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(NumRetsDup, "Number of return instructions duplicated");
68 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
69 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
70 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
71 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
73 static cl::opt<bool> DisableBranchOpts(
74 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
75 cl::desc("Disable branch optimizations in CodeGenPrepare"));
78 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
79 cl::desc("Disable GC optimizations in CodeGenPrepare"));
81 static cl::opt<bool> DisableSelectToBranch(
82 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
83 cl::desc("Disable select to branch conversion."));
85 static cl::opt<bool> AddrSinkUsingGEPs(
86 "addr-sink-using-gep", cl::Hidden, cl::init(false),
87 cl::desc("Address sinking in CGP using GEPs."));
89 static cl::opt<bool> EnableAndCmpSinking(
90 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
91 cl::desc("Enable sinkinig and/cmp into branches."));
93 static cl::opt<bool> DisableStoreExtract(
94 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
95 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
97 static cl::opt<bool> StressStoreExtract(
98 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
99 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
101 static cl::opt<bool> DisableExtLdPromotion(
102 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
103 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
106 static cl::opt<bool> StressExtLdPromotion(
107 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
108 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
109 "optimization in CodeGenPrepare"));
112 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
113 typedef PointerIntPair<Type *, 1, bool> TypeIsSExt;
114 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
115 class TypePromotionTransaction;
117 class CodeGenPrepare : public FunctionPass {
118 const TargetMachine *TM;
119 const TargetLowering *TLI;
120 const TargetTransformInfo *TTI;
121 const TargetLibraryInfo *TLInfo;
123 /// As we scan instructions optimizing them, this is the next instruction
124 /// to optimize. Transforms that can invalidate this should update it.
125 BasicBlock::iterator CurInstIterator;
127 /// Keeps track of non-local addresses that have been sunk into a block.
128 /// This allows us to avoid inserting duplicate code for blocks with
129 /// multiple load/stores of the same address.
130 ValueMap<Value*, Value*> SunkAddrs;
132 /// Keeps track of all instructions inserted for the current function.
133 SetOfInstrs InsertedInsts;
134 /// Keeps track of the type of the related instruction before their
135 /// promotion for the current function.
136 InstrToOrigTy PromotedInsts;
138 /// True if CFG is modified in any way.
141 /// True if optimizing for size.
144 /// DataLayout for the Function being processed.
145 const DataLayout *DL;
148 static char ID; // Pass identification, replacement for typeid
149 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
150 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr), DL(nullptr) {
151 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
153 bool runOnFunction(Function &F) override;
155 const char *getPassName() const override { return "CodeGen Prepare"; }
157 void getAnalysisUsage(AnalysisUsage &AU) const override {
158 AU.addPreserved<DominatorTreeWrapperPass>();
159 AU.addRequired<TargetLibraryInfoWrapperPass>();
160 AU.addRequired<TargetTransformInfoWrapperPass>();
164 bool eliminateFallThrough(Function &F);
165 bool eliminateMostlyEmptyBlocks(Function &F);
166 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
167 void eliminateMostlyEmptyBlock(BasicBlock *BB);
168 bool optimizeBlock(BasicBlock &BB, bool& ModifiedDT);
169 bool optimizeInst(Instruction *I, bool& ModifiedDT);
170 bool optimizeMemoryInst(Instruction *I, Value *Addr,
171 Type *AccessTy, unsigned AS);
172 bool optimizeInlineAsmInst(CallInst *CS);
173 bool optimizeCallInst(CallInst *CI, bool& ModifiedDT);
174 bool moveExtToFormExtLoad(Instruction *&I);
175 bool optimizeExtUses(Instruction *I);
176 bool optimizeSelectInst(SelectInst *SI);
177 bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
178 bool optimizeSwitchInst(SwitchInst *CI);
179 bool optimizeExtractElementInst(Instruction *Inst);
180 bool dupRetToEnableTailCallOpts(BasicBlock *BB);
181 bool placeDbgValues(Function &F);
182 bool sinkAndCmp(Function &F);
183 bool extLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
185 const SmallVectorImpl<Instruction *> &Exts,
186 unsigned CreatedInstCost);
187 bool splitBranchCondition(Function &F);
188 bool simplifyOffsetableRelocate(Instruction &I);
189 void stripInvariantGroupMetadata(Instruction &I);
193 char CodeGenPrepare::ID = 0;
194 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
195 "Optimize for code generation", false, false)
197 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
198 return new CodeGenPrepare(TM);
201 bool CodeGenPrepare::runOnFunction(Function &F) {
202 if (skipOptnoneFunction(F))
205 DL = &F.getParent()->getDataLayout();
207 bool EverMadeChange = false;
208 // Clear per function information.
209 InsertedInsts.clear();
210 PromotedInsts.clear();
214 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
215 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
216 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
217 OptSize = F.optForSize();
219 /// This optimization identifies DIV instructions that can be
220 /// profitably bypassed and carried out with a shorter, faster divide.
221 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
222 const DenseMap<unsigned int, unsigned int> &BypassWidths =
223 TLI->getBypassSlowDivWidths();
224 for (Function::iterator I = F.begin(); I != F.end(); I++)
225 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
228 // Eliminate blocks that contain only PHI nodes and an
229 // unconditional branch.
230 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
232 // llvm.dbg.value is far away from the value then iSel may not be able
233 // handle it properly. iSel will drop llvm.dbg.value if it can not
234 // find a node corresponding to the value.
235 EverMadeChange |= placeDbgValues(F);
237 // If there is a mask, compare against zero, and branch that can be combined
238 // into a single target instruction, push the mask and compare into branch
239 // users. Do this before OptimizeBlock -> OptimizeInst ->
240 // OptimizeCmpExpression, which perturbs the pattern being searched for.
241 if (!DisableBranchOpts) {
242 EverMadeChange |= sinkAndCmp(F);
243 EverMadeChange |= splitBranchCondition(F);
246 bool MadeChange = true;
249 for (Function::iterator I = F.begin(); I != F.end(); ) {
250 BasicBlock *BB = &*I++;
251 bool ModifiedDTOnIteration = false;
252 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
254 // Restart BB iteration if the dominator tree of the Function was changed
255 if (ModifiedDTOnIteration)
258 EverMadeChange |= MadeChange;
263 if (!DisableBranchOpts) {
265 SmallPtrSet<BasicBlock*, 8> WorkList;
266 for (BasicBlock &BB : F) {
267 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
268 MadeChange |= ConstantFoldTerminator(&BB, true);
269 if (!MadeChange) continue;
271 for (SmallVectorImpl<BasicBlock*>::iterator
272 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
273 if (pred_begin(*II) == pred_end(*II))
274 WorkList.insert(*II);
277 // Delete the dead blocks and any of their dead successors.
278 MadeChange |= !WorkList.empty();
279 while (!WorkList.empty()) {
280 BasicBlock *BB = *WorkList.begin();
282 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
286 for (SmallVectorImpl<BasicBlock*>::iterator
287 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
288 if (pred_begin(*II) == pred_end(*II))
289 WorkList.insert(*II);
292 // Merge pairs of basic blocks with unconditional branches, connected by
294 if (EverMadeChange || MadeChange)
295 MadeChange |= eliminateFallThrough(F);
297 EverMadeChange |= MadeChange;
300 if (!DisableGCOpts) {
301 SmallVector<Instruction *, 2> Statepoints;
302 for (BasicBlock &BB : F)
303 for (Instruction &I : BB)
305 Statepoints.push_back(&I);
306 for (auto &I : Statepoints)
307 EverMadeChange |= simplifyOffsetableRelocate(*I);
310 return EverMadeChange;
313 /// Merge basic blocks which are connected by a single edge, where one of the
314 /// basic blocks has a single successor pointing to the other basic block,
315 /// which has a single predecessor.
316 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
317 bool Changed = false;
318 // Scan all of the blocks in the function, except for the entry block.
319 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
320 BasicBlock *BB = &*I++;
321 // If the destination block has a single pred, then this is a trivial
322 // edge, just collapse it.
323 BasicBlock *SinglePred = BB->getSinglePredecessor();
325 // Don't merge if BB's address is taken.
326 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
328 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
329 if (Term && !Term->isConditional()) {
331 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
332 // Remember if SinglePred was the entry block of the function.
333 // If so, we will need to move BB back to the entry position.
334 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
335 MergeBasicBlockIntoOnlyPred(BB, nullptr);
337 if (isEntry && BB != &BB->getParent()->getEntryBlock())
338 BB->moveBefore(&BB->getParent()->getEntryBlock());
340 // We have erased a block. Update the iterator.
341 I = BB->getIterator();
347 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
348 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
349 /// edges in ways that are non-optimal for isel. Start by eliminating these
350 /// blocks so we can split them the way we want them.
351 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
352 bool MadeChange = false;
353 // Note that this intentionally skips the entry block.
354 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
355 BasicBlock *BB = &*I++;
357 // If this block doesn't end with an uncond branch, ignore it.
358 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
359 if (!BI || !BI->isUnconditional())
362 // If the instruction before the branch (skipping debug info) isn't a phi
363 // node, then other stuff is happening here.
364 BasicBlock::iterator BBI = BI->getIterator();
365 if (BBI != BB->begin()) {
367 while (isa<DbgInfoIntrinsic>(BBI)) {
368 if (BBI == BB->begin())
372 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
376 // Do not break infinite loops.
377 BasicBlock *DestBB = BI->getSuccessor(0);
381 if (!canMergeBlocks(BB, DestBB))
384 eliminateMostlyEmptyBlock(BB);
390 /// Return true if we can merge BB into DestBB if there is a single
391 /// unconditional branch between them, and BB contains no other non-phi
393 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
394 const BasicBlock *DestBB) const {
395 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
396 // the successor. If there are more complex condition (e.g. preheaders),
397 // don't mess around with them.
398 BasicBlock::const_iterator BBI = BB->begin();
399 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
400 for (const User *U : PN->users()) {
401 const Instruction *UI = cast<Instruction>(U);
402 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
404 // If User is inside DestBB block and it is a PHINode then check
405 // incoming value. If incoming value is not from BB then this is
406 // a complex condition (e.g. preheaders) we want to avoid here.
407 if (UI->getParent() == DestBB) {
408 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
409 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
410 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
411 if (Insn && Insn->getParent() == BB &&
412 Insn->getParent() != UPN->getIncomingBlock(I))
419 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
420 // and DestBB may have conflicting incoming values for the block. If so, we
421 // can't merge the block.
422 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
423 if (!DestBBPN) return true; // no conflict.
425 // Collect the preds of BB.
426 SmallPtrSet<const BasicBlock*, 16> BBPreds;
427 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
428 // It is faster to get preds from a PHI than with pred_iterator.
429 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
430 BBPreds.insert(BBPN->getIncomingBlock(i));
432 BBPreds.insert(pred_begin(BB), pred_end(BB));
435 // Walk the preds of DestBB.
436 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
437 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
438 if (BBPreds.count(Pred)) { // Common predecessor?
439 BBI = DestBB->begin();
440 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
441 const Value *V1 = PN->getIncomingValueForBlock(Pred);
442 const Value *V2 = PN->getIncomingValueForBlock(BB);
444 // If V2 is a phi node in BB, look up what the mapped value will be.
445 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
446 if (V2PN->getParent() == BB)
447 V2 = V2PN->getIncomingValueForBlock(Pred);
449 // If there is a conflict, bail out.
450 if (V1 != V2) return false;
459 /// Eliminate a basic block that has only phi's and an unconditional branch in
461 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
462 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
463 BasicBlock *DestBB = BI->getSuccessor(0);
465 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
467 // If the destination block has a single pred, then this is a trivial edge,
469 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
470 if (SinglePred != DestBB) {
471 // Remember if SinglePred was the entry block of the function. If so, we
472 // will need to move BB back to the entry position.
473 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
474 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
476 if (isEntry && BB != &BB->getParent()->getEntryBlock())
477 BB->moveBefore(&BB->getParent()->getEntryBlock());
479 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
484 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
485 // to handle the new incoming edges it is about to have.
487 for (BasicBlock::iterator BBI = DestBB->begin();
488 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
489 // Remove the incoming value for BB, and remember it.
490 Value *InVal = PN->removeIncomingValue(BB, false);
492 // Two options: either the InVal is a phi node defined in BB or it is some
493 // value that dominates BB.
494 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
495 if (InValPhi && InValPhi->getParent() == BB) {
496 // Add all of the input values of the input PHI as inputs of this phi.
497 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
498 PN->addIncoming(InValPhi->getIncomingValue(i),
499 InValPhi->getIncomingBlock(i));
501 // Otherwise, add one instance of the dominating value for each edge that
502 // we will be adding.
503 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
504 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
505 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
507 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
508 PN->addIncoming(InVal, *PI);
513 // The PHIs are now updated, change everything that refers to BB to use
514 // DestBB and remove BB.
515 BB->replaceAllUsesWith(DestBB);
516 BB->eraseFromParent();
519 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
522 // Computes a map of base pointer relocation instructions to corresponding
523 // derived pointer relocation instructions given a vector of all relocate calls
524 static void computeBaseDerivedRelocateMap(
525 const SmallVectorImpl<User *> &AllRelocateCalls,
526 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
528 // Collect information in two maps: one primarily for locating the base object
529 // while filling the second map; the second map is the final structure holding
530 // a mapping between Base and corresponding Derived relocate calls
531 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
532 for (auto &U : AllRelocateCalls) {
533 GCRelocateOperands ThisRelocate(U);
534 IntrinsicInst *I = cast<IntrinsicInst>(U);
535 auto K = std::make_pair(ThisRelocate.getBasePtrIndex(),
536 ThisRelocate.getDerivedPtrIndex());
537 RelocateIdxMap.insert(std::make_pair(K, I));
539 for (auto &Item : RelocateIdxMap) {
540 std::pair<unsigned, unsigned> Key = Item.first;
541 if (Key.first == Key.second)
542 // Base relocation: nothing to insert
545 IntrinsicInst *I = Item.second;
546 auto BaseKey = std::make_pair(Key.first, Key.first);
548 // We're iterating over RelocateIdxMap so we cannot modify it.
549 auto MaybeBase = RelocateIdxMap.find(BaseKey);
550 if (MaybeBase == RelocateIdxMap.end())
551 // TODO: We might want to insert a new base object relocate and gep off
552 // that, if there are enough derived object relocates.
555 RelocateInstMap[MaybeBase->second].push_back(I);
559 // Accepts a GEP and extracts the operands into a vector provided they're all
560 // small integer constants
561 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
562 SmallVectorImpl<Value *> &OffsetV) {
563 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
564 // Only accept small constant integer operands
565 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
566 if (!Op || Op->getZExtValue() > 20)
570 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
571 OffsetV.push_back(GEP->getOperand(i));
575 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
576 // replace, computes a replacement, and affects it.
578 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
579 const SmallVectorImpl<IntrinsicInst *> &Targets) {
580 bool MadeChange = false;
581 for (auto &ToReplace : Targets) {
582 GCRelocateOperands MasterRelocate(RelocatedBase);
583 GCRelocateOperands ThisRelocate(ToReplace);
585 assert(ThisRelocate.getBasePtrIndex() == MasterRelocate.getBasePtrIndex() &&
586 "Not relocating a derived object of the original base object");
587 if (ThisRelocate.getBasePtrIndex() == ThisRelocate.getDerivedPtrIndex()) {
588 // A duplicate relocate call. TODO: coalesce duplicates.
592 if (RelocatedBase->getParent() != ToReplace->getParent()) {
593 // Base and derived relocates are in different basic blocks.
594 // In this case transform is only valid when base dominates derived
595 // relocate. However it would be too expensive to check dominance
596 // for each such relocate, so we skip the whole transformation.
600 Value *Base = ThisRelocate.getBasePtr();
601 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.getDerivedPtr());
602 if (!Derived || Derived->getPointerOperand() != Base)
605 SmallVector<Value *, 2> OffsetV;
606 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
609 // Create a Builder and replace the target callsite with a gep
610 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
612 // Insert after RelocatedBase
613 IRBuilder<> Builder(RelocatedBase->getNextNode());
614 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
616 // If gc_relocate does not match the actual type, cast it to the right type.
617 // In theory, there must be a bitcast after gc_relocate if the type does not
618 // match, and we should reuse it to get the derived pointer. But it could be
622 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
627 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
631 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
632 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
634 // In this case, we can not find the bitcast any more. So we insert a new bitcast
635 // no matter there is already one or not. In this way, we can handle all cases, and
636 // the extra bitcast should be optimized away in later passes.
637 Instruction *ActualRelocatedBase = RelocatedBase;
638 if (RelocatedBase->getType() != Base->getType()) {
639 ActualRelocatedBase =
640 cast<Instruction>(Builder.CreateBitCast(RelocatedBase, Base->getType()));
642 Value *Replacement = Builder.CreateGEP(
643 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
644 Instruction *ReplacementInst = cast<Instruction>(Replacement);
645 Replacement->takeName(ToReplace);
646 // If the newly generated derived pointer's type does not match the original derived
647 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
648 Instruction *ActualReplacement = ReplacementInst;
649 if (ReplacementInst->getType() != ToReplace->getType()) {
651 cast<Instruction>(Builder.CreateBitCast(ReplacementInst, ToReplace->getType()));
653 ToReplace->replaceAllUsesWith(ActualReplacement);
654 ToReplace->eraseFromParent();
664 // %ptr = gep %base + 15
665 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
666 // %base' = relocate(%tok, i32 4, i32 4)
667 // %ptr' = relocate(%tok, i32 4, i32 5)
673 // %ptr = gep %base + 15
674 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
675 // %base' = gc.relocate(%tok, i32 4, i32 4)
676 // %ptr' = gep %base' + 15
678 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
679 bool MadeChange = false;
680 SmallVector<User *, 2> AllRelocateCalls;
682 for (auto *U : I.users())
683 if (isGCRelocate(dyn_cast<Instruction>(U)))
684 // Collect all the relocate calls associated with a statepoint
685 AllRelocateCalls.push_back(U);
687 // We need atleast one base pointer relocation + one derived pointer
688 // relocation to mangle
689 if (AllRelocateCalls.size() < 2)
692 // RelocateInstMap is a mapping from the base relocate instruction to the
693 // corresponding derived relocate instructions
694 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
695 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
696 if (RelocateInstMap.empty())
699 for (auto &Item : RelocateInstMap)
700 // Item.first is the RelocatedBase to offset against
701 // Item.second is the vector of Targets to replace
702 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
706 /// SinkCast - Sink the specified cast instruction into its user blocks
707 static bool SinkCast(CastInst *CI) {
708 BasicBlock *DefBB = CI->getParent();
710 /// InsertedCasts - Only insert a cast in each block once.
