1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/MemoryLocation.h"
25 #include "llvm/Analysis/TargetLibraryInfo.h"
26 #include "llvm/Analysis/TargetTransformInfo.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/IR/CallSite.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/Dominators.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/IR/GetElementPtrTypeIterator.h"
35 #include "llvm/IR/IRBuilder.h"
36 #include "llvm/IR/InlineAsm.h"
37 #include "llvm/IR/InstIterator.h"
38 #include "llvm/IR/InstrTypes.h"
39 #include "llvm/IR/Instructions.h"
40 #include "llvm/IR/IntrinsicInst.h"
41 #include "llvm/IR/MDBuilder.h"
42 #include "llvm/IR/NoFolder.h"
43 #include "llvm/IR/PatternMatch.h"
44 #include "llvm/IR/Statepoint.h"
45 #include "llvm/IR/ValueHandle.h"
46 #include "llvm/IR/ValueMap.h"
47 #include "llvm/Pass.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/raw_ostream.h"
51 #include "llvm/Target/TargetLowering.h"
52 #include "llvm/Target/TargetSubtargetInfo.h"
53 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
54 #include "llvm/Transforms/Utils/BuildLibCalls.h"
55 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
56 #include "llvm/Transforms/Utils/Local.h"
57 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
59 using namespace llvm::PatternMatch;
61 #define DEBUG_TYPE "codegenprepare"
63 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
64 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
65 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
66 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
68 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
70 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
71 "computations were sunk");
72 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
73 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
74 STATISTIC(NumAndsAdded,
75 "Number of and mask instructions added to form ext loads");
76 STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
77 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
78 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
79 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
80 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
81 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
83 static cl::opt<bool> DisableBranchOpts(
84 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
85 cl::desc("Disable branch optimizations in CodeGenPrepare"));
88 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
89 cl::desc("Disable GC optimizations in CodeGenPrepare"));
91 static cl::opt<bool> DisableSelectToBranch(
92 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
93 cl::desc("Disable select to branch conversion."));
95 static cl::opt<bool> AddrSinkUsingGEPs(
96 "addr-sink-using-gep", cl::Hidden, cl::init(false),
97 cl::desc("Address sinking in CGP using GEPs."));
99 static cl::opt<bool> EnableAndCmpSinking(
100 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
101 cl::desc("Enable sinkinig and/cmp into branches."));
103 static cl::opt<bool> DisableStoreExtract(
104 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
105 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
107 static cl::opt<bool> StressStoreExtract(
108 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
109 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
111 static cl::opt<bool> DisableExtLdPromotion(
112 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
113 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
116 static cl::opt<bool> StressExtLdPromotion(
117 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
118 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
119 "optimization in CodeGenPrepare"));
122 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
123 typedef PointerIntPair<Type *, 1, bool> TypeIsSExt;
124 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
125 class TypePromotionTransaction;
127 class CodeGenPrepare : public FunctionPass {
128 const TargetMachine *TM;
129 const TargetLowering *TLI;
130 const TargetTransformInfo *TTI;
131 const TargetLibraryInfo *TLInfo;
133 /// As we scan instructions optimizing them, this is the next instruction
134 /// to optimize. Transforms that can invalidate this should update it.
135 BasicBlock::iterator CurInstIterator;
137 /// Keeps track of non-local addresses that have been sunk into a block.
138 /// This allows us to avoid inserting duplicate code for blocks with
139 /// multiple load/stores of the same address.
140 ValueMap<Value*, Value*> SunkAddrs;
142 /// Keeps track of all instructions inserted for the current function.
143 SetOfInstrs InsertedInsts;
144 /// Keeps track of the type of the related instruction before their
145 /// promotion for the current function.
146 InstrToOrigTy PromotedInsts;
148 /// True if CFG is modified in any way.
151 /// True if optimizing for size.
154 /// DataLayout for the Function being processed.
155 const DataLayout *DL;
157 // XXX-comment:We need DominatorTree to figure out which instruction to
162 static char ID; // Pass identification, replacement for typeid
163 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
164 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr), DL(nullptr),
166 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
168 bool runOnFunction(Function &F) override;
170 const char *getPassName() const override { return "CodeGen Prepare"; }
172 void getAnalysisUsage(AnalysisUsage &AU) const override {
173 AU.addPreserved<DominatorTreeWrapperPass>();
174 AU.addRequired<TargetLibraryInfoWrapperPass>();
175 AU.addRequired<TargetTransformInfoWrapperPass>();
176 AU.addRequired<DominatorTreeWrapperPass>();
180 bool eliminateFallThrough(Function &F);
181 bool eliminateMostlyEmptyBlocks(Function &F);
182 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
183 void eliminateMostlyEmptyBlock(BasicBlock *BB);
184 bool optimizeBlock(BasicBlock &BB, bool& ModifiedDT);
185 bool optimizeInst(Instruction *I, bool& ModifiedDT);
186 bool optimizeMemoryInst(Instruction *I, Value *Addr,
187 Type *AccessTy, unsigned AS);
188 bool optimizeInlineAsmInst(CallInst *CS);
189 bool optimizeCallInst(CallInst *CI, bool& ModifiedDT);
190 bool moveExtToFormExtLoad(Instruction *&I);
191 bool optimizeExtUses(Instruction *I);
192 bool optimizeLoadExt(LoadInst *I);
193 bool optimizeSelectInst(SelectInst *SI);
194 bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
195 bool optimizeSwitchInst(SwitchInst *CI);
196 bool optimizeExtractElementInst(Instruction *Inst);
197 bool dupRetToEnableTailCallOpts(BasicBlock *BB);
198 bool placeDbgValues(Function &F);
199 bool sinkAndCmp(Function &F);
200 bool extLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
202 const SmallVectorImpl<Instruction *> &Exts,
203 unsigned CreatedInstCost);
204 bool splitBranchCondition(Function &F);
205 bool simplifyOffsetableRelocate(Instruction &I);
206 void stripInvariantGroupMetadata(Instruction &I);
210 char CodeGenPrepare::ID = 0;
211 INITIALIZE_TM_PASS_BEGIN(CodeGenPrepare, "codegenprepare",
212 "Optimize for code generation", false, false)
213 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
214 INITIALIZE_TM_PASS_END(CodeGenPrepare, "codegenprepare",
215 "Optimize for code generation", false, false)
217 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
218 return new CodeGenPrepare(TM);
223 bool StoreAddressDependOnValue(StoreInst* SI, Value* DepVal);
224 Value* GetUntaintedAddress(Value* CurrentAddress);
226 // The depth we trace down a variable to look for its dependence set.
227 const unsigned kDependenceDepth = 4;
229 // Recursively looks for variables that 'Val' depends on at the given depth
230 // 'Depth', and adds them in 'DepSet'. If 'InsertOnlyLeafNodes' is true, only
231 // inserts the leaf node values; otherwise, all visited nodes are included in
232 // 'DepSet'. Note that constants will be ignored.
233 template <typename SetType>
234 void recursivelyFindDependence(SetType* DepSet, Value* Val,
235 bool InsertOnlyLeafNodes = false,
236 unsigned Depth = kDependenceDepth) {
237 if (Val == nullptr) {
240 if (!InsertOnlyLeafNodes && !isa<Constant>(Val)) {
244 // Cannot go deeper. Insert the leaf nodes.
245 if (InsertOnlyLeafNodes && !isa<Constant>(Val)) {
251 // Go one step further to explore the dependence of the operands.
252 Instruction* I = nullptr;
253 if ((I = dyn_cast<Instruction>(Val))) {
254 if (isa<LoadInst>(I)) {
255 // A load is considerd the leaf load of the dependence tree. Done.
258 } else if (I->isBinaryOp()) {
259 BinaryOperator* I = dyn_cast<BinaryOperator>(Val);
260 Value *Op0 = I->getOperand(0), *Op1 = I->getOperand(1);
261 recursivelyFindDependence(DepSet, Op0, InsertOnlyLeafNodes, Depth - 1);
262 recursivelyFindDependence(DepSet, Op1, InsertOnlyLeafNodes, Depth - 1);
263 } else if (I->isCast()) {
264 Value* Op0 = I->getOperand(0);
265 recursivelyFindDependence(DepSet, Op0, InsertOnlyLeafNodes, Depth - 1);
266 } else if (I->getOpcode() == Instruction::Select) {
267 Value* Op0 = I->getOperand(0);
268 Value* Op1 = I->getOperand(1);
269 Value* Op2 = I->getOperand(2);
270 recursivelyFindDependence(DepSet, Op0, InsertOnlyLeafNodes, Depth - 1);
271 recursivelyFindDependence(DepSet, Op1, InsertOnlyLeafNodes, Depth - 1);
272 recursivelyFindDependence(DepSet, Op2, InsertOnlyLeafNodes, Depth - 1);
273 } else if (I->getOpcode() == Instruction::GetElementPtr) {
274 for (unsigned i = 0; i < I->getNumOperands(); i++) {
275 recursivelyFindDependence(DepSet, I->getOperand(i), InsertOnlyLeafNodes,
278 } else if (I->getOpcode() == Instruction::Store) {
279 auto* SI = dyn_cast<StoreInst>(Val);
280 recursivelyFindDependence(DepSet, SI->getPointerOperand(),
281 InsertOnlyLeafNodes, Depth - 1);
282 recursivelyFindDependence(DepSet, SI->getValueOperand(),
283 InsertOnlyLeafNodes, Depth - 1);
285 Value* Op0 = nullptr;
286 Value* Op1 = nullptr;
287 switch (I->getOpcode()) {
288 case Instruction::ICmp:
289 case Instruction::FCmp: {
290 Op0 = I->getOperand(0);
291 Op1 = I->getOperand(1);
292 recursivelyFindDependence(DepSet, Op0, InsertOnlyLeafNodes,
294 recursivelyFindDependence(DepSet, Op1, InsertOnlyLeafNodes,
298 case Instruction::PHI: {
299 for (int i = 0; i < I->getNumOperands(); i++) {
300 auto* op = I->getOperand(i);
301 if (DepSet->count(op) == 0) {
302 recursivelyFindDependence(DepSet, I->getOperand(i),
303 InsertOnlyLeafNodes, Depth - 1);
309 // Be conservative. Add it and be done with it.
315 } else if (isa<Constant>(Val)) {
316 // Not interested in constant values. Done.
319 // Be conservative. Add it and be done with it.
325 // Helper function to create a Cast instruction.
326 Value* createCast(IRBuilder<true, NoFolder>& Builder, Value* DepVal,
327 Type* TargetIntegerType) {
328 Instruction::CastOps CastOp = Instruction::BitCast;
329 switch (DepVal->getType()->getTypeID()) {
330 case Type::IntegerTyID: {
331 CastOp = Instruction::SExt;
334 case Type::FloatTyID:
335 case Type::DoubleTyID: {
336 CastOp = Instruction::FPToSI;
339 case Type::PointerTyID: {
340 CastOp = Instruction::PtrToInt;
346 return Builder.CreateCast(CastOp, DepVal, TargetIntegerType);
349 // Given a value, if it's a tainted address, this function returns the
350 // instruction that ORs the "dependence value" with the "original address".
351 // Otherwise, returns nullptr. This instruction is the first OR instruction
352 // where one of its operand is an AND instruction with an operand being 0.
354 // E.g., it returns '%4 = or i32 %3, %2' given 'CurrentAddress' is '%5'.
355 // %0 = load i32, i32* @y, align 4, !tbaa !1
356 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
357 // %1 = sext i1 %cmp to i32
358 // %2 = ptrtoint i32* @x to i32
359 // %3 = and i32 %1, 0
360 // %4 = or i32 %3, %2
361 // %5 = inttoptr i32 %4 to i32*
362 // store i32 1, i32* %5, align 4
363 Instruction* getOrAddress(Value* CurrentAddress) {
364 // Is it a cast from integer to pointer type.
365 Instruction* OrAddress = nullptr;
366 Instruction* AndDep = nullptr;
367 Instruction* CastToInt = nullptr;
368 Value* ActualAddress = nullptr;
369 Constant* ZeroConst = nullptr;
371 const Instruction* CastToPtr = dyn_cast<Instruction>(CurrentAddress);
372 if (CastToPtr && CastToPtr->getOpcode() == Instruction::IntToPtr) {
373 // Is it an OR instruction: %1 = or %and, %actualAddress.
374 if ((OrAddress = dyn_cast<Instruction>(CastToPtr->getOperand(0))) &&
375 OrAddress->getOpcode() == Instruction::Or) {
376 // The first operand should be and AND instruction.
377 AndDep = dyn_cast<Instruction>(OrAddress->getOperand(0));
378 if (AndDep && AndDep->getOpcode() == Instruction::And) {
379 // Also make sure its first operand of the "AND" is 0, or the "AND" is
380 // marked explicitly by "NoInstCombine".
381 if ((ZeroConst = dyn_cast<Constant>(AndDep->getOperand(1))) &&
382 ZeroConst->isNullValue()) {
388 // Looks like it's not been tainted.
392 // Given a value, if it's a tainted address, this function returns the
393 // instruction that taints the "dependence value". Otherwise, returns nullptr.
394 // This instruction is the last AND instruction where one of its operand is 0.
395 // E.g., it returns '%3' given 'CurrentAddress' is '%5'.
396 // %0 = load i32, i32* @y, align 4, !tbaa !1
397 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
398 // %1 = sext i1 %cmp to i32
399 // %2 = ptrtoint i32* @x to i32
400 // %3 = and i32 %1, 0
401 // %4 = or i32 %3, %2
402 // %5 = inttoptr i32 %4 to i32*
403 // store i32 1, i32* %5, align 4
404 Instruction* getAndDependence(Value* CurrentAddress) {
405 // If 'CurrentAddress' is tainted, get the OR instruction.
406 auto* OrAddress = getOrAddress(CurrentAddress);
407 if (OrAddress == nullptr) {
411 // No need to check the operands.
412 auto* AndDepInst = dyn_cast<Instruction>(OrAddress->getOperand(0));
417 // Given a value, if it's a tainted address, this function returns
418 // the "dependence value", which is the first operand in the AND instruction.
419 // E.g., it returns '%1' given 'CurrentAddress' is '%5'.
420 // %0 = load i32, i32* @y, align 4, !tbaa !1
421 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
422 // %1 = sext i1 %cmp to i32
423 // %2 = ptrtoint i32* @x to i32
424 // %3 = and i32 %1, 0
425 // %4 = or i32 %3, %2
426 // %5 = inttoptr i32 %4 to i32*
427 // store i32 1, i32* %5, align 4
428 Value* getDependence(Value* CurrentAddress) {
429 auto* AndInst = getAndDependence(CurrentAddress);
430 if (AndInst == nullptr) {
433 return AndInst->getOperand(0);
436 // Given an address that has been tainted, returns the only condition it depends
437 // on, if any; otherwise, returns nullptr.
438 Value* getConditionDependence(Value* Address) {
439 auto* Dep = getDependence(Address);
440 if (Dep == nullptr) {
441 // 'Address' has not been dependence-tainted.
445 Value* Operand = Dep;
447 auto* Inst = dyn_cast<Instruction>(Operand);
448 if (Inst == nullptr) {
449 // Non-instruction type does not have condition dependence.
452 if (Inst->getOpcode() == Instruction::ICmp) {
455 if (Inst->getNumOperands() != 1) {
458 Operand = Inst->getOperand(0);
464 // Conservatively decides whether the dependence set of 'Val1' includes the
465 // dependence set of 'Val2'. If 'ExpandSecondValue' is false, we do not expand
466 // 'Val2' and use that single value as its dependence set.
467 // If it returns true, it means the dependence set of 'Val1' includes that of
468 // 'Val2'; otherwise, it only means we cannot conclusively decide it.
469 bool dependenceSetInclusion(Value* Val1, Value* Val2,
470 int Val1ExpandLevel = 2 * kDependenceDepth,
471 int Val2ExpandLevel = kDependenceDepth) {
472 typedef SmallSet<Value*, 8> IncludingSet;
473 typedef SmallSet<Value*, 4> IncludedSet;
475 IncludingSet DepSet1;
477 // Look for more depths for the including set.
478 recursivelyFindDependence(&DepSet1, Val1, false /*Insert all visited nodes*/,
480 recursivelyFindDependence(&DepSet2, Val2, true /*Only insert leaf nodes*/,
483 auto set_inclusion = [](IncludingSet FullSet, IncludedSet Subset) {
484 for (auto* Dep : Subset) {
485 if (0 == FullSet.count(Dep)) {
491 bool inclusion = set_inclusion(DepSet1, DepSet2);
492 DEBUG(dbgs() << "[dependenceSetInclusion]: " << inclusion << "\n");
493 DEBUG(dbgs() << "Including set for: " << *Val1 << "\n");
494 DEBUG(for (const auto* Dep : DepSet1) { dbgs() << "\t\t" << *Dep << "\n"; });
495 DEBUG(dbgs() << "Included set for: " << *Val2 << "\n");
496 DEBUG(for (const auto* Dep : DepSet2) { dbgs() << "\t\t" << *Dep << "\n"; });
501 // Recursively iterates through the operands spawned from 'DepVal'. If there
502 // exists a single value that 'DepVal' only depends on, we call that value the
503 // root dependence of 'DepVal' and return it. Otherwise, return 'DepVal'.
504 Value* getRootDependence(Value* DepVal) {
505 SmallSet<Value*, 8> DepSet;
506 for (unsigned depth = kDependenceDepth; depth > 0; --depth) {
507 recursivelyFindDependence(&DepSet, DepVal, true /*Only insert leaf nodes*/,
509 if (DepSet.size() == 1) {
510 return *DepSet.begin();
517 // This function actually taints 'DepVal' to the address to 'SI'. If the
519 // of 'SI' already depends on whatever 'DepVal' depends on, this function
520 // doesn't do anything and returns false. Otherwise, returns true.
522 // This effect forces the store and any stores that comes later to depend on
523 // 'DepVal'. For example, we have a condition "cond", and a store instruction
524 // "s: STORE addr, val". If we want "s" (and any later store) to depend on
525 // "cond", we do the following:
526 // %conv = sext i1 %cond to i32
527 // %addrVal = ptrtoint i32* %addr to i32
528 // %andCond = and i32 conv, 0;
529 // %orAddr = or i32 %andCond, %addrVal;
530 // %NewAddr = inttoptr i32 %orAddr to i32*;
532 // This is a more concrete example:
534 // %0 = load i32, i32* @y, align 4, !tbaa !1
535 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
536 // %1 = sext i1 %cmp to i32
537 // %2 = ptrtoint i32* @x to i32
538 // %3 = and i32 %1, 0
539 // %4 = or i32 %3, %2
540 // %5 = inttoptr i32 %4 to i32*
541 // store i32 1, i32* %5, align 4
542 bool taintStoreAddress(StoreInst* SI, Value* DepVal) {
543 // Set the insertion point right after the 'DepVal'.
544 Instruction* Inst = nullptr;
545 IRBuilder<true, NoFolder> Builder(SI);
546 BasicBlock* BB = SI->getParent();
547 Value* Address = SI->getPointerOperand();
548 Type* TargetIntegerType =
549 IntegerType::get(Address->getContext(),
550 BB->getModule()->getDataLayout().getPointerSizeInBits());
552 // Does SI's address already depends on whatever 'DepVal' depends on?
553 if (StoreAddressDependOnValue(SI, DepVal)) {
557 // Figure out if there's a root variable 'DepVal' depends on. For example, we
558 // can extract "getelementptr inbounds %struct, %struct* %0, i64 0, i32 123"
559 // to be "%struct* %0" since all other operands are constant.
560 auto* RootVal = getRootDependence(DepVal);
561 auto* RootInst = dyn_cast<Instruction>(RootVal);
562 auto* DepValInst = dyn_cast<Instruction>(DepVal);
563 if (RootInst && DepValInst &&
564 RootInst->getParent() == DepValInst->getParent()) {
568 // Is this already a dependence-tainted store?
569 Value* OldDep = getDependence(Address);
571 // The address of 'SI' has already been tainted. Just need to absorb the
572 // DepVal to the existing dependence in the address of SI.
573 Instruction* AndDep = getAndDependence(Address);
574 IRBuilder<true, NoFolder> Builder(AndDep);
575 Value* NewDep = nullptr;
576 if (DepVal->getType() == AndDep->getType()) {
577 NewDep = Builder.CreateAnd(OldDep, DepVal);
579 NewDep = Builder.CreateAnd(
580 OldDep, createCast(Builder, DepVal, TargetIntegerType));
583 auto* NewDepInst = dyn_cast<Instruction>(NewDep);
585 // Use the new AND instruction as the dependence
586 AndDep->setOperand(0, NewDep);
590 // SI's address has not been tainted. Now taint it with 'DepVal'.
591 Value* CastDepToInt = createCast(Builder, DepVal, TargetIntegerType);
592 Value* PtrToIntCast = Builder.CreatePtrToInt(Address, TargetIntegerType);
594 Builder.CreateAnd(CastDepToInt, ConstantInt::get(TargetIntegerType, 0));
595 auto AndInst = dyn_cast<Instruction>(AndDepVal);
596 // XXX-comment: The original IR InstCombiner would change our and instruction
597 // to a select and then the back end optimize the condition out. We attach a
598 // flag to instructions and set it here to inform the InstCombiner to not to
599 // touch this and instruction at all.
600 Value* OrAddr = Builder.CreateOr(AndDepVal, PtrToIntCast);
601 Value* NewAddr = Builder.CreateIntToPtr(OrAddr, Address->getType());
603 DEBUG(dbgs() << "[taintStoreAddress]\n"
604 << "Original store: " << *SI << '\n');
605 SI->setOperand(1, NewAddr);
608 DEBUG(dbgs() << "\tTargetIntegerType: " << *TargetIntegerType << '\n'
609 << "\tCast dependence value to integer: " << *CastDepToInt
611 << "\tCast address to integer: " << *PtrToIntCast << '\n'
612 << "\tAnd dependence value: " << *AndDepVal << '\n'
613 << "\tOr address: " << *OrAddr << '\n'
614 << "\tCast or instruction to address: " << *NewAddr << "\n\n");
619 // Looks for the previous store in the if block --- 'BrBB', which makes the
620 // speculative store 'StoreToHoist' safe.
621 Value* getSpeculativeStoreInPrevBB(StoreInst* StoreToHoist, BasicBlock* BrBB) {
622 assert(StoreToHoist && "StoreToHoist must be a real store");
624 Value* StorePtr = StoreToHoist->getPointerOperand();
626 // Look for a store to the same pointer in BrBB.
627 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(), RE = BrBB->rend();
629 Instruction* CurI = &*RI;
631 StoreInst* SI = dyn_cast<StoreInst>(CurI);
632 // Found the previous store make sure it stores to the same location.
633 // XXX-update: If the previous store's original untainted address are the
634 // same as 'StorePtr', we are also good to hoist the store.
635 if (SI && (SI->getPointerOperand() == StorePtr ||
636 GetUntaintedAddress(SI->getPointerOperand()) == StorePtr)) {
637 // Found the previous store, return its value operand.
643 "We should not reach here since this store is safe to speculate");
646 // XXX-comment: Returns true if it changes the code, false otherwise (the branch
647 // condition already depends on 'DepVal'.
648 bool taintConditionalBranch(BranchInst* BI, Value* DepVal) {
649 assert(BI->isConditional());
650 auto* Cond = BI->getOperand(0);
651 if (dependenceSetInclusion(Cond, DepVal)) {
652 // The dependence/ordering is self-evident.
656 IRBuilder<true, NoFolder> Builder(BI);
658 Builder.CreateAnd(DepVal, ConstantInt::get(DepVal->getType(), 0));
660 Builder.CreateTrunc(AndDep, IntegerType::get(DepVal->getContext(), 1));
661 auto* OrCond = Builder.CreateOr(TruncAndDep, Cond);
662 BI->setOperand(0, OrCond);
665 DEBUG(dbgs() << "\tTainted branch condition:\n" << *BI->getParent());
670 bool ConditionalBranchDependsOnValue(BranchInst* BI, Value* DepVal) {
671 assert(BI->isConditional());
672 auto* Cond = BI->getOperand(0);
673 return dependenceSetInclusion(Cond, DepVal);
676 // XXX-update: For a relaxed load 'LI', find the first immediate atomic store or
677 // the first conditional branch. Returns nullptr if there's no such immediately
678 // following store/branch instructions, which we can only enforce the load with
679 // 'acquire'. 'ChainedBB' contains all the blocks chained together with
680 // unconditional branches from 'BB' to the block with the first store/cond
682 template <typename Vector>
683 Instruction* findFirstStoreCondBranchInst(LoadInst* LI, Vector* ChainedBB) {
684 // In some situations, relaxed loads can be left as is:
685 // 1. The relaxed load is used to calculate the address of the immediate
687 // 2. The relaxed load is used as a condition in the immediate following
688 // condition, and there are no stores in between. This is actually quite
690 // int r1 = x.load(relaxed);
692 // y.store(1, relaxed);
694 // However, in this function, we don't deal with them directly. Instead, we
695 // just find the immediate following store/condition branch and return it.
