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 assert(TargetIntegerType->getTypeID() == Type::IntegerTyID);
332 auto* FromType = dyn_cast<IntegerType>(DepVal->getType());
333 auto* ToType = dyn_cast<IntegerType>(TargetIntegerType);
334 assert(FromType && ToType);
335 if (FromType->getBitWidth() <= ToType->getBitWidth()) {
336 CastOp = Instruction::ZExt;
338 CastOp = Instruction::Trunc;
342 case Type::FloatTyID:
343 case Type::DoubleTyID: {
344 CastOp = Instruction::FPToSI;
347 case Type::PointerTyID: {
348 CastOp = Instruction::PtrToInt;
354 return Builder.CreateCast(CastOp, DepVal, TargetIntegerType);
357 // Given a value, if it's a tainted address, this function returns the
358 // instruction that ORs the "dependence value" with the "original address".
359 // Otherwise, returns nullptr. This instruction is the first OR instruction
360 // where one of its operand is an AND instruction with an operand being 0.
362 // E.g., it returns '%4 = or i32 %3, %2' given 'CurrentAddress' is '%5'.
363 // %0 = load i32, i32* @y, align 4, !tbaa !1
364 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
365 // %1 = sext i1 %cmp to i32
366 // %2 = ptrtoint i32* @x to i32
367 // %3 = and i32 %1, 0
368 // %4 = or i32 %3, %2
369 // %5 = inttoptr i32 %4 to i32*
370 // store i32 1, i32* %5, align 4
371 Instruction* getOrAddress(Value* CurrentAddress) {
372 // Is it a cast from integer to pointer type.
373 Instruction* OrAddress = nullptr;
374 Instruction* AndDep = nullptr;
375 Instruction* CastToInt = nullptr;
376 Value* ActualAddress = nullptr;
377 Constant* ZeroConst = nullptr;
379 const Instruction* CastToPtr = dyn_cast<Instruction>(CurrentAddress);
380 if (CastToPtr && CastToPtr->getOpcode() == Instruction::IntToPtr) {
381 // Is it an OR instruction: %1 = or %and, %actualAddress.
382 if ((OrAddress = dyn_cast<Instruction>(CastToPtr->getOperand(0))) &&
383 OrAddress->getOpcode() == Instruction::Or) {
384 // The first operand should be and AND instruction.
385 AndDep = dyn_cast<Instruction>(OrAddress->getOperand(0));
386 if (AndDep && AndDep->getOpcode() == Instruction::And) {
387 // Also make sure its first operand of the "AND" is 0, or the "AND" is
388 // marked explicitly by "NoInstCombine".
389 if ((ZeroConst = dyn_cast<Constant>(AndDep->getOperand(1))) &&
390 ZeroConst->isNullValue()) {
396 // Looks like it's not been tainted.
400 // Given a value, if it's a tainted address, this function returns the
401 // instruction that taints the "dependence value". Otherwise, returns nullptr.
402 // This instruction is the last AND instruction where one of its operand is 0.
403 // E.g., it returns '%3' given 'CurrentAddress' is '%5'.
404 // %0 = load i32, i32* @y, align 4, !tbaa !1
405 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
406 // %1 = sext i1 %cmp to i32
407 // %2 = ptrtoint i32* @x to i32
408 // %3 = and i32 %1, 0
409 // %4 = or i32 %3, %2
410 // %5 = inttoptr i32 %4 to i32*
411 // store i32 1, i32* %5, align 4
412 Instruction* getAndDependence(Value* CurrentAddress) {
413 // If 'CurrentAddress' is tainted, get the OR instruction.
414 auto* OrAddress = getOrAddress(CurrentAddress);
415 if (OrAddress == nullptr) {
419 // No need to check the operands.
420 auto* AndDepInst = dyn_cast<Instruction>(OrAddress->getOperand(0));
425 // Given a value, if it's a tainted address, this function returns
426 // the "dependence value", which is the first operand in the AND instruction.
427 // E.g., it returns '%1' given 'CurrentAddress' is '%5'.
428 // %0 = load i32, i32* @y, align 4, !tbaa !1
429 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
430 // %1 = sext i1 %cmp to i32
431 // %2 = ptrtoint i32* @x to i32
432 // %3 = and i32 %1, 0
433 // %4 = or i32 %3, %2
434 // %5 = inttoptr i32 %4 to i32*
435 // store i32 1, i32* %5, align 4
436 Value* getDependence(Value* CurrentAddress) {
437 auto* AndInst = getAndDependence(CurrentAddress);
438 if (AndInst == nullptr) {
441 return AndInst->getOperand(0);
444 // Given an address that has been tainted, returns the only condition it depends
445 // on, if any; otherwise, returns nullptr.
446 Value* getConditionDependence(Value* Address) {
447 auto* Dep = getDependence(Address);
448 if (Dep == nullptr) {
449 // 'Address' has not been dependence-tainted.
453 Value* Operand = Dep;
455 auto* Inst = dyn_cast<Instruction>(Operand);
456 if (Inst == nullptr) {
457 // Non-instruction type does not have condition dependence.
460 if (Inst->getOpcode() == Instruction::ICmp) {
463 if (Inst->getNumOperands() != 1) {
466 Operand = Inst->getOperand(0);
472 // Conservatively decides whether the dependence set of 'Val1' includes the
473 // dependence set of 'Val2'. If 'ExpandSecondValue' is false, we do not expand
474 // 'Val2' and use that single value as its dependence set.
475 // If it returns true, it means the dependence set of 'Val1' includes that of
476 // 'Val2'; otherwise, it only means we cannot conclusively decide it.
477 bool dependenceSetInclusion(Value* Val1, Value* Val2,
478 int Val1ExpandLevel = 2 * kDependenceDepth,
479 int Val2ExpandLevel = kDependenceDepth) {
480 typedef SmallSet<Value*, 8> IncludingSet;
481 typedef SmallSet<Value*, 4> IncludedSet;
483 IncludingSet DepSet1;
485 // Look for more depths for the including set.
486 recursivelyFindDependence(&DepSet1, Val1, false /*Insert all visited nodes*/,
488 recursivelyFindDependence(&DepSet2, Val2, true /*Only insert leaf nodes*/,
491 auto set_inclusion = [](IncludingSet FullSet, IncludedSet Subset) {
492 for (auto* Dep : Subset) {
493 if (0 == FullSet.count(Dep)) {
499 bool inclusion = set_inclusion(DepSet1, DepSet2);
500 DEBUG(dbgs() << "[dependenceSetInclusion]: " << inclusion << "\n");
501 DEBUG(dbgs() << "Including set for: " << *Val1 << "\n");
502 DEBUG(for (const auto* Dep : DepSet1) { dbgs() << "\t\t" << *Dep << "\n"; });
503 DEBUG(dbgs() << "Included set for: " << *Val2 << "\n");
504 DEBUG(for (const auto* Dep : DepSet2) { dbgs() << "\t\t" << *Dep << "\n"; });
509 // Recursively iterates through the operands spawned from 'DepVal'. If there
510 // exists a single value that 'DepVal' only depends on, we call that value the
511 // root dependence of 'DepVal' and return it. Otherwise, return 'DepVal'.
512 Value* getRootDependence(Value* DepVal) {
513 SmallSet<Value*, 8> DepSet;
514 for (unsigned depth = kDependenceDepth; depth > 0; --depth) {
515 recursivelyFindDependence(&DepSet, DepVal, true /*Only insert leaf nodes*/,
517 if (DepSet.size() == 1) {
518 return *DepSet.begin();
525 // This function actually taints 'DepVal' to the address to 'SI'. If the
527 // of 'SI' already depends on whatever 'DepVal' depends on, this function
528 // doesn't do anything and returns false. Otherwise, returns true.
530 // This effect forces the store and any stores that comes later to depend on
531 // 'DepVal'. For example, we have a condition "cond", and a store instruction
532 // "s: STORE addr, val". If we want "s" (and any later store) to depend on
533 // "cond", we do the following:
534 // %conv = sext i1 %cond to i32
535 // %addrVal = ptrtoint i32* %addr to i32
536 // %andCond = and i32 conv, 0;
537 // %orAddr = or i32 %andCond, %addrVal;
538 // %NewAddr = inttoptr i32 %orAddr to i32*;
540 // This is a more concrete example:
542 // %0 = load i32, i32* @y, align 4, !tbaa !1
543 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
544 // %1 = sext i1 %cmp to i32
545 // %2 = ptrtoint i32* @x to i32
546 // %3 = and i32 %1, 0
547 // %4 = or i32 %3, %2
548 // %5 = inttoptr i32 %4 to i32*
549 // store i32 1, i32* %5, align 4
550 bool taintStoreAddress(StoreInst* SI, Value* DepVal) {
551 // Set the insertion point right after the 'DepVal'.
552 Instruction* Inst = nullptr;
553 IRBuilder<true, NoFolder> Builder(SI);
554 BasicBlock* BB = SI->getParent();
555 Value* Address = SI->getPointerOperand();
556 Type* TargetIntegerType =
557 IntegerType::get(Address->getContext(),
558 BB->getModule()->getDataLayout().getPointerSizeInBits());
560 // Does SI's address already depends on whatever 'DepVal' depends on?
561 if (StoreAddressDependOnValue(SI, DepVal)) {
565 // Figure out if there's a root variable 'DepVal' depends on. For example, we
566 // can extract "getelementptr inbounds %struct, %struct* %0, i64 0, i32 123"
567 // to be "%struct* %0" since all other operands are constant.
568 auto* RootVal = getRootDependence(DepVal);
569 auto* RootInst = dyn_cast<Instruction>(RootVal);
570 auto* DepValInst = dyn_cast<Instruction>(DepVal);
571 if (RootInst && DepValInst &&
572 RootInst->getParent() == DepValInst->getParent()) {
576 // Is this already a dependence-tainted store?
577 Value* OldDep = getDependence(Address);
579 // The address of 'SI' has already been tainted. Just need to absorb the
580 // DepVal to the existing dependence in the address of SI.
581 Instruction* AndDep = getAndDependence(Address);
582 IRBuilder<true, NoFolder> Builder(AndDep);
583 Value* NewDep = nullptr;
584 if (DepVal->getType() == AndDep->getType()) {
585 NewDep = Builder.CreateAnd(OldDep, DepVal);
587 NewDep = Builder.CreateAnd(
588 OldDep, createCast(Builder, DepVal, TargetIntegerType));
591 auto* NewDepInst = dyn_cast<Instruction>(NewDep);
593 // Use the new AND instruction as the dependence
594 AndDep->setOperand(0, NewDep);
598 // SI's address has not been tainted. Now taint it with 'DepVal'.
599 Value* CastDepToInt = createCast(Builder, DepVal, TargetIntegerType);
600 Value* PtrToIntCast = Builder.CreatePtrToInt(Address, TargetIntegerType);
602 Builder.CreateAnd(CastDepToInt, ConstantInt::get(TargetIntegerType, 0));
603 auto AndInst = dyn_cast<Instruction>(AndDepVal);
604 // XXX-comment: The original IR InstCombiner would change our and instruction
605 // to a select and then the back end optimize the condition out. We attach a
606 // flag to instructions and set it here to inform the InstCombiner to not to
607 // touch this and instruction at all.
608 Value* OrAddr = Builder.CreateOr(AndDepVal, PtrToIntCast);
609 Value* NewAddr = Builder.CreateIntToPtr(OrAddr, Address->getType());
611 DEBUG(dbgs() << "[taintStoreAddress]\n"
612 << "Original store: " << *SI << '\n');
613 SI->setOperand(1, NewAddr);
616 DEBUG(dbgs() << "\tTargetIntegerType: " << *TargetIntegerType << '\n'
617 << "\tCast dependence value to integer: " << *CastDepToInt
619 << "\tCast address to integer: " << *PtrToIntCast << '\n'
620 << "\tAnd dependence value: " << *AndDepVal << '\n'
621 << "\tOr address: " << *OrAddr << '\n'
622 << "\tCast or instruction to address: " << *NewAddr << "\n\n");
627 // Looks for the previous store in the if block --- 'BrBB', which makes the
628 // speculative store 'StoreToHoist' safe.
629 Value* getSpeculativeStoreInPrevBB(StoreInst* StoreToHoist, BasicBlock* BrBB) {
630 assert(StoreToHoist && "StoreToHoist must be a real store");
632 Value* StorePtr = StoreToHoist->getPointerOperand();
634 // Look for a store to the same pointer in BrBB.
635 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(), RE = BrBB->rend();
637 Instruction* CurI = &*RI;
639 StoreInst* SI = dyn_cast<StoreInst>(CurI);
640 // Found the previous store make sure it stores to the same location.
641 // XXX-update: If the previous store's original untainted address are the
642 // same as 'StorePtr', we are also good to hoist the store.
643 if (SI && (SI->getPointerOperand() == StorePtr ||
644 GetUntaintedAddress(SI->getPointerOperand()) == StorePtr)) {
645 // Found the previous store, return its value operand.
651 "We should not reach here since this store is safe to speculate");
654 // XXX-comment: Returns true if it changes the code, false otherwise (the branch
655 // condition already depends on 'DepVal'.
656 bool taintConditionalBranch(BranchInst* BI, Value* DepVal) {
657 assert(BI->isConditional());
658 auto* Cond = BI->getOperand(0);
659 if (dependenceSetInclusion(Cond, DepVal)) {
660 // The dependence/ordering is self-evident.
664 IRBuilder<true, NoFolder> Builder(BI);
666 Builder.CreateAnd(DepVal, ConstantInt::get(DepVal->getType(), 0));
668 Builder.CreateTrunc(AndDep, IntegerType::get(DepVal->getContext(), 1));
669 auto* OrCond = Builder.CreateOr(TruncAndDep, Cond);
670 BI->setOperand(0, OrCond);
673 DEBUG(dbgs() << "\tTainted branch condition:\n" << *BI->getParent());
678 bool ConditionalBranchDependsOnValue(BranchInst* BI, Value* DepVal) {
679 assert(BI->isConditional());
680 auto* Cond = BI->getOperand(0);
681 return dependenceSetInclusion(Cond, DepVal);
684 // XXX-update: For a relaxed load 'LI', find the first immediate atomic store or
685 // the first conditional branch. Returns nullptr if there's no such immediately
686 // following store/branch instructions, which we can only enforce the load with
687 // 'acquire'. 'ChainedBB' contains all the blocks chained together with
688 // unconditional branches from 'BB' to the block with the first store/cond
690 template <typename Vector>
691 Instruction* findFirstStoreCondBranchInst(LoadInst* LI, Vector* ChainedBB) {
692 // In some situations, relaxed loads can be left as is:
693 // 1. The relaxed load is used to calculate the address of the immediate
695 // 2. The relaxed load is used as a condition in the immediate following
696 // condition, and there are no stores in between. This is actually quite
698 // int r1 = x.load(relaxed);
700 // y.store(1, relaxed);
702 // However, in this function, we don't deal with them directly. Instead, we
703 // just find the immediate following store/condition branch and return it.
705 assert(ChainedBB != nullptr && "Chained BB should not be nullptr");
706 auto* BB = LI->getParent();
707 ChainedBB->push_back(BB);
709 auto BBI = BasicBlock::iterator(LI);
712 for (; BBI != BE; BBI++) {
713 Instruction* Inst = &*BBI;
714 IntrinsicInst* II = dyn_cast<IntrinsicInst>(&*BBI);
715 if (II && II->getIntrinsicID() == Intrinsic::aarch64_stlxr) {
717 } else if (Inst->getOpcode() == Instruction::Store) {
719 } else if (Inst->getOpcode() == Instruction::Br) {
720 auto* BrInst = dyn_cast<BranchInst>(Inst);
721 if (BrInst->isConditional()) {
724 // Reinitialize iterators with the destination of the unconditional
726 BB = BrInst->getSuccessor(0);
727 ChainedBB->push_back(BB);
740 // XXX-update: Find the next node of the last relaxed load from 'FromInst' to
741 // 'ToInst'. If none, return 'ToInst'.
742 Instruction* findLastLoadNext(Instruction* FromInst, Instruction* ToInst) {
743 if (FromInst == ToInst) {
746 Instruction* LastLoad = ToInst;
747 auto* BB = FromInst->getParent();
749 auto BBI = BasicBlock::iterator(FromInst);
751 for (; BBI != BE && &*BBI != ToInst; BBI++) {
752 auto* LI = dyn_cast<LoadInst>(&*BBI);
753 if (LI == nullptr || !LI->isAtomic() || LI->getOrdering() != Monotonic) {
757 LastLoad = LastLoad->getNextNode();
762 // Inserts a fake conditional branch right after the instruction 'SplitInst',
763 // and the branch condition is 'Condition'. 'SplitInst' will be placed in the
764 // newly created block.
765 void AddFakeConditionalBranch(Instruction* SplitInst, Value* Condition) {
766 auto* BB = SplitInst->getParent();
767 TerminatorInst* ThenTerm = nullptr;
768 TerminatorInst* ElseTerm = nullptr;
769 SplitBlockAndInsertIfThenElse(Condition, SplitInst, &ThenTerm, &ElseTerm);
770 assert(ThenTerm && ElseTerm &&
771 "Then/Else terminators cannot be empty after basic block spliting");
772 auto* ThenBB = ThenTerm->getParent();
773 auto* ElseBB = ElseTerm->getParent();
774 auto* TailBB = ThenBB->getSingleSuccessor();
775 assert(TailBB && "Tail block cannot be empty after basic block spliting");
777 ThenBB->disableCanEliminateBlock();
778 ThenBB->disableCanEliminateBlock();
779 TailBB->disableCanEliminateBlock();
780 ThenBB->setName(BB->getName() + "Then.Fake");
781 ElseBB->setName(BB->getName() + "Else.Fake");
782 DEBUG(dbgs() << "Add fake conditional branch:\n"
784 << *ThenBB << "Else Block:\n"
788 // Returns true if the code is changed, and false otherwise.
789 void TaintRelaxedLoads(Instruction* UsageInst, Instruction* InsertPoint) {
790 // For better performance, we can add a "AND X 0" instruction before the
792 auto* BB = UsageInst->getParent();
793 if (InsertPoint == nullptr) {
794 InsertPoint = UsageInst->getNextNode();
796 // Insert instructions after PHI nodes.
797 while (dyn_cast<PHINode>(InsertPoint)) {
798 InsertPoint = InsertPoint->getNextNode();
800 // First thing is to cast 'UsageInst' to an integer type if necessary.
801 Value* AndTarget = nullptr;
802 Type* TargetIntegerType =
803 IntegerType::get(UsageInst->getContext(),
804 BB->getModule()->getDataLayout().getPointerSizeInBits());
806 // Check whether InsertPoint is a added fake conditional branch.
807 BranchInst* BI = nullptr;
808 if ((BI = dyn_cast<BranchInst>(InsertPoint)) && BI->isConditional()) {
809 auto* Cond = dyn_cast<Instruction>(BI->getOperand(0));
810 if (Cond && Cond->getOpcode() == Instruction::ICmp) {
811 auto* CmpInst = dyn_cast<ICmpInst>(Cond);
812 auto* Op0 = dyn_cast<Instruction>(Cond->getOperand(0));
813 auto* Op1 = dyn_cast<ConstantInt>(Cond->getOperand(1));
815 // %cmp = ICMP_NE %tmp, 0
818 // %tmp1 = And X, NewTaintedVal
819 // %tmp2 = And %tmp1, 0
820 // %cmp = ICMP_NE %tmp2, 0
822 if (CmpInst && CmpInst->getPredicate() == CmpInst::ICMP_NE && Op0 &&
823 Op0->getOpcode() == Instruction::And && Op1 && Op1->isZero()) {
824 auto* Op01 = dyn_cast<ConstantInt>(Op0->getOperand(1));
825 if (Op01 && Op01->isZero()) {
826 // Now we have a previously added fake cond branch.
827 auto* Op00 = Op0->getOperand(0);
828 IRBuilder<true, NoFolder> Builder(CmpInst);
829 if (Op00->getType() == UsageInst->getType()) {
830 AndTarget = UsageInst;
832 AndTarget = createCast(Builder, UsageInst, Op00->getType());
834 AndTarget = Builder.CreateAnd(Op00, AndTarget);
835 auto* AndZero = dyn_cast<Instruction>(Builder.CreateAnd(
836 AndTarget, Constant::getNullValue(AndTarget->getType())));
837 CmpInst->setOperand(0, AndZero);
844 IRBuilder<true, NoFolder> Builder(InsertPoint);
845 if (IntegerType::classof(UsageInst->getType())) {
846 AndTarget = UsageInst;
848 AndTarget = createCast(Builder, UsageInst, TargetIntegerType);
850 auto* AndZero = dyn_cast<Instruction>(
851 Builder.CreateAnd(AndTarget, Constant::getNullValue(AndTarget->getType())));
852 auto* FakeCondition = dyn_cast<Instruction>(Builder.CreateICmp(
853 CmpInst::ICMP_NE, AndZero, Constant::getNullValue(AndTarget->getType())));
854 AddFakeConditionalBranch(FakeCondition->getNextNode(), FakeCondition);
857 // XXX-comment: Finds the appropriate Value derived from an atomic load.
858 // 'ChainedBB' contains all the blocks chained together with unconditional
859 // branches from LI's parent BB to the block with the first store/cond branch.
860 // If we don't find any, it means 'LI' is not used at all (which should not
861 // happen in practice). We can simply set 'LI' to be acquire just to be safe.
862 template <typename Vector>
863 Instruction* findMostRecentDependenceUsage(LoadInst* LI, Instruction* LaterInst,
866 typedef SmallSet<Instruction*, 8> UsageSet;
867 typedef DenseMap<BasicBlock*, std::unique_ptr<UsageSet>> UsageMap;
868 assert(ChainedBB->size() >= 1 && "ChainedBB must have >=1 size");
869 // Mapping from basic block in 'ChainedBB' to the set of dependence usage of
870 // 'LI' in each block.
872 auto* LoadBB = LI->getParent();
873 usage_map[LoadBB] = make_unique<UsageSet>();
874 usage_map[LoadBB]->insert(LI);
876 for (auto* BB : *ChainedBB) {
877 if (usage_map[BB] == nullptr) {
878 usage_map[BB] = make_unique<UsageSet>();
880 auto& usage_set = usage_map[BB];
881 if (usage_set->size() == 0) {
882 // The value has not been used.
885 // Calculate the usage in the current BB first.
886 std::list<Value*> bb_usage_list;
887 std::copy(usage_set->begin(), usage_set->end(),
888 std::back_inserter(bb_usage_list));
889 for (auto list_iter = bb_usage_list.begin();
890 list_iter != bb_usage_list.end(); list_iter++) {
891 auto* val = *list_iter;
892 for (auto* U : val->users()) {
893 Instruction* Inst = nullptr;
894 if (!(Inst = dyn_cast<Instruction>(U))) {
897 assert(Inst && "Usage value must be an instruction");
899 std::find(ChainedBB->begin(), ChainedBB->end(), Inst->getParent());
900 if (iter == ChainedBB->end()) {
901 // Only care about usage within ChainedBB.
904 auto* UsageBB = *iter;
907 if (!usage_set->count(Inst)) {
908 bb_usage_list.push_back(Inst);
909 usage_set->insert(Inst);
913 if (usage_map[UsageBB] == nullptr) {
914 usage_map[UsageBB] = make_unique<UsageSet>();
916 usage_map[UsageBB]->insert(Inst);
922 // Pick one usage that is in LaterInst's block and that dominates 'LaterInst'.
