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/TargetTransformInfo.h"
22 #include "llvm/IR/CallSite.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/DerivedTypes.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/GetElementPtrTypeIterator.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/InlineAsm.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/PatternMatch.h"
34 #include "llvm/IR/ValueHandle.h"
35 #include "llvm/IR/ValueMap.h"
36 #include "llvm/Pass.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/Target/TargetLibraryInfo.h"
41 #include "llvm/Target/TargetLowering.h"
42 #include "llvm/Target/TargetSubtargetInfo.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/Transforms/Utils/BuildLibCalls.h"
45 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
46 #include "llvm/Transforms/Utils/Local.h"
48 using namespace llvm::PatternMatch;
50 #define DEBUG_TYPE "codegenprepare"
52 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
53 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
54 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
55 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
57 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
59 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
60 "computations were sunk");
61 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
62 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
63 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
64 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
65 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
66 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
67 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
69 static cl::opt<bool> DisableBranchOpts(
70 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
71 cl::desc("Disable branch optimizations in CodeGenPrepare"));
73 static cl::opt<bool> DisableSelectToBranch(
74 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
75 cl::desc("Disable select to branch conversion."));
77 static cl::opt<bool> AddrSinkUsingGEPs(
78 "addr-sink-using-gep", cl::Hidden, cl::init(false),
79 cl::desc("Address sinking in CGP using GEPs."));
81 static cl::opt<bool> EnableAndCmpSinking(
82 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
83 cl::desc("Enable sinkinig and/cmp into branches."));
85 static cl::opt<bool> DisableStoreExtract(
86 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
87 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
89 static cl::opt<bool> StressStoreExtract(
90 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
91 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
94 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
95 typedef DenseMap<Instruction *, Type *> InstrToOrigTy;
97 class CodeGenPrepare : public FunctionPass {
98 /// TLI - Keep a pointer of a TargetLowering to consult for determining
99 /// transformation profitability.
100 const TargetMachine *TM;
101 const TargetLowering *TLI;
102 const TargetTransformInfo *TTI;
103 const TargetLibraryInfo *TLInfo;
106 /// CurInstIterator - As we scan instructions optimizing them, this is the
107 /// next instruction to optimize. Xforms that can invalidate this should
109 BasicBlock::iterator CurInstIterator;
111 /// Keeps track of non-local addresses that have been sunk into a block.
112 /// This allows us to avoid inserting duplicate code for blocks with
113 /// multiple load/stores of the same address.
114 ValueMap<Value*, Value*> SunkAddrs;
116 /// Keeps track of all truncates inserted for the current function.
117 SetOfInstrs InsertedTruncsSet;
118 /// Keeps track of the type of the related instruction before their
119 /// promotion for the current function.
120 InstrToOrigTy PromotedInsts;
122 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
126 /// OptSize - True if optimizing for size.
130 static char ID; // Pass identification, replacement for typeid
131 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
132 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr) {
133 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
135 bool runOnFunction(Function &F) override;
137 const char *getPassName() const override { return "CodeGen Prepare"; }
139 void getAnalysisUsage(AnalysisUsage &AU) const override {
140 AU.addPreserved<DominatorTreeWrapperPass>();
141 AU.addRequired<TargetLibraryInfo>();
142 AU.addRequired<TargetTransformInfo>();
146 bool EliminateFallThrough(Function &F);
147 bool EliminateMostlyEmptyBlocks(Function &F);
148 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
149 void EliminateMostlyEmptyBlock(BasicBlock *BB);
150 bool OptimizeBlock(BasicBlock &BB);
151 bool OptimizeInst(Instruction *I);
152 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
153 bool OptimizeInlineAsmInst(CallInst *CS);
154 bool OptimizeCallInst(CallInst *CI);
155 bool MoveExtToFormExtLoad(Instruction *I);
156 bool OptimizeExtUses(Instruction *I);
157 bool OptimizeSelectInst(SelectInst *SI);
158 bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
159 bool OptimizeExtractElementInst(Instruction *Inst);
160 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
161 bool PlaceDbgValues(Function &F);
162 bool sinkAndCmp(Function &F);
166 char CodeGenPrepare::ID = 0;
167 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
168 "Optimize for code generation", false, false)
170 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
171 return new CodeGenPrepare(TM);
174 bool CodeGenPrepare::runOnFunction(Function &F) {
175 if (skipOptnoneFunction(F))
178 bool EverMadeChange = false;
179 // Clear per function information.
180 InsertedTruncsSet.clear();
181 PromotedInsts.clear();
185 TLI = TM->getSubtargetImpl()->getTargetLowering();
186 TLInfo = &getAnalysis<TargetLibraryInfo>();
187 TTI = &getAnalysis<TargetTransformInfo>();
188 DominatorTreeWrapperPass *DTWP =
189 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
190 DT = DTWP ? &DTWP->getDomTree() : nullptr;
191 OptSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
192 Attribute::OptimizeForSize);
194 /// This optimization identifies DIV instructions that can be
195 /// profitably bypassed and carried out with a shorter, faster divide.
196 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
197 const DenseMap<unsigned int, unsigned int> &BypassWidths =
198 TLI->getBypassSlowDivWidths();
199 for (Function::iterator I = F.begin(); I != F.end(); I++)
200 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
203 // Eliminate blocks that contain only PHI nodes and an
204 // unconditional branch.
205 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
207 // llvm.dbg.value is far away from the value then iSel may not be able
208 // handle it properly. iSel will drop llvm.dbg.value if it can not
209 // find a node corresponding to the value.
210 EverMadeChange |= PlaceDbgValues(F);
212 // If there is a mask, compare against zero, and branch that can be combined
213 // into a single target instruction, push the mask and compare into branch
214 // users. Do this before OptimizeBlock -> OptimizeInst ->
215 // OptimizeCmpExpression, which perturbs the pattern being searched for.
216 if (!DisableBranchOpts)
217 EverMadeChange |= sinkAndCmp(F);
219 bool MadeChange = true;
222 for (Function::iterator I = F.begin(); I != F.end(); ) {
223 BasicBlock *BB = I++;
224 MadeChange |= OptimizeBlock(*BB);
226 EverMadeChange |= MadeChange;
231 if (!DisableBranchOpts) {
233 SmallPtrSet<BasicBlock*, 8> WorkList;
234 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
235 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
236 MadeChange |= ConstantFoldTerminator(BB, true);
237 if (!MadeChange) continue;
239 for (SmallVectorImpl<BasicBlock*>::iterator
240 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
241 if (pred_begin(*II) == pred_end(*II))
242 WorkList.insert(*II);
245 // Delete the dead blocks and any of their dead successors.
246 MadeChange |= !WorkList.empty();
247 while (!WorkList.empty()) {
248 BasicBlock *BB = *WorkList.begin();
250 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
254 for (SmallVectorImpl<BasicBlock*>::iterator
255 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
256 if (pred_begin(*II) == pred_end(*II))
257 WorkList.insert(*II);
260 // Merge pairs of basic blocks with unconditional branches, connected by
262 if (EverMadeChange || MadeChange)
263 MadeChange |= EliminateFallThrough(F);
267 EverMadeChange |= MadeChange;
270 if (ModifiedDT && DT)
273 return EverMadeChange;
276 /// EliminateFallThrough - Merge basic blocks which are connected
277 /// by a single edge, where one of the basic blocks has a single successor
278 /// pointing to the other basic block, which has a single predecessor.
279 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
280 bool Changed = false;
281 // Scan all of the blocks in the function, except for the entry block.
282 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
283 BasicBlock *BB = I++;
284 // If the destination block has a single pred, then this is a trivial
285 // edge, just collapse it.
286 BasicBlock *SinglePred = BB->getSinglePredecessor();
288 // Don't merge if BB's address is taken.
289 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
291 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
292 if (Term && !Term->isConditional()) {
294 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
295 // Remember if SinglePred was the entry block of the function.
296 // If so, we will need to move BB back to the entry position.
297 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
298 MergeBasicBlockIntoOnlyPred(BB, this);
300 if (isEntry && BB != &BB->getParent()->getEntryBlock())
301 BB->moveBefore(&BB->getParent()->getEntryBlock());
303 // We have erased a block. Update the iterator.
310 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
311 /// debug info directives, and an unconditional branch. Passes before isel
312 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
313 /// isel. Start by eliminating these blocks so we can split them the way we
315 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
316 bool MadeChange = false;
317 // Note that this intentionally skips the entry block.
318 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
319 BasicBlock *BB = I++;
321 // If this block doesn't end with an uncond branch, ignore it.
322 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
323 if (!BI || !BI->isUnconditional())
326 // If the instruction before the branch (skipping debug info) isn't a phi
327 // node, then other stuff is happening here.
328 BasicBlock::iterator BBI = BI;
329 if (BBI != BB->begin()) {
331 while (isa<DbgInfoIntrinsic>(BBI)) {
332 if (BBI == BB->begin())
336 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
340 // Do not break infinite loops.
341 BasicBlock *DestBB = BI->getSuccessor(0);
345 if (!CanMergeBlocks(BB, DestBB))
348 EliminateMostlyEmptyBlock(BB);
354 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
355 /// single uncond branch between them, and BB contains no other non-phi
357 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
358 const BasicBlock *DestBB) const {
359 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
360 // the successor. If there are more complex condition (e.g. preheaders),
361 // don't mess around with them.
362 BasicBlock::const_iterator BBI = BB->begin();
363 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
364 for (const User *U : PN->users()) {
365 const Instruction *UI = cast<Instruction>(U);
366 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
368 // If User is inside DestBB block and it is a PHINode then check
369 // incoming value. If incoming value is not from BB then this is
370 // a complex condition (e.g. preheaders) we want to avoid here.
371 if (UI->getParent() == DestBB) {
372 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
373 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
374 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
375 if (Insn && Insn->getParent() == BB &&
376 Insn->getParent() != UPN->getIncomingBlock(I))
383 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
384 // and DestBB may have conflicting incoming values for the block. If so, we
385 // can't merge the block.
386 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
387 if (!DestBBPN) return true; // no conflict.
389 // Collect the preds of BB.
390 SmallPtrSet<const BasicBlock*, 16> BBPreds;
391 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
392 // It is faster to get preds from a PHI than with pred_iterator.
393 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
394 BBPreds.insert(BBPN->getIncomingBlock(i));
396 BBPreds.insert(pred_begin(BB), pred_end(BB));
399 // Walk the preds of DestBB.
400 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
401 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
402 if (BBPreds.count(Pred)) { // Common predecessor?
403 BBI = DestBB->begin();
404 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
405 const Value *V1 = PN->getIncomingValueForBlock(Pred);
406 const Value *V2 = PN->getIncomingValueForBlock(BB);
408 // If V2 is a phi node in BB, look up what the mapped value will be.
409 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
410 if (V2PN->getParent() == BB)
411 V2 = V2PN->getIncomingValueForBlock(Pred);
413 // If there is a conflict, bail out.
414 if (V1 != V2) return false;
423 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
424 /// an unconditional branch in it.
425 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
426 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
427 BasicBlock *DestBB = BI->getSuccessor(0);
429 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
431 // If the destination block has a single pred, then this is a trivial edge,
433 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
434 if (SinglePred != DestBB) {
435 // Remember if SinglePred was the entry block of the function. If so, we
436 // will need to move BB back to the entry position.
437 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
438 MergeBasicBlockIntoOnlyPred(DestBB, this);
440 if (isEntry && BB != &BB->getParent()->getEntryBlock())
441 BB->moveBefore(&BB->getParent()->getEntryBlock());
443 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
448 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
449 // to handle the new incoming edges it is about to have.
451 for (BasicBlock::iterator BBI = DestBB->begin();
452 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
453 // Remove the incoming value for BB, and remember it.
454 Value *InVal = PN->removeIncomingValue(BB, false);
456 // Two options: either the InVal is a phi node defined in BB or it is some
457 // value that dominates BB.
458 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
459 if (InValPhi && InValPhi->getParent() == BB) {
460 // Add all of the input values of the input PHI as inputs of this phi.
461 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
462 PN->addIncoming(InValPhi->getIncomingValue(i),
463 InValPhi->getIncomingBlock(i));
465 // Otherwise, add one instance of the dominating value for each edge that
466 // we will be adding.
467 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
468 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
469 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
471 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
472 PN->addIncoming(InVal, *PI);
477 // The PHIs are now updated, change everything that refers to BB to use
478 // DestBB and remove BB.
479 BB->replaceAllUsesWith(DestBB);
480 if (DT && !ModifiedDT) {
481 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
482 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
483 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
484 DT->changeImmediateDominator(DestBB, NewIDom);
487 BB->eraseFromParent();
490 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
493 /// SinkCast - Sink the specified cast instruction into its user blocks
494 static bool SinkCast(CastInst *CI) {
495 BasicBlock *DefBB = CI->getParent();
497 /// InsertedCasts - Only insert a cast in each block once.