697 assert(ChainedBB != nullptr && "Chained BB should not be nullptr");
698 auto* BB = LI->getParent();
699 ChainedBB->push_back(BB);
701 auto BBI = BasicBlock::iterator(LI);
704 for (; BBI != BE; BBI++) {
705 auto* Inst = dyn_cast<Instruction>(&*BBI);
706 if (Inst == nullptr) {
709 if (Inst->getOpcode() == Instruction::Store) {
711 } else if (Inst->getOpcode() == Instruction::Br) {
712 auto* BrInst = dyn_cast<BranchInst>(Inst);
713 if (BrInst->isConditional()) {
716 // Reinitialize iterators with the destination of the unconditional
718 BB = BrInst->getSuccessor(0);
719 ChainedBB->push_back(BB);
732 // XXX-comment: Returns whether the code has been changed.
733 bool taintMonotonicLoads(const SmallVector<LoadInst*, 1>& MonotonicLoadInsts) {
734 bool Changed = false;
735 for (auto* LI : MonotonicLoadInsts) {
736 SmallVector<BasicBlock*, 2> ChainedBB;
737 auto* FirstInst = findFirstStoreCondBranchInst(LI, &ChainedBB);
738 if (FirstInst == nullptr) {
739 // We don't seem to be able to taint a following store/conditional branch
740 // instruction. Simply make it acquire.
741 DEBUG(dbgs() << "[RelaxedLoad]: Transformed to acquire load\n"
743 LI->setOrdering(Acquire);
747 // Taint 'FirstInst', which could be a store or a condition branch
749 if (FirstInst->getOpcode() == Instruction::Store) {
750 Changed |= taintStoreAddress(dyn_cast<StoreInst>(FirstInst), LI);
751 } else if (FirstInst->getOpcode() == Instruction::Br) {
752 Changed |= taintConditionalBranch(dyn_cast<BranchInst>(FirstInst), LI);
754 assert(false && "findFirstStoreCondBranchInst() should return a "
755 "store/condition branch instruction");
761 // Inserts a fake conditional branch right after the instruction 'SplitInst',
762 // and the branch condition is 'Condition'. 'SplitInst' will be placed in the
763 // newly created block.
764 void AddFakeConditionalBranch(Instruction* SplitInst, Value* Condition) {
765 auto* BB = SplitInst->getParent();
766 TerminatorInst* ThenTerm = nullptr;
767 TerminatorInst* ElseTerm = nullptr;
768 SplitBlockAndInsertIfThenElse(Condition, SplitInst, &ThenTerm, &ElseTerm);
769 assert(ThenTerm && ElseTerm &&
770 "Then/Else terminators cannot be empty after basic block spliting");
771 auto* ThenBB = ThenTerm->getParent();
772 auto* ElseBB = ElseTerm->getParent();
773 auto* TailBB = ThenBB->getSingleSuccessor();
774 assert(TailBB && "Tail block cannot be empty after basic block spliting");
776 ThenBB->disableCanEliminateBlock();
777 ThenBB->disableCanEliminateBlock();
778 TailBB->disableCanEliminateBlock();
779 ThenBB->setName(BB->getName() + "Then.Fake");
780 ElseBB->setName(BB->getName() + "Else.Fake");
781 DEBUG(dbgs() << "Add fake conditional branch:\n"
783 << *ThenBB << "Else Block:\n"
787 // Returns true if the code is changed, and false otherwise.
788 void TaintRelaxedLoads(Instruction* UsageInst, Instruction* InsertPoint) {
789 // For better performance, we can add a "AND X 0" instruction before the
791 auto* BB = UsageInst->getParent();
792 if (InsertPoint == nullptr) {
793 InsertPoint = UsageInst->getNextNode();
795 // Insert instructions after PHI nodes.
796 while (dyn_cast<PHINode>(InsertPoint)) {
797 InsertPoint = InsertPoint->getNextNode();
799 // First thing is to cast 'UsageInst' to an integer type if necessary.
800 Value* AndTarget = nullptr;
801 Type* TargetIntegerType =
802 IntegerType::get(UsageInst->getContext(),
803 BB->getModule()->getDataLayout().getPointerSizeInBits());
805 // Check whether InsertPoint is a added fake conditional branch.
806 BranchInst* BI = nullptr;
807 if ((BI = dyn_cast<BranchInst>(InsertPoint)) && BI->isConditional()) {
808 auto* Cond = dyn_cast<Instruction>(BI->getOperand(0));
809 if (Cond && Cond->getOpcode() == Instruction::ICmp) {
810 auto* CmpInst = dyn_cast<ICmpInst>(Cond);
811 auto* Op0 = dyn_cast<Instruction>(Cond->getOperand(0));
812 auto* Op1 = dyn_cast<ConstantInt>(Cond->getOperand(1));
814 // %cmp = ICMP_NE %tmp, 0
817 // %tmp1 = And X, NewTaintedVal
818 // %tmp2 = And %tmp1, 0
819 // %cmp = ICMP_NE %tmp2, 0
821 if (CmpInst && CmpInst->getPredicate() == CmpInst::ICMP_NE && Op0 &&
822 Op0->getOpcode() == Instruction::And && Op1 && Op1->isZero()) {
823 auto* Op01 = dyn_cast<ConstantInt>(Op0->getOperand(1));
824 if (Op01 && Op01->isZero()) {
825 // Now we have a previously added fake cond branch.
826 auto* Op00 = Op0->getOperand(0);
827 IRBuilder<true, NoFolder> Builder(CmpInst);
828 if (UsageInst->getType() == TargetIntegerType) {
829 AndTarget = UsageInst;
831 AndTarget = createCast(Builder, UsageInst, TargetIntegerType);
833 AndTarget = Builder.CreateAnd(Op00, AndTarget);
834 auto* AndZero = dyn_cast<Instruction>(Builder.CreateAnd(
835 AndTarget, Constant::getNullValue(AndTarget->getType())));
836 CmpInst->setOperand(0, AndZero);
843 IRBuilder<true, NoFolder> Builder(InsertPoint);
844 if (UsageInst->getType() == TargetIntegerType) {
845 AndTarget = UsageInst;
847 AndTarget = createCast(Builder, UsageInst, TargetIntegerType);
849 auto* AndZero = dyn_cast<Instruction>(
850 Builder.CreateAnd(AndTarget, Constant::getNullValue(AndTarget->getType())));
851 auto* FakeCondition = dyn_cast<Instruction>(Builder.CreateICmp(
852 CmpInst::ICMP_NE, AndZero, Constant::getNullValue(AndTarget->getType())));
853 AddFakeConditionalBranch(FakeCondition->getNextNode(), FakeCondition);
856 // XXX-comment: Finds the appropriate Value derived from an atomic load.
857 // 'ChainedBB' contains all the blocks chained together with unconditional
858 // branches from LI's parent BB to the block with the first store/cond branch.
859 // If we don't find any, it means 'LI' is not used at all (which should not
860 // happen in practice). We can simply set 'LI' to be acquire just to be safe.
861 template <typename Vector>
862 Instruction* findMostRecentDependenceUsage(LoadInst* LI, Instruction* LaterInst,
865 typedef SmallSet<Instruction*, 8> UsageSet;
866 typedef DenseMap<BasicBlock*, std::unique_ptr<UsageSet>> UsageMap;
867 assert(ChainedBB->size() >= 1 && "ChainedBB must have >=1 size");
868 // Mapping from basic block in 'ChainedBB' to the set of dependence usage of
869 // 'LI' in each block.
871 auto* LoadBB = LI->getParent();
872 usage_map[LoadBB] = make_unique<UsageSet>();
873 usage_map[LoadBB]->insert(LI);
875 for (auto* BB : *ChainedBB) {
876 if (usage_map[BB] == nullptr) {
877 usage_map[BB] = make_unique<UsageSet>();
879 auto& usage_set = usage_map[BB];
880 if (usage_set->size() == 0) {
881 // The value has not been used.
884 // Calculate the usage in the current BB first.
885 std::list<Value*> bb_usage_list;
886 std::copy(usage_set->begin(), usage_set->end(),
887 std::back_inserter(bb_usage_list));
888 for (auto list_iter = bb_usage_list.begin();
889 list_iter != bb_usage_list.end(); list_iter++) {
890 auto* val = *list_iter;
891 for (auto* U : val->users()) {
892 Instruction* Inst = nullptr;
893 if (!(Inst = dyn_cast<Instruction>(U))) {
896 assert(Inst && "Usage value must be an instruction");
898 std::find(ChainedBB->begin(), ChainedBB->end(), Inst->getParent());
899 if (iter == ChainedBB->end()) {
900 // Only care about usage within ChainedBB.
903 auto* UsageBB = *iter;
906 if (!usage_set->count(Inst)) {
907 bb_usage_list.push_back(Inst);
908 usage_set->insert(Inst);
912 if (usage_map[UsageBB] == nullptr) {
913 usage_map[UsageBB] = make_unique<UsageSet>();
915 usage_map[UsageBB]->insert(Inst);
921 // Pick one usage that is in LaterInst's block and that dominates 'LaterInst'.
922 auto* LaterBB = LaterInst->getParent();
923 auto& usage_set = usage_map[LaterBB];
924 Instruction* usage_inst = nullptr;
925 for (auto* inst : *usage_set) {
926 if (DT->dominates(inst, LaterInst)) {
932 assert(usage_inst && "The usage instruction in the same block but after the "
933 "later instruction");
937 // XXX-comment: Returns whether the code has been changed.
938 bool AddFakeConditionalBranchAfterMonotonicLoads(
939 SmallSet<LoadInst*, 1>& MonotonicLoadInsts, DominatorTree* DT) {
940 bool Changed = false;
941 while (!MonotonicLoadInsts.empty()) {
942 auto* LI = *MonotonicLoadInsts.begin();
943 MonotonicLoadInsts.erase(LI);
944 SmallVector<BasicBlock*, 2> ChainedBB;
945 auto* FirstInst = findFirstStoreCondBranchInst(LI, &ChainedBB);
946 if (FirstInst != nullptr) {
947 if (FirstInst->getOpcode() == Instruction::Store) {
948 if (StoreAddressDependOnValue(dyn_cast<StoreInst>(FirstInst), LI)) {
951 } else if (FirstInst->getOpcode() == Instruction::Br) {
952 if (ConditionalBranchDependsOnValue(dyn_cast<BranchInst>(FirstInst),
957 dbgs() << "FirstInst=" << *FirstInst << "\n";
958 assert(false && "findFirstStoreCondBranchInst() should return a "
959 "store/condition branch instruction");
963 // We really need to process the relaxed load now.
964 StoreInst* SI = nullptr;;
965 if (FirstInst && (SI = dyn_cast<StoreInst>(FirstInst))) {
966 // For immediately coming stores, taint the address of the store.
967 if (SI->getParent() == LI->getParent() || DT->dominates(LI, SI)) {
968 TaintRelaxedLoads(LI, SI);
972 findMostRecentDependenceUsage(LI, FirstInst, &ChainedBB, DT);
974 LI->setOrdering(Acquire);
977 TaintRelaxedLoads(Inst, SI);
982 // No upcoming branch
984 TaintRelaxedLoads(LI, nullptr);
987 // For immediately coming branch, directly add a fake branch.
988 if (FirstInst->getParent() == LI->getParent() ||
989 DT->dominates(LI, FirstInst)) {
990 TaintRelaxedLoads(LI, FirstInst);
994 findMostRecentDependenceUsage(LI, FirstInst, &ChainedBB, DT);
996 TaintRelaxedLoads(Inst, FirstInst);
998 LI->setOrdering(Acquire);
1008 /**** Implementations of public methods for dependence tainting ****/
1009 Value* GetUntaintedAddress(Value* CurrentAddress) {
1010 auto* OrAddress = getOrAddress(CurrentAddress);
1011 if (OrAddress == nullptr) {
1012 // Is it tainted by a select instruction?
1013 auto* Inst = dyn_cast<Instruction>(CurrentAddress);
1014 if (nullptr != Inst && Inst->getOpcode() == Instruction::Select) {
1015 // A selection instruction.
1016 if (Inst->getOperand(1) == Inst->getOperand(2)) {
1017 return Inst->getOperand(1);
1021 return CurrentAddress;
1023 Value* ActualAddress = nullptr;
1025 auto* CastToInt = dyn_cast<Instruction>(OrAddress->getOperand(1));
1026 if (CastToInt && CastToInt->getOpcode() == Instruction::PtrToInt) {
1027 return CastToInt->getOperand(0);
1029 // This should be a IntToPtr constant expression.
1030 ConstantExpr* PtrToIntExpr =
1031 dyn_cast<ConstantExpr>(OrAddress->getOperand(1));
1032 if (PtrToIntExpr && PtrToIntExpr->getOpcode() == Instruction::PtrToInt) {
1033 return PtrToIntExpr->getOperand(0);
1037 // Looks like it's not been dependence-tainted. Returns itself.
1038 return CurrentAddress;
1041 MemoryLocation GetUntaintedMemoryLocation(StoreInst* SI) {
1043 SI->getAAMetadata(AATags);
1044 const auto& DL = SI->getModule()->getDataLayout();
1045 const auto* OriginalAddr = GetUntaintedAddress(SI->getPointerOperand());
1046 DEBUG(if (OriginalAddr != SI->getPointerOperand()) {
1047 dbgs() << "[GetUntaintedMemoryLocation]\n"
1048 << "Storing address: " << *SI->getPointerOperand()
1049 << "\nUntainted address: " << *OriginalAddr << "\n";
1051 return MemoryLocation(OriginalAddr,
1052 DL.getTypeStoreSize(SI->getValueOperand()->getType()),
1056 bool TaintDependenceToStore(StoreInst* SI, Value* DepVal) {
1057 if (dependenceSetInclusion(SI, DepVal)) {
1061 bool tainted = taintStoreAddress(SI, DepVal);
1066 bool TaintDependenceToStoreAddress(StoreInst* SI, Value* DepVal) {
1067 if (dependenceSetInclusion(SI->getPointerOperand(), DepVal)) {
1071 bool tainted = taintStoreAddress(SI, DepVal);
1076 bool CompressTaintedStore(BasicBlock* BB) {
1077 // This function looks for windows of adajcent stores in 'BB' that satisfy the
1078 // following condition (and then do optimization):
1079 // *Addr(d1) = v1, d1 is a condition and is the only dependence the store's
1080 // address depends on && Dep(v1) includes Dep(d1);
1081 // *Addr(d2) = v2, d2 is a condition and is the only dependnece the store's
1082 // address depends on && Dep(v2) includes Dep(d2) &&
1083 // Dep(d2) includes Dep(d1);
1085 // *Addr(dN) = vN, dN is a condition and is the only dependence the store's
1086 // address depends on && Dep(dN) includes Dep(d"N-1").
1088 // As a result, Dep(dN) includes [Dep(d1) V ... V Dep(d"N-1")], so we can
1089 // safely transform the above to the following. In between these stores, we
1090 // can omit untainted stores to the same address 'Addr' since they internally
1091 // have dependence on the previous stores on the same address.
1096 for (auto BI = BB->begin(), BE = BB->end(); BI != BE; BI++) {
1097 // Look for the first store in such a window of adajacent stores.
1098 auto* FirstSI = dyn_cast<StoreInst>(&*BI);
1103 // The first store in the window must be tainted.
1104 auto* UntaintedAddress = GetUntaintedAddress(FirstSI->getPointerOperand());
1105 if (UntaintedAddress == FirstSI->getPointerOperand()) {
1109 // The first store's address must directly depend on and only depend on a
1111 auto* FirstSIDepCond = getConditionDependence(FirstSI->getPointerOperand());
1112 if (nullptr == FirstSIDepCond) {
1116 // Dep(first store's storing value) includes Dep(tainted dependence).
1117 if (!dependenceSetInclusion(FirstSI->getValueOperand(), FirstSIDepCond)) {
1121 // Look for subsequent stores to the same address that satisfy the condition
1122 // of "compressing the dependence".
1123 SmallVector<StoreInst*, 8> AdajacentStores;
1124 AdajacentStores.push_back(FirstSI);
1125 auto BII = BasicBlock::iterator(FirstSI);
1126 for (BII++; BII != BE; BII++) {
1127 auto* CurrSI = dyn_cast<StoreInst>(&*BII);
1129 if (BII->mayHaveSideEffects()) {
1130 // Be conservative. Instructions with side effects are similar to
1137 auto* OrigAddress = GetUntaintedAddress(CurrSI->getPointerOperand());
1138 auto* CurrSIDepCond = getConditionDependence(CurrSI->getPointerOperand());
1139 // All other stores must satisfy either:
1140 // A. 'CurrSI' is an untainted store to the same address, or
1141 // B. the combination of the following 5 subconditions:
1143 // 2. Untainted address is the same as the group's address;
1144 // 3. The address is tainted with a sole value which is a condition;
1145 // 4. The storing value depends on the condition in 3.
1146 // 5. The condition in 3 depends on the previous stores dependence
1149 // Condition A. Should ignore this store directly.
1150 if (OrigAddress == CurrSI->getPointerOperand() &&
1151 OrigAddress == UntaintedAddress) {
1154 // Check condition B.
1155 Value* Cond = nullptr;
1156 if (OrigAddress == CurrSI->getPointerOperand() ||
1157 OrigAddress != UntaintedAddress || CurrSIDepCond == nullptr ||
1158 !dependenceSetInclusion(CurrSI->getValueOperand(), CurrSIDepCond)) {
1159 // Check condition 1, 2, 3 & 4.
1163 // Check condition 5.
1164 StoreInst* PrevSI = AdajacentStores[AdajacentStores.size() - 1];
1165 auto* PrevSIDepCond = getConditionDependence(PrevSI->getPointerOperand());
1166 assert(PrevSIDepCond &&
1167 "Store in the group must already depend on a condtion");
1168 if (!dependenceSetInclusion(CurrSIDepCond, PrevSIDepCond)) {
1172 AdajacentStores.push_back(CurrSI);
1175 if (AdajacentStores.size() == 1) {
1176 // The outer loop should keep looking from the next store.
1180 // Now we have such a group of tainted stores to the same address.
1181 DEBUG(dbgs() << "[CompressTaintedStore]\n");
1182 DEBUG(dbgs() << "Original BB\n");
1183 DEBUG(dbgs() << *BB << '\n');
1184 auto* LastSI = AdajacentStores[AdajacentStores.size() - 1];
1185 for (unsigned i = 0; i < AdajacentStores.size() - 1; ++i) {
1186 auto* SI = AdajacentStores[i];
1188 // Use the original address for stores before the last one.
1189 SI->setOperand(1, UntaintedAddress);
1191 DEBUG(dbgs() << "Store address has been reversed: " << *SI << '\n';);
1193 // XXX-comment: Try to make the last store use fewer registers.
1194 // If LastSI's storing value is a select based on the condition with which
1195 // its address is tainted, transform the tainted address to a select
1196 // instruction, as follows:
1197 // r1 = Select Cond ? A : B
1202 // r1 = Select Cond ? A : B
1203 // r2 = Select Cond ? Addr : Addr
1205 // The idea is that both Select instructions depend on the same condition,
1206 // so hopefully the backend can generate two cmov instructions for them (and
1207 // this saves the number of registers needed).
1208 auto* LastSIDep = getConditionDependence(LastSI->getPointerOperand());
1209 auto* LastSIValue = dyn_cast<Instruction>(LastSI->getValueOperand());
1210 if (LastSIValue && LastSIValue->getOpcode() == Instruction::Select &&
1211 LastSIValue->getOperand(0) == LastSIDep) {
1212 // XXX-comment: Maybe it's better for us to just leave it as an and/or
1213 // dependence pattern.
1215 IRBuilder<true, NoFolder> Builder(LastSI);
1217 Builder.CreateSelect(LastSIDep, UntaintedAddress, UntaintedAddress);
1218 LastSI->setOperand(1, Address);
1219 DEBUG(dbgs() << "The last store becomes :" << *LastSI << "\n\n";);
1227 bool PassDependenceToStore(Value* OldAddress, StoreInst* NewStore) {
1228 Value* OldDep = getDependence(OldAddress);
1229 // Return false when there's no dependence to pass from the OldAddress.
1234 // No need to pass the dependence to NewStore's address if it already depends
1235 // on whatever 'OldAddress' depends on.
1236 if (StoreAddressDependOnValue(NewStore, OldDep)) {
1239 return taintStoreAddress(NewStore, OldAddress);
1242 SmallSet<Value*, 8> FindDependence(Value* Val) {
1243 SmallSet<Value*, 8> DepSet;
1244 recursivelyFindDependence(&DepSet, Val, true /*Only insert leaf nodes*/);
1248 bool StoreAddressDependOnValue(StoreInst* SI, Value* DepVal) {
1249 return dependenceSetInclusion(SI->getPointerOperand(), DepVal);
1252 bool StoreDependOnValue(StoreInst* SI, Value* Dep) {
1253 return dependenceSetInclusion(SI, Dep);
1260 bool CodeGenPrepare::runOnFunction(Function &F) {
1261 bool EverMadeChange = false;
1263 if (skipOptnoneFunction(F))
1266 DL = &F.getParent()->getDataLayout();
1268 // Clear per function information.
1269 InsertedInsts.clear();
1270 PromotedInsts.clear();
1274 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
1275 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1276 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1277 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1278 OptSize = F.optForSize();
1280 /// This optimization identifies DIV instructions that can be
1281 /// profitably bypassed and carried out with a shorter, faster divide.
1282 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
1283 const DenseMap<unsigned int, unsigned int> &BypassWidths =
1284 TLI->getBypassSlowDivWidths();
1285 BasicBlock* BB = &*F.begin();
1286 while (BB != nullptr) {
1287 // bypassSlowDivision may create new BBs, but we don't want to reapply the
1288 // optimization to those blocks.
1289 BasicBlock* Next = BB->getNextNode();
1290 EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
1295 // Eliminate blocks that contain only PHI nodes and an
1296 // unconditional branch.
1297 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
1299 // llvm.dbg.value is far away from the value then iSel may not be able
1300 // handle it properly. iSel will drop llvm.dbg.value if it can not
1301 // find a node corresponding to the value.
1302 EverMadeChange |= placeDbgValues(F);
1304 // If there is a mask, compare against zero, and branch that can be combined
1305 // into a single target instruction, push the mask and compare into branch
1306 // users. Do this before OptimizeBlock -> OptimizeInst ->
1307 // OptimizeCmpExpression, which perturbs the pattern being searched for.
1308 if (!DisableBranchOpts) {
1309 EverMadeChange |= sinkAndCmp(F);
1310 EverMadeChange |= splitBranchCondition(F);
1313 bool MadeChange = true;
1314 while (MadeChange) {
1316 for (Function::iterator I = F.begin(); I != F.end(); ) {
1317 BasicBlock *BB = &*I++;
1318 bool ModifiedDTOnIteration = false;
1319 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
1321 // Restart BB iteration if the dominator tree of the Function was changed
1322 if (ModifiedDTOnIteration)
1325 EverMadeChange |= MadeChange;
1330 if (!DisableBranchOpts) {
1332 SmallPtrSet<BasicBlock*, 8> WorkList;
1333 for (BasicBlock &BB : F) {
1334 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
1335 MadeChange |= ConstantFoldTerminator(&BB, true);
1336 if (!MadeChange) continue;
1338 for (SmallVectorImpl<BasicBlock*>::iterator
1339 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
1340 if (pred_begin(*II) == pred_end(*II))
1341 WorkList.insert(*II);
1344 // Delete the dead blocks and any of their dead successors.
