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();
794 while (dyn_cast<PHINode>(InsertPoint)) {
795 InsertPoint = InsertPoint->getNextNode();
798 // First thing is to cast 'UsageInst' to an integer type if necessary.
799 Value* AndTarget = nullptr;
800 if (IntegerType::classof(UsageInst->getType())) {
801 AndTarget = UsageInst;
803 Type* TargetIntegerType = IntegerType::get(
804 UsageInst->getContext(),
805 BB->getModule()->getDataLayout().getPointerSizeInBits());
806 IRBuilder<true, NoFolder> Builder(UsageInst->getNextNode());
807 AndTarget = createCast(Builder, UsageInst, TargetIntegerType);
810 // Check whether InsertPoint is a added fake conditional branch.
811 BranchInst* BI = nullptr;
812 if ((BI = dyn_cast<BranchInst>(InsertPoint)) && BI->isConditional()) {
813 auto* Cond = dyn_cast<Instruction>(BI->getOperand(0));
814 if (Cond && Cond->getOpcode() == Instruction::ICmp) {
815 auto* CmpInst = dyn_cast<ICmpInst>(Cond);
816 auto* Op0 = dyn_cast<Instruction>(Cond->getOperand(0));
817 auto* Op1 = dyn_cast<ConstantInt>(Cond->getOperand(1));
819 // %cmp = ICMP_NE %tmp, 0
822 // %tmp1 = And X, NewTaintedVal
823 // %tmp2 = And %tmp1, 0
824 // %cmp = ICMP_NE %tmp2, 0
826 if (CmpInst && CmpInst->getPredicate() == CmpInst::ICMP_NE && Op0 &&
827 Op0->getOpcode() == Instruction::And && Op1 && Op1->isZero()) {
828 auto* Op01 = dyn_cast<ConstantInt>(Op0->getOperand(1));
829 if (Op01 && Op01->isZero()) {
830 // Now we have a previously added fake cond branch.
831 auto* Op00 = Op0->getOperand(0);
832 IRBuilder<true, NoFolder> Builder(CmpInst);
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 auto* AndZero = dyn_cast<Instruction>(
845 Builder.CreateAnd(AndTarget, Constant::getNullValue(AndTarget->getType())));
846 auto* FakeCondition = dyn_cast<Instruction>(Builder.CreateICmp(
847 CmpInst::ICMP_NE, AndZero, Constant::getNullValue(AndTarget->getType())));
848 AddFakeConditionalBranch(FakeCondition->getNextNode(), FakeCondition);
851 // XXX-comment: Finds the appropriate Value derived from an atomic load.
852 // 'ChainedBB' contains all the blocks chained together with unconditional
853 // branches from LI's parent BB to the block with the first store/cond branch.
854 // If we don't find any, it means 'LI' is not used at all (which should not
855 // happen in practice). We can simply set 'LI' to be acquire just to be safe.
856 template <typename Vector>
857 Instruction* findMostRecentDependenceUsage(LoadInst* LI, Instruction* LaterInst,
860 typedef SmallSet<Instruction*, 8> UsageSet;
861 typedef DenseMap<BasicBlock*, std::unique_ptr<UsageSet>> UsageMap;
862 assert(ChainedBB->size() >= 1 && "ChainedBB must have >=1 size");
863 // Mapping from basic block in 'ChainedBB' to the set of dependence usage of
864 // 'LI' in each block.
866 auto* LoadBB = LI->getParent();
867 usage_map[LoadBB] = make_unique<UsageSet>();
868 usage_map[LoadBB]->insert(LI);
870 for (auto* BB : *ChainedBB) {
871 if (usage_map[BB] == nullptr) {
872 usage_map[BB] = make_unique<UsageSet>();
874 auto& usage_set = usage_map[BB];
875 if (usage_set->size() == 0) {
876 // The value has not been used.
879 // Calculate the usage in the current BB first.
880 std::list<Value*> bb_usage_list;
881 std::copy(usage_set->begin(), usage_set->end(),
882 std::back_inserter(bb_usage_list));
883 for (auto list_iter = bb_usage_list.begin();
884 list_iter != bb_usage_list.end(); list_iter++) {
885 auto* val = *list_iter;
886 for (auto* U : val->users()) {
887 Instruction* Inst = nullptr;
888 if (!(Inst = dyn_cast<Instruction>(U))) {
891 assert(Inst && "Usage value must be an instruction");
893 std::find(ChainedBB->begin(), ChainedBB->end(), Inst->getParent());
894 if (iter == ChainedBB->end()) {
895 // Only care about usage within ChainedBB.
898 auto* UsageBB = *iter;
901 if (!usage_set->count(Inst)) {
902 bb_usage_list.push_back(Inst);
903 usage_set->insert(Inst);
907 if (usage_map[UsageBB] == nullptr) {
908 usage_map[UsageBB] = make_unique<UsageSet>();
910 usage_map[UsageBB]->insert(Inst);
916 // Pick one usage that is in LaterInst's block and that dominates 'LaterInst'.
917 auto* LaterBB = LaterInst->getParent();
918 auto& usage_set = usage_map[LaterBB];
919 Instruction* usage_inst = nullptr;
920 for (auto* inst : *usage_set) {
921 if (DT->dominates(inst, LaterInst)) {
927 assert(usage_inst && "The usage instruction in the same block but after the "
928 "later instruction");
932 // XXX-comment: Returns whether the code has been changed.
933 bool AddFakeConditionalBranchAfterMonotonicLoads(
934 SmallSet<LoadInst*, 1>& MonotonicLoadInsts, DominatorTree* DT) {
935 bool Changed = false;
936 while (!MonotonicLoadInsts.empty()) {
937 auto* LI = *MonotonicLoadInsts.begin();
938 MonotonicLoadInsts.erase(LI);
939 SmallVector<BasicBlock*, 2> ChainedBB;
940 auto* FirstInst = findFirstStoreCondBranchInst(LI, &ChainedBB);
941 if (FirstInst != nullptr) {
942 if (FirstInst->getOpcode() == Instruction::Store) {
943 if (StoreAddressDependOnValue(dyn_cast<StoreInst>(FirstInst), LI)) {
946 } else if (FirstInst->getOpcode() == Instruction::Br) {
947 if (ConditionalBranchDependsOnValue(dyn_cast<BranchInst>(FirstInst),
952 dbgs() << "FirstInst=" << *FirstInst << "\n";
953 assert(false && "findFirstStoreCondBranchInst() should return a "
954 "store/condition branch instruction");
958 // We really need to process the relaxed load now.
959 StoreInst* SI = nullptr;;
960 if (FirstInst && (SI = dyn_cast<StoreInst>(FirstInst))) {
961 // For immediately coming stores, taint the address of the store.
962 if (SI->getParent() == LI->getParent() || DT->dominates(LI, SI)) {
963 TaintRelaxedLoads(LI, SI);
967 findMostRecentDependenceUsage(LI, FirstInst, &ChainedBB, DT);
969 LI->setOrdering(Acquire);
972 TaintRelaxedLoads(Inst, SI);
977 // No upcoming branch
979 TaintRelaxedLoads(LI, nullptr);
982 // For immediately coming branch, directly add a fake branch.
983 if (FirstInst->getParent() == LI->getParent() ||
984 DT->dominates(LI, FirstInst)) {
985 TaintRelaxedLoads(LI, FirstInst);
989 findMostRecentDependenceUsage(LI, FirstInst, &ChainedBB, DT);
991 TaintRelaxedLoads(Inst, FirstInst);
993 LI->setOrdering(Acquire);
1003 /**** Implementations of public methods for dependence tainting ****/
1004 Value* GetUntaintedAddress(Value* CurrentAddress) {
1005 auto* OrAddress = getOrAddress(CurrentAddress);
1006 if (OrAddress == nullptr) {
1007 // Is it tainted by a select instruction?
1008 auto* Inst = dyn_cast<Instruction>(CurrentAddress);
1009 if (nullptr != Inst && Inst->getOpcode() == Instruction::Select) {
1010 // A selection instruction.
1011 if (Inst->getOperand(1) == Inst->getOperand(2)) {
1012 return Inst->getOperand(1);
1016 return CurrentAddress;
1018 Value* ActualAddress = nullptr;
1020 auto* CastToInt = dyn_cast<Instruction>(OrAddress->getOperand(1));
1021 if (CastToInt && CastToInt->getOpcode() == Instruction::PtrToInt) {
1022 return CastToInt->getOperand(0);
1024 // This should be a IntToPtr constant expression.
1025 ConstantExpr* PtrToIntExpr =
1026 dyn_cast<ConstantExpr>(OrAddress->getOperand(1));
1027 if (PtrToIntExpr && PtrToIntExpr->getOpcode() == Instruction::PtrToInt) {
1028 return PtrToIntExpr->getOperand(0);
1032 // Looks like it's not been dependence-tainted. Returns itself.
1033 return CurrentAddress;
1036 MemoryLocation GetUntaintedMemoryLocation(StoreInst* SI) {
1038 SI->getAAMetadata(AATags);
1039 const auto& DL = SI->getModule()->getDataLayout();
1040 const auto* OriginalAddr = GetUntaintedAddress(SI->getPointerOperand());
1041 DEBUG(if (OriginalAddr != SI->getPointerOperand()) {
1042 dbgs() << "[GetUntaintedMemoryLocation]\n"
1043 << "Storing address: " << *SI->getPointerOperand()
1044 << "\nUntainted address: " << *OriginalAddr << "\n";
1046 return MemoryLocation(OriginalAddr,
1047 DL.getTypeStoreSize(SI->getValueOperand()->getType()),
1051 bool TaintDependenceToStore(StoreInst* SI, Value* DepVal) {
1052 if (dependenceSetInclusion(SI, DepVal)) {
1056 bool tainted = taintStoreAddress(SI, DepVal);
1061 bool TaintDependenceToStoreAddress(StoreInst* SI, Value* DepVal) {
1062 if (dependenceSetInclusion(SI->getPointerOperand(), DepVal)) {
1066 bool tainted = taintStoreAddress(SI, DepVal);
1071 bool CompressTaintedStore(BasicBlock* BB) {
1072 // This function looks for windows of adajcent stores in 'BB' that satisfy the
1073 // following condition (and then do optimization):
1074 // *Addr(d1) = v1, d1 is a condition and is the only dependence the store's
1075 // address depends on && Dep(v1) includes Dep(d1);
1076 // *Addr(d2) = v2, d2 is a condition and is the only dependnece the store's
1077 // address depends on && Dep(v2) includes Dep(d2) &&
1078 // Dep(d2) includes Dep(d1);
1080 // *Addr(dN) = vN, dN is a condition and is the only dependence the store's
1081 // address depends on && Dep(dN) includes Dep(d"N-1").
1083 // As a result, Dep(dN) includes [Dep(d1) V ... V Dep(d"N-1")], so we can
1084 // safely transform the above to the following. In between these stores, we
1085 // can omit untainted stores to the same address 'Addr' since they internally
1086 // have dependence on the previous stores on the same address.
1091 for (auto BI = BB->begin(), BE = BB->end(); BI != BE; BI++) {
1092 // Look for the first store in such a window of adajacent stores.
1093 auto* FirstSI = dyn_cast<StoreInst>(&*BI);
1098 // The first store in the window must be tainted.
1099 auto* UntaintedAddress = GetUntaintedAddress(FirstSI->getPointerOperand());
1100 if (UntaintedAddress == FirstSI->getPointerOperand()) {
1104 // The first store's address must directly depend on and only depend on a
1106 auto* FirstSIDepCond = getConditionDependence(FirstSI->getPointerOperand());
1107 if (nullptr == FirstSIDepCond) {
1111 // Dep(first store's storing value) includes Dep(tainted dependence).
1112 if (!dependenceSetInclusion(FirstSI->getValueOperand(), FirstSIDepCond)) {
1116 // Look for subsequent stores to the same address that satisfy the condition
1117 // of "compressing the dependence".
1118 SmallVector<StoreInst*, 8> AdajacentStores;
1119 AdajacentStores.push_back(FirstSI);
1120 auto BII = BasicBlock::iterator(FirstSI);
1121 for (BII++; BII != BE; BII++) {
1122 auto* CurrSI = dyn_cast<StoreInst>(&*BII);
1124 if (BII->mayHaveSideEffects()) {
1125 // Be conservative. Instructions with side effects are similar to
1132 auto* OrigAddress = GetUntaintedAddress(CurrSI->getPointerOperand());
1133 auto* CurrSIDepCond = getConditionDependence(CurrSI->getPointerOperand());
1134 // All other stores must satisfy either:
1135 // A. 'CurrSI' is an untainted store to the same address, or
1136 // B. the combination of the following 5 subconditions:
1138 // 2. Untainted address is the same as the group's address;
1139 // 3. The address is tainted with a sole value which is a condition;
1140 // 4. The storing value depends on the condition in 3.
1141 // 5. The condition in 3 depends on the previous stores dependence
1144 // Condition A. Should ignore this store directly.
1145 if (OrigAddress == CurrSI->getPointerOperand() &&
1146 OrigAddress == UntaintedAddress) {
1149 // Check condition B.
1150 Value* Cond = nullptr;
1151 if (OrigAddress == CurrSI->getPointerOperand() ||
1152 OrigAddress != UntaintedAddress || CurrSIDepCond == nullptr ||
1153 !dependenceSetInclusion(CurrSI->getValueOperand(), CurrSIDepCond)) {
1154 // Check condition 1, 2, 3 & 4.
1158 // Check condition 5.
1159 StoreInst* PrevSI = AdajacentStores[AdajacentStores.size() - 1];
1160 auto* PrevSIDepCond = getConditionDependence(PrevSI->getPointerOperand());
1161 assert(PrevSIDepCond &&
1162 "Store in the group must already depend on a condtion");
1163 if (!dependenceSetInclusion(CurrSIDepCond, PrevSIDepCond)) {
1167 AdajacentStores.push_back(CurrSI);
1170 if (AdajacentStores.size() == 1) {
1171 // The outer loop should keep looking from the next store.
1175 // Now we have such a group of tainted stores to the same address.
1176 DEBUG(dbgs() << "[CompressTaintedStore]\n");
1177 DEBUG(dbgs() << "Original BB\n");
1178 DEBUG(dbgs() << *BB << '\n');
1179 auto* LastSI = AdajacentStores[AdajacentStores.size() - 1];
1180 for (unsigned i = 0; i < AdajacentStores.size() - 1; ++i) {
1181 auto* SI = AdajacentStores[i];
1183 // Use the original address for stores before the last one.
1184 SI->setOperand(1, UntaintedAddress);
1186 DEBUG(dbgs() << "Store address has been reversed: " << *SI << '\n';);
1188 // XXX-comment: Try to make the last store use fewer registers.
1189 // If LastSI's storing value is a select based on the condition with which
1190 // its address is tainted, transform the tainted address to a select
1191 // instruction, as follows:
1192 // r1 = Select Cond ? A : B
1197 // r1 = Select Cond ? A : B
1198 // r2 = Select Cond ? Addr : Addr
1200 // The idea is that both Select instructions depend on the same condition,
1201 // so hopefully the backend can generate two cmov instructions for them (and
1202 // this saves the number of registers needed).
1203 auto* LastSIDep = getConditionDependence(LastSI->getPointerOperand());
1204 auto* LastSIValue = dyn_cast<Instruction>(LastSI->getValueOperand());
1205 if (LastSIValue && LastSIValue->getOpcode() == Instruction::Select &&
1206 LastSIValue->getOperand(0) == LastSIDep) {
1207 // XXX-comment: Maybe it's better for us to just leave it as an and/or
1208 // dependence pattern.
1210 IRBuilder<true, NoFolder> Builder(LastSI);
1212 Builder.CreateSelect(LastSIDep, UntaintedAddress, UntaintedAddress);
1213 LastSI->setOperand(1, Address);
1214 DEBUG(dbgs() << "The last store becomes :" << *LastSI << "\n\n";);
1222 bool PassDependenceToStore(Value* OldAddress, StoreInst* NewStore) {
1223 Value* OldDep = getDependence(OldAddress);
1224 // Return false when there's no dependence to pass from the OldAddress.
1229 // No need to pass the dependence to NewStore's address if it already depends
1230 // on whatever 'OldAddress' depends on.
1231 if (StoreAddressDependOnValue(NewStore, OldDep)) {
1234 return taintStoreAddress(NewStore, OldAddress);
1237 SmallSet<Value*, 8> FindDependence(Value* Val) {
1238 SmallSet<Value*, 8> DepSet;
1239 recursivelyFindDependence(&DepSet, Val, true /*Only insert leaf nodes*/);
1243 bool StoreAddressDependOnValue(StoreInst* SI, Value* DepVal) {
1244 return dependenceSetInclusion(SI->getPointerOperand(), DepVal);
1247 bool StoreDependOnValue(StoreInst* SI, Value* Dep) {
1248 return dependenceSetInclusion(SI, Dep);
1255 bool CodeGenPrepare::runOnFunction(Function &F) {
1256 bool EverMadeChange = false;
1258 if (skipOptnoneFunction(F))
1261 DL = &F.getParent()->getDataLayout();
1263 // Clear per function information.
1264 InsertedInsts.clear();
1265 PromotedInsts.clear();
1269 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
1270 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1271 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1272 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1273 OptSize = F.optForSize();
1275 /// This optimization identifies DIV instructions that can be
1276 /// profitably bypassed and carried out with a shorter, faster divide.
1277 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
1278 const DenseMap<unsigned int, unsigned int> &BypassWidths =
1279 TLI->getBypassSlowDivWidths();
1280 BasicBlock* BB = &*F.begin();
1281 while (BB != nullptr) {
1282 // bypassSlowDivision may create new BBs, but we don't want to reapply the
1283 // optimization to those blocks.
1284 BasicBlock* Next = BB->getNextNode();
1285 EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
1290 // Eliminate blocks that contain only PHI nodes and an
1291 // unconditional branch.
1292 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
1294 // llvm.dbg.value is far away from the value then iSel may not be able
1295 // handle it properly. iSel will drop llvm.dbg.value if it can not
1296 // find a node corresponding to the value.
1297 EverMadeChange |= placeDbgValues(F);
1299 // If there is a mask, compare against zero, and branch that can be combined
1300 // into a single target instruction, push the mask and compare into branch
1301 // users. Do this before OptimizeBlock -> OptimizeInst ->
1302 // OptimizeCmpExpression, which perturbs the pattern being searched for.
1303 if (!DisableBranchOpts) {
1304 EverMadeChange |= sinkAndCmp(F);
1305 EverMadeChange |= splitBranchCondition(F);
1308 bool MadeChange = true;
1309 while (MadeChange) {
1311 for (Function::iterator I = F.begin(); I != F.end(); ) {
1312 BasicBlock *BB = &*I++;
1313 bool ModifiedDTOnIteration = false;
1314 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
1316 // Restart BB iteration if the dominator tree of the Function was changed
1317 if (ModifiedDTOnIteration)
1320 EverMadeChange |= MadeChange;
1325 if (!DisableBranchOpts) {
1327 SmallPtrSet<BasicBlock*, 8> WorkList;
1328 for (BasicBlock &BB : F) {
1329 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
1330 MadeChange |= ConstantFoldTerminator(&BB, true);
1331 if (!MadeChange) continue;
1333 for (SmallVectorImpl<BasicBlock*>::iterator
1334 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
1335 if (pred_begin(*II) == pred_end(*II))
1336 WorkList.insert(*II);
1339 // Delete the dead blocks and any of their dead successors.
1340 MadeChange |= !WorkList.empty();
1341 while (!WorkList.empty()) {
1342 BasicBlock *BB = *WorkList.begin();
1344 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
1346 DeleteDeadBlock(BB);
1348 for (SmallVectorImpl<BasicBlock*>::iterator
1349 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
1350 if (pred_begin(*II) == pred_end(*II))
1351 WorkList.insert(*II);
1354 // Merge pairs of basic blocks with unconditional branches, connected by
1356 if (EverMadeChange || MadeChange)
1357 MadeChange |= eliminateFallThrough(F);
1359 EverMadeChange |= MadeChange;
1362 if (!DisableGCOpts) {
1363 SmallVector<Instruction *, 2> Statepoints;
1364 for (BasicBlock &BB : F)
1365 for (Instruction &I : BB)
1366 if (isStatepoint(I))
1367 Statepoints.push_back(&I);
1368 for (auto &I : Statepoints)
1369 EverMadeChange |= simplifyOffsetableRelocate(*I);
1372 // XXX-comment: Delay dealing with relaxed loads in this function to avoid
1373 // further changes done by other passes (e.g., SimplifyCFG).
1374 // Collect all the relaxed loads.
1375 SmallSet<LoadInst*, 1> MonotonicLoadInsts;
1376 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
1377 if (I->isAtomic()) {
1378 switch (I->getOpcode()) {
1379 case Instruction::Load: {
1380 auto* LI = dyn_cast<LoadInst>(&*I);
1381 if (LI->getOrdering() == Monotonic) {
1382 MonotonicLoadInsts.insert(LI);
1393 AddFakeConditionalBranchAfterMonotonicLoads(MonotonicLoadInsts, DT);
1395 return EverMadeChange;
1398 /// Merge basic blocks which are connected by a single edge, where one of the
1399 /// basic blocks has a single successor pointing to the other basic block,
1400 /// which has a single predecessor.
