xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/CodeGenPrepare.cpp (revision 924226fba12cc9a228c73b956e1b7fa24c60b055)
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/ADT/APInt.h"
16 #include "llvm/ADT/ArrayRef.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/MapVector.h"
19 #include "llvm/ADT/PointerIntPair.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/Analysis/BlockFrequencyInfo.h"
25 #include "llvm/Analysis/BranchProbabilityInfo.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Analysis/InstructionSimplify.h"
28 #include "llvm/Analysis/LoopInfo.h"
29 #include "llvm/Analysis/MemoryBuiltins.h"
30 #include "llvm/Analysis/ProfileSummaryInfo.h"
31 #include "llvm/Analysis/TargetLibraryInfo.h"
32 #include "llvm/Analysis/TargetTransformInfo.h"
33 #include "llvm/Analysis/ValueTracking.h"
34 #include "llvm/Analysis/VectorUtils.h"
35 #include "llvm/CodeGen/Analysis.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/Config/llvm-config.h"
43 #include "llvm/IR/Argument.h"
44 #include "llvm/IR/Attributes.h"
45 #include "llvm/IR/BasicBlock.h"
46 #include "llvm/IR/Constant.h"
47 #include "llvm/IR/Constants.h"
48 #include "llvm/IR/DataLayout.h"
49 #include "llvm/IR/DebugInfo.h"
50 #include "llvm/IR/DerivedTypes.h"
51 #include "llvm/IR/Dominators.h"
52 #include "llvm/IR/Function.h"
53 #include "llvm/IR/GetElementPtrTypeIterator.h"
54 #include "llvm/IR/GlobalValue.h"
55 #include "llvm/IR/GlobalVariable.h"
56 #include "llvm/IR/IRBuilder.h"
57 #include "llvm/IR/InlineAsm.h"
58 #include "llvm/IR/InstrTypes.h"
59 #include "llvm/IR/Instruction.h"
60 #include "llvm/IR/Instructions.h"
61 #include "llvm/IR/IntrinsicInst.h"
62 #include "llvm/IR/Intrinsics.h"
63 #include "llvm/IR/IntrinsicsAArch64.h"
64 #include "llvm/IR/LLVMContext.h"
65 #include "llvm/IR/MDBuilder.h"
66 #include "llvm/IR/Module.h"
67 #include "llvm/IR/Operator.h"
68 #include "llvm/IR/PatternMatch.h"
69 #include "llvm/IR/Statepoint.h"
70 #include "llvm/IR/Type.h"
71 #include "llvm/IR/Use.h"
72 #include "llvm/IR/User.h"
73 #include "llvm/IR/Value.h"
74 #include "llvm/IR/ValueHandle.h"
75 #include "llvm/IR/ValueMap.h"
76 #include "llvm/InitializePasses.h"
77 #include "llvm/Pass.h"
78 #include "llvm/Support/BlockFrequency.h"
79 #include "llvm/Support/BranchProbability.h"
80 #include "llvm/Support/Casting.h"
81 #include "llvm/Support/CommandLine.h"
82 #include "llvm/Support/Compiler.h"
83 #include "llvm/Support/Debug.h"
84 #include "llvm/Support/ErrorHandling.h"
85 #include "llvm/Support/MachineValueType.h"
86 #include "llvm/Support/MathExtras.h"
87 #include "llvm/Support/raw_ostream.h"
88 #include "llvm/Target/TargetMachine.h"
89 #include "llvm/Target/TargetOptions.h"
90 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
91 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
92 #include "llvm/Transforms/Utils/Local.h"
93 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
94 #include "llvm/Transforms/Utils/SizeOpts.h"
95 #include <algorithm>
96 #include <cassert>
97 #include <cstdint>
98 #include <iterator>
99 #include <limits>
100 #include <memory>
101 #include <utility>
102 #include <vector>
103 
104 using namespace llvm;
105 using namespace llvm::PatternMatch;
106 
107 #define DEBUG_TYPE "codegenprepare"
108 
109 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
110 STATISTIC(NumPHIsElim,   "Number of trivial PHIs eliminated");
111 STATISTIC(NumGEPsElim,   "Number of GEPs converted to casts");
112 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
113                       "sunken Cmps");
114 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
115                        "of sunken Casts");
116 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
117                           "computations were sunk");
118 STATISTIC(NumMemoryInstsPhiCreated,
119           "Number of phis created when address "
120           "computations were sunk to memory instructions");
121 STATISTIC(NumMemoryInstsSelectCreated,
122           "Number of select created when address "
123           "computations were sunk to memory instructions");
124 STATISTIC(NumExtsMoved,  "Number of [s|z]ext instructions combined with loads");
125 STATISTIC(NumExtUses,    "Number of uses of [s|z]ext instructions optimized");
126 STATISTIC(NumAndsAdded,
127           "Number of and mask instructions added to form ext loads");
128 STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
129 STATISTIC(NumRetsDup,    "Number of return instructions duplicated");
130 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
131 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
132 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
133 
134 static cl::opt<bool> DisableBranchOpts(
135   "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
136   cl::desc("Disable branch optimizations in CodeGenPrepare"));
137 
138 static cl::opt<bool>
139     DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
140                   cl::desc("Disable GC optimizations in CodeGenPrepare"));
141 
142 static cl::opt<bool> DisableSelectToBranch(
143   "disable-cgp-select2branch", cl::Hidden, cl::init(false),
144   cl::desc("Disable select to branch conversion."));
145 
146 static cl::opt<bool> AddrSinkUsingGEPs(
147   "addr-sink-using-gep", cl::Hidden, cl::init(true),
148   cl::desc("Address sinking in CGP using GEPs."));
149 
150 static cl::opt<bool> EnableAndCmpSinking(
151    "enable-andcmp-sinking", cl::Hidden, cl::init(true),
152    cl::desc("Enable sinkinig and/cmp into branches."));
153 
154 static cl::opt<bool> DisableStoreExtract(
155     "disable-cgp-store-extract", cl::Hidden, cl::init(false),
156     cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
157 
158 static cl::opt<bool> StressStoreExtract(
159     "stress-cgp-store-extract", cl::Hidden, cl::init(false),
160     cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
161 
162 static cl::opt<bool> DisableExtLdPromotion(
163     "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
164     cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
165              "CodeGenPrepare"));
166 
167 static cl::opt<bool> StressExtLdPromotion(
168     "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
169     cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
170              "optimization in CodeGenPrepare"));
171 
172 static cl::opt<bool> DisablePreheaderProtect(
173     "disable-preheader-prot", cl::Hidden, cl::init(false),
174     cl::desc("Disable protection against removing loop preheaders"));
175 
176 static cl::opt<bool> ProfileGuidedSectionPrefix(
177     "profile-guided-section-prefix", cl::Hidden, cl::init(true), cl::ZeroOrMore,
178     cl::desc("Use profile info to add section prefix for hot/cold functions"));
179 
180 static cl::opt<bool> ProfileUnknownInSpecialSection(
181     "profile-unknown-in-special-section", cl::Hidden, cl::init(false),
182     cl::ZeroOrMore,
183     cl::desc("In profiling mode like sampleFDO, if a function doesn't have "
184              "profile, we cannot tell the function is cold for sure because "
185              "it may be a function newly added without ever being sampled. "
186              "With the flag enabled, compiler can put such profile unknown "
187              "functions into a special section, so runtime system can choose "
188              "to handle it in a different way than .text section, to save "
189              "RAM for example. "));
190 
191 static cl::opt<unsigned> FreqRatioToSkipMerge(
192     "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2),
193     cl::desc("Skip merging empty blocks if (frequency of empty block) / "
194              "(frequency of destination block) is greater than this ratio"));
195 
196 static cl::opt<bool> ForceSplitStore(
197     "force-split-store", cl::Hidden, cl::init(false),
198     cl::desc("Force store splitting no matter what the target query says."));
199 
200 static cl::opt<bool>
201 EnableTypePromotionMerge("cgp-type-promotion-merge", cl::Hidden,
202     cl::desc("Enable merging of redundant sexts when one is dominating"
203     " the other."), cl::init(true));
204 
205 static cl::opt<bool> DisableComplexAddrModes(
206     "disable-complex-addr-modes", cl::Hidden, cl::init(false),
207     cl::desc("Disables combining addressing modes with different parts "
208              "in optimizeMemoryInst."));
209 
210 static cl::opt<bool>
211 AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(false),
212                 cl::desc("Allow creation of Phis in Address sinking."));
213 
214 static cl::opt<bool>
215 AddrSinkNewSelects("addr-sink-new-select", cl::Hidden, cl::init(true),
216                    cl::desc("Allow creation of selects in Address sinking."));
217 
218 static cl::opt<bool> AddrSinkCombineBaseReg(
219     "addr-sink-combine-base-reg", cl::Hidden, cl::init(true),
220     cl::desc("Allow combining of BaseReg field in Address sinking."));
221 
222 static cl::opt<bool> AddrSinkCombineBaseGV(
223     "addr-sink-combine-base-gv", cl::Hidden, cl::init(true),
224     cl::desc("Allow combining of BaseGV field in Address sinking."));
225 
226 static cl::opt<bool> AddrSinkCombineBaseOffs(
227     "addr-sink-combine-base-offs", cl::Hidden, cl::init(true),
228     cl::desc("Allow combining of BaseOffs field in Address sinking."));
229 
230 static cl::opt<bool> AddrSinkCombineScaledReg(
231     "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(true),
232     cl::desc("Allow combining of ScaledReg field in Address sinking."));
233 
234 static cl::opt<bool>
235     EnableGEPOffsetSplit("cgp-split-large-offset-gep", cl::Hidden,
236                          cl::init(true),
237                          cl::desc("Enable splitting large offset of GEP."));
238 
239 static cl::opt<bool> EnableICMP_EQToICMP_ST(
240     "cgp-icmp-eq2icmp-st", cl::Hidden, cl::init(false),
241     cl::desc("Enable ICMP_EQ to ICMP_S(L|G)T conversion."));
242 
243 static cl::opt<bool>
244     VerifyBFIUpdates("cgp-verify-bfi-updates", cl::Hidden, cl::init(false),
245                      cl::desc("Enable BFI update verification for "
246                               "CodeGenPrepare."));
247 
248 static cl::opt<bool> OptimizePhiTypes(
249     "cgp-optimize-phi-types", cl::Hidden, cl::init(false),
250     cl::desc("Enable converting phi types in CodeGenPrepare"));
251 
252 namespace {
253 
254 enum ExtType {
255   ZeroExtension,   // Zero extension has been seen.
256   SignExtension,   // Sign extension has been seen.
257   BothExtension    // This extension type is used if we saw sext after
258                    // ZeroExtension had been set, or if we saw zext after
259                    // SignExtension had been set. It makes the type
260                    // information of a promoted instruction invalid.
261 };
262 
263 using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
264 using TypeIsSExt = PointerIntPair<Type *, 2, ExtType>;
265 using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
266 using SExts = SmallVector<Instruction *, 16>;
267 using ValueToSExts = DenseMap<Value *, SExts>;
268 
269 class TypePromotionTransaction;
270 
271   class CodeGenPrepare : public FunctionPass {
272     const TargetMachine *TM = nullptr;
273     const TargetSubtargetInfo *SubtargetInfo;
274     const TargetLowering *TLI = nullptr;
275     const TargetRegisterInfo *TRI;
276     const TargetTransformInfo *TTI = nullptr;
277     const TargetLibraryInfo *TLInfo;
278     const LoopInfo *LI;
279     std::unique_ptr<BlockFrequencyInfo> BFI;
280     std::unique_ptr<BranchProbabilityInfo> BPI;
281     ProfileSummaryInfo *PSI;
282 
283     /// As we scan instructions optimizing them, this is the next instruction
284     /// to optimize. Transforms that can invalidate this should update it.
285     BasicBlock::iterator CurInstIterator;
286 
287     /// Keeps track of non-local addresses that have been sunk into a block.
288     /// This allows us to avoid inserting duplicate code for blocks with
289     /// multiple load/stores of the same address. The usage of WeakTrackingVH
290     /// enables SunkAddrs to be treated as a cache whose entries can be
291     /// invalidated if a sunken address computation has been erased.
292     ValueMap<Value*, WeakTrackingVH> SunkAddrs;
293 
294     /// Keeps track of all instructions inserted for the current function.
295     SetOfInstrs InsertedInsts;
296 
297     /// Keeps track of the type of the related instruction before their
298     /// promotion for the current function.
299     InstrToOrigTy PromotedInsts;
300 
301     /// Keep track of instructions removed during promotion.
302     SetOfInstrs RemovedInsts;
303 
304     /// Keep track of sext chains based on their initial value.
305     DenseMap<Value *, Instruction *> SeenChainsForSExt;
306 
307     /// Keep track of GEPs accessing the same data structures such as structs or
308     /// arrays that are candidates to be split later because of their large
309     /// size.
310     MapVector<
311         AssertingVH<Value>,
312         SmallVector<std::pair<AssertingVH<GetElementPtrInst>, int64_t>, 32>>
313         LargeOffsetGEPMap;
314 
315     /// Keep track of new GEP base after splitting the GEPs having large offset.
316     SmallSet<AssertingVH<Value>, 2> NewGEPBases;
317 
318     /// Map serial numbers to Large offset GEPs.
319     DenseMap<AssertingVH<GetElementPtrInst>, int> LargeOffsetGEPID;
320 
321     /// Keep track of SExt promoted.
322     ValueToSExts ValToSExtendedUses;
323 
324     /// True if the function has the OptSize attribute.
325     bool OptSize;
326 
327     /// DataLayout for the Function being processed.
328     const DataLayout *DL = nullptr;
329 
330     /// Building the dominator tree can be expensive, so we only build it
331     /// lazily and update it when required.
332     std::unique_ptr<DominatorTree> DT;
333 
334   public:
335     static char ID; // Pass identification, replacement for typeid
336 
337     CodeGenPrepare() : FunctionPass(ID) {
338       initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
339     }
340 
341     bool runOnFunction(Function &F) override;
342 
343     StringRef getPassName() const override { return "CodeGen Prepare"; }
344 
345     void getAnalysisUsage(AnalysisUsage &AU) const override {
346       // FIXME: When we can selectively preserve passes, preserve the domtree.
347       AU.addRequired<ProfileSummaryInfoWrapperPass>();
348       AU.addRequired<TargetLibraryInfoWrapperPass>();
349       AU.addRequired<TargetPassConfig>();
350       AU.addRequired<TargetTransformInfoWrapperPass>();
351       AU.addRequired<LoopInfoWrapperPass>();
352     }
353 
354   private:
355     template <typename F>
356     void resetIteratorIfInvalidatedWhileCalling(BasicBlock *BB, F f) {
357       // Substituting can cause recursive simplifications, which can invalidate
358       // our iterator.  Use a WeakTrackingVH to hold onto it in case this
359       // happens.
360       Value *CurValue = &*CurInstIterator;
361       WeakTrackingVH IterHandle(CurValue);
362 
363       f();
364 
365       // If the iterator instruction was recursively deleted, start over at the
366       // start of the block.
367       if (IterHandle != CurValue) {
368         CurInstIterator = BB->begin();
369         SunkAddrs.clear();
370       }
371     }
372 
373     // Get the DominatorTree, building if necessary.
374     DominatorTree &getDT(Function &F) {
375       if (!DT)
376         DT = std::make_unique<DominatorTree>(F);
377       return *DT;
378     }
379 
380     void removeAllAssertingVHReferences(Value *V);
381     bool eliminateAssumptions(Function &F);
382     bool eliminateFallThrough(Function &F);
383     bool eliminateMostlyEmptyBlocks(Function &F);
384     BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
385     bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
386     void eliminateMostlyEmptyBlock(BasicBlock *BB);
387     bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
388                                        bool isPreheader);
389     bool makeBitReverse(Instruction &I);
390     bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
391     bool optimizeInst(Instruction *I, bool &ModifiedDT);
392     bool optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
393                             Type *AccessTy, unsigned AddrSpace);
394     bool optimizeGatherScatterInst(Instruction *MemoryInst, Value *Ptr);
395     bool optimizeInlineAsmInst(CallInst *CS);
396     bool optimizeCallInst(CallInst *CI, bool &ModifiedDT);
397     bool optimizeExt(Instruction *&I);
398     bool optimizeExtUses(Instruction *I);
399     bool optimizeLoadExt(LoadInst *Load);
400     bool optimizeShiftInst(BinaryOperator *BO);
401     bool optimizeFunnelShift(IntrinsicInst *Fsh);
402     bool optimizeSelectInst(SelectInst *SI);
403     bool optimizeShuffleVectorInst(ShuffleVectorInst *SVI);
404     bool optimizeSwitchInst(SwitchInst *SI);
405     bool optimizeExtractElementInst(Instruction *Inst);
406     bool dupRetToEnableTailCallOpts(BasicBlock *BB, bool &ModifiedDT);
407     bool fixupDbgValue(Instruction *I);
408     bool placeDbgValues(Function &F);
409     bool placePseudoProbes(Function &F);
410     bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
411                       LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
412     bool tryToPromoteExts(TypePromotionTransaction &TPT,
413                           const SmallVectorImpl<Instruction *> &Exts,
414                           SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
415                           unsigned CreatedInstsCost = 0);
416     bool mergeSExts(Function &F);
417     bool splitLargeGEPOffsets();
418     bool optimizePhiType(PHINode *Inst, SmallPtrSetImpl<PHINode *> &Visited,
419                          SmallPtrSetImpl<Instruction *> &DeletedInstrs);
420     bool optimizePhiTypes(Function &F);
421     bool performAddressTypePromotion(
422         Instruction *&Inst,
423         bool AllowPromotionWithoutCommonHeader,
424         bool HasPromoted, TypePromotionTransaction &TPT,
425         SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
426     bool splitBranchCondition(Function &F, bool &ModifiedDT);
427     bool simplifyOffsetableRelocate(GCStatepointInst &I);
428 
429     bool tryToSinkFreeOperands(Instruction *I);
430     bool replaceMathCmpWithIntrinsic(BinaryOperator *BO, Value *Arg0,
431                                      Value *Arg1, CmpInst *Cmp,
432                                      Intrinsic::ID IID);
433     bool optimizeCmp(CmpInst *Cmp, bool &ModifiedDT);
434     bool combineToUSubWithOverflow(CmpInst *Cmp, bool &ModifiedDT);
435     bool combineToUAddWithOverflow(CmpInst *Cmp, bool &ModifiedDT);
436     void verifyBFIUpdates(Function &F);
437   };
438 
439 } // end anonymous namespace
440 
441 char CodeGenPrepare::ID = 0;
442 
443 INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,
444                       "Optimize for code generation", false, false)
445 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
446 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
447 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
448 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
449 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
450 INITIALIZE_PASS_END(CodeGenPrepare, DEBUG_TYPE,
451                     "Optimize for code generation", false, false)
452 
453 FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); }
454 
455 bool CodeGenPrepare::runOnFunction(Function &F) {
456   if (skipFunction(F))
457     return false;
458 
459   DL = &F.getParent()->getDataLayout();
460 
461   bool EverMadeChange = false;
462   // Clear per function information.
463   InsertedInsts.clear();
464   PromotedInsts.clear();
465 
466   TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>();
467   SubtargetInfo = TM->getSubtargetImpl(F);
468   TLI = SubtargetInfo->getTargetLowering();
469   TRI = SubtargetInfo->getRegisterInfo();
470   TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
471   TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
472   LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
473   BPI.reset(new BranchProbabilityInfo(F, *LI));
474   BFI.reset(new BlockFrequencyInfo(F, *BPI, *LI));
475   PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
476   OptSize = F.hasOptSize();
477   if (ProfileGuidedSectionPrefix) {
478     // The hot attribute overwrites profile count based hotness while profile
479     // counts based hotness overwrite the cold attribute.
480     // This is a conservative behabvior.
481     if (F.hasFnAttribute(Attribute::Hot) ||
482         PSI->isFunctionHotInCallGraph(&F, *BFI))
483       F.setSectionPrefix("hot");
484     // If PSI shows this function is not hot, we will placed the function
485     // into unlikely section if (1) PSI shows this is a cold function, or
486     // (2) the function has a attribute of cold.
487     else if (PSI->isFunctionColdInCallGraph(&F, *BFI) ||
488              F.hasFnAttribute(Attribute::Cold))
489       F.setSectionPrefix("unlikely");
490     else if (ProfileUnknownInSpecialSection && PSI->hasPartialSampleProfile() &&
491              PSI->isFunctionHotnessUnknown(F))
492       F.setSectionPrefix("unknown");
493   }
494 
495   /// This optimization identifies DIV instructions that can be
496   /// profitably bypassed and carried out with a shorter, faster divide.
497   if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI->isSlowDivBypassed()) {
498     const DenseMap<unsigned int, unsigned int> &BypassWidths =
499         TLI->getBypassSlowDivWidths();
500     BasicBlock* BB = &*F.begin();
501     while (BB != nullptr) {
502       // bypassSlowDivision may create new BBs, but we don't want to reapply the
503       // optimization to those blocks.
504       BasicBlock* Next = BB->getNextNode();
505       // F.hasOptSize is already checked in the outer if statement.
506       if (!llvm::shouldOptimizeForSize(BB, PSI, BFI.get()))
507         EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
508       BB = Next;
509     }
510   }
511 
512   // Get rid of @llvm.assume builtins before attempting to eliminate empty
513   // blocks, since there might be blocks that only contain @llvm.assume calls
514   // (plus arguments that we can get rid of).
515   EverMadeChange |= eliminateAssumptions(F);
516 
517   // Eliminate blocks that contain only PHI nodes and an
518   // unconditional branch.
519   EverMadeChange |= eliminateMostlyEmptyBlocks(F);
520 
521   bool ModifiedDT = false;
522   if (!DisableBranchOpts)
523     EverMadeChange |= splitBranchCondition(F, ModifiedDT);
524 
525   // Split some critical edges where one of the sources is an indirect branch,
526   // to help generate sane code for PHIs involving such edges.
527   EverMadeChange |= SplitIndirectBrCriticalEdges(F);
528 
529   bool MadeChange = true;
530   while (MadeChange) {
531     MadeChange = false;
532     DT.reset();
533     for (BasicBlock &BB : llvm::make_early_inc_range(F)) {
534       bool ModifiedDTOnIteration = false;
535       MadeChange |= optimizeBlock(BB, ModifiedDTOnIteration);
536 
537       // Restart BB iteration if the dominator tree of the Function was changed
538       if (ModifiedDTOnIteration)
539         break;
540     }
541     if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
542       MadeChange |= mergeSExts(F);
543     if (!LargeOffsetGEPMap.empty())
544       MadeChange |= splitLargeGEPOffsets();
545     MadeChange |= optimizePhiTypes(F);
546 
547     if (MadeChange)
548       eliminateFallThrough(F);
549 
550     // Really free removed instructions during promotion.
551     for (Instruction *I : RemovedInsts)
552       I->deleteValue();
553 
554     EverMadeChange |= MadeChange;
555     SeenChainsForSExt.clear();
556     ValToSExtendedUses.clear();
557     RemovedInsts.clear();
558     LargeOffsetGEPMap.clear();
559     LargeOffsetGEPID.clear();
560   }
561 
562   NewGEPBases.clear();
563   SunkAddrs.clear();
564 
565   if (!DisableBranchOpts) {
566     MadeChange = false;
567     // Use a set vector to get deterministic iteration order. The order the
568     // blocks are removed may affect whether or not PHI nodes in successors
569     // are removed.
570     SmallSetVector<BasicBlock*, 8> WorkList;
571     for (BasicBlock &BB : F) {
572       SmallVector<BasicBlock *, 2> Successors(successors(&BB));
573       MadeChange |= ConstantFoldTerminator(&BB, true);
574       if (!MadeChange) continue;
575 
576       for (BasicBlock *Succ : Successors)
577         if (pred_empty(Succ))
578           WorkList.insert(Succ);
579     }
580 
581     // Delete the dead blocks and any of their dead successors.
582     MadeChange |= !WorkList.empty();
583     while (!WorkList.empty()) {
584       BasicBlock *BB = WorkList.pop_back_val();
585       SmallVector<BasicBlock*, 2> Successors(successors(BB));
586 
587       DeleteDeadBlock(BB);
588 
589       for (BasicBlock *Succ : Successors)
590         if (pred_empty(Succ))
591           WorkList.insert(Succ);
592     }
593 
594     // Merge pairs of basic blocks with unconditional branches, connected by
595     // a single edge.
596     if (EverMadeChange || MadeChange)
597       MadeChange |= eliminateFallThrough(F);
598 
599     EverMadeChange |= MadeChange;
600   }
601 
602   if (!DisableGCOpts) {
603     SmallVector<GCStatepointInst *, 2> Statepoints;
604     for (BasicBlock &BB : F)
605       for (Instruction &I : BB)
606         if (auto *SP = dyn_cast<GCStatepointInst>(&I))
607           Statepoints.push_back(SP);
608     for (auto &I : Statepoints)
609       EverMadeChange |= simplifyOffsetableRelocate(*I);
610   }
611 
612   // Do this last to clean up use-before-def scenarios introduced by other
613   // preparatory transforms.
614   EverMadeChange |= placeDbgValues(F);
615   EverMadeChange |= placePseudoProbes(F);
616 
617 #ifndef NDEBUG
618   if (VerifyBFIUpdates)
619     verifyBFIUpdates(F);
620 #endif
621 
622   return EverMadeChange;
623 }
624 
625 bool CodeGenPrepare::eliminateAssumptions(Function &F) {
626   bool MadeChange = false;
627   for (BasicBlock &BB : F) {
628     CurInstIterator = BB.begin();
629     while (CurInstIterator != BB.end()) {
630       Instruction *I = &*(CurInstIterator++);
631       if (auto *Assume = dyn_cast<AssumeInst>(I)) {
632         MadeChange = true;
633         Value *Operand = Assume->getOperand(0);
634         Assume->eraseFromParent();
635 
636         resetIteratorIfInvalidatedWhileCalling(&BB, [&]() {
637           RecursivelyDeleteTriviallyDeadInstructions(Operand, TLInfo, nullptr);
638         });
639       }
640     }
641   }
642   return MadeChange;
643 }
644 
645 /// An instruction is about to be deleted, so remove all references to it in our
646 /// GEP-tracking data strcutures.
647 void CodeGenPrepare::removeAllAssertingVHReferences(Value *V) {
648   LargeOffsetGEPMap.erase(V);
649   NewGEPBases.erase(V);
650 
651   auto GEP = dyn_cast<GetElementPtrInst>(V);
652   if (!GEP)
653     return;
654 
655   LargeOffsetGEPID.erase(GEP);
656 
657   auto VecI = LargeOffsetGEPMap.find(GEP->getPointerOperand());
658   if (VecI == LargeOffsetGEPMap.end())
659     return;
660 
661   auto &GEPVector = VecI->second;
662   llvm::erase_if(GEPVector, [=](auto &Elt) { return Elt.first == GEP; });
663 
664   if (GEPVector.empty())
665     LargeOffsetGEPMap.erase(VecI);
666 }
667 
668 // Verify BFI has been updated correctly by recomputing BFI and comparing them.
669 void LLVM_ATTRIBUTE_UNUSED CodeGenPrepare::verifyBFIUpdates(Function &F) {
670   DominatorTree NewDT(F);
671   LoopInfo NewLI(NewDT);
672   BranchProbabilityInfo NewBPI(F, NewLI, TLInfo);
673   BlockFrequencyInfo NewBFI(F, NewBPI, NewLI);
674   NewBFI.verifyMatch(*BFI);
675 }
676 
677 /// Merge basic blocks which are connected by a single edge, where one of the
678 /// basic blocks has a single successor pointing to the other basic block,
679 /// which has a single predecessor.
680 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
681   bool Changed = false;
682   // Scan all of the blocks in the function, except for the entry block.
683   // Use a temporary array to avoid iterator being invalidated when
684   // deleting blocks.
685   SmallVector<WeakTrackingVH, 16> Blocks;
686   for (auto &Block : llvm::drop_begin(F))
687     Blocks.push_back(&Block);
688 
689   SmallSet<WeakTrackingVH, 16> Preds;
690   for (auto &Block : Blocks) {
691     auto *BB = cast_or_null<BasicBlock>(Block);
692     if (!BB)
693       continue;
694     // If the destination block has a single pred, then this is a trivial
695     // edge, just collapse it.
696     BasicBlock *SinglePred = BB->getSinglePredecessor();
697 
698     // Don't merge if BB's address is taken.
699     if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
700 
701     BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
702     if (Term && !Term->isConditional()) {
703       Changed = true;
704       LLVM_DEBUG(dbgs() << "To merge:\n" << *BB << "\n\n\n");
705 
706       // Merge BB into SinglePred and delete it.
707       MergeBlockIntoPredecessor(BB);
708       Preds.insert(SinglePred);
709     }
710   }
711 
712   // (Repeatedly) merging blocks into their predecessors can create redundant
713   // debug intrinsics.
714   for (auto &Pred : Preds)
715     if (auto *BB = cast_or_null<BasicBlock>(Pred))
716       RemoveRedundantDbgInstrs(BB);
717 
718   return Changed;
719 }
720 
721 /// Find a destination block from BB if BB is mergeable empty block.
722 BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
723   // If this block doesn't end with an uncond branch, ignore it.
724   BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
725   if (!BI || !BI->isUnconditional())
726     return nullptr;
727 
728   // If the instruction before the branch (skipping debug info) isn't a phi
729   // node, then other stuff is happening here.
730   BasicBlock::iterator BBI = BI->getIterator();
731   if (BBI != BB->begin()) {
732     --BBI;
733     while (isa<DbgInfoIntrinsic>(BBI)) {
734       if (BBI == BB->begin())
735         break;
736       --BBI;
737     }
738     if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
739       return nullptr;
740   }
741 
742   // Do not break infinite loops.
743   BasicBlock *DestBB = BI->getSuccessor(0);
744   if (DestBB == BB)
745     return nullptr;
746 
747   if (!canMergeBlocks(BB, DestBB))
748     DestBB = nullptr;
749 
750   return DestBB;
751 }
752 
753 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
754 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
755 /// edges in ways that are non-optimal for isel. Start by eliminating these
756 /// blocks so we can split them the way we want them.
757 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
758   SmallPtrSet<BasicBlock *, 16> Preheaders;
759   SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
760   while (!LoopList.empty()) {
761     Loop *L = LoopList.pop_back_val();
762     llvm::append_range(LoopList, *L);
763     if (BasicBlock *Preheader = L->getLoopPreheader())
764       Preheaders.insert(Preheader);
765   }
766 
767   bool MadeChange = false;
768   // Copy blocks into a temporary array to avoid iterator invalidation issues
769   // as we remove them.
770   // Note that this intentionally skips the entry block.
771   SmallVector<WeakTrackingVH, 16> Blocks;
772   for (auto &Block : llvm::drop_begin(F))
773     Blocks.push_back(&Block);
774 
775   for (auto &Block : Blocks) {
776     BasicBlock *BB = cast_or_null<BasicBlock>(Block);
777     if (!BB)
778       continue;
779     BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
780     if (!DestBB ||
781         !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB)))
782       continue;
783 
784     eliminateMostlyEmptyBlock(BB);
785     MadeChange = true;
786   }
787   return MadeChange;
788 }
789 
790 bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
791                                                    BasicBlock *DestBB,
792                                                    bool isPreheader) {
793   // Do not delete loop preheaders if doing so would create a critical edge.
794   // Loop preheaders can be good locations to spill registers. If the
795   // preheader is deleted and we create a critical edge, registers may be
796   // spilled in the loop body instead.
797   if (!DisablePreheaderProtect && isPreheader &&
798       !(BB->getSinglePredecessor() &&
799         BB->getSinglePredecessor()->getSingleSuccessor()))
800     return false;
801 
802   // Skip merging if the block's successor is also a successor to any callbr
803   // that leads to this block.
804   // FIXME: Is this really needed? Is this a correctness issue?
805   for (BasicBlock *Pred : predecessors(BB)) {
806     if (auto *CBI = dyn_cast<CallBrInst>((Pred)->getTerminator()))
807       for (unsigned i = 0, e = CBI->getNumSuccessors(); i != e; ++i)
808         if (DestBB == CBI->getSuccessor(i))
809           return false;
810   }
811 
812   // Try to skip merging if the unique predecessor of BB is terminated by a
813   // switch or indirect branch instruction, and BB is used as an incoming block
814   // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
815   // add COPY instructions in the predecessor of BB instead of BB (if it is not
816   // merged). Note that the critical edge created by merging such blocks wont be
817   // split in MachineSink because the jump table is not analyzable. By keeping
818   // such empty block (BB), ISel will place COPY instructions in BB, not in the
819   // predecessor of BB.
820   BasicBlock *Pred = BB->getUniquePredecessor();
821   if (!Pred ||
822       !(isa<SwitchInst>(Pred->getTerminator()) ||
823         isa<IndirectBrInst>(Pred->getTerminator())))
824     return true;
825 
826   if (BB->getTerminator() != BB->getFirstNonPHIOrDbg())
827     return true;
828 
829   // We use a simple cost heuristic which determine skipping merging is
830   // profitable if the cost of skipping merging is less than the cost of
831   // merging : Cost(skipping merging) < Cost(merging BB), where the
832   // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
833   // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
834   // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
835   //   Freq(Pred) / Freq(BB) > 2.
836   // Note that if there are multiple empty blocks sharing the same incoming
837   // value for the PHIs in the DestBB, we consider them together. In such
838   // case, Cost(merging BB) will be the sum of their frequencies.
839 
840   if (!isa<PHINode>(DestBB->begin()))
841     return true;
842 
843   SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
844 
845   // Find all other incoming blocks from which incoming values of all PHIs in
846   // DestBB are the same as the ones from BB.
847   for (BasicBlock *DestBBPred : predecessors(DestBB)) {
848     if (DestBBPred == BB)
849       continue;
850 
851     if (llvm::all_of(DestBB->phis(), [&](const PHINode &DestPN) {
852           return DestPN.getIncomingValueForBlock(BB) ==
853                  DestPN.getIncomingValueForBlock(DestBBPred);
854         }))
855       SameIncomingValueBBs.insert(DestBBPred);
856   }
857 
858   // See if all BB's incoming values are same as the value from Pred. In this
859   // case, no reason to skip merging because COPYs are expected to be place in
860   // Pred already.
861   if (SameIncomingValueBBs.count(Pred))
862     return true;
863 
864   BlockFrequency PredFreq = BFI->getBlockFreq(Pred);
865   BlockFrequency BBFreq = BFI->getBlockFreq(BB);
866 
867   for (auto *SameValueBB : SameIncomingValueBBs)
868     if (SameValueBB->getUniquePredecessor() == Pred &&
869         DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB))
870       BBFreq += BFI->getBlockFreq(SameValueBB);
871 
872   return PredFreq.getFrequency() <=
873          BBFreq.getFrequency() * FreqRatioToSkipMerge;
874 }
875 
876 /// Return true if we can merge BB into DestBB if there is a single
877 /// unconditional branch between them, and BB contains no other non-phi
878 /// instructions.
879 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
880                                     const BasicBlock *DestBB) const {
881   // We only want to eliminate blocks whose phi nodes are used by phi nodes in
882   // the successor.  If there are more complex condition (e.g. preheaders),
883   // don't mess around with them.
884   for (const PHINode &PN : BB->phis()) {
885     for (const User *U : PN.users()) {
886       const Instruction *UI = cast<Instruction>(U);
887       if (UI->getParent() != DestBB || !isa<PHINode>(UI))
888         return false;
889       // If User is inside DestBB block and it is a PHINode then check
890       // incoming value. If incoming value is not from BB then this is
891       // a complex condition (e.g. preheaders) we want to avoid here.
892       if (UI->getParent() == DestBB) {
893         if (const PHINode *UPN = dyn_cast<PHINode>(UI))
894           for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
895             Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
896             if (Insn && Insn->getParent() == BB &&
897                 Insn->getParent() != UPN->getIncomingBlock(I))
898               return false;
899           }
900       }
901     }
902   }
903 
904   // If BB and DestBB contain any common predecessors, then the phi nodes in BB
905   // and DestBB may have conflicting incoming values for the block.  If so, we
906   // can't merge the block.
907   const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
908   if (!DestBBPN) return true;  // no conflict.
909 
910   // Collect the preds of BB.
911   SmallPtrSet<const BasicBlock*, 16> BBPreds;
912   if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
913     // It is faster to get preds from a PHI than with pred_iterator.
914     for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
915       BBPreds.insert(BBPN->getIncomingBlock(i));
916   } else {
917     BBPreds.insert(pred_begin(BB), pred_end(BB));
918   }
919 
920   // Walk the preds of DestBB.
921   for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
922     BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
923     if (BBPreds.count(Pred)) {   // Common predecessor?
924       for (const PHINode &PN : DestBB->phis()) {
925         const Value *V1 = PN.getIncomingValueForBlock(Pred);
926         const Value *V2 = PN.getIncomingValueForBlock(BB);
927 
928         // If V2 is a phi node in BB, look up what the mapped value will be.
929         if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
930           if (V2PN->getParent() == BB)
931             V2 = V2PN->getIncomingValueForBlock(Pred);
932 
933         // If there is a conflict, bail out.
934         if (V1 != V2) return false;
935       }
936     }
937   }
938 
939   return true;
940 }
941 
942 /// Eliminate a basic block that has only phi's and an unconditional branch in
943 /// it.
944 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
945   BranchInst *BI = cast<BranchInst>(BB->getTerminator());
946   BasicBlock *DestBB = BI->getSuccessor(0);
947 
948   LLVM_DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"
949                     << *BB << *DestBB);
950 
951   // If the destination block has a single pred, then this is a trivial edge,
952   // just collapse it.
953   if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
954     if (SinglePred != DestBB) {
955       assert(SinglePred == BB &&
956              "Single predecessor not the same as predecessor");
957       // Merge DestBB into SinglePred/BB and delete it.
958       MergeBlockIntoPredecessor(DestBB);
959       // Note: BB(=SinglePred) will not be deleted on this path.
960       // DestBB(=its single successor) is the one that was deleted.
961       LLVM_DEBUG(dbgs() << "AFTER:\n" << *SinglePred << "\n\n\n");
962       return;
963     }
964   }
965 
966   // Otherwise, we have multiple predecessors of BB.  Update the PHIs in DestBB
967   // to handle the new incoming edges it is about to have.
968   for (PHINode &PN : DestBB->phis()) {
969     // Remove the incoming value for BB, and remember it.
970     Value *InVal = PN.removeIncomingValue(BB, false);
971 
972     // Two options: either the InVal is a phi node defined in BB or it is some
973     // value that dominates BB.
974     PHINode *InValPhi = dyn_cast<PHINode>(InVal);
975     if (InValPhi && InValPhi->getParent() == BB) {
976       // Add all of the input values of the input PHI as inputs of this phi.
977       for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
978         PN.addIncoming(InValPhi->getIncomingValue(i),
979                        InValPhi->getIncomingBlock(i));
980     } else {
981       // Otherwise, add one instance of the dominating value for each edge that
982       // we will be adding.
983       if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
984         for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
985           PN.addIncoming(InVal, BBPN->getIncomingBlock(i));
986       } else {
987         for (BasicBlock *Pred : predecessors(BB))
988           PN.addIncoming(InVal, Pred);
989       }
990     }
991   }
992 
993   // The PHIs are now updated, change everything that refers to BB to use
994   // DestBB and remove BB.
995   BB->replaceAllUsesWith(DestBB);
996   BB->eraseFromParent();
997   ++NumBlocksElim;
998 
999   LLVM_DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1000 }
1001 
1002 // Computes a map of base pointer relocation instructions to corresponding
1003 // derived pointer relocation instructions given a vector of all relocate calls
1004 static void computeBaseDerivedRelocateMap(
1005     const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
1006     DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
1007         &RelocateInstMap) {
1008   // Collect information in two maps: one primarily for locating the base object
1009   // while filling the second map; the second map is the final structure holding
1010   // a mapping between Base and corresponding Derived relocate calls
1011   DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
1012   for (auto *ThisRelocate : AllRelocateCalls) {
1013     auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
1014                             ThisRelocate->getDerivedPtrIndex());
1015     RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
1016   }
1017   for (auto &Item : RelocateIdxMap) {
1018     std::pair<unsigned, unsigned> Key = Item.first;
1019     if (Key.first == Key.second)
1020       // Base relocation: nothing to insert
1021       continue;
1022 
1023     GCRelocateInst *I = Item.second;
1024     auto BaseKey = std::make_pair(Key.first, Key.first);
1025 
1026     // We're iterating over RelocateIdxMap so we cannot modify it.
1027     auto MaybeBase = RelocateIdxMap.find(BaseKey);
1028     if (MaybeBase == RelocateIdxMap.end())
1029       // TODO: We might want to insert a new base object relocate and gep off
1030       // that, if there are enough derived object relocates.
1031       continue;
1032 
1033     RelocateInstMap[MaybeBase->second].push_back(I);
1034   }
1035 }
1036 
1037 // Accepts a GEP and extracts the operands into a vector provided they're all
1038 // small integer constants
1039 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
1040                                           SmallVectorImpl<Value *> &OffsetV) {
1041   for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
1042     // Only accept small constant integer operands
1043     auto *Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
1044     if (!Op || Op->getZExtValue() > 20)
1045       return false;
1046   }
1047 
1048   for (unsigned i = 1; i < GEP->getNumOperands(); i++)
1049     OffsetV.push_back(GEP->getOperand(i));
1050   return true;
1051 }
1052 
1053 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
1054 // replace, computes a replacement, and affects it.
