LCOV - code coverage report
Current view: top level - lib/CodeGen - CodeGenPrepare.cpp (source / functions) Hit Total Coverage
Test: llvm-toolchain.info Lines: 1895 1977 95.9 %
Date: 2018-02-22 04:41:24 Functions: 134 144 93.1 %
Legend: Lines: hit not hit

          Line data    Source code
       1             : //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
       2             : //
       3             : //                     The LLVM Compiler Infrastructure
       4             : //
       5             : // This file is distributed under the University of Illinois Open Source
       6             : // License. See LICENSE.TXT for details.
       7             : //
       8             : //===----------------------------------------------------------------------===//
       9             : //
      10             : // This pass munges the code in the input function to better prepare it for
      11             : // SelectionDAG-based code generation. This works around limitations in it's
      12             : // basic-block-at-a-time approach. It should eventually be removed.
      13             : //
      14             : //===----------------------------------------------------------------------===//
      15             : 
      16             : #include "llvm/ADT/APInt.h"
      17             : #include "llvm/ADT/ArrayRef.h"
      18             : #include "llvm/ADT/DenseMap.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/CodeGen/Analysis.h"
      35             : #include "llvm/CodeGen/ISDOpcodes.h"
      36             : #include "llvm/CodeGen/MachineValueType.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/IR/Argument.h"
      43             : #include "llvm/IR/Attributes.h"
      44             : #include "llvm/IR/BasicBlock.h"
      45             : #include "llvm/IR/CallSite.h"
      46             : #include "llvm/IR/Constant.h"
      47             : #include "llvm/IR/Constants.h"
      48             : #include "llvm/IR/DataLayout.h"
      49             : #include "llvm/IR/DerivedTypes.h"
      50             : #include "llvm/IR/Dominators.h"
      51             : #include "llvm/IR/Function.h"
      52             : #include "llvm/IR/GetElementPtrTypeIterator.h"
      53             : #include "llvm/IR/GlobalValue.h"
      54             : #include "llvm/IR/GlobalVariable.h"
      55             : #include "llvm/IR/IRBuilder.h"
      56             : #include "llvm/IR/InlineAsm.h"
      57             : #include "llvm/IR/InstrTypes.h"
      58             : #include "llvm/IR/Instruction.h"
      59             : #include "llvm/IR/Instructions.h"
      60             : #include "llvm/IR/IntrinsicInst.h"
      61             : #include "llvm/IR/Intrinsics.h"
      62             : #include "llvm/IR/LLVMContext.h"
      63             : #include "llvm/IR/MDBuilder.h"
      64             : #include "llvm/IR/Module.h"
      65             : #include "llvm/IR/Operator.h"
      66             : #include "llvm/IR/PatternMatch.h"
      67             : #include "llvm/IR/Statepoint.h"
      68             : #include "llvm/IR/Type.h"
      69             : #include "llvm/IR/Use.h"
      70             : #include "llvm/IR/User.h"
      71             : #include "llvm/IR/Value.h"
      72             : #include "llvm/IR/ValueHandle.h"
      73             : #include "llvm/IR/ValueMap.h"
      74             : #include "llvm/Pass.h"
      75             : #include "llvm/Support/BlockFrequency.h"
      76             : #include "llvm/Support/BranchProbability.h"
      77             : #include "llvm/Support/Casting.h"
      78             : #include "llvm/Support/CommandLine.h"
      79             : #include "llvm/Support/Compiler.h"
      80             : #include "llvm/Support/Debug.h"
      81             : #include "llvm/Support/ErrorHandling.h"
      82             : #include "llvm/Support/MathExtras.h"
      83             : #include "llvm/Support/raw_ostream.h"
      84             : #include "llvm/Target/TargetMachine.h"
      85             : #include "llvm/Target/TargetOptions.h"
      86             : #include "llvm/Transforms/Utils/BasicBlockUtils.h"
      87             : #include "llvm/Transforms/Utils/BypassSlowDivision.h"
      88             : #include "llvm/Transforms/Utils/Local.h"
      89             : #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
      90             : #include <algorithm>
      91             : #include <cassert>
      92             : #include <cstdint>
      93             : #include <iterator>
      94             : #include <limits>
      95             : #include <memory>
      96             : #include <utility>
      97             : #include <vector>
      98             : 
      99             : using namespace llvm;
     100             : using namespace llvm::PatternMatch;
     101             : 
     102             : #define DEBUG_TYPE "codegenprepare"
     103             : 
     104             : STATISTIC(NumBlocksElim, "Number of blocks eliminated");
     105             : STATISTIC(NumPHIsElim,   "Number of trivial PHIs eliminated");
     106             : STATISTIC(NumGEPsElim,   "Number of GEPs converted to casts");
     107             : STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
     108             :                       "sunken Cmps");
     109             : STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
     110             :                        "of sunken Casts");
     111             : STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
     112             :                           "computations were sunk");
     113             : STATISTIC(NumMemoryInstsPhiCreated,
     114             :           "Number of phis created when address "
     115             :           "computations were sunk to memory instructions");
     116             : STATISTIC(NumMemoryInstsSelectCreated,
     117             :           "Number of select created when address "
     118             :           "computations were sunk to memory instructions");
     119             : STATISTIC(NumExtsMoved,  "Number of [s|z]ext instructions combined with loads");
     120             : STATISTIC(NumExtUses,    "Number of uses of [s|z]ext instructions optimized");
     121             : STATISTIC(NumAndsAdded,
     122             :           "Number of and mask instructions added to form ext loads");
     123             : STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
     124             : STATISTIC(NumRetsDup,    "Number of return instructions duplicated");
     125             : STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
     126             : STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
     127             : STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
     128             : 
     129       81626 : static cl::opt<bool> DisableBranchOpts(
     130      163252 :   "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
     131      163252 :   cl::desc("Disable branch optimizations in CodeGenPrepare"));
     132             : 
     133             : static cl::opt<bool>
     134      244878 :     DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
     135      163252 :                   cl::desc("Disable GC optimizations in CodeGenPrepare"));
     136             : 
     137       81626 : static cl::opt<bool> DisableSelectToBranch(
     138      163252 :   "disable-cgp-select2branch", cl::Hidden, cl::init(false),
     139      163252 :   cl::desc("Disable select to branch conversion."));
     140             : 
     141       81626 : static cl::opt<bool> AddrSinkUsingGEPs(
     142      163252 :   "addr-sink-using-gep", cl::Hidden, cl::init(true),
     143      163252 :   cl::desc("Address sinking in CGP using GEPs."));
     144             : 
     145       81626 : static cl::opt<bool> EnableAndCmpSinking(
     146      163252 :    "enable-andcmp-sinking", cl::Hidden, cl::init(true),
     147      163252 :    cl::desc("Enable sinkinig and/cmp into branches."));
     148             : 
     149       81626 : static cl::opt<bool> DisableStoreExtract(
     150      163252 :     "disable-cgp-store-extract", cl::Hidden, cl::init(false),
     151      163252 :     cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
     152             : 
     153       81626 : static cl::opt<bool> StressStoreExtract(
     154      163252 :     "stress-cgp-store-extract", cl::Hidden, cl::init(false),
     155      163252 :     cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
     156             : 
     157       81626 : static cl::opt<bool> DisableExtLdPromotion(
     158      163252 :     "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
     159       81626 :     cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
     160       81626 :              "CodeGenPrepare"));
     161             : 
     162       81626 : static cl::opt<bool> StressExtLdPromotion(
     163      163252 :     "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
     164       81626 :     cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
     165       81626 :              "optimization in CodeGenPrepare"));
     166             : 
     167       81626 : static cl::opt<bool> DisablePreheaderProtect(
     168      163252 :     "disable-preheader-prot", cl::Hidden, cl::init(false),
     169      163252 :     cl::desc("Disable protection against removing loop preheaders"));
     170             : 
     171       81626 : static cl::opt<bool> ProfileGuidedSectionPrefix(
     172      163252 :     "profile-guided-section-prefix", cl::Hidden, cl::init(true), cl::ZeroOrMore,
     173      244878 :     cl::desc("Use profile info to add section prefix for hot/cold functions"));
     174             : 
     175       81626 : static cl::opt<unsigned> FreqRatioToSkipMerge(
     176      163252 :     "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2),
     177       81626 :     cl::desc("Skip merging empty blocks if (frequency of empty block) / "
     178       81626 :              "(frequency of destination block) is greater than this ratio"));
     179             : 
     180       81626 : static cl::opt<bool> ForceSplitStore(
     181      163252 :     "force-split-store", cl::Hidden, cl::init(false),
     182      163252 :     cl::desc("Force store splitting no matter what the target query says."));
     183             : 
     184             : static cl::opt<bool>
     185       81626 : EnableTypePromotionMerge("cgp-type-promotion-merge", cl::Hidden,
     186       81626 :     cl::desc("Enable merging of redundant sexts when one is dominating"
     187      244878 :     " the other."), cl::init(true));
     188             : 
     189       81626 : static cl::opt<bool> DisableComplexAddrModes(
     190      163252 :     "disable-complex-addr-modes", cl::Hidden, cl::init(false),
     191       81626 :     cl::desc("Disables combining addressing modes with different parts "
     192       81626 :              "in optimizeMemoryInst."));
     193             : 
     194             : static cl::opt<bool>
     195      244878 : AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(false),
     196      163252 :                 cl::desc("Allow creation of Phis in Address sinking."));
     197             : 
     198             : static cl::opt<bool>
     199      244878 : AddrSinkNewSelects("addr-sink-new-select", cl::Hidden, cl::init(true),
     200      163252 :                    cl::desc("Allow creation of selects in Address sinking."));
     201             : 
     202       81626 : static cl::opt<bool> AddrSinkCombineBaseReg(
     203      163252 :     "addr-sink-combine-base-reg", cl::Hidden, cl::init(true),
     204      163252 :     cl::desc("Allow combining of BaseReg field in Address sinking."));
     205             : 
     206       81626 : static cl::opt<bool> AddrSinkCombineBaseGV(
     207      163252 :     "addr-sink-combine-base-gv", cl::Hidden, cl::init(true),
     208      163252 :     cl::desc("Allow combining of BaseGV field in Address sinking."));
     209             : 
     210       81626 : static cl::opt<bool> AddrSinkCombineBaseOffs(
     211      163252 :     "addr-sink-combine-base-offs", cl::Hidden, cl::init(true),
     212      163252 :     cl::desc("Allow combining of BaseOffs field in Address sinking."));
     213             : 
     214       81626 : static cl::opt<bool> AddrSinkCombineScaledReg(
     215      163252 :     "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(true),
     216      163252 :     cl::desc("Allow combining of ScaledReg field in Address sinking."));
     217             : 
     218             : namespace {
     219             : 
     220             : using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
     221             : using TypeIsSExt = PointerIntPair<Type *, 1, bool>;
     222             : using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
     223             : using SExts = SmallVector<Instruction *, 16>;
     224             : using ValueToSExts = DenseMap<Value *, SExts>;
     225             : 
     226             : class TypePromotionTransaction;
     227             : 
     228       69519 :   class CodeGenPrepare : public FunctionPass {
     229             :     const TargetMachine *TM = nullptr;
     230             :     const TargetSubtargetInfo *SubtargetInfo;
     231             :     const TargetLowering *TLI = nullptr;
     232             :     const TargetRegisterInfo *TRI;
     233             :     const TargetTransformInfo *TTI = nullptr;
     234             :     const TargetLibraryInfo *TLInfo;
     235             :     const LoopInfo *LI;
     236             :     std::unique_ptr<BlockFrequencyInfo> BFI;
     237             :     std::unique_ptr<BranchProbabilityInfo> BPI;
     238             : 
     239             :     /// As we scan instructions optimizing them, this is the next instruction
     240             :     /// to optimize. Transforms that can invalidate this should update it.
     241             :     BasicBlock::iterator CurInstIterator;
     242             : 
     243             :     /// Keeps track of non-local addresses that have been sunk into a block.
     244             :     /// This allows us to avoid inserting duplicate code for blocks with
     245             :     /// multiple load/stores of the same address. The usage of WeakTrackingVH
     246             :     /// enables SunkAddrs to be treated as a cache whose entries can be
     247             :     /// invalidated if a sunken address computation has been erased.
     248             :     ValueMap<Value*, WeakTrackingVH> SunkAddrs;
     249             : 
     250             :     /// Keeps track of all instructions inserted for the current function.
     251             :     SetOfInstrs InsertedInsts;
     252             : 
     253             :     /// Keeps track of the type of the related instruction before their
     254             :     /// promotion for the current function.
     255             :     InstrToOrigTy PromotedInsts;
     256             : 
     257             :     /// Keep track of instructions removed during promotion.
     258             :     SetOfInstrs RemovedInsts;
     259             : 
     260             :     /// Keep track of sext chains based on their initial value.
     261             :     DenseMap<Value *, Instruction *> SeenChainsForSExt;
     262             : 
     263             :     /// Keep track of SExt promoted.
     264             :     ValueToSExts ValToSExtendedUses;
     265             : 
     266             :     /// True if CFG is modified in any way.
     267             :     bool ModifiedDT;
     268             : 
     269             :     /// True if optimizing for size.
     270             :     bool OptSize;
     271             : 
     272             :     /// DataLayout for the Function being processed.
     273             :     const DataLayout *DL = nullptr;
     274             : 
     275             :   public:
     276             :     static char ID; // Pass identification, replacement for typeid
     277             : 
     278      104886 :     CodeGenPrepare() : FunctionPass(ID) {
     279       17481 :       initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
     280       17481 :     }
     281             : 
     282             :     bool runOnFunction(Function &F) override;
     283             : 
     284          28 :     StringRef getPassName() const override { return "CodeGen Prepare"; }
     285             : 
     286       17399 :     void getAnalysisUsage(AnalysisUsage &AU) const override {
     287             :       // FIXME: When we can selectively preserve passes, preserve the domtree.
     288             :       AU.addRequired<ProfileSummaryInfoWrapperPass>();
     289             :       AU.addRequired<TargetLibraryInfoWrapperPass>();
     290             :       AU.addRequired<TargetTransformInfoWrapperPass>();
     291             :       AU.addRequired<LoopInfoWrapperPass>();
     292       17399 :     }
     293             : 
     294             :   private:
     295             :     bool eliminateFallThrough(Function &F);
     296             :     bool eliminateMostlyEmptyBlocks(Function &F);
     297             :     BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
     298             :     bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
     299             :     void eliminateMostlyEmptyBlock(BasicBlock *BB);
     300             :     bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
     301             :                                        bool isPreheader);
     302             :     bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
     303             :     bool optimizeInst(Instruction *I, bool &ModifiedDT);
     304             :     bool optimizeMemoryInst(Instruction *I, Value *Addr,
     305             :                             Type *AccessTy, unsigned AS);
     306             :     bool optimizeInlineAsmInst(CallInst *CS);
     307             :     bool optimizeCallInst(CallInst *CI, bool &ModifiedDT);
     308             :     bool optimizeExt(Instruction *&I);
     309             :     bool optimizeExtUses(Instruction *I);
     310             :     bool optimizeLoadExt(LoadInst *I);
     311             :     bool optimizeSelectInst(SelectInst *SI);
     312             :     bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
     313             :     bool optimizeSwitchInst(SwitchInst *CI);
     314             :     bool optimizeExtractElementInst(Instruction *Inst);
     315             :     bool dupRetToEnableTailCallOpts(BasicBlock *BB);
     316             :     bool placeDbgValues(Function &F);
     317             :     bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
     318             :                       LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
     319             :     bool tryToPromoteExts(TypePromotionTransaction &TPT,
     320             :                           const SmallVectorImpl<Instruction *> &Exts,
     321             :                           SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
     322             :                           unsigned CreatedInstsCost = 0);
     323             :     bool mergeSExts(Function &F);
     324             :     bool performAddressTypePromotion(
     325             :         Instruction *&Inst,
     326             :         bool AllowPromotionWithoutCommonHeader,
     327             :         bool HasPromoted, TypePromotionTransaction &TPT,
     328             :         SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
     329             :     bool splitBranchCondition(Function &F);
     330             :     bool simplifyOffsetableRelocate(Instruction &I);
     331             :   };
     332             : 
     333             : } // end anonymous namespace
     334             : 
     335             : char CodeGenPrepare::ID = 0;
     336             : 
     337       29025 : INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,
     338             :                       "Optimize for code generation", false, false)
     339       29025 : INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
     340      194189 : INITIALIZE_PASS_END(CodeGenPrepare, DEBUG_TYPE,
     341             :                     "Optimize for code generation", false, false)
     342             : 
     343       17398 : FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); }
     344             : 
     345      160505 : bool CodeGenPrepare::runOnFunction(Function &F) {
     346      160505 :   if (skipFunction(F))
     347             :     return false;
     348             : 
     349      160412 :   DL = &F.getParent()->getDataLayout();
     350             : 
     351             :   bool EverMadeChange = false;
     352             :   // Clear per function information.
     353      160412 :   InsertedInsts.clear();
     354      160412 :   PromotedInsts.clear();
     355             : 
     356      160412 :   ModifiedDT = false;
     357      160412 :   if (auto *TPC = getAnalysisIfAvailable<TargetPassConfig>()) {
     358      160403 :     TM = &TPC->getTM<TargetMachine>();
     359      160403 :     SubtargetInfo = TM->getSubtargetImpl(F);
     360      160403 :     TLI = SubtargetInfo->getTargetLowering();
     361      160403 :     TRI = SubtargetInfo->getRegisterInfo();
     362             :   }
     363      320824 :   TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
     364      160412 :   TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
     365      320824 :   LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
     366      160412 :   BPI.reset(new BranchProbabilityInfo(F, *LI));
     367      160412 :   BFI.reset(new BlockFrequencyInfo(F, *BPI, *LI));
     368      160412 :   OptSize = F.optForSize();
     369             : 
     370             :   ProfileSummaryInfo *PSI =
     371      160412 :       getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
     372      160412 :   if (ProfileGuidedSectionPrefix) {
     373      160412 :     if (PSI->isFunctionHotInCallGraph(&F, *BFI))
     374          11 :       F.setSectionPrefix(".hot");
     375      160401 :     else if (PSI->isFunctionColdInCallGraph(&F, *BFI))
     376           6 :       F.setSectionPrefix(".unlikely");
     377             :   }
     378             : 
     379             :   /// This optimization identifies DIV instructions that can be
     380             :   /// profitably bypassed and carried out with a shorter, faster divide.
     381      315178 :   if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI &&
     382             :       TLI->isSlowDivBypassed()) {
     383             :     const DenseMap<unsigned int, unsigned int> &BypassWidths =
     384             :        TLI->getBypassSlowDivWidths();
     385             :     BasicBlock* BB = &*F.begin();
     386       36481 :     while (BB != nullptr) {
     387             :       // bypassSlowDivision may create new BBs, but we don't want to reapply the
     388             :       // optimization to those blocks.
     389             :       BasicBlock* Next = BB->getNextNode();
     390       12319 :       EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
     391             :       BB = Next;
     392             :     }
     393             :   }
     394             : 
     395             :   // Eliminate blocks that contain only PHI nodes and an
     396             :   // unconditional branch.
     397      160412 :   EverMadeChange |= eliminateMostlyEmptyBlocks(F);
     398             : 
     399             :   // llvm.dbg.value is far away from the value then iSel may not be able
     400             :   // handle it properly. iSel will drop llvm.dbg.value if it can not
     401             :   // find a node corresponding to the value.
     402      160412 :   EverMadeChange |= placeDbgValues(F);
     403             : 
     404      160412 :   if (!DisableBranchOpts)
     405      160349 :     EverMadeChange |= splitBranchCondition(F);
     406             : 
     407             :   // Split some critical edges where one of the sources is an indirect branch,
     408             :   // to help generate sane code for PHIs involving such edges.
     409      160412 :   EverMadeChange |= SplitIndirectBrCriticalEdges(F);
     410             : 
     411             :   bool MadeChange = true;
     412      496294 :   while (MadeChange) {
     413             :     MadeChange = false;
     414      167941 :     SeenChainsForSExt.clear();
     415      167941 :     ValToSExtendedUses.clear();
     416      167941 :     RemovedInsts.clear();
     417     1041479 :     for (Function::iterator I = F.begin(); I != F.end(); ) {
     418             :       BasicBlock *BB = &*I++;
     419      436804 :       bool ModifiedDTOnIteration = false;
     420      436804 :       MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
     421             : 
     422             :       // Restart BB iteration if the dominator tree of the Function was changed
     423      436804 :       if (ModifiedDTOnIteration)
     424             :         break;
     425             :     }
     426      335882 :     if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
     427          36 :       MadeChange |= mergeSExts(F);
     428             : 
     429             :     // Really free removed instructions during promotion.
     430      167941 :     for (Instruction *I : RemovedInsts)
     431         142 :       I->deleteValue();
     432             : 
     433      167941 :     EverMadeChange |= MadeChange;
     434             :   }
     435             : 
     436             :   SunkAddrs.clear();
     437             : 
     438      160412 :   if (!DisableBranchOpts) {
     439             :     MadeChange = false;
     440             :     SmallPtrSet<BasicBlock*, 8> WorkList;
     441      464780 :     for (BasicBlock &BB : F) {
     442      304431 :       SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
     443      304431 :       MadeChange |= ConstantFoldTerminator(&BB, true);
     444      304431 :       if (!MadeChange) continue;
     445             : 
     446        7262 :       for (SmallVectorImpl<BasicBlock*>::iterator
     447       11817 :              II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
     448       14524 :         if (pred_begin(*II) == pred_end(*II))
     449        1085 :           WorkList.insert(*II);
     450             :     }
     451             : 
     452             :     // Delete the dead blocks and any of their dead successors.
     453      160349 :     MadeChange |= !WorkList.empty();
     454      163731 :     while (!WorkList.empty()) {
     455        3382 :       BasicBlock *BB = *WorkList.begin();
     456             :       WorkList.erase(BB);
     457        1691 :       SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
     458             : 
     459        1691 :       DeleteDeadBlock(BB);
     460             : 
     461        1385 :       for (SmallVectorImpl<BasicBlock*>::iterator
     462        3076 :              II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
     463        2770 :         if (pred_begin(*II) == pred_end(*II))
     464         777 :           WorkList.insert(*II);
     465             :     }
     466             : 
     467             :     // Merge pairs of basic blocks with unconditional branches, connected by
     468             :     // a single edge.
     469      160349 :     if (EverMadeChange || MadeChange)
     470        8324 :       MadeChange |= eliminateFallThrough(F);
     471             : 
     472      160349 :     EverMadeChange |= MadeChange;
     473             :   }
     474             : 
     475      160412 :   if (!DisableGCOpts) {
     476             :     SmallVector<Instruction *, 2> Statepoints;
     477      461841 :     for (BasicBlock &BB : F)
     478     2405756 :       for (Instruction &I : BB)
     479     2104329 :         if (isStatepoint(I))
     480          75 :           Statepoints.push_back(&I);
     481      160562 :     for (auto &I : Statepoints)
     482          75 :       EverMadeChange |= simplifyOffsetableRelocate(*I);
     483             :   }
     484             : 
     485             :   return EverMadeChange;
     486             : }
     487             : 
     488             : /// Merge basic blocks which are connected by a single edge, where one of the
     489             : /// basic blocks has a single successor pointing to the other basic block,
     490             : /// which has a single predecessor.
     491        8324 : bool CodeGenPrepare::eliminateFallThrough(Function &F) {
     492             :   bool Changed = false;
     493             :   // Scan all of the blocks in the function, except for the entry block.
     494      135240 :   for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
     495             :     BasicBlock *BB = &*I++;
     496             :     // If the destination block has a single pred, then this is a trivial
     497             :     // edge, just collapse it.
     498             :     BasicBlock *SinglePred = BB->getSinglePredecessor();
     499             : 
     500             :     // Don't merge if BB's address is taken.
     501      213899 :     if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
     502             : 
     503             :     BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
     504       41046 :     if (Term && !Term->isConditional()) {
     505             :       Changed = true;
     506             :       DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
     507             :       // Remember if SinglePred was the entry block of the function.
     508             :       // If so, we will need to move BB back to the entry position.
     509        1427 :       bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
     510        1427 :       MergeBasicBlockIntoOnlyPred(BB, nullptr);
     511             : 
     512        1930 :       if (isEntry && BB != &BB->getParent()->getEntryBlock())
     513           0 :         BB->moveBefore(&BB->getParent()->getEntryBlock());
     514             : 
     515             :       // We have erased a block. Update the iterator.
     516        1427 :       I = BB->getIterator();
     517             :     }
     518             :   }
     519        8324 :   return Changed;
     520             : }
     521             : 
     522             : /// Find a destination block from BB if BB is mergeable empty block.
     523      147494 : BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
     524             :   // If this block doesn't end with an uncond branch, ignore it.
     525             :   BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
     526       94925 :   if (!BI || !BI->isUnconditional())
     527             :     return nullptr;
     528             : 
     529             :   // If the instruction before the branch (skipping debug info) isn't a phi
     530             :   // node, then other stuff is happening here.
     531       60927 :   BasicBlock::iterator BBI = BI->getIterator();
     532       60927 :   if (BBI != BB->begin()) {
     533             :     --BBI;
     534       55830 :     while (isa<DbgInfoIntrinsic>(BBI)) {
     535         360 :       if (BBI == BB->begin())
     536             :         break;
     537             :       --BBI;
     538             :     }
     539      110948 :     if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
     540             :       return nullptr;
     541             :   }
     542             : 
     543             :   // Do not break infinite loops.
     544             :   BasicBlock *DestBB = BI->getSuccessor(0);
     545        5750 :   if (DestBB == BB)
     546             :     return nullptr;
     547             : 
     548        5655 :   if (!canMergeBlocks(BB, DestBB))
     549             :     DestBB = nullptr;
     550             : 
     551             :   return DestBB;
     552             : }
     553             : 
     554             : /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
     555             : /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
     556             : /// edges in ways that are non-optimal for isel. Start by eliminating these
     557             : /// blocks so we can split them the way we want them.
     558      160412 : bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
     559             :   SmallPtrSet<BasicBlock *, 16> Preheaders;
     560      160412 :   SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
     561      166615 :   while (!LoopList.empty()) {
     562             :     Loop *L = LoopList.pop_back_val();
     563        6203 :     LoopList.insert(LoopList.end(), L->begin(), L->end());
     564        6203 :     if (BasicBlock *Preheader = L->getLoopPreheader())
     565        6167 :       Preheaders.insert(Preheader);
     566             :   }
     567             : 
     568             :   bool MadeChange = false;
     569             :   // Note that this intentionally skips the entry block.
     570      307899 :   for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
     571             :     BasicBlock *BB = &*I++;
     572      147487 :     BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
     573      152682 :     if (!DestBB ||
     574        5195 :         !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB)))
     575      143940 :       continue;
     576             : 
     577        3547 :     eliminateMostlyEmptyBlock(BB);
     578             :     MadeChange = true;
     579             :   }
     580      160412 :   return MadeChange;
     581             : }
     582             : 
     583        5195 : bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
     584             :                                                    BasicBlock *DestBB,
     585             :                                                    bool isPreheader) {
     586             :   // Do not delete loop preheaders if doing so would create a critical edge.
     587             :   // Loop preheaders can be good locations to spill registers. If the
     588             :   // preheader is deleted and we create a critical edge, registers may be
     589             :   // spilled in the loop body instead.
     590        6789 :   if (!DisablePreheaderProtect && isPreheader &&
     591        3090 :       !(BB->getSinglePredecessor() &&
     592        1496 :         BB->getSinglePredecessor()->getSingleSuccessor()))
     593             :     return false;
     594             : 
     595             :   // Try to skip merging if the unique predecessor of BB is terminated by a
     596             :   // switch or indirect branch instruction, and BB is used as an incoming block
     597             :   // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
     598             :   // add COPY instructions in the predecessor of BB instead of BB (if it is not
     599             :   // merged). Note that the critical edge created by merging such blocks wont be
     600             :   // split in MachineSink because the jump table is not analyzable. By keeping
     601             :   // such empty block (BB), ISel will place COPY instructions in BB, not in the
     602             :   // predecessor of BB.
     603        3606 :   BasicBlock *Pred = BB->getUniquePredecessor();
     604        6706 :   if (!Pred ||
     605        2679 :       !(isa<SwitchInst>(Pred->getTerminator()) ||
     606             :         isa<IndirectBrInst>(Pred->getTerminator())))
     607             :     return true;
     608             : 
     609         864 :   if (BB->getTerminator() != BB->getFirstNonPHI())
     610             :     return true;
     611             : 
     612             :   // We use a simple cost heuristic which determine skipping merging is
     613             :   // profitable if the cost of skipping merging is less than the cost of
     614             :   // merging : Cost(skipping merging) < Cost(merging BB), where the
     615             :   // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
     616             :   // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
     617             :   // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
     618             :   //   Freq(Pred) / Freq(BB) > 2.
     619             :   // Note that if there are multiple empty blocks sharing the same incoming
     620             :   // value for the PHIs in the DestBB, we consider them together. In such
     621             :   // case, Cost(merging BB) will be the sum of their frequencies.
     622             : 
     623         432 :   if (!isa<PHINode>(DestBB->begin()))
     624             :     return true;
     625             : 
     626             :   SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
     627             : 
     628             :   // Find all other incoming blocks from which incoming values of all PHIs in
     629             :   // DestBB are the same as the ones from BB.
     630         411 :   for (pred_iterator PI = pred_begin(DestBB), E = pred_end(DestBB); PI != E;
     631             :        ++PI) {
     632         263 :     BasicBlock *DestBBPred = *PI;
     633         263 :     if (DestBBPred == BB)
     634          74 :       continue;
     635             : 
     636         378 :     if (llvm::all_of(DestBB->phis(), [&](const PHINode &DestPN) {
     637         189 :           return DestPN.getIncomingValueForBlock(BB) ==
     638         189 :                  DestPN.getIncomingValueForBlock(DestBBPred);
     639         189 :         }))
     640          36 :       SameIncomingValueBBs.insert(DestBBPred);
     641             :   }
     642             : 
     643             :   // See if all BB's incoming values are same as the value from Pred. In this
     644             :   // case, no reason to skip merging because COPYs are expected to be place in
     645             :   // Pred already.
     646          74 :   if (SameIncomingValueBBs.count(Pred))
     647             :     return true;
     648             : 
     649          68 :   BlockFrequency PredFreq = BFI->getBlockFreq(Pred);
     650          68 :   BlockFrequency BBFreq = BFI->getBlockFreq(BB);
     651             : 
     652          68 :   for (auto SameValueBB : SameIncomingValueBBs)
     653          22 :     if (SameValueBB->getUniquePredecessor() == Pred &&
     654           7 :         DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB))
     655           4 :       BBFreq += BFI->getBlockFreq(SameValueBB);
     656             : 
     657             :   return PredFreq.getFrequency() <=
     658          68 :          BBFreq.getFrequency() * FreqRatioToSkipMerge;
     659             : }
     660             : 
     661             : /// Return true if we can merge BB into DestBB if there is a single
     662             : /// unconditional branch between them, and BB contains no other non-phi
     663             : /// instructions.
     664        5655 : bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
     665             :                                     const BasicBlock *DestBB) const {
     666             :   // We only want to eliminate blocks whose phi nodes are used by phi nodes in
     667             :   // the successor.  If there are more complex condition (e.g. preheaders),
     668             :   // don't mess around with them.
     669        5885 :   for (const PHINode &PN : BB->phis()) {
     670         559 :     for (const User *U : PN.users()) {
     671             :       const Instruction *UI = cast<Instruction>(U);
     672         589 :       if (UI->getParent() != DestBB || !isa<PHINode>(UI))
     673             :         return false;
     674             :       // If User is inside DestBB block and it is a PHINode then check
     675             :       // incoming value. If incoming value is not from BB then this is
     676             :       // a complex condition (e.g. preheaders) we want to avoid here.
     677             :       if (UI->getParent() == DestBB) {
     678             :         if (const PHINode *UPN = dyn_cast<PHINode>(UI))
     679        1409 :           for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
     680             :             Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
     681         618 :             if (Insn && Insn->getParent() == BB &&
     682             :                 Insn->getParent() != UPN->getIncomingBlock(I))
     683             :               return false;
     684             :           }
     685             :       }
     686             :     }
     687             :   }
     688             : 
     689             :   // If BB and DestBB contain any common predecessors, then the phi nodes in BB
     690             :   // and DestBB may have conflicting incoming values for the block.  If so, we
     691             :   // can't merge the block.
     692             :   const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
     693             :   if (!DestBBPN) return true;  // no conflict.
     694             : 
     695             :   // Collect the preds of BB.
     696             :   SmallPtrSet<const BasicBlock*, 16> BBPreds;
     697             :   if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
     698             :     // It is faster to get preds from a PHI than with pred_iterator.
     699        1121 :     for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
     700         462 :       BBPreds.insert(BBPN->getIncomingBlock(i));
     701             :   } else {
     702        2220 :     BBPreds.insert(pred_begin(BB), pred_end(BB));
     703             :   }
     704             : 
     705             :   // Walk the preds of DestBB.
     706       12999 :   for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
     707             :     BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
     708        5653 :     if (BBPreds.count(Pred)) {   // Common predecessor?
     709         418 :       for (const PHINode &PN : DestBB->phis()) {
     710         394 :         const Value *V1 = PN.getIncomingValueForBlock(Pred);
     711         394 :         const Value *V2 = PN.getIncomingValueForBlock(BB);
     712             : 
     713             :         // If V2 is a phi node in BB, look up what the mapped value will be.
     714             :         if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
     715          41 :           if (V2PN->getParent() == BB)
     716          26 :             V2 = V2PN->getIncomingValueForBlock(Pred);
     717             : 
     718             :         // If there is a conflict, bail out.
     719         394 :         if (V1 != V2) return false;
     720             :       }
     721             :     }
     722             :   }
     723             : 
     724             :   return true;
     725             : }
     726             : 
     727             : /// Eliminate a basic block that has only phi's and an unconditional branch in
     728             : /// it.
     729        3547 : void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
     730             :   BranchInst *BI = cast<BranchInst>(BB->getTerminator());
     731             :   BasicBlock *DestBB = BI->getSuccessor(0);
     732             : 
     733             :   DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
     734             : 
     735             :   // If the destination block has a single pred, then this is a trivial edge,
     736             :   // just collapse it.
     737        3547 :   if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
     738         112 :     if (SinglePred != DestBB) {
     739             :       // Remember if SinglePred was the entry block of the function.  If so, we
     740             :       // will need to move BB back to the entry position.
     741         112 :       bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
     742         112 :       MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
     743             : 
     744         112 :       if (isEntry && BB != &BB->getParent()->getEntryBlock())
     745           0 :         BB->moveBefore(&BB->getParent()->getEntryBlock());
     746             : 
     747             :       DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
     748             :       return;
     749             :     }
     750             :   }
     751             : 
     752             :   // Otherwise, we have multiple predecessors of BB.  Update the PHIs in DestBB
     753             :   // to handle the new incoming edges it is about to have.
     754        3435 :   for (PHINode &PN : DestBB->phis()) {
     755             :     // Remove the incoming value for BB, and remember it.
     756             :     Value *InVal = PN.removeIncomingValue(BB, false);
     757             : 
     758             :     // Two options: either the InVal is a phi node defined in BB or it is some
     759             :     // value that dominates BB.
     760             :     PHINode *InValPhi = dyn_cast<PHINode>(InVal);
     761         450 :     if (InValPhi && InValPhi->getParent() == BB) {
     762             :       // Add all of the input values of the input PHI as inputs of this phi.
     763         953 :       for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
     764         396 :         PN.addIncoming(InValPhi->getIncomingValue(i),
     765             :                        InValPhi->getIncomingBlock(i));
     766             :     } else {
     767             :       // Otherwise, add one instance of the dominating value for each edge that
     768             :       // we will be adding.
     769             :       if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
     770         100 :         for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
     771          41 :           PN.addIncoming(InVal, BBPN->getIncomingBlock(i));
     772             :       } else {
     773        3486 :         for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
     774        1190 :           PN.addIncoming(InVal, *PI);
     775             :       }
     776             :     }
     777             :   }
     778             : 
     779             :   // The PHIs are now updated, change everything that refers to BB to use
     780             :   // DestBB and remove BB.
     781        3435 :   BB->replaceAllUsesWith(DestBB);
     782        3435 :   BB->eraseFromParent();
     783             :   ++NumBlocksElim;
     784             : 
     785             :   DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
     786             : }
     787             : 
     788             : // Computes a map of base pointer relocation instructions to corresponding
     789             : // derived pointer relocation instructions given a vector of all relocate calls
     790          25 : static void computeBaseDerivedRelocateMap(
     791             :     const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
     792             :     DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
     793             :         &RelocateInstMap) {
     794             :   // Collect information in two maps: one primarily for locating the base object
     795             :   // while filling the second map; the second map is the final structure holding
     796             :   // a mapping between Base and corresponding Derived relocate calls
     797             :   DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
     798         149 :   for (auto *ThisRelocate : AllRelocateCalls) {
     799             :     auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
     800             :                             ThisRelocate->getDerivedPtrIndex());
     801         124 :     RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
     802             :   }
     803         174 :   for (auto &Item : RelocateIdxMap) {
     804          62 :     std::pair<unsigned, unsigned> Key = Item.first;
     805          62 :     if (Key.first == Key.second)
     806             :       // Base relocation: nothing to insert
     807          70 :       continue;
     808             : 
     809          28 :     GCRelocateInst *I = Item.second;
     810          28 :     auto BaseKey = std::make_pair(Key.first, Key.first);
     811             : 
     812             :     // We're iterating over RelocateIdxMap so we cannot modify it.
     813          28 :     auto MaybeBase = RelocateIdxMap.find(BaseKey);
     814          28 :     if (MaybeBase == RelocateIdxMap.end())
     815             :       // TODO: We might want to insert a new base object relocate and gep off
     816             :       // that, if there are enough derived object relocates.
