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/MDBuilder.h"
34 #include "llvm/IR/PatternMatch.h"
35 #include "llvm/IR/ValueHandle.h"
36 #include "llvm/IR/ValueMap.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Target/TargetLibraryInfo.h"
42 #include "llvm/Target/TargetLowering.h"
43 #include "llvm/Target/TargetSubtargetInfo.h"
44 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
45 #include "llvm/Transforms/Utils/BuildLibCalls.h"
46 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
47 #include "llvm/Transforms/Utils/Local.h"
49 using namespace llvm::PatternMatch;
51 #define DEBUG_TYPE "codegenprepare"
53 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
54 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
55 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
56 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
58 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
60 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
61 "computations were sunk");
62 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
63 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
64 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
65 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
66 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
67 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
68 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
70 static cl::opt<bool> DisableBranchOpts(
71 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
72 cl::desc("Disable branch optimizations in CodeGenPrepare"));
74 static cl::opt<bool> DisableSelectToBranch(
75 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
76 cl::desc("Disable select to branch conversion."));
78 static cl::opt<bool> AddrSinkUsingGEPs(
79 "addr-sink-using-gep", cl::Hidden, cl::init(false),
80 cl::desc("Address sinking in CGP using GEPs."));
82 static cl::opt<bool> EnableAndCmpSinking(
83 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
84 cl::desc("Enable sinkinig and/cmp into branches."));
86 static cl::opt<bool> DisableStoreExtract(
87 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
88 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
90 static cl::opt<bool> StressStoreExtract(
91 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
92 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
94 static cl::opt<bool> DisableExtLdPromotion(
95 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
96 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
99 static cl::opt<bool> StressExtLdPromotion(
100 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
101 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
102 "optimization in CodeGenPrepare"));
105 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
109 TypeIsSExt(Type *Ty, bool IsSExt) : Ty(Ty), IsSExt(IsSExt) {}
111 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
112 class TypePromotionTransaction;
114 class CodeGenPrepare : public FunctionPass {
115 /// TLI - Keep a pointer of a TargetLowering to consult for determining
116 /// transformation profitability.
117 const TargetMachine *TM;
118 const TargetLowering *TLI;
119 const TargetTransformInfo *TTI;
120 const TargetLibraryInfo *TLInfo;
123 /// CurInstIterator - As we scan instructions optimizing them, this is the
124 /// next instruction to optimize. Xforms that can invalidate this should
126 BasicBlock::iterator CurInstIterator;
128 /// Keeps track of non-local addresses that have been sunk into a block.
129 /// This allows us to avoid inserting duplicate code for blocks with
130 /// multiple load/stores of the same address.
131 ValueMap<Value*, Value*> SunkAddrs;
133 /// Keeps track of all truncates inserted for the current function.
134 SetOfInstrs InsertedTruncsSet;
135 /// Keeps track of the type of the related instruction before their
136 /// promotion for the current function.
137 InstrToOrigTy PromotedInsts;
139 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
143 /// OptSize - True if optimizing for size.
147 static char ID; // Pass identification, replacement for typeid
148 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
149 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr) {
150 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
152 bool runOnFunction(Function &F) override;
154 const char *getPassName() const override { return "CodeGen Prepare"; }
156 void getAnalysisUsage(AnalysisUsage &AU) const override {
157 AU.addPreserved<DominatorTreeWrapperPass>();
158 AU.addRequired<TargetLibraryInfo>();
159 AU.addRequired<TargetTransformInfo>();
163 bool EliminateFallThrough(Function &F);
164 bool EliminateMostlyEmptyBlocks(Function &F);
165 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
166 void EliminateMostlyEmptyBlock(BasicBlock *BB);
167 bool OptimizeBlock(BasicBlock &BB, bool& ModifiedDT);
168 bool OptimizeInst(Instruction *I, bool& ModifiedDT);
169 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
170 bool OptimizeInlineAsmInst(CallInst *CS);
171 bool OptimizeCallInst(CallInst *CI, bool& ModifiedDT);
172 bool MoveExtToFormExtLoad(Instruction *&I);
173 bool OptimizeExtUses(Instruction *I);
174 bool OptimizeSelectInst(SelectInst *SI);
175 bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
176 bool OptimizeExtractElementInst(Instruction *Inst);
177 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
178 bool PlaceDbgValues(Function &F);
179 bool sinkAndCmp(Function &F);
180 bool ExtLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
182 const SmallVectorImpl<Instruction *> &Exts,
183 unsigned CreatedInst);
184 bool splitBranchCondition(Function &F);
188 char CodeGenPrepare::ID = 0;
189 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
190 "Optimize for code generation", false, false)
192 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
193 return new CodeGenPrepare(TM);
196 bool CodeGenPrepare::runOnFunction(Function &F) {
197 if (skipOptnoneFunction(F))
200 bool EverMadeChange = false;
201 // Clear per function information.
202 InsertedTruncsSet.clear();
203 PromotedInsts.clear();
207 TLI = TM->getSubtargetImpl()->getTargetLowering();
208 TLInfo = &getAnalysis<TargetLibraryInfo>();
209 TTI = &getAnalysis<TargetTransformInfo>();
210 DominatorTreeWrapperPass *DTWP =
211 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
212 DT = DTWP ? &DTWP->getDomTree() : nullptr;
213 OptSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
214 Attribute::OptimizeForSize);
216 /// This optimization identifies DIV instructions that can be
217 /// profitably bypassed and carried out with a shorter, faster divide.
218 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
219 const DenseMap<unsigned int, unsigned int> &BypassWidths =
220 TLI->getBypassSlowDivWidths();
221 for (Function::iterator I = F.begin(); I != F.end(); I++)
222 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
225 // Eliminate blocks that contain only PHI nodes and an
226 // unconditional branch.
227 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
229 // llvm.dbg.value is far away from the value then iSel may not be able
230 // handle it properly. iSel will drop llvm.dbg.value if it can not
231 // find a node corresponding to the value.
232 EverMadeChange |= PlaceDbgValues(F);
234 // If there is a mask, compare against zero, and branch that can be combined
235 // into a single target instruction, push the mask and compare into branch
236 // users. Do this before OptimizeBlock -> OptimizeInst ->
237 // OptimizeCmpExpression, which perturbs the pattern being searched for.
238 if (!DisableBranchOpts) {
239 EverMadeChange |= sinkAndCmp(F);
240 EverMadeChange |= splitBranchCondition(F);
243 bool MadeChange = true;
246 for (Function::iterator I = F.begin(); I != F.end(); ) {
247 BasicBlock *BB = I++;
248 bool ModifiedDTOnIteration = false;
249 MadeChange |= OptimizeBlock(*BB, ModifiedDTOnIteration);
251 // Restart BB iteration if the dominator tree of the Function was changed
252 ModifiedDT |= ModifiedDTOnIteration;
253 if (ModifiedDTOnIteration)
256 EverMadeChange |= MadeChange;
261 if (!DisableBranchOpts) {
263 SmallPtrSet<BasicBlock*, 8> WorkList;
264 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
265 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
266 MadeChange |= ConstantFoldTerminator(BB, true);
267 if (!MadeChange) continue;
269 for (SmallVectorImpl<BasicBlock*>::iterator
270 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
271 if (pred_begin(*II) == pred_end(*II))
272 WorkList.insert(*II);
275 // Delete the dead blocks and any of their dead successors.
276 MadeChange |= !WorkList.empty();
277 while (!WorkList.empty()) {
278 BasicBlock *BB = *WorkList.begin();
280 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
284 for (SmallVectorImpl<BasicBlock*>::iterator
285 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
286 if (pred_begin(*II) == pred_end(*II))
287 WorkList.insert(*II);
290 // Merge pairs of basic blocks with unconditional branches, connected by
292 if (EverMadeChange || MadeChange)
293 MadeChange |= EliminateFallThrough(F);
297 EverMadeChange |= MadeChange;
300 if (ModifiedDT && DT)
303 return EverMadeChange;
306 /// EliminateFallThrough - Merge basic blocks which are connected
307 /// by a single edge, where one of the basic blocks has a single successor
308 /// pointing to the other basic block, which has a single predecessor.
309 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
310 bool Changed = false;
311 // Scan all of the blocks in the function, except for the entry block.
312 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
313 BasicBlock *BB = I++;
314 // If the destination block has a single pred, then this is a trivial
315 // edge, just collapse it.
316 BasicBlock *SinglePred = BB->getSinglePredecessor();
318 // Don't merge if BB's address is taken.
319 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
321 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
322 if (Term && !Term->isConditional()) {
324 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
325 // Remember if SinglePred was the entry block of the function.
326 // If so, we will need to move BB back to the entry position.
327 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
328 MergeBasicBlockIntoOnlyPred(BB, this);
330 if (isEntry && BB != &BB->getParent()->getEntryBlock())
331 BB->moveBefore(&BB->getParent()->getEntryBlock());
333 // We have erased a block. Update the iterator.
340 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
341 /// debug info directives, and an unconditional branch. Passes before isel
342 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
343 /// isel. Start by eliminating these blocks so we can split them the way we
345 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
346 bool MadeChange = false;
347 // Note that this intentionally skips the entry block.
348 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
349 BasicBlock *BB = I++;
351 // If this block doesn't end with an uncond branch, ignore it.
352 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
353 if (!BI || !BI->isUnconditional())
356 // If the instruction before the branch (skipping debug info) isn't a phi
357 // node, then other stuff is happening here.
358 BasicBlock::iterator BBI = BI;
359 if (BBI != BB->begin()) {
361 while (isa<DbgInfoIntrinsic>(BBI)) {
362 if (BBI == BB->begin())
366 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
370 // Do not break infinite loops.
371 BasicBlock *DestBB = BI->getSuccessor(0);
375 if (!CanMergeBlocks(BB, DestBB))
378 EliminateMostlyEmptyBlock(BB);
384 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
385 /// single uncond branch between them, and BB contains no other non-phi
387 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
388 const BasicBlock *DestBB) const {
389 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
390 // the successor. If there are more complex condition (e.g. preheaders),
391 // don't mess around with them.
392 BasicBlock::const_iterator BBI = BB->begin();
393 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
394 for (const User *U : PN->users()) {
395 const Instruction *UI = cast<Instruction>(U);
396 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
398 // If User is inside DestBB block and it is a PHINode then check
399 // incoming value. If incoming value is not from BB then this is
400 // a complex condition (e.g. preheaders) we want to avoid here.
401 if (UI->getParent() == DestBB) {
402 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
403 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
404 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
405 if (Insn && Insn->getParent() == BB &&
406 Insn->getParent() != UPN->getIncomingBlock(I))
413 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
414 // and DestBB may have conflicting incoming values for the block. If so, we
415 // can't merge the block.
416 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
417 if (!DestBBPN) return true; // no conflict.
419 // Collect the preds of BB.
420 SmallPtrSet<const BasicBlock*, 16> BBPreds;
421 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
422 // It is faster to get preds from a PHI than with pred_iterator.
423 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
424 BBPreds.insert(BBPN->getIncomingBlock(i));
426 BBPreds.insert(pred_begin(BB), pred_end(BB));
429 // Walk the preds of DestBB.
430 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
431 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
432 if (BBPreds.count(Pred)) { // Common predecessor?
433 BBI = DestBB->begin();
434 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
435 const Value *V1 = PN->getIncomingValueForBlock(Pred);
436 const Value *V2 = PN->getIncomingValueForBlock(BB);
438 // If V2 is a phi node in BB, look up what the mapped value will be.
439 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
440 if (V2PN->getParent() == BB)
441 V2 = V2PN->getIncomingValueForBlock(Pred);
443 // If there is a conflict, bail out.
444 if (V1 != V2) return false;
453 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
454 /// an unconditional branch in it.
455 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
456 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
457 BasicBlock *DestBB = BI->getSuccessor(0);
459 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
461 // If the destination block has a single pred, then this is a trivial edge,
463 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
464 if (SinglePred != DestBB) {
465 // Remember if SinglePred was the entry block of the function. If so, we
466 // will need to move BB back to the entry position.
467 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
468 MergeBasicBlockIntoOnlyPred(DestBB, this);
470 if (isEntry && BB != &BB->getParent()->getEntryBlock())
471 BB->moveBefore(&BB->getParent()->getEntryBlock());
473 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
478 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
479 // to handle the new incoming edges it is about to have.
481 for (BasicBlock::iterator BBI = DestBB->begin();
482 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
483 // Remove the incoming value for BB, and remember it.
484 Value *InVal = PN->removeIncomingValue(BB, false);
486 // Two options: either the InVal is a phi node defined in BB or it is some
487 // value that dominates BB.
488 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
489 if (InValPhi && InValPhi->getParent() == BB) {
490 // Add all of the input values of the input PHI as inputs of this phi.
491 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
492 PN->addIncoming(InValPhi->getIncomingValue(i),
493 InValPhi->getIncomingBlock(i));
495 // Otherwise, add one instance of the dominating value for each edge that
496 // we will be adding.
497 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
498 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
499 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
501 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
502 PN->addIncoming(InVal, *PI);
507 // The PHIs are now updated, change everything that refers to BB to use
508 // DestBB and remove BB.
509 BB->replaceAllUsesWith(DestBB);
510 if (DT && !ModifiedDT) {
511 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
512 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
513 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
514 DT->changeImmediateDominator(DestBB, NewIDom);
517 BB->eraseFromParent();
520 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
523 /// SinkCast - Sink the specified cast instruction into its user blocks
524 static bool SinkCast(CastInst *CI) {
525 BasicBlock *DefBB = CI->getParent();
527 /// InsertedCasts - Only insert a cast in each block once.
528 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
530 bool MadeChange = false;
531 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
533 Use &TheUse = UI.getUse();
534 Instruction *User = cast<Instruction>(*UI);
536 // Figure out which BB this cast is used in. For PHI's this is the
537 // appropriate predecessor block.
538 BasicBlock *UserBB = User->getParent();
539 if (PHINode *PN = dyn_cast<PHINode>(User)) {
540 UserBB = PN->getIncomingBlock(TheUse);
543 // Preincrement use iterator so we don't invalidate it.
546 // If this user is in the same block as the cast, don't change the cast.
547 if (UserBB == DefBB) continue;
549 // If we have already inserted a cast into this block, use it.
550 CastInst *&InsertedCast = InsertedCasts[UserBB];
553 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
555 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
560 // Replace a use of the cast with a use of the new cast.
561 TheUse = InsertedCast;
565 // If we removed all uses, nuke the cast.
566 if (CI->use_empty()) {
567 CI->eraseFromParent();
574 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
575 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
576 /// sink it into user blocks to reduce the number of virtual
577 /// registers that must be created and coalesced.
579 /// Return true if any changes are made.
581 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
582 // If this is a noop copy,
583 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
584 EVT DstVT = TLI.getValueType(CI->getType());
586 // This is an fp<->int conversion?
587 if (SrcVT.isInteger() != DstVT.isInteger())
590 // If this is an extension, it will be a zero or sign extension, which
592 if (SrcVT.bitsLT(DstVT)) return false;
594 // If these values will be promoted, find out what they will be promoted
595 // to. This helps us consider truncates on PPC as noop copies when they
597 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
598 TargetLowering::TypePromoteInteger)
599 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
600 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
601 TargetLowering::TypePromoteInteger)
602 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
604 // If, after promotion, these are the same types, this is a noop copy.
611 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
612 /// the number of virtual registers that must be created and coalesced. This is
613 /// a clear win except on targets with multiple condition code registers
614 /// (PowerPC), where it might lose; some adjustment may be wanted there.
616 /// Return true if any changes are made.
617 static bool OptimizeCmpExpression(CmpInst *CI) {
618 BasicBlock *DefBB = CI->getParent();
620 /// InsertedCmp - Only insert a cmp in each block once.
621 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
623 bool MadeChange = false;
624 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
626 Use &TheUse = UI.getUse();
627 Instruction *User = cast<Instruction>(*UI);
629 // Preincrement use iterator so we don't invalidate it.
632 // Don't bother for PHI nodes.
633 if (isa<PHINode>(User))
636 // Figure out which BB this cmp is used in.
637 BasicBlock *UserBB = User->getParent();
639 // If this user is in the same block as the cmp, don't change the cmp.