1401 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
1402 bool Changed = false;
1403 // Scan all of the blocks in the function, except for the entry block.
1404 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
1405 BasicBlock *BB = &*I++;
1406 // If the destination block has a single pred, then this is a trivial
1407 // edge, just collapse it.
1408 BasicBlock *SinglePred = BB->getSinglePredecessor();
1410 // Don't merge if BB's address is taken.
1411 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
1413 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
1414 if (Term && !Term->isConditional()) {
1416 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
1417 // Remember if SinglePred was the entry block of the function.
1418 // If so, we will need to move BB back to the entry position.
1419 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
1420 MergeBasicBlockIntoOnlyPred(BB, nullptr);
1422 if (isEntry && BB != &BB->getParent()->getEntryBlock())
1423 BB->moveBefore(&BB->getParent()->getEntryBlock());
1425 // We have erased a block. Update the iterator.
1426 I = BB->getIterator();
1432 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
1433 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
1434 /// edges in ways that are non-optimal for isel. Start by eliminating these
1435 /// blocks so we can split them the way we want them.
1436 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
1437 bool MadeChange = false;
1438 // Note that this intentionally skips the entry block.
1439 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
1440 BasicBlock *BB = &*I++;
1441 // If this block doesn't end with an uncond branch, ignore it.
1442 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
1443 if (!BI || !BI->isUnconditional())
1446 // If the instruction before the branch (skipping debug info) isn't a phi
1447 // node, then other stuff is happening here.
1448 BasicBlock::iterator BBI = BI->getIterator();
1449 if (BBI != BB->begin()) {
1451 while (isa<DbgInfoIntrinsic>(BBI)) {
1452 if (BBI == BB->begin())
1456 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
1460 // Do not break infinite loops.
1461 BasicBlock *DestBB = BI->getSuccessor(0);
1465 if (!canMergeBlocks(BB, DestBB))
1468 eliminateMostlyEmptyBlock(BB);
1474 /// Return true if we can merge BB into DestBB if there is a single
1475 /// unconditional branch between them, and BB contains no other non-phi
1477 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
1478 const BasicBlock *DestBB) const {
1479 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
1480 // the successor. If there are more complex condition (e.g. preheaders),
1481 // don't mess around with them.
1482 BasicBlock::const_iterator BBI = BB->begin();
1483 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
1484 for (const User *U : PN->users()) {
1485 const Instruction *UI = cast<Instruction>(U);
1486 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
1488 // IfUser is inside DestBB block and it is a PHINode then check
1489 // incoming value. If incoming value is not from BB then this is
1490 // a complex condition (e.g. preheaders) we want to avoid here.
1491 if (UI->getParent() == DestBB) {
1492 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
1493 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
1494 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
1495 if (Insn && Insn->getParent() == BB &&
1496 Insn->getParent() != UPN->getIncomingBlock(I))
1503 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
1504 // and DestBB may have conflicting incoming values for the block. If so, we
1505 // can't merge the block.
1506 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
1507 if (!DestBBPN) return true; // no conflict.
1509 // Collect the preds of BB.
1510 SmallPtrSet<const BasicBlock*, 16> BBPreds;
1511 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
1512 // It is faster to get preds from a PHI than with pred_iterator.
1513 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1514 BBPreds.insert(BBPN->getIncomingBlock(i));
1516 BBPreds.insert(pred_begin(BB), pred_end(BB));
1519 // Walk the preds of DestBB.
1520 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
1521 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
1522 if (BBPreds.count(Pred)) { // Common predecessor?
1523 BBI = DestBB->begin();
1524 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
1525 const Value *V1 = PN->getIncomingValueForBlock(Pred);
1526 const Value *V2 = PN->getIncomingValueForBlock(BB);
1528 // If V2 is a phi node in BB, look up what the mapped value will be.
1529 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
1530 if (V2PN->getParent() == BB)
1531 V2 = V2PN->getIncomingValueForBlock(Pred);
1533 // If there is a conflict, bail out.
1534 if (V1 != V2) return false;
1543 /// Eliminate a basic block that has only phi's and an unconditional branch in
1545 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
1546 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1547 BasicBlock *DestBB = BI->getSuccessor(0);
1549 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
1551 // If the destination block has a single pred, then this is a trivial edge,
1552 // just collapse it.
1553 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
1554 if (SinglePred != DestBB) {
1555 // Remember if SinglePred was the entry block of the function. If so, we
1556 // will need to move BB back to the entry position.
1557 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
1558 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
1560 if (isEntry && BB != &BB->getParent()->getEntryBlock())
1561 BB->moveBefore(&BB->getParent()->getEntryBlock());
1563 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1568 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
1569 // to handle the new incoming edges it is about to have.
1571 for (BasicBlock::iterator BBI = DestBB->begin();
1572 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1573 // Remove the incoming value for BB, and remember it.
1574 Value *InVal = PN->removeIncomingValue(BB, false);
1576 // Two options: either the InVal is a phi node defined in BB or it is some
1577 // value that dominates BB.
1578 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
1579 if (InValPhi && InValPhi->getParent() == BB) {
1580 // Add all of the input values of the input PHI as inputs of this phi.
1581 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
1582 PN->addIncoming(InValPhi->getIncomingValue(i),
1583 InValPhi->getIncomingBlock(i));
1585 // Otherwise, add one instance of the dominating value for each edge that
1586 // we will be adding.
1587 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
1588 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1589 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
1591 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
1592 PN->addIncoming(InVal, *PI);
1597 // The PHIs are now updated, change everything that refers to BB to use
1598 // DestBB and remove BB.
1599 BB->replaceAllUsesWith(DestBB);
1600 BB->eraseFromParent();
1603 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1606 // Computes a map of base pointer relocation instructions to corresponding
1607 // derived pointer relocation instructions given a vector of all relocate calls
1608 static void computeBaseDerivedRelocateMap(
1609 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
1610 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
1612 // Collect information in two maps: one primarily for locating the base object
1613 // while filling the second map; the second map is the final structure holding
1614 // a mapping between Base and corresponding Derived relocate calls
1615 DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
1616 for (auto *ThisRelocate : AllRelocateCalls) {
1617 auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
1618 ThisRelocate->getDerivedPtrIndex());
1619 RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
1621 for (auto &Item : RelocateIdxMap) {
1622 std::pair<unsigned, unsigned> Key = Item.first;
1623 if (Key.first == Key.second)
1624 // Base relocation: nothing to insert
1627 GCRelocateInst *I = Item.second;
1628 auto BaseKey = std::make_pair(Key.first, Key.first);
1630 // We're iterating over RelocateIdxMap so we cannot modify it.
1631 auto MaybeBase = RelocateIdxMap.find(BaseKey);
1632 if (MaybeBase == RelocateIdxMap.end())
1633 // TODO: We might want to insert a new base object relocate and gep off
1634 // that, if there are enough derived object relocates.
1637 RelocateInstMap[MaybeBase->second].push_back(I);
1641 // Accepts a GEP and extracts the operands into a vector provided they're all
1642 // small integer constants
1643 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
1644 SmallVectorImpl<Value *> &OffsetV) {
1645 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
1646 // Only accept small constant integer operands
1647 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
1648 if (!Op || Op->getZExtValue() > 20)
1652 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
1653 OffsetV.push_back(GEP->getOperand(i));
1657 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
1658 // replace, computes a replacement, and affects it.
1660 simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
1661 const SmallVectorImpl<GCRelocateInst *> &Targets) {
1662 bool MadeChange = false;
1663 for (GCRelocateInst *ToReplace : Targets) {
1664 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
1665 "Not relocating a derived object of the original base object");
1666 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
1667 // A duplicate relocate call. TODO: coalesce duplicates.
1671 if (RelocatedBase->getParent() != ToReplace->getParent()) {
1672 // Base and derived relocates are in different basic blocks.
1673 // In this case transform is only valid when base dominates derived
1674 // relocate. However it would be too expensive to check dominance
1675 // for each such relocate, so we skip the whole transformation.
1679 Value *Base = ToReplace->getBasePtr();
1680 auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
1681 if (!Derived || Derived->getPointerOperand() != Base)
1684 SmallVector<Value *, 2> OffsetV;
1685 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
1688 // Create a Builder and replace the target callsite with a gep
1689 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
1691 // Insert after RelocatedBase
1692 IRBuilder<> Builder(RelocatedBase->getNextNode());
1693 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1695 // If gc_relocate does not match the actual type, cast it to the right type.
1696 // In theory, there must be a bitcast after gc_relocate if the type does not
1697 // match, and we should reuse it to get the derived pointer. But it could be
1701 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1706 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1710 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
1711 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
1713 // In this case, we can not find the bitcast any more. So we insert a new bitcast
1714 // no matter there is already one or not. In this way, we can handle all cases, and
1715 // the extra bitcast should be optimized away in later passes.
1716 Value *ActualRelocatedBase = RelocatedBase;
1717 if (RelocatedBase->getType() != Base->getType()) {
1718 ActualRelocatedBase =
1719 Builder.CreateBitCast(RelocatedBase, Base->getType());
1721 Value *Replacement = Builder.CreateGEP(
1722 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
1723 Replacement->takeName(ToReplace);
1724 // If the newly generated derived pointer's type does not match the original derived
1725 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
1726 Value *ActualReplacement = Replacement;
1727 if (Replacement->getType() != ToReplace->getType()) {
1729 Builder.CreateBitCast(Replacement, ToReplace->getType());
1731 ToReplace->replaceAllUsesWith(ActualReplacement);
1732 ToReplace->eraseFromParent();
1742 // %ptr = gep %base + 15
1743 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1744 // %base' = relocate(%tok, i32 4, i32 4)
1745 // %ptr' = relocate(%tok, i32 4, i32 5)
1746 // %val = load %ptr'
1751 // %ptr = gep %base + 15
1752 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1753 // %base' = gc.relocate(%tok, i32 4, i32 4)
1754 // %ptr' = gep %base' + 15
1755 // %val = load %ptr'
1756 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
1757 bool MadeChange = false;
1758 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1760 for (auto *U : I.users())
1761 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1762 // Collect all the relocate calls associated with a statepoint
1763 AllRelocateCalls.push_back(Relocate);
1765 // We need atleast one base pointer relocation + one derived pointer
1766 // relocation to mangle
1767 if (AllRelocateCalls.size() < 2)
1770 // RelocateInstMap is a mapping from the base relocate instruction to the
1771 // corresponding derived relocate instructions
1772 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
1773 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1774 if (RelocateInstMap.empty())
1777 for (auto &Item : RelocateInstMap)
1778 // Item.first is the RelocatedBase to offset against
1779 // Item.second is the vector of Targets to replace
1780 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1784 /// SinkCast - Sink the specified cast instruction into its user blocks
1785 static bool SinkCast(CastInst *CI) {
1786 BasicBlock *DefBB = CI->getParent();
1788 /// InsertedCasts - Only insert a cast in each block once.
1789 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1791 bool MadeChange = false;
1792 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1794 Use &TheUse = UI.getUse();
1795 Instruction *User = cast<Instruction>(*UI);
1797 // Figure out which BB this cast is used in. For PHI's this is the
1798 // appropriate predecessor block.
1799 BasicBlock *UserBB = User->getParent();
1800 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1801 UserBB = PN->getIncomingBlock(TheUse);
1804 // Preincrement use iterator so we don't invalidate it.
1807 // If the block selected to receive the cast is an EH pad that does not
1808 // allow non-PHI instructions before the terminator, we can't sink the
1810 if (UserBB->getTerminator()->isEHPad())
1813 // If this user is in the same block as the cast, don't change the cast.
1814 if (UserBB == DefBB) continue;
1816 // If we have already inserted a cast into this block, use it.
1817 CastInst *&InsertedCast = InsertedCasts[UserBB];
1819 if (!InsertedCast) {
1820 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1821 assert(InsertPt != UserBB->end());
1822 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1823 CI->getType(), "", &*InsertPt);
1826 // Replace a use of the cast with a use of the new cast.
1827 TheUse = InsertedCast;
1832 // If we removed all uses, nuke the cast.
1833 if (CI->use_empty()) {
1834 CI->eraseFromParent();
1841 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1842 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1843 /// reduce the number of virtual registers that must be created and coalesced.
1845 /// Return true if any changes are made.
1847 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1848 const DataLayout &DL) {
1849 // If this is a noop copy,
1850 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1851 EVT DstVT = TLI.getValueType(DL, CI->getType());
1853 // This is an fp<->int conversion?
1854 if (SrcVT.isInteger() != DstVT.isInteger())
1857 // If this is an extension, it will be a zero or sign extension, which
1859 if (SrcVT.bitsLT(DstVT)) return false;
1861 // If these values will be promoted, find out what they will be promoted
1862 // to. This helps us consider truncates on PPC as noop copies when they
1864 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1865 TargetLowering::TypePromoteInteger)
1866 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1867 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1868 TargetLowering::TypePromoteInteger)
1869 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1871 // If, after promotion, these are the same types, this is a noop copy.
1875 return SinkCast(CI);
1878 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
1881 /// Return true if any changes were made.
1882 static bool CombineUAddWithOverflow(CmpInst *CI) {
1886 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
1889 Type *Ty = AddI->getType();
1890 if (!isa<IntegerType>(Ty))
1893 // We don't want to move around uses of condition values this late, so we we
1894 // check if it is legal to create the call to the intrinsic in the basic
1895 // block containing the icmp:
1897 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
1901 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
1903 if (AddI->hasOneUse())
1904 assert(*AddI->user_begin() == CI && "expected!");
1907 Module *M = CI->getModule();
1908 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1910 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
1912 auto *UAddWithOverflow =
1913 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
1914 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
1916 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
1918 CI->replaceAllUsesWith(Overflow);
1919 AddI->replaceAllUsesWith(UAdd);
1920 CI->eraseFromParent();
1921 AddI->eraseFromParent();
1925 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1926 /// registers that must be created and coalesced. This is a clear win except on
1927 /// targets with multiple condition code registers (PowerPC), where it might
1928 /// lose; some adjustment may be wanted there.
1930 /// Return true if any changes are made.
1931 static bool SinkCmpExpression(CmpInst *CI) {
1932 BasicBlock *DefBB = CI->getParent();
1934 /// Only insert a cmp in each block once.
1935 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1937 bool MadeChange = false;
1938 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1940 Use &TheUse = UI.getUse();
1941 Instruction *User = cast<Instruction>(*UI);
1943 // Preincrement use iterator so we don't invalidate it.
1946 // Don't bother for PHI nodes.
1947 if (isa<PHINode>(User))
1950 // Figure out which BB this cmp is used in.
1951 BasicBlock *UserBB = User->getParent();
1953 // If this user is in the same block as the cmp, don't change the cmp.
1954 if (UserBB == DefBB) continue;
1956 // If we have already inserted a cmp into this block, use it.
1957 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1960 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1961 assert(InsertPt != UserBB->end());
1963 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1964 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
1967 // Replace a use of the cmp with a use of the new cmp.
1968 TheUse = InsertedCmp;
1973 // If we removed all uses, nuke the cmp.
1974 if (CI->use_empty()) {
1975 CI->eraseFromParent();
1982 static bool OptimizeCmpExpression(CmpInst *CI) {
1983 if (SinkCmpExpression(CI))
1986 if (CombineUAddWithOverflow(CI))
1992 /// Check if the candidates could be combined with a shift instruction, which
1994 /// 1. Truncate instruction
1995 /// 2. And instruction and the imm is a mask of the low bits:
1996 /// imm & (imm+1) == 0
1997 static bool isExtractBitsCandidateUse(Instruction *User) {
1998 if (!isa<TruncInst>(User)) {
1999 if (User->getOpcode() != Instruction::And ||
2000 !isa<ConstantInt>(User->getOperand(1)))
2003 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
2005 if ((Cimm & (Cimm + 1)).getBoolValue())
2011 /// Sink both shift and truncate instruction to the use of truncate's BB.
2013 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
2014 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
2015 const TargetLowering &TLI, const DataLayout &DL) {
2016 BasicBlock *UserBB = User->getParent();
2017 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
2018 TruncInst *TruncI = dyn_cast<TruncInst>(User);
2019 bool MadeChange = false;
2021 for (Value::user_iterator TruncUI = TruncI->user_begin(),
2022 TruncE = TruncI->user_end();
2023 TruncUI != TruncE;) {
2025 Use &TruncTheUse = TruncUI.getUse();
2026 Instruction *TruncUser = cast<Instruction>(*TruncUI);
2027 // Preincrement use iterator so we don't invalidate it.
2031 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
2035 // If the use is actually a legal node, there will not be an
2036 // implicit truncate.
2037 // FIXME: always querying the result type is just an
2038 // approximation; some nodes' legality is determined by the
2039 // operand or other means. There's no good way to find out though.
2040 if (TLI.isOperationLegalOrCustom(
2041 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
2044 // Don't bother for PHI nodes.
2045 if (isa<PHINode>(TruncUser))
2048 BasicBlock *TruncUserBB = TruncUser->getParent();
2050 if (UserBB == TruncUserBB)
2053 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
2054 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
2056 if (!InsertedShift && !InsertedTrunc) {
2057 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
2058 assert(InsertPt != TruncUserBB->end());
2060 if (ShiftI->getOpcode() == Instruction::AShr)
2061 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
2064 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
2068 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
2070 assert(TruncInsertPt != TruncUserBB->end());
2072 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
2073 TruncI->getType(), "", &*TruncInsertPt);
2077 TruncTheUse = InsertedTrunc;
2083 /// Sink the shift *right* instruction into user blocks if the uses could
2084 /// potentially be combined with this shift instruction and generate BitExtract
2085 /// instruction. It will only be applied if the architecture supports BitExtract
2086 /// instruction. Here is an example:
2088 /// %x.extract.shift = lshr i64 %arg1, 32
2090 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
2094 /// %x.extract.shift.1 = lshr i64 %arg1, 32
2095 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
2097 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
2099 /// Return true if any changes are made.
2100 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
2101 const TargetLowering &TLI,
2102 const DataLayout &DL) {
2103 BasicBlock *DefBB = ShiftI->getParent();
2105 /// Only insert instructions in each block once.
2106 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
2108 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
2110 bool MadeChange = false;
2111 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
2113 Use &TheUse = UI.getUse();
2114 Instruction *User = cast<Instruction>(*UI);
2115 // Preincrement use iterator so we don't invalidate it.
2118 // Don't bother for PHI nodes.
2119 if (isa<PHINode>(User))
2122 if (!isExtractBitsCandidateUse(User))
2125 BasicBlock *UserBB = User->getParent();
2127 if (UserBB == DefBB) {
2128 // If the shift and truncate instruction are in the same BB. The use of
2129 // the truncate(TruncUse) may still introduce another truncate if not
2130 // legal. In this case, we would like to sink both shift and truncate
2131 // instruction to the BB of TruncUse.
2134 // i64 shift.result = lshr i64 opnd, imm
2135 // trunc.result = trunc shift.result to i16
2138 // ----> We will have an implicit truncate here if the architecture does
2139 // not have i16 compare.
2140 // cmp i16 trunc.result, opnd2
2142 if (isa<TruncInst>(User) && shiftIsLegal
2143 // If the type of the truncate is legal, no trucate will be
2144 // introduced in other basic blocks.
2146 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
2148 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
2152 // If we have already inserted a shift into this block, use it.
2153 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
2155 if (!InsertedShift) {
2156 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2157 assert(InsertPt != UserBB->end());
2159 if (ShiftI->getOpcode() == Instruction::AShr)
2160 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
2163 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
2169 // Replace a use of the shift with a use of the new shift.
2170 TheUse = InsertedShift;
2173 // If we removed all uses, nuke the shift.
2174 if (ShiftI->use_empty())
2175 ShiftI->eraseFromParent();
2180 // Translate a masked load intrinsic like
2181 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
2182 // <16 x i1> %mask, <16 x i32> %passthru)
2183 // to a chain of basic blocks, with loading element one-by-one if
2184 // the appropriate mask bit is set
2186 // %1 = bitcast i8* %addr to i32*
2187 // %2 = extractelement <16 x i1> %mask, i32 0
2188 // %3 = icmp eq i1 %2, true
2189 // br i1 %3, label %cond.load, label %else
2191 //cond.load: ; preds = %0
2192 // %4 = getelementptr i32* %1, i32 0
2193 // %5 = load i32* %4
2194 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
2197 //else: ; preds = %0, %cond.load
2198 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
2199 // %7 = extractelement <16 x i1> %mask, i32 1
2200 // %8 = icmp eq i1 %7, true
2201 // br i1 %8, label %cond.load1, label %else2
2203 //cond.load1: ; preds = %else
2204 // %9 = getelementptr i32* %1, i32 1
2205 // %10 = load i32* %9
2206 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
2209 //else2: ; preds = %else, %cond.load1
2210 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
2211 // %12 = extractelement <16 x i1> %mask, i32 2
2212 // %13 = icmp eq i1 %12, true
2213 // br i1 %13, label %cond.load4, label %else5
2215 static void ScalarizeMaskedLoad(CallInst *CI) {
2216 Value *Ptr = CI->getArgOperand(0);
2217 Value *Alignment = CI->getArgOperand(1);
2218 Value *Mask = CI->getArgOperand(2);
2219 Value *Src0 = CI->getArgOperand(3);
2221 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2222 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
2223 assert(VecType && "Unexpected return type of masked load intrinsic");
2225 Type *EltTy = CI->getType()->getVectorElementType();
2227 IRBuilder<> Builder(CI->getContext());
2228 Instruction *InsertPt = CI;
2229 BasicBlock *IfBlock = CI->getParent();
2230 BasicBlock *CondBlock = nullptr;
2231 BasicBlock *PrevIfBlock = CI->getParent();
2233 Builder.SetInsertPoint(InsertPt);
2234 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2236 // Short-cut if the mask is all-true.
