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 auto* Inst = dyn_cast<Instruction>(&*BBI);
714 if (Inst == nullptr) {
717 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 // XXX-comment: Returns whether the code has been changed.
763 bool taintMonotonicLoads(const SmallVector<LoadInst*, 1>& MonotonicLoadInsts) {
764 bool Changed = false;
765 for (auto* LI : MonotonicLoadInsts) {
766 SmallVector<BasicBlock*, 2> ChainedBB;
767 auto* FirstInst = findFirstStoreCondBranchInst(LI, &ChainedBB);
768 if (FirstInst == nullptr) {
769 // We don't seem to be able to taint a following store/conditional branch
770 // instruction. Simply make it acquire.
771 DEBUG(dbgs() << "[RelaxedLoad]: Transformed to acquire load\n"
773 LI->setOrdering(Acquire);
777 // Taint 'FirstInst', which could be a store or a condition branch
779 if (FirstInst->getOpcode() == Instruction::Store) {
780 Changed |= taintStoreAddress(dyn_cast<StoreInst>(FirstInst), LI);
781 } else if (FirstInst->getOpcode() == Instruction::Br) {
782 Changed |= taintConditionalBranch(dyn_cast<BranchInst>(FirstInst), LI);
784 assert(false && "findFirstStoreCondBranchInst() should return a "
785 "store/condition branch instruction");
791 // Inserts a fake conditional branch right after the instruction 'SplitInst',
792 // and the branch condition is 'Condition'. 'SplitInst' will be placed in the
793 // newly created block.
794 void AddFakeConditionalBranch(Instruction* SplitInst, Value* Condition) {
795 auto* BB = SplitInst->getParent();
796 TerminatorInst* ThenTerm = nullptr;
797 TerminatorInst* ElseTerm = nullptr;
798 SplitBlockAndInsertIfThenElse(Condition, SplitInst, &ThenTerm, &ElseTerm);
799 assert(ThenTerm && ElseTerm &&
800 "Then/Else terminators cannot be empty after basic block spliting");
801 auto* ThenBB = ThenTerm->getParent();
802 auto* ElseBB = ElseTerm->getParent();
803 auto* TailBB = ThenBB->getSingleSuccessor();
804 assert(TailBB && "Tail block cannot be empty after basic block spliting");
806 ThenBB->disableCanEliminateBlock();
807 ThenBB->disableCanEliminateBlock();
808 TailBB->disableCanEliminateBlock();
809 ThenBB->setName(BB->getName() + "Then.Fake");
810 ElseBB->setName(BB->getName() + "Else.Fake");
811 DEBUG(dbgs() << "Add fake conditional branch:\n"
813 << *ThenBB << "Else Block:\n"
817 // Returns true if the code is changed, and false otherwise.
818 void TaintRelaxedLoads(Instruction* UsageInst, Instruction* InsertPoint) {
819 // For better performance, we can add a "AND X 0" instruction before the
821 auto* BB = UsageInst->getParent();
822 if (InsertPoint == nullptr) {
823 InsertPoint = UsageInst->getNextNode();
825 // Insert instructions after PHI nodes.
826 while (dyn_cast<PHINode>(InsertPoint)) {
827 InsertPoint = InsertPoint->getNextNode();
829 // First thing is to cast 'UsageInst' to an integer type if necessary.
830 Value* AndTarget = nullptr;
831 Type* TargetIntegerType =
832 IntegerType::get(UsageInst->getContext(),
833 BB->getModule()->getDataLayout().getPointerSizeInBits());
835 // Check whether InsertPoint is a added fake conditional branch.
836 BranchInst* BI = nullptr;
837 if ((BI = dyn_cast<BranchInst>(InsertPoint)) && BI->isConditional()) {
838 auto* Cond = dyn_cast<Instruction>(BI->getOperand(0));
839 if (Cond && Cond->getOpcode() == Instruction::ICmp) {
840 auto* CmpInst = dyn_cast<ICmpInst>(Cond);
841 auto* Op0 = dyn_cast<Instruction>(Cond->getOperand(0));
842 auto* Op1 = dyn_cast<ConstantInt>(Cond->getOperand(1));
844 // %cmp = ICMP_NE %tmp, 0
847 // %tmp1 = And X, NewTaintedVal
848 // %tmp2 = And %tmp1, 0
849 // %cmp = ICMP_NE %tmp2, 0
851 if (CmpInst && CmpInst->getPredicate() == CmpInst::ICMP_NE && Op0 &&
852 Op0->getOpcode() == Instruction::And && Op1 && Op1->isZero()) {
853 auto* Op01 = dyn_cast<ConstantInt>(Op0->getOperand(1));
854 if (Op01 && Op01->isZero()) {
855 // Now we have a previously added fake cond branch.
856 auto* Op00 = Op0->getOperand(0);
857 IRBuilder<true, NoFolder> Builder(CmpInst);
858 if (Op00->getType() == UsageInst->getType()) {
859 AndTarget = UsageInst;
861 AndTarget = createCast(Builder, UsageInst, Op00->getType());
863 AndTarget = Builder.CreateAnd(Op00, AndTarget);
864 auto* AndZero = dyn_cast<Instruction>(Builder.CreateAnd(
865 AndTarget, Constant::getNullValue(AndTarget->getType())));
866 CmpInst->setOperand(0, AndZero);
873 IRBuilder<true, NoFolder> Builder(InsertPoint);
874 if (IntegerType::classof(UsageInst->getType())) {
875 AndTarget = UsageInst;
877 AndTarget = createCast(Builder, UsageInst, TargetIntegerType);
879 auto* AndZero = dyn_cast<Instruction>(
880 Builder.CreateAnd(AndTarget, Constant::getNullValue(AndTarget->getType())));
881 auto* FakeCondition = dyn_cast<Instruction>(Builder.CreateICmp(
882 CmpInst::ICMP_NE, AndZero, Constant::getNullValue(AndTarget->getType())));
883 AddFakeConditionalBranch(FakeCondition->getNextNode(), FakeCondition);
886 // XXX-comment: Finds the appropriate Value derived from an atomic load.
887 // 'ChainedBB' contains all the blocks chained together with unconditional
888 // branches from LI's parent BB to the block with the first store/cond branch.
889 // If we don't find any, it means 'LI' is not used at all (which should not
890 // happen in practice). We can simply set 'LI' to be acquire just to be safe.
891 template <typename Vector>
892 Instruction* findMostRecentDependenceUsage(LoadInst* LI, Instruction* LaterInst,
895 typedef SmallSet<Instruction*, 8> UsageSet;
896 typedef DenseMap<BasicBlock*, std::unique_ptr<UsageSet>> UsageMap;
897 assert(ChainedBB->size() >= 1 && "ChainedBB must have >=1 size");
898 // Mapping from basic block in 'ChainedBB' to the set of dependence usage of
899 // 'LI' in each block.
901 auto* LoadBB = LI->getParent();
902 usage_map[LoadBB] = make_unique<UsageSet>();
903 usage_map[LoadBB]->insert(LI);
905 for (auto* BB : *ChainedBB) {
906 if (usage_map[BB] == nullptr) {
907 usage_map[BB] = make_unique<UsageSet>();
909 auto& usage_set = usage_map[BB];
910 if (usage_set->size() == 0) {
911 // The value has not been used.
914 // Calculate the usage in the current BB first.
915 std::list<Value*> bb_usage_list;
916 std::copy(usage_set->begin(), usage_set->end(),
917 std::back_inserter(bb_usage_list));
918 for (auto list_iter = bb_usage_list.begin();
919 list_iter != bb_usage_list.end(); list_iter++) {
920 auto* val = *list_iter;
921 for (auto* U : val->users()) {
922 Instruction* Inst = nullptr;
923 if (!(Inst = dyn_cast<Instruction>(U))) {
926 assert(Inst && "Usage value must be an instruction");
928 std::find(ChainedBB->begin(), ChainedBB->end(), Inst->getParent());
929 if (iter == ChainedBB->end()) {
930 // Only care about usage within ChainedBB.
933 auto* UsageBB = *iter;
936 if (!usage_set->count(Inst)) {
937 bb_usage_list.push_back(Inst);
938 usage_set->insert(Inst);
942 if (usage_map[UsageBB] == nullptr) {
943 usage_map[UsageBB] = make_unique<UsageSet>();
945 usage_map[UsageBB]->insert(Inst);
951 // Pick one usage that is in LaterInst's block and that dominates 'LaterInst'.
952 auto* LaterBB = LaterInst->getParent();
953 auto& usage_set = usage_map[LaterBB];
954 Instruction* usage_inst = nullptr;
955 for (auto* inst : *usage_set) {
956 if (DT->dominates(inst, LaterInst)) {
962 assert(usage_inst && "The usage instruction in the same block but after the "
963 "later instruction");
967 // XXX-comment: Returns whether the code has been changed.
968 bool AddFakeConditionalBranchAfterMonotonicLoads(
969 SmallSet<LoadInst*, 1>& MonotonicLoadInsts, DominatorTree* DT) {
970 bool Changed = false;
971 while (!MonotonicLoadInsts.empty()) {
972 auto* LI = *MonotonicLoadInsts.begin();
973 MonotonicLoadInsts.erase(LI);
974 SmallVector<BasicBlock*, 2> ChainedBB;
975 auto* FirstInst = findFirstStoreCondBranchInst(LI, &ChainedBB);
976 if (FirstInst != nullptr) {
977 if (FirstInst->getOpcode() == Instruction::Store) {
978 if (StoreAddressDependOnValue(dyn_cast<StoreInst>(FirstInst), LI)) {
981 } else if (FirstInst->getOpcode() == Instruction::Br) {
982 if (ConditionalBranchDependsOnValue(dyn_cast<BranchInst>(FirstInst),
987 dbgs() << "FirstInst=" << *FirstInst << "\n";
988 assert(false && "findFirstStoreCondBranchInst() should return a "
989 "store/condition branch instruction");
993 // We really need to process the relaxed load now.
994 StoreInst* SI = nullptr;;
995 if (FirstInst && (SI = dyn_cast<StoreInst>(FirstInst))) {
996 // For immediately coming stores, taint the address of the store.
997 if (SI->getParent() == LI->getParent() || DT->dominates(LI, SI)) {
998 TaintRelaxedLoads(LI, SI);
1002 findMostRecentDependenceUsage(LI, FirstInst, &ChainedBB, DT);
1004 LI->setOrdering(Acquire);
1007 TaintRelaxedLoads(Inst, SI);
1012 // No upcoming branch
1014 TaintRelaxedLoads(LI, nullptr);
1017 // For immediately coming branch, directly add a fake branch.
1018 if (FirstInst->getParent() == LI->getParent() ||
1019 DT->dominates(LI, FirstInst)) {
1020 TaintRelaxedLoads(LI, FirstInst);
1024 findMostRecentDependenceUsage(LI, FirstInst, &ChainedBB, DT);
1026 TaintRelaxedLoads(Inst, FirstInst);
1028 LI->setOrdering(Acquire);
1038 /**** Implementations of public methods for dependence tainting ****/
1039 Value* GetUntaintedAddress(Value* CurrentAddress) {
1040 auto* OrAddress = getOrAddress(CurrentAddress);
1041 if (OrAddress == nullptr) {
1042 // Is it tainted by a select instruction?
1043 auto* Inst = dyn_cast<Instruction>(CurrentAddress);
1044 if (nullptr != Inst && Inst->getOpcode() == Instruction::Select) {
1045 // A selection instruction.
1046 if (Inst->getOperand(1) == Inst->getOperand(2)) {
1047 return Inst->getOperand(1);
1051 return CurrentAddress;
1053 Value* ActualAddress = nullptr;
1055 auto* CastToInt = dyn_cast<Instruction>(OrAddress->getOperand(1));
1056 if (CastToInt && CastToInt->getOpcode() == Instruction::PtrToInt) {
1057 return CastToInt->getOperand(0);
1059 // This should be a IntToPtr constant expression.
1060 ConstantExpr* PtrToIntExpr =
1061 dyn_cast<ConstantExpr>(OrAddress->getOperand(1));
1062 if (PtrToIntExpr && PtrToIntExpr->getOpcode() == Instruction::PtrToInt) {
1063 return PtrToIntExpr->getOperand(0);
1067 // Looks like it's not been dependence-tainted. Returns itself.
1068 return CurrentAddress;
1071 MemoryLocation GetUntaintedMemoryLocation(StoreInst* SI) {
1073 SI->getAAMetadata(AATags);
1074 const auto& DL = SI->getModule()->getDataLayout();
1075 const auto* OriginalAddr = GetUntaintedAddress(SI->getPointerOperand());
1076 DEBUG(if (OriginalAddr != SI->getPointerOperand()) {
1077 dbgs() << "[GetUntaintedMemoryLocation]\n"
1078 << "Storing address: " << *SI->getPointerOperand()
1079 << "\nUntainted address: " << *OriginalAddr << "\n";
1081 return MemoryLocation(OriginalAddr,
1082 DL.getTypeStoreSize(SI->getValueOperand()->getType()),
1086 bool TaintDependenceToStore(StoreInst* SI, Value* DepVal) {
1087 if (dependenceSetInclusion(SI, DepVal)) {
1091 bool tainted = taintStoreAddress(SI, DepVal);
1096 bool TaintDependenceToStoreAddress(StoreInst* SI, Value* DepVal) {
1097 if (dependenceSetInclusion(SI->getPointerOperand(), DepVal)) {
1101 bool tainted = taintStoreAddress(SI, DepVal);
1106 bool CompressTaintedStore(BasicBlock* BB) {
1107 // This function looks for windows of adajcent stores in 'BB' that satisfy the
1108 // following condition (and then do optimization):
1109 // *Addr(d1) = v1, d1 is a condition and is the only dependence the store's
1110 // address depends on && Dep(v1) includes Dep(d1);
1111 // *Addr(d2) = v2, d2 is a condition and is the only dependnece the store's
1112 // address depends on && Dep(v2) includes Dep(d2) &&
1113 // Dep(d2) includes Dep(d1);
1115 // *Addr(dN) = vN, dN is a condition and is the only dependence the store's
1116 // address depends on && Dep(dN) includes Dep(d"N-1").
1118 // As a result, Dep(dN) includes [Dep(d1) V ... V Dep(d"N-1")], so we can
1119 // safely transform the above to the following. In between these stores, we
1120 // can omit untainted stores to the same address 'Addr' since they internally
1121 // have dependence on the previous stores on the same address.
1126 for (auto BI = BB->begin(), BE = BB->end(); BI != BE; BI++) {
1127 // Look for the first store in such a window of adajacent stores.
1128 auto* FirstSI = dyn_cast<StoreInst>(&*BI);
1133 // The first store in the window must be tainted.
1134 auto* UntaintedAddress = GetUntaintedAddress(FirstSI->getPointerOperand());
1135 if (UntaintedAddress == FirstSI->getPointerOperand()) {
1139 // The first store's address must directly depend on and only depend on a
1141 auto* FirstSIDepCond = getConditionDependence(FirstSI->getPointerOperand());
1142 if (nullptr == FirstSIDepCond) {
1146 // Dep(first store's storing value) includes Dep(tainted dependence).
1147 if (!dependenceSetInclusion(FirstSI->getValueOperand(), FirstSIDepCond)) {
1151 // Look for subsequent stores to the same address that satisfy the condition
1152 // of "compressing the dependence".
1153 SmallVector<StoreInst*, 8> AdajacentStores;
1154 AdajacentStores.push_back(FirstSI);
1155 auto BII = BasicBlock::iterator(FirstSI);
1156 for (BII++; BII != BE; BII++) {
1157 auto* CurrSI = dyn_cast<StoreInst>(&*BII);
1159 if (BII->mayHaveSideEffects()) {
1160 // Be conservative. Instructions with side effects are similar to
1167 auto* OrigAddress = GetUntaintedAddress(CurrSI->getPointerOperand());
1168 auto* CurrSIDepCond = getConditionDependence(CurrSI->getPointerOperand());
1169 // All other stores must satisfy either:
1170 // A. 'CurrSI' is an untainted store to the same address, or
1171 // B. the combination of the following 5 subconditions:
1173 // 2. Untainted address is the same as the group's address;
1174 // 3. The address is tainted with a sole value which is a condition;
1175 // 4. The storing value depends on the condition in 3.
1176 // 5. The condition in 3 depends on the previous stores dependence
1179 // Condition A. Should ignore this store directly.
1180 if (OrigAddress == CurrSI->getPointerOperand() &&
1181 OrigAddress == UntaintedAddress) {
1184 // Check condition B.
1185 Value* Cond = nullptr;
1186 if (OrigAddress == CurrSI->getPointerOperand() ||
1187 OrigAddress != UntaintedAddress || CurrSIDepCond == nullptr ||
1188 !dependenceSetInclusion(CurrSI->getValueOperand(), CurrSIDepCond)) {
1189 // Check condition 1, 2, 3 & 4.
1193 // Check condition 5.
1194 StoreInst* PrevSI = AdajacentStores[AdajacentStores.size() - 1];
1195 auto* PrevSIDepCond = getConditionDependence(PrevSI->getPointerOperand());
1196 assert(PrevSIDepCond &&
1197 "Store in the group must already depend on a condtion");
1198 if (!dependenceSetInclusion(CurrSIDepCond, PrevSIDepCond)) {
1202 AdajacentStores.push_back(CurrSI);
1205 if (AdajacentStores.size() == 1) {
1206 // The outer loop should keep looking from the next store.
1210 // Now we have such a group of tainted stores to the same address.
1211 DEBUG(dbgs() << "[CompressTaintedStore]\n");
1212 DEBUG(dbgs() << "Original BB\n");
1213 DEBUG(dbgs() << *BB << '\n');
1214 auto* LastSI = AdajacentStores[AdajacentStores.size() - 1];
1215 for (unsigned i = 0; i < AdajacentStores.size() - 1; ++i) {
1216 auto* SI = AdajacentStores[i];
1218 // Use the original address for stores before the last one.
1219 SI->setOperand(1, UntaintedAddress);
1221 DEBUG(dbgs() << "Store address has been reversed: " << *SI << '\n';);
1223 // XXX-comment: Try to make the last store use fewer registers.
1224 // If LastSI's storing value is a select based on the condition with which
1225 // its address is tainted, transform the tainted address to a select
1226 // instruction, as follows:
1227 // r1 = Select Cond ? A : B
1232 // r1 = Select Cond ? A : B
1233 // r2 = Select Cond ? Addr : Addr
1235 // The idea is that both Select instructions depend on the same condition,
1236 // so hopefully the backend can generate two cmov instructions for them (and
1237 // this saves the number of registers needed).
1238 auto* LastSIDep = getConditionDependence(LastSI->getPointerOperand());
1239 auto* LastSIValue = dyn_cast<Instruction>(LastSI->getValueOperand());
1240 if (LastSIValue && LastSIValue->getOpcode() == Instruction::Select &&
1241 LastSIValue->getOperand(0) == LastSIDep) {
1242 // XXX-comment: Maybe it's better for us to just leave it as an and/or
1243 // dependence pattern.
1245 IRBuilder<true, NoFolder> Builder(LastSI);
1247 Builder.CreateSelect(LastSIDep, UntaintedAddress, UntaintedAddress);
1248 LastSI->setOperand(1, Address);
1249 DEBUG(dbgs() << "The last store becomes :" << *LastSI << "\n\n";);
1257 bool PassDependenceToStore(Value* OldAddress, StoreInst* NewStore) {
1258 Value* OldDep = getDependence(OldAddress);
1259 // Return false when there's no dependence to pass from the OldAddress.
1264 // No need to pass the dependence to NewStore's address if it already depends
1265 // on whatever 'OldAddress' depends on.
1266 if (StoreAddressDependOnValue(NewStore, OldDep)) {
1269 return taintStoreAddress(NewStore, OldAddress);
1272 SmallSet<Value*, 8> FindDependence(Value* Val) {
1273 SmallSet<Value*, 8> DepSet;
1274 recursivelyFindDependence(&DepSet, Val, true /*Only insert leaf nodes*/);
1278 bool StoreAddressDependOnValue(StoreInst* SI, Value* DepVal) {
1279 return dependenceSetInclusion(SI->getPointerOperand(), DepVal);
1282 bool StoreDependOnValue(StoreInst* SI, Value* Dep) {
1283 return dependenceSetInclusion(SI, Dep);
1290 bool CodeGenPrepare::runOnFunction(Function &F) {
1291 bool EverMadeChange = false;
1293 if (skipOptnoneFunction(F))
1296 DL = &F.getParent()->getDataLayout();
1298 // Clear per function information.
1299 InsertedInsts.clear();
1300 PromotedInsts.clear();
1304 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
1305 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1306 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1307 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1308 OptSize = F.optForSize();
1310 /// This optimization identifies DIV instructions that can be
1311 /// profitably bypassed and carried out with a shorter, faster divide.
1312 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
1313 const DenseMap<unsigned int, unsigned int> &BypassWidths =
1314 TLI->getBypassSlowDivWidths();
1315 BasicBlock* BB = &*F.begin();
1316 while (BB != nullptr) {
1317 // bypassSlowDivision may create new BBs, but we don't want to reapply the
1318 // optimization to those blocks.
1319 BasicBlock* Next = BB->getNextNode();
1320 EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
1325 // Eliminate blocks that contain only PHI nodes and an
1326 // unconditional branch.
1327 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
1329 // llvm.dbg.value is far away from the value then iSel may not be able
1330 // handle it properly. iSel will drop llvm.dbg.value if it can not
1331 // find a node corresponding to the value.