640 if (UserBB == DefBB) continue;
642 // If we have already inserted a cmp into this block, use it.
643 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
646 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
648 CmpInst::Create(CI->getOpcode(),
649 CI->getPredicate(), CI->getOperand(0),
650 CI->getOperand(1), "", InsertPt);
654 // Replace a use of the cmp with a use of the new cmp.
655 TheUse = InsertedCmp;
659 // If we removed all uses, nuke the cmp.
661 CI->eraseFromParent();
666 /// isExtractBitsCandidateUse - Check if the candidates could
667 /// be combined with shift instruction, which includes:
668 /// 1. Truncate instruction
669 /// 2. And instruction and the imm is a mask of the low bits:
670 /// imm & (imm+1) == 0
671 static bool isExtractBitsCandidateUse(Instruction *User) {
672 if (!isa<TruncInst>(User)) {
673 if (User->getOpcode() != Instruction::And ||
674 !isa<ConstantInt>(User->getOperand(1)))
677 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
679 if ((Cimm & (Cimm + 1)).getBoolValue())
685 /// SinkShiftAndTruncate - sink both shift and truncate instruction
686 /// to the use of truncate's BB.
688 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
689 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
690 const TargetLowering &TLI) {
691 BasicBlock *UserBB = User->getParent();
692 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
693 TruncInst *TruncI = dyn_cast<TruncInst>(User);
694 bool MadeChange = false;
696 for (Value::user_iterator TruncUI = TruncI->user_begin(),
697 TruncE = TruncI->user_end();
698 TruncUI != TruncE;) {
700 Use &TruncTheUse = TruncUI.getUse();
701 Instruction *TruncUser = cast<Instruction>(*TruncUI);
702 // Preincrement use iterator so we don't invalidate it.
706 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
710 // If the use is actually a legal node, there will not be an
711 // implicit truncate.
712 // FIXME: always querying the result type is just an
713 // approximation; some nodes' legality is determined by the
714 // operand or other means. There's no good way to find out though.
715 if (TLI.isOperationLegalOrCustom(
716 ISDOpcode, TLI.getValueType(TruncUser->getType(), true)))
719 // Don't bother for PHI nodes.
720 if (isa<PHINode>(TruncUser))
723 BasicBlock *TruncUserBB = TruncUser->getParent();
725 if (UserBB == TruncUserBB)
728 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
729 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
731 if (!InsertedShift && !InsertedTrunc) {
732 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
734 if (ShiftI->getOpcode() == Instruction::AShr)
736 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
739 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
742 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
745 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
746 TruncI->getType(), "", TruncInsertPt);
750 TruncTheUse = InsertedTrunc;
756 /// OptimizeExtractBits - sink the shift *right* instruction into user blocks if
757 /// the uses could potentially be combined with this shift instruction and
758 /// generate BitExtract instruction. It will only be applied if the architecture
759 /// supports BitExtract instruction. Here is an example:
761 /// %x.extract.shift = lshr i64 %arg1, 32
763 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
767 /// %x.extract.shift.1 = lshr i64 %arg1, 32
768 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
770 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
772 /// Return true if any changes are made.
773 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
774 const TargetLowering &TLI) {
775 BasicBlock *DefBB = ShiftI->getParent();
777 /// Only insert instructions in each block once.
778 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
780 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(ShiftI->getType()));
782 bool MadeChange = false;
783 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
785 Use &TheUse = UI.getUse();
786 Instruction *User = cast<Instruction>(*UI);
787 // Preincrement use iterator so we don't invalidate it.
790 // Don't bother for PHI nodes.
791 if (isa<PHINode>(User))
794 if (!isExtractBitsCandidateUse(User))
797 BasicBlock *UserBB = User->getParent();
799 if (UserBB == DefBB) {
800 // If the shift and truncate instruction are in the same BB. The use of
801 // the truncate(TruncUse) may still introduce another truncate if not
802 // legal. In this case, we would like to sink both shift and truncate
803 // instruction to the BB of TruncUse.
806 // i64 shift.result = lshr i64 opnd, imm
807 // trunc.result = trunc shift.result to i16
810 // ----> We will have an implicit truncate here if the architecture does
811 // not have i16 compare.
812 // cmp i16 trunc.result, opnd2
814 if (isa<TruncInst>(User) && shiftIsLegal
815 // If the type of the truncate is legal, no trucate will be
816 // introduced in other basic blocks.
817 && (!TLI.isTypeLegal(TLI.getValueType(User->getType()))))
819 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI);
823 // If we have already inserted a shift into this block, use it.
824 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
826 if (!InsertedShift) {
827 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
829 if (ShiftI->getOpcode() == Instruction::AShr)
831 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
834 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
839 // Replace a use of the shift with a use of the new shift.
840 TheUse = InsertedShift;
843 // If we removed all uses, nuke the shift.
844 if (ShiftI->use_empty())
845 ShiftI->eraseFromParent();
851 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
853 void replaceCall(Value *With) override {
854 CI->replaceAllUsesWith(With);
855 CI->eraseFromParent();
857 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const override {
858 if (ConstantInt *SizeCI =
859 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
860 return SizeCI->isAllOnesValue();
864 } // end anonymous namespace
866 // ScalarizeMaskedLoad() translates masked load intrinsic, like
867 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
868 // <16 x i1> %mask, <16 x i32> %passthru)
869 // to a chain of basic blocks, whith loading element one-by-one if
870 // the appropriate mask bit is set
872 // %1 = bitcast i8* %addr to i32*
873 // %2 = extractelement <16 x i1> %mask, i32 0
874 // %3 = icmp eq i1 %2, true
875 // br i1 %3, label %cond.load, label %else
877 //cond.load: ; preds = %0
878 // %4 = getelementptr i32* %1, i32 0
880 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
883 //else: ; preds = %0, %cond.load
884 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
885 // %7 = extractelement <16 x i1> %mask, i32 1
886 // %8 = icmp eq i1 %7, true
887 // br i1 %8, label %cond.load1, label %else2
889 //cond.load1: ; preds = %else
890 // %9 = getelementptr i32* %1, i32 1
891 // %10 = load i32* %9
892 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
895 //else2: ; preds = %else, %cond.load1
896 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
897 // %12 = extractelement <16 x i1> %mask, i32 2
898 // %13 = icmp eq i1 %12, true
899 // br i1 %13, label %cond.load4, label %else5
901 static void ScalarizeMaskedLoad(CallInst *CI) {
902 Value *Ptr = CI->getArgOperand(0);
903 Value *Src0 = CI->getArgOperand(3);
904 Value *Mask = CI->getArgOperand(2);
905 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
906 Type *EltTy = VecType->getElementType();
908 assert(VecType && "Unexpected return type of masked load intrinsic");
910 IRBuilder<> Builder(CI->getContext());
911 Instruction *InsertPt = CI;
912 BasicBlock *IfBlock = CI->getParent();
913 BasicBlock *CondBlock = nullptr;
914 BasicBlock *PrevIfBlock = CI->getParent();
915 Builder.SetInsertPoint(InsertPt);
917 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
919 // Bitcast %addr fron i8* to EltTy*
921 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
922 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
923 Value *UndefVal = UndefValue::get(VecType);
926 Value *VResult = UndefVal;
928 PHINode *Phi = nullptr;
929 Value *PrevPhi = UndefVal;
931 unsigned VectorWidth = VecType->getNumElements();
932 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
934 // Fill the "else" block, created in the previous iteration
936 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
937 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
938 // %to_load = icmp eq i1 %mask_1, true
939 // br i1 %to_load, label %cond.load, label %else
942 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
943 Phi->addIncoming(VResult, CondBlock);
944 Phi->addIncoming(PrevPhi, PrevIfBlock);
949 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
950 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
951 ConstantInt::get(Predicate->getType(), 1));
953 // Create "cond" block
955 // %EltAddr = getelementptr i32* %1, i32 0
956 // %Elt = load i32* %EltAddr
957 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
959 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
960 Builder.SetInsertPoint(InsertPt);
962 Value* Gep = Builder.CreateInBoundsGEP(FirstEltPtr, Builder.getInt32(Idx));
963 LoadInst* Load = Builder.CreateLoad(Gep, false);
964 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
966 // Create "else" block, fill it in the next iteration
967 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
968 Builder.SetInsertPoint(InsertPt);
969 Instruction *OldBr = IfBlock->getTerminator();
970 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
971 OldBr->eraseFromParent();
972 PrevIfBlock = IfBlock;
973 IfBlock = NewIfBlock;
976 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
977 Phi->addIncoming(VResult, CondBlock);
978 Phi->addIncoming(PrevPhi, PrevIfBlock);
979 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
980 CI->replaceAllUsesWith(NewI);
981 CI->eraseFromParent();
984 // ScalarizeMaskedStore() translates masked store intrinsic, like
985 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
987 // to a chain of basic blocks, that stores element one-by-one if
988 // the appropriate mask bit is set
990 // %1 = bitcast i8* %addr to i32*
991 // %2 = extractelement <16 x i1> %mask, i32 0
992 // %3 = icmp eq i1 %2, true
993 // br i1 %3, label %cond.store, label %else
995 // cond.store: ; preds = %0
996 // %4 = extractelement <16 x i32> %val, i32 0
997 // %5 = getelementptr i32* %1, i32 0
998 // store i32 %4, i32* %5
1001 // else: ; preds = %0, %cond.store
1002 // %6 = extractelement <16 x i1> %mask, i32 1
1003 // %7 = icmp eq i1 %6, true
1004 // br i1 %7, label %cond.store1, label %else2
1006 // cond.store1: ; preds = %else
1007 // %8 = extractelement <16 x i32> %val, i32 1
1008 // %9 = getelementptr i32* %1, i32 1
1009 // store i32 %8, i32* %9
1012 static void ScalarizeMaskedStore(CallInst *CI) {
1013 Value *Ptr = CI->getArgOperand(1);
1014 Value *Src = CI->getArgOperand(0);
1015 Value *Mask = CI->getArgOperand(3);
1017 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1018 Type *EltTy = VecType->getElementType();
1020 assert(VecType && "Unexpected data type in masked store intrinsic");
1022 IRBuilder<> Builder(CI->getContext());
1023 Instruction *InsertPt = CI;
1024 BasicBlock *IfBlock = CI->getParent();
1025 Builder.SetInsertPoint(InsertPt);
1026 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1028 // Bitcast %addr fron i8* to EltTy*
1030 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1031 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1033 unsigned VectorWidth = VecType->getNumElements();
1034 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1036 // Fill the "else" block, created in the previous iteration
1038 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1039 // %to_store = icmp eq i1 %mask_1, true
1040 // br i1 %to_load, label %cond.store, label %else
1042 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1043 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1044 ConstantInt::get(Predicate->getType(), 1));
1046 // Create "cond" block
1048 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1049 // %EltAddr = getelementptr i32* %1, i32 0
1050 // %store i32 %OneElt, i32* %EltAddr
1052 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1053 Builder.SetInsertPoint(InsertPt);
1055 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1056 Value* Gep = Builder.CreateInBoundsGEP(FirstEltPtr, Builder.getInt32(Idx));
1057 Builder.CreateStore(OneElt, Gep);
1059 // Create "else" block, fill it in the next iteration
1060 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1061 Builder.SetInsertPoint(InsertPt);
1062 Instruction *OldBr = IfBlock->getTerminator();
1063 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1064 OldBr->eraseFromParent();
1065 IfBlock = NewIfBlock;
1067 CI->eraseFromParent();
1070 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1071 BasicBlock *BB = CI->getParent();
1073 // Lower inline assembly if we can.
1074 // If we found an inline asm expession, and if the target knows how to
1075 // lower it to normal LLVM code, do so now.
1076 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1077 if (TLI->ExpandInlineAsm(CI)) {
1078 // Avoid invalidating the iterator.
1079 CurInstIterator = BB->begin();
1080 // Avoid processing instructions out of order, which could cause
1081 // reuse before a value is defined.
1085 // Sink address computing for memory operands into the block.
1086 if (OptimizeInlineAsmInst(CI))
1090 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1092 switch (II->getIntrinsicID()) {
1094 case Intrinsic::objectsize: {
1095 // Lower all uses of llvm.objectsize.*
1096 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1097 Type *ReturnTy = CI->getType();
1098 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1100 // Substituting this can cause recursive simplifications, which can
1101 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1103 WeakVH IterHandle(CurInstIterator);
1105 replaceAndRecursivelySimplify(CI, RetVal,
1106 TLI ? TLI->getDataLayout() : nullptr,
1107 TLInfo, ModifiedDT ? nullptr : DT);
1109 // If the iterator instruction was recursively deleted, start over at the
1110 // start of the block.
1111 if (IterHandle != CurInstIterator) {
1112 CurInstIterator = BB->begin();
1117 case Intrinsic::masked_load: {
1118 // Scalarize unsupported vector masked load
1119 if (!TTI->isLegalMaskedLoad(CI->getType(), 1)) {
1120 ScalarizeMaskedLoad(CI);
1126 case Intrinsic::masked_store: {
1127 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType(), 1)) {
1128 ScalarizeMaskedStore(CI);
1137 SmallVector<Value*, 2> PtrOps;
1139 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
1140 while (!PtrOps.empty())
1141 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
1146 // From here on out we're working with named functions.
1147 if (!CI->getCalledFunction()) return false;
1149 // We'll need DataLayout from here on out.
1150 const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
1151 if (!TD) return false;
1153 // Lower all default uses of _chk calls. This is very similar
1154 // to what InstCombineCalls does, but here we are only lowering calls
1155 // that have the default "don't know" as the objectsize. Anything else
1156 // should be left alone.
1157 CodeGenPrepareFortifiedLibCalls Simplifier;
1158 return Simplifier.fold(CI, TD, TLInfo);
1161 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
1162 /// instructions to the predecessor to enable tail call optimizations. The
1163 /// case it is currently looking for is:
1166 /// %tmp0 = tail call i32 @f0()
1167 /// br label %return
1169 /// %tmp1 = tail call i32 @f1()
1170 /// br label %return
1172 /// %tmp2 = tail call i32 @f2()
1173 /// br label %return
1175 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1183 /// %tmp0 = tail call i32 @f0()
1186 /// %tmp1 = tail call i32 @f1()
1189 /// %tmp2 = tail call i32 @f2()
1192 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
1196 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1200 PHINode *PN = nullptr;
1201 BitCastInst *BCI = nullptr;
1202 Value *V = RI->getReturnValue();
1204 BCI = dyn_cast<BitCastInst>(V);
1206 V = BCI->getOperand(0);
1208 PN = dyn_cast<PHINode>(V);
1213 if (PN && PN->getParent() != BB)
1216 // It's not safe to eliminate the sign / zero extension of the return value.
1217 // See llvm::isInTailCallPosition().
1218 const Function *F = BB->getParent();
1219 AttributeSet CallerAttrs = F->getAttributes();
1220 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1221 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1224 // Make sure there are no instructions between the PHI and return, or that the
1225 // return is the first instruction in the block.
1227 BasicBlock::iterator BI = BB->begin();
1228 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1230 // Also skip over the bitcast.
1235 BasicBlock::iterator BI = BB->begin();
1236 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1241 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1243 SmallVector<CallInst*, 4> TailCalls;
1245 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1246 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1247 // Make sure the phi value is indeed produced by the tail call.
1248 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1249 TLI->mayBeEmittedAsTailCall(CI))
1250 TailCalls.push_back(CI);
1253 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1254 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1255 if (!VisitedBBs.insert(*PI).second)
1258 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1259 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1260 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1261 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1265 CallInst *CI = dyn_cast<CallInst>(&*RI);
1266 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1267 TailCalls.push_back(CI);
1271 bool Changed = false;
1272 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1273 CallInst *CI = TailCalls[i];
1276 // Conservatively require the attributes of the call to match those of the
1277 // return. Ignore noalias because it doesn't affect the call sequence.
1278 AttributeSet CalleeAttrs = CS.getAttributes();
1279 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1280 removeAttribute(Attribute::NoAlias) !=
1281 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1282 removeAttribute(Attribute::NoAlias))
1285 // Make sure the call instruction is followed by an unconditional branch to
1286 // the return block.
