1//===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This pass munges the code in the input function to better prepare it for
10// SelectionDAG-based code generation. This works around limitations in it's
11// basic-block-at-a-time approach. It should eventually be removed.
12//
13//===----------------------------------------------------------------------===//
14
15#include "llvm/CodeGen/CodeGenPrepare.h"
16#include "llvm/ADT/APInt.h"
17#include "llvm/ADT/ArrayRef.h"
18#include "llvm/ADT/DenseMap.h"
19#include "llvm/ADT/MapVector.h"
20#include "llvm/ADT/PointerIntPair.h"
21#include "llvm/ADT/STLExtras.h"
22#include "llvm/ADT/SmallPtrSet.h"
23#include "llvm/ADT/SmallVector.h"
24#include "llvm/ADT/Statistic.h"
25#include "llvm/Analysis/BlockFrequencyInfo.h"
26#include "llvm/Analysis/BranchProbabilityInfo.h"
27#include "llvm/Analysis/FloatingPointPredicateUtils.h"
28#include "llvm/Analysis/InstructionSimplify.h"
29#include "llvm/Analysis/LoopInfo.h"
30#include "llvm/Analysis/ProfileSummaryInfo.h"
31#include "llvm/Analysis/ScalarEvolutionExpressions.h"
32#include "llvm/Analysis/TargetLibraryInfo.h"
33#include "llvm/Analysis/TargetTransformInfo.h"
34#include "llvm/Analysis/ValueTracking.h"
35#include "llvm/Analysis/VectorUtils.h"
36#include "llvm/CodeGen/Analysis.h"
37#include "llvm/CodeGen/BasicBlockSectionsProfileReader.h"
38#include "llvm/CodeGen/ISDOpcodes.h"
39#include "llvm/CodeGen/SelectionDAGNodes.h"
40#include "llvm/CodeGen/TargetLowering.h"
41#include "llvm/CodeGen/TargetPassConfig.h"
42#include "llvm/CodeGen/TargetSubtargetInfo.h"
43#include "llvm/CodeGen/ValueTypes.h"
44#include "llvm/CodeGenTypes/MachineValueType.h"
45#include "llvm/Config/llvm-config.h"
46#include "llvm/IR/Argument.h"
47#include "llvm/IR/Attributes.h"
48#include "llvm/IR/BasicBlock.h"
49#include "llvm/IR/Constant.h"
50#include "llvm/IR/Constants.h"
51#include "llvm/IR/DataLayout.h"
52#include "llvm/IR/DebugInfo.h"
53#include "llvm/IR/DerivedTypes.h"
54#include "llvm/IR/Dominators.h"
55#include "llvm/IR/Function.h"
56#include "llvm/IR/GetElementPtrTypeIterator.h"
57#include "llvm/IR/GlobalValue.h"
58#include "llvm/IR/GlobalVariable.h"
59#include "llvm/IR/IRBuilder.h"
60#include "llvm/IR/InlineAsm.h"
61#include "llvm/IR/InstrTypes.h"
62#include "llvm/IR/Instruction.h"
63#include "llvm/IR/Instructions.h"
64#include "llvm/IR/IntrinsicInst.h"
65#include "llvm/IR/Intrinsics.h"
66#include "llvm/IR/IntrinsicsAArch64.h"
67#include "llvm/IR/LLVMContext.h"
68#include "llvm/IR/MDBuilder.h"
69#include "llvm/IR/Module.h"
70#include "llvm/IR/Operator.h"
71#include "llvm/IR/PatternMatch.h"
72#include "llvm/IR/ProfDataUtils.h"
73#include "llvm/IR/Statepoint.h"
74#include "llvm/IR/Type.h"
75#include "llvm/IR/Use.h"
76#include "llvm/IR/User.h"
77#include "llvm/IR/Value.h"
78#include "llvm/IR/ValueHandle.h"
79#include "llvm/IR/ValueMap.h"
80#include "llvm/InitializePasses.h"
81#include "llvm/Pass.h"
82#include "llvm/Support/BlockFrequency.h"
83#include "llvm/Support/BranchProbability.h"
84#include "llvm/Support/Casting.h"
85#include "llvm/Support/CommandLine.h"
86#include "llvm/Support/Compiler.h"
87#include "llvm/Support/Debug.h"
88#include "llvm/Support/ErrorHandling.h"
89#include "llvm/Support/raw_ostream.h"
90#include "llvm/Target/TargetMachine.h"
91#include "llvm/Target/TargetOptions.h"
92#include "llvm/Transforms/Utils/BasicBlockUtils.h"
93#include "llvm/Transforms/Utils/BypassSlowDivision.h"
94#include "llvm/Transforms/Utils/Local.h"
95#include "llvm/Transforms/Utils/SimplifyLibCalls.h"
96#include "llvm/Transforms/Utils/SizeOpts.h"
97#include <algorithm>
98#include <cassert>
99#include <cstdint>
100#include <iterator>
101#include <limits>
102#include <memory>
103#include <optional>
104#include <utility>
105#include <vector>
106
107using namespace llvm;
108using namespace llvm::PatternMatch;
109
110#define DEBUG_TYPE "codegenprepare"
111
112STATISTIC(NumBlocksElim, "Number of blocks eliminated");
113STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
114STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
115STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
116 "sunken Cmps");
117STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
118 "of sunken Casts");
119STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
120 "computations were sunk");
121STATISTIC(NumMemoryInstsPhiCreated,
122 "Number of phis created when address "
123 "computations were sunk to memory instructions");
124STATISTIC(NumMemoryInstsSelectCreated,
125 "Number of select created when address "
126 "computations were sunk to memory instructions");
127STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
128STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
129STATISTIC(NumAndsAdded,
130 "Number of and mask instructions added to form ext loads");
131STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
132STATISTIC(NumRetsDup, "Number of return instructions duplicated");
133STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
134STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
135STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
136
137static cl::opt<bool> DisableBranchOpts(
138 "disable-cgp-branch-opts", cl::Hidden, cl::init(Val: false),
139 cl::desc("Disable branch optimizations in CodeGenPrepare"));
140
141static cl::opt<bool>
142 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(Val: false),
143 cl::desc("Disable GC optimizations in CodeGenPrepare"));
144
145static cl::opt<bool>
146 DisableSelectToBranch("disable-cgp-select2branch", cl::Hidden,
147 cl::init(Val: false),
148 cl::desc("Disable select to branch conversion."));
149
150static cl::opt<bool>
151 AddrSinkUsingGEPs("addr-sink-using-gep", cl::Hidden, cl::init(Val: true),
152 cl::desc("Address sinking in CGP using GEPs."));
153
154static cl::opt<bool>
155 EnableAndCmpSinking("enable-andcmp-sinking", cl::Hidden, cl::init(Val: true),
156 cl::desc("Enable sinking and/cmp into branches."));
157
158static cl::opt<bool> DisableStoreExtract(
159 "disable-cgp-store-extract", cl::Hidden, cl::init(Val: false),
160 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
161
162static cl::opt<bool> StressStoreExtract(
163 "stress-cgp-store-extract", cl::Hidden, cl::init(Val: false),
164 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
165
166static cl::opt<bool> DisableExtLdPromotion(
167 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(Val: false),
168 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
169 "CodeGenPrepare"));
170
171static cl::opt<bool> StressExtLdPromotion(
172 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(Val: false),
173 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
174 "optimization in CodeGenPrepare"));
175
176static cl::opt<bool> DisablePreheaderProtect(
177 "disable-preheader-prot", cl::Hidden, cl::init(Val: false),
178 cl::desc("Disable protection against removing loop preheaders"));
179
180static cl::opt<bool> ProfileGuidedSectionPrefix(
181 "profile-guided-section-prefix", cl::Hidden, cl::init(Val: true),
182 cl::desc("Use profile info to add section prefix for hot/cold functions"));
183
184static cl::opt<bool> ProfileUnknownInSpecialSection(
185 "profile-unknown-in-special-section", cl::Hidden,
186 cl::desc("In profiling mode like sampleFDO, if a function doesn't have "
187 "profile, we cannot tell the function is cold for sure because "
188 "it may be a function newly added without ever being sampled. "
189 "With the flag enabled, compiler can put such profile unknown "
190 "functions into a special section, so runtime system can choose "
191 "to handle it in a different way than .text section, to save "
192 "RAM for example. "));
193
194static cl::opt<bool> BBSectionsGuidedSectionPrefix(
195 "bbsections-guided-section-prefix", cl::Hidden, cl::init(Val: true),
196 cl::desc("Use the basic-block-sections profile to determine the text "
197 "section prefix for hot functions. Functions with "
198 "basic-block-sections profile will be placed in `.text.hot` "
199 "regardless of their FDO profile info. Other functions won't be "
200 "impacted, i.e., their prefixes will be decided by FDO/sampleFDO "
201 "profiles."));
202
203static cl::opt<uint64_t> FreqRatioToSkipMerge(
204 "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(Val: 2),
205 cl::desc("Skip merging empty blocks if (frequency of empty block) / "
206 "(frequency of destination block) is greater than this ratio"));
207
208static cl::opt<bool> ForceSplitStore(
209 "force-split-store", cl::Hidden, cl::init(Val: false),
210 cl::desc("Force store splitting no matter what the target query says."));
211
212static cl::opt<bool> EnableTypePromotionMerge(
213 "cgp-type-promotion-merge", cl::Hidden,
214 cl::desc("Enable merging of redundant sexts when one is dominating"
215 " the other."),
216 cl::init(Val: true));
217
218static cl::opt<bool> DisableComplexAddrModes(
219 "disable-complex-addr-modes", cl::Hidden, cl::init(Val: false),
220 cl::desc("Disables combining addressing modes with different parts "
221 "in optimizeMemoryInst."));
222
223static cl::opt<bool>
224 AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(Val: false),
225 cl::desc("Allow creation of Phis in Address sinking."));
226
227static cl::opt<bool> AddrSinkNewSelects(
228 "addr-sink-new-select", cl::Hidden, cl::init(Val: true),
229 cl::desc("Allow creation of selects in Address sinking."));
230
231static cl::opt<bool> AddrSinkCombineBaseReg(
232 "addr-sink-combine-base-reg", cl::Hidden, cl::init(Val: true),
233 cl::desc("Allow combining of BaseReg field in Address sinking."));
234
235static cl::opt<bool> AddrSinkCombineBaseGV(
236 "addr-sink-combine-base-gv", cl::Hidden, cl::init(Val: true),
237 cl::desc("Allow combining of BaseGV field in Address sinking."));
238
239static cl::opt<bool> AddrSinkCombineBaseOffs(
240 "addr-sink-combine-base-offs", cl::Hidden, cl::init(Val: true),
241 cl::desc("Allow combining of BaseOffs field in Address sinking."));
242
243static cl::opt<bool> AddrSinkCombineScaledReg(
244 "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(Val: true),
245 cl::desc("Allow combining of ScaledReg field in Address sinking."));
246
247static cl::opt<bool>
248 EnableGEPOffsetSplit("cgp-split-large-offset-gep", cl::Hidden,
249 cl::init(Val: true),
250 cl::desc("Enable splitting large offset of GEP."));
251
252static cl::opt<bool> EnableICMP_EQToICMP_ST(
253 "cgp-icmp-eq2icmp-st", cl::Hidden, cl::init(Val: false),
254 cl::desc("Enable ICMP_EQ to ICMP_S(L|G)T conversion."));
255
256static cl::opt<bool>
257 VerifyBFIUpdates("cgp-verify-bfi-updates", cl::Hidden, cl::init(Val: false),
258 cl::desc("Enable BFI update verification for "
259 "CodeGenPrepare."));
260
261static cl::opt<bool>
262 OptimizePhiTypes("cgp-optimize-phi-types", cl::Hidden, cl::init(Val: true),
263 cl::desc("Enable converting phi types in CodeGenPrepare"));
264
265static cl::opt<unsigned>
266 HugeFuncThresholdInCGPP("cgpp-huge-func", cl::init(Val: 10000), cl::Hidden,
267 cl::desc("Least BB number of huge function."));
268
269static cl::opt<unsigned>
270 MaxAddressUsersToScan("cgp-max-address-users-to-scan", cl::init(Val: 100),
271 cl::Hidden,
272 cl::desc("Max number of address users to look at"));
273
274static cl::opt<bool>
275 DisableDeletePHIs("disable-cgp-delete-phis", cl::Hidden, cl::init(Val: false),
276 cl::desc("Disable elimination of dead PHI nodes."));
277
278namespace {
279
280enum ExtType {
281 ZeroExtension, // Zero extension has been seen.
282 SignExtension, // Sign extension has been seen.
283 BothExtension // This extension type is used if we saw sext after
284 // ZeroExtension had been set, or if we saw zext after
285 // SignExtension had been set. It makes the type
286 // information of a promoted instruction invalid.
287};
288
289enum ModifyDT {
290 NotModifyDT, // Not Modify any DT.
291 ModifyBBDT, // Modify the Basic Block Dominator Tree.
292 ModifyInstDT // Modify the Instruction Dominator in a Basic Block,
293 // This usually means we move/delete/insert instruction
294 // in a Basic Block. So we should re-iterate instructions
295 // in such Basic Block.
296};
297
298using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
299using TypeIsSExt = PointerIntPair<Type *, 2, ExtType>;
300using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
301using SExts = SmallVector<Instruction *, 16>;
302using ValueToSExts = MapVector<Value *, SExts>;
303
304class TypePromotionTransaction;
305
306class CodeGenPrepare {
307 friend class CodeGenPrepareLegacyPass;
308 const TargetMachine *TM = nullptr;
309 const TargetSubtargetInfo *SubtargetInfo = nullptr;
310 const TargetLowering *TLI = nullptr;
311 const TargetRegisterInfo *TRI = nullptr;
312 const TargetTransformInfo *TTI = nullptr;
313 const BasicBlockSectionsProfileReader *BBSectionsProfileReader = nullptr;
314 const TargetLibraryInfo *TLInfo = nullptr;
315 LoopInfo *LI = nullptr;
316 BlockFrequencyInfo *BFI;
317 BranchProbabilityInfo *BPI;
318 ProfileSummaryInfo *PSI = nullptr;
319
320 /// As we scan instructions optimizing them, this is the next instruction
321 /// to optimize. Transforms that can invalidate this should update it.
322 BasicBlock::iterator CurInstIterator;
323
324 /// Keeps track of non-local addresses that have been sunk into a block.
325 /// This allows us to avoid inserting duplicate code for blocks with
326 /// multiple load/stores of the same address. The usage of WeakTrackingVH
327 /// enables SunkAddrs to be treated as a cache whose entries can be
328 /// invalidated if a sunken address computation has been erased.
329 ValueMap<Value *, WeakTrackingVH> SunkAddrs;
330
331 /// Keeps track of all instructions inserted for the current function.
332 SetOfInstrs InsertedInsts;
333
334 /// Keeps track of the type of the related instruction before their
335 /// promotion for the current function.
336 InstrToOrigTy PromotedInsts;
337
338 /// Keep track of instructions removed during promotion.
339 SetOfInstrs RemovedInsts;
340
341 /// Keep track of sext chains based on their initial value.
342 DenseMap<Value *, Instruction *> SeenChainsForSExt;
343
344 /// Keep track of GEPs accessing the same data structures such as structs or
345 /// arrays that are candidates to be split later because of their large
346 /// size.
347 MapVector<AssertingVH<Value>,
348 SmallVector<std::pair<AssertingVH<GetElementPtrInst>, int64_t>, 32>>
349 LargeOffsetGEPMap;
350
351 /// Keep track of new GEP base after splitting the GEPs having large offset.
352 SmallSet<AssertingVH<Value>, 2> NewGEPBases;
353
354 /// Map serial numbers to Large offset GEPs.
355 DenseMap<AssertingVH<GetElementPtrInst>, int> LargeOffsetGEPID;
356
357 /// Keep track of SExt promoted.
358 ValueToSExts ValToSExtendedUses;
359
360 /// True if the function has the OptSize attribute.
361 bool OptSize;
362
363 /// DataLayout for the Function being processed.
364 const DataLayout *DL = nullptr;
365
366 /// Building the dominator tree can be expensive, so we only build it
367 /// lazily and update it when required.
368 std::unique_ptr<DominatorTree> DT;
369
370public:
371 CodeGenPrepare() = default;
372 CodeGenPrepare(const TargetMachine *TM) : TM(TM){};
373 /// If encounter huge function, we need to limit the build time.
374 bool IsHugeFunc = false;
375
376 /// FreshBBs is like worklist, it collected the updated BBs which need
377 /// to be optimized again.
378 /// Note: Consider building time in this pass, when a BB updated, we need
379 /// to insert such BB into FreshBBs for huge function.
380 SmallPtrSet<BasicBlock *, 32> FreshBBs;
381
382 void releaseMemory() {
383 // Clear per function information.
384 InsertedInsts.clear();
385 PromotedInsts.clear();
386 FreshBBs.clear();
387 }
388
389 bool run(Function &F, FunctionAnalysisManager &AM);
390
391private:
392 template <typename F>
393 void resetIteratorIfInvalidatedWhileCalling(BasicBlock *BB, F f) {
394 // Substituting can cause recursive simplifications, which can invalidate
395 // our iterator. Use a WeakTrackingVH to hold onto it in case this
396 // happens.
397 Value *CurValue = &*CurInstIterator;
398 WeakTrackingVH IterHandle(CurValue);
399
400 f();
401
402 // If the iterator instruction was recursively deleted, start over at the
403 // start of the block.
404 if (IterHandle != CurValue) {
405 CurInstIterator = BB->begin();
406 SunkAddrs.clear();
407 }
408 }
409
410 // Get the DominatorTree, building if necessary.
411 DominatorTree &getDT(Function &F) {
412 if (!DT)
413 DT = std::make_unique<DominatorTree>(args&: F);
414 return *DT;
415 }
416
417 void removeAllAssertingVHReferences(Value *V);
418 bool eliminateAssumptions(Function &F);
419 bool eliminateFallThrough(Function &F, DominatorTree *DT = nullptr);
420 bool eliminateMostlyEmptyBlocks(Function &F);
421 BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
422 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
423 void eliminateMostlyEmptyBlock(BasicBlock *BB);
424 bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
425 bool isPreheader);
426 bool makeBitReverse(Instruction &I);
427 bool optimizeBlock(BasicBlock &BB, ModifyDT &ModifiedDT);
428 bool optimizeInst(Instruction *I, ModifyDT &ModifiedDT);
429 bool optimizeMemoryInst(Instruction *MemoryInst, Value *Addr, Type *AccessTy,
430 unsigned AddrSpace);
431 bool optimizeGatherScatterInst(Instruction *MemoryInst, Value *Ptr);
432 bool optimizeMulWithOverflow(Instruction *I, bool IsSigned,
433 ModifyDT &ModifiedDT);
434 bool optimizeInlineAsmInst(CallInst *CS);
435 bool optimizeCallInst(CallInst *CI, ModifyDT &ModifiedDT);
436 bool optimizeExt(Instruction *&I);
437 bool optimizeExtUses(Instruction *I);
438 bool optimizeLoadExt(LoadInst *Load);
439 bool optimizeShiftInst(BinaryOperator *BO);
440 bool optimizeFunnelShift(IntrinsicInst *Fsh);
441 bool optimizeSelectInst(SelectInst *SI);
442 bool optimizeShuffleVectorInst(ShuffleVectorInst *SVI);
443 bool optimizeSwitchType(SwitchInst *SI);
444 bool optimizeSwitchPhiConstants(SwitchInst *SI);
445 bool optimizeSwitchInst(SwitchInst *SI);
446 bool optimizeExtractElementInst(Instruction *Inst);
447 bool dupRetToEnableTailCallOpts(BasicBlock *BB, ModifyDT &ModifiedDT);
448 bool fixupDbgVariableRecord(DbgVariableRecord &I);
449 bool fixupDbgVariableRecordsOnInst(Instruction &I);
450 bool placeDbgValues(Function &F);
451 bool placePseudoProbes(Function &F);
452 bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
453 LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
454 bool tryToPromoteExts(TypePromotionTransaction &TPT,
455 const SmallVectorImpl<Instruction *> &Exts,
456 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
457 unsigned CreatedInstsCost = 0);
458 bool mergeSExts(Function &F);
459 bool splitLargeGEPOffsets();
460 bool optimizePhiType(PHINode *Inst, SmallPtrSetImpl<PHINode *> &Visited,
461 SmallPtrSetImpl<Instruction *> &DeletedInstrs);
462 bool optimizePhiTypes(Function &F);
463 bool performAddressTypePromotion(
464 Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
465 bool HasPromoted, TypePromotionTransaction &TPT,
466 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
467 bool splitBranchCondition(Function &F, ModifyDT &ModifiedDT);
468 bool simplifyOffsetableRelocate(GCStatepointInst &I);
469
470 bool tryToSinkFreeOperands(Instruction *I);
471 bool replaceMathCmpWithIntrinsic(BinaryOperator *BO, Value *Arg0, Value *Arg1,
472 CmpInst *Cmp, Intrinsic::ID IID);
473 bool optimizeCmp(CmpInst *Cmp, ModifyDT &ModifiedDT);
474 bool optimizeURem(Instruction *Rem);
475 bool combineToUSubWithOverflow(CmpInst *Cmp, ModifyDT &ModifiedDT);
476 bool combineToUAddWithOverflow(CmpInst *Cmp, ModifyDT &ModifiedDT);
477 bool unfoldPowerOf2Test(CmpInst *Cmp);
478 void verifyBFIUpdates(Function &F);
479 bool _run(Function &F);
480};
481
482class CodeGenPrepareLegacyPass : public FunctionPass {
483public:
484 static char ID; // Pass identification, replacement for typeid
485
486 CodeGenPrepareLegacyPass() : FunctionPass(ID) {}
487
488 bool runOnFunction(Function &F) override;
489
490 StringRef getPassName() const override { return "CodeGen Prepare"; }
491
492 void getAnalysisUsage(AnalysisUsage &AU) const override {
493 // FIXME: When we can selectively preserve passes, preserve the domtree.
494 AU.addRequired<ProfileSummaryInfoWrapperPass>();
495 AU.addRequired<TargetLibraryInfoWrapperPass>();
496 AU.addRequired<TargetPassConfig>();
497 AU.addRequired<TargetTransformInfoWrapperPass>();
498 AU.addRequired<LoopInfoWrapperPass>();
499 AU.addRequired<BranchProbabilityInfoWrapperPass>();
500 AU.addRequired<BlockFrequencyInfoWrapperPass>();
501 AU.addUsedIfAvailable<BasicBlockSectionsProfileReaderWrapperPass>();
502 }
503};
504
505} // end anonymous namespace
506
507char CodeGenPrepareLegacyPass::ID = 0;
508
509bool CodeGenPrepareLegacyPass::runOnFunction(Function &F) {
510 if (skipFunction(F))
511 return false;
512 auto TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>();
513 CodeGenPrepare CGP(TM);
514 CGP.DL = &F.getDataLayout();
515 CGP.SubtargetInfo = TM->getSubtargetImpl(F);
516 CGP.TLI = CGP.SubtargetInfo->getTargetLowering();
517 CGP.TRI = CGP.SubtargetInfo->getRegisterInfo();
518 CGP.TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
519 CGP.TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
520 CGP.LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
521 CGP.BPI = &getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI();
522 CGP.BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
523 CGP.PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
524 auto BBSPRWP =
525 getAnalysisIfAvailable<BasicBlockSectionsProfileReaderWrapperPass>();
526 CGP.BBSectionsProfileReader = BBSPRWP ? &BBSPRWP->getBBSPR() : nullptr;
527
528 return CGP._run(F);
529}
530
531INITIALIZE_PASS_BEGIN(CodeGenPrepareLegacyPass, DEBUG_TYPE,
532 "Optimize for code generation", false, false)
533INITIALIZE_PASS_DEPENDENCY(BasicBlockSectionsProfileReaderWrapperPass)
534INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
535INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
536INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
537INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
538INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
539INITIALIZE_PASS_END(CodeGenPrepareLegacyPass, DEBUG_TYPE,
540 "Optimize for code generation", false, false)
541
542FunctionPass *llvm::createCodeGenPrepareLegacyPass() {
543 return new CodeGenPrepareLegacyPass();
544}
545
546PreservedAnalyses CodeGenPreparePass::run(Function &F,
547 FunctionAnalysisManager &AM) {
548 CodeGenPrepare CGP(TM);
549
550 bool Changed = CGP.run(F, AM);
551 if (!Changed)
552 return PreservedAnalyses::all();
553
554 PreservedAnalyses PA;
555 PA.preserve<TargetLibraryAnalysis>();
556 PA.preserve<TargetIRAnalysis>();
557 return PA;
558}
559
560bool CodeGenPrepare::run(Function &F, FunctionAnalysisManager &AM) {
561 DL = &F.getDataLayout();
562 SubtargetInfo = TM->getSubtargetImpl(F);
563 TLI = SubtargetInfo->getTargetLowering();
564 TRI = SubtargetInfo->getRegisterInfo();
565 TLInfo = &AM.getResult<TargetLibraryAnalysis>(IR&: F);
566 TTI = &AM.getResult<TargetIRAnalysis>(IR&: F);
567 LI = &AM.getResult<LoopAnalysis>(IR&: F);
568 BPI = &AM.getResult<BranchProbabilityAnalysis>(IR&: F);
569 BFI = &AM.getResult<BlockFrequencyAnalysis>(IR&: F);
570 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(IR&: F);
571 PSI = MAMProxy.getCachedResult<ProfileSummaryAnalysis>(IR&: *F.getParent());
572 BBSectionsProfileReader =
573 AM.getCachedResult<BasicBlockSectionsProfileReaderAnalysis>(IR&: F);
574 return _run(F);
575}
576
577bool CodeGenPrepare::_run(Function &F) {
578 bool EverMadeChange = false;
579
580 OptSize = F.hasOptSize();
581 // Use the basic-block-sections profile to promote hot functions to .text.hot
582 // if requested.
583 if (BBSectionsGuidedSectionPrefix && BBSectionsProfileReader &&
584 BBSectionsProfileReader->isFunctionHot(FuncName: F.getName())) {
585 (void)F.setSectionPrefix("hot");
586 } else if (ProfileGuidedSectionPrefix) {
587 // The hot attribute overwrites profile count based hotness while profile
588 // counts based hotness overwrite the cold attribute.
589 // This is a conservative behabvior.
590 if (F.hasFnAttribute(Kind: Attribute::Hot) ||
591 PSI->isFunctionHotInCallGraph(F: &F, BFI&: *BFI))
592 (void)F.setSectionPrefix("hot");
593 // If PSI shows this function is not hot, we will placed the function
594 // into unlikely section if (1) PSI shows this is a cold function, or
595 // (2) the function has a attribute of cold.
596 else if (PSI->isFunctionColdInCallGraph(F: &F, BFI&: *BFI) ||
597 F.hasFnAttribute(Kind: Attribute::Cold))
598 (void)F.setSectionPrefix("unlikely");
599 else if (ProfileUnknownInSpecialSection && PSI->hasPartialSampleProfile() &&
600 PSI->isFunctionHotnessUnknown(F))
601 (void)F.setSectionPrefix("unknown");
602 }
603
604 /// This optimization identifies DIV instructions that can be
605 /// profitably bypassed and carried out with a shorter, faster divide.
606 if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI->isSlowDivBypassed()) {
607 const DenseMap<unsigned int, unsigned int> &BypassWidths =
608 TLI->getBypassSlowDivWidths();
609 BasicBlock *BB = &*F.begin();
610 while (BB != nullptr) {
611 // bypassSlowDivision may create new BBs, but we don't want to reapply the
612 // optimization to those blocks.
613 BasicBlock *Next = BB->getNextNode();
614 if (!llvm::shouldOptimizeForSize(BB, PSI, BFI))
615 EverMadeChange |= bypassSlowDivision(BB, BypassWidth: BypassWidths);
616 BB = Next;
617 }
618 }
619
620 // Get rid of @llvm.assume builtins before attempting to eliminate empty
621 // blocks, since there might be blocks that only contain @llvm.assume calls
622 // (plus arguments that we can get rid of).
623 EverMadeChange |= eliminateAssumptions(F);
624
625 // Eliminate blocks that contain only PHI nodes and an
626 // unconditional branch.
627 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
628
629 ModifyDT ModifiedDT = ModifyDT::NotModifyDT;
630 if (!DisableBranchOpts)
631 EverMadeChange |= splitBranchCondition(F, ModifiedDT);
632
633 // Split some critical edges where one of the sources is an indirect branch,
634 // to help generate sane code for PHIs involving such edges.
635 EverMadeChange |=
636 SplitIndirectBrCriticalEdges(F, /*IgnoreBlocksWithoutPHI=*/true);
637
638 // If we are optimzing huge function, we need to consider the build time.
639 // Because the basic algorithm's complex is near O(N!).
640 IsHugeFunc = F.size() > HugeFuncThresholdInCGPP;
641
642 // Transformations above may invalidate dominator tree and/or loop info.
643 DT.reset();
644 LI->releaseMemory();
645 LI->analyze(DomTree: getDT(F));
646
647 bool MadeChange = true;
648 bool FuncIterated = false;
649 while (MadeChange) {
650 MadeChange = false;
651
652 for (BasicBlock &BB : llvm::make_early_inc_range(Range&: F)) {
653 if (FuncIterated && !FreshBBs.contains(Ptr: &BB))
654 continue;
655
656 ModifyDT ModifiedDTOnIteration = ModifyDT::NotModifyDT;
657 bool Changed = optimizeBlock(BB, ModifiedDT&: ModifiedDTOnIteration);
658
659 if (ModifiedDTOnIteration == ModifyDT::ModifyBBDT)
660 DT.reset();
661
662 MadeChange |= Changed;
663 if (IsHugeFunc) {
664 // If the BB is updated, it may still has chance to be optimized.
665 // This usually happen at sink optimization.
666 // For example:
667 //
668 // bb0：
669 // %and = and i32 %a, 4
670 // %cmp = icmp eq i32 %and, 0
671 //
672 // If the %cmp sink to other BB, the %and will has chance to sink.
673 if (Changed)
674 FreshBBs.insert(Ptr: &BB);
675 else if (FuncIterated)
676 FreshBBs.erase(Ptr: &BB);
677 } else {
678 // For small/normal functions, we restart BB iteration if the dominator
679 // tree of the Function was changed.
680 if (ModifiedDTOnIteration != ModifyDT::NotModifyDT)
681 break;
682 }
683 }
684 // We have iterated all the BB in the (only work for huge) function.
685 FuncIterated = IsHugeFunc;
686
687 if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
688 MadeChange |= mergeSExts(F);
689 if (!LargeOffsetGEPMap.empty())
690 MadeChange |= splitLargeGEPOffsets();
691 MadeChange |= optimizePhiTypes(F);
692
693 if (MadeChange)
694 eliminateFallThrough(F, DT: DT.get());
695
696#ifndef NDEBUG
697 if (MadeChange && VerifyLoopInfo)
698 LI->verify(getDT(F));
699#endif
700
701 // Really free removed instructions during promotion.
702 for (Instruction *I : RemovedInsts)
703 I->deleteValue();
704
705 EverMadeChange |= MadeChange;
706 SeenChainsForSExt.clear();
707 ValToSExtendedUses.clear();
708 RemovedInsts.clear();
709 LargeOffsetGEPMap.clear();
710 LargeOffsetGEPID.clear();
711 }
712
713 NewGEPBases.clear();
714 SunkAddrs.clear();
715
716 if (!DisableBranchOpts) {
717 MadeChange = false;
718 // Use a set vector to get deterministic iteration order. The order the
719 // blocks are removed may affect whether or not PHI nodes in successors
720 // are removed.
721 SmallSetVector<BasicBlock *, 8> WorkList;
722 for (BasicBlock &BB : F) {
723 SmallVector<BasicBlock *, 2> Successors(successors(BB: &BB));
724 MadeChange |= ConstantFoldTerminator(BB: &BB, DeleteDeadConditions: true);
725 if (!MadeChange)
726 continue;
727
728 for (BasicBlock *Succ : Successors)
729 if (pred_empty(BB: Succ))
730 WorkList.insert(X: Succ);
731 }
732
733 // Delete the dead blocks and any of their dead successors.
734 MadeChange |= !WorkList.empty();
735 while (!WorkList.empty()) {
736 BasicBlock *BB = WorkList.pop_back_val();
737 SmallVector<BasicBlock *, 2> Successors(successors(BB));
738
739 DeleteDeadBlock(BB);
740
741 for (BasicBlock *Succ : Successors)
742 if (pred_empty(BB: Succ))
743 WorkList.insert(X: Succ);
744 }
745
746 // Merge pairs of basic blocks with unconditional branches, connected by
747 // a single edge.
748 if (EverMadeChange || MadeChange)
749 MadeChange |= eliminateFallThrough(F);
750
751 EverMadeChange |= MadeChange;
752 }
753
754 if (!DisableGCOpts) {
755 SmallVector<GCStatepointInst *, 2> Statepoints;
756 for (BasicBlock &BB : F)
757 for (Instruction &I : BB)
758 if (auto *SP = dyn_cast<GCStatepointInst>(Val: &I))
759 Statepoints.push_back(Elt: SP);
760 for (auto &I : Statepoints)
761 EverMadeChange |= simplifyOffsetableRelocate(I&: *I);
762 }
763
764 // Do this last to clean up use-before-def scenarios introduced by other
765 // preparatory transforms.
766 EverMadeChange |= placeDbgValues(F);
767 EverMadeChange |= placePseudoProbes(F);
768
769#ifndef NDEBUG
770 if (VerifyBFIUpdates)
771 verifyBFIUpdates(F);
772#endif
773
774 return EverMadeChange;
775}
776
777bool CodeGenPrepare::eliminateAssumptions(Function &F) {
778 bool MadeChange = false;
779 for (BasicBlock &BB : F) {
780 CurInstIterator = BB.begin();
781 while (CurInstIterator != BB.end()) {
782 Instruction *I = &*(CurInstIterator++);
783 if (auto *Assume = dyn_cast<AssumeInst>(Val: I)) {
784 MadeChange = true;
785 Value *Operand = Assume->getOperand(i_nocapture: 0);
786 Assume->eraseFromParent();
787
788 resetIteratorIfInvalidatedWhileCalling(BB: &BB, f: [&]() {
789 RecursivelyDeleteTriviallyDeadInstructions(V: Operand, TLI: TLInfo, MSSAU: nullptr);
790 });
791 }
792 }
793 }
794 return MadeChange;
795}
796
797/// An instruction is about to be deleted, so remove all references to it in our
798/// GEP-tracking data strcutures.
799void CodeGenPrepare::removeAllAssertingVHReferences(Value *V) {
800 LargeOffsetGEPMap.erase(Key: V);
801 NewGEPBases.erase(V);
802
803 auto GEP = dyn_cast<GetElementPtrInst>(Val: V);
804 if (!GEP)
805 return;
806
807 LargeOffsetGEPID.erase(Val: GEP);
808
809 auto VecI = LargeOffsetGEPMap.find(Key: GEP->getPointerOperand());
810 if (VecI == LargeOffsetGEPMap.end())
811 return;
812
813 auto &GEPVector = VecI->second;
814 llvm::erase_if(C&: GEPVector, P: [=](auto &Elt) { return Elt.first == GEP; });
815
816 if (GEPVector.empty())
817 LargeOffsetGEPMap.erase(Iterator: VecI);
818}
819
820// Verify BFI has been updated correctly by recomputing BFI and comparing them.
821[[maybe_unused]] void CodeGenPrepare::verifyBFIUpdates(Function &F) {
822 DominatorTree NewDT(F);
823 LoopInfo NewLI(NewDT);
824 BranchProbabilityInfo NewBPI(F, NewLI, TLInfo);
825 BlockFrequencyInfo NewBFI(F, NewBPI, NewLI);
826 NewBFI.verifyMatch(Other&: *BFI);
827}
828
829/// Merge basic blocks which are connected by a single edge, where one of the
830/// basic blocks has a single successor pointing to the other basic block,
831/// which has a single predecessor.
832bool CodeGenPrepare::eliminateFallThrough(Function &F, DominatorTree *DT) {
833 bool Changed = false;
834 // Scan all of the blocks in the function, except for the entry block.
835 // Use a temporary array to avoid iterator being invalidated when
836 // deleting blocks.
837 SmallVector<WeakTrackingVH, 16> Blocks(
838 llvm::make_pointer_range(Range: llvm::drop_begin(RangeOrContainer&: F)));
839
840 SmallSet<WeakTrackingVH, 16> Preds;
841 for (auto &Block : Blocks) {
842 auto *BB = cast_or_null<BasicBlock>(Val&: Block);
843 if (!BB)
844 continue;
845 // If the destination block has a single pred, then this is a trivial
846 // edge, just collapse it.
847 BasicBlock *SinglePred = BB->getSinglePredecessor();
848
849 // Don't merge if BB's address is taken.
850 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken())
851 continue;
852
853 // Make an effort to skip unreachable blocks.
854 if (DT && !DT->isReachableFromEntry(A: BB))
855 continue;
856
857 if (isa<UncondBrInst>(Val: SinglePred->getTerminator())) {
858 Changed = true;
859 LLVM_DEBUG(dbgs() << "To merge:\n" << *BB << "\n\n\n");
860
861 // Merge BB into SinglePred and delete it.
862 MergeBlockIntoPredecessor(BB, /* DTU */ nullptr, LI, /* MSSAU */ nullptr,
863 /* MemDep */ nullptr,
864 /* PredecessorWithTwoSuccessors */ false, DT);
865 Preds.insert(V: SinglePred);
866
867 if (IsHugeFunc) {
868 // Update FreshBBs to optimize the merged BB.
869 FreshBBs.insert(Ptr: SinglePred);
870 FreshBBs.erase(Ptr: BB);
871 }
872 }
873 }
874
875 // (Repeatedly) merging blocks into their predecessors can create redundant
876 // debug intrinsics.
877 for (const auto &Pred : Preds)
878 if (auto *BB = cast_or_null<BasicBlock>(Val: Pred))
879 RemoveRedundantDbgInstrs(BB);
880
881 return Changed;
882}
883
884/// Find a destination block from BB if BB is mergeable empty block.
885BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
886 // If this block doesn't end with an uncond branch, ignore it.
887 UncondBrInst *BI = dyn_cast<UncondBrInst>(Val: BB->getTerminator());
888 if (!BI)
889 return nullptr;
890
891 // If the instruction before the branch (skipping debug info) isn't a phi
892 // node, then other stuff is happening here.
893 BasicBlock::iterator BBI = BI->getIterator();
894 if (BBI != BB->begin()) {
895 --BBI;
896 if (!isa<PHINode>(Val: BBI))
897 return nullptr;
898 }
899
900 // Do not break infinite loops.
901 BasicBlock *DestBB = BI->getSuccessor();
902 if (DestBB == BB)
903 return nullptr;
904
905 if (!canMergeBlocks(BB, DestBB))
906 DestBB = nullptr;
907
908 return DestBB;
909}
910
911/// Eliminate blocks that contain only PHI nodes, debug info directives, and an
912/// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
913/// edges in ways that are non-optimal for isel. Start by eliminating these
914/// blocks so we can split them the way we want them.
915bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
916 SmallPtrSet<BasicBlock *, 16> Preheaders;
917 SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
918 while (!LoopList.empty()) {
919 Loop *L = LoopList.pop_back_val();
920 llvm::append_range(C&: LoopList, R&: *L);
921 if (BasicBlock *Preheader = L->getLoopPreheader())
922 Preheaders.insert(Ptr: Preheader);
923 }
924
925 bool MadeChange = false;
926 // Copy blocks into a temporary array to avoid iterator invalidation issues
927 // as we remove them.
928 // Note that this intentionally skips the entry block.
929 SmallVector<WeakTrackingVH, 16> Blocks;
930 for (auto &Block : llvm::drop_begin(RangeOrContainer&: F)) {
931 // Delete phi nodes that could block deleting other empty blocks.
932 if (!DisableDeletePHIs)
933 MadeChange |= DeleteDeadPHIs(BB: &Block, TLI: TLInfo);
934 Blocks.push_back(Elt: &Block);
935 }
936
937 for (auto &Block : Blocks) {
938 BasicBlock *BB = cast_or_null<BasicBlock>(Val&: Block);
939 if (!BB)
940 continue;
941 BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
942 if (!DestBB ||
943 !isMergingEmptyBlockProfitable(BB, DestBB, isPreheader: Preheaders.count(Ptr: BB)))
944 continue;
945
946 eliminateMostlyEmptyBlock(BB);
947 MadeChange = true;
948 }
949 return MadeChange;
950}
951
952bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
953 BasicBlock *DestBB,
954 bool isPreheader) {
955 // Do not delete loop preheaders if doing so would create a critical edge.
956 // Loop preheaders can be good locations to spill registers. If the
957 // preheader is deleted and we create a critical edge, registers may be
958 // spilled in the loop body instead.
959 if (!DisablePreheaderProtect && isPreheader &&
960 !(BB->getSinglePredecessor() &&
961 BB->getSinglePredecessor()->getSingleSuccessor()))
962 return false;
963
964 // Skip merging if the block's successor is also a successor to any callbr
965 // that leads to this block.
966 // FIXME: Is this really needed? Is this a correctness issue?
967 for (BasicBlock *Pred : predecessors(BB)) {
968 if (isa<CallBrInst>(Val: Pred->getTerminator()) &&
969 llvm::is_contained(Range: successors(BB: Pred), Element: DestBB))
970 return false;
971 }
972
973 // Try to skip merging if the unique predecessor of BB is terminated by a
974 // switch or indirect branch instruction, and BB is used as an incoming block
975 // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
976 // add COPY instructions in the predecessor of BB instead of BB (if it is not
977 // merged). Note that the critical edge created by merging such blocks wont be
978 // split in MachineSink because the jump table is not analyzable. By keeping
979 // such empty block (BB), ISel will place COPY instructions in BB, not in the
980 // predecessor of BB.
981 BasicBlock *Pred = BB->getUniquePredecessor();
982 if (!Pred || !(isa<SwitchInst>(Val: Pred->getTerminator()) ||
983 isa<IndirectBrInst>(Val: Pred->getTerminator())))
984 return true;
985
986 if (BB->getTerminator() != &*BB->getFirstNonPHIOrDbg())
987 return true;
988
989 // We use a simple cost heuristic which determine skipping merging is
990 // profitable if the cost of skipping merging is less than the cost of
991 // merging : Cost(skipping merging) < Cost(merging BB), where the
992 // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
993 // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
994 // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
995 // Freq(Pred) / Freq(BB) > 2.
996 // Note that if there are multiple empty blocks sharing the same incoming
997 // value for the PHIs in the DestBB, we consider them together. In such
998 // case, Cost(merging BB) will be the sum of their frequencies.
999
1000 if (!isa<PHINode>(Val: DestBB->begin()))
1001 return true;
1002
1003 SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
1004
1005 // Find all other incoming blocks from which incoming values of all PHIs in
1006 // DestBB are the same as the ones from BB.
1007 for (BasicBlock *DestBBPred : predecessors(BB: DestBB)) {
1008 if (DestBBPred == BB)
1009 continue;
1010
1011 if (llvm::all_of(Range: DestBB->phis(), P: [&](const PHINode &DestPN) {
1012 return DestPN.getIncomingValueForBlock(BB) ==
1013 DestPN.getIncomingValueForBlock(BB: DestBBPred);
1014 }))
1015 SameIncomingValueBBs.insert(Ptr: DestBBPred);
1016 }
1017
1018 // See if all BB's incoming values are same as the value from Pred. In this
1019 // case, no reason to skip merging because COPYs are expected to be place in
1020 // Pred already.
1021 if (SameIncomingValueBBs.count(Ptr: Pred))
1022 return true;
1023
1024 BlockFrequency PredFreq = BFI->getBlockFreq(BB: Pred);
1025 BlockFrequency BBFreq = BFI->getBlockFreq(BB);
1026
1027 for (auto *SameValueBB : SameIncomingValueBBs)
1028 if (SameValueBB->getUniquePredecessor() == Pred &&
1029 DestBB == findDestBlockOfMergeableEmptyBlock(BB: SameValueBB))
1030 BBFreq += BFI->getBlockFreq(BB: SameValueBB);
1031
1032 std::optional<BlockFrequency> Limit = BBFreq.mul(Factor: FreqRatioToSkipMerge);
1033 return !Limit || PredFreq <= *Limit;
1034}
1035
1036/// Return true if we can merge BB into DestBB if there is a single
1037/// unconditional branch between them, and BB contains no other non-phi
1038/// instructions.
1039bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
1040 const BasicBlock *DestBB) const {
1041 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
1042 // the successor. If there are more complex condition (e.g. preheaders),
1043 // don't mess around with them.
1044 for (const PHINode &PN : BB->phis()) {
1045 for (const User *U : PN.users()) {
1046 const Instruction *UI = cast<Instruction>(Val: U);
1047 if (UI->getParent() != DestBB || !isa<PHINode>(Val: UI))
1048 return false;
1049 // If User is inside DestBB block and it is a PHINode then check
1050 // incoming value. If incoming value is not from BB then this is
1051 // a complex condition (e.g. preheaders) we want to avoid here.
1052 if (UI->getParent() == DestBB) {
1053 if (const PHINode *UPN = dyn_cast<PHINode>(Val: UI))
1054 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
1055 Instruction *Insn = dyn_cast<Instruction>(Val: UPN->getIncomingValue(i: I));
1056 if (Insn && Insn->getParent() == BB &&
1057 Insn->getParent() != UPN->getIncomingBlock(i: I))
1058 return false;
1059 }
1060 }
1061 }
1062 }
1063
1064 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
1065 // and DestBB may have conflicting incoming values for the block. If so, we
1066 // can't merge the block.
1067 const PHINode *DestBBPN = dyn_cast<PHINode>(Val: DestBB->begin());
1068 if (!DestBBPN)
1069 return true; // no conflict.
1070
1071 // Collect the preds of BB.
1072 SmallPtrSet<const BasicBlock *, 16> BBPreds;
1073 if (const PHINode *BBPN = dyn_cast<PHINode>(Val: BB->begin())) {
1074 // It is faster to get preds from a PHI than with pred_iterator.
1075 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1076 BBPreds.insert(Ptr: BBPN->getIncomingBlock(i));
1077 } else {
1078 BBPreds.insert_range(R: predecessors(BB));
1079 }
1080
1081 // Walk the preds of DestBB.
1082 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
1083 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
1084 if (BBPreds.count(Ptr: Pred)) { // Common predecessor?
1085 for (const PHINode &PN : DestBB->phis()) {
1086 const Value *V1 = PN.getIncomingValueForBlock(BB: Pred);
1087 const Value *V2 = PN.getIncomingValueForBlock(BB);
1088
1089 // If V2 is a phi node in BB, look up what the mapped value will be.
1090 if (const PHINode *V2PN = dyn_cast<PHINode>(Val: V2))
1091 if (V2PN->getParent() == BB)
1092 V2 = V2PN->getIncomingValueForBlock(BB: Pred);
1093
1094 // If there is a conflict, bail out.
1095 if (V1 != V2)
1096 return false;
1097 }
1098 }
1099 }
1100
1101 return true;
1102}
1103
1104/// Replace all old uses with new ones, and push the updated BBs into FreshBBs.
1105static void replaceAllUsesWith(Value *Old, Value *New,
1106 SmallPtrSet<BasicBlock *, 32> &FreshBBs,
1107 bool IsHuge) {
1108 auto *OldI = dyn_cast<Instruction>(Val: Old);
1109 if (OldI) {
1110 for (Value::user_iterator UI = OldI->user_begin(), E = OldI->user_end();
1111 UI != E; ++UI) {
1112 Instruction *User = cast<Instruction>(Val: *UI);
1113 if (IsHuge)
1114 FreshBBs.insert(Ptr: User->getParent());
1115 }
1116 }
1117 Old->replaceAllUsesWith(V: New);
1118}
1119
1120/// Eliminate a basic block that has only phi's and an unconditional branch in
1121/// it.
1122void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
1123 UncondBrInst *BI = cast<UncondBrInst>(Val: BB->getTerminator());
1124 BasicBlock *DestBB = BI->getSuccessor();
1125
1126 LLVM_DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"
1127 << *BB << *DestBB);
1128
1129 // If the destination block has a single pred, then this is a trivial edge,
1130 // just collapse it.
1131 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
1132 if (SinglePred != DestBB) {
1133 assert(SinglePred == BB &&
1134 "Single predecessor not the same as predecessor");
1135 // Merge DestBB into SinglePred/BB and delete it.
1136 MergeBlockIntoPredecessor(BB: DestBB);
1137 // Note: BB(=SinglePred) will not be deleted on this path.
1138 // DestBB(=its single successor) is the one that was deleted.
1139 LLVM_DEBUG(dbgs() << "AFTER:\n" << *SinglePred << "\n\n\n");
1140
1141 if (IsHugeFunc) {
1142 // Update FreshBBs to optimize the merged BB.
1143 FreshBBs.insert(Ptr: SinglePred);
1144 FreshBBs.erase(Ptr: DestBB);
1145 }
1146 return;
1147 }
1148 }
1149
1150 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
1151 // to handle the new incoming edges it is about to have.
1152 for (PHINode &PN : DestBB->phis()) {
1153 // Remove the incoming value for BB, and remember it.
1154 Value *InVal = PN.removeIncomingValue(BB, DeletePHIIfEmpty: false);
1155
1156 // Two options: either the InVal is a phi node defined in BB or it is some
1157 // value that dominates BB.
1158 PHINode *InValPhi = dyn_cast<PHINode>(Val: InVal);
1159 if (InValPhi && InValPhi->getParent() == BB) {
1160 // Add all of the input values of the input PHI as inputs of this phi.
1161 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
1162 PN.addIncoming(V: InValPhi->getIncomingValue(i),
1163 BB: InValPhi->getIncomingBlock(i));
1164 } else {
1165 // Otherwise, add one instance of the dominating value for each edge that
1166 // we will be adding.
1167 if (PHINode *BBPN = dyn_cast<PHINode>(Val: BB->begin())) {
1168 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1169 PN.addIncoming(V: InVal, BB: BBPN->getIncomingBlock(i));
1170 } else {
1171 for (BasicBlock *Pred : predecessors(BB))
1172 PN.addIncoming(V: InVal, BB: Pred);
1173 }
1174 }
1175 }
1176
1177 // Preserve loop Metadata.
1178 if (BI->hasMetadata(KindID: LLVMContext::MD_loop)) {
1179 for (auto *Pred : predecessors(BB))
1180 Pred->getTerminator()->copyMetadata(SrcInst: *BI, WL: LLVMContext::MD_loop);
1181 }
1182
1183 // The PHIs are now updated, change everything that refers to BB to use
1184 // DestBB and remove BB.
1185 BB->replaceAllUsesWith(V: DestBB);
1186 BB->eraseFromParent();
1187 ++NumBlocksElim;
1188
1189 LLVM_DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1190}
1191
1192// Computes a map of base pointer relocation instructions to corresponding
1193// derived pointer relocation instructions given a vector of all relocate calls
1194static void computeBaseDerivedRelocateMap(
1195 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
1196 MapVector<GCRelocateInst *, SmallVector<GCRelocateInst *, 0>>
1197 &RelocateInstMap) {
1198 // Collect information in two maps: one primarily for locating the base object
1199 // while filling the second map; the second map is the final structure holding
1200 // a mapping between Base and corresponding Derived relocate calls
1201 MapVector<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
1202 for (auto *ThisRelocate : AllRelocateCalls) {
1203 auto K = std::make_pair(x: ThisRelocate->getBasePtrIndex(),
1204 y: ThisRelocate->getDerivedPtrIndex());
1205 RelocateIdxMap.insert(KV: std::make_pair(x&: K, y&: ThisRelocate));
1206 }
1207 for (auto &Item : RelocateIdxMap) {
1208 std::pair<unsigned, unsigned> Key = Item.first;
1209 if (Key.first == Key.second)
1210 // Base relocation: nothing to insert
1211 continue;
1212
1213 GCRelocateInst *I = Item.second;
1214 auto BaseKey = std::make_pair(x&: Key.first, y&: Key.first);
1215
1216 // We're iterating over RelocateIdxMap so we cannot modify it.
1217 auto MaybeBase = RelocateIdxMap.find(Key: BaseKey);
1218 if (MaybeBase == RelocateIdxMap.end())
1219 // TODO: We might want to insert a new base object relocate and gep off
1220 // that, if there are enough derived object relocates.