1345 MadeChange |= !WorkList.empty();
1346 while (!WorkList.empty()) {
1347 BasicBlock *BB = *WorkList.begin();
1349 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
1351 DeleteDeadBlock(BB);
1353 for (SmallVectorImpl<BasicBlock*>::iterator
1354 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
1355 if (pred_begin(*II) == pred_end(*II))
1356 WorkList.insert(*II);
1359 // Merge pairs of basic blocks with unconditional branches, connected by
1361 if (EverMadeChange || MadeChange)
1362 MadeChange |= eliminateFallThrough(F);
1364 EverMadeChange |= MadeChange;
1367 if (!DisableGCOpts) {
1368 SmallVector<Instruction *, 2> Statepoints;
1369 for (BasicBlock &BB : F)
1370 for (Instruction &I : BB)
1371 if (isStatepoint(I))
1372 Statepoints.push_back(&I);
1373 for (auto &I : Statepoints)
1374 EverMadeChange |= simplifyOffsetableRelocate(*I);
1377 // XXX-comment: Delay dealing with relaxed loads in this function to avoid
1378 // further changes done by other passes (e.g., SimplifyCFG).
1379 // Collect all the relaxed loads.
1380 SmallSet<LoadInst*, 1> MonotonicLoadInsts;
1381 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
1382 if (I->isAtomic()) {
1383 switch (I->getOpcode()) {
1384 case Instruction::Load: {
1385 auto* LI = dyn_cast<LoadInst>(&*I);
1386 if (LI->getOrdering() == Monotonic) {
1387 MonotonicLoadInsts.insert(LI);
1398 AddFakeConditionalBranchAfterMonotonicLoads(MonotonicLoadInsts, DT);
1400 return EverMadeChange;
1403 /// Merge basic blocks which are connected by a single edge, where one of the
1404 /// basic blocks has a single successor pointing to the other basic block,
1405 /// which has a single predecessor.
1406 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
1407 bool Changed = false;
1408 // Scan all of the blocks in the function, except for the entry block.
1409 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
1410 BasicBlock *BB = &*I++;
1411 // If the destination block has a single pred, then this is a trivial
1412 // edge, just collapse it.
1413 BasicBlock *SinglePred = BB->getSinglePredecessor();
1415 // Don't merge if BB's address is taken.
1416 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
1418 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
1419 if (Term && !Term->isConditional()) {
1421 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
1422 // Remember if SinglePred was the entry block of the function.
1423 // If so, we will need to move BB back to the entry position.
1424 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
1425 MergeBasicBlockIntoOnlyPred(BB, nullptr);
1427 if (isEntry && BB != &BB->getParent()->getEntryBlock())
1428 BB->moveBefore(&BB->getParent()->getEntryBlock());
1430 // We have erased a block. Update the iterator.
1431 I = BB->getIterator();
1437 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
1438 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
1439 /// edges in ways that are non-optimal for isel. Start by eliminating these
1440 /// blocks so we can split them the way we want them.
1441 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
1442 bool MadeChange = false;
1443 // Note that this intentionally skips the entry block.
1444 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
1445 BasicBlock *BB = &*I++;
1446 // If this block doesn't end with an uncond branch, ignore it.
1447 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
1448 if (!BI || !BI->isUnconditional())
1451 // If the instruction before the branch (skipping debug info) isn't a phi
1452 // node, then other stuff is happening here.
1453 BasicBlock::iterator BBI = BI->getIterator();
1454 if (BBI != BB->begin()) {
1456 while (isa<DbgInfoIntrinsic>(BBI)) {
1457 if (BBI == BB->begin())
1461 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
1465 // Do not break infinite loops.
1466 BasicBlock *DestBB = BI->getSuccessor(0);
1470 if (!canMergeBlocks(BB, DestBB))
1473 eliminateMostlyEmptyBlock(BB);
1479 /// Return true if we can merge BB into DestBB if there is a single
1480 /// unconditional branch between them, and BB contains no other non-phi
1482 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
1483 const BasicBlock *DestBB) const {
1484 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
1485 // the successor. If there are more complex condition (e.g. preheaders),
1486 // don't mess around with them.
1487 BasicBlock::const_iterator BBI = BB->begin();
1488 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
1489 for (const User *U : PN->users()) {
1490 const Instruction *UI = cast<Instruction>(U);
1491 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
1493 // IfUser is inside DestBB block and it is a PHINode then check
1494 // incoming value. If incoming value is not from BB then this is
1495 // a complex condition (e.g. preheaders) we want to avoid here.
1496 if (UI->getParent() == DestBB) {
1497 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
1498 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
1499 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
1500 if (Insn && Insn->getParent() == BB &&
1501 Insn->getParent() != UPN->getIncomingBlock(I))
1508 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
1509 // and DestBB may have conflicting incoming values for the block. If so, we
1510 // can't merge the block.
1511 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
1512 if (!DestBBPN) return true; // no conflict.
1514 // Collect the preds of BB.
1515 SmallPtrSet<const BasicBlock*, 16> BBPreds;
1516 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
1517 // It is faster to get preds from a PHI than with pred_iterator.
1518 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1519 BBPreds.insert(BBPN->getIncomingBlock(i));
1521 BBPreds.insert(pred_begin(BB), pred_end(BB));
1524 // Walk the preds of DestBB.
1525 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
1526 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
1527 if (BBPreds.count(Pred)) { // Common predecessor?
1528 BBI = DestBB->begin();
1529 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
1530 const Value *V1 = PN->getIncomingValueForBlock(Pred);
1531 const Value *V2 = PN->getIncomingValueForBlock(BB);
1533 // If V2 is a phi node in BB, look up what the mapped value will be.
1534 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
1535 if (V2PN->getParent() == BB)
1536 V2 = V2PN->getIncomingValueForBlock(Pred);
1538 // If there is a conflict, bail out.
1539 if (V1 != V2) return false;
1548 /// Eliminate a basic block that has only phi's and an unconditional branch in
1550 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
1551 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1552 BasicBlock *DestBB = BI->getSuccessor(0);
1554 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
1556 // If the destination block has a single pred, then this is a trivial edge,
1557 // just collapse it.
1558 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
1559 if (SinglePred != DestBB) {
1560 // Remember if SinglePred was the entry block of the function. If so, we
1561 // will need to move BB back to the entry position.
1562 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
1563 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
1565 if (isEntry && BB != &BB->getParent()->getEntryBlock())
1566 BB->moveBefore(&BB->getParent()->getEntryBlock());
1568 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1573 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
1574 // to handle the new incoming edges it is about to have.
1576 for (BasicBlock::iterator BBI = DestBB->begin();
1577 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1578 // Remove the incoming value for BB, and remember it.
1579 Value *InVal = PN->removeIncomingValue(BB, false);
1581 // Two options: either the InVal is a phi node defined in BB or it is some
1582 // value that dominates BB.
1583 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
1584 if (InValPhi && InValPhi->getParent() == BB) {
1585 // Add all of the input values of the input PHI as inputs of this phi.
1586 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
1587 PN->addIncoming(InValPhi->getIncomingValue(i),
1588 InValPhi->getIncomingBlock(i));
1590 // Otherwise, add one instance of the dominating value for each edge that
1591 // we will be adding.
1592 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
1593 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1594 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
1596 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
1597 PN->addIncoming(InVal, *PI);
1602 // The PHIs are now updated, change everything that refers to BB to use
1603 // DestBB and remove BB.
1604 BB->replaceAllUsesWith(DestBB);
1605 BB->eraseFromParent();
1608 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1611 // Computes a map of base pointer relocation instructions to corresponding
1612 // derived pointer relocation instructions given a vector of all relocate calls
1613 static void computeBaseDerivedRelocateMap(
1614 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
1615 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
1617 // Collect information in two maps: one primarily for locating the base object
1618 // while filling the second map; the second map is the final structure holding
1619 // a mapping between Base and corresponding Derived relocate calls
1620 DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
1621 for (auto *ThisRelocate : AllRelocateCalls) {
1622 auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
1623 ThisRelocate->getDerivedPtrIndex());
1624 RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
1626 for (auto &Item : RelocateIdxMap) {
1627 std::pair<unsigned, unsigned> Key = Item.first;
1628 if (Key.first == Key.second)
1629 // Base relocation: nothing to insert
1632 GCRelocateInst *I = Item.second;
1633 auto BaseKey = std::make_pair(Key.first, Key.first);
1635 // We're iterating over RelocateIdxMap so we cannot modify it.
1636 auto MaybeBase = RelocateIdxMap.find(BaseKey);
1637 if (MaybeBase == RelocateIdxMap.end())
1638 // TODO: We might want to insert a new base object relocate and gep off
1639 // that, if there are enough derived object relocates.
1642 RelocateInstMap[MaybeBase->second].push_back(I);
1646 // Accepts a GEP and extracts the operands into a vector provided they're all
1647 // small integer constants
1648 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
1649 SmallVectorImpl<Value *> &OffsetV) {
1650 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
1651 // Only accept small constant integer operands
1652 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
1653 if (!Op || Op->getZExtValue() > 20)
1657 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
1658 OffsetV.push_back(GEP->getOperand(i));
1662 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
1663 // replace, computes a replacement, and affects it.
1665 simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
1666 const SmallVectorImpl<GCRelocateInst *> &Targets) {
1667 bool MadeChange = false;
1668 for (GCRelocateInst *ToReplace : Targets) {
1669 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
1670 "Not relocating a derived object of the original base object");
1671 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
1672 // A duplicate relocate call. TODO: coalesce duplicates.
1676 if (RelocatedBase->getParent() != ToReplace->getParent()) {
1677 // Base and derived relocates are in different basic blocks.
1678 // In this case transform is only valid when base dominates derived
1679 // relocate. However it would be too expensive to check dominance
1680 // for each such relocate, so we skip the whole transformation.
1684 Value *Base = ToReplace->getBasePtr();
1685 auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
1686 if (!Derived || Derived->getPointerOperand() != Base)
1689 SmallVector<Value *, 2> OffsetV;
1690 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
1693 // Create a Builder and replace the target callsite with a gep
1694 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
1696 // Insert after RelocatedBase
1697 IRBuilder<> Builder(RelocatedBase->getNextNode());
1698 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1700 // If gc_relocate does not match the actual type, cast it to the right type.
1701 // In theory, there must be a bitcast after gc_relocate if the type does not
1702 // match, and we should reuse it to get the derived pointer. But it could be
1706 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1711 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1715 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
1716 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
1718 // In this case, we can not find the bitcast any more. So we insert a new bitcast
1719 // no matter there is already one or not. In this way, we can handle all cases, and
1720 // the extra bitcast should be optimized away in later passes.
1721 Value *ActualRelocatedBase = RelocatedBase;
1722 if (RelocatedBase->getType() != Base->getType()) {
1723 ActualRelocatedBase =
1724 Builder.CreateBitCast(RelocatedBase, Base->getType());
1726 Value *Replacement = Builder.CreateGEP(
1727 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
1728 Replacement->takeName(ToReplace);
1729 // If the newly generated derived pointer's type does not match the original derived
1730 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
1731 Value *ActualReplacement = Replacement;
1732 if (Replacement->getType() != ToReplace->getType()) {
1734 Builder.CreateBitCast(Replacement, ToReplace->getType());
1736 ToReplace->replaceAllUsesWith(ActualReplacement);
1737 ToReplace->eraseFromParent();
1747 // %ptr = gep %base + 15
1748 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1749 // %base' = relocate(%tok, i32 4, i32 4)
1750 // %ptr' = relocate(%tok, i32 4, i32 5)
1751 // %val = load %ptr'
1756 // %ptr = gep %base + 15
1757 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1758 // %base' = gc.relocate(%tok, i32 4, i32 4)
1759 // %ptr' = gep %base' + 15
1760 // %val = load %ptr'
1761 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
1762 bool MadeChange = false;
1763 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1765 for (auto *U : I.users())
1766 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1767 // Collect all the relocate calls associated with a statepoint
1768 AllRelocateCalls.push_back(Relocate);
1770 // We need atleast one base pointer relocation + one derived pointer
1771 // relocation to mangle
1772 if (AllRelocateCalls.size() < 2)
1775 // RelocateInstMap is a mapping from the base relocate instruction to the
1776 // corresponding derived relocate instructions
1777 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
1778 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1779 if (RelocateInstMap.empty())
1782 for (auto &Item : RelocateInstMap)
1783 // Item.first is the RelocatedBase to offset against
1784 // Item.second is the vector of Targets to replace
1785 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1789 /// SinkCast - Sink the specified cast instruction into its user blocks
1790 static bool SinkCast(CastInst *CI) {
1791 BasicBlock *DefBB = CI->getParent();
1793 /// InsertedCasts - Only insert a cast in each block once.
1794 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1796 bool MadeChange = false;
1797 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1799 Use &TheUse = UI.getUse();
1800 Instruction *User = cast<Instruction>(*UI);
1802 // Figure out which BB this cast is used in. For PHI's this is the
1803 // appropriate predecessor block.
1804 BasicBlock *UserBB = User->getParent();
1805 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1806 UserBB = PN->getIncomingBlock(TheUse);
1809 // Preincrement use iterator so we don't invalidate it.
1812 // If the block selected to receive the cast is an EH pad that does not
1813 // allow non-PHI instructions before the terminator, we can't sink the
1815 if (UserBB->getTerminator()->isEHPad())
1818 // If this user is in the same block as the cast, don't change the cast.
1819 if (UserBB == DefBB) continue;
1821 // If we have already inserted a cast into this block, use it.
1822 CastInst *&InsertedCast = InsertedCasts[UserBB];
1824 if (!InsertedCast) {
1825 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1826 assert(InsertPt != UserBB->end());
1827 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1828 CI->getType(), "", &*InsertPt);
1831 // Replace a use of the cast with a use of the new cast.
1832 TheUse = InsertedCast;
1837 // If we removed all uses, nuke the cast.
1838 if (CI->use_empty()) {
1839 CI->eraseFromParent();
1846 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1847 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1848 /// reduce the number of virtual registers that must be created and coalesced.
1850 /// Return true if any changes are made.
1852 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1853 const DataLayout &DL) {
1854 // If this is a noop copy,
1855 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1856 EVT DstVT = TLI.getValueType(DL, CI->getType());
1858 // This is an fp<->int conversion?
1859 if (SrcVT.isInteger() != DstVT.isInteger())
1862 // If this is an extension, it will be a zero or sign extension, which
1864 if (SrcVT.bitsLT(DstVT)) return false;
1866 // If these values will be promoted, find out what they will be promoted
1867 // to. This helps us consider truncates on PPC as noop copies when they
1869 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1870 TargetLowering::TypePromoteInteger)
1871 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1872 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1873 TargetLowering::TypePromoteInteger)
1874 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1876 // If, after promotion, these are the same types, this is a noop copy.
1880 return SinkCast(CI);
1883 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
1886 /// Return true if any changes were made.
1887 static bool CombineUAddWithOverflow(CmpInst *CI) {
1891 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
1894 Type *Ty = AddI->getType();
1895 if (!isa<IntegerType>(Ty))
1898 // We don't want to move around uses of condition values this late, so we we
1899 // check if it is legal to create the call to the intrinsic in the basic
1900 // block containing the icmp:
1902 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
1906 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
1908 if (AddI->hasOneUse())
1909 assert(*AddI->user_begin() == CI && "expected!");
1912 Module *M = CI->getModule();
1913 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1915 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
1917 auto *UAddWithOverflow =
1918 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
1919 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
1921 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
1923 CI->replaceAllUsesWith(Overflow);
1924 AddI->replaceAllUsesWith(UAdd);
1925 CI->eraseFromParent();
1926 AddI->eraseFromParent();
1930 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1931 /// registers that must be created and coalesced. This is a clear win except on
1932 /// targets with multiple condition code registers (PowerPC), where it might
1933 /// lose; some adjustment may be wanted there.
1935 /// Return true if any changes are made.
1936 static bool SinkCmpExpression(CmpInst *CI) {
1937 BasicBlock *DefBB = CI->getParent();
1939 /// Only insert a cmp in each block once.
1940 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1942 bool MadeChange = false;
1943 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1945 Use &TheUse = UI.getUse();
1946 Instruction *User = cast<Instruction>(*UI);
1948 // Preincrement use iterator so we don't invalidate it.
1951 // Don't bother for PHI nodes.
1952 if (isa<PHINode>(User))
1955 // Figure out which BB this cmp is used in.
1956 BasicBlock *UserBB = User->getParent();
1958 // If this user is in the same block as the cmp, don't change the cmp.
1959 if (UserBB == DefBB) continue;
1961 // If we have already inserted a cmp into this block, use it.
1962 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1965 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1966 assert(InsertPt != UserBB->end());
1968 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1969 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
1972 // Replace a use of the cmp with a use of the new cmp.
1973 TheUse = InsertedCmp;
1978 // If we removed all uses, nuke the cmp.
1979 if (CI->use_empty()) {
1980 CI->eraseFromParent();
1987 static bool OptimizeCmpExpression(CmpInst *CI) {
1988 if (SinkCmpExpression(CI))
1991 if (CombineUAddWithOverflow(CI))
1997 /// Check if the candidates could be combined with a shift instruction, which
1999 /// 1. Truncate instruction
2000 /// 2. And instruction and the imm is a mask of the low bits:
2001 /// imm & (imm+1) == 0
2002 static bool isExtractBitsCandidateUse(Instruction *User) {
2003 if (!isa<TruncInst>(User)) {
2004 if (User->getOpcode() != Instruction::And ||
2005 !isa<ConstantInt>(User->getOperand(1)))
2008 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
2010 if ((Cimm & (Cimm + 1)).getBoolValue())
2016 /// Sink both shift and truncate instruction to the use of truncate's BB.
2018 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
2019 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
2020 const TargetLowering &TLI, const DataLayout &DL) {
2021 BasicBlock *UserBB = User->getParent();
2022 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
2023 TruncInst *TruncI = dyn_cast<TruncInst>(User);
2024 bool MadeChange = false;
2026 for (Value::user_iterator TruncUI = TruncI->user_begin(),
2027 TruncE = TruncI->user_end();
2028 TruncUI != TruncE;) {
2030 Use &TruncTheUse = TruncUI.getUse();
2031 Instruction *TruncUser = cast<Instruction>(*TruncUI);
2032 // Preincrement use iterator so we don't invalidate it.
2036 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
2040 // If the use is actually a legal node, there will not be an
2041 // implicit truncate.
2042 // FIXME: always querying the result type is just an
2043 // approximation; some nodes' legality is determined by the
2044 // operand or other means. There's no good way to find out though.
2045 if (TLI.isOperationLegalOrCustom(
2046 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
2049 // Don't bother for PHI nodes.
2050 if (isa<PHINode>(TruncUser))
2053 BasicBlock *TruncUserBB = TruncUser->getParent();
2055 if (UserBB == TruncUserBB)
2058 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
2059 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
2061 if (!InsertedShift && !InsertedTrunc) {
2062 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
2063 assert(InsertPt != TruncUserBB->end());
2065 if (ShiftI->getOpcode() == Instruction::AShr)
2066 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
2069 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
2073 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
2075 assert(TruncInsertPt != TruncUserBB->end());
2077 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
2078 TruncI->getType(), "", &*TruncInsertPt);
2082 TruncTheUse = InsertedTrunc;
2088 /// Sink the shift *right* instruction into user blocks if the uses could
2089 /// potentially be combined with this shift instruction and generate BitExtract
2090 /// instruction. It will only be applied if the architecture supports BitExtract
2091 /// instruction. Here is an example:
2093 /// %x.extract.shift = lshr i64 %arg1, 32
2095 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
2099 /// %x.extract.shift.1 = lshr i64 %arg1, 32
2100 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
2102 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
2104 /// Return true if any changes are made.
2105 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
2106 const TargetLowering &TLI,
2107 const DataLayout &DL) {
2108 BasicBlock *DefBB = ShiftI->getParent();
2110 /// Only insert instructions in each block once.
2111 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
2113 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
2115 bool MadeChange = false;
2116 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
2118 Use &TheUse = UI.getUse();
2119 Instruction *User = cast<Instruction>(*UI);
2120 // Preincrement use iterator so we don't invalidate it.
2123 // Don't bother for PHI nodes.
2124 if (isa<PHINode>(User))
2127 if (!isExtractBitsCandidateUse(User))
2130 BasicBlock *UserBB = User->getParent();
2132 if (UserBB == DefBB) {
2133 // If the shift and truncate instruction are in the same BB. The use of
2134 // the truncate(TruncUse) may still introduce another truncate if not
2135 // legal. In this case, we would like to sink both shift and truncate
2136 // instruction to the BB of TruncUse.
2139 // i64 shift.result = lshr i64 opnd, imm
2140 // trunc.result = trunc shift.result to i16
2143 // ----> We will have an implicit truncate here if the architecture does
2144 // not have i16 compare.
2145 // cmp i16 trunc.result, opnd2
2147 if (isa<TruncInst>(User) && shiftIsLegal
2148 // If the type of the truncate is legal, no trucate will be
2149 // introduced in other basic blocks.
2151 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
2153 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
2157 // If we have already inserted a shift into this block, use it.
2158 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
2160 if (!InsertedShift) {
2161 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2162 assert(InsertPt != UserBB->end());
2164 if (ShiftI->getOpcode() == Instruction::AShr)
2165 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
2168 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
2174 // Replace a use of the shift with a use of the new shift.
2175 TheUse = InsertedShift;
2178 // If we removed all uses, nuke the shift.
2179 if (ShiftI->use_empty())
2180 ShiftI->eraseFromParent();
2185 // Translate a masked load intrinsic like
2186 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
2187 // <16 x i1> %mask, <16 x i32> %passthru)
2188 // to a chain of basic blocks, with loading element one-by-one if
2189 // the appropriate mask bit is set
2191 // %1 = bitcast i8* %addr to i32*
2192 // %2 = extractelement <16 x i1> %mask, i32 0
2193 // %3 = icmp eq i1 %2, true
2194 // br i1 %3, label %cond.load, label %else
2196 //cond.load: ; preds = %0
2197 // %4 = getelementptr i32* %1, i32 0
2198 // %5 = load i32* %4
2199 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
2202 //else: ; preds = %0, %cond.load
2203 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
2204 // %7 = extractelement <16 x i1> %mask, i32 1
2205 // %8 = icmp eq i1 %7, true
2206 // br i1 %8, label %cond.load1, label %else2
2208 //cond.load1: ; preds = %else
2209 // %9 = getelementptr i32* %1, i32 1
2210 // %10 = load i32* %9
2211 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
2214 //else2: ; preds = %else, %cond.load1
2215 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
2216 // %12 = extractelement <16 x i1> %mask, i32 2
2217 // %13 = icmp eq i1 %12, true
2218 // br i1 %13, label %cond.load4, label %else5
2220 static void ScalarizeMaskedLoad(CallInst *CI) {
2221 Value *Ptr = CI->getArgOperand(0);
2222 Value *Alignment = CI->getArgOperand(1);
2223 Value *Mask = CI->getArgOperand(2);
2224 Value *Src0 = CI->getArgOperand(3);
2226 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2227 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
2228 assert(VecType && "Unexpected return type of masked load intrinsic");
2230 Type *EltTy = CI->getType()->getVectorElementType();
2232 IRBuilder<> Builder(CI->getContext());
2233 Instruction *InsertPt = CI;
2234 BasicBlock *IfBlock = CI->getParent();
2235 BasicBlock *CondBlock = nullptr;
2236 BasicBlock *PrevIfBlock = CI->getParent();
2238 Builder.SetInsertPoint(InsertPt);
2239 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2241 // Short-cut if the mask is all-true.
2242 bool IsAllOnesMask = isa<Constant>(Mask) &&
2243 cast<Constant>(Mask)->isAllOnesValue();
2245 if (IsAllOnesMask) {
2246 Value *NewI = Builder.CreateAlignedLoad(Ptr, AlignVal);
2247 CI->replaceAllUsesWith(NewI);
2248 CI->eraseFromParent();
2252 // Adjust alignment for the scalar instruction.