1332 EverMadeChange |= placeDbgValues(F);
1334 // If there is a mask, compare against zero, and branch that can be combined
1335 // into a single target instruction, push the mask and compare into branch
1336 // users. Do this before OptimizeBlock -> OptimizeInst ->
1337 // OptimizeCmpExpression, which perturbs the pattern being searched for.
1338 if (!DisableBranchOpts) {
1339 EverMadeChange |= sinkAndCmp(F);
1340 EverMadeChange |= splitBranchCondition(F);
1343 bool MadeChange = true;
1344 while (MadeChange) {
1346 for (Function::iterator I = F.begin(); I != F.end(); ) {
1347 BasicBlock *BB = &*I++;
1348 bool ModifiedDTOnIteration = false;
1349 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
1351 // Restart BB iteration if the dominator tree of the Function was changed
1352 if (ModifiedDTOnIteration)
1355 EverMadeChange |= MadeChange;
1360 if (!DisableBranchOpts) {
1362 SmallPtrSet<BasicBlock*, 8> WorkList;
1363 for (BasicBlock &BB : F) {
1364 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
1365 MadeChange |= ConstantFoldTerminator(&BB, true);
1366 if (!MadeChange) continue;
1368 for (SmallVectorImpl<BasicBlock*>::iterator
1369 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
1370 if (pred_begin(*II) == pred_end(*II))
1371 WorkList.insert(*II);
1374 // Delete the dead blocks and any of their dead successors.
1375 MadeChange |= !WorkList.empty();
1376 while (!WorkList.empty()) {
1377 BasicBlock *BB = *WorkList.begin();
1379 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
1381 DeleteDeadBlock(BB);
1383 for (SmallVectorImpl<BasicBlock*>::iterator
1384 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
1385 if (pred_begin(*II) == pred_end(*II))
1386 WorkList.insert(*II);
1389 // Merge pairs of basic blocks with unconditional branches, connected by
1391 if (EverMadeChange || MadeChange)
1392 MadeChange |= eliminateFallThrough(F);
1394 EverMadeChange |= MadeChange;
1397 if (!DisableGCOpts) {
1398 SmallVector<Instruction *, 2> Statepoints;
1399 for (BasicBlock &BB : F)
1400 for (Instruction &I : BB)
1401 if (isStatepoint(I))
1402 Statepoints.push_back(&I);
1403 for (auto &I : Statepoints)
1404 EverMadeChange |= simplifyOffsetableRelocate(*I);
1407 // XXX-comment: Delay dealing with relaxed loads in this function to avoid
1408 // further changes done by other passes (e.g., SimplifyCFG).
1409 // Collect all the relaxed loads.
1410 SmallSet<LoadInst*, 1> MonotonicLoadInsts;
1411 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
1412 if (I->isAtomic()) {
1413 switch (I->getOpcode()) {
1414 case Instruction::Load: {
1415 auto* LI = dyn_cast<LoadInst>(&*I);
1416 if (LI->getOrdering() == Monotonic) {
1417 MonotonicLoadInsts.insert(LI);
1428 AddFakeConditionalBranchAfterMonotonicLoads(MonotonicLoadInsts, DT);
1430 return EverMadeChange;
1433 /// Merge basic blocks which are connected by a single edge, where one of the
1434 /// basic blocks has a single successor pointing to the other basic block,
1435 /// which has a single predecessor.
1436 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
1437 bool Changed = false;
1438 // Scan all of the blocks in the function, except for the entry block.
1439 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
1440 BasicBlock *BB = &*I++;
1441 // If the destination block has a single pred, then this is a trivial
1442 // edge, just collapse it.
1443 BasicBlock *SinglePred = BB->getSinglePredecessor();
1445 // Don't merge if BB's address is taken.
1446 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
1448 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
1449 if (Term && !Term->isConditional()) {
1451 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
1452 // Remember if SinglePred was the entry block of the function.
1453 // If so, we will need to move BB back to the entry position.
1454 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
1455 MergeBasicBlockIntoOnlyPred(BB, nullptr);
1457 if (isEntry && BB != &BB->getParent()->getEntryBlock())
1458 BB->moveBefore(&BB->getParent()->getEntryBlock());
1460 // We have erased a block. Update the iterator.
1461 I = BB->getIterator();
1467 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
1468 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
1469 /// edges in ways that are non-optimal for isel. Start by eliminating these
1470 /// blocks so we can split them the way we want them.
1471 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
1472 bool MadeChange = false;
1473 // Note that this intentionally skips the entry block.
1474 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
1475 BasicBlock *BB = &*I++;
1476 // If this block doesn't end with an uncond branch, ignore it.
1477 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
1478 if (!BI || !BI->isUnconditional())
1481 // If the instruction before the branch (skipping debug info) isn't a phi
1482 // node, then other stuff is happening here.
1483 BasicBlock::iterator BBI = BI->getIterator();
1484 if (BBI != BB->begin()) {
1486 while (isa<DbgInfoIntrinsic>(BBI)) {
1487 if (BBI == BB->begin())
1491 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
1495 // Do not break infinite loops.
1496 BasicBlock *DestBB = BI->getSuccessor(0);
1500 if (!canMergeBlocks(BB, DestBB))
1503 eliminateMostlyEmptyBlock(BB);
1509 /// Return true if we can merge BB into DestBB if there is a single
1510 /// unconditional branch between them, and BB contains no other non-phi
1512 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
1513 const BasicBlock *DestBB) const {
1514 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
1515 // the successor. If there are more complex condition (e.g. preheaders),
1516 // don't mess around with them.
1517 BasicBlock::const_iterator BBI = BB->begin();
1518 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
1519 for (const User *U : PN->users()) {
1520 const Instruction *UI = cast<Instruction>(U);
1521 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
1523 // IfUser is inside DestBB block and it is a PHINode then check
1524 // incoming value. If incoming value is not from BB then this is
1525 // a complex condition (e.g. preheaders) we want to avoid here.
1526 if (UI->getParent() == DestBB) {
1527 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
1528 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
1529 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
1530 if (Insn && Insn->getParent() == BB &&
1531 Insn->getParent() != UPN->getIncomingBlock(I))
1538 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
1539 // and DestBB may have conflicting incoming values for the block. If so, we
1540 // can't merge the block.
1541 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
1542 if (!DestBBPN) return true; // no conflict.
1544 // Collect the preds of BB.
1545 SmallPtrSet<const BasicBlock*, 16> BBPreds;
1546 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
1547 // It is faster to get preds from a PHI than with pred_iterator.
1548 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1549 BBPreds.insert(BBPN->getIncomingBlock(i));
1551 BBPreds.insert(pred_begin(BB), pred_end(BB));
1554 // Walk the preds of DestBB.
1555 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
1556 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
1557 if (BBPreds.count(Pred)) { // Common predecessor?
1558 BBI = DestBB->begin();
1559 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
1560 const Value *V1 = PN->getIncomingValueForBlock(Pred);
1561 const Value *V2 = PN->getIncomingValueForBlock(BB);
1563 // If V2 is a phi node in BB, look up what the mapped value will be.
1564 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
1565 if (V2PN->getParent() == BB)
1566 V2 = V2PN->getIncomingValueForBlock(Pred);
1568 // If there is a conflict, bail out.
1569 if (V1 != V2) return false;
1578 /// Eliminate a basic block that has only phi's and an unconditional branch in
1580 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
1581 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1582 BasicBlock *DestBB = BI->getSuccessor(0);
1584 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
1586 // If the destination block has a single pred, then this is a trivial edge,
1587 // just collapse it.
1588 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
1589 if (SinglePred != DestBB) {
1590 // Remember if SinglePred was the entry block of the function. If so, we
1591 // will need to move BB back to the entry position.
1592 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
1593 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
1595 if (isEntry && BB != &BB->getParent()->getEntryBlock())
1596 BB->moveBefore(&BB->getParent()->getEntryBlock());
1598 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1603 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
1604 // to handle the new incoming edges it is about to have.
1606 for (BasicBlock::iterator BBI = DestBB->begin();
1607 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1608 // Remove the incoming value for BB, and remember it.
1609 Value *InVal = PN->removeIncomingValue(BB, false);
1611 // Two options: either the InVal is a phi node defined in BB or it is some
1612 // value that dominates BB.
1613 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
1614 if (InValPhi && InValPhi->getParent() == BB) {
1615 // Add all of the input values of the input PHI as inputs of this phi.
1616 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
1617 PN->addIncoming(InValPhi->getIncomingValue(i),
1618 InValPhi->getIncomingBlock(i));
1620 // Otherwise, add one instance of the dominating value for each edge that
1621 // we will be adding.
1622 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
1623 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1624 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
1626 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
1627 PN->addIncoming(InVal, *PI);
1632 // The PHIs are now updated, change everything that refers to BB to use
1633 // DestBB and remove BB.
1634 BB->replaceAllUsesWith(DestBB);
1635 BB->eraseFromParent();
1638 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1641 // Computes a map of base pointer relocation instructions to corresponding
1642 // derived pointer relocation instructions given a vector of all relocate calls
1643 static void computeBaseDerivedRelocateMap(
1644 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
1645 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
1647 // Collect information in two maps: one primarily for locating the base object
1648 // while filling the second map; the second map is the final structure holding
1649 // a mapping between Base and corresponding Derived relocate calls
1650 DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
1651 for (auto *ThisRelocate : AllRelocateCalls) {
1652 auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
1653 ThisRelocate->getDerivedPtrIndex());
1654 RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
1656 for (auto &Item : RelocateIdxMap) {
1657 std::pair<unsigned, unsigned> Key = Item.first;
1658 if (Key.first == Key.second)
1659 // Base relocation: nothing to insert
1662 GCRelocateInst *I = Item.second;
1663 auto BaseKey = std::make_pair(Key.first, Key.first);
1665 // We're iterating over RelocateIdxMap so we cannot modify it.
1666 auto MaybeBase = RelocateIdxMap.find(BaseKey);
1667 if (MaybeBase == RelocateIdxMap.end())
1668 // TODO: We might want to insert a new base object relocate and gep off
1669 // that, if there are enough derived object relocates.
1672 RelocateInstMap[MaybeBase->second].push_back(I);
1676 // Accepts a GEP and extracts the operands into a vector provided they're all
1677 // small integer constants
1678 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
1679 SmallVectorImpl<Value *> &OffsetV) {
1680 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
1681 // Only accept small constant integer operands
1682 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
1683 if (!Op || Op->getZExtValue() > 20)
1687 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
1688 OffsetV.push_back(GEP->getOperand(i));
1692 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
1693 // replace, computes a replacement, and affects it.
1695 simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
1696 const SmallVectorImpl<GCRelocateInst *> &Targets) {
1697 bool MadeChange = false;
1698 for (GCRelocateInst *ToReplace : Targets) {
1699 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
1700 "Not relocating a derived object of the original base object");
1701 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
1702 // A duplicate relocate call. TODO: coalesce duplicates.
1706 if (RelocatedBase->getParent() != ToReplace->getParent()) {
1707 // Base and derived relocates are in different basic blocks.
1708 // In this case transform is only valid when base dominates derived
1709 // relocate. However it would be too expensive to check dominance
1710 // for each such relocate, so we skip the whole transformation.
1714 Value *Base = ToReplace->getBasePtr();
1715 auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
1716 if (!Derived || Derived->getPointerOperand() != Base)
1719 SmallVector<Value *, 2> OffsetV;
1720 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
1723 // Create a Builder and replace the target callsite with a gep
1724 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
1726 // Insert after RelocatedBase
1727 IRBuilder<> Builder(RelocatedBase->getNextNode());
1728 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1730 // If gc_relocate does not match the actual type, cast it to the right type.
1731 // In theory, there must be a bitcast after gc_relocate if the type does not
1732 // match, and we should reuse it to get the derived pointer. But it could be
1736 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1741 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1745 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
1746 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
1748 // In this case, we can not find the bitcast any more. So we insert a new bitcast
1749 // no matter there is already one or not. In this way, we can handle all cases, and
1750 // the extra bitcast should be optimized away in later passes.
1751 Value *ActualRelocatedBase = RelocatedBase;
1752 if (RelocatedBase->getType() != Base->getType()) {
1753 ActualRelocatedBase =
1754 Builder.CreateBitCast(RelocatedBase, Base->getType());
1756 Value *Replacement = Builder.CreateGEP(
1757 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
1758 Replacement->takeName(ToReplace);
1759 // If the newly generated derived pointer's type does not match the original derived
1760 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
1761 Value *ActualReplacement = Replacement;
1762 if (Replacement->getType() != ToReplace->getType()) {
1764 Builder.CreateBitCast(Replacement, ToReplace->getType());
1766 ToReplace->replaceAllUsesWith(ActualReplacement);
1767 ToReplace->eraseFromParent();
1777 // %ptr = gep %base + 15
1778 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1779 // %base' = relocate(%tok, i32 4, i32 4)
1780 // %ptr' = relocate(%tok, i32 4, i32 5)
1781 // %val = load %ptr'
1786 // %ptr = gep %base + 15
1787 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1788 // %base' = gc.relocate(%tok, i32 4, i32 4)
1789 // %ptr' = gep %base' + 15
1790 // %val = load %ptr'
1791 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
1792 bool MadeChange = false;
1793 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1795 for (auto *U : I.users())
1796 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1797 // Collect all the relocate calls associated with a statepoint
1798 AllRelocateCalls.push_back(Relocate);
1800 // We need atleast one base pointer relocation + one derived pointer
1801 // relocation to mangle
1802 if (AllRelocateCalls.size() < 2)
1805 // RelocateInstMap is a mapping from the base relocate instruction to the
1806 // corresponding derived relocate instructions
1807 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
1808 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1809 if (RelocateInstMap.empty())
1812 for (auto &Item : RelocateInstMap)
1813 // Item.first is the RelocatedBase to offset against
1814 // Item.second is the vector of Targets to replace
1815 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1819 /// SinkCast - Sink the specified cast instruction into its user blocks
1820 static bool SinkCast(CastInst *CI) {
1821 BasicBlock *DefBB = CI->getParent();
1823 /// InsertedCasts - Only insert a cast in each block once.
1824 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1826 bool MadeChange = false;
1827 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1829 Use &TheUse = UI.getUse();
1830 Instruction *User = cast<Instruction>(*UI);
1832 // Figure out which BB this cast is used in. For PHI's this is the
1833 // appropriate predecessor block.
1834 BasicBlock *UserBB = User->getParent();
1835 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1836 UserBB = PN->getIncomingBlock(TheUse);
1839 // Preincrement use iterator so we don't invalidate it.
1842 // If the block selected to receive the cast is an EH pad that does not
1843 // allow non-PHI instructions before the terminator, we can't sink the
1845 if (UserBB->getTerminator()->isEHPad())
1848 // If this user is in the same block as the cast, don't change the cast.
1849 if (UserBB == DefBB) continue;
1851 // If we have already inserted a cast into this block, use it.
1852 CastInst *&InsertedCast = InsertedCasts[UserBB];
1854 if (!InsertedCast) {
1855 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1856 assert(InsertPt != UserBB->end());
1857 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1858 CI->getType(), "", &*InsertPt);
1861 // Replace a use of the cast with a use of the new cast.
1862 TheUse = InsertedCast;
1867 // If we removed all uses, nuke the cast.
1868 if (CI->use_empty()) {
1869 CI->eraseFromParent();
1876 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1877 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1878 /// reduce the number of virtual registers that must be created and coalesced.
1880 /// Return true if any changes are made.
1882 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1883 const DataLayout &DL) {
1884 // If this is a noop copy,
1885 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1886 EVT DstVT = TLI.getValueType(DL, CI->getType());
1888 // This is an fp<->int conversion?
1889 if (SrcVT.isInteger() != DstVT.isInteger())
1892 // If this is an extension, it will be a zero or sign extension, which
1894 if (SrcVT.bitsLT(DstVT)) return false;
1896 // If these values will be promoted, find out what they will be promoted
1897 // to. This helps us consider truncates on PPC as noop copies when they
1899 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1900 TargetLowering::TypePromoteInteger)
1901 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1902 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1903 TargetLowering::TypePromoteInteger)
1904 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1906 // If, after promotion, these are the same types, this is a noop copy.
1910 return SinkCast(CI);
1913 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
1916 /// Return true if any changes were made.
1917 static bool CombineUAddWithOverflow(CmpInst *CI) {
1921 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
1924 Type *Ty = AddI->getType();
1925 if (!isa<IntegerType>(Ty))
1928 // We don't want to move around uses of condition values this late, so we we
1929 // check if it is legal to create the call to the intrinsic in the basic
1930 // block containing the icmp:
1932 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
1936 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
1938 if (AddI->hasOneUse())
1939 assert(*AddI->user_begin() == CI && "expected!");
1942 Module *M = CI->getModule();
1943 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1945 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
1947 auto *UAddWithOverflow =
1948 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
1949 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
1951 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
1953 CI->replaceAllUsesWith(Overflow);
1954 AddI->replaceAllUsesWith(UAdd);
1955 CI->eraseFromParent();
1956 AddI->eraseFromParent();
1960 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1961 /// registers that must be created and coalesced. This is a clear win except on
1962 /// targets with multiple condition code registers (PowerPC), where it might
1963 /// lose; some adjustment may be wanted there.
1965 /// Return true if any changes are made.
1966 static bool SinkCmpExpression(CmpInst *CI) {
1967 BasicBlock *DefBB = CI->getParent();
1969 /// Only insert a cmp in each block once.
1970 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1972 bool MadeChange = false;
1973 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1975 Use &TheUse = UI.getUse();
1976 Instruction *User = cast<Instruction>(*UI);
1978 // Preincrement use iterator so we don't invalidate it.
1981 // Don't bother for PHI nodes.
1982 if (isa<PHINode>(User))
1985 // Figure out which BB this cmp is used in.
1986 BasicBlock *UserBB = User->getParent();
1988 // If this user is in the same block as the cmp, don't change the cmp.
1989 if (UserBB == DefBB) continue;
1991 // If we have already inserted a cmp into this block, use it.
1992 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1995 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1996 assert(InsertPt != UserBB->end());
1998 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1999 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
2002 // Replace a use of the cmp with a use of the new cmp.
2003 TheUse = InsertedCmp;
2008 // If we removed all uses, nuke the cmp.
2009 if (CI->use_empty()) {
2010 CI->eraseFromParent();
2017 static bool OptimizeCmpExpression(CmpInst *CI) {
2018 if (SinkCmpExpression(CI))
2021 if (CombineUAddWithOverflow(CI))
2027 /// Check if the candidates could be combined with a shift instruction, which
2029 /// 1. Truncate instruction
2030 /// 2. And instruction and the imm is a mask of the low bits:
2031 /// imm & (imm+1) == 0
2032 static bool isExtractBitsCandidateUse(Instruction *User) {
2033 if (!isa<TruncInst>(User)) {
2034 if (User->getOpcode() != Instruction::And ||
2035 !isa<ConstantInt>(User->getOperand(1)))
2038 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
2040 if ((Cimm & (Cimm + 1)).getBoolValue())
2046 /// Sink both shift and truncate instruction to the use of truncate's BB.
2048 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
2049 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
2050 const TargetLowering &TLI, const DataLayout &DL) {
2051 BasicBlock *UserBB = User->getParent();
2052 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
2053 TruncInst *TruncI = dyn_cast<TruncInst>(User);
2054 bool MadeChange = false;
2056 for (Value::user_iterator TruncUI = TruncI->user_begin(),
2057 TruncE = TruncI->user_end();
2058 TruncUI != TruncE;) {
2060 Use &TruncTheUse = TruncUI.getUse();
2061 Instruction *TruncUser = cast<Instruction>(*TruncUI);
2062 // Preincrement use iterator so we don't invalidate it.
2066 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
2070 // If the use is actually a legal node, there will not be an
2071 // implicit truncate.
2072 // FIXME: always querying the result type is just an
2073 // approximation; some nodes' legality is determined by the
2074 // operand or other means. There's no good way to find out though.
2075 if (TLI.isOperationLegalOrCustom(
2076 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
2079 // Don't bother for PHI nodes.
2080 if (isa<PHINode>(TruncUser))
2083 BasicBlock *TruncUserBB = TruncUser->getParent();
2085 if (UserBB == TruncUserBB)
2088 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
2089 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
2091 if (!InsertedShift && !InsertedTrunc) {
2092 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
2093 assert(InsertPt != TruncUserBB->end());
2095 if (ShiftI->getOpcode() == Instruction::AShr)
2096 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
2099 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
2103 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
2105 assert(TruncInsertPt != TruncUserBB->end());
2107 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
2108 TruncI->getType(), "", &*TruncInsertPt);
2112 TruncTheUse = InsertedTrunc;
2118 /// Sink the shift *right* instruction into user blocks if the uses could
2119 /// potentially be combined with this shift instruction and generate BitExtract
2120 /// instruction. It will only be applied if the architecture supports BitExtract
2121 /// instruction. Here is an example:
2123 /// %x.extract.shift = lshr i64 %arg1, 32
2125 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
2129 /// %x.extract.shift.1 = lshr i64 %arg1, 32
2130 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
2132 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
2134 /// Return true if any changes are made.
2135 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
2136 const TargetLowering &TLI,
2137 const DataLayout &DL) {
2138 BasicBlock *DefBB = ShiftI->getParent();
2140 /// Only insert instructions in each block once.
