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"));
95 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
99 TypeIsSExt(Type *Ty, bool IsSExt) : Ty(Ty), IsSExt(IsSExt) {}
101 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
103 class CodeGenPrepare : public FunctionPass {
104 /// TLI - Keep a pointer of a TargetLowering to consult for determining
105 /// transformation profitability.
106 const TargetMachine *TM;
107 const TargetLowering *TLI;
108 const TargetTransformInfo *TTI;
109 const TargetLibraryInfo *TLInfo;
112 /// CurInstIterator - As we scan instructions optimizing them, this is the
113 /// next instruction to optimize. Xforms that can invalidate this should
115 BasicBlock::iterator CurInstIterator;
117 /// Keeps track of non-local addresses that have been sunk into a block.
118 /// This allows us to avoid inserting duplicate code for blocks with
119 /// multiple load/stores of the same address.
120 ValueMap<Value*, Value*> SunkAddrs;
122 /// Keeps track of all truncates inserted for the current function.
123 SetOfInstrs InsertedTruncsSet;
124 /// Keeps track of the type of the related instruction before their
125 /// promotion for the current function.
126 InstrToOrigTy PromotedInsts;
128 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
132 /// OptSize - True if optimizing for size.
136 static char ID; // Pass identification, replacement for typeid
137 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
138 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr) {
139 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
141 bool runOnFunction(Function &F) override;
143 const char *getPassName() const override { return "CodeGen Prepare"; }
145 void getAnalysisUsage(AnalysisUsage &AU) const override {
146 AU.addPreserved<DominatorTreeWrapperPass>();
147 AU.addRequired<TargetLibraryInfo>();
148 AU.addRequired<TargetTransformInfo>();
152 bool EliminateFallThrough(Function &F);
153 bool EliminateMostlyEmptyBlocks(Function &F);
154 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
155 void EliminateMostlyEmptyBlock(BasicBlock *BB);
156 bool OptimizeBlock(BasicBlock &BB);
157 bool OptimizeInst(Instruction *I);
158 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
159 bool OptimizeInlineAsmInst(CallInst *CS);
160 bool OptimizeCallInst(CallInst *CI);
161 bool MoveExtToFormExtLoad(Instruction *I);
162 bool OptimizeExtUses(Instruction *I);
163 bool OptimizeSelectInst(SelectInst *SI);
164 bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
165 bool OptimizeExtractElementInst(Instruction *Inst);
166 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
167 bool PlaceDbgValues(Function &F);
168 bool sinkAndCmp(Function &F);
169 bool splitBranchCondition(Function &F);
173 char CodeGenPrepare::ID = 0;
174 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
175 "Optimize for code generation", false, false)
177 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
178 return new CodeGenPrepare(TM);
181 bool CodeGenPrepare::runOnFunction(Function &F) {
182 if (skipOptnoneFunction(F))
185 bool EverMadeChange = false;
186 // Clear per function information.
187 InsertedTruncsSet.clear();
188 PromotedInsts.clear();
192 TLI = TM->getSubtargetImpl()->getTargetLowering();
193 TLInfo = &getAnalysis<TargetLibraryInfo>();
194 TTI = &getAnalysis<TargetTransformInfo>();
195 DominatorTreeWrapperPass *DTWP =
196 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
197 DT = DTWP ? &DTWP->getDomTree() : nullptr;
198 OptSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
199 Attribute::OptimizeForSize);
201 /// This optimization identifies DIV instructions that can be
202 /// profitably bypassed and carried out with a shorter, faster divide.
203 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
204 const DenseMap<unsigned int, unsigned int> &BypassWidths =
205 TLI->getBypassSlowDivWidths();
206 for (Function::iterator I = F.begin(); I != F.end(); I++)
207 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
210 // Eliminate blocks that contain only PHI nodes and an
211 // unconditional branch.
212 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
214 // llvm.dbg.value is far away from the value then iSel may not be able
215 // handle it properly. iSel will drop llvm.dbg.value if it can not
216 // find a node corresponding to the value.
217 EverMadeChange |= PlaceDbgValues(F);
219 // If there is a mask, compare against zero, and branch that can be combined
220 // into a single target instruction, push the mask and compare into branch
221 // users. Do this before OptimizeBlock -> OptimizeInst ->
222 // OptimizeCmpExpression, which perturbs the pattern being searched for.
223 if (!DisableBranchOpts) {
224 EverMadeChange |= sinkAndCmp(F);
225 EverMadeChange |= splitBranchCondition(F);
228 bool MadeChange = true;
231 for (Function::iterator I = F.begin(); I != F.end(); ) {
232 BasicBlock *BB = I++;
233 MadeChange |= OptimizeBlock(*BB);
235 EverMadeChange |= MadeChange;
240 if (!DisableBranchOpts) {
242 SmallPtrSet<BasicBlock*, 8> WorkList;
243 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
244 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
245 MadeChange |= ConstantFoldTerminator(BB, true);
246 if (!MadeChange) continue;
248 for (SmallVectorImpl<BasicBlock*>::iterator
249 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
250 if (pred_begin(*II) == pred_end(*II))
251 WorkList.insert(*II);
254 // Delete the dead blocks and any of their dead successors.
255 MadeChange |= !WorkList.empty();
256 while (!WorkList.empty()) {
257 BasicBlock *BB = *WorkList.begin();
259 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
263 for (SmallVectorImpl<BasicBlock*>::iterator
264 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
265 if (pred_begin(*II) == pred_end(*II))
266 WorkList.insert(*II);
269 // Merge pairs of basic blocks with unconditional branches, connected by
271 if (EverMadeChange || MadeChange)
272 MadeChange |= EliminateFallThrough(F);
276 EverMadeChange |= MadeChange;
279 if (ModifiedDT && DT)
282 return EverMadeChange;
285 /// EliminateFallThrough - Merge basic blocks which are connected
286 /// by a single edge, where one of the basic blocks has a single successor
287 /// pointing to the other basic block, which has a single predecessor.
288 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
289 bool Changed = false;
290 // Scan all of the blocks in the function, except for the entry block.
291 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
292 BasicBlock *BB = I++;
293 // If the destination block has a single pred, then this is a trivial
294 // edge, just collapse it.
295 BasicBlock *SinglePred = BB->getSinglePredecessor();
297 // Don't merge if BB's address is taken.
298 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
300 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
301 if (Term && !Term->isConditional()) {
303 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
304 // Remember if SinglePred was the entry block of the function.
305 // If so, we will need to move BB back to the entry position.
306 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
307 MergeBasicBlockIntoOnlyPred(BB, this);
309 if (isEntry && BB != &BB->getParent()->getEntryBlock())
310 BB->moveBefore(&BB->getParent()->getEntryBlock());
312 // We have erased a block. Update the iterator.
319 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
320 /// debug info directives, and an unconditional branch. Passes before isel
321 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
322 /// isel. Start by eliminating these blocks so we can split them the way we
324 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
325 bool MadeChange = false;
326 // Note that this intentionally skips the entry block.
327 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
328 BasicBlock *BB = I++;
330 // If this block doesn't end with an uncond branch, ignore it.
331 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
332 if (!BI || !BI->isUnconditional())
335 // If the instruction before the branch (skipping debug info) isn't a phi
336 // node, then other stuff is happening here.
337 BasicBlock::iterator BBI = BI;
338 if (BBI != BB->begin()) {
340 while (isa<DbgInfoIntrinsic>(BBI)) {
341 if (BBI == BB->begin())
345 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
349 // Do not break infinite loops.
350 BasicBlock *DestBB = BI->getSuccessor(0);
354 if (!CanMergeBlocks(BB, DestBB))
357 EliminateMostlyEmptyBlock(BB);
363 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
364 /// single uncond branch between them, and BB contains no other non-phi
366 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
367 const BasicBlock *DestBB) const {
368 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
369 // the successor. If there are more complex condition (e.g. preheaders),
370 // don't mess around with them.
371 BasicBlock::const_iterator BBI = BB->begin();
372 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
373 for (const User *U : PN->users()) {
374 const Instruction *UI = cast<Instruction>(U);
375 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
377 // If User is inside DestBB block and it is a PHINode then check
378 // incoming value. If incoming value is not from BB then this is
379 // a complex condition (e.g. preheaders) we want to avoid here.
380 if (UI->getParent() == DestBB) {
381 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
382 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
383 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
384 if (Insn && Insn->getParent() == BB &&
385 Insn->getParent() != UPN->getIncomingBlock(I))
392 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
393 // and DestBB may have conflicting incoming values for the block. If so, we
394 // can't merge the block.
395 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
396 if (!DestBBPN) return true; // no conflict.
398 // Collect the preds of BB.
399 SmallPtrSet<const BasicBlock*, 16> BBPreds;
400 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
401 // It is faster to get preds from a PHI than with pred_iterator.
402 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
403 BBPreds.insert(BBPN->getIncomingBlock(i));
405 BBPreds.insert(pred_begin(BB), pred_end(BB));
408 // Walk the preds of DestBB.
409 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
410 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
411 if (BBPreds.count(Pred)) { // Common predecessor?
412 BBI = DestBB->begin();
413 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
414 const Value *V1 = PN->getIncomingValueForBlock(Pred);
415 const Value *V2 = PN->getIncomingValueForBlock(BB);
417 // If V2 is a phi node in BB, look up what the mapped value will be.
418 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
419 if (V2PN->getParent() == BB)
420 V2 = V2PN->getIncomingValueForBlock(Pred);
422 // If there is a conflict, bail out.
423 if (V1 != V2) return false;
432 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
433 /// an unconditional branch in it.
434 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
435 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
436 BasicBlock *DestBB = BI->getSuccessor(0);
438 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
440 // If the destination block has a single pred, then this is a trivial edge,
442 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
443 if (SinglePred != DestBB) {
444 // Remember if SinglePred was the entry block of the function. If so, we
445 // will need to move BB back to the entry position.
446 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
447 MergeBasicBlockIntoOnlyPred(DestBB, this);
449 if (isEntry && BB != &BB->getParent()->getEntryBlock())
450 BB->moveBefore(&BB->getParent()->getEntryBlock());
452 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
457 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
458 // to handle the new incoming edges it is about to have.
460 for (BasicBlock::iterator BBI = DestBB->begin();
461 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
462 // Remove the incoming value for BB, and remember it.
463 Value *InVal = PN->removeIncomingValue(BB, false);
465 // Two options: either the InVal is a phi node defined in BB or it is some
466 // value that dominates BB.
467 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
468 if (InValPhi && InValPhi->getParent() == BB) {
469 // Add all of the input values of the input PHI as inputs of this phi.
470 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
471 PN->addIncoming(InValPhi->getIncomingValue(i),
472 InValPhi->getIncomingBlock(i));
474 // Otherwise, add one instance of the dominating value for each edge that
475 // we will be adding.
476 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
477 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
478 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
480 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
481 PN->addIncoming(InVal, *PI);
486 // The PHIs are now updated, change everything that refers to BB to use
487 // DestBB and remove BB.