2253 AlignVal = std::min(AlignVal, VecType->getScalarSizeInBits()/8);
2254 // Bitcast %addr fron i8* to EltTy*
2256 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
2257 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
2258 unsigned VectorWidth = VecType->getNumElements();
2260 Value *UndefVal = UndefValue::get(VecType);
2262 // The result vector
2263 Value *VResult = UndefVal;
2265 if (isa<ConstantVector>(Mask)) {
2266 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2267 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2270 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2271 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
2272 VResult = Builder.CreateInsertElement(VResult, Load,
2273 Builder.getInt32(Idx));
2275 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
2276 CI->replaceAllUsesWith(NewI);
2277 CI->eraseFromParent();
2281 PHINode *Phi = nullptr;
2282 Value *PrevPhi = UndefVal;
2284 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2286 // Fill the "else" block, created in the previous iteration
2288 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
2289 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
2290 // %to_load = icmp eq i1 %mask_1, true
2291 // br i1 %to_load, label %cond.load, label %else
2294 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
2295 Phi->addIncoming(VResult, CondBlock);
2296 Phi->addIncoming(PrevPhi, PrevIfBlock);
2301 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
2302 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2303 ConstantInt::get(Predicate->getType(), 1));
2305 // Create "cond" block
2307 // %EltAddr = getelementptr i32* %1, i32 0
2308 // %Elt = load i32* %EltAddr
2309 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
2311 CondBlock = IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.load");
2312 Builder.SetInsertPoint(InsertPt);
2315 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2316 LoadInst *Load = Builder.CreateAlignedLoad(Gep, AlignVal);
2317 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
2319 // Create "else" block, fill it in the next iteration
2320 BasicBlock *NewIfBlock =
2321 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
2322 Builder.SetInsertPoint(InsertPt);
2323 Instruction *OldBr = IfBlock->getTerminator();
2324 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2325 OldBr->eraseFromParent();
2326 PrevIfBlock = IfBlock;
2327 IfBlock = NewIfBlock;
2330 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
2331 Phi->addIncoming(VResult, CondBlock);
2332 Phi->addIncoming(PrevPhi, PrevIfBlock);
2333 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
2334 CI->replaceAllUsesWith(NewI);
2335 CI->eraseFromParent();
2338 // Translate a masked store intrinsic, like
2339 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
2341 // to a chain of basic blocks, that stores element one-by-one if
2342 // the appropriate mask bit is set
2344 // %1 = bitcast i8* %addr to i32*
2345 // %2 = extractelement <16 x i1> %mask, i32 0
2346 // %3 = icmp eq i1 %2, true
2347 // br i1 %3, label %cond.store, label %else
2349 // cond.store: ; preds = %0
2350 // %4 = extractelement <16 x i32> %val, i32 0
2351 // %5 = getelementptr i32* %1, i32 0
2352 // store i32 %4, i32* %5
2355 // else: ; preds = %0, %cond.store
2356 // %6 = extractelement <16 x i1> %mask, i32 1
2357 // %7 = icmp eq i1 %6, true
2358 // br i1 %7, label %cond.store1, label %else2
2360 // cond.store1: ; preds = %else
2361 // %8 = extractelement <16 x i32> %val, i32 1
2362 // %9 = getelementptr i32* %1, i32 1
2363 // store i32 %8, i32* %9
2366 static void ScalarizeMaskedStore(CallInst *CI) {
2367 Value *Src = CI->getArgOperand(0);
2368 Value *Ptr = CI->getArgOperand(1);
2369 Value *Alignment = CI->getArgOperand(2);
2370 Value *Mask = CI->getArgOperand(3);
2372 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2373 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
2374 assert(VecType && "Unexpected data type in masked store intrinsic");
2376 Type *EltTy = VecType->getElementType();
2378 IRBuilder<> Builder(CI->getContext());
2379 Instruction *InsertPt = CI;
2380 BasicBlock *IfBlock = CI->getParent();
2381 Builder.SetInsertPoint(InsertPt);
2382 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2384 // Short-cut if the mask is all-true.
2385 bool IsAllOnesMask = isa<Constant>(Mask) &&
2386 cast<Constant>(Mask)->isAllOnesValue();
2388 if (IsAllOnesMask) {
2389 Builder.CreateAlignedStore(Src, Ptr, AlignVal);
2390 CI->eraseFromParent();
2394 // Adjust alignment for the scalar instruction.
2395 AlignVal = std::max(AlignVal, VecType->getScalarSizeInBits()/8);
2396 // Bitcast %addr fron i8* to EltTy*
2398 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
2399 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
2400 unsigned VectorWidth = VecType->getNumElements();
2402 if (isa<ConstantVector>(Mask)) {
2403 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2404 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2406 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
2408 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2409 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
2411 CI->eraseFromParent();
2415 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2417 // Fill the "else" block, created in the previous iteration
2419 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
2420 // %to_store = icmp eq i1 %mask_1, true
2421 // br i1 %to_store, label %cond.store, label %else
2423 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
2424 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2425 ConstantInt::get(Predicate->getType(), 1));
2427 // Create "cond" block
2429 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
2430 // %EltAddr = getelementptr i32* %1, i32 0
2431 // %store i32 %OneElt, i32* %EltAddr
2433 BasicBlock *CondBlock =
2434 IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.store");
2435 Builder.SetInsertPoint(InsertPt);
2437 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
2439 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2440 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
2442 // Create "else" block, fill it in the next iteration
2443 BasicBlock *NewIfBlock =
2444 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
2445 Builder.SetInsertPoint(InsertPt);
2446 Instruction *OldBr = IfBlock->getTerminator();
2447 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2448 OldBr->eraseFromParent();
2449 IfBlock = NewIfBlock;
2451 CI->eraseFromParent();
2454 // Translate a masked gather intrinsic like
2455 // <16 x i32 > @llvm.masked.gather.v16i32( <16 x i32*> %Ptrs, i32 4,
2456 // <16 x i1> %Mask, <16 x i32> %Src)
2457 // to a chain of basic blocks, with loading element one-by-one if
2458 // the appropriate mask bit is set
2460 // % Ptrs = getelementptr i32, i32* %base, <16 x i64> %ind
2461 // % Mask0 = extractelement <16 x i1> %Mask, i32 0
2462 // % ToLoad0 = icmp eq i1 % Mask0, true
2463 // br i1 % ToLoad0, label %cond.load, label %else
2466 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
2467 // % Load0 = load i32, i32* % Ptr0, align 4
2468 // % Res0 = insertelement <16 x i32> undef, i32 % Load0, i32 0
2472 // %res.phi.else = phi <16 x i32>[% Res0, %cond.load], [undef, % 0]
2473 // % Mask1 = extractelement <16 x i1> %Mask, i32 1
2474 // % ToLoad1 = icmp eq i1 % Mask1, true
2475 // br i1 % ToLoad1, label %cond.load1, label %else2
2478 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2479 // % Load1 = load i32, i32* % Ptr1, align 4
2480 // % Res1 = insertelement <16 x i32> %res.phi.else, i32 % Load1, i32 1
2483 // % Result = select <16 x i1> %Mask, <16 x i32> %res.phi.select, <16 x i32> %Src
2484 // ret <16 x i32> %Result
2485 static void ScalarizeMaskedGather(CallInst *CI) {
2486 Value *Ptrs = CI->getArgOperand(0);
2487 Value *Alignment = CI->getArgOperand(1);
2488 Value *Mask = CI->getArgOperand(2);
2489 Value *Src0 = CI->getArgOperand(3);
2491 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
2493 assert(VecType && "Unexpected return type of masked load intrinsic");
2495 IRBuilder<> Builder(CI->getContext());
2496 Instruction *InsertPt = CI;
2497 BasicBlock *IfBlock = CI->getParent();
2498 BasicBlock *CondBlock = nullptr;
2499 BasicBlock *PrevIfBlock = CI->getParent();
2500 Builder.SetInsertPoint(InsertPt);
2501 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2503 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2505 Value *UndefVal = UndefValue::get(VecType);
2507 // The result vector
2508 Value *VResult = UndefVal;
2509 unsigned VectorWidth = VecType->getNumElements();
2511 // Shorten the way if the mask is a vector of constants.
2512 bool IsConstMask = isa<ConstantVector>(Mask);
2515 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2516 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2518 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2519 "Ptr" + Twine(Idx));
2520 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
2521 "Load" + Twine(Idx));
2522 VResult = Builder.CreateInsertElement(VResult, Load,
2523 Builder.getInt32(Idx),
2524 "Res" + Twine(Idx));
2526 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
2527 CI->replaceAllUsesWith(NewI);
2528 CI->eraseFromParent();
2532 PHINode *Phi = nullptr;
2533 Value *PrevPhi = UndefVal;
2535 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2537 // Fill the "else" block, created in the previous iteration
2539 // %Mask1 = extractelement <16 x i1> %Mask, i32 1
2540 // %ToLoad1 = icmp eq i1 %Mask1, true
2541 // br i1 %ToLoad1, label %cond.load, label %else
2544 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
2545 Phi->addIncoming(VResult, CondBlock);
2546 Phi->addIncoming(PrevPhi, PrevIfBlock);
2551 Value *Predicate = Builder.CreateExtractElement(Mask,
2552 Builder.getInt32(Idx),
2553 "Mask" + Twine(Idx));
2554 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2555 ConstantInt::get(Predicate->getType(), 1),
2556 "ToLoad" + Twine(Idx));
2558 // Create "cond" block
2560 // %EltAddr = getelementptr i32* %1, i32 0
2561 // %Elt = load i32* %EltAddr
2562 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
2564 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
2565 Builder.SetInsertPoint(InsertPt);
2567 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2568 "Ptr" + Twine(Idx));
2569 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
2570 "Load" + Twine(Idx));
2571 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx),
2572 "Res" + Twine(Idx));
2574 // Create "else" block, fill it in the next iteration
2575 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2576 Builder.SetInsertPoint(InsertPt);
2577 Instruction *OldBr = IfBlock->getTerminator();
2578 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2579 OldBr->eraseFromParent();
2580 PrevIfBlock = IfBlock;
2581 IfBlock = NewIfBlock;
2584 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
2585 Phi->addIncoming(VResult, CondBlock);
2586 Phi->addIncoming(PrevPhi, PrevIfBlock);
2587 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
2588 CI->replaceAllUsesWith(NewI);
2589 CI->eraseFromParent();
2592 // Translate a masked scatter intrinsic, like
2593 // void @llvm.masked.scatter.v16i32(<16 x i32> %Src, <16 x i32*>* %Ptrs, i32 4,
2595 // to a chain of basic blocks, that stores element one-by-one if
2596 // the appropriate mask bit is set.
2598 // % Ptrs = getelementptr i32, i32* %ptr, <16 x i64> %ind
2599 // % Mask0 = extractelement <16 x i1> % Mask, i32 0
2600 // % ToStore0 = icmp eq i1 % Mask0, true
2601 // br i1 %ToStore0, label %cond.store, label %else
2604 // % Elt0 = extractelement <16 x i32> %Src, i32 0
2605 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
2606 // store i32 %Elt0, i32* % Ptr0, align 4
2610 // % Mask1 = extractelement <16 x i1> % Mask, i32 1
2611 // % ToStore1 = icmp eq i1 % Mask1, true
2612 // br i1 % ToStore1, label %cond.store1, label %else2
2615 // % Elt1 = extractelement <16 x i32> %Src, i32 1
2616 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2617 // store i32 % Elt1, i32* % Ptr1, align 4
2620 static void ScalarizeMaskedScatter(CallInst *CI) {
2621 Value *Src = CI->getArgOperand(0);
2622 Value *Ptrs = CI->getArgOperand(1);
2623 Value *Alignment = CI->getArgOperand(2);
2624 Value *Mask = CI->getArgOperand(3);
2626 assert(isa<VectorType>(Src->getType()) &&
2627 "Unexpected data type in masked scatter intrinsic");
2628 assert(isa<VectorType>(Ptrs->getType()) &&
2629 isa<PointerType>(Ptrs->getType()->getVectorElementType()) &&
2630 "Vector of pointers is expected in masked scatter intrinsic");
2632 IRBuilder<> Builder(CI->getContext());
2633 Instruction *InsertPt = CI;
2634 BasicBlock *IfBlock = CI->getParent();
2635 Builder.SetInsertPoint(InsertPt);
2636 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2638 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2639 unsigned VectorWidth = Src->getType()->getVectorNumElements();
2641 // Shorten the way if the mask is a vector of constants.
2642 bool IsConstMask = isa<ConstantVector>(Mask);
2645 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2646 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2648 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
2649 "Elt" + Twine(Idx));
2650 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2651 "Ptr" + Twine(Idx));
2652 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
2654 CI->eraseFromParent();
2657 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2658 // Fill the "else" block, created in the previous iteration
2660 // % Mask1 = extractelement <16 x i1> % Mask, i32 Idx
2661 // % ToStore = icmp eq i1 % Mask1, true
2662 // br i1 % ToStore, label %cond.store, label %else
2664 Value *Predicate = Builder.CreateExtractElement(Mask,
2665 Builder.getInt32(Idx),
2666 "Mask" + Twine(Idx));
2668 Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2669 ConstantInt::get(Predicate->getType(), 1),
2670 "ToStore" + Twine(Idx));
2672 // Create "cond" block
2674 // % Elt1 = extractelement <16 x i32> %Src, i32 1
2675 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2676 // %store i32 % Elt1, i32* % Ptr1
2678 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2679 Builder.SetInsertPoint(InsertPt);
2681 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
2682 "Elt" + Twine(Idx));
2683 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2684 "Ptr" + Twine(Idx));
2685 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
2687 // Create "else" block, fill it in the next iteration
2688 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2689 Builder.SetInsertPoint(InsertPt);
2690 Instruction *OldBr = IfBlock->getTerminator();
2691 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2692 OldBr->eraseFromParent();
2693 IfBlock = NewIfBlock;
2695 CI->eraseFromParent();
2698 /// If counting leading or trailing zeros is an expensive operation and a zero
2699 /// input is defined, add a check for zero to avoid calling the intrinsic.
2701 /// We want to transform:
2702 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
2706 /// %cmpz = icmp eq i64 %A, 0
2707 /// br i1 %cmpz, label %cond.end, label %cond.false
2709 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
2710 /// br label %cond.end
2712 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
2714 /// If the transform is performed, return true and set ModifiedDT to true.
2715 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
2716 const TargetLowering *TLI,
2717 const DataLayout *DL,
2722 // If a zero input is undefined, it doesn't make sense to despeculate that.
2723 if (match(CountZeros->getOperand(1), m_One()))
2726 // If it's cheap to speculate, there's nothing to do.
2727 auto IntrinsicID = CountZeros->getIntrinsicID();
2728 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
2729 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
2732 // Only handle legal scalar cases. Anything else requires too much work.
2733 Type *Ty = CountZeros->getType();
2734 unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
2735 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSize())
2738 // The intrinsic will be sunk behind a compare against zero and branch.
2739 BasicBlock *StartBlock = CountZeros->getParent();
2740 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
2742 // Create another block after the count zero intrinsic. A PHI will be added
2743 // in this block to select the result of the intrinsic or the bit-width
2744 // constant if the input to the intrinsic is zero.
2745 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
2746 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
2748 // Set up a builder to create a compare, conditional branch, and PHI.
2749 IRBuilder<> Builder(CountZeros->getContext());
2750 Builder.SetInsertPoint(StartBlock->getTerminator());
2751 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
2753 // Replace the unconditional branch that was created by the first split with
2754 // a compare against zero and a conditional branch.
2755 Value *Zero = Constant::getNullValue(Ty);
2756 Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
2757 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
2758 StartBlock->getTerminator()->eraseFromParent();
2760 // Create a PHI in the end block to select either the output of the intrinsic
2761 // or the bit width of the operand.
2762 Builder.SetInsertPoint(&EndBlock->front());
2763 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
2764 CountZeros->replaceAllUsesWith(PN);
2765 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
2766 PN->addIncoming(BitWidth, StartBlock);
2767 PN->addIncoming(CountZeros, CallBlock);
2769 // We are explicitly handling the zero case, so we can set the intrinsic's
2770 // undefined zero argument to 'true'. This will also prevent reprocessing the
2771 // intrinsic; we only despeculate when a zero input is defined.
2772 CountZeros->setArgOperand(1, Builder.getTrue());
2777 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool& ModifiedDT) {
2778 BasicBlock *BB = CI->getParent();
2780 // Lower inline assembly if we can.
2781 // If we found an inline asm expession, and if the target knows how to
2782 // lower it to normal LLVM code, do so now.
2783 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
2784 if (TLI->ExpandInlineAsm(CI)) {
2785 // Avoid invalidating the iterator.
2786 CurInstIterator = BB->begin();
2787 // Avoid processing instructions out of order, which could cause
2788 // reuse before a value is defined.
2792 // Sink address computing for memory operands into the block.
2793 if (optimizeInlineAsmInst(CI))
2797 // Align the pointer arguments to this call if the target thinks it's a good
2799 unsigned MinSize, PrefAlign;
2800 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
2801 for (auto &Arg : CI->arg_operands()) {
2802 // We want to align both objects whose address is used directly and
2803 // objects whose address is used in casts and GEPs, though it only makes
2804 // sense for GEPs if the offset is a multiple of the desired alignment and
2805 // if size - offset meets the size threshold.
2806 if (!Arg->getType()->isPointerTy())
2808 APInt Offset(DL->getPointerSizeInBits(
2809 cast<PointerType>(Arg->getType())->getAddressSpace()),
2811 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
2812 uint64_t Offset2 = Offset.getLimitedValue();
2813 if ((Offset2 & (PrefAlign-1)) != 0)
2816 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
2817 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
2818 AI->setAlignment(PrefAlign);
2819 // Global variables can only be aligned if they are defined in this
2820 // object (i.e. they are uniquely initialized in this object), and
2821 // over-aligning global variables that have an explicit section is
2824 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
2825 GV->getAlignment() < PrefAlign &&
2826 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
2828 GV->setAlignment(PrefAlign);
2830 // If this is a memcpy (or similar) then we may be able to improve the
2832 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
2833 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
2834 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
2835 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
2836 if (Align > MI->getAlignment())
2837 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
2841 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
2843 switch (II->getIntrinsicID()) {
2845 case Intrinsic::objectsize: {
2846 // Lower all uses of llvm.objectsize.*
2847 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
2848 Type *ReturnTy = CI->getType();
2849 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
2851 // Substituting this can cause recursive simplifications, which can
2852 // invalidate our iterator. Use a WeakVH to hold onto it in case this
2854 WeakVH IterHandle(&*CurInstIterator);
2856 replaceAndRecursivelySimplify(CI, RetVal,
2859 // If the iterator instruction was recursively deleted, start over at the
2860 // start of the block.
2861 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
2862 CurInstIterator = BB->begin();
2867 case Intrinsic::masked_load: {
2868 // Scalarize unsupported vector masked load
2869 if (!TTI->isLegalMaskedLoad(CI->getType())) {
2870 ScalarizeMaskedLoad(CI);
2876 case Intrinsic::masked_store: {
2877 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType())) {
2878 ScalarizeMaskedStore(CI);
2884 case Intrinsic::masked_gather: {
2885 if (!TTI->isLegalMaskedGather(CI->getType())) {
2886 ScalarizeMaskedGather(CI);
2892 case Intrinsic::masked_scatter: {
2893 if (!TTI->isLegalMaskedScatter(CI->getArgOperand(0)->getType())) {
2894 ScalarizeMaskedScatter(CI);
2900 case Intrinsic::aarch64_stlxr:
2901 case Intrinsic::aarch64_stxr: {
2902 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
2903 if (!ExtVal || !ExtVal->hasOneUse() ||
2904 ExtVal->getParent() == CI->getParent())
2906 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
2907 ExtVal->moveBefore(CI);
2908 // Mark this instruction as "inserted by CGP", so that other
2909 // optimizations don't touch it.
2910 InsertedInsts.insert(ExtVal);
2913 case Intrinsic::invariant_group_barrier:
2914 II->replaceAllUsesWith(II->getArgOperand(0));
2915 II->eraseFromParent();
2918 case Intrinsic::cttz:
2919 case Intrinsic::ctlz:
2920 // If counting zeros is expensive, try to avoid it.
2921 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
2925 // Unknown address space.
2926 // TODO: Target hook to pick which address space the intrinsic cares
2928 unsigned AddrSpace = ~0u;
2929 SmallVector<Value*, 2> PtrOps;
2931 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
2932 while (!PtrOps.empty())
2933 if (optimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
2938 // From here on out we're working with named functions.
2939 if (!CI->getCalledFunction()) return false;
2941 // Lower all default uses of _chk calls. This is very similar
2942 // to what InstCombineCalls does, but here we are only lowering calls
2943 // to fortified library functions (e.g. __memcpy_chk) that have the default
2944 // "don't know" as the objectsize. Anything else should be left alone.
2945 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
2946 if (Value *V = Simplifier.optimizeCall(CI)) {
2947 CI->replaceAllUsesWith(V);
2948 CI->eraseFromParent();
2954 /// Look for opportunities to duplicate return instructions to the predecessor
2955 /// to enable tail call optimizations. The case it is currently looking for is:
2958 /// %tmp0 = tail call i32 @f0()
2959 /// br label %return
2961 /// %tmp1 = tail call i32 @f1()
2962 /// br label %return
2964 /// %tmp2 = tail call i32 @f2()
2965 /// br label %return
2967 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
2975 /// %tmp0 = tail call i32 @f0()
2978 /// %tmp1 = tail call i32 @f1()
2981 /// %tmp2 = tail call i32 @f2()
2984 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
2988 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
2992 PHINode *PN = nullptr;
2993 BitCastInst *BCI = nullptr;
2994 Value *V = RI->getReturnValue();
2996 BCI = dyn_cast<BitCastInst>(V);
2998 V = BCI->getOperand(0);
3000 PN = dyn_cast<PHINode>(V);
3005 if (PN && PN->getParent() != BB)
3008 // It's not safe to eliminate the sign / zero extension of the return value.
3009 // See llvm::isInTailCallPosition().
3010 const Function *F = BB->getParent();
3011 AttributeSet CallerAttrs = F->getAttributes();
3012 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
3013 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
3016 // Make sure there are no instructions between the PHI and return, or that the
3017 // return is the first instruction in the block.
3019 BasicBlock::iterator BI = BB->begin();
3020 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
3022 // Also skip over the bitcast.
3027 BasicBlock::iterator BI = BB->begin();
3028 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
3033 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
3035 SmallVector<CallInst*, 4> TailCalls;
3037 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
3038 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
3039 // Make sure the phi value is indeed produced by the tail call.
3040 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
3041 TLI->mayBeEmittedAsTailCall(CI))
3042 TailCalls.push_back(CI);
3045 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
3046 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
3047 if (!VisitedBBs.insert(*PI).second)
3050 BasicBlock::InstListType &InstList = (*PI)->getInstList();
3051 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
3052 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
3053 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
3057 CallInst *CI = dyn_cast<CallInst>(&*RI);
3058 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
3059 TailCalls.push_back(CI);
3063 bool Changed = false;
3064 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
3065 CallInst *CI = TailCalls[i];
3068 // Conservatively require the attributes of the call to match those of the
3069 // return. Ignore noalias because it doesn't affect the call sequence.
3070 AttributeSet CalleeAttrs = CS.getAttributes();
3071 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
3072 removeAttribute(Attribute::NoAlias) !=
3073 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
3074 removeAttribute(Attribute::NoAlias))
3077 // Make sure the call instruction is followed by an unconditional branch to
3078 // the return block.
3079 BasicBlock *CallBB = CI->getParent();
3080 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
3081 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
3084 // Duplicate the return into CallBB.
3085 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
3086 ModifiedDT = Changed = true;
3090 // If we eliminated all predecessors of the block, delete the block now.
3091 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
3092 BB->eraseFromParent();
3097 //===----------------------------------------------------------------------===//
3098 // Memory Optimization
3099 //===----------------------------------------------------------------------===//
3103 /// This is an extended version of TargetLowering::AddrMode
3104 /// which holds actual Value*'s for register values.
3105 struct ExtAddrMode : public TargetLowering::AddrMode {
3108 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
3109 void print(raw_ostream &OS) const;
3112 bool operator==(const ExtAddrMode& O) const {
3113 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
3114 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
3115 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
3120 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
3126 void ExtAddrMode::print(raw_ostream &OS) const {
3127 bool NeedPlus = false;
3130 OS << (NeedPlus ? " + " : "")
3132 BaseGV->printAsOperand(OS, /*PrintType=*/false);
3137 OS << (NeedPlus ? " + " : "")
3143 OS << (NeedPlus ? " + " : "")
3145 BaseReg->printAsOperand(OS, /*PrintType=*/false);
3149 OS << (NeedPlus ? " + " : "")
3151 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
3157 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3158 void ExtAddrMode::dump() const {
3164 /// \brief This class provides transaction based operation on the IR.
3165 /// Every change made through this class is recorded in the internal state and
3166 /// can be undone (rollback) until commit is called.
3167 class TypePromotionTransaction {
3169 /// \brief This represents the common interface of the individual transaction.
3170 /// Each class implements the logic for doing one specific modification on
3171 /// the IR via the TypePromotionTransaction.
3172 class TypePromotionAction {
3174 /// The Instruction modified.
