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/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/IR/CallSite.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/GetElementPtrTypeIterator.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/InlineAsm.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/PatternMatch.h"
37 #include "llvm/IR/Statepoint.h"
38 #include "llvm/IR/ValueHandle.h"
39 #include "llvm/IR/ValueMap.h"
40 #include "llvm/Pass.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Target/TargetLowering.h"
45 #include "llvm/Target/TargetSubtargetInfo.h"
46 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
47 #include "llvm/Transforms/Utils/BuildLibCalls.h"
48 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
52 using namespace llvm::PatternMatch;
54 #define DEBUG_TYPE "codegenprepare"
56 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
57 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
58 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
59 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
61 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
63 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
64 "computations were sunk");
65 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
66 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
67 STATISTIC(NumAndsAdded,
68 "Number of and mask instructions added to form ext loads");
69 STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
70 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
71 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
72 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
73 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
74 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
76 static cl::opt<bool> DisableBranchOpts(
77 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
78 cl::desc("Disable branch optimizations in CodeGenPrepare"));
81 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
82 cl::desc("Disable GC optimizations in CodeGenPrepare"));
84 static cl::opt<bool> DisableSelectToBranch(
85 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
86 cl::desc("Disable select to branch conversion."));
88 static cl::opt<bool> AddrSinkUsingGEPs(
89 "addr-sink-using-gep", cl::Hidden, cl::init(false),
90 cl::desc("Address sinking in CGP using GEPs."));
92 static cl::opt<bool> EnableAndCmpSinking(
93 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
94 cl::desc("Enable sinkinig and/cmp into branches."));
96 static cl::opt<bool> DisableStoreExtract(
97 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
98 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
100 static cl::opt<bool> StressStoreExtract(
101 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
102 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
104 static cl::opt<bool> DisableExtLdPromotion(
105 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
106 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
109 static cl::opt<bool> StressExtLdPromotion(
110 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
111 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
112 "optimization in CodeGenPrepare"));
115 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
116 typedef PointerIntPair<Type *, 1, bool> TypeIsSExt;
117 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
118 class TypePromotionTransaction;
120 class CodeGenPrepare : public FunctionPass {
121 const TargetMachine *TM;
122 const TargetLowering *TLI;
123 const TargetTransformInfo *TTI;
124 const TargetLibraryInfo *TLInfo;
126 /// As we scan instructions optimizing them, this is the next instruction
127 /// to optimize. Transforms that can invalidate this should update it.
128 BasicBlock::iterator CurInstIterator;
130 /// Keeps track of non-local addresses that have been sunk into a block.
131 /// This allows us to avoid inserting duplicate code for blocks with
132 /// multiple load/stores of the same address.
133 ValueMap<Value*, Value*> SunkAddrs;
135 /// Keeps track of all instructions inserted for the current function.
136 SetOfInstrs InsertedInsts;
137 /// Keeps track of the type of the related instruction before their
138 /// promotion for the current function.
139 InstrToOrigTy PromotedInsts;
141 /// True if CFG is modified in any way.
144 /// True if optimizing for size.
147 /// DataLayout for the Function being processed.
148 const DataLayout *DL;
151 static char ID; // Pass identification, replacement for typeid
152 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
153 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr), DL(nullptr) {
154 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
156 bool runOnFunction(Function &F) override;
158 const char *getPassName() const override { return "CodeGen Prepare"; }
160 void getAnalysisUsage(AnalysisUsage &AU) const override {
161 AU.addPreserved<DominatorTreeWrapperPass>();
162 AU.addRequired<TargetLibraryInfoWrapperPass>();
163 AU.addRequired<TargetTransformInfoWrapperPass>();
167 bool eliminateFallThrough(Function &F);
168 bool eliminateMostlyEmptyBlocks(Function &F);
169 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
170 void eliminateMostlyEmptyBlock(BasicBlock *BB);
171 bool optimizeBlock(BasicBlock &BB, bool& ModifiedDT);
172 bool optimizeInst(Instruction *I, bool& ModifiedDT);
173 bool optimizeMemoryInst(Instruction *I, Value *Addr,
174 Type *AccessTy, unsigned AS);
175 bool optimizeInlineAsmInst(CallInst *CS);
176 bool optimizeCallInst(CallInst *CI, bool& ModifiedDT);
177 bool moveExtToFormExtLoad(Instruction *&I);
178 bool optimizeExtUses(Instruction *I);
179 bool optimizeLoadExt(LoadInst *I);
180 bool optimizeSelectInst(SelectInst *SI);
181 bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
182 bool optimizeSwitchInst(SwitchInst *CI);
183 bool optimizeExtractElementInst(Instruction *Inst);
184 bool dupRetToEnableTailCallOpts(BasicBlock *BB);
185 bool placeDbgValues(Function &F);
186 bool sinkAndCmp(Function &F);
187 bool extLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
189 const SmallVectorImpl<Instruction *> &Exts,
190 unsigned CreatedInstCost);
191 bool splitBranchCondition(Function &F);
192 bool simplifyOffsetableRelocate(Instruction &I);
193 void stripInvariantGroupMetadata(Instruction &I);
197 char CodeGenPrepare::ID = 0;
198 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
199 "Optimize for code generation", false, false)
201 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
202 return new CodeGenPrepare(TM);
205 bool CodeGenPrepare::runOnFunction(Function &F) {
206 if (skipOptnoneFunction(F))
209 DL = &F.getParent()->getDataLayout();
211 bool EverMadeChange = false;
212 // Clear per function information.
213 InsertedInsts.clear();
214 PromotedInsts.clear();
218 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
219 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
220 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
221 OptSize = F.optForSize();
223 /// This optimization identifies DIV instructions that can be
224 /// profitably bypassed and carried out with a shorter, faster divide.
225 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
226 const DenseMap<unsigned int, unsigned int> &BypassWidths =
227 TLI->getBypassSlowDivWidths();
228 BasicBlock* BB = &*F.begin();
229 while (BB != nullptr) {
230 // bypassSlowDivision may create new BBs, but we don't want to reapply the
231 // optimization to those blocks.
232 BasicBlock* Next = BB->getNextNode();
233 EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
238 // Eliminate blocks that contain only PHI nodes and an
239 // unconditional branch.
240 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
242 // llvm.dbg.value is far away from the value then iSel may not be able
243 // handle it properly. iSel will drop llvm.dbg.value if it can not
244 // find a node corresponding to the value.
245 EverMadeChange |= placeDbgValues(F);
247 // If there is a mask, compare against zero, and branch that can be combined
248 // into a single target instruction, push the mask and compare into branch
249 // users. Do this before OptimizeBlock -> OptimizeInst ->
250 // OptimizeCmpExpression, which perturbs the pattern being searched for.
251 if (!DisableBranchOpts) {
252 EverMadeChange |= sinkAndCmp(F);
253 EverMadeChange |= splitBranchCondition(F);
256 bool MadeChange = true;
259 for (Function::iterator I = F.begin(); I != F.end(); ) {
260 BasicBlock *BB = &*I++;
261 bool ModifiedDTOnIteration = false;
262 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
264 // Restart BB iteration if the dominator tree of the Function was changed
265 if (ModifiedDTOnIteration)
268 EverMadeChange |= MadeChange;
273 if (!DisableBranchOpts) {
275 SmallPtrSet<BasicBlock*, 8> WorkList;
276 for (BasicBlock &BB : F) {
277 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
278 MadeChange |= ConstantFoldTerminator(&BB, true);
279 if (!MadeChange) continue;
281 for (SmallVectorImpl<BasicBlock*>::iterator
282 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
283 if (pred_begin(*II) == pred_end(*II))
284 WorkList.insert(*II);
287 // Delete the dead blocks and any of their dead successors.
288 MadeChange |= !WorkList.empty();
289 while (!WorkList.empty()) {
290 BasicBlock *BB = *WorkList.begin();
292 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
296 for (SmallVectorImpl<BasicBlock*>::iterator
297 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
298 if (pred_begin(*II) == pred_end(*II))
299 WorkList.insert(*II);
302 // Merge pairs of basic blocks with unconditional branches, connected by
304 if (EverMadeChange || MadeChange)
305 MadeChange |= eliminateFallThrough(F);
307 EverMadeChange |= MadeChange;
310 if (!DisableGCOpts) {
311 SmallVector<Instruction *, 2> Statepoints;
312 for (BasicBlock &BB : F)
313 for (Instruction &I : BB)
315 Statepoints.push_back(&I);
316 for (auto &I : Statepoints)
317 EverMadeChange |= simplifyOffsetableRelocate(*I);
320 return EverMadeChange;
323 /// Merge basic blocks which are connected by a single edge, where one of the
324 /// basic blocks has a single successor pointing to the other basic block,
325 /// which has a single predecessor.
326 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
327 bool Changed = false;
328 // Scan all of the blocks in the function, except for the entry block.
329 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
330 BasicBlock *BB = &*I++;
331 // If the destination block has a single pred, then this is a trivial
332 // edge, just collapse it.
333 BasicBlock *SinglePred = BB->getSinglePredecessor();
335 // Don't merge if BB's address is taken.
336 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
338 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
339 if (Term && !Term->isConditional()) {
341 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
342 // Remember if SinglePred was the entry block of the function.
343 // If so, we will need to move BB back to the entry position.
344 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
345 MergeBasicBlockIntoOnlyPred(BB, nullptr);
347 if (isEntry && BB != &BB->getParent()->getEntryBlock())
348 BB->moveBefore(&BB->getParent()->getEntryBlock());
350 // We have erased a block. Update the iterator.
351 I = BB->getIterator();
357 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
358 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
359 /// edges in ways that are non-optimal for isel. Start by eliminating these
360 /// blocks so we can split them the way we want them.
361 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
362 bool MadeChange = false;
363 // Note that this intentionally skips the entry block.
364 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
365 BasicBlock *BB = &*I++;
367 // If this block doesn't end with an uncond branch, ignore it.
368 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
369 if (!BI || !BI->isUnconditional())
372 // If the instruction before the branch (skipping debug info) isn't a phi
373 // node, then other stuff is happening here.
374 BasicBlock::iterator BBI = BI->getIterator();
375 if (BBI != BB->begin()) {
377 while (isa<DbgInfoIntrinsic>(BBI)) {
378 if (BBI == BB->begin())
382 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
386 // Do not break infinite loops.
387 BasicBlock *DestBB = BI->getSuccessor(0);
391 if (!canMergeBlocks(BB, DestBB))
394 eliminateMostlyEmptyBlock(BB);
400 /// Return true if we can merge BB into DestBB if there is a single
401 /// unconditional branch between them, and BB contains no other non-phi
403 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
404 const BasicBlock *DestBB) const {
405 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
406 // the successor. If there are more complex condition (e.g. preheaders),
407 // don't mess around with them.
408 BasicBlock::const_iterator BBI = BB->begin();
409 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
410 for (const User *U : PN->users()) {
411 const Instruction *UI = cast<Instruction>(U);
412 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
414 // If User is inside DestBB block and it is a PHINode then check
415 // incoming value. If incoming value is not from BB then this is
416 // a complex condition (e.g. preheaders) we want to avoid here.
417 if (UI->getParent() == DestBB) {
418 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
419 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
420 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
421 if (Insn && Insn->getParent() == BB &&
422 Insn->getParent() != UPN->getIncomingBlock(I))
429 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
430 // and DestBB may have conflicting incoming values for the block. If so, we
431 // can't merge the block.
432 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
433 if (!DestBBPN) return true; // no conflict.
435 // Collect the preds of BB.
436 SmallPtrSet<const BasicBlock*, 16> BBPreds;
437 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
438 // It is faster to get preds from a PHI than with pred_iterator.
439 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
440 BBPreds.insert(BBPN->getIncomingBlock(i));
442 BBPreds.insert(pred_begin(BB), pred_end(BB));
445 // Walk the preds of DestBB.
446 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
447 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
448 if (BBPreds.count(Pred)) { // Common predecessor?
449 BBI = DestBB->begin();
450 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
451 const Value *V1 = PN->getIncomingValueForBlock(Pred);
452 const Value *V2 = PN->getIncomingValueForBlock(BB);
454 // If V2 is a phi node in BB, look up what the mapped value will be.
455 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
456 if (V2PN->getParent() == BB)
457 V2 = V2PN->getIncomingValueForBlock(Pred);
459 // If there is a conflict, bail out.
460 if (V1 != V2) return false;
469 /// Eliminate a basic block that has only phi's and an unconditional branch in
471 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
472 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
473 BasicBlock *DestBB = BI->getSuccessor(0);
475 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
477 // If the destination block has a single pred, then this is a trivial edge,
479 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
480 if (SinglePred != DestBB) {
481 // Remember if SinglePred was the entry block of the function. If so, we
482 // will need to move BB back to the entry position.
483 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
484 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
486 if (isEntry && BB != &BB->getParent()->getEntryBlock())
487 BB->moveBefore(&BB->getParent()->getEntryBlock());
489 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
494 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
495 // to handle the new incoming edges it is about to have.
497 for (BasicBlock::iterator BBI = DestBB->begin();
498 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
499 // Remove the incoming value for BB, and remember it.
500 Value *InVal = PN->removeIncomingValue(BB, false);
502 // Two options: either the InVal is a phi node defined in BB or it is some
503 // value that dominates BB.
504 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
505 if (InValPhi && InValPhi->getParent() == BB) {
506 // Add all of the input values of the input PHI as inputs of this phi.
507 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
508 PN->addIncoming(InValPhi->getIncomingValue(i),
509 InValPhi->getIncomingBlock(i));
511 // Otherwise, add one instance of the dominating value for each edge that
512 // we will be adding.
513 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
514 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
515 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
517 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
518 PN->addIncoming(InVal, *PI);
523 // The PHIs are now updated, change everything that refers to BB to use
524 // DestBB and remove BB.
525 BB->replaceAllUsesWith(DestBB);
526 BB->eraseFromParent();
529 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
532 // Computes a map of base pointer relocation instructions to corresponding
533 // derived pointer relocation instructions given a vector of all relocate calls
534 static void computeBaseDerivedRelocateMap(
535 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
536 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
538 // Collect information in two maps: one primarily for locating the base object
539 // while filling the second map; the second map is the final structure holding
540 // a mapping between Base and corresponding Derived relocate calls
541 DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
542 for (auto *ThisRelocate : AllRelocateCalls) {
543 auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
544 ThisRelocate->getDerivedPtrIndex());
545 RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
547 for (auto &Item : RelocateIdxMap) {
548 std::pair<unsigned, unsigned> Key = Item.first;
549 if (Key.first == Key.second)
550 // Base relocation: nothing to insert
553 GCRelocateInst *I = Item.second;
554 auto BaseKey = std::make_pair(Key.first, Key.first);
556 // We're iterating over RelocateIdxMap so we cannot modify it.
557 auto MaybeBase = RelocateIdxMap.find(BaseKey);
558 if (MaybeBase == RelocateIdxMap.end())
559 // TODO: We might want to insert a new base object relocate and gep off
560 // that, if there are enough derived object relocates.
563 RelocateInstMap[MaybeBase->second].push_back(I);
567 // Accepts a GEP and extracts the operands into a vector provided they're all
568 // small integer constants
569 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
570 SmallVectorImpl<Value *> &OffsetV) {
571 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
572 // Only accept small constant integer operands
573 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
574 if (!Op || Op->getZExtValue() > 20)
578 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
579 OffsetV.push_back(GEP->getOperand(i));
583 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
584 // replace, computes a replacement, and affects it.
586 simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
587 const SmallVectorImpl<GCRelocateInst *> &Targets) {
588 bool MadeChange = false;
589 for (GCRelocateInst *ToReplace : Targets) {
590 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
591 "Not relocating a derived object of the original base object");
592 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
593 // A duplicate relocate call. TODO: coalesce duplicates.
597 if (RelocatedBase->getParent() != ToReplace->getParent()) {
598 // Base and derived relocates are in different basic blocks.
599 // In this case transform is only valid when base dominates derived
600 // relocate. However it would be too expensive to check dominance
601 // for each such relocate, so we skip the whole transformation.
605 Value *Base = ToReplace->getBasePtr();
606 auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
607 if (!Derived || Derived->getPointerOperand() != Base)
610 SmallVector<Value *, 2> OffsetV;
611 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
614 // Create a Builder and replace the target callsite with a gep
615 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
617 // Insert after RelocatedBase
618 IRBuilder<> Builder(RelocatedBase->getNextNode());
619 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
621 // If gc_relocate does not match the actual type, cast it to the right type.
622 // In theory, there must be a bitcast after gc_relocate if the type does not
623 // match, and we should reuse it to get the derived pointer. But it could be
627 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
632 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
636 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
637 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
639 // In this case, we can not find the bitcast any more. So we insert a new bitcast
640 // no matter there is already one or not. In this way, we can handle all cases, and
641 // the extra bitcast should be optimized away in later passes.
642 Value *ActualRelocatedBase = RelocatedBase;
643 if (RelocatedBase->getType() != Base->getType()) {
644 ActualRelocatedBase =
645 Builder.CreateBitCast(RelocatedBase, Base->getType());
647 Value *Replacement = Builder.CreateGEP(
648 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
649 Replacement->takeName(ToReplace);
650 // If the newly generated derived pointer's type does not match the original derived
651 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
652 Value *ActualReplacement = Replacement;
653 if (Replacement->getType() != ToReplace->getType()) {
655 Builder.CreateBitCast(Replacement, ToReplace->getType());
657 ToReplace->replaceAllUsesWith(ActualReplacement);
658 ToReplace->eraseFromParent();
668 // %ptr = gep %base + 15
669 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
670 // %base' = relocate(%tok, i32 4, i32 4)
671 // %ptr' = relocate(%tok, i32 4, i32 5)
677 // %ptr = gep %base + 15
678 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
679 // %base' = gc.relocate(%tok, i32 4, i32 4)
680 // %ptr' = gep %base' + 15
682 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
683 bool MadeChange = false;
684 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
686 for (auto *U : I.users())
687 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
688 // Collect all the relocate calls associated with a statepoint
689 AllRelocateCalls.push_back(Relocate);
691 // We need atleast one base pointer relocation + one derived pointer
692 // relocation to mangle
693 if (AllRelocateCalls.size() < 2)
696 // RelocateInstMap is a mapping from the base relocate instruction to the
697 // corresponding derived relocate instructions
698 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
699 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
700 if (RelocateInstMap.empty())
703 for (auto &Item : RelocateInstMap)
704 // Item.first is the RelocatedBase to offset against
705 // Item.second is the vector of Targets to replace
706 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
710 /// SinkCast - Sink the specified cast instruction into its user blocks
711 static bool SinkCast(CastInst *CI) {
712 BasicBlock *DefBB = CI->getParent();
714 /// InsertedCasts - Only insert a cast in each block once.
715 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
717 bool MadeChange = false;
718 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
720 Use &TheUse = UI.getUse();
721 Instruction *User = cast<Instruction>(*UI);
723 // Figure out which BB this cast is used in. For PHI's this is the
724 // appropriate predecessor block.