2237 bool IsAllOnesMask = isa<Constant>(Mask) &&
2238 cast<Constant>(Mask)->isAllOnesValue();
2240 if (IsAllOnesMask) {
2241 Value *NewI = Builder.CreateAlignedLoad(Ptr, AlignVal);
2242 CI->replaceAllUsesWith(NewI);
2243 CI->eraseFromParent();
2247 // Adjust alignment for the scalar instruction.
2248 AlignVal = std::min(AlignVal, VecType->getScalarSizeInBits()/8);
2249 // Bitcast %addr fron i8* to EltTy*
2251 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
2252 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
2253 unsigned VectorWidth = VecType->getNumElements();
2255 Value *UndefVal = UndefValue::get(VecType);
2257 // The result vector
2258 Value *VResult = UndefVal;
2260 if (isa<ConstantVector>(Mask)) {
2261 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2262 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2265 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2266 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
2267 VResult = Builder.CreateInsertElement(VResult, Load,
2268 Builder.getInt32(Idx));
2270 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
2271 CI->replaceAllUsesWith(NewI);
2272 CI->eraseFromParent();
2276 PHINode *Phi = nullptr;
2277 Value *PrevPhi = UndefVal;
2279 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2281 // Fill the "else" block, created in the previous iteration
2283 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
2284 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
2285 // %to_load = icmp eq i1 %mask_1, true
2286 // br i1 %to_load, label %cond.load, label %else
2289 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
2290 Phi->addIncoming(VResult, CondBlock);
2291 Phi->addIncoming(PrevPhi, PrevIfBlock);
2296 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
2297 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2298 ConstantInt::get(Predicate->getType(), 1));
2300 // Create "cond" block
2302 // %EltAddr = getelementptr i32* %1, i32 0
2303 // %Elt = load i32* %EltAddr
2304 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
2306 CondBlock = IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.load");
2307 Builder.SetInsertPoint(InsertPt);
2310 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2311 LoadInst *Load = Builder.CreateAlignedLoad(Gep, AlignVal);
2312 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
2314 // Create "else" block, fill it in the next iteration
2315 BasicBlock *NewIfBlock =
2316 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
2317 Builder.SetInsertPoint(InsertPt);
2318 Instruction *OldBr = IfBlock->getTerminator();
2319 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2320 OldBr->eraseFromParent();
2321 PrevIfBlock = IfBlock;
2322 IfBlock = NewIfBlock;
2325 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
2326 Phi->addIncoming(VResult, CondBlock);
2327 Phi->addIncoming(PrevPhi, PrevIfBlock);
2328 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
2329 CI->replaceAllUsesWith(NewI);
2330 CI->eraseFromParent();
2333 // Translate a masked store intrinsic, like
2334 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
2336 // to a chain of basic blocks, that stores element one-by-one if
2337 // the appropriate mask bit is set
2339 // %1 = bitcast i8* %addr to i32*
2340 // %2 = extractelement <16 x i1> %mask, i32 0
2341 // %3 = icmp eq i1 %2, true
2342 // br i1 %3, label %cond.store, label %else
2344 // cond.store: ; preds = %0
2345 // %4 = extractelement <16 x i32> %val, i32 0
2346 // %5 = getelementptr i32* %1, i32 0
2347 // store i32 %4, i32* %5
2350 // else: ; preds = %0, %cond.store
2351 // %6 = extractelement <16 x i1> %mask, i32 1
2352 // %7 = icmp eq i1 %6, true
2353 // br i1 %7, label %cond.store1, label %else2
2355 // cond.store1: ; preds = %else
2356 // %8 = extractelement <16 x i32> %val, i32 1
2357 // %9 = getelementptr i32* %1, i32 1
2358 // store i32 %8, i32* %9
2361 static void ScalarizeMaskedStore(CallInst *CI) {
2362 Value *Src = CI->getArgOperand(0);
2363 Value *Ptr = CI->getArgOperand(1);
2364 Value *Alignment = CI->getArgOperand(2);
2365 Value *Mask = CI->getArgOperand(3);
2367 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2368 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
2369 assert(VecType && "Unexpected data type in masked store intrinsic");
2371 Type *EltTy = VecType->getElementType();
2373 IRBuilder<> Builder(CI->getContext());
2374 Instruction *InsertPt = CI;
2375 BasicBlock *IfBlock = CI->getParent();
2376 Builder.SetInsertPoint(InsertPt);
2377 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2379 // Short-cut if the mask is all-true.
2380 bool IsAllOnesMask = isa<Constant>(Mask) &&
2381 cast<Constant>(Mask)->isAllOnesValue();
2383 if (IsAllOnesMask) {
2384 Builder.CreateAlignedStore(Src, Ptr, AlignVal);
2385 CI->eraseFromParent();
2389 // Adjust alignment for the scalar instruction.
2390 AlignVal = std::max(AlignVal, VecType->getScalarSizeInBits()/8);
2391 // Bitcast %addr fron i8* to EltTy*
2393 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
2394 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
2395 unsigned VectorWidth = VecType->getNumElements();
2397 if (isa<ConstantVector>(Mask)) {
2398 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2399 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2401 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
2403 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2404 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
2406 CI->eraseFromParent();
2410 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2412 // Fill the "else" block, created in the previous iteration
2414 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
2415 // %to_store = icmp eq i1 %mask_1, true
2416 // br i1 %to_store, label %cond.store, label %else
2418 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
2419 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2420 ConstantInt::get(Predicate->getType(), 1));
2422 // Create "cond" block
2424 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
2425 // %EltAddr = getelementptr i32* %1, i32 0
2426 // %store i32 %OneElt, i32* %EltAddr
2428 BasicBlock *CondBlock =
2429 IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.store");
2430 Builder.SetInsertPoint(InsertPt);
2432 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
2434 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2435 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
2437 // Create "else" block, fill it in the next iteration
2438 BasicBlock *NewIfBlock =
2439 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
2440 Builder.SetInsertPoint(InsertPt);
2441 Instruction *OldBr = IfBlock->getTerminator();
2442 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2443 OldBr->eraseFromParent();
2444 IfBlock = NewIfBlock;
2446 CI->eraseFromParent();
2449 // Translate a masked gather intrinsic like
2450 // <16 x i32 > @llvm.masked.gather.v16i32( <16 x i32*> %Ptrs, i32 4,
2451 // <16 x i1> %Mask, <16 x i32> %Src)
2452 // to a chain of basic blocks, with loading element one-by-one if
2453 // the appropriate mask bit is set
2455 // % Ptrs = getelementptr i32, i32* %base, <16 x i64> %ind
2456 // % Mask0 = extractelement <16 x i1> %Mask, i32 0
2457 // % ToLoad0 = icmp eq i1 % Mask0, true
2458 // br i1 % ToLoad0, label %cond.load, label %else
2461 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
2462 // % Load0 = load i32, i32* % Ptr0, align 4
2463 // % Res0 = insertelement <16 x i32> undef, i32 % Load0, i32 0
2467 // %res.phi.else = phi <16 x i32>[% Res0, %cond.load], [undef, % 0]
2468 // % Mask1 = extractelement <16 x i1> %Mask, i32 1
2469 // % ToLoad1 = icmp eq i1 % Mask1, true
2470 // br i1 % ToLoad1, label %cond.load1, label %else2
2473 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2474 // % Load1 = load i32, i32* % Ptr1, align 4
2475 // % Res1 = insertelement <16 x i32> %res.phi.else, i32 % Load1, i32 1
2478 // % Result = select <16 x i1> %Mask, <16 x i32> %res.phi.select, <16 x i32> %Src
2479 // ret <16 x i32> %Result
2480 static void ScalarizeMaskedGather(CallInst *CI) {
2481 Value *Ptrs = CI->getArgOperand(0);
2482 Value *Alignment = CI->getArgOperand(1);
2483 Value *Mask = CI->getArgOperand(2);
2484 Value *Src0 = CI->getArgOperand(3);
2486 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
2488 assert(VecType && "Unexpected return type of masked load intrinsic");
2490 IRBuilder<> Builder(CI->getContext());
2491 Instruction *InsertPt = CI;
2492 BasicBlock *IfBlock = CI->getParent();
2493 BasicBlock *CondBlock = nullptr;
2494 BasicBlock *PrevIfBlock = CI->getParent();
2495 Builder.SetInsertPoint(InsertPt);
2496 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2498 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2500 Value *UndefVal = UndefValue::get(VecType);
2502 // The result vector
2503 Value *VResult = UndefVal;
2504 unsigned VectorWidth = VecType->getNumElements();
2506 // Shorten the way if the mask is a vector of constants.
2507 bool IsConstMask = isa<ConstantVector>(Mask);
2510 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2511 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2513 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2514 "Ptr" + Twine(Idx));
2515 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
2516 "Load" + Twine(Idx));
2517 VResult = Builder.CreateInsertElement(VResult, Load,
2518 Builder.getInt32(Idx),
2519 "Res" + Twine(Idx));
2521 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
2522 CI->replaceAllUsesWith(NewI);
2523 CI->eraseFromParent();
2527 PHINode *Phi = nullptr;
2528 Value *PrevPhi = UndefVal;
2530 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2532 // Fill the "else" block, created in the previous iteration
2534 // %Mask1 = extractelement <16 x i1> %Mask, i32 1
2535 // %ToLoad1 = icmp eq i1 %Mask1, true
2536 // br i1 %ToLoad1, label %cond.load, label %else
2539 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
2540 Phi->addIncoming(VResult, CondBlock);
2541 Phi->addIncoming(PrevPhi, PrevIfBlock);
2546 Value *Predicate = Builder.CreateExtractElement(Mask,
2547 Builder.getInt32(Idx),
2548 "Mask" + Twine(Idx));
2549 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2550 ConstantInt::get(Predicate->getType(), 1),
2551 "ToLoad" + Twine(Idx));
2553 // Create "cond" block
2555 // %EltAddr = getelementptr i32* %1, i32 0
2556 // %Elt = load i32* %EltAddr
2557 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
2559 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
2560 Builder.SetInsertPoint(InsertPt);
2562 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2563 "Ptr" + Twine(Idx));
2564 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
2565 "Load" + Twine(Idx));
2566 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx),
2567 "Res" + Twine(Idx));
2569 // Create "else" block, fill it in the next iteration
2570 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2571 Builder.SetInsertPoint(InsertPt);
2572 Instruction *OldBr = IfBlock->getTerminator();
2573 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2574 OldBr->eraseFromParent();
2575 PrevIfBlock = IfBlock;
2576 IfBlock = NewIfBlock;
2579 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
2580 Phi->addIncoming(VResult, CondBlock);
2581 Phi->addIncoming(PrevPhi, PrevIfBlock);
2582 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
2583 CI->replaceAllUsesWith(NewI);
2584 CI->eraseFromParent();
2587 // Translate a masked scatter intrinsic, like
2588 // void @llvm.masked.scatter.v16i32(<16 x i32> %Src, <16 x i32*>* %Ptrs, i32 4,
2590 // to a chain of basic blocks, that stores element one-by-one if
2591 // the appropriate mask bit is set.
2593 // % Ptrs = getelementptr i32, i32* %ptr, <16 x i64> %ind
2594 // % Mask0 = extractelement <16 x i1> % Mask, i32 0
2595 // % ToStore0 = icmp eq i1 % Mask0, true
2596 // br i1 %ToStore0, label %cond.store, label %else
2599 // % Elt0 = extractelement <16 x i32> %Src, i32 0
2600 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
2601 // store i32 %Elt0, i32* % Ptr0, align 4
2605 // % Mask1 = extractelement <16 x i1> % Mask, i32 1
2606 // % ToStore1 = icmp eq i1 % Mask1, true
2607 // br i1 % ToStore1, label %cond.store1, label %else2
2610 // % Elt1 = extractelement <16 x i32> %Src, i32 1
2611 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2612 // store i32 % Elt1, i32* % Ptr1, align 4
2615 static void ScalarizeMaskedScatter(CallInst *CI) {
2616 Value *Src = CI->getArgOperand(0);
2617 Value *Ptrs = CI->getArgOperand(1);
2618 Value *Alignment = CI->getArgOperand(2);
2619 Value *Mask = CI->getArgOperand(3);
2621 assert(isa<VectorType>(Src->getType()) &&
2622 "Unexpected data type in masked scatter intrinsic");
2623 assert(isa<VectorType>(Ptrs->getType()) &&
2624 isa<PointerType>(Ptrs->getType()->getVectorElementType()) &&
2625 "Vector of pointers is expected in masked scatter intrinsic");
2627 IRBuilder<> Builder(CI->getContext());
2628 Instruction *InsertPt = CI;
2629 BasicBlock *IfBlock = CI->getParent();
2630 Builder.SetInsertPoint(InsertPt);
2631 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2633 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2634 unsigned VectorWidth = Src->getType()->getVectorNumElements();
2636 // Shorten the way if the mask is a vector of constants.
2637 bool IsConstMask = isa<ConstantVector>(Mask);
2640 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2641 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2643 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
2644 "Elt" + Twine(Idx));
2645 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2646 "Ptr" + Twine(Idx));
2647 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
2649 CI->eraseFromParent();
2652 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2653 // Fill the "else" block, created in the previous iteration
2655 // % Mask1 = extractelement <16 x i1> % Mask, i32 Idx
2656 // % ToStore = icmp eq i1 % Mask1, true
2657 // br i1 % ToStore, label %cond.store, label %else
2659 Value *Predicate = Builder.CreateExtractElement(Mask,
2660 Builder.getInt32(Idx),
2661 "Mask" + Twine(Idx));
2663 Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2664 ConstantInt::get(Predicate->getType(), 1),
2665 "ToStore" + Twine(Idx));
2667 // Create "cond" block
2669 // % Elt1 = extractelement <16 x i32> %Src, i32 1
2670 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2671 // %store i32 % Elt1, i32* % Ptr1
2673 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2674 Builder.SetInsertPoint(InsertPt);
2676 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
2677 "Elt" + Twine(Idx));
2678 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2679 "Ptr" + Twine(Idx));
2680 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
2682 // Create "else" block, fill it in the next iteration
2683 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2684 Builder.SetInsertPoint(InsertPt);
2685 Instruction *OldBr = IfBlock->getTerminator();
2686 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2687 OldBr->eraseFromParent();
2688 IfBlock = NewIfBlock;
2690 CI->eraseFromParent();
2693 /// If counting leading or trailing zeros is an expensive operation and a zero
2694 /// input is defined, add a check for zero to avoid calling the intrinsic.
2696 /// We want to transform:
2697 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
2701 /// %cmpz = icmp eq i64 %A, 0
2702 /// br i1 %cmpz, label %cond.end, label %cond.false
2704 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
2705 /// br label %cond.end
2707 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
2709 /// If the transform is performed, return true and set ModifiedDT to true.
2710 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
2711 const TargetLowering *TLI,
2712 const DataLayout *DL,
2717 // If a zero input is undefined, it doesn't make sense to despeculate that.
2718 if (match(CountZeros->getOperand(1), m_One()))
2721 // If it's cheap to speculate, there's nothing to do.
2722 auto IntrinsicID = CountZeros->getIntrinsicID();
2723 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
2724 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
2727 // Only handle legal scalar cases. Anything else requires too much work.
2728 Type *Ty = CountZeros->getType();
2729 unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
2730 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSize())
2733 // The intrinsic will be sunk behind a compare against zero and branch.
2734 BasicBlock *StartBlock = CountZeros->getParent();
2735 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
2737 // Create another block after the count zero intrinsic. A PHI will be added
2738 // in this block to select the result of the intrinsic or the bit-width
2739 // constant if the input to the intrinsic is zero.
2740 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
2741 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
2743 // Set up a builder to create a compare, conditional branch, and PHI.
2744 IRBuilder<> Builder(CountZeros->getContext());
2745 Builder.SetInsertPoint(StartBlock->getTerminator());
2746 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
2748 // Replace the unconditional branch that was created by the first split with
2749 // a compare against zero and a conditional branch.
2750 Value *Zero = Constant::getNullValue(Ty);
2751 Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
2752 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
2753 StartBlock->getTerminator()->eraseFromParent();
2755 // Create a PHI in the end block to select either the output of the intrinsic
2756 // or the bit width of the operand.
2757 Builder.SetInsertPoint(&EndBlock->front());
2758 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
2759 CountZeros->replaceAllUsesWith(PN);
2760 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
2761 PN->addIncoming(BitWidth, StartBlock);
2762 PN->addIncoming(CountZeros, CallBlock);
2764 // We are explicitly handling the zero case, so we can set the intrinsic's
2765 // undefined zero argument to 'true'. This will also prevent reprocessing the
2766 // intrinsic; we only despeculate when a zero input is defined.
2767 CountZeros->setArgOperand(1, Builder.getTrue());
2772 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool& ModifiedDT) {
2773 BasicBlock *BB = CI->getParent();
2775 // Lower inline assembly if we can.
2776 // If we found an inline asm expession, and if the target knows how to
2777 // lower it to normal LLVM code, do so now.
2778 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
2779 if (TLI->ExpandInlineAsm(CI)) {
2780 // Avoid invalidating the iterator.
2781 CurInstIterator = BB->begin();
2782 // Avoid processing instructions out of order, which could cause
2783 // reuse before a value is defined.
2787 // Sink address computing for memory operands into the block.
2788 if (optimizeInlineAsmInst(CI))
2792 // Align the pointer arguments to this call if the target thinks it's a good
2794 unsigned MinSize, PrefAlign;
2795 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
2796 for (auto &Arg : CI->arg_operands()) {
2797 // We want to align both objects whose address is used directly and
2798 // objects whose address is used in casts and GEPs, though it only makes
2799 // sense for GEPs if the offset is a multiple of the desired alignment and
2800 // if size - offset meets the size threshold.
2801 if (!Arg->getType()->isPointerTy())
2803 APInt Offset(DL->getPointerSizeInBits(
2804 cast<PointerType>(Arg->getType())->getAddressSpace()),
2806 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
2807 uint64_t Offset2 = Offset.getLimitedValue();
2808 if ((Offset2 & (PrefAlign-1)) != 0)
2811 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
2812 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
2813 AI->setAlignment(PrefAlign);
2814 // Global variables can only be aligned if they are defined in this
2815 // object (i.e. they are uniquely initialized in this object), and
2816 // over-aligning global variables that have an explicit section is
2819 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
2820 GV->getAlignment() < PrefAlign &&
2821 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
2823 GV->setAlignment(PrefAlign);
2825 // If this is a memcpy (or similar) then we may be able to improve the
2827 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
2828 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
2829 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
2830 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
2831 if (Align > MI->getAlignment())
2832 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
2836 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
2838 switch (II->getIntrinsicID()) {
2840 case Intrinsic::objectsize: {
2841 // Lower all uses of llvm.objectsize.*
2842 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
2843 Type *ReturnTy = CI->getType();
2844 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
2846 // Substituting this can cause recursive simplifications, which can
2847 // invalidate our iterator. Use a WeakVH to hold onto it in case this
2849 WeakVH IterHandle(&*CurInstIterator);
2851 replaceAndRecursivelySimplify(CI, RetVal,
2854 // If the iterator instruction was recursively deleted, start over at the
2855 // start of the block.
2856 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
2857 CurInstIterator = BB->begin();
2862 case Intrinsic::masked_load: {
2863 // Scalarize unsupported vector masked load
2864 if (!TTI->isLegalMaskedLoad(CI->getType())) {
2865 ScalarizeMaskedLoad(CI);
2871 case Intrinsic::masked_store: {
2872 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType())) {
2873 ScalarizeMaskedStore(CI);
2879 case Intrinsic::masked_gather: {
2880 if (!TTI->isLegalMaskedGather(CI->getType())) {
2881 ScalarizeMaskedGather(CI);
2887 case Intrinsic::masked_scatter: {
2888 if (!TTI->isLegalMaskedScatter(CI->getArgOperand(0)->getType())) {
2889 ScalarizeMaskedScatter(CI);
2895 case Intrinsic::aarch64_stlxr:
2896 case Intrinsic::aarch64_stxr: {
2897 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
2898 if (!ExtVal || !ExtVal->hasOneUse() ||
2899 ExtVal->getParent() == CI->getParent())
2901 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
2902 ExtVal->moveBefore(CI);
2903 // Mark this instruction as "inserted by CGP", so that other
2904 // optimizations don't touch it.