923 auto* LaterBB = LaterInst->getParent();
924 auto& usage_set = usage_map[LaterBB];
925 Instruction* usage_inst = nullptr;
926 for (auto* inst : *usage_set) {
927 if (DT->dominates(inst, LaterInst)) {
933 assert(usage_inst && "The usage instruction in the same block but after the "
934 "later instruction");
938 // XXX-comment: Returns whether the code has been changed.
939 bool AddFakeConditionalBranchAfterMonotonicLoads(
940 SmallSet<LoadInst*, 1>& MonotonicLoadInsts, DominatorTree* DT) {
941 bool Changed = false;
942 while (!MonotonicLoadInsts.empty()) {
943 auto* LI = *MonotonicLoadInsts.begin();
944 MonotonicLoadInsts.erase(LI);
945 SmallVector<BasicBlock*, 2> ChainedBB;
946 auto* FirstInst = findFirstStoreCondBranchInst(LI, &ChainedBB);
947 if (FirstInst != nullptr) {
948 if (FirstInst->getOpcode() == Instruction::Store) {
949 if (StoreAddressDependOnValue(dyn_cast<StoreInst>(FirstInst), LI)) {
952 } else if (FirstInst->getOpcode() == Instruction::Br) {
953 if (ConditionalBranchDependsOnValue(dyn_cast<BranchInst>(FirstInst),
958 IntrinsicInst* II = dyn_cast<IntrinsicInst>(FirstInst);
959 if (!II || II->getIntrinsicID() != Intrinsic::aarch64_stlxr) {
960 dbgs() << "FirstInst=" << *FirstInst << "\n";
961 assert(false && "findFirstStoreCondBranchInst() should return a "
962 "store/condition branch instruction");
967 // We really need to process the relaxed load now. Note that if the next
968 // instruction is a RMW, it will be transformed into a control block, so we
969 // can safely only taint upcoming store instructions.
970 StoreInst* SI = nullptr;
971 IntrinsicInst* II = nullptr;
973 SI = dyn_cast<StoreInst>(FirstInst);
974 II = dyn_cast<IntrinsicInst>(FirstInst);
976 if (FirstInst && SI) {
977 // For immediately coming stores, taint the address of the store.
978 if (FirstInst->getParent() == LI->getParent() ||
979 DT->dominates(LI, FirstInst)) {
980 Changed != taintStoreAddress(SI, LI);
984 findMostRecentDependenceUsage(LI, FirstInst, &ChainedBB, DT);
986 LI->setOrdering(Acquire);
989 Changed |= taintStoreAddress(SI, Inst);
993 // No upcoming branch
995 TaintRelaxedLoads(LI, nullptr);
998 // For immediately coming branch, directly add a fake branch.
999 if (FirstInst->getParent() == LI->getParent() ||
1000 DT->dominates(LI, FirstInst)) {
1001 TaintRelaxedLoads(LI, FirstInst);
1005 findMostRecentDependenceUsage(LI, FirstInst, &ChainedBB, DT);
1007 TaintRelaxedLoads(Inst, FirstInst);
1009 LI->setOrdering(Acquire);
1019 /**** Implementations of public methods for dependence tainting ****/
1020 Value* GetUntaintedAddress(Value* CurrentAddress) {
1021 auto* OrAddress = getOrAddress(CurrentAddress);
1022 if (OrAddress == nullptr) {
1023 // Is it tainted by a select instruction?
1024 auto* Inst = dyn_cast<Instruction>(CurrentAddress);
1025 if (nullptr != Inst && Inst->getOpcode() == Instruction::Select) {
1026 // A selection instruction.
1027 if (Inst->getOperand(1) == Inst->getOperand(2)) {
1028 return Inst->getOperand(1);
1032 return CurrentAddress;
1034 Value* ActualAddress = nullptr;
1036 auto* CastToInt = dyn_cast<Instruction>(OrAddress->getOperand(1));
1037 if (CastToInt && CastToInt->getOpcode() == Instruction::PtrToInt) {
1038 return CastToInt->getOperand(0);
1040 // This should be a IntToPtr constant expression.
1041 ConstantExpr* PtrToIntExpr =
1042 dyn_cast<ConstantExpr>(OrAddress->getOperand(1));
1043 if (PtrToIntExpr && PtrToIntExpr->getOpcode() == Instruction::PtrToInt) {
1044 return PtrToIntExpr->getOperand(0);
1048 // Looks like it's not been dependence-tainted. Returns itself.
1049 return CurrentAddress;
1052 MemoryLocation GetUntaintedMemoryLocation(StoreInst* SI) {
1054 SI->getAAMetadata(AATags);
1055 const auto& DL = SI->getModule()->getDataLayout();
1056 const auto* OriginalAddr = GetUntaintedAddress(SI->getPointerOperand());
1057 DEBUG(if (OriginalAddr != SI->getPointerOperand()) {
1058 dbgs() << "[GetUntaintedMemoryLocation]\n"
1059 << "Storing address: " << *SI->getPointerOperand()
1060 << "\nUntainted address: " << *OriginalAddr << "\n";
1062 return MemoryLocation(OriginalAddr,
1063 DL.getTypeStoreSize(SI->getValueOperand()->getType()),
1067 bool TaintDependenceToStore(StoreInst* SI, Value* DepVal) {
1068 if (dependenceSetInclusion(SI, DepVal)) {
1072 bool tainted = taintStoreAddress(SI, DepVal);
1077 bool TaintDependenceToStoreAddress(StoreInst* SI, Value* DepVal) {
1078 if (dependenceSetInclusion(SI->getPointerOperand(), DepVal)) {
1082 bool tainted = taintStoreAddress(SI, DepVal);
1087 bool CompressTaintedStore(BasicBlock* BB) {
1088 // This function looks for windows of adajcent stores in 'BB' that satisfy the
1089 // following condition (and then do optimization):
1090 // *Addr(d1) = v1, d1 is a condition and is the only dependence the store's
1091 // address depends on && Dep(v1) includes Dep(d1);
1092 // *Addr(d2) = v2, d2 is a condition and is the only dependnece the store's
1093 // address depends on && Dep(v2) includes Dep(d2) &&
1094 // Dep(d2) includes Dep(d1);
1096 // *Addr(dN) = vN, dN is a condition and is the only dependence the store's
1097 // address depends on && Dep(dN) includes Dep(d"N-1").
1099 // As a result, Dep(dN) includes [Dep(d1) V ... V Dep(d"N-1")], so we can
1100 // safely transform the above to the following. In between these stores, we
1101 // can omit untainted stores to the same address 'Addr' since they internally
1102 // have dependence on the previous stores on the same address.
1107 for (auto BI = BB->begin(), BE = BB->end(); BI != BE; BI++) {
1108 // Look for the first store in such a window of adajacent stores.
1109 auto* FirstSI = dyn_cast<StoreInst>(&*BI);
1114 // The first store in the window must be tainted.
1115 auto* UntaintedAddress = GetUntaintedAddress(FirstSI->getPointerOperand());
1116 if (UntaintedAddress == FirstSI->getPointerOperand()) {
1120 // The first store's address must directly depend on and only depend on a
1122 auto* FirstSIDepCond = getConditionDependence(FirstSI->getPointerOperand());
1123 if (nullptr == FirstSIDepCond) {
1127 // Dep(first store's storing value) includes Dep(tainted dependence).
1128 if (!dependenceSetInclusion(FirstSI->getValueOperand(), FirstSIDepCond)) {
1132 // Look for subsequent stores to the same address that satisfy the condition
1133 // of "compressing the dependence".
1134 SmallVector<StoreInst*, 8> AdajacentStores;
1135 AdajacentStores.push_back(FirstSI);
1136 auto BII = BasicBlock::iterator(FirstSI);
1137 for (BII++; BII != BE; BII++) {
1138 auto* CurrSI = dyn_cast<StoreInst>(&*BII);
1140 if (BII->mayHaveSideEffects()) {
1141 // Be conservative. Instructions with side effects are similar to
1148 auto* OrigAddress = GetUntaintedAddress(CurrSI->getPointerOperand());
1149 auto* CurrSIDepCond = getConditionDependence(CurrSI->getPointerOperand());
1150 // All other stores must satisfy either:
1151 // A. 'CurrSI' is an untainted store to the same address, or
1152 // B. the combination of the following 5 subconditions:
1154 // 2. Untainted address is the same as the group's address;
1155 // 3. The address is tainted with a sole value which is a condition;
1156 // 4. The storing value depends on the condition in 3.
1157 // 5. The condition in 3 depends on the previous stores dependence
1160 // Condition A. Should ignore this store directly.
1161 if (OrigAddress == CurrSI->getPointerOperand() &&
1162 OrigAddress == UntaintedAddress) {
1165 // Check condition B.
1166 Value* Cond = nullptr;
1167 if (OrigAddress == CurrSI->getPointerOperand() ||
1168 OrigAddress != UntaintedAddress || CurrSIDepCond == nullptr ||
1169 !dependenceSetInclusion(CurrSI->getValueOperand(), CurrSIDepCond)) {
1170 // Check condition 1, 2, 3 & 4.
1174 // Check condition 5.
1175 StoreInst* PrevSI = AdajacentStores[AdajacentStores.size() - 1];
1176 auto* PrevSIDepCond = getConditionDependence(PrevSI->getPointerOperand());
1177 assert(PrevSIDepCond &&
1178 "Store in the group must already depend on a condtion");
1179 if (!dependenceSetInclusion(CurrSIDepCond, PrevSIDepCond)) {
1183 AdajacentStores.push_back(CurrSI);
1186 if (AdajacentStores.size() == 1) {
1187 // The outer loop should keep looking from the next store.
1191 // Now we have such a group of tainted stores to the same address.
1192 DEBUG(dbgs() << "[CompressTaintedStore]\n");
1193 DEBUG(dbgs() << "Original BB\n");
1194 DEBUG(dbgs() << *BB << '\n');
1195 auto* LastSI = AdajacentStores[AdajacentStores.size() - 1];
1196 for (unsigned i = 0; i < AdajacentStores.size() - 1; ++i) {
1197 auto* SI = AdajacentStores[i];
1199 // Use the original address for stores before the last one.
1200 SI->setOperand(1, UntaintedAddress);
1202 DEBUG(dbgs() << "Store address has been reversed: " << *SI << '\n';);
1204 // XXX-comment: Try to make the last store use fewer registers.
1205 // If LastSI's storing value is a select based on the condition with which
1206 // its address is tainted, transform the tainted address to a select
1207 // instruction, as follows:
1208 // r1 = Select Cond ? A : B
1213 // r1 = Select Cond ? A : B
1214 // r2 = Select Cond ? Addr : Addr
1216 // The idea is that both Select instructions depend on the same condition,
1217 // so hopefully the backend can generate two cmov instructions for them (and
1218 // this saves the number of registers needed).
1219 auto* LastSIDep = getConditionDependence(LastSI->getPointerOperand());
1220 auto* LastSIValue = dyn_cast<Instruction>(LastSI->getValueOperand());
1221 if (LastSIValue && LastSIValue->getOpcode() == Instruction::Select &&
1222 LastSIValue->getOperand(0) == LastSIDep) {
1223 // XXX-comment: Maybe it's better for us to just leave it as an and/or
1224 // dependence pattern.
1226 IRBuilder<true, NoFolder> Builder(LastSI);
1228 Builder.CreateSelect(LastSIDep, UntaintedAddress, UntaintedAddress);
1229 LastSI->setOperand(1, Address);
1230 DEBUG(dbgs() << "The last store becomes :" << *LastSI << "\n\n";);
1238 bool PassDependenceToStore(Value* OldAddress, StoreInst* NewStore) {
1239 Value* OldDep = getDependence(OldAddress);
1240 // Return false when there's no dependence to pass from the OldAddress.
1245 // No need to pass the dependence to NewStore's address if it already depends
1246 // on whatever 'OldAddress' depends on.
1247 if (StoreAddressDependOnValue(NewStore, OldDep)) {
1250 return taintStoreAddress(NewStore, OldAddress);
1253 SmallSet<Value*, 8> FindDependence(Value* Val) {
1254 SmallSet<Value*, 8> DepSet;
1255 recursivelyFindDependence(&DepSet, Val, true /*Only insert leaf nodes*/);
1259 bool StoreAddressDependOnValue(StoreInst* SI, Value* DepVal) {
1260 return dependenceSetInclusion(SI->getPointerOperand(), DepVal);
1263 bool StoreDependOnValue(StoreInst* SI, Value* Dep) {
1264 return dependenceSetInclusion(SI, Dep);
1271 bool CodeGenPrepare::runOnFunction(Function &F) {
1272 bool EverMadeChange = false;
1274 if (skipOptnoneFunction(F))
1277 DL = &F.getParent()->getDataLayout();
1279 // Clear per function information.
1280 InsertedInsts.clear();
1281 PromotedInsts.clear();
1285 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
1286 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1287 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1288 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1289 OptSize = F.optForSize();
1291 /// This optimization identifies DIV instructions that can be
1292 /// profitably bypassed and carried out with a shorter, faster divide.
1293 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
1294 const DenseMap<unsigned int, unsigned int> &BypassWidths =
1295 TLI->getBypassSlowDivWidths();
1296 BasicBlock* BB = &*F.begin();
1297 while (BB != nullptr) {
1298 // bypassSlowDivision may create new BBs, but we don't want to reapply the
1299 // optimization to those blocks.
1300 BasicBlock* Next = BB->getNextNode();
1301 EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
1306 // Eliminate blocks that contain only PHI nodes and an
1307 // unconditional branch.
1308 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
1310 // llvm.dbg.value is far away from the value then iSel may not be able
1311 // handle it properly. iSel will drop llvm.dbg.value if it can not
1312 // find a node corresponding to the value.
1313 EverMadeChange |= placeDbgValues(F);
1315 // If there is a mask, compare against zero, and branch that can be combined
1316 // into a single target instruction, push the mask and compare into branch
1317 // users. Do this before OptimizeBlock -> OptimizeInst ->
1318 // OptimizeCmpExpression, which perturbs the pattern being searched for.
1319 if (!DisableBranchOpts) {
1320 EverMadeChange |= sinkAndCmp(F);
1321 EverMadeChange |= splitBranchCondition(F);
1324 bool MadeChange = true;
1325 while (MadeChange) {
1327 for (Function::iterator I = F.begin(); I != F.end(); ) {
1328 BasicBlock *BB = &*I++;
1329 bool ModifiedDTOnIteration = false;
1330 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
1332 // Restart BB iteration if the dominator tree of the Function was changed
1333 if (ModifiedDTOnIteration)
1336 EverMadeChange |= MadeChange;
1341 if (!DisableBranchOpts) {
1343 SmallPtrSet<BasicBlock*, 8> WorkList;
1344 for (BasicBlock &BB : F) {
1345 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
1346 MadeChange |= ConstantFoldTerminator(&BB, true);
1347 if (!MadeChange) continue;
1349 for (SmallVectorImpl<BasicBlock*>::iterator
1350 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
1351 if (pred_begin(*II) == pred_end(*II))
1352 WorkList.insert(*II);
1355 // Delete the dead blocks and any of their dead successors.
1356 MadeChange |= !WorkList.empty();
1357 while (!WorkList.empty()) {
1358 BasicBlock *BB = *WorkList.begin();
1360 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
1362 DeleteDeadBlock(BB);
1364 for (SmallVectorImpl<BasicBlock*>::iterator
1365 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
1366 if (pred_begin(*II) == pred_end(*II))
1367 WorkList.insert(*II);
1370 // Merge pairs of basic blocks with unconditional branches, connected by
1372 if (EverMadeChange || MadeChange)
1373 MadeChange |= eliminateFallThrough(F);
1375 EverMadeChange |= MadeChange;
1378 if (!DisableGCOpts) {
1379 SmallVector<Instruction *, 2> Statepoints;
1380 for (BasicBlock &BB : F)
1381 for (Instruction &I : BB)
1382 if (isStatepoint(I))
1383 Statepoints.push_back(&I);
1384 for (auto &I : Statepoints)
1385 EverMadeChange |= simplifyOffsetableRelocate(*I);
1388 // XXX-comment: Delay dealing with relaxed loads in this function to avoid
1389 // further changes done by other passes (e.g., SimplifyCFG).
1390 // Collect all the relaxed loads.
1391 SmallSet<LoadInst*, 1> MonotonicLoadInsts;
1392 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
1393 if (I->isAtomic()) {
1394 switch (I->getOpcode()) {
1395 case Instruction::Load: {
1396 auto* LI = dyn_cast<LoadInst>(&*I);
1397 if (LI->getOrdering() == Monotonic &&
1398 !LI->getHasSubsequentAcqlRMW()) {
1399 MonotonicLoadInsts.insert(LI);
1410 AddFakeConditionalBranchAfterMonotonicLoads(MonotonicLoadInsts, DT);
1412 return EverMadeChange;
1415 /// Merge basic blocks which are connected by a single edge, where one of the
1416 /// basic blocks has a single successor pointing to the other basic block,
1417 /// which has a single predecessor.
1418 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
1419 bool Changed = false;
1420 // Scan all of the blocks in the function, except for the entry block.
1421 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
1422 BasicBlock *BB = &*I++;
1423 // If the destination block has a single pred, then this is a trivial
1424 // edge, just collapse it.
1425 BasicBlock *SinglePred = BB->getSinglePredecessor();
1427 // Don't merge if BB's address is taken.
1428 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
1430 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
1431 if (Term && !Term->isConditional()) {
1433 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
1434 // Remember if SinglePred was the entry block of the function.
1435 // If so, we will need to move BB back to the entry position.
1436 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
1437 MergeBasicBlockIntoOnlyPred(BB, nullptr);
1439 if (isEntry && BB != &BB->getParent()->getEntryBlock())
1440 BB->moveBefore(&BB->getParent()->getEntryBlock());
1442 // We have erased a block. Update the iterator.
1443 I = BB->getIterator();
1449 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
1450 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
1451 /// edges in ways that are non-optimal for isel. Start by eliminating these
1452 /// blocks so we can split them the way we want them.
1453 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
1454 bool MadeChange = false;
1455 // Note that this intentionally skips the entry block.
1456 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
1457 BasicBlock *BB = &*I++;
1458 // If this block doesn't end with an uncond branch, ignore it.
1459 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
1460 if (!BI || !BI->isUnconditional())
1463 // If the instruction before the branch (skipping debug info) isn't a phi
1464 // node, then other stuff is happening here.
1465 BasicBlock::iterator BBI = BI->getIterator();
1466 if (BBI != BB->begin()) {
1468 while (isa<DbgInfoIntrinsic>(BBI)) {
1469 if (BBI == BB->begin())
1473 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
1477 // Do not break infinite loops.
1478 BasicBlock *DestBB = BI->getSuccessor(0);
1482 if (!canMergeBlocks(BB, DestBB))
1485 eliminateMostlyEmptyBlock(BB);
1491 /// Return true if we can merge BB into DestBB if there is a single
1492 /// unconditional branch between them, and BB contains no other non-phi
1494 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
1495 const BasicBlock *DestBB) const {
1496 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
1497 // the successor. If there are more complex condition (e.g. preheaders),
1498 // don't mess around with them.
1499 BasicBlock::const_iterator BBI = BB->begin();
1500 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
1501 for (const User *U : PN->users()) {
1502 const Instruction *UI = cast<Instruction>(U);
1503 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
1505 // IfUser is inside DestBB block and it is a PHINode then check
1506 // incoming value. If incoming value is not from BB then this is
1507 // a complex condition (e.g. preheaders) we want to avoid here.
1508 if (UI->getParent() == DestBB) {
1509 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
1510 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
1511 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
1512 if (Insn && Insn->getParent() == BB &&
1513 Insn->getParent() != UPN->getIncomingBlock(I))
1520 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
1521 // and DestBB may have conflicting incoming values for the block. If so, we
1522 // can't merge the block.
1523 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
1524 if (!DestBBPN) return true; // no conflict.
1526 // Collect the preds of BB.
1527 SmallPtrSet<const BasicBlock*, 16> BBPreds;
1528 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
1529 // It is faster to get preds from a PHI than with pred_iterator.
1530 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1531 BBPreds.insert(BBPN->getIncomingBlock(i));
1533 BBPreds.insert(pred_begin(BB), pred_end(BB));
1536 // Walk the preds of DestBB.
1537 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
1538 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
1539 if (BBPreds.count(Pred)) { // Common predecessor?
1540 BBI = DestBB->begin();
1541 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
1542 const Value *V1 = PN->getIncomingValueForBlock(Pred);
1543 const Value *V2 = PN->getIncomingValueForBlock(BB);
1545 // If V2 is a phi node in BB, look up what the mapped value will be.
1546 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
1547 if (V2PN->getParent() == BB)
1548 V2 = V2PN->getIncomingValueForBlock(Pred);
1550 // If there is a conflict, bail out.
1551 if (V1 != V2) return false;
1560 /// Eliminate a basic block that has only phi's and an unconditional branch in
1562 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
1563 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1564 BasicBlock *DestBB = BI->getSuccessor(0);
1566 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
1568 // If the destination block has a single pred, then this is a trivial edge,
1569 // just collapse it.
1570 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
1571 if (SinglePred != DestBB) {
1572 // Remember if SinglePred was the entry block of the function. If so, we
1573 // will need to move BB back to the entry position.
1574 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
1575 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
1577 if (isEntry && BB != &BB->getParent()->getEntryBlock())
1578 BB->moveBefore(&BB->getParent()->getEntryBlock());
1580 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1585 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
1586 // to handle the new incoming edges it is about to have.
1588 for (BasicBlock::iterator BBI = DestBB->begin();
1589 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1590 // Remove the incoming value for BB, and remember it.
1591 Value *InVal = PN->removeIncomingValue(BB, false);
1593 // Two options: either the InVal is a phi node defined in BB or it is some
1594 // value that dominates BB.
1595 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
1596 if (InValPhi && InValPhi->getParent() == BB) {
1597 // Add all of the input values of the input PHI as inputs of this phi.
1598 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
1599 PN->addIncoming(InValPhi->getIncomingValue(i),
1600 InValPhi->getIncomingBlock(i));
1602 // Otherwise, add one instance of the dominating value for each edge that
1603 // we will be adding.
1604 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
1605 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1606 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
1608 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
1609 PN->addIncoming(InVal, *PI);
1614 // The PHIs are now updated, change everything that refers to BB to use
1615 // DestBB and remove BB.
1616 BB->replaceAllUsesWith(DestBB);
1617 BB->eraseFromParent();
1620 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1623 // Computes a map of base pointer relocation instructions to corresponding
1624 // derived pointer relocation instructions given a vector of all relocate calls
1625 static void computeBaseDerivedRelocateMap(
1626 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
1627 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
1629 // Collect information in two maps: one primarily for locating the base object
1630 // while filling the second map; the second map is the final structure holding
1631 // a mapping between Base and corresponding Derived relocate calls
1632 DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
1633 for (auto *ThisRelocate : AllRelocateCalls) {
1634 auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
1635 ThisRelocate->getDerivedPtrIndex());
1636 RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
1638 for (auto &Item : RelocateIdxMap) {
1639 std::pair<unsigned, unsigned> Key = Item.first;
1640 if (Key.first == Key.second)
1641 // Base relocation: nothing to insert
1644 GCRelocateInst *I = Item.second;
1645 auto BaseKey = std::make_pair(Key.first, Key.first);
1647 // We're iterating over RelocateIdxMap so we cannot modify it.