1287 BasicBlock *CallBB = CI->getParent();
1288 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1289 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1292 // Duplicate the return into CallBB.
1293 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1294 ModifiedDT = Changed = true;
1298 // If we eliminated all predecessors of the block, delete the block now.
1299 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1300 BB->eraseFromParent();
1305 //===----------------------------------------------------------------------===//
1306 // Memory Optimization
1307 //===----------------------------------------------------------------------===//
1311 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1312 /// which holds actual Value*'s for register values.
1313 struct ExtAddrMode : public TargetLowering::AddrMode {
1316 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1317 void print(raw_ostream &OS) const;
1320 bool operator==(const ExtAddrMode& O) const {
1321 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1322 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1323 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1328 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1334 void ExtAddrMode::print(raw_ostream &OS) const {
1335 bool NeedPlus = false;
1338 OS << (NeedPlus ? " + " : "")
1340 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1345 OS << (NeedPlus ? " + " : "")
1351 OS << (NeedPlus ? " + " : "")
1353 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1357 OS << (NeedPlus ? " + " : "")
1359 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1365 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1366 void ExtAddrMode::dump() const {
1372 /// \brief This class provides transaction based operation on the IR.
1373 /// Every change made through this class is recorded in the internal state and
1374 /// can be undone (rollback) until commit is called.
1375 class TypePromotionTransaction {
1377 /// \brief This represents the common interface of the individual transaction.
1378 /// Each class implements the logic for doing one specific modification on
1379 /// the IR via the TypePromotionTransaction.
1380 class TypePromotionAction {
1382 /// The Instruction modified.
1386 /// \brief Constructor of the action.
1387 /// The constructor performs the related action on the IR.
1388 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1390 virtual ~TypePromotionAction() {}
1392 /// \brief Undo the modification done by this action.
1393 /// When this method is called, the IR must be in the same state as it was
1394 /// before this action was applied.
1395 /// \pre Undoing the action works if and only if the IR is in the exact same
1396 /// state as it was directly after this action was applied.
1397 virtual void undo() = 0;
1399 /// \brief Advocate every change made by this action.
1400 /// When the results on the IR of the action are to be kept, it is important
1401 /// to call this function, otherwise hidden information may be kept forever.
1402 virtual void commit() {
1403 // Nothing to be done, this action is not doing anything.
1407 /// \brief Utility to remember the position of an instruction.
1408 class InsertionHandler {
1409 /// Position of an instruction.
1410 /// Either an instruction:
1411 /// - Is the first in a basic block: BB is used.
1412 /// - Has a previous instructon: PrevInst is used.
1414 Instruction *PrevInst;
1417 /// Remember whether or not the instruction had a previous instruction.
1418 bool HasPrevInstruction;
1421 /// \brief Record the position of \p Inst.
1422 InsertionHandler(Instruction *Inst) {
1423 BasicBlock::iterator It = Inst;
1424 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1425 if (HasPrevInstruction)
1426 Point.PrevInst = --It;
1428 Point.BB = Inst->getParent();
1431 /// \brief Insert \p Inst at the recorded position.
1432 void insert(Instruction *Inst) {
1433 if (HasPrevInstruction) {
1434 if (Inst->getParent())
1435 Inst->removeFromParent();
1436 Inst->insertAfter(Point.PrevInst);
1438 Instruction *Position = Point.BB->getFirstInsertionPt();
1439 if (Inst->getParent())
1440 Inst->moveBefore(Position);
1442 Inst->insertBefore(Position);
1447 /// \brief Move an instruction before another.
1448 class InstructionMoveBefore : public TypePromotionAction {
1449 /// Original position of the instruction.
1450 InsertionHandler Position;
1453 /// \brief Move \p Inst before \p Before.
1454 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1455 : TypePromotionAction(Inst), Position(Inst) {
1456 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1457 Inst->moveBefore(Before);
1460 /// \brief Move the instruction back to its original position.
1461 void undo() override {
1462 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1463 Position.insert(Inst);
1467 /// \brief Set the operand of an instruction with a new value.
1468 class OperandSetter : public TypePromotionAction {
1469 /// Original operand of the instruction.
1471 /// Index of the modified instruction.
1475 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1476 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1477 : TypePromotionAction(Inst), Idx(Idx) {
1478 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1479 << "for:" << *Inst << "\n"
1480 << "with:" << *NewVal << "\n");
1481 Origin = Inst->getOperand(Idx);
1482 Inst->setOperand(Idx, NewVal);
1485 /// \brief Restore the original value of the instruction.
1486 void undo() override {
1487 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1488 << "for: " << *Inst << "\n"
1489 << "with: " << *Origin << "\n");
1490 Inst->setOperand(Idx, Origin);
1494 /// \brief Hide the operands of an instruction.
1495 /// Do as if this instruction was not using any of its operands.
1496 class OperandsHider : public TypePromotionAction {
1497 /// The list of original operands.
1498 SmallVector<Value *, 4> OriginalValues;
1501 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1502 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1503 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1504 unsigned NumOpnds = Inst->getNumOperands();
1505 OriginalValues.reserve(NumOpnds);
1506 for (unsigned It = 0; It < NumOpnds; ++It) {
1507 // Save the current operand.
1508 Value *Val = Inst->getOperand(It);
1509 OriginalValues.push_back(Val);
1511 // We could use OperandSetter here, but that would implied an overhead
1512 // that we are not willing to pay.
1513 Inst->setOperand(It, UndefValue::get(Val->getType()));
1517 /// \brief Restore the original list of uses.
1518 void undo() override {
1519 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1520 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1521 Inst->setOperand(It, OriginalValues[It]);
1525 /// \brief Build a truncate instruction.
1526 class TruncBuilder : public TypePromotionAction {
1529 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1531 /// trunc Opnd to Ty.
1532 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1533 IRBuilder<> Builder(Opnd);
1534 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
1535 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
1538 /// \brief Get the built value.
1539 Value *getBuiltValue() { return Val; }
1541 /// \brief Remove the built instruction.
1542 void undo() override {
1543 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
1544 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1545 IVal->eraseFromParent();
1549 /// \brief Build a sign extension instruction.
1550 class SExtBuilder : public TypePromotionAction {
1553 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1555 /// sext Opnd to Ty.
1556 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1557 : TypePromotionAction(InsertPt) {
1558 IRBuilder<> Builder(InsertPt);
1559 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
1560 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
1563 /// \brief Get the built value.
1564 Value *getBuiltValue() { return Val; }
1566 /// \brief Remove the built instruction.
1567 void undo() override {
1568 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
1569 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1570 IVal->eraseFromParent();
1574 /// \brief Build a zero extension instruction.
1575 class ZExtBuilder : public TypePromotionAction {
1578 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
1580 /// zext Opnd to Ty.
1581 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1582 : TypePromotionAction(InsertPt) {
1583 IRBuilder<> Builder(InsertPt);
1584 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
1585 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
1588 /// \brief Get the built value.
1589 Value *getBuiltValue() { return Val; }
1591 /// \brief Remove the built instruction.
1592 void undo() override {
1593 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
1594 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1595 IVal->eraseFromParent();
1599 /// \brief Mutate an instruction to another type.
1600 class TypeMutator : public TypePromotionAction {
1601 /// Record the original type.
1605 /// \brief Mutate the type of \p Inst into \p NewTy.
1606 TypeMutator(Instruction *Inst, Type *NewTy)
1607 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1608 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1610 Inst->mutateType(NewTy);
1613 /// \brief Mutate the instruction back to its original type.
1614 void undo() override {
1615 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1617 Inst->mutateType(OrigTy);
1621 /// \brief Replace the uses of an instruction by another instruction.
1622 class UsesReplacer : public TypePromotionAction {
1623 /// Helper structure to keep track of the replaced uses.
1624 struct InstructionAndIdx {
1625 /// The instruction using the instruction.
1627 /// The index where this instruction is used for Inst.
1629 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1630 : Inst(Inst), Idx(Idx) {}
1633 /// Keep track of the original uses (pair Instruction, Index).
1634 SmallVector<InstructionAndIdx, 4> OriginalUses;
1635 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1638 /// \brief Replace all the use of \p Inst by \p New.
1639 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1640 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1642 // Record the original uses.
1643 for (Use &U : Inst->uses()) {
1644 Instruction *UserI = cast<Instruction>(U.getUser());
1645 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1647 // Now, we can replace the uses.
1648 Inst->replaceAllUsesWith(New);
1651 /// \brief Reassign the original uses of Inst to Inst.
1652 void undo() override {
1653 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1654 for (use_iterator UseIt = OriginalUses.begin(),
1655 EndIt = OriginalUses.end();
1656 UseIt != EndIt; ++UseIt) {
1657 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1662 /// \brief Remove an instruction from the IR.
1663 class InstructionRemover : public TypePromotionAction {
1664 /// Original position of the instruction.
1665 InsertionHandler Inserter;
1666 /// Helper structure to hide all the link to the instruction. In other
1667 /// words, this helps to do as if the instruction was removed.
1668 OperandsHider Hider;
1669 /// Keep track of the uses replaced, if any.
1670 UsesReplacer *Replacer;
1673 /// \brief Remove all reference of \p Inst and optinally replace all its
1675 /// \pre If !Inst->use_empty(), then New != nullptr
1676 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1677 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1680 Replacer = new UsesReplacer(Inst, New);
1681 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1682 Inst->removeFromParent();
1685 ~InstructionRemover() { delete Replacer; }
1687 /// \brief Really remove the instruction.
1688 void commit() override { delete Inst; }
1690 /// \brief Resurrect the instruction and reassign it to the proper uses if
1691 /// new value was provided when build this action.
1692 void undo() override {
1693 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1694 Inserter.insert(Inst);
1702 /// Restoration point.
1703 /// The restoration point is a pointer to an action instead of an iterator
1704 /// because the iterator may be invalidated but not the pointer.
1705 typedef const TypePromotionAction *ConstRestorationPt;
1706 /// Advocate every changes made in that transaction.
1708 /// Undo all the changes made after the given point.
1709 void rollback(ConstRestorationPt Point);
1710 /// Get the current restoration point.
1711 ConstRestorationPt getRestorationPoint() const;
1713 /// \name API for IR modification with state keeping to support rollback.
1715 /// Same as Instruction::setOperand.
1716 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1717 /// Same as Instruction::eraseFromParent.
1718 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
1719 /// Same as Value::replaceAllUsesWith.
1720 void replaceAllUsesWith(Instruction *Inst, Value *New);
1721 /// Same as Value::mutateType.
1722 void mutateType(Instruction *Inst, Type *NewTy);
1723 /// Same as IRBuilder::createTrunc.
1724 Value *createTrunc(Instruction *Opnd, Type *Ty);
1725 /// Same as IRBuilder::createSExt.
1726 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1727 /// Same as IRBuilder::createZExt.
1728 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
1729 /// Same as Instruction::moveBefore.
1730 void moveBefore(Instruction *Inst, Instruction *Before);
1734 /// The ordered list of actions made so far.
1735 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
1736 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
1739 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1742 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
1745 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1748 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
1751 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1753 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
1756 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1757 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
1760 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1762 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
1763 Value *Val = Ptr->getBuiltValue();
1764 Actions.push_back(std::move(Ptr));
1768 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
1769 Value *Opnd, Type *Ty) {
1770 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
1771 Value *Val = Ptr->getBuiltValue();
1772 Actions.push_back(std::move(Ptr));
1776 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
1777 Value *Opnd, Type *Ty) {
1778 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
1779 Value *Val = Ptr->getBuiltValue();
1780 Actions.push_back(std::move(Ptr));
1784 void TypePromotionTransaction::moveBefore(Instruction *Inst,
1785 Instruction *Before) {
1787 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
1790 TypePromotionTransaction::ConstRestorationPt
1791 TypePromotionTransaction::getRestorationPoint() const {
1792 return !Actions.empty() ? Actions.back().get() : nullptr;
1795 void TypePromotionTransaction::commit() {
1796 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
1802 void TypePromotionTransaction::rollback(
1803 TypePromotionTransaction::ConstRestorationPt Point) {
1804 while (!Actions.empty() && Point != Actions.back().get()) {
1805 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
1810 /// \brief A helper class for matching addressing modes.
1812 /// This encapsulates the logic for matching the target-legal addressing modes.
1813 class AddressingModeMatcher {
1814 SmallVectorImpl<Instruction*> &AddrModeInsts;
1815 const TargetLowering &TLI;
1817 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
1818 /// the memory instruction that we're computing this address for.
1820 Instruction *MemoryInst;
1822 /// AddrMode - This is the addressing mode that we're building up. This is
1823 /// part of the return value of this addressing mode matching stuff.
1824 ExtAddrMode &AddrMode;
1826 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
1827 const SetOfInstrs &InsertedTruncs;
1828 /// A map from the instructions to their type before promotion.
1829 InstrToOrigTy &PromotedInsts;
1830 /// The ongoing transaction where every action should be registered.
1831 TypePromotionTransaction &TPT;
1833 /// IgnoreProfitability - This is set to true when we should not do
1834 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
1835 /// always returns true.
1836 bool IgnoreProfitability;
1838 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
1839 const TargetLowering &T, Type *AT,
1840 Instruction *MI, ExtAddrMode &AM,
1841 const SetOfInstrs &InsertedTruncs,
1842 InstrToOrigTy &PromotedInsts,
1843 TypePromotionTransaction &TPT)
1844 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM),
1845 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
1846 IgnoreProfitability = false;
1850 /// Match - Find the maximal addressing mode that a load/store of V can fold,
1851 /// give an access type of AccessTy. This returns a list of involved
1852 /// instructions in AddrModeInsts.
1853 /// \p InsertedTruncs The truncate instruction inserted by other
1856 /// \p PromotedInsts maps the instructions to their type before promotion.
1857 /// \p The ongoing transaction where every action should be registered.
1858 static ExtAddrMode Match(Value *V, Type *AccessTy,
1859 Instruction *MemoryInst,
1860 SmallVectorImpl<Instruction*> &AddrModeInsts,
1861 const TargetLowering &TLI,
1862 const SetOfInstrs &InsertedTruncs,
1863 InstrToOrigTy &PromotedInsts,
1864 TypePromotionTransaction &TPT) {
1867 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
1868 MemoryInst, Result, InsertedTruncs,
1869 PromotedInsts, TPT).MatchAddr(V, 0);
1870 (void)Success; assert(Success && "Couldn't select *anything*?");
1874 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
1875 bool MatchAddr(Value *V, unsigned Depth);
1876 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
1877 bool *MovedAway = nullptr);
1878 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
1879 ExtAddrMode &AMBefore,
1880 ExtAddrMode &AMAfter);
1881 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
1882 bool IsPromotionProfitable(unsigned MatchedSize, unsigned SizeWithPromotion,
1883 Value *PromotedOperand) const;
1886 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
1887 /// Return true and update AddrMode if this addr mode is legal for the target,
1889 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
1891 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
1892 // mode. Just process that directly.
1894 return MatchAddr(ScaleReg, Depth);
1896 // If the scale is 0, it takes nothing to add this.
1900 // If we already have a scale of this value, we can add to it, otherwise, we
1901 // need an available scale field.
1902 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
1905 ExtAddrMode TestAddrMode = AddrMode;
1907 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
1908 // [A+B + A*7] -> [B+A*8].
1909 TestAddrMode.Scale += Scale;
1910 TestAddrMode.ScaledReg = ScaleReg;
1912 // If the new address isn't legal, bail out.
1913 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
1916 // It was legal, so commit it.
1917 AddrMode = TestAddrMode;
1919 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
1920 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
1921 // X*Scale + C*Scale to addr mode.
1922 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
1923 if (isa<Instruction>(ScaleReg) && // not a constant expr.
1924 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
1925 TestAddrMode.ScaledReg = AddLHS;
1926 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
1928 // If this addressing mode is legal, commit it and remember that we folded
1929 // this instruction.
1930 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
1931 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
1932 AddrMode = TestAddrMode;
1937 // Otherwise, not (x+c)*scale, just return what we have.
1941 /// MightBeFoldableInst - This is a little filter, which returns true if an
1942 /// addressing computation involving I might be folded into a load/store
1943 /// accessing it. This doesn't need to be perfect, but needs to accept at least
1944 /// the set of instructions that MatchOperationAddr can.