     817           2 :       continue;
     818             : 
     819          52 :     RelocateInstMap[MaybeBase->second].push_back(I);
     820             :   }
     821          25 : }
     822             : 
     823             : // Accepts a GEP and extracts the operands into a vector provided they're all
     824             : // small integer constants
     825          11 : static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
     826             :                                           SmallVectorImpl<Value *> &OffsetV) {
     827          31 :   for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
     828             :     // Only accept small constant integer operands
     829             :     auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
     830          11 :     if (!Op || Op->getZExtValue() > 20)
     831             :       return false;
     832             :   }
     833             : 
     834          26 :   for (unsigned i = 1; i < GEP->getNumOperands(); i++)
     835          18 :     OffsetV.push_back(GEP->getOperand(i));
     836             :   return true;
     837             : }
     838             : 
     839             : // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
     840             : // replace, computes a replacement, and affects it.
     841             : static bool
     842          18 : simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
     843             :                           const SmallVectorImpl<GCRelocateInst *> &Targets) {
     844             :   bool MadeChange = false;
     845             :   // We must ensure the relocation of derived pointer is defined after
     846             :   // relocation of base pointer. If we find a relocation corresponding to base
     847             :   // defined earlier than relocation of base then we move relocation of base
     848             :   // right before found relocation. We consider only relocation in the same
     849             :   // basic block as relocation of base. Relocations from other basic block will
     850             :   // be skipped by optimization and we do not care about them.
     851          36 :   for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
     852          64 :        &*R != RelocatedBase; ++R)
     853          12 :     if (auto RI = dyn_cast<GCRelocateInst>(R))
     854          12 :       if (RI->getStatepoint() == RelocatedBase->getStatepoint())
     855           6 :         if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
     856           6 :           RelocatedBase->moveBefore(RI);
     857           6 :           break;
     858             :         }
     859             : 
     860          70 :   for (GCRelocateInst *ToReplace : Targets) {
     861             :     assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
     862             :            "Not relocating a derived object of the original base object");
     863          26 :     if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
     864             :       // A duplicate relocate call. TODO: coalesce duplicates.
     865          18 :       continue;
     866             :     }
     867             : 
     868          26 :     if (RelocatedBase->getParent() != ToReplace->getParent()) {
     869             :       // Base and derived relocates are in different basic blocks.
     870             :       // In this case transform is only valid when base dominates derived
     871             :       // relocate. However it would be too expensive to check dominance
     872             :       // for each such relocate, so we skip the whole transformation.
     873           1 :       continue;
     874             :     }
     875             : 
     876          25 :     Value *Base = ToReplace->getBasePtr();
     877          25 :     auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
     878          25 :     if (!Derived || Derived->getPointerOperand() != Base)
     879          14 :       continue;
     880             : 
     881             :     SmallVector<Value *, 2> OffsetV;
     882          14 :     if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
     883             :       continue;
     884             : 
     885             :     // Create a Builder and replace the target callsite with a gep
     886             :     assert(RelocatedBase->getNextNode() &&
     887             :            "Should always have one since it's not a terminator");
     888             : 
     889             :     // Insert after RelocatedBase
     890           8 :     IRBuilder<> Builder(RelocatedBase->getNextNode());
     891           8 :     Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
     892             : 
     893             :     // If gc_relocate does not match the actual type, cast it to the right type.
     894             :     // In theory, there must be a bitcast after gc_relocate if the type does not
     895             :     // match, and we should reuse it to get the derived pointer. But it could be
     896             :     // cases like this:
     897             :     // bb1:
     898             :     //  ...
     899             :     //  %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
     900             :     //  br label %merge
     901             :     //
     902             :     // bb2:
     903             :     //  ...
     904             :     //  %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
     905             :     //  br label %merge
     906             :     //
     907             :     // merge:
     908             :     //  %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
     909             :     //  %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
     910             :     //
     911             :     // In this case, we can not find the bitcast any more. So we insert a new bitcast
     912             :     // no matter there is already one or not. In this way, we can handle all cases, and
     913             :     // the extra bitcast should be optimized away in later passes.
     914             :     Value *ActualRelocatedBase = RelocatedBase;
     915           8 :     if (RelocatedBase->getType() != Base->getType()) {
     916             :       ActualRelocatedBase =
     917           0 :           Builder.CreateBitCast(RelocatedBase, Base->getType());
     918             :     }
     919          16 :     Value *Replacement = Builder.CreateGEP(
     920           8 :         Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
     921           8 :     Replacement->takeName(ToReplace);
     922             :     // If the newly generated derived pointer's type does not match the original derived
     923             :     // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
     924             :     Value *ActualReplacement = Replacement;
     925           8 :     if (Replacement->getType() != ToReplace->getType()) {
     926             :       ActualReplacement =
     927           0 :           Builder.CreateBitCast(Replacement, ToReplace->getType());
     928             :     }
     929           8 :     ToReplace->replaceAllUsesWith(ActualReplacement);
     930           8 :     ToReplace->eraseFromParent();
     931             : 
     932             :     MadeChange = true;
     933             :   }
     934          18 :   return MadeChange;
     935             : }
     936             : 
     937             : // Turns this:
     938             : //
     939             : // %base = ...
     940             : // %ptr = gep %base + 15
     941             : // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
     942             : // %base' = relocate(%tok, i32 4, i32 4)
     943             : // %ptr' = relocate(%tok, i32 4, i32 5)
     944             : // %val = load %ptr'
     945             : //
     946             : // into this:
     947             : //
     948             : // %base = ...
     949             : // %ptr = gep %base + 15
     950             : // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
     951             : // %base' = gc.relocate(%tok, i32 4, i32 4)
     952             : // %ptr' = gep %base' + 15
     953             : // %val = load %ptr'
     954          75 : bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
     955             :   bool MadeChange = false;
     956             :   SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
     957             : 
     958         177 :   for (auto *U : I.users())
     959         102 :     if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
     960             :       // Collect all the relocate calls associated with a statepoint
     961          78 :       AllRelocateCalls.push_back(Relocate);
     962             : 
     963             :   // We need atleast one base pointer relocation + one derived pointer
     964             :   // relocation to mangle
     965          75 :   if (AllRelocateCalls.size() < 2)
     966             :     return false;
     967             : 
     968             :   // RelocateInstMap is a mapping from the base relocate instruction to the
     969             :   // corresponding derived relocate instructions
     970             :   DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
     971          25 :   computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
     972          25 :   if (RelocateInstMap.empty())
     973             :     return false;
     974             : 
     975          54 :   for (auto &Item : RelocateInstMap)
     976             :     // Item.first is the RelocatedBase to offset against
     977             :     // Item.second is the vector of Targets to replace
     978          18 :     MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
     979             :   return MadeChange;
     980             : }
     981             : 
     982             : /// SinkCast - Sink the specified cast instruction into its user blocks
     983      217862 : static bool SinkCast(CastInst *CI) {
     984      217862 :   BasicBlock *DefBB = CI->getParent();
     985             : 
     986             :   /// InsertedCasts - Only insert a cast in each block once.
     987             :   DenseMap<BasicBlock*, CastInst*> InsertedCasts;
     988             : 
     989             :   bool MadeChange = false;
     990             :   for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
     991      489450 :        UI != E; ) {
     992             :     Use &TheUse = UI.getUse();
     993             :     Instruction *User = cast<Instruction>(*UI);
     994             : 
     995             :     // Figure out which BB this cast is used in.  For PHI's this is the
     996             :     // appropriate predecessor block.
     997      271588 :     BasicBlock *UserBB = User->getParent();
     998             :     if (PHINode *PN = dyn_cast<PHINode>(User)) {
     999        4538 :       UserBB = PN->getIncomingBlock(TheUse);
    1000             :     }
    1001             : 
    1002             :     // Preincrement use iterator so we don't invalidate it.
    1003             :     ++UI;
    1004             : 
    1005             :     // The first insertion point of a block containing an EH pad is after the
    1006             :     // pad.  If the pad is the user, we cannot sink the cast past the pad.
    1007           1 :     if (User->isEHPad())
    1008      225251 :       continue;
    1009             : 
    1010             :     // If the block selected to receive the cast is an EH pad that does not
    1011             :     // allow non-PHI instructions before the terminator, we can't sink the
    1012             :     // cast.
    1013      271587 :     if (UserBB->getTerminator()->isEHPad())
    1014           0 :       continue;
    1015             : 
    1016             :     // If this user is in the same block as the cast, don't change the cast.
    1017      496836 :     if (UserBB == DefBB) continue;
    1018             : 
    1019             :     // If we have already inserted a cast into this block, use it.
    1020             :     CastInst *&InsertedCast = InsertedCasts[UserBB];
    1021             : 
    1022       46338 :     if (!InsertedCast) {
    1023       44376 :       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
    1024             :       assert(InsertPt != UserBB->end());
    1025       88752 :       InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
    1026             :                                       CI->getType(), "", &*InsertPt);
    1027             :     }
    1028             : 
    1029             :     // Replace a use of the cast with a use of the new cast.
    1030       46338 :     TheUse = InsertedCast;
    1031             :     MadeChange = true;
    1032             :     ++NumCastUses;
    1033             :   }
    1034             : 
    1035             :   // If we removed all uses, nuke the cast.
    1036      217862 :   if (CI->use_empty()) {
    1037        7084 :     salvageDebugInfo(*CI);
    1038        7084 :     CI->eraseFromParent();
    1039             :     MadeChange = true;
    1040             :   }
    1041             : 
    1042      217862 :   return MadeChange;
    1043             : }
    1044             : 
    1045             : /// If the specified cast instruction is a noop copy (e.g. it's casting from
    1046             : /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
    1047             : /// reduce the number of virtual registers that must be created and coalesced.
    1048             : ///
    1049             : /// Return true if any changes are made.
    1050      289306 : static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
    1051             :                                        const DataLayout &DL) {
    1052             :   // Sink only "cheap" (or nop) address-space casts.  This is a weaker condition
    1053             :   // than sinking only nop casts, but is helpful on some platforms.
    1054             :   if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) {
    1055         554 :     if (!TLI.isCheapAddrSpaceCast(ASC->getSrcAddressSpace(),
    1056         277 :                                   ASC->getDestAddressSpace()))
    1057             :       return false;
    1058             :   }
    1059             : 
    1060             :   // If this is a noop copy,
    1061      578312 :   EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
    1062      289156 :   EVT DstVT = TLI.getValueType(DL, CI->getType());
    1063             : 
    1064             :   // This is an fp<->int conversion?
    1065      289156 :   if (SrcVT.isInteger() != DstVT.isInteger())
    1066             :     return false;
    1067             : 
    1068             :   // If this is an extension, it will be a zero or sign extension, which
    1069             :   // isn't a noop.
    1070      272467 :   if (SrcVT.bitsLT(DstVT)) return false;
    1071             : 
    1072             :   // If these values will be promoted, find out what they will be promoted
    1073             :   // to.  This helps us consider truncates on PPC as noop copies when they
    1074             :   // are.
    1075      476448 :   if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
    1076             :       TargetLowering::TypePromoteInteger)
    1077        2118 :     SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
    1078      476448 :   if (TLI.getTypeAction(CI->getContext(), DstVT) ==
    1079             :       TargetLowering::TypePromoteInteger)
    1080        9632 :     DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
    1081             : 
    1082             :   // If, after promotion, these are the same types, this is a noop copy.
    1083      238279 :   if (SrcVT != DstVT)
    1084             :     return false;
    1085             : 
    1086      214507 :   return SinkCast(CI);
    1087             : }
    1088             : 
    1089             : /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
    1090             : /// possible.
    1091             : ///
    1092             : /// Return true if any changes were made.
    1093       88476 : static bool CombineUAddWithOverflow(CmpInst *CI) {
    1094             :   Value *A, *B;
    1095             :   Instruction *AddI;
    1096       88476 :   if (!match(CI,
    1097       88476 :              m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
    1098             :     return false;
    1099             : 
    1100          53 :   Type *Ty = AddI->getType();
    1101          53 :   if (!isa<IntegerType>(Ty))
    1102             :     return false;
    1103             : 
    1104             :   // We don't want to move around uses of condition values this late, so we we
    1105             :   // check if it is legal to create the call to the intrinsic in the basic
    1106             :   // block containing the icmp:
    1107             : 
    1108          60 :   if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
    1109             :     return false;
    1110             : 
    1111             : #ifndef NDEBUG
    1112             :   // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
    1113             :   // for now:
    1114             :   if (AddI->hasOneUse())
    1115             :     assert(*AddI->user_begin() == CI && "expected!");
    1116             : #endif
    1117             : 
    1118          47 :   Module *M = CI->getModule();
    1119          47 :   Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
    1120             : 
    1121          94 :   auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
    1122             : 
    1123             :   auto *UAddWithOverflow =
    1124          94 :       CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
    1125          94 :   auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
    1126             :   auto *Overflow =
    1127          94 :       ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
    1128             : 
    1129          47 :   CI->replaceAllUsesWith(Overflow);
    1130          47 :   AddI->replaceAllUsesWith(UAdd);
    1131          47 :   CI->eraseFromParent();
    1132          47 :   AddI->eraseFromParent();
    1133          47 :   return true;
    1134             : }
    1135             : 
    1136             : /// Sink the given CmpInst into user blocks to reduce the number of virtual
    1137             : /// registers that must be created and coalesced. This is a clear win except on
    1138             : /// targets with multiple condition code registers (PowerPC), where it might
    1139             : /// lose; some adjustment may be wanted there.
    1140             : ///
    1141             : /// Return true if any changes are made.
    1142       89406 : static bool SinkCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
    1143       89406 :   BasicBlock *DefBB = CI->getParent();
    1144             : 
    1145             :   // Avoid sinking soft-FP comparisons, since this can move them into a loop.
    1146       89495 :   if (TLI && TLI->useSoftFloat() && isa<FCmpInst>(CI))
    1147             :     return false;
    1148             : 
    1149             :   // Only insert a cmp in each block once.
    1150             :   DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
    1151             : 
    1152             :   bool MadeChange = false;
    1153             :   for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
    1154      217893 :        UI != E; ) {
    1155             :     Use &TheUse = UI.getUse();
    1156             :     Instruction *User = cast<Instruction>(*UI);
    1157             : 
    1158             :     // Preincrement use iterator so we don't invalidate it.
    1159             :     ++UI;
    1160             : 
    1161             :     // Don't bother for PHI nodes.
    1162      128532 :     if (isa<PHINode>(User))
    1163      128050 :       continue;
    1164             : 
    1165             :     // Figure out which BB this cmp is used in.
    1166      128027 :     BasicBlock *UserBB = User->getParent();
    1167             : 
    1168             :     // If this user is in the same block as the cmp, don't change the cmp.
    1169      128027 :     if (UserBB == DefBB) continue;
    1170             : 
    1171             :     // If we have already inserted a cmp into this block, use it.
    1172             :     CmpInst *&InsertedCmp = InsertedCmps[UserBB];
    1173             : 
    1174         987 :     if (!InsertedCmp) {
    1175         828 :       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
    1176             :       assert(InsertPt != UserBB->end());
    1177         828 :       InsertedCmp =
    1178        1656 :           CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
    1179             :                           CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
    1180             :       // Propagate the debug info.
    1181         828 :       InsertedCmp->setDebugLoc(CI->getDebugLoc());
    1182             :     }
    1183             : 
    1184             :     // Replace a use of the cmp with a use of the new cmp.
    1185         987 :     TheUse = InsertedCmp;
    1186             :     MadeChange = true;
    1187             :     ++NumCmpUses;
    1188             :   }
    1189             : 
    1190             :   // If we removed all uses, nuke the cmp.
    1191       89361 :   if (CI->use_empty()) {
    1192         598 :     CI->eraseFromParent();
    1193             :     MadeChange = true;
    1194             :   }
    1195             : 
    1196             :   return MadeChange;
    1197             : }
    1198             : 
    1199       89406 : static bool OptimizeCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
    1200       89406 :   if (SinkCmpExpression(CI, TLI))
    1201             :     return true;
    1202             : 
    1203       88476 :   if (CombineUAddWithOverflow(CI))
    1204             :     return true;
    1205             : 
    1206       88429 :   return false;
    1207             : }
    1208             : 
    1209             : /// Duplicate and sink the given 'and' instruction into user blocks where it is
    1210             : /// used in a compare to allow isel to generate better code for targets where
    1211             : /// this operation can be combined.
    1212             : ///
    1213             : /// Return true if any changes are made.
    1214       15124 : static bool sinkAndCmp0Expression(Instruction *AndI,
    1215             :                                   const TargetLowering &TLI,
    1216             :                                   SetOfInstrs &InsertedInsts) {
    1217             :   // Double-check that we're not trying to optimize an instruction that was
    1218             :   // already optimized by some other part of this pass.
    1219             :   assert(!InsertedInsts.count(AndI) &&
    1220             :          "Attempting to optimize already optimized and instruction");
    1221             :   (void) InsertedInsts;
    1222             : 
    1223             :   // Nothing to do for single use in same basic block.
    1224       28455 :   if (AndI->hasOneUse() &&
    1225       26782 :       AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent())
    1226             :     return false;
    1227             : 
    1228             :   // Try to avoid cases where sinking/duplicating is likely to increase register
    1229             :   // pressure.
    1230        4019 :   if (!isa<ConstantInt>(AndI->getOperand(0)) &&
    1231         691 :       !isa<ConstantInt>(AndI->getOperand(1)) &&
    1232        2408 :       AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse())
    1233             :     return false;
    1234             : 
    1235        1810 :   for (auto *U : AndI->users()) {
    1236             :     Instruction *User = cast<Instruction>(U);
    1237             : 
    1238             :     // Only sink for and mask feeding icmp with 0.
    1239        1719 :     if (!isa<ICmpInst>(User))
    1240             :       return false;
    1241             : 
    1242         313 :     auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1));
    1243         108 :     if (!CmpC || !CmpC->isZero())
    1244             :       return false;
    1245             :   }
    1246             : 
    1247          91 :   if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI))
    1248             :     return false;
    1249             : 
    1250             :   DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n");
    1251             :   DEBUG(AndI->getParent()->dump());
    1252             : 
    1253             :   // Push the 'and' into the same block as the icmp 0.  There should only be
    1254             :   // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
    1255             :   // others, so we don't need to keep track of which BBs we insert into.
    1256             :   for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
    1257          84 :        UI != E; ) {
    1258             :     Use &TheUse = UI.getUse();
    1259             :     Instruction *User = cast<Instruction>(*UI);
    1260             : 
    1261             :     // Preincrement use iterator so we don't invalidate it.
    1262             :     ++UI;
    1263             : 
    1264             :     DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n");
    1265             : 
    1266             :     // Keep the 'and' in the same place if the use is already in the same block.
    1267             :     Instruction *InsertPt =
    1268          39 :         User->getParent() == AndI->getParent() ? AndI : User;
    1269             :     Instruction *InsertedAnd =
    1270          39 :         BinaryOperator::Create(Instruction::And, AndI->getOperand(0),
    1271          39 :                                AndI->getOperand(1), "", InsertPt);
    1272             :     // Propagate the debug info.
    1273          39 :     InsertedAnd->setDebugLoc(AndI->getDebugLoc());
    1274             : 
    1275             :     // Replace a use of the 'and' with a use of the new 'and'.
    1276             :     TheUse = InsertedAnd;
    1277             :     ++NumAndUses;
    1278             :     DEBUG(User->getParent()->dump());
    1279             :   }
    1280             : 
    1281             :   // We removed all uses, nuke the and.
    1282          45 :   AndI->eraseFromParent();
    1283             :   return true;
    1284             : }
    1285             : 
    1286             : /// Check if the candidates could be combined with a shift instruction, which
    1287             : /// includes:
    1288             : /// 1. Truncate instruction
    1289             : /// 2. And instruction and the imm is a mask of the low bits:
    1290             : /// imm & (imm+1) == 0
    1291        1194 : static bool isExtractBitsCandidateUse(Instruction *User) {
    1292        1194 :   if (!isa<TruncInst>(User)) {
    1293        1290 :     if (User->getOpcode() != Instruction::And ||
    1294         245 :         !isa<ConstantInt>(User->getOperand(1)))
    1295             :       return false;
    1296             : 
    1297             :     const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
    1298             : 
    1299         705 :     if ((Cimm & (Cimm + 1)).getBoolValue())
    1300             :       return false;
    1301             :   }
    1302             :   return true;
    1303             : }
    1304             : 
    1305             : /// Sink both shift and truncate instruction to the use of truncate's BB.
    1306             : static bool
    1307          46 : SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
    1308             :                      DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
    1309             :                      const TargetLowering &TLI, const DataLayout &DL) {
    1310          46 :   BasicBlock *UserBB = User->getParent();
    1311             :   DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
    1312             :   TruncInst *TruncI = dyn_cast<TruncInst>(User);
    1313             :   bool MadeChange = false;
    1314             : 
    1315             :   for (Value::user_iterator TruncUI = TruncI->user_begin(),
    1316             :                             TruncE = TruncI->user_end();
    1317          93 :        TruncUI != TruncE;) {
    1318             : 
    1319             :     Use &TruncTheUse = TruncUI.getUse();
    1320             :     Instruction *TruncUser = cast<Instruction>(*TruncUI);
    1321             :     // Preincrement use iterator so we don't invalidate it.
    1322             : 
    1323             :     ++TruncUI;
    1324             : 
    1325          94 :     int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
    1326          47 :     if (!ISDOpcode)
    1327          61 :       continue;
    1328             : 
    1329             :     // If the use is actually a legal node, there will not be an
    1330             :     // implicit truncate.
    1331             :     // FIXME: always querying the result type is just an
    1332             :     // approximation; some nodes' legality is determined by the
    1333             :     // operand or other means. There's no good way to find out though.
    1334          31 :     if (TLI.isOperationLegalOrCustom(
    1335             :             ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
    1336           5 :       continue;
    1337             : 
    1338             :     // Don't bother for PHI nodes.
    1339          26 :     if (isa<PHINode>(TruncUser))
    1340           0 :       continue;
    1341             : 
    1342          26 :     BasicBlock *TruncUserBB = TruncUser->getParent();
    1343             : 
    1344          26 :     if (UserBB == TruncUserBB)
    1345          24 :       continue;
    1346             : 
    1347             :     BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
    1348             :     CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
    1349             : 
    1350           2 :     if (!InsertedShift && !InsertedTrunc) {
    1351           2 :       BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
    1352             :       assert(InsertPt != TruncUserBB->end());
    1353             :       // Sink the shift
    1354           2 :       if (ShiftI->getOpcode() == Instruction::AShr)
    1355           0 :         InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
    1356             :                                                    "", &*InsertPt);
    1357             :       else
    1358           2 :         InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
    1359             :                                                    "", &*InsertPt);
    1360             : 
    1361             :       // Sink the trunc
    1362           2 :       BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
    1363             :       TruncInsertPt++;
    1364             :       assert(TruncInsertPt != TruncUserBB->end());
    1365             : 
    1366           4 :       InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
    1367             :                                        TruncI->getType(), "", &*TruncInsertPt);
    1368             : 
    1369             :       MadeChange = true;
    1370             : 
    1371             :       TruncTheUse = InsertedTrunc;
    1372             :     }
    1373             :   }
    1374          46 :   return MadeChange;
    1375             : }
    1376             : 
    1377             : /// Sink the shift *right* instruction into user blocks if the uses could
    1378             : /// potentially be combined with this shift instruction and generate BitExtract
    1379             : /// instruction. It will only be applied if the architecture supports BitExtract
    1380             : /// instruction. Here is an example:
    1381             : /// BB1:
    1382             : ///   %x.extract.shift = lshr i64 %arg1, 32
    1383             : /// BB2:
    1384             : ///   %x.extract.trunc = trunc i64 %x.extract.shift to i16
    1385             : /// ==>
    1386             : ///
    1387             : /// BB2:
    1388             : ///   %x.extract.shift.1 = lshr i64 %arg1, 32
    1389             : ///   %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
    1390             : ///
    1391             : /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
    1392             : /// instruction.
    1393             : /// Return true if any changes are made.
    1394        1122 : static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
    1395             :                                 const TargetLowering &TLI,
    1396             :                                 const DataLayout &DL) {
    1397        1122 :   BasicBlock *DefBB = ShiftI->getParent();
    1398             : 
    1399             :   /// Only insert instructions in each block once.
    1400             :   DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
    1401             : 
    1402        1122 :   bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
    1403             : 
    1404             :   bool MadeChange = false;
    1405             :   for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
    1406        2321 :        UI != E;) {
    1407             :     Use &TheUse = UI.getUse();
    1408             :     Instruction *User = cast<Instruction>(*UI);
    1409             :     // Preincrement use iterator so we don't invalidate it.
    1410             :     ++UI;
    1411             : 
    1412             :     // Don't bother for PHI nodes.
    1413        1199 :     if (isa<PHINode>(User))
    1414        1151 :       continue;
    1415             : 
    1416        1194 :     if (!isExtractBitsCandidateUse(User))
    1417         821 :       continue;
    1418             : 
    1419         373 :     BasicBlock *UserBB = User->getParent();
    1420             : 
    1421         693 :     if (UserBB == DefBB) {
    1422             :       // If the shift and truncate instruction are in the same BB. The use of
    1423             :       // the truncate(TruncUse) may still introduce another truncate if not
    1424             :       // legal. In this case, we would like to sink both shift and truncate
    1425             :       // instruction to the BB of TruncUse.
    1426             :       // for example:
    1427             :       // BB1:
    1428             :       // i64 shift.result = lshr i64 opnd, imm
    1429             :       // trunc.result = trunc shift.result to i16
    1430             :       //
    1431             :       // BB2:
    1432             :       //   ----> We will have an implicit truncate here if the architecture does
    1433             :       //   not have i16 compare.
    1434             :       // cmp i16 trunc.result, opnd2
    1435             :       //
    1436         146 :       if (isa<TruncInst>(User) && shiftIsLegal
    1437             :           // If the type of the truncate is legal, no trucate will be
    1438             :           // introduced in other basic blocks.
    1439         320 :           &&
    1440         110 :           (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
    1441          46 :         MadeChange =
    1442             :             SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
    1443             : 
    1444         320 :       continue;
    1445             :     }
    1446             :     // If we have already inserted a shift into this block, use it.
    1447             :     BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
    1448             : 
    1449          53 :     if (!InsertedShift) {
    1450          53 :       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
    1451             :       assert(InsertPt != UserBB->end());
    1452             : 
    1453          53 :       if (ShiftI->getOpcode() == Instruction::AShr)
    1454           8 :         InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
    1455             :                                                    "", &*InsertPt);
    1456             :       else
    1457          45 :         InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
    1458             :                                                    "", &*InsertPt);
    1459             : 
    1460             :       MadeChange = true;
    1461             :     }
    1462             : 
    1463             :     // Replace a use of the shift with a use of the new shift.
    1464          53 :     TheUse = InsertedShift;
    1465             :   }
    1466             : 
    1467             :   // If we removed all uses, nuke the shift.
    1468        1122 :   if (ShiftI->use_empty())
    1469          25 :     ShiftI->eraseFromParent();
    1470             : 
    1471        1122 :   return MadeChange;
    1472             : }
    1473             : 
    1474             : /// If counting leading or trailing zeros is an expensive operation and a zero
    1475             : /// input is defined, add a check for zero to avoid calling the intrinsic.
    1476             : ///
    1477             : /// We want to transform:
    1478             : ///     %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
    1479             : ///
    1480             : /// into:
    1481             : ///   entry:
    1482             : ///     %cmpz = icmp eq i64 %A, 0
    1483             : ///     br i1 %cmpz, label %cond.end, label %cond.false
    1484             : ///   cond.false:
    1485             : ///     %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
    1486             : ///     br label %cond.end
    1487             : ///   cond.end:
    1488             : ///     %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
    1489             : ///
    1490             : /// If the transform is performed, return true and set ModifiedDT to true.
    1491        1675 : static bool despeculateCountZeros(IntrinsicInst *CountZeros,
    1492             :                                   const TargetLowering *TLI,
    1493             :                                   const DataLayout *DL,
    1494             :                                   bool &ModifiedDT) {
    1495        1675 :   if (!TLI || !DL)
    1496             :     return false;
    1497             : 
    1498             :   // If a zero input is undefined, it doesn't make sense to despeculate that.
    1499        3350 :   if (match(CountZeros->getOperand(1), m_One()))
    1500             :     return false;
    1501             : 
    1502             :   // If it's cheap to speculate, there's nothing to do.
    1503             :   auto IntrinsicID = CountZeros->getIntrinsicID();
    1504         750 :   if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
    1505         417 :       (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
    1506             :     return false;
    1507             : 
    1508             :   // Only handle legal scalar cases. Anything else requires too much work.
    1509         455 :   Type *Ty = CountZeros->getType();
    1510         455 :   unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
    1511         455 :   if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits())
    1512             :     return false;
    1513             : 
    1514             :   // The intrinsic will be sunk behind a compare against zero and branch.
    1515          29 :   BasicBlock *StartBlock = CountZeros->getParent();
    1516          29 :   BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
    1517             : 
    1518             :   // Create another block after the count zero intrinsic. A PHI will be added
    1519             :   // in this block to select the result of the intrinsic or the bit-width
    1520             :   // constant if the input to the intrinsic is zero.
    1521          29 :   BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
    1522          29 :   BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
    1523             : 
    1524             :   // Set up a builder to create a compare, conditional branch, and PHI.
    1525          29 :   IRBuilder<> Builder(CountZeros->getContext());
    1526          29 :   Builder.SetInsertPoint(StartBlock->getTerminator());
    1527          29 :   Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
    1528             : 
    1529             :   // Replace the unconditional branch that was created by the first split with
    1530             :   // a compare against zero and a conditional branch.
    1531          29 :   Value *Zero = Constant::getNullValue(Ty);
    1532          29 :   Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
    1533          29 :   Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
    1534          29 :   StartBlock->getTerminator()->eraseFromParent();
    1535             : 
    1536             :   // Create a PHI in the end block to select either the output of the intrinsic
    1537             :   // or the bit width of the operand.
    1538          29 :   Builder.SetInsertPoint(&EndBlock->front());
    1539          29 :   PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
    1540          29 :   CountZeros->replaceAllUsesWith(PN);
    1541          87 :   Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
    1542          29 :   PN->addIncoming(BitWidth, StartBlock);
    1543          29 :   PN->addIncoming(CountZeros, CallBlock);
    1544             : 
    1545             :   // We are explicitly handling the zero case, so we can set the intrinsic's
    1546             :   // undefined zero argument to 'true'. This will also prevent reprocessing the
    1547             :   // intrinsic; we only despeculate when a zero input is defined.
    1548          29 :   CountZeros->setArgOperand(1, Builder.getTrue());
    1549          29 :   ModifiedDT = true;
    1550             :   return true;
    1551             : }
    1552             : 
    1553      506507 : bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) {
    1554      506507 :   BasicBlock *BB = CI->getParent();
    1555             : 
    1556             :   // Lower inline assembly if we can.
    1557             :   // If we found an inline asm expession, and if the target knows how to
    1558             :   // lower it to normal LLVM code, do so now.
    1559     1013009 :   if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
    1560        9307 :     if (TLI->ExpandInlineAsm(CI)) {
    1561             :       // Avoid invalidating the iterator.
    1562          49 :       CurInstIterator = BB->begin();
    1563             :       // Avoid processing instructions out of order, which could cause
    1564             :       // reuse before a value is defined.
    1565             :       SunkAddrs.clear();
    1566          49 :       return true;
    1567             :     }
    1568             :     // Sink address computing for memory operands into the block.
    1569        9258 :     if (optimizeInlineAsmInst(CI))
    1570             :       return true;
    1571             :   }
    1572             : 
    1573             :   // Align the pointer arguments to this call if the target thinks it's a good
    1574             :   // idea
    1575             :   unsigned MinSize, PrefAlign;
    1576      506443 :   if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
    1577        3320 :     for (auto &Arg : CI->arg_operands()) {
    1578             :       // We want to align both objects whose address is used directly and
    1579             :       // objects whose address is used in casts and GEPs, though it only makes
    1580             :       // sense for GEPs if the offset is a multiple of the desired alignment and
    1581             :       // if size - offset meets the size threshold.
    1582        5312 :       if (!Arg->getType()->isPointerTy())
    1583        3050 :         continue;
    1584        1154 :       APInt Offset(DL->getIndexSizeInBits(
    1585             :                        cast<PointerType>(Arg->getType())->getAddressSpace()),
    1586             :                    0);
    1587        1154 :       Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
    1588             :       uint64_t Offset2 = Offset.getLimitedValue();
    1589        1154 :       if ((Offset2 & (PrefAlign-1)) != 0)
    1590             :         continue;
    1591             :       AllocaInst *AI;
    1592         287 :       if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
    1593         138 :           DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
    1594          60 :         AI->setAlignment(PrefAlign);
    1595             :       // Global variables can only be aligned if they are defined in this
    1596             :       // object (i.e. they are uniquely initialized in this object), and
    1597             :       // over-aligning global variables that have an explicit section is
    1598             :       // forbidden.
    1599             :       GlobalVariable *GV;
    1600         296 :       if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
    1601         149 :           GV->getPointerAlignment(*DL) < PrefAlign &&
    1602          67 :           DL->getTypeAllocSize(GV->getValueType()) >=
    1603          67 :               MinSize + Offset2)
    1604          39 :         GV->setAlignment(PrefAlign);
    1605             :     }
    1606             :     // If this is a memcpy (or similar) then we may be able to improve the
    1607             :     // alignment
    1608             :     if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
    1609         664 :       unsigned DestAlign = getKnownAlignment(MI->getDest(), *DL);
    1610         664 :       if (DestAlign > MI->getDestAlignment())
    1611         454 :         MI->setDestAlignment(DestAlign);
    1612             :       if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
    1613         490 :         unsigned SrcAlign = getKnownAlignment(MTI->getSource(), *DL);
    1614         490 :         if (SrcAlign > MTI->getSourceAlignment())
    1615         329 :           MTI->setSourceAlignment(SrcAlign);
    1616             :       }
    1617             :     }
    1618             :   }
    1619             : 
    1620             :   // If we have a cold call site, try to sink addressing computation into the
    1621             :   // cold block.  This interacts with our handling for loads and stores to
    1622             :   // ensure that we can fold all uses of a potential addressing computation
    1623             :   // into their uses.  TODO: generalize this to work over profiling data
    1624     1006452 :   if (!OptSize && CI->hasFnAttr(Attribute::Cold))
    1625         132 :     for (auto &Arg : CI->arg_operands()) {
    1626         240 :       if (!Arg->getType()->isPointerTy())
    1627             :         continue;
    1628             :       unsigned AS = Arg->getType()->getPointerAddressSpace();
    1629         110 :       return optimizeMemoryInst(CI, Arg, Arg->getType(), AS);
    1630             :     }
    1631             : 
    1632             :   IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
    1633             :   if (II) {
    1634      242363 :     switch (II->getIntrinsicID()) {
    1635             :     default: break;
    1636          29 :     case Intrinsic::objectsize: {
    1637             :       // Lower all uses of llvm.objectsize.*
    1638             :       ConstantInt *RetVal =
    1639          29 :           lowerObjectSizeCall(II, *DL, TLInfo, /*MustSucceed=*/true);
    1640             :       // Substituting this can cause recursive simplifications, which can
    1641             :       // invalidate our iterator.  Use a WeakTrackingVH to hold onto it in case
    1642             :       // this
    1643             :       // happens.
    1644             :       Value *CurValue = &*CurInstIterator;
    1645             :       WeakTrackingVH IterHandle(CurValue);
    1646             : 
    1647          29 :       replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
    1648             : 
    1649             :       // If the iterator instruction was recursively deleted, start over at the
    1650             :       // start of the block.
    1651          29 :       if (IterHandle != CurValue) {
    1652           4 :         CurInstIterator = BB->begin();
    1653             :         SunkAddrs.clear();
    1654             :       }
    1655             :       return true;
    1656             :     }
    1657             :     case Intrinsic::aarch64_stlxr:
    1658             :     case Intrinsic::aarch64_stxr: {
    1659             :       ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
    1660         174 :       if (!ExtVal || !ExtVal->hasOneUse() ||
    1661          87 :           ExtVal->getParent() == CI->getParent())
    1662             :         return false;
    1663             :       // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
    1664           1 :       ExtVal->moveBefore(CI);
    1665             :       // Mark this instruction as "inserted by CGP", so that other
    1666             :       // optimizations don't touch it.
    1667           1 :       InsertedInsts.insert(ExtVal);
    1668           1 :       return true;
    1669             :     }
    1670           1 :     case Intrinsic::invariant_group_barrier:
    1671           2 :       II->replaceAllUsesWith(II->getArgOperand(0));
    1672           1 :       II->eraseFromParent();
    1673           1 :       return true;
    1674             : 
    1675        1675 :     case Intrinsic::cttz:
    1676             :     case Intrinsic::ctlz:
    1677             :       // If counting zeros is expensive, try to avoid it.
    1678        1675 :       return despeculateCountZeros(II, TLI, DL, ModifiedDT);
    1679             :     }
    1680             : 
    1681      240544 :     if (TLI) {
    1682             :       SmallVector<Value*, 2> PtrOps;
    1683             :       Type *AccessTy;
    1684      240543 :       if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
    1685         522 :         while (!PtrOps.empty()) {
    1686             :           Value *PtrVal = PtrOps.pop_back_val();
    1687         269 :           unsigned AS = PtrVal->getType()->getPointerAddressSpace();
    1688         269 :           if (optimizeMemoryInst(II, PtrVal, AccessTy, AS))
    1689             :             return true;
    1690             :         }
    1691             :     }
    1692             :   }
    1693             : 
    1694             :   // From here on out we're working with named functions.
    1695             :   if (!CI->getCalledFunction()) return false;
    1696             : 
    1697             :   // Lower all default uses of _chk calls.  This is very similar
    1698             :   // to what InstCombineCalls does, but here we are only lowering calls
    1699             :   // to fortified library functions (e.g. __memcpy_chk) that have the default
    1700             :   // "don't know" as the objectsize.  Anything else should be left alone.