1648 auto MaybeBase = RelocateIdxMap.find(BaseKey);
1649 if (MaybeBase == RelocateIdxMap.end())
1650 // TODO: We might want to insert a new base object relocate and gep off
1651 // that, if there are enough derived object relocates.
1654 RelocateInstMap[MaybeBase->second].push_back(I);
1658 // Accepts a GEP and extracts the operands into a vector provided they're all
1659 // small integer constants
1660 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
1661 SmallVectorImpl<Value *> &OffsetV) {
1662 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
1663 // Only accept small constant integer operands
1664 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
1665 if (!Op || Op->getZExtValue() > 20)
1669 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
1670 OffsetV.push_back(GEP->getOperand(i));
1674 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
1675 // replace, computes a replacement, and affects it.
1677 simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
1678 const SmallVectorImpl<GCRelocateInst *> &Targets) {
1679 bool MadeChange = false;
1680 for (GCRelocateInst *ToReplace : Targets) {
1681 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
1682 "Not relocating a derived object of the original base object");
1683 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
1684 // A duplicate relocate call. TODO: coalesce duplicates.
1688 if (RelocatedBase->getParent() != ToReplace->getParent()) {
1689 // Base and derived relocates are in different basic blocks.
1690 // In this case transform is only valid when base dominates derived
1691 // relocate. However it would be too expensive to check dominance
1692 // for each such relocate, so we skip the whole transformation.
1696 Value *Base = ToReplace->getBasePtr();
1697 auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
1698 if (!Derived || Derived->getPointerOperand() != Base)
1701 SmallVector<Value *, 2> OffsetV;
1702 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
1705 // Create a Builder and replace the target callsite with a gep
1706 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
1708 // Insert after RelocatedBase
1709 IRBuilder<> Builder(RelocatedBase->getNextNode());
1710 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1712 // If gc_relocate does not match the actual type, cast it to the right type.
1713 // In theory, there must be a bitcast after gc_relocate if the type does not
1714 // match, and we should reuse it to get the derived pointer. But it could be
1718 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1723 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1727 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
1728 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
1730 // In this case, we can not find the bitcast any more. So we insert a new bitcast
1731 // no matter there is already one or not. In this way, we can handle all cases, and
1732 // the extra bitcast should be optimized away in later passes.
1733 Value *ActualRelocatedBase = RelocatedBase;
1734 if (RelocatedBase->getType() != Base->getType()) {
1735 ActualRelocatedBase =
1736 Builder.CreateBitCast(RelocatedBase, Base->getType());
1738 Value *Replacement = Builder.CreateGEP(
1739 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
1740 Replacement->takeName(ToReplace);
1741 // If the newly generated derived pointer's type does not match the original derived
1742 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
1743 Value *ActualReplacement = Replacement;
1744 if (Replacement->getType() != ToReplace->getType()) {
1746 Builder.CreateBitCast(Replacement, ToReplace->getType());
1748 ToReplace->replaceAllUsesWith(ActualReplacement);
1749 ToReplace->eraseFromParent();
1759 // %ptr = gep %base + 15
1760 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1761 // %base' = relocate(%tok, i32 4, i32 4)
1762 // %ptr' = relocate(%tok, i32 4, i32 5)
1763 // %val = load %ptr'
1768 // %ptr = gep %base + 15
1769 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1770 // %base' = gc.relocate(%tok, i32 4, i32 4)
1771 // %ptr' = gep %base' + 15
1772 // %val = load %ptr'
1773 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
1774 bool MadeChange = false;
1775 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1777 for (auto *U : I.users())
1778 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1779 // Collect all the relocate calls associated with a statepoint
1780 AllRelocateCalls.push_back(Relocate);
1782 // We need atleast one base pointer relocation + one derived pointer
1783 // relocation to mangle
1784 if (AllRelocateCalls.size() < 2)
1787 // RelocateInstMap is a mapping from the base relocate instruction to the
1788 // corresponding derived relocate instructions
1789 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
1790 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1791 if (RelocateInstMap.empty())
1794 for (auto &Item : RelocateInstMap)
1795 // Item.first is the RelocatedBase to offset against
1796 // Item.second is the vector of Targets to replace
1797 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1801 /// SinkCast - Sink the specified cast instruction into its user blocks
1802 static bool SinkCast(CastInst *CI) {
1803 BasicBlock *DefBB = CI->getParent();
1805 /// InsertedCasts - Only insert a cast in each block once.
1806 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1808 bool MadeChange = false;
1809 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1811 Use &TheUse = UI.getUse();
1812 Instruction *User = cast<Instruction>(*UI);
1814 // Figure out which BB this cast is used in. For PHI's this is the
1815 // appropriate predecessor block.
1816 BasicBlock *UserBB = User->getParent();
1817 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1818 UserBB = PN->getIncomingBlock(TheUse);
1821 // Preincrement use iterator so we don't invalidate it.
1824 // If the block selected to receive the cast is an EH pad that does not
1825 // allow non-PHI instructions before the terminator, we can't sink the
1827 if (UserBB->getTerminator()->isEHPad())
1830 // If this user is in the same block as the cast, don't change the cast.
1831 if (UserBB == DefBB) continue;
1833 // If we have already inserted a cast into this block, use it.
1834 CastInst *&InsertedCast = InsertedCasts[UserBB];
1836 if (!InsertedCast) {
1837 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1838 assert(InsertPt != UserBB->end());
1839 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1840 CI->getType(), "", &*InsertPt);
1843 // Replace a use of the cast with a use of the new cast.
1844 TheUse = InsertedCast;
1849 // If we removed all uses, nuke the cast.
1850 if (CI->use_empty()) {
1851 CI->eraseFromParent();
1858 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1859 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1860 /// reduce the number of virtual registers that must be created and coalesced.
1862 /// Return true if any changes are made.
1864 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1865 const DataLayout &DL) {
1866 // If this is a noop copy,
1867 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1868 EVT DstVT = TLI.getValueType(DL, CI->getType());
1870 // This is an fp<->int conversion?
1871 if (SrcVT.isInteger() != DstVT.isInteger())
1874 // If this is an extension, it will be a zero or sign extension, which
1876 if (SrcVT.bitsLT(DstVT)) return false;
1878 // If these values will be promoted, find out what they will be promoted
1879 // to. This helps us consider truncates on PPC as noop copies when they
1881 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1882 TargetLowering::TypePromoteInteger)
1883 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1884 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1885 TargetLowering::TypePromoteInteger)
1886 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1888 // If, after promotion, these are the same types, this is a noop copy.
1892 return SinkCast(CI);
1895 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
1898 /// Return true if any changes were made.
1899 static bool CombineUAddWithOverflow(CmpInst *CI) {
1903 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
1906 Type *Ty = AddI->getType();
1907 if (!isa<IntegerType>(Ty))
1910 // We don't want to move around uses of condition values this late, so we we
1911 // check if it is legal to create the call to the intrinsic in the basic
1912 // block containing the icmp:
1914 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
1918 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
1920 if (AddI->hasOneUse())
1921 assert(*AddI->user_begin() == CI && "expected!");
1924 Module *M = CI->getModule();
1925 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1927 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
1929 auto *UAddWithOverflow =
1930 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
1931 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
1933 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
1935 CI->replaceAllUsesWith(Overflow);
1936 AddI->replaceAllUsesWith(UAdd);
1937 CI->eraseFromParent();
1938 AddI->eraseFromParent();
1942 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1943 /// registers that must be created and coalesced. This is a clear win except on
1944 /// targets with multiple condition code registers (PowerPC), where it might
1945 /// lose; some adjustment may be wanted there.
1947 /// Return true if any changes are made.
1948 static bool SinkCmpExpression(CmpInst *CI) {
1949 BasicBlock *DefBB = CI->getParent();
1951 /// Only insert a cmp in each block once.
1952 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1954 bool MadeChange = false;
1955 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1957 Use &TheUse = UI.getUse();
1958 Instruction *User = cast<Instruction>(*UI);
1960 // Preincrement use iterator so we don't invalidate it.
1963 // Don't bother for PHI nodes.
1964 if (isa<PHINode>(User))
1967 // Figure out which BB this cmp is used in.
1968 BasicBlock *UserBB = User->getParent();
1970 // If this user is in the same block as the cmp, don't change the cmp.
1971 if (UserBB == DefBB) continue;
1973 // If we have already inserted a cmp into this block, use it.
1974 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1977 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1978 assert(InsertPt != UserBB->end());
1980 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1981 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
1984 // Replace a use of the cmp with a use of the new cmp.
1985 TheUse = InsertedCmp;
1990 // If we removed all uses, nuke the cmp.
1991 if (CI->use_empty()) {
1992 CI->eraseFromParent();
1999 static bool OptimizeCmpExpression(CmpInst *CI) {
2000 if (SinkCmpExpression(CI))
2003 if (CombineUAddWithOverflow(CI))
2009 /// Check if the candidates could be combined with a shift instruction, which
2011 /// 1. Truncate instruction
2012 /// 2. And instruction and the imm is a mask of the low bits:
2013 /// imm & (imm+1) == 0
2014 static bool isExtractBitsCandidateUse(Instruction *User) {
2015 if (!isa<TruncInst>(User)) {
2016 if (User->getOpcode() != Instruction::And ||
2017 !isa<ConstantInt>(User->getOperand(1)))
2020 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
2022 if ((Cimm & (Cimm + 1)).getBoolValue())
2028 /// Sink both shift and truncate instruction to the use of truncate's BB.
2030 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
2031 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
2032 const TargetLowering &TLI, const DataLayout &DL) {
2033 BasicBlock *UserBB = User->getParent();
2034 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
2035 TruncInst *TruncI = dyn_cast<TruncInst>(User);
2036 bool MadeChange = false;
2038 for (Value::user_iterator TruncUI = TruncI->user_begin(),
2039 TruncE = TruncI->user_end();
2040 TruncUI != TruncE;) {
2042 Use &TruncTheUse = TruncUI.getUse();
2043 Instruction *TruncUser = cast<Instruction>(*TruncUI);
2044 // Preincrement use iterator so we don't invalidate it.
2048 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
2052 // If the use is actually a legal node, there will not be an
2053 // implicit truncate.
2054 // FIXME: always querying the result type is just an
2055 // approximation; some nodes' legality is determined by the
2056 // operand or other means. There's no good way to find out though.
2057 if (TLI.isOperationLegalOrCustom(
2058 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
2061 // Don't bother for PHI nodes.
2062 if (isa<PHINode>(TruncUser))
2065 BasicBlock *TruncUserBB = TruncUser->getParent();
2067 if (UserBB == TruncUserBB)
2070 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
2071 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
2073 if (!InsertedShift && !InsertedTrunc) {
2074 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
2075 assert(InsertPt != TruncUserBB->end());
2077 if (ShiftI->getOpcode() == Instruction::AShr)
2078 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
2081 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
2085 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
2087 assert(TruncInsertPt != TruncUserBB->end());
2089 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
2090 TruncI->getType(), "", &*TruncInsertPt);
2094 TruncTheUse = InsertedTrunc;
2100 /// Sink the shift *right* instruction into user blocks if the uses could
2101 /// potentially be combined with this shift instruction and generate BitExtract
2102 /// instruction. It will only be applied if the architecture supports BitExtract
2103 /// instruction. Here is an example:
2105 /// %x.extract.shift = lshr i64 %arg1, 32
2107 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
2111 /// %x.extract.shift.1 = lshr i64 %arg1, 32
2112 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
2114 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
2116 /// Return true if any changes are made.
2117 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
2118 const TargetLowering &TLI,
2119 const DataLayout &DL) {
2120 BasicBlock *DefBB = ShiftI->getParent();
2122 /// Only insert instructions in each block once.
2123 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
2125 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
2127 bool MadeChange = false;
2128 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
2130 Use &TheUse = UI.getUse();
2131 Instruction *User = cast<Instruction>(*UI);
2132 // Preincrement use iterator so we don't invalidate it.
2135 // Don't bother for PHI nodes.
2136 if (isa<PHINode>(User))
2139 if (!isExtractBitsCandidateUse(User))
2142 BasicBlock *UserBB = User->getParent();
2144 if (UserBB == DefBB) {
2145 // If the shift and truncate instruction are in the same BB. The use of
2146 // the truncate(TruncUse) may still introduce another truncate if not
2147 // legal. In this case, we would like to sink both shift and truncate
2148 // instruction to the BB of TruncUse.
2151 // i64 shift.result = lshr i64 opnd, imm
2152 // trunc.result = trunc shift.result to i16
2155 // ----> We will have an implicit truncate here if the architecture does
2156 // not have i16 compare.
2157 // cmp i16 trunc.result, opnd2
2159 if (isa<TruncInst>(User) && shiftIsLegal
2160 // If the type of the truncate is legal, no trucate will be
2161 // introduced in other basic blocks.
2163 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
2165 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
2169 // If we have already inserted a shift into this block, use it.
2170 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
2172 if (!InsertedShift) {
2173 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2174 assert(InsertPt != UserBB->end());
2176 if (ShiftI->getOpcode() == Instruction::AShr)
2177 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
2180 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
2186 // Replace a use of the shift with a use of the new shift.
2187 TheUse = InsertedShift;
2190 // If we removed all uses, nuke the shift.
2191 if (ShiftI->use_empty())
2192 ShiftI->eraseFromParent();
2197 // Translate a masked load intrinsic like
2198 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
2199 // <16 x i1> %mask, <16 x i32> %passthru)
2200 // to a chain of basic blocks, with loading element one-by-one if
2201 // the appropriate mask bit is set
2203 // %1 = bitcast i8* %addr to i32*
2204 // %2 = extractelement <16 x i1> %mask, i32 0
2205 // %3 = icmp eq i1 %2, true
2206 // br i1 %3, label %cond.load, label %else
2208 //cond.load: ; preds = %0
2209 // %4 = getelementptr i32* %1, i32 0
2210 // %5 = load i32* %4
2211 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
2214 //else: ; preds = %0, %cond.load
2215 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
2216 // %7 = extractelement <16 x i1> %mask, i32 1
2217 // %8 = icmp eq i1 %7, true
2218 // br i1 %8, label %cond.load1, label %else2
2220 //cond.load1: ; preds = %else
2221 // %9 = getelementptr i32* %1, i32 1
2222 // %10 = load i32* %9
2223 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
2226 //else2: ; preds = %else, %cond.load1
2227 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
2228 // %12 = extractelement <16 x i1> %mask, i32 2
2229 // %13 = icmp eq i1 %12, true
2230 // br i1 %13, label %cond.load4, label %else5
2232 static void ScalarizeMaskedLoad(CallInst *CI) {
2233 Value *Ptr = CI->getArgOperand(0);
2234 Value *Alignment = CI->getArgOperand(1);
2235 Value *Mask = CI->getArgOperand(2);
2236 Value *Src0 = CI->getArgOperand(3);
2238 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2239 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
2240 assert(VecType && "Unexpected return type of masked load intrinsic");
2242 Type *EltTy = CI->getType()->getVectorElementType();
2244 IRBuilder<> Builder(CI->getContext());
2245 Instruction *InsertPt = CI;
2246 BasicBlock *IfBlock = CI->getParent();
2247 BasicBlock *CondBlock = nullptr;
2248 BasicBlock *PrevIfBlock = CI->getParent();
2250 Builder.SetInsertPoint(InsertPt);
2251 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2253 // Short-cut if the mask is all-true.
2254 bool IsAllOnesMask = isa<Constant>(Mask) &&
2255 cast<Constant>(Mask)->isAllOnesValue();
2257 if (IsAllOnesMask) {
2258 Value *NewI = Builder.CreateAlignedLoad(Ptr, AlignVal);
2259 CI->replaceAllUsesWith(NewI);
2260 CI->eraseFromParent();
2264 // Adjust alignment for the scalar instruction.
2265 AlignVal = std::min(AlignVal, VecType->getScalarSizeInBits()/8);
2266 // Bitcast %addr fron i8* to EltTy*
2268 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
2269 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
2270 unsigned VectorWidth = VecType->getNumElements();
2272 Value *UndefVal = UndefValue::get(VecType);
2274 // The result vector
2275 Value *VResult = UndefVal;
2277 if (isa<ConstantVector>(Mask)) {
2278 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2279 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2282 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2283 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
2284 VResult = Builder.CreateInsertElement(VResult, Load,
2285 Builder.getInt32(Idx));
2287 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
2288 CI->replaceAllUsesWith(NewI);
2289 CI->eraseFromParent();
2293 PHINode *Phi = nullptr;
2294 Value *PrevPhi = UndefVal;
2296 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2298 // Fill the "else" block, created in the previous iteration
2300 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
2301 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
2302 // %to_load = icmp eq i1 %mask_1, true
2303 // br i1 %to_load, label %cond.load, label %else
2306 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
2307 Phi->addIncoming(VResult, CondBlock);
2308 Phi->addIncoming(PrevPhi, PrevIfBlock);
2313 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
2314 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2315 ConstantInt::get(Predicate->getType(), 1));
2317 // Create "cond" block
2319 // %EltAddr = getelementptr i32* %1, i32 0
2320 // %Elt = load i32* %EltAddr
2321 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
2323 CondBlock = IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.load");
2324 Builder.SetInsertPoint(InsertPt);
2327 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2328 LoadInst *Load = Builder.CreateAlignedLoad(Gep, AlignVal);
2329 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
2331 // Create "else" block, fill it in the next iteration
2332 BasicBlock *NewIfBlock =
2333 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
2334 Builder.SetInsertPoint(InsertPt);
2335 Instruction *OldBr = IfBlock->getTerminator();
2336 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2337 OldBr->eraseFromParent();
2338 PrevIfBlock = IfBlock;
2339 IfBlock = NewIfBlock;
2342 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
2343 Phi->addIncoming(VResult, CondBlock);
2344 Phi->addIncoming(PrevPhi, PrevIfBlock);
2345 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
2346 CI->replaceAllUsesWith(NewI);
2347 CI->eraseFromParent();
2350 // Translate a masked store intrinsic, like
2351 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
2353 // to a chain of basic blocks, that stores element one-by-one if
2354 // the appropriate mask bit is set
2356 // %1 = bitcast i8* %addr to i32*
2357 // %2 = extractelement <16 x i1> %mask, i32 0
2358 // %3 = icmp eq i1 %2, true
2359 // br i1 %3, label %cond.store, label %else
2361 // cond.store: ; preds = %0
2362 // %4 = extractelement <16 x i32> %val, i32 0
2363 // %5 = getelementptr i32* %1, i32 0
2364 // store i32 %4, i32* %5
2367 // else: ; preds = %0, %cond.store
2368 // %6 = extractelement <16 x i1> %mask, i32 1
2369 // %7 = icmp eq i1 %6, true
2370 // br i1 %7, label %cond.store1, label %else2
2372 // cond.store1: ; preds = %else
2373 // %8 = extractelement <16 x i32> %val, i32 1
2374 // %9 = getelementptr i32* %1, i32 1
2375 // store i32 %8, i32* %9
2378 static void ScalarizeMaskedStore(CallInst *CI) {
2379 Value *Src = CI->getArgOperand(0);
2380 Value *Ptr = CI->getArgOperand(1);
2381 Value *Alignment = CI->getArgOperand(2);
2382 Value *Mask = CI->getArgOperand(3);
2384 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2385 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
2386 assert(VecType && "Unexpected data type in masked store intrinsic");
2388 Type *EltTy = VecType->getElementType();
2390 IRBuilder<> Builder(CI->getContext());
2391 Instruction *InsertPt = CI;
2392 BasicBlock *IfBlock = CI->getParent();
2393 Builder.SetInsertPoint(InsertPt);
2394 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2396 // Short-cut if the mask is all-true.
2397 bool IsAllOnesMask = isa<Constant>(Mask) &&
2398 cast<Constant>(Mask)->isAllOnesValue();
2400 if (IsAllOnesMask) {
2401 Builder.CreateAlignedStore(Src, Ptr, AlignVal);
2402 CI->eraseFromParent();
2406 // Adjust alignment for the scalar instruction.
2407 AlignVal = std::max(AlignVal, VecType->getScalarSizeInBits()/8);
2408 // Bitcast %addr fron i8* to EltTy*
2410 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
2411 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
2412 unsigned VectorWidth = VecType->getNumElements();
2414 if (isa<ConstantVector>(Mask)) {
2415 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2416 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2418 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
2420 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2421 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
2423 CI->eraseFromParent();
2427 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2429 // Fill the "else" block, created in the previous iteration
2431 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
2432 // %to_store = icmp eq i1 %mask_1, true
2433 // br i1 %to_store, label %cond.store, label %else
2435 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
2436 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2437 ConstantInt::get(Predicate->getType(), 1));
2439 // Create "cond" block
2441 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
2442 // %EltAddr = getelementptr i32* %1, i32 0
2443 // %store i32 %OneElt, i32* %EltAddr
2445 BasicBlock *CondBlock =
2446 IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.store");
2447 Builder.SetInsertPoint(InsertPt);
2449 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
2451 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2452 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
2454 // Create "else" block, fill it in the next iteration
2455 BasicBlock *NewIfBlock =
2456 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
2457 Builder.SetInsertPoint(InsertPt);
2458 Instruction *OldBr = IfBlock->getTerminator();
2459 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2460 OldBr->eraseFromParent();
2461 IfBlock = NewIfBlock;
2463 CI->eraseFromParent();
2466 // Translate a masked gather intrinsic like
2467 // <16 x i32 > @llvm.masked.gather.v16i32( <16 x i32*> %Ptrs, i32 4,
2468 // <16 x i1> %Mask, <16 x i32> %Src)
2469 // to a chain of basic blocks, with loading element one-by-one if
2470 // the appropriate mask bit is set
2472 // % Ptrs = getelementptr i32, i32* %base, <16 x i64> %ind
2473 // % Mask0 = extractelement <16 x i1> %Mask, i32 0
2474 // % ToLoad0 = icmp eq i1 % Mask0, true
2475 // br i1 % ToLoad0, label %cond.load, label %else
2478 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
2479 // % Load0 = load i32, i32* % Ptr0, align 4
2480 // % Res0 = insertelement <16 x i32> undef, i32 % Load0, i32 0
2484 // %res.phi.else = phi <16 x i32>[% Res0, %cond.load], [undef, % 0]
2485 // % Mask1 = extractelement <16 x i1> %Mask, i32 1
2486 // % ToLoad1 = icmp eq i1 % Mask1, true
2487 // br i1 % ToLoad1, label %cond.load1, label %else2
2490 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2491 // % Load1 = load i32, i32* % Ptr1, align 4
2492 // % Res1 = insertelement <16 x i32> %res.phi.else, i32 % Load1, i32 1
2495 // % Result = select <16 x i1> %Mask, <16 x i32> %res.phi.select, <16 x i32> %Src
2496 // ret <16 x i32> %Result
2497 static void ScalarizeMaskedGather(CallInst *CI) {
2498 Value *Ptrs = CI->getArgOperand(0);
2499 Value *Alignment = CI->getArgOperand(1);
2500 Value *Mask = CI->getArgOperand(2);
2501 Value *Src0 = CI->getArgOperand(3);
2503 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
2505 assert(VecType && "Unexpected return type of masked load intrinsic");
2507 IRBuilder<> Builder(CI->getContext());
2508 Instruction *InsertPt = CI;
2509 BasicBlock *IfBlock = CI->getParent();
2510 BasicBlock *CondBlock = nullptr;
2511 BasicBlock *PrevIfBlock = CI->getParent();
2512 Builder.SetInsertPoint(InsertPt);
2513 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2515 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2517 Value *UndefVal = UndefValue::get(VecType);
2519 // The result vector
2520 Value *VResult = UndefVal;
2521 unsigned VectorWidth = VecType->getNumElements();
2523 // Shorten the way if the mask is a vector of constants.
