx = 0; Idx != EC.getKnownMinValue(); ++Idx) {
8329 if (Idx == ExtractIdx)
8330 ConstVec.push_back(Elt: Val);
8331 else
8332 ConstVec.push_back(Elt: PoisonVal);
8333 }
8334 return ConstantVector::get(V: ConstVec);
8335 } else
8336 llvm_unreachable(
8337 "Generate scalable vector for non-splat is unimplemented");
8338 }
8339
8340 /// Check if promoting to a vector type an operand at \p OperandIdx
8341 /// in \p Use can trigger undefined behavior.
8342 static bool canCauseUndefinedBehavior(const Instruction *Use,
8343 unsigned OperandIdx) {
8344 // This is not safe to introduce undef when the operand is on
8345 // the right hand side of a division-like instruction.
8346 if (OperandIdx != 1)
8347 return false;
8348 switch (Use->getOpcode()) {
8349 default:
8350 return false;
8351 case Instruction::SDiv:
8352 case Instruction::UDiv:
8353 case Instruction::SRem:
8354 case Instruction::URem:
8355 return true;
8356 case Instruction::FDiv:
8357 case Instruction::FRem:
8358 return !Use->hasNoNaNs();
8359 }
8360 llvm_unreachable(nullptr);
8361 }
8362
8363public:
8364 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
8365 const TargetTransformInfo &TTI, Instruction *Transition,
8366 unsigned CombineCost)
8367 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
8368 StoreExtractCombineCost(CombineCost) {
8369 assert(Transition && "Do not know how to promote null");
8370 }
8371
8372 /// Check if we can promote \p ToBePromoted to \p Type.
8373 bool canPromote(const Instruction *ToBePromoted) const {
8374 // We could support CastInst too.
8375 return isa<BinaryOperator>(Val: ToBePromoted);
8376 }
8377
8378 /// Check if it is profitable to promote \p ToBePromoted
8379 /// by moving downward the transition through.
8380 bool shouldPromote(const Instruction *ToBePromoted) const {
8381 // Promote only if all the operands can be statically expanded.
8382 // Indeed, we do not want to introduce any new kind of transitions.
8383 for (const Use &U : ToBePromoted->operands()) {
8384 const Value *Val = U.get();
8385 if (Val == getEndOfTransition()) {
8386 // If the use is a division and the transition is on the rhs,
8387 // we cannot promote the operation, otherwise we may create a
8388 // division by zero.
8389 if (canCauseUndefinedBehavior(Use: ToBePromoted, OperandIdx: U.getOperandNo()))
8390 return false;
8391 continue;
8392 }
8393 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
8394 !isa<ConstantFP>(Val))
8395 return false;
8396 }
8397 // Check that the resulting operation is legal.
8398 int ISDOpcode = TLI.InstructionOpcodeToISD(Opcode: ToBePromoted->getOpcode());
8399 if (!ISDOpcode)
8400 return false;
8401 return StressStoreExtract ||
8402 TLI.isOperationLegalOrCustom(
8403 Op: ISDOpcode, VT: TLI.getValueType(DL, Ty: getTransitionType(), AllowUnknown: true));
8404 }
8405
8406 /// Check whether or not \p Use can be combined
8407 /// with the transition.
8408 /// I.e., is it possible to do Use(Transition) => AnotherUse?
8409 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Val: Use); }
8410
8411 /// Record \p ToBePromoted as part of the chain to be promoted.
8412 void enqueueForPromotion(Instruction *ToBePromoted) {
8413 InstsToBePromoted.push_back(Elt: ToBePromoted);
8414 }
8415
8416 /// Set the instruction that will be combined with the transition.
8417 void recordCombineInstruction(Instruction *ToBeCombined) {
8418 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
8419 CombineInst = ToBeCombined;
8420 }
8421
8422 /// Promote all the instructions enqueued for promotion if it is
8423 /// is profitable.
8424 /// \return True if the promotion happened, false otherwise.
8425 bool promote() {
8426 // Check if there is something to promote.
8427 // Right now, if we do not have anything to combine with,
8428 // we assume the promotion is not profitable.
8429 if (InstsToBePromoted.empty() || !CombineInst)
8430 return false;
8431
8432 // Check cost.
8433 if (!StressStoreExtract && !isProfitableToPromote())
8434 return false;
8435
8436 // Promote.
8437 for (auto &ToBePromoted : InstsToBePromoted)
8438 promoteImpl(ToBePromoted);
8439 InstsToBePromoted.clear();
8440 return true;
8441 }
8442};
8443
8444} // end anonymous namespace
8445
8446void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
8447 // At this point, we know that all the operands of ToBePromoted but Def
8448 // can be statically promoted.
8449 // For Def, we need to use its parameter in ToBePromoted:
8450 // b = ToBePromoted ty1 a
8451 // Def = Transition ty1 b to ty2
8452 // Move the transition down.
8453 // 1. Replace all uses of the promoted operation by the transition.
8454 // = ... b => = ... Def.
8455 assert(ToBePromoted->getType() == Transition->getType() &&
8456 "The type of the result of the transition does not match "
8457 "the final type");
8458 ToBePromoted->replaceAllUsesWith(V: Transition);
8459 // 2. Update the type of the uses.
8460 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
8461 Type *TransitionTy = getTransitionType();
8462 ToBePromoted->mutateType(Ty: TransitionTy);
8463 // 3. Update all the operands of the promoted operation with promoted
8464 // operands.
8465 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
8466 for (Use &U : ToBePromoted->operands()) {
8467 Value *Val = U.get();
8468 Value *NewVal = nullptr;
8469 if (Val == Transition)
8470 NewVal = Transition->getOperand(i: getTransitionOriginalValueIdx());
8471 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
8472 isa<ConstantFP>(Val)) {
8473 // Use a splat constant if it is not safe to use undef.
8474 NewVal = getConstantVector(
8475 Val: cast<Constant>(Val),
8476 UseSplat: isa<UndefValue>(Val) ||
8477 canCauseUndefinedBehavior(Use: ToBePromoted, OperandIdx: U.getOperandNo()));
8478 } else
8479 llvm_unreachable("Did you modified shouldPromote and forgot to update "
8480 "this?");
8481 ToBePromoted->setOperand(i: U.getOperandNo(), Val: NewVal);
8482 }
8483 Transition->moveAfter(MovePos: ToBePromoted);
8484 Transition->setOperand(i: getTransitionOriginalValueIdx(), Val: ToBePromoted);
8485}
8486
8487/// Some targets can do store(extractelement) with one instruction.
8488/// Try to push the extractelement towards the stores when the target
8489/// has this feature and this is profitable.
8490bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
8491 unsigned CombineCost = std::numeric_limits<unsigned>::max();
8492 if (DisableStoreExtract ||
8493 (!StressStoreExtract &&
8494 !TLI->canCombineStoreAndExtract(VectorTy: Inst->getOperand(i: 0)->getType(),
8495 Idx: Inst->getOperand(i: 1), Cost&: CombineCost)))
8496 return false;
8497
8498 // At this point we know that Inst is a vector to scalar transition.