    1701      492474 :   FortifiedLibCallSimplifier Simplifier(TLInfo, true);
    1702      492474 :   if (Value *V = Simplifier.optimizeCall(CI)) {
    1703          38 :     CI->replaceAllUsesWith(V);
    1704          38 :     CI->eraseFromParent();
    1705          38 :     return true;
    1706             :   }
    1707             : 
    1708             :   return false;
    1709             : }
    1710             : 
    1711             : /// Look for opportunities to duplicate return instructions to the predecessor
    1712             : /// to enable tail call optimizations. The case it is currently looking for is:
    1713             : /// @code
    1714             : /// bb0:
    1715             : ///   %tmp0 = tail call i32 @f0()
    1716             : ///   br label %return
    1717             : /// bb1:
    1718             : ///   %tmp1 = tail call i32 @f1()
    1719             : ///   br label %return
    1720             : /// bb2:
    1721             : ///   %tmp2 = tail call i32 @f2()
    1722             : ///   br label %return
    1723             : /// return:
    1724             : ///   %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
    1725             : ///   ret i32 %retval
    1726             : /// @endcode
    1727             : ///
    1728             : /// =>
    1729             : ///
    1730             : /// @code
    1731             : /// bb0:
    1732             : ///   %tmp0 = tail call i32 @f0()
    1733             : ///   ret i32 %tmp0
    1734             : /// bb1:
    1735             : ///   %tmp1 = tail call i32 @f1()
    1736             : ///   ret i32 %tmp1
    1737             : /// bb2:
    1738             : ///   %tmp2 = tail call i32 @f2()
    1739             : ///   ret i32 %tmp2
    1740             : /// @endcode
    1741      436774 : bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
    1742      436774 :   if (!TLI)
    1743             :     return false;
    1744             : 
    1745             :   ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator());
    1746             :   if (!RetI)
    1747             :     return false;
    1748             : 
    1749             :   PHINode *PN = nullptr;
    1750             :   BitCastInst *BCI = nullptr;
    1751             :   Value *V = RetI->getReturnValue();
    1752             :   if (V) {
    1753             :     BCI = dyn_cast<BitCastInst>(V);
    1754             :     if (BCI)
    1755             :       V = BCI->getOperand(0);
    1756             : 
    1757             :     PN = dyn_cast<PHINode>(V);
    1758             :     if (!PN)
    1759             :       return false;
    1760             :   }
    1761             : 
    1762        2126 :   if (PN && PN->getParent() != BB)
    1763             :     return false;
    1764             : 
    1765             :   // Make sure there are no instructions between the PHI and return, or that the
    1766             :   // return is the first instruction in the block.
    1767       59271 :   if (PN) {
    1768        1972 :     BasicBlock::iterator BI = BB->begin();
    1769        1973 :     do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
    1770        1972 :     if (&*BI == BCI)
    1771             :       // Also skip over the bitcast.
    1772             :       ++BI;
    1773        1972 :     if (&*BI != RetI)
    1774        1009 :       return false;
    1775             :   } else {
    1776       57299 :     BasicBlock::iterator BI = BB->begin();
    1777       57497 :     while (isa<DbgInfoIntrinsic>(BI)) ++BI;
    1778       57299 :     if (&*BI != RetI)
    1779       51414 :       return false;
    1780             :   }
    1781             : 
    1782             :   /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
    1783             :   /// call.
    1784        6848 :   const Function *F = BB->getParent();
    1785             :   SmallVector<CallInst*, 4> TailCalls;
    1786        6848 :   if (PN) {
    1787        6155 :     for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
    1788        2596 :       CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
    1789             :       // Make sure the phi value is indeed produced by the tail call.
    1790         572 :       if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
    1791        2852 :           TLI->mayBeEmittedAsTailCall(CI) &&
    1792          82 :           attributesPermitTailCall(F, CI, RetI, *TLI))
    1793          82 :         TailCalls.push_back(CI);
    1794             :     }
    1795             :   } else {
    1796             :     SmallPtrSet<BasicBlock*, 4> VisitedBBs;
    1797       18636 :     for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
    1798        6866 :       if (!VisitedBBs.insert(*PI).second)
    1799        1234 :         continue;
    1800             : 
    1801             :       BasicBlock::InstListType &InstList = (*PI)->getInstList();
    1802             :       BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
    1803             :       BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
    1804        6671 :       do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
    1805        6592 :       if (RI == RE)
    1806             :         continue;
    1807             : 
    1808        5632 :       CallInst *CI = dyn_cast<CallInst>(&*RI);
    1809        5977 :       if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) &&
    1810         345 :           attributesPermitTailCall(F, CI, RetI, *TLI))
    1811         345 :         TailCalls.push_back(CI);
    1812             :     }
    1813             :   }
    1814             : 
    1815             :   bool Changed = false;
    1816        7275 :   for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
    1817         854 :     CallInst *CI = TailCalls[i];
    1818             :     CallSite CS(CI);
    1819             : 
    1820             :     // Conservatively require the attributes of the call to match those of the
    1821             :     // return. Ignore noalias because it doesn't affect the call sequence.
    1822         427 :     AttributeList CalleeAttrs = CS.getAttributes();
    1823        1281 :     if (AttrBuilder(CalleeAttrs, AttributeList::ReturnIndex)
    1824         427 :             .removeAttribute(Attribute::NoAlias) !=
    1825         854 :         AttrBuilder(CalleeAttrs, AttributeList::ReturnIndex)
    1826         427 :             .removeAttribute(Attribute::NoAlias))
    1827          18 :       continue;
    1828             : 
    1829             :     // Make sure the call instruction is followed by an unconditional branch to
    1830             :     // the return block.
    1831         427 :     BasicBlock *CallBB = CI->getParent();
    1832             :     BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
    1833         851 :     if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
    1834             :       continue;
    1835             : 
    1836             :     // Duplicate the return into CallBB.
    1837         409 :     (void)FoldReturnIntoUncondBranch(RetI, BB, CallBB);
    1838         409 :     ModifiedDT = Changed = true;
    1839             :     ++NumRetsDup;
    1840             :   }
    1841             : 
    1842             :   // If we eliminated all predecessors of the block, delete the block now.
    1843        7339 :   if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
    1844          26 :     BB->eraseFromParent();
    1845             : 
    1846             :   return Changed;
    1847             : }
    1848             : 
    1849             : //===----------------------------------------------------------------------===//
    1850             : // Memory Optimization
    1851             : //===----------------------------------------------------------------------===//
    1852             : 
    1853             : namespace {
    1854             : 
    1855             : /// This is an extended version of TargetLowering::AddrMode
    1856             : /// which holds actual Value*'s for register values.
    1857             : struct ExtAddrMode : public TargetLowering::AddrMode {
    1858             :   Value *BaseReg = nullptr;
    1859             :   Value *ScaledReg = nullptr;
    1860             :   Value *OriginalValue = nullptr;
    1861             : 
    1862             :   enum FieldName {
    1863             :     NoField        = 0x00,
    1864             :     BaseRegField   = 0x01,
    1865             :     BaseGVField    = 0x02,
    1866             :     BaseOffsField  = 0x04,
    1867             :     ScaledRegField = 0x08,
    1868             :     ScaleField     = 0x10,
    1869             :     MultipleFields = 0xff
    1870             :   };
    1871             : 
    1872             :   ExtAddrMode() = default;
    1873             : 
    1874             :   void print(raw_ostream &OS) const;
    1875             :   void dump() const;
    1876             : 
    1877        7135 :   FieldName compare(const ExtAddrMode &other) {
    1878             :     // First check that the types are the same on each field, as differing types
    1879             :     // is something we can't cope with later on.
    1880       13665 :     if (BaseReg && other.BaseReg &&
    1881        6530 :         BaseReg->getType() != other.BaseReg->getType())
    1882             :       return MultipleFields;
    1883        6966 :     if (BaseGV && other.BaseGV &&
    1884             :         BaseGV->getType() != other.BaseGV->getType())
    1885             :       return MultipleFields;
    1886        6573 :     if (ScaledReg && other.ScaledReg &&
    1887           8 :         ScaledReg->getType() != other.ScaledReg->getType())
    1888             :       return MultipleFields;
    1889             : 
    1890             :     // Check each field to see if it differs.
    1891             :     unsigned Result = NoField;
    1892        6565 :     if (BaseReg != other.BaseReg)
    1893             :       Result |= BaseRegField;
    1894        6565 :     if (BaseGV != other.BaseGV)
    1895         126 :       Result |= BaseGVField;
    1896        6565 :     if (BaseOffs != other.BaseOffs)
    1897        1473 :       Result |= BaseOffsField;
    1898        6565 :     if (ScaledReg != other.ScaledReg)
    1899         120 :       Result |= ScaledRegField;
    1900             :     // Don't count 0 as being a different scale, because that actually means
    1901             :     // unscaled (which will already be counted by having no ScaledReg).
    1902        6565 :     if (Scale && other.Scale && Scale != other.Scale)
    1903           1 :       Result |= ScaleField;
    1904             : 
    1905        6565 :     if (countPopulation(Result) > 1)
    1906             :       return MultipleFields;
    1907             :     else
    1908        5594 :       return static_cast<FieldName>(Result);
    1909             :   }
    1910             : 
    1911             :   // An AddrMode is trivial if it involves no calculation i.e. it is just a base
    1912             :   // with no offset.
    1913             :   bool isTrivial() {
    1914             :     // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
    1915             :     // trivial if at most one of these terms is nonzero, except that BaseGV and
    1916             :     // BaseReg both being zero actually means a null pointer value, which we
    1917             :     // consider to be 'non-zero' here.
    1918     1042865 :     return !BaseOffs && !Scale && !(BaseGV && BaseReg);
    1919             :   }
    1920             : 
    1921        1849 :   Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
    1922        1849 :     switch (Field) {
    1923             :     default:
    1924             :       return nullptr;
    1925         584 :     case BaseRegField:
    1926         584 :       return BaseReg;
    1927         152 :     case BaseGVField:
    1928         152 :       return BaseGV;
    1929          24 :     case ScaledRegField:
    1930          24 :       return ScaledReg;
    1931        1089 :     case BaseOffsField:
    1932        1089 :       return ConstantInt::get(IntPtrTy, BaseOffs);
    1933             :     }
    1934             :   }
    1935             : 
    1936          58 :   void SetCombinedField(FieldName Field, Value *V,
    1937             :                         const SmallVectorImpl<ExtAddrMode> &AddrModes) {
    1938          58 :     switch (Field) {
    1939           0 :     default:
    1940           0 :       llvm_unreachable("Unhandled fields are expected to be rejected earlier");
    1941             :       break;
    1942          39 :     case ExtAddrMode::BaseRegField:
    1943          39 :       BaseReg = V;
    1944          39 :       break;
    1945           0 :     case ExtAddrMode::BaseGVField:
    1946             :       // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
    1947             :       // in the BaseReg field.
    1948             :       assert(BaseReg == nullptr);
    1949           0 :       BaseReg = V;
    1950           0 :       BaseGV = nullptr;
    1951           0 :       break;
    1952           5 :     case ExtAddrMode::ScaledRegField:
    1953           5 :       ScaledReg = V;
    1954             :       // If we have a mix of scaled and unscaled addrmodes then we want scale
    1955             :       // to be the scale and not zero.
    1956           5 :       if (!Scale)
    1957           6 :         for (const ExtAddrMode &AM : AddrModes)
    1958           4 :           if (AM.Scale) {
    1959           2 :             Scale = AM.Scale;
    1960           2 :             break;
    1961             :           }
    1962             :       break;
    1963          14 :     case ExtAddrMode::BaseOffsField:
    1964             :       // The offset is no longer a constant, so it goes in ScaledReg with a
    1965             :       // scale of 1.
    1966             :       assert(ScaledReg == nullptr);
    1967          14 :       ScaledReg = V;
    1968          14 :       Scale = 1;
    1969          14 :       BaseOffs = 0;
    1970          14 :       break;
    1971             :     }
    1972          58 :   }
    1973             : };
    1974             : 
    1975             : } // end anonymous namespace
    1976             : 
    1977             : #ifndef NDEBUG
    1978             : static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
    1979             :   AM.print(OS);
    1980             :   return OS;
    1981             : }
    1982             : #endif
    1983             : 
    1984             : #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
    1985             : void ExtAddrMode::print(raw_ostream &OS) const {
    1986             :   bool NeedPlus = false;
    1987             :   OS << "[";
    1988             :   if (BaseGV) {
    1989             :     OS << (NeedPlus ? " + " : "")
    1990             :        << "GV:";
    1991             :     BaseGV->printAsOperand(OS, /*PrintType=*/false);
    1992             :     NeedPlus = true;
    1993             :   }
    1994             : 
    1995             :   if (BaseOffs) {
    1996             :     OS << (NeedPlus ? " + " : "")
    1997             :        << BaseOffs;
    1998             :     NeedPlus = true;
    1999             :   }
    2000             : 
    2001             :   if (BaseReg) {
    2002             :     OS << (NeedPlus ? " + " : "")
    2003             :        << "Base:";
    2004             :     BaseReg->printAsOperand(OS, /*PrintType=*/false);
    2005             :     NeedPlus = true;
    2006             :   }
    2007             :   if (Scale) {
    2008             :     OS << (NeedPlus ? " + " : "")
    2009             :        << Scale << "*";
    2010             :     ScaledReg->printAsOperand(OS, /*PrintType=*/false);
    2011             :   }
    2012             : 
    2013             :   OS << ']';
    2014             : }
    2015             : 
    2016             : LLVM_DUMP_METHOD void ExtAddrMode::dump() const {
    2017             :   print(dbgs());
    2018             :   dbgs() << '\n';
    2019             : }
    2020             : #endif
    2021             : 
    2022             : namespace {
    2023             : 
    2024             : /// \brief This class provides transaction based operation on the IR.
    2025             : /// Every change made through this class is recorded in the internal state and
    2026             : /// can be undone (rollback) until commit is called.
    2027     1070283 : class TypePromotionTransaction {
    2028             :   /// \brief This represents the common interface of the individual transaction.
    2029             :   /// Each class implements the logic for doing one specific modification on
    2030             :   /// the IR via the TypePromotionTransaction.
    2031             :   class TypePromotionAction {
    2032             :   protected:
    2033             :     /// The Instruction modified.
    2034             :     Instruction *Inst;
    2035             : 
    2036             :   public:
    2037             :     /// \brief Constructor of the action.
    2038             :     /// The constructor performs the related action on the IR.
    2039        4370 :     TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
    2040             : 
    2041             :     virtual ~TypePromotionAction() = default;
    2042             : 
    2043             :     /// \brief Undo the modification done by this action.
    2044             :     /// When this method is called, the IR must be in the same state as it was
    2045             :     /// before this action was applied.
    2046             :     /// \pre Undoing the action works if and only if the IR is in the exact same
    2047             :     /// state as it was directly after this action was applied.
    2048             :     virtual void undo() = 0;
    2049             : 
    2050             :     /// \brief Advocate every change made by this action.
    2051             :     /// When the results on the IR of the action are to be kept, it is important
    2052             :     /// to call this function, otherwise hidden information may be kept forever.
    2053           0 :     virtual void commit() {
    2054             :       // Nothing to be done, this action is not doing anything.
    2055           0 :     }
    2056             :   };
    2057             : 
    2058             :   /// \brief Utility to remember the position of an instruction.
    2059             :   class InsertionHandler {
    2060             :     /// Position of an instruction.
    2061             :     /// Either an instruction:
    2062             :     /// - Is the first in a basic block: BB is used.
    2063             :     /// - Has a previous instructon: PrevInst is used.
    2064             :     union {
    2065             :       Instruction *PrevInst;
    2066             :       BasicBlock *BB;
    2067             :     } Point;
    2068             : 
    2069             :     /// Remember whether or not the instruction had a previous instruction.
    2070             :     bool HasPrevInstruction;
    2071             : 
    2072             :   public:
    2073             :     /// \brief Record the position of \p Inst.
    2074         847 :     InsertionHandler(Instruction *Inst) {
    2075         847 :       BasicBlock::iterator It = Inst->getIterator();
    2076        1694 :       HasPrevInstruction = (It != (Inst->getParent()->begin()));
    2077         847 :       if (HasPrevInstruction)
    2078         762 :         Point.PrevInst = &*--It;
    2079             :       else
    2080          85 :         Point.BB = Inst->getParent();
    2081             :     }
    2082             : 
    2083             :     /// \brief Insert \p Inst at the recorded position.
    2084         481 :     void insert(Instruction *Inst) {
    2085         481 :       if (HasPrevInstruction) {
    2086         427 :         if (Inst->getParent())
    2087         343 :           Inst->removeFromParent();
    2088         427 :         Inst->insertAfter(Point.PrevInst);
    2089             :       } else {
    2090          54 :         Instruction *Position = &*Point.BB->getFirstInsertionPt();
    2091          54 :         if (Inst->getParent())
    2092          46 :           Inst->moveBefore(Position);
    2093             :         else
    2094           8 :           Inst->insertBefore(Position);
    2095             :       }
    2096         481 :     }
    2097             :   };
    2098             : 
    2099             :   /// \brief Move an instruction before another.
    2100         666 :   class InstructionMoveBefore : public TypePromotionAction {
    2101             :     /// Original position of the instruction.
    2102             :     InsertionHandler Position;
    2103             : 
    2104             :   public:
    2105             :     /// \brief Move \p Inst before \p Before.
    2106         666 :     InstructionMoveBefore(Instruction *Inst, Instruction *Before)
    2107        1332 :         : TypePromotionAction(Inst), Position(Inst) {
    2108             :       DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
    2109         666 :       Inst->moveBefore(Before);
    2110         666 :     }
    2111             : 
    2112             :     /// \brief Move the instruction back to its original position.
    2113         389 :     void undo() override {
    2114             :       DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
    2115         389 :       Position.insert(Inst);
    2116         389 :     }
    2117             :   };
    2118             : 
    2119             :   /// \brief Set the operand of an instruction with a new value.
    2120        1834 :   class OperandSetter : public TypePromotionAction {
    2121             :     /// Original operand of the instruction.
    2122             :     Value *Origin;
    2123             : 
    2124             :     /// Index of the modified instruction.
    2125             :     unsigned Idx;
    2126             : 
    2127             :   public:
    2128             :     /// \brief Set \p Idx operand of \p Inst with \p NewVal.
    2129        1834 :     OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
    2130        3668 :         : TypePromotionAction(Inst), Idx(Idx) {
    2131             :       DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
    2132             :                    << "for:" << *Inst << "\n"
    2133             :                    << "with:" << *NewVal << "\n");
    2134        3668 :       Origin = Inst->getOperand(Idx);
    2135        1834 :       Inst->setOperand(Idx, NewVal);
    2136        1834 :     }
    2137             : 
    2138             :     /// \brief Restore the original value of the instruction.
    2139        1090 :     void undo() override {
    2140             :       DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
    2141             :                    << "for: " << *Inst << "\n"
    2142             :                    << "with: " << *Origin << "\n");
    2143        1090 :       Inst->setOperand(Idx, Origin);
    2144        1090 :     }
    2145             :   };
    2146             : 
    2147             :   /// \brief Hide the operands of an instruction.
    2148             :   /// Do as if this instruction was not using any of its operands.
    2149         181 :   class OperandsHider : public TypePromotionAction {
    2150             :     /// The list of original operands.
    2151             :     SmallVector<Value *, 4> OriginalValues;
    2152             : 
    2153             :   public:
    2154             :     /// \brief Remove \p Inst from the uses of the operands of \p Inst.
    2155         362 :     OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
    2156             :       DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
    2157             :       unsigned NumOpnds = Inst->getNumOperands();
    2158         181 :       OriginalValues.reserve(NumOpnds);
    2159         543 :       for (unsigned It = 0; It < NumOpnds; ++It) {
    2160             :         // Save the current operand.
    2161         362 :         Value *Val = Inst->getOperand(It);
    2162         181 :         OriginalValues.push_back(Val);
    2163             :         // Set a dummy one.
    2164             :         // We could use OperandSetter here, but that would imply an overhead
    2165             :         // that we are not willing to pay.
    2166         181 :         Inst->setOperand(It, UndefValue::get(Val->getType()));
    2167             :       }
    2168         181 :     }
    2169             : 
    2170             :     /// \brief Restore the original list of uses.
    2171          92 :     void undo() override {
    2172             :       DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
    2173         184 :       for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
    2174         184 :         Inst->setOperand(It, OriginalValues[It]);
    2175          92 :     }
    2176             :   };
    2177             : 
    2178             :   /// \brief Build a truncate instruction.
    2179          46 :   class TruncBuilder : public TypePromotionAction {
    2180             :     Value *Val;
    2181             : 
    2182             :   public:
    2183             :     /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
    2184             :     /// result.
    2185             :     /// trunc Opnd to Ty.
    2186          92 :     TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
    2187          46 :       IRBuilder<> Builder(Opnd);
    2188          46 :       Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
    2189             :       DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
    2190          46 :     }
    2191             : 
    2192             :     /// \brief Get the built value.
    2193             :     Value *getBuiltValue() { return Val; }
    2194             : 
    2195             :     /// \brief Remove the built instruction.
    2196          38 :     void undo() override {
    2197             :       DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
    2198          38 :       if (Instruction *IVal = dyn_cast<Instruction>(Val))
    2199          38 :         IVal->eraseFromParent();
    2200          38 :     }
    2201             :   };
    2202             : 
    2203             :   /// \brief Build a sign extension instruction.
    2204          88 :   class SExtBuilder : public TypePromotionAction {
    2205             :     Value *Val;
    2206             : 
    2207             :   public:
    2208             :     /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
    2209             :     /// result.
    2210             :     /// sext Opnd to Ty.
    2211          88 :     SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
    2212         176 :         : TypePromotionAction(InsertPt) {
    2213          88 :       IRBuilder<> Builder(InsertPt);
    2214          88 :       Val = Builder.CreateSExt(Opnd, Ty, "promoted");
    2215             :       DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
    2216          88 :     }
    2217             : 
    2218             :     /// \brief Get the built value.
    2219             :     Value *getBuiltValue() { return Val; }
    2220             : 
    2221             :     /// \brief Remove the built instruction.
    2222          46 :     void undo() override {
    2223             :       DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
    2224          46 :       if (Instruction *IVal = dyn_cast<Instruction>(Val))
    2225          46 :         IVal->eraseFromParent();
    2226          46 :     }
    2227             :   };
    2228             : 
    2229             :   /// \brief Build a zero extension instruction.
    2230         125 :   class ZExtBuilder : public TypePromotionAction {
    2231             :     Value *Val;
    2232             : 
    2233             :   public:
    2234             :     /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
    2235             :     /// result.
    2236             :     /// zext Opnd to Ty.
    2237         125 :     ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
    2238         250 :         : TypePromotionAction(InsertPt) {
    2239         125 :       IRBuilder<> Builder(InsertPt);
    2240         125 :       Val = Builder.CreateZExt(Opnd, Ty, "promoted");
    2241             :       DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
    2242         125 :     }
    2243             : 
    2244             :     /// \brief Get the built value.
    2245             :     Value *getBuiltValue() { return Val; }
    2246             : 
    2247             :     /// \brief Remove the built instruction.
    2248          59 :     void undo() override {
    2249             :       DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
    2250          59 :       if (Instruction *IVal = dyn_cast<Instruction>(Val))
    2251          58 :         IVal->eraseFromParent();
    2252          59 :     }
    2253             :   };
    2254             : 
    2255             :   /// \brief Mutate an instruction to another type.
    2256         554 :   class TypeMutator : public TypePromotionAction {
    2257             :     /// Record the original type.
    2258             :     Type *OrigTy;
    2259             : 
    2260             :   public:
    2261             :     /// \brief Mutate the type of \p Inst into \p NewTy.
    2262             :     TypeMutator(Instruction *Inst, Type *NewTy)
    2263        1108 :         : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
    2264             :       DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
    2265             :                    << "\n");
    2266             :       Inst->mutateType(NewTy);
    2267             :     }
    2268             : 
    2269             :     /// \brief Mutate the instruction back to its original type.
    2270         326 :     void undo() override {
    2271             :       DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
    2272             :                    << "\n");
    2273         326 :       Inst->mutateType(OrigTy);
    2274         326 :     }
    2275             :   };
    2276             : 
    2277             :   /// \brief Replace the uses of an instruction by another instruction.
    2278        1390 :   class UsesReplacer : public TypePromotionAction {
    2279             :     /// Helper structure to keep track of the replaced uses.
    2280             :     struct InstructionAndIdx {
    2281             :       /// The instruction using the instruction.
    2282             :       Instruction *Inst;
    2283             : 
    2284             :       /// The index where this instruction is used for Inst.
    2285             :       unsigned Idx;
    2286             : 
    2287             :       InstructionAndIdx(Instruction *Inst, unsigned Idx)
    2288         888 :           : Inst(Inst), Idx(Idx) {}
    2289             :     };
    2290             : 
    2291             :     /// Keep track of the original uses (pair Instruction, Index).
    2292             :     SmallVector<InstructionAndIdx, 4> OriginalUses;
    2293             : 
    2294             :     using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
    2295             : 
    2296             :   public:
    2297             :     /// \brief Replace all the use of \p Inst by \p New.
    2298        1390 :     UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
    2299             :       DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
    2300             :                    << "\n");
    2301             :       // Record the original uses.
    2302        1583 :       for (Use &U : Inst->uses()) {
    2303         888 :         Instruction *UserI = cast<Instruction>(U.getUser());
    2304        1776 :         OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
    2305             :       }
    2306             :       // Now, we can replace the uses.
    2307         695 :       Inst->replaceAllUsesWith(New);
    2308         695 :     }
    2309             : 
    2310             :     /// \brief Reassign the original uses of Inst to Inst.
    2311         406 :     void undo() override {
    2312             :       DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
    2313         570 :       for (use_iterator UseIt = OriginalUses.begin(),
    2314             :                         EndIt = OriginalUses.end();
    2315         976 :            UseIt != EndIt; ++UseIt) {
    2316         570 :         UseIt->Inst->setOperand(UseIt->Idx, Inst);
    2317             :       }
    2318         406 :     }
    2319             :   };
    2320             : 
    2321             :   /// \brief Remove an instruction from the IR.
    2322             :   class InstructionRemover : public TypePromotionAction {
    2323             :     /// Original position of the instruction.
    2324             :     InsertionHandler Inserter;
    2325             : 
    2326             :     /// Helper structure to hide all the link to the instruction. In other
    2327             :     /// words, this helps to do as if the instruction was removed.
    2328             :     OperandsHider Hider;
    2329             : 
    2330             :     /// Keep track of the uses replaced, if any.
    2331             :     UsesReplacer *Replacer = nullptr;
    2332             : 
    2333             :     /// Keep track of instructions removed.
    2334             :     SetOfInstrs &RemovedInsts;
    2335             : 
    2336             :   public:
    2337             :     /// \brief Remove all reference of \p Inst and optinally replace all its
    2338             :     /// uses with New.
    2339             :     /// \p RemovedInsts Keep track of the instructions removed by this Action.
    2340             :     /// \pre If !Inst->use_empty(), then New != nullptr
    2341         181 :     InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
    2342             :                        Value *New = nullptr)
    2343         181 :         : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
    2344         362 :           RemovedInsts(RemovedInsts) {
    2345         181 :       if (New)
    2346           0 :         Replacer = new UsesReplacer(Inst, New);
    2347             :       DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
    2348         181 :       RemovedInsts.insert(Inst);
    2349             :       /// The instructions removed here will be freed after completing
    2350             :       /// optimizeBlock() for all blocks as we need to keep track of the
    2351             :       /// removed instructions during promotion.
    2352         181 :       Inst->removeFromParent();
    2353         181 :     }
    2354             : 
    2355         543 :     ~InstructionRemover() override { delete Replacer; }
    2356             : 
    2357             :     /// \brief Resurrect the instruction and reassign it to the proper uses if
    2358             :     /// new value was provided when build this action.
    2359          92 :     void undo() override {
    2360             :       DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
    2361          92 :       Inserter.insert(Inst);
    2362          92 :       if (Replacer)
    2363           0 :         Replacer->undo();
    2364          92 :       Hider.undo();
    2365          92 :       RemovedInsts.erase(Inst);
    2366          92 :     }
    2367             :   };
    2368             : 
    2369             : public:
    2370             :   /// Restoration point.
    2371             :   /// The restoration point is a pointer to an action instead of an iterator
    2372             :   /// because the iterator may be invalidated but not the pointer.
    2373             :   using ConstRestorationPt = const TypePromotionAction *;
    2374             : 
    2375             :   TypePromotionTransaction(SetOfInstrs &RemovedInsts)
    2376     2140566 :       : RemovedInsts(RemovedInsts) {}
    2377             : 
    2378             :   /// Advocate every changes made in that transaction.
    2379             :   void commit();
    2380             : 
    2381             :   /// Undo all the changes made after the given point.
    2382             :   void rollback(ConstRestorationPt Point);
    2383             : 
    2384             :   /// Get the current restoration point.
    2385             :   ConstRestorationPt getRestorationPoint() const;
    2386             : 
    2387             :   /// \name API for IR modification with state keeping to support rollback.
    2388             :   /// @{
    2389             :   /// Same as Instruction::setOperand.
    2390             :   void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
    2391             : 
    2392             :   /// Same as Instruction::eraseFromParent.
    2393             :   void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
    2394             : 
    2395             :   /// Same as Value::replaceAllUsesWith.
    2396             :   void replaceAllUsesWith(Instruction *Inst, Value *New);
    2397             : 
    2398             :   /// Same as Value::mutateType.
    2399             :   void mutateType(Instruction *Inst, Type *NewTy);
    2400             : 
    2401             :   /// Same as IRBuilder::createTrunc.
    2402             :   Value *createTrunc(Instruction *Opnd, Type *Ty);
    2403             : 
    2404             :   /// Same as IRBuilder::createSExt.
    2405             :   Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
    2406             : 
    2407             :   /// Same as IRBuilder::createZExt.
    2408             :   Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
    2409             : 
    2410             :   /// Same as Instruction::moveBefore.
    2411             :   void moveBefore(Instruction *Inst, Instruction *Before);
    2412             :   /// @}
    2413             : 
    2414             : private:
    2415             :   /// The ordered list of actions made so far.
    2416             :   SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
    2417             : 
    2418             :   using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
    2419             : 
    2420             :   SetOfInstrs &RemovedInsts;
    2421             : };
    2422             : 
    2423             : } // end anonymous namespace
    2424             : 
    2425        1834 : void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
    2426             :                                           Value *NewVal) {
    2427        5502 :   Actions.push_back(llvm::make_unique<TypePromotionTransaction::OperandSetter>(
    2428             :       Inst, Idx, NewVal));
    2429        1834 : }
    2430             : 
    2431         181 : void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
    2432             :                                                 Value *NewVal) {
    2433         543 :   Actions.push_back(
    2434         181 :       llvm::make_unique<TypePromotionTransaction::InstructionRemover>(
    2435         181 :           Inst, RemovedInsts, NewVal));
    2436         181 : }
    2437             : 
    2438         695 : void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
    2439             :                                                   Value *New) {
    2440        2085 :   Actions.push_back(
    2441             :       llvm::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
    2442         695 : }
    2443             : 
    2444         554 : void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
    2445        1662 :   Actions.push_back(
    2446             :       llvm::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
    2447         554 : }
    2448             : 
    2449          46 : Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
    2450             :                                              Type *Ty) {
    2451          46 :   std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
    2452          46 :   Value *Val = Ptr->getBuiltValue();
    2453         138 :   Actions.push_back(std::move(Ptr));
    2454          46 :   return Val;
    2455             : }
    2456             : 
    2457          88 : Value *TypePromotionTransaction::createSExt(Instruction *Inst,
    2458             :                                             Value *Opnd, Type *Ty) {
    2459          88 :   std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
    2460          88 :   Value *Val = Ptr->getBuiltValue();
    2461         264 :   Actions.push_back(std::move(Ptr));
    2462          88 :   return Val;
    2463             : }
    2464             : 
    2465         125 : Value *TypePromotionTransaction::createZExt(Instruction *Inst,
    2466             :                                             Value *Opnd, Type *Ty) {
    2467         125 :   std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
    2468         125 :   Value *Val = Ptr->getBuiltValue();
    2469         375 :   Actions.push_back(std::move(Ptr));
    2470         125 :   return Val;
    2471             : }
    2472             : 
    2473         666 : void TypePromotionTransaction::moveBefore(Instruction *Inst,
    2474             :                                           Instruction *Before) {
    2475        1998 :   Actions.push_back(
    2476             :       llvm::make_unique<TypePromotionTransaction::InstructionMoveBefore>(
    2477             :           Inst, Before));
    2478         666 : }
    2479             : 
    2480             : TypePromotionTransaction::ConstRestorationPt
    2481             : TypePromotionTransaction::getRestorationPoint() const {
    2482     2025167 :   return !Actions.empty() ? Actions.back().get() : nullptr;
    2483             : }
    2484             : 
    2485             : void TypePromotionTransaction::commit() {
    2486             :   for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
    2487             :        ++It)
    2488             :     (*It)->commit();
    2489             :   Actions.clear();
    2490             : }
    2491             : 
    2492       42787 : void TypePromotionTransaction::rollback(
    2493             :     TypePromotionTransaction::ConstRestorationPt Point) {
    2494       50603 :   while (!Actions.empty() && Point != Actions.back().get()) {
    2495             :     std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
    2496        2446 :     Curr->undo();
    2497             :   }
    2498       42787 : }
    2499             : 
    2500             : namespace {
    2501             : 
    2502             : /// \brief A helper class for matching addressing modes.
    2503             : ///
    2504             : /// This encapsulates the logic for matching the target-legal addressing modes.
    2505             : class AddressingModeMatcher {
    2506             :   SmallVectorImpl<Instruction*> &AddrModeInsts;
    2507             :   const TargetLowering &TLI;
    2508             :   const TargetRegisterInfo &TRI;
    2509             :   const DataLayout &DL;
    2510             : 
    2511             :   /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
    2512             :   /// the memory instruction that we're computing this address for.
    2513             :   Type *AccessTy;
    2514             :   unsigned AddrSpace;
    2515             :   Instruction *MemoryInst;
    2516             : 
    2517             :   /// This is the addressing mode that we're building up. This is
    2518             :   /// part of the return value of this addressing mode matching stuff.
    2519             :   ExtAddrMode &AddrMode;
    2520             : 
    2521             :   /// The instructions inserted by other CodeGenPrepare optimizations.
    2522             :   const SetOfInstrs &InsertedInsts;
    2523             : 
    2524             :   /// A map from the instructions to their type before promotion.
    2525             :   InstrToOrigTy &PromotedInsts;
    2526             : 
    2527             :   /// The ongoing transaction where every action should be registered.
    2528             :   TypePromotionTransaction &TPT;
    2529             : 
    2530             :   /// This is set to true when we should not do profitability checks.
    2531             :   /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
    2532             :   bool IgnoreProfitability;
    2533             : 
    2534             :   AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
    2535             :                         const TargetLowering &TLI,
    2536             :                         const TargetRegisterInfo &TRI,
    2537             :                         Type *AT, unsigned AS,
    2538             :                         Instruction *MI, ExtAddrMode &AM,
    2539             :                         const SetOfInstrs &InsertedInsts,
    2540             :                         InstrToOrigTy &PromotedInsts,
    2541             :                         TypePromotionTransaction &TPT)
    2542     1053073 :       : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
    2543     1053073 :         DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
    2544             :         MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
    2545     2106146 :         PromotedInsts(PromotedInsts), TPT(TPT) {
    2546     1048258 :     IgnoreProfitability = false;
    2547             :   }
    2548             : 
    2549             : public:
    2550             :   /// Find the maximal addressing mode that a load/store of V can fold,
    2551             :   /// give an access type of AccessTy.  This returns a list of involved
    2552             :   /// instructions in AddrModeInsts.
    2553             :   /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
    2554             :   /// optimizations.
    2555             :   /// \p PromotedInsts maps the instructions to their type before promotion.
    2556             :   /// \p The ongoing transaction where every action should be registered.
    2557     1048258 :   static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
    2558             :                            Instruction *MemoryInst,
    2559             :                            SmallVectorImpl<Instruction*> &AddrModeInsts,
    2560             :                            const TargetLowering &TLI,
    2561             :                            const TargetRegisterInfo &TRI,
    2562             :                            const SetOfInstrs &InsertedInsts,
    2563             :                            InstrToOrigTy &PromotedInsts,
    2564             :                            TypePromotionTransaction &TPT) {
    2565     1048258 :     ExtAddrMode Result;
    2566             : 
    2567     1048258 :     bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI,
    2568             :                                          AccessTy, AS,
    2569             :                                          MemoryInst, Result, InsertedInsts,
    2570     1048258 :                                          PromotedInsts, TPT).matchAddr(V, 0);
    2571             :     (void)Success; assert(Success && "Couldn't select *anything*?");
    2572     1048258 :     return Result;
    2573             :   }
    2574             : 
    2575             : private:
    2576             :   bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
    2577             :   bool matchAddr(Value *V, unsigned Depth);
    2578             :   bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
    2579             :                           bool *MovedAway = nullptr);
    2580             :   bool isProfitableToFoldIntoAddressingMode(Instruction *I,
    2581             :                                             ExtAddrMode &AMBefore,
    2582             :                                             ExtAddrMode &AMAfter);
    2583             :   bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
    2584             :   bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
    2585             :                              Value *PromotedOperand) const;
    2586             : };
    2587             : 
    2588             : /// \brief Keep track of simplification of Phi nodes.
    2589             : /// Accept the set of all phi nodes and erase phi node from this set
    2590             : /// if it is simplified.