2905 InsertedInsts.insert(ExtVal);
2908 case Intrinsic::invariant_group_barrier:
2909 II->replaceAllUsesWith(II->getArgOperand(0));
2910 II->eraseFromParent();
2913 case Intrinsic::cttz:
2914 case Intrinsic::ctlz:
2915 // If counting zeros is expensive, try to avoid it.
2916 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
2920 // Unknown address space.
2921 // TODO: Target hook to pick which address space the intrinsic cares
2923 unsigned AddrSpace = ~0u;
2924 SmallVector<Value*, 2> PtrOps;
2926 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
2927 while (!PtrOps.empty())
2928 if (optimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
2933 // From here on out we're working with named functions.
2934 if (!CI->getCalledFunction()) return false;
2936 // Lower all default uses of _chk calls. This is very similar
2937 // to what InstCombineCalls does, but here we are only lowering calls
2938 // to fortified library functions (e.g. __memcpy_chk) that have the default
2939 // "don't know" as the objectsize. Anything else should be left alone.
2940 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
2941 if (Value *V = Simplifier.optimizeCall(CI)) {
2942 CI->replaceAllUsesWith(V);
2943 CI->eraseFromParent();
2949 /// Look for opportunities to duplicate return instructions to the predecessor
2950 /// to enable tail call optimizations. The case it is currently looking for is:
2953 /// %tmp0 = tail call i32 @f0()
2954 /// br label %return
2956 /// %tmp1 = tail call i32 @f1()
2957 /// br label %return
2959 /// %tmp2 = tail call i32 @f2()
2960 /// br label %return
2962 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
2970 /// %tmp0 = tail call i32 @f0()
2973 /// %tmp1 = tail call i32 @f1()
2976 /// %tmp2 = tail call i32 @f2()
2979 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
2983 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
2987 PHINode *PN = nullptr;
2988 BitCastInst *BCI = nullptr;
2989 Value *V = RI->getReturnValue();
2991 BCI = dyn_cast<BitCastInst>(V);
2993 V = BCI->getOperand(0);
2995 PN = dyn_cast<PHINode>(V);
3000 if (PN && PN->getParent() != BB)
3003 // It's not safe to eliminate the sign / zero extension of the return value.
3004 // See llvm::isInTailCallPosition().
3005 const Function *F = BB->getParent();
3006 AttributeSet CallerAttrs = F->getAttributes();
3007 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
3008 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
3011 // Make sure there are no instructions between the PHI and return, or that the
3012 // return is the first instruction in the block.
3014 BasicBlock::iterator BI = BB->begin();
3015 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
3017 // Also skip over the bitcast.
3022 BasicBlock::iterator BI = BB->begin();
3023 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
3028 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
3030 SmallVector<CallInst*, 4> TailCalls;
3032 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
3033 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
3034 // Make sure the phi value is indeed produced by the tail call.
3035 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
3036 TLI->mayBeEmittedAsTailCall(CI))
3037 TailCalls.push_back(CI);
3040 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
3041 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
3042 if (!VisitedBBs.insert(*PI).second)
3045 BasicBlock::InstListType &InstList = (*PI)->getInstList();
3046 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
3047 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
3048 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
3052 CallInst *CI = dyn_cast<CallInst>(&*RI);
3053 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
3054 TailCalls.push_back(CI);
3058 bool Changed = false;
3059 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
3060 CallInst *CI = TailCalls[i];
3063 // Conservatively require the attributes of the call to match those of the
3064 // return. Ignore noalias because it doesn't affect the call sequence.
3065 AttributeSet CalleeAttrs = CS.getAttributes();
3066 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
3067 removeAttribute(Attribute::NoAlias) !=
3068 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
3069 removeAttribute(Attribute::NoAlias))
3072 // Make sure the call instruction is followed by an unconditional branch to
3073 // the return block.
3074 BasicBlock *CallBB = CI->getParent();
3075 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
3076 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
3079 // Duplicate the return into CallBB.
3080 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
3081 ModifiedDT = Changed = true;
3085 // If we eliminated all predecessors of the block, delete the block now.
3086 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
3087 BB->eraseFromParent();
3092 //===----------------------------------------------------------------------===//
3093 // Memory Optimization
3094 //===----------------------------------------------------------------------===//
3098 /// This is an extended version of TargetLowering::AddrMode
3099 /// which holds actual Value*'s for register values.
3100 struct ExtAddrMode : public TargetLowering::AddrMode {
3103 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
3104 void print(raw_ostream &OS) const;
3107 bool operator==(const ExtAddrMode& O) const {
3108 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
3109 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
3110 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
3115 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
3121 void ExtAddrMode::print(raw_ostream &OS) const {
3122 bool NeedPlus = false;
3125 OS << (NeedPlus ? " + " : "")
3127 BaseGV->printAsOperand(OS, /*PrintType=*/false);
3132 OS << (NeedPlus ? " + " : "")
3138 OS << (NeedPlus ? " + " : "")
3140 BaseReg->printAsOperand(OS, /*PrintType=*/false);
3144 OS << (NeedPlus ? " + " : "")
3146 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
3152 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3153 void ExtAddrMode::dump() const {
3159 /// \brief This class provides transaction based operation on the IR.
3160 /// Every change made through this class is recorded in the internal state and
3161 /// can be undone (rollback) until commit is called.
3162 class TypePromotionTransaction {
3164 /// \brief This represents the common interface of the individual transaction.
3165 /// Each class implements the logic for doing one specific modification on
3166 /// the IR via the TypePromotionTransaction.
3167 class TypePromotionAction {
3169 /// The Instruction modified.
3173 /// \brief Constructor of the action.
3174 /// The constructor performs the related action on the IR.
3175 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
3177 virtual ~TypePromotionAction() {}
3179 /// \brief Undo the modification done by this action.
3180 /// When this method is called, the IR must be in the same state as it was
3181 /// before this action was applied.
3182 /// \pre Undoing the action works if and only if the IR is in the exact same
3183 /// state as it was directly after this action was applied.
3184 virtual void undo() = 0;
3186 /// \brief Advocate every change made by this action.
3187 /// When the results on the IR of the action are to be kept, it is important
3188 /// to call this function, otherwise hidden information may be kept forever.
3189 virtual void commit() {
3190 // Nothing to be done, this action is not doing anything.
3194 /// \brief Utility to remember the position of an instruction.
3195 class InsertionHandler {
3196 /// Position of an instruction.
3197 /// Either an instruction:
3198 /// - Is the first in a basic block: BB is used.
3199 /// - Has a previous instructon: PrevInst is used.
3201 Instruction *PrevInst;
3204 /// Remember whether or not the instruction had a previous instruction.
3205 bool HasPrevInstruction;
3208 /// \brief Record the position of \p Inst.
3209 InsertionHandler(Instruction *Inst) {
3210 BasicBlock::iterator It = Inst->getIterator();
3211 HasPrevInstruction = (It != (Inst->getParent()->begin()));
3212 if (HasPrevInstruction)
3213 Point.PrevInst = &*--It;
3215 Point.BB = Inst->getParent();
3218 /// \brief Insert \p Inst at the recorded position.
3219 void insert(Instruction *Inst) {
3220 if (HasPrevInstruction) {
3221 if (Inst->getParent())
3222 Inst->removeFromParent();
3223 Inst->insertAfter(Point.PrevInst);
3225 Instruction *Position = &*Point.BB->getFirstInsertionPt();
3226 if (Inst->getParent())
3227 Inst->moveBefore(Position);
3229 Inst->insertBefore(Position);
3234 /// \brief Move an instruction before another.
3235 class InstructionMoveBefore : public TypePromotionAction {
3236 /// Original position of the instruction.
3237 InsertionHandler Position;
3240 /// \brief Move \p Inst before \p Before.
3241 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
3242 : TypePromotionAction(Inst), Position(Inst) {
3243 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
3244 Inst->moveBefore(Before);
3247 /// \brief Move the instruction back to its original position.
3248 void undo() override {
3249 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
3250 Position.insert(Inst);
3254 /// \brief Set the operand of an instruction with a new value.
3255 class OperandSetter : public TypePromotionAction {
3256 /// Original operand of the instruction.
3258 /// Index of the modified instruction.
3262 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
3263 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
3264 : TypePromotionAction(Inst), Idx(Idx) {
3265 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
3266 << "for:" << *Inst << "\n"
3267 << "with:" << *NewVal << "\n");
3268 Origin = Inst->getOperand(Idx);
3269 Inst->setOperand(Idx, NewVal);
3272 /// \brief Restore the original value of the instruction.
3273 void undo() override {
3274 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
3275 << "for: " << *Inst << "\n"
3276 << "with: " << *Origin << "\n");
3277 Inst->setOperand(Idx, Origin);
3281 /// \brief Hide the operands of an instruction.
3282 /// Do as if this instruction was not using any of its operands.
3283 class OperandsHider : public TypePromotionAction {
3284 /// The list of original operands.
3285 SmallVector<Value *, 4> OriginalValues;
3288 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
3289 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
3290 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
3291 unsigned NumOpnds = Inst->getNumOperands();
3292 OriginalValues.reserve(NumOpnds);
3293 for (unsigned It = 0; It < NumOpnds; ++It) {
3294 // Save the current operand.
3295 Value *Val = Inst->getOperand(It);
3296 OriginalValues.push_back(Val);
3298 // We could use OperandSetter here, but that would imply an overhead
3299 // that we are not willing to pay.
3300 Inst->setOperand(It, UndefValue::get(Val->getType()));
3304 /// \brief Restore the original list of uses.
3305 void undo() override {
3306 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
3307 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
3308 Inst->setOperand(It, OriginalValues[It]);
3312 /// \brief Build a truncate instruction.
3313 class TruncBuilder : public TypePromotionAction {
3316 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
3318 /// trunc Opnd to Ty.
3319 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
3320 IRBuilder<> Builder(Opnd);
3321 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
3322 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
3325 /// \brief Get the built value.
3326 Value *getBuiltValue() { return Val; }
3328 /// \brief Remove the built instruction.
3329 void undo() override {
3330 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
3331 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3332 IVal->eraseFromParent();
3336 /// \brief Build a sign extension instruction.
3337 class SExtBuilder : public TypePromotionAction {
3340 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
3342 /// sext Opnd to Ty.
3343 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3344 : TypePromotionAction(InsertPt) {
3345 IRBuilder<> Builder(InsertPt);
3346 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
3347 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
3350 /// \brief Get the built value.
3351 Value *getBuiltValue() { return Val; }
3353 /// \brief Remove the built instruction.
3354 void undo() override {
3355 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
3356 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3357 IVal->eraseFromParent();
3361 /// \brief Build a zero extension instruction.
3362 class ZExtBuilder : public TypePromotionAction {
3365 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
3367 /// zext Opnd to Ty.
3368 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3369 : TypePromotionAction(InsertPt) {
3370 IRBuilder<> Builder(InsertPt);
3371 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
3372 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
3375 /// \brief Get the built value.
3376 Value *getBuiltValue() { return Val; }
3378 /// \brief Remove the built instruction.
3379 void undo() override {
3380 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
3381 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3382 IVal->eraseFromParent();
3386 /// \brief Mutate an instruction to another type.
3387 class TypeMutator : public TypePromotionAction {
3388 /// Record the original type.
3392 /// \brief Mutate the type of \p Inst into \p NewTy.
3393 TypeMutator(Instruction *Inst, Type *NewTy)
3394 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
3395 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
3397 Inst->mutateType(NewTy);
3400 /// \brief Mutate the instruction back to its original type.
3401 void undo() override {
3402 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
3404 Inst->mutateType(OrigTy);
3408 /// \brief Replace the uses of an instruction by another instruction.
3409 class UsesReplacer : public TypePromotionAction {
3410 /// Helper structure to keep track of the replaced uses.
3411 struct InstructionAndIdx {
3412 /// The instruction using the instruction.
3414 /// The index where this instruction is used for Inst.
3416 InstructionAndIdx(Instruction *Inst, unsigned Idx)
3417 : Inst(Inst), Idx(Idx) {}
3420 /// Keep track of the original uses (pair Instruction, Index).
3421 SmallVector<InstructionAndIdx, 4> OriginalUses;
3422 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
3425 /// \brief Replace all the use of \p Inst by \p New.
3426 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
3427 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
3429 // Record the original uses.
3430 for (Use &U : Inst->uses()) {
3431 Instruction *UserI = cast<Instruction>(U.getUser());
3432 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
3434 // Now, we can replace the uses.
3435 Inst->replaceAllUsesWith(New);
3438 /// \brief Reassign the original uses of Inst to Inst.
3439 void undo() override {
3440 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
3441 for (use_iterator UseIt = OriginalUses.begin(),
3442 EndIt = OriginalUses.end();
3443 UseIt != EndIt; ++UseIt) {
3444 UseIt->Inst->setOperand(UseIt->Idx, Inst);
3449 /// \brief Remove an instruction from the IR.
3450 class InstructionRemover : public TypePromotionAction {
3451 /// Original position of the instruction.
3452 InsertionHandler Inserter;
3453 /// Helper structure to hide all the link to the instruction. In other
3454 /// words, this helps to do as if the instruction was removed.
3455 OperandsHider Hider;
3456 /// Keep track of the uses replaced, if any.
3457 UsesReplacer *Replacer;
3460 /// \brief Remove all reference of \p Inst and optinally replace all its
3462 /// \pre If !Inst->use_empty(), then New != nullptr
3463 InstructionRemover(Instruction *Inst, Value *New = nullptr)
3464 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
3467 Replacer = new UsesReplacer(Inst, New);
3468 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
3469 Inst->removeFromParent();
3472 ~InstructionRemover() override { delete Replacer; }
3474 /// \brief Really remove the instruction.
3475 void commit() override { delete Inst; }
3477 /// \brief Resurrect the instruction and reassign it to the proper uses if
3478 /// new value was provided when build this action.
3479 void undo() override {
3480 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
3481 Inserter.insert(Inst);
3489 /// Restoration point.
3490 /// The restoration point is a pointer to an action instead of an iterator
3491 /// because the iterator may be invalidated but not the pointer.
3492 typedef const TypePromotionAction *ConstRestorationPt;
3493 /// Advocate every changes made in that transaction.
3495 /// Undo all the changes made after the given point.
3496 void rollback(ConstRestorationPt Point);
3497 /// Get the current restoration point.
3498 ConstRestorationPt getRestorationPoint() const;
3500 /// \name API for IR modification with state keeping to support rollback.
3502 /// Same as Instruction::setOperand.
3503 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
3504 /// Same as Instruction::eraseFromParent.
3505 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
3506 /// Same as Value::replaceAllUsesWith.
3507 void replaceAllUsesWith(Instruction *Inst, Value *New);
3508 /// Same as Value::mutateType.
3509 void mutateType(Instruction *Inst, Type *NewTy);
3510 /// Same as IRBuilder::createTrunc.
3511 Value *createTrunc(Instruction *Opnd, Type *Ty);
3512 /// Same as IRBuilder::createSExt.
3513 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
3514 /// Same as IRBuilder::createZExt.
3515 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
3516 /// Same as Instruction::moveBefore.
3517 void moveBefore(Instruction *Inst, Instruction *Before);
3521 /// The ordered list of actions made so far.
3522 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
3523 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
3526 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
3529 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
3532 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
3535 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
3538 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
3540 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
3543 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
3544 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
3547 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
3549 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
3550 Value *Val = Ptr->getBuiltValue();
3551 Actions.push_back(std::move(Ptr));
3555 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
3556 Value *Opnd, Type *Ty) {
3557 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
3558 Value *Val = Ptr->getBuiltValue();
3559 Actions.push_back(std::move(Ptr));
3563 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
3564 Value *Opnd, Type *Ty) {
3565 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
3566 Value *Val = Ptr->getBuiltValue();
3567 Actions.push_back(std::move(Ptr));
3571 void TypePromotionTransaction::moveBefore(Instruction *Inst,
3572 Instruction *Before) {
3574 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
3577 TypePromotionTransaction::ConstRestorationPt
3578 TypePromotionTransaction::getRestorationPoint() const {
3579 return !Actions.empty() ? Actions.back().get() : nullptr;
3582 void TypePromotionTransaction::commit() {
3583 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
3589 void TypePromotionTransaction::rollback(
3590 TypePromotionTransaction::ConstRestorationPt Point) {
3591 while (!Actions.empty() && Point != Actions.back().get()) {
3592 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
3597 /// \brief A helper class for matching addressing modes.
3599 /// This encapsulates the logic for matching the target-legal addressing modes.
3600 class AddressingModeMatcher {
3601 SmallVectorImpl<Instruction*> &AddrModeInsts;
3602 const TargetMachine &TM;
3603 const TargetLowering &TLI;
3604 const DataLayout &DL;
3606 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
3607 /// the memory instruction that we're computing this address for.
3610 Instruction *MemoryInst;
3612 /// This is the addressing mode that we're building up. This is
3613 /// part of the return value of this addressing mode matching stuff.
3614 ExtAddrMode &AddrMode;
3616 /// The instructions inserted by other CodeGenPrepare optimizations.
3617 const SetOfInstrs &InsertedInsts;
3618 /// A map from the instructions to their type before promotion.
3619 InstrToOrigTy &PromotedInsts;
3620 /// The ongoing transaction where every action should be registered.
3621 TypePromotionTransaction &TPT;
3623 /// This is set to true when we should not do profitability checks.
3624 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
3625 bool IgnoreProfitability;
3627 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
3628 const TargetMachine &TM, Type *AT, unsigned AS,
3629 Instruction *MI, ExtAddrMode &AM,
3630 const SetOfInstrs &InsertedInsts,
3631 InstrToOrigTy &PromotedInsts,
3632 TypePromotionTransaction &TPT)
3633 : AddrModeInsts(AMI), TM(TM),
3634 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
3635 ->getTargetLowering()),
3636 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
3637 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
3638 PromotedInsts(PromotedInsts), TPT(TPT) {
3639 IgnoreProfitability = false;
3643 /// Find the maximal addressing mode that a load/store of V can fold,
3644 /// give an access type of AccessTy. This returns a list of involved
3645 /// instructions in AddrModeInsts.
3646 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
3648 /// \p PromotedInsts maps the instructions to their type before promotion.
3649 /// \p The ongoing transaction where every action should be registered.
3650 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
3651 Instruction *MemoryInst,
3652 SmallVectorImpl<Instruction*> &AddrModeInsts,
3653 const TargetMachine &TM,
3654 const SetOfInstrs &InsertedInsts,
3655 InstrToOrigTy &PromotedInsts,
3656 TypePromotionTransaction &TPT) {
3659 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
3660 MemoryInst, Result, InsertedInsts,
3661 PromotedInsts, TPT).matchAddr(V, 0);
3662 (void)Success; assert(Success && "Couldn't select *anything*?");
3666 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
3667 bool matchAddr(Value *V, unsigned Depth);
3668 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
3669 bool *MovedAway = nullptr);
3670 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
3671 ExtAddrMode &AMBefore,
3672 ExtAddrMode &AMAfter);
3673 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
3674 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
3675 Value *PromotedOperand) const;
3678 /// Try adding ScaleReg*Scale to the current addressing mode.
3679 /// Return true and update AddrMode if this addr mode is legal for the target,
3681 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3683 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3684 // mode. Just process that directly.
3686 return matchAddr(ScaleReg, Depth);
3688 // If the scale is 0, it takes nothing to add this.
3692 // If we already have a scale of this value, we can add to it, otherwise, we
3693 // need an available scale field.
3694 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3697 ExtAddrMode TestAddrMode = AddrMode;
3699 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3700 // [A+B + A*7] -> [B+A*8].
3701 TestAddrMode.Scale += Scale;
3702 TestAddrMode.ScaledReg = ScaleReg;
3704 // If the new address isn't legal, bail out.
3705 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3708 // It was legal, so commit it.
3709 AddrMode = TestAddrMode;
3711 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3712 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3713 // X*Scale + C*Scale to addr mode.
3714 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3715 if (isa<Instruction>(ScaleReg) && // not a constant expr.
3716 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3717 TestAddrMode.ScaledReg = AddLHS;
3718 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3720 // If this addressing mode is legal, commit it and remember that we folded
3721 // this instruction.
3722 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3723 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3724 AddrMode = TestAddrMode;
3729 // Otherwise, not (x+c)*scale, just return what we have.
3733 /// This is a little filter, which returns true if an addressing computation
3734 /// involving I might be folded into a load/store accessing it.
3735 /// This doesn't need to be perfect, but needs to accept at least
3736 /// the set of instructions that MatchOperationAddr can.
3737 static bool MightBeFoldableInst(Instruction *I) {
3738 switch (I->getOpcode()) {
3739 case Instruction::BitCast:
3740 case Instruction::AddrSpaceCast:
3741 // Don't touch identity bitcasts.