725 BasicBlock *UserBB = User->getParent();
726 if (PHINode *PN = dyn_cast<PHINode>(User)) {
727 UserBB = PN->getIncomingBlock(TheUse);
730 // Preincrement use iterator so we don't invalidate it.
733 // If the block selected to receive the cast is an EH pad that does not
734 // allow non-PHI instructions before the terminator, we can't sink the
736 if (UserBB->getTerminator()->isEHPad())
739 // If this user is in the same block as the cast, don't change the cast.
740 if (UserBB == DefBB) continue;
742 // If we have already inserted a cast into this block, use it.
743 CastInst *&InsertedCast = InsertedCasts[UserBB];
746 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
747 assert(InsertPt != UserBB->end());
748 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
749 CI->getType(), "", &*InsertPt);
752 // Replace a use of the cast with a use of the new cast.
753 TheUse = InsertedCast;
758 // If we removed all uses, nuke the cast.
759 if (CI->use_empty()) {
760 CI->eraseFromParent();
767 /// If the specified cast instruction is a noop copy (e.g. it's casting from
768 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
769 /// reduce the number of virtual registers that must be created and coalesced.
771 /// Return true if any changes are made.
773 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
774 const DataLayout &DL) {
775 // If this is a noop copy,
776 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
777 EVT DstVT = TLI.getValueType(DL, CI->getType());
779 // This is an fp<->int conversion?
780 if (SrcVT.isInteger() != DstVT.isInteger())
783 // If this is an extension, it will be a zero or sign extension, which
785 if (SrcVT.bitsLT(DstVT)) return false;
787 // If these values will be promoted, find out what they will be promoted
788 // to. This helps us consider truncates on PPC as noop copies when they
790 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
791 TargetLowering::TypePromoteInteger)
792 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
793 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
794 TargetLowering::TypePromoteInteger)
795 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
797 // If, after promotion, these are the same types, this is a noop copy.
804 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
807 /// Return true if any changes were made.
808 static bool CombineUAddWithOverflow(CmpInst *CI) {
812 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
815 Type *Ty = AddI->getType();
816 if (!isa<IntegerType>(Ty))
819 // We don't want to move around uses of condition values this late, so we we
820 // check if it is legal to create the call to the intrinsic in the basic
821 // block containing the icmp:
823 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
827 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
829 if (AddI->hasOneUse())
830 assert(*AddI->user_begin() == CI && "expected!");
833 Module *M = CI->getModule();
834 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
836 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
838 auto *UAddWithOverflow =
839 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
840 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
842 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
844 CI->replaceAllUsesWith(Overflow);
845 AddI->replaceAllUsesWith(UAdd);
846 CI->eraseFromParent();
847 AddI->eraseFromParent();
851 /// Sink the given CmpInst into user blocks to reduce the number of virtual
852 /// registers that must be created and coalesced. This is a clear win except on
853 /// targets with multiple condition code registers (PowerPC), where it might
854 /// lose; some adjustment may be wanted there.
856 /// Return true if any changes are made.
857 static bool SinkCmpExpression(CmpInst *CI) {
858 BasicBlock *DefBB = CI->getParent();
860 /// Only insert a cmp in each block once.
861 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
863 bool MadeChange = false;
864 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
866 Use &TheUse = UI.getUse();
867 Instruction *User = cast<Instruction>(*UI);
869 // Preincrement use iterator so we don't invalidate it.
872 // Don't bother for PHI nodes.
873 if (isa<PHINode>(User))
876 // Figure out which BB this cmp is used in.
877 BasicBlock *UserBB = User->getParent();
879 // If this user is in the same block as the cmp, don't change the cmp.
880 if (UserBB == DefBB) continue;
882 // If we have already inserted a cmp into this block, use it.
883 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
886 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
887 assert(InsertPt != UserBB->end());
889 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
890 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
893 // Replace a use of the cmp with a use of the new cmp.
894 TheUse = InsertedCmp;
899 // If we removed all uses, nuke the cmp.
900 if (CI->use_empty()) {
901 CI->eraseFromParent();
908 static bool OptimizeCmpExpression(CmpInst *CI) {
909 if (SinkCmpExpression(CI))
912 if (CombineUAddWithOverflow(CI))
918 /// Check if the candidates could be combined with a shift instruction, which
920 /// 1. Truncate instruction
921 /// 2. And instruction and the imm is a mask of the low bits:
922 /// imm & (imm+1) == 0
923 static bool isExtractBitsCandidateUse(Instruction *User) {
924 if (!isa<TruncInst>(User)) {
925 if (User->getOpcode() != Instruction::And ||
926 !isa<ConstantInt>(User->getOperand(1)))
929 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
931 if ((Cimm & (Cimm + 1)).getBoolValue())
937 /// Sink both shift and truncate instruction to the use of truncate's BB.
939 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
940 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
941 const TargetLowering &TLI, const DataLayout &DL) {
942 BasicBlock *UserBB = User->getParent();
943 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
944 TruncInst *TruncI = dyn_cast<TruncInst>(User);
945 bool MadeChange = false;
947 for (Value::user_iterator TruncUI = TruncI->user_begin(),
948 TruncE = TruncI->user_end();
949 TruncUI != TruncE;) {
951 Use &TruncTheUse = TruncUI.getUse();
952 Instruction *TruncUser = cast<Instruction>(*TruncUI);
953 // Preincrement use iterator so we don't invalidate it.
957 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
961 // If the use is actually a legal node, there will not be an
962 // implicit truncate.
963 // FIXME: always querying the result type is just an
964 // approximation; some nodes' legality is determined by the
965 // operand or other means. There's no good way to find out though.
966 if (TLI.isOperationLegalOrCustom(
967 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
970 // Don't bother for PHI nodes.
971 if (isa<PHINode>(TruncUser))
974 BasicBlock *TruncUserBB = TruncUser->getParent();
976 if (UserBB == TruncUserBB)
979 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
980 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
982 if (!InsertedShift && !InsertedTrunc) {
983 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
984 assert(InsertPt != TruncUserBB->end());
986 if (ShiftI->getOpcode() == Instruction::AShr)
987 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
990 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
994 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
996 assert(TruncInsertPt != TruncUserBB->end());
998 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
999 TruncI->getType(), "", &*TruncInsertPt);
1003 TruncTheUse = InsertedTrunc;
1009 /// Sink the shift *right* instruction into user blocks if the uses could
1010 /// potentially be combined with this shift instruction and generate BitExtract
1011 /// instruction. It will only be applied if the architecture supports BitExtract
1012 /// instruction. Here is an example:
1014 /// %x.extract.shift = lshr i64 %arg1, 32
1016 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1020 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1021 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1023 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1025 /// Return true if any changes are made.
1026 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1027 const TargetLowering &TLI,
1028 const DataLayout &DL) {
1029 BasicBlock *DefBB = ShiftI->getParent();
1031 /// Only insert instructions in each block once.
1032 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1034 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1036 bool MadeChange = false;
1037 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1039 Use &TheUse = UI.getUse();
1040 Instruction *User = cast<Instruction>(*UI);
1041 // Preincrement use iterator so we don't invalidate it.
1044 // Don't bother for PHI nodes.
1045 if (isa<PHINode>(User))
1048 if (!isExtractBitsCandidateUse(User))
1051 BasicBlock *UserBB = User->getParent();
1053 if (UserBB == DefBB) {
1054 // If the shift and truncate instruction are in the same BB. The use of
1055 // the truncate(TruncUse) may still introduce another truncate if not
1056 // legal. In this case, we would like to sink both shift and truncate
1057 // instruction to the BB of TruncUse.
1060 // i64 shift.result = lshr i64 opnd, imm
1061 // trunc.result = trunc shift.result to i16
1064 // ----> We will have an implicit truncate here if the architecture does
1065 // not have i16 compare.
1066 // cmp i16 trunc.result, opnd2
1068 if (isa<TruncInst>(User) && shiftIsLegal
1069 // If the type of the truncate is legal, no trucate will be
1070 // introduced in other basic blocks.
1072 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1074 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1078 // If we have already inserted a shift into this block, use it.
1079 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1081 if (!InsertedShift) {
1082 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1083 assert(InsertPt != UserBB->end());
1085 if (ShiftI->getOpcode() == Instruction::AShr)
1086 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1089 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1095 // Replace a use of the shift with a use of the new shift.
1096 TheUse = InsertedShift;
1099 // If we removed all uses, nuke the shift.
1100 if (ShiftI->use_empty())
1101 ShiftI->eraseFromParent();
1106 // Translate a masked load intrinsic like
1107 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1108 // <16 x i1> %mask, <16 x i32> %passthru)
1109 // to a chain of basic blocks, with loading element one-by-one if
1110 // the appropriate mask bit is set
1112 // %1 = bitcast i8* %addr to i32*
1113 // %2 = extractelement <16 x i1> %mask, i32 0
1114 // %3 = icmp eq i1 %2, true
1115 // br i1 %3, label %cond.load, label %else
1117 //cond.load: ; preds = %0
1118 // %4 = getelementptr i32* %1, i32 0
1119 // %5 = load i32* %4
1120 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1123 //else: ; preds = %0, %cond.load
1124 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1125 // %7 = extractelement <16 x i1> %mask, i32 1
1126 // %8 = icmp eq i1 %7, true
1127 // br i1 %8, label %cond.load1, label %else2
1129 //cond.load1: ; preds = %else
1130 // %9 = getelementptr i32* %1, i32 1
1131 // %10 = load i32* %9
1132 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1135 //else2: ; preds = %else, %cond.load1
1136 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1137 // %12 = extractelement <16 x i1> %mask, i32 2
1138 // %13 = icmp eq i1 %12, true
1139 // br i1 %13, label %cond.load4, label %else5
1141 static void ScalarizeMaskedLoad(CallInst *CI) {
1142 Value *Ptr = CI->getArgOperand(0);
1143 Value *Alignment = CI->getArgOperand(1);
1144 Value *Mask = CI->getArgOperand(2);
1145 Value *Src0 = CI->getArgOperand(3);
1147 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1148 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1149 assert(VecType && "Unexpected return type of masked load intrinsic");
1151 Type *EltTy = CI->getType()->getVectorElementType();
1153 IRBuilder<> Builder(CI->getContext());
1154 Instruction *InsertPt = CI;
1155 BasicBlock *IfBlock = CI->getParent();
1156 BasicBlock *CondBlock = nullptr;
1157 BasicBlock *PrevIfBlock = CI->getParent();
1159 Builder.SetInsertPoint(InsertPt);
1160 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1162 // Short-cut if the mask is all-true.
1163 bool IsAllOnesMask = isa<Constant>(Mask) &&
1164 cast<Constant>(Mask)->isAllOnesValue();
1166 if (IsAllOnesMask) {
1167 Value *NewI = Builder.CreateAlignedLoad(Ptr, AlignVal);
1168 CI->replaceAllUsesWith(NewI);
1169 CI->eraseFromParent();
1173 // Adjust alignment for the scalar instruction.
1174 AlignVal = std::min(AlignVal, VecType->getScalarSizeInBits()/8);
1175 // Bitcast %addr fron i8* to EltTy*
1177 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1178 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1179 unsigned VectorWidth = VecType->getNumElements();
1181 Value *UndefVal = UndefValue::get(VecType);
1183 // The result vector
1184 Value *VResult = UndefVal;
1186 if (isa<ConstantVector>(Mask)) {
1187 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1188 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1191 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1192 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1193 VResult = Builder.CreateInsertElement(VResult, Load,
1194 Builder.getInt32(Idx));
1196 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
1197 CI->replaceAllUsesWith(NewI);
1198 CI->eraseFromParent();
1202 PHINode *Phi = nullptr;
1203 Value *PrevPhi = UndefVal;
1205 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1207 // Fill the "else" block, created in the previous iteration
1209 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1210 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1211 // %to_load = icmp eq i1 %mask_1, true
1212 // br i1 %to_load, label %cond.load, label %else
1215 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1216 Phi->addIncoming(VResult, CondBlock);
1217 Phi->addIncoming(PrevPhi, PrevIfBlock);
1222 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1223 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1224 ConstantInt::get(Predicate->getType(), 1));
1226 // Create "cond" block
1228 // %EltAddr = getelementptr i32* %1, i32 0
1229 // %Elt = load i32* %EltAddr
1230 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1232 CondBlock = IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.load");
1233 Builder.SetInsertPoint(InsertPt);
1236 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1237 LoadInst *Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1238 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1240 // Create "else" block, fill it in the next iteration
1241 BasicBlock *NewIfBlock =
1242 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1243 Builder.SetInsertPoint(InsertPt);
1244 Instruction *OldBr = IfBlock->getTerminator();
1245 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1246 OldBr->eraseFromParent();
1247 PrevIfBlock = IfBlock;
1248 IfBlock = NewIfBlock;
1251 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1252 Phi->addIncoming(VResult, CondBlock);
1253 Phi->addIncoming(PrevPhi, PrevIfBlock);
1254 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1255 CI->replaceAllUsesWith(NewI);
1256 CI->eraseFromParent();
1259 // Translate a masked store intrinsic, like
1260 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1262 // to a chain of basic blocks, that stores element one-by-one if
1263 // the appropriate mask bit is set
1265 // %1 = bitcast i8* %addr to i32*
1266 // %2 = extractelement <16 x i1> %mask, i32 0
1267 // %3 = icmp eq i1 %2, true
1268 // br i1 %3, label %cond.store, label %else
1270 // cond.store: ; preds = %0
1271 // %4 = extractelement <16 x i32> %val, i32 0
1272 // %5 = getelementptr i32* %1, i32 0
1273 // store i32 %4, i32* %5
1276 // else: ; preds = %0, %cond.store
1277 // %6 = extractelement <16 x i1> %mask, i32 1
1278 // %7 = icmp eq i1 %6, true
1279 // br i1 %7, label %cond.store1, label %else2
1281 // cond.store1: ; preds = %else
1282 // %8 = extractelement <16 x i32> %val, i32 1
1283 // %9 = getelementptr i32* %1, i32 1
1284 // store i32 %8, i32* %9
1287 static void ScalarizeMaskedStore(CallInst *CI) {
1288 Value *Src = CI->getArgOperand(0);
1289 Value *Ptr = CI->getArgOperand(1);
1290 Value *Alignment = CI->getArgOperand(2);
1291 Value *Mask = CI->getArgOperand(3);
1293 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1294 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1295 assert(VecType && "Unexpected data type in masked store intrinsic");
1297 Type *EltTy = VecType->getElementType();
1299 IRBuilder<> Builder(CI->getContext());
1300 Instruction *InsertPt = CI;
1301 BasicBlock *IfBlock = CI->getParent();
1302 Builder.SetInsertPoint(InsertPt);
1303 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1305 // Short-cut if the mask is all-true.
1306 bool IsAllOnesMask = isa<Constant>(Mask) &&
1307 cast<Constant>(Mask)->isAllOnesValue();
1309 if (IsAllOnesMask) {
1310 Builder.CreateAlignedStore(Src, Ptr, AlignVal);
1311 CI->eraseFromParent();
1315 // Adjust alignment for the scalar instruction.
1316 AlignVal = std::max(AlignVal, VecType->getScalarSizeInBits()/8);
1317 // Bitcast %addr fron i8* to EltTy*
1319 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1320 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1321 unsigned VectorWidth = VecType->getNumElements();
1323 if (isa<ConstantVector>(Mask)) {
1324 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1325 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1327 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1329 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1330 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1332 CI->eraseFromParent();
1336 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1338 // Fill the "else" block, created in the previous iteration
1340 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1341 // %to_store = icmp eq i1 %mask_1, true
1342 // br i1 %to_store, label %cond.store, label %else
1344 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1345 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1346 ConstantInt::get(Predicate->getType(), 1));
1348 // Create "cond" block
1350 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1351 // %EltAddr = getelementptr i32* %1, i32 0
1352 // %store i32 %OneElt, i32* %EltAddr
1354 BasicBlock *CondBlock =
1355 IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.store");
1356 Builder.SetInsertPoint(InsertPt);
1358 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1360 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1361 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1363 // Create "else" block, fill it in the next iteration
1364 BasicBlock *NewIfBlock =
1365 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1366 Builder.SetInsertPoint(InsertPt);
1367 Instruction *OldBr = IfBlock->getTerminator();
1368 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1369 OldBr->eraseFromParent();
1370 IfBlock = NewIfBlock;
1372 CI->eraseFromParent();
1375 // Translate a masked gather intrinsic like
1376 // <16 x i32 > @llvm.masked.gather.v16i32( <16 x i32*> %Ptrs, i32 4,
1377 // <16 x i1> %Mask, <16 x i32> %Src)
1378 // to a chain of basic blocks, with loading element one-by-one if
1379 // the appropriate mask bit is set
1381 // % Ptrs = getelementptr i32, i32* %base, <16 x i64> %ind
1382 // % Mask0 = extractelement <16 x i1> %Mask, i32 0
1383 // % ToLoad0 = icmp eq i1 % Mask0, true
1384 // br i1 % ToLoad0, label %cond.load, label %else
1387 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
1388 // % Load0 = load i32, i32* % Ptr0, align 4
1389 // % Res0 = insertelement <16 x i32> undef, i32 % Load0, i32 0
1393 // %res.phi.else = phi <16 x i32>[% Res0, %cond.load], [undef, % 0]
1394 // % Mask1 = extractelement <16 x i1> %Mask, i32 1
1395 // % ToLoad1 = icmp eq i1 % Mask1, true
1396 // br i1 % ToLoad1, label %cond.load1, label %else2
1399 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1400 // % Load1 = load i32, i32* % Ptr1, align 4
1401 // % Res1 = insertelement <16 x i32> %res.phi.else, i32 % Load1, i32 1
1404 // % Result = select <16 x i1> %Mask, <16 x i32> %res.phi.select, <16 x i32> %Src
1405 // ret <16 x i32> %Result
1406 static void ScalarizeMaskedGather(CallInst *CI) {
1407 Value *Ptrs = CI->getArgOperand(0);
1408 Value *Alignment = CI->getArgOperand(1);
1409 Value *Mask = CI->getArgOperand(2);
1410 Value *Src0 = CI->getArgOperand(3);
1412 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1414 assert(VecType && "Unexpected return type of masked load intrinsic");
1416 IRBuilder<> Builder(CI->getContext());
1417 Instruction *InsertPt = CI;
1418 BasicBlock *IfBlock = CI->getParent();
1419 BasicBlock *CondBlock = nullptr;
1420 BasicBlock *PrevIfBlock = CI->getParent();
1421 Builder.SetInsertPoint(InsertPt);
1422 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1424 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1426 Value *UndefVal = UndefValue::get(VecType);
1428 // The result vector
1429 Value *VResult = UndefVal;
1430 unsigned VectorWidth = VecType->getNumElements();
1432 // Shorten the way if the mask is a vector of constants.