1221 continue;
1222
1223 RelocateInstMap[MaybeBase->second].push_back(Elt: I);
1224 }
1225}
1226
1227// Accepts a GEP and extracts the operands into a vector provided they're all
1228// small integer constants
1229static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
1230 SmallVectorImpl<Value *> &OffsetV) {
1231 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
1232 // Only accept small constant integer operands
1233 auto *Op = dyn_cast<ConstantInt>(Val: GEP->getOperand(i_nocapture: i));
1234 if (!Op || Op->getZExtValue() > 20)
1235 return false;
1236 }
1237
1238 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
1239 OffsetV.push_back(Elt: GEP->getOperand(i_nocapture: i));
1240 return true;
1241}
1242
1243// Takes a RelocatedBase (base pointer relocation instruction) and Targets to
1244// replace, computes a replacement, and affects it.
1245static bool
1246simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
1247 const SmallVectorImpl<GCRelocateInst *> &Targets) {
1248 bool MadeChange = false;
1249 // We must ensure the relocation of derived pointer is defined after
1250 // relocation of base pointer. If we find a relocation corresponding to base
1251 // defined earlier than relocation of base then we move relocation of base
1252 // right before found relocation. We consider only relocation in the same
1253 // basic block as relocation of base. Relocations from other basic block will
1254 // be skipped by optimization and we do not care about them.
1255 for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
1256 &*R != RelocatedBase; ++R)
1257 if (auto *RI = dyn_cast<GCRelocateInst>(Val&: R))
1258 if (RI->getStatepoint() == RelocatedBase->getStatepoint())
1259 if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
1260 RelocatedBase->moveBefore(InsertPos: RI->getIterator());
1261 MadeChange = true;
1262 break;
1263 }
1264
1265 for (GCRelocateInst *ToReplace : Targets) {
1266 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
1267 "Not relocating a derived object of the original base object");
1268 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
1269 // A duplicate relocate call. TODO: coalesce duplicates.
1270 continue;
1271 }
1272
1273 if (RelocatedBase->getParent() != ToReplace->getParent()) {
1274 // Base and derived relocates are in different basic blocks.
1275 // In this case transform is only valid when base dominates derived
1276 // relocate. However it would be too expensive to check dominance
1277 // for each such relocate, so we skip the whole transformation.
1278 continue;
1279 }
1280
1281 Value *Base = ToReplace->getBasePtr();
1282 auto *Derived = dyn_cast<GetElementPtrInst>(Val: ToReplace->getDerivedPtr());
1283 if (!Derived || Derived->getPointerOperand() != Base)
1284 continue;
1285
1286 SmallVector<Value *, 2> OffsetV;
1287 if (!getGEPSmallConstantIntOffsetV(GEP: Derived, OffsetV))
1288 continue;
1289
1290 // Create a Builder and replace the target callsite with a gep
1291 assert(RelocatedBase->getNextNode() &&
1292 "Should always have one since it's not a terminator");
1293
1294 // Insert after RelocatedBase
1295 IRBuilder<> Builder(RelocatedBase->getNextNode());
1296 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1297
1298 // If gc_relocate does not match the actual type, cast it to the right type.
1299 // In theory, there must be a bitcast after gc_relocate if the type does not
1300 // match, and we should reuse it to get the derived pointer. But it could be
1301 // cases like this:
1302 // bb1:
1303 // ...
1304 // %g1 = call coldcc i8 addrspace(1)*
1305 // @llvm.experimental.gc.relocate.p1i8(...) br label %merge
1306 //
1307 // bb2:
1308 // ...
1309 // %g2 = call coldcc i8 addrspace(1)*
1310 // @llvm.experimental.gc.relocate.p1i8(...) br label %merge
1311 //
1312 // merge:
1313 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
1314 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
1315 //
1316 // In this case, we can not find the bitcast any more. So we insert a new
1317 // bitcast no matter there is already one or not. In this way, we can handle
1318 // all cases, and the extra bitcast should be optimized away in later
1319 // passes.
1320 Value *ActualRelocatedBase = RelocatedBase;
1321 if (RelocatedBase->getType() != Base->getType()) {
1322 ActualRelocatedBase =
1323 Builder.CreateBitCast(V: RelocatedBase, DestTy: Base->getType());
1324 }
1325 Value *Replacement =
1326 Builder.CreateGEP(Ty: Derived->getSourceElementType(), Ptr: ActualRelocatedBase,
1327 IdxList: ArrayRef(OffsetV));
1328 Replacement->takeName(V: ToReplace);
1329 // If the newly generated derived pointer's type does not match the original
1330 // derived pointer's type, cast the new derived pointer to match it. Same
1331 // reasoning as above.
1332 Value *ActualReplacement = Replacement;
1333 if (Replacement->getType() != ToReplace->getType()) {
1334 ActualReplacement =
1335 Builder.CreateBitCast(V: Replacement, DestTy: ToReplace->getType());
1336 }
1337 ToReplace->replaceAllUsesWith(V: ActualReplacement);
1338 ToReplace->eraseFromParent();
1339
1340 MadeChange = true;
1341 }
1342 return MadeChange;
1343}
1344
1345// Turns this:
1346//
1347// %base = ...
1348// %ptr = gep %base + 15
1349// %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1350// %base' = relocate(%tok, i32 4, i32 4)
1351// %ptr' = relocate(%tok, i32 4, i32 5)
1352// %val = load %ptr'
1353//
1354// into this:
1355//
1356// %base = ...
1357// %ptr = gep %base + 15
1358// %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1359// %base' = gc.relocate(%tok, i32 4, i32 4)
1360// %ptr' = gep %base' + 15
1361// %val = load %ptr'
1362bool CodeGenPrepare::simplifyOffsetableRelocate(GCStatepointInst &I) {
1363 bool MadeChange = false;
1364 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1365 for (auto *U : I.users())
1366 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(Val: U))
1367 // Collect all the relocate calls associated with a statepoint
1368 AllRelocateCalls.push_back(Elt: Relocate);
1369
1370 // We need at least one base pointer relocation + one derived pointer
1371 // relocation to mangle
1372 if (AllRelocateCalls.size() < 2)
1373 return false;
1374
1375 // RelocateInstMap is a mapping from the base relocate instruction to the
1376 // corresponding derived relocate instructions
1377 MapVector<GCRelocateInst *, SmallVector<GCRelocateInst *, 0>> RelocateInstMap;
1378 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1379 if (RelocateInstMap.empty())
1380 return false;
1381
1382 for (auto &Item : RelocateInstMap)
1383 // Item.first is the RelocatedBase to offset against
1384 // Item.second is the vector of Targets to replace
1385 MadeChange = simplifyRelocatesOffABase(RelocatedBase: Item.first, Targets: Item.second);
1386 return MadeChange;
1387}
1388
1389/// Sink the specified cast instruction into its user blocks.
1390static bool SinkCast(CastInst *CI) {
1391 BasicBlock *DefBB = CI->getParent();
1392
1393 /// InsertedCasts - Only insert a cast in each block once.
1394 DenseMap<BasicBlock *, CastInst *> InsertedCasts;
1395
1396 bool MadeChange = false;
1397 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1398 UI != E;) {
1399 Use &TheUse = UI.getUse();
1400 Instruction *User = cast<Instruction>(Val: *UI);
1401
1402 // Figure out which BB this cast is used in. For PHI's this is the
1403 // appropriate predecessor block.
1404 BasicBlock *UserBB = User->getParent();
1405 if (PHINode *PN = dyn_cast<PHINode>(Val: User)) {
1406 UserBB = PN->getIncomingBlock(U: TheUse);
1407 }
1408
1409 // Preincrement use iterator so we don't invalidate it.
1410 ++UI;
1411
1412 // The first insertion point of a block containing an EH pad is after the
1413 // pad. If the pad is the user, we cannot sink the cast past the pad.
1414 if (User->isEHPad())
1415 continue;
1416
1417 // If the block selected to receive the cast is an EH pad that does not
1418 // allow non-PHI instructions before the terminator, we can't sink the
1419 // cast.
1420 if (UserBB->getTerminator()->isEHPad())
1421 continue;
1422
1423 // If this user is in the same block as the cast, don't change the cast.
1424 if (UserBB == DefBB)
1425 continue;
1426
1427 // If we have already inserted a cast into this block, use it.
1428 CastInst *&InsertedCast = InsertedCasts[UserBB];
1429
1430 if (!InsertedCast) {
1431 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1432 assert(InsertPt != UserBB->end());
1433 InsertedCast = cast<CastInst>(Val: CI->clone());
1434 InsertedCast->insertBefore(BB&: *UserBB, InsertPos: InsertPt);
1435 }
1436
1437 // Replace a use of the cast with a use of the new cast.
1438 TheUse = InsertedCast;
1439 MadeChange = true;
1440 ++NumCastUses;
1441 }
1442
1443 // If we removed all uses, nuke the cast.
1444 if (CI->use_empty()) {
1445 salvageDebugInfo(I&: *CI);
1446 CI->eraseFromParent();
1447 MadeChange = true;
1448 }
1449
1450 return MadeChange;
1451}
1452
1453/// If the specified cast instruction is a noop copy (e.g. it's casting from
1454/// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1455/// reduce the number of virtual registers that must be created and coalesced.
1456///
1457/// Return true if any changes are made.
1458static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1459 const DataLayout &DL) {
1460 // Sink only "cheap" (or nop) address-space casts. This is a weaker condition
1461 // than sinking only nop casts, but is helpful on some platforms.
1462 if (auto *ASC = dyn_cast<AddrSpaceCastInst>(Val: CI)) {
1463 if (!TLI.isFreeAddrSpaceCast(SrcAS: ASC->getSrcAddressSpace(),
1464 DestAS: ASC->getDestAddressSpace()))
1465 return false;
1466 }
1467
1468 // If this is a noop copy,
1469 EVT SrcVT = TLI.getValueType(DL, Ty: CI->getOperand(i_nocapture: 0)->getType());
1470 EVT DstVT = TLI.getValueType(DL, Ty: CI->getType());
1471
1472 // This is an fp<->int conversion?
1473 if (SrcVT.isInteger() != DstVT.isInteger())
1474 return false;
1475
1476 // If this is an extension, it will be a zero or sign extension, which
1477 // isn't a noop.
1478 if (SrcVT.bitsLT(VT: DstVT))
1479 return false;
1480
1481 // If these values will be promoted, find out what they will be promoted
1482 // to. This helps us consider truncates on PPC as noop copies when they
1483 // are.
1484 if (TLI.getTypeAction(Context&: CI->getContext(), VT: SrcVT) ==
1485 TargetLowering::TypePromoteInteger)
1486 SrcVT = TLI.getTypeToTransformTo(Context&: CI->getContext(), VT: SrcVT);
1487 if (TLI.getTypeAction(Context&: CI->getContext(), VT: DstVT) ==
1488 TargetLowering::TypePromoteInteger)
1489 DstVT = TLI.getTypeToTransformTo(Context&: CI->getContext(), VT: DstVT);
1490
1491 // If, after promotion, these are the same types, this is a noop copy.
1492 if (SrcVT != DstVT)
1493 return false;
1494
1495 return SinkCast(CI);
1496}
1497
1498// Match a simple increment by constant operation. Note that if a sub is
1499// matched, the step is negated (as if the step had been canonicalized to
1500// an add, even though we leave the instruction alone.)
1501static bool matchIncrement(const Instruction *IVInc, Instruction *&LHS,
1502 Constant *&Step) {
1503 if (match(V: IVInc, P: m_Add(L: m_Instruction(I&: LHS), R: m_Constant(C&: Step))) ||
1504 match(V: IVInc, P: m_ExtractValue<0>(V: m_Intrinsic<Intrinsic::uadd_with_overflow>(
1505 Op0: m_Instruction(I&: LHS), Op1: m_Constant(C&: Step)))))
1506 return true;
1507 if (match(V: IVInc, P: m_Sub(L: m_Instruction(I&: LHS), R: m_Constant(C&: Step))) ||
1508 match(V: IVInc, P: m_ExtractValue<0>(V: m_Intrinsic<Intrinsic::usub_with_overflow>(
1509 Op0: m_Instruction(I&: LHS), Op1: m_Constant(C&: Step))))) {
1510 Step = ConstantExpr::getNeg(C: Step);
1511 return true;
1512 }
1513 return false;
1514}
1515
1516/// If given \p PN is an inductive variable with value IVInc coming from the
1517/// backedge, and on each iteration it gets increased by Step, return pair
1518/// <IVInc, Step>. Otherwise, return std::nullopt.
1519static std::optional<std::pair<Instruction *, Constant *>>
1520getIVIncrement(const PHINode *PN, const LoopInfo *LI) {
1521 const Loop *L = LI->getLoopFor(BB: PN->getParent());
1522 if (!L || L->getHeader() != PN->getParent() || !L->getLoopLatch())
1523 return std::nullopt;
1524 auto *IVInc =
1525 dyn_cast<Instruction>(Val: PN->getIncomingValueForBlock(BB: L->getLoopLatch()));
1526 if (!IVInc || LI->getLoopFor(BB: IVInc->getParent()) != L)
1527 return std::nullopt;
1528 Instruction *LHS = nullptr;
1529 Constant *Step = nullptr;
1530 if (matchIncrement(IVInc, LHS, Step) && LHS == PN)
1531 return std::make_pair(x&: IVInc, y&: Step);
1532 return std::nullopt;
1533}
1534
1535static bool isIVIncrement(const Value *V, const LoopInfo *LI) {
1536 auto *I = dyn_cast<Instruction>(Val: V);
1537 if (!I)
1538 return false;
1539 Instruction *LHS = nullptr;
1540 Constant *Step = nullptr;
1541 if (!matchIncrement(IVInc: I, LHS, Step))
1542 return false;
1543 if (auto *PN = dyn_cast<PHINode>(Val: LHS))
1544 if (auto IVInc = getIVIncrement(PN, LI))
1545 return IVInc->first == I;
1546 return false;
1547}
1548
1549bool CodeGenPrepare::replaceMathCmpWithIntrinsic(BinaryOperator *BO,
1550 Value *Arg0, Value *Arg1,
1551 CmpInst *Cmp,
1552 Intrinsic::ID IID) {
1553 auto IsReplacableIVIncrement = [this, &Cmp](BinaryOperator *BO) {
1554 if (!isIVIncrement(V: BO, LI))
1555 return false;
1556 const Loop *L = LI->getLoopFor(BB: BO->getParent());
1557 assert(L && "L should not be null after isIVIncrement()");
1558 // Do not risk on moving increment into a child loop.
1559 if (LI->getLoopFor(BB: Cmp->getParent()) != L)
1560 return false;
1561
1562 // Finally, we need to ensure that the insert point will dominate all
1563 // existing uses of the increment.
1564
1565 auto &DT = getDT(F&: *BO->getParent()->getParent());
1566 if (DT.dominates(A: Cmp->getParent(), B: BO->getParent()))
1567 // If we're moving up the dom tree, all uses are trivially dominated.
1568 // (This is the common case for code produced by LSR.)
1569 return true;
1570
1571 // Otherwise, special case the single use in the phi recurrence.
1572 return BO->hasOneUse() && DT.dominates(A: Cmp->getParent(), B: L->getLoopLatch());
1573 };
1574 if (BO->getParent() != Cmp->getParent() && !IsReplacableIVIncrement(BO)) {
1575 // We used to use a dominator tree here to allow multi-block optimization.
1576 // But that was problematic because:
1577 // 1. It could cause a perf regression by hoisting the math op into the
1578 // critical path.
1579 // 2. It could cause a perf regression by creating a value that was live
1580 // across multiple blocks and increasing register pressure.
1581 // 3. Use of a dominator tree could cause large compile-time regression.
1582 // This is because we recompute the DT on every change in the main CGP
1583 // run-loop. The recomputing is probably unnecessary in many cases, so if
1584 // that was fixed, using a DT here would be ok.
1585 //
1586 // There is one important particular case we still want to handle: if BO is
1587 // the IV increment. Important properties that make it profitable:
1588 // - We can speculate IV increment anywhere in the loop (as long as the
1589 // indvar Phi is its only user);
1590 // - Upon computing Cmp, we effectively compute something equivalent to the
1591 // IV increment (despite it loops differently in the IR). So moving it up
1592 // to the cmp point does not really increase register pressure.
1593 return false;
1594 }
1595
1596 // We allow matching the canonical IR (add X, C) back to (usubo X, -C).
1597 if (BO->getOpcode() == Instruction::Add &&
1598 IID == Intrinsic::usub_with_overflow) {
1599 assert(isa<Constant>(Arg1) && "Unexpected input for usubo");
1600 Arg1 = ConstantExpr::getNeg(C: cast<Constant>(Val: Arg1));
1601 }
1602
1603 // Insert at the first instruction of the pair.
1604 Instruction *InsertPt = nullptr;
1605 for (Instruction &Iter : *Cmp->getParent()) {
1606 // If BO is an XOR, it is not guaranteed that it comes after both inputs to
1607 // the overflow intrinsic are defined.
1608 if ((BO->getOpcode() != Instruction::Xor && &Iter == BO) || &Iter == Cmp) {
1609 InsertPt = &Iter;
1610 break;
1611 }
1612 }
1613 assert(InsertPt != nullptr && "Parent block did not contain cmp or binop");
1614
1615 IRBuilder<> Builder(InsertPt);
1616 Value *MathOV = Builder.CreateBinaryIntrinsic(ID: IID, LHS: Arg0, RHS: Arg1);
1617 if (BO->getOpcode() != Instruction::Xor) {
1618 Value *Math = Builder.CreateExtractValue(Agg: MathOV, Idxs: 0, Name: "math");
1619 replaceAllUsesWith(Old: BO, New: Math, FreshBBs, IsHuge: IsHugeFunc);
1620 } else
1621 assert(BO->hasOneUse() &&
1622 "Patterns with XOr should use the BO only in the compare");
1623 Value *OV = Builder.CreateExtractValue(Agg: MathOV, Idxs: 1, Name: "ov");
1624 replaceAllUsesWith(Old: Cmp, New: OV, FreshBBs, IsHuge: IsHugeFunc);
1625 Cmp->eraseFromParent();
1626 BO->eraseFromParent();
1627 return true;
1628}
1629
1630/// Match special-case patterns that check for unsigned add overflow.
1631static bool matchUAddWithOverflowConstantEdgeCases(CmpInst *Cmp,
1632 BinaryOperator *&Add) {
1633 // Add = add A, 1; Cmp = icmp eq A,-1 (overflow if A is max val)
1634 // Add = add A,-1; Cmp = icmp ne A, 0 (overflow if A is non-zero)
1635 Value *A = Cmp->getOperand(i_nocapture: 0), *B = Cmp->getOperand(i_nocapture: 1);
1636
1637 // We are not expecting non-canonical/degenerate code. Just bail out.
1638 if (isa<Constant>(Val: A))
1639 return false;
1640
1641 ICmpInst::Predicate Pred = Cmp->getPredicate();
1642 if (Pred == ICmpInst::ICMP_EQ && match(V: B, P: m_AllOnes()))
1643 B = ConstantInt::get(Ty: B->getType(), V: 1);
1644 else if (Pred == ICmpInst::ICMP_NE && match(V: B, P: m_ZeroInt()))
1645 B = Constant::getAllOnesValue(Ty: B->getType());
1646 else
1647 return false;
1648
1649 // Check the users of the variable operand of the compare looking for an add
1650 // with the adjusted constant.
1651 for (User *U : A->users()) {
1652 if (match(V: U, P: m_Add(L: m_Specific(V: A), R: m_Specific(V: B)))) {
1653 Add = cast<BinaryOperator>(Val: U);
1654 return true;
1655 }
1656 }
1657 return false;
1658}
1659
1660/// Try to combine the compare into a call to the llvm.uadd.with.overflow
1661/// intrinsic. Return true if any changes were made.
1662bool CodeGenPrepare::combineToUAddWithOverflow(CmpInst *Cmp,
1663 ModifyDT &ModifiedDT) {
1664 bool EdgeCase = false;
1665 Value *A, *B;
1666 BinaryOperator *Add;
1667 if (!match(V: Cmp, P: m_UAddWithOverflow(L: m_Value(V&: A), R: m_Value(V&: B), S: m_BinOp(I&: Add)))) {
1668 if (!matchUAddWithOverflowConstantEdgeCases(Cmp, Add))
1669 return false;
1670 // Set A and B in case we match matchUAddWithOverflowConstantEdgeCases.
1671 A = Add->getOperand(i_nocapture: 0);
1672 B = Add->getOperand(i_nocapture: 1);
1673 EdgeCase = true;
1674 }
1675
1676 if (!TLI->shouldFormOverflowOp(Opcode: ISD::UADDO,
1677 VT: TLI->getValueType(DL: *DL, Ty: Add->getType()),
1678 MathUsed: Add->hasNUsesOrMore(N: EdgeCase ? 1 : 2)))
1679 return false;
1680
1681 // We don't want to move around uses of condition values this late, so we
1682 // check if it is legal to create the call to the intrinsic in the basic
1683 // block containing the icmp.
1684 if (Add->getParent() != Cmp->getParent() && !Add->hasOneUse())
1685 return false;
1686
1687 if (!replaceMathCmpWithIntrinsic(BO: Add, Arg0: A, Arg1: B, Cmp,
1688 IID: Intrinsic::uadd_with_overflow))
1689 return false;
1690
1691 // Reset callers - do not crash by iterating over a dead instruction.
1692 ModifiedDT = ModifyDT::ModifyInstDT;
1693 return true;
1694}
1695
1696bool CodeGenPrepare::combineToUSubWithOverflow(CmpInst *Cmp,
1697 ModifyDT &ModifiedDT) {
1698 // We are not expecting non-canonical/degenerate code. Just bail out.
1699 Value *A = Cmp->getOperand(i_nocapture: 0), *B = Cmp->getOperand(i_nocapture: 1);
1700 if (isa<Constant>(Val: A) && isa<Constant>(Val: B))
1701 return false;
1702
1703 // Convert (A u> B) to (A u< B) to simplify pattern matching.
1704 ICmpInst::Predicate Pred = Cmp->getPredicate();
1705 if (Pred == ICmpInst::ICMP_UGT) {
1706 std::swap(a&: A, b&: B);
1707 Pred = ICmpInst::ICMP_ULT;
1708 }
1709 // Convert special-case: (A == 0) is the same as (A u< 1).
1710 if (Pred == ICmpInst::ICMP_EQ && match(V: B, P: m_ZeroInt())) {
1711 B = ConstantInt::get(Ty: B->getType(), V: 1);
1712 Pred = ICmpInst::ICMP_ULT;
1713 }
1714 // Convert special-case: (A != 0) is the same as (0 u< A).
1715 if (Pred == ICmpInst::ICMP_NE && match(V: B, P: m_ZeroInt())) {
1716 std::swap(a&: A, b&: B);
1717 Pred = ICmpInst::ICMP_ULT;
1718 }
1719 if (Pred != ICmpInst::ICMP_ULT)
1720 return false;
1721
1722 // Walk the users of a variable operand of a compare looking for a subtract or
1723 // add with that same operand. Also match the 2nd operand of the compare to
1724 // the add/sub, but that may be a negated constant operand of an add.
1725 Value *CmpVariableOperand = isa<Constant>(Val: A) ? B : A;
1726 BinaryOperator *Sub = nullptr;
1727 for (User *U : CmpVariableOperand->users()) {
1728 // A - B, A u< B --> usubo(A, B)
1729 if (match(V: U, P: m_Sub(L: m_Specific(V: A), R: m_Specific(V: B)))) {
1730 Sub = cast<BinaryOperator>(Val: U);
1731 break;
1732 }
1733
1734 // A + (-C), A u< C (canonicalized form of (sub A, C))
1735 const APInt *CmpC, *AddC;
1736 if (match(V: U, P: m_Add(L: m_Specific(V: A), R: m_APInt(Res&: AddC))) &&
1737 match(V: B, P: m_APInt(Res&: CmpC)) && *AddC == -(*CmpC)) {
1738 Sub = cast<BinaryOperator>(Val: U);
1739 break;
1740 }
1741 }
1742 if (!Sub)
1743 return false;
1744
1745 if (!TLI->shouldFormOverflowOp(Opcode: ISD::USUBO,
1746 VT: TLI->getValueType(DL: *DL, Ty: Sub->getType()),
1747 MathUsed: Sub->hasNUsesOrMore(N: 1)))
1748 return false;
1749
1750 // We don't want to move around uses of condition values this late, so we
1751 // check if it is legal to create the call to the intrinsic in the basic
1752 // block containing the icmp.
1753 if (Sub->getParent() != Cmp->getParent() && !Sub->hasOneUse())
1754 return false;
1755
1756 if (!replaceMathCmpWithIntrinsic(BO: Sub, Arg0: Sub->getOperand(i_nocapture: 0), Arg1: Sub->getOperand(i_nocapture: 1),
1757 Cmp, IID: Intrinsic::usub_with_overflow))
1758 return false;
1759
1760 // Reset callers - do not crash by iterating over a dead instruction.
1761 ModifiedDT = ModifyDT::ModifyInstDT;
1762 return true;
1763}
1764
1765// Decanonicalizes icmp+ctpop power-of-two test if ctpop is slow.
1766// The same transformation exists in DAG combiner, but we repeat it here because
1767// DAG builder can break the pattern by moving icmp into a successor block.
1768bool CodeGenPrepare::unfoldPowerOf2Test(CmpInst *Cmp) {
1769 CmpPredicate Pred;
1770 Value *X;
1771 const APInt *C;
1772
1773 // (icmp (ctpop x), c)
1774 if (!match(V: Cmp, P: m_ICmp(Pred, L: m_Intrinsic<Intrinsic::ctpop>(Op0: m_Value(V&: X)),
1775 R: m_APIntAllowPoison(Res&: C))))
1776 return false;
1777
1778 // We're only interested in "is power of 2 [or zero]" patterns.
1779 bool IsStrictlyPowerOf2Test = ICmpInst::isEquality(P: Pred) && *C == 1;
1780 bool IsPowerOf2OrZeroTest = (Pred == CmpInst::ICMP_ULT && *C == 2) ||
1781 (Pred == CmpInst::ICMP_UGT && *C == 1);
1782 if (!IsStrictlyPowerOf2Test && !IsPowerOf2OrZeroTest)
1783 return false;
1784
1785 // Some targets have better codegen for `ctpop(x) u</u>= 2/1`than for
1786 // `ctpop(x) ==/!= 1`. If ctpop is fast, only try changing the comparison,
1787 // and otherwise expand ctpop into a few simple instructions.
1788 Type *OpTy = X->getType();
1789 if (TLI->isCtpopFast(VT: TLI->getValueType(DL: *DL, Ty: OpTy))) {
1790 // Look for `ctpop(x) ==/!= 1`, where `ctpop(x)` is known to be non-zero.
1791 if (!IsStrictlyPowerOf2Test || !isKnownNonZero(V: Cmp->getOperand(i_nocapture: 0), Q: *DL))
1792 return false;
1793
1794 // ctpop(x) == 1 -> ctpop(x) u< 2
1795 // ctpop(x) != 1 -> ctpop(x) u> 1
1796 if (Pred == ICmpInst::ICMP_EQ) {
1797 Cmp->setOperand(i_nocapture: 1, Val_nocapture: ConstantInt::get(Ty: OpTy, V: 2));
1798 Cmp->setPredicate(ICmpInst::ICMP_ULT);
1799 } else {
1800 Cmp->setPredicate(ICmpInst::ICMP_UGT);
1801 }
1802 return true;
1803 }
1804
1805 Value *NewCmp;
1806 if (IsPowerOf2OrZeroTest ||
1807 (IsStrictlyPowerOf2Test && isKnownNonZero(V: Cmp->getOperand(i_nocapture: 0), Q: *DL))) {
1808 // ctpop(x) u< 2 -> (x & (x - 1)) == 0
1809 // ctpop(x) u> 1 -> (x & (x - 1)) != 0
1810 IRBuilder<> Builder(Cmp);
1811 Value *Sub = Builder.CreateAdd(LHS: X, RHS: Constant::getAllOnesValue(Ty: OpTy));
1812 Value *And = Builder.CreateAnd(LHS: X, RHS: Sub);
1813 CmpInst::Predicate NewPred =
1814 (Pred == CmpInst::ICMP_ULT || Pred == CmpInst::ICMP_EQ)
1815 ? CmpInst::ICMP_EQ
1816 : CmpInst::ICMP_NE;
1817 NewCmp = Builder.CreateICmp(P: NewPred, LHS: And, RHS: ConstantInt::getNullValue(Ty: OpTy));
1818 } else {
1819 // ctpop(x) == 1 -> (x ^ (x - 1)) u> (x - 1)
1820 // ctpop(x) != 1 -> (x ^ (x - 1)) u<= (x - 1)
1821 IRBuilder<> Builder(Cmp);
1822 Value *Sub = Builder.CreateAdd(LHS: X, RHS: Constant::getAllOnesValue(Ty: OpTy));
1823 Value *Xor = Builder.CreateXor(LHS: X, RHS: Sub);
1824 CmpInst::Predicate NewPred =
1825 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT : CmpInst::ICMP_ULE;
1826 NewCmp = Builder.CreateICmp(P: NewPred, LHS: Xor, RHS: Sub);
1827 }
1828
1829 Cmp->replaceAllUsesWith(V: NewCmp);
1830 RecursivelyDeleteTriviallyDeadInstructions(V: Cmp);
1831 return true;
1832}
1833
1834/// Sink the given CmpInst into user blocks to reduce the number of virtual
1835/// registers that must be created and coalesced. This is a clear win except on
1836/// targets with multiple condition code registers (PowerPC), where it might
1837/// lose; some adjustment may be wanted there.
1838///
1839/// Return true if any changes are made.
1840static bool sinkCmpExpression(CmpInst *Cmp, const TargetLowering &TLI,
1841 const DataLayout &DL) {
1842 if (TLI.hasMultipleConditionRegisters(VT: EVT::getEVT(Ty: Cmp->getType())))
1843 return false;
1844
1845 // Avoid sinking soft-FP comparisons, since this can move them into a loop.
1846 if (TLI.useSoftFloat() && isa<FCmpInst>(Val: Cmp))
1847 return false;
1848
1849 bool UsedInPhiOrCurrentBlock = any_of(Range: Cmp->users(), P: [Cmp](User *U) {
1850 return isa<PHINode>(Val: U) ||
1851 cast<Instruction>(Val: U)->getParent() == Cmp->getParent();
1852 });
1853
1854 // Avoid sinking larger than legal integer comparisons unless its ONLY used in
1855 // another BB.
1856 if (UsedInPhiOrCurrentBlock && Cmp->getOperand(i_nocapture: 0)->getType()->isIntegerTy() &&
1857 Cmp->getOperand(i_nocapture: 0)->getType()->getScalarSizeInBits() >
1858 DL.getLargestLegalIntTypeSizeInBits())
1859 return false;
1860
1861 // Only insert a cmp in each block once.
1862 DenseMap<BasicBlock *, CmpInst *> InsertedCmps;
1863
1864 bool MadeChange = false;
1865 for (Value::user_iterator UI = Cmp->user_begin(), E = Cmp->user_end();
1866 UI != E;) {
1867 Use &TheUse = UI.getUse();
1868 Instruction *User = cast<Instruction>(Val: *UI);
1869
1870 // Preincrement use iterator so we don't invalidate it.
1871 ++UI;
1872
1873 // Don't bother for PHI nodes.
1874 if (isa<PHINode>(Val: User))
1875 continue;
1876
1877 // Figure out which BB this cmp is used in.
1878 BasicBlock *UserBB = User->getParent();
1879 BasicBlock *DefBB = Cmp->getParent();
1880
1881 // If this user is in the same block as the cmp, don't change the cmp.
1882 if (UserBB == DefBB)
1883 continue;
1884
1885 // If we have already inserted a cmp into this block, use it.
1886 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1887
1888 if (!InsertedCmp) {
1889 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1890 assert(InsertPt != UserBB->end());
1891 InsertedCmp = CmpInst::Create(Op: Cmp->getOpcode(), Pred: Cmp->getPredicate(),
1892 S1: Cmp->getOperand(i_nocapture: 0), S2: Cmp->getOperand(i_nocapture: 1), Name: "");
1893 InsertedCmp->insertBefore(BB&: *UserBB, InsertPos: InsertPt);
1894 // Propagate the debug info.
1895 InsertedCmp->setDebugLoc(Cmp->getDebugLoc());
1896 }
1897
1898 // Replace a use of the cmp with a use of the new cmp.
1899 TheUse = InsertedCmp;
1900 MadeChange = true;
1901 ++NumCmpUses;
1902 }
1903
1904 // If we removed all uses, nuke the cmp.
1905 if (Cmp->use_empty()) {
1906 Cmp->eraseFromParent();
1907 MadeChange = true;
1908 }
1909
1910 return MadeChange;
1911}
1912
1913/// For pattern like:
1914///
1915/// DomCond = icmp sgt/slt CmpOp0, CmpOp1 (might not be in DomBB)
1916/// ...
1917/// DomBB:
1918/// ...
1919/// br DomCond, TrueBB, CmpBB
1920/// CmpBB: (with DomBB being the single predecessor)
1921/// ...
1922/// Cmp = icmp eq CmpOp0, CmpOp1
1923/// ...
1924///
1925/// It would use two comparison on targets that lowering of icmp sgt/slt is
1926/// different from lowering of icmp eq (PowerPC). This function try to convert
1927/// 'Cmp = icmp eq CmpOp0, CmpOp1' to ' Cmp = icmp slt/sgt CmpOp0, CmpOp1'.
1928/// After that, DomCond and Cmp can use the same comparison so reduce one
1929/// comparison.
1930///
1931/// Return true if any changes are made.
1932static bool foldICmpWithDominatingICmp(CmpInst *Cmp,
1933 const TargetLowering &TLI) {
1934 if (!EnableICMP_EQToICMP_ST && TLI.isEqualityCmpFoldedWithSignedCmp())
1935 return false;
1936
1937 ICmpInst::Predicate Pred = Cmp->getPredicate();
1938 if (Pred != ICmpInst::ICMP_EQ)
1939 return false;
1940
1941 // If icmp eq has users other than CondBrInst and SelectInst, converting it to
1942 // icmp slt/sgt would introduce more redundant LLVM IR.
1943 for (User *U : Cmp->users()) {
1944 if (isa<CondBrInst>(Val: U))
1945 continue;
1946 if (isa<SelectInst>(Val: U) && cast<SelectInst>(Val: U)->getCondition() == Cmp)
1947 continue;
1948 return false;
1949 }
1950
1951 // This is a cheap/incomplete check for dominance - just match a single
1952 // predecessor with a conditional branch.
1953 BasicBlock *CmpBB = Cmp->getParent();
1954 BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1955 if (!DomBB)
1956 return false;
1957
1958 // We want to ensure that the only way control gets to the comparison of
1959 // interest is that a less/greater than comparison on the same operands is
1960 // false.
1961 Value *DomCond;
1962 BasicBlock *TrueBB, *FalseBB;
1963 if (!match(V: DomBB->getTerminator(), P: m_Br(C: m_Value(V&: DomCond), T&: TrueBB, F&: FalseBB)))
1964 return false;
1965 if (CmpBB != FalseBB)
1966 return false;
1967
1968 Value *CmpOp0 = Cmp->getOperand(i_nocapture: 0), *CmpOp1 = Cmp->getOperand(i_nocapture: 1);
1969 CmpPredicate DomPred;
1970 if (!match(V: DomCond, P: m_ICmp(Pred&: DomPred, L: m_Specific(V: CmpOp0), R: m_Specific(V: CmpOp1))))
1971 return false;
1972 if (DomPred != ICmpInst::ICMP_SGT && DomPred != ICmpInst::ICMP_SLT)
1973 return false;
1974
1975 // Convert the equality comparison to the opposite of the dominating
1976 // comparison and swap the direction for all branch/select users.
1977 // We have conceptually converted:
1978 // Res = (a < b) ? <LT_RES> : (a == b) ? <EQ_RES> : <GT_RES>;
1979 // to
1980 // Res = (a < b) ? <LT_RES> : (a > b) ? <GT_RES> : <EQ_RES>;
1981 // And similarly for branches.
1982 for (User *U : Cmp->users()) {
1983 if (auto *BI = dyn_cast<CondBrInst>(Val: U)) {
1984 BI->swapSuccessors();
1985 continue;
1986 }
1987 if (auto *SI = dyn_cast<SelectInst>(Val: U)) {
1988 // Swap operands
1989 SI->swapValues();
1990 SI->swapProfMetadata();
1991 continue;
1992 }
1993 llvm_unreachable("Must be a branch or a select");
1994 }
1995 Cmp->setPredicate(CmpInst::getSwappedPredicate(pred: DomPred));
1996 return true;
1997}
1998
1999/// Many architectures use the same instruction for both subtract and cmp. Try
2000/// to swap cmp operands to match subtract operations to allow for CSE.
2001static bool swapICmpOperandsToExposeCSEOpportunities(CmpInst *Cmp) {
2002 Value *Op0 = Cmp->getOperand(i_nocapture: 0);
2003 Value *Op1 = Cmp->getOperand(i_nocapture: 1);
2004 if (!Op0->getType()->isIntegerTy() || isa<Constant>(Val: Op0) ||
2005 isa<Constant>(Val: Op1) || Op0 == Op1)
2006 return false;
2007
2008 // If a subtract already has the same operands as a compare, swapping would be
2009 // bad. If a subtract has the same operands as a compare but in reverse order,
2010 // then swapping is good.
2011 int GoodToSwap = 0;
2012 unsigned NumInspected = 0;
2013 for (const User *U : Op0->users()) {
2014 // Avoid walking many users.
2015 if (++NumInspected > 128)
2016 return false;
2017 if (match(V: U, P: m_Sub(L: m_Specific(V: Op1), R: m_Specific(V: Op0))))
2018 GoodToSwap++;
2019 else if (match(V: U, P: m_Sub(L: m_Specific(V: Op0), R: m_Specific(V: Op1))))
2020 GoodToSwap--;
2021 }
2022
2023 if (GoodToSwap > 0) {
2024 Cmp->swapOperands();
2025 return true;
2026 }
2027 return false;
2028}
2029
2030static bool foldFCmpToFPClassTest(CmpInst *Cmp, const TargetLowering &TLI,
2031 const DataLayout &DL) {
2032 FCmpInst *FCmp = dyn_cast<FCmpInst>(Val: Cmp);
2033 if (!FCmp)
2034 return false;
2035
2036 // Don't fold if the target offers free fabs and the predicate is legal.
2037 EVT VT = TLI.getValueType(DL, Ty: Cmp->getOperand(i_nocapture: 0)->getType());
2038 if (TLI.isFAbsFree(VT) &&
2039 TLI.isCondCodeLegal(CC: getFCmpCondCode(Pred: FCmp->getPredicate()),
2040 VT: VT.getSimpleVT()))
2041 return false;
2042
2043 // Reverse the canonicalization if it is a FP class test
2044 auto ShouldReverseTransform = [](FPClassTest ClassTest) {
2045 return ClassTest == fcInf || ClassTest == (fcInf | fcNan);
2046 };
2047 auto [ClassVal, ClassTest] =
2048 fcmpToClassTest(Pred: FCmp->getPredicate(), F: *FCmp->getParent()->getParent(),
2049 LHS: FCmp->getOperand(i_nocapture: 0), RHS: FCmp->getOperand(i_nocapture: 1));
2050 if (!ClassVal)
2051 return false;
2052
2053 if (!ShouldReverseTransform(ClassTest) && !ShouldReverseTransform(~ClassTest))
2054 return false;
2055
2056 IRBuilder<> Builder(Cmp);
2057 Value *IsFPClass = Builder.createIsFPClass(FPNum: ClassVal, Test: ClassTest);
2058 Cmp->replaceAllUsesWith(V: IsFPClass);
2059 RecursivelyDeleteTriviallyDeadInstructions(V: Cmp);
2060 return true;
2061}
2062
2063static bool isRemOfLoopIncrementWithLoopInvariant(
2064 Instruction *Rem, const LoopInfo *LI, Value *&RemAmtOut, Value *&AddInstOut,
2065 Value *&AddOffsetOut, PHINode *&LoopIncrPNOut) {
2066 Value *Incr, *RemAmt;
2067 // NB: If RemAmt is a power of 2 it *should* have been transformed by now.
2068 if (!match(V: Rem, P: m_URem(L: m_Value(V&: Incr), R: m_Value(V&: RemAmt))))
2069 return false;
2070
2071 Value *AddInst, *AddOffset;
2072 // Find out loop increment PHI.
2073 auto *PN = dyn_cast<PHINode>(Val: Incr);
2074 if (PN != nullptr) {
2075 AddInst = nullptr;
2076 AddOffset = nullptr;
2077 } else {
2078 // Search through a NUW add on top of the loop increment.
2079 Value *V0, *V1;
2080 if (!match(V: Incr, P: m_NUWAdd(L: m_Value(V&: V0), R: m_Value(V&: V1))))
2081 return false;
2082
2083 AddInst = Incr;
2084 PN = dyn_cast<PHINode>(Val: V0);
2085 if (PN != nullptr) {
2086 AddOffset = V1;
2087 } else {
2088 PN = dyn_cast<PHINode>(Val: V1);
2089 AddOffset = V0;
2090 }
2091 }
2092
2093 if (!PN)
2094 return false;
2095
2096 // This isn't strictly necessary, what we really need is one increment and any
2097 // amount of initial values all being the same.
2098 if (PN->getNumIncomingValues() != 2)
2099 return false;
2100
2101 // Only trivially analyzable loops.
2102 Loop *L = LI->getLoopFor(BB: PN->getParent());
2103 if (!L || !L->getLoopPreheader() || !L->getLoopLatch())
2104 return false;
2105
2106 // Req that the remainder is in the loop
2107 if (!L->contains(Inst: Rem))
2108 return false;
2109
2110 // Only works if the remainder amount is a loop invaraint
2111 if (!L->isLoopInvariant(V: RemAmt))
2112 return false;
2113
2114 // Only works if the AddOffset is a loop invaraint
2115 if (AddOffset && !L->isLoopInvariant(V: AddOffset))
2116 return false;
2117
2118 // Is the PHI a loop increment?
2119 auto LoopIncrInfo = getIVIncrement(PN, LI);
2120 if (!LoopIncrInfo)
2121 return false;
2122
2123 // We need remainder_amount % increment_amount to be zero. Increment of one
2124 // satisfies that without any special logic and is overwhelmingly the common
2125 // case.
2126 if (!match(V: LoopIncrInfo->second, P: m_One()))
2127 return false;
2128
2129 // Need the increment to not overflow.
2130 if (!match(V: LoopIncrInfo->first, P: m_c_NUWAdd(L: m_Specific(V: PN), R: m_Value())))
2131 return false;
2132
2133 // Set output variables.
2134 RemAmtOut = RemAmt;
2135 LoopIncrPNOut = PN;
2136 AddInstOut = AddInst;
2137 AddOffsetOut = AddOffset;
2138
2139 return true;
2140}
2141
2142// Try to transform:
2143//
2144// for(i = Start; i < End; ++i)
2145// Rem = (i nuw+ IncrLoopInvariant) u% RemAmtLoopInvariant;
2146//
2147// ->
2148//
2149// Rem = (Start nuw+ IncrLoopInvariant) % RemAmtLoopInvariant;
2150// for(i = Start; i < End; ++i, ++rem)
2151// Rem = rem == RemAmtLoopInvariant ? 0 : Rem;
2152static bool foldURemOfLoopIncrement(Instruction *Rem, const DataLayout *DL,
2153 const LoopInfo *LI,
2154 SmallPtrSet<BasicBlock *, 32> &FreshBBs,
2155 bool IsHuge) {
2156 Value *AddOffset, *RemAmt, *AddInst;
2157 PHINode *LoopIncrPN;
2158 if (!isRemOfLoopIncrementWithLoopInvariant(Rem, LI, RemAmtOut&: RemAmt, AddInstOut&: AddInst,
2159 AddOffsetOut&: AddOffset, LoopIncrPNOut&: LoopIncrPN))
2160 return false;
2161
2162 // Only non-constant remainder as the extra IV is probably not profitable
2163 // in that case.
2164 //
2165 // Potential TODO(1): `urem` of a const ends up as `mul` + `shift` + `add`. If
2166 // we can rule out register pressure and ensure this `urem` is executed each
2167 // iteration, its probably profitable to handle the const case as well.
2168 //
2169 // Potential TODO(2): Should we have a check for how "nested" this remainder
2170 // operation is? The new code runs every iteration so if the remainder is
2171 // guarded behind unlikely conditions this might not be worth it.
2172 if (match(V: RemAmt, P: m_ImmConstant()))
2173 return false;
2174
2175 Loop *L = LI->getLoopFor(BB: LoopIncrPN->getParent());
2176 Value *Start = LoopIncrPN->getIncomingValueForBlock(BB: L->getLoopPreheader());
2177 // If we have add create initial value for remainder.
2178 // The logic here is:
2179 // (urem (add nuw Start, IncrLoopInvariant), RemAmtLoopInvariant
2180 //
2181 // Only proceed if the expression simplifies (otherwise we can't fully
2182 // optimize out the urem).
2183 if (AddInst) {
2184 assert(AddOffset && "We found an add but missing values");
2185 // Without dom-condition/assumption cache we aren't likely to get much out
2186 // of a context instruction.
2187 Start = simplifyAddInst(LHS: Start, RHS: AddOffset,
2188 IsNSW: match(V: AddInst, P: m_NSWAdd(L: m_Value(), R: m_Value())),
2189 /*IsNUW=*/true, Q: *DL);
2190 if (!Start)
2191 return false;
2192 }
2193
2194 // If we can't fully optimize out the `rem`, skip this transform.
2195 Start = simplifyURemInst(LHS: Start, RHS: RemAmt, Q: *DL);
2196 if (!Start)
2197 return false;
2198
2199 // Create new remainder with induction variable.
2200 Type *Ty = Rem->getType();
2201 IRBuilder<> Builder(Rem->getContext());
2202
2203 Builder.SetInsertPoint(LoopIncrPN);
2204 PHINode *NewRem = Builder.CreatePHI(Ty, NumReservedValues: 2);
2205
2206 Builder.SetInsertPoint(cast<Instruction>(
2207 Val: LoopIncrPN->getIncomingValueForBlock(BB: L->getLoopLatch())));
2208 // `(add (urem x, y), 1)` is always nuw.
2209 Value *RemAdd = Builder.CreateNUWAdd(LHS: NewRem, RHS: ConstantInt::get(Ty, V: 1));
2210 Value *RemCmp = Builder.CreateICmp(P: ICmpInst::ICMP_EQ, LHS: RemAdd, RHS: RemAmt);
2211 Value *RemSel =
2212 Builder.CreateSelect(C: RemCmp, True: Constant::getNullValue(Ty), False: RemAdd);
2213
2214 NewRem->addIncoming(V: Start, BB: L->getLoopPreheader());
2215 NewRem->addIncoming(V: RemSel, BB: L->getLoopLatch());
2216
2217 // Insert all touched BBs.
2218 FreshBBs.insert(Ptr: LoopIncrPN->getParent());
2219 FreshBBs.insert(Ptr: L->getLoopLatch());
2220 FreshBBs.insert(Ptr: Rem->getParent());
2221 if (AddInst)
2222 FreshBBs.insert(Ptr: cast<Instruction>(Val: AddInst)->getParent());
2223 replaceAllUsesWith(Old: Rem, New: NewRem, FreshBBs, IsHuge);
2224 Rem->eraseFromParent();
2225 if (AddInst && AddInst->use_empty())
2226 cast<Instruction>(Val: AddInst)->eraseFromParent();
2227 return true;
2228}
2229
2230bool CodeGenPrepare::optimizeURem(Instruction *Rem) {
2231 if (foldURemOfLoopIncrement(Rem, DL, LI, FreshBBs, IsHuge: IsHugeFunc))
2232 return true;
2233 return false;
2234}
2235
2236bool CodeGenPrepare::optimizeCmp(CmpInst *Cmp, ModifyDT &ModifiedDT) {
2237 if (sinkCmpExpression(Cmp, TLI: *TLI, DL: *DL))
2238 return true;
2239
2240 if (combineToUAddWithOverflow(Cmp, ModifiedDT))
2241 return true;
2242
2243 if (combineToUSubWithOverflow(Cmp, ModifiedDT))
2244 return true;
2245
2246 if (unfoldPowerOf2Test(Cmp))
2247 return true;
2248
2249 if (foldICmpWithDominatingICmp(Cmp, TLI: *TLI))
2250 return true;
2251
2252 if (swapICmpOperandsToExposeCSEOpportunities(Cmp))
2253 return true;
2254
2255 if (foldFCmpToFPClassTest(Cmp, TLI: *TLI, DL: *DL))
2256 return true;
2257
2258 return false;
2259}
2260
2261/// Duplicate and sink the given 'and' instruction into user blocks where it is
2262/// used in a compare to allow isel to generate better code for targets where
2263/// this operation can be combined.
2264///
2265/// Return true if any changes are made.
2266static bool sinkAndCmp0Expression(Instruction *AndI, const TargetLowering &TLI,
2267 SetOfInstrs &InsertedInsts) {
2268 // Double-check that we're not trying to optimize an instruction that was
2269 // already optimized by some other part of this pass.
2270 assert(!InsertedInsts.count(AndI) &&
2271 "Attempting to optimize already optimized and instruction");
2272 (void)InsertedInsts;
2273
2274 // Nothing to do for single use in same basic block.
2275 if (AndI->hasOneUse() &&
2276 AndI->getParent() == cast<Instruction>(Val: *AndI->user_begin())->getParent())
2277 return false;
2278
2279 // Try to avoid cases where sinking/duplicating is likely to increase register
2280 // pressure.
2281 if (!isa<ConstantInt>(Val: AndI->getOperand(i: 0)) &&
2282 !isa<ConstantInt>(Val: AndI->getOperand(i: 1)) &&
2283 AndI->getOperand(i: 0)->hasOneUse() && AndI->getOperand(i: 1)->hasOneUse())
2284 return false;
2285
2286 for (auto *U : AndI->users()) {
2287 Instruction *User = cast<Instruction>(Val: U);
2288
2289 // Only sink 'and' feeding icmp with 0.
2290 if (!isa<ICmpInst>(Val: User))
2291 return false;
2292
2293 auto *CmpC = dyn_cast<ConstantInt>(Val: User->getOperand(i: 1));
2294 if (!CmpC || !CmpC->isZero())
2295 return false;
2296 }
2297
2298 if (!TLI.isMaskAndCmp0FoldingBeneficial(AndI: *AndI))
2299 return false;
2300
2301 LLVM_DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n");
2302 LLVM_DEBUG(AndI->getParent()->dump());
2303
2304 // Push the 'and' into the same block as the icmp 0. There should only be
2305 // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
2306 // others, so we don't need to keep track of which BBs we insert into.
2307 for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
2308 UI != E;) {
2309 Use &TheUse = UI.getUse();
2310 Instruction *User = cast<Instruction>(Val: *UI);
2311
2312 // Preincrement use iterator so we don't invalidate it.
2313 ++UI;
2314
2315 LLVM_DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n");
2316
2317 // Keep the 'and' in the same place if the use is already in the same block.
2318 Instruction *InsertPt =
2319 User->getParent() == AndI->getParent() ? AndI : User;
2320 Instruction *InsertedAnd = BinaryOperator::Create(
2321 Op: Instruction::And, S1: AndI->getOperand(i: 0), S2: AndI->getOperand(i: 1), Name: "",
2322 InsertBefore: InsertPt->getIterator());
2323 // Propagate the debug info.
2324 InsertedAnd->setDebugLoc(AndI->getDebugLoc());
2325
2326 // Replace a use of the 'and' with a use of the new 'and'.
2327 TheUse = InsertedAnd;
2328 ++NumAndUses;
2329 LLVM_DEBUG(User->getParent()->dump());
2330 }
2331
2332 // We removed all uses, nuke the and.
2333 AndI->eraseFromParent();
2334 return true;
2335}
2336
2337/// Check if the candidates could be combined with a shift instruction, which
2338/// includes:
2339/// 1. Truncate instruction
2340/// 2. And instruction and the imm is a mask of the low bits:
2341/// imm & (imm+1) == 0
2342static bool isExtractBitsCandidateUse(Instruction *User) {
2343 if (!isa<TruncInst>(Val: User)) {
2344 if (User->getOpcode() != Instruction::And ||
2345 !isa<ConstantInt>(Val: User->getOperand(i: 1)))
2346 return false;
2347
2348 const APInt &Cimm = cast<ConstantInt>(Val: User->getOperand(i: 1))->getValue();
2349
2350 if ((Cimm & (Cimm + 1)).getBoolValue())
2351 return false;
2352 }
2353 return true;
2354}
2355
2356/// Sink both shift and truncate instruction to the use of truncate's BB.