2524 bool IsConstMask = isa<ConstantVector>(Mask);
2527 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2528 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2530 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2531 "Ptr" + Twine(Idx));
2532 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
2533 "Load" + Twine(Idx));
2534 VResult = Builder.CreateInsertElement(VResult, Load,
2535 Builder.getInt32(Idx),
2536 "Res" + Twine(Idx));
2538 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
2539 CI->replaceAllUsesWith(NewI);
2540 CI->eraseFromParent();
2544 PHINode *Phi = nullptr;
2545 Value *PrevPhi = UndefVal;
2547 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2549 // Fill the "else" block, created in the previous iteration
2551 // %Mask1 = extractelement <16 x i1> %Mask, i32 1
2552 // %ToLoad1 = icmp eq i1 %Mask1, true
2553 // br i1 %ToLoad1, label %cond.load, label %else
2556 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
2557 Phi->addIncoming(VResult, CondBlock);
2558 Phi->addIncoming(PrevPhi, PrevIfBlock);
2563 Value *Predicate = Builder.CreateExtractElement(Mask,
2564 Builder.getInt32(Idx),
2565 "Mask" + Twine(Idx));
2566 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2567 ConstantInt::get(Predicate->getType(), 1),
2568 "ToLoad" + Twine(Idx));
2570 // Create "cond" block
2572 // %EltAddr = getelementptr i32* %1, i32 0
2573 // %Elt = load i32* %EltAddr
2574 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
2576 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
2577 Builder.SetInsertPoint(InsertPt);
2579 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2580 "Ptr" + Twine(Idx));
2581 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
2582 "Load" + Twine(Idx));
2583 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx),
2584 "Res" + Twine(Idx));
2586 // Create "else" block, fill it in the next iteration
2587 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2588 Builder.SetInsertPoint(InsertPt);
2589 Instruction *OldBr = IfBlock->getTerminator();
2590 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2591 OldBr->eraseFromParent();
2592 PrevIfBlock = IfBlock;
2593 IfBlock = NewIfBlock;
2596 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
2597 Phi->addIncoming(VResult, CondBlock);
2598 Phi->addIncoming(PrevPhi, PrevIfBlock);
2599 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
2600 CI->replaceAllUsesWith(NewI);
2601 CI->eraseFromParent();
2604 // Translate a masked scatter intrinsic, like
2605 // void @llvm.masked.scatter.v16i32(<16 x i32> %Src, <16 x i32*>* %Ptrs, i32 4,
2607 // to a chain of basic blocks, that stores element one-by-one if
2608 // the appropriate mask bit is set.
2610 // % Ptrs = getelementptr i32, i32* %ptr, <16 x i64> %ind
2611 // % Mask0 = extractelement <16 x i1> % Mask, i32 0
2612 // % ToStore0 = icmp eq i1 % Mask0, true
2613 // br i1 %ToStore0, label %cond.store, label %else
2616 // % Elt0 = extractelement <16 x i32> %Src, i32 0
2617 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
2618 // store i32 %Elt0, i32* % Ptr0, align 4
2622 // % Mask1 = extractelement <16 x i1> % Mask, i32 1
2623 // % ToStore1 = icmp eq i1 % Mask1, true
2624 // br i1 % ToStore1, label %cond.store1, label %else2
2627 // % Elt1 = extractelement <16 x i32> %Src, i32 1
2628 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2629 // store i32 % Elt1, i32* % Ptr1, align 4
2632 static void ScalarizeMaskedScatter(CallInst *CI) {
2633 Value *Src = CI->getArgOperand(0);
2634 Value *Ptrs = CI->getArgOperand(1);
2635 Value *Alignment = CI->getArgOperand(2);
2636 Value *Mask = CI->getArgOperand(3);
2638 assert(isa<VectorType>(Src->getType()) &&
2639 "Unexpected data type in masked scatter intrinsic");
2640 assert(isa<VectorType>(Ptrs->getType()) &&
2641 isa<PointerType>(Ptrs->getType()->getVectorElementType()) &&
2642 "Vector of pointers is expected in masked scatter intrinsic");
2644 IRBuilder<> Builder(CI->getContext());
2645 Instruction *InsertPt = CI;
2646 BasicBlock *IfBlock = CI->getParent();
2647 Builder.SetInsertPoint(InsertPt);
2648 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2650 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2651 unsigned VectorWidth = Src->getType()->getVectorNumElements();
2653 // Shorten the way if the mask is a vector of constants.
2654 bool IsConstMask = isa<ConstantVector>(Mask);
2657 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2658 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2660 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
2661 "Elt" + Twine(Idx));
2662 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2663 "Ptr" + Twine(Idx));
2664 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
2666 CI->eraseFromParent();
2669 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2670 // Fill the "else" block, created in the previous iteration
2672 // % Mask1 = extractelement <16 x i1> % Mask, i32 Idx
2673 // % ToStore = icmp eq i1 % Mask1, true
2674 // br i1 % ToStore, label %cond.store, label %else
2676 Value *Predicate = Builder.CreateExtractElement(Mask,
2677 Builder.getInt32(Idx),
2678 "Mask" + Twine(Idx));
2680 Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2681 ConstantInt::get(Predicate->getType(), 1),
2682 "ToStore" + Twine(Idx));
2684 // Create "cond" block
2686 // % Elt1 = extractelement <16 x i32> %Src, i32 1
2687 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2688 // %store i32 % Elt1, i32* % Ptr1
2690 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2691 Builder.SetInsertPoint(InsertPt);
2693 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
2694 "Elt" + Twine(Idx));
2695 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2696 "Ptr" + Twine(Idx));
2697 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
2699 // Create "else" block, fill it in the next iteration
2700 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2701 Builder.SetInsertPoint(InsertPt);
2702 Instruction *OldBr = IfBlock->getTerminator();
2703 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2704 OldBr->eraseFromParent();
2705 IfBlock = NewIfBlock;
2707 CI->eraseFromParent();
2710 /// If counting leading or trailing zeros is an expensive operation and a zero
2711 /// input is defined, add a check for zero to avoid calling the intrinsic.
2713 /// We want to transform:
2714 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
2718 /// %cmpz = icmp eq i64 %A, 0
2719 /// br i1 %cmpz, label %cond.end, label %cond.false
2721 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
2722 /// br label %cond.end
2724 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
2726 /// If the transform is performed, return true and set ModifiedDT to true.
2727 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
2728 const TargetLowering *TLI,
2729 const DataLayout *DL,
2734 // If a zero input is undefined, it doesn't make sense to despeculate that.
2735 if (match(CountZeros->getOperand(1), m_One()))
2738 // If it's cheap to speculate, there's nothing to do.
2739 auto IntrinsicID = CountZeros->getIntrinsicID();
2740 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
2741 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
2744 // Only handle legal scalar cases. Anything else requires too much work.
2745 Type *Ty = CountZeros->getType();
2746 unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
2747 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSize())
2750 // The intrinsic will be sunk behind a compare against zero and branch.
2751 BasicBlock *StartBlock = CountZeros->getParent();
2752 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
2754 // Create another block after the count zero intrinsic. A PHI will be added
2755 // in this block to select the result of the intrinsic or the bit-width
2756 // constant if the input to the intrinsic is zero.
2757 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
2758 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
2760 // Set up a builder to create a compare, conditional branch, and PHI.
2761 IRBuilder<> Builder(CountZeros->getContext());
2762 Builder.SetInsertPoint(StartBlock->getTerminator());
2763 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
2765 // Replace the unconditional branch that was created by the first split with
2766 // a compare against zero and a conditional branch.
2767 Value *Zero = Constant::getNullValue(Ty);
2768 Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
2769 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
2770 StartBlock->getTerminator()->eraseFromParent();
2772 // Create a PHI in the end block to select either the output of the intrinsic
2773 // or the bit width of the operand.
2774 Builder.SetInsertPoint(&EndBlock->front());
2775 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
2776 CountZeros->replaceAllUsesWith(PN);
2777 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
2778 PN->addIncoming(BitWidth, StartBlock);
2779 PN->addIncoming(CountZeros, CallBlock);
2781 // We are explicitly handling the zero case, so we can set the intrinsic's
2782 // undefined zero argument to 'true'. This will also prevent reprocessing the
2783 // intrinsic; we only despeculate when a zero input is defined.
2784 CountZeros->setArgOperand(1, Builder.getTrue());
2789 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool& ModifiedDT) {
2790 BasicBlock *BB = CI->getParent();
2792 // Lower inline assembly if we can.
2793 // If we found an inline asm expession, and if the target knows how to
2794 // lower it to normal LLVM code, do so now.
2795 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
2796 if (TLI->ExpandInlineAsm(CI)) {
2797 // Avoid invalidating the iterator.
2798 CurInstIterator = BB->begin();
2799 // Avoid processing instructions out of order, which could cause
2800 // reuse before a value is defined.
2804 // Sink address computing for memory operands into the block.
2805 if (optimizeInlineAsmInst(CI))
2809 // Align the pointer arguments to this call if the target thinks it's a good
2811 unsigned MinSize, PrefAlign;
2812 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
2813 for (auto &Arg : CI->arg_operands()) {
2814 // We want to align both objects whose address is used directly and
2815 // objects whose address is used in casts and GEPs, though it only makes
2816 // sense for GEPs if the offset is a multiple of the desired alignment and
2817 // if size - offset meets the size threshold.
2818 if (!Arg->getType()->isPointerTy())
2820 APInt Offset(DL->getPointerSizeInBits(
2821 cast<PointerType>(Arg->getType())->getAddressSpace()),
2823 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
2824 uint64_t Offset2 = Offset.getLimitedValue();
2825 if ((Offset2 & (PrefAlign-1)) != 0)
2828 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
2829 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
2830 AI->setAlignment(PrefAlign);
2831 // Global variables can only be aligned if they are defined in this
2832 // object (i.e. they are uniquely initialized in this object), and
2833 // over-aligning global variables that have an explicit section is
2836 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
2837 GV->getAlignment() < PrefAlign &&
2838 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
2840 GV->setAlignment(PrefAlign);
2842 // If this is a memcpy (or similar) then we may be able to improve the
2844 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
2845 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
2846 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
2847 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
2848 if (Align > MI->getAlignment())
2849 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
2853 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
2855 switch (II->getIntrinsicID()) {
2857 case Intrinsic::objectsize: {
2858 // Lower all uses of llvm.objectsize.*
2859 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
2860 Type *ReturnTy = CI->getType();
2861 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
2863 // Substituting this can cause recursive simplifications, which can
2864 // invalidate our iterator. Use a WeakVH to hold onto it in case this
2866 WeakVH IterHandle(&*CurInstIterator);
2868 replaceAndRecursivelySimplify(CI, RetVal,
2871 // If the iterator instruction was recursively deleted, start over at the
2872 // start of the block.
2873 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
2874 CurInstIterator = BB->begin();
2879 case Intrinsic::masked_load: {
2880 // Scalarize unsupported vector masked load
2881 if (!TTI->isLegalMaskedLoad(CI->getType())) {
2882 ScalarizeMaskedLoad(CI);
2888 case Intrinsic::masked_store: {
2889 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType())) {
2890 ScalarizeMaskedStore(CI);
2896 case Intrinsic::masked_gather: {
2897 if (!TTI->isLegalMaskedGather(CI->getType())) {
2898 ScalarizeMaskedGather(CI);
2904 case Intrinsic::masked_scatter: {
2905 if (!TTI->isLegalMaskedScatter(CI->getArgOperand(0)->getType())) {
2906 ScalarizeMaskedScatter(CI);
2912 case Intrinsic::aarch64_stlxr:
2913 case Intrinsic::aarch64_stxr: {
2914 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
2915 if (!ExtVal || !ExtVal->hasOneUse() ||
2916 ExtVal->getParent() == CI->getParent())
2918 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
2919 ExtVal->moveBefore(CI);
2920 // Mark this instruction as "inserted by CGP", so that other
2921 // optimizations don't touch it.
2922 InsertedInsts.insert(ExtVal);
2925 case Intrinsic::invariant_group_barrier:
2926 II->replaceAllUsesWith(II->getArgOperand(0));
2927 II->eraseFromParent();
2930 case Intrinsic::cttz:
2931 case Intrinsic::ctlz:
2932 // If counting zeros is expensive, try to avoid it.
2933 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
2937 // Unknown address space.
2938 // TODO: Target hook to pick which address space the intrinsic cares
2940 unsigned AddrSpace = ~0u;
2941 SmallVector<Value*, 2> PtrOps;
2943 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
2944 while (!PtrOps.empty())
2945 if (optimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
2950 // From here on out we're working with named functions.
2951 if (!CI->getCalledFunction()) return false;
2953 // Lower all default uses of _chk calls. This is very similar
2954 // to what InstCombineCalls does, but here we are only lowering calls
2955 // to fortified library functions (e.g. __memcpy_chk) that have the default
2956 // "don't know" as the objectsize. Anything else should be left alone.
2957 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
2958 if (Value *V = Simplifier.optimizeCall(CI)) {
2959 CI->replaceAllUsesWith(V);
2960 CI->eraseFromParent();
2966 /// Look for opportunities to duplicate return instructions to the predecessor
2967 /// to enable tail call optimizations. The case it is currently looking for is:
2970 /// %tmp0 = tail call i32 @f0()
2971 /// br label %return
2973 /// %tmp1 = tail call i32 @f1()
2974 /// br label %return
2976 /// %tmp2 = tail call i32 @f2()
2977 /// br label %return
2979 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
2987 /// %tmp0 = tail call i32 @f0()
2990 /// %tmp1 = tail call i32 @f1()
2993 /// %tmp2 = tail call i32 @f2()
2996 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
3000 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
3004 PHINode *PN = nullptr;
3005 BitCastInst *BCI = nullptr;
3006 Value *V = RI->getReturnValue();
3008 BCI = dyn_cast<BitCastInst>(V);
3010 V = BCI->getOperand(0);
3012 PN = dyn_cast<PHINode>(V);
3017 if (PN && PN->getParent() != BB)
3020 // It's not safe to eliminate the sign / zero extension of the return value.
3021 // See llvm::isInTailCallPosition().
3022 const Function *F = BB->getParent();
3023 AttributeSet CallerAttrs = F->getAttributes();
3024 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
3025 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
3028 // Make sure there are no instructions between the PHI and return, or that the
3029 // return is the first instruction in the block.
3031 BasicBlock::iterator BI = BB->begin();
3032 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
3034 // Also skip over the bitcast.
3039 BasicBlock::iterator BI = BB->begin();
3040 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
3045 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
3047 SmallVector<CallInst*, 4> TailCalls;
3049 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
3050 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
3051 // Make sure the phi value is indeed produced by the tail call.
3052 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
3053 TLI->mayBeEmittedAsTailCall(CI))
3054 TailCalls.push_back(CI);
3057 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
3058 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
3059 if (!VisitedBBs.insert(*PI).second)
3062 BasicBlock::InstListType &InstList = (*PI)->getInstList();
3063 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
3064 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
3065 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
3069 CallInst *CI = dyn_cast<CallInst>(&*RI);
3070 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
3071 TailCalls.push_back(CI);
3075 bool Changed = false;
3076 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
3077 CallInst *CI = TailCalls[i];
3080 // Conservatively require the attributes of the call to match those of the
3081 // return. Ignore noalias because it doesn't affect the call sequence.
3082 AttributeSet CalleeAttrs = CS.getAttributes();
3083 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
3084 removeAttribute(Attribute::NoAlias) !=
3085 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
3086 removeAttribute(Attribute::NoAlias))
3089 // Make sure the call instruction is followed by an unconditional branch to
3090 // the return block.
3091 BasicBlock *CallBB = CI->getParent();
3092 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
3093 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
3096 // Duplicate the return into CallBB.
3097 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
3098 ModifiedDT = Changed = true;
3102 // If we eliminated all predecessors of the block, delete the block now.
3103 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
3104 BB->eraseFromParent();
3109 //===----------------------------------------------------------------------===//
3110 // Memory Optimization
3111 //===----------------------------------------------------------------------===//
3115 /// This is an extended version of TargetLowering::AddrMode
3116 /// which holds actual Value*'s for register values.
3117 struct ExtAddrMode : public TargetLowering::AddrMode {
3120 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
3121 void print(raw_ostream &OS) const;
3124 bool operator==(const ExtAddrMode& O) const {
3125 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
3126 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
3127 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
3132 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
3138 void ExtAddrMode::print(raw_ostream &OS) const {
3139 bool NeedPlus = false;
3142 OS << (NeedPlus ? " + " : "")
3144 BaseGV->printAsOperand(OS, /*PrintType=*/false);
3149 OS << (NeedPlus ? " + " : "")
3155 OS << (NeedPlus ? " + " : "")
3157 BaseReg->printAsOperand(OS, /*PrintType=*/false);
3161 OS << (NeedPlus ? " + " : "")
3163 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
3169 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3170 void ExtAddrMode::dump() const {
3176 /// \brief This class provides transaction based operation on the IR.
3177 /// Every change made through this class is recorded in the internal state and
3178 /// can be undone (rollback) until commit is called.
3179 class TypePromotionTransaction {
3181 /// \brief This represents the common interface of the individual transaction.
3182 /// Each class implements the logic for doing one specific modification on
3183 /// the IR via the TypePromotionTransaction.
3184 class TypePromotionAction {
3186 /// The Instruction modified.
3190 /// \brief Constructor of the action.
3191 /// The constructor performs the related action on the IR.
3192 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
3194 virtual ~TypePromotionAction() {}
3196 /// \brief Undo the modification done by this action.
3197 /// When this method is called, the IR must be in the same state as it was
3198 /// before this action was applied.
3199 /// \pre Undoing the action works if and only if the IR is in the exact same
3200 /// state as it was directly after this action was applied.
3201 virtual void undo() = 0;
3203 /// \brief Advocate every change made by this action.
3204 /// When the results on the IR of the action are to be kept, it is important
3205 /// to call this function, otherwise hidden information may be kept forever.
3206 virtual void commit() {
3207 // Nothing to be done, this action is not doing anything.
3211 /// \brief Utility to remember the position of an instruction.
3212 class InsertionHandler {
3213 /// Position of an instruction.
3214 /// Either an instruction:
3215 /// - Is the first in a basic block: BB is used.
3216 /// - Has a previous instructon: PrevInst is used.
3218 Instruction *PrevInst;
3221 /// Remember whether or not the instruction had a previous instruction.
3222 bool HasPrevInstruction;
3225 /// \brief Record the position of \p Inst.
3226 InsertionHandler(Instruction *Inst) {
3227 BasicBlock::iterator It = Inst->getIterator();
3228 HasPrevInstruction = (It != (Inst->getParent()->begin()));
3229 if (HasPrevInstruction)
3230 Point.PrevInst = &*--It;
3232 Point.BB = Inst->getParent();
3235 /// \brief Insert \p Inst at the recorded position.
3236 void insert(Instruction *Inst) {
3237 if (HasPrevInstruction) {
3238 if (Inst->getParent())
3239 Inst->removeFromParent();
3240 Inst->insertAfter(Point.PrevInst);
3242 Instruction *Position = &*Point.BB->getFirstInsertionPt();
3243 if (Inst->getParent())
3244 Inst->moveBefore(Position);
3246 Inst->insertBefore(Position);
3251 /// \brief Move an instruction before another.
3252 class InstructionMoveBefore : public TypePromotionAction {
3253 /// Original position of the instruction.
3254 InsertionHandler Position;
3257 /// \brief Move \p Inst before \p Before.
3258 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
3259 : TypePromotionAction(Inst), Position(Inst) {
3260 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
3261 Inst->moveBefore(Before);
3264 /// \brief Move the instruction back to its original position.
3265 void undo() override {
3266 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
3267 Position.insert(Inst);
3271 /// \brief Set the operand of an instruction with a new value.
3272 class OperandSetter : public TypePromotionAction {
3273 /// Original operand of the instruction.
3275 /// Index of the modified instruction.
3279 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
3280 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
3281 : TypePromotionAction(Inst), Idx(Idx) {
3282 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
3283 << "for:" << *Inst << "\n"
3284 << "with:" << *NewVal << "\n");
3285 Origin = Inst->getOperand(Idx);
3286 Inst->setOperand(Idx, NewVal);
3289 /// \brief Restore the original value of the instruction.
3290 void undo() override {
3291 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
3292 << "for: " << *Inst << "\n"
3293 << "with: " << *Origin << "\n");
3294 Inst->setOperand(Idx, Origin);
3298 /// \brief Hide the operands of an instruction.
3299 /// Do as if this instruction was not using any of its operands.
3300 class OperandsHider : public TypePromotionAction {
3301 /// The list of original operands.
3302 SmallVector<Value *, 4> OriginalValues;
3305 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
3306 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
3307 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
3308 unsigned NumOpnds = Inst->getNumOperands();
3309 OriginalValues.reserve(NumOpnds);
3310 for (unsigned It = 0; It < NumOpnds; ++It) {
3311 // Save the current operand.
3312 Value *Val = Inst->getOperand(It);
3313 OriginalValues.push_back(Val);
3315 // We could use OperandSetter here, but that would imply an overhead
3316 // that we are not willing to pay.
3317 Inst->setOperand(It, UndefValue::get(Val->getType()));
3321 /// \brief Restore the original list of uses.
3322 void undo() override {
3323 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
3324 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
3325 Inst->setOperand(It, OriginalValues[It]);
3329 /// \brief Build a truncate instruction.
3330 class TruncBuilder : public TypePromotionAction {
3333 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
3335 /// trunc Opnd to Ty.
3336 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
3337 IRBuilder<> Builder(Opnd);
3338 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
3339 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
3342 /// \brief Get the built value.
3343 Value *getBuiltValue() { return Val; }
3345 /// \brief Remove the built instruction.
3346 void undo() override {
3347 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
3348 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3349 IVal->eraseFromParent();
3353 /// \brief Build a sign extension instruction.
3354 class SExtBuilder : public TypePromotionAction {
3357 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
3359 /// sext Opnd to Ty.
3360 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3361 : TypePromotionAction(InsertPt) {
3362 IRBuilder<> Builder(InsertPt);
3363 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
3364 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
3367 /// \brief Get the built value.
3368 Value *getBuiltValue() { return Val; }
3370 /// \brief Remove the built instruction.
3371 void undo() override {
3372 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
3373 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3374 IVal->eraseFromParent();
3378 /// \brief Build a zero extension instruction.
3379 class ZExtBuilder : public TypePromotionAction {
3382 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
3384 /// zext Opnd to Ty.
3385 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3386 : TypePromotionAction(InsertPt) {
3387 IRBuilder<> Builder(InsertPt);
3388 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
3389 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
3392 /// \brief Get the built value.
3393 Value *getBuiltValue() { return Val; }
3395 /// \brief Remove the built instruction.
3396 void undo() override {
3397 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
3398 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3399 IVal->eraseFromParent();
3403 /// \brief Mutate an instruction to another type.
3404 class TypeMutator : public TypePromotionAction {
3405 /// Record the original type.