8499 // Try to move it down the def-use chain, until:
8500 // - We can combine the transition with its single use
8501 // => we got rid of the transition.
8502 // - We escape the current basic block
8503 // => we would need to check that we are moving it at a cheaper place and
8504 // we do not do that for now.
8505 BasicBlock *Parent = Inst->getParent();
8506 LLVM_DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
8507 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
8508 // If the transition has more than one use, assume this is not going to be
8509 // beneficial.
8510 while (Inst->hasOneUse()) {
8511 Instruction *ToBePromoted = cast<Instruction>(Val: *Inst->user_begin());
8512 LLVM_DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
8513
8514 if (ToBePromoted->getParent() != Parent) {
8515 LLVM_DEBUG(dbgs() << "Instruction to promote is in a different block ("
8516 << ToBePromoted->getParent()->getName()
8517 << ") than the transition (" << Parent->getName()
8518 << ").\n");
8519 return false;
8520 }
8521
8522 if (VPH.canCombine(Use: ToBePromoted)) {
8523 LLVM_DEBUG(dbgs() << "Assume " << *Inst << '\n'
8524 << "will be combined with: " << *ToBePromoted << '\n');
8525 VPH.recordCombineInstruction(ToBeCombined: ToBePromoted);
8526 bool Changed = VPH.promote();
8527 NumStoreExtractExposed += Changed;
8528 return Changed;
8529 }
8530
8531 LLVM_DEBUG(dbgs() << "Try promoting.\n");
8532 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
8533 return false;
8534
8535 LLVM_DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
8536
8537 VPH.enqueueForPromotion(ToBePromoted);
8538 Inst = ToBePromoted;
8539 }
8540 return false;
8541}
8542
8543/// For the instruction sequence of store below, F and I values
8544/// are bundled together as an i64 value before being stored into memory.
8545/// Sometimes it is more efficient to generate separate stores for F and I,
8546/// which can remove the bitwise instructions or sink them to colder places.
8547///
8548/// (store (or (zext (bitcast F to i32) to i64),
8549/// (shl (zext I to i64), 32)), addr) -->
8550/// (store F, addr) and (store I, addr+4)
8551///
8552/// Similarly, splitting for other merged store can also be beneficial, like:
8553/// For pair of {i32, i32}, i64 store --> two i32 stores.
8554/// For pair of {i32, i16}, i64 store --> two i32 stores.
8555/// For pair of {i16, i16}, i32 store --> two i16 stores.
8556/// For pair of {i16, i8}, i32 store --> two i16 stores.
8557/// For pair of {i8, i8}, i16 store --> two i8 stores.
8558///
8559/// We allow each target to determine specifically which kind of splitting is
8560/// supported.
8561///
8562/// The store patterns are commonly seen from the simple code snippet below
8563/// if only std::make_pair(...) is sroa transformed before inlined into hoo.
8564/// void goo(const std::pair<int, float> &);
8565/// hoo() {
8566/// ...
8567/// goo(std::make_pair(tmp, ftmp));
8568/// ...
8569/// }
8570///
8571/// Although we already have similar splitting in DAG Combine, we duplicate
8572/// it in CodeGenPrepare to catch the case in which pattern is across
8573/// multiple BBs. The logic in DAG Combine is kept to catch case generated
8574/// during code expansion.
8575static bool splitMergedValStore(StoreInst &SI, const DataLayout &DL,
8576 const TargetLowering &TLI) {
8577 // Handle simple but common cases only.
8578 Type *StoreType = SI.getValueOperand()->getType();
8579
8580 // The code below assumes shifting a value by <number of bits>,
8581 // whereas scalable vectors would have to be shifted by
8582 // <2log(vscale) + number of bits> in order to store the
8583 // low/high parts. Bailing out for now.
8584 if (StoreType->isScalableTy())
8585 return false;
8586
8587 if (!DL.typeSizeEqualsStoreSize(Ty: StoreType) ||
8588 DL.getTypeSizeInBits(Ty: StoreType) == 0)
8589 return false;
8590
8591 unsigned HalfValBitSize = DL.getTypeSizeInBits(Ty: StoreType) / 2;
8592 Type *SplitStoreType = Type::getIntNTy(C&: SI.getContext(), N: HalfValBitSize);
8593 if (!DL.typeSizeEqualsStoreSize(Ty: SplitStoreType))
8594 return false;
8595
8596 // Don't split the store if it is volatile.
8597 if (SI.isVolatile())
8598 return false;
8599
8600 // Match the following patterns:
8601 // (store (or (zext LValue to i64),
8602 // (shl (zext HValue to i64), 32)), HalfValBitSize)
8603 // or
8604 // (store (or (shl (zext HValue to i64), 32)), HalfValBitSize)
8605 // (zext LValue to i64),
8606 // Expect both operands of OR and the first operand of SHL have only
8607 // one use.
8608 Value *LValue, *HValue;
8609 if (!match(V: SI.getValueOperand(),
8610 P: m_c_Or(L: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: LValue))),
8611 R: m_OneUse(SubPattern: m_Shl(L: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: HValue))),
8612 R: m_SpecificInt(V: HalfValBitSize))))))
8613 return false;
8614
8615 // Check LValue and HValue are int with size less or equal than 32.
8616 if (!LValue->getType()->isIntegerTy() ||
8617 DL.getTypeSizeInBits(Ty: LValue->getType()) > HalfValBitSize ||
8618 !HValue->getType()->isIntegerTy() ||
8619 DL.getTypeSizeInBits(Ty: HValue->getType()) > HalfValBitSize)
8620 return false;
8621
8622 // If LValue/HValue is a bitcast instruction, use the EVT before bitcast
8623 // as the input of target query.
8624 auto *LBC = dyn_cast<BitCastInst>(Val: LValue);
8625 auto *HBC = dyn_cast<BitCastInst>(Val: HValue);
8626 EVT LowTy = LBC ? EVT::getEVT(Ty: LBC->getOperand(i_nocapture: 0)->getType())
8627 : EVT::getEVT(Ty: LValue->getType());
8628 EVT HighTy = HBC ? EVT::getEVT(Ty: HBC->getOperand(i_nocapture: 0)->getType())
8629 : EVT::getEVT(Ty: HValue->getType());
8630 if (!ForceSplitStore && !TLI.isMultiStoresCheaperThanBitsMerge(LTy: LowTy, HTy: HighTy))
8631 return false;
8632
8633 // Start to split store.
8634 IRBuilder<> Builder(SI.getContext());
8635 Builder.SetInsertPoint(&SI);
8636
8637 // If LValue/HValue is a bitcast in another BB, create a new one in current
8638 // BB so it may be merged with the splitted stores by dag combiner.