2141 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
2143 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
2145 bool MadeChange = false;
2146 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
2148 Use &TheUse = UI.getUse();
2149 Instruction *User = cast<Instruction>(*UI);
2150 // Preincrement use iterator so we don't invalidate it.
2153 // Don't bother for PHI nodes.
2154 if (isa<PHINode>(User))
2157 if (!isExtractBitsCandidateUse(User))
2160 BasicBlock *UserBB = User->getParent();
2162 if (UserBB == DefBB) {
2163 // If the shift and truncate instruction are in the same BB. The use of
2164 // the truncate(TruncUse) may still introduce another truncate if not
2165 // legal. In this case, we would like to sink both shift and truncate
2166 // instruction to the BB of TruncUse.
2169 // i64 shift.result = lshr i64 opnd, imm
2170 // trunc.result = trunc shift.result to i16
2173 // ----> We will have an implicit truncate here if the architecture does
2174 // not have i16 compare.
2175 // cmp i16 trunc.result, opnd2
2177 if (isa<TruncInst>(User) && shiftIsLegal
2178 // If the type of the truncate is legal, no trucate will be
2179 // introduced in other basic blocks.
2181 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
2183 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
2187 // If we have already inserted a shift into this block, use it.
2188 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
2190 if (!InsertedShift) {
2191 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2192 assert(InsertPt != UserBB->end());
2194 if (ShiftI->getOpcode() == Instruction::AShr)
2195 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
2198 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
2204 // Replace a use of the shift with a use of the new shift.
2205 TheUse = InsertedShift;
2208 // If we removed all uses, nuke the shift.
2209 if (ShiftI->use_empty())
2210 ShiftI->eraseFromParent();
2215 // Translate a masked load intrinsic like
2216 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
2217 // <16 x i1> %mask, <16 x i32> %passthru)
2218 // to a chain of basic blocks, with loading element one-by-one if
2219 // the appropriate mask bit is set
2221 // %1 = bitcast i8* %addr to i32*
2222 // %2 = extractelement <16 x i1> %mask, i32 0
2223 // %3 = icmp eq i1 %2, true
2224 // br i1 %3, label %cond.load, label %else
2226 //cond.load: ; preds = %0
2227 // %4 = getelementptr i32* %1, i32 0
2228 // %5 = load i32* %4
2229 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
2232 //else: ; preds = %0, %cond.load
2233 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
2234 // %7 = extractelement <16 x i1> %mask, i32 1
2235 // %8 = icmp eq i1 %7, true
2236 // br i1 %8, label %cond.load1, label %else2
2238 //cond.load1: ; preds = %else
2239 // %9 = getelementptr i32* %1, i32 1
2240 // %10 = load i32* %9
2241 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
2244 //else2: ; preds = %else, %cond.load1
2245 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
2246 // %12 = extractelement <16 x i1> %mask, i32 2
2247 // %13 = icmp eq i1 %12, true
2248 // br i1 %13, label %cond.load4, label %else5
2250 static void ScalarizeMaskedLoad(CallInst *CI) {
2251 Value *Ptr = CI->getArgOperand(0);
2252 Value *Alignment = CI->getArgOperand(1);
2253 Value *Mask = CI->getArgOperand(2);
2254 Value *Src0 = CI->getArgOperand(3);
2256 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2257 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
2258 assert(VecType && "Unexpected return type of masked load intrinsic");
2260 Type *EltTy = CI->getType()->getVectorElementType();
2262 IRBuilder<> Builder(CI->getContext());
2263 Instruction *InsertPt = CI;
2264 BasicBlock *IfBlock = CI->getParent();
2265 BasicBlock *CondBlock = nullptr;
2266 BasicBlock *PrevIfBlock = CI->getParent();
2268 Builder.SetInsertPoint(InsertPt);
2269 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2271 // Short-cut if the mask is all-true.
2272 bool IsAllOnesMask = isa<Constant>(Mask) &&
2273 cast<Constant>(Mask)->isAllOnesValue();
2275 if (IsAllOnesMask) {
2276 Value *NewI = Builder.CreateAlignedLoad(Ptr, AlignVal);
2277 CI->replaceAllUsesWith(NewI);
2278 CI->eraseFromParent();
2282 // Adjust alignment for the scalar instruction.
2283 AlignVal = std::min(AlignVal, VecType->getScalarSizeInBits()/8);
2284 // Bitcast %addr fron i8* to EltTy*
2286 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
2287 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
2288 unsigned VectorWidth = VecType->getNumElements();
2290 Value *UndefVal = UndefValue::get(VecType);
2292 // The result vector
2293 Value *VResult = UndefVal;
2295 if (isa<ConstantVector>(Mask)) {
2296 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2297 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2300 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2301 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
2302 VResult = Builder.CreateInsertElement(VResult, Load,
2303 Builder.getInt32(Idx));
2305 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
2306 CI->replaceAllUsesWith(NewI);
2307 CI->eraseFromParent();
2311 PHINode *Phi = nullptr;
2312 Value *PrevPhi = UndefVal;
2314 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2316 // Fill the "else" block, created in the previous iteration
2318 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
2319 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
2320 // %to_load = icmp eq i1 %mask_1, true
2321 // br i1 %to_load, label %cond.load, label %else
2324 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
2325 Phi->addIncoming(VResult, CondBlock);
2326 Phi->addIncoming(PrevPhi, PrevIfBlock);
2331 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
2332 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2333 ConstantInt::get(Predicate->getType(), 1));
2335 // Create "cond" block
2337 // %EltAddr = getelementptr i32* %1, i32 0
2338 // %Elt = load i32* %EltAddr
2339 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
2341 CondBlock = IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.load");
2342 Builder.SetInsertPoint(InsertPt);
2345 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2346 LoadInst *Load = Builder.CreateAlignedLoad(Gep, AlignVal);
2347 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
2349 // Create "else" block, fill it in the next iteration
2350 BasicBlock *NewIfBlock =
2351 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
2352 Builder.SetInsertPoint(InsertPt);
2353 Instruction *OldBr = IfBlock->getTerminator();
2354 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2355 OldBr->eraseFromParent();
2356 PrevIfBlock = IfBlock;
2357 IfBlock = NewIfBlock;
2360 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
2361 Phi->addIncoming(VResult, CondBlock);
2362 Phi->addIncoming(PrevPhi, PrevIfBlock);
2363 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
2364 CI->replaceAllUsesWith(NewI);
2365 CI->eraseFromParent();
2368 // Translate a masked store intrinsic, like
2369 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
2371 // to a chain of basic blocks, that stores element one-by-one if
2372 // the appropriate mask bit is set
2374 // %1 = bitcast i8* %addr to i32*
2375 // %2 = extractelement <16 x i1> %mask, i32 0
2376 // %3 = icmp eq i1 %2, true
2377 // br i1 %3, label %cond.store, label %else
2379 // cond.store: ; preds = %0
2380 // %4 = extractelement <16 x i32> %val, i32 0
2381 // %5 = getelementptr i32* %1, i32 0
2382 // store i32 %4, i32* %5
2385 // else: ; preds = %0, %cond.store
2386 // %6 = extractelement <16 x i1> %mask, i32 1
2387 // %7 = icmp eq i1 %6, true
2388 // br i1 %7, label %cond.store1, label %else2
2390 // cond.store1: ; preds = %else
2391 // %8 = extractelement <16 x i32> %val, i32 1
2392 // %9 = getelementptr i32* %1, i32 1
2393 // store i32 %8, i32* %9
2396 static void ScalarizeMaskedStore(CallInst *CI) {
2397 Value *Src = CI->getArgOperand(0);
2398 Value *Ptr = CI->getArgOperand(1);
2399 Value *Alignment = CI->getArgOperand(2);
2400 Value *Mask = CI->getArgOperand(3);
2402 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2403 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
2404 assert(VecType && "Unexpected data type in masked store intrinsic");
2406 Type *EltTy = VecType->getElementType();
2408 IRBuilder<> Builder(CI->getContext());
2409 Instruction *InsertPt = CI;
2410 BasicBlock *IfBlock = CI->getParent();
2411 Builder.SetInsertPoint(InsertPt);
2412 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2414 // Short-cut if the mask is all-true.
2415 bool IsAllOnesMask = isa<Constant>(Mask) &&
2416 cast<Constant>(Mask)->isAllOnesValue();
2418 if (IsAllOnesMask) {
2419 Builder.CreateAlignedStore(Src, Ptr, AlignVal);
2420 CI->eraseFromParent();
2424 // Adjust alignment for the scalar instruction.
2425 AlignVal = std::max(AlignVal, VecType->getScalarSizeInBits()/8);
2426 // Bitcast %addr fron i8* to EltTy*
2428 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
2429 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
2430 unsigned VectorWidth = VecType->getNumElements();
2432 if (isa<ConstantVector>(Mask)) {
2433 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2434 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2436 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
2438 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2439 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
2441 CI->eraseFromParent();
2445 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2447 // Fill the "else" block, created in the previous iteration
2449 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
2450 // %to_store = icmp eq i1 %mask_1, true
2451 // br i1 %to_store, label %cond.store, label %else
2453 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
2454 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2455 ConstantInt::get(Predicate->getType(), 1));
2457 // Create "cond" block
2459 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
2460 // %EltAddr = getelementptr i32* %1, i32 0
2461 // %store i32 %OneElt, i32* %EltAddr
2463 BasicBlock *CondBlock =
2464 IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.store");
2465 Builder.SetInsertPoint(InsertPt);
2467 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
2469 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2470 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
2472 // Create "else" block, fill it in the next iteration
2473 BasicBlock *NewIfBlock =
2474 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
2475 Builder.SetInsertPoint(InsertPt);
2476 Instruction *OldBr = IfBlock->getTerminator();
2477 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2478 OldBr->eraseFromParent();
2479 IfBlock = NewIfBlock;
2481 CI->eraseFromParent();
2484 // Translate a masked gather intrinsic like
2485 // <16 x i32 > @llvm.masked.gather.v16i32( <16 x i32*> %Ptrs, i32 4,
2486 // <16 x i1> %Mask, <16 x i32> %Src)
2487 // to a chain of basic blocks, with loading element one-by-one if
2488 // the appropriate mask bit is set
2490 // % Ptrs = getelementptr i32, i32* %base, <16 x i64> %ind
2491 // % Mask0 = extractelement <16 x i1> %Mask, i32 0
2492 // % ToLoad0 = icmp eq i1 % Mask0, true
2493 // br i1 % ToLoad0, label %cond.load, label %else
2496 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
2497 // % Load0 = load i32, i32* % Ptr0, align 4
2498 // % Res0 = insertelement <16 x i32> undef, i32 % Load0, i32 0
2502 // %res.phi.else = phi <16 x i32>[% Res0, %cond.load], [undef, % 0]
2503 // % Mask1 = extractelement <16 x i1> %Mask, i32 1
2504 // % ToLoad1 = icmp eq i1 % Mask1, true
2505 // br i1 % ToLoad1, label %cond.load1, label %else2
2508 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2509 // % Load1 = load i32, i32* % Ptr1, align 4
2510 // % Res1 = insertelement <16 x i32> %res.phi.else, i32 % Load1, i32 1
2513 // % Result = select <16 x i1> %Mask, <16 x i32> %res.phi.select, <16 x i32> %Src
2514 // ret <16 x i32> %Result
2515 static void ScalarizeMaskedGather(CallInst *CI) {
2516 Value *Ptrs = CI->getArgOperand(0);
2517 Value *Alignment = CI->getArgOperand(1);
2518 Value *Mask = CI->getArgOperand(2);
2519 Value *Src0 = CI->getArgOperand(3);
2521 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
2523 assert(VecType && "Unexpected return type of masked load intrinsic");
2525 IRBuilder<> Builder(CI->getContext());
2526 Instruction *InsertPt = CI;
2527 BasicBlock *IfBlock = CI->getParent();
2528 BasicBlock *CondBlock = nullptr;
2529 BasicBlock *PrevIfBlock = CI->getParent();
2530 Builder.SetInsertPoint(InsertPt);
2531 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2533 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2535 Value *UndefVal = UndefValue::get(VecType);
2537 // The result vector
2538 Value *VResult = UndefVal;
2539 unsigned VectorWidth = VecType->getNumElements();
2541 // Shorten the way if the mask is a vector of constants.
2542 bool IsConstMask = isa<ConstantVector>(Mask);
2545 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2546 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2548 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2549 "Ptr" + Twine(Idx));
2550 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
2551 "Load" + Twine(Idx));
2552 VResult = Builder.CreateInsertElement(VResult, Load,
2553 Builder.getInt32(Idx),
2554 "Res" + Twine(Idx));
2556 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
2557 CI->replaceAllUsesWith(NewI);
2558 CI->eraseFromParent();
2562 PHINode *Phi = nullptr;
2563 Value *PrevPhi = UndefVal;
2565 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2567 // Fill the "else" block, created in the previous iteration
2569 // %Mask1 = extractelement <16 x i1> %Mask, i32 1
2570 // %ToLoad1 = icmp eq i1 %Mask1, true
2571 // br i1 %ToLoad1, label %cond.load, label %else
2574 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
2575 Phi->addIncoming(VResult, CondBlock);
2576 Phi->addIncoming(PrevPhi, PrevIfBlock);
2581 Value *Predicate = Builder.CreateExtractElement(Mask,
2582 Builder.getInt32(Idx),
2583 "Mask" + Twine(Idx));
2584 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2585 ConstantInt::get(Predicate->getType(), 1),
2586 "ToLoad" + Twine(Idx));
2588 // Create "cond" block
2590 // %EltAddr = getelementptr i32* %1, i32 0
2591 // %Elt = load i32* %EltAddr
2592 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
2594 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
2595 Builder.SetInsertPoint(InsertPt);
2597 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2598 "Ptr" + Twine(Idx));
2599 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
2600 "Load" + Twine(Idx));
2601 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx),
2602 "Res" + Twine(Idx));
2604 // Create "else" block, fill it in the next iteration
2605 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2606 Builder.SetInsertPoint(InsertPt);
2607 Instruction *OldBr = IfBlock->getTerminator();
2608 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2609 OldBr->eraseFromParent();
2610 PrevIfBlock = IfBlock;
2611 IfBlock = NewIfBlock;
2614 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
2615 Phi->addIncoming(VResult, CondBlock);
2616 Phi->addIncoming(PrevPhi, PrevIfBlock);
2617 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
2618 CI->replaceAllUsesWith(NewI);
2619 CI->eraseFromParent();
2622 // Translate a masked scatter intrinsic, like
2623 // void @llvm.masked.scatter.v16i32(<16 x i32> %Src, <16 x i32*>* %Ptrs, i32 4,
2625 // to a chain of basic blocks, that stores element one-by-one if
2626 // the appropriate mask bit is set.
2628 // % Ptrs = getelementptr i32, i32* %ptr, <16 x i64> %ind
2629 // % Mask0 = extractelement <16 x i1> % Mask, i32 0
2630 // % ToStore0 = icmp eq i1 % Mask0, true
2631 // br i1 %ToStore0, label %cond.store, label %else
2634 // % Elt0 = extractelement <16 x i32> %Src, i32 0
2635 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
2636 // store i32 %Elt0, i32* % Ptr0, align 4
2640 // % Mask1 = extractelement <16 x i1> % Mask, i32 1
2641 // % ToStore1 = icmp eq i1 % Mask1, true
2642 // br i1 % ToStore1, label %cond.store1, label %else2
2645 // % Elt1 = extractelement <16 x i32> %Src, i32 1
2646 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2647 // store i32 % Elt1, i32* % Ptr1, align 4
2650 static void ScalarizeMaskedScatter(CallInst *CI) {
2651 Value *Src = CI->getArgOperand(0);
2652 Value *Ptrs = CI->getArgOperand(1);
2653 Value *Alignment = CI->getArgOperand(2);
2654 Value *Mask = CI->getArgOperand(3);
2656 assert(isa<VectorType>(Src->getType()) &&
2657 "Unexpected data type in masked scatter intrinsic");
2658 assert(isa<VectorType>(Ptrs->getType()) &&
2659 isa<PointerType>(Ptrs->getType()->getVectorElementType()) &&
2660 "Vector of pointers is expected in masked scatter intrinsic");
2662 IRBuilder<> Builder(CI->getContext());
2663 Instruction *InsertPt = CI;
2664 BasicBlock *IfBlock = CI->getParent();
2665 Builder.SetInsertPoint(InsertPt);
2666 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2668 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2669 unsigned VectorWidth = Src->getType()->getVectorNumElements();
2671 // Shorten the way if the mask is a vector of constants.
2672 bool IsConstMask = isa<ConstantVector>(Mask);
2675 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2676 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2678 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
2679 "Elt" + Twine(Idx));
2680 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2681 "Ptr" + Twine(Idx));
2682 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
2684 CI->eraseFromParent();
2687 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2688 // Fill the "else" block, created in the previous iteration
2690 // % Mask1 = extractelement <16 x i1> % Mask, i32 Idx
2691 // % ToStore = icmp eq i1 % Mask1, true
2692 // br i1 % ToStore, label %cond.store, label %else
2694 Value *Predicate = Builder.CreateExtractElement(Mask,
2695 Builder.getInt32(Idx),
2696 "Mask" + Twine(Idx));
2698 Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2699 ConstantInt::get(Predicate->getType(), 1),
2700 "ToStore" + Twine(Idx));
2702 // Create "cond" block
2704 // % Elt1 = extractelement <16 x i32> %Src, i32 1
2705 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2706 // %store i32 % Elt1, i32* % Ptr1
2708 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2709 Builder.SetInsertPoint(InsertPt);
2711 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
2712 "Elt" + Twine(Idx));
2713 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2714 "Ptr" + Twine(Idx));
2715 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
2717 // Create "else" block, fill it in the next iteration
2718 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2719 Builder.SetInsertPoint(InsertPt);
2720 Instruction *OldBr = IfBlock->getTerminator();
2721 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2722 OldBr->eraseFromParent();
2723 IfBlock = NewIfBlock;
2725 CI->eraseFromParent();
2728 /// If counting leading or trailing zeros is an expensive operation and a zero
2729 /// input is defined, add a check for zero to avoid calling the intrinsic.
2731 /// We want to transform:
2732 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
2736 /// %cmpz = icmp eq i64 %A, 0
2737 /// br i1 %cmpz, label %cond.end, label %cond.false
2739 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
2740 /// br label %cond.end
2742 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
2744 /// If the transform is performed, return true and set ModifiedDT to true.
2745 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
2746 const TargetLowering *TLI,
2747 const DataLayout *DL,
2752 // If a zero input is undefined, it doesn't make sense to despeculate that.
2753 if (match(CountZeros->getOperand(1), m_One()))
2756 // If it's cheap to speculate, there's nothing to do.
2757 auto IntrinsicID = CountZeros->getIntrinsicID();
2758 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
2759 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
2762 // Only handle legal scalar cases. Anything else requires too much work.
2763 Type *Ty = CountZeros->getType();
2764 unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
2765 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSize())
2768 // The intrinsic will be sunk behind a compare against zero and branch.
2769 BasicBlock *StartBlock = CountZeros->getParent();
2770 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
2772 // Create another block after the count zero intrinsic. A PHI will be added
2773 // in this block to select the result of the intrinsic or the bit-width
2774 // constant if the input to the intrinsic is zero.
2775 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
2776 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
2778 // Set up a builder to create a compare, conditional branch, and PHI.
2779 IRBuilder<> Builder(CountZeros->getContext());
2780 Builder.SetInsertPoint(StartBlock->getTerminator());
2781 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
2783 // Replace the unconditional branch that was created by the first split with
2784 // a compare against zero and a conditional branch.
2785 Value *Zero = Constant::getNullValue(Ty);
2786 Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
2787 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
2788 StartBlock->getTerminator()->eraseFromParent();
2790 // Create a PHI in the end block to select either the output of the intrinsic
2791 // or the bit width of the operand.
2792 Builder.SetInsertPoint(&EndBlock->front());
2793 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
2794 CountZeros->replaceAllUsesWith(PN);
2795 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
2796 PN->addIncoming(BitWidth, StartBlock);
2797 PN->addIncoming(CountZeros, CallBlock);
2799 // We are explicitly handling the zero case, so we can set the intrinsic's
2800 // undefined zero argument to 'true'. This will also prevent reprocessing the
2801 // intrinsic; we only despeculate when a zero input is defined.
2802 CountZeros->setArgOperand(1, Builder.getTrue());
2807 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool& ModifiedDT) {
2808 BasicBlock *BB = CI->getParent();
2810 // Lower inline assembly if we can.
2811 // If we found an inline asm expession, and if the target knows how to
2812 // lower it to normal LLVM code, do so now.
2813 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
2814 if (TLI->ExpandInlineAsm(CI)) {
2815 // Avoid invalidating the iterator.
2816 CurInstIterator = BB->begin();
2817 // Avoid processing instructions out of order, which could cause
2818 // reuse before a value is defined.
2822 // Sink address computing for memory operands into the block.
2823 if (optimizeInlineAsmInst(CI))
2827 // Align the pointer arguments to this call if the target thinks it's a good
2829 unsigned MinSize, PrefAlign;
2830 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
2831 for (auto &Arg : CI->arg_operands()) {
2832 // We want to align both objects whose address is used directly and
2833 // objects whose address is used in casts and GEPs, though it only makes
2834 // sense for GEPs if the offset is a multiple of the desired alignment and
2835 // if size - offset meets the size threshold.