498 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
500 bool MadeChange = false;
501 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
503 Use &TheUse = UI.getUse();
504 Instruction *User = cast<Instruction>(*UI);
506 // Figure out which BB this cast is used in. For PHI's this is the
507 // appropriate predecessor block.
508 BasicBlock *UserBB = User->getParent();
509 if (PHINode *PN = dyn_cast<PHINode>(User)) {
510 UserBB = PN->getIncomingBlock(TheUse);
513 // Preincrement use iterator so we don't invalidate it.
516 // If this user is in the same block as the cast, don't change the cast.
517 if (UserBB == DefBB) continue;
519 // If we have already inserted a cast into this block, use it.
520 CastInst *&InsertedCast = InsertedCasts[UserBB];
523 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
525 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
530 // Replace a use of the cast with a use of the new cast.
531 TheUse = InsertedCast;
535 // If we removed all uses, nuke the cast.
536 if (CI->use_empty()) {
537 CI->eraseFromParent();
544 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
545 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
546 /// sink it into user blocks to reduce the number of virtual
547 /// registers that must be created and coalesced.
549 /// Return true if any changes are made.
551 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
552 // If this is a noop copy,
553 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
554 EVT DstVT = TLI.getValueType(CI->getType());
556 // This is an fp<->int conversion?
557 if (SrcVT.isInteger() != DstVT.isInteger())
560 // If this is an extension, it will be a zero or sign extension, which
562 if (SrcVT.bitsLT(DstVT)) return false;
564 // If these values will be promoted, find out what they will be promoted
565 // to. This helps us consider truncates on PPC as noop copies when they
567 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
568 TargetLowering::TypePromoteInteger)
569 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
570 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
571 TargetLowering::TypePromoteInteger)
572 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
574 // If, after promotion, these are the same types, this is a noop copy.
581 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
582 /// the number of virtual registers that must be created and coalesced. This is
583 /// a clear win except on targets with multiple condition code registers
584 /// (PowerPC), where it might lose; some adjustment may be wanted there.
586 /// Return true if any changes are made.
587 static bool OptimizeCmpExpression(CmpInst *CI) {
588 BasicBlock *DefBB = CI->getParent();
590 /// InsertedCmp - Only insert a cmp in each block once.
591 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
593 bool MadeChange = false;
594 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
596 Use &TheUse = UI.getUse();
597 Instruction *User = cast<Instruction>(*UI);
599 // Preincrement use iterator so we don't invalidate it.
602 // Don't bother for PHI nodes.
603 if (isa<PHINode>(User))
606 // Figure out which BB this cmp is used in.
607 BasicBlock *UserBB = User->getParent();
609 // If this user is in the same block as the cmp, don't change the cmp.
610 if (UserBB == DefBB) continue;
612 // If we have already inserted a cmp into this block, use it.
613 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
616 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
618 CmpInst::Create(CI->getOpcode(),
619 CI->getPredicate(), CI->getOperand(0),
620 CI->getOperand(1), "", InsertPt);
624 // Replace a use of the cmp with a use of the new cmp.
625 TheUse = InsertedCmp;
629 // If we removed all uses, nuke the cmp.
631 CI->eraseFromParent();
636 /// isExtractBitsCandidateUse - Check if the candidates could
637 /// be combined with shift instruction, which includes:
638 /// 1. Truncate instruction
639 /// 2. And instruction and the imm is a mask of the low bits:
640 /// imm & (imm+1) == 0
641 static bool isExtractBitsCandidateUse(Instruction *User) {
642 if (!isa<TruncInst>(User)) {
643 if (User->getOpcode() != Instruction::And ||
644 !isa<ConstantInt>(User->getOperand(1)))
647 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
649 if ((Cimm & (Cimm + 1)).getBoolValue())
655 /// SinkShiftAndTruncate - sink both shift and truncate instruction
656 /// to the use of truncate's BB.
658 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
659 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
660 const TargetLowering &TLI) {
661 BasicBlock *UserBB = User->getParent();
662 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
663 TruncInst *TruncI = dyn_cast<TruncInst>(User);
664 bool MadeChange = false;
666 for (Value::user_iterator TruncUI = TruncI->user_begin(),
667 TruncE = TruncI->user_end();
668 TruncUI != TruncE;) {
670 Use &TruncTheUse = TruncUI.getUse();
671 Instruction *TruncUser = cast<Instruction>(*TruncUI);
672 // Preincrement use iterator so we don't invalidate it.
676 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
680 // If the use is actually a legal node, there will not be an
681 // implicit truncate.
682 // FIXME: always querying the result type is just an
683 // approximation; some nodes' legality is determined by the
684 // operand or other means. There's no good way to find out though.
685 if (TLI.isOperationLegalOrCustom(ISDOpcode,
686 EVT::getEVT(TruncUser->getType(), true)))
689 // Don't bother for PHI nodes.
690 if (isa<PHINode>(TruncUser))
693 BasicBlock *TruncUserBB = TruncUser->getParent();
695 if (UserBB == TruncUserBB)
698 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
699 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
701 if (!InsertedShift && !InsertedTrunc) {
702 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
704 if (ShiftI->getOpcode() == Instruction::AShr)
706 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
709 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
712 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
715 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
716 TruncI->getType(), "", TruncInsertPt);
720 TruncTheUse = InsertedTrunc;
726 /// OptimizeExtractBits - sink the shift *right* instruction into user blocks if
727 /// the uses could potentially be combined with this shift instruction and
728 /// generate BitExtract instruction. It will only be applied if the architecture
729 /// supports BitExtract instruction. Here is an example:
731 /// %x.extract.shift = lshr i64 %arg1, 32
733 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
737 /// %x.extract.shift.1 = lshr i64 %arg1, 32
738 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
740 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
742 /// Return true if any changes are made.
743 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
744 const TargetLowering &TLI) {
745 BasicBlock *DefBB = ShiftI->getParent();
747 /// Only insert instructions in each block once.
748 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
750 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(ShiftI->getType()));
752 bool MadeChange = false;
753 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
755 Use &TheUse = UI.getUse();
756 Instruction *User = cast<Instruction>(*UI);
757 // Preincrement use iterator so we don't invalidate it.
760 // Don't bother for PHI nodes.
761 if (isa<PHINode>(User))
764 if (!isExtractBitsCandidateUse(User))
767 BasicBlock *UserBB = User->getParent();
769 if (UserBB == DefBB) {
770 // If the shift and truncate instruction are in the same BB. The use of
771 // the truncate(TruncUse) may still introduce another truncate if not
772 // legal. In this case, we would like to sink both shift and truncate
773 // instruction to the BB of TruncUse.
776 // i64 shift.result = lshr i64 opnd, imm
777 // trunc.result = trunc shift.result to i16
780 // ----> We will have an implicit truncate here if the architecture does
781 // not have i16 compare.
782 // cmp i16 trunc.result, opnd2
784 if (isa<TruncInst>(User) && shiftIsLegal
785 // If the type of the truncate is legal, no trucate will be
786 // introduced in other basic blocks.
787 && (!TLI.isTypeLegal(TLI.getValueType(User->getType()))))
789 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI);
793 // If we have already inserted a shift into this block, use it.
794 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
796 if (!InsertedShift) {
797 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
799 if (ShiftI->getOpcode() == Instruction::AShr)
801 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
804 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
809 // Replace a use of the shift with a use of the new shift.
810 TheUse = InsertedShift;
813 // If we removed all uses, nuke the shift.
814 if (ShiftI->use_empty())
815 ShiftI->eraseFromParent();
821 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
823 void replaceCall(Value *With) override {
824 CI->replaceAllUsesWith(With);
825 CI->eraseFromParent();
827 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const override {
828 if (ConstantInt *SizeCI =
829 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
830 return SizeCI->isAllOnesValue();
834 } // end anonymous namespace
836 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
837 BasicBlock *BB = CI->getParent();
839 // Lower inline assembly if we can.
840 // If we found an inline asm expession, and if the target knows how to
841 // lower it to normal LLVM code, do so now.
842 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
843 if (TLI->ExpandInlineAsm(CI)) {
844 // Avoid invalidating the iterator.
845 CurInstIterator = BB->begin();
846 // Avoid processing instructions out of order, which could cause
847 // reuse before a value is defined.
851 // Sink address computing for memory operands into the block.
852 if (OptimizeInlineAsmInst(CI))
856 // Lower all uses of llvm.objectsize.*
857 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
858 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
859 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
860 Type *ReturnTy = CI->getType();
861 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
863 // Substituting this can cause recursive simplifications, which can
864 // invalidate our iterator. Use a WeakVH to hold onto it in case this
866 WeakVH IterHandle(CurInstIterator);
868 replaceAndRecursivelySimplify(CI, RetVal,
869 TLI ? TLI->getDataLayout() : nullptr,
870 TLInfo, ModifiedDT ? nullptr : DT);
872 // If the iterator instruction was recursively deleted, start over at the
873 // start of the block.
874 if (IterHandle != CurInstIterator) {
875 CurInstIterator = BB->begin();
882 SmallVector<Value*, 2> PtrOps;
884 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
885 while (!PtrOps.empty())
886 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
890 // From here on out we're working with named functions.
891 if (!CI->getCalledFunction()) return false;
893 // We'll need DataLayout from here on out.
894 const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
895 if (!TD) return false;
897 // Lower all default uses of _chk calls. This is very similar
898 // to what InstCombineCalls does, but here we are only lowering calls
899 // that have the default "don't know" as the objectsize. Anything else
900 // should be left alone.
901 CodeGenPrepareFortifiedLibCalls Simplifier;
902 return Simplifier.fold(CI, TD, TLInfo);
905 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
906 /// instructions to the predecessor to enable tail call optimizations. The
907 /// case it is currently looking for is:
910 /// %tmp0 = tail call i32 @f0()
913 /// %tmp1 = tail call i32 @f1()
916 /// %tmp2 = tail call i32 @f2()
919 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
927 /// %tmp0 = tail call i32 @f0()
930 /// %tmp1 = tail call i32 @f1()
933 /// %tmp2 = tail call i32 @f2()
936 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
940 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
944 PHINode *PN = nullptr;
945 BitCastInst *BCI = nullptr;
946 Value *V = RI->getReturnValue();
948 BCI = dyn_cast<BitCastInst>(V);
950 V = BCI->getOperand(0);
952 PN = dyn_cast<PHINode>(V);
957 if (PN && PN->getParent() != BB)
960 // It's not safe to eliminate the sign / zero extension of the return value.
961 // See llvm::isInTailCallPosition().
962 const Function *F = BB->getParent();
963 AttributeSet CallerAttrs = F->getAttributes();
964 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
965 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
968 // Make sure there are no instructions between the PHI and return, or that the
969 // return is the first instruction in the block.
971 BasicBlock::iterator BI = BB->begin();
972 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
974 // Also skip over the bitcast.
979 BasicBlock::iterator BI = BB->begin();
980 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
985 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
987 SmallVector<CallInst*, 4> TailCalls;
989 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
990 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
991 // Make sure the phi value is indeed produced by the tail call.
992 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
993 TLI->mayBeEmittedAsTailCall(CI))
994 TailCalls.push_back(CI);
997 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
998 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
999 if (!VisitedBBs.insert(*PI))
1002 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1003 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1004 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1005 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1009 CallInst *CI = dyn_cast<CallInst>(&*RI);
1010 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1011 TailCalls.push_back(CI);
1015 bool Changed = false;
1016 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1017 CallInst *CI = TailCalls[i];
1020 // Conservatively require the attributes of the call to match those of the
1021 // return. Ignore noalias because it doesn't affect the call sequence.
1022 AttributeSet CalleeAttrs = CS.getAttributes();
1023 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1024 removeAttribute(Attribute::NoAlias) !=
1025 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1026 removeAttribute(Attribute::NoAlias))
1029 // Make sure the call instruction is followed by an unconditional branch to
1030 // the return block.
1031 BasicBlock *CallBB = CI->getParent();
1032 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1033 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1036 // Duplicate the return into CallBB.
1037 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1038 ModifiedDT = Changed = true;
1042 // If we eliminated all predecessors of the block, delete the block now.
1043 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1044 BB->eraseFromParent();
1049 //===----------------------------------------------------------------------===//
1050 // Memory Optimization
1051 //===----------------------------------------------------------------------===//
1055 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1056 /// which holds actual Value*'s for register values.
1057 struct ExtAddrMode : public TargetLowering::AddrMode {
1060 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1061 void print(raw_ostream &OS) const;
1064 bool operator==(const ExtAddrMode& O) const {
1065 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1066 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1067 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1072 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1078 void ExtAddrMode::print(raw_ostream &OS) const {
1079 bool NeedPlus = false;
1082 OS << (NeedPlus ? " + " : "")
1084 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1089 OS << (NeedPlus ? " + " : "")
1095 OS << (NeedPlus ? " + " : "")
1097 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1101 OS << (NeedPlus ? " + " : "")
1103 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1109 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1110 void ExtAddrMode::dump() const {
1116 /// \brief This class provides transaction based operation on the IR.