3409 /// \brief Mutate the type of \p Inst into \p NewTy.
3410 TypeMutator(Instruction *Inst, Type *NewTy)
3411 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
3412 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
3414 Inst->mutateType(NewTy);
3417 /// \brief Mutate the instruction back to its original type.
3418 void undo() override {
3419 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
3421 Inst->mutateType(OrigTy);
3425 /// \brief Replace the uses of an instruction by another instruction.
3426 class UsesReplacer : public TypePromotionAction {
3427 /// Helper structure to keep track of the replaced uses.
3428 struct InstructionAndIdx {
3429 /// The instruction using the instruction.
3431 /// The index where this instruction is used for Inst.
3433 InstructionAndIdx(Instruction *Inst, unsigned Idx)
3434 : Inst(Inst), Idx(Idx) {}
3437 /// Keep track of the original uses (pair Instruction, Index).
3438 SmallVector<InstructionAndIdx, 4> OriginalUses;
3439 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
3442 /// \brief Replace all the use of \p Inst by \p New.
3443 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
3444 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
3446 // Record the original uses.
3447 for (Use &U : Inst->uses()) {
3448 Instruction *UserI = cast<Instruction>(U.getUser());
3449 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
3451 // Now, we can replace the uses.
3452 Inst->replaceAllUsesWith(New);
3455 /// \brief Reassign the original uses of Inst to Inst.
3456 void undo() override {
3457 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
3458 for (use_iterator UseIt = OriginalUses.begin(),
3459 EndIt = OriginalUses.end();
3460 UseIt != EndIt; ++UseIt) {
3461 UseIt->Inst->setOperand(UseIt->Idx, Inst);
3466 /// \brief Remove an instruction from the IR.
3467 class InstructionRemover : public TypePromotionAction {
3468 /// Original position of the instruction.
3469 InsertionHandler Inserter;
3470 /// Helper structure to hide all the link to the instruction. In other
3471 /// words, this helps to do as if the instruction was removed.
3472 OperandsHider Hider;
3473 /// Keep track of the uses replaced, if any.
3474 UsesReplacer *Replacer;
3477 /// \brief Remove all reference of \p Inst and optinally replace all its
3479 /// \pre If !Inst->use_empty(), then New != nullptr
3480 InstructionRemover(Instruction *Inst, Value *New = nullptr)
3481 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
3484 Replacer = new UsesReplacer(Inst, New);
3485 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
3486 Inst->removeFromParent();
3489 ~InstructionRemover() override { delete Replacer; }
3491 /// \brief Really remove the instruction.
3492 void commit() override { delete Inst; }
3494 /// \brief Resurrect the instruction and reassign it to the proper uses if
3495 /// new value was provided when build this action.
3496 void undo() override {
3497 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
3498 Inserter.insert(Inst);
3506 /// Restoration point.
3507 /// The restoration point is a pointer to an action instead of an iterator
3508 /// because the iterator may be invalidated but not the pointer.
3509 typedef const TypePromotionAction *ConstRestorationPt;
3510 /// Advocate every changes made in that transaction.
3512 /// Undo all the changes made after the given point.
3513 void rollback(ConstRestorationPt Point);
3514 /// Get the current restoration point.
3515 ConstRestorationPt getRestorationPoint() const;
3517 /// \name API for IR modification with state keeping to support rollback.
3519 /// Same as Instruction::setOperand.
3520 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
3521 /// Same as Instruction::eraseFromParent.
3522 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
3523 /// Same as Value::replaceAllUsesWith.
3524 void replaceAllUsesWith(Instruction *Inst, Value *New);
3525 /// Same as Value::mutateType.
3526 void mutateType(Instruction *Inst, Type *NewTy);
3527 /// Same as IRBuilder::createTrunc.
3528 Value *createTrunc(Instruction *Opnd, Type *Ty);
3529 /// Same as IRBuilder::createSExt.
3530 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
3531 /// Same as IRBuilder::createZExt.
3532 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
3533 /// Same as Instruction::moveBefore.
3534 void moveBefore(Instruction *Inst, Instruction *Before);
3538 /// The ordered list of actions made so far.
3539 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
3540 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
3543 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
3546 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
3549 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
3552 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
3555 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
3557 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
3560 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
3561 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
3564 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
3566 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
3567 Value *Val = Ptr->getBuiltValue();
3568 Actions.push_back(std::move(Ptr));
3572 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
3573 Value *Opnd, Type *Ty) {
3574 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
3575 Value *Val = Ptr->getBuiltValue();
3576 Actions.push_back(std::move(Ptr));
3580 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
3581 Value *Opnd, Type *Ty) {
3582 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
3583 Value *Val = Ptr->getBuiltValue();
3584 Actions.push_back(std::move(Ptr));
3588 void TypePromotionTransaction::moveBefore(Instruction *Inst,
3589 Instruction *Before) {
3591 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
3594 TypePromotionTransaction::ConstRestorationPt
3595 TypePromotionTransaction::getRestorationPoint() const {
3596 return !Actions.empty() ? Actions.back().get() : nullptr;
3599 void TypePromotionTransaction::commit() {
3600 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
3606 void TypePromotionTransaction::rollback(
3607 TypePromotionTransaction::ConstRestorationPt Point) {
3608 while (!Actions.empty() && Point != Actions.back().get()) {
3609 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
3614 /// \brief A helper class for matching addressing modes.
3616 /// This encapsulates the logic for matching the target-legal addressing modes.
3617 class AddressingModeMatcher {
3618 SmallVectorImpl<Instruction*> &AddrModeInsts;
3619 const TargetMachine &TM;
3620 const TargetLowering &TLI;
3621 const DataLayout &DL;
3623 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
3624 /// the memory instruction that we're computing this address for.
3627 Instruction *MemoryInst;
3629 /// This is the addressing mode that we're building up. This is
3630 /// part of the return value of this addressing mode matching stuff.
3631 ExtAddrMode &AddrMode;
3633 /// The instructions inserted by other CodeGenPrepare optimizations.
3634 const SetOfInstrs &InsertedInsts;
3635 /// A map from the instructions to their type before promotion.
3636 InstrToOrigTy &PromotedInsts;
3637 /// The ongoing transaction where every action should be registered.
3638 TypePromotionTransaction &TPT;
3640 /// This is set to true when we should not do profitability checks.
3641 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
3642 bool IgnoreProfitability;
3644 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
3645 const TargetMachine &TM, Type *AT, unsigned AS,
3646 Instruction *MI, ExtAddrMode &AM,
3647 const SetOfInstrs &InsertedInsts,
3648 InstrToOrigTy &PromotedInsts,
3649 TypePromotionTransaction &TPT)
3650 : AddrModeInsts(AMI), TM(TM),
3651 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
3652 ->getTargetLowering()),
3653 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
3654 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
3655 PromotedInsts(PromotedInsts), TPT(TPT) {
3656 IgnoreProfitability = false;
3660 /// Find the maximal addressing mode that a load/store of V can fold,
3661 /// give an access type of AccessTy. This returns a list of involved
3662 /// instructions in AddrModeInsts.
3663 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
3665 /// \p PromotedInsts maps the instructions to their type before promotion.
3666 /// \p The ongoing transaction where every action should be registered.
3667 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
3668 Instruction *MemoryInst,
3669 SmallVectorImpl<Instruction*> &AddrModeInsts,
3670 const TargetMachine &TM,
3671 const SetOfInstrs &InsertedInsts,
3672 InstrToOrigTy &PromotedInsts,
3673 TypePromotionTransaction &TPT) {
3676 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
3677 MemoryInst, Result, InsertedInsts,
3678 PromotedInsts, TPT).matchAddr(V, 0);
3679 (void)Success; assert(Success && "Couldn't select *anything*?");
3683 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
3684 bool matchAddr(Value *V, unsigned Depth);
3685 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
3686 bool *MovedAway = nullptr);
3687 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
3688 ExtAddrMode &AMBefore,
3689 ExtAddrMode &AMAfter);
3690 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
3691 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
3692 Value *PromotedOperand) const;
3695 /// Try adding ScaleReg*Scale to the current addressing mode.
3696 /// Return true and update AddrMode if this addr mode is legal for the target,
3698 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3700 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3701 // mode. Just process that directly.
3703 return matchAddr(ScaleReg, Depth);
3705 // If the scale is 0, it takes nothing to add this.
3709 // If we already have a scale of this value, we can add to it, otherwise, we
3710 // need an available scale field.
3711 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3714 ExtAddrMode TestAddrMode = AddrMode;
3716 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3717 // [A+B + A*7] -> [B+A*8].
3718 TestAddrMode.Scale += Scale;
3719 TestAddrMode.ScaledReg = ScaleReg;
3721 // If the new address isn't legal, bail out.
3722 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3725 // It was legal, so commit it.
3726 AddrMode = TestAddrMode;
3728 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3729 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3730 // X*Scale + C*Scale to addr mode.
3731 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3732 if (isa<Instruction>(ScaleReg) && // not a constant expr.
3733 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3734 TestAddrMode.ScaledReg = AddLHS;
3735 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3737 // If this addressing mode is legal, commit it and remember that we folded
3738 // this instruction.
3739 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3740 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3741 AddrMode = TestAddrMode;
3746 // Otherwise, not (x+c)*scale, just return what we have.
3750 /// This is a little filter, which returns true if an addressing computation
3751 /// involving I might be folded into a load/store accessing it.
3752 /// This doesn't need to be perfect, but needs to accept at least
3753 /// the set of instructions that MatchOperationAddr can.
3754 static bool MightBeFoldableInst(Instruction *I) {
3755 switch (I->getOpcode()) {
3756 case Instruction::BitCast:
3757 case Instruction::AddrSpaceCast:
3758 // Don't touch identity bitcasts.
3759 if (I->getType() == I->getOperand(0)->getType())
3761 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
3762 case Instruction::PtrToInt:
3763 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3765 case Instruction::IntToPtr:
3766 // We know the input is intptr_t, so this is foldable.
3768 case Instruction::Add:
3770 case Instruction::Mul:
3771 case Instruction::Shl:
3772 // Can only handle X*C and X << C.
3773 return isa<ConstantInt>(I->getOperand(1));
3774 case Instruction::GetElementPtr:
3781 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
3782 /// \note \p Val is assumed to be the product of some type promotion.
3783 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3784 /// to be legal, as the non-promoted value would have had the same state.
3785 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
3786 const DataLayout &DL, Value *Val) {
3787 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3790 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3791 // If the ISDOpcode is undefined, it was undefined before the promotion.
3794 // Otherwise, check if the promoted instruction is legal or not.
3795 return TLI.isOperationLegalOrCustom(
3796 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3799 /// \brief Hepler class to perform type promotion.
3800 class TypePromotionHelper {
3801 /// \brief Utility function to check whether or not a sign or zero extension
3802 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3803 /// either using the operands of \p Inst or promoting \p Inst.
3804 /// The type of the extension is defined by \p IsSExt.
3805 /// In other words, check if:
3806 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3807 /// #1 Promotion applies:
3808 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3809 /// #2 Operand reuses:
3810 /// ext opnd1 to ConsideredExtType.
3811 /// \p PromotedInsts maps the instructions to their type before promotion.
3812 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3813 const InstrToOrigTy &PromotedInsts, bool IsSExt);
3815 /// \brief Utility function to determine if \p OpIdx should be promoted when
3816 /// promoting \p Inst.
3817 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3818 return !(isa<SelectInst>(Inst) && OpIdx == 0);
3821 /// \brief Utility function to promote the operand of \p Ext when this
3822 /// operand is a promotable trunc or sext or zext.
3823 /// \p PromotedInsts maps the instructions to their type before promotion.
3824 /// \p CreatedInstsCost[out] contains the cost of all instructions
3825 /// created to promote the operand of Ext.
3826 /// Newly added extensions are inserted in \p Exts.
3827 /// Newly added truncates are inserted in \p Truncs.
3828 /// Should never be called directly.
3829 /// \return The promoted value which is used instead of Ext.
3830 static Value *promoteOperandForTruncAndAnyExt(
3831 Instruction *Ext, TypePromotionTransaction &TPT,
3832 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3833 SmallVectorImpl<Instruction *> *Exts,
3834 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3836 /// \brief Utility function to promote the operand of \p Ext when this
3837 /// operand is promotable and is not a supported trunc or sext.
3838 /// \p PromotedInsts maps the instructions to their type before promotion.
3839 /// \p CreatedInstsCost[out] contains the cost of all the instructions
3840 /// created to promote the operand of Ext.
3841 /// Newly added extensions are inserted in \p Exts.
3842 /// Newly added truncates are inserted in \p Truncs.
3843 /// Should never be called directly.
3844 /// \return The promoted value which is used instead of Ext.
3845 static Value *promoteOperandForOther(Instruction *Ext,
3846 TypePromotionTransaction &TPT,
3847 InstrToOrigTy &PromotedInsts,
3848 unsigned &CreatedInstsCost,
3849 SmallVectorImpl<Instruction *> *Exts,
3850 SmallVectorImpl<Instruction *> *Truncs,
3851 const TargetLowering &TLI, bool IsSExt);
3853 /// \see promoteOperandForOther.
3854 static Value *signExtendOperandForOther(
3855 Instruction *Ext, TypePromotionTransaction &TPT,
3856 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3857 SmallVectorImpl<Instruction *> *Exts,
3858 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3859 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3860 Exts, Truncs, TLI, true);
3863 /// \see promoteOperandForOther.
3864 static Value *zeroExtendOperandForOther(
3865 Instruction *Ext, TypePromotionTransaction &TPT,
3866 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3867 SmallVectorImpl<Instruction *> *Exts,
3868 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3869 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3870 Exts, Truncs, TLI, false);
3874 /// Type for the utility function that promotes the operand of Ext.
3875 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
3876 InstrToOrigTy &PromotedInsts,
3877 unsigned &CreatedInstsCost,
3878 SmallVectorImpl<Instruction *> *Exts,
3879 SmallVectorImpl<Instruction *> *Truncs,
3880 const TargetLowering &TLI);
3881 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
3882 /// action to promote the operand of \p Ext instead of using Ext.
3883 /// \return NULL if no promotable action is possible with the current
3885 /// \p InsertedInsts keeps track of all the instructions inserted by the
3886 /// other CodeGenPrepare optimizations. This information is important
3887 /// because we do not want to promote these instructions as CodeGenPrepare
3888 /// will reinsert them later. Thus creating an infinite loop: create/remove.
3889 /// \p PromotedInsts maps the instructions to their type before promotion.
3890 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3891 const TargetLowering &TLI,
3892 const InstrToOrigTy &PromotedInsts);
3895 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3896 Type *ConsideredExtType,
3897 const InstrToOrigTy &PromotedInsts,
3899 // The promotion helper does not know how to deal with vector types yet.
3900 // To be able to fix that, we would need to fix the places where we
3901 // statically extend, e.g., constants and such.
3902 if (Inst->getType()->isVectorTy())
3905 // We can always get through zext.
3906 if (isa<ZExtInst>(Inst))
3909 // sext(sext) is ok too.
3910 if (IsSExt && isa<SExtInst>(Inst))
3913 // We can get through binary operator, if it is legal. In other words, the
3914 // binary operator must have a nuw or nsw flag.
3915 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3916 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3917 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3918 (IsSExt && BinOp->hasNoSignedWrap())))
3921 // Check if we can do the following simplification.
3922 // ext(trunc(opnd)) --> ext(opnd)
3923 if (!isa<TruncInst>(Inst))
3926 Value *OpndVal = Inst->getOperand(0);
3927 // Check if we can use this operand in the extension.
3928 // If the type is larger than the result type of the extension, we cannot.
3929 if (!OpndVal->getType()->isIntegerTy() ||
3930 OpndVal->getType()->getIntegerBitWidth() >
3931 ConsideredExtType->getIntegerBitWidth())
3934 // If the operand of the truncate is not an instruction, we will not have
3935 // any information on the dropped bits.
3936 // (Actually we could for constant but it is not worth the extra logic).
3937 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3941 // Check if the source of the type is narrow enough.
3942 // I.e., check that trunc just drops extended bits of the same kind of
3944 // #1 get the type of the operand and check the kind of the extended bits.
3945 const Type *OpndType;
3946 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3947 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
3948 OpndType = It->second.getPointer();
3949 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3950 OpndType = Opnd->getOperand(0)->getType();
3954 // #2 check that the truncate just drops extended bits.
3955 return Inst->getType()->getIntegerBitWidth() >=
3956 OpndType->getIntegerBitWidth();
3959 TypePromotionHelper::Action TypePromotionHelper::getAction(
3960 Instruction *Ext, const SetOfInstrs &InsertedInsts,
3961 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3962 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
3963 "Unexpected instruction type");
3964 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3965 Type *ExtTy = Ext->getType();
3966 bool IsSExt = isa<SExtInst>(Ext);
3967 // If the operand of the extension is not an instruction, we cannot
3969 // If it, check we can get through.
3970 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3973 // Do not promote if the operand has been added by codegenprepare.
3974 // Otherwise, it means we are undoing an optimization that is likely to be
3975 // redone, thus causing potential infinite loop.
3976 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3979 // SExt or Trunc instructions.
3980 // Return the related handler.
3981 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3982 isa<ZExtInst>(ExtOpnd))
3983 return promoteOperandForTruncAndAnyExt;
3985 // Regular instruction.
3986 // Abort early if we will have to insert non-free instructions.
3987 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3989 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3992 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3993 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
3994 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3995 SmallVectorImpl<Instruction *> *Exts,
3996 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3997 // By construction, the operand of SExt is an instruction. Otherwise we cannot
3998 // get through it and this method should not be called.
3999 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
4000 Value *ExtVal = SExt;
4001 bool HasMergedNonFreeExt = false;
4002 if (isa<ZExtInst>(SExtOpnd)) {
4003 // Replace s|zext(zext(opnd))
4005 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
4007 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
4008 TPT.replaceAllUsesWith(SExt, ZExt);
4009 TPT.eraseInstruction(SExt);
4012 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
4014 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
4016 CreatedInstsCost = 0;
4018 // Remove dead code.
4019 if (SExtOpnd->use_empty())
4020 TPT.eraseInstruction(SExtOpnd);
4022 // Check if the extension is still needed.
4023 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
4024 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
4027 Exts->push_back(ExtInst);
4028 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
4033 // At this point we have: ext ty opnd to ty.
4034 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
4035 Value *NextVal = ExtInst->getOperand(0);
4036 TPT.eraseInstruction(ExtInst, NextVal);
4040 Value *TypePromotionHelper::promoteOperandForOther(
4041 Instruction *Ext, TypePromotionTransaction &TPT,
4042 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4043 SmallVectorImpl<Instruction *> *Exts,
4044 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
4046 // By construction, the operand of Ext is an instruction. Otherwise we cannot
4047 // get through it and this method should not be called.
4048 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
4049 CreatedInstsCost = 0;
4050 if (!ExtOpnd->hasOneUse()) {
4051 // ExtOpnd will be promoted.
4052 // All its uses, but Ext, will need to use a truncated value of the
4053 // promoted version.
4054 // Create the truncate now.
4055 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
4056 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
4057 ITrunc->removeFromParent();
4058 // Insert it just after the definition.
4059 ITrunc->insertAfter(ExtOpnd);
4061 Truncs->push_back(ITrunc);
4064 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
4065 // Restore the operand of Ext (which has been replaced by the previous call
4066 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
4067 TPT.setOperand(Ext, 0, ExtOpnd);
4070 // Get through the Instruction:
4071 // 1. Update its type.
4072 // 2. Replace the uses of Ext by Inst.
4073 // 3. Extend each operand that needs to be extended.
4075 // Remember the original type of the instruction before promotion.
4076 // This is useful to know that the high bits are sign extended bits.
4077 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
4078 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
4080 TPT.mutateType(ExtOpnd, Ext->getType());
4082 TPT.replaceAllUsesWith(Ext, ExtOpnd);
4084 Instruction *ExtForOpnd = Ext;
4086 DEBUG(dbgs() << "Propagate Ext to operands\n");
4087 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
4089 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
4090 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
4091 !shouldExtOperand(ExtOpnd, OpIdx)) {
4092 DEBUG(dbgs() << "No need to propagate\n");
4095 // Check if we can statically extend the operand.
4096 Value *Opnd = ExtOpnd->getOperand(OpIdx);
4097 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
4098 DEBUG(dbgs() << "Statically extend\n");
4099 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
4100 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
4101 : Cst->getValue().zext(BitWidth);
4102 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
4105 // UndefValue are typed, so we have to statically sign extend them.
4106 if (isa<UndefValue>(Opnd)) {
4107 DEBUG(dbgs() << "Statically extend\n");
4108 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
4112 // Otherwise we have to explicity sign extend the operand.
4113 // Check if Ext was reused to extend an operand.
4115 // If yes, create a new one.
4116 DEBUG(dbgs() << "More operands to ext\n");
4117 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
4118 : TPT.createZExt(Ext, Opnd, Ext->getType());
4119 if (!isa<Instruction>(ValForExtOpnd)) {
4120 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
4123 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
4126 Exts->push_back(ExtForOpnd);
4127 TPT.setOperand(ExtForOpnd, 0, Opnd);
4129 // Move the sign extension before the insertion point.
4130 TPT.moveBefore(ExtForOpnd, ExtOpnd);
4131 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
4132 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
4133 // If more sext are required, new instructions will have to be created.
4134 ExtForOpnd = nullptr;
4136 if (ExtForOpnd == Ext) {
4137 DEBUG(dbgs() << "Extension is useless now\n");
4138 TPT.eraseInstruction(Ext);
4143 /// Check whether or not promoting an instruction to a wider type is profitable.
4144 /// \p NewCost gives the cost of extension instructions created by the
4146 /// \p OldCost gives the cost of extension instructions before the promotion
4147 /// plus the number of instructions that have been
4148 /// matched in the addressing mode the promotion.
4149 /// \p PromotedOperand is the value that has been promoted.
4150 /// \return True if the promotion is profitable, false otherwise.
4151 bool AddressingModeMatcher::isPromotionProfitable(
4152 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
4153 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
4154 // The cost of the new extensions is greater than the cost of the
4155 // old extension plus what we folded.
4156 // This is not profitable.
4157 if (NewCost > OldCost)
4159 if (NewCost < OldCost)
4161 // The promotion is neutral but it may help folding the sign extension in
4162 // loads for instance.
4163 // Check that we did not create an illegal instruction.
4164 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
4167 /// Given an instruction or constant expr, see if we can fold the operation
4168 /// into the addressing mode. If so, update the addressing mode and return
4169 /// true, otherwise return false without modifying AddrMode.
4170 /// If \p MovedAway is not NULL, it contains the information of whether or
4171 /// not AddrInst has to be folded into the addressing mode on success.
4172 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
4173 /// because it has been moved away.
4174 /// Thus AddrInst must not be added in the matched instructions.
4175 /// This state can happen when AddrInst is a sext, since it may be moved away.
4176 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
4177 /// not be referenced anymore.
4178 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
4181 // Avoid exponential behavior on extremely deep expression trees.
4182 if (Depth >= 5) return false;
4184 // By default, all matched instructions stay in place.
4189 case Instruction::PtrToInt:
4190 // PtrToInt is always a noop, as we know that the int type is pointer sized.
4191 return matchAddr(AddrInst->getOperand(0), Depth);
4192 case Instruction::IntToPtr: {
4193 auto AS = AddrInst->getType()->getPointerAddressSpace();
4194 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
4195 // This inttoptr is a no-op if the integer type is pointer sized.
