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;
158 static char ID; // Pass identification, replacement for typeid
159 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
160 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr), DL(nullptr) {
161 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
163 bool runOnFunction(Function &F) override;
165 const char *getPassName() const override { return "CodeGen Prepare"; }
167 void getAnalysisUsage(AnalysisUsage &AU) const override {
168 AU.addPreserved<DominatorTreeWrapperPass>();
169 AU.addRequired<TargetLibraryInfoWrapperPass>();
170 AU.addRequired<TargetTransformInfoWrapperPass>();
174 bool eliminateFallThrough(Function &F);
175 bool eliminateMostlyEmptyBlocks(Function &F);
176 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
177 void eliminateMostlyEmptyBlock(BasicBlock *BB);
178 bool optimizeBlock(BasicBlock &BB, bool& ModifiedDT);
179 bool optimizeInst(Instruction *I, bool& ModifiedDT);
180 bool optimizeMemoryInst(Instruction *I, Value *Addr,
181 Type *AccessTy, unsigned AS);
182 bool optimizeInlineAsmInst(CallInst *CS);
183 bool optimizeCallInst(CallInst *CI, bool& ModifiedDT);
184 bool moveExtToFormExtLoad(Instruction *&I);
185 bool optimizeExtUses(Instruction *I);
186 bool optimizeLoadExt(LoadInst *I);
187 bool optimizeSelectInst(SelectInst *SI);
188 bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
189 bool optimizeSwitchInst(SwitchInst *CI);
190 bool optimizeExtractElementInst(Instruction *Inst);
191 bool dupRetToEnableTailCallOpts(BasicBlock *BB);
192 bool placeDbgValues(Function &F);
193 bool sinkAndCmp(Function &F);
194 bool extLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
196 const SmallVectorImpl<Instruction *> &Exts,
197 unsigned CreatedInstCost);
198 bool splitBranchCondition(Function &F);
199 bool simplifyOffsetableRelocate(Instruction &I);
200 void stripInvariantGroupMetadata(Instruction &I);
204 char CodeGenPrepare::ID = 0;
205 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
206 "Optimize for code generation", false, false)
208 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
209 return new CodeGenPrepare(TM);
214 bool StoreAddressDependOnValue(StoreInst* SI, Value* DepVal);
215 Value* GetUntaintedAddress(Value* CurrentAddress);
217 // The depth we trace down a variable to look for its dependence set.
218 const unsigned kDependenceDepth = 4;
220 // Recursively looks for variables that 'Val' depends on at the given depth
221 // 'Depth', and adds them in 'DepSet'. If 'InsertOnlyLeafNodes' is true, only
222 // inserts the leaf node values; otherwise, all visited nodes are included in
223 // 'DepSet'. Note that constants will be ignored.
224 template <typename SetType>
225 void recursivelyFindDependence(SetType* DepSet, Value* Val,
226 bool InsertOnlyLeafNodes = false,
227 unsigned Depth = kDependenceDepth) {
228 if (Val == nullptr) {
231 if (!InsertOnlyLeafNodes && !isa<Constant>(Val)) {
235 // Cannot go deeper. Insert the leaf nodes.
236 if (InsertOnlyLeafNodes && !isa<Constant>(Val)) {
242 // Go one step further to explore the dependence of the operands.
243 Instruction* I = nullptr;
244 if ((I = dyn_cast<Instruction>(Val))) {
245 if (isa<LoadInst>(I)) {
246 // A load is considerd the leaf load of the dependence tree. Done.
249 } else if (I->isBinaryOp()) {
250 BinaryOperator* I = dyn_cast<BinaryOperator>(Val);
251 Value *Op0 = I->getOperand(0), *Op1 = I->getOperand(1);
252 recursivelyFindDependence(DepSet, Op0, Depth - 1);
253 recursivelyFindDependence(DepSet, Op1, Depth - 1);
254 } else if (I->isCast()) {
255 Value* Op0 = I->getOperand(0);
256 recursivelyFindDependence(DepSet, Op0, Depth - 1);
257 } else if (I->getOpcode() == Instruction::Select) {
258 Value* Op0 = I->getOperand(0);
259 Value* Op1 = I->getOperand(1);
260 Value* Op2 = I->getOperand(2);
261 recursivelyFindDependence(DepSet, Op0, Depth - 1);
262 recursivelyFindDependence(DepSet, Op1, Depth - 1);
263 recursivelyFindDependence(DepSet, Op2, Depth - 1);
264 } else if (I->getOpcode() == Instruction::GetElementPtr) {
265 for (unsigned i = 0; i < I->getNumOperands(); i++) {
266 recursivelyFindDependence(DepSet, I->getOperand(i), Depth - 1);
268 } else if (I->getOpcode() == Instruction::Store) {
269 auto* SI = dyn_cast<StoreInst>(Val);
270 recursivelyFindDependence(DepSet, SI->getPointerOperand(), Depth - 1);
271 recursivelyFindDependence(DepSet, SI->getValueOperand(), Depth - 1);
273 Value* Op0 = nullptr;
274 Value* Op1 = nullptr;
275 switch (I->getOpcode()) {
276 case Instruction::ICmp:
277 case Instruction::FCmp: {
278 Op0 = I->getOperand(0);
279 Op1 = I->getOperand(1);
280 recursivelyFindDependence(DepSet, Op0, Depth - 1);
281 recursivelyFindDependence(DepSet, Op1, Depth - 1);
285 // Be conservative. Add it and be done with it.
291 } else if (isa<Constant>(Val)) {
292 // Not interested in constant values. Done.
295 // Be conservative. Add it and be done with it.
301 // Helper function to create a Cast instruction.
302 Value* createCast(IRBuilder<true, NoFolder>& Builder, Value* DepVal,
303 Type* TargetIntegerType) {
304 Instruction::CastOps CastOp = Instruction::BitCast;
305 switch (DepVal->getType()->getTypeID()) {
306 case Type::IntegerTyID: {
307 CastOp = Instruction::SExt;
310 case Type::FloatTyID:
311 case Type::DoubleTyID: {
312 CastOp = Instruction::FPToSI;
315 case Type::PointerTyID: {
316 CastOp = Instruction::PtrToInt;
322 return Builder.CreateCast(CastOp, DepVal, TargetIntegerType);
325 // Given a value, if it's a tainted address, this function returns the
326 // instruction that ORs the "dependence value" with the "original address".
327 // Otherwise, returns nullptr. This instruction is the first OR instruction
328 // where one of its operand is an AND instruction with an operand being 0.
330 // E.g., it returns '%4 = or i32 %3, %2' given 'CurrentAddress' is '%5'.
331 // %0 = load i32, i32* @y, align 4, !tbaa !1
332 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
333 // %1 = sext i1 %cmp to i32
334 // %2 = ptrtoint i32* @x to i32
335 // %3 = and i32 %1, 0
336 // %4 = or i32 %3, %2
337 // %5 = inttoptr i32 %4 to i32*
338 // store i32 1, i32* %5, align 4
339 Instruction* getOrAddress(Value* CurrentAddress) {
340 // Is it a cast from integer to pointer type.
341 Instruction* OrAddress = nullptr;
342 Instruction* AndDep = nullptr;
343 Instruction* CastToInt = nullptr;
344 Value* ActualAddress = nullptr;
345 Constant* ZeroConst = nullptr;
347 const Instruction* CastToPtr = dyn_cast<Instruction>(CurrentAddress);
348 if (CastToPtr && CastToPtr->getOpcode() == Instruction::IntToPtr) {
349 // Is it an OR instruction: %1 = or %and, %actualAddress.
350 if ((OrAddress = dyn_cast<Instruction>(CastToPtr->getOperand(0))) &&
351 OrAddress->getOpcode() == Instruction::Or) {
352 // The first operand should be and AND instruction.
353 AndDep = dyn_cast<Instruction>(OrAddress->getOperand(0));
354 if (AndDep && AndDep->getOpcode() == Instruction::And) {
355 // Also make sure its first operand of the "AND" is 0, or the "AND" is
356 // marked explicitly by "NoInstCombine".
357 if ((ZeroConst = dyn_cast<Constant>(AndDep->getOperand(1))) &&
358 ZeroConst->isNullValue()) {
364 // Looks like it's not been tainted.
368 // Given a value, if it's a tainted address, this function returns the
369 // instruction that taints the "dependence value". Otherwise, returns nullptr.
370 // This instruction is the last AND instruction where one of its operand is 0.
371 // E.g., it returns '%3' given 'CurrentAddress' is '%5'.
372 // %0 = load i32, i32* @y, align 4, !tbaa !1
373 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
374 // %1 = sext i1 %cmp to i32
375 // %2 = ptrtoint i32* @x to i32
376 // %3 = and i32 %1, 0
377 // %4 = or i32 %3, %2
378 // %5 = inttoptr i32 %4 to i32*
379 // store i32 1, i32* %5, align 4
380 Instruction* getAndDependence(Value* CurrentAddress) {
381 // If 'CurrentAddress' is tainted, get the OR instruction.
382 auto* OrAddress = getOrAddress(CurrentAddress);
383 if (OrAddress == nullptr) {
387 // No need to check the operands.
388 auto* AndDepInst = dyn_cast<Instruction>(OrAddress->getOperand(0));
393 // Given a value, if it's a tainted address, this function returns
394 // the "dependence value", which is the first operand in the AND instruction.
395 // E.g., it returns '%1' 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 Value* getDependence(Value* CurrentAddress) {
405 auto* AndInst = getAndDependence(CurrentAddress);
406 if (AndInst == nullptr) {
409 return AndInst->getOperand(0);
412 // Given an address that has been tainted, returns the only condition it depends
413 // on, if any; otherwise, returns nullptr.
414 Value* getConditionDependence(Value* Address) {
415 auto* Dep = getDependence(Address);
416 if (Dep == nullptr) {
417 // 'Address' has not been dependence-tainted.
421 Value* Operand = Dep;
423 auto* Inst = dyn_cast<Instruction>(Operand);
424 if (Inst == nullptr) {
425 // Non-instruction type does not have condition dependence.
428 if (Inst->getOpcode() == Instruction::ICmp) {
431 if (Inst->getNumOperands() != 1) {
434 Operand = Inst->getOperand(0);
440 // Conservatively decides whether the dependence set of 'Val1' includes the
441 // dependence set of 'Val2'. If 'ExpandSecondValue' is false, we do not expand
442 // 'Val2' and use that single value as its dependence set.
443 // If it returns true, it means the dependence set of 'Val1' includes that of
444 // 'Val2'; otherwise, it only means we cannot conclusively decide it.
445 bool dependenceSetInclusion(Value* Val1, Value* Val2,
446 int Val1ExpandLevel = 2 * kDependenceDepth,
447 int Val2ExpandLevel = kDependenceDepth) {
448 typedef SmallSet<Value*, 8> IncludingSet;
449 typedef SmallSet<Value*, 4> IncludedSet;
451 IncludingSet DepSet1;
453 // Look for more depths for the including set.
454 recursivelyFindDependence(&DepSet1, Val1, false /*Insert all visited nodes*/,
456 recursivelyFindDependence(&DepSet2, Val2, true /*Only insert leaf nodes*/,
459 auto set_inclusion = [](IncludingSet FullSet, IncludedSet Subset) {
460 for (auto* Dep : Subset) {
461 if (0 == FullSet.count(Dep)) {
467 bool inclusion = set_inclusion(DepSet1, DepSet2);
468 DEBUG(dbgs() << "[dependenceSetInclusion]: " << inclusion << "\n");
469 DEBUG(dbgs() << "Including set for: " << *Val1 << "\n");
470 DEBUG(for (const auto* Dep : DepSet1) { dbgs() << "\t\t" << *Dep << "\n"; });
471 DEBUG(dbgs() << "Included set for: " << *Val2 << "\n");
472 DEBUG(for (const auto* Dep : DepSet2) { dbgs() << "\t\t" << *Dep << "\n"; });
477 // Recursively iterates through the operands spawned from 'DepVal'. If there
478 // exists a single value that 'DepVal' only depends on, we call that value the
479 // root dependence of 'DepVal' and return it. Otherwise, return 'DepVal'.
480 Value* getRootDependence(Value* DepVal) {
481 SmallSet<Value*, 8> DepSet;
482 for (unsigned depth = kDependenceDepth; depth > 0; --depth) {
483 recursivelyFindDependence(&DepSet, DepVal, true /*Only insert leaf nodes*/,
485 if (DepSet.size() == 1) {
486 return *DepSet.begin();
493 // This function actually taints 'DepVal' to the address to 'SI'. If the
495 // of 'SI' already depends on whatever 'DepVal' depends on, this function
496 // doesn't do anything and returns false. Otherwise, returns true.
498 // This effect forces the store and any stores that comes later to depend on
499 // 'DepVal'. For example, we have a condition "cond", and a store instruction
500 // "s: STORE addr, val". If we want "s" (and any later store) to depend on
501 // "cond", we do the following:
502 // %conv = sext i1 %cond to i32
503 // %addrVal = ptrtoint i32* %addr to i32
504 // %andCond = and i32 conv, 0;
505 // %orAddr = or i32 %andCond, %addrVal;
506 // %NewAddr = inttoptr i32 %orAddr to i32*;
508 // This is a more concrete example:
510 // %0 = load i32, i32* @y, align 4, !tbaa !1
511 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
512 // %1 = sext i1 %cmp to i32
513 // %2 = ptrtoint i32* @x to i32
514 // %3 = and i32 %1, 0
515 // %4 = or i32 %3, %2
516 // %5 = inttoptr i32 %4 to i32*
517 // store i32 1, i32* %5, align 4
518 bool taintStoreAddress(StoreInst* SI, Value* DepVal,
519 const char* calling_func = __builtin_FUNCTION()) {
520 DEBUG(dbgs() << "Called from " << calling_func << '\n');
521 // Set the insertion point right after the 'DepVal'.
522 Instruction* Inst = nullptr;
523 IRBuilder<true, NoFolder> Builder(SI);
524 BasicBlock* BB = SI->getParent();
525 Value* Address = SI->getPointerOperand();
526 Type* TargetIntegerType =
527 IntegerType::get(Address->getContext(),
528 BB->getModule()->getDataLayout().getPointerSizeInBits());
530 // Does SI's address already depends on whatever 'DepVal' depends on?
531 if (StoreAddressDependOnValue(SI, DepVal)) {
535 // Figure out if there's a root variable 'DepVal' depends on. For example, we
536 // can extract "getelementptr inbounds %struct, %struct* %0, i64 0, i32 123"
537 // to be "%struct* %0" since all other operands are constant.
538 DepVal = getRootDependence(DepVal);
540 // Is this already a dependence-tainted store?
541 Value* OldDep = getDependence(Address);
543 // The address of 'SI' has already been tainted. Just need to absorb the
544 // DepVal to the existing dependence in the address of SI.
545 Instruction* AndDep = getAndDependence(Address);
546 IRBuilder<true, NoFolder> Builder(AndDep);
547 Value* NewDep = nullptr;
548 if (DepVal->getType() == AndDep->getType()) {
549 NewDep = Builder.CreateAnd(OldDep, DepVal);
551 NewDep = Builder.CreateAnd(
552 OldDep, createCast(Builder, DepVal, TargetIntegerType));
555 auto* NewDepInst = dyn_cast<Instruction>(NewDep);
557 // Use the new AND instruction as the dependence
558 AndDep->setOperand(0, NewDep);
562 // SI's address has not been tainted. Now taint it with 'DepVal'.
563 Value* CastDepToInt = createCast(Builder, DepVal, TargetIntegerType);
564 Value* PtrToIntCast = Builder.CreatePtrToInt(Address, TargetIntegerType);
566 Builder.CreateAnd(CastDepToInt, ConstantInt::get(TargetIntegerType, 0));
567 auto AndInst = dyn_cast<Instruction>(AndDepVal);
568 // XXX-comment: The original IR InstCombiner would change our and instruction
569 // to a select and then the back end optimize the condition out. We attach a
570 // flag to instructions and set it here to inform the InstCombiner to not to
571 // touch this and instruction at all.
572 Value* OrAddr = Builder.CreateOr(AndDepVal, PtrToIntCast);
573 Value* NewAddr = Builder.CreateIntToPtr(OrAddr, Address->getType());
575 DEBUG(dbgs() << "[taintStoreAddress]\n"
576 << "Original store: " << *SI << '\n');
577 SI->setOperand(1, NewAddr);
580 DEBUG(dbgs() << "\tTargetIntegerType: " << *TargetIntegerType << '\n'
581 << "\tCast dependence value to integer: " << *CastDepToInt
583 << "\tCast address to integer: " << *PtrToIntCast << '\n'
584 << "\tAnd dependence value: " << *AndDepVal << '\n'
585 << "\tOr address: " << *OrAddr << '\n'
586 << "\tCast or instruction to address: " << *NewAddr << "\n\n");
591 // Looks for the previous store in the if block --- 'BrBB', which makes the
592 // speculative store 'StoreToHoist' safe.
593 Value* getSpeculativeStoreInPrevBB(StoreInst* StoreToHoist, BasicBlock* BrBB) {
594 assert(StoreToHoist && "StoreToHoist must be a real store");
596 Value* StorePtr = StoreToHoist->getPointerOperand();
598 // Look for a store to the same pointer in BrBB.
599 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(), RE = BrBB->rend();
601 Instruction* CurI = &*RI;
603 StoreInst* SI = dyn_cast<StoreInst>(CurI);
604 // Found the previous store make sure it stores to the same location.
605 // XXX-update: If the previous store's original untainted address are the
606 // same as 'StorePtr', we are also good to hoist the store.
607 if (SI && (SI->getPointerOperand() == StorePtr ||
608 GetUntaintedAddress(SI->getPointerOperand()) == StorePtr)) {
609 // Found the previous store, return its value operand.
615 "We should not reach here since this store is safe to speculate");
618 // XXX-comment: Returns true if it changes the code, false otherwise (the branch
619 // condition already depends on 'DepVal'.
620 bool taintConditionalBranch(BranchInst* BI, Value* DepVal) {
621 assert(BI->isConditional());
622 auto* Cond = BI->getOperand(0);
623 if (dependenceSetInclusion(Cond, DepVal)) {
624 // The dependence/ordering is self-evident.
628 IRBuilder<true, NoFolder> Builder(BI);
630 Builder.CreateAnd(DepVal, ConstantInt::get(DepVal->getType(), 0));
632 Builder.CreateTrunc(AndDep, IntegerType::get(DepVal->getContext(), 1));
633 auto* OrCond = Builder.CreateOr(TruncAndDep, Cond);
634 BI->setOperand(0, OrCond);
637 DEBUG(dbgs() << "\tTainted branch condition:\n" << *BI->getParent());
642 bool ConditionalBranchDependsOnValue(BranchInst* BI, Value* DepVal) {
643 assert(BI->isConditional());
644 auto* Cond = BI->getOperand(0);
645 return dependenceSetInclusion(Cond, DepVal);
648 // XXX-update: For a relaxed load 'LI', find the first immediate atomic store or
649 // the first conditional branch. Returns nullptr if there's no such immediately
650 // following store/branch instructions, which we can only enforce the load with
652 Instruction* findFirstStoreCondBranchInst(LoadInst* LI) {
653 // In some situations, relaxed loads can be left as is:
654 // 1. The relaxed load is used to calculate the address of the immediate
656 // 2. The relaxed load is used as a condition in the immediate following
657 // condition, and there are no stores in between. This is actually quite
659 // int r1 = x.load(relaxed);
661 // y.store(1, relaxed);
663 // However, in this function, we don't deal with them directly. Instead, we
664 // just find the immediate following store/condition branch and return it.
666 auto* BB = LI->getParent();
668 auto BBI = BasicBlock::iterator(LI);
670 for (; BBI != BE; BBI++) {
671 auto* Inst = dyn_cast<Instruction>(&*BBI);
672 if (Inst == nullptr) {
675 if (Inst->getOpcode() == Instruction::Store) {
677 } else if (Inst->getOpcode() == Instruction::Br) {
678 auto* BrInst = dyn_cast<BranchInst>(Inst);
679 if (BrInst->isConditional()) {
689 // XXX-comment: Returns whether the code has been changed.
690 bool taintMonotonicLoads(const SmallVector<LoadInst*, 1>& MonotonicLoadInsts) {
691 bool Changed = false;
692 for (auto* LI : MonotonicLoadInsts) {
693 auto* FirstInst = findFirstStoreCondBranchInst(LI);
694 if (FirstInst == nullptr) {
695 // We don't seem to be able to taint a following store/conditional branch
696 // instruction. Simply make it acquire.
697 DEBUG(dbgs() << "[RelaxedLoad]: Transformed to acquire load\n"
699 LI->setOrdering(Acquire);
703 // Taint 'FirstInst', which could be a store or a condition branch
705 if (FirstInst->getOpcode() == Instruction::Store) {
706 Changed |= taintStoreAddress(dyn_cast<StoreInst>(FirstInst), LI);
707 } else if (FirstInst->getOpcode() == Instruction::Br) {
708 Changed |= taintConditionalBranch(dyn_cast<BranchInst>(FirstInst), LI);
710 assert(false && "findFirstStoreCondBranchInst() should return a "
711 "store/condition branch instruction");
717 // Inserts a fake conditional branch right after the instruction 'SplitInst',
718 // and the branch condition is 'Condition'. 'SplitInst' will be placed in the
719 // newly created block.
720 void AddFakeConditionalBranch(Instruction* SplitInst, Value* Condition) {
721 auto* BB = SplitInst->getParent();
722 TerminatorInst* ThenTerm = nullptr;
723 TerminatorInst* ElseTerm = nullptr;
724 SplitBlockAndInsertIfThenElse(Condition, SplitInst, &ThenTerm, &ElseTerm);
725 assert(ThenTerm && ElseTerm &&
726 "Then/Else terminators cannot be empty after basic block spliting");
727 auto* ThenBB = ThenTerm->getParent();
728 auto* ElseBB = ElseTerm->getParent();
729 auto* TailBB = ThenBB->getSingleSuccessor();
730 assert(TailBB && "Tail block cannot be empty after basic block spliting");
732 ThenBB->disableCanEliminateBlock();
733 ThenBB->disableCanEliminateBlock();
734 TailBB->disableCanEliminateBlock();
735 ThenBB->setName(BB->getName() + "Then.Fake");
736 ElseBB->setName(BB->getName() + "Else.Fake");
737 DEBUG(dbgs() << "Add fake conditional branch:\n"
739 << *ThenBB << "Else Block:\n"
743 // Returns true if the code is changed, and false otherwise.
744 void TaintRelaxedLoads(LoadInst* LI) {
745 // For better performance, we can add a "AND X 0" instruction before the
747 auto* FirstInst = findFirstStoreCondBranchInst(LI);
748 Instruction* InsertPoint = nullptr;
749 if (FirstInst == nullptr) {
750 InsertPoint = LI->getParent()->getTerminator();
751 InsertPoint = LI->getNextNode();
753 InsertPoint = LI->getNextNode();
755 IRBuilder<true, NoFolder> Builder(InsertPoint);
756 auto* AndZero = dyn_cast<Instruction>(
757 Builder.CreateAnd(LI, Constant::getNullValue(LI->getType())));
758 auto* FakeCondition = dyn_cast<Instruction>(Builder.CreateICmp(
759 CmpInst::ICMP_NE, AndZero, Constant::getNullValue(LI->getType())));
760 AddFakeConditionalBranch(FakeCondition->getNextNode(), FakeCondition);
763 // XXX-comment: Returns whether the code has been changed.