8639 if (LBC && LBC->getParent() != SI.getParent())
8640 LValue = Builder.CreateBitCast(V: LBC->getOperand(i_nocapture: 0), DestTy: LBC->getType());
8641 if (HBC && HBC->getParent() != SI.getParent())
8642 HValue = Builder.CreateBitCast(V: HBC->getOperand(i_nocapture: 0), DestTy: HBC->getType());
8643
8644 bool IsLE = SI.getDataLayout().isLittleEndian();
8645 auto CreateSplitStore = [&](Value *V, bool Upper) {
8646 V = Builder.CreateZExtOrBitCast(V, DestTy: SplitStoreType);
8647 Value *Addr = SI.getPointerOperand();
8648 Align Alignment = SI.getAlign();
8649 const bool IsOffsetStore = (IsLE && Upper) || (!IsLE && !Upper);
8650 if (IsOffsetStore) {
8651 Addr = Builder.CreateGEP(
8652 Ty: SplitStoreType, Ptr: Addr,
8653 IdxList: ConstantInt::get(Ty: Type::getInt32Ty(C&: SI.getContext()), V: 1));
8654
8655 // When splitting the store in half, naturally one half will retain the
8656 // alignment of the original wider store, regardless of whether it was
8657 // over-aligned or not, while the other will require adjustment.
8658 Alignment = commonAlignment(A: Alignment, Offset: HalfValBitSize / 8);
8659 }
8660 Builder.CreateAlignedStore(Val: V, Ptr: Addr, Align: Alignment);
8661 };
8662
8663 CreateSplitStore(LValue, false);
8664 CreateSplitStore(HValue, true);
8665
8666 // Delete the old store.
8667 SI.eraseFromParent();
8668 return true;
8669}
8670
8671// Return true if the GEP has two operands, the first operand is of a sequential
8672// type, and the second operand is a constant.
8673static bool GEPSequentialConstIndexed(GetElementPtrInst *GEP) {
8674 gep_type_iterator I = gep_type_begin(GEP: *GEP);
8675 return GEP->getNumOperands() == 2 && I.isSequential() &&
8676 isa<ConstantInt>(Val: GEP->getOperand(i_nocapture: 1));
8677}
8678
8679// Try unmerging GEPs to reduce liveness interference (register pressure) across
8680// IndirectBr edges. Since IndirectBr edges tend to touch on many blocks,
8681// reducing liveness interference across those edges benefits global register
8682// allocation. Currently handles only certain cases.
8683//
8684// For example, unmerge %GEPI and %UGEPI as below.
8685//
8686// ---------- BEFORE ----------
8687// SrcBlock:
8688// ...
8689// %GEPIOp = ...
8690// ...
8691// %GEPI = gep %GEPIOp, Idx
8692// ...
8693// indirectbr ... [ label %DstB0, label %DstB1, ... label %DstBi ... ]
8694// (* %GEPI is alive on the indirectbr edges due to other uses ahead)
8695// (* %GEPIOp is alive on the indirectbr edges only because of it's used by
8696// %UGEPI)
8697//
8698// DstB0: ... (there may be a gep similar to %UGEPI to be unmerged)
8699// DstB1: ... (there may be a gep similar to %UGEPI to be unmerged)
8700// ...
8701//
8702// DstBi:
8703// ...
8704// %UGEPI = gep %GEPIOp, UIdx
8705// ...
8706// ---------------------------
8707//
8708// ---------- AFTER ----------
8709// SrcBlock:
8710// ... (same as above)
8711// (* %GEPI is still alive on the indirectbr edges)
8712// (* %GEPIOp is no longer alive on the indirectbr edges as a result of the
8713// unmerging)
8714// ...
8715//
8716// DstBi:
8717// ...
8718// %UGEPI = gep %GEPI, (UIdx-Idx)
8719// ...
8720// ---------------------------
8721//
8722// The register pressure on the IndirectBr edges is reduced because %GEPIOp is
8723// no longer alive on them.
8724//
8725// We try to unmerge GEPs here in CodGenPrepare, as opposed to limiting merging
8726// of GEPs in the first place in InstCombiner::visitGetElementPtrInst() so as
8727// not to disable further simplications and optimizations as a result of GEP
8728// merging.
8729//
8730// Note this unmerging may increase the length of the data flow critical path
8731// (the path from %GEPIOp to %UGEPI would go through %GEPI), which is a tradeoff
8732// between the register pressure and the length of data-flow critical
8733// path. Restricting this to the uncommon IndirectBr case would minimize the
8734// impact of potentially longer critical path, if any, and the impact on compile
8735// time.
8736static bool tryUnmergingGEPsAcrossIndirectBr(GetElementPtrInst *GEPI,
8737 const TargetTransformInfo *TTI) {
8738 BasicBlock *SrcBlock = GEPI->getParent();
8739 // Check that SrcBlock ends with an IndirectBr. If not, give up. The common
8740 // (non-IndirectBr) cases exit early here.
8741 if (!isa<IndirectBrInst>(Val: SrcBlock->getTerminator()))
8742 return false;
8743 // Check that GEPI is a simple gep with a single constant index.
8744 if (!GEPSequentialConstIndexed(GEP: GEPI))
8745 return false;
8746 ConstantInt *GEPIIdx = cast<ConstantInt>(Val: GEPI->getOperand(i_nocapture: 1));
8747 // Check that GEPI is a cheap one.
8748 if (TTI->getIntImmCost(Imm: GEPIIdx->getValue(), Ty: GEPIIdx->getType(),
8749 CostKind: TargetTransformInfo::TCK_SizeAndLatency) >
8750 TargetTransformInfo::TCC_Basic)
8751 return false;
8752 Value *GEPIOp = GEPI->getOperand(i_nocapture: 0);
8753 // Check that GEPIOp is an instruction that's also defined in SrcBlock.
8754 if (!isa<Instruction>(Val: GEPIOp))
8755 return false;
8756 auto *GEPIOpI = cast<Instruction>(Val: GEPIOp);
8757 if (GEPIOpI->getParent() != SrcBlock)
8758 return false;
8759 // Check that GEP is used outside the block, meaning it's alive on the
8760 // IndirectBr edge(s).
8761 if (llvm::none_of(Range: GEPI->users(), P: [&](User *Usr) {
8762 if (auto *I = dyn_cast<Instruction>(Val: Usr)) {
8763 if (I->getParent() != SrcBlock) {
8764 return true;
8765 }
8766 }
8767 return false;
8768 }))
8769 return false;
8770 // The second elements of the GEP chains to be unmerged.
8771 std::vector<GetElementPtrInst *> UGEPIs;
8772 // Check each user of GEPIOp to check if unmerging would make GEPIOp not alive
8773 // on IndirectBr edges.
8774 for (User *Usr : GEPIOp->users()) {
8775 if (Usr == GEPI)
8776 continue;
8777 // Check if Usr is an Instruction. If not, give up.
8778 if (!isa<Instruction>(Val: Usr))
8779 return false;
8780 auto *UI = cast<Instruction>(Val: Usr);
8781 // Check if Usr in the same block as GEPIOp, which is fine, skip.
8782 if (UI->getParent() == SrcBlock)
8783 continue;
8784 // Check if Usr is a GEP. If not, give up.
8785 if (!isa<GetElementPtrInst>(Val: Usr))
8786 return false;
8787 auto *UGEPI = cast<GetElementPtrInst>(Val: Usr);
8788 // Check if UGEPI is a simple gep with a single constant index and GEPIOp is
8789 // the pointer operand to it. If so, record it in the vector. If not, give
8790 // up.