1945 static bool MightBeFoldableInst(Instruction *I) {
1946 switch (I->getOpcode()) {
1947 case Instruction::BitCast:
1948 case Instruction::AddrSpaceCast:
1949 // Don't touch identity bitcasts.
1950 if (I->getType() == I->getOperand(0)->getType())
1952 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
1953 case Instruction::PtrToInt:
1954 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1956 case Instruction::IntToPtr:
1957 // We know the input is intptr_t, so this is foldable.
1959 case Instruction::Add:
1961 case Instruction::Mul:
1962 case Instruction::Shl:
1963 // Can only handle X*C and X << C.
1964 return isa<ConstantInt>(I->getOperand(1));
1965 case Instruction::GetElementPtr:
1972 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
1973 /// \note \p Val is assumed to be the product of some type promotion.
1974 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
1975 /// to be legal, as the non-promoted value would have had the same state.
1976 static bool isPromotedInstructionLegal(const TargetLowering &TLI, Value *Val) {
1977 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
1980 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
1981 // If the ISDOpcode is undefined, it was undefined before the promotion.
1984 // Otherwise, check if the promoted instruction is legal or not.
1985 return TLI.isOperationLegalOrCustom(
1986 ISDOpcode, TLI.getValueType(PromotedInst->getType()));
1989 /// \brief Hepler class to perform type promotion.
1990 class TypePromotionHelper {
1991 /// \brief Utility function to check whether or not a sign or zero extension
1992 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
1993 /// either using the operands of \p Inst or promoting \p Inst.
1994 /// The type of the extension is defined by \p IsSExt.
1995 /// In other words, check if:
1996 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
1997 /// #1 Promotion applies:
1998 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
1999 /// #2 Operand reuses:
2000 /// ext opnd1 to ConsideredExtType.
2001 /// \p PromotedInsts maps the instructions to their type before promotion.
2002 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2003 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2005 /// \brief Utility function to determine if \p OpIdx should be promoted when
2006 /// promoting \p Inst.
2007 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2008 if (isa<SelectInst>(Inst) && OpIdx == 0)
2013 /// \brief Utility function to promote the operand of \p Ext when this
2014 /// operand is a promotable trunc or sext or zext.
2015 /// \p PromotedInsts maps the instructions to their type before promotion.
2016 /// \p CreatedInsts[out] contains how many non-free instructions have been
2017 /// created to promote the operand of Ext.
2018 /// Newly added extensions are inserted in \p Exts.
2019 /// Newly added truncates are inserted in \p Truncs.
2020 /// Should never be called directly.
2021 /// \return The promoted value which is used instead of Ext.
2022 static Value *promoteOperandForTruncAndAnyExt(
2023 Instruction *Ext, TypePromotionTransaction &TPT,
2024 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2025 SmallVectorImpl<Instruction *> *Exts,
2026 SmallVectorImpl<Instruction *> *Truncs);
2028 /// \brief Utility function to promote the operand of \p Ext when this
2029 /// operand is promotable and is not a supported trunc or sext.
2030 /// \p PromotedInsts maps the instructions to their type before promotion.
2031 /// \p CreatedInsts[out] contains how many non-free instructions have been
2032 /// created to promote the operand of Ext.
2033 /// Newly added extensions are inserted in \p Exts.
2034 /// Newly added truncates are inserted in \p Truncs.
2035 /// Should never be called directly.
2036 /// \return The promoted value which is used instead of Ext.
2038 promoteOperandForOther(Instruction *Ext, TypePromotionTransaction &TPT,
2039 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2040 SmallVectorImpl<Instruction *> *Exts,
2041 SmallVectorImpl<Instruction *> *Truncs, bool IsSExt);
2043 /// \see promoteOperandForOther.
2045 signExtendOperandForOther(Instruction *Ext, TypePromotionTransaction &TPT,
2046 InstrToOrigTy &PromotedInsts,
2047 unsigned &CreatedInsts,
2048 SmallVectorImpl<Instruction *> *Exts,
2049 SmallVectorImpl<Instruction *> *Truncs) {
2050 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInsts, Exts,
2054 /// \see promoteOperandForOther.
2056 zeroExtendOperandForOther(Instruction *Ext, TypePromotionTransaction &TPT,
2057 InstrToOrigTy &PromotedInsts,
2058 unsigned &CreatedInsts,
2059 SmallVectorImpl<Instruction *> *Exts,
2060 SmallVectorImpl<Instruction *> *Truncs) {
2061 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInsts, Exts,
2066 /// Type for the utility function that promotes the operand of Ext.
2067 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2068 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2069 SmallVectorImpl<Instruction *> *Exts,
2070 SmallVectorImpl<Instruction *> *Truncs);
2071 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2072 /// action to promote the operand of \p Ext instead of using Ext.
2073 /// \return NULL if no promotable action is possible with the current
2075 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
2076 /// the others CodeGenPrepare optimizations. This information is important
2077 /// because we do not want to promote these instructions as CodeGenPrepare
2078 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2079 /// \p PromotedInsts maps the instructions to their type before promotion.
2080 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2081 const TargetLowering &TLI,
2082 const InstrToOrigTy &PromotedInsts);
2085 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2086 Type *ConsideredExtType,
2087 const InstrToOrigTy &PromotedInsts,
2089 // The promotion helper does not know how to deal with vector types yet.
2090 // To be able to fix that, we would need to fix the places where we
2091 // statically extend, e.g., constants and such.
2092 if (Inst->getType()->isVectorTy())
2095 // We can always get through zext.
2096 if (isa<ZExtInst>(Inst))
2099 // sext(sext) is ok too.
2100 if (IsSExt && isa<SExtInst>(Inst))
2103 // We can get through binary operator, if it is legal. In other words, the
2104 // binary operator must have a nuw or nsw flag.
2105 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2106 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2107 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2108 (IsSExt && BinOp->hasNoSignedWrap())))
2111 // Check if we can do the following simplification.
2112 // ext(trunc(opnd)) --> ext(opnd)
2113 if (!isa<TruncInst>(Inst))
2116 Value *OpndVal = Inst->getOperand(0);
2117 // Check if we can use this operand in the extension.
2118 // If the type is larger than the result type of the extension,
2120 if (!OpndVal->getType()->isIntegerTy() ||
2121 OpndVal->getType()->getIntegerBitWidth() >
2122 ConsideredExtType->getIntegerBitWidth())
2125 // If the operand of the truncate is not an instruction, we will not have
2126 // any information on the dropped bits.
2127 // (Actually we could for constant but it is not worth the extra logic).
2128 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2132 // Check if the source of the type is narrow enough.
2133 // I.e., check that trunc just drops extended bits of the same kind of
2135 // #1 get the type of the operand and check the kind of the extended bits.
2136 const Type *OpndType;
2137 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2138 if (It != PromotedInsts.end() && It->second.IsSExt == IsSExt)
2139 OpndType = It->second.Ty;
2140 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2141 OpndType = Opnd->getOperand(0)->getType();
2145 // #2 check that the truncate just drop extended bits.
2146 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
2152 TypePromotionHelper::Action TypePromotionHelper::getAction(
2153 Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2154 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2155 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2156 "Unexpected instruction type");
2157 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2158 Type *ExtTy = Ext->getType();
2159 bool IsSExt = isa<SExtInst>(Ext);
2160 // If the operand of the extension is not an instruction, we cannot
2162 // If it, check we can get through.
2163 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2166 // Do not promote if the operand has been added by codegenprepare.
2167 // Otherwise, it means we are undoing an optimization that is likely to be
2168 // redone, thus causing potential infinite loop.
2169 if (isa<TruncInst>(ExtOpnd) && InsertedTruncs.count(ExtOpnd))
2172 // SExt or Trunc instructions.
2173 // Return the related handler.
2174 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2175 isa<ZExtInst>(ExtOpnd))
2176 return promoteOperandForTruncAndAnyExt;
2178 // Regular instruction.
2179 // Abort early if we will have to insert non-free instructions.
2180 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2182 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2185 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2186 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2187 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2188 SmallVectorImpl<Instruction *> *Exts,
2189 SmallVectorImpl<Instruction *> *Truncs) {
2190 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2191 // get through it and this method should not be called.
2192 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2193 Value *ExtVal = SExt;
2194 if (isa<ZExtInst>(SExtOpnd)) {
2195 // Replace s|zext(zext(opnd))
2198 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2199 TPT.replaceAllUsesWith(SExt, ZExt);
2200 TPT.eraseInstruction(SExt);
2203 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2205 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2209 // Remove dead code.
2210 if (SExtOpnd->use_empty())
2211 TPT.eraseInstruction(SExtOpnd);
2213 // Check if the extension is still needed.
2214 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2215 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2216 if (ExtInst && Exts)
2217 Exts->push_back(ExtInst);
2221 // At this point we have: ext ty opnd to ty.
2222 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2223 Value *NextVal = ExtInst->getOperand(0);
2224 TPT.eraseInstruction(ExtInst, NextVal);
2228 Value *TypePromotionHelper::promoteOperandForOther(
2229 Instruction *Ext, TypePromotionTransaction &TPT,
2230 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2231 SmallVectorImpl<Instruction *> *Exts,
2232 SmallVectorImpl<Instruction *> *Truncs, bool IsSExt) {
2233 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2234 // get through it and this method should not be called.
2235 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2237 if (!ExtOpnd->hasOneUse()) {
2238 // ExtOpnd will be promoted.
2239 // All its uses, but Ext, will need to use a truncated value of the
2240 // promoted version.
2241 // Create the truncate now.
2242 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2243 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2244 ITrunc->removeFromParent();
2245 // Insert it just after the definition.
2246 ITrunc->insertAfter(ExtOpnd);
2248 Truncs->push_back(ITrunc);
2251 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2252 // Restore the operand of Ext (which has been replace by the previous call
2253 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2254 TPT.setOperand(Ext, 0, ExtOpnd);
2257 // Get through the Instruction:
2258 // 1. Update its type.
2259 // 2. Replace the uses of Ext by Inst.
2260 // 3. Extend each operand that needs to be extended.
2262 // Remember the original type of the instruction before promotion.
2263 // This is useful to know that the high bits are sign extended bits.
2264 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2265 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2267 TPT.mutateType(ExtOpnd, Ext->getType());
2269 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2271 Instruction *ExtForOpnd = Ext;
2273 DEBUG(dbgs() << "Propagate Ext to operands\n");
2274 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2276 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2277 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2278 !shouldExtOperand(ExtOpnd, OpIdx)) {
2279 DEBUG(dbgs() << "No need to propagate\n");
2282 // Check if we can statically extend the operand.
2283 Value *Opnd = ExtOpnd->getOperand(OpIdx);
2284 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2285 DEBUG(dbgs() << "Statically extend\n");
2286 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2287 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
2288 : Cst->getValue().zext(BitWidth);
2289 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
2292 // UndefValue are typed, so we have to statically sign extend them.
2293 if (isa<UndefValue>(Opnd)) {
2294 DEBUG(dbgs() << "Statically extend\n");
2295 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2299 // Otherwise we have to explicity sign extend the operand.
2300 // Check if Ext was reused to extend an operand.
2302 // If yes, create a new one.
2303 DEBUG(dbgs() << "More operands to ext\n");
2304 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2305 : TPT.createZExt(Ext, Opnd, Ext->getType());
2306 if (!isa<Instruction>(ValForExtOpnd)) {
2307 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
2310 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
2314 Exts->push_back(ExtForOpnd);
2315 TPT.setOperand(ExtForOpnd, 0, Opnd);
2317 // Move the sign extension before the insertion point.
2318 TPT.moveBefore(ExtForOpnd, ExtOpnd);
2319 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2320 // If more sext are required, new instructions will have to be created.
2321 ExtForOpnd = nullptr;
2323 if (ExtForOpnd == Ext) {
2324 DEBUG(dbgs() << "Extension is useless now\n");
2325 TPT.eraseInstruction(Ext);
2330 /// IsPromotionProfitable - Check whether or not promoting an instruction
2331 /// to a wider type was profitable.
2332 /// \p MatchedSize gives the number of instructions that have been matched
2333 /// in the addressing mode after the promotion was applied.
2334 /// \p SizeWithPromotion gives the number of created instructions for
2335 /// the promotion plus the number of instructions that have been
2336 /// matched in the addressing mode before the promotion.
2337 /// \p PromotedOperand is the value that has been promoted.
2338 /// \return True if the promotion is profitable, false otherwise.
2340 AddressingModeMatcher::IsPromotionProfitable(unsigned MatchedSize,
2341 unsigned SizeWithPromotion,
2342 Value *PromotedOperand) const {
2343 // We folded less instructions than what we created to promote the operand.
2344 // This is not profitable.
2345 if (MatchedSize < SizeWithPromotion)
2347 if (MatchedSize > SizeWithPromotion)
2349 // The promotion is neutral but it may help folding the sign extension in
2350 // loads for instance.
2351 // Check that we did not create an illegal instruction.
2352 return isPromotedInstructionLegal(TLI, PromotedOperand);
2355 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
2356 /// fold the operation into the addressing mode. If so, update the addressing
2357 /// mode and return true, otherwise return false without modifying AddrMode.
2358 /// If \p MovedAway is not NULL, it contains the information of whether or
2359 /// not AddrInst has to be folded into the addressing mode on success.
2360 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2361 /// because it has been moved away.
2362 /// Thus AddrInst must not be added in the matched instructions.
2363 /// This state can happen when AddrInst is a sext, since it may be moved away.
2364 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2365 /// not be referenced anymore.
2366 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
2369 // Avoid exponential behavior on extremely deep expression trees.
2370 if (Depth >= 5) return false;
2372 // By default, all matched instructions stay in place.
2377 case Instruction::PtrToInt:
2378 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2379 return MatchAddr(AddrInst->getOperand(0), Depth);
2380 case Instruction::IntToPtr:
2381 // This inttoptr is a no-op if the integer type is pointer sized.
2382 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
2383 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
2384 return MatchAddr(AddrInst->getOperand(0), Depth);
2386 case Instruction::BitCast:
2387 case Instruction::AddrSpaceCast:
2388 // BitCast is always a noop, and we can handle it as long as it is
2389 // int->int or pointer->pointer (we don't want int<->fp or something).
2390 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2391 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2392 // Don't touch identity bitcasts. These were probably put here by LSR,
2393 // and we don't want to mess around with them. Assume it knows what it
2395 AddrInst->getOperand(0)->getType() != AddrInst->getType())
2396 return MatchAddr(AddrInst->getOperand(0), Depth);
2398 case Instruction::Add: {
2399 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2400 ExtAddrMode BackupAddrMode = AddrMode;
2401 unsigned OldSize = AddrModeInsts.size();
2402 // Start a transaction at this point.
2403 // The LHS may match but not the RHS.
2404 // Therefore, we need a higher level restoration point to undo partially
2405 // matched operation.
2406 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2407 TPT.getRestorationPoint();
2409 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2410 MatchAddr(AddrInst->getOperand(0), Depth+1))
2413 // Restore the old addr mode info.
2414 AddrMode = BackupAddrMode;
2415 AddrModeInsts.resize(OldSize);
2416 TPT.rollback(LastKnownGood);
2418 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2419 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2420 MatchAddr(AddrInst->getOperand(1), Depth+1))
2423 // Otherwise we definitely can't merge the ADD in.
2424 AddrMode = BackupAddrMode;
2425 AddrModeInsts.resize(OldSize);
2426 TPT.rollback(LastKnownGood);
2429 //case Instruction::Or:
2430 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2432 case Instruction::Mul:
2433 case Instruction::Shl: {
2434 // Can only handle X*C and X << C.
2435 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2438 int64_t Scale = RHS->getSExtValue();
2439 if (Opcode == Instruction::Shl)
2440 Scale = 1LL << Scale;
2442 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2444 case Instruction::GetElementPtr: {
2445 // Scan the GEP. We check it if it contains constant offsets and at most
2446 // one variable offset.