3178 /// \brief Constructor of the action.
3179 /// The constructor performs the related action on the IR.
3180 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
3182 virtual ~TypePromotionAction() {}
3184 /// \brief Undo the modification done by this action.
3185 /// When this method is called, the IR must be in the same state as it was
3186 /// before this action was applied.
3187 /// \pre Undoing the action works if and only if the IR is in the exact same
3188 /// state as it was directly after this action was applied.
3189 virtual void undo() = 0;
3191 /// \brief Advocate every change made by this action.
3192 /// When the results on the IR of the action are to be kept, it is important
3193 /// to call this function, otherwise hidden information may be kept forever.
3194 virtual void commit() {
3195 // Nothing to be done, this action is not doing anything.
3199 /// \brief Utility to remember the position of an instruction.
3200 class InsertionHandler {
3201 /// Position of an instruction.
3202 /// Either an instruction:
3203 /// - Is the first in a basic block: BB is used.
3204 /// - Has a previous instructon: PrevInst is used.
3206 Instruction *PrevInst;
3209 /// Remember whether or not the instruction had a previous instruction.
3210 bool HasPrevInstruction;
3213 /// \brief Record the position of \p Inst.
3214 InsertionHandler(Instruction *Inst) {
3215 BasicBlock::iterator It = Inst->getIterator();
3216 HasPrevInstruction = (It != (Inst->getParent()->begin()));
3217 if (HasPrevInstruction)
3218 Point.PrevInst = &*--It;
3220 Point.BB = Inst->getParent();
3223 /// \brief Insert \p Inst at the recorded position.
3224 void insert(Instruction *Inst) {
3225 if (HasPrevInstruction) {
3226 if (Inst->getParent())
3227 Inst->removeFromParent();
3228 Inst->insertAfter(Point.PrevInst);
3230 Instruction *Position = &*Point.BB->getFirstInsertionPt();
3231 if (Inst->getParent())
3232 Inst->moveBefore(Position);
3234 Inst->insertBefore(Position);
3239 /// \brief Move an instruction before another.
3240 class InstructionMoveBefore : public TypePromotionAction {
3241 /// Original position of the instruction.
3242 InsertionHandler Position;
3245 /// \brief Move \p Inst before \p Before.
3246 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
3247 : TypePromotionAction(Inst), Position(Inst) {
3248 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
3249 Inst->moveBefore(Before);
3252 /// \brief Move the instruction back to its original position.
3253 void undo() override {
3254 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
3255 Position.insert(Inst);
3259 /// \brief Set the operand of an instruction with a new value.
3260 class OperandSetter : public TypePromotionAction {
3261 /// Original operand of the instruction.
3263 /// Index of the modified instruction.
3267 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
3268 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
3269 : TypePromotionAction(Inst), Idx(Idx) {
3270 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
3271 << "for:" << *Inst << "\n"
3272 << "with:" << *NewVal << "\n");
3273 Origin = Inst->getOperand(Idx);
3274 Inst->setOperand(Idx, NewVal);
3277 /// \brief Restore the original value of the instruction.
3278 void undo() override {
3279 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
3280 << "for: " << *Inst << "\n"
3281 << "with: " << *Origin << "\n");
3282 Inst->setOperand(Idx, Origin);
3286 /// \brief Hide the operands of an instruction.
3287 /// Do as if this instruction was not using any of its operands.
3288 class OperandsHider : public TypePromotionAction {
3289 /// The list of original operands.
3290 SmallVector<Value *, 4> OriginalValues;
3293 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
3294 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
3295 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
3296 unsigned NumOpnds = Inst->getNumOperands();
3297 OriginalValues.reserve(NumOpnds);
3298 for (unsigned It = 0; It < NumOpnds; ++It) {
3299 // Save the current operand.
3300 Value *Val = Inst->getOperand(It);
3301 OriginalValues.push_back(Val);
3303 // We could use OperandSetter here, but that would imply an overhead
3304 // that we are not willing to pay.
3305 Inst->setOperand(It, UndefValue::get(Val->getType()));
3309 /// \brief Restore the original list of uses.
3310 void undo() override {
3311 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
3312 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
3313 Inst->setOperand(It, OriginalValues[It]);
3317 /// \brief Build a truncate instruction.
3318 class TruncBuilder : public TypePromotionAction {
3321 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
3323 /// trunc Opnd to Ty.
3324 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
3325 IRBuilder<> Builder(Opnd);
3326 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
3327 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
3330 /// \brief Get the built value.
3331 Value *getBuiltValue() { return Val; }
3333 /// \brief Remove the built instruction.
3334 void undo() override {
3335 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
3336 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3337 IVal->eraseFromParent();
3341 /// \brief Build a sign extension instruction.
3342 class SExtBuilder : public TypePromotionAction {
3345 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
3347 /// sext Opnd to Ty.
3348 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3349 : TypePromotionAction(InsertPt) {
3350 IRBuilder<> Builder(InsertPt);
3351 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
3352 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
3355 /// \brief Get the built value.
3356 Value *getBuiltValue() { return Val; }
3358 /// \brief Remove the built instruction.
3359 void undo() override {
3360 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
3361 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3362 IVal->eraseFromParent();
3366 /// \brief Build a zero extension instruction.
3367 class ZExtBuilder : public TypePromotionAction {
3370 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
3372 /// zext Opnd to Ty.
3373 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3374 : TypePromotionAction(InsertPt) {
3375 IRBuilder<> Builder(InsertPt);
3376 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
3377 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
3380 /// \brief Get the built value.
3381 Value *getBuiltValue() { return Val; }
3383 /// \brief Remove the built instruction.
3384 void undo() override {
3385 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
3386 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3387 IVal->eraseFromParent();
3391 /// \brief Mutate an instruction to another type.
3392 class TypeMutator : public TypePromotionAction {
3393 /// Record the original type.
3397 /// \brief Mutate the type of \p Inst into \p NewTy.
3398 TypeMutator(Instruction *Inst, Type *NewTy)
3399 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
3400 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
3402 Inst->mutateType(NewTy);
3405 /// \brief Mutate the instruction back to its original type.
3406 void undo() override {
3407 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
3409 Inst->mutateType(OrigTy);
3413 /// \brief Replace the uses of an instruction by another instruction.
3414 class UsesReplacer : public TypePromotionAction {
3415 /// Helper structure to keep track of the replaced uses.
3416 struct InstructionAndIdx {
3417 /// The instruction using the instruction.
3419 /// The index where this instruction is used for Inst.
3421 InstructionAndIdx(Instruction *Inst, unsigned Idx)
3422 : Inst(Inst), Idx(Idx) {}
3425 /// Keep track of the original uses (pair Instruction, Index).
3426 SmallVector<InstructionAndIdx, 4> OriginalUses;
3427 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
3430 /// \brief Replace all the use of \p Inst by \p New.
3431 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
3432 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
3434 // Record the original uses.
3435 for (Use &U : Inst->uses()) {
3436 Instruction *UserI = cast<Instruction>(U.getUser());
3437 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
3439 // Now, we can replace the uses.
3440 Inst->replaceAllUsesWith(New);
3443 /// \brief Reassign the original uses of Inst to Inst.
3444 void undo() override {
3445 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
3446 for (use_iterator UseIt = OriginalUses.begin(),
3447 EndIt = OriginalUses.end();
3448 UseIt != EndIt; ++UseIt) {
3449 UseIt->Inst->setOperand(UseIt->Idx, Inst);
3454 /// \brief Remove an instruction from the IR.
3455 class InstructionRemover : public TypePromotionAction {
3456 /// Original position of the instruction.
3457 InsertionHandler Inserter;
3458 /// Helper structure to hide all the link to the instruction. In other
3459 /// words, this helps to do as if the instruction was removed.
3460 OperandsHider Hider;
3461 /// Keep track of the uses replaced, if any.
3462 UsesReplacer *Replacer;
3465 /// \brief Remove all reference of \p Inst and optinally replace all its
3467 /// \pre If !Inst->use_empty(), then New != nullptr
3468 InstructionRemover(Instruction *Inst, Value *New = nullptr)
3469 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
3472 Replacer = new UsesReplacer(Inst, New);
3473 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
3474 Inst->removeFromParent();
3477 ~InstructionRemover() override { delete Replacer; }
3479 /// \brief Really remove the instruction.
3480 void commit() override { delete Inst; }
3482 /// \brief Resurrect the instruction and reassign it to the proper uses if
3483 /// new value was provided when build this action.
3484 void undo() override {
3485 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
3486 Inserter.insert(Inst);
3494 /// Restoration point.
3495 /// The restoration point is a pointer to an action instead of an iterator
3496 /// because the iterator may be invalidated but not the pointer.
3497 typedef const TypePromotionAction *ConstRestorationPt;
3498 /// Advocate every changes made in that transaction.
3500 /// Undo all the changes made after the given point.
3501 void rollback(ConstRestorationPt Point);
3502 /// Get the current restoration point.
3503 ConstRestorationPt getRestorationPoint() const;
3505 /// \name API for IR modification with state keeping to support rollback.
3507 /// Same as Instruction::setOperand.
3508 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
3509 /// Same as Instruction::eraseFromParent.
3510 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
3511 /// Same as Value::replaceAllUsesWith.
3512 void replaceAllUsesWith(Instruction *Inst, Value *New);
3513 /// Same as Value::mutateType.
3514 void mutateType(Instruction *Inst, Type *NewTy);
3515 /// Same as IRBuilder::createTrunc.
3516 Value *createTrunc(Instruction *Opnd, Type *Ty);
3517 /// Same as IRBuilder::createSExt.
3518 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
3519 /// Same as IRBuilder::createZExt.
3520 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
3521 /// Same as Instruction::moveBefore.
3522 void moveBefore(Instruction *Inst, Instruction *Before);
3526 /// The ordered list of actions made so far.
3527 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
3528 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
3531 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
3534 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
3537 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
3540 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
3543 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
3545 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
3548 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
3549 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
3552 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
3554 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
3555 Value *Val = Ptr->getBuiltValue();
3556 Actions.push_back(std::move(Ptr));
3560 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
3561 Value *Opnd, Type *Ty) {
3562 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
3563 Value *Val = Ptr->getBuiltValue();
3564 Actions.push_back(std::move(Ptr));
3568 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
3569 Value *Opnd, Type *Ty) {
3570 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
3571 Value *Val = Ptr->getBuiltValue();
3572 Actions.push_back(std::move(Ptr));
3576 void TypePromotionTransaction::moveBefore(Instruction *Inst,
3577 Instruction *Before) {
3579 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
3582 TypePromotionTransaction::ConstRestorationPt
3583 TypePromotionTransaction::getRestorationPoint() const {
3584 return !Actions.empty() ? Actions.back().get() : nullptr;
3587 void TypePromotionTransaction::commit() {
3588 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
3594 void TypePromotionTransaction::rollback(
3595 TypePromotionTransaction::ConstRestorationPt Point) {
3596 while (!Actions.empty() && Point != Actions.back().get()) {
3597 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
3602 /// \brief A helper class for matching addressing modes.
3604 /// This encapsulates the logic for matching the target-legal addressing modes.
3605 class AddressingModeMatcher {
3606 SmallVectorImpl<Instruction*> &AddrModeInsts;
3607 const TargetMachine &TM;
3608 const TargetLowering &TLI;
3609 const DataLayout &DL;
3611 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
3612 /// the memory instruction that we're computing this address for.
3615 Instruction *MemoryInst;
3617 /// This is the addressing mode that we're building up. This is
3618 /// part of the return value of this addressing mode matching stuff.
3619 ExtAddrMode &AddrMode;
3621 /// The instructions inserted by other CodeGenPrepare optimizations.
3622 const SetOfInstrs &InsertedInsts;
3623 /// A map from the instructions to their type before promotion.
3624 InstrToOrigTy &PromotedInsts;
3625 /// The ongoing transaction where every action should be registered.
3626 TypePromotionTransaction &TPT;
3628 /// This is set to true when we should not do profitability checks.
3629 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
3630 bool IgnoreProfitability;
3632 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
3633 const TargetMachine &TM, Type *AT, unsigned AS,
3634 Instruction *MI, ExtAddrMode &AM,
3635 const SetOfInstrs &InsertedInsts,
3636 InstrToOrigTy &PromotedInsts,
3637 TypePromotionTransaction &TPT)
3638 : AddrModeInsts(AMI), TM(TM),
3639 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
3640 ->getTargetLowering()),
3641 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
3642 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
3643 PromotedInsts(PromotedInsts), TPT(TPT) {
3644 IgnoreProfitability = false;
3648 /// Find the maximal addressing mode that a load/store of V can fold,
3649 /// give an access type of AccessTy. This returns a list of involved
3650 /// instructions in AddrModeInsts.
3651 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
3653 /// \p PromotedInsts maps the instructions to their type before promotion.
3654 /// \p The ongoing transaction where every action should be registered.
3655 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
3656 Instruction *MemoryInst,
3657 SmallVectorImpl<Instruction*> &AddrModeInsts,
3658 const TargetMachine &TM,
3659 const SetOfInstrs &InsertedInsts,
3660 InstrToOrigTy &PromotedInsts,
3661 TypePromotionTransaction &TPT) {
3664 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
3665 MemoryInst, Result, InsertedInsts,
3666 PromotedInsts, TPT).matchAddr(V, 0);
3667 (void)Success; assert(Success && "Couldn't select *anything*?");
3671 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
3672 bool matchAddr(Value *V, unsigned Depth);
3673 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
3674 bool *MovedAway = nullptr);
3675 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
3676 ExtAddrMode &AMBefore,
3677 ExtAddrMode &AMAfter);
3678 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
3679 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
3680 Value *PromotedOperand) const;
3683 /// Try adding ScaleReg*Scale to the current addressing mode.
3684 /// Return true and update AddrMode if this addr mode is legal for the target,
3686 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3688 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3689 // mode. Just process that directly.
3691 return matchAddr(ScaleReg, Depth);
3693 // If the scale is 0, it takes nothing to add this.
3697 // If we already have a scale of this value, we can add to it, otherwise, we
3698 // need an available scale field.
3699 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3702 ExtAddrMode TestAddrMode = AddrMode;
3704 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3705 // [A+B + A*7] -> [B+A*8].
3706 TestAddrMode.Scale += Scale;
3707 TestAddrMode.ScaledReg = ScaleReg;
3709 // If the new address isn't legal, bail out.
3710 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3713 // It was legal, so commit it.
3714 AddrMode = TestAddrMode;
3716 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3717 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3718 // X*Scale + C*Scale to addr mode.
3719 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3720 if (isa<Instruction>(ScaleReg) && // not a constant expr.
3721 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3722 TestAddrMode.ScaledReg = AddLHS;
3723 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3725 // If this addressing mode is legal, commit it and remember that we folded
3726 // this instruction.
3727 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3728 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3729 AddrMode = TestAddrMode;
3734 // Otherwise, not (x+c)*scale, just return what we have.
3738 /// This is a little filter, which returns true if an addressing computation
3739 /// involving I might be folded into a load/store accessing it.
3740 /// This doesn't need to be perfect, but needs to accept at least
3741 /// the set of instructions that MatchOperationAddr can.
3742 static bool MightBeFoldableInst(Instruction *I) {
3743 switch (I->getOpcode()) {
3744 case Instruction::BitCast:
3745 case Instruction::AddrSpaceCast:
3746 // Don't touch identity bitcasts.
3747 if (I->getType() == I->getOperand(0)->getType())
3749 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
3750 case Instruction::PtrToInt:
3751 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3753 case Instruction::IntToPtr:
3754 // We know the input is intptr_t, so this is foldable.
3756 case Instruction::Add:
3758 case Instruction::Mul:
3759 case Instruction::Shl:
3760 // Can only handle X*C and X << C.
3761 return isa<ConstantInt>(I->getOperand(1));
3762 case Instruction::GetElementPtr:
3769 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
3770 /// \note \p Val is assumed to be the product of some type promotion.
3771 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3772 /// to be legal, as the non-promoted value would have had the same state.
3773 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
3774 const DataLayout &DL, Value *Val) {
3775 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3778 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3779 // If the ISDOpcode is undefined, it was undefined before the promotion.
3782 // Otherwise, check if the promoted instruction is legal or not.
3783 return TLI.isOperationLegalOrCustom(
3784 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3787 /// \brief Hepler class to perform type promotion.
3788 class TypePromotionHelper {
3789 /// \brief Utility function to check whether or not a sign or zero extension
3790 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3791 /// either using the operands of \p Inst or promoting \p Inst.
3792 /// The type of the extension is defined by \p IsSExt.
3793 /// In other words, check if:
3794 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3795 /// #1 Promotion applies:
3796 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3797 /// #2 Operand reuses:
3798 /// ext opnd1 to ConsideredExtType.
3799 /// \p PromotedInsts maps the instructions to their type before promotion.
3800 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3801 const InstrToOrigTy &PromotedInsts, bool IsSExt);
3803 /// \brief Utility function to determine if \p OpIdx should be promoted when
3804 /// promoting \p Inst.
3805 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3806 return !(isa<SelectInst>(Inst) && OpIdx == 0);
3809 /// \brief Utility function to promote the operand of \p Ext when this
3810 /// operand is a promotable trunc or sext or zext.
3811 /// \p PromotedInsts maps the instructions to their type before promotion.
3812 /// \p CreatedInstsCost[out] contains the cost of all instructions
3813 /// created to promote the operand of Ext.
3814 /// Newly added extensions are inserted in \p Exts.
3815 /// Newly added truncates are inserted in \p Truncs.
3816 /// Should never be called directly.
3817 /// \return The promoted value which is used instead of Ext.
3818 static Value *promoteOperandForTruncAndAnyExt(
3819 Instruction *Ext, TypePromotionTransaction &TPT,
3820 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3821 SmallVectorImpl<Instruction *> *Exts,
3822 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3824 /// \brief Utility function to promote the operand of \p Ext when this
3825 /// operand is promotable and is not a supported trunc or sext.
3826 /// \p PromotedInsts maps the instructions to their type before promotion.
3827 /// \p CreatedInstsCost[out] contains the cost of all the instructions
3828 /// created to promote the operand of Ext.
3829 /// Newly added extensions are inserted in \p Exts.
3830 /// Newly added truncates are inserted in \p Truncs.
3831 /// Should never be called directly.
3832 /// \return The promoted value which is used instead of Ext.
3833 static Value *promoteOperandForOther(Instruction *Ext,
3834 TypePromotionTransaction &TPT,
3835 InstrToOrigTy &PromotedInsts,
3836 unsigned &CreatedInstsCost,
3837 SmallVectorImpl<Instruction *> *Exts,
3838 SmallVectorImpl<Instruction *> *Truncs,
3839 const TargetLowering &TLI, bool IsSExt);
3841 /// \see promoteOperandForOther.
3842 static Value *signExtendOperandForOther(
3843 Instruction *Ext, TypePromotionTransaction &TPT,
3844 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3845 SmallVectorImpl<Instruction *> *Exts,
3846 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3847 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3848 Exts, Truncs, TLI, true);
3851 /// \see promoteOperandForOther.
3852 static Value *zeroExtendOperandForOther(
3853 Instruction *Ext, TypePromotionTransaction &TPT,
3854 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3855 SmallVectorImpl<Instruction *> *Exts,
3856 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3857 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3858 Exts, Truncs, TLI, false);
3862 /// Type for the utility function that promotes the operand of Ext.
3863 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
3864 InstrToOrigTy &PromotedInsts,
3865 unsigned &CreatedInstsCost,
3866 SmallVectorImpl<Instruction *> *Exts,
3867 SmallVectorImpl<Instruction *> *Truncs,
3868 const TargetLowering &TLI);
3869 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
3870 /// action to promote the operand of \p Ext instead of using Ext.
3871 /// \return NULL if no promotable action is possible with the current
3873 /// \p InsertedInsts keeps track of all the instructions inserted by the
3874 /// other CodeGenPrepare optimizations. This information is important
3875 /// because we do not want to promote these instructions as CodeGenPrepare
3876 /// will reinsert them later. Thus creating an infinite loop: create/remove.
3877 /// \p PromotedInsts maps the instructions to their type before promotion.
3878 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3879 const TargetLowering &TLI,
3880 const InstrToOrigTy &PromotedInsts);
3883 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3884 Type *ConsideredExtType,
3885 const InstrToOrigTy &PromotedInsts,
3887 // The promotion helper does not know how to deal with vector types yet.
3888 // To be able to fix that, we would need to fix the places where we
3889 // statically extend, e.g., constants and such.
3890 if (Inst->getType()->isVectorTy())
3893 // We can always get through zext.
3894 if (isa<ZExtInst>(Inst))
3897 // sext(sext) is ok too.
3898 if (IsSExt && isa<SExtInst>(Inst))
3901 // We can get through binary operator, if it is legal. In other words, the
3902 // binary operator must have a nuw or nsw flag.
3903 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3904 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3905 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3906 (IsSExt && BinOp->hasNoSignedWrap())))
3909 // Check if we can do the following simplification.
3910 // ext(trunc(opnd)) --> ext(opnd)
3911 if (!isa<TruncInst>(Inst))
3914 Value *OpndVal = Inst->getOperand(0);
3915 // Check if we can use this operand in the extension.
3916 // If the type is larger than the result type of the extension, we cannot.
3917 if (!OpndVal->getType()->isIntegerTy() ||
3918 OpndVal->getType()->getIntegerBitWidth() >
3919 ConsideredExtType->getIntegerBitWidth())
3922 // If the operand of the truncate is not an instruction, we will not have
3923 // any information on the dropped bits.
3924 // (Actually we could for constant but it is not worth the extra logic).
3925 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3929 // Check if the source of the type is narrow enough.
3930 // I.e., check that trunc just drops extended bits of the same kind of
3932 // #1 get the type of the operand and check the kind of the extended bits.
3933 const Type *OpndType;
3934 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3935 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
3936 OpndType = It->second.getPointer();
3937 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3938 OpndType = Opnd->getOperand(0)->getType();
3942 // #2 check that the truncate just drops extended bits.
3943 return Inst->getType()->getIntegerBitWidth() >=
3944 OpndType->getIntegerBitWidth();
3947 TypePromotionHelper::Action TypePromotionHelper::getAction(
3948 Instruction *Ext, const SetOfInstrs &InsertedInsts,
3949 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3950 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
3951 "Unexpected instruction type");
3952 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3953 Type *ExtTy = Ext->getType();
3954 bool IsSExt = isa<SExtInst>(Ext);
3955 // If the operand of the extension is not an instruction, we cannot
3957 // If it, check we can get through.
3958 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3961 // Do not promote if the operand has been added by codegenprepare.
3962 // Otherwise, it means we are undoing an optimization that is likely to be
3963 // redone, thus causing potential infinite loop.
3964 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3967 // SExt or Trunc instructions.
3968 // Return the related handler.
3969 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3970 isa<ZExtInst>(ExtOpnd))
3971 return promoteOperandForTruncAndAnyExt;
3973 // Regular instruction.
3974 // Abort early if we will have to insert non-free instructions.
3975 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3977 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3980 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3981 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
3982 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3983 SmallVectorImpl<Instruction *> *Exts,
3984 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3985 // By construction, the operand of SExt is an instruction. Otherwise we cannot
3986 // get through it and this method should not be called.
3987 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
3988 Value *ExtVal = SExt;
3989 bool HasMergedNonFreeExt = false;
3990 if (isa<ZExtInst>(SExtOpnd)) {
3991 // Replace s|zext(zext(opnd))
3993 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
3995 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
3996 TPT.replaceAllUsesWith(SExt, ZExt);
3997 TPT.eraseInstruction(SExt);
4000 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
4002 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
4004 CreatedInstsCost = 0;
4006 // Remove dead code.
4007 if (SExtOpnd->use_empty())
4008 TPT.eraseInstruction(SExtOpnd);
4010 // Check if the extension is still needed.
4011 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
4012 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
4015 Exts->push_back(ExtInst);
4016 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
4021 // At this point we have: ext ty opnd to ty.
4022 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
4023 Value *NextVal = ExtInst->getOperand(0);
4024 TPT.eraseInstruction(ExtInst, NextVal);
4028 Value *TypePromotionHelper::promoteOperandForOther(
4029 Instruction *Ext, TypePromotionTransaction &TPT,
4030 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4031 SmallVectorImpl<Instruction *> *Exts,
4032 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
4034 // By construction, the operand of Ext is an instruction. Otherwise we cannot
4035 // get through it and this method should not be called.