8791 if (!GEPSequentialConstIndexed(GEP: UGEPI))
8792 return false;
8793 if (UGEPI->getOperand(i_nocapture: 0) != GEPIOp)
8794 return false;
8795 if (UGEPI->getSourceElementType() != GEPI->getSourceElementType())
8796 return false;
8797 if (GEPIIdx->getType() !=
8798 cast<ConstantInt>(Val: UGEPI->getOperand(i_nocapture: 1))->getType())
8799 return false;
8800 ConstantInt *UGEPIIdx = cast<ConstantInt>(Val: UGEPI->getOperand(i_nocapture: 1));
8801 if (TTI->getIntImmCost(Imm: UGEPIIdx->getValue(), Ty: UGEPIIdx->getType(),
8802 CostKind: TargetTransformInfo::TCK_SizeAndLatency) >
8803 TargetTransformInfo::TCC_Basic)
8804 return false;
8805 UGEPIs.push_back(x: UGEPI);
8806 }
8807 if (UGEPIs.size() == 0)
8808 return false;
8809 // Check the materializing cost of (Uidx-Idx).
8810 for (GetElementPtrInst *UGEPI : UGEPIs) {
8811 ConstantInt *UGEPIIdx = cast<ConstantInt>(Val: UGEPI->getOperand(i_nocapture: 1));
8812 APInt NewIdx = UGEPIIdx->getValue() - GEPIIdx->getValue();
8813 InstructionCost ImmCost = TTI->getIntImmCost(
8814 Imm: NewIdx, Ty: GEPIIdx->getType(), CostKind: TargetTransformInfo::TCK_SizeAndLatency);
8815 if (ImmCost > TargetTransformInfo::TCC_Basic)
8816 return false;
8817 }
8818 // Now unmerge between GEPI and UGEPIs.
8819 for (GetElementPtrInst *UGEPI : UGEPIs) {
8820 UGEPI->setOperand(i_nocapture: 0, Val_nocapture: GEPI);
8821 ConstantInt *UGEPIIdx = cast<ConstantInt>(Val: UGEPI->getOperand(i_nocapture: 1));
8822 Constant *NewUGEPIIdx = ConstantInt::get(
8823 Ty: GEPIIdx->getType(), V: UGEPIIdx->getValue() - GEPIIdx->getValue());
8824 UGEPI->setOperand(i_nocapture: 1, Val_nocapture: NewUGEPIIdx);
8825 // If GEPI is not inbounds but UGEPI is inbounds, change UGEPI to not
8826 // inbounds to avoid UB.
8827 if (!GEPI->isInBounds()) {
8828 UGEPI->setIsInBounds(false);
8829 }
8830 }
8831 // After unmerging, verify that GEPIOp is actually only used in SrcBlock (not
8832 // alive on IndirectBr edges).
8833 assert(llvm::none_of(GEPIOp->users(),
8834 [&](User *Usr) {
8835 return cast<Instruction>(Usr)->getParent() != SrcBlock;
8836 }) &&
8837 "GEPIOp is used outside SrcBlock");
8838 return true;
8839}
8840
8841static bool optimizeBranch(CondBrInst *Branch, const TargetLowering &TLI,
8842 SmallPtrSet<BasicBlock *, 32> &FreshBBs,
8843 bool IsHugeFunc) {
8844 // Try and convert
8845 // %c = icmp ult %x, 8
8846 // br %c, bla, blb
8847 // %tc = lshr %x, 3
8848 // to
8849 // %tc = lshr %x, 3
8850 // %c = icmp eq %tc, 0
8851 // br %c, bla, blb
8852 // Creating the cmp to zero can be better for the backend, especially if the
8853 // lshr produces flags that can be used automatically.
8854 if (!TLI.preferZeroCompareBranch())
8855 return false;
8856
8857 ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: Branch->getCondition());
8858 if (!Cmp || !isa<ConstantInt>(Val: Cmp->getOperand(i_nocapture: 1)) || !Cmp->hasOneUse())
8859 return false;
8860
8861 Value *X = Cmp->getOperand(i_nocapture: 0);
8862 if (!X->hasUseList())
8863 return false;
8864
8865 APInt CmpC = cast<ConstantInt>(Val: Cmp->getOperand(i_nocapture: 1))->getValue();
8866
8867 for (auto *U : X->users()) {
8868 Instruction *UI = dyn_cast<Instruction>(Val: U);
8869 // A quick dominance check
8870 if (!UI ||
8871 (UI->getParent() != Branch->getParent() &&
8872 UI->getParent() != Branch->getSuccessor(i: 0) &&
8873 UI->getParent() != Branch->getSuccessor(i: 1)) ||
8874 (UI->getParent() != Branch->getParent() &&
8875 !UI->getParent()->getSinglePredecessor()))
8876 continue;
8877
8878 if (CmpC.isPowerOf2() && Cmp->getPredicate() == ICmpInst::ICMP_ULT &&
8879 match(V: UI, P: m_Shr(L: m_Specific(V: X), R: m_SpecificInt(V: CmpC.logBase2())))) {
8880 IRBuilder<> Builder(Branch);
8881 if (UI->getParent() != Branch->getParent())
8882 UI->moveBefore(InsertPos: Branch->getIterator());
8883 UI->dropPoisonGeneratingFlags();
8884 Value *NewCmp = Builder.CreateCmp(Pred: ICmpInst::ICMP_EQ, LHS: UI,
8885 RHS: ConstantInt::get(Ty: UI->getType(), V: 0));
8886 LLVM_DEBUG(dbgs() << "Converting " << *Cmp << "\n");
8887 LLVM_DEBUG(dbgs() << " to compare on zero: " << *NewCmp << "\n");
8888 replaceAllUsesWith(Old: Cmp, New: NewCmp, FreshBBs, IsHuge: IsHugeFunc);
8889 return true;
8890 }
8891 if (Cmp->isEquality() &&
8892 (match(V: UI, P: m_Add(L: m_Specific(V: X), R: m_SpecificInt(V: -CmpC))) ||
8893 match(V: UI, P: m_Sub(L: m_Specific(V: X), R: m_SpecificInt(V: CmpC))) ||
8894 match(V: UI, P: m_Xor(L: m_Specific(V: X), R: m_SpecificInt(V: CmpC))))) {
8895 IRBuilder<> Builder(Branch);
8896 if (UI->getParent() != Branch->getParent())
8897 UI->moveBefore(InsertPos: Branch->getIterator());
8898 UI->dropPoisonGeneratingFlags();
8899 Value *NewCmp = Builder.CreateCmp(Pred: Cmp->getPredicate(), LHS: UI,
8900 RHS: ConstantInt::get(Ty: UI->getType(), V: 0));
8901 LLVM_DEBUG(dbgs() << "Converting " << *Cmp << "\n");
8902 LLVM_DEBUG(dbgs() << " to compare on zero: " << *NewCmp << "\n");
8903 replaceAllUsesWith(Old: Cmp, New: NewCmp, FreshBBs, IsHuge: IsHugeFunc);
8904 return true;
8905 }
8906 }
8907 return false;
8908}
8909
8910bool CodeGenPrepare::optimizeInst(Instruction *I, ModifyDT &ModifiedDT) {
8911 bool AnyChange = false;
8912 AnyChange = fixupDbgVariableRecordsOnInst(I&: *I);
8913
8914 // Bail out if we inserted the instruction to prevent optimizations from
8915 // stepping on each other's toes.
