1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/MemoryLocation.h"
25 #include "llvm/Analysis/TargetLibraryInfo.h"
26 #include "llvm/Analysis/TargetTransformInfo.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/IR/CallSite.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/Dominators.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/IR/GetElementPtrTypeIterator.h"
35 #include "llvm/IR/IRBuilder.h"
36 #include "llvm/IR/InlineAsm.h"
37 #include "llvm/IR/InstIterator.h"
38 #include "llvm/IR/InstrTypes.h"
39 #include "llvm/IR/Instructions.h"
40 #include "llvm/IR/IntrinsicInst.h"
41 #include "llvm/IR/MDBuilder.h"
42 #include "llvm/IR/NoFolder.h"
43 #include "llvm/IR/PatternMatch.h"
44 #include "llvm/IR/Statepoint.h"
45 #include "llvm/IR/ValueHandle.h"
46 #include "llvm/IR/ValueMap.h"
47 #include "llvm/Pass.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/raw_ostream.h"
51 #include "llvm/Target/TargetLowering.h"
52 #include "llvm/Target/TargetSubtargetInfo.h"
53 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
54 #include "llvm/Transforms/Utils/BuildLibCalls.h"
55 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
56 #include "llvm/Transforms/Utils/Local.h"
57 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
59 using namespace llvm::PatternMatch;
61 #define DEBUG_TYPE "codegenprepare"
63 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
64 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
65 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
66 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
68 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
70 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
71 "computations were sunk");
72 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
73 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
74 STATISTIC(NumAndsAdded,
75 "Number of and mask instructions added to form ext loads");
76 STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
77 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
78 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
79 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
80 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
81 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
83 static cl::opt<bool> DisableBranchOpts(
84 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
85 cl::desc("Disable branch optimizations in CodeGenPrepare"));
88 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
89 cl::desc("Disable GC optimizations in CodeGenPrepare"));
91 static cl::opt<bool> DisableSelectToBranch(
92 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
93 cl::desc("Disable select to branch conversion."));
95 static cl::opt<bool> AddrSinkUsingGEPs(
96 "addr-sink-using-gep", cl::Hidden, cl::init(false),
97 cl::desc("Address sinking in CGP using GEPs."));
99 static cl::opt<bool> EnableAndCmpSinking(
100 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
101 cl::desc("Enable sinkinig and/cmp into branches."));
103 static cl::opt<bool> DisableStoreExtract(
104 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
105 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
107 static cl::opt<bool> StressStoreExtract(
108 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
109 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
111 static cl::opt<bool> DisableExtLdPromotion(
112 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
113 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
116 static cl::opt<bool> StressExtLdPromotion(
117 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
118 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
119 "optimization in CodeGenPrepare"));
122 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
123 typedef PointerIntPair<Type *, 1, bool> TypeIsSExt;
124 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
125 class TypePromotionTransaction;
127 class CodeGenPrepare : public FunctionPass {
128 const TargetMachine *TM;
129 const TargetLowering *TLI;
130 const TargetTransformInfo *TTI;
131 const TargetLibraryInfo *TLInfo;
133 /// As we scan instructions optimizing them, this is the next instruction
134 /// to optimize. Transforms that can invalidate this should update it.
135 BasicBlock::iterator CurInstIterator;
137 /// Keeps track of non-local addresses that have been sunk into a block.
138 /// This allows us to avoid inserting duplicate code for blocks with
139 /// multiple load/stores of the same address.
140 ValueMap<Value*, Value*> SunkAddrs;
142 /// Keeps track of all instructions inserted for the current function.
143 SetOfInstrs InsertedInsts;
144 /// Keeps track of the type of the related instruction before their
145 /// promotion for the current function.
146 InstrToOrigTy PromotedInsts;
148 /// True if CFG is modified in any way.
151 /// True if optimizing for size.
154 /// DataLayout for the Function being processed.
155 const DataLayout *DL;
157 // XXX-comment:We need DominatorTree to figure out which instruction to
162 static char ID; // Pass identification, replacement for typeid
163 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
164 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr), DL(nullptr),
166 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
168 bool runOnFunction(Function &F) override;
170 const char *getPassName() const override { return "CodeGen Prepare"; }
172 void getAnalysisUsage(AnalysisUsage &AU) const override {
173 AU.addPreserved<DominatorTreeWrapperPass>();
174 AU.addRequired<TargetLibraryInfoWrapperPass>();
175 AU.addRequired<TargetTransformInfoWrapperPass>();
176 AU.addRequired<DominatorTreeWrapperPass>();
180 bool eliminateFallThrough(Function &F);
181 bool eliminateMostlyEmptyBlocks(Function &F);
182 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
183 void eliminateMostlyEmptyBlock(BasicBlock *BB);
184 bool optimizeBlock(BasicBlock &BB, bool& ModifiedDT);
185 bool optimizeInst(Instruction *I, bool& ModifiedDT);
186 bool optimizeMemoryInst(Instruction *I, Value *Addr,
187 Type *AccessTy, unsigned AS);
188 bool optimizeInlineAsmInst(CallInst *CS);
189 bool optimizeCallInst(CallInst *CI, bool& ModifiedDT);
190 bool moveExtToFormExtLoad(Instruction *&I);
191 bool optimizeExtUses(Instruction *I);
192 bool optimizeLoadExt(LoadInst *I);
193 bool optimizeSelectInst(SelectInst *SI);
194 bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
195 bool optimizeSwitchInst(SwitchInst *CI);
196 bool optimizeExtractElementInst(Instruction *Inst);
197 bool dupRetToEnableTailCallOpts(BasicBlock *BB);
198 bool placeDbgValues(Function &F);
199 bool sinkAndCmp(Function &F);
200 bool extLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
202 const SmallVectorImpl<Instruction *> &Exts,
203 unsigned CreatedInstCost);
204 bool splitBranchCondition(Function &F);
205 bool simplifyOffsetableRelocate(Instruction &I);
206 void stripInvariantGroupMetadata(Instruction &I);
210 char CodeGenPrepare::ID = 0;
211 INITIALIZE_TM_PASS_BEGIN(CodeGenPrepare, "codegenprepare",
212 "Optimize for code generation", false, false)
213 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
214 INITIALIZE_TM_PASS_END(CodeGenPrepare, "codegenprepare",
215 "Optimize for code generation", false, false)
217 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
218 return new CodeGenPrepare(TM);
223 bool StoreAddressDependOnValue(StoreInst* SI, Value* DepVal);
224 Value* GetUntaintedAddress(Value* CurrentAddress);
226 // The depth we trace down a variable to look for its dependence set.
227 const unsigned kDependenceDepth = 4;
229 // Recursively looks for variables that 'Val' depends on at the given depth
230 // 'Depth', and adds them in 'DepSet'. If 'InsertOnlyLeafNodes' is true, only
231 // inserts the leaf node values; otherwise, all visited nodes are included in
232 // 'DepSet'. Note that constants will be ignored.
233 template <typename SetType>
234 void recursivelyFindDependence(SetType* DepSet, Value* Val,
235 bool InsertOnlyLeafNodes = false,
236 unsigned Depth = kDependenceDepth) {
237 if (Val == nullptr) {
240 if (!InsertOnlyLeafNodes && !isa<Constant>(Val)) {
244 // Cannot go deeper. Insert the leaf nodes.
245 if (InsertOnlyLeafNodes && !isa<Constant>(Val)) {
251 // Go one step further to explore the dependence of the operands.
252 Instruction* I = nullptr;
253 if ((I = dyn_cast<Instruction>(Val))) {
254 if (isa<LoadInst>(I)) {
255 // A load is considerd the leaf load of the dependence tree. Done.
258 } else if (I->isBinaryOp()) {
259 BinaryOperator* I = dyn_cast<BinaryOperator>(Val);
260 Value *Op0 = I->getOperand(0), *Op1 = I->getOperand(1);
261 recursivelyFindDependence(DepSet, Op0, InsertOnlyLeafNodes, Depth - 1);
262 recursivelyFindDependence(DepSet, Op1, InsertOnlyLeafNodes, Depth - 1);
263 } else if (I->isCast()) {
264 Value* Op0 = I->getOperand(0);
265 recursivelyFindDependence(DepSet, Op0, InsertOnlyLeafNodes, Depth - 1);
266 } else if (I->getOpcode() == Instruction::Select) {
267 Value* Op0 = I->getOperand(0);
268 Value* Op1 = I->getOperand(1);
269 Value* Op2 = I->getOperand(2);
270 recursivelyFindDependence(DepSet, Op0, InsertOnlyLeafNodes, Depth - 1);
271 recursivelyFindDependence(DepSet, Op1, InsertOnlyLeafNodes, Depth - 1);
272 recursivelyFindDependence(DepSet, Op2, InsertOnlyLeafNodes, Depth - 1);
273 } else if (I->getOpcode() == Instruction::GetElementPtr) {
274 for (unsigned i = 0; i < I->getNumOperands(); i++) {
275 recursivelyFindDependence(DepSet, I->getOperand(i), InsertOnlyLeafNodes,
278 } else if (I->getOpcode() == Instruction::Store) {
279 auto* SI = dyn_cast<StoreInst>(Val);
280 recursivelyFindDependence(DepSet, SI->getPointerOperand(),
281 InsertOnlyLeafNodes, Depth - 1);
282 recursivelyFindDependence(DepSet, SI->getValueOperand(),
283 InsertOnlyLeafNodes, Depth - 1);
285 Value* Op0 = nullptr;
286 Value* Op1 = nullptr;
287 switch (I->getOpcode()) {
288 case Instruction::ICmp:
289 case Instruction::FCmp: {
290 Op0 = I->getOperand(0);
291 Op1 = I->getOperand(1);
292 recursivelyFindDependence(DepSet, Op0, InsertOnlyLeafNodes,
294 recursivelyFindDependence(DepSet, Op1, InsertOnlyLeafNodes,
298 case Instruction::PHI: {
299 for (int i = 0; i < I->getNumOperands(); i++) {
300 auto* op = I->getOperand(i);
301 if (DepSet->count(op) == 0) {
302 recursivelyFindDependence(DepSet, I->getOperand(i),
303 InsertOnlyLeafNodes, Depth - 1);
309 // Be conservative. Add it and be done with it.
315 } else if (isa<Constant>(Val)) {
316 // Not interested in constant values. Done.
319 // Be conservative. Add it and be done with it.
325 // Helper function to create a Cast instruction.
326 Value* createCast(IRBuilder<true, NoFolder>& Builder, Value* DepVal,
327 Type* TargetIntegerType) {
328 Instruction::CastOps CastOp = Instruction::BitCast;
329 switch (DepVal->getType()->getTypeID()) {
330 case Type::IntegerTyID: {
331 CastOp = Instruction::SExt;
334 case Type::FloatTyID:
335 case Type::DoubleTyID: {
336 CastOp = Instruction::FPToSI;
339 case Type::PointerTyID: {
340 CastOp = Instruction::PtrToInt;
346 return Builder.CreateCast(CastOp, DepVal, TargetIntegerType);
349 // Given a value, if it's a tainted address, this function returns the
350 // instruction that ORs the "dependence value" with the "original address".
351 // Otherwise, returns nullptr. This instruction is the first OR instruction
352 // where one of its operand is an AND instruction with an operand being 0.
354 // E.g., it returns '%4 = or i32 %3, %2' given 'CurrentAddress' is '%5'.
355 // %0 = load i32, i32* @y, align 4, !tbaa !1
356 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
357 // %1 = sext i1 %cmp to i32
358 // %2 = ptrtoint i32* @x to i32
359 // %3 = and i32 %1, 0
360 // %4 = or i32 %3, %2
361 // %5 = inttoptr i32 %4 to i32*
362 // store i32 1, i32* %5, align 4
363 Instruction* getOrAddress(Value* CurrentAddress) {
364 // Is it a cast from integer to pointer type.
365 Instruction* OrAddress = nullptr;
366 Instruction* AndDep = nullptr;
367 Instruction* CastToInt = nullptr;
368 Value* ActualAddress = nullptr;
369 Constant* ZeroConst = nullptr;
371 const Instruction* CastToPtr = dyn_cast<Instruction>(CurrentAddress);
372 if (CastToPtr && CastToPtr->getOpcode() == Instruction::IntToPtr) {
373 // Is it an OR instruction: %1 = or %and, %actualAddress.
374 if ((OrAddress = dyn_cast<Instruction>(CastToPtr->getOperand(0))) &&
375 OrAddress->getOpcode() == Instruction::Or) {
376 // The first operand should be and AND instruction.
377 AndDep = dyn_cast<Instruction>(OrAddress->getOperand(0));
378 if (AndDep && AndDep->getOpcode() == Instruction::And) {
379 // Also make sure its first operand of the "AND" is 0, or the "AND" is
380 // marked explicitly by "NoInstCombine".
381 if ((ZeroConst = dyn_cast<Constant>(AndDep->getOperand(1))) &&
382 ZeroConst->isNullValue()) {
388 // Looks like it's not been tainted.
392 // Given a value, if it's a tainted address, this function returns the
393 // instruction that taints the "dependence value". Otherwise, returns nullptr.
394 // This instruction is the last AND instruction where one of its operand is 0.
395 // E.g., it returns '%3' given 'CurrentAddress' is '%5'.
396 // %0 = load i32, i32* @y, align 4, !tbaa !1
397 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
398 // %1 = sext i1 %cmp to i32
399 // %2 = ptrtoint i32* @x to i32
400 // %3 = and i32 %1, 0
401 // %4 = or i32 %3, %2
402 // %5 = inttoptr i32 %4 to i32*
403 // store i32 1, i32* %5, align 4
404 Instruction* getAndDependence(Value* CurrentAddress) {
405 // If 'CurrentAddress' is tainted, get the OR instruction.
406 auto* OrAddress = getOrAddress(CurrentAddress);
407 if (OrAddress == nullptr) {
411 // No need to check the operands.
412 auto* AndDepInst = dyn_cast<Instruction>(OrAddress->getOperand(0));
417 // Given a value, if it's a tainted address, this function returns
418 // the "dependence value", which is the first operand in the AND instruction.
419 // E.g., it returns '%1' given 'CurrentAddress' is '%5'.
420 // %0 = load i32, i32* @y, align 4, !tbaa !1
421 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
422 // %1 = sext i1 %cmp to i32
423 // %2 = ptrtoint i32* @x to i32
424 // %3 = and i32 %1, 0
425 // %4 = or i32 %3, %2
426 // %5 = inttoptr i32 %4 to i32*
427 // store i32 1, i32* %5, align 4
428 Value* getDependence(Value* CurrentAddress) {
429 auto* AndInst = getAndDependence(CurrentAddress);
430 if (AndInst == nullptr) {
433 return AndInst->getOperand(0);
436 // Given an address that has been tainted, returns the only condition it depends
437 // on, if any; otherwise, returns nullptr.
438 Value* getConditionDependence(Value* Address) {
439 auto* Dep = getDependence(Address);
440 if (Dep == nullptr) {
441 // 'Address' has not been dependence-tainted.
445 Value* Operand = Dep;
447 auto* Inst = dyn_cast<Instruction>(Operand);
448 if (Inst == nullptr) {
449 // Non-instruction type does not have condition dependence.
452 if (Inst->getOpcode() == Instruction::ICmp) {
455 if (Inst->getNumOperands() != 1) {
458 Operand = Inst->getOperand(0);
464 // Conservatively decides whether the dependence set of 'Val1' includes the
465 // dependence set of 'Val2'. If 'ExpandSecondValue' is false, we do not expand
466 // 'Val2' and use that single value as its dependence set.
467 // If it returns true, it means the dependence set of 'Val1' includes that of
468 // 'Val2'; otherwise, it only means we cannot conclusively decide it.
469 bool dependenceSetInclusion(Value* Val1, Value* Val2,
470 int Val1ExpandLevel = 2 * kDependenceDepth,
471 int Val2ExpandLevel = kDependenceDepth) {
472 typedef SmallSet<Value*, 8> IncludingSet;
473 typedef SmallSet<Value*, 4> IncludedSet;
475 IncludingSet DepSet1;
477 // Look for more depths for the including set.
478 recursivelyFindDependence(&DepSet1, Val1, false /*Insert all visited nodes*/,
480 recursivelyFindDependence(&DepSet2, Val2, true /*Only insert leaf nodes*/,
483 auto set_inclusion = [](IncludingSet FullSet, IncludedSet Subset) {
484 for (auto* Dep : Subset) {
485 if (0 == FullSet.count(Dep)) {
491 bool inclusion = set_inclusion(DepSet1, DepSet2);
492 DEBUG(dbgs() << "[dependenceSetInclusion]: " << inclusion << "\n");
493 DEBUG(dbgs() << "Including set for: " << *Val1 << "\n");
494 DEBUG(for (const auto* Dep : DepSet1) { dbgs() << "\t\t" << *Dep << "\n"; });
495 DEBUG(dbgs() << "Included set for: " << *Val2 << "\n");
496 DEBUG(for (const auto* Dep : DepSet2) { dbgs() << "\t\t" << *Dep << "\n"; });
501 // Recursively iterates through the operands spawned from 'DepVal'. If there
502 // exists a single value that 'DepVal' only depends on, we call that value the
503 // root dependence of 'DepVal' and return it. Otherwise, return 'DepVal'.
504 Value* getRootDependence(Value* DepVal) {
505 SmallSet<Value*, 8> DepSet;
506 for (unsigned depth = kDependenceDepth; depth > 0; --depth) {
507 recursivelyFindDependence(&DepSet, DepVal, true /*Only insert leaf nodes*/,
509 if (DepSet.size() == 1) {
510 return *DepSet.begin();
517 // This function actually taints 'DepVal' to the address to 'SI'. If the
519 // of 'SI' already depends on whatever 'DepVal' depends on, this function
520 // doesn't do anything and returns false. Otherwise, returns true.
522 // This effect forces the store and any stores that comes later to depend on
523 // 'DepVal'. For example, we have a condition "cond", and a store instruction
524 // "s: STORE addr, val". If we want "s" (and any later store) to depend on
525 // "cond", we do the following:
526 // %conv = sext i1 %cond to i32
527 // %addrVal = ptrtoint i32* %addr to i32
528 // %andCond = and i32 conv, 0;
529 // %orAddr = or i32 %andCond, %addrVal;
530 // %NewAddr = inttoptr i32 %orAddr to i32*;
532 // This is a more concrete example:
534 // %0 = load i32, i32* @y, align 4, !tbaa !1
535 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
536 // %1 = sext i1 %cmp to i32
537 // %2 = ptrtoint i32* @x to i32
538 // %3 = and i32 %1, 0
539 // %4 = or i32 %3, %2
540 // %5 = inttoptr i32 %4 to i32*
541 // store i32 1, i32* %5, align 4
542 bool taintStoreAddress(StoreInst* SI, Value* DepVal) {
543 // Set the insertion point right after the 'DepVal'.
544 Instruction* Inst = nullptr;
545 IRBuilder<true, NoFolder> Builder(SI);
546 BasicBlock* BB = SI->getParent();
547 Value* Address = SI->getPointerOperand();
548 Type* TargetIntegerType =
549 IntegerType::get(Address->getContext(),
550 BB->getModule()->getDataLayout().getPointerSizeInBits());
552 // Does SI's address already depends on whatever 'DepVal' depends on?
553 if (StoreAddressDependOnValue(SI, DepVal)) {
557 // Figure out if there's a root variable 'DepVal' depends on. For example, we
558 // can extract "getelementptr inbounds %struct, %struct* %0, i64 0, i32 123"
559 // to be "%struct* %0" since all other operands are constant.
560 auto* RootVal = getRootDependence(DepVal);
561 auto* RootInst = dyn_cast<Instruction>(RootVal);
562 auto* DepValInst = dyn_cast<Instruction>(DepVal);
563 if (RootInst && DepValInst &&
564 RootInst->getParent() == DepValInst->getParent()) {
568 // Is this already a dependence-tainted store?
569 Value* OldDep = getDependence(Address);
571 // The address of 'SI' has already been tainted. Just need to absorb the
572 // DepVal to the existing dependence in the address of SI.
573 Instruction* AndDep = getAndDependence(Address);
574 IRBuilder<true, NoFolder> Builder(AndDep);
575 Value* NewDep = nullptr;
576 if (DepVal->getType() == AndDep->getType()) {
577 NewDep = Builder.CreateAnd(OldDep, DepVal);
579 NewDep = Builder.CreateAnd(
580 OldDep, createCast(Builder, DepVal, TargetIntegerType));
583 auto* NewDepInst = dyn_cast<Instruction>(NewDep);
585 // Use the new AND instruction as the dependence
586 AndDep->setOperand(0, NewDep);
590 // SI's address has not been tainted. Now taint it with 'DepVal'.
591 Value* CastDepToInt = createCast(Builder, DepVal, TargetIntegerType);
592 Value* PtrToIntCast = Builder.CreatePtrToInt(Address, TargetIntegerType);
594 Builder.CreateAnd(CastDepToInt, ConstantInt::get(TargetIntegerType, 0));
595 auto AndInst = dyn_cast<Instruction>(AndDepVal);
596 // XXX-comment: The original IR InstCombiner would change our and instruction
597 // to a select and then the back end optimize the condition out. We attach a
598 // flag to instructions and set it here to inform the InstCombiner to not to
599 // touch this and instruction at all.
600 Value* OrAddr = Builder.CreateOr(AndDepVal, PtrToIntCast);
601 Value* NewAddr = Builder.CreateIntToPtr(OrAddr, Address->getType());
603 DEBUG(dbgs() << "[taintStoreAddress]\n"
604 << "Original store: " << *SI << '\n');
605 SI->setOperand(1, NewAddr);
608 DEBUG(dbgs() << "\tTargetIntegerType: " << *TargetIntegerType << '\n'
609 << "\tCast dependence value to integer: " << *CastDepToInt
611 << "\tCast address to integer: " << *PtrToIntCast << '\n'
612 << "\tAnd dependence value: " << *AndDepVal << '\n'
613 << "\tOr address: " << *OrAddr << '\n'
614 << "\tCast or instruction to address: " << *NewAddr << "\n\n");
619 // Looks for the previous store in the if block --- 'BrBB', which makes the
620 // speculative store 'StoreToHoist' safe.
621 Value* getSpeculativeStoreInPrevBB(StoreInst* StoreToHoist, BasicBlock* BrBB) {
622 assert(StoreToHoist && "StoreToHoist must be a real store");
624 Value* StorePtr = StoreToHoist->getPointerOperand();
626 // Look for a store to the same pointer in BrBB.
627 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(), RE = BrBB->rend();
629 Instruction* CurI = &*RI;
631 StoreInst* SI = dyn_cast<StoreInst>(CurI);
632 // Found the previous store make sure it stores to the same location.
633 // XXX-update: If the previous store's original untainted address are the
634 // same as 'StorePtr', we are also good to hoist the store.
635 if (SI && (SI->getPointerOperand() == StorePtr ||
636 GetUntaintedAddress(SI->getPointerOperand()) == StorePtr)) {
637 // Found the previous store, return its value operand.
643 "We should not reach here since this store is safe to speculate");
646 // XXX-comment: Returns true if it changes the code, false otherwise (the branch
647 // condition already depends on 'DepVal'.
648 bool taintConditionalBranch(BranchInst* BI, Value* DepVal) {
649 assert(BI->isConditional());
650 auto* Cond = BI->getOperand(0);
651 if (dependenceSetInclusion(Cond, DepVal)) {
652 // The dependence/ordering is self-evident.
656 IRBuilder<true, NoFolder> Builder(BI);
658 Builder.CreateAnd(DepVal, ConstantInt::get(DepVal->getType(), 0));
660 Builder.CreateTrunc(AndDep, IntegerType::get(DepVal->getContext(), 1));
661 auto* OrCond = Builder.CreateOr(TruncAndDep, Cond);
662 BI->setOperand(0, OrCond);
665 DEBUG(dbgs() << "\tTainted branch condition:\n" << *BI->getParent());
670 bool ConditionalBranchDependsOnValue(BranchInst* BI, Value* DepVal) {
671 assert(BI->isConditional());
672 auto* Cond = BI->getOperand(0);
673 return dependenceSetInclusion(Cond, DepVal);
676 // XXX-update: For a relaxed load 'LI', find the first immediate atomic store or
677 // the first conditional branch. Returns nullptr if there's no such immediately
678 // following store/branch instructions, which we can only enforce the load with
679 // 'acquire'. 'ChainedBB' contains all the blocks chained together with
680 // unconditional branches from 'BB' to the block with the first store/cond
682 template <typename Vector>
683 Instruction* findFirstStoreCondBranchInst(LoadInst* LI, Vector* ChainedBB) {
684 // In some situations, relaxed loads can be left as is:
685 // 1. The relaxed load is used to calculate the address of the immediate
687 // 2. The relaxed load is used as a condition in the immediate following
688 // condition, and there are no stores in between. This is actually quite
690 // int r1 = x.load(relaxed);
692 // y.store(1, relaxed);
694 // However, in this function, we don't deal with them directly. Instead, we
695 // just find the immediate following store/condition branch and return it.