1433 bool IsConstMask = isa<ConstantVector>(Mask);
1436 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1437 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1439 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1440 "Ptr" + Twine(Idx));
1441 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
1442 "Load" + Twine(Idx));
1443 VResult = Builder.CreateInsertElement(VResult, Load,
1444 Builder.getInt32(Idx),
1445 "Res" + Twine(Idx));
1447 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
1448 CI->replaceAllUsesWith(NewI);
1449 CI->eraseFromParent();
1453 PHINode *Phi = nullptr;
1454 Value *PrevPhi = UndefVal;
1456 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1458 // Fill the "else" block, created in the previous iteration
1460 // %Mask1 = extractelement <16 x i1> %Mask, i32 1
1461 // %ToLoad1 = icmp eq i1 %Mask1, true
1462 // br i1 %ToLoad1, label %cond.load, label %else
1465 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1466 Phi->addIncoming(VResult, CondBlock);
1467 Phi->addIncoming(PrevPhi, PrevIfBlock);
1472 Value *Predicate = Builder.CreateExtractElement(Mask,
1473 Builder.getInt32(Idx),
1474 "Mask" + Twine(Idx));
1475 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1476 ConstantInt::get(Predicate->getType(), 1),
1477 "ToLoad" + Twine(Idx));
1479 // Create "cond" block
1481 // %EltAddr = getelementptr i32* %1, i32 0
1482 // %Elt = load i32* %EltAddr
1483 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1485 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1486 Builder.SetInsertPoint(InsertPt);
1488 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1489 "Ptr" + Twine(Idx));
1490 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
1491 "Load" + Twine(Idx));
1492 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx),
1493 "Res" + Twine(Idx));
1495 // Create "else" block, fill it in the next iteration
1496 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1497 Builder.SetInsertPoint(InsertPt);
1498 Instruction *OldBr = IfBlock->getTerminator();
1499 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1500 OldBr->eraseFromParent();
1501 PrevIfBlock = IfBlock;
1502 IfBlock = NewIfBlock;
1505 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1506 Phi->addIncoming(VResult, CondBlock);
1507 Phi->addIncoming(PrevPhi, PrevIfBlock);
1508 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1509 CI->replaceAllUsesWith(NewI);
1510 CI->eraseFromParent();
1513 // Translate a masked scatter intrinsic, like
1514 // void @llvm.masked.scatter.v16i32(<16 x i32> %Src, <16 x i32*>* %Ptrs, i32 4,
1516 // to a chain of basic blocks, that stores element one-by-one if
1517 // the appropriate mask bit is set.
1519 // % Ptrs = getelementptr i32, i32* %ptr, <16 x i64> %ind
1520 // % Mask0 = extractelement <16 x i1> % Mask, i32 0
1521 // % ToStore0 = icmp eq i1 % Mask0, true
1522 // br i1 %ToStore0, label %cond.store, label %else
1525 // % Elt0 = extractelement <16 x i32> %Src, i32 0
1526 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
1527 // store i32 %Elt0, i32* % Ptr0, align 4
1531 // % Mask1 = extractelement <16 x i1> % Mask, i32 1
1532 // % ToStore1 = icmp eq i1 % Mask1, true
1533 // br i1 % ToStore1, label %cond.store1, label %else2
1536 // % Elt1 = extractelement <16 x i32> %Src, i32 1
1537 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1538 // store i32 % Elt1, i32* % Ptr1, align 4
1541 static void ScalarizeMaskedScatter(CallInst *CI) {
1542 Value *Src = CI->getArgOperand(0);
1543 Value *Ptrs = CI->getArgOperand(1);
1544 Value *Alignment = CI->getArgOperand(2);
1545 Value *Mask = CI->getArgOperand(3);
1547 assert(isa<VectorType>(Src->getType()) &&
1548 "Unexpected data type in masked scatter intrinsic");
1549 assert(isa<VectorType>(Ptrs->getType()) &&
1550 isa<PointerType>(Ptrs->getType()->getVectorElementType()) &&
1551 "Vector of pointers is expected in masked scatter intrinsic");
1553 IRBuilder<> Builder(CI->getContext());
1554 Instruction *InsertPt = CI;
1555 BasicBlock *IfBlock = CI->getParent();
1556 Builder.SetInsertPoint(InsertPt);
1557 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1559 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1560 unsigned VectorWidth = Src->getType()->getVectorNumElements();
1562 // Shorten the way if the mask is a vector of constants.
1563 bool IsConstMask = isa<ConstantVector>(Mask);
1566 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1567 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1569 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
1570 "Elt" + Twine(Idx));
1571 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1572 "Ptr" + Twine(Idx));
1573 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
1575 CI->eraseFromParent();
1578 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1579 // Fill the "else" block, created in the previous iteration
1581 // % Mask1 = extractelement <16 x i1> % Mask, i32 Idx
1582 // % ToStore = icmp eq i1 % Mask1, true
1583 // br i1 % ToStore, label %cond.store, label %else
1585 Value *Predicate = Builder.CreateExtractElement(Mask,
1586 Builder.getInt32(Idx),
1587 "Mask" + Twine(Idx));
1589 Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1590 ConstantInt::get(Predicate->getType(), 1),
1591 "ToStore" + Twine(Idx));
1593 // Create "cond" block
1595 // % Elt1 = extractelement <16 x i32> %Src, i32 1
1596 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1597 // %store i32 % Elt1, i32* % Ptr1
1599 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1600 Builder.SetInsertPoint(InsertPt);
1602 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
1603 "Elt" + Twine(Idx));
1604 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1605 "Ptr" + Twine(Idx));
1606 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
1608 // Create "else" block, fill it in the next iteration
1609 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1610 Builder.SetInsertPoint(InsertPt);
1611 Instruction *OldBr = IfBlock->getTerminator();
1612 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1613 OldBr->eraseFromParent();
1614 IfBlock = NewIfBlock;
1616 CI->eraseFromParent();
1619 /// If counting leading or trailing zeros is an expensive operation and a zero
1620 /// input is defined, add a check for zero to avoid calling the intrinsic.
1622 /// We want to transform:
1623 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
1627 /// %cmpz = icmp eq i64 %A, 0
1628 /// br i1 %cmpz, label %cond.end, label %cond.false
1630 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
1631 /// br label %cond.end
1633 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
1635 /// If the transform is performed, return true and set ModifiedDT to true.
1636 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
1637 const TargetLowering *TLI,
1638 const DataLayout *DL,
1643 // If a zero input is undefined, it doesn't make sense to despeculate that.
1644 if (match(CountZeros->getOperand(1), m_One()))
1647 // If it's cheap to speculate, there's nothing to do.
1648 auto IntrinsicID = CountZeros->getIntrinsicID();
1649 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
1650 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
1653 // Only handle legal scalar cases. Anything else requires too much work.
1654 Type *Ty = CountZeros->getType();
1655 unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
1656 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSize())
1659 // The intrinsic will be sunk behind a compare against zero and branch.
1660 BasicBlock *StartBlock = CountZeros->getParent();
1661 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
1663 // Create another block after the count zero intrinsic. A PHI will be added
1664 // in this block to select the result of the intrinsic or the bit-width
1665 // constant if the input to the intrinsic is zero.
1666 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
1667 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
1669 // Set up a builder to create a compare, conditional branch, and PHI.
1670 IRBuilder<> Builder(CountZeros->getContext());
1671 Builder.SetInsertPoint(StartBlock->getTerminator());
1672 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
1674 // Replace the unconditional branch that was created by the first split with
1675 // a compare against zero and a conditional branch.
1676 Value *Zero = Constant::getNullValue(Ty);
1677 Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
1678 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
1679 StartBlock->getTerminator()->eraseFromParent();
1681 // Create a PHI in the end block to select either the output of the intrinsic
1682 // or the bit width of the operand.
1683 Builder.SetInsertPoint(&EndBlock->front());
1684 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
1685 CountZeros->replaceAllUsesWith(PN);
1686 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
1687 PN->addIncoming(BitWidth, StartBlock);
1688 PN->addIncoming(CountZeros, CallBlock);
1690 // We are explicitly handling the zero case, so we can set the intrinsic's
1691 // undefined zero argument to 'true'. This will also prevent reprocessing the
1692 // intrinsic; we only despeculate when a zero input is defined.
1693 CountZeros->setArgOperand(1, Builder.getTrue());
1698 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1699 BasicBlock *BB = CI->getParent();
1701 // Lower inline assembly if we can.
1702 // If we found an inline asm expession, and if the target knows how to
1703 // lower it to normal LLVM code, do so now.
1704 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1705 if (TLI->ExpandInlineAsm(CI)) {
1706 // Avoid invalidating the iterator.
1707 CurInstIterator = BB->begin();
1708 // Avoid processing instructions out of order, which could cause
1709 // reuse before a value is defined.
1713 // Sink address computing for memory operands into the block.
1714 if (optimizeInlineAsmInst(CI))
1718 // Align the pointer arguments to this call if the target thinks it's a good
1720 unsigned MinSize, PrefAlign;
1721 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1722 for (auto &Arg : CI->arg_operands()) {
1723 // We want to align both objects whose address is used directly and
1724 // objects whose address is used in casts and GEPs, though it only makes
1725 // sense for GEPs if the offset is a multiple of the desired alignment and
1726 // if size - offset meets the size threshold.
1727 if (!Arg->getType()->isPointerTy())
1729 APInt Offset(DL->getPointerSizeInBits(
1730 cast<PointerType>(Arg->getType())->getAddressSpace()),
1732 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
1733 uint64_t Offset2 = Offset.getLimitedValue();
1734 if ((Offset2 & (PrefAlign-1)) != 0)
1737 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1738 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1739 AI->setAlignment(PrefAlign);
1740 // Global variables can only be aligned if they are defined in this
1741 // object (i.e. they are uniquely initialized in this object), and
1742 // over-aligning global variables that have an explicit section is
1745 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->hasUniqueInitializer() &&
1746 !GV->hasSection() && GV->getAlignment() < PrefAlign &&
1747 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
1749 GV->setAlignment(PrefAlign);
1751 // If this is a memcpy (or similar) then we may be able to improve the
1753 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1754 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
1755 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1756 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
1757 if (Align > MI->getAlignment())
1758 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1762 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1764 switch (II->getIntrinsicID()) {
1766 case Intrinsic::objectsize: {
1767 // Lower all uses of llvm.objectsize.*
1768 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1769 Type *ReturnTy = CI->getType();
1770 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1772 // Substituting this can cause recursive simplifications, which can
1773 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1775 WeakVH IterHandle(&*CurInstIterator);
1777 replaceAndRecursivelySimplify(CI, RetVal,
1780 // If the iterator instruction was recursively deleted, start over at the
1781 // start of the block.
1782 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
1783 CurInstIterator = BB->begin();
1788 case Intrinsic::masked_load: {
1789 // Scalarize unsupported vector masked load
1790 if (!TTI->isLegalMaskedLoad(CI->getType())) {
1791 ScalarizeMaskedLoad(CI);
1797 case Intrinsic::masked_store: {
1798 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType())) {
1799 ScalarizeMaskedStore(CI);
1805 case Intrinsic::masked_gather: {
1806 if (!TTI->isLegalMaskedGather(CI->getType())) {
1807 ScalarizeMaskedGather(CI);
1813 case Intrinsic::masked_scatter: {
1814 if (!TTI->isLegalMaskedScatter(CI->getArgOperand(0)->getType())) {
1815 ScalarizeMaskedScatter(CI);
1821 case Intrinsic::aarch64_stlxr:
1822 case Intrinsic::aarch64_stxr: {
1823 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1824 if (!ExtVal || !ExtVal->hasOneUse() ||
1825 ExtVal->getParent() == CI->getParent())
1827 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1828 ExtVal->moveBefore(CI);
1829 // Mark this instruction as "inserted by CGP", so that other
1830 // optimizations don't touch it.
1831 InsertedInsts.insert(ExtVal);
1834 case Intrinsic::invariant_group_barrier:
1835 II->replaceAllUsesWith(II->getArgOperand(0));
1836 II->eraseFromParent();
1839 case Intrinsic::cttz:
1840 case Intrinsic::ctlz:
1841 // If counting zeros is expensive, try to avoid it.
1842 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
1846 // Unknown address space.
1847 // TODO: Target hook to pick which address space the intrinsic cares
1849 unsigned AddrSpace = ~0u;
1850 SmallVector<Value*, 2> PtrOps;
1852 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
1853 while (!PtrOps.empty())
1854 if (optimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
1859 // From here on out we're working with named functions.
1860 if (!CI->getCalledFunction()) return false;
1862 // Lower all default uses of _chk calls. This is very similar
1863 // to what InstCombineCalls does, but here we are only lowering calls
1864 // to fortified library functions (e.g. __memcpy_chk) that have the default
1865 // "don't know" as the objectsize. Anything else should be left alone.
1866 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1867 if (Value *V = Simplifier.optimizeCall(CI)) {
1868 CI->replaceAllUsesWith(V);
1869 CI->eraseFromParent();
1875 /// Look for opportunities to duplicate return instructions to the predecessor
1876 /// to enable tail call optimizations. The case it is currently looking for is:
1879 /// %tmp0 = tail call i32 @f0()
1880 /// br label %return
1882 /// %tmp1 = tail call i32 @f1()
1883 /// br label %return
1885 /// %tmp2 = tail call i32 @f2()
1886 /// br label %return
1888 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1896 /// %tmp0 = tail call i32 @f0()
1899 /// %tmp1 = tail call i32 @f1()
1902 /// %tmp2 = tail call i32 @f2()
1905 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1909 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1913 PHINode *PN = nullptr;
1914 BitCastInst *BCI = nullptr;
1915 Value *V = RI->getReturnValue();
1917 BCI = dyn_cast<BitCastInst>(V);
1919 V = BCI->getOperand(0);
1921 PN = dyn_cast<PHINode>(V);
1926 if (PN && PN->getParent() != BB)
1929 // It's not safe to eliminate the sign / zero extension of the return value.
1930 // See llvm::isInTailCallPosition().
1931 const Function *F = BB->getParent();
1932 AttributeSet CallerAttrs = F->getAttributes();
1933 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1934 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1937 // Make sure there are no instructions between the PHI and return, or that the
1938 // return is the first instruction in the block.
1940 BasicBlock::iterator BI = BB->begin();
1941 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1943 // Also skip over the bitcast.
1948 BasicBlock::iterator BI = BB->begin();
1949 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1954 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1956 SmallVector<CallInst*, 4> TailCalls;
1958 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1959 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1960 // Make sure the phi value is indeed produced by the tail call.
1961 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1962 TLI->mayBeEmittedAsTailCall(CI))
1963 TailCalls.push_back(CI);
1966 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1967 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1968 if (!VisitedBBs.insert(*PI).second)
1971 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1972 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1973 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1974 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1978 CallInst *CI = dyn_cast<CallInst>(&*RI);
1979 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1980 TailCalls.push_back(CI);
1984 bool Changed = false;
1985 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1986 CallInst *CI = TailCalls[i];
1989 // Conservatively require the attributes of the call to match those of the
1990 // return. Ignore noalias because it doesn't affect the call sequence.
1991 AttributeSet CalleeAttrs = CS.getAttributes();
1992 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1993 removeAttribute(Attribute::NoAlias) !=
1994 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1995 removeAttribute(Attribute::NoAlias))
1998 // Make sure the call instruction is followed by an unconditional branch to
1999 // the return block.
2000 BasicBlock *CallBB = CI->getParent();
2001 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
2002 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
2005 // Duplicate the return into CallBB.
2006 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
2007 ModifiedDT = Changed = true;
2011 // If we eliminated all predecessors of the block, delete the block now.
2012 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
2013 BB->eraseFromParent();
2018 //===----------------------------------------------------------------------===//
2019 // Memory Optimization
2020 //===----------------------------------------------------------------------===//
2024 /// This is an extended version of TargetLowering::AddrMode
2025 /// which holds actual Value*'s for register values.
2026 struct ExtAddrMode : public TargetLowering::AddrMode {
2029 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
2030 void print(raw_ostream &OS) const;
2033 bool operator==(const ExtAddrMode& O) const {
2034 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
2035 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
2036 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
2041 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2047 void ExtAddrMode::print(raw_ostream &OS) const {
2048 bool NeedPlus = false;
2051 OS << (NeedPlus ? " + " : "")
2053 BaseGV->printAsOperand(OS, /*PrintType=*/false);
2058 OS << (NeedPlus ? " + " : "")
2064 OS << (NeedPlus ? " + " : "")
2066 BaseReg->printAsOperand(OS, /*PrintType=*/false);
2070 OS << (NeedPlus ? " + " : "")
2072 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2078 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2079 void ExtAddrMode::dump() const {
2085 /// \brief This class provides transaction based operation on the IR.
2086 /// Every change made through this class is recorded in the internal state and
2087 /// can be undone (rollback) until commit is called.
2088 class TypePromotionTransaction {
2090 /// \brief This represents the common interface of the individual transaction.
2091 /// Each class implements the logic for doing one specific modification on
2092 /// the IR via the TypePromotionTransaction.
2093 class TypePromotionAction {
2095 /// The Instruction modified.
2099 /// \brief Constructor of the action.
2100 /// The constructor performs the related action on the IR.
2101 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2103 virtual ~TypePromotionAction() {}
2105 /// \brief Undo the modification done by this action.
2106 /// When this method is called, the IR must be in the same state as it was
2107 /// before this action was applied.
2108 /// \pre Undoing the action works if and only if the IR is in the exact same
2109 /// state as it was directly after this action was applied.
2110 virtual void undo() = 0;
2112 /// \brief Advocate every change made by this action.
2113 /// When the results on the IR of the action are to be kept, it is important
2114 /// to call this function, otherwise hidden information may be kept forever.
2115 virtual void commit() {
2116 // Nothing to be done, this action is not doing anything.
2120 /// \brief Utility to remember the position of an instruction.
2121 class InsertionHandler {
2122 /// Position of an instruction.
2123 /// Either an instruction:
2124 /// - Is the first in a basic block: BB is used.
2125 /// - Has a previous instructon: PrevInst is used.
2127 Instruction *PrevInst;
2130 /// Remember whether or not the instruction had a previous instruction.
2131 bool HasPrevInstruction;
2134 /// \brief Record the position of \p Inst.
2135 InsertionHandler(Instruction *Inst) {
2136 BasicBlock::iterator It = Inst->getIterator();
2137 HasPrevInstruction = (It != (Inst->getParent()->begin()));
2138 if (HasPrevInstruction)
2139 Point.PrevInst = &*--It;
2141 Point.BB = Inst->getParent();
2144 /// \brief Insert \p Inst at the recorded position.
2145 void insert(Instruction *Inst) {
2146 if (HasPrevInstruction) {
2147 if (Inst->getParent())
2148 Inst->removeFromParent();
2149 Inst->insertAfter(Point.PrevInst);
2151 Instruction *Position = &*Point.BB->getFirstInsertionPt();
2152 if (Inst->getParent())
2153 Inst->moveBefore(Position);
2155 Inst->insertBefore(Position);
2160 /// \brief Move an instruction before another.
2161 class InstructionMoveBefore : public TypePromotionAction {
2162 /// Original position of the instruction.
2163 InsertionHandler Position;
2166 /// \brief Move \p Inst before \p Before.
2167 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2168 : TypePromotionAction(Inst), Position(Inst) {
2169 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
2170 Inst->moveBefore(Before);
2173 /// \brief Move the instruction back to its original position.