2447 int VariableOperand = -1;
2448 unsigned VariableScale = 0;
2450 int64_t ConstantOffset = 0;
2451 const DataLayout *TD = TLI.getDataLayout();
2452 gep_type_iterator GTI = gep_type_begin(AddrInst);
2453 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2454 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2455 const StructLayout *SL = TD->getStructLayout(STy);
2457 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2458 ConstantOffset += SL->getElementOffset(Idx);
2460 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
2461 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2462 ConstantOffset += CI->getSExtValue()*TypeSize;
2463 } else if (TypeSize) { // Scales of zero don't do anything.
2464 // We only allow one variable index at the moment.
2465 if (VariableOperand != -1)
2468 // Remember the variable index.
2469 VariableOperand = i;
2470 VariableScale = TypeSize;
2475 // A common case is for the GEP to only do a constant offset. In this case,
2476 // just add it to the disp field and check validity.
2477 if (VariableOperand == -1) {
2478 AddrMode.BaseOffs += ConstantOffset;
2479 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
2480 // Check to see if we can fold the base pointer in too.
2481 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2484 AddrMode.BaseOffs -= ConstantOffset;
2488 // Save the valid addressing mode in case we can't match.
2489 ExtAddrMode BackupAddrMode = AddrMode;
2490 unsigned OldSize = AddrModeInsts.size();
2492 // See if the scale and offset amount is valid for this target.
2493 AddrMode.BaseOffs += ConstantOffset;
2495 // Match the base operand of the GEP.
2496 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2497 // If it couldn't be matched, just stuff the value in a register.
2498 if (AddrMode.HasBaseReg) {
2499 AddrMode = BackupAddrMode;
2500 AddrModeInsts.resize(OldSize);
2503 AddrMode.HasBaseReg = true;
2504 AddrMode.BaseReg = AddrInst->getOperand(0);
2507 // Match the remaining variable portion of the GEP.
2508 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2510 // If it couldn't be matched, try stuffing the base into a register
2511 // instead of matching it, and retrying the match of the scale.
2512 AddrMode = BackupAddrMode;
2513 AddrModeInsts.resize(OldSize);
2514 if (AddrMode.HasBaseReg)
2516 AddrMode.HasBaseReg = true;
2517 AddrMode.BaseReg = AddrInst->getOperand(0);
2518 AddrMode.BaseOffs += ConstantOffset;
2519 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2520 VariableScale, Depth)) {
2521 // If even that didn't work, bail.
2522 AddrMode = BackupAddrMode;
2523 AddrModeInsts.resize(OldSize);
2530 case Instruction::SExt:
2531 case Instruction::ZExt: {
2532 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
2536 // Try to move this ext out of the way of the addressing mode.
2537 // Ask for a method for doing so.
2538 TypePromotionHelper::Action TPH =
2539 TypePromotionHelper::getAction(Ext, InsertedTruncs, TLI, PromotedInsts);
2543 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2544 TPT.getRestorationPoint();
2545 unsigned CreatedInsts = 0;
2546 Value *PromotedOperand =
2547 TPH(Ext, TPT, PromotedInsts, CreatedInsts, nullptr, nullptr);
2548 // SExt has been moved away.
2549 // Thus either it will be rematched later in the recursive calls or it is
2550 // gone. Anyway, we must not fold it into the addressing mode at this point.
2554 // addr = gep base, idx
2556 // promotedOpnd = ext opnd <- no match here
2557 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2558 // addr = gep base, op <- match
2562 assert(PromotedOperand &&
2563 "TypePromotionHelper should have filtered out those cases");
2565 ExtAddrMode BackupAddrMode = AddrMode;
2566 unsigned OldSize = AddrModeInsts.size();
2568 if (!MatchAddr(PromotedOperand, Depth) ||
2569 !IsPromotionProfitable(AddrModeInsts.size(), OldSize + CreatedInsts,
2571 AddrMode = BackupAddrMode;
2572 AddrModeInsts.resize(OldSize);
2573 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2574 TPT.rollback(LastKnownGood);
2583 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2584 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2585 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2586 /// or intptr_t for the target.
2588 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2589 // Start a transaction at this point that we will rollback if the matching
2591 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2592 TPT.getRestorationPoint();
2593 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2594 // Fold in immediates if legal for the target.
2595 AddrMode.BaseOffs += CI->getSExtValue();
2596 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2598 AddrMode.BaseOffs -= CI->getSExtValue();
2599 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2600 // If this is a global variable, try to fold it into the addressing mode.
2601 if (!AddrMode.BaseGV) {
2602 AddrMode.BaseGV = GV;
2603 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2605 AddrMode.BaseGV = nullptr;
2607 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2608 ExtAddrMode BackupAddrMode = AddrMode;
2609 unsigned OldSize = AddrModeInsts.size();
2611 // Check to see if it is possible to fold this operation.
2612 bool MovedAway = false;
2613 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2614 // This instruction may have been move away. If so, there is nothing
2618 // Okay, it's possible to fold this. Check to see if it is actually
2619 // *profitable* to do so. We use a simple cost model to avoid increasing
2620 // register pressure too much.
2621 if (I->hasOneUse() ||
2622 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2623 AddrModeInsts.push_back(I);
2627 // It isn't profitable to do this, roll back.
2628 //cerr << "NOT FOLDING: " << *I;
2629 AddrMode = BackupAddrMode;
2630 AddrModeInsts.resize(OldSize);
2631 TPT.rollback(LastKnownGood);
2633 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2634 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2636 TPT.rollback(LastKnownGood);
2637 } else if (isa<ConstantPointerNull>(Addr)) {
2638 // Null pointer gets folded without affecting the addressing mode.
2642 // Worse case, the target should support [reg] addressing modes. :)
2643 if (!AddrMode.HasBaseReg) {
2644 AddrMode.HasBaseReg = true;
2645 AddrMode.BaseReg = Addr;
2646 // Still check for legality in case the target supports [imm] but not [i+r].
2647 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2649 AddrMode.HasBaseReg = false;
2650 AddrMode.BaseReg = nullptr;
2653 // If the base register is already taken, see if we can do [r+r].
2654 if (AddrMode.Scale == 0) {
2656 AddrMode.ScaledReg = Addr;
2657 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2660 AddrMode.ScaledReg = nullptr;
2663 TPT.rollback(LastKnownGood);
2667 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2668 /// inline asm call are due to memory operands. If so, return true, otherwise
2670 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2671 const TargetLowering &TLI) {
2672 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(ImmutableCallSite(CI));
2673 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2674 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2676 // Compute the constraint code and ConstraintType to use.
2677 TLI.ComputeConstraintToUse(OpInfo, SDValue());
2679 // If this asm operand is our Value*, and if it isn't an indirect memory
2680 // operand, we can't fold it!
2681 if (OpInfo.CallOperandVal == OpVal &&
2682 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2683 !OpInfo.isIndirect))
2690 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2691 /// memory use. If we find an obviously non-foldable instruction, return true.
2692 /// Add the ultimately found memory instructions to MemoryUses.
2693 static bool FindAllMemoryUses(Instruction *I,
2694 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
2695 SmallPtrSetImpl<Instruction*> &ConsideredInsts,
2696 const TargetLowering &TLI) {
2697 // If we already considered this instruction, we're done.
2698 if (!ConsideredInsts.insert(I).second)
2701 // If this is an obviously unfoldable instruction, bail out.
2702 if (!MightBeFoldableInst(I))
2705 // Loop over all the uses, recursively processing them.
2706 for (Use &U : I->uses()) {
2707 Instruction *UserI = cast<Instruction>(U.getUser());
2709 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2710 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2714 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2715 unsigned opNo = U.getOperandNo();
2716 if (opNo == 0) return true; // Storing addr, not into addr.
2717 MemoryUses.push_back(std::make_pair(SI, opNo));
2721 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2722 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2723 if (!IA) return true;
2725 // If this is a memory operand, we're cool, otherwise bail out.
2726 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
2731 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI))
2738 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2739 /// the use site that we're folding it into. If so, there is no cost to
2740 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2741 /// that we know are live at the instruction already.
2742 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2743 Value *KnownLive2) {
2744 // If Val is either of the known-live values, we know it is live!
2745 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
2748 // All values other than instructions and arguments (e.g. constants) are live.
2749 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2751 // If Val is a constant sized alloca in the entry block, it is live, this is
2752 // true because it is just a reference to the stack/frame pointer, which is
2753 // live for the whole function.
2754 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2755 if (AI->isStaticAlloca())
2758 // Check to see if this value is already used in the memory instruction's
2759 // block. If so, it's already live into the block at the very least, so we
2760 // can reasonably fold it.
2761 return Val->isUsedInBasicBlock(MemoryInst->getParent());
2764 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
2765 /// mode of the machine to fold the specified instruction into a load or store
2766 /// that ultimately uses it. However, the specified instruction has multiple
2767 /// uses. Given this, it may actually increase register pressure to fold it
2768 /// into the load. For example, consider this code:
2772 /// use(Y) -> nonload/store
2776 /// In this case, Y has multiple uses, and can be folded into the load of Z
2777 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
2778 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
2779 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
2780 /// number of computations either.
2782 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
2783 /// X was live across 'load Z' for other reasons, we actually *would* want to
2784 /// fold the addressing mode in the Z case. This would make Y die earlier.
2785 bool AddressingModeMatcher::
2786 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
2787 ExtAddrMode &AMAfter) {
2788 if (IgnoreProfitability) return true;
2790 // AMBefore is the addressing mode before this instruction was folded into it,
2791 // and AMAfter is the addressing mode after the instruction was folded. Get
2792 // the set of registers referenced by AMAfter and subtract out those
2793 // referenced by AMBefore: this is the set of values which folding in this
2794 // address extends the lifetime of.
2796 // Note that there are only two potential values being referenced here,
2797 // BaseReg and ScaleReg (global addresses are always available, as are any
2798 // folded immediates).
2799 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
2801 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
2802 // lifetime wasn't extended by adding this instruction.
2803 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2805 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2806 ScaledReg = nullptr;
2808 // If folding this instruction (and it's subexprs) didn't extend any live
2809 // ranges, we're ok with it.
2810 if (!BaseReg && !ScaledReg)
2813 // If all uses of this instruction are ultimately load/store/inlineasm's,
2814 // check to see if their addressing modes will include this instruction. If
2815 // so, we can fold it into all uses, so it doesn't matter if it has multiple
2817 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
2818 SmallPtrSet<Instruction*, 16> ConsideredInsts;
2819 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
2820 return false; // Has a non-memory, non-foldable use!
2822 // Now that we know that all uses of this instruction are part of a chain of
2823 // computation involving only operations that could theoretically be folded
2824 // into a memory use, loop over each of these uses and see if they could
2825 // *actually* fold the instruction.
2826 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
2827 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
2828 Instruction *User = MemoryUses[i].first;
2829 unsigned OpNo = MemoryUses[i].second;
2831 // Get the access type of this use. If the use isn't a pointer, we don't
2832 // know what it accesses.
2833 Value *Address = User->getOperand(OpNo);
2834 if (!Address->getType()->isPointerTy())
2836 Type *AddressAccessTy = Address->getType()->getPointerElementType();
2838 // Do a match against the root of this address, ignoring profitability. This
2839 // will tell us if the addressing mode for the memory operation will
2840 // *actually* cover the shared instruction.
2842 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2843 TPT.getRestorationPoint();
2844 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
2845 MemoryInst, Result, InsertedTruncs,
2846 PromotedInsts, TPT);
2847 Matcher.IgnoreProfitability = true;
2848 bool Success = Matcher.MatchAddr(Address, 0);
2849 (void)Success; assert(Success && "Couldn't select *anything*?");
2851 // The match was to check the profitability, the changes made are not
2852 // part of the original matcher. Therefore, they should be dropped
2853 // otherwise the original matcher will not present the right state.
2854 TPT.rollback(LastKnownGood);
2856 // If the match didn't cover I, then it won't be shared by it.
2857 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
2858 I) == MatchedAddrModeInsts.end())
2861 MatchedAddrModeInsts.clear();
2867 } // end anonymous namespace
2869 /// IsNonLocalValue - Return true if the specified values are defined in a
2870 /// different basic block than BB.
2871 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
2872 if (Instruction *I = dyn_cast<Instruction>(V))
2873 return I->getParent() != BB;
2877 /// OptimizeMemoryInst - Load and Store Instructions often have
2878 /// addressing modes that can do significant amounts of computation. As such,
2879 /// instruction selection will try to get the load or store to do as much
2880 /// computation as possible for the program. The problem is that isel can only
2881 /// see within a single block. As such, we sink as much legal addressing mode
2882 /// stuff into the block as possible.
2884 /// This method is used to optimize both load/store and inline asms with memory
2886 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
2890 // Try to collapse single-value PHI nodes. This is necessary to undo
2891 // unprofitable PRE transformations.
2892 SmallVector<Value*, 8> worklist;
2893 SmallPtrSet<Value*, 16> Visited;
2894 worklist.push_back(Addr);
2896 // Use a worklist to iteratively look through PHI nodes, and ensure that
2897 // the addressing mode obtained from the non-PHI roots of the graph
2899 Value *Consensus = nullptr;
2900 unsigned NumUsesConsensus = 0;
2901 bool IsNumUsesConsensusValid = false;
2902 SmallVector<Instruction*, 16> AddrModeInsts;
2903 ExtAddrMode AddrMode;
2904 TypePromotionTransaction TPT;
2905 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2906 TPT.getRestorationPoint();
2907 while (!worklist.empty()) {
2908 Value *V = worklist.back();
2909 worklist.pop_back();
2911 // Break use-def graph loops.
2912 if (!Visited.insert(V).second) {
2913 Consensus = nullptr;
2917 // For a PHI node, push all of its incoming values.
2918 if (PHINode *P = dyn_cast<PHINode>(V)) {
2919 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
2920 worklist.push_back(P->getIncomingValue(i));
2924 // For non-PHIs, determine the addressing mode being computed.
2925 SmallVector<Instruction*, 16> NewAddrModeInsts;
2926 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
2927 V, AccessTy, MemoryInst, NewAddrModeInsts, *TLI, InsertedTruncsSet,
2928 PromotedInsts, TPT);
2930 // This check is broken into two cases with very similar code to avoid using
2931 // getNumUses() as much as possible. Some values have a lot of uses, so
2932 // calling getNumUses() unconditionally caused a significant compile-time
2936 AddrMode = NewAddrMode;
2937 AddrModeInsts = NewAddrModeInsts;
2939 } else if (NewAddrMode == AddrMode) {
2940 if (!IsNumUsesConsensusValid) {
2941 NumUsesConsensus = Consensus->getNumUses();
2942 IsNumUsesConsensusValid = true;
2945 // Ensure that the obtained addressing mode is equivalent to that obtained
2946 // for all other roots of the PHI traversal. Also, when choosing one
2947 // such root as representative, select the one with the most uses in order
2948 // to keep the cost modeling heuristics in AddressingModeMatcher
2950 unsigned NumUses = V->getNumUses();
2951 if (NumUses > NumUsesConsensus) {
2953 NumUsesConsensus = NumUses;
2954 AddrModeInsts = NewAddrModeInsts;
2959 Consensus = nullptr;
2963 // If the addressing mode couldn't be determined, or if multiple different
2964 // ones were determined, bail out now.
2966 TPT.rollback(LastKnownGood);
2971 // Check to see if any of the instructions supersumed by this addr mode are
2972 // non-local to I's BB.
2973 bool AnyNonLocal = false;
2974 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
2975 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
2981 // If all the instructions matched are already in this BB, don't do anything.
2983 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
2987 // Insert this computation right after this user. Since our caller is
2988 // scanning from the top of the BB to the bottom, reuse of the expr are
2989 // guaranteed to happen later.
2990 IRBuilder<> Builder(MemoryInst);
2992 // Now that we determined the addressing expression we want to use and know
2993 // that we have to sink it into this block. Check to see if we have already
2994 // done this for some other load/store instr in this block. If so, reuse the
2996 Value *&SunkAddr = SunkAddrs[Addr];
2998 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
2999 << *MemoryInst << "\n");
3000 if (SunkAddr->getType() != Addr->getType())
3001 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3002 } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() &&
3003 TM && TM->getSubtarget<TargetSubtargetInfo>().useAA())) {
3004 // By default, we use the GEP-based method when AA is used later. This
3005 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3006 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3007 << *MemoryInst << "\n");
3008 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3009 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3011 // First, find the pointer.