697 assert(ChainedBB != nullptr && "Chained BB should not be nullptr");
698 auto* BB = LI->getParent();
699 ChainedBB->push_back(BB);
701 auto BBI = BasicBlock::iterator(LI);
704 for (; BBI != BE; BBI++) {
705 auto* Inst = dyn_cast<Instruction>(&*BBI);
706 if (Inst == nullptr) {
709 if (Inst->getOpcode() == Instruction::Store) {
711 } else if (Inst->getOpcode() == Instruction::Br) {
712 auto* BrInst = dyn_cast<BranchInst>(Inst);
713 if (BrInst->isConditional()) {
716 // Reinitialize iterators with the destination of the unconditional
718 BB = BrInst->getSuccessor(0);
719 ChainedBB->push_back(BB);
732 // XXX-comment: Returns whether the code has been changed.
733 bool taintMonotonicLoads(const SmallVector<LoadInst*, 1>& MonotonicLoadInsts) {
734 bool Changed = false;
735 for (auto* LI : MonotonicLoadInsts) {
736 SmallVector<BasicBlock*, 2> ChainedBB;
737 auto* FirstInst = findFirstStoreCondBranchInst(LI, &ChainedBB);
738 if (FirstInst == nullptr) {
739 // We don't seem to be able to taint a following store/conditional branch
740 // instruction. Simply make it acquire.
741 DEBUG(dbgs() << "[RelaxedLoad]: Transformed to acquire load\n"
743 LI->setOrdering(Acquire);
747 // Taint 'FirstInst', which could be a store or a condition branch
749 if (FirstInst->getOpcode() == Instruction::Store) {
750 Changed |= taintStoreAddress(dyn_cast<StoreInst>(FirstInst), LI);
751 } else if (FirstInst->getOpcode() == Instruction::Br) {
752 Changed |= taintConditionalBranch(dyn_cast<BranchInst>(FirstInst), LI);
754 assert(false && "findFirstStoreCondBranchInst() should return a "
755 "store/condition branch instruction");
761 // Inserts a fake conditional branch right after the instruction 'SplitInst',
762 // and the branch condition is 'Condition'. 'SplitInst' will be placed in the
763 // newly created block.
764 void AddFakeConditionalBranch(Instruction* SplitInst, Value* Condition) {
765 auto* BB = SplitInst->getParent();
766 TerminatorInst* ThenTerm = nullptr;
767 TerminatorInst* ElseTerm = nullptr;
768 SplitBlockAndInsertIfThenElse(Condition, SplitInst, &ThenTerm, &ElseTerm);
769 assert(ThenTerm && ElseTerm &&
770 "Then/Else terminators cannot be empty after basic block spliting");
771 auto* ThenBB = ThenTerm->getParent();
772 auto* ElseBB = ElseTerm->getParent();
773 auto* TailBB = ThenBB->getSingleSuccessor();
774 assert(TailBB && "Tail block cannot be empty after basic block spliting");
776 ThenBB->disableCanEliminateBlock();
777 ThenBB->disableCanEliminateBlock();
778 TailBB->disableCanEliminateBlock();
779 ThenBB->setName(BB->getName() + "Then.Fake");
780 ElseBB->setName(BB->getName() + "Else.Fake");
781 DEBUG(dbgs() << "Add fake conditional branch:\n"
783 << *ThenBB << "Else Block:\n"
787 // Returns true if the code is changed, and false otherwise.
788 void TaintRelaxedLoads(Instruction* UsageInst) {
789 // For better performance, we can add a "AND X 0" instruction before the
791 auto* BB = UsageInst->getParent();
792 auto* InsertPoint = UsageInst->getNextNode();
793 IRBuilder<true, NoFolder> Builder(InsertPoint);
794 // First thing is to cast 'UsageInst' to an integer type if necessary.
795 Value* AndTarget = nullptr;
796 if (IntegerType::classof(UsageInst->getType())) {
797 AndTarget = UsageInst;
799 Type* TargetIntegerType = IntegerType::get(
800 UsageInst->getContext(),
801 BB->getModule()->getDataLayout().getPointerSizeInBits());
802 AndTarget = createCast(Builder, UsageInst, TargetIntegerType);
805 auto* AndZero = dyn_cast<Instruction>(
806 Builder.CreateAnd(AndTarget, Constant::getNullValue(AndTarget->getType())));
807 auto* FakeCondition = dyn_cast<Instruction>(Builder.CreateICmp(
808 CmpInst::ICMP_NE, AndZero, Constant::getNullValue(AndTarget->getType())));
809 AddFakeConditionalBranch(FakeCondition->getNextNode(), FakeCondition);
812 // XXX-comment: Finds the appropriate Value derived from an atomic load.
813 // 'ChainedBB' contains all the blocks chained together with unconditional
814 // branches from LI's parent BB to the block with the first store/cond branch.
815 // If we don't find any, it means 'LI' is not used at all (which should not
816 // happen in practice). We can simply set 'LI' to be acquire just to be safe.
817 template <typename Vector>
818 Instruction* findMostRecentDependenceUsage(LoadInst* LI, Instruction* LaterInst,
821 typedef SmallSet<Instruction*, 8> UsageSet;
822 typedef DenseMap<BasicBlock*, std::unique_ptr<UsageSet>> UsageMap;
823 assert(ChainedBB->size() >= 1 && "ChainedBB must have >=1 size");
824 // Mapping from basic block in 'ChainedBB' to the set of dependence usage of
825 // 'LI' in each block.
827 auto* LoadBB = LI->getParent();
828 usage_map[LoadBB] = make_unique<UsageSet>();
829 usage_map[LoadBB]->insert(LI);
831 for (auto* BB : *ChainedBB) {
832 if (usage_map[BB] == nullptr) {
833 usage_map[BB] = make_unique<UsageSet>();
835 auto& usage_set = usage_map[BB];
836 if (usage_set->size() == 0) {
837 // The value has not been used.
840 // Calculate the usage in the current BB first.
841 std::list<Value*> bb_usage_list;
842 std::copy(usage_set->begin(), usage_set->end(),
843 std::back_inserter(bb_usage_list));
844 for (auto list_iter = bb_usage_list.begin();
845 list_iter != bb_usage_list.end(); list_iter++) {
846 auto* val = *list_iter;
847 for (auto* U : val->users()) {
848 Instruction* Inst = nullptr;
849 if (!(Inst = dyn_cast<Instruction>(U))) {
852 assert(Inst && "Usage value must be an instruction");
854 std::find(ChainedBB->begin(), ChainedBB->end(), Inst->getParent());
855 if (iter == ChainedBB->end()) {
856 // Only care about usage within ChainedBB.
859 auto* UsageBB = *iter;
862 if (!usage_set->count(Inst)) {
863 bb_usage_list.push_back(Inst);
864 usage_set->insert(Inst);
868 if (usage_map[UsageBB] == nullptr) {
869 usage_map[UsageBB] = make_unique<UsageSet>();
871 usage_map[UsageBB]->insert(Inst);
877 // Pick one usage that is in LaterInst's block and that dominates 'LaterInst'.
878 auto* LaterBB = LaterInst->getParent();
879 auto& usage_set = usage_map[LaterBB];
880 Instruction* usage_inst = nullptr;
881 for (auto* inst : *usage_set) {
882 if (DT->dominates(inst, LaterInst)) {
888 assert(usage_inst && "The usage instruction in the same block but after the "
889 "later instruction");
893 // XXX-comment: Returns whether the code has been changed.
894 bool AddFakeConditionalBranchAfterMonotonicLoads(
895 const SmallVector<LoadInst*, 1>& MonotonicLoadInsts, DominatorTree* DT) {
896 bool Changed = false;
897 for (auto* LI : MonotonicLoadInsts) {
898 SmallVector<BasicBlock*, 2> ChainedBB;
899 auto* FirstInst = findFirstStoreCondBranchInst(LI, &ChainedBB);
900 if (FirstInst != nullptr) {
901 if (FirstInst->getOpcode() == Instruction::Store) {
902 if (StoreAddressDependOnValue(dyn_cast<StoreInst>(FirstInst), LI)) {
905 } else if (FirstInst->getOpcode() == Instruction::Br) {
906 if (ConditionalBranchDependsOnValue(dyn_cast<BranchInst>(FirstInst),
911 dbgs() << "FirstInst=" << *FirstInst << "\n";
912 assert(false && "findFirstStoreCondBranchInst() should return a "
913 "store/condition branch instruction");
917 // We really need to process the relaxed load now.
918 StoreInst* SI = nullptr;;
919 if (FirstInst && (SI = dyn_cast<StoreInst>(FirstInst))) {
920 // For immediately coming stores, taint the address of the store.
921 if (SI->getParent() == LI->getParent() || DT->dominates(LI, SI)) {
922 Changed |= taintStoreAddress(SI, LI);
925 findMostRecentDependenceUsage(LI, FirstInst, &ChainedBB, DT);
927 LI->setOrdering(Acquire);
930 Changed |= taintStoreAddress(SI, Inst);
934 // No upcoming branch
936 TaintRelaxedLoads(LI);
939 // For immediately coming branch, directly add a fake branch.
940 if (FirstInst->getParent() == LI->getParent() ||
941 DT->dominates(LI, FirstInst)) {
942 TaintRelaxedLoads(LI);
946 findMostRecentDependenceUsage(LI, FirstInst, &ChainedBB, DT);
948 TaintRelaxedLoads(Inst);
950 LI->setOrdering(Acquire);
960 /**** Implementations of public methods for dependence tainting ****/
961 Value* GetUntaintedAddress(Value* CurrentAddress) {
962 auto* OrAddress = getOrAddress(CurrentAddress);
963 if (OrAddress == nullptr) {
964 // Is it tainted by a select instruction?
965 auto* Inst = dyn_cast<Instruction>(CurrentAddress);
966 if (nullptr != Inst && Inst->getOpcode() == Instruction::Select) {
967 // A selection instruction.
968 if (Inst->getOperand(1) == Inst->getOperand(2)) {
969 return Inst->getOperand(1);
973 return CurrentAddress;
975 Value* ActualAddress = nullptr;
977 auto* CastToInt = dyn_cast<Instruction>(OrAddress->getOperand(1));
978 if (CastToInt && CastToInt->getOpcode() == Instruction::PtrToInt) {
979 return CastToInt->getOperand(0);
981 // This should be a IntToPtr constant expression.
982 ConstantExpr* PtrToIntExpr =
983 dyn_cast<ConstantExpr>(OrAddress->getOperand(1));
984 if (PtrToIntExpr && PtrToIntExpr->getOpcode() == Instruction::PtrToInt) {
985 return PtrToIntExpr->getOperand(0);
989 // Looks like it's not been dependence-tainted. Returns itself.
990 return CurrentAddress;
993 MemoryLocation GetUntaintedMemoryLocation(StoreInst* SI) {
995 SI->getAAMetadata(AATags);
996 const auto& DL = SI->getModule()->getDataLayout();
997 const auto* OriginalAddr = GetUntaintedAddress(SI->getPointerOperand());
998 DEBUG(if (OriginalAddr != SI->getPointerOperand()) {
999 dbgs() << "[GetUntaintedMemoryLocation]\n"
1000 << "Storing address: " << *SI->getPointerOperand()
1001 << "\nUntainted address: " << *OriginalAddr << "\n";
1003 return MemoryLocation(OriginalAddr,
1004 DL.getTypeStoreSize(SI->getValueOperand()->getType()),
1008 bool TaintDependenceToStore(StoreInst* SI, Value* DepVal) {
1009 if (dependenceSetInclusion(SI, DepVal)) {
1013 bool tainted = taintStoreAddress(SI, DepVal);
1018 bool TaintDependenceToStoreAddress(StoreInst* SI, Value* DepVal) {
1019 if (dependenceSetInclusion(SI->getPointerOperand(), DepVal)) {
1023 bool tainted = taintStoreAddress(SI, DepVal);
1028 bool CompressTaintedStore(BasicBlock* BB) {
1029 // This function looks for windows of adajcent stores in 'BB' that satisfy the
1030 // following condition (and then do optimization):
1031 // *Addr(d1) = v1, d1 is a condition and is the only dependence the store's
1032 // address depends on && Dep(v1) includes Dep(d1);
1033 // *Addr(d2) = v2, d2 is a condition and is the only dependnece the store's
1034 // address depends on && Dep(v2) includes Dep(d2) &&
1035 // Dep(d2) includes Dep(d1);
1037 // *Addr(dN) = vN, dN is a condition and is the only dependence the store's
1038 // address depends on && Dep(dN) includes Dep(d"N-1").
1040 // As a result, Dep(dN) includes [Dep(d1) V ... V Dep(d"N-1")], so we can
1041 // safely transform the above to the following. In between these stores, we
1042 // can omit untainted stores to the same address 'Addr' since they internally
1043 // have dependence on the previous stores on the same address.
1048 for (auto BI = BB->begin(), BE = BB->end(); BI != BE; BI++) {
1049 // Look for the first store in such a window of adajacent stores.
1050 auto* FirstSI = dyn_cast<StoreInst>(&*BI);
1055 // The first store in the window must be tainted.
1056 auto* UntaintedAddress = GetUntaintedAddress(FirstSI->getPointerOperand());
1057 if (UntaintedAddress == FirstSI->getPointerOperand()) {
1061 // The first store's address must directly depend on and only depend on a
1063 auto* FirstSIDepCond = getConditionDependence(FirstSI->getPointerOperand());
1064 if (nullptr == FirstSIDepCond) {
1068 // Dep(first store's storing value) includes Dep(tainted dependence).
1069 if (!dependenceSetInclusion(FirstSI->getValueOperand(), FirstSIDepCond)) {
1073 // Look for subsequent stores to the same address that satisfy the condition
1074 // of "compressing the dependence".
1075 SmallVector<StoreInst*, 8> AdajacentStores;
1076 AdajacentStores.push_back(FirstSI);
1077 auto BII = BasicBlock::iterator(FirstSI);
1078 for (BII++; BII != BE; BII++) {
1079 auto* CurrSI = dyn_cast<StoreInst>(&*BII);
1081 if (BII->mayHaveSideEffects()) {
1082 // Be conservative. Instructions with side effects are similar to
1089 auto* OrigAddress = GetUntaintedAddress(CurrSI->getPointerOperand());
1090 auto* CurrSIDepCond = getConditionDependence(CurrSI->getPointerOperand());
1091 // All other stores must satisfy either:
1092 // A. 'CurrSI' is an untainted store to the same address, or
1093 // B. the combination of the following 5 subconditions:
1095 // 2. Untainted address is the same as the group's address;
1096 // 3. The address is tainted with a sole value which is a condition;
1097 // 4. The storing value depends on the condition in 3.
1098 // 5. The condition in 3 depends on the previous stores dependence
1101 // Condition A. Should ignore this store directly.
1102 if (OrigAddress == CurrSI->getPointerOperand() &&
1103 OrigAddress == UntaintedAddress) {
1106 // Check condition B.
1107 Value* Cond = nullptr;
1108 if (OrigAddress == CurrSI->getPointerOperand() ||
1109 OrigAddress != UntaintedAddress || CurrSIDepCond == nullptr ||
1110 !dependenceSetInclusion(CurrSI->getValueOperand(), CurrSIDepCond)) {
1111 // Check condition 1, 2, 3 & 4.
1115 // Check condition 5.
1116 StoreInst* PrevSI = AdajacentStores[AdajacentStores.size() - 1];
1117 auto* PrevSIDepCond = getConditionDependence(PrevSI->getPointerOperand());
1118 assert(PrevSIDepCond &&
1119 "Store in the group must already depend on a condtion");
1120 if (!dependenceSetInclusion(CurrSIDepCond, PrevSIDepCond)) {
1124 AdajacentStores.push_back(CurrSI);
1127 if (AdajacentStores.size() == 1) {
1128 // The outer loop should keep looking from the next store.
1132 // Now we have such a group of tainted stores to the same address.
1133 DEBUG(dbgs() << "[CompressTaintedStore]\n");
1134 DEBUG(dbgs() << "Original BB\n");
1135 DEBUG(dbgs() << *BB << '\n');
1136 auto* LastSI = AdajacentStores[AdajacentStores.size() - 1];
1137 for (unsigned i = 0; i < AdajacentStores.size() - 1; ++i) {
1138 auto* SI = AdajacentStores[i];
1140 // Use the original address for stores before the last one.
1141 SI->setOperand(1, UntaintedAddress);
1143 DEBUG(dbgs() << "Store address has been reversed: " << *SI << '\n';);
1145 // XXX-comment: Try to make the last store use fewer registers.
1146 // If LastSI's storing value is a select based on the condition with which
1147 // its address is tainted, transform the tainted address to a select
1148 // instruction, as follows:
1149 // r1 = Select Cond ? A : B
1154 // r1 = Select Cond ? A : B
1155 // r2 = Select Cond ? Addr : Addr
1157 // The idea is that both Select instructions depend on the same condition,
1158 // so hopefully the backend can generate two cmov instructions for them (and
1159 // this saves the number of registers needed).
1160 auto* LastSIDep = getConditionDependence(LastSI->getPointerOperand());
1161 auto* LastSIValue = dyn_cast<Instruction>(LastSI->getValueOperand());
1162 if (LastSIValue && LastSIValue->getOpcode() == Instruction::Select &&
1163 LastSIValue->getOperand(0) == LastSIDep) {
1164 // XXX-comment: Maybe it's better for us to just leave it as an and/or
1165 // dependence pattern.
1167 IRBuilder<true, NoFolder> Builder(LastSI);
1169 Builder.CreateSelect(LastSIDep, UntaintedAddress, UntaintedAddress);
1170 LastSI->setOperand(1, Address);
1171 DEBUG(dbgs() << "The last store becomes :" << *LastSI << "\n\n";);
1179 bool PassDependenceToStore(Value* OldAddress, StoreInst* NewStore) {
1180 Value* OldDep = getDependence(OldAddress);
1181 // Return false when there's no dependence to pass from the OldAddress.
1186 // No need to pass the dependence to NewStore's address if it already depends
1187 // on whatever 'OldAddress' depends on.
1188 if (StoreAddressDependOnValue(NewStore, OldDep)) {
1191 return taintStoreAddress(NewStore, OldAddress);
1194 SmallSet<Value*, 8> FindDependence(Value* Val) {
1195 SmallSet<Value*, 8> DepSet;
1196 recursivelyFindDependence(&DepSet, Val, true /*Only insert leaf nodes*/);
1200 bool StoreAddressDependOnValue(StoreInst* SI, Value* DepVal) {
1201 return dependenceSetInclusion(SI->getPointerOperand(), DepVal);
1204 bool StoreDependOnValue(StoreInst* SI, Value* Dep) {
1205 return dependenceSetInclusion(SI, Dep);
1212 bool CodeGenPrepare::runOnFunction(Function &F) {
1213 bool EverMadeChange = false;
1215 if (skipOptnoneFunction(F))
1218 DL = &F.getParent()->getDataLayout();
1220 // Clear per function information.
1221 InsertedInsts.clear();
1222 PromotedInsts.clear();
1226 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
1227 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1228 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1229 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1230 OptSize = F.optForSize();
1232 /// This optimization identifies DIV instructions that can be
1233 /// profitably bypassed and carried out with a shorter, faster divide.
1234 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
1235 const DenseMap<unsigned int, unsigned int> &BypassWidths =
1236 TLI->getBypassSlowDivWidths();
1237 BasicBlock* BB = &*F.begin();
1238 while (BB != nullptr) {
1239 // bypassSlowDivision may create new BBs, but we don't want to reapply the
1240 // optimization to those blocks.
1241 BasicBlock* Next = BB->getNextNode();
1242 EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
1247 // Eliminate blocks that contain only PHI nodes and an
1248 // unconditional branch.
1249 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
1251 // llvm.dbg.value is far away from the value then iSel may not be able
1252 // handle it properly. iSel will drop llvm.dbg.value if it can not
1253 // find a node corresponding to the value.
1254 EverMadeChange |= placeDbgValues(F);
1256 // If there is a mask, compare against zero, and branch that can be combined
1257 // into a single target instruction, push the mask and compare into branch
1258 // users. Do this before OptimizeBlock -> OptimizeInst ->
1259 // OptimizeCmpExpression, which perturbs the pattern being searched for.
1260 if (!DisableBranchOpts) {
1261 EverMadeChange |= sinkAndCmp(F);
1262 EverMadeChange |= splitBranchCondition(F);
1265 bool MadeChange = true;
1266 while (MadeChange) {
1268 for (Function::iterator I = F.begin(); I != F.end(); ) {
1269 BasicBlock *BB = &*I++;
1270 bool ModifiedDTOnIteration = false;
1271 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
1273 // Restart BB iteration if the dominator tree of the Function was changed
1274 if (ModifiedDTOnIteration)
1277 EverMadeChange |= MadeChange;
1282 if (!DisableBranchOpts) {
1284 SmallPtrSet<BasicBlock*, 8> WorkList;
1285 for (BasicBlock &BB : F) {
1286 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
1287 MadeChange |= ConstantFoldTerminator(&BB, true);
1288 if (!MadeChange) continue;
1290 for (SmallVectorImpl<BasicBlock*>::iterator
1291 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
1292 if (pred_begin(*II) == pred_end(*II))
1293 WorkList.insert(*II);
1296 // Delete the dead blocks and any of their dead successors.
1297 MadeChange |= !WorkList.empty();
1298 while (!WorkList.empty()) {
1299 BasicBlock *BB = *WorkList.begin();
1301 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
1303 DeleteDeadBlock(BB);
1305 for (SmallVectorImpl<BasicBlock*>::iterator
1306 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
1307 if (pred_begin(*II) == pred_end(*II))
1308 WorkList.insert(*II);
1311 // Merge pairs of basic blocks with unconditional branches, connected by
1313 if (EverMadeChange || MadeChange)
1314 MadeChange |= eliminateFallThrough(F);
1316 EverMadeChange |= MadeChange;
1319 if (!DisableGCOpts) {
1320 SmallVector<Instruction *, 2> Statepoints;
1321 for (BasicBlock &BB : F)
1322 for (Instruction &I : BB)
1323 if (isStatepoint(I))
1324 Statepoints.push_back(&I);
1325 for (auto &I : Statepoints)
1326 EverMadeChange |= simplifyOffsetableRelocate(*I);
1329 // XXX-comment: Delay dealing with relaxed loads in this function to avoid
1330 // further changes done by other passes (e.g., SimplifyCFG).
1331 // Collect all the relaxed loads.
1332 SmallVector<LoadInst*, 1> MonotonicLoadInsts;
1333 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
1334 if (I->isAtomic()) {
1335 switch (I->getOpcode()) {
1336 case Instruction::Load: {
1337 auto* LI = dyn_cast<LoadInst>(&*I);
1338 if (LI->getOrdering() == Monotonic) {
1339 MonotonicLoadInsts.push_back(LI);
1350 AddFakeConditionalBranchAfterMonotonicLoads(MonotonicLoadInsts, DT);
1352 return EverMadeChange;
1355 /// Merge basic blocks which are connected by a single edge, where one of the
1356 /// basic blocks has a single successor pointing to the other basic block,
1357 /// which has a single predecessor.
1358 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
1359 bool Changed = false;
1360 // Scan all of the blocks in the function, except for the entry block.
1361 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
1362 BasicBlock *BB = &*I++;
1363 // If the destination block has a single pred, then this is a trivial
1364 // edge, just collapse it.
1365 BasicBlock *SinglePred = BB->getSinglePredecessor();
1367 // Don't merge if BB's address is taken.
1368 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
1370 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
1371 if (Term && !Term->isConditional()) {
1373 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
1374 // Remember if SinglePred was the entry block of the function.
1375 // If so, we will need to move BB back to the entry position.