711 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
713 bool MadeChange = false;
714 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
716 Use &TheUse = UI.getUse();
717 Instruction *User = cast<Instruction>(*UI);
719 // Figure out which BB this cast is used in. For PHI's this is the
720 // appropriate predecessor block.
721 BasicBlock *UserBB = User->getParent();
722 if (PHINode *PN = dyn_cast<PHINode>(User)) {
723 UserBB = PN->getIncomingBlock(TheUse);
726 // Preincrement use iterator so we don't invalidate it.
729 // If this user is in the same block as the cast, don't change the cast.
730 if (UserBB == DefBB) continue;
732 // If we have already inserted a cast into this block, use it.
733 CastInst *&InsertedCast = InsertedCasts[UserBB];
736 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
737 assert(InsertPt != UserBB->end());
738 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
739 CI->getType(), "", &*InsertPt);
742 // Replace a use of the cast with a use of the new cast.
743 TheUse = InsertedCast;
748 // If we removed all uses, nuke the cast.
749 if (CI->use_empty()) {
750 CI->eraseFromParent();
757 /// If the specified cast instruction is a noop copy (e.g. it's casting from
758 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
759 /// reduce the number of virtual registers that must be created and coalesced.
761 /// Return true if any changes are made.
763 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
764 const DataLayout &DL) {
765 // If this is a noop copy,
766 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
767 EVT DstVT = TLI.getValueType(DL, CI->getType());
769 // This is an fp<->int conversion?
770 if (SrcVT.isInteger() != DstVT.isInteger())
773 // If this is an extension, it will be a zero or sign extension, which
775 if (SrcVT.bitsLT(DstVT)) return false;
777 // If these values will be promoted, find out what they will be promoted
778 // to. This helps us consider truncates on PPC as noop copies when they
780 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
781 TargetLowering::TypePromoteInteger)
782 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
783 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
784 TargetLowering::TypePromoteInteger)
785 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
787 // If, after promotion, these are the same types, this is a noop copy.
794 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
797 /// Return true if any changes were made.
798 static bool CombineUAddWithOverflow(CmpInst *CI) {
802 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
805 Type *Ty = AddI->getType();
806 if (!isa<IntegerType>(Ty))
809 // We don't want to move around uses of condition values this late, so we we
810 // check if it is legal to create the call to the intrinsic in the basic
811 // block containing the icmp:
813 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
817 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
819 if (AddI->hasOneUse())
820 assert(*AddI->user_begin() == CI && "expected!");
823 Module *M = CI->getParent()->getParent()->getParent();
824 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
826 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
828 auto *UAddWithOverflow =
829 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
830 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
832 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
834 CI->replaceAllUsesWith(Overflow);
835 AddI->replaceAllUsesWith(UAdd);
836 CI->eraseFromParent();
837 AddI->eraseFromParent();
841 /// Sink the given CmpInst into user blocks to reduce the number of virtual
842 /// registers that must be created and coalesced. This is a clear win except on
843 /// targets with multiple condition code registers (PowerPC), where it might
844 /// lose; some adjustment may be wanted there.
846 /// Return true if any changes are made.
847 static bool SinkCmpExpression(CmpInst *CI) {
848 BasicBlock *DefBB = CI->getParent();
850 /// Only insert a cmp in each block once.
851 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
853 bool MadeChange = false;
854 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
856 Use &TheUse = UI.getUse();
857 Instruction *User = cast<Instruction>(*UI);
859 // Preincrement use iterator so we don't invalidate it.
862 // Don't bother for PHI nodes.
863 if (isa<PHINode>(User))
866 // Figure out which BB this cmp is used in.
867 BasicBlock *UserBB = User->getParent();
869 // If this user is in the same block as the cmp, don't change the cmp.
870 if (UserBB == DefBB) continue;
872 // If we have already inserted a cmp into this block, use it.
873 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
876 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
877 assert(InsertPt != UserBB->end());
879 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
880 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
883 // Replace a use of the cmp with a use of the new cmp.
884 TheUse = InsertedCmp;
889 // If we removed all uses, nuke the cmp.
890 if (CI->use_empty()) {
891 CI->eraseFromParent();
898 static bool OptimizeCmpExpression(CmpInst *CI) {
899 if (SinkCmpExpression(CI))
902 if (CombineUAddWithOverflow(CI))
908 /// Check if the candidates could be combined with a shift instruction, which
910 /// 1. Truncate instruction
911 /// 2. And instruction and the imm is a mask of the low bits:
912 /// imm & (imm+1) == 0
913 static bool isExtractBitsCandidateUse(Instruction *User) {
914 if (!isa<TruncInst>(User)) {
915 if (User->getOpcode() != Instruction::And ||
916 !isa<ConstantInt>(User->getOperand(1)))
919 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
921 if ((Cimm & (Cimm + 1)).getBoolValue())
927 /// Sink both shift and truncate instruction to the use of truncate's BB.
929 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
930 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
931 const TargetLowering &TLI, const DataLayout &DL) {
932 BasicBlock *UserBB = User->getParent();
933 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
934 TruncInst *TruncI = dyn_cast<TruncInst>(User);
935 bool MadeChange = false;
937 for (Value::user_iterator TruncUI = TruncI->user_begin(),
938 TruncE = TruncI->user_end();
939 TruncUI != TruncE;) {
941 Use &TruncTheUse = TruncUI.getUse();
942 Instruction *TruncUser = cast<Instruction>(*TruncUI);
943 // Preincrement use iterator so we don't invalidate it.
947 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
951 // If the use is actually a legal node, there will not be an
952 // implicit truncate.
953 // FIXME: always querying the result type is just an
954 // approximation; some nodes' legality is determined by the
955 // operand or other means. There's no good way to find out though.
956 if (TLI.isOperationLegalOrCustom(
957 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
960 // Don't bother for PHI nodes.
961 if (isa<PHINode>(TruncUser))
964 BasicBlock *TruncUserBB = TruncUser->getParent();
966 if (UserBB == TruncUserBB)
969 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
970 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
972 if (!InsertedShift && !InsertedTrunc) {
973 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
974 assert(InsertPt != TruncUserBB->end());
976 if (ShiftI->getOpcode() == Instruction::AShr)
977 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
980 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
984 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
986 assert(TruncInsertPt != TruncUserBB->end());
988 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
989 TruncI->getType(), "", &*TruncInsertPt);
993 TruncTheUse = InsertedTrunc;
999 /// Sink the shift *right* instruction into user blocks if the uses could
1000 /// potentially be combined with this shift instruction and generate BitExtract
1001 /// instruction. It will only be applied if the architecture supports BitExtract
1002 /// instruction. Here is an example:
1004 /// %x.extract.shift = lshr i64 %arg1, 32
1006 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1010 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1011 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1013 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1015 /// Return true if any changes are made.
1016 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1017 const TargetLowering &TLI,
1018 const DataLayout &DL) {
1019 BasicBlock *DefBB = ShiftI->getParent();
1021 /// Only insert instructions in each block once.
1022 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1024 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1026 bool MadeChange = false;
1027 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1029 Use &TheUse = UI.getUse();
1030 Instruction *User = cast<Instruction>(*UI);
1031 // Preincrement use iterator so we don't invalidate it.
1034 // Don't bother for PHI nodes.
1035 if (isa<PHINode>(User))
1038 if (!isExtractBitsCandidateUse(User))
1041 BasicBlock *UserBB = User->getParent();
1043 if (UserBB == DefBB) {
1044 // If the shift and truncate instruction are in the same BB. The use of
1045 // the truncate(TruncUse) may still introduce another truncate if not
1046 // legal. In this case, we would like to sink both shift and truncate
1047 // instruction to the BB of TruncUse.
1050 // i64 shift.result = lshr i64 opnd, imm
1051 // trunc.result = trunc shift.result to i16
1054 // ----> We will have an implicit truncate here if the architecture does
1055 // not have i16 compare.
1056 // cmp i16 trunc.result, opnd2
1058 if (isa<TruncInst>(User) && shiftIsLegal
1059 // If the type of the truncate is legal, no trucate will be
1060 // introduced in other basic blocks.
1062 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1064 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1068 // If we have already inserted a shift into this block, use it.
1069 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1071 if (!InsertedShift) {
1072 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1073 assert(InsertPt != UserBB->end());
1075 if (ShiftI->getOpcode() == Instruction::AShr)
1076 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1079 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1085 // Replace a use of the shift with a use of the new shift.
1086 TheUse = InsertedShift;
1089 // If we removed all uses, nuke the shift.
1090 if (ShiftI->use_empty())
1091 ShiftI->eraseFromParent();
1096 // Translate a masked load intrinsic like
1097 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1098 // <16 x i1> %mask, <16 x i32> %passthru)
1099 // to a chain of basic blocks, with loading element one-by-one if
1100 // the appropriate mask bit is set
1102 // %1 = bitcast i8* %addr to i32*
1103 // %2 = extractelement <16 x i1> %mask, i32 0
1104 // %3 = icmp eq i1 %2, true
1105 // br i1 %3, label %cond.load, label %else
1107 //cond.load: ; preds = %0
1108 // %4 = getelementptr i32* %1, i32 0
1109 // %5 = load i32* %4
1110 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1113 //else: ; preds = %0, %cond.load
1114 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1115 // %7 = extractelement <16 x i1> %mask, i32 1
1116 // %8 = icmp eq i1 %7, true
1117 // br i1 %8, label %cond.load1, label %else2
1119 //cond.load1: ; preds = %else
1120 // %9 = getelementptr i32* %1, i32 1
1121 // %10 = load i32* %9
1122 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1125 //else2: ; preds = %else, %cond.load1
1126 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1127 // %12 = extractelement <16 x i1> %mask, i32 2
1128 // %13 = icmp eq i1 %12, true
1129 // br i1 %13, label %cond.load4, label %else5
1131 static void ScalarizeMaskedLoad(CallInst *CI) {
1132 Value *Ptr = CI->getArgOperand(0);
1133 Value *Alignment = CI->getArgOperand(1);
1134 Value *Mask = CI->getArgOperand(2);
1135 Value *Src0 = CI->getArgOperand(3);
1137 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1138 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1139 assert(VecType && "Unexpected return type of masked load intrinsic");
1141 Type *EltTy = CI->getType()->getVectorElementType();
1143 IRBuilder<> Builder(CI->getContext());
1144 Instruction *InsertPt = CI;
1145 BasicBlock *IfBlock = CI->getParent();
1146 BasicBlock *CondBlock = nullptr;
1147 BasicBlock *PrevIfBlock = CI->getParent();
1149 Builder.SetInsertPoint(InsertPt);
1150 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1152 // Short-cut if the mask is all-true.
1153 bool IsAllOnesMask = isa<Constant>(Mask) &&
1154 cast<Constant>(Mask)->isAllOnesValue();
1156 if (IsAllOnesMask) {
1157 Value *NewI = Builder.CreateAlignedLoad(Ptr, AlignVal);
1158 CI->replaceAllUsesWith(NewI);
1159 CI->eraseFromParent();
1163 // Adjust alignment for the scalar instruction.
1164 AlignVal = std::min(AlignVal, VecType->getScalarSizeInBits()/8);
1165 // Bitcast %addr fron i8* to EltTy*
1167 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1168 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1169 unsigned VectorWidth = VecType->getNumElements();
1171 Value *UndefVal = UndefValue::get(VecType);
1173 // The result vector
1174 Value *VResult = UndefVal;
1176 if (isa<ConstantVector>(Mask)) {
1177 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1178 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1181 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1182 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1183 VResult = Builder.CreateInsertElement(VResult, Load,
1184 Builder.getInt32(Idx));
1186 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
1187 CI->replaceAllUsesWith(NewI);
1188 CI->eraseFromParent();
1192 PHINode *Phi = nullptr;
1193 Value *PrevPhi = UndefVal;
1195 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1197 // Fill the "else" block, created in the previous iteration
1199 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1200 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1201 // %to_load = icmp eq i1 %mask_1, true
1202 // br i1 %to_load, label %cond.load, label %else
1205 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1206 Phi->addIncoming(VResult, CondBlock);
1207 Phi->addIncoming(PrevPhi, PrevIfBlock);
1212 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1213 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1214 ConstantInt::get(Predicate->getType(), 1));
1216 // Create "cond" block
1218 // %EltAddr = getelementptr i32* %1, i32 0
1219 // %Elt = load i32* %EltAddr
1220 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1222 CondBlock = IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.load");
1223 Builder.SetInsertPoint(InsertPt);
1226 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1227 LoadInst *Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1228 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1230 // Create "else" block, fill it in the next iteration
1231 BasicBlock *NewIfBlock =
1232 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1233 Builder.SetInsertPoint(InsertPt);
1234 Instruction *OldBr = IfBlock->getTerminator();
1235 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1236 OldBr->eraseFromParent();
1237 PrevIfBlock = IfBlock;
1238 IfBlock = NewIfBlock;
1241 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1242 Phi->addIncoming(VResult, CondBlock);
1243 Phi->addIncoming(PrevPhi, PrevIfBlock);
1244 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1245 CI->replaceAllUsesWith(NewI);
1246 CI->eraseFromParent();
1249 // Translate a masked store intrinsic, like
1250 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1252 // to a chain of basic blocks, that stores element one-by-one if
1253 // the appropriate mask bit is set
1255 // %1 = bitcast i8* %addr to i32*
1256 // %2 = extractelement <16 x i1> %mask, i32 0
1257 // %3 = icmp eq i1 %2, true
1258 // br i1 %3, label %cond.store, label %else
1260 // cond.store: ; preds = %0
1261 // %4 = extractelement <16 x i32> %val, i32 0
1262 // %5 = getelementptr i32* %1, i32 0
1263 // store i32 %4, i32* %5
1266 // else: ; preds = %0, %cond.store
1267 // %6 = extractelement <16 x i1> %mask, i32 1
1268 // %7 = icmp eq i1 %6, true
1269 // br i1 %7, label %cond.store1, label %else2
1271 // cond.store1: ; preds = %else
1272 // %8 = extractelement <16 x i32> %val, i32 1
1273 // %9 = getelementptr i32* %1, i32 1
1274 // store i32 %8, i32* %9
1277 static void ScalarizeMaskedStore(CallInst *CI) {
1278 Value *Src = CI->getArgOperand(0);
1279 Value *Ptr = CI->getArgOperand(1);
1280 Value *Alignment = CI->getArgOperand(2);
1281 Value *Mask = CI->getArgOperand(3);
1283 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1284 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1285 assert(VecType && "Unexpected data type in masked store intrinsic");
1287 Type *EltTy = VecType->getElementType();
1289 IRBuilder<> Builder(CI->getContext());
1290 Instruction *InsertPt = CI;
1291 BasicBlock *IfBlock = CI->getParent();
1292 Builder.SetInsertPoint(InsertPt);
1293 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1295 // Short-cut if the mask is all-true.
1296 bool IsAllOnesMask = isa<Constant>(Mask) &&
1297 cast<Constant>(Mask)->isAllOnesValue();
1299 if (IsAllOnesMask) {
1300 Builder.CreateAlignedStore(Src, Ptr, AlignVal);
1301 CI->eraseFromParent();
1305 // Adjust alignment for the scalar instruction.
1306 AlignVal = std::max(AlignVal, VecType->getScalarSizeInBits()/8);
1307 // Bitcast %addr fron i8* to EltTy*
1309 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1310 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1311 unsigned VectorWidth = VecType->getNumElements();
1313 if (isa<ConstantVector>(Mask)) {
1314 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1315 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1317 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1319 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1320 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1322 CI->eraseFromParent();
1326 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1328 // Fill the "else" block, created in the previous iteration
1330 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1331 // %to_store = icmp eq i1 %mask_1, true
1332 // br i1 %to_store, label %cond.store, label %else
1334 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1335 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1336 ConstantInt::get(Predicate->getType(), 1));
1338 // Create "cond" block
1340 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1341 // %EltAddr = getelementptr i32* %1, i32 0
1342 // %store i32 %OneElt, i32* %EltAddr
1344 BasicBlock *CondBlock =
1345 IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.store");
1346 Builder.SetInsertPoint(InsertPt);
1348 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1350 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1351 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1353 // Create "else" block, fill it in the next iteration
1354 BasicBlock *NewIfBlock =
1355 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1356 Builder.SetInsertPoint(InsertPt);
1357 Instruction *OldBr = IfBlock->getTerminator();
1358 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1359 OldBr->eraseFromParent();
1360 IfBlock = NewIfBlock;
1362 CI->eraseFromParent();
1365 // Translate a masked gather intrinsic like
1366 // <16 x i32 > @llvm.masked.gather.v16i32( <16 x i32*> %Ptrs, i32 4,
1367 // <16 x i1> %Mask, <16 x i32> %Src)
1368 // to a chain of basic blocks, with loading element one-by-one if
1369 // the appropriate mask bit is set
1371 // % Ptrs = getelementptr i32, i32* %base, <16 x i64> %ind
1372 // % Mask0 = extractelement <16 x i1> %Mask, i32 0
1373 // % ToLoad0 = icmp eq i1 % Mask0, true
1374 // br i1 % ToLoad0, label %cond.load, label %else
1377 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
1378 // % Load0 = load i32, i32* % Ptr0, align 4
1379 // % Res0 = insertelement <16 x i32> undef, i32 % Load0, i32 0
1383 // %res.phi.else = phi <16 x i32>[% Res0, %cond.load], [undef, % 0]
1384 // % Mask1 = extractelement <16 x i1> %Mask, i32 1
1385 // % ToLoad1 = icmp eq i1 % Mask1, true
1386 // br i1 % ToLoad1, label %cond.load1, label %else2
1389 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1390 // % Load1 = load i32, i32* % Ptr1, align 4
1391 // % Res1 = insertelement <16 x i32> %res.phi.else, i32 % Load1, i32 1
1394 // % Result = select <16 x i1> %Mask, <16 x i32> %res.phi.select, <16 x i32> %Src
1395 // ret <16 x i32> %Result
1396 static void ScalarizeMaskedGather(CallInst *CI) {
1397 Value *Ptrs = CI->getArgOperand(0);
1398 Value *Alignment = CI->getArgOperand(1);
1399 Value *Mask = CI->getArgOperand(2);
1400 Value *Src0 = CI->getArgOperand(3);
1402 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1404 assert(VecType && "Unexpected return type of masked load intrinsic");
1406 IRBuilder<> Builder(CI->getContext());
1407 Instruction *InsertPt = CI;
1408 BasicBlock *IfBlock = CI->getParent();
1409 BasicBlock *CondBlock = nullptr;
1410 BasicBlock *PrevIfBlock = CI->getParent();
1411 Builder.SetInsertPoint(InsertPt);
1412 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1414 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1416 Value *UndefVal = UndefValue::get(VecType);
1418 // The result vector
1419 Value *VResult = UndefVal;
1420 unsigned VectorWidth = VecType->getNumElements();
1422 // Shorten the way if the mask is a vector of constants.