1117 /// Every change made through this class is recorded in the internal state and
1118 /// can be undone (rollback) until commit is called.
1119 class TypePromotionTransaction {
1121 /// \brief This represents the common interface of the individual transaction.
1122 /// Each class implements the logic for doing one specific modification on
1123 /// the IR via the TypePromotionTransaction.
1124 class TypePromotionAction {
1126 /// The Instruction modified.
1130 /// \brief Constructor of the action.
1131 /// The constructor performs the related action on the IR.
1132 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1134 virtual ~TypePromotionAction() {}
1136 /// \brief Undo the modification done by this action.
1137 /// When this method is called, the IR must be in the same state as it was
1138 /// before this action was applied.
1139 /// \pre Undoing the action works if and only if the IR is in the exact same
1140 /// state as it was directly after this action was applied.
1141 virtual void undo() = 0;
1143 /// \brief Advocate every change made by this action.
1144 /// When the results on the IR of the action are to be kept, it is important
1145 /// to call this function, otherwise hidden information may be kept forever.
1146 virtual void commit() {
1147 // Nothing to be done, this action is not doing anything.
1151 /// \brief Utility to remember the position of an instruction.
1152 class InsertionHandler {
1153 /// Position of an instruction.
1154 /// Either an instruction:
1155 /// - Is the first in a basic block: BB is used.
1156 /// - Has a previous instructon: PrevInst is used.
1158 Instruction *PrevInst;
1161 /// Remember whether or not the instruction had a previous instruction.
1162 bool HasPrevInstruction;
1165 /// \brief Record the position of \p Inst.
1166 InsertionHandler(Instruction *Inst) {
1167 BasicBlock::iterator It = Inst;
1168 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1169 if (HasPrevInstruction)
1170 Point.PrevInst = --It;
1172 Point.BB = Inst->getParent();
1175 /// \brief Insert \p Inst at the recorded position.
1176 void insert(Instruction *Inst) {
1177 if (HasPrevInstruction) {
1178 if (Inst->getParent())
1179 Inst->removeFromParent();
1180 Inst->insertAfter(Point.PrevInst);
1182 Instruction *Position = Point.BB->getFirstInsertionPt();
1183 if (Inst->getParent())
1184 Inst->moveBefore(Position);
1186 Inst->insertBefore(Position);
1191 /// \brief Move an instruction before another.
1192 class InstructionMoveBefore : public TypePromotionAction {
1193 /// Original position of the instruction.
1194 InsertionHandler Position;
1197 /// \brief Move \p Inst before \p Before.
1198 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1199 : TypePromotionAction(Inst), Position(Inst) {
1200 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1201 Inst->moveBefore(Before);
1204 /// \brief Move the instruction back to its original position.
1205 void undo() override {
1206 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1207 Position.insert(Inst);
1211 /// \brief Set the operand of an instruction with a new value.
1212 class OperandSetter : public TypePromotionAction {
1213 /// Original operand of the instruction.
1215 /// Index of the modified instruction.
1219 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1220 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1221 : TypePromotionAction(Inst), Idx(Idx) {
1222 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1223 << "for:" << *Inst << "\n"
1224 << "with:" << *NewVal << "\n");
1225 Origin = Inst->getOperand(Idx);
1226 Inst->setOperand(Idx, NewVal);
1229 /// \brief Restore the original value of the instruction.
1230 void undo() override {
1231 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1232 << "for: " << *Inst << "\n"
1233 << "with: " << *Origin << "\n");
1234 Inst->setOperand(Idx, Origin);
1238 /// \brief Hide the operands of an instruction.
1239 /// Do as if this instruction was not using any of its operands.
1240 class OperandsHider : public TypePromotionAction {
1241 /// The list of original operands.
1242 SmallVector<Value *, 4> OriginalValues;
1245 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1246 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1247 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1248 unsigned NumOpnds = Inst->getNumOperands();
1249 OriginalValues.reserve(NumOpnds);
1250 for (unsigned It = 0; It < NumOpnds; ++It) {
1251 // Save the current operand.
1252 Value *Val = Inst->getOperand(It);
1253 OriginalValues.push_back(Val);
1255 // We could use OperandSetter here, but that would implied an overhead
1256 // that we are not willing to pay.
1257 Inst->setOperand(It, UndefValue::get(Val->getType()));
1261 /// \brief Restore the original list of uses.
1262 void undo() override {
1263 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1264 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1265 Inst->setOperand(It, OriginalValues[It]);
1269 /// \brief Build a truncate instruction.
1270 class TruncBuilder : public TypePromotionAction {
1273 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1275 /// trunc Opnd to Ty.
1276 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1277 IRBuilder<> Builder(Opnd);
1278 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
1279 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
1282 /// \brief Get the built value.
1283 Value *getBuiltValue() { return Val; }
1285 /// \brief Remove the built instruction.
1286 void undo() override {
1287 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
1288 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1289 IVal->eraseFromParent();
1293 /// \brief Build a sign extension instruction.
1294 class SExtBuilder : public TypePromotionAction {
1297 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1299 /// sext Opnd to Ty.
1300 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1301 : TypePromotionAction(InsertPt) {
1302 IRBuilder<> Builder(InsertPt);
1303 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
1304 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
1307 /// \brief Get the built value.
1308 Value *getBuiltValue() { return Val; }
1310 /// \brief Remove the built instruction.
1311 void undo() override {
1312 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
1313 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1314 IVal->eraseFromParent();
1318 /// \brief Build a zero extension instruction.
1319 class ZExtBuilder : public TypePromotionAction {
1322 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
1324 /// zext Opnd to Ty.
1325 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1326 : TypePromotionAction(InsertPt) {
1327 IRBuilder<> Builder(InsertPt);
1328 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
1329 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
1332 /// \brief Get the built value.
1333 Value *getBuiltValue() { return Val; }
1335 /// \brief Remove the built instruction.
1336 void undo() override {
1337 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
1338 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1339 IVal->eraseFromParent();
1343 /// \brief Mutate an instruction to another type.
1344 class TypeMutator : public TypePromotionAction {
1345 /// Record the original type.
1349 /// \brief Mutate the type of \p Inst into \p NewTy.
1350 TypeMutator(Instruction *Inst, Type *NewTy)
1351 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1352 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1354 Inst->mutateType(NewTy);
1357 /// \brief Mutate the instruction back to its original type.
1358 void undo() override {
1359 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1361 Inst->mutateType(OrigTy);
1365 /// \brief Replace the uses of an instruction by another instruction.
1366 class UsesReplacer : public TypePromotionAction {
1367 /// Helper structure to keep track of the replaced uses.
1368 struct InstructionAndIdx {
1369 /// The instruction using the instruction.
1371 /// The index where this instruction is used for Inst.
1373 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1374 : Inst(Inst), Idx(Idx) {}
1377 /// Keep track of the original uses (pair Instruction, Index).
1378 SmallVector<InstructionAndIdx, 4> OriginalUses;
1379 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1382 /// \brief Replace all the use of \p Inst by \p New.
1383 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1384 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1386 // Record the original uses.
1387 for (Use &U : Inst->uses()) {
1388 Instruction *UserI = cast<Instruction>(U.getUser());
1389 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1391 // Now, we can replace the uses.
1392 Inst->replaceAllUsesWith(New);
1395 /// \brief Reassign the original uses of Inst to Inst.
1396 void undo() override {
1397 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1398 for (use_iterator UseIt = OriginalUses.begin(),
1399 EndIt = OriginalUses.end();
1400 UseIt != EndIt; ++UseIt) {
1401 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1406 /// \brief Remove an instruction from the IR.
1407 class InstructionRemover : public TypePromotionAction {
1408 /// Original position of the instruction.
1409 InsertionHandler Inserter;
1410 /// Helper structure to hide all the link to the instruction. In other
1411 /// words, this helps to do as if the instruction was removed.
1412 OperandsHider Hider;
1413 /// Keep track of the uses replaced, if any.
1414 UsesReplacer *Replacer;
1417 /// \brief Remove all reference of \p Inst and optinally replace all its
1419 /// \pre If !Inst->use_empty(), then New != nullptr
1420 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1421 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1424 Replacer = new UsesReplacer(Inst, New);
1425 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1426 Inst->removeFromParent();
1429 ~InstructionRemover() { delete Replacer; }
1431 /// \brief Really remove the instruction.
1432 void commit() override { delete Inst; }
1434 /// \brief Resurrect the instruction and reassign it to the proper uses if
1435 /// new value was provided when build this action.
1436 void undo() override {
1437 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1438 Inserter.insert(Inst);
1446 /// Restoration point.
1447 /// The restoration point is a pointer to an action instead of an iterator
1448 /// because the iterator may be invalidated but not the pointer.
1449 typedef const TypePromotionAction *ConstRestorationPt;
1450 /// Advocate every changes made in that transaction.
1452 /// Undo all the changes made after the given point.
1453 void rollback(ConstRestorationPt Point);
1454 /// Get the current restoration point.
1455 ConstRestorationPt getRestorationPoint() const;
1457 /// \name API for IR modification with state keeping to support rollback.
1459 /// Same as Instruction::setOperand.
1460 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1461 /// Same as Instruction::eraseFromParent.
1462 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
1463 /// Same as Value::replaceAllUsesWith.
1464 void replaceAllUsesWith(Instruction *Inst, Value *New);
1465 /// Same as Value::mutateType.
1466 void mutateType(Instruction *Inst, Type *NewTy);
1467 /// Same as IRBuilder::createTrunc.
1468 Value *createTrunc(Instruction *Opnd, Type *Ty);
1469 /// Same as IRBuilder::createSExt.
1470 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1471 /// Same as IRBuilder::createZExt.
1472 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
1473 /// Same as Instruction::moveBefore.
1474 void moveBefore(Instruction *Inst, Instruction *Before);
1478 /// The ordered list of actions made so far.
1479 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
1480 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
1483 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1486 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
1489 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1492 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
1495 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1497 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
1500 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1501 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
1504 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1506 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
1507 Value *Val = Ptr->getBuiltValue();
1508 Actions.push_back(std::move(Ptr));
1512 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
1513 Value *Opnd, Type *Ty) {
1514 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
1515 Value *Val = Ptr->getBuiltValue();
1516 Actions.push_back(std::move(Ptr));
1520 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
1521 Value *Opnd, Type *Ty) {
1522 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
1523 Value *Val = Ptr->getBuiltValue();
1524 Actions.push_back(std::move(Ptr));
1528 void TypePromotionTransaction::moveBefore(Instruction *Inst,
1529 Instruction *Before) {
1531 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
1534 TypePromotionTransaction::ConstRestorationPt
1535 TypePromotionTransaction::getRestorationPoint() const {
1536 return !Actions.empty() ? Actions.back().get() : nullptr;
1539 void TypePromotionTransaction::commit() {
1540 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
1546 void TypePromotionTransaction::rollback(
1547 TypePromotionTransaction::ConstRestorationPt Point) {
1548 while (!Actions.empty() && Point != Actions.back().get()) {
1549 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
1554 /// \brief A helper class for matching addressing modes.
1556 /// This encapsulates the logic for matching the target-legal addressing modes.
1557 class AddressingModeMatcher {
1558 SmallVectorImpl<Instruction*> &AddrModeInsts;
1559 const TargetLowering &TLI;
1561 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
1562 /// the memory instruction that we're computing this address for.
1564 Instruction *MemoryInst;
1566 /// AddrMode - This is the addressing mode that we're building up. This is
1567 /// part of the return value of this addressing mode matching stuff.
1568 ExtAddrMode &AddrMode;
1570 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
1571 const SetOfInstrs &InsertedTruncs;
1572 /// A map from the instructions to their type before promotion.
1573 InstrToOrigTy &PromotedInsts;
1574 /// The ongoing transaction where every action should be registered.
1575 TypePromotionTransaction &TPT;
1577 /// IgnoreProfitability - This is set to true when we should not do
1578 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
1579 /// always returns true.
1580 bool IgnoreProfitability;
1582 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
1583 const TargetLowering &T, Type *AT,
1584 Instruction *MI, ExtAddrMode &AM,
1585 const SetOfInstrs &InsertedTruncs,
1586 InstrToOrigTy &PromotedInsts,
1587 TypePromotionTransaction &TPT)
1588 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM),
1589 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
1590 IgnoreProfitability = false;
1594 /// Match - Find the maximal addressing mode that a load/store of V can fold,
1595 /// give an access type of AccessTy. This returns a list of involved
1596 /// instructions in AddrModeInsts.
1597 /// \p InsertedTruncs The truncate instruction inserted by other
1600 /// \p PromotedInsts maps the instructions to their type before promotion.
1601 /// \p The ongoing transaction where every action should be registered.