1055 static bool
1056 simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
1057                           const SmallVectorImpl<GCRelocateInst *> &Targets) {
1058   bool MadeChange = false;
1059   // We must ensure the relocation of derived pointer is defined after
1060   // relocation of base pointer. If we find a relocation corresponding to base
1061   // defined earlier than relocation of base then we move relocation of base
1062   // right before found relocation. We consider only relocation in the same
1063   // basic block as relocation of base. Relocations from other basic block will
1064   // be skipped by optimization and we do not care about them.
1065   for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
1066        &*R != RelocatedBase; ++R)
1067     if (auto *RI = dyn_cast<GCRelocateInst>(R))
1068       if (RI->getStatepoint() == RelocatedBase->getStatepoint())
1069         if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
1070           RelocatedBase->moveBefore(RI);
1071           break;
1072         }
1073 
1074   for (GCRelocateInst *ToReplace : Targets) {
1075     assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
1076            "Not relocating a derived object of the original base object");
1077     if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
1078       // A duplicate relocate call. TODO: coalesce duplicates.
1079       continue;
1080     }
1081 
1082     if (RelocatedBase->getParent() != ToReplace->getParent()) {
1083       // Base and derived relocates are in different basic blocks.
1084       // In this case transform is only valid when base dominates derived
1085       // relocate. However it would be too expensive to check dominance
1086       // for each such relocate, so we skip the whole transformation.
1087       continue;
1088     }
1089 
1090     Value *Base = ToReplace->getBasePtr();
1091     auto *Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
1092     if (!Derived || Derived->getPointerOperand() != Base)
1093       continue;
1094 
1095     SmallVector<Value *, 2> OffsetV;
1096     if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
1097       continue;
1098 
1099     // Create a Builder and replace the target callsite with a gep
1100     assert(RelocatedBase->getNextNode() &&
1101            "Should always have one since it's not a terminator");
1102 
1103     // Insert after RelocatedBase
1104     IRBuilder<> Builder(RelocatedBase->getNextNode());
1105     Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1106 
1107     // If gc_relocate does not match the actual type, cast it to the right type.
1108     // In theory, there must be a bitcast after gc_relocate if the type does not
1109     // match, and we should reuse it to get the derived pointer. But it could be
1110     // cases like this:
1111     // bb1:
1112     //  ...
1113     //  %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1114     //  br label %merge
1115     //
1116     // bb2:
1117     //  ...
1118     //  %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1119     //  br label %merge
1120     //
1121     // merge:
1122     //  %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
1123     //  %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
1124     //
1125     // In this case, we can not find the bitcast any more. So we insert a new bitcast
1126     // no matter there is already one or not. In this way, we can handle all cases, and
1127     // the extra bitcast should be optimized away in later passes.
1128     Value *ActualRelocatedBase = RelocatedBase;
1129     if (RelocatedBase->getType() != Base->getType()) {
1130       ActualRelocatedBase =
1131           Builder.CreateBitCast(RelocatedBase, Base->getType());
1132     }
1133     Value *Replacement = Builder.CreateGEP(
1134         Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
1135     Replacement->takeName(ToReplace);
1136     // If the newly generated derived pointer's type does not match the original derived
1137     // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
1138     Value *ActualReplacement = Replacement;
1139     if (Replacement->getType() != ToReplace->getType()) {
1140       ActualReplacement =
1141           Builder.CreateBitCast(Replacement, ToReplace->getType());
1142     }
1143     ToReplace->replaceAllUsesWith(ActualReplacement);
1144     ToReplace->eraseFromParent();
1145 
1146     MadeChange = true;
1147   }
1148   return MadeChange;
1149 }
1150 
1151 // Turns this:
1152 //
1153 // %base = ...
1154 // %ptr = gep %base + 15
1155 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1156 // %base' = relocate(%tok, i32 4, i32 4)
1157 // %ptr' = relocate(%tok, i32 4, i32 5)
1158 // %val = load %ptr'
1159 //
1160 // into this:
1161 //
1162 // %base = ...
1163 // %ptr = gep %base + 15
1164 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1165 // %base' = gc.relocate(%tok, i32 4, i32 4)
1166 // %ptr' = gep %base' + 15
1167 // %val = load %ptr'
1168 bool CodeGenPrepare::simplifyOffsetableRelocate(GCStatepointInst &I) {
1169   bool MadeChange = false;
1170   SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1171   for (auto *U : I.users())
1172     if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1173       // Collect all the relocate calls associated with a statepoint
1174       AllRelocateCalls.push_back(Relocate);
1175 
1176   // We need at least one base pointer relocation + one derived pointer
1177   // relocation to mangle
1178   if (AllRelocateCalls.size() < 2)
1179     return false;
1180 
1181   // RelocateInstMap is a mapping from the base relocate instruction to the
1182   // corresponding derived relocate instructions
1183   DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
1184   computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1185   if (RelocateInstMap.empty())
1186     return false;
1187 
1188   for (auto &Item : RelocateInstMap)
1189     // Item.first is the RelocatedBase to offset against
1190     // Item.second is the vector of Targets to replace
1191     MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1192   return MadeChange;
1193 }
1194 
1195 /// Sink the specified cast instruction into its user blocks.
1196 static bool SinkCast(CastInst *CI) {
1197   BasicBlock *DefBB = CI->getParent();
1198 
1199   /// InsertedCasts - Only insert a cast in each block once.
1200   DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1201 
1202   bool MadeChange = false;
1203   for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1204        UI != E; ) {
1205     Use &TheUse = UI.getUse();
1206     Instruction *User = cast<Instruction>(*UI);
1207 
1208     // Figure out which BB this cast is used in.  For PHI's this is the
1209     // appropriate predecessor block.
1210     BasicBlock *UserBB = User->getParent();
1211     if (PHINode *PN = dyn_cast<PHINode>(User)) {
1212       UserBB = PN->getIncomingBlock(TheUse);
1213     }
1214 
1215     // Preincrement use iterator so we don't invalidate it.
1216     ++UI;
1217 
1218     // The first insertion point of a block containing an EH pad is after the
1219     // pad.  If the pad is the user, we cannot sink the cast past the pad.
1220     if (User->isEHPad())
1221       continue;
1222 
1223     // If the block selected to receive the cast is an EH pad that does not
1224     // allow non-PHI instructions before the terminator, we can't sink the
1225     // cast.
1226     if (UserBB->getTerminator()->isEHPad())
1227       continue;
1228 
1229     // If this user is in the same block as the cast, don't change the cast.
1230     if (UserBB == DefBB) continue;
1231 
1232     // If we have already inserted a cast into this block, use it.
1233     CastInst *&InsertedCast = InsertedCasts[UserBB];
1234 
1235     if (!InsertedCast) {
1236       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1237       assert(InsertPt != UserBB->end());
1238       InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1239                                       CI->getType(), "", &*InsertPt);
1240       InsertedCast->setDebugLoc(CI->getDebugLoc());
1241     }
1242 
1243     // Replace a use of the cast with a use of the new cast.
1244     TheUse = InsertedCast;
1245     MadeChange = true;
1246     ++NumCastUses;
1247   }
1248 
1249   // If we removed all uses, nuke the cast.
1250   if (CI->use_empty()) {
1251     salvageDebugInfo(*CI);
1252     CI->eraseFromParent();
1253     MadeChange = true;
1254   }
1255 
1256   return MadeChange;
1257 }
1258 
1259 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1260 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1261 /// reduce the number of virtual registers that must be created and coalesced.
1262 ///
1263 /// Return true if any changes are made.
1264 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1265                                        const DataLayout &DL) {
1266   // Sink only "cheap" (or nop) address-space casts.  This is a weaker condition
1267   // than sinking only nop casts, but is helpful on some platforms.
1268   if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) {
1269     if (!TLI.isFreeAddrSpaceCast(ASC->getSrcAddressSpace(),
1270                                  ASC->getDestAddressSpace()))
1271       return false;
1272   }
1273 
1274   // If this is a noop copy,
1275   EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1276   EVT DstVT = TLI.getValueType(DL, CI->getType());
1277 
1278   // This is an fp<->int conversion?
1279   if (SrcVT.isInteger() != DstVT.isInteger())
1280     return false;
1281 
1282   // If this is an extension, it will be a zero or sign extension, which
1283   // isn't a noop.
1284   if (SrcVT.bitsLT(DstVT)) return false;
1285 
1286   // If these values will be promoted, find out what they will be promoted
1287   // to.  This helps us consider truncates on PPC as noop copies when they
1288   // are.
1289   if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1290       TargetLowering::TypePromoteInteger)
1291     SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1292   if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1293       TargetLowering::TypePromoteInteger)
1294     DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1295 
1296   // If, after promotion, these are the same types, this is a noop copy.
1297   if (SrcVT != DstVT)
1298     return false;
1299 
1300   return SinkCast(CI);
1301 }
1302 
1303 // Match a simple increment by constant operation.  Note that if a sub is
1304 // matched, the step is negated (as if the step had been canonicalized to
1305 // an add, even though we leave the instruction alone.)
1306 bool matchIncrement(const Instruction* IVInc, Instruction *&LHS,
1307                     Constant *&Step) {
1308   if (match(IVInc, m_Add(m_Instruction(LHS), m_Constant(Step))) ||
1309       match(IVInc, m_ExtractValue<0>(m_Intrinsic<Intrinsic::uadd_with_overflow>(
1310                        m_Instruction(LHS), m_Constant(Step)))))
1311     return true;
1312   if (match(IVInc, m_Sub(m_Instruction(LHS), m_Constant(Step))) ||
1313       match(IVInc, m_ExtractValue<0>(m_Intrinsic<Intrinsic::usub_with_overflow>(
1314                        m_Instruction(LHS), m_Constant(Step))))) {
1315     Step = ConstantExpr::getNeg(Step);
1316     return true;
1317   }
1318   return false;
1319 }
1320 
1321 /// If given \p PN is an inductive variable with value IVInc coming from the
1322 /// backedge, and on each iteration it gets increased by Step, return pair
1323 /// <IVInc, Step>. Otherwise, return None.
1324 static Optional<std::pair<Instruction *, Constant *> >
1325 getIVIncrement(const PHINode *PN, const LoopInfo *LI) {
1326   const Loop *L = LI->getLoopFor(PN->getParent());
1327   if (!L || L->getHeader() != PN->getParent() || !L->getLoopLatch())
1328     return None;
1329   auto *IVInc =
1330       dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
1331   if (!IVInc || LI->getLoopFor(IVInc->getParent()) != L)
1332     return None;
1333   Instruction *LHS = nullptr;
1334   Constant *Step = nullptr;
1335   if (matchIncrement(IVInc, LHS, Step) && LHS == PN)
1336     return std::make_pair(IVInc, Step);
1337   return None;
1338 }
1339 
1340 static bool isIVIncrement(const Value *V, const LoopInfo *LI) {
1341   auto *I = dyn_cast<Instruction>(V);
1342   if (!I)
1343     return false;
1344   Instruction *LHS = nullptr;
1345   Constant *Step = nullptr;
1346   if (!matchIncrement(I, LHS, Step))
1347     return false;
1348   if (auto *PN = dyn_cast<PHINode>(LHS))
1349     if (auto IVInc = getIVIncrement(PN, LI))
1350       return IVInc->first == I;
1351   return false;
1352 }
1353 
1354 bool CodeGenPrepare::replaceMathCmpWithIntrinsic(BinaryOperator *BO,
1355                                                  Value *Arg0, Value *Arg1,
1356                                                  CmpInst *Cmp,
1357                                                  Intrinsic::ID IID) {
1358   auto IsReplacableIVIncrement = [this, &Cmp](BinaryOperator *BO) {
1359     if (!isIVIncrement(BO, LI))
1360       return false;
1361     const Loop *L = LI->getLoopFor(BO->getParent());
1362     assert(L && "L should not be null after isIVIncrement()");
1363     // Do not risk on moving increment into a child loop.
1364     if (LI->getLoopFor(Cmp->getParent()) != L)
1365       return false;
1366 
1367     // Finally, we need to ensure that the insert point will dominate all
1368     // existing uses of the increment.
1369 
1370     auto &DT = getDT(*BO->getParent()->getParent());
1371     if (DT.dominates(Cmp->getParent(), BO->getParent()))
1372       // If we're moving up the dom tree, all uses are trivially dominated.
1373       // (This is the common case for code produced by LSR.)
1374       return true;
1375 
1376     // Otherwise, special case the single use in the phi recurrence.
1377     return BO->hasOneUse() && DT.dominates(Cmp->getParent(), L->getLoopLatch());
1378   };
1379   if (BO->getParent() != Cmp->getParent() && !IsReplacableIVIncrement(BO)) {
1380     // We used to use a dominator tree here to allow multi-block optimization.
1381     // But that was problematic because:
1382     // 1. It could cause a perf regression by hoisting the math op into the
1383     //    critical path.
1384     // 2. It could cause a perf regression by creating a value that was live
1385     //    across multiple blocks and increasing register pressure.
1386     // 3. Use of a dominator tree could cause large compile-time regression.
1387     //    This is because we recompute the DT on every change in the main CGP
1388     //    run-loop. The recomputing is probably unnecessary in many cases, so if
1389     //    that was fixed, using a DT here would be ok.
1390     //
1391     // There is one important particular case we still want to handle: if BO is
1392     // the IV increment. Important properties that make it profitable:
1393     // - We can speculate IV increment anywhere in the loop (as long as the
1394     //   indvar Phi is its only user);
1395     // - Upon computing Cmp, we effectively compute something equivalent to the
1396     //   IV increment (despite it loops differently in the IR). So moving it up
1397     //   to the cmp point does not really increase register pressure.
1398     return false;
1399   }
1400 
1401   // We allow matching the canonical IR (add X, C) back to (usubo X, -C).
1402   if (BO->getOpcode() == Instruction::Add &&
1403       IID == Intrinsic::usub_with_overflow) {
1404     assert(isa<Constant>(Arg1) && "Unexpected input for usubo");
1405     Arg1 = ConstantExpr::getNeg(cast<Constant>(Arg1));
1406   }
1407 
1408   // Insert at the first instruction of the pair.
1409   Instruction *InsertPt = nullptr;
1410   for (Instruction &Iter : *Cmp->getParent()) {
1411     // If BO is an XOR, it is not guaranteed that it comes after both inputs to
1412     // the overflow intrinsic are defined.
1413     if ((BO->getOpcode() != Instruction::Xor && &Iter == BO) || &Iter == Cmp) {
1414       InsertPt = &Iter;
1415       break;
1416     }
1417   }
1418   assert(InsertPt != nullptr && "Parent block did not contain cmp or binop");
1419 
1420   IRBuilder<> Builder(InsertPt);
1421   Value *MathOV = Builder.CreateBinaryIntrinsic(IID, Arg0, Arg1);
1422   if (BO->getOpcode() != Instruction::Xor) {
1423     Value *Math = Builder.CreateExtractValue(MathOV, 0, "math");
1424     BO->replaceAllUsesWith(Math);
1425   } else
1426     assert(BO->hasOneUse() &&
1427            "Patterns with XOr should use the BO only in the compare");
1428   Value *OV = Builder.CreateExtractValue(MathOV, 1, "ov");
1429   Cmp->replaceAllUsesWith(OV);
1430   Cmp->eraseFromParent();
1431   BO->eraseFromParent();
1432   return true;
1433 }
1434 
1435 /// Match special-case patterns that check for unsigned add overflow.
1436 static bool matchUAddWithOverflowConstantEdgeCases(CmpInst *Cmp,
1437                                                    BinaryOperator *&Add) {
1438   // Add = add A, 1; Cmp = icmp eq A,-1 (overflow if A is max val)
1439   // Add = add A,-1; Cmp = icmp ne A, 0 (overflow if A is non-zero)
1440   Value *A = Cmp->getOperand(0), *B = Cmp->getOperand(1);
1441 
1442   // We are not expecting non-canonical/degenerate code. Just bail out.
1443   if (isa<Constant>(A))
1444     return false;
1445 
1446   ICmpInst::Predicate Pred = Cmp->getPredicate();
1447   if (Pred == ICmpInst::ICMP_EQ && match(B, m_AllOnes()))
1448     B = ConstantInt::get(B->getType(), 1);
1449   else if (Pred == ICmpInst::ICMP_NE && match(B, m_ZeroInt()))
1450     B = ConstantInt::get(B->getType(), -1);
1451   else
1452     return false;
1453 
1454   // Check the users of the variable operand of the compare looking for an add
1455   // with the adjusted constant.
1456   for (User *U : A->users()) {
1457     if (match(U, m_Add(m_Specific(A), m_Specific(B)))) {
1458       Add = cast<BinaryOperator>(U);
1459       return true;
1460     }
1461   }
1462   return false;
1463 }
1464 
1465 /// Try to combine the compare into a call to the llvm.uadd.with.overflow
1466 /// intrinsic. Return true if any changes were made.
1467 bool CodeGenPrepare::combineToUAddWithOverflow(CmpInst *Cmp,
1468                                                bool &ModifiedDT) {
1469   Value *A, *B;
1470   BinaryOperator *Add;
1471   if (!match(Cmp, m_UAddWithOverflow(m_Value(A), m_Value(B), m_BinOp(Add)))) {
1472     if (!matchUAddWithOverflowConstantEdgeCases(Cmp, Add))
1473       return false;
1474     // Set A and B in case we match matchUAddWithOverflowConstantEdgeCases.
1475     A = Add->getOperand(0);
1476     B = Add->getOperand(1);
1477   }
1478 
1479   if (!TLI->shouldFormOverflowOp(ISD::UADDO,
1480                                  TLI->getValueType(*DL, Add->getType()),
1481                                  Add->hasNUsesOrMore(2)))
1482     return false;
1483 
1484   // We don't want to move around uses of condition values this late, so we
1485   // check if it is legal to create the call to the intrinsic in the basic
1486   // block containing the icmp.
1487   if (Add->getParent() != Cmp->getParent() && !Add->hasOneUse())
1488     return false;
1489 
1490   if (!replaceMathCmpWithIntrinsic(Add, A, B, Cmp,
1491                                    Intrinsic::uadd_with_overflow))
1492     return false;
1493 
1494   // Reset callers - do not crash by iterating over a dead instruction.
1495   ModifiedDT = true;
1496   return true;
1497 }
1498 
1499 bool CodeGenPrepare::combineToUSubWithOverflow(CmpInst *Cmp,
1500                                                bool &ModifiedDT) {
1501   // We are not expecting non-canonical/degenerate code. Just bail out.
1502   Value *A = Cmp->getOperand(0), *B = Cmp->getOperand(1);
1503   if (isa<Constant>(A) && isa<Constant>(B))
1504     return false;
1505 
1506   // Convert (A u> B) to (A u< B) to simplify pattern matching.
1507   ICmpInst::Predicate Pred = Cmp->getPredicate();
1508   if (Pred == ICmpInst::ICMP_UGT) {
1509     std::swap(A, B);
1510     Pred = ICmpInst::ICMP_ULT;
1511   }
1512   // Convert special-case: (A == 0) is the same as (A u< 1).
1513   if (Pred == ICmpInst::ICMP_EQ && match(B, m_ZeroInt())) {
1514     B = ConstantInt::get(B->getType(), 1);
1515     Pred = ICmpInst::ICMP_ULT;
1516   }
1517   // Convert special-case: (A != 0) is the same as (0 u< A).
1518   if (Pred == ICmpInst::ICMP_NE && match(B, m_ZeroInt())) {
1519     std::swap(A, B);
1520     Pred = ICmpInst::ICMP_ULT;
1521   }
1522   if (Pred != ICmpInst::ICMP_ULT)
1523     return false;
1524 
1525   // Walk the users of a variable operand of a compare looking for a subtract or
1526   // add with that same operand. Also match the 2nd operand of the compare to
1527   // the add/sub, but that may be a negated constant operand of an add.
1528   Value *CmpVariableOperand = isa<Constant>(A) ? B : A;
1529   BinaryOperator *Sub = nullptr;
1530   for (User *U : CmpVariableOperand->users()) {
1531     // A - B, A u< B --> usubo(A, B)
1532     if (match(U, m_Sub(m_Specific(A), m_Specific(B)))) {
1533       Sub = cast<BinaryOperator>(U);
1534       break;
1535     }
1536 
1537     // A + (-C), A u< C (canonicalized form of (sub A, C))
1538     const APInt *CmpC, *AddC;
1539     if (match(U, m_Add(m_Specific(A), m_APInt(AddC))) &&
1540         match(B, m_APInt(CmpC)) && *AddC == -(*CmpC)) {
1541       Sub = cast<BinaryOperator>(U);
1542       break;
1543     }
1544   }
1545   if (!Sub)
1546     return false;
1547 
1548   if (!TLI->shouldFormOverflowOp(ISD::USUBO,
1549                                  TLI->getValueType(*DL, Sub->getType()),
1550                                  Sub->hasNUsesOrMore(2)))
1551     return false;
1552 
1553   if (!replaceMathCmpWithIntrinsic(Sub, Sub->getOperand(0), Sub->getOperand(1),
1554                                    Cmp, Intrinsic::usub_with_overflow))
1555     return false;
1556 
1557   // Reset callers - do not crash by iterating over a dead instruction.
1558   ModifiedDT = true;
1559   return true;
1560 }
1561 
1562 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1563 /// registers that must be created and coalesced. This is a clear win except on
1564 /// targets with multiple condition code registers (PowerPC), where it might
1565 /// lose; some adjustment may be wanted there.
1566 ///
1567 /// Return true if any changes are made.
1568 static bool sinkCmpExpression(CmpInst *Cmp, const TargetLowering &TLI) {
1569   if (TLI.hasMultipleConditionRegisters())
1570     return false;
1571 
1572   // Avoid sinking soft-FP comparisons, since this can move them into a loop.
1573   if (TLI.useSoftFloat() && isa<FCmpInst>(Cmp))
1574     return false;
1575 
1576   // Only insert a cmp in each block once.
1577   DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1578 
1579   bool MadeChange = false;
1580   for (Value::user_iterator UI = Cmp->user_begin(), E = Cmp->user_end();
1581        UI != E; ) {
1582     Use &TheUse = UI.getUse();
1583     Instruction *User = cast<Instruction>(*UI);
1584 
1585     // Preincrement use iterator so we don't invalidate it.
1586     ++UI;
1587 
1588     // Don't bother for PHI nodes.
1589     if (isa<PHINode>(User))
1590       continue;
1591 
1592     // Figure out which BB this cmp is used in.
1593     BasicBlock *UserBB = User->getParent();
1594     BasicBlock *DefBB = Cmp->getParent();
1595 
1596     // If this user is in the same block as the cmp, don't change the cmp.
1597     if (UserBB == DefBB) continue;
1598 
1599     // If we have already inserted a cmp into this block, use it.
1600     CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1601 
1602     if (!InsertedCmp) {
1603       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1604       assert(InsertPt != UserBB->end());
1605       InsertedCmp =
1606           CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(),
1607                           Cmp->getOperand(0), Cmp->getOperand(1), "",
1608                           &*InsertPt);
1609       // Propagate the debug info.
1610       InsertedCmp->setDebugLoc(Cmp->getDebugLoc());
1611     }
1612 
1613     // Replace a use of the cmp with a use of the new cmp.
1614     TheUse = InsertedCmp;
1615     MadeChange = true;
1616     ++NumCmpUses;
1617   }
1618 
1619   // If we removed all uses, nuke the cmp.
1620   if (Cmp->use_empty()) {
1621     Cmp->eraseFromParent();
1622     MadeChange = true;
1623   }
1624 
1625   return MadeChange;
1626 }
1627 
1628 /// For pattern like:
1629 ///
1630 ///   DomCond = icmp sgt/slt CmpOp0, CmpOp1 (might not be in DomBB)
1631 ///   ...
1632 /// DomBB:
1633 ///   ...
1634 ///   br DomCond, TrueBB, CmpBB
1635 /// CmpBB: (with DomBB being the single predecessor)
1636 ///   ...
1637 ///   Cmp = icmp eq CmpOp0, CmpOp1
1638 ///   ...
1639 ///
1640 /// It would use two comparison on targets that lowering of icmp sgt/slt is
1641 /// different from lowering of icmp eq (PowerPC). This function try to convert
1642 /// 'Cmp = icmp eq CmpOp0, CmpOp1' to ' Cmp = icmp slt/sgt CmpOp0, CmpOp1'.
1643 /// After that, DomCond and Cmp can use the same comparison so reduce one
1644 /// comparison.
1645 ///
1646 /// Return true if any changes are made.
1647 static bool foldICmpWithDominatingICmp(CmpInst *Cmp,
1648                                        const TargetLowering &TLI) {
1649   if (!EnableICMP_EQToICMP_ST && TLI.isEqualityCmpFoldedWithSignedCmp())
1650     return false;
1651 
1652   ICmpInst::Predicate Pred = Cmp->getPredicate();
1653   if (Pred != ICmpInst::ICMP_EQ)
1654     return false;
1655 
1656   // If icmp eq has users other than BranchInst and SelectInst, converting it to
1657   // icmp slt/sgt would introduce more redundant LLVM IR.
1658   for (User *U : Cmp->users()) {
1659     if (isa<BranchInst>(U))
1660       continue;
1661     if (isa<SelectInst>(U) && cast<SelectInst>(U)->getCondition() == Cmp)
1662       continue;
1663     return false;
1664   }
1665 
1666   // This is a cheap/incomplete check for dominance - just match a single
1667   // predecessor with a conditional branch.
1668   BasicBlock *CmpBB = Cmp->getParent();
1669   BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1670   if (!DomBB)
1671     return false;
1672 
1673   // We want to ensure that the only way control gets to the comparison of
1674   // interest is that a less/greater than comparison on the same operands is
1675   // false.
1676   Value *DomCond;
1677   BasicBlock *TrueBB, *FalseBB;
1678   if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1679     return false;
1680   if (CmpBB != FalseBB)
1681     return false;
1682 
1683   Value *CmpOp0 = Cmp->getOperand(0), *CmpOp1 = Cmp->getOperand(1);
1684   ICmpInst::Predicate DomPred;
1685   if (!match(DomCond, m_ICmp(DomPred, m_Specific(CmpOp0), m_Specific(CmpOp1))))
1686     return false;
1687   if (DomPred != ICmpInst::ICMP_SGT && DomPred != ICmpInst::ICMP_SLT)
1688     return false;
1689 
1690   // Convert the equality comparison to the opposite of the dominating
1691   // comparison and swap the direction for all branch/select users.
1692   // We have conceptually converted:
1693   // Res = (a < b) ? <LT_RES> : (a == b) ? <EQ_RES> : <GT_RES>;
1694   // to
1695   // Res = (a < b) ? <LT_RES> : (a > b)  ? <GT_RES> : <EQ_RES>;
1696   // And similarly for branches.
1697   for (User *U : Cmp->users()) {
1698     if (auto *BI = dyn_cast<BranchInst>(U)) {
1699       assert(BI->isConditional() && "Must be conditional");
1700       BI->swapSuccessors();
1701       continue;
1702     }
1703     if (auto *SI = dyn_cast<SelectInst>(U)) {
1704       // Swap operands
1705       SI->swapValues();
1706       SI->swapProfMetadata();
1707       continue;
1708     }
1709     llvm_unreachable("Must be a branch or a select");
1710   }
1711   Cmp->setPredicate(CmpInst::getSwappedPredicate(DomPred));
1712   return true;
1713 }
1714 
1715 bool CodeGenPrepare::optimizeCmp(CmpInst *Cmp, bool &ModifiedDT) {
1716   if (sinkCmpExpression(Cmp, *TLI))
1717     return true;
1718 
1719   if (combineToUAddWithOverflow(Cmp, ModifiedDT))
1720     return true;
1721 
1722   if (combineToUSubWithOverflow(Cmp, ModifiedDT))
1723     return true;
1724 
1725   if (foldICmpWithDominatingICmp(Cmp, *TLI))
1726     return true;
1727 
1728   return false;
1729 }
1730 
1731 /// Duplicate and sink the given 'and' instruction into user blocks where it is
1732 /// used in a compare to allow isel to generate better code for targets where
1733 /// this operation can be combined.
1734 ///
1735 /// Return true if any changes are made.
1736 static bool sinkAndCmp0Expression(Instruction *AndI,
1737                                   const TargetLowering &TLI,
1738                                   SetOfInstrs &InsertedInsts) {
1739   // Double-check that we're not trying to optimize an instruction that was
1740   // already optimized by some other part of this pass.
1741   assert(!InsertedInsts.count(AndI) &&
1742          "Attempting to optimize already optimized and instruction");
1743   (void) InsertedInsts;
1744 
1745   // Nothing to do for single use in same basic block.
1746   if (AndI->hasOneUse() &&
1747       AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent())
1748     return false;
1749 
1750   // Try to avoid cases where sinking/duplicating is likely to increase register
1751   // pressure.
1752   if (!isa<ConstantInt>(AndI->getOperand(0)) &&
1753       !isa<ConstantInt>(AndI->getOperand(1)) &&
1754       AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse())
1755     return false;
1756 
1757   for (auto *U : AndI->users()) {
1758     Instruction *User = cast<Instruction>(U);
1759 
1760     // Only sink 'and' feeding icmp with 0.
1761     if (!isa<ICmpInst>(User))
1762       return false;
1763 
1764     auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1));
1765     if (!CmpC || !CmpC->isZero())
1766       return false;
1767   }
1768 
1769   if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI))
1770     return false;
1771 
1772   LLVM_DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n");
1773   LLVM_DEBUG(AndI->getParent()->dump());
1774 
1775   // Push the 'and' into the same block as the icmp 0.  There should only be
1776   // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
1777   // others, so we don't need to keep track of which BBs we insert into.
1778   for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
1779        UI != E; ) {
1780     Use &TheUse = UI.getUse();
1781     Instruction *User = cast<Instruction>(*UI);
1782 
1783     // Preincrement use iterator so we don't invalidate it.
1784     ++UI;
1785 
1786     LLVM_DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n");
1787 
1788     // Keep the 'and' in the same place if the use is already in the same block.
1789     Instruction *InsertPt =
1790         User->getParent() == AndI->getParent() ? AndI : User;
1791     Instruction *InsertedAnd =
1792         BinaryOperator::Create(Instruction::And, AndI->getOperand(0),
1793                                AndI->getOperand(1), "", InsertPt);
1794     // Propagate the debug info.
1795     InsertedAnd->setDebugLoc(AndI->getDebugLoc());
1796 
1797     // Replace a use of the 'and' with a use of the new 'and'.
1798     TheUse = InsertedAnd;
1799     ++NumAndUses;
1800     LLVM_DEBUG(User->getParent()->dump());
1801   }
1802 
1803   // We removed all uses, nuke the and.
1804   AndI->eraseFromParent();
1805   return true;
1806 }
1807 
1808 /// Check if the candidates could be combined with a shift instruction, which
1809 /// includes:
1810 /// 1. Truncate instruction
1811 /// 2. And instruction and the imm is a mask of the low bits:
1812 /// imm & (imm+1) == 0
1813 static bool isExtractBitsCandidateUse(Instruction *User) {
1814   if (!isa<TruncInst>(User)) {
1815     if (User->getOpcode() != Instruction::And ||
1816         !isa<ConstantInt>(User->getOperand(1)))
1817       return false;
1818 
1819     const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
1820 
1821     if ((Cimm & (Cimm + 1)).getBoolValue())
1822       return false;
1823   }
1824   return true;
1825 }
1826 
1827 /// Sink both shift and truncate instruction to the use of truncate's BB.
1828 static bool
1829 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
1830                      DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
1831                      const TargetLowering &TLI, const DataLayout &DL) {
1832   BasicBlock *UserBB = User->getParent();
1833   DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
1834   auto *TruncI = cast<TruncInst>(User);
1835   bool MadeChange = false;
1836 
1837   for (Value::user_iterator TruncUI = TruncI->user_begin(),
1838                             TruncE = TruncI->user_end();
1839        TruncUI != TruncE;) {
1840 
1841     Use &TruncTheUse = TruncUI.getUse();
1842     Instruction *TruncUser = cast<Instruction>(*TruncUI);
1843     // Preincrement use iterator so we don't invalidate it.
1844 
1845     ++TruncUI;
1846 
1847     int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
1848     if (!ISDOpcode)
1849       continue;
1850 
1851     // If the use is actually a legal node, there will not be an
1852     // implicit truncate.
1853     // FIXME: always querying the result type is just an
1854     // approximation; some nodes' legality is determined by the
1855     // operand or other means. There's no good way to find out though.
1856     if (TLI.isOperationLegalOrCustom(
1857             ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
1858       continue;
1859 
1860     // Don't bother for PHI nodes.
1861     if (isa<PHINode>(TruncUser))
1862       continue;
1863 
1864     BasicBlock *TruncUserBB = TruncUser->getParent();
1865 
1866     if (UserBB == TruncUserBB)
1867       continue;
1868 
1869     BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
1870     CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
1871 
1872     if (!InsertedShift && !InsertedTrunc) {
1873       BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
1874       assert(InsertPt != TruncUserBB->end());
1875       // Sink the shift
1876       if (ShiftI->getOpcode() == Instruction::AShr)
1877         InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1878                                                    "", &*InsertPt);
1879       else
1880         InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1881                                                    "", &*InsertPt);
1882       InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
1883 
1884       // Sink the trunc
1885       BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
1886       TruncInsertPt++;
1887       assert(TruncInsertPt != TruncUserBB->end());
1888 
1889       InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
1890                                        TruncI->getType(), "", &*TruncInsertPt);
1891       InsertedTrunc->setDebugLoc(TruncI->getDebugLoc());
1892 
1893       MadeChange = true;
1894 
1895       TruncTheUse = InsertedTrunc;
1896     }
1897   }
1898   return MadeChange;
1899 }
1900 
1901 /// Sink the shift *right* instruction into user blocks if the uses could
1902 /// potentially be combined with this shift instruction and generate BitExtract
1903 /// instruction. It will only be applied if the architecture supports BitExtract
1904 /// instruction. Here is an example:
1905 /// BB1:
1906 ///   %x.extract.shift = lshr i64 %arg1, 32
1907 /// BB2:
1908 ///   %x.extract.trunc = trunc i64 %x.extract.shift to i16
1909 /// ==>
1910 ///
1911 /// BB2:
1912 ///   %x.extract.shift.1 = lshr i64 %arg1, 32
1913 ///   %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1914 ///
1915 /// CodeGen will recognize the pattern in BB2 and generate BitExtract
1916 /// instruction.
1917 /// Return true if any changes are made.
1918 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1919                                 const TargetLowering &TLI,
1920                                 const DataLayout &DL) {
1921   BasicBlock *DefBB = ShiftI->getParent();
1922 
1923   /// Only insert instructions in each block once.
1924   DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1925 
1926   bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1927 
1928   bool MadeChange = false;
1929   for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1930        UI != E;) {
1931     Use &TheUse = UI.getUse();
1932     Instruction *User = cast<Instruction>(*UI);
1933     // Preincrement use iterator so we don't invalidate it.
1934     ++UI;
1935 
1936     // Don't bother for PHI nodes.
1937     if (isa<PHINode>(User))
1938       continue;
1939 
1940     if (!isExtractBitsCandidateUse(User))
1941       continue;
1942 
1943     BasicBlock *UserBB = User->getParent();
1944 
1945     if (UserBB == DefBB) {
1946       // If the shift and truncate instruction are in the same BB. The use of
1947       // the truncate(TruncUse) may still introduce another truncate if not
1948       // legal. In this case, we would like to sink both shift and truncate
1949       // instruction to the BB of TruncUse.
1950       // for example:
1951       // BB1:
1952       // i64 shift.result = lshr i64 opnd, imm
1953       // trunc.result = trunc shift.result to i16
1954       //
1955       // BB2:
1956       //   ----> We will have an implicit truncate here if the architecture does
1957       //   not have i16 compare.
1958       // cmp i16 trunc.result, opnd2
1959       //
1960       if (isa<TruncInst>(User) && shiftIsLegal
1961           // If the type of the truncate is legal, no truncate will be
1962           // introduced in other basic blocks.
1963           &&
1964           (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1965         MadeChange =
1966             SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1967 
1968       continue;
1969     }
1970     // If we have already inserted a shift into this block, use it.
1971     BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1972 
1973     if (!InsertedShift) {
1974       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1975       assert(InsertPt != UserBB->end());
1976 
1977       if (ShiftI->getOpcode() == Instruction::AShr)
1978         InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1979                                                    "", &*InsertPt);
1980       else
1981         InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1982                                                    "", &*InsertPt);
1983       InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
1984 
1985       MadeChange = true;
1986     }
1987 
1988     // Replace a use of the shift with a use of the new shift.
1989     TheUse = InsertedShift;
1990   }
1991 
1992   // If we removed all uses, or there are none, nuke the shift.
1993   if (ShiftI->use_empty()) {
1994     salvageDebugInfo(*ShiftI);
1995     ShiftI->eraseFromParent();
1996     MadeChange = true;
1997   }
1998 
1999   return MadeChange;
2000 }
2001 
2002 /// If counting leading or trailing zeros is an expensive operation and a zero
2003 /// input is defined, add a check for zero to avoid calling the intrinsic.
2004 ///
2005 /// We want to transform:
2006 ///     %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
2007 ///
2008 /// into:
2009 ///   entry:
2010 ///     %cmpz = icmp eq i64 %A, 0
2011 ///     br i1 %cmpz, label %cond.end, label %cond.false
2012 ///   cond.false:
2013 ///     %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
2014 ///     br label %cond.end
2015 ///   cond.end:
2016 ///     %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
2017 ///
2018 /// If the transform is performed, return true and set ModifiedDT to true.
2019 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
2020                                   const TargetLowering *TLI,
2021                                   const DataLayout *DL,
2022                                   bool &ModifiedDT) {
2023   // If a zero input is undefined, it doesn't make sense to despeculate that.
2024   if (match(CountZeros->getOperand(1), m_One()))
2025     return false;
2026 
2027   // If it's cheap to speculate, there's nothing to do.
2028   auto IntrinsicID = CountZeros->getIntrinsicID();
2029   if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
2030       (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
2031     return false;
2032 
2033   // Only handle legal scalar cases. Anything else requires too much work.
2034   Type *Ty = CountZeros->getType();
2035   unsigned SizeInBits = Ty->getScalarSizeInBits();
2036   if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits())
2037     return false;
2038 
2039   // Bail if the value is never zero.
2040   if (llvm::isKnownNonZero(CountZeros->getOperand(0), *DL))
2041     return false;
2042 
2043   // The intrinsic will be sunk behind a compare against zero and branch.
2044   BasicBlock *StartBlock = CountZeros->getParent();
2045   BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
2046 
2047   // Create another block after the count zero intrinsic. A PHI will be added
2048   // in this block to select the result of the intrinsic or the bit-width
2049   // constant if the input to the intrinsic is zero.
2050   BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
2051   BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
2052 
2053   // Set up a builder to create a compare, conditional branch, and PHI.
2054   IRBuilder<> Builder(CountZeros->getContext());
2055   Builder.SetInsertPoint(StartBlock->getTerminator());
2056   Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
2057 
2058   // Replace the unconditional branch that was created by the first split with
2059   // a compare against zero and a conditional branch.
2060   Value *Zero = Constant::getNullValue(Ty);
2061   Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
2062   Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
2063   StartBlock->getTerminator()->eraseFromParent();
2064 
2065   // Create a PHI in the end block to select either the output of the intrinsic
2066   // or the bit width of the operand.
2067   Builder.SetInsertPoint(&EndBlock->front());
2068   PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
2069   CountZeros->replaceAllUsesWith(PN);
2070   Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
2071   PN->addIncoming(BitWidth, StartBlock);
2072   PN->addIncoming(CountZeros, CallBlock);
2073 
2074   // We are explicitly handling the zero case, so we can set the intrinsic's
2075   // undefined zero argument to 'true'. This will also prevent reprocessing the
2076   // intrinsic; we only despeculate when a zero input is defined.
2077   CountZeros->setArgOperand(1, Builder.getTrue());
2078   ModifiedDT = true;
2079   return true;
2080 }
2081 
2082 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) {
2083   BasicBlock *BB = CI->getParent();
2084 
2085   // Lower inline assembly if we can.
2086   // If we found an inline asm expession, and if the target knows how to
2087   // lower it to normal LLVM code, do so now.
2088   if (CI->isInlineAsm()) {
2089     if (TLI->ExpandInlineAsm(CI)) {
2090       // Avoid invalidating the iterator.
2091       CurInstIterator = BB->begin();
2092       // Avoid processing instructions out of order, which could cause
2093       // reuse before a value is defined.
2094       SunkAddrs.clear();
2095       return true;
2096     }
2097     // Sink address computing for memory operands into the block.
2098     if (optimizeInlineAsmInst(CI))
2099       return true;
2100   }
2101 
2102   // Align the pointer arguments to this call if the target thinks it's a good
2103   // idea
2104   unsigned MinSize, PrefAlign;
2105   if (TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
2106     for (auto &Arg : CI->args()) {
2107       // We want to align both objects whose address is used directly and
2108       // objects whose address is used in casts and GEPs, though it only makes
2109       // sense for GEPs if the offset is a multiple of the desired alignment and
2110       // if size - offset meets the size threshold.