2174 void undo() override {
2175 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2176 Position.insert(Inst);
2180 /// \brief Set the operand of an instruction with a new value.
2181 class OperandSetter : public TypePromotionAction {
2182 /// Original operand of the instruction.
2184 /// Index of the modified instruction.
2188 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
2189 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2190 : TypePromotionAction(Inst), Idx(Idx) {
2191 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2192 << "for:" << *Inst << "\n"
2193 << "with:" << *NewVal << "\n");
2194 Origin = Inst->getOperand(Idx);
2195 Inst->setOperand(Idx, NewVal);
2198 /// \brief Restore the original value of the instruction.
2199 void undo() override {
2200 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2201 << "for: " << *Inst << "\n"
2202 << "with: " << *Origin << "\n");
2203 Inst->setOperand(Idx, Origin);
2207 /// \brief Hide the operands of an instruction.
2208 /// Do as if this instruction was not using any of its operands.
2209 class OperandsHider : public TypePromotionAction {
2210 /// The list of original operands.
2211 SmallVector<Value *, 4> OriginalValues;
2214 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
2215 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2216 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2217 unsigned NumOpnds = Inst->getNumOperands();
2218 OriginalValues.reserve(NumOpnds);
2219 for (unsigned It = 0; It < NumOpnds; ++It) {
2220 // Save the current operand.
2221 Value *Val = Inst->getOperand(It);
2222 OriginalValues.push_back(Val);
2224 // We could use OperandSetter here, but that would imply an overhead
2225 // that we are not willing to pay.
2226 Inst->setOperand(It, UndefValue::get(Val->getType()));
2230 /// \brief Restore the original list of uses.
2231 void undo() override {
2232 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2233 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2234 Inst->setOperand(It, OriginalValues[It]);
2238 /// \brief Build a truncate instruction.
2239 class TruncBuilder : public TypePromotionAction {
2242 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
2244 /// trunc Opnd to Ty.
2245 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2246 IRBuilder<> Builder(Opnd);
2247 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2248 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2251 /// \brief Get the built value.
2252 Value *getBuiltValue() { return Val; }
2254 /// \brief Remove the built instruction.
2255 void undo() override {
2256 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2257 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2258 IVal->eraseFromParent();
2262 /// \brief Build a sign extension instruction.
2263 class SExtBuilder : public TypePromotionAction {
2266 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
2268 /// sext Opnd to Ty.
2269 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2270 : TypePromotionAction(InsertPt) {
2271 IRBuilder<> Builder(InsertPt);
2272 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2273 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2276 /// \brief Get the built value.
2277 Value *getBuiltValue() { return Val; }
2279 /// \brief Remove the built instruction.
2280 void undo() override {
2281 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2282 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2283 IVal->eraseFromParent();
2287 /// \brief Build a zero extension instruction.
2288 class ZExtBuilder : public TypePromotionAction {
2291 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
2293 /// zext Opnd to Ty.
2294 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2295 : TypePromotionAction(InsertPt) {
2296 IRBuilder<> Builder(InsertPt);
2297 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2298 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2301 /// \brief Get the built value.
2302 Value *getBuiltValue() { return Val; }
2304 /// \brief Remove the built instruction.
2305 void undo() override {
2306 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2307 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2308 IVal->eraseFromParent();
2312 /// \brief Mutate an instruction to another type.
2313 class TypeMutator : public TypePromotionAction {
2314 /// Record the original type.
2318 /// \brief Mutate the type of \p Inst into \p NewTy.
2319 TypeMutator(Instruction *Inst, Type *NewTy)
2320 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2321 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2323 Inst->mutateType(NewTy);
2326 /// \brief Mutate the instruction back to its original type.
2327 void undo() override {
2328 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2330 Inst->mutateType(OrigTy);
2334 /// \brief Replace the uses of an instruction by another instruction.
2335 class UsesReplacer : public TypePromotionAction {
2336 /// Helper structure to keep track of the replaced uses.
2337 struct InstructionAndIdx {
2338 /// The instruction using the instruction.
2340 /// The index where this instruction is used for Inst.
2342 InstructionAndIdx(Instruction *Inst, unsigned Idx)
2343 : Inst(Inst), Idx(Idx) {}
2346 /// Keep track of the original uses (pair Instruction, Index).
2347 SmallVector<InstructionAndIdx, 4> OriginalUses;
2348 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
2351 /// \brief Replace all the use of \p Inst by \p New.
2352 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2353 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2355 // Record the original uses.
2356 for (Use &U : Inst->uses()) {
2357 Instruction *UserI = cast<Instruction>(U.getUser());
2358 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2360 // Now, we can replace the uses.
2361 Inst->replaceAllUsesWith(New);
2364 /// \brief Reassign the original uses of Inst to Inst.
2365 void undo() override {
2366 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2367 for (use_iterator UseIt = OriginalUses.begin(),
2368 EndIt = OriginalUses.end();
2369 UseIt != EndIt; ++UseIt) {
2370 UseIt->Inst->setOperand(UseIt->Idx, Inst);
2375 /// \brief Remove an instruction from the IR.
2376 class InstructionRemover : public TypePromotionAction {
2377 /// Original position of the instruction.
2378 InsertionHandler Inserter;
2379 /// Helper structure to hide all the link to the instruction. In other
2380 /// words, this helps to do as if the instruction was removed.
2381 OperandsHider Hider;
2382 /// Keep track of the uses replaced, if any.
2383 UsesReplacer *Replacer;
2386 /// \brief Remove all reference of \p Inst and optinally replace all its
2388 /// \pre If !Inst->use_empty(), then New != nullptr
2389 InstructionRemover(Instruction *Inst, Value *New = nullptr)
2390 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2393 Replacer = new UsesReplacer(Inst, New);
2394 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2395 Inst->removeFromParent();
2398 ~InstructionRemover() override { delete Replacer; }
2400 /// \brief Really remove the instruction.
2401 void commit() override { delete Inst; }
2403 /// \brief Resurrect the instruction and reassign it to the proper uses if
2404 /// new value was provided when build this action.
2405 void undo() override {
2406 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2407 Inserter.insert(Inst);
2415 /// Restoration point.
2416 /// The restoration point is a pointer to an action instead of an iterator
2417 /// because the iterator may be invalidated but not the pointer.
2418 typedef const TypePromotionAction *ConstRestorationPt;
2419 /// Advocate every changes made in that transaction.
2421 /// Undo all the changes made after the given point.
2422 void rollback(ConstRestorationPt Point);
2423 /// Get the current restoration point.
2424 ConstRestorationPt getRestorationPoint() const;
2426 /// \name API for IR modification with state keeping to support rollback.
2428 /// Same as Instruction::setOperand.
2429 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2430 /// Same as Instruction::eraseFromParent.
2431 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2432 /// Same as Value::replaceAllUsesWith.
2433 void replaceAllUsesWith(Instruction *Inst, Value *New);
2434 /// Same as Value::mutateType.
2435 void mutateType(Instruction *Inst, Type *NewTy);
2436 /// Same as IRBuilder::createTrunc.
2437 Value *createTrunc(Instruction *Opnd, Type *Ty);
2438 /// Same as IRBuilder::createSExt.
2439 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2440 /// Same as IRBuilder::createZExt.
2441 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2442 /// Same as Instruction::moveBefore.
2443 void moveBefore(Instruction *Inst, Instruction *Before);
2447 /// The ordered list of actions made so far.
2448 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
2449 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
2452 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2455 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
2458 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2461 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
2464 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2466 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2469 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2470 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2473 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2475 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2476 Value *Val = Ptr->getBuiltValue();
2477 Actions.push_back(std::move(Ptr));
2481 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2482 Value *Opnd, Type *Ty) {
2483 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2484 Value *Val = Ptr->getBuiltValue();
2485 Actions.push_back(std::move(Ptr));
2489 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2490 Value *Opnd, Type *Ty) {
2491 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2492 Value *Val = Ptr->getBuiltValue();
2493 Actions.push_back(std::move(Ptr));
2497 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2498 Instruction *Before) {
2500 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
2503 TypePromotionTransaction::ConstRestorationPt
2504 TypePromotionTransaction::getRestorationPoint() const {
2505 return !Actions.empty() ? Actions.back().get() : nullptr;
2508 void TypePromotionTransaction::commit() {
2509 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2515 void TypePromotionTransaction::rollback(
2516 TypePromotionTransaction::ConstRestorationPt Point) {
2517 while (!Actions.empty() && Point != Actions.back().get()) {
2518 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2523 /// \brief A helper class for matching addressing modes.
2525 /// This encapsulates the logic for matching the target-legal addressing modes.
2526 class AddressingModeMatcher {
2527 SmallVectorImpl<Instruction*> &AddrModeInsts;
2528 const TargetMachine &TM;
2529 const TargetLowering &TLI;
2530 const DataLayout &DL;
2532 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2533 /// the memory instruction that we're computing this address for.
2536 Instruction *MemoryInst;
2538 /// This is the addressing mode that we're building up. This is
2539 /// part of the return value of this addressing mode matching stuff.
2540 ExtAddrMode &AddrMode;
2542 /// The instructions inserted by other CodeGenPrepare optimizations.
2543 const SetOfInstrs &InsertedInsts;
2544 /// A map from the instructions to their type before promotion.
2545 InstrToOrigTy &PromotedInsts;
2546 /// The ongoing transaction where every action should be registered.
2547 TypePromotionTransaction &TPT;
2549 /// This is set to true when we should not do profitability checks.
2550 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2551 bool IgnoreProfitability;
2553 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2554 const TargetMachine &TM, Type *AT, unsigned AS,
2555 Instruction *MI, ExtAddrMode &AM,
2556 const SetOfInstrs &InsertedInsts,
2557 InstrToOrigTy &PromotedInsts,
2558 TypePromotionTransaction &TPT)
2559 : AddrModeInsts(AMI), TM(TM),
2560 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
2561 ->getTargetLowering()),
2562 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2563 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2564 PromotedInsts(PromotedInsts), TPT(TPT) {
2565 IgnoreProfitability = false;
2569 /// Find the maximal addressing mode that a load/store of V can fold,
2570 /// give an access type of AccessTy. This returns a list of involved
2571 /// instructions in AddrModeInsts.
2572 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2574 /// \p PromotedInsts maps the instructions to their type before promotion.
2575 /// \p The ongoing transaction where every action should be registered.
2576 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2577 Instruction *MemoryInst,
2578 SmallVectorImpl<Instruction*> &AddrModeInsts,
2579 const TargetMachine &TM,
2580 const SetOfInstrs &InsertedInsts,
2581 InstrToOrigTy &PromotedInsts,
2582 TypePromotionTransaction &TPT) {
2585 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
2586 MemoryInst, Result, InsertedInsts,
2587 PromotedInsts, TPT).matchAddr(V, 0);
2588 (void)Success; assert(Success && "Couldn't select *anything*?");
2592 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2593 bool matchAddr(Value *V, unsigned Depth);
2594 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2595 bool *MovedAway = nullptr);
2596 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2597 ExtAddrMode &AMBefore,
2598 ExtAddrMode &AMAfter);
2599 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2600 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2601 Value *PromotedOperand) const;
2604 /// Try adding ScaleReg*Scale to the current addressing mode.
2605 /// Return true and update AddrMode if this addr mode is legal for the target,
2607 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
2609 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2610 // mode. Just process that directly.
2612 return matchAddr(ScaleReg, Depth);
2614 // If the scale is 0, it takes nothing to add this.
2618 // If we already have a scale of this value, we can add to it, otherwise, we
2619 // need an available scale field.
2620 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2623 ExtAddrMode TestAddrMode = AddrMode;
2625 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2626 // [A+B + A*7] -> [B+A*8].
2627 TestAddrMode.Scale += Scale;
2628 TestAddrMode.ScaledReg = ScaleReg;
2630 // If the new address isn't legal, bail out.
2631 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
2634 // It was legal, so commit it.
2635 AddrMode = TestAddrMode;
2637 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2638 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2639 // X*Scale + C*Scale to addr mode.
2640 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2641 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2642 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2643 TestAddrMode.ScaledReg = AddLHS;
2644 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2646 // If this addressing mode is legal, commit it and remember that we folded
2647 // this instruction.
2648 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
2649 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2650 AddrMode = TestAddrMode;
2655 // Otherwise, not (x+c)*scale, just return what we have.
2659 /// This is a little filter, which returns true if an addressing computation
2660 /// involving I might be folded into a load/store accessing it.
2661 /// This doesn't need to be perfect, but needs to accept at least
2662 /// the set of instructions that MatchOperationAddr can.
2663 static bool MightBeFoldableInst(Instruction *I) {
2664 switch (I->getOpcode()) {
2665 case Instruction::BitCast:
2666 case Instruction::AddrSpaceCast:
2667 // Don't touch identity bitcasts.
2668 if (I->getType() == I->getOperand(0)->getType())
2670 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2671 case Instruction::PtrToInt:
2672 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2674 case Instruction::IntToPtr:
2675 // We know the input is intptr_t, so this is foldable.
2677 case Instruction::Add:
2679 case Instruction::Mul:
2680 case Instruction::Shl:
2681 // Can only handle X*C and X << C.
2682 return isa<ConstantInt>(I->getOperand(1));
2683 case Instruction::GetElementPtr:
2690 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2691 /// \note \p Val is assumed to be the product of some type promotion.
2692 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2693 /// to be legal, as the non-promoted value would have had the same state.
2694 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
2695 const DataLayout &DL, Value *Val) {
2696 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2699 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2700 // If the ISDOpcode is undefined, it was undefined before the promotion.
2703 // Otherwise, check if the promoted instruction is legal or not.
2704 return TLI.isOperationLegalOrCustom(
2705 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
2708 /// \brief Hepler class to perform type promotion.
2709 class TypePromotionHelper {
2710 /// \brief Utility function to check whether or not a sign or zero extension
2711 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2712 /// either using the operands of \p Inst or promoting \p Inst.
2713 /// The type of the extension is defined by \p IsSExt.
2714 /// In other words, check if:
2715 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2716 /// #1 Promotion applies:
2717 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2718 /// #2 Operand reuses:
2719 /// ext opnd1 to ConsideredExtType.
2720 /// \p PromotedInsts maps the instructions to their type before promotion.
2721 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2722 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2724 /// \brief Utility function to determine if \p OpIdx should be promoted when
2725 /// promoting \p Inst.
2726 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2727 return !(isa<SelectInst>(Inst) && OpIdx == 0);
2730 /// \brief Utility function to promote the operand of \p Ext when this
2731 /// operand is a promotable trunc or sext or zext.
2732 /// \p PromotedInsts maps the instructions to their type before promotion.
2733 /// \p CreatedInstsCost[out] contains the cost of all instructions
2734 /// created to promote the operand of Ext.
2735 /// Newly added extensions are inserted in \p Exts.
2736 /// Newly added truncates are inserted in \p Truncs.
2737 /// Should never be called directly.
2738 /// \return The promoted value which is used instead of Ext.
2739 static Value *promoteOperandForTruncAndAnyExt(
2740 Instruction *Ext, TypePromotionTransaction &TPT,
2741 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2742 SmallVectorImpl<Instruction *> *Exts,
2743 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
2745 /// \brief Utility function to promote the operand of \p Ext when this
2746 /// operand is promotable and is not a supported trunc or sext.
2747 /// \p PromotedInsts maps the instructions to their type before promotion.
2748 /// \p CreatedInstsCost[out] contains the cost of all the instructions
2749 /// created to promote the operand of Ext.
2750 /// Newly added extensions are inserted in \p Exts.
2751 /// Newly added truncates are inserted in \p Truncs.
2752 /// Should never be called directly.
2753 /// \return The promoted value which is used instead of Ext.
2754 static Value *promoteOperandForOther(Instruction *Ext,
2755 TypePromotionTransaction &TPT,
2756 InstrToOrigTy &PromotedInsts,
2757 unsigned &CreatedInstsCost,
2758 SmallVectorImpl<Instruction *> *Exts,
2759 SmallVectorImpl<Instruction *> *Truncs,
2760 const TargetLowering &TLI, bool IsSExt);
2762 /// \see promoteOperandForOther.
2763 static Value *signExtendOperandForOther(
2764 Instruction *Ext, TypePromotionTransaction &TPT,
2765 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2766 SmallVectorImpl<Instruction *> *Exts,
2767 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2768 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2769 Exts, Truncs, TLI, true);
2772 /// \see promoteOperandForOther.
2773 static Value *zeroExtendOperandForOther(
2774 Instruction *Ext, TypePromotionTransaction &TPT,
2775 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2776 SmallVectorImpl<Instruction *> *Exts,
2777 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2778 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2779 Exts, Truncs, TLI, false);
2783 /// Type for the utility function that promotes the operand of Ext.
2784 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2785 InstrToOrigTy &PromotedInsts,
2786 unsigned &CreatedInstsCost,
2787 SmallVectorImpl<Instruction *> *Exts,
2788 SmallVectorImpl<Instruction *> *Truncs,
2789 const TargetLowering &TLI);
2790 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2791 /// action to promote the operand of \p Ext instead of using Ext.
2792 /// \return NULL if no promotable action is possible with the current
2794 /// \p InsertedInsts keeps track of all the instructions inserted by the
2795 /// other CodeGenPrepare optimizations. This information is important
2796 /// because we do not want to promote these instructions as CodeGenPrepare
2797 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2798 /// \p PromotedInsts maps the instructions to their type before promotion.
2799 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
2800 const TargetLowering &TLI,
2801 const InstrToOrigTy &PromotedInsts);
2804 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2805 Type *ConsideredExtType,
2806 const InstrToOrigTy &PromotedInsts,
2808 // The promotion helper does not know how to deal with vector types yet.
2809 // To be able to fix that, we would need to fix the places where we
2810 // statically extend, e.g., constants and such.
2811 if (Inst->getType()->isVectorTy())
2814 // We can always get through zext.
2815 if (isa<ZExtInst>(Inst))
2818 // sext(sext) is ok too.
2819 if (IsSExt && isa<SExtInst>(Inst))
2822 // We can get through binary operator, if it is legal. In other words, the
2823 // binary operator must have a nuw or nsw flag.
2824 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2825 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2826 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2827 (IsSExt && BinOp->hasNoSignedWrap())))
2830 // Check if we can do the following simplification.
2831 // ext(trunc(opnd)) --> ext(opnd)
2832 if (!isa<TruncInst>(Inst))
2835 Value *OpndVal = Inst->getOperand(0);
2836 // Check if we can use this operand in the extension.
2837 // If the type is larger than the result type of the extension, we cannot.
2838 if (!OpndVal->getType()->isIntegerTy() ||
2839 OpndVal->getType()->getIntegerBitWidth() >
2840 ConsideredExtType->getIntegerBitWidth())
2843 // If the operand of the truncate is not an instruction, we will not have
2844 // any information on the dropped bits.
2845 // (Actually we could for constant but it is not worth the extra logic).
2846 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2850 // Check if the source of the type is narrow enough.
2851 // I.e., check that trunc just drops extended bits of the same kind of
2853 // #1 get the type of the operand and check the kind of the extended bits.