488 BB->replaceAllUsesWith(DestBB);
489 if (DT && !ModifiedDT) {
490 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
491 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
492 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
493 DT->changeImmediateDominator(DestBB, NewIDom);
496 BB->eraseFromParent();
499 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
502 /// SinkCast - Sink the specified cast instruction into its user blocks
503 static bool SinkCast(CastInst *CI) {
504 BasicBlock *DefBB = CI->getParent();
506 /// InsertedCasts - Only insert a cast in each block once.
507 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
509 bool MadeChange = false;
510 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
512 Use &TheUse = UI.getUse();
513 Instruction *User = cast<Instruction>(*UI);
515 // Figure out which BB this cast is used in. For PHI's this is the
516 // appropriate predecessor block.
517 BasicBlock *UserBB = User->getParent();
518 if (PHINode *PN = dyn_cast<PHINode>(User)) {
519 UserBB = PN->getIncomingBlock(TheUse);
522 // Preincrement use iterator so we don't invalidate it.
525 // If this user is in the same block as the cast, don't change the cast.
526 if (UserBB == DefBB) continue;
528 // If we have already inserted a cast into this block, use it.
529 CastInst *&InsertedCast = InsertedCasts[UserBB];
532 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
534 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
539 // Replace a use of the cast with a use of the new cast.
540 TheUse = InsertedCast;
544 // If we removed all uses, nuke the cast.
545 if (CI->use_empty()) {
546 CI->eraseFromParent();
553 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
554 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
555 /// sink it into user blocks to reduce the number of virtual
556 /// registers that must be created and coalesced.
558 /// Return true if any changes are made.
560 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
561 // If this is a noop copy,
562 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
563 EVT DstVT = TLI.getValueType(CI->getType());
565 // This is an fp<->int conversion?
566 if (SrcVT.isInteger() != DstVT.isInteger())
569 // If this is an extension, it will be a zero or sign extension, which
571 if (SrcVT.bitsLT(DstVT)) return false;
573 // If these values will be promoted, find out what they will be promoted
574 // to. This helps us consider truncates on PPC as noop copies when they
576 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
577 TargetLowering::TypePromoteInteger)
578 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
579 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
580 TargetLowering::TypePromoteInteger)
581 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
583 // If, after promotion, these are the same types, this is a noop copy.
590 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
591 /// the number of virtual registers that must be created and coalesced. This is
592 /// a clear win except on targets with multiple condition code registers
593 /// (PowerPC), where it might lose; some adjustment may be wanted there.
595 /// Return true if any changes are made.
596 static bool OptimizeCmpExpression(CmpInst *CI) {
597 BasicBlock *DefBB = CI->getParent();
599 /// InsertedCmp - Only insert a cmp in each block once.
600 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
602 bool MadeChange = false;
603 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
605 Use &TheUse = UI.getUse();
606 Instruction *User = cast<Instruction>(*UI);
608 // Preincrement use iterator so we don't invalidate it.
611 // Don't bother for PHI nodes.
612 if (isa<PHINode>(User))
615 // Figure out which BB this cmp is used in.
616 BasicBlock *UserBB = User->getParent();
618 // If this user is in the same block as the cmp, don't change the cmp.
619 if (UserBB == DefBB) continue;
621 // If we have already inserted a cmp into this block, use it.
622 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
625 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
627 CmpInst::Create(CI->getOpcode(),
628 CI->getPredicate(), CI->getOperand(0),
629 CI->getOperand(1), "", InsertPt);
633 // Replace a use of the cmp with a use of the new cmp.
634 TheUse = InsertedCmp;
638 // If we removed all uses, nuke the cmp.
640 CI->eraseFromParent();
645 /// isExtractBitsCandidateUse - Check if the candidates could
646 /// be combined with shift instruction, which includes:
647 /// 1. Truncate instruction
648 /// 2. And instruction and the imm is a mask of the low bits:
649 /// imm & (imm+1) == 0
650 static bool isExtractBitsCandidateUse(Instruction *User) {
651 if (!isa<TruncInst>(User)) {
652 if (User->getOpcode() != Instruction::And ||
653 !isa<ConstantInt>(User->getOperand(1)))
656 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
658 if ((Cimm & (Cimm + 1)).getBoolValue())
664 /// SinkShiftAndTruncate - sink both shift and truncate instruction
665 /// to the use of truncate's BB.
667 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
668 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
669 const TargetLowering &TLI) {
670 BasicBlock *UserBB = User->getParent();
671 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
672 TruncInst *TruncI = dyn_cast<TruncInst>(User);
673 bool MadeChange = false;
675 for (Value::user_iterator TruncUI = TruncI->user_begin(),
676 TruncE = TruncI->user_end();
677 TruncUI != TruncE;) {
679 Use &TruncTheUse = TruncUI.getUse();
680 Instruction *TruncUser = cast<Instruction>(*TruncUI);
681 // Preincrement use iterator so we don't invalidate it.
685 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
689 // If the use is actually a legal node, there will not be an
690 // implicit truncate.
691 // FIXME: always querying the result type is just an
692 // approximation; some nodes' legality is determined by the
693 // operand or other means. There's no good way to find out though.
694 if (TLI.isOperationLegalOrCustom(
695 ISDOpcode, TLI.getValueType(TruncUser->getType(), true)))
698 // Don't bother for PHI nodes.
699 if (isa<PHINode>(TruncUser))
702 BasicBlock *TruncUserBB = TruncUser->getParent();
704 if (UserBB == TruncUserBB)
707 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
708 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
710 if (!InsertedShift && !InsertedTrunc) {
711 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
713 if (ShiftI->getOpcode() == Instruction::AShr)
715 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
718 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
721 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
724 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
725 TruncI->getType(), "", TruncInsertPt);
729 TruncTheUse = InsertedTrunc;
735 /// OptimizeExtractBits - sink the shift *right* instruction into user blocks if
736 /// the uses could potentially be combined with this shift instruction and
737 /// generate BitExtract instruction. It will only be applied if the architecture
738 /// supports BitExtract instruction. Here is an example:
740 /// %x.extract.shift = lshr i64 %arg1, 32
742 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
746 /// %x.extract.shift.1 = lshr i64 %arg1, 32
747 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
749 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
751 /// Return true if any changes are made.
752 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
753 const TargetLowering &TLI) {
754 BasicBlock *DefBB = ShiftI->getParent();
756 /// Only insert instructions in each block once.
757 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
759 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(ShiftI->getType()));
761 bool MadeChange = false;
762 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
764 Use &TheUse = UI.getUse();
765 Instruction *User = cast<Instruction>(*UI);
766 // Preincrement use iterator so we don't invalidate it.
769 // Don't bother for PHI nodes.
770 if (isa<PHINode>(User))
773 if (!isExtractBitsCandidateUse(User))
776 BasicBlock *UserBB = User->getParent();
778 if (UserBB == DefBB) {
779 // If the shift and truncate instruction are in the same BB. The use of
780 // the truncate(TruncUse) may still introduce another truncate if not
781 // legal. In this case, we would like to sink both shift and truncate
782 // instruction to the BB of TruncUse.
785 // i64 shift.result = lshr i64 opnd, imm
786 // trunc.result = trunc shift.result to i16
789 // ----> We will have an implicit truncate here if the architecture does
790 // not have i16 compare.
791 // cmp i16 trunc.result, opnd2
793 if (isa<TruncInst>(User) && shiftIsLegal
794 // If the type of the truncate is legal, no trucate will be
795 // introduced in other basic blocks.
796 && (!TLI.isTypeLegal(TLI.getValueType(User->getType()))))
798 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI);
802 // If we have already inserted a shift into this block, use it.
803 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
805 if (!InsertedShift) {
806 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
808 if (ShiftI->getOpcode() == Instruction::AShr)
810 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
813 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
818 // Replace a use of the shift with a use of the new shift.
819 TheUse = InsertedShift;
822 // If we removed all uses, nuke the shift.
823 if (ShiftI->use_empty())
824 ShiftI->eraseFromParent();
830 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
832 void replaceCall(Value *With) override {
833 CI->replaceAllUsesWith(With);
834 CI->eraseFromParent();
836 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const override {
837 if (ConstantInt *SizeCI =
838 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
839 return SizeCI->isAllOnesValue();
843 } // end anonymous namespace
845 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
846 BasicBlock *BB = CI->getParent();
848 // Lower inline assembly if we can.
849 // If we found an inline asm expession, and if the target knows how to
850 // lower it to normal LLVM code, do so now.
851 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
852 if (TLI->ExpandInlineAsm(CI)) {
853 // Avoid invalidating the iterator.
854 CurInstIterator = BB->begin();
855 // Avoid processing instructions out of order, which could cause
856 // reuse before a value is defined.
860 // Sink address computing for memory operands into the block.
861 if (OptimizeInlineAsmInst(CI))
865 // Lower all uses of llvm.objectsize.*
866 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
867 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
868 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
869 Type *ReturnTy = CI->getType();
870 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
872 // Substituting this can cause recursive simplifications, which can
873 // invalidate our iterator. Use a WeakVH to hold onto it in case this
875 WeakVH IterHandle(CurInstIterator);
877 replaceAndRecursivelySimplify(CI, RetVal,
878 TLI ? TLI->getDataLayout() : nullptr,
879 TLInfo, ModifiedDT ? nullptr : DT);
881 // If the iterator instruction was recursively deleted, start over at the
882 // start of the block.
883 if (IterHandle != CurInstIterator) {
884 CurInstIterator = BB->begin();
891 SmallVector<Value*, 2> PtrOps;
893 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
894 while (!PtrOps.empty())
895 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
899 // From here on out we're working with named functions.
900 if (!CI->getCalledFunction()) return false;
902 // We'll need DataLayout from here on out.
903 const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
904 if (!TD) return false;
906 // Lower all default uses of _chk calls. This is very similar
907 // to what InstCombineCalls does, but here we are only lowering calls
908 // that have the default "don't know" as the objectsize. Anything else
909 // should be left alone.
910 CodeGenPrepareFortifiedLibCalls Simplifier;
911 return Simplifier.fold(CI, TD, TLInfo);
914 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
915 /// instructions to the predecessor to enable tail call optimizations. The
916 /// case it is currently looking for is:
919 /// %tmp0 = tail call i32 @f0()
922 /// %tmp1 = tail call i32 @f1()
925 /// %tmp2 = tail call i32 @f2()
928 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
936 /// %tmp0 = tail call i32 @f0()
939 /// %tmp1 = tail call i32 @f1()
942 /// %tmp2 = tail call i32 @f2()
945 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
949 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
953 PHINode *PN = nullptr;
954 BitCastInst *BCI = nullptr;
955 Value *V = RI->getReturnValue();
957 BCI = dyn_cast<BitCastInst>(V);
959 V = BCI->getOperand(0);
961 PN = dyn_cast<PHINode>(V);
966 if (PN && PN->getParent() != BB)
969 // It's not safe to eliminate the sign / zero extension of the return value.
970 // See llvm::isInTailCallPosition().
971 const Function *F = BB->getParent();
972 AttributeSet CallerAttrs = F->getAttributes();
973 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
974 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
977 // Make sure there are no instructions between the PHI and return, or that the
978 // return is the first instruction in the block.
