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