764 bool AddFakeConditionalBranchAfterMonotonicLoads(
765 const SmallVector<LoadInst*, 1>& MonotonicLoadInsts) {
766 bool Changed = false;
767 for (auto* LI : MonotonicLoadInsts) {
768 auto* FirstInst = findFirstStoreCondBranchInst(LI);
769 if (FirstInst != nullptr) {
770 if (FirstInst->getOpcode() == Instruction::Store) {
771 if (StoreAddressDependOnValue(dyn_cast<StoreInst>(FirstInst), LI)) {
774 } else if (FirstInst->getOpcode() == Instruction::Br &&
775 isa<BranchInst>(FirstInst)) {
776 if (ConditionalBranchDependsOnValue(dyn_cast<BranchInst>(FirstInst),
781 dbgs() << "FirstInst=" << *FirstInst << "\n";
782 assert(false && "findFirstStoreCondBranchInst() should return a "
783 "store/condition branch instruction");
787 // We really need to process the relaxed load now.
788 StoreInst* SI = nullptr;;
789 if (FirstInst && (SI = dyn_cast<StoreInst>(FirstInst))) {
790 // For immediately coming stores, taint the address of the store.
791 taintStoreAddress(SI, LI);
793 // For immediately coming branch, directly add a fake branch.
794 TaintRelaxedLoads(LI);
801 /**** Implementations of public methods for dependence tainting ****/
802 Value* GetUntaintedAddress(Value* CurrentAddress) {
803 auto* OrAddress = getOrAddress(CurrentAddress);
804 if (OrAddress == nullptr) {
805 // Is it tainted by a select instruction?
806 auto* Inst = dyn_cast<Instruction>(CurrentAddress);
807 if (nullptr != Inst && Inst->getOpcode() == Instruction::Select) {
808 // A selection instruction.
809 if (Inst->getOperand(1) == Inst->getOperand(2)) {
810 return Inst->getOperand(1);
814 return CurrentAddress;
816 Value* ActualAddress = nullptr;
818 auto* CastToInt = dyn_cast<Instruction>(OrAddress->getOperand(1));
819 if (CastToInt && CastToInt->getOpcode() == Instruction::PtrToInt) {
820 return CastToInt->getOperand(0);
822 // This should be a IntToPtr constant expression.
823 ConstantExpr* PtrToIntExpr =
824 dyn_cast<ConstantExpr>(OrAddress->getOperand(1));
825 if (PtrToIntExpr && PtrToIntExpr->getOpcode() == Instruction::PtrToInt) {
826 return PtrToIntExpr->getOperand(0);
830 // Looks like it's not been dependence-tainted. Returns itself.
831 return CurrentAddress;
834 MemoryLocation GetUntaintedMemoryLocation(StoreInst* SI) {
836 SI->getAAMetadata(AATags);
837 const auto& DL = SI->getModule()->getDataLayout();
838 const auto* OriginalAddr = GetUntaintedAddress(SI->getPointerOperand());
839 DEBUG(if (OriginalAddr != SI->getPointerOperand()) {
840 dbgs() << "[GetUntaintedMemoryLocation]\n"
841 << "Storing address: " << *SI->getPointerOperand()
842 << "\nUntainted address: " << *OriginalAddr << "\n";
844 return MemoryLocation(OriginalAddr,
845 DL.getTypeStoreSize(SI->getValueOperand()->getType()),
849 bool TaintDependenceToStore(StoreInst* SI, Value* DepVal) {
850 if (dependenceSetInclusion(SI, DepVal)) {
854 bool tainted = taintStoreAddress(SI, DepVal);
859 bool TaintDependenceToStoreAddress(StoreInst* SI, Value* DepVal) {
860 if (dependenceSetInclusion(SI->getPointerOperand(), DepVal)) {
864 bool tainted = taintStoreAddress(SI, DepVal);
869 bool CompressTaintedStore(BasicBlock* BB) {
870 // This function looks for windows of adajcent stores in 'BB' that satisfy the
871 // following condition (and then do optimization):
872 // *Addr(d1) = v1, d1 is a condition and is the only dependence the store's
873 // address depends on && Dep(v1) includes Dep(d1);
874 // *Addr(d2) = v2, d2 is a condition and is the only dependnece the store's
875 // address depends on && Dep(v2) includes Dep(d2) &&
876 // Dep(d2) includes Dep(d1);
878 // *Addr(dN) = vN, dN is a condition and is the only dependence the store's
879 // address depends on && Dep(dN) includes Dep(d"N-1").
881 // As a result, Dep(dN) includes [Dep(d1) V ... V Dep(d"N-1")], so we can
882 // safely transform the above to the following. In between these stores, we
883 // can omit untainted stores to the same address 'Addr' since they internally
884 // have dependence on the previous stores on the same address.
889 for (auto BI = BB->begin(), BE = BB->end(); BI != BE; BI++) {
890 // Look for the first store in such a window of adajacent stores.
891 auto* FirstSI = dyn_cast<StoreInst>(&*BI);
896 // The first store in the window must be tainted.
897 auto* UntaintedAddress = GetUntaintedAddress(FirstSI->getPointerOperand());
898 if (UntaintedAddress == FirstSI->getPointerOperand()) {
902 // The first store's address must directly depend on and only depend on a
904 auto* FirstSIDepCond = getConditionDependence(FirstSI->getPointerOperand());
905 if (nullptr == FirstSIDepCond) {
909 // Dep(first store's storing value) includes Dep(tainted dependence).
910 if (!dependenceSetInclusion(FirstSI->getValueOperand(), FirstSIDepCond)) {
914 // Look for subsequent stores to the same address that satisfy the condition
915 // of "compressing the dependence".
916 SmallVector<StoreInst*, 8> AdajacentStores;
917 AdajacentStores.push_back(FirstSI);
918 auto BII = BasicBlock::iterator(FirstSI);
919 for (BII++; BII != BE; BII++) {
920 auto* CurrSI = dyn_cast<StoreInst>(&*BII);
922 if (BII->mayHaveSideEffects()) {
923 // Be conservative. Instructions with side effects are similar to
930 auto* OrigAddress = GetUntaintedAddress(CurrSI->getPointerOperand());
931 auto* CurrSIDepCond = getConditionDependence(CurrSI->getPointerOperand());
932 // All other stores must satisfy either:
933 // A. 'CurrSI' is an untainted store to the same address, or
934 // B. the combination of the following 5 subconditions:
936 // 2. Untainted address is the same as the group's address;
937 // 3. The address is tainted with a sole value which is a condition;
938 // 4. The storing value depends on the condition in 3.
939 // 5. The condition in 3 depends on the previous stores dependence
942 // Condition A. Should ignore this store directly.
943 if (OrigAddress == CurrSI->getPointerOperand() &&
944 OrigAddress == UntaintedAddress) {
947 // Check condition B.
948 Value* Cond = nullptr;
949 if (OrigAddress == CurrSI->getPointerOperand() ||
950 OrigAddress != UntaintedAddress || CurrSIDepCond == nullptr ||
951 !dependenceSetInclusion(CurrSI->getValueOperand(), CurrSIDepCond)) {
952 // Check condition 1, 2, 3 & 4.
956 // Check condition 5.
957 StoreInst* PrevSI = AdajacentStores[AdajacentStores.size() - 1];
958 auto* PrevSIDepCond = getConditionDependence(PrevSI->getPointerOperand());
959 assert(PrevSIDepCond &&
960 "Store in the group must already depend on a condtion");
961 if (!dependenceSetInclusion(CurrSIDepCond, PrevSIDepCond)) {
965 AdajacentStores.push_back(CurrSI);
968 if (AdajacentStores.size() == 1) {
969 // The outer loop should keep looking from the next store.
973 // Now we have such a group of tainted stores to the same address.
974 DEBUG(dbgs() << "[CompressTaintedStore]\n");
975 DEBUG(dbgs() << "Original BB\n");
976 DEBUG(dbgs() << *BB << '\n');
977 auto* LastSI = AdajacentStores[AdajacentStores.size() - 1];
978 for (unsigned i = 0; i < AdajacentStores.size() - 1; ++i) {
979 auto* SI = AdajacentStores[i];
981 // Use the original address for stores before the last one.
982 SI->setOperand(1, UntaintedAddress);
984 DEBUG(dbgs() << "Store address has been reversed: " << *SI << '\n';);
986 // XXX-comment: Try to make the last store use fewer registers.
987 // If LastSI's storing value is a select based on the condition with which
988 // its address is tainted, transform the tainted address to a select
989 // instruction, as follows:
990 // r1 = Select Cond ? A : B
995 // r1 = Select Cond ? A : B
996 // r2 = Select Cond ? Addr : Addr
998 // The idea is that both Select instructions depend on the same condition,
999 // so hopefully the backend can generate two cmov instructions for them (and
1000 // this saves the number of registers needed).
1001 auto* LastSIDep = getConditionDependence(LastSI->getPointerOperand());
1002 auto* LastSIValue = dyn_cast<Instruction>(LastSI->getValueOperand());
1003 if (LastSIValue && LastSIValue->getOpcode() == Instruction::Select &&
1004 LastSIValue->getOperand(0) == LastSIDep) {
1005 // XXX-comment: Maybe it's better for us to just leave it as an and/or
1006 // dependence pattern.
1008 IRBuilder<true, NoFolder> Builder(LastSI);
1010 Builder.CreateSelect(LastSIDep, UntaintedAddress, UntaintedAddress);
1011 LastSI->setOperand(1, Address);
1012 DEBUG(dbgs() << "The last store becomes :" << *LastSI << "\n\n";);
1020 bool PassDependenceToStore(Value* OldAddress, StoreInst* NewStore) {
1021 Value* OldDep = getDependence(OldAddress);
1022 // Return false when there's no dependence to pass from the OldAddress.
1027 // No need to pass the dependence to NewStore's address if it already depends
1028 // on whatever 'OldAddress' depends on.
1029 if (StoreAddressDependOnValue(NewStore, OldDep)) {
1032 return taintStoreAddress(NewStore, OldAddress);
1035 SmallSet<Value*, 8> FindDependence(Value* Val) {
1036 SmallSet<Value*, 8> DepSet;
1037 recursivelyFindDependence(&DepSet, Val, true /*Only insert leaf nodes*/);
1041 bool StoreAddressDependOnValue(StoreInst* SI, Value* DepVal) {
1042 return dependenceSetInclusion(SI->getPointerOperand(), DepVal);
1045 bool StoreDependOnValue(StoreInst* SI, Value* Dep) {
1046 return dependenceSetInclusion(SI, Dep);
1053 bool CodeGenPrepare::runOnFunction(Function &F) {
1054 bool EverMadeChange = false;
1056 if (skipOptnoneFunction(F))
1059 DL = &F.getParent()->getDataLayout();
1061 // Clear per function information.
1062 InsertedInsts.clear();
1063 PromotedInsts.clear();
1067 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
1068 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1069 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1070 OptSize = F.optForSize();
1072 /// This optimization identifies DIV instructions that can be
1073 /// profitably bypassed and carried out with a shorter, faster divide.
1074 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
1075 const DenseMap<unsigned int, unsigned int> &BypassWidths =
1076 TLI->getBypassSlowDivWidths();
1077 BasicBlock* BB = &*F.begin();
1078 while (BB != nullptr) {
1079 // bypassSlowDivision may create new BBs, but we don't want to reapply the
1080 // optimization to those blocks.
1081 BasicBlock* Next = BB->getNextNode();
1082 EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
1087 // Eliminate blocks that contain only PHI nodes and an
1088 // unconditional branch.
1089 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
1091 // llvm.dbg.value is far away from the value then iSel may not be able
1092 // handle it properly. iSel will drop llvm.dbg.value if it can not
1093 // find a node corresponding to the value.
1094 EverMadeChange |= placeDbgValues(F);
1096 // If there is a mask, compare against zero, and branch that can be combined
1097 // into a single target instruction, push the mask and compare into branch
1098 // users. Do this before OptimizeBlock -> OptimizeInst ->
1099 // OptimizeCmpExpression, which perturbs the pattern being searched for.
1100 if (!DisableBranchOpts) {
1101 EverMadeChange |= sinkAndCmp(F);
1102 EverMadeChange |= splitBranchCondition(F);
1105 bool MadeChange = true;
1106 while (MadeChange) {
1108 for (Function::iterator I = F.begin(); I != F.end(); ) {
1109 BasicBlock *BB = &*I++;
1110 bool ModifiedDTOnIteration = false;
1111 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
1113 // Restart BB iteration if the dominator tree of the Function was changed
1114 if (ModifiedDTOnIteration)
1117 EverMadeChange |= MadeChange;
1122 if (!DisableBranchOpts) {
1124 SmallPtrSet<BasicBlock*, 8> WorkList;
1125 for (BasicBlock &BB : F) {
1126 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
1127 MadeChange |= ConstantFoldTerminator(&BB, true);
1128 if (!MadeChange) continue;
1130 for (SmallVectorImpl<BasicBlock*>::iterator
1131 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
1132 if (pred_begin(*II) == pred_end(*II))
1133 WorkList.insert(*II);
1136 // Delete the dead blocks and any of their dead successors.
1137 MadeChange |= !WorkList.empty();
1138 while (!WorkList.empty()) {
1139 BasicBlock *BB = *WorkList.begin();
1141 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
1143 DeleteDeadBlock(BB);
1145 for (SmallVectorImpl<BasicBlock*>::iterator
1146 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
1147 if (pred_begin(*II) == pred_end(*II))
1148 WorkList.insert(*II);
1151 // Merge pairs of basic blocks with unconditional branches, connected by
1153 if (EverMadeChange || MadeChange)
1154 MadeChange |= eliminateFallThrough(F);
1156 EverMadeChange |= MadeChange;
1159 if (!DisableGCOpts) {
1160 SmallVector<Instruction *, 2> Statepoints;
1161 for (BasicBlock &BB : F)
1162 for (Instruction &I : BB)
1163 if (isStatepoint(I))
1164 Statepoints.push_back(&I);
1165 for (auto &I : Statepoints)
1166 EverMadeChange |= simplifyOffsetableRelocate(*I);
1169 // XXX-comment: Delay dealing with relaxed loads in this function to avoid
1170 // further changes done by other passes (e.g., SimplifyCFG).
1171 // Collect all the relaxed loads.
1172 SmallVector<LoadInst*, 1> MonotonicLoadInsts;
1173 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
1174 if (I->isAtomic()) {
1175 switch (I->getOpcode()) {
1176 case Instruction::Load: {
1177 auto* LI = dyn_cast<LoadInst>(&*I);
1178 if (LI->getOrdering() == Monotonic) {
1179 MonotonicLoadInsts.push_back(LI);
1190 AddFakeConditionalBranchAfterMonotonicLoads(MonotonicLoadInsts);
1192 return EverMadeChange;
1195 /// Merge basic blocks which are connected by a single edge, where one of the
1196 /// basic blocks has a single successor pointing to the other basic block,
1197 /// which has a single predecessor.
1198 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
1199 bool Changed = false;
1200 // Scan all of the blocks in the function, except for the entry block.
1201 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
1202 BasicBlock *BB = &*I++;
1203 // If the destination block has a single pred, then this is a trivial
1204 // edge, just collapse it.
1205 BasicBlock *SinglePred = BB->getSinglePredecessor();
1207 // Don't merge if BB's address is taken.
1208 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
1210 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
1211 if (Term && !Term->isConditional()) {
1213 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
1214 // Remember if SinglePred was the entry block of the function.
1215 // If so, we will need to move BB back to the entry position.
1216 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
1217 MergeBasicBlockIntoOnlyPred(BB, nullptr);
1219 if (isEntry && BB != &BB->getParent()->getEntryBlock())
1220 BB->moveBefore(&BB->getParent()->getEntryBlock());
1222 // We have erased a block. Update the iterator.
1223 I = BB->getIterator();
1229 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
1230 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
1231 /// edges in ways that are non-optimal for isel. Start by eliminating these
1232 /// blocks so we can split them the way we want them.
1233 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
1234 bool MadeChange = false;
1235 // Note that this intentionally skips the entry block.
1236 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
1237 BasicBlock *BB = &*I++;
1238 // If this block doesn't end with an uncond branch, ignore it.
1239 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
1240 if (!BI || !BI->isUnconditional())
1243 // If the instruction before the branch (skipping debug info) isn't a phi
1244 // node, then other stuff is happening here.
1245 BasicBlock::iterator BBI = BI->getIterator();
1246 if (BBI != BB->begin()) {
1248 while (isa<DbgInfoIntrinsic>(BBI)) {
1249 if (BBI == BB->begin())
1253 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
1257 // Do not break infinite loops.
1258 BasicBlock *DestBB = BI->getSuccessor(0);
1262 if (!canMergeBlocks(BB, DestBB))
1265 eliminateMostlyEmptyBlock(BB);
1271 /// Return true if we can merge BB into DestBB if there is a single
1272 /// unconditional branch between them, and BB contains no other non-phi
1274 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
1275 const BasicBlock *DestBB) const {
1276 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
1277 // the successor. If there are more complex condition (e.g. preheaders),
1278 // don't mess around with them.
1279 BasicBlock::const_iterator BBI = BB->begin();
1280 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
1281 for (const User *U : PN->users()) {
1282 const Instruction *UI = cast<Instruction>(U);
1283 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
1285 // If User is inside DestBB block and it is a PHINode then check
1286 // incoming value. If incoming value is not from BB then this is
1287 // a complex condition (e.g. preheaders) we want to avoid here.
1288 if (UI->getParent() == DestBB) {
1289 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
1290 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
1291 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
1292 if (Insn && Insn->getParent() == BB &&
1293 Insn->getParent() != UPN->getIncomingBlock(I))
1300 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
1301 // and DestBB may have conflicting incoming values for the block. If so, we
1302 // can't merge the block.
1303 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
1304 if (!DestBBPN) return true; // no conflict.
1306 // Collect the preds of BB.
1307 SmallPtrSet<const BasicBlock*, 16> BBPreds;
1308 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
1309 // It is faster to get preds from a PHI than with pred_iterator.
1310 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1311 BBPreds.insert(BBPN->getIncomingBlock(i));
1313 BBPreds.insert(pred_begin(BB), pred_end(BB));
1316 // Walk the preds of DestBB.
1317 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
1318 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
1319 if (BBPreds.count(Pred)) { // Common predecessor?
1320 BBI = DestBB->begin();
1321 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
1322 const Value *V1 = PN->getIncomingValueForBlock(Pred);
1323 const Value *V2 = PN->getIncomingValueForBlock(BB);
1325 // If V2 is a phi node in BB, look up what the mapped value will be.
1326 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
1327 if (V2PN->getParent() == BB)
1328 V2 = V2PN->getIncomingValueForBlock(Pred);
1330 // If there is a conflict, bail out.
1331 if (V1 != V2) return false;
1340 /// Eliminate a basic block that has only phi's and an unconditional branch in
1342 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
1343 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1344 BasicBlock *DestBB = BI->getSuccessor(0);
1346 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
1348 // If the destination block has a single pred, then this is a trivial edge,
1349 // just collapse it.
1350 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
1351 if (SinglePred != DestBB) {
1352 // Remember if SinglePred was the entry block of the function. If so, we
1353 // will need to move BB back to the entry position.
1354 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
1355 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
1357 if (isEntry && BB != &BB->getParent()->getEntryBlock())
1358 BB->moveBefore(&BB->getParent()->getEntryBlock());
1360 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1365 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
1366 // to handle the new incoming edges it is about to have.
1368 for (BasicBlock::iterator BBI = DestBB->begin();
1369 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1370 // Remove the incoming value for BB, and remember it.
1371 Value *InVal = PN->removeIncomingValue(BB, false);
1373 // Two options: either the InVal is a phi node defined in BB or it is some
1374 // value that dominates BB.
1375 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
1376 if (InValPhi && InValPhi->getParent() == BB) {
1377 // Add all of the input values of the input PHI as inputs of this phi.
1378 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
1379 PN->addIncoming(InValPhi->getIncomingValue(i),
1380 InValPhi->getIncomingBlock(i));
1382 // Otherwise, add one instance of the dominating value for each edge that
1383 // we will be adding.
1384 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
1385 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1386 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
1388 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
1389 PN->addIncoming(InVal, *PI);
1394 // The PHIs are now updated, change everything that refers to BB to use
1395 // DestBB and remove BB.
1396 BB->replaceAllUsesWith(DestBB);
1397 BB->eraseFromParent();
1400 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1403 // Computes a map of base pointer relocation instructions to corresponding
1404 // derived pointer relocation instructions given a vector of all relocate calls
1405 static void computeBaseDerivedRelocateMap(
1406 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
1407 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
1409 // Collect information in two maps: one primarily for locating the base object
1410 // while filling the second map; the second map is the final structure holding
1411 // a mapping between Base and corresponding Derived relocate calls
1412 DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
1413 for (auto *ThisRelocate : AllRelocateCalls) {
1414 auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
1415 ThisRelocate->getDerivedPtrIndex());
1416 RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
1418 for (auto &Item : RelocateIdxMap) {
1419 std::pair<unsigned, unsigned> Key = Item.first;
1420 if (Key.first == Key.second)
1421 // Base relocation: nothing to insert
1424 GCRelocateInst *I = Item.second;
1425 auto BaseKey = std::make_pair(Key.first, Key.first);
1427 // We're iterating over RelocateIdxMap so we cannot modify it.
1428 auto MaybeBase = RelocateIdxMap.find(BaseKey);
1429 if (MaybeBase == RelocateIdxMap.end())
1430 // TODO: We might want to insert a new base object relocate and gep off
1431 // that, if there are enough derived object relocates.
1434 RelocateInstMap[MaybeBase->second].push_back(I);
1438 // Accepts a GEP and extracts the operands into a vector provided they're all
1439 // small integer constants
1440 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
1441 SmallVectorImpl<Value *> &OffsetV) {
1442 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
1443 // Only accept small constant integer operands
1444 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
1445 if (!Op || Op->getZExtValue() > 20)
1449 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
1450 OffsetV.push_back(GEP->getOperand(i));
1454 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
1455 // replace, computes a replacement, and affects it.
1457 simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
1458 const SmallVectorImpl<GCRelocateInst *> &Targets) {
1459 bool MadeChange = false;
1460 for (GCRelocateInst *ToReplace : Targets) {
1461 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
1462 "Not relocating a derived object of the original base object");
1463 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
1464 // A duplicate relocate call. TODO: coalesce duplicates.
1468 if (RelocatedBase->getParent() != ToReplace->getParent()) {
1469 // Base and derived relocates are in different basic blocks.
1470 // In this case transform is only valid when base dominates derived
1471 // relocate. However it would be too expensive to check dominance
1472 // for each such relocate, so we skip the whole transformation.