2836 if (!Arg->getType()->isPointerTy())
2838 APInt Offset(DL->getPointerSizeInBits(
2839 cast<PointerType>(Arg->getType())->getAddressSpace()),
2841 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
2842 uint64_t Offset2 = Offset.getLimitedValue();
2843 if ((Offset2 & (PrefAlign-1)) != 0)
2846 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
2847 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
2848 AI->setAlignment(PrefAlign);
2849 // Global variables can only be aligned if they are defined in this
2850 // object (i.e. they are uniquely initialized in this object), and
2851 // over-aligning global variables that have an explicit section is
2854 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
2855 GV->getAlignment() < PrefAlign &&
2856 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
2858 GV->setAlignment(PrefAlign);
2860 // If this is a memcpy (or similar) then we may be able to improve the
2862 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
2863 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
2864 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
2865 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
2866 if (Align > MI->getAlignment())
2867 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
2871 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
2873 switch (II->getIntrinsicID()) {
2875 case Intrinsic::objectsize: {
2876 // Lower all uses of llvm.objectsize.*
2877 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
2878 Type *ReturnTy = CI->getType();
2879 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
2881 // Substituting this can cause recursive simplifications, which can
2882 // invalidate our iterator. Use a WeakVH to hold onto it in case this
2884 WeakVH IterHandle(&*CurInstIterator);
2886 replaceAndRecursivelySimplify(CI, RetVal,
2889 // If the iterator instruction was recursively deleted, start over at the
2890 // start of the block.
2891 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
2892 CurInstIterator = BB->begin();
2897 case Intrinsic::masked_load: {
2898 // Scalarize unsupported vector masked load
2899 if (!TTI->isLegalMaskedLoad(CI->getType())) {
2900 ScalarizeMaskedLoad(CI);
2906 case Intrinsic::masked_store: {
2907 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType())) {
2908 ScalarizeMaskedStore(CI);
2914 case Intrinsic::masked_gather: {
2915 if (!TTI->isLegalMaskedGather(CI->getType())) {
2916 ScalarizeMaskedGather(CI);
2922 case Intrinsic::masked_scatter: {
2923 if (!TTI->isLegalMaskedScatter(CI->getArgOperand(0)->getType())) {
2924 ScalarizeMaskedScatter(CI);
2930 case Intrinsic::aarch64_stlxr:
2931 case Intrinsic::aarch64_stxr: {
2932 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
2933 if (!ExtVal || !ExtVal->hasOneUse() ||
2934 ExtVal->getParent() == CI->getParent())
2936 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
2937 ExtVal->moveBefore(CI);
2938 // Mark this instruction as "inserted by CGP", so that other
2939 // optimizations don't touch it.
2940 InsertedInsts.insert(ExtVal);
2943 case Intrinsic::invariant_group_barrier:
2944 II->replaceAllUsesWith(II->getArgOperand(0));
2945 II->eraseFromParent();
2948 case Intrinsic::cttz:
2949 case Intrinsic::ctlz:
2950 // If counting zeros is expensive, try to avoid it.
2951 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
2955 // Unknown address space.
2956 // TODO: Target hook to pick which address space the intrinsic cares
2958 unsigned AddrSpace = ~0u;
2959 SmallVector<Value*, 2> PtrOps;
2961 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
2962 while (!PtrOps.empty())
2963 if (optimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
2968 // From here on out we're working with named functions.
2969 if (!CI->getCalledFunction()) return false;
2971 // Lower all default uses of _chk calls. This is very similar
2972 // to what InstCombineCalls does, but here we are only lowering calls
2973 // to fortified library functions (e.g. __memcpy_chk) that have the default
2974 // "don't know" as the objectsize. Anything else should be left alone.
2975 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
2976 if (Value *V = Simplifier.optimizeCall(CI)) {
2977 CI->replaceAllUsesWith(V);
2978 CI->eraseFromParent();
2984 /// Look for opportunities to duplicate return instructions to the predecessor
2985 /// to enable tail call optimizations. The case it is currently looking for is:
2988 /// %tmp0 = tail call i32 @f0()
2989 /// br label %return
2991 /// %tmp1 = tail call i32 @f1()
2992 /// br label %return
2994 /// %tmp2 = tail call i32 @f2()
2995 /// br label %return
2997 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
3005 /// %tmp0 = tail call i32 @f0()
3008 /// %tmp1 = tail call i32 @f1()
3011 /// %tmp2 = tail call i32 @f2()
3014 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
3018 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
3022 PHINode *PN = nullptr;
3023 BitCastInst *BCI = nullptr;
3024 Value *V = RI->getReturnValue();
3026 BCI = dyn_cast<BitCastInst>(V);
3028 V = BCI->getOperand(0);
3030 PN = dyn_cast<PHINode>(V);
3035 if (PN && PN->getParent() != BB)
3038 // It's not safe to eliminate the sign / zero extension of the return value.
3039 // See llvm::isInTailCallPosition().
3040 const Function *F = BB->getParent();
3041 AttributeSet CallerAttrs = F->getAttributes();
3042 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
3043 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
3046 // Make sure there are no instructions between the PHI and return, or that the
3047 // return is the first instruction in the block.
3049 BasicBlock::iterator BI = BB->begin();
3050 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
3052 // Also skip over the bitcast.
3057 BasicBlock::iterator BI = BB->begin();
3058 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
3063 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
3065 SmallVector<CallInst*, 4> TailCalls;
3067 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
3068 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
3069 // Make sure the phi value is indeed produced by the tail call.
3070 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
3071 TLI->mayBeEmittedAsTailCall(CI))
3072 TailCalls.push_back(CI);
3075 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
3076 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
3077 if (!VisitedBBs.insert(*PI).second)
3080 BasicBlock::InstListType &InstList = (*PI)->getInstList();
3081 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
3082 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
3083 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
3087 CallInst *CI = dyn_cast<CallInst>(&*RI);
3088 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
3089 TailCalls.push_back(CI);
3093 bool Changed = false;
3094 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
3095 CallInst *CI = TailCalls[i];
3098 // Conservatively require the attributes of the call to match those of the
3099 // return. Ignore noalias because it doesn't affect the call sequence.
3100 AttributeSet CalleeAttrs = CS.getAttributes();
3101 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
3102 removeAttribute(Attribute::NoAlias) !=
3103 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
3104 removeAttribute(Attribute::NoAlias))
3107 // Make sure the call instruction is followed by an unconditional branch to
3108 // the return block.
3109 BasicBlock *CallBB = CI->getParent();
3110 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
3111 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
3114 // Duplicate the return into CallBB.
3115 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
3116 ModifiedDT = Changed = true;
3120 // If we eliminated all predecessors of the block, delete the block now.
3121 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
3122 BB->eraseFromParent();
3127 //===----------------------------------------------------------------------===//
3128 // Memory Optimization
3129 //===----------------------------------------------------------------------===//
3133 /// This is an extended version of TargetLowering::AddrMode
3134 /// which holds actual Value*'s for register values.
3135 struct ExtAddrMode : public TargetLowering::AddrMode {
3138 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
3139 void print(raw_ostream &OS) const;
3142 bool operator==(const ExtAddrMode& O) const {
3143 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
3144 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
3145 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
3150 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
3156 void ExtAddrMode::print(raw_ostream &OS) const {
3157 bool NeedPlus = false;
3160 OS << (NeedPlus ? " + " : "")
3162 BaseGV->printAsOperand(OS, /*PrintType=*/false);
3167 OS << (NeedPlus ? " + " : "")
3173 OS << (NeedPlus ? " + " : "")
3175 BaseReg->printAsOperand(OS, /*PrintType=*/false);
3179 OS << (NeedPlus ? " + " : "")
3181 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
3187 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3188 void ExtAddrMode::dump() const {
3194 /// \brief This class provides transaction based operation on the IR.
3195 /// Every change made through this class is recorded in the internal state and
3196 /// can be undone (rollback) until commit is called.
3197 class TypePromotionTransaction {
3199 /// \brief This represents the common interface of the individual transaction.
3200 /// Each class implements the logic for doing one specific modification on
3201 /// the IR via the TypePromotionTransaction.
3202 class TypePromotionAction {
3204 /// The Instruction modified.
3208 /// \brief Constructor of the action.
3209 /// The constructor performs the related action on the IR.
3210 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
3212 virtual ~TypePromotionAction() {}
3214 /// \brief Undo the modification done by this action.
3215 /// When this method is called, the IR must be in the same state as it was
3216 /// before this action was applied.
3217 /// \pre Undoing the action works if and only if the IR is in the exact same
3218 /// state as it was directly after this action was applied.
3219 virtual void undo() = 0;
3221 /// \brief Advocate every change made by this action.
3222 /// When the results on the IR of the action are to be kept, it is important
3223 /// to call this function, otherwise hidden information may be kept forever.
3224 virtual void commit() {
3225 // Nothing to be done, this action is not doing anything.
3229 /// \brief Utility to remember the position of an instruction.
3230 class InsertionHandler {
3231 /// Position of an instruction.
3232 /// Either an instruction:
3233 /// - Is the first in a basic block: BB is used.
3234 /// - Has a previous instructon: PrevInst is used.
3236 Instruction *PrevInst;
3239 /// Remember whether or not the instruction had a previous instruction.
3240 bool HasPrevInstruction;
3243 /// \brief Record the position of \p Inst.
3244 InsertionHandler(Instruction *Inst) {
3245 BasicBlock::iterator It = Inst->getIterator();
3246 HasPrevInstruction = (It != (Inst->getParent()->begin()));
3247 if (HasPrevInstruction)
3248 Point.PrevInst = &*--It;
3250 Point.BB = Inst->getParent();
3253 /// \brief Insert \p Inst at the recorded position.
3254 void insert(Instruction *Inst) {
3255 if (HasPrevInstruction) {
3256 if (Inst->getParent())
3257 Inst->removeFromParent();
3258 Inst->insertAfter(Point.PrevInst);
3260 Instruction *Position = &*Point.BB->getFirstInsertionPt();
3261 if (Inst->getParent())
3262 Inst->moveBefore(Position);
3264 Inst->insertBefore(Position);
3269 /// \brief Move an instruction before another.
3270 class InstructionMoveBefore : public TypePromotionAction {
3271 /// Original position of the instruction.
3272 InsertionHandler Position;
3275 /// \brief Move \p Inst before \p Before.
3276 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
3277 : TypePromotionAction(Inst), Position(Inst) {
3278 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
3279 Inst->moveBefore(Before);
3282 /// \brief Move the instruction back to its original position.
3283 void undo() override {
3284 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
3285 Position.insert(Inst);
3289 /// \brief Set the operand of an instruction with a new value.
3290 class OperandSetter : public TypePromotionAction {
3291 /// Original operand of the instruction.
3293 /// Index of the modified instruction.
3297 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
3298 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
3299 : TypePromotionAction(Inst), Idx(Idx) {
3300 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
3301 << "for:" << *Inst << "\n"
3302 << "with:" << *NewVal << "\n");
3303 Origin = Inst->getOperand(Idx);
3304 Inst->setOperand(Idx, NewVal);
3307 /// \brief Restore the original value of the instruction.
3308 void undo() override {
3309 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
3310 << "for: " << *Inst << "\n"
3311 << "with: " << *Origin << "\n");
3312 Inst->setOperand(Idx, Origin);
3316 /// \brief Hide the operands of an instruction.
3317 /// Do as if this instruction was not using any of its operands.
3318 class OperandsHider : public TypePromotionAction {
3319 /// The list of original operands.
3320 SmallVector<Value *, 4> OriginalValues;
3323 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
3324 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
3325 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
3326 unsigned NumOpnds = Inst->getNumOperands();
3327 OriginalValues.reserve(NumOpnds);
3328 for (unsigned It = 0; It < NumOpnds; ++It) {
3329 // Save the current operand.
3330 Value *Val = Inst->getOperand(It);
3331 OriginalValues.push_back(Val);
3333 // We could use OperandSetter here, but that would imply an overhead
3334 // that we are not willing to pay.
3335 Inst->setOperand(It, UndefValue::get(Val->getType()));
3339 /// \brief Restore the original list of uses.
3340 void undo() override {
3341 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
3342 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
3343 Inst->setOperand(It, OriginalValues[It]);
3347 /// \brief Build a truncate instruction.
3348 class TruncBuilder : public TypePromotionAction {
3351 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
3353 /// trunc Opnd to Ty.
3354 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
3355 IRBuilder<> Builder(Opnd);
3356 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
3357 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
3360 /// \brief Get the built value.
3361 Value *getBuiltValue() { return Val; }
3363 /// \brief Remove the built instruction.
3364 void undo() override {
3365 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
3366 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3367 IVal->eraseFromParent();
3371 /// \brief Build a sign extension instruction.
3372 class SExtBuilder : public TypePromotionAction {
3375 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
3377 /// sext Opnd to Ty.
3378 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3379 : TypePromotionAction(InsertPt) {
3380 IRBuilder<> Builder(InsertPt);
3381 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
3382 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
3385 /// \brief Get the built value.
3386 Value *getBuiltValue() { return Val; }
3388 /// \brief Remove the built instruction.
3389 void undo() override {
3390 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
3391 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3392 IVal->eraseFromParent();
3396 /// \brief Build a zero extension instruction.
3397 class ZExtBuilder : public TypePromotionAction {
3400 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
3402 /// zext Opnd to Ty.
3403 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3404 : TypePromotionAction(InsertPt) {
3405 IRBuilder<> Builder(InsertPt);
3406 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
3407 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
3410 /// \brief Get the built value.
3411 Value *getBuiltValue() { return Val; }
3413 /// \brief Remove the built instruction.
3414 void undo() override {
3415 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
3416 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3417 IVal->eraseFromParent();
3421 /// \brief Mutate an instruction to another type.
3422 class TypeMutator : public TypePromotionAction {
3423 /// Record the original type.
3427 /// \brief Mutate the type of \p Inst into \p NewTy.
3428 TypeMutator(Instruction *Inst, Type *NewTy)
3429 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
3430 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
3432 Inst->mutateType(NewTy);
3435 /// \brief Mutate the instruction back to its original type.
3436 void undo() override {
3437 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
3439 Inst->mutateType(OrigTy);
3443 /// \brief Replace the uses of an instruction by another instruction.
3444 class UsesReplacer : public TypePromotionAction {
3445 /// Helper structure to keep track of the replaced uses.
3446 struct InstructionAndIdx {
3447 /// The instruction using the instruction.
3449 /// The index where this instruction is used for Inst.
3451 InstructionAndIdx(Instruction *Inst, unsigned Idx)
3452 : Inst(Inst), Idx(Idx) {}
3455 /// Keep track of the original uses (pair Instruction, Index).
3456 SmallVector<InstructionAndIdx, 4> OriginalUses;
3457 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
3460 /// \brief Replace all the use of \p Inst by \p New.
3461 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
3462 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
3464 // Record the original uses.
3465 for (Use &U : Inst->uses()) {
3466 Instruction *UserI = cast<Instruction>(U.getUser());
3467 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
3469 // Now, we can replace the uses.
3470 Inst->replaceAllUsesWith(New);
3473 /// \brief Reassign the original uses of Inst to Inst.
3474 void undo() override {
3475 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
3476 for (use_iterator UseIt = OriginalUses.begin(),
3477 EndIt = OriginalUses.end();
3478 UseIt != EndIt; ++UseIt) {
3479 UseIt->Inst->setOperand(UseIt->Idx, Inst);
3484 /// \brief Remove an instruction from the IR.
3485 class InstructionRemover : public TypePromotionAction {
3486 /// Original position of the instruction.
3487 InsertionHandler Inserter;
3488 /// Helper structure to hide all the link to the instruction. In other
3489 /// words, this helps to do as if the instruction was removed.
3490 OperandsHider Hider;
3491 /// Keep track of the uses replaced, if any.
3492 UsesReplacer *Replacer;
3495 /// \brief Remove all reference of \p Inst and optinally replace all its
3497 /// \pre If !Inst->use_empty(), then New != nullptr
3498 InstructionRemover(Instruction *Inst, Value *New = nullptr)
3499 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
3502 Replacer = new UsesReplacer(Inst, New);
3503 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
3504 Inst->removeFromParent();
3507 ~InstructionRemover() override { delete Replacer; }
3509 /// \brief Really remove the instruction.
3510 void commit() override { delete Inst; }
3512 /// \brief Resurrect the instruction and reassign it to the proper uses if
3513 /// new value was provided when build this action.
3514 void undo() override {
3515 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
3516 Inserter.insert(Inst);
3524 /// Restoration point.
3525 /// The restoration point is a pointer to an action instead of an iterator
3526 /// because the iterator may be invalidated but not the pointer.
3527 typedef const TypePromotionAction *ConstRestorationPt;
3528 /// Advocate every changes made in that transaction.
3530 /// Undo all the changes made after the given point.
3531 void rollback(ConstRestorationPt Point);
3532 /// Get the current restoration point.
3533 ConstRestorationPt getRestorationPoint() const;
3535 /// \name API for IR modification with state keeping to support rollback.
3537 /// Same as Instruction::setOperand.
3538 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
3539 /// Same as Instruction::eraseFromParent.
3540 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
3541 /// Same as Value::replaceAllUsesWith.
3542 void replaceAllUsesWith(Instruction *Inst, Value *New);
3543 /// Same as Value::mutateType.
3544 void mutateType(Instruction *Inst, Type *NewTy);
3545 /// Same as IRBuilder::createTrunc.
3546 Value *createTrunc(Instruction *Opnd, Type *Ty);
3547 /// Same as IRBuilder::createSExt.
3548 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
3549 /// Same as IRBuilder::createZExt.
3550 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
3551 /// Same as Instruction::moveBefore.
3552 void moveBefore(Instruction *Inst, Instruction *Before);
3556 /// The ordered list of actions made so far.
3557 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
3558 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
3561 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
3564 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
3567 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
3570 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
3573 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
3575 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
3578 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
3579 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
3582 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
3584 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
3585 Value *Val = Ptr->getBuiltValue();
3586 Actions.push_back(std::move(Ptr));
3590 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
3591 Value *Opnd, Type *Ty) {
3592 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
3593 Value *Val = Ptr->getBuiltValue();
3594 Actions.push_back(std::move(Ptr));
3598 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
3599 Value *Opnd, Type *Ty) {
3600 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
3601 Value *Val = Ptr->getBuiltValue();
3602 Actions.push_back(std::move(Ptr));
3606 void TypePromotionTransaction::moveBefore(Instruction *Inst,
3607 Instruction *Before) {
3609 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
3612 TypePromotionTransaction::ConstRestorationPt
3613 TypePromotionTransaction::getRestorationPoint() const {
3614 return !Actions.empty() ? Actions.back().get() : nullptr;
3617 void TypePromotionTransaction::commit() {
3618 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
3624 void TypePromotionTransaction::rollback(
3625 TypePromotionTransaction::ConstRestorationPt Point) {
3626 while (!Actions.empty() && Point != Actions.back().get()) {
3627 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
3632 /// \brief A helper class for matching addressing modes.
3634 /// This encapsulates the logic for matching the target-legal addressing modes.
3635 class AddressingModeMatcher {
3636 SmallVectorImpl<Instruction*> &AddrModeInsts;
3637 const TargetMachine &TM;
3638 const TargetLowering &TLI;
3639 const DataLayout &DL;
3641 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
3642 /// the memory instruction that we're computing this address for.
3645 Instruction *MemoryInst;
3647 /// This is the addressing mode that we're building up. This is
3648 /// part of the return value of this addressing mode matching stuff.
3649 ExtAddrMode &AddrMode;
3651 /// The instructions inserted by other CodeGenPrepare optimizations.
3652 const SetOfInstrs &InsertedInsts;
3653 /// A map from the instructions to their type before promotion.
3654 InstrToOrigTy &PromotedInsts;
3655 /// The ongoing transaction where every action should be registered.
3656 TypePromotionTransaction &TPT;
3658 /// This is set to true when we should not do profitability checks.
3659 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
3660 bool IgnoreProfitability;
3662 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
3663 const TargetMachine &TM, Type *AT, unsigned AS,
3664 Instruction *MI, ExtAddrMode &AM,
3665 const SetOfInstrs &InsertedInsts,
3666 InstrToOrigTy &PromotedInsts,
3667 TypePromotionTransaction &TPT)
3668 : AddrModeInsts(AMI), TM(TM),
3669 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
3670 ->getTargetLowering()),
3671 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
3672 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
3673 PromotedInsts(PromotedInsts), TPT(TPT) {
3674 IgnoreProfitability = false;
3678 /// Find the maximal addressing mode that a load/store of V can fold,
3679 /// give an access type of AccessTy. This returns a list of involved
3680 /// instructions in AddrModeInsts.
3681 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
3683 /// \p PromotedInsts maps the instructions to their type before promotion.
3684 /// \p The ongoing transaction where every action should be registered.
3685 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
3686 Instruction *MemoryInst,
3687 SmallVectorImpl<Instruction*> &AddrModeInsts,
3688 const TargetMachine &TM,
3689 const SetOfInstrs &InsertedInsts,
3690 InstrToOrigTy &PromotedInsts,
3691 TypePromotionTransaction &TPT) {
3694 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
3695 MemoryInst, Result, InsertedInsts,
3696 PromotedInsts, TPT).matchAddr(V, 0);
3697 (void)Success; assert(Success && "Couldn't select *anything*?");
3701 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
3702 bool matchAddr(Value *V, unsigned Depth);
3703 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
3704 bool *MovedAway = nullptr);
3705 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
3706 ExtAddrMode &AMBefore,
3707 ExtAddrMode &AMAfter);
3708 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
3709 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
3710 Value *PromotedOperand) const;
3713 /// Try adding ScaleReg*Scale to the current addressing mode.