2854 const Type *OpndType;
2855 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2856 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
2857 OpndType = It->second.getPointer();
2858 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2859 OpndType = Opnd->getOperand(0)->getType();
2863 // #2 check that the truncate just drops extended bits.
2864 return Inst->getType()->getIntegerBitWidth() >=
2865 OpndType->getIntegerBitWidth();
2868 TypePromotionHelper::Action TypePromotionHelper::getAction(
2869 Instruction *Ext, const SetOfInstrs &InsertedInsts,
2870 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2871 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2872 "Unexpected instruction type");
2873 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2874 Type *ExtTy = Ext->getType();
2875 bool IsSExt = isa<SExtInst>(Ext);
2876 // If the operand of the extension is not an instruction, we cannot
2878 // If it, check we can get through.
2879 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2882 // Do not promote if the operand has been added by codegenprepare.
2883 // Otherwise, it means we are undoing an optimization that is likely to be
2884 // redone, thus causing potential infinite loop.
2885 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
2888 // SExt or Trunc instructions.
2889 // Return the related handler.
2890 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2891 isa<ZExtInst>(ExtOpnd))
2892 return promoteOperandForTruncAndAnyExt;
2894 // Regular instruction.
2895 // Abort early if we will have to insert non-free instructions.
2896 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2898 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2901 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2902 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2903 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2904 SmallVectorImpl<Instruction *> *Exts,
2905 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2906 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2907 // get through it and this method should not be called.
2908 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2909 Value *ExtVal = SExt;
2910 bool HasMergedNonFreeExt = false;
2911 if (isa<ZExtInst>(SExtOpnd)) {
2912 // Replace s|zext(zext(opnd))
2914 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
2916 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2917 TPT.replaceAllUsesWith(SExt, ZExt);
2918 TPT.eraseInstruction(SExt);
2921 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2923 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2925 CreatedInstsCost = 0;
2927 // Remove dead code.
2928 if (SExtOpnd->use_empty())
2929 TPT.eraseInstruction(SExtOpnd);
2931 // Check if the extension is still needed.
2932 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2933 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2936 Exts->push_back(ExtInst);
2937 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
2942 // At this point we have: ext ty opnd to ty.
2943 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2944 Value *NextVal = ExtInst->getOperand(0);
2945 TPT.eraseInstruction(ExtInst, NextVal);
2949 Value *TypePromotionHelper::promoteOperandForOther(
2950 Instruction *Ext, TypePromotionTransaction &TPT,
2951 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2952 SmallVectorImpl<Instruction *> *Exts,
2953 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
2955 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2956 // get through it and this method should not be called.
2957 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2958 CreatedInstsCost = 0;
2959 if (!ExtOpnd->hasOneUse()) {
2960 // ExtOpnd will be promoted.
2961 // All its uses, but Ext, will need to use a truncated value of the
2962 // promoted version.
2963 // Create the truncate now.
2964 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2965 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2966 ITrunc->removeFromParent();
2967 // Insert it just after the definition.
2968 ITrunc->insertAfter(ExtOpnd);
2970 Truncs->push_back(ITrunc);
2973 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2974 // Restore the operand of Ext (which has been replaced by the previous call
2975 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2976 TPT.setOperand(Ext, 0, ExtOpnd);
2979 // Get through the Instruction:
2980 // 1. Update its type.
2981 // 2. Replace the uses of Ext by Inst.
2982 // 3. Extend each operand that needs to be extended.
2984 // Remember the original type of the instruction before promotion.
2985 // This is useful to know that the high bits are sign extended bits.
2986 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2987 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2989 TPT.mutateType(ExtOpnd, Ext->getType());
2991 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2993 Instruction *ExtForOpnd = Ext;
2995 DEBUG(dbgs() << "Propagate Ext to operands\n");
2996 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2998 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2999 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
3000 !shouldExtOperand(ExtOpnd, OpIdx)) {
3001 DEBUG(dbgs() << "No need to propagate\n");
3004 // Check if we can statically extend the operand.
3005 Value *Opnd = ExtOpnd->getOperand(OpIdx);
3006 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
3007 DEBUG(dbgs() << "Statically extend\n");
3008 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
3009 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3010 : Cst->getValue().zext(BitWidth);
3011 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3014 // UndefValue are typed, so we have to statically sign extend them.
3015 if (isa<UndefValue>(Opnd)) {
3016 DEBUG(dbgs() << "Statically extend\n");
3017 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3021 // Otherwise we have to explicity sign extend the operand.
3022 // Check if Ext was reused to extend an operand.
3024 // If yes, create a new one.
3025 DEBUG(dbgs() << "More operands to ext\n");
3026 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3027 : TPT.createZExt(Ext, Opnd, Ext->getType());
3028 if (!isa<Instruction>(ValForExtOpnd)) {
3029 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
3032 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
3035 Exts->push_back(ExtForOpnd);
3036 TPT.setOperand(ExtForOpnd, 0, Opnd);
3038 // Move the sign extension before the insertion point.
3039 TPT.moveBefore(ExtForOpnd, ExtOpnd);
3040 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
3041 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
3042 // If more sext are required, new instructions will have to be created.
3043 ExtForOpnd = nullptr;
3045 if (ExtForOpnd == Ext) {
3046 DEBUG(dbgs() << "Extension is useless now\n");
3047 TPT.eraseInstruction(Ext);
3052 /// Check whether or not promoting an instruction to a wider type is profitable.
3053 /// \p NewCost gives the cost of extension instructions created by the
3055 /// \p OldCost gives the cost of extension instructions before the promotion
3056 /// plus the number of instructions that have been
3057 /// matched in the addressing mode the promotion.
3058 /// \p PromotedOperand is the value that has been promoted.
3059 /// \return True if the promotion is profitable, false otherwise.
3060 bool AddressingModeMatcher::isPromotionProfitable(
3061 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
3062 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
3063 // The cost of the new extensions is greater than the cost of the
3064 // old extension plus what we folded.
3065 // This is not profitable.
3066 if (NewCost > OldCost)
3068 if (NewCost < OldCost)
3070 // The promotion is neutral but it may help folding the sign extension in
3071 // loads for instance.
3072 // Check that we did not create an illegal instruction.
3073 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
3076 /// Given an instruction or constant expr, see if we can fold the operation
3077 /// into the addressing mode. If so, update the addressing mode and return
3078 /// true, otherwise return false without modifying AddrMode.
3079 /// If \p MovedAway is not NULL, it contains the information of whether or
3080 /// not AddrInst has to be folded into the addressing mode on success.
3081 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
3082 /// because it has been moved away.
3083 /// Thus AddrInst must not be added in the matched instructions.
3084 /// This state can happen when AddrInst is a sext, since it may be moved away.
3085 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
3086 /// not be referenced anymore.
3087 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
3090 // Avoid exponential behavior on extremely deep expression trees.
3091 if (Depth >= 5) return false;
3093 // By default, all matched instructions stay in place.
3098 case Instruction::PtrToInt:
3099 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3100 return matchAddr(AddrInst->getOperand(0), Depth);
3101 case Instruction::IntToPtr: {
3102 auto AS = AddrInst->getType()->getPointerAddressSpace();
3103 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3104 // This inttoptr is a no-op if the integer type is pointer sized.
3105 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3106 return matchAddr(AddrInst->getOperand(0), Depth);
3109 case Instruction::BitCast:
3110 // BitCast is always a noop, and we can handle it as long as it is
3111 // int->int or pointer->pointer (we don't want int<->fp or something).
3112 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
3113 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
3114 // Don't touch identity bitcasts. These were probably put here by LSR,
3115 // and we don't want to mess around with them. Assume it knows what it
3117 AddrInst->getOperand(0)->getType() != AddrInst->getType())
3118 return matchAddr(AddrInst->getOperand(0), Depth);
3120 case Instruction::AddrSpaceCast: {
3122 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3123 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3124 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3125 return matchAddr(AddrInst->getOperand(0), Depth);
3128 case Instruction::Add: {
3129 // Check to see if we can merge in the RHS then the LHS. If so, we win.
3130 ExtAddrMode BackupAddrMode = AddrMode;
3131 unsigned OldSize = AddrModeInsts.size();
3132 // Start a transaction at this point.
3133 // The LHS may match but not the RHS.
3134 // Therefore, we need a higher level restoration point to undo partially
3135 // matched operation.
3136 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3137 TPT.getRestorationPoint();
3139 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
3140 matchAddr(AddrInst->getOperand(0), Depth+1))
3143 // Restore the old addr mode info.
3144 AddrMode = BackupAddrMode;
3145 AddrModeInsts.resize(OldSize);
3146 TPT.rollback(LastKnownGood);
3148 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
3149 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
3150 matchAddr(AddrInst->getOperand(1), Depth+1))
3153 // Otherwise we definitely can't merge the ADD in.
3154 AddrMode = BackupAddrMode;
3155 AddrModeInsts.resize(OldSize);
3156 TPT.rollback(LastKnownGood);
3159 //case Instruction::Or:
3160 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
3162 case Instruction::Mul:
3163 case Instruction::Shl: {
3164 // Can only handle X*C and X << C.
3165 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
3168 int64_t Scale = RHS->getSExtValue();
3169 if (Opcode == Instruction::Shl)
3170 Scale = 1LL << Scale;
3172 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
3174 case Instruction::GetElementPtr: {
3175 // Scan the GEP. We check it if it contains constant offsets and at most
3176 // one variable offset.
3177 int VariableOperand = -1;
3178 unsigned VariableScale = 0;
3180 int64_t ConstantOffset = 0;
3181 gep_type_iterator GTI = gep_type_begin(AddrInst);
3182 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
3183 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
3184 const StructLayout *SL = DL.getStructLayout(STy);
3186 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
3187 ConstantOffset += SL->getElementOffset(Idx);
3189 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
3190 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
3191 ConstantOffset += CI->getSExtValue()*TypeSize;
3192 } else if (TypeSize) { // Scales of zero don't do anything.
3193 // We only allow one variable index at the moment.
3194 if (VariableOperand != -1)
3197 // Remember the variable index.
3198 VariableOperand = i;
3199 VariableScale = TypeSize;
3204 // A common case is for the GEP to only do a constant offset. In this case,
3205 // just add it to the disp field and check validity.
3206 if (VariableOperand == -1) {
3207 AddrMode.BaseOffs += ConstantOffset;
3208 if (ConstantOffset == 0 ||
3209 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
3210 // Check to see if we can fold the base pointer in too.
3211 if (matchAddr(AddrInst->getOperand(0), Depth+1))
3214 AddrMode.BaseOffs -= ConstantOffset;
3218 // Save the valid addressing mode in case we can't match.
3219 ExtAddrMode BackupAddrMode = AddrMode;
3220 unsigned OldSize = AddrModeInsts.size();
3222 // See if the scale and offset amount is valid for this target.
3223 AddrMode.BaseOffs += ConstantOffset;
3225 // Match the base operand of the GEP.
3226 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
3227 // If it couldn't be matched, just stuff the value in a register.
3228 if (AddrMode.HasBaseReg) {
3229 AddrMode = BackupAddrMode;
3230 AddrModeInsts.resize(OldSize);
3233 AddrMode.HasBaseReg = true;
3234 AddrMode.BaseReg = AddrInst->getOperand(0);
3237 // Match the remaining variable portion of the GEP.
3238 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
3240 // If it couldn't be matched, try stuffing the base into a register
3241 // instead of matching it, and retrying the match of the scale.
3242 AddrMode = BackupAddrMode;
3243 AddrModeInsts.resize(OldSize);
3244 if (AddrMode.HasBaseReg)
3246 AddrMode.HasBaseReg = true;
3247 AddrMode.BaseReg = AddrInst->getOperand(0);
3248 AddrMode.BaseOffs += ConstantOffset;
3249 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
3250 VariableScale, Depth)) {
3251 // If even that didn't work, bail.
3252 AddrMode = BackupAddrMode;
3253 AddrModeInsts.resize(OldSize);
3260 case Instruction::SExt:
3261 case Instruction::ZExt: {
3262 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
3266 // Try to move this ext out of the way of the addressing mode.
3267 // Ask for a method for doing so.
3268 TypePromotionHelper::Action TPH =
3269 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
3273 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3274 TPT.getRestorationPoint();
3275 unsigned CreatedInstsCost = 0;
3276 unsigned ExtCost = !TLI.isExtFree(Ext);
3277 Value *PromotedOperand =
3278 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
3279 // SExt has been moved away.
3280 // Thus either it will be rematched later in the recursive calls or it is
3281 // gone. Anyway, we must not fold it into the addressing mode at this point.
3285 // addr = gep base, idx
3287 // promotedOpnd = ext opnd <- no match here
3288 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
3289 // addr = gep base, op <- match
3293 assert(PromotedOperand &&
3294 "TypePromotionHelper should have filtered out those cases");
3296 ExtAddrMode BackupAddrMode = AddrMode;
3297 unsigned OldSize = AddrModeInsts.size();
3299 if (!matchAddr(PromotedOperand, Depth) ||
3300 // The total of the new cost is equal to the cost of the created
3302 // The total of the old cost is equal to the cost of the extension plus
3303 // what we have saved in the addressing mode.
3304 !isPromotionProfitable(CreatedInstsCost,
3305 ExtCost + (AddrModeInsts.size() - OldSize),
3307 AddrMode = BackupAddrMode;
3308 AddrModeInsts.resize(OldSize);
3309 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
3310 TPT.rollback(LastKnownGood);
3319 /// If we can, try to add the value of 'Addr' into the current addressing mode.
3320 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
3321 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
3324 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
3325 // Start a transaction at this point that we will rollback if the matching
3327 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3328 TPT.getRestorationPoint();
3329 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
3330 // Fold in immediates if legal for the target.
3331 AddrMode.BaseOffs += CI->getSExtValue();
3332 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3334 AddrMode.BaseOffs -= CI->getSExtValue();
3335 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
3336 // If this is a global variable, try to fold it into the addressing mode.
3337 if (!AddrMode.BaseGV) {
3338 AddrMode.BaseGV = GV;
3339 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3341 AddrMode.BaseGV = nullptr;
3343 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
3344 ExtAddrMode BackupAddrMode = AddrMode;
3345 unsigned OldSize = AddrModeInsts.size();
3347 // Check to see if it is possible to fold this operation.
3348 bool MovedAway = false;
3349 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
3350 // This instruction may have been moved away. If so, there is nothing
3354 // Okay, it's possible to fold this. Check to see if it is actually
3355 // *profitable* to do so. We use a simple cost model to avoid increasing
3356 // register pressure too much.
3357 if (I->hasOneUse() ||
3358 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
3359 AddrModeInsts.push_back(I);
3363 // It isn't profitable to do this, roll back.
3364 //cerr << "NOT FOLDING: " << *I;
3365 AddrMode = BackupAddrMode;
3366 AddrModeInsts.resize(OldSize);
3367 TPT.rollback(LastKnownGood);
3369 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
3370 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
3372 TPT.rollback(LastKnownGood);
3373 } else if (isa<ConstantPointerNull>(Addr)) {
3374 // Null pointer gets folded without affecting the addressing mode.
3378 // Worse case, the target should support [reg] addressing modes. :)
3379 if (!AddrMode.HasBaseReg) {
3380 AddrMode.HasBaseReg = true;
3381 AddrMode.BaseReg = Addr;
3382 // Still check for legality in case the target supports [imm] but not [i+r].
3383 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3385 AddrMode.HasBaseReg = false;
3386 AddrMode.BaseReg = nullptr;
3389 // If the base register is already taken, see if we can do [r+r].
3390 if (AddrMode.Scale == 0) {
3392 AddrMode.ScaledReg = Addr;
3393 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3396 AddrMode.ScaledReg = nullptr;
3399 TPT.rollback(LastKnownGood);
3403 /// Check to see if all uses of OpVal by the specified inline asm call are due
3404 /// to memory operands. If so, return true, otherwise return false.
3405 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
3406 const TargetMachine &TM) {
3407 const Function *F = CI->getParent()->getParent();
3408 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
3409 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
3410 TargetLowering::AsmOperandInfoVector TargetConstraints =
3411 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
3412 ImmutableCallSite(CI));
3413 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3414 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3416 // Compute the constraint code and ConstraintType to use.
3417 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3419 // If this asm operand is our Value*, and if it isn't an indirect memory
3420 // operand, we can't fold it!
3421 if (OpInfo.CallOperandVal == OpVal &&
3422 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
3423 !OpInfo.isIndirect))
3430 /// Recursively walk all the uses of I until we find a memory use.
3431 /// If we find an obviously non-foldable instruction, return true.
3432 /// Add the ultimately found memory instructions to MemoryUses.
3433 static bool FindAllMemoryUses(
3435 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
3436 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
3437 // If we already considered this instruction, we're done.
3438 if (!ConsideredInsts.insert(I).second)
3441 // If this is an obviously unfoldable instruction, bail out.
3442 if (!MightBeFoldableInst(I))
3445 // Loop over all the uses, recursively processing them.
3446 for (Use &U : I->uses()) {
3447 Instruction *UserI = cast<Instruction>(U.getUser());
3449 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
3450 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
3454 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
3455 unsigned opNo = U.getOperandNo();
3456 if (opNo == 0) return true; // Storing addr, not into addr.
3457 MemoryUses.push_back(std::make_pair(SI, opNo));
3461 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
3462 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
3463 if (!IA) return true;
3465 // If this is a memory operand, we're cool, otherwise bail out.
3466 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
3471 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
3478 /// Return true if Val is already known to be live at the use site that we're
3479 /// folding it into. If so, there is no cost to include it in the addressing
3480 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
3481 /// instruction already.
3482 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
3483 Value *KnownLive2) {
3484 // If Val is either of the known-live values, we know it is live!
3485 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
3488 // All values other than instructions and arguments (e.g. constants) are live.
3489 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
3491 // If Val is a constant sized alloca in the entry block, it is live, this is
3492 // true because it is just a reference to the stack/frame pointer, which is
3493 // live for the whole function.
3494 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
3495 if (AI->isStaticAlloca())
3498 // Check to see if this value is already used in the memory instruction's
3499 // block. If so, it's already live into the block at the very least, so we
3500 // can reasonably fold it.
3501 return Val->isUsedInBasicBlock(MemoryInst->getParent());
3504 /// It is possible for the addressing mode of the machine to fold the specified
3505 /// instruction into a load or store that ultimately uses it.
3506 /// However, the specified instruction has multiple uses.
3507 /// Given this, it may actually increase register pressure to fold it
3508 /// into the load. For example, consider this code:
3512 /// use(Y) -> nonload/store
3516 /// In this case, Y has multiple uses, and can be folded into the load of Z
3517 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
3518 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
3519 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
3520 /// number of computations either.
3522 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
3523 /// X was live across 'load Z' for other reasons, we actually *would* want to
3524 /// fold the addressing mode in the Z case. This would make Y die earlier.