2357static bool
2358SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
2359 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
2360 const TargetLowering &TLI, const DataLayout &DL) {
2361 BasicBlock *UserBB = User->getParent();
2362 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
2363 auto *TruncI = cast<TruncInst>(Val: User);
2364 bool MadeChange = false;
2365
2366 for (Value::user_iterator TruncUI = TruncI->user_begin(),
2367 TruncE = TruncI->user_end();
2368 TruncUI != TruncE;) {
2369
2370 Use &TruncTheUse = TruncUI.getUse();
2371 Instruction *TruncUser = cast<Instruction>(Val: *TruncUI);
2372 // Preincrement use iterator so we don't invalidate it.
2373
2374 ++TruncUI;
2375
2376 int ISDOpcode = TLI.InstructionOpcodeToISD(Opcode: TruncUser->getOpcode());
2377 if (!ISDOpcode)
2378 continue;
2379
2380 // If the use is actually a legal node, there will not be an
2381 // implicit truncate.
2382 // FIXME: always querying the result type is just an
2383 // approximation; some nodes' legality is determined by the
2384 // operand or other means. There's no good way to find out though.
2385 if (TLI.isOperationLegalOrCustom(
2386 Op: ISDOpcode, VT: TLI.getValueType(DL, Ty: TruncUser->getType(), AllowUnknown: true)))
2387 continue;
2388
2389 // Don't bother for PHI nodes.
2390 if (isa<PHINode>(Val: TruncUser))
2391 continue;
2392
2393 BasicBlock *TruncUserBB = TruncUser->getParent();
2394
2395 if (UserBB == TruncUserBB)
2396 continue;
2397
2398 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
2399 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
2400
2401 if (!InsertedShift && !InsertedTrunc) {
2402 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
2403 assert(InsertPt != TruncUserBB->end());
2404 // Sink the shift
2405 if (ShiftI->getOpcode() == Instruction::AShr)
2406 InsertedShift =
2407 BinaryOperator::CreateAShr(V1: ShiftI->getOperand(i_nocapture: 0), V2: CI, Name: "");
2408 else
2409 InsertedShift =
2410 BinaryOperator::CreateLShr(V1: ShiftI->getOperand(i_nocapture: 0), V2: CI, Name: "");
2411 InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
2412 InsertedShift->insertBefore(BB&: *TruncUserBB, InsertPos: InsertPt);
2413
2414 // Sink the trunc
2415 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
2416 TruncInsertPt++;
2417 // It will go ahead of any debug-info.
2418 TruncInsertPt.setHeadBit(true);
2419 assert(TruncInsertPt != TruncUserBB->end());
2420
2421 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), S: InsertedShift,
2422 Ty: TruncI->getType(), Name: "");
2423 InsertedTrunc->insertBefore(BB&: *TruncUserBB, InsertPos: TruncInsertPt);
2424 InsertedTrunc->setDebugLoc(TruncI->getDebugLoc());
2425
2426 MadeChange = true;
2427
2428 TruncTheUse = InsertedTrunc;
2429 }
2430 }
2431 return MadeChange;
2432}
2433
2434/// Sink the shift *right* instruction into user blocks if the uses could
2435/// potentially be combined with this shift instruction and generate BitExtract
2436/// instruction. It will only be applied if the architecture supports BitExtract
2437/// instruction. Here is an example:
2438/// BB1:
2439/// %x.extract.shift = lshr i64 %arg1, 32
2440/// BB2:
2441/// %x.extract.trunc = trunc i64 %x.extract.shift to i16
2442/// ==>
2443///
2444/// BB2:
2445/// %x.extract.shift.1 = lshr i64 %arg1, 32
2446/// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
2447///
2448/// CodeGen will recognize the pattern in BB2 and generate BitExtract
2449/// instruction.
2450/// Return true if any changes are made.
2451static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
2452 const TargetLowering &TLI,
2453 const DataLayout &DL) {
2454 BasicBlock *DefBB = ShiftI->getParent();
2455
2456 /// Only insert instructions in each block once.
2457 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
2458
2459 bool shiftIsLegal = TLI.isTypeLegal(VT: TLI.getValueType(DL, Ty: ShiftI->getType()));
2460
2461 bool MadeChange = false;
2462 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
2463 UI != E;) {
2464 Use &TheUse = UI.getUse();
2465 Instruction *User = cast<Instruction>(Val: *UI);
2466 // Preincrement use iterator so we don't invalidate it.
2467 ++UI;
2468
2469 // Don't bother for PHI nodes.
2470 if (isa<PHINode>(Val: User))
2471 continue;
2472
2473 if (!isExtractBitsCandidateUse(User))
2474 continue;
2475
2476 BasicBlock *UserBB = User->getParent();
2477
2478 if (UserBB == DefBB) {
2479 // If the shift and truncate instruction are in the same BB. The use of
2480 // the truncate(TruncUse) may still introduce another truncate if not
2481 // legal. In this case, we would like to sink both shift and truncate
2482 // instruction to the BB of TruncUse.
2483 // for example:
2484 // BB1:
2485 // i64 shift.result = lshr i64 opnd, imm
2486 // trunc.result = trunc shift.result to i16
2487 //
2488 // BB2:
2489 // ----> We will have an implicit truncate here if the architecture does
2490 // not have i16 compare.
2491 // cmp i16 trunc.result, opnd2
2492 //
2493 if (isa<TruncInst>(Val: User) &&
2494 shiftIsLegal
2495 // If the type of the truncate is legal, no truncate will be
2496 // introduced in other basic blocks.
2497 && (!TLI.isTypeLegal(VT: TLI.getValueType(DL, Ty: User->getType()))))
2498 MadeChange =
2499 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
2500
2501 continue;
2502 }
2503 // If we have already inserted a shift into this block, use it.
2504 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
2505
2506 if (!InsertedShift) {
2507 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2508 assert(InsertPt != UserBB->end());
2509
2510 if (ShiftI->getOpcode() == Instruction::AShr)
2511 InsertedShift =
2512 BinaryOperator::CreateAShr(V1: ShiftI->getOperand(i_nocapture: 0), V2: CI, Name: "");
2513 else
2514 InsertedShift =
2515 BinaryOperator::CreateLShr(V1: ShiftI->getOperand(i_nocapture: 0), V2: CI, Name: "");
2516 InsertedShift->insertBefore(BB&: *UserBB, InsertPos: InsertPt);
2517 InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
2518
2519 MadeChange = true;
2520 }
2521
2522 // Replace a use of the shift with a use of the new shift.
2523 TheUse = InsertedShift;
2524 }
2525
2526 // If we removed all uses, or there are none, nuke the shift.
2527 if (ShiftI->use_empty()) {
2528 salvageDebugInfo(I&: *ShiftI);
2529 ShiftI->eraseFromParent();
2530 MadeChange = true;
2531 }
2532
2533 return MadeChange;
2534}
2535
2536/// If counting leading or trailing zeros is an expensive operation and a zero
2537/// input is defined, add a check for zero to avoid calling the intrinsic.
2538///
2539/// We want to transform:
2540/// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
2541///
2542/// into:
2543/// entry:
2544/// %cmpz = icmp eq i64 %A, 0
2545/// br i1 %cmpz, label %cond.end, label %cond.false
2546/// cond.false:
2547/// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
2548/// br label %cond.end
2549/// cond.end:
2550/// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
2551///
2552/// If the transform is performed, return true and set ModifiedDT to true.
2553static bool despeculateCountZeros(IntrinsicInst *CountZeros, LoopInfo &LI,
2554 const TargetLowering *TLI,
2555 const DataLayout *DL, ModifyDT &ModifiedDT,
2556 SmallPtrSet<BasicBlock *, 32> &FreshBBs,
2557 bool IsHugeFunc) {
2558 // If a zero input is undefined, it doesn't make sense to despeculate that.
2559 if (match(V: CountZeros->getOperand(i_nocapture: 1), P: m_One()))
2560 return false;
2561
2562 // If it's cheap to speculate, there's nothing to do.
2563 Type *Ty = CountZeros->getType();
2564 auto IntrinsicID = CountZeros->getIntrinsicID();
2565 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz(Ty)) ||
2566 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz(Ty)))
2567 return false;
2568
2569 // Only handle scalar cases. Anything else requires too much work.
2570 unsigned SizeInBits = Ty->getScalarSizeInBits();
2571 if (Ty->isVectorTy())
2572 return false;
2573
2574 // Bail if the value is never zero.
2575 Use &Op = CountZeros->getOperandUse(i: 0);
2576 if (isKnownNonZero(V: Op, Q: *DL))
2577 return false;
2578
2579 // The intrinsic will be sunk behind a compare against zero and branch.
2580 BasicBlock *StartBlock = CountZeros->getParent();
2581 BasicBlock *CallBlock = StartBlock->splitBasicBlock(I: CountZeros, BBName: "cond.false");
2582 if (IsHugeFunc)
2583 FreshBBs.insert(Ptr: CallBlock);
2584
2585 // Create another block after the count zero intrinsic. A PHI will be added
2586 // in this block to select the result of the intrinsic or the bit-width
2587 // constant if the input to the intrinsic is zero.
2588 BasicBlock::iterator SplitPt = std::next(x: BasicBlock::iterator(CountZeros));
2589 // Any debug-info after CountZeros should not be included.
2590 SplitPt.setHeadBit(true);
2591 BasicBlock *EndBlock = CallBlock->splitBasicBlock(I: SplitPt, BBName: "cond.end");
2592 if (IsHugeFunc)
2593 FreshBBs.insert(Ptr: EndBlock);
2594
2595 // Update the LoopInfo. The new blocks are in the same loop as the start
2596 // block.
2597 if (Loop *L = LI.getLoopFor(BB: StartBlock)) {
2598 L->addBasicBlockToLoop(NewBB: CallBlock, LI);
2599 L->addBasicBlockToLoop(NewBB: EndBlock, LI);
2600 }
2601
2602 // Set up a builder to create a compare, conditional branch, and PHI.
2603 IRBuilder<> Builder(CountZeros->getContext());
2604 Builder.SetInsertPoint(StartBlock->getTerminator());
2605 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
2606
2607 // Replace the unconditional branch that was created by the first split with
2608 // a compare against zero and a conditional branch.
2609 Value *Zero = Constant::getNullValue(Ty);
2610 // Avoid introducing branch on poison. This also replaces the ctz operand.
2611 if (!isGuaranteedNotToBeUndefOrPoison(V: Op))
2612 Op = Builder.CreateFreeze(V: Op, Name: Op->getName() + ".fr");
2613 Value *Cmp = Builder.CreateICmpEQ(LHS: Op, RHS: Zero, Name: "cmpz");
2614 Builder.CreateCondBr(Cond: Cmp, True: EndBlock, False: CallBlock);
2615 StartBlock->getTerminator()->eraseFromParent();
2616
2617 // Create a PHI in the end block to select either the output of the intrinsic
2618 // or the bit width of the operand.
2619 Builder.SetInsertPoint(TheBB: EndBlock, IP: EndBlock->begin());
2620 PHINode *PN = Builder.CreatePHI(Ty, NumReservedValues: 2, Name: "ctz");
2621 replaceAllUsesWith(Old: CountZeros, New: PN, FreshBBs, IsHuge: IsHugeFunc);
2622 Value *BitWidth = Builder.getInt(AI: APInt(SizeInBits, SizeInBits));
2623 PN->addIncoming(V: BitWidth, BB: StartBlock);
2624 PN->addIncoming(V: CountZeros, BB: CallBlock);
2625
2626 // We are explicitly handling the zero case, so we can set the intrinsic's
2627 // undefined zero argument to 'true'. This will also prevent reprocessing the
2628 // intrinsic; we only despeculate when a zero input is defined.
2629 CountZeros->setArgOperand(i: 1, v: Builder.getTrue());
2630 ModifiedDT = ModifyDT::ModifyBBDT;
2631 return true;
2632}
2633
2634bool CodeGenPrepare::optimizeCallInst(CallInst *CI, ModifyDT &ModifiedDT) {
2635 BasicBlock *BB = CI->getParent();
2636
2637 // Sink address computing for memory operands into the block.
2638 if (CI->isInlineAsm() && optimizeInlineAsmInst(CS: CI))
2639 return true;
2640
2641 // Align the pointer arguments to this call if the target thinks it's a good
2642 // idea
2643 unsigned MinSize;
2644 Align PrefAlign;
2645 if (TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
2646 for (auto &Arg : CI->args()) {
2647 // We want to align both objects whose address is used directly and
2648 // objects whose address is used in casts and GEPs, though it only makes
2649 // sense for GEPs if the offset is a multiple of the desired alignment and
2650 // if size - offset meets the size threshold.
2651 if (!Arg->getType()->isPointerTy())
2652 continue;
2653 APInt Offset(DL->getIndexSizeInBits(
2654 AS: cast<PointerType>(Val: Arg->getType())->getAddressSpace()),
2655 0);
2656 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(DL: *DL, Offset);
2657 uint64_t Offset2 = Offset.getLimitedValue();
2658 if (!isAligned(Lhs: PrefAlign, SizeInBytes: Offset2))
2659 continue;
2660 AllocaInst *AI;
2661 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlign() < PrefAlign) {
2662 std::optional<TypeSize> AllocaSize = AI->getAllocationSize(DL: *DL);
2663 if (AllocaSize && AllocaSize->getKnownMinValue() >= MinSize + Offset2)
2664 AI->setAlignment(PrefAlign);
2665 }
2666 // Global variables can only be aligned if they are defined in this
2667 // object (i.e. they are uniquely initialized in this object), and
2668 // over-aligning global variables that have an explicit section is
2669 // forbidden.
2670 GlobalVariable *GV;
2671 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
2672 GV->getPointerAlignment(DL: *DL) < PrefAlign &&
2673 GV->getGlobalSize(DL: *DL) >= MinSize + Offset2)
2674 GV->setAlignment(PrefAlign);
2675 }
2676 }
2677 // If this is a memcpy (or similar) then we may be able to improve the
2678 // alignment.
2679 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Val: CI)) {
2680 Align DestAlign = getKnownAlignment(V: MI->getDest(), DL: *DL);
2681 MaybeAlign MIDestAlign = MI->getDestAlign();
2682 if (!MIDestAlign || DestAlign > *MIDestAlign)
2683 MI->setDestAlignment(DestAlign);
2684 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(Val: MI)) {
2685 MaybeAlign MTISrcAlign = MTI->getSourceAlign();
2686 Align SrcAlign = getKnownAlignment(V: MTI->getSource(), DL: *DL);
2687 if (!MTISrcAlign || SrcAlign > *MTISrcAlign)
2688 MTI->setSourceAlignment(SrcAlign);
2689 }
2690 }
2691
2692 // If we have a cold call site, try to sink addressing computation into the
2693 // cold block. This interacts with our handling for loads and stores to
2694 // ensure that we can fold all uses of a potential addressing computation
2695 // into their uses. TODO: generalize this to work over profiling data
2696 if (CI->hasFnAttr(Kind: Attribute::Cold) &&
2697 !llvm::shouldOptimizeForSize(BB, PSI, BFI))
2698 for (auto &Arg : CI->args()) {
2699 if (!Arg->getType()->isPointerTy())
2700 continue;
2701 unsigned AS = Arg->getType()->getPointerAddressSpace();
2702 if (optimizeMemoryInst(MemoryInst: CI, Addr: Arg, AccessTy: Arg->getType(), AddrSpace: AS))
2703 return true;
2704 }
2705
2706 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: CI);
2707 if (II) {
2708 switch (II->getIntrinsicID()) {
2709 default:
2710 break;
2711 case Intrinsic::assume:
2712 llvm_unreachable("llvm.assume should have been removed already");
2713 case Intrinsic::allow_runtime_check:
2714 case Intrinsic::allow_ubsan_check:
2715 case Intrinsic::experimental_widenable_condition: {
2716 // Give up on future widening opportunities so that we can fold away dead
2717 // paths and merge blocks before going into block-local instruction
2718 // selection.
2719 if (II->use_empty()) {
2720 II->eraseFromParent();
2721 return true;
2722 }
2723 Constant *RetVal = ConstantInt::getTrue(Context&: II->getContext());
2724 resetIteratorIfInvalidatedWhileCalling(BB, f: [&]() {
2725 replaceAndRecursivelySimplify(I: CI, SimpleV: RetVal, TLI: TLInfo, DT: nullptr);
2726 });
2727 return true;
2728 }
2729 case Intrinsic::objectsize:
2730 llvm_unreachable("llvm.objectsize.* should have been lowered already");
2731 case Intrinsic::is_constant:
2732 llvm_unreachable("llvm.is.constant.* should have been lowered already");
2733 case Intrinsic::aarch64_stlxr:
2734 case Intrinsic::aarch64_stxr: {
2735 ZExtInst *ExtVal = dyn_cast<ZExtInst>(Val: CI->getArgOperand(i: 0));
2736 if (!ExtVal || !ExtVal->hasOneUse() ||
2737 ExtVal->getParent() == CI->getParent())
2738 return false;
2739 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
2740 ExtVal->moveBefore(InsertPos: CI->getIterator());
2741 // Mark this instruction as "inserted by CGP", so that other
2742 // optimizations don't touch it.
2743 InsertedInsts.insert(Ptr: ExtVal);
2744 return true;
2745 }
2746
2747 case Intrinsic::launder_invariant_group:
2748 case Intrinsic::strip_invariant_group: {
2749 Value *ArgVal = II->getArgOperand(i: 0);
2750 auto it = LargeOffsetGEPMap.find(Key: II);
2751 if (it != LargeOffsetGEPMap.end()) {
2752 // Merge entries in LargeOffsetGEPMap to reflect the RAUW.
2753 // Make sure not to have to deal with iterator invalidation
2754 // after possibly adding ArgVal to LargeOffsetGEPMap.
2755 auto GEPs = std::move(it->second);
2756 LargeOffsetGEPMap[ArgVal].append(in_start: GEPs.begin(), in_end: GEPs.end());
2757 LargeOffsetGEPMap.erase(Key: II);
2758 }
2759
2760 replaceAllUsesWith(Old: II, New: ArgVal, FreshBBs, IsHuge: IsHugeFunc);
2761 II->eraseFromParent();
2762 return true;
2763 }
2764 case Intrinsic::cttz:
2765 case Intrinsic::ctlz:
2766 // If counting zeros is expensive, try to avoid it.
2767 return despeculateCountZeros(CountZeros: II, LI&: *LI, TLI, DL, ModifiedDT, FreshBBs,
2768 IsHugeFunc);
2769 case Intrinsic::fshl:
2770 case Intrinsic::fshr:
2771 return optimizeFunnelShift(Fsh: II);
2772 case Intrinsic::masked_gather:
2773 return optimizeGatherScatterInst(MemoryInst: II, Ptr: II->getArgOperand(i: 0));
2774 case Intrinsic::masked_scatter:
2775 return optimizeGatherScatterInst(MemoryInst: II, Ptr: II->getArgOperand(i: 1));
2776 case Intrinsic::masked_load:
2777 // Treat v1X masked load as load X type.
2778 if (auto *VT = dyn_cast<FixedVectorType>(Val: II->getType())) {
2779 if (VT->getNumElements() == 1) {
2780 Value *PtrVal = II->getArgOperand(i: 0);
2781 unsigned AS = PtrVal->getType()->getPointerAddressSpace();
2782 if (optimizeMemoryInst(MemoryInst: II, Addr: PtrVal, AccessTy: VT->getElementType(), AddrSpace: AS))
2783 return true;
2784 }
2785 }
2786 return false;
2787 case Intrinsic::masked_store:
2788 // Treat v1X masked store as store X type.
2789 if (auto *VT =
2790 dyn_cast<FixedVectorType>(Val: II->getArgOperand(i: 0)->getType())) {
2791 if (VT->getNumElements() == 1) {
2792 Value *PtrVal = II->getArgOperand(i: 1);
2793 unsigned AS = PtrVal->getType()->getPointerAddressSpace();
2794 if (optimizeMemoryInst(MemoryInst: II, Addr: PtrVal, AccessTy: VT->getElementType(), AddrSpace: AS))
2795 return true;
2796 }
2797 }
2798 return false;
2799 case Intrinsic::umul_with_overflow:
2800 return optimizeMulWithOverflow(I: II, /*IsSigned=*/false, ModifiedDT);
2801 case Intrinsic::smul_with_overflow:
2802 return optimizeMulWithOverflow(I: II, /*IsSigned=*/true, ModifiedDT);
2803 }
2804
2805 SmallVector<Value *, 2> PtrOps;
2806 Type *AccessTy;
2807 if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
2808 while (!PtrOps.empty()) {
2809 Value *PtrVal = PtrOps.pop_back_val();
2810 unsigned AS = PtrVal->getType()->getPointerAddressSpace();
2811 if (optimizeMemoryInst(MemoryInst: II, Addr: PtrVal, AccessTy, AddrSpace: AS))
2812 return true;
2813 }
2814 }
2815
2816 // From here on out we're working with named functions.
2817 auto *Callee = CI->getCalledFunction();
2818 if (!Callee)
2819 return false;
2820
2821 // Lower all default uses of _chk calls. This is very similar
2822 // to what InstCombineCalls does, but here we are only lowering calls
2823 // to fortified library functions (e.g. __memcpy_chk) that have the default
2824 // "don't know" as the objectsize. Anything else should be left alone.
2825 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
2826 IRBuilder<> Builder(CI);
2827 if (Value *V = Simplifier.optimizeCall(CI, B&: Builder)) {
2828 replaceAllUsesWith(Old: CI, New: V, FreshBBs, IsHuge: IsHugeFunc);
2829 CI->eraseFromParent();
2830 return true;
2831 }
2832
2833 // SCCP may have propagated, among other things, C++ static variables across
2834 // calls. If this happens to be the case, we may want to undo it in order to
2835 // avoid redundant pointer computation of the constant, as the function method
2836 // returning the constant needs to be executed anyways.
2837 auto GetUniformReturnValue = [](const Function *F) -> GlobalVariable * {
2838 if (!F->getReturnType()->isPointerTy())
2839 return nullptr;
2840
2841 GlobalVariable *UniformValue = nullptr;
2842 for (auto &BB : *F) {
2843 if (auto *RI = dyn_cast<ReturnInst>(Val: BB.getTerminator())) {
2844 if (auto *V = dyn_cast<GlobalVariable>(Val: RI->getReturnValue())) {
2845 if (!UniformValue)
2846 UniformValue = V;
2847 else if (V != UniformValue)
2848 return nullptr;
2849 } else {
2850 return nullptr;
2851 }
2852 }
2853 }
2854
2855 return UniformValue;
2856 };
2857
2858 if (Callee->hasExactDefinition()) {
2859 if (GlobalVariable *RV = GetUniformReturnValue(Callee)) {
2860 bool MadeChange = false;
2861 for (Use &U : make_early_inc_range(Range: RV->uses())) {
2862 auto *I = dyn_cast<Instruction>(Val: U.getUser());
2863 if (!I || I->getParent() != CI->getParent()) {
2864 // Limit to the same basic block to avoid extending the call-site live
2865 // range, which otherwise could increase register pressure.
2866 continue;
2867 }
2868 if (CI->comesBefore(Other: I)) {
2869 U.set(CI);
2870 MadeChange = true;
2871 }
2872 }
2873
2874 return MadeChange;
2875 }
2876 }
2877
2878 return false;
2879}
2880
2881static bool isIntrinsicOrLFToBeTailCalled(const TargetLibraryInfo *TLInfo,
2882 const CallInst *CI) {
2883 assert(CI && CI->use_empty());
2884
2885 if (const auto *II = dyn_cast<IntrinsicInst>(Val: CI))
2886 switch (II->getIntrinsicID()) {
2887 case Intrinsic::memset:
2888 case Intrinsic::memcpy:
2889 case Intrinsic::memmove:
2890 return true;
2891 default:
2892 return false;
2893 }
2894
2895 LibFunc LF;
2896 Function *Callee = CI->getCalledFunction();
2897 if (Callee && TLInfo && TLInfo->getLibFunc(FDecl: *Callee, F&: LF))
2898 switch (LF) {
2899 case LibFunc_strcpy:
2900 case LibFunc_strncpy:
2901 case LibFunc_strcat:
2902 case LibFunc_strncat:
2903 return true;
2904 default:
2905 return false;
2906 }
2907
2908 return false;
2909}
2910
2911/// Look for opportunities to duplicate return instructions to the predecessor
2912/// to enable tail call optimizations. The case it is currently looking for is
2913/// the following one. Known intrinsics or library function that may be tail
2914/// called are taken into account as well.
2915/// @code
2916/// bb0:
2917/// %tmp0 = tail call i32 @f0()
2918/// br label %return
2919/// bb1:
2920/// %tmp1 = tail call i32 @f1()
2921/// br label %return
2922/// bb2:
2923/// %tmp2 = tail call i32 @f2()
2924/// br label %return
2925/// return:
2926/// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
2927/// ret i32 %retval
2928/// @endcode
2929///
2930/// =>
2931///
2932/// @code
2933/// bb0:
2934/// %tmp0 = tail call i32 @f0()
2935/// ret i32 %tmp0
2936/// bb1:
2937/// %tmp1 = tail call i32 @f1()
2938/// ret i32 %tmp1
2939/// bb2:
2940/// %tmp2 = tail call i32 @f2()
2941/// ret i32 %tmp2
2942/// @endcode
2943bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB,
2944 ModifyDT &ModifiedDT) {
2945 if (!BB->getTerminator())
2946 return false;
2947
2948 ReturnInst *RetI = dyn_cast<ReturnInst>(Val: BB->getTerminator());
2949 if (!RetI)
2950 return false;
2951
2952 assert(LI->getLoopFor(BB) == nullptr && "A return block cannot be in a loop");
2953
2954 PHINode *PN = nullptr;
2955 ExtractValueInst *EVI = nullptr;
2956 BitCastInst *BCI = nullptr;
2957 Value *V = RetI->getReturnValue();
2958 if (V) {
2959 BCI = dyn_cast<BitCastInst>(Val: V);
2960 if (BCI)
2961 V = BCI->getOperand(i_nocapture: 0);
2962
2963 EVI = dyn_cast<ExtractValueInst>(Val: V);
2964 if (EVI) {
2965 V = EVI->getOperand(i_nocapture: 0);
2966 if (!llvm::all_of(Range: EVI->indices(), P: equal_to(Arg: 0)))
2967 return false;
2968 }
2969
2970 PN = dyn_cast<PHINode>(Val: V);
2971 }
2972
2973 if (PN && PN->getParent() != BB)
2974 return false;
2975
2976 auto isLifetimeEndOrBitCastFor = [](const Instruction *Inst) {
2977 const BitCastInst *BC = dyn_cast<BitCastInst>(Val: Inst);
2978 if (BC && BC->hasOneUse())
2979 Inst = BC->user_back();
2980
2981 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: Inst))
2982 return II->getIntrinsicID() == Intrinsic::lifetime_end;
2983 return false;
2984 };
2985
2986 SmallVector<const IntrinsicInst *, 4> FakeUses;
2987
2988 auto isFakeUse = [&FakeUses](const Instruction *Inst) {
2989 if (auto *II = dyn_cast<IntrinsicInst>(Val: Inst);
2990 II && II->getIntrinsicID() == Intrinsic::fake_use) {
2991 // Record the instruction so it can be preserved when the exit block is
2992 // removed. Do not preserve the fake use that uses the result of the
2993 // PHI instruction.
2994 // Do not copy fake uses that use the result of a PHI node.
2995 // FIXME: If we do want to copy the fake use into the return blocks, we
2996 // have to figure out which of the PHI node operands to use for each
2997 // copy.
2998 if (!isa<PHINode>(Val: II->getOperand(i_nocapture: 0))) {
2999 FakeUses.push_back(Elt: II);
3000 }
3001 return true;
3002 }
3003
3004 return false;
3005 };
3006
3007 // Make sure there are no instructions between the first instruction
3008 // and return.
3009 BasicBlock::const_iterator BI = BB->getFirstNonPHIIt();
3010 // Skip over pseudo-probes and the bitcast.
3011 while (&*BI == BCI || &*BI == EVI || isa<PseudoProbeInst>(Val: BI) ||
3012 isLifetimeEndOrBitCastFor(&*BI) || isFakeUse(&*BI))
3013 BI = std::next(x: BI);
3014 if (&*BI != RetI)
3015 return false;
3016
3017 // Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
3018 // call.
3019 auto MayBePermittedAsTailCall = [&](const auto *CI) {
3020 return TLI->mayBeEmittedAsTailCall(CI) &&
3021 attributesPermitTailCall(BB->getParent(), CI, RetI, *TLI);
3022 };
3023
3024 SmallVector<BasicBlock *, 4> TailCallBBs;
3025 // Record the call instructions so we can insert any fake uses
3026 // that need to be preserved before them.
3027 SmallVector<CallInst *, 4> CallInsts;
3028 if (PN) {
3029 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
3030 // Look through bitcasts.
3031 Value *IncomingVal = PN->getIncomingValue(i: I)->stripPointerCasts();
3032 CallInst *CI = dyn_cast<CallInst>(Val: IncomingVal);
3033 BasicBlock *PredBB = PN->getIncomingBlock(i: I);
3034 // Make sure the phi value is indeed produced by the tail call.
3035 if (CI && CI->hasOneUse() && CI->getParent() == PredBB &&
3036 MayBePermittedAsTailCall(CI)) {
3037 TailCallBBs.push_back(Elt: PredBB);
3038 CallInsts.push_back(Elt: CI);
3039 } else {
3040 // Consider the cases in which the phi value is indirectly produced by
3041 // the tail call, for example when encountering memset(), memmove(),
3042 // strcpy(), whose return value may have been optimized out. In such
3043 // cases, the value needs to be the first function argument.
3044 //
3045 // bb0:
3046 // tail call void @llvm.memset.p0.i64(ptr %0, i8 0, i64 %1)
3047 // br label %return
3048 // return:
3049 // %phi = phi ptr [ %0, %bb0 ], [ %2, %entry ]
3050 if (PredBB && PredBB->getSingleSuccessor() == BB)
3051 CI = dyn_cast_or_null<CallInst>(
3052 Val: PredBB->getTerminator()->getPrevNode());
3053
3054 if (CI && CI->use_empty() &&
3055 isIntrinsicOrLFToBeTailCalled(TLInfo, CI) &&
3056 IncomingVal == CI->getArgOperand(i: 0) &&
3057 MayBePermittedAsTailCall(CI)) {
3058 TailCallBBs.push_back(Elt: PredBB);
3059 CallInsts.push_back(Elt: CI);
3060 }
3061 }
3062 }
3063 } else {
3064 SmallPtrSet<BasicBlock *, 4> VisitedBBs;
3065 for (BasicBlock *Pred : predecessors(BB)) {
3066 if (!VisitedBBs.insert(Ptr: Pred).second)
3067 continue;
3068 if (Instruction *I = Pred->rbegin()->getPrevNode()) {
3069 CallInst *CI = dyn_cast<CallInst>(Val: I);
3070 if (CI && CI->use_empty() && MayBePermittedAsTailCall(CI)) {
3071 // Either we return void or the return value must be the first
3072 // argument of a known intrinsic or library function.
3073 if (!V || isa<UndefValue>(Val: V) ||
3074 (isIntrinsicOrLFToBeTailCalled(TLInfo, CI) &&
3075 V == CI->getArgOperand(i: 0))) {
3076 TailCallBBs.push_back(Elt: Pred);
3077 CallInsts.push_back(Elt: CI);
3078 }
3079 }
3080 }
3081 }
3082 }
3083
3084 bool Changed = false;
3085 for (auto const &TailCallBB : TailCallBBs) {
3086 // Make sure the call instruction is followed by an unconditional branch to
3087 // the return block.
3088 UncondBrInst *BI = dyn_cast<UncondBrInst>(Val: TailCallBB->getTerminator());
3089 if (!BI || BI->getSuccessor() != BB)
3090 continue;
3091
3092 // Duplicate the return into TailCallBB.
3093 (void)FoldReturnIntoUncondBranch(RI: RetI, BB, Pred: TailCallBB);
3094 assert(!VerifyBFIUpdates ||
3095 BFI->getBlockFreq(BB) >= BFI->getBlockFreq(TailCallBB));
3096 BFI->setBlockFreq(BB,
3097 Freq: (BFI->getBlockFreq(BB) - BFI->getBlockFreq(BB: TailCallBB)));
3098 ModifiedDT = ModifyDT::ModifyBBDT;
3099 Changed = true;
3100 ++NumRetsDup;
3101 }
3102
3103 // If we eliminated all predecessors of the block, delete the block now.
3104 if (Changed && !BB->hasAddressTaken() && pred_empty(BB)) {
3105 // Copy the fake uses found in the original return block to all blocks
3106 // that contain tail calls.
3107 for (auto *CI : CallInsts) {
3108 for (auto const *FakeUse : FakeUses) {
3109 auto *ClonedInst = FakeUse->clone();
3110 ClonedInst->insertBefore(InsertPos: CI->getIterator());
3111 }
3112 }
3113 BB->eraseFromParent();
3114 }
3115
3116 return Changed;
3117}
3118
3119//===----------------------------------------------------------------------===//
3120// Memory Optimization
3121//===----------------------------------------------------------------------===//
3122
3123namespace {
3124
3125/// This is an extended version of TargetLowering::AddrMode
3126/// which holds actual Value*'s for register values.
3127struct ExtAddrMode : public TargetLowering::AddrMode {
3128 Value *BaseReg = nullptr;
3129 Value *ScaledReg = nullptr;
3130 Value *OriginalValue = nullptr;
3131 bool InBounds = true;
3132
3133 enum FieldName {
3134 NoField = 0x00,
3135 BaseRegField = 0x01,
3136 BaseGVField = 0x02,
3137 BaseOffsField = 0x04,
3138 ScaledRegField = 0x08,
3139 ScaleField = 0x10,
3140 MultipleFields = 0xff
3141 };
3142
3143 ExtAddrMode() = default;
3144
3145 void print(raw_ostream &OS) const;
3146 void dump() const;
3147
3148 // Replace From in ExtAddrMode with To.
3149 // E.g., SExt insts may be promoted and deleted. We should replace them with
3150 // the promoted values.
3151 void replaceWith(Value *From, Value *To) {
3152 if (ScaledReg == From)
3153 ScaledReg = To;
3154 }
3155
3156 FieldName compare(const ExtAddrMode &other) {
3157 // First check that the types are the same on each field, as differing types
3158 // is something we can't cope with later on.
3159 if (BaseReg && other.BaseReg &&
3160 BaseReg->getType() != other.BaseReg->getType())
3161 return MultipleFields;
3162 if (BaseGV && other.BaseGV && BaseGV->getType() != other.BaseGV->getType())
3163 return MultipleFields;
3164 if (ScaledReg && other.ScaledReg &&
3165 ScaledReg->getType() != other.ScaledReg->getType())
3166 return MultipleFields;
3167
3168 // Conservatively reject 'inbounds' mismatches.
3169 if (InBounds != other.InBounds)
3170 return MultipleFields;
3171
3172 // Check each field to see if it differs.
3173 unsigned Result = NoField;
3174 if (BaseReg != other.BaseReg)
3175 Result |= BaseRegField;
3176 if (BaseGV != other.BaseGV)
3177 Result |= BaseGVField;
3178 if (BaseOffs != other.BaseOffs)
3179 Result |= BaseOffsField;
3180 if (ScaledReg != other.ScaledReg)
3181 Result |= ScaledRegField;
3182 // Don't count 0 as being a different scale, because that actually means
3183 // unscaled (which will already be counted by having no ScaledReg).
3184 if (Scale && other.Scale && Scale != other.Scale)
3185 Result |= ScaleField;
3186
3187 if (llvm::popcount(Value: Result) > 1)
3188 return MultipleFields;
3189 else
3190 return static_cast<FieldName>(Result);
3191 }
3192
3193 // An AddrMode is trivial if it involves no calculation i.e. it is just a base
3194 // with no offset.
3195 bool isTrivial() {
3196 // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
3197 // trivial if at most one of these terms is nonzero, except that BaseGV and
3198 // BaseReg both being zero actually means a null pointer value, which we
3199 // consider to be 'non-zero' here.
3200 return !BaseOffs && !Scale && !(BaseGV && BaseReg);
3201 }
3202
3203 Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
3204 switch (Field) {
3205 default:
3206 return nullptr;
3207 case BaseRegField:
3208 return BaseReg;
3209 case BaseGVField:
3210 return BaseGV;
3211 case ScaledRegField:
3212 return ScaledReg;
3213 case BaseOffsField:
3214 return ConstantInt::getSigned(Ty: IntPtrTy, V: BaseOffs);
3215 }
3216 }
3217
3218 void SetCombinedField(FieldName Field, Value *V,
3219 const SmallVectorImpl<ExtAddrMode> &AddrModes) {
3220 switch (Field) {
3221 default:
3222 llvm_unreachable("Unhandled fields are expected to be rejected earlier");
3223 break;
3224 case ExtAddrMode::BaseRegField:
3225 BaseReg = V;
3226 break;
3227 case ExtAddrMode::BaseGVField:
3228 // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
3229 // in the BaseReg field.
3230 assert(BaseReg == nullptr);
3231 BaseReg = V;
3232 BaseGV = nullptr;
3233 break;
3234 case ExtAddrMode::ScaledRegField:
3235 ScaledReg = V;
3236 // If we have a mix of scaled and unscaled addrmodes then we want scale
3237 // to be the scale and not zero.
3238 if (!Scale)
3239 for (const ExtAddrMode &AM : AddrModes)
3240 if (AM.Scale) {
3241 Scale = AM.Scale;
3242 break;
3243 }
3244 break;
3245 case ExtAddrMode::BaseOffsField:
3246 // The offset is no longer a constant, so it goes in ScaledReg with a
3247 // scale of 1.
3248 assert(ScaledReg == nullptr);
3249 ScaledReg = V;
3250 Scale = 1;
3251 BaseOffs = 0;
3252 break;
3253 }
3254 }
3255};
3256
3257#ifndef NDEBUG
3258static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
3259 AM.print(OS);
3260 return OS;
3261}
3262#endif
3263
3264#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3265void ExtAddrMode::print(raw_ostream &OS) const {
3266 bool NeedPlus = false;
3267 OS << "[";
3268 if (InBounds)
3269 OS << "inbounds ";
3270 if (BaseGV) {
3271 OS << "GV:";
3272 BaseGV->printAsOperand(OS, /*PrintType=*/false);
3273 NeedPlus = true;
3274 }
3275
3276 if (BaseOffs) {
3277 OS << (NeedPlus ? " + " : "") << BaseOffs;
3278 NeedPlus = true;
3279 }
3280
3281 if (BaseReg) {
3282 OS << (NeedPlus ? " + " : "") << "Base:";
3283 BaseReg->printAsOperand(OS, /*PrintType=*/false);
3284 NeedPlus = true;
3285 }
3286 if (Scale) {
3287 OS << (NeedPlus ? " + " : "") << Scale << "*";
3288 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
3289 }
3290
3291 OS << ']';
3292}
3293
3294LLVM_DUMP_METHOD void ExtAddrMode::dump() const {
3295 print(dbgs());
3296 dbgs() << '\n';
3297}
3298#endif
3299
3300} // end anonymous namespace
3301
3302namespace {
3303
3304/// This class provides transaction based operation on the IR.
3305/// Every change made through this class is recorded in the internal state and
3306/// can be undone (rollback) until commit is called.
3307/// CGP does not check if instructions could be speculatively executed when
3308/// moved. Preserving the original location would pessimize the debugging
3309/// experience, as well as negatively impact the quality of sample PGO.
3310class TypePromotionTransaction {
3311 /// This represents the common interface of the individual transaction.
3312 /// Each class implements the logic for doing one specific modification on
3313 /// the IR via the TypePromotionTransaction.
3314 class TypePromotionAction {
3315 protected:
3316 /// The Instruction modified.
3317 Instruction *Inst;
3318
3319 public:
3320 /// Constructor of the action.
3321 /// The constructor performs the related action on the IR.
3322 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
3323
3324 virtual ~TypePromotionAction() = default;
3325
3326 /// Undo the modification done by this action.
3327 /// When this method is called, the IR must be in the same state as it was
3328 /// before this action was applied.
3329 /// \pre Undoing the action works if and only if the IR is in the exact same
3330 /// state as it was directly after this action was applied.
3331 virtual void undo() = 0;
3332
3333 /// Advocate every change made by this action.
3334 /// When the results on the IR of the action are to be kept, it is important
3335 /// to call this function, otherwise hidden information may be kept forever.
3336 virtual void commit() {
3337 // Nothing to be done, this action is not doing anything.
3338 }
3339 };
3340
3341 /// Utility to remember the position of an instruction.
3342 class InsertionHandler {
3343 /// Position of an instruction.
3344 /// Either an instruction:
3345 /// - Is the first in a basic block: BB is used.
3346 /// - Has a previous instruction: PrevInst is used.
3347 struct {
3348 BasicBlock::iterator PrevInst;
3349 BasicBlock *BB;
3350 } Point;
3351 std::optional<DbgRecord::self_iterator> BeforeDbgRecord = std::nullopt;
3352
3353 /// Remember whether or not the instruction had a previous instruction.
3354 bool HasPrevInstruction;
3355
3356 public:
3357 /// Record the position of \p Inst.
3358 InsertionHandler(Instruction *Inst) {
3359 HasPrevInstruction = (Inst != &*(Inst->getParent()->begin()));
3360 BasicBlock *BB = Inst->getParent();
3361
3362 // Record where we would have to re-insert the instruction in the sequence
3363 // of DbgRecords, if we ended up reinserting.
3364 BeforeDbgRecord = Inst->getDbgReinsertionPosition();
3365
3366 if (HasPrevInstruction) {
3367 Point.PrevInst = std::prev(x: Inst->getIterator());
3368 } else {
3369 Point.BB = BB;
3370 }
3371 }
3372
3373 /// Insert \p Inst at the recorded position.
3374 void insert(Instruction *Inst) {
3375 if (HasPrevInstruction) {
3376 if (Inst->getParent())
3377 Inst->removeFromParent();
3378 Inst->insertAfter(InsertPos: Point.PrevInst);
3379 } else {
3380 BasicBlock::iterator Position = Point.BB->getFirstInsertionPt();
3381 if (Inst->getParent())
3382 Inst->moveBefore(BB&: *Point.BB, I: Position);
3383 else
3384 Inst->insertBefore(BB&: *Point.BB, InsertPos: Position);
3385 }
3386
3387 Inst->getParent()->reinsertInstInDbgRecords(I: Inst, Pos: BeforeDbgRecord);
3388 }
3389 };
3390
3391 /// Move an instruction before another.
3392 class InstructionMoveBefore : public TypePromotionAction {
3393 /// Original position of the instruction.
3394 InsertionHandler Position;
3395
3396 public:
3397 /// Move \p Inst before \p Before.
3398 InstructionMoveBefore(Instruction *Inst, BasicBlock::iterator Before)
3399 : TypePromotionAction(Inst), Position(Inst) {
3400 LLVM_DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before
3401 << "\n");
3402 Inst->moveBefore(InsertPos: Before);
3403 }
3404
3405 /// Move the instruction back to its original position.
3406 void undo() override {
3407 LLVM_DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
3408 Position.insert(Inst);
3409 }
3410 };
3411
3412 /// Set the operand of an instruction with a new value.
3413 class OperandSetter : public TypePromotionAction {
3414 /// Original operand of the instruction.
3415 Value *Origin;
3416
3417 /// Index of the modified instruction.
3418 unsigned Idx;
3419
3420 public:
3421 /// Set \p Idx operand of \p Inst with \p NewVal.
3422 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
3423 : TypePromotionAction(Inst), Idx(Idx) {
3424 LLVM_DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
3425 << "for:" << *Inst << "\n"
3426 << "with:" << *NewVal << "\n");
3427 Origin = Inst->getOperand(i: Idx);
3428 Inst->setOperand(i: Idx, Val: NewVal);
3429 }
3430
3431 /// Restore the original value of the instruction.
3432 void undo() override {
3433 LLVM_DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
3434 << "for: " << *Inst << "\n"
3435 << "with: " << *Origin << "\n");
3436 Inst->setOperand(i: Idx, Val: Origin);
3437 }
3438 };
3439
3440 /// Hide the operands of an instruction.
3441 /// Do as if this instruction was not using any of its operands.
3442 class OperandsHider : public TypePromotionAction {
3443 /// The list of original operands.
3444 SmallVector<Value *, 4> OriginalValues;
3445
3446 public:
3447 /// Remove \p Inst from the uses of the operands of \p Inst.
3448 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
3449 LLVM_DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
3450 unsigned NumOpnds = Inst->getNumOperands();
3451 OriginalValues.reserve(N: NumOpnds);
3452 for (unsigned It = 0; It < NumOpnds; ++It) {
3453 // Save the current operand.
3454 Value *Val = Inst->getOperand(i: It);
3455 OriginalValues.push_back(Elt: Val);
3456 // Set a dummy one.
3457 // We could use OperandSetter here, but that would imply an overhead
3458 // that we are not willing to pay.
3459 Inst->setOperand(i: It, Val: PoisonValue::get(T: Val->getType()));
3460 }
3461 }
3462
3463 /// Restore the original list of uses.
3464 void undo() override {
3465 LLVM_DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
3466 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
3467 Inst->setOperand(i: It, Val: OriginalValues[It]);
3468 }
3469 };
3470
3471 /// Build a truncate instruction.
3472 class TruncBuilder : public TypePromotionAction {
3473 Value *Val;
3474
3475 public:
3476 /// Build a truncate instruction of \p Opnd producing a \p Ty
3477 /// result.
3478 /// trunc Opnd to Ty.
3479 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
3480 IRBuilder<> Builder(Opnd);
3481 Builder.SetCurrentDebugLocation(DebugLoc());
3482 Val = Builder.CreateTrunc(V: Opnd, DestTy: Ty, Name: "promoted");
3483 LLVM_DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
3484 }
3485
3486 /// Get the built value.
3487 Value *getBuiltValue() { return Val; }
3488
3489 /// Remove the built instruction.
3490 void undo() override {
3491 LLVM_DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
3492 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3493 IVal->eraseFromParent();
3494 }
3495 };
3496
3497 /// Build a sign extension instruction.
3498 class SExtBuilder : public TypePromotionAction {
3499 Value *Val;
3500
3501 public:
3502 /// Build a sign extension instruction of \p Opnd producing a \p Ty
3503 /// result.
3504 /// sext Opnd to Ty.
3505 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3506 : TypePromotionAction(InsertPt) {
3507 IRBuilder<> Builder(InsertPt);
3508 Val = Builder.CreateSExt(V: Opnd, DestTy: Ty, Name: "promoted");
3509 LLVM_DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
3510 }
3511
3512 /// Get the built value.
3513 Value *getBuiltValue() { return Val; }
3514
3515 /// Remove the built instruction.
3516 void undo() override {
3517 LLVM_DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
3518 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3519 IVal->eraseFromParent();
3520 }
3521 };
3522
3523 /// Build a zero extension instruction.
3524 class ZExtBuilder : public TypePromotionAction {
3525 Value *Val;
3526
3527 public:
3528 /// Build a zero extension instruction of \p Opnd producing a \p Ty
3529 /// result.
3530 /// zext Opnd to Ty.
3531 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3532 : TypePromotionAction(InsertPt) {
3533 IRBuilder<> Builder(InsertPt);
3534 Builder.SetCurrentDebugLocation(DebugLoc());
3535 Val = Builder.CreateZExt(V: Opnd, DestTy: Ty, Name: "promoted");
3536 LLVM_DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
3537 }
3538
3539 /// Get the built value.
3540 Value *getBuiltValue() { return Val; }
3541
3542 /// Remove the built instruction.
3543 void undo() override {
3544 LLVM_DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
3545 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3546 IVal->eraseFromParent();
3547 }
3548 };
3549
3550 /// Mutate an instruction to another type.
3551 class TypeMutator : public TypePromotionAction {
3552 /// Record the original type.
3553 Type *OrigTy;
3554
3555 public:
3556 /// Mutate the type of \p Inst into \p NewTy.
3557 TypeMutator(Instruction *Inst, Type *NewTy)
3558 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
3559 LLVM_DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
3560 << "\n");
3561 Inst->mutateType(Ty: NewTy);
3562 }
3563
3564 /// Mutate the instruction back to its original type.
3565 void undo() override {
3566 LLVM_DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
3567 << "\n");
3568 Inst->mutateType(Ty: OrigTy);
3569 }
3570 };
3571
3572 /// Replace the uses of an instruction by another instruction.
3573 class UsesReplacer : public TypePromotionAction {
3574 /// Helper structure to keep track of the replaced uses.
3575 struct InstructionAndIdx {
3576 /// The instruction using the instruction.
3577 Instruction *Inst;
3578
3579 /// The index where this instruction is used for Inst.
3580 unsigned Idx;
3581
3582 InstructionAndIdx(Instruction *Inst, unsigned Idx)
3583 : Inst(Inst), Idx(Idx) {}
3584 };
3585
3586 /// Keep track of the original uses (pair Instruction, Index).
3587 SmallVector<InstructionAndIdx, 4> OriginalUses;
3588 /// Keep track of the debug users.
3589 SmallVector<DbgVariableRecord *, 1> DbgVariableRecords;
3590
3591 /// Keep track of the new value so that we can undo it by replacing
3592 /// instances of the new value with the original value.
3593 Value *New;
3594
3595 using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
3596
3597 public:
3598 /// Replace all the use of \p Inst by \p New.
3599 UsesReplacer(Instruction *Inst, Value *New)
3600 : TypePromotionAction(Inst), New(New) {
3601 LLVM_DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
3602 << "\n");
3603 // Record the original uses.
3604 for (Use &U : Inst->uses()) {
3605 Instruction *UserI = cast<Instruction>(Val: U.getUser());
3606 OriginalUses.push_back(Elt: InstructionAndIdx(UserI, U.getOperandNo()));
3607 }
3608 // Record the debug uses separately. They are not in the instruction's
3609 // use list, but they are replaced by RAUW.
3610 findDbgValues(V: Inst, DbgVariableRecords);
3611
3612 // Now, we can replace the uses.
3613 Inst->replaceAllUsesWith(V: New);
3614 }
3615
3616 /// Reassign the original uses of Inst to Inst.
3617 void undo() override {
3618 LLVM_DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
3619 for (InstructionAndIdx &Use : OriginalUses)
3620 Use.Inst->setOperand(i: Use.Idx, Val: Inst);
3621 // RAUW has replaced all original uses with references to the new value,
3622 // including the debug uses. Since we are undoing the replacements,
3623 // the original debug uses must also be reinstated to maintain the
3624 // correctness and utility of debug value records.
3625 for (DbgVariableRecord *DVR : DbgVariableRecords)
3626 DVR->replaceVariableLocationOp(OldValue: New, NewValue: Inst);
3627 }
3628 };
3629
3630 /// Remove an instruction from the IR.
3631 class InstructionRemover : public TypePromotionAction {
3632 /// Original position of the instruction.
3633 InsertionHandler Inserter;
3634
3635 /// Helper structure to hide all the link to the instruction. In other
3636 /// words, this helps to do as if the instruction was removed.
3637 OperandsHider Hider;
3638
3639 /// Keep track of the uses replaced, if any.
3640 UsesReplacer *Replacer = nullptr;
3641
3642 /// Keep track of instructions removed.
3643 SetOfInstrs &RemovedInsts;
3644
3645 public:
3646 /// Remove all reference of \p Inst and optionally replace all its
3647 /// uses with New.
3648 /// \p RemovedInsts Keep track of the instructions removed by this Action.
3649 /// \pre If !Inst->use_empty(), then New != nullptr
3650 InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
3651 Value *New = nullptr)
3652 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
3653 RemovedInsts(RemovedInsts) {
3654 if (New)
3655 Replacer = new UsesReplacer(Inst, New);
3656 LLVM_DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
3657 RemovedInsts.insert(Ptr: Inst);
3658 /// The instructions removed here will be freed after completing
3659 /// optimizeBlock() for all blocks as we need to keep track of the
3660 /// removed instructions during promotion.
3661 Inst->removeFromParent();
3662 }
3663
3664 ~InstructionRemover() override { delete Replacer; }
3665
3666 InstructionRemover &operator=(const InstructionRemover &other) = delete;
3667 InstructionRemover(const InstructionRemover &other) = delete;
3668
3669 /// Resurrect the instruction and reassign it to the proper uses if
3670 /// new value was provided when build this action.
3671 void undo() override {
3672 LLVM_DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
3673 Inserter.insert(Inst);
3674 if (Replacer)
3675 Replacer->undo();
3676 Hider.undo();
3677 RemovedInsts.erase(Ptr: Inst);
3678 }
3679 };
3680
3681public:
3682 /// Restoration point.
3683 /// The restoration point is a pointer to an action instead of an iterator
3684 /// because the iterator may be invalidated but not the pointer.
3685 using ConstRestorationPt = const TypePromotionAction *;
3686
3687 TypePromotionTransaction(SetOfInstrs &RemovedInsts)
3688 : RemovedInsts(RemovedInsts) {}
3689
3690 /// Advocate every changes made in that transaction. Return true if any change
3691 /// happen.