1602 static ExtAddrMode Match(Value *V, Type *AccessTy,
1603 Instruction *MemoryInst,
1604 SmallVectorImpl<Instruction*> &AddrModeInsts,
1605 const TargetLowering &TLI,
1606 const SetOfInstrs &InsertedTruncs,
1607 InstrToOrigTy &PromotedInsts,
1608 TypePromotionTransaction &TPT) {
1611 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
1612 MemoryInst, Result, InsertedTruncs,
1613 PromotedInsts, TPT).MatchAddr(V, 0);
1614 (void)Success; assert(Success && "Couldn't select *anything*?");
1618 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
1619 bool MatchAddr(Value *V, unsigned Depth);
1620 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
1621 bool *MovedAway = nullptr);
1622 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
1623 ExtAddrMode &AMBefore,
1624 ExtAddrMode &AMAfter);
1625 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
1626 bool IsPromotionProfitable(unsigned MatchedSize, unsigned SizeWithPromotion,
1627 Value *PromotedOperand) const;
1630 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
1631 /// Return true and update AddrMode if this addr mode is legal for the target,
1633 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
1635 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
1636 // mode. Just process that directly.
1638 return MatchAddr(ScaleReg, Depth);
1640 // If the scale is 0, it takes nothing to add this.
1644 // If we already have a scale of this value, we can add to it, otherwise, we
1645 // need an available scale field.
1646 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
1649 ExtAddrMode TestAddrMode = AddrMode;
1651 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
1652 // [A+B + A*7] -> [B+A*8].
1653 TestAddrMode.Scale += Scale;
1654 TestAddrMode.ScaledReg = ScaleReg;
1656 // If the new address isn't legal, bail out.
1657 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
1660 // It was legal, so commit it.
1661 AddrMode = TestAddrMode;
1663 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
1664 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
1665 // X*Scale + C*Scale to addr mode.
1666 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
1667 if (isa<Instruction>(ScaleReg) && // not a constant expr.
1668 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
1669 TestAddrMode.ScaledReg = AddLHS;
1670 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
1672 // If this addressing mode is legal, commit it and remember that we folded
1673 // this instruction.
1674 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
1675 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
1676 AddrMode = TestAddrMode;
1681 // Otherwise, not (x+c)*scale, just return what we have.
1685 /// MightBeFoldableInst - This is a little filter, which returns true if an
1686 /// addressing computation involving I might be folded into a load/store
1687 /// accessing it. This doesn't need to be perfect, but needs to accept at least
1688 /// the set of instructions that MatchOperationAddr can.
1689 static bool MightBeFoldableInst(Instruction *I) {
1690 switch (I->getOpcode()) {
1691 case Instruction::BitCast:
1692 case Instruction::AddrSpaceCast:
1693 // Don't touch identity bitcasts.
1694 if (I->getType() == I->getOperand(0)->getType())
1696 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
1697 case Instruction::PtrToInt:
1698 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1700 case Instruction::IntToPtr:
1701 // We know the input is intptr_t, so this is foldable.
1703 case Instruction::Add:
1705 case Instruction::Mul:
1706 case Instruction::Shl:
1707 // Can only handle X*C and X << C.
1708 return isa<ConstantInt>(I->getOperand(1));
1709 case Instruction::GetElementPtr:
1716 /// \brief Hepler class to perform type promotion.
1717 class TypePromotionHelper {
1718 /// \brief Utility function to check whether or not a sign extension of
1719 /// \p Inst with \p ConsideredSExtType can be moved through \p Inst by either
1720 /// using the operands of \p Inst or promoting \p Inst.
1721 /// In other words, check if:
1722 /// sext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredSExtType.
1723 /// #1 Promotion applies:
1724 /// ConsideredSExtType Inst (sext opnd1 to ConsideredSExtType, ...).
1725 /// #2 Operand reuses:
1726 /// sext opnd1 to ConsideredSExtType.
1727 /// \p PromotedInsts maps the instructions to their type before promotion.
1728 static bool canGetThrough(const Instruction *Inst, Type *ConsideredSExtType,
1729 const InstrToOrigTy &PromotedInsts);
1731 /// \brief Utility function to determine if \p OpIdx should be promoted when
1732 /// promoting \p Inst.
1733 static bool shouldSExtOperand(const Instruction *Inst, int OpIdx) {
1734 if (isa<SelectInst>(Inst) && OpIdx == 0)
1739 /// \brief Utility function to promote the operand of \p SExt when this
1740 /// operand is a promotable trunc or sext or zext.
1741 /// \p PromotedInsts maps the instructions to their type before promotion.
1742 /// \p CreatedInsts[out] contains how many non-free instructions have been
1743 /// created to promote the operand of SExt.
1744 /// Should never be called directly.
1745 /// \return The promoted value which is used instead of SExt.
1746 static Value *promoteOperandForTruncAndAnyExt(Instruction *SExt,
1747 TypePromotionTransaction &TPT,
1748 InstrToOrigTy &PromotedInsts,
1749 unsigned &CreatedInsts);
1751 /// \brief Utility function to promote the operand of \p SExt when this
1752 /// operand is promotable and is not a supported trunc or sext.
1753 /// \p PromotedInsts maps the instructions to their type before promotion.
1754 /// \p CreatedInsts[out] contains how many non-free instructions have been
1755 /// created to promote the operand of SExt.
1756 /// Should never be called directly.
1757 /// \return The promoted value which is used instead of SExt.
1758 static Value *promoteOperandForOther(Instruction *SExt,
1759 TypePromotionTransaction &TPT,
1760 InstrToOrigTy &PromotedInsts,
1761 unsigned &CreatedInsts);
1764 /// Type for the utility function that promotes the operand of SExt.
1765 typedef Value *(*Action)(Instruction *SExt, TypePromotionTransaction &TPT,
1766 InstrToOrigTy &PromotedInsts,
1767 unsigned &CreatedInsts);
1768 /// \brief Given a sign extend instruction \p SExt, return the approriate
1769 /// action to promote the operand of \p SExt instead of using SExt.
1770 /// \return NULL if no promotable action is possible with the current
1772 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
1773 /// the others CodeGenPrepare optimizations. This information is important
1774 /// because we do not want to promote these instructions as CodeGenPrepare
1775 /// will reinsert them later. Thus creating an infinite loop: create/remove.
1776 /// \p PromotedInsts maps the instructions to their type before promotion.
1777 static Action getAction(Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1778 const TargetLowering &TLI,
1779 const InstrToOrigTy &PromotedInsts);
1782 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
1783 Type *ConsideredSExtType,
1784 const InstrToOrigTy &PromotedInsts) {
1785 // We can always get through sext or zext.
1786 if (isa<SExtInst>(Inst) || isa<ZExtInst>(Inst))
1789 // We can get through binary operator, if it is legal. In other words, the
1790 // binary operator must have a nuw or nsw flag.
1791 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
1792 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
1793 (BinOp->hasNoUnsignedWrap() || BinOp->hasNoSignedWrap()))
1796 // Check if we can do the following simplification.
1797 // sext(trunc(sext)) --> sext
1798 if (!isa<TruncInst>(Inst))
1801 Value *OpndVal = Inst->getOperand(0);
1802 // Check if we can use this operand in the sext.
1803 // If the type is larger than the result type of the sign extension,
1805 if (OpndVal->getType()->getIntegerBitWidth() >
1806 ConsideredSExtType->getIntegerBitWidth())
1809 // If the operand of the truncate is not an instruction, we will not have
1810 // any information on the dropped bits.
1811 // (Actually we could for constant but it is not worth the extra logic).
1812 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
1816 // Check if the source of the type is narrow enough.
1817 // I.e., check that trunc just drops sign extended bits.
1818 // #1 get the type of the operand.
1819 const Type *OpndType;
1820 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
1821 if (It != PromotedInsts.end())
1822 OpndType = It->second;
1823 else if (isa<SExtInst>(Opnd))
1824 OpndType = cast<Instruction>(Opnd)->getOperand(0)->getType();
1828 // #2 check that the truncate just drop sign extended bits.
1829 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
1835 TypePromotionHelper::Action TypePromotionHelper::getAction(
1836 Instruction *SExt, const SetOfInstrs &InsertedTruncs,
1837 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
1838 Instruction *SExtOpnd = dyn_cast<Instruction>(SExt->getOperand(0));
1839 Type *SExtTy = SExt->getType();
1840 // If the operand of the sign extension is not an instruction, we cannot
1842 // If it, check we can get through.
1843 if (!SExtOpnd || !canGetThrough(SExtOpnd, SExtTy, PromotedInsts))
1846 // Do not promote if the operand has been added by codegenprepare.
1847 // Otherwise, it means we are undoing an optimization that is likely to be
1848 // redone, thus causing potential infinite loop.
1849 if (isa<TruncInst>(SExtOpnd) && InsertedTruncs.count(SExtOpnd))
1852 // SExt or Trunc instructions.
1853 // Return the related handler.
1854 if (isa<SExtInst>(SExtOpnd) || isa<TruncInst>(SExtOpnd) ||
1855 isa<ZExtInst>(SExtOpnd))
1856 return promoteOperandForTruncAndAnyExt;
1858 // Regular instruction.
1859 // Abort early if we will have to insert non-free instructions.
1860 if (!SExtOpnd->hasOneUse() &&
1861 !TLI.isTruncateFree(SExtTy, SExtOpnd->getType()))
1863 return promoteOperandForOther;
1866 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
1867 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
1868 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts) {
1869 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1870 // get through it and this method should not be called.
1871 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1872 Value *ExtVal = SExt;
1873 if (isa<ZExtInst>(SExtOpnd)) {
1874 // Replace sext(zext(opnd))
1877 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
1878 TPT.replaceAllUsesWith(SExt, ZExt);
1879 TPT.eraseInstruction(SExt);
1882 // Replace sext(trunc(opnd)) or sext(sext(opnd))
1884 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
1888 // Remove dead code.
1889 if (SExtOpnd->use_empty())
1890 TPT.eraseInstruction(SExtOpnd);
1892 // Check if the extension is still needed.
1893 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
1894 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType())
1897 // At this point we have: ext ty opnd to ty.
1898 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
1899 Value *NextVal = ExtInst->getOperand(0);
1900 TPT.eraseInstruction(ExtInst, NextVal);
1905 TypePromotionHelper::promoteOperandForOther(Instruction *SExt,
1906 TypePromotionTransaction &TPT,
1907 InstrToOrigTy &PromotedInsts,
1908 unsigned &CreatedInsts) {
1909 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1910 // get through it and this method should not be called.
1911 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1913 if (!SExtOpnd->hasOneUse()) {
1914 // SExtOpnd will be promoted.
1915 // All its uses, but SExt, will need to use a truncated value of the
1916 // promoted version.
1917 // Create the truncate now.
1918 Value *Trunc = TPT.createTrunc(SExt, SExtOpnd->getType());
1919 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
1920 ITrunc->removeFromParent();
1921 // Insert it just after the definition.
1922 ITrunc->insertAfter(SExtOpnd);
1925 TPT.replaceAllUsesWith(SExtOpnd, Trunc);
1926 // Restore the operand of SExt (which has been replace by the previous call
1927 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
1928 TPT.setOperand(SExt, 0, SExtOpnd);
1931 // Get through the Instruction:
1932 // 1. Update its type.
1933 // 2. Replace the uses of SExt by Inst.
1934 // 3. Sign extend each operand that needs to be sign extended.
1936 // Remember the original type of the instruction before promotion.
1937 // This is useful to know that the high bits are sign extended bits.
1938 PromotedInsts.insert(
1939 std::pair<Instruction *, Type *>(SExtOpnd, SExtOpnd->getType()));
1941 TPT.mutateType(SExtOpnd, SExt->getType());
1943 TPT.replaceAllUsesWith(SExt, SExtOpnd);
1945 Instruction *SExtForOpnd = SExt;
1947 DEBUG(dbgs() << "Propagate SExt to operands\n");
1948 for (int OpIdx = 0, EndOpIdx = SExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
1950 DEBUG(dbgs() << "Operand:\n" << *(SExtOpnd->getOperand(OpIdx)) << '\n');
1951 if (SExtOpnd->getOperand(OpIdx)->getType() == SExt->getType() ||
1952 !shouldSExtOperand(SExtOpnd, OpIdx)) {
1953 DEBUG(dbgs() << "No need to propagate\n");
1956 // Check if we can statically sign extend the operand.
1957 Value *Opnd = SExtOpnd->getOperand(OpIdx);
1958 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
1959 DEBUG(dbgs() << "Statically sign extend\n");
1962 ConstantInt::getSigned(SExt->getType(), Cst->getSExtValue()));
1965 // UndefValue are typed, so we have to statically sign extend them.
1966 if (isa<UndefValue>(Opnd)) {
1967 DEBUG(dbgs() << "Statically sign extend\n");
1968 TPT.setOperand(SExtOpnd, OpIdx, UndefValue::get(SExt->getType()));
1972 // Otherwise we have to explicity sign extend the operand.
1973 // Check if SExt was reused to sign extend an operand.
1975 // If yes, create a new one.