2111       if (!Arg->getType()->isPointerTy())
2112         continue;
2113       APInt Offset(DL->getIndexSizeInBits(
2114                        cast<PointerType>(Arg->getType())->getAddressSpace()),
2115                    0);
2116       Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
2117       uint64_t Offset2 = Offset.getLimitedValue();
2118       if ((Offset2 & (PrefAlign-1)) != 0)
2119         continue;
2120       AllocaInst *AI;
2121       if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
2122           DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
2123         AI->setAlignment(Align(PrefAlign));
2124       // Global variables can only be aligned if they are defined in this
2125       // object (i.e. they are uniquely initialized in this object), and
2126       // over-aligning global variables that have an explicit section is
2127       // forbidden.
2128       GlobalVariable *GV;
2129       if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
2130           GV->getPointerAlignment(*DL) < PrefAlign &&
2131           DL->getTypeAllocSize(GV->getValueType()) >=
2132               MinSize + Offset2)
2133         GV->setAlignment(MaybeAlign(PrefAlign));
2134     }
2135     // If this is a memcpy (or similar) then we may be able to improve the
2136     // alignment
2137     if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
2138       Align DestAlign = getKnownAlignment(MI->getDest(), *DL);
2139       MaybeAlign MIDestAlign = MI->getDestAlign();
2140       if (!MIDestAlign || DestAlign > *MIDestAlign)
2141         MI->setDestAlignment(DestAlign);
2142       if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
2143         MaybeAlign MTISrcAlign = MTI->getSourceAlign();
2144         Align SrcAlign = getKnownAlignment(MTI->getSource(), *DL);
2145         if (!MTISrcAlign || SrcAlign > *MTISrcAlign)
2146           MTI->setSourceAlignment(SrcAlign);
2147       }
2148     }
2149   }
2150 
2151   // If we have a cold call site, try to sink addressing computation into the
2152   // cold block.  This interacts with our handling for loads and stores to
2153   // ensure that we can fold all uses of a potential addressing computation
2154   // into their uses.  TODO: generalize this to work over profiling data
2155   if (CI->hasFnAttr(Attribute::Cold) &&
2156       !OptSize && !llvm::shouldOptimizeForSize(BB, PSI, BFI.get()))
2157     for (auto &Arg : CI->args()) {
2158       if (!Arg->getType()->isPointerTy())
2159         continue;
2160       unsigned AS = Arg->getType()->getPointerAddressSpace();
2161       return optimizeMemoryInst(CI, Arg, Arg->getType(), AS);
2162     }
2163 
2164   IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
2165   if (II) {
2166     switch (II->getIntrinsicID()) {
2167     default: break;
2168     case Intrinsic::assume:
2169       llvm_unreachable("llvm.assume should have been removed already");
2170     case Intrinsic::experimental_widenable_condition: {
2171       // Give up on future widening oppurtunties so that we can fold away dead
2172       // paths and merge blocks before going into block-local instruction
2173       // selection.
2174       if (II->use_empty()) {
2175         II->eraseFromParent();
2176         return true;
2177       }
2178       Constant *RetVal = ConstantInt::getTrue(II->getContext());
2179       resetIteratorIfInvalidatedWhileCalling(BB, [&]() {
2180         replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
2181       });
2182       return true;
2183     }
2184     case Intrinsic::objectsize:
2185       llvm_unreachable("llvm.objectsize.* should have been lowered already");
2186     case Intrinsic::is_constant:
2187       llvm_unreachable("llvm.is.constant.* should have been lowered already");
2188     case Intrinsic::aarch64_stlxr:
2189     case Intrinsic::aarch64_stxr: {
2190       ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
2191       if (!ExtVal || !ExtVal->hasOneUse() ||
2192           ExtVal->getParent() == CI->getParent())
2193         return false;
2194       // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
2195       ExtVal->moveBefore(CI);
2196       // Mark this instruction as "inserted by CGP", so that other
2197       // optimizations don't touch it.
2198       InsertedInsts.insert(ExtVal);
2199       return true;
2200     }
2201 
2202     case Intrinsic::launder_invariant_group:
2203     case Intrinsic::strip_invariant_group: {
2204       Value *ArgVal = II->getArgOperand(0);
2205       auto it = LargeOffsetGEPMap.find(II);
2206       if (it != LargeOffsetGEPMap.end()) {
2207           // Merge entries in LargeOffsetGEPMap to reflect the RAUW.
2208           // Make sure not to have to deal with iterator invalidation
2209           // after possibly adding ArgVal to LargeOffsetGEPMap.
2210           auto GEPs = std::move(it->second);
2211           LargeOffsetGEPMap[ArgVal].append(GEPs.begin(), GEPs.end());
2212           LargeOffsetGEPMap.erase(II);
2213       }
2214 
2215       II->replaceAllUsesWith(ArgVal);
2216       II->eraseFromParent();
2217       return true;
2218     }
2219     case Intrinsic::cttz:
2220     case Intrinsic::ctlz:
2221       // If counting zeros is expensive, try to avoid it.
2222       return despeculateCountZeros(II, TLI, DL, ModifiedDT);
2223     case Intrinsic::fshl:
2224     case Intrinsic::fshr:
2225       return optimizeFunnelShift(II);
2226     case Intrinsic::dbg_value:
2227       return fixupDbgValue(II);
2228     case Intrinsic::vscale: {
2229       // If datalayout has no special restrictions on vector data layout,
2230       // replace `llvm.vscale` by an equivalent constant expression
2231       // to benefit from cheap constant propagation.
2232       Type *ScalableVectorTy =
2233           VectorType::get(Type::getInt8Ty(II->getContext()), 1, true);
2234       if (DL->getTypeAllocSize(ScalableVectorTy).getKnownMinSize() == 8) {
2235         auto *Null = Constant::getNullValue(ScalableVectorTy->getPointerTo());
2236         auto *One = ConstantInt::getSigned(II->getType(), 1);
2237         auto *CGep =
2238             ConstantExpr::getGetElementPtr(ScalableVectorTy, Null, One);
2239         II->replaceAllUsesWith(ConstantExpr::getPtrToInt(CGep, II->getType()));
2240         II->eraseFromParent();
2241         return true;
2242       }
2243       break;
2244     }
2245     case Intrinsic::masked_gather:
2246       return optimizeGatherScatterInst(II, II->getArgOperand(0));
2247     case Intrinsic::masked_scatter:
2248       return optimizeGatherScatterInst(II, II->getArgOperand(1));
2249     }
2250 
2251     SmallVector<Value *, 2> PtrOps;
2252     Type *AccessTy;
2253     if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
2254       while (!PtrOps.empty()) {
2255         Value *PtrVal = PtrOps.pop_back_val();
2256         unsigned AS = PtrVal->getType()->getPointerAddressSpace();
2257         if (optimizeMemoryInst(II, PtrVal, AccessTy, AS))
2258           return true;
2259       }
2260   }
2261 
2262   // From here on out we're working with named functions.
2263   if (!CI->getCalledFunction()) return false;
2264 
2265   // Lower all default uses of _chk calls.  This is very similar
2266   // to what InstCombineCalls does, but here we are only lowering calls
2267   // to fortified library functions (e.g. __memcpy_chk) that have the default
2268   // "don't know" as the objectsize.  Anything else should be left alone.
2269   FortifiedLibCallSimplifier Simplifier(TLInfo, true);
2270   IRBuilder<> Builder(CI);
2271   if (Value *V = Simplifier.optimizeCall(CI, Builder)) {
2272     CI->replaceAllUsesWith(V);
2273     CI->eraseFromParent();
2274     return true;
2275   }
2276 
2277   return false;
2278 }
2279 
2280 /// Look for opportunities to duplicate return instructions to the predecessor
2281 /// to enable tail call optimizations. The case it is currently looking for is:
2282 /// @code
2283 /// bb0:
2284 ///   %tmp0 = tail call i32 @f0()
2285 ///   br label %return
2286 /// bb1:
2287 ///   %tmp1 = tail call i32 @f1()
2288 ///   br label %return
2289 /// bb2:
2290 ///   %tmp2 = tail call i32 @f2()
2291 ///   br label %return
2292 /// return:
2293 ///   %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
2294 ///   ret i32 %retval
2295 /// @endcode
2296 ///
2297 /// =>
2298 ///
2299 /// @code
2300 /// bb0:
2301 ///   %tmp0 = tail call i32 @f0()
2302 ///   ret i32 %tmp0
2303 /// bb1:
2304 ///   %tmp1 = tail call i32 @f1()
2305 ///   ret i32 %tmp1
2306 /// bb2:
2307 ///   %tmp2 = tail call i32 @f2()
2308 ///   ret i32 %tmp2
2309 /// @endcode
2310 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB, bool &ModifiedDT) {
2311   ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator());
2312   if (!RetI)
2313     return false;
2314 
2315   PHINode *PN = nullptr;
2316   ExtractValueInst *EVI = nullptr;
2317   BitCastInst *BCI = nullptr;
2318   Value *V = RetI->getReturnValue();
2319   if (V) {
2320     BCI = dyn_cast<BitCastInst>(V);
2321     if (BCI)
2322       V = BCI->getOperand(0);
2323 
2324     EVI = dyn_cast<ExtractValueInst>(V);
2325     if (EVI) {
2326       V = EVI->getOperand(0);
2327       if (!llvm::all_of(EVI->indices(), [](unsigned idx) { return idx == 0; }))
2328         return false;
2329     }
2330 
2331     PN = dyn_cast<PHINode>(V);
2332     if (!PN)
2333       return false;
2334   }
2335 
2336   if (PN && PN->getParent() != BB)
2337     return false;
2338 
2339   auto isLifetimeEndOrBitCastFor = [](const Instruction *Inst) {
2340     const BitCastInst *BC = dyn_cast<BitCastInst>(Inst);
2341     if (BC && BC->hasOneUse())
2342       Inst = BC->user_back();
2343 
2344     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
2345       return II->getIntrinsicID() == Intrinsic::lifetime_end;
2346     return false;
2347   };
2348 
2349   // Make sure there are no instructions between the first instruction
2350   // and return.
2351   const Instruction *BI = BB->getFirstNonPHI();
2352   // Skip over debug and the bitcast.
2353   while (isa<DbgInfoIntrinsic>(BI) || BI == BCI || BI == EVI ||
2354          isa<PseudoProbeInst>(BI) || isLifetimeEndOrBitCastFor(BI))
2355     BI = BI->getNextNode();
2356   if (BI != RetI)
2357     return false;
2358 
2359   /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
2360   /// call.
2361   const Function *F = BB->getParent();
2362   SmallVector<BasicBlock*, 4> TailCallBBs;
2363   if (PN) {
2364     for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
2365       // Look through bitcasts.
2366       Value *IncomingVal = PN->getIncomingValue(I)->stripPointerCasts();
2367       CallInst *CI = dyn_cast<CallInst>(IncomingVal);
2368       BasicBlock *PredBB = PN->getIncomingBlock(I);
2369       // Make sure the phi value is indeed produced by the tail call.
2370       if (CI && CI->hasOneUse() && CI->getParent() == PredBB &&
2371           TLI->mayBeEmittedAsTailCall(CI) &&
2372           attributesPermitTailCall(F, CI, RetI, *TLI))
2373         TailCallBBs.push_back(PredBB);
2374     }
2375   } else {
2376     SmallPtrSet<BasicBlock*, 4> VisitedBBs;
2377     for (BasicBlock *Pred : predecessors(BB)) {
2378       if (!VisitedBBs.insert(Pred).second)
2379         continue;
2380       if (Instruction *I = Pred->rbegin()->getPrevNonDebugInstruction(true)) {
2381         CallInst *CI = dyn_cast<CallInst>(I);
2382         if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) &&
2383             attributesPermitTailCall(F, CI, RetI, *TLI))
2384           TailCallBBs.push_back(Pred);
2385       }
2386     }
2387   }
2388 
2389   bool Changed = false;
2390   for (auto const &TailCallBB : TailCallBBs) {
2391     // Make sure the call instruction is followed by an unconditional branch to
2392     // the return block.
2393     BranchInst *BI = dyn_cast<BranchInst>(TailCallBB->getTerminator());
2394     if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
2395       continue;
2396 
2397     // Duplicate the return into TailCallBB.
2398     (void)FoldReturnIntoUncondBranch(RetI, BB, TailCallBB);
2399     assert(!VerifyBFIUpdates ||
2400            BFI->getBlockFreq(BB) >= BFI->getBlockFreq(TailCallBB));
2401     BFI->setBlockFreq(
2402         BB,
2403         (BFI->getBlockFreq(BB) - BFI->getBlockFreq(TailCallBB)).getFrequency());
2404     ModifiedDT = Changed = true;
2405     ++NumRetsDup;
2406   }
2407 
2408   // If we eliminated all predecessors of the block, delete the block now.
2409   if (Changed && !BB->hasAddressTaken() && pred_empty(BB))
2410     BB->eraseFromParent();
2411 
2412   return Changed;
2413 }
2414 
2415 //===----------------------------------------------------------------------===//
2416 // Memory Optimization
2417 //===----------------------------------------------------------------------===//
2418 
2419 namespace {
2420 
2421 /// This is an extended version of TargetLowering::AddrMode
2422 /// which holds actual Value*'s for register values.
2423 struct ExtAddrMode : public TargetLowering::AddrMode {
2424   Value *BaseReg = nullptr;
2425   Value *ScaledReg = nullptr;
2426   Value *OriginalValue = nullptr;
2427   bool InBounds = true;
2428 
2429   enum FieldName {
2430     NoField        = 0x00,
2431     BaseRegField   = 0x01,
2432     BaseGVField    = 0x02,
2433     BaseOffsField  = 0x04,
2434     ScaledRegField = 0x08,
2435     ScaleField     = 0x10,
2436     MultipleFields = 0xff
2437   };
2438 
2439 
2440   ExtAddrMode() = default;
2441 
2442   void print(raw_ostream &OS) const;
2443   void dump() const;
2444 
2445   FieldName compare(const ExtAddrMode &other) {
2446     // First check that the types are the same on each field, as differing types
2447     // is something we can't cope with later on.
2448     if (BaseReg && other.BaseReg &&
2449         BaseReg->getType() != other.BaseReg->getType())
2450       return MultipleFields;
2451     if (BaseGV && other.BaseGV &&
2452         BaseGV->getType() != other.BaseGV->getType())
2453       return MultipleFields;
2454     if (ScaledReg && other.ScaledReg &&
2455         ScaledReg->getType() != other.ScaledReg->getType())
2456       return MultipleFields;
2457 
2458     // Conservatively reject 'inbounds' mismatches.
2459     if (InBounds != other.InBounds)
2460       return MultipleFields;
2461 
2462     // Check each field to see if it differs.
2463     unsigned Result = NoField;
2464     if (BaseReg != other.BaseReg)
2465       Result |= BaseRegField;
2466     if (BaseGV != other.BaseGV)
2467       Result |= BaseGVField;
2468     if (BaseOffs != other.BaseOffs)
2469       Result |= BaseOffsField;
2470     if (ScaledReg != other.ScaledReg)
2471       Result |= ScaledRegField;
2472     // Don't count 0 as being a different scale, because that actually means
2473     // unscaled (which will already be counted by having no ScaledReg).
2474     if (Scale && other.Scale && Scale != other.Scale)
2475       Result |= ScaleField;
2476 
2477     if (countPopulation(Result) > 1)
2478       return MultipleFields;
2479     else
2480       return static_cast<FieldName>(Result);
2481   }
2482 
2483   // An AddrMode is trivial if it involves no calculation i.e. it is just a base
2484   // with no offset.
2485   bool isTrivial() {
2486     // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
2487     // trivial if at most one of these terms is nonzero, except that BaseGV and
2488     // BaseReg both being zero actually means a null pointer value, which we
2489     // consider to be 'non-zero' here.
2490     return !BaseOffs && !Scale && !(BaseGV && BaseReg);
2491   }
2492 
2493   Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
2494     switch (Field) {
2495     default:
2496       return nullptr;
2497     case BaseRegField:
2498       return BaseReg;
2499     case BaseGVField:
2500       return BaseGV;
2501     case ScaledRegField:
2502       return ScaledReg;
2503     case BaseOffsField:
2504       return ConstantInt::get(IntPtrTy, BaseOffs);
2505     }
2506   }
2507 
2508   void SetCombinedField(FieldName Field, Value *V,
2509                         const SmallVectorImpl<ExtAddrMode> &AddrModes) {
2510     switch (Field) {
2511     default:
2512       llvm_unreachable("Unhandled fields are expected to be rejected earlier");
2513       break;
2514     case ExtAddrMode::BaseRegField:
2515       BaseReg = V;
2516       break;
2517     case ExtAddrMode::BaseGVField:
2518       // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
2519       // in the BaseReg field.
2520       assert(BaseReg == nullptr);
2521       BaseReg = V;
2522       BaseGV = nullptr;
2523       break;
2524     case ExtAddrMode::ScaledRegField:
2525       ScaledReg = V;
2526       // If we have a mix of scaled and unscaled addrmodes then we want scale
2527       // to be the scale and not zero.
2528       if (!Scale)
2529         for (const ExtAddrMode &AM : AddrModes)
2530           if (AM.Scale) {
2531             Scale = AM.Scale;
2532             break;
2533           }
2534       break;
2535     case ExtAddrMode::BaseOffsField:
2536       // The offset is no longer a constant, so it goes in ScaledReg with a
2537       // scale of 1.
2538       assert(ScaledReg == nullptr);
2539       ScaledReg = V;
2540       Scale = 1;
2541       BaseOffs = 0;
2542       break;
2543     }
2544   }
2545 };
2546 
2547 } // end anonymous namespace
2548 
2549 #ifndef NDEBUG
2550 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2551   AM.print(OS);
2552   return OS;
2553 }
2554 #endif
2555 
2556 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2557 void ExtAddrMode::print(raw_ostream &OS) const {
2558   bool NeedPlus = false;
2559   OS << "[";
2560   if (InBounds)
2561     OS << "inbounds ";
2562   if (BaseGV) {
2563     OS << (NeedPlus ? " + " : "")
2564        << "GV:";
2565     BaseGV->printAsOperand(OS, /*PrintType=*/false);
2566     NeedPlus = true;
2567   }
2568 
2569   if (BaseOffs) {
2570     OS << (NeedPlus ? " + " : "")
2571        << BaseOffs;
2572     NeedPlus = true;
2573   }
2574 
2575   if (BaseReg) {
2576     OS << (NeedPlus ? " + " : "")
2577        << "Base:";
2578     BaseReg->printAsOperand(OS, /*PrintType=*/false);
2579     NeedPlus = true;
2580   }
2581   if (Scale) {
2582     OS << (NeedPlus ? " + " : "")
2583        << Scale << "*";
2584     ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2585   }
2586 
2587   OS << ']';
2588 }
2589 
2590 LLVM_DUMP_METHOD void ExtAddrMode::dump() const {
2591   print(dbgs());
2592   dbgs() << '\n';
2593 }
2594 #endif
2595 
2596 namespace {
2597 
2598 /// This class provides transaction based operation on the IR.
2599 /// Every change made through this class is recorded in the internal state and
2600 /// can be undone (rollback) until commit is called.
2601 /// CGP does not check if instructions could be speculatively executed when
2602 /// moved. Preserving the original location would pessimize the debugging
2603 /// experience, as well as negatively impact the quality of sample PGO.
2604 class TypePromotionTransaction {
2605   /// This represents the common interface of the individual transaction.
2606   /// Each class implements the logic for doing one specific modification on
2607   /// the IR via the TypePromotionTransaction.
2608   class TypePromotionAction {
2609   protected:
2610     /// The Instruction modified.
2611     Instruction *Inst;
2612 
2613   public:
2614     /// Constructor of the action.
2615     /// The constructor performs the related action on the IR.
2616     TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2617 
2618     virtual ~TypePromotionAction() = default;
2619 
2620     /// Undo the modification done by this action.
2621     /// When this method is called, the IR must be in the same state as it was
2622     /// before this action was applied.
2623     /// \pre Undoing the action works if and only if the IR is in the exact same
2624     /// state as it was directly after this action was applied.
2625     virtual void undo() = 0;
2626 
2627     /// Advocate every change made by this action.
2628     /// When the results on the IR of the action are to be kept, it is important
2629     /// to call this function, otherwise hidden information may be kept forever.
2630     virtual void commit() {
2631       // Nothing to be done, this action is not doing anything.
2632     }
2633   };
2634 
2635   /// Utility to remember the position of an instruction.
2636   class InsertionHandler {
2637     /// Position of an instruction.
2638     /// Either an instruction:
2639     /// - Is the first in a basic block: BB is used.
2640     /// - Has a previous instruction: PrevInst is used.
2641     union {
2642       Instruction *PrevInst;
2643       BasicBlock *BB;
2644     } Point;
2645 
2646     /// Remember whether or not the instruction had a previous instruction.
2647     bool HasPrevInstruction;
2648 
2649   public:
2650     /// Record the position of \p Inst.
2651     InsertionHandler(Instruction *Inst) {
2652       BasicBlock::iterator It = Inst->getIterator();
2653       HasPrevInstruction = (It != (Inst->getParent()->begin()));
2654       if (HasPrevInstruction)
2655         Point.PrevInst = &*--It;
2656       else
2657         Point.BB = Inst->getParent();
2658     }
2659 
2660     /// Insert \p Inst at the recorded position.
2661     void insert(Instruction *Inst) {
2662       if (HasPrevInstruction) {
2663         if (Inst->getParent())
2664           Inst->removeFromParent();
2665         Inst->insertAfter(Point.PrevInst);
2666       } else {
2667         Instruction *Position = &*Point.BB->getFirstInsertionPt();
2668         if (Inst->getParent())
2669           Inst->moveBefore(Position);
2670         else
2671           Inst->insertBefore(Position);
2672       }
2673     }
2674   };
2675 
2676   /// Move an instruction before another.
2677   class InstructionMoveBefore : public TypePromotionAction {
2678     /// Original position of the instruction.
2679     InsertionHandler Position;
2680 
2681   public:
2682     /// Move \p Inst before \p Before.
2683     InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2684         : TypePromotionAction(Inst), Position(Inst) {
2685       LLVM_DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before
2686                         << "\n");
2687       Inst->moveBefore(Before);
2688     }
2689 
2690     /// Move the instruction back to its original position.
2691     void undo() override {
2692       LLVM_DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2693       Position.insert(Inst);
2694     }
2695   };
2696 
2697   /// Set the operand of an instruction with a new value.
2698   class OperandSetter : public TypePromotionAction {
2699     /// Original operand of the instruction.
2700     Value *Origin;
2701 
2702     /// Index of the modified instruction.
2703     unsigned Idx;
2704 
2705   public:
2706     /// Set \p Idx operand of \p Inst with \p NewVal.
2707     OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2708         : TypePromotionAction(Inst), Idx(Idx) {
2709       LLVM_DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2710                         << "for:" << *Inst << "\n"
2711                         << "with:" << *NewVal << "\n");
2712       Origin = Inst->getOperand(Idx);
2713       Inst->setOperand(Idx, NewVal);
2714     }
2715 
2716     /// Restore the original value of the instruction.
2717     void undo() override {
2718       LLVM_DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2719                         << "for: " << *Inst << "\n"
2720                         << "with: " << *Origin << "\n");
2721       Inst->setOperand(Idx, Origin);
2722     }
2723   };
2724 
2725   /// Hide the operands of an instruction.
2726   /// Do as if this instruction was not using any of its operands.
2727   class OperandsHider : public TypePromotionAction {
2728     /// The list of original operands.
2729     SmallVector<Value *, 4> OriginalValues;
2730 
2731   public:
2732     /// Remove \p Inst from the uses of the operands of \p Inst.
2733     OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2734       LLVM_DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2735       unsigned NumOpnds = Inst->getNumOperands();
2736       OriginalValues.reserve(NumOpnds);
2737       for (unsigned It = 0; It < NumOpnds; ++It) {
2738         // Save the current operand.
2739         Value *Val = Inst->getOperand(It);
2740         OriginalValues.push_back(Val);
2741         // Set a dummy one.
2742         // We could use OperandSetter here, but that would imply an overhead
2743         // that we are not willing to pay.
2744         Inst->setOperand(It, UndefValue::get(Val->getType()));
2745       }
2746     }
2747 
2748     /// Restore the original list of uses.
2749     void undo() override {
2750       LLVM_DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2751       for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2752         Inst->setOperand(It, OriginalValues[It]);
2753     }
2754   };
2755 
2756   /// Build a truncate instruction.
2757   class TruncBuilder : public TypePromotionAction {
2758     Value *Val;
2759 
2760   public:
2761     /// Build a truncate instruction of \p Opnd producing a \p Ty
2762     /// result.
2763     /// trunc Opnd to Ty.
2764     TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2765       IRBuilder<> Builder(Opnd);
2766       Builder.SetCurrentDebugLocation(DebugLoc());
2767       Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2768       LLVM_DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2769     }
2770 
2771     /// Get the built value.
2772     Value *getBuiltValue() { return Val; }
2773 
2774     /// Remove the built instruction.
2775     void undo() override {
2776       LLVM_DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2777       if (Instruction *IVal = dyn_cast<Instruction>(Val))
2778         IVal->eraseFromParent();
2779     }
2780   };
2781 
2782   /// Build a sign extension instruction.
2783   class SExtBuilder : public TypePromotionAction {
2784     Value *Val;
2785 
2786   public:
2787     /// Build a sign extension instruction of \p Opnd producing a \p Ty
2788     /// result.
2789     /// sext Opnd to Ty.
2790     SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2791         : TypePromotionAction(InsertPt) {
2792       IRBuilder<> Builder(InsertPt);
2793       Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2794       LLVM_DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2795     }
2796 
2797     /// Get the built value.
2798     Value *getBuiltValue() { return Val; }
2799 
2800     /// Remove the built instruction.
2801     void undo() override {
2802       LLVM_DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2803       if (Instruction *IVal = dyn_cast<Instruction>(Val))
2804         IVal->eraseFromParent();
2805     }
2806   };
2807 
2808   /// Build a zero extension instruction.
2809   class ZExtBuilder : public TypePromotionAction {
2810     Value *Val;
2811 
2812   public:
2813     /// Build a zero extension instruction of \p Opnd producing a \p Ty
2814     /// result.
2815     /// zext Opnd to Ty.
2816     ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2817         : TypePromotionAction(InsertPt) {
2818       IRBuilder<> Builder(InsertPt);
2819       Builder.SetCurrentDebugLocation(DebugLoc());
2820       Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2821       LLVM_DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2822     }
2823 
2824     /// Get the built value.
2825     Value *getBuiltValue() { return Val; }
2826 
2827     /// Remove the built instruction.
2828     void undo() override {
2829       LLVM_DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2830       if (Instruction *IVal = dyn_cast<Instruction>(Val))
2831         IVal->eraseFromParent();
2832     }
2833   };
2834 
2835   /// Mutate an instruction to another type.
2836   class TypeMutator : public TypePromotionAction {
2837     /// Record the original type.
2838     Type *OrigTy;
2839 
2840   public:
2841     /// Mutate the type of \p Inst into \p NewTy.
2842     TypeMutator(Instruction *Inst, Type *NewTy)
2843         : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2844       LLVM_DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2845                         << "\n");
2846       Inst->mutateType(NewTy);
2847     }
2848 
2849     /// Mutate the instruction back to its original type.
2850     void undo() override {
2851       LLVM_DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2852                         << "\n");
2853       Inst->mutateType(OrigTy);
2854     }
2855   };
2856 
2857   /// Replace the uses of an instruction by another instruction.
2858   class UsesReplacer : public TypePromotionAction {
2859     /// Helper structure to keep track of the replaced uses.
2860     struct InstructionAndIdx {
2861       /// The instruction using the instruction.
2862       Instruction *Inst;
2863 
2864       /// The index where this instruction is used for Inst.
2865       unsigned Idx;
2866 
2867       InstructionAndIdx(Instruction *Inst, unsigned Idx)
2868           : Inst(Inst), Idx(Idx) {}
2869     };
2870 
2871     /// Keep track of the original uses (pair Instruction, Index).
2872     SmallVector<InstructionAndIdx, 4> OriginalUses;
2873     /// Keep track of the debug users.
2874     SmallVector<DbgValueInst *, 1> DbgValues;
2875 
2876     /// Keep track of the new value so that we can undo it by replacing
2877     /// instances of the new value with the original value.
2878     Value *New;
2879 
2880     using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
2881 
2882   public:
2883     /// Replace all the use of \p Inst by \p New.
2884     UsesReplacer(Instruction *Inst, Value *New)
2885         : TypePromotionAction(Inst), New(New) {
2886       LLVM_DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2887                         << "\n");
2888       // Record the original uses.
2889       for (Use &U : Inst->uses()) {
2890         Instruction *UserI = cast<Instruction>(U.getUser());
2891         OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2892       }
2893       // Record the debug uses separately. They are not in the instruction's
2894       // use list, but they are replaced by RAUW.
2895       findDbgValues(DbgValues, Inst);
2896 
2897       // Now, we can replace the uses.
2898       Inst->replaceAllUsesWith(New);
2899     }
2900 
2901     /// Reassign the original uses of Inst to Inst.
2902     void undo() override {
2903       LLVM_DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2904       for (InstructionAndIdx &Use : OriginalUses)
2905         Use.Inst->setOperand(Use.Idx, Inst);
2906       // RAUW has replaced all original uses with references to the new value,
2907       // including the debug uses. Since we are undoing the replacements,
2908       // the original debug uses must also be reinstated to maintain the
2909       // correctness and utility of debug value instructions.
2910       for (auto *DVI : DbgValues)
2911         DVI->replaceVariableLocationOp(New, Inst);
2912     }
2913   };
2914 
2915   /// Remove an instruction from the IR.
2916   class InstructionRemover : public TypePromotionAction {
2917     /// Original position of the instruction.
2918     InsertionHandler Inserter;
2919 
2920     /// Helper structure to hide all the link to the instruction. In other
2921     /// words, this helps to do as if the instruction was removed.
2922     OperandsHider Hider;
2923 
2924     /// Keep track of the uses replaced, if any.
2925     UsesReplacer *Replacer = nullptr;
2926 
2927     /// Keep track of instructions removed.
2928     SetOfInstrs &RemovedInsts;
2929 
2930   public:
2931     /// Remove all reference of \p Inst and optionally replace all its
2932     /// uses with New.
2933     /// \p RemovedInsts Keep track of the instructions removed by this Action.
2934     /// \pre If !Inst->use_empty(), then New != nullptr
2935     InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
2936                        Value *New = nullptr)
2937         : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2938           RemovedInsts(RemovedInsts) {
2939       if (New)
2940         Replacer = new UsesReplacer(Inst, New);
2941       LLVM_DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2942       RemovedInsts.insert(Inst);
2943       /// The instructions removed here will be freed after completing
2944       /// optimizeBlock() for all blocks as we need to keep track of the
2945       /// removed instructions during promotion.
2946       Inst->removeFromParent();
2947     }
2948 
2949     ~InstructionRemover() override { delete Replacer; }
2950 
2951     /// Resurrect the instruction and reassign it to the proper uses if
2952     /// new value was provided when build this action.
2953     void undo() override {
2954       LLVM_DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2955       Inserter.insert(Inst);
2956       if (Replacer)
2957         Replacer->undo();
2958       Hider.undo();
2959       RemovedInsts.erase(Inst);
2960     }
2961   };
2962 
2963 public:
2964   /// Restoration point.
2965   /// The restoration point is a pointer to an action instead of an iterator
2966   /// because the iterator may be invalidated but not the pointer.
2967   using ConstRestorationPt = const TypePromotionAction *;
2968 
2969   TypePromotionTransaction(SetOfInstrs &RemovedInsts)
2970       : RemovedInsts(RemovedInsts) {}
2971 
2972   /// Advocate every changes made in that transaction. Return true if any change
2973   /// happen.
2974   bool commit();
2975 
2976   /// Undo all the changes made after the given point.
2977   void rollback(ConstRestorationPt Point);
2978 
2979   /// Get the current restoration point.
2980   ConstRestorationPt getRestorationPoint() const;
2981 
2982   /// \name API for IR modification with state keeping to support rollback.
2983   /// @{
2984   /// Same as Instruction::setOperand.
2985   void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2986 
2987   /// Same as Instruction::eraseFromParent.
2988   void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2989 
2990   /// Same as Value::replaceAllUsesWith.
2991   void replaceAllUsesWith(Instruction *Inst, Value *New);
2992 
2993   /// Same as Value::mutateType.
2994   void mutateType(Instruction *Inst, Type *NewTy);
2995 
2996   /// Same as IRBuilder::createTrunc.
2997   Value *createTrunc(Instruction *Opnd, Type *Ty);
2998 
2999   /// Same as IRBuilder::createSExt.
3000   Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
3001 
3002   /// Same as IRBuilder::createZExt.
3003   Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
3004 
3005   /// Same as Instruction::moveBefore.
3006   void moveBefore(Instruction *Inst, Instruction *Before);
3007   /// @}
3008 
3009 private:
3010   /// The ordered list of actions made so far.
3011   SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
3012 
3013   using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
3014 
3015   SetOfInstrs &RemovedInsts;
3016 };
3017 
3018 } // end anonymous namespace
3019 
3020 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
3021                                           Value *NewVal) {
3022   Actions.push_back(std::make_unique<TypePromotionTransaction::OperandSetter>(
3023       Inst, Idx, NewVal));
3024 }
3025 
3026 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
3027                                                 Value *NewVal) {
3028   Actions.push_back(
3029       std::make_unique<TypePromotionTransaction::InstructionRemover>(
3030           Inst, RemovedInsts, NewVal));
3031 }
3032 
3033 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
3034                                                   Value *New) {
3035   Actions.push_back(
3036       std::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
3037 }
3038 
3039 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
3040   Actions.push_back(
3041       std::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
3042 }
3043 
3044 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
3045                                              Type *Ty) {
3046   std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
3047   Value *Val = Ptr->getBuiltValue();
3048   Actions.push_back(std::move(Ptr));
3049   return Val;
3050 }
3051 
3052 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
3053                                             Value *Opnd, Type *Ty) {
3054   std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
3055   Value *Val = Ptr->getBuiltValue();
3056   Actions.push_back(std::move(Ptr));
3057   return Val;
3058 }
3059 
3060 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
3061                                             Value *Opnd, Type *Ty) {
3062   std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
3063   Value *Val = Ptr->getBuiltValue();
3064   Actions.push_back(std::move(Ptr));
3065   return Val;
3066 }
3067 
3068 void TypePromotionTransaction::moveBefore(Instruction *Inst,
3069                                           Instruction *Before) {
3070   Actions.push_back(
3071       std::make_unique<TypePromotionTransaction::InstructionMoveBefore>(
3072           Inst, Before));
3073 }
3074 
3075 TypePromotionTransaction::ConstRestorationPt
3076 TypePromotionTransaction::getRestorationPoint() const {
3077   return !Actions.empty() ? Actions.back().get() : nullptr;
3078 }
3079 
3080 bool TypePromotionTransaction::commit() {
3081   for (std::unique_ptr<TypePromotionAction> &Action : Actions)
3082     Action->commit();
3083   bool Modified = !Actions.empty();
3084   Actions.clear();
3085   return Modified;
3086 }
3087 
3088 void TypePromotionTransaction::rollback(
3089     TypePromotionTransaction::ConstRestorationPt Point) {
3090   while (!Actions.empty() && Point != Actions.back().get()) {
3091     std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
3092     Curr->undo();
3093   }
3094 }
3095 
3096 namespace {
3097 
3098 /// A helper class for matching addressing modes.
3099 ///
3100 /// This encapsulates the logic for matching the target-legal addressing modes.
3101 class AddressingModeMatcher {
3102   SmallVectorImpl<Instruction*> &AddrModeInsts;
3103   const TargetLowering &TLI;
3104   const TargetRegisterInfo &TRI;
3105   const DataLayout &DL;
3106   const LoopInfo &LI;
3107   const std::function<const DominatorTree &()> getDTFn;
3108 
3109   /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
3110   /// the memory instruction that we're computing this address for.
3111   Type *AccessTy;
3112   unsigned AddrSpace;
3113   Instruction *MemoryInst;
3114 
3115   /// This is the addressing mode that we're building up. This is
3116   /// part of the return value of this addressing mode matching stuff.
3117   ExtAddrMode &AddrMode;
3118 
3119   /// The instructions inserted by other CodeGenPrepare optimizations.
3120   const SetOfInstrs &InsertedInsts;
3121 
3122   /// A map from the instructions to their type before promotion.
3123   InstrToOrigTy &PromotedInsts;
3124 
3125   /// The ongoing transaction where every action should be registered.
3126   TypePromotionTransaction &TPT;
3127 
3128   // A GEP which has too large offset to be folded into the addressing mode.
3129   std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP;
3130 
3131   /// This is set to true when we should not do profitability checks.
3132   /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
3133   bool IgnoreProfitability;
3134 
3135   /// True if we are optimizing for size.
3136   bool OptSize;
3137 
3138   ProfileSummaryInfo *PSI;
3139   BlockFrequencyInfo *BFI;
3140 
3141   AddressingModeMatcher(
3142       SmallVectorImpl<Instruction *> &AMI, const TargetLowering &TLI,
3143       const TargetRegisterInfo &TRI, const LoopInfo &LI,
3144       const std::function<const DominatorTree &()> getDTFn,
3145       Type *AT, unsigned AS, Instruction *MI, ExtAddrMode &AM,
3146       const SetOfInstrs &InsertedInsts, InstrToOrigTy &PromotedInsts,
3147       TypePromotionTransaction &TPT,
3148       std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP,
3149       bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI)
3150       : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
3151         DL(MI->getModule()->getDataLayout()), LI(LI), getDTFn(getDTFn),
3152         AccessTy(AT), AddrSpace(AS), MemoryInst(MI), AddrMode(AM),
3153         InsertedInsts(InsertedInsts), PromotedInsts(PromotedInsts), TPT(TPT),
3154         LargeOffsetGEP(LargeOffsetGEP), OptSize(OptSize), PSI(PSI), BFI(BFI) {
3155     IgnoreProfitability = false;
3156   }
3157 
3158 public:
3159   /// Find the maximal addressing mode that a load/store of V can fold,
3160   /// give an access type of AccessTy.  This returns a list of involved
3161   /// instructions in AddrModeInsts.
3162   /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
3163   /// optimizations.
3164   /// \p PromotedInsts maps the instructions to their type before promotion.
3165   /// \p The ongoing transaction where every action should be registered.
3166   static ExtAddrMode
3167   Match(Value *V, Type *AccessTy, unsigned AS, Instruction *MemoryInst,
3168         SmallVectorImpl<Instruction *> &AddrModeInsts,
3169         const TargetLowering &TLI, const LoopInfo &LI,
3170         const std::function<const DominatorTree &()> getDTFn,
3171         const TargetRegisterInfo &TRI, const SetOfInstrs &InsertedInsts,
3172         InstrToOrigTy &PromotedInsts, TypePromotionTransaction &TPT,
3173         std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP,
3174         bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) {
3175     ExtAddrMode Result;
3176 
3177     bool Success = AddressingModeMatcher(
3178         AddrModeInsts, TLI, TRI, LI, getDTFn, AccessTy, AS, MemoryInst, Result,
3179         InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP, OptSize, PSI,
3180         BFI).matchAddr(V, 0);
3181     (void)Success; assert(Success && "Couldn't select *anything*?");
3182     return Result;
3183   }
3184 
3185 private:
3186   bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
3187   bool matchAddr(Value *Addr, unsigned Depth);
3188   bool matchOperationAddr(User *AddrInst, unsigned Opcode, unsigned Depth,
3189                           bool *MovedAway = nullptr);
3190   bool isProfitableToFoldIntoAddressingMode(Instruction *I,
3191                                             ExtAddrMode &AMBefore,
3192                                             ExtAddrMode &AMAfter);
3193   bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
3194   bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
3195                              Value *PromotedOperand) const;
3196 };
3197 
3198 class PhiNodeSet;
3199 
3200 /// An iterator for PhiNodeSet.
3201 class PhiNodeSetIterator {
3202   PhiNodeSet * const Set;
3203   size_t CurrentIndex = 0;
3204 
3205 public:
3206   /// The constructor. Start should point to either a valid element, or be equal
3207   /// to the size of the underlying SmallVector of the PhiNodeSet.
3208   PhiNodeSetIterator(PhiNodeSet * const Set, size_t Start);
3209   PHINode * operator*() const;
3210   PhiNodeSetIterator& operator++();
3211   bool operator==(const PhiNodeSetIterator &RHS) const;
3212   bool operator!=(const PhiNodeSetIterator &RHS) const;
3213 };
3214 
3215 /// Keeps a set of PHINodes.
3216 ///
3217 /// This is a minimal set implementation for a specific use case:
3218 /// It is very fast when there are very few elements, but also provides good
3219 /// performance when there are many. It is similar to SmallPtrSet, but also
3220 /// provides iteration by insertion order, which is deterministic and stable
3221 /// across runs. It is also similar to SmallSetVector, but provides removing
3222 /// elements in O(1) time. This is achieved by not actually removing the element
3223 /// from the underlying vector, so comes at the cost of using more memory, but
3224 /// that is fine, since PhiNodeSets are used as short lived objects.
3225 class PhiNodeSet {
3226   friend class PhiNodeSetIterator;
3227 
3228   using MapType = SmallDenseMap<PHINode *, size_t, 32>;
3229   using iterator =  PhiNodeSetIterator;
3230 
3231   /// Keeps the elements in the order of their insertion in the underlying
3232   /// vector. To achieve constant time removal, it never deletes any element.
3233   SmallVector<PHINode *, 32> NodeList;
3234 
3235   /// Keeps the elements in the underlying set implementation. This (and not the
3236   /// NodeList defined above) is the source of truth on whether an element
3237   /// is actually in the collection.