1376 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
1377 MergeBasicBlockIntoOnlyPred(BB, nullptr);
1379 if (isEntry && BB != &BB->getParent()->getEntryBlock())
1380 BB->moveBefore(&BB->getParent()->getEntryBlock());
1382 // We have erased a block. Update the iterator.
1383 I = BB->getIterator();
1389 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
1390 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
1391 /// edges in ways that are non-optimal for isel. Start by eliminating these
1392 /// blocks so we can split them the way we want them.
1393 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
1394 bool MadeChange = false;
1395 // Note that this intentionally skips the entry block.
1396 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
1397 BasicBlock *BB = &*I++;
1398 // If this block doesn't end with an uncond branch, ignore it.
1399 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
1400 if (!BI || !BI->isUnconditional())
1403 // If the instruction before the branch (skipping debug info) isn't a phi
1404 // node, then other stuff is happening here.
1405 BasicBlock::iterator BBI = BI->getIterator();
1406 if (BBI != BB->begin()) {
1408 while (isa<DbgInfoIntrinsic>(BBI)) {
1409 if (BBI == BB->begin())
1413 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
1417 // Do not break infinite loops.
1418 BasicBlock *DestBB = BI->getSuccessor(0);
1422 if (!canMergeBlocks(BB, DestBB))
1425 eliminateMostlyEmptyBlock(BB);
1431 /// Return true if we can merge BB into DestBB if there is a single
1432 /// unconditional branch between them, and BB contains no other non-phi
1434 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
1435 const BasicBlock *DestBB) const {
1436 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
1437 // the successor. If there are more complex condition (e.g. preheaders),
1438 // don't mess around with them.
1439 BasicBlock::const_iterator BBI = BB->begin();
1440 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
1441 for (const User *U : PN->users()) {
1442 const Instruction *UI = cast<Instruction>(U);
1443 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
1445 // IfUser is inside DestBB block and it is a PHINode then check
1446 // incoming value. If incoming value is not from BB then this is
1447 // a complex condition (e.g. preheaders) we want to avoid here.
1448 if (UI->getParent() == DestBB) {
1449 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
1450 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
1451 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
1452 if (Insn && Insn->getParent() == BB &&
1453 Insn->getParent() != UPN->getIncomingBlock(I))
1460 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
1461 // and DestBB may have conflicting incoming values for the block. If so, we
1462 // can't merge the block.
1463 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
1464 if (!DestBBPN) return true; // no conflict.
1466 // Collect the preds of BB.
1467 SmallPtrSet<const BasicBlock*, 16> BBPreds;
1468 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
1469 // It is faster to get preds from a PHI than with pred_iterator.
1470 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1471 BBPreds.insert(BBPN->getIncomingBlock(i));
1473 BBPreds.insert(pred_begin(BB), pred_end(BB));
1476 // Walk the preds of DestBB.
1477 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
1478 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
1479 if (BBPreds.count(Pred)) { // Common predecessor?
1480 BBI = DestBB->begin();
1481 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
1482 const Value *V1 = PN->getIncomingValueForBlock(Pred);
1483 const Value *V2 = PN->getIncomingValueForBlock(BB);
1485 // If V2 is a phi node in BB, look up what the mapped value will be.
1486 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
1487 if (V2PN->getParent() == BB)
1488 V2 = V2PN->getIncomingValueForBlock(Pred);
1490 // If there is a conflict, bail out.
1491 if (V1 != V2) return false;
1500 /// Eliminate a basic block that has only phi's and an unconditional branch in
1502 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
1503 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1504 BasicBlock *DestBB = BI->getSuccessor(0);
1506 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
1508 // If the destination block has a single pred, then this is a trivial edge,
1509 // just collapse it.
1510 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
1511 if (SinglePred != DestBB) {
1512 // Remember if SinglePred was the entry block of the function. If so, we
1513 // will need to move BB back to the entry position.
1514 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
1515 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
1517 if (isEntry && BB != &BB->getParent()->getEntryBlock())
1518 BB->moveBefore(&BB->getParent()->getEntryBlock());
1520 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1525 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
1526 // to handle the new incoming edges it is about to have.
1528 for (BasicBlock::iterator BBI = DestBB->begin();
1529 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1530 // Remove the incoming value for BB, and remember it.
1531 Value *InVal = PN->removeIncomingValue(BB, false);
1533 // Two options: either the InVal is a phi node defined in BB or it is some
1534 // value that dominates BB.
1535 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
1536 if (InValPhi && InValPhi->getParent() == BB) {
1537 // Add all of the input values of the input PHI as inputs of this phi.
1538 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
1539 PN->addIncoming(InValPhi->getIncomingValue(i),
1540 InValPhi->getIncomingBlock(i));
1542 // Otherwise, add one instance of the dominating value for each edge that
1543 // we will be adding.
1544 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
1545 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1546 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
1548 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
1549 PN->addIncoming(InVal, *PI);
1554 // The PHIs are now updated, change everything that refers to BB to use
1555 // DestBB and remove BB.
1556 BB->replaceAllUsesWith(DestBB);
1557 BB->eraseFromParent();
1560 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1563 // Computes a map of base pointer relocation instructions to corresponding
1564 // derived pointer relocation instructions given a vector of all relocate calls
1565 static void computeBaseDerivedRelocateMap(
1566 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
1567 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
1569 // Collect information in two maps: one primarily for locating the base object
1570 // while filling the second map; the second map is the final structure holding
1571 // a mapping between Base and corresponding Derived relocate calls
1572 DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
1573 for (auto *ThisRelocate : AllRelocateCalls) {
1574 auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
1575 ThisRelocate->getDerivedPtrIndex());
1576 RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
1578 for (auto &Item : RelocateIdxMap) {
1579 std::pair<unsigned, unsigned> Key = Item.first;
1580 if (Key.first == Key.second)
1581 // Base relocation: nothing to insert
1584 GCRelocateInst *I = Item.second;
1585 auto BaseKey = std::make_pair(Key.first, Key.first);
1587 // We're iterating over RelocateIdxMap so we cannot modify it.
1588 auto MaybeBase = RelocateIdxMap.find(BaseKey);
1589 if (MaybeBase == RelocateIdxMap.end())
1590 // TODO: We might want to insert a new base object relocate and gep off
1591 // that, if there are enough derived object relocates.
1594 RelocateInstMap[MaybeBase->second].push_back(I);
1598 // Accepts a GEP and extracts the operands into a vector provided they're all
1599 // small integer constants
1600 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
1601 SmallVectorImpl<Value *> &OffsetV) {
1602 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
1603 // Only accept small constant integer operands
1604 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
1605 if (!Op || Op->getZExtValue() > 20)
1609 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
1610 OffsetV.push_back(GEP->getOperand(i));
1614 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
1615 // replace, computes a replacement, and affects it.
1617 simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
1618 const SmallVectorImpl<GCRelocateInst *> &Targets) {
1619 bool MadeChange = false;
1620 for (GCRelocateInst *ToReplace : Targets) {
1621 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
1622 "Not relocating a derived object of the original base object");
1623 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
1624 // A duplicate relocate call. TODO: coalesce duplicates.
1628 if (RelocatedBase->getParent() != ToReplace->getParent()) {
1629 // Base and derived relocates are in different basic blocks.
1630 // In this case transform is only valid when base dominates derived
1631 // relocate. However it would be too expensive to check dominance
1632 // for each such relocate, so we skip the whole transformation.
1636 Value *Base = ToReplace->getBasePtr();
1637 auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
1638 if (!Derived || Derived->getPointerOperand() != Base)
1641 SmallVector<Value *, 2> OffsetV;
1642 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
1645 // Create a Builder and replace the target callsite with a gep
1646 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
1648 // Insert after RelocatedBase
1649 IRBuilder<> Builder(RelocatedBase->getNextNode());
1650 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1652 // If gc_relocate does not match the actual type, cast it to the right type.
1653 // In theory, there must be a bitcast after gc_relocate if the type does not
1654 // match, and we should reuse it to get the derived pointer. But it could be
1658 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1663 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1667 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
1668 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
1670 // In this case, we can not find the bitcast any more. So we insert a new bitcast
1671 // no matter there is already one or not. In this way, we can handle all cases, and
1672 // the extra bitcast should be optimized away in later passes.
1673 Value *ActualRelocatedBase = RelocatedBase;
1674 if (RelocatedBase->getType() != Base->getType()) {
1675 ActualRelocatedBase =
1676 Builder.CreateBitCast(RelocatedBase, Base->getType());
1678 Value *Replacement = Builder.CreateGEP(
1679 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
1680 Replacement->takeName(ToReplace);
1681 // If the newly generated derived pointer's type does not match the original derived
1682 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
1683 Value *ActualReplacement = Replacement;
1684 if (Replacement->getType() != ToReplace->getType()) {
1686 Builder.CreateBitCast(Replacement, ToReplace->getType());
1688 ToReplace->replaceAllUsesWith(ActualReplacement);
1689 ToReplace->eraseFromParent();
1699 // %ptr = gep %base + 15
1700 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1701 // %base' = relocate(%tok, i32 4, i32 4)
1702 // %ptr' = relocate(%tok, i32 4, i32 5)
1703 // %val = load %ptr'
1708 // %ptr = gep %base + 15
1709 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1710 // %base' = gc.relocate(%tok, i32 4, i32 4)
1711 // %ptr' = gep %base' + 15
1712 // %val = load %ptr'
1713 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
1714 bool MadeChange = false;
1715 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1717 for (auto *U : I.users())
1718 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1719 // Collect all the relocate calls associated with a statepoint
1720 AllRelocateCalls.push_back(Relocate);
1722 // We need atleast one base pointer relocation + one derived pointer
1723 // relocation to mangle
1724 if (AllRelocateCalls.size() < 2)
1727 // RelocateInstMap is a mapping from the base relocate instruction to the
1728 // corresponding derived relocate instructions
1729 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
1730 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1731 if (RelocateInstMap.empty())
1734 for (auto &Item : RelocateInstMap)
1735 // Item.first is the RelocatedBase to offset against
1736 // Item.second is the vector of Targets to replace
1737 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1741 /// SinkCast - Sink the specified cast instruction into its user blocks
1742 static bool SinkCast(CastInst *CI) {
1743 BasicBlock *DefBB = CI->getParent();
1745 /// InsertedCasts - Only insert a cast in each block once.
1746 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1748 bool MadeChange = false;
1749 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1751 Use &TheUse = UI.getUse();
1752 Instruction *User = cast<Instruction>(*UI);
1754 // Figure out which BB this cast is used in. For PHI's this is the
1755 // appropriate predecessor block.
1756 BasicBlock *UserBB = User->getParent();
1757 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1758 UserBB = PN->getIncomingBlock(TheUse);
1761 // Preincrement use iterator so we don't invalidate it.
1764 // If the block selected to receive the cast is an EH pad that does not
1765 // allow non-PHI instructions before the terminator, we can't sink the
1767 if (UserBB->getTerminator()->isEHPad())
1770 // If this user is in the same block as the cast, don't change the cast.
1771 if (UserBB == DefBB) continue;
1773 // If we have already inserted a cast into this block, use it.
1774 CastInst *&InsertedCast = InsertedCasts[UserBB];
1776 if (!InsertedCast) {
1777 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1778 assert(InsertPt != UserBB->end());
1779 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1780 CI->getType(), "", &*InsertPt);
1783 // Replace a use of the cast with a use of the new cast.
1784 TheUse = InsertedCast;
1789 // If we removed all uses, nuke the cast.
1790 if (CI->use_empty()) {
1791 CI->eraseFromParent();
1798 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1799 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1800 /// reduce the number of virtual registers that must be created and coalesced.
1802 /// Return true if any changes are made.
1804 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1805 const DataLayout &DL) {
1806 // If this is a noop copy,
1807 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1808 EVT DstVT = TLI.getValueType(DL, CI->getType());
1810 // This is an fp<->int conversion?
1811 if (SrcVT.isInteger() != DstVT.isInteger())
1814 // If this is an extension, it will be a zero or sign extension, which
1816 if (SrcVT.bitsLT(DstVT)) return false;
1818 // If these values will be promoted, find out what they will be promoted
1819 // to. This helps us consider truncates on PPC as noop copies when they
1821 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1822 TargetLowering::TypePromoteInteger)
1823 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1824 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1825 TargetLowering::TypePromoteInteger)
1826 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1828 // If, after promotion, these are the same types, this is a noop copy.
1832 return SinkCast(CI);
1835 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
1838 /// Return true if any changes were made.
1839 static bool CombineUAddWithOverflow(CmpInst *CI) {
1843 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
1846 Type *Ty = AddI->getType();
1847 if (!isa<IntegerType>(Ty))
1850 // We don't want to move around uses of condition values this late, so we we
1851 // check if it is legal to create the call to the intrinsic in the basic
1852 // block containing the icmp:
1854 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
1858 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
1860 if (AddI->hasOneUse())
1861 assert(*AddI->user_begin() == CI && "expected!");
1864 Module *M = CI->getModule();
1865 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1867 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
1869 auto *UAddWithOverflow =
1870 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
1871 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
1873 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
1875 CI->replaceAllUsesWith(Overflow);
1876 AddI->replaceAllUsesWith(UAdd);
1877 CI->eraseFromParent();
1878 AddI->eraseFromParent();
1882 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1883 /// registers that must be created and coalesced. This is a clear win except on
1884 /// targets with multiple condition code registers (PowerPC), where it might
1885 /// lose; some adjustment may be wanted there.
1887 /// Return true if any changes are made.
1888 static bool SinkCmpExpression(CmpInst *CI) {
1889 BasicBlock *DefBB = CI->getParent();
1891 /// Only insert a cmp in each block once.
1892 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1894 bool MadeChange = false;
1895 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1897 Use &TheUse = UI.getUse();
1898 Instruction *User = cast<Instruction>(*UI);
1900 // Preincrement use iterator so we don't invalidate it.
1903 // Don't bother for PHI nodes.
1904 if (isa<PHINode>(User))
1907 // Figure out which BB this cmp is used in.
1908 BasicBlock *UserBB = User->getParent();
1910 // If this user is in the same block as the cmp, don't change the cmp.
1911 if (UserBB == DefBB) continue;
1913 // If we have already inserted a cmp into this block, use it.
1914 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1917 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1918 assert(InsertPt != UserBB->end());
1920 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1921 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
1924 // Replace a use of the cmp with a use of the new cmp.
1925 TheUse = InsertedCmp;
1930 // If we removed all uses, nuke the cmp.
1931 if (CI->use_empty()) {
1932 CI->eraseFromParent();
1939 static bool OptimizeCmpExpression(CmpInst *CI) {
1940 if (SinkCmpExpression(CI))
1943 if (CombineUAddWithOverflow(CI))
1949 /// Check if the candidates could be combined with a shift instruction, which
1951 /// 1. Truncate instruction
1952 /// 2. And instruction and the imm is a mask of the low bits:
1953 /// imm & (imm+1) == 0
1954 static bool isExtractBitsCandidateUse(Instruction *User) {
1955 if (!isa<TruncInst>(User)) {
1956 if (User->getOpcode() != Instruction::And ||
1957 !isa<ConstantInt>(User->getOperand(1)))
1960 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
1962 if ((Cimm & (Cimm + 1)).getBoolValue())
1968 /// Sink both shift and truncate instruction to the use of truncate's BB.
1970 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
1971 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
1972 const TargetLowering &TLI, const DataLayout &DL) {
1973 BasicBlock *UserBB = User->getParent();
1974 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
1975 TruncInst *TruncI = dyn_cast<TruncInst>(User);
1976 bool MadeChange = false;
1978 for (Value::user_iterator TruncUI = TruncI->user_begin(),
1979 TruncE = TruncI->user_end();
1980 TruncUI != TruncE;) {
1982 Use &TruncTheUse = TruncUI.getUse();
1983 Instruction *TruncUser = cast<Instruction>(*TruncUI);
1984 // Preincrement use iterator so we don't invalidate it.
1988 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
1992 // If the use is actually a legal node, there will not be an
1993 // implicit truncate.
1994 // FIXME: always querying the result type is just an
1995 // approximation; some nodes' legality is determined by the
1996 // operand or other means. There's no good way to find out though.
1997 if (TLI.isOperationLegalOrCustom(
1998 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
2001 // Don't bother for PHI nodes.
2002 if (isa<PHINode>(TruncUser))
2005 BasicBlock *TruncUserBB = TruncUser->getParent();
2007 if (UserBB == TruncUserBB)
2010 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
2011 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
2013 if (!InsertedShift && !InsertedTrunc) {
2014 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
2015 assert(InsertPt != TruncUserBB->end());
2017 if (ShiftI->getOpcode() == Instruction::AShr)
2018 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
2021 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
2025 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
2027 assert(TruncInsertPt != TruncUserBB->end());
2029 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
2030 TruncI->getType(), "", &*TruncInsertPt);
2034 TruncTheUse = InsertedTrunc;
2040 /// Sink the shift *right* instruction into user blocks if the uses could
2041 /// potentially be combined with this shift instruction and generate BitExtract
2042 /// instruction. It will only be applied if the architecture supports BitExtract
2043 /// instruction. Here is an example:
2045 /// %x.extract.shift = lshr i64 %arg1, 32
2047 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
2051 /// %x.extract.shift.1 = lshr i64 %arg1, 32
2052 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
2054 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
2056 /// Return true if any changes are made.
2057 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
2058 const TargetLowering &TLI,
2059 const DataLayout &DL) {
2060 BasicBlock *DefBB = ShiftI->getParent();
2062 /// Only insert instructions in each block once.
2063 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
2065 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
2067 bool MadeChange = false;
2068 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
2070 Use &TheUse = UI.getUse();
2071 Instruction *User = cast<Instruction>(*UI);
2072 // Preincrement use iterator so we don't invalidate it.
2075 // Don't bother for PHI nodes.
2076 if (isa<PHINode>(User))
2079 if (!isExtractBitsCandidateUse(User))
2082 BasicBlock *UserBB = User->getParent();
2084 if (UserBB == DefBB) {
2085 // If the shift and truncate instruction are in the same BB. The use of
2086 // the truncate(TruncUse) may still introduce another truncate if not
2087 // legal. In this case, we would like to sink both shift and truncate
2088 // instruction to the BB of TruncUse.
2091 // i64 shift.result = lshr i64 opnd, imm
2092 // trunc.result = trunc shift.result to i16
2095 // ----> We will have an implicit truncate here if the architecture does
2096 // not have i16 compare.
2097 // cmp i16 trunc.result, opnd2
2099 if (isa<TruncInst>(User) && shiftIsLegal
2100 // If the type of the truncate is legal, no trucate will be
2101 // introduced in other basic blocks.
2103 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
2105 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
2109 // If we have already inserted a shift into this block, use it.
2110 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
2112 if (!InsertedShift) {
2113 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2114 assert(InsertPt != UserBB->end());
2116 if (ShiftI->getOpcode() == Instruction::AShr)
2117 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
2120 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
2126 // Replace a use of the shift with a use of the new shift.
2127 TheUse = InsertedShift;
2130 // If we removed all uses, nuke the shift.
2131 if (ShiftI->use_empty())
2132 ShiftI->eraseFromParent();
2137 // Translate a masked load intrinsic like
2138 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
2139 // <16 x i1> %mask, <16 x i32> %passthru)
2140 // to a chain of basic blocks, with loading element one-by-one if
2141 // the appropriate mask bit is set
2143 // %1 = bitcast i8* %addr to i32*
2144 // %2 = extractelement <16 x i1> %mask, i32 0
2145 // %3 = icmp eq i1 %2, true
2146 // br i1 %3, label %cond.load, label %else
2148 //cond.load: ; preds = %0
2149 // %4 = getelementptr i32* %1, i32 0
2150 // %5 = load i32* %4
2151 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
2154 //else: ; preds = %0, %cond.load
2155 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
2156 // %7 = extractelement <16 x i1> %mask, i32 1
2157 // %8 = icmp eq i1 %7, true
2158 // br i1 %8, label %cond.load1, label %else2
2160 //cond.load1: ; preds = %else
2161 // %9 = getelementptr i32* %1, i32 1
2162 // %10 = load i32* %9
2163 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
2166 //else2: ; preds = %else, %cond.load1
2167 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
2168 // %12 = extractelement <16 x i1> %mask, i32 2
2169 // %13 = icmp eq i1 %12, true
2170 // br i1 %13, label %cond.load4, label %else5
2172 static void ScalarizeMaskedLoad(CallInst *CI) {
2173 Value *Ptr = CI->getArgOperand(0);
2174 Value *Alignment = CI->getArgOperand(1);
2175 Value *Mask = CI->getArgOperand(2);
2176 Value *Src0 = CI->getArgOperand(3);
2178 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2179 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
2180 assert(VecType && "Unexpected return type of masked load intrinsic");
2182 Type *EltTy = CI->getType()->getVectorElementType();
2184 IRBuilder<> Builder(CI->getContext());
2185 Instruction *InsertPt = CI;
2186 BasicBlock *IfBlock = CI->getParent();
2187 BasicBlock *CondBlock = nullptr;
2188 BasicBlock *PrevIfBlock = CI->getParent();
2190 Builder.SetInsertPoint(InsertPt);
2191 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2193 // Short-cut if the mask is all-true.
2194 bool IsAllOnesMask = isa<Constant>(Mask) &&
2195 cast<Constant>(Mask)->isAllOnesValue();
2197 if (IsAllOnesMask) {
2198 Value *NewI = Builder.CreateAlignedLoad(Ptr, AlignVal);
2199 CI->replaceAllUsesWith(NewI);
2200 CI->eraseFromParent();
2204 // Adjust alignment for the scalar instruction.
2205 AlignVal = std::min(AlignVal, VecType->getScalarSizeInBits()/8);
2206 // Bitcast %addr fron i8* to EltTy*
2208 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
2209 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
2210 unsigned VectorWidth = VecType->getNumElements();
2212 Value *UndefVal = UndefValue::get(VecType);
2214 // The result vector
2215 Value *VResult = UndefVal;
2217 if (isa<ConstantVector>(Mask)) {
2218 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2219 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2222 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2223 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
2224 VResult = Builder.CreateInsertElement(VResult, Load,
2225 Builder.getInt32(Idx));
2227 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
2228 CI->replaceAllUsesWith(NewI);
2229 CI->eraseFromParent();
2233 PHINode *Phi = nullptr;
2234 Value *PrevPhi = UndefVal;
2236 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2238 // Fill the "else" block, created in the previous iteration
2240 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
2241 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
2242 // %to_load = icmp eq i1 %mask_1, true
2243 // br i1 %to_load, label %cond.load, label %else
2246 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
2247 Phi->addIncoming(VResult, CondBlock);
2248 Phi->addIncoming(PrevPhi, PrevIfBlock);
2253 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
2254 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2255 ConstantInt::get(Predicate->getType(), 1));
2257 // Create "cond" block
2259 // %EltAddr = getelementptr i32* %1, i32 0
2260 // %Elt = load i32* %EltAddr
2261 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
2263 CondBlock = IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.load");
2264 Builder.SetInsertPoint(InsertPt);
2267 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2268 LoadInst *Load = Builder.CreateAlignedLoad(Gep, AlignVal);
2269 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
2271 // Create "else" block, fill it in the next iteration
2272 BasicBlock *NewIfBlock =
2273 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
2274 Builder.SetInsertPoint(InsertPt);
2275 Instruction *OldBr = IfBlock->getTerminator();
2276 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2277 OldBr->eraseFromParent();
2278 PrevIfBlock = IfBlock;
2279 IfBlock = NewIfBlock;
2282 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
2283 Phi->addIncoming(VResult, CondBlock);
2284 Phi->addIncoming(PrevPhi, PrevIfBlock);
2285 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
2286 CI->replaceAllUsesWith(NewI);
2287 CI->eraseFromParent();
2290 // Translate a masked store intrinsic, like
2291 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
2293 // to a chain of basic blocks, that stores element one-by-one if
2294 // the appropriate mask bit is set
2296 // %1 = bitcast i8* %addr to i32*
2297 // %2 = extractelement <16 x i1> %mask, i32 0
2298 // %3 = icmp eq i1 %2, true
2299 // br i1 %3, label %cond.store, label %else
2301 // cond.store: ; preds = %0
2302 // %4 = extractelement <16 x i32> %val, i32 0
2303 // %5 = getelementptr i32* %1, i32 0
2304 // store i32 %4, i32* %5
2307 // else: ; preds = %0, %cond.store
2308 // %6 = extractelement <16 x i1> %mask, i32 1
2309 // %7 = icmp eq i1 %6, true
2310 // br i1 %7, label %cond.store1, label %else2
2312 // cond.store1: ; preds = %else
2313 // %8 = extractelement <16 x i32> %val, i32 1
2314 // %9 = getelementptr i32* %1, i32 1
2315 // store i32 %8, i32* %9
2318 static void ScalarizeMaskedStore(CallInst *CI) {
2319 Value *Src = CI->getArgOperand(0);
2320 Value *Ptr = CI->getArgOperand(1);
2321 Value *Alignment = CI->getArgOperand(2);
2322 Value *Mask = CI->getArgOperand(3);
2324 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2325 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
2326 assert(VecType && "Unexpected data type in masked store intrinsic");
2328 Type *EltTy = VecType->getElementType();
2330 IRBuilder<> Builder(CI->getContext());
2331 Instruction *InsertPt = CI;
2332 BasicBlock *IfBlock = CI->getParent();
2333 Builder.SetInsertPoint(InsertPt);
2334 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2336 // Short-cut if the mask is all-true.