3012 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3013 ResultPtr = AddrMode.BaseReg;
3014 AddrMode.BaseReg = nullptr;
3017 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3018 // We can't add more than one pointer together, nor can we scale a
3019 // pointer (both of which seem meaningless).
3020 if (ResultPtr || AddrMode.Scale != 1)
3023 ResultPtr = AddrMode.ScaledReg;
3027 if (AddrMode.BaseGV) {
3031 ResultPtr = AddrMode.BaseGV;
3034 // If the real base value actually came from an inttoptr, then the matcher
3035 // will look through it and provide only the integer value. In that case,
3037 if (!ResultPtr && AddrMode.BaseReg) {
3039 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3040 AddrMode.BaseReg = nullptr;
3041 } else if (!ResultPtr && AddrMode.Scale == 1) {
3043 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3048 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3049 SunkAddr = Constant::getNullValue(Addr->getType());
3050 } else if (!ResultPtr) {
3054 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3056 // Start with the base register. Do this first so that subsequent address
3057 // matching finds it last, which will prevent it from trying to match it
3058 // as the scaled value in case it happens to be a mul. That would be
3059 // problematic if we've sunk a different mul for the scale, because then
3060 // we'd end up sinking both muls.
3061 if (AddrMode.BaseReg) {
3062 Value *V = AddrMode.BaseReg;
3063 if (V->getType() != IntPtrTy)
3064 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3069 // Add the scale value.
3070 if (AddrMode.Scale) {
3071 Value *V = AddrMode.ScaledReg;
3072 if (V->getType() == IntPtrTy) {
3074 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3075 cast<IntegerType>(V->getType())->getBitWidth()) {
3076 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3078 // It is only safe to sign extend the BaseReg if we know that the math
3079 // required to create it did not overflow before we extend it. Since
3080 // the original IR value was tossed in favor of a constant back when
3081 // the AddrMode was created we need to bail out gracefully if widths
3082 // do not match instead of extending it.
3083 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3084 if (I && (ResultIndex != AddrMode.BaseReg))
3085 I->eraseFromParent();
3089 if (AddrMode.Scale != 1)
3090 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3093 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3098 // Add in the Base Offset if present.
3099 if (AddrMode.BaseOffs) {
3100 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3102 // We need to add this separately from the scale above to help with
3103 // SDAG consecutive load/store merging.
3104 if (ResultPtr->getType() != I8PtrTy)
3105 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3106 ResultPtr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
3113 SunkAddr = ResultPtr;
3115 if (ResultPtr->getType() != I8PtrTy)
3116 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3117 SunkAddr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
3120 if (SunkAddr->getType() != Addr->getType())
3121 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3124 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3125 << *MemoryInst << "\n");
3126 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3127 Value *Result = nullptr;
3129 // Start with the base register. Do this first so that subsequent address
3130 // matching finds it last, which will prevent it from trying to match it
3131 // as the scaled value in case it happens to be a mul. That would be
3132 // problematic if we've sunk a different mul for the scale, because then
3133 // we'd end up sinking both muls.
3134 if (AddrMode.BaseReg) {
3135 Value *V = AddrMode.BaseReg;
3136 if (V->getType()->isPointerTy())
3137 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3138 if (V->getType() != IntPtrTy)
3139 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3143 // Add the scale value.
3144 if (AddrMode.Scale) {
3145 Value *V = AddrMode.ScaledReg;
3146 if (V->getType() == IntPtrTy) {
3148 } else if (V->getType()->isPointerTy()) {
3149 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3150 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3151 cast<IntegerType>(V->getType())->getBitWidth()) {
3152 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3154 // It is only safe to sign extend the BaseReg if we know that the math
3155 // required to create it did not overflow before we extend it. Since
3156 // the original IR value was tossed in favor of a constant back when
3157 // the AddrMode was created we need to bail out gracefully if widths
3158 // do not match instead of extending it.
3159 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3160 if (I && (Result != AddrMode.BaseReg))
3161 I->eraseFromParent();
3164 if (AddrMode.Scale != 1)
3165 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3168 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3173 // Add in the BaseGV if present.
3174 if (AddrMode.BaseGV) {
3175 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3177 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3182 // Add in the Base Offset if present.
3183 if (AddrMode.BaseOffs) {
3184 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3186 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3192 SunkAddr = Constant::getNullValue(Addr->getType());
3194 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3197 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3199 // If we have no uses, recursively delete the value and all dead instructions
3201 if (Repl->use_empty()) {
3202 // This can cause recursive deletion, which can invalidate our iterator.
3203 // Use a WeakVH to hold onto it in case this happens.
3204 WeakVH IterHandle(CurInstIterator);
3205 BasicBlock *BB = CurInstIterator->getParent();
3207 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3209 if (IterHandle != CurInstIterator) {
3210 // If the iterator instruction was recursively deleted, start over at the
3211 // start of the block.
3212 CurInstIterator = BB->begin();
3220 /// OptimizeInlineAsmInst - If there are any memory operands, use
3221 /// OptimizeMemoryInst to sink their address computing into the block when
3222 /// possible / profitable.
3223 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
3224 bool MadeChange = false;
3226 TargetLowering::AsmOperandInfoVector
3227 TargetConstraints = TLI->ParseConstraints(CS);
3229 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3230 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3232 // Compute the constraint code and ConstraintType to use.
3233 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3235 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3236 OpInfo.isIndirect) {
3237 Value *OpVal = CS->getArgOperand(ArgNo++);
3238 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
3239 } else if (OpInfo.Type == InlineAsm::isInput)
3246 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3247 /// sign extensions.
3248 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3249 assert(!Inst->use_empty() && "Input must have at least one use");
3250 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3251 bool IsSExt = isa<SExtInst>(FirstUser);
3252 Type *ExtTy = FirstUser->getType();
3253 for (const User *U : Inst->users()) {
3254 const Instruction *UI = cast<Instruction>(U);
3255 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3257 Type *CurTy = UI->getType();
3258 // Same input and output types: Same instruction after CSE.
3262 // If IsSExt is true, we are in this situation:
3264 // b = sext ty1 a to ty2
3265 // c = sext ty1 a to ty3
3266 // Assuming ty2 is shorter than ty3, this could be turned into:
3268 // b = sext ty1 a to ty2
3269 // c = sext ty2 b to ty3
3270 // However, the last sext is not free.
3274 // This is a ZExt, maybe this is free to extend from one type to another.
3275 // In that case, we would not account for a different use.
3278 if (ExtTy->getScalarType()->getIntegerBitWidth() >
3279 CurTy->getScalarType()->getIntegerBitWidth()) {
3287 if (!TLI.isZExtFree(NarrowTy, LargeTy))
3290 // All uses are the same or can be derived from one another for free.
3294 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
3295 /// load instruction.
3296 /// If an ext(load) can be formed, it is returned via \p LI for the load
3297 /// and \p Inst for the extension.
3298 /// Otherwise LI == nullptr and Inst == nullptr.
3299 /// When some promotion happened, \p TPT contains the proper state to
3302 /// \return true when promoting was necessary to expose the ext(load)
3303 /// opportunity, false otherwise.
3307 /// %ld = load i32* %addr
3308 /// %add = add nuw i32 %ld, 4
3309 /// %zext = zext i32 %add to i64
3313 /// %ld = load i32* %addr
3314 /// %zext = zext i32 %ld to i64
3315 /// %add = add nuw i64 %zext, 4
3317 /// Thanks to the promotion, we can match zext(load i32*) to i64.
3318 bool CodeGenPrepare::ExtLdPromotion(TypePromotionTransaction &TPT,
3319 LoadInst *&LI, Instruction *&Inst,
3320 const SmallVectorImpl<Instruction *> &Exts,
3321 unsigned CreatedInsts = 0) {
3322 // Iterate over all the extensions to see if one form an ext(load).
3323 for (auto I : Exts) {
3324 // Check if we directly have ext(load).
3325 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
3327 // No promotion happened here.
3330 // Check whether or not we want to do any promotion.
3331 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
3333 // Get the action to perform the promotion.
3334 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
3335 I, InsertedTruncsSet, *TLI, PromotedInsts);
3336 // Check if we can promote.
3339 // Save the current state.
3340 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3341 TPT.getRestorationPoint();
3342 SmallVector<Instruction *, 4> NewExts;
3343 unsigned NewCreatedInsts = 0;
3345 Value *PromotedVal =
3346 TPH(I, TPT, PromotedInsts, NewCreatedInsts, &NewExts, nullptr);
3347 assert(PromotedVal &&
3348 "TypePromotionHelper should have filtered out those cases");
3350 // We would be able to merge only one extension in a load.
3351 // Therefore, if we have more than 1 new extension we heuristically
3352 // cut this search path, because it means we degrade the code quality.
3353 // With exactly 2, the transformation is neutral, because we will merge
3354 // one extension but leave one. However, we optimistically keep going,
3355 // because the new extension may be removed too.
3356 unsigned TotalCreatedInsts = CreatedInsts + NewCreatedInsts;
3357 if (!StressExtLdPromotion &&
3358 (TotalCreatedInsts > 1 ||
3359 !isPromotedInstructionLegal(*TLI, PromotedVal))) {
3360 // The promotion is not profitable, rollback to the previous state.
3361 TPT.rollback(LastKnownGood);
3364 // The promotion is profitable.
3365 // Check if it exposes an ext(load).
3366 (void)ExtLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInsts);
3367 if (LI && (StressExtLdPromotion || NewCreatedInsts == 0 ||
3368 // If we have created a new extension, i.e., now we have two
3369 // extensions. We must make sure one of them is merged with
3370 // the load, otherwise we may degrade the code quality.
3371 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
3372 // Promotion happened.
3374 // If this does not help to expose an ext(load) then, rollback.
3375 TPT.rollback(LastKnownGood);
3377 // None of the extension can form an ext(load).
3383 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
3384 /// basic block as the load, unless conditions are unfavorable. This allows
3385 /// SelectionDAG to fold the extend into the load.
3386 /// \p I[in/out] the extension may be modified during the process if some
3387 /// promotions apply.
3389 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *&I) {
3390 // Try to promote a chain of computation if it allows to form
3391 // an extended load.
3392 TypePromotionTransaction TPT;
3393 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3394 TPT.getRestorationPoint();
3395 SmallVector<Instruction *, 1> Exts;
3397 // Look for a load being extended.
3398 LoadInst *LI = nullptr;
3399 Instruction *OldExt = I;
3400 bool HasPromoted = ExtLdPromotion(TPT, LI, I, Exts);
3402 assert(!HasPromoted && !LI && "If we did not match any load instruction "
3403 "the code must remain the same");
3408 // If they're already in the same block, there's nothing to do.
3409 // Make the cheap checks first if we did not promote.
3410 // If we promoted, we need to check if it is indeed profitable.
3411 if (!HasPromoted && LI->getParent() == I->getParent())
3414 EVT VT = TLI->getValueType(I->getType());
3415 EVT LoadVT = TLI->getValueType(LI->getType());
3417 // If the load has other users and the truncate is not free, this probably
3418 // isn't worthwhile.
3419 if (!LI->hasOneUse() && TLI &&
3420 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
3421 !TLI->isTruncateFree(I->getType(), LI->getType())) {
3423 TPT.rollback(LastKnownGood);
3427 // Check whether the target supports casts folded into loads.
3429 if (isa<ZExtInst>(I))
3430 LType = ISD::ZEXTLOAD;
3432 assert(isa<SExtInst>(I) && "Unexpected ext type!");
3433 LType = ISD::SEXTLOAD;
3435 if (TLI && !TLI->isLoadExtLegal(LType, LoadVT)) {
3437 TPT.rollback(LastKnownGood);
3441 // Move the extend into the same block as the load, so that SelectionDAG
3444 I->removeFromParent();
3450 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
3451 BasicBlock *DefBB = I->getParent();
3453 // If the result of a {s|z}ext and its source are both live out, rewrite all
3454 // other uses of the source with result of extension.
3455 Value *Src = I->getOperand(0);
3456 if (Src->hasOneUse())
3459 // Only do this xform if truncating is free.
3460 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
3463 // Only safe to perform the optimization if the source is also defined in
3465 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
3468 bool DefIsLiveOut = false;
3469 for (User *U : I->users()) {
3470 Instruction *UI = cast<Instruction>(U);
3472 // Figure out which BB this ext is used in.
3473 BasicBlock *UserBB = UI->getParent();
3474 if (UserBB == DefBB) continue;
3475 DefIsLiveOut = true;
3481 // Make sure none of the uses are PHI nodes.
3482 for (User *U : Src->users()) {
3483 Instruction *UI = cast<Instruction>(U);
3484 BasicBlock *UserBB = UI->getParent();
3485 if (UserBB == DefBB) continue;
3486 // Be conservative. We don't want this xform to end up introducing
3487 // reloads just before load / store instructions.
3488 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
3492 // InsertedTruncs - Only insert one trunc in each block once.
3493 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
3495 bool MadeChange = false;
3496 for (Use &U : Src->uses()) {
3497 Instruction *User = cast<Instruction>(U.getUser());
3499 // Figure out which BB this ext is used in.
3500 BasicBlock *UserBB = User->getParent();
3501 if (UserBB == DefBB) continue;
3503 // Both src and def are live in this block. Rewrite the use.
3504 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
3506 if (!InsertedTrunc) {
3507 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3508 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
3509 InsertedTruncsSet.insert(InsertedTrunc);
3512 // Replace a use of the {s|z}ext source with a use of the result.
3521 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
3522 /// turned into an explicit branch.
3523 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
3524 // FIXME: This should use the same heuristics as IfConversion to determine
3525 // whether a select is better represented as a branch. This requires that
3526 // branch probability metadata is preserved for the select, which is not the
3529 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3531 // If the branch is predicted right, an out of order CPU can avoid blocking on
3532 // the compare. Emit cmovs on compares with a memory operand as branches to
3533 // avoid stalls on the load from memory. If the compare has more than one use
3534 // there's probably another cmov or setcc around so it's not worth emitting a
3539 Value *CmpOp0 = Cmp->getOperand(0);
3540 Value *CmpOp1 = Cmp->getOperand(1);
3542 // We check that the memory operand has one use to avoid uses of the loaded
3543 // value directly after the compare, making branches unprofitable.
3544 return Cmp->hasOneUse() &&
3545 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3546 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
3550 /// If we have a SelectInst that will likely profit from branch prediction,
3551 /// turn it into a branch.
3552 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
3553 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3555 // Can we convert the 'select' to CF ?
3556 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3559 TargetLowering::SelectSupportKind SelectKind;
3561 SelectKind = TargetLowering::VectorMaskSelect;
3562 else if (SI->getType()->isVectorTy())
3563 SelectKind = TargetLowering::ScalarCondVectorVal;
3565 SelectKind = TargetLowering::ScalarValSelect;
3567 // Do we have efficient codegen support for this kind of 'selects' ?
3568 if (TLI->isSelectSupported(SelectKind)) {
3569 // We have efficient codegen support for the select instruction.
3570 // Check if it is profitable to keep this 'select'.
3571 if (!TLI->isPredictableSelectExpensive() ||
3572 !isFormingBranchFromSelectProfitable(SI))
3578 // First, we split the block containing the select into 2 blocks.
3579 BasicBlock *StartBlock = SI->getParent();
3580 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3581 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3583 // Create a new block serving as the landing pad for the branch.
3584 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3585 NextBlock->getParent(), NextBlock);
3587 // Move the unconditional branch from the block with the select in it into our
3588 // landing pad block.
3589 StartBlock->getTerminator()->eraseFromParent();
3590 BranchInst::Create(NextBlock, SmallBlock);
3592 // Insert the real conditional branch based on the original condition.
3593 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3595 // The select itself is replaced with a PHI Node.
3596 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3598 PN->addIncoming(SI->getTrueValue(), StartBlock);
3599 PN->addIncoming(SI->getFalseValue(), SmallBlock);
3600 SI->replaceAllUsesWith(PN);
3601 SI->eraseFromParent();
3603 // Instruct OptimizeBlock to skip to the next block.