    2591         904 : class SimplificationTracker {
    2592             :   DenseMap<Value *, Value *> Storage;
    2593             :   const SimplifyQuery &SQ;
    2594             :   SmallPtrSetImpl<PHINode *> &AllPhiNodes;
    2595             :   SmallPtrSetImpl<SelectInst *> &AllSelectNodes;
    2596             : 
    2597             : public:
    2598             :   SimplificationTracker(const SimplifyQuery &sq,
    2599             :                         SmallPtrSetImpl<PHINode *> &APN,
    2600             :                         SmallPtrSetImpl<SelectInst *> &ASN)
    2601        1808 :       : SQ(sq), AllPhiNodes(APN), AllSelectNodes(ASN) {}
    2602             : 
    2603             :   Value *Get(Value *V) {
    2604             :     do {
    2605        4600 :       auto SV = Storage.find(V);
    2606        4600 :       if (SV == Storage.end())
    2607        4099 :         return V;
    2608         501 :       V = SV->second;
    2609             :     } while (true);
    2610             :   }
    2611             : 
    2612        1339 :   Value *Simplify(Value *Val) {
    2613             :     SmallVector<Value *, 32> WorkList;
    2614             :     SmallPtrSet<Value *, 32> Visited;
    2615        1339 :     WorkList.push_back(Val);
    2616        2734 :     while (!WorkList.empty()) {
    2617             :       auto P = WorkList.pop_back_val();
    2618        1395 :       if (!Visited.insert(P).second)
    2619          35 :         continue;
    2620             :       if (auto *PI = dyn_cast<Instruction>(P))
    2621        1360 :         if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) {
    2622         462 :           for (auto *U : PI->users())
    2623          56 :             WorkList.push_back(cast<Value>(U));
    2624             :           Put(PI, V);
    2625         406 :           PI->replaceAllUsesWith(V);
    2626             :           if (auto *PHI = dyn_cast<PHINode>(PI))
    2627         406 :             AllPhiNodes.erase(PHI);
    2628             :           if (auto *Select = dyn_cast<SelectInst>(PI))
    2629           0 :             AllSelectNodes.erase(Select);
    2630         406 :           PI->eraseFromParent();
    2631             :         }
    2632             :     }
    2633        2678 :     return Get(Val);
    2634             :   }
    2635             : 
    2636             :   void Put(Value *From, Value *To) {
    2637         860 :     Storage.insert({ From, To });
    2638             :   }
    2639             : };
    2640             : 
    2641             : /// \brief A helper class for combining addressing modes.
    2642     1041123 : class AddressingModeCombiner {
    2643             :   typedef std::pair<Value *, BasicBlock *> ValueInBB;
    2644             :   typedef DenseMap<ValueInBB, Value *> FoldAddrToValueMapping;
    2645             :   typedef std::pair<PHINode *, PHINode *> PHIPair;
    2646             : 
    2647             : private:
    2648             :   /// The addressing modes we've collected.
    2649             :   SmallVector<ExtAddrMode, 16> AddrModes;
    2650             : 
    2651             :   /// The field in which the AddrModes differ, when we have more than one.
    2652             :   ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
    2653             : 
    2654             :   /// Are the AddrModes that we have all just equal to their original values?
    2655             :   bool AllAddrModesTrivial = true;
    2656             : 
    2657             :   /// Common Type for all different fields in addressing modes.
    2658             :   Type *CommonType;
    2659             : 
    2660             :   /// SimplifyQuery for simplifyInstruction utility.
    2661             :   const SimplifyQuery &SQ;
    2662             : 
    2663             :   /// Original Address.
    2664             :   ValueInBB Original;
    2665             : 
    2666             : public:
    2667             :   AddressingModeCombiner(const SimplifyQuery &_SQ, ValueInBB OriginalValue)
    2668     2082246 :       : CommonType(nullptr), SQ(_SQ), Original(OriginalValue) {}
    2669             : 
    2670             :   /// \brief Get the combined AddrMode
    2671             :   const ExtAddrMode &getAddrMode() const {
    2672             :     return AddrModes[0];
    2673             :   }
    2674             : 
    2675             :   /// \brief Add a new AddrMode if it's compatible with the AddrModes we already
    2676             :   /// have.
    2677             :   /// \return True iff we succeeded in doing so.
    2678     1048258 :   bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
    2679             :     // Take note of if we have any non-trivial AddrModes, as we need to detect
    2680             :     // when all AddrModes are trivial as then we would introduce a phi or select
    2681             :     // which just duplicates what's already there.
    2682     2096516 :     AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
    2683             : 
    2684             :     // If this is the first addrmode then everything is fine.
    2685     1048258 :     if (AddrModes.empty()) {
    2686     1041123 :       AddrModes.emplace_back(NewAddrMode);
    2687     1041123 :       return true;
    2688             :     }
    2689             : 
    2690             :     // Figure out how different this is from the other address modes, which we
    2691             :     // can do just by comparing against the first one given that we only care
    2692             :     // about the cumulative difference.
    2693             :     ExtAddrMode::FieldName ThisDifferentField =
    2694        7135 :       AddrModes[0].compare(NewAddrMode);
    2695        7135 :     if (DifferentField == ExtAddrMode::NoField)
    2696        6946 :       DifferentField = ThisDifferentField;
    2697         189 :     else if (DifferentField != ThisDifferentField)
    2698          72 :       DifferentField = ExtAddrMode::MultipleFields;
    2699             : 
    2700             :     // If NewAddrMode differs in more than one dimension we cannot handle it.
    2701        7135 :     bool CanHandle = DifferentField != ExtAddrMode::MultipleFields;
    2702             : 
    2703             :     // If Scale Field is different then we reject.
    2704        7135 :     CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField;
    2705             : 
    2706             :     // We also must reject the case when base offset is different and
    2707             :     // scale reg is not null, we cannot handle this case due to merge of
    2708             :     // different offsets will be used as ScaleReg.
    2709        6149 :     CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField ||
    2710         560 :                               !NewAddrMode.ScaledReg);
    2711             : 
    2712             :     // We also must reject the case when GV is different and BaseReg installed
    2713             :     // due to we want to use base reg as a merge of GV values.
    2714        5703 :     CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField ||
    2715         115 :                               !NewAddrMode.HasBaseReg);
    2716             : 
    2717             :     // Even if NewAddMode is the same we still need to collect it due to
    2718             :     // original value is different. And later we will need all original values
    2719             :     // as anchors during finding the common Phi node.
    2720             :     if (CanHandle)
    2721        5587 :       AddrModes.emplace_back(NewAddrMode);
    2722             :     else
    2723             :       AddrModes.clear();
    2724             : 
    2725             :     return CanHandle;
    2726             :   }
    2727             : 
    2728             :   /// \brief Combine the addressing modes we've collected into a single
    2729             :   /// addressing mode.
    2730             :   /// \return True iff we successfully combined them or we only had one so
    2731             :   /// didn't need to combine them anyway.
    2732     1041123 :   bool combineAddrModes() {
    2733             :     // If we have no AddrModes then they can't be combined.
    2734     1041123 :     if (AddrModes.size() == 0)
    2735             :       return false;
    2736             : 
    2737             :     // A single AddrMode can trivially be combined.
    2738     1039575 :     if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
    2739             :       return true;
    2740             : 
    2741             :     // If the AddrModes we collected are all just equal to the value they are
    2742             :     // derived from then combining them wouldn't do anything useful.
    2743        1584 :     if (AllAddrModesTrivial)
    2744             :       return false;
    2745             : 
    2746         908 :     if (!addrModeCombiningAllowed())
    2747             :       return false;
    2748             : 
    2749             :     // Build a map between <original value, basic block where we saw it> to
    2750             :     // value of base register.
    2751             :     // Bail out if there is no common type.
    2752             :     FoldAddrToValueMapping Map;
    2753         908 :     if (!initializeMap(Map))
    2754             :       return false;
    2755             : 
    2756         904 :     Value *CommonValue = findCommon(Map);
    2757         904 :     if (CommonValue)
    2758          58 :       AddrModes[0].SetCombinedField(DifferentField, CommonValue, AddrModes);
    2759         904 :     return CommonValue != nullptr;
    2760             :   }
    2761             : 
    2762             : private:
    2763             :   /// \brief Initialize Map with anchor values. For address seen in some BB
    2764             :   /// we set the value of different field saw in this address.
    2765             :   /// If address is not an instruction than basic block is set to null.
    2766             :   /// At the same time we find a common type for different field we will
    2767             :   /// use to create new Phi/Select nodes. Keep it in CommonType field.
    2768             :   /// Return false if there is no common type found.
    2769         908 :   bool initializeMap(FoldAddrToValueMapping &Map) {
    2770             :     // Keep track of keys where the value is null. We will need to replace it
    2771             :     // with constant null when we know the common type.
    2772             :     SmallVector<ValueInBB, 2> NullValue;
    2773        1816 :     Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
    2774        4598 :     for (auto &AM : AddrModes) {
    2775             :       BasicBlock *BB = nullptr;
    2776        1849 :       if (Instruction *I = dyn_cast<Instruction>(AM.OriginalValue))
    2777         439 :         BB = I->getParent();
    2778             : 
    2779        1849 :       Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy);
    2780        1849 :       if (DV) {
    2781        1836 :         auto *Type = DV->getType();
    2782        1836 :         if (CommonType && CommonType != Type)
    2783             :           return false;
    2784        1832 :         CommonType = Type;
    2785        1832 :         Map[{ AM.OriginalValue, BB }] = DV;
    2786             :       } else {
    2787          13 :         NullValue.push_back({ AM.OriginalValue, BB });
    2788             :       }
    2789             :     }
    2790             :     assert(CommonType && "At least one non-null value must be!");
    2791         922 :     for (auto VIBB : NullValue)
    2792          18 :       Map[VIBB] = Constant::getNullValue(CommonType);
    2793         904 :     return true;
    2794             :   }
    2795             : 
    2796             :   /// \brief We have mapping between value A and basic block where value A
    2797             :   /// seen to other value B where B was a field in addressing mode represented
    2798             :   /// by A. Also we have an original value C representin an address in some
    2799             :   /// basic block. Traversing from C through phi and selects we ended up with
    2800             :   /// A's in a map. This utility function tries to find a value V which is a
    2801             :   /// field in addressing mode C and traversing through phi nodes and selects
    2802             :   /// we will end up in corresponded values B in a map.
    2803             :   /// The utility will create a new Phi/Selects if needed.
    2804             :   // The simple example looks as follows:
    2805             :   // BB1:
    2806             :   //   p1 = b1 + 40
    2807             :   //   br cond BB2, BB3
    2808             :   // BB2:
    2809             :   //   p2 = b2 + 40
    2810             :   //   br BB3
    2811             :   // BB3:
    2812             :   //   p = phi [p1, BB1], [p2, BB2]
    2813             :   //   v = load p
    2814             :   // Map is
    2815             :   //   <p1, BB1> -> b1
    2816             :   //   <p2, BB2> -> b2
    2817             :   // Request is
    2818             :   //   <p, BB3> -> ?
    2819             :   // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3
    2820         904 :   Value *findCommon(FoldAddrToValueMapping &Map) {
    2821             :     // Tracks newly created Phi nodes.
    2822             :     SmallPtrSet<PHINode *, 32> NewPhiNodes;
    2823             :     // Tracks newly created Select nodes.
    2824             :     SmallPtrSet<SelectInst *, 32> NewSelectNodes;
    2825             :     // Tracks the simplification of newly created phi nodes. The reason we use
    2826             :     // this mapping is because we will add new created Phi nodes in AddrToBase.
    2827             :     // Simplification of Phi nodes is recursive, so some Phi node may
    2828             :     // be simplified after we added it to AddrToBase.
    2829             :     // Using this mapping we can find the current value in AddrToBase.
    2830         904 :     SimplificationTracker ST(SQ, NewPhiNodes, NewSelectNodes);
    2831             : 
    2832             :     // First step, DFS to create PHI nodes for all intermediate blocks.
    2833             :     // Also fill traverse order for the second step.
    2834             :     SmallVector<ValueInBB, 32> TraverseOrder;
    2835         904 :     InsertPlaceholders(Map, TraverseOrder, NewPhiNodes, NewSelectNodes);
    2836             : 
    2837             :     // Second Step, fill new nodes by merged values and simplify if possible.
    2838         904 :     FillPlaceholders(Map, TraverseOrder, ST);
    2839             : 
    2840         904 :     if (!AddrSinkNewSelects && NewSelectNodes.size() > 0) {
    2841           0 :       DestroyNodes(NewPhiNodes);
    2842           0 :       DestroyNodes(NewSelectNodes);
    2843           0 :       return nullptr;
    2844             :     }
    2845             : 
    2846             :     // Now we'd like to match New Phi nodes to existed ones.
    2847         904 :     unsigned PhiNotMatchedCount = 0;
    2848         904 :     if (!MatchPhiSet(NewPhiNodes, ST, AddrSinkNewPhis, PhiNotMatchedCount)) {
    2849         846 :       DestroyNodes(NewPhiNodes);
    2850         846 :       DestroyNodes(NewSelectNodes);
    2851         846 :       return nullptr;
    2852             :     }
    2853             : 
    2854         116 :     auto *Result = ST.Get(Map.find(Original)->second);
    2855             :     if (Result) {
    2856             :       NumMemoryInstsPhiCreated += NewPhiNodes.size() + PhiNotMatchedCount;
    2857             :       NumMemoryInstsSelectCreated += NewSelectNodes.size();
    2858             :     }
    2859             :     return Result;
    2860             :   }
    2861             : 
    2862             :   /// \brief Destroy nodes from a set.
    2863        1692 :   template <typename T> void DestroyNodes(SmallPtrSetImpl<T *> &Instructions) {
    2864             :     // For safe erasing, replace the Phi with dummy value first.
    2865        1692 :     auto Dummy = UndefValue::get(CommonType);
    2866        1692 :     for (auto I : Instructions) {
    2867         861 :       I->replaceAllUsesWith(Dummy);
    2868         861 :       I->eraseFromParent();
    2869             :     }
    2870        1692 :   }
    2871             : 
    2872             :   /// \brief Try to match PHI node to Candidate.
    2873             :   /// Matcher tracks the matched Phi nodes.
    2874        1586 :   bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
    2875             :                     DenseSet<PHIPair> &Matcher,
    2876             :                     SmallPtrSetImpl<PHINode *> &PhiNodesToMatch) {
    2877             :     SmallVector<PHIPair, 8> WorkList;
    2878        1586 :     Matcher.insert({ PHI, Candidate });
    2879        1586 :     WorkList.push_back({ PHI, Candidate });
    2880        1586 :     SmallSet<PHIPair, 8> Visited;
    2881        1615 :     while (!WorkList.empty()) {
    2882        1593 :       auto Item = WorkList.pop_back_val();
    2883        1593 :       if (!Visited.insert(Item).second)
    2884           0 :         continue;
    2885             :       // We iterate over all incoming values to Phi to compare them.
    2886             :       // If values are different and both of them Phi and the first one is a
    2887             :       // Phi we added (subject to match) and both of them is in the same basic
    2888             :       // block then we can match our pair if values match. So we state that
    2889             :       // these values match and add it to work list to verify that.
    2890        3308 :       for (auto B : Item.first->blocks()) {
    2891        1625 :         Value *FirstValue = Item.first->getIncomingValueForBlock(B);
    2892        1625 :         Value *SecondValue = Item.second->getIncomingValueForBlock(B);
    2893        1625 :         if (FirstValue == SecondValue)
    2894             :           continue;
    2895             : 
    2896             :         PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue);
    2897             :         PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue);
    2898             : 
    2899             :         // One of them is not Phi or
    2900             :         // The first one is not Phi node from the set we'd like to match or
    2901             :         // Phi nodes from different basic blocks then
    2902             :         // we will not be able to match.
    2903        1587 :         if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) ||
    2904          13 :             FirstPhi->getParent() != SecondPhi->getParent())
    2905             :           return false;
    2906             : 
    2907             :         // If we already matched them then continue.
    2908          10 :         if (Matcher.count({ FirstPhi, SecondPhi }))
    2909             :           continue;
    2910             :         // So the values are different and does not match. So we need them to
    2911             :         // match.
    2912           8 :         Matcher.insert({ FirstPhi, SecondPhi });
    2913             :         // But me must check it.
    2914           8 :         WorkList.push_back({ FirstPhi, SecondPhi });
    2915             :       }
    2916             :     }
    2917             :     return true;
    2918             :   }
    2919             : 
    2920             :   /// \brief For the given set of PHI nodes try to find their equivalents.
    2921             :   /// Returns false if this matching fails and creation of new Phi is disabled.
    2922         904 :   bool MatchPhiSet(SmallPtrSetImpl<PHINode *> &PhiNodesToMatch,
    2923             :                    SimplificationTracker &ST, bool AllowNewPhiNodes,
    2924             :                    unsigned &PhiNotMatchedCount) {
    2925             :     DenseSet<PHIPair> Matched;
    2926             :     SmallPtrSet<PHINode *, 8> WillNotMatch;
    2927         949 :     while (PhiNodesToMatch.size()) {
    2928        1782 :       PHINode *PHI = *PhiNodesToMatch.begin();
    2929             : 
    2930             :       // Add us, if no Phi nodes in the basic block we do not match.
    2931         891 :       WillNotMatch.clear();
    2932         891 :       WillNotMatch.insert(PHI);
    2933             : 
    2934             :       // Traverse all Phis until we found equivalent or fail to do that.
    2935             :       bool IsMatched = false;
    2936         891 :       for (auto &P : PHI->getParent()->phis()) {
    2937        2477 :         if (&P == PHI)
    2938         891 :           continue;
    2939        1586 :         if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch)))
    2940             :           break;
    2941             :         // If it does not match, collect all Phi nodes from matcher.
    2942             :         // if we end up with no match, them all these Phi nodes will not match
    2943             :         // later.
    2944        4698 :         for (auto M : Matched)
    2945        1570 :           WillNotMatch.insert(M.first);
    2946             :         Matched.clear();
    2947             :       }
    2948         891 :       if (IsMatched) {
    2949             :         // Replace all matched values and erase them.
    2950          68 :         for (auto MV : Matched) {
    2951          24 :           MV.first->replaceAllUsesWith(MV.second);
    2952             :           PhiNodesToMatch.erase(MV.first);
    2953             :           ST.Put(MV.first, MV.second);
    2954          24 :           MV.first->eraseFromParent();
    2955             :         }
    2956             :         Matched.clear();
    2957          22 :         continue;
    2958             :       }
    2959             :       // If we are not allowed to create new nodes then bail out.
    2960         869 :       if (!AllowNewPhiNodes)
    2961             :         return false;
    2962             :       // Just remove all seen values in matcher. They will not match anything.
    2963          46 :       PhiNotMatchedCount += WillNotMatch.size();
    2964          23 :       for (auto *P : WillNotMatch)
    2965             :         PhiNodesToMatch.erase(P);
    2966             :     }
    2967             :     return true;
    2968             :   }
    2969             :   /// \brief Fill the placeholder with values from predecessors and simplify it.
    2970         904 :   void FillPlaceholders(FoldAddrToValueMapping &Map,
    2971             :                         SmallVectorImpl<ValueInBB> &TraverseOrder,
    2972             :                         SimplificationTracker &ST) {
    2973        2243 :     while (!TraverseOrder.empty()) {
    2974        1339 :       auto Current = TraverseOrder.pop_back_val();
    2975             :       assert(Map.find(Current) != Map.end() && "No node to fill!!!");
    2976             :       Value *CurrentValue = Current.first;
    2977             :       BasicBlock *CurrentBlock = Current.second;
    2978        1339 :       Value *V = Map[Current];
    2979             : 
    2980             :       if (SelectInst *Select = dyn_cast<SelectInst>(V)) {
    2981             :         // CurrentValue also must be Select.
    2982             :         auto *CurrentSelect = cast<SelectInst>(CurrentValue);
    2983             :         auto *TrueValue = CurrentSelect->getTrueValue();
    2984             :         ValueInBB TrueItem = { TrueValue, isa<Instruction>(TrueValue)
    2985          24 :                                               ? CurrentBlock
    2986             :                                               : nullptr };
    2987             :         assert(Map.find(TrueItem) != Map.end() && "No True Value!");
    2988          24 :         Select->setTrueValue(ST.Get(Map[TrueItem]));
    2989             :         auto *FalseValue = CurrentSelect->getFalseValue();
    2990             :         ValueInBB FalseItem = { FalseValue, isa<Instruction>(FalseValue)
    2991          24 :                                                 ? CurrentBlock
    2992             :                                                 : nullptr };
    2993             :         assert(Map.find(FalseItem) != Map.end() && "No False Value!");
    2994          24 :         Select->setFalseValue(ST.Get(Map[FalseItem]));
    2995             :       } else {
    2996             :         // Must be a Phi node then.
    2997             :         PHINode *PHI = cast<PHINode>(V);
    2998             :         // Fill the Phi node with values from predecessors.
    2999             :         bool IsDefinedInThisBB =
    3000        1315 :             cast<Instruction>(CurrentValue)->getParent() == CurrentBlock;
    3001             :         auto *CurrentPhi = dyn_cast<PHINode>(CurrentValue);
    3002        5284 :         for (auto B : predecessors(CurrentBlock)) {
    3003             :           Value *PV = IsDefinedInThisBB
    3004        2654 :                           ? CurrentPhi->getIncomingValueForBlock(B)
    3005             :                           : CurrentValue;
    3006        2654 :           ValueInBB item = { PV, isa<Instruction>(PV) ? B : nullptr };
    3007             :           assert(Map.find(item) != Map.end() && "No predecessor Value!");
    3008        5308 :           PHI->addIncoming(ST.Get(Map[item]), B);
    3009             :         }
    3010             :       }
    3011             :       // Simplify if possible.
    3012        1339 :       Map[Current] = ST.Simplify(V);
    3013             :     }
    3014         904 :   }
    3015             : 
    3016             :   /// Starting from value recursively iterates over predecessors up to known
    3017             :   /// ending values represented in a map. For each traversed block inserts
    3018             :   /// a placeholder Phi or Select.
    3019             :   /// Reports all new created Phi/Select nodes by adding them to set.
    3020             :   /// Also reports and order in what basic blocks have been traversed.
    3021         904 :   void InsertPlaceholders(FoldAddrToValueMapping &Map,
    3022             :                           SmallVectorImpl<ValueInBB> &TraverseOrder,
    3023             :                           SmallPtrSetImpl<PHINode *> &NewPhiNodes,
    3024             :                           SmallPtrSetImpl<SelectInst *> &NewSelectNodes) {
    3025             :     SmallVector<ValueInBB, 32> Worklist;
    3026             :     assert((isa<PHINode>(Original.first) || isa<SelectInst>(Original.first)) &&
    3027             :            "Address must be a Phi or Select node");
    3028         904 :     auto *Dummy = UndefValue::get(CommonType);
    3029         904 :     Worklist.push_back(Original);
    3030        4510 :     while (!Worklist.empty()) {
    3031        3606 :       auto Current = Worklist.pop_back_val();
    3032             :       // If value is not an instruction it is something global, constant,
    3033             :       // parameter and we can say that this value is observable in any block.
    3034             :       // Set block to null to denote it.
    3035             :       // Also please take into account that it is how we build anchors.
    3036        3606 :       if (!isa<Instruction>(Current.first))
    3037        1702 :         Current.second = nullptr;
    3038             :       // if it is already visited or it is an ending value then skip it.
    3039        3606 :       if (Map.find(Current) != Map.end())
    3040        4940 :         continue;
    3041        1339 :       TraverseOrder.push_back(Current);
    3042             : 
    3043        1339 :       Value *CurrentValue = Current.first;
    3044        1339 :       BasicBlock *CurrentBlock = Current.second;
    3045             :       // CurrentValue must be a Phi node or select. All others must be covered
    3046             :       // by anchors.
    3047             :       Instruction *CurrentI = cast<Instruction>(CurrentValue);
    3048        1339 :       bool IsDefinedInThisBB = CurrentI->getParent() == CurrentBlock;
    3049             : 
    3050             :       unsigned PredCount =
    3051        1339 :           std::distance(pred_begin(CurrentBlock), pred_end(CurrentBlock));
    3052             :       // if Current Value is not defined in this basic block we are interested
    3053             :       // in values in predecessors.
    3054        1339 :       if (!IsDefinedInThisBB) {
    3055             :         assert(PredCount && "Unreachable block?!");
    3056         812 :         PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
    3057             :                                        &CurrentBlock->front());
    3058         406 :         Map[Current] = PHI;
    3059         406 :         NewPhiNodes.insert(PHI);
    3060             :         // Add all predecessors in work list.
    3061        1339 :         for (auto B : predecessors(CurrentBlock))
    3062         527 :           Worklist.push_back({ CurrentValue, B });
    3063         406 :         continue;
    3064             :       }
    3065             :       // Value is defined in this basic block.
    3066             :       if (SelectInst *OrigSelect = dyn_cast<SelectInst>(CurrentI)) {
    3067             :         // Is it OK to get metadata from OrigSelect?!
    3068             :         // Create a Select placeholder with dummy value.
    3069             :         SelectInst *Select =
    3070          24 :             SelectInst::Create(OrigSelect->getCondition(), Dummy, Dummy,
    3071          72 :                                OrigSelect->getName(), OrigSelect, OrigSelect);
    3072          24 :         Map[Current] = Select;
    3073          24 :         NewSelectNodes.insert(Select);
    3074             :         // We are interested in True and False value in this basic block.
    3075          24 :         Worklist.push_back({ OrigSelect->getTrueValue(), CurrentBlock });
    3076          24 :         Worklist.push_back({ OrigSelect->getFalseValue(), CurrentBlock });
    3077             :       } else {
    3078             :         // It must be a Phi node then.
    3079             :         auto *CurrentPhi = cast<PHINode>(CurrentI);
    3080             :         // Create new Phi node for merge of bases.
    3081             :         assert(PredCount && "Unreachable block?!");
    3082        1818 :         PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
    3083             :                                        &CurrentBlock->front());
    3084         909 :         Map[Current] = PHI;
    3085         909 :         NewPhiNodes.insert(PHI);
    3086             : 
    3087             :         // Add all predecessors in work list.
    3088        3945 :         for (auto B : predecessors(CurrentBlock))
    3089        4254 :           Worklist.push_back({ CurrentPhi->getIncomingValueForBlock(B), B });
    3090             :       }
    3091             :     }
    3092         904 :   }
    3093             : 
    3094         908 :   bool addrModeCombiningAllowed() {
    3095         908 :     if (DisableComplexAddrModes)
    3096             :       return false;
    3097         908 :     switch (DifferentField) {
    3098             :     default:
    3099             :       return false;
    3100             :     case ExtAddrMode::BaseRegField:
    3101             :       return AddrSinkCombineBaseReg;
    3102             :     case ExtAddrMode::BaseGVField:
    3103             :       return AddrSinkCombineBaseGV;
    3104             :     case ExtAddrMode::BaseOffsField:
    3105             :       return AddrSinkCombineBaseOffs;
    3106             :     case ExtAddrMode::ScaledRegField:
    3107             :       return AddrSinkCombineScaledReg;
    3108             :     }
    3109             :   }
    3110             : };
    3111             : } // end anonymous namespace
    3112             : 
    3113             : /// Try adding ScaleReg*Scale to the current addressing mode.
    3114             : /// Return true and update AddrMode if this addr mode is legal for the target,
    3115             : /// false if not.
    3116      180035 : bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
    3117             :                                              unsigned Depth) {
    3118             :   // If Scale is 1, then this is the same as adding ScaleReg to the addressing
    3119             :   // mode.  Just process that directly.
    3120      180035 :   if (Scale == 1)
    3121       10010 :     return matchAddr(ScaleReg, Depth);
    3122             : 
    3123             :   // If the scale is 0, it takes nothing to add this.
    3124      170025 :   if (Scale == 0)
    3125             :     return true;
    3126             : 
    3127             :   // If we already have a scale of this value, we can add to it, otherwise, we
    3128             :   // need an available scale field.
    3129      170021 :   if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
    3130             :     return false;
    3131             : 
    3132      169657 :   ExtAddrMode TestAddrMode = AddrMode;
    3133             : 
    3134             :   // Add scale to turn X*4+X*3 -> X*7.  This could also do things like
    3135             :   // [A+B + A*7] -> [B+A*8].
    3136      169657 :   TestAddrMode.Scale += Scale;
    3137      169657 :   TestAddrMode.ScaledReg = ScaleReg;
    3138             : 
    3139             :   // If the new address isn't legal, bail out.
    3140      169657 :   if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
    3141             :     return false;
    3142             : 
    3143             :   // It was legal, so commit it.
    3144      104010 :   AddrMode = TestAddrMode;
    3145             : 
    3146             :   // Okay, we decided that we can add ScaleReg+Scale to AddrMode.  Check now
    3147             :   // to see if ScaleReg is actually X+C.  If so, we can turn this into adding
    3148             :   // X*Scale + C*Scale to addr mode.
    3149      104010 :   ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
    3150      207348 :   if (isa<Instruction>(ScaleReg) &&  // not a constant expr.
    3151      207348 :       match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
    3152        1058 :     TestAddrMode.ScaledReg = AddLHS;
    3153        2116 :     TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
    3154             : 
    3155             :     // If this addressing mode is legal, commit it and remember that we folded
    3156             :     // this instruction.
    3157        1058 :     if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
    3158         967 :       AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
    3159         967 :       AddrMode = TestAddrMode;
    3160         967 :       return true;
    3161             :     }
    3162             :   }
    3163             : 
    3164             :   // Otherwise, not (x+c)*scale, just return what we have.
    3165             :   return true;
    3166             : }
    3167             : 
    3168             : /// This is a little filter, which returns true if an addressing computation
    3169             : /// involving I might be folded into a load/store accessing it.
    3170             : /// This doesn't need to be perfect, but needs to accept at least
    3171             : /// the set of instructions that MatchOperationAddr can.
    3172       13530 : static bool MightBeFoldableInst(Instruction *I) {
    3173             :   switch (I->getOpcode()) {
    3174        4054 :   case Instruction::BitCast:
    3175             :   case Instruction::AddrSpaceCast:
    3176             :     // Don't touch identity bitcasts.
    3177        8108 :     if (I->getType() == I->getOperand(0)->getType())
    3178             :       return false;
    3179        4054 :     return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
    3180             :   case Instruction::PtrToInt:
    3181             :     // PtrToInt is always a noop, as we know that the int type is pointer sized.
    3182             :     return true;
    3183             :   case Instruction::IntToPtr:
    3184             :     // We know the input is intptr_t, so this is foldable.
    3185             :     return true;
    3186             :   case Instruction::Add:
    3187             :     return true;
    3188          30 :   case Instruction::Mul:
    3189             :   case Instruction::Shl:
    3190             :     // Can only handle X*C and X << C.
    3191          30 :     return isa<ConstantInt>(I->getOperand(1));
    3192             :   case Instruction::GetElementPtr:
    3193             :     return true;
    3194        1298 :   default:
    3195        1298 :     return false;
    3196             :   }
    3197             : }
    3198             : 
    3199             : /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
    3200             : /// \note \p Val is assumed to be the product of some type promotion.
    3201             : /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
    3202             : /// to be legal, as the non-promoted value would have had the same state.
    3203         437 : static bool isPromotedInstructionLegal(const TargetLowering &TLI,
    3204             :                                        const DataLayout &DL, Value *Val) {
    3205             :   Instruction *PromotedInst = dyn_cast<Instruction>(Val);
    3206             :   if (!PromotedInst)
    3207             :     return false;
    3208         872 :   int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
    3209             :   // If the ISDOpcode is undefined, it was undefined before the promotion.
    3210         436 :   if (!ISDOpcode)
    3211             :     return true;
    3212             :   // Otherwise, check if the promoted instruction is legal or not.
    3213         436 :   return TLI.isOperationLegalOrCustom(
    3214         436 :       ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
    3215             : }
    3216             : 
    3217             : namespace {
    3218             : 
    3219             : /// \brief Hepler class to perform type promotion.
    3220             : class TypePromotionHelper {
    3221             :   /// \brief Utility function to check whether or not a sign or zero extension
    3222             :   /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
    3223             :   /// either using the operands of \p Inst or promoting \p Inst.
    3224             :   /// The type of the extension is defined by \p IsSExt.
    3225             :   /// In other words, check if:
    3226             :   /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
    3227             :   /// #1 Promotion applies:
    3228             :   /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
    3229             :   /// #2 Operand reuses:
    3230             :   /// ext opnd1 to ConsideredExtType.
    3231             :   /// \p PromotedInsts maps the instructions to their type before promotion.
    3232             :   static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
    3233             :                             const InstrToOrigTy &PromotedInsts, bool IsSExt);
    3234             : 
    3235             :   /// \brief Utility function to determine if \p OpIdx should be promoted when
    3236             :   /// promoting \p Inst.
    3237             :   static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
    3238        1108 :     return !(isa<SelectInst>(Inst) && OpIdx == 0);
    3239             :   }
    3240             : 
    3241             :   /// \brief Utility function to promote the operand of \p Ext when this
    3242             :   /// operand is a promotable trunc or sext or zext.
    3243             :   /// \p PromotedInsts maps the instructions to their type before promotion.
    3244             :   /// \p CreatedInstsCost[out] contains the cost of all instructions
    3245             :   /// created to promote the operand of Ext.
    3246             :   /// Newly added extensions are inserted in \p Exts.
    3247             :   /// Newly added truncates are inserted in \p Truncs.
    3248             :   /// Should never be called directly.
    3249             :   /// \return The promoted value which is used instead of Ext.
    3250             :   static Value *promoteOperandForTruncAndAnyExt(
    3251             :       Instruction *Ext, TypePromotionTransaction &TPT,
    3252             :       InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
    3253             :       SmallVectorImpl<Instruction *> *Exts,
    3254             :       SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
    3255             : 
    3256             :   /// \brief Utility function to promote the operand of \p Ext when this
    3257             :   /// operand is promotable and is not a supported trunc or sext.
    3258             :   /// \p PromotedInsts maps the instructions to their type before promotion.
    3259             :   /// \p CreatedInstsCost[out] contains the cost of all the instructions
    3260             :   /// created to promote the operand of Ext.
    3261             :   /// Newly added extensions are inserted in \p Exts.
    3262             :   /// Newly added truncates are inserted in \p Truncs.
    3263             :   /// Should never be called directly.
    3264             :   /// \return The promoted value which is used instead of Ext.
    3265             :   static Value *promoteOperandForOther(Instruction *Ext,
    3266             :                                        TypePromotionTransaction &TPT,
    3267             :                                        InstrToOrigTy &PromotedInsts,
    3268             :                                        unsigned &CreatedInstsCost,
    3269             :                                        SmallVectorImpl<Instruction *> *Exts,
    3270             :                                        SmallVectorImpl<Instruction *> *Truncs,
    3271             :                                        const TargetLowering &TLI, bool IsSExt);
    3272             : 
    3273             :   /// \see promoteOperandForOther.
    3274         271 :   static Value *signExtendOperandForOther(
    3275             :       Instruction *Ext, TypePromotionTransaction &TPT,
    3276             :       InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
    3277             :       SmallVectorImpl<Instruction *> *Exts,
    3278             :       SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
    3279             :     return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
    3280         271 :                                   Exts, Truncs, TLI, true);
    3281             :   }
    3282             : 
    3283             :   /// \see promoteOperandForOther.
    3284         283 :   static Value *zeroExtendOperandForOther(
    3285             :       Instruction *Ext, TypePromotionTransaction &TPT,
    3286             :       InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
    3287             :       SmallVectorImpl<Instruction *> *Exts,
    3288             :       SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
    3289             :     return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
    3290         283 :                                   Exts, Truncs, TLI, false);
    3291             :   }
    3292             : 
    3293             : public:
    3294             :   /// Type for the utility function that promotes the operand of Ext.
    3295             :   using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
    3296             :                             InstrToOrigTy &PromotedInsts,
    3297             :                             unsigned &CreatedInstsCost,
    3298             :                             SmallVectorImpl<Instruction *> *Exts,
    3299             :                             SmallVectorImpl<Instruction *> *Truncs,
    3300             :                             const TargetLowering &TLI);
    3301             : 
    3302             :   /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
    3303             :   /// action to promote the operand of \p Ext instead of using Ext.
    3304             :   /// \return NULL if no promotable action is possible with the current
    3305             :   /// sign extension.
    3306             :   /// \p InsertedInsts keeps track of all the instructions inserted by the
    3307             :   /// other CodeGenPrepare optimizations. This information is important
    3308             :   /// because we do not want to promote these instructions as CodeGenPrepare
    3309             :   /// will reinsert them later. Thus creating an infinite loop: create/remove.
    3310             :   /// \p PromotedInsts maps the instructions to their type before promotion.
    3311             :   static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
    3312             :                           const TargetLowering &TLI,
    3313             :                           const InstrToOrigTy &PromotedInsts);
    3314             : };
    3315             : 
    3316             : } // end anonymous namespace
    3317             : 
    3318        9787 : bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
    3319             :                                         Type *ConsideredExtType,
    3320             :                                         const InstrToOrigTy &PromotedInsts,
    3321             :                                         bool IsSExt) {
    3322             :   // The promotion helper does not know how to deal with vector types yet.
    3323             :   // To be able to fix that, we would need to fix the places where we
    3324             :   // statically extend, e.g., constants and such.
    3325       19574 :   if (Inst->getType()->isVectorTy())
    3326             :     return false;
    3327             : 
    3328             :   // We can always get through zext.
    3329        6534 :   if (isa<ZExtInst>(Inst))
    3330             :     return true;
    3331             : 
    3332             :   // sext(sext) is ok too.
    3333        6439 :   if (IsSExt && isa<SExtInst>(Inst))
    3334             :     return true;
    3335             : 
    3336             :   // We can get through binary operator, if it is legal. In other words, the
    3337             :   // binary operator must have a nuw or nsw flag.
    3338             :   const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
    3339         921 :   if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
    3340         538 :       ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
    3341         383 :        (IsSExt && BinOp->hasNoSignedWrap())))
    3342             :     return true;
    3343             : 
    3344             :   // Check if we can do the following simplification.
