LCOV - code coverage report
Current view: top level - lib/CodeGen - CodeGenPrepare.cpp (source / functions) Hit Total Coverage
Test: llvm-toolchain.info Lines: 1997 2075 96.2 %
Date: 2018-06-17 00:07:59 Functions: 138 148 93.2 %
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/Transforms/Utils/Local.h"
      34             : #include "llvm/Analysis/ValueTracking.h"
      35             : #include "llvm/CodeGen/Analysis.h"
      36             : #include "llvm/CodeGen/ISDOpcodes.h"
      37             : #include "llvm/CodeGen/SelectionDAGNodes.h"
      38             : #include "llvm/CodeGen/TargetLowering.h"
      39             : #include "llvm/CodeGen/TargetPassConfig.h"
      40             : #include "llvm/CodeGen/TargetSubtargetInfo.h"
      41             : #include "llvm/CodeGen/ValueTypes.h"
      42             : #include "llvm/Config/llvm-config.h"
      43             : #include "llvm/IR/Argument.h"
      44             : #include "llvm/IR/Attributes.h"
      45             : #include "llvm/IR/BasicBlock.h"
      46             : #include "llvm/IR/CallSite.h"
      47             : #include "llvm/IR/Constant.h"
      48             : #include "llvm/IR/Constants.h"
      49             : #include "llvm/IR/DataLayout.h"
      50             : #include "llvm/IR/DerivedTypes.h"
      51             : #include "llvm/IR/Dominators.h"
      52             : #include "llvm/IR/Function.h"
      53             : #include "llvm/IR/GetElementPtrTypeIterator.h"
      54             : #include "llvm/IR/GlobalValue.h"
      55             : #include "llvm/IR/GlobalVariable.h"
      56             : #include "llvm/IR/IRBuilder.h"
      57             : #include "llvm/IR/InlineAsm.h"
      58             : #include "llvm/IR/InstrTypes.h"
      59             : #include "llvm/IR/Instruction.h"
      60             : #include "llvm/IR/Instructions.h"
      61             : #include "llvm/IR/IntrinsicInst.h"
      62             : #include "llvm/IR/Intrinsics.h"
      63             : #include "llvm/IR/LLVMContext.h"
      64             : #include "llvm/IR/MDBuilder.h"
      65             : #include "llvm/IR/Module.h"
      66             : #include "llvm/IR/Operator.h"
      67             : #include "llvm/IR/PatternMatch.h"
      68             : #include "llvm/IR/Statepoint.h"
      69             : #include "llvm/IR/Type.h"
      70             : #include "llvm/IR/Use.h"
      71             : #include "llvm/IR/User.h"
      72             : #include "llvm/IR/Value.h"
      73             : #include "llvm/IR/ValueHandle.h"
      74             : #include "llvm/IR/ValueMap.h"
      75             : #include "llvm/Pass.h"
      76             : #include "llvm/Support/BlockFrequency.h"
      77             : #include "llvm/Support/BranchProbability.h"
      78             : #include "llvm/Support/Casting.h"
      79             : #include "llvm/Support/CommandLine.h"
      80             : #include "llvm/Support/Compiler.h"
      81             : #include "llvm/Support/Debug.h"
      82             : #include "llvm/Support/ErrorHandling.h"
      83             : #include "llvm/Support/MachineValueType.h"
      84             : #include "llvm/Support/MathExtras.h"
      85             : #include "llvm/Support/raw_ostream.h"
      86             : #include "llvm/Target/TargetMachine.h"
      87             : #include "llvm/Target/TargetOptions.h"
      88             : #include "llvm/Transforms/Utils/BasicBlockUtils.h"
      89             : #include "llvm/Transforms/Utils/BypassSlowDivision.h"
      90             : #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
      91             : #include <algorithm>
      92             : #include <cassert>
      93             : #include <cstdint>
      94             : #include <iterator>
      95             : #include <limits>
      96             : #include <memory>
      97             : #include <utility>
      98             : #include <vector>
      99             : 
     100             : using namespace llvm;
     101             : using namespace llvm::PatternMatch;
     102             : 
     103             : #define DEBUG_TYPE "codegenprepare"
     104             : 
     105             : STATISTIC(NumBlocksElim, "Number of blocks eliminated");
     106             : STATISTIC(NumPHIsElim,   "Number of trivial PHIs eliminated");
     107             : STATISTIC(NumGEPsElim,   "Number of GEPs converted to casts");
     108             : STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
     109             :                       "sunken Cmps");
     110             : STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
     111             :                        "of sunken Casts");
     112             : STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
     113             :                           "computations were sunk");
     114             : STATISTIC(NumMemoryInstsPhiCreated,
     115             :           "Number of phis created when address "
     116             :           "computations were sunk to memory instructions");
     117             : STATISTIC(NumMemoryInstsSelectCreated,
     118             :           "Number of select created when address "
     119             :           "computations were sunk to memory instructions");
     120             : STATISTIC(NumExtsMoved,  "Number of [s|z]ext instructions combined with loads");
     121             : STATISTIC(NumExtUses,    "Number of uses of [s|z]ext instructions optimized");
     122             : STATISTIC(NumAndsAdded,
     123             :           "Number of and mask instructions added to form ext loads");
     124             : STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
     125             : STATISTIC(NumRetsDup,    "Number of return instructions duplicated");
     126             : STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
     127             : STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
     128             : STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
     129             : 
     130      101169 : static cl::opt<bool> DisableBranchOpts(
     131      202338 :   "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
     132      202338 :   cl::desc("Disable branch optimizations in CodeGenPrepare"));
     133             : 
     134             : static cl::opt<bool>
     135      303507 :     DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
     136      202338 :                   cl::desc("Disable GC optimizations in CodeGenPrepare"));
     137             : 
     138      101169 : static cl::opt<bool> DisableSelectToBranch(
     139      202338 :   "disable-cgp-select2branch", cl::Hidden, cl::init(false),
     140      202338 :   cl::desc("Disable select to branch conversion."));
     141             : 
     142      101169 : static cl::opt<bool> AddrSinkUsingGEPs(
     143      202338 :   "addr-sink-using-gep", cl::Hidden, cl::init(true),
     144      202338 :   cl::desc("Address sinking in CGP using GEPs."));
     145             : 
     146      101169 : static cl::opt<bool> EnableAndCmpSinking(
     147      202338 :    "enable-andcmp-sinking", cl::Hidden, cl::init(true),
     148      202338 :    cl::desc("Enable sinkinig and/cmp into branches."));
     149             : 
     150      101169 : static cl::opt<bool> DisableStoreExtract(
     151      202338 :     "disable-cgp-store-extract", cl::Hidden, cl::init(false),
     152      202338 :     cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
     153             : 
     154      101169 : static cl::opt<bool> StressStoreExtract(
     155      202338 :     "stress-cgp-store-extract", cl::Hidden, cl::init(false),
     156      202338 :     cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
     157             : 
     158      101169 : static cl::opt<bool> DisableExtLdPromotion(
     159      202338 :     "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
     160      101169 :     cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
     161      101169 :              "CodeGenPrepare"));
     162             : 
     163      101169 : static cl::opt<bool> StressExtLdPromotion(
     164      202338 :     "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
     165      101169 :     cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
     166      101169 :              "optimization in CodeGenPrepare"));
     167             : 
     168      101169 : static cl::opt<bool> DisablePreheaderProtect(
     169      202338 :     "disable-preheader-prot", cl::Hidden, cl::init(false),
     170      202338 :     cl::desc("Disable protection against removing loop preheaders"));
     171             : 
     172      101169 : static cl::opt<bool> ProfileGuidedSectionPrefix(
     173      202338 :     "profile-guided-section-prefix", cl::Hidden, cl::init(true), cl::ZeroOrMore,
     174      303507 :     cl::desc("Use profile info to add section prefix for hot/cold functions"));
     175             : 
     176      101169 : static cl::opt<unsigned> FreqRatioToSkipMerge(
     177      202338 :     "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2),
     178      101169 :     cl::desc("Skip merging empty blocks if (frequency of empty block) / "
     179      101169 :              "(frequency of destination block) is greater than this ratio"));
     180             : 
     181      101169 : static cl::opt<bool> ForceSplitStore(
     182      202338 :     "force-split-store", cl::Hidden, cl::init(false),
     183      202338 :     cl::desc("Force store splitting no matter what the target query says."));
     184             : 
     185             : static cl::opt<bool>
     186      101169 : EnableTypePromotionMerge("cgp-type-promotion-merge", cl::Hidden,
     187      101169 :     cl::desc("Enable merging of redundant sexts when one is dominating"
     188      303507 :     " the other."), cl::init(true));
     189             : 
     190      101169 : static cl::opt<bool> DisableComplexAddrModes(
     191      202338 :     "disable-complex-addr-modes", cl::Hidden, cl::init(false),
     192      101169 :     cl::desc("Disables combining addressing modes with different parts "
     193      101169 :              "in optimizeMemoryInst."));
     194             : 
     195             : static cl::opt<bool>
     196      303507 : AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(false),
     197      202338 :                 cl::desc("Allow creation of Phis in Address sinking."));
     198             : 
     199             : static cl::opt<bool>
     200      303507 : AddrSinkNewSelects("addr-sink-new-select", cl::Hidden, cl::init(true),
     201      202338 :                    cl::desc("Allow creation of selects in Address sinking."));
     202             : 
     203      101169 : static cl::opt<bool> AddrSinkCombineBaseReg(
     204      202338 :     "addr-sink-combine-base-reg", cl::Hidden, cl::init(true),
     205      202338 :     cl::desc("Allow combining of BaseReg field in Address sinking."));
     206             : 
     207      101169 : static cl::opt<bool> AddrSinkCombineBaseGV(
     208      202338 :     "addr-sink-combine-base-gv", cl::Hidden, cl::init(true),
     209      202338 :     cl::desc("Allow combining of BaseGV field in Address sinking."));
     210             : 
     211      101169 : static cl::opt<bool> AddrSinkCombineBaseOffs(
     212      202338 :     "addr-sink-combine-base-offs", cl::Hidden, cl::init(true),
     213      202338 :     cl::desc("Allow combining of BaseOffs field in Address sinking."));
     214             : 
     215      101169 : static cl::opt<bool> AddrSinkCombineScaledReg(
     216      202338 :     "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(true),
     217      202338 :     cl::desc("Allow combining of ScaledReg field in Address sinking."));
     218             : 
     219             : static cl::opt<bool>
     220      101169 :     EnableGEPOffsetSplit("cgp-split-large-offset-gep", cl::Hidden,
     221      202338 :                          cl::init(true),
     222      202338 :                          cl::desc("Enable splitting large offset of GEP."));
     223             : 
     224             : namespace {
     225             : 
     226             : using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
     227             : using TypeIsSExt = PointerIntPair<Type *, 1, bool>;
     228             : using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
     229             : using SExts = SmallVector<Instruction *, 16>;
     230             : using ValueToSExts = DenseMap<Value *, SExts>;
     231             : 
     232             : class TypePromotionTransaction;
     233             : 
     234       95505 :   class CodeGenPrepare : public FunctionPass {
     235             :     const TargetMachine *TM = nullptr;
     236             :     const TargetSubtargetInfo *SubtargetInfo;
     237             :     const TargetLowering *TLI = nullptr;
     238             :     const TargetRegisterInfo *TRI;
     239             :     const TargetTransformInfo *TTI = nullptr;
     240             :     const TargetLibraryInfo *TLInfo;
     241             :     const LoopInfo *LI;
     242             :     std::unique_ptr<BlockFrequencyInfo> BFI;
     243             :     std::unique_ptr<BranchProbabilityInfo> BPI;
     244             : 
     245             :     /// As we scan instructions optimizing them, this is the next instruction
     246             :     /// to optimize. Transforms that can invalidate this should update it.
     247             :     BasicBlock::iterator CurInstIterator;
     248             : 
     249             :     /// Keeps track of non-local addresses that have been sunk into a block.
     250             :     /// This allows us to avoid inserting duplicate code for blocks with
     251             :     /// multiple load/stores of the same address. The usage of WeakTrackingVH
     252             :     /// enables SunkAddrs to be treated as a cache whose entries can be
     253             :     /// invalidated if a sunken address computation has been erased.
     254             :     ValueMap<Value*, WeakTrackingVH> SunkAddrs;
     255             : 
     256             :     /// Keeps track of all instructions inserted for the current function.
     257             :     SetOfInstrs InsertedInsts;
     258             : 
     259             :     /// Keeps track of the type of the related instruction before their
     260             :     /// promotion for the current function.
     261             :     InstrToOrigTy PromotedInsts;
     262             : 
     263             :     /// Keep track of instructions removed during promotion.
     264             :     SetOfInstrs RemovedInsts;
     265             : 
     266             :     /// Keep track of sext chains based on their initial value.
     267             :     DenseMap<Value *, Instruction *> SeenChainsForSExt;
     268             : 
     269             :     /// Keep track of GEPs accessing the same data structures such as structs or
     270             :     /// arrays that are candidates to be split later because of their large
     271             :     /// size.
     272             :     DenseMap<
     273             :         AssertingVH<Value>,
     274             :         SmallVector<std::pair<AssertingVH<GetElementPtrInst>, int64_t>, 32>>
     275             :         LargeOffsetGEPMap;
     276             : 
     277             :     /// Keep track of new GEP base after splitting the GEPs having large offset.
     278             :     SmallSet<AssertingVH<Value>, 2> NewGEPBases;
     279             : 
     280             :     /// Map serial numbers to Large offset GEPs.
     281             :     DenseMap<AssertingVH<GetElementPtrInst>, int> LargeOffsetGEPID;
     282             : 
     283             :     /// Keep track of SExt promoted.
     284             :     ValueToSExts ValToSExtendedUses;
     285             : 
     286             :     /// True if CFG is modified in any way.
     287             :     bool ModifiedDT;
     288             : 
     289             :     /// True if optimizing for size.
     290             :     bool OptSize;
     291             : 
     292             :     /// DataLayout for the Function being processed.
     293             :     const DataLayout *DL = nullptr;
     294             : 
     295             :   public:
     296             :     static char ID; // Pass identification, replacement for typeid
     297             : 
     298      153632 :     CodeGenPrepare() : FunctionPass(ID) {
     299       19204 :       initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
     300       19204 :     }
     301             : 
     302             :     bool runOnFunction(Function &F) override;
     303             : 
     304          32 :     StringRef getPassName() const override { return "CodeGen Prepare"; }
     305             : 
     306       19090 :     void getAnalysisUsage(AnalysisUsage &AU) const override {
     307             :       // FIXME: When we can selectively preserve passes, preserve the domtree.
     308             :       AU.addRequired<ProfileSummaryInfoWrapperPass>();
     309             :       AU.addRequired<TargetLibraryInfoWrapperPass>();
     310             :       AU.addRequired<TargetTransformInfoWrapperPass>();
     311             :       AU.addRequired<LoopInfoWrapperPass>();
     312       19090 :     }
     313             : 
     314             :   private:
     315             :     bool eliminateFallThrough(Function &F);
     316             :     bool eliminateMostlyEmptyBlocks(Function &F);
     317             :     BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
     318             :     bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
     319             :     void eliminateMostlyEmptyBlock(BasicBlock *BB);
     320             :     bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
     321             :                                        bool isPreheader);
     322             :     bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
     323             :     bool optimizeInst(Instruction *I, bool &ModifiedDT);
     324             :     bool optimizeMemoryInst(Instruction *I, Value *Addr,
     325             :                             Type *AccessTy, unsigned AS);
     326             :     bool optimizeInlineAsmInst(CallInst *CS);
     327             :     bool optimizeCallInst(CallInst *CI, bool &ModifiedDT);
     328             :     bool optimizeExt(Instruction *&I);
     329             :     bool optimizeExtUses(Instruction *I);
     330             :     bool optimizeLoadExt(LoadInst *I);
     331             :     bool optimizeSelectInst(SelectInst *SI);
     332             :     bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
     333             :     bool optimizeSwitchInst(SwitchInst *CI);
     334             :     bool optimizeExtractElementInst(Instruction *Inst);
     335             :     bool dupRetToEnableTailCallOpts(BasicBlock *BB);
     336             :     bool placeDbgValues(Function &F);
     337             :     bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
     338             :                       LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
     339             :     bool tryToPromoteExts(TypePromotionTransaction &TPT,
     340             :                           const SmallVectorImpl<Instruction *> &Exts,
     341             :                           SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
     342             :                           unsigned CreatedInstsCost = 0);
     343             :     bool mergeSExts(Function &F);
     344             :     bool splitLargeGEPOffsets();
     345             :     bool performAddressTypePromotion(
     346             :         Instruction *&Inst,
     347             :         bool AllowPromotionWithoutCommonHeader,
     348             :         bool HasPromoted, TypePromotionTransaction &TPT,
     349             :         SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
     350             :     bool splitBranchCondition(Function &F);
     351             :     bool simplifyOffsetableRelocate(Instruction &I);
     352             :   };
     353             : 
     354             : } // end anonymous namespace
     355             : 
     356             : char CodeGenPrepare::ID = 0;
     357             : 
     358       33750 : INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,
     359             :                       "Optimize for code generation", false, false)
     360       33750 : INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
     361      220974 : INITIALIZE_PASS_END(CodeGenPrepare, DEBUG_TYPE,
     362             :                     "Optimize for code generation", false, false)
     363             : 
     364       19118 : FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); }
     365             : 
     366      182244 : bool CodeGenPrepare::runOnFunction(Function &F) {
     367      182244 :   if (skipFunction(F))
     368             :     return false;
     369             : 
     370      182088 :   DL = &F.getParent()->getDataLayout();
     371             : 
     372             :   bool EverMadeChange = false;
     373             :   // Clear per function information.
     374      182088 :   InsertedInsts.clear();
     375      182088 :   PromotedInsts.clear();
     376             : 
     377      182088 :   ModifiedDT = false;
     378      182088 :   if (auto *TPC = getAnalysisIfAvailable<TargetPassConfig>()) {
     379      182079 :     TM = &TPC->getTM<TargetMachine>();
     380      182079 :     SubtargetInfo = TM->getSubtargetImpl(F);
     381      182079 :     TLI = SubtargetInfo->getTargetLowering();
     382      182079 :     TRI = SubtargetInfo->getRegisterInfo();
     383             :   }
     384      364176 :   TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
     385      182088 :   TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
     386      364176 :   LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
     387      182088 :   BPI.reset(new BranchProbabilityInfo(F, *LI));
     388      182088 :   BFI.reset(new BlockFrequencyInfo(F, *BPI, *LI));
     389      182088 :   OptSize = F.optForSize();
     390             : 
     391             :   ProfileSummaryInfo *PSI =
     392      182088 :       getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
     393      182088 :   if (ProfileGuidedSectionPrefix) {
     394      182088 :     if (PSI->isFunctionHotInCallGraph(&F, *BFI))
     395          11 :       F.setSectionPrefix(".hot");
     396      182077 :     else if (PSI->isFunctionColdInCallGraph(&F, *BFI))
     397           6 :       F.setSectionPrefix(".unlikely");
     398             :   }
     399             : 
     400             :   /// This optimization identifies DIV instructions that can be
     401             :   /// profitably bypassed and carried out with a shorter, faster divide.
     402      358432 :   if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI &&
     403             :       TLI->isSlowDivBypassed()) {
     404             :     const DenseMap<unsigned int, unsigned int> &BypassWidths =
     405             :        TLI->getBypassSlowDivWidths();
     406             :     BasicBlock* BB = &*F.begin();
     407       34628 :     while (BB != nullptr) {
     408             :       // bypassSlowDivision may create new BBs, but we don't want to reapply the
     409             :       // optimization to those blocks.
     410             :       BasicBlock* Next = BB->getNextNode();
     411       11709 :       EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
     412             :       BB = Next;
     413             :     }
     414             :   }
     415             : 
     416             :   // Eliminate blocks that contain only PHI nodes and an
     417             :   // unconditional branch.
     418      182088 :   EverMadeChange |= eliminateMostlyEmptyBlocks(F);
     419             : 
     420             :   // llvm.dbg.value is far away from the value then iSel may not be able
     421             :   // handle it properly. iSel will drop llvm.dbg.value if it can not
     422             :   // find a node corresponding to the value.
     423      182088 :   EverMadeChange |= placeDbgValues(F);
     424             : 
     425      182088 :   if (!DisableBranchOpts)
     426      182025 :     EverMadeChange |= splitBranchCondition(F);
     427             : 
     428             :   // Split some critical edges where one of the sources is an indirect branch,
     429             :   // to help generate sane code for PHIs involving such edges.
     430      182088 :   EverMadeChange |= SplitIndirectBrCriticalEdges(F);
     431             : 
     432             :   bool MadeChange = true;
     433      562560 :   while (MadeChange) {
     434             :     MadeChange = false;
     435      190236 :     SeenChainsForSExt.clear();
     436      190236 :     ValToSExtendedUses.clear();
     437      190236 :     RemovedInsts.clear();
     438      190236 :     LargeOffsetGEPMap.clear();
     439      190236 :     LargeOffsetGEPID.clear();
     440     1129032 :     for (Function::iterator I = F.begin(); I != F.end(); ) {
     441             :       BasicBlock *BB = &*I++;
     442      469435 :       bool ModifiedDTOnIteration = false;
     443      469435 :       MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
     444             : 
     445             :       // Restart BB iteration if the dominator tree of the Function was changed
     446      469435 :       if (ModifiedDTOnIteration)
     447             :         break;
     448             :     }
     449      380472 :     if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
     450          36 :       MadeChange |= mergeSExts(F);
     451      190236 :     if (!LargeOffsetGEPMap.empty())
     452           5 :       MadeChange |= splitLargeGEPOffsets();
     453             : 
     454             :     // Really free removed instructions during promotion.
     455      190236 :     for (Instruction *I : RemovedInsts)
     456         233 :       I->deleteValue();
     457             : 
     458      190236 :     EverMadeChange |= MadeChange;
     459             :   }
     460             : 
     461             :   SunkAddrs.clear();
     462             : 
     463      182088 :   if (!DisableBranchOpts) {
     464             :     MadeChange = false;
     465             :     SmallPtrSet<BasicBlock*, 8> WorkList;
     466      513755 :     for (BasicBlock &BB : F) {
     467      331730 :       SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
     468      331730 :       MadeChange |= ConstantFoldTerminator(&BB, true);
     469      331730 :       if (!MadeChange) continue;
     470             : 
     471        7393 :       for (SmallVectorImpl<BasicBlock*>::iterator
     472       12047 :              II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
     473       14786 :         if (pred_begin(*II) == pred_end(*II))
     474        1089 :           WorkList.insert(*II);
     475             :     }
     476             : 
     477             :     // Delete the dead blocks and any of their dead successors.
     478      182025 :     MadeChange |= !WorkList.empty();
     479      185415 :     while (!WorkList.empty()) {
     480        3390 :       BasicBlock *BB = *WorkList.begin();
     481             :       WorkList.erase(BB);
     482        1695 :       SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
     483             : 
     484        1695 :       DeleteDeadBlock(BB);
     485             : 
     486        1386 :       for (SmallVectorImpl<BasicBlock*>::iterator
     487        3081 :              II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
     488        2772 :         if (pred_begin(*II) == pred_end(*II))
     489         777 :           WorkList.insert(*II);
     490             :     }
     491             : 
     492             :     // Merge pairs of basic blocks with unconditional branches, connected by
     493             :     // a single edge.
     494      182025 :     if (EverMadeChange || MadeChange)
     495        9009 :       MadeChange |= eliminateFallThrough(F);
     496             : 
     497      182025 :     EverMadeChange |= MadeChange;
     498             :   }
     499             : 
     500      182088 :   if (!DisableGCOpts) {
     501             :     SmallVector<Instruction *, 2> Statepoints;
     502      510744 :     for (BasicBlock &BB : F)
     503     2566292 :       for (Instruction &I : BB)
     504     2237636 :         if (isStatepoint(I))
     505          76 :           Statepoints.push_back(&I);
     506      182240 :     for (auto &I : Statepoints)
     507          76 :       EverMadeChange |= simplifyOffsetableRelocate(*I);
     508             :   }
     509             : 
     510             :   return EverMadeChange;
     511             : }
     512             : 
     513             : /// Merge basic blocks which are connected by a single edge, where one of the
     514             : /// basic blocks has a single successor pointing to the other basic block,
     515             : /// which has a single predecessor.
     516        9009 : bool CodeGenPrepare::eliminateFallThrough(Function &F) {
     517             :   bool Changed = false;
     518             :   // Scan all of the blocks in the function, except for the entry block.
     519      140722 :   for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
     520             :     BasicBlock *BB = &*I++;
     521             :     // If the destination block has a single pred, then this is a trivial
     522             :     // edge, just collapse it.
     523             :     BasicBlock *SinglePred = BB->getSinglePredecessor();
     524             : 
     525             :     // Don't merge if BB's address is taken.
     526      222086 :     if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
     527             : 
     528             :     BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
     529       42782 :     if (Term && !Term->isConditional()) {
     530             :       Changed = true;
     531             :       LLVM_DEBUG(dbgs() << "To merge:\n" << *SinglePred << "\n\n\n");
     532             :       // Remember if SinglePred was the entry block of the function.
     533             :       // If so, we will need to move BB back to the entry position.
     534        1494 :       bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
     535        1494 :       MergeBasicBlockIntoOnlyPred(BB, nullptr);
     536             : 
     537        2026 :       if (isEntry && BB != &BB->getParent()->getEntryBlock())
     538           0 :         BB->moveBefore(&BB->getParent()->getEntryBlock());
     539             : 
     540             :       // We have erased a block. Update the iterator.
     541        1494 :       I = BB->getIterator();
     542             :     }
     543             :   }
     544        9009 :   return Changed;
     545             : }
     546             : 
     547             : /// Find a destination block from BB if BB is mergeable empty block.
     548      153298 : BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
     549             :   // If this block doesn't end with an uncond branch, ignore it.
     550             :   BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
     551       98342 :   if (!BI || !BI->isUnconditional())
     552             :     return nullptr;
     553             : 
     554             :   // If the instruction before the branch (skipping debug info) isn't a phi
     555             :   // node, then other stuff is happening here.
     556       63071 :   BasicBlock::iterator BBI = BI->getIterator();
     557       63071 :   if (BBI != BB->begin()) {
     558             :     --BBI;
     559       57680 :     while (isa<DbgInfoIntrinsic>(BBI)) {
     560         340 :       if (BBI == BB->begin())
     561             :         break;
     562             :       --BBI;
     563             :     }
     564      114689 :     if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
     565             :       return nullptr;
     566             :   }
     567             : 
     568             :   // Do not break infinite loops.
     569             :   BasicBlock *DestBB = BI->getSuccessor(0);
     570        6054 :   if (DestBB == BB)
     571             :     return nullptr;
     572             : 
     573        5953 :   if (!canMergeBlocks(BB, DestBB))
     574             :     DestBB = nullptr;
     575             : 
     576             :   return DestBB;
     577             : }
     578             : 
     579             : /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
     580             : /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
     581             : /// edges in ways that are non-optimal for isel. Start by eliminating these
     582             : /// blocks so we can split them the way we want them.
     583      182088 : bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
     584             :   SmallPtrSet<BasicBlock *, 16> Preheaders;
     585      182088 :   SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
     586      188771 :   while (!LoopList.empty()) {
     587             :     Loop *L = LoopList.pop_back_val();
     588        6683 :     LoopList.insert(LoopList.end(), L->begin(), L->end());
     589        6683 :     if (BasicBlock *Preheader = L->getLoopPreheader())
     590        6647 :       Preheaders.insert(Preheader);
     591             :   }
     592             : 
     593             :   bool MadeChange = false;
     594             :   // Note that this intentionally skips the entry block.
     595      335379 :   for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
     596             :     BasicBlock *BB = &*I++;
     597      153291 :     BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
     598      158740 :     if (!DestBB ||
     599        5449 :         !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB)))
     600      149552 :       continue;
     601             : 
     602        3739 :     eliminateMostlyEmptyBlock(BB);
     603             :     MadeChange = true;
     604             :   }
     605      182088 :   return MadeChange;
     606             : }
     607             : 
     608        5449 : bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
     609             :                                                    BasicBlock *DestBB,
     610             :                                                    bool isPreheader) {
     611             :   // Do not delete loop preheaders if doing so would create a critical edge.
     612             :   // Loop preheaders can be good locations to spill registers. If the
     613             :   // preheader is deleted and we create a critical edge, registers may be
     614             :   // spilled in the loop body instead.
     615        7128 :   if (!DisablePreheaderProtect && isPreheader &&
     616        3255 :       !(BB->getSinglePredecessor() &&
     617        1576 :         BB->getSinglePredecessor()->getSingleSuccessor()))
     618             :     return false;
     619             : 
     620             :   // Try to skip merging if the unique predecessor of BB is terminated by a
     621             :   // switch or indirect branch instruction, and BB is used as an incoming block
     622             :   // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
     623             :   // add COPY instructions in the predecessor of BB instead of BB (if it is not
     624             :   // merged). Note that the critical edge created by merging such blocks wont be
     625             :   // split in MachineSink because the jump table is not analyzable. By keeping
     626             :   // such empty block (BB), ISel will place COPY instructions in BB, not in the
     627             :   // predecessor of BB.
     628        3782 :   BasicBlock *Pred = BB->getUniquePredecessor();
     629        7034 :   if (!Pred ||
     630        2854 :       !(isa<SwitchInst>(Pred->getTerminator()) ||
     631             :         isa<IndirectBrInst>(Pred->getTerminator())))
     632             :     return true;
     633             : 
     634         818 :   if (BB->getTerminator() != BB->getFirstNonPHI())
     635             :     return true;
     636             : 
     637             :   // We use a simple cost heuristic which determine skipping merging is
     638             :   // profitable if the cost of skipping merging is less than the cost of
     639             :   // merging : Cost(skipping merging) < Cost(merging BB), where the
     640             :   // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
     641             :   // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
     642             :   // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
     643             :   //   Freq(Pred) / Freq(BB) > 2.
     644             :   // Note that if there are multiple empty blocks sharing the same incoming
     645             :   // value for the PHIs in the DestBB, we consider them together. In such
     646             :   // case, Cost(merging BB) will be the sum of their frequencies.
     647             : 
     648         409 :   if (!isa<PHINode>(DestBB->begin()))
     649             :     return true;
     650             : 
     651             :   SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
     652             : 
     653             :   // Find all other incoming blocks from which incoming values of all PHIs in
     654             :   // DestBB are the same as the ones from BB.
     655         359 :   for (pred_iterator PI = pred_begin(DestBB), E = pred_end(DestBB); PI != E;
     656             :        ++PI) {
     657         237 :     BasicBlock *DestBBPred = *PI;
     658         237 :     if (DestBBPred == BB)
     659          61 :       continue;
     660             : 
     661         352 :     if (llvm::all_of(DestBB->phis(), [&](const PHINode &DestPN) {
     662         176 :           return DestPN.getIncomingValueForBlock(BB) ==
     663         176 :                  DestPN.getIncomingValueForBlock(DestBBPred);
     664         176 :         }))
     665          36 :       SameIncomingValueBBs.insert(DestBBPred);
     666             :   }
     667             : 
     668             :   // See if all BB's incoming values are same as the value from Pred. In this
     669             :   // case, no reason to skip merging because COPYs are expected to be place in
     670             :   // Pred already.
     671          61 :   if (SameIncomingValueBBs.count(Pred))
     672             :     return true;
     673             : 
     674          55 :   BlockFrequency PredFreq = BFI->getBlockFreq(Pred);
     675          55 :   BlockFrequency BBFreq = BFI->getBlockFreq(BB);
     676             : 
     677          55 :   for (auto SameValueBB : SameIncomingValueBBs)
     678          22 :     if (SameValueBB->getUniquePredecessor() == Pred &&
     679           7 :         DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB))
     680           4 :       BBFreq += BFI->getBlockFreq(SameValueBB);
     681             : 
     682             :   return PredFreq.getFrequency() <=
     683          55 :          BBFreq.getFrequency() * FreqRatioToSkipMerge;
     684             : }
     685             : 
     686             : /// Return true if we can merge BB into DestBB if there is a single
     687             : /// unconditional branch between them, and BB contains no other non-phi
     688             : /// instructions.
     689        5953 : bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
     690             :                                     const BasicBlock *DestBB) const {
     691             :   // We only want to eliminate blocks whose phi nodes are used by phi nodes in
     692             :   // the successor.  If there are more complex condition (e.g. preheaders),
     693             :   // don't mess around with them.
     694        6203 :   for (const PHINode &PN : BB->phis()) {
     695         617 :     for (const User *U : PN.users()) {
     696             :       const Instruction *UI = cast<Instruction>(U);
     697         651 :       if (UI->getParent() != DestBB || !isa<PHINode>(UI))
     698             :         return false;
     699             :       // If User is inside DestBB block and it is a PHINode then check
     700             :       // incoming value. If incoming value is not from BB then this is
     701             :       // a complex condition (e.g. preheaders) we want to avoid here.
     702             :       if (UI->getParent() == DestBB) {
     703             :         if (const PHINode *UPN = dyn_cast<PHINode>(UI))
     704        1539 :           for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
     705             :             Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
     706         673 :             if (Insn && Insn->getParent() == BB &&
     707             :                 Insn->getParent() != UPN->getIncomingBlock(I))
     708             :               return false;
     709             :           }
     710             :       }
     711             :     }
     712             :   }
     713             : 
     714             :   // If BB and DestBB contain any common predecessors, then the phi nodes in BB
     715             :   // and DestBB may have conflicting incoming values for the block.  If so, we
     716             :   // can't merge the block.
     717             :   const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
     718             :   if (!DestBBPN) return true;  // no conflict.
     719             : 
     720             :   // Collect the preds of BB.
     721             :   SmallPtrSet<const BasicBlock*, 16> BBPreds;
     722             :   if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
     723             :     // It is faster to get preds from a PHI than with pred_iterator.
     724        1185 :     for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
     725         487 :       BBPreds.insert(BBPN->getIncomingBlock(i));
     726             :   } else {
     727        2340 :     BBPreds.insert(pred_begin(BB), pred_end(BB));
     728             :   }
     729             : 
     730             :   // Walk the preds of DestBB.
     731       13683 :   for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
     732             :     BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
     733        5956 :     if (BBPreds.count(Pred)) {   // Common predecessor?
     734         447 :       for (const PHINode &PN : DestBB->phis()) {
     735         424 :         const Value *V1 = PN.getIncomingValueForBlock(Pred);
     736         424 :         const Value *V2 = PN.getIncomingValueForBlock(BB);
     737             : 
     738             :         // If V2 is a phi node in BB, look up what the mapped value will be.
     739             :         if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
     740          43 :           if (V2PN->getParent() == BB)
     741          27 :             V2 = V2PN->getIncomingValueForBlock(Pred);
     742             : 
     743             :         // If there is a conflict, bail out.
     744         424 :         if (V1 != V2) return false;
     745             :       }
     746             :     }
     747             :   }
     748             : 
     749             :   return true;
     750             : }
     751             : 
     752             : /// Eliminate a basic block that has only phi's and an unconditional branch in
     753             : /// it.
     754        3739 : void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
     755             :   BranchInst *BI = cast<BranchInst>(BB->getTerminator());
     756             :   BasicBlock *DestBB = BI->getSuccessor(0);
     757             : 
     758             :   LLVM_DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"
     759             :                     << *BB << *DestBB);
     760             : 
     761             :   // If the destination block has a single pred, then this is a trivial edge,
     762             :   // just collapse it.
     763        3739 :   if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
     764         129 :     if (SinglePred != DestBB) {
     765             :       // Remember if SinglePred was the entry block of the function.  If so, we
     766             :       // will need to move BB back to the entry position.
     767         129 :       bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
     768         129 :       MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
     769             : 
     770         129 :       if (isEntry && BB != &BB->getParent()->getEntryBlock())
     771           0 :         BB->moveBefore(&BB->getParent()->getEntryBlock());
     772             : 
     773             :       LLVM_DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
     774             :       return;
     775             :     }
     776             :   }
     777             : 
     778             :   // Otherwise, we have multiple predecessors of BB.  Update the PHIs in DestBB
     779             :   // to handle the new incoming edges it is about to have.
     780        3610 :   for (PHINode &PN : DestBB->phis()) {
     781             :     // Remove the incoming value for BB, and remember it.
     782             :     Value *InVal = PN.removeIncomingValue(BB, false);
     783             : 
     784             :     // Two options: either the InVal is a phi node defined in BB or it is some
     785             :     // value that dominates BB.
     786             :     PHINode *InValPhi = dyn_cast<PHINode>(InVal);
     787         487 :     if (InValPhi && InValPhi->getParent() == BB) {
     788             :       // Add all of the input values of the input PHI as inputs of this phi.
     789        1015 :       for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
     790         420 :         PN.addIncoming(InValPhi->getIncomingValue(i),
     791             :                        InValPhi->getIncomingBlock(i));
     792             :     } else {
     793             :       // Otherwise, add one instance of the dominating value for each edge that
     794             :       // we will be adding.
     795             :       if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
     796         124 :         for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
     797          51 :           PN.addIncoming(InVal, BBPN->getIncomingBlock(i));
     798             :       } else {
     799        3704 :         for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
     800        1264 :           PN.addIncoming(InVal, *PI);
     801             :       }
     802             :     }
     803             :   }
     804             : 
     805             :   // The PHIs are now updated, change everything that refers to BB to use
     806             :   // DestBB and remove BB.
     807        3610 :   BB->replaceAllUsesWith(DestBB);
     808        3610 :   BB->eraseFromParent();
     809             :   ++NumBlocksElim;
     810             : 
     811             :   LLVM_DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
     812             : }
     813             : 
     814             : // Computes a map of base pointer relocation instructions to corresponding
     815             : // derived pointer relocation instructions given a vector of all relocate calls
     816          25 : static void computeBaseDerivedRelocateMap(
     817             :     const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
     818             :     DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
     819             :         &RelocateInstMap) {
     820             :   // Collect information in two maps: one primarily for locating the base object
     821             :   // while filling the second map; the second map is the final structure holding
     822             :   // a mapping between Base and corresponding Derived relocate calls
     823             :   DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
     824         149 :   for (auto *ThisRelocate : AllRelocateCalls) {
     825             :     auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
     826             :                             ThisRelocate->getDerivedPtrIndex());
     827         124 :     RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
     828             :   }
     829         174 :   for (auto &Item : RelocateIdxMap) {
     830          62 :     std::pair<unsigned, unsigned> Key = Item.first;
     831          62 :     if (Key.first == Key.second)
     832             :       // Base relocation: nothing to insert
     833          70 :       continue;
     834             : 
     835          28 :     GCRelocateInst *I = Item.second;
     836          28 :     auto BaseKey = std::make_pair(Key.first, Key.first);
     837             : 
     838             :     // We're iterating over RelocateIdxMap so we cannot modify it.
     839          28 :     auto MaybeBase = RelocateIdxMap.find(BaseKey);
     840          28 :     if (MaybeBase == RelocateIdxMap.end())
     841             :       // TODO: We might want to insert a new base object relocate and gep off
     842             :       // that, if there are enough derived object relocates.
     843           2 :       continue;
     844             : 
     845          52 :     RelocateInstMap[MaybeBase->second].push_back(I);
     846             :   }
     847          25 : }
     848             : 
     849             : // Accepts a GEP and extracts the operands into a vector provided they're all
     850             : // small integer constants
     851          11 : static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
     852             :                                           SmallVectorImpl<Value *> &OffsetV) {
     853          31 :   for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
     854             :     // Only accept small constant integer operands
     855             :     auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
     856          11 :     if (!Op || Op->getZExtValue() > 20)
     857             :       return false;
     858             :   }
     859             : 
     860          26 :   for (unsigned i = 1; i < GEP->getNumOperands(); i++)
     861          18 :     OffsetV.push_back(GEP->getOperand(i));
     862             :   return true;
     863             : }
     864             : 
     865             : // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
     866             : // replace, computes a replacement, and affects it.