1476 Value *Base = ToReplace->getBasePtr();
1477 auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
1478 if (!Derived || Derived->getPointerOperand() != Base)
1481 SmallVector<Value *, 2> OffsetV;
1482 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
1485 // Create a Builder and replace the target callsite with a gep
1486 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
1488 // Insert after RelocatedBase
1489 IRBuilder<> Builder(RelocatedBase->getNextNode());
1490 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1492 // If gc_relocate does not match the actual type, cast it to the right type.
1493 // In theory, there must be a bitcast after gc_relocate if the type does not
1494 // match, and we should reuse it to get the derived pointer. But it could be
1498 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1503 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1507 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
1508 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
1510 // In this case, we can not find the bitcast any more. So we insert a new bitcast
1511 // no matter there is already one or not. In this way, we can handle all cases, and
1512 // the extra bitcast should be optimized away in later passes.
1513 Value *ActualRelocatedBase = RelocatedBase;
1514 if (RelocatedBase->getType() != Base->getType()) {
1515 ActualRelocatedBase =
1516 Builder.CreateBitCast(RelocatedBase, Base->getType());
1518 Value *Replacement = Builder.CreateGEP(
1519 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
1520 Replacement->takeName(ToReplace);
1521 // If the newly generated derived pointer's type does not match the original derived
1522 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
1523 Value *ActualReplacement = Replacement;
1524 if (Replacement->getType() != ToReplace->getType()) {
1526 Builder.CreateBitCast(Replacement, ToReplace->getType());
1528 ToReplace->replaceAllUsesWith(ActualReplacement);
1529 ToReplace->eraseFromParent();
1539 // %ptr = gep %base + 15
1540 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1541 // %base' = relocate(%tok, i32 4, i32 4)
1542 // %ptr' = relocate(%tok, i32 4, i32 5)
1543 // %val = load %ptr'
1548 // %ptr = gep %base + 15
1549 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1550 // %base' = gc.relocate(%tok, i32 4, i32 4)
1551 // %ptr' = gep %base' + 15
1552 // %val = load %ptr'
1553 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
1554 bool MadeChange = false;
1555 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1557 for (auto *U : I.users())
1558 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1559 // Collect all the relocate calls associated with a statepoint
1560 AllRelocateCalls.push_back(Relocate);
1562 // We need atleast one base pointer relocation + one derived pointer
1563 // relocation to mangle
1564 if (AllRelocateCalls.size() < 2)
1567 // RelocateInstMap is a mapping from the base relocate instruction to the
1568 // corresponding derived relocate instructions
1569 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
1570 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1571 if (RelocateInstMap.empty())
1574 for (auto &Item : RelocateInstMap)
1575 // Item.first is the RelocatedBase to offset against
1576 // Item.second is the vector of Targets to replace
1577 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1581 /// SinkCast - Sink the specified cast instruction into its user blocks
1582 static bool SinkCast(CastInst *CI) {
1583 BasicBlock *DefBB = CI->getParent();
1585 /// InsertedCasts - Only insert a cast in each block once.
1586 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1588 bool MadeChange = false;
1589 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1591 Use &TheUse = UI.getUse();
1592 Instruction *User = cast<Instruction>(*UI);
1594 // Figure out which BB this cast is used in. For PHI's this is the
1595 // appropriate predecessor block.
1596 BasicBlock *UserBB = User->getParent();
1597 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1598 UserBB = PN->getIncomingBlock(TheUse);
1601 // Preincrement use iterator so we don't invalidate it.
1604 // If the block selected to receive the cast is an EH pad that does not
1605 // allow non-PHI instructions before the terminator, we can't sink the
1607 if (UserBB->getTerminator()->isEHPad())
1610 // If this user is in the same block as the cast, don't change the cast.
1611 if (UserBB == DefBB) continue;
1613 // If we have already inserted a cast into this block, use it.
1614 CastInst *&InsertedCast = InsertedCasts[UserBB];
1616 if (!InsertedCast) {
1617 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1618 assert(InsertPt != UserBB->end());
1619 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1620 CI->getType(), "", &*InsertPt);
1623 // Replace a use of the cast with a use of the new cast.
1624 TheUse = InsertedCast;
1629 // If we removed all uses, nuke the cast.
1630 if (CI->use_empty()) {
1631 CI->eraseFromParent();
1638 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1639 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1640 /// reduce the number of virtual registers that must be created and coalesced.
1642 /// Return true if any changes are made.
1644 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1645 const DataLayout &DL) {
1646 // If this is a noop copy,
1647 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1648 EVT DstVT = TLI.getValueType(DL, CI->getType());
1650 // This is an fp<->int conversion?
1651 if (SrcVT.isInteger() != DstVT.isInteger())
1654 // If this is an extension, it will be a zero or sign extension, which
1656 if (SrcVT.bitsLT(DstVT)) return false;
1658 // If these values will be promoted, find out what they will be promoted
1659 // to. This helps us consider truncates on PPC as noop copies when they
1661 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1662 TargetLowering::TypePromoteInteger)
1663 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1664 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1665 TargetLowering::TypePromoteInteger)
1666 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1668 // If, after promotion, these are the same types, this is a noop copy.
1672 return SinkCast(CI);
1675 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
1678 /// Return true if any changes were made.
1679 static bool CombineUAddWithOverflow(CmpInst *CI) {
1683 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
1686 Type *Ty = AddI->getType();
1687 if (!isa<IntegerType>(Ty))
1690 // We don't want to move around uses of condition values this late, so we we
1691 // check if it is legal to create the call to the intrinsic in the basic
1692 // block containing the icmp:
1694 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
1698 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
1700 if (AddI->hasOneUse())
1701 assert(*AddI->user_begin() == CI && "expected!");
1704 Module *M = CI->getModule();
1705 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1707 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
1709 auto *UAddWithOverflow =
1710 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
1711 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
1713 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
1715 CI->replaceAllUsesWith(Overflow);
1716 AddI->replaceAllUsesWith(UAdd);
1717 CI->eraseFromParent();
1718 AddI->eraseFromParent();
1722 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1723 /// registers that must be created and coalesced. This is a clear win except on
1724 /// targets with multiple condition code registers (PowerPC), where it might
1725 /// lose; some adjustment may be wanted there.
1727 /// Return true if any changes are made.
1728 static bool SinkCmpExpression(CmpInst *CI) {
1729 BasicBlock *DefBB = CI->getParent();
1731 /// Only insert a cmp in each block once.
1732 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1734 bool MadeChange = false;
1735 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1737 Use &TheUse = UI.getUse();
1738 Instruction *User = cast<Instruction>(*UI);
1740 // Preincrement use iterator so we don't invalidate it.
1743 // Don't bother for PHI nodes.
1744 if (isa<PHINode>(User))
1747 // Figure out which BB this cmp is used in.
1748 BasicBlock *UserBB = User->getParent();
1750 // If this user is in the same block as the cmp, don't change the cmp.
1751 if (UserBB == DefBB) continue;
1753 // If we have already inserted a cmp into this block, use it.
1754 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1757 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1758 assert(InsertPt != UserBB->end());
1760 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1761 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
1764 // Replace a use of the cmp with a use of the new cmp.
1765 TheUse = InsertedCmp;
1770 // If we removed all uses, nuke the cmp.
1771 if (CI->use_empty()) {
1772 CI->eraseFromParent();
1779 static bool OptimizeCmpExpression(CmpInst *CI) {
1780 if (SinkCmpExpression(CI))
1783 if (CombineUAddWithOverflow(CI))
1789 /// Check if the candidates could be combined with a shift instruction, which
1791 /// 1. Truncate instruction
1792 /// 2. And instruction and the imm is a mask of the low bits:
1793 /// imm & (imm+1) == 0
1794 static bool isExtractBitsCandidateUse(Instruction *User) {
1795 if (!isa<TruncInst>(User)) {
1796 if (User->getOpcode() != Instruction::And ||
1797 !isa<ConstantInt>(User->getOperand(1)))
1800 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
1802 if ((Cimm & (Cimm + 1)).getBoolValue())
1808 /// Sink both shift and truncate instruction to the use of truncate's BB.
1810 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
1811 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
1812 const TargetLowering &TLI, const DataLayout &DL) {
1813 BasicBlock *UserBB = User->getParent();
1814 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
1815 TruncInst *TruncI = dyn_cast<TruncInst>(User);
1816 bool MadeChange = false;
1818 for (Value::user_iterator TruncUI = TruncI->user_begin(),
1819 TruncE = TruncI->user_end();
1820 TruncUI != TruncE;) {
1822 Use &TruncTheUse = TruncUI.getUse();
1823 Instruction *TruncUser = cast<Instruction>(*TruncUI);
1824 // Preincrement use iterator so we don't invalidate it.
1828 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
1832 // If the use is actually a legal node, there will not be an
1833 // implicit truncate.
1834 // FIXME: always querying the result type is just an
1835 // approximation; some nodes' legality is determined by the
1836 // operand or other means. There's no good way to find out though.
1837 if (TLI.isOperationLegalOrCustom(
1838 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
1841 // Don't bother for PHI nodes.
1842 if (isa<PHINode>(TruncUser))
1845 BasicBlock *TruncUserBB = TruncUser->getParent();
1847 if (UserBB == TruncUserBB)
1850 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
1851 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
1853 if (!InsertedShift && !InsertedTrunc) {
1854 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
1855 assert(InsertPt != TruncUserBB->end());
1857 if (ShiftI->getOpcode() == Instruction::AShr)
1858 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1861 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1865 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
1867 assert(TruncInsertPt != TruncUserBB->end());
1869 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
1870 TruncI->getType(), "", &*TruncInsertPt);
1874 TruncTheUse = InsertedTrunc;
1880 /// Sink the shift *right* instruction into user blocks if the uses could
1881 /// potentially be combined with this shift instruction and generate BitExtract
1882 /// instruction. It will only be applied if the architecture supports BitExtract
1883 /// instruction. Here is an example:
1885 /// %x.extract.shift = lshr i64 %arg1, 32
1887 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1891 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1892 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1894 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1896 /// Return true if any changes are made.
1897 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1898 const TargetLowering &TLI,
1899 const DataLayout &DL) {
1900 BasicBlock *DefBB = ShiftI->getParent();
1902 /// Only insert instructions in each block once.
1903 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1905 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1907 bool MadeChange = false;
1908 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1910 Use &TheUse = UI.getUse();
1911 Instruction *User = cast<Instruction>(*UI);
1912 // Preincrement use iterator so we don't invalidate it.
1915 // Don't bother for PHI nodes.
1916 if (isa<PHINode>(User))
1919 if (!isExtractBitsCandidateUse(User))
1922 BasicBlock *UserBB = User->getParent();
1924 if (UserBB == DefBB) {
1925 // If the shift and truncate instruction are in the same BB. The use of
1926 // the truncate(TruncUse) may still introduce another truncate if not
1927 // legal. In this case, we would like to sink both shift and truncate
1928 // instruction to the BB of TruncUse.
1931 // i64 shift.result = lshr i64 opnd, imm
1932 // trunc.result = trunc shift.result to i16
1935 // ----> We will have an implicit truncate here if the architecture does
1936 // not have i16 compare.
1937 // cmp i16 trunc.result, opnd2
1939 if (isa<TruncInst>(User) && shiftIsLegal
1940 // If the type of the truncate is legal, no trucate will be
1941 // introduced in other basic blocks.
1943 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1945 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1949 // If we have already inserted a shift into this block, use it.
1950 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1952 if (!InsertedShift) {
1953 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1954 assert(InsertPt != UserBB->end());
1956 if (ShiftI->getOpcode() == Instruction::AShr)
1957 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1960 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1966 // Replace a use of the shift with a use of the new shift.
1967 TheUse = InsertedShift;
1970 // If we removed all uses, nuke the shift.
1971 if (ShiftI->use_empty())
1972 ShiftI->eraseFromParent();
1977 // Translate a masked load intrinsic like
1978 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1979 // <16 x i1> %mask, <16 x i32> %passthru)
1980 // to a chain of basic blocks, with loading element one-by-one if
1981 // the appropriate mask bit is set
1983 // %1 = bitcast i8* %addr to i32*
1984 // %2 = extractelement <16 x i1> %mask, i32 0
1985 // %3 = icmp eq i1 %2, true
1986 // br i1 %3, label %cond.load, label %else
1988 //cond.load: ; preds = %0
1989 // %4 = getelementptr i32* %1, i32 0
1990 // %5 = load i32* %4
1991 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1994 //else: ; preds = %0, %cond.load
1995 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1996 // %7 = extractelement <16 x i1> %mask, i32 1
1997 // %8 = icmp eq i1 %7, true
1998 // br i1 %8, label %cond.load1, label %else2
2000 //cond.load1: ; preds = %else
2001 // %9 = getelementptr i32* %1, i32 1
2002 // %10 = load i32* %9
2003 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
2006 //else2: ; preds = %else, %cond.load1
2007 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
2008 // %12 = extractelement <16 x i1> %mask, i32 2
2009 // %13 = icmp eq i1 %12, true
2010 // br i1 %13, label %cond.load4, label %else5
2012 static void ScalarizeMaskedLoad(CallInst *CI) {
2013 Value *Ptr = CI->getArgOperand(0);
2014 Value *Alignment = CI->getArgOperand(1);
2015 Value *Mask = CI->getArgOperand(2);
2016 Value *Src0 = CI->getArgOperand(3);
2018 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2019 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
2020 assert(VecType && "Unexpected return type of masked load intrinsic");
2022 Type *EltTy = CI->getType()->getVectorElementType();
2024 IRBuilder<> Builder(CI->getContext());
2025 Instruction *InsertPt = CI;
2026 BasicBlock *IfBlock = CI->getParent();
2027 BasicBlock *CondBlock = nullptr;
2028 BasicBlock *PrevIfBlock = CI->getParent();
2030 Builder.SetInsertPoint(InsertPt);
2031 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2033 // Short-cut if the mask is all-true.
2034 bool IsAllOnesMask = isa<Constant>(Mask) &&
2035 cast<Constant>(Mask)->isAllOnesValue();
2037 if (IsAllOnesMask) {
2038 Value *NewI = Builder.CreateAlignedLoad(Ptr, AlignVal);
2039 CI->replaceAllUsesWith(NewI);
2040 CI->eraseFromParent();
2044 // Adjust alignment for the scalar instruction.
2045 AlignVal = std::min(AlignVal, VecType->getScalarSizeInBits()/8);
2046 // Bitcast %addr fron i8* to EltTy*
2048 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
2049 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
2050 unsigned VectorWidth = VecType->getNumElements();
2052 Value *UndefVal = UndefValue::get(VecType);
2054 // The result vector
2055 Value *VResult = UndefVal;
2057 if (isa<ConstantVector>(Mask)) {
2058 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2059 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2062 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2063 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
2064 VResult = Builder.CreateInsertElement(VResult, Load,
2065 Builder.getInt32(Idx));
2067 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
2068 CI->replaceAllUsesWith(NewI);
2069 CI->eraseFromParent();
2073 PHINode *Phi = nullptr;
2074 Value *PrevPhi = UndefVal;
2076 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2078 // Fill the "else" block, created in the previous iteration
2080 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
2081 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
2082 // %to_load = icmp eq i1 %mask_1, true
2083 // br i1 %to_load, label %cond.load, label %else
2086 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
2087 Phi->addIncoming(VResult, CondBlock);
2088 Phi->addIncoming(PrevPhi, PrevIfBlock);
2093 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
2094 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2095 ConstantInt::get(Predicate->getType(), 1));
2097 // Create "cond" block
2099 // %EltAddr = getelementptr i32* %1, i32 0
2100 // %Elt = load i32* %EltAddr
2101 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
2103 CondBlock = IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.load");
2104 Builder.SetInsertPoint(InsertPt);
2107 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2108 LoadInst *Load = Builder.CreateAlignedLoad(Gep, AlignVal);
2109 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
2111 // Create "else" block, fill it in the next iteration
2112 BasicBlock *NewIfBlock =
2113 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
2114 Builder.SetInsertPoint(InsertPt);
2115 Instruction *OldBr = IfBlock->getTerminator();
2116 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2117 OldBr->eraseFromParent();
2118 PrevIfBlock = IfBlock;
2119 IfBlock = NewIfBlock;
2122 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
2123 Phi->addIncoming(VResult, CondBlock);
2124 Phi->addIncoming(PrevPhi, PrevIfBlock);
2125 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
2126 CI->replaceAllUsesWith(NewI);
2127 CI->eraseFromParent();
2130 // Translate a masked store intrinsic, like
2131 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
2133 // to a chain of basic blocks, that stores element one-by-one if
2134 // the appropriate mask bit is set
2136 // %1 = bitcast i8* %addr to i32*
2137 // %2 = extractelement <16 x i1> %mask, i32 0
2138 // %3 = icmp eq i1 %2, true
2139 // br i1 %3, label %cond.store, label %else
2141 // cond.store: ; preds = %0
2142 // %4 = extractelement <16 x i32> %val, i32 0
2143 // %5 = getelementptr i32* %1, i32 0
2144 // store i32 %4, i32* %5
2147 // else: ; preds = %0, %cond.store
2148 // %6 = extractelement <16 x i1> %mask, i32 1
2149 // %7 = icmp eq i1 %6, true
2150 // br i1 %7, label %cond.store1, label %else2
2152 // cond.store1: ; preds = %else
2153 // %8 = extractelement <16 x i32> %val, i32 1
2154 // %9 = getelementptr i32* %1, i32 1
2155 // store i32 %8, i32* %9
2158 static void ScalarizeMaskedStore(CallInst *CI) {
2159 Value *Src = CI->getArgOperand(0);
2160 Value *Ptr = CI->getArgOperand(1);
2161 Value *Alignment = CI->getArgOperand(2);
2162 Value *Mask = CI->getArgOperand(3);
2164 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2165 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
2166 assert(VecType && "Unexpected data type in masked store intrinsic");
2168 Type *EltTy = VecType->getElementType();
2170 IRBuilder<> Builder(CI->getContext());
2171 Instruction *InsertPt = CI;
2172 BasicBlock *IfBlock = CI->getParent();
2173 Builder.SetInsertPoint(InsertPt);
2174 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2176 // Short-cut if the mask is all-true.
2177 bool IsAllOnesMask = isa<Constant>(Mask) &&
2178 cast<Constant>(Mask)->isAllOnesValue();
2180 if (IsAllOnesMask) {
2181 Builder.CreateAlignedStore(Src, Ptr, AlignVal);
2182 CI->eraseFromParent();
2186 // Adjust alignment for the scalar instruction.
2187 AlignVal = std::max(AlignVal, VecType->getScalarSizeInBits()/8);
2188 // Bitcast %addr fron i8* to EltTy*
2190 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
2191 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
2192 unsigned VectorWidth = VecType->getNumElements();
2194 if (isa<ConstantVector>(Mask)) {
2195 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2196 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2198 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
2200 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2201 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
2203 CI->eraseFromParent();
2207 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2209 // Fill the "else" block, created in the previous iteration
2211 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
2212 // %to_store = icmp eq i1 %mask_1, true
2213 // br i1 %to_store, label %cond.store, label %else
2215 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
2216 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2217 ConstantInt::get(Predicate->getType(), 1));
2219 // Create "cond" block
2221 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
2222 // %EltAddr = getelementptr i32* %1, i32 0
2223 // %store i32 %OneElt, i32* %EltAddr
2225 BasicBlock *CondBlock =
2226 IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.store");
2227 Builder.SetInsertPoint(InsertPt);
2229 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
2231 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2232 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
2234 // Create "else" block, fill it in the next iteration
2235 BasicBlock *NewIfBlock =
2236 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
2237 Builder.SetInsertPoint(InsertPt);
2238 Instruction *OldBr = IfBlock->getTerminator();
2239 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2240 OldBr->eraseFromParent();
2241 IfBlock = NewIfBlock;
2243 CI->eraseFromParent();
2246 // Translate a masked gather intrinsic like
2247 // <16 x i32 > @llvm.masked.gather.v16i32( <16 x i32*> %Ptrs, i32 4,
2248 // <16 x i1> %Mask, <16 x i32> %Src)
2249 // to a chain of basic blocks, with loading element one-by-one if
2250 // the appropriate mask bit is set
2252 // % Ptrs = getelementptr i32, i32* %base, <16 x i64> %ind
2253 // % Mask0 = extractelement <16 x i1> %Mask, i32 0
2254 // % ToLoad0 = icmp eq i1 % Mask0, true
2255 // br i1 % ToLoad0, label %cond.load, label %else
2258 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
2259 // % Load0 = load i32, i32* % Ptr0, align 4
2260 // % Res0 = insertelement <16 x i32> undef, i32 % Load0, i32 0
2264 // %res.phi.else = phi <16 x i32>[% Res0, %cond.load], [undef, % 0]
2265 // % Mask1 = extractelement <16 x i1> %Mask, i32 1
2266 // % ToLoad1 = icmp eq i1 % Mask1, true
2267 // br i1 % ToLoad1, label %cond.load1, label %else2
2270 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2271 // % Load1 = load i32, i32* % Ptr1, align 4
2272 // % Res1 = insertelement <16 x i32> %res.phi.else, i32 % Load1, i32 1
2275 // % Result = select <16 x i1> %Mask, <16 x i32> %res.phi.select, <16 x i32> %Src
2276 // ret <16 x i32> %Result
2277 static void ScalarizeMaskedGather(CallInst *CI) {
2278 Value *Ptrs = CI->getArgOperand(0);
2279 Value *Alignment = CI->getArgOperand(1);
2280 Value *Mask = CI->getArgOperand(2);
2281 Value *Src0 = CI->getArgOperand(3);
2283 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
2285 assert(VecType && "Unexpected return type of masked load intrinsic");
2287 IRBuilder<> Builder(CI->getContext());
2288 Instruction *InsertPt = CI;
2289 BasicBlock *IfBlock = CI->getParent();
2290 BasicBlock *CondBlock = nullptr;
2291 BasicBlock *PrevIfBlock = CI->getParent();
2292 Builder.SetInsertPoint(InsertPt);
2293 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2295 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2297 Value *UndefVal = UndefValue::get(VecType);
2299 // The result vector
2300 Value *VResult = UndefVal;
2301 unsigned VectorWidth = VecType->getNumElements();
2303 // Shorten the way if the mask is a vector of constants.