1976 DEBUG(dbgs() << "More operands to sext\n");
1978 cast<Instruction>(TPT.createSExt(SExt, Opnd, SExt->getType()));
1982 TPT.setOperand(SExtForOpnd, 0, Opnd);
1984 // Move the sign extension before the insertion point.
1985 TPT.moveBefore(SExtForOpnd, SExtOpnd);
1986 TPT.setOperand(SExtOpnd, OpIdx, SExtForOpnd);
1987 // If more sext are required, new instructions will have to be created.
1988 SExtForOpnd = nullptr;
1990 if (SExtForOpnd == SExt) {
1991 DEBUG(dbgs() << "Sign extension is useless now\n");
1992 TPT.eraseInstruction(SExt);
1997 /// IsPromotionProfitable - Check whether or not promoting an instruction
1998 /// to a wider type was profitable.
1999 /// \p MatchedSize gives the number of instructions that have been matched
2000 /// in the addressing mode after the promotion was applied.
2001 /// \p SizeWithPromotion gives the number of created instructions for
2002 /// the promotion plus the number of instructions that have been
2003 /// matched in the addressing mode before the promotion.
2004 /// \p PromotedOperand is the value that has been promoted.
2005 /// \return True if the promotion is profitable, false otherwise.
2007 AddressingModeMatcher::IsPromotionProfitable(unsigned MatchedSize,
2008 unsigned SizeWithPromotion,
2009 Value *PromotedOperand) const {
2010 // We folded less instructions than what we created to promote the operand.
2011 // This is not profitable.
2012 if (MatchedSize < SizeWithPromotion)
2014 if (MatchedSize > SizeWithPromotion)
2016 // The promotion is neutral but it may help folding the sign extension in
2017 // loads for instance.
2018 // Check that we did not create an illegal instruction.
2019 Instruction *PromotedInst = dyn_cast<Instruction>(PromotedOperand);
2022 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2023 // If the ISDOpcode is undefined, it was undefined before the promotion.
2026 // Otherwise, check if the promoted instruction is legal or not.
2027 return TLI.isOperationLegalOrCustom(ISDOpcode,
2028 EVT::getEVT(PromotedInst->getType()));
2031 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
2032 /// fold the operation into the addressing mode. If so, update the addressing
2033 /// mode and return true, otherwise return false without modifying AddrMode.
2034 /// If \p MovedAway is not NULL, it contains the information of whether or
2035 /// not AddrInst has to be folded into the addressing mode on success.
2036 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2037 /// because it has been moved away.
2038 /// Thus AddrInst must not be added in the matched instructions.
2039 /// This state can happen when AddrInst is a sext, since it may be moved away.
2040 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2041 /// not be referenced anymore.
2042 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
2045 // Avoid exponential behavior on extremely deep expression trees.
2046 if (Depth >= 5) return false;
2048 // By default, all matched instructions stay in place.
2053 case Instruction::PtrToInt:
2054 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2055 return MatchAddr(AddrInst->getOperand(0), Depth);
2056 case Instruction::IntToPtr:
2057 // This inttoptr is a no-op if the integer type is pointer sized.
2058 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
2059 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
2060 return MatchAddr(AddrInst->getOperand(0), Depth);
2062 case Instruction::BitCast:
2063 case Instruction::AddrSpaceCast:
2064 // BitCast is always a noop, and we can handle it as long as it is
2065 // int->int or pointer->pointer (we don't want int<->fp or something).
2066 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2067 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2068 // Don't touch identity bitcasts. These were probably put here by LSR,
2069 // and we don't want to mess around with them. Assume it knows what it
2071 AddrInst->getOperand(0)->getType() != AddrInst->getType())
2072 return MatchAddr(AddrInst->getOperand(0), Depth);
2074 case Instruction::Add: {
2075 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2076 ExtAddrMode BackupAddrMode = AddrMode;
2077 unsigned OldSize = AddrModeInsts.size();
2078 // Start a transaction at this point.
2079 // The LHS may match but not the RHS.
2080 // Therefore, we need a higher level restoration point to undo partially
2081 // matched operation.
2082 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2083 TPT.getRestorationPoint();
2085 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2086 MatchAddr(AddrInst->getOperand(0), Depth+1))
2089 // Restore the old addr mode info.
2090 AddrMode = BackupAddrMode;
2091 AddrModeInsts.resize(OldSize);
2092 TPT.rollback(LastKnownGood);
2094 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2095 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2096 MatchAddr(AddrInst->getOperand(1), Depth+1))
2099 // Otherwise we definitely can't merge the ADD in.
2100 AddrMode = BackupAddrMode;
2101 AddrModeInsts.resize(OldSize);
2102 TPT.rollback(LastKnownGood);
2105 //case Instruction::Or:
2106 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2108 case Instruction::Mul:
2109 case Instruction::Shl: {
2110 // Can only handle X*C and X << C.
2111 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2114 int64_t Scale = RHS->getSExtValue();
2115 if (Opcode == Instruction::Shl)
2116 Scale = 1LL << Scale;
2118 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2120 case Instruction::GetElementPtr: {
2121 // Scan the GEP. We check it if it contains constant offsets and at most
2122 // one variable offset.
2123 int VariableOperand = -1;
2124 unsigned VariableScale = 0;
2126 int64_t ConstantOffset = 0;
2127 const DataLayout *TD = TLI.getDataLayout();
2128 gep_type_iterator GTI = gep_type_begin(AddrInst);
2129 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2130 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2131 const StructLayout *SL = TD->getStructLayout(STy);
2133 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2134 ConstantOffset += SL->getElementOffset(Idx);
2136 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
2137 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2138 ConstantOffset += CI->getSExtValue()*TypeSize;
2139 } else if (TypeSize) { // Scales of zero don't do anything.
2140 // We only allow one variable index at the moment.
2141 if (VariableOperand != -1)
2144 // Remember the variable index.
2145 VariableOperand = i;
2146 VariableScale = TypeSize;
2151 // A common case is for the GEP to only do a constant offset. In this case,
2152 // just add it to the disp field and check validity.
2153 if (VariableOperand == -1) {
2154 AddrMode.BaseOffs += ConstantOffset;
2155 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
2156 // Check to see if we can fold the base pointer in too.
2157 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2160 AddrMode.BaseOffs -= ConstantOffset;
2164 // Save the valid addressing mode in case we can't match.
2165 ExtAddrMode BackupAddrMode = AddrMode;
2166 unsigned OldSize = AddrModeInsts.size();
2168 // See if the scale and offset amount is valid for this target.
2169 AddrMode.BaseOffs += ConstantOffset;
2171 // Match the base operand of the GEP.
2172 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2173 // If it couldn't be matched, just stuff the value in a register.
2174 if (AddrMode.HasBaseReg) {
2175 AddrMode = BackupAddrMode;
2176 AddrModeInsts.resize(OldSize);
2179 AddrMode.HasBaseReg = true;
2180 AddrMode.BaseReg = AddrInst->getOperand(0);
2183 // Match the remaining variable portion of the GEP.
2184 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2186 // If it couldn't be matched, try stuffing the base into a register
2187 // instead of matching it, and retrying the match of the scale.
2188 AddrMode = BackupAddrMode;
2189 AddrModeInsts.resize(OldSize);
2190 if (AddrMode.HasBaseReg)
2192 AddrMode.HasBaseReg = true;
2193 AddrMode.BaseReg = AddrInst->getOperand(0);
2194 AddrMode.BaseOffs += ConstantOffset;
2195 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2196 VariableScale, Depth)) {
2197 // If even that didn't work, bail.
2198 AddrMode = BackupAddrMode;
2199 AddrModeInsts.resize(OldSize);
2206 case Instruction::SExt: {
2207 Instruction *SExt = dyn_cast<Instruction>(AddrInst);
2211 // Try to move this sext out of the way of the addressing mode.
2212 // Ask for a method for doing so.
2213 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
2214 SExt, InsertedTruncs, TLI, PromotedInsts);
2218 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2219 TPT.getRestorationPoint();
2220 unsigned CreatedInsts = 0;
2221 Value *PromotedOperand = TPH(SExt, TPT, PromotedInsts, CreatedInsts);
2222 // SExt has been moved away.
2223 // Thus either it will be rematched later in the recursive calls or it is
2224 // gone. Anyway, we must not fold it into the addressing mode at this point.
2228 // addr = gep base, idx
2230 // promotedOpnd = sext opnd <- no match here
2231 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2232 // addr = gep base, op <- match
2236 assert(PromotedOperand &&
2237 "TypePromotionHelper should have filtered out those cases");
2239 ExtAddrMode BackupAddrMode = AddrMode;
2240 unsigned OldSize = AddrModeInsts.size();
2242 if (!MatchAddr(PromotedOperand, Depth) ||
2243 !IsPromotionProfitable(AddrModeInsts.size(), OldSize + CreatedInsts,
2245 AddrMode = BackupAddrMode;
2246 AddrModeInsts.resize(OldSize);
2247 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2248 TPT.rollback(LastKnownGood);
2257 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2258 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2259 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2260 /// or intptr_t for the target.
2262 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2263 // Start a transaction at this point that we will rollback if the matching
2265 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2266 TPT.getRestorationPoint();
2267 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2268 // Fold in immediates if legal for the target.
2269 AddrMode.BaseOffs += CI->getSExtValue();
2270 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2272 AddrMode.BaseOffs -= CI->getSExtValue();
2273 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2274 // If this is a global variable, try to fold it into the addressing mode.
2275 if (!AddrMode.BaseGV) {
2276 AddrMode.BaseGV = GV;
2277 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2279 AddrMode.BaseGV = nullptr;
2281 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2282 ExtAddrMode BackupAddrMode = AddrMode;
2283 unsigned OldSize = AddrModeInsts.size();
2285 // Check to see if it is possible to fold this operation.
2286 bool MovedAway = false;
2287 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2288 // This instruction may have been move away. If so, there is nothing
2292 // Okay, it's possible to fold this. Check to see if it is actually
2293 // *profitable* to do so. We use a simple cost model to avoid increasing
2294 // register pressure too much.
2295 if (I->hasOneUse() ||
2296 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2297 AddrModeInsts.push_back(I);
2301 // It isn't profitable to do this, roll back.
2302 //cerr << "NOT FOLDING: " << *I;
2303 AddrMode = BackupAddrMode;
2304 AddrModeInsts.resize(OldSize);
2305 TPT.rollback(LastKnownGood);
2307 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2308 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2310 TPT.rollback(LastKnownGood);
2311 } else if (isa<ConstantPointerNull>(Addr)) {
2312 // Null pointer gets folded without affecting the addressing mode.
2316 // Worse case, the target should support [reg] addressing modes. :)
2317 if (!AddrMode.HasBaseReg) {
2318 AddrMode.HasBaseReg = true;
2319 AddrMode.BaseReg = Addr;
2320 // Still check for legality in case the target supports [imm] but not [i+r].
2321 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2323 AddrMode.HasBaseReg = false;
2324 AddrMode.BaseReg = nullptr;
2327 // If the base register is already taken, see if we can do [r+r].
2328 if (AddrMode.Scale == 0) {
2330 AddrMode.ScaledReg = Addr;
2331 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2334 AddrMode.ScaledReg = nullptr;
2337 TPT.rollback(LastKnownGood);
2341 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2342 /// inline asm call are due to memory operands. If so, return true, otherwise
2344 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2345 const TargetLowering &TLI) {
2346 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(ImmutableCallSite(CI));
2347 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2348 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2350 // Compute the constraint code and ConstraintType to use.
2351 TLI.ComputeConstraintToUse(OpInfo, SDValue());
2353 // If this asm operand is our Value*, and if it isn't an indirect memory
2354 // operand, we can't fold it!
2355 if (OpInfo.CallOperandVal == OpVal &&
2356 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2357 !OpInfo.isIndirect))
2364 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2365 /// memory use. If we find an obviously non-foldable instruction, return true.
2366 /// Add the ultimately found memory instructions to MemoryUses.
2367 static bool FindAllMemoryUses(Instruction *I,
2368 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
2369 SmallPtrSetImpl<Instruction*> &ConsideredInsts,
2370 const TargetLowering &TLI) {
2371 // If we already considered this instruction, we're done.
2372 if (!ConsideredInsts.insert(I))
2375 // If this is an obviously unfoldable instruction, bail out.
2376 if (!MightBeFoldableInst(I))
2379 // Loop over all the uses, recursively processing them.
2380 for (Use &U : I->uses()) {
2381 Instruction *UserI = cast<Instruction>(U.getUser());
2383 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2384 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2388 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2389 unsigned opNo = U.getOperandNo();
2390 if (opNo == 0) return true; // Storing addr, not into addr.
2391 MemoryUses.push_back(std::make_pair(SI, opNo));
2395 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2396 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2397 if (!IA) return true;
2399 // If this is a memory operand, we're cool, otherwise bail out.
2400 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
2405 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI))
2412 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2413 /// the use site that we're folding it into. If so, there is no cost to
2414 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2415 /// that we know are live at the instruction already.