     867             : static bool
     868          18 : simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
     869             :                           const SmallVectorImpl<GCRelocateInst *> &Targets) {
     870             :   bool MadeChange = false;
     871             :   // We must ensure the relocation of derived pointer is defined after
     872             :   // relocation of base pointer. If we find a relocation corresponding to base
     873             :   // defined earlier than relocation of base then we move relocation of base
     874             :   // right before found relocation. We consider only relocation in the same
     875             :   // basic block as relocation of base. Relocations from other basic block will
     876             :   // be skipped by optimization and we do not care about them.
     877          36 :   for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
     878          64 :        &*R != RelocatedBase; ++R)
     879          12 :     if (auto RI = dyn_cast<GCRelocateInst>(R))
     880          12 :       if (RI->getStatepoint() == RelocatedBase->getStatepoint())
     881           6 :         if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
     882           6 :           RelocatedBase->moveBefore(RI);
     883           6 :           break;
     884             :         }
     885             : 
     886          70 :   for (GCRelocateInst *ToReplace : Targets) {
     887             :     assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
     888             :            "Not relocating a derived object of the original base object");
     889          26 :     if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
     890             :       // A duplicate relocate call. TODO: coalesce duplicates.
     891          18 :       continue;
     892             :     }
     893             : 
     894          26 :     if (RelocatedBase->getParent() != ToReplace->getParent()) {
     895             :       // Base and derived relocates are in different basic blocks.
     896             :       // In this case transform is only valid when base dominates derived
     897             :       // relocate. However it would be too expensive to check dominance
     898             :       // for each such relocate, so we skip the whole transformation.
     899           1 :       continue;
     900             :     }
     901             : 
     902          25 :     Value *Base = ToReplace->getBasePtr();
     903          25 :     auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
     904          25 :     if (!Derived || Derived->getPointerOperand() != Base)
     905          14 :       continue;
     906             : 
     907             :     SmallVector<Value *, 2> OffsetV;
     908          14 :     if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
     909             :       continue;
     910             : 
     911             :     // Create a Builder and replace the target callsite with a gep
     912             :     assert(RelocatedBase->getNextNode() &&
     913             :            "Should always have one since it's not a terminator");
     914             : 
     915             :     // Insert after RelocatedBase
     916           8 :     IRBuilder<> Builder(RelocatedBase->getNextNode());
     917           8 :     Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
     918             : 
     919             :     // If gc_relocate does not match the actual type, cast it to the right type.
     920             :     // In theory, there must be a bitcast after gc_relocate if the type does not
     921             :     // match, and we should reuse it to get the derived pointer. But it could be
     922             :     // cases like this:
     923             :     // bb1:
     924             :     //  ...
     925             :     //  %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
     926             :     //  br label %merge
     927             :     //
     928             :     // bb2:
     929             :     //  ...
     930             :     //  %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
     931             :     //  br label %merge
     932             :     //
     933             :     // merge:
     934             :     //  %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
     935             :     //  %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
     936             :     //
     937             :     // In this case, we can not find the bitcast any more. So we insert a new bitcast
     938             :     // no matter there is already one or not. In this way, we can handle all cases, and
     939             :     // the extra bitcast should be optimized away in later passes.
     940             :     Value *ActualRelocatedBase = RelocatedBase;
     941           8 :     if (RelocatedBase->getType() != Base->getType()) {
     942             :       ActualRelocatedBase =
     943           0 :           Builder.CreateBitCast(RelocatedBase, Base->getType());
     944             :     }
     945          16 :     Value *Replacement = Builder.CreateGEP(
     946           8 :         Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
     947           8 :     Replacement->takeName(ToReplace);
     948             :     // If the newly generated derived pointer's type does not match the original derived
     949             :     // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
     950             :     Value *ActualReplacement = Replacement;
     951           8 :     if (Replacement->getType() != ToReplace->getType()) {
     952             :       ActualReplacement =
     953           0 :           Builder.CreateBitCast(Replacement, ToReplace->getType());
     954             :     }
     955           8 :     ToReplace->replaceAllUsesWith(ActualReplacement);
     956           8 :     ToReplace->eraseFromParent();
     957             : 
     958             :     MadeChange = true;
     959             :   }
     960          18 :   return MadeChange;
     961             : }
     962             : 
     963             : // Turns this:
     964             : //
     965             : // %base = ...
     966             : // %ptr = gep %base + 15
     967             : // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
     968             : // %base' = relocate(%tok, i32 4, i32 4)
     969             : // %ptr' = relocate(%tok, i32 4, i32 5)
     970             : // %val = load %ptr'
     971             : //
     972             : // into this:
     973             : //
     974             : // %base = ...
     975             : // %ptr = gep %base + 15
     976             : // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
     977             : // %base' = gc.relocate(%tok, i32 4, i32 4)
     978             : // %ptr' = gep %base' + 15
     979             : // %val = load %ptr'
     980          76 : bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
     981             :   bool MadeChange = false;
     982             :   SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
     983             : 
     984         178 :   for (auto *U : I.users())
     985         102 :     if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
     986             :       // Collect all the relocate calls associated with a statepoint
     987          78 :       AllRelocateCalls.push_back(Relocate);
     988             : 
     989             :   // We need atleast one base pointer relocation + one derived pointer
     990             :   // relocation to mangle
     991          76 :   if (AllRelocateCalls.size() < 2)
     992             :     return false;
     993             : 
     994             :   // RelocateInstMap is a mapping from the base relocate instruction to the
     995             :   // corresponding derived relocate instructions
     996             :   DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
     997          25 :   computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
     998          25 :   if (RelocateInstMap.empty())
     999             :     return false;
    1000             : 
    1001          54 :   for (auto &Item : RelocateInstMap)
    1002             :     // Item.first is the RelocatedBase to offset against
    1003             :     // Item.second is the vector of Targets to replace
    1004          18 :     MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
    1005             :   return MadeChange;
    1006             : }
    1007             : 
    1008             : /// SinkCast - Sink the specified cast instruction into its user blocks
    1009      227614 : static bool SinkCast(CastInst *CI) {
    1010      227614 :   BasicBlock *DefBB = CI->getParent();
    1011             : 
    1012             :   /// InsertedCasts - Only insert a cast in each block once.
    1013             :   DenseMap<BasicBlock*, CastInst*> InsertedCasts;
    1014             : 
    1015             :   bool MadeChange = false;
    1016             :   for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
    1017      510845 :        UI != E; ) {
    1018             :     Use &TheUse = UI.getUse();
    1019             :     Instruction *User = cast<Instruction>(*UI);
    1020             : 
    1021             :     // Figure out which BB this cast is used in.  For PHI's this is the
    1022             :     // appropriate predecessor block.
    1023      283231 :     BasicBlock *UserBB = User->getParent();
    1024             :     if (PHINode *PN = dyn_cast<PHINode>(User)) {
    1025        4889 :       UserBB = PN->getIncomingBlock(TheUse);
    1026             :     }
    1027             : 
    1028             :     // Preincrement use iterator so we don't invalidate it.
    1029             :     ++UI;
    1030             : 
    1031             :     // The first insertion point of a block containing an EH pad is after the
    1032             :     // pad.  If the pad is the user, we cannot sink the cast past the pad.
    1033           1 :     if (User->isEHPad())
    1034      235406 :       continue;
    1035             : 
    1036             :     // If the block selected to receive the cast is an EH pad that does not
    1037             :     // allow non-PHI instructions before the terminator, we can't sink the
    1038             :     // cast.
    1039      283230 :     if (UserBB->getTerminator()->isEHPad())
    1040           0 :       continue;
    1041             : 
    1042             :     // If this user is in the same block as the cast, don't change the cast.
    1043      518634 :     if (UserBB == DefBB) continue;
    1044             : 
    1045             :     // If we have already inserted a cast into this block, use it.
    1046             :     CastInst *&InsertedCast = InsertedCasts[UserBB];
    1047             : 
    1048       47826 :     if (!InsertedCast) {
    1049       45787 :       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
    1050             :       assert(InsertPt != UserBB->end());
    1051       91574 :       InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
    1052             :                                       CI->getType(), "", &*InsertPt);
    1053       45787 :       InsertedCast->setDebugLoc(CI->getDebugLoc());
    1054             :     }
    1055             : 
    1056             :     // Replace a use of the cast with a use of the new cast.
    1057       47826 :     TheUse = InsertedCast;
    1058             :     MadeChange = true;
    1059             :     ++NumCastUses;
    1060             :   }
    1061             : 
    1062             :   // If we removed all uses, nuke the cast.
    1063      227614 :   if (CI->use_empty()) {
    1064        7284 :     salvageDebugInfo(*CI);
    1065        7284 :     CI->eraseFromParent();
    1066             :     MadeChange = true;
    1067             :   }
    1068             : 
    1069      227614 :   return MadeChange;
    1070             : }
    1071             : 
    1072             : /// If the specified cast instruction is a noop copy (e.g. it's casting from
    1073             : /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
    1074             : /// reduce the number of virtual registers that must be created and coalesced.
    1075             : ///
    1076             : /// Return true if any changes are made.
    1077      313624 : static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
    1078             :                                        const DataLayout &DL) {
    1079             :   // Sink only "cheap" (or nop) address-space casts.  This is a weaker condition
    1080             :   // than sinking only nop casts, but is helpful on some platforms.
    1081             :   if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) {
    1082         554 :     if (!TLI.isCheapAddrSpaceCast(ASC->getSrcAddressSpace(),
    1083         277 :                                   ASC->getDestAddressSpace()))
    1084             :       return false;
    1085             :   }
    1086             : 
    1087             :   // If this is a noop copy,
    1088      626952 :   EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
    1089      313476 :   EVT DstVT = TLI.getValueType(DL, CI->getType());
    1090             : 
    1091             :   // This is an fp<->int conversion?
    1092      313476 :   if (SrcVT.isInteger() != DstVT.isInteger())
    1093             :     return false;
    1094             : 
    1095             :   // If this is an extension, it will be a zero or sign extension, which
    1096             :   // isn't a noop.
    1097      294283 :   if (SrcVT.bitsLT(DstVT)) return false;
    1098             : 
    1099             :   // If these values will be promoted, find out what they will be promoted
    1100             :   // to.  This helps us consider truncates on PPC as noop copies when they
    1101             :   // are.
    1102      513382 :   if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
    1103             :       TargetLowering::TypePromoteInteger)
    1104        2180 :     SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
    1105      513382 :   if (TLI.getTypeAction(CI->getContext(), DstVT) ==
    1106             :       TargetLowering::TypePromoteInteger)
    1107       10216 :     DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
    1108             : 
    1109             :   // If, after promotion, these are the same types, this is a noop copy.
    1110      256751 :   if (SrcVT != DstVT)
    1111             :     return false;
    1112             : 
    1113      224126 :   return SinkCast(CI);
    1114             : }
    1115             : 
    1116             : /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
    1117             : /// possible.
    1118             : ///
    1119             : /// Return true if any changes were made.
    1120       95415 : static bool CombineUAddWithOverflow(CmpInst *CI) {
    1121             :   Value *A, *B;
    1122             :   Instruction *AddI;
    1123       95415 :   if (!match(CI,
    1124       95415 :              m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
    1125             :     return false;
    1126             : 
    1127          55 :   Type *Ty = AddI->getType();
    1128          55 :   if (!isa<IntegerType>(Ty))
    1129             :     return false;
    1130             : 
    1131             :   // We don't want to move around uses of condition values this late, so we we
    1132             :   // check if it is legal to create the call to the intrinsic in the basic
    1133             :   // block containing the icmp:
    1134             : 
    1135          62 :   if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
    1136             :     return false;
    1137             : 
    1138             : #ifndef NDEBUG
    1139             :   // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
    1140             :   // for now:
    1141             :   if (AddI->hasOneUse())
    1142             :     assert(*AddI->user_begin() == CI && "expected!");
    1143             : #endif
    1144             : 
    1145          49 :   Module *M = CI->getModule();
    1146          49 :   Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
    1147             : 
    1148          98 :   auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
    1149             : 
    1150             :   auto *UAddWithOverflow =
    1151          98 :       CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
    1152          98 :   auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
    1153             :   auto *Overflow =
    1154          98 :       ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
    1155             : 
    1156          49 :   CI->replaceAllUsesWith(Overflow);
    1157          49 :   AddI->replaceAllUsesWith(UAdd);
    1158          49 :   CI->eraseFromParent();
    1159          49 :   AddI->eraseFromParent();
    1160          49 :   return true;
    1161             : }
    1162             : 
    1163             : /// Sink the given CmpInst into user blocks to reduce the number of virtual
    1164             : /// registers that must be created and coalesced. This is a clear win except on
    1165             : /// targets with multiple condition code registers (PowerPC), where it might
    1166             : /// lose; some adjustment may be wanted there.
    1167             : ///
    1168             : /// Return true if any changes are made.
    1169       96373 : static bool SinkCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
    1170       96373 :   BasicBlock *DefBB = CI->getParent();
    1171             : 
    1172             :   // Avoid sinking soft-FP comparisons, since this can move them into a loop.
    1173       96462 :   if (TLI && TLI->useSoftFloat() && isa<FCmpInst>(CI))
    1174             :     return false;
    1175             : 
    1176             :   // Only insert a cmp in each block once.
    1177             :   DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
    1178             : 
    1179             :   bool MadeChange = false;
    1180             :   for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
    1181      233583 :        UI != E; ) {
    1182             :     Use &TheUse = UI.getUse();
    1183             :     Instruction *User = cast<Instruction>(*UI);
    1184             : 
    1185             :     // Preincrement use iterator so we don't invalidate it.
    1186             :     ++UI;
    1187             : 
    1188             :     // Don't bother for PHI nodes.
    1189      137255 :     if (isa<PHINode>(User))
    1190      136738 :       continue;
    1191             : 
    1192             :     // Figure out which BB this cmp is used in.
    1193      136756 :     BasicBlock *UserBB = User->getParent();
    1194             : 
    1195             :     // If this user is in the same block as the cmp, don't change the cmp.
    1196      136756 :     if (UserBB == DefBB) continue;
    1197             : 
    1198             :     // If we have already inserted a cmp into this block, use it.
    1199             :     CmpInst *&InsertedCmp = InsertedCmps[UserBB];
    1200             : 
    1201        1016 :     if (!InsertedCmp) {
    1202         857 :       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
    1203             :       assert(InsertPt != UserBB->end());
    1204         857 :       InsertedCmp =
    1205        1714 :           CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
    1206             :                           CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
    1207             :       // Propagate the debug info.
    1208         857 :       InsertedCmp->setDebugLoc(CI->getDebugLoc());
    1209             :     }
    1210             : 
    1211             :     // Replace a use of the cmp with a use of the new cmp.
    1212        1016 :     TheUse = InsertedCmp;
    1213             :     MadeChange = true;
    1214             :     ++NumCmpUses;
    1215             :   }
    1216             : 
    1217             :   // If we removed all uses, nuke the cmp.
    1218       96328 :   if (CI->use_empty()) {
    1219         631 :     CI->eraseFromParent();
    1220             :     MadeChange = true;
    1221             :   }
    1222             : 
    1223             :   return MadeChange;
    1224             : }
    1225             : 
    1226       96373 : static bool OptimizeCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
    1227       96373 :   if (SinkCmpExpression(CI, TLI))
    1228             :     return true;
    1229             : 
    1230       95415 :   if (CombineUAddWithOverflow(CI))
    1231             :     return true;
    1232             : 
    1233       95366 :   return false;
    1234             : }
    1235             : 
    1236             : /// Duplicate and sink the given 'and' instruction into user blocks where it is
    1237             : /// used in a compare to allow isel to generate better code for targets where
    1238             : /// this operation can be combined.
    1239             : ///
    1240             : /// Return true if any changes are made.
    1241       18764 : static bool sinkAndCmp0Expression(Instruction *AndI,
    1242             :                                   const TargetLowering &TLI,
    1243             :                                   SetOfInstrs &InsertedInsts) {
    1244             :   // Double-check that we're not trying to optimize an instruction that was
    1245             :   // already optimized by some other part of this pass.
    1246             :   assert(!InsertedInsts.count(AndI) &&
    1247             :          "Attempting to optimize already optimized and instruction");
    1248             :   (void) InsertedInsts;
    1249             : 
    1250             :   // Nothing to do for single use in same basic block.
    1251       35327 :   if (AndI->hasOneUse() &&
    1252       33330 :       AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent())
    1253             :     return false;
    1254             : 
    1255             :   // Try to avoid cases where sinking/duplicating is likely to increase register
    1256             :   // pressure.
    1257        4794 :   if (!isa<ConstantInt>(AndI->getOperand(0)) &&
    1258        1090 :       !isa<ConstantInt>(AndI->getOperand(1)) &&
    1259        2901 :       AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse())
    1260             :     return false;
    1261             : 
    1262        2107 :   for (auto *U : AndI->users()) {
    1263             :     Instruction *User = cast<Instruction>(U);
    1264             : 
    1265             :     // Only sink for and mask feeding icmp with 0.
    1266        1967 :     if (!isa<ICmpInst>(User))
    1267             :       return false;
    1268             : 
    1269         325 :     auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1));
    1270         118 :     if (!CmpC || !CmpC->isZero())
    1271             :       return false;
    1272             :   }
    1273             : 
    1274         140 :   if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI))
    1275             :     return false;
    1276             : 
    1277             :   LLVM_DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n");
    1278             :   LLVM_DEBUG(AndI->getParent()->dump());
    1279             : 
    1280             :   // Push the 'and' into the same block as the icmp 0.  There should only be
    1281             :   // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
    1282             :   // others, so we don't need to keep track of which BBs we insert into.
    1283             :   for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
    1284          91 :        UI != E; ) {
    1285             :     Use &TheUse = UI.getUse();
    1286             :     Instruction *User = cast<Instruction>(*UI);
    1287             : 
    1288             :     // Preincrement use iterator so we don't invalidate it.
    1289             :     ++UI;
    1290             : 
    1291             :     LLVM_DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n");
    1292             : 
    1293             :     // Keep the 'and' in the same place if the use is already in the same block.
    1294             :     Instruction *InsertPt =
    1295          42 :         User->getParent() == AndI->getParent() ? AndI : User;
    1296             :     Instruction *InsertedAnd =
    1297          42 :         BinaryOperator::Create(Instruction::And, AndI->getOperand(0),
    1298          42 :                                AndI->getOperand(1), "", InsertPt);
    1299             :     // Propagate the debug info.
    1300          42 :     InsertedAnd->setDebugLoc(AndI->getDebugLoc());
    1301             : 
    1302             :     // Replace a use of the 'and' with a use of the new 'and'.
    1303             :     TheUse = InsertedAnd;
    1304             :     ++NumAndUses;
    1305             :     LLVM_DEBUG(User->getParent()->dump());
    1306             :   }
    1307             : 
    1308             :   // We removed all uses, nuke the and.
    1309          49 :   AndI->eraseFromParent();
    1310             :   return true;
    1311             : }
    1312             : 
    1313             : /// Check if the candidates could be combined with a shift instruction, which
    1314             : /// includes:
    1315             : /// 1. Truncate instruction
    1316             : /// 2. And instruction and the imm is a mask of the low bits:
    1317             : /// imm & (imm+1) == 0
    1318        1218 : static bool isExtractBitsCandidateUse(Instruction *User) {
    1319        1218 :   if (!isa<TruncInst>(User)) {
    1320        1298 :     if (User->getOpcode() != Instruction::And ||
    1321         247 :         !isa<ConstantInt>(User->getOperand(1)))
    1322             :       return false;
    1323             : 
    1324             :     const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
    1325             : 
    1326         711 :     if ((Cimm & (Cimm + 1)).getBoolValue())
    1327             :       return false;
    1328             :   }
    1329             :   return true;
    1330             : }
    1331             : 
    1332             : /// Sink both shift and truncate instruction to the use of truncate's BB.
    1333             : static bool
    1334          49 : SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
    1335             :                      DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
    1336             :                      const TargetLowering &TLI, const DataLayout &DL) {
    1337          49 :   BasicBlock *UserBB = User->getParent();
    1338             :   DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
    1339             :   TruncInst *TruncI = dyn_cast<TruncInst>(User);
    1340             :   bool MadeChange = false;
    1341             : 
    1342             :   for (Value::user_iterator TruncUI = TruncI->user_begin(),
    1343             :                             TruncE = TruncI->user_end();
    1344          99 :        TruncUI != TruncE;) {
    1345             : 
    1346             :     Use &TruncTheUse = TruncUI.getUse();
    1347             :     Instruction *TruncUser = cast<Instruction>(*TruncUI);
    1348             :     // Preincrement use iterator so we don't invalidate it.
    1349             : 
    1350             :     ++TruncUI;
    1351             : 
    1352         100 :     int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
    1353          50 :     if (!ISDOpcode)
    1354          64 :       continue;
    1355             : 
    1356             :     // If the use is actually a legal node, there will not be an
    1357             :     // implicit truncate.
    1358             :     // FIXME: always querying the result type is just an
    1359             :     // approximation; some nodes' legality is determined by the
    1360             :     // operand or other means. There's no good way to find out though.
    1361          34 :     if (TLI.isOperationLegalOrCustom(
    1362             :             ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
    1363           5 :       continue;
    1364             : 
    1365             :     // Don't bother for PHI nodes.
    1366          29 :     if (isa<PHINode>(TruncUser))
    1367           0 :       continue;
    1368             : 
    1369          29 :     BasicBlock *TruncUserBB = TruncUser->getParent();
    1370             : 
    1371          29 :     if (UserBB == TruncUserBB)
    1372          27 :       continue;
    1373             : 
    1374             :     BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
    1375             :     CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
    1376             : 
    1377           2 :     if (!InsertedShift && !InsertedTrunc) {
    1378           2 :       BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
    1379             :       assert(InsertPt != TruncUserBB->end());
    1380             :       // Sink the shift
    1381           2 :       if (ShiftI->getOpcode() == Instruction::AShr)
    1382           0 :         InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
    1383             :                                                    "", &*InsertPt);
    1384             :       else
    1385           2 :         InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
    1386             :                                                    "", &*InsertPt);
    1387             : 
    1388             :       // Sink the trunc
    1389           2 :       BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
    1390             :       TruncInsertPt++;
    1391             :       assert(TruncInsertPt != TruncUserBB->end());
    1392             : 
    1393           4 :       InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
    1394             :                                        TruncI->getType(), "", &*TruncInsertPt);
    1395             : 
    1396             :       MadeChange = true;
    1397             : 
    1398             :       TruncTheUse = InsertedTrunc;
    1399             :     }
    1400             :   }
    1401          49 :   return MadeChange;
    1402             : }
    1403             : 
    1404             : /// Sink the shift *right* instruction into user blocks if the uses could
    1405             : /// potentially be combined with this shift instruction and generate BitExtract
    1406             : /// instruction. It will only be applied if the architecture supports BitExtract
    1407             : /// instruction. Here is an example:
    1408             : /// BB1:
    1409             : ///   %x.extract.shift = lshr i64 %arg1, 32
    1410             : /// BB2:
    1411             : ///   %x.extract.trunc = trunc i64 %x.extract.shift to i16
    1412             : /// ==>
    1413             : ///
    1414             : /// BB2:
    1415             : ///   %x.extract.shift.1 = lshr i64 %arg1, 32
    1416             : ///   %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
    1417             : ///
    1418             : /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
    1419             : /// instruction.
    1420             : /// Return true if any changes are made.
    1421        1146 : static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
    1422             :                                 const TargetLowering &TLI,
    1423             :                                 const DataLayout &DL) {
    1424        1146 :   BasicBlock *DefBB = ShiftI->getParent();
    1425             : 
    1426             :   /// Only insert instructions in each block once.
    1427             :   DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
    1428             : 
    1429        1146 :   bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
    1430             : 
    1431             :   bool MadeChange = false;
    1432             :   for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
    1433        2369 :        UI != E;) {
    1434             :     Use &TheUse = UI.getUse();
    1435             :     Instruction *User = cast<Instruction>(*UI);
    1436             :     // Preincrement use iterator so we don't invalidate it.
    1437             :     ++UI;
    1438             : 
    1439             :     // Don't bother for PHI nodes.
    1440        1223 :     if (isa<PHINode>(User))
    1441        1175 :       continue;
    1442             : 
    1443        1218 :     if (!isExtractBitsCandidateUse(User))
    1444         825 :       continue;
    1445             : 
    1446         393 :     BasicBlock *UserBB = User->getParent();
    1447             : 
    1448         733 :     if (UserBB == DefBB) {
    1449             :       // If the shift and truncate instruction are in the same BB. The use of
    1450             :       // the truncate(TruncUse) may still introduce another truncate if not
    1451             :       // legal. In this case, we would like to sink both shift and truncate
    1452             :       // instruction to the BB of TruncUse.
    1453             :       // for example:
    1454             :       // BB1:
    1455             :       // i64 shift.result = lshr i64 opnd, imm
    1456             :       // trunc.result = trunc shift.result to i16
    1457             :       //
    1458             :       // BB2:
    1459             :       //   ----> We will have an implicit truncate here if the architecture does
    1460             :       //   not have i16 compare.
    1461             :       // cmp i16 trunc.result, opnd2
    1462             :       //
    1463         164 :       if (isa<TruncInst>(User) && shiftIsLegal
    1464             :           // If the type of the truncate is legal, no trucate will be
    1465             :           // introduced in other basic blocks.
    1466         340 :           &&
    1467         127 :           (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
    1468          49 :         MadeChange =
    1469             :             SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
    1470             : 
    1471         340 :       continue;
    1472             :     }
    1473             :     // If we have already inserted a shift into this block, use it.
    1474             :     BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
    1475             : 
    1476          53 :     if (!InsertedShift) {
    1477          53 :       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
    1478             :       assert(InsertPt != UserBB->end());
    1479             : 
    1480          53 :       if (ShiftI->getOpcode() == Instruction::AShr)
    1481           8 :         InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
    1482             :                                                    "", &*InsertPt);
    1483             :       else
    1484          45 :         InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
    1485             :                                                    "", &*InsertPt);
    1486             : 
    1487             :       MadeChange = true;
    1488             :     }
    1489             : 
    1490             :     // Replace a use of the shift with a use of the new shift.
    1491          53 :     TheUse = InsertedShift;
    1492             :   }
    1493             : 
    1494             :   // If we removed all uses, nuke the shift.
    1495        1146 :   if (ShiftI->use_empty())
    1496          25 :     ShiftI->eraseFromParent();
    1497             : 
    1498        1146 :   return MadeChange;
    1499             : }
    1500             : 
    1501             : /// If counting leading or trailing zeros is an expensive operation and a zero
    1502             : /// input is defined, add a check for zero to avoid calling the intrinsic.
    1503             : ///
    1504             : /// We want to transform:
    1505             : ///     %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
    1506             : ///
    1507             : /// into:
    1508             : ///   entry:
    1509             : ///     %cmpz = icmp eq i64 %A, 0
    1510             : ///     br i1 %cmpz, label %cond.end, label %cond.false
    1511             : ///   cond.false:
    1512             : ///     %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
    1513             : ///     br label %cond.end
    1514             : ///   cond.end:
    1515             : ///     %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
    1516             : ///
    1517             : /// If the transform is performed, return true and set ModifiedDT to true.
    1518        1733 : static bool despeculateCountZeros(IntrinsicInst *CountZeros,
    1519             :                                   const TargetLowering *TLI,
    1520             :                                   const DataLayout *DL,
    1521             :                                   bool &ModifiedDT) {
    1522        1733 :   if (!TLI || !DL)
    1523             :     return false;
    1524             : 
    1525             :   // If a zero input is undefined, it doesn't make sense to despeculate that.
    1526        3466 :   if (match(CountZeros->getOperand(1), m_One()))
    1527             :     return false;
    1528             : 
    1529             :   // If it's cheap to speculate, there's nothing to do.
    1530             :   auto IntrinsicID = CountZeros->getIntrinsicID();
    1531         794 :   if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
    1532         457 :       (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
    1533             :     return false;
    1534             : 
    1535             :   // Only handle legal scalar cases. Anything else requires too much work.
    1536         509 :   Type *Ty = CountZeros->getType();
    1537         509 :   unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
    1538         509 :   if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits())
    1539             :     return false;
    1540             : 
    1541             :   // The intrinsic will be sunk behind a compare against zero and branch.
    1542          31 :   BasicBlock *StartBlock = CountZeros->getParent();
    1543          31 :   BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
    1544             : 
    1545             :   // Create another block after the count zero intrinsic. A PHI will be added
    1546             :   // in this block to select the result of the intrinsic or the bit-width
    1547             :   // constant if the input to the intrinsic is zero.
    1548          31 :   BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
    1549          31 :   BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
    1550             : 
    1551             :   // Set up a builder to create a compare, conditional branch, and PHI.
    1552          31 :   IRBuilder<> Builder(CountZeros->getContext());
    1553          31 :   Builder.SetInsertPoint(StartBlock->getTerminator());
    1554          31 :   Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
    1555             : 
    1556             :   // Replace the unconditional branch that was created by the first split with
    1557             :   // a compare against zero and a conditional branch.
    1558          31 :   Value *Zero = Constant::getNullValue(Ty);
    1559          31 :   Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
    1560          31 :   Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
    1561          31 :   StartBlock->getTerminator()->eraseFromParent();
    1562             : 
    1563             :   // Create a PHI in the end block to select either the output of the intrinsic
    1564             :   // or the bit width of the operand.
    1565          31 :   Builder.SetInsertPoint(&EndBlock->front());
    1566          31 :   PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
    1567          31 :   CountZeros->replaceAllUsesWith(PN);
    1568          93 :   Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
    1569          31 :   PN->addIncoming(BitWidth, StartBlock);
    1570          31 :   PN->addIncoming(CountZeros, CallBlock);
    1571             : 
    1572             :   // We are explicitly handling the zero case, so we can set the intrinsic's
    1573             :   // undefined zero argument to 'true'. This will also prevent reprocessing the
    1574             :   // intrinsic; we only despeculate when a zero input is defined.
    1575          31 :   CountZeros->setArgOperand(1, Builder.getTrue());
    1576          31 :   ModifiedDT = true;
    1577             :   return true;
    1578             : }
    1579             : 
    1580      494910 : bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) {
    1581      494910 :   BasicBlock *BB = CI->getParent();
    1582             : 
    1583             :   // Lower inline assembly if we can.
    1584             :   // If we found an inline asm expession, and if the target knows how to
    1585             :   // lower it to normal LLVM code, do so now.
    1586      989815 :   if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
    1587       13731 :     if (TLI->ExpandInlineAsm(CI)) {
    1588             :       // Avoid invalidating the iterator.
    1589          49 :       CurInstIterator = BB->begin();
    1590             :       // Avoid processing instructions out of order, which could cause
    1591             :       // reuse before a value is defined.
    1592             :       SunkAddrs.clear();
    1593          49 :       return true;
    1594             :     }
    1595             :     // Sink address computing for memory operands into the block.
    1596       13682 :     if (optimizeInlineAsmInst(CI))
    1597             :       return true;
    1598             :   }
    1599             : 
    1600             :   // Align the pointer arguments to this call if the target thinks it's a good
    1601             :   // idea
    1602             :   unsigned MinSize, PrefAlign;
    1603      494846 :   if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
    1604        3320 :     for (auto &Arg : CI->arg_operands()) {
    1605             :       // We want to align both objects whose address is used directly and
    1606             :       // objects whose address is used in casts and GEPs, though it only makes
    1607             :       // sense for GEPs if the offset is a multiple of the desired alignment and
    1608             :       // if size - offset meets the size threshold.
    1609        5312 :       if (!Arg->getType()->isPointerTy())
    1610        3050 :         continue;
    1611        1154 :       APInt Offset(DL->getIndexSizeInBits(
    1612             :                        cast<PointerType>(Arg->getType())->getAddressSpace()),
    1613             :                    0);
    1614        1154 :       Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
    1615             :       uint64_t Offset2 = Offset.getLimitedValue();
    1616        1154 :       if ((Offset2 & (PrefAlign-1)) != 0)
    1617             :         continue;
    1618             :       AllocaInst *AI;
    1619         287 :       if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
    1620         138 :           DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
    1621          60 :         AI->setAlignment(PrefAlign);
    1622             :       // Global variables can only be aligned if they are defined in this
    1623             :       // object (i.e. they are uniquely initialized in this object), and
    1624             :       // over-aligning global variables that have an explicit section is
    1625             :       // forbidden.
    1626             :       GlobalVariable *GV;
    1627         296 :       if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
    1628         149 :           GV->getPointerAlignment(*DL) < PrefAlign &&
    1629          67 :           DL->getTypeAllocSize(GV->getValueType()) >=
    1630          67 :               MinSize + Offset2)
    1631          39 :         GV->setAlignment(PrefAlign);
    1632             :     }
    1633             :     // If this is a memcpy (or similar) then we may be able to improve the
    1634             :     // alignment
    1635             :     if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
    1636         664 :       unsigned DestAlign = getKnownAlignment(MI->getDest(), *DL);
    1637         664 :       if (DestAlign > MI->getDestAlignment())
    1638         454 :         MI->setDestAlignment(DestAlign);
    1639             :       if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
    1640         490 :         unsigned SrcAlign = getKnownAlignment(MTI->getSource(), *DL);
    1641         490 :         if (SrcAlign > MTI->getSourceAlignment())
    1642         329 :           MTI->setSourceAlignment(SrcAlign);
    1643             :       }
    1644             :     }
    1645             :   }
    1646             : 
    1647             :   // If we have a cold call site, try to sink addressing computation into the
    1648             :   // cold block.  This interacts with our handling for loads and stores to
    1649             :   // ensure that we can fold all uses of a potential addressing computation
    1650             :   // into their uses.  TODO: generalize this to work over profiling data
    1651      983116 :   if (!OptSize && CI->hasFnAttr(Attribute::Cold))
    1652         132 :     for (auto &Arg : CI->arg_operands()) {
    1653         240 :       if (!Arg->getType()->isPointerTy())
    1654             :         continue;
    1655             :       unsigned AS = Arg->getType()->getPointerAddressSpace();
    1656         110 :       return optimizeMemoryInst(CI, Arg, Arg->getType(), AS);
    1657             :     }
    1658             : 
    1659             :   IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
    1660             :   if (II) {
    1661      262294 :     switch (II->getIntrinsicID()) {
    1662             :     default: break;
    1663          29 :     case Intrinsic::objectsize: {
    1664             :       // Lower all uses of llvm.objectsize.*
    1665             :       ConstantInt *RetVal =
    1666          29 :           lowerObjectSizeCall(II, *DL, TLInfo, /*MustSucceed=*/true);
    1667             :       // Substituting this can cause recursive simplifications, which can
    1668             :       // invalidate our iterator.  Use a WeakTrackingVH to hold onto it in case
    1669             :       // this
    1670             :       // happens.
    1671             :       Value *CurValue = &*CurInstIterator;
    1672             :       WeakTrackingVH IterHandle(CurValue);
    1673             : 
    1674          29 :       replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
    1675             : 
    1676             :       // If the iterator instruction was recursively deleted, start over at the
    1677             :       // start of the block.
    1678          29 :       if (IterHandle != CurValue) {
    1679           4 :         CurInstIterator = BB->begin();
    1680             :         SunkAddrs.clear();
    1681             :       }
    1682             :       return true;
    1683             :     }
    1684         114 :     case Intrinsic::aarch64_stlxr:
    1685             :     case Intrinsic::aarch64_stxr: {
    1686         114 :       ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
    1687         174 :       if (!ExtVal || !ExtVal->hasOneUse() ||
    1688          87 :           ExtVal->getParent() == CI->getParent())
    1689             :         return false;
    1690             :       // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
    1691           1 :       ExtVal->moveBefore(CI);
    1692             :       // Mark this instruction as "inserted by CGP", so that other
    1693             :       // optimizations don't touch it.
    1694           1 :       InsertedInsts.insert(ExtVal);
    1695           1 :       return true;
    1696             :     }
    1697           2 :     case Intrinsic::launder_invariant_group:
    1698           4 :       II->replaceAllUsesWith(II->getArgOperand(0));
    1699           2 :       II->eraseFromParent();
    1700           2 :       return true;
    1701             : 
    1702        1733 :     case Intrinsic::cttz:
    1703             :     case Intrinsic::ctlz:
    1704             :       // If counting zeros is expensive, try to avoid it.
    1705        1733 :       return despeculateCountZeros(II, TLI, DL, ModifiedDT);
    1706             :     }
    1707             : 
    1708      260416 :     if (TLI) {
    1709             :       SmallVector<Value*, 2> PtrOps;
    1710             :       Type *AccessTy;
    1711      260415 :       if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
    1712         522 :         while (!PtrOps.empty()) {
    1713             :           Value *PtrVal = PtrOps.pop_back_val();
    1714         269 :           unsigned AS = PtrVal->getType()->getPointerAddressSpace();
    1715         269 :           if (optimizeMemoryInst(II, PtrVal, AccessTy, AS))
    1716             :             return true;
    1717             :         }
    1718             :     }
    1719             :   }
    1720             : 
    1721             :   // From here on out we're working with named functions.
    1722             :   if (!CI->getCalledFunction()) return false;
    1723             : 
    1724             :   // Lower all default uses of _chk calls.  This is very similar
    1725             :   // to what InstCombineCalls does, but here we are only lowering calls
    1726             :   // to fortified library functions (e.g. __memcpy_chk) that have the default
    1727             :   // "don't know" as the objectsize.  Anything else should be left alone.
    1728      476277 :   FortifiedLibCallSimplifier Simplifier(TLInfo, true);
    1729      476277 :   if (Value *V = Simplifier.optimizeCall(CI)) {
    1730          38 :     CI->replaceAllUsesWith(V);
    1731          38 :     CI->eraseFromParent();
    1732          38 :     return true;
    1733             :   }
    1734             : 
    1735             :   return false;
    1736             : }
    1737             : 
    1738             : /// Look for opportunities to duplicate return instructions to the predecessor
    1739             : /// to enable tail call optimizations. The case it is currently looking for is:
    1740             : /// @code
    1741             : /// bb0:
    1742             : ///   %tmp0 = tail call i32 @f0()
    1743             : ///   br label %return
    1744             : /// bb1:
    1745             : ///   %tmp1 = tail call i32 @f1()
    1746             : ///   br label %return
    1747             : /// bb2:
    1748             : ///   %tmp2 = tail call i32 @f2()
    1749             : ///   br label %return
    1750             : /// return:
    1751             : ///   %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
    1752             : ///   ret i32 %retval
    1753             : /// @endcode
    1754             : ///
    1755             : /// =>
    1756             : ///
    1757             : /// @code
    1758             : /// bb0:
    1759             : ///   %tmp0 = tail call i32 @f0()
    1760             : ///   ret i32 %tmp0
    1761             : /// bb1:
    1762             : ///   %tmp1 = tail call i32 @f1()
    1763             : ///   ret i32 %tmp1
    1764             : /// bb2:
    1765             : ///   %tmp2 = tail call i32 @f2()
    1766             : ///   ret i32 %tmp2
    1767             : /// @endcode
    1768      469404 : bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
    1769      469404 :   if (!TLI)
    1770             :     return false;
    1771             : 
    1772             :   ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator());
    1773             :   if (!RetI)
    1774             :     return false;
    1775             : 
    1776             :   PHINode *PN = nullptr;
    1777             :   BitCastInst *BCI = nullptr;
    1778             :   Value *V = RetI->getReturnValue();
    1779             :   if (V) {
    1780             :     BCI = dyn_cast<BitCastInst>(V);
    1781             :     if (BCI)
    1782             :       V = BCI->getOperand(0);
    1783             : 
    1784             :     PN = dyn_cast<PHINode>(V);
    1785             :     if (!PN)
    1786             :       return false;
    1787             :   }
    1788             : 
    1789        2248 :   if (PN && PN->getParent() != BB)
    1790             :     return false;
    1791             : 
    1792             :   // Make sure there are no instructions between the PHI and return, or that the
    1793             :   // return is the first instruction in the block.
    1794       63679 :   if (PN) {
    1795        2083 :     BasicBlock::iterator BI = BB->begin();
    1796        2100 :     do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
    1797        2083 :     if (&*BI == BCI)
    1798             :       // Also skip over the bitcast.