980 BasicBlock::iterator BI = BB->begin();
981 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
983 // Also skip over the bitcast.
988 BasicBlock::iterator BI = BB->begin();
989 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
994 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
996 SmallVector<CallInst*, 4> TailCalls;
998 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
999 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1000 // Make sure the phi value is indeed produced by the tail call.
1001 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1002 TLI->mayBeEmittedAsTailCall(CI))
1003 TailCalls.push_back(CI);
1006 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1007 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1008 if (!VisitedBBs.insert(*PI).second)
1011 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1012 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1013 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1014 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1018 CallInst *CI = dyn_cast<CallInst>(&*RI);
1019 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1020 TailCalls.push_back(CI);
1024 bool Changed = false;
1025 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1026 CallInst *CI = TailCalls[i];
1029 // Conservatively require the attributes of the call to match those of the
1030 // return. Ignore noalias because it doesn't affect the call sequence.
1031 AttributeSet CalleeAttrs = CS.getAttributes();
1032 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1033 removeAttribute(Attribute::NoAlias) !=
1034 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1035 removeAttribute(Attribute::NoAlias))
1038 // Make sure the call instruction is followed by an unconditional branch to
1039 // the return block.
1040 BasicBlock *CallBB = CI->getParent();
1041 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1042 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1045 // Duplicate the return into CallBB.
1046 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1047 ModifiedDT = Changed = true;
1051 // If we eliminated all predecessors of the block, delete the block now.
1052 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1053 BB->eraseFromParent();
1058 //===----------------------------------------------------------------------===//
1059 // Memory Optimization
1060 //===----------------------------------------------------------------------===//
1064 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1065 /// which holds actual Value*'s for register values.
1066 struct ExtAddrMode : public TargetLowering::AddrMode {
1069 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1070 void print(raw_ostream &OS) const;
1073 bool operator==(const ExtAddrMode& O) const {
1074 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1075 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1076 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1081 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1087 void ExtAddrMode::print(raw_ostream &OS) const {
1088 bool NeedPlus = false;
1091 OS << (NeedPlus ? " + " : "")
1093 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1098 OS << (NeedPlus ? " + " : "")
1104 OS << (NeedPlus ? " + " : "")
1106 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1110 OS << (NeedPlus ? " + " : "")
1112 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1118 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1119 void ExtAddrMode::dump() const {
1125 /// \brief This class provides transaction based operation on the IR.
1126 /// Every change made through this class is recorded in the internal state and
1127 /// can be undone (rollback) until commit is called.
1128 class TypePromotionTransaction {
1130 /// \brief This represents the common interface of the individual transaction.
1131 /// Each class implements the logic for doing one specific modification on
1132 /// the IR via the TypePromotionTransaction.
1133 class TypePromotionAction {
1135 /// The Instruction modified.
1139 /// \brief Constructor of the action.
1140 /// The constructor performs the related action on the IR.
1141 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1143 virtual ~TypePromotionAction() {}
1145 /// \brief Undo the modification done by this action.
1146 /// When this method is called, the IR must be in the same state as it was
1147 /// before this action was applied.
1148 /// \pre Undoing the action works if and only if the IR is in the exact same
1149 /// state as it was directly after this action was applied.
1150 virtual void undo() = 0;
1152 /// \brief Advocate every change made by this action.
1153 /// When the results on the IR of the action are to be kept, it is important
1154 /// to call this function, otherwise hidden information may be kept forever.
1155 virtual void commit() {
1156 // Nothing to be done, this action is not doing anything.
1160 /// \brief Utility to remember the position of an instruction.
1161 class InsertionHandler {
1162 /// Position of an instruction.
1163 /// Either an instruction:
1164 /// - Is the first in a basic block: BB is used.
1165 /// - Has a previous instructon: PrevInst is used.
1167 Instruction *PrevInst;
1170 /// Remember whether or not the instruction had a previous instruction.
1171 bool HasPrevInstruction;
1174 /// \brief Record the position of \p Inst.
1175 InsertionHandler(Instruction *Inst) {
1176 BasicBlock::iterator It = Inst;
1177 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1178 if (HasPrevInstruction)
1179 Point.PrevInst = --It;
1181 Point.BB = Inst->getParent();
1184 /// \brief Insert \p Inst at the recorded position.
1185 void insert(Instruction *Inst) {
1186 if (HasPrevInstruction) {
1187 if (Inst->getParent())
1188 Inst->removeFromParent();
1189 Inst->insertAfter(Point.PrevInst);
1191 Instruction *Position = Point.BB->getFirstInsertionPt();
1192 if (Inst->getParent())
1193 Inst->moveBefore(Position);
1195 Inst->insertBefore(Position);
1200 /// \brief Move an instruction before another.
1201 class InstructionMoveBefore : public TypePromotionAction {
1202 /// Original position of the instruction.
1203 InsertionHandler Position;
1206 /// \brief Move \p Inst before \p Before.
1207 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1208 : TypePromotionAction(Inst), Position(Inst) {
1209 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1210 Inst->moveBefore(Before);
1213 /// \brief Move the instruction back to its original position.
1214 void undo() override {
1215 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1216 Position.insert(Inst);
1220 /// \brief Set the operand of an instruction with a new value.
1221 class OperandSetter : public TypePromotionAction {
1222 /// Original operand of the instruction.
1224 /// Index of the modified instruction.
1228 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1229 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1230 : TypePromotionAction(Inst), Idx(Idx) {
1231 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1232 << "for:" << *Inst << "\n"
1233 << "with:" << *NewVal << "\n");
1234 Origin = Inst->getOperand(Idx);
1235 Inst->setOperand(Idx, NewVal);
1238 /// \brief Restore the original value of the instruction.
1239 void undo() override {
1240 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1241 << "for: " << *Inst << "\n"
1242 << "with: " << *Origin << "\n");
1243 Inst->setOperand(Idx, Origin);
1247 /// \brief Hide the operands of an instruction.
1248 /// Do as if this instruction was not using any of its operands.
1249 class OperandsHider : public TypePromotionAction {
1250 /// The list of original operands.
1251 SmallVector<Value *, 4> OriginalValues;
1254 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1255 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1256 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1257 unsigned NumOpnds = Inst->getNumOperands();
1258 OriginalValues.reserve(NumOpnds);
1259 for (unsigned It = 0; It < NumOpnds; ++It) {
1260 // Save the current operand.
1261 Value *Val = Inst->getOperand(It);
1262 OriginalValues.push_back(Val);
1264 // We could use OperandSetter here, but that would implied an overhead
1265 // that we are not willing to pay.
1266 Inst->setOperand(It, UndefValue::get(Val->getType()));
1270 /// \brief Restore the original list of uses.
1271 void undo() override {
1272 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1273 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1274 Inst->setOperand(It, OriginalValues[It]);
1278 /// \brief Build a truncate instruction.
1279 class TruncBuilder : public TypePromotionAction {
1282 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1284 /// trunc Opnd to Ty.
1285 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1286 IRBuilder<> Builder(Opnd);
1287 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
1288 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
1291 /// \brief Get the built value.
1292 Value *getBuiltValue() { return Val; }
1294 /// \brief Remove the built instruction.
1295 void undo() override {
1296 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
1297 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1298 IVal->eraseFromParent();
1302 /// \brief Build a sign extension instruction.
1303 class SExtBuilder : public TypePromotionAction {
1306 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1308 /// sext Opnd to Ty.
1309 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1310 : TypePromotionAction(InsertPt) {
1311 IRBuilder<> Builder(InsertPt);
1312 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
1313 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
1316 /// \brief Get the built value.
1317 Value *getBuiltValue() { return Val; }
1319 /// \brief Remove the built instruction.
1320 void undo() override {
1321 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
1322 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1323 IVal->eraseFromParent();
1327 /// \brief Build a zero extension instruction.
1328 class ZExtBuilder : public TypePromotionAction {
1331 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
1333 /// zext Opnd to Ty.
1334 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1335 : TypePromotionAction(InsertPt) {
1336 IRBuilder<> Builder(InsertPt);
1337 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
1338 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
1341 /// \brief Get the built value.
1342 Value *getBuiltValue() { return Val; }
1344 /// \brief Remove the built instruction.
1345 void undo() override {
1346 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
1347 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1348 IVal->eraseFromParent();
1352 /// \brief Mutate an instruction to another type.
1353 class TypeMutator : public TypePromotionAction {
1354 /// Record the original type.
1358 /// \brief Mutate the type of \p Inst into \p NewTy.
1359 TypeMutator(Instruction *Inst, Type *NewTy)
1360 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1361 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1363 Inst->mutateType(NewTy);
1366 /// \brief Mutate the instruction back to its original type.
1367 void undo() override {
1368 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1370 Inst->mutateType(OrigTy);
1374 /// \brief Replace the uses of an instruction by another instruction.
1375 class UsesReplacer : public TypePromotionAction {
1376 /// Helper structure to keep track of the replaced uses.
1377 struct InstructionAndIdx {
1378 /// The instruction using the instruction.
1380 /// The index where this instruction is used for Inst.
1382 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1383 : Inst(Inst), Idx(Idx) {}
1386 /// Keep track of the original uses (pair Instruction, Index).
1387 SmallVector<InstructionAndIdx, 4> OriginalUses;
1388 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1391 /// \brief Replace all the use of \p Inst by \p New.
1392 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1393 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1395 // Record the original uses.
1396 for (Use &U : Inst->uses()) {
1397 Instruction *UserI = cast<Instruction>(U.getUser());
1398 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1400 // Now, we can replace the uses.
1401 Inst->replaceAllUsesWith(New);
1404 /// \brief Reassign the original uses of Inst to Inst.
1405 void undo() override {
1406 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1407 for (use_iterator UseIt = OriginalUses.begin(),
1408 EndIt = OriginalUses.end();
1409 UseIt != EndIt; ++UseIt) {
1410 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1415 /// \brief Remove an instruction from the IR.
1416 class InstructionRemover : public TypePromotionAction {
1417 /// Original position of the instruction.
1418 InsertionHandler Inserter;
1419 /// Helper structure to hide all the link to the instruction. In other
1420 /// words, this helps to do as if the instruction was removed.
1421 OperandsHider Hider;
1422 /// Keep track of the uses replaced, if any.
1423 UsesReplacer *Replacer;
1426 /// \brief Remove all reference of \p Inst and optinally replace all its
1428 /// \pre If !Inst->use_empty(), then New != nullptr
1429 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1430 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1433 Replacer = new UsesReplacer(Inst, New);
1434 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1435 Inst->removeFromParent();
1438 ~InstructionRemover() { delete Replacer; }
1440 /// \brief Really remove the instruction.
1441 void commit() override { delete Inst; }
1443 /// \brief Resurrect the instruction and reassign it to the proper uses if
1444 /// new value was provided when build this action.
1445 void undo() override {
1446 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1447 Inserter.insert(Inst);
1455 /// Restoration point.
1456 /// The restoration point is a pointer to an action instead of an iterator
1457 /// because the iterator may be invalidated but not the pointer.