2416 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2417 Value *KnownLive2) {
2418 // If Val is either of the known-live values, we know it is live!
2419 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
2422 // All values other than instructions and arguments (e.g. constants) are live.
2423 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2425 // If Val is a constant sized alloca in the entry block, it is live, this is
2426 // true because it is just a reference to the stack/frame pointer, which is
2427 // live for the whole function.
2428 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2429 if (AI->isStaticAlloca())
2432 // Check to see if this value is already used in the memory instruction's
2433 // block. If so, it's already live into the block at the very least, so we
2434 // can reasonably fold it.
2435 return Val->isUsedInBasicBlock(MemoryInst->getParent());
2438 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
2439 /// mode of the machine to fold the specified instruction into a load or store
2440 /// that ultimately uses it. However, the specified instruction has multiple
2441 /// uses. Given this, it may actually increase register pressure to fold it
2442 /// into the load. For example, consider this code:
2446 /// use(Y) -> nonload/store
2450 /// In this case, Y has multiple uses, and can be folded into the load of Z
2451 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
2452 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
2453 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
2454 /// number of computations either.
2456 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
2457 /// X was live across 'load Z' for other reasons, we actually *would* want to
2458 /// fold the addressing mode in the Z case. This would make Y die earlier.
2459 bool AddressingModeMatcher::
2460 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
2461 ExtAddrMode &AMAfter) {
2462 if (IgnoreProfitability) return true;
2464 // AMBefore is the addressing mode before this instruction was folded into it,
2465 // and AMAfter is the addressing mode after the instruction was folded. Get
2466 // the set of registers referenced by AMAfter and subtract out those
2467 // referenced by AMBefore: this is the set of values which folding in this
2468 // address extends the lifetime of.
2470 // Note that there are only two potential values being referenced here,
2471 // BaseReg and ScaleReg (global addresses are always available, as are any
2472 // folded immediates).
2473 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
2475 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
2476 // lifetime wasn't extended by adding this instruction.
2477 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2479 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2480 ScaledReg = nullptr;
2482 // If folding this instruction (and it's subexprs) didn't extend any live
2483 // ranges, we're ok with it.
2484 if (!BaseReg && !ScaledReg)
2487 // If all uses of this instruction are ultimately load/store/inlineasm's,
2488 // check to see if their addressing modes will include this instruction. If
2489 // so, we can fold it into all uses, so it doesn't matter if it has multiple
2491 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
2492 SmallPtrSet<Instruction*, 16> ConsideredInsts;
2493 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
2494 return false; // Has a non-memory, non-foldable use!
2496 // Now that we know that all uses of this instruction are part of a chain of
2497 // computation involving only operations that could theoretically be folded
2498 // into a memory use, loop over each of these uses and see if they could
2499 // *actually* fold the instruction.
2500 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
2501 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
2502 Instruction *User = MemoryUses[i].first;
2503 unsigned OpNo = MemoryUses[i].second;
2505 // Get the access type of this use. If the use isn't a pointer, we don't
2506 // know what it accesses.
2507 Value *Address = User->getOperand(OpNo);
2508 if (!Address->getType()->isPointerTy())
2510 Type *AddressAccessTy = Address->getType()->getPointerElementType();
2512 // Do a match against the root of this address, ignoring profitability. This
2513 // will tell us if the addressing mode for the memory operation will
2514 // *actually* cover the shared instruction.
2516 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2517 TPT.getRestorationPoint();
2518 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
2519 MemoryInst, Result, InsertedTruncs,
2520 PromotedInsts, TPT);
2521 Matcher.IgnoreProfitability = true;
2522 bool Success = Matcher.MatchAddr(Address, 0);
2523 (void)Success; assert(Success && "Couldn't select *anything*?");
2525 // The match was to check the profitability, the changes made are not
2526 // part of the original matcher. Therefore, they should be dropped
2527 // otherwise the original matcher will not present the right state.
2528 TPT.rollback(LastKnownGood);
2530 // If the match didn't cover I, then it won't be shared by it.
2531 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
2532 I) == MatchedAddrModeInsts.end())
2535 MatchedAddrModeInsts.clear();
2541 } // end anonymous namespace
2543 /// IsNonLocalValue - Return true if the specified values are defined in a
2544 /// different basic block than BB.
2545 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
2546 if (Instruction *I = dyn_cast<Instruction>(V))
2547 return I->getParent() != BB;
2551 /// OptimizeMemoryInst - Load and Store Instructions often have
2552 /// addressing modes that can do significant amounts of computation. As such,
2553 /// instruction selection will try to get the load or store to do as much
2554 /// computation as possible for the program. The problem is that isel can only
2555 /// see within a single block. As such, we sink as much legal addressing mode
2556 /// stuff into the block as possible.
2558 /// This method is used to optimize both load/store and inline asms with memory
2560 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
2564 // Try to collapse single-value PHI nodes. This is necessary to undo
2565 // unprofitable PRE transformations.
2566 SmallVector<Value*, 8> worklist;
2567 SmallPtrSet<Value*, 16> Visited;
2568 worklist.push_back(Addr);
2570 // Use a worklist to iteratively look through PHI nodes, and ensure that
2571 // the addressing mode obtained from the non-PHI roots of the graph
2573 Value *Consensus = nullptr;
2574 unsigned NumUsesConsensus = 0;
2575 bool IsNumUsesConsensusValid = false;
2576 SmallVector<Instruction*, 16> AddrModeInsts;
2577 ExtAddrMode AddrMode;
2578 TypePromotionTransaction TPT;
2579 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2580 TPT.getRestorationPoint();
2581 while (!worklist.empty()) {
2582 Value *V = worklist.back();
2583 worklist.pop_back();
2585 // Break use-def graph loops.
2586 if (!Visited.insert(V)) {
2587 Consensus = nullptr;
2591 // For a PHI node, push all of its incoming values.
2592 if (PHINode *P = dyn_cast<PHINode>(V)) {
2593 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
2594 worklist.push_back(P->getIncomingValue(i));
2598 // For non-PHIs, determine the addressing mode being computed.
2599 SmallVector<Instruction*, 16> NewAddrModeInsts;
2600 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
2601 V, AccessTy, MemoryInst, NewAddrModeInsts, *TLI, InsertedTruncsSet,
2602 PromotedInsts, TPT);
2604 // This check is broken into two cases with very similar code to avoid using
2605 // getNumUses() as much as possible. Some values have a lot of uses, so
2606 // calling getNumUses() unconditionally caused a significant compile-time
2610 AddrMode = NewAddrMode;
2611 AddrModeInsts = NewAddrModeInsts;
2613 } else if (NewAddrMode == AddrMode) {
2614 if (!IsNumUsesConsensusValid) {
2615 NumUsesConsensus = Consensus->getNumUses();
2616 IsNumUsesConsensusValid = true;
2619 // Ensure that the obtained addressing mode is equivalent to that obtained
2620 // for all other roots of the PHI traversal. Also, when choosing one
2621 // such root as representative, select the one with the most uses in order
2622 // to keep the cost modeling heuristics in AddressingModeMatcher
2624 unsigned NumUses = V->getNumUses();
2625 if (NumUses > NumUsesConsensus) {
2627 NumUsesConsensus = NumUses;
2628 AddrModeInsts = NewAddrModeInsts;
2633 Consensus = nullptr;
2637 // If the addressing mode couldn't be determined, or if multiple different
2638 // ones were determined, bail out now.
2640 TPT.rollback(LastKnownGood);
2645 // Check to see if any of the instructions supersumed by this addr mode are
2646 // non-local to I's BB.
2647 bool AnyNonLocal = false;
2648 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
2649 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
2655 // If all the instructions matched are already in this BB, don't do anything.
2657 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
2661 // Insert this computation right after this user. Since our caller is
2662 // scanning from the top of the BB to the bottom, reuse of the expr are
2663 // guaranteed to happen later.
2664 IRBuilder<> Builder(MemoryInst);
2666 // Now that we determined the addressing expression we want to use and know
2667 // that we have to sink it into this block. Check to see if we have already
2668 // done this for some other load/store instr in this block. If so, reuse the
2670 Value *&SunkAddr = SunkAddrs[Addr];
2672 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
2673 << *MemoryInst << "\n");
2674 if (SunkAddr->getType() != Addr->getType())
2675 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
2676 } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() &&
2677 TM && TM->getSubtarget<TargetSubtargetInfo>().useAA())) {
2678 // By default, we use the GEP-based method when AA is used later. This
2679 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
2680 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
2681 << *MemoryInst << "\n");
2682 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
2683 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
2685 // First, find the pointer.
2686 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
2687 ResultPtr = AddrMode.BaseReg;
2688 AddrMode.BaseReg = nullptr;
2691 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
2692 // We can't add more than one pointer together, nor can we scale a
2693 // pointer (both of which seem meaningless).
2694 if (ResultPtr || AddrMode.Scale != 1)
2697 ResultPtr = AddrMode.ScaledReg;
2701 if (AddrMode.BaseGV) {
2705 ResultPtr = AddrMode.BaseGV;
2708 // If the real base value actually came from an inttoptr, then the matcher
2709 // will look through it and provide only the integer value. In that case,
2711 if (!ResultPtr && AddrMode.BaseReg) {
2713 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
2714 AddrMode.BaseReg = nullptr;
2715 } else if (!ResultPtr && AddrMode.Scale == 1) {
2717 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
2722 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
2723 SunkAddr = Constant::getNullValue(Addr->getType());
2724 } else if (!ResultPtr) {
2728 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
2730 // Start with the base register. Do this first so that subsequent address
2731 // matching finds it last, which will prevent it from trying to match it
2732 // as the scaled value in case it happens to be a mul. That would be
2733 // problematic if we've sunk a different mul for the scale, because then
2734 // we'd end up sinking both muls.
2735 if (AddrMode.BaseReg) {
2736 Value *V = AddrMode.BaseReg;
2737 if (V->getType() != IntPtrTy)
2738 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
2743 // Add the scale value.
2744 if (AddrMode.Scale) {
2745 Value *V = AddrMode.ScaledReg;
2746 if (V->getType() == IntPtrTy) {
2748 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
2749 cast<IntegerType>(V->getType())->getBitWidth()) {
2750 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
2752 // It is only safe to sign extend the BaseReg if we know that the math
2753 // required to create it did not overflow before we extend it. Since
2754 // the original IR value was tossed in favor of a constant back when
2755 // the AddrMode was created we need to bail out gracefully if widths
2756 // do not match instead of extending it.
2757 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
2758 if (I && (ResultIndex != AddrMode.BaseReg))
2759 I->eraseFromParent();
2763 if (AddrMode.Scale != 1)
2764 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
2767 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
2772 // Add in the Base Offset if present.
2773 if (AddrMode.BaseOffs) {
2774 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
2776 // We need to add this separately from the scale above to help with
2777 // SDAG consecutive load/store merging.
2778 if (ResultPtr->getType() != I8PtrTy)
2779 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
2780 ResultPtr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
2787 SunkAddr = ResultPtr;
2789 if (ResultPtr->getType() != I8PtrTy)
2790 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
2791 SunkAddr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
2794 if (SunkAddr->getType() != Addr->getType())
2795 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
2798 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
2799 << *MemoryInst << "\n");
2800 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
2801 Value *Result = nullptr;
2803 // Start with the base register. Do this first so that subsequent address
2804 // matching finds it last, which will prevent it from trying to match it
2805 // as the scaled value in case it happens to be a mul. That would be
2806 // problematic if we've sunk a different mul for the scale, because then
2807 // we'd end up sinking both muls.
2808 if (AddrMode.BaseReg) {
2809 Value *V = AddrMode.BaseReg;
2810 if (V->getType()->isPointerTy())
2811 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2812 if (V->getType() != IntPtrTy)
2813 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
2817 // Add the scale value.
2818 if (AddrMode.Scale) {
2819 Value *V = AddrMode.ScaledReg;
2820 if (V->getType() == IntPtrTy) {
2822 } else if (V->getType()->isPointerTy()) {
2823 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2824 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
2825 cast<IntegerType>(V->getType())->getBitWidth()) {
2826 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
2828 // It is only safe to sign extend the BaseReg if we know that the math
2829 // required to create it did not overflow before we extend it. Since
2830 // the original IR value was tossed in favor of a constant back when
2831 // the AddrMode was created we need to bail out gracefully if widths
2832 // do not match instead of extending it.
2833 Instruction *I = dyn_cast_or_null<Instruction>(Result);
2834 if (I && (Result != AddrMode.BaseReg))
2835 I->eraseFromParent();
2838 if (AddrMode.Scale != 1)
2839 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
2842 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2847 // Add in the BaseGV if present.
2848 if (AddrMode.BaseGV) {
2849 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
2851 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2856 // Add in the Base Offset if present.