    1799             :       ++BI;
    1800        2083 :     if (&*BI != RetI)
    1801        1072 :       return false;
    1802             :   } else {
    1803       61596 :     BasicBlock::iterator BI = BB->begin();
    1804       61811 :     while (isa<DbgInfoIntrinsic>(BI)) ++BI;
    1805       61596 :     if (&*BI != RetI)
    1806       55233 :       return false;
    1807             :   }
    1808             : 
    1809             :   /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
    1810             :   /// call.
    1811        7374 :   const Function *F = BB->getParent();
    1812             :   SmallVector<CallInst*, 4> TailCalls;
    1813        7374 :   if (PN) {
    1814        6497 :     for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
    1815        2743 :       CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
    1816             :       // Make sure the phi value is indeed produced by the tail call.
    1817         587 :       if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
    1818        3006 :           TLI->mayBeEmittedAsTailCall(CI) &&
    1819          84 :           attributesPermitTailCall(F, CI, RetI, *TLI))
    1820          84 :         TailCalls.push_back(CI);
    1821             :     }
    1822             :   } else {
    1823             :     SmallPtrSet<BasicBlock*, 4> VisitedBBs;
    1824       20202 :     for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
    1825        7476 :       if (!VisitedBBs.insert(*PI).second)
    1826        1293 :         continue;
    1827             : 
    1828             :       BasicBlock::InstListType &InstList = (*PI)->getInstList();
    1829             :       BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
    1830             :       BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
    1831        7291 :       do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
    1832        7199 :       if (RI == RE)
    1833             :         continue;
    1834             : 
    1835        6183 :       CallInst *CI = dyn_cast<CallInst>(&*RI);
    1836        6622 :       if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) &&
    1837         439 :           attributesPermitTailCall(F, CI, RetI, *TLI))
    1838         439 :         TailCalls.push_back(CI);
    1839             :     }
    1840             :   }
    1841             : 
    1842             :   bool Changed = false;
    1843        7897 :   for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
    1844        1046 :     CallInst *CI = TailCalls[i];
    1845             :     CallSite CS(CI);
    1846             : 
    1847             :     // Conservatively require the attributes of the call to match those of the
    1848             :     // return. Ignore noalias because it doesn't affect the call sequence.
    1849         523 :     AttributeList CalleeAttrs = CS.getAttributes();
    1850        1569 :     if (AttrBuilder(CalleeAttrs, AttributeList::ReturnIndex)
    1851         523 :             .removeAttribute(Attribute::NoAlias) !=
    1852        1046 :         AttrBuilder(CalleeAttrs, AttributeList::ReturnIndex)
    1853         523 :             .removeAttribute(Attribute::NoAlias))
    1854          19 :       continue;
    1855             : 
    1856             :     // Make sure the call instruction is followed by an unconditional branch to
    1857             :     // the return block.
    1858         523 :     BasicBlock *CallBB = CI->getParent();
    1859             :     BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
    1860        1043 :     if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
    1861             :       continue;
    1862             : 
    1863             :     // Duplicate the return into CallBB.
    1864         504 :     (void)FoldReturnIntoUncondBranch(RetI, BB, CallBB);
    1865         504 :     ModifiedDT = Changed = true;
    1866             :     ++NumRetsDup;
    1867             :   }
    1868             : 
    1869             :   // If we eliminated all predecessors of the block, delete the block now.
    1870        8051 :   if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
    1871          27 :     BB->eraseFromParent();
    1872             : 
    1873             :   return Changed;
    1874             : }
    1875             : 
    1876             : //===----------------------------------------------------------------------===//
    1877             : // Memory Optimization
    1878             : //===----------------------------------------------------------------------===//
    1879             : 
    1880             : namespace {
    1881             : 
    1882             : /// This is an extended version of TargetLowering::AddrMode
    1883             : /// which holds actual Value*'s for register values.
    1884             : struct ExtAddrMode : public TargetLowering::AddrMode {
    1885             :   Value *BaseReg = nullptr;
    1886             :   Value *ScaledReg = nullptr;
    1887             :   Value *OriginalValue = nullptr;
    1888             : 
    1889             :   enum FieldName {
    1890             :     NoField        = 0x00,
    1891             :     BaseRegField   = 0x01,
    1892             :     BaseGVField    = 0x02,
    1893             :     BaseOffsField  = 0x04,
    1894             :     ScaledRegField = 0x08,
    1895             :     ScaleField     = 0x10,
    1896             :     MultipleFields = 0xff
    1897             :   };
    1898             : 
    1899             :   ExtAddrMode() = default;
    1900             : 
    1901             :   void print(raw_ostream &OS) const;
    1902             :   void dump() const;
    1903             : 
    1904        7706 :   FieldName compare(const ExtAddrMode &other) {
    1905             :     // First check that the types are the same on each field, as differing types
    1906             :     // is something we can't cope with later on.
    1907       14757 :     if (BaseReg && other.BaseReg &&
    1908        7051 :         BaseReg->getType() != other.BaseReg->getType())
    1909             :       return MultipleFields;
    1910        7501 :     if (BaseGV && other.BaseGV &&
    1911             :         BaseGV->getType() != other.BaseGV->getType())
    1912             :       return MultipleFields;
    1913        7077 :     if (ScaledReg && other.ScaledReg &&
    1914           9 :         ScaledReg->getType() != other.ScaledReg->getType())
    1915             :       return MultipleFields;
    1916             : 
    1917             :     // Check each field to see if it differs.
    1918             :     unsigned Result = NoField;
    1919        7068 :     if (BaseReg != other.BaseReg)
    1920             :       Result |= BaseRegField;
    1921        7068 :     if (BaseGV != other.BaseGV)
    1922         147 :       Result |= BaseGVField;
    1923        7068 :     if (BaseOffs != other.BaseOffs)
    1924        1710 :       Result |= BaseOffsField;
    1925        7068 :     if (ScaledReg != other.ScaledReg)
    1926         116 :       Result |= ScaledRegField;
    1927             :     // Don't count 0 as being a different scale, because that actually means
    1928             :     // unscaled (which will already be counted by having no ScaledReg).
    1929        7068 :     if (Scale && other.Scale && Scale != other.Scale)
    1930           1 :       Result |= ScaleField;
    1931             : 
    1932        7068 :     if (countPopulation(Result) > 1)
    1933             :       return MultipleFields;
    1934             :     else
    1935        5926 :       return static_cast<FieldName>(Result);
    1936             :   }
    1937             : 
    1938             :   // An AddrMode is trivial if it involves no calculation i.e. it is just a base
    1939             :   // with no offset.
    1940             :   bool isTrivial() {
    1941             :     // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
    1942             :     // trivial if at most one of these terms is nonzero, except that BaseGV and
    1943             :     // BaseReg both being zero actually means a null pointer value, which we
    1944             :     // consider to be 'non-zero' here.
    1945     1092268 :     return !BaseOffs && !Scale && !(BaseGV && BaseReg);
    1946             :   }
    1947             : 
    1948        2053 :   Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
    1949        2053 :     switch (Field) {
    1950             :     default:
    1951             :       return nullptr;
    1952         640 :     case BaseRegField:
    1953         640 :       return BaseReg;
    1954         160 :     case BaseGVField:
    1955         160 :       return BaseGV;
    1956          24 :     case ScaledRegField:
    1957          24 :       return ScaledReg;
    1958        1229 :     case BaseOffsField:
    1959        1229 :       return ConstantInt::get(IntPtrTy, BaseOffs);
    1960             :     }
    1961             :   }
    1962             : 
    1963          68 :   void SetCombinedField(FieldName Field, Value *V,
    1964             :                         const SmallVectorImpl<ExtAddrMode> &AddrModes) {
    1965          68 :     switch (Field) {
    1966           0 :     default:
    1967           0 :       llvm_unreachable("Unhandled fields are expected to be rejected earlier");
    1968             :       break;
    1969          49 :     case ExtAddrMode::BaseRegField:
    1970          49 :       BaseReg = V;
    1971          49 :       break;
    1972           0 :     case ExtAddrMode::BaseGVField:
    1973             :       // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
    1974             :       // in the BaseReg field.
    1975             :       assert(BaseReg == nullptr);
    1976           0 :       BaseReg = V;
    1977           0 :       BaseGV = nullptr;
    1978           0 :       break;
    1979           5 :     case ExtAddrMode::ScaledRegField:
    1980           5 :       ScaledReg = V;
    1981             :       // If we have a mix of scaled and unscaled addrmodes then we want scale
    1982             :       // to be the scale and not zero.
    1983           5 :       if (!Scale)
    1984           6 :         for (const ExtAddrMode &AM : AddrModes)
    1985           4 :           if (AM.Scale) {
    1986           2 :             Scale = AM.Scale;
    1987           2 :             break;
    1988             :           }
    1989             :       break;
    1990          14 :     case ExtAddrMode::BaseOffsField:
    1991             :       // The offset is no longer a constant, so it goes in ScaledReg with a
    1992             :       // scale of 1.
    1993             :       assert(ScaledReg == nullptr);
    1994          14 :       ScaledReg = V;
    1995          14 :       Scale = 1;
    1996          14 :       BaseOffs = 0;
    1997          14 :       break;
    1998             :     }
    1999          68 :   }
    2000             : };
    2001             : 
    2002             : } // end anonymous namespace
    2003             : 
    2004             : #ifndef NDEBUG
    2005             : static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
    2006             :   AM.print(OS);
    2007             :   return OS;
    2008             : }
    2009             : #endif
    2010             : 
    2011             : #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
    2012             : void ExtAddrMode::print(raw_ostream &OS) const {
    2013             :   bool NeedPlus = false;
    2014             :   OS << "[";
    2015             :   if (BaseGV) {
    2016             :     OS << (NeedPlus ? " + " : "")
    2017             :        << "GV:";
    2018             :     BaseGV->printAsOperand(OS, /*PrintType=*/false);
    2019             :     NeedPlus = true;
    2020             :   }
    2021             : 
    2022             :   if (BaseOffs) {
    2023             :     OS << (NeedPlus ? " + " : "")
    2024             :        << BaseOffs;
    2025             :     NeedPlus = true;
    2026             :   }
    2027             : 
    2028             :   if (BaseReg) {
    2029             :     OS << (NeedPlus ? " + " : "")
    2030             :        << "Base:";
    2031             :     BaseReg->printAsOperand(OS, /*PrintType=*/false);
    2032             :     NeedPlus = true;
    2033             :   }
    2034             :   if (Scale) {
    2035             :     OS << (NeedPlus ? " + " : "")
    2036             :        << Scale << "*";
    2037             :     ScaledReg->printAsOperand(OS, /*PrintType=*/false);
    2038             :   }
    2039             : 
    2040             :   OS << ']';
    2041             : }
    2042             : 
    2043             : LLVM_DUMP_METHOD void ExtAddrMode::dump() const {
    2044             :   print(dbgs());
    2045             :   dbgs() << '\n';
    2046             : }
    2047             : #endif
    2048             : 
    2049             : namespace {
    2050             : 
    2051             : /// This class provides transaction based operation on the IR.
    2052             : /// Every change made through this class is recorded in the internal state and
    2053             : /// can be undone (rollback) until commit is called.
    2054     1122463 : class TypePromotionTransaction {
    2055             :   /// This represents the common interface of the individual transaction.
    2056             :   /// Each class implements the logic for doing one specific modification on
    2057             :   /// the IR via the TypePromotionTransaction.
    2058             :   class TypePromotionAction {
    2059             :   protected:
    2060             :     /// The Instruction modified.
    2061             :     Instruction *Inst;
    2062             : 
    2063             :   public:
    2064             :     /// Constructor of the action.
    2065             :     /// The constructor performs the related action on the IR.
    2066        8833 :     TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
    2067             : 
    2068             :     virtual ~TypePromotionAction() = default;
    2069             : 
    2070             :     /// Undo the modification done by this action.
    2071             :     /// When this method is called, the IR must be in the same state as it was
    2072             :     /// before this action was applied.
    2073             :     /// \pre Undoing the action works if and only if the IR is in the exact same
    2074             :     /// state as it was directly after this action was applied.
    2075             :     virtual void undo() = 0;
    2076             : 
    2077             :     /// Advocate every change made by this action.
    2078             :     /// When the results on the IR of the action are to be kept, it is important
    2079             :     /// to call this function, otherwise hidden information may be kept forever.
    2080           0 :     virtual void commit() {
    2081             :       // Nothing to be done, this action is not doing anything.
    2082           0 :     }
    2083             :   };
    2084             : 
    2085             :   /// Utility to remember the position of an instruction.
    2086             :   class InsertionHandler {
    2087             :     /// Position of an instruction.
    2088             :     /// Either an instruction:
    2089             :     /// - Is the first in a basic block: BB is used.
    2090             :     /// - Has a previous instructon: PrevInst is used.
    2091             :     union {
    2092             :       Instruction *PrevInst;
    2093             :       BasicBlock *BB;
    2094             :     } Point;
    2095             : 
    2096             :     /// Remember whether or not the instruction had a previous instruction.
    2097             :     bool HasPrevInstruction;
    2098             : 
    2099             :   public:
    2100             :     /// Record the position of \p Inst.
    2101        1693 :     InsertionHandler(Instruction *Inst) {
    2102        1693 :       BasicBlock::iterator It = Inst->getIterator();
    2103        3386 :       HasPrevInstruction = (It != (Inst->getParent()->begin()));
    2104        1693 :       if (HasPrevInstruction)
    2105        1601 :         Point.PrevInst = &*--It;
    2106             :       else
    2107          92 :         Point.BB = Inst->getParent();
    2108             :     }
    2109             : 
    2110             :     /// Insert \p Inst at the recorded position.
    2111        1070 :     void insert(Instruction *Inst) {
    2112        1070 :       if (HasPrevInstruction) {
    2113        1015 :         if (Inst->getParent())
    2114         920 :           Inst->removeFromParent();
    2115        1015 :         Inst->insertAfter(Point.PrevInst);
    2116             :       } else {
    2117          55 :         Instruction *Position = &*Point.BB->getFirstInsertionPt();
    2118          55 :         if (Inst->getParent())
    2119          46 :           Inst->moveBefore(Position);
    2120             :         else
    2121           9 :           Inst->insertBefore(Position);
    2122             :       }
    2123        1070 :     }
    2124             :   };
    2125             : 
    2126             :   /// Move an instruction before another.
    2127        1409 :   class InstructionMoveBefore : public TypePromotionAction {
    2128             :     /// Original position of the instruction.
    2129             :     InsertionHandler Position;
    2130             : 
    2131             :   public:
    2132             :     /// Move \p Inst before \p Before.
    2133        1409 :     InstructionMoveBefore(Instruction *Inst, Instruction *Before)
    2134        2818 :         : TypePromotionAction(Inst), Position(Inst) {
    2135             :       LLVM_DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before
    2136             :                         << "\n");
    2137        1409 :       Inst->moveBefore(Before);
    2138        1409 :     }
    2139             : 
    2140             :     /// Move the instruction back to its original position.
    2141         966 :     void undo() override {
    2142             :       LLVM_DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
    2143         966 :       Position.insert(Inst);
    2144         966 :     }
    2145             :   };
    2146             : 
    2147             :   /// Set the operand of an instruction with a new value.
    2148        3809 :   class OperandSetter : public TypePromotionAction {
    2149             :     /// Original operand of the instruction.
    2150             :     Value *Origin;
    2151             : 
    2152             :     /// Index of the modified instruction.
    2153             :     unsigned Idx;
    2154             : 
    2155             :   public:
    2156             :     /// Set \p Idx operand of \p Inst with \p NewVal.
    2157        3809 :     OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
    2158        7618 :         : TypePromotionAction(Inst), Idx(Idx) {
    2159             :       LLVM_DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
    2160             :                         << "for:" << *Inst << "\n"
    2161             :                         << "with:" << *NewVal << "\n");
    2162        7618 :       Origin = Inst->getOperand(Idx);
    2163        3809 :       Inst->setOperand(Idx, NewVal);
    2164        3809 :     }
    2165             : 
    2166             :     /// Restore the original value of the instruction.
    2167        2631 :     void undo() override {
    2168             :       LLVM_DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
    2169             :                         << "for: " << *Inst << "\n"
    2170             :                         << "with: " << *Origin << "\n");
    2171        2631 :       Inst->setOperand(Idx, Origin);
    2172        2631 :     }
    2173             :   };
    2174             : 
    2175             :   /// Hide the operands of an instruction.
    2176             :   /// Do as if this instruction was not using any of its operands.
    2177         284 :   class OperandsHider : public TypePromotionAction {
    2178             :     /// The list of original operands.
    2179             :     SmallVector<Value *, 4> OriginalValues;
    2180             : 
    2181             :   public:
    2182             :     /// Remove \p Inst from the uses of the operands of \p Inst.
    2183         568 :     OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
    2184             :       LLVM_DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
    2185             :       unsigned NumOpnds = Inst->getNumOperands();
    2186         284 :       OriginalValues.reserve(NumOpnds);
    2187         852 :       for (unsigned It = 0; It < NumOpnds; ++It) {
    2188             :         // Save the current operand.
    2189         568 :         Value *Val = Inst->getOperand(It);
    2190         284 :         OriginalValues.push_back(Val);
    2191             :         // Set a dummy one.
    2192             :         // We could use OperandSetter here, but that would imply an overhead
    2193             :         // that we are not willing to pay.
    2194         284 :         Inst->setOperand(It, UndefValue::get(Val->getType()));
    2195             :       }
    2196         284 :     }
    2197             : 
    2198             :     /// Restore the original list of uses.
    2199         104 :     void undo() override {
    2200             :       LLVM_DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
    2201         208 :       for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
    2202         208 :         Inst->setOperand(It, OriginalValues[It]);
    2203         104 :     }
    2204             :   };
    2205             : 
    2206             :   /// Build a truncate instruction.
    2207          99 :   class TruncBuilder : public TypePromotionAction {
    2208             :     Value *Val;
    2209             : 
    2210             :   public:
    2211             :     /// Build a truncate instruction of \p Opnd producing a \p Ty
    2212             :     /// result.
    2213             :     /// trunc Opnd to Ty.
    2214         198 :     TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
    2215          99 :       IRBuilder<> Builder(Opnd);
    2216          99 :       Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
    2217             :       LLVM_DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
    2218          99 :     }
    2219             : 
    2220             :     /// Get the built value.
    2221             :     Value *getBuiltValue() { return Val; }
    2222             : 
    2223             :     /// Remove the built instruction.
    2224          75 :     void undo() override {
    2225             :       LLVM_DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
    2226          75 :       if (Instruction *IVal = dyn_cast<Instruction>(Val))
    2227          75 :         IVal->eraseFromParent();
    2228          75 :     }
    2229             :   };
    2230             : 
    2231             :   /// Build a sign extension instruction.
    2232          99 :   class SExtBuilder : public TypePromotionAction {
    2233             :     Value *Val;
    2234             : 
    2235             :   public:
    2236             :     /// Build a sign extension instruction of \p Opnd producing a \p Ty
    2237             :     /// result.
    2238             :     /// sext Opnd to Ty.
    2239          99 :     SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
    2240         198 :         : TypePromotionAction(InsertPt) {
    2241          99 :       IRBuilder<> Builder(InsertPt);
    2242          99 :       Val = Builder.CreateSExt(Opnd, Ty, "promoted");
    2243             :       LLVM_DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
    2244          99 :     }
    2245             : 
    2246             :     /// Get the built value.
    2247             :     Value *getBuiltValue() { return Val; }
    2248             : 
    2249             :     /// Remove the built instruction.
    2250          56 :     void undo() override {
    2251             :       LLVM_DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
    2252          56 :       if (Instruction *IVal = dyn_cast<Instruction>(Val))
    2253          56 :         IVal->eraseFromParent();
    2254          56 :     }
    2255             :   };
    2256             : 
    2257             :   /// Build a zero extension instruction.
    2258         319 :   class ZExtBuilder : public TypePromotionAction {
    2259             :     Value *Val;
    2260             : 
    2261             :   public:
    2262             :     /// Build a zero extension instruction of \p Opnd producing a \p Ty
    2263             :     /// result.
    2264             :     /// zext Opnd to Ty.
    2265         319 :     ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
    2266         638 :         : TypePromotionAction(InsertPt) {
    2267         319 :       IRBuilder<> Builder(InsertPt);
    2268         319 :       Val = Builder.CreateZExt(Opnd, Ty, "promoted");
    2269             :       LLVM_DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
    2270         319 :     }
    2271             : 
    2272             :     /// Get the built value.
    2273             :     Value *getBuiltValue() { return Val; }
    2274             : 
    2275             :     /// Remove the built instruction.
    2276         166 :     void undo() override {
    2277             :       LLVM_DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
    2278         166 :       if (Instruction *IVal = dyn_cast<Instruction>(Val))
    2279         165 :         IVal->eraseFromParent();
    2280         166 :     }
    2281             :   };
    2282             : 
    2283             :   /// Mutate an instruction to another type.
    2284        1141 :   class TypeMutator : public TypePromotionAction {
    2285             :     /// Record the original type.
    2286             :     Type *OrigTy;
    2287             : 
    2288             :   public:
    2289             :     /// Mutate the type of \p Inst into \p NewTy.
    2290             :     TypeMutator(Instruction *Inst, Type *NewTy)
    2291        2282 :         : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
    2292             :       LLVM_DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
    2293             :                         << "\n");
    2294             :       Inst->mutateType(NewTy);
    2295             :     }
    2296             : 
    2297             :     /// Mutate the instruction back to its original type.
    2298         787 :     void undo() override {
    2299             :       LLVM_DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
    2300             :                         << "\n");
    2301         787 :       Inst->mutateType(OrigTy);
    2302         787 :     }
    2303             :   };
    2304             : 
    2305             :   /// Replace the uses of an instruction by another instruction.
    2306        2778 :   class UsesReplacer : public TypePromotionAction {
    2307             :     /// Helper structure to keep track of the replaced uses.
    2308             :     struct InstructionAndIdx {
    2309             :       /// The instruction using the instruction.
    2310             :       Instruction *Inst;
    2311             : 
    2312             :       /// The index where this instruction is used for Inst.
    2313             :       unsigned Idx;
    2314             : 
    2315             :       InstructionAndIdx(Instruction *Inst, unsigned Idx)
    2316        1814 :           : Inst(Inst), Idx(Idx) {}
    2317             :     };
    2318             : 
    2319             :     /// Keep track of the original uses (pair Instruction, Index).
    2320             :     SmallVector<InstructionAndIdx, 4> OriginalUses;
    2321             : 
    2322             :     using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
    2323             : 
    2324             :   public:
    2325             :     /// Replace all the use of \p Inst by \p New.
    2326        2778 :     UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
    2327             :       LLVM_DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
    2328             :                         << "\n");
    2329             :       // Record the original uses.
    2330        3203 :       for (Use &U : Inst->uses()) {
    2331        1814 :         Instruction *UserI = cast<Instruction>(U.getUser());
    2332        3628 :         OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
    2333             :       }
    2334             :       // Now, we can replace the uses.
    2335        1389 :       Inst->replaceAllUsesWith(New);
    2336        1389 :     }
    2337             : 
    2338             :     /// Reassign the original uses of Inst to Inst.
    2339         910 :     void undo() override {
    2340             :       LLVM_DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
    2341        1251 :       for (use_iterator UseIt = OriginalUses.begin(),
    2342             :                         EndIt = OriginalUses.end();
    2343        2161 :            UseIt != EndIt; ++UseIt) {
    2344        1251 :         UseIt->Inst->setOperand(UseIt->Idx, Inst);
    2345             :       }
    2346         910 :     }
    2347             :   };
    2348             : 
    2349             :   /// Remove an instruction from the IR.
    2350             :   class InstructionRemover : public TypePromotionAction {
    2351             :     /// Original position of the instruction.
    2352             :     InsertionHandler Inserter;
    2353             : 
    2354             :     /// Helper structure to hide all the link to the instruction. In other
    2355             :     /// words, this helps to do as if the instruction was removed.
    2356             :     OperandsHider Hider;
    2357             : 
    2358             :     /// Keep track of the uses replaced, if any.
    2359             :     UsesReplacer *Replacer = nullptr;
    2360             : 
    2361             :     /// Keep track of instructions removed.
    2362             :     SetOfInstrs &RemovedInsts;
    2363             : 
    2364             :   public:
    2365             :     /// Remove all reference of \p Inst and optinally replace all its
    2366             :     /// uses with New.
    2367             :     /// \p RemovedInsts Keep track of the instructions removed by this Action.
    2368             :     /// \pre If !Inst->use_empty(), then New != nullptr
    2369         284 :     InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
    2370             :                        Value *New = nullptr)
    2371         284 :         : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
    2372         568 :           RemovedInsts(RemovedInsts) {
    2373         284 :       if (New)
    2374           5 :         Replacer = new UsesReplacer(Inst, New);
    2375             :       LLVM_DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
    2376         284 :       RemovedInsts.insert(Inst);
    2377             :       /// The instructions removed here will be freed after completing
    2378             :       /// optimizeBlock() for all blocks as we need to keep track of the
    2379             :       /// removed instructions during promotion.
    2380         284 :       Inst->removeFromParent();
    2381         284 :     }
    2382             : 
    2383         852 :     ~InstructionRemover() override { delete Replacer; }
    2384             : 
    2385             :     /// Resurrect the instruction and reassign it to the proper uses if
    2386             :     /// new value was provided when build this action.
    2387         104 :     void undo() override {
    2388             :       LLVM_DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
    2389         104 :       Inserter.insert(Inst);
    2390         104 :       if (Replacer)
    2391           5 :         Replacer->undo();
    2392         104 :       Hider.undo();
    2393         104 :       RemovedInsts.erase(Inst);
    2394         104 :     }
    2395             :   };
    2396             : 
    2397             : public:
    2398             :   /// Restoration point.
    2399             :   /// The restoration point is a pointer to an action instead of an iterator
    2400             :   /// because the iterator may be invalidated but not the pointer.
    2401             :   using ConstRestorationPt = const TypePromotionAction *;
    2402             : 
    2403             :   TypePromotionTransaction(SetOfInstrs &RemovedInsts)
    2404     2244926 :       : RemovedInsts(RemovedInsts) {}
    2405             : 
    2406             :   /// Advocate every changes made in that transaction.
    2407             :   void commit();
    2408             : 
    2409             :   /// Undo all the changes made after the given point.
    2410             :   void rollback(ConstRestorationPt Point);
    2411             : 
    2412             :   /// Get the current restoration point.
    2413             :   ConstRestorationPt getRestorationPoint() const;
    2414             : 
    2415             :   /// \name API for IR modification with state keeping to support rollback.
    2416             :   /// @{
    2417             :   /// Same as Instruction::setOperand.
    2418             :   void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
    2419             : 
    2420             :   /// Same as Instruction::eraseFromParent.
    2421             :   void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
    2422             : 
    2423             :   /// Same as Value::replaceAllUsesWith.
    2424             :   void replaceAllUsesWith(Instruction *Inst, Value *New);
    2425             : 
    2426             :   /// Same as Value::mutateType.
    2427             :   void mutateType(Instruction *Inst, Type *NewTy);
    2428             : 
    2429             :   /// Same as IRBuilder::createTrunc.
    2430             :   Value *createTrunc(Instruction *Opnd, Type *Ty);
    2431             : 
    2432             :   /// Same as IRBuilder::createSExt.
    2433             :   Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
    2434             : 
    2435             :   /// Same as IRBuilder::createZExt.
    2436             :   Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
    2437             : 
    2438             :   /// Same as Instruction::moveBefore.
    2439             :   void moveBefore(Instruction *Inst, Instruction *Before);
    2440             :   /// @}
    2441             : 
    2442             : private:
    2443             :   /// The ordered list of actions made so far.
    2444             :   SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
    2445             : 
    2446             :   using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
    2447             : 
    2448             :   SetOfInstrs &RemovedInsts;
    2449             : };
    2450             : 
    2451             : } // end anonymous namespace
    2452             : 
    2453        3809 : void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
    2454             :                                           Value *NewVal) {
    2455       11427 :   Actions.push_back(llvm::make_unique<TypePromotionTransaction::OperandSetter>(
    2456             :       Inst, Idx, NewVal));
    2457        3809 : }
    2458             : 
    2459         284 : void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
    2460             :                                                 Value *NewVal) {
    2461         852 :   Actions.push_back(
    2462         284 :       llvm::make_unique<TypePromotionTransaction::InstructionRemover>(
    2463         284 :           Inst, RemovedInsts, NewVal));
    2464         284 : }
    2465             : 
    2466        1384 : void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
    2467             :                                                   Value *New) {
    2468        4152 :   Actions.push_back(
    2469             :       llvm::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
    2470        1384 : }
    2471             : 
    2472        1141 : void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
    2473        3423 :   Actions.push_back(
    2474             :       llvm::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
    2475        1141 : }
    2476             : 
    2477          99 : Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
    2478             :                                              Type *Ty) {
    2479          99 :   std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
    2480          99 :   Value *Val = Ptr->getBuiltValue();
    2481         297 :   Actions.push_back(std::move(Ptr));
    2482          99 :   return Val;
    2483             : }
    2484             : 
    2485          99 : Value *TypePromotionTransaction::createSExt(Instruction *Inst,
    2486             :                                             Value *Opnd, Type *Ty) {
    2487          99 :   std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
    2488          99 :   Value *Val = Ptr->getBuiltValue();
    2489         297 :   Actions.push_back(std::move(Ptr));
    2490          99 :   return Val;
    2491             : }
    2492             : 
    2493         319 : Value *TypePromotionTransaction::createZExt(Instruction *Inst,
    2494             :                                             Value *Opnd, Type *Ty) {
    2495         319 :   std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
    2496         319 :   Value *Val = Ptr->getBuiltValue();
    2497         957 :   Actions.push_back(std::move(Ptr));
    2498         319 :   return Val;
    2499             : }
    2500             : 
    2501        1409 : void TypePromotionTransaction::moveBefore(Instruction *Inst,
    2502             :                                           Instruction *Before) {
    2503        4227 :   Actions.push_back(
    2504             :       llvm::make_unique<TypePromotionTransaction::InstructionMoveBefore>(
    2505             :           Inst, Before));
    2506        1409 : }
    2507             : 
    2508             : TypePromotionTransaction::ConstRestorationPt
    2509             : TypePromotionTransaction::getRestorationPoint() const {
    2510     2117628 :   return !Actions.empty() ? Actions.back().get() : nullptr;
    2511             : }
    2512             : 
    2513             : void TypePromotionTransaction::commit() {
    2514             :   for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
    2515             :        ++It)
    2516             :     (*It)->commit();
    2517             :   Actions.clear();
    2518             : }
    2519             : 
    2520       48408 : void TypePromotionTransaction::rollback(
    2521             :     TypePromotionTransaction::ConstRestorationPt Point) {
    2522       65987 :   while (!Actions.empty() && Point != Actions.back().get()) {
    2523             :     std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
    2524        5690 :     Curr->undo();
    2525             :   }
    2526       48408 : }
    2527             : 
    2528             : namespace {
    2529             : 
    2530             : /// A helper class for matching addressing modes.
    2531             : ///
    2532             : /// This encapsulates the logic for matching the target-legal addressing modes.
    2533             : class AddressingModeMatcher {
    2534             :   SmallVectorImpl<Instruction*> &AddrModeInsts;
    2535             :   const TargetLowering &TLI;
    2536             :   const TargetRegisterInfo &TRI;
    2537             :   const DataLayout &DL;
    2538             : 
    2539             :   /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
    2540             :   /// the memory instruction that we're computing this address for.
    2541             :   Type *AccessTy;
    2542             :   unsigned AddrSpace;
    2543             :   Instruction *MemoryInst;
    2544             : 
    2545             :   /// This is the addressing mode that we're building up. This is
    2546             :   /// part of the return value of this addressing mode matching stuff.
    2547             :   ExtAddrMode &AddrMode;
    2548             : 
    2549             :   /// The instructions inserted by other CodeGenPrepare optimizations.
    2550             :   const SetOfInstrs &InsertedInsts;
    2551             : 
    2552             :   /// A map from the instructions to their type before promotion.
    2553             :   InstrToOrigTy &PromotedInsts;
    2554             : 
    2555             :   /// The ongoing transaction where every action should be registered.
    2556             :   TypePromotionTransaction &TPT;
    2557             : 
    2558             :   // A GEP which has too large offset to be folded into the addressing mode.
    2559             :   std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP;
    2560             : 
    2561             :   /// This is set to true when we should not do profitability checks.
    2562             :   /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
    2563             :   bool IgnoreProfitability;
    2564             : 
    2565             :   AddressingModeMatcher(
    2566             :       SmallVectorImpl<Instruction *> &AMI, const TargetLowering &TLI,
    2567             :       const TargetRegisterInfo &TRI, Type *AT, unsigned AS, Instruction *MI,
    2568             :       ExtAddrMode &AM, const SetOfInstrs &InsertedInsts,
    2569             :       InstrToOrigTy &PromotedInsts, TypePromotionTransaction &TPT,
    2570             :       std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP)
    2571     1103040 :       : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
    2572     1103040 :         DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
    2573             :         MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
    2574     2206080 :         PromotedInsts(PromotedInsts), TPT(TPT), LargeOffsetGEP(LargeOffsetGEP) {
    2575     1097985 :     IgnoreProfitability = false;
    2576             :   }
    2577             : 
    2578             : public:
    2579             :   /// Find the maximal addressing mode that a load/store of V can fold,
    2580             :   /// give an access type of AccessTy.  This returns a list of involved
    2581             :   /// instructions in AddrModeInsts.
    2582             :   /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
    2583             :   /// optimizations.
    2584             :   /// \p PromotedInsts maps the instructions to their type before promotion.
    2585             :   /// \p The ongoing transaction where every action should be registered.
    2586             :   static ExtAddrMode
    2587     1097985 :   Match(Value *V, Type *AccessTy, unsigned AS, Instruction *MemoryInst,
    2588             :         SmallVectorImpl<Instruction *> &AddrModeInsts,
    2589             :         const TargetLowering &TLI, const TargetRegisterInfo &TRI,
    2590             :         const SetOfInstrs &InsertedInsts, InstrToOrigTy &PromotedInsts,
    2591             :         TypePromotionTransaction &TPT,
    2592             :         std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP) {
    2593     1097985 :     ExtAddrMode Result;
    2594             : 
    2595     1097985 :     bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI, AccessTy, AS,
    2596             :                                          MemoryInst, Result, InsertedInsts,
    2597             :                                          PromotedInsts, TPT, LargeOffsetGEP)
    2598     1097985 :                        .matchAddr(V, 0);
    2599             :     (void)Success; assert(Success && "Couldn't select *anything*?");
    2600     1097985 :     return Result;
    2601             :   }
    2602             : 
    2603             : private:
    2604             :   bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
    2605             :   bool matchAddr(Value *V, unsigned Depth);
    2606             :   bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
    2607             :                           bool *MovedAway = nullptr);
    2608             :   bool isProfitableToFoldIntoAddressingMode(Instruction *I,
    2609             :                                             ExtAddrMode &AMBefore,
    2610             :                                             ExtAddrMode &AMAfter);
    2611             :   bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
    2612             :   bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
    2613             :                              Value *PromotedOperand) const;
    2614             : };
    2615             : 
    2616             : /// Keep track of simplification of Phi nodes.
    2617             : /// Accept the set of all phi nodes and erase phi node from this set
    2618             : /// if it is simplified.
    2619        2012 : class SimplificationTracker {
    2620             :   DenseMap<Value *, Value *> Storage;
    2621             :   const SimplifyQuery &SQ;
    2622             :   // Tracks newly created Phi nodes. We use a SetVector to get deterministic
    2623             :   // order when iterating over the set in MatchPhiSet.
    2624             :   SmallSetVector<PHINode *, 32> AllPhiNodes;
    2625             :   // Tracks newly created Select nodes.
    2626             :   SmallPtrSet<SelectInst *, 32> AllSelectNodes;
    2627             : 
    2628             : public:
    2629        1006 :   SimplificationTracker(const SimplifyQuery &sq)
    2630        3018 :       : SQ(sq) {}
    2631             : 
    2632             :   Value *Get(Value *V) {
    2633             :     do {
    2634        5275 :       auto SV = Storage.find(V);
    2635        5275 :       if (SV == Storage.end())
    2636        4689 :         return V;
    2637         586 :       V = SV->second;
    2638             :     } while (true);
    2639             :   }
    2640             : 
    2641        1522 :   Value *Simplify(Value *Val) {
    2642             :     SmallVector<Value *, 32> WorkList;
    2643             :     SmallPtrSet<Value *, 32> Visited;
    2644        1522 :     WorkList.push_back(Val);
    2645        3144 :     while (!WorkList.empty()) {
    2646             :       auto P = WorkList.pop_back_val();
    2647        1622 :       if (!Visited.insert(P).second)
    2648          51 :         continue;
    2649             :       if (auto *PI = dyn_cast<Instruction>(P))
    2650        1571 :         if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) {
    2651         586 :           for (auto *U : PI->users())
    2652         100 :             WorkList.push_back(cast<Value>(U));
    2653             :           Put(PI, V);
    2654         486 :           PI->replaceAllUsesWith(V);
    2655         486 :           if (auto *PHI = dyn_cast<PHINode>(PI))
    2656         486 :             AllPhiNodes.remove(PHI);
    2657             :           if (auto *Select = dyn_cast<SelectInst>(PI))
    2658             :             AllSelectNodes.erase(Select);
    2659         486 :           PI->eraseFromParent();
    2660             :         }
    2661             :     }
    2662        3044 :     return Get(Val);
    2663             :   }
    2664             : 
    2665             :   void Put(Value *From, Value *To) {
    2666        1008 :     Storage.insert({ From, To });
    2667             :   }
    2668             : 
    2669          36 :   void ReplacePhi(PHINode *From, PHINode *To) {
    2670          36 :     Value* OldReplacement = Get(From);
    2671          38 :     while (OldReplacement != From) {
    2672           1 :       From = To;
    2673             :       To = dyn_cast<PHINode>(OldReplacement);
    2674             :       OldReplacement = Get(From);
    2675             :     }
    2676             :     assert(Get(To) == To && "Replacement PHI node is already replaced.");
    2677             :     Put(From, To);
    2678          36 :     From->replaceAllUsesWith(To);
    2679          36 :     AllPhiNodes.remove(From);
    2680          36 :     From->eraseFromParent();
    2681          36 :   }
    2682             : 
    2683        1006 :   SmallSetVector<PHINode *, 32>& newPhiNodes() { return AllPhiNodes; }
    2684             : 
    2685        1498 :   void insertNewPhi(PHINode *PN) { AllPhiNodes.insert(PN); }
    2686             : 
    2687          24 :   void insertNewSelect(SelectInst *SI) { AllSelectNodes.insert(SI); }
    2688             : 
    2689             :   unsigned countNewPhiNodes() const { return AllPhiNodes.size(); }
    2690             : 
    2691           0 :   unsigned countNewSelectNodes() const { return AllSelectNodes.size(); }
    2692             : 
    2693         938 :   void destroyNewNodes(Type *CommonType) {
    2694             :     // For safe erasing, replace the uses with dummy value first.
    2695         938 :     auto Dummy = UndefValue::get(CommonType);
    2696        2840 :     for (auto I : AllPhiNodes) {
    2697         951 :       I->replaceAllUsesWith(Dummy);
    2698         951 :       I->eraseFromParent();
    2699             :     }
    2700             :     AllPhiNodes.clear();
    2701         938 :     for (auto I : AllSelectNodes) {
    2702           2 :       I->replaceAllUsesWith(Dummy);
    2703           2 :       I->eraseFromParent();
    2704             :     }
    2705         938 :     AllSelectNodes.clear();
    2706         938 :   }
    2707             : };
    2708             : 
    2709             : /// A helper class for combining addressing modes.
    2710     1090279 : class AddressingModeCombiner {
    2711             :   typedef std::pair<Value *, BasicBlock *> ValueInBB;
    2712             :   typedef DenseMap<ValueInBB, Value *> FoldAddrToValueMapping;
    2713             :   typedef std::pair<PHINode *, PHINode *> PHIPair;
    2714             : 
    2715             : private:
    2716             :   /// The addressing modes we've collected.