    3345             :   // ext(trunc(opnd)) --> ext(opnd)
    3346        5878 :   if (!isa<TruncInst>(Inst))
    3347             :     return false;
    3348             : 
    3349         595 :   Value *OpndVal = Inst->getOperand(0);
    3350             :   // Check if we can use this operand in the extension.
    3351             :   // If the type is larger than the result type of the extension, we cannot.
    3352        1785 :   if (!OpndVal->getType()->isIntegerTy() ||
    3353             :       OpndVal->getType()->getIntegerBitWidth() >
    3354             :           ConsideredExtType->getIntegerBitWidth())
    3355             :     return false;
    3356             : 
    3357             :   // If the operand of the truncate is not an instruction, we will not have
    3358             :   // any information on the dropped bits.
    3359             :   // (Actually we could for constant but it is not worth the extra logic).
    3360             :   Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
    3361             :   if (!Opnd)
    3362             :     return false;
    3363             : 
    3364             :   // Check if the source of the type is narrow enough.
    3365             :   // I.e., check that trunc just drops extended bits of the same kind of
    3366             :   // the extension.
    3367             :   // #1 get the type of the operand and check the kind of the extended bits.
    3368             :   const Type *OpndType;
    3369         499 :   InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
    3370         499 :   if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
    3371             :     OpndType = It->second.getPointer();
    3372         998 :   else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
    3373          98 :     OpndType = Opnd->getOperand(0)->getType();
    3374             :   else
    3375             :     return false;
    3376             : 
    3377             :   // #2 check that the truncate just drops extended bits.
    3378          49 :   return Inst->getType()->getIntegerBitWidth() >=
    3379          49 :          OpndType->getIntegerBitWidth();
    3380             : }
    3381             : 
    3382       12839 : TypePromotionHelper::Action TypePromotionHelper::getAction(
    3383             :     Instruction *Ext, const SetOfInstrs &InsertedInsts,
    3384             :     const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
    3385             :   assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
    3386             :          "Unexpected instruction type");
    3387       12839 :   Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
    3388       12839 :   Type *ExtTy = Ext->getType();
    3389             :   bool IsSExt = isa<SExtInst>(Ext);
    3390             :   // If the operand of the extension is not an instruction, we cannot
    3391             :   // get through.
    3392             :   // If it, check we can get through.
    3393       12839 :   if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
    3394             :     return nullptr;
    3395             : 
    3396             :   // Do not promote if the operand has been added by codegenprepare.
    3397             :   // Otherwise, it means we are undoing an optimization that is likely to be
    3398             :   // redone, thus causing potential infinite loop.
    3399         703 :   if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
    3400             :     return nullptr;
    3401             : 
    3402             :   // SExt or Trunc instructions.
    3403             :   // Return the related handler.
    3404         663 :   if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
    3405             :       isa<ZExtInst>(ExtOpnd))
    3406             :     return promoteOperandForTruncAndAnyExt;
    3407             : 
    3408             :   // Regular instruction.
    3409             :   // Abort early if we will have to insert non-free instructions.
    3410         554 :   if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
    3411             :     return nullptr;
    3412         554 :   return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
    3413             : }
    3414             : 
    3415         109 : Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
    3416             :     Instruction *SExt, TypePromotionTransaction &TPT,
    3417             :     InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
    3418             :     SmallVectorImpl<Instruction *> *Exts,
    3419             :     SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
    3420             :   // By construction, the operand of SExt is an instruction. Otherwise we cannot
    3421             :   // get through it and this method should not be called.
    3422         109 :   Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
    3423             :   Value *ExtVal = SExt;
    3424             :   bool HasMergedNonFreeExt = false;
    3425         109 :   if (isa<ZExtInst>(SExtOpnd)) {
    3426             :     // Replace s|zext(zext(opnd))
    3427             :     // => zext(opnd).
    3428          95 :     HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
    3429             :     Value *ZExt =
    3430         190 :         TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
    3431          95 :     TPT.replaceAllUsesWith(SExt, ZExt);
    3432          95 :     TPT.eraseInstruction(SExt);
    3433             :     ExtVal = ZExt;
    3434             :   } else {
    3435             :     // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
    3436             :     // => z|sext(opnd).
    3437          28 :     TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
    3438             :   }
    3439         109 :   CreatedInstsCost = 0;
    3440             : 
    3441             :   // Remove dead code.
    3442         109 :   if (SExtOpnd->use_empty())
    3443          86 :     TPT.eraseInstruction(SExtOpnd);
    3444             : 
    3445             :   // Check if the extension is still needed.
    3446         109 :   Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
    3447         216 :   if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
    3448         109 :     if (ExtInst) {
    3449         107 :       if (Exts)
    3450          95 :         Exts->push_back(ExtInst);
    3451         107 :       CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
    3452             :     }
    3453             :     return ExtVal;
    3454             :   }
    3455             : 
    3456             :   // At this point we have: ext ty opnd to ty.
    3457             :   // Reassign the uses of ExtInst to the opnd and remove ExtInst.
    3458             :   Value *NextVal = ExtInst->getOperand(0);
    3459           0 :   TPT.eraseInstruction(ExtInst, NextVal);
    3460           0 :   return NextVal;
    3461             : }
    3462             : 
    3463         554 : Value *TypePromotionHelper::promoteOperandForOther(
    3464             :     Instruction *Ext, TypePromotionTransaction &TPT,
    3465             :     InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
    3466             :     SmallVectorImpl<Instruction *> *Exts,
    3467             :     SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
    3468             :     bool IsSExt) {
    3469             :   // By construction, the operand of Ext is an instruction. Otherwise we cannot
    3470             :   // get through it and this method should not be called.
    3471         554 :   Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
    3472         554 :   CreatedInstsCost = 0;
    3473         554 :   if (!ExtOpnd->hasOneUse()) {
    3474             :     // ExtOpnd will be promoted.
    3475             :     // All its uses, but Ext, will need to use a truncated value of the
    3476             :     // promoted version.
    3477             :     // Create the truncate now.
    3478          46 :     Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
    3479          46 :     if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
    3480             :       // Insert it just after the definition.
    3481          46 :       ITrunc->moveAfter(ExtOpnd);
    3482          46 :       if (Truncs)
    3483           0 :         Truncs->push_back(ITrunc);
    3484             :     }
    3485             : 
    3486          46 :     TPT.replaceAllUsesWith(ExtOpnd, Trunc);
    3487             :     // Restore the operand of Ext (which has been replaced by the previous call
    3488             :     // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
    3489          46 :     TPT.setOperand(Ext, 0, ExtOpnd);
    3490             :   }
    3491             : 
    3492             :   // Get through the Instruction:
    3493             :   // 1. Update its type.
    3494             :   // 2. Replace the uses of Ext by Inst.
    3495             :   // 3. Extend each operand that needs to be extended.
    3496             : 
    3497             :   // Remember the original type of the instruction before promotion.
    3498             :   // This is useful to know that the high bits are sign extended bits.
    3499         554 :   PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
    3500         554 :       ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
    3501             :   // Step #1.
    3502         554 :   TPT.mutateType(ExtOpnd, Ext->getType());
    3503             :   // Step #2.
    3504         554 :   TPT.replaceAllUsesWith(Ext, ExtOpnd);
    3505             :   // Step #3.
    3506         554 :   Instruction *ExtForOpnd = Ext;
    3507             : 
    3508             :   DEBUG(dbgs() << "Propagate Ext to operands\n");
    3509        1662 :   for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
    3510             :        ++OpIdx) {
    3511             :     DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
    3512        2216 :     if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
    3513             :         !shouldExtOperand(ExtOpnd, OpIdx)) {
    3514             :       DEBUG(dbgs() << "No need to propagate\n");
    3515           0 :       continue;
    3516             :     }
    3517             :     // Check if we can statically extend the operand.
    3518             :     Value *Opnd = ExtOpnd->getOperand(OpIdx);
    3519             :     if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
    3520             :       DEBUG(dbgs() << "Statically extend\n");
    3521             :       unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
    3522             :       APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
    3523         872 :                             : Cst->getValue().zext(BitWidth);
    3524         436 :       TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
    3525             :       continue;
    3526             :     }
    3527             :     // UndefValue are typed, so we have to statically sign extend them.
    3528         672 :     if (isa<UndefValue>(Opnd)) {
    3529             :       DEBUG(dbgs() << "Statically extend\n");
    3530           0 :       TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
    3531           0 :       continue;
    3532             :     }
    3533             : 
    3534             :     // Otherwise we have to explicity sign extend the operand.
    3535             :     // Check if Ext was reused to extend an operand.
    3536         672 :     if (!ExtForOpnd) {
    3537             :       // If yes, create a new one.
    3538             :       DEBUG(dbgs() << "More operands to ext\n");
    3539         118 :       Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
    3540             :         : TPT.createZExt(Ext, Opnd, Ext->getType());
    3541         124 :       if (!isa<Instruction>(ValForExtOpnd)) {
    3542           6 :         TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
    3543             :         continue;
    3544             :       }
    3545         112 :       ExtForOpnd = cast<Instruction>(ValForExtOpnd);
    3546             :     }
    3547         666 :     if (Exts)
    3548         447 :       Exts->push_back(ExtForOpnd);
    3549         666 :     TPT.setOperand(ExtForOpnd, 0, Opnd);
    3550             : 
    3551             :     // Move the sign extension before the insertion point.
    3552         666 :     TPT.moveBefore(ExtForOpnd, ExtOpnd);
    3553         666 :     TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
    3554         666 :     CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
    3555             :     // If more sext are required, new instructions will have to be created.
    3556         666 :     ExtForOpnd = nullptr;
    3557             :   }
    3558         554 :   if (ExtForOpnd == Ext) {
    3559             :     DEBUG(dbgs() << "Extension is useless now\n");
    3560           0 :     TPT.eraseInstruction(Ext);
    3561             :   }
    3562         554 :   return ExtOpnd;
    3563             : }
    3564             : 
    3565             : /// Check whether or not promoting an instruction to a wider type is profitable.
    3566             : /// \p NewCost gives the cost of extension instructions created by the
    3567             : /// promotion.
    3568             : /// \p OldCost gives the cost of extension instructions before the promotion
    3569             : /// plus the number of instructions that have been
    3570             : /// matched in the addressing mode the promotion.
    3571             : /// \p PromotedOperand is the value that has been promoted.
    3572             : /// \return True if the promotion is profitable, false otherwise.
    3573             : bool AddressingModeMatcher::isPromotionProfitable(
    3574             :     unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
    3575             :   DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
    3576             :   // The cost of the new extensions is greater than the cost of the
    3577             :   // old extension plus what we folded.
    3578             :   // This is not profitable.
    3579         159 :   if (NewCost > OldCost)
    3580             :     return false;
    3581         148 :   if (NewCost < OldCost)
    3582             :     return true;
    3583             :   // The promotion is neutral but it may help folding the sign extension in
    3584             :   // loads for instance.
    3585             :   // Check that we did not create an illegal instruction.
    3586          58 :   return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
    3587             : }
    3588             : 
    3589             : /// Given an instruction or constant expr, see if we can fold the operation
    3590             : /// into the addressing mode. If so, update the addressing mode and return
    3591             : /// true, otherwise return false without modifying AddrMode.
    3592             : /// If \p MovedAway is not NULL, it contains the information of whether or
    3593             : /// not AddrInst has to be folded into the addressing mode on success.
    3594             : /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
    3595             : /// because it has been moved away.
    3596             : /// Thus AddrInst must not be added in the matched instructions.
    3597             : /// This state can happen when AddrInst is a sext, since it may be moved away.
    3598             : /// Therefore, AddrInst may not be valid when MovedAway is true and it must
    3599             : /// not be referenced anymore.
    3600     1072957 : bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
    3601             :                                                unsigned Depth,
    3602             :                                                bool *MovedAway) {
    3603             :   // Avoid exponential behavior on extremely deep expression trees.
    3604     1072957 :   if (Depth >= 5) return false;
    3605             : 
    3606             :   // By default, all matched instructions stay in place.
    3607     1072905 :   if (MovedAway)
    3608      434689 :     *MovedAway = false;
    3609             : 
    3610     1072905 :   switch (Opcode) {
    3611             :   case Instruction::PtrToInt:
    3612             :     // PtrToInt is always a noop, as we know that the int type is pointer sized.
    3613         668 :     return matchAddr(AddrInst->getOperand(0), Depth);
    3614        3379 :   case Instruction::IntToPtr: {
    3615        3379 :     auto AS = AddrInst->getType()->getPointerAddressSpace();
    3616        6758 :     auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
    3617             :     // This inttoptr is a no-op if the integer type is pointer sized.
    3618       10137 :     if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
    3619        3339 :       return matchAddr(AddrInst->getOperand(0), Depth);
    3620             :     return false;
    3621             :   }
    3622             :   case Instruction::BitCast:
    3623             :     // BitCast is always a noop, and we can handle it as long as it is
    3624             :     // int->int or pointer->pointer (we don't want int<->fp or something).
    3625      364021 :     if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
    3626      727772 :          AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
    3627             :         // Don't touch identity bitcasts.  These were probably put here by LSR,
    3628             :         // and we don't want to mess around with them.  Assume it knows what it
    3629             :         // is doing.
    3630      363884 :         AddrInst->getOperand(0)->getType() != AddrInst->getType())
    3631      362486 :       return matchAddr(AddrInst->getOperand(0), Depth);
    3632             :     return false;
    3633             :   case Instruction::AddrSpaceCast: {
    3634             :     unsigned SrcAS
    3635         296 :       = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
    3636         296 :     unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
    3637         296 :     if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
    3638          93 :       return matchAddr(AddrInst->getOperand(0), Depth);
    3639             :     return false;
    3640             :   }
    3641        2435 :   case Instruction::Add: {
    3642             :     // Check to see if we can merge in the RHS then the LHS.  If so, we win.
    3643        2435 :     ExtAddrMode BackupAddrMode = AddrMode;
    3644        2435 :     unsigned OldSize = AddrModeInsts.size();
    3645             :     // Start a transaction at this point.
    3646             :     // The LHS may match but not the RHS.
    3647             :     // Therefore, we need a higher level restoration point to undo partially
    3648             :     // matched operation.
    3649             :     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
    3650        2435 :         TPT.getRestorationPoint();
    3651             : 
    3652        7270 :     if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
    3653        2400 :         matchAddr(AddrInst->getOperand(0), Depth+1))
    3654             :       return true;
    3655             : 
    3656             :     // Restore the old addr mode info.
    3657         491 :     AddrMode = BackupAddrMode;
    3658         491 :     AddrModeInsts.resize(OldSize);
    3659         491 :     TPT.rollback(LastKnownGood);
    3660             : 
    3661             :     // Otherwise this was over-aggressive.  Try merging in the LHS then the RHS.
    3662         862 :     if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
    3663         371 :         matchAddr(AddrInst->getOperand(1), Depth+1))
    3664             :       return true;
    3665             : 
    3666             :     // Otherwise we definitely can't merge the ADD in.
    3667         489 :     AddrMode = BackupAddrMode;
    3668         489 :     AddrModeInsts.resize(OldSize);
    3669         489 :     TPT.rollback(LastKnownGood);
    3670             :     break;
    3671             :   }
    3672             :   //case Instruction::Or:
    3673             :   // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
    3674             :   //break;
    3675             :   case Instruction::Mul:
    3676             :   case Instruction::Shl: {
    3677             :     // Can only handle X*C and X << C.
    3678             :     ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
    3679        2414 :     if (!RHS || RHS->getBitWidth() > 64)
    3680             :       return false;
    3681             :     int64_t Scale = RHS->getSExtValue();
    3682        2413 :     if (Opcode == Instruction::Shl)
    3683        1374 :       Scale = 1LL << Scale;
    3684             : 
    3685        2413 :     return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
    3686             :   }
    3687      585691 :   case Instruction::GetElementPtr: {
    3688             :     // Scan the GEP.  We check it if it contains constant offsets and at most
    3689             :     // one variable offset.
    3690             :     int VariableOperand = -1;
    3691             :     unsigned VariableScale = 0;
    3692             : 
    3693             :     int64_t ConstantOffset = 0;
    3694      585691 :     gep_type_iterator GTI = gep_type_begin(AddrInst);
    3695     2890455 :     for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
    3696       85569 :       if (StructType *STy = GTI.getStructTypeOrNull()) {
    3697       85569 :         const StructLayout *SL = DL.getStructLayout(STy);
    3698             :         unsigned Idx =
    3699       85569 :           cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
    3700       85569 :         ConstantOffset += SL->getElementOffset(Idx);
    3701             :       } else {
    3702     1067768 :         uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
    3703             :         if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
    3704      940167 :           ConstantOffset += CI->getSExtValue() * TypeSize;
    3705      127601 :         } else if (TypeSize) {  // Scales of zero don't do anything.
    3706             :           // We only allow one variable index at the moment.
    3707      127595 :           if (VariableOperand != -1)
    3708             :             return false;
    3709             : 
    3710             :           // Remember the variable index.
    3711      126640 :           VariableOperand = i;
    3712      126640 :           VariableScale = TypeSize;
    3713             :         }
    3714             :       }
    3715             :     }
    3716             : 
    3717             :     // A common case is for the GEP to only do a constant offset.  In this case,
    3718             :     // just add it to the disp field and check validity.
    3719      584736 :     if (VariableOperand == -1) {
    3720      459051 :       AddrMode.BaseOffs += ConstantOffset;
    3721      843429 :       if (ConstantOffset == 0 ||
    3722      384378 :           TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
    3723             :         // Check to see if we can fold the base pointer in too.
    3724      911976 :         if (matchAddr(AddrInst->getOperand(0), Depth+1))
    3725             :           return true;
    3726             :       }
    3727        3063 :       AddrMode.BaseOffs -= ConstantOffset;
    3728        3063 :       return false;
    3729             :     }
    3730             : 
    3731             :     // Save the valid addressing mode in case we can't match.
    3732      125685 :     ExtAddrMode BackupAddrMode = AddrMode;
    3733      125685 :     unsigned OldSize = AddrModeInsts.size();
    3734             : 
    3735             :     // See if the scale and offset amount is valid for this target.
    3736      125685 :     AddrMode.BaseOffs += ConstantOffset;
    3737             : 
    3738             :     // Match the base operand of the GEP.
    3739      251370 :     if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
    3740             :       // If it couldn't be matched, just stuff the value in a register.
    3741           5 :       if (AddrMode.HasBaseReg) {
    3742           0 :         AddrMode = BackupAddrMode;
    3743           0 :         AddrModeInsts.resize(OldSize);
    3744           0 :         return false;
    3745             :       }
    3746           5 :       AddrMode.HasBaseReg = true;
    3747           5 :       AddrMode.BaseReg = AddrInst->getOperand(0);
    3748             :     }
    3749             : 
    3750             :     // Match the remaining variable portion of the GEP.
    3751      251370 :     if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
    3752             :                           Depth)) {
    3753             :       // If it couldn't be matched, try stuffing the base into a register
    3754             :       // instead of matching it, and retrying the match of the scale.
    3755       51937 :       AddrMode = BackupAddrMode;
    3756       51937 :       AddrModeInsts.resize(OldSize);
    3757       51937 :       if (AddrMode.HasBaseReg)
    3758             :         return false;
    3759       51937 :       AddrMode.HasBaseReg = true;
    3760       51937 :       AddrMode.BaseReg = AddrInst->getOperand(0);
    3761       51937 :       AddrMode.BaseOffs += ConstantOffset;
    3762       51937 :       if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
    3763             :                             VariableScale, Depth)) {
    3764             :         // If even that didn't work, bail.
    3765       14672 :         AddrMode = BackupAddrMode;
    3766       14672 :         AddrModeInsts.resize(OldSize);
    3767       14672 :         return false;
    3768             :       }
    3769             :     }
    3770             : 
    3771             :     return true;
    3772             :   }
    3773             :   case Instruction::SExt:
    3774             :   case Instruction::ZExt: {
    3775             :     Instruction *Ext = dyn_cast<Instruction>(AddrInst);
    3776             :     if (!Ext)
    3777             :       return false;
    3778             : 
    3779             :     // Try to move this ext out of the way of the addressing mode.
    3780             :     // Ask for a method for doing so.
    3781             :     TypePromotionHelper::Action TPH =
    3782         996 :         TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
    3783         996 :     if (!TPH)
    3784             :       return false;
    3785             : 
    3786             :     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
    3787         215 :         TPT.getRestorationPoint();
    3788         215 :     unsigned CreatedInstsCost = 0;
    3789         215 :     unsigned ExtCost = !TLI.isExtFree(Ext);
    3790             :     Value *PromotedOperand =
    3791         215 :         TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
    3792             :     // SExt has been moved away.
    3793             :     // Thus either it will be rematched later in the recursive calls or it is
    3794             :     // gone. Anyway, we must not fold it into the addressing mode at this point.
    3795             :     // E.g.,
    3796             :     // op = add opnd, 1
    3797             :     // idx = ext op
    3798             :     // addr = gep base, idx
    3799             :     // is now:
    3800             :     // promotedOpnd = ext opnd            <- no match here
    3801             :     // op = promoted_add promotedOpnd, 1  <- match (later in recursive calls)
    3802             :     // addr = gep base, op                <- match
    3803         215 :     if (MovedAway)
    3804         215 :       *MovedAway = true;
    3805             : 
    3806             :     assert(PromotedOperand &&
    3807             :            "TypePromotionHelper should have filtered out those cases");
    3808             : 
    3809         215 :     ExtAddrMode BackupAddrMode = AddrMode;
    3810         430 :     unsigned OldSize = AddrModeInsts.size();
    3811             : 
    3812         273 :     if (!matchAddr(PromotedOperand, Depth) ||
    3813             :         // The total of the new cost is equal to the cost of the created
    3814             :         // instructions.
    3815             :         // The total of the old cost is equal to the cost of the extension plus
    3816             :         // what we have saved in the addressing mode.
    3817         159 :         !isPromotionProfitable(CreatedInstsCost,
    3818         318 :                                ExtCost + (AddrModeInsts.size() - OldSize),
    3819             :                                PromotedOperand)) {
    3820          97 :       AddrMode = BackupAddrMode;
    3821          97 :       AddrModeInsts.resize(OldSize);
    3822             :       DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
    3823          97 :       TPT.rollback(LastKnownGood);
    3824          97 :       return false;
    3825             :     }
    3826             :     return true;
    3827             :   }
    3828             :   }
    3829             :   return false;
    3830             : }
    3831             : 
    3832             : /// If we can, try to add the value of 'Addr' into the current addressing mode.
    3833             : /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
    3834             : /// unmodified. This assumes that Addr is either a pointer type or intptr_t
    3835             : /// for the target.
    3836             : ///
    3837     2017254 : bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
    3838             :   // Start a transaction at this point that we will rollback if the matching
    3839             :   // fails.
    3840             :   TypePromotionTransaction::ConstRestorationPt LastKnownGood =
    3841     2017254 :       TPT.getRestorationPoint();
    3842             :   if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
    3843             :     // Fold in immediates if legal for the target.
    3844        1977 :     AddrMode.BaseOffs += CI->getSExtValue();
    3845        1977 :     if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
    3846             :       return true;
    3847         514 :     AddrMode.BaseOffs -= CI->getSExtValue();
    3848             :   } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
    3849             :     // If this is a global variable, try to fold it into the addressing mode.
    3850      817074 :     if (!AddrMode.BaseGV) {
    3851      817074 :       AddrMode.BaseGV = GV;
    3852      817074 :       if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
    3853             :         return true;
    3854      169829 :       AddrMode.BaseGV = nullptr;
    3855             :     }
    3856     1198203 :   } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
    3857      434741 :     ExtAddrMode BackupAddrMode = AddrMode;
    3858      434741 :     unsigned OldSize = AddrModeInsts.size();
    3859             : 
    3860             :     // Check to see if it is possible to fold this operation.
    3861      434741 :     bool MovedAway = false;
    3862      434741 :     if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
    3863             :       // This instruction may have been moved away. If so, there is nothing
    3864             :       // to check here.
    3865      298723 :       if (MovedAway)
    3866      296527 :         return true;
    3867             :       // Okay, it's possible to fold this.  Check to see if it is actually
    3868             :       // *profitable* to do so.  We use a simple cost model to avoid increasing
    3869             :       // register pressure too much.
    3870      736714 :       if (I->hasOneUse() ||
    3871      139504 :           isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
    3872      296409 :         AddrModeInsts.push_back(I);
    3873      296409 :         return true;
    3874             :       }
    3875             : 
    3876             :       // It isn't profitable to do this, roll back.
    3877             :       //cerr << "NOT FOLDING: " << *I;
    3878        2196 :       AddrMode = BackupAddrMode;
    3879        2196 :       AddrModeInsts.resize(OldSize);
    3880        2196 :       TPT.rollback(LastKnownGood);
    3881             :     }
    3882             :   } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
    3883      638216 :     if (matchOperationAddr(CE, CE->getOpcode(), Depth))
    3884             :       return true;
    3885         192 :     TPT.rollback(LastKnownGood);
    3886      125246 :   } else if (isa<ConstantPointerNull>(Addr)) {
    3887             :     // Null pointer gets folded without affecting the addressing mode.
    3888             :     return true;
    3889             :   }
    3890             : 
    3891             :   // Worse case, the target should support [reg] addressing modes. :)
    3892      432832 :   if (!AddrMode.HasBaseReg) {
    3893      421923 :     AddrMode.HasBaseReg = true;
    3894      421923 :     AddrMode.BaseReg = Addr;
    3895             :     // Still check for legality in case the target supports [imm] but not [i+r].
    3896      421923 :     if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
    3897             :       return true;
    3898          10 :     AddrMode.HasBaseReg = false;
    3899          10 :     AddrMode.BaseReg = nullptr;
    3900             :   }
    3901             : 
    3902             :   // If the base register is already taken, see if we can do [r+r].
    3903       10919 :   if (AddrMode.Scale == 0) {
    3904       10350 :     AddrMode.Scale = 1;
    3905       10350 :     AddrMode.ScaledReg = Addr;
    3906       10350 :     if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
    3907             :       return true;
    3908        2387 :     AddrMode.Scale = 0;
    3909        2387 :     AddrMode.ScaledReg = nullptr;
    3910             :   }
    3911             :   // Couldn't match.
    3912        2956 :   TPT.rollback(LastKnownGood);
    3913        2956 :   return false;
    3914             : }
    3915             : 
    3916             : /// Check to see if all uses of OpVal by the specified inline asm call are due
    3917             : /// to memory operands. If so, return true, otherwise return false.
    3918          14 : static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
    3919             :                                     const TargetLowering &TLI,
    3920             :                                     const TargetRegisterInfo &TRI) {
    3921          14 :   const Function *F = CI->getFunction();
    3922             :   TargetLowering::AsmOperandInfoVector TargetConstraints =
    3923             :       TLI.ParseConstraints(F->getParent()->getDataLayout(), &TRI,
    3924          42 :                             ImmutableCallSite(CI));
    3925             : 
    3926          82 :   for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
    3927          66 :     TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
    3928             : 
    3929             :     // Compute the constraint code and ConstraintType to use.
    3930          66 :     TLI.ComputeConstraintToUse(OpInfo, SDValue());
    3931             : 
    3932             :     // If this asm operand is our Value*, and if it isn't an indirect memory
    3933             :     // operand, we can't fold it!
    3934          94 :     if (OpInfo.CallOperandVal == OpVal &&
    3935          44 :         (OpInfo.ConstraintType != TargetLowering::C_Memory ||
    3936          16 :          !OpInfo.isIndirect))
    3937             :       return false;
    3938             :   }
    3939             : 
    3940             :   return true;
    3941             : }
    3942             : 
    3943             : // Max number of memory uses to look at before aborting the search to conserve
    3944             : // compile time.
    3945             : static constexpr int MaxMemoryUsesToScan = 20;
    3946             : 
    3947             : /// Recursively walk all the uses of I until we find a memory use.
    3948             : /// If we find an obviously non-foldable instruction, return true.
    3949             : /// Add the ultimately found memory instructions to MemoryUses.
    3950       13530 : static bool FindAllMemoryUses(
    3951             :     Instruction *I,
    3952             :     SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
    3953             :     SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
    3954             :     const TargetRegisterInfo &TRI, int SeenInsts = 0) {
    3955             :   // If we already considered this instruction, we're done.
    3956       13530 :   if (!ConsideredInsts.insert(I).second)
    3957             :     return false;
    3958             : 
    3959             :   // If this is an obviously unfoldable instruction, bail out.
    3960       13530 :   if (!MightBeFoldableInst(I))
    3961             :     return true;
    3962             : 
    3963       12232 :   const bool OptSize = I->getFunction()->optForSize();
    3964             : 
    3965             :   // Loop over all the uses, recursively processing them.
    3966       29023 :   for (Use &U : I->uses()) {
    3967             :     // Conservatively return true if we're seeing a large number or a deep chain
    3968             :     // of users. This avoids excessive compilation times in pathological cases.
    3969       19945 :     if (SeenInsts++ >= MaxMemoryUsesToScan)
    3970             :       return true;
    3971             : 
    3972       19865 :     Instruction *UserI = cast<Instruction>(U.getUser());
    3973        6153 :     if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
    3974       12306 :       MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
    3975             :       continue;
    3976             :     }
    3977             : 
    3978             :     if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
    3979        3195 :       unsigned opNo = U.getOperandNo();
    3980        3195 :       if (opNo != StoreInst::getPointerOperandIndex())
    3981             :         return true; // Storing addr, not into addr.
    3982        6350 :       MemoryUses.push_back(std::make_pair(SI, opNo));
    3983        3175 :       continue;
    3984             :     }
    3985             : 
    3986             :     if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
    3987          52 :       unsigned opNo = U.getOperandNo();
    3988          52 :       if (opNo != AtomicRMWInst::getPointerOperandIndex())
    3989             :         return true; // Storing addr, not into addr.
    3990         104 :       MemoryUses.push_back(std::make_pair(RMW, opNo));
    3991          52 :       continue;
    3992             :     }
    3993             : 
    3994             :     if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
    3995           2 :       unsigned opNo = U.getOperandNo();
    3996           2 :       if (opNo != AtomicCmpXchgInst::getPointerOperandIndex())
    3997             :         return true; // Storing addr, not into addr.
    3998           4 :       MemoryUses.push_back(std::make_pair(CmpX, opNo));
    3999           2 :       continue;
    4000             :     }
    4001             : 
    4002             :     if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
    4003             :       // If this is a cold call, we can sink the addressing calculation into
    4004             :       // the cold path.  See optimizeCallInst
    4005        1109 :       if (!OptSize && CI->hasFnAttr(Attribute::Cold))
    4006           7 :         continue;
    4007             : 
    4008             :       InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
    4009             :       if (!IA) return true;
    4010             : 
    4011             :       // If this is a memory operand, we're cool, otherwise bail out.
    4012          14 :       if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI))
    4013             :         return true;
    4014           2 :       continue;
    4015             :     }
    4016             : 
    4017        9911 :     if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI,
    4018             :                           SeenInsts))
    4019             :       return true;
    4020             :   }
    4021             : 
    4022             :   return false;
    4023             : }
    4024             : 
    4025             : /// Return true if Val is already known to be live at the use site that we're
    4026             : /// folding it into. If so, there is no cost to include it in the addressing
    4027             : /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
    4028             : /// instruction already.
    4029      269006 : bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
    4030             :                                                    Value *KnownLive2) {
    4031             :   // If Val is either of the known-live values, we know it is live!
    4032      269006 :   if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
    4033             :     return true;
    4034             : 
    4035             :   // All values other than instructions and arguments (e.g. constants) are live.
    4036      175216 :   if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
    4037             : 
    4038             :   // If Val is a constant sized alloca in the entry block, it is live, this is
    4039             :   // true because it is just a reference to the stack/frame pointer, which is
    4040             :   // live for the whole function.
    4041             :   if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
    4042       17872 :     if (AI->isStaticAlloca())
    4043             :       return true;
    4044             : 
    4045             :   // Check to see if this value is already used in the memory instruction's
    4046             :   // block.  If so, it's already live into the block at the very least, so we
    4047             :   // can reasonably fold it.
    4048      120114 :   return Val->isUsedInBasicBlock(MemoryInst->getParent());
    4049             : }
    4050             : 
    4051             : /// It is possible for the addressing mode of the machine to fold the specified
    4052             : /// instruction into a load or store that ultimately uses it.
    4053             : /// However, the specified instruction has multiple uses.
    4054             : /// Given this, it may actually increase register pressure to fold it
    4055             : /// into the load. For example, consider this code:
    4056             : ///
    4057             : ///     X = ...
    4058             : ///     Y = X+1
    4059             : ///     use(Y)   -> nonload/store
    4060             : ///     Z = Y+1
    4061             : ///     load Z
    4062             : ///
    4063             : /// In this case, Y has multiple uses, and can be folded into the load of Z
    4064             : /// (yielding load [X+2]).  However, doing this will cause both "X" and "X+1" to
    4065             : /// be live at the use(Y) line.  If we don't fold Y into load Z, we use one
    4066             : /// fewer register.  Since Y can't be folded into "use(Y)" we don't increase the
    4067             : /// number of computations either.
    4068             : ///
    4069             : /// Note that this (like most of CodeGenPrepare) is just a rough heuristic.  If
    4070             : /// X was live across 'load Z' for other reasons, we actually *would* want to
    4071             : /// fold the addressing mode in the Z case.  This would make Y die earlier.
    4072      139504 : bool AddressingModeMatcher::
    4073             : isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
    4074             :                                      ExtAddrMode &AMAfter) {
    4075      139504 :   if (IgnoreProfitability) return true;
    4076             : 
    4077             :   // AMBefore is the addressing mode before this instruction was folded into it,
    4078             :   // and AMAfter is the addressing mode after the instruction was folded.  Get
    4079             :   // the set of registers referenced by AMAfter and subtract out those
    4080             :   // referenced by AMBefore: this is the set of values which folding in this
    4081             :   // address extends the lifetime of.
    4082             :   //
    4083             :   // Note that there are only two potential values being referenced here,
    4084             :   // BaseReg and ScaleReg (global addresses are always available, as are any
    4085             :   // folded immediates).
    4086      134503 :   Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
    4087             : 
    4088             :   // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
    4089             :   // lifetime wasn't extended by adding this instruction.
    4090      134503 :   if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
    4091             :     BaseReg = nullptr;
    4092      134503 :   if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
    4093             :     ScaledReg = nullptr;
    4094             : 
    4095             :   // If folding this instruction (and it's subexprs) didn't extend any live
    4096             :   // ranges, we're ok with it.
    4097      134503 :   if (!BaseReg && !ScaledReg)
    4098             :     return true;
    4099             : 
    4100             :   // If all uses of this instruction can have the address mode sunk into them,
    4101             :   // we can remove the addressing mode and effectively trade one live register
    4102             :   // for another (at worst.)  In this context, folding an addressing mode into
    4103             :   // the use is just a particularly nice way of sinking it.
    4104             :   SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
    4105             :   SmallPtrSet<Instruction*, 16> ConsideredInsts;
    4106        3619 :   if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI))
    4107             :     return false;  // Has a non-memory, non-foldable use!
    4108             : 
    4109             :   // Now that we know that all uses of this instruction are part of a chain of
    4110             :   // computation involving only operations that could theoretically be folded
    4111             :   // into a memory use, loop over each of these memory operation uses and see
    4112             :   // if they could  *actually* fold the instruction.  The assumption is that
    4113             :   // addressing modes are cheap and that duplicating the computation involved
    4114             :   // many times is worthwhile, even on a fastpath. For sinking candidates
    4115             :   // (i.e. cold call sites), this serves as a way to prevent excessive code
    4116             :   // growth since most architectures have some reasonable small and fast way to
    4117             :   // compute an effective address.  (i.e LEA on x86)
    4118             :   SmallVector<Instruction*, 32> MatchedAddrModeInsts;
    4119        6238 :   for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
    4120        9630 :     Instruction *User = MemoryUses[i].first;
    4121        4815 :     unsigned OpNo = MemoryUses[i].second;
    4122             : 
    4123             :     // Get the access type of this use.  If the use isn't a pointer, we don't
    4124             :     // know what it accesses.
    4125        4815 :     Value *Address = User->getOperand(OpNo);
    4126        4815 :     PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
    4127             :     if (!AddrTy)
    4128         255 :       return false;
    4129        4815 :     Type *AddressAccessTy = AddrTy->getElementType();
    4130             :     unsigned AS = AddrTy->getAddressSpace();
    4131             : 
    4132             :     // Do a match against the root of this address, ignoring profitability. This
    4133             :     // will tell us if the addressing mode for the memory operation will
    4134             :     // *actually* cover the shared instruction.
    4135        4815 :     ExtAddrMode Result;
    4136             :     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
    4137        4815 :         TPT.getRestorationPoint();
    4138             :     AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, TRI,
    4139             :                                   AddressAccessTy, AS,
    4140             :                                   MemoryInst, Result, InsertedInsts,
    4141        4815 :                                   PromotedInsts, TPT);
    4142        4815 :     Matcher.IgnoreProfitability = true;
    4143        4815 :     bool Success = Matcher.matchAddr(Address, 0);
    4144             :     (void)Success; assert(Success && "Couldn't select *anything*?");
    4145             : 
    4146             :     // The match was to check the profitability, the changes made are not
    4147             :     // part of the original matcher. Therefore, they should be dropped
    4148             :     // otherwise the original matcher will not present the right state.
    4149        4815 :     TPT.rollback(LastKnownGood);
    4150             : 
    4151             :     // If the match didn't cover I, then it won't be shared by it.
    4152        4815 :     if (!is_contained(MatchedAddrModeInsts, I))
    4153             :       return false;
    4154             : 
    4155             :     MatchedAddrModeInsts.clear();
    4156             :   }
    4157             : 
    4158             :   return true;
    4159             : }
    4160             : 
    4161             : /// Return true if the specified values are defined in a
    4162             : /// different basic block than BB.
