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