    2717             :   SmallVector<ExtAddrMode, 16> AddrModes;
    2718             : 
    2719             :   /// The field in which the AddrModes differ, when we have more than one.
    2720             :   ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
    2721             : 
    2722             :   /// Are the AddrModes that we have all just equal to their original values?
    2723             :   bool AllAddrModesTrivial = true;
    2724             : 
    2725             :   /// Common Type for all different fields in addressing modes.
    2726             :   Type *CommonType;
    2727             : 
    2728             :   /// SimplifyQuery for simplifyInstruction utility.
    2729             :   const SimplifyQuery &SQ;
    2730             : 
    2731             :   /// Original Address.
    2732             :   ValueInBB Original;
    2733             : 
    2734             : public:
    2735             :   AddressingModeCombiner(const SimplifyQuery &_SQ, ValueInBB OriginalValue)
    2736     2180558 :       : CommonType(nullptr), SQ(_SQ), Original(OriginalValue) {}
    2737             : 
    2738             :   /// Get the combined AddrMode
    2739             :   const ExtAddrMode &getAddrMode() const {
    2740             :     return AddrModes[0];
    2741             :   }
    2742             : 
    2743             :   /// Add a new AddrMode if it's compatible with the AddrModes we already
    2744             :   /// have.
    2745             :   /// \return True iff we succeeded in doing so.
    2746     1097985 :   bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
    2747             :     // Take note of if we have any non-trivial AddrModes, as we need to detect
    2748             :     // when all AddrModes are trivial as then we would introduce a phi or select
    2749             :     // which just duplicates what's already there.
    2750     2195970 :     AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
    2751             : 
    2752             :     // If this is the first addrmode then everything is fine.
    2753     1097985 :     if (AddrModes.empty()) {
    2754     1090279 :       AddrModes.emplace_back(NewAddrMode);
    2755     1090279 :       return true;
    2756             :     }
    2757             : 
    2758             :     // Figure out how different this is from the other address modes, which we
    2759             :     // can do just by comparing against the first one given that we only care
    2760             :     // about the cumulative difference.
    2761             :     ExtAddrMode::FieldName ThisDifferentField =
    2762        7706 :       AddrModes[0].compare(NewAddrMode);
    2763        7706 :     if (DifferentField == ExtAddrMode::NoField)
    2764        7513 :       DifferentField = ThisDifferentField;
    2765         193 :     else if (DifferentField != ThisDifferentField)
    2766          76 :       DifferentField = ExtAddrMode::MultipleFields;
    2767             : 
    2768             :     // If NewAddrMode differs in more than one dimension we cannot handle it.
    2769        7706 :     bool CanHandle = DifferentField != ExtAddrMode::MultipleFields;
    2770             : 
    2771             :     // If Scale Field is different then we reject.
    2772        7706 :     CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField;
    2773             : 
    2774             :     // We also must reject the case when base offset is different and
    2775             :     // scale reg is not null, we cannot handle this case due to merge of
    2776             :     // different offsets will be used as ScaleReg.
    2777        6549 :     CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField ||
    2778         628 :                               !NewAddrMode.ScaledReg);
    2779             : 
    2780             :     // We also must reject the case when GV is different and BaseReg installed
    2781             :     // due to we want to use base reg as a merge of GV values.
    2782        6059 :     CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField ||
    2783         139 :                               !NewAddrMode.HasBaseReg);
    2784             : 
    2785             :     // Even if NewAddMode is the same we still need to collect it due to
    2786             :     // original value is different. And later we will need all original values
    2787             :     // as anchors during finding the common Phi node.
    2788             :     if (CanHandle)
    2789        5919 :       AddrModes.emplace_back(NewAddrMode);
    2790             :     else
    2791             :       AddrModes.clear();
    2792             : 
    2793             :     return CanHandle;
    2794             :   }
    2795             : 
    2796             :   /// Combine the addressing modes we've collected into a single
    2797             :   /// addressing mode.
    2798             :   /// \return True iff we successfully combined them or we only had one so
    2799             :   /// didn't need to combine them anyway.
    2800     1090279 :   bool combineAddrModes() {
    2801             :     // If we have no AddrModes then they can't be combined.
    2802     1090279 :     if (AddrModes.size() == 0)
    2803             :       return false;
    2804             : 
    2805             :     // A single AddrMode can trivially be combined.
    2806     1088492 :     if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
    2807             :       return true;
    2808             : 
    2809             :     // If the AddrModes we collected are all just equal to the value they are
    2810             :     // derived from then combining them wouldn't do anything useful.
    2811        1740 :     if (AllAddrModesTrivial)
    2812             :       return false;
    2813             : 
    2814        1010 :     if (!addrModeCombiningAllowed())
    2815             :       return false;
    2816             : 
    2817             :     // Build a map between <original value, basic block where we saw it> to
    2818             :     // value of base register.
    2819             :     // Bail out if there is no common type.
    2820             :     FoldAddrToValueMapping Map;
    2821        1010 :     if (!initializeMap(Map))
    2822             :       return false;
    2823             : 
    2824        1006 :     Value *CommonValue = findCommon(Map);
    2825        1006 :     if (CommonValue)
    2826          68 :       AddrModes[0].SetCombinedField(DifferentField, CommonValue, AddrModes);
    2827        1006 :     return CommonValue != nullptr;
    2828             :   }
    2829             : 
    2830             : private:
    2831             :   /// Initialize Map with anchor values. For address seen in some BB
    2832             :   /// we set the value of different field saw in this address.
    2833             :   /// If address is not an instruction than basic block is set to null.
    2834             :   /// At the same time we find a common type for different field we will
    2835             :   /// use to create new Phi/Select nodes. Keep it in CommonType field.
    2836             :   /// Return false if there is no common type found.
    2837        1010 :   bool initializeMap(FoldAddrToValueMapping &Map) {
    2838             :     // Keep track of keys where the value is null. We will need to replace it
    2839             :     // with constant null when we know the common type.
    2840             :     SmallVector<ValueInBB, 2> NullValue;
    2841        2020 :     Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
    2842        5108 :     for (auto &AM : AddrModes) {
    2843             :       BasicBlock *BB = nullptr;
    2844        2053 :       if (Instruction *I = dyn_cast<Instruction>(AM.OriginalValue))
    2845         483 :         BB = I->getParent();
    2846             : 
    2847        2053 :       Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy);
    2848        2053 :       if (DV) {
    2849        2039 :         auto *Type = DV->getType();
    2850        2039 :         if (CommonType && CommonType != Type)
    2851             :           return false;
    2852        2035 :         CommonType = Type;
    2853        2035 :         Map[{ AM.OriginalValue, BB }] = DV;
    2854             :       } else {
    2855          14 :         NullValue.push_back({ AM.OriginalValue, BB });
    2856             :       }
    2857             :     }
    2858             :     assert(CommonType && "At least one non-null value must be!");
    2859        1026 :     for (auto VIBB : NullValue)
    2860          20 :       Map[VIBB] = Constant::getNullValue(CommonType);
    2861        1006 :     return true;
    2862             :   }
    2863             : 
    2864             :   /// We have mapping between value A and basic block where value A
    2865             :   /// seen to other value B where B was a field in addressing mode represented
    2866             :   /// by A. Also we have an original value C representin an address in some
    2867             :   /// basic block. Traversing from C through phi and selects we ended up with
    2868             :   /// A's in a map. This utility function tries to find a value V which is a
    2869             :   /// field in addressing mode C and traversing through phi nodes and selects
    2870             :   /// we will end up in corresponded values B in a map.
    2871             :   /// The utility will create a new Phi/Selects if needed.
    2872             :   // The simple example looks as follows:
    2873             :   // BB1:
    2874             :   //   p1 = b1 + 40
    2875             :   //   br cond BB2, BB3
    2876             :   // BB2:
    2877             :   //   p2 = b2 + 40
    2878             :   //   br BB3
    2879             :   // BB3:
    2880             :   //   p = phi [p1, BB1], [p2, BB2]
    2881             :   //   v = load p
    2882             :   // Map is
    2883             :   //   <p1, BB1> -> b1
    2884             :   //   <p2, BB2> -> b2
    2885             :   // Request is
    2886             :   //   <p, BB3> -> ?
    2887             :   // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3
    2888        1006 :   Value *findCommon(FoldAddrToValueMapping &Map) {
    2889             :     // Tracks the simplification of newly created phi nodes. The reason we use
    2890             :     // this mapping is because we will add new created Phi nodes in AddrToBase.
    2891             :     // Simplification of Phi nodes is recursive, so some Phi node may
    2892             :     // be simplified after we added it to AddrToBase.
    2893             :     // Using this mapping we can find the current value in AddrToBase.
    2894        2012 :     SimplificationTracker ST(SQ);
    2895             : 
    2896             :     // First step, DFS to create PHI nodes for all intermediate blocks.
    2897             :     // Also fill traverse order for the second step.
    2898             :     SmallVector<ValueInBB, 32> TraverseOrder;
    2899        1006 :     InsertPlaceholders(Map, TraverseOrder, ST);
    2900             : 
    2901             :     // Second Step, fill new nodes by merged values and simplify if possible.
    2902        1006 :     FillPlaceholders(Map, TraverseOrder, ST);
    2903             : 
    2904        1006 :     if (!AddrSinkNewSelects && ST.countNewSelectNodes() > 0) {
    2905           0 :       ST.destroyNewNodes(CommonType);
    2906           0 :       return nullptr;
    2907             :     }
    2908             : 
    2909             :     // Now we'd like to match New Phi nodes to existed ones.
    2910        1006 :     unsigned PhiNotMatchedCount = 0;
    2911        1006 :     if (!MatchPhiSet(ST, AddrSinkNewPhis, PhiNotMatchedCount)) {
    2912         938 :       ST.destroyNewNodes(CommonType);
    2913         938 :       return nullptr;
    2914             :     }
    2915             : 
    2916         136 :     auto *Result = ST.Get(Map.find(Original)->second);
    2917             :     if (Result) {
    2918             :       NumMemoryInstsPhiCreated += ST.countNewPhiNodes() + PhiNotMatchedCount;
    2919             :       NumMemoryInstsSelectCreated += ST.countNewSelectNodes();
    2920             :     }
    2921             :     return Result;
    2922             :   }
    2923             : 
    2924             :   /// Try to match PHI node to Candidate.
    2925             :   /// Matcher tracks the matched Phi nodes.
    2926        1758 :   bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
    2927             :                     SmallSetVector<PHIPair, 8> &Matcher,
    2928             :                     SmallSetVector<PHINode *, 32> &PhiNodesToMatch) {
    2929             :     SmallVector<PHIPair, 8> WorkList;
    2930        3516 :     Matcher.insert({ PHI, Candidate });
    2931        1758 :     WorkList.push_back({ PHI, Candidate });
    2932        1758 :     SmallSet<PHIPair, 8> Visited;
    2933        1799 :     while (!WorkList.empty()) {
    2934        1767 :       auto Item = WorkList.pop_back_val();
    2935        1767 :       if (!Visited.insert(Item).second)
    2936           0 :         continue;
    2937             :       // We iterate over all incoming values to Phi to compare them.
    2938             :       // If values are different and both of them Phi and the first one is a
    2939             :       // Phi we added (subject to match) and both of them is in the same basic
    2940             :       // block then we can match our pair if values match. So we state that
    2941             :       // these values match and add it to work list to verify that.
    2942        3704 :       for (auto B : Item.first->blocks()) {
    2943        1811 :         Value *FirstValue = Item.first->getIncomingValueForBlock(B);
    2944        1811 :         Value *SecondValue = Item.second->getIncomingValueForBlock(B);
    2945        1811 :         if (FirstValue == SecondValue)
    2946             :           continue;
    2947             : 
    2948             :         PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue);
    2949             :         PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue);
    2950             : 
    2951             :         // One of them is not Phi or
    2952             :         // The first one is not Phi node from the set we'd like to match or
    2953             :         // Phi nodes from different basic blocks then
    2954             :         // we will not be able to match.
    2955        1805 :         if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) ||
    2956          16 :             FirstPhi->getParent() != SecondPhi->getParent())
    2957             :           return false;
    2958             : 
    2959             :         // If we already matched them then continue.
    2960          13 :         if (Matcher.count({ FirstPhi, SecondPhi }))
    2961             :           continue;
    2962             :         // So the values are different and does not match. So we need them to
    2963             :         // match.
    2964          10 :         Matcher.insert({ FirstPhi, SecondPhi });
    2965             :         // But me must check it.
    2966          10 :         WorkList.push_back({ FirstPhi, SecondPhi });
    2967             :       }
    2968             :     }
    2969             :     return true;
    2970             :   }
    2971             : 
    2972             :   /// For the given set of PHI nodes (in the SimplificationTracker) try
    2973             :   /// to find their equivalents.
    2974             :   /// Returns false if this matching fails and creation of new Phi is disabled.
    2975        1006 :   bool MatchPhiSet(SimplificationTracker &ST, bool AllowNewPhiNodes,
    2976             :                    unsigned &PhiNotMatchedCount) {
    2977             :     // Use a SetVector for Matched to make sure we do replacements (ReplacePhi)
    2978             :     // in a deterministic order below.
    2979             :     SmallSetVector<PHIPair, 8> Matched;
    2980             :     SmallPtrSet<PHINode *, 8> WillNotMatch;
    2981             :     SmallSetVector<PHINode *, 32> &PhiNodesToMatch = ST.newPhiNodes();
    2982        1061 :     while (PhiNodesToMatch.size()) {
    2983         993 :       PHINode *PHI = *PhiNodesToMatch.begin();
    2984             : 
    2985             :       // Add us, if no Phi nodes in the basic block we do not match.
    2986         993 :       WillNotMatch.clear();
    2987         993 :       WillNotMatch.insert(PHI);
    2988             : 
    2989             :       // Traverse all Phis until we found equivalent or fail to do that.
    2990             :       bool IsMatched = false;
    2991         993 :       for (auto &P : PHI->getParent()->phis()) {
    2992        2751 :         if (&P == PHI)
    2993         993 :           continue;
    2994        1758 :         if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch)))
    2995             :           break;
    2996             :         // If it does not match, collect all Phi nodes from matcher.
    2997             :         // if we end up with no match, them all these Phi nodes will not match
    2998             :         // later.
    2999        5190 :         for (auto M : Matched)
    3000        1732 :           WillNotMatch.insert(M.first);
    3001             :         Matched.clear();
    3002             :       }
    3003         993 :       if (IsMatched) {
    3004             :         // Replace all matched values and erase them.
    3005         104 :         for (auto MV : Matched)
    3006          36 :           ST.ReplacePhi(MV.first, MV.second);
    3007             :         Matched.clear();
    3008          32 :         continue;
    3009             :       }
    3010             :       // If we are not allowed to create new nodes then bail out.
    3011         961 :       if (!AllowNewPhiNodes)
    3012             :         return false;
    3013             :       // Just remove all seen values in matcher. They will not match anything.
    3014          46 :       PhiNotMatchedCount += WillNotMatch.size();
    3015          49 :       for (auto *P : WillNotMatch)
    3016          26 :         PhiNodesToMatch.remove(P);
    3017             :     }
    3018             :     return true;
    3019             :   }
    3020             :   /// Fill the placeholder with values from predecessors and simplify it.
    3021        1006 :   void FillPlaceholders(FoldAddrToValueMapping &Map,
    3022             :                         SmallVectorImpl<ValueInBB> &TraverseOrder,
    3023             :                         SimplificationTracker &ST) {
    3024        2528 :     while (!TraverseOrder.empty()) {
    3025        1522 :       auto Current = TraverseOrder.pop_back_val();
    3026             :       assert(Map.find(Current) != Map.end() && "No node to fill!!!");
    3027             :       Value *CurrentValue = Current.first;
    3028             :       BasicBlock *CurrentBlock = Current.second;
    3029        1522 :       Value *V = Map[Current];
    3030             : 
    3031             :       if (SelectInst *Select = dyn_cast<SelectInst>(V)) {
    3032             :         // CurrentValue also must be Select.
    3033             :         auto *CurrentSelect = cast<SelectInst>(CurrentValue);
    3034             :         auto *TrueValue = CurrentSelect->getTrueValue();
    3035             :         ValueInBB TrueItem = { TrueValue, isa<Instruction>(TrueValue)
    3036          24 :                                               ? CurrentBlock
    3037             :                                               : nullptr };
    3038             :         assert(Map.find(TrueItem) != Map.end() && "No True Value!");
    3039          24 :         Select->setTrueValue(ST.Get(Map[TrueItem]));
    3040             :         auto *FalseValue = CurrentSelect->getFalseValue();
    3041             :         ValueInBB FalseItem = { FalseValue, isa<Instruction>(FalseValue)
    3042          24 :                                                 ? CurrentBlock
    3043             :                                                 : nullptr };
    3044             :         assert(Map.find(FalseItem) != Map.end() && "No False Value!");
    3045          24 :         Select->setFalseValue(ST.Get(Map[FalseItem]));
    3046             :       } else {
    3047             :         // Must be a Phi node then.
    3048             :         PHINode *PHI = cast<PHINode>(V);
    3049             :         // Fill the Phi node with values from predecessors.
    3050             :         bool IsDefinedInThisBB =
    3051        1498 :             cast<Instruction>(CurrentValue)->getParent() == CurrentBlock;
    3052             :         auto *CurrentPhi = dyn_cast<PHINode>(CurrentValue);
    3053        6010 :         for (auto B : predecessors(CurrentBlock)) {
    3054             :           Value *PV = IsDefinedInThisBB
    3055        3014 :                           ? CurrentPhi->getIncomingValueForBlock(B)
    3056             :                           : CurrentValue;
    3057        3014 :           ValueInBB item = { PV, isa<Instruction>(PV) ? B : nullptr };
    3058             :           assert(Map.find(item) != Map.end() && "No predecessor Value!");
    3059        6028 :           PHI->addIncoming(ST.Get(Map[item]), B);
    3060             :         }
    3061             :       }
    3062             :       // Simplify if possible.
    3063        1522 :       Map[Current] = ST.Simplify(V);
    3064             :     }
    3065        1006 :   }
    3066             : 
    3067             :   /// Starting from value recursively iterates over predecessors up to known
    3068             :   /// ending values represented in a map. For each traversed block inserts
    3069             :   /// a placeholder Phi or Select.
    3070             :   /// Reports all new created Phi/Select nodes by adding them to set.
    3071             :   /// Also reports and order in what basic blocks have been traversed.
    3072        1006 :   void InsertPlaceholders(FoldAddrToValueMapping &Map,
    3073             :                           SmallVectorImpl<ValueInBB> &TraverseOrder,
    3074             :                           SimplificationTracker &ST) {
    3075             :     SmallVector<ValueInBB, 32> Worklist;
    3076             :     assert((isa<PHINode>(Original.first) || isa<SelectInst>(Original.first)) &&
    3077             :            "Address must be a Phi or Select node");
    3078        1006 :     auto *Dummy = UndefValue::get(CommonType);
    3079        1006 :     Worklist.push_back(Original);
    3080        5074 :     while (!Worklist.empty()) {
    3081        4068 :       auto Current = Worklist.pop_back_val();
    3082             :       // If value is not an instruction it is something global, constant,
    3083             :       // parameter and we can say that this value is observable in any block.
    3084             :       // Set block to null to denote it.
    3085             :       // Also please take into account that it is how we build anchors.
    3086        4068 :       if (!isa<Instruction>(Current.first))
    3087        1898 :         Current.second = nullptr;
    3088             :       // if it is already visited or it is an ending value then skip it.
    3089        4068 :       if (Map.find(Current) != Map.end())
    3090        5578 :         continue;
    3091        1522 :       TraverseOrder.push_back(Current);
    3092             : 
    3093        1522 :       Value *CurrentValue = Current.first;
    3094        1522 :       BasicBlock *CurrentBlock = Current.second;
    3095             :       // CurrentValue must be a Phi node or select. All others must be covered
    3096             :       // by anchors.
    3097             :       Instruction *CurrentI = cast<Instruction>(CurrentValue);
    3098        1522 :       bool IsDefinedInThisBB = CurrentI->getParent() == CurrentBlock;
    3099             : 
    3100        1522 :       unsigned PredCount = pred_size(CurrentBlock);
    3101             :       // if Current Value is not defined in this basic block we are interested
    3102             :       // in values in predecessors.
    3103        1522 :       if (!IsDefinedInThisBB) {
    3104             :         assert(PredCount && "Unreachable block?!");
    3105         972 :         PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
    3106             :                                        &CurrentBlock->front());
    3107         486 :         Map[Current] = PHI;
    3108             :         ST.insertNewPhi(PHI);
    3109             :         // Add all predecessors in work list.
    3110        1617 :         for (auto B : predecessors(CurrentBlock))
    3111         645 :           Worklist.push_back({ CurrentValue, B });
    3112         486 :         continue;
    3113             :       }
    3114             :       // Value is defined in this basic block.
    3115             :       if (SelectInst *OrigSelect = dyn_cast<SelectInst>(CurrentI)) {
    3116             :         // Is it OK to get metadata from OrigSelect?!
    3117             :         // Create a Select placeholder with dummy value.
    3118             :         SelectInst *Select =
    3119          24 :             SelectInst::Create(OrigSelect->getCondition(), Dummy, Dummy,
    3120          72 :                                OrigSelect->getName(), OrigSelect, OrigSelect);
    3121          24 :         Map[Current] = Select;
    3122             :         ST.insertNewSelect(Select);
    3123             :         // We are interested in True and False value in this basic block.
    3124          24 :         Worklist.push_back({ OrigSelect->getTrueValue(), CurrentBlock });
    3125          24 :         Worklist.push_back({ OrigSelect->getFalseValue(), CurrentBlock });
    3126             :       } else {
    3127             :         // It must be a Phi node then.
    3128             :         auto *CurrentPhi = cast<PHINode>(CurrentI);
    3129             :         // Create new Phi node for merge of bases.
    3130             :         assert(PredCount && "Unreachable block?!");
    3131        2024 :         PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
    3132             :                                        &CurrentBlock->front());
    3133        1012 :         Map[Current] = PHI;
    3134             :         ST.insertNewPhi(PHI);
    3135             : 
    3136             :         // Add all predecessors in work list.
    3137        4393 :         for (auto B : predecessors(CurrentBlock))
    3138        4738 :           Worklist.push_back({ CurrentPhi->getIncomingValueForBlock(B), B });
    3139             :       }
    3140             :     }
    3141        1006 :   }
    3142             : 
    3143        1010 :   bool addrModeCombiningAllowed() {
    3144        1010 :     if (DisableComplexAddrModes)
    3145             :       return false;
    3146        1010 :     switch (DifferentField) {
    3147             :     default:
    3148             :       return false;
    3149             :     case ExtAddrMode::BaseRegField:
    3150             :       return AddrSinkCombineBaseReg;
    3151             :     case ExtAddrMode::BaseGVField:
    3152             :       return AddrSinkCombineBaseGV;
    3153             :     case ExtAddrMode::BaseOffsField:
    3154             :       return AddrSinkCombineBaseOffs;
    3155             :     case ExtAddrMode::ScaledRegField:
    3156             :       return AddrSinkCombineScaledReg;
    3157             :     }
    3158             :   }
    3159             : };
    3160             : } // end anonymous namespace
    3161             : 
    3162             : /// Try adding ScaleReg*Scale to the current addressing mode.
    3163             : /// Return true and update AddrMode if this addr mode is legal for the target,
    3164             : /// false if not.
    3165      186983 : bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
    3166             :                                              unsigned Depth) {
    3167             :   // If Scale is 1, then this is the same as adding ScaleReg to the addressing
    3168             :   // mode.  Just process that directly.
    3169      186983 :   if (Scale == 1)
    3170       10471 :     return matchAddr(ScaleReg, Depth);
    3171             : 
    3172             :   // If the scale is 0, it takes nothing to add this.
    3173      176512 :   if (Scale == 0)
    3174             :     return true;
    3175             : 
    3176             :   // If we already have a scale of this value, we can add to it, otherwise, we
    3177             :   // need an available scale field.
    3178      176508 :   if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
    3179             :     return false;
    3180             : 
    3181      176146 :   ExtAddrMode TestAddrMode = AddrMode;
    3182             : 
    3183             :   // Add scale to turn X*4+X*3 -> X*7.  This could also do things like
    3184             :   // [A+B + A*7] -> [B+A*8].
    3185      176146 :   TestAddrMode.Scale += Scale;
    3186      176146 :   TestAddrMode.ScaledReg = ScaleReg;
    3187             : 
    3188             :   // If the new address isn't legal, bail out.
    3189      176146 :   if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
    3190             :     return false;
    3191             : 
    3192             :   // It was legal, so commit it.
    3193      107423 :   AddrMode = TestAddrMode;
    3194             : 
    3195             :   // Okay, we decided that we can add ScaleReg+Scale to AddrMode.  Check now
    3196             :   // to see if ScaleReg is actually X+C.  If so, we can turn this into adding
    3197             :   // X*Scale + C*Scale to addr mode.
    3198      107423 :   ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
    3199      214132 :   if (isa<Instruction>(ScaleReg) &&  // not a constant expr.
    3200      214132 :       match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
    3201         855 :     TestAddrMode.ScaledReg = AddLHS;
    3202        1710 :     TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
    3203             : 
    3204             :     // If this addressing mode is legal, commit it and remember that we folded
    3205             :     // this instruction.
    3206         855 :     if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
    3207         764 :       AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
    3208         764 :       AddrMode = TestAddrMode;
    3209         764 :       return true;
    3210             :     }
    3211             :   }
    3212             : 
    3213             :   // Otherwise, not (x+c)*scale, just return what we have.
    3214             :   return true;
    3215             : }
    3216             : 
    3217             : /// This is a little filter, which returns true if an addressing computation
    3218             : /// involving I might be folded into a load/store accessing it.
    3219             : /// This doesn't need to be perfect, but needs to accept at least
    3220             : /// the set of instructions that MatchOperationAddr can.
    3221       13936 : static bool MightBeFoldableInst(Instruction *I) {
    3222             :   switch (I->getOpcode()) {
    3223        4176 :   case Instruction::BitCast:
    3224             :   case Instruction::AddrSpaceCast:
    3225             :     // Don't touch identity bitcasts.
    3226        8352 :     if (I->getType() == I->getOperand(0)->getType())
    3227             :       return false;
    3228        4176 :     return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
    3229             :   case Instruction::PtrToInt:
    3230             :     // PtrToInt is always a noop, as we know that the int type is pointer sized.
    3231             :     return true;
    3232             :   case Instruction::IntToPtr:
    3233             :     // We know the input is intptr_t, so this is foldable.
    3234             :     return true;
    3235             :   case Instruction::Add:
    3236             :     return true;
    3237          30 :   case Instruction::Mul:
    3238             :   case Instruction::Shl:
    3239             :     // Can only handle X*C and X << C.
    3240          30 :     return isa<ConstantInt>(I->getOperand(1));
    3241             :   case Instruction::GetElementPtr:
    3242             :     return true;
    3243        1336 :   default:
    3244        1336 :     return false;
    3245             :   }
    3246             : }
    3247             : 
    3248             : /// Check whether or not \p Val is a legal instruction for \p TLI.
    3249             : /// \note \p Val is assumed to be the product of some type promotion.
    3250             : /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
    3251             : /// to be legal, as the non-promoted value would have had the same state.
    3252        1048 : static bool isPromotedInstructionLegal(const TargetLowering &TLI,
    3253             :                                        const DataLayout &DL, Value *Val) {
    3254             :   Instruction *PromotedInst = dyn_cast<Instruction>(Val);
    3255             :   if (!PromotedInst)
    3256             :     return false;
    3257        2094 :   int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
    3258             :   // If the ISDOpcode is undefined, it was undefined before the promotion.
    3259        1047 :   if (!ISDOpcode)
    3260             :     return true;
    3261             :   // Otherwise, check if the promoted instruction is legal or not.
    3262        1047 :   return TLI.isOperationLegalOrCustom(
    3263        1047 :       ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
    3264             : }
    3265             : 
    3266             : namespace {
    3267             : 
    3268             : /// Hepler class to perform type promotion.
    3269             : class TypePromotionHelper {
    3270             :   /// Utility function to check whether or not a sign or zero extension
    3271             :   /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
    3272             :   /// either using the operands of \p Inst or promoting \p Inst.
    3273             :   /// The type of the extension is defined by \p IsSExt.
    3274             :   /// In other words, check if:
    3275             :   /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
    3276             :   /// #1 Promotion applies:
    3277             :   /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
    3278             :   /// #2 Operand reuses:
    3279             :   /// ext opnd1 to ConsideredExtType.
    3280             :   /// \p PromotedInsts maps the instructions to their type before promotion.
    3281             :   static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
    3282             :                             const InstrToOrigTy &PromotedInsts, bool IsSExt);
    3283             : 
    3284             :   /// Utility function to determine if \p OpIdx should be promoted when
    3285             :   /// promoting \p Inst.
    3286             :   static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
    3287        2282 :     return !(isa<SelectInst>(Inst) && OpIdx == 0);
    3288             :   }
    3289             : 
    3290             :   /// Utility function to promote the operand of \p Ext when this
    3291             :   /// operand is a promotable trunc or sext or zext.
    3292             :   /// \p PromotedInsts maps the instructions to their type before promotion.
    3293             :   /// \p CreatedInstsCost[out] contains the cost of all instructions
    3294             :   /// created to promote the operand of Ext.
    3295             :   /// Newly added extensions are inserted in \p Exts.
    3296             :   /// Newly added truncates are inserted in \p Truncs.
    3297             :   /// Should never be called directly.
    3298             :   /// \return The promoted value which is used instead of Ext.
    3299             :   static Value *promoteOperandForTruncAndAnyExt(
    3300             :       Instruction *Ext, TypePromotionTransaction &TPT,
    3301             :       InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
    3302             :       SmallVectorImpl<Instruction *> *Exts,
    3303             :       SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
    3304             : 
    3305             :   /// Utility function to promote the operand of \p Ext when this
    3306             :   /// operand is promotable and is not a supported trunc or sext.
    3307             :   /// \p PromotedInsts maps the instructions to their type before promotion.
    3308             :   /// \p CreatedInstsCost[out] contains the cost of all the instructions
    3309             :   /// created to promote the operand of Ext.
    3310             :   /// Newly added extensions are inserted in \p Exts.
    3311             :   /// Newly added truncates are inserted in \p Truncs.
    3312             :   /// Should never be called directly.
    3313             :   /// \return The promoted value which is used instead of Ext.
    3314             :   static Value *promoteOperandForOther(Instruction *Ext,
    3315             :                                        TypePromotionTransaction &TPT,
    3316             :                                        InstrToOrigTy &PromotedInsts,
    3317             :                                        unsigned &CreatedInstsCost,
    3318             :                                        SmallVectorImpl<Instruction *> *Exts,
    3319             :                                        SmallVectorImpl<Instruction *> *Truncs,
    3320             :                                        const TargetLowering &TLI, bool IsSExt);
    3321             : 
    3322             :   /// \see promoteOperandForOther.
    3323         296 :   static Value *signExtendOperandForOther(
    3324             :       Instruction *Ext, TypePromotionTransaction &TPT,
    3325             :       InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
    3326             :       SmallVectorImpl<Instruction *> *Exts,
    3327             :       SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
    3328             :     return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
    3329         296 :                                   Exts, Truncs, TLI, true);
    3330             :   }
    3331             : 
    3332             :   /// \see promoteOperandForOther.
    3333         845 :   static Value *zeroExtendOperandForOther(
    3334             :       Instruction *Ext, TypePromotionTransaction &TPT,
    3335             :       InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
    3336             :       SmallVectorImpl<Instruction *> *Exts,
    3337             :       SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
    3338             :     return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
    3339         845 :                                   Exts, Truncs, TLI, false);
    3340             :   }
    3341             : 
    3342             : public:
    3343             :   /// Type for the utility function that promotes the operand of Ext.
    3344             :   using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
    3345             :                             InstrToOrigTy &PromotedInsts,
    3346             :                             unsigned &CreatedInstsCost,
    3347             :                             SmallVectorImpl<Instruction *> *Exts,
    3348             :                             SmallVectorImpl<Instruction *> *Truncs,
    3349             :                             const TargetLowering &TLI);
    3350             : 
    3351             :   /// Given a sign/zero extend instruction \p Ext, return the approriate
    3352             :   /// action to promote the operand of \p Ext instead of using Ext.
    3353             :   /// \return NULL if no promotable action is possible with the current
    3354             :   /// sign extension.
    3355             :   /// \p InsertedInsts keeps track of all the instructions inserted by the
    3356             :   /// other CodeGenPrepare optimizations. This information is important
    3357             :   /// because we do not want to promote these instructions as CodeGenPrepare
    3358             :   /// will reinsert them later. Thus creating an infinite loop: create/remove.
    3359             :   /// \p PromotedInsts maps the instructions to their type before promotion.
    3360             :   static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
    3361             :                           const TargetLowering &TLI,
    3362             :                           const InstrToOrigTy &PromotedInsts);
    3363             : };
    3364             : 
    3365             : } // end anonymous namespace
    3366             : 
    3367       11177 : bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
    3368             :                                         Type *ConsideredExtType,
    3369             :                                         const InstrToOrigTy &PromotedInsts,
    3370             :                                         bool IsSExt) {
    3371             :   // The promotion helper does not know how to deal with vector types yet.
    3372             :   // To be able to fix that, we would need to fix the places where we
    3373             :   // statically extend, e.g., constants and such.
    3374       22354 :   if (Inst->getType()->isVectorTy())
    3375             :     return false;
    3376             : 
    3377             :   // We can always get through zext.
    3378        7265 :   if (isa<ZExtInst>(Inst))
    3379             :     return true;
    3380             : 
    3381             :   // sext(sext) is ok too.
    3382        7121 :   if (IsSExt && isa<SExtInst>(Inst))
    3383             :     return true;
    3384             : 
    3385             :   // We can get through binary operator, if it is legal. In other words, the
    3386             :   // binary operator must have a nuw or nsw flag.
    3387             :   const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
    3388        1134 :   if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
    3389         739 :       ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
    3390         395 :        (IsSExt && BinOp->hasNoSignedWrap())))
    3391             :     return true;
    3392             : 
    3393             :   // ext(and(opnd, cst)) --> and(ext(opnd), ext(cst))
    3394        6565 :   if ((Inst->getOpcode() == Instruction::And ||
    3395             :        Inst->getOpcode() == Instruction::Or))
    3396             :     return true;
    3397             : 
    3398             :   // ext(xor(opnd, cst)) --> xor(ext(opnd), ext(cst))
    3399        6185 :   if (Inst->getOpcode() == Instruction::Xor) {
    3400         875 :     const ConstantInt *Cst = dyn_cast<ConstantInt>(Inst->getOperand(1));
    3401             :     // Make sure it is not a NOT.
    3402         821 :     if (Cst && !Cst->getValue().isAllOnesValue())
    3403             :       return true;
    3404             :   }
    3405             : 
    3406             :   // zext(shrl(opnd, cst)) --> shrl(zext(opnd), zext(cst))
    3407             :   // It may change a poisoned value into a regular value, like
    3408             :   //     zext i32 (shrl i8 %val, 12)  -->  shrl i32 (zext i8 %val), 12
    3409             :   //          poisoned value                    regular value
    3410             :   // It should be OK since undef covers valid value.
    3411        6176 :   if (Inst->getOpcode() == Instruction::LShr && !IsSExt)
    3412             :     return true;
    3413             : 
    3414             :   // and(ext(shl(opnd, cst)), cst) --> and(shl(ext(opnd), ext(cst)), cst)
    3415             :   // It may change a poisoned value into a regular value, like
    3416             :   //     zext i32 (shl i8 %val, 12)  -->  shl i32 (zext i8 %val), 12
    3417             :   //          poisoned value                    regular value
    3418             :   // It should be OK since undef covers valid value.
    3419        6034 :   if (Inst->getOpcode() == Instruction::Shl && Inst->hasOneUse()) {
    3420             :     const Instruction *ExtInst =
    3421             :         dyn_cast<const Instruction>(*Inst->user_begin());
    3422          46 :     if (ExtInst->hasOneUse()) {
    3423             :       const Instruction *AndInst =
    3424             :           dyn_cast<const Instruction>(*ExtInst->user_begin());
    3425          45 :       if (AndInst && AndInst->getOpcode() == Instruction::And) {
    3426           5 :         const ConstantInt *Cst = dyn_cast<ConstantInt>(AndInst->getOperand(1));
    3427           5 :         if (Cst &&
    3428           5 :             Cst->getValue().isIntN(Inst->getType()->getIntegerBitWidth()))
    3429             :           return true;
    3430             :       }
    3431             :     }
    3432             :   }
    3433             : 
    3434             :   // Check if we can do the following simplification.
    3435             :   // ext(trunc(opnd)) --> ext(opnd)
    3436        5973 :   if (!isa<TruncInst>(Inst))
    3437             :     return false;
    3438             : 
    3439         732 :   Value *OpndVal = Inst->getOperand(0);
    3440             :   // Check if we can use this operand in the extension.
    3441             :   // If the type is larger than the result type of the extension, we cannot.
    3442        2196 :   if (!OpndVal->getType()->isIntegerTy() ||
    3443             :       OpndVal->getType()->getIntegerBitWidth() >
    3444             :           ConsideredExtType->getIntegerBitWidth())
    3445             :     return false;
    3446             : 
    3447             :   // If the operand of the truncate is not an instruction, we will not have
    3448             :   // any information on the dropped bits.
    3449             :   // (Actually we could for constant but it is not worth the extra logic).
    3450             :   Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
    3451             :   if (!Opnd)
    3452             :     return false;
    3453             : 
    3454             :   // Check if the source of the type is narrow enough.
    3455             :   // I.e., check that trunc just drops extended bits of the same kind of
    3456             :   // the extension.
    3457             :   // #1 get the type of the operand and check the kind of the extended bits.
    3458             :   const Type *OpndType;
    3459         652 :   InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
    3460         657 :   if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
    3461             :     OpndType = It->second.getPointer();
    3462        1294 :   else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
    3463         182 :     OpndType = Opnd->getOperand(0)->getType();
    3464             :   else
    3465             :     return false;
    3466             : 
    3467             :   // #2 check that the truncate just drops extended bits.
    3468          96 :   return Inst->getType()->getIntegerBitWidth() >=
    3469          96 :          OpndType->getIntegerBitWidth();
    3470             : }
    3471             : 
    3472       14656 : TypePromotionHelper::Action TypePromotionHelper::getAction(
    3473             :     Instruction *Ext, const SetOfInstrs &InsertedInsts,
    3474             :     const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
    3475             :   assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
    3476             :          "Unexpected instruction type");
    3477       14656 :   Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
    3478       14656 :   Type *ExtTy = Ext->getType();
    3479             :   bool IsSExt = isa<SExtInst>(Ext);
    3480             :   // If the operand of the extension is not an instruction, we cannot
    3481             :   // get through.
    3482             :   // If it, check we can get through.
    3483       14656 :   if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
    3484             :     return nullptr;
    3485             : 
    3486             :   // Do not promote if the operand has been added by codegenprepare.
    3487             :   // Otherwise, it means we are undoing an optimization that is likely to be
    3488             :   // redone, thus causing potential infinite loop.
    3489        1386 :   if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
    3490             :     return nullptr;
    3491             : 
    3492             :   // SExt or Trunc instructions.
    3493             :   // Return the related handler.
    3494        1304 :   if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
    3495             :       isa<ZExtInst>(ExtOpnd))
    3496             :     return promoteOperandForTruncAndAnyExt;
    3497             : 
    3498             :   // Regular instruction.
    3499             :   // Abort early if we will have to insert non-free instructions.