2304 bool IsConstMask = isa<ConstantVector>(Mask);
2307 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2308 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2310 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2311 "Ptr" + Twine(Idx));
2312 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
2313 "Load" + Twine(Idx));
2314 VResult = Builder.CreateInsertElement(VResult, Load,
2315 Builder.getInt32(Idx),
2316 "Res" + Twine(Idx));
2318 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
2319 CI->replaceAllUsesWith(NewI);
2320 CI->eraseFromParent();
2324 PHINode *Phi = nullptr;
2325 Value *PrevPhi = UndefVal;
2327 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2329 // Fill the "else" block, created in the previous iteration
2331 // %Mask1 = extractelement <16 x i1> %Mask, i32 1
2332 // %ToLoad1 = icmp eq i1 %Mask1, true
2333 // br i1 %ToLoad1, label %cond.load, label %else
2336 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
2337 Phi->addIncoming(VResult, CondBlock);
2338 Phi->addIncoming(PrevPhi, PrevIfBlock);
2343 Value *Predicate = Builder.CreateExtractElement(Mask,
2344 Builder.getInt32(Idx),
2345 "Mask" + Twine(Idx));
2346 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2347 ConstantInt::get(Predicate->getType(), 1),
2348 "ToLoad" + Twine(Idx));
2350 // Create "cond" block
2352 // %EltAddr = getelementptr i32* %1, i32 0
2353 // %Elt = load i32* %EltAddr
2354 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
2356 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
2357 Builder.SetInsertPoint(InsertPt);
2359 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2360 "Ptr" + Twine(Idx));
2361 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
2362 "Load" + Twine(Idx));
2363 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx),
2364 "Res" + Twine(Idx));
2366 // Create "else" block, fill it in the next iteration
2367 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2368 Builder.SetInsertPoint(InsertPt);
2369 Instruction *OldBr = IfBlock->getTerminator();
2370 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2371 OldBr->eraseFromParent();
2372 PrevIfBlock = IfBlock;
2373 IfBlock = NewIfBlock;
2376 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
2377 Phi->addIncoming(VResult, CondBlock);
2378 Phi->addIncoming(PrevPhi, PrevIfBlock);
2379 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
2380 CI->replaceAllUsesWith(NewI);
2381 CI->eraseFromParent();
2384 // Translate a masked scatter intrinsic, like
2385 // void @llvm.masked.scatter.v16i32(<16 x i32> %Src, <16 x i32*>* %Ptrs, i32 4,
2387 // to a chain of basic blocks, that stores element one-by-one if
2388 // the appropriate mask bit is set.
2390 // % Ptrs = getelementptr i32, i32* %ptr, <16 x i64> %ind
2391 // % Mask0 = extractelement <16 x i1> % Mask, i32 0
2392 // % ToStore0 = icmp eq i1 % Mask0, true
2393 // br i1 %ToStore0, label %cond.store, label %else
2396 // % Elt0 = extractelement <16 x i32> %Src, i32 0
2397 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
2398 // store i32 %Elt0, i32* % Ptr0, align 4
2402 // % Mask1 = extractelement <16 x i1> % Mask, i32 1
2403 // % ToStore1 = icmp eq i1 % Mask1, true
2404 // br i1 % ToStore1, label %cond.store1, label %else2
2407 // % Elt1 = extractelement <16 x i32> %Src, i32 1
2408 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2409 // store i32 % Elt1, i32* % Ptr1, align 4
2412 static void ScalarizeMaskedScatter(CallInst *CI) {
2413 Value *Src = CI->getArgOperand(0);
2414 Value *Ptrs = CI->getArgOperand(1);
2415 Value *Alignment = CI->getArgOperand(2);
2416 Value *Mask = CI->getArgOperand(3);
2418 assert(isa<VectorType>(Src->getType()) &&
2419 "Unexpected data type in masked scatter intrinsic");
2420 assert(isa<VectorType>(Ptrs->getType()) &&
2421 isa<PointerType>(Ptrs->getType()->getVectorElementType()) &&
2422 "Vector of pointers is expected in masked scatter intrinsic");
2424 IRBuilder<> Builder(CI->getContext());
2425 Instruction *InsertPt = CI;
2426 BasicBlock *IfBlock = CI->getParent();
2427 Builder.SetInsertPoint(InsertPt);
2428 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2430 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2431 unsigned VectorWidth = Src->getType()->getVectorNumElements();
2433 // Shorten the way if the mask is a vector of constants.
2434 bool IsConstMask = isa<ConstantVector>(Mask);
2437 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2438 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2440 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
2441 "Elt" + Twine(Idx));
2442 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2443 "Ptr" + Twine(Idx));
2444 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
2446 CI->eraseFromParent();
2449 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2450 // Fill the "else" block, created in the previous iteration
2452 // % Mask1 = extractelement <16 x i1> % Mask, i32 Idx
2453 // % ToStore = icmp eq i1 % Mask1, true
2454 // br i1 % ToStore, label %cond.store, label %else
2456 Value *Predicate = Builder.CreateExtractElement(Mask,
2457 Builder.getInt32(Idx),
2458 "Mask" + Twine(Idx));
2460 Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2461 ConstantInt::get(Predicate->getType(), 1),
2462 "ToStore" + Twine(Idx));
2464 // Create "cond" block
2466 // % Elt1 = extractelement <16 x i32> %Src, i32 1
2467 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2468 // %store i32 % Elt1, i32* % Ptr1
2470 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2471 Builder.SetInsertPoint(InsertPt);
2473 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
2474 "Elt" + Twine(Idx));
2475 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2476 "Ptr" + Twine(Idx));
2477 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
2479 // Create "else" block, fill it in the next iteration
2480 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2481 Builder.SetInsertPoint(InsertPt);
2482 Instruction *OldBr = IfBlock->getTerminator();
2483 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2484 OldBr->eraseFromParent();
2485 IfBlock = NewIfBlock;
2487 CI->eraseFromParent();
2490 /// If counting leading or trailing zeros is an expensive operation and a zero
2491 /// input is defined, add a check for zero to avoid calling the intrinsic.
2493 /// We want to transform:
2494 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
2498 /// %cmpz = icmp eq i64 %A, 0
2499 /// br i1 %cmpz, label %cond.end, label %cond.false
2501 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
2502 /// br label %cond.end
2504 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
2506 /// If the transform is performed, return true and set ModifiedDT to true.
2507 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
2508 const TargetLowering *TLI,
2509 const DataLayout *DL,
2514 // If a zero input is undefined, it doesn't make sense to despeculate that.
2515 if (match(CountZeros->getOperand(1), m_One()))
2518 // If it's cheap to speculate, there's nothing to do.
2519 auto IntrinsicID = CountZeros->getIntrinsicID();
2520 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
2521 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
2524 // Only handle legal scalar cases. Anything else requires too much work.
2525 Type *Ty = CountZeros->getType();
2526 unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
2527 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSize())
2530 // The intrinsic will be sunk behind a compare against zero and branch.
2531 BasicBlock *StartBlock = CountZeros->getParent();
2532 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
2534 // Create another block after the count zero intrinsic. A PHI will be added
2535 // in this block to select the result of the intrinsic or the bit-width
2536 // constant if the input to the intrinsic is zero.
2537 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
2538 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
2540 // Set up a builder to create a compare, conditional branch, and PHI.
2541 IRBuilder<> Builder(CountZeros->getContext());
2542 Builder.SetInsertPoint(StartBlock->getTerminator());
2543 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
2545 // Replace the unconditional branch that was created by the first split with
2546 // a compare against zero and a conditional branch.
2547 Value *Zero = Constant::getNullValue(Ty);
2548 Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
2549 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
2550 StartBlock->getTerminator()->eraseFromParent();
2552 // Create a PHI in the end block to select either the output of the intrinsic
2553 // or the bit width of the operand.
2554 Builder.SetInsertPoint(&EndBlock->front());
2555 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
2556 CountZeros->replaceAllUsesWith(PN);
2557 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
2558 PN->addIncoming(BitWidth, StartBlock);
2559 PN->addIncoming(CountZeros, CallBlock);
2561 // We are explicitly handling the zero case, so we can set the intrinsic's
2562 // undefined zero argument to 'true'. This will also prevent reprocessing the
2563 // intrinsic; we only despeculate when a zero input is defined.
2564 CountZeros->setArgOperand(1, Builder.getTrue());
2569 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool& ModifiedDT) {
2570 BasicBlock *BB = CI->getParent();
2572 // Lower inline assembly if we can.
2573 // If we found an inline asm expession, and if the target knows how to
2574 // lower it to normal LLVM code, do so now.
2575 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
2576 if (TLI->ExpandInlineAsm(CI)) {
2577 // Avoid invalidating the iterator.
2578 CurInstIterator = BB->begin();
2579 // Avoid processing instructions out of order, which could cause
2580 // reuse before a value is defined.
2584 // Sink address computing for memory operands into the block.
2585 if (optimizeInlineAsmInst(CI))
2589 // Align the pointer arguments to this call if the target thinks it's a good
2591 unsigned MinSize, PrefAlign;
2592 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
2593 for (auto &Arg : CI->arg_operands()) {
2594 // We want to align both objects whose address is used directly and
2595 // objects whose address is used in casts and GEPs, though it only makes
2596 // sense for GEPs if the offset is a multiple of the desired alignment and
2597 // if size - offset meets the size threshold.
2598 if (!Arg->getType()->isPointerTy())
2600 APInt Offset(DL->getPointerSizeInBits(
2601 cast<PointerType>(Arg->getType())->getAddressSpace()),
2603 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
2604 uint64_t Offset2 = Offset.getLimitedValue();
2605 if ((Offset2 & (PrefAlign-1)) != 0)
2608 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
2609 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
2610 AI->setAlignment(PrefAlign);
2611 // Global variables can only be aligned if they are defined in this
2612 // object (i.e. they are uniquely initialized in this object), and
2613 // over-aligning global variables that have an explicit section is
2616 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
2617 GV->getAlignment() < PrefAlign &&
2618 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
2620 GV->setAlignment(PrefAlign);
2622 // If this is a memcpy (or similar) then we may be able to improve the
2624 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
2625 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
2626 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
2627 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
2628 if (Align > MI->getAlignment())
2629 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
2633 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
2635 switch (II->getIntrinsicID()) {
2637 case Intrinsic::objectsize: {
2638 // Lower all uses of llvm.objectsize.*
2639 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
2640 Type *ReturnTy = CI->getType();
2641 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
2643 // Substituting this can cause recursive simplifications, which can
2644 // invalidate our iterator. Use a WeakVH to hold onto it in case this
2646 WeakVH IterHandle(&*CurInstIterator);
2648 replaceAndRecursivelySimplify(CI, RetVal,
2651 // If the iterator instruction was recursively deleted, start over at the
2652 // start of the block.
2653 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
2654 CurInstIterator = BB->begin();
2659 case Intrinsic::masked_load: {
2660 // Scalarize unsupported vector masked load
2661 if (!TTI->isLegalMaskedLoad(CI->getType())) {
2662 ScalarizeMaskedLoad(CI);
2668 case Intrinsic::masked_store: {
2669 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType())) {
2670 ScalarizeMaskedStore(CI);
2676 case Intrinsic::masked_gather: {
2677 if (!TTI->isLegalMaskedGather(CI->getType())) {
2678 ScalarizeMaskedGather(CI);
2684 case Intrinsic::masked_scatter: {
2685 if (!TTI->isLegalMaskedScatter(CI->getArgOperand(0)->getType())) {
2686 ScalarizeMaskedScatter(CI);
2692 case Intrinsic::aarch64_stlxr:
2693 case Intrinsic::aarch64_stxr: {
2694 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
2695 if (!ExtVal || !ExtVal->hasOneUse() ||
2696 ExtVal->getParent() == CI->getParent())
2698 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
2699 ExtVal->moveBefore(CI);
2700 // Mark this instruction as "inserted by CGP", so that other
2701 // optimizations don't touch it.
2702 InsertedInsts.insert(ExtVal);
2705 case Intrinsic::invariant_group_barrier:
2706 II->replaceAllUsesWith(II->getArgOperand(0));
2707 II->eraseFromParent();
2710 case Intrinsic::cttz:
2711 case Intrinsic::ctlz:
2712 // If counting zeros is expensive, try to avoid it.
2713 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
2717 // Unknown address space.
2718 // TODO: Target hook to pick which address space the intrinsic cares
2720 unsigned AddrSpace = ~0u;
2721 SmallVector<Value*, 2> PtrOps;
2723 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
2724 while (!PtrOps.empty())
2725 if (optimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
2730 // From here on out we're working with named functions.
2731 if (!CI->getCalledFunction()) return false;
2733 // Lower all default uses of _chk calls. This is very similar
2734 // to what InstCombineCalls does, but here we are only lowering calls
2735 // to fortified library functions (e.g. __memcpy_chk) that have the default
2736 // "don't know" as the objectsize. Anything else should be left alone.
2737 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
2738 if (Value *V = Simplifier.optimizeCall(CI)) {
2739 CI->replaceAllUsesWith(V);
2740 CI->eraseFromParent();
2746 /// Look for opportunities to duplicate return instructions to the predecessor
2747 /// to enable tail call optimizations. The case it is currently looking for is:
2750 /// %tmp0 = tail call i32 @f0()
2751 /// br label %return
2753 /// %tmp1 = tail call i32 @f1()
2754 /// br label %return
2756 /// %tmp2 = tail call i32 @f2()
2757 /// br label %return
2759 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
2767 /// %tmp0 = tail call i32 @f0()
2770 /// %tmp1 = tail call i32 @f1()
2773 /// %tmp2 = tail call i32 @f2()
2776 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
2780 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
2784 PHINode *PN = nullptr;
2785 BitCastInst *BCI = nullptr;
2786 Value *V = RI->getReturnValue();
2788 BCI = dyn_cast<BitCastInst>(V);
2790 V = BCI->getOperand(0);
2792 PN = dyn_cast<PHINode>(V);
2797 if (PN && PN->getParent() != BB)
2800 // It's not safe to eliminate the sign / zero extension of the return value.
2801 // See llvm::isInTailCallPosition().
2802 const Function *F = BB->getParent();
2803 AttributeSet CallerAttrs = F->getAttributes();
2804 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
2805 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
2808 // Make sure there are no instructions between the PHI and return, or that the
2809 // return is the first instruction in the block.
2811 BasicBlock::iterator BI = BB->begin();
2812 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
2814 // Also skip over the bitcast.
2819 BasicBlock::iterator BI = BB->begin();
2820 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
2825 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
2827 SmallVector<CallInst*, 4> TailCalls;
2829 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
2830 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
2831 // Make sure the phi value is indeed produced by the tail call.
2832 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
2833 TLI->mayBeEmittedAsTailCall(CI))
2834 TailCalls.push_back(CI);
2837 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
2838 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
2839 if (!VisitedBBs.insert(*PI).second)
2842 BasicBlock::InstListType &InstList = (*PI)->getInstList();
2843 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
2844 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
2845 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
2849 CallInst *CI = dyn_cast<CallInst>(&*RI);
2850 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
2851 TailCalls.push_back(CI);
2855 bool Changed = false;
2856 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
2857 CallInst *CI = TailCalls[i];
2860 // Conservatively require the attributes of the call to match those of the
2861 // return. Ignore noalias because it doesn't affect the call sequence.
2862 AttributeSet CalleeAttrs = CS.getAttributes();
2863 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
2864 removeAttribute(Attribute::NoAlias) !=
2865 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
2866 removeAttribute(Attribute::NoAlias))
2869 // Make sure the call instruction is followed by an unconditional branch to
2870 // the return block.
2871 BasicBlock *CallBB = CI->getParent();
2872 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
2873 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
2876 // Duplicate the return into CallBB.
2877 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
2878 ModifiedDT = Changed = true;
2882 // If we eliminated all predecessors of the block, delete the block now.
2883 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
2884 BB->eraseFromParent();
2889 //===----------------------------------------------------------------------===//
2890 // Memory Optimization
2891 //===----------------------------------------------------------------------===//
2895 /// This is an extended version of TargetLowering::AddrMode
2896 /// which holds actual Value*'s for register values.
2897 struct ExtAddrMode : public TargetLowering::AddrMode {
2900 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
2901 void print(raw_ostream &OS) const;
2904 bool operator==(const ExtAddrMode& O) const {
2905 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
2906 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
2907 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
2912 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2918 void ExtAddrMode::print(raw_ostream &OS) const {
2919 bool NeedPlus = false;
2922 OS << (NeedPlus ? " + " : "")
2924 BaseGV->printAsOperand(OS, /*PrintType=*/false);
2929 OS << (NeedPlus ? " + " : "")
2935 OS << (NeedPlus ? " + " : "")
2937 BaseReg->printAsOperand(OS, /*PrintType=*/false);
2941 OS << (NeedPlus ? " + " : "")
2943 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2949 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2950 void ExtAddrMode::dump() const {
2956 /// \brief This class provides transaction based operation on the IR.
2957 /// Every change made through this class is recorded in the internal state and
2958 /// can be undone (rollback) until commit is called.
2959 class TypePromotionTransaction {
2961 /// \brief This represents the common interface of the individual transaction.
2962 /// Each class implements the logic for doing one specific modification on
2963 /// the IR via the TypePromotionTransaction.
2964 class TypePromotionAction {
2966 /// The Instruction modified.
2970 /// \brief Constructor of the action.
2971 /// The constructor performs the related action on the IR.
2972 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2974 virtual ~TypePromotionAction() {}
2976 /// \brief Undo the modification done by this action.
2977 /// When this method is called, the IR must be in the same state as it was
2978 /// before this action was applied.
2979 /// \pre Undoing the action works if and only if the IR is in the exact same
2980 /// state as it was directly after this action was applied.
2981 virtual void undo() = 0;
2983 /// \brief Advocate every change made by this action.
2984 /// When the results on the IR of the action are to be kept, it is important
2985 /// to call this function, otherwise hidden information may be kept forever.
2986 virtual void commit() {
2987 // Nothing to be done, this action is not doing anything.
2991 /// \brief Utility to remember the position of an instruction.
2992 class InsertionHandler {
2993 /// Position of an instruction.
2994 /// Either an instruction:
2995 /// - Is the first in a basic block: BB is used.
2996 /// - Has a previous instructon: PrevInst is used.
2998 Instruction *PrevInst;
3001 /// Remember whether or not the instruction had a previous instruction.
3002 bool HasPrevInstruction;
3005 /// \brief Record the position of \p Inst.
3006 InsertionHandler(Instruction *Inst) {
3007 BasicBlock::iterator It = Inst->getIterator();
3008 HasPrevInstruction = (It != (Inst->getParent()->begin()));
3009 if (HasPrevInstruction)
3010 Point.PrevInst = &*--It;
3012 Point.BB = Inst->getParent();
3015 /// \brief Insert \p Inst at the recorded position.
3016 void insert(Instruction *Inst) {
3017 if (HasPrevInstruction) {
3018 if (Inst->getParent())
3019 Inst->removeFromParent();
3020 Inst->insertAfter(Point.PrevInst);
3022 Instruction *Position = &*Point.BB->getFirstInsertionPt();
3023 if (Inst->getParent())
3024 Inst->moveBefore(Position);
3026 Inst->insertBefore(Position);
3031 /// \brief Move an instruction before another.
3032 class InstructionMoveBefore : public TypePromotionAction {
3033 /// Original position of the instruction.
3034 InsertionHandler Position;
3037 /// \brief Move \p Inst before \p Before.
3038 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
3039 : TypePromotionAction(Inst), Position(Inst) {
3040 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
3041 Inst->moveBefore(Before);
3044 /// \brief Move the instruction back to its original position.
3045 void undo() override {
3046 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
3047 Position.insert(Inst);
3051 /// \brief Set the operand of an instruction with a new value.
3052 class OperandSetter : public TypePromotionAction {
3053 /// Original operand of the instruction.
3055 /// Index of the modified instruction.
3059 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
3060 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
3061 : TypePromotionAction(Inst), Idx(Idx) {
3062 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
3063 << "for:" << *Inst << "\n"
3064 << "with:" << *NewVal << "\n");
3065 Origin = Inst->getOperand(Idx);
3066 Inst->setOperand(Idx, NewVal);
3069 /// \brief Restore the original value of the instruction.
3070 void undo() override {
3071 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
3072 << "for: " << *Inst << "\n"
3073 << "with: " << *Origin << "\n");
3074 Inst->setOperand(Idx, Origin);
3078 /// \brief Hide the operands of an instruction.
3079 /// Do as if this instruction was not using any of its operands.
3080 class OperandsHider : public TypePromotionAction {
3081 /// The list of original operands.
3082 SmallVector<Value *, 4> OriginalValues;
3085 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
3086 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
3087 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
3088 unsigned NumOpnds = Inst->getNumOperands();
3089 OriginalValues.reserve(NumOpnds);
3090 for (unsigned It = 0; It < NumOpnds; ++It) {
3091 // Save the current operand.
3092 Value *Val = Inst->getOperand(It);
3093 OriginalValues.push_back(Val);
3095 // We could use OperandSetter here, but that would imply an overhead
3096 // that we are not willing to pay.
3097 Inst->setOperand(It, UndefValue::get(Val->getType()));
3101 /// \brief Restore the original list of uses.
3102 void undo() override {
3103 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
3104 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
3105 Inst->setOperand(It, OriginalValues[It]);
3109 /// \brief Build a truncate instruction.
3110 class TruncBuilder : public TypePromotionAction {
3113 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
3115 /// trunc Opnd to Ty.
3116 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
3117 IRBuilder<> Builder(Opnd);
3118 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
3119 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
3122 /// \brief Get the built value.
3123 Value *getBuiltValue() { return Val; }
3125 /// \brief Remove the built instruction.
3126 void undo() override {
3127 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
3128 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3129 IVal->eraseFromParent();
3133 /// \brief Build a sign extension instruction.
3134 class SExtBuilder : public TypePromotionAction {
3137 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
3139 /// sext Opnd to Ty.
3140 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3141 : TypePromotionAction(InsertPt) {
3142 IRBuilder<> Builder(InsertPt);
3143 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
3144 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
3147 /// \brief Get the built value.
3148 Value *getBuiltValue() { return Val; }
3150 /// \brief Remove the built instruction.
3151 void undo() override {
3152 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
3153 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3154 IVal->eraseFromParent();
3158 /// \brief Build a zero extension instruction.
3159 class ZExtBuilder : public TypePromotionAction {
3162 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
3164 /// zext Opnd to Ty.
3165 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3166 : TypePromotionAction(InsertPt) {
3167 IRBuilder<> Builder(InsertPt);
3168 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
3169 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
3172 /// \brief Get the built value.
3173 Value *getBuiltValue() { return Val; }
3175 /// \brief Remove the built instruction.
3176 void undo() override {
3177 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
3178 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3179 IVal->eraseFromParent();
3183 /// \brief Mutate an instruction to another type.
3184 class TypeMutator : public TypePromotionAction {
3185 /// Record the original type.
3189 /// \brief Mutate the type of \p Inst into \p NewTy.
3190 TypeMutator(Instruction *Inst, Type *NewTy)
3191 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
3192 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
3194 Inst->mutateType(NewTy);
3197 /// \brief Mutate the instruction back to its original type.
3198 void undo() override {
3199 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
3201 Inst->mutateType(OrigTy);
3205 /// \brief Replace the uses of an instruction by another instruction.
3206 class UsesReplacer : public TypePromotionAction {
3207 /// Helper structure to keep track of the replaced uses.
3208 struct InstructionAndIdx {
3209 /// The instruction using the instruction.
3211 /// The index where this instruction is used for Inst.
3213 InstructionAndIdx(Instruction *Inst, unsigned Idx)
3214 : Inst(Inst), Idx(Idx) {}
3217 /// Keep track of the original uses (pair Instruction, Index).
3218 SmallVector<InstructionAndIdx, 4> OriginalUses;
3219 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
3222 /// \brief Replace all the use of \p Inst by \p New.
3223 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
3224 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
3226 // Record the original uses.