3238   MapType NodeMap;
3239 
3240   /// Points to the first valid (not deleted) element when the set is not empty
3241   /// and the value is not zero. Equals to the size of the underlying vector
3242   /// when the set is empty. When the value is 0, as in the beginning, the
3243   /// first element may or may not be valid.
3244   size_t FirstValidElement = 0;
3245 
3246 public:
3247   /// Inserts a new element to the collection.
3248   /// \returns true if the element is actually added, i.e. was not in the
3249   /// collection before the operation.
3250   bool insert(PHINode *Ptr) {
3251     if (NodeMap.insert(std::make_pair(Ptr, NodeList.size())).second) {
3252       NodeList.push_back(Ptr);
3253       return true;
3254     }
3255     return false;
3256   }
3257 
3258   /// Removes the element from the collection.
3259   /// \returns whether the element is actually removed, i.e. was in the
3260   /// collection before the operation.
3261   bool erase(PHINode *Ptr) {
3262     if (NodeMap.erase(Ptr)) {
3263       SkipRemovedElements(FirstValidElement);
3264       return true;
3265     }
3266     return false;
3267   }
3268 
3269   /// Removes all elements and clears the collection.
3270   void clear() {
3271     NodeMap.clear();
3272     NodeList.clear();
3273     FirstValidElement = 0;
3274   }
3275 
3276   /// \returns an iterator that will iterate the elements in the order of
3277   /// insertion.
3278   iterator begin() {
3279     if (FirstValidElement == 0)
3280       SkipRemovedElements(FirstValidElement);
3281     return PhiNodeSetIterator(this, FirstValidElement);
3282   }
3283 
3284   /// \returns an iterator that points to the end of the collection.
3285   iterator end() { return PhiNodeSetIterator(this, NodeList.size()); }
3286 
3287   /// Returns the number of elements in the collection.
3288   size_t size() const {
3289     return NodeMap.size();
3290   }
3291 
3292   /// \returns 1 if the given element is in the collection, and 0 if otherwise.
3293   size_t count(PHINode *Ptr) const {
3294     return NodeMap.count(Ptr);
3295   }
3296 
3297 private:
3298   /// Updates the CurrentIndex so that it will point to a valid element.
3299   ///
3300   /// If the element of NodeList at CurrentIndex is valid, it does not
3301   /// change it. If there are no more valid elements, it updates CurrentIndex
3302   /// to point to the end of the NodeList.
3303   void SkipRemovedElements(size_t &CurrentIndex) {
3304     while (CurrentIndex < NodeList.size()) {
3305       auto it = NodeMap.find(NodeList[CurrentIndex]);
3306       // If the element has been deleted and added again later, NodeMap will
3307       // point to a different index, so CurrentIndex will still be invalid.
3308       if (it != NodeMap.end() && it->second == CurrentIndex)
3309         break;
3310       ++CurrentIndex;
3311     }
3312   }
3313 };
3314 
3315 PhiNodeSetIterator::PhiNodeSetIterator(PhiNodeSet *const Set, size_t Start)
3316     : Set(Set), CurrentIndex(Start) {}
3317 
3318 PHINode * PhiNodeSetIterator::operator*() const {
3319   assert(CurrentIndex < Set->NodeList.size() &&
3320          "PhiNodeSet access out of range");
3321   return Set->NodeList[CurrentIndex];
3322 }
3323 
3324 PhiNodeSetIterator& PhiNodeSetIterator::operator++() {
3325   assert(CurrentIndex < Set->NodeList.size() &&
3326          "PhiNodeSet access out of range");
3327   ++CurrentIndex;
3328   Set->SkipRemovedElements(CurrentIndex);
3329   return *this;
3330 }
3331 
3332 bool PhiNodeSetIterator::operator==(const PhiNodeSetIterator &RHS) const {
3333   return CurrentIndex == RHS.CurrentIndex;
3334 }
3335 
3336 bool PhiNodeSetIterator::operator!=(const PhiNodeSetIterator &RHS) const {
3337   return !((*this) == RHS);
3338 }
3339 
3340 /// Keep track of simplification of Phi nodes.
3341 /// Accept the set of all phi nodes and erase phi node from this set
3342 /// if it is simplified.
3343 class SimplificationTracker {
3344   DenseMap<Value *, Value *> Storage;
3345   const SimplifyQuery &SQ;
3346   // Tracks newly created Phi nodes. The elements are iterated by insertion
3347   // order.
3348   PhiNodeSet AllPhiNodes;
3349   // Tracks newly created Select nodes.
3350   SmallPtrSet<SelectInst *, 32> AllSelectNodes;
3351 
3352 public:
3353   SimplificationTracker(const SimplifyQuery &sq)
3354       : SQ(sq) {}
3355 
3356   Value *Get(Value *V) {
3357     do {
3358       auto SV = Storage.find(V);
3359       if (SV == Storage.end())
3360         return V;
3361       V = SV->second;
3362     } while (true);
3363   }
3364 
3365   Value *Simplify(Value *Val) {
3366     SmallVector<Value *, 32> WorkList;
3367     SmallPtrSet<Value *, 32> Visited;
3368     WorkList.push_back(Val);
3369     while (!WorkList.empty()) {
3370       auto *P = WorkList.pop_back_val();
3371       if (!Visited.insert(P).second)
3372         continue;
3373       if (auto *PI = dyn_cast<Instruction>(P))
3374         if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) {
3375           for (auto *U : PI->users())
3376             WorkList.push_back(cast<Value>(U));
3377           Put(PI, V);
3378           PI->replaceAllUsesWith(V);
3379           if (auto *PHI = dyn_cast<PHINode>(PI))
3380             AllPhiNodes.erase(PHI);
3381           if (auto *Select = dyn_cast<SelectInst>(PI))
3382             AllSelectNodes.erase(Select);
3383           PI->eraseFromParent();
3384         }
3385     }
3386     return Get(Val);
3387   }
3388 
3389   void Put(Value *From, Value *To) {
3390     Storage.insert({ From, To });
3391   }
3392 
3393   void ReplacePhi(PHINode *From, PHINode *To) {
3394     Value* OldReplacement = Get(From);
3395     while (OldReplacement != From) {
3396       From = To;
3397       To = dyn_cast<PHINode>(OldReplacement);
3398       OldReplacement = Get(From);
3399     }
3400     assert(To && Get(To) == To && "Replacement PHI node is already replaced.");
3401     Put(From, To);
3402     From->replaceAllUsesWith(To);
3403     AllPhiNodes.erase(From);
3404     From->eraseFromParent();
3405   }
3406 
3407   PhiNodeSet& newPhiNodes() { return AllPhiNodes; }
3408 
3409   void insertNewPhi(PHINode *PN) { AllPhiNodes.insert(PN); }
3410 
3411   void insertNewSelect(SelectInst *SI) { AllSelectNodes.insert(SI); }
3412 
3413   unsigned countNewPhiNodes() const { return AllPhiNodes.size(); }
3414 
3415   unsigned countNewSelectNodes() const { return AllSelectNodes.size(); }
3416 
3417   void destroyNewNodes(Type *CommonType) {
3418     // For safe erasing, replace the uses with dummy value first.
3419     auto *Dummy = UndefValue::get(CommonType);
3420     for (auto *I : AllPhiNodes) {
3421       I->replaceAllUsesWith(Dummy);
3422       I->eraseFromParent();
3423     }
3424     AllPhiNodes.clear();
3425     for (auto *I : AllSelectNodes) {
3426       I->replaceAllUsesWith(Dummy);
3427       I->eraseFromParent();
3428     }
3429     AllSelectNodes.clear();
3430   }
3431 };
3432 
3433 /// A helper class for combining addressing modes.
3434 class AddressingModeCombiner {
3435   typedef DenseMap<Value *, Value *> FoldAddrToValueMapping;
3436   typedef std::pair<PHINode *, PHINode *> PHIPair;
3437 
3438 private:
3439   /// The addressing modes we've collected.
3440   SmallVector<ExtAddrMode, 16> AddrModes;
3441 
3442   /// The field in which the AddrModes differ, when we have more than one.
3443   ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
3444 
3445   /// Are the AddrModes that we have all just equal to their original values?
3446   bool AllAddrModesTrivial = true;
3447 
3448   /// Common Type for all different fields in addressing modes.
3449   Type *CommonType = nullptr;
3450 
3451   /// SimplifyQuery for simplifyInstruction utility.
3452   const SimplifyQuery &SQ;
3453 
3454   /// Original Address.
3455   Value *Original;
3456 
3457 public:
3458   AddressingModeCombiner(const SimplifyQuery &_SQ, Value *OriginalValue)
3459       : SQ(_SQ), Original(OriginalValue) {}
3460 
3461   /// Get the combined AddrMode
3462   const ExtAddrMode &getAddrMode() const {
3463     return AddrModes[0];
3464   }
3465 
3466   /// Add a new AddrMode if it's compatible with the AddrModes we already
3467   /// have.
3468   /// \return True iff we succeeded in doing so.
3469   bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
3470     // Take note of if we have any non-trivial AddrModes, as we need to detect
3471     // when all AddrModes are trivial as then we would introduce a phi or select
3472     // which just duplicates what's already there.
3473     AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
3474 
3475     // If this is the first addrmode then everything is fine.
3476     if (AddrModes.empty()) {
3477       AddrModes.emplace_back(NewAddrMode);
3478       return true;
3479     }
3480 
3481     // Figure out how different this is from the other address modes, which we
3482     // can do just by comparing against the first one given that we only care
3483     // about the cumulative difference.
3484     ExtAddrMode::FieldName ThisDifferentField =
3485       AddrModes[0].compare(NewAddrMode);
3486     if (DifferentField == ExtAddrMode::NoField)
3487       DifferentField = ThisDifferentField;
3488     else if (DifferentField != ThisDifferentField)
3489       DifferentField = ExtAddrMode::MultipleFields;
3490 
3491     // If NewAddrMode differs in more than one dimension we cannot handle it.
3492     bool CanHandle = DifferentField != ExtAddrMode::MultipleFields;
3493 
3494     // If Scale Field is different then we reject.
3495     CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField;
3496 
3497     // We also must reject the case when base offset is different and
3498     // scale reg is not null, we cannot handle this case due to merge of
3499     // different offsets will be used as ScaleReg.
3500     CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField ||
3501                               !NewAddrMode.ScaledReg);
3502 
3503     // We also must reject the case when GV is different and BaseReg installed
3504     // due to we want to use base reg as a merge of GV values.
3505     CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField ||
3506                               !NewAddrMode.HasBaseReg);
3507 
3508     // Even if NewAddMode is the same we still need to collect it due to
3509     // original value is different. And later we will need all original values
3510     // as anchors during finding the common Phi node.
3511     if (CanHandle)
3512       AddrModes.emplace_back(NewAddrMode);
3513     else
3514       AddrModes.clear();
3515 
3516     return CanHandle;
3517   }
3518 
3519   /// Combine the addressing modes we've collected into a single
3520   /// addressing mode.
3521   /// \return True iff we successfully combined them or we only had one so
3522   /// didn't need to combine them anyway.
3523   bool combineAddrModes() {
3524     // If we have no AddrModes then they can't be combined.
3525     if (AddrModes.size() == 0)
3526       return false;
3527 
3528     // A single AddrMode can trivially be combined.
3529     if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
3530       return true;
3531 
3532     // If the AddrModes we collected are all just equal to the value they are
3533     // derived from then combining them wouldn't do anything useful.
3534     if (AllAddrModesTrivial)
3535       return false;
3536 
3537     if (!addrModeCombiningAllowed())
3538       return false;
3539 
3540     // Build a map between <original value, basic block where we saw it> to
3541     // value of base register.
3542     // Bail out if there is no common type.
3543     FoldAddrToValueMapping Map;
3544     if (!initializeMap(Map))
3545       return false;
3546 
3547     Value *CommonValue = findCommon(Map);
3548     if (CommonValue)
3549       AddrModes[0].SetCombinedField(DifferentField, CommonValue, AddrModes);
3550     return CommonValue != nullptr;
3551   }
3552 
3553 private:
3554   /// Initialize Map with anchor values. For address seen
3555   /// we set the value of different field saw in this address.
3556   /// At the same time we find a common type for different field we will
3557   /// use to create new Phi/Select nodes. Keep it in CommonType field.
3558   /// Return false if there is no common type found.
3559   bool initializeMap(FoldAddrToValueMapping &Map) {
3560     // Keep track of keys where the value is null. We will need to replace it
3561     // with constant null when we know the common type.
3562     SmallVector<Value *, 2> NullValue;
3563     Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
3564     for (auto &AM : AddrModes) {
3565       Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy);
3566       if (DV) {
3567         auto *Type = DV->getType();
3568         if (CommonType && CommonType != Type)
3569           return false;
3570         CommonType = Type;
3571         Map[AM.OriginalValue] = DV;
3572       } else {
3573         NullValue.push_back(AM.OriginalValue);
3574       }
3575     }
3576     assert(CommonType && "At least one non-null value must be!");
3577     for (auto *V : NullValue)
3578       Map[V] = Constant::getNullValue(CommonType);
3579     return true;
3580   }
3581 
3582   /// We have mapping between value A and other value B where B was a field in
3583   /// addressing mode represented by A. Also we have an original value C
3584   /// representing an address we start with. Traversing from C through phi and
3585   /// selects we ended up with A's in a map. This utility function tries to find
3586   /// a value V which is a field in addressing mode C and traversing through phi
3587   /// nodes and selects we will end up in corresponded values B in a map.
3588   /// The utility will create a new Phi/Selects if needed.
3589   // The simple example looks as follows:
3590   // BB1:
3591   //   p1 = b1 + 40
3592   //   br cond BB2, BB3
3593   // BB2:
3594   //   p2 = b2 + 40
3595   //   br BB3
3596   // BB3:
3597   //   p = phi [p1, BB1], [p2, BB2]
3598   //   v = load p
3599   // Map is
3600   //   p1 -> b1
3601   //   p2 -> b2
3602   // Request is
3603   //   p -> ?
3604   // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3.
3605   Value *findCommon(FoldAddrToValueMapping &Map) {
3606     // Tracks the simplification of newly created phi nodes. The reason we use
3607     // this mapping is because we will add new created Phi nodes in AddrToBase.
3608     // Simplification of Phi nodes is recursive, so some Phi node may
3609     // be simplified after we added it to AddrToBase. In reality this
3610     // simplification is possible only if original phi/selects were not
3611     // simplified yet.
3612     // Using this mapping we can find the current value in AddrToBase.
3613     SimplificationTracker ST(SQ);
3614 
3615     // First step, DFS to create PHI nodes for all intermediate blocks.
3616     // Also fill traverse order for the second step.
3617     SmallVector<Value *, 32> TraverseOrder;
3618     InsertPlaceholders(Map, TraverseOrder, ST);
3619 
3620     // Second Step, fill new nodes by merged values and simplify if possible.
3621     FillPlaceholders(Map, TraverseOrder, ST);
3622 
3623     if (!AddrSinkNewSelects && ST.countNewSelectNodes() > 0) {
3624       ST.destroyNewNodes(CommonType);
3625       return nullptr;
3626     }
3627 
3628     // Now we'd like to match New Phi nodes to existed ones.
3629     unsigned PhiNotMatchedCount = 0;
3630     if (!MatchPhiSet(ST, AddrSinkNewPhis, PhiNotMatchedCount)) {
3631       ST.destroyNewNodes(CommonType);
3632       return nullptr;
3633     }
3634 
3635     auto *Result = ST.Get(Map.find(Original)->second);
3636     if (Result) {
3637       NumMemoryInstsPhiCreated += ST.countNewPhiNodes() + PhiNotMatchedCount;
3638       NumMemoryInstsSelectCreated += ST.countNewSelectNodes();
3639     }
3640     return Result;
3641   }
3642 
3643   /// Try to match PHI node to Candidate.
3644   /// Matcher tracks the matched Phi nodes.
3645   bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
3646                     SmallSetVector<PHIPair, 8> &Matcher,
3647                     PhiNodeSet &PhiNodesToMatch) {
3648     SmallVector<PHIPair, 8> WorkList;
3649     Matcher.insert({ PHI, Candidate });
3650     SmallSet<PHINode *, 8> MatchedPHIs;
3651     MatchedPHIs.insert(PHI);
3652     WorkList.push_back({ PHI, Candidate });
3653     SmallSet<PHIPair, 8> Visited;
3654     while (!WorkList.empty()) {
3655       auto Item = WorkList.pop_back_val();
3656       if (!Visited.insert(Item).second)
3657         continue;
3658       // We iterate over all incoming values to Phi to compare them.
3659       // If values are different and both of them Phi and the first one is a
3660       // Phi we added (subject to match) and both of them is in the same basic
3661       // block then we can match our pair if values match. So we state that
3662       // these values match and add it to work list to verify that.
3663       for (auto B : Item.first->blocks()) {
3664         Value *FirstValue = Item.first->getIncomingValueForBlock(B);
3665         Value *SecondValue = Item.second->getIncomingValueForBlock(B);
3666         if (FirstValue == SecondValue)
3667           continue;
3668 
3669         PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue);
3670         PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue);
3671 
3672         // One of them is not Phi or
3673         // The first one is not Phi node from the set we'd like to match or
3674         // Phi nodes from different basic blocks then
3675         // we will not be able to match.
3676         if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) ||
3677             FirstPhi->getParent() != SecondPhi->getParent())
3678           return false;
3679 
3680         // If we already matched them then continue.
3681         if (Matcher.count({ FirstPhi, SecondPhi }))
3682           continue;
3683         // So the values are different and does not match. So we need them to
3684         // match. (But we register no more than one match per PHI node, so that
3685         // we won't later try to replace them twice.)
3686         if (MatchedPHIs.insert(FirstPhi).second)
3687           Matcher.insert({ FirstPhi, SecondPhi });
3688         // But me must check it.
3689         WorkList.push_back({ FirstPhi, SecondPhi });
3690       }
3691     }
3692     return true;
3693   }
3694 
3695   /// For the given set of PHI nodes (in the SimplificationTracker) try
3696   /// to find their equivalents.
3697   /// Returns false if this matching fails and creation of new Phi is disabled.
3698   bool MatchPhiSet(SimplificationTracker &ST, bool AllowNewPhiNodes,
3699                    unsigned &PhiNotMatchedCount) {
3700     // Matched and PhiNodesToMatch iterate their elements in a deterministic
3701     // order, so the replacements (ReplacePhi) are also done in a deterministic
3702     // order.
3703     SmallSetVector<PHIPair, 8> Matched;
3704     SmallPtrSet<PHINode *, 8> WillNotMatch;
3705     PhiNodeSet &PhiNodesToMatch = ST.newPhiNodes();
3706     while (PhiNodesToMatch.size()) {
3707       PHINode *PHI = *PhiNodesToMatch.begin();
3708 
3709       // Add us, if no Phi nodes in the basic block we do not match.
3710       WillNotMatch.clear();
3711       WillNotMatch.insert(PHI);
3712 
3713       // Traverse all Phis until we found equivalent or fail to do that.
3714       bool IsMatched = false;
3715       for (auto &P : PHI->getParent()->phis()) {
3716         // Skip new Phi nodes.
3717         if (PhiNodesToMatch.count(&P))
3718           continue;
3719         if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch)))
3720           break;
3721         // If it does not match, collect all Phi nodes from matcher.
3722         // if we end up with no match, them all these Phi nodes will not match
3723         // later.
3724         for (auto M : Matched)
3725           WillNotMatch.insert(M.first);
3726         Matched.clear();
3727       }
3728       if (IsMatched) {
3729         // Replace all matched values and erase them.
3730         for (auto MV : Matched)
3731           ST.ReplacePhi(MV.first, MV.second);
3732         Matched.clear();
3733         continue;
3734       }
3735       // If we are not allowed to create new nodes then bail out.
3736       if (!AllowNewPhiNodes)
3737         return false;
3738       // Just remove all seen values in matcher. They will not match anything.
3739       PhiNotMatchedCount += WillNotMatch.size();
3740       for (auto *P : WillNotMatch)
3741         PhiNodesToMatch.erase(P);
3742     }
3743     return true;
3744   }
3745   /// Fill the placeholders with values from predecessors and simplify them.
3746   void FillPlaceholders(FoldAddrToValueMapping &Map,
3747                         SmallVectorImpl<Value *> &TraverseOrder,
3748                         SimplificationTracker &ST) {
3749     while (!TraverseOrder.empty()) {
3750       Value *Current = TraverseOrder.pop_back_val();
3751       assert(Map.find(Current) != Map.end() && "No node to fill!!!");
3752       Value *V = Map[Current];
3753 
3754       if (SelectInst *Select = dyn_cast<SelectInst>(V)) {
3755         // CurrentValue also must be Select.
3756         auto *CurrentSelect = cast<SelectInst>(Current);
3757         auto *TrueValue = CurrentSelect->getTrueValue();
3758         assert(Map.find(TrueValue) != Map.end() && "No True Value!");
3759         Select->setTrueValue(ST.Get(Map[TrueValue]));
3760         auto *FalseValue = CurrentSelect->getFalseValue();
3761         assert(Map.find(FalseValue) != Map.end() && "No False Value!");
3762         Select->setFalseValue(ST.Get(Map[FalseValue]));
3763       } else {
3764         // Must be a Phi node then.
3765         auto *PHI = cast<PHINode>(V);
3766         // Fill the Phi node with values from predecessors.
3767         for (auto *B : predecessors(PHI->getParent())) {
3768           Value *PV = cast<PHINode>(Current)->getIncomingValueForBlock(B);
3769           assert(Map.find(PV) != Map.end() && "No predecessor Value!");
3770           PHI->addIncoming(ST.Get(Map[PV]), B);
3771         }
3772       }
3773       Map[Current] = ST.Simplify(V);
3774     }
3775   }
3776 
3777   /// Starting from original value recursively iterates over def-use chain up to
3778   /// known ending values represented in a map. For each traversed phi/select
3779   /// inserts a placeholder Phi or Select.
3780   /// Reports all new created Phi/Select nodes by adding them to set.
3781   /// Also reports and order in what values have been traversed.
3782   void InsertPlaceholders(FoldAddrToValueMapping &Map,
3783                           SmallVectorImpl<Value *> &TraverseOrder,
3784                           SimplificationTracker &ST) {
3785     SmallVector<Value *, 32> Worklist;
3786     assert((isa<PHINode>(Original) || isa<SelectInst>(Original)) &&
3787            "Address must be a Phi or Select node");
3788     auto *Dummy = UndefValue::get(CommonType);
3789     Worklist.push_back(Original);
3790     while (!Worklist.empty()) {
3791       Value *Current = Worklist.pop_back_val();
3792       // if it is already visited or it is an ending value then skip it.
3793       if (Map.find(Current) != Map.end())
3794         continue;
3795       TraverseOrder.push_back(Current);
3796 
3797       // CurrentValue must be a Phi node or select. All others must be covered
3798       // by anchors.
3799       if (SelectInst *CurrentSelect = dyn_cast<SelectInst>(Current)) {
3800         // Is it OK to get metadata from OrigSelect?!
3801         // Create a Select placeholder with dummy value.
3802         SelectInst *Select = SelectInst::Create(
3803             CurrentSelect->getCondition(), Dummy, Dummy,
3804             CurrentSelect->getName(), CurrentSelect, CurrentSelect);
3805         Map[Current] = Select;
3806         ST.insertNewSelect(Select);
3807         // We are interested in True and False values.
3808         Worklist.push_back(CurrentSelect->getTrueValue());
3809         Worklist.push_back(CurrentSelect->getFalseValue());
3810       } else {
3811         // It must be a Phi node then.
3812         PHINode *CurrentPhi = cast<PHINode>(Current);
3813         unsigned PredCount = CurrentPhi->getNumIncomingValues();
3814         PHINode *PHI =
3815             PHINode::Create(CommonType, PredCount, "sunk_phi", CurrentPhi);
3816         Map[Current] = PHI;
3817         ST.insertNewPhi(PHI);
3818         append_range(Worklist, CurrentPhi->incoming_values());
3819       }
3820     }
3821   }
3822 
3823   bool addrModeCombiningAllowed() {
3824     if (DisableComplexAddrModes)
3825       return false;
3826     switch (DifferentField) {
3827     default:
3828       return false;
3829     case ExtAddrMode::BaseRegField:
3830       return AddrSinkCombineBaseReg;
3831     case ExtAddrMode::BaseGVField:
3832       return AddrSinkCombineBaseGV;
3833     case ExtAddrMode::BaseOffsField:
3834       return AddrSinkCombineBaseOffs;
3835     case ExtAddrMode::ScaledRegField:
3836       return AddrSinkCombineScaledReg;
3837     }
3838   }
3839 };
3840 } // end anonymous namespace
3841 
3842 /// Try adding ScaleReg*Scale to the current addressing mode.
3843 /// Return true and update AddrMode if this addr mode is legal for the target,
3844 /// false if not.
3845 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3846                                              unsigned Depth) {
3847   // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3848   // mode.  Just process that directly.
3849   if (Scale == 1)
3850     return matchAddr(ScaleReg, Depth);
3851 
3852   // If the scale is 0, it takes nothing to add this.
3853   if (Scale == 0)
3854     return true;
3855 
3856   // If we already have a scale of this value, we can add to it, otherwise, we
3857   // need an available scale field.
3858   if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3859     return false;
3860 
3861   ExtAddrMode TestAddrMode = AddrMode;
3862 
3863   // Add scale to turn X*4+X*3 -> X*7.  This could also do things like
3864   // [A+B + A*7] -> [B+A*8].
3865   TestAddrMode.Scale += Scale;
3866   TestAddrMode.ScaledReg = ScaleReg;
3867 
3868   // If the new address isn't legal, bail out.
3869   if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3870     return false;
3871 
3872   // It was legal, so commit it.
3873   AddrMode = TestAddrMode;
3874 
3875   // Okay, we decided that we can add ScaleReg+Scale to AddrMode.  Check now
3876   // to see if ScaleReg is actually X+C.  If so, we can turn this into adding
3877   // X*Scale + C*Scale to addr mode. If we found available IV increment, do not
3878   // go any further: we can reuse it and cannot eliminate it.
3879   ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3880   if (isa<Instruction>(ScaleReg) && // not a constant expr.
3881       match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI))) &&
3882       !isIVIncrement(ScaleReg, &LI) && CI->getValue().isSignedIntN(64)) {
3883     TestAddrMode.InBounds = false;
3884     TestAddrMode.ScaledReg = AddLHS;
3885     TestAddrMode.BaseOffs += CI->getSExtValue() * TestAddrMode.Scale;
3886 
3887     // If this addressing mode is legal, commit it and remember that we folded
3888     // this instruction.
3889     if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3890       AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3891       AddrMode = TestAddrMode;
3892       return true;
3893     }
3894     // Restore status quo.
3895     TestAddrMode = AddrMode;
3896   }
3897 
3898   // If this is an add recurrence with a constant step, return the increment
3899   // instruction and the canonicalized step.
3900   auto GetConstantStep = [this](const Value * V)
3901       ->Optional<std::pair<Instruction *, APInt> > {
3902     auto *PN = dyn_cast<PHINode>(V);
3903     if (!PN)
3904       return None;
3905     auto IVInc = getIVIncrement(PN, &LI);
3906     if (!IVInc)
3907       return None;
3908     // TODO: The result of the intrinsics above is two-compliment. However when
3909     // IV inc is expressed as add or sub, iv.next is potentially a poison value.
3910     // If it has nuw or nsw flags, we need to make sure that these flags are
3911     // inferrable at the point of memory instruction. Otherwise we are replacing
3912     // well-defined two-compliment computation with poison. Currently, to avoid
3913     // potentially complex analysis needed to prove this, we reject such cases.
3914     if (auto *OIVInc = dyn_cast<OverflowingBinaryOperator>(IVInc->first))
3915       if (OIVInc->hasNoSignedWrap() || OIVInc->hasNoUnsignedWrap())
3916         return None;
3917     if (auto *ConstantStep = dyn_cast<ConstantInt>(IVInc->second))
3918       return std::make_pair(IVInc->first, ConstantStep->getValue());
3919     return None;
3920   };
3921 
3922   // Try to account for the following special case:
3923   // 1. ScaleReg is an inductive variable;
3924   // 2. We use it with non-zero offset;
3925   // 3. IV's increment is available at the point of memory instruction.
3926   //
3927   // In this case, we may reuse the IV increment instead of the IV Phi to
3928   // achieve the following advantages:
3929   // 1. If IV step matches the offset, we will have no need in the offset;
3930   // 2. Even if they don't match, we will reduce the overlap of living IV
3931   //    and IV increment, that will potentially lead to better register
3932   //    assignment.
3933   if (AddrMode.BaseOffs) {
3934     if (auto IVStep = GetConstantStep(ScaleReg)) {
3935       Instruction *IVInc = IVStep->first;
3936       // The following assert is important to ensure a lack of infinite loops.
3937       // This transforms is (intentionally) the inverse of the one just above.
3938       // If they don't agree on the definition of an increment, we'd alternate
3939       // back and forth indefinitely.
3940       assert(isIVIncrement(IVInc, &LI) && "implied by GetConstantStep");
3941       APInt Step = IVStep->second;
3942       APInt Offset = Step * AddrMode.Scale;
3943       if (Offset.isSignedIntN(64)) {
3944         TestAddrMode.InBounds = false;
3945         TestAddrMode.ScaledReg = IVInc;
3946         TestAddrMode.BaseOffs -= Offset.getLimitedValue();
3947         // If this addressing mode is legal, commit it..
3948         // (Note that we defer the (expensive) domtree base legality check
3949         // to the very last possible point.)
3950         if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace) &&
3951             getDTFn().dominates(IVInc, MemoryInst)) {
3952           AddrModeInsts.push_back(cast<Instruction>(IVInc));
3953           AddrMode = TestAddrMode;
3954           return true;
3955         }
3956         // Restore status quo.
3957         TestAddrMode = AddrMode;
3958       }
3959     }
3960   }
3961 
3962   // Otherwise, just return what we have.
3963   return true;
3964 }
3965 
3966 /// This is a little filter, which returns true if an addressing computation
3967 /// involving I might be folded into a load/store accessing it.
3968 /// This doesn't need to be perfect, but needs to accept at least
3969 /// the set of instructions that MatchOperationAddr can.
3970 static bool MightBeFoldableInst(Instruction *I) {
3971   switch (I->getOpcode()) {
3972   case Instruction::BitCast:
3973   case Instruction::AddrSpaceCast:
3974     // Don't touch identity bitcasts.
3975     if (I->getType() == I->getOperand(0)->getType())
3976       return false;
3977     return I->getType()->isIntOrPtrTy();
3978   case Instruction::PtrToInt:
3979     // PtrToInt is always a noop, as we know that the int type is pointer sized.
3980     return true;
3981   case Instruction::IntToPtr:
3982     // We know the input is intptr_t, so this is foldable.
3983     return true;
3984   case Instruction::Add:
3985     return true;
3986   case Instruction::Mul:
3987   case Instruction::Shl:
3988     // Can only handle X*C and X << C.
3989     return isa<ConstantInt>(I->getOperand(1));
3990   case Instruction::GetElementPtr:
3991     return true;
3992   default:
3993     return false;
3994   }
3995 }
3996 
3997 /// Check whether or not \p Val is a legal instruction for \p TLI.
3998 /// \note \p Val is assumed to be the product of some type promotion.
3999 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
4000 /// to be legal, as the non-promoted value would have had the same state.
4001 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
4002                                        const DataLayout &DL, Value *Val) {
4003   Instruction *PromotedInst = dyn_cast<Instruction>(Val);
4004   if (!PromotedInst)
4005     return false;
4006   int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
4007   // If the ISDOpcode is undefined, it was undefined before the promotion.
4008   if (!ISDOpcode)
4009     return true;
4010   // Otherwise, check if the promoted instruction is legal or not.
4011   return TLI.isOperationLegalOrCustom(
4012       ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
4013 }
4014 
4015 namespace {
4016 
4017 /// Hepler class to perform type promotion.
4018 class TypePromotionHelper {
4019   /// Utility function to add a promoted instruction \p ExtOpnd to
4020   /// \p PromotedInsts and record the type of extension we have seen.
4021   static void addPromotedInst(InstrToOrigTy &PromotedInsts,
4022                               Instruction *ExtOpnd,
4023                               bool IsSExt) {
4024     ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
4025     InstrToOrigTy::iterator It = PromotedInsts.find(ExtOpnd);
4026     if (It != PromotedInsts.end()) {
4027       // If the new extension is same as original, the information in
4028       // PromotedInsts[ExtOpnd] is still correct.
4029       if (It->second.getInt() == ExtTy)
4030         return;
4031 
4032       // Now the new extension is different from old extension, we make
4033       // the type information invalid by setting extension type to
4034       // BothExtension.
4035       ExtTy = BothExtension;
4036     }
4037     PromotedInsts[ExtOpnd] = TypeIsSExt(ExtOpnd->getType(), ExtTy);
4038   }
4039 
4040   /// Utility function to query the original type of instruction \p Opnd
4041   /// with a matched extension type. If the extension doesn't match, we
4042   /// cannot use the information we had on the original type.
4043   /// BothExtension doesn't match any extension type.
4044   static const Type *getOrigType(const InstrToOrigTy &PromotedInsts,
4045                                  Instruction *Opnd,
4046                                  bool IsSExt) {
4047     ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
4048     InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
4049     if (It != PromotedInsts.end() && It->second.getInt() == ExtTy)
4050       return It->second.getPointer();
4051     return nullptr;
4052   }
4053 
4054   /// Utility function to check whether or not a sign or zero extension
4055   /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
4056   /// either using the operands of \p Inst or promoting \p Inst.
4057   /// The type of the extension is defined by \p IsSExt.
4058   /// In other words, check if:
4059   /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
4060   /// #1 Promotion applies:
4061   /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
4062   /// #2 Operand reuses:
4063   /// ext opnd1 to ConsideredExtType.
4064   /// \p PromotedInsts maps the instructions to their type before promotion.
4065   static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
4066                             const InstrToOrigTy &PromotedInsts, bool IsSExt);
4067 
4068   /// Utility function to determine if \p OpIdx should be promoted when
4069   /// promoting \p Inst.
4070   static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
4071     return !(isa<SelectInst>(Inst) && OpIdx == 0);
4072   }
4073 
4074   /// Utility function to promote the operand of \p Ext when this
4075   /// operand is a promotable trunc or sext or zext.
4076   /// \p PromotedInsts maps the instructions to their type before promotion.
4077   /// \p CreatedInstsCost[out] contains the cost of all instructions
4078   /// created to promote the operand of Ext.
4079   /// Newly added extensions are inserted in \p Exts.
4080   /// Newly added truncates are inserted in \p Truncs.
4081   /// Should never be called directly.
4082   /// \return The promoted value which is used instead of Ext.
4083   static Value *promoteOperandForTruncAndAnyExt(
4084       Instruction *Ext, TypePromotionTransaction &TPT,
4085       InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4086       SmallVectorImpl<Instruction *> *Exts,
4087       SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
4088 
4089   /// Utility function to promote the operand of \p Ext when this
4090   /// operand is promotable and is not a supported trunc or sext.
4091   /// \p PromotedInsts maps the instructions to their type before promotion.
4092   /// \p CreatedInstsCost[out] contains the cost of all the instructions
4093   /// created to promote the operand of Ext.
4094   /// Newly added extensions are inserted in \p Exts.
4095   /// Newly added truncates are inserted in \p Truncs.
4096   /// Should never be called directly.
4097   /// \return The promoted value which is used instead of Ext.
4098   static Value *promoteOperandForOther(Instruction *Ext,
4099                                        TypePromotionTransaction &TPT,
4100                                        InstrToOrigTy &PromotedInsts,
4101                                        unsigned &CreatedInstsCost,
4102                                        SmallVectorImpl<Instruction *> *Exts,
4103                                        SmallVectorImpl<Instruction *> *Truncs,
4104                                        const TargetLowering &TLI, bool IsSExt);
4105 
4106   /// \see promoteOperandForOther.
4107   static Value *signExtendOperandForOther(
4108       Instruction *Ext, TypePromotionTransaction &TPT,
4109       InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4110       SmallVectorImpl<Instruction *> *Exts,
4111       SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4112     return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
4113                                   Exts, Truncs, TLI, true);
4114   }
4115 
4116   /// \see promoteOperandForOther.
4117   static Value *zeroExtendOperandForOther(
4118       Instruction *Ext, TypePromotionTransaction &TPT,
4119       InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4120       SmallVectorImpl<Instruction *> *Exts,
4121       SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4122     return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
4123                                   Exts, Truncs, TLI, false);
4124   }
4125 
4126 public:
4127   /// Type for the utility function that promotes the operand of Ext.
4128   using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
4129                             InstrToOrigTy &PromotedInsts,
4130                             unsigned &CreatedInstsCost,
4131                             SmallVectorImpl<Instruction *> *Exts,
4132                             SmallVectorImpl<Instruction *> *Truncs,
4133                             const TargetLowering &TLI);
4134 
4135   /// Given a sign/zero extend instruction \p Ext, return the appropriate
4136   /// action to promote the operand of \p Ext instead of using Ext.
4137   /// \return NULL if no promotable action is possible with the current
4138   /// sign extension.
4139   /// \p InsertedInsts keeps track of all the instructions inserted by the
4140   /// other CodeGenPrepare optimizations. This information is important
4141   /// because we do not want to promote these instructions as CodeGenPrepare
4142   /// will reinsert them later. Thus creating an infinite loop: create/remove.
4143   /// \p PromotedInsts maps the instructions to their type before promotion.
4144   static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
4145                           const TargetLowering &TLI,
4146                           const InstrToOrigTy &PromotedInsts);
4147 };
4148 
4149 } // end anonymous namespace
4150 
4151 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
4152                                         Type *ConsideredExtType,
4153                                         const InstrToOrigTy &PromotedInsts,
4154                                         bool IsSExt) {
4155   // The promotion helper does not know how to deal with vector types yet.
4156   // To be able to fix that, we would need to fix the places where we
4157   // statically extend, e.g., constants and such.
4158   if (Inst->getType()->isVectorTy())
4159     return false;
4160 
4161   // We can always get through zext.
4162   if (isa<ZExtInst>(Inst))
4163     return true;
4164 
4165   // sext(sext) is ok too.
4166   if (IsSExt && isa<SExtInst>(Inst))
4167     return true;
4168 
4169   // We can get through binary operator, if it is legal. In other words, the
4170   // binary operator must have a nuw or nsw flag.
4171   if (const auto *BinOp = dyn_cast<BinaryOperator>(Inst))
4172     if (isa<OverflowingBinaryOperator>(BinOp) &&
4173         ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
4174          (IsSExt && BinOp->hasNoSignedWrap())))
4175       return true;
4176 
4177   // ext(and(opnd, cst)) --> and(ext(opnd), ext(cst))
4178   if ((Inst->getOpcode() == Instruction::And ||
4179        Inst->getOpcode() == Instruction::Or))
4180     return true;
4181 
4182   // ext(xor(opnd, cst)) --> xor(ext(opnd), ext(cst))
4183   if (Inst->getOpcode() == Instruction::Xor) {
4184     // Make sure it is not a NOT.
4185     if (const auto *Cst = dyn_cast<ConstantInt>(Inst->getOperand(1)))
4186       if (!Cst->getValue().isAllOnes())
4187         return true;
4188   }
4189 
4190   // zext(shrl(opnd, cst)) --> shrl(zext(opnd), zext(cst))
4191   // It may change a poisoned value into a regular value, like
4192   //     zext i32 (shrl i8 %val, 12)  -->  shrl i32 (zext i8 %val), 12
4193   //          poisoned value                    regular value
4194   // It should be OK since undef covers valid value.
4195   if (Inst->getOpcode() == Instruction::LShr && !IsSExt)
4196     return true;
4197 
4198   // and(ext(shl(opnd, cst)), cst) --> and(shl(ext(opnd), ext(cst)), cst)
4199   // It may change a poisoned value into a regular value, like
4200   //     zext i32 (shl i8 %val, 12)  -->  shl i32 (zext i8 %val), 12
4201   //          poisoned value                    regular value
4202   // It should be OK since undef covers valid value.
4203   if (Inst->getOpcode() == Instruction::Shl && Inst->hasOneUse()) {
4204     const auto *ExtInst = cast<const Instruction>(*Inst->user_begin());
4205     if (ExtInst->hasOneUse()) {
4206       const auto *AndInst = dyn_cast<const Instruction>(*ExtInst->user_begin());
4207       if (AndInst && AndInst->getOpcode() == Instruction::And) {
4208         const auto *Cst = dyn_cast<ConstantInt>(AndInst->getOperand(1));
4209         if (Cst &&
4210             Cst->getValue().isIntN(Inst->getType()->getIntegerBitWidth()))
4211           return true;
4212       }
4213     }
4214   }
4215 
4216   // Check if we can do the following simplification.
4217   // ext(trunc(opnd)) --> ext(opnd)
4218   if (!isa<TruncInst>(Inst))
4219     return false;
4220 
4221   Value *OpndVal = Inst->getOperand(0);
4222   // Check if we can use this operand in the extension.
4223   // If the type is larger than the result type of the extension, we cannot.
4224   if (!OpndVal->getType()->isIntegerTy() ||
4225       OpndVal->getType()->getIntegerBitWidth() >
4226           ConsideredExtType->getIntegerBitWidth())
4227     return false;
4228 
4229   // If the operand of the truncate is not an instruction, we will not have
4230   // any information on the dropped bits.