4036 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
4037 CreatedInstsCost = 0;
4038 if (!ExtOpnd->hasOneUse()) {
4039 // ExtOpnd will be promoted.
4040 // All its uses, but Ext, will need to use a truncated value of the
4041 // promoted version.
4042 // Create the truncate now.
4043 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
4044 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
4045 ITrunc->removeFromParent();
4046 // Insert it just after the definition.
4047 ITrunc->insertAfter(ExtOpnd);
4049 Truncs->push_back(ITrunc);
4052 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
4053 // Restore the operand of Ext (which has been replaced by the previous call
4054 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
4055 TPT.setOperand(Ext, 0, ExtOpnd);
4058 // Get through the Instruction:
4059 // 1. Update its type.
4060 // 2. Replace the uses of Ext by Inst.
4061 // 3. Extend each operand that needs to be extended.
4063 // Remember the original type of the instruction before promotion.
4064 // This is useful to know that the high bits are sign extended bits.
4065 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
4066 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
4068 TPT.mutateType(ExtOpnd, Ext->getType());
4070 TPT.replaceAllUsesWith(Ext, ExtOpnd);
4072 Instruction *ExtForOpnd = Ext;
4074 DEBUG(dbgs() << "Propagate Ext to operands\n");
4075 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
4077 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
4078 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
4079 !shouldExtOperand(ExtOpnd, OpIdx)) {
4080 DEBUG(dbgs() << "No need to propagate\n");
4083 // Check if we can statically extend the operand.
4084 Value *Opnd = ExtOpnd->getOperand(OpIdx);
4085 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
4086 DEBUG(dbgs() << "Statically extend\n");
4087 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
4088 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
4089 : Cst->getValue().zext(BitWidth);
4090 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
4093 // UndefValue are typed, so we have to statically sign extend them.
4094 if (isa<UndefValue>(Opnd)) {
4095 DEBUG(dbgs() << "Statically extend\n");
4096 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
4100 // Otherwise we have to explicity sign extend the operand.
4101 // Check if Ext was reused to extend an operand.
4103 // If yes, create a new one.
4104 DEBUG(dbgs() << "More operands to ext\n");
4105 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
4106 : TPT.createZExt(Ext, Opnd, Ext->getType());
4107 if (!isa<Instruction>(ValForExtOpnd)) {
4108 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
4111 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
4114 Exts->push_back(ExtForOpnd);
4115 TPT.setOperand(ExtForOpnd, 0, Opnd);
4117 // Move the sign extension before the insertion point.
4118 TPT.moveBefore(ExtForOpnd, ExtOpnd);
4119 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
4120 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
4121 // If more sext are required, new instructions will have to be created.
4122 ExtForOpnd = nullptr;
4124 if (ExtForOpnd == Ext) {
4125 DEBUG(dbgs() << "Extension is useless now\n");
4126 TPT.eraseInstruction(Ext);
4131 /// Check whether or not promoting an instruction to a wider type is profitable.
4132 /// \p NewCost gives the cost of extension instructions created by the
4134 /// \p OldCost gives the cost of extension instructions before the promotion
4135 /// plus the number of instructions that have been
4136 /// matched in the addressing mode the promotion.
4137 /// \p PromotedOperand is the value that has been promoted.
4138 /// \return True if the promotion is profitable, false otherwise.
4139 bool AddressingModeMatcher::isPromotionProfitable(
4140 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
4141 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
4142 // The cost of the new extensions is greater than the cost of the
4143 // old extension plus what we folded.
4144 // This is not profitable.
4145 if (NewCost > OldCost)
4147 if (NewCost < OldCost)
4149 // The promotion is neutral but it may help folding the sign extension in
4150 // loads for instance.
4151 // Check that we did not create an illegal instruction.
4152 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
4155 /// Given an instruction or constant expr, see if we can fold the operation
4156 /// into the addressing mode. If so, update the addressing mode and return
4157 /// true, otherwise return false without modifying AddrMode.
4158 /// If \p MovedAway is not NULL, it contains the information of whether or
4159 /// not AddrInst has to be folded into the addressing mode on success.
4160 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
4161 /// because it has been moved away.
4162 /// Thus AddrInst must not be added in the matched instructions.
4163 /// This state can happen when AddrInst is a sext, since it may be moved away.
4164 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
4165 /// not be referenced anymore.
4166 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
4169 // Avoid exponential behavior on extremely deep expression trees.
4170 if (Depth >= 5) return false;
4172 // By default, all matched instructions stay in place.
4177 case Instruction::PtrToInt:
4178 // PtrToInt is always a noop, as we know that the int type is pointer sized.
4179 return matchAddr(AddrInst->getOperand(0), Depth);
4180 case Instruction::IntToPtr: {
4181 auto AS = AddrInst->getType()->getPointerAddressSpace();
4182 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
4183 // This inttoptr is a no-op if the integer type is pointer sized.
4184 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
4185 return matchAddr(AddrInst->getOperand(0), Depth);
4188 case Instruction::BitCast:
4189 // BitCast is always a noop, and we can handle it as long as it is
4190 // int->int or pointer->pointer (we don't want int<->fp or something).
4191 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
4192 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
4193 // Don't touch identity bitcasts. These were probably put here by LSR,
4194 // and we don't want to mess around with them. Assume it knows what it
4196 AddrInst->getOperand(0)->getType() != AddrInst->getType())
4197 return matchAddr(AddrInst->getOperand(0), Depth);
4199 case Instruction::AddrSpaceCast: {
4201 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
4202 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
4203 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
4204 return matchAddr(AddrInst->getOperand(0), Depth);
4207 case Instruction::Add: {
4208 // Check to see if we can merge in the RHS then the LHS. If so, we win.
4209 ExtAddrMode BackupAddrMode = AddrMode;
4210 unsigned OldSize = AddrModeInsts.size();
4211 // Start a transaction at this point.
4212 // The LHS may match but not the RHS.
4213 // Therefore, we need a higher level restoration point to undo partially
4214 // matched operation.
4215 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4216 TPT.getRestorationPoint();
4218 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
4219 matchAddr(AddrInst->getOperand(0), Depth+1))
4222 // Restore the old addr mode info.
4223 AddrMode = BackupAddrMode;
4224 AddrModeInsts.resize(OldSize);
4225 TPT.rollback(LastKnownGood);
4227 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
4228 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
4229 matchAddr(AddrInst->getOperand(1), Depth+1))
4232 // Otherwise we definitely can't merge the ADD in.
4233 AddrMode = BackupAddrMode;
4234 AddrModeInsts.resize(OldSize);
4235 TPT.rollback(LastKnownGood);
4238 //case Instruction::Or:
4239 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
4241 case Instruction::Mul:
4242 case Instruction::Shl: {
4243 // Can only handle X*C and X << C.
4244 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
4247 int64_t Scale = RHS->getSExtValue();
4248 if (Opcode == Instruction::Shl)
4249 Scale = 1LL << Scale;
4251 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
4253 case Instruction::GetElementPtr: {
4254 // Scan the GEP. We check it if it contains constant offsets and at most
4255 // one variable offset.
4256 int VariableOperand = -1;
4257 unsigned VariableScale = 0;
4259 int64_t ConstantOffset = 0;
4260 gep_type_iterator GTI = gep_type_begin(AddrInst);
4261 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
4262 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
4263 const StructLayout *SL = DL.getStructLayout(STy);
4265 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
4266 ConstantOffset += SL->getElementOffset(Idx);
4268 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
4269 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
4270 ConstantOffset += CI->getSExtValue()*TypeSize;
4271 } else if (TypeSize) { // Scales of zero don't do anything.
4272 // We only allow one variable index at the moment.
4273 if (VariableOperand != -1)
4276 // Remember the variable index.
4277 VariableOperand = i;
4278 VariableScale = TypeSize;
4283 // A common case is for the GEP to only do a constant offset. In this case,
4284 // just add it to the disp field and check validity.
4285 if (VariableOperand == -1) {
4286 AddrMode.BaseOffs += ConstantOffset;
4287 if (ConstantOffset == 0 ||
4288 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
4289 // Check to see if we can fold the base pointer in too.
4290 if (matchAddr(AddrInst->getOperand(0), Depth+1))
4293 AddrMode.BaseOffs -= ConstantOffset;
4297 // Save the valid addressing mode in case we can't match.
4298 ExtAddrMode BackupAddrMode = AddrMode;
4299 unsigned OldSize = AddrModeInsts.size();
4301 // See if the scale and offset amount is valid for this target.
4302 AddrMode.BaseOffs += ConstantOffset;
4304 // Match the base operand of the GEP.
4305 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
4306 // If it couldn't be matched, just stuff the value in a register.
4307 if (AddrMode.HasBaseReg) {
4308 AddrMode = BackupAddrMode;
4309 AddrModeInsts.resize(OldSize);
4312 AddrMode.HasBaseReg = true;
4313 AddrMode.BaseReg = AddrInst->getOperand(0);
4316 // Match the remaining variable portion of the GEP.
4317 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
4319 // If it couldn't be matched, try stuffing the base into a register
4320 // instead of matching it, and retrying the match of the scale.
4321 AddrMode = BackupAddrMode;
4322 AddrModeInsts.resize(OldSize);
4323 if (AddrMode.HasBaseReg)
4325 AddrMode.HasBaseReg = true;
4326 AddrMode.BaseReg = AddrInst->getOperand(0);
4327 AddrMode.BaseOffs += ConstantOffset;
4328 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
4329 VariableScale, Depth)) {
4330 // If even that didn't work, bail.
4331 AddrMode = BackupAddrMode;
4332 AddrModeInsts.resize(OldSize);
4339 case Instruction::SExt:
4340 case Instruction::ZExt: {
4341 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
4345 // Try to move this ext out of the way of the addressing mode.
4346 // Ask for a method for doing so.
4347 TypePromotionHelper::Action TPH =
4348 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
4352 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4353 TPT.getRestorationPoint();
4354 unsigned CreatedInstsCost = 0;
4355 unsigned ExtCost = !TLI.isExtFree(Ext);
4356 Value *PromotedOperand =
4357 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
4358 // SExt has been moved away.
4359 // Thus either it will be rematched later in the recursive calls or it is
4360 // gone. Anyway, we must not fold it into the addressing mode at this point.
4364 // addr = gep base, idx
4366 // promotedOpnd = ext opnd <- no match here
4367 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
4368 // addr = gep base, op <- match
4372 assert(PromotedOperand &&
4373 "TypePromotionHelper should have filtered out those cases");
4375 ExtAddrMode BackupAddrMode = AddrMode;
4376 unsigned OldSize = AddrModeInsts.size();
4378 if (!matchAddr(PromotedOperand, Depth) ||
4379 // The total of the new cost is equal to the cost of the created
4381 // The total of the old cost is equal to the cost of the extension plus
4382 // what we have saved in the addressing mode.
4383 !isPromotionProfitable(CreatedInstsCost,
4384 ExtCost + (AddrModeInsts.size() - OldSize),
4386 AddrMode = BackupAddrMode;
4387 AddrModeInsts.resize(OldSize);
4388 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
4389 TPT.rollback(LastKnownGood);
4398 /// If we can, try to add the value of 'Addr' into the current addressing mode.
4399 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
4400 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
4403 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
4404 // Start a transaction at this point that we will rollback if the matching
4406 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4407 TPT.getRestorationPoint();
4408 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
4409 // Fold in immediates if legal for the target.
4410 AddrMode.BaseOffs += CI->getSExtValue();
4411 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4413 AddrMode.BaseOffs -= CI->getSExtValue();
4414 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
4415 // If this is a global variable, try to fold it into the addressing mode.
4416 if (!AddrMode.BaseGV) {
4417 AddrMode.BaseGV = GV;
4418 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4420 AddrMode.BaseGV = nullptr;
4422 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
4423 ExtAddrMode BackupAddrMode = AddrMode;
4424 unsigned OldSize = AddrModeInsts.size();
4426 // Check to see if it is possible to fold this operation.
4427 bool MovedAway = false;
4428 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
4429 // This instruction may have been moved away. If so, there is nothing
4433 // Okay, it's possible to fold this. Check to see if it is actually
4434 // *profitable* to do so. We use a simple cost model to avoid increasing
4435 // register pressure too much.
4436 if (I->hasOneUse() ||
4437 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
4438 AddrModeInsts.push_back(I);
4442 // It isn't profitable to do this, roll back.
4443 //cerr << "NOT FOLDING: " << *I;
4444 AddrMode = BackupAddrMode;
4445 AddrModeInsts.resize(OldSize);
4446 TPT.rollback(LastKnownGood);
4448 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
4449 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
4451 TPT.rollback(LastKnownGood);
4452 } else if (isa<ConstantPointerNull>(Addr)) {
4453 // Null pointer gets folded without affecting the addressing mode.
4457 // Worse case, the target should support [reg] addressing modes. :)
4458 if (!AddrMode.HasBaseReg) {
4459 AddrMode.HasBaseReg = true;
4460 AddrMode.BaseReg = Addr;
4461 // Still check for legality in case the target supports [imm] but not [i+r].
4462 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4464 AddrMode.HasBaseReg = false;
4465 AddrMode.BaseReg = nullptr;
4468 // If the base register is already taken, see if we can do [r+r].
4469 if (AddrMode.Scale == 0) {
4471 AddrMode.ScaledReg = Addr;
4472 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4475 AddrMode.ScaledReg = nullptr;
4478 TPT.rollback(LastKnownGood);
4482 /// Check to see if all uses of OpVal by the specified inline asm call are due
4483 /// to memory operands. If so, return true, otherwise return false.
4484 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
4485 const TargetMachine &TM) {
4486 const Function *F = CI->getParent()->getParent();
4487 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
4488 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
4489 TargetLowering::AsmOperandInfoVector TargetConstraints =
4490 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
4491 ImmutableCallSite(CI));
4492 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4493 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4495 // Compute the constraint code and ConstraintType to use.
4496 TLI->ComputeConstraintToUse(OpInfo, SDValue());
4498 // If this asm operand is our Value*, and if it isn't an indirect memory
4499 // operand, we can't fold it!
4500 if (OpInfo.CallOperandVal == OpVal &&
4501 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
4502 !OpInfo.isIndirect))
4509 /// Recursively walk all the uses of I until we find a memory use.
4510 /// If we find an obviously non-foldable instruction, return true.
4511 /// Add the ultimately found memory instructions to MemoryUses.
4512 static bool FindAllMemoryUses(
4514 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
4515 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
4516 // If we already considered this instruction, we're done.
4517 if (!ConsideredInsts.insert(I).second)
4520 // If this is an obviously unfoldable instruction, bail out.
4521 if (!MightBeFoldableInst(I))
4524 // Loop over all the uses, recursively processing them.
4525 for (Use &U : I->uses()) {
4526 Instruction *UserI = cast<Instruction>(U.getUser());
4528 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4529 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
4533 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4534 unsigned opNo = U.getOperandNo();
4535 if (opNo == 0) return true; // Storing addr, not into addr.
4536 MemoryUses.push_back(std::make_pair(SI, opNo));
4540 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4541 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
4542 if (!IA) return true;
4544 // If this is a memory operand, we're cool, otherwise bail out.
4545 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
4550 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
4557 /// Return true if Val is already known to be live at the use site that we're
4558 /// folding it into. If so, there is no cost to include it in the addressing
4559 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4560 /// instruction already.
4561 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4562 Value *KnownLive2) {
4563 // If Val is either of the known-live values, we know it is live!
4564 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4567 // All values other than instructions and arguments (e.g. constants) are live.
4568 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4570 // If Val is a constant sized alloca in the entry block, it is live, this is
4571 // true because it is just a reference to the stack/frame pointer, which is
4572 // live for the whole function.
4573 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4574 if (AI->isStaticAlloca())
4577 // Check to see if this value is already used in the memory instruction's
4578 // block. If so, it's already live into the block at the very least, so we
4579 // can reasonably fold it.
4580 return Val->isUsedInBasicBlock(MemoryInst->getParent());
4583 /// It is possible for the addressing mode of the machine to fold the specified
4584 /// instruction into a load or store that ultimately uses it.
4585 /// However, the specified instruction has multiple uses.
4586 /// Given this, it may actually increase register pressure to fold it
4587 /// into the load. For example, consider this code:
4591 /// use(Y) -> nonload/store
4595 /// In this case, Y has multiple uses, and can be folded into the load of Z
4596 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4597 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4598 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4599 /// number of computations either.
4601 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4602 /// X was live across 'load Z' for other reasons, we actually *would* want to
4603 /// fold the addressing mode in the Z case. This would make Y die earlier.
4604 bool AddressingModeMatcher::
4605 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4606 ExtAddrMode &AMAfter) {
4607 if (IgnoreProfitability) return true;
4609 // AMBefore is the addressing mode before this instruction was folded into it,
4610 // and AMAfter is the addressing mode after the instruction was folded. Get
4611 // the set of registers referenced by AMAfter and subtract out those
4612 // referenced by AMBefore: this is the set of values which folding in this
4613 // address extends the lifetime of.
4615 // Note that there are only two potential values being referenced here,
4616 // BaseReg and ScaleReg (global addresses are always available, as are any
4617 // folded immediates).
4618 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4620 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4621 // lifetime wasn't extended by adding this instruction.
4622 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4624 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4625 ScaledReg = nullptr;
4627 // If folding this instruction (and it's subexprs) didn't extend any live
4628 // ranges, we're ok with it.
4629 if (!BaseReg && !ScaledReg)
4632 // If all uses of this instruction are ultimately load/store/inlineasm's,
4633 // check to see if their addressing modes will include this instruction. If
4634 // so, we can fold it into all uses, so it doesn't matter if it has multiple
4636 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
4637 SmallPtrSet<Instruction*, 16> ConsideredInsts;
4638 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
4639 return false; // Has a non-memory, non-foldable use!
4641 // Now that we know that all uses of this instruction are part of a chain of
4642 // computation involving only operations that could theoretically be folded
4643 // into a memory use, loop over each of these uses and see if they could
4644 // *actually* fold the instruction.
4645 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4646 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4647 Instruction *User = MemoryUses[i].first;
4648 unsigned OpNo = MemoryUses[i].second;
4650 // Get the access type of this use. If the use isn't a pointer, we don't
4651 // know what it accesses.
4652 Value *Address = User->getOperand(OpNo);
4653 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4656 Type *AddressAccessTy = AddrTy->getElementType();
4657 unsigned AS = AddrTy->getAddressSpace();
4659 // Do a match against the root of this address, ignoring profitability. This
4660 // will tell us if the addressing mode for the memory operation will
4661 // *actually* cover the shared instruction.
4663 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4664 TPT.getRestorationPoint();
4665 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
4666 MemoryInst, Result, InsertedInsts,
4667 PromotedInsts, TPT);
4668 Matcher.IgnoreProfitability = true;
4669 bool Success = Matcher.matchAddr(Address, 0);
4670 (void)Success; assert(Success && "Couldn't select *anything*?");
4672 // The match was to check the profitability, the changes made are not
4673 // part of the original matcher. Therefore, they should be dropped
4674 // otherwise the original matcher will not present the right state.
4675 TPT.rollback(LastKnownGood);
4677 // If the match didn't cover I, then it won't be shared by it.
4678 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
4679 I) == MatchedAddrModeInsts.end())
4682 MatchedAddrModeInsts.clear();
4688 } // end anonymous namespace
4690 /// Return true if the specified values are defined in a
4691 /// different basic block than BB.
4692 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4693 if (Instruction *I = dyn_cast<Instruction>(V))
4694 return I->getParent() != BB;
4698 /// Load and Store Instructions often have addressing modes that can do
4699 /// significant amounts of computation. As such, instruction selection will try
4700 /// to get the load or store to do as much computation as possible for the
4701 /// program. The problem is that isel can only see within a single block. As
4702 /// such, we sink as much legal addressing mode work into the block as possible.
4704 /// This method is used to optimize both load/store and inline asms with memory
4706 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4707 Type *AccessTy, unsigned AddrSpace) {
4710 // Try to collapse single-value PHI nodes. This is necessary to undo
4711 // unprofitable PRE transformations.
4712 SmallVector<Value*, 8> worklist;
4713 SmallPtrSet<Value*, 16> Visited;
4714 worklist.push_back(Addr);
4716 // Use a worklist to iteratively look through PHI nodes, and ensure that
4717 // the addressing mode obtained from the non-PHI roots of the graph
4719 Value *Consensus = nullptr;
4720 unsigned NumUsesConsensus = 0;
4721 bool IsNumUsesConsensusValid = false;
4722 SmallVector<Instruction*, 16> AddrModeInsts;
4723 ExtAddrMode AddrMode;
4724 TypePromotionTransaction TPT;
4725 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4726 TPT.getRestorationPoint();
4727 while (!worklist.empty()) {
4728 Value *V = worklist.back();
4729 worklist.pop_back();
4731 // Break use-def graph loops.
4732 if (!Visited.insert(V).second) {
4733 Consensus = nullptr;
4737 // For a PHI node, push all of its incoming values.
4738 if (PHINode *P = dyn_cast<PHINode>(V)) {
4739 for (Value *IncValue : P->incoming_values())
4740 worklist.push_back(IncValue);
4744 // For non-PHIs, determine the addressing mode being computed.
4745 SmallVector<Instruction*, 16> NewAddrModeInsts;
4746 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4747 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
4748 InsertedInsts, PromotedInsts, TPT);
4750 // This check is broken into two cases with very similar code to avoid using
4751 // getNumUses() as much as possible. Some values have a lot of uses, so
4752 // calling getNumUses() unconditionally caused a significant compile-time
4756 AddrMode = NewAddrMode;
4757 AddrModeInsts = NewAddrModeInsts;
4759 } else if (NewAddrMode == AddrMode) {
4760 if (!IsNumUsesConsensusValid) {
4761 NumUsesConsensus = Consensus->getNumUses();
4762 IsNumUsesConsensusValid = true;
4765 // Ensure that the obtained addressing mode is equivalent to that obtained
4766 // for all other roots of the PHI traversal. Also, when choosing one
4767 // such root as representative, select the one with the most uses in order
4768 // to keep the cost modeling heuristics in AddressingModeMatcher
4770 unsigned NumUses = V->getNumUses();
4771 if (NumUses > NumUsesConsensus) {
4773 NumUsesConsensus = NumUses;
4774 AddrModeInsts = NewAddrModeInsts;
4779 Consensus = nullptr;
4783 // If the addressing mode couldn't be determined, or if multiple different
4784 // ones were determined, bail out now.
4786 TPT.rollback(LastKnownGood);
4791 // Check to see if any of the instructions supersumed by this addr mode are
4792 // non-local to I's BB.
4793 bool AnyNonLocal = false;
4794 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
4795 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
4801 // If all the instructions matched are already in this BB, don't do anything.
4803 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
4807 // Insert this computation right after this user. Since our caller is
4808 // scanning from the top of the BB to the bottom, reuse of the expr are
4809 // guaranteed to happen later.
4810 IRBuilder<> Builder(MemoryInst);
4812 // Now that we determined the addressing expression we want to use and know
4813 // that we have to sink it into this block. Check to see if we have already
4814 // done this for some other load/store instr in this block. If so, reuse the
4816 Value *&SunkAddr = SunkAddrs[Addr];
4818 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
4819 << *MemoryInst << "\n");
4820 if (SunkAddr->getType() != Addr->getType())
4821 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
4822 } else if (AddrSinkUsingGEPs ||
4823 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
4824 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
4826 // By default, we use the GEP-based method when AA is used later. This
4827 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4828 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4829 << *MemoryInst << "\n");
4830 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4831 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4833 // First, find the pointer.
4834 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4835 ResultPtr = AddrMode.BaseReg;
4836 AddrMode.BaseReg = nullptr;
4839 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4840 // We can't add more than one pointer together, nor can we scale a
4841 // pointer (both of which seem meaningless).
4842 if (ResultPtr || AddrMode.Scale != 1)
4845 ResultPtr = AddrMode.ScaledReg;
4849 if (AddrMode.BaseGV) {
4853 ResultPtr = AddrMode.BaseGV;
4856 // If the real base value actually came from an inttoptr, then the matcher
4857 // will look through it and provide only the integer value. In that case,
4859 if (!ResultPtr && AddrMode.BaseReg) {
4861 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
4862 AddrMode.BaseReg = nullptr;
4863 } else if (!ResultPtr && AddrMode.Scale == 1) {
4865 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
4870 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4871 SunkAddr = Constant::getNullValue(Addr->getType());
4872 } else if (!ResultPtr) {
4876 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4877 Type *I8Ty = Builder.getInt8Ty();
4879 // Start with the base register. Do this first so that subsequent address
4880 // matching finds it last, which will prevent it from trying to match it
4881 // as the scaled value in case it happens to be a mul. That would be
4882 // problematic if we've sunk a different mul for the scale, because then
4883 // we'd end up sinking both muls.