1458 typedef const TypePromotionAction *ConstRestorationPt;
1459 /// Advocate every changes made in that transaction.
1461 /// Undo all the changes made after the given point.
1462 void rollback(ConstRestorationPt Point);
1463 /// Get the current restoration point.
1464 ConstRestorationPt getRestorationPoint() const;
1466 /// \name API for IR modification with state keeping to support rollback.
1468 /// Same as Instruction::setOperand.
1469 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1470 /// Same as Instruction::eraseFromParent.
1471 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
1472 /// Same as Value::replaceAllUsesWith.
1473 void replaceAllUsesWith(Instruction *Inst, Value *New);
1474 /// Same as Value::mutateType.
1475 void mutateType(Instruction *Inst, Type *NewTy);
1476 /// Same as IRBuilder::createTrunc.
1477 Value *createTrunc(Instruction *Opnd, Type *Ty);
1478 /// Same as IRBuilder::createSExt.
1479 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1480 /// Same as IRBuilder::createZExt.
1481 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
1482 /// Same as Instruction::moveBefore.
1483 void moveBefore(Instruction *Inst, Instruction *Before);
1487 /// The ordered list of actions made so far.
1488 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
1489 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
1492 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1495 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
1498 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1501 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
1504 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1506 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
1509 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1510 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
1513 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1515 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
1516 Value *Val = Ptr->getBuiltValue();
1517 Actions.push_back(std::move(Ptr));
1521 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
1522 Value *Opnd, Type *Ty) {
1523 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
1524 Value *Val = Ptr->getBuiltValue();
1525 Actions.push_back(std::move(Ptr));
1529 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
1530 Value *Opnd, Type *Ty) {
1531 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
1532 Value *Val = Ptr->getBuiltValue();
1533 Actions.push_back(std::move(Ptr));
1537 void TypePromotionTransaction::moveBefore(Instruction *Inst,
1538 Instruction *Before) {
1540 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
1543 TypePromotionTransaction::ConstRestorationPt
1544 TypePromotionTransaction::getRestorationPoint() const {
1545 return !Actions.empty() ? Actions.back().get() : nullptr;
1548 void TypePromotionTransaction::commit() {
1549 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
1555 void TypePromotionTransaction::rollback(
1556 TypePromotionTransaction::ConstRestorationPt Point) {
1557 while (!Actions.empty() && Point != Actions.back().get()) {
1558 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
1563 /// \brief A helper class for matching addressing modes.
1565 /// This encapsulates the logic for matching the target-legal addressing modes.
1566 class AddressingModeMatcher {
1567 SmallVectorImpl<Instruction*> &AddrModeInsts;
1568 const TargetLowering &TLI;
1570 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
1571 /// the memory instruction that we're computing this address for.
1573 Instruction *MemoryInst;
1575 /// AddrMode - This is the addressing mode that we're building up. This is
1576 /// part of the return value of this addressing mode matching stuff.
1577 ExtAddrMode &AddrMode;
1579 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
1580 const SetOfInstrs &InsertedTruncs;
1581 /// A map from the instructions to their type before promotion.
1582 InstrToOrigTy &PromotedInsts;
1583 /// The ongoing transaction where every action should be registered.
1584 TypePromotionTransaction &TPT;
1586 /// IgnoreProfitability - This is set to true when we should not do
1587 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
1588 /// always returns true.
1589 bool IgnoreProfitability;
1591 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
1592 const TargetLowering &T, Type *AT,
1593 Instruction *MI, ExtAddrMode &AM,
1594 const SetOfInstrs &InsertedTruncs,
1595 InstrToOrigTy &PromotedInsts,
1596 TypePromotionTransaction &TPT)
1597 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM),
1598 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
1599 IgnoreProfitability = false;
1603 /// Match - Find the maximal addressing mode that a load/store of V can fold,
1604 /// give an access type of AccessTy. This returns a list of involved
1605 /// instructions in AddrModeInsts.
1606 /// \p InsertedTruncs The truncate instruction inserted by other
1609 /// \p PromotedInsts maps the instructions to their type before promotion.
1610 /// \p The ongoing transaction where every action should be registered.
1611 static ExtAddrMode Match(Value *V, Type *AccessTy,
1612 Instruction *MemoryInst,
1613 SmallVectorImpl<Instruction*> &AddrModeInsts,
1614 const TargetLowering &TLI,
1615 const SetOfInstrs &InsertedTruncs,
1616 InstrToOrigTy &PromotedInsts,
1617 TypePromotionTransaction &TPT) {
1620 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
1621 MemoryInst, Result, InsertedTruncs,
1622 PromotedInsts, TPT).MatchAddr(V, 0);
1623 (void)Success; assert(Success && "Couldn't select *anything*?");
1627 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
1628 bool MatchAddr(Value *V, unsigned Depth);
1629 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
1630 bool *MovedAway = nullptr);
1631 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
1632 ExtAddrMode &AMBefore,
1633 ExtAddrMode &AMAfter);
1634 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
1635 bool IsPromotionProfitable(unsigned MatchedSize, unsigned SizeWithPromotion,
1636 Value *PromotedOperand) const;
1639 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
1640 /// Return true and update AddrMode if this addr mode is legal for the target,
1642 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
1644 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
1645 // mode. Just process that directly.
1647 return MatchAddr(ScaleReg, Depth);
1649 // If the scale is 0, it takes nothing to add this.
1653 // If we already have a scale of this value, we can add to it, otherwise, we
1654 // need an available scale field.
1655 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
1658 ExtAddrMode TestAddrMode = AddrMode;
1660 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
1661 // [A+B + A*7] -> [B+A*8].
1662 TestAddrMode.Scale += Scale;
1663 TestAddrMode.ScaledReg = ScaleReg;
1665 // If the new address isn't legal, bail out.
1666 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
1669 // It was legal, so commit it.
1670 AddrMode = TestAddrMode;
1672 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
1673 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
1674 // X*Scale + C*Scale to addr mode.
1675 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
1676 if (isa<Instruction>(ScaleReg) && // not a constant expr.
1677 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
1678 TestAddrMode.ScaledReg = AddLHS;
1679 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
1681 // If this addressing mode is legal, commit it and remember that we folded
1682 // this instruction.
1683 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
1684 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
1685 AddrMode = TestAddrMode;
1690 // Otherwise, not (x+c)*scale, just return what we have.
1694 /// MightBeFoldableInst - This is a little filter, which returns true if an
1695 /// addressing computation involving I might be folded into a load/store
1696 /// accessing it. This doesn't need to be perfect, but needs to accept at least
1697 /// the set of instructions that MatchOperationAddr can.
1698 static bool MightBeFoldableInst(Instruction *I) {
1699 switch (I->getOpcode()) {
1700 case Instruction::BitCast:
1701 case Instruction::AddrSpaceCast:
1702 // Don't touch identity bitcasts.
1703 if (I->getType() == I->getOperand(0)->getType())
1705 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
1706 case Instruction::PtrToInt:
1707 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1709 case Instruction::IntToPtr:
1710 // We know the input is intptr_t, so this is foldable.
1712 case Instruction::Add:
1714 case Instruction::Mul:
1715 case Instruction::Shl:
1716 // Can only handle X*C and X << C.
1717 return isa<ConstantInt>(I->getOperand(1));
1718 case Instruction::GetElementPtr:
1725 /// \brief Hepler class to perform type promotion.
1726 class TypePromotionHelper {
1727 /// \brief Utility function to check whether or not a sign or zero extension
1728 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
1729 /// either using the operands of \p Inst or promoting \p Inst.
1730 /// The type of the extension is defined by \p IsSExt.
1731 /// In other words, check if:
1732 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
1733 /// #1 Promotion applies:
1734 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
1735 /// #2 Operand reuses:
1736 /// ext opnd1 to ConsideredExtType.
1737 /// \p PromotedInsts maps the instructions to their type before promotion.
1738 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
1739 const InstrToOrigTy &PromotedInsts, bool IsSExt);
1741 /// \brief Utility function to determine if \p OpIdx should be promoted when
1742 /// promoting \p Inst.
1743 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
1744 if (isa<SelectInst>(Inst) && OpIdx == 0)
1749 /// \brief Utility function to promote the operand of \p Ext when this
1750 /// operand is a promotable trunc or sext or zext.
1751 /// \p PromotedInsts maps the instructions to their type before promotion.
1752 /// \p CreatedInsts[out] contains how many non-free instructions have been
1753 /// created to promote the operand of Ext.
1754 /// Should never be called directly.
1755 /// \return The promoted value which is used instead of Ext.
1756 static Value *promoteOperandForTruncAndAnyExt(Instruction *Ext,
1757 TypePromotionTransaction &TPT,
1758 InstrToOrigTy &PromotedInsts,
1759 unsigned &CreatedInsts);
1761 /// \brief Utility function to promote the operand of \p Ext when this
1762 /// operand is promotable and is not a supported trunc or sext.
1763 /// \p PromotedInsts maps the instructions to their type before promotion.
1764 /// \p CreatedInsts[out] contains how many non-free instructions have been
1765 /// created to promote the operand of Ext.
1766 /// Should never be called directly.
1767 /// \return The promoted value which is used instead of Ext.
1768 static Value *promoteOperandForOther(Instruction *Ext,
1769 TypePromotionTransaction &TPT,
1770 InstrToOrigTy &PromotedInsts,
1771 unsigned &CreatedInsts, bool IsSExt);
1773 /// \see promoteOperandForOther.
1774 static Value *signExtendOperandForOther(Instruction *Ext,
1775 TypePromotionTransaction &TPT,
1776 InstrToOrigTy &PromotedInsts,
1777 unsigned &CreatedInsts) {
1778 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInsts, true);
1781 /// \see promoteOperandForOther.
1782 static Value *zeroExtendOperandForOther(Instruction *Ext,
1783 TypePromotionTransaction &TPT,
1784 InstrToOrigTy &PromotedInsts,
1785 unsigned &CreatedInsts) {
1786 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInsts, false);
1790 /// Type for the utility function that promotes the operand of Ext.
1791 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
1792 InstrToOrigTy &PromotedInsts,
1793 unsigned &CreatedInsts);
1794 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
1795 /// action to promote the operand of \p Ext instead of using Ext.
1796 /// \return NULL if no promotable action is possible with the current
1798 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
1799 /// the others CodeGenPrepare optimizations. This information is important
1800 /// because we do not want to promote these instructions as CodeGenPrepare
1801 /// will reinsert them later. Thus creating an infinite loop: create/remove.
1802 /// \p PromotedInsts maps the instructions to their type before promotion.
1803 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedTruncs,
1804 const TargetLowering &TLI,
1805 const InstrToOrigTy &PromotedInsts);
1808 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
1809 Type *ConsideredExtType,
1810 const InstrToOrigTy &PromotedInsts,
1812 // We can always get through zext.
1813 if (isa<ZExtInst>(Inst))
1816 // sext(sext) is ok too.
1817 if (IsSExt && isa<SExtInst>(Inst))
1820 // We can get through binary operator, if it is legal. In other words, the
1821 // binary operator must have a nuw or nsw flag.