1423 bool IsConstMask = isa<ConstantVector>(Mask);
1426 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1427 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1429 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1430 "Ptr" + Twine(Idx));
1431 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
1432 "Load" + Twine(Idx));
1433 VResult = Builder.CreateInsertElement(VResult, Load,
1434 Builder.getInt32(Idx),
1435 "Res" + Twine(Idx));
1437 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
1438 CI->replaceAllUsesWith(NewI);
1439 CI->eraseFromParent();
1443 PHINode *Phi = nullptr;
1444 Value *PrevPhi = UndefVal;
1446 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1448 // Fill the "else" block, created in the previous iteration
1450 // %Mask1 = extractelement <16 x i1> %Mask, i32 1
1451 // %ToLoad1 = icmp eq i1 %Mask1, true
1452 // br i1 %ToLoad1, label %cond.load, label %else
1455 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1456 Phi->addIncoming(VResult, CondBlock);
1457 Phi->addIncoming(PrevPhi, PrevIfBlock);
1462 Value *Predicate = Builder.CreateExtractElement(Mask,
1463 Builder.getInt32(Idx),
1464 "Mask" + Twine(Idx));
1465 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1466 ConstantInt::get(Predicate->getType(), 1),
1467 "ToLoad" + Twine(Idx));
1469 // Create "cond" block
1471 // %EltAddr = getelementptr i32* %1, i32 0
1472 // %Elt = load i32* %EltAddr
1473 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1475 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1476 Builder.SetInsertPoint(InsertPt);
1478 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1479 "Ptr" + Twine(Idx));
1480 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
1481 "Load" + Twine(Idx));
1482 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx),
1483 "Res" + Twine(Idx));
1485 // Create "else" block, fill it in the next iteration
1486 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1487 Builder.SetInsertPoint(InsertPt);
1488 Instruction *OldBr = IfBlock->getTerminator();
1489 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1490 OldBr->eraseFromParent();
1491 PrevIfBlock = IfBlock;
1492 IfBlock = NewIfBlock;
1495 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1496 Phi->addIncoming(VResult, CondBlock);
1497 Phi->addIncoming(PrevPhi, PrevIfBlock);
1498 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1499 CI->replaceAllUsesWith(NewI);
1500 CI->eraseFromParent();
1503 // Translate a masked scatter intrinsic, like
1504 // void @llvm.masked.scatter.v16i32(<16 x i32> %Src, <16 x i32*>* %Ptrs, i32 4,
1506 // to a chain of basic blocks, that stores element one-by-one if
1507 // the appropriate mask bit is set.
1509 // % Ptrs = getelementptr i32, i32* %ptr, <16 x i64> %ind
1510 // % Mask0 = extractelement <16 x i1> % Mask, i32 0
1511 // % ToStore0 = icmp eq i1 % Mask0, true
1512 // br i1 %ToStore0, label %cond.store, label %else
1515 // % Elt0 = extractelement <16 x i32> %Src, i32 0
1516 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
1517 // store i32 %Elt0, i32* % Ptr0, align 4
1521 // % Mask1 = extractelement <16 x i1> % Mask, i32 1
1522 // % ToStore1 = icmp eq i1 % Mask1, true
1523 // br i1 % ToStore1, label %cond.store1, label %else2
1526 // % Elt1 = extractelement <16 x i32> %Src, i32 1
1527 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1528 // store i32 % Elt1, i32* % Ptr1, align 4
1531 static void ScalarizeMaskedScatter(CallInst *CI) {
1532 Value *Src = CI->getArgOperand(0);
1533 Value *Ptrs = CI->getArgOperand(1);
1534 Value *Alignment = CI->getArgOperand(2);
1535 Value *Mask = CI->getArgOperand(3);
1537 assert(isa<VectorType>(Src->getType()) &&
1538 "Unexpected data type in masked scatter intrinsic");
1539 assert(isa<VectorType>(Ptrs->getType()) &&
1540 isa<PointerType>(Ptrs->getType()->getVectorElementType()) &&
1541 "Vector of pointers is expected in masked scatter intrinsic");
1543 IRBuilder<> Builder(CI->getContext());
1544 Instruction *InsertPt = CI;
1545 BasicBlock *IfBlock = CI->getParent();
1546 Builder.SetInsertPoint(InsertPt);
1547 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1549 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1550 unsigned VectorWidth = Src->getType()->getVectorNumElements();
1552 // Shorten the way if the mask is a vector of constants.
1553 bool IsConstMask = isa<ConstantVector>(Mask);
1556 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1557 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1559 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
1560 "Elt" + Twine(Idx));
1561 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1562 "Ptr" + Twine(Idx));
1563 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
1565 CI->eraseFromParent();
1568 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1569 // Fill the "else" block, created in the previous iteration
1571 // % Mask1 = extractelement <16 x i1> % Mask, i32 Idx
1572 // % ToStore = icmp eq i1 % Mask1, true
1573 // br i1 % ToStore, label %cond.store, label %else
1575 Value *Predicate = Builder.CreateExtractElement(Mask,
1576 Builder.getInt32(Idx),
1577 "Mask" + Twine(Idx));
1579 Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1580 ConstantInt::get(Predicate->getType(), 1),
1581 "ToStore" + Twine(Idx));
1583 // Create "cond" block
1585 // % Elt1 = extractelement <16 x i32> %Src, i32 1
1586 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1587 // %store i32 % Elt1, i32* % Ptr1
1589 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1590 Builder.SetInsertPoint(InsertPt);
1592 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
1593 "Elt" + Twine(Idx));
1594 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1595 "Ptr" + Twine(Idx));
1596 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
1598 // Create "else" block, fill it in the next iteration
1599 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1600 Builder.SetInsertPoint(InsertPt);
1601 Instruction *OldBr = IfBlock->getTerminator();
1602 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1603 OldBr->eraseFromParent();
1604 IfBlock = NewIfBlock;
1606 CI->eraseFromParent();
1609 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1610 BasicBlock *BB = CI->getParent();
1612 // Lower inline assembly if we can.
1613 // If we found an inline asm expession, and if the target knows how to
1614 // lower it to normal LLVM code, do so now.
1615 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1616 if (TLI->ExpandInlineAsm(CI)) {
1617 // Avoid invalidating the iterator.
1618 CurInstIterator = BB->begin();
1619 // Avoid processing instructions out of order, which could cause
1620 // reuse before a value is defined.
1624 // Sink address computing for memory operands into the block.
1625 if (optimizeInlineAsmInst(CI))
1629 // Align the pointer arguments to this call if the target thinks it's a good
1631 unsigned MinSize, PrefAlign;
1632 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1633 for (auto &Arg : CI->arg_operands()) {
1634 // We want to align both objects whose address is used directly and
1635 // objects whose address is used in casts and GEPs, though it only makes
1636 // sense for GEPs if the offset is a multiple of the desired alignment and
1637 // if size - offset meets the size threshold.
1638 if (!Arg->getType()->isPointerTy())
1640 APInt Offset(DL->getPointerSizeInBits(
1641 cast<PointerType>(Arg->getType())->getAddressSpace()),
1643 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
1644 uint64_t Offset2 = Offset.getLimitedValue();
1645 if ((Offset2 & (PrefAlign-1)) != 0)
1648 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1649 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1650 AI->setAlignment(PrefAlign);
1651 // Global variables can only be aligned if they are defined in this
1652 // object (i.e. they are uniquely initialized in this object), and
1653 // over-aligning global variables that have an explicit section is
1656 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->hasUniqueInitializer() &&
1657 !GV->hasSection() && GV->getAlignment() < PrefAlign &&
1658 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
1660 GV->setAlignment(PrefAlign);
1662 // If this is a memcpy (or similar) then we may be able to improve the
1664 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1665 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
1666 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1667 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
1668 if (Align > MI->getAlignment())
1669 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1673 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1675 switch (II->getIntrinsicID()) {
1677 case Intrinsic::objectsize: {
1678 // Lower all uses of llvm.objectsize.*
1679 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1680 Type *ReturnTy = CI->getType();
1681 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1683 // Substituting this can cause recursive simplifications, which can
1684 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1686 WeakVH IterHandle(&*CurInstIterator);
1688 replaceAndRecursivelySimplify(CI, RetVal,
1691 // If the iterator instruction was recursively deleted, start over at the
1692 // start of the block.
1693 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
1694 CurInstIterator = BB->begin();
1699 case Intrinsic::masked_load: {
1700 // Scalarize unsupported vector masked load
1701 if (!TTI->isLegalMaskedLoad(CI->getType())) {
1702 ScalarizeMaskedLoad(CI);
1708 case Intrinsic::masked_store: {
1709 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType())) {
1710 ScalarizeMaskedStore(CI);
1716 case Intrinsic::masked_gather: {
1717 if (!TTI->isLegalMaskedGather(CI->getType())) {
1718 ScalarizeMaskedGather(CI);
1724 case Intrinsic::masked_scatter: {
1725 if (!TTI->isLegalMaskedScatter(CI->getArgOperand(0)->getType())) {
1726 ScalarizeMaskedScatter(CI);
1732 case Intrinsic::aarch64_stlxr:
1733 case Intrinsic::aarch64_stxr: {
1734 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1735 if (!ExtVal || !ExtVal->hasOneUse() ||
1736 ExtVal->getParent() == CI->getParent())
1738 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1739 ExtVal->moveBefore(CI);
1740 // Mark this instruction as "inserted by CGP", so that other
1741 // optimizations don't touch it.
1742 InsertedInsts.insert(ExtVal);
1745 case Intrinsic::invariant_group_barrier:
1746 II->replaceAllUsesWith(II->getArgOperand(0));
1747 II->eraseFromParent();
1752 // Unknown address space.
1753 // TODO: Target hook to pick which address space the intrinsic cares
1755 unsigned AddrSpace = ~0u;
1756 SmallVector<Value*, 2> PtrOps;
1758 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
1759 while (!PtrOps.empty())
1760 if (optimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
1765 // From here on out we're working with named functions.
1766 if (!CI->getCalledFunction()) return false;
1768 // Lower all default uses of _chk calls. This is very similar
1769 // to what InstCombineCalls does, but here we are only lowering calls
1770 // to fortified library functions (e.g. __memcpy_chk) that have the default
1771 // "don't know" as the objectsize. Anything else should be left alone.
1772 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1773 if (Value *V = Simplifier.optimizeCall(CI)) {
1774 CI->replaceAllUsesWith(V);
1775 CI->eraseFromParent();
1781 /// Look for opportunities to duplicate return instructions to the predecessor
1782 /// to enable tail call optimizations. The case it is currently looking for is:
1785 /// %tmp0 = tail call i32 @f0()
1786 /// br label %return
1788 /// %tmp1 = tail call i32 @f1()
1789 /// br label %return
1791 /// %tmp2 = tail call i32 @f2()
1792 /// br label %return
1794 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1802 /// %tmp0 = tail call i32 @f0()
1805 /// %tmp1 = tail call i32 @f1()
1808 /// %tmp2 = tail call i32 @f2()
1811 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1815 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1819 PHINode *PN = nullptr;
1820 BitCastInst *BCI = nullptr;
1821 Value *V = RI->getReturnValue();
1823 BCI = dyn_cast<BitCastInst>(V);
1825 V = BCI->getOperand(0);
1827 PN = dyn_cast<PHINode>(V);
1832 if (PN && PN->getParent() != BB)
1835 // It's not safe to eliminate the sign / zero extension of the return value.
1836 // See llvm::isInTailCallPosition().
1837 const Function *F = BB->getParent();
1838 AttributeSet CallerAttrs = F->getAttributes();
1839 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1840 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1843 // Make sure there are no instructions between the PHI and return, or that the
1844 // return is the first instruction in the block.
1846 BasicBlock::iterator BI = BB->begin();
1847 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1849 // Also skip over the bitcast.
1854 BasicBlock::iterator BI = BB->begin();
1855 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1860 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1862 SmallVector<CallInst*, 4> TailCalls;
1864 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1865 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1866 // Make sure the phi value is indeed produced by the tail call.
1867 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1868 TLI->mayBeEmittedAsTailCall(CI))
1869 TailCalls.push_back(CI);
1872 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1873 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1874 if (!VisitedBBs.insert(*PI).second)
1877 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1878 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1879 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1880 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1884 CallInst *CI = dyn_cast<CallInst>(&*RI);
1885 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1886 TailCalls.push_back(CI);
1890 bool Changed = false;
1891 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1892 CallInst *CI = TailCalls[i];
1895 // Conservatively require the attributes of the call to match those of the
1896 // return. Ignore noalias because it doesn't affect the call sequence.
1897 AttributeSet CalleeAttrs = CS.getAttributes();
1898 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1899 removeAttribute(Attribute::NoAlias) !=
1900 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1901 removeAttribute(Attribute::NoAlias))
1904 // Make sure the call instruction is followed by an unconditional branch to
1905 // the return block.
1906 BasicBlock *CallBB = CI->getParent();
1907 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1908 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1911 // Duplicate the return into CallBB.
1912 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1913 ModifiedDT = Changed = true;
1917 // If we eliminated all predecessors of the block, delete the block now.
1918 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1919 BB->eraseFromParent();
1924 //===----------------------------------------------------------------------===//
1925 // Memory Optimization
1926 //===----------------------------------------------------------------------===//
1930 /// This is an extended version of TargetLowering::AddrMode
1931 /// which holds actual Value*'s for register values.
1932 struct ExtAddrMode : public TargetLowering::AddrMode {
1935 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1936 void print(raw_ostream &OS) const;
1939 bool operator==(const ExtAddrMode& O) const {
1940 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1941 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1942 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1947 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1953 void ExtAddrMode::print(raw_ostream &OS) const {
1954 bool NeedPlus = false;
1957 OS << (NeedPlus ? " + " : "")
1959 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1964 OS << (NeedPlus ? " + " : "")
1970 OS << (NeedPlus ? " + " : "")
1972 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1976 OS << (NeedPlus ? " + " : "")
1978 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1984 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1985 void ExtAddrMode::dump() const {
1991 /// \brief This class provides transaction based operation on the IR.
1992 /// Every change made through this class is recorded in the internal state and
1993 /// can be undone (rollback) until commit is called.
1994 class TypePromotionTransaction {
1996 /// \brief This represents the common interface of the individual transaction.
1997 /// Each class implements the logic for doing one specific modification on
1998 /// the IR via the TypePromotionTransaction.
1999 class TypePromotionAction {
2001 /// The Instruction modified.
2005 /// \brief Constructor of the action.
2006 /// The constructor performs the related action on the IR.
2007 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2009 virtual ~TypePromotionAction() {}
2011 /// \brief Undo the modification done by this action.
2012 /// When this method is called, the IR must be in the same state as it was
2013 /// before this action was applied.
2014 /// \pre Undoing the action works if and only if the IR is in the exact same
2015 /// state as it was directly after this action was applied.
2016 virtual void undo() = 0;
2018 /// \brief Advocate every change made by this action.
2019 /// When the results on the IR of the action are to be kept, it is important
2020 /// to call this function, otherwise hidden information may be kept forever.
2021 virtual void commit() {
2022 // Nothing to be done, this action is not doing anything.
2026 /// \brief Utility to remember the position of an instruction.
2027 class InsertionHandler {
2028 /// Position of an instruction.
2029 /// Either an instruction:
2030 /// - Is the first in a basic block: BB is used.
2031 /// - Has a previous instructon: PrevInst is used.
2033 Instruction *PrevInst;
2036 /// Remember whether or not the instruction had a previous instruction.
2037 bool HasPrevInstruction;
2040 /// \brief Record the position of \p Inst.
2041 InsertionHandler(Instruction *Inst) {
2042 BasicBlock::iterator It = Inst->getIterator();
2043 HasPrevInstruction = (It != (Inst->getParent()->begin()));
2044 if (HasPrevInstruction)
2045 Point.PrevInst = &*--It;
2047 Point.BB = Inst->getParent();
2050 /// \brief Insert \p Inst at the recorded position.
2051 void insert(Instruction *Inst) {
2052 if (HasPrevInstruction) {
2053 if (Inst->getParent())
2054 Inst->removeFromParent();
2055 Inst->insertAfter(Point.PrevInst);
2057 Instruction *Position = &*Point.BB->getFirstInsertionPt();
2058 if (Inst->getParent())
2059 Inst->moveBefore(Position);
2061 Inst->insertBefore(Position);
2066 /// \brief Move an instruction before another.
2067 class InstructionMoveBefore : public TypePromotionAction {
2068 /// Original position of the instruction.
2069 InsertionHandler Position;
2072 /// \brief Move \p Inst before \p Before.
2073 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2074 : TypePromotionAction(Inst), Position(Inst) {
2075 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
2076 Inst->moveBefore(Before);
2079 /// \brief Move the instruction back to its original position.
2080 void undo() override {
2081 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2082 Position.insert(Inst);
2086 /// \brief Set the operand of an instruction with a new value.