    3500        1141 :   if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
    3501             :     return nullptr;
    3502        1141 :   return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
    3503             : }
    3504             : 
    3505         163 : Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
    3506             :     Instruction *SExt, TypePromotionTransaction &TPT,
    3507             :     InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
    3508             :     SmallVectorImpl<Instruction *> *Exts,
    3509             :     SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
    3510             :   // By construction, the operand of SExt is an instruction. Otherwise we cannot
    3511             :   // get through it and this method should not be called.
    3512         163 :   Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
    3513             :   Value *ExtVal = SExt;
    3514             :   bool HasMergedNonFreeExt = false;
    3515         163 :   if (isa<ZExtInst>(SExtOpnd)) {
    3516             :     // Replace s|zext(zext(opnd))
    3517             :     // => zext(opnd).
    3518         144 :     HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
    3519             :     Value *ZExt =
    3520         288 :         TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
    3521         144 :     TPT.replaceAllUsesWith(SExt, ZExt);
    3522         144 :     TPT.eraseInstruction(SExt);
    3523             :     ExtVal = ZExt;
    3524             :   } else {
    3525             :     // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
    3526             :     // => z|sext(opnd).
    3527          38 :     TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
    3528             :   }
    3529         163 :   CreatedInstsCost = 0;
    3530             : 
    3531             :   // Remove dead code.
    3532         163 :   if (SExtOpnd->use_empty())
    3533         135 :     TPT.eraseInstruction(SExtOpnd);
    3534             : 
    3535             :   // Check if the extension is still needed.
    3536         163 :   Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
    3537         324 :   if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
    3538         158 :     if (ExtInst) {
    3539         156 :       if (Exts)
    3540         144 :         Exts->push_back(ExtInst);
    3541         156 :       CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
    3542             :     }
    3543             :     return ExtVal;
    3544             :   }
    3545             : 
    3546             :   // At this point we have: ext ty opnd to ty.
    3547             :   // Reassign the uses of ExtInst to the opnd and remove ExtInst.
    3548             :   Value *NextVal = ExtInst->getOperand(0);
    3549           5 :   TPT.eraseInstruction(ExtInst, NextVal);
    3550           5 :   return NextVal;
    3551             : }
    3552             : 
    3553        1141 : Value *TypePromotionHelper::promoteOperandForOther(
    3554             :     Instruction *Ext, TypePromotionTransaction &TPT,
    3555             :     InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
    3556             :     SmallVectorImpl<Instruction *> *Exts,
    3557             :     SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
    3558             :     bool IsSExt) {
    3559             :   // By construction, the operand of Ext is an instruction. Otherwise we cannot
    3560             :   // get through it and this method should not be called.
    3561        1141 :   Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
    3562        1141 :   CreatedInstsCost = 0;
    3563        1141 :   if (!ExtOpnd->hasOneUse()) {
    3564             :     // ExtOpnd will be promoted.
    3565             :     // All its uses, but Ext, will need to use a truncated value of the
    3566             :     // promoted version.
    3567             :     // Create the truncate now.
    3568          99 :     Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
    3569          99 :     if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
    3570             :       // Insert it just after the definition.
    3571          99 :       ITrunc->moveAfter(ExtOpnd);
    3572          99 :       if (Truncs)
    3573           0 :         Truncs->push_back(ITrunc);
    3574             :     }
    3575             : 
    3576          99 :     TPT.replaceAllUsesWith(ExtOpnd, Trunc);
    3577             :     // Restore the operand of Ext (which has been replaced by the previous call
    3578             :     // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
    3579          99 :     TPT.setOperand(Ext, 0, ExtOpnd);
    3580             :   }
    3581             : 
    3582             :   // Get through the Instruction:
    3583             :   // 1. Update its type.
    3584             :   // 2. Replace the uses of Ext by Inst.
    3585             :   // 3. Extend each operand that needs to be extended.
    3586             : 
    3587             :   // Remember the original type of the instruction before promotion.
    3588             :   // This is useful to know that the high bits are sign extended bits.
    3589        1141 :   PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
    3590        1141 :       ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
    3591             :   // Step #1.
    3592        1141 :   TPT.mutateType(ExtOpnd, Ext->getType());
    3593             :   // Step #2.
    3594        1141 :   TPT.replaceAllUsesWith(Ext, ExtOpnd);
    3595             :   // Step #3.
    3596        1141 :   Instruction *ExtForOpnd = Ext;
    3597             : 
    3598             :   LLVM_DEBUG(dbgs() << "Propagate Ext to operands\n");
    3599        3423 :   for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
    3600             :        ++OpIdx) {
    3601             :     LLVM_DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
    3602        4564 :     if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
    3603             :         !shouldExtOperand(ExtOpnd, OpIdx)) {
    3604             :       LLVM_DEBUG(dbgs() << "No need to propagate\n");
    3605           0 :       continue;
    3606             :     }
    3607             :     // Check if we can statically extend the operand.
    3608             :     Value *Opnd = ExtOpnd->getOperand(OpIdx);
    3609             :     if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
    3610             :       LLVM_DEBUG(dbgs() << "Statically extend\n");
    3611             :       unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
    3612             :       APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
    3613        1734 :                             : Cst->getValue().zext(BitWidth);
    3614         867 :       TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
    3615             :       continue;
    3616             :     }
    3617             :     // UndefValue are typed, so we have to statically sign extend them.
    3618        1415 :     if (isa<UndefValue>(Opnd)) {
    3619             :       LLVM_DEBUG(dbgs() << "Statically extend\n");
    3620           0 :       TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
    3621           0 :       continue;
    3622             :     }
    3623             : 
    3624             :     // Otherwise we have to explicity sign extend the operand.
    3625             :     // Check if Ext was reused to extend an operand.
    3626        1415 :     if (!ExtForOpnd) {
    3627             :       // If yes, create a new one.
    3628             :       LLVM_DEBUG(dbgs() << "More operands to ext\n");
    3629         274 :       Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
    3630             :         : TPT.createZExt(Ext, Opnd, Ext->getType());
    3631         280 :       if (!isa<Instruction>(ValForExtOpnd)) {
    3632           6 :         TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
    3633             :         continue;
    3634             :       }
    3635         268 :       ExtForOpnd = cast<Instruction>(ValForExtOpnd);
    3636             :     }
    3637        1409 :     if (Exts)
    3638        1175 :       Exts->push_back(ExtForOpnd);
    3639        1409 :     TPT.setOperand(ExtForOpnd, 0, Opnd);
    3640             : 
    3641             :     // Move the sign extension before the insertion point.
    3642        1409 :     TPT.moveBefore(ExtForOpnd, ExtOpnd);
    3643        1409 :     TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
    3644        1409 :     CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
    3645             :     // If more sext are required, new instructions will have to be created.
    3646        1409 :     ExtForOpnd = nullptr;
    3647             :   }
    3648        1141 :   if (ExtForOpnd == Ext) {
    3649             :     LLVM_DEBUG(dbgs() << "Extension is useless now\n");
    3650           0 :     TPT.eraseInstruction(Ext);
    3651             :   }
    3652        1141 :   return ExtOpnd;
    3653             : }
    3654             : 
    3655             : /// Check whether or not promoting an instruction to a wider type is profitable.
    3656             : /// \p NewCost gives the cost of extension instructions created by the
    3657             : /// promotion.
    3658             : /// \p OldCost gives the cost of extension instructions before the promotion
    3659             : /// plus the number of instructions that have been
    3660             : /// matched in the addressing mode the promotion.
    3661             : /// \p PromotedOperand is the value that has been promoted.
    3662             : /// \return True if the promotion is profitable, false otherwise.
    3663             : bool AddressingModeMatcher::isPromotionProfitable(
    3664             :     unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
    3665             :   LLVM_DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost
    3666             :                     << '\n');
    3667             :   // The cost of the new extensions is greater than the cost of the
    3668             :   // old extension plus what we folded.
    3669             :   // This is not profitable.
    3670         162 :   if (NewCost > OldCost)
    3671             :     return false;
    3672         151 :   if (NewCost < OldCost)
    3673             :     return true;
    3674             :   // The promotion is neutral but it may help folding the sign extension in
    3675             :   // loads for instance.
    3676             :   // Check that we did not create an illegal instruction.
    3677          61 :   return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
    3678             : }
    3679             : 
    3680             : /// Given an instruction or constant expr, see if we can fold the operation
    3681             : /// into the addressing mode. If so, update the addressing mode and return
    3682             : /// true, otherwise return false without modifying AddrMode.
    3683             : /// If \p MovedAway is not NULL, it contains the information of whether or
    3684             : /// not AddrInst has to be folded into the addressing mode on success.
    3685             : /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
    3686             : /// because it has been moved away.
    3687             : /// Thus AddrInst must not be added in the matched instructions.
    3688             : /// This state can happen when AddrInst is a sext, since it may be moved away.
    3689             : /// Therefore, AddrInst may not be valid when MovedAway is true and it must
    3690             : /// not be referenced anymore.
    3691     1119836 : bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
    3692             :                                                unsigned Depth,
    3693             :                                                bool *MovedAway) {
    3694             :   // Avoid exponential behavior on extremely deep expression trees.
    3695     1119836 :   if (Depth >= 5) return false;
    3696             : 
    3697             :   // By default, all matched instructions stay in place.
    3698     1119784 :   if (MovedAway)
    3699      456505 :     *MovedAway = false;
    3700             : 
    3701     1119784 :   switch (Opcode) {
    3702         773 :   case Instruction::PtrToInt:
    3703             :     // PtrToInt is always a noop, as we know that the int type is pointer sized.
    3704         773 :     return matchAddr(AddrInst->getOperand(0), Depth);
    3705        3577 :   case Instruction::IntToPtr: {
    3706        3577 :     auto AS = AddrInst->getType()->getPointerAddressSpace();
    3707        7154 :     auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
    3708             :     // This inttoptr is a no-op if the integer type is pointer sized.
    3709       10731 :     if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
    3710        3522 :       return matchAddr(AddrInst->getOperand(0), Depth);
    3711             :     return false;
    3712             :   }
    3713      381098 :   case Instruction::BitCast:
    3714             :     // BitCast is always a noop, and we can handle it as long as it is
    3715             :     // int->int or pointer->pointer (we don't want int<->fp or something).
    3716      381231 :     if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
    3717      762192 :          AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
    3718             :         // Don't touch identity bitcasts.  These were probably put here by LSR,
    3719             :         // and we don't want to mess around with them.  Assume it knows what it
    3720             :         // is doing.
    3721      381094 :         AddrInst->getOperand(0)->getType() != AddrInst->getType())
    3722      379278 :       return matchAddr(AddrInst->getOperand(0), Depth);
    3723             :     return false;
    3724         295 :   case Instruction::AddrSpaceCast: {
    3725             :     unsigned SrcAS
    3726         295 :       = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
    3727         295 :     unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
    3728         295 :     if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
    3729          94 :       return matchAddr(AddrInst->getOperand(0), Depth);
    3730             :     return false;
    3731             :   }
    3732        3001 :   case Instruction::Add: {
    3733             :     // Check to see if we can merge in the RHS then the LHS.  If so, we win.
    3734        3001 :     ExtAddrMode BackupAddrMode = AddrMode;
    3735        3001 :     unsigned OldSize = AddrModeInsts.size();
    3736             :     // Start a transaction at this point.
    3737             :     // The LHS may match but not the RHS.
    3738             :     // Therefore, we need a higher level restoration point to undo partially
    3739             :     // matched operation.
    3740             :     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
    3741        3001 :         TPT.getRestorationPoint();
    3742             : 
    3743        8688 :     if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
    3744        2686 :         matchAddr(AddrInst->getOperand(0), Depth+1))
    3745             :       return true;
    3746             : 
    3747             :     // Restore the old addr mode info.
    3748         917 :     AddrMode = BackupAddrMode;
    3749         917 :     AddrModeInsts.resize(OldSize);
    3750         917 :     TPT.rollback(LastKnownGood);
    3751             : 
    3752             :     // Otherwise this was over-aggressive.  Try merging in the LHS then the RHS.
    3753        1348 :     if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
    3754         431 :         matchAddr(AddrInst->getOperand(1), Depth+1))
    3755             :       return true;
    3756             : 
    3757             :     // Otherwise we definitely can't merge the ADD in.
    3758         915 :     AddrMode = BackupAddrMode;
    3759         915 :     AddrModeInsts.resize(OldSize);
    3760         915 :     TPT.rollback(LastKnownGood);
    3761             :     break;
    3762             :   }
    3763             :   //case Instruction::Or:
    3764             :   // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
    3765             :   //break;
    3766        3035 :   case Instruction::Mul:
    3767             :   case Instruction::Shl: {
    3768             :     // Can only handle X*C and X << C.
    3769             :     ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
    3770        2797 :     if (!RHS || RHS->getBitWidth() > 64)
    3771             :       return false;
    3772             :     int64_t Scale = RHS->getSExtValue();
    3773        2796 :     if (Opcode == Instruction::Shl)
    3774        1660 :       Scale = 1LL << Scale;
    3775             : 
    3776        2796 :     return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
    3777             :   }
    3778      608154 :   case Instruction::GetElementPtr: {
    3779             :     // Scan the GEP.  We check it if it contains constant offsets and at most
    3780             :     // one variable offset.
    3781             :     int VariableOperand = -1;
    3782             :     unsigned VariableScale = 0;
    3783             : 
    3784             :     int64_t ConstantOffset = 0;
    3785      608154 :     gep_type_iterator GTI = gep_type_begin(AddrInst);
    3786     2999912 :     for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
    3787       90374 :       if (StructType *STy = GTI.getStructTypeOrNull()) {
    3788       90374 :         const StructLayout *SL = DL.getStructLayout(STy);
    3789             :         unsigned Idx =
    3790       90374 :           cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
    3791       90374 :         ConstantOffset += SL->getElementOffset(Idx);
    3792             :       } else {
    3793     1106440 :         uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
    3794             :         if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
    3795      974365 :           ConstantOffset += CI->getSExtValue() * TypeSize;
    3796      132075 :         } else if (TypeSize) {  // Scales of zero don't do anything.
    3797             :           // We only allow one variable index at the moment.
    3798      132069 :           if (VariableOperand != -1)
    3799             :             return false;
    3800             : 
    3801             :           // Remember the variable index.
    3802      131134 :           VariableOperand = i;
    3803      131134 :           VariableScale = TypeSize;
    3804             :         }
    3805             :       }
    3806             :     }
    3807             : 
    3808             :     // A common case is for the GEP to only do a constant offset.  In this case,
    3809             :     // just add it to the disp field and check validity.
    3810      607219 :     if (VariableOperand == -1) {
    3811      477020 :       AddrMode.BaseOffs += ConstantOffset;
    3812      876241 :       if (ConstantOffset == 0 ||
    3813      399221 :           TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
    3814             :         // Check to see if we can fold the base pointer in too.
    3815      947252 :         if (matchAddr(AddrInst->getOperand(0), Depth+1))
    3816             :           return true;
    3817        3280 :       } else if (EnableGEPOffsetSplit && isa<GetElementPtrInst>(AddrInst) &&
    3818        6730 :                  TLI.shouldConsiderGEPOffsetSplit() && Depth == 0 &&
    3819             :                  ConstantOffset > 0) {
    3820             :         // Record GEPs with non-zero offsets as candidates for splitting in the
    3821             :         // event that the offset cannot fit into the r+i addressing mode.
    3822             :         // Simple and common case that only one GEP is used in calculating the
    3823             :         // address for the memory access.
    3824             :         Value *Base = AddrInst->getOperand(0);
    3825             :         auto *BaseI = dyn_cast<Instruction>(Base);
    3826             :         auto *GEP = cast<GetElementPtrInst>(AddrInst);
    3827          59 :         if (isa<Argument>(Base) || isa<GlobalValue>(Base) ||
    3828          23 :             (BaseI && !isa<CastInst>(BaseI) &&
    3829             :              !isa<GetElementPtrInst>(BaseI))) {
    3830             :           // If the base is an instruction, make sure the GEP is not in the same
    3831             :           // basic block as the base. If the base is an argument or global
    3832             :           // value, make sure the GEP is not in the entry block.  Otherwise,
    3833             :           // instruction selection can undo the split.  Also make sure the
    3834             :           // parent block allows inserting non-PHI instructions before the
    3835             :           // terminator.
    3836             :           BasicBlock *Parent =
    3837          21 :               BaseI ? BaseI->getParent() : &GEP->getFunction()->getEntryBlock();
    3838          21 :           if (GEP->getParent() != Parent && !Parent->getTerminator()->isEHPad())
    3839          10 :             LargeOffsetGEP = std::make_pair(GEP, ConstantOffset);
    3840             :         }
    3841             :       }
    3842        3394 :       AddrMode.BaseOffs -= ConstantOffset;
    3843        3394 :       return false;
    3844             :     }
    3845             : 
    3846             :     // Save the valid addressing mode in case we can't match.
    3847      130199 :     ExtAddrMode BackupAddrMode = AddrMode;
    3848      130199 :     unsigned OldSize = AddrModeInsts.size();
    3849             : 
    3850             :     // See if the scale and offset amount is valid for this target.
    3851      130199 :     AddrMode.BaseOffs += ConstantOffset;
    3852             : 
    3853             :     // Match the base operand of the GEP.
    3854      260398 :     if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
    3855             :       // If it couldn't be matched, just stuff the value in a register.
    3856           5 :       if (AddrMode.HasBaseReg) {
    3857           0 :         AddrMode = BackupAddrMode;
    3858           0 :         AddrModeInsts.resize(OldSize);
    3859           0 :         return false;
    3860             :       }
    3861           5 :       AddrMode.HasBaseReg = true;
    3862           5 :       AddrMode.BaseReg = AddrInst->getOperand(0);
    3863             :     }
    3864             : 
    3865             :     // Match the remaining variable portion of the GEP.
    3866      260398 :     if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
    3867             :                           Depth)) {
    3868             :       // If it couldn't be matched, try stuffing the base into a register
    3869             :       // instead of matching it, and retrying the match of the scale.
    3870       53988 :       AddrMode = BackupAddrMode;
    3871       53988 :       AddrModeInsts.resize(OldSize);
    3872       53988 :       if (AddrMode.HasBaseReg)
    3873             :         return false;
    3874       53988 :       AddrMode.HasBaseReg = true;
    3875       53988 :       AddrMode.BaseReg = AddrInst->getOperand(0);
    3876       53988 :       AddrMode.BaseOffs += ConstantOffset;
    3877       53988 :       if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
    3878             :                             VariableScale, Depth)) {
    3879             :         // If even that didn't work, bail.
    3880       15511 :         AddrMode = BackupAddrMode;
    3881       15511 :         AddrModeInsts.resize(OldSize);
    3882       15511 :         return false;
    3883             :       }
    3884             :     }
    3885             : 
    3886             :     return true;
    3887             :   }
    3888         986 :   case Instruction::SExt:
    3889             :   case Instruction::ZExt: {
    3890             :     Instruction *Ext = dyn_cast<Instruction>(AddrInst);
    3891             :     if (!Ext)
    3892             :       return false;
    3893             : 
    3894             :     // Try to move this ext out of the way of the addressing mode.
    3895             :     // Ask for a method for doing so.
    3896             :     TypePromotionHelper::Action TPH =
    3897         985 :         TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
    3898         985 :     if (!TPH)
    3899             :       return false;
    3900             : 
    3901             :     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
    3902         230 :         TPT.getRestorationPoint();
    3903         230 :     unsigned CreatedInstsCost = 0;
    3904         230 :     unsigned ExtCost = !TLI.isExtFree(Ext);
    3905             :     Value *PromotedOperand =
    3906         230 :         TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
    3907             :     // SExt has been moved away.
    3908             :     // Thus either it will be rematched later in the recursive calls or it is
    3909             :     // gone. Anyway, we must not fold it into the addressing mode at this point.
    3910             :     // E.g.,
    3911             :     // op = add opnd, 1
    3912             :     // idx = ext op
    3913             :     // addr = gep base, idx
    3914             :     // is now:
    3915             :     // promotedOpnd = ext opnd            <- no match here
    3916             :     // op = promoted_add promotedOpnd, 1  <- match (later in recursive calls)
    3917             :     // addr = gep base, op                <- match
    3918         230 :     if (MovedAway)
    3919         230 :       *MovedAway = true;
    3920             : 
    3921             :     assert(PromotedOperand &&
    3922             :            "TypePromotionHelper should have filtered out those cases");
    3923             : 
    3924         230 :     ExtAddrMode BackupAddrMode = AddrMode;
    3925         460 :     unsigned OldSize = AddrModeInsts.size();
    3926             : 
    3927         291 :     if (!matchAddr(PromotedOperand, Depth) ||
    3928             :         // The total of the new cost is equal to the cost of the created
    3929             :         // instructions.
    3930             :         // The total of the old cost is equal to the cost of the extension plus
    3931             :         // what we have saved in the addressing mode.
    3932         162 :         !isPromotionProfitable(CreatedInstsCost,
    3933         324 :                                ExtCost + (AddrModeInsts.size() - OldSize),
    3934             :                                PromotedOperand)) {
    3935         109 :       AddrMode = BackupAddrMode;
    3936         109 :       AddrModeInsts.resize(OldSize);
    3937             :       LLVM_DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
    3938         109 :       TPT.rollback(LastKnownGood);
    3939         109 :       return false;
    3940             :     }
    3941             :     return true;
    3942             :   }
    3943             :   }
    3944             :   return false;
    3945             : }
    3946             : 
    3947             : /// If we can, try to add the value of 'Addr' into the current addressing mode.
    3948             : /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
    3949             : /// unmodified. This assumes that Addr is either a pointer type or intptr_t
    3950             : /// for the target.
    3951             : ///
    3952     2108268 : bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
    3953             :   // Start a transaction at this point that we will rollback if the matching
    3954             :   // fails.
    3955             :   TypePromotionTransaction::ConstRestorationPt LastKnownGood =
    3956     2108268 :       TPT.getRestorationPoint();
    3957             :   if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
    3958             :     // Fold in immediates if legal for the target.
    3959        2313 :     AddrMode.BaseOffs += CI->getSExtValue();
    3960        2313 :     if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
    3961             :       return true;
    3962         647 :     AddrMode.BaseOffs -= CI->getSExtValue();
    3963             :   } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
    3964             :     // If this is a global variable, try to fold it into the addressing mode.
    3965      848696 :     if (!AddrMode.BaseGV) {
    3966      848696 :       AddrMode.BaseGV = GV;
    3967      848696 :       if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
    3968             :         return true;
    3969      175412 :       AddrMode.BaseGV = nullptr;
    3970             :     }
    3971     1257259 :   } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
    3972      456557 :     ExtAddrMode BackupAddrMode = AddrMode;
    3973      456557 :     unsigned OldSize = AddrModeInsts.size();
    3974             : 
    3975             :     // Check to see if it is possible to fold this operation.
    3976      456557 :     bool MovedAway = false;
    3977      456557 :     if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
    3978             :       // This instruction may have been moved away. If so, there is nothing
    3979             :       // to check here.
    3980      312370 :       if (MovedAway)
    3981      310086 :         return true;
    3982             :       // Okay, it's possible to fold this.  Check to see if it is actually
    3983             :       // *profitable* to do so.  We use a simple cost model to avoid increasing
    3984             :       // register pressure too much.
    3985      769327 :       if (I->hasOneUse() ||
    3986      144829 :           isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
    3987      309965 :         AddrModeInsts.push_back(I);
    3988      309965 :         return true;
    3989             :       }
    3990             : 
    3991             :       // It isn't profitable to do this, roll back.
    3992             :       //cerr << "NOT FOLDING: " << *I;
    3993        2284 :       AddrMode = BackupAddrMode;
    3994        2284 :       AddrModeInsts.resize(OldSize);
    3995        2284 :       TPT.rollback(LastKnownGood);
    3996             :     }
    3997             :   } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
    3998      663279 :     if (matchOperationAddr(CE, CE->getOpcode(), Depth))
    3999             :       return true;
    4000         288 :     TPT.rollback(LastKnownGood);
    4001      137423 :   } else if (isa<ConstantPointerNull>(Addr)) {
    4002             :     // Null pointer gets folded without affecting the addressing mode.
    4003             :     return true;
    4004             :   }
    4005             : 
    4006             :   // Worse case, the target should support [reg] addressing modes. :)
    4007      458904 :   if (!AddrMode.HasBaseReg) {
    4008      446792 :     AddrMode.HasBaseReg = true;
    4009      446792 :     AddrMode.BaseReg = Addr;
    4010             :     // Still check for legality in case the target supports [imm] but not [i+r].
    4011      446792 :     if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
    4012             :       return true;
    4013          10 :     AddrMode.HasBaseReg = false;
    4014          10 :     AddrMode.BaseReg = nullptr;
    4015             :   }
    4016             : 
    4017             :   // If the base register is already taken, see if we can do [r+r].
    4018       12122 :   if (AddrMode.Scale == 0) {
    4019       11521 :     AddrMode.Scale = 1;
    4020       11521 :     AddrMode.ScaledReg = Addr;
    4021       11521 :     if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
    4022             :       return true;
    4023        3341 :     AddrMode.Scale = 0;
    4024        3341 :     AddrMode.ScaledReg = nullptr;
    4025             :   }
    4026             :   // Couldn't match.
    4027        3942 :   TPT.rollback(LastKnownGood);
    4028        3942 :   return false;
    4029             : }
    4030             : 
    4031             : /// Check to see if all uses of OpVal by the specified inline asm call are due
    4032             : /// to memory operands. If so, return true, otherwise return false.
    4033          14 : static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
    4034             :                                     const TargetLowering &TLI,
    4035             :                                     const TargetRegisterInfo &TRI) {
    4036          14 :   const Function *F = CI->getFunction();
    4037             :   TargetLowering::AsmOperandInfoVector TargetConstraints =
    4038             :       TLI.ParseConstraints(F->getParent()->getDataLayout(), &TRI,
    4039          42 :                             ImmutableCallSite(CI));
    4040             : 
    4041          82 :   for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
    4042          66 :     TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
    4043             : 
    4044             :     // Compute the constraint code and ConstraintType to use.
    4045          66 :     TLI.ComputeConstraintToUse(OpInfo, SDValue());
    4046             : 
    4047             :     // If this asm operand is our Value*, and if it isn't an indirect memory
    4048             :     // operand, we can't fold it!
    4049          94 :     if (OpInfo.CallOperandVal == OpVal &&
    4050          44 :         (OpInfo.ConstraintType != TargetLowering::C_Memory ||
    4051          16 :          !OpInfo.isIndirect))
    4052             :       return false;
    4053             :   }
    4054             : 
    4055             :   return true;
    4056             : }
    4057             : 
    4058             : // Max number of memory uses to look at before aborting the search to conserve
    4059             : // compile time.
    4060             : static constexpr int MaxMemoryUsesToScan = 20;
    4061             : 
    4062             : /// Recursively walk all the uses of I until we find a memory use.
    4063             : /// If we find an obviously non-foldable instruction, return true.
    4064             : /// Add the ultimately found memory instructions to MemoryUses.
    4065       13936 : static bool FindAllMemoryUses(
    4066             :     Instruction *I,
    4067             :     SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
    4068             :     SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
    4069             :     const TargetRegisterInfo &TRI, int SeenInsts = 0) {
    4070             :   // If we already considered this instruction, we're done.
    4071       13936 :   if (!ConsideredInsts.insert(I).second)
    4072             :     return false;
    4073             : 
    4074             :   // If this is an obviously unfoldable instruction, bail out.
    4075       13936 :   if (!MightBeFoldableInst(I))
    4076             :     return true;
    4077             : 
    4078       12600 :   const bool OptSize = I->getFunction()->optForSize();
    4079             : 
    4080             :   // Loop over all the uses, recursively processing them.
    4081       30029 :   for (Use &U : I->uses()) {
    4082             :     // Conservatively return true if we're seeing a large number or a deep chain
    4083             :     // of users. This avoids excessive compilation times in pathological cases.
    4084       20677 :     if (SeenInsts++ >= MaxMemoryUsesToScan)
    4085             :       return true;
    4086             : 
    4087       20597 :     Instruction *UserI = cast<Instruction>(U.getUser());
    4088        6472 :     if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
    4089       12944 :       MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
    4090             :       continue;
    4091             :     }
    4092             : 
    4093             :     if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
    4094        3341 :       unsigned opNo = U.getOperandNo();
    4095        3341 :       if (opNo != StoreInst::getPointerOperandIndex())
    4096             :         return true; // Storing addr, not into addr.
    4097        6634 :       MemoryUses.push_back(std::make_pair(SI, opNo));
    4098        3317 :       continue;
    4099             :     }
    4100             : 
    4101             :     if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
    4102          54 :       unsigned opNo = U.getOperandNo();
    4103          54 :       if (opNo != AtomicRMWInst::getPointerOperandIndex())
    4104             :         return true; // Storing addr, not into addr.
    4105         108 :       MemoryUses.push_back(std::make_pair(RMW, opNo));
    4106          54 :       continue;
    4107             :     }
    4108             : 
    4109             :     if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
    4110           2 :       unsigned opNo = U.getOperandNo();
    4111           2 :       if (opNo != AtomicCmpXchgInst::getPointerOperandIndex())
    4112             :         return true; // Storing addr, not into addr.
    4113           4 :       MemoryUses.push_back(std::make_pair(CmpX, opNo));
    4114           2 :       continue;
    4115             :     }
    4116             : 
    4117             :     if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
    4118             :       // If this is a cold call, we can sink the addressing calculation into
    4119             :       // the cold path.  See optimizeCallInst
    4120        1137 :       if (!OptSize && CI->hasFnAttr(Attribute::Cold))
    4121           7 :         continue;
    4122             : 
    4123             :       InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
    4124             :       if (!IA) return true;
    4125             : 
    4126             :       // If this is a memory operand, we're cool, otherwise bail out.
    4127          14 :       if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI))
    4128             :         return true;
    4129           2 :       continue;
    4130             :     }
    4131             : 
    4132       10162 :     if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI,
    4133             :                           SeenInsts))
    4134             :       return true;
    4135             :   }
    4136             : 
    4137             :   return false;
    4138             : }
    4139             : 
    4140             : /// Return true if Val is already known to be live at the use site that we're
    4141             : /// folding it into. If so, there is no cost to include it in the addressing
    4142             : /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
    4143             : /// instruction already.
    4144      279176 : bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
    4145             :                                                    Value *KnownLive2) {
    4146             :   // If Val is either of the known-live values, we know it is live!
    4147      279176 :   if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
    4148             :     return true;
    4149             : 
    4150             :   // All values other than instructions and arguments (e.g. constants) are live.
    4151      181736 :   if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
    4152             : 
    4153             :   // If Val is a constant sized alloca in the entry block, it is live, this is
    4154             :   // true because it is just a reference to the stack/frame pointer, which is
    4155             :   // live for the whole function.
    4156             :   if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
    4157       18390 :     if (AI->isStaticAlloca())
    4158             :       return true;
    4159             : 
    4160             :   // Check to see if this value is already used in the memory instruction's
    4161             :   // block.  If so, it's already live into the block at the very least, so we
    4162             :   // can reasonably fold it.
    4163      124801 :   return Val->isUsedInBasicBlock(MemoryInst->getParent());
    4164             : }
    4165             : 
    4166             : /// It is possible for the addressing mode of the machine to fold the specified
    4167             : /// instruction into a load or store that ultimately uses it.
    4168             : /// However, the specified instruction has multiple uses.
    4169             : /// Given this, it may actually increase register pressure to fold it
    4170             : /// into the load. For example, consider this code:
    4171             : ///
    4172             : ///     X = ...
    4173             : ///     Y = X+1
    4174             : ///     use(Y)   -> nonload/store
    4175             : ///     Z = Y+1
    4176             : ///     load Z
    4177             : ///
    4178             : /// In this case, Y has multiple uses, and can be folded into the load of Z
    4179             : /// (yielding load [X+2]).  However, doing this will cause both "X" and "X+1" to
    4180             : /// be live at the use(Y) line.  If we don't fold Y into load Z, we use one
    4181             : /// fewer register.  Since Y can't be folded into "use(Y)" we don't increase the
    4182             : /// number of computations either.
    4183             : ///
    4184             : /// Note that this (like most of CodeGenPrepare) is just a rough heuristic.  If
    4185             : /// X was live across 'load Z' for other reasons, we actually *would* want to
    4186             : /// fold the addressing mode in the Z case.  This would make Y die earlier.
    4187      144829 : bool AddressingModeMatcher::
    4188             : isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
    4189             :                                      ExtAddrMode &AMAfter) {
    4190      144829 :   if (IgnoreProfitability) return true;
    4191             : 
    4192             :   // AMBefore is the addressing mode before this instruction was folded into it,
    4193             :   // and AMAfter is the addressing mode after the instruction was folded.  Get
    4194             :   // the set of registers referenced by AMAfter and subtract out those
    4195             :   // referenced by AMBefore: this is the set of values which folding in this
    4196             :   // address extends the lifetime of.
    4197             :   //
    4198             :   // Note that there are only two potential values being referenced here,
    4199             :   // BaseReg and ScaleReg (global addresses are always available, as are any
    4200             :   // folded immediates).
    4201      139588 :   Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
    4202             : 
    4203             :   // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
    4204             :   // lifetime wasn't extended by adding this instruction.
    4205      139588 :   if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
    4206             :     BaseReg = nullptr;
    4207      139588 :   if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
    4208             :     ScaledReg = nullptr;
    4209             : 
    4210             :   // If folding this instruction (and it's subexprs) didn't extend any live
    4211             :   // ranges, we're ok with it.
    4212      139588 :   if (!BaseReg && !ScaledReg)
    4213             :     return true;
    4214             : 
    4215             :   // If all uses of this instruction can have the address mode sunk into them,
    4216             :   // we can remove the addressing mode and effectively trade one live register
    4217             :   // for another (at worst.)  In this context, folding an addressing mode into
    4218             :   // the use is just a particularly nice way of sinking it.
    4219             :   SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
    4220             :   SmallPtrSet<Instruction*, 16> ConsideredInsts;
    4221        3774 :   if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI))
    4222             :     return false;  // Has a non-memory, non-foldable use!
    4223             : 
    4224             :   // Now that we know that all uses of this instruction are part of a chain of
    4225             :   // computation involving only operations that could theoretically be folded
    4226             :   // into a memory use, loop over each of these memory operation uses and see
    4227             :   // if they could  *actually* fold the instruction.  The assumption is that
    4228             :   // addressing modes are cheap and that duplicating the computation involved
    4229             :   // many times is worthwhile, even on a fastpath. For sinking candidates
    4230             :   // (i.e. cold call sites), this serves as a way to prevent excessive code
    4231             :   // growth since most architectures have some reasonable small and fast way to
    4232             :   // compute an effective address.  (i.e LEA on x86)
    4233             :   SmallVector<Instruction*, 32> MatchedAddrModeInsts;
    4234        6545 :   for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
    4235       10110 :     Instruction *User = MemoryUses[i].first;
    4236        5055 :     unsigned OpNo = MemoryUses[i].second;
    4237             : 
    4238             :     // Get the access type of this use.  If the use isn't a pointer, we don't
    4239             :     // know what it accesses.
    4240        5055 :     Value *Address = User->getOperand(OpNo);
    4241        5055 :     PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
    4242             :     if (!AddrTy)
    4243         287 :       return false;
    4244        5055 :     Type *AddressAccessTy = AddrTy->getElementType();
    4245             :     unsigned AS = AddrTy->getAddressSpace();
    4246             : 
    4247             :     // Do a match against the root of this address, ignoring profitability. This
    4248             :     // will tell us if the addressing mode for the memory operation will
    4249             :     // *actually* cover the shared instruction.
    4250        5055 :     ExtAddrMode Result;
    4251             :     std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
    4252             :                                                                       0);
    4253             :     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
    4254        5055 :         TPT.getRestorationPoint();
    4255             :     AddressingModeMatcher Matcher(
    4256             :         MatchedAddrModeInsts, TLI, TRI, AddressAccessTy, AS, MemoryInst, Result,
    4257        5055 :         InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP);
    4258        5055 :     Matcher.IgnoreProfitability = true;
    4259        5055 :     bool Success = Matcher.matchAddr(Address, 0);
    4260             :     (void)Success; assert(Success && "Couldn't select *anything*?");
    4261             : 
    4262             :     // The match was to check the profitability, the changes made are not
    4263             :     // part of the original matcher. Therefore, they should be dropped
    4264             :     // otherwise the original matcher will not present the right state.
    4265        5055 :     TPT.rollback(LastKnownGood);
    4266             : 
    4267             :     // If the match didn't cover I, then it won't be shared by it.
    4268        5055 :     if (!is_contained(MatchedAddrModeInsts, I))
    4269             :       return false;
    4270             : 
    4271             :     MatchedAddrModeInsts.clear();
    4272             :   }
    4273             : 
    4274             :   return true;
    4275             : }
    4276             : 
    4277             : /// Return true if the specified values are defined in a
    4278             : /// different basic block than BB.
    4279             : static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
    4280             :   if (Instruction *I = dyn_cast<Instruction>(V))
    4281      282930 :     return I->getParent() != BB;
    4282             :   return false;
    4283             : }
    4284             : 
    4285             : /// Sink addressing mode computation immediate before MemoryInst if doing so
    4286             : /// can be done without increasing register pressure.  The need for the
    4287             : /// register pressure constraint means this can end up being an all or nothing
    4288             : /// decision for all uses of the same addressing computation.
    4289             : ///
    4290             : /// Load and Store Instructions often have addressing modes that can do
    4291             : /// significant amounts of computation. As such, instruction selection will try
    4292             : /// to get the load or store to do as much computation as possible for the
    4293             : /// program. The problem is that isel can only see within a single block. As
    4294             : /// such, we sink as much legal addressing mode work into the block as possible.
    4295             : ///
    4296             : /// This method is used to optimize both load/store and inline asms with memory
    4297             : /// operands.  It's also used to sink addressing computations feeding into cold
    4298             : /// call sites into their (cold) basic block.
    4299             : ///
    4300             : /// The motivation for handling sinking into cold blocks is that doing so can
    4301             : /// both enable other address mode sinking (by satisfying the register pressure
    4302             : /// constraint above), and reduce register pressure globally (by removing the
    4303             : /// addressing mode computation from the fast path entirely.).
    4304     1090279 : bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
    4305             :                                         Type *AccessTy, unsigned AddrSpace) {
    4306     1090279 :   Value *Repl = Addr;
    4307             : 
    4308             :   // Try to collapse single-value PHI nodes.  This is necessary to undo
    4309             :   // unprofitable PRE transformations.
    4310             :   SmallVector<Value*, 8> worklist;
    4311             :   SmallPtrSet<Value*, 16> Visited;
    4312     1090279 :   worklist.push_back(Addr);
    4313             : 
    4314             :   // Use a worklist to iteratively look through PHI and select nodes, and
    4315             :   // ensure that the addressing mode obtained from the non-PHI/select roots of
    4316             :   // the graph are compatible.
    4317             :   bool PhiOrSelectSeen = false;
    4318             :   SmallVector<Instruction*, 16> AddrModeInsts;
    4319     1090279 :   const SimplifyQuery SQ(*DL, TLInfo);
    4320     1090279 :   AddressingModeCombiner AddrModes(SQ, { Addr, MemoryInst->getParent() });
    4321     1090279 :   TypePromotionTransaction TPT(RemovedInsts);
    4322             :   TypePromotionTransaction::ConstRestorationPt LastKnownGood =
    4323             :       TPT.getRestorationPoint();
    4324     2195644 :   while (!worklist.empty()) {
    4325     1107152 :     Value *V = worklist.back();
    4326             :     worklist.pop_back();
    4327             : 
    4328             :     // We allow traversing cyclic Phi nodes.
    4329             :     // In case of success after this loop we ensure that traversing through
    4330             :     // Phi nodes ends up with all cases to compute address of the form
    4331             :     //    BaseGV + Base + Scale * Index + Offset
    4332             :     // where Scale and Offset are constans and BaseGV, Base and Index
    4333             :     // are exactly the same Values in all cases.
    4334             :     // It means that BaseGV, Scale and Offset dominate our memory instruction
    4335             :     // and have the same value as they had in address computation represented
    4336             :     // as Phi. So we can safely sink address computation to memory instruction.