2337 bool IsAllOnesMask = isa<Constant>(Mask) &&
2338 cast<Constant>(Mask)->isAllOnesValue();
2340 if (IsAllOnesMask) {
2341 Builder.CreateAlignedStore(Src, Ptr, AlignVal);
2342 CI->eraseFromParent();
2346 // Adjust alignment for the scalar instruction.
2347 AlignVal = std::max(AlignVal, VecType->getScalarSizeInBits()/8);
2348 // Bitcast %addr fron i8* to EltTy*
2350 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
2351 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
2352 unsigned VectorWidth = VecType->getNumElements();
2354 if (isa<ConstantVector>(Mask)) {
2355 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2356 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2358 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
2360 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2361 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
2363 CI->eraseFromParent();
2367 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2369 // Fill the "else" block, created in the previous iteration
2371 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
2372 // %to_store = icmp eq i1 %mask_1, true
2373 // br i1 %to_store, label %cond.store, label %else
2375 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
2376 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2377 ConstantInt::get(Predicate->getType(), 1));
2379 // Create "cond" block
2381 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
2382 // %EltAddr = getelementptr i32* %1, i32 0
2383 // %store i32 %OneElt, i32* %EltAddr
2385 BasicBlock *CondBlock =
2386 IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.store");
2387 Builder.SetInsertPoint(InsertPt);
2389 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
2391 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2392 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
2394 // Create "else" block, fill it in the next iteration
2395 BasicBlock *NewIfBlock =
2396 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
2397 Builder.SetInsertPoint(InsertPt);
2398 Instruction *OldBr = IfBlock->getTerminator();
2399 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2400 OldBr->eraseFromParent();
2401 IfBlock = NewIfBlock;
2403 CI->eraseFromParent();
2406 // Translate a masked gather intrinsic like
2407 // <16 x i32 > @llvm.masked.gather.v16i32( <16 x i32*> %Ptrs, i32 4,
2408 // <16 x i1> %Mask, <16 x i32> %Src)
2409 // to a chain of basic blocks, with loading element one-by-one if
2410 // the appropriate mask bit is set
2412 // % Ptrs = getelementptr i32, i32* %base, <16 x i64> %ind
2413 // % Mask0 = extractelement <16 x i1> %Mask, i32 0
2414 // % ToLoad0 = icmp eq i1 % Mask0, true
2415 // br i1 % ToLoad0, label %cond.load, label %else
2418 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
2419 // % Load0 = load i32, i32* % Ptr0, align 4
2420 // % Res0 = insertelement <16 x i32> undef, i32 % Load0, i32 0
2424 // %res.phi.else = phi <16 x i32>[% Res0, %cond.load], [undef, % 0]
2425 // % Mask1 = extractelement <16 x i1> %Mask, i32 1
2426 // % ToLoad1 = icmp eq i1 % Mask1, true
2427 // br i1 % ToLoad1, label %cond.load1, label %else2
2430 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2431 // % Load1 = load i32, i32* % Ptr1, align 4
2432 // % Res1 = insertelement <16 x i32> %res.phi.else, i32 % Load1, i32 1
2435 // % Result = select <16 x i1> %Mask, <16 x i32> %res.phi.select, <16 x i32> %Src
2436 // ret <16 x i32> %Result
2437 static void ScalarizeMaskedGather(CallInst *CI) {
2438 Value *Ptrs = CI->getArgOperand(0);
2439 Value *Alignment = CI->getArgOperand(1);
2440 Value *Mask = CI->getArgOperand(2);
2441 Value *Src0 = CI->getArgOperand(3);
2443 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
2445 assert(VecType && "Unexpected return type of masked load intrinsic");
2447 IRBuilder<> Builder(CI->getContext());
2448 Instruction *InsertPt = CI;
2449 BasicBlock *IfBlock = CI->getParent();
2450 BasicBlock *CondBlock = nullptr;
2451 BasicBlock *PrevIfBlock = CI->getParent();
2452 Builder.SetInsertPoint(InsertPt);
2453 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2455 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2457 Value *UndefVal = UndefValue::get(VecType);
2459 // The result vector
2460 Value *VResult = UndefVal;
2461 unsigned VectorWidth = VecType->getNumElements();
2463 // Shorten the way if the mask is a vector of constants.
2464 bool IsConstMask = isa<ConstantVector>(Mask);
2467 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2468 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2470 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2471 "Ptr" + Twine(Idx));
2472 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
2473 "Load" + Twine(Idx));
2474 VResult = Builder.CreateInsertElement(VResult, Load,
2475 Builder.getInt32(Idx),
2476 "Res" + Twine(Idx));
2478 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
2479 CI->replaceAllUsesWith(NewI);
2480 CI->eraseFromParent();
2484 PHINode *Phi = nullptr;
2485 Value *PrevPhi = UndefVal;
2487 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2489 // Fill the "else" block, created in the previous iteration
2491 // %Mask1 = extractelement <16 x i1> %Mask, i32 1
2492 // %ToLoad1 = icmp eq i1 %Mask1, true
2493 // br i1 %ToLoad1, label %cond.load, label %else
2496 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
2497 Phi->addIncoming(VResult, CondBlock);
2498 Phi->addIncoming(PrevPhi, PrevIfBlock);
2503 Value *Predicate = Builder.CreateExtractElement(Mask,
2504 Builder.getInt32(Idx),
2505 "Mask" + Twine(Idx));
2506 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2507 ConstantInt::get(Predicate->getType(), 1),
2508 "ToLoad" + Twine(Idx));
2510 // Create "cond" block
2512 // %EltAddr = getelementptr i32* %1, i32 0
2513 // %Elt = load i32* %EltAddr
2514 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
2516 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
2517 Builder.SetInsertPoint(InsertPt);
2519 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2520 "Ptr" + Twine(Idx));
2521 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
2522 "Load" + Twine(Idx));
2523 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx),
2524 "Res" + Twine(Idx));
2526 // Create "else" block, fill it in the next iteration
2527 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2528 Builder.SetInsertPoint(InsertPt);
2529 Instruction *OldBr = IfBlock->getTerminator();
2530 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2531 OldBr->eraseFromParent();
2532 PrevIfBlock = IfBlock;
2533 IfBlock = NewIfBlock;
2536 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
2537 Phi->addIncoming(VResult, CondBlock);
2538 Phi->addIncoming(PrevPhi, PrevIfBlock);
2539 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
2540 CI->replaceAllUsesWith(NewI);
2541 CI->eraseFromParent();
2544 // Translate a masked scatter intrinsic, like
2545 // void @llvm.masked.scatter.v16i32(<16 x i32> %Src, <16 x i32*>* %Ptrs, i32 4,
2547 // to a chain of basic blocks, that stores element one-by-one if
2548 // the appropriate mask bit is set.
2550 // % Ptrs = getelementptr i32, i32* %ptr, <16 x i64> %ind
2551 // % Mask0 = extractelement <16 x i1> % Mask, i32 0
2552 // % ToStore0 = icmp eq i1 % Mask0, true
2553 // br i1 %ToStore0, label %cond.store, label %else
2556 // % Elt0 = extractelement <16 x i32> %Src, i32 0
2557 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
2558 // store i32 %Elt0, i32* % Ptr0, align 4
2562 // % Mask1 = extractelement <16 x i1> % Mask, i32 1
2563 // % ToStore1 = icmp eq i1 % Mask1, true
2564 // br i1 % ToStore1, label %cond.store1, label %else2
2567 // % Elt1 = extractelement <16 x i32> %Src, i32 1
2568 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2569 // store i32 % Elt1, i32* % Ptr1, align 4
2572 static void ScalarizeMaskedScatter(CallInst *CI) {
2573 Value *Src = CI->getArgOperand(0);
2574 Value *Ptrs = CI->getArgOperand(1);
2575 Value *Alignment = CI->getArgOperand(2);
2576 Value *Mask = CI->getArgOperand(3);
2578 assert(isa<VectorType>(Src->getType()) &&
2579 "Unexpected data type in masked scatter intrinsic");
2580 assert(isa<VectorType>(Ptrs->getType()) &&
2581 isa<PointerType>(Ptrs->getType()->getVectorElementType()) &&
2582 "Vector of pointers is expected in masked scatter intrinsic");
2584 IRBuilder<> Builder(CI->getContext());
2585 Instruction *InsertPt = CI;
2586 BasicBlock *IfBlock = CI->getParent();
2587 Builder.SetInsertPoint(InsertPt);
2588 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2590 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2591 unsigned VectorWidth = Src->getType()->getVectorNumElements();
2593 // Shorten the way if the mask is a vector of constants.
2594 bool IsConstMask = isa<ConstantVector>(Mask);
2597 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2598 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2600 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
2601 "Elt" + Twine(Idx));
2602 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2603 "Ptr" + Twine(Idx));
2604 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
2606 CI->eraseFromParent();
2609 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2610 // Fill the "else" block, created in the previous iteration
2612 // % Mask1 = extractelement <16 x i1> % Mask, i32 Idx
2613 // % ToStore = icmp eq i1 % Mask1, true
2614 // br i1 % ToStore, label %cond.store, label %else
2616 Value *Predicate = Builder.CreateExtractElement(Mask,
2617 Builder.getInt32(Idx),
2618 "Mask" + Twine(Idx));
2620 Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2621 ConstantInt::get(Predicate->getType(), 1),
2622 "ToStore" + Twine(Idx));
2624 // Create "cond" block
2626 // % Elt1 = extractelement <16 x i32> %Src, i32 1
2627 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2628 // %store i32 % Elt1, i32* % Ptr1
2630 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2631 Builder.SetInsertPoint(InsertPt);
2633 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
2634 "Elt" + Twine(Idx));
2635 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2636 "Ptr" + Twine(Idx));
2637 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
2639 // Create "else" block, fill it in the next iteration
2640 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2641 Builder.SetInsertPoint(InsertPt);
2642 Instruction *OldBr = IfBlock->getTerminator();
2643 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2644 OldBr->eraseFromParent();
2645 IfBlock = NewIfBlock;
2647 CI->eraseFromParent();
2650 /// If counting leading or trailing zeros is an expensive operation and a zero
2651 /// input is defined, add a check for zero to avoid calling the intrinsic.
2653 /// We want to transform:
2654 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
2658 /// %cmpz = icmp eq i64 %A, 0
2659 /// br i1 %cmpz, label %cond.end, label %cond.false
2661 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
2662 /// br label %cond.end
2664 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
2666 /// If the transform is performed, return true and set ModifiedDT to true.
2667 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
2668 const TargetLowering *TLI,
2669 const DataLayout *DL,
2674 // If a zero input is undefined, it doesn't make sense to despeculate that.
2675 if (match(CountZeros->getOperand(1), m_One()))
2678 // If it's cheap to speculate, there's nothing to do.
2679 auto IntrinsicID = CountZeros->getIntrinsicID();
2680 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
2681 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
2684 // Only handle legal scalar cases. Anything else requires too much work.
2685 Type *Ty = CountZeros->getType();
2686 unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
2687 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSize())
2690 // The intrinsic will be sunk behind a compare against zero and branch.
2691 BasicBlock *StartBlock = CountZeros->getParent();
2692 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
2694 // Create another block after the count zero intrinsic. A PHI will be added
2695 // in this block to select the result of the intrinsic or the bit-width
2696 // constant if the input to the intrinsic is zero.
2697 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
2698 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
2700 // Set up a builder to create a compare, conditional branch, and PHI.
2701 IRBuilder<> Builder(CountZeros->getContext());
2702 Builder.SetInsertPoint(StartBlock->getTerminator());
2703 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
2705 // Replace the unconditional branch that was created by the first split with
2706 // a compare against zero and a conditional branch.
2707 Value *Zero = Constant::getNullValue(Ty);
2708 Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
2709 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
2710 StartBlock->getTerminator()->eraseFromParent();
2712 // Create a PHI in the end block to select either the output of the intrinsic
2713 // or the bit width of the operand.
2714 Builder.SetInsertPoint(&EndBlock->front());
2715 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
2716 CountZeros->replaceAllUsesWith(PN);
2717 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
2718 PN->addIncoming(BitWidth, StartBlock);
2719 PN->addIncoming(CountZeros, CallBlock);
2721 // We are explicitly handling the zero case, so we can set the intrinsic's
2722 // undefined zero argument to 'true'. This will also prevent reprocessing the
2723 // intrinsic; we only despeculate when a zero input is defined.
2724 CountZeros->setArgOperand(1, Builder.getTrue());
2729 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool& ModifiedDT) {
2730 BasicBlock *BB = CI->getParent();
2732 // Lower inline assembly if we can.
2733 // If we found an inline asm expession, and if the target knows how to
2734 // lower it to normal LLVM code, do so now.
2735 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
2736 if (TLI->ExpandInlineAsm(CI)) {
2737 // Avoid invalidating the iterator.
2738 CurInstIterator = BB->begin();
2739 // Avoid processing instructions out of order, which could cause
2740 // reuse before a value is defined.
2744 // Sink address computing for memory operands into the block.
2745 if (optimizeInlineAsmInst(CI))
2749 // Align the pointer arguments to this call if the target thinks it's a good
2751 unsigned MinSize, PrefAlign;
2752 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
2753 for (auto &Arg : CI->arg_operands()) {
2754 // We want to align both objects whose address is used directly and
2755 // objects whose address is used in casts and GEPs, though it only makes
2756 // sense for GEPs if the offset is a multiple of the desired alignment and
2757 // if size - offset meets the size threshold.
2758 if (!Arg->getType()->isPointerTy())
2760 APInt Offset(DL->getPointerSizeInBits(
2761 cast<PointerType>(Arg->getType())->getAddressSpace()),
2763 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
2764 uint64_t Offset2 = Offset.getLimitedValue();
2765 if ((Offset2 & (PrefAlign-1)) != 0)
2768 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
2769 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
2770 AI->setAlignment(PrefAlign);
2771 // Global variables can only be aligned if they are defined in this
2772 // object (i.e. they are uniquely initialized in this object), and
2773 // over-aligning global variables that have an explicit section is
2776 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
2777 GV->getAlignment() < PrefAlign &&
2778 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
2780 GV->setAlignment(PrefAlign);
2782 // If this is a memcpy (or similar) then we may be able to improve the
2784 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
2785 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
2786 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
2787 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
2788 if (Align > MI->getAlignment())
2789 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
2793 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
2795 switch (II->getIntrinsicID()) {
2797 case Intrinsic::objectsize: {
2798 // Lower all uses of llvm.objectsize.*
2799 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
2800 Type *ReturnTy = CI->getType();
2801 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
2803 // Substituting this can cause recursive simplifications, which can
2804 // invalidate our iterator. Use a WeakVH to hold onto it in case this
2806 WeakVH IterHandle(&*CurInstIterator);
2808 replaceAndRecursivelySimplify(CI, RetVal,
2811 // If the iterator instruction was recursively deleted, start over at the
2812 // start of the block.
2813 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
2814 CurInstIterator = BB->begin();
2819 case Intrinsic::masked_load: {
2820 // Scalarize unsupported vector masked load
2821 if (!TTI->isLegalMaskedLoad(CI->getType())) {
2822 ScalarizeMaskedLoad(CI);
2828 case Intrinsic::masked_store: {
2829 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType())) {
2830 ScalarizeMaskedStore(CI);
2836 case Intrinsic::masked_gather: {
2837 if (!TTI->isLegalMaskedGather(CI->getType())) {
2838 ScalarizeMaskedGather(CI);
2844 case Intrinsic::masked_scatter: {
2845 if (!TTI->isLegalMaskedScatter(CI->getArgOperand(0)->getType())) {
2846 ScalarizeMaskedScatter(CI);
2852 case Intrinsic::aarch64_stlxr:
2853 case Intrinsic::aarch64_stxr: {
2854 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
2855 if (!ExtVal || !ExtVal->hasOneUse() ||
2856 ExtVal->getParent() == CI->getParent())
2858 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
2859 ExtVal->moveBefore(CI);
2860 // Mark this instruction as "inserted by CGP", so that other
2861 // optimizations don't touch it.
2862 InsertedInsts.insert(ExtVal);
2865 case Intrinsic::invariant_group_barrier:
2866 II->replaceAllUsesWith(II->getArgOperand(0));
2867 II->eraseFromParent();
2870 case Intrinsic::cttz:
2871 case Intrinsic::ctlz:
2872 // If counting zeros is expensive, try to avoid it.
2873 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
2877 // Unknown address space.
2878 // TODO: Target hook to pick which address space the intrinsic cares
2880 unsigned AddrSpace = ~0u;
2881 SmallVector<Value*, 2> PtrOps;
2883 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
2884 while (!PtrOps.empty())
2885 if (optimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
2890 // From here on out we're working with named functions.
2891 if (!CI->getCalledFunction()) return false;
2893 // Lower all default uses of _chk calls. This is very similar
2894 // to what InstCombineCalls does, but here we are only lowering calls
2895 // to fortified library functions (e.g. __memcpy_chk) that have the default
2896 // "don't know" as the objectsize. Anything else should be left alone.
2897 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
2898 if (Value *V = Simplifier.optimizeCall(CI)) {
2899 CI->replaceAllUsesWith(V);
2900 CI->eraseFromParent();
2906 /// Look for opportunities to duplicate return instructions to the predecessor
2907 /// to enable tail call optimizations. The case it is currently looking for is:
2910 /// %tmp0 = tail call i32 @f0()
2911 /// br label %return
2913 /// %tmp1 = tail call i32 @f1()
2914 /// br label %return
2916 /// %tmp2 = tail call i32 @f2()
2917 /// br label %return
2919 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
2927 /// %tmp0 = tail call i32 @f0()
2930 /// %tmp1 = tail call i32 @f1()
2933 /// %tmp2 = tail call i32 @f2()
2936 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
2940 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
2944 PHINode *PN = nullptr;
2945 BitCastInst *BCI = nullptr;
2946 Value *V = RI->getReturnValue();
2948 BCI = dyn_cast<BitCastInst>(V);
2950 V = BCI->getOperand(0);
2952 PN = dyn_cast<PHINode>(V);
2957 if (PN && PN->getParent() != BB)
2960 // It's not safe to eliminate the sign / zero extension of the return value.
2961 // See llvm::isInTailCallPosition().
2962 const Function *F = BB->getParent();
2963 AttributeSet CallerAttrs = F->getAttributes();
2964 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
2965 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
2968 // Make sure there are no instructions between the PHI and return, or that the
2969 // return is the first instruction in the block.
2971 BasicBlock::iterator BI = BB->begin();
2972 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
2974 // Also skip over the bitcast.
2979 BasicBlock::iterator BI = BB->begin();
2980 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
2985 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
2987 SmallVector<CallInst*, 4> TailCalls;
2989 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
2990 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
2991 // Make sure the phi value is indeed produced by the tail call.
2992 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
2993 TLI->mayBeEmittedAsTailCall(CI))
2994 TailCalls.push_back(CI);
2997 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
2998 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
2999 if (!VisitedBBs.insert(*PI).second)
3002 BasicBlock::InstListType &InstList = (*PI)->getInstList();
3003 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
3004 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
3005 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
3009 CallInst *CI = dyn_cast<CallInst>(&*RI);
3010 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
3011 TailCalls.push_back(CI);
3015 bool Changed = false;
3016 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
3017 CallInst *CI = TailCalls[i];
3020 // Conservatively require the attributes of the call to match those of the
3021 // return. Ignore noalias because it doesn't affect the call sequence.
3022 AttributeSet CalleeAttrs = CS.getAttributes();
3023 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
3024 removeAttribute(Attribute::NoAlias) !=
3025 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
3026 removeAttribute(Attribute::NoAlias))
3029 // Make sure the call instruction is followed by an unconditional branch to
3030 // the return block.
3031 BasicBlock *CallBB = CI->getParent();
3032 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
3033 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
3036 // Duplicate the return into CallBB.
3037 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
3038 ModifiedDT = Changed = true;
3042 // If we eliminated all predecessors of the block, delete the block now.
3043 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
3044 BB->eraseFromParent();
3049 //===----------------------------------------------------------------------===//
3050 // Memory Optimization
3051 //===----------------------------------------------------------------------===//
3055 /// This is an extended version of TargetLowering::AddrMode
3056 /// which holds actual Value*'s for register values.
3057 struct ExtAddrMode : public TargetLowering::AddrMode {
3060 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
3061 void print(raw_ostream &OS) const;
3064 bool operator==(const ExtAddrMode& O) const {
3065 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
3066 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
3067 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
3072 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
3078 void ExtAddrMode::print(raw_ostream &OS) const {
3079 bool NeedPlus = false;
3082 OS << (NeedPlus ? " + " : "")
3084 BaseGV->printAsOperand(OS, /*PrintType=*/false);
3089 OS << (NeedPlus ? " + " : "")
3095 OS << (NeedPlus ? " + " : "")
3097 BaseReg->printAsOperand(OS, /*PrintType=*/false);
3101 OS << (NeedPlus ? " + " : "")
3103 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
3109 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3110 void ExtAddrMode::dump() const {
3116 /// \brief This class provides transaction based operation on the IR.
3117 /// Every change made through this class is recorded in the internal state and
3118 /// can be undone (rollback) until commit is called.
3119 class TypePromotionTransaction {
3121 /// \brief This represents the common interface of the individual transaction.
3122 /// Each class implements the logic for doing one specific modification on
3123 /// the IR via the TypePromotionTransaction.
3124 class TypePromotionAction {
3126 /// The Instruction modified.
3130 /// \brief Constructor of the action.
3131 /// The constructor performs the related action on the IR.
3132 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
3134 virtual ~TypePromotionAction() {}
3136 /// \brief Undo the modification done by this action.
3137 /// When this method is called, the IR must be in the same state as it was
3138 /// before this action was applied.
3139 /// \pre Undoing the action works if and only if the IR is in the exact same
3140 /// state as it was directly after this action was applied.
3141 virtual void undo() = 0;
3143 /// \brief Advocate every change made by this action.
3144 /// When the results on the IR of the action are to be kept, it is important
3145 /// to call this function, otherwise hidden information may be kept forever.
3146 virtual void commit() {
3147 // Nothing to be done, this action is not doing anything.
3151 /// \brief Utility to remember the position of an instruction.
3152 class InsertionHandler {
3153 /// Position of an instruction.
3154 /// Either an instruction:
3155 /// - Is the first in a basic block: BB is used.
3156 /// - Has a previous instructon: PrevInst is used.
3158 Instruction *PrevInst;
3161 /// Remember whether or not the instruction had a previous instruction.
3162 bool HasPrevInstruction;
3165 /// \brief Record the position of \p Inst.
3166 InsertionHandler(Instruction *Inst) {
3167 BasicBlock::iterator It = Inst->getIterator();
3168 HasPrevInstruction = (It != (Inst->getParent()->begin()));
3169 if (HasPrevInstruction)
3170 Point.PrevInst = &*--It;
3172 Point.BB = Inst->getParent();
3175 /// \brief Insert \p Inst at the recorded position.
3176 void insert(Instruction *Inst) {
3177 if (HasPrevInstruction) {
3178 if (Inst->getParent())
3179 Inst->removeFromParent();
3180 Inst->insertAfter(Point.PrevInst);
3182 Instruction *Position = &*Point.BB->getFirstInsertionPt();
3183 if (Inst->getParent())
3184 Inst->moveBefore(Position);
3186 Inst->insertBefore(Position);
3191 /// \brief Move an instruction before another.