4196 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
4197 return matchAddr(AddrInst->getOperand(0), Depth);
4200 case Instruction::BitCast:
4201 // BitCast is always a noop, and we can handle it as long as it is
4202 // int->int or pointer->pointer (we don't want int<->fp or something).
4203 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
4204 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
4205 // Don't touch identity bitcasts. These were probably put here by LSR,
4206 // and we don't want to mess around with them. Assume it knows what it
4208 AddrInst->getOperand(0)->getType() != AddrInst->getType())
4209 return matchAddr(AddrInst->getOperand(0), Depth);
4211 case Instruction::AddrSpaceCast: {
4213 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
4214 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
4215 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
4216 return matchAddr(AddrInst->getOperand(0), Depth);
4219 case Instruction::Add: {
4220 // Check to see if we can merge in the RHS then the LHS. If so, we win.
4221 ExtAddrMode BackupAddrMode = AddrMode;
4222 unsigned OldSize = AddrModeInsts.size();
4223 // Start a transaction at this point.
4224 // The LHS may match but not the RHS.
4225 // Therefore, we need a higher level restoration point to undo partially
4226 // matched operation.
4227 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4228 TPT.getRestorationPoint();
4230 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
4231 matchAddr(AddrInst->getOperand(0), Depth+1))
4234 // Restore the old addr mode info.
4235 AddrMode = BackupAddrMode;
4236 AddrModeInsts.resize(OldSize);
4237 TPT.rollback(LastKnownGood);
4239 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
4240 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
4241 matchAddr(AddrInst->getOperand(1), Depth+1))
4244 // Otherwise we definitely can't merge the ADD in.
4245 AddrMode = BackupAddrMode;
4246 AddrModeInsts.resize(OldSize);
4247 TPT.rollback(LastKnownGood);
4250 //case Instruction::Or:
4251 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
4253 case Instruction::Mul:
4254 case Instruction::Shl: {
4255 // Can only handle X*C and X << C.
4256 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
4259 int64_t Scale = RHS->getSExtValue();
4260 if (Opcode == Instruction::Shl)
4261 Scale = 1LL << Scale;
4263 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
4265 case Instruction::GetElementPtr: {
4266 // Scan the GEP. We check it if it contains constant offsets and at most
4267 // one variable offset.
4268 int VariableOperand = -1;
4269 unsigned VariableScale = 0;
4271 int64_t ConstantOffset = 0;
4272 gep_type_iterator GTI = gep_type_begin(AddrInst);
4273 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
4274 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
4275 const StructLayout *SL = DL.getStructLayout(STy);
4277 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
4278 ConstantOffset += SL->getElementOffset(Idx);
4280 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
4281 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
4282 ConstantOffset += CI->getSExtValue()*TypeSize;
4283 } else if (TypeSize) { // Scales of zero don't do anything.
4284 // We only allow one variable index at the moment.
4285 if (VariableOperand != -1)
4288 // Remember the variable index.
4289 VariableOperand = i;
4290 VariableScale = TypeSize;
4295 // A common case is for the GEP to only do a constant offset. In this case,
4296 // just add it to the disp field and check validity.
4297 if (VariableOperand == -1) {
4298 AddrMode.BaseOffs += ConstantOffset;
4299 if (ConstantOffset == 0 ||
4300 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
4301 // Check to see if we can fold the base pointer in too.
4302 if (matchAddr(AddrInst->getOperand(0), Depth+1))
4305 AddrMode.BaseOffs -= ConstantOffset;
4309 // Save the valid addressing mode in case we can't match.
4310 ExtAddrMode BackupAddrMode = AddrMode;
4311 unsigned OldSize = AddrModeInsts.size();
4313 // See if the scale and offset amount is valid for this target.
4314 AddrMode.BaseOffs += ConstantOffset;
4316 // Match the base operand of the GEP.
4317 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
4318 // If it couldn't be matched, just stuff the value in a register.
4319 if (AddrMode.HasBaseReg) {
4320 AddrMode = BackupAddrMode;
4321 AddrModeInsts.resize(OldSize);
4324 AddrMode.HasBaseReg = true;
4325 AddrMode.BaseReg = AddrInst->getOperand(0);
4328 // Match the remaining variable portion of the GEP.
4329 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
4331 // If it couldn't be matched, try stuffing the base into a register
4332 // instead of matching it, and retrying the match of the scale.
4333 AddrMode = BackupAddrMode;
4334 AddrModeInsts.resize(OldSize);
4335 if (AddrMode.HasBaseReg)
4337 AddrMode.HasBaseReg = true;
4338 AddrMode.BaseReg = AddrInst->getOperand(0);
4339 AddrMode.BaseOffs += ConstantOffset;
4340 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
4341 VariableScale, Depth)) {
4342 // If even that didn't work, bail.
4343 AddrMode = BackupAddrMode;
4344 AddrModeInsts.resize(OldSize);
4351 case Instruction::SExt:
4352 case Instruction::ZExt: {
4353 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
4357 // Try to move this ext out of the way of the addressing mode.
4358 // Ask for a method for doing so.
4359 TypePromotionHelper::Action TPH =
4360 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
4364 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4365 TPT.getRestorationPoint();
4366 unsigned CreatedInstsCost = 0;
4367 unsigned ExtCost = !TLI.isExtFree(Ext);
4368 Value *PromotedOperand =
4369 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
4370 // SExt has been moved away.
4371 // Thus either it will be rematched later in the recursive calls or it is
4372 // gone. Anyway, we must not fold it into the addressing mode at this point.
4376 // addr = gep base, idx
4378 // promotedOpnd = ext opnd <- no match here
4379 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
4380 // addr = gep base, op <- match
4384 assert(PromotedOperand &&
4385 "TypePromotionHelper should have filtered out those cases");
4387 ExtAddrMode BackupAddrMode = AddrMode;
4388 unsigned OldSize = AddrModeInsts.size();
4390 if (!matchAddr(PromotedOperand, Depth) ||
4391 // The total of the new cost is equal to the cost of the created
4393 // The total of the old cost is equal to the cost of the extension plus
4394 // what we have saved in the addressing mode.
4395 !isPromotionProfitable(CreatedInstsCost,
4396 ExtCost + (AddrModeInsts.size() - OldSize),
4398 AddrMode = BackupAddrMode;
4399 AddrModeInsts.resize(OldSize);
4400 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
4401 TPT.rollback(LastKnownGood);
4410 /// If we can, try to add the value of 'Addr' into the current addressing mode.
4411 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
4412 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
4415 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
4416 // Start a transaction at this point that we will rollback if the matching
4418 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4419 TPT.getRestorationPoint();
4420 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
4421 // Fold in immediates if legal for the target.
4422 AddrMode.BaseOffs += CI->getSExtValue();
4423 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4425 AddrMode.BaseOffs -= CI->getSExtValue();
4426 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
4427 // If this is a global variable, try to fold it into the addressing mode.
4428 if (!AddrMode.BaseGV) {
4429 AddrMode.BaseGV = GV;
4430 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4432 AddrMode.BaseGV = nullptr;
4434 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
4435 ExtAddrMode BackupAddrMode = AddrMode;
4436 unsigned OldSize = AddrModeInsts.size();
4438 // Check to see if it is possible to fold this operation.
4439 bool MovedAway = false;
4440 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
4441 // This instruction may have been moved away. If so, there is nothing
4445 // Okay, it's possible to fold this. Check to see if it is actually
4446 // *profitable* to do so. We use a simple cost model to avoid increasing
4447 // register pressure too much.
4448 if (I->hasOneUse() ||
4449 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
4450 AddrModeInsts.push_back(I);
4454 // It isn't profitable to do this, roll back.
4455 //cerr << "NOT FOLDING: " << *I;
4456 AddrMode = BackupAddrMode;
4457 AddrModeInsts.resize(OldSize);
4458 TPT.rollback(LastKnownGood);
4460 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
4461 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
4463 TPT.rollback(LastKnownGood);
4464 } else if (isa<ConstantPointerNull>(Addr)) {
4465 // Null pointer gets folded without affecting the addressing mode.
4469 // Worse case, the target should support [reg] addressing modes. :)
4470 if (!AddrMode.HasBaseReg) {
4471 AddrMode.HasBaseReg = true;
4472 AddrMode.BaseReg = Addr;
4473 // Still check for legality in case the target supports [imm] but not [i+r].
4474 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4476 AddrMode.HasBaseReg = false;
4477 AddrMode.BaseReg = nullptr;
4480 // If the base register is already taken, see if we can do [r+r].
4481 if (AddrMode.Scale == 0) {
4483 AddrMode.ScaledReg = Addr;
4484 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4487 AddrMode.ScaledReg = nullptr;
4490 TPT.rollback(LastKnownGood);
4494 /// Check to see if all uses of OpVal by the specified inline asm call are due
4495 /// to memory operands. If so, return true, otherwise return false.
4496 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
4497 const TargetMachine &TM) {
4498 const Function *F = CI->getParent()->getParent();
4499 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
4500 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
4501 TargetLowering::AsmOperandInfoVector TargetConstraints =
4502 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
4503 ImmutableCallSite(CI));
4504 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4505 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4507 // Compute the constraint code and ConstraintType to use.
4508 TLI->ComputeConstraintToUse(OpInfo, SDValue());
4510 // If this asm operand is our Value*, and if it isn't an indirect memory
4511 // operand, we can't fold it!
4512 if (OpInfo.CallOperandVal == OpVal &&
4513 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
4514 !OpInfo.isIndirect))
4521 /// Recursively walk all the uses of I until we find a memory use.
4522 /// If we find an obviously non-foldable instruction, return true.
4523 /// Add the ultimately found memory instructions to MemoryUses.
4524 static bool FindAllMemoryUses(
4526 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
4527 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
4528 // If we already considered this instruction, we're done.
4529 if (!ConsideredInsts.insert(I).second)
4532 // If this is an obviously unfoldable instruction, bail out.
4533 if (!MightBeFoldableInst(I))
4536 // Loop over all the uses, recursively processing them.
4537 for (Use &U : I->uses()) {
4538 Instruction *UserI = cast<Instruction>(U.getUser());
4540 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4541 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
4545 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4546 unsigned opNo = U.getOperandNo();
4547 if (opNo == 0) return true; // Storing addr, not into addr.
4548 MemoryUses.push_back(std::make_pair(SI, opNo));
4552 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4553 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
4554 if (!IA) return true;
4556 // If this is a memory operand, we're cool, otherwise bail out.
4557 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
4562 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
4569 /// Return true if Val is already known to be live at the use site that we're
4570 /// folding it into. If so, there is no cost to include it in the addressing
4571 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4572 /// instruction already.
4573 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4574 Value *KnownLive2) {
4575 // If Val is either of the known-live values, we know it is live!
4576 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4579 // All values other than instructions and arguments (e.g. constants) are live.
4580 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4582 // If Val is a constant sized alloca in the entry block, it is live, this is
4583 // true because it is just a reference to the stack/frame pointer, which is
4584 // live for the whole function.
4585 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4586 if (AI->isStaticAlloca())
4589 // Check to see if this value is already used in the memory instruction's
4590 // block. If so, it's already live into the block at the very least, so we
4591 // can reasonably fold it.
4592 return Val->isUsedInBasicBlock(MemoryInst->getParent());
4595 /// It is possible for the addressing mode of the machine to fold the specified
4596 /// instruction into a load or store that ultimately uses it.
4597 /// However, the specified instruction has multiple uses.
4598 /// Given this, it may actually increase register pressure to fold it
4599 /// into the load. For example, consider this code:
4603 /// use(Y) -> nonload/store
4607 /// In this case, Y has multiple uses, and can be folded into the load of Z
4608 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4609 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4610 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4611 /// number of computations either.
4613 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4614 /// X was live across 'load Z' for other reasons, we actually *would* want to
4615 /// fold the addressing mode in the Z case. This would make Y die earlier.
4616 bool AddressingModeMatcher::
4617 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4618 ExtAddrMode &AMAfter) {
4619 if (IgnoreProfitability) return true;
4621 // AMBefore is the addressing mode before this instruction was folded into it,
4622 // and AMAfter is the addressing mode after the instruction was folded. Get
4623 // the set of registers referenced by AMAfter and subtract out those
4624 // referenced by AMBefore: this is the set of values which folding in this
4625 // address extends the lifetime of.
4627 // Note that there are only two potential values being referenced here,
4628 // BaseReg and ScaleReg (global addresses are always available, as are any
4629 // folded immediates).
4630 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4632 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4633 // lifetime wasn't extended by adding this instruction.
4634 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4636 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4637 ScaledReg = nullptr;
4639 // If folding this instruction (and it's subexprs) didn't extend any live
4640 // ranges, we're ok with it.
4641 if (!BaseReg && !ScaledReg)
4644 // If all uses of this instruction are ultimately load/store/inlineasm's,
4645 // check to see if their addressing modes will include this instruction. If
4646 // so, we can fold it into all uses, so it doesn't matter if it has multiple
4648 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
4649 SmallPtrSet<Instruction*, 16> ConsideredInsts;
4650 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
4651 return false; // Has a non-memory, non-foldable use!
4653 // Now that we know that all uses of this instruction are part of a chain of
4654 // computation involving only operations that could theoretically be folded
4655 // into a memory use, loop over each of these uses and see if they could
4656 // *actually* fold the instruction.
4657 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4658 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4659 Instruction *User = MemoryUses[i].first;
4660 unsigned OpNo = MemoryUses[i].second;
4662 // Get the access type of this use. If the use isn't a pointer, we don't
4663 // know what it accesses.
4664 Value *Address = User->getOperand(OpNo);
4665 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4668 Type *AddressAccessTy = AddrTy->getElementType();
4669 unsigned AS = AddrTy->getAddressSpace();
4671 // Do a match against the root of this address, ignoring profitability. This
4672 // will tell us if the addressing mode for the memory operation will
4673 // *actually* cover the shared instruction.
4675 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4676 TPT.getRestorationPoint();
4677 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
4678 MemoryInst, Result, InsertedInsts,
4679 PromotedInsts, TPT);
4680 Matcher.IgnoreProfitability = true;
4681 bool Success = Matcher.matchAddr(Address, 0);
4682 (void)Success; assert(Success && "Couldn't select *anything*?");
4684 // The match was to check the profitability, the changes made are not
4685 // part of the original matcher. Therefore, they should be dropped
4686 // otherwise the original matcher will not present the right state.
4687 TPT.rollback(LastKnownGood);
4689 // If the match didn't cover I, then it won't be shared by it.
4690 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
4691 I) == MatchedAddrModeInsts.end())
4694 MatchedAddrModeInsts.clear();
4700 } // end anonymous namespace
4702 /// Return true if the specified values are defined in a
4703 /// different basic block than BB.
4704 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4705 if (Instruction *I = dyn_cast<Instruction>(V))
4706 return I->getParent() != BB;
4710 /// Load and Store Instructions often have addressing modes that can do
4711 /// significant amounts of computation. As such, instruction selection will try
4712 /// to get the load or store to do as much computation as possible for the
4713 /// program. The problem is that isel can only see within a single block. As
4714 /// such, we sink as much legal addressing mode work into the block as possible.
4716 /// This method is used to optimize both load/store and inline asms with memory
4718 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4719 Type *AccessTy, unsigned AddrSpace) {
4722 // Try to collapse single-value PHI nodes. This is necessary to undo
4723 // unprofitable PRE transformations.
4724 SmallVector<Value*, 8> worklist;
4725 SmallPtrSet<Value*, 16> Visited;
4726 worklist.push_back(Addr);
4728 // Use a worklist to iteratively look through PHI nodes, and ensure that
4729 // the addressing mode obtained from the non-PHI roots of the graph
4731 Value *Consensus = nullptr;
4732 unsigned NumUsesConsensus = 0;
4733 bool IsNumUsesConsensusValid = false;
4734 SmallVector<Instruction*, 16> AddrModeInsts;
4735 ExtAddrMode AddrMode;
4736 TypePromotionTransaction TPT;
4737 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4738 TPT.getRestorationPoint();
4739 while (!worklist.empty()) {
4740 Value *V = worklist.back();
4741 worklist.pop_back();
4743 // Break use-def graph loops.
4744 if (!Visited.insert(V).second) {
4745 Consensus = nullptr;
4749 // For a PHI node, push all of its incoming values.
4750 if (PHINode *P = dyn_cast<PHINode>(V)) {
4751 for (Value *IncValue : P->incoming_values())
4752 worklist.push_back(IncValue);
4756 // For non-PHIs, determine the addressing mode being computed.
4757 SmallVector<Instruction*, 16> NewAddrModeInsts;
4758 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4759 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
4760 InsertedInsts, PromotedInsts, TPT);
4762 // This check is broken into two cases with very similar code to avoid using
4763 // getNumUses() as much as possible. Some values have a lot of uses, so
4764 // calling getNumUses() unconditionally caused a significant compile-time
4768 AddrMode = NewAddrMode;
4769 AddrModeInsts = NewAddrModeInsts;
4771 } else if (NewAddrMode == AddrMode) {
4772 if (!IsNumUsesConsensusValid) {
4773 NumUsesConsensus = Consensus->getNumUses();
4774 IsNumUsesConsensusValid = true;
4777 // Ensure that the obtained addressing mode is equivalent to that obtained
4778 // for all other roots of the PHI traversal. Also, when choosing one
4779 // such root as representative, select the one with the most uses in order
4780 // to keep the cost modeling heuristics in AddressingModeMatcher
4782 unsigned NumUses = V->getNumUses();
4783 if (NumUses > NumUsesConsensus) {
4785 NumUsesConsensus = NumUses;
4786 AddrModeInsts = NewAddrModeInsts;
4791 Consensus = nullptr;
4795 // If the addressing mode couldn't be determined, or if multiple different
4796 // ones were determined, bail out now.
4798 TPT.rollback(LastKnownGood);
4803 // Check to see if any of the instructions supersumed by this addr mode are
4804 // non-local to I's BB.
4805 bool AnyNonLocal = false;
4806 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
4807 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
4813 // If all the instructions matched are already in this BB, don't do anything.
4815 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
4819 // Insert this computation right after this user. Since our caller is
4820 // scanning from the top of the BB to the bottom, reuse of the expr are
4821 // guaranteed to happen later.
4822 IRBuilder<> Builder(MemoryInst);
4824 // Now that we determined the addressing expression we want to use and know
4825 // that we have to sink it into this block. Check to see if we have already
4826 // done this for some other load/store instr in this block. If so, reuse the
4828 Value *&SunkAddr = SunkAddrs[Addr];
4830 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
4831 << *MemoryInst << "\n");
4832 if (SunkAddr->getType() != Addr->getType())
4833 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
4834 } else if (AddrSinkUsingGEPs ||
4835 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
4836 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
4838 // By default, we use the GEP-based method when AA is used later. This
4839 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4840 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4841 << *MemoryInst << "\n");
4842 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4843 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4845 // First, find the pointer.
4846 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4847 ResultPtr = AddrMode.BaseReg;
4848 AddrMode.BaseReg = nullptr;
4851 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4852 // We can't add more than one pointer together, nor can we scale a
4853 // pointer (both of which seem meaningless).
4854 if (ResultPtr || AddrMode.Scale != 1)
4857 ResultPtr = AddrMode.ScaledReg;
4861 if (AddrMode.BaseGV) {
4865 ResultPtr = AddrMode.BaseGV;
4868 // If the real base value actually came from an inttoptr, then the matcher
4869 // will look through it and provide only the integer value. In that case,
4871 if (!ResultPtr && AddrMode.BaseReg) {
4873 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
4874 AddrMode.BaseReg = nullptr;
4875 } else if (!ResultPtr && AddrMode.Scale == 1) {
4877 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
4882 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4883 SunkAddr = Constant::getNullValue(Addr->getType());
4884 } else if (!ResultPtr) {
4888 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4889 Type *I8Ty = Builder.getInt8Ty();
4891 // Start with the base register. Do this first so that subsequent address
4892 // matching finds it last, which will prevent it from trying to match it
4893 // as the scaled value in case it happens to be a mul. That would be
4894 // problematic if we've sunk a different mul for the scale, because then
4895 // we'd end up sinking both muls.
4896 if (AddrMode.BaseReg) {
4897 Value *V = AddrMode.BaseReg;
4898 if (V->getType() != IntPtrTy)
4899 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4904 // Add the scale value.
4905 if (AddrMode.Scale) {
4906 Value *V = AddrMode.ScaledReg;
4907 if (V->getType() == IntPtrTy) {
4909 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4910 cast<IntegerType>(V->getType())->getBitWidth()) {
4911 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4913 // It is only safe to sign extend the BaseReg if we know that the math
4914 // required to create it did not overflow before we extend it. Since
4915 // the original IR value was tossed in favor of a constant back when
4916 // the AddrMode was created we need to bail out gracefully if widths
4917 // do not match instead of extending it.
4918 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
4919 if (I && (ResultIndex != AddrMode.BaseReg))
4920 I->eraseFromParent();
4924 if (AddrMode.Scale != 1)
4925 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4928 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4933 // Add in the Base Offset if present.
4934 if (AddrMode.BaseOffs) {
4935 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4937 // We need to add this separately from the scale above to help with
4938 // SDAG consecutive load/store merging.
4939 if (ResultPtr->getType() != I8PtrTy)
4940 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
4941 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4948 SunkAddr = ResultPtr;
4950 if (ResultPtr->getType() != I8PtrTy)
4951 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
4952 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4955 if (SunkAddr->getType() != Addr->getType())
4956 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
4959 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4960 << *MemoryInst << "\n");
4961 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4962 Value *Result = nullptr;
4964 // Start with the base register. Do this first so that subsequent address
4965 // matching finds it last, which will prevent it from trying to match it
4966 // as the scaled value in case it happens to be a mul. That would be
4967 // problematic if we've sunk a different mul for the scale, because then
4968 // we'd end up sinking both muls.
4969 if (AddrMode.BaseReg) {
4970 Value *V = AddrMode.BaseReg;
4971 if (V->getType()->isPointerTy())
4972 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4973 if (V->getType() != IntPtrTy)
4974 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4978 // Add the scale value.
4979 if (AddrMode.Scale) {
4980 Value *V = AddrMode.ScaledReg;
4981 if (V->getType() == IntPtrTy) {
4983 } else if (V->getType()->isPointerTy()) {
4984 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4985 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4986 cast<IntegerType>(V->getType())->getBitWidth()) {
4987 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4989 // It is only safe to sign extend the BaseReg if we know that the math
4990 // required to create it did not overflow before we extend it. Since
4991 // the original IR value was tossed in favor of a constant back when
4992 // the AddrMode was created we need to bail out gracefully if widths
4993 // do not match instead of extending it.
4994 Instruction *I = dyn_cast_or_null<Instruction>(Result);
4995 if (I && (Result != AddrMode.BaseReg))
4996 I->eraseFromParent();
4999 if (AddrMode.Scale != 1)
5000 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
5003 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5008 // Add in the BaseGV if present.
5009 if (AddrMode.BaseGV) {
5010 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
5012 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5017 // Add in the Base Offset if present.
5018 if (AddrMode.BaseOffs) {
5019 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
5021 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5027 SunkAddr = Constant::getNullValue(Addr->getType());
5029 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
5032 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
5034 // If we have no uses, recursively delete the value and all dead instructions
5036 if (Repl->use_empty()) {
5037 // This can cause recursive deletion, which can invalidate our iterator.