8916 if (InsertedInsts.count(Ptr: I))
8917 return AnyChange;
8918
8919 // TODO: Move into the switch on opcode below here.
8920 if (PHINode *P = dyn_cast<PHINode>(Val: I)) {
8921 // It is possible for very late stage optimizations (such as SimplifyCFG)
8922 // to introduce PHI nodes too late to be cleaned up. If we detect such a
8923 // trivial PHI, go ahead and zap it here.
8924 if (Value *V = simplifyInstruction(I: P, Q: {*DL, TLInfo})) {
8925 LargeOffsetGEPMap.erase(Key: P);
8926 replaceAllUsesWith(Old: P, New: V, FreshBBs, IsHuge: IsHugeFunc);
8927 P->eraseFromParent();
8928 ++NumPHIsElim;
8929 return true;
8930 }
8931 return AnyChange;
8932 }
8933
8934 if (CastInst *CI = dyn_cast<CastInst>(Val: I)) {
8935 // If the source of the cast is a constant, then this should have
8936 // already been constant folded. The only reason NOT to constant fold
8937 // it is if something (e.g. LSR) was careful to place the constant
8938 // evaluation in a block other than then one that uses it (e.g. to hoist
8939 // the address of globals out of a loop). If this is the case, we don't
8940 // want to forward-subst the cast.
8941 if (isa<Constant>(Val: CI->getOperand(i_nocapture: 0)))
8942 return AnyChange;
8943
8944 if (OptimizeNoopCopyExpression(CI, TLI: *TLI, DL: *DL))
8945 return true;
8946
8947 if ((isa<UIToFPInst>(Val: I) || isa<SIToFPInst>(Val: I) || isa<FPToUIInst>(Val: I) ||
8948 isa<TruncInst>(Val: I)) &&
8949 TLI->optimizeExtendOrTruncateConversion(
8950 I, L: LI->getLoopFor(BB: I->getParent()), TTI: *TTI))
8951 return true;
8952
8953 if (isa<ZExtInst>(Val: I) || isa<SExtInst>(Val: I)) {
8954 /// Sink a zext or sext into its user blocks if the target type doesn't
8955 /// fit in one register
8956 if (TLI->getTypeAction(Context&: CI->getContext(),
8957 VT: TLI->getValueType(DL: *DL, Ty: CI->getType())) ==
8958 TargetLowering::TypeExpandInteger) {
8959 return SinkCast(CI);
8960 } else {
8961 if (TLI->optimizeExtendOrTruncateConversion(
8962 I, L: LI->getLoopFor(BB: I->getParent()), TTI: *TTI))
8963 return true;
8964
8965 bool MadeChange = optimizeExt(Inst&: I);
8966 return MadeChange | optimizeExtUses(I);
8967 }
8968 }
8969 return AnyChange;
8970 }
8971
8972 if (auto *Cmp = dyn_cast<CmpInst>(Val: I))
8973 if (optimizeCmp(Cmp, ModifiedDT))
8974 return true;
8975
8976 if (match(V: I, P: m_URem(L: m_Value(), R: m_Value())))
8977 if (optimizeURem(Rem: I))
8978 return true;
8979
8980 if (LoadInst *LI = dyn_cast<LoadInst>(Val: I)) {
8981 LI->setMetadata(KindID: LLVMContext::MD_invariant_group, Node: nullptr);
8982 bool Modified = optimizeLoadExt(Load: LI);
8983 unsigned AS = LI->getPointerAddressSpace();
8984 Modified |= optimizeMemoryInst(MemoryInst: I, Addr: I->getOperand(i: 0), AccessTy: LI->getType(), AddrSpace: AS);
8985 return Modified;
8986 }
8987
8988 if (StoreInst *SI = dyn_cast<StoreInst>(Val: I)) {
8989 if (splitMergedValStore(SI&: *SI, DL: *DL, TLI: *TLI))
8990 return true;
8991 SI->setMetadata(KindID: LLVMContext::MD_invariant_group, Node: nullptr);
8992 unsigned AS = SI->getPointerAddressSpace();
8993 return optimizeMemoryInst(MemoryInst: I, Addr: SI->getOperand(i_nocapture: 1),
8994 AccessTy: SI->getOperand(i_nocapture: 0)->getType(), AddrSpace: AS);
8995 }
8996
8997 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: I)) {
8998 unsigned AS = RMW->getPointerAddressSpace();
8999 return optimizeMemoryInst(MemoryInst: I, Addr: RMW->getPointerOperand(), AccessTy: RMW->getType(), AddrSpace: AS);
9000 }
9001
9002 if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Val: I)) {
9003 unsigned AS = CmpX->getPointerAddressSpace();
9004 return optimizeMemoryInst(MemoryInst: I, Addr: CmpX->getPointerOperand(),
9005 AccessTy: CmpX->getCompareOperand()->getType(), AddrSpace: AS);
9006 }
9007
9008 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Val: I);
9009
9010 if (BinOp && BinOp->getOpcode() == Instruction::And && EnableAndCmpSinking &&
9011 sinkAndCmp0Expression(AndI: BinOp, TLI: *TLI, InsertedInsts))
9012 return true;
9013
9014 // TODO: Move this into the switch on opcode - it handles shifts already.
9015 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
9016 BinOp->getOpcode() == Instruction::LShr)) {
9017 ConstantInt *CI = dyn_cast<ConstantInt>(Val: BinOp->getOperand(i_nocapture: 1));
9018 if (CI && TLI->hasExtractBitsInsn())
9019 if (OptimizeExtractBits(ShiftI: BinOp, CI, TLI: *TLI, DL: *DL))
9020 return true;
9021 }
9022
9023 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Val: I)) {
9024 if (GEPI->hasAllZeroIndices()) {
9025 /// The GEP operand must be a pointer, so must its result -> BitCast
9026 Instruction *NC = new BitCastInst(GEPI->getOperand(i_nocapture: 0), GEPI->getType(),
9027 GEPI->getName(), GEPI->getIterator());
9028 NC->setDebugLoc(GEPI->getDebugLoc());
9029 replaceAllUsesWith(Old: GEPI, New: NC, FreshBBs, IsHuge: IsHugeFunc);
9030 RecursivelyDeleteTriviallyDeadInstructions(
9031 V: GEPI, TLI: TLInfo, MSSAU: nullptr,
9032 AboutToDeleteCallback: [&](Value *V) { removeAllAssertingVHReferences(V); });
9033 ++NumGEPsElim;
9034 optimizeInst(I: NC, ModifiedDT);
9035 return true;
9036 }
9037 if (tryUnmergingGEPsAcrossIndirectBr(GEPI, TTI)) {
9038 return true;
9039 }
9040 }
9041
9042 if (FreezeInst *FI = dyn_cast<FreezeInst>(Val: I)) {
9043 // freeze(icmp a, const)) -> icmp (freeze a), const
9044 // This helps generate efficient conditional jumps.