3742 if (I->getType() == I->getOperand(0)->getType())
3744 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
3745 case Instruction::PtrToInt:
3746 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3748 case Instruction::IntToPtr:
3749 // We know the input is intptr_t, so this is foldable.
3751 case Instruction::Add:
3753 case Instruction::Mul:
3754 case Instruction::Shl:
3755 // Can only handle X*C and X << C.
3756 return isa<ConstantInt>(I->getOperand(1));
3757 case Instruction::GetElementPtr:
3764 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
3765 /// \note \p Val is assumed to be the product of some type promotion.
3766 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3767 /// to be legal, as the non-promoted value would have had the same state.
3768 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
3769 const DataLayout &DL, Value *Val) {
3770 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3773 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3774 // If the ISDOpcode is undefined, it was undefined before the promotion.
3777 // Otherwise, check if the promoted instruction is legal or not.
3778 return TLI.isOperationLegalOrCustom(
3779 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3782 /// \brief Hepler class to perform type promotion.
3783 class TypePromotionHelper {
3784 /// \brief Utility function to check whether or not a sign or zero extension
3785 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3786 /// either using the operands of \p Inst or promoting \p Inst.
3787 /// The type of the extension is defined by \p IsSExt.
3788 /// In other words, check if:
3789 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3790 /// #1 Promotion applies:
3791 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3792 /// #2 Operand reuses:
3793 /// ext opnd1 to ConsideredExtType.
3794 /// \p PromotedInsts maps the instructions to their type before promotion.
3795 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3796 const InstrToOrigTy &PromotedInsts, bool IsSExt);
3798 /// \brief Utility function to determine if \p OpIdx should be promoted when
3799 /// promoting \p Inst.
3800 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3801 return !(isa<SelectInst>(Inst) && OpIdx == 0);
3804 /// \brief Utility function to promote the operand of \p Ext when this
3805 /// operand is a promotable trunc or sext or zext.
3806 /// \p PromotedInsts maps the instructions to their type before promotion.
3807 /// \p CreatedInstsCost[out] contains the cost of all instructions
3808 /// created to promote the operand of Ext.
3809 /// Newly added extensions are inserted in \p Exts.
3810 /// Newly added truncates are inserted in \p Truncs.
3811 /// Should never be called directly.
3812 /// \return The promoted value which is used instead of Ext.
3813 static Value *promoteOperandForTruncAndAnyExt(
3814 Instruction *Ext, TypePromotionTransaction &TPT,
3815 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3816 SmallVectorImpl<Instruction *> *Exts,
3817 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3819 /// \brief Utility function to promote the operand of \p Ext when this
3820 /// operand is promotable and is not a supported trunc or sext.
3821 /// \p PromotedInsts maps the instructions to their type before promotion.
3822 /// \p CreatedInstsCost[out] contains the cost of all the instructions
3823 /// created to promote the operand of Ext.
3824 /// Newly added extensions are inserted in \p Exts.
3825 /// Newly added truncates are inserted in \p Truncs.
3826 /// Should never be called directly.
3827 /// \return The promoted value which is used instead of Ext.
3828 static Value *promoteOperandForOther(Instruction *Ext,
3829 TypePromotionTransaction &TPT,
3830 InstrToOrigTy &PromotedInsts,
3831 unsigned &CreatedInstsCost,
3832 SmallVectorImpl<Instruction *> *Exts,
3833 SmallVectorImpl<Instruction *> *Truncs,
3834 const TargetLowering &TLI, bool IsSExt);
3836 /// \see promoteOperandForOther.
3837 static Value *signExtendOperandForOther(
3838 Instruction *Ext, TypePromotionTransaction &TPT,
3839 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3840 SmallVectorImpl<Instruction *> *Exts,
3841 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3842 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3843 Exts, Truncs, TLI, true);
3846 /// \see promoteOperandForOther.
3847 static Value *zeroExtendOperandForOther(
3848 Instruction *Ext, TypePromotionTransaction &TPT,
3849 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3850 SmallVectorImpl<Instruction *> *Exts,
3851 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3852 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3853 Exts, Truncs, TLI, false);
3857 /// Type for the utility function that promotes the operand of Ext.
3858 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
3859 InstrToOrigTy &PromotedInsts,
3860 unsigned &CreatedInstsCost,
3861 SmallVectorImpl<Instruction *> *Exts,
3862 SmallVectorImpl<Instruction *> *Truncs,
3863 const TargetLowering &TLI);
3864 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
3865 /// action to promote the operand of \p Ext instead of using Ext.
3866 /// \return NULL if no promotable action is possible with the current
3868 /// \p InsertedInsts keeps track of all the instructions inserted by the
3869 /// other CodeGenPrepare optimizations. This information is important
3870 /// because we do not want to promote these instructions as CodeGenPrepare
3871 /// will reinsert them later. Thus creating an infinite loop: create/remove.
3872 /// \p PromotedInsts maps the instructions to their type before promotion.
3873 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3874 const TargetLowering &TLI,
3875 const InstrToOrigTy &PromotedInsts);
3878 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3879 Type *ConsideredExtType,
3880 const InstrToOrigTy &PromotedInsts,
3882 // The promotion helper does not know how to deal with vector types yet.
3883 // To be able to fix that, we would need to fix the places where we
3884 // statically extend, e.g., constants and such.
3885 if (Inst->getType()->isVectorTy())
3888 // We can always get through zext.
3889 if (isa<ZExtInst>(Inst))
3892 // sext(sext) is ok too.
3893 if (IsSExt && isa<SExtInst>(Inst))
3896 // We can get through binary operator, if it is legal. In other words, the
3897 // binary operator must have a nuw or nsw flag.
3898 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3899 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3900 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3901 (IsSExt && BinOp->hasNoSignedWrap())))
3904 // Check if we can do the following simplification.
3905 // ext(trunc(opnd)) --> ext(opnd)
3906 if (!isa<TruncInst>(Inst))
3909 Value *OpndVal = Inst->getOperand(0);
3910 // Check if we can use this operand in the extension.
3911 // If the type is larger than the result type of the extension, we cannot.
3912 if (!OpndVal->getType()->isIntegerTy() ||
3913 OpndVal->getType()->getIntegerBitWidth() >
3914 ConsideredExtType->getIntegerBitWidth())
3917 // If the operand of the truncate is not an instruction, we will not have
3918 // any information on the dropped bits.
3919 // (Actually we could for constant but it is not worth the extra logic).
3920 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3924 // Check if the source of the type is narrow enough.
3925 // I.e., check that trunc just drops extended bits of the same kind of
3927 // #1 get the type of the operand and check the kind of the extended bits.
3928 const Type *OpndType;
3929 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3930 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
3931 OpndType = It->second.getPointer();
3932 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3933 OpndType = Opnd->getOperand(0)->getType();
3937 // #2 check that the truncate just drops extended bits.
3938 return Inst->getType()->getIntegerBitWidth() >=
3939 OpndType->getIntegerBitWidth();
3942 TypePromotionHelper::Action TypePromotionHelper::getAction(
3943 Instruction *Ext, const SetOfInstrs &InsertedInsts,
3944 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3945 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
3946 "Unexpected instruction type");
3947 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3948 Type *ExtTy = Ext->getType();
3949 bool IsSExt = isa<SExtInst>(Ext);
3950 // If the operand of the extension is not an instruction, we cannot
3952 // If it, check we can get through.
3953 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3956 // Do not promote if the operand has been added by codegenprepare.
3957 // Otherwise, it means we are undoing an optimization that is likely to be
3958 // redone, thus causing potential infinite loop.
3959 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3962 // SExt or Trunc instructions.
3963 // Return the related handler.
3964 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3965 isa<ZExtInst>(ExtOpnd))
3966 return promoteOperandForTruncAndAnyExt;
3968 // Regular instruction.
3969 // Abort early if we will have to insert non-free instructions.
3970 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3972 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3975 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3976 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
3977 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3978 SmallVectorImpl<Instruction *> *Exts,
3979 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3980 // By construction, the operand of SExt is an instruction. Otherwise we cannot
3981 // get through it and this method should not be called.
3982 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
3983 Value *ExtVal = SExt;
3984 bool HasMergedNonFreeExt = false;
3985 if (isa<ZExtInst>(SExtOpnd)) {
3986 // Replace s|zext(zext(opnd))
3988 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
3990 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
3991 TPT.replaceAllUsesWith(SExt, ZExt);
3992 TPT.eraseInstruction(SExt);
3995 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
3997 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
3999 CreatedInstsCost = 0;
4001 // Remove dead code.
4002 if (SExtOpnd->use_empty())
4003 TPT.eraseInstruction(SExtOpnd);
4005 // Check if the extension is still needed.
4006 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
4007 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
4010 Exts->push_back(ExtInst);
4011 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
4016 // At this point we have: ext ty opnd to ty.
4017 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
4018 Value *NextVal = ExtInst->getOperand(0);
4019 TPT.eraseInstruction(ExtInst, NextVal);
4023 Value *TypePromotionHelper::promoteOperandForOther(
4024 Instruction *Ext, TypePromotionTransaction &TPT,
4025 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4026 SmallVectorImpl<Instruction *> *Exts,
4027 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
4029 // By construction, the operand of Ext is an instruction. Otherwise we cannot
4030 // get through it and this method should not be called.
4031 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
4032 CreatedInstsCost = 0;
4033 if (!ExtOpnd->hasOneUse()) {
4034 // ExtOpnd will be promoted.
4035 // All its uses, but Ext, will need to use a truncated value of the
4036 // promoted version.
4037 // Create the truncate now.
4038 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
4039 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
4040 ITrunc->removeFromParent();
4041 // Insert it just after the definition.
4042 ITrunc->insertAfter(ExtOpnd);
4044 Truncs->push_back(ITrunc);
4047 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
4048 // Restore the operand of Ext (which has been replaced by the previous call
4049 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
4050 TPT.setOperand(Ext, 0, ExtOpnd);
4053 // Get through the Instruction:
4054 // 1. Update its type.
4055 // 2. Replace the uses of Ext by Inst.
4056 // 3. Extend each operand that needs to be extended.
4058 // Remember the original type of the instruction before promotion.
4059 // This is useful to know that the high bits are sign extended bits.
4060 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
4061 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
4063 TPT.mutateType(ExtOpnd, Ext->getType());
4065 TPT.replaceAllUsesWith(Ext, ExtOpnd);
4067 Instruction *ExtForOpnd = Ext;
4069 DEBUG(dbgs() << "Propagate Ext to operands\n");
4070 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
4072 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
4073 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
4074 !shouldExtOperand(ExtOpnd, OpIdx)) {
4075 DEBUG(dbgs() << "No need to propagate\n");
4078 // Check if we can statically extend the operand.
4079 Value *Opnd = ExtOpnd->getOperand(OpIdx);
4080 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
4081 DEBUG(dbgs() << "Statically extend\n");
4082 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
4083 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
4084 : Cst->getValue().zext(BitWidth);
4085 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
4088 // UndefValue are typed, so we have to statically sign extend them.
4089 if (isa<UndefValue>(Opnd)) {
4090 DEBUG(dbgs() << "Statically extend\n");
4091 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
4095 // Otherwise we have to explicity sign extend the operand.
4096 // Check if Ext was reused to extend an operand.
4098 // If yes, create a new one.
4099 DEBUG(dbgs() << "More operands to ext\n");
4100 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
4101 : TPT.createZExt(Ext, Opnd, Ext->getType());
4102 if (!isa<Instruction>(ValForExtOpnd)) {
4103 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
4106 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
4109 Exts->push_back(ExtForOpnd);
4110 TPT.setOperand(ExtForOpnd, 0, Opnd);
4112 // Move the sign extension before the insertion point.
4113 TPT.moveBefore(ExtForOpnd, ExtOpnd);
4114 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
4115 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
4116 // If more sext are required, new instructions will have to be created.
4117 ExtForOpnd = nullptr;
4119 if (ExtForOpnd == Ext) {
4120 DEBUG(dbgs() << "Extension is useless now\n");
4121 TPT.eraseInstruction(Ext);
4126 /// Check whether or not promoting an instruction to a wider type is profitable.
4127 /// \p NewCost gives the cost of extension instructions created by the
4129 /// \p OldCost gives the cost of extension instructions before the promotion
4130 /// plus the number of instructions that have been
4131 /// matched in the addressing mode the promotion.
4132 /// \p PromotedOperand is the value that has been promoted.
4133 /// \return True if the promotion is profitable, false otherwise.
4134 bool AddressingModeMatcher::isPromotionProfitable(
4135 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
4136 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
4137 // The cost of the new extensions is greater than the cost of the
4138 // old extension plus what we folded.
4139 // This is not profitable.
4140 if (NewCost > OldCost)
4142 if (NewCost < OldCost)
4144 // The promotion is neutral but it may help folding the sign extension in
4145 // loads for instance.
4146 // Check that we did not create an illegal instruction.
4147 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
4150 /// Given an instruction or constant expr, see if we can fold the operation
4151 /// into the addressing mode. If so, update the addressing mode and return
4152 /// true, otherwise return false without modifying AddrMode.
4153 /// If \p MovedAway is not NULL, it contains the information of whether or
4154 /// not AddrInst has to be folded into the addressing mode on success.
4155 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
4156 /// because it has been moved away.
4157 /// Thus AddrInst must not be added in the matched instructions.
4158 /// This state can happen when AddrInst is a sext, since it may be moved away.
4159 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
4160 /// not be referenced anymore.
4161 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
4164 // Avoid exponential behavior on extremely deep expression trees.
4165 if (Depth >= 5) return false;
4167 // By default, all matched instructions stay in place.
4172 case Instruction::PtrToInt:
4173 // PtrToInt is always a noop, as we know that the int type is pointer sized.
4174 return matchAddr(AddrInst->getOperand(0), Depth);
4175 case Instruction::IntToPtr: {
4176 auto AS = AddrInst->getType()->getPointerAddressSpace();
4177 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
4178 // This inttoptr is a no-op if the integer type is pointer sized.
4179 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
4180 return matchAddr(AddrInst->getOperand(0), Depth);
4183 case Instruction::BitCast:
4184 // BitCast is always a noop, and we can handle it as long as it is
4185 // int->int or pointer->pointer (we don't want int<->fp or something).
4186 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
4187 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
4188 // Don't touch identity bitcasts. These were probably put here by LSR,
4189 // and we don't want to mess around with them. Assume it knows what it
4191 AddrInst->getOperand(0)->getType() != AddrInst->getType())
4192 return matchAddr(AddrInst->getOperand(0), Depth);
4194 case Instruction::AddrSpaceCast: {
4196 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
4197 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
4198 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
4199 return matchAddr(AddrInst->getOperand(0), Depth);
4202 case Instruction::Add: {
4203 // Check to see if we can merge in the RHS then the LHS. If so, we win.
4204 ExtAddrMode BackupAddrMode = AddrMode;
4205 unsigned OldSize = AddrModeInsts.size();
4206 // Start a transaction at this point.
4207 // The LHS may match but not the RHS.
4208 // Therefore, we need a higher level restoration point to undo partially
4209 // matched operation.
4210 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4211 TPT.getRestorationPoint();
4213 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
4214 matchAddr(AddrInst->getOperand(0), Depth+1))
4217 // Restore the old addr mode info.
4218 AddrMode = BackupAddrMode;
4219 AddrModeInsts.resize(OldSize);
4220 TPT.rollback(LastKnownGood);
4222 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
4223 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
4224 matchAddr(AddrInst->getOperand(1), Depth+1))
4227 // Otherwise we definitely can't merge the ADD in.
4228 AddrMode = BackupAddrMode;
4229 AddrModeInsts.resize(OldSize);
4230 TPT.rollback(LastKnownGood);
4233 //case Instruction::Or:
4234 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
4236 case Instruction::Mul:
4237 case Instruction::Shl: {
4238 // Can only handle X*C and X << C.
4239 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
4242 int64_t Scale = RHS->getSExtValue();
4243 if (Opcode == Instruction::Shl)
4244 Scale = 1LL << Scale;
4246 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
4248 case Instruction::GetElementPtr: {
4249 // Scan the GEP. We check it if it contains constant offsets and at most
4250 // one variable offset.
4251 int VariableOperand = -1;
4252 unsigned VariableScale = 0;
4254 int64_t ConstantOffset = 0;
4255 gep_type_iterator GTI = gep_type_begin(AddrInst);
4256 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
4257 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
4258 const StructLayout *SL = DL.getStructLayout(STy);
4260 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
4261 ConstantOffset += SL->getElementOffset(Idx);
4263 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
4264 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
4265 ConstantOffset += CI->getSExtValue()*TypeSize;
4266 } else if (TypeSize) { // Scales of zero don't do anything.
4267 // We only allow one variable index at the moment.
4268 if (VariableOperand != -1)
4271 // Remember the variable index.
4272 VariableOperand = i;
4273 VariableScale = TypeSize;
4278 // A common case is for the GEP to only do a constant offset. In this case,
4279 // just add it to the disp field and check validity.
4280 if (VariableOperand == -1) {
4281 AddrMode.BaseOffs += ConstantOffset;
4282 if (ConstantOffset == 0 ||
4283 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
4284 // Check to see if we can fold the base pointer in too.
4285 if (matchAddr(AddrInst->getOperand(0), Depth+1))
4288 AddrMode.BaseOffs -= ConstantOffset;
4292 // Save the valid addressing mode in case we can't match.
4293 ExtAddrMode BackupAddrMode = AddrMode;
4294 unsigned OldSize = AddrModeInsts.size();
4296 // See if the scale and offset amount is valid for this target.
4297 AddrMode.BaseOffs += ConstantOffset;
4299 // Match the base operand of the GEP.
4300 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
4301 // If it couldn't be matched, just stuff the value in a register.
4302 if (AddrMode.HasBaseReg) {
4303 AddrMode = BackupAddrMode;
4304 AddrModeInsts.resize(OldSize);
4307 AddrMode.HasBaseReg = true;
4308 AddrMode.BaseReg = AddrInst->getOperand(0);
4311 // Match the remaining variable portion of the GEP.
4312 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
4314 // If it couldn't be matched, try stuffing the base into a register
4315 // instead of matching it, and retrying the match of the scale.
4316 AddrMode = BackupAddrMode;
4317 AddrModeInsts.resize(OldSize);
4318 if (AddrMode.HasBaseReg)
4320 AddrMode.HasBaseReg = true;
4321 AddrMode.BaseReg = AddrInst->getOperand(0);
4322 AddrMode.BaseOffs += ConstantOffset;
4323 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
4324 VariableScale, Depth)) {
4325 // If even that didn't work, bail.
4326 AddrMode = BackupAddrMode;
4327 AddrModeInsts.resize(OldSize);
4334 case Instruction::SExt:
4335 case Instruction::ZExt: {
4336 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
4340 // Try to move this ext out of the way of the addressing mode.
4341 // Ask for a method for doing so.
4342 TypePromotionHelper::Action TPH =
4343 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
4347 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4348 TPT.getRestorationPoint();
4349 unsigned CreatedInstsCost = 0;
4350 unsigned ExtCost = !TLI.isExtFree(Ext);
4351 Value *PromotedOperand =
4352 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
4353 // SExt has been moved away.
4354 // Thus either it will be rematched later in the recursive calls or it is
4355 // gone. Anyway, we must not fold it into the addressing mode at this point.
4359 // addr = gep base, idx
4361 // promotedOpnd = ext opnd <- no match here
4362 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
4363 // addr = gep base, op <- match
4367 assert(PromotedOperand &&
4368 "TypePromotionHelper should have filtered out those cases");
4370 ExtAddrMode BackupAddrMode = AddrMode;
4371 unsigned OldSize = AddrModeInsts.size();
4373 if (!matchAddr(PromotedOperand, Depth) ||
4374 // The total of the new cost is equal to the cost of the created
4376 // The total of the old cost is equal to the cost of the extension plus
4377 // what we have saved in the addressing mode.
4378 !isPromotionProfitable(CreatedInstsCost,
4379 ExtCost + (AddrModeInsts.size() - OldSize),
4381 AddrMode = BackupAddrMode;
4382 AddrModeInsts.resize(OldSize);
4383 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
4384 TPT.rollback(LastKnownGood);
4393 /// If we can, try to add the value of 'Addr' into the current addressing mode.
4394 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
4395 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
4398 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
4399 // Start a transaction at this point that we will rollback if the matching
4401 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4402 TPT.getRestorationPoint();
4403 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
4404 // Fold in immediates if legal for the target.
4405 AddrMode.BaseOffs += CI->getSExtValue();
4406 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4408 AddrMode.BaseOffs -= CI->getSExtValue();
4409 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
4410 // If this is a global variable, try to fold it into the addressing mode.
4411 if (!AddrMode.BaseGV) {
4412 AddrMode.BaseGV = GV;
4413 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4415 AddrMode.BaseGV = nullptr;
4417 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
4418 ExtAddrMode BackupAddrMode = AddrMode;
4419 unsigned OldSize = AddrModeInsts.size();
4421 // Check to see if it is possible to fold this operation.