4231   // (Actually we could for constant but it is not worth the extra logic).
4232   Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
4233   if (!Opnd)
4234     return false;
4235 
4236   // Check if the source of the type is narrow enough.
4237   // I.e., check that trunc just drops extended bits of the same kind of
4238   // the extension.
4239   // #1 get the type of the operand and check the kind of the extended bits.
4240   const Type *OpndType = getOrigType(PromotedInsts, Opnd, IsSExt);
4241   if (OpndType)
4242     ;
4243   else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
4244     OpndType = Opnd->getOperand(0)->getType();
4245   else
4246     return false;
4247 
4248   // #2 check that the truncate just drops extended bits.
4249   return Inst->getType()->getIntegerBitWidth() >=
4250          OpndType->getIntegerBitWidth();
4251 }
4252 
4253 TypePromotionHelper::Action TypePromotionHelper::getAction(
4254     Instruction *Ext, const SetOfInstrs &InsertedInsts,
4255     const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
4256   assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
4257          "Unexpected instruction type");
4258   Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
4259   Type *ExtTy = Ext->getType();
4260   bool IsSExt = isa<SExtInst>(Ext);
4261   // If the operand of the extension is not an instruction, we cannot
4262   // get through.
4263   // If it, check we can get through.
4264   if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
4265     return nullptr;
4266 
4267   // Do not promote if the operand has been added by codegenprepare.
4268   // Otherwise, it means we are undoing an optimization that is likely to be
4269   // redone, thus causing potential infinite loop.
4270   if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
4271     return nullptr;
4272 
4273   // SExt or Trunc instructions.
4274   // Return the related handler.
4275   if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
4276       isa<ZExtInst>(ExtOpnd))
4277     return promoteOperandForTruncAndAnyExt;
4278 
4279   // Regular instruction.
4280   // Abort early if we will have to insert non-free instructions.
4281   if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
4282     return nullptr;
4283   return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
4284 }
4285 
4286 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
4287     Instruction *SExt, TypePromotionTransaction &TPT,
4288     InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4289     SmallVectorImpl<Instruction *> *Exts,
4290     SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4291   // By construction, the operand of SExt is an instruction. Otherwise we cannot
4292   // get through it and this method should not be called.
4293   Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
4294   Value *ExtVal = SExt;
4295   bool HasMergedNonFreeExt = false;
4296   if (isa<ZExtInst>(SExtOpnd)) {
4297     // Replace s|zext(zext(opnd))
4298     // => zext(opnd).
4299     HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
4300     Value *ZExt =
4301         TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
4302     TPT.replaceAllUsesWith(SExt, ZExt);
4303     TPT.eraseInstruction(SExt);
4304     ExtVal = ZExt;
4305   } else {
4306     // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
4307     // => z|sext(opnd).
4308     TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
4309   }
4310   CreatedInstsCost = 0;
4311 
4312   // Remove dead code.
4313   if (SExtOpnd->use_empty())
4314     TPT.eraseInstruction(SExtOpnd);
4315 
4316   // Check if the extension is still needed.
4317   Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
4318   if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
4319     if (ExtInst) {
4320       if (Exts)
4321         Exts->push_back(ExtInst);
4322       CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
4323     }
4324     return ExtVal;
4325   }
4326 
4327   // At this point we have: ext ty opnd to ty.
4328   // Reassign the uses of ExtInst to the opnd and remove ExtInst.
4329   Value *NextVal = ExtInst->getOperand(0);
4330   TPT.eraseInstruction(ExtInst, NextVal);
4331   return NextVal;
4332 }
4333 
4334 Value *TypePromotionHelper::promoteOperandForOther(
4335     Instruction *Ext, TypePromotionTransaction &TPT,
4336     InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4337     SmallVectorImpl<Instruction *> *Exts,
4338     SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
4339     bool IsSExt) {
4340   // By construction, the operand of Ext is an instruction. Otherwise we cannot
4341   // get through it and this method should not be called.
4342   Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
4343   CreatedInstsCost = 0;
4344   if (!ExtOpnd->hasOneUse()) {
4345     // ExtOpnd will be promoted.
4346     // All its uses, but Ext, will need to use a truncated value of the
4347     // promoted version.
4348     // Create the truncate now.
4349     Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
4350     if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
4351       // Insert it just after the definition.
4352       ITrunc->moveAfter(ExtOpnd);
4353       if (Truncs)
4354         Truncs->push_back(ITrunc);
4355     }
4356 
4357     TPT.replaceAllUsesWith(ExtOpnd, Trunc);
4358     // Restore the operand of Ext (which has been replaced by the previous call
4359     // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
4360     TPT.setOperand(Ext, 0, ExtOpnd);
4361   }
4362 
4363   // Get through the Instruction:
4364   // 1. Update its type.
4365   // 2. Replace the uses of Ext by Inst.
4366   // 3. Extend each operand that needs to be extended.
4367 
4368   // Remember the original type of the instruction before promotion.
4369   // This is useful to know that the high bits are sign extended bits.
4370   addPromotedInst(PromotedInsts, ExtOpnd, IsSExt);
4371   // Step #1.
4372   TPT.mutateType(ExtOpnd, Ext->getType());
4373   // Step #2.
4374   TPT.replaceAllUsesWith(Ext, ExtOpnd);
4375   // Step #3.
4376   Instruction *ExtForOpnd = Ext;
4377 
4378   LLVM_DEBUG(dbgs() << "Propagate Ext to operands\n");
4379   for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
4380        ++OpIdx) {
4381     LLVM_DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
4382     if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
4383         !shouldExtOperand(ExtOpnd, OpIdx)) {
4384       LLVM_DEBUG(dbgs() << "No need to propagate\n");
4385       continue;
4386     }
4387     // Check if we can statically extend the operand.
4388     Value *Opnd = ExtOpnd->getOperand(OpIdx);
4389     if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
4390       LLVM_DEBUG(dbgs() << "Statically extend\n");
4391       unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
4392       APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
4393                             : Cst->getValue().zext(BitWidth);
4394       TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
4395       continue;
4396     }
4397     // UndefValue are typed, so we have to statically sign extend them.
4398     if (isa<UndefValue>(Opnd)) {
4399       LLVM_DEBUG(dbgs() << "Statically extend\n");
4400       TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
4401       continue;
4402     }
4403 
4404     // Otherwise we have to explicitly sign extend the operand.
4405     // Check if Ext was reused to extend an operand.
4406     if (!ExtForOpnd) {
4407       // If yes, create a new one.
4408       LLVM_DEBUG(dbgs() << "More operands to ext\n");
4409       Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
4410         : TPT.createZExt(Ext, Opnd, Ext->getType());
4411       if (!isa<Instruction>(ValForExtOpnd)) {
4412         TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
4413         continue;
4414       }
4415       ExtForOpnd = cast<Instruction>(ValForExtOpnd);
4416     }
4417     if (Exts)
4418       Exts->push_back(ExtForOpnd);
4419     TPT.setOperand(ExtForOpnd, 0, Opnd);
4420 
4421     // Move the sign extension before the insertion point.
4422     TPT.moveBefore(ExtForOpnd, ExtOpnd);
4423     TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
4424     CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
4425     // If more sext are required, new instructions will have to be created.
4426     ExtForOpnd = nullptr;
4427   }
4428   if (ExtForOpnd == Ext) {
4429     LLVM_DEBUG(dbgs() << "Extension is useless now\n");
4430     TPT.eraseInstruction(Ext);
4431   }
4432   return ExtOpnd;
4433 }
4434 
4435 /// Check whether or not promoting an instruction to a wider type is profitable.
4436 /// \p NewCost gives the cost of extension instructions created by the
4437 /// promotion.
4438 /// \p OldCost gives the cost of extension instructions before the promotion
4439 /// plus the number of instructions that have been
4440 /// matched in the addressing mode the promotion.
4441 /// \p PromotedOperand is the value that has been promoted.
4442 /// \return True if the promotion is profitable, false otherwise.
4443 bool AddressingModeMatcher::isPromotionProfitable(
4444     unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
4445   LLVM_DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost
4446                     << '\n');
4447   // The cost of the new extensions is greater than the cost of the
4448   // old extension plus what we folded.
4449   // This is not profitable.
4450   if (NewCost > OldCost)
4451     return false;
4452   if (NewCost < OldCost)
4453     return true;
4454   // The promotion is neutral but it may help folding the sign extension in
4455   // loads for instance.
4456   // Check that we did not create an illegal instruction.
4457   return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
4458 }
4459 
4460 /// Given an instruction or constant expr, see if we can fold the operation
4461 /// into the addressing mode. If so, update the addressing mode and return
4462 /// true, otherwise return false without modifying AddrMode.
4463 /// If \p MovedAway is not NULL, it contains the information of whether or
4464 /// not AddrInst has to be folded into the addressing mode on success.
4465 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
4466 /// because it has been moved away.
4467 /// Thus AddrInst must not be added in the matched instructions.
4468 /// This state can happen when AddrInst is a sext, since it may be moved away.
4469 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
4470 /// not be referenced anymore.
4471 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
4472                                                unsigned Depth,
4473                                                bool *MovedAway) {
4474   // Avoid exponential behavior on extremely deep expression trees.
4475   if (Depth >= 5) return false;
4476 
4477   // By default, all matched instructions stay in place.
4478   if (MovedAway)
4479     *MovedAway = false;
4480 
4481   switch (Opcode) {
4482   case Instruction::PtrToInt:
4483     // PtrToInt is always a noop, as we know that the int type is pointer sized.
4484     return matchAddr(AddrInst->getOperand(0), Depth);
4485   case Instruction::IntToPtr: {
4486     auto AS = AddrInst->getType()->getPointerAddressSpace();
4487     auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
4488     // This inttoptr is a no-op if the integer type is pointer sized.
4489     if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
4490       return matchAddr(AddrInst->getOperand(0), Depth);
4491     return false;
4492   }
4493   case Instruction::BitCast:
4494     // BitCast is always a noop, and we can handle it as long as it is
4495     // int->int or pointer->pointer (we don't want int<->fp or something).
4496     if (AddrInst->getOperand(0)->getType()->isIntOrPtrTy() &&
4497         // Don't touch identity bitcasts.  These were probably put here by LSR,
4498         // and we don't want to mess around with them.  Assume it knows what it
4499         // is doing.
4500         AddrInst->getOperand(0)->getType() != AddrInst->getType())
4501       return matchAddr(AddrInst->getOperand(0), Depth);
4502     return false;
4503   case Instruction::AddrSpaceCast: {
4504     unsigned SrcAS
4505       = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
4506     unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
4507     if (TLI.getTargetMachine().isNoopAddrSpaceCast(SrcAS, DestAS))
4508       return matchAddr(AddrInst->getOperand(0), Depth);
4509     return false;
4510   }
4511   case Instruction::Add: {
4512     // Check to see if we can merge in the RHS then the LHS.  If so, we win.
4513     ExtAddrMode BackupAddrMode = AddrMode;
4514     unsigned OldSize = AddrModeInsts.size();
4515     // Start a transaction at this point.
4516     // The LHS may match but not the RHS.
4517     // Therefore, we need a higher level restoration point to undo partially
4518     // matched operation.
4519     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4520         TPT.getRestorationPoint();
4521 
4522     AddrMode.InBounds = false;
4523     if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
4524         matchAddr(AddrInst->getOperand(0), Depth+1))
4525       return true;
4526 
4527     // Restore the old addr mode info.
4528     AddrMode = BackupAddrMode;
4529     AddrModeInsts.resize(OldSize);
4530     TPT.rollback(LastKnownGood);
4531 
4532     // Otherwise this was over-aggressive.  Try merging in the LHS then the RHS.
4533     if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
4534         matchAddr(AddrInst->getOperand(1), Depth+1))
4535       return true;
4536 
4537     // Otherwise we definitely can't merge the ADD in.
4538     AddrMode = BackupAddrMode;
4539     AddrModeInsts.resize(OldSize);
4540     TPT.rollback(LastKnownGood);
4541     break;
4542   }
4543   //case Instruction::Or:
4544   // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
4545   //break;
4546   case Instruction::Mul:
4547   case Instruction::Shl: {
4548     // Can only handle X*C and X << C.
4549     AddrMode.InBounds = false;
4550     ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
4551     if (!RHS || RHS->getBitWidth() > 64)
4552       return false;
4553     int64_t Scale = RHS->getSExtValue();
4554     if (Opcode == Instruction::Shl)
4555       Scale = 1LL << Scale;
4556 
4557     return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
4558   }
4559   case Instruction::GetElementPtr: {
4560     // Scan the GEP.  We check it if it contains constant offsets and at most
4561     // one variable offset.
4562     int VariableOperand = -1;
4563     unsigned VariableScale = 0;
4564 
4565     int64_t ConstantOffset = 0;
4566     gep_type_iterator GTI = gep_type_begin(AddrInst);
4567     for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
4568       if (StructType *STy = GTI.getStructTypeOrNull()) {
4569         const StructLayout *SL = DL.getStructLayout(STy);
4570         unsigned Idx =
4571           cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
4572         ConstantOffset += SL->getElementOffset(Idx);
4573       } else {
4574         TypeSize TS = DL.getTypeAllocSize(GTI.getIndexedType());
4575         if (TS.isNonZero()) {
4576           // The optimisations below currently only work for fixed offsets.
4577           if (TS.isScalable())
4578             return false;
4579           int64_t TypeSize = TS.getFixedSize();
4580           if (ConstantInt *CI =
4581                   dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
4582             const APInt &CVal = CI->getValue();
4583             if (CVal.getMinSignedBits() <= 64) {
4584               ConstantOffset += CVal.getSExtValue() * TypeSize;
4585               continue;
4586             }
4587           }
4588           // We only allow one variable index at the moment.
4589           if (VariableOperand != -1)
4590             return false;
4591 
4592           // Remember the variable index.
4593           VariableOperand = i;
4594           VariableScale = TypeSize;
4595         }
4596       }
4597     }
4598 
4599     // A common case is for the GEP to only do a constant offset.  In this case,
4600     // just add it to the disp field and check validity.
4601     if (VariableOperand == -1) {
4602       AddrMode.BaseOffs += ConstantOffset;
4603       if (ConstantOffset == 0 ||
4604           TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
4605         // Check to see if we can fold the base pointer in too.
4606         if (matchAddr(AddrInst->getOperand(0), Depth+1)) {
4607           if (!cast<GEPOperator>(AddrInst)->isInBounds())
4608             AddrMode.InBounds = false;
4609           return true;
4610         }
4611       } else if (EnableGEPOffsetSplit && isa<GetElementPtrInst>(AddrInst) &&
4612                  TLI.shouldConsiderGEPOffsetSplit() && Depth == 0 &&
4613                  ConstantOffset > 0) {
4614         // Record GEPs with non-zero offsets as candidates for splitting in the
4615         // event that the offset cannot fit into the r+i addressing mode.
4616         // Simple and common case that only one GEP is used in calculating the
4617         // address for the memory access.
4618         Value *Base = AddrInst->getOperand(0);
4619         auto *BaseI = dyn_cast<Instruction>(Base);
4620         auto *GEP = cast<GetElementPtrInst>(AddrInst);
4621         if (isa<Argument>(Base) || isa<GlobalValue>(Base) ||
4622             (BaseI && !isa<CastInst>(BaseI) &&
4623              !isa<GetElementPtrInst>(BaseI))) {
4624           // Make sure the parent block allows inserting non-PHI instructions
4625           // before the terminator.
4626           BasicBlock *Parent =
4627               BaseI ? BaseI->getParent() : &GEP->getFunction()->getEntryBlock();
4628           if (!Parent->getTerminator()->isEHPad())
4629             LargeOffsetGEP = std::make_pair(GEP, ConstantOffset);
4630         }
4631       }
4632       AddrMode.BaseOffs -= ConstantOffset;
4633       return false;
4634     }
4635 
4636     // Save the valid addressing mode in case we can't match.
4637     ExtAddrMode BackupAddrMode = AddrMode;
4638     unsigned OldSize = AddrModeInsts.size();
4639 
4640     // See if the scale and offset amount is valid for this target.
4641     AddrMode.BaseOffs += ConstantOffset;
4642     if (!cast<GEPOperator>(AddrInst)->isInBounds())
4643       AddrMode.InBounds = false;
4644 
4645     // Match the base operand of the GEP.
4646     if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
4647       // If it couldn't be matched, just stuff the value in a register.
4648       if (AddrMode.HasBaseReg) {
4649         AddrMode = BackupAddrMode;
4650         AddrModeInsts.resize(OldSize);
4651         return false;
4652       }
4653       AddrMode.HasBaseReg = true;
4654       AddrMode.BaseReg = AddrInst->getOperand(0);
4655     }
4656 
4657     // Match the remaining variable portion of the GEP.
4658     if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
4659                           Depth)) {
4660       // If it couldn't be matched, try stuffing the base into a register
4661       // instead of matching it, and retrying the match of the scale.
4662       AddrMode = BackupAddrMode;
4663       AddrModeInsts.resize(OldSize);
4664       if (AddrMode.HasBaseReg)
4665         return false;
4666       AddrMode.HasBaseReg = true;
4667       AddrMode.BaseReg = AddrInst->getOperand(0);
4668       AddrMode.BaseOffs += ConstantOffset;
4669       if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
4670                             VariableScale, Depth)) {
4671         // If even that didn't work, bail.
4672         AddrMode = BackupAddrMode;
4673         AddrModeInsts.resize(OldSize);
4674         return false;
4675       }
4676     }
4677 
4678     return true;
4679   }
4680   case Instruction::SExt:
4681   case Instruction::ZExt: {
4682     Instruction *Ext = dyn_cast<Instruction>(AddrInst);
4683     if (!Ext)
4684       return false;
4685 
4686     // Try to move this ext out of the way of the addressing mode.
4687     // Ask for a method for doing so.
4688     TypePromotionHelper::Action TPH =
4689         TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
4690     if (!TPH)
4691       return false;
4692 
4693     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4694         TPT.getRestorationPoint();
4695     unsigned CreatedInstsCost = 0;
4696     unsigned ExtCost = !TLI.isExtFree(Ext);
4697     Value *PromotedOperand =
4698         TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
4699     // SExt has been moved away.
4700     // Thus either it will be rematched later in the recursive calls or it is
4701     // gone. Anyway, we must not fold it into the addressing mode at this point.
4702     // E.g.,
4703     // op = add opnd, 1
4704     // idx = ext op
4705     // addr = gep base, idx
4706     // is now:
4707     // promotedOpnd = ext opnd            <- no match here
4708     // op = promoted_add promotedOpnd, 1  <- match (later in recursive calls)
4709     // addr = gep base, op                <- match
4710     if (MovedAway)
4711       *MovedAway = true;
4712 
4713     assert(PromotedOperand &&
4714            "TypePromotionHelper should have filtered out those cases");
4715 
4716     ExtAddrMode BackupAddrMode = AddrMode;
4717     unsigned OldSize = AddrModeInsts.size();
4718 
4719     if (!matchAddr(PromotedOperand, Depth) ||
4720         // The total of the new cost is equal to the cost of the created
4721         // instructions.
4722         // The total of the old cost is equal to the cost of the extension plus
4723         // what we have saved in the addressing mode.
4724         !isPromotionProfitable(CreatedInstsCost,
4725                                ExtCost + (AddrModeInsts.size() - OldSize),
4726                                PromotedOperand)) {
4727       AddrMode = BackupAddrMode;
4728       AddrModeInsts.resize(OldSize);
4729       LLVM_DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
4730       TPT.rollback(LastKnownGood);
4731       return false;
4732     }
4733     return true;
4734   }
4735   }
4736   return false;
4737 }
4738 
4739 /// If we can, try to add the value of 'Addr' into the current addressing mode.
4740 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
4741 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
4742 /// for the target.
4743 ///
4744 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
4745   // Start a transaction at this point that we will rollback if the matching
4746   // fails.
4747   TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4748       TPT.getRestorationPoint();
4749   if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
4750     if (CI->getValue().isSignedIntN(64)) {
4751       // Fold in immediates if legal for the target.
4752       AddrMode.BaseOffs += CI->getSExtValue();
4753       if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4754         return true;
4755       AddrMode.BaseOffs -= CI->getSExtValue();
4756     }
4757   } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
4758     // If this is a global variable, try to fold it into the addressing mode.
4759     if (!AddrMode.BaseGV) {
4760       AddrMode.BaseGV = GV;
4761       if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4762         return true;
4763       AddrMode.BaseGV = nullptr;
4764     }
4765   } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
4766     ExtAddrMode BackupAddrMode = AddrMode;
4767     unsigned OldSize = AddrModeInsts.size();
4768 
4769     // Check to see if it is possible to fold this operation.
4770     bool MovedAway = false;
4771     if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
4772       // This instruction may have been moved away. If so, there is nothing
4773       // to check here.
4774       if (MovedAway)
4775         return true;
4776       // Okay, it's possible to fold this.  Check to see if it is actually
4777       // *profitable* to do so.  We use a simple cost model to avoid increasing
4778       // register pressure too much.
4779       if (I->hasOneUse() ||
4780           isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
4781         AddrModeInsts.push_back(I);
4782         return true;
4783       }
4784 
4785       // It isn't profitable to do this, roll back.
4786       //cerr << "NOT FOLDING: " << *I;
4787       AddrMode = BackupAddrMode;
4788       AddrModeInsts.resize(OldSize);
4789       TPT.rollback(LastKnownGood);
4790     }
4791   } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
4792     if (matchOperationAddr(CE, CE->getOpcode(), Depth))
4793       return true;
4794     TPT.rollback(LastKnownGood);
4795   } else if (isa<ConstantPointerNull>(Addr)) {
4796     // Null pointer gets folded without affecting the addressing mode.
4797     return true;
4798   }
4799 
4800   // Worse case, the target should support [reg] addressing modes. :)
4801   if (!AddrMode.HasBaseReg) {
4802     AddrMode.HasBaseReg = true;
4803     AddrMode.BaseReg = Addr;
4804     // Still check for legality in case the target supports [imm] but not [i+r].
4805     if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4806       return true;
4807     AddrMode.HasBaseReg = false;
4808     AddrMode.BaseReg = nullptr;
4809   }
4810 
4811   // If the base register is already taken, see if we can do [r+r].
4812   if (AddrMode.Scale == 0) {
4813     AddrMode.Scale = 1;
4814     AddrMode.ScaledReg = Addr;
4815     if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4816       return true;
4817     AddrMode.Scale = 0;
4818     AddrMode.ScaledReg = nullptr;
4819   }
4820   // Couldn't match.
4821   TPT.rollback(LastKnownGood);
4822   return false;
4823 }
4824 
4825 /// Check to see if all uses of OpVal by the specified inline asm call are due
4826 /// to memory operands. If so, return true, otherwise return false.
4827 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
4828                                     const TargetLowering &TLI,
4829                                     const TargetRegisterInfo &TRI) {
4830   const Function *F = CI->getFunction();
4831   TargetLowering::AsmOperandInfoVector TargetConstraints =
4832       TLI.ParseConstraints(F->getParent()->getDataLayout(), &TRI, *CI);
4833 
4834   for (TargetLowering::AsmOperandInfo &OpInfo : TargetConstraints) {
4835     // Compute the constraint code and ConstraintType to use.
4836     TLI.ComputeConstraintToUse(OpInfo, SDValue());
4837 
4838     // If this asm operand is our Value*, and if it isn't an indirect memory
4839     // operand, we can't fold it!
4840     if (OpInfo.CallOperandVal == OpVal &&
4841         (OpInfo.ConstraintType != TargetLowering::C_Memory ||
4842          !OpInfo.isIndirect))
4843       return false;
4844   }
4845 
4846   return true;
4847 }
4848 
4849 // Max number of memory uses to look at before aborting the search to conserve
4850 // compile time.
4851 static constexpr int MaxMemoryUsesToScan = 20;
4852 
4853 /// Recursively walk all the uses of I until we find a memory use.
4854 /// If we find an obviously non-foldable instruction, return true.
4855 /// Add accessed addresses and types to MemoryUses.
4856 static bool FindAllMemoryUses(
4857     Instruction *I, SmallVectorImpl<std::pair<Value *, Type *>> &MemoryUses,
4858     SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
4859     const TargetRegisterInfo &TRI, bool OptSize, ProfileSummaryInfo *PSI,
4860     BlockFrequencyInfo *BFI, int SeenInsts = 0) {
4861   // If we already considered this instruction, we're done.
4862   if (!ConsideredInsts.insert(I).second)
4863     return false;
4864 
4865   // If this is an obviously unfoldable instruction, bail out.
4866   if (!MightBeFoldableInst(I))
4867     return true;
4868 
4869   // Loop over all the uses, recursively processing them.
4870   for (Use &U : I->uses()) {
4871     // Conservatively return true if we're seeing a large number or a deep chain
4872     // of users. This avoids excessive compilation times in pathological cases.
4873     if (SeenInsts++ >= MaxMemoryUsesToScan)
4874       return true;
4875 
4876     Instruction *UserI = cast<Instruction>(U.getUser());
4877     if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4878       MemoryUses.push_back({U.get(), LI->getType()});
4879       continue;
4880     }
4881 
4882     if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4883       if (U.getOperandNo() != StoreInst::getPointerOperandIndex())
4884         return true; // Storing addr, not into addr.
4885       MemoryUses.push_back({U.get(), SI->getValueOperand()->getType()});
4886       continue;
4887     }
4888 
4889     if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
4890       if (U.getOperandNo() != AtomicRMWInst::getPointerOperandIndex())
4891         return true; // Storing addr, not into addr.
4892       MemoryUses.push_back({U.get(), RMW->getValOperand()->getType()});
4893       continue;
4894     }
4895 
4896     if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
4897       if (U.getOperandNo() != AtomicCmpXchgInst::getPointerOperandIndex())
4898         return true; // Storing addr, not into addr.
4899       MemoryUses.push_back({U.get(), CmpX->getCompareOperand()->getType()});
4900       continue;
4901     }
4902 
4903     if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4904       if (CI->hasFnAttr(Attribute::Cold)) {
4905         // If this is a cold call, we can sink the addressing calculation into
4906         // the cold path.  See optimizeCallInst
4907         bool OptForSize = OptSize ||
4908           llvm::shouldOptimizeForSize(CI->getParent(), PSI, BFI);
4909         if (!OptForSize)
4910           continue;
4911       }
4912 
4913       InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledOperand());
4914       if (!IA) return true;
4915 
4916       // If this is a memory operand, we're cool, otherwise bail out.
4917       if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI))
4918         return true;
4919       continue;
4920     }
4921 
4922     if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI, OptSize,
4923                           PSI, BFI, SeenInsts))
4924       return true;
4925   }
4926 
4927   return false;
4928 }
4929 
4930 /// Return true if Val is already known to be live at the use site that we're
4931 /// folding it into. If so, there is no cost to include it in the addressing
4932 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4933 /// instruction already.
4934 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4935                                                    Value *KnownLive2) {
4936   // If Val is either of the known-live values, we know it is live!
4937   if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4938     return true;
4939 
4940   // All values other than instructions and arguments (e.g. constants) are live.
4941   if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4942 
4943   // If Val is a constant sized alloca in the entry block, it is live, this is
4944   // true because it is just a reference to the stack/frame pointer, which is
4945   // live for the whole function.
4946   if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4947     if (AI->isStaticAlloca())
4948       return true;
4949 
4950   // Check to see if this value is already used in the memory instruction's
4951   // block.  If so, it's already live into the block at the very least, so we
4952   // can reasonably fold it.
4953   return Val->isUsedInBasicBlock(MemoryInst->getParent());
4954 }
4955 
4956 /// It is possible for the addressing mode of the machine to fold the specified
4957 /// instruction into a load or store that ultimately uses it.
4958 /// However, the specified instruction has multiple uses.
4959 /// Given this, it may actually increase register pressure to fold it
4960 /// into the load. For example, consider this code:
4961 ///
4962 ///     X = ...
4963 ///     Y = X+1
4964 ///     use(Y)   -> nonload/store
4965 ///     Z = Y+1
4966 ///     load Z
4967 ///
4968 /// In this case, Y has multiple uses, and can be folded into the load of Z
4969 /// (yielding load [X+2]).  However, doing this will cause both "X" and "X+1" to
4970 /// be live at the use(Y) line.  If we don't fold Y into load Z, we use one
4971 /// fewer register.  Since Y can't be folded into "use(Y)" we don't increase the
4972 /// number of computations either.
4973 ///
4974 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic.  If
4975 /// X was live across 'load Z' for other reasons, we actually *would* want to
4976 /// fold the addressing mode in the Z case.  This would make Y die earlier.
4977 bool AddressingModeMatcher::
4978 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4979                                      ExtAddrMode &AMAfter) {
4980   if (IgnoreProfitability) return true;
4981 
4982   // AMBefore is the addressing mode before this instruction was folded into it,
4983   // and AMAfter is the addressing mode after the instruction was folded.  Get
4984   // the set of registers referenced by AMAfter and subtract out those
4985   // referenced by AMBefore: this is the set of values which folding in this
4986   // address extends the lifetime of.
4987   //
4988   // Note that there are only two potential values being referenced here,
4989   // BaseReg and ScaleReg (global addresses are always available, as are any
4990   // folded immediates).
4991   Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4992 
4993   // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4994   // lifetime wasn't extended by adding this instruction.
4995   if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4996     BaseReg = nullptr;
4997   if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4998     ScaledReg = nullptr;
4999 
5000   // If folding this instruction (and it's subexprs) didn't extend any live
5001   // ranges, we're ok with it.
5002   if (!BaseReg && !ScaledReg)
5003     return true;
5004 
5005   // If all uses of this instruction can have the address mode sunk into them,
5006   // we can remove the addressing mode and effectively trade one live register
5007   // for another (at worst.)  In this context, folding an addressing mode into
5008   // the use is just a particularly nice way of sinking it.
5009   SmallVector<std::pair<Value *, Type *>, 16> MemoryUses;
5010   SmallPtrSet<Instruction*, 16> ConsideredInsts;
5011   if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI, OptSize,
5012                         PSI, BFI))
5013     return false;  // Has a non-memory, non-foldable use!
5014 
5015   // Now that we know that all uses of this instruction are part of a chain of
5016   // computation involving only operations that could theoretically be folded
5017   // into a memory use, loop over each of these memory operation uses and see
5018   // if they could  *actually* fold the instruction.  The assumption is that
5019   // addressing modes are cheap and that duplicating the computation involved
5020   // many times is worthwhile, even on a fastpath. For sinking candidates
5021   // (i.e. cold call sites), this serves as a way to prevent excessive code
5022   // growth since most architectures have some reasonable small and fast way to
5023   // compute an effective address.  (i.e LEA on x86)
5024   SmallVector<Instruction*, 32> MatchedAddrModeInsts;
5025   for (const std::pair<Value *, Type *> &Pair : MemoryUses) {
5026     Value *Address = Pair.first;
5027     Type *AddressAccessTy = Pair.second;
5028     unsigned AS = Address->getType()->getPointerAddressSpace();
5029 
5030     // Do a match against the root of this address, ignoring profitability. This
5031     // will tell us if the addressing mode for the memory operation will
5032     // *actually* cover the shared instruction.
5033     ExtAddrMode Result;
5034     std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
5035                                                                       0);
5036     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5037         TPT.getRestorationPoint();
5038     AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, TRI, LI, getDTFn,
5039                                   AddressAccessTy, AS, MemoryInst, Result,
5040                                   InsertedInsts, PromotedInsts, TPT,
5041                                   LargeOffsetGEP, OptSize, PSI, BFI);
5042     Matcher.IgnoreProfitability = true;
5043     bool Success = Matcher.matchAddr(Address, 0);
5044     (void)Success; assert(Success && "Couldn't select *anything*?");
5045 
5046     // The match was to check the profitability, the changes made are not
5047     // part of the original matcher. Therefore, they should be dropped
5048     // otherwise the original matcher will not present the right state.
5049     TPT.rollback(LastKnownGood);
5050 
5051     // If the match didn't cover I, then it won't be shared by it.
5052     if (!is_contained(MatchedAddrModeInsts, I))
5053       return false;
5054 
5055     MatchedAddrModeInsts.clear();
5056   }
5057 
5058   return true;
5059 }
5060 
5061 /// Return true if the specified values are defined in a
5062 /// different basic block than BB.
5063 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
5064   if (Instruction *I = dyn_cast<Instruction>(V))
5065     return I->getParent() != BB;
5066   return false;
5067 }
5068 
5069 /// Sink addressing mode computation immediate before MemoryInst if doing so
5070 /// can be done without increasing register pressure.  The need for the
5071 /// register pressure constraint means this can end up being an all or nothing
5072 /// decision for all uses of the same addressing computation.
5073 ///
5074 /// Load and Store Instructions often have addressing modes that can do
5075 /// significant amounts of computation. As such, instruction selection will try
5076 /// to get the load or store to do as much computation as possible for the
5077 /// program. The problem is that isel can only see within a single block. As
5078 /// such, we sink as much legal addressing mode work into the block as possible.
5079 ///
5080 /// This method is used to optimize both load/store and inline asms with memory
5081 /// operands.  It's also used to sink addressing computations feeding into cold
5082 /// call sites into their (cold) basic block.
5083 ///
5084 /// The motivation for handling sinking into cold blocks is that doing so can
5085 /// both enable other address mode sinking (by satisfying the register pressure
5086 /// constraint above), and reduce register pressure globally (by removing the
5087 /// addressing mode computation from the fast path entirely.).
5088 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
5089                                         Type *AccessTy, unsigned AddrSpace) {
5090   Value *Repl = Addr;
5091 
5092   // Try to collapse single-value PHI nodes.  This is necessary to undo
5093   // unprofitable PRE transformations.
5094   SmallVector<Value*, 8> worklist;
5095   SmallPtrSet<Value*, 16> Visited;
5096   worklist.push_back(Addr);
5097 
5098   // Use a worklist to iteratively look through PHI and select nodes, and
5099   // ensure that the addressing mode obtained from the non-PHI/select roots of
5100   // the graph are compatible.
5101   bool PhiOrSelectSeen = false;
5102   SmallVector<Instruction*, 16> AddrModeInsts;
5103   const SimplifyQuery SQ(*DL, TLInfo);
5104   AddressingModeCombiner AddrModes(SQ, Addr);
5105   TypePromotionTransaction TPT(RemovedInsts);
5106   TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5107       TPT.getRestorationPoint();
5108   while (!worklist.empty()) {
5109     Value *V = worklist.pop_back_val();
5110 
5111     // We allow traversing cyclic Phi nodes.
5112     // In case of success after this loop we ensure that traversing through
5113     // Phi nodes ends up with all cases to compute address of the form
5114     //    BaseGV + Base + Scale * Index + Offset
5115     // where Scale and Offset are constans and BaseGV, Base and Index
5116     // are exactly the same Values in all cases.
5117     // It means that BaseGV, Scale and Offset dominate our memory instruction
5118     // and have the same value as they had in address computation represented
5119     // as Phi. So we can safely sink address computation to memory instruction.
5120     if (!Visited.insert(V).second)
5121       continue;
5122 
5123     // For a PHI node, push all of its incoming values.
5124     if (PHINode *P = dyn_cast<PHINode>(V)) {
5125       append_range(worklist, P->incoming_values());
5126       PhiOrSelectSeen = true;
5127       continue;
5128     }
5129     // Similar for select.
5130     if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
5131       worklist.push_back(SI->getFalseValue());
5132       worklist.push_back(SI->getTrueValue());
5133       PhiOrSelectSeen = true;
5134       continue;
5135     }
5136 
5137     // For non-PHIs, determine the addressing mode being computed.  Note that
5138     // the result may differ depending on what other uses our candidate
5139     // addressing instructions might have.
5140     AddrModeInsts.clear();
5141     std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
5142                                                                       0);
5143     // Defer the query (and possible computation of) the dom tree to point of
5144     // actual use.  It's expected that most address matches don't actually need
5145     // the domtree.
5146     auto getDTFn = [MemoryInst, this]() -> const DominatorTree & {
5147       Function *F = MemoryInst->getParent()->getParent();
5148       return this->getDT(*F);
5149     };
5150     ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
5151         V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *LI, getDTFn,
5152         *TRI, InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP, OptSize, PSI,
5153         BFI.get());
5154 
5155     GetElementPtrInst *GEP = LargeOffsetGEP.first;
5156     if (GEP && !NewGEPBases.count(GEP)) {
5157       // If splitting the underlying data structure can reduce the offset of a
5158       // GEP, collect the GEP.  Skip the GEPs that are the new bases of
5159       // previously split data structures.
5160       LargeOffsetGEPMap[GEP->getPointerOperand()].push_back(LargeOffsetGEP);
5161       if (LargeOffsetGEPID.find(GEP) == LargeOffsetGEPID.end())
5162         LargeOffsetGEPID[GEP] = LargeOffsetGEPID.size();
5163     }
5164 
5165     NewAddrMode.OriginalValue = V;
5166     if (!AddrModes.addNewAddrMode(NewAddrMode))
5167       break;
5168   }
5169 
5170   // Try to combine the AddrModes we've collected. If we couldn't collect any,
5171   // or we have multiple but either couldn't combine them or combining them
5172   // wouldn't do anything useful, bail out now.
5173   if (!AddrModes.combineAddrModes()) {
5174     TPT.rollback(LastKnownGood);
5175     return false;
5176   }
5177   bool Modified = TPT.commit();
5178 
5179   // Get the combined AddrMode (or the only AddrMode, if we only had one).
5180   ExtAddrMode AddrMode = AddrModes.getAddrMode();
5181 
5182   // If all the instructions matched are already in this BB, don't do anything.
5183   // If we saw a Phi node then it is not local definitely, and if we saw a select
5184   // then we want to push the address calculation past it even if it's already
5185   // in this BB.
5186   if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) {
5187         return IsNonLocalValue(V, MemoryInst->getParent());
5188                   })) {
5189     LLVM_DEBUG(dbgs() << "CGP: Found      local addrmode: " << AddrMode
5190                       << "\n");
5191     return Modified;
5192   }
5193 
5194   // Insert this computation right after this user.  Since our caller is
5195   // scanning from the top of the BB to the bottom, reuse of the expr are
5196   // guaranteed to happen later.
5197   IRBuilder<> Builder(MemoryInst);
5198 
5199   // Now that we determined the addressing expression we want to use and know
5200   // that we have to sink it into this block.  Check to see if we have already
5201   // done this for some other load/store instr in this block.  If so, reuse
5202   // the computation.  Before attempting reuse, check if the address is valid
5203   // as it may have been erased.
5204 
5205   WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
5206 
5207   Value * SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
5208   if (SunkAddr) {
5209     LLVM_DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode
5210                       << " for " << *MemoryInst << "\n");
5211     if (SunkAddr->getType() != Addr->getType())
5212       SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
5213   } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() &&
5214                                    SubtargetInfo->addrSinkUsingGEPs())) {
5215     // By default, we use the GEP-based method when AA is used later. This
5216     // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
5217     LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
5218                       << " for " << *MemoryInst << "\n");
5219     Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
5220     Value *ResultPtr = nullptr, *ResultIndex = nullptr;
5221 
5222     // First, find the pointer.
5223     if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
5224       ResultPtr = AddrMode.BaseReg;
5225       AddrMode.BaseReg = nullptr;
5226     }
5227 
5228     if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
5229       // We can't add more than one pointer together, nor can we scale a
5230       // pointer (both of which seem meaningless).
5231       if (ResultPtr || AddrMode.Scale != 1)
5232         return Modified;
5233 
5234       ResultPtr = AddrMode.ScaledReg;
5235       AddrMode.Scale = 0;
5236     }
5237 
5238     // It is only safe to sign extend the BaseReg if we know that the math
5239     // required to create it did not overflow before we extend it. Since
5240     // the original IR value was tossed in favor of a constant back when
5241     // the AddrMode was created we need to bail out gracefully if widths
5242     // do not match instead of extending it.
5243     //
5244     // (See below for code to add the scale.)
5245     if (AddrMode.Scale) {
5246       Type *ScaledRegTy = AddrMode.ScaledReg->getType();
5247       if (cast<IntegerType>(IntPtrTy)->getBitWidth() >
5248           cast<IntegerType>(ScaledRegTy)->getBitWidth())
5249         return Modified;
5250     }
5251 
5252     if (AddrMode.BaseGV) {
5253       if (ResultPtr)
5254         return Modified;
5255 
5256       ResultPtr = AddrMode.BaseGV;
5257     }
5258 
5259     // If the real base value actually came from an inttoptr, then the matcher
5260     // will look through it and provide only the integer value. In that case,
5261     // use it here.
5262     if (!DL->isNonIntegralPointerType(Addr->getType())) {
5263       if (!ResultPtr && AddrMode.BaseReg) {
5264         ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(),
5265                                            "sunkaddr");
5266         AddrMode.BaseReg = nullptr;
5267       } else if (!ResultPtr && AddrMode.Scale == 1) {
5268         ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(),
5269                                            "sunkaddr");
5270         AddrMode.Scale = 0;
5271       }
5272     }
5273 
5274     if (!ResultPtr &&
5275         !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
5276       SunkAddr = Constant::getNullValue(Addr->getType());
5277     } else if (!ResultPtr) {
5278       return Modified;
5279     } else {
5280       Type *I8PtrTy =
5281           Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
5282       Type *I8Ty = Builder.getInt8Ty();
5283 
5284       // Start with the base register. Do this first so that subsequent address
5285       // matching finds it last, which will prevent it from trying to match it
5286       // as the scaled value in case it happens to be a mul. That would be
5287       // problematic if we've sunk a different mul for the scale, because then
5288       // we'd end up sinking both muls.