3227 for (Use &U : Inst->uses()) {
3228 Instruction *UserI = cast<Instruction>(U.getUser());
3229 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
3231 // Now, we can replace the uses.
3232 Inst->replaceAllUsesWith(New);
3235 /// \brief Reassign the original uses of Inst to Inst.
3236 void undo() override {
3237 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
3238 for (use_iterator UseIt = OriginalUses.begin(),
3239 EndIt = OriginalUses.end();
3240 UseIt != EndIt; ++UseIt) {
3241 UseIt->Inst->setOperand(UseIt->Idx, Inst);
3246 /// \brief Remove an instruction from the IR.
3247 class InstructionRemover : public TypePromotionAction {
3248 /// Original position of the instruction.
3249 InsertionHandler Inserter;
3250 /// Helper structure to hide all the link to the instruction. In other
3251 /// words, this helps to do as if the instruction was removed.
3252 OperandsHider Hider;
3253 /// Keep track of the uses replaced, if any.
3254 UsesReplacer *Replacer;
3257 /// \brief Remove all reference of \p Inst and optinally replace all its
3259 /// \pre If !Inst->use_empty(), then New != nullptr
3260 InstructionRemover(Instruction *Inst, Value *New = nullptr)
3261 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
3264 Replacer = new UsesReplacer(Inst, New);
3265 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
3266 Inst->removeFromParent();
3269 ~InstructionRemover() override { delete Replacer; }
3271 /// \brief Really remove the instruction.
3272 void commit() override { delete Inst; }
3274 /// \brief Resurrect the instruction and reassign it to the proper uses if
3275 /// new value was provided when build this action.
3276 void undo() override {
3277 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
3278 Inserter.insert(Inst);
3286 /// Restoration point.
3287 /// The restoration point is a pointer to an action instead of an iterator
3288 /// because the iterator may be invalidated but not the pointer.
3289 typedef const TypePromotionAction *ConstRestorationPt;
3290 /// Advocate every changes made in that transaction.
3292 /// Undo all the changes made after the given point.
3293 void rollback(ConstRestorationPt Point);
3294 /// Get the current restoration point.
3295 ConstRestorationPt getRestorationPoint() const;
3297 /// \name API for IR modification with state keeping to support rollback.
3299 /// Same as Instruction::setOperand.
3300 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
3301 /// Same as Instruction::eraseFromParent.
3302 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
3303 /// Same as Value::replaceAllUsesWith.
3304 void replaceAllUsesWith(Instruction *Inst, Value *New);
3305 /// Same as Value::mutateType.
3306 void mutateType(Instruction *Inst, Type *NewTy);
3307 /// Same as IRBuilder::createTrunc.
3308 Value *createTrunc(Instruction *Opnd, Type *Ty);
3309 /// Same as IRBuilder::createSExt.
3310 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
3311 /// Same as IRBuilder::createZExt.
3312 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
3313 /// Same as Instruction::moveBefore.
3314 void moveBefore(Instruction *Inst, Instruction *Before);
3318 /// The ordered list of actions made so far.
3319 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
3320 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
3323 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
3326 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
3329 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
3332 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
3335 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
3337 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
3340 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
3341 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
3344 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
3346 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
3347 Value *Val = Ptr->getBuiltValue();
3348 Actions.push_back(std::move(Ptr));
3352 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
3353 Value *Opnd, Type *Ty) {
3354 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
3355 Value *Val = Ptr->getBuiltValue();
3356 Actions.push_back(std::move(Ptr));
3360 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
3361 Value *Opnd, Type *Ty) {
3362 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
3363 Value *Val = Ptr->getBuiltValue();
3364 Actions.push_back(std::move(Ptr));
3368 void TypePromotionTransaction::moveBefore(Instruction *Inst,
3369 Instruction *Before) {
3371 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
3374 TypePromotionTransaction::ConstRestorationPt
3375 TypePromotionTransaction::getRestorationPoint() const {
3376 return !Actions.empty() ? Actions.back().get() : nullptr;
3379 void TypePromotionTransaction::commit() {
3380 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
3386 void TypePromotionTransaction::rollback(
3387 TypePromotionTransaction::ConstRestorationPt Point) {
3388 while (!Actions.empty() && Point != Actions.back().get()) {
3389 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
3394 /// \brief A helper class for matching addressing modes.
3396 /// This encapsulates the logic for matching the target-legal addressing modes.
3397 class AddressingModeMatcher {
3398 SmallVectorImpl<Instruction*> &AddrModeInsts;
3399 const TargetMachine &TM;
3400 const TargetLowering &TLI;
3401 const DataLayout &DL;
3403 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
3404 /// the memory instruction that we're computing this address for.
3407 Instruction *MemoryInst;
3409 /// This is the addressing mode that we're building up. This is
3410 /// part of the return value of this addressing mode matching stuff.
3411 ExtAddrMode &AddrMode;
3413 /// The instructions inserted by other CodeGenPrepare optimizations.
3414 const SetOfInstrs &InsertedInsts;
3415 /// A map from the instructions to their type before promotion.
3416 InstrToOrigTy &PromotedInsts;
3417 /// The ongoing transaction where every action should be registered.
3418 TypePromotionTransaction &TPT;
3420 /// This is set to true when we should not do profitability checks.
3421 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
3422 bool IgnoreProfitability;
3424 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
3425 const TargetMachine &TM, Type *AT, unsigned AS,
3426 Instruction *MI, ExtAddrMode &AM,
3427 const SetOfInstrs &InsertedInsts,
3428 InstrToOrigTy &PromotedInsts,
3429 TypePromotionTransaction &TPT)
3430 : AddrModeInsts(AMI), TM(TM),
3431 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
3432 ->getTargetLowering()),
3433 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
3434 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
3435 PromotedInsts(PromotedInsts), TPT(TPT) {
3436 IgnoreProfitability = false;
3440 /// Find the maximal addressing mode that a load/store of V can fold,
3441 /// give an access type of AccessTy. This returns a list of involved
3442 /// instructions in AddrModeInsts.
3443 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
3445 /// \p PromotedInsts maps the instructions to their type before promotion.
3446 /// \p The ongoing transaction where every action should be registered.
3447 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
3448 Instruction *MemoryInst,
3449 SmallVectorImpl<Instruction*> &AddrModeInsts,
3450 const TargetMachine &TM,
3451 const SetOfInstrs &InsertedInsts,
3452 InstrToOrigTy &PromotedInsts,
3453 TypePromotionTransaction &TPT) {
3456 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
3457 MemoryInst, Result, InsertedInsts,
3458 PromotedInsts, TPT).matchAddr(V, 0);
3459 (void)Success; assert(Success && "Couldn't select *anything*?");
3463 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
3464 bool matchAddr(Value *V, unsigned Depth);
3465 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
3466 bool *MovedAway = nullptr);
3467 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
3468 ExtAddrMode &AMBefore,
3469 ExtAddrMode &AMAfter);
3470 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
3471 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
3472 Value *PromotedOperand) const;
3475 /// Try adding ScaleReg*Scale to the current addressing mode.
3476 /// Return true and update AddrMode if this addr mode is legal for the target,
3478 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3480 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3481 // mode. Just process that directly.
3483 return matchAddr(ScaleReg, Depth);
3485 // If the scale is 0, it takes nothing to add this.
3489 // If we already have a scale of this value, we can add to it, otherwise, we
3490 // need an available scale field.
3491 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3494 ExtAddrMode TestAddrMode = AddrMode;
3496 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3497 // [A+B + A*7] -> [B+A*8].
3498 TestAddrMode.Scale += Scale;
3499 TestAddrMode.ScaledReg = ScaleReg;
3501 // If the new address isn't legal, bail out.
3502 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3505 // It was legal, so commit it.
3506 AddrMode = TestAddrMode;
3508 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3509 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3510 // X*Scale + C*Scale to addr mode.
3511 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3512 if (isa<Instruction>(ScaleReg) && // not a constant expr.
3513 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3514 TestAddrMode.ScaledReg = AddLHS;
3515 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3517 // If this addressing mode is legal, commit it and remember that we folded
3518 // this instruction.
3519 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3520 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3521 AddrMode = TestAddrMode;
3526 // Otherwise, not (x+c)*scale, just return what we have.
3530 /// This is a little filter, which returns true if an addressing computation
3531 /// involving I might be folded into a load/store accessing it.
3532 /// This doesn't need to be perfect, but needs to accept at least
3533 /// the set of instructions that MatchOperationAddr can.
3534 static bool MightBeFoldableInst(Instruction *I) {
3535 switch (I->getOpcode()) {
3536 case Instruction::BitCast:
3537 case Instruction::AddrSpaceCast:
3538 // Don't touch identity bitcasts.
3539 if (I->getType() == I->getOperand(0)->getType())
3541 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
3542 case Instruction::PtrToInt:
3543 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3545 case Instruction::IntToPtr:
3546 // We know the input is intptr_t, so this is foldable.
3548 case Instruction::Add:
3550 case Instruction::Mul:
3551 case Instruction::Shl:
3552 // Can only handle X*C and X << C.
3553 return isa<ConstantInt>(I->getOperand(1));
3554 case Instruction::GetElementPtr:
3561 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
3562 /// \note \p Val is assumed to be the product of some type promotion.
3563 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3564 /// to be legal, as the non-promoted value would have had the same state.
3565 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
3566 const DataLayout &DL, Value *Val) {
3567 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3570 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3571 // If the ISDOpcode is undefined, it was undefined before the promotion.
3574 // Otherwise, check if the promoted instruction is legal or not.
3575 return TLI.isOperationLegalOrCustom(
3576 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3579 /// \brief Hepler class to perform type promotion.
3580 class TypePromotionHelper {
3581 /// \brief Utility function to check whether or not a sign or zero extension
3582 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3583 /// either using the operands of \p Inst or promoting \p Inst.
3584 /// The type of the extension is defined by \p IsSExt.
3585 /// In other words, check if:
3586 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3587 /// #1 Promotion applies:
3588 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3589 /// #2 Operand reuses:
3590 /// ext opnd1 to ConsideredExtType.
3591 /// \p PromotedInsts maps the instructions to their type before promotion.
3592 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3593 const InstrToOrigTy &PromotedInsts, bool IsSExt);
3595 /// \brief Utility function to determine if \p OpIdx should be promoted when
3596 /// promoting \p Inst.
3597 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3598 return !(isa<SelectInst>(Inst) && OpIdx == 0);
3601 /// \brief Utility function to promote the operand of \p Ext when this
3602 /// operand is a promotable trunc or sext or zext.
3603 /// \p PromotedInsts maps the instructions to their type before promotion.
3604 /// \p CreatedInstsCost[out] contains the cost of all instructions
3605 /// created to promote the operand of Ext.
3606 /// Newly added extensions are inserted in \p Exts.
3607 /// Newly added truncates are inserted in \p Truncs.
3608 /// Should never be called directly.
3609 /// \return The promoted value which is used instead of Ext.
3610 static Value *promoteOperandForTruncAndAnyExt(
3611 Instruction *Ext, TypePromotionTransaction &TPT,
3612 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3613 SmallVectorImpl<Instruction *> *Exts,
3614 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3616 /// \brief Utility function to promote the operand of \p Ext when this
3617 /// operand is promotable and is not a supported trunc or sext.
3618 /// \p PromotedInsts maps the instructions to their type before promotion.
3619 /// \p CreatedInstsCost[out] contains the cost of all the instructions
3620 /// created to promote the operand of Ext.
3621 /// Newly added extensions are inserted in \p Exts.
3622 /// Newly added truncates are inserted in \p Truncs.
3623 /// Should never be called directly.
3624 /// \return The promoted value which is used instead of Ext.
3625 static Value *promoteOperandForOther(Instruction *Ext,
3626 TypePromotionTransaction &TPT,
3627 InstrToOrigTy &PromotedInsts,
3628 unsigned &CreatedInstsCost,
3629 SmallVectorImpl<Instruction *> *Exts,
3630 SmallVectorImpl<Instruction *> *Truncs,
3631 const TargetLowering &TLI, bool IsSExt);
3633 /// \see promoteOperandForOther.
3634 static Value *signExtendOperandForOther(
3635 Instruction *Ext, TypePromotionTransaction &TPT,
3636 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3637 SmallVectorImpl<Instruction *> *Exts,
3638 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3639 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3640 Exts, Truncs, TLI, true);
3643 /// \see promoteOperandForOther.
3644 static Value *zeroExtendOperandForOther(
3645 Instruction *Ext, TypePromotionTransaction &TPT,
3646 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3647 SmallVectorImpl<Instruction *> *Exts,
3648 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3649 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3650 Exts, Truncs, TLI, false);
3654 /// Type for the utility function that promotes the operand of Ext.
3655 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
3656 InstrToOrigTy &PromotedInsts,
3657 unsigned &CreatedInstsCost,
3658 SmallVectorImpl<Instruction *> *Exts,
3659 SmallVectorImpl<Instruction *> *Truncs,
3660 const TargetLowering &TLI);
3661 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
3662 /// action to promote the operand of \p Ext instead of using Ext.
3663 /// \return NULL if no promotable action is possible with the current
3665 /// \p InsertedInsts keeps track of all the instructions inserted by the
3666 /// other CodeGenPrepare optimizations. This information is important
3667 /// because we do not want to promote these instructions as CodeGenPrepare
3668 /// will reinsert them later. Thus creating an infinite loop: create/remove.
3669 /// \p PromotedInsts maps the instructions to their type before promotion.
3670 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3671 const TargetLowering &TLI,
3672 const InstrToOrigTy &PromotedInsts);
3675 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3676 Type *ConsideredExtType,
3677 const InstrToOrigTy &PromotedInsts,
3679 // The promotion helper does not know how to deal with vector types yet.
3680 // To be able to fix that, we would need to fix the places where we
3681 // statically extend, e.g., constants and such.
3682 if (Inst->getType()->isVectorTy())
3685 // We can always get through zext.
3686 if (isa<ZExtInst>(Inst))
3689 // sext(sext) is ok too.
3690 if (IsSExt && isa<SExtInst>(Inst))
3693 // We can get through binary operator, if it is legal. In other words, the
3694 // binary operator must have a nuw or nsw flag.
3695 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3696 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3697 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3698 (IsSExt && BinOp->hasNoSignedWrap())))
3701 // Check if we can do the following simplification.
3702 // ext(trunc(opnd)) --> ext(opnd)
3703 if (!isa<TruncInst>(Inst))
3706 Value *OpndVal = Inst->getOperand(0);
3707 // Check if we can use this operand in the extension.
3708 // If the type is larger than the result type of the extension, we cannot.
3709 if (!OpndVal->getType()->isIntegerTy() ||
3710 OpndVal->getType()->getIntegerBitWidth() >
3711 ConsideredExtType->getIntegerBitWidth())
3714 // If the operand of the truncate is not an instruction, we will not have
3715 // any information on the dropped bits.
3716 // (Actually we could for constant but it is not worth the extra logic).
3717 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3721 // Check if the source of the type is narrow enough.
3722 // I.e., check that trunc just drops extended bits of the same kind of
3724 // #1 get the type of the operand and check the kind of the extended bits.
3725 const Type *OpndType;
3726 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3727 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
3728 OpndType = It->second.getPointer();
3729 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3730 OpndType = Opnd->getOperand(0)->getType();
3734 // #2 check that the truncate just drops extended bits.
3735 return Inst->getType()->getIntegerBitWidth() >=
3736 OpndType->getIntegerBitWidth();
3739 TypePromotionHelper::Action TypePromotionHelper::getAction(
3740 Instruction *Ext, const SetOfInstrs &InsertedInsts,
3741 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3742 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
3743 "Unexpected instruction type");
3744 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3745 Type *ExtTy = Ext->getType();
3746 bool IsSExt = isa<SExtInst>(Ext);
3747 // If the operand of the extension is not an instruction, we cannot
3749 // If it, check we can get through.
3750 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3753 // Do not promote if the operand has been added by codegenprepare.
3754 // Otherwise, it means we are undoing an optimization that is likely to be
3755 // redone, thus causing potential infinite loop.
3756 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3759 // SExt or Trunc instructions.
3760 // Return the related handler.
3761 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3762 isa<ZExtInst>(ExtOpnd))
3763 return promoteOperandForTruncAndAnyExt;
3765 // Regular instruction.
3766 // Abort early if we will have to insert non-free instructions.
3767 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3769 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3772 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3773 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
3774 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3775 SmallVectorImpl<Instruction *> *Exts,
3776 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3777 // By construction, the operand of SExt is an instruction. Otherwise we cannot
3778 // get through it and this method should not be called.
3779 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
3780 Value *ExtVal = SExt;
3781 bool HasMergedNonFreeExt = false;
3782 if (isa<ZExtInst>(SExtOpnd)) {
3783 // Replace s|zext(zext(opnd))
3785 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
3787 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
3788 TPT.replaceAllUsesWith(SExt, ZExt);
3789 TPT.eraseInstruction(SExt);
3792 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
3794 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
3796 CreatedInstsCost = 0;
3798 // Remove dead code.
3799 if (SExtOpnd->use_empty())
3800 TPT.eraseInstruction(SExtOpnd);
3802 // Check if the extension is still needed.
3803 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
3804 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
3807 Exts->push_back(ExtInst);
3808 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
3813 // At this point we have: ext ty opnd to ty.
3814 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
3815 Value *NextVal = ExtInst->getOperand(0);
3816 TPT.eraseInstruction(ExtInst, NextVal);
3820 Value *TypePromotionHelper::promoteOperandForOther(
3821 Instruction *Ext, TypePromotionTransaction &TPT,
3822 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3823 SmallVectorImpl<Instruction *> *Exts,
3824 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
3826 // By construction, the operand of Ext is an instruction. Otherwise we cannot
3827 // get through it and this method should not be called.
3828 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
3829 CreatedInstsCost = 0;
3830 if (!ExtOpnd->hasOneUse()) {
3831 // ExtOpnd will be promoted.
3832 // All its uses, but Ext, will need to use a truncated value of the
3833 // promoted version.
3834 // Create the truncate now.
3835 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
3836 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
3837 ITrunc->removeFromParent();
3838 // Insert it just after the definition.
3839 ITrunc->insertAfter(ExtOpnd);
3841 Truncs->push_back(ITrunc);
3844 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
3845 // Restore the operand of Ext (which has been replaced by the previous call
3846 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
3847 TPT.setOperand(Ext, 0, ExtOpnd);
3850 // Get through the Instruction:
3851 // 1. Update its type.
3852 // 2. Replace the uses of Ext by Inst.
3853 // 3. Extend each operand that needs to be extended.
3855 // Remember the original type of the instruction before promotion.
3856 // This is useful to know that the high bits are sign extended bits.
3857 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
3858 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
3860 TPT.mutateType(ExtOpnd, Ext->getType());
3862 TPT.replaceAllUsesWith(Ext, ExtOpnd);
3864 Instruction *ExtForOpnd = Ext;
3866 DEBUG(dbgs() << "Propagate Ext to operands\n");
3867 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
3869 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
3870 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
3871 !shouldExtOperand(ExtOpnd, OpIdx)) {
3872 DEBUG(dbgs() << "No need to propagate\n");
3875 // Check if we can statically extend the operand.
3876 Value *Opnd = ExtOpnd->getOperand(OpIdx);
3877 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
3878 DEBUG(dbgs() << "Statically extend\n");
3879 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
3880 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3881 : Cst->getValue().zext(BitWidth);
3882 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3885 // UndefValue are typed, so we have to statically sign extend them.
3886 if (isa<UndefValue>(Opnd)) {
3887 DEBUG(dbgs() << "Statically extend\n");
3888 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3892 // Otherwise we have to explicity sign extend the operand.
3893 // Check if Ext was reused to extend an operand.
3895 // If yes, create a new one.
3896 DEBUG(dbgs() << "More operands to ext\n");
3897 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3898 : TPT.createZExt(Ext, Opnd, Ext->getType());
3899 if (!isa<Instruction>(ValForExtOpnd)) {
3900 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
3903 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
3906 Exts->push_back(ExtForOpnd);
3907 TPT.setOperand(ExtForOpnd, 0, Opnd);
3909 // Move the sign extension before the insertion point.
3910 TPT.moveBefore(ExtForOpnd, ExtOpnd);
3911 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
3912 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
3913 // If more sext are required, new instructions will have to be created.
3914 ExtForOpnd = nullptr;
3916 if (ExtForOpnd == Ext) {
3917 DEBUG(dbgs() << "Extension is useless now\n");
3918 TPT.eraseInstruction(Ext);
3923 /// Check whether or not promoting an instruction to a wider type is profitable.
3924 /// \p NewCost gives the cost of extension instructions created by the
3926 /// \p OldCost gives the cost of extension instructions before the promotion
3927 /// plus the number of instructions that have been
3928 /// matched in the addressing mode the promotion.
3929 /// \p PromotedOperand is the value that has been promoted.
3930 /// \return True if the promotion is profitable, false otherwise.
3931 bool AddressingModeMatcher::isPromotionProfitable(
3932 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
3933 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
3934 // The cost of the new extensions is greater than the cost of the
3935 // old extension plus what we folded.
3936 // This is not profitable.
3937 if (NewCost > OldCost)
3939 if (NewCost < OldCost)
3941 // The promotion is neutral but it may help folding the sign extension in
3942 // loads for instance.
3943 // Check that we did not create an illegal instruction.
3944 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
3947 /// Given an instruction or constant expr, see if we can fold the operation
3948 /// into the addressing mode. If so, update the addressing mode and return
3949 /// true, otherwise return false without modifying AddrMode.
3950 /// If \p MovedAway is not NULL, it contains the information of whether or
3951 /// not AddrInst has to be folded into the addressing mode on success.
3952 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
3953 /// because it has been moved away.
3954 /// Thus AddrInst must not be added in the matched instructions.
3955 /// This state can happen when AddrInst is a sext, since it may be moved away.
3956 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
3957 /// not be referenced anymore.
3958 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
3961 // Avoid exponential behavior on extremely deep expression trees.
3962 if (Depth >= 5) return false;
3964 // By default, all matched instructions stay in place.
3969 case Instruction::PtrToInt:
3970 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3971 return matchAddr(AddrInst->getOperand(0), Depth);
3972 case Instruction::IntToPtr: {
3973 auto AS = AddrInst->getType()->getPointerAddressSpace();
3974 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3975 // This inttoptr is a no-op if the integer type is pointer sized.
3976 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3977 return matchAddr(AddrInst->getOperand(0), Depth);
3980 case Instruction::BitCast:
3981 // BitCast is always a noop, and we can handle it as long as it is
3982 // int->int or pointer->pointer (we don't want int<->fp or something).
3983 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
3984 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
3985 // Don't touch identity bitcasts. These were probably put here by LSR,
3986 // and we don't want to mess around with them. Assume it knows what it
3988 AddrInst->getOperand(0)->getType() != AddrInst->getType())
3989 return matchAddr(AddrInst->getOperand(0), Depth);
3991 case Instruction::AddrSpaceCast: {
3993 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3994 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3995 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3996 return matchAddr(AddrInst->getOperand(0), Depth);
3999 case Instruction::Add: {
4000 // Check to see if we can merge in the RHS then the LHS. If so, we win.