    4337     1107152 :     if (!Visited.insert(V).second)
    4338       10072 :       continue;
    4339             : 
    4340             :     // For a PHI node, push all of its incoming values.
    4341             :     if (PHINode *P = dyn_cast<PHINode>(V)) {
    4342       41136 :       for (Value *IncValue : P->incoming_values())
    4343       16667 :         worklist.push_back(IncValue);
    4344             :       PhiOrSelectSeen = true;
    4345        7802 :       continue;
    4346             :     }
    4347             :     // Similar for select.
    4348         460 :     if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
    4349         460 :       worklist.push_back(SI->getFalseValue());
    4350         460 :       worklist.push_back(SI->getTrueValue());
    4351             :       PhiOrSelectSeen = true;
    4352         460 :       continue;
    4353             :     }
    4354             : 
    4355             :     // For non-PHIs, determine the addressing mode being computed.  Note that
    4356             :     // the result may differ depending on what other uses our candidate
    4357             :     // addressing instructions might have.
    4358             :     AddrModeInsts.clear();
    4359             :     std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
    4360             :                                                                       0);
    4361             :     ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
    4362     1097985 :         V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *TRI,
    4363     2195970 :         InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP);
    4364             : 
    4365             :     GetElementPtrInst *GEP = LargeOffsetGEP.first;
    4366     1097995 :     if (GEP && GEP->getParent() != MemoryInst->getParent() &&
    4367     1097995 :         !NewGEPBases.count(GEP)) {
    4368             :       // If splitting the underlying data structure can reduce the offset of a
    4369             :       // GEP, collect the GEP.  Skip the GEPs that are the new bases of
    4370             :       // previously split data structures.
    4371          20 :       LargeOffsetGEPMap[GEP->getPointerOperand()].push_back(LargeOffsetGEP);
    4372          20 :       if (LargeOffsetGEPID.find(GEP) == LargeOffsetGEPID.end())
    4373          10 :         LargeOffsetGEPID[GEP] = LargeOffsetGEPID.size();
    4374             :     }
    4375             : 
    4376     1097985 :     NewAddrMode.OriginalValue = V;
    4377     1097985 :     if (!AddrModes.addNewAddrMode(NewAddrMode))
    4378             :       break;
    4379             :   }
    4380             : 
    4381             :   // Try to combine the AddrModes we've collected. If we couldn't collect any,
    4382             :   // or we have multiple but either couldn't combine them or combining them
    4383             :   // wouldn't do anything useful, bail out now.
    4384     1090279 :   if (!AddrModes.combineAddrModes()) {
    4385        3459 :     TPT.rollback(LastKnownGood);
    4386        3459 :     return false;
    4387             :   }
    4388             :   TPT.commit();
    4389             : 
    4390             :   // Get the combined AddrMode (or the only AddrMode, if we only had one).
    4391     1086820 :   ExtAddrMode AddrMode = AddrModes.getAddrMode();
    4392             : 
    4393             :   // If all the instructions matched are already in this BB, don't do anything.
    4394             :   // If we saw a Phi node then it is not local definitely, and if we saw a select
    4395             :   // then we want to push the address calculation past it even if it's already
    4396             :   // in this BB.
    4397     2169552 :   if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) {
    4398      274663 :         return IsNonLocalValue(V, MemoryInst->getParent());
    4399             :                   })) {
    4400             :     LLVM_DEBUG(dbgs() << "CGP: Found      local addrmode: " << AddrMode
    4401             :                       << "\n");
    4402             :     return false;
    4403             :   }
    4404             : 
    4405             :   // Insert this computation right after this user.  Since our caller is
    4406             :   // scanning from the top of the BB to the bottom, reuse of the expr are
    4407             :   // guaranteed to happen later.
    4408       16833 :   IRBuilder<> Builder(MemoryInst);
    4409             : 
    4410             :   // Now that we determined the addressing expression we want to use and know
    4411             :   // that we have to sink it into this block.  Check to see if we have already
    4412             :   // done this for some other load/store instr in this block.  If so, reuse
    4413             :   // the computation.  Before attempting reuse, check if the address is valid
    4414             :   // as it may have been erased.
    4415             : 
    4416       16833 :   WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
    4417             : 
    4418       16833 :   Value * SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
    4419       16833 :   if (SunkAddr) {
    4420             :     LLVM_DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode
    4421             :                       << " for " << *MemoryInst << "\n");
    4422         923 :     if (SunkAddr->getType() != Addr->getType())
    4423           0 :       SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
    4424       15915 :   } else if (AddrSinkUsingGEPs ||
    4425           5 :              (!AddrSinkUsingGEPs.getNumOccurrences() && TM && TTI->useAA())) {
    4426             :     // By default, we use the GEP-based method when AA is used later. This
    4427             :     // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
    4428             :     LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
    4429             :                       << " for " << *MemoryInst << "\n");
    4430       15905 :     Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
    4431             :     Value *ResultPtr = nullptr, *ResultIndex = nullptr;
    4432             : 
    4433             :     // First, find the pointer.
    4434       47621 :     if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
    4435             :       ResultPtr = AddrMode.BaseReg;
    4436             :       AddrMode.BaseReg = nullptr;
    4437             :     }
    4438             : 
    4439       17212 :     if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
    4440             :       // We can't add more than one pointer together, nor can we scale a
    4441             :       // pointer (both of which seem meaningless).
    4442           0 :       if (ResultPtr || AddrMode.Scale != 1)
    4443             :         return false;
    4444             : 
    4445             :       ResultPtr = AddrMode.ScaledReg;
    4446             :       AddrMode.Scale = 0;
    4447             :     }
    4448             : 
    4449             :     // It is only safe to sign extend the BaseReg if we know that the math
    4450             :     // required to create it did not overflow before we extend it. Since
    4451             :     // the original IR value was tossed in favor of a constant back when
    4452             :     // the AddrMode was created we need to bail out gracefully if widths
    4453             :     // do not match instead of extending it.
    4454             :     //
    4455             :     // (See below for code to add the scale.)
    4456       15905 :     if (AddrMode.Scale) {
    4457        1307 :       Type *ScaledRegTy = AddrMode.ScaledReg->getType();
    4458        1307 :       if (cast<IntegerType>(IntPtrTy)->getBitWidth() >
    4459             :           cast<IntegerType>(ScaledRegTy)->getBitWidth())
    4460             :         return false;
    4461             :     }
    4462             : 
    4463       15895 :     if (AddrMode.BaseGV) {
    4464          39 :       if (ResultPtr)
    4465             :         return false;
    4466             : 
    4467             :       ResultPtr = AddrMode.BaseGV;
    4468             :     }
    4469             : 
    4470             :     // If the real base value actually came from an inttoptr, then the matcher
    4471             :     // will look through it and provide only the integer value. In that case,
    4472             :     // use it here.
    4473       15895 :     if (!DL->isNonIntegralPointerType(Addr->getType())) {
    4474       15892 :       if (!ResultPtr && AddrMode.BaseReg) {
    4475         244 :         ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(),
    4476             :                                            "sunkaddr");
    4477         122 :         AddrMode.BaseReg = nullptr;
    4478       15770 :       } else if (!ResultPtr && AddrMode.Scale == 1) {
    4479           0 :         ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(),
    4480             :                                            "sunkaddr");
    4481             :         AddrMode.Scale = 0;
    4482             :       }
    4483             :     }
    4484             : 
    4485       15895 :     if (!ResultPtr &&
    4486           9 :         !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
    4487           3 :       SunkAddr = Constant::getNullValue(Addr->getType());
    4488       15892 :     } else if (!ResultPtr) {
    4489             :       return false;
    4490             :     } else {
    4491             :       Type *I8PtrTy =
    4492       31768 :           Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
    4493       15884 :       Type *I8Ty = Builder.getInt8Ty();
    4494             : 
    4495             :       // Start with the base register. Do this first so that subsequent address
    4496             :       // matching finds it last, which will prevent it from trying to match it
    4497             :       // as the scaled value in case it happens to be a mul. That would be
    4498             :       // problematic if we've sunk a different mul for the scale, because then
    4499             :       // we'd end up sinking both muls.
    4500       15884 :       if (AddrMode.BaseReg) {
    4501             :         Value *V = AddrMode.BaseReg;
    4502           2 :         if (V->getType() != IntPtrTy)
    4503           2 :           V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
    4504             : 
    4505             :         ResultIndex = V;
    4506             :       }
    4507             : 
    4508             :       // Add the scale value.
    4509       15884 :       if (AddrMode.Scale) {
    4510             :         Value *V = AddrMode.ScaledReg;
    4511        1297 :         if (V->getType() == IntPtrTy) {
    4512             :           // done.
    4513             :         } else {
    4514             :           assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <
    4515             :                  cast<IntegerType>(V->getType())->getBitWidth() &&
    4516             :                  "We can't transform if ScaledReg is too narrow");
    4517           1 :           V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
    4518             :         }
    4519             : 
    4520        1297 :         if (AddrMode.Scale != 1)
    4521         747 :           V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
    4522             :                                 "sunkaddr");
    4523        1297 :         if (ResultIndex)
    4524           0 :           ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
    4525             :         else
    4526             :           ResultIndex = V;
    4527             :       }
    4528             : 
    4529             :       // Add in the Base Offset if present.
    4530       15884 :       if (AddrMode.BaseOffs) {
    4531       14995 :         Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
    4532       14995 :         if (ResultIndex) {
    4533             :           // We need to add this separately from the scale above to help with
    4534             :           // SDAG consecutive load/store merging.
    4535         851 :           if (ResultPtr->getType() != I8PtrTy)
    4536         545 :             ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
    4537         851 :           ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
    4538             :         }
    4539             : 
    4540             :         ResultIndex = V;
    4541             :       }
    4542             : 
    4543       15884 :       if (!ResultIndex) {
    4544             :         SunkAddr = ResultPtr;
    4545             :       } else {
    4546       15443 :         if (ResultPtr->getType() != I8PtrTy)
    4547       14408 :           ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
    4548       15443 :         SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
    4549             :       }
    4550             : 
    4551       15884 :       if (SunkAddr->getType() != Addr->getType())
    4552       15038 :         SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
    4553             :     }
    4554             :   } else {
    4555             :     // We'd require a ptrtoint/inttoptr down the line, which we can't do for
    4556             :     // non-integral pointers, so in that case bail out now.
    4557          10 :     Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
    4558           5 :     Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
    4559             :     PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy);
    4560             :     PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy);
    4561           6 :     if (DL->isNonIntegralPointerType(Addr->getType()) ||
    4562           1 :         (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) ||
    4563           0 :         (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) ||
    4564           0 :         (AddrMode.BaseGV &&
    4565           0 :          DL->isNonIntegralPointerType(AddrMode.BaseGV->getType())))
    4566             :       return false;
    4567             : 
    4568             :     LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
    4569             :                       << " for " << *MemoryInst << "\n");
    4570           0 :     Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
    4571             :     Value *Result = nullptr;
    4572             : 
    4573             :     // Start with the base register. Do this first so that subsequent address
    4574             :     // matching finds it last, which will prevent it from trying to match it
    4575             :     // as the scaled value in case it happens to be a mul. That would be
    4576             :     // problematic if we've sunk a different mul for the scale, because then
    4577             :     // we'd end up sinking both muls.
    4578           0 :     if (AddrMode.BaseReg) {
    4579             :       Value *V = AddrMode.BaseReg;
    4580           0 :       if (V->getType()->isPointerTy())
    4581           0 :         V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
    4582           0 :       if (V->getType() != IntPtrTy)
    4583           0 :         V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
    4584             :       Result = V;
    4585             :     }
    4586             : 
    4587             :     // Add the scale value.
    4588           0 :     if (AddrMode.Scale) {
    4589             :       Value *V = AddrMode.ScaledReg;
    4590           0 :       if (V->getType() == IntPtrTy) {
    4591             :         // done.
    4592           0 :       } else if (V->getType()->isPointerTy()) {
    4593           0 :         V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
    4594           0 :       } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
    4595             :                  cast<IntegerType>(V->getType())->getBitWidth()) {
    4596           0 :         V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
    4597             :       } else {
    4598             :         // It is only safe to sign extend the BaseReg if we know that the math
    4599             :         // required to create it did not overflow before we extend it. Since
    4600             :         // the original IR value was tossed in favor of a constant back when
    4601             :         // the AddrMode was created we need to bail out gracefully if widths
    4602             :         // do not match instead of extending it.
    4603             :         Instruction *I = dyn_cast_or_null<Instruction>(Result);
    4604           0 :         if (I && (Result != AddrMode.BaseReg))
    4605           0 :           I->eraseFromParent();
    4606             :         return false;
    4607             :       }
    4608           0 :       if (AddrMode.Scale != 1)
    4609           0 :         V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
    4610             :                               "sunkaddr");
    4611           0 :       if (Result)
    4612           0 :         Result = Builder.CreateAdd(Result, V, "sunkaddr");
    4613             :       else
    4614             :         Result = V;
    4615             :     }
    4616             : 
    4617             :     // Add in the BaseGV if present.
    4618           0 :     if (AddrMode.BaseGV) {
    4619           0 :       Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
    4620           0 :       if (Result)
    4621           0 :         Result = Builder.CreateAdd(Result, V, "sunkaddr");
    4622             :       else
    4623             :         Result = V;
    4624             :     }
    4625             : 
    4626             :     // Add in the Base Offset if present.
    4627           0 :     if (AddrMode.BaseOffs) {
    4628           0 :       Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
    4629           0 :       if (Result)
    4630           0 :         Result = Builder.CreateAdd(Result, V, "sunkaddr");
    4631             :       else
    4632             :         Result = V;
    4633             :     }
    4634             : 
    4635           0 :     if (!Result)
    4636           0 :       SunkAddr = Constant::getNullValue(Addr->getType());
    4637             :     else
    4638           0 :       SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
    4639             :   }
    4640             : 
    4641       16810 :   MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
    4642             :   // Store the newly computed address into the cache. In the case we reused a
    4643             :   // value, this should be idempotent.
    4644       33620 :   SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
    4645             : 
    4646             :   // If we have no uses, recursively delete the value and all dead instructions
    4647             :   // using it.
    4648       16810 :   if (Repl->use_empty()) {
    4649             :     // This can cause recursive deletion, which can invalidate our iterator.
    4650             :     // Use a WeakTrackingVH to hold onto it in case this happens.
    4651             :     Value *CurValue = &*CurInstIterator;
    4652             :     WeakTrackingVH IterHandle(CurValue);
    4653        6072 :     BasicBlock *BB = CurInstIterator->getParent();
    4654             : 
    4655        6072 :     RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
    4656             : 
    4657        6072 :     if (IterHandle != CurValue) {
    4658             :       // If the iterator instruction was recursively deleted, start over at the
    4659             :       // start of the block.
    4660           1 :       CurInstIterator = BB->begin();
    4661             :       SunkAddrs.clear();
    4662             :     }
    4663             :   }
    4664             :   ++NumMemoryInsts;
    4665             :   return true;
    4666             : }
    4667             : 
    4668             : /// If there are any memory operands, use OptimizeMemoryInst to sink their
    4669             : /// address computing into the block when possible / profitable.
    4670       13682 : bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
    4671             :   bool MadeChange = false;
    4672             : 
    4673             :   const TargetRegisterInfo *TRI =
    4674       27364 :       TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
    4675             :   TargetLowering::AsmOperandInfoVector TargetConstraints =
    4676       41046 :       TLI->ParseConstraints(*DL, TRI, CS);
    4677             :   unsigned ArgNo = 0;
    4678       92305 :   for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
    4679       64941 :     TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
    4680             : 
    4681             :     // Compute the constraint code and ConstraintType to use.
    4682       64941 :     TLI->ComputeConstraintToUse(OpInfo, SDValue());
    4683             : 
    4684       68936 :     if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
    4685        3995 :         OpInfo.isIndirect) {
    4686        3093 :       Value *OpVal = CS->getArgOperand(ArgNo++);
    4687        3093 :       MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
    4688       61848 :     } else if (OpInfo.Type == InlineAsm::isInput)
    4689        8060 :       ArgNo++;
    4690             :   }
    4691             : 
    4692       13682 :   return MadeChange;
    4693             : }
    4694             : 
    4695             : /// Check if all the uses of \p Val are equivalent (or free) zero or
    4696             : /// sign extensions.
    4697          32 : static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
    4698             :   assert(!Val->use_empty() && "Input must have at least one use");
    4699             :   const Instruction *FirstUser = cast<Instruction>(*Val->user_begin());
    4700             :   bool IsSExt = isa<SExtInst>(FirstUser);
    4701          32 :   Type *ExtTy = FirstUser->getType();
    4702          86 :   for (const User *U : Val->users()) {
    4703             :     const Instruction *UI = cast<Instruction>(U);
    4704         146 :     if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
    4705             :       return false;
    4706          55 :     Type *CurTy = UI->getType();
    4707             :     // Same input and output types: Same instruction after CSE.
    4708          55 :     if (CurTy == ExtTy)
    4709             :       continue;
    4710             : 
    4711             :     // If IsSExt is true, we are in this situation:
    4712             :     // a = Val
    4713             :     // b = sext ty1 a to ty2
    4714             :     // c = sext ty1 a to ty3
    4715             :     // Assuming ty2 is shorter than ty3, this could be turned into:
    4716             :     // a = Val
    4717             :     // b = sext ty1 a to ty2
    4718             :     // c = sext ty2 b to ty3
    4719             :     // However, the last sext is not free.
    4720          10 :     if (IsSExt)
    4721             :       return false;
    4722             : 
    4723             :     // This is a ZExt, maybe this is free to extend from one type to another.
    4724             :     // In that case, we would not account for a different use.
    4725             :     Type *NarrowTy;
    4726             :     Type *LargeTy;
    4727           9 :     if (ExtTy->getScalarType()->getIntegerBitWidth() >
    4728             :         CurTy->getScalarType()->getIntegerBitWidth()) {
    4729             :       NarrowTy = CurTy;
    4730             :       LargeTy = ExtTy;
    4731             :     } else {
    4732             :       NarrowTy = ExtTy;
    4733             :       LargeTy = CurTy;
    4734             :     }
    4735             : 
    4736           9 :     if (!TLI.isZExtFree(NarrowTy, LargeTy))
    4737             :       return false;
    4738             :   }
    4739             :   // All uses are the same or can be derived from one another for free.
    4740             :   return true;
    4741             : }
    4742             : 
    4743             : /// Try to speculatively promote extensions in \p Exts and continue
    4744             : /// promoting through newly promoted operands recursively as far as doing so is
    4745             : /// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
    4746             : /// When some promotion happened, \p TPT contains the proper state to revert
    4747             : /// them.
    4748             : ///
    4749             : /// \return true if some promotion happened, false otherwise.
    4750       33227 : bool CodeGenPrepare::tryToPromoteExts(
    4751             :     TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
    4752             :     SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
    4753             :     unsigned CreatedInstsCost) {
    4754             :   bool Promoted = false;
    4755             : 
    4756             :   // Iterate over all the extensions to try to promote them.
    4757       80607 :   for (auto I : Exts) {
    4758             :     // Early check if we directly have ext(load).
    4759       43462 :     if (isa<LoadInst>(I->getOperand(0))) {
    4760       10019 :       ProfitablyMovedExts.push_back(I);
    4761       32672 :       continue;
    4762             :     }
    4763             : 
    4764             :     // Check whether or not we want to do any promotion.  The reason we have
    4765             :     // this check inside the for loop is to catch the case where an extension
    4766             :     // is directly fed by a load because in such case the extension can be moved
    4767             :     // up without any promotion on its operands.
    4768       37133 :     if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
    4769        9753 :       return false;
    4770             : 
    4771             :     // Get the action to perform the promotion.
    4772             :     TypePromotionHelper::Action TPH =
    4773       13671 :         TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
    4774             :     // Check if we can promote.
    4775       26268 :     if (!TPH) {
    4776             :       // Save the current extension as we cannot move up through its operand.
    4777       12597 :       ProfitablyMovedExts.push_back(I);
    4778       12597 :       continue;
    4779             :     }
    4780             : 
    4781             :     // Save the current state.
    4782             :     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
    4783             :         TPT.getRestorationPoint();
    4784             :     SmallVector<Instruction *, 4> NewExts;
    4785        1074 :     unsigned NewCreatedInstsCost = 0;
    4786        1074 :     unsigned ExtCost = !TLI->isExtFree(I);
    4787             :     // Promote.
    4788        1074 :     Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
    4789        2148 :                              &NewExts, nullptr, *TLI);
    4790             :     assert(PromotedVal &&
    4791             :            "TypePromotionHelper should have filtered out those cases");
    4792             : 
    4793             :     // We would be able to merge only one extension in a load.
    4794             :     // Therefore, if we have more than 1 new extension we heuristically
    4795             :     // cut this search path, because it means we degrade the code quality.
    4796             :     // With exactly 2, the transformation is neutral, because we will merge
    4797             :     // one extension but leave one. However, we optimistically keep going,
    4798             :     // because the new extension may be removed too.
    4799        1074 :     long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
    4800             :     // FIXME: It would be possible to propagate a negative value instead of
    4801             :     // conservatively ceiling it to 0.
    4802        1074 :     TotalCreatedInstsCost =
    4803        2148 :         std::max((long long)0, (TotalCreatedInstsCost - ExtCost));
    4804        1105 :     if (!StressExtLdPromotion &&
    4805         987 :         (TotalCreatedInstsCost > 1 ||
    4806         987 :          !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
    4807             :       // This promotion is not profitable, rollback to the previous state, and
    4808             :       // save the current extension in ProfitablyMovedExts as the latest
    4809             :       // speculative promotion turned out to be unprofitable.
    4810          31 :       TPT.rollback(LastKnownGood);
    4811          31 :       ProfitablyMovedExts.push_back(I);
    4812          31 :       continue;
    4813             :     }
    4814             :     // Continue promoting NewExts as far as doing so is profitable.
    4815             :     SmallVector<Instruction *, 2> NewlyMovedExts;
    4816        1043 :     (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost);
    4817             :     bool NewPromoted = false;
    4818        3585 :     for (auto ExtInst : NewlyMovedExts) {
    4819        1271 :       Instruction *MovedExt = cast<Instruction>(ExtInst);
    4820        1271 :       Value *ExtOperand = MovedExt->getOperand(0);
    4821             :       // If we have reached to a load, we need this extra profitability check
    4822             :       // as it could potentially be merged into an ext(load).
    4823         369 :       if (isa<LoadInst>(ExtOperand) &&
    4824         367 :           !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
    4825          32 :             (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI))))
    4826          19 :         continue;
    4827             : 
    4828        1252 :       ProfitablyMovedExts.push_back(MovedExt);
    4829             :       NewPromoted = true;
    4830             :     }
    4831             : 
    4832             :     // If none of speculative promotions for NewExts is profitable, rollback
    4833             :     // and save the current extension (I) as the last profitable extension.
    4834        1049 :     if (!NewPromoted) {
    4835           6 :       TPT.rollback(LastKnownGood);
    4836           6 :       ProfitablyMovedExts.push_back(I);
    4837             :       continue;
    4838             :     }
    4839             :     // The promotion is profitable.
    4840             :     Promoted = true;
    4841             :   }
    4842       23474 :   return Promoted;
    4843             : }
    4844             : 
    4845             : /// Merging redundant sexts when one is dominating the other.
    4846          36 : bool CodeGenPrepare::mergeSExts(Function &F) {
    4847          36 :   DominatorTree DT(F);
    4848             :   bool Changed = false;
    4849         122 :   for (auto &Entry : ValToSExtendedUses) {
    4850             :     SExts &Insts = Entry.second;
    4851             :     SExts CurPts;
    4852         256 :     for (Instruction *Inst : Insts) {
    4853         309 :       if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) ||
    4854         206 :           Inst->getOperand(0) != Entry.first)
    4855           0 :         continue;
    4856             :       bool inserted = false;
    4857         103 :       for (auto &Pt : CurPts) {
    4858          53 :         if (DT.dominates(Inst, Pt)) {
    4859          30 :           Pt->replaceAllUsesWith(Inst);
    4860          30 :           RemovedInsts.insert(Pt);
    4861          30 :           Pt->removeFromParent();
    4862          30 :           Pt = Inst;
    4863             :           inserted = true;
    4864             :           Changed = true;
    4865          30 :           break;
    4866             :         }
    4867          23 :         if (!DT.dominates(Pt, Inst))
    4868             :           // Give up if we need to merge in a common dominator as the
    4869             :           // expermients show it is not profitable.
    4870             :           continue;
    4871          23 :         Inst->replaceAllUsesWith(Pt);
    4872          23 :         RemovedInsts.insert(Inst);
    4873          23 :         Inst->removeFromParent();
    4874             :         inserted = true;
    4875             :         Changed = true;
    4876          23 :         break;
    4877             :       }
    4878         103 :       if (!inserted)
    4879          50 :         CurPts.push_back(Inst);
    4880             :     }
    4881             :   }
    4882          36 :   return Changed;
    4883             : }
    4884             : 
    4885             : // Spliting large data structures so that the GEPs accessing them can have
    4886             : // smaller offsets so that they can be sunk to the same blocks as their users.
    4887             : // For example, a large struct starting from %base is splitted into two parts
    4888             : // where the second part starts from %new_base.
    4889             : //
    4890             : // Before:
    4891             : // BB0:
    4892             : //   %base     =
    4893             : //
    4894             : // BB1:
    4895             : //   %gep0     = gep %base, off0
    4896             : //   %gep1     = gep %base, off1
    4897             : //   %gep2     = gep %base, off2
    4898             : //
    4899             : // BB2:
    4900             : //   %load1    = load %gep0
    4901             : //   %load2    = load %gep1
    4902             : //   %load3    = load %gep2
    4903             : //
    4904             : // After:
    4905             : // BB0:
    4906             : //   %base     =
    4907             : //   %new_base = gep %base, off0
    4908             : //
    4909             : // BB1:
    4910             : //   %new_gep0 = %new_base
    4911             : //   %new_gep1 = gep %new_base, off1 - off0
    4912             : //   %new_gep2 = gep %new_base, off2 - off0
    4913             : //
    4914             : // BB2:
    4915             : //   %load1    = load i32, i32* %new_gep0
    4916             : //   %load2    = load i32, i32* %new_gep1
    4917             : //   %load3    = load i32, i32* %new_gep2
    4918             : //
    4919             : // %new_gep1 and %new_gep2 can be sunk to BB2 now after the splitting because
    4920             : // their offsets are smaller enough to fit into the addressing mode.
    4921           5 : bool CodeGenPrepare::splitLargeGEPOffsets() {
    4922             :   bool Changed = false;
    4923          15 :   for (auto &Entry : LargeOffsetGEPMap) {
    4924             :     Value *OldBase = Entry.first;
    4925             :     SmallVectorImpl<std::pair<AssertingVH<GetElementPtrInst>, int64_t>>
    4926             :         &LargeOffsetGEPs = Entry.second;
    4927             :     auto compareGEPOffset =
    4928             :         [&](const std::pair<GetElementPtrInst *, int64_t> &LHS,
    4929          10 :             const std::pair<GetElementPtrInst *, int64_t> &RHS) {
    4930          10 :           if (LHS.first == RHS.first)
    4931             :             return false;
    4932          10 :           if (LHS.second != RHS.second)
    4933          10 :             return LHS.second < RHS.second;
    4934           0 :           return LargeOffsetGEPID[LHS.first] < LargeOffsetGEPID[RHS.first];
    4935             :         };
    4936             :     // Sorting all the GEPs of the same data structures based on the offsets.
    4937             :     llvm::sort(LargeOffsetGEPs.begin(), LargeOffsetGEPs.end(),
    4938             :                compareGEPOffset);
    4939             :     LargeOffsetGEPs.erase(
    4940             :         std::unique(LargeOffsetGEPs.begin(), LargeOffsetGEPs.end()),
    4941             :         LargeOffsetGEPs.end());
    4942             :     // Skip if all the GEPs have the same offsets.
    4943           5 :     if (LargeOffsetGEPs.front().second == LargeOffsetGEPs.back().second)
    4944             :       continue;
    4945             :     GetElementPtrInst *BaseGEP = LargeOffsetGEPs.begin()->first;
    4946           5 :     int64_t BaseOffset = LargeOffsetGEPs.begin()->second;
    4947             :     Value *NewBaseGEP = nullptr;
    4948             : 
    4949             :     auto LargeOffsetGEP = LargeOffsetGEPs.begin();
    4950          25 :     while (LargeOffsetGEP != LargeOffsetGEPs.end()) {
    4951             :       GetElementPtrInst *GEP = LargeOffsetGEP->first;
    4952          10 :       int64_t Offset = LargeOffsetGEP->second;
    4953          10 :       if (Offset != BaseOffset) {
    4954           5 :         TargetLowering::AddrMode AddrMode;
    4955           5 :         AddrMode.BaseOffs = Offset - BaseOffset;
    4956             :         // The result type of the GEP might not be the type of the memory
    4957             :         // access.
    4958          10 :         if (!TLI->isLegalAddressingMode(*DL, AddrMode,
    4959             :                                         GEP->getResultElementType(),
    4960           5 :                                         GEP->getAddressSpace())) {
    4961             :           // We need to create a new base if the offset to the current base is
    4962             :           // too large to fit into the addressing mode. So, a very large struct
    4963             :           // may be splitted into several parts.
    4964             :           BaseGEP = GEP;
    4965             :           BaseOffset = Offset;
    4966             :           NewBaseGEP = nullptr;
    4967             :         }
    4968             :       }
    4969             : 
    4970             :       // Generate a new GEP to replace the current one.
    4971          10 :       IRBuilder<> Builder(GEP);
    4972          10 :       Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
    4973             :       Type *I8PtrTy =
    4974          20 :           Builder.getInt8PtrTy(GEP->getType()->getPointerAddressSpace());
    4975          10 :       Type *I8Ty = Builder.getInt8Ty();
    4976             : 
    4977          10 :       if (!NewBaseGEP) {
    4978             :         // Create a new base if we don't have one yet.  Find the insertion
    4979             :         // pointer for the new base first.
    4980             :         BasicBlock::iterator NewBaseInsertPt;
    4981             :         BasicBlock *NewBaseInsertBB;
    4982             :         if (auto *BaseI = dyn_cast<Instruction>(OldBase)) {
    4983             :           // If the base of the struct is an instruction, the new base will be
    4984             :           // inserted close to it.
    4985           4 :           NewBaseInsertBB = BaseI->getParent();
    4986           4 :           if (isa<PHINode>(BaseI))
    4987           0 :             NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
    4988             :           else if (InvokeInst *Invoke = dyn_cast<InvokeInst>(BaseI)) {
    4989           0 :             NewBaseInsertBB =
    4990             :                 SplitEdge(NewBaseInsertBB, Invoke->getNormalDest());
    4991           0 :             NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
    4992             :           } else
    4993           8 :             NewBaseInsertPt = std::next(BaseI->getIterator());
    4994             :         } else {
    4995             :           // If the current base is an argument or global value, the new base
    4996             :           // will be inserted to the entry block.
    4997           1 :           NewBaseInsertBB = &BaseGEP->getFunction()->getEntryBlock();
    4998           1 :           NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
    4999             :         }
    5000           5 :         IRBuilder<> NewBaseBuilder(NewBaseInsertBB, NewBaseInsertPt);
    5001             :         // Create a new base.
    5002           5 :         Value *BaseIndex = ConstantInt::get(IntPtrTy, BaseOffset);
    5003             :         NewBaseGEP = OldBase;
    5004           5 :         if (NewBaseGEP->getType() != I8PtrTy)
    5005           5 :           NewBaseGEP = NewBaseBuilder.CreatePointerCast(NewBaseGEP, I8PtrTy);
    5006           5 :         NewBaseGEP =
    5007           5 :             NewBaseBuilder.CreateGEP(I8Ty, NewBaseGEP, BaseIndex, "splitgep");
    5008          10 :         NewGEPBases.insert(NewBaseGEP);
    5009             :       }
    5010             : 
    5011             :       Value *NewGEP = NewBaseGEP;
    5012          10 :       if (Offset == BaseOffset) {
    5013           5 :         if (GEP->getType() != I8PtrTy)
    5014           5 :           NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
    5015             :       } else {
    5016             :         // Calculate the new offset for the new GEP.
    5017           5 :         Value *Index = ConstantInt::get(IntPtrTy, Offset - BaseOffset);
    5018           5 :         NewGEP = Builder.CreateGEP(I8Ty, NewBaseGEP, Index);
    5019             : 
    5020           5 :         if (GEP->getType() != I8PtrTy)
    5021           5 :           NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
    5022             :       }
    5023          10 :       GEP->replaceAllUsesWith(NewGEP);
    5024          20 :       LargeOffsetGEPID.erase(GEP);
    5025             :       LargeOffsetGEP = LargeOffsetGEPs.erase(LargeOffsetGEP);
    5026          10 :       GEP->eraseFromParent();
    5027             :       Changed = true;
    5028             :     }
    5029             :   }
    5030           5 :   return Changed;
    5031             : }
    5032             : 
    5033             : /// Return true, if an ext(load) can be formed from an extension in
    5034             : /// \p MovedExts.
    5035       32154 : bool CodeGenPrepare::canFormExtLd(
    5036             :     const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
    5037             :     Instruction *&Inst, bool HasPromoted) {
    5038       57276 :   for (auto *MovedExtInst : MovedExts) {
    5039       22534 :     if (isa<LoadInst>(MovedExtInst->getOperand(0))) {
    5040        9973 :       LI = cast<LoadInst>(MovedExtInst->getOperand(0));
    5041        9973 :       Inst = MovedExtInst;
    5042        9973 :       break;
    5043             :     }
    5044             :   }
    5045       32154 :   if (!LI)
    5046             :     return false;
    5047             : 
    5048             :   // If they're already in the same block, there's nothing to do.
    5049             :   // Make the cheap checks first if we did not promote.
    5050             :   // If we promoted, we need to check if it is indeed profitable.
    5051        9973 :   if (!HasPromoted && LI->getParent() == Inst->getParent())
    5052             :     return false;
    5053             : 
    5054         705 :   return TLI->isExtLoad(LI, Inst, *DL);
    5055             : }
    5056             : 
    5057             : /// Move a zext or sext fed by a load into the same basic block as the load,
    5058             : /// unless conditions are unfavorable. This allows SelectionDAG to fold the
    5059             : /// extend into the load.
    5060             : ///
    5061             : /// E.g.,
    5062             : /// \code
    5063             : /// %ld = load i32* %addr
    5064             : /// %add = add nuw i32 %ld, 4
    5065             : /// %zext = zext i32 %add to i64
    5066             : // \endcode
    5067             : /// =>
    5068             : /// \code
    5069             : /// %ld = load i32* %addr
    5070             : /// %zext = zext i32 %ld to i64
    5071             : /// %add = add nuw i64 %zext, 4
    5072             : /// \encode
    5073             : /// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
    5074             : /// allow us to match zext(load i32*) to i64.
    5075             : ///
    5076             : /// Also, try to promote the computations used to obtain a sign extended
    5077             : /// value used into memory accesses.
    5078             : /// E.g.,
    5079             : /// \code
    5080             : /// a = add nsw i32 b, 3
    5081             : /// d = sext i32 a to i64
    5082             : /// e = getelementptr ..., i64 d
    5083             : /// \endcode
    5084             : /// =>
    5085             : /// \code
    5086             : /// f = sext i32 b to i64
    5087             : /// a = add nsw i64 f, 3
    5088             : /// e = getelementptr ..., i64 a
    5089             : /// \endcode
    5090             : ///
    5091             : /// \p Inst[in/out] the extension may be modified during the process if some
    5092             : /// promotions apply.
    5093       32156 : bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
    5094             :   // ExtLoad formation and address type promotion infrastructure requires TLI to
    5095             :   // be effective.
    5096       32156 :   if (!TLI)
    5097             :     return false;
    5098             : 
    5099       32154 :   bool AllowPromotionWithoutCommonHeader = false;
    5100             :   /// See if it is an interesting sext operations for the address type
    5101             :   /// promotion before trying to promote it, e.g., the ones with the right
    5102             :   /// type and used in memory accesses.
    5103       32154 :   bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
    5104       32154 :       *Inst, AllowPromotionWithoutCommonHeader);
    5105       32154 :   TypePromotionTransaction TPT(RemovedInsts);
    5106             :   TypePromotionTransaction::ConstRestorationPt LastKnownGood =
    5107             :       TPT.getRestorationPoint();
    5108             :   SmallVector<Instruction *, 1> Exts;
    5109             :   SmallVector<Instruction *, 2> SpeculativelyMovedExts;
    5110       32154 :   Exts.push_back(Inst);
    5111             : 
    5112       32154 :   bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts);
    5113             : 
    5114             :   // Look for a load being extended.
    5115       32154 :   LoadInst *LI = nullptr;
    5116             :   Instruction *ExtFedByLoad;
    5117             : 
    5118             :   // Try to promote a chain of computation if it allows to form an extended
    5119             :   // load.
    5120       32154 :   if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) {
    5121             :     assert(LI && ExtFedByLoad && "Expect a valid load and extension");
    5122             :     TPT.commit();
    5123             :     // Move the extend into the same block as the load
    5124         700 :     ExtFedByLoad->moveAfter(LI);
    5125             :     // CGP does not check if the zext would be speculatively executed when moved
    5126             :     // to the same basic block as the load. Preserving its original location
    5127             :     // would pessimize the debugging experience, as well as negatively impact
    5128             :     // the quality of sample pgo. We don't want to use "line 0" as that has a
    5129             :     // size cost in the line-table section and logically the zext can be seen as
    5130             :     // part of the load. Therefore we conservatively reuse the same debug
    5131             :     // location for the load and the zext.
    5132        1400 :     ExtFedByLoad->setDebugLoc(LI->getDebugLoc());
    5133             :     ++NumExtsMoved;
    5134         700 :     Inst = ExtFedByLoad;
    5135         700 :     return true;
    5136             :   }
    5137             : 
    5138             :   // Continue promoting SExts if known as considerable depending on targets.
    5139       31600 :   if (ATPConsiderable &&
    5140         146 :       performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
    5141             :                                   HasPromoted, TPT, SpeculativelyMovedExts))
    5142             :     return true;
    5143             : 
    5144       31402 :   TPT.rollback(LastKnownGood);
    5145       31402 :   return false;
    5146             : }
    5147             : 
    5148             : // Perform address type promotion if doing so is profitable.
    5149             : // If AllowPromotionWithoutCommonHeader == false, we should find other sext
    5150             : // instructions that sign extended the same initial value. However, if
    5151             : // AllowPromotionWithoutCommonHeader == true, we expect promoting the
    5152             : // extension is just profitable.
    5153         146 : bool CodeGenPrepare::performAddressTypePromotion(
    5154             :     Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
    5155             :     bool HasPromoted, TypePromotionTransaction &TPT,
    5156             :     SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
    5157             :   bool Promoted = false;
    5158             :   SmallPtrSet<Instruction *, 1> UnhandledExts;
    5159             :   bool AllSeenFirst = true;
    5160         446 :   for (auto I : SpeculativelyMovedExts) {
    5161         150 :     Value *HeadOfChain = I->getOperand(0);
    5162             :     DenseMap<Value *, Instruction *>::iterator AlreadySeen =
    5163         150 :         SeenChainsForSExt.find(HeadOfChain);
    5164             :     // If there is an unhandled SExt which has the same header, try to promote
    5165             :     // it as well.
    5166         150 :     if (AlreadySeen != SeenChainsForSExt.end()) {
    5167          53 :       if (AlreadySeen->second != nullptr)
    5168          30 :         UnhandledExts.insert(AlreadySeen->second);
    5169             :       AllSeenFirst = false;
    5170             :     }
    5171             :   }
    5172             : 
    5173         163 :   if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
    5174             :                         SpeculativelyMovedExts.size() == 1)) {
    5175             :     TPT.commit();
    5176          70 :     if (HasPromoted)
    5177             :       Promoted = true;
    5178         212 :     for (auto I : SpeculativelyMovedExts) {
    5179         142 :       Value *HeadOfChain = I->getOperand(0);
    5180         142 :       SeenChainsForSExt[HeadOfChain] = nullptr;
    5181         142 :       ValToSExtendedUses[HeadOfChain].push_back(I);
    5182             :     }
    5183             :     // Update Inst as promotion happen.