3604 CurInstIterator = StartBlock->end();
3605 ++NumSelectsExpanded;
3609 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3610 SmallVector<int, 16> Mask(SVI->getShuffleMask());
3612 for (unsigned i = 0; i < Mask.size(); ++i) {
3613 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3615 SplatElem = Mask[i];
3621 /// Some targets have expensive vector shifts if the lanes aren't all the same
3622 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3623 /// it's often worth sinking a shufflevector splat down to its use so that
3624 /// codegen can spot all lanes are identical.
3625 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3626 BasicBlock *DefBB = SVI->getParent();
3628 // Only do this xform if variable vector shifts are particularly expensive.
3629 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3632 // We only expect better codegen by sinking a shuffle if we can recognise a
3634 if (!isBroadcastShuffle(SVI))
3637 // InsertedShuffles - Only insert a shuffle in each block once.
3638 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3640 bool MadeChange = false;
3641 for (User *U : SVI->users()) {
3642 Instruction *UI = cast<Instruction>(U);
3644 // Figure out which BB this ext is used in.
3645 BasicBlock *UserBB = UI->getParent();
3646 if (UserBB == DefBB) continue;
3648 // For now only apply this when the splat is used by a shift instruction.
3649 if (!UI->isShift()) continue;
3651 // Everything checks out, sink the shuffle if the user's block doesn't
3652 // already have a copy.
3653 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3655 if (!InsertedShuffle) {
3656 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3657 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3659 SVI->getOperand(2), "", InsertPt);
3662 UI->replaceUsesOfWith(SVI, InsertedShuffle);
3666 // If we removed all uses, nuke the shuffle.
3667 if (SVI->use_empty()) {
3668 SVI->eraseFromParent();
3676 /// \brief Helper class to promote a scalar operation to a vector one.
3677 /// This class is used to move downward extractelement transition.
3679 /// a = vector_op <2 x i32>
3680 /// b = extractelement <2 x i32> a, i32 0
3685 /// a = vector_op <2 x i32>
3686 /// c = vector_op a (equivalent to scalar_op on the related lane)
3687 /// * d = extractelement <2 x i32> c, i32 0
3689 /// Assuming both extractelement and store can be combine, we get rid of the
3691 class VectorPromoteHelper {
3692 /// Used to perform some checks on the legality of vector operations.
3693 const TargetLowering &TLI;
3695 /// Used to estimated the cost of the promoted chain.
3696 const TargetTransformInfo &TTI;
3698 /// The transition being moved downwards.
3699 Instruction *Transition;
3700 /// The sequence of instructions to be promoted.
3701 SmallVector<Instruction *, 4> InstsToBePromoted;
3702 /// Cost of combining a store and an extract.
3703 unsigned StoreExtractCombineCost;
3704 /// Instruction that will be combined with the transition.
3705 Instruction *CombineInst;
3707 /// \brief The instruction that represents the current end of the transition.
3708 /// Since we are faking the promotion until we reach the end of the chain
3709 /// of computation, we need a way to get the current end of the transition.
3710 Instruction *getEndOfTransition() const {
3711 if (InstsToBePromoted.empty())
3713 return InstsToBePromoted.back();
3716 /// \brief Return the index of the original value in the transition.
3717 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
3718 /// c, is at index 0.
3719 unsigned getTransitionOriginalValueIdx() const {
3720 assert(isa<ExtractElementInst>(Transition) &&
3721 "Other kind of transitions are not supported yet");
3725 /// \brief Return the index of the index in the transition.
3726 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
3728 unsigned getTransitionIdx() const {
3729 assert(isa<ExtractElementInst>(Transition) &&
3730 "Other kind of transitions are not supported yet");
3734 /// \brief Get the type of the transition.
3735 /// This is the type of the original value.
3736 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
3737 /// transition is <2 x i32>.
3738 Type *getTransitionType() const {
3739 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
3742 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
3743 /// I.e., we have the following sequence:
3744 /// Def = Transition <ty1> a to <ty2>
3745 /// b = ToBePromoted <ty2> Def, ...
3747 /// b = ToBePromoted <ty1> a, ...
3748 /// Def = Transition <ty1> ToBePromoted to <ty2>
3749 void promoteImpl(Instruction *ToBePromoted);
3751 /// \brief Check whether or not it is profitable to promote all the
3752 /// instructions enqueued to be promoted.
3753 bool isProfitableToPromote() {
3754 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
3755 unsigned Index = isa<ConstantInt>(ValIdx)
3756 ? cast<ConstantInt>(ValIdx)->getZExtValue()
3758 Type *PromotedType = getTransitionType();
3760 StoreInst *ST = cast<StoreInst>(CombineInst);
3761 unsigned AS = ST->getPointerAddressSpace();
3762 unsigned Align = ST->getAlignment();
3763 // Check if this store is supported.
3764 if (!TLI.allowsMisalignedMemoryAccesses(
3765 TLI.getValueType(ST->getValueOperand()->getType()), AS, Align)) {
3766 // If this is not supported, there is no way we can combine
3767 // the extract with the store.
3771 // The scalar chain of computation has to pay for the transition
3772 // scalar to vector.
3773 // The vector chain has to account for the combining cost.
3774 uint64_t ScalarCost =
3775 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
3776 uint64_t VectorCost = StoreExtractCombineCost;
3777 for (const auto &Inst : InstsToBePromoted) {
3778 // Compute the cost.
3779 // By construction, all instructions being promoted are arithmetic ones.
3780 // Moreover, one argument is a constant that can be viewed as a splat
3782 Value *Arg0 = Inst->getOperand(0);
3783 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
3784 isa<ConstantFP>(Arg0);
3785 TargetTransformInfo::OperandValueKind Arg0OVK =
3786 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3787 : TargetTransformInfo::OK_AnyValue;
3788 TargetTransformInfo::OperandValueKind Arg1OVK =
3789 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3790 : TargetTransformInfo::OK_AnyValue;
3791 ScalarCost += TTI.getArithmeticInstrCost(
3792 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
3793 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
3796 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
3797 << ScalarCost << "\nVector: " << VectorCost << '\n');
3798 return ScalarCost > VectorCost;
3801 /// \brief Generate a constant vector with \p Val with the same
3802 /// number of elements as the transition.
3803 /// \p UseSplat defines whether or not \p Val should be replicated
3804 /// accross the whole vector.
3805 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
3806 /// otherwise we generate a vector with as many undef as possible:
3807 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
3808 /// used at the index of the extract.
3809 Value *getConstantVector(Constant *Val, bool UseSplat) const {
3810 unsigned ExtractIdx = UINT_MAX;
3812 // If we cannot determine where the constant must be, we have to
3813 // use a splat constant.
3814 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
3815 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
3816 ExtractIdx = CstVal->getSExtValue();
3821 unsigned End = getTransitionType()->getVectorNumElements();
3823 return ConstantVector::getSplat(End, Val);
3825 SmallVector<Constant *, 4> ConstVec;
3826 UndefValue *UndefVal = UndefValue::get(Val->getType());
3827 for (unsigned Idx = 0; Idx != End; ++Idx) {
3828 if (Idx == ExtractIdx)
3829 ConstVec.push_back(Val);
3831 ConstVec.push_back(UndefVal);
3833 return ConstantVector::get(ConstVec);
3836 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
3837 /// in \p Use can trigger undefined behavior.
3838 static bool canCauseUndefinedBehavior(const Instruction *Use,
3839 unsigned OperandIdx) {
3840 // This is not safe to introduce undef when the operand is on
3841 // the right hand side of a division-like instruction.
3842 if (OperandIdx != 1)
3844 switch (Use->getOpcode()) {
3847 case Instruction::SDiv:
3848 case Instruction::UDiv:
3849 case Instruction::SRem:
3850 case Instruction::URem:
3852 case Instruction::FDiv:
3853 case Instruction::FRem:
3854 return !Use->hasNoNaNs();
3856 llvm_unreachable(nullptr);
3860 VectorPromoteHelper(const TargetLowering &TLI, const TargetTransformInfo &TTI,
3861 Instruction *Transition, unsigned CombineCost)
3862 : TLI(TLI), TTI(TTI), Transition(Transition),
3863 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
3864 assert(Transition && "Do not know how to promote null");
3867 /// \brief Check if we can promote \p ToBePromoted to \p Type.
3868 bool canPromote(const Instruction *ToBePromoted) const {
3869 // We could support CastInst too.
3870 return isa<BinaryOperator>(ToBePromoted);
3873 /// \brief Check if it is profitable to promote \p ToBePromoted
3874 /// by moving downward the transition through.
3875 bool shouldPromote(const Instruction *ToBePromoted) const {
3876 // Promote only if all the operands can be statically expanded.
3877 // Indeed, we do not want to introduce any new kind of transitions.
3878 for (const Use &U : ToBePromoted->operands()) {
3879 const Value *Val = U.get();
3880 if (Val == getEndOfTransition()) {
3881 // If the use is a division and the transition is on the rhs,
3882 // we cannot promote the operation, otherwise we may create a
3883 // division by zero.
3884 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
3888 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
3889 !isa<ConstantFP>(Val))
3892 // Check that the resulting operation is legal.
3893 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
3896 return StressStoreExtract ||
3897 TLI.isOperationLegalOrCustom(
3898 ISDOpcode, TLI.getValueType(getTransitionType(), true));
3901 /// \brief Check whether or not \p Use can be combined
3902 /// with the transition.
3903 /// I.e., is it possible to do Use(Transition) => AnotherUse?
3904 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
3906 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
3907 void enqueueForPromotion(Instruction *ToBePromoted) {
3908 InstsToBePromoted.push_back(ToBePromoted);
3911 /// \brief Set the instruction that will be combined with the transition.
3912 void recordCombineInstruction(Instruction *ToBeCombined) {
3913 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
3914 CombineInst = ToBeCombined;
3917 /// \brief Promote all the instructions enqueued for promotion if it is
3919 /// \return True if the promotion happened, false otherwise.
3921 // Check if there is something to promote.
3922 // Right now, if we do not have anything to combine with,
3923 // we assume the promotion is not profitable.
3924 if (InstsToBePromoted.empty() || !CombineInst)
3928 if (!StressStoreExtract && !isProfitableToPromote())
3932 for (auto &ToBePromoted : InstsToBePromoted)
3933 promoteImpl(ToBePromoted);
3934 InstsToBePromoted.clear();
3938 } // End of anonymous namespace.
3940 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
3941 // At this point, we know that all the operands of ToBePromoted but Def
3942 // can be statically promoted.
3943 // For Def, we need to use its parameter in ToBePromoted:
3944 // b = ToBePromoted ty1 a
3945 // Def = Transition ty1 b to ty2
3946 // Move the transition down.
3947 // 1. Replace all uses of the promoted operation by the transition.
3948 // = ... b => = ... Def.
3949 assert(ToBePromoted->getType() == Transition->getType() &&
3950 "The type of the result of the transition does not match "
3952 ToBePromoted->replaceAllUsesWith(Transition);
3953 // 2. Update the type of the uses.
3954 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
3955 Type *TransitionTy = getTransitionType();
3956 ToBePromoted->mutateType(TransitionTy);
3957 // 3. Update all the operands of the promoted operation with promoted
3959 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
3960 for (Use &U : ToBePromoted->operands()) {
3961 Value *Val = U.get();
3962 Value *NewVal = nullptr;
3963 if (Val == Transition)
3964 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
3965 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
3966 isa<ConstantFP>(Val)) {
3967 // Use a splat constant if it is not safe to use undef.
3968 NewVal = getConstantVector(
3969 cast<Constant>(Val),
3970 isa<UndefValue>(Val) ||
3971 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
3973 llvm_unreachable("Did you modified shouldPromote and forgot to update "
3975 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
3977 Transition->removeFromParent();
3978 Transition->insertAfter(ToBePromoted);
3979 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
3982 // See if we can speculate calls to intrinsic cttz/ctlz.
3987 // %cmp = icmp eq i64 %val, 0
3988 // br i1 %cmp, label %end.bb, label %then.bb
3991 // %c = tail call i64 @llvm.cttz.i64(i64 %val, i1 true)
3995 // %cond = phi i64 [ %c, %then.bb ], [ 64, %entry ]
4001 // %c = tail call i64 @llvm.cttz.i64(i64 %val, i1 false)
4003 static bool OptimizeBranchInst(BranchInst *BrInst, const TargetLowering &TLI) {
4004 assert(BrInst->isConditional() && "Expected a conditional branch!");
4005 BasicBlock *ThenBB = BrInst->getSuccessor(1);
4006 BasicBlock *EndBB = BrInst->getSuccessor(0);
4008 // See if ThenBB contains only one instruction (excluding the
4009 // terminator and DbgInfoIntrinsic calls).
4010 IntrinsicInst *II = nullptr;
4011 CastInst *CI = nullptr;
4012 for (BasicBlock::iterator I = ThenBB->begin(),
4013 E = std::prev(ThenBB->end()); I != E; ++I) {
4015 if (isa<DbgInfoIntrinsic>(I))
4018 // Check if this is a zero extension or a truncate of a previously
4019 // matched call to intrinsic cttz/ctlz.
4021 // Early exit if we already found a "free" zero extend/truncate.
4025 Type *SrcTy = II->getType();
4026 Type *DestTy = I->getType();
4029 if (match(cast<Instruction>(I), m_ZExt(m_Value(V))) && V == II) {
4030 // Speculate this zero extend only if it is "free" for the target.
4031 if (TLI.isZExtFree(SrcTy, DestTy)) {
4032 CI = cast<CastInst>(I);
4035 } else if (match(cast<Instruction>(I), m_Trunc(m_Value(V))) && V == II) {
4036 // Speculate this truncate only if it is "free" for the target.
4037 if (TLI.isTruncateFree(SrcTy, DestTy)) {
4038 CI = cast<CastInst>(I);
4042 // Avoid speculating more than one instruction.
4047 // See if this is a call to intrinsic cttz/ctlz.
4048 if (match(cast<Instruction>(I), m_Intrinsic<Intrinsic::cttz>())) {
4049 // Avoid speculating expensive intrinsic calls.
4050 if (!TLI.isCheapToSpeculateCttz())
4053 else if (match(cast<Instruction>(I), m_Intrinsic<Intrinsic::ctlz>())) {
4054 // Avoid speculating expensive intrinsic calls.
4055 if (!TLI.isCheapToSpeculateCtlz())
4060 II = cast<IntrinsicInst>(I);
4063 // Look for PHI nodes with 'II' as the incoming value from 'ThenBB'.
4064 BasicBlock *EntryBB = BrInst->getParent();
4065 for (BasicBlock::iterator I = EndBB->begin();
4066 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
4067 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
4068 Value *OrigV = PN->getIncomingValueForBlock(EntryBB);
4073 if (ThenV != II && (!CI || ThenV != CI))
4076 if (ConstantInt *CInt = dyn_cast<ConstantInt>(OrigV)) {
4077 unsigned BitWidth = II->getType()->getIntegerBitWidth();
4079 // Don't try to simplify this phi node if 'ThenV' is a cttz/ctlz
4080 // intrinsic call, but 'OrigV' is not equal to the 'size-of' in bits
4081 // of the value in input to the cttz/ctlz.
4082 if (CInt->getValue() != BitWidth)
4085 // Hoist the call to cttz/ctlz from ThenBB into EntryBB.
4086 EntryBB->getInstList().splice(BrInst, ThenBB->getInstList(),
4087 ThenBB->begin(), std::prev(ThenBB->end()));
4089 // Update PN setting ThenV as the incoming value from both 'EntryBB'
4090 // and 'ThenBB'. Eventually, method 'OptimizeInst' will fold this
4091 // phi node if all the incoming values are the same.
4092 PN->setIncomingValue(PN->getBasicBlockIndex(EntryBB), ThenV);
4093 PN->setIncomingValue(PN->getBasicBlockIndex(ThenBB), ThenV);
4095 // Clear the 'undef on zero' flag of the cttz/ctlz intrinsic call.