4001 ExtAddrMode BackupAddrMode = AddrMode;
4002 unsigned OldSize = AddrModeInsts.size();
4003 // Start a transaction at this point.
4004 // The LHS may match but not the RHS.
4005 // Therefore, we need a higher level restoration point to undo partially
4006 // matched operation.
4007 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4008 TPT.getRestorationPoint();
4010 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
4011 matchAddr(AddrInst->getOperand(0), Depth+1))
4014 // Restore the old addr mode info.
4015 AddrMode = BackupAddrMode;
4016 AddrModeInsts.resize(OldSize);
4017 TPT.rollback(LastKnownGood);
4019 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
4020 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
4021 matchAddr(AddrInst->getOperand(1), Depth+1))
4024 // Otherwise we definitely can't merge the ADD in.
4025 AddrMode = BackupAddrMode;
4026 AddrModeInsts.resize(OldSize);
4027 TPT.rollback(LastKnownGood);
4030 //case Instruction::Or:
4031 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
4033 case Instruction::Mul:
4034 case Instruction::Shl: {
4035 // Can only handle X*C and X << C.
4036 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
4039 int64_t Scale = RHS->getSExtValue();
4040 if (Opcode == Instruction::Shl)
4041 Scale = 1LL << Scale;
4043 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
4045 case Instruction::GetElementPtr: {
4046 // Scan the GEP. We check it if it contains constant offsets and at most
4047 // one variable offset.
4048 int VariableOperand = -1;
4049 unsigned VariableScale = 0;
4051 int64_t ConstantOffset = 0;
4052 gep_type_iterator GTI = gep_type_begin(AddrInst);
4053 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
4054 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
4055 const StructLayout *SL = DL.getStructLayout(STy);
4057 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
4058 ConstantOffset += SL->getElementOffset(Idx);
4060 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
4061 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
4062 ConstantOffset += CI->getSExtValue()*TypeSize;
4063 } else if (TypeSize) { // Scales of zero don't do anything.
4064 // We only allow one variable index at the moment.
4065 if (VariableOperand != -1)
4068 // Remember the variable index.
4069 VariableOperand = i;
4070 VariableScale = TypeSize;
4075 // A common case is for the GEP to only do a constant offset. In this case,
4076 // just add it to the disp field and check validity.
4077 if (VariableOperand == -1) {
4078 AddrMode.BaseOffs += ConstantOffset;
4079 if (ConstantOffset == 0 ||
4080 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
4081 // Check to see if we can fold the base pointer in too.
4082 if (matchAddr(AddrInst->getOperand(0), Depth+1))
4085 AddrMode.BaseOffs -= ConstantOffset;
4089 // Save the valid addressing mode in case we can't match.
4090 ExtAddrMode BackupAddrMode = AddrMode;
4091 unsigned OldSize = AddrModeInsts.size();
4093 // See if the scale and offset amount is valid for this target.
4094 AddrMode.BaseOffs += ConstantOffset;
4096 // Match the base operand of the GEP.
4097 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
4098 // If it couldn't be matched, just stuff the value in a register.
4099 if (AddrMode.HasBaseReg) {
4100 AddrMode = BackupAddrMode;
4101 AddrModeInsts.resize(OldSize);
4104 AddrMode.HasBaseReg = true;
4105 AddrMode.BaseReg = AddrInst->getOperand(0);
4108 // Match the remaining variable portion of the GEP.
4109 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
4111 // If it couldn't be matched, try stuffing the base into a register
4112 // instead of matching it, and retrying the match of the scale.
4113 AddrMode = BackupAddrMode;
4114 AddrModeInsts.resize(OldSize);
4115 if (AddrMode.HasBaseReg)
4117 AddrMode.HasBaseReg = true;
4118 AddrMode.BaseReg = AddrInst->getOperand(0);
4119 AddrMode.BaseOffs += ConstantOffset;
4120 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
4121 VariableScale, Depth)) {
4122 // If even that didn't work, bail.
4123 AddrMode = BackupAddrMode;
4124 AddrModeInsts.resize(OldSize);
4131 case Instruction::SExt:
4132 case Instruction::ZExt: {
4133 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
4137 // Try to move this ext out of the way of the addressing mode.
4138 // Ask for a method for doing so.
4139 TypePromotionHelper::Action TPH =
4140 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
4144 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4145 TPT.getRestorationPoint();
4146 unsigned CreatedInstsCost = 0;
4147 unsigned ExtCost = !TLI.isExtFree(Ext);
4148 Value *PromotedOperand =
4149 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
4150 // SExt has been moved away.
4151 // Thus either it will be rematched later in the recursive calls or it is
4152 // gone. Anyway, we must not fold it into the addressing mode at this point.
4156 // addr = gep base, idx
4158 // promotedOpnd = ext opnd <- no match here
4159 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
4160 // addr = gep base, op <- match
4164 assert(PromotedOperand &&
4165 "TypePromotionHelper should have filtered out those cases");
4167 ExtAddrMode BackupAddrMode = AddrMode;
4168 unsigned OldSize = AddrModeInsts.size();
4170 if (!matchAddr(PromotedOperand, Depth) ||
4171 // The total of the new cost is equal to the cost of the created
4173 // The total of the old cost is equal to the cost of the extension plus
4174 // what we have saved in the addressing mode.
4175 !isPromotionProfitable(CreatedInstsCost,
4176 ExtCost + (AddrModeInsts.size() - OldSize),
4178 AddrMode = BackupAddrMode;
4179 AddrModeInsts.resize(OldSize);
4180 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
4181 TPT.rollback(LastKnownGood);
4190 /// If we can, try to add the value of 'Addr' into the current addressing mode.
4191 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
4192 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
4195 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
4196 // Start a transaction at this point that we will rollback if the matching
4198 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4199 TPT.getRestorationPoint();
4200 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
4201 // Fold in immediates if legal for the target.
4202 AddrMode.BaseOffs += CI->getSExtValue();
4203 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4205 AddrMode.BaseOffs -= CI->getSExtValue();
4206 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
4207 // If this is a global variable, try to fold it into the addressing mode.
4208 if (!AddrMode.BaseGV) {
4209 AddrMode.BaseGV = GV;
4210 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4212 AddrMode.BaseGV = nullptr;
4214 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
4215 ExtAddrMode BackupAddrMode = AddrMode;
4216 unsigned OldSize = AddrModeInsts.size();
4218 // Check to see if it is possible to fold this operation.
4219 bool MovedAway = false;
4220 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
4221 // This instruction may have been moved away. If so, there is nothing
4225 // Okay, it's possible to fold this. Check to see if it is actually
4226 // *profitable* to do so. We use a simple cost model to avoid increasing
4227 // register pressure too much.
4228 if (I->hasOneUse() ||
4229 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
4230 AddrModeInsts.push_back(I);
4234 // It isn't profitable to do this, roll back.
4235 //cerr << "NOT FOLDING: " << *I;
4236 AddrMode = BackupAddrMode;
4237 AddrModeInsts.resize(OldSize);
4238 TPT.rollback(LastKnownGood);
4240 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
4241 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
4243 TPT.rollback(LastKnownGood);
4244 } else if (isa<ConstantPointerNull>(Addr)) {
4245 // Null pointer gets folded without affecting the addressing mode.
4249 // Worse case, the target should support [reg] addressing modes. :)
4250 if (!AddrMode.HasBaseReg) {
4251 AddrMode.HasBaseReg = true;
4252 AddrMode.BaseReg = Addr;
4253 // Still check for legality in case the target supports [imm] but not [i+r].
4254 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4256 AddrMode.HasBaseReg = false;
4257 AddrMode.BaseReg = nullptr;
4260 // If the base register is already taken, see if we can do [r+r].
4261 if (AddrMode.Scale == 0) {
4263 AddrMode.ScaledReg = Addr;
4264 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4267 AddrMode.ScaledReg = nullptr;
4270 TPT.rollback(LastKnownGood);
4274 /// Check to see if all uses of OpVal by the specified inline asm call are due
4275 /// to memory operands. If so, return true, otherwise return false.
4276 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
4277 const TargetMachine &TM) {
4278 const Function *F = CI->getParent()->getParent();
4279 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
4280 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
4281 TargetLowering::AsmOperandInfoVector TargetConstraints =
4282 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
4283 ImmutableCallSite(CI));
4284 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4285 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4287 // Compute the constraint code and ConstraintType to use.
4288 TLI->ComputeConstraintToUse(OpInfo, SDValue());
4290 // If this asm operand is our Value*, and if it isn't an indirect memory
4291 // operand, we can't fold it!
4292 if (OpInfo.CallOperandVal == OpVal &&
4293 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
4294 !OpInfo.isIndirect))
4301 /// Recursively walk all the uses of I until we find a memory use.
4302 /// If we find an obviously non-foldable instruction, return true.
4303 /// Add the ultimately found memory instructions to MemoryUses.
4304 static bool FindAllMemoryUses(
4306 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
4307 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
4308 // If we already considered this instruction, we're done.
4309 if (!ConsideredInsts.insert(I).second)
4312 // If this is an obviously unfoldable instruction, bail out.
4313 if (!MightBeFoldableInst(I))
4316 // Loop over all the uses, recursively processing them.
4317 for (Use &U : I->uses()) {
4318 Instruction *UserI = cast<Instruction>(U.getUser());
4320 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4321 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
4325 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4326 unsigned opNo = U.getOperandNo();
4327 if (opNo == 0) return true; // Storing addr, not into addr.
4328 MemoryUses.push_back(std::make_pair(SI, opNo));
4332 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4333 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
4334 if (!IA) return true;
4336 // If this is a memory operand, we're cool, otherwise bail out.
4337 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
4342 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
4349 /// Return true if Val is already known to be live at the use site that we're
4350 /// folding it into. If so, there is no cost to include it in the addressing
4351 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4352 /// instruction already.
4353 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4354 Value *KnownLive2) {
4355 // If Val is either of the known-live values, we know it is live!
4356 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4359 // All values other than instructions and arguments (e.g. constants) are live.
4360 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4362 // If Val is a constant sized alloca in the entry block, it is live, this is
4363 // true because it is just a reference to the stack/frame pointer, which is
4364 // live for the whole function.
4365 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4366 if (AI->isStaticAlloca())
4369 // Check to see if this value is already used in the memory instruction's
4370 // block. If so, it's already live into the block at the very least, so we
4371 // can reasonably fold it.
4372 return Val->isUsedInBasicBlock(MemoryInst->getParent());
4375 /// It is possible for the addressing mode of the machine to fold the specified
4376 /// instruction into a load or store that ultimately uses it.
4377 /// However, the specified instruction has multiple uses.
4378 /// Given this, it may actually increase register pressure to fold it
4379 /// into the load. For example, consider this code:
4383 /// use(Y) -> nonload/store
4387 /// In this case, Y has multiple uses, and can be folded into the load of Z
4388 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4389 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4390 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4391 /// number of computations either.
4393 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4394 /// X was live across 'load Z' for other reasons, we actually *would* want to
4395 /// fold the addressing mode in the Z case. This would make Y die earlier.
4396 bool AddressingModeMatcher::
4397 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4398 ExtAddrMode &AMAfter) {
4399 if (IgnoreProfitability) return true;
4401 // AMBefore is the addressing mode before this instruction was folded into it,
4402 // and AMAfter is the addressing mode after the instruction was folded. Get
4403 // the set of registers referenced by AMAfter and subtract out those
4404 // referenced by AMBefore: this is the set of values which folding in this
4405 // address extends the lifetime of.
4407 // Note that there are only two potential values being referenced here,
4408 // BaseReg and ScaleReg (global addresses are always available, as are any
4409 // folded immediates).
4410 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4412 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4413 // lifetime wasn't extended by adding this instruction.
4414 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4416 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4417 ScaledReg = nullptr;
4419 // If folding this instruction (and it's subexprs) didn't extend any live
4420 // ranges, we're ok with it.
4421 if (!BaseReg && !ScaledReg)
4424 // If all uses of this instruction are ultimately load/store/inlineasm's,
4425 // check to see if their addressing modes will include this instruction. If
4426 // so, we can fold it into all uses, so it doesn't matter if it has multiple
4428 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
4429 SmallPtrSet<Instruction*, 16> ConsideredInsts;
4430 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
4431 return false; // Has a non-memory, non-foldable use!
4433 // Now that we know that all uses of this instruction are part of a chain of
4434 // computation involving only operations that could theoretically be folded
4435 // into a memory use, loop over each of these uses and see if they could
4436 // *actually* fold the instruction.
4437 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4438 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4439 Instruction *User = MemoryUses[i].first;
4440 unsigned OpNo = MemoryUses[i].second;
4442 // Get the access type of this use. If the use isn't a pointer, we don't
4443 // know what it accesses.
4444 Value *Address = User->getOperand(OpNo);
4445 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4448 Type *AddressAccessTy = AddrTy->getElementType();
4449 unsigned AS = AddrTy->getAddressSpace();
4451 // Do a match against the root of this address, ignoring profitability. This
4452 // will tell us if the addressing mode for the memory operation will
4453 // *actually* cover the shared instruction.
4455 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4456 TPT.getRestorationPoint();
4457 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
4458 MemoryInst, Result, InsertedInsts,
4459 PromotedInsts, TPT);
4460 Matcher.IgnoreProfitability = true;
4461 bool Success = Matcher.matchAddr(Address, 0);
4462 (void)Success; assert(Success && "Couldn't select *anything*?");
4464 // The match was to check the profitability, the changes made are not
4465 // part of the original matcher. Therefore, they should be dropped
4466 // otherwise the original matcher will not present the right state.
4467 TPT.rollback(LastKnownGood);
4469 // If the match didn't cover I, then it won't be shared by it.
4470 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
4471 I) == MatchedAddrModeInsts.end())
4474 MatchedAddrModeInsts.clear();
4480 } // end anonymous namespace
4482 /// Return true if the specified values are defined in a
4483 /// different basic block than BB.
4484 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4485 if (Instruction *I = dyn_cast<Instruction>(V))
4486 return I->getParent() != BB;
4490 /// Load and Store Instructions often have addressing modes that can do
4491 /// significant amounts of computation. As such, instruction selection will try
4492 /// to get the load or store to do as much computation as possible for the
4493 /// program. The problem is that isel can only see within a single block. As
4494 /// such, we sink as much legal addressing mode work into the block as possible.
4496 /// This method is used to optimize both load/store and inline asms with memory
4498 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4499 Type *AccessTy, unsigned AddrSpace) {
4502 // Try to collapse single-value PHI nodes. This is necessary to undo
4503 // unprofitable PRE transformations.
4504 SmallVector<Value*, 8> worklist;
4505 SmallPtrSet<Value*, 16> Visited;
4506 worklist.push_back(Addr);
4508 // Use a worklist to iteratively look through PHI nodes, and ensure that
4509 // the addressing mode obtained from the non-PHI roots of the graph
4511 Value *Consensus = nullptr;
4512 unsigned NumUsesConsensus = 0;
4513 bool IsNumUsesConsensusValid = false;
4514 SmallVector<Instruction*, 16> AddrModeInsts;
4515 ExtAddrMode AddrMode;
4516 TypePromotionTransaction TPT;
4517 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4518 TPT.getRestorationPoint();
4519 while (!worklist.empty()) {
4520 Value *V = worklist.back();
4521 worklist.pop_back();
4523 // Break use-def graph loops.
4524 if (!Visited.insert(V).second) {
4525 Consensus = nullptr;
4529 // For a PHI node, push all of its incoming values.
4530 if (PHINode *P = dyn_cast<PHINode>(V)) {
4531 for (Value *IncValue : P->incoming_values())
4532 worklist.push_back(IncValue);
4536 // For non-PHIs, determine the addressing mode being computed.
4537 SmallVector<Instruction*, 16> NewAddrModeInsts;
4538 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4539 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
4540 InsertedInsts, PromotedInsts, TPT);
4542 // This check is broken into two cases with very similar code to avoid using
4543 // getNumUses() as much as possible. Some values have a lot of uses, so
4544 // calling getNumUses() unconditionally caused a significant compile-time
4548 AddrMode = NewAddrMode;
4549 AddrModeInsts = NewAddrModeInsts;
4551 } else if (NewAddrMode == AddrMode) {
4552 if (!IsNumUsesConsensusValid) {
4553 NumUsesConsensus = Consensus->getNumUses();
4554 IsNumUsesConsensusValid = true;
4557 // Ensure that the obtained addressing mode is equivalent to that obtained
4558 // for all other roots of the PHI traversal. Also, when choosing one
4559 // such root as representative, select the one with the most uses in order
4560 // to keep the cost modeling heuristics in AddressingModeMatcher
4562 unsigned NumUses = V->getNumUses();
4563 if (NumUses > NumUsesConsensus) {
4565 NumUsesConsensus = NumUses;
4566 AddrModeInsts = NewAddrModeInsts;
4571 Consensus = nullptr;
4575 // If the addressing mode couldn't be determined, or if multiple different
4576 // ones were determined, bail out now.
4578 TPT.rollback(LastKnownGood);
4583 // Check to see if any of the instructions supersumed by this addr mode are
4584 // non-local to I's BB.
4585 bool AnyNonLocal = false;
4586 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
4587 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
4593 // If all the instructions matched are already in this BB, don't do anything.
4595 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
4599 // Insert this computation right after this user. Since our caller is
4600 // scanning from the top of the BB to the bottom, reuse of the expr are
4601 // guaranteed to happen later.
4602 IRBuilder<> Builder(MemoryInst);
4604 // Now that we determined the addressing expression we want to use and know
4605 // that we have to sink it into this block. Check to see if we have already
4606 // done this for some other load/store instr in this block. If so, reuse the
4608 Value *&SunkAddr = SunkAddrs[Addr];
4610 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
4611 << *MemoryInst << "\n");
4612 if (SunkAddr->getType() != Addr->getType())
4613 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
4614 } else if (AddrSinkUsingGEPs ||
4615 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
4616 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
4618 // By default, we use the GEP-based method when AA is used later. This
4619 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4620 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4621 << *MemoryInst << "\n");
4622 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4623 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4625 // First, find the pointer.
4626 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4627 ResultPtr = AddrMode.BaseReg;
4628 AddrMode.BaseReg = nullptr;
4631 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4632 // We can't add more than one pointer together, nor can we scale a
4633 // pointer (both of which seem meaningless).
4634 if (ResultPtr || AddrMode.Scale != 1)
4637 ResultPtr = AddrMode.ScaledReg;
4641 if (AddrMode.BaseGV) {
4645 ResultPtr = AddrMode.BaseGV;
4648 // If the real base value actually came from an inttoptr, then the matcher
4649 // will look through it and provide only the integer value. In that case,
4651 if (!ResultPtr && AddrMode.BaseReg) {
4653 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
4654 AddrMode.BaseReg = nullptr;
4655 } else if (!ResultPtr && AddrMode.Scale == 1) {
4657 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
4662 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4663 SunkAddr = Constant::getNullValue(Addr->getType());
4664 } else if (!ResultPtr) {
4668 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4669 Type *I8Ty = Builder.getInt8Ty();
4671 // Start with the base register. Do this first so that subsequent address
4672 // matching finds it last, which will prevent it from trying to match it
4673 // as the scaled value in case it happens to be a mul. That would be
4674 // problematic if we've sunk a different mul for the scale, because then
4675 // we'd end up sinking both muls.
4676 if (AddrMode.BaseReg) {
4677 Value *V = AddrMode.BaseReg;
4678 if (V->getType() != IntPtrTy)
4679 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4684 // Add the scale value.
4685 if (AddrMode.Scale) {
4686 Value *V = AddrMode.ScaledReg;
4687 if (V->getType() == IntPtrTy) {
4689 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4690 cast<IntegerType>(V->getType())->getBitWidth()) {
4691 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4693 // It is only safe to sign extend the BaseReg if we know that the math
4694 // required to create it did not overflow before we extend it. Since
4695 // the original IR value was tossed in favor of a constant back when
4696 // the AddrMode was created we need to bail out gracefully if widths
4697 // do not match instead of extending it.
4698 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
4699 if (I && (ResultIndex != AddrMode.BaseReg))
4700 I->eraseFromParent();
4704 if (AddrMode.Scale != 1)
4705 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4708 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4713 // Add in the Base Offset if present.
4714 if (AddrMode.BaseOffs) {
4715 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4717 // We need to add this separately from the scale above to help with
4718 // SDAG consecutive load/store merging.
4719 if (ResultPtr->getType() != I8PtrTy)
4720 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
4721 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4728 SunkAddr = ResultPtr;
4730 if (ResultPtr->getType() != I8PtrTy)
4731 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
4732 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4735 if (SunkAddr->getType() != Addr->getType())
4736 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
4739 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4740 << *MemoryInst << "\n");
4741 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4742 Value *Result = nullptr;
4744 // Start with the base register. Do this first so that subsequent address
4745 // matching finds it last, which will prevent it from trying to match it
4746 // as the scaled value in case it happens to be a mul. That would be
4747 // problematic if we've sunk a different mul for the scale, because then
4748 // we'd end up sinking both muls.
4749 if (AddrMode.BaseReg) {
4750 Value *V = AddrMode.BaseReg;
4751 if (V->getType()->isPointerTy())
4752 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4753 if (V->getType() != IntPtrTy)
4754 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4758 // Add the scale value.
4759 if (AddrMode.Scale) {
4760 Value *V = AddrMode.ScaledReg;
4761 if (V->getType() == IntPtrTy) {
4763 } else if (V->getType()->isPointerTy()) {
4764 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4765 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4766 cast<IntegerType>(V->getType())->getBitWidth()) {
4767 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4769 // It is only safe to sign extend the BaseReg if we know that the math
4770 // required to create it did not overflow before we extend it. Since
4771 // the original IR value was tossed in favor of a constant back when
4772 // the AddrMode was created we need to bail out gracefully if widths
4773 // do not match instead of extending it.
4774 Instruction *I = dyn_cast_or_null<Instruction>(Result);
4775 if (I && (Result != AddrMode.BaseReg))
4776 I->eraseFromParent();
4779 if (AddrMode.Scale != 1)
4780 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4783 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4788 // Add in the BaseGV if present.
4789 if (AddrMode.BaseGV) {
4790 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
4792 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4797 // Add in the Base Offset if present.
4798 if (AddrMode.BaseOffs) {
4799 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4801 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4807 SunkAddr = Constant::getNullValue(Addr->getType());
4809 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
4812 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
4814 // If we have no uses, recursively delete the value and all dead instructions
4816 if (Repl->use_empty()) {
4817 // This can cause recursive deletion, which can invalidate our iterator.
4818 // Use a WeakVH to hold onto it in case this happens.
4819 WeakVH IterHandle(&*CurInstIterator);
4820 BasicBlock *BB = CurInstIterator->getParent();
4822 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
4824 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
4825 // If the iterator instruction was recursively deleted, start over at the
4826 // start of the block.
4827 CurInstIterator = BB->begin();
4835 /// If there are any memory operands, use OptimizeMemoryInst to sink their
4836 /// address computing into the block when possible / profitable.