3692 bool commit();
3693
3694 /// Undo all the changes made after the given point.
3695 void rollback(ConstRestorationPt Point);
3696
3697 /// Get the current restoration point.
3698 ConstRestorationPt getRestorationPoint() const;
3699
3700 /// \name API for IR modification with state keeping to support rollback.
3701 /// @{
3702 /// Same as Instruction::setOperand.
3703 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
3704
3705 /// Same as Instruction::eraseFromParent.
3706 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
3707
3708 /// Same as Value::replaceAllUsesWith.
3709 void replaceAllUsesWith(Instruction *Inst, Value *New);
3710
3711 /// Same as Value::mutateType.
3712 void mutateType(Instruction *Inst, Type *NewTy);
3713
3714 /// Same as IRBuilder::createTrunc.
3715 Value *createTrunc(Instruction *Opnd, Type *Ty);
3716
3717 /// Same as IRBuilder::createSExt.
3718 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
3719
3720 /// Same as IRBuilder::createZExt.
3721 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
3722
3723private:
3724 /// The ordered list of actions made so far.
3725 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
3726
3727 using CommitPt =
3728 SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
3729
3730 SetOfInstrs &RemovedInsts;
3731};
3732
3733} // end anonymous namespace
3734
3735void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
3736 Value *NewVal) {
3737 Actions.push_back(Elt: std::make_unique<TypePromotionTransaction::OperandSetter>(
3738 args&: Inst, args&: Idx, args&: NewVal));
3739}
3740
3741void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
3742 Value *NewVal) {
3743 Actions.push_back(
3744 Elt: std::make_unique<TypePromotionTransaction::InstructionRemover>(
3745 args&: Inst, args&: RemovedInsts, args&: NewVal));
3746}
3747
3748void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
3749 Value *New) {
3750 Actions.push_back(
3751 Elt: std::make_unique<TypePromotionTransaction::UsesReplacer>(args&: Inst, args&: New));
3752}
3753
3754void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
3755 Actions.push_back(
3756 Elt: std::make_unique<TypePromotionTransaction::TypeMutator>(args&: Inst, args&: NewTy));
3757}
3758
3759Value *TypePromotionTransaction::createTrunc(Instruction *Opnd, Type *Ty) {
3760 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
3761 Value *Val = Ptr->getBuiltValue();
3762 Actions.push_back(Elt: std::move(Ptr));
3763 return Val;
3764}
3765
3766Value *TypePromotionTransaction::createSExt(Instruction *Inst, Value *Opnd,
3767 Type *Ty) {
3768 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
3769 Value *Val = Ptr->getBuiltValue();
3770 Actions.push_back(Elt: std::move(Ptr));
3771 return Val;
3772}
3773
3774Value *TypePromotionTransaction::createZExt(Instruction *Inst, Value *Opnd,
3775 Type *Ty) {
3776 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
3777 Value *Val = Ptr->getBuiltValue();
3778 Actions.push_back(Elt: std::move(Ptr));
3779 return Val;
3780}
3781
3782TypePromotionTransaction::ConstRestorationPt
3783TypePromotionTransaction::getRestorationPoint() const {
3784 return !Actions.empty() ? Actions.back().get() : nullptr;
3785}
3786
3787bool TypePromotionTransaction::commit() {
3788 for (std::unique_ptr<TypePromotionAction> &Action : Actions)
3789 Action->commit();
3790 bool Modified = !Actions.empty();
3791 Actions.clear();
3792 return Modified;
3793}
3794
3795void TypePromotionTransaction::rollback(
3796 TypePromotionTransaction::ConstRestorationPt Point) {
3797 while (!Actions.empty() && Point != Actions.back().get()) {
3798 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
3799 Curr->undo();
3800 }
3801}
3802
3803namespace {
3804
3805/// A helper class for matching addressing modes.
3806///
3807/// This encapsulates the logic for matching the target-legal addressing modes.
3808class AddressingModeMatcher {
3809 SmallVectorImpl<Instruction *> &AddrModeInsts;
3810 const TargetLowering &TLI;
3811 const TargetRegisterInfo &TRI;
3812 const DataLayout &DL;
3813 const LoopInfo &LI;
3814 const std::function<const DominatorTree &()> getDTFn;
3815
3816 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
3817 /// the memory instruction that we're computing this address for.
3818 Type *AccessTy;
3819 unsigned AddrSpace;
3820 Instruction *MemoryInst;
3821
3822 /// This is the addressing mode that we're building up. This is
3823 /// part of the return value of this addressing mode matching stuff.
3824 ExtAddrMode &AddrMode;
3825
3826 /// The instructions inserted by other CodeGenPrepare optimizations.
3827 const SetOfInstrs &InsertedInsts;
3828
3829 /// A map from the instructions to their type before promotion.
3830 InstrToOrigTy &PromotedInsts;
3831
3832 /// The ongoing transaction where every action should be registered.
3833 TypePromotionTransaction &TPT;
3834
3835 // A GEP which has too large offset to be folded into the addressing mode.
3836 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP;
3837
3838 /// This is set to true when we should not do profitability checks.
3839 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
3840 bool IgnoreProfitability;
3841
3842 /// True if we are optimizing for size.
3843 bool OptSize = false;
3844
3845 ProfileSummaryInfo *PSI;
3846 BlockFrequencyInfo *BFI;
3847
3848 AddressingModeMatcher(
3849 SmallVectorImpl<Instruction *> &AMI, const TargetLowering &TLI,
3850 const TargetRegisterInfo &TRI, const LoopInfo &LI,
3851 const std::function<const DominatorTree &()> getDTFn, Type *AT,
3852 unsigned AS, Instruction *MI, ExtAddrMode &AM,
3853 const SetOfInstrs &InsertedInsts, InstrToOrigTy &PromotedInsts,
3854 TypePromotionTransaction &TPT,
3855 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP,
3856 bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI)
3857 : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
3858 DL(MI->getDataLayout()), LI(LI), getDTFn(getDTFn),
3859 AccessTy(AT), AddrSpace(AS), MemoryInst(MI), AddrMode(AM),
3860 InsertedInsts(InsertedInsts), PromotedInsts(PromotedInsts), TPT(TPT),
3861 LargeOffsetGEP(LargeOffsetGEP), OptSize(OptSize), PSI(PSI), BFI(BFI) {
3862 IgnoreProfitability = false;
3863 }
3864
3865public:
3866 /// Find the maximal addressing mode that a load/store of V can fold,
3867 /// give an access type of AccessTy. This returns a list of involved
3868 /// instructions in AddrModeInsts.
3869 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
3870 /// optimizations.
3871 /// \p PromotedInsts maps the instructions to their type before promotion.
3872 /// \p The ongoing transaction where every action should be registered.
3873 static ExtAddrMode
3874 Match(Value *V, Type *AccessTy, unsigned AS, Instruction *MemoryInst,
3875 SmallVectorImpl<Instruction *> &AddrModeInsts,
3876 const TargetLowering &TLI, const LoopInfo &LI,
3877 const std::function<const DominatorTree &()> getDTFn,
3878 const TargetRegisterInfo &TRI, const SetOfInstrs &InsertedInsts,
3879 InstrToOrigTy &PromotedInsts, TypePromotionTransaction &TPT,
3880 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP,
3881 bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) {
3882 ExtAddrMode Result;
3883
3884 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI, LI, getDTFn,
3885 AccessTy, AS, MemoryInst, Result,
3886 InsertedInsts, PromotedInsts, TPT,
3887 LargeOffsetGEP, OptSize, PSI, BFI)
3888 .matchAddr(Addr: V, Depth: 0);
3889 (void)Success;
3890 assert(Success && "Couldn't select *anything*?");
3891 return Result;
3892 }
3893
3894private:
3895 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
3896 bool matchAddr(Value *Addr, unsigned Depth);
3897 bool matchOperationAddr(User *AddrInst, unsigned Opcode, unsigned Depth,
3898 bool *MovedAway = nullptr);
3899 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
3900 ExtAddrMode &AMBefore,
3901 ExtAddrMode &AMAfter);
3902 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
3903 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
3904 Value *PromotedOperand) const;
3905};
3906
3907class PhiNodeSet;
3908
3909/// An iterator for PhiNodeSet.
3910class PhiNodeSetIterator {
3911 PhiNodeSet *const Set;
3912 size_t CurrentIndex = 0;
3913
3914public:
3915 /// The constructor. Start should point to either a valid element, or be equal
3916 /// to the size of the underlying SmallVector of the PhiNodeSet.
3917 PhiNodeSetIterator(PhiNodeSet *const Set, size_t Start);
3918 PHINode *operator*() const;
3919 PhiNodeSetIterator &operator++();
3920 bool operator==(const PhiNodeSetIterator &RHS) const;
3921 bool operator!=(const PhiNodeSetIterator &RHS) const;
3922};
3923
3924/// Keeps a set of PHINodes.
3925///
3926/// This is a minimal set implementation for a specific use case:
3927/// It is very fast when there are very few elements, but also provides good
3928/// performance when there are many. It is similar to SmallPtrSet, but also
3929/// provides iteration by insertion order, which is deterministic and stable
3930/// across runs. It is also similar to SmallSetVector, but provides removing
3931/// elements in O(1) time. This is achieved by not actually removing the element
3932/// from the underlying vector, so comes at the cost of using more memory, but
3933/// that is fine, since PhiNodeSets are used as short lived objects.
3934class PhiNodeSet {
3935 friend class PhiNodeSetIterator;
3936
3937 using MapType = SmallDenseMap<PHINode *, size_t, 32>;
3938 using iterator = PhiNodeSetIterator;
3939
3940 /// Keeps the elements in the order of their insertion in the underlying
3941 /// vector. To achieve constant time removal, it never deletes any element.
3942 SmallVector<PHINode *, 32> NodeList;
3943
3944 /// Keeps the elements in the underlying set implementation. This (and not the
3945 /// NodeList defined above) is the source of truth on whether an element
3946 /// is actually in the collection.
3947 MapType NodeMap;
3948
3949 /// Points to the first valid (not deleted) element when the set is not empty
3950 /// and the value is not zero. Equals to the size of the underlying vector
3951 /// when the set is empty. When the value is 0, as in the beginning, the
3952 /// first element may or may not be valid.
3953 size_t FirstValidElement = 0;
3954
3955public:
3956 /// Inserts a new element to the collection.
3957 /// \returns true if the element is actually added, i.e. was not in the
3958 /// collection before the operation.
3959 bool insert(PHINode *Ptr) {
3960 if (NodeMap.insert(KV: std::make_pair(x&: Ptr, y: NodeList.size())).second) {
3961 NodeList.push_back(Elt: Ptr);
3962 return true;
3963 }
3964 return false;
3965 }
3966
3967 /// Removes the element from the collection.
3968 /// \returns whether the element is actually removed, i.e. was in the
3969 /// collection before the operation.
3970 bool erase(PHINode *Ptr) {
3971 if (NodeMap.erase(Val: Ptr)) {
3972 SkipRemovedElements(CurrentIndex&: FirstValidElement);
3973 return true;
3974 }
3975 return false;
3976 }
3977
3978 /// Removes all elements and clears the collection.
3979 void clear() {
3980 NodeMap.clear();
3981 NodeList.clear();
3982 FirstValidElement = 0;
3983 }
3984
3985 /// \returns an iterator that will iterate the elements in the order of
3986 /// insertion.
3987 iterator begin() {
3988 if (FirstValidElement == 0)
3989 SkipRemovedElements(CurrentIndex&: FirstValidElement);
3990 return PhiNodeSetIterator(this, FirstValidElement);
3991 }
3992
3993 /// \returns an iterator that points to the end of the collection.
3994 iterator end() { return PhiNodeSetIterator(this, NodeList.size()); }
3995
3996 /// Returns the number of elements in the collection.
3997 size_t size() const { return NodeMap.size(); }
3998
3999 /// \returns 1 if the given element is in the collection, and 0 if otherwise.
4000 size_t count(PHINode *Ptr) const { return NodeMap.count(Val: Ptr); }
4001
4002private:
4003 /// Updates the CurrentIndex so that it will point to a valid element.
4004 ///
4005 /// If the element of NodeList at CurrentIndex is valid, it does not
4006 /// change it. If there are no more valid elements, it updates CurrentIndex
4007 /// to point to the end of the NodeList.
4008 void SkipRemovedElements(size_t &CurrentIndex) {
4009 while (CurrentIndex < NodeList.size()) {
4010 auto it = NodeMap.find(Val: NodeList[CurrentIndex]);
4011 // If the element has been deleted and added again later, NodeMap will
4012 // point to a different index, so CurrentIndex will still be invalid.
4013 if (it != NodeMap.end() && it->second == CurrentIndex)
4014 break;
4015 ++CurrentIndex;
4016 }
4017 }
4018};
4019
4020PhiNodeSetIterator::PhiNodeSetIterator(PhiNodeSet *const Set, size_t Start)
4021 : Set(Set), CurrentIndex(Start) {}
4022
4023PHINode *PhiNodeSetIterator::operator*() const {
4024 assert(CurrentIndex < Set->NodeList.size() &&
4025 "PhiNodeSet access out of range");
4026 return Set->NodeList[CurrentIndex];
4027}
4028
4029PhiNodeSetIterator &PhiNodeSetIterator::operator++() {
4030 assert(CurrentIndex < Set->NodeList.size() &&
4031 "PhiNodeSet access out of range");
4032 ++CurrentIndex;
4033 Set->SkipRemovedElements(CurrentIndex);
4034 return *this;
4035}
4036
4037bool PhiNodeSetIterator::operator==(const PhiNodeSetIterator &RHS) const {
4038 return CurrentIndex == RHS.CurrentIndex;
4039}
4040
4041bool PhiNodeSetIterator::operator!=(const PhiNodeSetIterator &RHS) const {
4042 return !((*this) == RHS);
4043}
4044
4045/// Keep track of simplification of Phi nodes.
4046/// Accept the set of all phi nodes and erase phi node from this set
4047/// if it is simplified.
4048class SimplificationTracker {
4049 DenseMap<Value *, Value *> Storage;
4050 // Tracks newly created Phi nodes. The elements are iterated by insertion
4051 // order.
4052 PhiNodeSet AllPhiNodes;
4053 // Tracks newly created Select nodes.
4054 SmallPtrSet<SelectInst *, 32> AllSelectNodes;
4055
4056public:
4057 Value *Get(Value *V) {
4058 do {
4059 auto SV = Storage.find(Val: V);
4060 if (SV == Storage.end())
4061 return V;
4062 V = SV->second;
4063 } while (true);
4064 }
4065
4066 void Put(Value *From, Value *To) { Storage.insert(KV: {From, To}); }
4067
4068 void ReplacePhi(PHINode *From, PHINode *To) {
4069 Value *OldReplacement = Get(V: From);
4070 while (OldReplacement != From) {
4071 From = To;
4072 To = dyn_cast<PHINode>(Val: OldReplacement);
4073 OldReplacement = Get(V: From);
4074 }
4075 assert(To && Get(To) == To && "Replacement PHI node is already replaced.");
4076 Put(From, To);
4077 From->replaceAllUsesWith(V: To);
4078 AllPhiNodes.erase(Ptr: From);
4079 From->eraseFromParent();
4080 }
4081
4082 PhiNodeSet &newPhiNodes() { return AllPhiNodes; }
4083
4084 void insertNewPhi(PHINode *PN) { AllPhiNodes.insert(Ptr: PN); }
4085
4086 void insertNewSelect(SelectInst *SI) { AllSelectNodes.insert(Ptr: SI); }
4087
4088 unsigned countNewPhiNodes() const { return AllPhiNodes.size(); }
4089
4090 unsigned countNewSelectNodes() const { return AllSelectNodes.size(); }
4091
4092 void destroyNewNodes(Type *CommonType) {
4093 // For safe erasing, replace the uses with dummy value first.
4094 auto *Dummy = PoisonValue::get(T: CommonType);
4095 for (auto *I : AllPhiNodes) {
4096 I->replaceAllUsesWith(V: Dummy);
4097 I->eraseFromParent();
4098 }
4099 AllPhiNodes.clear();
4100 for (auto *I : AllSelectNodes) {
4101 I->replaceAllUsesWith(V: Dummy);
4102 I->eraseFromParent();
4103 }
4104 AllSelectNodes.clear();
4105 }
4106};
4107
4108/// A helper class for combining addressing modes.
4109class AddressingModeCombiner {
4110 typedef DenseMap<Value *, Value *> FoldAddrToValueMapping;
4111 typedef std::pair<PHINode *, PHINode *> PHIPair;
4112
4113private:
4114 /// The addressing modes we've collected.
4115 SmallVector<ExtAddrMode, 16> AddrModes;
4116
4117 /// The field in which the AddrModes differ, when we have more than one.
4118 ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
4119
4120 /// Are the AddrModes that we have all just equal to their original values?
4121 bool AllAddrModesTrivial = true;
4122
4123 /// Common Type for all different fields in addressing modes.
4124 Type *CommonType = nullptr;
4125
4126 const DataLayout &DL;
4127
4128 /// Original Address.
4129 Value *Original;
4130
4131 /// Common value among addresses
4132 Value *CommonValue = nullptr;
4133
4134public:
4135 AddressingModeCombiner(const DataLayout &DL, Value *OriginalValue)
4136 : DL(DL), Original(OriginalValue) {}
4137
4138 ~AddressingModeCombiner() { eraseCommonValueIfDead(); }
4139
4140 /// Get the combined AddrMode
4141 const ExtAddrMode &getAddrMode() const { return AddrModes[0]; }
4142
4143 /// Add a new AddrMode if it's compatible with the AddrModes we already
4144 /// have.
4145 /// \return True iff we succeeded in doing so.
4146 bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
4147 // Take note of if we have any non-trivial AddrModes, as we need to detect
4148 // when all AddrModes are trivial as then we would introduce a phi or select
4149 // which just duplicates what's already there.
4150 AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
4151
4152 // If this is the first addrmode then everything is fine.
4153 if (AddrModes.empty()) {
4154 AddrModes.emplace_back(Args&: NewAddrMode);
4155 return true;
4156 }
4157
4158 // Figure out how different this is from the other address modes, which we
4159 // can do just by comparing against the first one given that we only care
4160 // about the cumulative difference.
4161 ExtAddrMode::FieldName ThisDifferentField =
4162 AddrModes[0].compare(other: NewAddrMode);
4163 if (DifferentField == ExtAddrMode::NoField)
4164 DifferentField = ThisDifferentField;
4165 else if (DifferentField != ThisDifferentField)
4166 DifferentField = ExtAddrMode::MultipleFields;
4167
4168 // If NewAddrMode differs in more than one dimension we cannot handle it.
4169 bool CanHandle = DifferentField != ExtAddrMode::MultipleFields;
4170
4171 // If Scale Field is different then we reject.
4172 CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField;
4173
4174 // We also must reject the case when base offset is different and
4175 // scale reg is not null, we cannot handle this case due to merge of
4176 // different offsets will be used as ScaleReg.
4177 CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField ||
4178 !NewAddrMode.ScaledReg);
4179
4180 // We also must reject the case when GV is different and BaseReg installed
4181 // due to we want to use base reg as a merge of GV values.
4182 CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField ||
4183 !NewAddrMode.HasBaseReg);
4184
4185 // Even if NewAddMode is the same we still need to collect it due to
4186 // original value is different. And later we will need all original values
4187 // as anchors during finding the common Phi node.
4188 if (CanHandle)
4189 AddrModes.emplace_back(Args&: NewAddrMode);
4190 else
4191 AddrModes.clear();
4192
4193 return CanHandle;
4194 }
4195
4196 /// Combine the addressing modes we've collected into a single
4197 /// addressing mode.
4198 /// \return True iff we successfully combined them or we only had one so
4199 /// didn't need to combine them anyway.
4200 bool combineAddrModes() {
4201 // If we have no AddrModes then they can't be combined.
4202 if (AddrModes.size() == 0)
4203 return false;
4204
4205 // A single AddrMode can trivially be combined.
4206 if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
4207 return true;
4208
4209 // If the AddrModes we collected are all just equal to the value they are
4210 // derived from then combining them wouldn't do anything useful.
4211 if (AllAddrModesTrivial)
4212 return false;
4213
4214 if (!addrModeCombiningAllowed())
4215 return false;
4216
4217 // Build a map between <original value, basic block where we saw it> to
4218 // value of base register.
4219 // Bail out if there is no common type.
4220 FoldAddrToValueMapping Map;
4221 if (!initializeMap(Map))
4222 return false;
4223
4224 CommonValue = findCommon(Map);
4225 if (CommonValue)
4226 AddrModes[0].SetCombinedField(Field: DifferentField, V: CommonValue, AddrModes);
4227 return CommonValue != nullptr;
4228 }
4229
4230private:
4231 /// `CommonValue` may be a placeholder inserted by us.
4232 /// If the placeholder is not used, we should remove this dead instruction.
4233 void eraseCommonValueIfDead() {
4234 if (CommonValue && CommonValue->use_empty())
4235 if (Instruction *CommonInst = dyn_cast<Instruction>(Val: CommonValue))
4236 CommonInst->eraseFromParent();
4237 }
4238
4239 /// Initialize Map with anchor values. For address seen
4240 /// we set the value of different field saw in this address.
4241 /// At the same time we find a common type for different field we will
4242 /// use to create new Phi/Select nodes. Keep it in CommonType field.
4243 /// Return false if there is no common type found.
4244 bool initializeMap(FoldAddrToValueMapping &Map) {
4245 // Keep track of keys where the value is null. We will need to replace it
4246 // with constant null when we know the common type.
4247 SmallVector<Value *, 2> NullValue;
4248 Type *IntPtrTy = DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
4249 for (auto &AM : AddrModes) {
4250 Value *DV = AM.GetFieldAsValue(Field: DifferentField, IntPtrTy);
4251 if (DV) {
4252 auto *Type = DV->getType();
4253 if (CommonType && CommonType != Type)
4254 return false;
4255 CommonType = Type;
4256 Map[AM.OriginalValue] = DV;
4257 } else {
4258 NullValue.push_back(Elt: AM.OriginalValue);
4259 }
4260 }
4261 assert(CommonType && "At least one non-null value must be!");
4262 for (auto *V : NullValue)
4263 Map[V] = Constant::getNullValue(Ty: CommonType);
4264 return true;
4265 }
4266
4267 /// We have mapping between value A and other value B where B was a field in
4268 /// addressing mode represented by A. Also we have an original value C
4269 /// representing an address we start with. Traversing from C through phi and
4270 /// selects we ended up with A's in a map. This utility function tries to find
4271 /// a value V which is a field in addressing mode C and traversing through phi
4272 /// nodes and selects we will end up in corresponded values B in a map.
4273 /// The utility will create a new Phi/Selects if needed.
4274 // The simple example looks as follows:
4275 // BB1:
4276 // p1 = b1 + 40
4277 // br cond BB2, BB3
4278 // BB2:
4279 // p2 = b2 + 40
4280 // br BB3
4281 // BB3:
4282 // p = phi [p1, BB1], [p2, BB2]
4283 // v = load p
4284 // Map is
4285 // p1 -> b1
4286 // p2 -> b2
4287 // Request is
4288 // p -> ?
4289 // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3.
4290 Value *findCommon(FoldAddrToValueMapping &Map) {
4291 // Tracks the simplification of newly created phi nodes. The reason we use
4292 // this mapping is because we will add new created Phi nodes in AddrToBase.
4293 // Simplification of Phi nodes is recursive, so some Phi node may
4294 // be simplified after we added it to AddrToBase. In reality this
4295 // simplification is possible only if original phi/selects were not
4296 // simplified yet.
4297 // Using this mapping we can find the current value in AddrToBase.
4298 SimplificationTracker ST;
4299
4300 // First step, DFS to create PHI nodes for all intermediate blocks.
4301 // Also fill traverse order for the second step.
4302 SmallVector<Value *, 32> TraverseOrder;
4303 InsertPlaceholders(Map, TraverseOrder, ST);
4304
4305 // Second Step, fill new nodes by merged values and simplify if possible.
4306 FillPlaceholders(Map, TraverseOrder, ST);
4307
4308 if (!AddrSinkNewSelects && ST.countNewSelectNodes() > 0) {
4309 ST.destroyNewNodes(CommonType);
4310 return nullptr;
4311 }
4312
4313 // Now we'd like to match New Phi nodes to existed ones.
4314 unsigned PhiNotMatchedCount = 0;
4315 if (!MatchPhiSet(ST, AllowNewPhiNodes: AddrSinkNewPhis, PhiNotMatchedCount)) {
4316 ST.destroyNewNodes(CommonType);
4317 return nullptr;
4318 }
4319
4320 auto *Result = ST.Get(V: Map.find(Val: Original)->second);
4321 if (Result) {
4322 NumMemoryInstsPhiCreated += ST.countNewPhiNodes() + PhiNotMatchedCount;
4323 NumMemoryInstsSelectCreated += ST.countNewSelectNodes();
4324 }
4325 return Result;
4326 }
4327
4328 /// Try to match PHI node to Candidate.
4329 /// Matcher tracks the matched Phi nodes.
4330 bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
4331 SmallSetVector<PHIPair, 8> &Matcher,
4332 PhiNodeSet &PhiNodesToMatch) {
4333 SmallVector<PHIPair, 8> WorkList;
4334 Matcher.insert(X: {PHI, Candidate});
4335 SmallPtrSet<PHINode *, 8> MatchedPHIs;
4336 MatchedPHIs.insert(Ptr: PHI);
4337 WorkList.push_back(Elt: {PHI, Candidate});
4338 SmallSet<PHIPair, 8> Visited;
4339 while (!WorkList.empty()) {
4340 auto Item = WorkList.pop_back_val();
4341 if (!Visited.insert(V: Item).second)
4342 continue;
4343 // We iterate over all incoming values to Phi to compare them.
4344 // If values are different and both of them Phi and the first one is a
4345 // Phi we added (subject to match) and both of them is in the same basic
4346 // block then we can match our pair if values match. So we state that
4347 // these values match and add it to work list to verify that.
4348 for (auto *B : Item.first->blocks()) {
4349 Value *FirstValue = Item.first->getIncomingValueForBlock(BB: B);
4350 Value *SecondValue = Item.second->getIncomingValueForBlock(BB: B);
4351 if (FirstValue == SecondValue)
4352 continue;
4353
4354 PHINode *FirstPhi = dyn_cast<PHINode>(Val: FirstValue);
4355 PHINode *SecondPhi = dyn_cast<PHINode>(Val: SecondValue);
4356
4357 // One of them is not Phi or
4358 // The first one is not Phi node from the set we'd like to match or
4359 // Phi nodes from different basic blocks then
4360 // we will not be able to match.
4361 if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(Ptr: FirstPhi) ||
4362 FirstPhi->getParent() != SecondPhi->getParent())
4363 return false;
4364
4365 // If we already matched them then continue.
4366 if (Matcher.count(key: {FirstPhi, SecondPhi}))
4367 continue;
4368 // So the values are different and does not match. So we need them to
4369 // match. (But we register no more than one match per PHI node, so that
4370 // we won't later try to replace them twice.)
4371 if (MatchedPHIs.insert(Ptr: FirstPhi).second)
4372 Matcher.insert(X: {FirstPhi, SecondPhi});
4373 // But me must check it.
4374 WorkList.push_back(Elt: {FirstPhi, SecondPhi});
4375 }
4376 }
4377 return true;
4378 }
4379
4380 /// For the given set of PHI nodes (in the SimplificationTracker) try
4381 /// to find their equivalents.
4382 /// Returns false if this matching fails and creation of new Phi is disabled.
4383 bool MatchPhiSet(SimplificationTracker &ST, bool AllowNewPhiNodes,
4384 unsigned &PhiNotMatchedCount) {
4385 // Matched and PhiNodesToMatch iterate their elements in a deterministic
4386 // order, so the replacements (ReplacePhi) are also done in a deterministic
4387 // order.
4388 SmallSetVector<PHIPair, 8> Matched;
4389 SmallPtrSet<PHINode *, 8> WillNotMatch;
4390 PhiNodeSet &PhiNodesToMatch = ST.newPhiNodes();
4391 while (PhiNodesToMatch.size()) {
4392 PHINode *PHI = *PhiNodesToMatch.begin();
4393
4394 // Add us, if no Phi nodes in the basic block we do not match.
4395 WillNotMatch.clear();
4396 WillNotMatch.insert(Ptr: PHI);
4397
4398 // Traverse all Phis until we found equivalent or fail to do that.
4399 bool IsMatched = false;
4400 for (auto &P : PHI->getParent()->phis()) {
4401 // Skip new Phi nodes.
4402 if (PhiNodesToMatch.count(Ptr: &P))
4403 continue;
4404 if ((IsMatched = MatchPhiNode(PHI, Candidate: &P, Matcher&: Matched, PhiNodesToMatch)))
4405 break;
4406 // If it does not match, collect all Phi nodes from matcher.
4407 // if we end up with no match, them all these Phi nodes will not match
4408 // later.
4409 WillNotMatch.insert_range(R: llvm::make_first_range(c&: Matched));
4410 Matched.clear();
4411 }
4412 if (IsMatched) {
4413 // Replace all matched values and erase them.
4414 for (auto MV : Matched)
4415 ST.ReplacePhi(From: MV.first, To: MV.second);
4416 Matched.clear();
4417 continue;
4418 }
4419 // If we are not allowed to create new nodes then bail out.
4420 if (!AllowNewPhiNodes)
4421 return false;
4422 // Just remove all seen values in matcher. They will not match anything.
4423 PhiNotMatchedCount += WillNotMatch.size();
4424 for (auto *P : WillNotMatch)
4425 PhiNodesToMatch.erase(Ptr: P);
4426 }
4427 return true;
4428 }
4429 /// Fill the placeholders with values from predecessors and simplify them.
4430 void FillPlaceholders(FoldAddrToValueMapping &Map,
4431 SmallVectorImpl<Value *> &TraverseOrder,
4432 SimplificationTracker &ST) {
4433 while (!TraverseOrder.empty()) {
4434 Value *Current = TraverseOrder.pop_back_val();
4435 assert(Map.contains(Current) && "No node to fill!!!");
4436 Value *V = Map[Current];
4437
4438 if (SelectInst *Select = dyn_cast<SelectInst>(Val: V)) {
4439 // CurrentValue also must be Select.
4440 auto *CurrentSelect = cast<SelectInst>(Val: Current);
4441 auto *TrueValue = CurrentSelect->getTrueValue();
4442 assert(Map.contains(TrueValue) && "No True Value!");
4443 Select->setTrueValue(ST.Get(V: Map[TrueValue]));
4444 auto *FalseValue = CurrentSelect->getFalseValue();
4445 assert(Map.contains(FalseValue) && "No False Value!");
4446 Select->setFalseValue(ST.Get(V: Map[FalseValue]));
4447 } else {
4448 // Must be a Phi node then.
4449 auto *PHI = cast<PHINode>(Val: V);
4450 // Fill the Phi node with values from predecessors.
4451 for (auto *B : predecessors(BB: PHI->getParent())) {
4452 Value *PV = cast<PHINode>(Val: Current)->getIncomingValueForBlock(BB: B);
4453 assert(Map.contains(PV) && "No predecessor Value!");
4454 PHI->addIncoming(V: ST.Get(V: Map[PV]), BB: B);
4455 }
4456 }
4457 }
4458 }
4459
4460 /// Starting from original value recursively iterates over def-use chain up to
4461 /// known ending values represented in a map. For each traversed phi/select
4462 /// inserts a placeholder Phi or Select.
4463 /// Reports all new created Phi/Select nodes by adding them to set.
4464 /// Also reports and order in what values have been traversed.
4465 void InsertPlaceholders(FoldAddrToValueMapping &Map,
4466 SmallVectorImpl<Value *> &TraverseOrder,
4467 SimplificationTracker &ST) {
4468 SmallVector<Value *, 32> Worklist;
4469 assert((isa<PHINode>(Original) || isa<SelectInst>(Original)) &&
4470 "Address must be a Phi or Select node");
4471 auto *Dummy = PoisonValue::get(T: CommonType);
4472 Worklist.push_back(Elt: Original);
4473 while (!Worklist.empty()) {
4474 Value *Current = Worklist.pop_back_val();
4475 // if it is already visited or it is an ending value then skip it.
4476 if (Map.contains(Val: Current))
4477 continue;
4478 TraverseOrder.push_back(Elt: Current);
4479
4480 // CurrentValue must be a Phi node or select. All others must be covered
4481 // by anchors.
4482 if (SelectInst *CurrentSelect = dyn_cast<SelectInst>(Val: Current)) {
4483 // Is it OK to get metadata from OrigSelect?!
4484 // Create a Select placeholder with dummy value.
4485 SelectInst *Select =
4486 SelectInst::Create(C: CurrentSelect->getCondition(), S1: Dummy, S2: Dummy,
4487 NameStr: CurrentSelect->getName(),
4488 InsertBefore: CurrentSelect->getIterator(), MDFrom: CurrentSelect);
4489 Map[Current] = Select;
4490 ST.insertNewSelect(SI: Select);
4491 // We are interested in True and False values.
4492 Worklist.push_back(Elt: CurrentSelect->getTrueValue());
4493 Worklist.push_back(Elt: CurrentSelect->getFalseValue());
4494 } else {
4495 // It must be a Phi node then.
4496 PHINode *CurrentPhi = cast<PHINode>(Val: Current);
4497 unsigned PredCount = CurrentPhi->getNumIncomingValues();
4498 PHINode *PHI =
4499 PHINode::Create(Ty: CommonType, NumReservedValues: PredCount, NameStr: "sunk_phi", InsertBefore: CurrentPhi->getIterator());
4500 Map[Current] = PHI;
4501 ST.insertNewPhi(PN: PHI);
4502 append_range(C&: Worklist, R: CurrentPhi->incoming_values());
4503 }
4504 }
4505 }
4506
4507 bool addrModeCombiningAllowed() {
4508 if (DisableComplexAddrModes)
4509 return false;
4510 switch (DifferentField) {
4511 default:
4512 return false;
4513 case ExtAddrMode::BaseRegField:
4514 return AddrSinkCombineBaseReg;
4515 case ExtAddrMode::BaseGVField:
4516 return AddrSinkCombineBaseGV;
4517 case ExtAddrMode::BaseOffsField:
4518 return AddrSinkCombineBaseOffs;
4519 case ExtAddrMode::ScaledRegField:
4520 return AddrSinkCombineScaledReg;
4521 }
4522 }
4523};
4524} // end anonymous namespace
4525
4526/// Try adding ScaleReg*Scale to the current addressing mode.
4527/// Return true and update AddrMode if this addr mode is legal for the target,
4528/// false if not.
4529bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
4530 unsigned Depth) {
4531 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
4532 // mode. Just process that directly.
4533 if (Scale == 1)
4534 return matchAddr(Addr: ScaleReg, Depth);
4535
4536 // If the scale is 0, it takes nothing to add this.
4537 if (Scale == 0)
4538 return true;
4539
4540 // If we already have a scale of this value, we can add to it, otherwise, we
4541 // need an available scale field.
4542 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
4543 return false;
4544
4545 ExtAddrMode TestAddrMode = AddrMode;
4546
4547 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
4548 // [A+B + A*7] -> [B+A*8].
4549 TestAddrMode.Scale += Scale;
4550 TestAddrMode.ScaledReg = ScaleReg;
4551
4552 // If the new address isn't legal, bail out.
4553 if (!TLI.isLegalAddressingMode(DL, AM: TestAddrMode, Ty: AccessTy, AddrSpace))
4554 return false;
4555
4556 // It was legal, so commit it.
4557 AddrMode = TestAddrMode;
4558
4559 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
4560 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
4561 // X*Scale + C*Scale to addr mode. If we found available IV increment, do not
4562 // go any further: we can reuse it and cannot eliminate it.
4563 ConstantInt *CI = nullptr;
4564 Value *AddLHS = nullptr;
4565 if (isa<Instruction>(Val: ScaleReg) && // not a constant expr.
4566 match(V: ScaleReg, P: m_Add(L: m_Value(V&: AddLHS), R: m_ConstantInt(CI))) &&
4567 !isIVIncrement(V: ScaleReg, LI: &LI) && CI->getValue().isSignedIntN(N: 64)) {
4568 TestAddrMode.InBounds = false;
4569 TestAddrMode.ScaledReg = AddLHS;
4570 TestAddrMode.BaseOffs += CI->getSExtValue() * TestAddrMode.Scale;
4571
4572 // If this addressing mode is legal, commit it and remember that we folded
4573 // this instruction.
4574 if (TLI.isLegalAddressingMode(DL, AM: TestAddrMode, Ty: AccessTy, AddrSpace)) {
4575 AddrModeInsts.push_back(Elt: cast<Instruction>(Val: ScaleReg));
4576 AddrMode = TestAddrMode;
4577 return true;
4578 }
4579 // Restore status quo.
4580 TestAddrMode = AddrMode;
4581 }
4582
4583 // If this is an add recurrence with a constant step, return the increment
4584 // instruction and the canonicalized step.
4585 auto GetConstantStep =
4586 [this](const Value *V) -> std::optional<std::pair<Instruction *, APInt>> {
4587 auto *PN = dyn_cast<PHINode>(Val: V);
4588 if (!PN)
4589 return std::nullopt;
4590 auto IVInc = getIVIncrement(PN, LI: &LI);
4591 if (!IVInc)
4592 return std::nullopt;
4593 // TODO: The result of the intrinsics above is two-complement. However when
4594 // IV inc is expressed as add or sub, iv.next is potentially a poison value.
4595 // If it has nuw or nsw flags, we need to make sure that these flags are
4596 // inferrable at the point of memory instruction. Otherwise we are replacing
4597 // well-defined two-complement computation with poison. Currently, to avoid
4598 // potentially complex analysis needed to prove this, we reject such cases.
4599 if (auto *OIVInc = dyn_cast<OverflowingBinaryOperator>(Val: IVInc->first))
4600 if (OIVInc->hasNoSignedWrap() || OIVInc->hasNoUnsignedWrap())
4601 return std::nullopt;
4602 if (auto *ConstantStep = dyn_cast<ConstantInt>(Val: IVInc->second))
4603 return std::make_pair(x&: IVInc->first, y: ConstantStep->getValue());
4604 return std::nullopt;
4605 };
4606
4607 // Try to account for the following special case:
4608 // 1. ScaleReg is an inductive variable;
4609 // 2. We use it with non-zero offset;
4610 // 3. IV's increment is available at the point of memory instruction.
4611 //
4612 // In this case, we may reuse the IV increment instead of the IV Phi to
4613 // achieve the following advantages:
4614 // 1. If IV step matches the offset, we will have no need in the offset;
4615 // 2. Even if they don't match, we will reduce the overlap of living IV
4616 // and IV increment, that will potentially lead to better register
4617 // assignment.
4618 if (AddrMode.BaseOffs) {
4619 if (auto IVStep = GetConstantStep(ScaleReg)) {
4620 Instruction *IVInc = IVStep->first;
4621 // The following assert is important to ensure a lack of infinite loops.
4622 // This transforms is (intentionally) the inverse of the one just above.
4623 // If they don't agree on the definition of an increment, we'd alternate
4624 // back and forth indefinitely.
4625 assert(isIVIncrement(IVInc, &LI) && "implied by GetConstantStep");
4626 APInt Step = IVStep->second;
4627 APInt Offset = Step * AddrMode.Scale;
4628 if (Offset.isSignedIntN(N: 64)) {
4629 TestAddrMode.InBounds = false;
4630 TestAddrMode.ScaledReg = IVInc;
4631 TestAddrMode.BaseOffs -= Offset.getLimitedValue();
4632 // If this addressing mode is legal, commit it..
4633 // (Note that we defer the (expensive) domtree base legality check
4634 // to the very last possible point.)
4635 if (TLI.isLegalAddressingMode(DL, AM: TestAddrMode, Ty: AccessTy, AddrSpace) &&
4636 getDTFn().dominates(Def: IVInc, User: MemoryInst)) {
4637 AddrModeInsts.push_back(Elt: cast<Instruction>(Val: IVInc));
4638 AddrMode = TestAddrMode;
4639 return true;
4640 }
4641 // Restore status quo.
4642 TestAddrMode = AddrMode;
4643 }
4644 }
4645 }
4646
4647 // Otherwise, just return what we have.
4648 return true;
4649}
4650
4651/// This is a little filter, which returns true if an addressing computation
4652/// involving I might be folded into a load/store accessing it.
4653/// This doesn't need to be perfect, but needs to accept at least
4654/// the set of instructions that MatchOperationAddr can.
4655static bool MightBeFoldableInst(Instruction *I) {
4656 switch (I->getOpcode()) {
4657 case Instruction::BitCast:
4658 case Instruction::AddrSpaceCast:
4659 // Don't touch identity bitcasts.
4660 if (I->getType() == I->getOperand(i: 0)->getType())
4661 return false;
4662 return I->getType()->isIntOrPtrTy();
4663 case Instruction::PtrToInt:
4664 // PtrToInt is always a noop, as we know that the int type is pointer sized.
4665 return true;
4666 case Instruction::IntToPtr:
4667 // We know the input is intptr_t, so this is foldable.
4668 return true;
4669 case Instruction::Add:
4670 return true;
4671 case Instruction::Mul:
4672 case Instruction::Shl:
4673 // Can only handle X*C and X << C.
4674 return isa<ConstantInt>(Val: I->getOperand(i: 1));
4675 case Instruction::GetElementPtr:
4676 return true;
4677 default:
4678 return false;
4679 }
4680}
4681
4682/// Check whether or not \p Val is a legal instruction for \p TLI.
4683/// \note \p Val is assumed to be the product of some type promotion.
4684/// Therefore if \p Val has an undefined state in \p TLI, this is assumed
4685/// to be legal, as the non-promoted value would have had the same state.
4686static bool isPromotedInstructionLegal(const TargetLowering &TLI,
4687 const DataLayout &DL, Value *Val) {
4688 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
4689 if (!PromotedInst)
4690 return false;
4691 int ISDOpcode = TLI.InstructionOpcodeToISD(Opcode: PromotedInst->getOpcode());
4692 // If the ISDOpcode is undefined, it was undefined before the promotion.
4693 if (!ISDOpcode)
4694 return true;
4695 // Otherwise, check if the promoted instruction is legal or not.
4696 return TLI.isOperationLegalOrCustom(
4697 Op: ISDOpcode, VT: TLI.getValueType(DL, Ty: PromotedInst->getType()));
4698}
4699
4700namespace {
4701
4702/// Hepler class to perform type promotion.
4703class TypePromotionHelper {
4704 /// Utility function to add a promoted instruction \p ExtOpnd to
4705 /// \p PromotedInsts and record the type of extension we have seen.
4706 static void addPromotedInst(InstrToOrigTy &PromotedInsts,
4707 Instruction *ExtOpnd, bool IsSExt) {
4708 ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
4709 auto [It, Inserted] = PromotedInsts.try_emplace(Key: ExtOpnd);
4710 if (!Inserted) {
4711 // If the new extension is same as original, the information in
4712 // PromotedInsts[ExtOpnd] is still correct.
4713 if (It->second.getInt() == ExtTy)
4714 return;
4715
4716 // Now the new extension is different from old extension, we make
4717 // the type information invalid by setting extension type to
4718 // BothExtension.
4719 ExtTy = BothExtension;
4720 }
4721 It->second = TypeIsSExt(ExtOpnd->getType(), ExtTy);
4722 }
4723
4724 /// Utility function to query the original type of instruction \p Opnd
4725 /// with a matched extension type. If the extension doesn't match, we
4726 /// cannot use the information we had on the original type.
4727 /// BothExtension doesn't match any extension type.
4728 static const Type *getOrigType(const InstrToOrigTy &PromotedInsts,
4729 Instruction *Opnd, bool IsSExt) {
4730 ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
4731 InstrToOrigTy::const_iterator It = PromotedInsts.find(Val: Opnd);
4732 if (It != PromotedInsts.end() && It->second.getInt() == ExtTy)
4733 return It->second.getPointer();
4734 return nullptr;
4735 }
4736
4737 /// Utility function to check whether or not a sign or zero extension
4738 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
4739 /// either using the operands of \p Inst or promoting \p Inst.
4740 /// The type of the extension is defined by \p IsSExt.
4741 /// In other words, check if:
4742 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
4743 /// #1 Promotion applies:
4744 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
4745 /// #2 Operand reuses:
4746 /// ext opnd1 to ConsideredExtType.
4747 /// \p PromotedInsts maps the instructions to their type before promotion.
4748 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
4749 const InstrToOrigTy &PromotedInsts, bool IsSExt);
4750
4751 /// Utility function to determine if \p OpIdx should be promoted when
4752 /// promoting \p Inst.
4753 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
4754 return !(isa<SelectInst>(Val: Inst) && OpIdx == 0);
4755 }
4756
4757 /// Utility function to promote the operand of \p Ext when this
4758 /// operand is a promotable trunc or sext or zext.
4759 /// \p PromotedInsts maps the instructions to their type before promotion.
4760 /// \p CreatedInstsCost[out] contains the cost of all instructions
4761 /// created to promote the operand of Ext.
4762 /// Newly added extensions are inserted in \p Exts.
4763 /// Newly added truncates are inserted in \p Truncs.
4764 /// Should never be called directly.
4765 /// \return The promoted value which is used instead of Ext.
4766 static Value *promoteOperandForTruncAndAnyExt(
4767 Instruction *Ext, TypePromotionTransaction &TPT,
4768 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4769 SmallVectorImpl<Instruction *> *Exts,
4770 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
4771
4772 /// Utility function to promote the operand of \p Ext when this
4773 /// operand is promotable and is not a supported trunc or sext.
4774 /// \p PromotedInsts maps the instructions to their type before promotion.
4775 /// \p CreatedInstsCost[out] contains the cost of all the instructions
4776 /// created to promote the operand of Ext.
4777 /// Newly added extensions are inserted in \p Exts.
4778 /// Newly added truncates are inserted in \p Truncs.
4779 /// Should never be called directly.
4780 /// \return The promoted value which is used instead of Ext.
4781 static Value *promoteOperandForOther(Instruction *Ext,
4782 TypePromotionTransaction &TPT,
4783 InstrToOrigTy &PromotedInsts,
4784 unsigned &CreatedInstsCost,
4785 SmallVectorImpl<Instruction *> *Exts,
4786 SmallVectorImpl<Instruction *> *Truncs,
4787 const TargetLowering &TLI, bool IsSExt);
4788
4789 /// \see promoteOperandForOther.
4790 static Value *signExtendOperandForOther(
4791 Instruction *Ext, TypePromotionTransaction &TPT,
4792 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4793 SmallVectorImpl<Instruction *> *Exts,
4794 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4795 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
4796 Exts, Truncs, TLI, IsSExt: true);
4797 }
4798
4799 /// \see promoteOperandForOther.
4800 static Value *zeroExtendOperandForOther(
4801 Instruction *Ext, TypePromotionTransaction &TPT,
4802 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4803 SmallVectorImpl<Instruction *> *Exts,
4804 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4805 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
4806 Exts, Truncs, TLI, IsSExt: false);
4807 }
4808
4809public:
4810 /// Type for the utility function that promotes the operand of Ext.
4811 using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
4812 InstrToOrigTy &PromotedInsts,
4813 unsigned &CreatedInstsCost,
4814 SmallVectorImpl<Instruction *> *Exts,
4815 SmallVectorImpl<Instruction *> *Truncs,
4816 const TargetLowering &TLI);
4817
4818 /// Given a sign/zero extend instruction \p Ext, return the appropriate
4819 /// action to promote the operand of \p Ext instead of using Ext.
4820 /// \return NULL if no promotable action is possible with the current
4821 /// sign extension.
4822 /// \p InsertedInsts keeps track of all the instructions inserted by the
4823 /// other CodeGenPrepare optimizations. This information is important
4824 /// because we do not want to promote these instructions as CodeGenPrepare
4825 /// will reinsert them later. Thus creating an infinite loop: create/remove.
4826 /// \p PromotedInsts maps the instructions to their type before promotion.
4827 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
4828 const TargetLowering &TLI,
4829 const InstrToOrigTy &PromotedInsts);
4830};
4831
4832} // end anonymous namespace
4833
4834bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
4835 Type *ConsideredExtType,
4836 const InstrToOrigTy &PromotedInsts,
4837 bool IsSExt) {
4838 // The promotion helper does not know how to deal with vector types yet.
4839 // To be able to fix that, we would need to fix the places where we
4840 // statically extend, e.g., constants and such.
4841 if (Inst->getType()->isVectorTy())
4842 return false;
4843
4844 // We can always get through zext.
4845 if (isa<ZExtInst>(Val: Inst))
4846 return true;
4847
4848 // sext(sext) is ok too.
4849 if (IsSExt && isa<SExtInst>(Val: Inst))
4850 return true;
4851
4852 // We can get through binary operator, if it is legal. In other words, the
4853 // binary operator must have a nuw or nsw flag.
4854 if (const auto *BinOp = dyn_cast<BinaryOperator>(Val: Inst))
4855 if (isa<OverflowingBinaryOperator>(Val: BinOp) &&
4856 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
4857 (IsSExt && BinOp->hasNoSignedWrap())))
4858 return true;
4859
4860 // ext(and(opnd, cst)) --> and(ext(opnd), ext(cst))
4861 if ((Inst->getOpcode() == Instruction::And ||
4862 Inst->getOpcode() == Instruction::Or))
4863 return true;
4864
4865 // ext(xor(opnd, cst)) --> xor(ext(opnd), ext(cst))
4866 if (Inst->getOpcode() == Instruction::Xor) {
4867 // Make sure it is not a NOT.
4868 if (const auto *Cst = dyn_cast<ConstantInt>(Val: Inst->getOperand(i: 1)))
4869 if (!Cst->getValue().isAllOnes())
4870 return true;
4871 }
4872
4873 // zext(shrl(opnd, cst)) --> shrl(zext(opnd), zext(cst))
4874 // It may change a poisoned value into a regular value, like
4875 // zext i32 (shrl i8 %val, 12) --> shrl i32 (zext i8 %val), 12
4876 // poisoned value regular value
4877 // It should be OK since undef covers valid value.
4878 if (Inst->getOpcode() == Instruction::LShr && !IsSExt)
4879 return true;
4880
4881 // and(ext(shl(opnd, cst)), cst) --> and(shl(ext(opnd), ext(cst)), cst)
4882 // It may change a poisoned value into a regular value, like
4883 // zext i32 (shl i8 %val, 12) --> shl i32 (zext i8 %val), 12
4884 // poisoned value regular value
4885 // It should be OK since undef covers valid value.
4886 if (Inst->getOpcode() == Instruction::Shl && Inst->hasOneUse()) {
4887 const auto *ExtInst = cast<const Instruction>(Val: *Inst->user_begin());
4888 if (ExtInst->hasOneUse()) {
4889 const auto *AndInst = dyn_cast<const Instruction>(Val: *ExtInst->user_begin());
4890 if (AndInst && AndInst->getOpcode() == Instruction::And) {
4891 const auto *Cst = dyn_cast<ConstantInt>(Val: AndInst->getOperand(i: 1));
4892 if (Cst &&
4893 Cst->getValue().isIntN(N: Inst->getType()->getIntegerBitWidth()))
4894 return true;
4895 }
4896 }
4897 }
4898
4899 // Check if we can do the following simplification.
4900 // ext(trunc(opnd)) --> ext(opnd)
4901 if (!isa<TruncInst>(Val: Inst))
4902 return false;
4903
4904 Value *OpndVal = Inst->getOperand(i: 0);
4905 // Check if we can use this operand in the extension.
4906 // If the type is larger than the result type of the extension, we cannot.
4907 if (!OpndVal->getType()->isIntegerTy() ||
4908 OpndVal->getType()->getIntegerBitWidth() >
4909 ConsideredExtType->getIntegerBitWidth())
4910 return false;
4911
4912 // If the operand of the truncate is not an instruction, we will not have
4913 // any information on the dropped bits.
4914 // (Actually we could for constant but it is not worth the extra logic).
4915 Instruction *Opnd = dyn_cast<Instruction>(Val: OpndVal);
4916 if (!Opnd)
4917 return false;
4918
4919 // Check if the source of the type is narrow enough.
4920 // I.e., check that trunc just drops extended bits of the same kind of
4921 // the extension.
4922 // #1 get the type of the operand and check the kind of the extended bits.
4923 const Type *OpndType = getOrigType(PromotedInsts, Opnd, IsSExt);
4924 if (OpndType)
4925 ;
4926 else if ((IsSExt && isa<SExtInst>(Val: Opnd)) || (!IsSExt && isa<ZExtInst>(Val: Opnd)))
4927 OpndType = Opnd->getOperand(i: 0)->getType();
4928 else
4929 return false;
4930
4931 // #2 check that the truncate just drops extended bits.