1822 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
1823 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
1824 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
1825 (IsSExt && BinOp->hasNoSignedWrap())))
1828 // Check if we can do the following simplification.
1829 // ext(trunc(opnd)) --> ext(opnd)
1830 if (!isa<TruncInst>(Inst))
1833 Value *OpndVal = Inst->getOperand(0);
1834 // Check if we can use this operand in the extension.
1835 // If the type is larger than the result type of the extension,
1837 if (OpndVal->getType()->getIntegerBitWidth() >
1838 ConsideredExtType->getIntegerBitWidth())
1841 // If the operand of the truncate is not an instruction, we will not have
1842 // any information on the dropped bits.
1843 // (Actually we could for constant but it is not worth the extra logic).
1844 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
1848 // Check if the source of the type is narrow enough.
1849 // I.e., check that trunc just drops extended bits of the same kind of
1851 // #1 get the type of the operand and check the kind of the extended bits.
1852 const Type *OpndType;
1853 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
1854 if (It != PromotedInsts.end() && It->second.IsSExt == IsSExt)
1855 OpndType = It->second.Ty;
1856 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
1857 OpndType = Opnd->getOperand(0)->getType();
1861 // #2 check that the truncate just drop extended bits.
1862 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
1868 TypePromotionHelper::Action TypePromotionHelper::getAction(
1869 Instruction *Ext, const SetOfInstrs &InsertedTruncs,
1870 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
1871 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
1872 "Unexpected instruction type");
1873 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
1874 Type *ExtTy = Ext->getType();
1875 bool IsSExt = isa<SExtInst>(Ext);
1876 // If the operand of the extension is not an instruction, we cannot
1878 // If it, check we can get through.
1879 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
1882 // Do not promote if the operand has been added by codegenprepare.
1883 // Otherwise, it means we are undoing an optimization that is likely to be
1884 // redone, thus causing potential infinite loop.
1885 if (isa<TruncInst>(ExtOpnd) && InsertedTruncs.count(ExtOpnd))
1888 // SExt or Trunc instructions.
1889 // Return the related handler.
1890 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
1891 isa<ZExtInst>(ExtOpnd))
1892 return promoteOperandForTruncAndAnyExt;
1894 // Regular instruction.
1895 // Abort early if we will have to insert non-free instructions.
1896 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
1898 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
1901 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
1902 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
1903 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts) {
1904 // By construction, the operand of SExt is an instruction. Otherwise we cannot
1905 // get through it and this method should not be called.
1906 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
1907 Value *ExtVal = SExt;
1908 if (isa<ZExtInst>(SExtOpnd)) {
1909 // Replace s|zext(zext(opnd))
1912 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
1913 TPT.replaceAllUsesWith(SExt, ZExt);
1914 TPT.eraseInstruction(SExt);
1917 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
1919 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
1923 // Remove dead code.
1924 if (SExtOpnd->use_empty())
1925 TPT.eraseInstruction(SExtOpnd);
1927 // Check if the extension is still needed.
1928 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
1929 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType())
1932 // At this point we have: ext ty opnd to ty.
1933 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
1934 Value *NextVal = ExtInst->getOperand(0);
1935 TPT.eraseInstruction(ExtInst, NextVal);
1939 Value *TypePromotionHelper::promoteOperandForOther(
1940 Instruction *Ext, TypePromotionTransaction &TPT,
1941 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts, bool IsSExt) {
1942 // By construction, the operand of Ext is an instruction. Otherwise we cannot
1943 // get through it and this method should not be called.
1944 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
1946 if (!ExtOpnd->hasOneUse()) {
1947 // ExtOpnd will be promoted.
1948 // All its uses, but Ext, will need to use a truncated value of the
1949 // promoted version.
1950 // Create the truncate now.
1951 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
1952 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
1953 ITrunc->removeFromParent();
1954 // Insert it just after the definition.
1955 ITrunc->insertAfter(ExtOpnd);
1958 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
1959 // Restore the operand of Ext (which has been replace by the previous call
1960 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
1961 TPT.setOperand(Ext, 0, ExtOpnd);
1964 // Get through the Instruction:
1965 // 1. Update its type.
1966 // 2. Replace the uses of Ext by Inst.
1967 // 3. Extend each operand that needs to be extended.
1969 // Remember the original type of the instruction before promotion.
1970 // This is useful to know that the high bits are sign extended bits.
1971 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
1972 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
1974 TPT.mutateType(ExtOpnd, Ext->getType());
1976 TPT.replaceAllUsesWith(Ext, ExtOpnd);
1978 Instruction *ExtForOpnd = Ext;
1980 DEBUG(dbgs() << "Propagate Ext to operands\n");
1981 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
1983 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
1984 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
1985 !shouldExtOperand(ExtOpnd, OpIdx)) {
1986 DEBUG(dbgs() << "No need to propagate\n");
1989 // Check if we can statically extend the operand.
1990 Value *Opnd = ExtOpnd->getOperand(OpIdx);
1991 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
1992 DEBUG(dbgs() << "Statically extend\n");
1993 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
1994 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
1995 : Cst->getValue().zext(BitWidth);
1996 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
1999 // UndefValue are typed, so we have to statically sign extend them.
2000 if (isa<UndefValue>(Opnd)) {
2001 DEBUG(dbgs() << "Statically extend\n");
2002 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2006 // Otherwise we have to explicity sign extend the operand.
2007 // Check if Ext was reused to extend an operand.
2009 // If yes, create a new one.
2010 DEBUG(dbgs() << "More operands to ext\n");
2012 cast<Instruction>(IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2013 : TPT.createZExt(Ext, Opnd, Ext->getType()));
2017 TPT.setOperand(ExtForOpnd, 0, Opnd);
2019 // Move the sign extension before the insertion point.
2020 TPT.moveBefore(ExtForOpnd, ExtOpnd);
2021 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2022 // If more sext are required, new instructions will have to be created.
2023 ExtForOpnd = nullptr;
2025 if (ExtForOpnd == Ext) {
2026 DEBUG(dbgs() << "Extension is useless now\n");
2027 TPT.eraseInstruction(Ext);
2032 /// IsPromotionProfitable - Check whether or not promoting an instruction
2033 /// to a wider type was profitable.
2034 /// \p MatchedSize gives the number of instructions that have been matched
2035 /// in the addressing mode after the promotion was applied.
2036 /// \p SizeWithPromotion gives the number of created instructions for
2037 /// the promotion plus the number of instructions that have been
2038 /// matched in the addressing mode before the promotion.
2039 /// \p PromotedOperand is the value that has been promoted.
2040 /// \return True if the promotion is profitable, false otherwise.
2042 AddressingModeMatcher::IsPromotionProfitable(unsigned MatchedSize,
2043 unsigned SizeWithPromotion,
2044 Value *PromotedOperand) const {
2045 // We folded less instructions than what we created to promote the operand.
2046 // This is not profitable.
2047 if (MatchedSize < SizeWithPromotion)
2049 if (MatchedSize > SizeWithPromotion)
2051 // The promotion is neutral but it may help folding the sign extension in
2052 // loads for instance.
2053 // Check that we did not create an illegal instruction.
2054 Instruction *PromotedInst = dyn_cast<Instruction>(PromotedOperand);
2057 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2058 // If the ISDOpcode is undefined, it was undefined before the promotion.
2061 // Otherwise, check if the promoted instruction is legal or not.
2062 return TLI.isOperationLegalOrCustom(
2063 ISDOpcode, TLI.getValueType(PromotedInst->getType()));
2066 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
2067 /// fold the operation into the addressing mode. If so, update the addressing
2068 /// mode and return true, otherwise return false without modifying AddrMode.
2069 /// If \p MovedAway is not NULL, it contains the information of whether or
2070 /// not AddrInst has to be folded into the addressing mode on success.
2071 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2072 /// because it has been moved away.
2073 /// Thus AddrInst must not be added in the matched instructions.
2074 /// This state can happen when AddrInst is a sext, since it may be moved away.
2075 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2076 /// not be referenced anymore.
2077 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
2080 // Avoid exponential behavior on extremely deep expression trees.
2081 if (Depth >= 5) return false;
2083 // By default, all matched instructions stay in place.
2088 case Instruction::PtrToInt:
2089 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2090 return MatchAddr(AddrInst->getOperand(0), Depth);
2091 case Instruction::IntToPtr:
2092 // This inttoptr is a no-op if the integer type is pointer sized.
2093 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
2094 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
2095 return MatchAddr(AddrInst->getOperand(0), Depth);
2097 case Instruction::BitCast:
2098 case Instruction::AddrSpaceCast:
2099 // BitCast is always a noop, and we can handle it as long as it is
2100 // int->int or pointer->pointer (we don't want int<->fp or something).
2101 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2102 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2103 // Don't touch identity bitcasts. These were probably put here by LSR,
2104 // and we don't want to mess around with them. Assume it knows what it
2106 AddrInst->getOperand(0)->getType() != AddrInst->getType())
2107 return MatchAddr(AddrInst->getOperand(0), Depth);
2109 case Instruction::Add: {
2110 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2111 ExtAddrMode BackupAddrMode = AddrMode;
2112 unsigned OldSize = AddrModeInsts.size();
2113 // Start a transaction at this point.
2114 // The LHS may match but not the RHS.
2115 // Therefore, we need a higher level restoration point to undo partially
2116 // matched operation.
2117 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2118 TPT.getRestorationPoint();
2120 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2121 MatchAddr(AddrInst->getOperand(0), Depth+1))
2124 // Restore the old addr mode info.
2125 AddrMode = BackupAddrMode;
2126 AddrModeInsts.resize(OldSize);
2127 TPT.rollback(LastKnownGood);
2129 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2130 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2131 MatchAddr(AddrInst->getOperand(1), Depth+1))
2134 // Otherwise we definitely can't merge the ADD in.
2135 AddrMode = BackupAddrMode;
2136 AddrModeInsts.resize(OldSize);
2137 TPT.rollback(LastKnownGood);
2140 //case Instruction::Or:
2141 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2143 case Instruction::Mul:
2144 case Instruction::Shl: {
2145 // Can only handle X*C and X << C.
2146 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2149 int64_t Scale = RHS->getSExtValue();
2150 if (Opcode == Instruction::Shl)
2151 Scale = 1LL << Scale;
2153 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2155 case Instruction::GetElementPtr: {
2156 // Scan the GEP. We check it if it contains constant offsets and at most
2157 // one variable offset.
2158 int VariableOperand = -1;
2159 unsigned VariableScale = 0;
2161 int64_t ConstantOffset = 0;
2162 const DataLayout *TD = TLI.getDataLayout();
2163 gep_type_iterator GTI = gep_type_begin(AddrInst);
2164 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2165 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2166 const StructLayout *SL = TD->getStructLayout(STy);
2168 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2169 ConstantOffset += SL->getElementOffset(Idx);
2171 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
2172 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2173 ConstantOffset += CI->getSExtValue()*TypeSize;
2174 } else if (TypeSize) { // Scales of zero don't do anything.
2175 // We only allow one variable index at the moment.