4837 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
4838 bool MadeChange = false;
4840 const TargetRegisterInfo *TRI =
4841 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
4842 TargetLowering::AsmOperandInfoVector TargetConstraints =
4843 TLI->ParseConstraints(*DL, TRI, CS);
4845 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4846 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4848 // Compute the constraint code and ConstraintType to use.
4849 TLI->ComputeConstraintToUse(OpInfo, SDValue());
4851 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
4852 OpInfo.isIndirect) {
4853 Value *OpVal = CS->getArgOperand(ArgNo++);
4854 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
4855 } else if (OpInfo.Type == InlineAsm::isInput)
4862 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
4863 /// sign extensions.
4864 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
4865 assert(!Inst->use_empty() && "Input must have at least one use");
4866 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
4867 bool IsSExt = isa<SExtInst>(FirstUser);
4868 Type *ExtTy = FirstUser->getType();
4869 for (const User *U : Inst->users()) {
4870 const Instruction *UI = cast<Instruction>(U);
4871 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
4873 Type *CurTy = UI->getType();
4874 // Same input and output types: Same instruction after CSE.
4878 // If IsSExt is true, we are in this situation:
4880 // b = sext ty1 a to ty2
4881 // c = sext ty1 a to ty3
4882 // Assuming ty2 is shorter than ty3, this could be turned into:
4884 // b = sext ty1 a to ty2
4885 // c = sext ty2 b to ty3
4886 // However, the last sext is not free.
4890 // This is a ZExt, maybe this is free to extend from one type to another.
4891 // In that case, we would not account for a different use.
4894 if (ExtTy->getScalarType()->getIntegerBitWidth() >
4895 CurTy->getScalarType()->getIntegerBitWidth()) {
4903 if (!TLI.isZExtFree(NarrowTy, LargeTy))
4906 // All uses are the same or can be derived from one another for free.
4910 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
4911 /// load instruction.
4912 /// If an ext(load) can be formed, it is returned via \p LI for the load
4913 /// and \p Inst for the extension.
4914 /// Otherwise LI == nullptr and Inst == nullptr.
4915 /// When some promotion happened, \p TPT contains the proper state to
4918 /// \return true when promoting was necessary to expose the ext(load)
4919 /// opportunity, false otherwise.
4923 /// %ld = load i32* %addr
4924 /// %add = add nuw i32 %ld, 4
4925 /// %zext = zext i32 %add to i64
4929 /// %ld = load i32* %addr
4930 /// %zext = zext i32 %ld to i64
4931 /// %add = add nuw i64 %zext, 4
4933 /// Thanks to the promotion, we can match zext(load i32*) to i64.
4934 bool CodeGenPrepare::extLdPromotion(TypePromotionTransaction &TPT,
4935 LoadInst *&LI, Instruction *&Inst,
4936 const SmallVectorImpl<Instruction *> &Exts,
4937 unsigned CreatedInstsCost = 0) {
4938 // Iterate over all the extensions to see if one form an ext(load).
4939 for (auto I : Exts) {
4940 // Check if we directly have ext(load).
4941 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
4943 // No promotion happened here.
4946 // Check whether or not we want to do any promotion.
4947 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
4949 // Get the action to perform the promotion.
4950 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
4951 I, InsertedInsts, *TLI, PromotedInsts);
4952 // Check if we can promote.
4955 // Save the current state.
4956 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4957 TPT.getRestorationPoint();
4958 SmallVector<Instruction *, 4> NewExts;
4959 unsigned NewCreatedInstsCost = 0;
4960 unsigned ExtCost = !TLI->isExtFree(I);
4962 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
4963 &NewExts, nullptr, *TLI);
4964 assert(PromotedVal &&
4965 "TypePromotionHelper should have filtered out those cases");
4967 // We would be able to merge only one extension in a load.
4968 // Therefore, if we have more than 1 new extension we heuristically
4969 // cut this search path, because it means we degrade the code quality.
4970 // With exactly 2, the transformation is neutral, because we will merge
4971 // one extension but leave one. However, we optimistically keep going,
4972 // because the new extension may be removed too.
4973 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4974 TotalCreatedInstsCost -= ExtCost;
4975 if (!StressExtLdPromotion &&
4976 (TotalCreatedInstsCost > 1 ||
4977 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4978 // The promotion is not profitable, rollback to the previous state.
4979 TPT.rollback(LastKnownGood);
4982 // The promotion is profitable.
4983 // Check if it exposes an ext(load).
4984 (void)extLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
4985 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4986 // If we have created a new extension, i.e., now we have two
4987 // extensions. We must make sure one of them is merged with
4988 // the load, otherwise we may degrade the code quality.
4989 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
4990 // Promotion happened.
4992 // If this does not help to expose an ext(load) then, rollback.
4993 TPT.rollback(LastKnownGood);
4995 // None of the extension can form an ext(load).
5001 /// Move a zext or sext fed by a load into the same basic block as the load,
5002 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
5003 /// extend into the load.
5004 /// \p I[in/out] the extension may be modified during the process if some
5005 /// promotions apply.
5007 bool CodeGenPrepare::moveExtToFormExtLoad(Instruction *&I) {
5008 // Try to promote a chain of computation if it allows to form
5009 // an extended load.
5010 TypePromotionTransaction TPT;
5011 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5012 TPT.getRestorationPoint();
5013 SmallVector<Instruction *, 1> Exts;
5015 // Look for a load being extended.
5016 LoadInst *LI = nullptr;
5017 Instruction *OldExt = I;
5018 bool HasPromoted = extLdPromotion(TPT, LI, I, Exts);
5020 assert(!HasPromoted && !LI && "If we did not match any load instruction "
5021 "the code must remain the same");
5026 // If they're already in the same block, there's nothing to do.
5027 // Make the cheap checks first if we did not promote.
5028 // If we promoted, we need to check if it is indeed profitable.
5029 if (!HasPromoted && LI->getParent() == I->getParent())
5032 EVT VT = TLI->getValueType(*DL, I->getType());
5033 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
5035 // If the load has other users and the truncate is not free, this probably
5036 // isn't worthwhile.
5037 if (!LI->hasOneUse() && TLI &&
5038 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
5039 !TLI->isTruncateFree(I->getType(), LI->getType())) {
5041 TPT.rollback(LastKnownGood);
5045 // Check whether the target supports casts folded into loads.
5047 if (isa<ZExtInst>(I))
5048 LType = ISD::ZEXTLOAD;
5050 assert(isa<SExtInst>(I) && "Unexpected ext type!");
5051 LType = ISD::SEXTLOAD;
5053 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
5055 TPT.rollback(LastKnownGood);
5059 // Move the extend into the same block as the load, so that SelectionDAG
5062 I->removeFromParent();
5068 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
5069 BasicBlock *DefBB = I->getParent();
5071 // If the result of a {s|z}ext and its source are both live out, rewrite all
5072 // other uses of the source with result of extension.
5073 Value *Src = I->getOperand(0);
5074 if (Src->hasOneUse())
5077 // Only do this xform if truncating is free.
5078 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
5081 // Only safe to perform the optimization if the source is also defined in
5083 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
5086 bool DefIsLiveOut = false;
5087 for (User *U : I->users()) {
5088 Instruction *UI = cast<Instruction>(U);
5090 // Figure out which BB this ext is used in.
5091 BasicBlock *UserBB = UI->getParent();
5092 if (UserBB == DefBB) continue;
5093 DefIsLiveOut = true;
5099 // Make sure none of the uses are PHI nodes.
5100 for (User *U : Src->users()) {
5101 Instruction *UI = cast<Instruction>(U);
5102 BasicBlock *UserBB = UI->getParent();
5103 if (UserBB == DefBB) continue;
5104 // Be conservative. We don't want this xform to end up introducing
5105 // reloads just before load / store instructions.
5106 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
5110 // InsertedTruncs - Only insert one trunc in each block once.
5111 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
5113 bool MadeChange = false;
5114 for (Use &U : Src->uses()) {
5115 Instruction *User = cast<Instruction>(U.getUser());
5117 // Figure out which BB this ext is used in.
5118 BasicBlock *UserBB = User->getParent();
5119 if (UserBB == DefBB) continue;
5121 // Both src and def are live in this block. Rewrite the use.
5122 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
5124 if (!InsertedTrunc) {
5125 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5126 assert(InsertPt != UserBB->end());
5127 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5128 InsertedInsts.insert(InsertedTrunc);
5131 // Replace a use of the {s|z}ext source with a use of the result.
5140 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
5141 // just after the load if the target can fold this into one extload instruction,
5142 // with the hope of eliminating some of the other later "and" instructions using
5143 // the loaded value. "and"s that are made trivially redundant by the insertion
5144 // of the new "and" are removed by this function, while others (e.g. those whose
5145 // path from the load goes through a phi) are left for isel to potentially
5178 // becomes (after a call to optimizeLoadExt for each load):
5182 // x1' = and x1, 0xff
5186 // x2' = and x2, 0xff
5193 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5195 if (!Load->isSimple() ||
5196 !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
5199 // Skip loads we've already transformed or have no reason to transform.
5200 if (Load->hasOneUse()) {
5201 User *LoadUser = *Load->user_begin();
5202 if (cast<Instruction>(LoadUser)->getParent() == Load->getParent() &&
5203 !dyn_cast<PHINode>(LoadUser))
5207 // Look at all uses of Load, looking through phis, to determine how many bits
5208 // of the loaded value are needed.
5209 SmallVector<Instruction *, 8> WorkList;
5210 SmallPtrSet<Instruction *, 16> Visited;
5211 SmallVector<Instruction *, 8> AndsToMaybeRemove;
5212 for (auto *U : Load->users())
5213 WorkList.push_back(cast<Instruction>(U));
5215 EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5216 unsigned BitWidth = LoadResultVT.getSizeInBits();
5217 APInt DemandBits(BitWidth, 0);
5218 APInt WidestAndBits(BitWidth, 0);
5220 while (!WorkList.empty()) {
5221 Instruction *I = WorkList.back();
5222 WorkList.pop_back();
5224 // Break use-def graph loops.
5225 if (!Visited.insert(I).second)
5228 // For a PHI node, push all of its users.
5229 if (auto *Phi = dyn_cast<PHINode>(I)) {
5230 for (auto *U : Phi->users())
5231 WorkList.push_back(cast<Instruction>(U));
5235 switch (I->getOpcode()) {
5236 case llvm::Instruction::And: {
5237 auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5240 APInt AndBits = AndC->getValue();
5241 DemandBits |= AndBits;
5242 // Keep track of the widest and mask we see.
5243 if (AndBits.ugt(WidestAndBits))
5244 WidestAndBits = AndBits;
5245 if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5246 AndsToMaybeRemove.push_back(I);
5250 case llvm::Instruction::Shl: {
5251 auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5254 uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5255 auto ShlDemandBits = APInt::getAllOnesValue(BitWidth).lshr(ShiftAmt);
5256 DemandBits |= ShlDemandBits;
5260 case llvm::Instruction::Trunc: {
5261 EVT TruncVT = TLI->getValueType(*DL, I->getType());
5262 unsigned TruncBitWidth = TruncVT.getSizeInBits();
5263 auto TruncBits = APInt::getAllOnesValue(TruncBitWidth).zext(BitWidth);
5264 DemandBits |= TruncBits;
5273 uint32_t ActiveBits = DemandBits.getActiveBits();
5274 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5275 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5276 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5277 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5278 // followed by an AND.
5279 // TODO: Look into removing this restriction by fixing backends to either
5280 // return false for isLoadExtLegal for i1 or have them select this pattern to
5281 // a single instruction.
5283 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5284 // mask, since these are the only ands that will be removed by isel.
5285 if (ActiveBits <= 1 || !APIntOps::isMask(ActiveBits, DemandBits) ||
5286 WidestAndBits != DemandBits)
5289 LLVMContext &Ctx = Load->getType()->getContext();
5290 Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5291 EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5293 // Reject cases that won't be matched as extloads.
5294 if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5295 !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5298 IRBuilder<> Builder(Load->getNextNode());
5299 auto *NewAnd = dyn_cast<Instruction>(
5300 Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5302 // Replace all uses of load with new and (except for the use of load in the
5304 Load->replaceAllUsesWith(NewAnd);
5305 NewAnd->setOperand(0, Load);
5307 // Remove any and instructions that are now redundant.
5308 for (auto *And : AndsToMaybeRemove)
5309 // Check that the and mask is the same as the one we decided to put on the
5311 if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5312 And->replaceAllUsesWith(NewAnd);
5313 if (&*CurInstIterator == And)
5314 CurInstIterator = std::next(And->getIterator());
5315 And->eraseFromParent();
5323 /// Check if V (an operand of a select instruction) is an expensive instruction
5324 /// that is only used once.
5325 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5326 auto *I = dyn_cast<Instruction>(V);
5327 // If it's safe to speculatively execute, then it should not have side
5328 // effects; therefore, it's safe to sink and possibly *not* execute.
5329 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5330 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
5333 /// Returns true if a SelectInst should be turned into an explicit branch.
5334 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
5336 // FIXME: This should use the same heuristics as IfConversion to determine
5337 // whether a select is better represented as a branch. This requires that
5338 // branch probability metadata is preserved for the select, which is not the
5341 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5343 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5344 // comparison condition. If the compare has more than one use, there's
5345 // probably another cmov or setcc around, so it's not worth emitting a branch.
5346 if (!Cmp || !Cmp->hasOneUse())
5349 Value *CmpOp0 = Cmp->getOperand(0);
5350 Value *CmpOp1 = Cmp->getOperand(1);
5352 // Emit "cmov on compare with a memory operand" as a branch to avoid stalls
5353 // on a load from memory. But if the load is used more than once, do not
5354 // change the select to a branch because the load is probably needed
5355 // regardless of whether the branch is taken or not.
5356 if ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
5357 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()))
5360 // If either operand of the select is expensive and only needed on one side
5361 // of the select, we should form a branch.
5362 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5363 sinkSelectOperand(TTI, SI->getFalseValue()))
5370 /// If we have a SelectInst that will likely profit from branch prediction,
5371 /// turn it into a branch.
5372 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5373 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
5375 // Can we convert the 'select' to CF ?
5376 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
5379 TargetLowering::SelectSupportKind SelectKind;
5381 SelectKind = TargetLowering::VectorMaskSelect;
5382 else if (SI->getType()->isVectorTy())
5383 SelectKind = TargetLowering::ScalarCondVectorVal;
5385 SelectKind = TargetLowering::ScalarValSelect;
5387 // Do we have efficient codegen support for this kind of 'selects' ?
5388 if (TLI->isSelectSupported(SelectKind)) {
5389 // We have efficient codegen support for the select instruction.
5390 // Check if it is profitable to keep this 'select'.
5391 if (!TLI->isPredictableSelectExpensive() ||
5392 !isFormingBranchFromSelectProfitable(TTI, SI))
5398 // Transform a sequence like this:
5400 // %cmp = cmp uge i32 %a, %b
5401 // %sel = select i1 %cmp, i32 %c, i32 %d
5405 // %cmp = cmp uge i32 %a, %b
5406 // br i1 %cmp, label %select.true, label %select.false
5408 // br label %select.end
5410 // br label %select.end
5412 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
5414 // In addition, we may sink instructions that produce %c or %d from
5415 // the entry block into the destination(s) of the new branch.
5416 // If the true or false blocks do not contain a sunken instruction, that
5417 // block and its branch may be optimized away. In that case, one side of the
5418 // first branch will point directly to select.end, and the corresponding PHI
5419 // predecessor block will be the start block.
5421 // First, we split the block containing the select into 2 blocks.
5422 BasicBlock *StartBlock = SI->getParent();
5423 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
5424 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
5426 // Delete the unconditional branch that was just created by the split.
5427 StartBlock->getTerminator()->eraseFromParent();
5429 // These are the new basic blocks for the conditional branch.
5430 // At least one will become an actual new basic block.
5431 BasicBlock *TrueBlock = nullptr;
5432 BasicBlock *FalseBlock = nullptr;
5434 // Sink expensive instructions into the conditional blocks to avoid executing
5435 // them speculatively.
5436 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
5437 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
5438 EndBlock->getParent(), EndBlock);
5439 auto *TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
5440 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
5441 TrueInst->moveBefore(TrueBranch);
5443 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
5444 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
5445 EndBlock->getParent(), EndBlock);
5446 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5447 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
5448 FalseInst->moveBefore(FalseBranch);
5451 // If there was nothing to sink, then arbitrarily choose the 'false' side
5452 // for a new input value to the PHI.
5453 if (TrueBlock == FalseBlock) {
5454 assert(TrueBlock == nullptr &&
5455 "Unexpected basic block transform while optimizing select");
5457 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
5458 EndBlock->getParent(), EndBlock);
5459 BranchInst::Create(EndBlock, FalseBlock);
5462 // Insert the real conditional branch based on the original condition.
5463 // If we did not create a new block for one of the 'true' or 'false' paths
5464 // of the condition, it means that side of the branch goes to the end block
5465 // directly and the path originates from the start block from the point of
5466 // view of the new PHI.
5467 if (TrueBlock == nullptr) {
5468 BranchInst::Create(EndBlock, FalseBlock, SI->getCondition(), SI);
5469 TrueBlock = StartBlock;
5470 } else if (FalseBlock == nullptr) {
5471 BranchInst::Create(TrueBlock, EndBlock, SI->getCondition(), SI);
5472 FalseBlock = StartBlock;
5474 BranchInst::Create(TrueBlock, FalseBlock, SI->getCondition(), SI);
5477 // The select itself is replaced with a PHI Node.
5478 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
5480 PN->addIncoming(SI->getTrueValue(), TrueBlock);
5481 PN->addIncoming(SI->getFalseValue(), FalseBlock);
5483 SI->replaceAllUsesWith(PN);
5484 SI->eraseFromParent();
5486 // Instruct OptimizeBlock to skip to the next block.
5487 CurInstIterator = StartBlock->end();
5488 ++NumSelectsExpanded;
5492 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
5493 SmallVector<int, 16> Mask(SVI->getShuffleMask());
5495 for (unsigned i = 0; i < Mask.size(); ++i) {
5496 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
5498 SplatElem = Mask[i];
5504 /// Some targets have expensive vector shifts if the lanes aren't all the same
5505 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
5506 /// it's often worth sinking a shufflevector splat down to its use so that
5507 /// codegen can spot all lanes are identical.
5508 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
5509 BasicBlock *DefBB = SVI->getParent();
5511 // Only do this xform if variable vector shifts are particularly expensive.
5512 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
5515 // We only expect better codegen by sinking a shuffle if we can recognise a
5517 if (!isBroadcastShuffle(SVI))
5520 // InsertedShuffles - Only insert a shuffle in each block once.
5521 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
5523 bool MadeChange = false;
5524 for (User *U : SVI->users()) {
5525 Instruction *UI = cast<Instruction>(U);
5527 // Figure out which BB this ext is used in.
5528 BasicBlock *UserBB = UI->getParent();
5529 if (UserBB == DefBB) continue;
5531 // For now only apply this when the splat is used by a shift instruction.
5532 if (!UI->isShift()) continue;
5534 // Everything checks out, sink the shuffle if the user's block doesn't
5535 // already have a copy.
5536 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
5538 if (!InsertedShuffle) {
5539 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5540 assert(InsertPt != UserBB->end());
5542 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5543 SVI->getOperand(2), "", &*InsertPt);
5546 UI->replaceUsesOfWith(SVI, InsertedShuffle);
5550 // If we removed all uses, nuke the shuffle.
5551 if (SVI->use_empty()) {
5552 SVI->eraseFromParent();
5559 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
5563 Value *Cond = SI->getCondition();
5564 Type *OldType = Cond->getType();
5565 LLVMContext &Context = Cond->getContext();
5566 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
5567 unsigned RegWidth = RegType.getSizeInBits();
5569 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
5572 // If the register width is greater than the type width, expand the condition
5573 // of the switch instruction and each case constant to the width of the
5574 // register. By widening the type of the switch condition, subsequent
5575 // comparisons (for case comparisons) will not need to be extended to the
5576 // preferred register width, so we will potentially eliminate N-1 extends,
5577 // where N is the number of cases in the switch.
5578 auto *NewType = Type::getIntNTy(Context, RegWidth);
5580 // Zero-extend the switch condition and case constants unless the switch
5581 // condition is a function argument that is already being sign-extended.
5582 // In that case, we can avoid an unnecessary mask/extension by sign-extending
5583 // everything instead.
5584 Instruction::CastOps ExtType = Instruction::ZExt;
5585 if (auto *Arg = dyn_cast<Argument>(Cond))
5586 if (Arg->hasSExtAttr())
5587 ExtType = Instruction::SExt;
5589 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
5590 ExtInst->insertBefore(SI);
5591 SI->setCondition(ExtInst);
5592 for (SwitchInst::CaseIt Case : SI->cases()) {
5593 APInt NarrowConst = Case.getCaseValue()->getValue();
5594 APInt WideConst = (ExtType == Instruction::ZExt) ?
5595 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
5596 Case.setValue(ConstantInt::get(Context, WideConst));
5603 /// \brief Helper class to promote a scalar operation to a vector one.
5604 /// This class is used to move downward extractelement transition.
5606 /// a = vector_op <2 x i32>
5607 /// b = extractelement <2 x i32> a, i32 0
5612 /// a = vector_op <2 x i32>
5613 /// c = vector_op a (equivalent to scalar_op on the related lane)
5614 /// * d = extractelement <2 x i32> c, i32 0
5616 /// Assuming both extractelement and store can be combine, we get rid of the
5618 class VectorPromoteHelper {
5619 /// DataLayout associated with the current module.
5620 const DataLayout &DL;
5622 /// Used to perform some checks on the legality of vector operations.
5623 const TargetLowering &TLI;
5625 /// Used to estimated the cost of the promoted chain.
5626 const TargetTransformInfo &TTI;
5628 /// The transition being moved downwards.
5629 Instruction *Transition;
5630 /// The sequence of instructions to be promoted.
5631 SmallVector<Instruction *, 4> InstsToBePromoted;
5632 /// Cost of combining a store and an extract.
5633 unsigned StoreExtractCombineCost;
5634 /// Instruction that will be combined with the transition.
5635 Instruction *CombineInst;
5637 /// \brief The instruction that represents the current end of the transition.
5638 /// Since we are faking the promotion until we reach the end of the chain
5639 /// of computation, we need a way to get the current end of the transition.
5640 Instruction *getEndOfTransition() const {
5641 if (InstsToBePromoted.empty())
5643 return InstsToBePromoted.back();
5646 /// \brief Return the index of the original value in the transition.
5647 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
5648 /// c, is at index 0.
5649 unsigned getTransitionOriginalValueIdx() const {
5650 assert(isa<ExtractElementInst>(Transition) &&
5651 "Other kind of transitions are not supported yet");
5655 /// \brief Return the index of the index in the transition.
5656 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
5658 unsigned getTransitionIdx() const {
5659 assert(isa<ExtractElementInst>(Transition) &&
5660 "Other kind of transitions are not supported yet");
5664 /// \brief Get the type of the transition.
5665 /// This is the type of the original value.
5666 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
5667 /// transition is <2 x i32>.