4932 return Inst->getType()->getIntegerBitWidth() >=
4933 OpndType->getIntegerBitWidth();
4934}
4935
4936TypePromotionHelper::Action TypePromotionHelper::getAction(
4937 Instruction *Ext, const SetOfInstrs &InsertedInsts,
4938 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
4939 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
4940 "Unexpected instruction type");
4941 Instruction *ExtOpnd = dyn_cast<Instruction>(Val: Ext->getOperand(i: 0));
4942 Type *ExtTy = Ext->getType();
4943 bool IsSExt = isa<SExtInst>(Val: Ext);
4944 // If the operand of the extension is not an instruction, we cannot
4945 // get through.
4946 // If it, check we can get through.
4947 if (!ExtOpnd || !canGetThrough(Inst: ExtOpnd, ConsideredExtType: ExtTy, PromotedInsts, IsSExt))
4948 return nullptr;
4949
4950 // Do not promote if the operand has been added by codegenprepare.
4951 // Otherwise, it means we are undoing an optimization that is likely to be
4952 // redone, thus causing potential infinite loop.
4953 if (isa<TruncInst>(Val: ExtOpnd) && InsertedInsts.count(Ptr: ExtOpnd))
4954 return nullptr;
4955
4956 // SExt or Trunc instructions.
4957 // Return the related handler.
4958 if (isa<SExtInst>(Val: ExtOpnd) || isa<TruncInst>(Val: ExtOpnd) ||
4959 isa<ZExtInst>(Val: ExtOpnd))
4960 return promoteOperandForTruncAndAnyExt;
4961
4962 // Regular instruction.
4963 // Abort early if we will have to insert non-free instructions.
4964 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(FromTy: ExtTy, ToTy: ExtOpnd->getType()))
4965 return nullptr;
4966 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
4967}
4968
4969Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
4970 Instruction *SExt, TypePromotionTransaction &TPT,
4971 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4972 SmallVectorImpl<Instruction *> *Exts,
4973 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4974 // By construction, the operand of SExt is an instruction. Otherwise we cannot
4975 // get through it and this method should not be called.
4976 Instruction *SExtOpnd = cast<Instruction>(Val: SExt->getOperand(i: 0));
4977 Value *ExtVal = SExt;
4978 bool HasMergedNonFreeExt = false;
4979 if (isa<ZExtInst>(Val: SExtOpnd)) {
4980 // Replace s|zext(zext(opnd))
4981 // => zext(opnd).
4982 HasMergedNonFreeExt = !TLI.isExtFree(I: SExtOpnd);
4983 Value *ZExt =
4984 TPT.createZExt(Inst: SExt, Opnd: SExtOpnd->getOperand(i: 0), Ty: SExt->getType());
4985 TPT.replaceAllUsesWith(Inst: SExt, New: ZExt);
4986 TPT.eraseInstruction(Inst: SExt);
4987 ExtVal = ZExt;
4988 } else {
4989 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
4990 // => z|sext(opnd).
4991 TPT.setOperand(Inst: SExt, Idx: 0, NewVal: SExtOpnd->getOperand(i: 0));
4992 }
4993 CreatedInstsCost = 0;
4994
4995 // Remove dead code.
4996 if (SExtOpnd->use_empty())
4997 TPT.eraseInstruction(Inst: SExtOpnd);
4998
4999 // Check if the extension is still needed.
5000 Instruction *ExtInst = dyn_cast<Instruction>(Val: ExtVal);
5001 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(i: 0)->getType()) {
5002 if (ExtInst) {
5003 if (Exts)
5004 Exts->push_back(Elt: ExtInst);
5005 CreatedInstsCost = !TLI.isExtFree(I: ExtInst) && !HasMergedNonFreeExt;
5006 }
5007 return ExtVal;
5008 }
5009
5010 // At this point we have: ext ty opnd to ty.
5011 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
5012 Value *NextVal = ExtInst->getOperand(i: 0);
5013 TPT.eraseInstruction(Inst: ExtInst, NewVal: NextVal);
5014 return NextVal;
5015}
5016
5017Value *TypePromotionHelper::promoteOperandForOther(
5018 Instruction *Ext, TypePromotionTransaction &TPT,
5019 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
5020 SmallVectorImpl<Instruction *> *Exts,
5021 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
5022 bool IsSExt) {
5023 // By construction, the operand of Ext is an instruction. Otherwise we cannot
5024 // get through it and this method should not be called.
5025 Instruction *ExtOpnd = cast<Instruction>(Val: Ext->getOperand(i: 0));
5026 CreatedInstsCost = 0;
5027 if (!ExtOpnd->hasOneUse()) {
5028 // ExtOpnd will be promoted.
5029 // All its uses, but Ext, will need to use a truncated value of the
5030 // promoted version.
5031 // Create the truncate now.
5032 Value *Trunc = TPT.createTrunc(Opnd: Ext, Ty: ExtOpnd->getType());
5033 if (Instruction *ITrunc = dyn_cast<Instruction>(Val: Trunc)) {
5034 // Insert it just after the definition.
5035 ITrunc->moveAfter(MovePos: ExtOpnd);
5036 if (Truncs)
5037 Truncs->push_back(Elt: ITrunc);
5038 }
5039
5040 TPT.replaceAllUsesWith(Inst: ExtOpnd, New: Trunc);
5041 // Restore the operand of Ext (which has been replaced by the previous call
5042 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
5043 TPT.setOperand(Inst: Ext, Idx: 0, NewVal: ExtOpnd);
5044 }
5045
5046 // Get through the Instruction:
5047 // 1. Update its type.
5048 // 2. Replace the uses of Ext by Inst.
5049 // 3. Extend each operand that needs to be extended.
5050
5051 // Remember the original type of the instruction before promotion.
5052 // This is useful to know that the high bits are sign extended bits.
5053 addPromotedInst(PromotedInsts, ExtOpnd, IsSExt);
5054 // Step #1.
5055 TPT.mutateType(Inst: ExtOpnd, NewTy: Ext->getType());
5056 // Step #2.
5057 TPT.replaceAllUsesWith(Inst: Ext, New: ExtOpnd);
5058 // Step #3.
5059 LLVM_DEBUG(dbgs() << "Propagate Ext to operands\n");
5060 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
5061 ++OpIdx) {
5062 LLVM_DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
5063 if (ExtOpnd->getOperand(i: OpIdx)->getType() == Ext->getType() ||
5064 !shouldExtOperand(Inst: ExtOpnd, OpIdx)) {
5065 LLVM_DEBUG(dbgs() << "No need to propagate\n");
5066 continue;
5067 }
5068 // Check if we can statically extend the operand.
5069 Value *Opnd = ExtOpnd->getOperand(i: OpIdx);
5070 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Val: Opnd)) {
5071 LLVM_DEBUG(dbgs() << "Statically extend\n");
5072 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
5073 APInt CstVal = IsSExt ? Cst->getValue().sext(width: BitWidth)
5074 : Cst->getValue().zext(width: BitWidth);
5075 TPT.setOperand(Inst: ExtOpnd, Idx: OpIdx, NewVal: ConstantInt::get(Ty: Ext->getType(), V: CstVal));
5076 continue;
5077 }
5078 // UndefValue are typed, so we have to statically sign extend them.
5079 if (isa<UndefValue>(Val: Opnd)) {
5080 LLVM_DEBUG(dbgs() << "Statically extend\n");
5081 TPT.setOperand(Inst: ExtOpnd, Idx: OpIdx, NewVal: UndefValue::get(T: Ext->getType()));
5082 continue;
5083 }
5084
5085 // Otherwise we have to explicitly sign extend the operand.
5086 Value *ValForExtOpnd = IsSExt
5087 ? TPT.createSExt(Inst: ExtOpnd, Opnd, Ty: Ext->getType())
5088 : TPT.createZExt(Inst: ExtOpnd, Opnd, Ty: Ext->getType());
5089 TPT.setOperand(Inst: ExtOpnd, Idx: OpIdx, NewVal: ValForExtOpnd);
5090 Instruction *InstForExtOpnd = dyn_cast<Instruction>(Val: ValForExtOpnd);
5091 if (!InstForExtOpnd)
5092 continue;
5093
5094 if (Exts)
5095 Exts->push_back(Elt: InstForExtOpnd);
5096
5097 CreatedInstsCost += !TLI.isExtFree(I: InstForExtOpnd);
5098 }
5099 LLVM_DEBUG(dbgs() << "Extension is useless now\n");
5100 TPT.eraseInstruction(Inst: Ext);
5101 return ExtOpnd;
5102}
5103
5104/// Check whether or not promoting an instruction to a wider type is profitable.
5105/// \p NewCost gives the cost of extension instructions created by the
5106/// promotion.
5107/// \p OldCost gives the cost of extension instructions before the promotion
5108/// plus the number of instructions that have been
5109/// matched in the addressing mode the promotion.
5110/// \p PromotedOperand is the value that has been promoted.
5111/// \return True if the promotion is profitable, false otherwise.
5112bool AddressingModeMatcher::isPromotionProfitable(
5113 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
5114 LLVM_DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost
5115 << '\n');
5116 // The cost of the new extensions is greater than the cost of the
5117 // old extension plus what we folded.
5118 // This is not profitable.
5119 if (NewCost > OldCost)
5120 return false;
5121 if (NewCost < OldCost)
5122 return true;
5123 // The promotion is neutral but it may help folding the sign extension in
5124 // loads for instance.
5125 // Check that we did not create an illegal instruction.
5126 return isPromotedInstructionLegal(TLI, DL, Val: PromotedOperand);
5127}
5128
5129/// Given an instruction or constant expr, see if we can fold the operation
5130/// into the addressing mode. If so, update the addressing mode and return
5131/// true, otherwise return false without modifying AddrMode.
5132/// If \p MovedAway is not NULL, it contains the information of whether or
5133/// not AddrInst has to be folded into the addressing mode on success.
5134/// If \p MovedAway == true, \p AddrInst will not be part of the addressing
5135/// because it has been moved away.
5136/// Thus AddrInst must not be added in the matched instructions.
5137/// This state can happen when AddrInst is a sext, since it may be moved away.
5138/// Therefore, AddrInst may not be valid when MovedAway is true and it must
5139/// not be referenced anymore.
5140bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
5141 unsigned Depth,
5142 bool *MovedAway) {
5143 // Avoid exponential behavior on extremely deep expression trees.
5144 if (Depth >= 5)
5145 return false;
5146
5147 // By default, all matched instructions stay in place.
5148 if (MovedAway)
5149 *MovedAway = false;
5150
5151 switch (Opcode) {
5152 case Instruction::PtrToInt:
5153 // PtrToInt is always a noop, as we know that the int type is pointer sized.
5154 return matchAddr(Addr: AddrInst->getOperand(i: 0), Depth);
5155 case Instruction::IntToPtr: {
5156 auto AS = AddrInst->getType()->getPointerAddressSpace();
5157 auto PtrTy = MVT::getIntegerVT(BitWidth: DL.getPointerSizeInBits(AS));
5158 // This inttoptr is a no-op if the integer type is pointer sized.
5159 if (TLI.getValueType(DL, Ty: AddrInst->getOperand(i: 0)->getType()) == PtrTy)
5160 return matchAddr(Addr: AddrInst->getOperand(i: 0), Depth);
5161 return false;
5162 }
5163 case Instruction::BitCast:
5164 // BitCast is always a noop, and we can handle it as long as it is
5165 // int->int or pointer->pointer (we don't want int<->fp or something).
5166 if (AddrInst->getOperand(i: 0)->getType()->isIntOrPtrTy() &&
5167 // Don't touch identity bitcasts. These were probably put here by LSR,
5168 // and we don't want to mess around with them. Assume it knows what it
5169 // is doing.
5170 AddrInst->getOperand(i: 0)->getType() != AddrInst->getType())
5171 return matchAddr(Addr: AddrInst->getOperand(i: 0), Depth);
5172 return false;
5173 case Instruction::AddrSpaceCast: {
5174 unsigned SrcAS =
5175 AddrInst->getOperand(i: 0)->getType()->getPointerAddressSpace();
5176 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
5177 if (TLI.getTargetMachine().isNoopAddrSpaceCast(SrcAS, DestAS))
5178 return matchAddr(Addr: AddrInst->getOperand(i: 0), Depth);
5179 return false;
5180 }
5181 case Instruction::Add: {
5182 // Check to see if we can merge in one operand, then the other. If so, we
5183 // win.
5184 ExtAddrMode BackupAddrMode = AddrMode;
5185 unsigned OldSize = AddrModeInsts.size();
5186 // Start a transaction at this point.
5187 // The LHS may match but not the RHS.
5188 // Therefore, we need a higher level restoration point to undo partially
5189 // matched operation.
5190 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5191 TPT.getRestorationPoint();
5192
5193 // Try to match an integer constant second to increase its chance of ending
5194 // up in `BaseOffs`, resp. decrease its chance of ending up in `BaseReg`.
5195 int First = 0, Second = 1;
5196 if (isa<ConstantInt>(Val: AddrInst->getOperand(i: First))
5197 && !isa<ConstantInt>(Val: AddrInst->getOperand(i: Second)))
5198 std::swap(a&: First, b&: Second);
5199 AddrMode.InBounds = false;
5200 if (matchAddr(Addr: AddrInst->getOperand(i: First), Depth: Depth + 1) &&
5201 matchAddr(Addr: AddrInst->getOperand(i: Second), Depth: Depth + 1))
5202 return true;
5203
5204 // Restore the old addr mode info.
5205 AddrMode = BackupAddrMode;
5206 AddrModeInsts.resize(N: OldSize);
5207 TPT.rollback(Point: LastKnownGood);
5208
5209 // Otherwise this was over-aggressive. Try merging operands in the opposite
5210 // order.
5211 if (matchAddr(Addr: AddrInst->getOperand(i: Second), Depth: Depth + 1) &&
5212 matchAddr(Addr: AddrInst->getOperand(i: First), Depth: Depth + 1))
5213 return true;
5214
5215 // Otherwise we definitely can't merge the ADD in.
5216 AddrMode = BackupAddrMode;
5217 AddrModeInsts.resize(N: OldSize);
5218 TPT.rollback(Point: LastKnownGood);
5219 break;
5220 }
5221 // case Instruction::Or:
5222 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
5223 // break;
5224 case Instruction::Mul:
5225 case Instruction::Shl: {
5226 // Can only handle X*C and X << C.
5227 AddrMode.InBounds = false;
5228 ConstantInt *RHS = dyn_cast<ConstantInt>(Val: AddrInst->getOperand(i: 1));
5229 if (!RHS || RHS->getBitWidth() > 64)
5230 return false;
5231 int64_t Scale = Opcode == Instruction::Shl
5232 ? 1LL << RHS->getLimitedValue(Limit: RHS->getBitWidth() - 1)
5233 : RHS->getSExtValue();
5234
5235 return matchScaledValue(ScaleReg: AddrInst->getOperand(i: 0), Scale, Depth);
5236 }
5237 case Instruction::GetElementPtr: {
5238 // Scan the GEP. We check it if it contains constant offsets and at most
5239 // one variable offset.
5240 int VariableOperand = -1;
5241 unsigned VariableScale = 0;
5242
5243 int64_t ConstantOffset = 0;
5244 gep_type_iterator GTI = gep_type_begin(GEP: AddrInst);
5245 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
5246 if (StructType *STy = GTI.getStructTypeOrNull()) {
5247 const StructLayout *SL = DL.getStructLayout(Ty: STy);
5248 unsigned Idx =
5249 cast<ConstantInt>(Val: AddrInst->getOperand(i))->getZExtValue();
5250 ConstantOffset += SL->getElementOffset(Idx);
5251 } else {
5252 TypeSize TS = GTI.getSequentialElementStride(DL);
5253 if (TS.isNonZero()) {
5254 // The optimisations below currently only work for fixed offsets.
5255 if (TS.isScalable())
5256 return false;
5257 int64_t TypeSize = TS.getFixedValue();
5258 if (ConstantInt *CI =
5259 dyn_cast<ConstantInt>(Val: AddrInst->getOperand(i))) {
5260 const APInt &CVal = CI->getValue();
5261 if (CVal.getSignificantBits() <= 64) {
5262 ConstantOffset += CVal.getSExtValue() * TypeSize;
5263 continue;
5264 }
5265 }
5266 // We only allow one variable index at the moment.
5267 if (VariableOperand != -1)
5268 return false;
5269
5270 // Remember the variable index.
5271 VariableOperand = i;
5272 VariableScale = TypeSize;
5273 }
5274 }
5275 }
5276
5277 // A common case is for the GEP to only do a constant offset. In this case,
5278 // just add it to the disp field and check validity.
5279 if (VariableOperand == -1) {
5280 AddrMode.BaseOffs += ConstantOffset;
5281 if (matchAddr(Addr: AddrInst->getOperand(i: 0), Depth: Depth + 1)) {
5282 if (!cast<GEPOperator>(Val: AddrInst)->isInBounds())
5283 AddrMode.InBounds = false;
5284 return true;
5285 }
5286 AddrMode.BaseOffs -= ConstantOffset;
5287
5288 if (EnableGEPOffsetSplit && isa<GetElementPtrInst>(Val: AddrInst) &&
5289 TLI.shouldConsiderGEPOffsetSplit() && Depth == 0 &&
5290 ConstantOffset > 0) {
5291 // Record GEPs with non-zero offsets as candidates for splitting in
5292 // the event that the offset cannot fit into the r+i addressing mode.
5293 // Simple and common case that only one GEP is used in calculating the
5294 // address for the memory access.
5295 Value *Base = AddrInst->getOperand(i: 0);
5296 auto *BaseI = dyn_cast<Instruction>(Val: Base);
5297 auto *GEP = cast<GetElementPtrInst>(Val: AddrInst);
5298 if (isa<Argument>(Val: Base) || isa<GlobalValue>(Val: Base) ||
5299 (BaseI && !isa<CastInst>(Val: BaseI) &&
5300 !isa<GetElementPtrInst>(Val: BaseI))) {
5301 // Make sure the parent block allows inserting non-PHI instructions
5302 // before the terminator.
5303 BasicBlock *Parent = BaseI ? BaseI->getParent()
5304 : &GEP->getFunction()->getEntryBlock();
5305 if (!Parent->getTerminator()->isEHPad())
5306 LargeOffsetGEP = std::make_pair(x&: GEP, y&: ConstantOffset);
5307 }
5308 }
5309
5310 return false;
5311 }
5312
5313 // Save the valid addressing mode in case we can't match.
5314 ExtAddrMode BackupAddrMode = AddrMode;
5315 unsigned OldSize = AddrModeInsts.size();
5316
5317 // See if the scale and offset amount is valid for this target.
5318 AddrMode.BaseOffs += ConstantOffset;
5319 if (!cast<GEPOperator>(Val: AddrInst)->isInBounds())
5320 AddrMode.InBounds = false;
5321
5322 // Match the base operand of the GEP.
5323 if (!matchAddr(Addr: AddrInst->getOperand(i: 0), Depth: Depth + 1)) {
5324 // If it couldn't be matched, just stuff the value in a register.
5325 if (AddrMode.HasBaseReg) {
5326 AddrMode = BackupAddrMode;
5327 AddrModeInsts.resize(N: OldSize);
5328 return false;
5329 }
5330 AddrMode.HasBaseReg = true;
5331 AddrMode.BaseReg = AddrInst->getOperand(i: 0);
5332 }
5333
5334 // Match the remaining variable portion of the GEP.
5335 if (!matchScaledValue(ScaleReg: AddrInst->getOperand(i: VariableOperand), Scale: VariableScale,
5336 Depth)) {
5337 // If it couldn't be matched, try stuffing the base into a register
5338 // instead of matching it, and retrying the match of the scale.
5339 AddrMode = BackupAddrMode;
5340 AddrModeInsts.resize(N: OldSize);
5341 if (AddrMode.HasBaseReg)
5342 return false;
5343 AddrMode.HasBaseReg = true;
5344 AddrMode.BaseReg = AddrInst->getOperand(i: 0);
5345 AddrMode.BaseOffs += ConstantOffset;
5346 if (!matchScaledValue(ScaleReg: AddrInst->getOperand(i: VariableOperand),
5347 Scale: VariableScale, Depth)) {
5348 // If even that didn't work, bail.
5349 AddrMode = BackupAddrMode;
5350 AddrModeInsts.resize(N: OldSize);
5351 return false;
5352 }
5353 }
5354
5355 return true;
5356 }
5357 case Instruction::SExt:
5358 case Instruction::ZExt: {
5359 Instruction *Ext = dyn_cast<Instruction>(Val: AddrInst);
5360 if (!Ext)
5361 return false;
5362
5363 // Try to move this ext out of the way of the addressing mode.
5364 // Ask for a method for doing so.
5365 TypePromotionHelper::Action TPH =
5366 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
5367 if (!TPH)
5368 return false;
5369
5370 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5371 TPT.getRestorationPoint();
5372 unsigned CreatedInstsCost = 0;
5373 unsigned ExtCost = !TLI.isExtFree(I: Ext);
5374 Value *PromotedOperand =
5375 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
5376 // SExt has been moved away.
5377 // Thus either it will be rematched later in the recursive calls or it is
5378 // gone. Anyway, we must not fold it into the addressing mode at this point.
5379 // E.g.,
5380 // op = add opnd, 1
5381 // idx = ext op
5382 // addr = gep base, idx
5383 // is now:
5384 // promotedOpnd = ext opnd <- no match here
5385 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
5386 // addr = gep base, op <- match
5387 if (MovedAway)
5388 *MovedAway = true;
5389
5390 assert(PromotedOperand &&
5391 "TypePromotionHelper should have filtered out those cases");
5392
5393 ExtAddrMode BackupAddrMode = AddrMode;
5394 unsigned OldSize = AddrModeInsts.size();
5395
5396 if (!matchAddr(Addr: PromotedOperand, Depth) ||
5397 // The total of the new cost is equal to the cost of the created
5398 // instructions.
5399 // The total of the old cost is equal to the cost of the extension plus
5400 // what we have saved in the addressing mode.
5401 !isPromotionProfitable(NewCost: CreatedInstsCost,
5402 OldCost: ExtCost + (AddrModeInsts.size() - OldSize),
5403 PromotedOperand)) {
5404 AddrMode = BackupAddrMode;
5405 AddrModeInsts.resize(N: OldSize);
5406 LLVM_DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
5407 TPT.rollback(Point: LastKnownGood);
5408 return false;
5409 }
5410
5411 // SExt has been deleted. Make sure it is not referenced by the AddrMode.
5412 AddrMode.replaceWith(From: Ext, To: PromotedOperand);
5413 return true;
5414 }
5415 case Instruction::Call:
5416 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: AddrInst)) {
5417 if (II->getIntrinsicID() == Intrinsic::threadlocal_address) {
5418 GlobalValue &GV = cast<GlobalValue>(Val&: *II->getArgOperand(i: 0));
5419 if (TLI.addressingModeSupportsTLS(GV))
5420 return matchAddr(Addr: AddrInst->getOperand(i: 0), Depth);
5421 }
5422 }
5423 break;
5424 }
5425 return false;
5426}
5427
5428/// If we can, try to add the value of 'Addr' into the current addressing mode.
5429/// If Addr can't be added to AddrMode this returns false and leaves AddrMode
5430/// unmodified. This assumes that Addr is either a pointer type or intptr_t
5431/// for the target.
5432///
5433bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
5434 // Start a transaction at this point that we will rollback if the matching
5435 // fails.
5436 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5437 TPT.getRestorationPoint();
5438 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: Addr)) {
5439 if (CI->getValue().isSignedIntN(N: 64)) {
5440 // Check if the addition would result in a signed overflow.
5441 int64_t Result;
5442 bool Overflow =
5443 AddOverflow(X: AddrMode.BaseOffs, Y: CI->getSExtValue(), Result);
5444 if (!Overflow) {
5445 // Fold in immediates if legal for the target.
5446 AddrMode.BaseOffs = Result;
5447 if (TLI.isLegalAddressingMode(DL, AM: AddrMode, Ty: AccessTy, AddrSpace))
5448 return true;
5449 AddrMode.BaseOffs -= CI->getSExtValue();
5450 }
5451 }
5452 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Val: Addr)) {
5453 // If this is a global variable, try to fold it into the addressing mode.
5454 if (!AddrMode.BaseGV) {
5455 AddrMode.BaseGV = GV;
5456 if (TLI.isLegalAddressingMode(DL, AM: AddrMode, Ty: AccessTy, AddrSpace))
5457 return true;
5458 AddrMode.BaseGV = nullptr;
5459 }
5460 } else if (Instruction *I = dyn_cast<Instruction>(Val: Addr)) {
5461 ExtAddrMode BackupAddrMode = AddrMode;
5462 unsigned OldSize = AddrModeInsts.size();
5463
5464 // Check to see if it is possible to fold this operation.
5465 bool MovedAway = false;
5466 if (matchOperationAddr(AddrInst: I, Opcode: I->getOpcode(), Depth, MovedAway: &MovedAway)) {
5467 // This instruction may have been moved away. If so, there is nothing
5468 // to check here.
5469 if (MovedAway)
5470 return true;
5471 // Okay, it's possible to fold this. Check to see if it is actually
5472 // *profitable* to do so. We use a simple cost model to avoid increasing
5473 // register pressure too much.
5474 if (I->hasOneUse() ||
5475 isProfitableToFoldIntoAddressingMode(I, AMBefore&: BackupAddrMode, AMAfter&: AddrMode)) {
5476 AddrModeInsts.push_back(Elt: I);
5477 return true;
5478 }
5479
5480 // It isn't profitable to do this, roll back.
5481 AddrMode = BackupAddrMode;
5482 AddrModeInsts.resize(N: OldSize);
5483 TPT.rollback(Point: LastKnownGood);
5484 }
5485 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Val: Addr)) {
5486 if (matchOperationAddr(AddrInst: CE, Opcode: CE->getOpcode(), Depth))
5487 return true;
5488 TPT.rollback(Point: LastKnownGood);
5489 } else if (isa<ConstantPointerNull>(Val: Addr)) {
5490 // Null pointer gets folded without affecting the addressing mode.
5491 return true;
5492 }
5493
5494 // Worse case, the target should support [reg] addressing modes. :)
5495 if (!AddrMode.HasBaseReg) {
5496 AddrMode.HasBaseReg = true;
5497 AddrMode.BaseReg = Addr;
5498 // Still check for legality in case the target supports [imm] but not [i+r].
5499 if (TLI.isLegalAddressingMode(DL, AM: AddrMode, Ty: AccessTy, AddrSpace))
5500 return true;
5501 AddrMode.HasBaseReg = false;
5502 AddrMode.BaseReg = nullptr;
5503 }
5504
5505 // If the base register is already taken, see if we can do [r+r].
5506 if (AddrMode.Scale == 0) {
5507 AddrMode.Scale = 1;
5508 AddrMode.ScaledReg = Addr;
5509 if (TLI.isLegalAddressingMode(DL, AM: AddrMode, Ty: AccessTy, AddrSpace))
5510 return true;
5511 AddrMode.Scale = 0;
5512 AddrMode.ScaledReg = nullptr;
5513 }
5514 // Couldn't match.
5515 TPT.rollback(Point: LastKnownGood);
5516 return false;
5517}
5518
5519/// Check to see if all uses of OpVal by the specified inline asm call are due
5520/// to memory operands. If so, return true, otherwise return false.
5521static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
5522 const TargetLowering &TLI,
5523 const TargetRegisterInfo &TRI) {
5524 const Function *F = CI->getFunction();
5525 TargetLowering::AsmOperandInfoVector TargetConstraints =
5526 TLI.ParseConstraints(DL: F->getDataLayout(), TRI: &TRI, Call: *CI);
5527
5528 for (TargetLowering::AsmOperandInfo &OpInfo : TargetConstraints) {
5529 // Compute the constraint code and ConstraintType to use.
5530 TLI.ComputeConstraintToUse(OpInfo, Op: SDValue());
5531
5532 // If this asm operand is our Value*, and if it isn't an indirect memory
5533 // operand, we can't fold it! TODO: Also handle C_Address?
5534 if (OpInfo.CallOperandVal == OpVal &&
5535 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
5536 !OpInfo.isIndirect))
5537 return false;
5538 }
5539
5540 return true;
5541}
5542
5543/// Recursively walk all the uses of I until we find a memory use.
5544/// If we find an obviously non-foldable instruction, return true.
5545/// Add accessed addresses and types to MemoryUses.
5546static bool FindAllMemoryUses(
5547 Instruction *I, SmallVectorImpl<std::pair<Use *, Type *>> &MemoryUses,
5548 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
5549 const TargetRegisterInfo &TRI, bool OptSize, ProfileSummaryInfo *PSI,
5550 BlockFrequencyInfo *BFI, unsigned &SeenInsts) {
5551 // If we already considered this instruction, we're done.
5552 if (!ConsideredInsts.insert(Ptr: I).second)
5553 return false;
5554
5555 // If this is an obviously unfoldable instruction, bail out.
5556 if (!MightBeFoldableInst(I))
5557 return true;
5558
5559 // Loop over all the uses, recursively processing them.
5560 for (Use &U : I->uses()) {
5561 // Conservatively return true if we're seeing a large number or a deep chain
5562 // of users. This avoids excessive compilation times in pathological cases.
5563 if (SeenInsts++ >= MaxAddressUsersToScan)
5564 return true;
5565
5566 Instruction *UserI = cast<Instruction>(Val: U.getUser());
5567 if (LoadInst *LI = dyn_cast<LoadInst>(Val: UserI)) {
5568 MemoryUses.push_back(Elt: {&U, LI->getType()});
5569 continue;
5570 }
5571
5572 if (StoreInst *SI = dyn_cast<StoreInst>(Val: UserI)) {
5573 if (U.getOperandNo() != StoreInst::getPointerOperandIndex())
5574 return true; // Storing addr, not into addr.
5575 MemoryUses.push_back(Elt: {&U, SI->getValueOperand()->getType()});
5576 continue;
5577 }
5578
5579 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: UserI)) {
5580 if (U.getOperandNo() != AtomicRMWInst::getPointerOperandIndex())
5581 return true; // Storing addr, not into addr.
5582 MemoryUses.push_back(Elt: {&U, RMW->getValOperand()->getType()});
5583 continue;
5584 }
5585
5586 if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Val: UserI)) {
5587 if (U.getOperandNo() != AtomicCmpXchgInst::getPointerOperandIndex())
5588 return true; // Storing addr, not into addr.
5589 MemoryUses.push_back(Elt: {&U, CmpX->getCompareOperand()->getType()});
5590 continue;
5591 }
5592
5593 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: UserI)) {
5594 SmallVector<Value *, 2> PtrOps;
5595 Type *AccessTy;
5596 if (!TLI.getAddrModeArguments(II, PtrOps, AccessTy))
5597 return true;
5598
5599 if (!find(Range&: PtrOps, Val: U.get()))
5600 return true;
5601
5602 MemoryUses.push_back(Elt: {&U, AccessTy});
5603 continue;
5604 }
5605
5606 if (CallInst *CI = dyn_cast<CallInst>(Val: UserI)) {
5607 if (CI->hasFnAttr(Kind: Attribute::Cold)) {
5608 // If this is a cold call, we can sink the addressing calculation into
5609 // the cold path. See optimizeCallInst
5610 if (!llvm::shouldOptimizeForSize(BB: CI->getParent(), PSI, BFI))
5611 continue;
5612 }
5613
5614 InlineAsm *IA = dyn_cast<InlineAsm>(Val: CI->getCalledOperand());
5615 if (!IA)
5616 return true;
5617
5618 // If this is a memory operand, we're cool, otherwise bail out.
5619 if (!IsOperandAMemoryOperand(CI, IA, OpVal: I, TLI, TRI))
5620 return true;
5621 continue;
5622 }
5623
5624 if (FindAllMemoryUses(I: UserI, MemoryUses, ConsideredInsts, TLI, TRI, OptSize,
5625 PSI, BFI, SeenInsts))
5626 return true;
5627 }
5628
5629 return false;
5630}
5631
5632static bool FindAllMemoryUses(
5633 Instruction *I, SmallVectorImpl<std::pair<Use *, Type *>> &MemoryUses,
5634 const TargetLowering &TLI, const TargetRegisterInfo &TRI, bool OptSize,
5635 ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) {
5636 unsigned SeenInsts = 0;
5637 SmallPtrSet<Instruction *, 16> ConsideredInsts;
5638 return FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI, OptSize,
5639 PSI, BFI, SeenInsts);
5640}
5641
5642
5643/// Return true if Val is already known to be live at the use site that we're
5644/// folding it into. If so, there is no cost to include it in the addressing
5645/// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
5646/// instruction already.
5647bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,
5648 Value *KnownLive1,
5649 Value *KnownLive2) {
5650 // If Val is either of the known-live values, we know it is live!
5651 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
5652 return true;
5653
5654 // All values other than instructions and arguments (e.g. constants) are live.
5655 if (!isa<Instruction>(Val) && !isa<Argument>(Val))
5656 return true;
5657
5658 // If Val is a constant sized alloca in the entry block, it is live, this is
5659 // true because it is just a reference to the stack/frame pointer, which is
5660 // live for the whole function.
5661 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
5662 if (AI->isStaticAlloca())
5663 return true;
5664
5665 // Check to see if this value is already used in the memory instruction's
5666 // block. If so, it's already live into the block at the very least, so we
5667 // can reasonably fold it.
5668 return Val->isUsedInBasicBlock(BB: MemoryInst->getParent());
5669}
5670
5671/// It is possible for the addressing mode of the machine to fold the specified
5672/// instruction into a load or store that ultimately uses it.
5673/// However, the specified instruction has multiple uses.
5674/// Given this, it may actually increase register pressure to fold it
5675/// into the load. For example, consider this code:
5676///
5677/// X = ...
5678/// Y = X+1
5679/// use(Y) -> nonload/store
5680/// Z = Y+1
5681/// load Z
5682///
5683/// In this case, Y has multiple uses, and can be folded into the load of Z
5684/// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
5685/// be live at the use(Y) line. If we don't fold Y into load Z, we use one
5686/// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
5687/// number of computations either.
5688///
5689/// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
5690/// X was live across 'load Z' for other reasons, we actually *would* want to
5691/// fold the addressing mode in the Z case. This would make Y die earlier.
5692bool AddressingModeMatcher::isProfitableToFoldIntoAddressingMode(
5693 Instruction *I, ExtAddrMode &AMBefore, ExtAddrMode &AMAfter) {
5694 if (IgnoreProfitability)
5695 return true;
5696
5697 // AMBefore is the addressing mode before this instruction was folded into it,
5698 // and AMAfter is the addressing mode after the instruction was folded. Get
5699 // the set of registers referenced by AMAfter and subtract out those
5700 // referenced by AMBefore: this is the set of values which folding in this
5701 // address extends the lifetime of.
5702 //
5703 // Note that there are only two potential values being referenced here,
5704 // BaseReg and ScaleReg (global addresses are always available, as are any
5705 // folded immediates).
5706 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
5707
5708 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
5709 // lifetime wasn't extended by adding this instruction.
5710 if (valueAlreadyLiveAtInst(Val: BaseReg, KnownLive1: AMBefore.BaseReg, KnownLive2: AMBefore.ScaledReg))
5711 BaseReg = nullptr;
5712 if (valueAlreadyLiveAtInst(Val: ScaledReg, KnownLive1: AMBefore.BaseReg, KnownLive2: AMBefore.ScaledReg))
5713 ScaledReg = nullptr;
5714
5715 // If folding this instruction (and it's subexprs) didn't extend any live
5716 // ranges, we're ok with it.
5717 if (!BaseReg && !ScaledReg)
5718 return true;
5719
5720 // If all uses of this instruction can have the address mode sunk into them,
5721 // we can remove the addressing mode and effectively trade one live register
5722 // for another (at worst.) In this context, folding an addressing mode into
5723 // the use is just a particularly nice way of sinking it.
5724 SmallVector<std::pair<Use *, Type *>, 16> MemoryUses;
5725 if (FindAllMemoryUses(I, MemoryUses, TLI, TRI, OptSize, PSI, BFI))
5726 return false; // Has a non-memory, non-foldable use!
5727
5728 // Now that we know that all uses of this instruction are part of a chain of
5729 // computation involving only operations that could theoretically be folded
5730 // into a memory use, loop over each of these memory operation uses and see
5731 // if they could *actually* fold the instruction. The assumption is that
5732 // addressing modes are cheap and that duplicating the computation involved
5733 // many times is worthwhile, even on a fastpath. For sinking candidates
5734 // (i.e. cold call sites), this serves as a way to prevent excessive code
5735 // growth since most architectures have some reasonable small and fast way to
5736 // compute an effective address. (i.e LEA on x86)
5737 SmallVector<Instruction *, 32> MatchedAddrModeInsts;
5738 for (const std::pair<Use *, Type *> &Pair : MemoryUses) {
5739 Value *Address = Pair.first->get();
5740 Instruction *UserI = cast<Instruction>(Val: Pair.first->getUser());
5741 Type *AddressAccessTy = Pair.second;
5742 unsigned AS = Address->getType()->getPointerAddressSpace();
5743
5744 // Do a match against the root of this address, ignoring profitability. This
5745 // will tell us if the addressing mode for the memory operation will
5746 // *actually* cover the shared instruction.
5747 ExtAddrMode Result;
5748 std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
5749 0);
5750 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5751 TPT.getRestorationPoint();
5752 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, TRI, LI, getDTFn,
5753 AddressAccessTy, AS, UserI, Result,
5754 InsertedInsts, PromotedInsts, TPT,
5755 LargeOffsetGEP, OptSize, PSI, BFI);
5756 Matcher.IgnoreProfitability = true;
5757 bool Success = Matcher.matchAddr(Addr: Address, Depth: 0);
5758 (void)Success;
5759 assert(Success && "Couldn't select *anything*?");
5760
5761 // The match was to check the profitability, the changes made are not
5762 // part of the original matcher. Therefore, they should be dropped
5763 // otherwise the original matcher will not present the right state.
5764 TPT.rollback(Point: LastKnownGood);
5765
5766 // If the match didn't cover I, then it won't be shared by it.
5767 if (!is_contained(Range&: MatchedAddrModeInsts, Element: I))
5768 return false;
5769
5770 MatchedAddrModeInsts.clear();
5771 }
5772
5773 return true;
5774}
5775
5776/// Return true if the specified values are defined in a
5777/// different basic block than BB.
5778static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
5779 if (Instruction *I = dyn_cast<Instruction>(Val: V))
5780 return I->getParent() != BB;
5781 return false;
5782}
5783
5784// Find an insert position of Addr for MemoryInst. We can't guarantee MemoryInst
5785// is the first instruction that will use Addr. So we need to find the first
5786// user of Addr in current BB.
5787static BasicBlock::iterator findInsertPos(Value *Addr, Instruction *MemoryInst,
5788 Value *SunkAddr) {
5789 if (Addr->hasOneUse())
5790 return MemoryInst->getIterator();
5791
5792 // We already have a SunkAddr in current BB, but we may need to insert cast
5793 // instruction after it.
5794 if (SunkAddr) {
5795 if (Instruction *AddrInst = dyn_cast<Instruction>(Val: SunkAddr))
5796 return std::next(x: AddrInst->getIterator());
5797 }
5798
5799 // Find the first user of Addr in current BB.
5800 Instruction *Earliest = MemoryInst;
5801 for (User *U : Addr->users()) {
5802 Instruction *UserInst = dyn_cast<Instruction>(Val: U);
5803 if (UserInst && UserInst->getParent() == MemoryInst->getParent()) {
5804 if (isa<PHINode>(Val: UserInst) || UserInst->isDebugOrPseudoInst())
5805 continue;
5806 if (UserInst->comesBefore(Other: Earliest))
5807 Earliest = UserInst;
5808 }
5809 }
5810 return Earliest->getIterator();
5811}
5812
5813/// Sink addressing mode computation immediate before MemoryInst if doing so
5814/// can be done without increasing register pressure. The need for the
5815/// register pressure constraint means this can end up being an all or nothing
5816/// decision for all uses of the same addressing computation.
5817///
5818/// Load and Store Instructions often have addressing modes that can do
5819/// significant amounts of computation. As such, instruction selection will try
5820/// to get the load or store to do as much computation as possible for the
5821/// program. The problem is that isel can only see within a single block. As
5822/// such, we sink as much legal addressing mode work into the block as possible.
5823///
5824/// This method is used to optimize both load/store and inline asms with memory
5825/// operands. It's also used to sink addressing computations feeding into cold
5826/// call sites into their (cold) basic block.
5827///
5828/// The motivation for handling sinking into cold blocks is that doing so can
5829/// both enable other address mode sinking (by satisfying the register pressure
5830/// constraint above), and reduce register pressure globally (by removing the
5831/// addressing mode computation from the fast path entirely.).
5832bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
5833 Type *AccessTy, unsigned AddrSpace) {
5834 Value *Repl = Addr;
5835
5836 // Try to collapse single-value PHI nodes. This is necessary to undo
5837 // unprofitable PRE transformations.
5838 SmallVector<Value *, 8> worklist;
5839 SmallPtrSet<Value *, 16> Visited;
5840 worklist.push_back(Elt: Addr);
5841
5842 // Use a worklist to iteratively look through PHI and select nodes, and
5843 // ensure that the addressing mode obtained from the non-PHI/select roots of
5844 // the graph are compatible.
5845 bool PhiOrSelectSeen = false;
5846 SmallVector<Instruction *, 16> AddrModeInsts;
5847 AddressingModeCombiner AddrModes(*DL, Addr);
5848 TypePromotionTransaction TPT(RemovedInsts);
5849 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5850 TPT.getRestorationPoint();
5851 while (!worklist.empty()) {
5852 Value *V = worklist.pop_back_val();
5853
5854 // We allow traversing cyclic Phi nodes.
5855 // In case of success after this loop we ensure that traversing through
5856 // Phi nodes ends up with all cases to compute address of the form
5857 // BaseGV + Base + Scale * Index + Offset
5858 // where Scale and Offset are constans and BaseGV, Base and Index
5859 // are exactly the same Values in all cases.
5860 // It means that BaseGV, Scale and Offset dominate our memory instruction
5861 // and have the same value as they had in address computation represented
5862 // as Phi. So we can safely sink address computation to memory instruction.
5863 if (!Visited.insert(Ptr: V).second)
5864 continue;
5865
5866 // For a PHI node, push all of its incoming values.
5867 if (PHINode *P = dyn_cast<PHINode>(Val: V)) {
5868 append_range(C&: worklist, R: P->incoming_values());
5869 PhiOrSelectSeen = true;
5870 continue;
5871 }
5872 // Similar for select.
5873 if (SelectInst *SI = dyn_cast<SelectInst>(Val: V)) {
5874 worklist.push_back(Elt: SI->getFalseValue());
5875 worklist.push_back(Elt: SI->getTrueValue());
5876 PhiOrSelectSeen = true;
5877 continue;
5878 }
5879
5880 // For non-PHIs, determine the addressing mode being computed. Note that
5881 // the result may differ depending on what other uses our candidate
5882 // addressing instructions might have.
5883 AddrModeInsts.clear();
5884 std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
5885 0);
5886 // Defer the query (and possible computation of) the dom tree to point of
5887 // actual use. It's expected that most address matches don't actually need
5888 // the domtree.
5889 auto getDTFn = [MemoryInst, this]() -> const DominatorTree & {
5890 Function *F = MemoryInst->getParent()->getParent();
5891 return this->getDT(F&: *F);
5892 };
5893 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
5894 V, AccessTy, AS: AddrSpace, MemoryInst, AddrModeInsts, TLI: *TLI, LI: *LI, getDTFn,
5895 TRI: *TRI, InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP, OptSize, PSI,
5896 BFI);
5897
5898 GetElementPtrInst *GEP = LargeOffsetGEP.first;
5899 if (GEP && !NewGEPBases.count(V: GEP)) {
5900 // If splitting the underlying data structure can reduce the offset of a
5901 // GEP, collect the GEP. Skip the GEPs that are the new bases of
5902 // previously split data structures.
5903 LargeOffsetGEPMap[GEP->getPointerOperand()].push_back(Elt: LargeOffsetGEP);
5904 LargeOffsetGEPID.insert(KV: std::make_pair(x&: GEP, y: LargeOffsetGEPID.size()));
5905 }
5906
5907 NewAddrMode.OriginalValue = V;
5908 if (!AddrModes.addNewAddrMode(NewAddrMode))
5909 break;
5910 }
5911
5912 // Try to combine the AddrModes we've collected. If we couldn't collect any,
5913 // or we have multiple but either couldn't combine them or combining them
5914 // wouldn't do anything useful, bail out now.
5915 if (!AddrModes.combineAddrModes()) {
5916 TPT.rollback(Point: LastKnownGood);
5917 return false;
5918 }
5919 bool Modified = TPT.commit();
5920
5921 // Get the combined AddrMode (or the only AddrMode, if we only had one).
5922 ExtAddrMode AddrMode = AddrModes.getAddrMode();
5923
5924 // If all the instructions matched are already in this BB, don't do anything.
5925 // If we saw a Phi node then it is not local definitely, and if we saw a
5926 // select then we want to push the address calculation past it even if it's
5927 // already in this BB.
5928 if (!PhiOrSelectSeen && none_of(Range&: AddrModeInsts, P: [&](Value *V) {
5929 return IsNonLocalValue(V, BB: MemoryInst->getParent());
5930 })) {
5931 LLVM_DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode
5932 << "\n");
5933 return Modified;
5934 }
5935
5936 // Now that we determined the addressing expression we want to use and know
5937 // that we have to sink it into this block. Check to see if we have already
5938 // done this for some other load/store instr in this block. If so, reuse
5939 // the computation. Before attempting reuse, check if the address is valid
5940 // as it may have been erased.
5941
5942 WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
5943
5944 Value *SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
5945 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
5946
5947 // The current BB may be optimized multiple times, we can't guarantee the
5948 // reuse of Addr happens later, call findInsertPos to find an appropriate
5949 // insert position.
5950 auto InsertPos = findInsertPos(Addr, MemoryInst, SunkAddr);
5951
5952 // TODO: Adjust insert point considering (Base|Scaled)Reg if possible.
5953 if (!SunkAddr) {
5954 auto &DT = getDT(F&: *MemoryInst->getFunction());
5955 if ((AddrMode.BaseReg && !DT.dominates(Def: AddrMode.BaseReg, User: &*InsertPos)) ||
5956 (AddrMode.ScaledReg && !DT.dominates(Def: AddrMode.ScaledReg, User: &*InsertPos)))
5957 return Modified;
5958 }
5959
5960 IRBuilder<> Builder(MemoryInst->getParent(), InsertPos);
5961
5962 if (SunkAddr) {
5963 LLVM_DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode
5964 << " for " << *MemoryInst << "\n");
5965 if (SunkAddr->getType() != Addr->getType()) {
5966 if (SunkAddr->getType()->getPointerAddressSpace() !=
5967 Addr->getType()->getPointerAddressSpace() &&
5968 !DL->isNonIntegralPointerType(Ty: Addr->getType())) {
5969 // There are two reasons the address spaces might not match: a no-op
5970 // addrspacecast, or a ptrtoint/inttoptr pair. Either way, we emit a
5971 // ptrtoint/inttoptr pair to ensure we match the original semantics.
5972 // TODO: allow bitcast between different address space pointers with the
5973 // same size.
5974 SunkAddr = Builder.CreatePtrToInt(V: SunkAddr, DestTy: IntPtrTy, Name: "sunkaddr");
5975 SunkAddr =
5976 Builder.CreateIntToPtr(V: SunkAddr, DestTy: Addr->getType(), Name: "sunkaddr");
5977 } else
5978 SunkAddr = Builder.CreatePointerCast(V: SunkAddr, DestTy: Addr->getType());
5979 }
5980 } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() &&
5981 SubtargetInfo->addrSinkUsingGEPs())) {
5982 // By default, we use the GEP-based method when AA is used later. This
5983 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
5984 LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
5985 << " for " << *MemoryInst << "\n");
5986 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
5987
5988 // First, find the pointer.
5989 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
5990 ResultPtr = AddrMode.BaseReg;
5991 AddrMode.BaseReg = nullptr;
5992 }
5993
5994 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
5995 // We can't add more than one pointer together, nor can we scale a
5996 // pointer (both of which seem meaningless).
5997 if (ResultPtr || AddrMode.Scale != 1)
5998 return Modified;
5999
6000 ResultPtr = AddrMode.ScaledReg;
6001 AddrMode.Scale = 0;
6002 }
6003
6004 // It is only safe to sign extend the BaseReg if we know that the math
6005 // required to create it did not overflow before we extend it. Since
6006 // the original IR value was tossed in favor of a constant back when
6007 // the AddrMode was created we need to bail out gracefully if widths
6008 // do not match instead of extending it.
6009 //
6010 // (See below for code to add the scale.)
6011 if (AddrMode.Scale) {
6012 Type *ScaledRegTy = AddrMode.ScaledReg->getType();
6013 if (cast<IntegerType>(Val: IntPtrTy)->getBitWidth() >
6014 cast<IntegerType>(Val: ScaledRegTy)->getBitWidth())
6015 return Modified;
6016 }
6017
6018 GlobalValue *BaseGV = AddrMode.BaseGV;
6019 if (BaseGV != nullptr) {
6020 if (ResultPtr)
6021 return Modified;
6022
6023 if (BaseGV->isThreadLocal()) {
6024 ResultPtr = Builder.CreateThreadLocalAddress(Ptr: BaseGV);
6025 } else {
6026 ResultPtr = BaseGV;
6027 }
6028 }
6029
6030 // If the real base value actually came from an inttoptr, then the matcher
6031 // will look through it and provide only the integer value. In that case,
6032 // use it here.
6033 if (!DL->isNonIntegralPointerType(Ty: Addr->getType())) {
6034 if (!ResultPtr && AddrMode.BaseReg) {
6035 ResultPtr = Builder.CreateIntToPtr(V: AddrMode.BaseReg, DestTy: Addr->getType(),
6036 Name: "sunkaddr");
6037 AddrMode.BaseReg = nullptr;
6038 } else if (!ResultPtr && AddrMode.Scale == 1) {
6039 ResultPtr = Builder.CreateIntToPtr(V: AddrMode.ScaledReg, DestTy: Addr->getType(),
6040 Name: "sunkaddr");
6041 AddrMode.Scale = 0;
6042 }
6043 }
6044
6045 if (!ResultPtr && !AddrMode.BaseReg && !AddrMode.Scale &&
6046 !AddrMode.BaseOffs) {
6047 SunkAddr = Constant::getNullValue(Ty: Addr->getType());
6048 } else if (!ResultPtr) {
6049 return Modified;
6050 } else {
6051 Type *I8PtrTy =
6052 Builder.getPtrTy(AddrSpace: Addr->getType()->getPointerAddressSpace());
6053
6054 // Start with the base register. Do this first so that subsequent address
6055 // matching finds it last, which will prevent it from trying to match it
6056 // as the scaled value in case it happens to be a mul. That would be
6057 // problematic if we've sunk a different mul for the scale, because then
6058 // we'd end up sinking both muls.
6059 if (AddrMode.BaseReg) {
6060 Value *V = AddrMode.BaseReg;
6061 if (V->getType() != IntPtrTy)
6062 V = Builder.CreateIntCast(V, DestTy: IntPtrTy, /*isSigned=*/true, Name: "sunkaddr");
6063
6064 ResultIndex = V;
6065 }
6066
6067 // Add the scale value.
6068 if (AddrMode.Scale) {
6069 Value *V = AddrMode.ScaledReg;
6070 if (V->getType() == IntPtrTy) {
6071 // done.
6072 } else {
6073 assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <
6074 cast<IntegerType>(V->getType())->getBitWidth() &&
6075 "We can't transform if ScaledReg is too narrow");
6076 V = Builder.CreateTrunc(V, DestTy: IntPtrTy, Name: "sunkaddr");
6077 }
6078
6079 if (AddrMode.Scale != 1)
6080 V = Builder.CreateMul(
6081 LHS: V, RHS: ConstantInt::getSigned(Ty: IntPtrTy, V: AddrMode.Scale), Name: "sunkaddr");
6082 if (ResultIndex)
6083 ResultIndex = Builder.CreateAdd(LHS: ResultIndex, RHS: V, Name: "sunkaddr");
6084 else
6085 ResultIndex = V;
6086 }
6087
6088 // Add in the Base Offset if present.
6089 if (AddrMode.BaseOffs) {
6090 Value *V = ConstantInt::getSigned(Ty: IntPtrTy, V: AddrMode.BaseOffs);
6091 if (ResultIndex) {
6092 // We need to add this separately from the scale above to help with
6093 // SDAG consecutive load/store merging.