3525 bool AddressingModeMatcher::
3526 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
3527 ExtAddrMode &AMAfter) {
3528 if (IgnoreProfitability) return true;
3530 // AMBefore is the addressing mode before this instruction was folded into it,
3531 // and AMAfter is the addressing mode after the instruction was folded. Get
3532 // the set of registers referenced by AMAfter and subtract out those
3533 // referenced by AMBefore: this is the set of values which folding in this
3534 // address extends the lifetime of.
3536 // Note that there are only two potential values being referenced here,
3537 // BaseReg and ScaleReg (global addresses are always available, as are any
3538 // folded immediates).
3539 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
3541 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
3542 // lifetime wasn't extended by adding this instruction.
3543 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3545 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3546 ScaledReg = nullptr;
3548 // If folding this instruction (and it's subexprs) didn't extend any live
3549 // ranges, we're ok with it.
3550 if (!BaseReg && !ScaledReg)
3553 // If all uses of this instruction are ultimately load/store/inlineasm's,
3554 // check to see if their addressing modes will include this instruction. If
3555 // so, we can fold it into all uses, so it doesn't matter if it has multiple
3557 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
3558 SmallPtrSet<Instruction*, 16> ConsideredInsts;
3559 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
3560 return false; // Has a non-memory, non-foldable use!
3562 // Now that we know that all uses of this instruction are part of a chain of
3563 // computation involving only operations that could theoretically be folded
3564 // into a memory use, loop over each of these uses and see if they could
3565 // *actually* fold the instruction.
3566 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
3567 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
3568 Instruction *User = MemoryUses[i].first;
3569 unsigned OpNo = MemoryUses[i].second;
3571 // Get the access type of this use. If the use isn't a pointer, we don't
3572 // know what it accesses.
3573 Value *Address = User->getOperand(OpNo);
3574 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
3577 Type *AddressAccessTy = AddrTy->getElementType();
3578 unsigned AS = AddrTy->getAddressSpace();
3580 // Do a match against the root of this address, ignoring profitability. This
3581 // will tell us if the addressing mode for the memory operation will
3582 // *actually* cover the shared instruction.
3584 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3585 TPT.getRestorationPoint();
3586 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
3587 MemoryInst, Result, InsertedInsts,
3588 PromotedInsts, TPT);
3589 Matcher.IgnoreProfitability = true;
3590 bool Success = Matcher.matchAddr(Address, 0);
3591 (void)Success; assert(Success && "Couldn't select *anything*?");
3593 // The match was to check the profitability, the changes made are not
3594 // part of the original matcher. Therefore, they should be dropped
3595 // otherwise the original matcher will not present the right state.
3596 TPT.rollback(LastKnownGood);
3598 // If the match didn't cover I, then it won't be shared by it.
3599 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3600 I) == MatchedAddrModeInsts.end())
3603 MatchedAddrModeInsts.clear();
3609 } // end anonymous namespace
3611 /// Return true if the specified values are defined in a
3612 /// different basic block than BB.
3613 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3614 if (Instruction *I = dyn_cast<Instruction>(V))
3615 return I->getParent() != BB;
3619 /// Load and Store Instructions often have addressing modes that can do
3620 /// significant amounts of computation. As such, instruction selection will try
3621 /// to get the load or store to do as much computation as possible for the
3622 /// program. The problem is that isel can only see within a single block. As
3623 /// such, we sink as much legal addressing mode work into the block as possible.
3625 /// This method is used to optimize both load/store and inline asms with memory
3627 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3628 Type *AccessTy, unsigned AddrSpace) {
3631 // Try to collapse single-value PHI nodes. This is necessary to undo
3632 // unprofitable PRE transformations.
3633 SmallVector<Value*, 8> worklist;
3634 SmallPtrSet<Value*, 16> Visited;
3635 worklist.push_back(Addr);
3637 // Use a worklist to iteratively look through PHI nodes, and ensure that
3638 // the addressing mode obtained from the non-PHI roots of the graph
3640 Value *Consensus = nullptr;
3641 unsigned NumUsesConsensus = 0;
3642 bool IsNumUsesConsensusValid = false;
3643 SmallVector<Instruction*, 16> AddrModeInsts;
3644 ExtAddrMode AddrMode;
3645 TypePromotionTransaction TPT;
3646 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3647 TPT.getRestorationPoint();
3648 while (!worklist.empty()) {
3649 Value *V = worklist.back();
3650 worklist.pop_back();
3652 // Break use-def graph loops.
3653 if (!Visited.insert(V).second) {
3654 Consensus = nullptr;
3658 // For a PHI node, push all of its incoming values.
3659 if (PHINode *P = dyn_cast<PHINode>(V)) {
3660 for (Value *IncValue : P->incoming_values())
3661 worklist.push_back(IncValue);
3665 // For non-PHIs, determine the addressing mode being computed.
3666 SmallVector<Instruction*, 16> NewAddrModeInsts;
3667 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3668 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
3669 InsertedInsts, PromotedInsts, TPT);
3671 // This check is broken into two cases with very similar code to avoid using
3672 // getNumUses() as much as possible. Some values have a lot of uses, so
3673 // calling getNumUses() unconditionally caused a significant compile-time
3677 AddrMode = NewAddrMode;
3678 AddrModeInsts = NewAddrModeInsts;
3680 } else if (NewAddrMode == AddrMode) {
3681 if (!IsNumUsesConsensusValid) {
3682 NumUsesConsensus = Consensus->getNumUses();
3683 IsNumUsesConsensusValid = true;
3686 // Ensure that the obtained addressing mode is equivalent to that obtained
3687 // for all other roots of the PHI traversal. Also, when choosing one
3688 // such root as representative, select the one with the most uses in order
3689 // to keep the cost modeling heuristics in AddressingModeMatcher
3691 unsigned NumUses = V->getNumUses();
3692 if (NumUses > NumUsesConsensus) {
3694 NumUsesConsensus = NumUses;
3695 AddrModeInsts = NewAddrModeInsts;
3700 Consensus = nullptr;
3704 // If the addressing mode couldn't be determined, or if multiple different
3705 // ones were determined, bail out now.
3707 TPT.rollback(LastKnownGood);
3712 // Check to see if any of the instructions supersumed by this addr mode are
3713 // non-local to I's BB.
3714 bool AnyNonLocal = false;
3715 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3716 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3722 // If all the instructions matched are already in this BB, don't do anything.
3724 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3728 // Insert this computation right after this user. Since our caller is
3729 // scanning from the top of the BB to the bottom, reuse of the expr are
3730 // guaranteed to happen later.
3731 IRBuilder<> Builder(MemoryInst);
3733 // Now that we determined the addressing expression we want to use and know
3734 // that we have to sink it into this block. Check to see if we have already
3735 // done this for some other load/store instr in this block. If so, reuse the
3737 Value *&SunkAddr = SunkAddrs[Addr];
3739 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3740 << *MemoryInst << "\n");
3741 if (SunkAddr->getType() != Addr->getType())
3742 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3743 } else if (AddrSinkUsingGEPs ||
3744 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3745 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3747 // By default, we use the GEP-based method when AA is used later. This
3748 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3749 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3750 << *MemoryInst << "\n");
3751 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3752 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3754 // First, find the pointer.
3755 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3756 ResultPtr = AddrMode.BaseReg;
3757 AddrMode.BaseReg = nullptr;
3760 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3761 // We can't add more than one pointer together, nor can we scale a
3762 // pointer (both of which seem meaningless).
3763 if (ResultPtr || AddrMode.Scale != 1)
3766 ResultPtr = AddrMode.ScaledReg;
3770 if (AddrMode.BaseGV) {
3774 ResultPtr = AddrMode.BaseGV;
3777 // If the real base value actually came from an inttoptr, then the matcher
3778 // will look through it and provide only the integer value. In that case,
3780 if (!ResultPtr && AddrMode.BaseReg) {
3782 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3783 AddrMode.BaseReg = nullptr;
3784 } else if (!ResultPtr && AddrMode.Scale == 1) {
3786 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3791 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3792 SunkAddr = Constant::getNullValue(Addr->getType());
3793 } else if (!ResultPtr) {
3797 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3798 Type *I8Ty = Builder.getInt8Ty();
3800 // Start with the base register. Do this first so that subsequent address
3801 // matching finds it last, which will prevent it from trying to match it
3802 // as the scaled value in case it happens to be a mul. That would be
3803 // problematic if we've sunk a different mul for the scale, because then
3804 // we'd end up sinking both muls.
3805 if (AddrMode.BaseReg) {
3806 Value *V = AddrMode.BaseReg;
3807 if (V->getType() != IntPtrTy)
3808 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3813 // Add the scale value.
3814 if (AddrMode.Scale) {
3815 Value *V = AddrMode.ScaledReg;
3816 if (V->getType() == IntPtrTy) {
3818 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3819 cast<IntegerType>(V->getType())->getBitWidth()) {
3820 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3822 // It is only safe to sign extend the BaseReg if we know that the math
3823 // required to create it did not overflow before we extend it. Since
3824 // the original IR value was tossed in favor of a constant back when
3825 // the AddrMode was created we need to bail out gracefully if widths
3826 // do not match instead of extending it.
3827 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3828 if (I && (ResultIndex != AddrMode.BaseReg))
3829 I->eraseFromParent();
3833 if (AddrMode.Scale != 1)
3834 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3837 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3842 // Add in the Base Offset if present.
3843 if (AddrMode.BaseOffs) {
3844 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3846 // We need to add this separately from the scale above to help with
3847 // SDAG consecutive load/store merging.
3848 if (ResultPtr->getType() != I8PtrTy)
3849 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3850 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3857 SunkAddr = ResultPtr;
3859 if (ResultPtr->getType() != I8PtrTy)
3860 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3861 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3864 if (SunkAddr->getType() != Addr->getType())
3865 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3868 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3869 << *MemoryInst << "\n");
3870 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3871 Value *Result = nullptr;
3873 // Start with the base register. Do this first so that subsequent address
3874 // matching finds it last, which will prevent it from trying to match it
3875 // as the scaled value in case it happens to be a mul. That would be
3876 // problematic if we've sunk a different mul for the scale, because then
3877 // we'd end up sinking both muls.
3878 if (AddrMode.BaseReg) {
3879 Value *V = AddrMode.BaseReg;
3880 if (V->getType()->isPointerTy())
3881 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3882 if (V->getType() != IntPtrTy)
3883 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3887 // Add the scale value.
3888 if (AddrMode.Scale) {
3889 Value *V = AddrMode.ScaledReg;
3890 if (V->getType() == IntPtrTy) {
3892 } else if (V->getType()->isPointerTy()) {
3893 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3894 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3895 cast<IntegerType>(V->getType())->getBitWidth()) {
3896 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3898 // It is only safe to sign extend the BaseReg if we know that the math
3899 // required to create it did not overflow before we extend it. Since
3900 // the original IR value was tossed in favor of a constant back when
3901 // the AddrMode was created we need to bail out gracefully if widths
3902 // do not match instead of extending it.
3903 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3904 if (I && (Result != AddrMode.BaseReg))
3905 I->eraseFromParent();
3908 if (AddrMode.Scale != 1)
3909 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3912 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3917 // Add in the BaseGV if present.
3918 if (AddrMode.BaseGV) {
3919 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3921 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3926 // Add in the Base Offset if present.
3927 if (AddrMode.BaseOffs) {
3928 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3930 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3936 SunkAddr = Constant::getNullValue(Addr->getType());
3938 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3941 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3943 // If we have no uses, recursively delete the value and all dead instructions
3945 if (Repl->use_empty()) {
3946 // This can cause recursive deletion, which can invalidate our iterator.
3947 // Use a WeakVH to hold onto it in case this happens.
3948 WeakVH IterHandle(&*CurInstIterator);
3949 BasicBlock *BB = CurInstIterator->getParent();
3951 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3953 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
3954 // If the iterator instruction was recursively deleted, start over at the
3955 // start of the block.
3956 CurInstIterator = BB->begin();
3964 /// If there are any memory operands, use OptimizeMemoryInst to sink their
3965 /// address computing into the block when possible / profitable.
3966 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
3967 bool MadeChange = false;
3969 const TargetRegisterInfo *TRI =
3970 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3971 TargetLowering::AsmOperandInfoVector TargetConstraints =
3972 TLI->ParseConstraints(*DL, TRI, CS);
3974 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3975 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3977 // Compute the constraint code and ConstraintType to use.
3978 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3980 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3981 OpInfo.isIndirect) {
3982 Value *OpVal = CS->getArgOperand(ArgNo++);
3983 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
3984 } else if (OpInfo.Type == InlineAsm::isInput)
3991 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3992 /// sign extensions.
3993 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3994 assert(!Inst->use_empty() && "Input must have at least one use");
3995 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3996 bool IsSExt = isa<SExtInst>(FirstUser);
3997 Type *ExtTy = FirstUser->getType();
3998 for (const User *U : Inst->users()) {
3999 const Instruction *UI = cast<Instruction>(U);
4000 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
4002 Type *CurTy = UI->getType();
4003 // Same input and output types: Same instruction after CSE.
4007 // If IsSExt is true, we are in this situation:
4009 // b = sext ty1 a to ty2
4010 // c = sext ty1 a to ty3
4011 // Assuming ty2 is shorter than ty3, this could be turned into:
4013 // b = sext ty1 a to ty2
4014 // c = sext ty2 b to ty3
4015 // However, the last sext is not free.
4019 // This is a ZExt, maybe this is free to extend from one type to another.
4020 // In that case, we would not account for a different use.
4023 if (ExtTy->getScalarType()->getIntegerBitWidth() >
4024 CurTy->getScalarType()->getIntegerBitWidth()) {
4032 if (!TLI.isZExtFree(NarrowTy, LargeTy))
4035 // All uses are the same or can be derived from one another for free.
4039 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
4040 /// load instruction.
4041 /// If an ext(load) can be formed, it is returned via \p LI for the load
4042 /// and \p Inst for the extension.
4043 /// Otherwise LI == nullptr and Inst == nullptr.
4044 /// When some promotion happened, \p TPT contains the proper state to
4047 /// \return true when promoting was necessary to expose the ext(load)
4048 /// opportunity, false otherwise.
4052 /// %ld = load i32* %addr
4053 /// %add = add nuw i32 %ld, 4
4054 /// %zext = zext i32 %add to i64
4058 /// %ld = load i32* %addr
4059 /// %zext = zext i32 %ld to i64
4060 /// %add = add nuw i64 %zext, 4
4062 /// Thanks to the promotion, we can match zext(load i32*) to i64.
4063 bool CodeGenPrepare::extLdPromotion(TypePromotionTransaction &TPT,
4064 LoadInst *&LI, Instruction *&Inst,
4065 const SmallVectorImpl<Instruction *> &Exts,
4066 unsigned CreatedInstsCost = 0) {
4067 // Iterate over all the extensions to see if one form an ext(load).
4068 for (auto I : Exts) {
4069 // Check if we directly have ext(load).
4070 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
4072 // No promotion happened here.
4075 // Check whether or not we want to do any promotion.
4076 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
4078 // Get the action to perform the promotion.
4079 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
4080 I, InsertedInsts, *TLI, PromotedInsts);
4081 // Check if we can promote.
4084 // Save the current state.
4085 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4086 TPT.getRestorationPoint();
4087 SmallVector<Instruction *, 4> NewExts;
4088 unsigned NewCreatedInstsCost = 0;
4089 unsigned ExtCost = !TLI->isExtFree(I);
4091 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
4092 &NewExts, nullptr, *TLI);
4093 assert(PromotedVal &&
4094 "TypePromotionHelper should have filtered out those cases");
4096 // We would be able to merge only one extension in a load.
4097 // Therefore, if we have more than 1 new extension we heuristically
4098 // cut this search path, because it means we degrade the code quality.
4099 // With exactly 2, the transformation is neutral, because we will merge
4100 // one extension but leave one. However, we optimistically keep going,
4101 // because the new extension may be removed too.
4102 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4103 TotalCreatedInstsCost -= ExtCost;
4104 if (!StressExtLdPromotion &&
4105 (TotalCreatedInstsCost > 1 ||
4106 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4107 // The promotion is not profitable, rollback to the previous state.
4108 TPT.rollback(LastKnownGood);
4111 // The promotion is profitable.
4112 // Check if it exposes an ext(load).
4113 (void)extLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
4114 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4115 // If we have created a new extension, i.e., now we have two
4116 // extensions. We must make sure one of them is merged with
4117 // the load, otherwise we may degrade the code quality.
4118 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
4119 // Promotion happened.
4121 // If this does not help to expose an ext(load) then, rollback.
4122 TPT.rollback(LastKnownGood);
4124 // None of the extension can form an ext(load).
4130 /// Move a zext or sext fed by a load into the same basic block as the load,
4131 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
4132 /// extend into the load.
4133 /// \p I[in/out] the extension may be modified during the process if some
4134 /// promotions apply.
4136 bool CodeGenPrepare::moveExtToFormExtLoad(Instruction *&I) {
4137 // Try to promote a chain of computation if it allows to form
4138 // an extended load.
4139 TypePromotionTransaction TPT;
4140 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4141 TPT.getRestorationPoint();
4142 SmallVector<Instruction *, 1> Exts;
4144 // Look for a load being extended.
4145 LoadInst *LI = nullptr;
4146 Instruction *OldExt = I;
4147 bool HasPromoted = extLdPromotion(TPT, LI, I, Exts);
4149 assert(!HasPromoted && !LI && "If we did not match any load instruction "
4150 "the code must remain the same");
4155 // If they're already in the same block, there's nothing to do.
4156 // Make the cheap checks first if we did not promote.
4157 // If we promoted, we need to check if it is indeed profitable.
4158 if (!HasPromoted && LI->getParent() == I->getParent())
4161 EVT VT = TLI->getValueType(*DL, I->getType());
4162 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
4164 // If the load has other users and the truncate is not free, this probably
4165 // isn't worthwhile.
4166 if (!LI->hasOneUse() && TLI &&
4167 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
4168 !TLI->isTruncateFree(I->getType(), LI->getType())) {
4170 TPT.rollback(LastKnownGood);
4174 // Check whether the target supports casts folded into loads.
4176 if (isa<ZExtInst>(I))
4177 LType = ISD::ZEXTLOAD;
4179 assert(isa<SExtInst>(I) && "Unexpected ext type!");
4180 LType = ISD::SEXTLOAD;
4182 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
4184 TPT.rollback(LastKnownGood);
4188 // Move the extend into the same block as the load, so that SelectionDAG
4191 I->removeFromParent();
4197 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
4198 BasicBlock *DefBB = I->getParent();
4200 // If the result of a {s|z}ext and its source are both live out, rewrite all
4201 // other uses of the source with result of extension.
4202 Value *Src = I->getOperand(0);
4203 if (Src->hasOneUse())
4206 // Only do this xform if truncating is free.
4207 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
4210 // Only safe to perform the optimization if the source is also defined in
4212 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
4215 bool DefIsLiveOut = false;
4216 for (User *U : I->users()) {
4217 Instruction *UI = cast<Instruction>(U);
4219 // Figure out which BB this ext is used in.
4220 BasicBlock *UserBB = UI->getParent();
4221 if (UserBB == DefBB) continue;
4222 DefIsLiveOut = true;
4228 // Make sure none of the uses are PHI nodes.