5289       if (AddrMode.BaseReg) {
5290         Value *V = AddrMode.BaseReg;
5291         if (V->getType() != IntPtrTy)
5292           V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
5293 
5294         ResultIndex = V;
5295       }
5296 
5297       // Add the scale value.
5298       if (AddrMode.Scale) {
5299         Value *V = AddrMode.ScaledReg;
5300         if (V->getType() == IntPtrTy) {
5301           // done.
5302         } else {
5303           assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <
5304                  cast<IntegerType>(V->getType())->getBitWidth() &&
5305                  "We can't transform if ScaledReg is too narrow");
5306           V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
5307         }
5308 
5309         if (AddrMode.Scale != 1)
5310           V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
5311                                 "sunkaddr");
5312         if (ResultIndex)
5313           ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
5314         else
5315           ResultIndex = V;
5316       }
5317 
5318       // Add in the Base Offset if present.
5319       if (AddrMode.BaseOffs) {
5320         Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
5321         if (ResultIndex) {
5322           // We need to add this separately from the scale above to help with
5323           // SDAG consecutive load/store merging.
5324           if (ResultPtr->getType() != I8PtrTy)
5325             ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
5326           ResultPtr =
5327               AddrMode.InBounds
5328                   ? Builder.CreateInBoundsGEP(I8Ty, ResultPtr, ResultIndex,
5329                                               "sunkaddr")
5330                   : Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
5331         }
5332 
5333         ResultIndex = V;
5334       }
5335 
5336       if (!ResultIndex) {
5337         SunkAddr = ResultPtr;
5338       } else {
5339         if (ResultPtr->getType() != I8PtrTy)
5340           ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
5341         SunkAddr =
5342             AddrMode.InBounds
5343                 ? Builder.CreateInBoundsGEP(I8Ty, ResultPtr, ResultIndex,
5344                                             "sunkaddr")
5345                 : Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
5346       }
5347 
5348       if (SunkAddr->getType() != Addr->getType())
5349         SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
5350     }
5351   } else {
5352     // We'd require a ptrtoint/inttoptr down the line, which we can't do for
5353     // non-integral pointers, so in that case bail out now.
5354     Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
5355     Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
5356     PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy);
5357     PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy);
5358     if (DL->isNonIntegralPointerType(Addr->getType()) ||
5359         (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) ||
5360         (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) ||
5361         (AddrMode.BaseGV &&
5362          DL->isNonIntegralPointerType(AddrMode.BaseGV->getType())))
5363       return Modified;
5364 
5365     LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
5366                       << " for " << *MemoryInst << "\n");
5367     Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
5368     Value *Result = nullptr;
5369 
5370     // Start with the base register. Do this first so that subsequent address
5371     // matching finds it last, which will prevent it from trying to match it
5372     // as the scaled value in case it happens to be a mul. That would be
5373     // problematic if we've sunk a different mul for the scale, because then
5374     // we'd end up sinking both muls.
5375     if (AddrMode.BaseReg) {
5376       Value *V = AddrMode.BaseReg;
5377       if (V->getType()->isPointerTy())
5378         V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
5379       if (V->getType() != IntPtrTy)
5380         V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
5381       Result = V;
5382     }
5383 
5384     // Add the scale value.
5385     if (AddrMode.Scale) {
5386       Value *V = AddrMode.ScaledReg;
5387       if (V->getType() == IntPtrTy) {
5388         // done.
5389       } else if (V->getType()->isPointerTy()) {
5390         V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
5391       } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
5392                  cast<IntegerType>(V->getType())->getBitWidth()) {
5393         V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
5394       } else {
5395         // It is only safe to sign extend the BaseReg if we know that the math
5396         // required to create it did not overflow before we extend it. Since
5397         // the original IR value was tossed in favor of a constant back when
5398         // the AddrMode was created we need to bail out gracefully if widths
5399         // do not match instead of extending it.
5400         Instruction *I = dyn_cast_or_null<Instruction>(Result);
5401         if (I && (Result != AddrMode.BaseReg))
5402           I->eraseFromParent();
5403         return Modified;
5404       }
5405       if (AddrMode.Scale != 1)
5406         V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
5407                               "sunkaddr");
5408       if (Result)
5409         Result = Builder.CreateAdd(Result, V, "sunkaddr");
5410       else
5411         Result = V;
5412     }
5413 
5414     // Add in the BaseGV if present.
5415     if (AddrMode.BaseGV) {
5416       Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
5417       if (Result)
5418         Result = Builder.CreateAdd(Result, V, "sunkaddr");
5419       else
5420         Result = V;
5421     }
5422 
5423     // Add in the Base Offset if present.
5424     if (AddrMode.BaseOffs) {
5425       Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
5426       if (Result)
5427         Result = Builder.CreateAdd(Result, V, "sunkaddr");
5428       else
5429         Result = V;
5430     }
5431 
5432     if (!Result)
5433       SunkAddr = Constant::getNullValue(Addr->getType());
5434     else
5435       SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
5436   }
5437 
5438   MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
5439   // Store the newly computed address into the cache. In the case we reused a
5440   // value, this should be idempotent.
5441   SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
5442 
5443   // If we have no uses, recursively delete the value and all dead instructions
5444   // using it.
5445   if (Repl->use_empty()) {
5446     resetIteratorIfInvalidatedWhileCalling(CurInstIterator->getParent(), [&]() {
5447       RecursivelyDeleteTriviallyDeadInstructions(
5448           Repl, TLInfo, nullptr,
5449           [&](Value *V) { removeAllAssertingVHReferences(V); });
5450     });
5451   }
5452   ++NumMemoryInsts;
5453   return true;
5454 }
5455 
5456 /// Rewrite GEP input to gather/scatter to enable SelectionDAGBuilder to find
5457 /// a uniform base to use for ISD::MGATHER/MSCATTER. SelectionDAGBuilder can
5458 /// only handle a 2 operand GEP in the same basic block or a splat constant
5459 /// vector. The 2 operands to the GEP must have a scalar pointer and a vector
5460 /// index.
5461 ///
5462 /// If the existing GEP has a vector base pointer that is splat, we can look
5463 /// through the splat to find the scalar pointer. If we can't find a scalar
5464 /// pointer there's nothing we can do.
5465 ///
5466 /// If we have a GEP with more than 2 indices where the middle indices are all
5467 /// zeroes, we can replace it with 2 GEPs where the second has 2 operands.
5468 ///
5469 /// If the final index isn't a vector or is a splat, we can emit a scalar GEP
5470 /// followed by a GEP with an all zeroes vector index. This will enable
5471 /// SelectionDAGBuilder to use the scalar GEP as the uniform base and have a
5472 /// zero index.
5473 bool CodeGenPrepare::optimizeGatherScatterInst(Instruction *MemoryInst,
5474                                                Value *Ptr) {
5475   Value *NewAddr;
5476 
5477   if (const auto *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
5478     // Don't optimize GEPs that don't have indices.
5479     if (!GEP->hasIndices())
5480       return false;
5481 
5482     // If the GEP and the gather/scatter aren't in the same BB, don't optimize.
5483     // FIXME: We should support this by sinking the GEP.
5484     if (MemoryInst->getParent() != GEP->getParent())
5485       return false;
5486 
5487     SmallVector<Value *, 2> Ops(GEP->operands());
5488 
5489     bool RewriteGEP = false;
5490 
5491     if (Ops[0]->getType()->isVectorTy()) {
5492       Ops[0] = getSplatValue(Ops[0]);
5493       if (!Ops[0])
5494         return false;
5495       RewriteGEP = true;
5496     }
5497 
5498     unsigned FinalIndex = Ops.size() - 1;
5499 
5500     // Ensure all but the last index is 0.
5501     // FIXME: This isn't strictly required. All that's required is that they are
5502     // all scalars or splats.
5503     for (unsigned i = 1; i < FinalIndex; ++i) {
5504       auto *C = dyn_cast<Constant>(Ops[i]);
5505       if (!C)
5506         return false;
5507       if (isa<VectorType>(C->getType()))
5508         C = C->getSplatValue();
5509       auto *CI = dyn_cast_or_null<ConstantInt>(C);
5510       if (!CI || !CI->isZero())
5511         return false;
5512       // Scalarize the index if needed.
5513       Ops[i] = CI;
5514     }
5515 
5516     // Try to scalarize the final index.
5517     if (Ops[FinalIndex]->getType()->isVectorTy()) {
5518       if (Value *V = getSplatValue(Ops[FinalIndex])) {
5519         auto *C = dyn_cast<ConstantInt>(V);
5520         // Don't scalarize all zeros vector.
5521         if (!C || !C->isZero()) {
5522           Ops[FinalIndex] = V;
5523           RewriteGEP = true;
5524         }
5525       }
5526     }
5527 
5528     // If we made any changes or the we have extra operands, we need to generate
5529     // new instructions.
5530     if (!RewriteGEP && Ops.size() == 2)
5531       return false;
5532 
5533     auto NumElts = cast<VectorType>(Ptr->getType())->getElementCount();
5534 
5535     IRBuilder<> Builder(MemoryInst);
5536 
5537     Type *SourceTy = GEP->getSourceElementType();
5538     Type *ScalarIndexTy = DL->getIndexType(Ops[0]->getType()->getScalarType());
5539 
5540     // If the final index isn't a vector, emit a scalar GEP containing all ops
5541     // and a vector GEP with all zeroes final index.
5542     if (!Ops[FinalIndex]->getType()->isVectorTy()) {
5543       NewAddr = Builder.CreateGEP(SourceTy, Ops[0],
5544                                   makeArrayRef(Ops).drop_front());
5545       auto *IndexTy = VectorType::get(ScalarIndexTy, NumElts);
5546       auto *SecondTy = GetElementPtrInst::getIndexedType(
5547           SourceTy, makeArrayRef(Ops).drop_front());
5548       NewAddr =
5549           Builder.CreateGEP(SecondTy, NewAddr, Constant::getNullValue(IndexTy));
5550     } else {
5551       Value *Base = Ops[0];
5552       Value *Index = Ops[FinalIndex];
5553 
5554       // Create a scalar GEP if there are more than 2 operands.
5555       if (Ops.size() != 2) {
5556         // Replace the last index with 0.
5557         Ops[FinalIndex] = Constant::getNullValue(ScalarIndexTy);
5558         Base = Builder.CreateGEP(SourceTy, Base,
5559                                  makeArrayRef(Ops).drop_front());
5560         SourceTy = GetElementPtrInst::getIndexedType(
5561             SourceTy, makeArrayRef(Ops).drop_front());
5562       }
5563 
5564       // Now create the GEP with scalar pointer and vector index.
5565       NewAddr = Builder.CreateGEP(SourceTy, Base, Index);
5566     }
5567   } else if (!isa<Constant>(Ptr)) {
5568     // Not a GEP, maybe its a splat and we can create a GEP to enable
5569     // SelectionDAGBuilder to use it as a uniform base.
5570     Value *V = getSplatValue(Ptr);
5571     if (!V)
5572       return false;
5573 
5574     auto NumElts = cast<VectorType>(Ptr->getType())->getElementCount();
5575 
5576     IRBuilder<> Builder(MemoryInst);
5577 
5578     // Emit a vector GEP with a scalar pointer and all 0s vector index.
5579     Type *ScalarIndexTy = DL->getIndexType(V->getType()->getScalarType());
5580     auto *IndexTy = VectorType::get(ScalarIndexTy, NumElts);
5581     Type *ScalarTy;
5582     if (cast<IntrinsicInst>(MemoryInst)->getIntrinsicID() ==
5583         Intrinsic::masked_gather) {
5584       ScalarTy = MemoryInst->getType()->getScalarType();
5585     } else {
5586       assert(cast<IntrinsicInst>(MemoryInst)->getIntrinsicID() ==
5587              Intrinsic::masked_scatter);
5588       ScalarTy = MemoryInst->getOperand(0)->getType()->getScalarType();
5589     }
5590     NewAddr = Builder.CreateGEP(ScalarTy, V, Constant::getNullValue(IndexTy));
5591   } else {
5592     // Constant, SelectionDAGBuilder knows to check if its a splat.
5593     return false;
5594   }
5595 
5596   MemoryInst->replaceUsesOfWith(Ptr, NewAddr);
5597 
5598   // If we have no uses, recursively delete the value and all dead instructions
5599   // using it.
5600   if (Ptr->use_empty())
5601     RecursivelyDeleteTriviallyDeadInstructions(
5602         Ptr, TLInfo, nullptr,
5603         [&](Value *V) { removeAllAssertingVHReferences(V); });
5604 
5605   return true;
5606 }
5607 
5608 /// If there are any memory operands, use OptimizeMemoryInst to sink their
5609 /// address computing into the block when possible / profitable.
5610 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
5611   bool MadeChange = false;
5612 
5613   const TargetRegisterInfo *TRI =
5614       TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
5615   TargetLowering::AsmOperandInfoVector TargetConstraints =
5616       TLI->ParseConstraints(*DL, TRI, *CS);
5617   unsigned ArgNo = 0;
5618   for (TargetLowering::AsmOperandInfo &OpInfo : TargetConstraints) {
5619     // Compute the constraint code and ConstraintType to use.
5620     TLI->ComputeConstraintToUse(OpInfo, SDValue());
5621 
5622     if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5623         OpInfo.isIndirect) {
5624       Value *OpVal = CS->getArgOperand(ArgNo++);
5625       MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
5626     } else if (OpInfo.Type == InlineAsm::isInput)
5627       ArgNo++;
5628   }
5629 
5630   return MadeChange;
5631 }
5632 
5633 /// Check if all the uses of \p Val are equivalent (or free) zero or
5634 /// sign extensions.
5635 static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
5636   assert(!Val->use_empty() && "Input must have at least one use");
5637   const Instruction *FirstUser = cast<Instruction>(*Val->user_begin());
5638   bool IsSExt = isa<SExtInst>(FirstUser);
5639   Type *ExtTy = FirstUser->getType();
5640   for (const User *U : Val->users()) {
5641     const Instruction *UI = cast<Instruction>(U);
5642     if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
5643       return false;
5644     Type *CurTy = UI->getType();
5645     // Same input and output types: Same instruction after CSE.
5646     if (CurTy == ExtTy)
5647       continue;
5648 
5649     // If IsSExt is true, we are in this situation:
5650     // a = Val
5651     // b = sext ty1 a to ty2
5652     // c = sext ty1 a to ty3
5653     // Assuming ty2 is shorter than ty3, this could be turned into:
5654     // a = Val
5655     // b = sext ty1 a to ty2
5656     // c = sext ty2 b to ty3
5657     // However, the last sext is not free.
5658     if (IsSExt)
5659       return false;
5660 
5661     // This is a ZExt, maybe this is free to extend from one type to another.
5662     // In that case, we would not account for a different use.
5663     Type *NarrowTy;
5664     Type *LargeTy;
5665     if (ExtTy->getScalarType()->getIntegerBitWidth() >
5666         CurTy->getScalarType()->getIntegerBitWidth()) {
5667       NarrowTy = CurTy;
5668       LargeTy = ExtTy;
5669     } else {
5670       NarrowTy = ExtTy;
5671       LargeTy = CurTy;
5672     }
5673 
5674     if (!TLI.isZExtFree(NarrowTy, LargeTy))
5675       return false;
5676   }
5677   // All uses are the same or can be derived from one another for free.
5678   return true;
5679 }
5680 
5681 /// Try to speculatively promote extensions in \p Exts and continue
5682 /// promoting through newly promoted operands recursively as far as doing so is
5683 /// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
5684 /// When some promotion happened, \p TPT contains the proper state to revert
5685 /// them.
5686 ///
5687 /// \return true if some promotion happened, false otherwise.
5688 bool CodeGenPrepare::tryToPromoteExts(
5689     TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
5690     SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
5691     unsigned CreatedInstsCost) {
5692   bool Promoted = false;
5693 
5694   // Iterate over all the extensions to try to promote them.
5695   for (auto *I : Exts) {
5696     // Early check if we directly have ext(load).
5697     if (isa<LoadInst>(I->getOperand(0))) {
5698       ProfitablyMovedExts.push_back(I);
5699       continue;
5700     }
5701 
5702     // Check whether or not we want to do any promotion.  The reason we have
5703     // this check inside the for loop is to catch the case where an extension
5704     // is directly fed by a load because in such case the extension can be moved
5705     // up without any promotion on its operands.
5706     if (!TLI->enableExtLdPromotion() || DisableExtLdPromotion)
5707       return false;
5708 
5709     // Get the action to perform the promotion.
5710     TypePromotionHelper::Action TPH =
5711         TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
5712     // Check if we can promote.
5713     if (!TPH) {
5714       // Save the current extension as we cannot move up through its operand.
5715       ProfitablyMovedExts.push_back(I);
5716       continue;
5717     }
5718 
5719     // Save the current state.
5720     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5721         TPT.getRestorationPoint();
5722     SmallVector<Instruction *, 4> NewExts;
5723     unsigned NewCreatedInstsCost = 0;
5724     unsigned ExtCost = !TLI->isExtFree(I);
5725     // Promote.
5726     Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
5727                              &NewExts, nullptr, *TLI);
5728     assert(PromotedVal &&
5729            "TypePromotionHelper should have filtered out those cases");
5730 
5731     // We would be able to merge only one extension in a load.
5732     // Therefore, if we have more than 1 new extension we heuristically
5733     // cut this search path, because it means we degrade the code quality.
5734     // With exactly 2, the transformation is neutral, because we will merge
5735     // one extension but leave one. However, we optimistically keep going,
5736     // because the new extension may be removed too.
5737     long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
5738     // FIXME: It would be possible to propagate a negative value instead of
5739     // conservatively ceiling it to 0.
5740     TotalCreatedInstsCost =
5741         std::max((long long)0, (TotalCreatedInstsCost - ExtCost));
5742     if (!StressExtLdPromotion &&
5743         (TotalCreatedInstsCost > 1 ||
5744          !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
5745       // This promotion is not profitable, rollback to the previous state, and
5746       // save the current extension in ProfitablyMovedExts as the latest
5747       // speculative promotion turned out to be unprofitable.
5748       TPT.rollback(LastKnownGood);
5749       ProfitablyMovedExts.push_back(I);
5750       continue;
5751     }
5752     // Continue promoting NewExts as far as doing so is profitable.
5753     SmallVector<Instruction *, 2> NewlyMovedExts;
5754     (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost);
5755     bool NewPromoted = false;
5756     for (auto *ExtInst : NewlyMovedExts) {
5757       Instruction *MovedExt = cast<Instruction>(ExtInst);
5758       Value *ExtOperand = MovedExt->getOperand(0);
5759       // If we have reached to a load, we need this extra profitability check
5760       // as it could potentially be merged into an ext(load).
5761       if (isa<LoadInst>(ExtOperand) &&
5762           !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
5763             (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI))))
5764         continue;
5765 
5766       ProfitablyMovedExts.push_back(MovedExt);
5767       NewPromoted = true;
5768     }
5769 
5770     // If none of speculative promotions for NewExts is profitable, rollback
5771     // and save the current extension (I) as the last profitable extension.
5772     if (!NewPromoted) {
5773       TPT.rollback(LastKnownGood);
5774       ProfitablyMovedExts.push_back(I);
5775       continue;
5776     }
5777     // The promotion is profitable.
5778     Promoted = true;
5779   }
5780   return Promoted;
5781 }
5782 
5783 /// Merging redundant sexts when one is dominating the other.
5784 bool CodeGenPrepare::mergeSExts(Function &F) {
5785   bool Changed = false;
5786   for (auto &Entry : ValToSExtendedUses) {
5787     SExts &Insts = Entry.second;
5788     SExts CurPts;
5789     for (Instruction *Inst : Insts) {
5790       if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) ||
5791           Inst->getOperand(0) != Entry.first)
5792         continue;
5793       bool inserted = false;
5794       for (auto &Pt : CurPts) {
5795         if (getDT(F).dominates(Inst, Pt)) {
5796           Pt->replaceAllUsesWith(Inst);
5797           RemovedInsts.insert(Pt);
5798           Pt->removeFromParent();
5799           Pt = Inst;
5800           inserted = true;
5801           Changed = true;
5802           break;
5803         }
5804         if (!getDT(F).dominates(Pt, Inst))
5805           // Give up if we need to merge in a common dominator as the
5806           // experiments show it is not profitable.
5807           continue;
5808         Inst->replaceAllUsesWith(Pt);
5809         RemovedInsts.insert(Inst);
5810         Inst->removeFromParent();
5811         inserted = true;
5812         Changed = true;
5813         break;
5814       }
5815       if (!inserted)
5816         CurPts.push_back(Inst);
5817     }
5818   }
5819   return Changed;
5820 }
5821 
5822 // Splitting large data structures so that the GEPs accessing them can have
5823 // smaller offsets so that they can be sunk to the same blocks as their users.
5824 // For example, a large struct starting from %base is split into two parts
5825 // where the second part starts from %new_base.
5826 //
5827 // Before:
5828 // BB0:
5829 //   %base     =
5830 //
5831 // BB1:
5832 //   %gep0     = gep %base, off0
5833 //   %gep1     = gep %base, off1
5834 //   %gep2     = gep %base, off2
5835 //
5836 // BB2:
5837 //   %load1    = load %gep0
5838 //   %load2    = load %gep1
5839 //   %load3    = load %gep2
5840 //
5841 // After:
5842 // BB0:
5843 //   %base     =
5844 //   %new_base = gep %base, off0
5845 //
5846 // BB1:
5847 //   %new_gep0 = %new_base
5848 //   %new_gep1 = gep %new_base, off1 - off0
5849 //   %new_gep2 = gep %new_base, off2 - off0
5850 //
5851 // BB2:
5852 //   %load1    = load i32, i32* %new_gep0
5853 //   %load2    = load i32, i32* %new_gep1
5854 //   %load3    = load i32, i32* %new_gep2
5855 //
5856 // %new_gep1 and %new_gep2 can be sunk to BB2 now after the splitting because
5857 // their offsets are smaller enough to fit into the addressing mode.
5858 bool CodeGenPrepare::splitLargeGEPOffsets() {
5859   bool Changed = false;
5860   for (auto &Entry : LargeOffsetGEPMap) {
5861     Value *OldBase = Entry.first;
5862     SmallVectorImpl<std::pair<AssertingVH<GetElementPtrInst>, int64_t>>
5863         &LargeOffsetGEPs = Entry.second;
5864     auto compareGEPOffset =
5865         [&](const std::pair<GetElementPtrInst *, int64_t> &LHS,
5866             const std::pair<GetElementPtrInst *, int64_t> &RHS) {
5867           if (LHS.first == RHS.first)
5868             return false;
5869           if (LHS.second != RHS.second)
5870             return LHS.second < RHS.second;
5871           return LargeOffsetGEPID[LHS.first] < LargeOffsetGEPID[RHS.first];
5872         };
5873     // Sorting all the GEPs of the same data structures based on the offsets.
5874     llvm::sort(LargeOffsetGEPs, compareGEPOffset);
5875     LargeOffsetGEPs.erase(
5876         std::unique(LargeOffsetGEPs.begin(), LargeOffsetGEPs.end()),
5877         LargeOffsetGEPs.end());
5878     // Skip if all the GEPs have the same offsets.
5879     if (LargeOffsetGEPs.front().second == LargeOffsetGEPs.back().second)
5880       continue;
5881     GetElementPtrInst *BaseGEP = LargeOffsetGEPs.begin()->first;
5882     int64_t BaseOffset = LargeOffsetGEPs.begin()->second;
5883     Value *NewBaseGEP = nullptr;
5884 
5885     auto *LargeOffsetGEP = LargeOffsetGEPs.begin();
5886     while (LargeOffsetGEP != LargeOffsetGEPs.end()) {
5887       GetElementPtrInst *GEP = LargeOffsetGEP->first;
5888       int64_t Offset = LargeOffsetGEP->second;
5889       if (Offset != BaseOffset) {
5890         TargetLowering::AddrMode AddrMode;
5891         AddrMode.BaseOffs = Offset - BaseOffset;
5892         // The result type of the GEP might not be the type of the memory
5893         // access.
5894         if (!TLI->isLegalAddressingMode(*DL, AddrMode,
5895                                         GEP->getResultElementType(),
5896                                         GEP->getAddressSpace())) {
5897           // We need to create a new base if the offset to the current base is
5898           // too large to fit into the addressing mode. So, a very large struct
5899           // may be split into several parts.
5900           BaseGEP = GEP;
5901           BaseOffset = Offset;
5902           NewBaseGEP = nullptr;
5903         }
5904       }
5905 
5906       // Generate a new GEP to replace the current one.
5907       LLVMContext &Ctx = GEP->getContext();
5908       Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
5909       Type *I8PtrTy =
5910           Type::getInt8PtrTy(Ctx, GEP->getType()->getPointerAddressSpace());
5911       Type *I8Ty = Type::getInt8Ty(Ctx);
5912 
5913       if (!NewBaseGEP) {
5914         // Create a new base if we don't have one yet.  Find the insertion
5915         // pointer for the new base first.
5916         BasicBlock::iterator NewBaseInsertPt;
5917         BasicBlock *NewBaseInsertBB;
5918         if (auto *BaseI = dyn_cast<Instruction>(OldBase)) {
5919           // If the base of the struct is an instruction, the new base will be
5920           // inserted close to it.
5921           NewBaseInsertBB = BaseI->getParent();
5922           if (isa<PHINode>(BaseI))
5923             NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5924           else if (InvokeInst *Invoke = dyn_cast<InvokeInst>(BaseI)) {
5925             NewBaseInsertBB =
5926                 SplitEdge(NewBaseInsertBB, Invoke->getNormalDest());
5927             NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5928           } else
5929             NewBaseInsertPt = std::next(BaseI->getIterator());
5930         } else {
5931           // If the current base is an argument or global value, the new base
5932           // will be inserted to the entry block.
5933           NewBaseInsertBB = &BaseGEP->getFunction()->getEntryBlock();
5934           NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5935         }
5936         IRBuilder<> NewBaseBuilder(NewBaseInsertBB, NewBaseInsertPt);
5937         // Create a new base.
5938         Value *BaseIndex = ConstantInt::get(IntPtrTy, BaseOffset);
5939         NewBaseGEP = OldBase;
5940         if (NewBaseGEP->getType() != I8PtrTy)
5941           NewBaseGEP = NewBaseBuilder.CreatePointerCast(NewBaseGEP, I8PtrTy);
5942         NewBaseGEP =
5943             NewBaseBuilder.CreateGEP(I8Ty, NewBaseGEP, BaseIndex, "splitgep");
5944         NewGEPBases.insert(NewBaseGEP);
5945       }
5946 
5947       IRBuilder<> Builder(GEP);
5948       Value *NewGEP = NewBaseGEP;
5949       if (Offset == BaseOffset) {
5950         if (GEP->getType() != I8PtrTy)
5951           NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
5952       } else {
5953         // Calculate the new offset for the new GEP.
5954         Value *Index = ConstantInt::get(IntPtrTy, Offset - BaseOffset);
5955         NewGEP = Builder.CreateGEP(I8Ty, NewBaseGEP, Index);
5956 
5957         if (GEP->getType() != I8PtrTy)
5958           NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
5959       }
5960       GEP->replaceAllUsesWith(NewGEP);
5961       LargeOffsetGEPID.erase(GEP);
5962       LargeOffsetGEP = LargeOffsetGEPs.erase(LargeOffsetGEP);
5963       GEP->eraseFromParent();
5964       Changed = true;
5965     }
5966   }
5967   return Changed;
5968 }
5969 
5970 bool CodeGenPrepare::optimizePhiType(
5971     PHINode *I, SmallPtrSetImpl<PHINode *> &Visited,
5972     SmallPtrSetImpl<Instruction *> &DeletedInstrs) {
5973   // We are looking for a collection on interconnected phi nodes that together
5974   // only use loads/bitcasts and are used by stores/bitcasts, and the bitcasts
5975   // are of the same type. Convert the whole set of nodes to the type of the
5976   // bitcast.
5977   Type *PhiTy = I->getType();
5978   Type *ConvertTy = nullptr;
5979   if (Visited.count(I) ||
5980       (!I->getType()->isIntegerTy() && !I->getType()->isFloatingPointTy()))
5981     return false;
5982 
5983   SmallVector<Instruction *, 4> Worklist;
5984   Worklist.push_back(cast<Instruction>(I));
5985   SmallPtrSet<PHINode *, 4> PhiNodes;
5986   PhiNodes.insert(I);
5987   Visited.insert(I);
5988   SmallPtrSet<Instruction *, 4> Defs;
5989   SmallPtrSet<Instruction *, 4> Uses;
5990   // This works by adding extra bitcasts between load/stores and removing
5991   // existing bicasts. If we have a phi(bitcast(load)) or a store(bitcast(phi))
5992   // we can get in the situation where we remove a bitcast in one iteration
5993   // just to add it again in the next. We need to ensure that at least one
5994   // bitcast we remove are anchored to something that will not change back.
5995   bool AnyAnchored = false;
5996 
5997   while (!Worklist.empty()) {
5998     Instruction *II = Worklist.pop_back_val();
5999 
6000     if (auto *Phi = dyn_cast<PHINode>(II)) {
6001       // Handle Defs, which might also be PHI's
6002       for (Value *V : Phi->incoming_values()) {
6003         if (auto *OpPhi = dyn_cast<PHINode>(V)) {
6004           if (!PhiNodes.count(OpPhi)) {
6005             if (Visited.count(OpPhi))
6006               return false;
6007             PhiNodes.insert(OpPhi);
6008             Visited.insert(OpPhi);
6009             Worklist.push_back(OpPhi);
6010           }
6011         } else if (auto *OpLoad = dyn_cast<LoadInst>(V)) {
6012           if (!OpLoad->isSimple())
6013             return false;
6014           if (!Defs.count(OpLoad)) {
6015             Defs.insert(OpLoad);
6016             Worklist.push_back(OpLoad);
6017           }
6018         } else if (auto *OpEx = dyn_cast<ExtractElementInst>(V)) {
6019           if (!Defs.count(OpEx)) {
6020             Defs.insert(OpEx);
6021             Worklist.push_back(OpEx);
6022           }
6023         } else if (auto *OpBC = dyn_cast<BitCastInst>(V)) {
6024           if (!ConvertTy)
6025             ConvertTy = OpBC->getOperand(0)->getType();
6026           if (OpBC->getOperand(0)->getType() != ConvertTy)
6027             return false;
6028           if (!Defs.count(OpBC)) {
6029             Defs.insert(OpBC);
6030             Worklist.push_back(OpBC);
6031             AnyAnchored |= !isa<LoadInst>(OpBC->getOperand(0)) &&
6032                            !isa<ExtractElementInst>(OpBC->getOperand(0));
6033           }
6034         } else if (!isa<UndefValue>(V)) {
6035           return false;
6036         }
6037       }
6038     }
6039 
6040     // Handle uses which might also be phi's
6041     for (User *V : II->users()) {
6042       if (auto *OpPhi = dyn_cast<PHINode>(V)) {
6043         if (!PhiNodes.count(OpPhi)) {
6044           if (Visited.count(OpPhi))
6045             return false;
6046           PhiNodes.insert(OpPhi);
6047           Visited.insert(OpPhi);
6048           Worklist.push_back(OpPhi);
6049         }
6050       } else if (auto *OpStore = dyn_cast<StoreInst>(V)) {
6051         if (!OpStore->isSimple() || OpStore->getOperand(0) != II)
6052           return false;
6053         Uses.insert(OpStore);
6054       } else if (auto *OpBC = dyn_cast<BitCastInst>(V)) {
6055         if (!ConvertTy)
6056           ConvertTy = OpBC->getType();
6057         if (OpBC->getType() != ConvertTy)
6058           return false;
6059         Uses.insert(OpBC);
6060         AnyAnchored |=
6061             any_of(OpBC->users(), [](User *U) { return !isa<StoreInst>(U); });
6062       } else {
6063         return false;
6064       }
6065     }
6066   }
6067 
6068   if (!ConvertTy || !AnyAnchored || !TLI->shouldConvertPhiType(PhiTy, ConvertTy))
6069     return false;
6070 
6071   LLVM_DEBUG(dbgs() << "Converting " << *I << "\n  and connected nodes to "
6072                     << *ConvertTy << "\n");
6073 
6074   // Create all the new phi nodes of the new type, and bitcast any loads to the
6075   // correct type.
6076   ValueToValueMap ValMap;
6077   ValMap[UndefValue::get(PhiTy)] = UndefValue::get(ConvertTy);
6078   for (Instruction *D : Defs) {
6079     if (isa<BitCastInst>(D)) {
6080       ValMap[D] = D->getOperand(0);
6081       DeletedInstrs.insert(D);
6082     } else {
6083       ValMap[D] =
6084           new BitCastInst(D, ConvertTy, D->getName() + ".bc", D->getNextNode());
6085     }
6086   }
6087   for (PHINode *Phi : PhiNodes)
6088     ValMap[Phi] = PHINode::Create(ConvertTy, Phi->getNumIncomingValues(),
6089                                   Phi->getName() + ".tc", Phi);
6090   // Pipe together all the PhiNodes.
6091   for (PHINode *Phi : PhiNodes) {
6092     PHINode *NewPhi = cast<PHINode>(ValMap[Phi]);
6093     for (int i = 0, e = Phi->getNumIncomingValues(); i < e; i++)
6094       NewPhi->addIncoming(ValMap[Phi->getIncomingValue(i)],
6095                           Phi->getIncomingBlock(i));
6096     Visited.insert(NewPhi);
6097   }
6098   // And finally pipe up the stores and bitcasts
6099   for (Instruction *U : Uses) {
6100     if (isa<BitCastInst>(U)) {
6101       DeletedInstrs.insert(U);
6102       U->replaceAllUsesWith(ValMap[U->getOperand(0)]);
6103     } else {
6104       U->setOperand(0,
6105                     new BitCastInst(ValMap[U->getOperand(0)], PhiTy, "bc", U));
6106     }
6107   }
6108 
6109   // Save the removed phis to be deleted later.
6110   for (PHINode *Phi : PhiNodes)
6111     DeletedInstrs.insert(Phi);
6112   return true;
6113 }
6114 
6115 bool CodeGenPrepare::optimizePhiTypes(Function &F) {
6116   if (!OptimizePhiTypes)
6117     return false;
6118 
6119   bool Changed = false;
6120   SmallPtrSet<PHINode *, 4> Visited;
6121   SmallPtrSet<Instruction *, 4> DeletedInstrs;
6122 
6123   // Attempt to optimize all the phis in the functions to the correct type.
6124   for (auto &BB : F)
6125     for (auto &Phi : BB.phis())
6126       Changed |= optimizePhiType(&Phi, Visited, DeletedInstrs);
6127 
6128   // Remove any old phi's that have been converted.
6129   for (auto *I : DeletedInstrs) {
6130     I->replaceAllUsesWith(UndefValue::get(I->getType()));
6131     I->eraseFromParent();
6132   }
6133 
6134   return Changed;
6135 }
6136 
6137 /// Return true, if an ext(load) can be formed from an extension in
6138 /// \p MovedExts.
6139 bool CodeGenPrepare::canFormExtLd(
6140     const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
6141     Instruction *&Inst, bool HasPromoted) {
6142   for (auto *MovedExtInst : MovedExts) {
6143     if (isa<LoadInst>(MovedExtInst->getOperand(0))) {
6144       LI = cast<LoadInst>(MovedExtInst->getOperand(0));
6145       Inst = MovedExtInst;
6146       break;
6147     }
6148   }
6149   if (!LI)
6150     return false;
6151 
6152   // If they're already in the same block, there's nothing to do.
6153   // Make the cheap checks first if we did not promote.
6154   // If we promoted, we need to check if it is indeed profitable.
6155   if (!HasPromoted && LI->getParent() == Inst->getParent())
6156     return false;
6157 
6158   return TLI->isExtLoad(LI, Inst, *DL);
6159 }
6160 
6161 /// Move a zext or sext fed by a load into the same basic block as the load,
6162 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
6163 /// extend into the load.
6164 ///
6165 /// E.g.,
6166 /// \code
6167 /// %ld = load i32* %addr
6168 /// %add = add nuw i32 %ld, 4
6169 /// %zext = zext i32 %add to i64
6170 // \endcode
6171 /// =>
6172 /// \code
6173 /// %ld = load i32* %addr
6174 /// %zext = zext i32 %ld to i64
6175 /// %add = add nuw i64 %zext, 4
6176 /// \encode
6177 /// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
6178 /// allow us to match zext(load i32*) to i64.
6179 ///
6180 /// Also, try to promote the computations used to obtain a sign extended
6181 /// value used into memory accesses.
6182 /// E.g.,
6183 /// \code
6184 /// a = add nsw i32 b, 3
6185 /// d = sext i32 a to i64
6186 /// e = getelementptr ..., i64 d
6187 /// \endcode
6188 /// =>
6189 /// \code
6190 /// f = sext i32 b to i64
6191 /// a = add nsw i64 f, 3
6192 /// e = getelementptr ..., i64 a
6193 /// \endcode
6194 ///
6195 /// \p Inst[in/out] the extension may be modified during the process if some
6196 /// promotions apply.
6197 bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
6198   bool AllowPromotionWithoutCommonHeader = false;
6199   /// See if it is an interesting sext operations for the address type
6200   /// promotion before trying to promote it, e.g., the ones with the right
6201   /// type and used in memory accesses.
6202   bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
6203       *Inst, AllowPromotionWithoutCommonHeader);
6204   TypePromotionTransaction TPT(RemovedInsts);
6205   TypePromotionTransaction::ConstRestorationPt LastKnownGood =
6206       TPT.getRestorationPoint();
6207   SmallVector<Instruction *, 1> Exts;
6208   SmallVector<Instruction *, 2> SpeculativelyMovedExts;
6209   Exts.push_back(Inst);
6210 
6211   bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts);
6212 
6213   // Look for a load being extended.
6214   LoadInst *LI = nullptr;
6215   Instruction *ExtFedByLoad;
6216 
6217   // Try to promote a chain of computation if it allows to form an extended
6218   // load.
6219   if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) {
6220     assert(LI && ExtFedByLoad && "Expect a valid load and extension");
6221     TPT.commit();
6222     // Move the extend into the same block as the load.
6223     ExtFedByLoad->moveAfter(LI);
6224     ++NumExtsMoved;
6225     Inst = ExtFedByLoad;
6226     return true;
6227   }
6228 
6229   // Continue promoting SExts if known as considerable depending on targets.
6230   if (ATPConsiderable &&
6231       performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
6232                                   HasPromoted, TPT, SpeculativelyMovedExts))
6233     return true;
6234 
6235   TPT.rollback(LastKnownGood);
6236   return false;
6237 }
6238 
6239 // Perform address type promotion if doing so is profitable.
6240 // If AllowPromotionWithoutCommonHeader == false, we should find other sext
6241 // instructions that sign extended the same initial value. However, if
6242 // AllowPromotionWithoutCommonHeader == true, we expect promoting the
6243 // extension is just profitable.
6244 bool CodeGenPrepare::performAddressTypePromotion(
6245     Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
6246     bool HasPromoted, TypePromotionTransaction &TPT,
6247     SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
6248   bool Promoted = false;
6249   SmallPtrSet<Instruction *, 1> UnhandledExts;
6250   bool AllSeenFirst = true;
6251   for (auto *I : SpeculativelyMovedExts) {
6252     Value *HeadOfChain = I->getOperand(0);
6253     DenseMap<Value *, Instruction *>::iterator AlreadySeen =
6254         SeenChainsForSExt.find(HeadOfChain);
6255     // If there is an unhandled SExt which has the same header, try to promote
6256     // it as well.
6257     if (AlreadySeen != SeenChainsForSExt.end()) {
6258       if (AlreadySeen->second != nullptr)
6259         UnhandledExts.insert(AlreadySeen->second);
6260       AllSeenFirst = false;
6261     }
6262   }
6263 
6264   if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
6265                         SpeculativelyMovedExts.size() == 1)) {
6266     TPT.commit();
6267     if (HasPromoted)
6268       Promoted = true;
6269     for (auto *I : SpeculativelyMovedExts) {
6270       Value *HeadOfChain = I->getOperand(0);
6271       SeenChainsForSExt[HeadOfChain] = nullptr;
6272       ValToSExtendedUses[HeadOfChain].push_back(I);
6273     }
6274     // Update Inst as promotion happen.
6275     Inst = SpeculativelyMovedExts.pop_back_val();
6276   } else {
6277     // This is the first chain visited from the header, keep the current chain
6278     // as unhandled. Defer to promote this until we encounter another SExt
6279     // chain derived from the same header.
6280     for (auto *I : SpeculativelyMovedExts) {
6281       Value *HeadOfChain = I->getOperand(0);
6282       SeenChainsForSExt[HeadOfChain] = Inst;
6283     }
6284     return false;
6285   }
6286 
6287   if (!AllSeenFirst && !UnhandledExts.empty())
6288     for (auto *VisitedSExt : UnhandledExts) {
6289       if (RemovedInsts.count(VisitedSExt))
6290         continue;
6291       TypePromotionTransaction TPT(RemovedInsts);
6292       SmallVector<Instruction *, 1> Exts;
6293       SmallVector<Instruction *, 2> Chains;
6294       Exts.push_back(VisitedSExt);
6295       bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains);
6296       TPT.commit();
6297       if (HasPromoted)
6298         Promoted = true;
6299       for (auto *I : Chains) {
6300         Value *HeadOfChain = I->getOperand(0);
6301         // Mark this as handled.