9045 Instruction *CmpI = nullptr;
9046 if (ICmpInst *II = dyn_cast<ICmpInst>(Val: FI->getOperand(i_nocapture: 0)))
9047 CmpI = II;
9048 else if (FCmpInst *F = dyn_cast<FCmpInst>(Val: FI->getOperand(i_nocapture: 0)))
9049 CmpI = F->getFastMathFlags().none() ? F : nullptr;
9050
9051 if (CmpI && CmpI->hasOneUse()) {
9052 auto Op0 = CmpI->getOperand(i: 0), Op1 = CmpI->getOperand(i: 1);
9053 bool Const0 = isa<ConstantInt>(Val: Op0) || isa<ConstantFP>(Val: Op0) ||
9054 isa<ConstantPointerNull>(Val: Op0);
9055 bool Const1 = isa<ConstantInt>(Val: Op1) || isa<ConstantFP>(Val: Op1) ||
9056 isa<ConstantPointerNull>(Val: Op1);
9057 if (Const0 || Const1) {
9058 if (!Const0 || !Const1) {
9059 auto *F = new FreezeInst(Const0 ? Op1 : Op0, "", CmpI->getIterator());
9060 F->takeName(V: FI);
9061 CmpI->setOperand(i: Const0 ? 1 : 0, Val: F);
9062 }
9063 replaceAllUsesWith(Old: FI, New: CmpI, FreshBBs, IsHuge: IsHugeFunc);
9064 FI->eraseFromParent();
9065 return true;
9066 }
9067 }
9068 return AnyChange;
9069 }
9070
9071 if (tryToSinkFreeOperands(I))
9072 return true;
9073
9074 switch (I->getOpcode()) {
9075 case Instruction::Shl:
9076 case Instruction::LShr:
9077 case Instruction::AShr:
9078 return optimizeShiftInst(Shift: cast<BinaryOperator>(Val: I));
9079 case Instruction::Call:
9080 return optimizeCallInst(CI: cast<CallInst>(Val: I), ModifiedDT);
9081 case Instruction::Select:
9082 return optimizeSelectInst(SI: cast<SelectInst>(Val: I));
9083 case Instruction::ShuffleVector:
9084 return optimizeShuffleVectorInst(SVI: cast<ShuffleVectorInst>(Val: I));
9085 case Instruction::Switch:
9086 return optimizeSwitchInst(SI: cast<SwitchInst>(Val: I));
9087 case Instruction::ExtractElement:
9088 return optimizeExtractElementInst(Inst: cast<ExtractElementInst>(Val: I));
9089 case Instruction::CondBr:
9090 return optimizeBranch(Branch: cast<CondBrInst>(Val: I), TLI: *TLI, FreshBBs, IsHugeFunc);
9091 }
9092
9093 return AnyChange;
9094}
9095
9096/// Given an OR instruction, check to see if this is a bitreverse
9097/// idiom. If so, insert the new intrinsic and return true.
9098bool CodeGenPrepare::makeBitReverse(Instruction &I) {
9099 if (!I.getType()->isIntegerTy() ||
9100 !TLI->isOperationLegalOrCustom(Op: ISD::BITREVERSE,
9101 VT: TLI->getValueType(DL: *DL, Ty: I.getType(), AllowUnknown: true)))
9102 return false;
9103
9104 SmallVector<Instruction *, 4> Insts;
9105 if (!recognizeBSwapOrBitReverseIdiom(I: &I, MatchBSwaps: false, MatchBitReversals: true, InsertedInsts&: Insts))
9106 return false;
9107 Instruction *LastInst = Insts.back();
9108 replaceAllUsesWith(Old: &I, New: LastInst, FreshBBs, IsHuge: IsHugeFunc);
9109 RecursivelyDeleteTriviallyDeadInstructions(
9110 V: &I, TLI: TLInfo, MSSAU: nullptr,
9111 AboutToDeleteCallback: [&](Value *V) { removeAllAssertingVHReferences(V); });
9112 return true;
9113}
9114
9115// In this pass we look for GEP and cast instructions that are used
9116// across basic blocks and rewrite them to improve basic-block-at-a-time
9117// selection.
9118bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, ModifyDT &ModifiedDT) {
9119 SunkAddrs.clear();
9120 bool MadeChange = false;
9121
9122 do {
9123 CurInstIterator = BB.begin();
9124 ModifiedDT = ModifyDT::NotModifyDT;
9125 while (CurInstIterator != BB.end()) {
9126 MadeChange |= optimizeInst(I: &*CurInstIterator++, ModifiedDT);
9127 if (ModifiedDT != ModifyDT::NotModifyDT) {
9128 // For huge function we tend to quickly go though the inner optmization
9129 // opportunities in the BB. So we go back to the BB head to re-optimize
9130 // each instruction instead of go back to the function head.
9131 if (IsHugeFunc)
9132 break;
9133 return true;
9134 }
9135 }
9136 } while (ModifiedDT == ModifyDT::ModifyInstDT);
9137
9138 bool MadeBitReverse = true;
9139 while (MadeBitReverse) {
9140 MadeBitReverse = false;
9141 for (auto &I : reverse(C&: BB)) {
9142 if (makeBitReverse(I)) {
9143 MadeBitReverse = MadeChange = true;
9144 break;
9145 }
9146 }
9147 }
9148 MadeChange |= dupRetToEnableTailCallOpts(BB: &BB, ModifiedDT);
9149
9150 return MadeChange;
9151}
9152
9153bool CodeGenPrepare::fixupDbgVariableRecordsOnInst(Instruction &I) {
9154 bool AnyChange = false;
9155 for (DbgVariableRecord &DVR : filterDbgVars(R: I.getDbgRecordRange()))
9156 AnyChange |= fixupDbgVariableRecord(I&: DVR);
9157 return AnyChange;
9158}
9159
9160// FIXME: should updating debug-info really cause the "changed" flag to fire,
9161// which can cause a function to be reprocessed?
9162bool CodeGenPrepare::fixupDbgVariableRecord(DbgVariableRecord &DVR) {
9163 if (DVR.Type != DbgVariableRecord::LocationType::Value &&
9164 DVR.Type != DbgVariableRecord::LocationType::Assign)
9165 return false;
9166
9167 // Does this DbgVariableRecord refer to a sunk address calculation?
9168 bool AnyChange = false;
9169 SmallDenseSet<Value *> LocationOps(DVR.location_ops().begin(),
9170 DVR.location_ops().end());
9171 for (Value *Location : LocationOps) {
9172 WeakTrackingVH SunkAddrVH = SunkAddrs[Location];
9173 Value *SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
9174 if (SunkAddr) {
9175 // Point dbg.value at locally computed address, which should give the best
9176 // opportunity to be accurately lowered. This update may change the type
9177 // of pointer being referred to; however this makes no difference to
9178 // debugging information, and we can't generate bitcasts that may affect
9179 // codegen.