4229 for (User *U : Src->users()) {
4230 Instruction *UI = cast<Instruction>(U);
4231 BasicBlock *UserBB = UI->getParent();
4232 if (UserBB == DefBB) continue;
4233 // Be conservative. We don't want this xform to end up introducing
4234 // reloads just before load / store instructions.
4235 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
4239 // InsertedTruncs - Only insert one trunc in each block once.
4240 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
4242 bool MadeChange = false;
4243 for (Use &U : Src->uses()) {
4244 Instruction *User = cast<Instruction>(U.getUser());
4246 // Figure out which BB this ext is used in.
4247 BasicBlock *UserBB = User->getParent();
4248 if (UserBB == DefBB) continue;
4250 // Both src and def are live in this block. Rewrite the use.
4251 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
4253 if (!InsertedTrunc) {
4254 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4255 assert(InsertPt != UserBB->end());
4256 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
4257 InsertedInsts.insert(InsertedTrunc);
4260 // Replace a use of the {s|z}ext source with a use of the result.
4269 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
4270 // just after the load if the target can fold this into one extload instruction,
4271 // with the hope of eliminating some of the other later "and" instructions using
4272 // the loaded value. "and"s that are made trivially redundant by the insertion
4273 // of the new "and" are removed by this function, while others (e.g. those whose
4274 // path from the load goes through a phi) are left for isel to potentially
4307 // becomes (after a call to optimizeLoadExt for each load):
4311 // x1' = and x1, 0xff
4315 // x2' = and x2, 0xff
4322 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
4324 if (!Load->isSimple() ||
4325 !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
4328 // Skip loads we've already transformed or have no reason to transform.
4329 if (Load->hasOneUse()) {
4330 User *LoadUser = *Load->user_begin();
4331 if (cast<Instruction>(LoadUser)->getParent() == Load->getParent() &&
4332 !dyn_cast<PHINode>(LoadUser))
4336 // Look at all uses of Load, looking through phis, to determine how many bits
4337 // of the loaded value are needed.
4338 SmallVector<Instruction *, 8> WorkList;
4339 SmallPtrSet<Instruction *, 16> Visited;
4340 SmallVector<Instruction *, 8> AndsToMaybeRemove;
4341 for (auto *U : Load->users())
4342 WorkList.push_back(cast<Instruction>(U));
4344 EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
4345 unsigned BitWidth = LoadResultVT.getSizeInBits();
4346 APInt DemandBits(BitWidth, 0);
4347 APInt WidestAndBits(BitWidth, 0);
4349 while (!WorkList.empty()) {
4350 Instruction *I = WorkList.back();
4351 WorkList.pop_back();
4353 // Break use-def graph loops.
4354 if (!Visited.insert(I).second)
4357 // For a PHI node, push all of its users.
4358 if (auto *Phi = dyn_cast<PHINode>(I)) {
4359 for (auto *U : Phi->users())
4360 WorkList.push_back(cast<Instruction>(U));
4364 switch (I->getOpcode()) {
4365 case llvm::Instruction::And: {
4366 auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
4369 APInt AndBits = AndC->getValue();
4370 DemandBits |= AndBits;
4371 // Keep track of the widest and mask we see.
4372 if (AndBits.ugt(WidestAndBits))
4373 WidestAndBits = AndBits;
4374 if (AndBits == WidestAndBits && I->getOperand(0) == Load)
4375 AndsToMaybeRemove.push_back(I);
4379 case llvm::Instruction::Shl: {
4380 auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
4383 uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
4384 auto ShlDemandBits = APInt::getAllOnesValue(BitWidth).lshr(ShiftAmt);
4385 DemandBits |= ShlDemandBits;
4389 case llvm::Instruction::Trunc: {
4390 EVT TruncVT = TLI->getValueType(*DL, I->getType());
4391 unsigned TruncBitWidth = TruncVT.getSizeInBits();
4392 auto TruncBits = APInt::getAllOnesValue(TruncBitWidth).zext(BitWidth);
4393 DemandBits |= TruncBits;
4402 uint32_t ActiveBits = DemandBits.getActiveBits();
4403 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
4404 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
4405 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
4406 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
4407 // followed by an AND.
4408 // TODO: Look into removing this restriction by fixing backends to either
4409 // return false for isLoadExtLegal for i1 or have them select this pattern to
4410 // a single instruction.
4412 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
4413 // mask, since these are the only ands that will be removed by isel.
4414 if (ActiveBits <= 1 || !APIntOps::isMask(ActiveBits, DemandBits) ||
4415 WidestAndBits != DemandBits)
4418 LLVMContext &Ctx = Load->getType()->getContext();
4419 Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
4420 EVT TruncVT = TLI->getValueType(*DL, TruncTy);
4422 // Reject cases that won't be matched as extloads.
4423 if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
4424 !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
4427 IRBuilder<> Builder(Load->getNextNode());
4428 auto *NewAnd = dyn_cast<Instruction>(
4429 Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
4431 // Replace all uses of load with new and (except for the use of load in the
4433 Load->replaceAllUsesWith(NewAnd);
4434 NewAnd->setOperand(0, Load);
4436 // Remove any and instructions that are now redundant.
4437 for (auto *And : AndsToMaybeRemove)
4438 // Check that the and mask is the same as the one we decided to put on the
4440 if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
4441 And->replaceAllUsesWith(NewAnd);
4442 if (&*CurInstIterator == And)
4443 CurInstIterator = std::next(And->getIterator());
4444 And->eraseFromParent();
4452 /// Check if V (an operand of a select instruction) is an expensive instruction
4453 /// that is only used once.
4454 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
4455 auto *I = dyn_cast<Instruction>(V);
4456 // If it's safe to speculatively execute, then it should not have side
4457 // effects; therefore, it's safe to sink and possibly *not* execute.
4458 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
4459 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
4462 /// Returns true if a SelectInst should be turned into an explicit branch.
4463 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
4465 // FIXME: This should use the same heuristics as IfConversion to determine
4466 // whether a select is better represented as a branch. This requires that
4467 // branch probability metadata is preserved for the select, which is not the
4470 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
4472 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
4473 // comparison condition. If the compare has more than one use, there's
4474 // probably another cmov or setcc around, so it's not worth emitting a branch.
4475 if (!Cmp || !Cmp->hasOneUse())
4478 Value *CmpOp0 = Cmp->getOperand(0);
4479 Value *CmpOp1 = Cmp->getOperand(1);
4481 // Emit "cmov on compare with a memory operand" as a branch to avoid stalls
4482 // on a load from memory. But if the load is used more than once, do not
4483 // change the select to a branch because the load is probably needed
4484 // regardless of whether the branch is taken or not.
4485 if ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
4486 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()))
4489 // If either operand of the select is expensive and only needed on one side
4490 // of the select, we should form a branch.
4491 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
4492 sinkSelectOperand(TTI, SI->getFalseValue()))
4499 /// If we have a SelectInst that will likely profit from branch prediction,
4500 /// turn it into a branch.
4501 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
4502 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
4504 // Can we convert the 'select' to CF ?
4505 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
4508 TargetLowering::SelectSupportKind SelectKind;
4510 SelectKind = TargetLowering::VectorMaskSelect;
4511 else if (SI->getType()->isVectorTy())
4512 SelectKind = TargetLowering::ScalarCondVectorVal;
4514 SelectKind = TargetLowering::ScalarValSelect;
4516 // Do we have efficient codegen support for this kind of 'selects' ?
4517 if (TLI->isSelectSupported(SelectKind)) {
4518 // We have efficient codegen support for the select instruction.
4519 // Check if it is profitable to keep this 'select'.
4520 if (!TLI->isPredictableSelectExpensive() ||
4521 !isFormingBranchFromSelectProfitable(TTI, SI))
4527 // Transform a sequence like this:
4529 // %cmp = cmp uge i32 %a, %b
4530 // %sel = select i1 %cmp, i32 %c, i32 %d
4534 // %cmp = cmp uge i32 %a, %b
4535 // br i1 %cmp, label %select.true, label %select.false
4537 // br label %select.end
4539 // br label %select.end
4541 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
4543 // In addition, we may sink instructions that produce %c or %d from
4544 // the entry block into the destination(s) of the new branch.
4545 // If the true or false blocks do not contain a sunken instruction, that
4546 // block and its branch may be optimized away. In that case, one side of the
4547 // first branch will point directly to select.end, and the corresponding PHI
4548 // predecessor block will be the start block.
4550 // First, we split the block containing the select into 2 blocks.
4551 BasicBlock *StartBlock = SI->getParent();
4552 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
4553 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
4555 // Delete the unconditional branch that was just created by the split.
4556 StartBlock->getTerminator()->eraseFromParent();
4558 // These are the new basic blocks for the conditional branch.
4559 // At least one will become an actual new basic block.
4560 BasicBlock *TrueBlock = nullptr;
4561 BasicBlock *FalseBlock = nullptr;
4563 // Sink expensive instructions into the conditional blocks to avoid executing
4564 // them speculatively.
4565 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
4566 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
4567 EndBlock->getParent(), EndBlock);
4568 auto *TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
4569 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
4570 TrueInst->moveBefore(TrueBranch);
4572 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
4573 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
4574 EndBlock->getParent(), EndBlock);
4575 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
4576 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
4577 FalseInst->moveBefore(FalseBranch);
4580 // If there was nothing to sink, then arbitrarily choose the 'false' side
4581 // for a new input value to the PHI.
4582 if (TrueBlock == FalseBlock) {
4583 assert(TrueBlock == nullptr &&
4584 "Unexpected basic block transform while optimizing select");
4586 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
4587 EndBlock->getParent(), EndBlock);
4588 BranchInst::Create(EndBlock, FalseBlock);
4591 // Insert the real conditional branch based on the original condition.
4592 // If we did not create a new block for one of the 'true' or 'false' paths
4593 // of the condition, it means that side of the branch goes to the end block
4594 // directly and the path originates from the start block from the point of
4595 // view of the new PHI.
4596 if (TrueBlock == nullptr) {
4597 BranchInst::Create(EndBlock, FalseBlock, SI->getCondition(), SI);
4598 TrueBlock = StartBlock;
4599 } else if (FalseBlock == nullptr) {
4600 BranchInst::Create(TrueBlock, EndBlock, SI->getCondition(), SI);
4601 FalseBlock = StartBlock;
4603 BranchInst::Create(TrueBlock, FalseBlock, SI->getCondition(), SI);
4606 // The select itself is replaced with a PHI Node.
4607 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
4609 PN->addIncoming(SI->getTrueValue(), TrueBlock);
4610 PN->addIncoming(SI->getFalseValue(), FalseBlock);
4612 SI->replaceAllUsesWith(PN);
4613 SI->eraseFromParent();
4615 // Instruct OptimizeBlock to skip to the next block.
4616 CurInstIterator = StartBlock->end();
4617 ++NumSelectsExpanded;
4621 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
4622 SmallVector<int, 16> Mask(SVI->getShuffleMask());
4624 for (unsigned i = 0; i < Mask.size(); ++i) {
4625 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
4627 SplatElem = Mask[i];
4633 /// Some targets have expensive vector shifts if the lanes aren't all the same
4634 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
4635 /// it's often worth sinking a shufflevector splat down to its use so that
4636 /// codegen can spot all lanes are identical.
4637 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
4638 BasicBlock *DefBB = SVI->getParent();
4640 // Only do this xform if variable vector shifts are particularly expensive.
4641 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
4644 // We only expect better codegen by sinking a shuffle if we can recognise a
4646 if (!isBroadcastShuffle(SVI))
4649 // InsertedShuffles - Only insert a shuffle in each block once.
4650 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
4652 bool MadeChange = false;
4653 for (User *U : SVI->users()) {
4654 Instruction *UI = cast<Instruction>(U);
4656 // Figure out which BB this ext is used in.
4657 BasicBlock *UserBB = UI->getParent();
4658 if (UserBB == DefBB) continue;
4660 // For now only apply this when the splat is used by a shift instruction.
4661 if (!UI->isShift()) continue;
4663 // Everything checks out, sink the shuffle if the user's block doesn't
4664 // already have a copy.
4665 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
4667 if (!InsertedShuffle) {
4668 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4669 assert(InsertPt != UserBB->end());
4671 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4672 SVI->getOperand(2), "", &*InsertPt);
4675 UI->replaceUsesOfWith(SVI, InsertedShuffle);
4679 // If we removed all uses, nuke the shuffle.
4680 if (SVI->use_empty()) {
4681 SVI->eraseFromParent();
4688 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
4692 Value *Cond = SI->getCondition();
4693 Type *OldType = Cond->getType();
4694 LLVMContext &Context = Cond->getContext();
4695 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
4696 unsigned RegWidth = RegType.getSizeInBits();
4698 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
4701 // If the register width is greater than the type width, expand the condition
4702 // of the switch instruction and each case constant to the width of the
4703 // register. By widening the type of the switch condition, subsequent
4704 // comparisons (for case comparisons) will not need to be extended to the
4705 // preferred register width, so we will potentially eliminate N-1 extends,
4706 // where N is the number of cases in the switch.
4707 auto *NewType = Type::getIntNTy(Context, RegWidth);
4709 // Zero-extend the switch condition and case constants unless the switch
4710 // condition is a function argument that is already being sign-extended.
4711 // In that case, we can avoid an unnecessary mask/extension by sign-extending
4712 // everything instead.
4713 Instruction::CastOps ExtType = Instruction::ZExt;
4714 if (auto *Arg = dyn_cast<Argument>(Cond))
4715 if (Arg->hasSExtAttr())
4716 ExtType = Instruction::SExt;
4718 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
4719 ExtInst->insertBefore(SI);
4720 SI->setCondition(ExtInst);
4721 for (SwitchInst::CaseIt Case : SI->cases()) {
4722 APInt NarrowConst = Case.getCaseValue()->getValue();
4723 APInt WideConst = (ExtType == Instruction::ZExt) ?
4724 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
4725 Case.setValue(ConstantInt::get(Context, WideConst));
4732 /// \brief Helper class to promote a scalar operation to a vector one.
4733 /// This class is used to move downward extractelement transition.
4735 /// a = vector_op <2 x i32>
4736 /// b = extractelement <2 x i32> a, i32 0
4741 /// a = vector_op <2 x i32>
4742 /// c = vector_op a (equivalent to scalar_op on the related lane)
4743 /// * d = extractelement <2 x i32> c, i32 0
4745 /// Assuming both extractelement and store can be combine, we get rid of the
4747 class VectorPromoteHelper {
4748 /// DataLayout associated with the current module.
4749 const DataLayout &DL;
4751 /// Used to perform some checks on the legality of vector operations.
4752 const TargetLowering &TLI;
4754 /// Used to estimated the cost of the promoted chain.
4755 const TargetTransformInfo &TTI;
4757 /// The transition being moved downwards.
4758 Instruction *Transition;
4759 /// The sequence of instructions to be promoted.
4760 SmallVector<Instruction *, 4> InstsToBePromoted;
4761 /// Cost of combining a store and an extract.
4762 unsigned StoreExtractCombineCost;
4763 /// Instruction that will be combined with the transition.
4764 Instruction *CombineInst;
4766 /// \brief The instruction that represents the current end of the transition.
4767 /// Since we are faking the promotion until we reach the end of the chain
4768 /// of computation, we need a way to get the current end of the transition.
4769 Instruction *getEndOfTransition() const {
4770 if (InstsToBePromoted.empty())
4772 return InstsToBePromoted.back();
4775 /// \brief Return the index of the original value in the transition.
4776 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
4777 /// c, is at index 0.
4778 unsigned getTransitionOriginalValueIdx() const {
4779 assert(isa<ExtractElementInst>(Transition) &&
4780 "Other kind of transitions are not supported yet");
4784 /// \brief Return the index of the index in the transition.
4785 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
4787 unsigned getTransitionIdx() const {
4788 assert(isa<ExtractElementInst>(Transition) &&
4789 "Other kind of transitions are not supported yet");
4793 /// \brief Get the type of the transition.
4794 /// This is the type of the original value.
4795 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
4796 /// transition is <2 x i32>.
4797 Type *getTransitionType() const {
4798 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
4801 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
4802 /// I.e., we have the following sequence:
4803 /// Def = Transition <ty1> a to <ty2>
4804 /// b = ToBePromoted <ty2> Def, ...
4806 /// b = ToBePromoted <ty1> a, ...
4807 /// Def = Transition <ty1> ToBePromoted to <ty2>
4808 void promoteImpl(Instruction *ToBePromoted);
4810 /// \brief Check whether or not it is profitable to promote all the
4811 /// instructions enqueued to be promoted.
4812 bool isProfitableToPromote() {
4813 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
4814 unsigned Index = isa<ConstantInt>(ValIdx)
4815 ? cast<ConstantInt>(ValIdx)->getZExtValue()
4817 Type *PromotedType = getTransitionType();
4819 StoreInst *ST = cast<StoreInst>(CombineInst);
4820 unsigned AS = ST->getPointerAddressSpace();
4821 unsigned Align = ST->getAlignment();
4822 // Check if this store is supported.
4823 if (!TLI.allowsMisalignedMemoryAccesses(
4824 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
4826 // If this is not supported, there is no way we can combine
4827 // the extract with the store.
4831 // The scalar chain of computation has to pay for the transition
4832 // scalar to vector.
4833 // The vector chain has to account for the combining cost.
4834 uint64_t ScalarCost =
4835 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
4836 uint64_t VectorCost = StoreExtractCombineCost;
4837 for (const auto &Inst : InstsToBePromoted) {
4838 // Compute the cost.
4839 // By construction, all instructions being promoted are arithmetic ones.
4840 // Moreover, one argument is a constant that can be viewed as a splat
4842 Value *Arg0 = Inst->getOperand(0);
4843 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
4844 isa<ConstantFP>(Arg0);
4845 TargetTransformInfo::OperandValueKind Arg0OVK =
4846 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4847 : TargetTransformInfo::OK_AnyValue;
4848 TargetTransformInfo::OperandValueKind Arg1OVK =
4849 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4850 : TargetTransformInfo::OK_AnyValue;
4851 ScalarCost += TTI.getArithmeticInstrCost(
4852 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
4853 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
4856 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
4857 << ScalarCost << "\nVector: " << VectorCost << '\n');
4858 return ScalarCost > VectorCost;
4861 /// \brief Generate a constant vector with \p Val with the same
4862 /// number of elements as the transition.
4863 /// \p UseSplat defines whether or not \p Val should be replicated
4864 /// across the whole vector.
4865 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
4866 /// otherwise we generate a vector with as many undef as possible:
4867 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
4868 /// used at the index of the extract.
4869 Value *getConstantVector(Constant *Val, bool UseSplat) const {
4870 unsigned ExtractIdx = UINT_MAX;
4872 // If we cannot determine where the constant must be, we have to
4873 // use a splat constant.