5038 // Use a WeakVH to hold onto it in case this happens.
5039 WeakVH IterHandle(&*CurInstIterator);
5040 BasicBlock *BB = CurInstIterator->getParent();
5042 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
5044 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
5045 // If the iterator instruction was recursively deleted, start over at the
5046 // start of the block.
5047 CurInstIterator = BB->begin();
5055 /// If there are any memory operands, use OptimizeMemoryInst to sink their
5056 /// address computing into the block when possible / profitable.
5057 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
5058 bool MadeChange = false;
5060 const TargetRegisterInfo *TRI =
5061 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
5062 TargetLowering::AsmOperandInfoVector TargetConstraints =
5063 TLI->ParseConstraints(*DL, TRI, CS);
5065 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5066 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
5068 // Compute the constraint code and ConstraintType to use.
5069 TLI->ComputeConstraintToUse(OpInfo, SDValue());
5071 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5072 OpInfo.isIndirect) {
5073 Value *OpVal = CS->getArgOperand(ArgNo++);
5074 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
5075 } else if (OpInfo.Type == InlineAsm::isInput)
5082 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
5083 /// sign extensions.
5084 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
5085 assert(!Inst->use_empty() && "Input must have at least one use");
5086 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
5087 bool IsSExt = isa<SExtInst>(FirstUser);
5088 Type *ExtTy = FirstUser->getType();
5089 for (const User *U : Inst->users()) {
5090 const Instruction *UI = cast<Instruction>(U);
5091 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
5093 Type *CurTy = UI->getType();
5094 // Same input and output types: Same instruction after CSE.
5098 // If IsSExt is true, we are in this situation:
5100 // b = sext ty1 a to ty2
5101 // c = sext ty1 a to ty3
5102 // Assuming ty2 is shorter than ty3, this could be turned into:
5104 // b = sext ty1 a to ty2
5105 // c = sext ty2 b to ty3
5106 // However, the last sext is not free.
5110 // This is a ZExt, maybe this is free to extend from one type to another.
5111 // In that case, we would not account for a different use.
5114 if (ExtTy->getScalarType()->getIntegerBitWidth() >
5115 CurTy->getScalarType()->getIntegerBitWidth()) {
5123 if (!TLI.isZExtFree(NarrowTy, LargeTy))
5126 // All uses are the same or can be derived from one another for free.
5130 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
5131 /// load instruction.
5132 /// If an ext(load) can be formed, it is returned via \p LI for the load
5133 /// and \p Inst for the extension.
5134 /// Otherwise LI == nullptr and Inst == nullptr.
5135 /// When some promotion happened, \p TPT contains the proper state to
5138 /// \return true when promoting was necessary to expose the ext(load)
5139 /// opportunity, false otherwise.
5143 /// %ld = load i32* %addr
5144 /// %add = add nuw i32 %ld, 4
5145 /// %zext = zext i32 %add to i64
5149 /// %ld = load i32* %addr
5150 /// %zext = zext i32 %ld to i64
5151 /// %add = add nuw i64 %zext, 4
5153 /// Thanks to the promotion, we can match zext(load i32*) to i64.
5154 bool CodeGenPrepare::extLdPromotion(TypePromotionTransaction &TPT,
5155 LoadInst *&LI, Instruction *&Inst,
5156 const SmallVectorImpl<Instruction *> &Exts,
5157 unsigned CreatedInstsCost = 0) {
5158 // Iterate over all the extensions to see if one form an ext(load).
5159 for (auto I : Exts) {
5160 // Check if we directly have ext(load).
5161 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
5163 // No promotion happened here.
5166 // Check whether or not we want to do any promotion.
5167 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
5169 // Get the action to perform the promotion.
5170 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
5171 I, InsertedInsts, *TLI, PromotedInsts);
5172 // Check if we can promote.
5175 // Save the current state.
5176 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5177 TPT.getRestorationPoint();
5178 SmallVector<Instruction *, 4> NewExts;
5179 unsigned NewCreatedInstsCost = 0;
5180 unsigned ExtCost = !TLI->isExtFree(I);
5182 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
5183 &NewExts, nullptr, *TLI);
5184 assert(PromotedVal &&
5185 "TypePromotionHelper should have filtered out those cases");
5187 // We would be able to merge only one extension in a load.
5188 // Therefore, if we have more than 1 new extension we heuristically
5189 // cut this search path, because it means we degrade the code quality.
5190 // With exactly 2, the transformation is neutral, because we will merge
5191 // one extension but leave one. However, we optimistically keep going,
5192 // because the new extension may be removed too.
5193 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
5194 TotalCreatedInstsCost -= ExtCost;
5195 if (!StressExtLdPromotion &&
5196 (TotalCreatedInstsCost > 1 ||
5197 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
5198 // The promotion is not profitable, rollback to the previous state.
5199 TPT.rollback(LastKnownGood);
5202 // The promotion is profitable.
5203 // Check if it exposes an ext(load).
5204 (void)extLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
5205 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
5206 // If we have created a new extension, i.e., now we have two
5207 // extensions. We must make sure one of them is merged with
5208 // the load, otherwise we may degrade the code quality.
5209 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
5210 // Promotion happened.
5212 // If this does not help to expose an ext(load) then, rollback.
5213 TPT.rollback(LastKnownGood);
5215 // None of the extension can form an ext(load).
5221 /// Move a zext or sext fed by a load into the same basic block as the load,
5222 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
5223 /// extend into the load.
5224 /// \p I[in/out] the extension may be modified during the process if some
5225 /// promotions apply.
5227 bool CodeGenPrepare::moveExtToFormExtLoad(Instruction *&I) {
5228 // Try to promote a chain of computation if it allows to form
5229 // an extended load.
5230 TypePromotionTransaction TPT;
5231 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5232 TPT.getRestorationPoint();
5233 SmallVector<Instruction *, 1> Exts;
5235 // Look for a load being extended.
5236 LoadInst *LI = nullptr;
5237 Instruction *OldExt = I;
5238 bool HasPromoted = extLdPromotion(TPT, LI, I, Exts);
5240 assert(!HasPromoted && !LI && "If we did not match any load instruction "
5241 "the code must remain the same");
5246 // If they're already in the same block, there's nothing to do.
5247 // Make the cheap checks first if we did not promote.
5248 // If we promoted, we need to check if it is indeed profitable.
5249 if (!HasPromoted && LI->getParent() == I->getParent())
5252 EVT VT = TLI->getValueType(*DL, I->getType());
5253 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
5255 // If the load has other users and the truncate is not free, this probably
5256 // isn't worthwhile.
5257 if (!LI->hasOneUse() && TLI &&
5258 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
5259 !TLI->isTruncateFree(I->getType(), LI->getType())) {
5261 TPT.rollback(LastKnownGood);
5265 // Check whether the target supports casts folded into loads.
5267 if (isa<ZExtInst>(I))
5268 LType = ISD::ZEXTLOAD;
5270 assert(isa<SExtInst>(I) && "Unexpected ext type!");
5271 LType = ISD::SEXTLOAD;
5273 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
5275 TPT.rollback(LastKnownGood);
5279 // Move the extend into the same block as the load, so that SelectionDAG
5282 I->removeFromParent();
5288 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
5289 BasicBlock *DefBB = I->getParent();
5291 // If the result of a {s|z}ext and its source are both live out, rewrite all
5292 // other uses of the source with result of extension.
5293 Value *Src = I->getOperand(0);
5294 if (Src->hasOneUse())
5297 // Only do this xform if truncating is free.
5298 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
5301 // Only safe to perform the optimization if the source is also defined in
5303 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
5306 bool DefIsLiveOut = false;
5307 for (User *U : I->users()) {
5308 Instruction *UI = cast<Instruction>(U);
5310 // Figure out which BB this ext is used in.
5311 BasicBlock *UserBB = UI->getParent();
5312 if (UserBB == DefBB) continue;
5313 DefIsLiveOut = true;
5319 // Make sure none of the uses are PHI nodes.
5320 for (User *U : Src->users()) {
5321 Instruction *UI = cast<Instruction>(U);
5322 BasicBlock *UserBB = UI->getParent();
5323 if (UserBB == DefBB) continue;
5324 // Be conservative. We don't want this xform to end up introducing
5325 // reloads just before load / store instructions.
5326 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
5330 // InsertedTruncs - Only insert one trunc in each block once.
5331 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
5333 bool MadeChange = false;
5334 for (Use &U : Src->uses()) {
5335 Instruction *User = cast<Instruction>(U.getUser());
5337 // Figure out which BB this ext is used in.
5338 BasicBlock *UserBB = User->getParent();
5339 if (UserBB == DefBB) continue;
5341 // Both src and def are live in this block. Rewrite the use.
5342 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
5344 if (!InsertedTrunc) {
5345 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5346 assert(InsertPt != UserBB->end());
5347 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5348 InsertedInsts.insert(InsertedTrunc);
5351 // Replace a use of the {s|z}ext source with a use of the result.
5360 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
5361 // just after the load if the target can fold this into one extload instruction,
5362 // with the hope of eliminating some of the other later "and" instructions using
5363 // the loaded value. "and"s that are made trivially redundant by the insertion
5364 // of the new "and" are removed by this function, while others (e.g. those whose
5365 // path from the load goes through a phi) are left for isel to potentially
5398 // becomes (after a call to optimizeLoadExt for each load):
5402 // x1' = and x1, 0xff
5406 // x2' = and x2, 0xff
5413 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5415 if (!Load->isSimple() ||
5416 !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
5419 // Skip loads we've already transformed or have no reason to transform.
5420 if (Load->hasOneUse()) {
5421 User *LoadUser = *Load->user_begin();
5422 if (cast<Instruction>(LoadUser)->getParent() == Load->getParent() &&
5423 !dyn_cast<PHINode>(LoadUser))
5427 // Look at all uses of Load, looking through phis, to determine how many bits
5428 // of the loaded value are needed.
5429 SmallVector<Instruction *, 8> WorkList;
5430 SmallPtrSet<Instruction *, 16> Visited;
5431 SmallVector<Instruction *, 8> AndsToMaybeRemove;
5432 for (auto *U : Load->users())
5433 WorkList.push_back(cast<Instruction>(U));
5435 EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5436 unsigned BitWidth = LoadResultVT.getSizeInBits();
5437 APInt DemandBits(BitWidth, 0);
5438 APInt WidestAndBits(BitWidth, 0);
5440 while (!WorkList.empty()) {
5441 Instruction *I = WorkList.back();
5442 WorkList.pop_back();
5444 // Break use-def graph loops.
5445 if (!Visited.insert(I).second)
5448 // For a PHI node, push all of its users.
5449 if (auto *Phi = dyn_cast<PHINode>(I)) {
5450 for (auto *U : Phi->users())
5451 WorkList.push_back(cast<Instruction>(U));
5455 switch (I->getOpcode()) {
5456 case llvm::Instruction::And: {
5457 auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5460 APInt AndBits = AndC->getValue();
5461 DemandBits |= AndBits;
5462 // Keep track of the widest and mask we see.
5463 if (AndBits.ugt(WidestAndBits))
5464 WidestAndBits = AndBits;
5465 if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5466 AndsToMaybeRemove.push_back(I);
5470 case llvm::Instruction::Shl: {
5471 auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5474 uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5475 auto ShlDemandBits = APInt::getAllOnesValue(BitWidth).lshr(ShiftAmt);
5476 DemandBits |= ShlDemandBits;
5480 case llvm::Instruction::Trunc: {
5481 EVT TruncVT = TLI->getValueType(*DL, I->getType());
5482 unsigned TruncBitWidth = TruncVT.getSizeInBits();
5483 auto TruncBits = APInt::getAllOnesValue(TruncBitWidth).zext(BitWidth);
5484 DemandBits |= TruncBits;
5493 uint32_t ActiveBits = DemandBits.getActiveBits();
5494 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5495 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5496 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5497 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5498 // followed by an AND.
5499 // TODO: Look into removing this restriction by fixing backends to either
5500 // return false for isLoadExtLegal for i1 or have them select this pattern to
5501 // a single instruction.
5503 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5504 // mask, since these are the only ands that will be removed by isel.
5505 if (ActiveBits <= 1 || !APIntOps::isMask(ActiveBits, DemandBits) ||
5506 WidestAndBits != DemandBits)
5509 LLVMContext &Ctx = Load->getType()->getContext();
5510 Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5511 EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5513 // Reject cases that won't be matched as extloads.
5514 if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5515 !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5518 IRBuilder<> Builder(Load->getNextNode());
5519 auto *NewAnd = dyn_cast<Instruction>(
5520 Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5522 // Replace all uses of load with new and (except for the use of load in the
5524 Load->replaceAllUsesWith(NewAnd);
5525 NewAnd->setOperand(0, Load);
5527 // Remove any and instructions that are now redundant.
5528 for (auto *And : AndsToMaybeRemove)
5529 // Check that the and mask is the same as the one we decided to put on the
5531 if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5532 And->replaceAllUsesWith(NewAnd);
5533 if (&*CurInstIterator == And)
5534 CurInstIterator = std::next(And->getIterator());
5535 And->eraseFromParent();
5543 /// Check if V (an operand of a select instruction) is an expensive instruction
5544 /// that is only used once.
5545 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5546 auto *I = dyn_cast<Instruction>(V);
5547 // If it's safe to speculatively execute, then it should not have side
5548 // effects; therefore, it's safe to sink and possibly *not* execute.
5549 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5550 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
5553 /// Returns true if a SelectInst should be turned into an explicit branch.
5554 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
5556 // FIXME: This should use the same heuristics as IfConversion to determine
5557 // whether a select is better represented as a branch. This requires that
5558 // branch probability metadata is preserved for the select, which is not the
5561 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5563 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5564 // comparison condition. If the compare has more than one use, there's
5565 // probably another cmov or setcc around, so it's not worth emitting a branch.
5566 if (!Cmp || !Cmp->hasOneUse())
5569 Value *CmpOp0 = Cmp->getOperand(0);
5570 Value *CmpOp1 = Cmp->getOperand(1);
5572 // Emit "cmov on compare with a memory operand" as a branch to avoid stalls
5573 // on a load from memory. But if the load is used more than once, do not
5574 // change the select to a branch because the load is probably needed
5575 // regardless of whether the branch is taken or not.
5576 if ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
5577 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()))
5580 // If either operand of the select is expensive and only needed on one side
5581 // of the select, we should form a branch.
5582 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5583 sinkSelectOperand(TTI, SI->getFalseValue()))
5590 /// If we have a SelectInst that will likely profit from branch prediction,
5591 /// turn it into a branch.
5592 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5593 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
5595 // Can we convert the 'select' to CF ?
5596 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
5599 TargetLowering::SelectSupportKind SelectKind;
5601 SelectKind = TargetLowering::VectorMaskSelect;
5602 else if (SI->getType()->isVectorTy())
5603 SelectKind = TargetLowering::ScalarCondVectorVal;
5605 SelectKind = TargetLowering::ScalarValSelect;
5607 // Do we have efficient codegen support for this kind of 'selects' ?
5608 if (TLI->isSelectSupported(SelectKind)) {
5609 // We have efficient codegen support for the select instruction.
5610 // Check if it is profitable to keep this 'select'.
5611 if (!TLI->isPredictableSelectExpensive() ||
5612 !isFormingBranchFromSelectProfitable(TTI, SI))
5618 // Transform a sequence like this:
5620 // %cmp = cmp uge i32 %a, %b
5621 // %sel = select i1 %cmp, i32 %c, i32 %d
5625 // %cmp = cmp uge i32 %a, %b
5626 // br i1 %cmp, label %select.true, label %select.false
5628 // br label %select.end
5630 // br label %select.end
5632 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
5634 // In addition, we may sink instructions that produce %c or %d from
5635 // the entry block into the destination(s) of the new branch.
5636 // If the true or false blocks do not contain a sunken instruction, that
5637 // block and its branch may be optimized away. In that case, one side of the
5638 // first branch will point directly to select.end, and the corresponding PHI
5639 // predecessor block will be the start block.
5641 // First, we split the block containing the select into 2 blocks.
5642 BasicBlock *StartBlock = SI->getParent();
5643 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
5644 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
5646 // Delete the unconditional branch that was just created by the split.
5647 StartBlock->getTerminator()->eraseFromParent();
5649 // These are the new basic blocks for the conditional branch.
5650 // At least one will become an actual new basic block.
5651 BasicBlock *TrueBlock = nullptr;
5652 BasicBlock *FalseBlock = nullptr;
5654 // Sink expensive instructions into the conditional blocks to avoid executing
5655 // them speculatively.
5656 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
5657 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
5658 EndBlock->getParent(), EndBlock);
5659 auto *TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
5660 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
5661 TrueInst->moveBefore(TrueBranch);
5663 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
5664 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
5665 EndBlock->getParent(), EndBlock);
5666 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5667 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
5668 FalseInst->moveBefore(FalseBranch);
5671 // If there was nothing to sink, then arbitrarily choose the 'false' side
5672 // for a new input value to the PHI.
5673 if (TrueBlock == FalseBlock) {
5674 assert(TrueBlock == nullptr &&
5675 "Unexpected basic block transform while optimizing select");
5677 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
5678 EndBlock->getParent(), EndBlock);
5679 BranchInst::Create(EndBlock, FalseBlock);
5682 // Insert the real conditional branch based on the original condition.
5683 // If we did not create a new block for one of the 'true' or 'false' paths
5684 // of the condition, it means that side of the branch goes to the end block
5685 // directly and the path originates from the start block from the point of
5686 // view of the new PHI.
5687 if (TrueBlock == nullptr) {
5688 BranchInst::Create(EndBlock, FalseBlock, SI->getCondition(), SI);
5689 TrueBlock = StartBlock;
5690 } else if (FalseBlock == nullptr) {
5691 BranchInst::Create(TrueBlock, EndBlock, SI->getCondition(), SI);
5692 FalseBlock = StartBlock;
5694 BranchInst::Create(TrueBlock, FalseBlock, SI->getCondition(), SI);
5697 // The select itself is replaced with a PHI Node.
5698 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
5700 PN->addIncoming(SI->getTrueValue(), TrueBlock);
5701 PN->addIncoming(SI->getFalseValue(), FalseBlock);
5703 SI->replaceAllUsesWith(PN);
5704 SI->eraseFromParent();
5706 // Instruct OptimizeBlock to skip to the next block.
5707 CurInstIterator = StartBlock->end();
5708 ++NumSelectsExpanded;
5712 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
5713 SmallVector<int, 16> Mask(SVI->getShuffleMask());
5715 for (unsigned i = 0; i < Mask.size(); ++i) {
5716 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
5718 SplatElem = Mask[i];
5724 /// Some targets have expensive vector shifts if the lanes aren't all the same
5725 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
5726 /// it's often worth sinking a shufflevector splat down to its use so that
5727 /// codegen can spot all lanes are identical.
5728 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
5729 BasicBlock *DefBB = SVI->getParent();
5731 // Only do this xform if variable vector shifts are particularly expensive.
5732 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
5735 // We only expect better codegen by sinking a shuffle if we can recognise a
5737 if (!isBroadcastShuffle(SVI))
5740 // InsertedShuffles - Only insert a shuffle in each block once.
5741 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
5743 bool MadeChange = false;
5744 for (User *U : SVI->users()) {
5745 Instruction *UI = cast<Instruction>(U);
5747 // Figure out which BB this ext is used in.
5748 BasicBlock *UserBB = UI->getParent();
5749 if (UserBB == DefBB) continue;
5751 // For now only apply this when the splat is used by a shift instruction.
5752 if (!UI->isShift()) continue;
5754 // Everything checks out, sink the shuffle if the user's block doesn't
5755 // already have a copy.
5756 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
5758 if (!InsertedShuffle) {
5759 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5760 assert(InsertPt != UserBB->end());
5762 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5763 SVI->getOperand(2), "", &*InsertPt);
5766 UI->replaceUsesOfWith(SVI, InsertedShuffle);
5770 // If we removed all uses, nuke the shuffle.
5771 if (SVI->use_empty()) {
5772 SVI->eraseFromParent();
5779 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
5783 Value *Cond = SI->getCondition();
5784 Type *OldType = Cond->getType();
5785 LLVMContext &Context = Cond->getContext();
5786 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
5787 unsigned RegWidth = RegType.getSizeInBits();
5789 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
5792 // If the register width is greater than the type width, expand the condition
5793 // of the switch instruction and each case constant to the width of the
5794 // register. By widening the type of the switch condition, subsequent
5795 // comparisons (for case comparisons) will not need to be extended to the
5796 // preferred register width, so we will potentially eliminate N-1 extends,
5797 // where N is the number of cases in the switch.
5798 auto *NewType = Type::getIntNTy(Context, RegWidth);
5800 // Zero-extend the switch condition and case constants unless the switch
5801 // condition is a function argument that is already being sign-extended.
5802 // In that case, we can avoid an unnecessary mask/extension by sign-extending
5803 // everything instead.
5804 Instruction::CastOps ExtType = Instruction::ZExt;
5805 if (auto *Arg = dyn_cast<Argument>(Cond))
5806 if (Arg->hasSExtAttr())
5807 ExtType = Instruction::SExt;
5809 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
5810 ExtInst->insertBefore(SI);
5811 SI->setCondition(ExtInst);
5812 for (SwitchInst::CaseIt Case : SI->cases()) {
5813 APInt NarrowConst = Case.getCaseValue()->getValue();
5814 APInt WideConst = (ExtType == Instruction::ZExt) ?
5815 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
5816 Case.setValue(ConstantInt::get(Context, WideConst));
5823 /// \brief Helper class to promote a scalar operation to a vector one.
5824 /// This class is used to move downward extractelement transition.
5826 /// a = vector_op <2 x i32>
5827 /// b = extractelement <2 x i32> a, i32 0
5832 /// a = vector_op <2 x i32>
5833 /// c = vector_op a (equivalent to scalar_op on the related lane)
5834 /// * d = extractelement <2 x i32> c, i32 0
5836 /// Assuming both extractelement and store can be combine, we get rid of the
5838 class VectorPromoteHelper {
5839 /// DataLayout associated with the current module.
5840 const DataLayout &DL;
5842 /// Used to perform some checks on the legality of vector operations.
5843 const TargetLowering &TLI;
5845 /// Used to estimated the cost of the promoted chain.
5846 const TargetTransformInfo &TTI;
5848 /// The transition being moved downwards.
5849 Instruction *Transition;
5850 /// The sequence of instructions to be promoted.
5851 SmallVector<Instruction *, 4> InstsToBePromoted;
5852 /// Cost of combining a store and an extract.
5853 unsigned StoreExtractCombineCost;
5854 /// Instruction that will be combined with the transition.
5855 Instruction *CombineInst;
5857 /// \brief The instruction that represents the current end of the transition.
5858 /// Since we are faking the promotion until we reach the end of the chain
5859 /// of computation, we need a way to get the current end of the transition.
5860 Instruction *getEndOfTransition() const {
5861 if (InstsToBePromoted.empty())
5863 return InstsToBePromoted.back();
5866 /// \brief Return the index of the original value in the transition.
5867 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
5868 /// c, is at index 0.
5869 unsigned getTransitionOriginalValueIdx() const {
5870 assert(isa<ExtractElementInst>(Transition) &&
5871 "Other kind of transitions are not supported yet");
5875 /// \brief Return the index of the index in the transition.
5876 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
5878 unsigned getTransitionIdx() const {
5879 assert(isa<ExtractElementInst>(Transition) &&
5880 "Other kind of transitions are not supported yet");
5884 /// \brief Get the type of the transition.