    4163             : static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
    4164             :   if (Instruction *I = dyn_cast<Instruction>(V))
    4165      270805 :     return I->getParent() != BB;
    4166             :   return false;
    4167             : }
    4168             : 
    4169             : /// Sink addressing mode computation immediate before MemoryInst if doing so
    4170             : /// can be done without increasing register pressure.  The need for the
    4171             : /// register pressure constraint means this can end up being an all or nothing
    4172             : /// decision for all uses of the same addressing computation.
    4173             : ///
    4174             : /// Load and Store Instructions often have addressing modes that can do
    4175             : /// significant amounts of computation. As such, instruction selection will try
    4176             : /// to get the load or store to do as much computation as possible for the
    4177             : /// program. The problem is that isel can only see within a single block. As
    4178             : /// such, we sink as much legal addressing mode work into the block as possible.
    4179             : ///
    4180             : /// This method is used to optimize both load/store and inline asms with memory
    4181             : /// operands.  It's also used to sink addressing computations feeding into cold
    4182             : /// call sites into their (cold) basic block.
    4183             : ///
    4184             : /// The motivation for handling sinking into cold blocks is that doing so can
    4185             : /// both enable other address mode sinking (by satisfying the register pressure
    4186             : /// constraint above), and reduce register pressure globally (by removing the
    4187             : /// addressing mode computation from the fast path entirely.).
    4188     1041123 : bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
    4189             :                                         Type *AccessTy, unsigned AddrSpace) {
    4190     1041123 :   Value *Repl = Addr;
    4191             : 
    4192             :   // Try to collapse single-value PHI nodes.  This is necessary to undo
    4193             :   // unprofitable PRE transformations.
    4194             :   SmallVector<Value*, 8> worklist;
    4195             :   SmallPtrSet<Value*, 16> Visited;
    4196     1041123 :   worklist.push_back(Addr);
    4197             : 
    4198             :   // Use a worklist to iteratively look through PHI and select nodes, and
    4199             :   // ensure that the addressing mode obtained from the non-PHI/select roots of
    4200             :   // the graph are compatible.
    4201             :   bool PhiOrSelectSeen = false;
    4202             :   SmallVector<Instruction*, 16> AddrModeInsts;
    4203     1041123 :   const SimplifyQuery SQ(*DL, TLInfo);
    4204     1041123 :   AddressingModeCombiner AddrModes(SQ, { Addr, MemoryInst->getParent() });
    4205     1041123 :   TypePromotionTransaction TPT(RemovedInsts);
    4206             :   TypePromotionTransaction::ConstRestorationPt LastKnownGood =
    4207             :       TPT.getRestorationPoint();
    4208     2096407 :   while (!worklist.empty()) {
    4209     1056832 :     Value *V = worklist.back();
    4210             :     worklist.pop_back();
    4211             : 
    4212             :     // We allow traversing cyclic Phi nodes.
    4213             :     // In case of success after this loop we ensure that traversing through
    4214             :     // Phi nodes ends up with all cases to compute address of the form
    4215             :     //    BaseGV + Base + Scale * Index + Offset
    4216             :     // where Scale and Offset are constans and BaseGV, Base and Index
    4217             :     // are exactly the same Values in all cases.
    4218             :     // It means that BaseGV, Scale and Offset dominate our memory instruction
    4219             :     // and have the same value as they had in address computation represented
    4220             :     // as Phi. So we can safely sink address computation to memory instruction.
    4221     1056832 :     if (!Visited.insert(V).second)
    4222        9454 :       continue;
    4223             : 
    4224             :     // For a PHI node, push all of its incoming values.
    4225             :     if (PHINode *P = dyn_cast<PHINode>(V)) {
    4226       38246 :       for (Value *IncValue : P->incoming_values())
    4227       15496 :         worklist.push_back(IncValue);
    4228             :       PhiOrSelectSeen = true;
    4229        7254 :       continue;
    4230             :     }
    4231             :     // Similar for select.
    4232         440 :     if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
    4233         440 :       worklist.push_back(SI->getFalseValue());
    4234         440 :       worklist.push_back(SI->getTrueValue());
    4235             :       PhiOrSelectSeen = true;
    4236         440 :       continue;
    4237             :     }
    4238             : 
    4239             :     // For non-PHIs, determine the addressing mode being computed.  Note that
    4240             :     // the result may differ depending on what other uses our candidate
    4241             :     // addressing instructions might have.
    4242             :     AddrModeInsts.clear();
    4243             :     ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
    4244     1048258 :         V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *TRI,
    4245     2096516 :         InsertedInsts, PromotedInsts, TPT);
    4246     1048258 :     NewAddrMode.OriginalValue = V;
    4247             : 
    4248     1048258 :     if (!AddrModes.addNewAddrMode(NewAddrMode))
    4249             :       break;
    4250             :   }
    4251             : 
    4252             :   // Try to combine the AddrModes we've collected. If we couldn't collect any,
    4253             :   // or we have multiple but either couldn't combine them or combining them
    4254             :   // wouldn't do anything useful, bail out now.
    4255     1041123 :   if (!AddrModes.combineAddrModes()) {
    4256        3074 :     TPT.rollback(LastKnownGood);
    4257        3074 :     return false;
    4258             :   }
    4259             :   TPT.commit();
    4260             : 
    4261             :   // Get the combined AddrMode (or the only AddrMode, if we only had one).
    4262     1038049 :   ExtAddrMode AddrMode = AddrModes.getAddrMode();
    4263             : 
    4264             :   // If all the instructions matched are already in this BB, don't do anything.
    4265             :   // If we saw a Phi node then it is not local definitely, and if we saw a select
    4266             :   // then we want to push the address calculation past it even if it's already
    4267             :   // in this BB.
    4268     2072138 :   if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) {
    4269      262859 :         return IsNonLocalValue(V, MemoryInst->getParent());
    4270             :                   })) {
    4271             :     DEBUG(dbgs() << "CGP: Found      local addrmode: " << AddrMode << "\n");
    4272             :     return false;
    4273             :   }
    4274             : 
    4275             :   // Insert this computation right after this user.  Since our caller is
    4276             :   // scanning from the top of the BB to the bottom, reuse of the expr are
    4277             :   // guaranteed to happen later.
    4278       16176 :   IRBuilder<> Builder(MemoryInst);
    4279             : 
    4280             :   // Now that we determined the addressing expression we want to use and know
    4281             :   // that we have to sink it into this block.  Check to see if we have already
    4282             :   // done this for some other load/store instr in this block.  If so, reuse
    4283             :   // the computation.  Before attempting reuse, check if the address is valid
    4284             :   // as it may have been erased.
    4285             : 
    4286       16176 :   WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
    4287             : 
    4288       16176 :   Value * SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
    4289       16176 :   if (SunkAddr) {
    4290             :     DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
    4291             :                  << *MemoryInst << "\n");
    4292         867 :     if (SunkAddr->getType() != Addr->getType())
    4293           0 :       SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
    4294       15314 :   } else if (AddrSinkUsingGEPs ||
    4295           5 :              (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
    4296           0 :               SubtargetInfo->useAA())) {
    4297             :     // By default, we use the GEP-based method when AA is used later. This
    4298             :     // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
    4299             :     DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
    4300             :                  << *MemoryInst << "\n");
    4301       15304 :     Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
    4302             :     Value *ResultPtr = nullptr, *ResultIndex = nullptr;
    4303             : 
    4304             :     // First, find the pointer.
    4305       45822 :     if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
    4306             :       ResultPtr = AddrMode.BaseReg;
    4307             :       AddrMode.BaseReg = nullptr;
    4308             :     }
    4309             : 
    4310       16597 :     if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
    4311             :       // We can't add more than one pointer together, nor can we scale a
    4312             :       // pointer (both of which seem meaningless).
    4313           0 :       if (ResultPtr || AddrMode.Scale != 1)
    4314             :         return false;
    4315             : 
    4316             :       ResultPtr = AddrMode.ScaledReg;
    4317             :       AddrMode.Scale = 0;
    4318             :     }
    4319             : 
    4320             :     // It is only safe to sign extend the BaseReg if we know that the math
    4321             :     // required to create it did not overflow before we extend it. Since
    4322             :     // the original IR value was tossed in favor of a constant back when
    4323             :     // the AddrMode was created we need to bail out gracefully if widths
    4324             :     // do not match instead of extending it.
    4325             :     //
    4326             :     // (See below for code to add the scale.)
    4327       15304 :     if (AddrMode.Scale) {
    4328        1293 :       Type *ScaledRegTy = AddrMode.ScaledReg->getType();
    4329        1293 :       if (cast<IntegerType>(IntPtrTy)->getBitWidth() >
    4330             :           cast<IntegerType>(ScaledRegTy)->getBitWidth())
    4331             :         return false;
    4332             :     }
    4333             : 
    4334       15294 :     if (AddrMode.BaseGV) {
    4335          37 :       if (ResultPtr)
    4336             :         return false;
    4337             : 
    4338             :       ResultPtr = AddrMode.BaseGV;
    4339             :     }
    4340             : 
    4341             :     // If the real base value actually came from an inttoptr, then the matcher
    4342             :     // will look through it and provide only the integer value. In that case,
    4343             :     // use it here.
    4344       15294 :     if (!DL->isNonIntegralPointerType(Addr->getType())) {
    4345       15291 :       if (!ResultPtr && AddrMode.BaseReg) {
    4346         244 :         ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(),
    4347             :                                            "sunkaddr");
    4348         122 :         AddrMode.BaseReg = nullptr;
    4349       15169 :       } else if (!ResultPtr && AddrMode.Scale == 1) {
    4350           0 :         ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(),
    4351             :                                            "sunkaddr");
    4352             :         AddrMode.Scale = 0;
    4353             :       }
    4354             :     }
    4355             : 
    4356       15294 :     if (!ResultPtr &&
    4357           9 :         !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
    4358           3 :       SunkAddr = Constant::getNullValue(Addr->getType());
    4359       15291 :     } else if (!ResultPtr) {
    4360             :       return false;
    4361             :     } else {
    4362             :       Type *I8PtrTy =
    4363       30566 :           Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
    4364       15283 :       Type *I8Ty = Builder.getInt8Ty();
    4365             : 
    4366             :       // Start with the base register. Do this first so that subsequent address
    4367             :       // matching finds it last, which will prevent it from trying to match it
    4368             :       // as the scaled value in case it happens to be a mul. That would be
    4369             :       // problematic if we've sunk a different mul for the scale, because then
    4370             :       // we'd end up sinking both muls.
    4371       15283 :       if (AddrMode.BaseReg) {
    4372             :         Value *V = AddrMode.BaseReg;
    4373           2 :         if (V->getType() != IntPtrTy)
    4374           2 :           V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
    4375             : 
    4376             :         ResultIndex = V;
    4377             :       }
    4378             : 
    4379             :       // Add the scale value.
    4380       15283 :       if (AddrMode.Scale) {
    4381             :         Value *V = AddrMode.ScaledReg;
    4382        1283 :         if (V->getType() == IntPtrTy) {
    4383             :           // done.
    4384             :         } else {
    4385             :           assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <
    4386             :                  cast<IntegerType>(V->getType())->getBitWidth() &&
    4387             :                  "We can't transform if ScaledReg is too narrow");
    4388           1 :           V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
    4389             :         }
    4390             : 
    4391        1283 :         if (AddrMode.Scale != 1)
    4392         719 :           V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
    4393             :                                 "sunkaddr");
    4394        1283 :         if (ResultIndex)
    4395           0 :           ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
    4396             :         else
    4397             :           ResultIndex = V;
    4398             :       }
    4399             : 
    4400             :       // Add in the Base Offset if present.
    4401       15283 :       if (AddrMode.BaseOffs) {
    4402       14430 :         Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
    4403       14430 :         if (ResultIndex) {
    4404             :           // We need to add this separately from the scale above to help with
    4405             :           // SDAG consecutive load/store merging.
    4406         860 :           if (ResultPtr->getType() != I8PtrTy)
    4407         536 :             ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
    4408         860 :           ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
    4409             :         }
    4410             : 
    4411             :         ResultIndex = V;
    4412             :       }
    4413             : 
    4414       15283 :       if (!ResultIndex) {
    4415             :         SunkAddr = ResultPtr;
    4416             :       } else {
    4417       14855 :         if (ResultPtr->getType() != I8PtrTy)
    4418       13810 :           ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
    4419       14855 :         SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
    4420             :       }
    4421             : 
    4422       15283 :       if (SunkAddr->getType() != Addr->getType())
    4423       14434 :         SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
    4424             :     }
    4425             :   } else {
    4426             :     // We'd require a ptrtoint/inttoptr down the line, which we can't do for
    4427             :     // non-integral pointers, so in that case bail out now.
    4428          10 :     Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
    4429           5 :     Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
    4430             :     PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy);
    4431             :     PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy);
    4432           6 :     if (DL->isNonIntegralPointerType(Addr->getType()) ||
    4433           1 :         (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) ||
    4434           0 :         (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) ||
    4435           0 :         (AddrMode.BaseGV &&
    4436           0 :          DL->isNonIntegralPointerType(AddrMode.BaseGV->getType())))
    4437             :       return false;
    4438             : 
    4439             :     DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
    4440             :                  << *MemoryInst << "\n");
    4441           0 :     Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
    4442             :     Value *Result = nullptr;
    4443             : 
    4444             :     // Start with the base register. Do this first so that subsequent address
    4445             :     // matching finds it last, which will prevent it from trying to match it
    4446             :     // as the scaled value in case it happens to be a mul. That would be
    4447             :     // problematic if we've sunk a different mul for the scale, because then
    4448             :     // we'd end up sinking both muls.
    4449           0 :     if (AddrMode.BaseReg) {
    4450             :       Value *V = AddrMode.BaseReg;
    4451           0 :       if (V->getType()->isPointerTy())
    4452           0 :         V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
    4453           0 :       if (V->getType() != IntPtrTy)
    4454           0 :         V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
    4455             :       Result = V;
    4456             :     }
    4457             : 
    4458             :     // Add the scale value.
    4459           0 :     if (AddrMode.Scale) {
    4460             :       Value *V = AddrMode.ScaledReg;
    4461           0 :       if (V->getType() == IntPtrTy) {
    4462             :         // done.
    4463           0 :       } else if (V->getType()->isPointerTy()) {
    4464           0 :         V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
    4465           0 :       } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
    4466             :                  cast<IntegerType>(V->getType())->getBitWidth()) {
    4467           0 :         V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
    4468             :       } else {
    4469             :         // It is only safe to sign extend the BaseReg if we know that the math
    4470             :         // required to create it did not overflow before we extend it. Since
    4471             :         // the original IR value was tossed in favor of a constant back when
    4472             :         // the AddrMode was created we need to bail out gracefully if widths
    4473             :         // do not match instead of extending it.
    4474             :         Instruction *I = dyn_cast_or_null<Instruction>(Result);
    4475           0 :         if (I && (Result != AddrMode.BaseReg))
    4476           0 :           I->eraseFromParent();
    4477             :         return false;
    4478             :       }
    4479           0 :       if (AddrMode.Scale != 1)
    4480           0 :         V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
    4481             :                               "sunkaddr");
    4482           0 :       if (Result)
    4483           0 :         Result = Builder.CreateAdd(Result, V, "sunkaddr");
    4484             :       else
    4485             :         Result = V;
    4486             :     }
    4487             : 
    4488             :     // Add in the BaseGV if present.
    4489           0 :     if (AddrMode.BaseGV) {
    4490           0 :       Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
    4491           0 :       if (Result)
    4492           0 :         Result = Builder.CreateAdd(Result, V, "sunkaddr");
    4493             :       else
    4494             :         Result = V;
    4495             :     }
    4496             : 
    4497             :     // Add in the Base Offset if present.
    4498           0 :     if (AddrMode.BaseOffs) {
    4499           0 :       Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
    4500           0 :       if (Result)
    4501           0 :         Result = Builder.CreateAdd(Result, V, "sunkaddr");
    4502             :       else
    4503             :         Result = V;
    4504             :     }
    4505             : 
    4506           0 :     if (!Result)
    4507           0 :       SunkAddr = Constant::getNullValue(Addr->getType());
    4508             :     else
    4509           0 :       SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
    4510             :   }
    4511             : 
    4512       16153 :   MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
    4513             :   // Store the newly computed address into the cache. In the case we reused a
    4514             :   // value, this should be idempotent.
    4515       32306 :   SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
    4516             : 
    4517             :   // If we have no uses, recursively delete the value and all dead instructions
    4518             :   // using it.
    4519       16153 :   if (Repl->use_empty()) {
    4520             :     // This can cause recursive deletion, which can invalidate our iterator.
    4521             :     // Use a WeakTrackingVH to hold onto it in case this happens.
    4522             :     Value *CurValue = &*CurInstIterator;
    4523             :     WeakTrackingVH IterHandle(CurValue);
    4524        5903 :     BasicBlock *BB = CurInstIterator->getParent();
    4525             : 
    4526        5903 :     RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
    4527             : 
    4528        5903 :     if (IterHandle != CurValue) {
    4529             :       // If the iterator instruction was recursively deleted, start over at the
    4530             :       // start of the block.
    4531           1 :       CurInstIterator = BB->begin();
    4532             :       SunkAddrs.clear();
    4533             :     }
    4534             :   }
    4535             :   ++NumMemoryInsts;
    4536             :   return true;
    4537             : }
    4538             : 
    4539             : /// If there are any memory operands, use OptimizeMemoryInst to sink their
    4540             : /// address computing into the block when possible / profitable.
    4541        9258 : bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
    4542             :   bool MadeChange = false;
    4543             : 
    4544             :   const TargetRegisterInfo *TRI =
    4545       18516 :       TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
    4546             :   TargetLowering::AsmOperandInfoVector TargetConstraints =
    4547       27774 :       TLI->ParseConstraints(*DL, TRI, CS);
    4548             :   unsigned ArgNo = 0;
    4549       82196 :   for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
    4550       63680 :     TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
    4551             : 
    4552             :     // Compute the constraint code and ConstraintType to use.
    4553       63680 :     TLI->ComputeConstraintToUse(OpInfo, SDValue());
    4554             : 
    4555       67642 :     if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
    4556        3962 :         OpInfo.isIndirect) {
    4557        3074 :       Value *OpVal = CS->getArgOperand(ArgNo++);
    4558        3074 :       MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
    4559       60606 :     } else if (OpInfo.Type == InlineAsm::isInput)
    4560        7495 :       ArgNo++;
    4561             :   }
    4562             : 
    4563        9258 :   return MadeChange;
    4564             : }
    4565             : 
    4566             : /// \brief Check if all the uses of \p Val are equivalent (or free) zero or
    4567             : /// sign extensions.
    4568          31 : static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
    4569             :   assert(!Val->use_empty() && "Input must have at least one use");
    4570             :   const Instruction *FirstUser = cast<Instruction>(*Val->user_begin());
    4571             :   bool IsSExt = isa<SExtInst>(FirstUser);
    4572          31 :   Type *ExtTy = FirstUser->getType();
    4573          84 :   for (const User *U : Val->users()) {
    4574             :     const Instruction *UI = cast<Instruction>(U);
    4575         142 :     if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
    4576             :       return false;
    4577          54 :     Type *CurTy = UI->getType();
    4578             :     // Same input and output types: Same instruction after CSE.
    4579          54 :     if (CurTy == ExtTy)
    4580             :       continue;
    4581             : 
    4582             :     // If IsSExt is true, we are in this situation:
    4583             :     // a = Val
    4584             :     // b = sext ty1 a to ty2
    4585             :     // c = sext ty1 a to ty3
    4586             :     // Assuming ty2 is shorter than ty3, this could be turned into:
    4587             :     // a = Val
    4588             :     // b = sext ty1 a to ty2
    4589             :     // c = sext ty2 b to ty3
    4590             :     // However, the last sext is not free.
    4591          10 :     if (IsSExt)
    4592             :       return false;
    4593             : 
    4594             :     // This is a ZExt, maybe this is free to extend from one type to another.
    4595             :     // In that case, we would not account for a different use.
    4596             :     Type *NarrowTy;
    4597             :     Type *LargeTy;
    4598           9 :     if (ExtTy->getScalarType()->getIntegerBitWidth() >
    4599             :         CurTy->getScalarType()->getIntegerBitWidth()) {
    4600             :       NarrowTy = CurTy;
    4601             :       LargeTy = ExtTy;
    4602             :     } else {
    4603             :       NarrowTy = ExtTy;
    4604             :       LargeTy = CurTy;
    4605             :     }
    4606             : 
    4607           9 :     if (!TLI.isZExtFree(NarrowTy, LargeTy))
    4608             :       return false;
    4609             :   }
    4610             :   // All uses are the same or can be derived from one another for free.
    4611             :   return true;
    4612             : }
    4613             : 
    4614             : /// \brief Try to speculatively promote extensions in \p Exts and continue
    4615             : /// promoting through newly promoted operands recursively as far as doing so is
    4616             : /// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
    4617             : /// When some promotion happened, \p TPT contains the proper state to revert
    4618             : /// them.
    4619             : ///
    4620             : /// \return true if some promotion happened, false otherwise.
    4621       29595 : bool CodeGenPrepare::tryToPromoteExts(
    4622             :     TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
    4623             :     SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
    4624             :     unsigned CreatedInstsCost) {
    4625             :   bool Promoted = false;
    4626             : 
    4627             :   // Iterate over all the extensions to try to promote them.
    4628       71077 :   for (auto I : Exts) {
    4629             :     // Early check if we directly have ext(load).
    4630       38576 :     if (isa<LoadInst>(I->getOperand(0))) {
    4631        8898 :       ProfitablyMovedExts.push_back(I);
    4632       29204 :       continue;
    4633             :     }
    4634             : 
    4635             :     // Check whether or not we want to do any promotion.  The reason we have
    4636             :     // this check inside the for loop is to catch the case where an extension
    4637             :     // is directly fed by a load because in such case the extension can be moved
    4638             :     // up without any promotion on its operands.
    4639       32661 :     if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
    4640        8937 :       return false;
    4641             : 
    4642             :     // Get the action to perform the promotion.
    4643             :     TypePromotionHelper::Action TPH =
    4644       11843 :         TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
    4645             :     // Check if we can promote.
    4646       23238 :     if (!TPH) {
    4647             :       // Save the current extension as we cannot move up through its operand.
    4648       11395 :       ProfitablyMovedExts.push_back(I);
    4649       11395 :       continue;
    4650             :     }
    4651             : 
    4652             :     // Save the current state.
    4653             :     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
    4654             :         TPT.getRestorationPoint();
    4655             :     SmallVector<Instruction *, 4> NewExts;
    4656         448 :     unsigned NewCreatedInstsCost = 0;
    4657         448 :     unsigned ExtCost = !TLI->isExtFree(I);
    4658             :     // Promote.
    4659         448 :     Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
    4660         896 :                              &NewExts, nullptr, *TLI);
    4661             :     assert(PromotedVal &&
    4662             :            "TypePromotionHelper should have filtered out those cases");
    4663             : 
    4664             :     // We would be able to merge only one extension in a load.
    4665             :     // Therefore, if we have more than 1 new extension we heuristically
    4666             :     // cut this search path, because it means we degrade the code quality.
    4667             :     // With exactly 2, the transformation is neutral, because we will merge
    4668             :     // one extension but leave one. However, we optimistically keep going,
    4669             :     // because the new extension may be removed too.
    4670         448 :     long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
    4671             :     // FIXME: It would be possible to propagate a negative value instead of
    4672             :     // conservatively ceiling it to 0.
    4673         448 :     TotalCreatedInstsCost =
    4674         896 :         std::max((long long)0, (TotalCreatedInstsCost - ExtCost));
    4675         461 :     if (!StressExtLdPromotion &&
    4676         379 :         (TotalCreatedInstsCost > 1 ||
    4677         379 :          !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
    4678             :       // This promotion is not profitable, rollback to the previous state, and
    4679             :       // save the current extension in ProfitablyMovedExts as the latest
    4680             :       // speculative promotion turned out to be unprofitable.
    4681          13 :       TPT.rollback(LastKnownGood);
    4682          13 :       ProfitablyMovedExts.push_back(I);
    4683          13 :       continue;
    4684             :     }
    4685             :     // Continue promoting NewExts as far as doing so is profitable.
    4686             :     SmallVector<Instruction *, 2> NewlyMovedExts;
    4687         435 :     (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost);
    4688             :     bool NewPromoted = false;
    4689        1475 :     for (auto ExtInst : NewlyMovedExts) {
    4690         520 :       Instruction *MovedExt = cast<Instruction>(ExtInst);
    4691         520 :       Value *ExtOperand = MovedExt->getOperand(0);
    4692             :       // If we have reached to a load, we need this extra profitability check
    4693             :       // as it could potentially be merged into an ext(load).
    4694         170 :       if (isa<LoadInst>(ExtOperand) &&
    4695         133 :           !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
    4696          31 :             (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI))))
    4697          18 :         continue;
    4698             : 
    4699         502 :       ProfitablyMovedExts.push_back(MovedExt);
    4700             :       NewPromoted = true;
    4701             :     }
    4702             : 
    4703             :     // If none of speculative promotions for NewExts is profitable, rollback
    4704             :     // and save the current extension (I) as the last profitable extension.
    4705         435 :     if (!NewPromoted) {
    4706           0 :       TPT.rollback(LastKnownGood);
    4707           0 :       ProfitablyMovedExts.push_back(I);
    4708             :       continue;
    4709             :     }
    4710             :     // The promotion is profitable.
    4711             :     Promoted = true;
    4712             :   }
    4713       20658 :   return Promoted;
    4714             : }
    4715             : 
    4716             : /// Merging redundant sexts when one is dominating the other.
    4717          36 : bool CodeGenPrepare::mergeSExts(Function &F) {
    4718          36 :   DominatorTree DT(F);
    4719             :   bool Changed = false;
    4720         122 :   for (auto &Entry : ValToSExtendedUses) {
    4721             :     SExts &Insts = Entry.second;
    4722             :     SExts CurPts;
    4723         256 :     for (Instruction *Inst : Insts) {
    4724         309 :       if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) ||
    4725         206 :           Inst->getOperand(0) != Entry.first)
    4726           0 :         continue;
    4727             :       bool inserted = false;
    4728         103 :       for (auto &Pt : CurPts) {
    4729          53 :         if (DT.dominates(Inst, Pt)) {
    4730          30 :           Pt->replaceAllUsesWith(Inst);
    4731          30 :           RemovedInsts.insert(Pt);
    4732          30 :           Pt->removeFromParent();
    4733          30 :           Pt = Inst;
    4734             :           inserted = true;
    4735             :           Changed = true;
    4736          30 :           break;
    4737             :         }
    4738          23 :         if (!DT.dominates(Pt, Inst))
    4739             :           // Give up if we need to merge in a common dominator as the
    4740             :           // expermients show it is not profitable.
    4741             :           continue;
    4742          23 :         Inst->replaceAllUsesWith(Pt);
    4743          23 :         RemovedInsts.insert(Inst);
    4744          23 :         Inst->removeFromParent();
    4745             :         inserted = true;
    4746             :         Changed = true;
    4747          23 :         break;
    4748             :       }
    4749         103 :       if (!inserted)
    4750          50 :         CurPts.push_back(Inst);
    4751             :     }
    4752             :   }
    4753          36 :   return Changed;
    4754             : }
    4755             : 
    4756             : /// Return true, if an ext(load) can be formed from an extension in
    4757             : /// \p MovedExts.
    4758       29130 : bool CodeGenPrepare::canFormExtLd(
    4759             :     const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
    4760             :     Instruction *&Inst, bool HasPromoted) {
    4761       51832 :   for (auto *MovedExtInst : MovedExts) {
    4762       20225 :     if (isa<LoadInst>(MovedExtInst->getOperand(0))) {
    4763        8874 :       LI = cast<LoadInst>(MovedExtInst->getOperand(0));
    4764        8874 :       Inst = MovedExtInst;
    4765        8874 :       break;
    4766             :     }
    4767             :   }
    4768       29130 :   if (!LI)
    4769             :     return false;
    4770             : 
    4771             :   // If they're already in the same block, there's nothing to do.
    4772             :   // Make the cheap checks first if we did not promote.
    4773             :   // If we promoted, we need to check if it is indeed profitable.
    4774        8874 :   if (!HasPromoted && LI->getParent() == Inst->getParent())
    4775             :     return false;
    4776             : 
    4777         619 :   return TLI->isExtLoad(LI, Inst, *DL);
    4778             : }
    4779             : 
    4780             : /// Move a zext or sext fed by a load into the same basic block as the load,
    4781             : /// unless conditions are unfavorable. This allows SelectionDAG to fold the
    4782             : /// extend into the load.
    4783             : ///
    4784             : /// E.g.,
    4785             : /// \code
    4786             : /// %ld = load i32* %addr
    4787             : /// %add = add nuw i32 %ld, 4
    4788             : /// %zext = zext i32 %add to i64
    4789             : // \endcode
    4790             : /// =>
    4791             : /// \code
    4792             : /// %ld = load i32* %addr
    4793             : /// %zext = zext i32 %ld to i64
    4794             : /// %add = add nuw i64 %zext, 4
    4795             : /// \encode
    4796             : /// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
    4797             : /// allow us to match zext(load i32*) to i64.
    4798             : ///
    4799             : /// Also, try to promote the computations used to obtain a sign extended
    4800             : /// value used into memory accesses.
    4801             : /// E.g.,
    4802             : /// \code
    4803             : /// a = add nsw i32 b, 3
    4804             : /// d = sext i32 a to i64
    4805             : /// e = getelementptr ..., i64 d
    4806             : /// \endcode
    4807             : /// =>
    4808             : /// \code
    4809             : /// f = sext i32 b to i64
    4810             : /// a = add nsw i64 f, 3
    4811             : /// e = getelementptr ..., i64 a
    4812             : /// \endcode
    4813             : ///
    4814             : /// \p Inst[in/out] the extension may be modified during the process if some
    4815             : /// promotions apply.
    4816       29132 : bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
    4817             :   // ExtLoad formation and address type promotion infrastructure requires TLI to
    4818             :   // be effective.
    4819       29132 :   if (!TLI)
    4820             :     return false;
    4821             : 
    4822       29130 :   bool AllowPromotionWithoutCommonHeader = false;
    4823             :   /// See if it is an interesting sext operations for the address type
    4824             :   /// promotion before trying to promote it, e.g., the ones with the right
    4825             :   /// type and used in memory accesses.
    4826       29130 :   bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
    4827       29130 :       *Inst, AllowPromotionWithoutCommonHeader);
    4828       29130 :   TypePromotionTransaction TPT(RemovedInsts);
    4829             :   TypePromotionTransaction::ConstRestorationPt LastKnownGood =
    4830             :       TPT.getRestorationPoint();
    4831             :   SmallVector<Instruction *, 1> Exts;
    4832             :   SmallVector<Instruction *, 2> SpeculativelyMovedExts;
    4833       29130 :   Exts.push_back(Inst);
    4834             : 
    4835       29130 :   bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts);
    4836             : 
    4837             :   // Look for a load being extended.
    4838       29130 :   LoadInst *LI = nullptr;
    4839             :   Instruction *ExtFedByLoad;
    4840             : 
    4841             :   // Try to promote a chain of computation if it allows to form an extended
    4842             :   // load.
    4843       29130 :   if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) {
    4844             :     assert(LI && ExtFedByLoad && "Expect a valid load and extension");
    4845             :     TPT.commit();
    4846             :     // Move the extend into the same block as the load
    4847         614 :     ExtFedByLoad->moveAfter(LI);
    4848             :     // CGP does not check if the zext would be speculatively executed when moved
    4849             :     // to the same basic block as the load. Preserving its original location
    4850             :     // would pessimize the debugging experience, as well as negatively impact
    4851             :     // the quality of sample pgo. We don't want to use "line 0" as that has a
    4852             :     // size cost in the line-table section and logically the zext can be seen as
    4853             :     // part of the load. Therefore we conservatively reuse the same debug
    4854             :     // location for the load and the zext.
    4855        1228 :     ExtFedByLoad->setDebugLoc(LI->getDebugLoc());
    4856             :     ++NumExtsMoved;
    4857         614 :     Inst = ExtFedByLoad;
    4858         614 :     return true;
    4859             :   }
    4860             : 
    4861             :   // Continue promoting SExts if known as considerable depending on targets.
    4862       28659 :   if (ATPConsiderable &&
    4863         143 :       performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
    4864             :                                   HasPromoted, TPT, SpeculativelyMovedExts))
    4865             :     return true;
    4866             : 
    4867       28464 :   TPT.rollback(LastKnownGood);
    4868       28464 :   return false;
    4869             : }
    4870             : 
    4871             : // Perform address type promotion if doing so is profitable.
    4872             : // If AllowPromotionWithoutCommonHeader == false, we should find other sext
    4873             : // instructions that sign extended the same initial value. However, if
    4874             : // AllowPromotionWithoutCommonHeader == true, we expect promoting the
    4875             : // extension is just profitable.
    4876         143 : bool CodeGenPrepare::performAddressTypePromotion(
    4877             :     Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
    4878             :     bool HasPromoted, TypePromotionTransaction &TPT,
    4879             :     SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
    4880             :   bool Promoted = false;
    4881             :   SmallPtrSet<Instruction *, 1> UnhandledExts;
    4882             :   bool AllSeenFirst = true;
    4883         435 :   for (auto I : SpeculativelyMovedExts) {
    4884         146 :     Value *HeadOfChain = I->getOperand(0);
    4885             :     DenseMap<Value *, Instruction *>::iterator AlreadySeen =
    4886         146 :         SeenChainsForSExt.find(HeadOfChain);
    4887             :     // If there is an unhandled SExt which has the same header, try to promote
    4888             :     // it as well.
    4889         146 :     if (AlreadySeen != SeenChainsForSExt.end()) {
    4890          53 :       if (AlreadySeen->second != nullptr)
    4891          30 :         UnhandledExts.insert(AlreadySeen->second);
    4892             :       AllSeenFirst = false;
    4893             :     }
    4894             :   }
    4895             : 
    4896         160 :   if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
    4897             :                         SpeculativelyMovedExts.size() == 1)) {
    4898             :     TPT.commit();
    4899          70 :     if (HasPromoted)
    4900             :       Promoted = true;
    4901         212 :     for (auto I : SpeculativelyMovedExts) {
    4902         142 :       Value *HeadOfChain = I->getOperand(0);
    4903         142 :       SeenChainsForSExt[HeadOfChain] = nullptr;
    4904         142 :       ValToSExtendedUses[HeadOfChain].push_back(I);
    4905             :     }
    4906             :     // Update Inst as promotion happen.
    4907          70 :     Inst = SpeculativelyMovedExts.pop_back_val();
    4908             :   } else {
    4909             :     // This is the first chain visited from the header, keep the current chain
    4910             :     // as unhandled. Defer to promote this until we encounter another SExt
    4911             :     // chain derived from the same header.
    4912         223 :     for (auto I : SpeculativelyMovedExts) {
    4913         150 :       Value *HeadOfChain = I->getOperand(0);
    4914         150 :       SeenChainsForSExt[HeadOfChain] = Inst;
    4915             :     }
    4916             :     return false;
    4917             :   }
    4918             : 
    4919         123 :   if (!AllSeenFirst && !UnhandledExts.empty())
    4920          60 :     for (auto VisitedSExt : UnhandledExts) {
    4921          30 :       if (RemovedInsts.count(VisitedSExt))
    4922           0 :         continue;
    4923          30 :       TypePromotionTransaction TPT(RemovedInsts);
    4924             :       SmallVector<Instruction *, 1> Exts;
    4925             :       SmallVector<Instruction *, 2> Chains;
    4926          30 :       Exts.push_back(VisitedSExt);
    4927          30 :       bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains);
    4928             :       TPT.commit();
    4929          30 :       if (HasPromoted)
    4930             :         Promoted = true;
    4931          94 :       for (auto I : Chains) {
    4932          64 :         Value *HeadOfChain = I->getOperand(0);
    4933             :         // Mark this as handled.
    4934          64 :         SeenChainsForSExt[HeadOfChain] = nullptr;
    4935          64 :         ValToSExtendedUses[HeadOfChain].push_back(I);
    4936             :       }
    4937             :     }
    4938             :   return Promoted;
    4939             : }
    4940             : 
    4941       29132 : bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
    4942       29132 :   BasicBlock *DefBB = I->getParent();
    4943             : 
    4944             :   // If the result of a {s|z}ext and its source are both live out, rewrite all
    4945             :   // other uses of the source with result of extension.
    4946       29132 :   Value *Src = I->getOperand(0);
    4947       29132 :   if (Src->hasOneUse())
    4948             :     return false;
    4949             : 
    4950             :   // Only do this xform if truncating is free.
    4951        3535 :   if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
    4952             :     return false;
    4953             : 
    4954             :   // Only safe to perform the optimization if the source is also defined in
    4955             :   // this block.
    4956        3300 :   if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
    4957             :     return false;
    4958             : 
    4959             :   bool DefIsLiveOut = false;
    4960        4144 :   for (User *U : I->users()) {
    4961             :     Instruction *UI = cast<Instruction>(U);
    4962             : 
    4963             :     // Figure out which BB this ext is used in.
    4964        2684 :     BasicBlock *UserBB = UI->getParent();
    4965        2684 :     if (UserBB == DefBB) continue;
    4966             :     DefIsLiveOut = true;
    4967             :     break;
    4968             :   }
    4969        2594 :   if (!DefIsLiveOut)
    4970             :     return false;
    4971             : 
    4972             :   // Make sure none of the uses are PHI nodes.
    4973        3897 :   for (User *U : Src->users()) {
    4974             :     Instruction *UI = cast<Instruction>(U);
    4975        2786 :     BasicBlock *UserBB = UI->getParent();
    4976        2786 :     if (UserBB == DefBB) continue;
    4977             :     // Be conservative. We don't want this xform to end up introducing
    4978             :     // reloads just before load / store instructions.
    4979         550 :     if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
    4980             :       return false;
    4981             :   }
    4982             : 
    4983             :   // InsertedTruncs - Only insert one trunc in each block once.
    4984             :   DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
    4985             : 
    4986             :   bool MadeChange = false;
    4987        2921 :   for (Use &U : Src->uses()) {
    4988        1810 :     Instruction *User = cast<Instruction>(U.getUser());
    4989             : 
    4990             :     // Figure out which BB this ext is used in.