4874 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
4875 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
4876 ExtractIdx = CstVal->getSExtValue();
4881 unsigned End = getTransitionType()->getVectorNumElements();
4883 return ConstantVector::getSplat(End, Val);
4885 SmallVector<Constant *, 4> ConstVec;
4886 UndefValue *UndefVal = UndefValue::get(Val->getType());
4887 for (unsigned Idx = 0; Idx != End; ++Idx) {
4888 if (Idx == ExtractIdx)
4889 ConstVec.push_back(Val);
4891 ConstVec.push_back(UndefVal);
4893 return ConstantVector::get(ConstVec);
4896 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
4897 /// in \p Use can trigger undefined behavior.
4898 static bool canCauseUndefinedBehavior(const Instruction *Use,
4899 unsigned OperandIdx) {
4900 // This is not safe to introduce undef when the operand is on
4901 // the right hand side of a division-like instruction.
4902 if (OperandIdx != 1)
4904 switch (Use->getOpcode()) {
4907 case Instruction::SDiv:
4908 case Instruction::UDiv:
4909 case Instruction::SRem:
4910 case Instruction::URem:
4912 case Instruction::FDiv:
4913 case Instruction::FRem:
4914 return !Use->hasNoNaNs();
4916 llvm_unreachable(nullptr);
4920 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
4921 const TargetTransformInfo &TTI, Instruction *Transition,
4922 unsigned CombineCost)
4923 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
4924 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4925 assert(Transition && "Do not know how to promote null");
4928 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4929 bool canPromote(const Instruction *ToBePromoted) const {
4930 // We could support CastInst too.
4931 return isa<BinaryOperator>(ToBePromoted);
4934 /// \brief Check if it is profitable to promote \p ToBePromoted
4935 /// by moving downward the transition through.
4936 bool shouldPromote(const Instruction *ToBePromoted) const {
4937 // Promote only if all the operands can be statically expanded.
4938 // Indeed, we do not want to introduce any new kind of transitions.
4939 for (const Use &U : ToBePromoted->operands()) {
4940 const Value *Val = U.get();
4941 if (Val == getEndOfTransition()) {
4942 // If the use is a division and the transition is on the rhs,
4943 // we cannot promote the operation, otherwise we may create a
4944 // division by zero.
4945 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4949 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4950 !isa<ConstantFP>(Val))
4953 // Check that the resulting operation is legal.
4954 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4957 return StressStoreExtract ||
4958 TLI.isOperationLegalOrCustom(
4959 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
4962 /// \brief Check whether or not \p Use can be combined
4963 /// with the transition.
4964 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4965 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4967 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4968 void enqueueForPromotion(Instruction *ToBePromoted) {
4969 InstsToBePromoted.push_back(ToBePromoted);
4972 /// \brief Set the instruction that will be combined with the transition.
4973 void recordCombineInstruction(Instruction *ToBeCombined) {
4974 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4975 CombineInst = ToBeCombined;
4978 /// \brief Promote all the instructions enqueued for promotion if it is
4980 /// \return True if the promotion happened, false otherwise.
4982 // Check if there is something to promote.
4983 // Right now, if we do not have anything to combine with,
4984 // we assume the promotion is not profitable.
4985 if (InstsToBePromoted.empty() || !CombineInst)
4989 if (!StressStoreExtract && !isProfitableToPromote())
4993 for (auto &ToBePromoted : InstsToBePromoted)
4994 promoteImpl(ToBePromoted);
4995 InstsToBePromoted.clear();
4999 } // End of anonymous namespace.
5001 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
5002 // At this point, we know that all the operands of ToBePromoted but Def
5003 // can be statically promoted.
5004 // For Def, we need to use its parameter in ToBePromoted:
5005 // b = ToBePromoted ty1 a
5006 // Def = Transition ty1 b to ty2
5007 // Move the transition down.
5008 // 1. Replace all uses of the promoted operation by the transition.
5009 // = ... b => = ... Def.
5010 assert(ToBePromoted->getType() == Transition->getType() &&
5011 "The type of the result of the transition does not match "
5013 ToBePromoted->replaceAllUsesWith(Transition);
5014 // 2. Update the type of the uses.
5015 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
5016 Type *TransitionTy = getTransitionType();
5017 ToBePromoted->mutateType(TransitionTy);
5018 // 3. Update all the operands of the promoted operation with promoted
5020 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
5021 for (Use &U : ToBePromoted->operands()) {
5022 Value *Val = U.get();
5023 Value *NewVal = nullptr;
5024 if (Val == Transition)
5025 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
5026 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
5027 isa<ConstantFP>(Val)) {
5028 // Use a splat constant if it is not safe to use undef.
5029 NewVal = getConstantVector(
5030 cast<Constant>(Val),
5031 isa<UndefValue>(Val) ||
5032 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
5034 llvm_unreachable("Did you modified shouldPromote and forgot to update "
5036 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
5038 Transition->removeFromParent();
5039 Transition->insertAfter(ToBePromoted);
5040 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
5043 /// Some targets can do store(extractelement) with one instruction.
5044 /// Try to push the extractelement towards the stores when the target
5045 /// has this feature and this is profitable.
5046 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
5047 unsigned CombineCost = UINT_MAX;
5048 if (DisableStoreExtract || !TLI ||
5049 (!StressStoreExtract &&
5050 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
5051 Inst->getOperand(1), CombineCost)))
5054 // At this point we know that Inst is a vector to scalar transition.
5055 // Try to move it down the def-use chain, until:
5056 // - We can combine the transition with its single use
5057 // => we got rid of the transition.
5058 // - We escape the current basic block
5059 // => we would need to check that we are moving it at a cheaper place and
5060 // we do not do that for now.
5061 BasicBlock *Parent = Inst->getParent();
5062 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
5063 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
5064 // If the transition has more than one use, assume this is not going to be
5066 while (Inst->hasOneUse()) {
5067 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
5068 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
5070 if (ToBePromoted->getParent() != Parent) {
5071 DEBUG(dbgs() << "Instruction to promote is in a different block ("
5072 << ToBePromoted->getParent()->getName()
5073 << ") than the transition (" << Parent->getName() << ").\n");
5077 if (VPH.canCombine(ToBePromoted)) {
5078 DEBUG(dbgs() << "Assume " << *Inst << '\n'
5079 << "will be combined with: " << *ToBePromoted << '\n');
5080 VPH.recordCombineInstruction(ToBePromoted);
5081 bool Changed = VPH.promote();
5082 NumStoreExtractExposed += Changed;
5086 DEBUG(dbgs() << "Try promoting.\n");
5087 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
5090 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
5092 VPH.enqueueForPromotion(ToBePromoted);
5093 Inst = ToBePromoted;
5098 bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
5099 // Bail out if we inserted the instruction to prevent optimizations from
5100 // stepping on each other's toes.
5101 if (InsertedInsts.count(I))
5104 if (PHINode *P = dyn_cast<PHINode>(I)) {
5105 // It is possible for very late stage optimizations (such as SimplifyCFG)
5106 // to introduce PHI nodes too late to be cleaned up. If we detect such a
5107 // trivial PHI, go ahead and zap it here.
5108 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
5109 P->replaceAllUsesWith(V);
5110 P->eraseFromParent();
5117 if (CastInst *CI = dyn_cast<CastInst>(I)) {
5118 // If the source of the cast is a constant, then this should have
5119 // already been constant folded. The only reason NOT to constant fold
5120 // it is if something (e.g. LSR) was careful to place the constant
5121 // evaluation in a block other than then one that uses it (e.g. to hoist
5122 // the address of globals out of a loop). If this is the case, we don't
5123 // want to forward-subst the cast.
5124 if (isa<Constant>(CI->getOperand(0)))
5127 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
5130 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
5131 /// Sink a zext or sext into its user blocks if the target type doesn't
5132 /// fit in one register
5134 TLI->getTypeAction(CI->getContext(),
5135 TLI->getValueType(*DL, CI->getType())) ==
5136 TargetLowering::TypeExpandInteger) {
5137 return SinkCast(CI);
5139 bool MadeChange = moveExtToFormExtLoad(I);
5140 return MadeChange | optimizeExtUses(I);
5146 if (CmpInst *CI = dyn_cast<CmpInst>(I))
5147 if (!TLI || !TLI->hasMultipleConditionRegisters())
5148 return OptimizeCmpExpression(CI);
5150 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
5151 stripInvariantGroupMetadata(*LI);
5153 bool Modified = optimizeLoadExt(LI);
5154 unsigned AS = LI->getPointerAddressSpace();
5155 Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
5161 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
5162 stripInvariantGroupMetadata(*SI);
5164 unsigned AS = SI->getPointerAddressSpace();
5165 return optimizeMemoryInst(I, SI->getOperand(1),
5166 SI->getOperand(0)->getType(), AS);
5171 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
5173 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
5174 BinOp->getOpcode() == Instruction::LShr)) {
5175 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
5176 if (TLI && CI && TLI->hasExtractBitsInsn())
5177 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
5182 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
5183 if (GEPI->hasAllZeroIndices()) {
5184 /// The GEP operand must be a pointer, so must its result -> BitCast
5185 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
5186 GEPI->getName(), GEPI);
5187 GEPI->replaceAllUsesWith(NC);
5188 GEPI->eraseFromParent();
5190 optimizeInst(NC, ModifiedDT);
5196 if (CallInst *CI = dyn_cast<CallInst>(I))
5197 return optimizeCallInst(CI, ModifiedDT);
5199 if (SelectInst *SI = dyn_cast<SelectInst>(I))
5200 return optimizeSelectInst(SI);
5202 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
5203 return optimizeShuffleVectorInst(SVI);
5205 if (auto *Switch = dyn_cast<SwitchInst>(I))
5206 return optimizeSwitchInst(Switch);
5208 if (isa<ExtractElementInst>(I))
5209 return optimizeExtractElementInst(I);
5214 // In this pass we look for GEP and cast instructions that are used
5215 // across basic blocks and rewrite them to improve basic-block-at-a-time
5217 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
5219 bool MadeChange = false;
5221 CurInstIterator = BB.begin();
5222 while (CurInstIterator != BB.end()) {
5223 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
5227 MadeChange |= dupRetToEnableTailCallOpts(&BB);
5232 // llvm.dbg.value is far away from the value then iSel may not be able
5233 // handle it properly. iSel will drop llvm.dbg.value if it can not
5234 // find a node corresponding to the value.
5235 bool CodeGenPrepare::placeDbgValues(Function &F) {
5236 bool MadeChange = false;
5237 for (BasicBlock &BB : F) {
5238 Instruction *PrevNonDbgInst = nullptr;
5239 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
5240 Instruction *Insn = &*BI++;
5241 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
5242 // Leave dbg.values that refer to an alloca alone. These
5243 // instrinsics describe the address of a variable (= the alloca)
5244 // being taken. They should not be moved next to the alloca
5245 // (and to the beginning of the scope), but rather stay close to
5246 // where said address is used.
5247 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
5248 PrevNonDbgInst = Insn;
5252 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
5253 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
5254 // If VI is a phi in a block with an EHPad terminator, we can't insert
5256 if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
5258 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
5259 DVI->removeFromParent();
5260 if (isa<PHINode>(VI))
5261 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
5263 DVI->insertAfter(VI);
5272 // If there is a sequence that branches based on comparing a single bit
5273 // against zero that can be combined into a single instruction, and the
5274 // target supports folding these into a single instruction, sink the
5275 // mask and compare into the branch uses. Do this before OptimizeBlock ->
5276 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
5278 bool CodeGenPrepare::sinkAndCmp(Function &F) {
5279 if (!EnableAndCmpSinking)
5281 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
5283 bool MadeChange = false;
5284 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
5285 BasicBlock *BB = &*I++;
5287 // Does this BB end with the following?
5288 // %andVal = and %val, #single-bit-set
5289 // %icmpVal = icmp %andResult, 0
5290 // br i1 %cmpVal label %dest1, label %dest2"
5291 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
5292 if (!Brcc || !Brcc->isConditional())
5294 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
5295 if (!Cmp || Cmp->getParent() != BB)
5297 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
5298 if (!Zero || !Zero->isZero())
5300 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
5301 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
5303 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
5304 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
5306 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
5308 // Push the "and; icmp" for any users that are conditional branches.
5309 // Since there can only be one branch use per BB, we don't need to keep
5310 // track of which BBs we insert into.
5311 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
5315 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
5317 if (!BrccUser || !BrccUser->isConditional())
5319 BasicBlock *UserBB = BrccUser->getParent();
5320 if (UserBB == BB) continue;
5321 DEBUG(dbgs() << "found Brcc use\n");
5323 // Sink the "and; icmp" to use.
5325 BinaryOperator *NewAnd =
5326 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
5329 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
5333 DEBUG(BrccUser->getParent()->dump());
5339 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
5340 /// success, or returns false if no or invalid metadata was found.
5341 static bool extractBranchMetadata(BranchInst *BI,
5342 uint64_t &ProbTrue, uint64_t &ProbFalse) {
5343 assert(BI->isConditional() &&
5344 "Looking for probabilities on unconditional branch?");
5345 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
5346 if (!ProfileData || ProfileData->getNumOperands() != 3)
5349 const auto *CITrue =
5350 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
5351 const auto *CIFalse =
5352 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
5353 if (!CITrue || !CIFalse)
5356 ProbTrue = CITrue->getValue().getZExtValue();
5357 ProbFalse = CIFalse->getValue().getZExtValue();
5362 /// \brief Scale down both weights to fit into uint32_t.
5363 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
5364 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
5365 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
5366 NewTrue = NewTrue / Scale;
5367 NewFalse = NewFalse / Scale;
5370 /// \brief Some targets prefer to split a conditional branch like:
5372 /// %0 = icmp ne i32 %a, 0
5373 /// %1 = icmp ne i32 %b, 0
5374 /// %or.cond = or i1 %0, %1
5375 /// br i1 %or.cond, label %TrueBB, label %FalseBB
5377 /// into multiple branch instructions like:
5380 /// %0 = icmp ne i32 %a, 0
5381 /// br i1 %0, label %TrueBB, label %bb2
5383 /// %1 = icmp ne i32 %b, 0
5384 /// br i1 %1, label %TrueBB, label %FalseBB
5386 /// This usually allows instruction selection to do even further optimizations
5387 /// and combine the compare with the branch instruction. Currently this is
5388 /// applied for targets which have "cheap" jump instructions.
5390 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
5392 bool CodeGenPrepare::splitBranchCondition(Function &F) {
5393 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
5396 bool MadeChange = false;
5397 for (auto &BB : F) {
5398 // Does this BB end with the following?
5399 // %cond1 = icmp|fcmp|binary instruction ...
5400 // %cond2 = icmp|fcmp|binary instruction ...
5401 // %cond.or = or|and i1 %cond1, cond2
5402 // br i1 %cond.or label %dest1, label %dest2"
5403 BinaryOperator *LogicOp;
5404 BasicBlock *TBB, *FBB;
5405 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
5408 auto *Br1 = cast<BranchInst>(BB.getTerminator());
5409 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
5413 Value *Cond1, *Cond2;
5414 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
5415 m_OneUse(m_Value(Cond2)))))
5416 Opc = Instruction::And;
5417 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
5418 m_OneUse(m_Value(Cond2)))))
5419 Opc = Instruction::Or;
5423 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
5424 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
5427 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
5430 auto *InsertBefore = std::next(Function::iterator(BB))
5431 .getNodePtrUnchecked();
5432 auto TmpBB = BasicBlock::Create(BB.getContext(),
5433 BB.getName() + ".cond.split",
5434 BB.getParent(), InsertBefore);
5436 // Update original basic block by using the first condition directly by the
5437 // branch instruction and removing the no longer needed and/or instruction.
5438 Br1->setCondition(Cond1);
5439 LogicOp->eraseFromParent();
5441 // Depending on the conditon we have to either replace the true or the false
5442 // successor of the original branch instruction.
5443 if (Opc == Instruction::And)
5444 Br1->setSuccessor(0, TmpBB);
5446 Br1->setSuccessor(1, TmpBB);
5448 // Fill in the new basic block.
5449 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
5450 if (auto *I = dyn_cast<Instruction>(Cond2)) {
5451 I->removeFromParent();
5452 I->insertBefore(Br2);
5455 // Update PHI nodes in both successors. The original BB needs to be
5456 // replaced in one succesor's PHI nodes, because the branch comes now from
5457 // the newly generated BB (NewBB). In the other successor we need to add one
5458 // incoming edge to the PHI nodes, because both branch instructions target
5459 // now the same successor. Depending on the original branch condition
5460 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
5461 // we perfrom the correct update for the PHI nodes.
5462 // This doesn't change the successor order of the just created branch
5463 // instruction (or any other instruction).
5464 if (Opc == Instruction::Or)
5465 std::swap(TBB, FBB);
5467 // Replace the old BB with the new BB.
5468 for (auto &I : *TBB) {
5469 PHINode *PN = dyn_cast<PHINode>(&I);
5473 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
5474 PN->setIncomingBlock(i, TmpBB);
5477 // Add another incoming edge form the new BB.
5478 for (auto &I : *FBB) {
5479 PHINode *PN = dyn_cast<PHINode>(&I);
5482 auto *Val = PN->getIncomingValueForBlock(&BB);
5483 PN->addIncoming(Val, TmpBB);
5486 // Update the branch weights (from SelectionDAGBuilder::
5487 // FindMergedConditions).
5488 if (Opc == Instruction::Or) {
5489 // Codegen X | Y as:
5498 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
5499 // The requirement is that
5500 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
5501 // = TrueProb for orignal BB.
5502 // Assuming the orignal weights are A and B, one choice is to set BB1's
5503 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
5505 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
5506 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
5507 // TmpBB, but the math is more complicated.
5508 uint64_t TrueWeight, FalseWeight;
5509 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
5510 uint64_t NewTrueWeight = TrueWeight;
5511 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
5512 scaleWeights(NewTrueWeight, NewFalseWeight);
5513 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
5514 .createBranchWeights(TrueWeight, FalseWeight));
5516 NewTrueWeight = TrueWeight;
5517 NewFalseWeight = 2 * FalseWeight;
5518 scaleWeights(NewTrueWeight, NewFalseWeight);
5519 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
5520 .createBranchWeights(TrueWeight, FalseWeight));
5523 // Codegen X & Y as:
5531 // This requires creation of TmpBB after CurBB.
5533 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
5534 // The requirement is that
5535 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
5536 // = FalseProb for orignal BB.
5537 // Assuming the orignal weights are A and B, one choice is to set BB1's
5538 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
5540 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
5541 uint64_t TrueWeight, FalseWeight;
5542 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
5543 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
5544 uint64_t NewFalseWeight = FalseWeight;
5545 scaleWeights(NewTrueWeight, NewFalseWeight);
5546 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
5547 .createBranchWeights(TrueWeight, FalseWeight));
5549 NewTrueWeight = 2 * TrueWeight;
5550 NewFalseWeight = FalseWeight;
5551 scaleWeights(NewTrueWeight, NewFalseWeight);
5552 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
5553 .createBranchWeights(TrueWeight, FalseWeight));
5557 // Note: No point in getting fancy here, since the DT info is never
5558 // available to CodeGenPrepare.
5563 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
5569 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
5570 if (auto *InvariantMD = I.getMetadata(LLVMContext::MD_invariant_group))
5571 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());