4422 bool MovedAway = false;
4423 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
4424 // This instruction may have been moved away. If so, there is nothing
4428 // Okay, it's possible to fold this. Check to see if it is actually
4429 // *profitable* to do so. We use a simple cost model to avoid increasing
4430 // register pressure too much.
4431 if (I->hasOneUse() ||
4432 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
4433 AddrModeInsts.push_back(I);
4437 // It isn't profitable to do this, roll back.
4438 //cerr << "NOT FOLDING: " << *I;
4439 AddrMode = BackupAddrMode;
4440 AddrModeInsts.resize(OldSize);
4441 TPT.rollback(LastKnownGood);
4443 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
4444 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
4446 TPT.rollback(LastKnownGood);
4447 } else if (isa<ConstantPointerNull>(Addr)) {
4448 // Null pointer gets folded without affecting the addressing mode.
4452 // Worse case, the target should support [reg] addressing modes. :)
4453 if (!AddrMode.HasBaseReg) {
4454 AddrMode.HasBaseReg = true;
4455 AddrMode.BaseReg = Addr;
4456 // Still check for legality in case the target supports [imm] but not [i+r].
4457 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4459 AddrMode.HasBaseReg = false;
4460 AddrMode.BaseReg = nullptr;
4463 // If the base register is already taken, see if we can do [r+r].
4464 if (AddrMode.Scale == 0) {
4466 AddrMode.ScaledReg = Addr;
4467 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4470 AddrMode.ScaledReg = nullptr;
4473 TPT.rollback(LastKnownGood);
4477 /// Check to see if all uses of OpVal by the specified inline asm call are due
4478 /// to memory operands. If so, return true, otherwise return false.
4479 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
4480 const TargetMachine &TM) {
4481 const Function *F = CI->getParent()->getParent();
4482 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
4483 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
4484 TargetLowering::AsmOperandInfoVector TargetConstraints =
4485 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
4486 ImmutableCallSite(CI));
4487 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4488 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4490 // Compute the constraint code and ConstraintType to use.
4491 TLI->ComputeConstraintToUse(OpInfo, SDValue());
4493 // If this asm operand is our Value*, and if it isn't an indirect memory
4494 // operand, we can't fold it!
4495 if (OpInfo.CallOperandVal == OpVal &&
4496 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
4497 !OpInfo.isIndirect))
4504 /// Recursively walk all the uses of I until we find a memory use.
4505 /// If we find an obviously non-foldable instruction, return true.
4506 /// Add the ultimately found memory instructions to MemoryUses.
4507 static bool FindAllMemoryUses(
4509 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
4510 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
4511 // If we already considered this instruction, we're done.
4512 if (!ConsideredInsts.insert(I).second)
4515 // If this is an obviously unfoldable instruction, bail out.
4516 if (!MightBeFoldableInst(I))
4519 // Loop over all the uses, recursively processing them.
4520 for (Use &U : I->uses()) {
4521 Instruction *UserI = cast<Instruction>(U.getUser());
4523 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4524 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
4528 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4529 unsigned opNo = U.getOperandNo();
4530 if (opNo == 0) return true; // Storing addr, not into addr.
4531 MemoryUses.push_back(std::make_pair(SI, opNo));
4535 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4536 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
4537 if (!IA) return true;
4539 // If this is a memory operand, we're cool, otherwise bail out.
4540 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
4545 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
4552 /// Return true if Val is already known to be live at the use site that we're
4553 /// folding it into. If so, there is no cost to include it in the addressing
4554 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4555 /// instruction already.
4556 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4557 Value *KnownLive2) {
4558 // If Val is either of the known-live values, we know it is live!
4559 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4562 // All values other than instructions and arguments (e.g. constants) are live.
4563 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4565 // If Val is a constant sized alloca in the entry block, it is live, this is
4566 // true because it is just a reference to the stack/frame pointer, which is
4567 // live for the whole function.
4568 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4569 if (AI->isStaticAlloca())
4572 // Check to see if this value is already used in the memory instruction's
4573 // block. If so, it's already live into the block at the very least, so we
4574 // can reasonably fold it.
4575 return Val->isUsedInBasicBlock(MemoryInst->getParent());
4578 /// It is possible for the addressing mode of the machine to fold the specified
4579 /// instruction into a load or store that ultimately uses it.
4580 /// However, the specified instruction has multiple uses.
4581 /// Given this, it may actually increase register pressure to fold it
4582 /// into the load. For example, consider this code:
4586 /// use(Y) -> nonload/store
4590 /// In this case, Y has multiple uses, and can be folded into the load of Z
4591 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4592 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4593 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4594 /// number of computations either.
4596 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4597 /// X was live across 'load Z' for other reasons, we actually *would* want to
4598 /// fold the addressing mode in the Z case. This would make Y die earlier.
4599 bool AddressingModeMatcher::
4600 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4601 ExtAddrMode &AMAfter) {
4602 if (IgnoreProfitability) return true;
4604 // AMBefore is the addressing mode before this instruction was folded into it,
4605 // and AMAfter is the addressing mode after the instruction was folded. Get
4606 // the set of registers referenced by AMAfter and subtract out those
4607 // referenced by AMBefore: this is the set of values which folding in this
4608 // address extends the lifetime of.
4610 // Note that there are only two potential values being referenced here,
4611 // BaseReg and ScaleReg (global addresses are always available, as are any
4612 // folded immediates).
4613 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4615 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4616 // lifetime wasn't extended by adding this instruction.
4617 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4619 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4620 ScaledReg = nullptr;
4622 // If folding this instruction (and it's subexprs) didn't extend any live
4623 // ranges, we're ok with it.
4624 if (!BaseReg && !ScaledReg)
4627 // If all uses of this instruction are ultimately load/store/inlineasm's,
4628 // check to see if their addressing modes will include this instruction. If
4629 // so, we can fold it into all uses, so it doesn't matter if it has multiple
4631 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
4632 SmallPtrSet<Instruction*, 16> ConsideredInsts;
4633 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
4634 return false; // Has a non-memory, non-foldable use!
4636 // Now that we know that all uses of this instruction are part of a chain of
4637 // computation involving only operations that could theoretically be folded
4638 // into a memory use, loop over each of these uses and see if they could
4639 // *actually* fold the instruction.
4640 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4641 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4642 Instruction *User = MemoryUses[i].first;
4643 unsigned OpNo = MemoryUses[i].second;
4645 // Get the access type of this use. If the use isn't a pointer, we don't
4646 // know what it accesses.
4647 Value *Address = User->getOperand(OpNo);
4648 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4651 Type *AddressAccessTy = AddrTy->getElementType();
4652 unsigned AS = AddrTy->getAddressSpace();
4654 // Do a match against the root of this address, ignoring profitability. This
4655 // will tell us if the addressing mode for the memory operation will
4656 // *actually* cover the shared instruction.
4658 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4659 TPT.getRestorationPoint();
4660 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
4661 MemoryInst, Result, InsertedInsts,
4662 PromotedInsts, TPT);
4663 Matcher.IgnoreProfitability = true;
4664 bool Success = Matcher.matchAddr(Address, 0);
4665 (void)Success; assert(Success && "Couldn't select *anything*?");
4667 // The match was to check the profitability, the changes made are not
4668 // part of the original matcher. Therefore, they should be dropped
4669 // otherwise the original matcher will not present the right state.
4670 TPT.rollback(LastKnownGood);
4672 // If the match didn't cover I, then it won't be shared by it.
4673 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
4674 I) == MatchedAddrModeInsts.end())
4677 MatchedAddrModeInsts.clear();
4683 } // end anonymous namespace
4685 /// Return true if the specified values are defined in a
4686 /// different basic block than BB.
4687 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4688 if (Instruction *I = dyn_cast<Instruction>(V))
4689 return I->getParent() != BB;
4693 /// Load and Store Instructions often have addressing modes that can do
4694 /// significant amounts of computation. As such, instruction selection will try
4695 /// to get the load or store to do as much computation as possible for the
4696 /// program. The problem is that isel can only see within a single block. As
4697 /// such, we sink as much legal addressing mode work into the block as possible.
4699 /// This method is used to optimize both load/store and inline asms with memory
4701 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4702 Type *AccessTy, unsigned AddrSpace) {
4705 // Try to collapse single-value PHI nodes. This is necessary to undo
4706 // unprofitable PRE transformations.
4707 SmallVector<Value*, 8> worklist;
4708 SmallPtrSet<Value*, 16> Visited;
4709 worklist.push_back(Addr);
4711 // Use a worklist to iteratively look through PHI nodes, and ensure that
4712 // the addressing mode obtained from the non-PHI roots of the graph
4714 Value *Consensus = nullptr;
4715 unsigned NumUsesConsensus = 0;
4716 bool IsNumUsesConsensusValid = false;
4717 SmallVector<Instruction*, 16> AddrModeInsts;
4718 ExtAddrMode AddrMode;
4719 TypePromotionTransaction TPT;
4720 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4721 TPT.getRestorationPoint();
4722 while (!worklist.empty()) {
4723 Value *V = worklist.back();
4724 worklist.pop_back();
4726 // Break use-def graph loops.
4727 if (!Visited.insert(V).second) {
4728 Consensus = nullptr;
4732 // For a PHI node, push all of its incoming values.
4733 if (PHINode *P = dyn_cast<PHINode>(V)) {
4734 for (Value *IncValue : P->incoming_values())
4735 worklist.push_back(IncValue);
4739 // For non-PHIs, determine the addressing mode being computed.
4740 SmallVector<Instruction*, 16> NewAddrModeInsts;
4741 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4742 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
4743 InsertedInsts, PromotedInsts, TPT);
4745 // This check is broken into two cases with very similar code to avoid using
4746 // getNumUses() as much as possible. Some values have a lot of uses, so
4747 // calling getNumUses() unconditionally caused a significant compile-time
4751 AddrMode = NewAddrMode;
4752 AddrModeInsts = NewAddrModeInsts;
4754 } else if (NewAddrMode == AddrMode) {
4755 if (!IsNumUsesConsensusValid) {
4756 NumUsesConsensus = Consensus->getNumUses();
4757 IsNumUsesConsensusValid = true;
4760 // Ensure that the obtained addressing mode is equivalent to that obtained
4761 // for all other roots of the PHI traversal. Also, when choosing one
4762 // such root as representative, select the one with the most uses in order
4763 // to keep the cost modeling heuristics in AddressingModeMatcher
4765 unsigned NumUses = V->getNumUses();
4766 if (NumUses > NumUsesConsensus) {
4768 NumUsesConsensus = NumUses;
4769 AddrModeInsts = NewAddrModeInsts;
4774 Consensus = nullptr;
4778 // If the addressing mode couldn't be determined, or if multiple different
4779 // ones were determined, bail out now.
4781 TPT.rollback(LastKnownGood);
4786 // Check to see if any of the instructions supersumed by this addr mode are
4787 // non-local to I's BB.
4788 bool AnyNonLocal = false;
4789 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
4790 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
4796 // If all the instructions matched are already in this BB, don't do anything.
4798 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
4802 // Insert this computation right after this user. Since our caller is
4803 // scanning from the top of the BB to the bottom, reuse of the expr are
4804 // guaranteed to happen later.
4805 IRBuilder<> Builder(MemoryInst);
4807 // Now that we determined the addressing expression we want to use and know
4808 // that we have to sink it into this block. Check to see if we have already
4809 // done this for some other load/store instr in this block. If so, reuse the
4811 Value *&SunkAddr = SunkAddrs[Addr];
4813 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
4814 << *MemoryInst << "\n");
4815 if (SunkAddr->getType() != Addr->getType())
4816 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
4817 } else if (AddrSinkUsingGEPs ||
4818 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
4819 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
4821 // By default, we use the GEP-based method when AA is used later. This
4822 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4823 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4824 << *MemoryInst << "\n");
4825 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4826 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4828 // First, find the pointer.
4829 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4830 ResultPtr = AddrMode.BaseReg;
4831 AddrMode.BaseReg = nullptr;
4834 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4835 // We can't add more than one pointer together, nor can we scale a
4836 // pointer (both of which seem meaningless).
4837 if (ResultPtr || AddrMode.Scale != 1)
4840 ResultPtr = AddrMode.ScaledReg;
4844 if (AddrMode.BaseGV) {
4848 ResultPtr = AddrMode.BaseGV;
4851 // If the real base value actually came from an inttoptr, then the matcher
4852 // will look through it and provide only the integer value. In that case,
4854 if (!ResultPtr && AddrMode.BaseReg) {
4856 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
4857 AddrMode.BaseReg = nullptr;
4858 } else if (!ResultPtr && AddrMode.Scale == 1) {
4860 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
4865 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4866 SunkAddr = Constant::getNullValue(Addr->getType());
4867 } else if (!ResultPtr) {
4871 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4872 Type *I8Ty = Builder.getInt8Ty();
4874 // Start with the base register. Do this first so that subsequent address
4875 // matching finds it last, which will prevent it from trying to match it
4876 // as the scaled value in case it happens to be a mul. That would be
4877 // problematic if we've sunk a different mul for the scale, because then
4878 // we'd end up sinking both muls.
4879 if (AddrMode.BaseReg) {
4880 Value *V = AddrMode.BaseReg;
4881 if (V->getType() != IntPtrTy)
4882 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4887 // Add the scale value.
4888 if (AddrMode.Scale) {
4889 Value *V = AddrMode.ScaledReg;
4890 if (V->getType() == IntPtrTy) {
4892 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4893 cast<IntegerType>(V->getType())->getBitWidth()) {
4894 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4896 // It is only safe to sign extend the BaseReg if we know that the math
4897 // required to create it did not overflow before we extend it. Since
4898 // the original IR value was tossed in favor of a constant back when
4899 // the AddrMode was created we need to bail out gracefully if widths
4900 // do not match instead of extending it.
4901 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
4902 if (I && (ResultIndex != AddrMode.BaseReg))
4903 I->eraseFromParent();
4907 if (AddrMode.Scale != 1)
4908 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4911 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4916 // Add in the Base Offset if present.
4917 if (AddrMode.BaseOffs) {
4918 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4920 // We need to add this separately from the scale above to help with
4921 // SDAG consecutive load/store merging.
4922 if (ResultPtr->getType() != I8PtrTy)
4923 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
4924 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4931 SunkAddr = ResultPtr;
4933 if (ResultPtr->getType() != I8PtrTy)
4934 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
4935 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4938 if (SunkAddr->getType() != Addr->getType())
4939 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
4942 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4943 << *MemoryInst << "\n");
4944 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4945 Value *Result = nullptr;
4947 // Start with the base register. Do this first so that subsequent address
4948 // matching finds it last, which will prevent it from trying to match it
4949 // as the scaled value in case it happens to be a mul. That would be
4950 // problematic if we've sunk a different mul for the scale, because then
4951 // we'd end up sinking both muls.
4952 if (AddrMode.BaseReg) {
4953 Value *V = AddrMode.BaseReg;
4954 if (V->getType()->isPointerTy())
4955 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4956 if (V->getType() != IntPtrTy)
4957 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4961 // Add the scale value.
4962 if (AddrMode.Scale) {
4963 Value *V = AddrMode.ScaledReg;
4964 if (V->getType() == IntPtrTy) {
4966 } else if (V->getType()->isPointerTy()) {
4967 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4968 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4969 cast<IntegerType>(V->getType())->getBitWidth()) {
4970 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4972 // It is only safe to sign extend the BaseReg if we know that the math
4973 // required to create it did not overflow before we extend it. Since
4974 // the original IR value was tossed in favor of a constant back when
4975 // the AddrMode was created we need to bail out gracefully if widths
4976 // do not match instead of extending it.
4977 Instruction *I = dyn_cast_or_null<Instruction>(Result);
4978 if (I && (Result != AddrMode.BaseReg))
4979 I->eraseFromParent();
4982 if (AddrMode.Scale != 1)
4983 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4986 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4991 // Add in the BaseGV if present.
4992 if (AddrMode.BaseGV) {
4993 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
4995 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5000 // Add in the Base Offset if present.
5001 if (AddrMode.BaseOffs) {
5002 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
5004 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5010 SunkAddr = Constant::getNullValue(Addr->getType());
5012 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
5015 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
5017 // If we have no uses, recursively delete the value and all dead instructions
5019 if (Repl->use_empty()) {
5020 // This can cause recursive deletion, which can invalidate our iterator.
5021 // Use a WeakVH to hold onto it in case this happens.
5022 WeakVH IterHandle(&*CurInstIterator);
5023 BasicBlock *BB = CurInstIterator->getParent();
5025 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
5027 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
5028 // If the iterator instruction was recursively deleted, start over at the
5029 // start of the block.
5030 CurInstIterator = BB->begin();
5038 /// If there are any memory operands, use OptimizeMemoryInst to sink their
5039 /// address computing into the block when possible / profitable.
5040 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
5041 bool MadeChange = false;
5043 const TargetRegisterInfo *TRI =
5044 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
5045 TargetLowering::AsmOperandInfoVector TargetConstraints =
5046 TLI->ParseConstraints(*DL, TRI, CS);
5048 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5049 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
5051 // Compute the constraint code and ConstraintType to use.
5052 TLI->ComputeConstraintToUse(OpInfo, SDValue());
5054 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5055 OpInfo.isIndirect) {
5056 Value *OpVal = CS->getArgOperand(ArgNo++);
5057 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
5058 } else if (OpInfo.Type == InlineAsm::isInput)
5065 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
5066 /// sign extensions.
5067 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
5068 assert(!Inst->use_empty() && "Input must have at least one use");
5069 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
5070 bool IsSExt = isa<SExtInst>(FirstUser);
5071 Type *ExtTy = FirstUser->getType();
5072 for (const User *U : Inst->users()) {
5073 const Instruction *UI = cast<Instruction>(U);
5074 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
5076 Type *CurTy = UI->getType();
5077 // Same input and output types: Same instruction after CSE.
5081 // If IsSExt is true, we are in this situation:
5083 // b = sext ty1 a to ty2
5084 // c = sext ty1 a to ty3
5085 // Assuming ty2 is shorter than ty3, this could be turned into:
5087 // b = sext ty1 a to ty2
5088 // c = sext ty2 b to ty3
5089 // However, the last sext is not free.
5093 // This is a ZExt, maybe this is free to extend from one type to another.
5094 // In that case, we would not account for a different use.
5097 if (ExtTy->getScalarType()->getIntegerBitWidth() >
5098 CurTy->getScalarType()->getIntegerBitWidth()) {
5106 if (!TLI.isZExtFree(NarrowTy, LargeTy))
5109 // All uses are the same or can be derived from one another for free.
5113 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
5114 /// load instruction.
5115 /// If an ext(load) can be formed, it is returned via \p LI for the load
5116 /// and \p Inst for the extension.
5117 /// Otherwise LI == nullptr and Inst == nullptr.
5118 /// When some promotion happened, \p TPT contains the proper state to
5121 /// \return true when promoting was necessary to expose the ext(load)
5122 /// opportunity, false otherwise.
5126 /// %ld = load i32* %addr
5127 /// %add = add nuw i32 %ld, 4
5128 /// %zext = zext i32 %add to i64
5132 /// %ld = load i32* %addr
5133 /// %zext = zext i32 %ld to i64
5134 /// %add = add nuw i64 %zext, 4
5136 /// Thanks to the promotion, we can match zext(load i32*) to i64.
5137 bool CodeGenPrepare::extLdPromotion(TypePromotionTransaction &TPT,
5138 LoadInst *&LI, Instruction *&Inst,
5139 const SmallVectorImpl<Instruction *> &Exts,
5140 unsigned CreatedInstsCost = 0) {
5141 // Iterate over all the extensions to see if one form an ext(load).
5142 for (auto I : Exts) {
5143 // Check if we directly have ext(load).
5144 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
5146 // No promotion happened here.
5149 // Check whether or not we want to do any promotion.
5150 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
5152 // Get the action to perform the promotion.
5153 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
5154 I, InsertedInsts, *TLI, PromotedInsts);
5155 // Check if we can promote.
5158 // Save the current state.
5159 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5160 TPT.getRestorationPoint();
5161 SmallVector<Instruction *, 4> NewExts;
5162 unsigned NewCreatedInstsCost = 0;
5163 unsigned ExtCost = !TLI->isExtFree(I);
5165 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
5166 &NewExts, nullptr, *TLI);
5167 assert(PromotedVal &&
5168 "TypePromotionHelper should have filtered out those cases");
5170 // We would be able to merge only one extension in a load.
5171 // Therefore, if we have more than 1 new extension we heuristically
5172 // cut this search path, because it means we degrade the code quality.
5173 // With exactly 2, the transformation is neutral, because we will merge
5174 // one extension but leave one. However, we optimistically keep going,
5175 // because the new extension may be removed too.
5176 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
5177 TotalCreatedInstsCost -= ExtCost;
5178 if (!StressExtLdPromotion &&
5179 (TotalCreatedInstsCost > 1 ||
5180 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
5181 // The promotion is not profitable, rollback to the previous state.
5182 TPT.rollback(LastKnownGood);
5185 // The promotion is profitable.
5186 // Check if it exposes an ext(load).