    4991        1810 :     BasicBlock *UserBB = User->getParent();
    4992        1810 :     if (UserBB == DefBB) continue;
    4993             : 
    4994             :     // Both src and def are live in this block. Rewrite the use.
    4995             :     Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
    4996             : 
    4997         484 :     if (!InsertedTrunc) {
    4998         484 :       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
    4999             :       assert(InsertPt != UserBB->end());
    5000         968 :       InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
    5001         484 :       InsertedInsts.insert(InsertedTrunc);
    5002             :     }
    5003             : 
    5004             :     // Replace a use of the {s|z}ext source with a use of the result.
    5005         484 :     U = InsertedTrunc;
    5006             :     ++NumExtUses;
    5007             :     MadeChange = true;
    5008             :   }
    5009             : 
    5010             :   return MadeChange;
    5011             : }
    5012             : 
    5013             : // Find loads whose uses only use some of the loaded value's bits.  Add an "and"
    5014             : // just after the load if the target can fold this into one extload instruction,
    5015             : // with the hope of eliminating some of the other later "and" instructions using
    5016             : // the loaded value.  "and"s that are made trivially redundant by the insertion
    5017             : // of the new "and" are removed by this function, while others (e.g. those whose
    5018             : // path from the load goes through a phi) are left for isel to potentially
    5019             : // remove.
    5020             : //
    5021             : // For example:
    5022             : //
    5023             : // b0:
    5024             : //   x = load i32
    5025             : //   ...
    5026             : // b1:
    5027             : //   y = and x, 0xff
    5028             : //   z = use y
    5029             : //
    5030             : // becomes:
    5031             : //
    5032             : // b0:
    5033             : //   x = load i32
    5034             : //   x' = and x, 0xff
    5035             : //   ...
    5036             : // b1:
    5037             : //   z = use x'
    5038             : //
    5039             : // whereas:
    5040             : //
    5041             : // b0:
    5042             : //   x1 = load i32
    5043             : //   ...
    5044             : // b1:
    5045             : //   x2 = load i32
    5046             : //   ...
    5047             : // b2:
    5048             : //   x = phi x1, x2
    5049             : //   y = and x, 0xff
    5050             : //
    5051             : // becomes (after a call to optimizeLoadExt for each load):
    5052             : //
    5053             : // b0:
    5054             : //   x1 = load i32
    5055             : //   x1' = and x1, 0xff
    5056             : //   ...
    5057             : // b1:
    5058             : //   x2 = load i32
    5059             : //   x2' = and x2, 0xff
    5060             : //   ...
    5061             : // b2:
    5062             : //   x = phi x1', x2'
    5063             : //   y = and x, 0xff
    5064      507148 : bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
    5065      499012 :   if (!Load->isSimple() ||
    5066      711102 :       !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
    5067             :     return false;
    5068             : 
    5069             :   // Skip loads we've already transformed.
    5070      629443 :   if (Load->hasOneUse() &&
    5071      604644 :       InsertedInsts.count(cast<Instruction>(*Load->user_begin())))
    5072             :     return false;
    5073             : 
    5074             :   // Look at all uses of Load, looking through phis, to determine how many bits
    5075             :   // of the loaded value are needed.
    5076             :   SmallVector<Instruction *, 8> WorkList;
    5077             :   SmallPtrSet<Instruction *, 16> Visited;
    5078             :   SmallVector<Instruction *, 8> AndsToMaybeRemove;
    5079      696833 :   for (auto *U : Load->users())
    5080      369112 :     WorkList.push_back(cast<Instruction>(U));
    5081             : 
    5082      327721 :   EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
    5083      327721 :   unsigned BitWidth = LoadResultVT.getSizeInBits();
    5084             :   APInt DemandBits(BitWidth, 0);
    5085             :   APInt WidestAndBits(BitWidth, 0);
    5086             : 
    5087      337386 :   while (!WorkList.empty()) {
    5088      334903 :     Instruction *I = WorkList.back();
    5089             :     WorkList.pop_back();
    5090             : 
    5091             :     // Break use-def graph loops.
    5092      334903 :     if (!Visited.insert(I).second)
    5093        7470 :       continue;
    5094             : 
    5095             :     // For a PHI node, push all of its users.
    5096      334837 :     if (auto *Phi = dyn_cast<PHINode>(I)) {
    5097       17692 :       for (auto *U : Phi->users())
    5098       10354 :         WorkList.push_back(cast<Instruction>(U));
    5099        7338 :       continue;
    5100             :     }
    5101             : 
    5102      327499 :     switch (I->getOpcode()) {
    5103        2286 :     case Instruction::And: {
    5104        2286 :       auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
    5105             :       if (!AndC)
    5106        1239 :         return false;
    5107             :       APInt AndBits = AndC->getValue();
    5108             :       DemandBits |= AndBits;
    5109             :       // Keep track of the widest and mask we see.
    5110        1047 :       if (AndBits.ugt(WidestAndBits))
    5111         981 :         WidestAndBits = AndBits;
    5112        2028 :       if (AndBits == WidestAndBits && I->getOperand(0) == Load)
    5113         805 :         AndsToMaybeRemove.push_back(I);
    5114             :       break;
    5115             :     }
    5116             : 
    5117        1141 :     case Instruction::Shl: {
    5118        1141 :       auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
    5119             :       if (!ShlC)
    5120             :         return false;
    5121         869 :       uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
    5122         869 :       DemandBits.setLowBits(BitWidth - ShiftAmt);
    5123             :       break;
    5124             :     }
    5125             : 
    5126         345 :     case Instruction::Trunc: {
    5127         345 :       EVT TruncVT = TLI->getValueType(*DL, I->getType());
    5128         345 :       unsigned TruncBitWidth = TruncVT.getSizeInBits();
    5129             :       DemandBits.setLowBits(TruncBitWidth);
    5130             :       break;
    5131             :     }
    5132             : 
    5133             :     default:
    5134             :       return false;
    5135             :     }
    5136             :   }
    5137             : 
    5138             :   uint32_t ActiveBits = DemandBits.getActiveBits();
    5139             :   // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
    5140             :   // target even if isLoadExtLegal says an i1 EXTLOAD is valid.  For example,
    5141             :   // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
    5142             :   // (and (load x) 1) is not matched as a single instruction, rather as a LDR
    5143             :   // followed by an AND.
    5144             :   // TODO: Look into removing this restriction by fixing backends to either
    5145             :   // return false for isLoadExtLegal for i1 or have them select this pattern to
    5146             :   // a single instruction.
    5147             :   //
    5148             :   // Also avoid hoisting if we didn't see any ands with the exact DemandBits
    5149             :   // mask, since these are the only ands that will be removed by isel.
    5150        3718 :   if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) ||
    5151             :       WidestAndBits != DemandBits)
    5152             :     return false;
    5153             : 
    5154         131 :   LLVMContext &Ctx = Load->getType()->getContext();
    5155         131 :   Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
    5156         131 :   EVT TruncVT = TLI->getValueType(*DL, TruncTy);
    5157             : 
    5158             :   // Reject cases that won't be matched as extloads.
    5159         194 :   if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
    5160          63 :       !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
    5161             :     return false;
    5162             : 
    5163          63 :   IRBuilder<> Builder(Load->getNextNode());
    5164         126 :   auto *NewAnd = dyn_cast<Instruction>(
    5165          63 :       Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
    5166             :   // Mark this instruction as "inserted by CGP", so that other
    5167             :   // optimizations don't touch it.
    5168          63 :   InsertedInsts.insert(NewAnd);
    5169             : 
    5170             :   // Replace all uses of load with new and (except for the use of load in the
    5171             :   // new and itself).
    5172          63 :   Load->replaceAllUsesWith(NewAnd);
    5173          63 :   NewAnd->setOperand(0, Load);
    5174             : 
    5175             :   // Remove any and instructions that are now redundant.
    5176         173 :   for (auto *And : AndsToMaybeRemove)
    5177             :     // Check that the and mask is the same as the one we decided to put on the
    5178             :     // new and.
    5179         110 :     if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
    5180          55 :       And->replaceAllUsesWith(NewAnd);
    5181          55 :       if (&*CurInstIterator == And)
    5182          72 :         CurInstIterator = std::next(And->getIterator());
    5183          55 :       And->eraseFromParent();
    5184             :       ++NumAndUses;
    5185             :     }
    5186             : 
    5187             :   ++NumAndsAdded;
    5188             :   return true;
    5189             : }
    5190             : 
    5191             : /// Check if V (an operand of a select instruction) is an expensive instruction
    5192             : /// that is only used once.
    5193        8114 : static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
    5194             :   auto *I = dyn_cast<Instruction>(V);
    5195             :   // If it's safe to speculatively execute, then it should not have side
    5196             :   // effects; therefore, it's safe to sink and possibly *not* execute.
    5197        1756 :   return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
    5198        8552 :          TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
    5199             : }
    5200             : 
    5201             : /// Returns true if a SelectInst should be turned into an explicit branch.
    5202       56515 : static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
    5203             :                                                 const TargetLowering *TLI,
    5204             :                                                 SelectInst *SI) {
    5205             :   // If even a predictable select is cheap, then a branch can't be cheaper.
    5206       56515 :   if (!TLI->isPredictableSelectExpensive())
    5207             :     return false;
    5208             : 
    5209             :   // FIXME: This should use the same heuristics as IfConversion to determine
    5210             :   // whether a select is better represented as a branch.
    5211             : 
    5212             :   // If metadata tells us that the select condition is obviously predictable,
    5213             :   // then we want to replace the select with a branch.
    5214             :   uint64_t TrueWeight, FalseWeight;
    5215       50118 :   if (SI->extractProfMetadata(TrueWeight, FalseWeight)) {
    5216           9 :     uint64_t Max = std::max(TrueWeight, FalseWeight);
    5217           9 :     uint64_t Sum = TrueWeight + FalseWeight;
    5218           9 :     if (Sum != 0) {
    5219           8 :       auto Probability = BranchProbability::getBranchProbability(Max, Sum);
    5220           8 :       if (Probability > TLI->getPredictableBranchThreshold())
    5221           7 :         return true;
    5222             :     }
    5223             :   }
    5224             : 
    5225             :   CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
    5226             : 
    5227             :   // If a branch is predictable, an out-of-order CPU can avoid blocking on its
    5228             :   // comparison condition. If the compare has more than one use, there's
    5229             :   // probably another cmov or setcc around, so it's not worth emitting a branch.
    5230       42057 :   if (!Cmp || !Cmp->hasOneUse())
    5231             :     return false;
    5232             : 
    5233             :   // If either operand of the select is expensive and only needed on one side
    5234             :   // of the select, we should form a branch.
    5235        8068 :   if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
    5236        4033 :       sinkSelectOperand(TTI, SI->getFalseValue()))
    5237             :     return true;
    5238             : 
    5239             :   return false;
    5240             : }
    5241             : 
    5242             : /// If \p isTrue is true, return the true value of \p SI, otherwise return
    5243             : /// false value of \p SI. If the true/false value of \p SI is defined by any
    5244             : /// select instructions in \p Selects, look through the defining select
    5245             : /// instruction until the true/false value is not defined in \p Selects.
    5246          46 : static Value *getTrueOrFalseValue(
    5247             :     SelectInst *SI, bool isTrue,
    5248             :     const SmallPtrSet<const Instruction *, 2> &Selects) {
    5249             :   Value *V;
    5250             : 
    5251          95 :   for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI);
    5252             :        DefSI = dyn_cast<SelectInst>(V)) {
    5253             :     assert(DefSI->getCondition() == SI->getCondition() &&
    5254             :            "The condition of DefSI does not match with SI");
    5255          49 :     V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
    5256             :   }
    5257          46 :   return V;
    5258             : }
    5259             : 
    5260             : /// If we have a SelectInst that will likely profit from branch prediction,
    5261             : /// turn it into a branch.
    5262       70657 : bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
    5263             :   // Find all consecutive select instructions that share the same condition.
    5264             :   SmallVector<SelectInst *, 2> ASI;
    5265       70657 :   ASI.push_back(SI);
    5266       70657 :   for (BasicBlock::iterator It = ++BasicBlock::iterator(SI);
    5267      142210 :        It != SI->getParent()->end(); ++It) {
    5268       71105 :     SelectInst *I = dyn_cast<SelectInst>(&*It);
    5269       71856 :     if (I && SI->getCondition() == I->getCondition()) {
    5270         448 :       ASI.push_back(I);
    5271             :     } else {
    5272             :       break;
    5273             :     }
    5274             :   }
    5275             : 
    5276       70657 :   SelectInst *LastSI = ASI.back();
    5277             :   // Increment the current iterator to skip all the rest of select instructions
    5278             :   // because they will be either "not lowered" or "all lowered" to branch.
    5279      141314 :   CurInstIterator = std::next(LastSI->getIterator());
    5280             : 
    5281       70657 :   bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
    5282             : 
    5283             :   // Can we convert the 'select' to CF ?
    5284      118300 :   if (DisableSelectToBranch || OptSize || !TLI || VectorCond ||
    5285       56526 :       SI->getMetadata(LLVMContext::MD_unpredictable))
    5286             :     return false;
    5287             : 
    5288             :   TargetLowering::SelectSupportKind SelectKind;
    5289             :   if (VectorCond)
    5290             :     SelectKind = TargetLowering::VectorMaskSelect;
    5291      113050 :   else if (SI->getType()->isVectorTy())
    5292             :     SelectKind = TargetLowering::ScalarCondVectorVal;
    5293             :   else
    5294             :     SelectKind = TargetLowering::ScalarValSelect;
    5295             : 
    5296      113040 :   if (TLI->isSelectSupported(SelectKind) &&
    5297       56515 :       !isFormingBranchFromSelectProfitable(TTI, TLI, SI))
    5298             :     return false;
    5299             : 
    5300          21 :   ModifiedDT = true;
    5301             : 
    5302             :   // Transform a sequence like this:
    5303             :   //    start:
    5304             :   //       %cmp = cmp uge i32 %a, %b
    5305             :   //       %sel = select i1 %cmp, i32 %c, i32 %d
    5306             :   //
    5307             :   // Into:
    5308             :   //    start:
    5309             :   //       %cmp = cmp uge i32 %a, %b
    5310             :   //       br i1 %cmp, label %select.true, label %select.false
    5311             :   //    select.true:
    5312             :   //       br label %select.end
    5313             :   //    select.false:
    5314             :   //       br label %select.end
    5315             :   //    select.end:
    5316             :   //       %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
    5317             :   //
    5318             :   // In addition, we may sink instructions that produce %c or %d from
    5319             :   // the entry block into the destination(s) of the new branch.
    5320             :   // If the true or false blocks do not contain a sunken instruction, that
    5321             :   // block and its branch may be optimized away. In that case, one side of the
    5322             :   // first branch will point directly to select.end, and the corresponding PHI
    5323             :   // predecessor block will be the start block.
    5324             : 
    5325             :   // First, we split the block containing the select into 2 blocks.
    5326          21 :   BasicBlock *StartBlock = SI->getParent();
    5327          21 :   BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
    5328          21 :   BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
    5329             : 
    5330             :   // Delete the unconditional branch that was just created by the split.
    5331          21 :   StartBlock->getTerminator()->eraseFromParent();
    5332             : 
    5333             :   // These are the new basic blocks for the conditional branch.
    5334             :   // At least one will become an actual new basic block.
    5335             :   BasicBlock *TrueBlock = nullptr;
    5336             :   BasicBlock *FalseBlock = nullptr;
    5337             :   BranchInst *TrueBranch = nullptr;
    5338             :   BranchInst *FalseBranch = nullptr;
    5339             : 
    5340             :   // Sink expensive instructions into the conditional blocks to avoid executing
    5341             :   // them speculatively.
    5342          67 :   for (SelectInst *SI : ASI) {
    5343          23 :     if (sinkSelectOperand(TTI, SI->getTrueValue())) {
    5344           2 :       if (TrueBlock == nullptr) {
    5345           6 :         TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
    5346             :                                        EndBlock->getParent(), EndBlock);
    5347             :         TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
    5348             :       }
    5349             :       auto *TrueInst = cast<Instruction>(SI->getTrueValue());
    5350           2 :       TrueInst->moveBefore(TrueBranch);
    5351             :     }
    5352          23 :     if (sinkSelectOperand(TTI, SI->getFalseValue())) {
    5353           3 :       if (FalseBlock == nullptr) {
    5354           9 :         FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
    5355             :                                         EndBlock->getParent(), EndBlock);
    5356             :         FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
    5357             :       }
    5358             :       auto *FalseInst = cast<Instruction>(SI->getFalseValue());
    5359           3 :       FalseInst->moveBefore(FalseBranch);
    5360             :     }
    5361             :   }
    5362             : 
    5363             :   // If there was nothing to sink, then arbitrarily choose the 'false' side
    5364             :   // for a new input value to the PHI.
    5365          21 :   if (TrueBlock == FalseBlock) {
    5366             :     assert(TrueBlock == nullptr &&
    5367             :            "Unexpected basic block transform while optimizing select");
    5368             : 
    5369          51 :     FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
    5370             :                                     EndBlock->getParent(), EndBlock);
    5371          17 :     BranchInst::Create(EndBlock, FalseBlock);
    5372             :   }
    5373             : 
    5374             :   // Insert the real conditional branch based on the original condition.
    5375             :   // If we did not create a new block for one of the 'true' or 'false' paths
    5376             :   // of the condition, it means that side of the branch goes to the end block
    5377             :   // directly and the path originates from the start block from the point of
    5378             :   // view of the new PHI.
    5379             :   BasicBlock *TT, *FT;
    5380          21 :   if (TrueBlock == nullptr) {
    5381             :     TT = EndBlock;
    5382             :     FT = FalseBlock;
    5383             :     TrueBlock = StartBlock;
    5384           2 :   } else if (FalseBlock == nullptr) {
    5385             :     TT = TrueBlock;
    5386             :     FT = EndBlock;
    5387             :     FalseBlock = StartBlock;
    5388             :   } else {
    5389             :     TT = TrueBlock;
    5390             :     FT = FalseBlock;
    5391             :   }
    5392          63 :   IRBuilder<>(SI).CreateCondBr(SI->getCondition(), TT, FT, SI);
    5393             : 
    5394             :   SmallPtrSet<const Instruction *, 2> INS;
    5395             :   INS.insert(ASI.begin(), ASI.end());
    5396             :   // Use reverse iterator because later select may use the value of the
    5397             :   // earlier select, and we need to propagate value through earlier select
    5398             :   // to get the PHI operand.
    5399          21 :   for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) {
    5400          23 :     SelectInst *SI = *It;
    5401             :     // The select itself is replaced with a PHI Node.
    5402          46 :     PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
    5403          23 :     PN->takeName(SI);
    5404          23 :     PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock);
    5405          23 :     PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock);
    5406             : 
    5407          23 :     SI->replaceAllUsesWith(PN);
    5408          23 :     SI->eraseFromParent();
    5409             :     INS.erase(SI);
    5410             :     ++NumSelectsExpanded;
    5411             :   }
    5412             : 
    5413             :   // Instruct OptimizeBlock to skip to the next block.
    5414          21 :   CurInstIterator = StartBlock->end();
    5415             :   return true;
    5416             : }
    5417             : 
    5418       11664 : static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
    5419             :   SmallVector<int, 16> Mask(SVI->getShuffleMask());
    5420             :   int SplatElem = -1;
    5421       92247 :   for (unsigned i = 0; i < Mask.size(); ++i) {
    5422       50841 :     if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
    5423             :       return false;
    5424       22973 :     SplatElem = Mask[i];
    5425             :   }
    5426             : 
    5427             :   return true;
    5428             : }
    5429             : 
    5430             : /// Some targets have expensive vector shifts if the lanes aren't all the same
    5431             : /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
    5432             : /// it's often worth sinking a shufflevector splat down to its use so that
    5433             : /// codegen can spot all lanes are identical.
    5434       35581 : bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
    5435       35581 :   BasicBlock *DefBB = SVI->getParent();
    5436             : 
    5437             :   // Only do this xform if variable vector shifts are particularly expensive.
    5438       71162 :   if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
    5439             :     return false;
    5440             : 
    5441             :   // We only expect better codegen by sinking a shuffle if we can recognise a
    5442             :   // constant splat.
    5443       11664 :   if (!isBroadcastShuffle(SVI))
    5444             :     return false;
    5445             : 
    5446             :   // InsertedShuffles - Only insert a shuffle in each block once.
    5447             :   DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
    5448             : 
    5449             :   bool MadeChange = false;
    5450        2866 :   for (User *U : SVI->users()) {
    5451             :     Instruction *UI = cast<Instruction>(U);
    5452             : 
    5453             :     // Figure out which BB this ext is used in.
    5454        1563 :     BasicBlock *UserBB = UI->getParent();
    5455        3122 :     if (UserBB == DefBB) continue;
    5456             : 
    5457             :     // For now only apply this when the splat is used by a shift instruction.
    5458          50 :     if (!UI->isShift()) continue;
    5459             : 
    5460             :     // Everything checks out, sink the shuffle if the user's block doesn't
    5461             :     // already have a copy.
    5462             :     Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
    5463             : 
    5464           4 :     if (!InsertedShuffle) {
    5465           4 :       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
    5466             :       assert(InsertPt != UserBB->end());
    5467           4 :       InsertedShuffle =
    5468             :           new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
    5469           4 :                                 SVI->getOperand(2), "", &*InsertPt);
    5470             :     }
    5471             : 
    5472           4 :     UI->replaceUsesOfWith(SVI, InsertedShuffle);
    5473             :     MadeChange = true;
    5474             :   }
    5475             : 
    5476             :   // If we removed all uses, nuke the shuffle.
    5477        1303 :   if (SVI->use_empty()) {
    5478           2 :     SVI->eraseFromParent();
    5479             :     MadeChange = true;
    5480             :   }
    5481             : 
    5482             :   return MadeChange;
    5483             : }
    5484             : 
    5485        1365 : bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
    5486        1365 :   if (!TLI || !DL)
    5487             :     return false;
    5488             : 
    5489             :   Value *Cond = SI->getCondition();
    5490        1361 :   Type *OldType = Cond->getType();
    5491        1361 :   LLVMContext &Context = Cond->getContext();
    5492        1361 :   MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
    5493        1361 :   unsigned RegWidth = RegType.getSizeInBits();
    5494             : 
    5495        1361 :   if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
    5496             :     return false;
    5497             : 
    5498             :   // If the register width is greater than the type width, expand the condition
    5499             :   // of the switch instruction and each case constant to the width of the
    5500             :   // register. By widening the type of the switch condition, subsequent
    5501             :   // comparisons (for case comparisons) will not need to be extended to the
    5502             :   // preferred register width, so we will potentially eliminate N-1 extends,
    5503             :   // where N is the number of cases in the switch.
    5504          51 :   auto *NewType = Type::getIntNTy(Context, RegWidth);
    5505             : 
    5506             :   // Zero-extend the switch condition and case constants unless the switch
    5507             :   // condition is a function argument that is already being sign-extended.
    5508             :   // In that case, we can avoid an unnecessary mask/extension by sign-extending
    5509             :   // everything instead.
    5510             :   Instruction::CastOps ExtType = Instruction::ZExt;
    5511             :   if (auto *Arg = dyn_cast<Argument>(Cond))
    5512           3 :     if (Arg->hasSExtAttr())
    5513             :       ExtType = Instruction::SExt;
    5514             : 
    5515          51 :   auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
    5516          51 :   ExtInst->insertBefore(SI);
    5517             :   SI->setCondition(ExtInst);
    5518         179 :   for (auto Case : SI->cases()) {
    5519             :     APInt NarrowConst = Case.getCaseValue()->getValue();
    5520             :     APInt WideConst = (ExtType == Instruction::ZExt) ?
    5521         128 :                       NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
    5522         128 :     Case.setValue(ConstantInt::get(Context, WideConst));
    5523             :   }
    5524             : 
    5525             :   return true;
    5526             : }
    5527             : 
    5528             : 
    5529             : namespace {
    5530             : 
    5531             : /// \brief Helper class to promote a scalar operation to a vector one.
    5532             : /// This class is used to move downward extractelement transition.
    5533             : /// E.g.,
    5534             : /// a = vector_op <2 x i32>
    5535             : /// b = extractelement <2 x i32> a, i32 0
    5536             : /// c = scalar_op b
    5537             : /// store c
    5538             : ///
    5539             : /// =>
    5540             : /// a = vector_op <2 x i32>
    5541             : /// c = vector_op a (equivalent to scalar_op on the related lane)
    5542             : /// * d = extractelement <2 x i32> c, i32 0
    5543             : /// * store d
    5544             : /// Assuming both extractelement and store can be combine, we get rid of the
    5545             : /// transition.
    5546             : class VectorPromoteHelper {
    5547             :   /// DataLayout associated with the current module.
    5548             :   const DataLayout &DL;
    5549             : 
    5550             :   /// Used to perform some checks on the legality of vector operations.
    5551             :   const TargetLowering &TLI;
    5552             : 
    5553             :   /// Used to estimated the cost of the promoted chain.
    5554             :   const TargetTransformInfo &TTI;
    5555             : 
    5556             :   /// The transition being moved downwards.
    5557             :   Instruction *Transition;
    5558             : 
    5559             :   /// The sequence of instructions to be promoted.
    5560             :   SmallVector<Instruction *, 4> InstsToBePromoted;
    5561             : 
    5562             :   /// Cost of combining a store and an extract.
    5563             :   unsigned StoreExtractCombineCost;
    5564             : 
    5565             :   /// Instruction that will be combined with the transition.
    5566             :   Instruction *CombineInst = nullptr;
    5567             : 
    5568             :   /// \brief The instruction that represents the current end of the transition.
    5569             :   /// Since we are faking the promotion until we reach the end of the chain
    5570             :   /// of computation, we need a way to get the current end of the transition.
    5571             :   Instruction *getEndOfTransition() const {
    5572         114 :     if (InstsToBePromoted.empty())
    5573          78 :       return Transition;
    5574          36 :     return InstsToBePromoted.back();
    5575             :   }
    5576             : 
    5577             :   /// \brief Return the index of the original value in the transition.
    5578             :   /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
    5579             :   /// c, is at index 0.
    5580             :   unsigned getTransitionOriginalValueIdx() const {
    5581             :     assert(isa<ExtractElementInst>(Transition) &&
    5582             :            "Other kind of transitions are not supported yet");
    5583             :     return 0;
    5584             :   }
    5585             : 
    5586             :   /// \brief Return the index of the index in the transition.
    5587             :   /// E.g., for "extractelement <2 x i32> c, i32 0" the index
    5588             :   /// is at index 1.
    5589             :   unsigned getTransitionIdx() const {
    5590             :     assert(isa<ExtractElementInst>(Transition) &&
    5591             :            "Other kind of transitions are not supported yet");
    5592             :     return 1;
    5593             :   }
    5594             : 
    5595             :   /// \brief Get the type of the transition.
    5596             :   /// This is the type of the original value.
    5597             :   /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
    5598             :   /// transition is <2 x i32>.
    5599             :   Type *getTransitionType() const {
    5600         218 :     return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
    5601             :   }
    5602             : 
    5603             :   /// \brief Promote \p ToBePromoted by moving \p Def downward through.
    5604             :   /// I.e., we have the following sequence:
    5605             :   /// Def = Transition <ty1> a to <ty2>
    5606             :   /// b = ToBePromoted <ty2> Def, ...
    5607             :   /// =>
    5608             :   /// b = ToBePromoted <ty1> a, ...
    5609             :   /// Def = Transition <ty1> ToBePromoted to <ty2>
    5610             :   void promoteImpl(Instruction *ToBePromoted);
    5611             : 
    5612             :   /// \brief Check whether or not it is profitable to promote all the
    5613             :   /// instructions enqueued to be promoted.
    5614           6 :   bool isProfitableToPromote() {
    5615           6 :     Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
    5616             :     unsigned Index = isa<ConstantInt>(ValIdx)
    5617             :                          ? cast<ConstantInt>(ValIdx)->getZExtValue()
    5618           6 :                          : -1;
    5619             :     Type *PromotedType = getTransitionType();
    5620             : 
    5621           6 :     StoreInst *ST = cast<StoreInst>(CombineInst);
    5622             :     unsigned AS = ST->getPointerAddressSpace();
    5623             :     unsigned Align = ST->getAlignment();
    5624             :     // Check if this store is supported.
    5625          12 :     if (!TLI.allowsMisalignedMemoryAccesses(
    5626             :             TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
    5627           6 :             Align)) {
    5628             :       // If this is not supported, there is no way we can combine
    5629             :       // the extract with the store.
    5630             :       return false;
    5631             :     }
    5632             : 
    5633             :     // The scalar chain of computation has to pay for the transition
    5634             :     // scalar to vector.
    5635             :     // The vector chain has to account for the combining cost.
    5636             :     uint64_t ScalarCost =
    5637          12 :         TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
    5638           6 :     uint64_t VectorCost = StoreExtractCombineCost;
    5639          42 :     for (const auto &Inst : InstsToBePromoted) {
    5640             :       // Compute the cost.
    5641             :       // By construction, all instructions being promoted are arithmetic ones.
    5642             :       // Moreover, one argument is a constant that can be viewed as a splat
    5643             :       // constant.
    5644          18 :       Value *Arg0 = Inst->getOperand(0);
    5645          18 :       bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
    5646             :                             isa<ConstantFP>(Arg0);
    5647             :       TargetTransformInfo::OperandValueKind Arg0OVK =
    5648             :           IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
    5649             :                          : TargetTransformInfo::OK_AnyValue;
    5650             :       TargetTransformInfo::OperandValueKind Arg1OVK =
    5651          18 :           !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
    5652             :                           : TargetTransformInfo::OK_AnyValue;
    5653          54 :       ScalarCost += TTI.getArithmeticInstrCost(
    5654          18 :           Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
    5655          54 :       VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
    5656          18 :                                                Arg0OVK, Arg1OVK);
    5657             :     }
    5658             :     DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
    5659             :                  << ScalarCost << "\nVector: " << VectorCost << '\n');
    5660           6 :     return ScalarCost > VectorCost;
    5661             :   }
    5662             : 
    5663             :   /// \brief Generate a constant vector with \p Val with the same
    5664             :   /// number of elements as the transition.
    5665             :   /// \p UseSplat defines whether or not \p Val should be replicated
    5666             :   /// across the whole vector.
    5667             :   /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
    5668             :   /// otherwise we generate a vector with as many undef as possible:
    5669             :   /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
    5670             :   /// used at the index of the extract.
    5671          38 :   Value *getConstantVector(Constant *Val, bool UseSplat) const {
    5672             :     unsigned ExtractIdx = std::numeric_limits<unsigned>::max();
    5673          38 :     if (!UseSplat) {
    5674             :       // If we cannot determine where the constant must be, we have to
    5675             :       // use a splat constant.
    5676          32 :       Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
    5677             :       if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
    5678          31 :         ExtractIdx = CstVal->getSExtValue();
    5679             :       else
    5680             :         UseSplat = true;
    5681             :     }
    5682             : 
    5683          38 :     unsigned End = getTransitionType()->getVectorNumElements();
    5684          38 :     if (UseSplat)
    5685           7 :       return ConstantVector::getSplat(End, Val);
    5686             : 
    5687             :     SmallVector<Constant *, 4> ConstVec;
    5688          31 :     UndefValue *UndefVal = UndefValue::get(Val->getType());
    5689         179 :     for (unsigned Idx = 0; Idx != End; ++Idx) {
    5690          74 :       if (Idx == ExtractIdx)
    5691          31 :         ConstVec.push_back(Val);
    5692             :       else
    5693          43 :         ConstVec.push_back(UndefVal);
    5694             :     }
    5695          31 :     return ConstantVector::get(ConstVec);
    5696             :   }
    5697             : 
    5698             :   /// \brief Check if promoting to a vector type an operand at \p OperandIdx
    5699             :   /// in \p Use can trigger undefined behavior.
    5700          95 :   static bool canCauseUndefinedBehavior(const Instruction *Use,
    5701             :                                         unsigned OperandIdx) {
    5702             :     // This is not safe to introduce undef when the operand is on
    5703             :     // the right hand side of a division-like instruction.
    5704          95 :     if (OperandIdx != 1)
    5705             :       return false;
    5706             :     switch (Use->getOpcode()) {
    5707             :     default:
    5708             :       return false;
    5709          10 :     case Instruction::SDiv:
    5710             :     case Instruction::UDiv:
    5711             :     case Instruction::SRem:
    5712             :     case Instruction::URem:
    5713          10 :       return true;
    5714           4 :     case Instruction::FDiv:
    5715             :     case Instruction::FRem:
    5716           4 :       return !Use->hasNoNaNs();
    5717             :     }
    5718             :     llvm_unreachable(nullptr);
    5719             :   }
    5720             : 
    5721             : public:
    5722             :   VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
    5723             :                       const TargetTransformInfo &TTI, Instruction *Transition,
    5724             :                       unsigned CombineCost)
    5725         136 :       : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
    5726         272 :         StoreExtractCombineCost(CombineCost) {
    5727             :     assert(Transition && "Do not know how to promote null");
    5728             :   }
    5729             : 
    5730             :   /// \brief Check if we can promote \p ToBePromoted to \p Type.
    5731             :   bool canPromote(const Instruction *ToBePromoted) const {
    5732             :     // We could support CastInst too.
    5733             :     return isa<BinaryOperator>(ToBePromoted);
    5734             :   }
    5735             : 
    5736             :   /// \brief Check if it is profitable to promote \p ToBePromoted
    5737             :   /// by moving downward the transition through.
    5738          57 :   bool shouldPromote(const Instruction *ToBePromoted) const {
    5739             :     // Promote only if all the operands can be statically expanded.
    5740             :     // Indeed, we do not want to introduce any new kind of transitions.
    5741         330 :     for (const Use &U : ToBePromoted->operands()) {
    5742         114 :       const Value *Val = U.get();
    5743         165 :       if (Val == getEndOfTransition()) {
    5744             :         // If the use is a division and the transition is on the rhs,
    5745             :         // we cannot promote the operation, otherwise we may create a
    5746             :         // division by zero.
    5747          57 :         if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
    5748             :           return false;
    5749             :         continue;
    5750             :       }
    5751          57 :       if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
    5752             :           !isa<ConstantFP>(Val))
    5753             :         return false;
    5754             :     }
    5755             :     // Check that the resulting operation is legal.
    5756         102 :     int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
    5757          51 :     if (!ISDOpcode)
    5758             :       return false;
    5759          81 :     return StressStoreExtract ||
    5760          60 :            TLI.isOperationLegalOrCustom(
    5761             :                ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
    5762             :   }
    5763             : 
    5764             :   /// \brief Check whether or not \p Use can be combined
    5765             :   /// with the transition.
    5766             :   /// I.e., is it possible to do Use(Transition) => AnotherUse?
    5767             :   bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
    5768             : 
    5769             :   /// \brief Record \p ToBePromoted as part of the chain to be promoted.
    5770             :   void enqueueForPromotion(Instruction *ToBePromoted) {
    5771          41 :     InstsToBePromoted.push_back(ToBePromoted);
    5772             :   }
    5773             : 
    5774             :   /// \brief Set the instruction that will be combined with the transition.
    5775             :   void recordCombineInstruction(Instruction *ToBeCombined) {
    5776             :     assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
    5777          73 :     CombineInst = ToBeCombined;
    5778             :   }
    5779             : 
    5780             :   /// \brief Promote all the instructions enqueued for promotion if it is
    5781             :   /// is profitable.
    5782             :   /// \return True if the promotion happened, false otherwise.
    5783          73 :   bool promote() {
    5784             :     // Check if there is something to promote.
    5785             :     // Right now, if we do not have anything to combine with,
    5786             :     // we assume the promotion is not profitable.
    5787          73 :     if (InstsToBePromoted.empty() || !CombineInst)
    5788             :       return false;
    5789             : 
    5790             :     // Check cost.
    5791          20 :     if (!StressStoreExtract && !isProfitableToPromote())
    5792             :       return false;
    5793             : 
    5794             :     // Promote.
    5795          96 :     for (auto &ToBePromoted : InstsToBePromoted)
    5796          38 :       promoteImpl(ToBePromoted);
    5797             :     InstsToBePromoted.clear();
    5798          20 :     return true;
    5799             :   }
    5800             : };
    5801             : 
    5802             : } // end anonymous namespace
    5803             : 
    5804          38 : void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
    5805             :   // At this point, we know that all the operands of ToBePromoted but Def
    5806             :   // can be statically promoted.
    5807             :   // For Def, we need to use its parameter in ToBePromoted:
    5808             :   // b = ToBePromoted ty1 a
    5809             :   // Def = Transition ty1 b to ty2
    5810             :   // Move the transition down.
    5811             :   // 1. Replace all uses of the promoted operation by the transition.
    5812             :   // = ... b => = ... Def.
    5813             :   assert(ToBePromoted->getType() == Transition->getType() &&
    5814             :          "The type of the result of the transition does not match "
    5815             :          "the final type");
    5816          38 :   ToBePromoted->replaceAllUsesWith(Transition);
    5817             :   // 2. Update the type of the uses.
    5818             :   // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
    5819          38 :   Type *TransitionTy = getTransitionType();
    5820             :   ToBePromoted->mutateType(TransitionTy);
    5821             :   // 3. Update all the operands of the promoted operation with promoted
    5822             :   // operands.
    5823             :   // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
    5824         228 :   for (Use &U : ToBePromoted->operands()) {
    5825          76 :     Value *Val = U.get();
    5826             :     Value *NewVal = nullptr;
    5827          76 :     if (Val == Transition)
    5828          38 :       NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
    5829          38 :     else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
    5830             :              isa<ConstantFP>(Val)) {
    5831             :       // Use a splat constant if it is not safe to use undef.