6094 if (ResultPtr->getType() != I8PtrTy)
6095 ResultPtr = Builder.CreatePointerCast(V: ResultPtr, DestTy: I8PtrTy);
6096 ResultPtr = Builder.CreatePtrAdd(Ptr: ResultPtr, Offset: ResultIndex, Name: "sunkaddr",
6097 NW: AddrMode.InBounds);
6098 }
6099
6100 ResultIndex = V;
6101 }
6102
6103 if (!ResultIndex) {
6104 auto PtrInst = dyn_cast<Instruction>(Val: ResultPtr);
6105 // We know that we have a pointer without any offsets. If this pointer
6106 // originates from a different basic block than the current one, we
6107 // must be able to recreate it in the current basic block.
6108 // We do not support the recreation of any instructions yet.
6109 if (PtrInst && PtrInst->getParent() != MemoryInst->getParent())
6110 return Modified;
6111 SunkAddr = ResultPtr;
6112 } else {
6113 if (ResultPtr->getType() != I8PtrTy)
6114 ResultPtr = Builder.CreatePointerCast(V: ResultPtr, DestTy: I8PtrTy);
6115 SunkAddr = Builder.CreatePtrAdd(Ptr: ResultPtr, Offset: ResultIndex, Name: "sunkaddr",
6116 NW: AddrMode.InBounds);
6117 }
6118
6119 if (SunkAddr->getType() != Addr->getType()) {
6120 if (SunkAddr->getType()->getPointerAddressSpace() !=
6121 Addr->getType()->getPointerAddressSpace() &&
6122 !DL->isNonIntegralPointerType(Ty: Addr->getType())) {
6123 // There are two reasons the address spaces might not match: a no-op
6124 // addrspacecast, or a ptrtoint/inttoptr pair. Either way, we emit a
6125 // ptrtoint/inttoptr pair to ensure we match the original semantics.
6126 // TODO: allow bitcast between different address space pointers with
6127 // the same size.
6128 SunkAddr = Builder.CreatePtrToInt(V: SunkAddr, DestTy: IntPtrTy, Name: "sunkaddr");
6129 SunkAddr =
6130 Builder.CreateIntToPtr(V: SunkAddr, DestTy: Addr->getType(), Name: "sunkaddr");
6131 } else
6132 SunkAddr = Builder.CreatePointerCast(V: SunkAddr, DestTy: Addr->getType());
6133 }
6134 }
6135 } else {
6136 // We'd require a ptrtoint/inttoptr down the line, which we can't do for
6137 // non-integral pointers, so in that case bail out now.
6138 Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
6139 Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
6140 PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(Val: BaseTy);
6141 PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(Val: ScaleTy);
6142 if (DL->isNonIntegralPointerType(Ty: Addr->getType()) ||
6143 (BasePtrTy && DL->isNonIntegralPointerType(PT: BasePtrTy)) ||
6144 (ScalePtrTy && DL->isNonIntegralPointerType(PT: ScalePtrTy)) ||
6145 (AddrMode.BaseGV &&
6146 DL->isNonIntegralPointerType(PT: AddrMode.BaseGV->getType())))
6147 return Modified;
6148
6149 LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
6150 << " for " << *MemoryInst << "\n");
6151 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
6152 Value *Result = nullptr;
6153
6154 // Start with the base register. Do this first so that subsequent address
6155 // matching finds it last, which will prevent it from trying to match it
6156 // as the scaled value in case it happens to be a mul. That would be
6157 // problematic if we've sunk a different mul for the scale, because then
6158 // we'd end up sinking both muls.
6159 if (AddrMode.BaseReg) {
6160 Value *V = AddrMode.BaseReg;
6161 if (V->getType()->isPointerTy())
6162 V = Builder.CreatePtrToInt(V, DestTy: IntPtrTy, Name: "sunkaddr");
6163 if (V->getType() != IntPtrTy)
6164 V = Builder.CreateIntCast(V, DestTy: IntPtrTy, /*isSigned=*/true, Name: "sunkaddr");
6165 Result = V;
6166 }
6167
6168 // Add the scale value.
6169 if (AddrMode.Scale) {
6170 Value *V = AddrMode.ScaledReg;
6171 if (V->getType() == IntPtrTy) {
6172 // done.
6173 } else if (V->getType()->isPointerTy()) {
6174 V = Builder.CreatePtrToInt(V, DestTy: IntPtrTy, Name: "sunkaddr");
6175 } else if (cast<IntegerType>(Val: IntPtrTy)->getBitWidth() <
6176 cast<IntegerType>(Val: V->getType())->getBitWidth()) {
6177 V = Builder.CreateTrunc(V, DestTy: IntPtrTy, Name: "sunkaddr");
6178 } else {
6179 // It is only safe to sign extend the BaseReg if we know that the math
6180 // required to create it did not overflow before we extend it. Since
6181 // the original IR value was tossed in favor of a constant back when
6182 // the AddrMode was created we need to bail out gracefully if widths
6183 // do not match instead of extending it.
6184 Instruction *I = dyn_cast_or_null<Instruction>(Val: Result);
6185 if (I && (Result != AddrMode.BaseReg))
6186 I->eraseFromParent();
6187 return Modified;
6188 }
6189 if (AddrMode.Scale != 1)
6190 V = Builder.CreateMul(
6191 LHS: V, RHS: ConstantInt::getSigned(Ty: IntPtrTy, V: AddrMode.Scale), Name: "sunkaddr");
6192 if (Result)
6193 Result = Builder.CreateAdd(LHS: Result, RHS: V, Name: "sunkaddr");
6194 else
6195 Result = V;
6196 }
6197
6198 // Add in the BaseGV if present.
6199 GlobalValue *BaseGV = AddrMode.BaseGV;
6200 if (BaseGV != nullptr) {
6201 Value *BaseGVPtr;
6202 if (BaseGV->isThreadLocal()) {
6203 BaseGVPtr = Builder.CreateThreadLocalAddress(Ptr: BaseGV);
6204 } else {
6205 BaseGVPtr = BaseGV;
6206 }
6207 Value *V = Builder.CreatePtrToInt(V: BaseGVPtr, DestTy: IntPtrTy, Name: "sunkaddr");
6208 if (Result)
6209 Result = Builder.CreateAdd(LHS: Result, RHS: V, Name: "sunkaddr");
6210 else
6211 Result = V;
6212 }
6213
6214 // Add in the Base Offset if present.
6215 if (AddrMode.BaseOffs) {
6216 Value *V = ConstantInt::getSigned(Ty: IntPtrTy, V: AddrMode.BaseOffs);
6217 if (Result)
6218 Result = Builder.CreateAdd(LHS: Result, RHS: V, Name: "sunkaddr");
6219 else
6220 Result = V;
6221 }
6222
6223 if (!Result)
6224 SunkAddr = Constant::getNullValue(Ty: Addr->getType());
6225 else
6226 SunkAddr = Builder.CreateIntToPtr(V: Result, DestTy: Addr->getType(), Name: "sunkaddr");
6227 }
6228
6229 MemoryInst->replaceUsesOfWith(From: Repl, To: SunkAddr);
6230 // Store the newly computed address into the cache. In the case we reused a
6231 // value, this should be idempotent.
6232 SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
6233
6234 // If we have no uses, recursively delete the value and all dead instructions
6235 // using it.
6236 if (Repl->use_empty()) {
6237 resetIteratorIfInvalidatedWhileCalling(BB: CurInstIterator->getParent(), f: [&]() {
6238 RecursivelyDeleteTriviallyDeadInstructions(
6239 V: Repl, TLI: TLInfo, MSSAU: nullptr,
6240 AboutToDeleteCallback: [&](Value *V) { removeAllAssertingVHReferences(V); });
6241 });
6242 }
6243 ++NumMemoryInsts;
6244 return true;
6245}
6246
6247/// Rewrite GEP input to gather/scatter to enable SelectionDAGBuilder to find
6248/// a uniform base to use for ISD::MGATHER/MSCATTER. SelectionDAGBuilder can
6249/// only handle a 2 operand GEP in the same basic block or a splat constant
6250/// vector. The 2 operands to the GEP must have a scalar pointer and a vector
6251/// index.
6252///
6253/// If the existing GEP has a vector base pointer that is splat, we can look
6254/// through the splat to find the scalar pointer. If we can't find a scalar
6255/// pointer there's nothing we can do.
6256///
6257/// If we have a GEP with more than 2 indices where the middle indices are all
6258/// zeroes, we can replace it with 2 GEPs where the second has 2 operands.
6259///
6260/// If the final index isn't a vector or is a splat, we can emit a scalar GEP
6261/// followed by a GEP with an all zeroes vector index. This will enable
6262/// SelectionDAGBuilder to use the scalar GEP as the uniform base and have a
6263/// zero index.
6264bool CodeGenPrepare::optimizeGatherScatterInst(Instruction *MemoryInst,
6265 Value *Ptr) {
6266 Value *NewAddr;
6267
6268 if (const auto *GEP = dyn_cast<GetElementPtrInst>(Val: Ptr)) {
6269 // Don't optimize GEPs that don't have indices.
6270 if (!GEP->hasIndices())
6271 return false;
6272
6273 // If the GEP and the gather/scatter aren't in the same BB, don't optimize.
6274 // FIXME: We should support this by sinking the GEP.
6275 if (MemoryInst->getParent() != GEP->getParent())
6276 return false;
6277
6278 SmallVector<Value *, 2> Ops(GEP->operands());
6279
6280 bool RewriteGEP = false;
6281
6282 if (Ops[0]->getType()->isVectorTy()) {
6283 Ops[0] = getSplatValue(V: Ops[0]);
6284 if (!Ops[0])
6285 return false;
6286 RewriteGEP = true;
6287 }
6288
6289 unsigned FinalIndex = Ops.size() - 1;
6290
6291 // Ensure all but the last index is 0.
6292 // FIXME: This isn't strictly required. All that's required is that they are
6293 // all scalars or splats.
6294 for (unsigned i = 1; i < FinalIndex; ++i) {
6295 auto *C = dyn_cast<Constant>(Val: Ops[i]);
6296 if (!C)
6297 return false;
6298 if (isa<VectorType>(Val: C->getType()))
6299 C = C->getSplatValue();
6300 auto *CI = dyn_cast_or_null<ConstantInt>(Val: C);
6301 if (!CI || !CI->isZero())
6302 return false;
6303 // Scalarize the index if needed.
6304 Ops[i] = CI;
6305 }
6306
6307 // Try to scalarize the final index.
6308 if (Ops[FinalIndex]->getType()->isVectorTy()) {
6309 if (Value *V = getSplatValue(V: Ops[FinalIndex])) {
6310 auto *C = dyn_cast<ConstantInt>(Val: V);
6311 // Don't scalarize all zeros vector.
6312 if (!C || !C->isZero()) {
6313 Ops[FinalIndex] = V;
6314 RewriteGEP = true;
6315 }
6316 }
6317 }
6318
6319 // If we made any changes or the we have extra operands, we need to generate
6320 // new instructions.
6321 if (!RewriteGEP && Ops.size() == 2)
6322 return false;
6323
6324 auto NumElts = cast<VectorType>(Val: Ptr->getType())->getElementCount();
6325
6326 IRBuilder<> Builder(MemoryInst);
6327
6328 Type *SourceTy = GEP->getSourceElementType();
6329 Type *ScalarIndexTy = DL->getIndexType(PtrTy: Ops[0]->getType()->getScalarType());
6330
6331 // If the final index isn't a vector, emit a scalar GEP containing all ops
6332 // and a vector GEP with all zeroes final index.
6333 if (!Ops[FinalIndex]->getType()->isVectorTy()) {
6334 NewAddr = Builder.CreateGEP(Ty: SourceTy, Ptr: Ops[0], IdxList: ArrayRef(Ops).drop_front());
6335 auto *IndexTy = VectorType::get(ElementType: ScalarIndexTy, EC: NumElts);
6336 auto *SecondTy = GetElementPtrInst::getIndexedType(
6337 Ty: SourceTy, IdxList: ArrayRef(Ops).drop_front());
6338 NewAddr =
6339 Builder.CreateGEP(Ty: SecondTy, Ptr: NewAddr, IdxList: Constant::getNullValue(Ty: IndexTy));
6340 } else {
6341 Value *Base = Ops[0];
6342 Value *Index = Ops[FinalIndex];
6343
6344 // Create a scalar GEP if there are more than 2 operands.
6345 if (Ops.size() != 2) {
6346 // Replace the last index with 0.
6347 Ops[FinalIndex] =
6348 Constant::getNullValue(Ty: Ops[FinalIndex]->getType()->getScalarType());
6349 Base = Builder.CreateGEP(Ty: SourceTy, Ptr: Base, IdxList: ArrayRef(Ops).drop_front());
6350 SourceTy = GetElementPtrInst::getIndexedType(
6351 Ty: SourceTy, IdxList: ArrayRef(Ops).drop_front());
6352 }
6353
6354 // Now create the GEP with scalar pointer and vector index.
6355 NewAddr = Builder.CreateGEP(Ty: SourceTy, Ptr: Base, IdxList: Index);
6356 }
6357 } else if (!isa<Constant>(Val: Ptr)) {
6358 // Not a GEP, maybe its a splat and we can create a GEP to enable
6359 // SelectionDAGBuilder to use it as a uniform base.
6360 Value *V = getSplatValue(V: Ptr);
6361 if (!V)
6362 return false;
6363
6364 auto NumElts = cast<VectorType>(Val: Ptr->getType())->getElementCount();
6365
6366 IRBuilder<> Builder(MemoryInst);
6367
6368 // Emit a vector GEP with a scalar pointer and all 0s vector index.
6369 Type *ScalarIndexTy = DL->getIndexType(PtrTy: V->getType()->getScalarType());
6370 auto *IndexTy = VectorType::get(ElementType: ScalarIndexTy, EC: NumElts);
6371 Type *ScalarTy;
6372 if (cast<IntrinsicInst>(Val: MemoryInst)->getIntrinsicID() ==
6373 Intrinsic::masked_gather) {
6374 ScalarTy = MemoryInst->getType()->getScalarType();
6375 } else {
6376 assert(cast<IntrinsicInst>(MemoryInst)->getIntrinsicID() ==
6377 Intrinsic::masked_scatter);
6378 ScalarTy = MemoryInst->getOperand(i: 0)->getType()->getScalarType();
6379 }
6380 NewAddr = Builder.CreateGEP(Ty: ScalarTy, Ptr: V, IdxList: Constant::getNullValue(Ty: IndexTy));
6381 } else {
6382 // Constant, SelectionDAGBuilder knows to check if its a splat.
6383 return false;
6384 }
6385
6386 MemoryInst->replaceUsesOfWith(From: Ptr, To: NewAddr);
6387
6388 // If we have no uses, recursively delete the value and all dead instructions
6389 // using it.
6390 if (Ptr->use_empty())
6391 RecursivelyDeleteTriviallyDeadInstructions(
6392 V: Ptr, TLI: TLInfo, MSSAU: nullptr,
6393 AboutToDeleteCallback: [&](Value *V) { removeAllAssertingVHReferences(V); });
6394
6395 return true;
6396}
6397
6398// This is a helper for CodeGenPrepare::optimizeMulWithOverflow.
6399// Check the pattern we are interested in where there are maximum 2 uses
6400// of the intrinsic which are the extract instructions.
6401static bool matchOverflowPattern(Instruction *&I, ExtractValueInst *&MulExtract,
6402 ExtractValueInst *&OverflowExtract) {
6403 // Bail out if it's more than 2 users:
6404 if (I->hasNUsesOrMore(N: 3))
6405 return false;
6406
6407 for (User *U : I->users()) {
6408 auto *Extract = dyn_cast<ExtractValueInst>(Val: U);
6409 if (!Extract || Extract->getNumIndices() != 1)
6410 return false;
6411
6412 unsigned Index = Extract->getIndices()[0];
6413 if (Index == 0)
6414 MulExtract = Extract;
6415 else if (Index == 1)
6416 OverflowExtract = Extract;
6417 else
6418 return false;
6419 }
6420 return true;
6421}
6422
6423// Rewrite the mul_with_overflow intrinsic by checking if both of the
6424// operands' value ranges are within the legal type. If so, we can optimize the
6425// multiplication algorithm. This code is supposed to be written during the step
6426// of type legalization, but given that we need to reconstruct the IR which is
6427// not doable there, we do it here.
6428// The IR after the optimization will look like:
6429// entry:
6430// if signed:
6431// ( (lhs_lo>>BW-1) ^ lhs_hi) || ( (rhs_lo>>BW-1) ^ rhs_hi) ? overflow,
6432// overflow_no
6433// else:
6434// (lhs_hi != 0) || (rhs_hi != 0) ? overflow, overflow_no
6435// overflow_no:
6436// overflow:
6437// overflow.res:
6438// \returns true if optimization was applied
6439// TODO: This optimization can be further improved to optimize branching on
6440// overflow where the 'overflow_no' BB can branch directly to the false
6441// successor of overflow, but that would add additional complexity so we leave
6442// it for future work.
6443bool CodeGenPrepare::optimizeMulWithOverflow(Instruction *I, bool IsSigned,
6444 ModifyDT &ModifiedDT) {
6445 // Check if target supports this optimization.
6446 if (!TLI->shouldOptimizeMulOverflowWithZeroHighBits(
6447 Context&: I->getContext(),
6448 VT: TLI->getValueType(DL: *DL, Ty: I->getType()->getContainedType(i: 0))))
6449 return false;
6450
6451 ExtractValueInst *MulExtract = nullptr, *OverflowExtract = nullptr;
6452 if (!matchOverflowPattern(I, MulExtract, OverflowExtract))
6453 return false;
6454
6455 // Keep track of the instruction to stop reoptimizing it again.
6456 InsertedInsts.insert(Ptr: I);
6457
6458 Value *LHS = I->getOperand(i: 0);
6459 Value *RHS = I->getOperand(i: 1);
6460 Type *Ty = LHS->getType();
6461 unsigned VTHalfBitWidth = Ty->getScalarSizeInBits() / 2;
6462 Type *LegalTy = Ty->getWithNewBitWidth(NewBitWidth: VTHalfBitWidth);
6463
6464 // New BBs:
6465 BasicBlock *OverflowEntryBB = I->getParent()->splitBasicBlockBefore(I, BBName: "");
6466 OverflowEntryBB->takeName(V: I->getParent());
6467 // Keep the 'br' instruction that is generated as a result of the split to be
6468 // erased/replaced later.
6469 Instruction *OldTerminator = OverflowEntryBB->getTerminator();
6470 BasicBlock *NoOverflowBB =
6471 BasicBlock::Create(Context&: I->getContext(), Name: "overflow.no", Parent: I->getFunction());
6472 NoOverflowBB->moveAfter(MovePos: OverflowEntryBB);
6473 BasicBlock *OverflowBB =
6474 BasicBlock::Create(Context&: I->getContext(), Name: "overflow", Parent: I->getFunction());
6475 OverflowBB->moveAfter(MovePos: NoOverflowBB);
6476
6477 // BB overflow.entry:
6478 IRBuilder<> Builder(OverflowEntryBB);
6479 // Extract low and high halves of LHS:
6480 Value *LoLHS = Builder.CreateTrunc(V: LHS, DestTy: LegalTy, Name: "lo.lhs");
6481 Value *HiLHS = Builder.CreateLShr(LHS, RHS: VTHalfBitWidth, Name: "lhs.lsr");
6482 HiLHS = Builder.CreateTrunc(V: HiLHS, DestTy: LegalTy, Name: "hi.lhs");
6483
6484 // Extract low and high halves of RHS:
6485 Value *LoRHS = Builder.CreateTrunc(V: RHS, DestTy: LegalTy, Name: "lo.rhs");
6486 Value *HiRHS = Builder.CreateLShr(LHS: RHS, RHS: VTHalfBitWidth, Name: "rhs.lsr");
6487 HiRHS = Builder.CreateTrunc(V: HiRHS, DestTy: LegalTy, Name: "hi.rhs");
6488
6489 Value *IsAnyBitTrue;
6490 if (IsSigned) {
6491 Value *SignLoLHS =
6492 Builder.CreateAShr(LHS: LoLHS, RHS: VTHalfBitWidth - 1, Name: "sign.lo.lhs");
6493 Value *SignLoRHS =
6494 Builder.CreateAShr(LHS: LoRHS, RHS: VTHalfBitWidth - 1, Name: "sign.lo.rhs");
6495 Value *XorLHS = Builder.CreateXor(LHS: HiLHS, RHS: SignLoLHS);
6496 Value *XorRHS = Builder.CreateXor(LHS: HiRHS, RHS: SignLoRHS);
6497 Value *Or = Builder.CreateOr(LHS: XorLHS, RHS: XorRHS, Name: "or.lhs.rhs");
6498 IsAnyBitTrue = Builder.CreateCmp(Pred: ICmpInst::ICMP_NE, LHS: Or,
6499 RHS: ConstantInt::getNullValue(Ty: Or->getType()));
6500 } else {
6501 Value *CmpLHS = Builder.CreateCmp(Pred: ICmpInst::ICMP_NE, LHS: HiLHS,
6502 RHS: ConstantInt::getNullValue(Ty: LegalTy));
6503 Value *CmpRHS = Builder.CreateCmp(Pred: ICmpInst::ICMP_NE, LHS: HiRHS,
6504 RHS: ConstantInt::getNullValue(Ty: LegalTy));
6505 IsAnyBitTrue = Builder.CreateOr(LHS: CmpLHS, RHS: CmpRHS, Name: "or.lhs.rhs");
6506 }
6507 Builder.CreateCondBr(Cond: IsAnyBitTrue, True: OverflowBB, False: NoOverflowBB);
6508
6509 // BB overflow.no:
6510 Builder.SetInsertPoint(NoOverflowBB);
6511 Value *ExtLoLHS, *ExtLoRHS;
6512 if (IsSigned) {
6513 ExtLoLHS = Builder.CreateSExt(V: LoLHS, DestTy: Ty, Name: "lo.lhs.ext");
6514 ExtLoRHS = Builder.CreateSExt(V: LoRHS, DestTy: Ty, Name: "lo.rhs.ext");
6515 } else {
6516 ExtLoLHS = Builder.CreateZExt(V: LoLHS, DestTy: Ty, Name: "lo.lhs.ext");
6517 ExtLoRHS = Builder.CreateZExt(V: LoRHS, DestTy: Ty, Name: "lo.rhs.ext");
6518 }
6519
6520 Value *Mul = Builder.CreateMul(LHS: ExtLoLHS, RHS: ExtLoRHS, Name: "mul.overflow.no");
6521
6522 // Create the 'overflow.res' BB to merge the results of
6523 // the two paths:
6524 BasicBlock *OverflowResBB = I->getParent();
6525 OverflowResBB->setName("overflow.res");
6526
6527 // BB overflow.no: jump to overflow.res BB
6528 Builder.CreateBr(Dest: OverflowResBB);
6529 // No we don't need the old terminator in overflow.entry BB, erase it:
6530 OldTerminator->eraseFromParent();
6531
6532 // BB overflow.res:
6533 Builder.SetInsertPoint(TheBB: OverflowResBB, IP: OverflowResBB->getFirstInsertionPt());
6534 // Create PHI nodes to merge results from no.overflow BB and overflow BB to
6535 // replace the extract instructions.
6536 PHINode *OverflowResPHI = Builder.CreatePHI(Ty, NumReservedValues: 2),
6537 *OverflowFlagPHI =
6538 Builder.CreatePHI(Ty: IntegerType::getInt1Ty(C&: I->getContext()), NumReservedValues: 2);
6539
6540 // Add the incoming values from no.overflow BB and later from overflow BB.
6541 OverflowResPHI->addIncoming(V: Mul, BB: NoOverflowBB);
6542 OverflowFlagPHI->addIncoming(V: ConstantInt::getFalse(Context&: I->getContext()),
6543 BB: NoOverflowBB);
6544
6545 // Replace all users of MulExtract and OverflowExtract to use the PHI nodes.
6546 if (MulExtract) {
6547 MulExtract->replaceAllUsesWith(V: OverflowResPHI);
6548 MulExtract->eraseFromParent();
6549 }
6550 if (OverflowExtract) {
6551 OverflowExtract->replaceAllUsesWith(V: OverflowFlagPHI);
6552 OverflowExtract->eraseFromParent();
6553 }
6554
6555 // Remove the intrinsic from parent (overflow.res BB) as it will be part of
6556 // overflow BB
6557 I->removeFromParent();
6558 // BB overflow:
6559 I->insertInto(ParentBB: OverflowBB, It: OverflowBB->end());
6560 Builder.SetInsertPoint(TheBB: OverflowBB, IP: OverflowBB->end());
6561 Value *MulOverflow = Builder.CreateExtractValue(Agg: I, Idxs: {0}, Name: "mul.overflow");
6562 Value *OverflowFlag = Builder.CreateExtractValue(Agg: I, Idxs: {1}, Name: "overflow.flag");
6563 Builder.CreateBr(Dest: OverflowResBB);
6564
6565 // Add The Extracted values to the PHINodes in the overflow.res BB.
6566 OverflowResPHI->addIncoming(V: MulOverflow, BB: OverflowBB);
6567 OverflowFlagPHI->addIncoming(V: OverflowFlag, BB: OverflowBB);
6568
6569 ModifiedDT = ModifyDT::ModifyBBDT;
6570 return true;
6571}
6572
6573/// If there are any memory operands, use OptimizeMemoryInst to sink their
6574/// address computing into the block when possible / profitable.
6575bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
6576 bool MadeChange = false;
6577
6578 const TargetRegisterInfo *TRI =
6579 TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
6580 TargetLowering::AsmOperandInfoVector TargetConstraints =
6581 TLI->ParseConstraints(DL: *DL, TRI, Call: *CS);
6582 unsigned ArgNo = 0;
6583 for (TargetLowering::AsmOperandInfo &OpInfo : TargetConstraints) {
6584 // Compute the constraint code and ConstraintType to use.
6585 TLI->ComputeConstraintToUse(OpInfo, Op: SDValue());
6586
6587 // TODO: Also handle C_Address?
6588 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6589 OpInfo.isIndirect) {
6590 Value *OpVal = CS->getArgOperand(i: ArgNo++);
6591 MadeChange |= optimizeMemoryInst(MemoryInst: CS, Addr: OpVal, AccessTy: OpVal->getType(), AddrSpace: ~0u);
6592 } else if (OpInfo.Type == InlineAsm::isInput)
6593 ArgNo++;
6594 }
6595
6596 return MadeChange;
6597}
6598
6599/// Check if all the uses of \p Val are equivalent (or free) zero or
6600/// sign extensions.
6601static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
6602 assert(!Val->use_empty() && "Input must have at least one use");
6603 const Instruction *FirstUser = cast<Instruction>(Val: *Val->user_begin());
6604 bool IsSExt = isa<SExtInst>(Val: FirstUser);
6605 Type *ExtTy = FirstUser->getType();
6606 for (const User *U : Val->users()) {
6607 const Instruction *UI = cast<Instruction>(Val: U);
6608 if ((IsSExt && !isa<SExtInst>(Val: UI)) || (!IsSExt && !isa<ZExtInst>(Val: UI)))
6609 return false;
6610 Type *CurTy = UI->getType();
6611 // Same input and output types: Same instruction after CSE.
6612 if (CurTy == ExtTy)
6613 continue;
6614
6615 // If IsSExt is true, we are in this situation:
6616 // a = Val
6617 // b = sext ty1 a to ty2
6618 // c = sext ty1 a to ty3
6619 // Assuming ty2 is shorter than ty3, this could be turned into:
6620 // a = Val
6621 // b = sext ty1 a to ty2
6622 // c = sext ty2 b to ty3
6623 // However, the last sext is not free.
6624 if (IsSExt)
6625 return false;
6626
6627 // This is a ZExt, maybe this is free to extend from one type to another.
6628 // In that case, we would not account for a different use.
6629 Type *NarrowTy;
6630 Type *LargeTy;
6631 if (ExtTy->getScalarType()->getIntegerBitWidth() >
6632 CurTy->getScalarType()->getIntegerBitWidth()) {
6633 NarrowTy = CurTy;
6634 LargeTy = ExtTy;
6635 } else {
6636 NarrowTy = ExtTy;
6637 LargeTy = CurTy;
6638 }
6639
6640 if (!TLI.isZExtFree(FromTy: NarrowTy, ToTy: LargeTy))
6641 return false;
6642 }
6643 // All uses are the same or can be derived from one another for free.
6644 return true;
6645}
6646
6647/// Try to speculatively promote extensions in \p Exts and continue
6648/// promoting through newly promoted operands recursively as far as doing so is
6649/// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
6650/// When some promotion happened, \p TPT contains the proper state to revert
6651/// them.
6652///
6653/// \return true if some promotion happened, false otherwise.
6654bool CodeGenPrepare::tryToPromoteExts(
6655 TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
6656 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
6657 unsigned CreatedInstsCost) {
6658 bool Promoted = false;
6659
6660 // Iterate over all the extensions to try to promote them.
6661 for (auto *I : Exts) {
6662 // Early check if we directly have ext(load).
6663 if (isa<LoadInst>(Val: I->getOperand(i: 0))) {
6664 ProfitablyMovedExts.push_back(Elt: I);
6665 continue;
6666 }
6667
6668 // Check whether or not we want to do any promotion. The reason we have
6669 // this check inside the for loop is to catch the case where an extension
6670 // is directly fed by a load because in such case the extension can be moved
6671 // up without any promotion on its operands.
6672 if (!TLI->enableExtLdPromotion() || DisableExtLdPromotion)
6673 return false;
6674
6675 // Get the action to perform the promotion.
6676 TypePromotionHelper::Action TPH =
6677 TypePromotionHelper::getAction(Ext: I, InsertedInsts, TLI: *TLI, PromotedInsts);
6678 // Check if we can promote.
6679 if (!TPH) {
6680 // Save the current extension as we cannot move up through its operand.
6681 ProfitablyMovedExts.push_back(Elt: I);
6682 continue;
6683 }
6684
6685 // Save the current state.
6686 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
6687 TPT.getRestorationPoint();
6688 SmallVector<Instruction *, 4> NewExts;
6689 unsigned NewCreatedInstsCost = 0;
6690 unsigned ExtCost = !TLI->isExtFree(I);
6691 // Promote.
6692 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
6693 &NewExts, nullptr, *TLI);
6694 assert(PromotedVal &&
6695 "TypePromotionHelper should have filtered out those cases");
6696
6697 // We would be able to merge only one extension in a load.
6698 // Therefore, if we have more than 1 new extension we heuristically
6699 // cut this search path, because it means we degrade the code quality.
6700 // With exactly 2, the transformation is neutral, because we will merge
6701 // one extension but leave one. However, we optimistically keep going,
6702 // because the new extension may be removed too. Also avoid replacing a
6703 // single free extension with multiple extensions, as this increases the
6704 // number of IR instructions while not providing any savings.
6705 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
6706 // FIXME: It would be possible to propagate a negative value instead of
6707 // conservatively ceiling it to 0.
6708 TotalCreatedInstsCost =
6709 std::max(a: (long long)0, b: (TotalCreatedInstsCost - ExtCost));
6710 if (!StressExtLdPromotion &&
6711 (TotalCreatedInstsCost > 1 ||
6712 !isPromotedInstructionLegal(TLI: *TLI, DL: *DL, Val: PromotedVal) ||
6713 (ExtCost == 0 && NewExts.size() > 1))) {
6714 // This promotion is not profitable, rollback to the previous state, and
6715 // save the current extension in ProfitablyMovedExts as the latest
6716 // speculative promotion turned out to be unprofitable.
6717 TPT.rollback(Point: LastKnownGood);
6718 ProfitablyMovedExts.push_back(Elt: I);
6719 continue;
6720 }
6721 // Continue promoting NewExts as far as doing so is profitable.
6722 SmallVector<Instruction *, 2> NewlyMovedExts;
6723 (void)tryToPromoteExts(TPT, Exts: NewExts, ProfitablyMovedExts&: NewlyMovedExts, CreatedInstsCost: TotalCreatedInstsCost);
6724 bool NewPromoted = false;
6725 for (auto *ExtInst : NewlyMovedExts) {
6726 Instruction *MovedExt = cast<Instruction>(Val: ExtInst);
6727 Value *ExtOperand = MovedExt->getOperand(i: 0);
6728 // If we have reached to a load, we need this extra profitability check
6729 // as it could potentially be merged into an ext(load).
6730 if (isa<LoadInst>(Val: ExtOperand) &&
6731 !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
6732 (ExtOperand->hasOneUse() || hasSameExtUse(Val: ExtOperand, TLI: *TLI))))
6733 continue;
6734
6735 ProfitablyMovedExts.push_back(Elt: MovedExt);
6736 NewPromoted = true;
6737 }
6738
6739 // If none of speculative promotions for NewExts is profitable, rollback
6740 // and save the current extension (I) as the last profitable extension.
6741 if (!NewPromoted) {
6742 TPT.rollback(Point: LastKnownGood);
6743 ProfitablyMovedExts.push_back(Elt: I);
6744 continue;
6745 }
6746 // The promotion is profitable.
6747 Promoted = true;
6748 }
6749 return Promoted;
6750}
6751
6752/// Merging redundant sexts when one is dominating the other.
6753bool CodeGenPrepare::mergeSExts(Function &F) {
6754 bool Changed = false;
6755 for (auto &Entry : ValToSExtendedUses) {
6756 SExts &Insts = Entry.second;
6757 SExts CurPts;
6758 for (Instruction *Inst : Insts) {
6759 if (RemovedInsts.count(Ptr: Inst) || !isa<SExtInst>(Val: Inst) ||
6760 Inst->getOperand(i: 0) != Entry.first)
6761 continue;
6762 bool inserted = false;
6763 for (auto &Pt : CurPts) {
6764 if (getDT(F).dominates(Def: Inst, User: Pt)) {
6765 replaceAllUsesWith(Old: Pt, New: Inst, FreshBBs, IsHuge: IsHugeFunc);
6766 RemovedInsts.insert(Ptr: Pt);
6767 Pt->removeFromParent();
6768 Pt = Inst;
6769 inserted = true;
6770 Changed = true;
6771 break;
6772 }
6773 if (!getDT(F).dominates(Def: Pt, User: Inst))
6774 // Give up if we need to merge in a common dominator as the
6775 // experiments show it is not profitable.
6776 continue;
6777 replaceAllUsesWith(Old: Inst, New: Pt, FreshBBs, IsHuge: IsHugeFunc);
6778 RemovedInsts.insert(Ptr: Inst);
6779 Inst->removeFromParent();
6780 inserted = true;
6781 Changed = true;
6782 break;
6783 }
6784 if (!inserted)
6785 CurPts.push_back(Elt: Inst);
6786 }
6787 }
6788 return Changed;
6789}
6790
6791// Splitting large data structures so that the GEPs accessing them can have
6792// smaller offsets so that they can be sunk to the same blocks as their users.
6793// For example, a large struct starting from %base is split into two parts
6794// where the second part starts from %new_base.
6795//
6796// Before:
6797// BB0:
6798// %base =
6799//
6800// BB1:
6801// %gep0 = gep %base, off0
6802// %gep1 = gep %base, off1
6803// %gep2 = gep %base, off2
6804//
6805// BB2:
6806// %load1 = load %gep0
6807// %load2 = load %gep1
6808// %load3 = load %gep2
6809//
6810// After:
6811// BB0:
6812// %base =
6813// %new_base = gep %base, off0
6814//
6815// BB1:
6816// %new_gep0 = %new_base
6817// %new_gep1 = gep %new_base, off1 - off0
6818// %new_gep2 = gep %new_base, off2 - off0
6819//
6820// BB2:
6821// %load1 = load i32, i32* %new_gep0
6822// %load2 = load i32, i32* %new_gep1
6823// %load3 = load i32, i32* %new_gep2
6824//
6825// %new_gep1 and %new_gep2 can be sunk to BB2 now after the splitting because
6826// their offsets are smaller enough to fit into the addressing mode.
6827bool CodeGenPrepare::splitLargeGEPOffsets() {
6828 bool Changed = false;
6829 for (auto &Entry : LargeOffsetGEPMap) {
6830 Value *OldBase = Entry.first;
6831 SmallVectorImpl<std::pair<AssertingVH<GetElementPtrInst>, int64_t>>
6832 &LargeOffsetGEPs = Entry.second;
6833 auto compareGEPOffset =
6834 [&](const std::pair<GetElementPtrInst *, int64_t> &LHS,
6835 const std::pair<GetElementPtrInst *, int64_t> &RHS) {
6836 if (LHS.first == RHS.first)
6837 return false;
6838 if (LHS.second != RHS.second)
6839 return LHS.second < RHS.second;
6840 return LargeOffsetGEPID[LHS.first] < LargeOffsetGEPID[RHS.first];
6841 };
6842 // Sorting all the GEPs of the same data structures based on the offsets.
6843 llvm::sort(C&: LargeOffsetGEPs, Comp: compareGEPOffset);
6844 LargeOffsetGEPs.erase(CS: llvm::unique(R&: LargeOffsetGEPs), CE: LargeOffsetGEPs.end());
6845 // Skip if all the GEPs have the same offsets.
6846 if (LargeOffsetGEPs.front().second == LargeOffsetGEPs.back().second)
6847 continue;
6848 GetElementPtrInst *BaseGEP = LargeOffsetGEPs.begin()->first;
6849 int64_t BaseOffset = LargeOffsetGEPs.begin()->second;
6850 Value *NewBaseGEP = nullptr;
6851
6852 auto createNewBase = [&](int64_t BaseOffset, Value *OldBase,
6853 GetElementPtrInst *GEP) {
6854 LLVMContext &Ctx = GEP->getContext();
6855 Type *PtrIdxTy = DL->getIndexType(PtrTy: GEP->getType());
6856 Type *I8PtrTy =
6857 PointerType::get(C&: Ctx, AddressSpace: GEP->getType()->getPointerAddressSpace());
6858
6859 BasicBlock::iterator NewBaseInsertPt;
6860 BasicBlock *NewBaseInsertBB;
6861 if (auto *BaseI = dyn_cast<Instruction>(Val: OldBase)) {
6862 // If the base of the struct is an instruction, the new base will be
6863 // inserted close to it.
6864 NewBaseInsertBB = BaseI->getParent();
6865 if (isa<PHINode>(Val: BaseI))
6866 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
6867 else if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Val: BaseI)) {
6868 NewBaseInsertBB =
6869 SplitEdge(From: NewBaseInsertBB, To: Invoke->getNormalDest(),
6870 DT: &getDT(F&: *NewBaseInsertBB->getParent()), LI);
6871 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
6872 } else
6873 NewBaseInsertPt = std::next(x: BaseI->getIterator());
6874 } else {
6875 // If the current base is an argument or global value, the new base
6876 // will be inserted to the entry block.
6877 NewBaseInsertBB = &BaseGEP->getFunction()->getEntryBlock();
6878 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
6879 }
6880 IRBuilder<> NewBaseBuilder(NewBaseInsertBB, NewBaseInsertPt);
6881 // Create a new base.
6882 // TODO: Avoid implicit trunc?
6883 // See https://github.com/llvm/llvm-project/issues/112510.
6884 Value *BaseIndex =
6885 ConstantInt::getSigned(Ty: PtrIdxTy, V: BaseOffset, /*ImplicitTrunc=*/true);
6886 NewBaseGEP = OldBase;
6887 if (NewBaseGEP->getType() != I8PtrTy)
6888 NewBaseGEP = NewBaseBuilder.CreatePointerCast(V: NewBaseGEP, DestTy: I8PtrTy);
6889 NewBaseGEP =
6890 NewBaseBuilder.CreatePtrAdd(Ptr: NewBaseGEP, Offset: BaseIndex, Name: "splitgep");
6891 NewGEPBases.insert(V: NewBaseGEP);
6892 return;
6893 };
6894
6895 // Check whether all the offsets can be encoded with prefered common base.
6896 if (int64_t PreferBase = TLI->getPreferredLargeGEPBaseOffset(
6897 MinOffset: LargeOffsetGEPs.front().second, MaxOffset: LargeOffsetGEPs.back().second)) {
6898 BaseOffset = PreferBase;
6899 // Create a new base if the offset of the BaseGEP can be decoded with one
6900 // instruction.
6901 createNewBase(BaseOffset, OldBase, BaseGEP);
6902 }
6903
6904 auto *LargeOffsetGEP = LargeOffsetGEPs.begin();
6905 while (LargeOffsetGEP != LargeOffsetGEPs.end()) {
6906 GetElementPtrInst *GEP = LargeOffsetGEP->first;
6907 int64_t Offset = LargeOffsetGEP->second;
6908 if (Offset != BaseOffset) {
6909 TargetLowering::AddrMode AddrMode;
6910 AddrMode.HasBaseReg = true;
6911 AddrMode.BaseOffs = Offset - BaseOffset;
6912 // The result type of the GEP might not be the type of the memory
6913 // access.
6914 if (!TLI->isLegalAddressingMode(DL: *DL, AM: AddrMode,
6915 Ty: GEP->getResultElementType(),
6916 AddrSpace: GEP->getAddressSpace())) {
6917 // We need to create a new base if the offset to the current base is
6918 // too large to fit into the addressing mode. So, a very large struct
6919 // may be split into several parts.
6920 BaseGEP = GEP;
6921 BaseOffset = Offset;
6922 NewBaseGEP = nullptr;
6923 }
6924 }
6925
6926 // Generate a new GEP to replace the current one.
6927 Type *PtrIdxTy = DL->getIndexType(PtrTy: GEP->getType());
6928
6929 if (!NewBaseGEP) {
6930 // Create a new base if we don't have one yet. Find the insertion
6931 // pointer for the new base first.
6932 createNewBase(BaseOffset, OldBase, GEP);
6933 }
6934
6935 IRBuilder<> Builder(GEP);
6936 Value *NewGEP = NewBaseGEP;
6937 if (Offset != BaseOffset) {
6938 // Calculate the new offset for the new GEP.
6939 Value *Index = ConstantInt::get(Ty: PtrIdxTy, V: Offset - BaseOffset);
6940 NewGEP = Builder.CreatePtrAdd(Ptr: NewBaseGEP, Offset: Index);
6941 }
6942 replaceAllUsesWith(Old: GEP, New: NewGEP, FreshBBs, IsHuge: IsHugeFunc);
6943 LargeOffsetGEPID.erase(Val: GEP);
6944 LargeOffsetGEP = LargeOffsetGEPs.erase(CI: LargeOffsetGEP);
6945 GEP->eraseFromParent();
6946 Changed = true;
6947 }
6948 }
6949 return Changed;
6950}
6951
6952bool CodeGenPrepare::optimizePhiType(
6953 PHINode *I, SmallPtrSetImpl<PHINode *> &Visited,
6954 SmallPtrSetImpl<Instruction *> &DeletedInstrs) {
6955 // We are looking for a collection on interconnected phi nodes that together
6956 // only use loads/bitcasts and are used by stores/bitcasts, and the bitcasts
6957 // are of the same type. Convert the whole set of nodes to the type of the
6958 // bitcast.
6959 Type *PhiTy = I->getType();
6960 Type *ConvertTy = nullptr;
6961 if (Visited.count(Ptr: I) ||
6962 (!I->getType()->isIntegerTy() && !I->getType()->isFloatingPointTy()))
6963 return false;
6964
6965 SmallVector<Instruction *, 4> Worklist;
6966 Worklist.push_back(Elt: cast<Instruction>(Val: I));
6967 SmallPtrSet<PHINode *, 4> PhiNodes;
6968 SmallPtrSet<ConstantData *, 4> Constants;
6969 PhiNodes.insert(Ptr: I);
6970 Visited.insert(Ptr: I);
6971 SmallPtrSet<Instruction *, 4> Defs;
6972 SmallPtrSet<Instruction *, 4> Uses;
6973 // This works by adding extra bitcasts between load/stores and removing
6974 // existing bitcasts. If we have a phi(bitcast(load)) or a store(bitcast(phi))
6975 // we can get in the situation where we remove a bitcast in one iteration
6976 // just to add it again in the next. We need to ensure that at least one
6977 // bitcast we remove are anchored to something that will not change back.
6978 bool AnyAnchored = false;
6979
6980 while (!Worklist.empty()) {
6981 Instruction *II = Worklist.pop_back_val();
6982
6983 if (auto *Phi = dyn_cast<PHINode>(Val: II)) {
6984 // Handle Defs, which might also be PHI's
6985 for (Value *V : Phi->incoming_values()) {
6986 if (auto *OpPhi = dyn_cast<PHINode>(Val: V)) {
6987 if (!PhiNodes.count(Ptr: OpPhi)) {
6988 if (!Visited.insert(Ptr: OpPhi).second)
6989 return false;
6990 PhiNodes.insert(Ptr: OpPhi);
6991 Worklist.push_back(Elt: OpPhi);
6992 }
6993 } else if (auto *OpLoad = dyn_cast<LoadInst>(Val: V)) {
6994 if (!OpLoad->isSimple())
6995 return false;
6996 if (Defs.insert(Ptr: OpLoad).second)
6997 Worklist.push_back(Elt: OpLoad);
6998 } else if (auto *OpEx = dyn_cast<ExtractElementInst>(Val: V)) {
6999 if (Defs.insert(Ptr: OpEx).second)
7000 Worklist.push_back(Elt: OpEx);
7001 } else if (auto *OpBC = dyn_cast<BitCastInst>(Val: V)) {
7002 if (!ConvertTy)
7003 ConvertTy = OpBC->getOperand(i_nocapture: 0)->getType();
7004 if (OpBC->getOperand(i_nocapture: 0)->getType() != ConvertTy)
7005 return false;
7006 if (Defs.insert(Ptr: OpBC).second) {
7007 Worklist.push_back(Elt: OpBC);
7008 AnyAnchored |= !isa<LoadInst>(Val: OpBC->getOperand(i_nocapture: 0)) &&
7009 !isa<ExtractElementInst>(Val: OpBC->getOperand(i_nocapture: 0));
7010 }
7011 } else if (auto *OpC = dyn_cast<ConstantData>(Val: V))
7012 Constants.insert(Ptr: OpC);
7013 else
7014 return false;
7015 }
7016 }
7017
7018 // Handle uses which might also be phi's
7019 for (User *V : II->users()) {
7020 if (auto *OpPhi = dyn_cast<PHINode>(Val: V)) {
7021 if (!PhiNodes.count(Ptr: OpPhi)) {
7022 if (Visited.count(Ptr: OpPhi))
7023 return false;
7024 PhiNodes.insert(Ptr: OpPhi);
7025 Visited.insert(Ptr: OpPhi);
7026 Worklist.push_back(Elt: OpPhi);
7027 }
7028 } else if (auto *OpStore = dyn_cast<StoreInst>(Val: V)) {
7029 if (!OpStore->isSimple() || OpStore->getOperand(i_nocapture: 0) != II)
7030 return false;
7031 Uses.insert(Ptr: OpStore);
7032 } else if (auto *OpBC = dyn_cast<BitCastInst>(Val: V)) {
7033 if (!ConvertTy)
7034 ConvertTy = OpBC->getType();
7035 if (OpBC->getType() != ConvertTy)
7036 return false;
7037 Uses.insert(Ptr: OpBC);
7038 AnyAnchored |=
7039 any_of(Range: OpBC->users(), P: [](User *U) { return !isa<StoreInst>(Val: U); });
7040 } else {
7041 return false;
7042 }
7043 }
7044 }
7045
7046 if (!ConvertTy || !AnyAnchored || PhiTy == ConvertTy ||
7047 !TLI->shouldConvertPhiType(From: PhiTy, To: ConvertTy))
7048 return false;
7049
7050 LLVM_DEBUG(dbgs() << "Converting " << *I << "\n and connected nodes to "
7051 << *ConvertTy << "\n");
7052
7053 // Create all the new phi nodes of the new type, and bitcast any loads to the
7054 // correct type.
7055 ValueToValueMap ValMap;
7056 for (ConstantData *C : Constants)
7057 ValMap[C] = ConstantExpr::getBitCast(C, Ty: ConvertTy);
7058 for (Instruction *D : Defs) {
7059 if (isa<BitCastInst>(Val: D)) {
7060 ValMap[D] = D->getOperand(i: 0);
7061 DeletedInstrs.insert(Ptr: D);
7062 } else {
7063 BasicBlock::iterator insertPt = std::next(x: D->getIterator());
7064 ValMap[D] = new BitCastInst(D, ConvertTy, D->getName() + ".bc", insertPt);
7065 }
7066 }
7067 for (PHINode *Phi : PhiNodes)
7068 ValMap[Phi] = PHINode::Create(Ty: ConvertTy, NumReservedValues: Phi->getNumIncomingValues(),
7069 NameStr: Phi->getName() + ".tc", InsertBefore: Phi->getIterator());
7070 // Pipe together all the PhiNodes.
7071 for (PHINode *Phi : PhiNodes) {
7072 PHINode *NewPhi = cast<PHINode>(Val: ValMap[Phi]);
7073 for (int i = 0, e = Phi->getNumIncomingValues(); i < e; i++)
7074 NewPhi->addIncoming(V: ValMap[Phi->getIncomingValue(i)],
7075 BB: Phi->getIncomingBlock(i));
7076 Visited.insert(Ptr: NewPhi);
7077 }
7078 // And finally pipe up the stores and bitcasts
7079 for (Instruction *U : Uses) {
7080 if (isa<BitCastInst>(Val: U)) {
7081 DeletedInstrs.insert(Ptr: U);
7082 replaceAllUsesWith(Old: U, New: ValMap[U->getOperand(i: 0)], FreshBBs, IsHuge: IsHugeFunc);
7083 } else {
7084 U->setOperand(i: 0, Val: new BitCastInst(ValMap[U->getOperand(i: 0)], PhiTy, "bc",
7085 U->getIterator()));
7086 }
7087 }
7088
7089 // Save the removed phis to be deleted later.
7090 DeletedInstrs.insert_range(R&: PhiNodes);
7091 return true;
7092}
7093
7094bool CodeGenPrepare::optimizePhiTypes(Function &F) {
7095 if (!OptimizePhiTypes)
7096 return false;
7097
7098 bool Changed = false;
7099 SmallPtrSet<PHINode *, 4> Visited;
7100 SmallPtrSet<Instruction *, 4> DeletedInstrs;
7101
7102 // Attempt to optimize all the phis in the functions to the correct type.
7103 for (auto &BB : F)
7104 for (auto &Phi : BB.phis())
7105 Changed |= optimizePhiType(I: &Phi, Visited, DeletedInstrs);
7106
7107 // Remove any old phi's that have been converted.
7108 for (auto *I : DeletedInstrs) {
7109 replaceAllUsesWith(Old: I, New: PoisonValue::get(T: I->getType()), FreshBBs, IsHuge: IsHugeFunc);
7110 I->eraseFromParent();
7111 }
7112
7113 return Changed;
7114}
7115
7116/// Return true, if an ext(load) can be formed from an extension in
7117/// \p MovedExts.
7118bool CodeGenPrepare::canFormExtLd(
7119 const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
7120 Instruction *&Inst, bool HasPromoted) {
7121 for (auto *MovedExtInst : MovedExts) {
7122 if (isa<LoadInst>(Val: MovedExtInst->getOperand(i: 0))) {
7123 LI = cast<LoadInst>(Val: MovedExtInst->getOperand(i: 0));
7124 Inst = MovedExtInst;
7125 break;
7126 }
7127 }
7128 if (!LI)
7129 return false;
7130
7131 // If they're already in the same block, there's nothing to do.
7132 // Make the cheap checks first if we did not promote.
7133 // If we promoted, we need to check if it is indeed profitable.
7134 if (!HasPromoted && LI->getParent() == Inst->getParent())
7135 return false;
7136
7137 return TLI->isExtLoad(Load: LI, Ext: Inst, DL: *DL);
7138}
7139
7140/// Move a zext or sext fed by a load into the same basic block as the load,
7141/// unless conditions are unfavorable. This allows SelectionDAG to fold the
7142/// extend into the load.
7143///
7144/// E.g.,
7145/// \code
7146/// %ld = load i32* %addr
7147/// %add = add nuw i32 %ld, 4
7148/// %zext = zext i32 %add to i64
7149// \endcode
7150/// =>
7151/// \code
7152/// %ld = load i32* %addr
7153/// %zext = zext i32 %ld to i64
7154/// %add = add nuw i64 %zext, 4
7155/// \encode
7156/// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
7157/// allow us to match zext(load i32*) to i64.
7158///
7159/// Also, try to promote the computations used to obtain a sign extended
7160/// value used into memory accesses.
7161/// E.g.,
7162/// \code
7163/// a = add nsw i32 b, 3
7164/// d = sext i32 a to i64
7165/// e = getelementptr ..., i64 d
7166/// \endcode
7167/// =>
7168/// \code
7169/// f = sext i32 b to i64
7170/// a = add nsw i64 f, 3
7171/// e = getelementptr ..., i64 a
7172/// \endcode
7173///
7174/// \p Inst[in/out] the extension may be modified during the process if some
7175/// promotions apply.