6302         SeenChainsForSExt[HeadOfChain] = nullptr;
6303         ValToSExtendedUses[HeadOfChain].push_back(I);
6304       }
6305     }
6306   return Promoted;
6307 }
6308 
6309 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
6310   BasicBlock *DefBB = I->getParent();
6311 
6312   // If the result of a {s|z}ext and its source are both live out, rewrite all
6313   // other uses of the source with result of extension.
6314   Value *Src = I->getOperand(0);
6315   if (Src->hasOneUse())
6316     return false;
6317 
6318   // Only do this xform if truncating is free.
6319   if (!TLI->isTruncateFree(I->getType(), Src->getType()))
6320     return false;
6321 
6322   // Only safe to perform the optimization if the source is also defined in
6323   // this block.
6324   if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
6325     return false;
6326 
6327   bool DefIsLiveOut = false;
6328   for (User *U : I->users()) {
6329     Instruction *UI = cast<Instruction>(U);
6330 
6331     // Figure out which BB this ext is used in.
6332     BasicBlock *UserBB = UI->getParent();
6333     if (UserBB == DefBB) continue;
6334     DefIsLiveOut = true;
6335     break;
6336   }
6337   if (!DefIsLiveOut)
6338     return false;
6339 
6340   // Make sure none of the uses are PHI nodes.
6341   for (User *U : Src->users()) {
6342     Instruction *UI = cast<Instruction>(U);
6343     BasicBlock *UserBB = UI->getParent();
6344     if (UserBB == DefBB) continue;
6345     // Be conservative. We don't want this xform to end up introducing
6346     // reloads just before load / store instructions.
6347     if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
6348       return false;
6349   }
6350 
6351   // InsertedTruncs - Only insert one trunc in each block once.
6352   DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
6353 
6354   bool MadeChange = false;
6355   for (Use &U : Src->uses()) {
6356     Instruction *User = cast<Instruction>(U.getUser());
6357 
6358     // Figure out which BB this ext is used in.
6359     BasicBlock *UserBB = User->getParent();
6360     if (UserBB == DefBB) continue;
6361 
6362     // Both src and def are live in this block. Rewrite the use.
6363     Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
6364 
6365     if (!InsertedTrunc) {
6366       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
6367       assert(InsertPt != UserBB->end());
6368       InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
6369       InsertedInsts.insert(InsertedTrunc);
6370     }
6371 
6372     // Replace a use of the {s|z}ext source with a use of the result.
6373     U = InsertedTrunc;
6374     ++NumExtUses;
6375     MadeChange = true;
6376   }
6377 
6378   return MadeChange;
6379 }
6380 
6381 // Find loads whose uses only use some of the loaded value's bits.  Add an "and"
6382 // just after the load if the target can fold this into one extload instruction,
6383 // with the hope of eliminating some of the other later "and" instructions using
6384 // the loaded value.  "and"s that are made trivially redundant by the insertion
6385 // of the new "and" are removed by this function, while others (e.g. those whose
6386 // path from the load goes through a phi) are left for isel to potentially
6387 // remove.
6388 //
6389 // For example:
6390 //
6391 // b0:
6392 //   x = load i32
6393 //   ...
6394 // b1:
6395 //   y = and x, 0xff
6396 //   z = use y
6397 //
6398 // becomes:
6399 //
6400 // b0:
6401 //   x = load i32
6402 //   x' = and x, 0xff
6403 //   ...
6404 // b1:
6405 //   z = use x'
6406 //
6407 // whereas:
6408 //
6409 // b0:
6410 //   x1 = load i32
6411 //   ...
6412 // b1:
6413 //   x2 = load i32
6414 //   ...
6415 // b2:
6416 //   x = phi x1, x2
6417 //   y = and x, 0xff
6418 //
6419 // becomes (after a call to optimizeLoadExt for each load):
6420 //
6421 // b0:
6422 //   x1 = load i32
6423 //   x1' = and x1, 0xff
6424 //   ...
6425 // b1:
6426 //   x2 = load i32
6427 //   x2' = and x2, 0xff
6428 //   ...
6429 // b2:
6430 //   x = phi x1', x2'
6431 //   y = and x, 0xff
6432 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
6433   if (!Load->isSimple() || !Load->getType()->isIntOrPtrTy())
6434     return false;
6435 
6436   // Skip loads we've already transformed.
6437   if (Load->hasOneUse() &&
6438       InsertedInsts.count(cast<Instruction>(*Load->user_begin())))
6439     return false;
6440 
6441   // Look at all uses of Load, looking through phis, to determine how many bits
6442   // of the loaded value are needed.
6443   SmallVector<Instruction *, 8> WorkList;
6444   SmallPtrSet<Instruction *, 16> Visited;
6445   SmallVector<Instruction *, 8> AndsToMaybeRemove;
6446   for (auto *U : Load->users())
6447     WorkList.push_back(cast<Instruction>(U));
6448 
6449   EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
6450   unsigned BitWidth = LoadResultVT.getSizeInBits();
6451   // If the BitWidth is 0, do not try to optimize the type
6452   if (BitWidth == 0)
6453     return false;
6454 
6455   APInt DemandBits(BitWidth, 0);
6456   APInt WidestAndBits(BitWidth, 0);
6457 
6458   while (!WorkList.empty()) {
6459     Instruction *I = WorkList.pop_back_val();
6460 
6461     // Break use-def graph loops.
6462     if (!Visited.insert(I).second)
6463       continue;
6464 
6465     // For a PHI node, push all of its users.
6466     if (auto *Phi = dyn_cast<PHINode>(I)) {
6467       for (auto *U : Phi->users())
6468         WorkList.push_back(cast<Instruction>(U));
6469       continue;
6470     }
6471 
6472     switch (I->getOpcode()) {
6473     case Instruction::And: {
6474       auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
6475       if (!AndC)
6476         return false;
6477       APInt AndBits = AndC->getValue();
6478       DemandBits |= AndBits;
6479       // Keep track of the widest and mask we see.
6480       if (AndBits.ugt(WidestAndBits))
6481         WidestAndBits = AndBits;
6482       if (AndBits == WidestAndBits && I->getOperand(0) == Load)
6483         AndsToMaybeRemove.push_back(I);
6484       break;
6485     }
6486 
6487     case Instruction::Shl: {
6488       auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
6489       if (!ShlC)
6490         return false;
6491       uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
6492       DemandBits.setLowBits(BitWidth - ShiftAmt);
6493       break;
6494     }
6495 
6496     case Instruction::Trunc: {
6497       EVT TruncVT = TLI->getValueType(*DL, I->getType());
6498       unsigned TruncBitWidth = TruncVT.getSizeInBits();
6499       DemandBits.setLowBits(TruncBitWidth);
6500       break;
6501     }
6502 
6503     default:
6504       return false;
6505     }
6506   }
6507 
6508   uint32_t ActiveBits = DemandBits.getActiveBits();
6509   // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
6510   // target even if isLoadExtLegal says an i1 EXTLOAD is valid.  For example,
6511   // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
6512   // (and (load x) 1) is not matched as a single instruction, rather as a LDR
6513   // followed by an AND.
6514   // TODO: Look into removing this restriction by fixing backends to either
6515   // return false for isLoadExtLegal for i1 or have them select this pattern to
6516   // a single instruction.
6517   //
6518   // Also avoid hoisting if we didn't see any ands with the exact DemandBits
6519   // mask, since these are the only ands that will be removed by isel.
6520   if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) ||
6521       WidestAndBits != DemandBits)
6522     return false;
6523 
6524   LLVMContext &Ctx = Load->getType()->getContext();
6525   Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
6526   EVT TruncVT = TLI->getValueType(*DL, TruncTy);
6527 
6528   // Reject cases that won't be matched as extloads.
6529   if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
6530       !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
6531     return false;
6532 
6533   IRBuilder<> Builder(Load->getNextNode());
6534   auto *NewAnd = cast<Instruction>(
6535       Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
6536   // Mark this instruction as "inserted by CGP", so that other
6537   // optimizations don't touch it.
6538   InsertedInsts.insert(NewAnd);
6539 
6540   // Replace all uses of load with new and (except for the use of load in the
6541   // new and itself).
6542   Load->replaceAllUsesWith(NewAnd);
6543   NewAnd->setOperand(0, Load);
6544 
6545   // Remove any and instructions that are now redundant.
6546   for (auto *And : AndsToMaybeRemove)
6547     // Check that the and mask is the same as the one we decided to put on the
6548     // new and.
6549     if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
6550       And->replaceAllUsesWith(NewAnd);
6551       if (&*CurInstIterator == And)
6552         CurInstIterator = std::next(And->getIterator());
6553       And->eraseFromParent();
6554       ++NumAndUses;
6555     }
6556 
6557   ++NumAndsAdded;
6558   return true;
6559 }
6560 
6561 /// Check if V (an operand of a select instruction) is an expensive instruction
6562 /// that is only used once.
6563 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
6564   auto *I = dyn_cast<Instruction>(V);
6565   // If it's safe to speculatively execute, then it should not have side
6566   // effects; therefore, it's safe to sink and possibly *not* execute.
6567   return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
6568          TTI->getUserCost(I, TargetTransformInfo::TCK_SizeAndLatency) >=
6569          TargetTransformInfo::TCC_Expensive;
6570 }
6571 
6572 /// Returns true if a SelectInst should be turned into an explicit branch.
6573 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
6574                                                 const TargetLowering *TLI,
6575                                                 SelectInst *SI) {
6576   // If even a predictable select is cheap, then a branch can't be cheaper.
6577   if (!TLI->isPredictableSelectExpensive())
6578     return false;
6579 
6580   // FIXME: This should use the same heuristics as IfConversion to determine
6581   // whether a select is better represented as a branch.
6582 
6583   // If metadata tells us that the select condition is obviously predictable,
6584   // then we want to replace the select with a branch.
6585   uint64_t TrueWeight, FalseWeight;
6586   if (SI->extractProfMetadata(TrueWeight, FalseWeight)) {
6587     uint64_t Max = std::max(TrueWeight, FalseWeight);
6588     uint64_t Sum = TrueWeight + FalseWeight;
6589     if (Sum != 0) {
6590       auto Probability = BranchProbability::getBranchProbability(Max, Sum);
6591       if (Probability > TTI->getPredictableBranchThreshold())
6592         return true;
6593     }
6594   }
6595 
6596   CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
6597 
6598   // If a branch is predictable, an out-of-order CPU can avoid blocking on its
6599   // comparison condition. If the compare has more than one use, there's
6600   // probably another cmov or setcc around, so it's not worth emitting a branch.
6601   if (!Cmp || !Cmp->hasOneUse())
6602     return false;
6603 
6604   // If either operand of the select is expensive and only needed on one side
6605   // of the select, we should form a branch.
6606   if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
6607       sinkSelectOperand(TTI, SI->getFalseValue()))
6608     return true;
6609 
6610   return false;
6611 }
6612 
6613 /// If \p isTrue is true, return the true value of \p SI, otherwise return
6614 /// false value of \p SI. If the true/false value of \p SI is defined by any
6615 /// select instructions in \p Selects, look through the defining select
6616 /// instruction until the true/false value is not defined in \p Selects.
6617 static Value *getTrueOrFalseValue(
6618     SelectInst *SI, bool isTrue,
6619     const SmallPtrSet<const Instruction *, 2> &Selects) {
6620   Value *V = nullptr;
6621 
6622   for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI);
6623        DefSI = dyn_cast<SelectInst>(V)) {
6624     assert(DefSI->getCondition() == SI->getCondition() &&
6625            "The condition of DefSI does not match with SI");
6626     V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
6627   }
6628 
6629   assert(V && "Failed to get select true/false value");
6630   return V;
6631 }
6632 
6633 bool CodeGenPrepare::optimizeShiftInst(BinaryOperator *Shift) {
6634   assert(Shift->isShift() && "Expected a shift");
6635 
6636   // If this is (1) a vector shift, (2) shifts by scalars are cheaper than
6637   // general vector shifts, and (3) the shift amount is a select-of-splatted
6638   // values, hoist the shifts before the select:
6639   //   shift Op0, (select Cond, TVal, FVal) -->
6640   //   select Cond, (shift Op0, TVal), (shift Op0, FVal)
6641   //
6642   // This is inverting a generic IR transform when we know that the cost of a
6643   // general vector shift is more than the cost of 2 shift-by-scalars.
6644   // We can't do this effectively in SDAG because we may not be able to
6645   // determine if the select operands are splats from within a basic block.
6646   Type *Ty = Shift->getType();
6647   if (!Ty->isVectorTy() || !TLI->isVectorShiftByScalarCheap(Ty))
6648     return false;
6649   Value *Cond, *TVal, *FVal;
6650   if (!match(Shift->getOperand(1),
6651              m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))))
6652     return false;
6653   if (!isSplatValue(TVal) || !isSplatValue(FVal))
6654     return false;
6655 
6656   IRBuilder<> Builder(Shift);
6657   BinaryOperator::BinaryOps Opcode = Shift->getOpcode();
6658   Value *NewTVal = Builder.CreateBinOp(Opcode, Shift->getOperand(0), TVal);
6659   Value *NewFVal = Builder.CreateBinOp(Opcode, Shift->getOperand(0), FVal);
6660   Value *NewSel = Builder.CreateSelect(Cond, NewTVal, NewFVal);
6661   Shift->replaceAllUsesWith(NewSel);
6662   Shift->eraseFromParent();
6663   return true;
6664 }
6665 
6666 bool CodeGenPrepare::optimizeFunnelShift(IntrinsicInst *Fsh) {
6667   Intrinsic::ID Opcode = Fsh->getIntrinsicID();
6668   assert((Opcode == Intrinsic::fshl || Opcode == Intrinsic::fshr) &&
6669          "Expected a funnel shift");
6670 
6671   // If this is (1) a vector funnel shift, (2) shifts by scalars are cheaper
6672   // than general vector shifts, and (3) the shift amount is select-of-splatted
6673   // values, hoist the funnel shifts before the select:
6674   //   fsh Op0, Op1, (select Cond, TVal, FVal) -->
6675   //   select Cond, (fsh Op0, Op1, TVal), (fsh Op0, Op1, FVal)
6676   //
6677   // This is inverting a generic IR transform when we know that the cost of a
6678   // general vector shift is more than the cost of 2 shift-by-scalars.
6679   // We can't do this effectively in SDAG because we may not be able to
6680   // determine if the select operands are splats from within a basic block.
6681   Type *Ty = Fsh->getType();
6682   if (!Ty->isVectorTy() || !TLI->isVectorShiftByScalarCheap(Ty))
6683     return false;
6684   Value *Cond, *TVal, *FVal;
6685   if (!match(Fsh->getOperand(2),
6686              m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))))
6687     return false;
6688   if (!isSplatValue(TVal) || !isSplatValue(FVal))
6689     return false;
6690 
6691   IRBuilder<> Builder(Fsh);
6692   Value *X = Fsh->getOperand(0), *Y = Fsh->getOperand(1);
6693   Value *NewTVal = Builder.CreateIntrinsic(Opcode, Ty, { X, Y, TVal });
6694   Value *NewFVal = Builder.CreateIntrinsic(Opcode, Ty, { X, Y, FVal });
6695   Value *NewSel = Builder.CreateSelect(Cond, NewTVal, NewFVal);
6696   Fsh->replaceAllUsesWith(NewSel);
6697   Fsh->eraseFromParent();
6698   return true;
6699 }
6700 
6701 /// If we have a SelectInst that will likely profit from branch prediction,
6702 /// turn it into a branch.
6703 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
6704   if (DisableSelectToBranch)
6705     return false;
6706 
6707   // Find all consecutive select instructions that share the same condition.
6708   SmallVector<SelectInst *, 2> ASI;
6709   ASI.push_back(SI);
6710   for (BasicBlock::iterator It = ++BasicBlock::iterator(SI);
6711        It != SI->getParent()->end(); ++It) {
6712     SelectInst *I = dyn_cast<SelectInst>(&*It);
6713     if (I && SI->getCondition() == I->getCondition()) {
6714       ASI.push_back(I);
6715     } else {
6716       break;
6717     }
6718   }
6719 
6720   SelectInst *LastSI = ASI.back();
6721   // Increment the current iterator to skip all the rest of select instructions
6722   // because they will be either "not lowered" or "all lowered" to branch.
6723   CurInstIterator = std::next(LastSI->getIterator());
6724 
6725   bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
6726 
6727   // Can we convert the 'select' to CF ?
6728   if (VectorCond || SI->getMetadata(LLVMContext::MD_unpredictable))
6729     return false;
6730 
6731   TargetLowering::SelectSupportKind SelectKind;
6732   if (VectorCond)
6733     SelectKind = TargetLowering::VectorMaskSelect;
6734   else if (SI->getType()->isVectorTy())
6735     SelectKind = TargetLowering::ScalarCondVectorVal;
6736   else
6737     SelectKind = TargetLowering::ScalarValSelect;
6738 
6739   if (TLI->isSelectSupported(SelectKind) &&
6740       (!isFormingBranchFromSelectProfitable(TTI, TLI, SI) || OptSize ||
6741        llvm::shouldOptimizeForSize(SI->getParent(), PSI, BFI.get())))
6742     return false;
6743 
6744   // The DominatorTree needs to be rebuilt by any consumers after this
6745   // transformation. We simply reset here rather than setting the ModifiedDT
6746   // flag to avoid restarting the function walk in runOnFunction for each
6747   // select optimized.
6748   DT.reset();
6749 
6750   // Transform a sequence like this:
6751   //    start:
6752   //       %cmp = cmp uge i32 %a, %b
6753   //       %sel = select i1 %cmp, i32 %c, i32 %d
6754   //
6755   // Into:
6756   //    start:
6757   //       %cmp = cmp uge i32 %a, %b
6758   //       %cmp.frozen = freeze %cmp
6759   //       br i1 %cmp.frozen, label %select.true, label %select.false
6760   //    select.true:
6761   //       br label %select.end
6762   //    select.false:
6763   //       br label %select.end
6764   //    select.end:
6765   //       %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
6766   //
6767   // %cmp should be frozen, otherwise it may introduce undefined behavior.
6768   // In addition, we may sink instructions that produce %c or %d from
6769   // the entry block into the destination(s) of the new branch.
6770   // If the true or false blocks do not contain a sunken instruction, that
6771   // block and its branch may be optimized away. In that case, one side of the
6772   // first branch will point directly to select.end, and the corresponding PHI
6773   // predecessor block will be the start block.
6774 
6775   // First, we split the block containing the select into 2 blocks.
6776   BasicBlock *StartBlock = SI->getParent();
6777   BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
6778   BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
6779   BFI->setBlockFreq(EndBlock, BFI->getBlockFreq(StartBlock).getFrequency());
6780 
6781   // Delete the unconditional branch that was just created by the split.
6782   StartBlock->getTerminator()->eraseFromParent();
6783 
6784   // These are the new basic blocks for the conditional branch.
6785   // At least one will become an actual new basic block.
6786   BasicBlock *TrueBlock = nullptr;
6787   BasicBlock *FalseBlock = nullptr;
6788   BranchInst *TrueBranch = nullptr;
6789   BranchInst *FalseBranch = nullptr;
6790 
6791   // Sink expensive instructions into the conditional blocks to avoid executing
6792   // them speculatively.
6793   for (SelectInst *SI : ASI) {
6794     if (sinkSelectOperand(TTI, SI->getTrueValue())) {
6795       if (TrueBlock == nullptr) {
6796         TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
6797                                        EndBlock->getParent(), EndBlock);
6798         TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
6799         TrueBranch->setDebugLoc(SI->getDebugLoc());
6800       }
6801       auto *TrueInst = cast<Instruction>(SI->getTrueValue());
6802       TrueInst->moveBefore(TrueBranch);
6803     }
6804     if (sinkSelectOperand(TTI, SI->getFalseValue())) {
6805       if (FalseBlock == nullptr) {
6806         FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
6807                                         EndBlock->getParent(), EndBlock);
6808         FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
6809         FalseBranch->setDebugLoc(SI->getDebugLoc());
6810       }
6811       auto *FalseInst = cast<Instruction>(SI->getFalseValue());
6812       FalseInst->moveBefore(FalseBranch);
6813     }
6814   }
6815 
6816   // If there was nothing to sink, then arbitrarily choose the 'false' side
6817   // for a new input value to the PHI.
6818   if (TrueBlock == FalseBlock) {
6819     assert(TrueBlock == nullptr &&
6820            "Unexpected basic block transform while optimizing select");
6821 
6822     FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
6823                                     EndBlock->getParent(), EndBlock);
6824     auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
6825     FalseBranch->setDebugLoc(SI->getDebugLoc());
6826   }
6827 
6828   // Insert the real conditional branch based on the original condition.
6829   // If we did not create a new block for one of the 'true' or 'false' paths
6830   // of the condition, it means that side of the branch goes to the end block
6831   // directly and the path originates from the start block from the point of
6832   // view of the new PHI.
6833   BasicBlock *TT, *FT;
6834   if (TrueBlock == nullptr) {
6835     TT = EndBlock;
6836     FT = FalseBlock;
6837     TrueBlock = StartBlock;
6838   } else if (FalseBlock == nullptr) {
6839     TT = TrueBlock;
6840     FT = EndBlock;
6841     FalseBlock = StartBlock;
6842   } else {
6843     TT = TrueBlock;
6844     FT = FalseBlock;
6845   }
6846   IRBuilder<> IB(SI);
6847   auto *CondFr = IB.CreateFreeze(SI->getCondition(), SI->getName() + ".frozen");
6848   IB.CreateCondBr(CondFr, TT, FT, SI);
6849 
6850   SmallPtrSet<const Instruction *, 2> INS;
6851   INS.insert(ASI.begin(), ASI.end());
6852   // Use reverse iterator because later select may use the value of the
6853   // earlier select, and we need to propagate value through earlier select
6854   // to get the PHI operand.
6855   for (SelectInst *SI : llvm::reverse(ASI)) {
6856     // The select itself is replaced with a PHI Node.
6857     PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
6858     PN->takeName(SI);
6859     PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock);
6860     PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock);
6861     PN->setDebugLoc(SI->getDebugLoc());
6862 
6863     SI->replaceAllUsesWith(PN);
6864     SI->eraseFromParent();
6865     INS.erase(SI);
6866     ++NumSelectsExpanded;
6867   }
6868 
6869   // Instruct OptimizeBlock to skip to the next block.
6870   CurInstIterator = StartBlock->end();
6871   return true;
6872 }
6873 
6874 /// Some targets only accept certain types for splat inputs. For example a VDUP
6875 /// in MVE takes a GPR (integer) register, and the instruction that incorporate
6876 /// a VDUP (such as a VADD qd, qm, rm) also require a gpr register.
6877 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
6878   // Accept shuf(insertelem(undef/poison, val, 0), undef/poison, <0,0,..>) only
6879   if (!match(SVI, m_Shuffle(m_InsertElt(m_Undef(), m_Value(), m_ZeroInt()),
6880                             m_Undef(), m_ZeroMask())))
6881     return false;
6882   Type *NewType = TLI->shouldConvertSplatType(SVI);
6883   if (!NewType)
6884     return false;
6885 
6886   auto *SVIVecType = cast<FixedVectorType>(SVI->getType());
6887   assert(!NewType->isVectorTy() && "Expected a scalar type!");
6888   assert(NewType->getScalarSizeInBits() == SVIVecType->getScalarSizeInBits() &&
6889          "Expected a type of the same size!");
6890   auto *NewVecType =
6891       FixedVectorType::get(NewType, SVIVecType->getNumElements());
6892 
6893   // Create a bitcast (shuffle (insert (bitcast(..))))
6894   IRBuilder<> Builder(SVI->getContext());
6895   Builder.SetInsertPoint(SVI);
6896   Value *BC1 = Builder.CreateBitCast(
6897       cast<Instruction>(SVI->getOperand(0))->getOperand(1), NewType);
6898   Value *Shuffle = Builder.CreateVectorSplat(NewVecType->getNumElements(), BC1);
6899   Value *BC2 = Builder.CreateBitCast(Shuffle, SVIVecType);
6900 
6901   SVI->replaceAllUsesWith(BC2);
6902   RecursivelyDeleteTriviallyDeadInstructions(
6903       SVI, TLInfo, nullptr, [&](Value *V) { removeAllAssertingVHReferences(V); });
6904 
6905   // Also hoist the bitcast up to its operand if it they are not in the same
6906   // block.
6907   if (auto *BCI = dyn_cast<Instruction>(BC1))
6908     if (auto *Op = dyn_cast<Instruction>(BCI->getOperand(0)))
6909       if (BCI->getParent() != Op->getParent() && !isa<PHINode>(Op) &&
6910           !Op->isTerminator() && !Op->isEHPad())
6911         BCI->moveAfter(Op);
6912 
6913   return true;
6914 }
6915 
6916 bool CodeGenPrepare::tryToSinkFreeOperands(Instruction *I) {
6917   // If the operands of I can be folded into a target instruction together with
6918   // I, duplicate and sink them.
6919   SmallVector<Use *, 4> OpsToSink;
6920   if (!TLI->shouldSinkOperands(I, OpsToSink))
6921     return false;
6922 
6923   // OpsToSink can contain multiple uses in a use chain (e.g.
6924   // (%u1 with %u1 = shufflevector), (%u2 with %u2 = zext %u1)). The dominating
6925   // uses must come first, so we process the ops in reverse order so as to not
6926   // create invalid IR.
6927   BasicBlock *TargetBB = I->getParent();
6928   bool Changed = false;
6929   SmallVector<Use *, 4> ToReplace;
6930   Instruction *InsertPoint = I;
6931   DenseMap<const Instruction *, unsigned long> InstOrdering;
6932   unsigned long InstNumber = 0;
6933   for (const auto &I : *TargetBB)
6934     InstOrdering[&I] = InstNumber++;
6935 
6936   for (Use *U : reverse(OpsToSink)) {
6937     auto *UI = cast<Instruction>(U->get());
6938     if (isa<PHINode>(UI))
6939       continue;
6940     if (UI->getParent() == TargetBB) {
6941       if (InstOrdering[UI] < InstOrdering[InsertPoint])
6942         InsertPoint = UI;
6943       continue;
6944     }
6945     ToReplace.push_back(U);
6946   }
6947 
6948   SetVector<Instruction *> MaybeDead;
6949   DenseMap<Instruction *, Instruction *> NewInstructions;
6950   for (Use *U : ToReplace) {
6951     auto *UI = cast<Instruction>(U->get());
6952     Instruction *NI = UI->clone();
6953     NewInstructions[UI] = NI;
6954     MaybeDead.insert(UI);
6955     LLVM_DEBUG(dbgs() << "Sinking " << *UI << " to user " << *I << "\n");
6956     NI->insertBefore(InsertPoint);
6957     InsertPoint = NI;
6958     InsertedInsts.insert(NI);
6959 
6960     // Update the use for the new instruction, making sure that we update the
6961     // sunk instruction uses, if it is part of a chain that has already been
6962     // sunk.
6963     Instruction *OldI = cast<Instruction>(U->getUser());
6964     if (NewInstructions.count(OldI))
6965       NewInstructions[OldI]->setOperand(U->getOperandNo(), NI);
6966     else
6967       U->set(NI);
6968     Changed = true;
6969   }
6970 
6971   // Remove instructions that are dead after sinking.
6972   for (auto *I : MaybeDead) {
6973     if (!I->hasNUsesOrMore(1)) {
6974       LLVM_DEBUG(dbgs() << "Removing dead instruction: " << *I << "\n");
6975       I->eraseFromParent();
6976     }
6977   }
6978 
6979   return Changed;
6980 }
6981 
6982 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
6983   Value *Cond = SI->getCondition();
6984   Type *OldType = Cond->getType();
6985   LLVMContext &Context = Cond->getContext();
6986   EVT OldVT = TLI->getValueType(*DL, OldType);
6987   MVT RegType = TLI->getRegisterType(Context, OldVT);
6988   unsigned RegWidth = RegType.getSizeInBits();
6989 
6990   if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
6991     return false;
6992 
6993   // If the register width is greater than the type width, expand the condition
6994   // of the switch instruction and each case constant to the width of the
6995   // register. By widening the type of the switch condition, subsequent
6996   // comparisons (for case comparisons) will not need to be extended to the
6997   // preferred register width, so we will potentially eliminate N-1 extends,
6998   // where N is the number of cases in the switch.
6999   auto *NewType = Type::getIntNTy(Context, RegWidth);
7000 
7001   // Extend the switch condition and case constants using the target preferred
7002   // extend unless the switch condition is a function argument with an extend
7003   // attribute. In that case, we can avoid an unnecessary mask/extension by
7004   // matching the argument extension instead.
7005   Instruction::CastOps ExtType = Instruction::ZExt;
7006   // Some targets prefer SExt over ZExt.
7007   if (TLI->isSExtCheaperThanZExt(OldVT, RegType))
7008     ExtType = Instruction::SExt;
7009 
7010   if (auto *Arg = dyn_cast<Argument>(Cond)) {
7011     if (Arg->hasSExtAttr())
7012       ExtType = Instruction::SExt;
7013     if (Arg->hasZExtAttr())
7014       ExtType = Instruction::ZExt;
7015   }
7016 
7017   auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
7018   ExtInst->insertBefore(SI);
7019   ExtInst->setDebugLoc(SI->getDebugLoc());
7020   SI->setCondition(ExtInst);
7021   for (auto Case : SI->cases()) {
7022     APInt NarrowConst = Case.getCaseValue()->getValue();
7023     APInt WideConst = (ExtType == Instruction::ZExt) ?
7024                       NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
7025     Case.setValue(ConstantInt::get(Context, WideConst));
7026   }
7027 
7028   return true;
7029 }
7030 
7031 
7032 namespace {
7033 
7034 /// Helper class to promote a scalar operation to a vector one.
7035 /// This class is used to move downward extractelement transition.
7036 /// E.g.,
7037 /// a = vector_op <2 x i32>
7038 /// b = extractelement <2 x i32> a, i32 0
7039 /// c = scalar_op b
7040 /// store c
7041 ///
7042 /// =>
7043 /// a = vector_op <2 x i32>
7044 /// c = vector_op a (equivalent to scalar_op on the related lane)
7045 /// * d = extractelement <2 x i32> c, i32 0
7046 /// * store d
7047 /// Assuming both extractelement and store can be combine, we get rid of the
7048 /// transition.
7049 class VectorPromoteHelper {
7050   /// DataLayout associated with the current module.
7051   const DataLayout &DL;
7052 
7053   /// Used to perform some checks on the legality of vector operations.
7054   const TargetLowering &TLI;
7055 
7056   /// Used to estimated the cost of the promoted chain.
7057   const TargetTransformInfo &TTI;
7058 
7059   /// The transition being moved downwards.
7060   Instruction *Transition;
7061 
7062   /// The sequence of instructions to be promoted.
7063   SmallVector<Instruction *, 4> InstsToBePromoted;
7064 
7065   /// Cost of combining a store and an extract.
7066   unsigned StoreExtractCombineCost;
7067 
7068   /// Instruction that will be combined with the transition.
7069   Instruction *CombineInst = nullptr;
7070 
7071   /// The instruction that represents the current end of the transition.
7072   /// Since we are faking the promotion until we reach the end of the chain
7073   /// of computation, we need a way to get the current end of the transition.
7074   Instruction *getEndOfTransition() const {
7075     if (InstsToBePromoted.empty())
7076       return Transition;
7077     return InstsToBePromoted.back();
7078   }
7079 
7080   /// Return the index of the original value in the transition.
7081   /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
7082   /// c, is at index 0.
7083   unsigned getTransitionOriginalValueIdx() const {
7084     assert(isa<ExtractElementInst>(Transition) &&
7085            "Other kind of transitions are not supported yet");
7086     return 0;
7087   }
7088 
7089   /// Return the index of the index in the transition.
7090   /// E.g., for "extractelement <2 x i32> c, i32 0" the index
7091   /// is at index 1.
7092   unsigned getTransitionIdx() const {
7093     assert(isa<ExtractElementInst>(Transition) &&
7094            "Other kind of transitions are not supported yet");
7095     return 1;
7096   }
7097 
7098   /// Get the type of the transition.
7099   /// This is the type of the original value.
7100   /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
7101   /// transition is <2 x i32>.
7102   Type *getTransitionType() const {
7103     return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
7104   }
7105 
7106   /// Promote \p ToBePromoted by moving \p Def downward through.
7107   /// I.e., we have the following sequence:
7108   /// Def = Transition <ty1> a to <ty2>
7109   /// b = ToBePromoted <ty2> Def, ...
7110   /// =>
7111   /// b = ToBePromoted <ty1> a, ...
7112   /// Def = Transition <ty1> ToBePromoted to <ty2>
7113   void promoteImpl(Instruction *ToBePromoted);
7114 
7115   /// Check whether or not it is profitable to promote all the
7116   /// instructions enqueued to be promoted.
7117   bool isProfitableToPromote() {
7118     Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
7119     unsigned Index = isa<ConstantInt>(ValIdx)
7120                          ? cast<ConstantInt>(ValIdx)->getZExtValue()
7121                          : -1;
7122     Type *PromotedType = getTransitionType();
7123 
7124     StoreInst *ST = cast<StoreInst>(CombineInst);
7125     unsigned AS = ST->getPointerAddressSpace();
7126     // Check if this store is supported.
7127     if (!TLI.allowsMisalignedMemoryAccesses(
7128             TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
7129             ST->getAlign())) {
7130       // If this is not supported, there is no way we can combine
7131       // the extract with the store.
7132       return false;
7133     }
7134 
7135     // The scalar chain of computation has to pay for the transition
7136     // scalar to vector.
7137     // The vector chain has to account for the combining cost.
7138     InstructionCost ScalarCost =
7139         TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
7140     InstructionCost VectorCost = StoreExtractCombineCost;
7141     enum TargetTransformInfo::TargetCostKind CostKind =
7142       TargetTransformInfo::TCK_RecipThroughput;
7143     for (const auto &Inst : InstsToBePromoted) {
7144       // Compute the cost.
7145       // By construction, all instructions being promoted are arithmetic ones.
7146       // Moreover, one argument is a constant that can be viewed as a splat
7147       // constant.
7148       Value *Arg0 = Inst->getOperand(0);
7149       bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
7150                             isa<ConstantFP>(Arg0);
7151       TargetTransformInfo::OperandValueKind Arg0OVK =
7152           IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
7153                          : TargetTransformInfo::OK_AnyValue;
7154       TargetTransformInfo::OperandValueKind Arg1OVK =
7155           !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
7156                           : TargetTransformInfo::OK_AnyValue;
7157       ScalarCost += TTI.getArithmeticInstrCost(
7158           Inst->getOpcode(), Inst->getType(), CostKind, Arg0OVK, Arg1OVK);
7159       VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
7160                                                CostKind,
7161                                                Arg0OVK, Arg1OVK);
7162     }
7163     LLVM_DEBUG(
7164         dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
7165                << ScalarCost << "\nVector: " << VectorCost << '\n');
7166     return ScalarCost > VectorCost;
7167   }
7168 
7169   /// Generate a constant vector with \p Val with the same
7170   /// number of elements as the transition.
7171   /// \p UseSplat defines whether or not \p Val should be replicated
7172   /// across the whole vector.
7173   /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
7174   /// otherwise we generate a vector with as many undef as possible:
7175   /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
7176   /// used at the index of the extract.
7177   Value *getConstantVector(Constant *Val, bool UseSplat) const {
7178     unsigned ExtractIdx = std::numeric_limits<unsigned>::max();
7179     if (!UseSplat) {
7180       // If we cannot determine where the constant must be, we have to
7181       // use a splat constant.
7182       Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
7183       if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
7184         ExtractIdx = CstVal->getSExtValue();
7185       else
7186         UseSplat = true;
7187     }
7188 
7189     ElementCount EC = cast<VectorType>(getTransitionType())->getElementCount();
7190     if (UseSplat)
7191       return ConstantVector::getSplat(EC, Val);
7192 
7193     if (!EC.isScalable()) {
7194       SmallVector<Constant *, 4> ConstVec;
7195       UndefValue *UndefVal = UndefValue::get(Val->getType());
7196       for (unsigned Idx = 0; Idx != EC.getKnownMinValue(); ++Idx) {
7197         if (Idx == ExtractIdx)
7198           ConstVec.push_back(Val);
7199         else
7200           ConstVec.push_back(UndefVal);
7201       }
7202       return ConstantVector::get(ConstVec);
7203     } else
7204       llvm_unreachable(
7205           "Generate scalable vector for non-splat is unimplemented");
7206   }
7207 
7208   /// Check if promoting to a vector type an operand at \p OperandIdx
7209   /// in \p Use can trigger undefined behavior.
7210   static bool canCauseUndefinedBehavior(const Instruction *Use,
7211                                         unsigned OperandIdx) {
7212     // This is not safe to introduce undef when the operand is on
7213     // the right hand side of a division-like instruction.
7214     if (OperandIdx != 1)
7215       return false;
7216     switch (Use->getOpcode()) {
7217     default:
7218       return false;
7219     case Instruction::SDiv:
7220     case Instruction::UDiv:
7221     case Instruction::SRem:
7222     case Instruction::URem:
7223       return true;
7224     case Instruction::FDiv:
7225     case Instruction::FRem:
7226       return !Use->hasNoNaNs();
7227     }
7228     llvm_unreachable(nullptr);
7229   }
7230 
7231 public:
7232   VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
7233                       const TargetTransformInfo &TTI, Instruction *Transition,
7234                       unsigned CombineCost)
7235       : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
7236         StoreExtractCombineCost(CombineCost) {
7237     assert(Transition && "Do not know how to promote null");
7238   }
7239 
7240   /// Check if we can promote \p ToBePromoted to \p Type.
7241   bool canPromote(const Instruction *ToBePromoted) const {
7242     // We could support CastInst too.
7243     return isa<BinaryOperator>(ToBePromoted);
7244   }
7245 
7246   /// Check if it is profitable to promote \p ToBePromoted
7247   /// by moving downward the transition through.
7248   bool shouldPromote(const Instruction *ToBePromoted) const {
7249     // Promote only if all the operands can be statically expanded.
7250     // Indeed, we do not want to introduce any new kind of transitions.
7251     for (const Use &U : ToBePromoted->operands()) {
7252       const Value *Val = U.get();
7253       if (Val == getEndOfTransition()) {
7254         // If the use is a division and the transition is on the rhs,
7255         // we cannot promote the operation, otherwise we may create a
7256         // division by zero.
7257         if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
7258           return false;
7259         continue;
7260       }
7261       if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
7262           !isa<ConstantFP>(Val))
7263         return false;
7264     }
7265     // Check that the resulting operation is legal.
7266     int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
7267     if (!ISDOpcode)
7268       return false;
7269     return StressStoreExtract ||
7270            TLI.isOperationLegalOrCustom(
7271                ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
7272   }
7273 
7274   /// Check whether or not \p Use can be combined
7275   /// with the transition.
7276   /// I.e., is it possible to do Use(Transition) => AnotherUse?
7277   bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
7278 
7279   /// Record \p ToBePromoted as part of the chain to be promoted.
7280   void enqueueForPromotion(Instruction *ToBePromoted) {
7281     InstsToBePromoted.push_back(ToBePromoted);
7282   }
7283 
7284   /// Set the instruction that will be combined with the transition.
7285   void recordCombineInstruction(Instruction *ToBeCombined) {
7286     assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
7287     CombineInst = ToBeCombined;
7288   }
7289 
7290   /// Promote all the instructions enqueued for promotion if it is
7291   /// is profitable.
7292   /// \return True if the promotion happened, false otherwise.
7293   bool promote() {
7294     // Check if there is something to promote.
7295     // Right now, if we do not have anything to combine with,
7296     // we assume the promotion is not profitable.
7297     if (InstsToBePromoted.empty() || !CombineInst)
7298       return false;
7299 
7300     // Check cost.
7301     if (!StressStoreExtract && !isProfitableToPromote())
7302       return false;
7303 
7304     // Promote.
7305     for (auto &ToBePromoted : InstsToBePromoted)
7306       promoteImpl(ToBePromoted);
7307     InstsToBePromoted.clear();
7308     return true;
7309   }
7310 };
7311 
7312 } // end anonymous namespace
7313 
7314 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
7315   // At this point, we know that all the operands of ToBePromoted but Def
7316   // can be statically promoted.
7317   // For Def, we need to use its parameter in ToBePromoted:
7318   // b = ToBePromoted ty1 a
7319   // Def = Transition ty1 b to ty2
7320   // Move the transition down.
7321   // 1. Replace all uses of the promoted operation by the transition.
7322   // = ... b => = ... Def.
7323   assert(ToBePromoted->getType() == Transition->getType() &&
7324          "The type of the result of the transition does not match "
7325          "the final type");
7326   ToBePromoted->replaceAllUsesWith(Transition);
7327   // 2. Update the type of the uses.
7328   // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
7329   Type *TransitionTy = getTransitionType();
7330   ToBePromoted->mutateType(TransitionTy);
7331   // 3. Update all the operands of the promoted operation with promoted
7332   // operands.
7333   // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
7334   for (Use &U : ToBePromoted->operands()) {
7335     Value *Val = U.get();
7336     Value *NewVal = nullptr;
7337     if (Val == Transition)
7338       NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
7339     else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
7340              isa<ConstantFP>(Val)) {
7341       // Use a splat constant if it is not safe to use undef.