5187 (void)extLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
5188 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
5189 // If we have created a new extension, i.e., now we have two
5190 // extensions. We must make sure one of them is merged with
5191 // the load, otherwise we may degrade the code quality.
5192 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
5193 // Promotion happened.
5195 // If this does not help to expose an ext(load) then, rollback.
5196 TPT.rollback(LastKnownGood);
5198 // None of the extension can form an ext(load).
5204 /// Move a zext or sext fed by a load into the same basic block as the load,
5205 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
5206 /// extend into the load.
5207 /// \p I[in/out] the extension may be modified during the process if some
5208 /// promotions apply.
5210 bool CodeGenPrepare::moveExtToFormExtLoad(Instruction *&I) {
5211 // Try to promote a chain of computation if it allows to form
5212 // an extended load.
5213 TypePromotionTransaction TPT;
5214 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5215 TPT.getRestorationPoint();
5216 SmallVector<Instruction *, 1> Exts;
5218 // Look for a load being extended.
5219 LoadInst *LI = nullptr;
5220 Instruction *OldExt = I;
5221 bool HasPromoted = extLdPromotion(TPT, LI, I, Exts);
5223 assert(!HasPromoted && !LI && "If we did not match any load instruction "
5224 "the code must remain the same");
5229 // If they're already in the same block, there's nothing to do.
5230 // Make the cheap checks first if we did not promote.
5231 // If we promoted, we need to check if it is indeed profitable.
5232 if (!HasPromoted && LI->getParent() == I->getParent())
5235 EVT VT = TLI->getValueType(*DL, I->getType());
5236 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
5238 // If the load has other users and the truncate is not free, this probably
5239 // isn't worthwhile.
5240 if (!LI->hasOneUse() && TLI &&
5241 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
5242 !TLI->isTruncateFree(I->getType(), LI->getType())) {
5244 TPT.rollback(LastKnownGood);
5248 // Check whether the target supports casts folded into loads.
5250 if (isa<ZExtInst>(I))
5251 LType = ISD::ZEXTLOAD;
5253 assert(isa<SExtInst>(I) && "Unexpected ext type!");
5254 LType = ISD::SEXTLOAD;
5256 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
5258 TPT.rollback(LastKnownGood);
5262 // Move the extend into the same block as the load, so that SelectionDAG
5265 I->removeFromParent();
5271 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
5272 BasicBlock *DefBB = I->getParent();
5274 // If the result of a {s|z}ext and its source are both live out, rewrite all
5275 // other uses of the source with result of extension.
5276 Value *Src = I->getOperand(0);
5277 if (Src->hasOneUse())
5280 // Only do this xform if truncating is free.
5281 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
5284 // Only safe to perform the optimization if the source is also defined in
5286 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
5289 bool DefIsLiveOut = false;
5290 for (User *U : I->users()) {
5291 Instruction *UI = cast<Instruction>(U);
5293 // Figure out which BB this ext is used in.
5294 BasicBlock *UserBB = UI->getParent();
5295 if (UserBB == DefBB) continue;
5296 DefIsLiveOut = true;
5302 // Make sure none of the uses are PHI nodes.
5303 for (User *U : Src->users()) {
5304 Instruction *UI = cast<Instruction>(U);
5305 BasicBlock *UserBB = UI->getParent();
5306 if (UserBB == DefBB) continue;
5307 // Be conservative. We don't want this xform to end up introducing
5308 // reloads just before load / store instructions.
5309 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
5313 // InsertedTruncs - Only insert one trunc in each block once.
5314 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
5316 bool MadeChange = false;
5317 for (Use &U : Src->uses()) {
5318 Instruction *User = cast<Instruction>(U.getUser());
5320 // Figure out which BB this ext is used in.
5321 BasicBlock *UserBB = User->getParent();
5322 if (UserBB == DefBB) continue;
5324 // Both src and def are live in this block. Rewrite the use.
5325 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
5327 if (!InsertedTrunc) {
5328 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5329 assert(InsertPt != UserBB->end());
5330 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5331 InsertedInsts.insert(InsertedTrunc);
5334 // Replace a use of the {s|z}ext source with a use of the result.
5343 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
5344 // just after the load if the target can fold this into one extload instruction,
5345 // with the hope of eliminating some of the other later "and" instructions using
5346 // the loaded value. "and"s that are made trivially redundant by the insertion
5347 // of the new "and" are removed by this function, while others (e.g. those whose
5348 // path from the load goes through a phi) are left for isel to potentially
5381 // becomes (after a call to optimizeLoadExt for each load):
5385 // x1' = and x1, 0xff
5389 // x2' = and x2, 0xff
5396 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5398 if (!Load->isSimple() ||
5399 !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
5402 // Skip loads we've already transformed or have no reason to transform.
5403 if (Load->hasOneUse()) {
5404 User *LoadUser = *Load->user_begin();
5405 if (cast<Instruction>(LoadUser)->getParent() == Load->getParent() &&
5406 !dyn_cast<PHINode>(LoadUser))
5410 // Look at all uses of Load, looking through phis, to determine how many bits
5411 // of the loaded value are needed.
5412 SmallVector<Instruction *, 8> WorkList;
5413 SmallPtrSet<Instruction *, 16> Visited;
5414 SmallVector<Instruction *, 8> AndsToMaybeRemove;
5415 for (auto *U : Load->users())
5416 WorkList.push_back(cast<Instruction>(U));
5418 EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5419 unsigned BitWidth = LoadResultVT.getSizeInBits();
5420 APInt DemandBits(BitWidth, 0);
5421 APInt WidestAndBits(BitWidth, 0);
5423 while (!WorkList.empty()) {
5424 Instruction *I = WorkList.back();
5425 WorkList.pop_back();
5427 // Break use-def graph loops.
5428 if (!Visited.insert(I).second)
5431 // For a PHI node, push all of its users.
5432 if (auto *Phi = dyn_cast<PHINode>(I)) {
5433 for (auto *U : Phi->users())
5434 WorkList.push_back(cast<Instruction>(U));
5438 switch (I->getOpcode()) {
5439 case llvm::Instruction::And: {
5440 auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5443 APInt AndBits = AndC->getValue();
5444 DemandBits |= AndBits;
5445 // Keep track of the widest and mask we see.
5446 if (AndBits.ugt(WidestAndBits))
5447 WidestAndBits = AndBits;
5448 if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5449 AndsToMaybeRemove.push_back(I);
5453 case llvm::Instruction::Shl: {
5454 auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5457 uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5458 auto ShlDemandBits = APInt::getAllOnesValue(BitWidth).lshr(ShiftAmt);
5459 DemandBits |= ShlDemandBits;
5463 case llvm::Instruction::Trunc: {
5464 EVT TruncVT = TLI->getValueType(*DL, I->getType());
5465 unsigned TruncBitWidth = TruncVT.getSizeInBits();
5466 auto TruncBits = APInt::getAllOnesValue(TruncBitWidth).zext(BitWidth);
5467 DemandBits |= TruncBits;
5476 uint32_t ActiveBits = DemandBits.getActiveBits();
5477 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5478 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5479 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5480 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5481 // followed by an AND.
5482 // TODO: Look into removing this restriction by fixing backends to either
5483 // return false for isLoadExtLegal for i1 or have them select this pattern to
5484 // a single instruction.
5486 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5487 // mask, since these are the only ands that will be removed by isel.
5488 if (ActiveBits <= 1 || !APIntOps::isMask(ActiveBits, DemandBits) ||
5489 WidestAndBits != DemandBits)
5492 LLVMContext &Ctx = Load->getType()->getContext();
5493 Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5494 EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5496 // Reject cases that won't be matched as extloads.
5497 if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5498 !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5501 IRBuilder<> Builder(Load->getNextNode());
5502 auto *NewAnd = dyn_cast<Instruction>(
5503 Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5505 // Replace all uses of load with new and (except for the use of load in the
5507 Load->replaceAllUsesWith(NewAnd);
5508 NewAnd->setOperand(0, Load);
5510 // Remove any and instructions that are now redundant.
5511 for (auto *And : AndsToMaybeRemove)
5512 // Check that the and mask is the same as the one we decided to put on the
5514 if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5515 And->replaceAllUsesWith(NewAnd);
5516 if (&*CurInstIterator == And)
5517 CurInstIterator = std::next(And->getIterator());
5518 And->eraseFromParent();
5526 /// Check if V (an operand of a select instruction) is an expensive instruction
5527 /// that is only used once.
5528 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5529 auto *I = dyn_cast<Instruction>(V);
5530 // If it's safe to speculatively execute, then it should not have side
5531 // effects; therefore, it's safe to sink and possibly *not* execute.
5532 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5533 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
5536 /// Returns true if a SelectInst should be turned into an explicit branch.
5537 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
5539 // FIXME: This should use the same heuristics as IfConversion to determine
5540 // whether a select is better represented as a branch. This requires that
5541 // branch probability metadata is preserved for the select, which is not the
5544 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5546 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5547 // comparison condition. If the compare has more than one use, there's
5548 // probably another cmov or setcc around, so it's not worth emitting a branch.
5549 if (!Cmp || !Cmp->hasOneUse())
5552 Value *CmpOp0 = Cmp->getOperand(0);
5553 Value *CmpOp1 = Cmp->getOperand(1);
5555 // Emit "cmov on compare with a memory operand" as a branch to avoid stalls
5556 // on a load from memory. But if the load is used more than once, do not
5557 // change the select to a branch because the load is probably needed
5558 // regardless of whether the branch is taken or not.
5559 if ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
5560 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()))
5563 // If either operand of the select is expensive and only needed on one side
5564 // of the select, we should form a branch.
5565 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5566 sinkSelectOperand(TTI, SI->getFalseValue()))
5573 /// If we have a SelectInst that will likely profit from branch prediction,
5574 /// turn it into a branch.
5575 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5576 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
5578 // Can we convert the 'select' to CF ?
5579 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
5582 TargetLowering::SelectSupportKind SelectKind;
5584 SelectKind = TargetLowering::VectorMaskSelect;
5585 else if (SI->getType()->isVectorTy())
5586 SelectKind = TargetLowering::ScalarCondVectorVal;
5588 SelectKind = TargetLowering::ScalarValSelect;
5590 // Do we have efficient codegen support for this kind of 'selects' ?
5591 if (TLI->isSelectSupported(SelectKind)) {
5592 // We have efficient codegen support for the select instruction.
5593 // Check if it is profitable to keep this 'select'.
5594 if (!TLI->isPredictableSelectExpensive() ||
5595 !isFormingBranchFromSelectProfitable(TTI, SI))
5601 // Transform a sequence like this:
5603 // %cmp = cmp uge i32 %a, %b
5604 // %sel = select i1 %cmp, i32 %c, i32 %d
5608 // %cmp = cmp uge i32 %a, %b
5609 // br i1 %cmp, label %select.true, label %select.false
5611 // br label %select.end
5613 // br label %select.end
5615 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
5617 // In addition, we may sink instructions that produce %c or %d from
5618 // the entry block into the destination(s) of the new branch.
5619 // If the true or false blocks do not contain a sunken instruction, that
5620 // block and its branch may be optimized away. In that case, one side of the
5621 // first branch will point directly to select.end, and the corresponding PHI
5622 // predecessor block will be the start block.
5624 // First, we split the block containing the select into 2 blocks.
5625 BasicBlock *StartBlock = SI->getParent();
5626 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
5627 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
5629 // Delete the unconditional branch that was just created by the split.
5630 StartBlock->getTerminator()->eraseFromParent();
5632 // These are the new basic blocks for the conditional branch.
5633 // At least one will become an actual new basic block.
5634 BasicBlock *TrueBlock = nullptr;
5635 BasicBlock *FalseBlock = nullptr;
5637 // Sink expensive instructions into the conditional blocks to avoid executing
5638 // them speculatively.
5639 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
5640 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
5641 EndBlock->getParent(), EndBlock);
5642 auto *TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
5643 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
5644 TrueInst->moveBefore(TrueBranch);
5646 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
5647 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
5648 EndBlock->getParent(), EndBlock);
5649 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5650 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
5651 FalseInst->moveBefore(FalseBranch);
5654 // If there was nothing to sink, then arbitrarily choose the 'false' side
5655 // for a new input value to the PHI.
5656 if (TrueBlock == FalseBlock) {
5657 assert(TrueBlock == nullptr &&
5658 "Unexpected basic block transform while optimizing select");
5660 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
5661 EndBlock->getParent(), EndBlock);
5662 BranchInst::Create(EndBlock, FalseBlock);
5665 // Insert the real conditional branch based on the original condition.
5666 // If we did not create a new block for one of the 'true' or 'false' paths
5667 // of the condition, it means that side of the branch goes to the end block
5668 // directly and the path originates from the start block from the point of
5669 // view of the new PHI.
5670 if (TrueBlock == nullptr) {
5671 BranchInst::Create(EndBlock, FalseBlock, SI->getCondition(), SI);
5672 TrueBlock = StartBlock;
5673 } else if (FalseBlock == nullptr) {
5674 BranchInst::Create(TrueBlock, EndBlock, SI->getCondition(), SI);
5675 FalseBlock = StartBlock;
5677 BranchInst::Create(TrueBlock, FalseBlock, SI->getCondition(), SI);
5680 // The select itself is replaced with a PHI Node.
5681 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
5683 PN->addIncoming(SI->getTrueValue(), TrueBlock);
5684 PN->addIncoming(SI->getFalseValue(), FalseBlock);
5686 SI->replaceAllUsesWith(PN);
5687 SI->eraseFromParent();
5689 // Instruct OptimizeBlock to skip to the next block.
5690 CurInstIterator = StartBlock->end();
5691 ++NumSelectsExpanded;
5695 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
5696 SmallVector<int, 16> Mask(SVI->getShuffleMask());
5698 for (unsigned i = 0; i < Mask.size(); ++i) {
5699 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
5701 SplatElem = Mask[i];
5707 /// Some targets have expensive vector shifts if the lanes aren't all the same
5708 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
5709 /// it's often worth sinking a shufflevector splat down to its use so that
5710 /// codegen can spot all lanes are identical.
5711 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
5712 BasicBlock *DefBB = SVI->getParent();
5714 // Only do this xform if variable vector shifts are particularly expensive.
5715 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
5718 // We only expect better codegen by sinking a shuffle if we can recognise a
5720 if (!isBroadcastShuffle(SVI))
5723 // InsertedShuffles - Only insert a shuffle in each block once.
5724 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
5726 bool MadeChange = false;
5727 for (User *U : SVI->users()) {
5728 Instruction *UI = cast<Instruction>(U);
5730 // Figure out which BB this ext is used in.
5731 BasicBlock *UserBB = UI->getParent();
5732 if (UserBB == DefBB) continue;
5734 // For now only apply this when the splat is used by a shift instruction.
5735 if (!UI->isShift()) continue;
5737 // Everything checks out, sink the shuffle if the user's block doesn't
5738 // already have a copy.
5739 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
5741 if (!InsertedShuffle) {
5742 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5743 assert(InsertPt != UserBB->end());
5745 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5746 SVI->getOperand(2), "", &*InsertPt);
5749 UI->replaceUsesOfWith(SVI, InsertedShuffle);
5753 // If we removed all uses, nuke the shuffle.
5754 if (SVI->use_empty()) {
5755 SVI->eraseFromParent();
5762 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
5766 Value *Cond = SI->getCondition();
5767 Type *OldType = Cond->getType();
5768 LLVMContext &Context = Cond->getContext();
5769 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
5770 unsigned RegWidth = RegType.getSizeInBits();
5772 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
5775 // If the register width is greater than the type width, expand the condition
5776 // of the switch instruction and each case constant to the width of the
5777 // register. By widening the type of the switch condition, subsequent
5778 // comparisons (for case comparisons) will not need to be extended to the
5779 // preferred register width, so we will potentially eliminate N-1 extends,
5780 // where N is the number of cases in the switch.
5781 auto *NewType = Type::getIntNTy(Context, RegWidth);
5783 // Zero-extend the switch condition and case constants unless the switch
5784 // condition is a function argument that is already being sign-extended.
5785 // In that case, we can avoid an unnecessary mask/extension by sign-extending
5786 // everything instead.
5787 Instruction::CastOps ExtType = Instruction::ZExt;
5788 if (auto *Arg = dyn_cast<Argument>(Cond))
5789 if (Arg->hasSExtAttr())
5790 ExtType = Instruction::SExt;
5792 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
5793 ExtInst->insertBefore(SI);
5794 SI->setCondition(ExtInst);
5795 for (SwitchInst::CaseIt Case : SI->cases()) {
5796 APInt NarrowConst = Case.getCaseValue()->getValue();
5797 APInt WideConst = (ExtType == Instruction::ZExt) ?
5798 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
5799 Case.setValue(ConstantInt::get(Context, WideConst));
5806 /// \brief Helper class to promote a scalar operation to a vector one.
5807 /// This class is used to move downward extractelement transition.
5809 /// a = vector_op <2 x i32>
5810 /// b = extractelement <2 x i32> a, i32 0
5815 /// a = vector_op <2 x i32>
5816 /// c = vector_op a (equivalent to scalar_op on the related lane)
5817 /// * d = extractelement <2 x i32> c, i32 0
5819 /// Assuming both extractelement and store can be combine, we get rid of the
5821 class VectorPromoteHelper {
5822 /// DataLayout associated with the current module.
5823 const DataLayout &DL;
5825 /// Used to perform some checks on the legality of vector operations.
5826 const TargetLowering &TLI;
5828 /// Used to estimated the cost of the promoted chain.
5829 const TargetTransformInfo &TTI;
5831 /// The transition being moved downwards.
5832 Instruction *Transition;
5833 /// The sequence of instructions to be promoted.
5834 SmallVector<Instruction *, 4> InstsToBePromoted;
5835 /// Cost of combining a store and an extract.
5836 unsigned StoreExtractCombineCost;
5837 /// Instruction that will be combined with the transition.
5838 Instruction *CombineInst;
5840 /// \brief The instruction that represents the current end of the transition.
5841 /// Since we are faking the promotion until we reach the end of the chain
5842 /// of computation, we need a way to get the current end of the transition.
5843 Instruction *getEndOfTransition() const {
5844 if (InstsToBePromoted.empty())
5846 return InstsToBePromoted.back();
5849 /// \brief Return the index of the original value in the transition.
5850 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
5851 /// c, is at index 0.
5852 unsigned getTransitionOriginalValueIdx() const {
5853 assert(isa<ExtractElementInst>(Transition) &&
5854 "Other kind of transitions are not supported yet");
5858 /// \brief Return the index of the index in the transition.
5859 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
5861 unsigned getTransitionIdx() const {
5862 assert(isa<ExtractElementInst>(Transition) &&
5863 "Other kind of transitions are not supported yet");
5867 /// \brief Get the type of the transition.
5868 /// This is the type of the original value.
5869 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
5870 /// transition is <2 x i32>.
5871 Type *getTransitionType() const {
5872 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
5875 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
5876 /// I.e., we have the following sequence:
5877 /// Def = Transition <ty1> a to <ty2>
5878 /// b = ToBePromoted <ty2> Def, ...
5880 /// b = ToBePromoted <ty1> a, ...
5881 /// Def = Transition <ty1> ToBePromoted to <ty2>
5882 void promoteImpl(Instruction *ToBePromoted);
5884 /// \brief Check whether or not it is profitable to promote all the
5885 /// instructions enqueued to be promoted.
5886 bool isProfitableToPromote() {
5887 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
5888 unsigned Index = isa<ConstantInt>(ValIdx)
5889 ? cast<ConstantInt>(ValIdx)->getZExtValue()
5891 Type *PromotedType = getTransitionType();
5893 StoreInst *ST = cast<StoreInst>(CombineInst);
5894 unsigned AS = ST->getPointerAddressSpace();
5895 unsigned Align = ST->getAlignment();
5896 // Check if this store is supported.
5897 if (!TLI.allowsMisalignedMemoryAccesses(
5898 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
5900 // If this is not supported, there is no way we can combine
5901 // the extract with the store.
5905 // The scalar chain of computation has to pay for the transition
5906 // scalar to vector.
5907 // The vector chain has to account for the combining cost.
5908 uint64_t ScalarCost =
5909 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
5910 uint64_t VectorCost = StoreExtractCombineCost;
5911 for (const auto &Inst : InstsToBePromoted) {
5912 // Compute the cost.
5913 // By construction, all instructions being promoted are arithmetic ones.
5914 // Moreover, one argument is a constant that can be viewed as a splat
5916 Value *Arg0 = Inst->getOperand(0);
5917 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
5918 isa<ConstantFP>(Arg0);
5919 TargetTransformInfo::OperandValueKind Arg0OVK =
5920 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
5921 : TargetTransformInfo::OK_AnyValue;
5922 TargetTransformInfo::OperandValueKind Arg1OVK =
5923 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
5924 : TargetTransformInfo::OK_AnyValue;
5925 ScalarCost += TTI.getArithmeticInstrCost(
5926 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
5927 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
5930 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
5931 << ScalarCost << "\nVector: " << VectorCost << '\n');
5932 return ScalarCost > VectorCost;
5935 /// \brief Generate a constant vector with \p Val with the same
5936 /// number of elements as the transition.