3714 /// Return true and update AddrMode if this addr mode is legal for the target,
3716 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3718 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3719 // mode. Just process that directly.
3721 return matchAddr(ScaleReg, Depth);
3723 // If the scale is 0, it takes nothing to add this.
3727 // If we already have a scale of this value, we can add to it, otherwise, we
3728 // need an available scale field.
3729 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3732 ExtAddrMode TestAddrMode = AddrMode;
3734 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3735 // [A+B + A*7] -> [B+A*8].
3736 TestAddrMode.Scale += Scale;
3737 TestAddrMode.ScaledReg = ScaleReg;
3739 // If the new address isn't legal, bail out.
3740 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3743 // It was legal, so commit it.
3744 AddrMode = TestAddrMode;
3746 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3747 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3748 // X*Scale + C*Scale to addr mode.
3749 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3750 if (isa<Instruction>(ScaleReg) && // not a constant expr.
3751 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3752 TestAddrMode.ScaledReg = AddLHS;
3753 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3755 // If this addressing mode is legal, commit it and remember that we folded
3756 // this instruction.
3757 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3758 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3759 AddrMode = TestAddrMode;
3764 // Otherwise, not (x+c)*scale, just return what we have.
3768 /// This is a little filter, which returns true if an addressing computation
3769 /// involving I might be folded into a load/store accessing it.
3770 /// This doesn't need to be perfect, but needs to accept at least
3771 /// the set of instructions that MatchOperationAddr can.
3772 static bool MightBeFoldableInst(Instruction *I) {
3773 switch (I->getOpcode()) {
3774 case Instruction::BitCast:
3775 case Instruction::AddrSpaceCast:
3776 // Don't touch identity bitcasts.
3777 if (I->getType() == I->getOperand(0)->getType())
3779 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
3780 case Instruction::PtrToInt:
3781 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3783 case Instruction::IntToPtr:
3784 // We know the input is intptr_t, so this is foldable.
3786 case Instruction::Add:
3788 case Instruction::Mul:
3789 case Instruction::Shl:
3790 // Can only handle X*C and X << C.
3791 return isa<ConstantInt>(I->getOperand(1));
3792 case Instruction::GetElementPtr:
3799 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
3800 /// \note \p Val is assumed to be the product of some type promotion.
3801 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3802 /// to be legal, as the non-promoted value would have had the same state.
3803 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
3804 const DataLayout &DL, Value *Val) {
3805 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3808 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3809 // If the ISDOpcode is undefined, it was undefined before the promotion.
3812 // Otherwise, check if the promoted instruction is legal or not.
3813 return TLI.isOperationLegalOrCustom(
3814 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3817 /// \brief Hepler class to perform type promotion.
3818 class TypePromotionHelper {
3819 /// \brief Utility function to check whether or not a sign or zero extension
3820 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3821 /// either using the operands of \p Inst or promoting \p Inst.
3822 /// The type of the extension is defined by \p IsSExt.
3823 /// In other words, check if:
3824 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3825 /// #1 Promotion applies:
3826 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3827 /// #2 Operand reuses:
3828 /// ext opnd1 to ConsideredExtType.
3829 /// \p PromotedInsts maps the instructions to their type before promotion.
3830 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3831 const InstrToOrigTy &PromotedInsts, bool IsSExt);
3833 /// \brief Utility function to determine if \p OpIdx should be promoted when
3834 /// promoting \p Inst.
3835 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3836 return !(isa<SelectInst>(Inst) && OpIdx == 0);
3839 /// \brief Utility function to promote the operand of \p Ext when this
3840 /// operand is a promotable trunc or sext or zext.
3841 /// \p PromotedInsts maps the instructions to their type before promotion.
3842 /// \p CreatedInstsCost[out] contains the cost of all instructions
3843 /// created to promote the operand of Ext.
3844 /// Newly added extensions are inserted in \p Exts.
3845 /// Newly added truncates are inserted in \p Truncs.
3846 /// Should never be called directly.
3847 /// \return The promoted value which is used instead of Ext.
3848 static Value *promoteOperandForTruncAndAnyExt(
3849 Instruction *Ext, TypePromotionTransaction &TPT,
3850 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3851 SmallVectorImpl<Instruction *> *Exts,
3852 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3854 /// \brief Utility function to promote the operand of \p Ext when this
3855 /// operand is promotable and is not a supported trunc or sext.
3856 /// \p PromotedInsts maps the instructions to their type before promotion.
3857 /// \p CreatedInstsCost[out] contains the cost of all the instructions
3858 /// created to promote the operand of Ext.
3859 /// Newly added extensions are inserted in \p Exts.
3860 /// Newly added truncates are inserted in \p Truncs.
3861 /// Should never be called directly.
3862 /// \return The promoted value which is used instead of Ext.
3863 static Value *promoteOperandForOther(Instruction *Ext,
3864 TypePromotionTransaction &TPT,
3865 InstrToOrigTy &PromotedInsts,
3866 unsigned &CreatedInstsCost,
3867 SmallVectorImpl<Instruction *> *Exts,
3868 SmallVectorImpl<Instruction *> *Truncs,
3869 const TargetLowering &TLI, bool IsSExt);
3871 /// \see promoteOperandForOther.
3872 static Value *signExtendOperandForOther(
3873 Instruction *Ext, TypePromotionTransaction &TPT,
3874 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3875 SmallVectorImpl<Instruction *> *Exts,
3876 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3877 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3878 Exts, Truncs, TLI, true);
3881 /// \see promoteOperandForOther.
3882 static Value *zeroExtendOperandForOther(
3883 Instruction *Ext, TypePromotionTransaction &TPT,
3884 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3885 SmallVectorImpl<Instruction *> *Exts,
3886 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3887 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3888 Exts, Truncs, TLI, false);
3892 /// Type for the utility function that promotes the operand of Ext.
3893 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
3894 InstrToOrigTy &PromotedInsts,
3895 unsigned &CreatedInstsCost,
3896 SmallVectorImpl<Instruction *> *Exts,
3897 SmallVectorImpl<Instruction *> *Truncs,
3898 const TargetLowering &TLI);
3899 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
3900 /// action to promote the operand of \p Ext instead of using Ext.
3901 /// \return NULL if no promotable action is possible with the current
3903 /// \p InsertedInsts keeps track of all the instructions inserted by the
3904 /// other CodeGenPrepare optimizations. This information is important
3905 /// because we do not want to promote these instructions as CodeGenPrepare
3906 /// will reinsert them later. Thus creating an infinite loop: create/remove.
3907 /// \p PromotedInsts maps the instructions to their type before promotion.
3908 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3909 const TargetLowering &TLI,
3910 const InstrToOrigTy &PromotedInsts);
3913 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3914 Type *ConsideredExtType,
3915 const InstrToOrigTy &PromotedInsts,
3917 // The promotion helper does not know how to deal with vector types yet.
3918 // To be able to fix that, we would need to fix the places where we
3919 // statically extend, e.g., constants and such.
3920 if (Inst->getType()->isVectorTy())
3923 // We can always get through zext.
3924 if (isa<ZExtInst>(Inst))
3927 // sext(sext) is ok too.
3928 if (IsSExt && isa<SExtInst>(Inst))
3931 // We can get through binary operator, if it is legal. In other words, the
3932 // binary operator must have a nuw or nsw flag.
3933 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3934 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3935 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3936 (IsSExt && BinOp->hasNoSignedWrap())))
3939 // Check if we can do the following simplification.
3940 // ext(trunc(opnd)) --> ext(opnd)
3941 if (!isa<TruncInst>(Inst))
3944 Value *OpndVal = Inst->getOperand(0);
3945 // Check if we can use this operand in the extension.
3946 // If the type is larger than the result type of the extension, we cannot.
3947 if (!OpndVal->getType()->isIntegerTy() ||
3948 OpndVal->getType()->getIntegerBitWidth() >
3949 ConsideredExtType->getIntegerBitWidth())
3952 // If the operand of the truncate is not an instruction, we will not have
3953 // any information on the dropped bits.
3954 // (Actually we could for constant but it is not worth the extra logic).
3955 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3959 // Check if the source of the type is narrow enough.
3960 // I.e., check that trunc just drops extended bits of the same kind of
3962 // #1 get the type of the operand and check the kind of the extended bits.
3963 const Type *OpndType;
3964 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3965 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
3966 OpndType = It->second.getPointer();
3967 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3968 OpndType = Opnd->getOperand(0)->getType();
3972 // #2 check that the truncate just drops extended bits.
3973 return Inst->getType()->getIntegerBitWidth() >=
3974 OpndType->getIntegerBitWidth();
3977 TypePromotionHelper::Action TypePromotionHelper::getAction(
3978 Instruction *Ext, const SetOfInstrs &InsertedInsts,
3979 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3980 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
3981 "Unexpected instruction type");
3982 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3983 Type *ExtTy = Ext->getType();
3984 bool IsSExt = isa<SExtInst>(Ext);
3985 // If the operand of the extension is not an instruction, we cannot
3987 // If it, check we can get through.
3988 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3991 // Do not promote if the operand has been added by codegenprepare.
3992 // Otherwise, it means we are undoing an optimization that is likely to be
3993 // redone, thus causing potential infinite loop.
3994 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3997 // SExt or Trunc instructions.
3998 // Return the related handler.
3999 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
4000 isa<ZExtInst>(ExtOpnd))
4001 return promoteOperandForTruncAndAnyExt;
4003 // Regular instruction.
4004 // Abort early if we will have to insert non-free instructions.
4005 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
4007 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
4010 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
4011 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
4012 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4013 SmallVectorImpl<Instruction *> *Exts,
4014 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4015 // By construction, the operand of SExt is an instruction. Otherwise we cannot
4016 // get through it and this method should not be called.
4017 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
4018 Value *ExtVal = SExt;
4019 bool HasMergedNonFreeExt = false;
4020 if (isa<ZExtInst>(SExtOpnd)) {
4021 // Replace s|zext(zext(opnd))
4023 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
4025 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
4026 TPT.replaceAllUsesWith(SExt, ZExt);
4027 TPT.eraseInstruction(SExt);
4030 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
4032 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
4034 CreatedInstsCost = 0;
4036 // Remove dead code.
4037 if (SExtOpnd->use_empty())
4038 TPT.eraseInstruction(SExtOpnd);
4040 // Check if the extension is still needed.
4041 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
4042 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
4045 Exts->push_back(ExtInst);
4046 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
4051 // At this point we have: ext ty opnd to ty.
4052 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
4053 Value *NextVal = ExtInst->getOperand(0);
4054 TPT.eraseInstruction(ExtInst, NextVal);
4058 Value *TypePromotionHelper::promoteOperandForOther(
4059 Instruction *Ext, TypePromotionTransaction &TPT,
4060 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4061 SmallVectorImpl<Instruction *> *Exts,
4062 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
4064 // By construction, the operand of Ext is an instruction. Otherwise we cannot
4065 // get through it and this method should not be called.
4066 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
4067 CreatedInstsCost = 0;
4068 if (!ExtOpnd->hasOneUse()) {
4069 // ExtOpnd will be promoted.
4070 // All its uses, but Ext, will need to use a truncated value of the
4071 // promoted version.
4072 // Create the truncate now.
4073 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
4074 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
4075 ITrunc->removeFromParent();
4076 // Insert it just after the definition.
4077 ITrunc->insertAfter(ExtOpnd);
4079 Truncs->push_back(ITrunc);
4082 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
4083 // Restore the operand of Ext (which has been replaced by the previous call
4084 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
4085 TPT.setOperand(Ext, 0, ExtOpnd);
4088 // Get through the Instruction:
4089 // 1. Update its type.
4090 // 2. Replace the uses of Ext by Inst.
4091 // 3. Extend each operand that needs to be extended.
4093 // Remember the original type of the instruction before promotion.
4094 // This is useful to know that the high bits are sign extended bits.
4095 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
4096 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
4098 TPT.mutateType(ExtOpnd, Ext->getType());
4100 TPT.replaceAllUsesWith(Ext, ExtOpnd);
4102 Instruction *ExtForOpnd = Ext;
4104 DEBUG(dbgs() << "Propagate Ext to operands\n");
4105 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
4107 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
4108 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
4109 !shouldExtOperand(ExtOpnd, OpIdx)) {
4110 DEBUG(dbgs() << "No need to propagate\n");
4113 // Check if we can statically extend the operand.
4114 Value *Opnd = ExtOpnd->getOperand(OpIdx);
4115 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
4116 DEBUG(dbgs() << "Statically extend\n");
4117 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
4118 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
4119 : Cst->getValue().zext(BitWidth);
4120 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
4123 // UndefValue are typed, so we have to statically sign extend them.
4124 if (isa<UndefValue>(Opnd)) {
4125 DEBUG(dbgs() << "Statically extend\n");
4126 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
4130 // Otherwise we have to explicity sign extend the operand.
4131 // Check if Ext was reused to extend an operand.
4133 // If yes, create a new one.
4134 DEBUG(dbgs() << "More operands to ext\n");
4135 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
4136 : TPT.createZExt(Ext, Opnd, Ext->getType());
4137 if (!isa<Instruction>(ValForExtOpnd)) {
4138 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
4141 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
4144 Exts->push_back(ExtForOpnd);
4145 TPT.setOperand(ExtForOpnd, 0, Opnd);
4147 // Move the sign extension before the insertion point.
4148 TPT.moveBefore(ExtForOpnd, ExtOpnd);
4149 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
4150 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
4151 // If more sext are required, new instructions will have to be created.
4152 ExtForOpnd = nullptr;
4154 if (ExtForOpnd == Ext) {
4155 DEBUG(dbgs() << "Extension is useless now\n");
4156 TPT.eraseInstruction(Ext);
4161 /// Check whether or not promoting an instruction to a wider type is profitable.
4162 /// \p NewCost gives the cost of extension instructions created by the
4164 /// \p OldCost gives the cost of extension instructions before the promotion
4165 /// plus the number of instructions that have been
4166 /// matched in the addressing mode the promotion.
4167 /// \p PromotedOperand is the value that has been promoted.
4168 /// \return True if the promotion is profitable, false otherwise.
4169 bool AddressingModeMatcher::isPromotionProfitable(
4170 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
4171 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
4172 // The cost of the new extensions is greater than the cost of the
4173 // old extension plus what we folded.
4174 // This is not profitable.
4175 if (NewCost > OldCost)
4177 if (NewCost < OldCost)
4179 // The promotion is neutral but it may help folding the sign extension in
4180 // loads for instance.
4181 // Check that we did not create an illegal instruction.
4182 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
4185 /// Given an instruction or constant expr, see if we can fold the operation
4186 /// into the addressing mode. If so, update the addressing mode and return
4187 /// true, otherwise return false without modifying AddrMode.
4188 /// If \p MovedAway is not NULL, it contains the information of whether or
4189 /// not AddrInst has to be folded into the addressing mode on success.
4190 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
4191 /// because it has been moved away.
4192 /// Thus AddrInst must not be added in the matched instructions.
4193 /// This state can happen when AddrInst is a sext, since it may be moved away.
4194 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
4195 /// not be referenced anymore.
4196 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
4199 // Avoid exponential behavior on extremely deep expression trees.
4200 if (Depth >= 5) return false;
4202 // By default, all matched instructions stay in place.
4207 case Instruction::PtrToInt:
4208 // PtrToInt is always a noop, as we know that the int type is pointer sized.
4209 return matchAddr(AddrInst->getOperand(0), Depth);
4210 case Instruction::IntToPtr: {
4211 auto AS = AddrInst->getType()->getPointerAddressSpace();
4212 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
4213 // This inttoptr is a no-op if the integer type is pointer sized.
4214 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
4215 return matchAddr(AddrInst->getOperand(0), Depth);
4218 case Instruction::BitCast:
4219 // BitCast is always a noop, and we can handle it as long as it is
4220 // int->int or pointer->pointer (we don't want int<->fp or something).
4221 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
4222 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
4223 // Don't touch identity bitcasts. These were probably put here by LSR,
4224 // and we don't want to mess around with them. Assume it knows what it
4226 AddrInst->getOperand(0)->getType() != AddrInst->getType())
4227 return matchAddr(AddrInst->getOperand(0), Depth);
4229 case Instruction::AddrSpaceCast: {
4231 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
4232 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
4233 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
4234 return matchAddr(AddrInst->getOperand(0), Depth);
4237 case Instruction::Add: {
4238 // Check to see if we can merge in the RHS then the LHS. If so, we win.
4239 ExtAddrMode BackupAddrMode = AddrMode;
4240 unsigned OldSize = AddrModeInsts.size();
4241 // Start a transaction at this point.
4242 // The LHS may match but not the RHS.
4243 // Therefore, we need a higher level restoration point to undo partially
4244 // matched operation.
4245 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4246 TPT.getRestorationPoint();
4248 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
4249 matchAddr(AddrInst->getOperand(0), Depth+1))
4252 // Restore the old addr mode info.
4253 AddrMode = BackupAddrMode;
4254 AddrModeInsts.resize(OldSize);
4255 TPT.rollback(LastKnownGood);
4257 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
4258 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
4259 matchAddr(AddrInst->getOperand(1), Depth+1))
4262 // Otherwise we definitely can't merge the ADD in.
4263 AddrMode = BackupAddrMode;
4264 AddrModeInsts.resize(OldSize);
4265 TPT.rollback(LastKnownGood);
4268 //case Instruction::Or:
4269 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
4271 case Instruction::Mul:
4272 case Instruction::Shl: {
4273 // Can only handle X*C and X << C.
4274 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
4277 int64_t Scale = RHS->getSExtValue();
4278 if (Opcode == Instruction::Shl)
4279 Scale = 1LL << Scale;
4281 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
4283 case Instruction::GetElementPtr: {
4284 // Scan the GEP. We check it if it contains constant offsets and at most
4285 // one variable offset.
4286 int VariableOperand = -1;
4287 unsigned VariableScale = 0;
4289 int64_t ConstantOffset = 0;
4290 gep_type_iterator GTI = gep_type_begin(AddrInst);
4291 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
4292 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
4293 const StructLayout *SL = DL.getStructLayout(STy);
4295 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
4296 ConstantOffset += SL->getElementOffset(Idx);
4298 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
4299 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
4300 ConstantOffset += CI->getSExtValue()*TypeSize;
4301 } else if (TypeSize) { // Scales of zero don't do anything.
4302 // We only allow one variable index at the moment.
4303 if (VariableOperand != -1)
4306 // Remember the variable index.
4307 VariableOperand = i;
4308 VariableScale = TypeSize;
4313 // A common case is for the GEP to only do a constant offset. In this case,
4314 // just add it to the disp field and check validity.
4315 if (VariableOperand == -1) {
4316 AddrMode.BaseOffs += ConstantOffset;
4317 if (ConstantOffset == 0 ||
4318 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
4319 // Check to see if we can fold the base pointer in too.
4320 if (matchAddr(AddrInst->getOperand(0), Depth+1))
4323 AddrMode.BaseOffs -= ConstantOffset;
4327 // Save the valid addressing mode in case we can't match.
4328 ExtAddrMode BackupAddrMode = AddrMode;
4329 unsigned OldSize = AddrModeInsts.size();
4331 // See if the scale and offset amount is valid for this target.
4332 AddrMode.BaseOffs += ConstantOffset;
4334 // Match the base operand of the GEP.
4335 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
4336 // If it couldn't be matched, just stuff the value in a register.
4337 if (AddrMode.HasBaseReg) {
4338 AddrMode = BackupAddrMode;
4339 AddrModeInsts.resize(OldSize);
4342 AddrMode.HasBaseReg = true;
4343 AddrMode.BaseReg = AddrInst->getOperand(0);
4346 // Match the remaining variable portion of the GEP.
4347 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
4349 // If it couldn't be matched, try stuffing the base into a register
4350 // instead of matching it, and retrying the match of the scale.
4351 AddrMode = BackupAddrMode;
4352 AddrModeInsts.resize(OldSize);
4353 if (AddrMode.HasBaseReg)
4355 AddrMode.HasBaseReg = true;
4356 AddrMode.BaseReg = AddrInst->getOperand(0);
4357 AddrMode.BaseOffs += ConstantOffset;
4358 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
4359 VariableScale, Depth)) {
4360 // If even that didn't work, bail.
4361 AddrMode = BackupAddrMode;
4362 AddrModeInsts.resize(OldSize);
4369 case Instruction::SExt:
4370 case Instruction::ZExt: {
4371 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
4375 // Try to move this ext out of the way of the addressing mode.
4376 // Ask for a method for doing so.
4377 TypePromotionHelper::Action TPH =
4378 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
4382 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4383 TPT.getRestorationPoint();
4384 unsigned CreatedInstsCost = 0;
4385 unsigned ExtCost = !TLI.isExtFree(Ext);
4386 Value *PromotedOperand =
4387 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
4388 // SExt has been moved away.
4389 // Thus either it will be rematched later in the recursive calls or it is
4390 // gone. Anyway, we must not fold it into the addressing mode at this point.
4394 // addr = gep base, idx
4396 // promotedOpnd = ext opnd <- no match here
4397 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
4398 // addr = gep base, op <- match
4402 assert(PromotedOperand &&
4403 "TypePromotionHelper should have filtered out those cases");
4405 ExtAddrMode BackupAddrMode = AddrMode;
4406 unsigned OldSize = AddrModeInsts.size();
4408 if (!matchAddr(PromotedOperand, Depth) ||
4409 // The total of the new cost is equal to the cost of the created
4411 // The total of the old cost is equal to the cost of the extension plus
4412 // what we have saved in the addressing mode.