4884 if (AddrMode.BaseReg) {
4885 Value *V = AddrMode.BaseReg;
4886 if (V->getType() != IntPtrTy)
4887 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4892 // Add the scale value.
4893 if (AddrMode.Scale) {
4894 Value *V = AddrMode.ScaledReg;
4895 if (V->getType() == IntPtrTy) {
4897 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4898 cast<IntegerType>(V->getType())->getBitWidth()) {
4899 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4901 // It is only safe to sign extend the BaseReg if we know that the math
4902 // required to create it did not overflow before we extend it. Since
4903 // the original IR value was tossed in favor of a constant back when
4904 // the AddrMode was created we need to bail out gracefully if widths
4905 // do not match instead of extending it.
4906 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
4907 if (I && (ResultIndex != AddrMode.BaseReg))
4908 I->eraseFromParent();
4912 if (AddrMode.Scale != 1)
4913 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4916 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4921 // Add in the Base Offset if present.
4922 if (AddrMode.BaseOffs) {
4923 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4925 // We need to add this separately from the scale above to help with
4926 // SDAG consecutive load/store merging.
4927 if (ResultPtr->getType() != I8PtrTy)
4928 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
4929 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4936 SunkAddr = ResultPtr;
4938 if (ResultPtr->getType() != I8PtrTy)
4939 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
4940 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4943 if (SunkAddr->getType() != Addr->getType())
4944 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
4947 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4948 << *MemoryInst << "\n");
4949 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4950 Value *Result = nullptr;
4952 // Start with the base register. Do this first so that subsequent address
4953 // matching finds it last, which will prevent it from trying to match it
4954 // as the scaled value in case it happens to be a mul. That would be
4955 // problematic if we've sunk a different mul for the scale, because then
4956 // we'd end up sinking both muls.
4957 if (AddrMode.BaseReg) {
4958 Value *V = AddrMode.BaseReg;
4959 if (V->getType()->isPointerTy())
4960 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4961 if (V->getType() != IntPtrTy)
4962 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4966 // Add the scale value.
4967 if (AddrMode.Scale) {
4968 Value *V = AddrMode.ScaledReg;
4969 if (V->getType() == IntPtrTy) {
4971 } else if (V->getType()->isPointerTy()) {
4972 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4973 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4974 cast<IntegerType>(V->getType())->getBitWidth()) {
4975 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4977 // It is only safe to sign extend the BaseReg if we know that the math
4978 // required to create it did not overflow before we extend it. Since
4979 // the original IR value was tossed in favor of a constant back when
4980 // the AddrMode was created we need to bail out gracefully if widths
4981 // do not match instead of extending it.
4982 Instruction *I = dyn_cast_or_null<Instruction>(Result);
4983 if (I && (Result != AddrMode.BaseReg))
4984 I->eraseFromParent();
4987 if (AddrMode.Scale != 1)
4988 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4991 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4996 // Add in the BaseGV if present.
4997 if (AddrMode.BaseGV) {
4998 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
5000 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5005 // Add in the Base Offset if present.
5006 if (AddrMode.BaseOffs) {
5007 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
5009 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5015 SunkAddr = Constant::getNullValue(Addr->getType());
5017 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
5020 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
5022 // If we have no uses, recursively delete the value and all dead instructions
5024 if (Repl->use_empty()) {
5025 // This can cause recursive deletion, which can invalidate our iterator.
5026 // Use a WeakVH to hold onto it in case this happens.
5027 WeakVH IterHandle(&*CurInstIterator);
5028 BasicBlock *BB = CurInstIterator->getParent();
5030 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
5032 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
5033 // If the iterator instruction was recursively deleted, start over at the
5034 // start of the block.
5035 CurInstIterator = BB->begin();
5043 /// If there are any memory operands, use OptimizeMemoryInst to sink their
5044 /// address computing into the block when possible / profitable.
5045 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
5046 bool MadeChange = false;
5048 const TargetRegisterInfo *TRI =
5049 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
5050 TargetLowering::AsmOperandInfoVector TargetConstraints =
5051 TLI->ParseConstraints(*DL, TRI, CS);
5053 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5054 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
5056 // Compute the constraint code and ConstraintType to use.
5057 TLI->ComputeConstraintToUse(OpInfo, SDValue());
5059 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5060 OpInfo.isIndirect) {
5061 Value *OpVal = CS->getArgOperand(ArgNo++);
5062 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
5063 } else if (OpInfo.Type == InlineAsm::isInput)
5070 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
5071 /// sign extensions.
5072 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
5073 assert(!Inst->use_empty() && "Input must have at least one use");
5074 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
5075 bool IsSExt = isa<SExtInst>(FirstUser);
5076 Type *ExtTy = FirstUser->getType();
5077 for (const User *U : Inst->users()) {
5078 const Instruction *UI = cast<Instruction>(U);
5079 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
5081 Type *CurTy = UI->getType();
5082 // Same input and output types: Same instruction after CSE.
5086 // If IsSExt is true, we are in this situation:
5088 // b = sext ty1 a to ty2
5089 // c = sext ty1 a to ty3
5090 // Assuming ty2 is shorter than ty3, this could be turned into:
5092 // b = sext ty1 a to ty2
5093 // c = sext ty2 b to ty3
5094 // However, the last sext is not free.
5098 // This is a ZExt, maybe this is free to extend from one type to another.
5099 // In that case, we would not account for a different use.
5102 if (ExtTy->getScalarType()->getIntegerBitWidth() >
5103 CurTy->getScalarType()->getIntegerBitWidth()) {
5111 if (!TLI.isZExtFree(NarrowTy, LargeTy))
5114 // All uses are the same or can be derived from one another for free.
5118 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
5119 /// load instruction.
5120 /// If an ext(load) can be formed, it is returned via \p LI for the load
5121 /// and \p Inst for the extension.
5122 /// Otherwise LI == nullptr and Inst == nullptr.
5123 /// When some promotion happened, \p TPT contains the proper state to
5126 /// \return true when promoting was necessary to expose the ext(load)
5127 /// opportunity, false otherwise.
5131 /// %ld = load i32* %addr
5132 /// %add = add nuw i32 %ld, 4
5133 /// %zext = zext i32 %add to i64
5137 /// %ld = load i32* %addr
5138 /// %zext = zext i32 %ld to i64
5139 /// %add = add nuw i64 %zext, 4
5141 /// Thanks to the promotion, we can match zext(load i32*) to i64.
5142 bool CodeGenPrepare::extLdPromotion(TypePromotionTransaction &TPT,
5143 LoadInst *&LI, Instruction *&Inst,
5144 const SmallVectorImpl<Instruction *> &Exts,
5145 unsigned CreatedInstsCost = 0) {
5146 // Iterate over all the extensions to see if one form an ext(load).
5147 for (auto I : Exts) {
5148 // Check if we directly have ext(load).
5149 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
5151 // No promotion happened here.
5154 // Check whether or not we want to do any promotion.
5155 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
5157 // Get the action to perform the promotion.
5158 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
5159 I, InsertedInsts, *TLI, PromotedInsts);
5160 // Check if we can promote.
5163 // Save the current state.
5164 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5165 TPT.getRestorationPoint();
5166 SmallVector<Instruction *, 4> NewExts;
5167 unsigned NewCreatedInstsCost = 0;
5168 unsigned ExtCost = !TLI->isExtFree(I);
5170 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
5171 &NewExts, nullptr, *TLI);
5172 assert(PromotedVal &&
5173 "TypePromotionHelper should have filtered out those cases");
5175 // We would be able to merge only one extension in a load.
5176 // Therefore, if we have more than 1 new extension we heuristically
5177 // cut this search path, because it means we degrade the code quality.
5178 // With exactly 2, the transformation is neutral, because we will merge
5179 // one extension but leave one. However, we optimistically keep going,
5180 // because the new extension may be removed too.
5181 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
5182 TotalCreatedInstsCost -= ExtCost;
5183 if (!StressExtLdPromotion &&
5184 (TotalCreatedInstsCost > 1 ||
5185 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
5186 // The promotion is not profitable, rollback to the previous state.
5187 TPT.rollback(LastKnownGood);
5190 // The promotion is profitable.
5191 // Check if it exposes an ext(load).
5192 (void)extLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
5193 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
5194 // If we have created a new extension, i.e., now we have two
5195 // extensions. We must make sure one of them is merged with
5196 // the load, otherwise we may degrade the code quality.
5197 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
5198 // Promotion happened.
5200 // If this does not help to expose an ext(load) then, rollback.
5201 TPT.rollback(LastKnownGood);
5203 // None of the extension can form an ext(load).
5209 /// Move a zext or sext fed by a load into the same basic block as the load,
5210 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
5211 /// extend into the load.
5212 /// \p I[in/out] the extension may be modified during the process if some
5213 /// promotions apply.
5215 bool CodeGenPrepare::moveExtToFormExtLoad(Instruction *&I) {
5216 // Try to promote a chain of computation if it allows to form
5217 // an extended load.
5218 TypePromotionTransaction TPT;
5219 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5220 TPT.getRestorationPoint();
5221 SmallVector<Instruction *, 1> Exts;
5223 // Look for a load being extended.
5224 LoadInst *LI = nullptr;
5225 Instruction *OldExt = I;
5226 bool HasPromoted = extLdPromotion(TPT, LI, I, Exts);
5228 assert(!HasPromoted && !LI && "If we did not match any load instruction "
5229 "the code must remain the same");
5234 // If they're already in the same block, there's nothing to do.
5235 // Make the cheap checks first if we did not promote.
5236 // If we promoted, we need to check if it is indeed profitable.
5237 if (!HasPromoted && LI->getParent() == I->getParent())
5240 EVT VT = TLI->getValueType(*DL, I->getType());
5241 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
5243 // If the load has other users and the truncate is not free, this probably
5244 // isn't worthwhile.
5245 if (!LI->hasOneUse() && TLI &&
5246 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
5247 !TLI->isTruncateFree(I->getType(), LI->getType())) {
5249 TPT.rollback(LastKnownGood);
5253 // Check whether the target supports casts folded into loads.
5255 if (isa<ZExtInst>(I))
5256 LType = ISD::ZEXTLOAD;
5258 assert(isa<SExtInst>(I) && "Unexpected ext type!");
5259 LType = ISD::SEXTLOAD;
5261 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
5263 TPT.rollback(LastKnownGood);
5267 // Move the extend into the same block as the load, so that SelectionDAG
5270 I->removeFromParent();
5276 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
5277 BasicBlock *DefBB = I->getParent();
5279 // If the result of a {s|z}ext and its source are both live out, rewrite all
5280 // other uses of the source with result of extension.
5281 Value *Src = I->getOperand(0);
5282 if (Src->hasOneUse())
5285 // Only do this xform if truncating is free.
5286 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
5289 // Only safe to perform the optimization if the source is also defined in
5291 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
5294 bool DefIsLiveOut = false;
5295 for (User *U : I->users()) {
5296 Instruction *UI = cast<Instruction>(U);
5298 // Figure out which BB this ext is used in.
5299 BasicBlock *UserBB = UI->getParent();
5300 if (UserBB == DefBB) continue;
5301 DefIsLiveOut = true;
5307 // Make sure none of the uses are PHI nodes.
5308 for (User *U : Src->users()) {
5309 Instruction *UI = cast<Instruction>(U);
5310 BasicBlock *UserBB = UI->getParent();
5311 if (UserBB == DefBB) continue;
5312 // Be conservative. We don't want this xform to end up introducing
5313 // reloads just before load / store instructions.
5314 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
5318 // InsertedTruncs - Only insert one trunc in each block once.
5319 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
5321 bool MadeChange = false;
5322 for (Use &U : Src->uses()) {
5323 Instruction *User = cast<Instruction>(U.getUser());
5325 // Figure out which BB this ext is used in.
5326 BasicBlock *UserBB = User->getParent();
5327 if (UserBB == DefBB) continue;
5329 // Both src and def are live in this block. Rewrite the use.
5330 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
5332 if (!InsertedTrunc) {
5333 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5334 assert(InsertPt != UserBB->end());
5335 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5336 InsertedInsts.insert(InsertedTrunc);
5339 // Replace a use of the {s|z}ext source with a use of the result.
5348 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
5349 // just after the load if the target can fold this into one extload instruction,
5350 // with the hope of eliminating some of the other later "and" instructions using
5351 // the loaded value. "and"s that are made trivially redundant by the insertion
5352 // of the new "and" are removed by this function, while others (e.g. those whose
5353 // path from the load goes through a phi) are left for isel to potentially
5386 // becomes (after a call to optimizeLoadExt for each load):
5390 // x1' = and x1, 0xff
5394 // x2' = and x2, 0xff
5401 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5403 if (!Load->isSimple() ||
5404 !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
5407 // Skip loads we've already transformed or have no reason to transform.
5408 if (Load->hasOneUse()) {
5409 User *LoadUser = *Load->user_begin();
5410 if (cast<Instruction>(LoadUser)->getParent() == Load->getParent() &&
5411 !dyn_cast<PHINode>(LoadUser))
5415 // Look at all uses of Load, looking through phis, to determine how many bits
5416 // of the loaded value are needed.
5417 SmallVector<Instruction *, 8> WorkList;
5418 SmallPtrSet<Instruction *, 16> Visited;
5419 SmallVector<Instruction *, 8> AndsToMaybeRemove;
5420 for (auto *U : Load->users())
5421 WorkList.push_back(cast<Instruction>(U));
5423 EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5424 unsigned BitWidth = LoadResultVT.getSizeInBits();
5425 APInt DemandBits(BitWidth, 0);
5426 APInt WidestAndBits(BitWidth, 0);
5428 while (!WorkList.empty()) {
5429 Instruction *I = WorkList.back();
5430 WorkList.pop_back();
5432 // Break use-def graph loops.
5433 if (!Visited.insert(I).second)
5436 // For a PHI node, push all of its users.
5437 if (auto *Phi = dyn_cast<PHINode>(I)) {
5438 for (auto *U : Phi->users())
5439 WorkList.push_back(cast<Instruction>(U));
5443 switch (I->getOpcode()) {
5444 case llvm::Instruction::And: {
5445 auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5448 APInt AndBits = AndC->getValue();
5449 DemandBits |= AndBits;
5450 // Keep track of the widest and mask we see.
5451 if (AndBits.ugt(WidestAndBits))
5452 WidestAndBits = AndBits;
5453 if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5454 AndsToMaybeRemove.push_back(I);
5458 case llvm::Instruction::Shl: {
5459 auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5462 uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5463 auto ShlDemandBits = APInt::getAllOnesValue(BitWidth).lshr(ShiftAmt);
5464 DemandBits |= ShlDemandBits;
5468 case llvm::Instruction::Trunc: {
5469 EVT TruncVT = TLI->getValueType(*DL, I->getType());
5470 unsigned TruncBitWidth = TruncVT.getSizeInBits();
5471 auto TruncBits = APInt::getAllOnesValue(TruncBitWidth).zext(BitWidth);
5472 DemandBits |= TruncBits;
5481 uint32_t ActiveBits = DemandBits.getActiveBits();
5482 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5483 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5484 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5485 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5486 // followed by an AND.
5487 // TODO: Look into removing this restriction by fixing backends to either
5488 // return false for isLoadExtLegal for i1 or have them select this pattern to
5489 // a single instruction.
5491 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5492 // mask, since these are the only ands that will be removed by isel.
5493 if (ActiveBits <= 1 || !APIntOps::isMask(ActiveBits, DemandBits) ||
5494 WidestAndBits != DemandBits)
5497 LLVMContext &Ctx = Load->getType()->getContext();
5498 Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5499 EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5501 // Reject cases that won't be matched as extloads.
5502 if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5503 !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5506 IRBuilder<> Builder(Load->getNextNode());
5507 auto *NewAnd = dyn_cast<Instruction>(
5508 Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5510 // Replace all uses of load with new and (except for the use of load in the
5512 Load->replaceAllUsesWith(NewAnd);
5513 NewAnd->setOperand(0, Load);
5515 // Remove any and instructions that are now redundant.
5516 for (auto *And : AndsToMaybeRemove)
5517 // Check that the and mask is the same as the one we decided to put on the
5519 if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5520 And->replaceAllUsesWith(NewAnd);
5521 if (&*CurInstIterator == And)
5522 CurInstIterator = std::next(And->getIterator());
5523 And->eraseFromParent();
5531 /// Check if V (an operand of a select instruction) is an expensive instruction
5532 /// that is only used once.
5533 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5534 auto *I = dyn_cast<Instruction>(V);
5535 // If it's safe to speculatively execute, then it should not have side
5536 // effects; therefore, it's safe to sink and possibly *not* execute.
5537 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5538 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
5541 /// Returns true if a SelectInst should be turned into an explicit branch.
5542 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
5544 // FIXME: This should use the same heuristics as IfConversion to determine
5545 // whether a select is better represented as a branch. This requires that
5546 // branch probability metadata is preserved for the select, which is not the
5549 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5551 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5552 // comparison condition. If the compare has more than one use, there's
5553 // probably another cmov or setcc around, so it's not worth emitting a branch.
5554 if (!Cmp || !Cmp->hasOneUse())
5557 Value *CmpOp0 = Cmp->getOperand(0);
5558 Value *CmpOp1 = Cmp->getOperand(1);
5560 // Emit "cmov on compare with a memory operand" as a branch to avoid stalls
5561 // on a load from memory. But if the load is used more than once, do not
5562 // change the select to a branch because the load is probably needed
5563 // regardless of whether the branch is taken or not.
5564 if ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
5565 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()))
5568 // If either operand of the select is expensive and only needed on one side
5569 // of the select, we should form a branch.
5570 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5571 sinkSelectOperand(TTI, SI->getFalseValue()))
5578 /// If we have a SelectInst that will likely profit from branch prediction,
5579 /// turn it into a branch.
5580 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5581 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
5583 // Can we convert the 'select' to CF ?
5584 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
5587 TargetLowering::SelectSupportKind SelectKind;
5589 SelectKind = TargetLowering::VectorMaskSelect;
5590 else if (SI->getType()->isVectorTy())
5591 SelectKind = TargetLowering::ScalarCondVectorVal;
5593 SelectKind = TargetLowering::ScalarValSelect;
5595 // Do we have efficient codegen support for this kind of 'selects' ?
5596 if (TLI->isSelectSupported(SelectKind)) {
5597 // We have efficient codegen support for the select instruction.
5598 // Check if it is profitable to keep this 'select'.
5599 if (!TLI->isPredictableSelectExpensive() ||
5600 !isFormingBranchFromSelectProfitable(TTI, SI))
5606 // Transform a sequence like this:
5608 // %cmp = cmp uge i32 %a, %b
5609 // %sel = select i1 %cmp, i32 %c, i32 %d
5613 // %cmp = cmp uge i32 %a, %b
5614 // br i1 %cmp, label %select.true, label %select.false
5616 // br label %select.end
5618 // br label %select.end
5620 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
5622 // In addition, we may sink instructions that produce %c or %d from
5623 // the entry block into the destination(s) of the new branch.
5624 // If the true or false blocks do not contain a sunken instruction, that
5625 // block and its branch may be optimized away. In that case, one side of the
5626 // first branch will point directly to select.end, and the corresponding PHI
5627 // predecessor block will be the start block.
5629 // First, we split the block containing the select into 2 blocks.
5630 BasicBlock *StartBlock = SI->getParent();
5631 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
5632 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
5634 // Delete the unconditional branch that was just created by the split.
5635 StartBlock->getTerminator()->eraseFromParent();
5637 // These are the new basic blocks for the conditional branch.
5638 // At least one will become an actual new basic block.
5639 BasicBlock *TrueBlock = nullptr;
5640 BasicBlock *FalseBlock = nullptr;
5642 // Sink expensive instructions into the conditional blocks to avoid executing
5643 // them speculatively.
5644 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
5645 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
5646 EndBlock->getParent(), EndBlock);
5647 auto *TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
5648 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
5649 TrueInst->moveBefore(TrueBranch);
5651 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
5652 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
5653 EndBlock->getParent(), EndBlock);
5654 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5655 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
5656 FalseInst->moveBefore(FalseBranch);
5659 // If there was nothing to sink, then arbitrarily choose the 'false' side
5660 // for a new input value to the PHI.
5661 if (TrueBlock == FalseBlock) {
5662 assert(TrueBlock == nullptr &&
5663 "Unexpected basic block transform while optimizing select");
5665 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
5666 EndBlock->getParent(), EndBlock);
5667 BranchInst::Create(EndBlock, FalseBlock);
5670 // Insert the real conditional branch based on the original condition.
5671 // If we did not create a new block for one of the 'true' or 'false' paths
5672 // of the condition, it means that side of the branch goes to the end block
5673 // directly and the path originates from the start block from the point of
5674 // view of the new PHI.
5675 if (TrueBlock == nullptr) {
5676 BranchInst::Create(EndBlock, FalseBlock, SI->getCondition(), SI);
5677 TrueBlock = StartBlock;
5678 } else if (FalseBlock == nullptr) {
5679 BranchInst::Create(TrueBlock, EndBlock, SI->getCondition(), SI);
5680 FalseBlock = StartBlock;
5682 BranchInst::Create(TrueBlock, FalseBlock, SI->getCondition(), SI);
5685 // The select itself is replaced with a PHI Node.
5686 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
5688 PN->addIncoming(SI->getTrueValue(), TrueBlock);
5689 PN->addIncoming(SI->getFalseValue(), FalseBlock);
5691 SI->replaceAllUsesWith(PN);
5692 SI->eraseFromParent();
5694 // Instruct OptimizeBlock to skip to the next block.
5695 CurInstIterator = StartBlock->end();
5696 ++NumSelectsExpanded;
5700 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
5701 SmallVector<int, 16> Mask(SVI->getShuffleMask());
5703 for (unsigned i = 0; i < Mask.size(); ++i) {
5704 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
5706 SplatElem = Mask[i];
5712 /// Some targets have expensive vector shifts if the lanes aren't all the same
5713 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
5714 /// it's often worth sinking a shufflevector splat down to its use so that
5715 /// codegen can spot all lanes are identical.
5716 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
5717 BasicBlock *DefBB = SVI->getParent();
5719 // Only do this xform if variable vector shifts are particularly expensive.
5720 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
5723 // We only expect better codegen by sinking a shuffle if we can recognise a
5725 if (!isBroadcastShuffle(SVI))
5728 // InsertedShuffles - Only insert a shuffle in each block once.
5729 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
5731 bool MadeChange = false;
5732 for (User *U : SVI->users()) {
5733 Instruction *UI = cast<Instruction>(U);
5735 // Figure out which BB this ext is used in.
5736 BasicBlock *UserBB = UI->getParent();
5737 if (UserBB == DefBB) continue;
5739 // For now only apply this when the splat is used by a shift instruction.
5740 if (!UI->isShift()) continue;
5742 // Everything checks out, sink the shuffle if the user's block doesn't
5743 // already have a copy.
5744 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
5746 if (!InsertedShuffle) {
5747 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5748 assert(InsertPt != UserBB->end());
5750 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5751 SVI->getOperand(2), "", &*InsertPt);
5754 UI->replaceUsesOfWith(SVI, InsertedShuffle);
5758 // If we removed all uses, nuke the shuffle.
5759 if (SVI->use_empty()) {
5760 SVI->eraseFromParent();
5767 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
5771 Value *Cond = SI->getCondition();
5772 Type *OldType = Cond->getType();
5773 LLVMContext &Context = Cond->getContext();
5774 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
5775 unsigned RegWidth = RegType.getSizeInBits();
5777 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
5780 // If the register width is greater than the type width, expand the condition
5781 // of the switch instruction and each case constant to the width of the
5782 // register. By widening the type of the switch condition, subsequent
5783 // comparisons (for case comparisons) will not need to be extended to the
5784 // preferred register width, so we will potentially eliminate N-1 extends,
5785 // where N is the number of cases in the switch.
5786 auto *NewType = Type::getIntNTy(Context, RegWidth);
5788 // Zero-extend the switch condition and case constants unless the switch
5789 // condition is a function argument that is already being sign-extended.
5790 // In that case, we can avoid an unnecessary mask/extension by sign-extending
5791 // everything instead.