3192 class InstructionMoveBefore : public TypePromotionAction {
3193 /// Original position of the instruction.
3194 InsertionHandler Position;
3197 /// \brief Move \p Inst before \p Before.
3198 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
3199 : TypePromotionAction(Inst), Position(Inst) {
3200 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
3201 Inst->moveBefore(Before);
3204 /// \brief Move the instruction back to its original position.
3205 void undo() override {
3206 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
3207 Position.insert(Inst);
3211 /// \brief Set the operand of an instruction with a new value.
3212 class OperandSetter : public TypePromotionAction {
3213 /// Original operand of the instruction.
3215 /// Index of the modified instruction.
3219 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
3220 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
3221 : TypePromotionAction(Inst), Idx(Idx) {
3222 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
3223 << "for:" << *Inst << "\n"
3224 << "with:" << *NewVal << "\n");
3225 Origin = Inst->getOperand(Idx);
3226 Inst->setOperand(Idx, NewVal);
3229 /// \brief Restore the original value of the instruction.
3230 void undo() override {
3231 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
3232 << "for: " << *Inst << "\n"
3233 << "with: " << *Origin << "\n");
3234 Inst->setOperand(Idx, Origin);
3238 /// \brief Hide the operands of an instruction.
3239 /// Do as if this instruction was not using any of its operands.
3240 class OperandsHider : public TypePromotionAction {
3241 /// The list of original operands.
3242 SmallVector<Value *, 4> OriginalValues;
3245 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
3246 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
3247 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
3248 unsigned NumOpnds = Inst->getNumOperands();
3249 OriginalValues.reserve(NumOpnds);
3250 for (unsigned It = 0; It < NumOpnds; ++It) {
3251 // Save the current operand.
3252 Value *Val = Inst->getOperand(It);
3253 OriginalValues.push_back(Val);
3255 // We could use OperandSetter here, but that would imply an overhead
3256 // that we are not willing to pay.
3257 Inst->setOperand(It, UndefValue::get(Val->getType()));
3261 /// \brief Restore the original list of uses.
3262 void undo() override {
3263 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
3264 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
3265 Inst->setOperand(It, OriginalValues[It]);
3269 /// \brief Build a truncate instruction.
3270 class TruncBuilder : public TypePromotionAction {
3273 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
3275 /// trunc Opnd to Ty.
3276 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
3277 IRBuilder<> Builder(Opnd);
3278 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
3279 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
3282 /// \brief Get the built value.
3283 Value *getBuiltValue() { return Val; }
3285 /// \brief Remove the built instruction.
3286 void undo() override {
3287 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
3288 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3289 IVal->eraseFromParent();
3293 /// \brief Build a sign extension instruction.
3294 class SExtBuilder : public TypePromotionAction {
3297 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
3299 /// sext Opnd to Ty.
3300 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3301 : TypePromotionAction(InsertPt) {
3302 IRBuilder<> Builder(InsertPt);
3303 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
3304 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
3307 /// \brief Get the built value.
3308 Value *getBuiltValue() { return Val; }
3310 /// \brief Remove the built instruction.
3311 void undo() override {
3312 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
3313 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3314 IVal->eraseFromParent();
3318 /// \brief Build a zero extension instruction.
3319 class ZExtBuilder : public TypePromotionAction {
3322 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
3324 /// zext Opnd to Ty.
3325 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3326 : TypePromotionAction(InsertPt) {
3327 IRBuilder<> Builder(InsertPt);
3328 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
3329 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
3332 /// \brief Get the built value.
3333 Value *getBuiltValue() { return Val; }
3335 /// \brief Remove the built instruction.
3336 void undo() override {
3337 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
3338 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3339 IVal->eraseFromParent();
3343 /// \brief Mutate an instruction to another type.
3344 class TypeMutator : public TypePromotionAction {
3345 /// Record the original type.
3349 /// \brief Mutate the type of \p Inst into \p NewTy.
3350 TypeMutator(Instruction *Inst, Type *NewTy)
3351 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
3352 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
3354 Inst->mutateType(NewTy);
3357 /// \brief Mutate the instruction back to its original type.
3358 void undo() override {
3359 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
3361 Inst->mutateType(OrigTy);
3365 /// \brief Replace the uses of an instruction by another instruction.
3366 class UsesReplacer : public TypePromotionAction {
3367 /// Helper structure to keep track of the replaced uses.
3368 struct InstructionAndIdx {
3369 /// The instruction using the instruction.
3371 /// The index where this instruction is used for Inst.
3373 InstructionAndIdx(Instruction *Inst, unsigned Idx)
3374 : Inst(Inst), Idx(Idx) {}
3377 /// Keep track of the original uses (pair Instruction, Index).
3378 SmallVector<InstructionAndIdx, 4> OriginalUses;
3379 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
3382 /// \brief Replace all the use of \p Inst by \p New.
3383 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
3384 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
3386 // Record the original uses.
3387 for (Use &U : Inst->uses()) {
3388 Instruction *UserI = cast<Instruction>(U.getUser());
3389 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
3391 // Now, we can replace the uses.
3392 Inst->replaceAllUsesWith(New);
3395 /// \brief Reassign the original uses of Inst to Inst.
3396 void undo() override {
3397 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
3398 for (use_iterator UseIt = OriginalUses.begin(),
3399 EndIt = OriginalUses.end();
3400 UseIt != EndIt; ++UseIt) {
3401 UseIt->Inst->setOperand(UseIt->Idx, Inst);
3406 /// \brief Remove an instruction from the IR.
3407 class InstructionRemover : public TypePromotionAction {
3408 /// Original position of the instruction.
3409 InsertionHandler Inserter;
3410 /// Helper structure to hide all the link to the instruction. In other
3411 /// words, this helps to do as if the instruction was removed.
3412 OperandsHider Hider;
3413 /// Keep track of the uses replaced, if any.
3414 UsesReplacer *Replacer;
3417 /// \brief Remove all reference of \p Inst and optinally replace all its
3419 /// \pre If !Inst->use_empty(), then New != nullptr
3420 InstructionRemover(Instruction *Inst, Value *New = nullptr)
3421 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
3424 Replacer = new UsesReplacer(Inst, New);
3425 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
3426 Inst->removeFromParent();
3429 ~InstructionRemover() override { delete Replacer; }
3431 /// \brief Really remove the instruction.
3432 void commit() override { delete Inst; }
3434 /// \brief Resurrect the instruction and reassign it to the proper uses if
3435 /// new value was provided when build this action.
3436 void undo() override {
3437 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
3438 Inserter.insert(Inst);
3446 /// Restoration point.
3447 /// The restoration point is a pointer to an action instead of an iterator
3448 /// because the iterator may be invalidated but not the pointer.
3449 typedef const TypePromotionAction *ConstRestorationPt;
3450 /// Advocate every changes made in that transaction.
3452 /// Undo all the changes made after the given point.
3453 void rollback(ConstRestorationPt Point);
3454 /// Get the current restoration point.
3455 ConstRestorationPt getRestorationPoint() const;
3457 /// \name API for IR modification with state keeping to support rollback.
3459 /// Same as Instruction::setOperand.
3460 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
3461 /// Same as Instruction::eraseFromParent.
3462 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
3463 /// Same as Value::replaceAllUsesWith.
3464 void replaceAllUsesWith(Instruction *Inst, Value *New);
3465 /// Same as Value::mutateType.
3466 void mutateType(Instruction *Inst, Type *NewTy);
3467 /// Same as IRBuilder::createTrunc.
3468 Value *createTrunc(Instruction *Opnd, Type *Ty);
3469 /// Same as IRBuilder::createSExt.
3470 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
3471 /// Same as IRBuilder::createZExt.
3472 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
3473 /// Same as Instruction::moveBefore.
3474 void moveBefore(Instruction *Inst, Instruction *Before);
3478 /// The ordered list of actions made so far.
3479 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
3480 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
3483 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
3486 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
3489 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
3492 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
3495 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
3497 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
3500 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
3501 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
3504 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
3506 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
3507 Value *Val = Ptr->getBuiltValue();
3508 Actions.push_back(std::move(Ptr));
3512 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
3513 Value *Opnd, Type *Ty) {
3514 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
3515 Value *Val = Ptr->getBuiltValue();
3516 Actions.push_back(std::move(Ptr));
3520 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
3521 Value *Opnd, Type *Ty) {
3522 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
3523 Value *Val = Ptr->getBuiltValue();
3524 Actions.push_back(std::move(Ptr));
3528 void TypePromotionTransaction::moveBefore(Instruction *Inst,
3529 Instruction *Before) {
3531 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
3534 TypePromotionTransaction::ConstRestorationPt
3535 TypePromotionTransaction::getRestorationPoint() const {
3536 return !Actions.empty() ? Actions.back().get() : nullptr;
3539 void TypePromotionTransaction::commit() {
3540 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
3546 void TypePromotionTransaction::rollback(
3547 TypePromotionTransaction::ConstRestorationPt Point) {
3548 while (!Actions.empty() && Point != Actions.back().get()) {
3549 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
3554 /// \brief A helper class for matching addressing modes.
3556 /// This encapsulates the logic for matching the target-legal addressing modes.
3557 class AddressingModeMatcher {
3558 SmallVectorImpl<Instruction*> &AddrModeInsts;
3559 const TargetMachine &TM;
3560 const TargetLowering &TLI;
3561 const DataLayout &DL;
3563 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
3564 /// the memory instruction that we're computing this address for.
3567 Instruction *MemoryInst;
3569 /// This is the addressing mode that we're building up. This is
3570 /// part of the return value of this addressing mode matching stuff.
3571 ExtAddrMode &AddrMode;
3573 /// The instructions inserted by other CodeGenPrepare optimizations.
3574 const SetOfInstrs &InsertedInsts;
3575 /// A map from the instructions to their type before promotion.
3576 InstrToOrigTy &PromotedInsts;
3577 /// The ongoing transaction where every action should be registered.
3578 TypePromotionTransaction &TPT;
3580 /// This is set to true when we should not do profitability checks.
3581 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
3582 bool IgnoreProfitability;
3584 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
3585 const TargetMachine &TM, Type *AT, unsigned AS,
3586 Instruction *MI, ExtAddrMode &AM,
3587 const SetOfInstrs &InsertedInsts,
3588 InstrToOrigTy &PromotedInsts,
3589 TypePromotionTransaction &TPT)
3590 : AddrModeInsts(AMI), TM(TM),
3591 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
3592 ->getTargetLowering()),
3593 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
3594 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
3595 PromotedInsts(PromotedInsts), TPT(TPT) {
3596 IgnoreProfitability = false;
3600 /// Find the maximal addressing mode that a load/store of V can fold,
3601 /// give an access type of AccessTy. This returns a list of involved
3602 /// instructions in AddrModeInsts.
3603 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
3605 /// \p PromotedInsts maps the instructions to their type before promotion.
3606 /// \p The ongoing transaction where every action should be registered.
3607 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
3608 Instruction *MemoryInst,
3609 SmallVectorImpl<Instruction*> &AddrModeInsts,
3610 const TargetMachine &TM,
3611 const SetOfInstrs &InsertedInsts,
3612 InstrToOrigTy &PromotedInsts,
3613 TypePromotionTransaction &TPT) {
3616 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
3617 MemoryInst, Result, InsertedInsts,
3618 PromotedInsts, TPT).matchAddr(V, 0);
3619 (void)Success; assert(Success && "Couldn't select *anything*?");
3623 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
3624 bool matchAddr(Value *V, unsigned Depth);
3625 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
3626 bool *MovedAway = nullptr);
3627 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
3628 ExtAddrMode &AMBefore,
3629 ExtAddrMode &AMAfter);
3630 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
3631 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
3632 Value *PromotedOperand) const;
3635 /// Try adding ScaleReg*Scale to the current addressing mode.
3636 /// Return true and update AddrMode if this addr mode is legal for the target,
3638 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3640 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3641 // mode. Just process that directly.
3643 return matchAddr(ScaleReg, Depth);
3645 // If the scale is 0, it takes nothing to add this.
3649 // If we already have a scale of this value, we can add to it, otherwise, we
3650 // need an available scale field.
3651 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3654 ExtAddrMode TestAddrMode = AddrMode;
3656 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3657 // [A+B + A*7] -> [B+A*8].
3658 TestAddrMode.Scale += Scale;
3659 TestAddrMode.ScaledReg = ScaleReg;
3661 // If the new address isn't legal, bail out.
3662 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3665 // It was legal, so commit it.
3666 AddrMode = TestAddrMode;
3668 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3669 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3670 // X*Scale + C*Scale to addr mode.
3671 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3672 if (isa<Instruction>(ScaleReg) && // not a constant expr.
3673 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3674 TestAddrMode.ScaledReg = AddLHS;
3675 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3677 // If this addressing mode is legal, commit it and remember that we folded
3678 // this instruction.
3679 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3680 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3681 AddrMode = TestAddrMode;
3686 // Otherwise, not (x+c)*scale, just return what we have.
3690 /// This is a little filter, which returns true if an addressing computation
3691 /// involving I might be folded into a load/store accessing it.
3692 /// This doesn't need to be perfect, but needs to accept at least
3693 /// the set of instructions that MatchOperationAddr can.
3694 static bool MightBeFoldableInst(Instruction *I) {
3695 switch (I->getOpcode()) {
3696 case Instruction::BitCast:
3697 case Instruction::AddrSpaceCast:
3698 // Don't touch identity bitcasts.
3699 if (I->getType() == I->getOperand(0)->getType())
3701 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
3702 case Instruction::PtrToInt:
3703 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3705 case Instruction::IntToPtr:
3706 // We know the input is intptr_t, so this is foldable.
3708 case Instruction::Add:
3710 case Instruction::Mul:
3711 case Instruction::Shl:
3712 // Can only handle X*C and X << C.
3713 return isa<ConstantInt>(I->getOperand(1));
3714 case Instruction::GetElementPtr:
3721 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
3722 /// \note \p Val is assumed to be the product of some type promotion.
3723 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3724 /// to be legal, as the non-promoted value would have had the same state.
3725 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
3726 const DataLayout &DL, Value *Val) {
3727 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3730 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3731 // If the ISDOpcode is undefined, it was undefined before the promotion.
3734 // Otherwise, check if the promoted instruction is legal or not.
3735 return TLI.isOperationLegalOrCustom(
3736 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3739 /// \brief Hepler class to perform type promotion.
3740 class TypePromotionHelper {
3741 /// \brief Utility function to check whether or not a sign or zero extension
3742 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3743 /// either using the operands of \p Inst or promoting \p Inst.
3744 /// The type of the extension is defined by \p IsSExt.
3745 /// In other words, check if:
3746 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3747 /// #1 Promotion applies:
3748 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3749 /// #2 Operand reuses:
3750 /// ext opnd1 to ConsideredExtType.
3751 /// \p PromotedInsts maps the instructions to their type before promotion.
3752 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3753 const InstrToOrigTy &PromotedInsts, bool IsSExt);
3755 /// \brief Utility function to determine if \p OpIdx should be promoted when
3756 /// promoting \p Inst.
3757 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3758 return !(isa<SelectInst>(Inst) && OpIdx == 0);
3761 /// \brief Utility function to promote the operand of \p Ext when this
3762 /// operand is a promotable trunc or sext or zext.
3763 /// \p PromotedInsts maps the instructions to their type before promotion.
3764 /// \p CreatedInstsCost[out] contains the cost of all instructions
3765 /// created to promote the operand of Ext.
3766 /// Newly added extensions are inserted in \p Exts.
3767 /// Newly added truncates are inserted in \p Truncs.
3768 /// Should never be called directly.
3769 /// \return The promoted value which is used instead of Ext.
3770 static Value *promoteOperandForTruncAndAnyExt(
3771 Instruction *Ext, TypePromotionTransaction &TPT,
3772 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3773 SmallVectorImpl<Instruction *> *Exts,
3774 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3776 /// \brief Utility function to promote the operand of \p Ext when this
3777 /// operand is promotable and is not a supported trunc or sext.
3778 /// \p PromotedInsts maps the instructions to their type before promotion.
3779 /// \p CreatedInstsCost[out] contains the cost of all the instructions
3780 /// created to promote the operand of Ext.
3781 /// Newly added extensions are inserted in \p Exts.
3782 /// Newly added truncates are inserted in \p Truncs.
3783 /// Should never be called directly.
3784 /// \return The promoted value which is used instead of Ext.
3785 static Value *promoteOperandForOther(Instruction *Ext,
3786 TypePromotionTransaction &TPT,
3787 InstrToOrigTy &PromotedInsts,
3788 unsigned &CreatedInstsCost,
3789 SmallVectorImpl<Instruction *> *Exts,
3790 SmallVectorImpl<Instruction *> *Truncs,
3791 const TargetLowering &TLI, bool IsSExt);
3793 /// \see promoteOperandForOther.
3794 static Value *signExtendOperandForOther(
3795 Instruction *Ext, TypePromotionTransaction &TPT,
3796 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3797 SmallVectorImpl<Instruction *> *Exts,
3798 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3799 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3800 Exts, Truncs, TLI, true);
3803 /// \see promoteOperandForOther.
3804 static Value *zeroExtendOperandForOther(
3805 Instruction *Ext, TypePromotionTransaction &TPT,
3806 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3807 SmallVectorImpl<Instruction *> *Exts,
3808 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3809 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3810 Exts, Truncs, TLI, false);
3814 /// Type for the utility function that promotes the operand of Ext.
3815 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
3816 InstrToOrigTy &PromotedInsts,
3817 unsigned &CreatedInstsCost,
3818 SmallVectorImpl<Instruction *> *Exts,
3819 SmallVectorImpl<Instruction *> *Truncs,
3820 const TargetLowering &TLI);
3821 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
3822 /// action to promote the operand of \p Ext instead of using Ext.
3823 /// \return NULL if no promotable action is possible with the current
3825 /// \p InsertedInsts keeps track of all the instructions inserted by the
3826 /// other CodeGenPrepare optimizations. This information is important
3827 /// because we do not want to promote these instructions as CodeGenPrepare
3828 /// will reinsert them later. Thus creating an infinite loop: create/remove.
3829 /// \p PromotedInsts maps the instructions to their type before promotion.
3830 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3831 const TargetLowering &TLI,
3832 const InstrToOrigTy &PromotedInsts);
3835 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3836 Type *ConsideredExtType,
3837 const InstrToOrigTy &PromotedInsts,
3839 // The promotion helper does not know how to deal with vector types yet.
3840 // To be able to fix that, we would need to fix the places where we
3841 // statically extend, e.g., constants and such.
3842 if (Inst->getType()->isVectorTy())
3845 // We can always get through zext.
3846 if (isa<ZExtInst>(Inst))
3849 // sext(sext) is ok too.
3850 if (IsSExt && isa<SExtInst>(Inst))
3853 // We can get through binary operator, if it is legal. In other words, the
3854 // binary operator must have a nuw or nsw flag.
3855 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3856 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3857 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3858 (IsSExt && BinOp->hasNoSignedWrap())))
3861 // Check if we can do the following simplification.
3862 // ext(trunc(opnd)) --> ext(opnd)
3863 if (!isa<TruncInst>(Inst))
3866 Value *OpndVal = Inst->getOperand(0);
3867 // Check if we can use this operand in the extension.
3868 // If the type is larger than the result type of the extension, we cannot.
3869 if (!OpndVal->getType()->isIntegerTy() ||
3870 OpndVal->getType()->getIntegerBitWidth() >
3871 ConsideredExtType->getIntegerBitWidth())
3874 // If the operand of the truncate is not an instruction, we will not have
3875 // any information on the dropped bits.
3876 // (Actually we could for constant but it is not worth the extra logic).
3877 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3881 // Check if the source of the type is narrow enough.
3882 // I.e., check that trunc just drops extended bits of the same kind of
3884 // #1 get the type of the operand and check the kind of the extended bits.
3885 const Type *OpndType;
3886 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3887 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
3888 OpndType = It->second.getPointer();
3889 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3890 OpndType = Opnd->getOperand(0)->getType();
3894 // #2 check that the truncate just drops extended bits.
3895 return Inst->getType()->getIntegerBitWidth() >=
3896 OpndType->getIntegerBitWidth();
3899 TypePromotionHelper::Action TypePromotionHelper::getAction(
3900 Instruction *Ext, const SetOfInstrs &InsertedInsts,
3901 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3902 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
3903 "Unexpected instruction type");
3904 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3905 Type *ExtTy = Ext->getType();
3906 bool IsSExt = isa<SExtInst>(Ext);
3907 // If the operand of the extension is not an instruction, we cannot
3909 // If it, check we can get through.
3910 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3913 // Do not promote if the operand has been added by codegenprepare.
3914 // Otherwise, it means we are undoing an optimization that is likely to be
3915 // redone, thus causing potential infinite loop.
3916 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3919 // SExt or Trunc instructions.
3920 // Return the related handler.
3921 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3922 isa<ZExtInst>(ExtOpnd))
3923 return promoteOperandForTruncAndAnyExt;
3925 // Regular instruction.
3926 // Abort early if we will have to insert non-free instructions.
3927 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3929 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3932 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3933 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
3934 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3935 SmallVectorImpl<Instruction *> *Exts,
3936 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3937 // By construction, the operand of SExt is an instruction. Otherwise we cannot
3938 // get through it and this method should not be called.
3939 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
3940 Value *ExtVal = SExt;
3941 bool HasMergedNonFreeExt = false;
3942 if (isa<ZExtInst>(SExtOpnd)) {
3943 // Replace s|zext(zext(opnd))
3945 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
3947 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
3948 TPT.replaceAllUsesWith(SExt, ZExt);
3949 TPT.eraseInstruction(SExt);
3952 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
3954 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
3956 CreatedInstsCost = 0;
3958 // Remove dead code.
3959 if (SExtOpnd->use_empty())
3960 TPT.eraseInstruction(SExtOpnd);
3962 // Check if the extension is still needed.
3963 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
3964 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
3967 Exts->push_back(ExtInst);
3968 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
3973 // At this point we have: ext ty opnd to ty.
3974 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
3975 Value *NextVal = ExtInst->getOperand(0);
3976 TPT.eraseInstruction(ExtInst, NextVal);
3980 Value *TypePromotionHelper::promoteOperandForOther(
3981 Instruction *Ext, TypePromotionTransaction &TPT,
3982 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3983 SmallVectorImpl<Instruction *> *Exts,
3984 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
3986 // By construction, the operand of Ext is an instruction. Otherwise we cannot
3987 // get through it and this method should not be called.
3988 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
3989 CreatedInstsCost = 0;
3990 if (!ExtOpnd->hasOneUse()) {
3991 // ExtOpnd will be promoted.
3992 // All its uses, but Ext, will need to use a truncated value of the
3993 // promoted version.
3994 // Create the truncate now.
3995 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
3996 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
3997 ITrunc->removeFromParent();
3998 // Insert it just after the definition.
3999 ITrunc->insertAfter(ExtOpnd);
4001 Truncs->push_back(ITrunc);
4004 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
4005 // Restore the operand of Ext (which has been replaced by the previous call
4006 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
4007 TPT.setOperand(Ext, 0, ExtOpnd);
4010 // Get through the Instruction:
4011 // 1. Update its type.
4012 // 2. Replace the uses of Ext by Inst.
4013 // 3. Extend each operand that needs to be extended.
4015 // Remember the original type of the instruction before promotion.
4016 // This is useful to know that the high bits are sign extended bits.
4017 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
4018 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
4020 TPT.mutateType(ExtOpnd, Ext->getType());
4022 TPT.replaceAllUsesWith(Ext, ExtOpnd);
4024 Instruction *ExtForOpnd = Ext;
4026 DEBUG(dbgs() << "Propagate Ext to operands\n");
4027 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
4029 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
4030 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
4031 !shouldExtOperand(ExtOpnd, OpIdx)) {
4032 DEBUG(dbgs() << "No need to propagate\n");
4035 // Check if we can statically extend the operand.
4036 Value *Opnd = ExtOpnd->getOperand(OpIdx);
4037 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
4038 DEBUG(dbgs() << "Statically extend\n");
4039 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
4040 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
4041 : Cst->getValue().zext(BitWidth);
4042 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
4045 // UndefValue are typed, so we have to statically sign extend them.