4413 !isPromotionProfitable(CreatedInstsCost,
4414 ExtCost + (AddrModeInsts.size() - OldSize),
4416 AddrMode = BackupAddrMode;
4417 AddrModeInsts.resize(OldSize);
4418 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
4419 TPT.rollback(LastKnownGood);
4428 /// If we can, try to add the value of 'Addr' into the current addressing mode.
4429 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
4430 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
4433 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
4434 // Start a transaction at this point that we will rollback if the matching
4436 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4437 TPT.getRestorationPoint();
4438 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
4439 // Fold in immediates if legal for the target.
4440 AddrMode.BaseOffs += CI->getSExtValue();
4441 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4443 AddrMode.BaseOffs -= CI->getSExtValue();
4444 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
4445 // If this is a global variable, try to fold it into the addressing mode.
4446 if (!AddrMode.BaseGV) {
4447 AddrMode.BaseGV = GV;
4448 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4450 AddrMode.BaseGV = nullptr;
4452 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
4453 ExtAddrMode BackupAddrMode = AddrMode;
4454 unsigned OldSize = AddrModeInsts.size();
4456 // Check to see if it is possible to fold this operation.
4457 bool MovedAway = false;
4458 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
4459 // This instruction may have been moved away. If so, there is nothing
4463 // Okay, it's possible to fold this. Check to see if it is actually
4464 // *profitable* to do so. We use a simple cost model to avoid increasing
4465 // register pressure too much.
4466 if (I->hasOneUse() ||
4467 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
4468 AddrModeInsts.push_back(I);
4472 // It isn't profitable to do this, roll back.
4473 //cerr << "NOT FOLDING: " << *I;
4474 AddrMode = BackupAddrMode;
4475 AddrModeInsts.resize(OldSize);
4476 TPT.rollback(LastKnownGood);
4478 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
4479 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
4481 TPT.rollback(LastKnownGood);
4482 } else if (isa<ConstantPointerNull>(Addr)) {
4483 // Null pointer gets folded without affecting the addressing mode.
4487 // Worse case, the target should support [reg] addressing modes. :)
4488 if (!AddrMode.HasBaseReg) {
4489 AddrMode.HasBaseReg = true;
4490 AddrMode.BaseReg = Addr;
4491 // Still check for legality in case the target supports [imm] but not [i+r].
4492 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4494 AddrMode.HasBaseReg = false;
4495 AddrMode.BaseReg = nullptr;
4498 // If the base register is already taken, see if we can do [r+r].
4499 if (AddrMode.Scale == 0) {
4501 AddrMode.ScaledReg = Addr;
4502 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4505 AddrMode.ScaledReg = nullptr;
4508 TPT.rollback(LastKnownGood);
4512 /// Check to see if all uses of OpVal by the specified inline asm call are due
4513 /// to memory operands. If so, return true, otherwise return false.
4514 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
4515 const TargetMachine &TM) {
4516 const Function *F = CI->getParent()->getParent();
4517 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
4518 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
4519 TargetLowering::AsmOperandInfoVector TargetConstraints =
4520 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
4521 ImmutableCallSite(CI));
4522 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4523 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4525 // Compute the constraint code and ConstraintType to use.
4526 TLI->ComputeConstraintToUse(OpInfo, SDValue());
4528 // If this asm operand is our Value*, and if it isn't an indirect memory
4529 // operand, we can't fold it!
4530 if (OpInfo.CallOperandVal == OpVal &&
4531 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
4532 !OpInfo.isIndirect))
4539 /// Recursively walk all the uses of I until we find a memory use.
4540 /// If we find an obviously non-foldable instruction, return true.
4541 /// Add the ultimately found memory instructions to MemoryUses.
4542 static bool FindAllMemoryUses(
4544 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
4545 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
4546 // If we already considered this instruction, we're done.
4547 if (!ConsideredInsts.insert(I).second)
4550 // If this is an obviously unfoldable instruction, bail out.
4551 if (!MightBeFoldableInst(I))
4554 // Loop over all the uses, recursively processing them.
4555 for (Use &U : I->uses()) {
4556 Instruction *UserI = cast<Instruction>(U.getUser());
4558 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4559 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
4563 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4564 unsigned opNo = U.getOperandNo();
4565 if (opNo == 0) return true; // Storing addr, not into addr.
4566 MemoryUses.push_back(std::make_pair(SI, opNo));
4570 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4571 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
4572 if (!IA) return true;
4574 // If this is a memory operand, we're cool, otherwise bail out.
4575 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
4580 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
4587 /// Return true if Val is already known to be live at the use site that we're
4588 /// folding it into. If so, there is no cost to include it in the addressing
4589 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4590 /// instruction already.
4591 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4592 Value *KnownLive2) {
4593 // If Val is either of the known-live values, we know it is live!
4594 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4597 // All values other than instructions and arguments (e.g. constants) are live.
4598 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4600 // If Val is a constant sized alloca in the entry block, it is live, this is
4601 // true because it is just a reference to the stack/frame pointer, which is
4602 // live for the whole function.
4603 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4604 if (AI->isStaticAlloca())
4607 // Check to see if this value is already used in the memory instruction's
4608 // block. If so, it's already live into the block at the very least, so we
4609 // can reasonably fold it.
4610 return Val->isUsedInBasicBlock(MemoryInst->getParent());
4613 /// It is possible for the addressing mode of the machine to fold the specified
4614 /// instruction into a load or store that ultimately uses it.
4615 /// However, the specified instruction has multiple uses.
4616 /// Given this, it may actually increase register pressure to fold it
4617 /// into the load. For example, consider this code:
4621 /// use(Y) -> nonload/store
4625 /// In this case, Y has multiple uses, and can be folded into the load of Z
4626 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4627 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4628 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4629 /// number of computations either.
4631 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4632 /// X was live across 'load Z' for other reasons, we actually *would* want to
4633 /// fold the addressing mode in the Z case. This would make Y die earlier.
4634 bool AddressingModeMatcher::
4635 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4636 ExtAddrMode &AMAfter) {
4637 if (IgnoreProfitability) return true;
4639 // AMBefore is the addressing mode before this instruction was folded into it,
4640 // and AMAfter is the addressing mode after the instruction was folded. Get
4641 // the set of registers referenced by AMAfter and subtract out those
4642 // referenced by AMBefore: this is the set of values which folding in this
4643 // address extends the lifetime of.
4645 // Note that there are only two potential values being referenced here,
4646 // BaseReg and ScaleReg (global addresses are always available, as are any
4647 // folded immediates).
4648 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4650 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4651 // lifetime wasn't extended by adding this instruction.
4652 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4654 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4655 ScaledReg = nullptr;
4657 // If folding this instruction (and it's subexprs) didn't extend any live
4658 // ranges, we're ok with it.
4659 if (!BaseReg && !ScaledReg)
4662 // If all uses of this instruction are ultimately load/store/inlineasm's,
4663 // check to see if their addressing modes will include this instruction. If
4664 // so, we can fold it into all uses, so it doesn't matter if it has multiple
4666 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
4667 SmallPtrSet<Instruction*, 16> ConsideredInsts;
4668 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
4669 return false; // Has a non-memory, non-foldable use!
4671 // Now that we know that all uses of this instruction are part of a chain of
4672 // computation involving only operations that could theoretically be folded
4673 // into a memory use, loop over each of these uses and see if they could
4674 // *actually* fold the instruction.
4675 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4676 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4677 Instruction *User = MemoryUses[i].first;
4678 unsigned OpNo = MemoryUses[i].second;
4680 // Get the access type of this use. If the use isn't a pointer, we don't
4681 // know what it accesses.
4682 Value *Address = User->getOperand(OpNo);
4683 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4686 Type *AddressAccessTy = AddrTy->getElementType();
4687 unsigned AS = AddrTy->getAddressSpace();
4689 // Do a match against the root of this address, ignoring profitability. This
4690 // will tell us if the addressing mode for the memory operation will
4691 // *actually* cover the shared instruction.
4693 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4694 TPT.getRestorationPoint();
4695 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
4696 MemoryInst, Result, InsertedInsts,
4697 PromotedInsts, TPT);
4698 Matcher.IgnoreProfitability = true;
4699 bool Success = Matcher.matchAddr(Address, 0);
4700 (void)Success; assert(Success && "Couldn't select *anything*?");
4702 // The match was to check the profitability, the changes made are not
4703 // part of the original matcher. Therefore, they should be dropped
4704 // otherwise the original matcher will not present the right state.
4705 TPT.rollback(LastKnownGood);
4707 // If the match didn't cover I, then it won't be shared by it.
4708 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
4709 I) == MatchedAddrModeInsts.end())
4712 MatchedAddrModeInsts.clear();
4718 } // end anonymous namespace
4720 /// Return true if the specified values are defined in a
4721 /// different basic block than BB.
4722 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4723 if (Instruction *I = dyn_cast<Instruction>(V))
4724 return I->getParent() != BB;
4728 /// Load and Store Instructions often have addressing modes that can do
4729 /// significant amounts of computation. As such, instruction selection will try
4730 /// to get the load or store to do as much computation as possible for the
4731 /// program. The problem is that isel can only see within a single block. As
4732 /// such, we sink as much legal addressing mode work into the block as possible.
4734 /// This method is used to optimize both load/store and inline asms with memory
4736 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4737 Type *AccessTy, unsigned AddrSpace) {
4740 // Try to collapse single-value PHI nodes. This is necessary to undo
4741 // unprofitable PRE transformations.
4742 SmallVector<Value*, 8> worklist;
4743 SmallPtrSet<Value*, 16> Visited;
4744 worklist.push_back(Addr);
4746 // Use a worklist to iteratively look through PHI nodes, and ensure that
4747 // the addressing mode obtained from the non-PHI roots of the graph
4749 Value *Consensus = nullptr;
4750 unsigned NumUsesConsensus = 0;
4751 bool IsNumUsesConsensusValid = false;
4752 SmallVector<Instruction*, 16> AddrModeInsts;
4753 ExtAddrMode AddrMode;
4754 TypePromotionTransaction TPT;
4755 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4756 TPT.getRestorationPoint();
4757 while (!worklist.empty()) {
4758 Value *V = worklist.back();
4759 worklist.pop_back();
4761 // Break use-def graph loops.
4762 if (!Visited.insert(V).second) {
4763 Consensus = nullptr;
4767 // For a PHI node, push all of its incoming values.
4768 if (PHINode *P = dyn_cast<PHINode>(V)) {
4769 for (Value *IncValue : P->incoming_values())
4770 worklist.push_back(IncValue);
4774 // For non-PHIs, determine the addressing mode being computed.
4775 SmallVector<Instruction*, 16> NewAddrModeInsts;
4776 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4777 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
4778 InsertedInsts, PromotedInsts, TPT);
4780 // This check is broken into two cases with very similar code to avoid using
4781 // getNumUses() as much as possible. Some values have a lot of uses, so
4782 // calling getNumUses() unconditionally caused a significant compile-time
4786 AddrMode = NewAddrMode;
4787 AddrModeInsts = NewAddrModeInsts;
4789 } else if (NewAddrMode == AddrMode) {
4790 if (!IsNumUsesConsensusValid) {
4791 NumUsesConsensus = Consensus->getNumUses();
4792 IsNumUsesConsensusValid = true;
4795 // Ensure that the obtained addressing mode is equivalent to that obtained
4796 // for all other roots of the PHI traversal. Also, when choosing one
4797 // such root as representative, select the one with the most uses in order
4798 // to keep the cost modeling heuristics in AddressingModeMatcher
4800 unsigned NumUses = V->getNumUses();
4801 if (NumUses > NumUsesConsensus) {
4803 NumUsesConsensus = NumUses;
4804 AddrModeInsts = NewAddrModeInsts;
4809 Consensus = nullptr;
4813 // If the addressing mode couldn't be determined, or if multiple different
4814 // ones were determined, bail out now.
4816 TPT.rollback(LastKnownGood);
4821 // Check to see if any of the instructions supersumed by this addr mode are
4822 // non-local to I's BB.
4823 bool AnyNonLocal = false;
4824 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
4825 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
4831 // If all the instructions matched are already in this BB, don't do anything.
4833 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
4837 // Insert this computation right after this user. Since our caller is
4838 // scanning from the top of the BB to the bottom, reuse of the expr are
4839 // guaranteed to happen later.
4840 IRBuilder<> Builder(MemoryInst);
4842 // Now that we determined the addressing expression we want to use and know
4843 // that we have to sink it into this block. Check to see if we have already
4844 // done this for some other load/store instr in this block. If so, reuse the
4846 Value *&SunkAddr = SunkAddrs[Addr];
4848 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
4849 << *MemoryInst << "\n");
4850 if (SunkAddr->getType() != Addr->getType())
4851 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
4852 } else if (AddrSinkUsingGEPs ||
4853 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
4854 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
4856 // By default, we use the GEP-based method when AA is used later. This
4857 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4858 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4859 << *MemoryInst << "\n");
4860 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4861 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4863 // First, find the pointer.
4864 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4865 ResultPtr = AddrMode.BaseReg;
4866 AddrMode.BaseReg = nullptr;
4869 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4870 // We can't add more than one pointer together, nor can we scale a
4871 // pointer (both of which seem meaningless).
4872 if (ResultPtr || AddrMode.Scale != 1)
4875 ResultPtr = AddrMode.ScaledReg;
4879 if (AddrMode.BaseGV) {
4883 ResultPtr = AddrMode.BaseGV;
4886 // If the real base value actually came from an inttoptr, then the matcher
4887 // will look through it and provide only the integer value. In that case,
4889 if (!ResultPtr && AddrMode.BaseReg) {
4891 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
4892 AddrMode.BaseReg = nullptr;
4893 } else if (!ResultPtr && AddrMode.Scale == 1) {
4895 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
4900 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4901 SunkAddr = Constant::getNullValue(Addr->getType());
4902 } else if (!ResultPtr) {
4906 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4907 Type *I8Ty = Builder.getInt8Ty();
4909 // Start with the base register. Do this first so that subsequent address
4910 // matching finds it last, which will prevent it from trying to match it
4911 // as the scaled value in case it happens to be a mul. That would be
4912 // problematic if we've sunk a different mul for the scale, because then
4913 // we'd end up sinking both muls.
4914 if (AddrMode.BaseReg) {
4915 Value *V = AddrMode.BaseReg;
4916 if (V->getType() != IntPtrTy)
4917 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4922 // Add the scale value.
4923 if (AddrMode.Scale) {
4924 Value *V = AddrMode.ScaledReg;
4925 if (V->getType() == IntPtrTy) {
4927 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4928 cast<IntegerType>(V->getType())->getBitWidth()) {
4929 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4931 // It is only safe to sign extend the BaseReg if we know that the math
4932 // required to create it did not overflow before we extend it. Since
4933 // the original IR value was tossed in favor of a constant back when
4934 // the AddrMode was created we need to bail out gracefully if widths
4935 // do not match instead of extending it.
4936 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
4937 if (I && (ResultIndex != AddrMode.BaseReg))
4938 I->eraseFromParent();
4942 if (AddrMode.Scale != 1)
4943 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4946 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4951 // Add in the Base Offset if present.
4952 if (AddrMode.BaseOffs) {
4953 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4955 // We need to add this separately from the scale above to help with
4956 // SDAG consecutive load/store merging.
4957 if (ResultPtr->getType() != I8PtrTy)
4958 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
4959 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4966 SunkAddr = ResultPtr;
4968 if (ResultPtr->getType() != I8PtrTy)
4969 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
4970 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4973 if (SunkAddr->getType() != Addr->getType())
4974 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
4977 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4978 << *MemoryInst << "\n");
4979 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4980 Value *Result = nullptr;
4982 // Start with the base register. Do this first so that subsequent address
4983 // matching finds it last, which will prevent it from trying to match it
4984 // as the scaled value in case it happens to be a mul. That would be
4985 // problematic if we've sunk a different mul for the scale, because then
4986 // we'd end up sinking both muls.
4987 if (AddrMode.BaseReg) {
4988 Value *V = AddrMode.BaseReg;
4989 if (V->getType()->isPointerTy())
4990 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4991 if (V->getType() != IntPtrTy)
4992 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4996 // Add the scale value.
4997 if (AddrMode.Scale) {
4998 Value *V = AddrMode.ScaledReg;
4999 if (V->getType() == IntPtrTy) {
5001 } else if (V->getType()->isPointerTy()) {
5002 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
5003 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
5004 cast<IntegerType>(V->getType())->getBitWidth()) {
5005 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
5007 // It is only safe to sign extend the BaseReg if we know that the math
5008 // required to create it did not overflow before we extend it. Since
5009 // the original IR value was tossed in favor of a constant back when
5010 // the AddrMode was created we need to bail out gracefully if widths
5011 // do not match instead of extending it.
5012 Instruction *I = dyn_cast_or_null<Instruction>(Result);
5013 if (I && (Result != AddrMode.BaseReg))
5014 I->eraseFromParent();
5017 if (AddrMode.Scale != 1)
5018 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
5021 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5026 // Add in the BaseGV if present.
5027 if (AddrMode.BaseGV) {
5028 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
5030 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5035 // Add in the Base Offset if present.
5036 if (AddrMode.BaseOffs) {
5037 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
5039 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5045 SunkAddr = Constant::getNullValue(Addr->getType());
5047 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
5050 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
5052 // If we have no uses, recursively delete the value and all dead instructions
5054 if (Repl->use_empty()) {
5055 // This can cause recursive deletion, which can invalidate our iterator.
5056 // Use a WeakVH to hold onto it in case this happens.
5057 WeakVH IterHandle(&*CurInstIterator);
5058 BasicBlock *BB = CurInstIterator->getParent();
5060 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
5062 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
5063 // If the iterator instruction was recursively deleted, start over at the
5064 // start of the block.
5065 CurInstIterator = BB->begin();
5073 /// If there are any memory operands, use OptimizeMemoryInst to sink their
5074 /// address computing into the block when possible / profitable.
5075 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
5076 bool MadeChange = false;
5078 const TargetRegisterInfo *TRI =
5079 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
5080 TargetLowering::AsmOperandInfoVector TargetConstraints =
5081 TLI->ParseConstraints(*DL, TRI, CS);
5083 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5084 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
5086 // Compute the constraint code and ConstraintType to use.
5087 TLI->ComputeConstraintToUse(OpInfo, SDValue());
5089 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5090 OpInfo.isIndirect) {
5091 Value *OpVal = CS->getArgOperand(ArgNo++);
5092 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
5093 } else if (OpInfo.Type == InlineAsm::isInput)
5100 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
5101 /// sign extensions.
5102 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
5103 assert(!Inst->use_empty() && "Input must have at least one use");
5104 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
5105 bool IsSExt = isa<SExtInst>(FirstUser);
5106 Type *ExtTy = FirstUser->getType();
5107 for (const User *U : Inst->users()) {
5108 const Instruction *UI = cast<Instruction>(U);
5109 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
5111 Type *CurTy = UI->getType();
5112 // Same input and output types: Same instruction after CSE.
5116 // If IsSExt is true, we are in this situation:
5118 // b = sext ty1 a to ty2
5119 // c = sext ty1 a to ty3
5120 // Assuming ty2 is shorter than ty3, this could be turned into:
5122 // b = sext ty1 a to ty2
5123 // c = sext ty2 b to ty3
5124 // However, the last sext is not free.
5128 // This is a ZExt, maybe this is free to extend from one type to another.
5129 // In that case, we would not account for a different use.
5132 if (ExtTy->getScalarType()->getIntegerBitWidth() >
5133 CurTy->getScalarType()->getIntegerBitWidth()) {
5141 if (!TLI.isZExtFree(NarrowTy, LargeTy))
5144 // All uses are the same or can be derived from one another for free.
5148 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
5149 /// load instruction.
5150 /// If an ext(load) can be formed, it is returned via \p LI for the load
5151 /// and \p Inst for the extension.
5152 /// Otherwise LI == nullptr and Inst == nullptr.
5153 /// When some promotion happened, \p TPT contains the proper state to
5156 /// \return true when promoting was necessary to expose the ext(load)
5157 /// opportunity, false otherwise.
5161 /// %ld = load i32* %addr
5162 /// %add = add nuw i32 %ld, 4
5163 /// %zext = zext i32 %add to i64
5167 /// %ld = load i32* %addr
5168 /// %zext = zext i32 %ld to i64
5169 /// %add = add nuw i64 %zext, 4
5171 /// Thanks to the promotion, we can match zext(load i32*) to i64.
5172 bool CodeGenPrepare::extLdPromotion(TypePromotionTransaction &TPT,
5173 LoadInst *&LI, Instruction *&Inst,
5174 const SmallVectorImpl<Instruction *> &Exts,
5175 unsigned CreatedInstsCost = 0) {
5176 // Iterate over all the extensions to see if one form an ext(load).
5177 for (auto I : Exts) {
5178 // Check if we directly have ext(load).
5179 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
5181 // No promotion happened here.
5184 // Check whether or not we want to do any promotion.
5185 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
5187 // Get the action to perform the promotion.
5188 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
5189 I, InsertedInsts, *TLI, PromotedInsts);
5190 // Check if we can promote.
5193 // Save the current state.
5194 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5195 TPT.getRestorationPoint();
5196 SmallVector<Instruction *, 4> NewExts;
5197 unsigned NewCreatedInstsCost = 0;
5198 unsigned ExtCost = !TLI->isExtFree(I);
5200 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
5201 &NewExts, nullptr, *TLI);
5202 assert(PromotedVal &&
5203 "TypePromotionHelper should have filtered out those cases");
5205 // We would be able to merge only one extension in a load.