    5832          38 :       NewVal = getConstantVector(
    5833             :           cast<Constant>(Val),
    5834          76 :           isa<UndefValue>(Val) ||
    5835          38 :               canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
    5836             :     } else
    5837           0 :       llvm_unreachable("Did you modified shouldPromote and forgot to update "
    5838             :                        "this?");
    5839          76 :     ToBePromoted->setOperand(U.getOperandNo(), NewVal);
    5840             :   }
    5841          38 :   Transition->moveAfter(ToBePromoted);
    5842          38 :   Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
    5843          38 : }
    5844             : 
    5845             : /// Some targets can do store(extractelement) with one instruction.
    5846             : /// Try to push the extractelement towards the stores when the target
    5847             : /// has this feature and this is profitable.
    5848       18342 : bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
    5849       18342 :   unsigned CombineCost = std::numeric_limits<unsigned>::max();
    5850       36684 :   if (DisableStoreExtract || !TLI ||
    5851       18295 :       (!StressStoreExtract &&
    5852       54885 :        !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
    5853       18295 :                                        Inst->getOperand(1), CombineCost)))
    5854             :     return false;
    5855             : 
    5856             :   // At this point we know that Inst is a vector to scalar transition.
    5857             :   // Try to move it down the def-use chain, until:
    5858             :   // - We can combine the transition with its single use
    5859             :   //   => we got rid of the transition.
    5860             :   // - We escape the current basic block
    5861             :   //   => we would need to check that we are moving it at a cheaper place and
    5862             :   //      we do not do that for now.
    5863         136 :   BasicBlock *Parent = Inst->getParent();
    5864             :   DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
    5865         136 :   VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
    5866             :   // If the transition has more than one use, assume this is not going to be
    5867             :   // beneficial.
    5868         218 :   while (Inst->hasOneUse()) {
    5869             :     Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
    5870             :     DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
    5871             : 
    5872         164 :     if (ToBePromoted->getParent() != Parent) {
    5873             :       DEBUG(dbgs() << "Instruction to promote is in a different block ("
    5874             :                    << ToBePromoted->getParent()->getName()
    5875             :                    << ") than the transition (" << Parent->getName() << ").\n");
    5876             :       return false;
    5877             :     }
    5878             : 
    5879         161 :     if (VPH.canCombine(ToBePromoted)) {
    5880             :       DEBUG(dbgs() << "Assume " << *Inst << '\n'
    5881             :                    << "will be combined with: " << *ToBePromoted << '\n');
    5882             :       VPH.recordCombineInstruction(ToBePromoted);
    5883          73 :       bool Changed = VPH.promote();
    5884             :       NumStoreExtractExposed += Changed;
    5885          73 :       return Changed;
    5886             :     }
    5887             : 
    5888             :     DEBUG(dbgs() << "Try promoting.\n");
    5889          88 :     if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
    5890             :       return false;
    5891             : 
    5892             :     DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
    5893             : 
    5894             :     VPH.enqueueForPromotion(ToBePromoted);
    5895             :     Inst = ToBePromoted;
    5896             :   }
    5897             :   return false;
    5898             : }
    5899             : 
    5900             : /// For the instruction sequence of store below, F and I values
    5901             : /// are bundled together as an i64 value before being stored into memory.
    5902             : /// Sometimes it is more efficent to generate separate stores for F and I,
    5903             : /// which can remove the bitwise instructions or sink them to colder places.
    5904             : ///
    5905             : ///   (store (or (zext (bitcast F to i32) to i64),
    5906             : ///              (shl (zext I to i64), 32)), addr)  -->
    5907             : ///   (store F, addr) and (store I, addr+4)
    5908             : ///
    5909             : /// Similarly, splitting for other merged store can also be beneficial, like:
    5910             : /// For pair of {i32, i32}, i64 store --> two i32 stores.
    5911             : /// For pair of {i32, i16}, i64 store --> two i32 stores.
    5912             : /// For pair of {i16, i16}, i32 store --> two i16 stores.
    5913             : /// For pair of {i16, i8},  i32 store --> two i16 stores.
    5914             : /// For pair of {i8, i8},   i16 store --> two i8 stores.
    5915             : ///
    5916             : /// We allow each target to determine specifically which kind of splitting is
    5917             : /// supported.
    5918             : ///
    5919             : /// The store patterns are commonly seen from the simple code snippet below
    5920             : /// if only std::make_pair(...) is sroa transformed before inlined into hoo.
    5921             : ///   void goo(const std::pair<int, float> &);
    5922             : ///   hoo() {
    5923             : ///     ...
    5924             : ///     goo(std::make_pair(tmp, ftmp));
    5925             : ///     ...
    5926             : ///   }
    5927             : ///
    5928             : /// Although we already have similar splitting in DAG Combine, we duplicate
    5929             : /// it in CodeGenPrepare to catch the case in which pattern is across
    5930             : /// multiple BBs. The logic in DAG Combine is kept to catch case generated
    5931             : /// during code expansion.
    5932      524479 : static bool splitMergedValStore(StoreInst &SI, const DataLayout &DL,
    5933             :                                 const TargetLowering &TLI) {
    5934             :   // Handle simple but common cases only.
    5935      524479 :   Type *StoreType = SI.getValueOperand()->getType();
    5936     1047480 :   if (DL.getTypeStoreSizeInBits(StoreType) != DL.getTypeSizeInBits(StoreType) ||
    5937      523001 :       DL.getTypeSizeInBits(StoreType) == 0)
    5938             :     return false;
    5939             : 
    5940      522980 :   unsigned HalfValBitSize = DL.getTypeSizeInBits(StoreType) / 2;
    5941      522980 :   Type *SplitStoreType = Type::getIntNTy(SI.getContext(), HalfValBitSize);
    5942      522980 :   if (DL.getTypeStoreSizeInBits(SplitStoreType) !=
    5943      522980 :       DL.getTypeSizeInBits(SplitStoreType))
    5944             :     return false;
    5945             : 
    5946             :   // Match the following patterns:
    5947             :   // (store (or (zext LValue to i64),
    5948             :   //            (shl (zext HValue to i64), 32)), HalfValBitSize)
    5949             :   //  or
    5950             :   // (store (or (shl (zext HValue to i64), 32)), HalfValBitSize)
    5951             :   //            (zext LValue to i64),
    5952             :   // Expect both operands of OR and the first operand of SHL have only
    5953             :   // one use.
    5954             :   Value *LValue, *HValue;
    5955      512542 :   if (!match(SI.getValueOperand(),
    5956      512542 :              m_c_Or(m_OneUse(m_ZExt(m_Value(LValue))),
    5957             :                     m_OneUse(m_Shl(m_OneUse(m_ZExt(m_Value(HValue))),
    5958             :                                    m_SpecificInt(HalfValBitSize))))))
    5959             :     return false;
    5960             : 
    5961             :   // Check LValue and HValue are int with size less or equal than 32.
    5962         136 :   if (!LValue->getType()->isIntegerTy() ||
    5963         136 :       DL.getTypeSizeInBits(LValue->getType()) > HalfValBitSize ||
    5964         204 :       !HValue->getType()->isIntegerTy() ||
    5965          68 :       DL.getTypeSizeInBits(HValue->getType()) > HalfValBitSize)
    5966             :     return false;
    5967             : 
    5968             :   // If LValue/HValue is a bitcast instruction, use the EVT before bitcast
    5969             :   // as the input of target query.
    5970          68 :   auto *LBC = dyn_cast<BitCastInst>(LValue);
    5971          68 :   auto *HBC = dyn_cast<BitCastInst>(HValue);
    5972           1 :   EVT LowTy = LBC ? EVT::getEVT(LBC->getOperand(0)->getType())
    5973          69 :                   : EVT::getEVT(LValue->getType());
    5974           6 :   EVT HighTy = HBC ? EVT::getEVT(HBC->getOperand(0)->getType())
    5975          74 :                    : EVT::getEVT(HValue->getType());
    5976          68 :   if (!ForceSplitStore && !TLI.isMultiStoresCheaperThanBitsMerge(LowTy, HighTy))
    5977             :     return false;
    5978             : 
    5979             :   // Start to split store.
    5980          14 :   IRBuilder<> Builder(SI.getContext());
    5981          14 :   Builder.SetInsertPoint(&SI);
    5982             : 
    5983             :   // If LValue/HValue is a bitcast in another BB, create a new one in current
    5984             :   // BB so it may be merged with the splitted stores by dag combiner.
    5985          14 :   if (LBC && LBC->getParent() != SI.getParent())
    5986           0 :     LValue = Builder.CreateBitCast(LBC->getOperand(0), LBC->getType());
    5987          14 :   if (HBC && HBC->getParent() != SI.getParent())
    5988           6 :     HValue = Builder.CreateBitCast(HBC->getOperand(0), HBC->getType());
    5989             : 
    5990          28 :   auto CreateSplitStore = [&](Value *V, bool Upper) {
    5991         140 :     V = Builder.CreateZExtOrBitCast(V, SplitStoreType);
    5992          28 :     Value *Addr = Builder.CreateBitCast(
    5993          98 :         SI.getOperand(1),
    5994          56 :         SplitStoreType->getPointerTo(SI.getPointerAddressSpace()));
    5995          28 :     if (Upper)
    5996          14 :       Addr = Builder.CreateGEP(
    5997             :           SplitStoreType, Addr,
    5998          14 :           ConstantInt::get(Type::getInt32Ty(SI.getContext()), 1));
    5999          42 :     Builder.CreateAlignedStore(
    6000             :         V, Addr, Upper ? SI.getAlignment() / 2 : SI.getAlignment());
    6001          42 :   };
    6002             : 
    6003          14 :   CreateSplitStore(LValue, false);
    6004          14 :   CreateSplitStore(HValue, true);
    6005             : 
    6006             :   // Delete the old store.
    6007          14 :   SI.eraseFromParent();
    6008             :   return true;
    6009             : }
    6010             : 
    6011             : // Return true if the GEP has two operands, the first operand is of a sequential
    6012             : // type, and the second operand is a constant.
    6013          78 : static bool GEPSequentialConstIndexed(GetElementPtrInst *GEP) {
    6014          78 :   gep_type_iterator I = gep_type_begin(*GEP);
    6015          21 :   return GEP->getNumOperands() == 2 &&
    6016          99 :       I.isSequential() &&
    6017          78 :       isa<ConstantInt>(GEP->getOperand(1));
    6018             : }
    6019             : 
    6020             : // Try unmerging GEPs to reduce liveness interference (register pressure) across
    6021             : // IndirectBr edges. Since IndirectBr edges tend to touch on many blocks,
    6022             : // reducing liveness interference across those edges benefits global register
    6023             : // allocation. Currently handles only certain cases.
    6024             : //
    6025             : // For example, unmerge %GEPI and %UGEPI as below.
    6026             : //
    6027             : // ---------- BEFORE ----------
    6028             : // SrcBlock:
    6029             : //   ...
    6030             : //   %GEPIOp = ...
    6031             : //   ...
    6032             : //   %GEPI = gep %GEPIOp, Idx
    6033             : //   ...
    6034             : //   indirectbr ... [ label %DstB0, label %DstB1, ... label %DstBi ... ]
    6035             : //   (* %GEPI is alive on the indirectbr edges due to other uses ahead)
    6036             : //   (* %GEPIOp is alive on the indirectbr edges only because of it's used by
    6037             : //   %UGEPI)
    6038             : //
    6039             : // DstB0: ... (there may be a gep similar to %UGEPI to be unmerged)
    6040             : // DstB1: ... (there may be a gep similar to %UGEPI to be unmerged)
    6041             : // ...
    6042             : //
    6043             : // DstBi:
    6044             : //   ...
    6045             : //   %UGEPI = gep %GEPIOp, UIdx
    6046             : // ...
    6047             : // ---------------------------
    6048             : //
    6049             : // ---------- AFTER ----------
    6050             : // SrcBlock:
    6051             : //   ... (same as above)
    6052             : //    (* %GEPI is still alive on the indirectbr edges)
    6053             : //    (* %GEPIOp is no longer alive on the indirectbr edges as a result of the
    6054             : //    unmerging)
    6055             : // ...
    6056             : //
    6057             : // DstBi:
    6058             : //   ...
    6059             : //   %UGEPI = gep %GEPI, (UIdx-Idx)
    6060             : //   ...
    6061             : // ---------------------------
    6062             : //
    6063             : // The register pressure on the IndirectBr edges is reduced because %GEPIOp is
    6064             : // no longer alive on them.
    6065             : //
    6066             : // We try to unmerge GEPs here in CodGenPrepare, as opposed to limiting merging
    6067             : // of GEPs in the first place in InstCombiner::visitGetElementPtrInst() so as
    6068             : // not to disable further simplications and optimizations as a result of GEP
    6069             : // merging.
    6070             : //
    6071             : // Note this unmerging may increase the length of the data flow critical path
    6072             : // (the path from %GEPIOp to %UGEPI would go through %GEPI), which is a tradeoff
    6073             : // between the register pressure and the length of data-flow critical
    6074             : // path. Restricting this to the uncommon IndirectBr case would minimize the
    6075             : // impact of potentially longer critical path, if any, and the impact on compile
    6076             : // time.
    6077      154231 : static bool tryUnmergingGEPsAcrossIndirectBr(GetElementPtrInst *GEPI,
    6078             :                                              const TargetTransformInfo *TTI) {
    6079      154231 :   BasicBlock *SrcBlock = GEPI->getParent();
    6080             :   // Check that SrcBlock ends with an IndirectBr. If not, give up. The common
    6081             :   // (non-IndirectBr) cases exit early here.
    6082      154231 :   if (!isa<IndirectBrInst>(SrcBlock->getTerminator()))
    6083             :     return false;
    6084             :   // Check that GEPI is a simple gep with a single constant index.
    6085          75 :   if (!GEPSequentialConstIndexed(GEPI))
    6086             :     return false;
    6087             :   ConstantInt *GEPIIdx = cast<ConstantInt>(GEPI->getOperand(1));
    6088             :   // Check that GEPI is a cheap one.
    6089           8 :   if (TTI->getIntImmCost(GEPIIdx->getValue(), GEPIIdx->getType())
    6090             :       > TargetTransformInfo::TCC_Basic)
    6091             :     return false;
    6092             :   Value *GEPIOp = GEPI->getOperand(0);
    6093             :   // Check that GEPIOp is an instruction that's also defined in SrcBlock.
    6094           8 :   if (!isa<Instruction>(GEPIOp))
    6095             :     return false;
    6096             :   auto *GEPIOpI = cast<Instruction>(GEPIOp);
    6097           8 :   if (GEPIOpI->getParent() != SrcBlock)
    6098             :     return false;
    6099             :   // Check that GEP is used outside the block, meaning it's alive on the
    6100             :   // IndirectBr edge(s).
    6101           8 :   if (find_if(GEPI->users(), [&](User *Usr) {
    6102             :         if (auto *I = dyn_cast<Instruction>(Usr)) {
    6103           9 :           if (I->getParent() != SrcBlock) {
    6104             :             return true;
    6105             :           }
    6106             :         }
    6107             :         return false;
    6108             :       }) == GEPI->users().end())
    6109             :     return false;
    6110             :   // The second elements of the GEP chains to be unmerged.
    6111             :   std::vector<GetElementPtrInst *> UGEPIs;
    6112             :   // Check each user of GEPIOp to check if unmerging would make GEPIOp not alive
    6113             :   // on IndirectBr edges.
    6114          21 :   for (User *Usr : GEPIOp->users()) {
    6115          27 :     if (Usr == GEPI) continue;
    6116             :     // Check if Usr is an Instruction. If not, give up.
    6117           9 :     if (!isa<Instruction>(Usr))
    6118           0 :       return false;
    6119             :     auto *UI = cast<Instruction>(Usr);
    6120             :     // Check if Usr in the same block as GEPIOp, which is fine, skip.
    6121           9 :     if (UI->getParent() == SrcBlock)
    6122           6 :       continue;
    6123             :     // Check if Usr is a GEP. If not, give up.
    6124             :     if (!isa<GetElementPtrInst>(Usr))
    6125             :       return false;
    6126           3 :     auto *UGEPI = cast<GetElementPtrInst>(Usr);
    6127             :     // Check if UGEPI is a simple gep with a single constant index and GEPIOp is
    6128             :     // the pointer operand to it. If so, record it in the vector. If not, give
    6129             :     // up.
    6130           3 :     if (!GEPSequentialConstIndexed(UGEPI))
    6131             :       return false;
    6132           6 :     if (UGEPI->getOperand(0) != GEPIOp)
    6133             :       return false;
    6134           3 :     if (GEPIIdx->getType() !=
    6135             :         cast<ConstantInt>(UGEPI->getOperand(1))->getType())
    6136             :       return false;
    6137             :     ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
    6138           3 :     if (TTI->getIntImmCost(UGEPIIdx->getValue(), UGEPIIdx->getType())
    6139             :         > TargetTransformInfo::TCC_Basic)
    6140             :       return false;
    6141           3 :     UGEPIs.push_back(UGEPI);
    6142             :   }
    6143          12 :   if (UGEPIs.size() == 0)
    6144             :     return false;
    6145             :   // Check the materializing cost of (Uidx-Idx).
    6146           4 :   for (GetElementPtrInst *UGEPI : UGEPIs) {
    6147             :     ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
    6148           3 :     APInt NewIdx = UGEPIIdx->getValue() - GEPIIdx->getValue();
    6149           3 :     unsigned ImmCost = TTI->getIntImmCost(NewIdx, GEPIIdx->getType());
    6150           3 :     if (ImmCost > TargetTransformInfo::TCC_Basic)
    6151             :       return false;
    6152             :   }
    6153             :   // Now unmerge between GEPI and UGEPIs.
    6154           4 :   for (GetElementPtrInst *UGEPI : UGEPIs) {
    6155             :     UGEPI->setOperand(0, GEPI);
    6156             :     ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
    6157             :     Constant *NewUGEPIIdx =
    6158             :         ConstantInt::get(GEPIIdx->getType(),
    6159           9 :                          UGEPIIdx->getValue() - GEPIIdx->getValue());
    6160             :     UGEPI->setOperand(1, NewUGEPIIdx);
    6161             :     // If GEPI is not inbounds but UGEPI is inbounds, change UGEPI to not
    6162             :     // inbounds to avoid UB.
    6163           3 :     if (!GEPI->isInBounds()) {
    6164           3 :       UGEPI->setIsInBounds(false);
    6165             :     }
    6166             :   }
    6167             :   // After unmerging, verify that GEPIOp is actually only used in SrcBlock (not
    6168             :   // alive on IndirectBr edges).
    6169             :   assert(find_if(GEPIOp->users(), [&](User *Usr) {
    6170             :         return cast<Instruction>(Usr)->getParent() != SrcBlock;
    6171             :       }) == GEPIOp->users().end() && "GEPIOp is used outside SrcBlock");
    6172             :   return true;
    6173             : }
    6174             : 
    6175     3444007 : bool CodeGenPrepare::optimizeInst(Instruction *I, bool &ModifiedDT) {
    6176             :   // Bail out if we inserted the instruction to prevent optimizations from
    6177             :   // stepping on each other's toes.
    6178     3444007 :   if (InsertedInsts.count(I))
    6179             :     return false;
    6180             : 
    6181     3443333 :   if (PHINode *P = dyn_cast<PHINode>(I)) {
    6182             :     // It is possible for very late stage optimizations (such as SimplifyCFG)
    6183             :     // to introduce PHI nodes too late to be cleaned up.  If we detect such a
    6184             :     // trivial PHI, go ahead and zap it here.
    6185      160338 :     if (Value *V = SimplifyInstruction(P, {*DL, TLInfo})) {
    6186         408 :       P->replaceAllUsesWith(V);
    6187         408 :       P->eraseFromParent();
    6188             :       ++NumPHIsElim;
    6189         408 :       return true;
    6190             :     }
    6191             :     return false;
    6192             :   }
    6193             : 
    6194             :   if (CastInst *CI = dyn_cast<CastInst>(I)) {
    6195             :     // If the source of the cast is a constant, then this should have
    6196             :     // already been constant folded.  The only reason NOT to constant fold
    6197             :     // it is if something (e.g. LSR) was careful to place the constant
    6198             :     // evaluation in a block other than then one that uses it (e.g. to hoist
    6199             :     // the address of globals out of a loop).  If this is the case, we don't
    6200             :     // want to forward-subst the cast.
    6201      292438 :     if (isa<Constant>(CI->getOperand(0)))
    6202             :       return false;
    6203             : 
    6204      289314 :     if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
    6205             :       return true;
    6206             : 
    6207      535958 :     if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
    6208             :       /// Sink a zext or sext into its user blocks if the target type doesn't
    6209             :       /// fit in one register
    6210       64972 :       if (TLI &&
    6211       64970 :           TLI->getTypeAction(CI->getContext(),
    6212       32485 :                              TLI->getValueType(*DL, CI->getType())) ==
    6213             :               TargetLowering::TypeExpandInteger) {
    6214        3355 :         return SinkCast(CI);
    6215             :       } else {
    6216       29132 :         bool MadeChange = optimizeExt(I);
    6217       29132 :         return MadeChange | optimizeExtUses(I);
    6218             :       }
    6219             :     }
    6220             :     return false;
    6221             :   }
    6222             : 
    6223             :   if (CmpInst *CI = dyn_cast<CmpInst>(I))
    6224       96205 :     if (!TLI || !TLI->hasMultipleConditionRegisters())
    6225       89406 :       return OptimizeCmpExpression(CI, TLI);
    6226             : 
    6227             :   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
    6228      507173 :     LI->setMetadata(LLVMContext::MD_invariant_group, nullptr);
    6229      507173 :     if (TLI) {
    6230      507148 :       bool Modified = optimizeLoadExt(LI);
    6231             :       unsigned AS = LI->getPointerAddressSpace();
    6232     1014296 :       Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
    6233      507148 :       return Modified;
    6234             :     }
    6235             :     return false;
    6236             :   }
    6237             : 
    6238             :   if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
    6239      524496 :     if (TLI && splitMergedValStore(*SI, *DL, *TLI))
    6240             :       return true;
    6241      524482 :     SI->setMetadata(LLVMContext::MD_invariant_group, nullptr);
    6242      524482 :     if (TLI) {
    6243             :       unsigned AS = SI->getPointerAddressSpace();
    6244      524465 :       return optimizeMemoryInst(I, SI->getOperand(1),
    6245      524465 :                                 SI->getOperand(0)->getType(), AS);
    6246             :     }
    6247             :     return false;
    6248             :   }
    6249             : 
    6250             :   if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
    6251             :       unsigned AS = RMW->getPointerAddressSpace();
    6252        4727 :       return optimizeMemoryInst(I, RMW->getPointerOperand(),
    6253        4727 :                                 RMW->getType(), AS);
    6254             :   }
    6255             : 
    6256             :   if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(I)) {
    6257             :       unsigned AS = CmpX->getPointerAddressSpace();
    6258        1330 :       return optimizeMemoryInst(I, CmpX->getPointerOperand(),
    6259        1330 :                                 CmpX->getCompareOperand()->getType(), AS);
    6260             :   }
    6261             : 
    6262             :   BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
    6263             : 
    6264      516418 :   if (BinOp && (BinOp->getOpcode() == Instruction::And) &&
    6265       15124 :       EnableAndCmpSinking && TLI)
    6266       15124 :     return sinkAndCmp0Expression(BinOp, *TLI, InsertedInsts);
    6267             : 
    6268     2414640 :   if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
    6269             :                 BinOp->getOpcode() == Instruction::LShr)) {
    6270             :     ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
    6271       12216 :     if (TLI && CI && TLI->hasExtractBitsInsn())
    6272        1122 :       return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
    6273             : 
    6274             :     return false;
    6275             :   }
    6276             : 
    6277             :   if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
    6278      169271 :     if (GEPI->hasAllZeroIndices()) {
    6279             :       /// The GEP operand must be a pointer, so must its result -> BitCast
    6280       15040 :       Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
    6281       30080 :                                         GEPI->getName(), GEPI);
    6282       15040 :       GEPI->replaceAllUsesWith(NC);
    6283       15040 :       GEPI->eraseFromParent();
    6284             :       ++NumGEPsElim;
    6285       15040 :       optimizeInst(NC, ModifiedDT);
    6286       15040 :       return true;
    6287             :     }
    6288      154231 :     if (tryUnmergingGEPsAcrossIndirectBr(GEPI, TTI)) {
    6289             :       return true;
    6290             :     }
    6291      154230 :     return false;
    6292             :   }
    6293             : 
    6294             :   if (CallInst *CI = dyn_cast<CallInst>(I))
    6295      506507 :     return optimizeCallInst(CI, ModifiedDT);
    6296             : 
    6297             :   if (SelectInst *SI = dyn_cast<SelectInst>(I))
    6298       70657 :     return optimizeSelectInst(SI);
    6299             : 
    6300             :   if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
    6301       35581 :     return optimizeShuffleVectorInst(SVI);
    6302             : 
    6303             :   if (auto *Switch = dyn_cast<SwitchInst>(I))
    6304        1365 :     return optimizeSwitchInst(Switch);
    6305             : 
    6306     1132873 :   if (isa<ExtractElementInst>(I))
    6307       18342 :     return optimizeExtractElementInst(I);
    6308             : 
    6309             :   return false;
    6310             : }
    6311             : 
    6312             : /// Given an OR instruction, check to see if this is a bitreverse
    6313             : /// idiom. If so, insert the new intrinsic and return true.
    6314     3461271 : static bool makeBitReverse(Instruction &I, const DataLayout &DL,
    6315             :                            const TargetLowering &TLI) {
    6316     6922542 :   if (!I.getType()->isIntegerTy() ||
    6317      906046 :       !TLI.isOperationLegalOrCustom(ISD::BITREVERSE,
    6318             :                                     TLI.getValueType(DL, I.getType(), true)))
    6319             :     return false;
    6320             : 
    6321             :   SmallVector<Instruction*, 4> Insts;
    6322       47943 :   if (!recognizeBSwapOrBitReverseIdiom(&I, false, true, Insts))
    6323             :     return false;
    6324           6 :   Instruction *LastInst = Insts.back();
    6325           6 :   I.replaceAllUsesWith(LastInst);
    6326           6 :   RecursivelyDeleteTriviallyDeadInstructions(&I);
    6327           6 :   return true;
    6328             : }
    6329             : 
    6330             : // In this pass we look for GEP and cast instructions that are used
    6331             : // across basic blocks and rewrite them to improve basic-block-at-a-time
    6332             : // selection.
    6333      436804 : bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool &ModifiedDT) {
    6334             :   SunkAddrs.clear();
    6335             :   bool MadeChange = false;
    6336             : 
    6337      436804 :   CurInstIterator = BB.begin();
    6338     3865743 :   while (CurInstIterator != BB.end()) {
    6339     3428968 :     MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
    6340     3428968 :     if (ModifiedDT)
    6341             :       return true;
    6342             :   }
    6343             : 
    6344             :   bool MadeBitReverse = true;
    6345      873394 :   while (TLI && MadeBitReverse) {
    6346             :     MadeBitReverse = false;
    6347     3897885 :     for (auto &I : reverse(BB)) {
    6348     3461272 :       if (makeBitReverse(I, *DL, *TLI)) {
    6349             :         MadeBitReverse = MadeChange = true;
    6350           6 :         ModifiedDT = true;
    6351           6 :         break;
    6352             :       }
    6353             :     }
    6354             :   }
    6355      436775 :   MadeChange |= dupRetToEnableTailCallOpts(&BB);
    6356             : 
    6357      436775 :   return MadeChange;
    6358             : }
    6359             : 
    6360             : // llvm.dbg.value is far away from the value then iSel may not be able
    6361             : // handle it properly. iSel will drop llvm.dbg.value if it can not
    6362             : // find a node corresponding to the value.
    6363      160412 : bool CodeGenPrepare::placeDbgValues(Function &F) {
    6364             :   bool MadeChange = false;
    6365      464764 :   for (BasicBlock &BB : F) {
    6366             :     Instruction *PrevNonDbgInst = nullptr;
    6367     2339972 :     for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
    6368             :       Instruction *Insn = &*BI++;
    6369             :       DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
    6370             :       // Leave dbg.values that refer to an alloca alone. These
    6371             :       // intrinsics describe the address of a variable (= the alloca)
    6372             :       // being taken.  They should not be moved next to the alloca
    6373             :       // (and to the beginning of the scope), but rather stay close to
    6374             :       // where said address is used.
    6375     2042890 :       if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
    6376             :         PrevNonDbgInst = Insn;
    6377             :         continue;
    6378             :       }
    6379             : 
    6380             :       Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
    6381       40533 :       if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
    6382             :         // If VI is a phi in a block with an EHPad terminator, we can't insert
    6383             :         // after it.
    6384        8922 :         if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
    6385             :           continue;
    6386             :         DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
    6387        8920 :         DVI->removeFromParent();
    6388        8920 :         if (isa<PHINode>(VI))
    6389        5266 :           DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
    6390             :         else
    6391        6287 :           DVI->insertAfter(VI);
    6392             :         MadeChange = true;
    6393             :         ++NumDbgValueMoved;
    6394             :       }
    6395             :     }
    6396             :   }
    6397      160412 :   return MadeChange;
    6398             : }
    6399             : 
    6400             : /// \brief Scale down both weights to fit into uint32_t.
    6401             : static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
    6402             :   uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
    6403             :   uint32_t Scale = (NewMax / std::numeric_limits<uint32_t>::max()) + 1;
    6404             :   NewTrue = NewTrue / Scale;
    6405             :   NewFalse = NewFalse / Scale;
    6406             : }
    6407             : 
    6408             : /// \brief Some targets prefer to split a conditional branch like:
    6409             : /// \code
    6410             : ///   %0 = icmp ne i32 %a, 0
    6411             : ///   %1 = icmp ne i32 %b, 0
    6412             : ///   %or.cond = or i1 %0, %1
    6413             : ///   br i1 %or.cond, label %TrueBB, label %FalseBB
    6414             : /// \endcode
    6415             : /// into multiple branch instructions like:
    6416             : /// \code
    6417             : ///   bb1:
    6418             : ///     %0 = icmp ne i32 %a, 0
    6419             : ///     br i1 %0, label %TrueBB, label %bb2
    6420             : ///   bb2:
    6421             : ///     %1 = icmp ne i32 %b, 0
    6422             : ///     br i1 %1, label %TrueBB, label %FalseBB
    6423             : /// \endcode
    6424             : /// This usually allows instruction selection to do even further optimizations
    6425             : /// and combine the compare with the branch instruction. Currently this is
    6426             : /// applied for targets which have "cheap" jump instructions.
    6427             : ///
    6428             : /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
    6429             : ///
    6430      160349 : bool CodeGenPrepare::splitBranchCondition(Function &F) {
    6431      160349 :   if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
    6432             :     return false;
    6433             : 
    6434             :   bool MadeChange = false;
    6435        8555 :   for (auto &BB : F) {
    6436             :     // Does this BB end with the following?
    6437             :     //   %cond1 = icmp|fcmp|binary instruction ...
    6438             :     //   %cond2 = icmp|fcmp|binary instruction ...
    6439             :     //   %cond.or = or|and i1 %cond1, cond2
    6440             :     //   br i1 %cond.or label %dest1, label %dest2"
    6441             :     BinaryOperator *LogicOp;
    6442             :     BasicBlock *TBB, *FBB;
    6443        4462 :     if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
    6444        8916 :       continue;
    6445             : 
    6446             :     auto *Br1 = cast<BranchInst>(BB.getTerminator());
    6447           9 :     if (Br1->getMetadata(LLVMContext::MD_unpredictable))
    6448           2 :       continue;
    6449             : 
    6450             :     unsigned Opc;
    6451             :     Value *Cond1, *Cond2;
    6452           6 :     if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
    6453             :                              m_OneUse(m_Value(Cond2)))))
    6454             :       Opc = Instruction::And;
    6455           2 :     else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
    6456             :                                  m_OneUse(m_Value(Cond2)))))
    6457             :       Opc = Instruction::Or;
    6458             :     else
    6459           0 :       continue;
    6460             : 
    6461           3 :     if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
    6462           3 :         !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp()))   )
    6463           0 :       continue;
    6464             : 
    6465             :     DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
    6466             : 
    6467             :     // Create a new BB.
    6468             :     auto TmpBB =
    6469           9 :         BasicBlock::Create(BB.getContext(), BB.getName() + ".cond.split",
    6470             :                            BB.getParent(), BB.getNextNode());
    6471             : 
    6472             :     // Update original basic block by using the first condition directly by the
    6473             :     // branch instruction and removing the no longer needed and/or instruction.
    6474           3 :     Br1->setCondition(Cond1);
    6475           3 :     LogicOp->eraseFromParent();
    6476             : 
    6477             :     // Depending on the conditon we have to either replace the true or the false
    6478             :     // successor of the original branch instruction.
    6479           3 :     if (Opc == Instruction::And)
    6480             :       Br1->setSuccessor(0, TmpBB);
    6481             :     else
    6482             :       Br1->setSuccessor(1, TmpBB);
    6483             : 
    6484             :     // Fill in the new basic block.
    6485           6 :     auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
    6486           3 :     if (auto *I = dyn_cast<Instruction>(Cond2)) {
    6487           3 :       I->removeFromParent();
    6488           3 :       I->insertBefore(Br2);
    6489             :     }
    6490             : 
    6491             :     // Update PHI nodes in both successors. The original BB needs to be
    6492             :     // replaced in one successor's PHI nodes, because the branch comes now from
    6493             :     // the newly generated BB (NewBB). In the other successor we need to add one
    6494             :     // incoming edge to the PHI nodes, because both branch instructions target
    6495             :     // now the same successor. Depending on the original branch condition
    6496             :     // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
    6497             :     // we perform the correct update for the PHI nodes.
    6498             :     // This doesn't change the successor order of the just created branch
    6499             :     // instruction (or any other instruction).
    6500           3 :     if (Opc == Instruction::Or)
    6501             :       std::swap(TBB, FBB);
    6502             : 
    6503             :     // Replace the old BB with the new BB.
    6504           3 :     for (PHINode &PN : TBB->phis()) {
    6505             :       int i;
    6506           0 :       while ((i = PN.getBasicBlockIndex(&BB)) >= 0)
    6507           0 :         PN.setIncomingBlock(i, TmpBB);
    6508             :     }
    6509             : 
    6510             :     // Add another incoming edge form the new BB.
    6511           3 :     for (PHINode &PN : FBB->phis()) {
    6512           0 :       auto *Val = PN.getIncomingValueForBlock(&BB);
    6513           0 :       PN.addIncoming(Val, TmpBB);
    6514             :     }
    6515             : 
    6516             :     // Update the branch weights (from SelectionDAGBuilder::
    6517             :     // FindMergedConditions).
    6518           3 :     if (Opc == Instruction::Or) {
    6519             :       // Codegen X | Y as:
    6520             :       // BB1:
    6521             :       //   jmp_if_X TBB
    6522             :       //   jmp TmpBB
    6523             :       // TmpBB:
    6524             :       //   jmp_if_Y TBB
    6525             :       //   jmp FBB
    6526             :       //
    6527             : 
    6528             :       // We have flexibility in setting Prob for BB1 and Prob for NewBB.
    6529             :       // The requirement is that
    6530             :       //   TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
    6531             :       //     = TrueProb for orignal BB.
    6532             :       // Assuming the orignal weights are A and B, one choice is to set BB1's
    6533             :       // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
    6534             :       // assumes that
    6535             :       //   TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
    6536             :       // Another choice is to assume TrueProb for BB1 equals to TrueProb for
    6537             :       // TmpBB, but the math is more complicated.
    6538             :       uint64_t TrueWeight, FalseWeight;
    6539           1 :       if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) {
    6540             :         uint64_t NewTrueWeight = TrueWeight;
    6541             :         uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
    6542             :         scaleWeights(NewTrueWeight, NewFalseWeight);
    6543           2 :         Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
    6544             :                          .createBranchWeights(TrueWeight, FalseWeight));
    6545             : 
    6546             :         NewTrueWeight = TrueWeight;
    6547             :         NewFalseWeight = 2 * FalseWeight;
    6548             :         scaleWeights(NewTrueWeight, NewFalseWeight);
    6549           2 :         Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
    6550             :                          .createBranchWeights(TrueWeight, FalseWeight));
    6551             :       }
    6552             :     } else {
    6553             :       // Codegen X & Y as:
    6554             :       // BB1:
    6555             :       //   jmp_if_X TmpBB
    6556             :       //   jmp FBB
    6557             :       // TmpBB:
    6558             :       //   jmp_if_Y TBB
    6559             :       //   jmp FBB
    6560             :       //
    6561             :       //  This requires creation of TmpBB after CurBB.
    6562             : 
    6563             :       // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
    6564             :       // The requirement is that
    6565             :       //   FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
    6566             :       //     = FalseProb for orignal BB.
    6567             :       // Assuming the orignal weights are A and B, one choice is to set BB1's
    6568             :       // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
    6569             :       // assumes that
    6570             :       //   FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
    6571             :       uint64_t TrueWeight, FalseWeight;
    6572           2 :       if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) {
    6573             :         uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
    6574             :         uint64_t NewFalseWeight = FalseWeight;
    6575             :         scaleWeights(NewTrueWeight, NewFalseWeight);
    6576           2 :         Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
    6577             :                          .createBranchWeights(TrueWeight, FalseWeight));
    6578             : 
    6579             :         NewTrueWeight = 2 * TrueWeight;
    6580             :         NewFalseWeight = FalseWeight;
    6581             :         scaleWeights(NewTrueWeight, NewFalseWeight);
    6582           2 :         Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
    6583             :                          .createBranchWeights(TrueWeight, FalseWeight));
    6584             :       }
    6585             :     }
    6586             : 
    6587             :     // Note: No point in getting fancy here, since the DT info is never
    6588             :     // available to CodeGenPrepare.
    6589           3 :     ModifiedDT = true;
    6590             : 
    6591             :     MadeChange = true;
    6592             : 
    6593             :     DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
    6594             :           TmpBB->dump());
    6595             :   }
    6596             :   return MadeChange;
    6597      244878 : }

Generated by: LCOV version 1.13