4046 if (isa<UndefValue>(Opnd)) {
4047 DEBUG(dbgs() << "Statically extend\n");
4048 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
4052 // Otherwise we have to explicity sign extend the operand.
4053 // Check if Ext was reused to extend an operand.
4055 // If yes, create a new one.
4056 DEBUG(dbgs() << "More operands to ext\n");
4057 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
4058 : TPT.createZExt(Ext, Opnd, Ext->getType());
4059 if (!isa<Instruction>(ValForExtOpnd)) {
4060 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
4063 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
4066 Exts->push_back(ExtForOpnd);
4067 TPT.setOperand(ExtForOpnd, 0, Opnd);
4069 // Move the sign extension before the insertion point.
4070 TPT.moveBefore(ExtForOpnd, ExtOpnd);
4071 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
4072 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
4073 // If more sext are required, new instructions will have to be created.
4074 ExtForOpnd = nullptr;
4076 if (ExtForOpnd == Ext) {
4077 DEBUG(dbgs() << "Extension is useless now\n");
4078 TPT.eraseInstruction(Ext);
4083 /// Check whether or not promoting an instruction to a wider type is profitable.
4084 /// \p NewCost gives the cost of extension instructions created by the
4086 /// \p OldCost gives the cost of extension instructions before the promotion
4087 /// plus the number of instructions that have been
4088 /// matched in the addressing mode the promotion.
4089 /// \p PromotedOperand is the value that has been promoted.
4090 /// \return True if the promotion is profitable, false otherwise.
4091 bool AddressingModeMatcher::isPromotionProfitable(
4092 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
4093 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
4094 // The cost of the new extensions is greater than the cost of the
4095 // old extension plus what we folded.
4096 // This is not profitable.
4097 if (NewCost > OldCost)
4099 if (NewCost < OldCost)
4101 // The promotion is neutral but it may help folding the sign extension in
4102 // loads for instance.
4103 // Check that we did not create an illegal instruction.
4104 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
4107 /// Given an instruction or constant expr, see if we can fold the operation
4108 /// into the addressing mode. If so, update the addressing mode and return
4109 /// true, otherwise return false without modifying AddrMode.
4110 /// If \p MovedAway is not NULL, it contains the information of whether or
4111 /// not AddrInst has to be folded into the addressing mode on success.
4112 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
4113 /// because it has been moved away.
4114 /// Thus AddrInst must not be added in the matched instructions.
4115 /// This state can happen when AddrInst is a sext, since it may be moved away.
4116 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
4117 /// not be referenced anymore.
4118 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
4121 // Avoid exponential behavior on extremely deep expression trees.
4122 if (Depth >= 5) return false;
4124 // By default, all matched instructions stay in place.
4129 case Instruction::PtrToInt:
4130 // PtrToInt is always a noop, as we know that the int type is pointer sized.
4131 return matchAddr(AddrInst->getOperand(0), Depth);
4132 case Instruction::IntToPtr: {
4133 auto AS = AddrInst->getType()->getPointerAddressSpace();
4134 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
4135 // This inttoptr is a no-op if the integer type is pointer sized.
4136 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
4137 return matchAddr(AddrInst->getOperand(0), Depth);
4140 case Instruction::BitCast:
4141 // BitCast is always a noop, and we can handle it as long as it is
4142 // int->int or pointer->pointer (we don't want int<->fp or something).
4143 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
4144 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
4145 // Don't touch identity bitcasts. These were probably put here by LSR,
4146 // and we don't want to mess around with them. Assume it knows what it
4148 AddrInst->getOperand(0)->getType() != AddrInst->getType())
4149 return matchAddr(AddrInst->getOperand(0), Depth);
4151 case Instruction::AddrSpaceCast: {
4153 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
4154 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
4155 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
4156 return matchAddr(AddrInst->getOperand(0), Depth);
4159 case Instruction::Add: {
4160 // Check to see if we can merge in the RHS then the LHS. If so, we win.
4161 ExtAddrMode BackupAddrMode = AddrMode;
4162 unsigned OldSize = AddrModeInsts.size();
4163 // Start a transaction at this point.
4164 // The LHS may match but not the RHS.
4165 // Therefore, we need a higher level restoration point to undo partially
4166 // matched operation.
4167 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4168 TPT.getRestorationPoint();
4170 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
4171 matchAddr(AddrInst->getOperand(0), Depth+1))
4174 // Restore the old addr mode info.
4175 AddrMode = BackupAddrMode;
4176 AddrModeInsts.resize(OldSize);
4177 TPT.rollback(LastKnownGood);
4179 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
4180 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
4181 matchAddr(AddrInst->getOperand(1), Depth+1))
4184 // Otherwise we definitely can't merge the ADD in.
4185 AddrMode = BackupAddrMode;
4186 AddrModeInsts.resize(OldSize);
4187 TPT.rollback(LastKnownGood);
4190 //case Instruction::Or:
4191 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
4193 case Instruction::Mul:
4194 case Instruction::Shl: {
4195 // Can only handle X*C and X << C.
4196 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
4199 int64_t Scale = RHS->getSExtValue();
4200 if (Opcode == Instruction::Shl)
4201 Scale = 1LL << Scale;
4203 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
4205 case Instruction::GetElementPtr: {
4206 // Scan the GEP. We check it if it contains constant offsets and at most
4207 // one variable offset.
4208 int VariableOperand = -1;
4209 unsigned VariableScale = 0;
4211 int64_t ConstantOffset = 0;
4212 gep_type_iterator GTI = gep_type_begin(AddrInst);
4213 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
4214 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
4215 const StructLayout *SL = DL.getStructLayout(STy);
4217 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
4218 ConstantOffset += SL->getElementOffset(Idx);
4220 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
4221 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
4222 ConstantOffset += CI->getSExtValue()*TypeSize;
4223 } else if (TypeSize) { // Scales of zero don't do anything.
4224 // We only allow one variable index at the moment.
4225 if (VariableOperand != -1)
4228 // Remember the variable index.
4229 VariableOperand = i;
4230 VariableScale = TypeSize;
4235 // A common case is for the GEP to only do a constant offset. In this case,
4236 // just add it to the disp field and check validity.
4237 if (VariableOperand == -1) {
4238 AddrMode.BaseOffs += ConstantOffset;
4239 if (ConstantOffset == 0 ||
4240 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
4241 // Check to see if we can fold the base pointer in too.
4242 if (matchAddr(AddrInst->getOperand(0), Depth+1))
4245 AddrMode.BaseOffs -= ConstantOffset;
4249 // Save the valid addressing mode in case we can't match.
4250 ExtAddrMode BackupAddrMode = AddrMode;
4251 unsigned OldSize = AddrModeInsts.size();
4253 // See if the scale and offset amount is valid for this target.
4254 AddrMode.BaseOffs += ConstantOffset;
4256 // Match the base operand of the GEP.
4257 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
4258 // If it couldn't be matched, just stuff the value in a register.
4259 if (AddrMode.HasBaseReg) {
4260 AddrMode = BackupAddrMode;
4261 AddrModeInsts.resize(OldSize);
4264 AddrMode.HasBaseReg = true;
4265 AddrMode.BaseReg = AddrInst->getOperand(0);
4268 // Match the remaining variable portion of the GEP.
4269 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
4271 // If it couldn't be matched, try stuffing the base into a register
4272 // instead of matching it, and retrying the match of the scale.
4273 AddrMode = BackupAddrMode;
4274 AddrModeInsts.resize(OldSize);
4275 if (AddrMode.HasBaseReg)
4277 AddrMode.HasBaseReg = true;
4278 AddrMode.BaseReg = AddrInst->getOperand(0);
4279 AddrMode.BaseOffs += ConstantOffset;
4280 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
4281 VariableScale, Depth)) {
4282 // If even that didn't work, bail.
4283 AddrMode = BackupAddrMode;
4284 AddrModeInsts.resize(OldSize);
4291 case Instruction::SExt:
4292 case Instruction::ZExt: {
4293 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
4297 // Try to move this ext out of the way of the addressing mode.
4298 // Ask for a method for doing so.
4299 TypePromotionHelper::Action TPH =
4300 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
4304 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4305 TPT.getRestorationPoint();
4306 unsigned CreatedInstsCost = 0;
4307 unsigned ExtCost = !TLI.isExtFree(Ext);
4308 Value *PromotedOperand =
4309 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
4310 // SExt has been moved away.
4311 // Thus either it will be rematched later in the recursive calls or it is
4312 // gone. Anyway, we must not fold it into the addressing mode at this point.
4316 // addr = gep base, idx
4318 // promotedOpnd = ext opnd <- no match here
4319 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
4320 // addr = gep base, op <- match
4324 assert(PromotedOperand &&
4325 "TypePromotionHelper should have filtered out those cases");
4327 ExtAddrMode BackupAddrMode = AddrMode;
4328 unsigned OldSize = AddrModeInsts.size();
4330 if (!matchAddr(PromotedOperand, Depth) ||
4331 // The total of the new cost is equal to the cost of the created
4333 // The total of the old cost is equal to the cost of the extension plus
4334 // what we have saved in the addressing mode.
4335 !isPromotionProfitable(CreatedInstsCost,
4336 ExtCost + (AddrModeInsts.size() - OldSize),
4338 AddrMode = BackupAddrMode;
4339 AddrModeInsts.resize(OldSize);
4340 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
4341 TPT.rollback(LastKnownGood);
4350 /// If we can, try to add the value of 'Addr' into the current addressing mode.
4351 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
4352 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
4355 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
4356 // Start a transaction at this point that we will rollback if the matching
4358 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4359 TPT.getRestorationPoint();
4360 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
4361 // Fold in immediates if legal for the target.
4362 AddrMode.BaseOffs += CI->getSExtValue();
4363 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4365 AddrMode.BaseOffs -= CI->getSExtValue();
4366 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
4367 // If this is a global variable, try to fold it into the addressing mode.
4368 if (!AddrMode.BaseGV) {
4369 AddrMode.BaseGV = GV;
4370 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4372 AddrMode.BaseGV = nullptr;
4374 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
4375 ExtAddrMode BackupAddrMode = AddrMode;
4376 unsigned OldSize = AddrModeInsts.size();
4378 // Check to see if it is possible to fold this operation.
4379 bool MovedAway = false;
4380 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
4381 // This instruction may have been moved away. If so, there is nothing
4385 // Okay, it's possible to fold this. Check to see if it is actually
4386 // *profitable* to do so. We use a simple cost model to avoid increasing
4387 // register pressure too much.
4388 if (I->hasOneUse() ||
4389 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
4390 AddrModeInsts.push_back(I);
4394 // It isn't profitable to do this, roll back.
4395 //cerr << "NOT FOLDING: " << *I;
4396 AddrMode = BackupAddrMode;
4397 AddrModeInsts.resize(OldSize);
4398 TPT.rollback(LastKnownGood);
4400 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
4401 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
4403 TPT.rollback(LastKnownGood);
4404 } else if (isa<ConstantPointerNull>(Addr)) {
4405 // Null pointer gets folded without affecting the addressing mode.
4409 // Worse case, the target should support [reg] addressing modes. :)
4410 if (!AddrMode.HasBaseReg) {
4411 AddrMode.HasBaseReg = true;
4412 AddrMode.BaseReg = Addr;
4413 // Still check for legality in case the target supports [imm] but not [i+r].
4414 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4416 AddrMode.HasBaseReg = false;
4417 AddrMode.BaseReg = nullptr;
4420 // If the base register is already taken, see if we can do [r+r].
4421 if (AddrMode.Scale == 0) {
4423 AddrMode.ScaledReg = Addr;
4424 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4427 AddrMode.ScaledReg = nullptr;
4430 TPT.rollback(LastKnownGood);
4434 /// Check to see if all uses of OpVal by the specified inline asm call are due
4435 /// to memory operands. If so, return true, otherwise return false.
4436 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
4437 const TargetMachine &TM) {
4438 const Function *F = CI->getParent()->getParent();
4439 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
4440 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
4441 TargetLowering::AsmOperandInfoVector TargetConstraints =
4442 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
4443 ImmutableCallSite(CI));
4444 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4445 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4447 // Compute the constraint code and ConstraintType to use.
4448 TLI->ComputeConstraintToUse(OpInfo, SDValue());
4450 // If this asm operand is our Value*, and if it isn't an indirect memory
4451 // operand, we can't fold it!
4452 if (OpInfo.CallOperandVal == OpVal &&
4453 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
4454 !OpInfo.isIndirect))
4461 /// Recursively walk all the uses of I until we find a memory use.
4462 /// If we find an obviously non-foldable instruction, return true.
4463 /// Add the ultimately found memory instructions to MemoryUses.
4464 static bool FindAllMemoryUses(
4466 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
4467 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
4468 // If we already considered this instruction, we're done.
4469 if (!ConsideredInsts.insert(I).second)
4472 // If this is an obviously unfoldable instruction, bail out.
4473 if (!MightBeFoldableInst(I))
4476 // Loop over all the uses, recursively processing them.
4477 for (Use &U : I->uses()) {
4478 Instruction *UserI = cast<Instruction>(U.getUser());
4480 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4481 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
4485 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4486 unsigned opNo = U.getOperandNo();
4487 if (opNo == 0) return true; // Storing addr, not into addr.
4488 MemoryUses.push_back(std::make_pair(SI, opNo));
4492 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4493 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
4494 if (!IA) return true;
4496 // If this is a memory operand, we're cool, otherwise bail out.
4497 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
4502 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
4509 /// Return true if Val is already known to be live at the use site that we're
4510 /// folding it into. If so, there is no cost to include it in the addressing
4511 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4512 /// instruction already.
4513 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4514 Value *KnownLive2) {
4515 // If Val is either of the known-live values, we know it is live!
4516 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4519 // All values other than instructions and arguments (e.g. constants) are live.
4520 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4522 // If Val is a constant sized alloca in the entry block, it is live, this is
4523 // true because it is just a reference to the stack/frame pointer, which is
4524 // live for the whole function.
4525 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4526 if (AI->isStaticAlloca())
4529 // Check to see if this value is already used in the memory instruction's
4530 // block. If so, it's already live into the block at the very least, so we
4531 // can reasonably fold it.
4532 return Val->isUsedInBasicBlock(MemoryInst->getParent());
4535 /// It is possible for the addressing mode of the machine to fold the specified
4536 /// instruction into a load or store that ultimately uses it.
4537 /// However, the specified instruction has multiple uses.
4538 /// Given this, it may actually increase register pressure to fold it
4539 /// into the load. For example, consider this code:
4543 /// use(Y) -> nonload/store
4547 /// In this case, Y has multiple uses, and can be folded into the load of Z
4548 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4549 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4550 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4551 /// number of computations either.
4553 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4554 /// X was live across 'load Z' for other reasons, we actually *would* want to
4555 /// fold the addressing mode in the Z case. This would make Y die earlier.
4556 bool AddressingModeMatcher::
4557 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4558 ExtAddrMode &AMAfter) {
4559 if (IgnoreProfitability) return true;
4561 // AMBefore is the addressing mode before this instruction was folded into it,
4562 // and AMAfter is the addressing mode after the instruction was folded. Get
4563 // the set of registers referenced by AMAfter and subtract out those
4564 // referenced by AMBefore: this is the set of values which folding in this
4565 // address extends the lifetime of.
4567 // Note that there are only two potential values being referenced here,
4568 // BaseReg and ScaleReg (global addresses are always available, as are any
4569 // folded immediates).
4570 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4572 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4573 // lifetime wasn't extended by adding this instruction.
4574 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4576 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4577 ScaledReg = nullptr;
4579 // If folding this instruction (and it's subexprs) didn't extend any live
4580 // ranges, we're ok with it.
4581 if (!BaseReg && !ScaledReg)
4584 // If all uses of this instruction are ultimately load/store/inlineasm's,
4585 // check to see if their addressing modes will include this instruction. If
4586 // so, we can fold it into all uses, so it doesn't matter if it has multiple
4588 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
4589 SmallPtrSet<Instruction*, 16> ConsideredInsts;
4590 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
4591 return false; // Has a non-memory, non-foldable use!
4593 // Now that we know that all uses of this instruction are part of a chain of
4594 // computation involving only operations that could theoretically be folded
4595 // into a memory use, loop over each of these uses and see if they could
4596 // *actually* fold the instruction.
4597 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4598 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4599 Instruction *User = MemoryUses[i].first;
4600 unsigned OpNo = MemoryUses[i].second;
4602 // Get the access type of this use. If the use isn't a pointer, we don't
4603 // know what it accesses.
4604 Value *Address = User->getOperand(OpNo);
4605 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4608 Type *AddressAccessTy = AddrTy->getElementType();
4609 unsigned AS = AddrTy->getAddressSpace();
4611 // Do a match against the root of this address, ignoring profitability. This
4612 // will tell us if the addressing mode for the memory operation will
4613 // *actually* cover the shared instruction.
4615 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4616 TPT.getRestorationPoint();
4617 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
4618 MemoryInst, Result, InsertedInsts,
4619 PromotedInsts, TPT);
4620 Matcher.IgnoreProfitability = true;
4621 bool Success = Matcher.matchAddr(Address, 0);
4622 (void)Success; assert(Success && "Couldn't select *anything*?");
4624 // The match was to check the profitability, the changes made are not
4625 // part of the original matcher. Therefore, they should be dropped
4626 // otherwise the original matcher will not present the right state.
4627 TPT.rollback(LastKnownGood);
4629 // If the match didn't cover I, then it won't be shared by it.
4630 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
4631 I) == MatchedAddrModeInsts.end())
4634 MatchedAddrModeInsts.clear();
4640 } // end anonymous namespace
4642 /// Return true if the specified values are defined in a
4643 /// different basic block than BB.
4644 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4645 if (Instruction *I = dyn_cast<Instruction>(V))
4646 return I->getParent() != BB;
4650 /// Load and Store Instructions often have addressing modes that can do
4651 /// significant amounts of computation. As such, instruction selection will try
4652 /// to get the load or store to do as much computation as possible for the
4653 /// program. The problem is that isel can only see within a single block. As
4654 /// such, we sink as much legal addressing mode work into the block as possible.
4656 /// This method is used to optimize both load/store and inline asms with memory
4658 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4659 Type *AccessTy, unsigned AddrSpace) {
4662 // Try to collapse single-value PHI nodes. This is necessary to undo
4663 // unprofitable PRE transformations.
4664 SmallVector<Value*, 8> worklist;
4665 SmallPtrSet<Value*, 16> Visited;
4666 worklist.push_back(Addr);
4668 // Use a worklist to iteratively look through PHI nodes, and ensure that
4669 // the addressing mode obtained from the non-PHI roots of the graph
4671 Value *Consensus = nullptr;
4672 unsigned NumUsesConsensus = 0;
4673 bool IsNumUsesConsensusValid = false;
4674 SmallVector<Instruction*, 16> AddrModeInsts;
4675 ExtAddrMode AddrMode;
4676 TypePromotionTransaction TPT;
4677 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4678 TPT.getRestorationPoint();
4679 while (!worklist.empty()) {
4680 Value *V = worklist.back();
4681 worklist.pop_back();
4683 // Break use-def graph loops.
4684 if (!Visited.insert(V).second) {
4685 Consensus = nullptr;
4689 // For a PHI node, push all of its incoming values.
4690 if (PHINode *P = dyn_cast<PHINode>(V)) {
4691 for (Value *IncValue : P->incoming_values())
4692 worklist.push_back(IncValue);
4696 // For non-PHIs, determine the addressing mode being computed.
4697 SmallVector<Instruction*, 16> NewAddrModeInsts;
4698 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4699 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
4700 InsertedInsts, PromotedInsts, TPT);
4702 // This check is broken into two cases with very similar code to avoid using
4703 // getNumUses() as much as possible. Some values have a lot of uses, so
4704 // calling getNumUses() unconditionally caused a significant compile-time
4708 AddrMode = NewAddrMode;
4709 AddrModeInsts = NewAddrModeInsts;
4711 } else if (NewAddrMode == AddrMode) {
4712 if (!IsNumUsesConsensusValid) {
4713 NumUsesConsensus = Consensus->getNumUses();
4714 IsNumUsesConsensusValid = true;
4717 // Ensure that the obtained addressing mode is equivalent to that obtained
4718 // for all other roots of the PHI traversal. Also, when choosing one
4719 // such root as representative, select the one with the most uses in order
4720 // to keep the cost modeling heuristics in AddressingModeMatcher
4722 unsigned NumUses = V->getNumUses();
4723 if (NumUses > NumUsesConsensus) {
4725 NumUsesConsensus = NumUses;
4726 AddrModeInsts = NewAddrModeInsts;
4731 Consensus = nullptr;
4735 // If the addressing mode couldn't be determined, or if multiple different
4736 // ones were determined, bail out now.
4738 TPT.rollback(LastKnownGood);
4743 // Check to see if any of the instructions supersumed by this addr mode are
4744 // non-local to I's BB.
4745 bool AnyNonLocal = false;
4746 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
4747 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
4753 // If all the instructions matched are already in this BB, don't do anything.
4755 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
4759 // Insert this computation right after this user. Since our caller is
4760 // scanning from the top of the BB to the bottom, reuse of the expr are
4761 // guaranteed to happen later.
4762 IRBuilder<> Builder(MemoryInst);
4764 // Now that we determined the addressing expression we want to use and know
4765 // that we have to sink it into this block. Check to see if we have already
4766 // done this for some other load/store instr in this block. If so, reuse the
4768 Value *&SunkAddr = SunkAddrs[Addr];
4770 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
4771 << *MemoryInst << "\n");
4772 if (SunkAddr->getType() != Addr->getType())
4773 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
4774 } else if (AddrSinkUsingGEPs ||
4775 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
4776 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
4778 // By default, we use the GEP-based method when AA is used later. This
4779 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4780 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4781 << *MemoryInst << "\n");
4782 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4783 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4785 // First, find the pointer.
4786 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4787 ResultPtr = AddrMode.BaseReg;
4788 AddrMode.BaseReg = nullptr;
4791 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4792 // We can't add more than one pointer together, nor can we scale a
4793 // pointer (both of which seem meaningless).
4794 if (ResultPtr || AddrMode.Scale != 1)
4797 ResultPtr = AddrMode.ScaledReg;
4801 if (AddrMode.BaseGV) {
4805 ResultPtr = AddrMode.BaseGV;
4808 // If the real base value actually came from an inttoptr, then the matcher
4809 // will look through it and provide only the integer value. In that case,
4811 if (!ResultPtr && AddrMode.BaseReg) {
4813 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
4814 AddrMode.BaseReg = nullptr;
4815 } else if (!ResultPtr && AddrMode.Scale == 1) {
4817 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
4822 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4823 SunkAddr = Constant::getNullValue(Addr->getType());
4824 } else if (!ResultPtr) {
4828 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4829 Type *I8Ty = Builder.getInt8Ty();
4831 // Start with the base register. Do this first so that subsequent address
4832 // matching finds it last, which will prevent it from trying to match it
4833 // as the scaled value in case it happens to be a mul. That would be
4834 // problematic if we've sunk a different mul for the scale, because then
4835 // we'd end up sinking both muls.
4836 if (AddrMode.BaseReg) {
4837 Value *V = AddrMode.BaseReg;
4838 if (V->getType() != IntPtrTy)
4839 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4844 // Add the scale value.
4845 if (AddrMode.Scale) {
4846 Value *V = AddrMode.ScaledReg;
4847 if (V->getType() == IntPtrTy) {
4849 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4850 cast<IntegerType>(V->getType())->getBitWidth()) {
4851 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4853 // It is only safe to sign extend the BaseReg if we know that the math
4854 // required to create it did not overflow before we extend it. Since
4855 // the original IR value was tossed in favor of a constant back when
4856 // the AddrMode was created we need to bail out gracefully if widths
4857 // do not match instead of extending it.
4858 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
4859 if (I && (ResultIndex != AddrMode.BaseReg))
4860 I->eraseFromParent();
4864 if (AddrMode.Scale != 1)
4865 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4868 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4873 // Add in the Base Offset if present.
4874 if (AddrMode.BaseOffs) {
4875 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4877 // We need to add this separately from the scale above to help with
4878 // SDAG consecutive load/store merging.