5668 Type *getTransitionType() const {
5669 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
5672 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
5673 /// I.e., we have the following sequence:
5674 /// Def = Transition <ty1> a to <ty2>
5675 /// b = ToBePromoted <ty2> Def, ...
5677 /// b = ToBePromoted <ty1> a, ...
5678 /// Def = Transition <ty1> ToBePromoted to <ty2>
5679 void promoteImpl(Instruction *ToBePromoted);
5681 /// \brief Check whether or not it is profitable to promote all the
5682 /// instructions enqueued to be promoted.
5683 bool isProfitableToPromote() {
5684 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
5685 unsigned Index = isa<ConstantInt>(ValIdx)
5686 ? cast<ConstantInt>(ValIdx)->getZExtValue()
5688 Type *PromotedType = getTransitionType();
5690 StoreInst *ST = cast<StoreInst>(CombineInst);
5691 unsigned AS = ST->getPointerAddressSpace();
5692 unsigned Align = ST->getAlignment();
5693 // Check if this store is supported.
5694 if (!TLI.allowsMisalignedMemoryAccesses(
5695 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
5697 // If this is not supported, there is no way we can combine
5698 // the extract with the store.
5702 // The scalar chain of computation has to pay for the transition
5703 // scalar to vector.
5704 // The vector chain has to account for the combining cost.
5705 uint64_t ScalarCost =
5706 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
5707 uint64_t VectorCost = StoreExtractCombineCost;
5708 for (const auto &Inst : InstsToBePromoted) {
5709 // Compute the cost.
5710 // By construction, all instructions being promoted are arithmetic ones.
5711 // Moreover, one argument is a constant that can be viewed as a splat
5713 Value *Arg0 = Inst->getOperand(0);
5714 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
5715 isa<ConstantFP>(Arg0);
5716 TargetTransformInfo::OperandValueKind Arg0OVK =
5717 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
5718 : TargetTransformInfo::OK_AnyValue;
5719 TargetTransformInfo::OperandValueKind Arg1OVK =
5720 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
5721 : TargetTransformInfo::OK_AnyValue;
5722 ScalarCost += TTI.getArithmeticInstrCost(
5723 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
5724 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
5727 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
5728 << ScalarCost << "\nVector: " << VectorCost << '\n');
5729 return ScalarCost > VectorCost;
5732 /// \brief Generate a constant vector with \p Val with the same
5733 /// number of elements as the transition.
5734 /// \p UseSplat defines whether or not \p Val should be replicated
5735 /// across the whole vector.
5736 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
5737 /// otherwise we generate a vector with as many undef as possible:
5738 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
5739 /// used at the index of the extract.
5740 Value *getConstantVector(Constant *Val, bool UseSplat) const {
5741 unsigned ExtractIdx = UINT_MAX;
5743 // If we cannot determine where the constant must be, we have to
5744 // use a splat constant.
5745 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
5746 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
5747 ExtractIdx = CstVal->getSExtValue();
5752 unsigned End = getTransitionType()->getVectorNumElements();
5754 return ConstantVector::getSplat(End, Val);
5756 SmallVector<Constant *, 4> ConstVec;
5757 UndefValue *UndefVal = UndefValue::get(Val->getType());
5758 for (unsigned Idx = 0; Idx != End; ++Idx) {
5759 if (Idx == ExtractIdx)
5760 ConstVec.push_back(Val);
5762 ConstVec.push_back(UndefVal);
5764 return ConstantVector::get(ConstVec);
5767 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
5768 /// in \p Use can trigger undefined behavior.
5769 static bool canCauseUndefinedBehavior(const Instruction *Use,
5770 unsigned OperandIdx) {
5771 // This is not safe to introduce undef when the operand is on
5772 // the right hand side of a division-like instruction.
5773 if (OperandIdx != 1)
5775 switch (Use->getOpcode()) {
5778 case Instruction::SDiv:
5779 case Instruction::UDiv:
5780 case Instruction::SRem:
5781 case Instruction::URem:
5783 case Instruction::FDiv:
5784 case Instruction::FRem:
5785 return !Use->hasNoNaNs();
5787 llvm_unreachable(nullptr);
5791 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
5792 const TargetTransformInfo &TTI, Instruction *Transition,
5793 unsigned CombineCost)
5794 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
5795 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
5796 assert(Transition && "Do not know how to promote null");
5799 /// \brief Check if we can promote \p ToBePromoted to \p Type.
5800 bool canPromote(const Instruction *ToBePromoted) const {
5801 // We could support CastInst too.
5802 return isa<BinaryOperator>(ToBePromoted);
5805 /// \brief Check if it is profitable to promote \p ToBePromoted
5806 /// by moving downward the transition through.
5807 bool shouldPromote(const Instruction *ToBePromoted) const {
5808 // Promote only if all the operands can be statically expanded.
5809 // Indeed, we do not want to introduce any new kind of transitions.
5810 for (const Use &U : ToBePromoted->operands()) {
5811 const Value *Val = U.get();
5812 if (Val == getEndOfTransition()) {
5813 // If the use is a division and the transition is on the rhs,
5814 // we cannot promote the operation, otherwise we may create a
5815 // division by zero.
5816 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
5820 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
5821 !isa<ConstantFP>(Val))
5824 // Check that the resulting operation is legal.
5825 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
5828 return StressStoreExtract ||
5829 TLI.isOperationLegalOrCustom(
5830 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
5833 /// \brief Check whether or not \p Use can be combined
5834 /// with the transition.
5835 /// I.e., is it possible to do Use(Transition) => AnotherUse?
5836 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
5838 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
5839 void enqueueForPromotion(Instruction *ToBePromoted) {
5840 InstsToBePromoted.push_back(ToBePromoted);
5843 /// \brief Set the instruction that will be combined with the transition.
5844 void recordCombineInstruction(Instruction *ToBeCombined) {
5845 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
5846 CombineInst = ToBeCombined;
5849 /// \brief Promote all the instructions enqueued for promotion if it is
5851 /// \return True if the promotion happened, false otherwise.
5853 // Check if there is something to promote.
5854 // Right now, if we do not have anything to combine with,
5855 // we assume the promotion is not profitable.
5856 if (InstsToBePromoted.empty() || !CombineInst)
5860 if (!StressStoreExtract && !isProfitableToPromote())
5864 for (auto &ToBePromoted : InstsToBePromoted)
5865 promoteImpl(ToBePromoted);
5866 InstsToBePromoted.clear();
5870 } // End of anonymous namespace.
5872 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
5873 // At this point, we know that all the operands of ToBePromoted but Def
5874 // can be statically promoted.
5875 // For Def, we need to use its parameter in ToBePromoted:
5876 // b = ToBePromoted ty1 a
5877 // Def = Transition ty1 b to ty2
5878 // Move the transition down.
5879 // 1. Replace all uses of the promoted operation by the transition.
5880 // = ... b => = ... Def.
5881 assert(ToBePromoted->getType() == Transition->getType() &&
5882 "The type of the result of the transition does not match "
5884 ToBePromoted->replaceAllUsesWith(Transition);
5885 // 2. Update the type of the uses.
5886 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
5887 Type *TransitionTy = getTransitionType();
5888 ToBePromoted->mutateType(TransitionTy);
5889 // 3. Update all the operands of the promoted operation with promoted
5891 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
5892 for (Use &U : ToBePromoted->operands()) {
5893 Value *Val = U.get();
5894 Value *NewVal = nullptr;
5895 if (Val == Transition)
5896 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
5897 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
5898 isa<ConstantFP>(Val)) {
5899 // Use a splat constant if it is not safe to use undef.
5900 NewVal = getConstantVector(
5901 cast<Constant>(Val),
5902 isa<UndefValue>(Val) ||
5903 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
5905 llvm_unreachable("Did you modified shouldPromote and forgot to update "
5907 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
5909 Transition->removeFromParent();
5910 Transition->insertAfter(ToBePromoted);
5911 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
5914 /// Some targets can do store(extractelement) with one instruction.
5915 /// Try to push the extractelement towards the stores when the target
5916 /// has this feature and this is profitable.
5917 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
5918 unsigned CombineCost = UINT_MAX;
5919 if (DisableStoreExtract || !TLI ||
5920 (!StressStoreExtract &&
5921 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
5922 Inst->getOperand(1), CombineCost)))
5925 // At this point we know that Inst is a vector to scalar transition.
5926 // Try to move it down the def-use chain, until:
5927 // - We can combine the transition with its single use
5928 // => we got rid of the transition.
5929 // - We escape the current basic block
5930 // => we would need to check that we are moving it at a cheaper place and
5931 // we do not do that for now.
5932 BasicBlock *Parent = Inst->getParent();
5933 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
5934 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
5935 // If the transition has more than one use, assume this is not going to be
5937 while (Inst->hasOneUse()) {
5938 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
5939 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
5941 if (ToBePromoted->getParent() != Parent) {
5942 DEBUG(dbgs() << "Instruction to promote is in a different block ("
5943 << ToBePromoted->getParent()->getName()
5944 << ") than the transition (" << Parent->getName() << ").\n");
5948 if (VPH.canCombine(ToBePromoted)) {
5949 DEBUG(dbgs() << "Assume " << *Inst << '\n'
5950 << "will be combined with: " << *ToBePromoted << '\n');
5951 VPH.recordCombineInstruction(ToBePromoted);
5952 bool Changed = VPH.promote();
5953 NumStoreExtractExposed += Changed;
5957 DEBUG(dbgs() << "Try promoting.\n");
5958 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
5961 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
5963 VPH.enqueueForPromotion(ToBePromoted);
5964 Inst = ToBePromoted;
5969 bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
5970 // Bail out if we inserted the instruction to prevent optimizations from
5971 // stepping on each other's toes.
5972 if (InsertedInsts.count(I))
5975 if (PHINode *P = dyn_cast<PHINode>(I)) {
5976 // It is possible for very late stage optimizations (such as SimplifyCFG)
5977 // to introduce PHI nodes too late to be cleaned up. If we detect such a
5978 // trivial PHI, go ahead and zap it here.
5979 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
5980 P->replaceAllUsesWith(V);
5981 P->eraseFromParent();
5988 if (CastInst *CI = dyn_cast<CastInst>(I)) {
5989 // If the source of the cast is a constant, then this should have
5990 // already been constant folded. The only reason NOT to constant fold
5991 // it is if something (e.g. LSR) was careful to place the constant
5992 // evaluation in a block other than then one that uses it (e.g. to hoist
5993 // the address of globals out of a loop). If this is the case, we don't
5994 // want to forward-subst the cast.
5995 if (isa<Constant>(CI->getOperand(0)))
5998 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
6001 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6002 /// Sink a zext or sext into its user blocks if the target type doesn't
6003 /// fit in one register
6005 TLI->getTypeAction(CI->getContext(),
6006 TLI->getValueType(*DL, CI->getType())) ==
6007 TargetLowering::TypeExpandInteger) {
6008 return SinkCast(CI);
6010 bool MadeChange = moveExtToFormExtLoad(I);
6011 return MadeChange | optimizeExtUses(I);
6017 if (CmpInst *CI = dyn_cast<CmpInst>(I))
6018 if (!TLI || !TLI->hasMultipleConditionRegisters())
6019 return OptimizeCmpExpression(CI);
6021 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
6022 stripInvariantGroupMetadata(*LI);
6024 bool Modified = optimizeLoadExt(LI);
6025 unsigned AS = LI->getPointerAddressSpace();
6026 Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
6032 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
6033 stripInvariantGroupMetadata(*SI);
6035 unsigned AS = SI->getPointerAddressSpace();
6036 return optimizeMemoryInst(I, SI->getOperand(1),
6037 SI->getOperand(0)->getType(), AS);
6042 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
6044 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
6045 BinOp->getOpcode() == Instruction::LShr)) {
6046 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
6047 if (TLI && CI && TLI->hasExtractBitsInsn())
6048 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
6053 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
6054 if (GEPI->hasAllZeroIndices()) {
6055 /// The GEP operand must be a pointer, so must its result -> BitCast
6056 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
6057 GEPI->getName(), GEPI);
6058 GEPI->replaceAllUsesWith(NC);
6059 GEPI->eraseFromParent();
6061 optimizeInst(NC, ModifiedDT);
6067 if (CallInst *CI = dyn_cast<CallInst>(I))
6068 return optimizeCallInst(CI, ModifiedDT);
6070 if (SelectInst *SI = dyn_cast<SelectInst>(I))
6071 return optimizeSelectInst(SI);
6073 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
6074 return optimizeShuffleVectorInst(SVI);
6076 if (auto *Switch = dyn_cast<SwitchInst>(I))
6077 return optimizeSwitchInst(Switch);
6079 if (isa<ExtractElementInst>(I))
6080 return optimizeExtractElementInst(I);
6085 /// Given an OR instruction, check to see if this is a bitreverse
6086 /// idiom. If so, insert the new intrinsic and return true.
6087 static bool makeBitReverse(Instruction &I, const DataLayout &DL,
6088 const TargetLowering &TLI) {
6089 if (!I.getType()->isIntegerTy() ||
6090 !TLI.isOperationLegalOrCustom(ISD::BITREVERSE,
6091 TLI.getValueType(DL, I.getType(), true)))
6094 SmallVector<Instruction*, 4> Insts;
6095 if (!recognizeBitReverseOrBSwapIdiom(&I, false, true, Insts))
6097 Instruction *LastInst = Insts.back();
6098 I.replaceAllUsesWith(LastInst);
6099 RecursivelyDeleteTriviallyDeadInstructions(&I);
6103 // In this pass we look for GEP and cast instructions that are used
6104 // across basic blocks and rewrite them to improve basic-block-at-a-time
6106 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
6108 bool MadeChange = false;
6110 CurInstIterator = BB.begin();
6111 while (CurInstIterator != BB.end()) {
6112 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
6117 bool MadeBitReverse = true;
6118 while (TLI && MadeBitReverse) {
6119 MadeBitReverse = false;
6120 for (auto &I : reverse(BB)) {
6121 if (makeBitReverse(I, *DL, *TLI)) {
6122 MadeBitReverse = MadeChange = true;
6127 MadeChange |= dupRetToEnableTailCallOpts(&BB);
6132 // llvm.dbg.value is far away from the value then iSel may not be able
6133 // handle it properly. iSel will drop llvm.dbg.value if it can not
6134 // find a node corresponding to the value.
6135 bool CodeGenPrepare::placeDbgValues(Function &F) {
6136 bool MadeChange = false;
6137 for (BasicBlock &BB : F) {
6138 Instruction *PrevNonDbgInst = nullptr;
6139 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
6140 Instruction *Insn = &*BI++;
6141 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
6142 // Leave dbg.values that refer to an alloca alone. These
6143 // instrinsics describe the address of a variable (= the alloca)
6144 // being taken. They should not be moved next to the alloca
6145 // (and to the beginning of the scope), but rather stay close to
6146 // where said address is used.
6147 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
6148 PrevNonDbgInst = Insn;
6152 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
6153 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
6154 // If VI is a phi in a block with an EHPad terminator, we can't insert
6156 if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
6158 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
6159 DVI->removeFromParent();
6160 if (isa<PHINode>(VI))
6161 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
6163 DVI->insertAfter(VI);
6172 // If there is a sequence that branches based on comparing a single bit
6173 // against zero that can be combined into a single instruction, and the
6174 // target supports folding these into a single instruction, sink the
6175 // mask and compare into the branch uses. Do this before OptimizeBlock ->
6176 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
6178 bool CodeGenPrepare::sinkAndCmp(Function &F) {
6179 if (!EnableAndCmpSinking)
6181 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
6183 bool MadeChange = false;
6184 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
6185 BasicBlock *BB = &*I++;
6187 // Does this BB end with the following?
6188 // %andVal = and %val, #single-bit-set
6189 // %icmpVal = icmp %andResult, 0
6190 // br i1 %cmpVal label %dest1, label %dest2"
6191 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
6192 if (!Brcc || !Brcc->isConditional())
6194 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
6195 if (!Cmp || Cmp->getParent() != BB)
6197 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
6198 if (!Zero || !Zero->isZero())
6200 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
6201 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
6203 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
6204 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
6206 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
6208 // Push the "and; icmp" for any users that are conditional branches.
6209 // Since there can only be one branch use per BB, we don't need to keep
6210 // track of which BBs we insert into.
6211 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
6215 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
6217 if (!BrccUser || !BrccUser->isConditional())
6219 BasicBlock *UserBB = BrccUser->getParent();
6220 if (UserBB == BB) continue;
6221 DEBUG(dbgs() << "found Brcc use\n");
6223 // Sink the "and; icmp" to use.
6225 BinaryOperator *NewAnd =
6226 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
6229 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
6233 DEBUG(BrccUser->getParent()->dump());
6239 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
6240 /// success, or returns false if no or invalid metadata was found.
6241 static bool extractBranchMetadata(BranchInst *BI,
6242 uint64_t &ProbTrue, uint64_t &ProbFalse) {
6243 assert(BI->isConditional() &&
6244 "Looking for probabilities on unconditional branch?");
6245 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
6246 if (!ProfileData || ProfileData->getNumOperands() != 3)
6249 const auto *CITrue =
6250 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
6251 const auto *CIFalse =
6252 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
6253 if (!CITrue || !CIFalse)
6256 ProbTrue = CITrue->getValue().getZExtValue();
6257 ProbFalse = CIFalse->getValue().getZExtValue();
6262 /// \brief Scale down both weights to fit into uint32_t.
6263 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
6264 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
6265 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
6266 NewTrue = NewTrue / Scale;
6267 NewFalse = NewFalse / Scale;
6270 /// \brief Some targets prefer to split a conditional branch like:
6272 /// %0 = icmp ne i32 %a, 0
6273 /// %1 = icmp ne i32 %b, 0
6274 /// %or.cond = or i1 %0, %1
6275 /// br i1 %or.cond, label %TrueBB, label %FalseBB
6277 /// into multiple branch instructions like:
6280 /// %0 = icmp ne i32 %a, 0
6281 /// br i1 %0, label %TrueBB, label %bb2
6283 /// %1 = icmp ne i32 %b, 0
6284 /// br i1 %1, label %TrueBB, label %FalseBB
6286 /// This usually allows instruction selection to do even further optimizations
6287 /// and combine the compare with the branch instruction. Currently this is
6288 /// applied for targets which have "cheap" jump instructions.
6290 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
6292 bool CodeGenPrepare::splitBranchCondition(Function &F) {
6293 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
6296 bool MadeChange = false;
6297 for (auto &BB : F) {
6298 // Does this BB end with the following?
6299 // %cond1 = icmp|fcmp|binary instruction ...
6300 // %cond2 = icmp|fcmp|binary instruction ...
6301 // %cond.or = or|and i1 %cond1, cond2
6302 // br i1 %cond.or label %dest1, label %dest2"
6303 BinaryOperator *LogicOp;
6304 BasicBlock *TBB, *FBB;
6305 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
6308 auto *Br1 = cast<BranchInst>(BB.getTerminator());
6309 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
6313 Value *Cond1, *Cond2;
6314 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
6315 m_OneUse(m_Value(Cond2)))))
6316 Opc = Instruction::And;
6317 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
6318 m_OneUse(m_Value(Cond2)))))
6319 Opc = Instruction::Or;
6323 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
6324 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
6327 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
6330 auto *InsertBefore = std::next(Function::iterator(BB))
6331 .getNodePtrUnchecked();
6332 auto TmpBB = BasicBlock::Create(BB.getContext(),
6333 BB.getName() + ".cond.split",
6334 BB.getParent(), InsertBefore);
6336 // Update original basic block by using the first condition directly by the
6337 // branch instruction and removing the no longer needed and/or instruction.
6338 Br1->setCondition(Cond1);
6339 LogicOp->eraseFromParent();
6341 // Depending on the conditon we have to either replace the true or the false
6342 // successor of the original branch instruction.
6343 if (Opc == Instruction::And)
6344 Br1->setSuccessor(0, TmpBB);
6346 Br1->setSuccessor(1, TmpBB);
6348 // Fill in the new basic block.
6349 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
6350 if (auto *I = dyn_cast<Instruction>(Cond2)) {
6351 I->removeFromParent();
6352 I->insertBefore(Br2);
6355 // Update PHI nodes in both successors. The original BB needs to be
6356 // replaced in one succesor's PHI nodes, because the branch comes now from
6357 // the newly generated BB (NewBB). In the other successor we need to add one
6358 // incoming edge to the PHI nodes, because both branch instructions target
6359 // now the same successor. Depending on the original branch condition
6360 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
6361 // we perfrom the correct update for the PHI nodes.
6362 // This doesn't change the successor order of the just created branch
6363 // instruction (or any other instruction).
6364 if (Opc == Instruction::Or)
6365 std::swap(TBB, FBB);
6367 // Replace the old BB with the new BB.
6368 for (auto &I : *TBB) {
6369 PHINode *PN = dyn_cast<PHINode>(&I);
6373 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
6374 PN->setIncomingBlock(i, TmpBB);
6377 // Add another incoming edge form the new BB.
6378 for (auto &I : *FBB) {
6379 PHINode *PN = dyn_cast<PHINode>(&I);
6382 auto *Val = PN->getIncomingValueForBlock(&BB);
6383 PN->addIncoming(Val, TmpBB);
6386 // Update the branch weights (from SelectionDAGBuilder::
6387 // FindMergedConditions).
6388 if (Opc == Instruction::Or) {
6389 // Codegen X | Y as:
6398 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
6399 // The requirement is that
6400 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
6401 // = TrueProb for orignal BB.
6402 // Assuming the orignal weights are A and B, one choice is to set BB1's
6403 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
6405 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
6406 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
6407 // TmpBB, but the math is more complicated.
6408 uint64_t TrueWeight, FalseWeight;
6409 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
6410 uint64_t NewTrueWeight = TrueWeight;
6411 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
6412 scaleWeights(NewTrueWeight, NewFalseWeight);
6413 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6414 .createBranchWeights(TrueWeight, FalseWeight));
6416 NewTrueWeight = TrueWeight;
6417 NewFalseWeight = 2 * FalseWeight;
6418 scaleWeights(NewTrueWeight, NewFalseWeight);
6419 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6420 .createBranchWeights(TrueWeight, FalseWeight));
6423 // Codegen X & Y as:
6431 // This requires creation of TmpBB after CurBB.
6433 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
6434 // The requirement is that
6435 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
6436 // = FalseProb for orignal BB.
6437 // Assuming the orignal weights are A and B, one choice is to set BB1's
6438 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
6440 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
6441 uint64_t TrueWeight, FalseWeight;
6442 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
6443 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
6444 uint64_t NewFalseWeight = FalseWeight;
6445 scaleWeights(NewTrueWeight, NewFalseWeight);
6446 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6447 .createBranchWeights(TrueWeight, FalseWeight));
6449 NewTrueWeight = 2 * TrueWeight;
6450 NewFalseWeight = FalseWeight;
6451 scaleWeights(NewTrueWeight, NewFalseWeight);
6452 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6453 .createBranchWeights(TrueWeight, FalseWeight));
6457 // Note: No point in getting fancy here, since the DT info is never
6458 // available to CodeGenPrepare.
6463 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
6469 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
6470 if (auto *InvariantMD = I.getMetadata(LLVMContext::MD_invariant_group))
6471 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());