    5184          70 :     Inst = SpeculativelyMovedExts.pop_back_val();
    5185             :   } else {
    5186             :     // This is the first chain visited from the header, keep the current chain
    5187             :     // as unhandled. Defer to promote this until we encounter another SExt
    5188             :     // chain derived from the same header.
    5189         234 :     for (auto I : SpeculativelyMovedExts) {
    5190         158 :       Value *HeadOfChain = I->getOperand(0);
    5191         158 :       SeenChainsForSExt[HeadOfChain] = Inst;
    5192             :     }
    5193             :     return false;
    5194             :   }
    5195             : 
    5196         123 :   if (!AllSeenFirst && !UnhandledExts.empty())
    5197          60 :     for (auto VisitedSExt : UnhandledExts) {
    5198          30 :       if (RemovedInsts.count(VisitedSExt))
    5199           0 :         continue;
    5200          30 :       TypePromotionTransaction TPT(RemovedInsts);
    5201             :       SmallVector<Instruction *, 1> Exts;
    5202             :       SmallVector<Instruction *, 2> Chains;
    5203          30 :       Exts.push_back(VisitedSExt);
    5204          30 :       bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains);
    5205             :       TPT.commit();
    5206          30 :       if (HasPromoted)
    5207             :         Promoted = true;
    5208          94 :       for (auto I : Chains) {
    5209          64 :         Value *HeadOfChain = I->getOperand(0);
    5210             :         // Mark this as handled.
    5211          64 :         SeenChainsForSExt[HeadOfChain] = nullptr;
    5212          64 :         ValToSExtendedUses[HeadOfChain].push_back(I);
    5213             :       }
    5214             :     }
    5215             :   return Promoted;
    5216             : }
    5217             : 
    5218       32156 : bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
    5219       32156 :   BasicBlock *DefBB = I->getParent();
    5220             : 
    5221             :   // If the result of a {s|z}ext and its source are both live out, rewrite all
    5222             :   // other uses of the source with result of extension.
    5223       32156 :   Value *Src = I->getOperand(0);
    5224       32156 :   if (Src->hasOneUse())
    5225             :     return false;
    5226             : 
    5227             :   // Only do this xform if truncating is free.
    5228        3756 :   if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
    5229             :     return false;
    5230             : 
    5231             :   // Only safe to perform the optimization if the source is also defined in
    5232             :   // this block.
    5233        3441 :   if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
    5234             :     return false;
    5235             : 
    5236             :   bool DefIsLiveOut = false;
    5237        4332 :   for (User *U : I->users()) {
    5238             :     Instruction *UI = cast<Instruction>(U);
    5239             : 
    5240             :     // Figure out which BB this ext is used in.
    5241        2794 :     BasicBlock *UserBB = UI->getParent();
    5242        2794 :     if (UserBB == DefBB) continue;
    5243             :     DefIsLiveOut = true;
    5244             :     break;
    5245             :   }
    5246        2704 :   if (!DefIsLiveOut)
    5247             :     return false;
    5248             : 
    5249             :   // Make sure none of the uses are PHI nodes.
    5250        3985 :   for (User *U : Src->users()) {
    5251             :     Instruction *UI = cast<Instruction>(U);
    5252        2840 :     BasicBlock *UserBB = UI->getParent();
    5253        2840 :     if (UserBB == DefBB) continue;
    5254             :     // Be conservative. We don't want this xform to end up introducing
    5255             :     // reloads just before load / store instructions.
    5256         548 :     if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
    5257             :       return false;
    5258             :   }
    5259             : 
    5260             :   // InsertedTruncs - Only insert one trunc in each block once.
    5261             :   DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
    5262             : 
    5263             :   bool MadeChange = false;
    5264        3015 :   for (Use &U : Src->uses()) {
    5265        1870 :     Instruction *User = cast<Instruction>(U.getUser());
    5266             : 
    5267             :     // Figure out which BB this ext is used in.
    5268        1870 :     BasicBlock *UserBB = User->getParent();
    5269        1870 :     if (UserBB == DefBB) continue;
    5270             : 
    5271             :     // Both src and def are live in this block. Rewrite the use.
    5272             :     Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
    5273             : 
    5274         486 :     if (!InsertedTrunc) {
    5275         486 :       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
    5276             :       assert(InsertPt != UserBB->end());
    5277         972 :       InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
    5278         486 :       InsertedInsts.insert(InsertedTrunc);
    5279             :     }
    5280             : 
    5281             :     // Replace a use of the {s|z}ext source with a use of the result.
    5282         486 :     U = InsertedTrunc;
    5283             :     ++NumExtUses;
    5284             :     MadeChange = true;
    5285             :   }
    5286             : 
    5287             :   return MadeChange;
    5288             : }
    5289             : 
    5290             : // Find loads whose uses only use some of the loaded value's bits.  Add an "and"
    5291             : // just after the load if the target can fold this into one extload instruction,
    5292             : // with the hope of eliminating some of the other later "and" instructions using
    5293             : // the loaded value.  "and"s that are made trivially redundant by the insertion
    5294             : // of the new "and" are removed by this function, while others (e.g. those whose
    5295             : // path from the load goes through a phi) are left for isel to potentially
    5296             : // remove.
    5297             : //
    5298             : // For example:
    5299             : //
    5300             : // b0:
    5301             : //   x = load i32
    5302             : //   ...
    5303             : // b1:
    5304             : //   y = and x, 0xff
    5305             : //   z = use y
    5306             : //
    5307             : // becomes:
    5308             : //
    5309             : // b0:
    5310             : //   x = load i32
    5311             : //   x' = and x, 0xff
    5312             : //   ...
    5313             : // b1:
    5314             : //   z = use x'
    5315             : //
    5316             : // whereas:
    5317             : //
    5318             : // b0:
    5319             : //   x1 = load i32
    5320             : //   ...
    5321             : // b1:
    5322             : //   x2 = load i32
    5323             : //   ...
    5324             : // b2:
    5325             : //   x = phi x1, x2
    5326             : //   y = and x, 0xff
    5327             : //
    5328             : // becomes (after a call to optimizeLoadExt for each load):
    5329             : //
    5330             : // b0:
    5331             : //   x1 = load i32
    5332             : //   x1' = and x1, 0xff
    5333             : //   ...
    5334             : // b1:
    5335             : //   x2 = load i32
    5336             : //   x2' = and x2, 0xff
    5337             : //   ...
    5338             : // b2:
    5339             : //   x = phi x1', x2'
    5340             : //   y = and x, 0xff
    5341      533616 : bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
    5342      524912 :   if (!Load->isSimple() ||
    5343      750344 :       !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
    5344             :     return false;
    5345             : 
    5346             :   // Skip loads we've already transformed.
    5347      658078 :   if (Load->hasOneUse() &&
    5348      632232 :       InsertedInsts.count(cast<Instruction>(*Load->user_begin())))
    5349             :     return false;
    5350             : 
    5351             :   // Look at all uses of Load, looking through phis, to determine how many bits
    5352             :   // of the loaded value are needed.
    5353             :   SmallVector<Instruction *, 8> WorkList;
    5354             :   SmallPtrSet<Instruction *, 16> Visited;
    5355             :   SmallVector<Instruction *, 8> AndsToMaybeRemove;
    5356      728153 :   for (auto *U : Load->users())
    5357      385570 :     WorkList.push_back(cast<Instruction>(U));
    5358             : 
    5359      342583 :   EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
    5360      342583 :   unsigned BitWidth = LoadResultVT.getSizeInBits();
    5361             :   APInt DemandBits(BitWidth, 0);
    5362             :   APInt WidestAndBits(BitWidth, 0);
    5363             : 
    5364      352610 :   while (!WorkList.empty()) {
    5365      349920 :     Instruction *I = WorkList.back();
    5366             :     WorkList.pop_back();
    5367             : 
    5368             :     // Break use-def graph loops.
    5369      349920 :     if (!Visited.insert(I).second)
    5370        7605 :       continue;
    5371             : 
    5372             :     // For a PHI node, push all of its users.
    5373      349854 :     if (auto *Phi = dyn_cast<PHINode>(I)) {
    5374       18031 :       for (auto *U : Phi->users())
    5375       10558 :         WorkList.push_back(cast<Instruction>(U));
    5376        7473 :       continue;
    5377             :     }
    5378             : 
    5379      342381 :     switch (I->getOpcode()) {
    5380        2643 :     case Instruction::And: {
    5381        2643 :       auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
    5382             :       if (!AndC)
    5383        1412 :         return false;
    5384             :       APInt AndBits = AndC->getValue();
    5385             :       DemandBits |= AndBits;
    5386             :       // Keep track of the widest and mask we see.
    5387        1231 :       if (AndBits.ugt(WidestAndBits))
    5388        1163 :         WidestAndBits = AndBits;
    5389        2394 :       if (AndBits == WidestAndBits && I->getOperand(0) == Load)
    5390         978 :         AndsToMaybeRemove.push_back(I);
    5391             :       break;
    5392             :     }
    5393             : 
    5394        1184 :     case Instruction::Shl: {
    5395        1184 :       auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
    5396             :       if (!ShlC)
    5397             :         return false;
    5398         862 :       uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
    5399         862 :       DemandBits.setLowBits(BitWidth - ShiftAmt);
    5400             :       break;
    5401             :     }
    5402             : 
    5403         395 :     case Instruction::Trunc: {
    5404         395 :       EVT TruncVT = TLI->getValueType(*DL, I->getType());
    5405         395 :       unsigned TruncBitWidth = TruncVT.getSizeInBits();
    5406             :       DemandBits.setLowBits(TruncBitWidth);
    5407             :       break;
    5408             :     }
    5409             : 
    5410             :     default:
    5411             :       return false;
    5412             :     }
    5413             :   }
    5414             : 
    5415             :   uint32_t ActiveBits = DemandBits.getActiveBits();
    5416             :   // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
    5417             :   // target even if isLoadExtLegal says an i1 EXTLOAD is valid.  For example,
    5418             :   // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
    5419             :   // (and (load x) 1) is not matched as a single instruction, rather as a LDR
    5420             :   // followed by an AND.
    5421             :   // TODO: Look into removing this restriction by fixing backends to either
    5422             :   // return false for isLoadExtLegal for i1 or have them select this pattern to
    5423             :   // a single instruction.
    5424             :   //
    5425             :   // Also avoid hoisting if we didn't see any ands with the exact DemandBits
    5426             :   // mask, since these are the only ands that will be removed by isel.
    5427        4031 :   if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) ||
    5428             :       WidestAndBits != DemandBits)
    5429             :     return false;
    5430             : 
    5431         211 :   LLVMContext &Ctx = Load->getType()->getContext();
    5432         211 :   Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
    5433         211 :   EVT TruncVT = TLI->getValueType(*DL, TruncTy);
    5434             : 
    5435             :   // Reject cases that won't be matched as extloads.
    5436         274 :   if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
    5437          63 :       !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
    5438             :     return false;
    5439             : 
    5440          63 :   IRBuilder<> Builder(Load->getNextNode());
    5441         126 :   auto *NewAnd = dyn_cast<Instruction>(
    5442          63 :       Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
    5443             :   // Mark this instruction as "inserted by CGP", so that other
    5444             :   // optimizations don't touch it.
    5445          63 :   InsertedInsts.insert(NewAnd);
    5446             : 
    5447             :   // Replace all uses of load with new and (except for the use of load in the
    5448             :   // new and itself).
    5449          63 :   Load->replaceAllUsesWith(NewAnd);
    5450          63 :   NewAnd->setOperand(0, Load);
    5451             : 
    5452             :   // Remove any and instructions that are now redundant.
    5453         173 :   for (auto *And : AndsToMaybeRemove)
    5454             :     // Check that the and mask is the same as the one we decided to put on the
    5455             :     // new and.
    5456         110 :     if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
    5457          55 :       And->replaceAllUsesWith(NewAnd);
    5458          55 :       if (&*CurInstIterator == And)
    5459          72 :         CurInstIterator = std::next(And->getIterator());
    5460          55 :       And->eraseFromParent();
    5461             :       ++NumAndUses;
    5462             :     }
    5463             : 
    5464             :   ++NumAndsAdded;
    5465             :   return true;
    5466             : }
    5467             : 
    5468             : /// Check if V (an operand of a select instruction) is an expensive instruction
    5469             : /// that is only used once.
    5470        8605 : static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
    5471             :   auto *I = dyn_cast<Instruction>(V);
    5472             :   // If it's safe to speculatively execute, then it should not have side
    5473             :   // effects; therefore, it's safe to sink and possibly *not* execute.
    5474        2139 :   return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
    5475        9216 :          TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
    5476             : }
    5477             : 
    5478             : /// Returns true if a SelectInst should be turned into an explicit branch.
    5479       59377 : static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
    5480             :                                                 const TargetLowering *TLI,
    5481             :                                                 SelectInst *SI) {
    5482             :   // If even a predictable select is cheap, then a branch can't be cheaper.
    5483       59377 :   if (!TLI->isPredictableSelectExpensive())
    5484             :     return false;
    5485             : 
    5486             :   // FIXME: This should use the same heuristics as IfConversion to determine
    5487             :   // whether a select is better represented as a branch.
    5488             : 
    5489             :   // If metadata tells us that the select condition is obviously predictable,
    5490             :   // then we want to replace the select with a branch.
    5491             :   uint64_t TrueWeight, FalseWeight;
    5492       52337 :   if (SI->extractProfMetadata(TrueWeight, FalseWeight)) {
    5493           9 :     uint64_t Max = std::max(TrueWeight, FalseWeight);
    5494           9 :     uint64_t Sum = TrueWeight + FalseWeight;
    5495           9 :     if (Sum != 0) {
    5496           8 :       auto Probability = BranchProbability::getBranchProbability(Max, Sum);
    5497           8 :       if (Probability > TLI->getPredictableBranchThreshold())
    5498           7 :         return true;
    5499             :     }
    5500             :   }
    5501             : 
    5502             :   CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
    5503             : 
    5504             :   // If a branch is predictable, an out-of-order CPU can avoid blocking on its
    5505             :   // comparison condition. If the compare has more than one use, there's
    5506             :   // probably another cmov or setcc around, so it's not worth emitting a branch.
    5507       44043 :   if (!Cmp || !Cmp->hasOneUse())
    5508             :     return false;
    5509             : 
    5510             :   // If either operand of the select is expensive and only needed on one side
    5511             :   // of the select, we should form a branch.
    5512        8557 :   if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
    5513        4277 :       sinkSelectOperand(TTI, SI->getFalseValue()))
    5514             :     return true;
    5515             : 
    5516             :   return false;
    5517             : }
    5518             : 
    5519             : /// If \p isTrue is true, return the true value of \p SI, otherwise return
    5520             : /// false value of \p SI. If the true/false value of \p SI is defined by any
    5521             : /// select instructions in \p Selects, look through the defining select
    5522             : /// instruction until the true/false value is not defined in \p Selects.
    5523          48 : static Value *getTrueOrFalseValue(
    5524             :     SelectInst *SI, bool isTrue,
    5525             :     const SmallPtrSet<const Instruction *, 2> &Selects) {
    5526             :   Value *V;
    5527             : 
    5528          99 :   for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI);
    5529             :        DefSI = dyn_cast<SelectInst>(V)) {
    5530             :     assert(DefSI->getCondition() == SI->getCondition() &&
    5531             :            "The condition of DefSI does not match with SI");
    5532          51 :     V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
    5533             :   }
    5534          48 :   return V;
    5535             : }
    5536             : 
    5537             : /// If we have a SelectInst that will likely profit from branch prediction,
    5538             : /// turn it into a branch.
    5539       79330 : bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
    5540             :   // Find all consecutive select instructions that share the same condition.
    5541             :   SmallVector<SelectInst *, 2> ASI;
    5542       79330 :   ASI.push_back(SI);
    5543       79330 :   for (BasicBlock::iterator It = ++BasicBlock::iterator(SI);
    5544      159608 :        It != SI->getParent()->end(); ++It) {
    5545       79804 :     SelectInst *I = dyn_cast<SelectInst>(&*It);
    5546       80586 :     if (I && SI->getCondition() == I->getCondition()) {
    5547         474 :       ASI.push_back(I);
    5548             :     } else {
    5549             :       break;
    5550             :     }
    5551             :   }
    5552             : 
    5553       79330 :   SelectInst *LastSI = ASI.back();
    5554             :   // Increment the current iterator to skip all the rest of select instructions
    5555             :   // because they will be either "not lowered" or "all lowered" to branch.
    5556      158660 :   CurInstIterator = std::next(LastSI->getIterator());
    5557             : 
    5558       79330 :   bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
    5559             : 
    5560             :   // Can we convert the 'select' to CF ?
    5561      128944 :   if (DisableSelectToBranch || OptSize || !TLI || VectorCond ||
    5562       59388 :       SI->getMetadata(LLVMContext::MD_unpredictable))
    5563             :     return false;
    5564             : 
    5565             :   TargetLowering::SelectSupportKind SelectKind;
    5566             :   if (VectorCond)
    5567             :     SelectKind = TargetLowering::VectorMaskSelect;
    5568      118774 :   else if (SI->getType()->isVectorTy())
    5569             :     SelectKind = TargetLowering::ScalarCondVectorVal;
    5570             :   else
    5571             :     SelectKind = TargetLowering::ScalarValSelect;
    5572             : 
    5573      118764 :   if (TLI->isSelectSupported(SelectKind) &&
    5574       59377 :       !isFormingBranchFromSelectProfitable(TTI, TLI, SI))
    5575             :     return false;
    5576             : 
    5577          22 :   ModifiedDT = true;
    5578             : 
    5579             :   // Transform a sequence like this:
    5580             :   //    start:
    5581             :   //       %cmp = cmp uge i32 %a, %b
    5582             :   //       %sel = select i1 %cmp, i32 %c, i32 %d
    5583             :   //
    5584             :   // Into:
    5585             :   //    start:
    5586             :   //       %cmp = cmp uge i32 %a, %b
    5587             :   //       br i1 %cmp, label %select.true, label %select.false
    5588             :   //    select.true:
    5589             :   //       br label %select.end
    5590             :   //    select.false:
    5591             :   //       br label %select.end
    5592             :   //    select.end:
    5593             :   //       %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
    5594             :   //
    5595             :   // In addition, we may sink instructions that produce %c or %d from
    5596             :   // the entry block into the destination(s) of the new branch.
    5597             :   // If the true or false blocks do not contain a sunken instruction, that
    5598             :   // block and its branch may be optimized away. In that case, one side of the
    5599             :   // first branch will point directly to select.end, and the corresponding PHI
    5600             :   // predecessor block will be the start block.
    5601             : 
    5602             :   // First, we split the block containing the select into 2 blocks.
    5603          22 :   BasicBlock *StartBlock = SI->getParent();
    5604          22 :   BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
    5605          22 :   BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
    5606             : 
    5607             :   // Delete the unconditional branch that was just created by the split.
    5608          22 :   StartBlock->getTerminator()->eraseFromParent();
    5609             : 
    5610             :   // These are the new basic blocks for the conditional branch.
    5611             :   // At least one will become an actual new basic block.
    5612             :   BasicBlock *TrueBlock = nullptr;
    5613             :   BasicBlock *FalseBlock = nullptr;
    5614             :   BranchInst *TrueBranch = nullptr;
    5615             :   BranchInst *FalseBranch = nullptr;
    5616             : 
    5617             :   // Sink expensive instructions into the conditional blocks to avoid executing
    5618             :   // them speculatively.
    5619          70 :   for (SelectInst *SI : ASI) {
    5620          24 :     if (sinkSelectOperand(TTI, SI->getTrueValue())) {
    5621           3 :       if (TrueBlock == nullptr) {
    5622           9 :         TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
    5623             :                                        EndBlock->getParent(), EndBlock);
    5624             :         TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
    5625             :       }
    5626             :       auto *TrueInst = cast<Instruction>(SI->getTrueValue());
    5627           3 :       TrueInst->moveBefore(TrueBranch);
    5628             :     }
    5629          24 :     if (sinkSelectOperand(TTI, SI->getFalseValue())) {
    5630           3 :       if (FalseBlock == nullptr) {
    5631           9 :         FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
    5632             :                                         EndBlock->getParent(), EndBlock);
    5633             :         FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
    5634             :       }
    5635             :       auto *FalseInst = cast<Instruction>(SI->getFalseValue());
    5636           3 :       FalseInst->moveBefore(FalseBranch);
    5637             :     }
    5638             :   }
    5639             : 
    5640             :   // If there was nothing to sink, then arbitrarily choose the 'false' side
    5641             :   // for a new input value to the PHI.
    5642          22 :   if (TrueBlock == FalseBlock) {
    5643             :     assert(TrueBlock == nullptr &&
    5644             :            "Unexpected basic block transform while optimizing select");
    5645             : 
    5646          51 :     FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
    5647             :                                     EndBlock->getParent(), EndBlock);
    5648          17 :     BranchInst::Create(EndBlock, FalseBlock);
    5649             :   }
    5650             : 
    5651             :   // Insert the real conditional branch based on the original condition.
    5652             :   // If we did not create a new block for one of the 'true' or 'false' paths
    5653             :   // of the condition, it means that side of the branch goes to the end block
    5654             :   // directly and the path originates from the start block from the point of
    5655             :   // view of the new PHI.
    5656             :   BasicBlock *TT, *FT;
    5657          22 :   if (TrueBlock == nullptr) {
    5658             :     TT = EndBlock;
    5659             :     FT = FalseBlock;
    5660             :     TrueBlock = StartBlock;
    5661           3 :   } else if (FalseBlock == nullptr) {
    5662             :     TT = TrueBlock;
    5663             :     FT = EndBlock;
    5664             :     FalseBlock = StartBlock;
    5665             :   } else {
    5666             :     TT = TrueBlock;
    5667             :     FT = FalseBlock;
    5668             :   }
    5669          66 :   IRBuilder<>(SI).CreateCondBr(SI->getCondition(), TT, FT, SI);
    5670             : 
    5671             :   SmallPtrSet<const Instruction *, 2> INS;
    5672             :   INS.insert(ASI.begin(), ASI.end());
    5673             :   // Use reverse iterator because later select may use the value of the
    5674             :   // earlier select, and we need to propagate value through earlier select
    5675             :   // to get the PHI operand.
    5676          22 :   for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) {
    5677          24 :     SelectInst *SI = *It;
    5678             :     // The select itself is replaced with a PHI Node.
    5679          48 :     PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
    5680          24 :     PN->takeName(SI);
    5681          24 :     PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock);
    5682          24 :     PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock);
    5683             : 
    5684          24 :     SI->replaceAllUsesWith(PN);
    5685          24 :     SI->eraseFromParent();
    5686             :     INS.erase(SI);
    5687             :     ++NumSelectsExpanded;
    5688             :   }
    5689             : 
    5690             :   // Instruct OptimizeBlock to skip to the next block.
    5691          22 :   CurInstIterator = StartBlock->end();
    5692             :   return true;
    5693             : }
    5694             : 
    5695       16563 : static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
    5696             :   SmallVector<int, 16> Mask(SVI->getShuffleMask());
    5697             :   int SplatElem = -1;
    5698      143967 :   for (unsigned i = 0; i < Mask.size(); ++i) {
    5699       74290 :     if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
    5700             :       return false;
    5701       36947 :     SplatElem = Mask[i];
    5702             :   }
    5703             : 
    5704             :   return true;
    5705             : }
    5706             : 
    5707             : /// Some targets have expensive vector shifts if the lanes aren't all the same
    5708             : /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
    5709             : /// it's often worth sinking a shufflevector splat down to its use so that
    5710             : /// codegen can spot all lanes are identical.
    5711       44889 : bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
    5712       44889 :   BasicBlock *DefBB = SVI->getParent();
    5713             : 
    5714             :   // Only do this xform if variable vector shifts are particularly expensive.
    5715       89778 :   if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
    5716             :     return false;
    5717             : 
    5718             :   // We only expect better codegen by sinking a shuffle if we can recognise a
    5719             :   // constant splat.
    5720       16563 :   if (!isBroadcastShuffle(SVI))
    5721             :     return false;
    5722             : 
    5723             :   // InsertedShuffles - Only insert a shuffle in each block once.
    5724             :   DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
    5725             : 
    5726             :   bool MadeChange = false;
    5727        4575 :   for (User *U : SVI->users()) {
    5728             :     Instruction *UI = cast<Instruction>(U);
    5729             : 
    5730             :     // Figure out which BB this ext is used in.
    5731        2557 :     BasicBlock *UserBB = UI->getParent();
    5732        5110 :     if (UserBB == DefBB) continue;
    5733             : 
    5734             :     // For now only apply this when the splat is used by a shift instruction.
    5735          50 :     if (!UI->isShift()) continue;
    5736             : 
    5737             :     // Everything checks out, sink the shuffle if the user's block doesn't
    5738             :     // already have a copy.
    5739             :     Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
    5740             : 
    5741           4 :     if (!InsertedShuffle) {
    5742           4 :       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
    5743             :       assert(InsertPt != UserBB->end());
    5744           4 :       InsertedShuffle =
    5745             :           new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
    5746           4 :                                 SVI->getOperand(2), "", &*InsertPt);
    5747             :     }
    5748             : 
    5749           4 :     UI->replaceUsesOfWith(SVI, InsertedShuffle);
    5750             :     MadeChange = true;
    5751             :   }
    5752             : 
    5753             :   // If we removed all uses, nuke the shuffle.
    5754        2018 :   if (SVI->use_empty()) {
    5755           2 :     SVI->eraseFromParent();
    5756             :     MadeChange = true;
    5757             :   }
    5758             : 
    5759             :   return MadeChange;
    5760             : }
    5761             : 
    5762        1405 : bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
    5763        1405 :   if (!TLI || !DL)
    5764             :     return false;
    5765             : 
    5766             :   Value *Cond = SI->getCondition();
    5767        1401 :   Type *OldType = Cond->getType();
    5768        1401 :   LLVMContext &Context = Cond->getContext();
    5769        1401 :   MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
    5770        1401 :   unsigned RegWidth = RegType.getSizeInBits();
    5771             : 
    5772        1401 :   if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
    5773             :     return false;
    5774             : 
    5775             :   // If the register width is greater than the type width, expand the condition
    5776             :   // of the switch instruction and each case constant to the width of the
    5777             :   // register. By widening the type of the switch condition, subsequent
    5778             :   // comparisons (for case comparisons) will not need to be extended to the
    5779             :   // preferred register width, so we will potentially eliminate N-1 extends,
    5780             :   // where N is the number of cases in the switch.
    5781          51 :   auto *NewType = Type::getIntNTy(Context, RegWidth);
    5782             : 
    5783             :   // Zero-extend the switch condition and case constants unless the switch
    5784             :   // condition is a function argument that is already being sign-extended.
    5785             :   // In that case, we can avoid an unnecessary mask/extension by sign-extending
    5786             :   // everything instead.
    5787             :   Instruction::CastOps ExtType = Instruction::ZExt;
    5788             :   if (auto *Arg = dyn_cast<Argument>(Cond))
    5789           3 :     if (Arg->hasSExtAttr())
    5790             :       ExtType = Instruction::SExt;
    5791             : 
    5792          51 :   auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
    5793          51 :   ExtInst->insertBefore(SI);
    5794             :   SI->setCondition(ExtInst);
    5795         181 :   for (auto Case : SI->cases()) {
    5796             :     APInt NarrowConst = Case.getCaseValue()->getValue();
    5797             :     APInt WideConst = (ExtType == Instruction::ZExt) ?
    5798         130 :                       NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
    5799         130 :     Case.setValue(ConstantInt::get(Context, WideConst));
    5800             :   }
    5801             : 
    5802             :   return true;
    5803             : }
    5804             : 
    5805             : 
    5806             : namespace {
    5807             : 
    5808             : /// Helper class to promote a scalar operation to a vector one.
    5809             : /// This class is used to move downward extractelement transition.
    5810             : /// E.g.,
    5811             : /// a = vector_op <2 x i32>
    5812             : /// b = extractelement <2 x i32> a, i32 0
    5813             : /// c = scalar_op b
    5814             : /// store c
    5815             : ///
    5816             : /// =>
    5817             : /// a = vector_op <2 x i32>
    5818             : /// c = vector_op a (equivalent to scalar_op on the related lane)
    5819             : /// * d = extractelement <2 x i32> c, i32 0
    5820             : /// * store d
    5821             : /// Assuming both extractelement and store can be combine, we get rid of the
    5822             : /// transition.
    5823             : class VectorPromoteHelper {
    5824             :   /// DataLayout associated with the current module.
    5825             :   const DataLayout &DL;
    5826             : 
    5827             :   /// Used to perform some checks on the legality of vector operations.
    5828             :   const TargetLowering &TLI;
    5829             : 
    5830             :   /// Used to estimated the cost of the promoted chain.
    5831             :   const TargetTransformInfo &TTI;
    5832             : 
    5833             :   /// The transition being moved downwards.
    5834             :   Instruction *Transition;
    5835             : 
    5836             :   /// The sequence of instructions to be promoted.
    5837             :   SmallVector<Instruction *, 4> InstsToBePromoted;
    5838             : 
    5839             :   /// Cost of combining a store and an extract.
    5840             :   unsigned StoreExtractCombineCost;
    5841             : 
    5842             :   /// Instruction that will be combined with the transition.
    5843             :   Instruction *CombineInst = nullptr;
    5844             : 
    5845             :   /// The instruction that represents the current end of the transition.
    5846             :   /// Since we are faking the promotion until we reach the end of the chain
    5847             :   /// of computation, we need a way to get the current end of the transition.
    5848             :   Instruction *getEndOfTransition() const {
    5849         114 :     if (InstsToBePromoted.empty())
    5850          78 :       return Transition;
    5851          36 :     return InstsToBePromoted.back();
    5852             :   }
    5853             : 
    5854             :   /// Return the index of the original value in the transition.
    5855             :   /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
    5856             :   /// c, is at index 0.
    5857             :   unsigned getTransitionOriginalValueIdx() const {
    5858             :     assert(isa<ExtractElementInst>(Transition) &&
    5859             :            "Other kind of transitions are not supported yet");
    5860             :     return 0;
    5861             :   }
    5862             : 
    5863             :   /// Return the index of the index in the transition.
    5864             :   /// E.g., for "extractelement <2 x i32> c, i32 0" the index
    5865             :   /// is at index 1.
    5866             :   unsigned getTransitionIdx() const {
    5867             :     assert(isa<ExtractElementInst>(Transition) &&
    5868             :            "Other kind of transitions are not supported yet");
    5869             :     return 1;
    5870             :   }
    5871             : 
    5872             :   /// Get the type of the transition.
    5873             :   /// This is the type of the original value.
    5874             :   /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
    5875             :   /// transition is <2 x i32>.
    5876             :   Type *getTransitionType() const {
    5877         218 :     return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
    5878             :   }
    5879             : 
    5880             :   /// Promote \p ToBePromoted by moving \p Def downward through.
    5881             :   /// I.e., we have the following sequence:
    5882             :   /// Def = Transition <ty1> a to <ty2>
    5883             :   /// b = ToBePromoted <ty2> Def, ...
    5884             :   /// =>
    5885             :   /// b = ToBePromoted <ty1> a, ...
    5886             :   /// Def = Transition <ty1> ToBePromoted to <ty2>
    5887             :   void promoteImpl(Instruction *ToBePromoted);
    5888             : 
    5889             :   /// Check whether or not it is profitable to promote all the
    5890             :   /// instructions enqueued to be promoted.
    5891           6 :   bool isProfitableToPromote() {
    5892           6 :     Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
    5893             :     unsigned Index = isa<ConstantInt>(ValIdx)
    5894             :                          ? cast<ConstantInt>(ValIdx)->getZExtValue()
    5895           6 :                          : -1;
    5896             :     Type *PromotedType = getTransitionType();
    5897             : 
    5898           6 :     StoreInst *ST = cast<StoreInst>(CombineInst);
    5899             :     unsigned AS = ST->getPointerAddressSpace();
    5900             :     unsigned Align = ST->getAlignment();
    5901             :     // Check if this store is supported.
    5902          12 :     if (!TLI.allowsMisalignedMemoryAccesses(
    5903             :             TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
    5904           6 :             Align)) {
    5905             :       // If this is not supported, there is no way we can combine
    5906             :       // the extract with the store.
    5907             :       return false;
    5908             :     }
    5909             : 
    5910             :     // The scalar chain of computation has to pay for the transition
    5911             :     // scalar to vector.
    5912             :     // The vector chain has to account for the combining cost.
    5913             :     uint64_t ScalarCost =
    5914          12 :         TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
    5915           6 :     uint64_t VectorCost = StoreExtractCombineCost;
    5916          42 :     for (const auto &Inst : InstsToBePromoted) {
    5917             :       // Compute the cost.
    5918             :       // By construction, all instructions being promoted are arithmetic ones.
    5919             :       // Moreover, one argument is a constant that can be viewed as a splat
    5920             :       // constant.
    5921          18 :       Value *Arg0 = Inst->getOperand(0);
    5922          18 :       bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
    5923             :                             isa<ConstantFP>(Arg0);
    5924             :       TargetTransformInfo::OperandValueKind Arg0OVK =
    5925             :           IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
    5926             :                          : TargetTransformInfo::OK_AnyValue;
    5927             :       TargetTransformInfo::OperandValueKind Arg1OVK =
    5928          18 :           !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
    5929             :                           : TargetTransformInfo::OK_AnyValue;
    5930          54 :       ScalarCost += TTI.getArithmeticInstrCost(
    5931          18 :           Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
    5932          54 :       VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
    5933          18 :                                                Arg0OVK, Arg1OVK);
    5934             :     }
    5935             :     LLVM_DEBUG(
    5936             :         dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
    5937             :                << ScalarCost << "\nVector: " << VectorCost << '\n');
    5938           6 :     return ScalarCost > VectorCost;
    5939             :   }
    5940             : 
    5941             :   /// Generate a constant vector with \p Val with the same
    5942             :   /// number of elements as the transition.
    5943             :   /// \p UseSplat defines whether or not \p Val should be replicated
    5944             :   /// across the whole vector.
    5945             :   /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
    5946             :   /// otherwise we generate a vector with as many undef as possible:
    5947             :   /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
    5948             :   /// used at the index of the extract.
    5949          38 :   Value *getConstantVector(Constant *Val, bool UseSplat) const {
    5950             :     unsigned ExtractIdx = std::numeric_limits<unsigned>::max();
    5951          38 :     if (!UseSplat) {
    5952             :       // If we cannot determine where the constant must be, we have to
    5953             :       // use a splat constant.
    5954          32 :       Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
    5955             :       if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
    5956          31 :         ExtractIdx = CstVal->getSExtValue();
    5957             :       else
    5958             :         UseSplat = true;
    5959             :     }
    5960             : 
    5961          38 :     unsigned End = getTransitionType()->getVectorNumElements();
    5962          38 :     if (UseSplat)
    5963           7 :       return ConstantVector::getSplat(End, Val);
    5964             : 
    5965             :     SmallVector<Constant *, 4> ConstVec;
    5966          31 :     UndefValue *UndefVal = UndefValue::get(Val->getType());
    5967         179 :     for (unsigned Idx = 0; Idx != End; ++Idx) {
    5968          74 :       if (Idx == ExtractIdx)
    5969          31 :         ConstVec.push_back(Val);
    5970             :       else
    5971          43 :         ConstVec.push_back(UndefVal);
    5972             :     }
    5973          31 :     return ConstantVector::get(ConstVec);
    5974             :   }
    5975             : 
    5976             :   /// Check if promoting to a vector type an operand at \p OperandIdx
    5977             :   /// in \p Use can trigger undefined behavior.
    5978          95 :   static bool canCauseUndefinedBehavior(const Instruction *Use,
    5979             :                                         unsigned OperandIdx) {
    5980             :     // This is not safe to introduce undef when the operand is on
    5981             :     // the right hand side of a division-like instruction.
    5982          95 :     if (OperandIdx != 1)
    5983             :       return false;
    5984             :     switch (Use->getOpcode()) {
    5985             :     default:
    5986             :       return false;
    5987          10 :     case Instruction::SDiv:
    5988             :     case Instruction::UDiv:
    5989             :     case Instruction::SRem:
    5990             :     case Instruction::URem:
    5991          10 :       return true;
    5992           4 :     case Instruction::FDiv:
    5993             :     case Instruction::FRem:
    5994           4 :       return !Use->hasNoNaNs();
    5995             :     }
    5996             :     llvm_unreachable(nullptr);
    5997             :   }
    5998             : 
    5999             : public:
    6000             :   VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
    6001             :                       const TargetTransformInfo &TTI, Instruction *Transition,
    6002             :                       unsigned CombineCost)
    6003         136 :       : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
    6004         272 :         StoreExtractCombineCost(CombineCost) {
    6005             :     assert(Transition && "Do not know how to promote null");
    6006             :   }
    6007             : 
    6008             :   /// Check if we can promote \p ToBePromoted to \p Type.
    6009             :   bool canPromote(const Instruction *ToBePromoted) const {
    6010             :     // We could support CastInst too.
    6011             :     return isa<BinaryOperator>(ToBePromoted);
    6012             :   }
    6013             : 
    6014             :   /// Check if it is profitable to promote \p ToBePromoted
    6015             :   /// by moving downward the transition through.
    6016          57 :   bool shouldPromote(const Instruction *ToBePromoted) const {
    6017             :     // Promote only if all the operands can be statically expanded.
    6018             :     // Indeed, we do not want to introduce any new kind of transitions.
    6019         330 :     for (const Use &U : ToBePromoted->operands()) {
    6020         114 :       const Value *Val = U.get();
    6021         165 :       if (Val == getEndOfTransition()) {
    6022             :         // If the use is a division and the transition is on the rhs,
    6023             :         // we cannot promote the operation, otherwise we may create a
    6024             :         // division by zero.
    6025          57 :         if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
    6026             :           return false;
    6027             :         continue;
    6028             :       }
    6029          57 :       if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
    6030             :           !isa<ConstantFP>(Val))
    6031             :         return false;
    6032             :     }
    6033             :     // Check that the resulting operation is legal.
    6034         102 :     int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
    6035          51 :     if (!ISDOpcode)
    6036             :       return false;
    6037          81 :     return StressStoreExtract ||
    6038          60 :            TLI.isOperationLegalOrCustom(
    6039             :                ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
    6040             :   }
    6041             : 
    6042             :   /// Check whether or not \p Use can be combined
    6043             :   /// with the transition.
    6044             :   /// I.e., is it possible to do Use(Transition) => AnotherUse?
    6045             :   bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
    6046             : 
    6047             :   /// Record \p ToBePromoted as part of the chain to be promoted.
    6048             :   void enqueueForPromotion(Instruction *ToBePromoted) {
    6049          41 :     InstsToBePromoted.push_back(ToBePromoted);
    6050             :   }
    6051             : 
    6052             :   /// Set the instruction that will be combined with the transition.
    6053             :   void recordCombineInstruction(Instruction *ToBeCombined) {
    6054             :     assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
    6055          73 :     CombineInst = ToBeCombined;
    6056             :   }
    6057             : 
    6058             :   /// Promote all the instructions enqueued for promotion if it is
    6059             :   /// is profitable.
    6060             :   /// \return True if the promotion happened, false otherwise.
    6061          73 :   bool promote() {
    6062             :     // Check if there is something to promote.
    6063             :     // Right now, if we do not have anything to combine with,
    6064             :     // we assume the promotion is not profitable.
    6065          73 :     if (InstsToBePromoted.empty() || !CombineInst)
    6066             :       return false;
    6067             : 
    6068             :     // Check cost.
    6069          20 :     if (!StressStoreExtract && !isProfitableToPromote())
    6070             :       return false;
    6071             : 
    6072             :     // Promote.