4879 if (ResultPtr->getType() != I8PtrTy)
4880 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
4881 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4888 SunkAddr = ResultPtr;
4890 if (ResultPtr->getType() != I8PtrTy)
4891 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
4892 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4895 if (SunkAddr->getType() != Addr->getType())
4896 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
4899 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4900 << *MemoryInst << "\n");
4901 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4902 Value *Result = nullptr;
4904 // Start with the base register. Do this first so that subsequent address
4905 // matching finds it last, which will prevent it from trying to match it
4906 // as the scaled value in case it happens to be a mul. That would be
4907 // problematic if we've sunk a different mul for the scale, because then
4908 // we'd end up sinking both muls.
4909 if (AddrMode.BaseReg) {
4910 Value *V = AddrMode.BaseReg;
4911 if (V->getType()->isPointerTy())
4912 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4913 if (V->getType() != IntPtrTy)
4914 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4918 // Add the scale value.
4919 if (AddrMode.Scale) {
4920 Value *V = AddrMode.ScaledReg;
4921 if (V->getType() == IntPtrTy) {
4923 } else if (V->getType()->isPointerTy()) {
4924 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4925 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4926 cast<IntegerType>(V->getType())->getBitWidth()) {
4927 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4929 // It is only safe to sign extend the BaseReg if we know that the math
4930 // required to create it did not overflow before we extend it. Since
4931 // the original IR value was tossed in favor of a constant back when
4932 // the AddrMode was created we need to bail out gracefully if widths
4933 // do not match instead of extending it.
4934 Instruction *I = dyn_cast_or_null<Instruction>(Result);
4935 if (I && (Result != AddrMode.BaseReg))
4936 I->eraseFromParent();
4939 if (AddrMode.Scale != 1)
4940 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4943 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4948 // Add in the BaseGV if present.
4949 if (AddrMode.BaseGV) {
4950 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
4952 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4957 // Add in the Base Offset if present.
4958 if (AddrMode.BaseOffs) {
4959 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4961 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4967 SunkAddr = Constant::getNullValue(Addr->getType());
4969 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
4972 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
4974 // If we have no uses, recursively delete the value and all dead instructions
4976 if (Repl->use_empty()) {
4977 // This can cause recursive deletion, which can invalidate our iterator.
4978 // Use a WeakVH to hold onto it in case this happens.
4979 WeakVH IterHandle(&*CurInstIterator);
4980 BasicBlock *BB = CurInstIterator->getParent();
4982 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
4984 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
4985 // If the iterator instruction was recursively deleted, start over at the
4986 // start of the block.
4987 CurInstIterator = BB->begin();
4995 /// If there are any memory operands, use OptimizeMemoryInst to sink their
4996 /// address computing into the block when possible / profitable.
4997 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
4998 bool MadeChange = false;
5000 const TargetRegisterInfo *TRI =
5001 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
5002 TargetLowering::AsmOperandInfoVector TargetConstraints =
5003 TLI->ParseConstraints(*DL, TRI, CS);
5005 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5006 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
5008 // Compute the constraint code and ConstraintType to use.
5009 TLI->ComputeConstraintToUse(OpInfo, SDValue());
5011 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5012 OpInfo.isIndirect) {
5013 Value *OpVal = CS->getArgOperand(ArgNo++);
5014 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
5015 } else if (OpInfo.Type == InlineAsm::isInput)
5022 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
5023 /// sign extensions.
5024 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
5025 assert(!Inst->use_empty() && "Input must have at least one use");
5026 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
5027 bool IsSExt = isa<SExtInst>(FirstUser);
5028 Type *ExtTy = FirstUser->getType();
5029 for (const User *U : Inst->users()) {
5030 const Instruction *UI = cast<Instruction>(U);
5031 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
5033 Type *CurTy = UI->getType();
5034 // Same input and output types: Same instruction after CSE.
5038 // If IsSExt is true, we are in this situation:
5040 // b = sext ty1 a to ty2
5041 // c = sext ty1 a to ty3
5042 // Assuming ty2 is shorter than ty3, this could be turned into:
5044 // b = sext ty1 a to ty2
5045 // c = sext ty2 b to ty3
5046 // However, the last sext is not free.
5050 // This is a ZExt, maybe this is free to extend from one type to another.
5051 // In that case, we would not account for a different use.
5054 if (ExtTy->getScalarType()->getIntegerBitWidth() >
5055 CurTy->getScalarType()->getIntegerBitWidth()) {
5063 if (!TLI.isZExtFree(NarrowTy, LargeTy))
5066 // All uses are the same or can be derived from one another for free.
5070 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
5071 /// load instruction.
5072 /// If an ext(load) can be formed, it is returned via \p LI for the load
5073 /// and \p Inst for the extension.
5074 /// Otherwise LI == nullptr and Inst == nullptr.
5075 /// When some promotion happened, \p TPT contains the proper state to
5078 /// \return true when promoting was necessary to expose the ext(load)
5079 /// opportunity, false otherwise.
5083 /// %ld = load i32* %addr
5084 /// %add = add nuw i32 %ld, 4
5085 /// %zext = zext i32 %add to i64
5089 /// %ld = load i32* %addr
5090 /// %zext = zext i32 %ld to i64
5091 /// %add = add nuw i64 %zext, 4
5093 /// Thanks to the promotion, we can match zext(load i32*) to i64.
5094 bool CodeGenPrepare::extLdPromotion(TypePromotionTransaction &TPT,
5095 LoadInst *&LI, Instruction *&Inst,
5096 const SmallVectorImpl<Instruction *> &Exts,
5097 unsigned CreatedInstsCost = 0) {
5098 // Iterate over all the extensions to see if one form an ext(load).
5099 for (auto I : Exts) {
5100 // Check if we directly have ext(load).
5101 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
5103 // No promotion happened here.
5106 // Check whether or not we want to do any promotion.
5107 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
5109 // Get the action to perform the promotion.
5110 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
5111 I, InsertedInsts, *TLI, PromotedInsts);
5112 // Check if we can promote.
5115 // Save the current state.
5116 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5117 TPT.getRestorationPoint();
5118 SmallVector<Instruction *, 4> NewExts;
5119 unsigned NewCreatedInstsCost = 0;
5120 unsigned ExtCost = !TLI->isExtFree(I);
5122 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
5123 &NewExts, nullptr, *TLI);
5124 assert(PromotedVal &&
5125 "TypePromotionHelper should have filtered out those cases");
5127 // We would be able to merge only one extension in a load.
5128 // Therefore, if we have more than 1 new extension we heuristically
5129 // cut this search path, because it means we degrade the code quality.
5130 // With exactly 2, the transformation is neutral, because we will merge
5131 // one extension but leave one. However, we optimistically keep going,
5132 // because the new extension may be removed too.
5133 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
5134 TotalCreatedInstsCost -= ExtCost;
5135 if (!StressExtLdPromotion &&
5136 (TotalCreatedInstsCost > 1 ||
5137 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
5138 // The promotion is not profitable, rollback to the previous state.
5139 TPT.rollback(LastKnownGood);
5142 // The promotion is profitable.
5143 // Check if it exposes an ext(load).
5144 (void)extLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
5145 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
5146 // If we have created a new extension, i.e., now we have two
5147 // extensions. We must make sure one of them is merged with
5148 // the load, otherwise we may degrade the code quality.
5149 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
5150 // Promotion happened.
5152 // If this does not help to expose an ext(load) then, rollback.
5153 TPT.rollback(LastKnownGood);
5155 // None of the extension can form an ext(load).
5161 /// Move a zext or sext fed by a load into the same basic block as the load,
5162 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
5163 /// extend into the load.
5164 /// \p I[in/out] the extension may be modified during the process if some
5165 /// promotions apply.
5167 bool CodeGenPrepare::moveExtToFormExtLoad(Instruction *&I) {
5168 // Try to promote a chain of computation if it allows to form
5169 // an extended load.
5170 TypePromotionTransaction TPT;
5171 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5172 TPT.getRestorationPoint();
5173 SmallVector<Instruction *, 1> Exts;
5175 // Look for a load being extended.
5176 LoadInst *LI = nullptr;
5177 Instruction *OldExt = I;
5178 bool HasPromoted = extLdPromotion(TPT, LI, I, Exts);
5180 assert(!HasPromoted && !LI && "If we did not match any load instruction "
5181 "the code must remain the same");
5186 // If they're already in the same block, there's nothing to do.
5187 // Make the cheap checks first if we did not promote.
5188 // If we promoted, we need to check if it is indeed profitable.
5189 if (!HasPromoted && LI->getParent() == I->getParent())
5192 EVT VT = TLI->getValueType(*DL, I->getType());
5193 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
5195 // If the load has other users and the truncate is not free, this probably
5196 // isn't worthwhile.
5197 if (!LI->hasOneUse() && TLI &&
5198 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
5199 !TLI->isTruncateFree(I->getType(), LI->getType())) {
5201 TPT.rollback(LastKnownGood);
5205 // Check whether the target supports casts folded into loads.
5207 if (isa<ZExtInst>(I))
5208 LType = ISD::ZEXTLOAD;
5210 assert(isa<SExtInst>(I) && "Unexpected ext type!");
5211 LType = ISD::SEXTLOAD;
5213 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
5215 TPT.rollback(LastKnownGood);
5219 // Move the extend into the same block as the load, so that SelectionDAG
5222 I->removeFromParent();
5228 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
5229 BasicBlock *DefBB = I->getParent();
5231 // If the result of a {s|z}ext and its source are both live out, rewrite all
5232 // other uses of the source with result of extension.
5233 Value *Src = I->getOperand(0);
5234 if (Src->hasOneUse())
5237 // Only do this xform if truncating is free.
5238 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
5241 // Only safe to perform the optimization if the source is also defined in
5243 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
5246 bool DefIsLiveOut = false;
5247 for (User *U : I->users()) {
5248 Instruction *UI = cast<Instruction>(U);
5250 // Figure out which BB this ext is used in.
5251 BasicBlock *UserBB = UI->getParent();
5252 if (UserBB == DefBB) continue;
5253 DefIsLiveOut = true;
5259 // Make sure none of the uses are PHI nodes.
5260 for (User *U : Src->users()) {
5261 Instruction *UI = cast<Instruction>(U);
5262 BasicBlock *UserBB = UI->getParent();
5263 if (UserBB == DefBB) continue;
5264 // Be conservative. We don't want this xform to end up introducing
5265 // reloads just before load / store instructions.
5266 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
5270 // InsertedTruncs - Only insert one trunc in each block once.
5271 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
5273 bool MadeChange = false;
5274 for (Use &U : Src->uses()) {
5275 Instruction *User = cast<Instruction>(U.getUser());
5277 // Figure out which BB this ext is used in.
5278 BasicBlock *UserBB = User->getParent();
5279 if (UserBB == DefBB) continue;
5281 // Both src and def are live in this block. Rewrite the use.
5282 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
5284 if (!InsertedTrunc) {
5285 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5286 assert(InsertPt != UserBB->end());
5287 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5288 InsertedInsts.insert(InsertedTrunc);
5291 // Replace a use of the {s|z}ext source with a use of the result.
5300 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
5301 // just after the load if the target can fold this into one extload instruction,
5302 // with the hope of eliminating some of the other later "and" instructions using
5303 // the loaded value. "and"s that are made trivially redundant by the insertion
5304 // of the new "and" are removed by this function, while others (e.g. those whose
5305 // path from the load goes through a phi) are left for isel to potentially
5338 // becomes (after a call to optimizeLoadExt for each load):
5342 // x1' = and x1, 0xff
5346 // x2' = and x2, 0xff
5353 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5355 if (!Load->isSimple() ||
5356 !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
5359 // Skip loads we've already transformed or have no reason to transform.
5360 if (Load->hasOneUse()) {
5361 User *LoadUser = *Load->user_begin();
5362 if (cast<Instruction>(LoadUser)->getParent() == Load->getParent() &&
5363 !dyn_cast<PHINode>(LoadUser))
5367 // Look at all uses of Load, looking through phis, to determine how many bits
5368 // of the loaded value are needed.
5369 SmallVector<Instruction *, 8> WorkList;
5370 SmallPtrSet<Instruction *, 16> Visited;
5371 SmallVector<Instruction *, 8> AndsToMaybeRemove;
5372 for (auto *U : Load->users())
5373 WorkList.push_back(cast<Instruction>(U));
5375 EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5376 unsigned BitWidth = LoadResultVT.getSizeInBits();
5377 APInt DemandBits(BitWidth, 0);
5378 APInt WidestAndBits(BitWidth, 0);
5380 while (!WorkList.empty()) {
5381 Instruction *I = WorkList.back();
5382 WorkList.pop_back();
5384 // Break use-def graph loops.
5385 if (!Visited.insert(I).second)
5388 // For a PHI node, push all of its users.
5389 if (auto *Phi = dyn_cast<PHINode>(I)) {
5390 for (auto *U : Phi->users())
5391 WorkList.push_back(cast<Instruction>(U));
5395 switch (I->getOpcode()) {
5396 case llvm::Instruction::And: {
5397 auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5400 APInt AndBits = AndC->getValue();
5401 DemandBits |= AndBits;
5402 // Keep track of the widest and mask we see.
5403 if (AndBits.ugt(WidestAndBits))
5404 WidestAndBits = AndBits;
5405 if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5406 AndsToMaybeRemove.push_back(I);
5410 case llvm::Instruction::Shl: {
5411 auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5414 uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5415 auto ShlDemandBits = APInt::getAllOnesValue(BitWidth).lshr(ShiftAmt);
5416 DemandBits |= ShlDemandBits;
5420 case llvm::Instruction::Trunc: {
5421 EVT TruncVT = TLI->getValueType(*DL, I->getType());
5422 unsigned TruncBitWidth = TruncVT.getSizeInBits();
5423 auto TruncBits = APInt::getAllOnesValue(TruncBitWidth).zext(BitWidth);
5424 DemandBits |= TruncBits;
5433 uint32_t ActiveBits = DemandBits.getActiveBits();
5434 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5435 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5436 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5437 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5438 // followed by an AND.
5439 // TODO: Look into removing this restriction by fixing backends to either
5440 // return false for isLoadExtLegal for i1 or have them select this pattern to
5441 // a single instruction.
5443 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5444 // mask, since these are the only ands that will be removed by isel.
5445 if (ActiveBits <= 1 || !APIntOps::isMask(ActiveBits, DemandBits) ||
5446 WidestAndBits != DemandBits)
5449 LLVMContext &Ctx = Load->getType()->getContext();
5450 Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5451 EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5453 // Reject cases that won't be matched as extloads.
5454 if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5455 !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5458 IRBuilder<> Builder(Load->getNextNode());
5459 auto *NewAnd = dyn_cast<Instruction>(
5460 Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5462 // Replace all uses of load with new and (except for the use of load in the
5464 Load->replaceAllUsesWith(NewAnd);
5465 NewAnd->setOperand(0, Load);
5467 // Remove any and instructions that are now redundant.
5468 for (auto *And : AndsToMaybeRemove)
5469 // Check that the and mask is the same as the one we decided to put on the
5471 if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5472 And->replaceAllUsesWith(NewAnd);
5473 if (&*CurInstIterator == And)
5474 CurInstIterator = std::next(And->getIterator());
5475 And->eraseFromParent();
5483 /// Check if V (an operand of a select instruction) is an expensive instruction
5484 /// that is only used once.
5485 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5486 auto *I = dyn_cast<Instruction>(V);
5487 // If it's safe to speculatively execute, then it should not have side
5488 // effects; therefore, it's safe to sink and possibly *not* execute.
5489 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5490 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
5493 /// Returns true if a SelectInst should be turned into an explicit branch.
5494 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
5496 // FIXME: This should use the same heuristics as IfConversion to determine
5497 // whether a select is better represented as a branch. This requires that
5498 // branch probability metadata is preserved for the select, which is not the
5501 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5503 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5504 // comparison condition. If the compare has more than one use, there's
5505 // probably another cmov or setcc around, so it's not worth emitting a branch.
5506 if (!Cmp || !Cmp->hasOneUse())
5509 Value *CmpOp0 = Cmp->getOperand(0);
5510 Value *CmpOp1 = Cmp->getOperand(1);
5512 // Emit "cmov on compare with a memory operand" as a branch to avoid stalls
5513 // on a load from memory. But if the load is used more than once, do not
5514 // change the select to a branch because the load is probably needed
5515 // regardless of whether the branch is taken or not.
5516 if ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
5517 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()))
5520 // If either operand of the select is expensive and only needed on one side
5521 // of the select, we should form a branch.
5522 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5523 sinkSelectOperand(TTI, SI->getFalseValue()))
5530 /// If we have a SelectInst that will likely profit from branch prediction,
5531 /// turn it into a branch.
5532 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5533 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
5535 // Can we convert the 'select' to CF ?
5536 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
5539 TargetLowering::SelectSupportKind SelectKind;
5541 SelectKind = TargetLowering::VectorMaskSelect;
5542 else if (SI->getType()->isVectorTy())
5543 SelectKind = TargetLowering::ScalarCondVectorVal;
5545 SelectKind = TargetLowering::ScalarValSelect;
5547 // Do we have efficient codegen support for this kind of 'selects' ?
5548 if (TLI->isSelectSupported(SelectKind)) {
5549 // We have efficient codegen support for the select instruction.
5550 // Check if it is profitable to keep this 'select'.
5551 if (!TLI->isPredictableSelectExpensive() ||
5552 !isFormingBranchFromSelectProfitable(TTI, SI))
5558 // Transform a sequence like this:
5560 // %cmp = cmp uge i32 %a, %b
5561 // %sel = select i1 %cmp, i32 %c, i32 %d
5565 // %cmp = cmp uge i32 %a, %b
5566 // br i1 %cmp, label %select.true, label %select.false
5568 // br label %select.end
5570 // br label %select.end
5572 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
5574 // In addition, we may sink instructions that produce %c or %d from
5575 // the entry block into the destination(s) of the new branch.
5576 // If the true or false blocks do not contain a sunken instruction, that
5577 // block and its branch may be optimized away. In that case, one side of the
5578 // first branch will point directly to select.end, and the corresponding PHI
5579 // predecessor block will be the start block.
5581 // First, we split the block containing the select into 2 blocks.
5582 BasicBlock *StartBlock = SI->getParent();
5583 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
5584 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
5586 // Delete the unconditional branch that was just created by the split.
5587 StartBlock->getTerminator()->eraseFromParent();
5589 // These are the new basic blocks for the conditional branch.
5590 // At least one will become an actual new basic block.
5591 BasicBlock *TrueBlock = nullptr;
5592 BasicBlock *FalseBlock = nullptr;
5594 // Sink expensive instructions into the conditional blocks to avoid executing
5595 // them speculatively.
5596 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
5597 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
5598 EndBlock->getParent(), EndBlock);
5599 auto *TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
5600 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
5601 TrueInst->moveBefore(TrueBranch);
5603 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
5604 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
5605 EndBlock->getParent(), EndBlock);
5606 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5607 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
5608 FalseInst->moveBefore(FalseBranch);
5611 // If there was nothing to sink, then arbitrarily choose the 'false' side
5612 // for a new input value to the PHI.
5613 if (TrueBlock == FalseBlock) {
5614 assert(TrueBlock == nullptr &&
5615 "Unexpected basic block transform while optimizing select");
5617 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
5618 EndBlock->getParent(), EndBlock);
5619 BranchInst::Create(EndBlock, FalseBlock);
5622 // Insert the real conditional branch based on the original condition.
5623 // If we did not create a new block for one of the 'true' or 'false' paths
5624 // of the condition, it means that side of the branch goes to the end block
5625 // directly and the path originates from the start block from the point of
5626 // view of the new PHI.
5627 if (TrueBlock == nullptr) {
5628 BranchInst::Create(EndBlock, FalseBlock, SI->getCondition(), SI);
5629 TrueBlock = StartBlock;
5630 } else if (FalseBlock == nullptr) {
5631 BranchInst::Create(TrueBlock, EndBlock, SI->getCondition(), SI);
5632 FalseBlock = StartBlock;
5634 BranchInst::Create(TrueBlock, FalseBlock, SI->getCondition(), SI);
5637 // The select itself is replaced with a PHI Node.
5638 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
5640 PN->addIncoming(SI->getTrueValue(), TrueBlock);
5641 PN->addIncoming(SI->getFalseValue(), FalseBlock);
5643 SI->replaceAllUsesWith(PN);
5644 SI->eraseFromParent();
5646 // Instruct OptimizeBlock to skip to the next block.
5647 CurInstIterator = StartBlock->end();
5648 ++NumSelectsExpanded;
5652 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
5653 SmallVector<int, 16> Mask(SVI->getShuffleMask());
5655 for (unsigned i = 0; i < Mask.size(); ++i) {
5656 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
5658 SplatElem = Mask[i];
5664 /// Some targets have expensive vector shifts if the lanes aren't all the same
5665 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
5666 /// it's often worth sinking a shufflevector splat down to its use so that
5667 /// codegen can spot all lanes are identical.
5668 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
5669 BasicBlock *DefBB = SVI->getParent();
5671 // Only do this xform if variable vector shifts are particularly expensive.
5672 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
5675 // We only expect better codegen by sinking a shuffle if we can recognise a
5677 if (!isBroadcastShuffle(SVI))
5680 // InsertedShuffles - Only insert a shuffle in each block once.
5681 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
5683 bool MadeChange = false;
5684 for (User *U : SVI->users()) {
5685 Instruction *UI = cast<Instruction>(U);
5687 // Figure out which BB this ext is used in.
5688 BasicBlock *UserBB = UI->getParent();
5689 if (UserBB == DefBB) continue;
5691 // For now only apply this when the splat is used by a shift instruction.
5692 if (!UI->isShift()) continue;
5694 // Everything checks out, sink the shuffle if the user's block doesn't
5695 // already have a copy.
5696 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
5698 if (!InsertedShuffle) {
5699 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5700 assert(InsertPt != UserBB->end());
5702 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5703 SVI->getOperand(2), "", &*InsertPt);
5706 UI->replaceUsesOfWith(SVI, InsertedShuffle);
5710 // If we removed all uses, nuke the shuffle.
5711 if (SVI->use_empty()) {
5712 SVI->eraseFromParent();
5719 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
5723 Value *Cond = SI->getCondition();
5724 Type *OldType = Cond->getType();
5725 LLVMContext &Context = Cond->getContext();
5726 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
5727 unsigned RegWidth = RegType.getSizeInBits();
5729 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
5732 // If the register width is greater than the type width, expand the condition
5733 // of the switch instruction and each case constant to the width of the
5734 // register. By widening the type of the switch condition, subsequent
5735 // comparisons (for case comparisons) will not need to be extended to the
5736 // preferred register width, so we will potentially eliminate N-1 extends,
5737 // where N is the number of cases in the switch.
5738 auto *NewType = Type::getIntNTy(Context, RegWidth);
5740 // Zero-extend the switch condition and case constants unless the switch
5741 // condition is a function argument that is already being sign-extended.
5742 // In that case, we can avoid an unnecessary mask/extension by sign-extending
5743 // everything instead.
5744 Instruction::CastOps ExtType = Instruction::ZExt;
5745 if (auto *Arg = dyn_cast<Argument>(Cond))
5746 if (Arg->hasSExtAttr())
5747 ExtType = Instruction::SExt;
5749 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
5750 ExtInst->insertBefore(SI);
5751 SI->setCondition(ExtInst);
5752 for (SwitchInst::CaseIt Case : SI->cases()) {
5753 APInt NarrowConst = Case.getCaseValue()->getValue();
5754 APInt WideConst = (ExtType == Instruction::ZExt) ?
5755 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
5756 Case.setValue(ConstantInt::get(Context, WideConst));
5763 /// \brief Helper class to promote a scalar operation to a vector one.
5764 /// This class is used to move downward extractelement transition.
5766 /// a = vector_op <2 x i32>
5767 /// b = extractelement <2 x i32> a, i32 0
5772 /// a = vector_op <2 x i32>
5773 /// c = vector_op a (equivalent to scalar_op on the related lane)
5774 /// * d = extractelement <2 x i32> c, i32 0
5776 /// Assuming both extractelement and store can be combine, we get rid of the
5778 class VectorPromoteHelper {
5779 /// DataLayout associated with the current module.
5780 const DataLayout &DL;
5782 /// Used to perform some checks on the legality of vector operations.
5783 const TargetLowering &TLI;
5785 /// Used to estimated the cost of the promoted chain.
5786 const TargetTransformInfo &TTI;
5788 /// The transition being moved downwards.
5789 Instruction *Transition;
5790 /// The sequence of instructions to be promoted.