9180 DVR.replaceVariableLocationOp(OldValue: Location, NewValue: SunkAddr);
9181 AnyChange = true;
9182 }
9183 }
9184 return AnyChange;
9185}
9186
9187static void DbgInserterHelper(DbgVariableRecord *DVR, BasicBlock::iterator VI) {
9188 DVR->removeFromParent();
9189 BasicBlock *VIBB = VI->getParent();
9190 if (isa<PHINode>(Val: VI))
9191 VIBB->insertDbgRecordBefore(DR: DVR, Here: VIBB->getFirstInsertionPt());
9192 else
9193 VIBB->insertDbgRecordAfter(DR: DVR, I: &*VI);
9194}
9195
9196// A llvm.dbg.value may be using a value before its definition, due to
9197// optimizations in this pass and others. Scan for such dbg.values, and rescue
9198// them by moving the dbg.value to immediately after the value definition.
9199// FIXME: Ideally this should never be necessary, and this has the potential
9200// to re-order dbg.value intrinsics.
9201bool CodeGenPrepare::placeDbgValues(Function &F) {
9202 bool MadeChange = false;
9203 DominatorTree &DT = getDT();
9204
9205 auto DbgProcessor = [&](auto *DbgItem, Instruction *Position) {
9206 SmallVector<Instruction *, 4> VIs;
9207 for (Value *V : DbgItem->location_ops())
9208 if (Instruction *VI = dyn_cast_or_null<Instruction>(Val: V))
9209 VIs.push_back(Elt: VI);
9210
9211 // This item may depend on multiple instructions, complicating any
9212 // potential sink. This block takes the defensive approach, opting to
9213 // "undef" the item if it has more than one instruction and any of them do
9214 // not dominate iem.
9215 for (Instruction *VI : VIs) {
9216 if (VI->isTerminator())
9217 continue;
9218
9219 // If VI is a phi in a block with an EHPad terminator, we can't insert
9220 // after it.
9221 if (isa<PHINode>(Val: VI) && VI->getParent()->getTerminator()->isEHPad())
9222 continue;
9223
9224 // If the defining instruction dominates the dbg.value, we do not need
9225 // to move the dbg.value.
9226 if (DT.dominates(Def: VI, User: Position))
9227 continue;
9228
9229 // If we depend on multiple instructions and any of them doesn't
9230 // dominate this DVI, we probably can't salvage it: moving it to
9231 // after any of the instructions could cause us to lose the others.
9232 if (VIs.size() > 1) {
9233 LLVM_DEBUG(
9234 dbgs()
9235 << "Unable to find valid location for Debug Value, undefing:\n"
9236 << *DbgItem);
9237 DbgItem->setKillLocation();
9238 break;
9239 }
9240
9241 LLVM_DEBUG(dbgs() << "Moving Debug Value before :\n"
9242 << *DbgItem << ' ' << *VI);
9243 DbgInserterHelper(DbgItem, VI->getIterator());
9244 MadeChange = true;
9245 ++NumDbgValueMoved;
9246 }
9247 };
9248
9249 for (BasicBlock &BB : F) {
9250 for (Instruction &Insn : llvm::make_early_inc_range(Range&: BB)) {
9251 // Process any DbgVariableRecord records attached to this
9252 // instruction.
9253 for (DbgVariableRecord &DVR : llvm::make_early_inc_range(
9254 Range: filterDbgVars(R: Insn.getDbgRecordRange()))) {
9255 if (DVR.Type != DbgVariableRecord::LocationType::Value)
9256 continue;
9257 DbgProcessor(&DVR, &Insn);
9258 }
9259 }
9260 }
9261
9262 return MadeChange;
9263}
9264
9265// Group scattered pseudo probes in a block to favor SelectionDAG. Scattered
9266// probes can be chained dependencies of other regular DAG nodes and block DAG
9267// combine optimizations.
9268bool CodeGenPrepare::placePseudoProbes(Function &F) {
9269 bool MadeChange = false;
9270 for (auto &Block : F) {
9271 // Move the rest probes to the beginning of the block.
9272 auto FirstInst = Block.getFirstInsertionPt();
9273 while (FirstInst != Block.end() && FirstInst->isDebugOrPseudoInst())
9274 ++FirstInst;
9275 BasicBlock::iterator I(FirstInst);
9276 I++;
9277 while (I != Block.end()) {
9278 if (auto *II = dyn_cast<PseudoProbeInst>(Val: I++)) {
9279 II->moveBefore(InsertPos: FirstInst);
9280 MadeChange = true;
9281 }
9282 }
9283 }
9284 return MadeChange;
9285}
9286
9287/// Scale down both weights to fit into uint32_t.
9288static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
9289 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
9290 uint32_t Scale = (NewMax / std::numeric_limits<uint32_t>::max()) + 1;
9291 NewTrue = NewTrue / Scale;
9292 NewFalse = NewFalse / Scale;
9293}
9294
9295/// Some targets prefer to split a conditional branch like:
9296/// \code
9297/// %0 = icmp ne i32 %a, 0
9298/// %1 = icmp ne i32 %b, 0
9299/// %or.cond = or i1 %0, %1
9300/// br i1 %or.cond, label %TrueBB, label %FalseBB
9301/// \endcode
9302/// into multiple branch instructions like:
9303/// \code
9304/// bb1:
9305/// %0 = icmp ne i32 %a, 0
9306/// br i1 %0, label %TrueBB, label %bb2
9307/// bb2:
9308/// %1 = icmp ne i32 %b, 0
9309/// br i1 %1, label %TrueBB, label %FalseBB
9310/// \endcode
9311/// This usually allows instruction selection to do even further optimizations
9312/// and combine the compare with the branch instruction. Currently this is
9313/// applied for targets which have "cheap" jump instructions.
9314///
9315/// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
9316///
9317bool CodeGenPrepare::splitBranchCondition(Function &F) {
9318 if (!TM->Options.EnableFastISel || TLI->isJumpExpensive())
9319 return false;
9320
9321 bool MadeChange = false;
9322 for (auto &BB : F) {
9323 // Does this BB end with the following?
9324 // %cond1 = icmp|fcmp|binary instruction ...
9325 // %cond2 = icmp|fcmp|binary instruction ...
9326 // %cond.or = or|and i1 %cond1, cond2
9327 // br i1 %cond.or label %dest1, label %dest2"
9328 Instruction *LogicOp;
9329 BasicBlock *TBB, *FBB;
9330 if (!match(V: BB.getTerminator(),
9331 P: m_Br(C: m_OneUse(SubPattern: m_Instruction(I&: LogicOp)), T&: TBB, F&: FBB)))
9332 continue;
9333
9334 auto *Br1 = cast<CondBrInst>(Val: BB.getTerminator());
9335 if (Br1->getMetadata(KindID: LLVMContext::MD_unpredictable))
9336 continue;
9337
9338 // The merging of mostly empty BB can cause a degenerate branch.