5937 /// \p UseSplat defines whether or not \p Val should be replicated
5938 /// across the whole vector.
5939 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
5940 /// otherwise we generate a vector with as many undef as possible:
5941 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
5942 /// used at the index of the extract.
5943 Value *getConstantVector(Constant *Val, bool UseSplat) const {
5944 unsigned ExtractIdx = UINT_MAX;
5946 // If we cannot determine where the constant must be, we have to
5947 // use a splat constant.
5948 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
5949 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
5950 ExtractIdx = CstVal->getSExtValue();
5955 unsigned End = getTransitionType()->getVectorNumElements();
5957 return ConstantVector::getSplat(End, Val);
5959 SmallVector<Constant *, 4> ConstVec;
5960 UndefValue *UndefVal = UndefValue::get(Val->getType());
5961 for (unsigned Idx = 0; Idx != End; ++Idx) {
5962 if (Idx == ExtractIdx)
5963 ConstVec.push_back(Val);
5965 ConstVec.push_back(UndefVal);
5967 return ConstantVector::get(ConstVec);
5970 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
5971 /// in \p Use can trigger undefined behavior.
5972 static bool canCauseUndefinedBehavior(const Instruction *Use,
5973 unsigned OperandIdx) {
5974 // This is not safe to introduce undef when the operand is on
5975 // the right hand side of a division-like instruction.
5976 if (OperandIdx != 1)
5978 switch (Use->getOpcode()) {
5981 case Instruction::SDiv:
5982 case Instruction::UDiv:
5983 case Instruction::SRem:
5984 case Instruction::URem:
5986 case Instruction::FDiv:
5987 case Instruction::FRem:
5988 return !Use->hasNoNaNs();
5990 llvm_unreachable(nullptr);
5994 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
5995 const TargetTransformInfo &TTI, Instruction *Transition,
5996 unsigned CombineCost)
5997 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
5998 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
5999 assert(Transition && "Do not know how to promote null");
6002 /// \brief Check if we can promote \p ToBePromoted to \p Type.
6003 bool canPromote(const Instruction *ToBePromoted) const {
6004 // We could support CastInst too.
6005 return isa<BinaryOperator>(ToBePromoted);
6008 /// \brief Check if it is profitable to promote \p ToBePromoted
6009 /// by moving downward the transition through.
6010 bool shouldPromote(const Instruction *ToBePromoted) const {
6011 // Promote only if all the operands can be statically expanded.
6012 // Indeed, we do not want to introduce any new kind of transitions.
6013 for (const Use &U : ToBePromoted->operands()) {
6014 const Value *Val = U.get();
6015 if (Val == getEndOfTransition()) {
6016 // If the use is a division and the transition is on the rhs,
6017 // we cannot promote the operation, otherwise we may create a
6018 // division by zero.
6019 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
6023 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
6024 !isa<ConstantFP>(Val))
6027 // Check that the resulting operation is legal.
6028 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
6031 return StressStoreExtract ||
6032 TLI.isOperationLegalOrCustom(
6033 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
6036 /// \brief Check whether or not \p Use can be combined
6037 /// with the transition.
6038 /// I.e., is it possible to do Use(Transition) => AnotherUse?
6039 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
6041 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
6042 void enqueueForPromotion(Instruction *ToBePromoted) {
6043 InstsToBePromoted.push_back(ToBePromoted);
6046 /// \brief Set the instruction that will be combined with the transition.
6047 void recordCombineInstruction(Instruction *ToBeCombined) {
6048 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
6049 CombineInst = ToBeCombined;
6052 /// \brief Promote all the instructions enqueued for promotion if it is
6054 /// \return True if the promotion happened, false otherwise.
6056 // Check if there is something to promote.
6057 // Right now, if we do not have anything to combine with,
6058 // we assume the promotion is not profitable.
6059 if (InstsToBePromoted.empty() || !CombineInst)
6063 if (!StressStoreExtract && !isProfitableToPromote())
6067 for (auto &ToBePromoted : InstsToBePromoted)
6068 promoteImpl(ToBePromoted);
6069 InstsToBePromoted.clear();
6073 } // End of anonymous namespace.
6075 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
6076 // At this point, we know that all the operands of ToBePromoted but Def
6077 // can be statically promoted.
6078 // For Def, we need to use its parameter in ToBePromoted:
6079 // b = ToBePromoted ty1 a
6080 // Def = Transition ty1 b to ty2
6081 // Move the transition down.
6082 // 1. Replace all uses of the promoted operation by the transition.
6083 // = ... b => = ... Def.
6084 assert(ToBePromoted->getType() == Transition->getType() &&
6085 "The type of the result of the transition does not match "
6087 ToBePromoted->replaceAllUsesWith(Transition);
6088 // 2. Update the type of the uses.
6089 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
6090 Type *TransitionTy = getTransitionType();
6091 ToBePromoted->mutateType(TransitionTy);
6092 // 3. Update all the operands of the promoted operation with promoted
6094 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
6095 for (Use &U : ToBePromoted->operands()) {
6096 Value *Val = U.get();
6097 Value *NewVal = nullptr;
6098 if (Val == Transition)
6099 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
6100 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
6101 isa<ConstantFP>(Val)) {
6102 // Use a splat constant if it is not safe to use undef.
6103 NewVal = getConstantVector(
6104 cast<Constant>(Val),
6105 isa<UndefValue>(Val) ||
6106 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
6108 llvm_unreachable("Did you modified shouldPromote and forgot to update "
6110 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
6112 Transition->removeFromParent();
6113 Transition->insertAfter(ToBePromoted);
6114 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
6117 /// Some targets can do store(extractelement) with one instruction.
6118 /// Try to push the extractelement towards the stores when the target
6119 /// has this feature and this is profitable.
6120 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
6121 unsigned CombineCost = UINT_MAX;
6122 if (DisableStoreExtract || !TLI ||
6123 (!StressStoreExtract &&
6124 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
6125 Inst->getOperand(1), CombineCost)))
6128 // At this point we know that Inst is a vector to scalar transition.
6129 // Try to move it down the def-use chain, until:
6130 // - We can combine the transition with its single use
6131 // => we got rid of the transition.
6132 // - We escape the current basic block
6133 // => we would need to check that we are moving it at a cheaper place and
6134 // we do not do that for now.
6135 BasicBlock *Parent = Inst->getParent();
6136 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
6137 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
6138 // If the transition has more than one use, assume this is not going to be
6140 while (Inst->hasOneUse()) {
6141 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
6142 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
6144 if (ToBePromoted->getParent() != Parent) {
6145 DEBUG(dbgs() << "Instruction to promote is in a different block ("
6146 << ToBePromoted->getParent()->getName()
6147 << ") than the transition (" << Parent->getName() << ").\n");
6151 if (VPH.canCombine(ToBePromoted)) {
6152 DEBUG(dbgs() << "Assume " << *Inst << '\n'
6153 << "will be combined with: " << *ToBePromoted << '\n');
6154 VPH.recordCombineInstruction(ToBePromoted);
6155 bool Changed = VPH.promote();
6156 NumStoreExtractExposed += Changed;
6160 DEBUG(dbgs() << "Try promoting.\n");
6161 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
6164 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
6166 VPH.enqueueForPromotion(ToBePromoted);
6167 Inst = ToBePromoted;
6172 bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
6173 // Bail out if we inserted the instruction to prevent optimizations from
6174 // stepping on each other's toes.
6175 if (InsertedInsts.count(I))
6178 if (PHINode *P = dyn_cast<PHINode>(I)) {
6179 // It is possible for very late stage optimizations (such as SimplifyCFG)
6180 // to introduce PHI nodes too late to be cleaned up. If we detect such a
6181 // trivial PHI, go ahead and zap it here.
6182 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
6183 P->replaceAllUsesWith(V);
6184 P->eraseFromParent();
6191 if (CastInst *CI = dyn_cast<CastInst>(I)) {
6192 // If the source of the cast is a constant, then this should have
6193 // already been constant folded. The only reason NOT to constant fold
6194 // it is if something (e.g. LSR) was careful to place the constant
6195 // evaluation in a block other than then one that uses it (e.g. to hoist
6196 // the address of globals out of a loop). If this is the case, we don't
6197 // want to forward-subst the cast.
6198 if (isa<Constant>(CI->getOperand(0)))
6201 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
6204 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6205 /// Sink a zext or sext into its user blocks if the target type doesn't
6206 /// fit in one register
6208 TLI->getTypeAction(CI->getContext(),
6209 TLI->getValueType(*DL, CI->getType())) ==
6210 TargetLowering::TypeExpandInteger) {
6211 return SinkCast(CI);
6213 bool MadeChange = moveExtToFormExtLoad(I);
6214 return MadeChange | optimizeExtUses(I);
6220 if (CmpInst *CI = dyn_cast<CmpInst>(I))
6221 if (!TLI || !TLI->hasMultipleConditionRegisters())
6222 return OptimizeCmpExpression(CI);
6224 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
6225 stripInvariantGroupMetadata(*LI);
6227 bool Modified = optimizeLoadExt(LI);
6228 unsigned AS = LI->getPointerAddressSpace();
6229 Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
6235 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
6236 stripInvariantGroupMetadata(*SI);
6238 unsigned AS = SI->getPointerAddressSpace();
6239 return optimizeMemoryInst(I, SI->getOperand(1),
6240 SI->getOperand(0)->getType(), AS);
6245 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
6247 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
6248 BinOp->getOpcode() == Instruction::LShr)) {
6249 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
6250 if (TLI && CI && TLI->hasExtractBitsInsn())
6251 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
6256 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
6257 if (GEPI->hasAllZeroIndices()) {
6258 /// The GEP operand must be a pointer, so must its result -> BitCast
6259 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
6260 GEPI->getName(), GEPI);
6261 GEPI->replaceAllUsesWith(NC);
6262 GEPI->eraseFromParent();
6264 optimizeInst(NC, ModifiedDT);
6270 if (CallInst *CI = dyn_cast<CallInst>(I))
6271 return optimizeCallInst(CI, ModifiedDT);
6273 if (SelectInst *SI = dyn_cast<SelectInst>(I))
6274 return optimizeSelectInst(SI);
6276 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
6277 return optimizeShuffleVectorInst(SVI);
6279 if (auto *Switch = dyn_cast<SwitchInst>(I))
6280 return optimizeSwitchInst(Switch);
6282 if (isa<ExtractElementInst>(I))
6283 return optimizeExtractElementInst(I);
6288 /// Given an OR instruction, check to see if this is a bitreverse
6289 /// idiom. If so, insert the new intrinsic and return true.
6290 static bool makeBitReverse(Instruction &I, const DataLayout &DL,
6291 const TargetLowering &TLI) {
6292 if (!I.getType()->isIntegerTy() ||
6293 !TLI.isOperationLegalOrCustom(ISD::BITREVERSE,
6294 TLI.getValueType(DL, I.getType(), true)))
6297 SmallVector<Instruction*, 4> Insts;
6298 if (!recognizeBitReverseOrBSwapIdiom(&I, false, true, Insts))
6300 Instruction *LastInst = Insts.back();
6301 I.replaceAllUsesWith(LastInst);
6302 RecursivelyDeleteTriviallyDeadInstructions(&I);
6306 // In this pass we look for GEP and cast instructions that are used
6307 // across basic blocks and rewrite them to improve basic-block-at-a-time
6309 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
6311 bool MadeChange = false;
6313 CurInstIterator = BB.begin();
6314 while (CurInstIterator != BB.end()) {
6315 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
6320 bool MadeBitReverse = true;
6321 while (TLI && MadeBitReverse) {
6322 MadeBitReverse = false;
6323 for (auto &I : reverse(BB)) {
6324 if (makeBitReverse(I, *DL, *TLI)) {
6325 MadeBitReverse = MadeChange = true;
6330 MadeChange |= dupRetToEnableTailCallOpts(&BB);
6335 // llvm.dbg.value is far away from the value then iSel may not be able
6336 // handle it properly. iSel will drop llvm.dbg.value if it can not
6337 // find a node corresponding to the value.
6338 bool CodeGenPrepare::placeDbgValues(Function &F) {
6339 bool MadeChange = false;
6340 for (BasicBlock &BB : F) {
6341 Instruction *PrevNonDbgInst = nullptr;
6342 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
6343 Instruction *Insn = &*BI++;
6344 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
6345 // Leave dbg.values that refer to an alloca alone. These
6346 // instrinsics describe the address of a variable (= the alloca)
6347 // being taken. They should not be moved next to the alloca
6348 // (and to the beginning of the scope), but rather stay close to
6349 // where said address is used.
6350 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
6351 PrevNonDbgInst = Insn;
6355 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
6356 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
6357 // If VI is a phi in a block with an EHPad terminator, we can't insert
6359 if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
6361 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
6362 DVI->removeFromParent();
6363 if (isa<PHINode>(VI))
6364 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
6366 DVI->insertAfter(VI);
6375 // If there is a sequence that branches based on comparing a single bit
6376 // against zero that can be combined into a single instruction, and the
6377 // target supports folding these into a single instruction, sink the
6378 // mask and compare into the branch uses. Do this before OptimizeBlock ->
6379 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
6381 bool CodeGenPrepare::sinkAndCmp(Function &F) {
6382 if (!EnableAndCmpSinking)
6384 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
6386 bool MadeChange = false;
6387 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
6388 BasicBlock *BB = &*I++;
6390 // Does this BB end with the following?
6391 // %andVal = and %val, #single-bit-set
6392 // %icmpVal = icmp %andResult, 0
6393 // br i1 %cmpVal label %dest1, label %dest2"
6394 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
6395 if (!Brcc || !Brcc->isConditional())
6397 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
6398 if (!Cmp || Cmp->getParent() != BB)
6400 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
6401 if (!Zero || !Zero->isZero())
6403 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
6404 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
6406 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
6407 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
6409 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
6411 // Push the "and; icmp" for any users that are conditional branches.
6412 // Since there can only be one branch use per BB, we don't need to keep
6413 // track of which BBs we insert into.
6414 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
6418 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
6420 if (!BrccUser || !BrccUser->isConditional())
6422 BasicBlock *UserBB = BrccUser->getParent();
6423 if (UserBB == BB) continue;
6424 DEBUG(dbgs() << "found Brcc use\n");
6426 // Sink the "and; icmp" to use.
6428 BinaryOperator *NewAnd =
6429 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
6432 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
6436 DEBUG(BrccUser->getParent()->dump());
6442 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
6443 /// success, or returns false if no or invalid metadata was found.
6444 static bool extractBranchMetadata(BranchInst *BI,
6445 uint64_t &ProbTrue, uint64_t &ProbFalse) {
6446 assert(BI->isConditional() &&
6447 "Looking for probabilities on unconditional branch?");
6448 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
6449 if (!ProfileData || ProfileData->getNumOperands() != 3)
6452 const auto *CITrue =
6453 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
6454 const auto *CIFalse =
6455 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
6456 if (!CITrue || !CIFalse)
6459 ProbTrue = CITrue->getValue().getZExtValue();
6460 ProbFalse = CIFalse->getValue().getZExtValue();
6465 /// \brief Scale down both weights to fit into uint32_t.
6466 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
6467 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
6468 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
6469 NewTrue = NewTrue / Scale;
6470 NewFalse = NewFalse / Scale;
6473 /// \brief Some targets prefer to split a conditional branch like:
6475 /// %0 = icmp ne i32 %a, 0
6476 /// %1 = icmp ne i32 %b, 0
6477 /// %or.cond = or i1 %0, %1
6478 /// br i1 %or.cond, label %TrueBB, label %FalseBB
6480 /// into multiple branch instructions like:
6483 /// %0 = icmp ne i32 %a, 0
6484 /// br i1 %0, label %TrueBB, label %bb2
6486 /// %1 = icmp ne i32 %b, 0
6487 /// br i1 %1, label %TrueBB, label %FalseBB
6489 /// This usually allows instruction selection to do even further optimizations
6490 /// and combine the compare with the branch instruction. Currently this is
6491 /// applied for targets which have "cheap" jump instructions.
6493 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
6495 bool CodeGenPrepare::splitBranchCondition(Function &F) {
6496 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
6499 bool MadeChange = false;
6500 for (auto &BB : F) {
6501 // Does this BB end with the following?
6502 // %cond1 = icmp|fcmp|binary instruction ...
6503 // %cond2 = icmp|fcmp|binary instruction ...
6504 // %cond.or = or|and i1 %cond1, cond2
6505 // br i1 %cond.or label %dest1, label %dest2"
6506 BinaryOperator *LogicOp;
6507 BasicBlock *TBB, *FBB;
6508 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
6511 auto *Br1 = cast<BranchInst>(BB.getTerminator());
6512 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
6516 Value *Cond1, *Cond2;
6517 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
6518 m_OneUse(m_Value(Cond2)))))
6519 Opc = Instruction::And;
6520 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
6521 m_OneUse(m_Value(Cond2)))))
6522 Opc = Instruction::Or;
6526 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
6527 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
6530 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
6533 auto *InsertBefore = std::next(Function::iterator(BB))
6534 .getNodePtrUnchecked();
6535 auto TmpBB = BasicBlock::Create(BB.getContext(),
6536 BB.getName() + ".cond.split",
6537 BB.getParent(), InsertBefore);
6539 // Update original basic block by using the first condition directly by the
6540 // branch instruction and removing the no longer needed and/or instruction.
6541 Br1->setCondition(Cond1);
6542 LogicOp->eraseFromParent();
6544 // Depending on the conditon we have to either replace the true or the false
6545 // successor of the original branch instruction.
6546 if (Opc == Instruction::And)
6547 Br1->setSuccessor(0, TmpBB);
6549 Br1->setSuccessor(1, TmpBB);
6551 // Fill in the new basic block.
6552 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
6553 if (auto *I = dyn_cast<Instruction>(Cond2)) {
6554 I->removeFromParent();
6555 I->insertBefore(Br2);
6558 // Update PHI nodes in both successors. The original BB needs to be
6559 // replaced in one succesor's PHI nodes, because the branch comes now from
6560 // the newly generated BB (NewBB). In the other successor we need to add one
6561 // incoming edge to the PHI nodes, because both branch instructions target
6562 // now the same successor. Depending on the original branch condition
6563 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
6564 // we perfrom the correct update for the PHI nodes.
6565 // This doesn't change the successor order of the just created branch
6566 // instruction (or any other instruction).
6567 if (Opc == Instruction::Or)
6568 std::swap(TBB, FBB);
6570 // Replace the old BB with the new BB.
6571 for (auto &I : *TBB) {
6572 PHINode *PN = dyn_cast<PHINode>(&I);
6576 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
6577 PN->setIncomingBlock(i, TmpBB);
6580 // Add another incoming edge form the new BB.
6581 for (auto &I : *FBB) {
6582 PHINode *PN = dyn_cast<PHINode>(&I);
6585 auto *Val = PN->getIncomingValueForBlock(&BB);
6586 PN->addIncoming(Val, TmpBB);
6589 // Update the branch weights (from SelectionDAGBuilder::
6590 // FindMergedConditions).
6591 if (Opc == Instruction::Or) {
6592 // Codegen X | Y as:
6601 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
6602 // The requirement is that
6603 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
6604 // = TrueProb for orignal BB.
6605 // Assuming the orignal weights are A and B, one choice is to set BB1's
6606 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
6608 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
6609 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
6610 // TmpBB, but the math is more complicated.
6611 uint64_t TrueWeight, FalseWeight;
6612 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
6613 uint64_t NewTrueWeight = TrueWeight;
6614 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
6615 scaleWeights(NewTrueWeight, NewFalseWeight);
6616 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6617 .createBranchWeights(TrueWeight, FalseWeight));
6619 NewTrueWeight = TrueWeight;
6620 NewFalseWeight = 2 * FalseWeight;
6621 scaleWeights(NewTrueWeight, NewFalseWeight);
6622 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6623 .createBranchWeights(TrueWeight, FalseWeight));
6626 // Codegen X & Y as:
6634 // This requires creation of TmpBB after CurBB.
6636 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
6637 // The requirement is that
6638 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
6639 // = FalseProb for orignal BB.
6640 // Assuming the orignal weights are A and B, one choice is to set BB1's
6641 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
6643 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
6644 uint64_t TrueWeight, FalseWeight;
6645 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
6646 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
6647 uint64_t NewFalseWeight = FalseWeight;
6648 scaleWeights(NewTrueWeight, NewFalseWeight);
6649 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6650 .createBranchWeights(TrueWeight, FalseWeight));
6652 NewTrueWeight = 2 * TrueWeight;
6653 NewFalseWeight = FalseWeight;
6654 scaleWeights(NewTrueWeight, NewFalseWeight);
6655 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6656 .createBranchWeights(TrueWeight, FalseWeight));
6660 // Note: No point in getting fancy here, since the DT info is never
6661 // available to CodeGenPrepare.
6666 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
6672 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
6673 if (auto *InvariantMD = I.getMetadata(LLVMContext::MD_invariant_group))
6674 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());