5792 Instruction::CastOps ExtType = Instruction::ZExt;
5793 if (auto *Arg = dyn_cast<Argument>(Cond))
5794 if (Arg->hasSExtAttr())
5795 ExtType = Instruction::SExt;
5797 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
5798 ExtInst->insertBefore(SI);
5799 SI->setCondition(ExtInst);
5800 for (SwitchInst::CaseIt Case : SI->cases()) {
5801 APInt NarrowConst = Case.getCaseValue()->getValue();
5802 APInt WideConst = (ExtType == Instruction::ZExt) ?
5803 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
5804 Case.setValue(ConstantInt::get(Context, WideConst));
5811 /// \brief Helper class to promote a scalar operation to a vector one.
5812 /// This class is used to move downward extractelement transition.
5814 /// a = vector_op <2 x i32>
5815 /// b = extractelement <2 x i32> a, i32 0
5820 /// a = vector_op <2 x i32>
5821 /// c = vector_op a (equivalent to scalar_op on the related lane)
5822 /// * d = extractelement <2 x i32> c, i32 0
5824 /// Assuming both extractelement and store can be combine, we get rid of the
5826 class VectorPromoteHelper {
5827 /// DataLayout associated with the current module.
5828 const DataLayout &DL;
5830 /// Used to perform some checks on the legality of vector operations.
5831 const TargetLowering &TLI;
5833 /// Used to estimated the cost of the promoted chain.
5834 const TargetTransformInfo &TTI;
5836 /// The transition being moved downwards.
5837 Instruction *Transition;
5838 /// The sequence of instructions to be promoted.
5839 SmallVector<Instruction *, 4> InstsToBePromoted;
5840 /// Cost of combining a store and an extract.
5841 unsigned StoreExtractCombineCost;
5842 /// Instruction that will be combined with the transition.
5843 Instruction *CombineInst;
5845 /// \brief 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 if (InstsToBePromoted.empty())
5851 return InstsToBePromoted.back();
5854 /// \brief 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");
5863 /// \brief Return the index of the index in the transition.
5864 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
5866 unsigned getTransitionIdx() const {
5867 assert(isa<ExtractElementInst>(Transition) &&
5868 "Other kind of transitions are not supported yet");
5872 /// \brief 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 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
5880 /// \brief 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, ...
5885 /// b = ToBePromoted <ty1> a, ...
5886 /// Def = Transition <ty1> ToBePromoted to <ty2>
5887 void promoteImpl(Instruction *ToBePromoted);
5889 /// \brief Check whether or not it is profitable to promote all the
5890 /// instructions enqueued to be promoted.
5891 bool isProfitableToPromote() {
5892 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
5893 unsigned Index = isa<ConstantInt>(ValIdx)
5894 ? cast<ConstantInt>(ValIdx)->getZExtValue()
5896 Type *PromotedType = getTransitionType();
5898 StoreInst *ST = cast<StoreInst>(CombineInst);
5899 unsigned AS = ST->getPointerAddressSpace();
5900 unsigned Align = ST->getAlignment();
5901 // Check if this store is supported.
5902 if (!TLI.allowsMisalignedMemoryAccesses(
5903 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
5905 // If this is not supported, there is no way we can combine
5906 // the extract with the store.
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 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
5915 uint64_t VectorCost = StoreExtractCombineCost;
5916 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
5921 Value *Arg0 = Inst->getOperand(0);
5922 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 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
5929 : TargetTransformInfo::OK_AnyValue;
5930 ScalarCost += TTI.getArithmeticInstrCost(
5931 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
5932 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
5935 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
5936 << ScalarCost << "\nVector: " << VectorCost << '\n');
5937 return ScalarCost > VectorCost;
5940 /// \brief Generate a constant vector with \p Val with the same
5941 /// number of elements as the transition.
5942 /// \p UseSplat defines whether or not \p Val should be replicated
5943 /// across the whole vector.
5944 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
5945 /// otherwise we generate a vector with as many undef as possible:
5946 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
5947 /// used at the index of the extract.
5948 Value *getConstantVector(Constant *Val, bool UseSplat) const {
5949 unsigned ExtractIdx = UINT_MAX;
5951 // If we cannot determine where the constant must be, we have to
5952 // use a splat constant.
5953 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
5954 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
5955 ExtractIdx = CstVal->getSExtValue();
5960 unsigned End = getTransitionType()->getVectorNumElements();
5962 return ConstantVector::getSplat(End, Val);
5964 SmallVector<Constant *, 4> ConstVec;
5965 UndefValue *UndefVal = UndefValue::get(Val->getType());
5966 for (unsigned Idx = 0; Idx != End; ++Idx) {
5967 if (Idx == ExtractIdx)
5968 ConstVec.push_back(Val);
5970 ConstVec.push_back(UndefVal);
5972 return ConstantVector::get(ConstVec);
5975 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
5976 /// in \p Use can trigger undefined behavior.
5977 static bool canCauseUndefinedBehavior(const Instruction *Use,
5978 unsigned OperandIdx) {
5979 // This is not safe to introduce undef when the operand is on
5980 // the right hand side of a division-like instruction.
5981 if (OperandIdx != 1)
5983 switch (Use->getOpcode()) {
5986 case Instruction::SDiv:
5987 case Instruction::UDiv:
5988 case Instruction::SRem:
5989 case Instruction::URem:
5991 case Instruction::FDiv:
5992 case Instruction::FRem:
5993 return !Use->hasNoNaNs();
5995 llvm_unreachable(nullptr);
5999 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
6000 const TargetTransformInfo &TTI, Instruction *Transition,
6001 unsigned CombineCost)
6002 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
6003 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
6004 assert(Transition && "Do not know how to promote null");
6007 /// \brief Check if we can promote \p ToBePromoted to \p Type.
6008 bool canPromote(const Instruction *ToBePromoted) const {
6009 // We could support CastInst too.
6010 return isa<BinaryOperator>(ToBePromoted);
6013 /// \brief Check if it is profitable to promote \p ToBePromoted
6014 /// by moving downward the transition through.
6015 bool shouldPromote(const Instruction *ToBePromoted) const {
6016 // Promote only if all the operands can be statically expanded.
6017 // Indeed, we do not want to introduce any new kind of transitions.
6018 for (const Use &U : ToBePromoted->operands()) {
6019 const Value *Val = U.get();
6020 if (Val == getEndOfTransition()) {
6021 // If the use is a division and the transition is on the rhs,
6022 // we cannot promote the operation, otherwise we may create a
6023 // division by zero.
6024 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
6028 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
6029 !isa<ConstantFP>(Val))
6032 // Check that the resulting operation is legal.
6033 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
6036 return StressStoreExtract ||
6037 TLI.isOperationLegalOrCustom(
6038 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
6041 /// \brief Check whether or not \p Use can be combined
6042 /// with the transition.
6043 /// I.e., is it possible to do Use(Transition) => AnotherUse?
6044 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
6046 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
6047 void enqueueForPromotion(Instruction *ToBePromoted) {
6048 InstsToBePromoted.push_back(ToBePromoted);
6051 /// \brief Set the instruction that will be combined with the transition.
6052 void recordCombineInstruction(Instruction *ToBeCombined) {
6053 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
6054 CombineInst = ToBeCombined;
6057 /// \brief Promote all the instructions enqueued for promotion if it is
6059 /// \return True if the promotion happened, false otherwise.
6061 // Check if there is something to promote.
6062 // Right now, if we do not have anything to combine with,
6063 // we assume the promotion is not profitable.
6064 if (InstsToBePromoted.empty() || !CombineInst)
6068 if (!StressStoreExtract && !isProfitableToPromote())
6072 for (auto &ToBePromoted : InstsToBePromoted)
6073 promoteImpl(ToBePromoted);
6074 InstsToBePromoted.clear();
6078 } // End of anonymous namespace.
6080 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
6081 // At this point, we know that all the operands of ToBePromoted but Def
6082 // can be statically promoted.
6083 // For Def, we need to use its parameter in ToBePromoted:
6084 // b = ToBePromoted ty1 a
6085 // Def = Transition ty1 b to ty2
6086 // Move the transition down.
6087 // 1. Replace all uses of the promoted operation by the transition.
6088 // = ... b => = ... Def.
6089 assert(ToBePromoted->getType() == Transition->getType() &&
6090 "The type of the result of the transition does not match "
6092 ToBePromoted->replaceAllUsesWith(Transition);
6093 // 2. Update the type of the uses.
6094 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
6095 Type *TransitionTy = getTransitionType();
6096 ToBePromoted->mutateType(TransitionTy);
6097 // 3. Update all the operands of the promoted operation with promoted
6099 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
6100 for (Use &U : ToBePromoted->operands()) {
6101 Value *Val = U.get();
6102 Value *NewVal = nullptr;
6103 if (Val == Transition)
6104 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
6105 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
6106 isa<ConstantFP>(Val)) {
6107 // Use a splat constant if it is not safe to use undef.
6108 NewVal = getConstantVector(
6109 cast<Constant>(Val),
6110 isa<UndefValue>(Val) ||
6111 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
6113 llvm_unreachable("Did you modified shouldPromote and forgot to update "
6115 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
6117 Transition->removeFromParent();
6118 Transition->insertAfter(ToBePromoted);
6119 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
6122 /// Some targets can do store(extractelement) with one instruction.
6123 /// Try to push the extractelement towards the stores when the target
6124 /// has this feature and this is profitable.
6125 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
6126 unsigned CombineCost = UINT_MAX;
6127 if (DisableStoreExtract || !TLI ||
6128 (!StressStoreExtract &&
6129 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
6130 Inst->getOperand(1), CombineCost)))
6133 // At this point we know that Inst is a vector to scalar transition.
6134 // Try to move it down the def-use chain, until:
6135 // - We can combine the transition with its single use
6136 // => we got rid of the transition.
6137 // - We escape the current basic block
6138 // => we would need to check that we are moving it at a cheaper place and
6139 // we do not do that for now.
6140 BasicBlock *Parent = Inst->getParent();
6141 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
6142 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
6143 // If the transition has more than one use, assume this is not going to be
6145 while (Inst->hasOneUse()) {
6146 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
6147 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
6149 if (ToBePromoted->getParent() != Parent) {
6150 DEBUG(dbgs() << "Instruction to promote is in a different block ("
6151 << ToBePromoted->getParent()->getName()
6152 << ") than the transition (" << Parent->getName() << ").\n");
6156 if (VPH.canCombine(ToBePromoted)) {
6157 DEBUG(dbgs() << "Assume " << *Inst << '\n'
6158 << "will be combined with: " << *ToBePromoted << '\n');
6159 VPH.recordCombineInstruction(ToBePromoted);
6160 bool Changed = VPH.promote();
6161 NumStoreExtractExposed += Changed;
6165 DEBUG(dbgs() << "Try promoting.\n");
6166 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
6169 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
6171 VPH.enqueueForPromotion(ToBePromoted);
6172 Inst = ToBePromoted;
6177 bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
6178 // Bail out if we inserted the instruction to prevent optimizations from
6179 // stepping on each other's toes.
6180 if (InsertedInsts.count(I))
6183 if (PHINode *P = dyn_cast<PHINode>(I)) {
6184 // It is possible for very late stage optimizations (such as SimplifyCFG)
6185 // to introduce PHI nodes too late to be cleaned up. If we detect such a
6186 // trivial PHI, go ahead and zap it here.
6187 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
6188 P->replaceAllUsesWith(V);
6189 P->eraseFromParent();
6196 if (CastInst *CI = dyn_cast<CastInst>(I)) {
6197 // If the source of the cast is a constant, then this should have
6198 // already been constant folded. The only reason NOT to constant fold
6199 // it is if something (e.g. LSR) was careful to place the constant
6200 // evaluation in a block other than then one that uses it (e.g. to hoist
6201 // the address of globals out of a loop). If this is the case, we don't
6202 // want to forward-subst the cast.
6203 if (isa<Constant>(CI->getOperand(0)))
6206 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
6209 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6210 /// Sink a zext or sext into its user blocks if the target type doesn't
6211 /// fit in one register
6213 TLI->getTypeAction(CI->getContext(),
6214 TLI->getValueType(*DL, CI->getType())) ==
6215 TargetLowering::TypeExpandInteger) {
6216 return SinkCast(CI);
6218 bool MadeChange = moveExtToFormExtLoad(I);
6219 return MadeChange | optimizeExtUses(I);
6225 if (CmpInst *CI = dyn_cast<CmpInst>(I))
6226 if (!TLI || !TLI->hasMultipleConditionRegisters())
6227 return OptimizeCmpExpression(CI);
6229 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
6230 stripInvariantGroupMetadata(*LI);
6232 bool Modified = optimizeLoadExt(LI);
6233 unsigned AS = LI->getPointerAddressSpace();
6234 Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
6240 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
6241 stripInvariantGroupMetadata(*SI);
6243 unsigned AS = SI->getPointerAddressSpace();
6244 return optimizeMemoryInst(I, SI->getOperand(1),
6245 SI->getOperand(0)->getType(), AS);
6250 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
6252 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
6253 BinOp->getOpcode() == Instruction::LShr)) {
6254 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
6255 if (TLI && CI && TLI->hasExtractBitsInsn())
6256 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
6261 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
6262 if (GEPI->hasAllZeroIndices()) {
6263 /// The GEP operand must be a pointer, so must its result -> BitCast
6264 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
6265 GEPI->getName(), GEPI);
6266 GEPI->replaceAllUsesWith(NC);
6267 GEPI->eraseFromParent();
6269 optimizeInst(NC, ModifiedDT);
6275 if (CallInst *CI = dyn_cast<CallInst>(I))
6276 return optimizeCallInst(CI, ModifiedDT);
6278 if (SelectInst *SI = dyn_cast<SelectInst>(I))
6279 return optimizeSelectInst(SI);
6281 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
6282 return optimizeShuffleVectorInst(SVI);
6284 if (auto *Switch = dyn_cast<SwitchInst>(I))
6285 return optimizeSwitchInst(Switch);
6287 if (isa<ExtractElementInst>(I))
6288 return optimizeExtractElementInst(I);
6293 /// Given an OR instruction, check to see if this is a bitreverse
6294 /// idiom. If so, insert the new intrinsic and return true.
6295 static bool makeBitReverse(Instruction &I, const DataLayout &DL,
6296 const TargetLowering &TLI) {
6297 if (!I.getType()->isIntegerTy() ||
6298 !TLI.isOperationLegalOrCustom(ISD::BITREVERSE,
6299 TLI.getValueType(DL, I.getType(), true)))
6302 SmallVector<Instruction*, 4> Insts;
6303 if (!recognizeBitReverseOrBSwapIdiom(&I, false, true, Insts))
6305 Instruction *LastInst = Insts.back();
6306 I.replaceAllUsesWith(LastInst);
6307 RecursivelyDeleteTriviallyDeadInstructions(&I);
6311 // In this pass we look for GEP and cast instructions that are used
6312 // across basic blocks and rewrite them to improve basic-block-at-a-time
6314 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
6316 bool MadeChange = false;
6318 CurInstIterator = BB.begin();
6319 while (CurInstIterator != BB.end()) {
6320 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
6325 bool MadeBitReverse = true;
6326 while (TLI && MadeBitReverse) {
6327 MadeBitReverse = false;
6328 for (auto &I : reverse(BB)) {
6329 if (makeBitReverse(I, *DL, *TLI)) {
6330 MadeBitReverse = MadeChange = true;
6335 MadeChange |= dupRetToEnableTailCallOpts(&BB);
6340 // llvm.dbg.value is far away from the value then iSel may not be able
6341 // handle it properly. iSel will drop llvm.dbg.value if it can not
6342 // find a node corresponding to the value.
6343 bool CodeGenPrepare::placeDbgValues(Function &F) {
6344 bool MadeChange = false;
6345 for (BasicBlock &BB : F) {
6346 Instruction *PrevNonDbgInst = nullptr;
6347 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
6348 Instruction *Insn = &*BI++;
6349 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
6350 // Leave dbg.values that refer to an alloca alone. These
6351 // instrinsics describe the address of a variable (= the alloca)
6352 // being taken. They should not be moved next to the alloca
6353 // (and to the beginning of the scope), but rather stay close to
6354 // where said address is used.
6355 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
6356 PrevNonDbgInst = Insn;
6360 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
6361 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
6362 // If VI is a phi in a block with an EHPad terminator, we can't insert
6364 if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
6366 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
6367 DVI->removeFromParent();
6368 if (isa<PHINode>(VI))
6369 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
6371 DVI->insertAfter(VI);
6380 // If there is a sequence that branches based on comparing a single bit
6381 // against zero that can be combined into a single instruction, and the
6382 // target supports folding these into a single instruction, sink the
6383 // mask and compare into the branch uses. Do this before OptimizeBlock ->
6384 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
6386 bool CodeGenPrepare::sinkAndCmp(Function &F) {
6387 if (!EnableAndCmpSinking)
6389 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
6391 bool MadeChange = false;
6392 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
6393 BasicBlock *BB = &*I++;
6395 // Does this BB end with the following?
6396 // %andVal = and %val, #single-bit-set
6397 // %icmpVal = icmp %andResult, 0
6398 // br i1 %cmpVal label %dest1, label %dest2"
6399 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
6400 if (!Brcc || !Brcc->isConditional())
6402 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
6403 if (!Cmp || Cmp->getParent() != BB)
6405 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
6406 if (!Zero || !Zero->isZero())
6408 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
6409 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
6411 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
6412 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
6414 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
6416 // Push the "and; icmp" for any users that are conditional branches.
6417 // Since there can only be one branch use per BB, we don't need to keep
6418 // track of which BBs we insert into.
6419 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
6423 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
6425 if (!BrccUser || !BrccUser->isConditional())
6427 BasicBlock *UserBB = BrccUser->getParent();
6428 if (UserBB == BB) continue;
6429 DEBUG(dbgs() << "found Brcc use\n");
6431 // Sink the "and; icmp" to use.
6433 BinaryOperator *NewAnd =
6434 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
6437 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
6441 DEBUG(BrccUser->getParent()->dump());
6447 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
6448 /// success, or returns false if no or invalid metadata was found.
6449 static bool extractBranchMetadata(BranchInst *BI,
6450 uint64_t &ProbTrue, uint64_t &ProbFalse) {
6451 assert(BI->isConditional() &&
6452 "Looking for probabilities on unconditional branch?");
6453 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
6454 if (!ProfileData || ProfileData->getNumOperands() != 3)
6457 const auto *CITrue =
6458 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
6459 const auto *CIFalse =
6460 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
6461 if (!CITrue || !CIFalse)
6464 ProbTrue = CITrue->getValue().getZExtValue();
6465 ProbFalse = CIFalse->getValue().getZExtValue();
6470 /// \brief Scale down both weights to fit into uint32_t.
6471 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
6472 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
6473 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
6474 NewTrue = NewTrue / Scale;
6475 NewFalse = NewFalse / Scale;
6478 /// \brief Some targets prefer to split a conditional branch like:
6480 /// %0 = icmp ne i32 %a, 0
6481 /// %1 = icmp ne i32 %b, 0
6482 /// %or.cond = or i1 %0, %1
6483 /// br i1 %or.cond, label %TrueBB, label %FalseBB
6485 /// into multiple branch instructions like:
6488 /// %0 = icmp ne i32 %a, 0
6489 /// br i1 %0, label %TrueBB, label %bb2
6491 /// %1 = icmp ne i32 %b, 0
6492 /// br i1 %1, label %TrueBB, label %FalseBB
6494 /// This usually allows instruction selection to do even further optimizations
6495 /// and combine the compare with the branch instruction. Currently this is
6496 /// applied for targets which have "cheap" jump instructions.
6498 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
6500 bool CodeGenPrepare::splitBranchCondition(Function &F) {
6501 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
6504 bool MadeChange = false;
6505 for (auto &BB : F) {
6506 // Does this BB end with the following?
6507 // %cond1 = icmp|fcmp|binary instruction ...
6508 // %cond2 = icmp|fcmp|binary instruction ...
6509 // %cond.or = or|and i1 %cond1, cond2
6510 // br i1 %cond.or label %dest1, label %dest2"
6511 BinaryOperator *LogicOp;
6512 BasicBlock *TBB, *FBB;
6513 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
6516 auto *Br1 = cast<BranchInst>(BB.getTerminator());
6517 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
6521 Value *Cond1, *Cond2;
6522 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
6523 m_OneUse(m_Value(Cond2)))))
6524 Opc = Instruction::And;
6525 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
6526 m_OneUse(m_Value(Cond2)))))
6527 Opc = Instruction::Or;
6531 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
6532 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
6535 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
6538 auto *InsertBefore = std::next(Function::iterator(BB))
6539 .getNodePtrUnchecked();
6540 auto TmpBB = BasicBlock::Create(BB.getContext(),
6541 BB.getName() + ".cond.split",
6542 BB.getParent(), InsertBefore);
6544 // Update original basic block by using the first condition directly by the
6545 // branch instruction and removing the no longer needed and/or instruction.
6546 Br1->setCondition(Cond1);
6547 LogicOp->eraseFromParent();
6549 // Depending on the conditon we have to either replace the true or the false
6550 // successor of the original branch instruction.
6551 if (Opc == Instruction::And)
6552 Br1->setSuccessor(0, TmpBB);
6554 Br1->setSuccessor(1, TmpBB);
6556 // Fill in the new basic block.
6557 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
6558 if (auto *I = dyn_cast<Instruction>(Cond2)) {
6559 I->removeFromParent();
6560 I->insertBefore(Br2);
6563 // Update PHI nodes in both successors. The original BB needs to be
6564 // replaced in one succesor's PHI nodes, because the branch comes now from
6565 // the newly generated BB (NewBB). In the other successor we need to add one
6566 // incoming edge to the PHI nodes, because both branch instructions target
6567 // now the same successor. Depending on the original branch condition
6568 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
6569 // we perfrom the correct update for the PHI nodes.
6570 // This doesn't change the successor order of the just created branch
6571 // instruction (or any other instruction).
6572 if (Opc == Instruction::Or)
6573 std::swap(TBB, FBB);
6575 // Replace the old BB with the new BB.
6576 for (auto &I : *TBB) {
6577 PHINode *PN = dyn_cast<PHINode>(&I);
6581 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
6582 PN->setIncomingBlock(i, TmpBB);
6585 // Add another incoming edge form the new BB.
6586 for (auto &I : *FBB) {
6587 PHINode *PN = dyn_cast<PHINode>(&I);
6590 auto *Val = PN->getIncomingValueForBlock(&BB);
6591 PN->addIncoming(Val, TmpBB);
6594 // Update the branch weights (from SelectionDAGBuilder::
6595 // FindMergedConditions).
6596 if (Opc == Instruction::Or) {
6597 // Codegen X | Y as:
6606 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
6607 // The requirement is that
6608 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
6609 // = TrueProb for orignal BB.
6610 // Assuming the orignal weights are A and B, one choice is to set BB1's
6611 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
6613 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
6614 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
6615 // TmpBB, but the math is more complicated.
6616 uint64_t TrueWeight, FalseWeight;
6617 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
6618 uint64_t NewTrueWeight = TrueWeight;
6619 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
6620 scaleWeights(NewTrueWeight, NewFalseWeight);
6621 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6622 .createBranchWeights(TrueWeight, FalseWeight));
6624 NewTrueWeight = TrueWeight;
6625 NewFalseWeight = 2 * FalseWeight;
6626 scaleWeights(NewTrueWeight, NewFalseWeight);
6627 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6628 .createBranchWeights(TrueWeight, FalseWeight));
6631 // Codegen X & Y as:
6639 // This requires creation of TmpBB after CurBB.
6641 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
6642 // The requirement is that
6643 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
6644 // = FalseProb for orignal BB.
6645 // Assuming the orignal weights are A and B, one choice is to set BB1's
6646 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
6648 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
6649 uint64_t TrueWeight, FalseWeight;
6650 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
6651 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
6652 uint64_t NewFalseWeight = FalseWeight;
6653 scaleWeights(NewTrueWeight, NewFalseWeight);
6654 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6655 .createBranchWeights(TrueWeight, FalseWeight));
6657 NewTrueWeight = 2 * TrueWeight;
6658 NewFalseWeight = FalseWeight;
6659 scaleWeights(NewTrueWeight, NewFalseWeight);
6660 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6661 .createBranchWeights(TrueWeight, FalseWeight));
6665 // Note: No point in getting fancy here, since the DT info is never
6666 // available to CodeGenPrepare.
6671 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
6677 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
6678 if (auto *InvariantMD = I.getMetadata(LLVMContext::MD_invariant_group))
6679 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());