5206 // Therefore, if we have more than 1 new extension we heuristically
5207 // cut this search path, because it means we degrade the code quality.
5208 // With exactly 2, the transformation is neutral, because we will merge
5209 // one extension but leave one. However, we optimistically keep going,
5210 // because the new extension may be removed too.
5211 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
5212 TotalCreatedInstsCost -= ExtCost;
5213 if (!StressExtLdPromotion &&
5214 (TotalCreatedInstsCost > 1 ||
5215 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
5216 // The promotion is not profitable, rollback to the previous state.
5217 TPT.rollback(LastKnownGood);
5220 // The promotion is profitable.
5221 // Check if it exposes an ext(load).
5222 (void)extLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
5223 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
5224 // If we have created a new extension, i.e., now we have two
5225 // extensions. We must make sure one of them is merged with
5226 // the load, otherwise we may degrade the code quality.
5227 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
5228 // Promotion happened.
5230 // If this does not help to expose an ext(load) then, rollback.
5231 TPT.rollback(LastKnownGood);
5233 // None of the extension can form an ext(load).
5239 /// Move a zext or sext fed by a load into the same basic block as the load,
5240 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
5241 /// extend into the load.
5242 /// \p I[in/out] the extension may be modified during the process if some
5243 /// promotions apply.
5245 bool CodeGenPrepare::moveExtToFormExtLoad(Instruction *&I) {
5246 // Try to promote a chain of computation if it allows to form
5247 // an extended load.
5248 TypePromotionTransaction TPT;
5249 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5250 TPT.getRestorationPoint();
5251 SmallVector<Instruction *, 1> Exts;
5253 // Look for a load being extended.
5254 LoadInst *LI = nullptr;
5255 Instruction *OldExt = I;
5256 bool HasPromoted = extLdPromotion(TPT, LI, I, Exts);
5258 assert(!HasPromoted && !LI && "If we did not match any load instruction "
5259 "the code must remain the same");
5264 // If they're already in the same block, there's nothing to do.
5265 // Make the cheap checks first if we did not promote.
5266 // If we promoted, we need to check if it is indeed profitable.
5267 if (!HasPromoted && LI->getParent() == I->getParent())
5270 EVT VT = TLI->getValueType(*DL, I->getType());
5271 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
5273 // If the load has other users and the truncate is not free, this probably
5274 // isn't worthwhile.
5275 if (!LI->hasOneUse() && TLI &&
5276 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
5277 !TLI->isTruncateFree(I->getType(), LI->getType())) {
5279 TPT.rollback(LastKnownGood);
5283 // Check whether the target supports casts folded into loads.
5285 if (isa<ZExtInst>(I))
5286 LType = ISD::ZEXTLOAD;
5288 assert(isa<SExtInst>(I) && "Unexpected ext type!");
5289 LType = ISD::SEXTLOAD;
5291 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
5293 TPT.rollback(LastKnownGood);
5297 // Move the extend into the same block as the load, so that SelectionDAG
5300 I->removeFromParent();
5306 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
5307 BasicBlock *DefBB = I->getParent();
5309 // If the result of a {s|z}ext and its source are both live out, rewrite all
5310 // other uses of the source with result of extension.
5311 Value *Src = I->getOperand(0);
5312 if (Src->hasOneUse())
5315 // Only do this xform if truncating is free.
5316 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
5319 // Only safe to perform the optimization if the source is also defined in
5321 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
5324 bool DefIsLiveOut = false;
5325 for (User *U : I->users()) {
5326 Instruction *UI = cast<Instruction>(U);
5328 // Figure out which BB this ext is used in.
5329 BasicBlock *UserBB = UI->getParent();
5330 if (UserBB == DefBB) continue;
5331 DefIsLiveOut = true;
5337 // Make sure none of the uses are PHI nodes.
5338 for (User *U : Src->users()) {
5339 Instruction *UI = cast<Instruction>(U);
5340 BasicBlock *UserBB = UI->getParent();
5341 if (UserBB == DefBB) continue;
5342 // Be conservative. We don't want this xform to end up introducing
5343 // reloads just before load / store instructions.
5344 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
5348 // InsertedTruncs - Only insert one trunc in each block once.
5349 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
5351 bool MadeChange = false;
5352 for (Use &U : Src->uses()) {
5353 Instruction *User = cast<Instruction>(U.getUser());
5355 // Figure out which BB this ext is used in.
5356 BasicBlock *UserBB = User->getParent();
5357 if (UserBB == DefBB) continue;
5359 // Both src and def are live in this block. Rewrite the use.
5360 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
5362 if (!InsertedTrunc) {
5363 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5364 assert(InsertPt != UserBB->end());
5365 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5366 InsertedInsts.insert(InsertedTrunc);
5369 // Replace a use of the {s|z}ext source with a use of the result.
5378 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
5379 // just after the load if the target can fold this into one extload instruction,
5380 // with the hope of eliminating some of the other later "and" instructions using
5381 // the loaded value. "and"s that are made trivially redundant by the insertion
5382 // of the new "and" are removed by this function, while others (e.g. those whose
5383 // path from the load goes through a phi) are left for isel to potentially
5416 // becomes (after a call to optimizeLoadExt for each load):
5420 // x1' = and x1, 0xff
5424 // x2' = and x2, 0xff
5431 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5433 if (!Load->isSimple() ||
5434 !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
5437 // Skip loads we've already transformed or have no reason to transform.
5438 if (Load->hasOneUse()) {
5439 User *LoadUser = *Load->user_begin();
5440 if (cast<Instruction>(LoadUser)->getParent() == Load->getParent() &&
5441 !dyn_cast<PHINode>(LoadUser))
5445 // Look at all uses of Load, looking through phis, to determine how many bits
5446 // of the loaded value are needed.
5447 SmallVector<Instruction *, 8> WorkList;
5448 SmallPtrSet<Instruction *, 16> Visited;
5449 SmallVector<Instruction *, 8> AndsToMaybeRemove;
5450 for (auto *U : Load->users())
5451 WorkList.push_back(cast<Instruction>(U));
5453 EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5454 unsigned BitWidth = LoadResultVT.getSizeInBits();
5455 APInt DemandBits(BitWidth, 0);
5456 APInt WidestAndBits(BitWidth, 0);
5458 while (!WorkList.empty()) {
5459 Instruction *I = WorkList.back();
5460 WorkList.pop_back();
5462 // Break use-def graph loops.
5463 if (!Visited.insert(I).second)
5466 // For a PHI node, push all of its users.
5467 if (auto *Phi = dyn_cast<PHINode>(I)) {
5468 for (auto *U : Phi->users())
5469 WorkList.push_back(cast<Instruction>(U));
5473 switch (I->getOpcode()) {
5474 case llvm::Instruction::And: {
5475 auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5478 APInt AndBits = AndC->getValue();
5479 DemandBits |= AndBits;
5480 // Keep track of the widest and mask we see.
5481 if (AndBits.ugt(WidestAndBits))
5482 WidestAndBits = AndBits;
5483 if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5484 AndsToMaybeRemove.push_back(I);
5488 case llvm::Instruction::Shl: {
5489 auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5492 uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5493 auto ShlDemandBits = APInt::getAllOnesValue(BitWidth).lshr(ShiftAmt);
5494 DemandBits |= ShlDemandBits;
5498 case llvm::Instruction::Trunc: {
5499 EVT TruncVT = TLI->getValueType(*DL, I->getType());
5500 unsigned TruncBitWidth = TruncVT.getSizeInBits();
5501 auto TruncBits = APInt::getAllOnesValue(TruncBitWidth).zext(BitWidth);
5502 DemandBits |= TruncBits;
5511 uint32_t ActiveBits = DemandBits.getActiveBits();
5512 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5513 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5514 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5515 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5516 // followed by an AND.
5517 // TODO: Look into removing this restriction by fixing backends to either
5518 // return false for isLoadExtLegal for i1 or have them select this pattern to
5519 // a single instruction.
5521 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5522 // mask, since these are the only ands that will be removed by isel.
5523 if (ActiveBits <= 1 || !APIntOps::isMask(ActiveBits, DemandBits) ||
5524 WidestAndBits != DemandBits)
5527 LLVMContext &Ctx = Load->getType()->getContext();
5528 Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5529 EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5531 // Reject cases that won't be matched as extloads.
5532 if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5533 !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5536 IRBuilder<> Builder(Load->getNextNode());
5537 auto *NewAnd = dyn_cast<Instruction>(
5538 Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5540 // Replace all uses of load with new and (except for the use of load in the
5542 Load->replaceAllUsesWith(NewAnd);
5543 NewAnd->setOperand(0, Load);
5545 // Remove any and instructions that are now redundant.
5546 for (auto *And : AndsToMaybeRemove)
5547 // Check that the and mask is the same as the one we decided to put on the
5549 if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5550 And->replaceAllUsesWith(NewAnd);
5551 if (&*CurInstIterator == And)
5552 CurInstIterator = std::next(And->getIterator());
5553 And->eraseFromParent();
5561 /// Check if V (an operand of a select instruction) is an expensive instruction
5562 /// that is only used once.
5563 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5564 auto *I = dyn_cast<Instruction>(V);
5565 // If it's safe to speculatively execute, then it should not have side
5566 // effects; therefore, it's safe to sink and possibly *not* execute.
5567 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5568 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
5571 /// Returns true if a SelectInst should be turned into an explicit branch.
5572 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
5574 // FIXME: This should use the same heuristics as IfConversion to determine
5575 // whether a select is better represented as a branch. This requires that
5576 // branch probability metadata is preserved for the select, which is not the
5579 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5581 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5582 // comparison condition. If the compare has more than one use, there's
5583 // probably another cmov or setcc around, so it's not worth emitting a branch.
5584 if (!Cmp || !Cmp->hasOneUse())
5587 Value *CmpOp0 = Cmp->getOperand(0);
5588 Value *CmpOp1 = Cmp->getOperand(1);
5590 // Emit "cmov on compare with a memory operand" as a branch to avoid stalls
5591 // on a load from memory. But if the load is used more than once, do not
5592 // change the select to a branch because the load is probably needed
5593 // regardless of whether the branch is taken or not.
5594 if ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
5595 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()))
5598 // If either operand of the select is expensive and only needed on one side
5599 // of the select, we should form a branch.
5600 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5601 sinkSelectOperand(TTI, SI->getFalseValue()))
5608 /// If we have a SelectInst that will likely profit from branch prediction,
5609 /// turn it into a branch.
5610 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5611 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
5613 // Can we convert the 'select' to CF ?
5614 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
5617 TargetLowering::SelectSupportKind SelectKind;
5619 SelectKind = TargetLowering::VectorMaskSelect;
5620 else if (SI->getType()->isVectorTy())
5621 SelectKind = TargetLowering::ScalarCondVectorVal;
5623 SelectKind = TargetLowering::ScalarValSelect;
5625 // Do we have efficient codegen support for this kind of 'selects' ?
5626 if (TLI->isSelectSupported(SelectKind)) {
5627 // We have efficient codegen support for the select instruction.
5628 // Check if it is profitable to keep this 'select'.
5629 if (!TLI->isPredictableSelectExpensive() ||
5630 !isFormingBranchFromSelectProfitable(TTI, SI))
5636 // Transform a sequence like this:
5638 // %cmp = cmp uge i32 %a, %b
5639 // %sel = select i1 %cmp, i32 %c, i32 %d
5643 // %cmp = cmp uge i32 %a, %b
5644 // br i1 %cmp, label %select.true, label %select.false
5646 // br label %select.end
5648 // br label %select.end
5650 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
5652 // In addition, we may sink instructions that produce %c or %d from
5653 // the entry block into the destination(s) of the new branch.
5654 // If the true or false blocks do not contain a sunken instruction, that
5655 // block and its branch may be optimized away. In that case, one side of the
5656 // first branch will point directly to select.end, and the corresponding PHI
5657 // predecessor block will be the start block.
5659 // First, we split the block containing the select into 2 blocks.
5660 BasicBlock *StartBlock = SI->getParent();
5661 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
5662 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
5664 // Delete the unconditional branch that was just created by the split.
5665 StartBlock->getTerminator()->eraseFromParent();
5667 // These are the new basic blocks for the conditional branch.
5668 // At least one will become an actual new basic block.
5669 BasicBlock *TrueBlock = nullptr;
5670 BasicBlock *FalseBlock = nullptr;
5672 // Sink expensive instructions into the conditional blocks to avoid executing
5673 // them speculatively.
5674 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
5675 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
5676 EndBlock->getParent(), EndBlock);
5677 auto *TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
5678 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
5679 TrueInst->moveBefore(TrueBranch);
5681 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
5682 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
5683 EndBlock->getParent(), EndBlock);
5684 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5685 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
5686 FalseInst->moveBefore(FalseBranch);
5689 // If there was nothing to sink, then arbitrarily choose the 'false' side
5690 // for a new input value to the PHI.
5691 if (TrueBlock == FalseBlock) {
5692 assert(TrueBlock == nullptr &&
5693 "Unexpected basic block transform while optimizing select");
5695 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
5696 EndBlock->getParent(), EndBlock);
5697 BranchInst::Create(EndBlock, FalseBlock);
5700 // Insert the real conditional branch based on the original condition.
5701 // If we did not create a new block for one of the 'true' or 'false' paths
5702 // of the condition, it means that side of the branch goes to the end block
5703 // directly and the path originates from the start block from the point of
5704 // view of the new PHI.
5705 if (TrueBlock == nullptr) {
5706 BranchInst::Create(EndBlock, FalseBlock, SI->getCondition(), SI);
5707 TrueBlock = StartBlock;
5708 } else if (FalseBlock == nullptr) {
5709 BranchInst::Create(TrueBlock, EndBlock, SI->getCondition(), SI);
5710 FalseBlock = StartBlock;
5712 BranchInst::Create(TrueBlock, FalseBlock, SI->getCondition(), SI);
5715 // The select itself is replaced with a PHI Node.
5716 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
5718 PN->addIncoming(SI->getTrueValue(), TrueBlock);
5719 PN->addIncoming(SI->getFalseValue(), FalseBlock);
5721 SI->replaceAllUsesWith(PN);
5722 SI->eraseFromParent();
5724 // Instruct OptimizeBlock to skip to the next block.
5725 CurInstIterator = StartBlock->end();
5726 ++NumSelectsExpanded;
5730 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
5731 SmallVector<int, 16> Mask(SVI->getShuffleMask());
5733 for (unsigned i = 0; i < Mask.size(); ++i) {
5734 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
5736 SplatElem = Mask[i];
5742 /// Some targets have expensive vector shifts if the lanes aren't all the same
5743 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
5744 /// it's often worth sinking a shufflevector splat down to its use so that
5745 /// codegen can spot all lanes are identical.
5746 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
5747 BasicBlock *DefBB = SVI->getParent();
5749 // Only do this xform if variable vector shifts are particularly expensive.
5750 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
5753 // We only expect better codegen by sinking a shuffle if we can recognise a
5755 if (!isBroadcastShuffle(SVI))
5758 // InsertedShuffles - Only insert a shuffle in each block once.
5759 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
5761 bool MadeChange = false;
5762 for (User *U : SVI->users()) {
5763 Instruction *UI = cast<Instruction>(U);
5765 // Figure out which BB this ext is used in.
5766 BasicBlock *UserBB = UI->getParent();
5767 if (UserBB == DefBB) continue;
5769 // For now only apply this when the splat is used by a shift instruction.
5770 if (!UI->isShift()) continue;
5772 // Everything checks out, sink the shuffle if the user's block doesn't
5773 // already have a copy.
5774 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
5776 if (!InsertedShuffle) {
5777 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5778 assert(InsertPt != UserBB->end());
5780 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5781 SVI->getOperand(2), "", &*InsertPt);
5784 UI->replaceUsesOfWith(SVI, InsertedShuffle);
5788 // If we removed all uses, nuke the shuffle.
5789 if (SVI->use_empty()) {
5790 SVI->eraseFromParent();
5797 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
5801 Value *Cond = SI->getCondition();
5802 Type *OldType = Cond->getType();
5803 LLVMContext &Context = Cond->getContext();
5804 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
5805 unsigned RegWidth = RegType.getSizeInBits();
5807 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
5810 // If the register width is greater than the type width, expand the condition
5811 // of the switch instruction and each case constant to the width of the
5812 // register. By widening the type of the switch condition, subsequent
5813 // comparisons (for case comparisons) will not need to be extended to the
5814 // preferred register width, so we will potentially eliminate N-1 extends,
5815 // where N is the number of cases in the switch.
5816 auto *NewType = Type::getIntNTy(Context, RegWidth);
5818 // Zero-extend the switch condition and case constants unless the switch
5819 // condition is a function argument that is already being sign-extended.
5820 // In that case, we can avoid an unnecessary mask/extension by sign-extending
5821 // everything instead.
5822 Instruction::CastOps ExtType = Instruction::ZExt;
5823 if (auto *Arg = dyn_cast<Argument>(Cond))
5824 if (Arg->hasSExtAttr())
5825 ExtType = Instruction::SExt;
5827 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
5828 ExtInst->insertBefore(SI);
5829 SI->setCondition(ExtInst);
5830 for (SwitchInst::CaseIt Case : SI->cases()) {
5831 APInt NarrowConst = Case.getCaseValue()->getValue();
5832 APInt WideConst = (ExtType == Instruction::ZExt) ?
5833 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
5834 Case.setValue(ConstantInt::get(Context, WideConst));
5841 /// \brief Helper class to promote a scalar operation to a vector one.
5842 /// This class is used to move downward extractelement transition.
5844 /// a = vector_op <2 x i32>
5845 /// b = extractelement <2 x i32> a, i32 0
5850 /// a = vector_op <2 x i32>
5851 /// c = vector_op a (equivalent to scalar_op on the related lane)
5852 /// * d = extractelement <2 x i32> c, i32 0
5854 /// Assuming both extractelement and store can be combine, we get rid of the
5856 class VectorPromoteHelper {
5857 /// DataLayout associated with the current module.
5858 const DataLayout &DL;
5860 /// Used to perform some checks on the legality of vector operations.
5861 const TargetLowering &TLI;
5863 /// Used to estimated the cost of the promoted chain.
5864 const TargetTransformInfo &TTI;
5866 /// The transition being moved downwards.
5867 Instruction *Transition;
5868 /// The sequence of instructions to be promoted.
5869 SmallVector<Instruction *, 4> InstsToBePromoted;
5870 /// Cost of combining a store and an extract.
5871 unsigned StoreExtractCombineCost;
5872 /// Instruction that will be combined with the transition.
5873 Instruction *CombineInst;
5875 /// \brief The instruction that represents the current end of the transition.
5876 /// Since we are faking the promotion until we reach the end of the chain
5877 /// of computation, we need a way to get the current end of the transition.
5878 Instruction *getEndOfTransition() const {
5879 if (InstsToBePromoted.empty())
5881 return InstsToBePromoted.back();
5884 /// \brief Return the index of the original value in the transition.
5885 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
5886 /// c, is at index 0.
5887 unsigned getTransitionOriginalValueIdx() const {
5888 assert(isa<ExtractElementInst>(Transition) &&
5889 "Other kind of transitions are not supported yet");
5893 /// \brief Return the index of the index in the transition.
5894 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
5896 unsigned getTransitionIdx() const {
5897 assert(isa<ExtractElementInst>(Transition) &&
5898 "Other kind of transitions are not supported yet");
5902 /// \brief Get the type of the transition.
5903 /// This is the type of the original value.
5904 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
5905 /// transition is <2 x i32>.
5906 Type *getTransitionType() const {
5907 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
5910 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
5911 /// I.e., we have the following sequence:
5912 /// Def = Transition <ty1> a to <ty2>
5913 /// b = ToBePromoted <ty2> Def, ...
5915 /// b = ToBePromoted <ty1> a, ...
5916 /// Def = Transition <ty1> ToBePromoted to <ty2>
5917 void promoteImpl(Instruction *ToBePromoted);
5919 /// \brief Check whether or not it is profitable to promote all the
5920 /// instructions enqueued to be promoted.
5921 bool isProfitableToPromote() {
5922 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
5923 unsigned Index = isa<ConstantInt>(ValIdx)
5924 ? cast<ConstantInt>(ValIdx)->getZExtValue()
5926 Type *PromotedType = getTransitionType();
5928 StoreInst *ST = cast<StoreInst>(CombineInst);
5929 unsigned AS = ST->getPointerAddressSpace();
5930 unsigned Align = ST->getAlignment();
5931 // Check if this store is supported.
5932 if (!TLI.allowsMisalignedMemoryAccesses(
5933 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
5935 // If this is not supported, there is no way we can combine
5936 // the extract with the store.
5940 // The scalar chain of computation has to pay for the transition
5941 // scalar to vector.
5942 // The vector chain has to account for the combining cost.
5943 uint64_t ScalarCost =
5944 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
5945 uint64_t VectorCost = StoreExtractCombineCost;
5946 for (const auto &Inst : InstsToBePromoted) {
5947 // Compute the cost.
5948 // By construction, all instructions being promoted are arithmetic ones.