4096 if (cast<ConstantInt>(II->getArgOperand(1))->isOne()) {
4097 Type *Ty = II->getArgOperand(0)->getType();
4098 Value *Args[] = { II->getArgOperand(0),
4099 ConstantInt::getFalse(II->getContext()) };
4100 Module *M = EntryBB->getParent()->getParent();
4101 Value *IF = Intrinsic::getDeclaration(M, II->getIntrinsicID(), Ty);
4102 IRBuilder<> Builder(II);
4103 Instruction *NewI = Builder.CreateCall(IF, Args);
4105 // Replace the old call to cttz/ctlz.
4106 II->replaceAllUsesWith(NewI);
4107 II->eraseFromParent();
4110 // Update BrInst condition so that the branch to EndBB is always taken.
4111 // Later on, method 'ConstantFoldTerminator' will simplify this branch
4112 // replacing it with a direct branch to 'EndBB'.
4113 // As a side effect, CodeGenPrepare will attempt to simplify the control
4114 // flow graph by deleting basic block 'ThenBB' and merging 'EntryBB' into
4115 // 'EndBB' (calling method 'EliminateFallThrough').
4116 BrInst->setCondition(ConstantInt::getTrue(BrInst->getContext()));
4124 /// Some targets can do store(extractelement) with one instruction.
4125 /// Try to push the extractelement towards the stores when the target
4126 /// has this feature and this is profitable.
4127 bool CodeGenPrepare::OptimizeExtractElementInst(Instruction *Inst) {
4128 unsigned CombineCost = UINT_MAX;
4129 if (DisableStoreExtract || !TLI ||
4130 (!StressStoreExtract &&
4131 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4132 Inst->getOperand(1), CombineCost)))
4135 // At this point we know that Inst is a vector to scalar transition.
4136 // Try to move it down the def-use chain, until:
4137 // - We can combine the transition with its single use
4138 // => we got rid of the transition.
4139 // - We escape the current basic block
4140 // => we would need to check that we are moving it at a cheaper place and
4141 // we do not do that for now.
4142 BasicBlock *Parent = Inst->getParent();
4143 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4144 VectorPromoteHelper VPH(*TLI, *TTI, Inst, CombineCost);
4145 // If the transition has more than one use, assume this is not going to be
4147 while (Inst->hasOneUse()) {
4148 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4149 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4151 if (ToBePromoted->getParent() != Parent) {
4152 DEBUG(dbgs() << "Instruction to promote is in a different block ("
4153 << ToBePromoted->getParent()->getName()
4154 << ") than the transition (" << Parent->getName() << ").\n");
4158 if (VPH.canCombine(ToBePromoted)) {
4159 DEBUG(dbgs() << "Assume " << *Inst << '\n'
4160 << "will be combined with: " << *ToBePromoted << '\n');
4161 VPH.recordCombineInstruction(ToBePromoted);
4162 bool Changed = VPH.promote();
4163 NumStoreExtractExposed += Changed;
4167 DEBUG(dbgs() << "Try promoting.\n");
4168 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4171 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4173 VPH.enqueueForPromotion(ToBePromoted);
4174 Inst = ToBePromoted;
4179 bool CodeGenPrepare::OptimizeInst(Instruction *I, bool& ModifiedDT) {
4180 if (PHINode *P = dyn_cast<PHINode>(I)) {
4181 // It is possible for very late stage optimizations (such as SimplifyCFG)
4182 // to introduce PHI nodes too late to be cleaned up. If we detect such a
4183 // trivial PHI, go ahead and zap it here.
4184 if (Value *V = SimplifyInstruction(P, TLI ? TLI->getDataLayout() : nullptr,
4186 P->replaceAllUsesWith(V);
4187 P->eraseFromParent();
4194 if (CastInst *CI = dyn_cast<CastInst>(I)) {
4195 // If the source of the cast is a constant, then this should have
4196 // already been constant folded. The only reason NOT to constant fold
4197 // it is if something (e.g. LSR) was careful to place the constant
4198 // evaluation in a block other than then one that uses it (e.g. to hoist
4199 // the address of globals out of a loop). If this is the case, we don't
4200 // want to forward-subst the cast.
4201 if (isa<Constant>(CI->getOperand(0)))
4204 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
4207 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4208 /// Sink a zext or sext into its user blocks if the target type doesn't
4209 /// fit in one register
4210 if (TLI && TLI->getTypeAction(CI->getContext(),
4211 TLI->getValueType(CI->getType())) ==
4212 TargetLowering::TypeExpandInteger) {
4213 return SinkCast(CI);
4215 bool MadeChange = MoveExtToFormExtLoad(I);
4216 return MadeChange | OptimizeExtUses(I);
4222 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4223 if (!TLI || !TLI->hasMultipleConditionRegisters())
4224 return OptimizeCmpExpression(CI);
4226 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4228 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
4232 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4234 return OptimizeMemoryInst(I, SI->getOperand(1),
4235 SI->getOperand(0)->getType());
4239 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4241 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4242 BinOp->getOpcode() == Instruction::LShr)) {
4243 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4244 if (TLI && CI && TLI->hasExtractBitsInsn())
4245 return OptimizeExtractBits(BinOp, CI, *TLI);
4250 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4251 if (GEPI->hasAllZeroIndices()) {
4252 /// The GEP operand must be a pointer, so must its result -> BitCast
4253 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4254 GEPI->getName(), GEPI);
4255 GEPI->replaceAllUsesWith(NC);
4256 GEPI->eraseFromParent();
4258 OptimizeInst(NC, ModifiedDT);
4264 if (CallInst *CI = dyn_cast<CallInst>(I))
4265 return OptimizeCallInst(CI, ModifiedDT);
4267 if (SelectInst *SI = dyn_cast<SelectInst>(I))
4268 return OptimizeSelectInst(SI);
4270 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
4271 return OptimizeShuffleVectorInst(SVI);
4273 if (isa<ExtractElementInst>(I))
4274 return OptimizeExtractElementInst(I);
4276 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
4277 if (TLI && BI->isConditional() && BI->getCondition()->hasOneUse()) {
4278 // Check if the branch condition compares a value agaist zero.
4279 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
4280 if (ICI->getPredicate() == ICmpInst::ICMP_EQ &&
4281 match(ICI->getOperand(1), m_Zero())) {
4282 BasicBlock *ThenBB = BI->getSuccessor(1);
4283 BasicBlock *EndBB = BI->getSuccessor(0);
4285 // Check if ThenBB is only reachable from this basic block; also,
4286 // check if EndBB has more than one predecessor.
4287 if (ThenBB->getSinglePredecessor() &&
4288 !EndBB->getSinglePredecessor()) {
4289 TerminatorInst *TI = ThenBB->getTerminator();
4291 if (TI->getNumSuccessors() == 1 && TI->getSuccessor(0) == EndBB &&
4292 // Try to speculate calls to intrinsic cttz/ctlz from 'ThenBB'.
4293 OptimizeBranchInst(BI, *TLI)) {
4307 // In this pass we look for GEP and cast instructions that are used
4308 // across basic blocks and rewrite them to improve basic-block-at-a-time
4310 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
4312 bool MadeChange = false;
4314 CurInstIterator = BB.begin();
4315 while (CurInstIterator != BB.end()) {
4316 MadeChange |= OptimizeInst(CurInstIterator++, ModifiedDT);
4320 MadeChange |= DupRetToEnableTailCallOpts(&BB);
4325 // llvm.dbg.value is far away from the value then iSel may not be able
4326 // handle it properly. iSel will drop llvm.dbg.value if it can not
4327 // find a node corresponding to the value.
4328 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
4329 bool MadeChange = false;
4330 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
4331 Instruction *PrevNonDbgInst = nullptr;
4332 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
4333 Instruction *Insn = BI; ++BI;
4334 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
4335 // Leave dbg.values that refer to an alloca alone. These
4336 // instrinsics describe the address of a variable (= the alloca)
4337 // being taken. They should not be moved next to the alloca
4338 // (and to the beginning of the scope), but rather stay close to
4339 // where said address is used.
4340 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
4341 PrevNonDbgInst = Insn;
4345 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
4346 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
4347 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
4348 DVI->removeFromParent();
4349 if (isa<PHINode>(VI))
4350 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
4352 DVI->insertAfter(VI);
4361 // If there is a sequence that branches based on comparing a single bit
4362 // against zero that can be combined into a single instruction, and the
4363 // target supports folding these into a single instruction, sink the
4364 // mask and compare into the branch uses. Do this before OptimizeBlock ->
4365 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
4367 bool CodeGenPrepare::sinkAndCmp(Function &F) {
4368 if (!EnableAndCmpSinking)
4370 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
4372 bool MadeChange = false;
4373 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
4374 BasicBlock *BB = I++;
4376 // Does this BB end with the following?
4377 // %andVal = and %val, #single-bit-set
4378 // %icmpVal = icmp %andResult, 0
4379 // br i1 %cmpVal label %dest1, label %dest2"
4380 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
4381 if (!Brcc || !Brcc->isConditional())
4383 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
4384 if (!Cmp || Cmp->getParent() != BB)
4386 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
4387 if (!Zero || !Zero->isZero())
4389 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
4390 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
4392 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
4393 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
4395 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
4397 // Push the "and; icmp" for any users that are conditional branches.
4398 // Since there can only be one branch use per BB, we don't need to keep
4399 // track of which BBs we insert into.
4400 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
4404 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
4406 if (!BrccUser || !BrccUser->isConditional())
4408 BasicBlock *UserBB = BrccUser->getParent();
4409 if (UserBB == BB) continue;
4410 DEBUG(dbgs() << "found Brcc use\n");
4412 // Sink the "and; icmp" to use.
4414 BinaryOperator *NewAnd =
4415 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
4418 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
4422 DEBUG(BrccUser->getParent()->dump());
4428 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
4429 /// success, or returns false if no or invalid metadata was found.
4430 static bool extractBranchMetadata(BranchInst *BI,
4431 uint64_t &ProbTrue, uint64_t &ProbFalse) {
4432 assert(BI->isConditional() &&
4433 "Looking for probabilities on unconditional branch?");
4434 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
4435 if (!ProfileData || ProfileData->getNumOperands() != 3)
4438 const auto *CITrue =
4439 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
4440 const auto *CIFalse =
4441 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
4442 if (!CITrue || !CIFalse)
4445 ProbTrue = CITrue->getValue().getZExtValue();
4446 ProbFalse = CIFalse->getValue().getZExtValue();
4451 /// \brief Scale down both weights to fit into uint32_t.
4452 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
4453 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
4454 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
4455 NewTrue = NewTrue / Scale;
4456 NewFalse = NewFalse / Scale;
4459 /// \brief Some targets prefer to split a conditional branch like:
4461 /// %0 = icmp ne i32 %a, 0
4462 /// %1 = icmp ne i32 %b, 0
4463 /// %or.cond = or i1 %0, %1
4464 /// br i1 %or.cond, label %TrueBB, label %FalseBB
4466 /// into multiple branch instructions like:
4469 /// %0 = icmp ne i32 %a, 0
4470 /// br i1 %0, label %TrueBB, label %bb2
4472 /// %1 = icmp ne i32 %b, 0
4473 /// br i1 %1, label %TrueBB, label %FalseBB
4475 /// This usually allows instruction selection to do even further optimizations
4476 /// and combine the compare with the branch instruction. Currently this is
4477 /// applied for targets which have "cheap" jump instructions.
4479 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
4481 bool CodeGenPrepare::splitBranchCondition(Function &F) {
4482 if (!TM || TM->Options.EnableFastISel != true ||
4483 !TLI || TLI->isJumpExpensive())
4486 bool MadeChange = false;
4487 for (auto &BB : F) {
4488 // Does this BB end with the following?
4489 // %cond1 = icmp|fcmp|binary instruction ...
4490 // %cond2 = icmp|fcmp|binary instruction ...
4491 // %cond.or = or|and i1 %cond1, cond2
4492 // br i1 %cond.or label %dest1, label %dest2"
4493 BinaryOperator *LogicOp;
4494 BasicBlock *TBB, *FBB;
4495 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
4499 Value *Cond1, *Cond2;
4500 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
4501 m_OneUse(m_Value(Cond2)))))
4502 Opc = Instruction::And;
4503 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
4504 m_OneUse(m_Value(Cond2)))))
4505 Opc = Instruction::Or;
4509 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
4510 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
4513 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
4516 auto *InsertBefore = std::next(Function::iterator(BB))
4517 .getNodePtrUnchecked();
4518 auto TmpBB = BasicBlock::Create(BB.getContext(),
4519 BB.getName() + ".cond.split",
4520 BB.getParent(), InsertBefore);
4522 // Update original basic block by using the first condition directly by the
4523 // branch instruction and removing the no longer needed and/or instruction.
4524 auto *Br1 = cast<BranchInst>(BB.getTerminator());
4525 Br1->setCondition(Cond1);
4526 LogicOp->eraseFromParent();
4528 // Depending on the conditon we have to either replace the true or the false
4529 // successor of the original branch instruction.
4530 if (Opc == Instruction::And)
4531 Br1->setSuccessor(0, TmpBB);
4533 Br1->setSuccessor(1, TmpBB);
4535 // Fill in the new basic block.
4536 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
4537 if (auto *I = dyn_cast<Instruction>(Cond2)) {
4538 I->removeFromParent();
4539 I->insertBefore(Br2);
4542 // Update PHI nodes in both successors. The original BB needs to be
4543 // replaced in one succesor's PHI nodes, because the branch comes now from
4544 // the newly generated BB (NewBB). In the other successor we need to add one
4545 // incoming edge to the PHI nodes, because both branch instructions target
4546 // now the same successor. Depending on the original branch condition
4547 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
4548 // we perfrom the correct update for the PHI nodes.
4549 // This doesn't change the successor order of the just created branch
4550 // instruction (or any other instruction).
4551 if (Opc == Instruction::Or)
4552 std::swap(TBB, FBB);
4554 // Replace the old BB with the new BB.
4555 for (auto &I : *TBB) {
4556 PHINode *PN = dyn_cast<PHINode>(&I);
4560 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
4561 PN->setIncomingBlock(i, TmpBB);
4564 // Add another incoming edge form the new BB.
4565 for (auto &I : *FBB) {
4566 PHINode *PN = dyn_cast<PHINode>(&I);
4569 auto *Val = PN->getIncomingValueForBlock(&BB);
4570 PN->addIncoming(Val, TmpBB);
4573 // Update the branch weights (from SelectionDAGBuilder::
4574 // FindMergedConditions).
4575 if (Opc == Instruction::Or) {
4576 // Codegen X | Y as:
4585 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
4586 // The requirement is that
4587 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
4588 // = TrueProb for orignal BB.
4589 // Assuming the orignal weights are A and B, one choice is to set BB1's
4590 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
4592 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
4593 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
4594 // TmpBB, but the math is more complicated.
4595 uint64_t TrueWeight, FalseWeight;
4596 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4597 uint64_t NewTrueWeight = TrueWeight;
4598 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
4599 scaleWeights(NewTrueWeight, NewFalseWeight);
4600 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4601 .createBranchWeights(TrueWeight, FalseWeight));
4603 NewTrueWeight = TrueWeight;
4604 NewFalseWeight = 2 * FalseWeight;
4605 scaleWeights(NewTrueWeight, NewFalseWeight);
4606 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4607 .createBranchWeights(TrueWeight, FalseWeight));
4610 // Codegen X & Y as:
4618 // This requires creation of TmpBB after CurBB.
4620 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
4621 // The requirement is that
4622 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
4623 // = FalseProb for orignal BB.
4624 // Assuming the orignal weights are A and B, one choice is to set BB1's
4625 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
4627 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
4628 uint64_t TrueWeight, FalseWeight;
4629 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4630 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
4631 uint64_t NewFalseWeight = FalseWeight;
4632 scaleWeights(NewTrueWeight, NewFalseWeight);
4633 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4634 .createBranchWeights(TrueWeight, FalseWeight));
4636 NewTrueWeight = 2 * TrueWeight;
4637 NewFalseWeight = FalseWeight;
4638 scaleWeights(NewTrueWeight, NewFalseWeight);
4639 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4640 .createBranchWeights(TrueWeight, FalseWeight));
4644 // Request DOM Tree update.
4645 // Note: No point in getting fancy here, since the DT info is never
4646 // available to CodeGenPrepare and the existing update code is broken
4652 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();