2087 class OperandSetter : public TypePromotionAction {
2088 /// Original operand of the instruction.
2090 /// Index of the modified instruction.
2094 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
2095 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2096 : TypePromotionAction(Inst), Idx(Idx) {
2097 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2098 << "for:" << *Inst << "\n"
2099 << "with:" << *NewVal << "\n");
2100 Origin = Inst->getOperand(Idx);
2101 Inst->setOperand(Idx, NewVal);
2104 /// \brief Restore the original value of the instruction.
2105 void undo() override {
2106 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2107 << "for: " << *Inst << "\n"
2108 << "with: " << *Origin << "\n");
2109 Inst->setOperand(Idx, Origin);
2113 /// \brief Hide the operands of an instruction.
2114 /// Do as if this instruction was not using any of its operands.
2115 class OperandsHider : public TypePromotionAction {
2116 /// The list of original operands.
2117 SmallVector<Value *, 4> OriginalValues;
2120 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
2121 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2122 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2123 unsigned NumOpnds = Inst->getNumOperands();
2124 OriginalValues.reserve(NumOpnds);
2125 for (unsigned It = 0; It < NumOpnds; ++It) {
2126 // Save the current operand.
2127 Value *Val = Inst->getOperand(It);
2128 OriginalValues.push_back(Val);
2130 // We could use OperandSetter here, but that would imply an overhead
2131 // that we are not willing to pay.
2132 Inst->setOperand(It, UndefValue::get(Val->getType()));
2136 /// \brief Restore the original list of uses.
2137 void undo() override {
2138 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2139 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2140 Inst->setOperand(It, OriginalValues[It]);
2144 /// \brief Build a truncate instruction.
2145 class TruncBuilder : public TypePromotionAction {
2148 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
2150 /// trunc Opnd to Ty.
2151 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2152 IRBuilder<> Builder(Opnd);
2153 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2154 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2157 /// \brief Get the built value.
2158 Value *getBuiltValue() { return Val; }
2160 /// \brief Remove the built instruction.
2161 void undo() override {
2162 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2163 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2164 IVal->eraseFromParent();
2168 /// \brief Build a sign extension instruction.
2169 class SExtBuilder : public TypePromotionAction {
2172 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
2174 /// sext Opnd to Ty.
2175 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2176 : TypePromotionAction(InsertPt) {
2177 IRBuilder<> Builder(InsertPt);
2178 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2179 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2182 /// \brief Get the built value.
2183 Value *getBuiltValue() { return Val; }
2185 /// \brief Remove the built instruction.
2186 void undo() override {
2187 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2188 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2189 IVal->eraseFromParent();
2193 /// \brief Build a zero extension instruction.
2194 class ZExtBuilder : public TypePromotionAction {
2197 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
2199 /// zext Opnd to Ty.
2200 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2201 : TypePromotionAction(InsertPt) {
2202 IRBuilder<> Builder(InsertPt);
2203 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2204 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2207 /// \brief Get the built value.
2208 Value *getBuiltValue() { return Val; }
2210 /// \brief Remove the built instruction.
2211 void undo() override {
2212 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2213 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2214 IVal->eraseFromParent();
2218 /// \brief Mutate an instruction to another type.
2219 class TypeMutator : public TypePromotionAction {
2220 /// Record the original type.
2224 /// \brief Mutate the type of \p Inst into \p NewTy.
2225 TypeMutator(Instruction *Inst, Type *NewTy)
2226 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2227 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2229 Inst->mutateType(NewTy);
2232 /// \brief Mutate the instruction back to its original type.
2233 void undo() override {
2234 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2236 Inst->mutateType(OrigTy);
2240 /// \brief Replace the uses of an instruction by another instruction.
2241 class UsesReplacer : public TypePromotionAction {
2242 /// Helper structure to keep track of the replaced uses.
2243 struct InstructionAndIdx {
2244 /// The instruction using the instruction.
2246 /// The index where this instruction is used for Inst.
2248 InstructionAndIdx(Instruction *Inst, unsigned Idx)
2249 : Inst(Inst), Idx(Idx) {}
2252 /// Keep track of the original uses (pair Instruction, Index).
2253 SmallVector<InstructionAndIdx, 4> OriginalUses;
2254 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
2257 /// \brief Replace all the use of \p Inst by \p New.
2258 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2259 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2261 // Record the original uses.
2262 for (Use &U : Inst->uses()) {
2263 Instruction *UserI = cast<Instruction>(U.getUser());
2264 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2266 // Now, we can replace the uses.
2267 Inst->replaceAllUsesWith(New);
2270 /// \brief Reassign the original uses of Inst to Inst.
2271 void undo() override {
2272 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2273 for (use_iterator UseIt = OriginalUses.begin(),
2274 EndIt = OriginalUses.end();
2275 UseIt != EndIt; ++UseIt) {
2276 UseIt->Inst->setOperand(UseIt->Idx, Inst);
2281 /// \brief Remove an instruction from the IR.
2282 class InstructionRemover : public TypePromotionAction {
2283 /// Original position of the instruction.
2284 InsertionHandler Inserter;
2285 /// Helper structure to hide all the link to the instruction. In other
2286 /// words, this helps to do as if the instruction was removed.
2287 OperandsHider Hider;
2288 /// Keep track of the uses replaced, if any.
2289 UsesReplacer *Replacer;
2292 /// \brief Remove all reference of \p Inst and optinally replace all its
2294 /// \pre If !Inst->use_empty(), then New != nullptr
2295 InstructionRemover(Instruction *Inst, Value *New = nullptr)
2296 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2299 Replacer = new UsesReplacer(Inst, New);
2300 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2301 Inst->removeFromParent();
2304 ~InstructionRemover() override { delete Replacer; }
2306 /// \brief Really remove the instruction.
2307 void commit() override { delete Inst; }
2309 /// \brief Resurrect the instruction and reassign it to the proper uses if
2310 /// new value was provided when build this action.
2311 void undo() override {
2312 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2313 Inserter.insert(Inst);
2321 /// Restoration point.
2322 /// The restoration point is a pointer to an action instead of an iterator
2323 /// because the iterator may be invalidated but not the pointer.
2324 typedef const TypePromotionAction *ConstRestorationPt;
2325 /// Advocate every changes made in that transaction.
2327 /// Undo all the changes made after the given point.
2328 void rollback(ConstRestorationPt Point);
2329 /// Get the current restoration point.
2330 ConstRestorationPt getRestorationPoint() const;
2332 /// \name API for IR modification with state keeping to support rollback.
2334 /// Same as Instruction::setOperand.
2335 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2336 /// Same as Instruction::eraseFromParent.
2337 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2338 /// Same as Value::replaceAllUsesWith.
2339 void replaceAllUsesWith(Instruction *Inst, Value *New);
2340 /// Same as Value::mutateType.
2341 void mutateType(Instruction *Inst, Type *NewTy);
2342 /// Same as IRBuilder::createTrunc.
2343 Value *createTrunc(Instruction *Opnd, Type *Ty);
2344 /// Same as IRBuilder::createSExt.
2345 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2346 /// Same as IRBuilder::createZExt.
2347 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2348 /// Same as Instruction::moveBefore.
2349 void moveBefore(Instruction *Inst, Instruction *Before);
2353 /// The ordered list of actions made so far.
2354 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
2355 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
2358 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2361 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
2364 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2367 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
2370 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2372 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2375 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2376 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2379 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2381 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2382 Value *Val = Ptr->getBuiltValue();
2383 Actions.push_back(std::move(Ptr));
2387 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2388 Value *Opnd, Type *Ty) {
2389 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2390 Value *Val = Ptr->getBuiltValue();
2391 Actions.push_back(std::move(Ptr));
2395 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2396 Value *Opnd, Type *Ty) {
2397 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2398 Value *Val = Ptr->getBuiltValue();
2399 Actions.push_back(std::move(Ptr));
2403 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2404 Instruction *Before) {
2406 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
2409 TypePromotionTransaction::ConstRestorationPt
2410 TypePromotionTransaction::getRestorationPoint() const {
2411 return !Actions.empty() ? Actions.back().get() : nullptr;
2414 void TypePromotionTransaction::commit() {
2415 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2421 void TypePromotionTransaction::rollback(
2422 TypePromotionTransaction::ConstRestorationPt Point) {
2423 while (!Actions.empty() && Point != Actions.back().get()) {
2424 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2429 /// \brief A helper class for matching addressing modes.
2431 /// This encapsulates the logic for matching the target-legal addressing modes.
2432 class AddressingModeMatcher {
2433 SmallVectorImpl<Instruction*> &AddrModeInsts;
2434 const TargetMachine &TM;
2435 const TargetLowering &TLI;
2436 const DataLayout &DL;
2438 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2439 /// the memory instruction that we're computing this address for.
2442 Instruction *MemoryInst;
2444 /// This is the addressing mode that we're building up. This is
2445 /// part of the return value of this addressing mode matching stuff.
2446 ExtAddrMode &AddrMode;
2448 /// The instructions inserted by other CodeGenPrepare optimizations.
2449 const SetOfInstrs &InsertedInsts;
2450 /// A map from the instructions to their type before promotion.
2451 InstrToOrigTy &PromotedInsts;
2452 /// The ongoing transaction where every action should be registered.
2453 TypePromotionTransaction &TPT;
2455 /// This is set to true when we should not do profitability checks.
2456 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2457 bool IgnoreProfitability;
2459 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2460 const TargetMachine &TM, Type *AT, unsigned AS,
2461 Instruction *MI, ExtAddrMode &AM,
2462 const SetOfInstrs &InsertedInsts,
2463 InstrToOrigTy &PromotedInsts,
2464 TypePromotionTransaction &TPT)
2465 : AddrModeInsts(AMI), TM(TM),
2466 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
2467 ->getTargetLowering()),
2468 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2469 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2470 PromotedInsts(PromotedInsts), TPT(TPT) {
2471 IgnoreProfitability = false;
2475 /// Find the maximal addressing mode that a load/store of V can fold,
2476 /// give an access type of AccessTy. This returns a list of involved
2477 /// instructions in AddrModeInsts.
2478 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2480 /// \p PromotedInsts maps the instructions to their type before promotion.
2481 /// \p The ongoing transaction where every action should be registered.
2482 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2483 Instruction *MemoryInst,
2484 SmallVectorImpl<Instruction*> &AddrModeInsts,
2485 const TargetMachine &TM,
2486 const SetOfInstrs &InsertedInsts,
2487 InstrToOrigTy &PromotedInsts,
2488 TypePromotionTransaction &TPT) {
2491 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
2492 MemoryInst, Result, InsertedInsts,
2493 PromotedInsts, TPT).matchAddr(V, 0);
2494 (void)Success; assert(Success && "Couldn't select *anything*?");
2498 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2499 bool matchAddr(Value *V, unsigned Depth);
2500 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2501 bool *MovedAway = nullptr);
2502 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2503 ExtAddrMode &AMBefore,
2504 ExtAddrMode &AMAfter);
2505 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2506 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2507 Value *PromotedOperand) const;
2510 /// Try adding ScaleReg*Scale to the current addressing mode.
2511 /// Return true and update AddrMode if this addr mode is legal for the target,
2513 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
2515 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2516 // mode. Just process that directly.
2518 return matchAddr(ScaleReg, Depth);
2520 // If the scale is 0, it takes nothing to add this.
2524 // If we already have a scale of this value, we can add to it, otherwise, we
2525 // need an available scale field.
2526 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2529 ExtAddrMode TestAddrMode = AddrMode;
2531 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2532 // [A+B + A*7] -> [B+A*8].
2533 TestAddrMode.Scale += Scale;
2534 TestAddrMode.ScaledReg = ScaleReg;
2536 // If the new address isn't legal, bail out.
2537 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
2540 // It was legal, so commit it.
2541 AddrMode = TestAddrMode;
2543 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2544 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2545 // X*Scale + C*Scale to addr mode.
2546 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2547 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2548 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2549 TestAddrMode.ScaledReg = AddLHS;
2550 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2552 // If this addressing mode is legal, commit it and remember that we folded
2553 // this instruction.
2554 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
2555 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2556 AddrMode = TestAddrMode;
2561 // Otherwise, not (x+c)*scale, just return what we have.
2565 /// This is a little filter, which returns true if an addressing computation
2566 /// involving I might be folded into a load/store accessing it.
2567 /// This doesn't need to be perfect, but needs to accept at least
2568 /// the set of instructions that MatchOperationAddr can.
2569 static bool MightBeFoldableInst(Instruction *I) {
2570 switch (I->getOpcode()) {
2571 case Instruction::BitCast:
2572 case Instruction::AddrSpaceCast:
2573 // Don't touch identity bitcasts.
2574 if (I->getType() == I->getOperand(0)->getType())
2576 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2577 case Instruction::PtrToInt:
2578 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2580 case Instruction::IntToPtr:
2581 // We know the input is intptr_t, so this is foldable.
2583 case Instruction::Add:
2585 case Instruction::Mul:
2586 case Instruction::Shl:
2587 // Can only handle X*C and X << C.
2588 return isa<ConstantInt>(I->getOperand(1));
2589 case Instruction::GetElementPtr:
2596 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2597 /// \note \p Val is assumed to be the product of some type promotion.
2598 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2599 /// to be legal, as the non-promoted value would have had the same state.
2600 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
2601 const DataLayout &DL, Value *Val) {
2602 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2605 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2606 // If the ISDOpcode is undefined, it was undefined before the promotion.
2609 // Otherwise, check if the promoted instruction is legal or not.
2610 return TLI.isOperationLegalOrCustom(
2611 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
2614 /// \brief Hepler class to perform type promotion.
2615 class TypePromotionHelper {
2616 /// \brief Utility function to check whether or not a sign or zero extension
2617 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2618 /// either using the operands of \p Inst or promoting \p Inst.
2619 /// The type of the extension is defined by \p IsSExt.
2620 /// In other words, check if:
2621 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2622 /// #1 Promotion applies:
2623 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2624 /// #2 Operand reuses:
2625 /// ext opnd1 to ConsideredExtType.
2626 /// \p PromotedInsts maps the instructions to their type before promotion.
2627 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2628 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2630 /// \brief Utility function to determine if \p OpIdx should be promoted when
2631 /// promoting \p Inst.
2632 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2633 return !(isa<SelectInst>(Inst) && OpIdx == 0);
2636 /// \brief Utility function to promote the operand of \p Ext when this
2637 /// operand is a promotable trunc or sext or zext.
2638 /// \p PromotedInsts maps the instructions to their type before promotion.
2639 /// \p CreatedInstsCost[out] contains the cost of all instructions
2640 /// created to promote the operand of Ext.
2641 /// Newly added extensions are inserted in \p Exts.
2642 /// Newly added truncates are inserted in \p Truncs.
2643 /// Should never be called directly.
2644 /// \return The promoted value which is used instead of Ext.
2645 static Value *promoteOperandForTruncAndAnyExt(
2646 Instruction *Ext, TypePromotionTransaction &TPT,
2647 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2648 SmallVectorImpl<Instruction *> *Exts,
2649 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
2651 /// \brief Utility function to promote the operand of \p Ext when this
2652 /// operand is promotable and is not a supported trunc or sext.
2653 /// \p PromotedInsts maps the instructions to their type before promotion.
2654 /// \p CreatedInstsCost[out] contains the cost of all the instructions
2655 /// created to promote the operand of Ext.
2656 /// Newly added extensions are inserted in \p Exts.
2657 /// Newly added truncates are inserted in \p Truncs.
2658 /// Should never be called directly.
2659 /// \return The promoted value which is used instead of Ext.
2660 static Value *promoteOperandForOther(Instruction *Ext,
2661 TypePromotionTransaction &TPT,
2662 InstrToOrigTy &PromotedInsts,
2663 unsigned &CreatedInstsCost,
2664 SmallVectorImpl<Instruction *> *Exts,
2665 SmallVectorImpl<Instruction *> *Truncs,
2666 const TargetLowering &TLI, bool IsSExt);
2668 /// \see promoteOperandForOther.
2669 static Value *signExtendOperandForOther(
2670 Instruction *Ext, TypePromotionTransaction &TPT,
2671 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2672 SmallVectorImpl<Instruction *> *Exts,
2673 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2674 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2675 Exts, Truncs, TLI, true);
2678 /// \see promoteOperandForOther.
2679 static Value *zeroExtendOperandForOther(
2680 Instruction *Ext, TypePromotionTransaction &TPT,
2681 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2682 SmallVectorImpl<Instruction *> *Exts,
2683 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2684 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2685 Exts, Truncs, TLI, false);
2689 /// Type for the utility function that promotes the operand of Ext.
2690 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2691 InstrToOrigTy &PromotedInsts,
2692 unsigned &CreatedInstsCost,
2693 SmallVectorImpl<Instruction *> *Exts,
2694 SmallVectorImpl<Instruction *> *Truncs,
2695 const TargetLowering &TLI);
2696 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2697 /// action to promote the operand of \p Ext instead of using Ext.
2698 /// \return NULL if no promotable action is possible with the current
2700 /// \p InsertedInsts keeps track of all the instructions inserted by the
2701 /// other CodeGenPrepare optimizations. This information is important
2702 /// because we do not want to promote these instructions as CodeGenPrepare
2703 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2704 /// \p PromotedInsts maps the instructions to their type before promotion.
2705 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
2706 const TargetLowering &TLI,
2707 const InstrToOrigTy &PromotedInsts);
2710 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2711 Type *ConsideredExtType,
2712 const InstrToOrigTy &PromotedInsts,
2714 // The promotion helper does not know how to deal with vector types yet.
2715 // To be able to fix that, we would need to fix the places where we
2716 // statically extend, e.g., constants and such.
2717 if (Inst->getType()->isVectorTy())
2720 // We can always get through zext.
2721 if (isa<ZExtInst>(Inst))
2724 // sext(sext) is ok too.
2725 if (IsSExt && isa<SExtInst>(Inst))
2728 // We can get through binary operator, if it is legal. In other words, the
2729 // binary operator must have a nuw or nsw flag.
2730 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2731 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2732 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2733 (IsSExt && BinOp->hasNoSignedWrap())))
2736 // Check if we can do the following simplification.