7342       NewVal = getConstantVector(
7343           cast<Constant>(Val),
7344           isa<UndefValue>(Val) ||
7345               canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
7346     } else
7347       llvm_unreachable("Did you modified shouldPromote and forgot to update "
7348                        "this?");
7349     ToBePromoted->setOperand(U.getOperandNo(), NewVal);
7350   }
7351   Transition->moveAfter(ToBePromoted);
7352   Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
7353 }
7354 
7355 /// Some targets can do store(extractelement) with one instruction.
7356 /// Try to push the extractelement towards the stores when the target
7357 /// has this feature and this is profitable.
7358 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
7359   unsigned CombineCost = std::numeric_limits<unsigned>::max();
7360   if (DisableStoreExtract ||
7361       (!StressStoreExtract &&
7362        !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
7363                                        Inst->getOperand(1), CombineCost)))
7364     return false;
7365 
7366   // At this point we know that Inst is a vector to scalar transition.
7367   // Try to move it down the def-use chain, until:
7368   // - We can combine the transition with its single use
7369   //   => we got rid of the transition.
7370   // - We escape the current basic block
7371   //   => we would need to check that we are moving it at a cheaper place and
7372   //      we do not do that for now.
7373   BasicBlock *Parent = Inst->getParent();
7374   LLVM_DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
7375   VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
7376   // If the transition has more than one use, assume this is not going to be
7377   // beneficial.
7378   while (Inst->hasOneUse()) {
7379     Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
7380     LLVM_DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
7381 
7382     if (ToBePromoted->getParent() != Parent) {
7383       LLVM_DEBUG(dbgs() << "Instruction to promote is in a different block ("
7384                         << ToBePromoted->getParent()->getName()
7385                         << ") than the transition (" << Parent->getName()
7386                         << ").\n");
7387       return false;
7388     }
7389 
7390     if (VPH.canCombine(ToBePromoted)) {
7391       LLVM_DEBUG(dbgs() << "Assume " << *Inst << '\n'
7392                         << "will be combined with: " << *ToBePromoted << '\n');
7393       VPH.recordCombineInstruction(ToBePromoted);
7394       bool Changed = VPH.promote();
7395       NumStoreExtractExposed += Changed;
7396       return Changed;
7397     }
7398 
7399     LLVM_DEBUG(dbgs() << "Try promoting.\n");
7400     if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
7401       return false;
7402 
7403     LLVM_DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
7404 
7405     VPH.enqueueForPromotion(ToBePromoted);
7406     Inst = ToBePromoted;
7407   }
7408   return false;
7409 }
7410 
7411 /// For the instruction sequence of store below, F and I values
7412 /// are bundled together as an i64 value before being stored into memory.
7413 /// Sometimes it is more efficient to generate separate stores for F and I,
7414 /// which can remove the bitwise instructions or sink them to colder places.
7415 ///
7416 ///   (store (or (zext (bitcast F to i32) to i64),
7417 ///              (shl (zext I to i64), 32)), addr)  -->
7418 ///   (store F, addr) and (store I, addr+4)
7419 ///
7420 /// Similarly, splitting for other merged store can also be beneficial, like:
7421 /// For pair of {i32, i32}, i64 store --> two i32 stores.
7422 /// For pair of {i32, i16}, i64 store --> two i32 stores.
7423 /// For pair of {i16, i16}, i32 store --> two i16 stores.
7424 /// For pair of {i16, i8},  i32 store --> two i16 stores.
7425 /// For pair of {i8, i8},   i16 store --> two i8 stores.
7426 ///
7427 /// We allow each target to determine specifically which kind of splitting is
7428 /// supported.
7429 ///
7430 /// The store patterns are commonly seen from the simple code snippet below
7431 /// if only std::make_pair(...) is sroa transformed before inlined into hoo.
7432 ///   void goo(const std::pair<int, float> &);
7433 ///   hoo() {
7434 ///     ...
7435 ///     goo(std::make_pair(tmp, ftmp));
7436 ///     ...
7437 ///   }
7438 ///
7439 /// Although we already have similar splitting in DAG Combine, we duplicate
7440 /// it in CodeGenPrepare to catch the case in which pattern is across
7441 /// multiple BBs. The logic in DAG Combine is kept to catch case generated
7442 /// during code expansion.
7443 static bool splitMergedValStore(StoreInst &SI, const DataLayout &DL,
7444                                 const TargetLowering &TLI) {
7445   // Handle simple but common cases only.
7446   Type *StoreType = SI.getValueOperand()->getType();
7447 
7448   // The code below assumes shifting a value by <number of bits>,
7449   // whereas scalable vectors would have to be shifted by
7450   // <2log(vscale) + number of bits> in order to store the
7451   // low/high parts. Bailing out for now.
7452   if (isa<ScalableVectorType>(StoreType))
7453     return false;
7454 
7455   if (!DL.typeSizeEqualsStoreSize(StoreType) ||
7456       DL.getTypeSizeInBits(StoreType) == 0)
7457     return false;
7458 
7459   unsigned HalfValBitSize = DL.getTypeSizeInBits(StoreType) / 2;
7460   Type *SplitStoreType = Type::getIntNTy(SI.getContext(), HalfValBitSize);
7461   if (!DL.typeSizeEqualsStoreSize(SplitStoreType))
7462     return false;
7463 
7464   // Don't split the store if it is volatile.
7465   if (SI.isVolatile())
7466     return false;
7467 
7468   // Match the following patterns:
7469   // (store (or (zext LValue to i64),
7470   //            (shl (zext HValue to i64), 32)), HalfValBitSize)
7471   //  or
7472   // (store (or (shl (zext HValue to i64), 32)), HalfValBitSize)
7473   //            (zext LValue to i64),
7474   // Expect both operands of OR and the first operand of SHL have only
7475   // one use.
7476   Value *LValue, *HValue;
7477   if (!match(SI.getValueOperand(),
7478              m_c_Or(m_OneUse(m_ZExt(m_Value(LValue))),
7479                     m_OneUse(m_Shl(m_OneUse(m_ZExt(m_Value(HValue))),
7480                                    m_SpecificInt(HalfValBitSize))))))
7481     return false;
7482 
7483   // Check LValue and HValue are int with size less or equal than 32.
7484   if (!LValue->getType()->isIntegerTy() ||
7485       DL.getTypeSizeInBits(LValue->getType()) > HalfValBitSize ||
7486       !HValue->getType()->isIntegerTy() ||
7487       DL.getTypeSizeInBits(HValue->getType()) > HalfValBitSize)
7488     return false;
7489 
7490   // If LValue/HValue is a bitcast instruction, use the EVT before bitcast
7491   // as the input of target query.
7492   auto *LBC = dyn_cast<BitCastInst>(LValue);
7493   auto *HBC = dyn_cast<BitCastInst>(HValue);
7494   EVT LowTy = LBC ? EVT::getEVT(LBC->getOperand(0)->getType())
7495                   : EVT::getEVT(LValue->getType());
7496   EVT HighTy = HBC ? EVT::getEVT(HBC->getOperand(0)->getType())
7497                    : EVT::getEVT(HValue->getType());
7498   if (!ForceSplitStore && !TLI.isMultiStoresCheaperThanBitsMerge(LowTy, HighTy))
7499     return false;
7500 
7501   // Start to split store.
7502   IRBuilder<> Builder(SI.getContext());
7503   Builder.SetInsertPoint(&SI);
7504 
7505   // If LValue/HValue is a bitcast in another BB, create a new one in current
7506   // BB so it may be merged with the splitted stores by dag combiner.
7507   if (LBC && LBC->getParent() != SI.getParent())
7508     LValue = Builder.CreateBitCast(LBC->getOperand(0), LBC->getType());
7509   if (HBC && HBC->getParent() != SI.getParent())
7510     HValue = Builder.CreateBitCast(HBC->getOperand(0), HBC->getType());
7511 
7512   bool IsLE = SI.getModule()->getDataLayout().isLittleEndian();
7513   auto CreateSplitStore = [&](Value *V, bool Upper) {
7514     V = Builder.CreateZExtOrBitCast(V, SplitStoreType);
7515     Value *Addr = Builder.CreateBitCast(
7516         SI.getOperand(1),
7517         SplitStoreType->getPointerTo(SI.getPointerAddressSpace()));
7518     Align Alignment = SI.getAlign();
7519     const bool IsOffsetStore = (IsLE && Upper) || (!IsLE && !Upper);
7520     if (IsOffsetStore) {
7521       Addr = Builder.CreateGEP(
7522           SplitStoreType, Addr,
7523           ConstantInt::get(Type::getInt32Ty(SI.getContext()), 1));
7524 
7525       // When splitting the store in half, naturally one half will retain the
7526       // alignment of the original wider store, regardless of whether it was
7527       // over-aligned or not, while the other will require adjustment.
7528       Alignment = commonAlignment(Alignment, HalfValBitSize / 8);
7529     }
7530     Builder.CreateAlignedStore(V, Addr, Alignment);
7531   };
7532 
7533   CreateSplitStore(LValue, false);
7534   CreateSplitStore(HValue, true);
7535 
7536   // Delete the old store.
7537   SI.eraseFromParent();
7538   return true;
7539 }
7540 
7541 // Return true if the GEP has two operands, the first operand is of a sequential
7542 // type, and the second operand is a constant.
7543 static bool GEPSequentialConstIndexed(GetElementPtrInst *GEP) {
7544   gep_type_iterator I = gep_type_begin(*GEP);
7545   return GEP->getNumOperands() == 2 &&
7546       I.isSequential() &&
7547       isa<ConstantInt>(GEP->getOperand(1));
7548 }
7549 
7550 // Try unmerging GEPs to reduce liveness interference (register pressure) across
7551 // IndirectBr edges. Since IndirectBr edges tend to touch on many blocks,
7552 // reducing liveness interference across those edges benefits global register
7553 // allocation. Currently handles only certain cases.
7554 //
7555 // For example, unmerge %GEPI and %UGEPI as below.
7556 //
7557 // ---------- BEFORE ----------
7558 // SrcBlock:
7559 //   ...
7560 //   %GEPIOp = ...
7561 //   ...
7562 //   %GEPI = gep %GEPIOp, Idx
7563 //   ...
7564 //   indirectbr ... [ label %DstB0, label %DstB1, ... label %DstBi ... ]
7565 //   (* %GEPI is alive on the indirectbr edges due to other uses ahead)
7566 //   (* %GEPIOp is alive on the indirectbr edges only because of it's used by
7567 //   %UGEPI)
7568 //
7569 // DstB0: ... (there may be a gep similar to %UGEPI to be unmerged)
7570 // DstB1: ... (there may be a gep similar to %UGEPI to be unmerged)
7571 // ...
7572 //
7573 // DstBi:
7574 //   ...
7575 //   %UGEPI = gep %GEPIOp, UIdx
7576 // ...
7577 // ---------------------------
7578 //
7579 // ---------- AFTER ----------
7580 // SrcBlock:
7581 //   ... (same as above)
7582 //    (* %GEPI is still alive on the indirectbr edges)
7583 //    (* %GEPIOp is no longer alive on the indirectbr edges as a result of the
7584 //    unmerging)
7585 // ...
7586 //
7587 // DstBi:
7588 //   ...
7589 //   %UGEPI = gep %GEPI, (UIdx-Idx)
7590 //   ...
7591 // ---------------------------
7592 //
7593 // The register pressure on the IndirectBr edges is reduced because %GEPIOp is
7594 // no longer alive on them.
7595 //
7596 // We try to unmerge GEPs here in CodGenPrepare, as opposed to limiting merging
7597 // of GEPs in the first place in InstCombiner::visitGetElementPtrInst() so as
7598 // not to disable further simplications and optimizations as a result of GEP
7599 // merging.
7600 //
7601 // Note this unmerging may increase the length of the data flow critical path
7602 // (the path from %GEPIOp to %UGEPI would go through %GEPI), which is a tradeoff
7603 // between the register pressure and the length of data-flow critical
7604 // path. Restricting this to the uncommon IndirectBr case would minimize the
7605 // impact of potentially longer critical path, if any, and the impact on compile
7606 // time.
7607 static bool tryUnmergingGEPsAcrossIndirectBr(GetElementPtrInst *GEPI,
7608                                              const TargetTransformInfo *TTI) {
7609   BasicBlock *SrcBlock = GEPI->getParent();
7610   // Check that SrcBlock ends with an IndirectBr. If not, give up. The common
7611   // (non-IndirectBr) cases exit early here.
7612   if (!isa<IndirectBrInst>(SrcBlock->getTerminator()))
7613     return false;
7614   // Check that GEPI is a simple gep with a single constant index.
7615   if (!GEPSequentialConstIndexed(GEPI))
7616     return false;
7617   ConstantInt *GEPIIdx = cast<ConstantInt>(GEPI->getOperand(1));
7618   // Check that GEPI is a cheap one.
7619   if (TTI->getIntImmCost(GEPIIdx->getValue(), GEPIIdx->getType(),
7620                          TargetTransformInfo::TCK_SizeAndLatency)
7621       > TargetTransformInfo::TCC_Basic)
7622     return false;
7623   Value *GEPIOp = GEPI->getOperand(0);
7624   // Check that GEPIOp is an instruction that's also defined in SrcBlock.
7625   if (!isa<Instruction>(GEPIOp))
7626     return false;
7627   auto *GEPIOpI = cast<Instruction>(GEPIOp);
7628   if (GEPIOpI->getParent() != SrcBlock)
7629     return false;
7630   // Check that GEP is used outside the block, meaning it's alive on the
7631   // IndirectBr edge(s).
7632   if (find_if(GEPI->users(), [&](User *Usr) {
7633         if (auto *I = dyn_cast<Instruction>(Usr)) {
7634           if (I->getParent() != SrcBlock) {
7635             return true;
7636           }
7637         }
7638         return false;
7639       }) == GEPI->users().end())
7640     return false;
7641   // The second elements of the GEP chains to be unmerged.
7642   std::vector<GetElementPtrInst *> UGEPIs;
7643   // Check each user of GEPIOp to check if unmerging would make GEPIOp not alive
7644   // on IndirectBr edges.
7645   for (User *Usr : GEPIOp->users()) {
7646     if (Usr == GEPI) continue;
7647     // Check if Usr is an Instruction. If not, give up.
7648     if (!isa<Instruction>(Usr))
7649       return false;
7650     auto *UI = cast<Instruction>(Usr);
7651     // Check if Usr in the same block as GEPIOp, which is fine, skip.
7652     if (UI->getParent() == SrcBlock)
7653       continue;
7654     // Check if Usr is a GEP. If not, give up.
7655     if (!isa<GetElementPtrInst>(Usr))
7656       return false;
7657     auto *UGEPI = cast<GetElementPtrInst>(Usr);
7658     // Check if UGEPI is a simple gep with a single constant index and GEPIOp is
7659     // the pointer operand to it. If so, record it in the vector. If not, give
7660     // up.
7661     if (!GEPSequentialConstIndexed(UGEPI))
7662       return false;
7663     if (UGEPI->getOperand(0) != GEPIOp)
7664       return false;
7665     if (GEPIIdx->getType() !=
7666         cast<ConstantInt>(UGEPI->getOperand(1))->getType())
7667       return false;
7668     ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
7669     if (TTI->getIntImmCost(UGEPIIdx->getValue(), UGEPIIdx->getType(),
7670                            TargetTransformInfo::TCK_SizeAndLatency)
7671         > TargetTransformInfo::TCC_Basic)
7672       return false;
7673     UGEPIs.push_back(UGEPI);
7674   }
7675   if (UGEPIs.size() == 0)
7676     return false;
7677   // Check the materializing cost of (Uidx-Idx).
7678   for (GetElementPtrInst *UGEPI : UGEPIs) {
7679     ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
7680     APInt NewIdx = UGEPIIdx->getValue() - GEPIIdx->getValue();
7681     InstructionCost ImmCost = TTI->getIntImmCost(
7682         NewIdx, GEPIIdx->getType(), TargetTransformInfo::TCK_SizeAndLatency);
7683     if (ImmCost > TargetTransformInfo::TCC_Basic)
7684       return false;
7685   }
7686   // Now unmerge between GEPI and UGEPIs.
7687   for (GetElementPtrInst *UGEPI : UGEPIs) {
7688     UGEPI->setOperand(0, GEPI);
7689     ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
7690     Constant *NewUGEPIIdx =
7691         ConstantInt::get(GEPIIdx->getType(),
7692                          UGEPIIdx->getValue() - GEPIIdx->getValue());
7693     UGEPI->setOperand(1, NewUGEPIIdx);
7694     // If GEPI is not inbounds but UGEPI is inbounds, change UGEPI to not
7695     // inbounds to avoid UB.
7696     if (!GEPI->isInBounds()) {
7697       UGEPI->setIsInBounds(false);
7698     }
7699   }
7700   // After unmerging, verify that GEPIOp is actually only used in SrcBlock (not
7701   // alive on IndirectBr edges).
7702   assert(find_if(GEPIOp->users(), [&](User *Usr) {
7703         return cast<Instruction>(Usr)->getParent() != SrcBlock;
7704       }) == GEPIOp->users().end() && "GEPIOp is used outside SrcBlock");
7705   return true;
7706 }
7707 
7708 static bool optimizeBranch(BranchInst *Branch, const TargetLowering &TLI) {
7709   // Try and convert
7710   //  %c = icmp ult %x, 8
7711   //  br %c, bla, blb
7712   //  %tc = lshr %x, 3
7713   // to
7714   //  %tc = lshr %x, 3
7715   //  %c = icmp eq %tc, 0
7716   //  br %c, bla, blb
7717   // Creating the cmp to zero can be better for the backend, especially if the
7718   // lshr produces flags that can be used automatically.
7719   if (!TLI.preferZeroCompareBranch() || !Branch->isConditional())
7720     return false;
7721 
7722   ICmpInst *Cmp = dyn_cast<ICmpInst>(Branch->getCondition());
7723   if (!Cmp || !isa<ConstantInt>(Cmp->getOperand(1)) || !Cmp->hasOneUse())
7724     return false;
7725 
7726   Value *X = Cmp->getOperand(0);
7727   APInt CmpC = cast<ConstantInt>(Cmp->getOperand(1))->getValue();
7728 
7729   for (auto *U : X->users()) {
7730     Instruction *UI = dyn_cast<Instruction>(U);
7731     // A quick dominance check
7732     if (!UI ||
7733         (UI->getParent() != Branch->getParent() &&
7734          UI->getParent() != Branch->getSuccessor(0) &&
7735          UI->getParent() != Branch->getSuccessor(1)) ||
7736         (UI->getParent() != Branch->getParent() &&
7737          !UI->getParent()->getSinglePredecessor()))
7738       continue;
7739 
7740     if (CmpC.isPowerOf2() && Cmp->getPredicate() == ICmpInst::ICMP_ULT &&
7741         match(UI, m_Shr(m_Specific(X), m_SpecificInt(CmpC.logBase2())))) {
7742       IRBuilder<> Builder(Branch);
7743       if (UI->getParent() != Branch->getParent())
7744         UI->moveBefore(Branch);
7745       Value *NewCmp = Builder.CreateCmp(ICmpInst::ICMP_EQ, UI,
7746                                         ConstantInt::get(UI->getType(), 0));
7747       LLVM_DEBUG(dbgs() << "Converting " << *Cmp << "\n");
7748       LLVM_DEBUG(dbgs() << " to compare on zero: " << *NewCmp << "\n");
7749       Cmp->replaceAllUsesWith(NewCmp);
7750       return true;
7751     }
7752     if (Cmp->isEquality() &&
7753         (match(UI, m_Add(m_Specific(X), m_SpecificInt(-CmpC))) ||
7754          match(UI, m_Sub(m_Specific(X), m_SpecificInt(CmpC))))) {
7755       IRBuilder<> Builder(Branch);
7756       if (UI->getParent() != Branch->getParent())
7757         UI->moveBefore(Branch);
7758       Value *NewCmp = Builder.CreateCmp(Cmp->getPredicate(), UI,
7759                                         ConstantInt::get(UI->getType(), 0));
7760       LLVM_DEBUG(dbgs() << "Converting " << *Cmp << "\n");
7761       LLVM_DEBUG(dbgs() << " to compare on zero: " << *NewCmp << "\n");
7762       Cmp->replaceAllUsesWith(NewCmp);
7763       return true;
7764     }
7765   }
7766   return false;
7767 }
7768 
7769 bool CodeGenPrepare::optimizeInst(Instruction *I, bool &ModifiedDT) {
7770   // Bail out if we inserted the instruction to prevent optimizations from
7771   // stepping on each other's toes.
7772   if (InsertedInsts.count(I))
7773     return false;
7774 
7775   // TODO: Move into the switch on opcode below here.
7776   if (PHINode *P = dyn_cast<PHINode>(I)) {
7777     // It is possible for very late stage optimizations (such as SimplifyCFG)
7778     // to introduce PHI nodes too late to be cleaned up.  If we detect such a
7779     // trivial PHI, go ahead and zap it here.
7780     if (Value *V = SimplifyInstruction(P, {*DL, TLInfo})) {
7781       LargeOffsetGEPMap.erase(P);
7782       P->replaceAllUsesWith(V);
7783       P->eraseFromParent();
7784       ++NumPHIsElim;
7785       return true;
7786     }
7787     return false;
7788   }
7789 
7790   if (CastInst *CI = dyn_cast<CastInst>(I)) {
7791     // If the source of the cast is a constant, then this should have
7792     // already been constant folded.  The only reason NOT to constant fold
7793     // it is if something (e.g. LSR) was careful to place the constant
7794     // evaluation in a block other than then one that uses it (e.g. to hoist
7795     // the address of globals out of a loop).  If this is the case, we don't
7796     // want to forward-subst the cast.
7797     if (isa<Constant>(CI->getOperand(0)))
7798       return false;
7799 
7800     if (OptimizeNoopCopyExpression(CI, *TLI, *DL))
7801       return true;
7802 
7803     if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7804       /// Sink a zext or sext into its user blocks if the target type doesn't
7805       /// fit in one register
7806       if (TLI->getTypeAction(CI->getContext(),
7807                              TLI->getValueType(*DL, CI->getType())) ==
7808           TargetLowering::TypeExpandInteger) {
7809         return SinkCast(CI);
7810       } else {
7811         bool MadeChange = optimizeExt(I);
7812         return MadeChange | optimizeExtUses(I);
7813       }
7814     }
7815     return false;
7816   }
7817 
7818   if (auto *Cmp = dyn_cast<CmpInst>(I))
7819     if (optimizeCmp(Cmp, ModifiedDT))
7820       return true;
7821 
7822   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7823     LI->setMetadata(LLVMContext::MD_invariant_group, nullptr);
7824     bool Modified = optimizeLoadExt(LI);
7825     unsigned AS = LI->getPointerAddressSpace();
7826     Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
7827     return Modified;
7828   }
7829 
7830   if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
7831     if (splitMergedValStore(*SI, *DL, *TLI))
7832       return true;
7833     SI->setMetadata(LLVMContext::MD_invariant_group, nullptr);
7834     unsigned AS = SI->getPointerAddressSpace();
7835     return optimizeMemoryInst(I, SI->getOperand(1),
7836                               SI->getOperand(0)->getType(), AS);
7837   }
7838 
7839   if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
7840       unsigned AS = RMW->getPointerAddressSpace();
7841       return optimizeMemoryInst(I, RMW->getPointerOperand(),
7842                                 RMW->getType(), AS);
7843   }
7844 
7845   if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(I)) {
7846       unsigned AS = CmpX->getPointerAddressSpace();
7847       return optimizeMemoryInst(I, CmpX->getPointerOperand(),
7848                                 CmpX->getCompareOperand()->getType(), AS);
7849   }
7850 
7851   BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
7852 
7853   if (BinOp && BinOp->getOpcode() == Instruction::And && EnableAndCmpSinking &&
7854       sinkAndCmp0Expression(BinOp, *TLI, InsertedInsts))
7855     return true;
7856 
7857   // TODO: Move this into the switch on opcode - it handles shifts already.
7858   if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
7859                 BinOp->getOpcode() == Instruction::LShr)) {
7860     ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
7861     if (CI && TLI->hasExtractBitsInsn())
7862       if (OptimizeExtractBits(BinOp, CI, *TLI, *DL))
7863         return true;
7864   }
7865 
7866   if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
7867     if (GEPI->hasAllZeroIndices()) {
7868       /// The GEP operand must be a pointer, so must its result -> BitCast
7869       Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
7870                                         GEPI->getName(), GEPI);
7871       NC->setDebugLoc(GEPI->getDebugLoc());
7872       GEPI->replaceAllUsesWith(NC);
7873       GEPI->eraseFromParent();
7874       ++NumGEPsElim;
7875       optimizeInst(NC, ModifiedDT);
7876       return true;
7877     }
7878     if (tryUnmergingGEPsAcrossIndirectBr(GEPI, TTI)) {
7879       return true;
7880     }
7881     return false;
7882   }
7883 
7884   if (FreezeInst *FI = dyn_cast<FreezeInst>(I)) {
7885     // freeze(icmp a, const)) -> icmp (freeze a), const
7886     // This helps generate efficient conditional jumps.
7887     Instruction *CmpI = nullptr;
7888     if (ICmpInst *II = dyn_cast<ICmpInst>(FI->getOperand(0)))
7889       CmpI = II;
7890     else if (FCmpInst *F = dyn_cast<FCmpInst>(FI->getOperand(0)))
7891       CmpI = F->getFastMathFlags().none() ? F : nullptr;
7892 
7893     if (CmpI && CmpI->hasOneUse()) {
7894       auto Op0 = CmpI->getOperand(0), Op1 = CmpI->getOperand(1);
7895       bool Const0 = isa<ConstantInt>(Op0) || isa<ConstantFP>(Op0) ||
7896                     isa<ConstantPointerNull>(Op0);
7897       bool Const1 = isa<ConstantInt>(Op1) || isa<ConstantFP>(Op1) ||
7898                     isa<ConstantPointerNull>(Op1);
7899       if (Const0 || Const1) {
7900         if (!Const0 || !Const1) {
7901           auto *F = new FreezeInst(Const0 ? Op1 : Op0, "", CmpI);
7902           F->takeName(FI);
7903           CmpI->setOperand(Const0 ? 1 : 0, F);
7904         }
7905         FI->replaceAllUsesWith(CmpI);
7906         FI->eraseFromParent();
7907         return true;
7908       }
7909     }
7910     return false;
7911   }
7912 
7913   if (tryToSinkFreeOperands(I))
7914     return true;
7915 
7916   switch (I->getOpcode()) {
7917   case Instruction::Shl:
7918   case Instruction::LShr:
7919   case Instruction::AShr:
7920     return optimizeShiftInst(cast<BinaryOperator>(I));
7921   case Instruction::Call:
7922     return optimizeCallInst(cast<CallInst>(I), ModifiedDT);
7923   case Instruction::Select:
7924     return optimizeSelectInst(cast<SelectInst>(I));
7925   case Instruction::ShuffleVector:
7926     return optimizeShuffleVectorInst(cast<ShuffleVectorInst>(I));
7927   case Instruction::Switch:
7928     return optimizeSwitchInst(cast<SwitchInst>(I));
7929   case Instruction::ExtractElement:
7930     return optimizeExtractElementInst(cast<ExtractElementInst>(I));
7931   case Instruction::Br:
7932     return optimizeBranch(cast<BranchInst>(I), *TLI);
7933   }
7934 
7935   return false;
7936 }
7937 
7938 /// Given an OR instruction, check to see if this is a bitreverse
7939 /// idiom. If so, insert the new intrinsic and return true.
7940 bool CodeGenPrepare::makeBitReverse(Instruction &I) {
7941   if (!I.getType()->isIntegerTy() ||
7942       !TLI->isOperationLegalOrCustom(ISD::BITREVERSE,
7943                                      TLI->getValueType(*DL, I.getType(), true)))
7944     return false;
7945 
7946   SmallVector<Instruction*, 4> Insts;
7947   if (!recognizeBSwapOrBitReverseIdiom(&I, false, true, Insts))
7948     return false;
7949   Instruction *LastInst = Insts.back();
7950   I.replaceAllUsesWith(LastInst);
7951   RecursivelyDeleteTriviallyDeadInstructions(
7952       &I, TLInfo, nullptr, [&](Value *V) { removeAllAssertingVHReferences(V); });
7953   return true;
7954 }
7955 
7956 // In this pass we look for GEP and cast instructions that are used
7957 // across basic blocks and rewrite them to improve basic-block-at-a-time
7958 // selection.
7959 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool &ModifiedDT) {
7960   SunkAddrs.clear();
7961   bool MadeChange = false;
7962 
7963   CurInstIterator = BB.begin();
7964   while (CurInstIterator != BB.end()) {
7965     MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
7966     if (ModifiedDT)
7967       return true;
7968   }
7969 
7970   bool MadeBitReverse = true;
7971   while (MadeBitReverse) {
7972     MadeBitReverse = false;
7973     for (auto &I : reverse(BB)) {
7974       if (makeBitReverse(I)) {
7975         MadeBitReverse = MadeChange = true;
7976         break;
7977       }
7978     }
7979   }
7980   MadeChange |= dupRetToEnableTailCallOpts(&BB, ModifiedDT);
7981 
7982   return MadeChange;
7983 }
7984 
7985 // Some CGP optimizations may move or alter what's computed in a block. Check
7986 // whether a dbg.value intrinsic could be pointed at a more appropriate operand.
7987 bool CodeGenPrepare::fixupDbgValue(Instruction *I) {
7988   assert(isa<DbgValueInst>(I));
7989   DbgValueInst &DVI = *cast<DbgValueInst>(I);
7990 
7991   // Does this dbg.value refer to a sunk address calculation?
7992   bool AnyChange = false;
7993   SmallDenseSet<Value *> LocationOps(DVI.location_ops().begin(),
7994                                      DVI.location_ops().end());
7995   for (Value *Location : LocationOps) {
7996     WeakTrackingVH SunkAddrVH = SunkAddrs[Location];
7997     Value *SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
7998     if (SunkAddr) {
7999       // Point dbg.value at locally computed address, which should give the best
8000       // opportunity to be accurately lowered. This update may change the type
8001       // of pointer being referred to; however this makes no difference to
8002       // debugging information, and we can't generate bitcasts that may affect
8003       // codegen.
8004       DVI.replaceVariableLocationOp(Location, SunkAddr);
8005       AnyChange = true;
8006     }
8007   }
8008   return AnyChange;
8009 }
8010 
8011 // A llvm.dbg.value may be using a value before its definition, due to
8012 // optimizations in this pass and others. Scan for such dbg.values, and rescue
8013 // them by moving the dbg.value to immediately after the value definition.
8014 // FIXME: Ideally this should never be necessary, and this has the potential
8015 // to re-order dbg.value intrinsics.
8016 bool CodeGenPrepare::placeDbgValues(Function &F) {
8017   bool MadeChange = false;
8018   DominatorTree DT(F);
8019 
8020   for (BasicBlock &BB : F) {
8021     for (Instruction &Insn : llvm::make_early_inc_range(BB)) {
8022       DbgValueInst *DVI = dyn_cast<DbgValueInst>(&Insn);
8023       if (!DVI)
8024         continue;
8025 
8026       SmallVector<Instruction *, 4> VIs;
8027       for (Value *V : DVI->getValues())
8028         if (Instruction *VI = dyn_cast_or_null<Instruction>(V))
8029           VIs.push_back(VI);
8030 
8031       // This DVI may depend on multiple instructions, complicating any
8032       // potential sink. This block takes the defensive approach, opting to
8033       // "undef" the DVI if it has more than one instruction and any of them do
8034       // not dominate DVI.
8035       for (Instruction *VI : VIs) {
8036         if (VI->isTerminator())
8037           continue;
8038 
8039         // If VI is a phi in a block with an EHPad terminator, we can't insert
8040         // after it.
8041         if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
8042           continue;
8043 
8044         // If the defining instruction dominates the dbg.value, we do not need
8045         // to move the dbg.value.
8046         if (DT.dominates(VI, DVI))
8047           continue;
8048 
8049         // If we depend on multiple instructions and any of them doesn't
8050         // dominate this DVI, we probably can't salvage it: moving it to
8051         // after any of the instructions could cause us to lose the others.
8052         if (VIs.size() > 1) {
8053           LLVM_DEBUG(
8054               dbgs()
8055               << "Unable to find valid location for Debug Value, undefing:\n"
8056               << *DVI);
8057           DVI->setUndef();
8058           break;
8059         }
8060 
8061         LLVM_DEBUG(dbgs() << "Moving Debug Value before :\n"
8062                           << *DVI << ' ' << *VI);
8063         DVI->removeFromParent();
8064         if (isa<PHINode>(VI))
8065           DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
8066         else
8067           DVI->insertAfter(VI);
8068         MadeChange = true;
8069         ++NumDbgValueMoved;
8070       }
8071     }
8072   }
8073   return MadeChange;
8074 }
8075 
8076 // Group scattered pseudo probes in a block to favor SelectionDAG. Scattered
8077 // probes can be chained dependencies of other regular DAG nodes and block DAG
8078 // combine optimizations.
8079 bool CodeGenPrepare::placePseudoProbes(Function &F) {
8080   bool MadeChange = false;
8081   for (auto &Block : F) {
8082     // Move the rest probes to the beginning of the block.
8083     auto FirstInst = Block.getFirstInsertionPt();
8084     while (FirstInst != Block.end() && FirstInst->isDebugOrPseudoInst())
8085       ++FirstInst;
8086     BasicBlock::iterator I(FirstInst);
8087     I++;
8088     while (I != Block.end()) {
8089       if (auto *II = dyn_cast<PseudoProbeInst>(I++)) {
8090         II->moveBefore(&*FirstInst);
8091         MadeChange = true;
8092       }
8093     }
8094   }
8095   return MadeChange;
8096 }
8097 
8098 /// Scale down both weights to fit into uint32_t.
8099 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
8100   uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
8101   uint32_t Scale = (NewMax / std::numeric_limits<uint32_t>::max()) + 1;
8102   NewTrue = NewTrue / Scale;
8103   NewFalse = NewFalse / Scale;
8104 }
8105 
8106 /// Some targets prefer to split a conditional branch like:
8107 /// \code
8108 ///   %0 = icmp ne i32 %a, 0
8109 ///   %1 = icmp ne i32 %b, 0
8110 ///   %or.cond = or i1 %0, %1
8111 ///   br i1 %or.cond, label %TrueBB, label %FalseBB
8112 /// \endcode
8113 /// into multiple branch instructions like:
8114 /// \code
8115 ///   bb1:
8116 ///     %0 = icmp ne i32 %a, 0
8117 ///     br i1 %0, label %TrueBB, label %bb2
8118 ///   bb2:
8119 ///     %1 = icmp ne i32 %b, 0
8120 ///     br i1 %1, label %TrueBB, label %FalseBB
8121 /// \endcode
8122 /// This usually allows instruction selection to do even further optimizations
8123 /// and combine the compare with the branch instruction. Currently this is
8124 /// applied for targets which have "cheap" jump instructions.
8125 ///
8126 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
8127 ///
8128 bool CodeGenPrepare::splitBranchCondition(Function &F, bool &ModifiedDT) {
8129   if (!TM->Options.EnableFastISel || TLI->isJumpExpensive())
8130     return false;
8131 
8132   bool MadeChange = false;
8133   for (auto &BB : F) {
8134     // Does this BB end with the following?
8135     //   %cond1 = icmp|fcmp|binary instruction ...
8136     //   %cond2 = icmp|fcmp|binary instruction ...
8137     //   %cond.or = or|and i1 %cond1, cond2
8138     //   br i1 %cond.or label %dest1, label %dest2"
8139     Instruction *LogicOp;
8140     BasicBlock *TBB, *FBB;
8141     if (!match(BB.getTerminator(),
8142                m_Br(m_OneUse(m_Instruction(LogicOp)), TBB, FBB)))
8143       continue;
8144 
8145     auto *Br1 = cast<BranchInst>(BB.getTerminator());
8146     if (Br1->getMetadata(LLVMContext::MD_unpredictable))
8147       continue;
8148 
8149     // The merging of mostly empty BB can cause a degenerate branch.
8150     if (TBB == FBB)
8151       continue;
8152 
8153     unsigned Opc;
8154     Value *Cond1, *Cond2;
8155     if (match(LogicOp,
8156               m_LogicalAnd(m_OneUse(m_Value(Cond1)), m_OneUse(m_Value(Cond2)))))
8157       Opc = Instruction::And;
8158     else if (match(LogicOp, m_LogicalOr(m_OneUse(m_Value(Cond1)),
8159                                         m_OneUse(m_Value(Cond2)))))
8160       Opc = Instruction::Or;
8161     else
8162       continue;
8163 
8164     auto IsGoodCond = [](Value *Cond) {
8165       return match(
8166           Cond,
8167           m_CombineOr(m_Cmp(), m_CombineOr(m_LogicalAnd(m_Value(), m_Value()),
8168                                            m_LogicalOr(m_Value(), m_Value()))));
8169     };
8170     if (!IsGoodCond(Cond1) || !IsGoodCond(Cond2))
8171       continue;
8172 
8173     LLVM_DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
8174 
8175     // Create a new BB.
8176     auto *TmpBB =
8177         BasicBlock::Create(BB.getContext(), BB.getName() + ".cond.split",
8178                            BB.getParent(), BB.getNextNode());
8179 
8180     // Update original basic block by using the first condition directly by the
8181     // branch instruction and removing the no longer needed and/or instruction.
8182     Br1->setCondition(Cond1);
8183     LogicOp->eraseFromParent();
8184 
8185     // Depending on the condition we have to either replace the true or the
8186     // false successor of the original branch instruction.
8187     if (Opc == Instruction::And)
8188       Br1->setSuccessor(0, TmpBB);
8189     else
8190       Br1->setSuccessor(1, TmpBB);
8191 
8192     // Fill in the new basic block.
8193     auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
8194     if (auto *I = dyn_cast<Instruction>(Cond2)) {
8195       I->removeFromParent();
8196       I->insertBefore(Br2);
8197     }
8198 
8199     // Update PHI nodes in both successors. The original BB needs to be
8200     // replaced in one successor's PHI nodes, because the branch comes now from
8201     // the newly generated BB (NewBB). In the other successor we need to add one
8202     // incoming edge to the PHI nodes, because both branch instructions target
8203     // now the same successor. Depending on the original branch condition
8204     // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
8205     // we perform the correct update for the PHI nodes.
8206     // This doesn't change the successor order of the just created branch
8207     // instruction (or any other instruction).
8208     if (Opc == Instruction::Or)
8209       std::swap(TBB, FBB);
8210 
8211     // Replace the old BB with the new BB.
8212     TBB->replacePhiUsesWith(&BB, TmpBB);
8213 
8214     // Add another incoming edge form the new BB.
8215     for (PHINode &PN : FBB->phis()) {
8216       auto *Val = PN.getIncomingValueForBlock(&BB);
8217       PN.addIncoming(Val, TmpBB);
8218     }
8219 
8220     // Update the branch weights (from SelectionDAGBuilder::
8221     // FindMergedConditions).
8222     if (Opc == Instruction::Or) {
8223       // Codegen X | Y as:
8224       // BB1:
8225       //   jmp_if_X TBB
8226       //   jmp TmpBB
8227       // TmpBB:
8228       //   jmp_if_Y TBB
8229       //   jmp FBB
8230       //
8231 
8232       // We have flexibility in setting Prob for BB1 and Prob for NewBB.
8233       // The requirement is that
8234       //   TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
8235       //     = TrueProb for original BB.
8236       // Assuming the original weights are A and B, one choice is to set BB1's
8237       // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
8238       // assumes that
8239       //   TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
8240       // Another choice is to assume TrueProb for BB1 equals to TrueProb for
8241       // TmpBB, but the math is more complicated.
8242       uint64_t TrueWeight, FalseWeight;
8243       if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) {
8244         uint64_t NewTrueWeight = TrueWeight;
8245         uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
8246         scaleWeights(NewTrueWeight, NewFalseWeight);
8247         Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
8248                          .createBranchWeights(TrueWeight, FalseWeight));
8249 
8250         NewTrueWeight = TrueWeight;
8251         NewFalseWeight = 2 * FalseWeight;
8252         scaleWeights(NewTrueWeight, NewFalseWeight);
8253         Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
8254                          .createBranchWeights(TrueWeight, FalseWeight));
8255       }
8256     } else {
8257       // Codegen X & Y as:
8258       // BB1:
8259       //   jmp_if_X TmpBB
8260       //   jmp FBB
8261       // TmpBB:
8262       //   jmp_if_Y TBB
8263       //   jmp FBB
8264       //
8265       //  This requires creation of TmpBB after CurBB.
8266 
8267       // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
8268       // The requirement is that
8269       //   FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
8270       //     = FalseProb for original BB.
8271       // Assuming the original weights are A and B, one choice is to set BB1's
8272       // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
8273       // assumes that
8274       //   FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
8275       uint64_t TrueWeight, FalseWeight;
8276       if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) {
8277         uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
8278         uint64_t NewFalseWeight = FalseWeight;
8279         scaleWeights(NewTrueWeight, NewFalseWeight);
8280         Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
8281                          .createBranchWeights(TrueWeight, FalseWeight));
8282 
8283         NewTrueWeight = 2 * TrueWeight;
8284         NewFalseWeight = FalseWeight;
8285         scaleWeights(NewTrueWeight, NewFalseWeight);
8286         Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
8287                          .createBranchWeights(TrueWeight, FalseWeight));
8288       }
8289     }
8290 
8291     ModifiedDT = true;
8292     MadeChange = true;
8293 
8294     LLVM_DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
8295                TmpBB->dump());
8296   }
8297   return MadeChange;
8298 }
8299