7176bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
7177 bool AllowPromotionWithoutCommonHeader = false;
7178 /// See if it is an interesting sext operations for the address type
7179 /// promotion before trying to promote it, e.g., the ones with the right
7180 /// type and used in memory accesses.
7181 bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
7182 I: *Inst, AllowPromotionWithoutCommonHeader);
7183 TypePromotionTransaction TPT(RemovedInsts);
7184 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
7185 TPT.getRestorationPoint();
7186 SmallVector<Instruction *, 1> Exts;
7187 SmallVector<Instruction *, 2> SpeculativelyMovedExts;
7188 Exts.push_back(Elt: Inst);
7189
7190 bool HasPromoted = tryToPromoteExts(TPT, Exts, ProfitablyMovedExts&: SpeculativelyMovedExts);
7191
7192 // Look for a load being extended.
7193 LoadInst *LI = nullptr;
7194 Instruction *ExtFedByLoad;
7195
7196 // Try to promote a chain of computation if it allows to form an extended
7197 // load.
7198 if (canFormExtLd(MovedExts: SpeculativelyMovedExts, LI, Inst&: ExtFedByLoad, HasPromoted)) {
7199 assert(LI && ExtFedByLoad && "Expect a valid load and extension");
7200 TPT.commit();
7201 // Move the extend into the same block as the load.
7202 ExtFedByLoad->moveAfter(MovePos: LI);
7203 ++NumExtsMoved;
7204 Inst = ExtFedByLoad;
7205 return true;
7206 }
7207
7208 // Continue promoting SExts if known as considerable depending on targets.
7209 if (ATPConsiderable &&
7210 performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
7211 HasPromoted, TPT, SpeculativelyMovedExts))
7212 return true;
7213
7214 TPT.rollback(Point: LastKnownGood);
7215 return false;
7216}
7217
7218// Perform address type promotion if doing so is profitable.
7219// If AllowPromotionWithoutCommonHeader == false, we should find other sext
7220// instructions that sign extended the same initial value. However, if
7221// AllowPromotionWithoutCommonHeader == true, we expect promoting the
7222// extension is just profitable.
7223bool CodeGenPrepare::performAddressTypePromotion(
7224 Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
7225 bool HasPromoted, TypePromotionTransaction &TPT,
7226 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
7227 bool Promoted = false;
7228 SmallPtrSet<Instruction *, 1> UnhandledExts;
7229 bool AllSeenFirst = true;
7230 for (auto *I : SpeculativelyMovedExts) {
7231 Value *HeadOfChain = I->getOperand(i: 0);
7232 DenseMap<Value *, Instruction *>::iterator AlreadySeen =
7233 SeenChainsForSExt.find(Val: HeadOfChain);
7234 // If there is an unhandled SExt which has the same header, try to promote
7235 // it as well.
7236 if (AlreadySeen != SeenChainsForSExt.end()) {
7237 if (AlreadySeen->second != nullptr)
7238 UnhandledExts.insert(Ptr: AlreadySeen->second);
7239 AllSeenFirst = false;
7240 }
7241 }
7242
7243 if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
7244 SpeculativelyMovedExts.size() == 1)) {
7245 TPT.commit();
7246 if (HasPromoted)
7247 Promoted = true;
7248 for (auto *I : SpeculativelyMovedExts) {
7249 Value *HeadOfChain = I->getOperand(i: 0);
7250 SeenChainsForSExt[HeadOfChain] = nullptr;
7251 ValToSExtendedUses[HeadOfChain].push_back(Elt: I);
7252 }
7253 // Update Inst as promotion happen.
7254 Inst = SpeculativelyMovedExts.pop_back_val();
7255 } else {
7256 // This is the first chain visited from the header, keep the current chain
7257 // as unhandled. Defer to promote this until we encounter another SExt
7258 // chain derived from the same header.
7259 for (auto *I : SpeculativelyMovedExts) {
7260 Value *HeadOfChain = I->getOperand(i: 0);
7261 SeenChainsForSExt[HeadOfChain] = Inst;
7262 }
7263 return false;
7264 }
7265
7266 if (!AllSeenFirst && !UnhandledExts.empty())
7267 for (auto *VisitedSExt : UnhandledExts) {
7268 if (RemovedInsts.count(Ptr: VisitedSExt))
7269 continue;
7270 TypePromotionTransaction TPT(RemovedInsts);
7271 SmallVector<Instruction *, 1> Exts;
7272 SmallVector<Instruction *, 2> Chains;
7273 Exts.push_back(Elt: VisitedSExt);
7274 bool HasPromoted = tryToPromoteExts(TPT, Exts, ProfitablyMovedExts&: Chains);
7275 TPT.commit();
7276 if (HasPromoted)
7277 Promoted = true;
7278 for (auto *I : Chains) {
7279 Value *HeadOfChain = I->getOperand(i: 0);
7280 // Mark this as handled.
7281 SeenChainsForSExt[HeadOfChain] = nullptr;
7282 ValToSExtendedUses[HeadOfChain].push_back(Elt: I);
7283 }
7284 }
7285 return Promoted;
7286}
7287
7288bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
7289 BasicBlock *DefBB = I->getParent();
7290
7291 // If the result of a {s|z}ext and its source are both live out, rewrite all
7292 // other uses of the source with result of extension.
7293 Value *Src = I->getOperand(i: 0);
7294 if (Src->hasOneUse())
7295 return false;
7296
7297 // Only do this xform if truncating is free.
7298 if (!TLI->isTruncateFree(FromTy: I->getType(), ToTy: Src->getType()))
7299 return false;
7300
7301 // Only safe to perform the optimization if the source is also defined in
7302 // this block.
7303 if (!isa<Instruction>(Val: Src) || DefBB != cast<Instruction>(Val: Src)->getParent())
7304 return false;
7305
7306 bool DefIsLiveOut = false;
7307 for (User *U : I->users()) {
7308 Instruction *UI = cast<Instruction>(Val: U);
7309
7310 // Figure out which BB this ext is used in.
7311 BasicBlock *UserBB = UI->getParent();
7312 if (UserBB == DefBB)
7313 continue;
7314 DefIsLiveOut = true;
7315 break;
7316 }
7317 if (!DefIsLiveOut)
7318 return false;
7319
7320 // Make sure none of the uses are PHI nodes.
7321 for (User *U : Src->users()) {
7322 Instruction *UI = cast<Instruction>(Val: U);
7323 BasicBlock *UserBB = UI->getParent();
7324 if (UserBB == DefBB)
7325 continue;
7326 // Be conservative. We don't want this xform to end up introducing
7327 // reloads just before load / store instructions.
7328 if (isa<PHINode>(Val: UI) || isa<LoadInst>(Val: UI) || isa<StoreInst>(Val: UI))
7329 return false;
7330 }
7331
7332 // InsertedTruncs - Only insert one trunc in each block once.
7333 DenseMap<BasicBlock *, Instruction *> InsertedTruncs;
7334
7335 bool MadeChange = false;
7336 for (Use &U : Src->uses()) {
7337 Instruction *User = cast<Instruction>(Val: U.getUser());
7338
7339 // Figure out which BB this ext is used in.
7340 BasicBlock *UserBB = User->getParent();
7341 if (UserBB == DefBB)
7342 continue;
7343
7344 // Both src and def are live in this block. Rewrite the use.
7345 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
7346
7347 if (!InsertedTrunc) {
7348 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
7349 assert(InsertPt != UserBB->end());
7350 InsertedTrunc = new TruncInst(I, Src->getType(), "");
7351 InsertedTrunc->insertBefore(BB&: *UserBB, InsertPos: InsertPt);
7352 InsertedInsts.insert(Ptr: InsertedTrunc);
7353 }
7354
7355 // Replace a use of the {s|z}ext source with a use of the result.
7356 U = InsertedTrunc;
7357 ++NumExtUses;
7358 MadeChange = true;
7359 }
7360
7361 return MadeChange;
7362}
7363
7364// Find loads whose uses only use some of the loaded value's bits. Add an "and"
7365// just after the load if the target can fold this into one extload instruction,
7366// with the hope of eliminating some of the other later "and" instructions using
7367// the loaded value. "and"s that are made trivially redundant by the insertion
7368// of the new "and" are removed by this function, while others (e.g. those whose
7369// path from the load goes through a phi) are left for isel to potentially
7370// remove.
7371//
7372// For example:
7373//
7374// b0:
7375// x = load i32
7376// ...
7377// b1:
7378// y = and x, 0xff
7379// z = use y
7380//
7381// becomes:
7382//
7383// b0:
7384// x = load i32
7385// x' = and x, 0xff
7386// ...
7387// b1:
7388// z = use x'
7389//
7390// whereas:
7391//
7392// b0:
7393// x1 = load i32
7394// ...
7395// b1:
7396// x2 = load i32
7397// ...
7398// b2:
7399// x = phi x1, x2
7400// y = and x, 0xff
7401//
7402// becomes (after a call to optimizeLoadExt for each load):
7403//
7404// b0:
7405// x1 = load i32
7406// x1' = and x1, 0xff
7407// ...
7408// b1:
7409// x2 = load i32
7410// x2' = and x2, 0xff
7411// ...
7412// b2:
7413// x = phi x1', x2'
7414// y = and x, 0xff
7415bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
7416 if (!Load->isSimple() || !Load->getType()->isIntOrPtrTy())
7417 return false;
7418
7419 // Skip loads we've already transformed.
7420 if (Load->hasOneUse() &&
7421 InsertedInsts.count(Ptr: cast<Instruction>(Val: *Load->user_begin())))
7422 return false;
7423
7424 // Look at all uses of Load, looking through phis, to determine how many bits
7425 // of the loaded value are needed.
7426 SmallVector<Instruction *, 8> WorkList;
7427 SmallPtrSet<Instruction *, 16> Visited;
7428 SmallVector<Instruction *, 8> AndsToMaybeRemove;
7429 SmallVector<Instruction *, 8> DropFlags;
7430 for (auto *U : Load->users())
7431 WorkList.push_back(Elt: cast<Instruction>(Val: U));
7432
7433 EVT LoadResultVT = TLI->getValueType(DL: *DL, Ty: Load->getType());
7434 unsigned BitWidth = LoadResultVT.getSizeInBits();
7435 // If the BitWidth is 0, do not try to optimize the type
7436 if (BitWidth == 0)
7437 return false;
7438
7439 APInt DemandBits(BitWidth, 0);
7440 APInt WidestAndBits(BitWidth, 0);
7441
7442 while (!WorkList.empty()) {
7443 Instruction *I = WorkList.pop_back_val();
7444
7445 // Break use-def graph loops.
7446 if (!Visited.insert(Ptr: I).second)
7447 continue;
7448
7449 // For a PHI node, push all of its users.
7450 if (auto *Phi = dyn_cast<PHINode>(Val: I)) {
7451 for (auto *U : Phi->users())
7452 WorkList.push_back(Elt: cast<Instruction>(Val: U));
7453 continue;
7454 }
7455
7456 switch (I->getOpcode()) {
7457 case Instruction::And: {
7458 auto *AndC = dyn_cast<ConstantInt>(Val: I->getOperand(i: 1));
7459 if (!AndC)
7460 return false;
7461 APInt AndBits = AndC->getValue();
7462 DemandBits |= AndBits;
7463 // Keep track of the widest and mask we see.
7464 if (AndBits.ugt(RHS: WidestAndBits))
7465 WidestAndBits = AndBits;
7466 if (AndBits == WidestAndBits && I->getOperand(i: 0) == Load)
7467 AndsToMaybeRemove.push_back(Elt: I);
7468 break;
7469 }
7470
7471 case Instruction::Shl: {
7472 auto *ShlC = dyn_cast<ConstantInt>(Val: I->getOperand(i: 1));
7473 if (!ShlC)
7474 return false;
7475 uint64_t ShiftAmt = ShlC->getLimitedValue(Limit: BitWidth - 1);
7476 DemandBits.setLowBits(BitWidth - ShiftAmt);
7477 DropFlags.push_back(Elt: I);
7478 break;
7479 }
7480
7481 case Instruction::Trunc: {
7482 EVT TruncVT = TLI->getValueType(DL: *DL, Ty: I->getType());
7483 unsigned TruncBitWidth = TruncVT.getSizeInBits();
7484 DemandBits.setLowBits(TruncBitWidth);
7485 DropFlags.push_back(Elt: I);
7486 break;
7487 }
7488
7489 default:
7490 return false;
7491 }
7492 }
7493
7494 uint32_t ActiveBits = DemandBits.getActiveBits();
7495 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
7496 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
7497 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
7498 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
7499 // followed by an AND.
7500 // TODO: Look into removing this restriction by fixing backends to either
7501 // return false for isLoadExtLegal for i1 or have them select this pattern to
7502 // a single instruction.
7503 //
7504 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
7505 // mask, since these are the only ands that will be removed by isel.
7506 if (ActiveBits <= 1 || !DemandBits.isMask(numBits: ActiveBits) ||
7507 WidestAndBits != DemandBits)
7508 return false;
7509
7510 LLVMContext &Ctx = Load->getType()->getContext();
7511 Type *TruncTy = Type::getIntNTy(C&: Ctx, N: ActiveBits);
7512 EVT TruncVT = TLI->getValueType(DL: *DL, Ty: TruncTy);
7513
7514 // Reject cases that won't be matched as extloads.
7515 if (!LoadResultVT.bitsGT(VT: TruncVT) || !TruncVT.isRound() ||
7516 !TLI->isLoadExtLegal(ExtType: ISD::ZEXTLOAD, ValVT: LoadResultVT, MemVT: TruncVT))
7517 return false;
7518
7519 IRBuilder<> Builder(Load->getNextNode());
7520 auto *NewAnd = cast<Instruction>(
7521 Val: Builder.CreateAnd(LHS: Load, RHS: ConstantInt::get(Context&: Ctx, V: DemandBits)));
7522 // Mark this instruction as "inserted by CGP", so that other
7523 // optimizations don't touch it.
7524 InsertedInsts.insert(Ptr: NewAnd);
7525
7526 // Replace all uses of load with new and (except for the use of load in the
7527 // new and itself).
7528 replaceAllUsesWith(Old: Load, New: NewAnd, FreshBBs, IsHuge: IsHugeFunc);
7529 NewAnd->setOperand(i: 0, Val: Load);
7530
7531 // Remove any and instructions that are now redundant.
7532 for (auto *And : AndsToMaybeRemove)
7533 // Check that the and mask is the same as the one we decided to put on the
7534 // new and.
7535 if (cast<ConstantInt>(Val: And->getOperand(i: 1))->getValue() == DemandBits) {
7536 replaceAllUsesWith(Old: And, New: NewAnd, FreshBBs, IsHuge: IsHugeFunc);
7537 if (&*CurInstIterator == And)
7538 CurInstIterator = std::next(x: And->getIterator());
7539 And->eraseFromParent();
7540 ++NumAndUses;
7541 }
7542
7543 // NSW flags may not longer hold.
7544 for (auto *Inst : DropFlags)
7545 Inst->setHasNoSignedWrap(false);
7546
7547 ++NumAndsAdded;
7548 return true;
7549}
7550
7551/// Check if V (an operand of a select instruction) is an expensive instruction
7552/// that is only used once.
7553static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
7554 auto *I = dyn_cast<Instruction>(Val: V);
7555 // If it's safe to speculatively execute, then it should not have side
7556 // effects; therefore, it's safe to sink and possibly *not* execute.
7557 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
7558 TTI->isExpensiveToSpeculativelyExecute(I);
7559}
7560
7561/// Returns true if a SelectInst should be turned into an explicit branch.
7562static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
7563 const TargetLowering *TLI,
7564 SelectInst *SI) {
7565 // If even a predictable select is cheap, then a branch can't be cheaper.
7566 if (!TLI->isPredictableSelectExpensive())
7567 return false;
7568
7569 // FIXME: This should use the same heuristics as IfConversion to determine
7570 // whether a select is better represented as a branch.
7571
7572 // If metadata tells us that the select condition is obviously predictable,
7573 // then we want to replace the select with a branch.
7574 uint64_t TrueWeight, FalseWeight;
7575 if (extractBranchWeights(I: *SI, TrueVal&: TrueWeight, FalseVal&: FalseWeight)) {
7576 uint64_t Max = std::max(a: TrueWeight, b: FalseWeight);
7577 uint64_t Sum = TrueWeight + FalseWeight;
7578 if (Sum != 0) {
7579 auto Probability = BranchProbability::getBranchProbability(Numerator: Max, Denominator: Sum);
7580 if (Probability > TTI->getPredictableBranchThreshold())
7581 return true;
7582 }
7583 }
7584
7585 CmpInst *Cmp = dyn_cast<CmpInst>(Val: SI->getCondition());
7586
7587 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
7588 // comparison condition. If the compare has more than one use, there's
7589 // probably another cmov or setcc around, so it's not worth emitting a branch.
7590 if (!Cmp || !Cmp->hasOneUse())
7591 return false;
7592
7593 // If either operand of the select is expensive and only needed on one side
7594 // of the select, we should form a branch.
7595 if (sinkSelectOperand(TTI, V: SI->getTrueValue()) ||
7596 sinkSelectOperand(TTI, V: SI->getFalseValue()))
7597 return true;
7598
7599 return false;
7600}
7601
7602/// If \p isTrue is true, return the true value of \p SI, otherwise return
7603/// false value of \p SI. If the true/false value of \p SI is defined by any
7604/// select instructions in \p Selects, look through the defining select
7605/// instruction until the true/false value is not defined in \p Selects.
7606static Value *
7607getTrueOrFalseValue(SelectInst *SI, bool isTrue,
7608 const SmallPtrSet<const Instruction *, 2> &Selects) {
7609 Value *V = nullptr;
7610
7611 for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(Ptr: DefSI);
7612 DefSI = dyn_cast<SelectInst>(Val: V)) {
7613 assert(DefSI->getCondition() == SI->getCondition() &&
7614 "The condition of DefSI does not match with SI");
7615 V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
7616 }
7617
7618 assert(V && "Failed to get select true/false value");
7619 return V;
7620}
7621
7622bool CodeGenPrepare::optimizeShiftInst(BinaryOperator *Shift) {
7623 assert(Shift->isShift() && "Expected a shift");
7624
7625 // If this is (1) a vector shift, (2) shifts by scalars are cheaper than
7626 // general vector shifts, and (3) the shift amount is a select-of-splatted
7627 // values, hoist the shifts before the select:
7628 // shift Op0, (select Cond, TVal, FVal) -->
7629 // select Cond, (shift Op0, TVal), (shift Op0, FVal)
7630 //
7631 // This is inverting a generic IR transform when we know that the cost of a
7632 // general vector shift is more than the cost of 2 shift-by-scalars.
7633 // We can't do this effectively in SDAG because we may not be able to
7634 // determine if the select operands are splats from within a basic block.
7635 Type *Ty = Shift->getType();
7636 if (!Ty->isVectorTy() || !TTI->isVectorShiftByScalarCheap(Ty))
7637 return false;
7638 Value *Cond, *TVal, *FVal;
7639 if (!match(V: Shift->getOperand(i_nocapture: 1),
7640 P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: TVal), R: m_Value(V&: FVal)))))
7641 return false;
7642 if (!isSplatValue(V: TVal) || !isSplatValue(V: FVal))
7643 return false;
7644
7645 IRBuilder<> Builder(Shift);
7646 BinaryOperator::BinaryOps Opcode = Shift->getOpcode();
7647 Value *NewTVal = Builder.CreateBinOp(Opc: Opcode, LHS: Shift->getOperand(i_nocapture: 0), RHS: TVal);
7648 Value *NewFVal = Builder.CreateBinOp(Opc: Opcode, LHS: Shift->getOperand(i_nocapture: 0), RHS: FVal);
7649 Value *NewSel = Builder.CreateSelect(C: Cond, True: NewTVal, False: NewFVal);
7650 replaceAllUsesWith(Old: Shift, New: NewSel, FreshBBs, IsHuge: IsHugeFunc);
7651 Shift->eraseFromParent();
7652 return true;
7653}
7654
7655bool CodeGenPrepare::optimizeFunnelShift(IntrinsicInst *Fsh) {
7656 Intrinsic::ID Opcode = Fsh->getIntrinsicID();
7657 assert((Opcode == Intrinsic::fshl || Opcode == Intrinsic::fshr) &&
7658 "Expected a funnel shift");
7659
7660 // If this is (1) a vector funnel shift, (2) shifts by scalars are cheaper
7661 // than general vector shifts, and (3) the shift amount is select-of-splatted
7662 // values, hoist the funnel shifts before the select:
7663 // fsh Op0, Op1, (select Cond, TVal, FVal) -->
7664 // select Cond, (fsh Op0, Op1, TVal), (fsh Op0, Op1, FVal)
7665 //
7666 // This is inverting a generic IR transform when we know that the cost of a
7667 // general vector shift is more than the cost of 2 shift-by-scalars.
7668 // We can't do this effectively in SDAG because we may not be able to
7669 // determine if the select operands are splats from within a basic block.
7670 Type *Ty = Fsh->getType();
7671 if (!Ty->isVectorTy() || !TTI->isVectorShiftByScalarCheap(Ty))
7672 return false;
7673 Value *Cond, *TVal, *FVal;
7674 if (!match(V: Fsh->getOperand(i_nocapture: 2),
7675 P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: TVal), R: m_Value(V&: FVal)))))
7676 return false;
7677 if (!isSplatValue(V: TVal) || !isSplatValue(V: FVal))
7678 return false;
7679
7680 IRBuilder<> Builder(Fsh);
7681 Value *X = Fsh->getOperand(i_nocapture: 0), *Y = Fsh->getOperand(i_nocapture: 1);
7682 Value *NewTVal = Builder.CreateIntrinsic(ID: Opcode, Types: Ty, Args: {X, Y, TVal});
7683 Value *NewFVal = Builder.CreateIntrinsic(ID: Opcode, Types: Ty, Args: {X, Y, FVal});
7684 Value *NewSel = Builder.CreateSelect(C: Cond, True: NewTVal, False: NewFVal);
7685 replaceAllUsesWith(Old: Fsh, New: NewSel, FreshBBs, IsHuge: IsHugeFunc);
7686 Fsh->eraseFromParent();
7687 return true;
7688}
7689
7690/// If we have a SelectInst that will likely profit from branch prediction,
7691/// turn it into a branch.
7692bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
7693 if (DisableSelectToBranch)
7694 return false;
7695
7696 // If the SelectOptimize pass is enabled, selects have already been optimized.
7697 if (!getCGPassBuilderOption().DisableSelectOptimize)
7698 return false;
7699
7700 // Find all consecutive select instructions that share the same condition.
7701 SmallVector<SelectInst *, 2> ASI;
7702 ASI.push_back(Elt: SI);
7703 for (BasicBlock::iterator It = ++BasicBlock::iterator(SI);
7704 It != SI->getParent()->end(); ++It) {
7705 SelectInst *I = dyn_cast<SelectInst>(Val: &*It);
7706 if (I && SI->getCondition() == I->getCondition()) {
7707 ASI.push_back(Elt: I);
7708 } else {
7709 break;
7710 }
7711 }
7712
7713 SelectInst *LastSI = ASI.back();
7714 // Increment the current iterator to skip all the rest of select instructions
7715 // because they will be either "not lowered" or "all lowered" to branch.
7716 CurInstIterator = std::next(x: LastSI->getIterator());
7717 // Examine debug-info attached to the consecutive select instructions. They
7718 // won't be individually optimised by optimizeInst, so we need to perform
7719 // DbgVariableRecord maintenence here instead.
7720 for (SelectInst *SI : ArrayRef(ASI).drop_front())
7721 fixupDbgVariableRecordsOnInst(I&: *SI);
7722
7723 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(Bitwidth: 1);
7724
7725 // Can we convert the 'select' to CF ?
7726 if (VectorCond || SI->getMetadata(KindID: LLVMContext::MD_unpredictable))
7727 return false;
7728
7729 TargetLowering::SelectSupportKind SelectKind;
7730 if (SI->getType()->isVectorTy())
7731 SelectKind = TargetLowering::ScalarCondVectorVal;
7732 else
7733 SelectKind = TargetLowering::ScalarValSelect;
7734
7735 if (TLI->isSelectSupported(SelectKind) &&
7736 (!isFormingBranchFromSelectProfitable(TTI, TLI, SI) ||
7737 llvm::shouldOptimizeForSize(BB: SI->getParent(), PSI, BFI)))
7738 return false;
7739
7740 // The DominatorTree needs to be rebuilt by any consumers after this
7741 // transformation. We simply reset here rather than setting the ModifiedDT
7742 // flag to avoid restarting the function walk in runOnFunction for each
7743 // select optimized.
7744 DT.reset();
7745
7746 // Transform a sequence like this:
7747 // start:
7748 // %cmp = cmp uge i32 %a, %b
7749 // %sel = select i1 %cmp, i32 %c, i32 %d
7750 //
7751 // Into:
7752 // start:
7753 // %cmp = cmp uge i32 %a, %b
7754 // %cmp.frozen = freeze %cmp
7755 // br i1 %cmp.frozen, label %select.true, label %select.false
7756 // select.true:
7757 // br label %select.end
7758 // select.false:
7759 // br label %select.end
7760 // select.end:
7761 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
7762 //
7763 // %cmp should be frozen, otherwise it may introduce undefined behavior.
7764 // In addition, we may sink instructions that produce %c or %d from
7765 // the entry block into the destination(s) of the new branch.
7766 // If the true or false blocks do not contain a sunken instruction, that
7767 // block and its branch may be optimized away. In that case, one side of the
7768 // first branch will point directly to select.end, and the corresponding PHI
7769 // predecessor block will be the start block.
7770
7771 // Collect values that go on the true side and the values that go on the false
7772 // side.
7773 SmallVector<Instruction *> TrueInstrs, FalseInstrs;
7774 for (SelectInst *SI : ASI) {
7775 if (Value *V = SI->getTrueValue(); sinkSelectOperand(TTI, V))
7776 TrueInstrs.push_back(Elt: cast<Instruction>(Val: V));
7777 if (Value *V = SI->getFalseValue(); sinkSelectOperand(TTI, V))
7778 FalseInstrs.push_back(Elt: cast<Instruction>(Val: V));
7779 }
7780
7781 // Split the select block, according to how many (if any) values go on each
7782 // side.
7783 BasicBlock *StartBlock = SI->getParent();
7784 BasicBlock::iterator SplitPt = std::next(x: BasicBlock::iterator(LastSI));
7785 // We should split before any debug-info.
7786 SplitPt.setHeadBit(true);
7787
7788 IRBuilder<> IB(SI);
7789 auto *CondFr = IB.CreateFreeze(V: SI->getCondition(), Name: SI->getName() + ".frozen");
7790
7791 BasicBlock *TrueBlock = nullptr;
7792 BasicBlock *FalseBlock = nullptr;
7793 BasicBlock *EndBlock = nullptr;
7794 UncondBrInst *TrueBranch = nullptr;
7795 UncondBrInst *FalseBranch = nullptr;
7796 if (TrueInstrs.size() == 0) {
7797 FalseBranch = cast<UncondBrInst>(Val: SplitBlockAndInsertIfElse(
7798 Cond: CondFr, SplitBefore: SplitPt, Unreachable: false, BranchWeights: nullptr, DTU: nullptr, LI));
7799 FalseBlock = FalseBranch->getParent();
7800 EndBlock = cast<BasicBlock>(Val: FalseBranch->getOperand(i_nocapture: 0));
7801 } else if (FalseInstrs.size() == 0) {
7802 TrueBranch = cast<UncondBrInst>(Val: SplitBlockAndInsertIfThen(
7803 Cond: CondFr, SplitBefore: SplitPt, Unreachable: false, BranchWeights: nullptr, DTU: nullptr, LI));
7804 TrueBlock = TrueBranch->getParent();
7805 EndBlock = TrueBranch->getSuccessor();
7806 } else {
7807 Instruction *ThenTerm = nullptr;
7808 Instruction *ElseTerm = nullptr;
7809 SplitBlockAndInsertIfThenElse(Cond: CondFr, SplitBefore: SplitPt, ThenTerm: &ThenTerm, ElseTerm: &ElseTerm,
7810 BranchWeights: nullptr, DTU: nullptr, LI);
7811 TrueBranch = cast<UncondBrInst>(Val: ThenTerm);
7812 FalseBranch = cast<UncondBrInst>(Val: ElseTerm);
7813 TrueBlock = TrueBranch->getParent();
7814 FalseBlock = FalseBranch->getParent();
7815 EndBlock = TrueBranch->getSuccessor();
7816 }
7817
7818 EndBlock->setName("select.end");
7819 if (TrueBlock)
7820 TrueBlock->setName("select.true.sink");
7821 if (FalseBlock)
7822 FalseBlock->setName(FalseInstrs.size() == 0 ? "select.false"
7823 : "select.false.sink");
7824
7825 if (IsHugeFunc) {
7826 if (TrueBlock)
7827 FreshBBs.insert(Ptr: TrueBlock);
7828 if (FalseBlock)
7829 FreshBBs.insert(Ptr: FalseBlock);
7830 FreshBBs.insert(Ptr: EndBlock);
7831 }
7832
7833 BFI->setBlockFreq(BB: EndBlock, Freq: BFI->getBlockFreq(BB: StartBlock));
7834
7835 static const unsigned MD[] = {
7836 LLVMContext::MD_prof, LLVMContext::MD_unpredictable,
7837 LLVMContext::MD_make_implicit, LLVMContext::MD_dbg};
7838 StartBlock->getTerminator()->copyMetadata(SrcInst: *SI, WL: MD);
7839
7840 // Sink expensive instructions into the conditional blocks to avoid executing
7841 // them speculatively.
7842 for (Instruction *I : TrueInstrs)
7843 I->moveBefore(InsertPos: TrueBranch->getIterator());
7844 for (Instruction *I : FalseInstrs)
7845 I->moveBefore(InsertPos: FalseBranch->getIterator());
7846
7847 // If we did not create a new block for one of the 'true' or 'false' paths
7848 // of the condition, it means that side of the branch goes to the end block
7849 // directly and the path originates from the start block from the point of
7850 // view of the new PHI.
7851 if (TrueBlock == nullptr)
7852 TrueBlock = StartBlock;
7853 else if (FalseBlock == nullptr)
7854 FalseBlock = StartBlock;
7855
7856 SmallPtrSet<const Instruction *, 2> INS(llvm::from_range, ASI);
7857 // Use reverse iterator because later select may use the value of the
7858 // earlier select, and we need to propagate value through earlier select
7859 // to get the PHI operand.
7860 for (SelectInst *SI : llvm::reverse(C&: ASI)) {
7861 // The select itself is replaced with a PHI Node.
7862 PHINode *PN = PHINode::Create(Ty: SI->getType(), NumReservedValues: 2, NameStr: "");
7863 PN->insertBefore(InsertPos: EndBlock->begin());
7864 PN->takeName(V: SI);
7865 PN->addIncoming(V: getTrueOrFalseValue(SI, isTrue: true, Selects: INS), BB: TrueBlock);
7866 PN->addIncoming(V: getTrueOrFalseValue(SI, isTrue: false, Selects: INS), BB: FalseBlock);
7867 PN->setDebugLoc(SI->getDebugLoc());
7868
7869 replaceAllUsesWith(Old: SI, New: PN, FreshBBs, IsHuge: IsHugeFunc);
7870 SI->eraseFromParent();
7871 INS.erase(Ptr: SI);
7872 ++NumSelectsExpanded;
7873 }
7874
7875 // Instruct OptimizeBlock to skip to the next block.
7876 CurInstIterator = StartBlock->end();
7877 return true;
7878}
7879
7880/// Some targets only accept certain types for splat inputs. For example a VDUP
7881/// in MVE takes a GPR (integer) register, and the instruction that incorporate
7882/// a VDUP (such as a VADD qd, qm, rm) also require a gpr register.
7883bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
7884 // Accept shuf(insertelem(undef/poison, val, 0), undef/poison, <0,0,..>) only
7885 if (!match(V: SVI, P: m_Shuffle(v1: m_InsertElt(Val: m_Undef(), Elt: m_Value(), Idx: m_ZeroInt()),
7886 v2: m_Undef(), mask: m_ZeroMask())))
7887 return false;
7888 Type *NewType = TLI->shouldConvertSplatType(SVI);
7889 if (!NewType)
7890 return false;
7891
7892 auto *SVIVecType = cast<FixedVectorType>(Val: SVI->getType());
7893 assert(!NewType->isVectorTy() && "Expected a scalar type!");
7894 assert(NewType->getScalarSizeInBits() == SVIVecType->getScalarSizeInBits() &&
7895 "Expected a type of the same size!");
7896 auto *NewVecType =
7897 FixedVectorType::get(ElementType: NewType, NumElts: SVIVecType->getNumElements());
7898
7899 // Create a bitcast (shuffle (insert (bitcast(..))))
7900 IRBuilder<> Builder(SVI->getContext());
7901 Builder.SetInsertPoint(SVI);
7902 Value *BC1 = Builder.CreateBitCast(
7903 V: cast<Instruction>(Val: SVI->getOperand(i_nocapture: 0))->getOperand(i: 1), DestTy: NewType);
7904 Value *Shuffle = Builder.CreateVectorSplat(NumElts: NewVecType->getNumElements(), V: BC1);
7905 Value *BC2 = Builder.CreateBitCast(V: Shuffle, DestTy: SVIVecType);
7906
7907 replaceAllUsesWith(Old: SVI, New: BC2, FreshBBs, IsHuge: IsHugeFunc);
7908 RecursivelyDeleteTriviallyDeadInstructions(
7909 V: SVI, TLI: TLInfo, MSSAU: nullptr,
7910 AboutToDeleteCallback: [&](Value *V) { removeAllAssertingVHReferences(V); });
7911
7912 // Also hoist the bitcast up to its operand if it they are not in the same
7913 // block.
7914 if (auto *BCI = dyn_cast<Instruction>(Val: BC1))
7915 if (auto *Op = dyn_cast<Instruction>(Val: BCI->getOperand(i: 0)))
7916 if (BCI->getParent() != Op->getParent() && !isa<PHINode>(Val: Op) &&
7917 !Op->isTerminator() && !Op->isEHPad())
7918 BCI->moveAfter(MovePos: Op);
7919
7920 return true;
7921}
7922
7923bool CodeGenPrepare::tryToSinkFreeOperands(Instruction *I) {
7924 // If the operands of I can be folded into a target instruction together with
7925 // I, duplicate and sink them.
7926 SmallVector<Use *, 4> OpsToSink;
7927 if (!TTI->isProfitableToSinkOperands(I, Ops&: OpsToSink))
7928 return false;
7929
7930 // OpsToSink can contain multiple uses in a use chain (e.g.
7931 // (%u1 with %u1 = shufflevector), (%u2 with %u2 = zext %u1)). The dominating
7932 // uses must come first, so we process the ops in reverse order so as to not
7933 // create invalid IR.
7934 BasicBlock *TargetBB = I->getParent();
7935 bool Changed = false;
7936 SmallVector<Use *, 4> ToReplace;
7937 Instruction *InsertPoint = I;
7938 DenseMap<const Instruction *, unsigned long> InstOrdering;
7939 unsigned long InstNumber = 0;
7940 for (const auto &I : *TargetBB)
7941 InstOrdering[&I] = InstNumber++;
7942
7943 for (Use *U : reverse(C&: OpsToSink)) {
7944 auto *UI = cast<Instruction>(Val: U->get());
7945 if (isa<PHINode>(Val: UI) || UI->mayHaveSideEffects() || UI->mayReadFromMemory())
7946 continue;
7947 if (UI->getParent() == TargetBB) {
7948 if (InstOrdering[UI] < InstOrdering[InsertPoint])
7949 InsertPoint = UI;
7950 continue;
7951 }
7952 ToReplace.push_back(Elt: U);
7953 }
7954
7955 SetVector<Instruction *> MaybeDead;
7956 DenseMap<Instruction *, Instruction *> NewInstructions;
7957 for (Use *U : ToReplace) {
7958 auto *UI = cast<Instruction>(Val: U->get());
7959 Instruction *NI = UI->clone();
7960
7961 if (IsHugeFunc) {
7962 // Now we clone an instruction, its operands' defs may sink to this BB
7963 // now. So we put the operands defs' BBs into FreshBBs to do optimization.
7964 for (Value *Op : NI->operands())
7965 if (auto *OpDef = dyn_cast<Instruction>(Val: Op))
7966 FreshBBs.insert(Ptr: OpDef->getParent());
7967 }
7968
7969 NewInstructions[UI] = NI;
7970 MaybeDead.insert(X: UI);
7971 LLVM_DEBUG(dbgs() << "Sinking " << *UI << " to user " << *I << "\n");
7972 NI->insertBefore(InsertPos: InsertPoint->getIterator());
7973 InsertPoint = NI;
7974 InsertedInsts.insert(Ptr: NI);
7975
7976 // Update the use for the new instruction, making sure that we update the
7977 // sunk instruction uses, if it is part of a chain that has already been
7978 // sunk.
7979 Instruction *OldI = cast<Instruction>(Val: U->getUser());
7980 if (auto It = NewInstructions.find(Val: OldI); It != NewInstructions.end())
7981 It->second->setOperand(i: U->getOperandNo(), Val: NI);
7982 else
7983 U->set(NI);
7984 Changed = true;
7985 }
7986
7987 // Remove instructions that are dead after sinking.
7988 for (auto *I : MaybeDead) {
7989 if (!I->hasNUsesOrMore(N: 1)) {
7990 LLVM_DEBUG(dbgs() << "Removing dead instruction: " << *I << "\n");
7991 I->eraseFromParent();
7992 }
7993 }
7994
7995 return Changed;
7996}
7997
7998bool CodeGenPrepare::optimizeSwitchType(SwitchInst *SI) {
7999 Value *Cond = SI->getCondition();
8000 Type *OldType = Cond->getType();
8001 LLVMContext &Context = Cond->getContext();
8002 EVT OldVT = TLI->getValueType(DL: *DL, Ty: OldType);
8003 MVT RegType = TLI->getPreferredSwitchConditionType(Context, ConditionVT: OldVT);
8004 unsigned RegWidth = RegType.getSizeInBits();
8005
8006 if (RegWidth <= cast<IntegerType>(Val: OldType)->getBitWidth())
8007 return false;
8008
8009 // If the register width is greater than the type width, expand the condition
8010 // of the switch instruction and each case constant to the width of the
8011 // register. By widening the type of the switch condition, subsequent
8012 // comparisons (for case comparisons) will not need to be extended to the
8013 // preferred register width, so we will potentially eliminate N-1 extends,
8014 // where N is the number of cases in the switch.
8015 auto *NewType = Type::getIntNTy(C&: Context, N: RegWidth);
8016
8017 // Extend the switch condition and case constants using the target preferred
8018 // extend unless the switch condition is a function argument with an extend
8019 // attribute. In that case, we can avoid an unnecessary mask/extension by
8020 // matching the argument extension instead.
8021 Instruction::CastOps ExtType = Instruction::ZExt;
8022 // Some targets prefer SExt over ZExt.
8023 if (TLI->isSExtCheaperThanZExt(FromTy: OldVT, ToTy: RegType))
8024 ExtType = Instruction::SExt;
8025
8026 if (auto *Arg = dyn_cast<Argument>(Val: Cond)) {
8027 if (Arg->hasSExtAttr())
8028 ExtType = Instruction::SExt;
8029 if (Arg->hasZExtAttr())
8030 ExtType = Instruction::ZExt;
8031 }
8032
8033 auto *ExtInst = CastInst::Create(ExtType, S: Cond, Ty: NewType);
8034 ExtInst->insertBefore(InsertPos: SI->getIterator());
8035 ExtInst->setDebugLoc(SI->getDebugLoc());
8036 SI->setCondition(ExtInst);
8037 for (auto Case : SI->cases()) {
8038 const APInt &NarrowConst = Case.getCaseValue()->getValue();
8039 APInt WideConst = (ExtType == Instruction::ZExt)
8040 ? NarrowConst.zext(width: RegWidth)
8041 : NarrowConst.sext(width: RegWidth);
8042 Case.setValue(ConstantInt::get(Context, V: WideConst));
8043 }
8044
8045 return true;
8046}
8047
8048bool CodeGenPrepare::optimizeSwitchPhiConstants(SwitchInst *SI) {
8049 // The SCCP optimization tends to produce code like this:
8050 // switch(x) { case 42: phi(42, ...) }
8051 // Materializing the constant for the phi-argument needs instructions; So we
8052 // change the code to:
8053 // switch(x) { case 42: phi(x, ...) }
8054
8055 Value *Condition = SI->getCondition();
8056 // Avoid endless loop in degenerate case.
8057 if (isa<ConstantInt>(Val: *Condition))
8058 return false;
8059
8060 bool Changed = false;
8061 BasicBlock *SwitchBB = SI->getParent();
8062 Type *ConditionType = Condition->getType();
8063
8064 for (const SwitchInst::CaseHandle &Case : SI->cases()) {
8065 ConstantInt *CaseValue = Case.getCaseValue();
8066 BasicBlock *CaseBB = Case.getCaseSuccessor();
8067 // Set to true if we previously checked that `CaseBB` is only reached by
8068 // a single case from this switch.
8069 bool CheckedForSinglePred = false;
8070 for (PHINode &PHI : CaseBB->phis()) {
8071 Type *PHIType = PHI.getType();
8072 // If ZExt is free then we can also catch patterns like this:
8073 // switch((i32)x) { case 42: phi((i64)42, ...); }
8074 // and replace `(i64)42` with `zext i32 %x to i64`.
8075 bool TryZExt =
8076 PHIType->isIntegerTy() &&
8077 PHIType->getIntegerBitWidth() > ConditionType->getIntegerBitWidth() &&
8078 TLI->isZExtFree(FromTy: ConditionType, ToTy: PHIType);
8079 if (PHIType == ConditionType || TryZExt) {
8080 // Set to true to skip this case because of multiple preds.
8081 bool SkipCase = false;
8082 Value *Replacement = nullptr;
8083 for (unsigned I = 0, E = PHI.getNumIncomingValues(); I != E; I++) {
8084 Value *PHIValue = PHI.getIncomingValue(i: I);
8085 if (PHIValue != CaseValue) {
8086 if (!TryZExt)
8087 continue;
8088 ConstantInt *PHIValueInt = dyn_cast<ConstantInt>(Val: PHIValue);
8089 if (!PHIValueInt ||
8090 PHIValueInt->getValue() !=
8091 CaseValue->getValue().zext(width: PHIType->getIntegerBitWidth()))
8092 continue;
8093 }
8094 if (PHI.getIncomingBlock(i: I) != SwitchBB)
8095 continue;
8096 // We cannot optimize if there are multiple case labels jumping to
8097 // this block. This check may get expensive when there are many
8098 // case labels so we test for it last.
8099 if (!CheckedForSinglePred) {
8100 CheckedForSinglePred = true;
8101 if (SI->findCaseDest(BB: CaseBB) == nullptr) {
8102 SkipCase = true;
8103 break;
8104 }
8105 }
8106
8107 if (Replacement == nullptr) {
8108 if (PHIValue == CaseValue) {
8109 Replacement = Condition;
8110 } else {
8111 IRBuilder<> Builder(SI);
8112 Replacement = Builder.CreateZExt(V: Condition, DestTy: PHIType);
8113 }
8114 }
8115 PHI.setIncomingValue(i: I, V: Replacement);
8116 Changed = true;
8117 }
8118 if (SkipCase)
8119 break;
8120 }
8121 }
8122 }
8123 return Changed;
8124}
8125
8126bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
8127 bool Changed = optimizeSwitchType(SI);
8128 Changed |= optimizeSwitchPhiConstants(SI);
8129 return Changed;
8130}
8131
8132namespace {
8133
8134/// Helper class to promote a scalar operation to a vector one.
8135/// This class is used to move downward extractelement transition.
8136/// E.g.,
8137/// a = vector_op <2 x i32>
8138/// b = extractelement <2 x i32> a, i32 0
8139/// c = scalar_op b
8140/// store c
8141///
8142/// =>
8143/// a = vector_op <2 x i32>
8144/// c = vector_op a (equivalent to scalar_op on the related lane)
8145/// * d = extractelement <2 x i32> c, i32 0
8146/// * store d
8147/// Assuming both extractelement and store can be combine, we get rid of the
8148/// transition.
8149class VectorPromoteHelper {
8150 /// DataLayout associated with the current module.
8151 const DataLayout &DL;
8152
8153 /// Used to perform some checks on the legality of vector operations.
8154 const TargetLowering &TLI;
8155
8156 /// Used to estimated the cost of the promoted chain.
8157 const TargetTransformInfo &TTI;
8158
8159 /// The transition being moved downwards.
8160 Instruction *Transition;
8161
8162 /// The sequence of instructions to be promoted.
8163 SmallVector<Instruction *, 4> InstsToBePromoted;
8164
8165 /// Cost of combining a store and an extract.
8166 unsigned StoreExtractCombineCost;
8167
8168 /// Instruction that will be combined with the transition.
8169 Instruction *CombineInst = nullptr;
8170
8171 /// The instruction that represents the current end of the transition.
8172 /// Since we are faking the promotion until we reach the end of the chain
8173 /// of computation, we need a way to get the current end of the transition.
8174 Instruction *getEndOfTransition() const {
8175 if (InstsToBePromoted.empty())
8176 return Transition;
8177 return InstsToBePromoted.back();
8178 }
8179
8180 /// Return the index of the original value in the transition.
8181 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
8182 /// c, is at index 0.
8183 unsigned getTransitionOriginalValueIdx() const {
8184 assert(isa<ExtractElementInst>(Transition) &&
8185 "Other kind of transitions are not supported yet");
8186 return 0;
8187 }
8188
8189 /// Return the index of the index in the transition.
8190 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
8191 /// is at index 1.
8192 unsigned getTransitionIdx() const {
8193 assert(isa<ExtractElementInst>(Transition) &&
8194 "Other kind of transitions are not supported yet");
8195 return 1;
8196 }
8197
8198 /// Get the type of the transition.
8199 /// This is the type of the original value.
8200 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
8201 /// transition is <2 x i32>.
8202 Type *getTransitionType() const {
8203 return Transition->getOperand(i: getTransitionOriginalValueIdx())->getType();
8204 }
8205
8206 /// Promote \p ToBePromoted by moving \p Def downward through.
8207 /// I.e., we have the following sequence:
8208 /// Def = Transition <ty1> a to <ty2>
8209 /// b = ToBePromoted <ty2> Def, ...
8210 /// =>
8211 /// b = ToBePromoted <ty1> a, ...
8212 /// Def = Transition <ty1> ToBePromoted to <ty2>
8213 void promoteImpl(Instruction *ToBePromoted);
8214
8215 /// Check whether or not it is profitable to promote all the
8216 /// instructions enqueued to be promoted.
8217 bool isProfitableToPromote() {
8218 Value *ValIdx = Transition->getOperand(i: getTransitionOriginalValueIdx());
8219 unsigned Index = isa<ConstantInt>(Val: ValIdx)
8220 ? cast<ConstantInt>(Val: ValIdx)->getZExtValue()
8221 : -1;
8222 Type *PromotedType = getTransitionType();
8223
8224 StoreInst *ST = cast<StoreInst>(Val: CombineInst);
8225 unsigned AS = ST->getPointerAddressSpace();
8226 // Check if this store is supported.
8227 if (!TLI.allowsMisalignedMemoryAccesses(
8228 TLI.getValueType(DL, Ty: ST->getValueOperand()->getType()), AddrSpace: AS,
8229 Alignment: ST->getAlign())) {
8230 // If this is not supported, there is no way we can combine
8231 // the extract with the store.
8232 return false;
8233 }
8234
8235 // The scalar chain of computation has to pay for the transition
8236 // scalar to vector.
8237 // The vector chain has to account for the combining cost.
8238 enum TargetTransformInfo::TargetCostKind CostKind =
8239 TargetTransformInfo::TCK_RecipThroughput;
8240 InstructionCost ScalarCost =
8241 TTI.getVectorInstrCost(I: *Transition, Val: PromotedType, CostKind, Index);
8242 InstructionCost VectorCost = StoreExtractCombineCost;
8243 for (const auto &Inst : InstsToBePromoted) {
8244 // Compute the cost.
8245 // By construction, all instructions being promoted are arithmetic ones.
8246 // Moreover, one argument is a constant that can be viewed as a splat
8247 // constant.
8248 Value *Arg0 = Inst->getOperand(i: 0);
8249 bool IsArg0Constant = isa<UndefValue>(Val: Arg0) || isa<ConstantInt>(Val: Arg0) ||
8250 isa<ConstantFP>(Val: Arg0);
8251 TargetTransformInfo::OperandValueInfo Arg0Info, Arg1Info;
8252 if (IsArg0Constant)
8253 Arg0Info.Kind = TargetTransformInfo::OK_UniformConstantValue;
8254 else
8255 Arg1Info.Kind = TargetTransformInfo::OK_UniformConstantValue;
8256
8257 ScalarCost += TTI.getArithmeticInstrCost(
8258 Opcode: Inst->getOpcode(), Ty: Inst->getType(), CostKind, Opd1Info: Arg0Info, Opd2Info: Arg1Info);
8259 VectorCost += TTI.getArithmeticInstrCost(Opcode: Inst->getOpcode(), Ty: PromotedType,
8260 CostKind, Opd1Info: Arg0Info, Opd2Info: Arg1Info);
8261 }
8262 LLVM_DEBUG(
8263 dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
8264 << ScalarCost << "\nVector: " << VectorCost << '\n');
8265 return ScalarCost > VectorCost;
8266 }
8267
8268 /// Generate a constant vector with \p Val with the same
8269 /// number of elements as the transition.
8270 /// \p UseSplat defines whether or not \p Val should be replicated
8271 /// across the whole vector.
8272 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
8273 /// otherwise we generate a vector with as many poison as possible:
8274 /// <poison, ..., poison, Val, poison, ..., poison> where \p Val is only
8275 /// used at the index of the extract.
8276 Value *getConstantVector(Constant *Val, bool UseSplat) const {
8277 unsigned ExtractIdx = std::numeric_limits<unsigned>::max();
8278 if (!UseSplat) {
8279 // If we cannot determine where the constant must be, we have to
8280 // use a splat constant.
8281 Value *ValExtractIdx = Transition->getOperand(i: getTransitionIdx());
8282 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Val: ValExtractIdx))
8283 ExtractIdx = CstVal->getSExtValue();
8284 else
8285 UseSplat = true;
8286 }
8287
8288 ElementCount EC = cast<VectorType>(Val: getTransitionType())->getElementCount();
8289 if (UseSplat)
8290 return ConstantVector::getSplat(EC, Elt: Val);
8291
8292 if (!EC.isScalable()) {
8293 SmallVector<Constant *, 4> ConstVec;
8294 PoisonValue *PoisonVal = PoisonValue::get(T: Val->getType());
8295 for (unsigned Idx = 0; Idx != EC.getKnownMinValue(); ++Idx) {
8296 if (Idx == ExtractIdx)
8297 ConstVec.push_back(Elt: Val);
8298 else
8299 ConstVec.push_back(Elt: PoisonVal);
8300 }
8301 return ConstantVector::get(V: ConstVec);
8302 } else
8303 llvm_unreachable(
8304 "Generate scalable vector for non-splat is unimplemented");
8305 }
8306
8307 /// Check if promoting to a vector type an operand at \p OperandIdx
8308 /// in \p Use can trigger undefined behavior.
8309 static bool canCauseUndefinedBehavior(const Instruction *Use,
8310 unsigned OperandIdx) {
8311 // This is not safe to introduce undef when the operand is on
8312 // the right hand side of a division-like instruction.
8313 if (OperandIdx != 1)
8314 return false;
8315 switch (Use->getOpcode()) {
8316 default:
8317 return false;
8318 case Instruction::SDiv:
8319 case Instruction::UDiv:
8320 case Instruction::SRem:
8321 case Instruction::URem:
8322 return true;
8323 case Instruction::FDiv:
8324 case Instruction::FRem:
8325 return !Use->hasNoNaNs();
8326 }
8327 llvm_unreachable(nullptr);
8328 }
8329
8330public:
8331 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
8332 const TargetTransformInfo &TTI, Instruction *Transition,
8333 unsigned CombineCost)
8334 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
8335 StoreExtractCombineCost(CombineCost) {
8336 assert(Transition && "Do not know how to promote null");
8337 }
8338
8339 /// Check if we can promote \p ToBePromoted to \p Type.
8340 bool canPromote(const Instruction *ToBePromoted) const {
8341 // We could support CastInst too.
8342 return isa<BinaryOperator>(Val: ToBePromoted);
8343 }
8344
8345 /// Check if it is profitable to promote \p ToBePromoted
8346 /// by moving downward the transition through.
8347 bool shouldPromote(const Instruction *ToBePromoted) const {
8348 // Promote only if all the operands can be statically expanded.
8349 // Indeed, we do not want to introduce any new kind of transitions.