2176 if (VariableOperand != -1)
2179 // Remember the variable index.
2180 VariableOperand = i;
2181 VariableScale = TypeSize;
2186 // A common case is for the GEP to only do a constant offset. In this case,
2187 // just add it to the disp field and check validity.
2188 if (VariableOperand == -1) {
2189 AddrMode.BaseOffs += ConstantOffset;
2190 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
2191 // Check to see if we can fold the base pointer in too.
2192 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2195 AddrMode.BaseOffs -= ConstantOffset;
2199 // Save the valid addressing mode in case we can't match.
2200 ExtAddrMode BackupAddrMode = AddrMode;
2201 unsigned OldSize = AddrModeInsts.size();
2203 // See if the scale and offset amount is valid for this target.
2204 AddrMode.BaseOffs += ConstantOffset;
2206 // Match the base operand of the GEP.
2207 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2208 // If it couldn't be matched, just stuff the value in a register.
2209 if (AddrMode.HasBaseReg) {
2210 AddrMode = BackupAddrMode;
2211 AddrModeInsts.resize(OldSize);
2214 AddrMode.HasBaseReg = true;
2215 AddrMode.BaseReg = AddrInst->getOperand(0);
2218 // Match the remaining variable portion of the GEP.
2219 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2221 // If it couldn't be matched, try stuffing the base into a register
2222 // instead of matching it, and retrying the match of the scale.
2223 AddrMode = BackupAddrMode;
2224 AddrModeInsts.resize(OldSize);
2225 if (AddrMode.HasBaseReg)
2227 AddrMode.HasBaseReg = true;
2228 AddrMode.BaseReg = AddrInst->getOperand(0);
2229 AddrMode.BaseOffs += ConstantOffset;
2230 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2231 VariableScale, Depth)) {
2232 // If even that didn't work, bail.
2233 AddrMode = BackupAddrMode;
2234 AddrModeInsts.resize(OldSize);
2241 case Instruction::SExt:
2242 case Instruction::ZExt: {
2243 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
2247 // Try to move this ext out of the way of the addressing mode.
2248 // Ask for a method for doing so.
2249 TypePromotionHelper::Action TPH =
2250 TypePromotionHelper::getAction(Ext, InsertedTruncs, TLI, PromotedInsts);
2254 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2255 TPT.getRestorationPoint();
2256 unsigned CreatedInsts = 0;
2257 Value *PromotedOperand = TPH(Ext, TPT, PromotedInsts, CreatedInsts);
2258 // SExt has been moved away.
2259 // Thus either it will be rematched later in the recursive calls or it is
2260 // gone. Anyway, we must not fold it into the addressing mode at this point.
2264 // addr = gep base, idx
2266 // promotedOpnd = ext opnd <- no match here
2267 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2268 // addr = gep base, op <- match
2272 assert(PromotedOperand &&
2273 "TypePromotionHelper should have filtered out those cases");
2275 ExtAddrMode BackupAddrMode = AddrMode;
2276 unsigned OldSize = AddrModeInsts.size();
2278 if (!MatchAddr(PromotedOperand, Depth) ||
2279 !IsPromotionProfitable(AddrModeInsts.size(), OldSize + CreatedInsts,
2281 AddrMode = BackupAddrMode;
2282 AddrModeInsts.resize(OldSize);
2283 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2284 TPT.rollback(LastKnownGood);
2293 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2294 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2295 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2296 /// or intptr_t for the target.
2298 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2299 // Start a transaction at this point that we will rollback if the matching
2301 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2302 TPT.getRestorationPoint();
2303 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2304 // Fold in immediates if legal for the target.
2305 AddrMode.BaseOffs += CI->getSExtValue();
2306 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2308 AddrMode.BaseOffs -= CI->getSExtValue();
2309 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2310 // If this is a global variable, try to fold it into the addressing mode.
2311 if (!AddrMode.BaseGV) {
2312 AddrMode.BaseGV = GV;
2313 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2315 AddrMode.BaseGV = nullptr;
2317 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2318 ExtAddrMode BackupAddrMode = AddrMode;
2319 unsigned OldSize = AddrModeInsts.size();
2321 // Check to see if it is possible to fold this operation.
2322 bool MovedAway = false;
2323 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2324 // This instruction may have been move away. If so, there is nothing
2328 // Okay, it's possible to fold this. Check to see if it is actually
2329 // *profitable* to do so. We use a simple cost model to avoid increasing
2330 // register pressure too much.
2331 if (I->hasOneUse() ||
2332 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2333 AddrModeInsts.push_back(I);
2337 // It isn't profitable to do this, roll back.
2338 //cerr << "NOT FOLDING: " << *I;
2339 AddrMode = BackupAddrMode;
2340 AddrModeInsts.resize(OldSize);
2341 TPT.rollback(LastKnownGood);
2343 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2344 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2346 TPT.rollback(LastKnownGood);
2347 } else if (isa<ConstantPointerNull>(Addr)) {
2348 // Null pointer gets folded without affecting the addressing mode.
2352 // Worse case, the target should support [reg] addressing modes. :)
2353 if (!AddrMode.HasBaseReg) {
2354 AddrMode.HasBaseReg = true;
2355 AddrMode.BaseReg = Addr;
2356 // Still check for legality in case the target supports [imm] but not [i+r].
2357 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2359 AddrMode.HasBaseReg = false;
2360 AddrMode.BaseReg = nullptr;
2363 // If the base register is already taken, see if we can do [r+r].
2364 if (AddrMode.Scale == 0) {
2366 AddrMode.ScaledReg = Addr;
2367 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2370 AddrMode.ScaledReg = nullptr;
2373 TPT.rollback(LastKnownGood);
2377 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2378 /// inline asm call are due to memory operands. If so, return true, otherwise
2380 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2381 const TargetLowering &TLI) {
2382 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(ImmutableCallSite(CI));
2383 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2384 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2386 // Compute the constraint code and ConstraintType to use.
2387 TLI.ComputeConstraintToUse(OpInfo, SDValue());
2389 // If this asm operand is our Value*, and if it isn't an indirect memory
2390 // operand, we can't fold it!
2391 if (OpInfo.CallOperandVal == OpVal &&
2392 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2393 !OpInfo.isIndirect))
2400 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2401 /// memory use. If we find an obviously non-foldable instruction, return true.
2402 /// Add the ultimately found memory instructions to MemoryUses.
2403 static bool FindAllMemoryUses(Instruction *I,
2404 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
2405 SmallPtrSetImpl<Instruction*> &ConsideredInsts,
2406 const TargetLowering &TLI) {
2407 // If we already considered this instruction, we're done.
2408 if (!ConsideredInsts.insert(I).second)
2411 // If this is an obviously unfoldable instruction, bail out.
2412 if (!MightBeFoldableInst(I))
2415 // Loop over all the uses, recursively processing them.
2416 for (Use &U : I->uses()) {
2417 Instruction *UserI = cast<Instruction>(U.getUser());
2419 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2420 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2424 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2425 unsigned opNo = U.getOperandNo();
2426 if (opNo == 0) return true; // Storing addr, not into addr.
2427 MemoryUses.push_back(std::make_pair(SI, opNo));
2431 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2432 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2433 if (!IA) return true;
2435 // If this is a memory operand, we're cool, otherwise bail out.
2436 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
2441 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI))
2448 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2449 /// the use site that we're folding it into. If so, there is no cost to
2450 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2451 /// that we know are live at the instruction already.
2452 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2453 Value *KnownLive2) {
2454 // If Val is either of the known-live values, we know it is live!
2455 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
2458 // All values other than instructions and arguments (e.g. constants) are live.
2459 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2461 // If Val is a constant sized alloca in the entry block, it is live, this is
2462 // true because it is just a reference to the stack/frame pointer, which is
2463 // live for the whole function.
2464 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2465 if (AI->isStaticAlloca())
2468 // Check to see if this value is already used in the memory instruction's
2469 // block. If so, it's already live into the block at the very least, so we
2470 // can reasonably fold it.
2471 return Val->isUsedInBasicBlock(MemoryInst->getParent());
2474 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
2475 /// mode of the machine to fold the specified instruction into a load or store
2476 /// that ultimately uses it. However, the specified instruction has multiple
2477 /// uses. Given this, it may actually increase register pressure to fold it
2478 /// into the load. For example, consider this code:
2482 /// use(Y) -> nonload/store
2486 /// In this case, Y has multiple uses, and can be folded into the load of Z
2487 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
2488 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
2489 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
2490 /// number of computations either.
2492 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
2493 /// X was live across 'load Z' for other reasons, we actually *would* want to
2494 /// fold the addressing mode in the Z case. This would make Y die earlier.
2495 bool AddressingModeMatcher::
2496 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
2497 ExtAddrMode &AMAfter) {
2498 if (IgnoreProfitability) return true;
2500 // AMBefore is the addressing mode before this instruction was folded into it,
2501 // and AMAfter is the addressing mode after the instruction was folded. Get
2502 // the set of registers referenced by AMAfter and subtract out those
2503 // referenced by AMBefore: this is the set of values which folding in this
2504 // address extends the lifetime of.
2506 // Note that there are only two potential values being referenced here,
2507 // BaseReg and ScaleReg (global addresses are always available, as are any
2508 // folded immediates).
2509 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
2511 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
2512 // lifetime wasn't extended by adding this instruction.
2513 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2515 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2516 ScaledReg = nullptr;
2518 // If folding this instruction (and it's subexprs) didn't extend any live
2519 // ranges, we're ok with it.
2520 if (!BaseReg && !ScaledReg)
2523 // If all uses of this instruction are ultimately load/store/inlineasm's,
2524 // check to see if their addressing modes will include this instruction. If
2525 // so, we can fold it into all uses, so it doesn't matter if it has multiple
2527 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
2528 SmallPtrSet<Instruction*, 16> ConsideredInsts;
2529 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
2530 return false; // Has a non-memory, non-foldable use!
2532 // Now that we know that all uses of this instruction are part of a chain of
2533 // computation involving only operations that could theoretically be folded
2534 // into a memory use, loop over each of these uses and see if they could
2535 // *actually* fold the instruction.
2536 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
2537 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
2538 Instruction *User = MemoryUses[i].first;
2539 unsigned OpNo = MemoryUses[i].second;
2541 // Get the access type of this use. If the use isn't a pointer, we don't
2542 // know what it accesses.
2543 Value *Address = User->getOperand(OpNo);
2544 if (!Address->getType()->isPointerTy())
2546 Type *AddressAccessTy = Address->getType()->getPointerElementType();
2548 // Do a match against the root of this address, ignoring profitability. This
2549 // will tell us if the addressing mode for the memory operation will
2550 // *actually* cover the shared instruction.
2552 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2553 TPT.getRestorationPoint();
2554 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
2555 MemoryInst, Result, InsertedTruncs,
2556 PromotedInsts, TPT);
2557 Matcher.IgnoreProfitability = true;
2558 bool Success = Matcher.MatchAddr(Address, 0);
2559 (void)Success; assert(Success && "Couldn't select *anything*?");
2561 // The match was to check the profitability, the changes made are not
2562 // part of the original matcher. Therefore, they should be dropped
2563 // otherwise the original matcher will not present the right state.