5791 SmallVector<Instruction *, 4> InstsToBePromoted;
5792 /// Cost of combining a store and an extract.
5793 unsigned StoreExtractCombineCost;
5794 /// Instruction that will be combined with the transition.
5795 Instruction *CombineInst;
5797 /// \brief The instruction that represents the current end of the transition.
5798 /// Since we are faking the promotion until we reach the end of the chain
5799 /// of computation, we need a way to get the current end of the transition.
5800 Instruction *getEndOfTransition() const {
5801 if (InstsToBePromoted.empty())
5803 return InstsToBePromoted.back();
5806 /// \brief Return the index of the original value in the transition.
5807 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
5808 /// c, is at index 0.
5809 unsigned getTransitionOriginalValueIdx() const {
5810 assert(isa<ExtractElementInst>(Transition) &&
5811 "Other kind of transitions are not supported yet");
5815 /// \brief Return the index of the index in the transition.
5816 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
5818 unsigned getTransitionIdx() const {
5819 assert(isa<ExtractElementInst>(Transition) &&
5820 "Other kind of transitions are not supported yet");
5824 /// \brief Get the type of the transition.
5825 /// This is the type of the original value.
5826 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
5827 /// transition is <2 x i32>.
5828 Type *getTransitionType() const {
5829 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
5832 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
5833 /// I.e., we have the following sequence:
5834 /// Def = Transition <ty1> a to <ty2>
5835 /// b = ToBePromoted <ty2> Def, ...
5837 /// b = ToBePromoted <ty1> a, ...
5838 /// Def = Transition <ty1> ToBePromoted to <ty2>
5839 void promoteImpl(Instruction *ToBePromoted);
5841 /// \brief Check whether or not it is profitable to promote all the
5842 /// instructions enqueued to be promoted.
5843 bool isProfitableToPromote() {
5844 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
5845 unsigned Index = isa<ConstantInt>(ValIdx)
5846 ? cast<ConstantInt>(ValIdx)->getZExtValue()
5848 Type *PromotedType = getTransitionType();
5850 StoreInst *ST = cast<StoreInst>(CombineInst);
5851 unsigned AS = ST->getPointerAddressSpace();
5852 unsigned Align = ST->getAlignment();
5853 // Check if this store is supported.
5854 if (!TLI.allowsMisalignedMemoryAccesses(
5855 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
5857 // If this is not supported, there is no way we can combine
5858 // the extract with the store.
5862 // The scalar chain of computation has to pay for the transition
5863 // scalar to vector.
5864 // The vector chain has to account for the combining cost.
5865 uint64_t ScalarCost =
5866 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
5867 uint64_t VectorCost = StoreExtractCombineCost;
5868 for (const auto &Inst : InstsToBePromoted) {
5869 // Compute the cost.
5870 // By construction, all instructions being promoted are arithmetic ones.
5871 // Moreover, one argument is a constant that can be viewed as a splat
5873 Value *Arg0 = Inst->getOperand(0);
5874 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
5875 isa<ConstantFP>(Arg0);
5876 TargetTransformInfo::OperandValueKind Arg0OVK =
5877 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
5878 : TargetTransformInfo::OK_AnyValue;
5879 TargetTransformInfo::OperandValueKind Arg1OVK =
5880 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
5881 : TargetTransformInfo::OK_AnyValue;
5882 ScalarCost += TTI.getArithmeticInstrCost(
5883 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
5884 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
5887 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
5888 << ScalarCost << "\nVector: " << VectorCost << '\n');
5889 return ScalarCost > VectorCost;
5892 /// \brief Generate a constant vector with \p Val with the same
5893 /// number of elements as the transition.
5894 /// \p UseSplat defines whether or not \p Val should be replicated
5895 /// across the whole vector.
5896 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
5897 /// otherwise we generate a vector with as many undef as possible:
5898 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
5899 /// used at the index of the extract.
5900 Value *getConstantVector(Constant *Val, bool UseSplat) const {
5901 unsigned ExtractIdx = UINT_MAX;
5903 // If we cannot determine where the constant must be, we have to
5904 // use a splat constant.
5905 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
5906 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
5907 ExtractIdx = CstVal->getSExtValue();
5912 unsigned End = getTransitionType()->getVectorNumElements();
5914 return ConstantVector::getSplat(End, Val);
5916 SmallVector<Constant *, 4> ConstVec;
5917 UndefValue *UndefVal = UndefValue::get(Val->getType());
5918 for (unsigned Idx = 0; Idx != End; ++Idx) {
5919 if (Idx == ExtractIdx)
5920 ConstVec.push_back(Val);
5922 ConstVec.push_back(UndefVal);
5924 return ConstantVector::get(ConstVec);
5927 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
5928 /// in \p Use can trigger undefined behavior.
5929 static bool canCauseUndefinedBehavior(const Instruction *Use,
5930 unsigned OperandIdx) {
5931 // This is not safe to introduce undef when the operand is on
5932 // the right hand side of a division-like instruction.
5933 if (OperandIdx != 1)
5935 switch (Use->getOpcode()) {
5938 case Instruction::SDiv:
5939 case Instruction::UDiv:
5940 case Instruction::SRem:
5941 case Instruction::URem:
5943 case Instruction::FDiv:
5944 case Instruction::FRem:
5945 return !Use->hasNoNaNs();
5947 llvm_unreachable(nullptr);
5951 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
5952 const TargetTransformInfo &TTI, Instruction *Transition,
5953 unsigned CombineCost)
5954 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
5955 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
5956 assert(Transition && "Do not know how to promote null");
5959 /// \brief Check if we can promote \p ToBePromoted to \p Type.
5960 bool canPromote(const Instruction *ToBePromoted) const {
5961 // We could support CastInst too.
5962 return isa<BinaryOperator>(ToBePromoted);
5965 /// \brief Check if it is profitable to promote \p ToBePromoted
5966 /// by moving downward the transition through.
5967 bool shouldPromote(const Instruction *ToBePromoted) const {
5968 // Promote only if all the operands can be statically expanded.
5969 // Indeed, we do not want to introduce any new kind of transitions.
5970 for (const Use &U : ToBePromoted->operands()) {
5971 const Value *Val = U.get();
5972 if (Val == getEndOfTransition()) {
5973 // If the use is a division and the transition is on the rhs,
5974 // we cannot promote the operation, otherwise we may create a
5975 // division by zero.
5976 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
5980 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
5981 !isa<ConstantFP>(Val))
5984 // Check that the resulting operation is legal.
5985 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
5988 return StressStoreExtract ||
5989 TLI.isOperationLegalOrCustom(
5990 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
5993 /// \brief Check whether or not \p Use can be combined
5994 /// with the transition.
5995 /// I.e., is it possible to do Use(Transition) => AnotherUse?
5996 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
5998 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
5999 void enqueueForPromotion(Instruction *ToBePromoted) {
6000 InstsToBePromoted.push_back(ToBePromoted);
6003 /// \brief Set the instruction that will be combined with the transition.
6004 void recordCombineInstruction(Instruction *ToBeCombined) {
6005 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
6006 CombineInst = ToBeCombined;
6009 /// \brief Promote all the instructions enqueued for promotion if it is
6011 /// \return True if the promotion happened, false otherwise.
6013 // Check if there is something to promote.
6014 // Right now, if we do not have anything to combine with,
6015 // we assume the promotion is not profitable.
6016 if (InstsToBePromoted.empty() || !CombineInst)
6020 if (!StressStoreExtract && !isProfitableToPromote())
6024 for (auto &ToBePromoted : InstsToBePromoted)
6025 promoteImpl(ToBePromoted);
6026 InstsToBePromoted.clear();
6030 } // End of anonymous namespace.
6032 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
6033 // At this point, we know that all the operands of ToBePromoted but Def
6034 // can be statically promoted.
6035 // For Def, we need to use its parameter in ToBePromoted:
6036 // b = ToBePromoted ty1 a
6037 // Def = Transition ty1 b to ty2
6038 // Move the transition down.
6039 // 1. Replace all uses of the promoted operation by the transition.
6040 // = ... b => = ... Def.
6041 assert(ToBePromoted->getType() == Transition->getType() &&
6042 "The type of the result of the transition does not match "
6044 ToBePromoted->replaceAllUsesWith(Transition);
6045 // 2. Update the type of the uses.
6046 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
6047 Type *TransitionTy = getTransitionType();
6048 ToBePromoted->mutateType(TransitionTy);
6049 // 3. Update all the operands of the promoted operation with promoted
6051 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
6052 for (Use &U : ToBePromoted->operands()) {
6053 Value *Val = U.get();
6054 Value *NewVal = nullptr;
6055 if (Val == Transition)
6056 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
6057 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
6058 isa<ConstantFP>(Val)) {
6059 // Use a splat constant if it is not safe to use undef.
6060 NewVal = getConstantVector(
6061 cast<Constant>(Val),
6062 isa<UndefValue>(Val) ||
6063 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
6065 llvm_unreachable("Did you modified shouldPromote and forgot to update "
6067 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
6069 Transition->removeFromParent();
6070 Transition->insertAfter(ToBePromoted);
6071 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
6074 /// Some targets can do store(extractelement) with one instruction.
6075 /// Try to push the extractelement towards the stores when the target
6076 /// has this feature and this is profitable.
6077 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
6078 unsigned CombineCost = UINT_MAX;
6079 if (DisableStoreExtract || !TLI ||
6080 (!StressStoreExtract &&
6081 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
6082 Inst->getOperand(1), CombineCost)))
6085 // At this point we know that Inst is a vector to scalar transition.
6086 // Try to move it down the def-use chain, until:
6087 // - We can combine the transition with its single use
6088 // => we got rid of the transition.
6089 // - We escape the current basic block
6090 // => we would need to check that we are moving it at a cheaper place and
6091 // we do not do that for now.
6092 BasicBlock *Parent = Inst->getParent();
6093 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
6094 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
6095 // If the transition has more than one use, assume this is not going to be
6097 while (Inst->hasOneUse()) {
6098 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
6099 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
6101 if (ToBePromoted->getParent() != Parent) {
6102 DEBUG(dbgs() << "Instruction to promote is in a different block ("
6103 << ToBePromoted->getParent()->getName()
6104 << ") than the transition (" << Parent->getName() << ").\n");
6108 if (VPH.canCombine(ToBePromoted)) {
6109 DEBUG(dbgs() << "Assume " << *Inst << '\n'
6110 << "will be combined with: " << *ToBePromoted << '\n');
6111 VPH.recordCombineInstruction(ToBePromoted);
6112 bool Changed = VPH.promote();
6113 NumStoreExtractExposed += Changed;
6117 DEBUG(dbgs() << "Try promoting.\n");
6118 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
6121 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
6123 VPH.enqueueForPromotion(ToBePromoted);
6124 Inst = ToBePromoted;
6129 bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
6130 // Bail out if we inserted the instruction to prevent optimizations from
6131 // stepping on each other's toes.
6132 if (InsertedInsts.count(I))
6135 if (PHINode *P = dyn_cast<PHINode>(I)) {
6136 // It is possible for very late stage optimizations (such as SimplifyCFG)
6137 // to introduce PHI nodes too late to be cleaned up. If we detect such a
6138 // trivial PHI, go ahead and zap it here.
6139 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
6140 P->replaceAllUsesWith(V);
6141 P->eraseFromParent();
6148 if (CastInst *CI = dyn_cast<CastInst>(I)) {
6149 // If the source of the cast is a constant, then this should have
6150 // already been constant folded. The only reason NOT to constant fold
6151 // it is if something (e.g. LSR) was careful to place the constant
6152 // evaluation in a block other than then one that uses it (e.g. to hoist
6153 // the address of globals out of a loop). If this is the case, we don't
6154 // want to forward-subst the cast.
6155 if (isa<Constant>(CI->getOperand(0)))
6158 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
6161 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6162 /// Sink a zext or sext into its user blocks if the target type doesn't
6163 /// fit in one register
6165 TLI->getTypeAction(CI->getContext(),
6166 TLI->getValueType(*DL, CI->getType())) ==
6167 TargetLowering::TypeExpandInteger) {
6168 return SinkCast(CI);
6170 bool MadeChange = moveExtToFormExtLoad(I);
6171 return MadeChange | optimizeExtUses(I);
6177 if (CmpInst *CI = dyn_cast<CmpInst>(I))
6178 if (!TLI || !TLI->hasMultipleConditionRegisters())
6179 return OptimizeCmpExpression(CI);
6181 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
6182 stripInvariantGroupMetadata(*LI);
6184 bool Modified = optimizeLoadExt(LI);
6185 unsigned AS = LI->getPointerAddressSpace();
6186 Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
6192 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
6193 stripInvariantGroupMetadata(*SI);
6195 unsigned AS = SI->getPointerAddressSpace();
6196 return optimizeMemoryInst(I, SI->getOperand(1),
6197 SI->getOperand(0)->getType(), AS);
6202 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
6204 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
6205 BinOp->getOpcode() == Instruction::LShr)) {
6206 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
6207 if (TLI && CI && TLI->hasExtractBitsInsn())
6208 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
6213 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
6214 if (GEPI->hasAllZeroIndices()) {
6215 /// The GEP operand must be a pointer, so must its result -> BitCast
6216 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
6217 GEPI->getName(), GEPI);
6218 GEPI->replaceAllUsesWith(NC);
6219 GEPI->eraseFromParent();
6221 optimizeInst(NC, ModifiedDT);
6227 if (CallInst *CI = dyn_cast<CallInst>(I))
6228 return optimizeCallInst(CI, ModifiedDT);
6230 if (SelectInst *SI = dyn_cast<SelectInst>(I))
6231 return optimizeSelectInst(SI);
6233 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
6234 return optimizeShuffleVectorInst(SVI);
6236 if (auto *Switch = dyn_cast<SwitchInst>(I))
6237 return optimizeSwitchInst(Switch);
6239 if (isa<ExtractElementInst>(I))
6240 return optimizeExtractElementInst(I);
6245 /// Given an OR instruction, check to see if this is a bitreverse
6246 /// idiom. If so, insert the new intrinsic and return true.
6247 static bool makeBitReverse(Instruction &I, const DataLayout &DL,
6248 const TargetLowering &TLI) {
6249 if (!I.getType()->isIntegerTy() ||
6250 !TLI.isOperationLegalOrCustom(ISD::BITREVERSE,
6251 TLI.getValueType(DL, I.getType(), true)))
6254 SmallVector<Instruction*, 4> Insts;
6255 if (!recognizeBitReverseOrBSwapIdiom(&I, false, true, Insts))
6257 Instruction *LastInst = Insts.back();
6258 I.replaceAllUsesWith(LastInst);
6259 RecursivelyDeleteTriviallyDeadInstructions(&I);
6263 // In this pass we look for GEP and cast instructions that are used
6264 // across basic blocks and rewrite them to improve basic-block-at-a-time
6266 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
6268 bool MadeChange = false;
6270 CurInstIterator = BB.begin();
6271 while (CurInstIterator != BB.end()) {
6272 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
6277 bool MadeBitReverse = true;
6278 while (TLI && MadeBitReverse) {
6279 MadeBitReverse = false;
6280 for (auto &I : reverse(BB)) {
6281 if (makeBitReverse(I, *DL, *TLI)) {
6282 MadeBitReverse = MadeChange = true;
6287 MadeChange |= dupRetToEnableTailCallOpts(&BB);
6292 // llvm.dbg.value is far away from the value then iSel may not be able
6293 // handle it properly. iSel will drop llvm.dbg.value if it can not
6294 // find a node corresponding to the value.
6295 bool CodeGenPrepare::placeDbgValues(Function &F) {
6296 bool MadeChange = false;
6297 for (BasicBlock &BB : F) {
6298 Instruction *PrevNonDbgInst = nullptr;
6299 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
6300 Instruction *Insn = &*BI++;
6301 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
6302 // Leave dbg.values that refer to an alloca alone. These
6303 // instrinsics describe the address of a variable (= the alloca)
6304 // being taken. They should not be moved next to the alloca
6305 // (and to the beginning of the scope), but rather stay close to
6306 // where said address is used.
6307 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
6308 PrevNonDbgInst = Insn;
6312 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
6313 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
6314 // If VI is a phi in a block with an EHPad terminator, we can't insert
6316 if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
6318 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
6319 DVI->removeFromParent();
6320 if (isa<PHINode>(VI))
6321 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
6323 DVI->insertAfter(VI);
6332 // If there is a sequence that branches based on comparing a single bit
6333 // against zero that can be combined into a single instruction, and the
6334 // target supports folding these into a single instruction, sink the
6335 // mask and compare into the branch uses. Do this before OptimizeBlock ->
6336 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
6338 bool CodeGenPrepare::sinkAndCmp(Function &F) {
6339 if (!EnableAndCmpSinking)
6341 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
6343 bool MadeChange = false;
6344 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
6345 BasicBlock *BB = &*I++;
6347 // Does this BB end with the following?
6348 // %andVal = and %val, #single-bit-set
6349 // %icmpVal = icmp %andResult, 0
6350 // br i1 %cmpVal label %dest1, label %dest2"
6351 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
6352 if (!Brcc || !Brcc->isConditional())
6354 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
6355 if (!Cmp || Cmp->getParent() != BB)
6357 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
6358 if (!Zero || !Zero->isZero())
6360 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
6361 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
6363 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
6364 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
6366 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
6368 // Push the "and; icmp" for any users that are conditional branches.
6369 // Since there can only be one branch use per BB, we don't need to keep
6370 // track of which BBs we insert into.
6371 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
6375 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
6377 if (!BrccUser || !BrccUser->isConditional())
6379 BasicBlock *UserBB = BrccUser->getParent();
6380 if (UserBB == BB) continue;
6381 DEBUG(dbgs() << "found Brcc use\n");
6383 // Sink the "and; icmp" to use.
6385 BinaryOperator *NewAnd =
6386 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
6389 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
6393 DEBUG(BrccUser->getParent()->dump());
6399 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
6400 /// success, or returns false if no or invalid metadata was found.
6401 static bool extractBranchMetadata(BranchInst *BI,
6402 uint64_t &ProbTrue, uint64_t &ProbFalse) {
6403 assert(BI->isConditional() &&
6404 "Looking for probabilities on unconditional branch?");
6405 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
6406 if (!ProfileData || ProfileData->getNumOperands() != 3)
6409 const auto *CITrue =
6410 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
6411 const auto *CIFalse =
6412 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
6413 if (!CITrue || !CIFalse)
6416 ProbTrue = CITrue->getValue().getZExtValue();
6417 ProbFalse = CIFalse->getValue().getZExtValue();
6422 /// \brief Scale down both weights to fit into uint32_t.
6423 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
6424 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
6425 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
6426 NewTrue = NewTrue / Scale;
6427 NewFalse = NewFalse / Scale;
6430 /// \brief Some targets prefer to split a conditional branch like:
6432 /// %0 = icmp ne i32 %a, 0
6433 /// %1 = icmp ne i32 %b, 0
6434 /// %or.cond = or i1 %0, %1
6435 /// br i1 %or.cond, label %TrueBB, label %FalseBB
6437 /// into multiple branch instructions like:
6440 /// %0 = icmp ne i32 %a, 0
6441 /// br i1 %0, label %TrueBB, label %bb2
6443 /// %1 = icmp ne i32 %b, 0
6444 /// br i1 %1, label %TrueBB, label %FalseBB
6446 /// This usually allows instruction selection to do even further optimizations
6447 /// and combine the compare with the branch instruction. Currently this is
6448 /// applied for targets which have "cheap" jump instructions.
6450 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
6452 bool CodeGenPrepare::splitBranchCondition(Function &F) {
6453 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
6456 bool MadeChange = false;
6457 for (auto &BB : F) {
6458 // Does this BB end with the following?
6459 // %cond1 = icmp|fcmp|binary instruction ...
6460 // %cond2 = icmp|fcmp|binary instruction ...
6461 // %cond.or = or|and i1 %cond1, cond2
6462 // br i1 %cond.or label %dest1, label %dest2"
6463 BinaryOperator *LogicOp;
6464 BasicBlock *TBB, *FBB;
6465 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
6468 auto *Br1 = cast<BranchInst>(BB.getTerminator());
6469 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
6473 Value *Cond1, *Cond2;
6474 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
6475 m_OneUse(m_Value(Cond2)))))
6476 Opc = Instruction::And;
6477 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
6478 m_OneUse(m_Value(Cond2)))))
6479 Opc = Instruction::Or;
6483 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
6484 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
6487 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
6490 auto *InsertBefore = std::next(Function::iterator(BB))
6491 .getNodePtrUnchecked();
6492 auto TmpBB = BasicBlock::Create(BB.getContext(),
6493 BB.getName() + ".cond.split",
6494 BB.getParent(), InsertBefore);
6496 // Update original basic block by using the first condition directly by the
6497 // branch instruction and removing the no longer needed and/or instruction.
6498 Br1->setCondition(Cond1);
6499 LogicOp->eraseFromParent();
6501 // Depending on the conditon we have to either replace the true or the false
6502 // successor of the original branch instruction.
6503 if (Opc == Instruction::And)
6504 Br1->setSuccessor(0, TmpBB);
6506 Br1->setSuccessor(1, TmpBB);
6508 // Fill in the new basic block.
6509 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
6510 if (auto *I = dyn_cast<Instruction>(Cond2)) {
6511 I->removeFromParent();
6512 I->insertBefore(Br2);
6515 // Update PHI nodes in both successors. The original BB needs to be
6516 // replaced in one succesor's PHI nodes, because the branch comes now from
6517 // the newly generated BB (NewBB). In the other successor we need to add one
6518 // incoming edge to the PHI nodes, because both branch instructions target
6519 // now the same successor. Depending on the original branch condition
6520 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
6521 // we perfrom the correct update for the PHI nodes.
6522 // This doesn't change the successor order of the just created branch
6523 // instruction (or any other instruction).
6524 if (Opc == Instruction::Or)
6525 std::swap(TBB, FBB);
6527 // Replace the old BB with the new BB.
6528 for (auto &I : *TBB) {
6529 PHINode *PN = dyn_cast<PHINode>(&I);
6533 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
6534 PN->setIncomingBlock(i, TmpBB);
6537 // Add another incoming edge form the new BB.
6538 for (auto &I : *FBB) {
6539 PHINode *PN = dyn_cast<PHINode>(&I);
6542 auto *Val = PN->getIncomingValueForBlock(&BB);
6543 PN->addIncoming(Val, TmpBB);
6546 // Update the branch weights (from SelectionDAGBuilder::
6547 // FindMergedConditions).
6548 if (Opc == Instruction::Or) {
6549 // Codegen X | Y as:
6558 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
6559 // The requirement is that
6560 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
6561 // = TrueProb for orignal BB.
6562 // Assuming the orignal weights are A and B, one choice is to set BB1's
6563 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
6565 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
6566 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
6567 // TmpBB, but the math is more complicated.
6568 uint64_t TrueWeight, FalseWeight;
6569 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
6570 uint64_t NewTrueWeight = TrueWeight;
6571 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
6572 scaleWeights(NewTrueWeight, NewFalseWeight);
6573 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6574 .createBranchWeights(TrueWeight, FalseWeight));
6576 NewTrueWeight = TrueWeight;
6577 NewFalseWeight = 2 * FalseWeight;
6578 scaleWeights(NewTrueWeight, NewFalseWeight);
6579 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6580 .createBranchWeights(TrueWeight, FalseWeight));
6583 // Codegen X & Y as:
6591 // This requires creation of TmpBB after CurBB.
6593 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
6594 // The requirement is that
6595 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
6596 // = FalseProb for orignal BB.
6597 // Assuming the orignal weights are A and B, one choice is to set BB1's
6598 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
6600 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
6601 uint64_t TrueWeight, FalseWeight;
6602 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
6603 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
6604 uint64_t NewFalseWeight = FalseWeight;
6605 scaleWeights(NewTrueWeight, NewFalseWeight);
6606 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6607 .createBranchWeights(TrueWeight, FalseWeight));
6609 NewTrueWeight = 2 * TrueWeight;
6610 NewFalseWeight = FalseWeight;
6611 scaleWeights(NewTrueWeight, NewFalseWeight);
6612 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6613 .createBranchWeights(TrueWeight, FalseWeight));
6617 // Note: No point in getting fancy here, since the DT info is never
6618 // available to CodeGenPrepare.
6623 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
6629 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
6630 if (auto *InvariantMD = I.getMetadata(LLVMContext::MD_invariant_group))
6631 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());