2737 // ext(trunc(opnd)) --> ext(opnd)
2738 if (!isa<TruncInst>(Inst))
2741 Value *OpndVal = Inst->getOperand(0);
2742 // Check if we can use this operand in the extension.
2743 // If the type is larger than the result type of the extension, we cannot.
2744 if (!OpndVal->getType()->isIntegerTy() ||
2745 OpndVal->getType()->getIntegerBitWidth() >
2746 ConsideredExtType->getIntegerBitWidth())
2749 // If the operand of the truncate is not an instruction, we will not have
2750 // any information on the dropped bits.
2751 // (Actually we could for constant but it is not worth the extra logic).
2752 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2756 // Check if the source of the type is narrow enough.
2757 // I.e., check that trunc just drops extended bits of the same kind of
2759 // #1 get the type of the operand and check the kind of the extended bits.
2760 const Type *OpndType;
2761 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2762 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
2763 OpndType = It->second.getPointer();
2764 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2765 OpndType = Opnd->getOperand(0)->getType();
2769 // #2 check that the truncate just drops extended bits.
2770 return Inst->getType()->getIntegerBitWidth() >=
2771 OpndType->getIntegerBitWidth();
2774 TypePromotionHelper::Action TypePromotionHelper::getAction(
2775 Instruction *Ext, const SetOfInstrs &InsertedInsts,
2776 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2777 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2778 "Unexpected instruction type");
2779 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2780 Type *ExtTy = Ext->getType();
2781 bool IsSExt = isa<SExtInst>(Ext);
2782 // If the operand of the extension is not an instruction, we cannot
2784 // If it, check we can get through.
2785 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2788 // Do not promote if the operand has been added by codegenprepare.
2789 // Otherwise, it means we are undoing an optimization that is likely to be
2790 // redone, thus causing potential infinite loop.
2791 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
2794 // SExt or Trunc instructions.
2795 // Return the related handler.
2796 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2797 isa<ZExtInst>(ExtOpnd))
2798 return promoteOperandForTruncAndAnyExt;
2800 // Regular instruction.
2801 // Abort early if we will have to insert non-free instructions.
2802 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2804 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2807 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2808 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2809 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2810 SmallVectorImpl<Instruction *> *Exts,
2811 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2812 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2813 // get through it and this method should not be called.
2814 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2815 Value *ExtVal = SExt;
2816 bool HasMergedNonFreeExt = false;
2817 if (isa<ZExtInst>(SExtOpnd)) {
2818 // Replace s|zext(zext(opnd))
2820 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
2822 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2823 TPT.replaceAllUsesWith(SExt, ZExt);
2824 TPT.eraseInstruction(SExt);
2827 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2829 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2831 CreatedInstsCost = 0;
2833 // Remove dead code.
2834 if (SExtOpnd->use_empty())
2835 TPT.eraseInstruction(SExtOpnd);
2837 // Check if the extension is still needed.
2838 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2839 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2842 Exts->push_back(ExtInst);
2843 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
2848 // At this point we have: ext ty opnd to ty.
2849 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2850 Value *NextVal = ExtInst->getOperand(0);
2851 TPT.eraseInstruction(ExtInst, NextVal);
2855 Value *TypePromotionHelper::promoteOperandForOther(
2856 Instruction *Ext, TypePromotionTransaction &TPT,
2857 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2858 SmallVectorImpl<Instruction *> *Exts,
2859 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
2861 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2862 // get through it and this method should not be called.
2863 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2864 CreatedInstsCost = 0;
2865 if (!ExtOpnd->hasOneUse()) {
2866 // ExtOpnd will be promoted.
2867 // All its uses, but Ext, will need to use a truncated value of the
2868 // promoted version.
2869 // Create the truncate now.
2870 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2871 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2872 ITrunc->removeFromParent();
2873 // Insert it just after the definition.
2874 ITrunc->insertAfter(ExtOpnd);
2876 Truncs->push_back(ITrunc);
2879 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2880 // Restore the operand of Ext (which has been replaced by the previous call
2881 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2882 TPT.setOperand(Ext, 0, ExtOpnd);
2885 // Get through the Instruction:
2886 // 1. Update its type.
2887 // 2. Replace the uses of Ext by Inst.
2888 // 3. Extend each operand that needs to be extended.
2890 // Remember the original type of the instruction before promotion.
2891 // This is useful to know that the high bits are sign extended bits.
2892 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2893 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2895 TPT.mutateType(ExtOpnd, Ext->getType());
2897 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2899 Instruction *ExtForOpnd = Ext;
2901 DEBUG(dbgs() << "Propagate Ext to operands\n");
2902 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2904 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2905 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2906 !shouldExtOperand(ExtOpnd, OpIdx)) {
2907 DEBUG(dbgs() << "No need to propagate\n");
2910 // Check if we can statically extend the operand.
2911 Value *Opnd = ExtOpnd->getOperand(OpIdx);
2912 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2913 DEBUG(dbgs() << "Statically extend\n");
2914 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2915 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
2916 : Cst->getValue().zext(BitWidth);
2917 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
2920 // UndefValue are typed, so we have to statically sign extend them.
2921 if (isa<UndefValue>(Opnd)) {
2922 DEBUG(dbgs() << "Statically extend\n");
2923 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2927 // Otherwise we have to explicity sign extend the operand.
2928 // Check if Ext was reused to extend an operand.
2930 // If yes, create a new one.
2931 DEBUG(dbgs() << "More operands to ext\n");
2932 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2933 : TPT.createZExt(Ext, Opnd, Ext->getType());
2934 if (!isa<Instruction>(ValForExtOpnd)) {
2935 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
2938 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
2941 Exts->push_back(ExtForOpnd);
2942 TPT.setOperand(ExtForOpnd, 0, Opnd);
2944 // Move the sign extension before the insertion point.
2945 TPT.moveBefore(ExtForOpnd, ExtOpnd);
2946 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2947 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
2948 // If more sext are required, new instructions will have to be created.
2949 ExtForOpnd = nullptr;
2951 if (ExtForOpnd == Ext) {
2952 DEBUG(dbgs() << "Extension is useless now\n");
2953 TPT.eraseInstruction(Ext);
2958 /// Check whether or not promoting an instruction to a wider type is profitable.
2959 /// \p NewCost gives the cost of extension instructions created by the
2961 /// \p OldCost gives the cost of extension instructions before the promotion
2962 /// plus the number of instructions that have been
2963 /// matched in the addressing mode the promotion.
2964 /// \p PromotedOperand is the value that has been promoted.
2965 /// \return True if the promotion is profitable, false otherwise.
2966 bool AddressingModeMatcher::isPromotionProfitable(
2967 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
2968 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
2969 // The cost of the new extensions is greater than the cost of the
2970 // old extension plus what we folded.
2971 // This is not profitable.
2972 if (NewCost > OldCost)
2974 if (NewCost < OldCost)
2976 // The promotion is neutral but it may help folding the sign extension in
2977 // loads for instance.
2978 // Check that we did not create an illegal instruction.
2979 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
2982 /// Given an instruction or constant expr, see if we can fold the operation
2983 /// into the addressing mode. If so, update the addressing mode and return
2984 /// true, otherwise return false without modifying AddrMode.
2985 /// If \p MovedAway is not NULL, it contains the information of whether or
2986 /// not AddrInst has to be folded into the addressing mode on success.
2987 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2988 /// because it has been moved away.
2989 /// Thus AddrInst must not be added in the matched instructions.
2990 /// This state can happen when AddrInst is a sext, since it may be moved away.
2991 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2992 /// not be referenced anymore.
2993 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
2996 // Avoid exponential behavior on extremely deep expression trees.
2997 if (Depth >= 5) return false;
2999 // By default, all matched instructions stay in place.
3004 case Instruction::PtrToInt:
3005 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3006 return matchAddr(AddrInst->getOperand(0), Depth);
3007 case Instruction::IntToPtr: {
3008 auto AS = AddrInst->getType()->getPointerAddressSpace();
3009 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3010 // This inttoptr is a no-op if the integer type is pointer sized.
3011 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3012 return matchAddr(AddrInst->getOperand(0), Depth);
3015 case Instruction::BitCast:
3016 // BitCast is always a noop, and we can handle it as long as it is
3017 // int->int or pointer->pointer (we don't want int<->fp or something).
3018 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
3019 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
3020 // Don't touch identity bitcasts. These were probably put here by LSR,
3021 // and we don't want to mess around with them. Assume it knows what it
3023 AddrInst->getOperand(0)->getType() != AddrInst->getType())
3024 return matchAddr(AddrInst->getOperand(0), Depth);
3026 case Instruction::AddrSpaceCast: {
3028 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3029 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3030 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3031 return matchAddr(AddrInst->getOperand(0), Depth);
3034 case Instruction::Add: {
3035 // Check to see if we can merge in the RHS then the LHS. If so, we win.
3036 ExtAddrMode BackupAddrMode = AddrMode;
3037 unsigned OldSize = AddrModeInsts.size();
3038 // Start a transaction at this point.
3039 // The LHS may match but not the RHS.
3040 // Therefore, we need a higher level restoration point to undo partially
3041 // matched operation.
3042 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3043 TPT.getRestorationPoint();
3045 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
3046 matchAddr(AddrInst->getOperand(0), Depth+1))
3049 // Restore the old addr mode info.
3050 AddrMode = BackupAddrMode;
3051 AddrModeInsts.resize(OldSize);
3052 TPT.rollback(LastKnownGood);
3054 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
3055 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
3056 matchAddr(AddrInst->getOperand(1), Depth+1))
3059 // Otherwise we definitely can't merge the ADD in.
3060 AddrMode = BackupAddrMode;
3061 AddrModeInsts.resize(OldSize);
3062 TPT.rollback(LastKnownGood);
3065 //case Instruction::Or:
3066 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
3068 case Instruction::Mul:
3069 case Instruction::Shl: {
3070 // Can only handle X*C and X << C.
3071 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
3074 int64_t Scale = RHS->getSExtValue();
3075 if (Opcode == Instruction::Shl)
3076 Scale = 1LL << Scale;
3078 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
3080 case Instruction::GetElementPtr: {
3081 // Scan the GEP. We check it if it contains constant offsets and at most
3082 // one variable offset.
3083 int VariableOperand = -1;
3084 unsigned VariableScale = 0;
3086 int64_t ConstantOffset = 0;
3087 gep_type_iterator GTI = gep_type_begin(AddrInst);
3088 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
3089 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
3090 const StructLayout *SL = DL.getStructLayout(STy);
3092 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
3093 ConstantOffset += SL->getElementOffset(Idx);
3095 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
3096 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
3097 ConstantOffset += CI->getSExtValue()*TypeSize;
3098 } else if (TypeSize) { // Scales of zero don't do anything.
3099 // We only allow one variable index at the moment.
3100 if (VariableOperand != -1)
3103 // Remember the variable index.
3104 VariableOperand = i;
3105 VariableScale = TypeSize;
3110 // A common case is for the GEP to only do a constant offset. In this case,
3111 // just add it to the disp field and check validity.
3112 if (VariableOperand == -1) {
3113 AddrMode.BaseOffs += ConstantOffset;
3114 if (ConstantOffset == 0 ||
3115 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
3116 // Check to see if we can fold the base pointer in too.
3117 if (matchAddr(AddrInst->getOperand(0), Depth+1))
3120 AddrMode.BaseOffs -= ConstantOffset;
3124 // Save the valid addressing mode in case we can't match.
3125 ExtAddrMode BackupAddrMode = AddrMode;
3126 unsigned OldSize = AddrModeInsts.size();
3128 // See if the scale and offset amount is valid for this target.
3129 AddrMode.BaseOffs += ConstantOffset;
3131 // Match the base operand of the GEP.
3132 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
3133 // If it couldn't be matched, just stuff the value in a register.
3134 if (AddrMode.HasBaseReg) {
3135 AddrMode = BackupAddrMode;
3136 AddrModeInsts.resize(OldSize);
3139 AddrMode.HasBaseReg = true;
3140 AddrMode.BaseReg = AddrInst->getOperand(0);
3143 // Match the remaining variable portion of the GEP.
3144 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
3146 // If it couldn't be matched, try stuffing the base into a register
3147 // instead of matching it, and retrying the match of the scale.
3148 AddrMode = BackupAddrMode;
3149 AddrModeInsts.resize(OldSize);
3150 if (AddrMode.HasBaseReg)
3152 AddrMode.HasBaseReg = true;
3153 AddrMode.BaseReg = AddrInst->getOperand(0);
3154 AddrMode.BaseOffs += ConstantOffset;
3155 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
3156 VariableScale, Depth)) {
3157 // If even that didn't work, bail.
3158 AddrMode = BackupAddrMode;
3159 AddrModeInsts.resize(OldSize);
3166 case Instruction::SExt:
3167 case Instruction::ZExt: {
3168 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
3172 // Try to move this ext out of the way of the addressing mode.
3173 // Ask for a method for doing so.
3174 TypePromotionHelper::Action TPH =
3175 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
3179 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3180 TPT.getRestorationPoint();
3181 unsigned CreatedInstsCost = 0;
3182 unsigned ExtCost = !TLI.isExtFree(Ext);
3183 Value *PromotedOperand =
3184 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
3185 // SExt has been moved away.
3186 // Thus either it will be rematched later in the recursive calls or it is
3187 // gone. Anyway, we must not fold it into the addressing mode at this point.
3191 // addr = gep base, idx
3193 // promotedOpnd = ext opnd <- no match here
3194 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
3195 // addr = gep base, op <- match
3199 assert(PromotedOperand &&
3200 "TypePromotionHelper should have filtered out those cases");
3202 ExtAddrMode BackupAddrMode = AddrMode;
3203 unsigned OldSize = AddrModeInsts.size();
3205 if (!matchAddr(PromotedOperand, Depth) ||
3206 // The total of the new cost is equal to the cost of the created
3208 // The total of the old cost is equal to the cost of the extension plus
3209 // what we have saved in the addressing mode.
3210 !isPromotionProfitable(CreatedInstsCost,
3211 ExtCost + (AddrModeInsts.size() - OldSize),
3213 AddrMode = BackupAddrMode;
3214 AddrModeInsts.resize(OldSize);
3215 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
3216 TPT.rollback(LastKnownGood);
3225 /// If we can, try to add the value of 'Addr' into the current addressing mode.
3226 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
3227 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
3230 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
3231 // Start a transaction at this point that we will rollback if the matching
3233 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3234 TPT.getRestorationPoint();
3235 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
3236 // Fold in immediates if legal for the target.
3237 AddrMode.BaseOffs += CI->getSExtValue();
3238 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3240 AddrMode.BaseOffs -= CI->getSExtValue();
3241 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
3242 // If this is a global variable, try to fold it into the addressing mode.
3243 if (!AddrMode.BaseGV) {
3244 AddrMode.BaseGV = GV;
3245 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3247 AddrMode.BaseGV = nullptr;
3249 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
3250 ExtAddrMode BackupAddrMode = AddrMode;
3251 unsigned OldSize = AddrModeInsts.size();
3253 // Check to see if it is possible to fold this operation.
3254 bool MovedAway = false;
3255 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
3256 // This instruction may have been moved away. If so, there is nothing
3260 // Okay, it's possible to fold this. Check to see if it is actually
3261 // *profitable* to do so. We use a simple cost model to avoid increasing
3262 // register pressure too much.
3263 if (I->hasOneUse() ||
3264 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
3265 AddrModeInsts.push_back(I);
3269 // It isn't profitable to do this, roll back.
3270 //cerr << "NOT FOLDING: " << *I;
3271 AddrMode = BackupAddrMode;
3272 AddrModeInsts.resize(OldSize);
3273 TPT.rollback(LastKnownGood);
3275 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
3276 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
3278 TPT.rollback(LastKnownGood);
3279 } else if (isa<ConstantPointerNull>(Addr)) {
3280 // Null pointer gets folded without affecting the addressing mode.
3284 // Worse case, the target should support [reg] addressing modes. :)
3285 if (!AddrMode.HasBaseReg) {
3286 AddrMode.HasBaseReg = true;
3287 AddrMode.BaseReg = Addr;
3288 // Still check for legality in case the target supports [imm] but not [i+r].
3289 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3291 AddrMode.HasBaseReg = false;
3292 AddrMode.BaseReg = nullptr;
3295 // If the base register is already taken, see if we can do [r+r].
3296 if (AddrMode.Scale == 0) {
3298 AddrMode.ScaledReg = Addr;
3299 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3302 AddrMode.ScaledReg = nullptr;
3305 TPT.rollback(LastKnownGood);
3309 /// Check to see if all uses of OpVal by the specified inline asm call are due
3310 /// to memory operands. If so, return true, otherwise return false.
3311 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
3312 const TargetMachine &TM) {
3313 const Function *F = CI->getParent()->getParent();
3314 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
3315 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
3316 TargetLowering::AsmOperandInfoVector TargetConstraints =
3317 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
3318 ImmutableCallSite(CI));
3319 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3320 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3322 // Compute the constraint code and ConstraintType to use.
3323 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3325 // If this asm operand is our Value*, and if it isn't an indirect memory
3326 // operand, we can't fold it!
3327 if (OpInfo.CallOperandVal == OpVal &&
3328 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
3329 !OpInfo.isIndirect))
3336 /// Recursively walk all the uses of I until we find a memory use.
3337 /// If we find an obviously non-foldable instruction, return true.
3338 /// Add the ultimately found memory instructions to MemoryUses.
3339 static bool FindAllMemoryUses(
3341 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
3342 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
3343 // If we already considered this instruction, we're done.
3344 if (!ConsideredInsts.insert(I).second)
3347 // If this is an obviously unfoldable instruction, bail out.
3348 if (!MightBeFoldableInst(I))
3351 // Loop over all the uses, recursively processing them.
3352 for (Use &U : I->uses()) {
3353 Instruction *UserI = cast<Instruction>(U.getUser());
3355 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
3356 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
3360 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
3361 unsigned opNo = U.getOperandNo();
3362 if (opNo == 0) return true; // Storing addr, not into addr.