2857 if (AddrMode.BaseOffs) {
2858 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
2860 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2866 SunkAddr = Constant::getNullValue(Addr->getType());
2868 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
2871 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
2873 // If we have no uses, recursively delete the value and all dead instructions
2875 if (Repl->use_empty()) {
2876 // This can cause recursive deletion, which can invalidate our iterator.
2877 // Use a WeakVH to hold onto it in case this happens.
2878 WeakVH IterHandle(CurInstIterator);
2879 BasicBlock *BB = CurInstIterator->getParent();
2881 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
2883 if (IterHandle != CurInstIterator) {
2884 // If the iterator instruction was recursively deleted, start over at the
2885 // start of the block.
2886 CurInstIterator = BB->begin();
2894 /// OptimizeInlineAsmInst - If there are any memory operands, use
2895 /// OptimizeMemoryInst to sink their address computing into the block when
2896 /// possible / profitable.
2897 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
2898 bool MadeChange = false;
2900 TargetLowering::AsmOperandInfoVector
2901 TargetConstraints = TLI->ParseConstraints(CS);
2903 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2904 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2906 // Compute the constraint code and ConstraintType to use.
2907 TLI->ComputeConstraintToUse(OpInfo, SDValue());
2909 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
2910 OpInfo.isIndirect) {
2911 Value *OpVal = CS->getArgOperand(ArgNo++);
2912 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
2913 } else if (OpInfo.Type == InlineAsm::isInput)
2920 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
2921 /// basic block as the load, unless conditions are unfavorable. This allows
2922 /// SelectionDAG to fold the extend into the load.
2924 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
2925 // Look for a load being extended.
2926 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
2927 if (!LI) return false;
2929 // If they're already in the same block, there's nothing to do.
2930 if (LI->getParent() == I->getParent())
2933 // If the load has other users and the truncate is not free, this probably
2934 // isn't worthwhile.
2935 if (!LI->hasOneUse() &&
2936 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
2937 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
2938 !TLI->isTruncateFree(I->getType(), LI->getType()))
2941 // Check whether the target supports casts folded into loads.
2943 if (isa<ZExtInst>(I))
2944 LType = ISD::ZEXTLOAD;
2946 assert(isa<SExtInst>(I) && "Unexpected ext type!");
2947 LType = ISD::SEXTLOAD;
2949 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
2952 // Move the extend into the same block as the load, so that SelectionDAG
2954 I->removeFromParent();
2960 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
2961 BasicBlock *DefBB = I->getParent();
2963 // If the result of a {s|z}ext and its source are both live out, rewrite all
2964 // other uses of the source with result of extension.
2965 Value *Src = I->getOperand(0);
2966 if (Src->hasOneUse())
2969 // Only do this xform if truncating is free.
2970 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
2973 // Only safe to perform the optimization if the source is also defined in
2975 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
2978 bool DefIsLiveOut = false;
2979 for (User *U : I->users()) {
2980 Instruction *UI = cast<Instruction>(U);
2982 // Figure out which BB this ext is used in.
2983 BasicBlock *UserBB = UI->getParent();
2984 if (UserBB == DefBB) continue;
2985 DefIsLiveOut = true;
2991 // Make sure none of the uses are PHI nodes.
2992 for (User *U : Src->users()) {
2993 Instruction *UI = cast<Instruction>(U);
2994 BasicBlock *UserBB = UI->getParent();
2995 if (UserBB == DefBB) continue;
2996 // Be conservative. We don't want this xform to end up introducing
2997 // reloads just before load / store instructions.
2998 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
3002 // InsertedTruncs - Only insert one trunc in each block once.
3003 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
3005 bool MadeChange = false;
3006 for (Use &U : Src->uses()) {
3007 Instruction *User = cast<Instruction>(U.getUser());
3009 // Figure out which BB this ext is used in.
3010 BasicBlock *UserBB = User->getParent();
3011 if (UserBB == DefBB) continue;
3013 // Both src and def are live in this block. Rewrite the use.
3014 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
3016 if (!InsertedTrunc) {
3017 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3018 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
3019 InsertedTruncsSet.insert(InsertedTrunc);
3022 // Replace a use of the {s|z}ext source with a use of the result.
3031 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
3032 /// turned into an explicit branch.
3033 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
3034 // FIXME: This should use the same heuristics as IfConversion to determine
3035 // whether a select is better represented as a branch. This requires that
3036 // branch probability metadata is preserved for the select, which is not the
3039 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3041 // If the branch is predicted right, an out of order CPU can avoid blocking on
3042 // the compare. Emit cmovs on compares with a memory operand as branches to
3043 // avoid stalls on the load from memory. If the compare has more than one use
3044 // there's probably another cmov or setcc around so it's not worth emitting a
3049 Value *CmpOp0 = Cmp->getOperand(0);
3050 Value *CmpOp1 = Cmp->getOperand(1);
3052 // We check that the memory operand has one use to avoid uses of the loaded
3053 // value directly after the compare, making branches unprofitable.
3054 return Cmp->hasOneUse() &&
3055 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3056 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
3060 /// If we have a SelectInst that will likely profit from branch prediction,
3061 /// turn it into a branch.
3062 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
3063 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3065 // Can we convert the 'select' to CF ?
3066 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3069 TargetLowering::SelectSupportKind SelectKind;
3071 SelectKind = TargetLowering::VectorMaskSelect;
3072 else if (SI->getType()->isVectorTy())
3073 SelectKind = TargetLowering::ScalarCondVectorVal;
3075 SelectKind = TargetLowering::ScalarValSelect;
3077 // Do we have efficient codegen support for this kind of 'selects' ?
3078 if (TLI->isSelectSupported(SelectKind)) {
3079 // We have efficient codegen support for the select instruction.
3080 // Check if it is profitable to keep this 'select'.
3081 if (!TLI->isPredictableSelectExpensive() ||
3082 !isFormingBranchFromSelectProfitable(SI))
3088 // First, we split the block containing the select into 2 blocks.
3089 BasicBlock *StartBlock = SI->getParent();
3090 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3091 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3093 // Create a new block serving as the landing pad for the branch.
3094 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3095 NextBlock->getParent(), NextBlock);
3097 // Move the unconditional branch from the block with the select in it into our
3098 // landing pad block.
3099 StartBlock->getTerminator()->eraseFromParent();
3100 BranchInst::Create(NextBlock, SmallBlock);
3102 // Insert the real conditional branch based on the original condition.
3103 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3105 // The select itself is replaced with a PHI Node.
3106 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3108 PN->addIncoming(SI->getTrueValue(), StartBlock);
3109 PN->addIncoming(SI->getFalseValue(), SmallBlock);
3110 SI->replaceAllUsesWith(PN);
3111 SI->eraseFromParent();
3113 // Instruct OptimizeBlock to skip to the next block.
3114 CurInstIterator = StartBlock->end();
3115 ++NumSelectsExpanded;
3119 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3120 SmallVector<int, 16> Mask(SVI->getShuffleMask());
3122 for (unsigned i = 0; i < Mask.size(); ++i) {
3123 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3125 SplatElem = Mask[i];
3131 /// Some targets have expensive vector shifts if the lanes aren't all the same
3132 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3133 /// it's often worth sinking a shufflevector splat down to its use so that
3134 /// codegen can spot all lanes are identical.
3135 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3136 BasicBlock *DefBB = SVI->getParent();
3138 // Only do this xform if variable vector shifts are particularly expensive.
3139 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3142 // We only expect better codegen by sinking a shuffle if we can recognise a
3144 if (!isBroadcastShuffle(SVI))
3147 // InsertedShuffles - Only insert a shuffle in each block once.
3148 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3150 bool MadeChange = false;
3151 for (User *U : SVI->users()) {
3152 Instruction *UI = cast<Instruction>(U);
3154 // Figure out which BB this ext is used in.
3155 BasicBlock *UserBB = UI->getParent();
3156 if (UserBB == DefBB) continue;
3158 // For now only apply this when the splat is used by a shift instruction.
3159 if (!UI->isShift()) continue;
3161 // Everything checks out, sink the shuffle if the user's block doesn't
3162 // already have a copy.
3163 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3165 if (!InsertedShuffle) {
3166 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3167 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3169 SVI->getOperand(2), "", InsertPt);
3172 UI->replaceUsesOfWith(SVI, InsertedShuffle);
3176 // If we removed all uses, nuke the shuffle.
3177 if (SVI->use_empty()) {
3178 SVI->eraseFromParent();
3186 /// \brief Helper class to promote a scalar operation to a vector one.
3187 /// This class is used to move downward extractelement transition.
3189 /// a = vector_op <2 x i32>
3190 /// b = extractelement <2 x i32> a, i32 0
3195 /// a = vector_op <2 x i32>
3196 /// c = vector_op a (equivalent to scalar_op on the related lane)
3197 /// * d = extractelement <2 x i32> c, i32 0
3199 /// Assuming both extractelement and store can be combine, we get rid of the
3201 class VectorPromoteHelper {
3202 /// Used to perform some checks on the legality of vector operations.
3203 const TargetLowering &TLI;
3205 /// Used to estimated the cost of the promoted chain.
3206 const TargetTransformInfo &TTI;
3208 /// The transition being moved downwards.
3209 Instruction *Transition;
3210 /// The sequence of instructions to be promoted.
3211 SmallVector<Instruction *, 4> InstsToBePromoted;
3212 /// Cost of combining a store and an extract.
3213 unsigned StoreExtractCombineCost;
3214 /// Instruction that will be combined with the transition.
3215 Instruction *CombineInst;
3217 /// \brief The instruction that represents the current end of the transition.
3218 /// Since we are faking the promotion until we reach the end of the chain
3219 /// of computation, we need a way to get the current end of the transition.
3220 Instruction *getEndOfTransition() const {
3221 if (InstsToBePromoted.empty())
3223 return InstsToBePromoted.back();
3226 /// \brief Return the index of the original value in the transition.
3227 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
3228 /// c, is at index 0.
3229 unsigned getTransitionOriginalValueIdx() const {
3230 assert(isa<ExtractElementInst>(Transition) &&
3231 "Other kind of transitions are not supported yet");
3235 /// \brief Return the index of the index in the transition.
3236 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
3238 unsigned getTransitionIdx() const {
3239 assert(isa<ExtractElementInst>(Transition) &&
3240 "Other kind of transitions are not supported yet");
3244 /// \brief Get the type of the transition.
3245 /// This is the type of the original value.
3246 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
3247 /// transition is <2 x i32>.
3248 Type *getTransitionType() const {
3249 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
3252 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
3253 /// I.e., we have the following sequence:
3254 /// Def = Transition <ty1> a to <ty2>
3255 /// b = ToBePromoted <ty2> Def, ...
3257 /// b = ToBePromoted <ty1> a, ...
3258 /// Def = Transition <ty1> ToBePromoted to <ty2>
3259 void promoteImpl(Instruction *ToBePromoted);
3261 /// \brief Check whether or not it is profitable to promote all the
3262 /// instructions enqueued to be promoted.
3263 bool isProfitableToPromote() {
3264 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
3265 unsigned Index = isa<ConstantInt>(ValIdx)
3266 ? cast<ConstantInt>(ValIdx)->getZExtValue()
3268 Type *PromotedType = getTransitionType();
3270 StoreInst *ST = cast<StoreInst>(CombineInst);
3271 unsigned AS = ST->getPointerAddressSpace();
3272 unsigned Align = ST->getAlignment();
3273 // Check if this store is supported.
3274 if (!TLI.allowsMisalignedMemoryAccesses(
3275 EVT::getEVT(ST->getValueOperand()->getType()), AS, Align)) {
3276 // If this is not supported, there is no way we can combine
3277 // the extract with the store.
3281 // The scalar chain of computation has to pay for the transition
3282 // scalar to vector.
3283 // The vector chain has to account for the combining cost.
3284 uint64_t ScalarCost =
3285 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
3286 uint64_t VectorCost = StoreExtractCombineCost;
3287 for (const auto &Inst : InstsToBePromoted) {
3288 // Compute the cost.
3289 // By construction, all instructions being promoted are arithmetic ones.
3290 // Moreover, one argument is a constant that can be viewed as a splat
3292 Value *Arg0 = Inst->getOperand(0);
3293 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
3294 isa<ConstantFP>(Arg0);
3295 TargetTransformInfo::OperandValueKind Arg0OVK =
3296 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3297 : TargetTransformInfo::OK_AnyValue;
3298 TargetTransformInfo::OperandValueKind Arg1OVK =
3299 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3300 : TargetTransformInfo::OK_AnyValue;
3301 ScalarCost += TTI.getArithmeticInstrCost(
3302 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
3303 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
3306 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
3307 << ScalarCost << "\nVector: " << VectorCost << '\n');
3308 return ScalarCost > VectorCost;
3311 /// \brief Generate a constant vector with \p Val with the same
3312 /// number of elements as the transition.
3313 /// \p UseSplat defines whether or not \p Val should be replicated
3314 /// accross the whole vector.
3315 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
3316 /// otherwise we generate a vector with as many undef as possible:
3317 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
3318 /// used at the index of the extract.