2564 TPT.rollback(LastKnownGood);
2566 // If the match didn't cover I, then it won't be shared by it.
2567 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
2568 I) == MatchedAddrModeInsts.end())
2571 MatchedAddrModeInsts.clear();
2577 } // end anonymous namespace
2579 /// IsNonLocalValue - Return true if the specified values are defined in a
2580 /// different basic block than BB.
2581 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
2582 if (Instruction *I = dyn_cast<Instruction>(V))
2583 return I->getParent() != BB;
2587 /// OptimizeMemoryInst - Load and Store Instructions often have
2588 /// addressing modes that can do significant amounts of computation. As such,
2589 /// instruction selection will try to get the load or store to do as much
2590 /// computation as possible for the program. The problem is that isel can only
2591 /// see within a single block. As such, we sink as much legal addressing mode
2592 /// stuff into the block as possible.
2594 /// This method is used to optimize both load/store and inline asms with memory
2596 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
2600 // Try to collapse single-value PHI nodes. This is necessary to undo
2601 // unprofitable PRE transformations.
2602 SmallVector<Value*, 8> worklist;
2603 SmallPtrSet<Value*, 16> Visited;
2604 worklist.push_back(Addr);
2606 // Use a worklist to iteratively look through PHI nodes, and ensure that
2607 // the addressing mode obtained from the non-PHI roots of the graph
2609 Value *Consensus = nullptr;
2610 unsigned NumUsesConsensus = 0;
2611 bool IsNumUsesConsensusValid = false;
2612 SmallVector<Instruction*, 16> AddrModeInsts;
2613 ExtAddrMode AddrMode;
2614 TypePromotionTransaction TPT;
2615 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2616 TPT.getRestorationPoint();
2617 while (!worklist.empty()) {
2618 Value *V = worklist.back();
2619 worklist.pop_back();
2621 // Break use-def graph loops.
2622 if (!Visited.insert(V).second) {
2623 Consensus = nullptr;
2627 // For a PHI node, push all of its incoming values.
2628 if (PHINode *P = dyn_cast<PHINode>(V)) {
2629 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
2630 worklist.push_back(P->getIncomingValue(i));
2634 // For non-PHIs, determine the addressing mode being computed.
2635 SmallVector<Instruction*, 16> NewAddrModeInsts;
2636 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
2637 V, AccessTy, MemoryInst, NewAddrModeInsts, *TLI, InsertedTruncsSet,
2638 PromotedInsts, TPT);
2640 // This check is broken into two cases with very similar code to avoid using
2641 // getNumUses() as much as possible. Some values have a lot of uses, so
2642 // calling getNumUses() unconditionally caused a significant compile-time
2646 AddrMode = NewAddrMode;
2647 AddrModeInsts = NewAddrModeInsts;
2649 } else if (NewAddrMode == AddrMode) {
2650 if (!IsNumUsesConsensusValid) {
2651 NumUsesConsensus = Consensus->getNumUses();
2652 IsNumUsesConsensusValid = true;
2655 // Ensure that the obtained addressing mode is equivalent to that obtained
2656 // for all other roots of the PHI traversal. Also, when choosing one
2657 // such root as representative, select the one with the most uses in order
2658 // to keep the cost modeling heuristics in AddressingModeMatcher
2660 unsigned NumUses = V->getNumUses();
2661 if (NumUses > NumUsesConsensus) {
2663 NumUsesConsensus = NumUses;
2664 AddrModeInsts = NewAddrModeInsts;
2669 Consensus = nullptr;
2673 // If the addressing mode couldn't be determined, or if multiple different
2674 // ones were determined, bail out now.
2676 TPT.rollback(LastKnownGood);
2681 // Check to see if any of the instructions supersumed by this addr mode are
2682 // non-local to I's BB.
2683 bool AnyNonLocal = false;
2684 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
2685 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
2691 // If all the instructions matched are already in this BB, don't do anything.
2693 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
2697 // Insert this computation right after this user. Since our caller is
2698 // scanning from the top of the BB to the bottom, reuse of the expr are
2699 // guaranteed to happen later.
2700 IRBuilder<> Builder(MemoryInst);
2702 // Now that we determined the addressing expression we want to use and know
2703 // that we have to sink it into this block. Check to see if we have already
2704 // done this for some other load/store instr in this block. If so, reuse the
2706 Value *&SunkAddr = SunkAddrs[Addr];
2708 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
2709 << *MemoryInst << "\n");
2710 if (SunkAddr->getType() != Addr->getType())
2711 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
2712 } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() &&
2713 TM && TM->getSubtarget<TargetSubtargetInfo>().useAA())) {
2714 // By default, we use the GEP-based method when AA is used later. This
2715 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
2716 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
2717 << *MemoryInst << "\n");
2718 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
2719 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
2721 // First, find the pointer.
2722 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
2723 ResultPtr = AddrMode.BaseReg;
2724 AddrMode.BaseReg = nullptr;
2727 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
2728 // We can't add more than one pointer together, nor can we scale a
2729 // pointer (both of which seem meaningless).
2730 if (ResultPtr || AddrMode.Scale != 1)
2733 ResultPtr = AddrMode.ScaledReg;
2737 if (AddrMode.BaseGV) {
2741 ResultPtr = AddrMode.BaseGV;
2744 // If the real base value actually came from an inttoptr, then the matcher
2745 // will look through it and provide only the integer value. In that case,
2747 if (!ResultPtr && AddrMode.BaseReg) {
2749 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
2750 AddrMode.BaseReg = nullptr;
2751 } else if (!ResultPtr && AddrMode.Scale == 1) {
2753 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
2758 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
2759 SunkAddr = Constant::getNullValue(Addr->getType());
2760 } else if (!ResultPtr) {
2764 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
2766 // Start with the base register. Do this first so that subsequent address
2767 // matching finds it last, which will prevent it from trying to match it
2768 // as the scaled value in case it happens to be a mul. That would be
2769 // problematic if we've sunk a different mul for the scale, because then
2770 // we'd end up sinking both muls.
2771 if (AddrMode.BaseReg) {
2772 Value *V = AddrMode.BaseReg;
2773 if (V->getType() != IntPtrTy)
2774 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
2779 // Add the scale value.
2780 if (AddrMode.Scale) {
2781 Value *V = AddrMode.ScaledReg;
2782 if (V->getType() == IntPtrTy) {
2784 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
2785 cast<IntegerType>(V->getType())->getBitWidth()) {
2786 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
2788 // It is only safe to sign extend the BaseReg if we know that the math
2789 // required to create it did not overflow before we extend it. Since
2790 // the original IR value was tossed in favor of a constant back when
2791 // the AddrMode was created we need to bail out gracefully if widths
2792 // do not match instead of extending it.
2793 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
2794 if (I && (ResultIndex != AddrMode.BaseReg))
2795 I->eraseFromParent();
2799 if (AddrMode.Scale != 1)
2800 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
2803 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
2808 // Add in the Base Offset if present.
2809 if (AddrMode.BaseOffs) {
2810 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
2812 // We need to add this separately from the scale above to help with
2813 // SDAG consecutive load/store merging.
2814 if (ResultPtr->getType() != I8PtrTy)
2815 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
2816 ResultPtr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
2823 SunkAddr = ResultPtr;
2825 if (ResultPtr->getType() != I8PtrTy)
2826 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
2827 SunkAddr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
2830 if (SunkAddr->getType() != Addr->getType())
2831 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
2834 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
2835 << *MemoryInst << "\n");
2836 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
2837 Value *Result = nullptr;
2839 // Start with the base register. Do this first so that subsequent address
2840 // matching finds it last, which will prevent it from trying to match it
2841 // as the scaled value in case it happens to be a mul. That would be
2842 // problematic if we've sunk a different mul for the scale, because then
2843 // we'd end up sinking both muls.
2844 if (AddrMode.BaseReg) {
2845 Value *V = AddrMode.BaseReg;
2846 if (V->getType()->isPointerTy())
2847 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2848 if (V->getType() != IntPtrTy)
2849 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
2853 // Add the scale value.
2854 if (AddrMode.Scale) {
2855 Value *V = AddrMode.ScaledReg;
2856 if (V->getType() == IntPtrTy) {
2858 } else if (V->getType()->isPointerTy()) {
2859 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
2860 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
2861 cast<IntegerType>(V->getType())->getBitWidth()) {
2862 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
2864 // It is only safe to sign extend the BaseReg if we know that the math
2865 // required to create it did not overflow before we extend it. Since
2866 // the original IR value was tossed in favor of a constant back when
2867 // the AddrMode was created we need to bail out gracefully if widths
2868 // do not match instead of extending it.
2869 Instruction *I = dyn_cast_or_null<Instruction>(Result);
2870 if (I && (Result != AddrMode.BaseReg))
2871 I->eraseFromParent();
2874 if (AddrMode.Scale != 1)
2875 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
2878 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2883 // Add in the BaseGV if present.
2884 if (AddrMode.BaseGV) {
2885 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
2887 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2892 // Add in the Base Offset if present.
2893 if (AddrMode.BaseOffs) {
2894 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
2896 Result = Builder.CreateAdd(Result, V, "sunkaddr");
2902 SunkAddr = Constant::getNullValue(Addr->getType());
2904 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
2907 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
2909 // If we have no uses, recursively delete the value and all dead instructions
2911 if (Repl->use_empty()) {
2912 // This can cause recursive deletion, which can invalidate our iterator.
2913 // Use a WeakVH to hold onto it in case this happens.
2914 WeakVH IterHandle(CurInstIterator);
2915 BasicBlock *BB = CurInstIterator->getParent();
2917 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
2919 if (IterHandle != CurInstIterator) {
2920 // If the iterator instruction was recursively deleted, start over at the
2921 // start of the block.
2922 CurInstIterator = BB->begin();
2930 /// OptimizeInlineAsmInst - If there are any memory operands, use
2931 /// OptimizeMemoryInst to sink their address computing into the block when
2932 /// possible / profitable.
2933 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
2934 bool MadeChange = false;
2936 TargetLowering::AsmOperandInfoVector
2937 TargetConstraints = TLI->ParseConstraints(CS);
2939 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2940 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2942 // Compute the constraint code and ConstraintType to use.
2943 TLI->ComputeConstraintToUse(OpInfo, SDValue());
2945 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
2946 OpInfo.isIndirect) {
2947 Value *OpVal = CS->getArgOperand(ArgNo++);
2948 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
2949 } else if (OpInfo.Type == InlineAsm::isInput)
2956 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
2957 /// basic block as the load, unless conditions are unfavorable. This allows
2958 /// SelectionDAG to fold the extend into the load.
2960 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
2961 // Look for a load being extended.
2962 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
2963 if (!LI) return false;
2965 // If they're already in the same block, there's nothing to do.
2966 if (LI->getParent() == I->getParent())
2969 EVT VT = TLI->getValueType(I->getType());
2970 EVT LoadVT = TLI->getValueType(LI->getType());
2972 // If the load has other users and the truncate is not free, this probably
2973 // isn't worthwhile.