3363 MemoryUses.push_back(std::make_pair(SI, opNo));
3367 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
3368 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
3369 if (!IA) return true;
3371 // If this is a memory operand, we're cool, otherwise bail out.
3372 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
3377 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
3384 /// Return true if Val is already known to be live at the use site that we're
3385 /// folding it into. If so, there is no cost to include it in the addressing
3386 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
3387 /// instruction already.
3388 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
3389 Value *KnownLive2) {
3390 // If Val is either of the known-live values, we know it is live!
3391 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
3394 // All values other than instructions and arguments (e.g. constants) are live.
3395 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
3397 // If Val is a constant sized alloca in the entry block, it is live, this is
3398 // true because it is just a reference to the stack/frame pointer, which is
3399 // live for the whole function.
3400 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
3401 if (AI->isStaticAlloca())
3404 // Check to see if this value is already used in the memory instruction's
3405 // block. If so, it's already live into the block at the very least, so we
3406 // can reasonably fold it.
3407 return Val->isUsedInBasicBlock(MemoryInst->getParent());
3410 /// It is possible for the addressing mode of the machine to fold the specified
3411 /// instruction into a load or store that ultimately uses it.
3412 /// However, the specified instruction has multiple uses.
3413 /// Given this, it may actually increase register pressure to fold it
3414 /// into the load. For example, consider this code:
3418 /// use(Y) -> nonload/store
3422 /// In this case, Y has multiple uses, and can be folded into the load of Z
3423 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
3424 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
3425 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
3426 /// number of computations either.
3428 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
3429 /// X was live across 'load Z' for other reasons, we actually *would* want to
3430 /// fold the addressing mode in the Z case. This would make Y die earlier.
3431 bool AddressingModeMatcher::
3432 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
3433 ExtAddrMode &AMAfter) {
3434 if (IgnoreProfitability) return true;
3436 // AMBefore is the addressing mode before this instruction was folded into it,
3437 // and AMAfter is the addressing mode after the instruction was folded. Get
3438 // the set of registers referenced by AMAfter and subtract out those
3439 // referenced by AMBefore: this is the set of values which folding in this
3440 // address extends the lifetime of.
3442 // Note that there are only two potential values being referenced here,
3443 // BaseReg and ScaleReg (global addresses are always available, as are any
3444 // folded immediates).
3445 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
3447 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
3448 // lifetime wasn't extended by adding this instruction.
3449 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3451 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3452 ScaledReg = nullptr;
3454 // If folding this instruction (and it's subexprs) didn't extend any live
3455 // ranges, we're ok with it.
3456 if (!BaseReg && !ScaledReg)
3459 // If all uses of this instruction are ultimately load/store/inlineasm's,
3460 // check to see if their addressing modes will include this instruction. If
3461 // so, we can fold it into all uses, so it doesn't matter if it has multiple
3463 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
3464 SmallPtrSet<Instruction*, 16> ConsideredInsts;
3465 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
3466 return false; // Has a non-memory, non-foldable use!
3468 // Now that we know that all uses of this instruction are part of a chain of
3469 // computation involving only operations that could theoretically be folded
3470 // into a memory use, loop over each of these uses and see if they could
3471 // *actually* fold the instruction.
3472 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
3473 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
3474 Instruction *User = MemoryUses[i].first;
3475 unsigned OpNo = MemoryUses[i].second;
3477 // Get the access type of this use. If the use isn't a pointer, we don't
3478 // know what it accesses.
3479 Value *Address = User->getOperand(OpNo);
3480 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
3483 Type *AddressAccessTy = AddrTy->getElementType();
3484 unsigned AS = AddrTy->getAddressSpace();
3486 // Do a match against the root of this address, ignoring profitability. This
3487 // will tell us if the addressing mode for the memory operation will
3488 // *actually* cover the shared instruction.
3490 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3491 TPT.getRestorationPoint();
3492 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
3493 MemoryInst, Result, InsertedInsts,
3494 PromotedInsts, TPT);
3495 Matcher.IgnoreProfitability = true;
3496 bool Success = Matcher.matchAddr(Address, 0);
3497 (void)Success; assert(Success && "Couldn't select *anything*?");
3499 // The match was to check the profitability, the changes made are not
3500 // part of the original matcher. Therefore, they should be dropped
3501 // otherwise the original matcher will not present the right state.
3502 TPT.rollback(LastKnownGood);
3504 // If the match didn't cover I, then it won't be shared by it.
3505 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3506 I) == MatchedAddrModeInsts.end())
3509 MatchedAddrModeInsts.clear();
3515 } // end anonymous namespace
3517 /// Return true if the specified values are defined in a
3518 /// different basic block than BB.
3519 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3520 if (Instruction *I = dyn_cast<Instruction>(V))
3521 return I->getParent() != BB;
3525 /// Load and Store Instructions often have addressing modes that can do
3526 /// significant amounts of computation. As such, instruction selection will try
3527 /// to get the load or store to do as much computation as possible for the
3528 /// program. The problem is that isel can only see within a single block. As
3529 /// such, we sink as much legal addressing mode work into the block as possible.
3531 /// This method is used to optimize both load/store and inline asms with memory
3533 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3534 Type *AccessTy, unsigned AddrSpace) {
3537 // Try to collapse single-value PHI nodes. This is necessary to undo
3538 // unprofitable PRE transformations.
3539 SmallVector<Value*, 8> worklist;
3540 SmallPtrSet<Value*, 16> Visited;
3541 worklist.push_back(Addr);
3543 // Use a worklist to iteratively look through PHI nodes, and ensure that
3544 // the addressing mode obtained from the non-PHI roots of the graph
3546 Value *Consensus = nullptr;
3547 unsigned NumUsesConsensus = 0;
3548 bool IsNumUsesConsensusValid = false;
3549 SmallVector<Instruction*, 16> AddrModeInsts;
3550 ExtAddrMode AddrMode;
3551 TypePromotionTransaction TPT;
3552 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3553 TPT.getRestorationPoint();
3554 while (!worklist.empty()) {
3555 Value *V = worklist.back();
3556 worklist.pop_back();
3558 // Break use-def graph loops.
3559 if (!Visited.insert(V).second) {
3560 Consensus = nullptr;
3564 // For a PHI node, push all of its incoming values.
3565 if (PHINode *P = dyn_cast<PHINode>(V)) {
3566 for (Value *IncValue : P->incoming_values())
3567 worklist.push_back(IncValue);
3571 // For non-PHIs, determine the addressing mode being computed.
3572 SmallVector<Instruction*, 16> NewAddrModeInsts;
3573 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3574 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
3575 InsertedInsts, PromotedInsts, TPT);
3577 // This check is broken into two cases with very similar code to avoid using
3578 // getNumUses() as much as possible. Some values have a lot of uses, so
3579 // calling getNumUses() unconditionally caused a significant compile-time
3583 AddrMode = NewAddrMode;
3584 AddrModeInsts = NewAddrModeInsts;
3586 } else if (NewAddrMode == AddrMode) {
3587 if (!IsNumUsesConsensusValid) {
3588 NumUsesConsensus = Consensus->getNumUses();
3589 IsNumUsesConsensusValid = true;
3592 // Ensure that the obtained addressing mode is equivalent to that obtained
3593 // for all other roots of the PHI traversal. Also, when choosing one
3594 // such root as representative, select the one with the most uses in order
3595 // to keep the cost modeling heuristics in AddressingModeMatcher
3597 unsigned NumUses = V->getNumUses();
3598 if (NumUses > NumUsesConsensus) {
3600 NumUsesConsensus = NumUses;
3601 AddrModeInsts = NewAddrModeInsts;
3606 Consensus = nullptr;
3610 // If the addressing mode couldn't be determined, or if multiple different
3611 // ones were determined, bail out now.
3613 TPT.rollback(LastKnownGood);
3618 // Check to see if any of the instructions supersumed by this addr mode are
3619 // non-local to I's BB.
3620 bool AnyNonLocal = false;
3621 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3622 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3628 // If all the instructions matched are already in this BB, don't do anything.
3630 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3634 // Insert this computation right after this user. Since our caller is
3635 // scanning from the top of the BB to the bottom, reuse of the expr are
3636 // guaranteed to happen later.
3637 IRBuilder<> Builder(MemoryInst);
3639 // Now that we determined the addressing expression we want to use and know
3640 // that we have to sink it into this block. Check to see if we have already
3641 // done this for some other load/store instr in this block. If so, reuse the
3643 Value *&SunkAddr = SunkAddrs[Addr];
3645 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3646 << *MemoryInst << "\n");
3647 if (SunkAddr->getType() != Addr->getType())
3648 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3649 } else if (AddrSinkUsingGEPs ||
3650 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3651 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3653 // By default, we use the GEP-based method when AA is used later. This
3654 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3655 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3656 << *MemoryInst << "\n");
3657 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3658 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3660 // First, find the pointer.
3661 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3662 ResultPtr = AddrMode.BaseReg;
3663 AddrMode.BaseReg = nullptr;
3666 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3667 // We can't add more than one pointer together, nor can we scale a
3668 // pointer (both of which seem meaningless).
3669 if (ResultPtr || AddrMode.Scale != 1)
3672 ResultPtr = AddrMode.ScaledReg;
3676 if (AddrMode.BaseGV) {
3680 ResultPtr = AddrMode.BaseGV;
3683 // If the real base value actually came from an inttoptr, then the matcher
3684 // will look through it and provide only the integer value. In that case,
3686 if (!ResultPtr && AddrMode.BaseReg) {
3688 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3689 AddrMode.BaseReg = nullptr;
3690 } else if (!ResultPtr && AddrMode.Scale == 1) {
3692 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3697 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3698 SunkAddr = Constant::getNullValue(Addr->getType());
3699 } else if (!ResultPtr) {
3703 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3704 Type *I8Ty = Builder.getInt8Ty();
3706 // Start with the base register. Do this first so that subsequent address
3707 // matching finds it last, which will prevent it from trying to match it
3708 // as the scaled value in case it happens to be a mul. That would be
3709 // problematic if we've sunk a different mul for the scale, because then
3710 // we'd end up sinking both muls.
3711 if (AddrMode.BaseReg) {
3712 Value *V = AddrMode.BaseReg;
3713 if (V->getType() != IntPtrTy)
3714 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3719 // Add the scale value.
3720 if (AddrMode.Scale) {
3721 Value *V = AddrMode.ScaledReg;
3722 if (V->getType() == IntPtrTy) {
3724 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3725 cast<IntegerType>(V->getType())->getBitWidth()) {
3726 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3728 // It is only safe to sign extend the BaseReg if we know that the math
3729 // required to create it did not overflow before we extend it. Since
3730 // the original IR value was tossed in favor of a constant back when
3731 // the AddrMode was created we need to bail out gracefully if widths
3732 // do not match instead of extending it.
3733 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3734 if (I && (ResultIndex != AddrMode.BaseReg))
3735 I->eraseFromParent();
3739 if (AddrMode.Scale != 1)
3740 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3743 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3748 // Add in the Base Offset if present.
3749 if (AddrMode.BaseOffs) {
3750 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3752 // We need to add this separately from the scale above to help with
3753 // SDAG consecutive load/store merging.
3754 if (ResultPtr->getType() != I8PtrTy)
3755 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3756 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3763 SunkAddr = ResultPtr;
3765 if (ResultPtr->getType() != I8PtrTy)
3766 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3767 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3770 if (SunkAddr->getType() != Addr->getType())
3771 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3774 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3775 << *MemoryInst << "\n");
3776 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3777 Value *Result = nullptr;
3779 // Start with the base register. Do this first so that subsequent address
3780 // matching finds it last, which will prevent it from trying to match it
3781 // as the scaled value in case it happens to be a mul. That would be
3782 // problematic if we've sunk a different mul for the scale, because then
3783 // we'd end up sinking both muls.
3784 if (AddrMode.BaseReg) {
3785 Value *V = AddrMode.BaseReg;
3786 if (V->getType()->isPointerTy())
3787 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3788 if (V->getType() != IntPtrTy)
3789 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3793 // Add the scale value.
3794 if (AddrMode.Scale) {
3795 Value *V = AddrMode.ScaledReg;
3796 if (V->getType() == IntPtrTy) {
3798 } else if (V->getType()->isPointerTy()) {
3799 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3800 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3801 cast<IntegerType>(V->getType())->getBitWidth()) {
3802 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3804 // It is only safe to sign extend the BaseReg if we know that the math
3805 // required to create it did not overflow before we extend it. Since
3806 // the original IR value was tossed in favor of a constant back when
3807 // the AddrMode was created we need to bail out gracefully if widths
3808 // do not match instead of extending it.
3809 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3810 if (I && (Result != AddrMode.BaseReg))
3811 I->eraseFromParent();
3814 if (AddrMode.Scale != 1)
3815 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3818 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3823 // Add in the BaseGV if present.
3824 if (AddrMode.BaseGV) {
3825 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3827 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3832 // Add in the Base Offset if present.
3833 if (AddrMode.BaseOffs) {
3834 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3836 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3842 SunkAddr = Constant::getNullValue(Addr->getType());
3844 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3847 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3849 // If we have no uses, recursively delete the value and all dead instructions
3851 if (Repl->use_empty()) {
3852 // This can cause recursive deletion, which can invalidate our iterator.
3853 // Use a WeakVH to hold onto it in case this happens.
3854 WeakVH IterHandle(&*CurInstIterator);
3855 BasicBlock *BB = CurInstIterator->getParent();
3857 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3859 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
3860 // If the iterator instruction was recursively deleted, start over at the
3861 // start of the block.
3862 CurInstIterator = BB->begin();
3870 /// If there are any memory operands, use OptimizeMemoryInst to sink their
3871 /// address computing into the block when possible / profitable.
3872 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
3873 bool MadeChange = false;
3875 const TargetRegisterInfo *TRI =
3876 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3877 TargetLowering::AsmOperandInfoVector TargetConstraints =
3878 TLI->ParseConstraints(*DL, TRI, CS);
3880 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3881 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3883 // Compute the constraint code and ConstraintType to use.
3884 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3886 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3887 OpInfo.isIndirect) {
3888 Value *OpVal = CS->getArgOperand(ArgNo++);
3889 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
3890 } else if (OpInfo.Type == InlineAsm::isInput)
3897 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3898 /// sign extensions.
3899 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3900 assert(!Inst->use_empty() && "Input must have at least one use");
3901 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3902 bool IsSExt = isa<SExtInst>(FirstUser);
3903 Type *ExtTy = FirstUser->getType();
3904 for (const User *U : Inst->users()) {
3905 const Instruction *UI = cast<Instruction>(U);
3906 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3908 Type *CurTy = UI->getType();
3909 // Same input and output types: Same instruction after CSE.
3913 // If IsSExt is true, we are in this situation:
3915 // b = sext ty1 a to ty2
3916 // c = sext ty1 a to ty3
3917 // Assuming ty2 is shorter than ty3, this could be turned into:
3919 // b = sext ty1 a to ty2
3920 // c = sext ty2 b to ty3
3921 // However, the last sext is not free.
3925 // This is a ZExt, maybe this is free to extend from one type to another.
3926 // In that case, we would not account for a different use.
3929 if (ExtTy->getScalarType()->getIntegerBitWidth() >
3930 CurTy->getScalarType()->getIntegerBitWidth()) {
3938 if (!TLI.isZExtFree(NarrowTy, LargeTy))
3941 // All uses are the same or can be derived from one another for free.
3945 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
3946 /// load instruction.
3947 /// If an ext(load) can be formed, it is returned via \p LI for the load
3948 /// and \p Inst for the extension.
3949 /// Otherwise LI == nullptr and Inst == nullptr.
3950 /// When some promotion happened, \p TPT contains the proper state to
3953 /// \return true when promoting was necessary to expose the ext(load)
3954 /// opportunity, false otherwise.
3958 /// %ld = load i32* %addr
3959 /// %add = add nuw i32 %ld, 4
3960 /// %zext = zext i32 %add to i64
3964 /// %ld = load i32* %addr
3965 /// %zext = zext i32 %ld to i64
3966 /// %add = add nuw i64 %zext, 4
3968 /// Thanks to the promotion, we can match zext(load i32*) to i64.
3969 bool CodeGenPrepare::extLdPromotion(TypePromotionTransaction &TPT,
3970 LoadInst *&LI, Instruction *&Inst,
3971 const SmallVectorImpl<Instruction *> &Exts,
3972 unsigned CreatedInstsCost = 0) {
3973 // Iterate over all the extensions to see if one form an ext(load).
3974 for (auto I : Exts) {
3975 // Check if we directly have ext(load).
3976 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
3978 // No promotion happened here.
3981 // Check whether or not we want to do any promotion.
3982 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
3984 // Get the action to perform the promotion.
3985 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
3986 I, InsertedInsts, *TLI, PromotedInsts);
3987 // Check if we can promote.
3990 // Save the current state.
3991 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3992 TPT.getRestorationPoint();
3993 SmallVector<Instruction *, 4> NewExts;
3994 unsigned NewCreatedInstsCost = 0;
3995 unsigned ExtCost = !TLI->isExtFree(I);
3997 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
3998 &NewExts, nullptr, *TLI);
3999 assert(PromotedVal &&
4000 "TypePromotionHelper should have filtered out those cases");
4002 // We would be able to merge only one extension in a load.
4003 // Therefore, if we have more than 1 new extension we heuristically
4004 // cut this search path, because it means we degrade the code quality.
4005 // With exactly 2, the transformation is neutral, because we will merge
4006 // one extension but leave one. However, we optimistically keep going,
4007 // because the new extension may be removed too.
4008 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4009 TotalCreatedInstsCost -= ExtCost;
4010 if (!StressExtLdPromotion &&
4011 (TotalCreatedInstsCost > 1 ||
4012 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4013 // The promotion is not profitable, rollback to the previous state.
4014 TPT.rollback(LastKnownGood);
4017 // The promotion is profitable.
4018 // Check if it exposes an ext(load).
4019 (void)extLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
4020 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4021 // If we have created a new extension, i.e., now we have two
4022 // extensions. We must make sure one of them is merged with
4023 // the load, otherwise we may degrade the code quality.
4024 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
4025 // Promotion happened.
4027 // If this does not help to expose an ext(load) then, rollback.