3319 Value *getConstantVector(Constant *Val, bool UseSplat) const {
3320 unsigned ExtractIdx = UINT_MAX;
3322 // If we cannot determine where the constant must be, we have to
3323 // use a splat constant.
3324 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
3325 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
3326 ExtractIdx = CstVal->getSExtValue();
3331 unsigned End = getTransitionType()->getVectorNumElements();
3333 return ConstantVector::getSplat(End, Val);
3335 SmallVector<Constant *, 4> ConstVec;
3336 UndefValue *UndefVal = UndefValue::get(Val->getType());
3337 for (unsigned Idx = 0; Idx != End; ++Idx) {
3338 if (Idx == ExtractIdx)
3339 ConstVec.push_back(Val);
3341 ConstVec.push_back(UndefVal);
3343 return ConstantVector::get(ConstVec);
3346 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
3347 /// in \p Use can trigger undefined behavior.
3348 static bool canCauseUndefinedBehavior(const Instruction *Use,
3349 unsigned OperandIdx) {
3350 // This is not safe to introduce undef when the operand is on
3351 // the right hand side of a division-like instruction.
3352 if (OperandIdx != 1)
3354 switch (Use->getOpcode()) {
3357 case Instruction::SDiv:
3358 case Instruction::UDiv:
3359 case Instruction::SRem:
3360 case Instruction::URem:
3362 case Instruction::FDiv:
3363 case Instruction::FRem:
3364 return !Use->hasNoNaNs();
3366 llvm_unreachable(nullptr);
3370 VectorPromoteHelper(const TargetLowering &TLI, const TargetTransformInfo &TTI,
3371 Instruction *Transition, unsigned CombineCost)
3372 : TLI(TLI), TTI(TTI), Transition(Transition),
3373 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
3374 assert(Transition && "Do not know how to promote null");
3377 /// \brief Check if we can promote \p ToBePromoted to \p Type.
3378 bool canPromote(const Instruction *ToBePromoted) const {
3379 // We could support CastInst too.
3380 return isa<BinaryOperator>(ToBePromoted);
3383 /// \brief Check if it is profitable to promote \p ToBePromoted
3384 /// by moving downward the transition through.
3385 bool shouldPromote(const Instruction *ToBePromoted) const {
3386 // Promote only if all the operands can be statically expanded.
3387 // Indeed, we do not want to introduce any new kind of transitions.
3388 for (const Use &U : ToBePromoted->operands()) {
3389 const Value *Val = U.get();
3390 if (Val == getEndOfTransition()) {
3391 // If the use is a division and the transition is on the rhs,
3392 // we cannot promote the operation, otherwise we may create a
3393 // division by zero.
3394 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
3398 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
3399 !isa<ConstantFP>(Val))
3402 // Check that the resulting operation is legal.
3403 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
3406 return StressStoreExtract ||
3407 TLI.isOperationLegalOrCustom(ISDOpcode,
3408 EVT::getEVT(getTransitionType(), true));
3411 /// \brief Check whether or not \p Use can be combined
3412 /// with the transition.
3413 /// I.e., is it possible to do Use(Transition) => AnotherUse?
3414 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
3416 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
3417 void enqueueForPromotion(Instruction *ToBePromoted) {
3418 InstsToBePromoted.push_back(ToBePromoted);
3421 /// \brief Set the instruction that will be combined with the transition.
3422 void recordCombineInstruction(Instruction *ToBeCombined) {
3423 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
3424 CombineInst = ToBeCombined;
3427 /// \brief Promote all the instructions enqueued for promotion if it is
3429 /// \return True if the promotion happened, false otherwise.
3431 // Check if there is something to promote.
3432 // Right now, if we do not have anything to combine with,
3433 // we assume the promotion is not profitable.
3434 if (InstsToBePromoted.empty() || !CombineInst)
3438 if (!StressStoreExtract && !isProfitableToPromote())
3442 for (auto &ToBePromoted : InstsToBePromoted)
3443 promoteImpl(ToBePromoted);
3444 InstsToBePromoted.clear();
3448 } // End of anonymous namespace.
3450 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
3451 // At this point, we know that all the operands of ToBePromoted but Def
3452 // can be statically promoted.
3453 // For Def, we need to use its parameter in ToBePromoted:
3454 // b = ToBePromoted ty1 a
3455 // Def = Transition ty1 b to ty2
3456 // Move the transition down.
3457 // 1. Replace all uses of the promoted operation by the transition.
3458 // = ... b => = ... Def.
3459 assert(ToBePromoted->getType() == Transition->getType() &&
3460 "The type of the result of the transition does not match "
3462 ToBePromoted->replaceAllUsesWith(Transition);
3463 // 2. Update the type of the uses.
3464 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
3465 Type *TransitionTy = getTransitionType();
3466 ToBePromoted->mutateType(TransitionTy);
3467 // 3. Update all the operands of the promoted operation with promoted
3469 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
3470 for (Use &U : ToBePromoted->operands()) {
3471 Value *Val = U.get();
3472 Value *NewVal = nullptr;
3473 if (Val == Transition)
3474 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
3475 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
3476 isa<ConstantFP>(Val)) {
3477 // Use a splat constant if it is not safe to use undef.
3478 NewVal = getConstantVector(
3479 cast<Constant>(Val),
3480 isa<UndefValue>(Val) ||
3481 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
3483 assert(0 && "Did you modified shouldPromote and forgot to update this?");
3484 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
3486 Transition->removeFromParent();
3487 Transition->insertAfter(ToBePromoted);
3488 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
3491 /// Some targets can do store(extractelement) with one instruction.
3492 /// Try to push the extractelement towards the stores when the target
3493 /// has this feature and this is profitable.
3494 bool CodeGenPrepare::OptimizeExtractElementInst(Instruction *Inst) {
3495 unsigned CombineCost = UINT_MAX;
3496 if (DisableStoreExtract || !TLI ||
3497 (!StressStoreExtract &&
3498 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
3499 Inst->getOperand(1), CombineCost)))
3502 // At this point we know that Inst is a vector to scalar transition.
3503 // Try to move it down the def-use chain, until:
3504 // - We can combine the transition with its single use
3505 // => we got rid of the transition.
3506 // - We escape the current basic block
3507 // => we would need to check that we are moving it at a cheaper place and
3508 // we do not do that for now.
3509 BasicBlock *Parent = Inst->getParent();
3510 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
3511 VectorPromoteHelper VPH(*TLI, *TTI, Inst, CombineCost);
3512 // If the transition has more than one use, assume this is not going to be
3514 while (Inst->hasOneUse()) {
3515 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
3516 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
3518 if (ToBePromoted->getParent() != Parent) {
3519 DEBUG(dbgs() << "Instruction to promote is in a different block ("
3520 << ToBePromoted->getParent()->getName()
3521 << ") than the transition (" << Parent->getName() << ").\n");
3525 if (VPH.canCombine(ToBePromoted)) {
3526 DEBUG(dbgs() << "Assume " << *Inst << '\n'
3527 << "will be combined with: " << *ToBePromoted << '\n');
3528 VPH.recordCombineInstruction(ToBePromoted);
3529 bool Changed = VPH.promote();
3530 NumStoreExtractExposed += Changed;
3534 DEBUG(dbgs() << "Try promoting.\n");
3535 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
3538 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
3540 VPH.enqueueForPromotion(ToBePromoted);
3541 Inst = ToBePromoted;
3546 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
3547 if (PHINode *P = dyn_cast<PHINode>(I)) {
3548 // It is possible for very late stage optimizations (such as SimplifyCFG)
3549 // to introduce PHI nodes too late to be cleaned up. If we detect such a
3550 // trivial PHI, go ahead and zap it here.
3551 if (Value *V = SimplifyInstruction(P, TLI ? TLI->getDataLayout() : nullptr,
3553 P->replaceAllUsesWith(V);
3554 P->eraseFromParent();
3561 if (CastInst *CI = dyn_cast<CastInst>(I)) {
3562 // If the source of the cast is a constant, then this should have
3563 // already been constant folded. The only reason NOT to constant fold
3564 // it is if something (e.g. LSR) was careful to place the constant
3565 // evaluation in a block other than then one that uses it (e.g. to hoist
3566 // the address of globals out of a loop). If this is the case, we don't
3567 // want to forward-subst the cast.
3568 if (isa<Constant>(CI->getOperand(0)))
3571 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
3574 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
3575 /// Sink a zext or sext into its user blocks if the target type doesn't
3576 /// fit in one register
3577 if (TLI && TLI->getTypeAction(CI->getContext(),
3578 TLI->getValueType(CI->getType())) ==
3579 TargetLowering::TypeExpandInteger) {
3580 return SinkCast(CI);
3582 bool MadeChange = MoveExtToFormExtLoad(I);
3583 return MadeChange | OptimizeExtUses(I);
3589 if (CmpInst *CI = dyn_cast<CmpInst>(I))
3590 if (!TLI || !TLI->hasMultipleConditionRegisters())
3591 return OptimizeCmpExpression(CI);
3593 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3595 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
3599 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3601 return OptimizeMemoryInst(I, SI->getOperand(1),
3602 SI->getOperand(0)->getType());
3606 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
3608 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
3609 BinOp->getOpcode() == Instruction::LShr)) {
3610 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
3611 if (TLI && CI && TLI->hasExtractBitsInsn())
3612 return OptimizeExtractBits(BinOp, CI, *TLI);
3617 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
3618 if (GEPI->hasAllZeroIndices()) {
3619 /// The GEP operand must be a pointer, so must its result -> BitCast
3620 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
3621 GEPI->getName(), GEPI);
3622 GEPI->replaceAllUsesWith(NC);
3623 GEPI->eraseFromParent();
3631 if (CallInst *CI = dyn_cast<CallInst>(I))
3632 return OptimizeCallInst(CI);
3634 if (SelectInst *SI = dyn_cast<SelectInst>(I))
3635 return OptimizeSelectInst(SI);
3637 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
3638 return OptimizeShuffleVectorInst(SVI);
3640 if (isa<ExtractElementInst>(I))
3641 return OptimizeExtractElementInst(I);
3646 // In this pass we look for GEP and cast instructions that are used
3647 // across basic blocks and rewrite them to improve basic-block-at-a-time
3649 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
3651 bool MadeChange = false;
3653 CurInstIterator = BB.begin();
3654 while (CurInstIterator != BB.end())
3655 MadeChange |= OptimizeInst(CurInstIterator++);
3657 MadeChange |= DupRetToEnableTailCallOpts(&BB);
3662 // llvm.dbg.value is far away from the value then iSel may not be able
3663 // handle it properly. iSel will drop llvm.dbg.value if it can not
3664 // find a node corresponding to the value.
3665 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
3666 bool MadeChange = false;
3667 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
3668 Instruction *PrevNonDbgInst = nullptr;
3669 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
3670 Instruction *Insn = BI; ++BI;
3671 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
3672 // Leave dbg.values that refer to an alloca alone. These
3673 // instrinsics describe the address of a variable (= the alloca)
3674 // being taken. They should not be moved next to the alloca
3675 // (and to the beginning of the scope), but rather stay close to
3676 // where said address is used.
3677 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
3678 PrevNonDbgInst = Insn;
3682 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
3683 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
3684 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
3685 DVI->removeFromParent();
3686 if (isa<PHINode>(VI))
3687 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
3689 DVI->insertAfter(VI);
3698 // If there is a sequence that branches based on comparing a single bit
3699 // against zero that can be combined into a single instruction, and the
3700 // target supports folding these into a single instruction, sink the
3701 // mask and compare into the branch uses. Do this before OptimizeBlock ->
3702 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
3704 bool CodeGenPrepare::sinkAndCmp(Function &F) {
3705 if (!EnableAndCmpSinking)
3707 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
3709 bool MadeChange = false;
3710 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
3711 BasicBlock *BB = I++;
3713 // Does this BB end with the following?
3714 // %andVal = and %val, #single-bit-set
3715 // %icmpVal = icmp %andResult, 0
3716 // br i1 %cmpVal label %dest1, label %dest2"
3717 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
3718 if (!Brcc || !Brcc->isConditional())
3720 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
3721 if (!Cmp || Cmp->getParent() != BB)
3723 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
3724 if (!Zero || !Zero->isZero())
3726 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
3727 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
3729 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
3730 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
3732 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
3734 // Push the "and; icmp" for any users that are conditional branches.
3735 // Since there can only be one branch use per BB, we don't need to keep
3736 // track of which BBs we insert into.
3737 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
3741 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
3743 if (!BrccUser || !BrccUser->isConditional())
3745 BasicBlock *UserBB = BrccUser->getParent();
3746 if (UserBB == BB) continue;
3747 DEBUG(dbgs() << "found Brcc use\n");
3749 // Sink the "and; icmp" to use.
3751 BinaryOperator *NewAnd =
3752 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
3755 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
3759 DEBUG(BrccUser->getParent()->dump());