8350 for (const Use &U : ToBePromoted->operands()) {
8351 const Value *Val = U.get();
8352 if (Val == getEndOfTransition()) {
8353 // If the use is a division and the transition is on the rhs,
8354 // we cannot promote the operation, otherwise we may create a
8355 // division by zero.
8356 if (canCauseUndefinedBehavior(Use: ToBePromoted, OperandIdx: U.getOperandNo()))
8357 return false;
8358 continue;
8359 }
8360 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
8361 !isa<ConstantFP>(Val))
8362 return false;
8363 }
8364 // Check that the resulting operation is legal.
8365 int ISDOpcode = TLI.InstructionOpcodeToISD(Opcode: ToBePromoted->getOpcode());
8366 if (!ISDOpcode)
8367 return false;
8368 return StressStoreExtract ||
8369 TLI.isOperationLegalOrCustom(
8370 Op: ISDOpcode, VT: TLI.getValueType(DL, Ty: getTransitionType(), AllowUnknown: true));
8371 }
8372
8373 /// Check whether or not \p Use can be combined
8374 /// with the transition.
8375 /// I.e., is it possible to do Use(Transition) => AnotherUse?
8376 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Val: Use); }
8377
8378 /// Record \p ToBePromoted as part of the chain to be promoted.
8379 void enqueueForPromotion(Instruction *ToBePromoted) {
8380 InstsToBePromoted.push_back(Elt: ToBePromoted);
8381 }
8382
8383 /// Set the instruction that will be combined with the transition.
8384 void recordCombineInstruction(Instruction *ToBeCombined) {
8385 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
8386 CombineInst = ToBeCombined;
8387 }
8388
8389 /// Promote all the instructions enqueued for promotion if it is
8390 /// is profitable.
8391 /// \return True if the promotion happened, false otherwise.
8392 bool promote() {
8393 // Check if there is something to promote.
8394 // Right now, if we do not have anything to combine with,
8395 // we assume the promotion is not profitable.
8396 if (InstsToBePromoted.empty() || !CombineInst)
8397 return false;
8398
8399 // Check cost.
8400 if (!StressStoreExtract && !isProfitableToPromote())
8401 return false;
8402
8403 // Promote.
8404 for (auto &ToBePromoted : InstsToBePromoted)
8405 promoteImpl(ToBePromoted);
8406 InstsToBePromoted.clear();
8407 return true;
8408 }
8409};
8410
8411} // end anonymous namespace
8412
8413void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
8414 // At this point, we know that all the operands of ToBePromoted but Def
8415 // can be statically promoted.
8416 // For Def, we need to use its parameter in ToBePromoted:
8417 // b = ToBePromoted ty1 a
8418 // Def = Transition ty1 b to ty2
8419 // Move the transition down.
8420 // 1. Replace all uses of the promoted operation by the transition.
8421 // = ... b => = ... Def.
8422 assert(ToBePromoted->getType() == Transition->getType() &&
8423 "The type of the result of the transition does not match "
8424 "the final type");
8425 ToBePromoted->replaceAllUsesWith(V: Transition);
8426 // 2. Update the type of the uses.
8427 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
8428 Type *TransitionTy = getTransitionType();
8429 ToBePromoted->mutateType(Ty: TransitionTy);
8430 // 3. Update all the operands of the promoted operation with promoted
8431 // operands.
8432 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
8433 for (Use &U : ToBePromoted->operands()) {
8434 Value *Val = U.get();
8435 Value *NewVal = nullptr;
8436 if (Val == Transition)
8437 NewVal = Transition->getOperand(i: getTransitionOriginalValueIdx());
8438 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
8439 isa<ConstantFP>(Val)) {
8440 // Use a splat constant if it is not safe to use undef.
8441 NewVal = getConstantVector(
8442 Val: cast<Constant>(Val),
8443 UseSplat: isa<UndefValue>(Val) ||
8444 canCauseUndefinedBehavior(Use: ToBePromoted, OperandIdx: U.getOperandNo()));
8445 } else
8446 llvm_unreachable("Did you modified shouldPromote and forgot to update "
8447 "this?");
8448 ToBePromoted->setOperand(i: U.getOperandNo(), Val: NewVal);
8449 }
8450 Transition->moveAfter(MovePos: ToBePromoted);
8451 Transition->setOperand(i: getTransitionOriginalValueIdx(), Val: ToBePromoted);
8452}
8453
8454/// Some targets can do store(extractelement) with one instruction.
8455/// Try to push the extractelement towards the stores when the target
8456/// has this feature and this is profitable.
8457bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
8458 unsigned CombineCost = std::numeric_limits<unsigned>::max();
8459 if (DisableStoreExtract ||
8460 (!StressStoreExtract &&
8461 !TLI->canCombineStoreAndExtract(VectorTy: Inst->getOperand(i: 0)->getType(),
8462 Idx: Inst->getOperand(i: 1), Cost&: CombineCost)))
8463 return false;
8464
8465 // At this point we know that Inst is a vector to scalar transition.
8466 // Try to move it down the def-use chain, until:
8467 // - We can combine the transition with its single use
8468 // => we got rid of the transition.
8469 // - We escape the current basic block
8470 // => we would need to check that we are moving it at a cheaper place and
8471 // we do not do that for now.
8472 BasicBlock *Parent = Inst->getParent();
8473 LLVM_DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
8474 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
8475 // If the transition has more than one use, assume this is not going to be
8476 // beneficial.
8477 while (Inst->hasOneUse()) {
8478 Instruction *ToBePromoted = cast<Instruction>(Val: *Inst->user_begin());
8479 LLVM_DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
8480
8481 if (ToBePromoted->getParent() != Parent) {
8482 LLVM_DEBUG(dbgs() << "Instruction to promote is in a different block ("
8483 << ToBePromoted->getParent()->getName()
8484 << ") than the transition (" << Parent->getName()
8485 << ").\n");
8486 return false;
8487 }
8488
8489 if (VPH.canCombine(Use: ToBePromoted)) {
8490 LLVM_DEBUG(dbgs() << "Assume " << *Inst << '\n'
8491 << "will be combined with: " << *ToBePromoted << '\n');
8492 VPH.recordCombineInstruction(ToBeCombined: ToBePromoted);
8493 bool Changed = VPH.promote();
8494 NumStoreExtractExposed += Changed;
8495 return Changed;
8496 }
8497
8498 LLVM_DEBUG(dbgs() << "Try promoting.\n");
8499 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
8500 return false;
8501
8502 LLVM_DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
8503
8504 VPH.enqueueForPromotion(ToBePromoted);
8505 Inst = ToBePromoted;
8506 }
8507 return false;
8508}
8509
8510/// For the instruction sequence of store below, F and I values
8511/// are bundled together as an i64 value before being stored into memory.
8512/// Sometimes it is more efficient to generate separate stores for F and I,
8513/// which can remove the bitwise instructions or sink them to colder places.
8514///
8515/// (store (or (zext (bitcast F to i32) to i64),
8516/// (shl (zext I to i64), 32)), addr) -->
8517/// (store F, addr) and (store I, addr+4)
8518///
8519/// Similarly, splitting for other merged store can also be beneficial, like:
8520/// For pair of {i32, i32}, i64 store --> two i32 stores.
8521/// For pair of {i32, i16}, i64 store --> two i32 stores.
8522/// For pair of {i16, i16}, i32 store --> two i16 stores.
8523/// For pair of {i16, i8}, i32 store --> two i16 stores.
8524/// For pair of {i8, i8}, i16 store --> two i8 stores.
8525///
8526/// We allow each target to determine specifically which kind of splitting is
8527/// supported.
8528///
8529/// The store patterns are commonly seen from the simple code snippet below
8530/// if only std::make_pair(...) is sroa transformed before inlined into hoo.
8531/// void goo(const std::pair<int, float> &);
8532/// hoo() {
8533/// ...
8534/// goo(std::make_pair(tmp, ftmp));
8535/// ...
8536/// }
8537///
8538/// Although we already have similar splitting in DAG Combine, we duplicate
8539/// it in CodeGenPrepare to catch the case in which pattern is across
8540/// multiple BBs. The logic in DAG Combine is kept to catch case generated
8541/// during code expansion.
8542static bool splitMergedValStore(StoreInst &SI, const DataLayout &DL,
8543 const TargetLowering &TLI) {
8544 // Handle simple but common cases only.
8545 Type *StoreType = SI.getValueOperand()->getType();
8546
8547 // The code below assumes shifting a value by <number of bits>,
8548 // whereas scalable vectors would have to be shifted by
8549 // <2log(vscale) + number of bits> in order to store the
8550 // low/high parts. Bailing out for now.
8551 if (StoreType->isScalableTy())
8552 return false;
8553
8554 if (!DL.typeSizeEqualsStoreSize(Ty: StoreType) ||
8555 DL.getTypeSizeInBits(Ty: StoreType) == 0)
8556 return false;
8557
8558 unsigned HalfValBitSize = DL.getTypeSizeInBits(Ty: StoreType) / 2;
8559 Type *SplitStoreType = Type::getIntNTy(C&: SI.getContext(), N: HalfValBitSize);
8560 if (!DL.typeSizeEqualsStoreSize(Ty: SplitStoreType))
8561 return false;
8562
8563 // Don't split the store if it is volatile.
8564 if (SI.isVolatile())
8565 return false;
8566
8567 // Match the following patterns:
8568 // (store (or (zext LValue to i64),
8569 // (shl (zext HValue to i64), 32)), HalfValBitSize)
8570 // or
8571 // (store (or (shl (zext HValue to i64), 32)), HalfValBitSize)
8572 // (zext LValue to i64),
8573 // Expect both operands of OR and the first operand of SHL have only
8574 // one use.
8575 Value *LValue, *HValue;
8576 if (!match(V: SI.getValueOperand(),
8577 P: m_c_Or(L: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: LValue))),
8578 R: m_OneUse(SubPattern: m_Shl(L: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: HValue))),
8579 R: m_SpecificInt(V: HalfValBitSize))))))
8580 return false;
8581
8582 // Check LValue and HValue are int with size less or equal than 32.
8583 if (!LValue->getType()->isIntegerTy() ||
8584 DL.getTypeSizeInBits(Ty: LValue->getType()) > HalfValBitSize ||
8585 !HValue->getType()->isIntegerTy() ||
8586 DL.getTypeSizeInBits(Ty: HValue->getType()) > HalfValBitSize)
8587 return false;
8588
8589 // If LValue/HValue is a bitcast instruction, use the EVT before bitcast
8590 // as the input of target query.
8591 auto *LBC = dyn_cast<BitCastInst>(Val: LValue);
8592 auto *HBC = dyn_cast<BitCastInst>(Val: HValue);
8593 EVT LowTy = LBC ? EVT::getEVT(Ty: LBC->getOperand(i_nocapture: 0)->getType())
8594 : EVT::getEVT(Ty: LValue->getType());
8595 EVT HighTy = HBC ? EVT::getEVT(Ty: HBC->getOperand(i_nocapture: 0)->getType())
8596 : EVT::getEVT(Ty: HValue->getType());
8597 if (!ForceSplitStore && !TLI.isMultiStoresCheaperThanBitsMerge(LTy: LowTy, HTy: HighTy))
8598 return false;
8599
8600 // Start to split store.
8601 IRBuilder<> Builder(SI.getContext());
8602 Builder.SetInsertPoint(&SI);
8603
8604 // If LValue/HValue is a bitcast in another BB, create a new one in current
8605 // BB so it may be merged with the splitted stores by dag combiner.
8606 if (LBC && LBC->getParent() != SI.getParent())
8607 LValue = Builder.CreateBitCast(V: LBC->getOperand(i_nocapture: 0), DestTy: LBC->getType());
8608 if (HBC && HBC->getParent() != SI.getParent())
8609 HValue = Builder.CreateBitCast(V: HBC->getOperand(i_nocapture: 0), DestTy: HBC->getType());
8610
8611 bool IsLE = SI.getDataLayout().isLittleEndian();
8612 auto CreateSplitStore = [&](Value *V, bool Upper) {
8613 V = Builder.CreateZExtOrBitCast(V, DestTy: SplitStoreType);
8614 Value *Addr = SI.getPointerOperand();
8615 Align Alignment = SI.getAlign();
8616 const bool IsOffsetStore = (IsLE && Upper) || (!IsLE && !Upper);
8617 if (IsOffsetStore) {
8618 Addr = Builder.CreateGEP(
8619 Ty: SplitStoreType, Ptr: Addr,
8620 IdxList: ConstantInt::get(Ty: Type::getInt32Ty(C&: SI.getContext()), V: 1));
8621
8622 // When splitting the store in half, naturally one half will retain the
8623 // alignment of the original wider store, regardless of whether it was
8624 // over-aligned or not, while the other will require adjustment.
8625 Alignment = commonAlignment(A: Alignment, Offset: HalfValBitSize / 8);
8626 }
8627 Builder.CreateAlignedStore(Val: V, Ptr: Addr, Align: Alignment);
8628 };
8629
8630 CreateSplitStore(LValue, false);
8631 CreateSplitStore(HValue, true);
8632
8633 // Delete the old store.
8634 SI.eraseFromParent();
8635 return true;
8636}
8637
8638// Return true if the GEP has two operands, the first operand is of a sequential
8639// type, and the second operand is a constant.
8640static bool GEPSequentialConstIndexed(GetElementPtrInst *GEP) {
8641 gep_type_iterator I = gep_type_begin(GEP: *GEP);
8642 return GEP->getNumOperands() == 2 && I.isSequential() &&
8643 isa<ConstantInt>(Val: GEP->getOperand(i_nocapture: 1));
8644}
8645
8646// Try unmerging GEPs to reduce liveness interference (register pressure) across
8647// IndirectBr edges. Since IndirectBr edges tend to touch on many blocks,
8648// reducing liveness interference across those edges benefits global register
8649// allocation. Currently handles only certain cases.
8650//
8651// For example, unmerge %GEPI and %UGEPI as below.
8652//
8653// ---------- BEFORE ----------
8654// SrcBlock:
8655// ...
8656// %GEPIOp = ...
8657// ...
8658// %GEPI = gep %GEPIOp, Idx
8659// ...
8660// indirectbr ... [ label %DstB0, label %DstB1, ... label %DstBi ... ]
8661// (* %GEPI is alive on the indirectbr edges due to other uses ahead)
8662// (* %GEPIOp is alive on the indirectbr edges only because of it's used by
8663// %UGEPI)
8664//
8665// DstB0: ... (there may be a gep similar to %UGEPI to be unmerged)
8666// DstB1: ... (there may be a gep similar to %UGEPI to be unmerged)
8667// ...
8668//
8669// DstBi:
8670// ...
8671// %UGEPI = gep %GEPIOp, UIdx
8672// ...
8673// ---------------------------
8674//
8675// ---------- AFTER ----------
8676// SrcBlock:
8677// ... (same as above)
8678// (* %GEPI is still alive on the indirectbr edges)
8679// (* %GEPIOp is no longer alive on the indirectbr edges as a result of the
8680// unmerging)
8681// ...
8682//
8683// DstBi:
8684// ...
8685// %UGEPI = gep %GEPI, (UIdx-Idx)
8686// ...
8687// ---------------------------
8688//
8689// The register pressure on the IndirectBr edges is reduced because %GEPIOp is
8690// no longer alive on them.
8691//
8692// We try to unmerge GEPs here in CodGenPrepare, as opposed to limiting merging
8693// of GEPs in the first place in InstCombiner::visitGetElementPtrInst() so as
8694// not to disable further simplications and optimizations as a result of GEP
8695// merging.
8696//
8697// Note this unmerging may increase the length of the data flow critical path
8698// (the path from %GEPIOp to %UGEPI would go through %GEPI), which is a tradeoff
8699// between the register pressure and the length of data-flow critical
8700// path. Restricting this to the uncommon IndirectBr case would minimize the
8701// impact of potentially longer critical path, if any, and the impact on compile
8702// time.
8703static bool tryUnmergingGEPsAcrossIndirectBr(GetElementPtrInst *GEPI,
8704 const TargetTransformInfo *TTI) {
8705 BasicBlock *SrcBlock = GEPI->getParent();
8706 // Check that SrcBlock ends with an IndirectBr. If not, give up. The common
8707 // (non-IndirectBr) cases exit early here.
8708 if (!isa<IndirectBrInst>(Val: SrcBlock->getTerminator()))
8709 return false;
8710 // Check that GEPI is a simple gep with a single constant index.
8711 if (!GEPSequentialConstIndexed(GEP: GEPI))
8712 return false;
8713 ConstantInt *GEPIIdx = cast<ConstantInt>(Val: GEPI->getOperand(i_nocapture: 1));
8714 // Check that GEPI is a cheap one.
8715 if (TTI->getIntImmCost(Imm: GEPIIdx->getValue(), Ty: GEPIIdx->getType(),
8716 CostKind: TargetTransformInfo::TCK_SizeAndLatency) >
8717 TargetTransformInfo::TCC_Basic)
8718 return false;
8719 Value *GEPIOp = GEPI->getOperand(i_nocapture: 0);
8720 // Check that GEPIOp is an instruction that's also defined in SrcBlock.
8721 if (!isa<Instruction>(Val: GEPIOp))
8722 return false;
8723 auto *GEPIOpI = cast<Instruction>(Val: GEPIOp);
8724 if (GEPIOpI->getParent() != SrcBlock)
8725 return false;
8726 // Check that GEP is used outside the block, meaning it's alive on the
8727 // IndirectBr edge(s).
8728 if (llvm::none_of(Range: GEPI->users(), P: [&](User *Usr) {
8729 if (auto *I = dyn_cast<Instruction>(Val: Usr)) {
8730 if (I->getParent() != SrcBlock) {
8731 return true;
8732 }
8733 }
8734 return false;
8735 }))
8736 return false;
8737 // The second elements of the GEP chains to be unmerged.
8738 std::vector<GetElementPtrInst *> UGEPIs;
8739 // Check each user of GEPIOp to check if unmerging would make GEPIOp not alive
8740 // on IndirectBr edges.
8741 for (User *Usr : GEPIOp->users()) {
8742 if (Usr == GEPI)
8743 continue;
8744 // Check if Usr is an Instruction. If not, give up.
8745 if (!isa<Instruction>(Val: Usr))
8746 return false;
8747 auto *UI = cast<Instruction>(Val: Usr);
8748 // Check if Usr in the same block as GEPIOp, which is fine, skip.
8749 if (UI->getParent() == SrcBlock)
8750 continue;
8751 // Check if Usr is a GEP. If not, give up.
8752 if (!isa<GetElementPtrInst>(Val: Usr))
8753 return false;
8754 auto *UGEPI = cast<GetElementPtrInst>(Val: Usr);
8755 // Check if UGEPI is a simple gep with a single constant index and GEPIOp is
8756 // the pointer operand to it. If so, record it in the vector. If not, give
8757 // up.
8758 if (!GEPSequentialConstIndexed(GEP: UGEPI))
8759 return false;
8760 if (UGEPI->getOperand(i_nocapture: 0) != GEPIOp)
8761 return false;
8762 if (UGEPI->getSourceElementType() != GEPI->getSourceElementType())
8763 return false;
8764 if (GEPIIdx->getType() !=
8765 cast<ConstantInt>(Val: UGEPI->getOperand(i_nocapture: 1))->getType())
8766 return false;
8767 ConstantInt *UGEPIIdx = cast<ConstantInt>(Val: UGEPI->getOperand(i_nocapture: 1));
8768 if (TTI->getIntImmCost(Imm: UGEPIIdx->getValue(), Ty: UGEPIIdx->getType(),
8769 CostKind: TargetTransformInfo::TCK_SizeAndLatency) >
8770 TargetTransformInfo::TCC_Basic)
8771 return false;
8772 UGEPIs.push_back(x: UGEPI);
8773 }
8774 if (UGEPIs.size() == 0)
8775 return false;
8776 // Check the materializing cost of (Uidx-Idx).
8777 for (GetElementPtrInst *UGEPI : UGEPIs) {
8778 ConstantInt *UGEPIIdx = cast<ConstantInt>(Val: UGEPI->getOperand(i_nocapture: 1));
8779 APInt NewIdx = UGEPIIdx->getValue() - GEPIIdx->getValue();
8780 InstructionCost ImmCost = TTI->getIntImmCost(
8781 Imm: NewIdx, Ty: GEPIIdx->getType(), CostKind: TargetTransformInfo::TCK_SizeAndLatency);
8782 if (ImmCost > TargetTransformInfo::TCC_Basic)
8783 return false;
8784 }
8785 // Now unmerge between GEPI and UGEPIs.
8786 for (GetElementPtrInst *UGEPI : UGEPIs) {
8787 UGEPI->setOperand(i_nocapture: 0, Val_nocapture: GEPI);
8788 ConstantInt *UGEPIIdx = cast<ConstantInt>(Val: UGEPI->getOperand(i_nocapture: 1));
8789 Constant *NewUGEPIIdx = ConstantInt::get(
8790 Ty: GEPIIdx->getType(), V: UGEPIIdx->getValue() - GEPIIdx->getValue());
8791 UGEPI->setOperand(i_nocapture: 1, Val_nocapture: NewUGEPIIdx);
8792 // If GEPI is not inbounds but UGEPI is inbounds, change UGEPI to not
8793 // inbounds to avoid UB.
8794 if (!GEPI->isInBounds()) {
8795 UGEPI->setIsInBounds(false);
8796 }
8797 }
8798 // After unmerging, verify that GEPIOp is actually only used in SrcBlock (not
8799 // alive on IndirectBr edges).
8800 assert(llvm::none_of(GEPIOp->users(),
8801 [&](User *Usr) {
8802 return cast<Instruction>(Usr)->getParent() != SrcBlock;
8803 }) &&
8804 "GEPIOp is used outside SrcBlock");
8805 return true;
8806}
8807
8808static bool optimizeBranch(CondBrInst *Branch, const TargetLowering &TLI,
8809 SmallPtrSet<BasicBlock *, 32> &FreshBBs,
8810 bool IsHugeFunc) {
8811 // Try and convert
8812 // %c = icmp ult %x, 8
8813 // br %c, bla, blb
8814 // %tc = lshr %x, 3
8815 // to
8816 // %tc = lshr %x, 3
8817 // %c = icmp eq %tc, 0
8818 // br %c, bla, blb
8819 // Creating the cmp to zero can be better for the backend, especially if the
8820 // lshr produces flags that can be used automatically.
8821 if (!TLI.preferZeroCompareBranch())
8822 return false;
8823
8824 ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: Branch->getCondition());
8825 if (!Cmp || !isa<ConstantInt>(Val: Cmp->getOperand(i_nocapture: 1)) || !Cmp->hasOneUse())
8826 return false;
8827
8828 Value *X = Cmp->getOperand(i_nocapture: 0);
8829 if (!X->hasUseList())
8830 return false;
8831
8832 APInt CmpC = cast<ConstantInt>(Val: Cmp->getOperand(i_nocapture: 1))->getValue();
8833
8834 for (auto *U : X->users()) {
8835 Instruction *UI = dyn_cast<Instruction>(Val: U);
8836 // A quick dominance check
8837 if (!UI ||
8838 (UI->getParent() != Branch->getParent() &&
8839 UI->getParent() != Branch->getSuccessor(i: 0) &&
8840 UI->getParent() != Branch->getSuccessor(i: 1)) ||
8841 (UI->getParent() != Branch->getParent() &&
8842 !UI->getParent()->getSinglePredecessor()))
8843 continue;
8844
8845 if (CmpC.isPowerOf2() && Cmp->getPredicate() == ICmpInst::ICMP_ULT &&
8846 match(V: UI, P: m_Shr(L: m_Specific(V: X), R: m_SpecificInt(V: CmpC.logBase2())))) {
8847 IRBuilder<> Builder(Branch);
8848 if (UI->getParent() != Branch->getParent())
8849 UI->moveBefore(InsertPos: Branch->getIterator());
8850 UI->dropPoisonGeneratingFlags();
8851 Value *NewCmp = Builder.CreateCmp(Pred: ICmpInst::ICMP_EQ, LHS: UI,
8852 RHS: ConstantInt::get(Ty: UI->getType(), V: 0));
8853 LLVM_DEBUG(dbgs() << "Converting " << *Cmp << "\n");
8854 LLVM_DEBUG(dbgs() << " to compare on zero: " << *NewCmp << "\n");
8855 replaceAllUsesWith(Old: Cmp, New: NewCmp, FreshBBs, IsHuge: IsHugeFunc);
8856 return true;
8857 }
8858 if (Cmp->isEquality() &&
8859 (match(V: UI, P: m_Add(L: m_Specific(V: X), R: m_SpecificInt(V: -CmpC))) ||
8860 match(V: UI, P: m_Sub(L: m_Specific(V: X), R: m_SpecificInt(V: CmpC))) ||
8861 match(V: UI, P: m_Xor(L: m_Specific(V: X), R: m_SpecificInt(V: CmpC))))) {
8862 IRBuilder<> Builder(Branch);
8863 if (UI->getParent() != Branch->getParent())
8864 UI->moveBefore(InsertPos: Branch->getIterator());
8865 UI->dropPoisonGeneratingFlags();
8866 Value *NewCmp = Builder.CreateCmp(Pred: Cmp->getPredicate(), LHS: UI,
8867 RHS: ConstantInt::get(Ty: UI->getType(), V: 0));
8868 LLVM_DEBUG(dbgs() << "Converting " << *Cmp << "\n");
8869 LLVM_DEBUG(dbgs() << " to compare on zero: " << *NewCmp << "\n");
8870 replaceAllUsesWith(Old: Cmp, New: NewCmp, FreshBBs, IsHuge: IsHugeFunc);
8871 return true;
8872 }
8873 }
8874 return false;
8875}
8876
8877bool CodeGenPrepare::optimizeInst(Instruction *I, ModifyDT &ModifiedDT) {
8878 bool AnyChange = false;
8879 AnyChange = fixupDbgVariableRecordsOnInst(I&: *I);
8880
8881 // Bail out if we inserted the instruction to prevent optimizations from
8882 // stepping on each other's toes.
8883 if (InsertedInsts.count(Ptr: I))
8884 return AnyChange;
8885
8886 // TODO: Move into the switch on opcode below here.
8887 if (PHINode *P = dyn_cast<PHINode>(Val: I)) {
8888 // It is possible for very late stage optimizations (such as SimplifyCFG)
8889 // to introduce PHI nodes too late to be cleaned up. If we detect such a
8890 // trivial PHI, go ahead and zap it here.
8891 if (Value *V = simplifyInstruction(I: P, Q: {*DL, TLInfo})) {
8892 LargeOffsetGEPMap.erase(Key: P);
8893 replaceAllUsesWith(Old: P, New: V, FreshBBs, IsHuge: IsHugeFunc);
8894 P->eraseFromParent();
8895 ++NumPHIsElim;
8896 return true;
8897 }
8898 return AnyChange;
8899 }
8900
8901 if (CastInst *CI = dyn_cast<CastInst>(Val: I)) {
8902 // If the source of the cast is a constant, then this should have
8903 // already been constant folded. The only reason NOT to constant fold
8904 // it is if something (e.g. LSR) was careful to place the constant
8905 // evaluation in a block other than then one that uses it (e.g. to hoist
8906 // the address of globals out of a loop). If this is the case, we don't
8907 // want to forward-subst the cast.
8908 if (isa<Constant>(Val: CI->getOperand(i_nocapture: 0)))
8909 return AnyChange;
8910
8911 if (OptimizeNoopCopyExpression(CI, TLI: *TLI, DL: *DL))
8912 return true;
8913
8914 if ((isa<UIToFPInst>(Val: I) || isa<SIToFPInst>(Val: I) || isa<FPToUIInst>(Val: I) ||
8915 isa<TruncInst>(Val: I)) &&
8916 TLI->optimizeExtendOrTruncateConversion(
8917 I, L: LI->getLoopFor(BB: I->getParent()), TTI: *TTI))
8918 return true;
8919
8920 if (isa<ZExtInst>(Val: I) || isa<SExtInst>(Val: I)) {
8921 /// Sink a zext or sext into its user blocks if the target type doesn't
8922 /// fit in one register
8923 if (TLI->getTypeAction(Context&: CI->getContext(),
8924 VT: TLI->getValueType(DL: *DL, Ty: CI->getType())) ==
8925 TargetLowering::TypeExpandInteger) {
8926 return SinkCast(CI);
8927 } else {
8928 if (TLI->optimizeExtendOrTruncateConversion(
8929 I, L: LI->getLoopFor(BB: I->getParent()), TTI: *TTI))
8930 return true;
8931
8932 bool MadeChange = optimizeExt(Inst&: I);
8933 return MadeChange | optimizeExtUses(I);
8934 }
8935 }
8936 return AnyChange;
8937 }
8938
8939 if (auto *Cmp = dyn_cast<CmpInst>(Val: I))
8940 if (optimizeCmp(Cmp, ModifiedDT))
8941 return true;
8942
8943 if (match(V: I, P: m_URem(L: m_Value(), R: m_Value())))
8944 if (optimizeURem(Rem: I))
8945 return true;
8946
8947 if (LoadInst *LI = dyn_cast<LoadInst>(Val: I)) {
8948 LI->setMetadata(KindID: LLVMContext::MD_invariant_group, Node: nullptr);
8949 bool Modified = optimizeLoadExt(Load: LI);
8950 unsigned AS = LI->getPointerAddressSpace();
8951 Modified |= optimizeMemoryInst(MemoryInst: I, Addr: I->getOperand(i: 0), AccessTy: LI->getType(), AddrSpace: AS);
8952 return Modified;
8953 }
8954
8955 if (StoreInst *SI = dyn_cast<StoreInst>(Val: I)) {
8956 if (splitMergedValStore(SI&: *SI, DL: *DL, TLI: *TLI))
8957 return true;
8958 SI->setMetadata(KindID: LLVMContext::MD_invariant_group, Node: nullptr);
8959 unsigned AS = SI->getPointerAddressSpace();
8960 return optimizeMemoryInst(MemoryInst: I, Addr: SI->getOperand(i_nocapture: 1),
8961 AccessTy: SI->getOperand(i_nocapture: 0)->getType(), AddrSpace: AS);
8962 }
8963
8964 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: I)) {
8965 unsigned AS = RMW->getPointerAddressSpace();
8966 return optimizeMemoryInst(MemoryInst: I, Addr: RMW->getPointerOperand(), AccessTy: RMW->getType(), AddrSpace: AS);
8967 }
8968
8969 if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Val: I)) {
8970 unsigned AS = CmpX->getPointerAddressSpace();
8971 return optimizeMemoryInst(MemoryInst: I, Addr: CmpX->getPointerOperand(),
8972 AccessTy: CmpX->getCompareOperand()->getType(), AddrSpace: AS);
8973 }
8974
8975 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Val: I);
8976
8977 if (BinOp && BinOp->getOpcode() == Instruction::And && EnableAndCmpSinking &&
8978 sinkAndCmp0Expression(AndI: BinOp, TLI: *TLI, InsertedInsts))
8979 return true;
8980
8981 // TODO: Move this into the switch on opcode - it handles shifts already.
8982 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
8983 BinOp->getOpcode() == Instruction::LShr)) {
8984 ConstantInt *CI = dyn_cast<ConstantInt>(Val: BinOp->getOperand(i_nocapture: 1));
8985 if (CI && TLI->hasExtractBitsInsn())
8986 if (OptimizeExtractBits(ShiftI: BinOp, CI, TLI: *TLI, DL: *DL))
8987 return true;
8988 }
8989
8990 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Val: I)) {
8991 if (GEPI->hasAllZeroIndices()) {
8992 /// The GEP operand must be a pointer, so must its result -> BitCast
8993 Instruction *NC = new BitCastInst(GEPI->getOperand(i_nocapture: 0), GEPI->getType(),
8994 GEPI->getName(), GEPI->getIterator());
8995 NC->setDebugLoc(GEPI->getDebugLoc());
8996 replaceAllUsesWith(Old: GEPI, New: NC, FreshBBs, IsHuge: IsHugeFunc);
8997 RecursivelyDeleteTriviallyDeadInstructions(
8998 V: GEPI, TLI: TLInfo, MSSAU: nullptr,
8999 AboutToDeleteCallback: [&](Value *V) { removeAllAssertingVHReferences(V); });
9000 ++NumGEPsElim;
9001 optimizeInst(I: NC, ModifiedDT);
9002 return true;
9003 }
9004 if (tryUnmergingGEPsAcrossIndirectBr(GEPI, TTI)) {
9005 return true;
9006 }
9007 }
9008
9009 if (FreezeInst *FI = dyn_cast<FreezeInst>(Val: I)) {
9010 // freeze(icmp a, const)) -> icmp (freeze a), const
9011 // This helps generate efficient conditional jumps.
9012 Instruction *CmpI = nullptr;
9013 if (ICmpInst *II = dyn_cast<ICmpInst>(Val: FI->getOperand(i_nocapture: 0)))
9014 CmpI = II;
9015 else if (FCmpInst *F = dyn_cast<FCmpInst>(Val: FI->getOperand(i_nocapture: 0)))
9016 CmpI = F->getFastMathFlags().none() ? F : nullptr;
9017
9018 if (CmpI && CmpI->hasOneUse()) {
9019 auto Op0 = CmpI->getOperand(i: 0), Op1 = CmpI->getOperand(i: 1);
9020 bool Const0 = isa<ConstantInt>(Val: Op0) || isa<ConstantFP>(Val: Op0) ||
9021 isa<ConstantPointerNull>(Val: Op0);
9022 bool Const1 = isa<ConstantInt>(Val: Op1) || isa<ConstantFP>(Val: Op1) ||
9023 isa<ConstantPointerNull>(Val: Op1);
9024 if (Const0 || Const1) {
9025 if (!Const0 || !Const1) {
9026 auto *F = new FreezeInst(Const0 ? Op1 : Op0, "", CmpI->getIterator());
9027 F->takeName(V: FI);
9028 CmpI->setOperand(i: Const0 ? 1 : 0, Val: F);
9029 }
9030 replaceAllUsesWith(Old: FI, New: CmpI, FreshBBs, IsHuge: IsHugeFunc);
9031 FI->eraseFromParent();
9032 return true;
9033 }
9034 }
9035 return AnyChange;
9036 }
9037
9038 if (tryToSinkFreeOperands(I))
9039 return true;
9040
9041 switch (I->getOpcode()) {
9042 case Instruction::Shl:
9043 case Instruction::LShr:
9044 case Instruction::AShr:
9045 return optimizeShiftInst(Shift: cast<BinaryOperator>(Val: I));
9046 case Instruction::Call:
9047 return optimizeCallInst(CI: cast<CallInst>(Val: I), ModifiedDT);
9048 case Instruction::Select:
9049 return optimizeSelectInst(SI: cast<SelectInst>(Val: I));
9050 case Instruction::ShuffleVector:
9051 return optimizeShuffleVectorInst(SVI: cast<ShuffleVectorInst>(Val: I));
9052 case Instruction::Switch:
9053 return optimizeSwitchInst(SI: cast<SwitchInst>(Val: I));
9054 case Instruction::ExtractElement:
9055 return optimizeExtractElementInst(Inst: cast<ExtractElementInst>(Val: I));
9056 case Instruction::CondBr:
9057 return optimizeBranch(Branch: cast<CondBrInst>(Val: I), TLI: *TLI, FreshBBs, IsHugeFunc);
9058 }
9059
9060 return AnyChange;
9061}
9062
9063/// Given an OR instruction, check to see if this is a bitreverse
9064/// idiom. If so, insert the new intrinsic and return true.
9065bool CodeGenPrepare::makeBitReverse(Instruction &I) {
9066 if (!I.getType()->isIntegerTy() ||
9067 !TLI->isOperationLegalOrCustom(Op: ISD::BITREVERSE,
9068 VT: TLI->getValueType(DL: *DL, Ty: I.getType(), AllowUnknown: true)))
9069 return false;
9070
9071 SmallVector<Instruction *, 4> Insts;
9072 if (!recognizeBSwapOrBitReverseIdiom(I: &I, MatchBSwaps: false, MatchBitReversals: true, InsertedInsts&: Insts))
9073 return false;
9074 Instruction *LastInst = Insts.back();
9075 replaceAllUsesWith(Old: &I, New: LastInst, FreshBBs, IsHuge: IsHugeFunc);
9076 RecursivelyDeleteTriviallyDeadInstructions(
9077 V: &I, TLI: TLInfo, MSSAU: nullptr,
9078 AboutToDeleteCallback: [&](Value *V) { removeAllAssertingVHReferences(V); });
9079 return true;
9080}
9081
9082// In this pass we look for GEP and cast instructions that are used
9083// across basic blocks and rewrite them to improve basic-block-at-a-time
9084// selection.
9085bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, ModifyDT &ModifiedDT) {
9086 SunkAddrs.clear();
9087 bool MadeChange = false;
9088
9089 do {
9090 CurInstIterator = BB.begin();
9091 ModifiedDT = ModifyDT::NotModifyDT;
9092 while (CurInstIterator != BB.end()) {
9093 MadeChange |= optimizeInst(I: &*CurInstIterator++, ModifiedDT);
9094 if (ModifiedDT != ModifyDT::NotModifyDT) {
9095 // For huge function we tend to quickly go though the inner optmization
9096 // opportunities in the BB. So we go back to the BB head to re-optimize
9097 // each instruction instead of go back to the function head.
9098 if (IsHugeFunc) {
9099 DT.reset();
9100 getDT(F&: *BB.getParent());
9101 break;
9102 } else {
9103 return true;
9104 }
9105 }
9106 }
9107 } while (ModifiedDT == ModifyDT::ModifyInstDT);
9108
9109 bool MadeBitReverse = true;
9110 while (MadeBitReverse) {
9111 MadeBitReverse = false;
9112 for (auto &I : reverse(C&: BB)) {
9113 if (makeBitReverse(I)) {
9114 MadeBitReverse = MadeChange = true;
9115 break;
9116 }
9117 }
9118 }
9119 MadeChange |= dupRetToEnableTailCallOpts(BB: &BB, ModifiedDT);
9120
9121 return MadeChange;
9122}
9123
9124bool CodeGenPrepare::fixupDbgVariableRecordsOnInst(Instruction &I) {
9125 bool AnyChange = false;
9126 for (DbgVariableRecord &DVR : filterDbgVars(R: I.getDbgRecordRange()))
9127 AnyChange |= fixupDbgVariableRecord(I&: DVR);
9128 return AnyChange;
9129}
9130
9131// FIXME: should updating debug-info really cause the "changed" flag to fire,
9132// which can cause a function to be reprocessed?
9133bool CodeGenPrepare::fixupDbgVariableRecord(DbgVariableRecord &DVR) {
9134 if (DVR.Type != DbgVariableRecord::LocationType::Value &&
9135 DVR.Type != DbgVariableRecord::LocationType::Assign)
9136 return false;
9137
9138 // Does this DbgVariableRecord refer to a sunk address calculation?
9139 bool AnyChange = false;
9140 SmallDenseSet<Value *> LocationOps(DVR.location_ops().begin(),
9141 DVR.location_ops().end());
9142 for (Value *Location : LocationOps) {
9143 WeakTrackingVH SunkAddrVH = SunkAddrs[Location];
9144 Value *SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
9145 if (SunkAddr) {
9146 // Point dbg.value at locally computed address, which should give the best
9147 // opportunity to be accurately lowered. This update may change the type
9148 // of pointer being referred to; however this makes no difference to
9149 // debugging information, and we can't generate bitcasts that may affect
9150 // codegen.
9151 DVR.replaceVariableLocationOp(OldValue: Location, NewValue: SunkAddr);
9152 AnyChange = true;
9153 }
9154 }
9155 return AnyChange;
9156}
9157
9158static void DbgInserterHelper(DbgVariableRecord *DVR, BasicBlock::iterator VI) {
9159 DVR->removeFromParent();
9160 BasicBlock *VIBB = VI->getParent();
9161 if (isa<PHINode>(Val: VI))
9162 VIBB->insertDbgRecordBefore(DR: DVR, Here: VIBB->getFirstInsertionPt());
9163 else
9164 VIBB->insertDbgRecordAfter(DR: DVR, I: &*VI);
9165}
9166
9167// A llvm.dbg.value may be using a value before its definition, due to
9168// optimizations in this pass and others. Scan for such dbg.values, and rescue
9169// them by moving the dbg.value to immediately after the value definition.
9170// FIXME: Ideally this should never be necessary, and this has the potential
9171// to re-order dbg.value intrinsics.
9172bool CodeGenPrepare::placeDbgValues(Function &F) {
9173 bool MadeChange = false;
9174 DominatorTree DT(F);
9175
9176 auto DbgProcessor = [&](auto *DbgItem, Instruction *Position) {
9177 SmallVector<Instruction *, 4> VIs;
9178 for (Value *V : DbgItem->location_ops())
9179 if (Instruction *VI = dyn_cast_or_null<Instruction>(Val: V))
9180 VIs.push_back(Elt: VI);
9181
9182 // This item may depend on multiple instructions, complicating any
9183 // potential sink. This block takes the defensive approach, opting to
9184 // "undef" the item if it has more than one instruction and any of them do
9185 // not dominate iem.
9186 for (Instruction *VI : VIs) {
9187 if (VI->isTerminator())
9188 continue;
9189
9190 // If VI is a phi in a block with an EHPad terminator, we can't insert
9191 // after it.
9192 if (isa<PHINode>(Val: VI) && VI->getParent()->getTerminator()->isEHPad())
9193 continue;
9194
9195 // If the defining instruction dominates the dbg.value, we do not need
9196 // to move the dbg.value.
9197 if (DT.dominates(Def: VI, User: Position))
9198 continue;
9199
9200 // If we depend on multiple instructions and any of them doesn't
9201 // dominate this DVI, we probably can't salvage it: moving it to
9202 // after any of the instructions could cause us to lose the others.
9203 if (VIs.size() > 1) {
9204 LLVM_DEBUG(
9205 dbgs()
9206 << "Unable to find valid location for Debug Value, undefing:\n"
9207 << *DbgItem);
9208 DbgItem->setKillLocation();
9209 break;
9210 }
9211
9212 LLVM_DEBUG(dbgs() << "Moving Debug Value before :\n"
9213 << *DbgItem << ' ' << *VI);
9214 DbgInserterHelper(DbgItem, VI->getIterator());
9215 MadeChange = true;
9216 ++NumDbgValueMoved;
9217 }
9218 };
9219
9220 for (BasicBlock &BB : F) {
9221 for (Instruction &Insn : llvm::make_early_inc_range(Range&: BB)) {
9222 // Process any DbgVariableRecord records attached to this
9223 // instruction.
9224 for (DbgVariableRecord &DVR : llvm::make_early_inc_range(
9225 Range: filterDbgVars(R: Insn.getDbgRecordRange()))) {
9226 if (DVR.Type != DbgVariableRecord::LocationType::Value)
9227 continue;
9228 DbgProcessor(&DVR, &Insn);
9229 }
9230 }
9231 }
9232
9233 return MadeChange;
9234}
9235
9236// Group scattered pseudo probes in a block to favor SelectionDAG. Scattered
9237// probes can be chained dependencies of other regular DAG nodes and block DAG
9238// combine optimizations.
9239bool CodeGenPrepare::placePseudoProbes(Function &F) {
9240 bool MadeChange = false;
9241 for (auto &Block : F) {
9242 // Move the rest probes to the beginning of the block.
9243 auto FirstInst = Block.getFirstInsertionPt();
9244 while (FirstInst != Block.end() && FirstInst->isDebugOrPseudoInst())
9245 ++FirstInst;
9246 BasicBlock::iterator I(FirstInst);
9247 I++;
9248 while (I != Block.end()) {
9249 if (auto *II = dyn_cast<PseudoProbeInst>(Val: I++)) {
9250 II->moveBefore(InsertPos: FirstInst);
9251 MadeChange = true;
9252 }
9253 }
9254 }
9255 return MadeChange;
9256}
9257
9258/// Scale down both weights to fit into uint32_t.
9259static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
9260 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
9261 uint32_t Scale = (NewMax / std::numeric_limits<uint32_t>::max()) + 1;
9262 NewTrue = NewTrue / Scale;
9263 NewFalse = NewFalse / Scale;
9264}
9265
9266/// Some targets prefer to split a conditional branch like:
9267/// \code
9268/// %0 = icmp ne i32 %a, 0
9269/// %1 = icmp ne i32 %b, 0
9270/// %or.cond = or i1 %0, %1
9271/// br i1 %or.cond, label %TrueBB, label %FalseBB
9272/// \endcode
9273/// into multiple branch instructions like:
9274/// \code
9275/// bb1:
9276/// %0 = icmp ne i32 %a, 0
9277/// br i1 %0, label %TrueBB, label %bb2
9278/// bb2:
9279/// %1 = icmp ne i32 %b, 0
9280/// br i1 %1, label %TrueBB, label %FalseBB
9281/// \endcode
9282/// This usually allows instruction selection to do even further optimizations
9283/// and combine the compare with the branch instruction. Currently this is
9284/// applied for targets which have "cheap" jump instructions.
9285///
9286/// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
9287///
9288bool CodeGenPrepare::splitBranchCondition(Function &F, ModifyDT &ModifiedDT) {
9289 if (!TM->Options.EnableFastISel || TLI->isJumpExpensive())
9290 return false;
9291
9292 bool MadeChange = false;
9293 for (auto &BB : F) {
9294 // Does this BB end with the following?
9295 // %cond1 = icmp|fcmp|binary instruction ...
9296 // %cond2 = icmp|fcmp|binary instruction ...
9297 // %cond.or = or|and i1 %cond1, cond2
9298 // br i1 %cond.or label %dest1, label %dest2"
9299 Instruction *LogicOp;
9300 BasicBlock *TBB, *FBB;
9301 if (!match(V: BB.getTerminator(),
9302 P: m_Br(C: m_OneUse(SubPattern: m_Instruction(I&: LogicOp)), T&: TBB, F&: FBB)))
9303 continue;
9304
9305 auto *Br1 = cast<CondBrInst>(Val: BB.getTerminator());
9306 if (Br1->getMetadata(KindID: LLVMContext::MD_unpredictable))
9307 continue;
9308
9309 // The merging of mostly empty BB can cause a degenerate branch.
9310 if (TBB == FBB)
9311 continue;
9312
9313 unsigned Opc;
9314 Value *Cond1, *Cond2;
9315 if (match(V: LogicOp,
9316 P: m_LogicalAnd(L: m_OneUse(SubPattern: m_Value(V&: Cond1)), R: m_OneUse(SubPattern: m_Value(V&: Cond2)))))
9317 Opc = Instruction::And;
9318 else if (match(V: LogicOp, P: m_LogicalOr(L: m_OneUse(SubPattern: m_Value(V&: Cond1)),
9319 R: m_OneUse(SubPattern: m_Value(V&: Cond2)))))
9320 Opc = Instruction::Or;
9321 else
9322 continue;
9323
9324 auto IsGoodCond = [](Value *Cond) {
9325 return match(
9326 V: Cond,
9327 P: m_CombineOr(L: m_Cmp(), R: m_CombineOr(L: m_LogicalAnd(L: m_Value(), R: m_Value()),
9328 R: m_LogicalOr(L: m_Value(), R: m_Value()))));
9329 };
9330 if (!IsGoodCond(Cond1) || !IsGoodCond(Cond2))
9331 continue;
9332
9333 LLVM_DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
9334
9335 // Create a new BB.
9336 auto *TmpBB =
9337 BasicBlock::Create(Context&: BB.getContext(), Name: BB.getName() + ".cond.split",
9338 Parent: BB.getParent(), InsertBefore: BB.getNextNode());
9339 if (IsHugeFunc)
9340 FreshBBs.insert(Ptr: TmpBB);
9341
9342 // Update original basic block by using the first condition directly by the
9343 // branch instruction and removing the no longer needed and/or instruction.
9344 Br1->setCondition(Cond1);
9345 LogicOp->eraseFromParent();
9346
9347 // Depending on the condition we have to either replace the true or the
9348 // false successor of the original branch instruction.
9349 if (Opc == Instruction::And)
9350 Br1->setSuccessor(idx: 0, NewSucc: TmpBB);
9351 else
9352 Br1->setSuccessor(idx: 1, NewSucc: TmpBB);
9353
9354 // Fill in the new basic block.
9355 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond: Cond2, True: TBB, False: FBB);
9356 if (auto *I = dyn_cast<Instruction>(Val: Cond2)) {
9357 I->removeFromParent();
9358 I->insertBefore(InsertPos: Br2->getIterator());
9359 }
9360
9361 // Update PHI nodes in both successors. The original BB needs to be
9362 // replaced in one successor's PHI nodes, because the branch comes now from
9363 // the newly generated BB (NewBB). In the other successor we need to add one
9364 // incoming edge to the PHI nodes, because both branch instructions target
9365 // now the same successor. Depending on the original branch condition
9366 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
9367 // we perform the correct update for the PHI nodes.
9368 // This doesn't change the successor order of the just created branch
9369 // instruction (or any other instruction).
9370 if (Opc == Instruction::Or)
9371 std::swap(a&: TBB, b&: FBB);
9372
9373 // Replace the old BB with the new BB.
9374 TBB->replacePhiUsesWith(Old: &BB, New: TmpBB);
9375
9376 // Add another incoming edge from the new BB.
9377 for (PHINode &PN : FBB->phis()) {
9378 auto *Val = PN.getIncomingValueForBlock(BB: &BB);
9379 PN.addIncoming(V: Val, BB: TmpBB);
9380 }
9381
9382 // Update the branch weights (from SelectionDAGBuilder::
9383 // FindMergedConditions).
9384 if (Opc == Instruction::Or) {
9385 // Codegen X | Y as:
9386 // BB1:
9387 // jmp_if_X TBB
9388 // jmp TmpBB
9389 // TmpBB:
9390 // jmp_if_Y TBB
9391 // jmp FBB
9392 //
9393
9394 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
9395 // The requirement is that
9396 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
9397 // = TrueProb for original BB.
9398 // Assuming the original weights are A and B, one choice is to set BB1's
9399 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
9400 // assumes that
9401 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
9402 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
9403 // TmpBB, but the math is more complicated.
9404 uint64_t TrueWeight, FalseWeight;
9405 if (extractBranchWeights(I: *Br1, TrueVal&: TrueWeight, FalseVal&: FalseWeight)) {
9406 uint64_t NewTrueWeight = TrueWeight;
9407 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
9408 scaleWeights(NewTrue&: NewTrueWeight, NewFalse&: NewFalseWeight);
9409 Br1->setMetadata(KindID: LLVMContext::MD_prof,
9410 Node: MDBuilder(Br1->getContext())
9411 .createBranchWeights(TrueWeight, FalseWeight,
9412 IsExpected: hasBranchWeightOrigin(I: *Br1)));
9413
9414 NewTrueWeight = TrueWeight;
9415 NewFalseWeight = 2 * FalseWeight;
9416 scaleWeights(NewTrue&: NewTrueWeight, NewFalse&: NewFalseWeight);
9417 Br2->setMetadata(KindID: LLVMContext::MD_prof,
9418 Node: MDBuilder(Br2->getContext())
9419 .createBranchWeights(TrueWeight, FalseWeight));
9420 }
9421 } else {
9422 // Codegen X & Y as:
9423 // BB1:
9424 // jmp_if_X TmpBB
9425 // jmp FBB
9426 // TmpBB:
9427 // jmp_if_Y TBB
9428 // jmp FBB
9429 //
9430 // This requires creation of TmpBB after CurBB.
9431
9432 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
9433 // The requirement is that
9434 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
9435 // = FalseProb for original BB.
9436 // Assuming the original weights are A and B, one choice is to set BB1's
9437 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
9438 // assumes that
9439 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
9440 uint64_t TrueWeight, FalseWeight;
9441 if (extractBranchWeights(I: *Br1, TrueVal&: TrueWeight, FalseVal&: FalseWeight)) {
9442 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
9443 uint64_t NewFalseWeight = FalseWeight;
9444 scaleWeights(NewTrue&: NewTrueWeight, NewFalse&: NewFalseWeight);
9445 Br1->setMetadata(KindID: LLVMContext::MD_prof,
9446 Node: MDBuilder(Br1->getContext())
9447 .createBranchWeights(TrueWeight, FalseWeight));
9448
9449 NewTrueWeight = 2 * TrueWeight;
9450 NewFalseWeight = FalseWeight;
9451 scaleWeights(NewTrue&: NewTrueWeight, NewFalse&: NewFalseWeight);
9452 Br2->setMetadata(KindID: LLVMContext::MD_prof,
9453 Node: MDBuilder(Br2->getContext())
9454 .createBranchWeights(TrueWeight, FalseWeight));
9455 }
9456 }
9457
9458 ModifiedDT = ModifyDT::ModifyBBDT;
9459 MadeChange = true;
9460
9461 LLVM_DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
9462 TmpBB->dump());
9463 }
9464 return MadeChange;
9465}
9466