4028 TPT.rollback(LastKnownGood);
4030 // None of the extension can form an ext(load).
4036 /// Move a zext or sext fed by a load into the same basic block as the load,
4037 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
4038 /// extend into the load.
4039 /// \p I[in/out] the extension may be modified during the process if some
4040 /// promotions apply.
4042 bool CodeGenPrepare::moveExtToFormExtLoad(Instruction *&I) {
4043 // Try to promote a chain of computation if it allows to form
4044 // an extended load.
4045 TypePromotionTransaction TPT;
4046 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4047 TPT.getRestorationPoint();
4048 SmallVector<Instruction *, 1> Exts;
4050 // Look for a load being extended.
4051 LoadInst *LI = nullptr;
4052 Instruction *OldExt = I;
4053 bool HasPromoted = extLdPromotion(TPT, LI, I, Exts);
4055 assert(!HasPromoted && !LI && "If we did not match any load instruction "
4056 "the code must remain the same");
4061 // If they're already in the same block, there's nothing to do.
4062 // Make the cheap checks first if we did not promote.
4063 // If we promoted, we need to check if it is indeed profitable.
4064 if (!HasPromoted && LI->getParent() == I->getParent())
4067 EVT VT = TLI->getValueType(*DL, I->getType());
4068 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
4070 // If the load has other users and the truncate is not free, this probably
4071 // isn't worthwhile.
4072 if (!LI->hasOneUse() && TLI &&
4073 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
4074 !TLI->isTruncateFree(I->getType(), LI->getType())) {
4076 TPT.rollback(LastKnownGood);
4080 // Check whether the target supports casts folded into loads.
4082 if (isa<ZExtInst>(I))
4083 LType = ISD::ZEXTLOAD;
4085 assert(isa<SExtInst>(I) && "Unexpected ext type!");
4086 LType = ISD::SEXTLOAD;
4088 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
4090 TPT.rollback(LastKnownGood);
4094 // Move the extend into the same block as the load, so that SelectionDAG
4097 I->removeFromParent();
4103 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
4104 BasicBlock *DefBB = I->getParent();
4106 // If the result of a {s|z}ext and its source are both live out, rewrite all
4107 // other uses of the source with result of extension.
4108 Value *Src = I->getOperand(0);
4109 if (Src->hasOneUse())
4112 // Only do this xform if truncating is free.
4113 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
4116 // Only safe to perform the optimization if the source is also defined in
4118 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
4121 bool DefIsLiveOut = false;
4122 for (User *U : I->users()) {
4123 Instruction *UI = cast<Instruction>(U);
4125 // Figure out which BB this ext is used in.
4126 BasicBlock *UserBB = UI->getParent();
4127 if (UserBB == DefBB) continue;
4128 DefIsLiveOut = true;
4134 // Make sure none of the uses are PHI nodes.
4135 for (User *U : Src->users()) {
4136 Instruction *UI = cast<Instruction>(U);
4137 BasicBlock *UserBB = UI->getParent();
4138 if (UserBB == DefBB) continue;
4139 // Be conservative. We don't want this xform to end up introducing
4140 // reloads just before load / store instructions.
4141 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
4145 // InsertedTruncs - Only insert one trunc in each block once.
4146 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
4148 bool MadeChange = false;
4149 for (Use &U : Src->uses()) {
4150 Instruction *User = cast<Instruction>(U.getUser());
4152 // Figure out which BB this ext is used in.
4153 BasicBlock *UserBB = User->getParent();
4154 if (UserBB == DefBB) continue;
4156 // Both src and def are live in this block. Rewrite the use.
4157 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
4159 if (!InsertedTrunc) {
4160 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4161 assert(InsertPt != UserBB->end());
4162 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
4163 InsertedInsts.insert(InsertedTrunc);
4166 // Replace a use of the {s|z}ext source with a use of the result.
4175 /// Check if V (an operand of a select instruction) is an expensive instruction
4176 /// that is only used once.
4177 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
4178 auto *I = dyn_cast<Instruction>(V);
4179 // If it's safe to speculatively execute, then it should not have side
4180 // effects; therefore, it's safe to sink and possibly *not* execute.
4181 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
4182 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
4185 /// Returns true if a SelectInst should be turned into an explicit branch.
4186 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
4188 // FIXME: This should use the same heuristics as IfConversion to determine
4189 // whether a select is better represented as a branch. This requires that
4190 // branch probability metadata is preserved for the select, which is not the
4193 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
4195 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
4196 // comparison condition. If the compare has more than one use, there's
4197 // probably another cmov or setcc around, so it's not worth emitting a branch.
4198 if (!Cmp || !Cmp->hasOneUse())
4201 Value *CmpOp0 = Cmp->getOperand(0);
4202 Value *CmpOp1 = Cmp->getOperand(1);
4204 // Emit "cmov on compare with a memory operand" as a branch to avoid stalls
4205 // on a load from memory. But if the load is used more than once, do not
4206 // change the select to a branch because the load is probably needed
4207 // regardless of whether the branch is taken or not.
4208 if ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
4209 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()))
4212 // If either operand of the select is expensive and only needed on one side
4213 // of the select, we should form a branch.
4214 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
4215 sinkSelectOperand(TTI, SI->getFalseValue()))
4222 /// If we have a SelectInst that will likely profit from branch prediction,
4223 /// turn it into a branch.
4224 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
4225 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
4227 // Can we convert the 'select' to CF ?
4228 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
4231 TargetLowering::SelectSupportKind SelectKind;
4233 SelectKind = TargetLowering::VectorMaskSelect;
4234 else if (SI->getType()->isVectorTy())
4235 SelectKind = TargetLowering::ScalarCondVectorVal;
4237 SelectKind = TargetLowering::ScalarValSelect;
4239 // Do we have efficient codegen support for this kind of 'selects' ?
4240 if (TLI->isSelectSupported(SelectKind)) {
4241 // We have efficient codegen support for the select instruction.
4242 // Check if it is profitable to keep this 'select'.
4243 if (!TLI->isPredictableSelectExpensive() ||
4244 !isFormingBranchFromSelectProfitable(TTI, SI))
4250 // Transform a sequence like this:
4252 // %cmp = cmp uge i32 %a, %b
4253 // %sel = select i1 %cmp, i32 %c, i32 %d
4257 // %cmp = cmp uge i32 %a, %b
4258 // br i1 %cmp, label %select.true, label %select.false
4260 // br label %select.end
4262 // br label %select.end
4264 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
4266 // In addition, we may sink instructions that produce %c or %d from
4267 // the entry block into the destination(s) of the new branch.
4268 // If the true or false blocks do not contain a sunken instruction, that
4269 // block and its branch may be optimized away. In that case, one side of the
4270 // first branch will point directly to select.end, and the corresponding PHI
4271 // predecessor block will be the start block.
4273 // First, we split the block containing the select into 2 blocks.
4274 BasicBlock *StartBlock = SI->getParent();
4275 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
4276 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
4278 // Delete the unconditional branch that was just created by the split.
4279 StartBlock->getTerminator()->eraseFromParent();
4281 // These are the new basic blocks for the conditional branch.
4282 // At least one will become an actual new basic block.
4283 BasicBlock *TrueBlock = nullptr;
4284 BasicBlock *FalseBlock = nullptr;
4286 // Sink expensive instructions into the conditional blocks to avoid executing
4287 // them speculatively.
4288 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
4289 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
4290 EndBlock->getParent(), EndBlock);
4291 auto *TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
4292 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
4293 TrueInst->moveBefore(TrueBranch);
4295 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
4296 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
4297 EndBlock->getParent(), EndBlock);
4298 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
4299 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
4300 FalseInst->moveBefore(FalseBranch);
4303 // If there was nothing to sink, then arbitrarily choose the 'false' side
4304 // for a new input value to the PHI.
4305 if (TrueBlock == FalseBlock) {
4306 assert(TrueBlock == nullptr &&
4307 "Unexpected basic block transform while optimizing select");
4309 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
4310 EndBlock->getParent(), EndBlock);
4311 BranchInst::Create(EndBlock, FalseBlock);
4314 // Insert the real conditional branch based on the original condition.
4315 // If we did not create a new block for one of the 'true' or 'false' paths
4316 // of the condition, it means that side of the branch goes to the end block
4317 // directly and the path originates from the start block from the point of
4318 // view of the new PHI.
4319 if (TrueBlock == nullptr) {
4320 BranchInst::Create(EndBlock, FalseBlock, SI->getCondition(), SI);
4321 TrueBlock = StartBlock;
4322 } else if (FalseBlock == nullptr) {
4323 BranchInst::Create(TrueBlock, EndBlock, SI->getCondition(), SI);
4324 FalseBlock = StartBlock;
4326 BranchInst::Create(TrueBlock, FalseBlock, SI->getCondition(), SI);
4329 // The select itself is replaced with a PHI Node.
4330 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
4332 PN->addIncoming(SI->getTrueValue(), TrueBlock);
4333 PN->addIncoming(SI->getFalseValue(), FalseBlock);
4335 SI->replaceAllUsesWith(PN);
4336 SI->eraseFromParent();
4338 // Instruct OptimizeBlock to skip to the next block.
4339 CurInstIterator = StartBlock->end();
4340 ++NumSelectsExpanded;
4344 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
4345 SmallVector<int, 16> Mask(SVI->getShuffleMask());
4347 for (unsigned i = 0; i < Mask.size(); ++i) {
4348 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
4350 SplatElem = Mask[i];
4356 /// Some targets have expensive vector shifts if the lanes aren't all the same
4357 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
4358 /// it's often worth sinking a shufflevector splat down to its use so that
4359 /// codegen can spot all lanes are identical.
4360 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
4361 BasicBlock *DefBB = SVI->getParent();
4363 // Only do this xform if variable vector shifts are particularly expensive.
4364 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
4367 // We only expect better codegen by sinking a shuffle if we can recognise a
4369 if (!isBroadcastShuffle(SVI))
4372 // InsertedShuffles - Only insert a shuffle in each block once.
4373 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
4375 bool MadeChange = false;
4376 for (User *U : SVI->users()) {
4377 Instruction *UI = cast<Instruction>(U);
4379 // Figure out which BB this ext is used in.
4380 BasicBlock *UserBB = UI->getParent();
4381 if (UserBB == DefBB) continue;
4383 // For now only apply this when the splat is used by a shift instruction.
4384 if (!UI->isShift()) continue;
4386 // Everything checks out, sink the shuffle if the user's block doesn't
4387 // already have a copy.
4388 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
4390 if (!InsertedShuffle) {
4391 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4392 assert(InsertPt != UserBB->end());
4394 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4395 SVI->getOperand(2), "", &*InsertPt);
4398 UI->replaceUsesOfWith(SVI, InsertedShuffle);
4402 // If we removed all uses, nuke the shuffle.
4403 if (SVI->use_empty()) {
4404 SVI->eraseFromParent();
4411 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
4415 Value *Cond = SI->getCondition();
4416 Type *OldType = Cond->getType();
4417 LLVMContext &Context = Cond->getContext();
4418 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
4419 unsigned RegWidth = RegType.getSizeInBits();
4421 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
4424 // If the register width is greater than the type width, expand the condition
4425 // of the switch instruction and each case constant to the width of the
4426 // register. By widening the type of the switch condition, subsequent
4427 // comparisons (for case comparisons) will not need to be extended to the
4428 // preferred register width, so we will potentially eliminate N-1 extends,
4429 // where N is the number of cases in the switch.
4430 auto *NewType = Type::getIntNTy(Context, RegWidth);
4432 // Zero-extend the switch condition and case constants unless the switch
4433 // condition is a function argument that is already being sign-extended.
4434 // In that case, we can avoid an unnecessary mask/extension by sign-extending
4435 // everything instead.
4436 Instruction::CastOps ExtType = Instruction::ZExt;
4437 if (auto *Arg = dyn_cast<Argument>(Cond))
4438 if (Arg->hasSExtAttr())
4439 ExtType = Instruction::SExt;
4441 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
4442 ExtInst->insertBefore(SI);
4443 SI->setCondition(ExtInst);
4444 for (SwitchInst::CaseIt Case : SI->cases()) {
4445 APInt NarrowConst = Case.getCaseValue()->getValue();
4446 APInt WideConst = (ExtType == Instruction::ZExt) ?
4447 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
4448 Case.setValue(ConstantInt::get(Context, WideConst));
4455 /// \brief Helper class to promote a scalar operation to a vector one.
4456 /// This class is used to move downward extractelement transition.
4458 /// a = vector_op <2 x i32>
4459 /// b = extractelement <2 x i32> a, i32 0
4464 /// a = vector_op <2 x i32>
4465 /// c = vector_op a (equivalent to scalar_op on the related lane)
4466 /// * d = extractelement <2 x i32> c, i32 0
4468 /// Assuming both extractelement and store can be combine, we get rid of the
4470 class VectorPromoteHelper {
4471 /// DataLayout associated with the current module.
4472 const DataLayout &DL;
4474 /// Used to perform some checks on the legality of vector operations.
4475 const TargetLowering &TLI;
4477 /// Used to estimated the cost of the promoted chain.
4478 const TargetTransformInfo &TTI;
4480 /// The transition being moved downwards.
4481 Instruction *Transition;
4482 /// The sequence of instructions to be promoted.
4483 SmallVector<Instruction *, 4> InstsToBePromoted;
4484 /// Cost of combining a store and an extract.
4485 unsigned StoreExtractCombineCost;
4486 /// Instruction that will be combined with the transition.
4487 Instruction *CombineInst;
4489 /// \brief The instruction that represents the current end of the transition.
4490 /// Since we are faking the promotion until we reach the end of the chain
4491 /// of computation, we need a way to get the current end of the transition.
4492 Instruction *getEndOfTransition() const {
4493 if (InstsToBePromoted.empty())
4495 return InstsToBePromoted.back();
4498 /// \brief Return the index of the original value in the transition.
4499 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
4500 /// c, is at index 0.
4501 unsigned getTransitionOriginalValueIdx() const {
4502 assert(isa<ExtractElementInst>(Transition) &&
4503 "Other kind of transitions are not supported yet");
4507 /// \brief Return the index of the index in the transition.
4508 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
4510 unsigned getTransitionIdx() const {
4511 assert(isa<ExtractElementInst>(Transition) &&
4512 "Other kind of transitions are not supported yet");
4516 /// \brief Get the type of the transition.
4517 /// This is the type of the original value.
4518 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
4519 /// transition is <2 x i32>.
4520 Type *getTransitionType() const {
4521 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
4524 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
4525 /// I.e., we have the following sequence:
4526 /// Def = Transition <ty1> a to <ty2>
4527 /// b = ToBePromoted <ty2> Def, ...
4529 /// b = ToBePromoted <ty1> a, ...
4530 /// Def = Transition <ty1> ToBePromoted to <ty2>
4531 void promoteImpl(Instruction *ToBePromoted);
4533 /// \brief Check whether or not it is profitable to promote all the
4534 /// instructions enqueued to be promoted.
4535 bool isProfitableToPromote() {
4536 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
4537 unsigned Index = isa<ConstantInt>(ValIdx)
4538 ? cast<ConstantInt>(ValIdx)->getZExtValue()
4540 Type *PromotedType = getTransitionType();
4542 StoreInst *ST = cast<StoreInst>(CombineInst);
4543 unsigned AS = ST->getPointerAddressSpace();
4544 unsigned Align = ST->getAlignment();
4545 // Check if this store is supported.
4546 if (!TLI.allowsMisalignedMemoryAccesses(
4547 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
4549 // If this is not supported, there is no way we can combine
4550 // the extract with the store.
4554 // The scalar chain of computation has to pay for the transition
4555 // scalar to vector.
4556 // The vector chain has to account for the combining cost.
4557 uint64_t ScalarCost =
4558 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
4559 uint64_t VectorCost = StoreExtractCombineCost;
4560 for (const auto &Inst : InstsToBePromoted) {
4561 // Compute the cost.
4562 // By construction, all instructions being promoted are arithmetic ones.
4563 // Moreover, one argument is a constant that can be viewed as a splat
4565 Value *Arg0 = Inst->getOperand(0);
4566 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
4567 isa<ConstantFP>(Arg0);
4568 TargetTransformInfo::OperandValueKind Arg0OVK =
4569 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4570 : TargetTransformInfo::OK_AnyValue;
4571 TargetTransformInfo::OperandValueKind Arg1OVK =
4572 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4573 : TargetTransformInfo::OK_AnyValue;
4574 ScalarCost += TTI.getArithmeticInstrCost(
4575 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
4576 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
4579 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
4580 << ScalarCost << "\nVector: " << VectorCost << '\n');
4581 return ScalarCost > VectorCost;
4584 /// \brief Generate a constant vector with \p Val with the same
4585 /// number of elements as the transition.
4586 /// \p UseSplat defines whether or not \p Val should be replicated
4587 /// across the whole vector.
4588 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
4589 /// otherwise we generate a vector with as many undef as possible:
4590 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
4591 /// used at the index of the extract.
4592 Value *getConstantVector(Constant *Val, bool UseSplat) const {
4593 unsigned ExtractIdx = UINT_MAX;
4595 // If we cannot determine where the constant must be, we have to
4596 // use a splat constant.
4597 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
4598 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
4599 ExtractIdx = CstVal->getSExtValue();
4604 unsigned End = getTransitionType()->getVectorNumElements();
4606 return ConstantVector::getSplat(End, Val);
4608 SmallVector<Constant *, 4> ConstVec;
4609 UndefValue *UndefVal = UndefValue::get(Val->getType());
4610 for (unsigned Idx = 0; Idx != End; ++Idx) {
4611 if (Idx == ExtractIdx)
4612 ConstVec.push_back(Val);
4614 ConstVec.push_back(UndefVal);
4616 return ConstantVector::get(ConstVec);
4619 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
4620 /// in \p Use can trigger undefined behavior.
4621 static bool canCauseUndefinedBehavior(const Instruction *Use,
4622 unsigned OperandIdx) {
4623 // This is not safe to introduce undef when the operand is on
4624 // the right hand side of a division-like instruction.