5949 // Moreover, one argument is a constant that can be viewed as a splat
5951 Value *Arg0 = Inst->getOperand(0);
5952 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
5953 isa<ConstantFP>(Arg0);
5954 TargetTransformInfo::OperandValueKind Arg0OVK =
5955 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
5956 : TargetTransformInfo::OK_AnyValue;
5957 TargetTransformInfo::OperandValueKind Arg1OVK =
5958 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
5959 : TargetTransformInfo::OK_AnyValue;
5960 ScalarCost += TTI.getArithmeticInstrCost(
5961 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
5962 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
5965 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
5966 << ScalarCost << "\nVector: " << VectorCost << '\n');
5967 return ScalarCost > VectorCost;
5970 /// \brief Generate a constant vector with \p Val with the same
5971 /// number of elements as the transition.
5972 /// \p UseSplat defines whether or not \p Val should be replicated
5973 /// across the whole vector.
5974 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
5975 /// otherwise we generate a vector with as many undef as possible:
5976 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
5977 /// used at the index of the extract.
5978 Value *getConstantVector(Constant *Val, bool UseSplat) const {
5979 unsigned ExtractIdx = UINT_MAX;
5981 // If we cannot determine where the constant must be, we have to
5982 // use a splat constant.
5983 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
5984 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
5985 ExtractIdx = CstVal->getSExtValue();
5990 unsigned End = getTransitionType()->getVectorNumElements();
5992 return ConstantVector::getSplat(End, Val);
5994 SmallVector<Constant *, 4> ConstVec;
5995 UndefValue *UndefVal = UndefValue::get(Val->getType());
5996 for (unsigned Idx = 0; Idx != End; ++Idx) {
5997 if (Idx == ExtractIdx)
5998 ConstVec.push_back(Val);
6000 ConstVec.push_back(UndefVal);
6002 return ConstantVector::get(ConstVec);
6005 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
6006 /// in \p Use can trigger undefined behavior.
6007 static bool canCauseUndefinedBehavior(const Instruction *Use,
6008 unsigned OperandIdx) {
6009 // This is not safe to introduce undef when the operand is on
6010 // the right hand side of a division-like instruction.
6011 if (OperandIdx != 1)
6013 switch (Use->getOpcode()) {
6016 case Instruction::SDiv:
6017 case Instruction::UDiv:
6018 case Instruction::SRem:
6019 case Instruction::URem:
6021 case Instruction::FDiv:
6022 case Instruction::FRem:
6023 return !Use->hasNoNaNs();
6025 llvm_unreachable(nullptr);
6029 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
6030 const TargetTransformInfo &TTI, Instruction *Transition,
6031 unsigned CombineCost)
6032 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
6033 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
6034 assert(Transition && "Do not know how to promote null");
6037 /// \brief Check if we can promote \p ToBePromoted to \p Type.
6038 bool canPromote(const Instruction *ToBePromoted) const {
6039 // We could support CastInst too.
6040 return isa<BinaryOperator>(ToBePromoted);
6043 /// \brief Check if it is profitable to promote \p ToBePromoted
6044 /// by moving downward the transition through.
6045 bool shouldPromote(const Instruction *ToBePromoted) const {
6046 // Promote only if all the operands can be statically expanded.
6047 // Indeed, we do not want to introduce any new kind of transitions.
6048 for (const Use &U : ToBePromoted->operands()) {
6049 const Value *Val = U.get();
6050 if (Val == getEndOfTransition()) {
6051 // If the use is a division and the transition is on the rhs,
6052 // we cannot promote the operation, otherwise we may create a
6053 // division by zero.
6054 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
6058 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
6059 !isa<ConstantFP>(Val))
6062 // Check that the resulting operation is legal.
6063 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
6066 return StressStoreExtract ||
6067 TLI.isOperationLegalOrCustom(
6068 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
6071 /// \brief Check whether or not \p Use can be combined
6072 /// with the transition.
6073 /// I.e., is it possible to do Use(Transition) => AnotherUse?
6074 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
6076 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
6077 void enqueueForPromotion(Instruction *ToBePromoted) {
6078 InstsToBePromoted.push_back(ToBePromoted);
6081 /// \brief Set the instruction that will be combined with the transition.
6082 void recordCombineInstruction(Instruction *ToBeCombined) {
6083 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
6084 CombineInst = ToBeCombined;
6087 /// \brief Promote all the instructions enqueued for promotion if it is
6089 /// \return True if the promotion happened, false otherwise.
6091 // Check if there is something to promote.
6092 // Right now, if we do not have anything to combine with,
6093 // we assume the promotion is not profitable.
6094 if (InstsToBePromoted.empty() || !CombineInst)
6098 if (!StressStoreExtract && !isProfitableToPromote())
6102 for (auto &ToBePromoted : InstsToBePromoted)
6103 promoteImpl(ToBePromoted);
6104 InstsToBePromoted.clear();
6108 } // End of anonymous namespace.
6110 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
6111 // At this point, we know that all the operands of ToBePromoted but Def
6112 // can be statically promoted.
6113 // For Def, we need to use its parameter in ToBePromoted:
6114 // b = ToBePromoted ty1 a
6115 // Def = Transition ty1 b to ty2
6116 // Move the transition down.
6117 // 1. Replace all uses of the promoted operation by the transition.
6118 // = ... b => = ... Def.
6119 assert(ToBePromoted->getType() == Transition->getType() &&
6120 "The type of the result of the transition does not match "
6122 ToBePromoted->replaceAllUsesWith(Transition);
6123 // 2. Update the type of the uses.
6124 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
6125 Type *TransitionTy = getTransitionType();
6126 ToBePromoted->mutateType(TransitionTy);
6127 // 3. Update all the operands of the promoted operation with promoted
6129 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
6130 for (Use &U : ToBePromoted->operands()) {
6131 Value *Val = U.get();
6132 Value *NewVal = nullptr;
6133 if (Val == Transition)
6134 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
6135 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
6136 isa<ConstantFP>(Val)) {
6137 // Use a splat constant if it is not safe to use undef.
6138 NewVal = getConstantVector(
6139 cast<Constant>(Val),
6140 isa<UndefValue>(Val) ||
6141 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
6143 llvm_unreachable("Did you modified shouldPromote and forgot to update "
6145 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
6147 Transition->removeFromParent();
6148 Transition->insertAfter(ToBePromoted);
6149 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
6152 /// Some targets can do store(extractelement) with one instruction.
6153 /// Try to push the extractelement towards the stores when the target
6154 /// has this feature and this is profitable.
6155 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
6156 unsigned CombineCost = UINT_MAX;
6157 if (DisableStoreExtract || !TLI ||
6158 (!StressStoreExtract &&
6159 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
6160 Inst->getOperand(1), CombineCost)))
6163 // At this point we know that Inst is a vector to scalar transition.
6164 // Try to move it down the def-use chain, until:
6165 // - We can combine the transition with its single use
6166 // => we got rid of the transition.
6167 // - We escape the current basic block
6168 // => we would need to check that we are moving it at a cheaper place and
6169 // we do not do that for now.
6170 BasicBlock *Parent = Inst->getParent();
6171 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
6172 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
6173 // If the transition has more than one use, assume this is not going to be
6175 while (Inst->hasOneUse()) {
6176 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
6177 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
6179 if (ToBePromoted->getParent() != Parent) {
6180 DEBUG(dbgs() << "Instruction to promote is in a different block ("
6181 << ToBePromoted->getParent()->getName()
6182 << ") than the transition (" << Parent->getName() << ").\n");
6186 if (VPH.canCombine(ToBePromoted)) {
6187 DEBUG(dbgs() << "Assume " << *Inst << '\n'
6188 << "will be combined with: " << *ToBePromoted << '\n');
6189 VPH.recordCombineInstruction(ToBePromoted);
6190 bool Changed = VPH.promote();
6191 NumStoreExtractExposed += Changed;
6195 DEBUG(dbgs() << "Try promoting.\n");
6196 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
6199 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
6201 VPH.enqueueForPromotion(ToBePromoted);
6202 Inst = ToBePromoted;
6207 bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
6208 // Bail out if we inserted the instruction to prevent optimizations from
6209 // stepping on each other's toes.
6210 if (InsertedInsts.count(I))
6213 if (PHINode *P = dyn_cast<PHINode>(I)) {
6214 // It is possible for very late stage optimizations (such as SimplifyCFG)
6215 // to introduce PHI nodes too late to be cleaned up. If we detect such a
6216 // trivial PHI, go ahead and zap it here.
6217 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
6218 P->replaceAllUsesWith(V);
6219 P->eraseFromParent();
6226 if (CastInst *CI = dyn_cast<CastInst>(I)) {
6227 // If the source of the cast is a constant, then this should have
6228 // already been constant folded. The only reason NOT to constant fold
6229 // it is if something (e.g. LSR) was careful to place the constant
6230 // evaluation in a block other than then one that uses it (e.g. to hoist
6231 // the address of globals out of a loop). If this is the case, we don't
6232 // want to forward-subst the cast.
6233 if (isa<Constant>(CI->getOperand(0)))
6236 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
6239 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6240 /// Sink a zext or sext into its user blocks if the target type doesn't
6241 /// fit in one register
6243 TLI->getTypeAction(CI->getContext(),
6244 TLI->getValueType(*DL, CI->getType())) ==
6245 TargetLowering::TypeExpandInteger) {
6246 return SinkCast(CI);
6248 bool MadeChange = moveExtToFormExtLoad(I);
6249 return MadeChange | optimizeExtUses(I);
6255 if (CmpInst *CI = dyn_cast<CmpInst>(I))
6256 if (!TLI || !TLI->hasMultipleConditionRegisters())
6257 return OptimizeCmpExpression(CI);
6259 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
6260 stripInvariantGroupMetadata(*LI);
6262 bool Modified = optimizeLoadExt(LI);
6263 unsigned AS = LI->getPointerAddressSpace();
6264 Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
6270 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
6271 stripInvariantGroupMetadata(*SI);
6273 unsigned AS = SI->getPointerAddressSpace();
6274 return optimizeMemoryInst(I, SI->getOperand(1),
6275 SI->getOperand(0)->getType(), AS);
6280 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
6282 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
6283 BinOp->getOpcode() == Instruction::LShr)) {
6284 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
6285 if (TLI && CI && TLI->hasExtractBitsInsn())
6286 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
6291 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
6292 if (GEPI->hasAllZeroIndices()) {
6293 /// The GEP operand must be a pointer, so must its result -> BitCast
6294 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
6295 GEPI->getName(), GEPI);
6296 GEPI->replaceAllUsesWith(NC);
6297 GEPI->eraseFromParent();
6299 optimizeInst(NC, ModifiedDT);
6305 if (CallInst *CI = dyn_cast<CallInst>(I))
6306 return optimizeCallInst(CI, ModifiedDT);
6308 if (SelectInst *SI = dyn_cast<SelectInst>(I))
6309 return optimizeSelectInst(SI);
6311 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
6312 return optimizeShuffleVectorInst(SVI);
6314 if (auto *Switch = dyn_cast<SwitchInst>(I))
6315 return optimizeSwitchInst(Switch);
6317 if (isa<ExtractElementInst>(I))
6318 return optimizeExtractElementInst(I);
6323 /// Given an OR instruction, check to see if this is a bitreverse
6324 /// idiom. If so, insert the new intrinsic and return true.
6325 static bool makeBitReverse(Instruction &I, const DataLayout &DL,
6326 const TargetLowering &TLI) {
6327 if (!I.getType()->isIntegerTy() ||
6328 !TLI.isOperationLegalOrCustom(ISD::BITREVERSE,
6329 TLI.getValueType(DL, I.getType(), true)))
6332 SmallVector<Instruction*, 4> Insts;
6333 if (!recognizeBitReverseOrBSwapIdiom(&I, false, true, Insts))
6335 Instruction *LastInst = Insts.back();
6336 I.replaceAllUsesWith(LastInst);
6337 RecursivelyDeleteTriviallyDeadInstructions(&I);
6341 // In this pass we look for GEP and cast instructions that are used
6342 // across basic blocks and rewrite them to improve basic-block-at-a-time
6344 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
6346 bool MadeChange = false;
6348 CurInstIterator = BB.begin();
6349 while (CurInstIterator != BB.end()) {
6350 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
6355 bool MadeBitReverse = true;
6356 while (TLI && MadeBitReverse) {
6357 MadeBitReverse = false;
6358 for (auto &I : reverse(BB)) {
6359 if (makeBitReverse(I, *DL, *TLI)) {
6360 MadeBitReverse = MadeChange = true;
6365 MadeChange |= dupRetToEnableTailCallOpts(&BB);
6370 // llvm.dbg.value is far away from the value then iSel may not be able
6371 // handle it properly. iSel will drop llvm.dbg.value if it can not
6372 // find a node corresponding to the value.
6373 bool CodeGenPrepare::placeDbgValues(Function &F) {
6374 bool MadeChange = false;
6375 for (BasicBlock &BB : F) {
6376 Instruction *PrevNonDbgInst = nullptr;
6377 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
6378 Instruction *Insn = &*BI++;
6379 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
6380 // Leave dbg.values that refer to an alloca alone. These
6381 // instrinsics describe the address of a variable (= the alloca)
6382 // being taken. They should not be moved next to the alloca
6383 // (and to the beginning of the scope), but rather stay close to
6384 // where said address is used.
6385 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
6386 PrevNonDbgInst = Insn;
6390 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
6391 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
6392 // If VI is a phi in a block with an EHPad terminator, we can't insert
6394 if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
6396 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
6397 DVI->removeFromParent();
6398 if (isa<PHINode>(VI))
6399 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
6401 DVI->insertAfter(VI);
6410 // If there is a sequence that branches based on comparing a single bit
6411 // against zero that can be combined into a single instruction, and the
6412 // target supports folding these into a single instruction, sink the
6413 // mask and compare into the branch uses. Do this before OptimizeBlock ->
6414 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
6416 bool CodeGenPrepare::sinkAndCmp(Function &F) {
6417 if (!EnableAndCmpSinking)
6419 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
6421 bool MadeChange = false;
6422 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
6423 BasicBlock *BB = &*I++;
6425 // Does this BB end with the following?
6426 // %andVal = and %val, #single-bit-set
6427 // %icmpVal = icmp %andResult, 0
6428 // br i1 %cmpVal label %dest1, label %dest2"
6429 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
6430 if (!Brcc || !Brcc->isConditional())
6432 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
6433 if (!Cmp || Cmp->getParent() != BB)
6435 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
6436 if (!Zero || !Zero->isZero())
6438 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
6439 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
6441 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
6442 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
6444 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
6446 // Push the "and; icmp" for any users that are conditional branches.
6447 // Since there can only be one branch use per BB, we don't need to keep
6448 // track of which BBs we insert into.
6449 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
6453 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
6455 if (!BrccUser || !BrccUser->isConditional())
6457 BasicBlock *UserBB = BrccUser->getParent();
6458 if (UserBB == BB) continue;
6459 DEBUG(dbgs() << "found Brcc use\n");
6461 // Sink the "and; icmp" to use.
6463 BinaryOperator *NewAnd =
6464 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
6467 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
6471 DEBUG(BrccUser->getParent()->dump());
6477 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
6478 /// success, or returns false if no or invalid metadata was found.
6479 static bool extractBranchMetadata(BranchInst *BI,
6480 uint64_t &ProbTrue, uint64_t &ProbFalse) {
6481 assert(BI->isConditional() &&
6482 "Looking for probabilities on unconditional branch?");
6483 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
6484 if (!ProfileData || ProfileData->getNumOperands() != 3)
6487 const auto *CITrue =
6488 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
6489 const auto *CIFalse =
6490 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
6491 if (!CITrue || !CIFalse)
6494 ProbTrue = CITrue->getValue().getZExtValue();
6495 ProbFalse = CIFalse->getValue().getZExtValue();
6500 /// \brief Scale down both weights to fit into uint32_t.
6501 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
6502 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
6503 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
6504 NewTrue = NewTrue / Scale;
6505 NewFalse = NewFalse / Scale;
6508 /// \brief Some targets prefer to split a conditional branch like:
6510 /// %0 = icmp ne i32 %a, 0
6511 /// %1 = icmp ne i32 %b, 0
6512 /// %or.cond = or i1 %0, %1
6513 /// br i1 %or.cond, label %TrueBB, label %FalseBB
6515 /// into multiple branch instructions like:
6518 /// %0 = icmp ne i32 %a, 0
6519 /// br i1 %0, label %TrueBB, label %bb2
6521 /// %1 = icmp ne i32 %b, 0
6522 /// br i1 %1, label %TrueBB, label %FalseBB
6524 /// This usually allows instruction selection to do even further optimizations
6525 /// and combine the compare with the branch instruction. Currently this is
6526 /// applied for targets which have "cheap" jump instructions.
6528 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
6530 bool CodeGenPrepare::splitBranchCondition(Function &F) {
6531 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
6534 bool MadeChange = false;
6535 for (auto &BB : F) {
6536 // Does this BB end with the following?
6537 // %cond1 = icmp|fcmp|binary instruction ...
6538 // %cond2 = icmp|fcmp|binary instruction ...
6539 // %cond.or = or|and i1 %cond1, cond2
6540 // br i1 %cond.or label %dest1, label %dest2"
6541 BinaryOperator *LogicOp;
6542 BasicBlock *TBB, *FBB;
6543 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
6546 auto *Br1 = cast<BranchInst>(BB.getTerminator());
6547 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
6551 Value *Cond1, *Cond2;
6552 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
6553 m_OneUse(m_Value(Cond2)))))
6554 Opc = Instruction::And;
6555 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
6556 m_OneUse(m_Value(Cond2)))))
6557 Opc = Instruction::Or;
6561 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
6562 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
6565 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
6568 auto *InsertBefore = std::next(Function::iterator(BB))
6569 .getNodePtrUnchecked();
6570 auto TmpBB = BasicBlock::Create(BB.getContext(),
6571 BB.getName() + ".cond.split",
6572 BB.getParent(), InsertBefore);
6574 // Update original basic block by using the first condition directly by the
6575 // branch instruction and removing the no longer needed and/or instruction.
6576 Br1->setCondition(Cond1);
6577 LogicOp->eraseFromParent();
6579 // Depending on the conditon we have to either replace the true or the false
6580 // successor of the original branch instruction.
6581 if (Opc == Instruction::And)
6582 Br1->setSuccessor(0, TmpBB);
6584 Br1->setSuccessor(1, TmpBB);
6586 // Fill in the new basic block.
6587 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
6588 if (auto *I = dyn_cast<Instruction>(Cond2)) {
6589 I->removeFromParent();
6590 I->insertBefore(Br2);
6593 // Update PHI nodes in both successors. The original BB needs to be
6594 // replaced in one succesor's PHI nodes, because the branch comes now from
6595 // the newly generated BB (NewBB). In the other successor we need to add one
6596 // incoming edge to the PHI nodes, because both branch instructions target
6597 // now the same successor. Depending on the original branch condition
6598 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
6599 // we perfrom the correct update for the PHI nodes.
6600 // This doesn't change the successor order of the just created branch
6601 // instruction (or any other instruction).
6602 if (Opc == Instruction::Or)
6603 std::swap(TBB, FBB);
6605 // Replace the old BB with the new BB.
6606 for (auto &I : *TBB) {
6607 PHINode *PN = dyn_cast<PHINode>(&I);
6611 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
6612 PN->setIncomingBlock(i, TmpBB);
6615 // Add another incoming edge form the new BB.
6616 for (auto &I : *FBB) {
6617 PHINode *PN = dyn_cast<PHINode>(&I);
6620 auto *Val = PN->getIncomingValueForBlock(&BB);
6621 PN->addIncoming(Val, TmpBB);
6624 // Update the branch weights (from SelectionDAGBuilder::
6625 // FindMergedConditions).
6626 if (Opc == Instruction::Or) {
6627 // Codegen X | Y as:
6636 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
6637 // The requirement is that
6638 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
6639 // = TrueProb for orignal BB.
6640 // Assuming the orignal weights are A and B, one choice is to set BB1's
6641 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
6643 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
6644 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
6645 // TmpBB, but the math is more complicated.
6646 uint64_t TrueWeight, FalseWeight;
6647 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
6648 uint64_t NewTrueWeight = TrueWeight;
6649 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
6650 scaleWeights(NewTrueWeight, NewFalseWeight);
6651 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6652 .createBranchWeights(TrueWeight, FalseWeight));
6654 NewTrueWeight = TrueWeight;
6655 NewFalseWeight = 2 * FalseWeight;
6656 scaleWeights(NewTrueWeight, NewFalseWeight);
6657 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6658 .createBranchWeights(TrueWeight, FalseWeight));
6661 // Codegen X & Y as:
6669 // This requires creation of TmpBB after CurBB.
6671 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
6672 // The requirement is that
6673 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
6674 // = FalseProb for orignal BB.
6675 // Assuming the orignal weights are A and B, one choice is to set BB1's
6676 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
6678 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
6679 uint64_t TrueWeight, FalseWeight;
6680 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
6681 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
6682 uint64_t NewFalseWeight = FalseWeight;
6683 scaleWeights(NewTrueWeight, NewFalseWeight);
6684 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6685 .createBranchWeights(TrueWeight, FalseWeight));
6687 NewTrueWeight = 2 * TrueWeight;
6688 NewFalseWeight = FalseWeight;
6689 scaleWeights(NewTrueWeight, NewFalseWeight);
6690 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6691 .createBranchWeights(TrueWeight, FalseWeight));
6695 // Note: No point in getting fancy here, since the DT info is never
6696 // available to CodeGenPrepare.
6701 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
6707 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
6708 if (auto *InvariantMD = I.getMetadata(LLVMContext::MD_invariant_group))
6709 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());