5885 /// This is the type of the original value.
5886 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
5887 /// transition is <2 x i32>.
5888 Type *getTransitionType() const {
5889 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
5892 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
5893 /// I.e., we have the following sequence:
5894 /// Def = Transition <ty1> a to <ty2>
5895 /// b = ToBePromoted <ty2> Def, ...
5897 /// b = ToBePromoted <ty1> a, ...
5898 /// Def = Transition <ty1> ToBePromoted to <ty2>
5899 void promoteImpl(Instruction *ToBePromoted);
5901 /// \brief Check whether or not it is profitable to promote all the
5902 /// instructions enqueued to be promoted.
5903 bool isProfitableToPromote() {
5904 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
5905 unsigned Index = isa<ConstantInt>(ValIdx)
5906 ? cast<ConstantInt>(ValIdx)->getZExtValue()
5908 Type *PromotedType = getTransitionType();
5910 StoreInst *ST = cast<StoreInst>(CombineInst);
5911 unsigned AS = ST->getPointerAddressSpace();
5912 unsigned Align = ST->getAlignment();
5913 // Check if this store is supported.
5914 if (!TLI.allowsMisalignedMemoryAccesses(
5915 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
5917 // If this is not supported, there is no way we can combine
5918 // the extract with the store.
5922 // The scalar chain of computation has to pay for the transition
5923 // scalar to vector.
5924 // The vector chain has to account for the combining cost.
5925 uint64_t ScalarCost =
5926 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
5927 uint64_t VectorCost = StoreExtractCombineCost;
5928 for (const auto &Inst : InstsToBePromoted) {
5929 // Compute the cost.
5930 // By construction, all instructions being promoted are arithmetic ones.
5931 // Moreover, one argument is a constant that can be viewed as a splat
5933 Value *Arg0 = Inst->getOperand(0);
5934 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
5935 isa<ConstantFP>(Arg0);
5936 TargetTransformInfo::OperandValueKind Arg0OVK =
5937 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
5938 : TargetTransformInfo::OK_AnyValue;
5939 TargetTransformInfo::OperandValueKind Arg1OVK =
5940 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
5941 : TargetTransformInfo::OK_AnyValue;
5942 ScalarCost += TTI.getArithmeticInstrCost(
5943 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
5944 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
5947 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
5948 << ScalarCost << "\nVector: " << VectorCost << '\n');
5949 return ScalarCost > VectorCost;
5952 /// \brief Generate a constant vector with \p Val with the same
5953 /// number of elements as the transition.
5954 /// \p UseSplat defines whether or not \p Val should be replicated
5955 /// across the whole vector.
5956 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
5957 /// otherwise we generate a vector with as many undef as possible:
5958 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
5959 /// used at the index of the extract.
5960 Value *getConstantVector(Constant *Val, bool UseSplat) const {
5961 unsigned ExtractIdx = UINT_MAX;
5963 // If we cannot determine where the constant must be, we have to
5964 // use a splat constant.
5965 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
5966 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
5967 ExtractIdx = CstVal->getSExtValue();
5972 unsigned End = getTransitionType()->getVectorNumElements();
5974 return ConstantVector::getSplat(End, Val);
5976 SmallVector<Constant *, 4> ConstVec;
5977 UndefValue *UndefVal = UndefValue::get(Val->getType());
5978 for (unsigned Idx = 0; Idx != End; ++Idx) {
5979 if (Idx == ExtractIdx)
5980 ConstVec.push_back(Val);
5982 ConstVec.push_back(UndefVal);
5984 return ConstantVector::get(ConstVec);
5987 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
5988 /// in \p Use can trigger undefined behavior.
5989 static bool canCauseUndefinedBehavior(const Instruction *Use,
5990 unsigned OperandIdx) {
5991 // This is not safe to introduce undef when the operand is on
5992 // the right hand side of a division-like instruction.
5993 if (OperandIdx != 1)
5995 switch (Use->getOpcode()) {
5998 case Instruction::SDiv:
5999 case Instruction::UDiv:
6000 case Instruction::SRem:
6001 case Instruction::URem:
6003 case Instruction::FDiv:
6004 case Instruction::FRem:
6005 return !Use->hasNoNaNs();
6007 llvm_unreachable(nullptr);
6011 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
6012 const TargetTransformInfo &TTI, Instruction *Transition,
6013 unsigned CombineCost)
6014 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
6015 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
6016 assert(Transition && "Do not know how to promote null");
6019 /// \brief Check if we can promote \p ToBePromoted to \p Type.
6020 bool canPromote(const Instruction *ToBePromoted) const {
6021 // We could support CastInst too.
6022 return isa<BinaryOperator>(ToBePromoted);
6025 /// \brief Check if it is profitable to promote \p ToBePromoted
6026 /// by moving downward the transition through.
6027 bool shouldPromote(const Instruction *ToBePromoted) const {
6028 // Promote only if all the operands can be statically expanded.
6029 // Indeed, we do not want to introduce any new kind of transitions.
6030 for (const Use &U : ToBePromoted->operands()) {
6031 const Value *Val = U.get();
6032 if (Val == getEndOfTransition()) {
6033 // If the use is a division and the transition is on the rhs,
6034 // we cannot promote the operation, otherwise we may create a
6035 // division by zero.
6036 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
6040 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
6041 !isa<ConstantFP>(Val))
6044 // Check that the resulting operation is legal.
6045 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
6048 return StressStoreExtract ||
6049 TLI.isOperationLegalOrCustom(
6050 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
6053 /// \brief Check whether or not \p Use can be combined
6054 /// with the transition.
6055 /// I.e., is it possible to do Use(Transition) => AnotherUse?
6056 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
6058 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
6059 void enqueueForPromotion(Instruction *ToBePromoted) {
6060 InstsToBePromoted.push_back(ToBePromoted);
6063 /// \brief Set the instruction that will be combined with the transition.
6064 void recordCombineInstruction(Instruction *ToBeCombined) {
6065 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
6066 CombineInst = ToBeCombined;
6069 /// \brief Promote all the instructions enqueued for promotion if it is
6071 /// \return True if the promotion happened, false otherwise.
6073 // Check if there is something to promote.
6074 // Right now, if we do not have anything to combine with,
6075 // we assume the promotion is not profitable.
6076 if (InstsToBePromoted.empty() || !CombineInst)
6080 if (!StressStoreExtract && !isProfitableToPromote())
6084 for (auto &ToBePromoted : InstsToBePromoted)
6085 promoteImpl(ToBePromoted);
6086 InstsToBePromoted.clear();
6090 } // End of anonymous namespace.
6092 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
6093 // At this point, we know that all the operands of ToBePromoted but Def
6094 // can be statically promoted.
6095 // For Def, we need to use its parameter in ToBePromoted:
6096 // b = ToBePromoted ty1 a
6097 // Def = Transition ty1 b to ty2
6098 // Move the transition down.
6099 // 1. Replace all uses of the promoted operation by the transition.
6100 // = ... b => = ... Def.
6101 assert(ToBePromoted->getType() == Transition->getType() &&
6102 "The type of the result of the transition does not match "
6104 ToBePromoted->replaceAllUsesWith(Transition);
6105 // 2. Update the type of the uses.
6106 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
6107 Type *TransitionTy = getTransitionType();
6108 ToBePromoted->mutateType(TransitionTy);
6109 // 3. Update all the operands of the promoted operation with promoted
6111 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
6112 for (Use &U : ToBePromoted->operands()) {
6113 Value *Val = U.get();
6114 Value *NewVal = nullptr;
6115 if (Val == Transition)
6116 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
6117 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
6118 isa<ConstantFP>(Val)) {
6119 // Use a splat constant if it is not safe to use undef.
6120 NewVal = getConstantVector(
6121 cast<Constant>(Val),
6122 isa<UndefValue>(Val) ||
6123 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
6125 llvm_unreachable("Did you modified shouldPromote and forgot to update "
6127 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
6129 Transition->removeFromParent();
6130 Transition->insertAfter(ToBePromoted);
6131 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
6134 /// Some targets can do store(extractelement) with one instruction.
6135 /// Try to push the extractelement towards the stores when the target
6136 /// has this feature and this is profitable.
6137 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
6138 unsigned CombineCost = UINT_MAX;
6139 if (DisableStoreExtract || !TLI ||
6140 (!StressStoreExtract &&
6141 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
6142 Inst->getOperand(1), CombineCost)))
6145 // At this point we know that Inst is a vector to scalar transition.
6146 // Try to move it down the def-use chain, until:
6147 // - We can combine the transition with its single use
6148 // => we got rid of the transition.
6149 // - We escape the current basic block
6150 // => we would need to check that we are moving it at a cheaper place and
6151 // we do not do that for now.
6152 BasicBlock *Parent = Inst->getParent();
6153 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
6154 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
6155 // If the transition has more than one use, assume this is not going to be
6157 while (Inst->hasOneUse()) {
6158 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
6159 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
6161 if (ToBePromoted->getParent() != Parent) {
6162 DEBUG(dbgs() << "Instruction to promote is in a different block ("
6163 << ToBePromoted->getParent()->getName()
6164 << ") than the transition (" << Parent->getName() << ").\n");
6168 if (VPH.canCombine(ToBePromoted)) {
6169 DEBUG(dbgs() << "Assume " << *Inst << '\n'
6170 << "will be combined with: " << *ToBePromoted << '\n');
6171 VPH.recordCombineInstruction(ToBePromoted);
6172 bool Changed = VPH.promote();
6173 NumStoreExtractExposed += Changed;
6177 DEBUG(dbgs() << "Try promoting.\n");
6178 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
6181 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
6183 VPH.enqueueForPromotion(ToBePromoted);
6184 Inst = ToBePromoted;
6189 bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
6190 // Bail out if we inserted the instruction to prevent optimizations from
6191 // stepping on each other's toes.
6192 if (InsertedInsts.count(I))
6195 if (PHINode *P = dyn_cast<PHINode>(I)) {
6196 // It is possible for very late stage optimizations (such as SimplifyCFG)
6197 // to introduce PHI nodes too late to be cleaned up. If we detect such a
6198 // trivial PHI, go ahead and zap it here.
6199 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
6200 P->replaceAllUsesWith(V);
6201 P->eraseFromParent();
6208 if (CastInst *CI = dyn_cast<CastInst>(I)) {
6209 // If the source of the cast is a constant, then this should have
6210 // already been constant folded. The only reason NOT to constant fold
6211 // it is if something (e.g. LSR) was careful to place the constant
6212 // evaluation in a block other than then one that uses it (e.g. to hoist
6213 // the address of globals out of a loop). If this is the case, we don't
6214 // want to forward-subst the cast.
6215 if (isa<Constant>(CI->getOperand(0)))
6218 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
6221 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6222 /// Sink a zext or sext into its user blocks if the target type doesn't
6223 /// fit in one register
6225 TLI->getTypeAction(CI->getContext(),
6226 TLI->getValueType(*DL, CI->getType())) ==
6227 TargetLowering::TypeExpandInteger) {
6228 return SinkCast(CI);
6230 bool MadeChange = moveExtToFormExtLoad(I);
6231 return MadeChange | optimizeExtUses(I);
6237 if (CmpInst *CI = dyn_cast<CmpInst>(I))
6238 if (!TLI || !TLI->hasMultipleConditionRegisters())
6239 return OptimizeCmpExpression(CI);
6241 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
6242 stripInvariantGroupMetadata(*LI);
6244 bool Modified = optimizeLoadExt(LI);
6245 unsigned AS = LI->getPointerAddressSpace();
6246 Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
6252 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
6253 stripInvariantGroupMetadata(*SI);
6255 unsigned AS = SI->getPointerAddressSpace();
6256 return optimizeMemoryInst(I, SI->getOperand(1),
6257 SI->getOperand(0)->getType(), AS);
6262 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
6264 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
6265 BinOp->getOpcode() == Instruction::LShr)) {
6266 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
6267 if (TLI && CI && TLI->hasExtractBitsInsn())
6268 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
6273 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
6274 if (GEPI->hasAllZeroIndices()) {
6275 /// The GEP operand must be a pointer, so must its result -> BitCast
6276 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
6277 GEPI->getName(), GEPI);
6278 GEPI->replaceAllUsesWith(NC);
6279 GEPI->eraseFromParent();
6281 optimizeInst(NC, ModifiedDT);
6287 if (CallInst *CI = dyn_cast<CallInst>(I))
6288 return optimizeCallInst(CI, ModifiedDT);
6290 if (SelectInst *SI = dyn_cast<SelectInst>(I))
6291 return optimizeSelectInst(SI);
6293 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
6294 return optimizeShuffleVectorInst(SVI);
6296 if (auto *Switch = dyn_cast<SwitchInst>(I))
6297 return optimizeSwitchInst(Switch);
6299 if (isa<ExtractElementInst>(I))
6300 return optimizeExtractElementInst(I);
6305 /// Given an OR instruction, check to see if this is a bitreverse
6306 /// idiom. If so, insert the new intrinsic and return true.
6307 static bool makeBitReverse(Instruction &I, const DataLayout &DL,
6308 const TargetLowering &TLI) {
6309 if (!I.getType()->isIntegerTy() ||
6310 !TLI.isOperationLegalOrCustom(ISD::BITREVERSE,
6311 TLI.getValueType(DL, I.getType(), true)))
6314 SmallVector<Instruction*, 4> Insts;
6315 if (!recognizeBitReverseOrBSwapIdiom(&I, false, true, Insts))
6317 Instruction *LastInst = Insts.back();
6318 I.replaceAllUsesWith(LastInst);
6319 RecursivelyDeleteTriviallyDeadInstructions(&I);
6323 // In this pass we look for GEP and cast instructions that are used
6324 // across basic blocks and rewrite them to improve basic-block-at-a-time
6326 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
6328 bool MadeChange = false;
6330 CurInstIterator = BB.begin();
6331 while (CurInstIterator != BB.end()) {
6332 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
6337 bool MadeBitReverse = true;
6338 while (TLI && MadeBitReverse) {
6339 MadeBitReverse = false;
6340 for (auto &I : reverse(BB)) {
6341 if (makeBitReverse(I, *DL, *TLI)) {
6342 MadeBitReverse = MadeChange = true;
6347 MadeChange |= dupRetToEnableTailCallOpts(&BB);
6352 // llvm.dbg.value is far away from the value then iSel may not be able
6353 // handle it properly. iSel will drop llvm.dbg.value if it can not
6354 // find a node corresponding to the value.
6355 bool CodeGenPrepare::placeDbgValues(Function &F) {
6356 bool MadeChange = false;
6357 for (BasicBlock &BB : F) {
6358 Instruction *PrevNonDbgInst = nullptr;
6359 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
6360 Instruction *Insn = &*BI++;
6361 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
6362 // Leave dbg.values that refer to an alloca alone. These
6363 // instrinsics describe the address of a variable (= the alloca)
6364 // being taken. They should not be moved next to the alloca
6365 // (and to the beginning of the scope), but rather stay close to
6366 // where said address is used.
6367 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
6368 PrevNonDbgInst = Insn;
6372 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
6373 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
6374 // If VI is a phi in a block with an EHPad terminator, we can't insert
6376 if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
6378 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
6379 DVI->removeFromParent();
6380 if (isa<PHINode>(VI))
6381 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
6383 DVI->insertAfter(VI);
6392 // If there is a sequence that branches based on comparing a single bit
6393 // against zero that can be combined into a single instruction, and the
6394 // target supports folding these into a single instruction, sink the
6395 // mask and compare into the branch uses. Do this before OptimizeBlock ->
6396 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
6398 bool CodeGenPrepare::sinkAndCmp(Function &F) {
6399 if (!EnableAndCmpSinking)
6401 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
6403 bool MadeChange = false;
6404 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
6405 BasicBlock *BB = &*I++;
6407 // Does this BB end with the following?
6408 // %andVal = and %val, #single-bit-set
6409 // %icmpVal = icmp %andResult, 0
6410 // br i1 %cmpVal label %dest1, label %dest2"
6411 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
6412 if (!Brcc || !Brcc->isConditional())
6414 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
6415 if (!Cmp || Cmp->getParent() != BB)
6417 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
6418 if (!Zero || !Zero->isZero())
6420 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
6421 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
6423 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
6424 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
6426 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
6428 // Push the "and; icmp" for any users that are conditional branches.
6429 // Since there can only be one branch use per BB, we don't need to keep
6430 // track of which BBs we insert into.
6431 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
6435 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
6437 if (!BrccUser || !BrccUser->isConditional())
6439 BasicBlock *UserBB = BrccUser->getParent();
6440 if (UserBB == BB) continue;
6441 DEBUG(dbgs() << "found Brcc use\n");
6443 // Sink the "and; icmp" to use.
6445 BinaryOperator *NewAnd =
6446 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
6449 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
6453 DEBUG(BrccUser->getParent()->dump());
6459 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
6460 /// success, or returns false if no or invalid metadata was found.
6461 static bool extractBranchMetadata(BranchInst *BI,
6462 uint64_t &ProbTrue, uint64_t &ProbFalse) {
6463 assert(BI->isConditional() &&
6464 "Looking for probabilities on unconditional branch?");
6465 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
6466 if (!ProfileData || ProfileData->getNumOperands() != 3)
6469 const auto *CITrue =
6470 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
6471 const auto *CIFalse =
6472 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
6473 if (!CITrue || !CIFalse)
6476 ProbTrue = CITrue->getValue().getZExtValue();
6477 ProbFalse = CIFalse->getValue().getZExtValue();
6482 /// \brief Scale down both weights to fit into uint32_t.
6483 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
6484 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
6485 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
6486 NewTrue = NewTrue / Scale;
6487 NewFalse = NewFalse / Scale;
6490 /// \brief Some targets prefer to split a conditional branch like:
6492 /// %0 = icmp ne i32 %a, 0
6493 /// %1 = icmp ne i32 %b, 0
6494 /// %or.cond = or i1 %0, %1
6495 /// br i1 %or.cond, label %TrueBB, label %FalseBB
6497 /// into multiple branch instructions like:
6500 /// %0 = icmp ne i32 %a, 0
6501 /// br i1 %0, label %TrueBB, label %bb2
6503 /// %1 = icmp ne i32 %b, 0
6504 /// br i1 %1, label %TrueBB, label %FalseBB
6506 /// This usually allows instruction selection to do even further optimizations
6507 /// and combine the compare with the branch instruction. Currently this is
6508 /// applied for targets which have "cheap" jump instructions.
6510 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
6512 bool CodeGenPrepare::splitBranchCondition(Function &F) {
6513 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
6516 bool MadeChange = false;
6517 for (auto &BB : F) {
6518 // Does this BB end with the following?
6519 // %cond1 = icmp|fcmp|binary instruction ...
6520 // %cond2 = icmp|fcmp|binary instruction ...
6521 // %cond.or = or|and i1 %cond1, cond2
6522 // br i1 %cond.or label %dest1, label %dest2"
6523 BinaryOperator *LogicOp;
6524 BasicBlock *TBB, *FBB;
6525 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
6528 auto *Br1 = cast<BranchInst>(BB.getTerminator());
6529 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
6533 Value *Cond1, *Cond2;
6534 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
6535 m_OneUse(m_Value(Cond2)))))
6536 Opc = Instruction::And;
6537 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
6538 m_OneUse(m_Value(Cond2)))))
6539 Opc = Instruction::Or;
6543 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
6544 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
6547 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
6550 auto *InsertBefore = std::next(Function::iterator(BB))
6551 .getNodePtrUnchecked();
6552 auto TmpBB = BasicBlock::Create(BB.getContext(),
6553 BB.getName() + ".cond.split",
6554 BB.getParent(), InsertBefore);
6556 // Update original basic block by using the first condition directly by the
6557 // branch instruction and removing the no longer needed and/or instruction.
6558 Br1->setCondition(Cond1);
6559 LogicOp->eraseFromParent();
6561 // Depending on the conditon we have to either replace the true or the false
6562 // successor of the original branch instruction.
6563 if (Opc == Instruction::And)
6564 Br1->setSuccessor(0, TmpBB);
6566 Br1->setSuccessor(1, TmpBB);
6568 // Fill in the new basic block.
6569 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
6570 if (auto *I = dyn_cast<Instruction>(Cond2)) {
6571 I->removeFromParent();
6572 I->insertBefore(Br2);
6575 // Update PHI nodes in both successors. The original BB needs to be
6576 // replaced in one succesor's PHI nodes, because the branch comes now from
6577 // the newly generated BB (NewBB). In the other successor we need to add one
6578 // incoming edge to the PHI nodes, because both branch instructions target
6579 // now the same successor. Depending on the original branch condition
6580 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
6581 // we perfrom the correct update for the PHI nodes.
6582 // This doesn't change the successor order of the just created branch
6583 // instruction (or any other instruction).
6584 if (Opc == Instruction::Or)
6585 std::swap(TBB, FBB);
6587 // Replace the old BB with the new BB.
6588 for (auto &I : *TBB) {
6589 PHINode *PN = dyn_cast<PHINode>(&I);
6593 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
6594 PN->setIncomingBlock(i, TmpBB);
6597 // Add another incoming edge form the new BB.
6598 for (auto &I : *FBB) {
6599 PHINode *PN = dyn_cast<PHINode>(&I);
6602 auto *Val = PN->getIncomingValueForBlock(&BB);
6603 PN->addIncoming(Val, TmpBB);
6606 // Update the branch weights (from SelectionDAGBuilder::
6607 // FindMergedConditions).
6608 if (Opc == Instruction::Or) {
6609 // Codegen X | Y as:
6618 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
6619 // The requirement is that
6620 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
6621 // = TrueProb for orignal BB.
6622 // Assuming the orignal weights are A and B, one choice is to set BB1's
6623 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
6625 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
6626 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
6627 // TmpBB, but the math is more complicated.
6628 uint64_t TrueWeight, FalseWeight;
6629 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
6630 uint64_t NewTrueWeight = TrueWeight;
6631 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
6632 scaleWeights(NewTrueWeight, NewFalseWeight);
6633 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6634 .createBranchWeights(TrueWeight, FalseWeight));
6636 NewTrueWeight = TrueWeight;
6637 NewFalseWeight = 2 * FalseWeight;
6638 scaleWeights(NewTrueWeight, NewFalseWeight);
6639 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6640 .createBranchWeights(TrueWeight, FalseWeight));
6643 // Codegen X & Y as:
6651 // This requires creation of TmpBB after CurBB.
6653 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
6654 // The requirement is that
6655 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
6656 // = FalseProb for orignal BB.
6657 // Assuming the orignal weights are A and B, one choice is to set BB1's
6658 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
6660 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
6661 uint64_t TrueWeight, FalseWeight;
6662 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
6663 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
6664 uint64_t NewFalseWeight = FalseWeight;
6665 scaleWeights(NewTrueWeight, NewFalseWeight);
6666 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6667 .createBranchWeights(TrueWeight, FalseWeight));
6669 NewTrueWeight = 2 * TrueWeight;
6670 NewFalseWeight = FalseWeight;
6671 scaleWeights(NewTrueWeight, NewFalseWeight);
6672 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6673 .createBranchWeights(TrueWeight, FalseWeight));
6677 // Note: No point in getting fancy here, since the DT info is never
6678 // available to CodeGenPrepare.
6683 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
6689 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
6690 if (auto *InvariantMD = I.getMetadata(LLVMContext::MD_invariant_group))
6691 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());