    6073          96 :     for (auto &ToBePromoted : InstsToBePromoted)
    6074          38 :       promoteImpl(ToBePromoted);
    6075             :     InstsToBePromoted.clear();
    6076          20 :     return true;
    6077             :   }
    6078             : };
    6079             : 
    6080             : } // end anonymous namespace
    6081             : 
    6082          38 : void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
    6083             :   // At this point, we know that all the operands of ToBePromoted but Def
    6084             :   // can be statically promoted.
    6085             :   // For Def, we need to use its parameter in ToBePromoted:
    6086             :   // b = ToBePromoted ty1 a
    6087             :   // Def = Transition ty1 b to ty2
    6088             :   // Move the transition down.
    6089             :   // 1. Replace all uses of the promoted operation by the transition.
    6090             :   // = ... b => = ... Def.
    6091             :   assert(ToBePromoted->getType() == Transition->getType() &&
    6092             :          "The type of the result of the transition does not match "
    6093             :          "the final type");
    6094          38 :   ToBePromoted->replaceAllUsesWith(Transition);
    6095             :   // 2. Update the type of the uses.
    6096             :   // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
    6097          38 :   Type *TransitionTy = getTransitionType();
    6098             :   ToBePromoted->mutateType(TransitionTy);
    6099             :   // 3. Update all the operands of the promoted operation with promoted
    6100             :   // operands.
    6101             :   // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
    6102         228 :   for (Use &U : ToBePromoted->operands()) {
    6103          76 :     Value *Val = U.get();
    6104             :     Value *NewVal = nullptr;
    6105          76 :     if (Val == Transition)
    6106          38 :       NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
    6107          38 :     else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
    6108             :              isa<ConstantFP>(Val)) {
    6109             :       // Use a splat constant if it is not safe to use undef.
    6110          38 :       NewVal = getConstantVector(
    6111             :           cast<Constant>(Val),
    6112          76 :           isa<UndefValue>(Val) ||
    6113          38 :               canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
    6114             :     } else
    6115           0 :       llvm_unreachable("Did you modified shouldPromote and forgot to update "
    6116             :                        "this?");
    6117          76 :     ToBePromoted->setOperand(U.getOperandNo(), NewVal);
    6118             :   }
    6119          38 :   Transition->moveAfter(ToBePromoted);
    6120          38 :   Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
    6121          38 : }
    6122             : 
    6123             : /// Some targets can do store(extractelement) with one instruction.
    6124             : /// Try to push the extractelement towards the stores when the target
    6125             : /// has this feature and this is profitable.
    6126       25304 : bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
    6127       25304 :   unsigned CombineCost = std::numeric_limits<unsigned>::max();
    6128       50608 :   if (DisableStoreExtract || !TLI ||
    6129       25257 :       (!StressStoreExtract &&
    6130       75771 :        !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
    6131       25257 :                                        Inst->getOperand(1), CombineCost)))
    6132             :     return false;
    6133             : 
    6134             :   // At this point we know that Inst is a vector to scalar transition.
    6135             :   // Try to move it down the def-use chain, until:
    6136             :   // - We can combine the transition with its single use
    6137             :   //   => we got rid of the transition.
    6138             :   // - We escape the current basic block
    6139             :   //   => we would need to check that we are moving it at a cheaper place and
    6140             :   //      we do not do that for now.
    6141         136 :   BasicBlock *Parent = Inst->getParent();
    6142             :   LLVM_DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
    6143         136 :   VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
    6144             :   // If the transition has more than one use, assume this is not going to be
    6145             :   // beneficial.
    6146         218 :   while (Inst->hasOneUse()) {
    6147             :     Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
    6148             :     LLVM_DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
    6149             : 
    6150         164 :     if (ToBePromoted->getParent() != Parent) {
    6151             :       LLVM_DEBUG(dbgs() << "Instruction to promote is in a different block ("
    6152             :                         << ToBePromoted->getParent()->getName()
    6153             :                         << ") than the transition (" << Parent->getName()
    6154             :                         << ").\n");
    6155             :       return false;
    6156             :     }
    6157             : 
    6158         161 :     if (VPH.canCombine(ToBePromoted)) {
    6159             :       LLVM_DEBUG(dbgs() << "Assume " << *Inst << '\n'
    6160             :                         << "will be combined with: " << *ToBePromoted << '\n');
    6161             :       VPH.recordCombineInstruction(ToBePromoted);
    6162          73 :       bool Changed = VPH.promote();
    6163             :       NumStoreExtractExposed += Changed;
    6164          73 :       return Changed;
    6165             :     }
    6166             : 
    6167             :     LLVM_DEBUG(dbgs() << "Try promoting.\n");
    6168          88 :     if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
    6169             :       return false;
    6170             : 
    6171             :     LLVM_DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
    6172             : 
    6173             :     VPH.enqueueForPromotion(ToBePromoted);
    6174             :     Inst = ToBePromoted;
    6175             :   }
    6176             :   return false;
    6177             : }
    6178             : 
    6179             : /// For the instruction sequence of store below, F and I values
    6180             : /// are bundled together as an i64 value before being stored into memory.
    6181             : /// Sometimes it is more efficent to generate separate stores for F and I,
    6182             : /// which can remove the bitwise instructions or sink them to colder places.
    6183             : ///
    6184             : ///   (store (or (zext (bitcast F to i32) to i64),
    6185             : ///              (shl (zext I to i64), 32)), addr)  -->
    6186             : ///   (store F, addr) and (store I, addr+4)
    6187             : ///
    6188             : /// Similarly, splitting for other merged store can also be beneficial, like:
    6189             : /// For pair of {i32, i32}, i64 store --> two i32 stores.
    6190             : /// For pair of {i32, i16}, i64 store --> two i32 stores.
    6191             : /// For pair of {i16, i16}, i32 store --> two i16 stores.
    6192             : /// For pair of {i16, i8},  i32 store --> two i16 stores.
    6193             : /// For pair of {i8, i8},   i16 store --> two i8 stores.
    6194             : ///
    6195             : /// We allow each target to determine specifically which kind of splitting is
    6196             : /// supported.
    6197             : ///
    6198             : /// The store patterns are commonly seen from the simple code snippet below
    6199             : /// if only std::make_pair(...) is sroa transformed before inlined into hoo.
    6200             : ///   void goo(const std::pair<int, float> &);
    6201             : ///   hoo() {
    6202             : ///     ...
    6203             : ///     goo(std::make_pair(tmp, ftmp));
    6204             : ///     ...
    6205             : ///   }
    6206             : ///
    6207             : /// Although we already have similar splitting in DAG Combine, we duplicate
    6208             : /// it in CodeGenPrepare to catch the case in which pattern is across
    6209             : /// multiple BBs. The logic in DAG Combine is kept to catch case generated
    6210             : /// during code expansion.
    6211      547115 : static bool splitMergedValStore(StoreInst &SI, const DataLayout &DL,
    6212             :                                 const TargetLowering &TLI) {
    6213             :   // Handle simple but common cases only.
    6214      547115 :   Type *StoreType = SI.getValueOperand()->getType();
    6215     1092711 :   if (DL.getTypeStoreSizeInBits(StoreType) != DL.getTypeSizeInBits(StoreType) ||
    6216      545596 :       DL.getTypeSizeInBits(StoreType) == 0)
    6217             :     return false;
    6218             : 
    6219      545575 :   unsigned HalfValBitSize = DL.getTypeSizeInBits(StoreType) / 2;
    6220      545575 :   Type *SplitStoreType = Type::getIntNTy(SI.getContext(), HalfValBitSize);
    6221      545575 :   if (DL.getTypeStoreSizeInBits(SplitStoreType) !=
    6222      545575 :       DL.getTypeSizeInBits(SplitStoreType))
    6223             :     return false;
    6224             : 
    6225             :   // Match the following patterns:
    6226             :   // (store (or (zext LValue to i64),
    6227             :   //            (shl (zext HValue to i64), 32)), HalfValBitSize)
    6228             :   //  or
    6229             :   // (store (or (shl (zext HValue to i64), 32)), HalfValBitSize)
    6230             :   //            (zext LValue to i64),
    6231             :   // Expect both operands of OR and the first operand of SHL have only
    6232             :   // one use.
    6233             :   Value *LValue, *HValue;
    6234      534785 :   if (!match(SI.getValueOperand(),
    6235      534785 :              m_c_Or(m_OneUse(m_ZExt(m_Value(LValue))),
    6236             :                     m_OneUse(m_Shl(m_OneUse(m_ZExt(m_Value(HValue))),
    6237             :                                    m_SpecificInt(HalfValBitSize))))))
    6238             :     return false;
    6239             : 
    6240             :   // Check LValue and HValue are int with size less or equal than 32.
    6241         138 :   if (!LValue->getType()->isIntegerTy() ||
    6242         138 :       DL.getTypeSizeInBits(LValue->getType()) > HalfValBitSize ||
    6243         207 :       !HValue->getType()->isIntegerTy() ||
    6244          69 :       DL.getTypeSizeInBits(HValue->getType()) > HalfValBitSize)
    6245             :     return false;
    6246             : 
    6247             :   // If LValue/HValue is a bitcast instruction, use the EVT before bitcast
    6248             :   // as the input of target query.
    6249          69 :   auto *LBC = dyn_cast<BitCastInst>(LValue);
    6250          69 :   auto *HBC = dyn_cast<BitCastInst>(HValue);
    6251           1 :   EVT LowTy = LBC ? EVT::getEVT(LBC->getOperand(0)->getType())
    6252          70 :                   : EVT::getEVT(LValue->getType());
    6253           6 :   EVT HighTy = HBC ? EVT::getEVT(HBC->getOperand(0)->getType())
    6254          75 :                    : EVT::getEVT(HValue->getType());
    6255          69 :   if (!ForceSplitStore && !TLI.isMultiStoresCheaperThanBitsMerge(LowTy, HighTy))
    6256             :     return false;
    6257             : 
    6258             :   // Start to split store.
    6259          15 :   IRBuilder<> Builder(SI.getContext());
    6260          15 :   Builder.SetInsertPoint(&SI);
    6261             : 
    6262             :   // If LValue/HValue is a bitcast in another BB, create a new one in current
    6263             :   // BB so it may be merged with the splitted stores by dag combiner.
    6264          15 :   if (LBC && LBC->getParent() != SI.getParent())
    6265           0 :     LValue = Builder.CreateBitCast(LBC->getOperand(0), LBC->getType());
    6266          15 :   if (HBC && HBC->getParent() != SI.getParent())
    6267           6 :     HValue = Builder.CreateBitCast(HBC->getOperand(0), HBC->getType());
    6268             : 
    6269          30 :   bool IsLE = SI.getModule()->getDataLayout().isLittleEndian();
    6270          30 :   auto CreateSplitStore = [&](Value *V, bool Upper) {
    6271         150 :     V = Builder.CreateZExtOrBitCast(V, SplitStoreType);
    6272          30 :     Value *Addr = Builder.CreateBitCast(
    6273         105 :         SI.getOperand(1),
    6274          60 :         SplitStoreType->getPointerTo(SI.getPointerAddressSpace()));
    6275          30 :     if ((IsLE && Upper) || (!IsLE && !Upper))
    6276          15 :       Addr = Builder.CreateGEP(
    6277             :           SplitStoreType, Addr,
    6278          15 :           ConstantInt::get(Type::getInt32Ty(SI.getContext()), 1));
    6279          45 :     Builder.CreateAlignedStore(
    6280             :         V, Addr, Upper ? SI.getAlignment() / 2 : SI.getAlignment());
    6281          45 :   };
    6282             : 
    6283          15 :   CreateSplitStore(LValue, false);
    6284          15 :   CreateSplitStore(HValue, true);
    6285             : 
    6286             :   // Delete the old store.
    6287          15 :   SI.eraseFromParent();
    6288             :   return true;
    6289             : }
    6290             : 
    6291             : // Return true if the GEP has two operands, the first operand is of a sequential
    6292             : // type, and the second operand is a constant.
    6293          80 : static bool GEPSequentialConstIndexed(GetElementPtrInst *GEP) {
    6294          80 :   gep_type_iterator I = gep_type_begin(*GEP);
    6295          21 :   return GEP->getNumOperands() == 2 &&
    6296         101 :       I.isSequential() &&
    6297          80 :       isa<ConstantInt>(GEP->getOperand(1));
    6298             : }
    6299             : 
    6300             : // Try unmerging GEPs to reduce liveness interference (register pressure) across
    6301             : // IndirectBr edges. Since IndirectBr edges tend to touch on many blocks,
    6302             : // reducing liveness interference across those edges benefits global register
    6303             : // allocation. Currently handles only certain cases.
    6304             : //
    6305             : // For example, unmerge %GEPI and %UGEPI as below.
    6306             : //
    6307             : // ---------- BEFORE ----------
    6308             : // SrcBlock:
    6309             : //   ...
    6310             : //   %GEPIOp = ...
    6311             : //   ...
    6312             : //   %GEPI = gep %GEPIOp, Idx
    6313             : //   ...
    6314             : //   indirectbr ... [ label %DstB0, label %DstB1, ... label %DstBi ... ]
    6315             : //   (* %GEPI is alive on the indirectbr edges due to other uses ahead)
    6316             : //   (* %GEPIOp is alive on the indirectbr edges only because of it's used by
    6317             : //   %UGEPI)
    6318             : //
    6319             : // DstB0: ... (there may be a gep similar to %UGEPI to be unmerged)
    6320             : // DstB1: ... (there may be a gep similar to %UGEPI to be unmerged)
    6321             : // ...
    6322             : //
    6323             : // DstBi:
    6324             : //   ...
    6325             : //   %UGEPI = gep %GEPIOp, UIdx
    6326             : // ...
    6327             : // ---------------------------
    6328             : //
    6329             : // ---------- AFTER ----------
    6330             : // SrcBlock:
    6331             : //   ... (same as above)
    6332             : //    (* %GEPI is still alive on the indirectbr edges)
    6333             : //    (* %GEPIOp is no longer alive on the indirectbr edges as a result of the
    6334             : //    unmerging)
    6335             : // ...
    6336             : //
    6337             : // DstBi:
    6338             : //   ...
    6339             : //   %UGEPI = gep %GEPI, (UIdx-Idx)
    6340             : //   ...
    6341             : // ---------------------------
    6342             : //
    6343             : // The register pressure on the IndirectBr edges is reduced because %GEPIOp is
    6344             : // no longer alive on them.
    6345             : //
    6346             : // We try to unmerge GEPs here in CodGenPrepare, as opposed to limiting merging
    6347             : // of GEPs in the first place in InstCombiner::visitGetElementPtrInst() so as
    6348             : // not to disable further simplications and optimizations as a result of GEP
    6349             : // merging.
    6350             : //
    6351             : // Note this unmerging may increase the length of the data flow critical path
    6352             : // (the path from %GEPIOp to %UGEPI would go through %GEPI), which is a tradeoff
    6353             : // between the register pressure and the length of data-flow critical
    6354             : // path. Restricting this to the uncommon IndirectBr case would minimize the
    6355             : // impact of potentially longer critical path, if any, and the impact on compile
    6356             : // time.
    6357      162016 : static bool tryUnmergingGEPsAcrossIndirectBr(GetElementPtrInst *GEPI,
    6358             :                                              const TargetTransformInfo *TTI) {
    6359      162016 :   BasicBlock *SrcBlock = GEPI->getParent();
    6360             :   // Check that SrcBlock ends with an IndirectBr. If not, give up. The common
    6361             :   // (non-IndirectBr) cases exit early here.
    6362      162016 :   if (!isa<IndirectBrInst>(SrcBlock->getTerminator()))
    6363             :     return false;
    6364             :   // Check that GEPI is a simple gep with a single constant index.
    6365          77 :   if (!GEPSequentialConstIndexed(GEPI))
    6366             :     return false;
    6367             :   ConstantInt *GEPIIdx = cast<ConstantInt>(GEPI->getOperand(1));
    6368             :   // Check that GEPI is a cheap one.
    6369           8 :   if (TTI->getIntImmCost(GEPIIdx->getValue(), GEPIIdx->getType())
    6370             :       > TargetTransformInfo::TCC_Basic)
    6371             :     return false;
    6372             :   Value *GEPIOp = GEPI->getOperand(0);
    6373             :   // Check that GEPIOp is an instruction that's also defined in SrcBlock.
    6374           8 :   if (!isa<Instruction>(GEPIOp))
    6375             :     return false;
    6376             :   auto *GEPIOpI = cast<Instruction>(GEPIOp);
    6377           8 :   if (GEPIOpI->getParent() != SrcBlock)
    6378             :     return false;
    6379             :   // Check that GEP is used outside the block, meaning it's alive on the
    6380             :   // IndirectBr edge(s).
    6381           8 :   if (find_if(GEPI->users(), [&](User *Usr) {
    6382             :         if (auto *I = dyn_cast<Instruction>(Usr)) {
    6383           9 :           if (I->getParent() != SrcBlock) {
    6384             :             return true;
    6385             :           }
    6386             :         }
    6387             :         return false;
    6388             :       }) == GEPI->users().end())
    6389             :     return false;
    6390             :   // The second elements of the GEP chains to be unmerged.
    6391             :   std::vector<GetElementPtrInst *> UGEPIs;
    6392             :   // Check each user of GEPIOp to check if unmerging would make GEPIOp not alive
    6393             :   // on IndirectBr edges.
    6394          21 :   for (User *Usr : GEPIOp->users()) {
    6395          27 :     if (Usr == GEPI) continue;
    6396             :     // Check if Usr is an Instruction. If not, give up.
    6397           9 :     if (!isa<Instruction>(Usr))
    6398           0 :       return false;
    6399             :     auto *UI = cast<Instruction>(Usr);
    6400             :     // Check if Usr in the same block as GEPIOp, which is fine, skip.
    6401           9 :     if (UI->getParent() == SrcBlock)
    6402           6 :       continue;
    6403             :     // Check if Usr is a GEP. If not, give up.
    6404             :     if (!isa<GetElementPtrInst>(Usr))
    6405             :       return false;
    6406           3 :     auto *UGEPI = cast<GetElementPtrInst>(Usr);
    6407             :     // Check if UGEPI is a simple gep with a single constant index and GEPIOp is
    6408             :     // the pointer operand to it. If so, record it in the vector. If not, give
    6409             :     // up.
    6410           3 :     if (!GEPSequentialConstIndexed(UGEPI))
    6411             :       return false;
    6412           6 :     if (UGEPI->getOperand(0) != GEPIOp)
    6413             :       return false;
    6414           3 :     if (GEPIIdx->getType() !=
    6415             :         cast<ConstantInt>(UGEPI->getOperand(1))->getType())
    6416             :       return false;
    6417             :     ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
    6418           3 :     if (TTI->getIntImmCost(UGEPIIdx->getValue(), UGEPIIdx->getType())
    6419             :         > TargetTransformInfo::TCC_Basic)
    6420             :       return false;
    6421           3 :     UGEPIs.push_back(UGEPI);
    6422             :   }
    6423          12 :   if (UGEPIs.size() == 0)
    6424             :     return false;
    6425             :   // Check the materializing cost of (Uidx-Idx).
    6426           4 :   for (GetElementPtrInst *UGEPI : UGEPIs) {
    6427             :     ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
    6428           3 :     APInt NewIdx = UGEPIIdx->getValue() - GEPIIdx->getValue();
    6429           3 :     unsigned ImmCost = TTI->getIntImmCost(NewIdx, GEPIIdx->getType());
    6430           3 :     if (ImmCost > TargetTransformInfo::TCC_Basic)
    6431             :       return false;
    6432             :   }
    6433             :   // Now unmerge between GEPI and UGEPIs.
    6434           4 :   for (GetElementPtrInst *UGEPI : UGEPIs) {
    6435             :     UGEPI->setOperand(0, GEPI);
    6436             :     ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
    6437             :     Constant *NewUGEPIIdx =
    6438             :         ConstantInt::get(GEPIIdx->getType(),
    6439           9 :                          UGEPIIdx->getValue() - GEPIIdx->getValue());
    6440             :     UGEPI->setOperand(1, NewUGEPIIdx);
    6441             :     // If GEPI is not inbounds but UGEPI is inbounds, change UGEPI to not
    6442             :     // inbounds to avoid UB.
    6443           3 :     if (!GEPI->isInBounds()) {
    6444           3 :       UGEPI->setIsInBounds(false);
    6445             :     }
    6446             :   }
    6447             :   // After unmerging, verify that GEPIOp is actually only used in SrcBlock (not
    6448             :   // alive on IndirectBr edges).
    6449             :   assert(find_if(GEPIOp->users(), [&](User *Usr) {
    6450             :         return cast<Instruction>(Usr)->getParent() != SrcBlock;
    6451             :       }) == GEPIOp->users().end() && "GEPIOp is used outside SrcBlock");
    6452             :   return true;
    6453             : }
    6454             : 
    6455     3631001 : bool CodeGenPrepare::optimizeInst(Instruction *I, bool &ModifiedDT) {
    6456             :   // Bail out if we inserted the instruction to prevent optimizations from
    6457             :   // stepping on each other's toes.
    6458     3631001 :   if (InsertedInsts.count(I))
    6459             :     return false;
    6460             : 
    6461     3630324 :   if (PHINode *P = dyn_cast<PHINode>(I)) {
    6462             :     // It is possible for very late stage optimizations (such as SimplifyCFG)
    6463             :     // to introduce PHI nodes too late to be cleaned up.  If we detect such a
    6464             :     // trivial PHI, go ahead and zap it here.
    6465      166862 :     if (Value *V = SimplifyInstruction(P, {*DL, TLInfo})) {
    6466         442 :       P->replaceAllUsesWith(V);
    6467         442 :       P->eraseFromParent();
    6468             :       ++NumPHIsElim;
    6469         442 :       return true;
    6470             :     }
    6471             :     return false;
    6472             :   }
    6473             : 
    6474             :   if (CastInst *CI = dyn_cast<CastInst>(I)) {
    6475             :     // If the source of the cast is a constant, then this should have
    6476             :     // already been constant folded.  The only reason NOT to constant fold
    6477             :     // it is if something (e.g. LSR) was careful to place the constant
    6478             :     // evaluation in a block other than then one that uses it (e.g. to hoist
    6479             :     // the address of globals out of a loop).  If this is the case, we don't
    6480             :     // want to forward-subst the cast.
    6481      316863 :     if (isa<Constant>(CI->getOperand(0)))
    6482             :       return false;
    6483             : 
    6484      313632 :     if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
    6485             :       return true;
    6486             : 
    6487      583180 :     if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
    6488             :       /// Sink a zext or sext into its user blocks if the target type doesn't
    6489             :       /// fit in one register
    6490       71286 :       if (TLI &&
    6491       71284 :           TLI->getTypeAction(CI->getContext(),
    6492       35642 :                              TLI->getValueType(*DL, CI->getType())) ==
    6493             :               TargetLowering::TypeExpandInteger) {
    6494        3488 :         return SinkCast(CI);
    6495             :       } else {
    6496       32156 :         bool MadeChange = optimizeExt(I);
    6497       32156 :         return MadeChange | optimizeExtUses(I);
    6498             :       }
    6499             :     }
    6500             :     return false;
    6501             :   }
    6502             : 
    6503             :   if (CmpInst *CI = dyn_cast<CmpInst>(I))
    6504      103327 :     if (!TLI || !TLI->hasMultipleConditionRegisters())
    6505       96373 :       return OptimizeCmpExpression(CI, TLI);
    6506             : 
    6507             :   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
    6508      533641 :     LI->setMetadata(LLVMContext::MD_invariant_group, nullptr);
    6509      533641 :     if (TLI) {
    6510      533616 :       bool Modified = optimizeLoadExt(LI);
    6511             :       unsigned AS = LI->getPointerAddressSpace();
    6512     1067232 :       Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
    6513      533616 :       return Modified;
    6514             :     }
    6515             :     return false;
    6516             :   }
    6517             : 
    6518             :   if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
    6519      547132 :     if (TLI && splitMergedValStore(*SI, *DL, *TLI))
    6520             :       return true;
    6521      547117 :     SI->setMetadata(LLVMContext::MD_invariant_group, nullptr);
    6522      547117 :     if (TLI) {
    6523             :       unsigned AS = SI->getPointerAddressSpace();
    6524      547100 :       return optimizeMemoryInst(I, SI->getOperand(1),
    6525      547100 :                                 SI->getOperand(0)->getType(), AS);
    6526             :     }
    6527             :     return false;
    6528             :   }
    6529             : 
    6530             :   if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
    6531             :       unsigned AS = RMW->getPointerAddressSpace();
    6532        4757 :       return optimizeMemoryInst(I, RMW->getPointerOperand(),
    6533        4757 :                                 RMW->getType(), AS);
    6534             :   }
    6535             : 
    6536             :   if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(I)) {
    6537             :       unsigned AS = CmpX->getPointerAddressSpace();
    6538        1334 :       return optimizeMemoryInst(I, CmpX->getPointerOperand(),
    6539        1334 :                                 CmpX->getCompareOperand()->getType(), AS);
    6540             :   }
    6541             : 
    6542             :   BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
    6543             : 
    6544      557032 :   if (BinOp && (BinOp->getOpcode() == Instruction::And) &&
    6545       18764 :       EnableAndCmpSinking && TLI)
    6546       18764 :     return sinkAndCmp0Expression(BinOp, *TLI, InsertedInsts);
    6547             : 
    6548     2547533 :   if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
    6549             :                 BinOp->getOpcode() == Instruction::LShr)) {
    6550             :     ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
    6551       13305 :     if (TLI && CI && TLI->hasExtractBitsInsn())
    6552        1146 :       return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
    6553             : 
    6554             :     return false;
    6555             :   }
    6556             : 
    6557             :   if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
    6558      177936 :     if (GEPI->hasAllZeroIndices()) {
    6559             :       /// The GEP operand must be a pointer, so must its result -> BitCast
    6560       15920 :       Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
    6561       31840 :                                         GEPI->getName(), GEPI);
    6562       15920 :       NC->setDebugLoc(GEPI->getDebugLoc());
    6563       15920 :       GEPI->replaceAllUsesWith(NC);
    6564       15920 :       GEPI->eraseFromParent();
    6565             :       ++NumGEPsElim;
    6566       15920 :       optimizeInst(NC, ModifiedDT);
    6567       15920 :       return true;
    6568             :     }
    6569      162016 :     if (tryUnmergingGEPsAcrossIndirectBr(GEPI, TTI)) {
    6570             :       return true;
    6571             :     }
    6572      162015 :     return false;
    6573             :   }
    6574             : 
    6575             :   if (CallInst *CI = dyn_cast<CallInst>(I))
    6576      494910 :     return optimizeCallInst(CI, ModifiedDT);
    6577             : 
    6578             :   if (SelectInst *SI = dyn_cast<SelectInst>(I))
    6579       79330 :     return optimizeSelectInst(SI);
    6580             : 
    6581             :   if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
    6582       44889 :     return optimizeShuffleVectorInst(SVI);
    6583             : 
    6584             :   if (auto *Switch = dyn_cast<SwitchInst>(I))
    6585        1405 :     return optimizeSwitchInst(Switch);
    6586             : 
    6587     1216254 :   if (isa<ExtractElementInst>(I))
    6588       25304 :     return optimizeExtractElementInst(I);
    6589             : 
    6590             :   return false;
    6591             : }
    6592             : 
    6593             : /// Given an OR instruction, check to see if this is a bitreverse
    6594             : /// idiom. If so, insert the new intrinsic and return true.
    6595     3648892 : static bool makeBitReverse(Instruction &I, const DataLayout &DL,
    6596             :                            const TargetLowering &TLI) {
    6597     7297784 :   if (!I.getType()->isIntegerTy() ||
    6598      953850 :       !TLI.isOperationLegalOrCustom(ISD::BITREVERSE,
    6599             :                                     TLI.getValueType(DL, I.getType(), true)))
    6600             :     return false;
    6601             : 
    6602             :   SmallVector<Instruction*, 4> Insts;
    6603       54300 :   if (!recognizeBSwapOrBitReverseIdiom(&I, false, true, Insts))
    6604             :     return false;
    6605           6 :   Instruction *LastInst = Insts.back();
    6606           6 :   I.replaceAllUsesWith(LastInst);
    6607           6 :   RecursivelyDeleteTriviallyDeadInstructions(&I);
    6608           6 :   return true;
    6609             : }
    6610             : 
    6611             : // In this pass we look for GEP and cast instructions that are used
    6612             : // across basic blocks and rewrite them to improve basic-block-at-a-time
    6613             : // selection.
    6614      469435 : bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool &ModifiedDT) {
    6615             :   SunkAddrs.clear();
    6616             :   bool MadeChange = false;
    6617             : 
    6618      469435 :   CurInstIterator = BB.begin();
    6619     4084485 :   while (CurInstIterator != BB.end()) {
    6620     3615081 :     MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
    6621     3615081 :     if (ModifiedDT)
    6622             :       return true;
    6623             :   }
    6624             : 
    6625             :   bool MadeBitReverse = true;
    6626      938652 :   while (TLI && MadeBitReverse) {
    6627             :     MadeBitReverse = false;
    6628     4118134 :     for (auto &I : reverse(BB)) {
    6629     3648892 :       if (makeBitReverse(I, *DL, *TLI)) {
    6630             :         MadeBitReverse = MadeChange = true;
    6631           6 :         ModifiedDT = true;
    6632           6 :         break;
    6633             :       }
    6634             :     }
    6635             :   }
    6636      469404 :   MadeChange |= dupRetToEnableTailCallOpts(&BB);
    6637             : 
    6638      469404 :   return MadeChange;
    6639             : }
    6640             : 
    6641             : // llvm.dbg.value is far away from the value then iSel may not be able
    6642             : // handle it properly. iSel will drop llvm.dbg.value if it can not
    6643             : // find a node corresponding to the value.
    6644      182088 : bool CodeGenPrepare::placeDbgValues(Function &F) {
    6645             :   bool MadeChange = false;
    6646      513728 :   for (BasicBlock &BB : F) {
    6647             :     Instruction *PrevNonDbgInst = nullptr;
    6648     2497825 :     for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
    6649             :       Instruction *Insn = &*BI++;
    6650             :       DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
    6651             :       // Leave dbg.values that refer to an alloca alone. These
    6652             :       // intrinsics describe the address of a variable (= the alloca)
    6653             :       // being taken.  They should not be moved next to the alloca
    6654             :       // (and to the beginning of the scope), but rather stay close to
    6655             :       // where said address is used.
    6656     2173778 :       if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
    6657             :         PrevNonDbgInst = Insn;
    6658             :         continue;
    6659             :       }
    6660             : 
    6661             :       Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
    6662       42390 :       if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
    6663             :         // If VI is a phi in a block with an EHPad terminator, we can't insert
    6664             :         // after it.
    6665        9505 :         if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
    6666             :           continue;
    6667             :         LLVM_DEBUG(dbgs() << "Moving Debug Value before :\n"
    6668             :                           << *DVI << ' ' << *VI);
    6669        9503 :         DVI->removeFromParent();
    6670        9503 :         if (isa<PHINode>(VI))
    6671        5844 :           DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
    6672             :         else
    6673        6581 :           DVI->insertAfter(VI);
    6674             :         MadeChange = true;
    6675             :         ++NumDbgValueMoved;
    6676             :       }
    6677             :     }
    6678             :   }
    6679      182088 :   return MadeChange;
    6680             : }
    6681             : 
    6682             : /// Scale down both weights to fit into uint32_t.
    6683             : static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
    6684             :   uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
    6685             :   uint32_t Scale = (NewMax / std::numeric_limits<uint32_t>::max()) + 1;
    6686             :   NewTrue = NewTrue / Scale;
    6687             :   NewFalse = NewFalse / Scale;
    6688             : }
    6689             : 
    6690             : /// Some targets prefer to split a conditional branch like:
    6691             : /// \code
    6692             : ///   %0 = icmp ne i32 %a, 0
    6693             : ///   %1 = icmp ne i32 %b, 0
    6694             : ///   %or.cond = or i1 %0, %1
    6695             : ///   br i1 %or.cond, label %TrueBB, label %FalseBB
    6696             : /// \endcode
    6697             : /// into multiple branch instructions like:
    6698             : /// \code
    6699             : ///   bb1:
    6700             : ///     %0 = icmp ne i32 %a, 0
    6701             : ///     br i1 %0, label %TrueBB, label %bb2
    6702             : ///   bb2:
    6703             : ///     %1 = icmp ne i32 %b, 0
    6704             : ///     br i1 %1, label %TrueBB, label %FalseBB
    6705             : /// \endcode
    6706             : /// This usually allows instruction selection to do even further optimizations
    6707             : /// and combine the compare with the branch instruction. Currently this is
    6708             : /// applied for targets which have "cheap" jump instructions.
    6709             : ///
    6710             : /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
    6711             : ///
    6712      182025 : bool CodeGenPrepare::splitBranchCondition(Function &F) {
    6713      182025 :   if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
    6714             :     return false;
    6715             : 
    6716             :   bool MadeChange = false;
    6717       14907 :   for (auto &BB : F) {
    6718             :     // Does this BB end with the following?
    6719             :     //   %cond1 = icmp|fcmp|binary instruction ...
    6720             :     //   %cond2 = icmp|fcmp|binary instruction ...
    6721             :     //   %cond.or = or|and i1 %cond1, cond2
    6722             :     //   br i1 %cond.or label %dest1, label %dest2"
    6723             :     BinaryOperator *LogicOp;
    6724             :     BasicBlock *TBB, *FBB;
    6725        7641 :     if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
    6726       15274 :       continue;
    6727             : 
    6728             :     auto *Br1 = cast<BranchInst>(BB.getTerminator());
    6729           9 :     if (Br1->getMetadata(LLVMContext::MD_unpredictable))
    6730           2 :       continue;
    6731             : 
    6732             :     unsigned Opc;
    6733             :     Value *Cond1, *Cond2;
    6734           6 :     if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
    6735             :                              m_OneUse(m_Value(Cond2)))))
    6736             :       Opc = Instruction::And;
    6737           2 :     else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
    6738             :                                  m_OneUse(m_Value(Cond2)))))
    6739             :       Opc = Instruction::Or;
    6740             :     else
    6741           0 :       continue;
    6742             : 
    6743           3 :     if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
    6744           3 :         !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp()))   )
    6745           0 :       continue;
    6746             : 
    6747             :     LLVM_DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
    6748             : 
    6749             :     // Create a new BB.
    6750             :     auto TmpBB =
    6751           9 :         BasicBlock::Create(BB.getContext(), BB.getName() + ".cond.split",
    6752             :                            BB.getParent(), BB.getNextNode());
    6753             : 
    6754             :     // Update original basic block by using the first condition directly by the
    6755             :     // branch instruction and removing the no longer needed and/or instruction.
    6756           3 :     Br1->setCondition(Cond1);
    6757           3 :     LogicOp->eraseFromParent();
    6758             : 
    6759             :     // Depending on the conditon we have to either replace the true or the false
    6760             :     // successor of the original branch instruction.
    6761           3 :     if (Opc == Instruction::And)
    6762             :       Br1->setSuccessor(0, TmpBB);
    6763             :     else
    6764             :       Br1->setSuccessor(1, TmpBB);
    6765             : 
    6766             :     // Fill in the new basic block.
    6767           6 :     auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
    6768           3 :     if (auto *I = dyn_cast<Instruction>(Cond2)) {
    6769           3 :       I->removeFromParent();
    6770           3 :       I->insertBefore(Br2);
    6771             :     }
    6772             : 
    6773             :     // Update PHI nodes in both successors. The original BB needs to be
    6774             :     // replaced in one successor's PHI nodes, because the branch comes now from
    6775             :     // the newly generated BB (NewBB). In the other successor we need to add one
    6776             :     // incoming edge to the PHI nodes, because both branch instructions target
    6777             :     // now the same successor. Depending on the original branch condition
    6778             :     // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
    6779             :     // we perform the correct update for the PHI nodes.
    6780             :     // This doesn't change the successor order of the just created branch
    6781             :     // instruction (or any other instruction).
    6782           3 :     if (Opc == Instruction::Or)
    6783             :       std::swap(TBB, FBB);
    6784             : 
    6785             :     // Replace the old BB with the new BB.
    6786           3 :     for (PHINode &PN : TBB->phis()) {
    6787             :       int i;
    6788           0 :       while ((i = PN.getBasicBlockIndex(&BB)) >= 0)
    6789           0 :         PN.setIncomingBlock(i, TmpBB);
    6790             :     }
    6791             : 
    6792             :     // Add another incoming edge form the new BB.
    6793           3 :     for (PHINode &PN : FBB->phis()) {
    6794           0 :       auto *Val = PN.getIncomingValueForBlock(&BB);
    6795           0 :       PN.addIncoming(Val, TmpBB);
    6796             :     }
    6797             : 
    6798             :     // Update the branch weights (from SelectionDAGBuilder::
    6799             :     // FindMergedConditions).
    6800           3 :     if (Opc == Instruction::Or) {
    6801             :       // Codegen X | Y as:
    6802             :       // BB1:
    6803             :       //   jmp_if_X TBB
    6804             :       //   jmp TmpBB
    6805             :       // TmpBB:
    6806             :       //   jmp_if_Y TBB
    6807             :       //   jmp FBB
    6808             :       //
    6809             : 
    6810             :       // We have flexibility in setting Prob for BB1 and Prob for NewBB.
    6811             :       // The requirement is that
    6812             :       //   TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
    6813             :       //     = TrueProb for orignal BB.
    6814             :       // Assuming the orignal weights are A and B, one choice is to set BB1's
    6815             :       // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
    6816             :       // assumes that
    6817             :       //   TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
    6818             :       // Another choice is to assume TrueProb for BB1 equals to TrueProb for
    6819             :       // TmpBB, but the math is more complicated.
    6820             :       uint64_t TrueWeight, FalseWeight;
    6821           1 :       if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) {
    6822             :         uint64_t NewTrueWeight = TrueWeight;
    6823             :         uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
    6824             :         scaleWeights(NewTrueWeight, NewFalseWeight);
    6825           2 :         Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
    6826             :                          .createBranchWeights(TrueWeight, FalseWeight));
    6827             : 
    6828             :         NewTrueWeight = TrueWeight;
    6829             :         NewFalseWeight = 2 * FalseWeight;
    6830             :         scaleWeights(NewTrueWeight, NewFalseWeight);
    6831           2 :         Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
    6832             :                          .createBranchWeights(TrueWeight, FalseWeight));
    6833             :       }
    6834             :     } else {
    6835             :       // Codegen X & Y as:
    6836             :       // BB1:
    6837             :       //   jmp_if_X TmpBB
    6838             :       //   jmp FBB
    6839             :       // TmpBB:
    6840             :       //   jmp_if_Y TBB
    6841             :       //   jmp FBB
    6842             :       //
    6843             :       //  This requires creation of TmpBB after CurBB.
    6844             : 
    6845             :       // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
    6846             :       // The requirement is that
    6847             :       //   FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
    6848             :       //     = FalseProb for orignal BB.
    6849             :       // Assuming the orignal weights are A and B, one choice is to set BB1's
    6850             :       // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
    6851             :       // assumes that
    6852             :       //   FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
    6853             :       uint64_t TrueWeight, FalseWeight;
    6854           2 :       if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) {
    6855             :         uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
    6856             :         uint64_t NewFalseWeight = FalseWeight;
    6857             :         scaleWeights(NewTrueWeight, NewFalseWeight);
    6858           2 :         Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
    6859             :                          .createBranchWeights(TrueWeight, FalseWeight));
    6860             : 
    6861             :         NewTrueWeight = 2 * TrueWeight;
    6862             :         NewFalseWeight = FalseWeight;
    6863             :         scaleWeights(NewTrueWeight, NewFalseWeight);
    6864           2 :         Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
    6865             :                          .createBranchWeights(TrueWeight, FalseWeight));
    6866             :       }
    6867             :     }
    6868             : 
    6869             :     // Note: No point in getting fancy here, since the DT info is never
    6870             :     // available to CodeGenPrepare.
    6871           3 :     ModifiedDT = true;
    6872             : 
    6873             :     MadeChange = true;
    6874             : 
    6875             :     LLVM_DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
    6876             :                TmpBB->dump());
    6877             :   }
    6878             :   return MadeChange;
    6879      303507 : }

Generated by: LCOV version 1.13