4625 if (OperandIdx != 1)
4627 switch (Use->getOpcode()) {
4630 case Instruction::SDiv:
4631 case Instruction::UDiv:
4632 case Instruction::SRem:
4633 case Instruction::URem:
4635 case Instruction::FDiv:
4636 case Instruction::FRem:
4637 return !Use->hasNoNaNs();
4639 llvm_unreachable(nullptr);
4643 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
4644 const TargetTransformInfo &TTI, Instruction *Transition,
4645 unsigned CombineCost)
4646 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
4647 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4648 assert(Transition && "Do not know how to promote null");
4651 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4652 bool canPromote(const Instruction *ToBePromoted) const {
4653 // We could support CastInst too.
4654 return isa<BinaryOperator>(ToBePromoted);
4657 /// \brief Check if it is profitable to promote \p ToBePromoted
4658 /// by moving downward the transition through.
4659 bool shouldPromote(const Instruction *ToBePromoted) const {
4660 // Promote only if all the operands can be statically expanded.
4661 // Indeed, we do not want to introduce any new kind of transitions.
4662 for (const Use &U : ToBePromoted->operands()) {
4663 const Value *Val = U.get();
4664 if (Val == getEndOfTransition()) {
4665 // If the use is a division and the transition is on the rhs,
4666 // we cannot promote the operation, otherwise we may create a
4667 // division by zero.
4668 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4672 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4673 !isa<ConstantFP>(Val))
4676 // Check that the resulting operation is legal.
4677 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4680 return StressStoreExtract ||
4681 TLI.isOperationLegalOrCustom(
4682 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
4685 /// \brief Check whether or not \p Use can be combined
4686 /// with the transition.
4687 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4688 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4690 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4691 void enqueueForPromotion(Instruction *ToBePromoted) {
4692 InstsToBePromoted.push_back(ToBePromoted);
4695 /// \brief Set the instruction that will be combined with the transition.
4696 void recordCombineInstruction(Instruction *ToBeCombined) {
4697 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4698 CombineInst = ToBeCombined;
4701 /// \brief Promote all the instructions enqueued for promotion if it is
4703 /// \return True if the promotion happened, false otherwise.
4705 // Check if there is something to promote.
4706 // Right now, if we do not have anything to combine with,
4707 // we assume the promotion is not profitable.
4708 if (InstsToBePromoted.empty() || !CombineInst)
4712 if (!StressStoreExtract && !isProfitableToPromote())
4716 for (auto &ToBePromoted : InstsToBePromoted)
4717 promoteImpl(ToBePromoted);
4718 InstsToBePromoted.clear();
4722 } // End of anonymous namespace.
4724 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4725 // At this point, we know that all the operands of ToBePromoted but Def
4726 // can be statically promoted.
4727 // For Def, we need to use its parameter in ToBePromoted:
4728 // b = ToBePromoted ty1 a
4729 // Def = Transition ty1 b to ty2
4730 // Move the transition down.
4731 // 1. Replace all uses of the promoted operation by the transition.
4732 // = ... b => = ... Def.
4733 assert(ToBePromoted->getType() == Transition->getType() &&
4734 "The type of the result of the transition does not match "
4736 ToBePromoted->replaceAllUsesWith(Transition);
4737 // 2. Update the type of the uses.
4738 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
4739 Type *TransitionTy = getTransitionType();
4740 ToBePromoted->mutateType(TransitionTy);
4741 // 3. Update all the operands of the promoted operation with promoted
4743 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
4744 for (Use &U : ToBePromoted->operands()) {
4745 Value *Val = U.get();
4746 Value *NewVal = nullptr;
4747 if (Val == Transition)
4748 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
4749 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
4750 isa<ConstantFP>(Val)) {
4751 // Use a splat constant if it is not safe to use undef.
4752 NewVal = getConstantVector(
4753 cast<Constant>(Val),
4754 isa<UndefValue>(Val) ||
4755 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
4757 llvm_unreachable("Did you modified shouldPromote and forgot to update "
4759 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
4761 Transition->removeFromParent();
4762 Transition->insertAfter(ToBePromoted);
4763 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
4766 /// Some targets can do store(extractelement) with one instruction.
4767 /// Try to push the extractelement towards the stores when the target
4768 /// has this feature and this is profitable.
4769 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
4770 unsigned CombineCost = UINT_MAX;
4771 if (DisableStoreExtract || !TLI ||
4772 (!StressStoreExtract &&
4773 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4774 Inst->getOperand(1), CombineCost)))
4777 // At this point we know that Inst is a vector to scalar transition.
4778 // Try to move it down the def-use chain, until:
4779 // - We can combine the transition with its single use
4780 // => we got rid of the transition.
4781 // - We escape the current basic block
4782 // => we would need to check that we are moving it at a cheaper place and
4783 // we do not do that for now.
4784 BasicBlock *Parent = Inst->getParent();
4785 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4786 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
4787 // If the transition has more than one use, assume this is not going to be
4789 while (Inst->hasOneUse()) {
4790 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4791 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4793 if (ToBePromoted->getParent() != Parent) {
4794 DEBUG(dbgs() << "Instruction to promote is in a different block ("
4795 << ToBePromoted->getParent()->getName()
4796 << ") than the transition (" << Parent->getName() << ").\n");
4800 if (VPH.canCombine(ToBePromoted)) {
4801 DEBUG(dbgs() << "Assume " << *Inst << '\n'
4802 << "will be combined with: " << *ToBePromoted << '\n');
4803 VPH.recordCombineInstruction(ToBePromoted);
4804 bool Changed = VPH.promote();
4805 NumStoreExtractExposed += Changed;
4809 DEBUG(dbgs() << "Try promoting.\n");
4810 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4813 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4815 VPH.enqueueForPromotion(ToBePromoted);
4816 Inst = ToBePromoted;
4821 bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
4822 // Bail out if we inserted the instruction to prevent optimizations from
4823 // stepping on each other's toes.
4824 if (InsertedInsts.count(I))
4827 if (PHINode *P = dyn_cast<PHINode>(I)) {
4828 // It is possible for very late stage optimizations (such as SimplifyCFG)
4829 // to introduce PHI nodes too late to be cleaned up. If we detect such a
4830 // trivial PHI, go ahead and zap it here.
4831 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
4832 P->replaceAllUsesWith(V);
4833 P->eraseFromParent();
4840 if (CastInst *CI = dyn_cast<CastInst>(I)) {
4841 // If the source of the cast is a constant, then this should have
4842 // already been constant folded. The only reason NOT to constant fold
4843 // it is if something (e.g. LSR) was careful to place the constant
4844 // evaluation in a block other than then one that uses it (e.g. to hoist
4845 // the address of globals out of a loop). If this is the case, we don't
4846 // want to forward-subst the cast.
4847 if (isa<Constant>(CI->getOperand(0)))
4850 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
4853 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4854 /// Sink a zext or sext into its user blocks if the target type doesn't
4855 /// fit in one register
4857 TLI->getTypeAction(CI->getContext(),
4858 TLI->getValueType(*DL, CI->getType())) ==
4859 TargetLowering::TypeExpandInteger) {
4860 return SinkCast(CI);
4862 bool MadeChange = moveExtToFormExtLoad(I);
4863 return MadeChange | optimizeExtUses(I);
4869 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4870 if (!TLI || !TLI->hasMultipleConditionRegisters())
4871 return OptimizeCmpExpression(CI);
4873 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4874 stripInvariantGroupMetadata(*LI);
4876 unsigned AS = LI->getPointerAddressSpace();
4877 return optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
4882 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4883 stripInvariantGroupMetadata(*SI);
4885 unsigned AS = SI->getPointerAddressSpace();
4886 return optimizeMemoryInst(I, SI->getOperand(1),
4887 SI->getOperand(0)->getType(), AS);
4892 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4894 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4895 BinOp->getOpcode() == Instruction::LShr)) {
4896 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4897 if (TLI && CI && TLI->hasExtractBitsInsn())
4898 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
4903 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4904 if (GEPI->hasAllZeroIndices()) {
4905 /// The GEP operand must be a pointer, so must its result -> BitCast
4906 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4907 GEPI->getName(), GEPI);
4908 GEPI->replaceAllUsesWith(NC);
4909 GEPI->eraseFromParent();
4911 optimizeInst(NC, ModifiedDT);
4917 if (CallInst *CI = dyn_cast<CallInst>(I))
4918 return optimizeCallInst(CI, ModifiedDT);
4920 if (SelectInst *SI = dyn_cast<SelectInst>(I))
4921 return optimizeSelectInst(SI);
4923 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
4924 return optimizeShuffleVectorInst(SVI);
4926 if (auto *Switch = dyn_cast<SwitchInst>(I))
4927 return optimizeSwitchInst(Switch);
4929 if (isa<ExtractElementInst>(I))
4930 return optimizeExtractElementInst(I);
4935 // In this pass we look for GEP and cast instructions that are used
4936 // across basic blocks and rewrite them to improve basic-block-at-a-time
4938 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
4940 bool MadeChange = false;
4942 CurInstIterator = BB.begin();
4943 while (CurInstIterator != BB.end()) {
4944 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
4948 MadeChange |= dupRetToEnableTailCallOpts(&BB);
4953 // llvm.dbg.value is far away from the value then iSel may not be able
4954 // handle it properly. iSel will drop llvm.dbg.value if it can not
4955 // find a node corresponding to the value.
4956 bool CodeGenPrepare::placeDbgValues(Function &F) {
4957 bool MadeChange = false;
4958 for (BasicBlock &BB : F) {
4959 Instruction *PrevNonDbgInst = nullptr;
4960 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
4961 Instruction *Insn = &*BI++;
4962 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
4963 // Leave dbg.values that refer to an alloca alone. These
4964 // instrinsics describe the address of a variable (= the alloca)
4965 // being taken. They should not be moved next to the alloca
4966 // (and to the beginning of the scope), but rather stay close to
4967 // where said address is used.
4968 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
4969 PrevNonDbgInst = Insn;
4973 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
4974 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
4975 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
4976 DVI->removeFromParent();
4977 if (isa<PHINode>(VI))
4978 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
4980 DVI->insertAfter(VI);
4989 // If there is a sequence that branches based on comparing a single bit
4990 // against zero that can be combined into a single instruction, and the
4991 // target supports folding these into a single instruction, sink the
4992 // mask and compare into the branch uses. Do this before OptimizeBlock ->
4993 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
4995 bool CodeGenPrepare::sinkAndCmp(Function &F) {
4996 if (!EnableAndCmpSinking)
4998 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
5000 bool MadeChange = false;
5001 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
5002 BasicBlock *BB = &*I++;
5004 // Does this BB end with the following?
5005 // %andVal = and %val, #single-bit-set
5006 // %icmpVal = icmp %andResult, 0
5007 // br i1 %cmpVal label %dest1, label %dest2"
5008 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
5009 if (!Brcc || !Brcc->isConditional())
5011 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
5012 if (!Cmp || Cmp->getParent() != BB)
5014 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
5015 if (!Zero || !Zero->isZero())
5017 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
5018 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
5020 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
5021 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
5023 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
5025 // Push the "and; icmp" for any users that are conditional branches.
5026 // Since there can only be one branch use per BB, we don't need to keep
5027 // track of which BBs we insert into.
5028 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
5032 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
5034 if (!BrccUser || !BrccUser->isConditional())
5036 BasicBlock *UserBB = BrccUser->getParent();
5037 if (UserBB == BB) continue;
5038 DEBUG(dbgs() << "found Brcc use\n");
5040 // Sink the "and; icmp" to use.
5042 BinaryOperator *NewAnd =
5043 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
5046 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
5050 DEBUG(BrccUser->getParent()->dump());
5056 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
5057 /// success, or returns false if no or invalid metadata was found.
5058 static bool extractBranchMetadata(BranchInst *BI,
5059 uint64_t &ProbTrue, uint64_t &ProbFalse) {
5060 assert(BI->isConditional() &&
5061 "Looking for probabilities on unconditional branch?");
5062 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
5063 if (!ProfileData || ProfileData->getNumOperands() != 3)
5066 const auto *CITrue =
5067 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
5068 const auto *CIFalse =
5069 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
5070 if (!CITrue || !CIFalse)
5073 ProbTrue = CITrue->getValue().getZExtValue();
5074 ProbFalse = CIFalse->getValue().getZExtValue();
5079 /// \brief Scale down both weights to fit into uint32_t.
5080 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
5081 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
5082 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
5083 NewTrue = NewTrue / Scale;
5084 NewFalse = NewFalse / Scale;
5087 /// \brief Some targets prefer to split a conditional branch like:
5089 /// %0 = icmp ne i32 %a, 0
5090 /// %1 = icmp ne i32 %b, 0
5091 /// %or.cond = or i1 %0, %1
5092 /// br i1 %or.cond, label %TrueBB, label %FalseBB
5094 /// into multiple branch instructions like:
5097 /// %0 = icmp ne i32 %a, 0
5098 /// br i1 %0, label %TrueBB, label %bb2
5100 /// %1 = icmp ne i32 %b, 0
5101 /// br i1 %1, label %TrueBB, label %FalseBB
5103 /// This usually allows instruction selection to do even further optimizations
5104 /// and combine the compare with the branch instruction. Currently this is
5105 /// applied for targets which have "cheap" jump instructions.
5107 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
5109 bool CodeGenPrepare::splitBranchCondition(Function &F) {
5110 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
5113 bool MadeChange = false;
5114 for (auto &BB : F) {
5115 // Does this BB end with the following?
5116 // %cond1 = icmp|fcmp|binary instruction ...
5117 // %cond2 = icmp|fcmp|binary instruction ...
5118 // %cond.or = or|and i1 %cond1, cond2
5119 // br i1 %cond.or label %dest1, label %dest2"
5120 BinaryOperator *LogicOp;
5121 BasicBlock *TBB, *FBB;
5122 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
5125 auto *Br1 = cast<BranchInst>(BB.getTerminator());
5126 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
5130 Value *Cond1, *Cond2;
5131 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
5132 m_OneUse(m_Value(Cond2)))))
5133 Opc = Instruction::And;
5134 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
5135 m_OneUse(m_Value(Cond2)))))
5136 Opc = Instruction::Or;
5140 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
5141 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
5144 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
5147 auto *InsertBefore = std::next(Function::iterator(BB))
5148 .getNodePtrUnchecked();
5149 auto TmpBB = BasicBlock::Create(BB.getContext(),
5150 BB.getName() + ".cond.split",
5151 BB.getParent(), InsertBefore);
5153 // Update original basic block by using the first condition directly by the
5154 // branch instruction and removing the no longer needed and/or instruction.
5155 Br1->setCondition(Cond1);
5156 LogicOp->eraseFromParent();
5158 // Depending on the conditon we have to either replace the true or the false
5159 // successor of the original branch instruction.
5160 if (Opc == Instruction::And)
5161 Br1->setSuccessor(0, TmpBB);
5163 Br1->setSuccessor(1, TmpBB);
5165 // Fill in the new basic block.
5166 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
5167 if (auto *I = dyn_cast<Instruction>(Cond2)) {
5168 I->removeFromParent();
5169 I->insertBefore(Br2);
5172 // Update PHI nodes in both successors. The original BB needs to be
5173 // replaced in one succesor's PHI nodes, because the branch comes now from
5174 // the newly generated BB (NewBB). In the other successor we need to add one
5175 // incoming edge to the PHI nodes, because both branch instructions target
5176 // now the same successor. Depending on the original branch condition
5177 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
5178 // we perfrom the correct update for the PHI nodes.
5179 // This doesn't change the successor order of the just created branch
5180 // instruction (or any other instruction).
5181 if (Opc == Instruction::Or)
5182 std::swap(TBB, FBB);
5184 // Replace the old BB with the new BB.
5185 for (auto &I : *TBB) {
5186 PHINode *PN = dyn_cast<PHINode>(&I);
5190 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
5191 PN->setIncomingBlock(i, TmpBB);
5194 // Add another incoming edge form the new BB.
5195 for (auto &I : *FBB) {
5196 PHINode *PN = dyn_cast<PHINode>(&I);
5199 auto *Val = PN->getIncomingValueForBlock(&BB);
5200 PN->addIncoming(Val, TmpBB);
5203 // Update the branch weights (from SelectionDAGBuilder::
5204 // FindMergedConditions).
5205 if (Opc == Instruction::Or) {
5206 // Codegen X | Y as:
5215 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
5216 // The requirement is that
5217 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
5218 // = TrueProb for orignal BB.
5219 // Assuming the orignal weights are A and B, one choice is to set BB1's
5220 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
5222 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
5223 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
5224 // TmpBB, but the math is more complicated.
5225 uint64_t TrueWeight, FalseWeight;
5226 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
5227 uint64_t NewTrueWeight = TrueWeight;
5228 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
5229 scaleWeights(NewTrueWeight, NewFalseWeight);
5230 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
5231 .createBranchWeights(TrueWeight, FalseWeight));
5233 NewTrueWeight = TrueWeight;
5234 NewFalseWeight = 2 * FalseWeight;
5235 scaleWeights(NewTrueWeight, NewFalseWeight);
5236 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
5237 .createBranchWeights(TrueWeight, FalseWeight));
5240 // Codegen X & Y as:
5248 // This requires creation of TmpBB after CurBB.
5250 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
5251 // The requirement is that
5252 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
5253 // = FalseProb for orignal BB.
5254 // Assuming the orignal weights are A and B, one choice is to set BB1's
5255 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
5257 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
5258 uint64_t TrueWeight, FalseWeight;
5259 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
5260 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
5261 uint64_t NewFalseWeight = FalseWeight;
5262 scaleWeights(NewTrueWeight, NewFalseWeight);
5263 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
5264 .createBranchWeights(TrueWeight, FalseWeight));
5266 NewTrueWeight = 2 * TrueWeight;
5267 NewFalseWeight = FalseWeight;
5268 scaleWeights(NewTrueWeight, NewFalseWeight);
5269 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
5270 .createBranchWeights(TrueWeight, FalseWeight));
5274 // Note: No point in getting fancy here, since the DT info is never
5275 // available to CodeGenPrepare.
5280 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
5286 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
5287 if (auto *InvariantMD = I.getMetadata(LLVMContext::MD_invariant_group))
5288 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());