9339 if (TBB == FBB)
9340 continue;
9341
9342 unsigned Opc;
9343 Value *Cond1, *Cond2;
9344 if (match(V: LogicOp,
9345 P: m_LogicalAnd(L: m_OneUse(SubPattern: m_Value(V&: Cond1)), R: m_OneUse(SubPattern: m_Value(V&: Cond2)))))
9346 Opc = Instruction::And;
9347 else if (match(V: LogicOp, P: m_LogicalOr(L: m_OneUse(SubPattern: m_Value(V&: Cond1)),
9348 R: m_OneUse(SubPattern: m_Value(V&: Cond2)))))
9349 Opc = Instruction::Or;
9350 else
9351 continue;
9352
9353 auto IsGoodCond = [](Value *Cond) {
9354 return match(
9355 V: Cond,
9356 P: m_CombineOr(L: m_Cmp(), R: m_CombineOr(L: m_LogicalAnd(L: m_Value(), R: m_Value()),
9357 R: m_LogicalOr(L: m_Value(), R: m_Value()))));
9358 };
9359 if (!IsGoodCond(Cond1) || !IsGoodCond(Cond2))
9360 continue;
9361
9362 LLVM_DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
9363
9364 // Create a new BB.
9365 auto *TmpBB =
9366 BasicBlock::Create(Context&: BB.getContext(), Name: BB.getName() + ".cond.split",
9367 Parent: BB.getParent(), InsertBefore: BB.getNextNode());
9368 if (IsHugeFunc)
9369 FreshBBs.insert(Ptr: TmpBB);
9370
9371 // Update original basic block by using the first condition directly by the
9372 // branch instruction and removing the no longer needed and/or instruction.
9373 Br1->setCondition(Cond1);
9374 LogicOp->eraseFromParent();
9375
9376 // Depending on the condition we have to either replace the true or the
9377 // false successor of the original branch instruction.
9378 if (Opc == Instruction::And)
9379 Br1->setSuccessor(idx: 0, NewSucc: TmpBB);
9380 else
9381 Br1->setSuccessor(idx: 1, NewSucc: TmpBB);
9382
9383 // Fill in the new basic block.
9384 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond: Cond2, True: TBB, False: FBB);
9385 if (auto *I = dyn_cast<Instruction>(Val: Cond2)) {
9386 I->removeFromParent();
9387 I->insertBefore(InsertPos: Br2->getIterator());
9388 }
9389
9390 // Update PHI nodes in both successors. The original BB needs to be
9391 // replaced in one successor's PHI nodes, because the branch comes now from
9392 // the newly generated BB (NewBB). In the other successor we need to add one
9393 // incoming edge to the PHI nodes, because both branch instructions target
9394 // now the same successor. Depending on the original branch condition
9395 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
9396 // we perform the correct update for the PHI nodes.
9397 // This doesn't change the successor order of the just created branch
9398 // instruction (or any other instruction).
9399 if (Opc == Instruction::Or)
9400 std::swap(a&: TBB, b&: FBB);
9401
9402 // Replace the old BB with the new BB.
9403 TBB->replacePhiUsesWith(Old: &BB, New: TmpBB);
9404
9405 // Add another incoming edge from the new BB.
9406 for (PHINode &PN : FBB->phis()) {
9407 auto *Val = PN.getIncomingValueForBlock(BB: &BB);
9408 PN.addIncoming(V: Val, BB: TmpBB);
9409 }
9410
9411 if (Loop *L = LI->getLoopFor(BB: &BB))
9412 L->addBasicBlockToLoop(NewBB: TmpBB, LI&: *LI);
9413
9414 // The edge we need to delete starts at BB and ends at whatever TBB ends
9415 // up pointing to.
9416 DTU->applyUpdates(Updates: {{DominatorTree::Insert, &BB, TmpBB},
9417 {DominatorTree::Insert, TmpBB, TBB},
9418 {DominatorTree::Insert, TmpBB, FBB},
9419 {DominatorTree::Delete, &BB, TBB}});
9420
9421 // Update the branch weights (from SelectionDAGBuilder::
9422 // FindMergedConditions).
9423 if (Opc == Instruction::Or) {
9424 // Codegen X | Y as:
9425 // BB1:
9426 // jmp_if_X TBB
9427 // jmp TmpBB
9428 // TmpBB:
9429 // jmp_if_Y TBB
9430 // jmp FBB
9431 //
9432
9433 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
9434 // The requirement is that
9435 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
9436 // = TrueProb for original BB.
9437 // Assuming the original weights are A and B, one choice is to set BB1's
9438 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
9439 // assumes that
9440 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
9441 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
9442 // TmpBB, but the math is more complicated.
9443 uint64_t TrueWeight, FalseWeight;
9444 if (extractBranchWeights(I: *Br1, TrueVal&: TrueWeight, FalseVal&: FalseWeight)) {
9445 uint64_t NewTrueWeight = TrueWeight;
9446 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
9447 scaleWeights(NewTrue&: NewTrueWeight, NewFalse&: NewFalseWeight);
9448 Br1->setMetadata(KindID: LLVMContext::MD_prof,
9449 Node: MDBuilder(Br1->getContext())
9450 .createBranchWeights(TrueWeight, FalseWeight,
9451 IsExpected: hasBranchWeightOrigin(I: *Br1)));
9452
9453 NewTrueWeight = TrueWeight;
9454 NewFalseWeight = 2 * FalseWeight;
9455 scaleWeights(NewTrue&: NewTrueWeight, NewFalse&: NewFalseWeight);
9456 Br2->setMetadata(KindID: LLVMContext::MD_prof,
9457 Node: MDBuilder(Br2->getContext())
9458 .createBranchWeights(TrueWeight, FalseWeight));
9459 }
9460 } else {
9461 // Codegen X & Y as:
9462 // BB1:
9463 // jmp_if_X TmpBB
9464 // jmp FBB
9465 // TmpBB:
9466 // jmp_if_Y TBB
9467 // jmp FBB
9468 //
9469 // This requires creation of TmpBB after CurBB.
9470
9471 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
9472 // The requirement is that
9473 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
9474 // = FalseProb for original BB.
9475 // Assuming the original weights are A and B, one choice is to set BB1's
9476 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
9477 // assumes that
9478 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
9479 uint64_t TrueWeight, FalseWeight;
9480 if (extractBranchWeights(I: *Br1, TrueVal&: TrueWeight, FalseVal&: FalseWeight)) {
9481 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
9482 uint64_t NewFalseWeight = FalseWeight;
9483 scaleWeights(NewTrue&: NewTrueWeight, NewFalse&: NewFalseWeight);
9484 Br1->setMetadata(KindID: LLVMContext::MD_prof,
9485 Node: MDBuilder(Br1->getContext())
9486 .createBranchWeights(TrueWeight, FalseWeight));
9487
9488 NewTrueWeight = 2 * TrueWeight;
9489 NewFalseWeight = FalseWeight;
9490 scaleWeights(NewTrue&: NewTrueWeight, NewFalse&: NewFalseWeight);
9491 Br2->setMetadata(KindID: LLVMContext::MD_prof,
9492 Node: MDBuilder(Br2->getContext())
9493 .createBranchWeights(TrueWeight, FalseWeight));
9494 }
9495 }
9496
9497 MadeChange = true;
9498
9499 LLVM_DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
9500 TmpBB->dump());
9501 }
9502 return MadeChange;
9503}
9504