2974 if (!LI->hasOneUse() && TLI &&
2975 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
2976 !TLI->isTruncateFree(I->getType(), LI->getType()))
2979 // Check whether the target supports casts folded into loads.
2981 if (isa<ZExtInst>(I))
2982 LType = ISD::ZEXTLOAD;
2984 assert(isa<SExtInst>(I) && "Unexpected ext type!");
2985 LType = ISD::SEXTLOAD;
2987 if (TLI && !TLI->isLoadExtLegal(LType, LoadVT))
2990 // Move the extend into the same block as the load, so that SelectionDAG
2992 I->removeFromParent();
2998 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
2999 BasicBlock *DefBB = I->getParent();
3001 // If the result of a {s|z}ext and its source are both live out, rewrite all
3002 // other uses of the source with result of extension.
3003 Value *Src = I->getOperand(0);
3004 if (Src->hasOneUse())
3007 // Only do this xform if truncating is free.
3008 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
3011 // Only safe to perform the optimization if the source is also defined in
3013 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
3016 bool DefIsLiveOut = false;
3017 for (User *U : I->users()) {
3018 Instruction *UI = cast<Instruction>(U);
3020 // Figure out which BB this ext is used in.
3021 BasicBlock *UserBB = UI->getParent();
3022 if (UserBB == DefBB) continue;
3023 DefIsLiveOut = true;
3029 // Make sure none of the uses are PHI nodes.
3030 for (User *U : Src->users()) {
3031 Instruction *UI = cast<Instruction>(U);
3032 BasicBlock *UserBB = UI->getParent();
3033 if (UserBB == DefBB) continue;
3034 // Be conservative. We don't want this xform to end up introducing
3035 // reloads just before load / store instructions.
3036 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
3040 // InsertedTruncs - Only insert one trunc in each block once.
3041 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
3043 bool MadeChange = false;
3044 for (Use &U : Src->uses()) {
3045 Instruction *User = cast<Instruction>(U.getUser());
3047 // Figure out which BB this ext is used in.
3048 BasicBlock *UserBB = User->getParent();
3049 if (UserBB == DefBB) continue;
3051 // Both src and def are live in this block. Rewrite the use.
3052 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
3054 if (!InsertedTrunc) {
3055 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3056 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
3057 InsertedTruncsSet.insert(InsertedTrunc);
3060 // Replace a use of the {s|z}ext source with a use of the result.
3069 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
3070 /// turned into an explicit branch.
3071 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
3072 // FIXME: This should use the same heuristics as IfConversion to determine
3073 // whether a select is better represented as a branch. This requires that
3074 // branch probability metadata is preserved for the select, which is not the
3077 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3079 // If the branch is predicted right, an out of order CPU can avoid blocking on
3080 // the compare. Emit cmovs on compares with a memory operand as branches to
3081 // avoid stalls on the load from memory. If the compare has more than one use
3082 // there's probably another cmov or setcc around so it's not worth emitting a
3087 Value *CmpOp0 = Cmp->getOperand(0);
3088 Value *CmpOp1 = Cmp->getOperand(1);
3090 // We check that the memory operand has one use to avoid uses of the loaded
3091 // value directly after the compare, making branches unprofitable.
3092 return Cmp->hasOneUse() &&
3093 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3094 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
3098 /// If we have a SelectInst that will likely profit from branch prediction,
3099 /// turn it into a branch.
3100 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
3101 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3103 // Can we convert the 'select' to CF ?
3104 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3107 TargetLowering::SelectSupportKind SelectKind;
3109 SelectKind = TargetLowering::VectorMaskSelect;
3110 else if (SI->getType()->isVectorTy())
3111 SelectKind = TargetLowering::ScalarCondVectorVal;
3113 SelectKind = TargetLowering::ScalarValSelect;
3115 // Do we have efficient codegen support for this kind of 'selects' ?
3116 if (TLI->isSelectSupported(SelectKind)) {
3117 // We have efficient codegen support for the select instruction.
3118 // Check if it is profitable to keep this 'select'.
3119 if (!TLI->isPredictableSelectExpensive() ||
3120 !isFormingBranchFromSelectProfitable(SI))
3126 // First, we split the block containing the select into 2 blocks.
3127 BasicBlock *StartBlock = SI->getParent();
3128 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3129 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3131 // Create a new block serving as the landing pad for the branch.
3132 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3133 NextBlock->getParent(), NextBlock);
3135 // Move the unconditional branch from the block with the select in it into our
3136 // landing pad block.
3137 StartBlock->getTerminator()->eraseFromParent();
3138 BranchInst::Create(NextBlock, SmallBlock);
3140 // Insert the real conditional branch based on the original condition.
3141 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3143 // The select itself is replaced with a PHI Node.
3144 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3146 PN->addIncoming(SI->getTrueValue(), StartBlock);
3147 PN->addIncoming(SI->getFalseValue(), SmallBlock);
3148 SI->replaceAllUsesWith(PN);
3149 SI->eraseFromParent();
3151 // Instruct OptimizeBlock to skip to the next block.
3152 CurInstIterator = StartBlock->end();
3153 ++NumSelectsExpanded;
3157 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3158 SmallVector<int, 16> Mask(SVI->getShuffleMask());
3160 for (unsigned i = 0; i < Mask.size(); ++i) {
3161 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3163 SplatElem = Mask[i];
3169 /// Some targets have expensive vector shifts if the lanes aren't all the same
3170 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3171 /// it's often worth sinking a shufflevector splat down to its use so that
3172 /// codegen can spot all lanes are identical.
3173 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3174 BasicBlock *DefBB = SVI->getParent();
3176 // Only do this xform if variable vector shifts are particularly expensive.
3177 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3180 // We only expect better codegen by sinking a shuffle if we can recognise a
3182 if (!isBroadcastShuffle(SVI))
3185 // InsertedShuffles - Only insert a shuffle in each block once.
3186 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3188 bool MadeChange = false;
3189 for (User *U : SVI->users()) {
3190 Instruction *UI = cast<Instruction>(U);
3192 // Figure out which BB this ext is used in.
3193 BasicBlock *UserBB = UI->getParent();
3194 if (UserBB == DefBB) continue;
3196 // For now only apply this when the splat is used by a shift instruction.
3197 if (!UI->isShift()) continue;
3199 // Everything checks out, sink the shuffle if the user's block doesn't
3200 // already have a copy.
3201 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3203 if (!InsertedShuffle) {
3204 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3205 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3207 SVI->getOperand(2), "", InsertPt);
3210 UI->replaceUsesOfWith(SVI, InsertedShuffle);
3214 // If we removed all uses, nuke the shuffle.
3215 if (SVI->use_empty()) {
3216 SVI->eraseFromParent();
3224 /// \brief Helper class to promote a scalar operation to a vector one.
3225 /// This class is used to move downward extractelement transition.
3227 /// a = vector_op <2 x i32>
3228 /// b = extractelement <2 x i32> a, i32 0
3233 /// a = vector_op <2 x i32>
3234 /// c = vector_op a (equivalent to scalar_op on the related lane)
3235 /// * d = extractelement <2 x i32> c, i32 0
3237 /// Assuming both extractelement and store can be combine, we get rid of the
3239 class VectorPromoteHelper {
3240 /// Used to perform some checks on the legality of vector operations.
3241 const TargetLowering &TLI;
3243 /// Used to estimated the cost of the promoted chain.
3244 const TargetTransformInfo &TTI;
3246 /// The transition being moved downwards.
3247 Instruction *Transition;
3248 /// The sequence of instructions to be promoted.
3249 SmallVector<Instruction *, 4> InstsToBePromoted;
3250 /// Cost of combining a store and an extract.
3251 unsigned StoreExtractCombineCost;
3252 /// Instruction that will be combined with the transition.
3253 Instruction *CombineInst;
3255 /// \brief The instruction that represents the current end of the transition.
3256 /// Since we are faking the promotion until we reach the end of the chain
3257 /// of computation, we need a way to get the current end of the transition.
3258 Instruction *getEndOfTransition() const {
3259 if (InstsToBePromoted.empty())
3261 return InstsToBePromoted.back();
3264 /// \brief Return the index of the original value in the transition.
3265 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
3266 /// c, is at index 0.
3267 unsigned getTransitionOriginalValueIdx() const {
3268 assert(isa<ExtractElementInst>(Transition) &&
3269 "Other kind of transitions are not supported yet");
3273 /// \brief Return the index of the index in the transition.
3274 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
3276 unsigned getTransitionIdx() const {
3277 assert(isa<ExtractElementInst>(Transition) &&
3278 "Other kind of transitions are not supported yet");
3282 /// \brief Get the type of the transition.
3283 /// This is the type of the original value.
3284 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
3285 /// transition is <2 x i32>.
3286 Type *getTransitionType() const {
3287 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
3290 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
3291 /// I.e., we have the following sequence:
3292 /// Def = Transition <ty1> a to <ty2>
3293 /// b = ToBePromoted <ty2> Def, ...
3295 /// b = ToBePromoted <ty1> a, ...
3296 /// Def = Transition <ty1> ToBePromoted to <ty2>
3297 void promoteImpl(Instruction *ToBePromoted);
3299 /// \brief Check whether or not it is profitable to promote all the
3300 /// instructions enqueued to be promoted.
3301 bool isProfitableToPromote() {
3302 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
3303 unsigned Index = isa<ConstantInt>(ValIdx)
3304 ? cast<ConstantInt>(ValIdx)->getZExtValue()
3306 Type *PromotedType = getTransitionType();
3308 StoreInst *ST = cast<StoreInst>(CombineInst);
3309 unsigned AS = ST->getPointerAddressSpace();
3310 unsigned Align = ST->getAlignment();
3311 // Check if this store is supported.
3312 if (!TLI.allowsMisalignedMemoryAccesses(
3313 TLI.getValueType(ST->getValueOperand()->getType()), AS, Align)) {
3314 // If this is not supported, there is no way we can combine
3315 // the extract with the store.
3319 // The scalar chain of computation has to pay for the transition
3320 // scalar to vector.
3321 // The vector chain has to account for the combining cost.
3322 uint64_t ScalarCost =
3323 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
3324 uint64_t VectorCost = StoreExtractCombineCost;
3325 for (const auto &Inst : InstsToBePromoted) {
3326 // Compute the cost.
3327 // By construction, all instructions being promoted are arithmetic ones.
3328 // Moreover, one argument is a constant that can be viewed as a splat
3330 Value *Arg0 = Inst->getOperand(0);
3331 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
3332 isa<ConstantFP>(Arg0);
3333 TargetTransformInfo::OperandValueKind Arg0OVK =
3334 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3335 : TargetTransformInfo::OK_AnyValue;
3336 TargetTransformInfo::OperandValueKind Arg1OVK =
3337 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3338 : TargetTransformInfo::OK_AnyValue;
3339 ScalarCost += TTI.getArithmeticInstrCost(
3340 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
3341 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
3344 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
3345 << ScalarCost << "\nVector: " << VectorCost << '\n');
3346 return ScalarCost > VectorCost;
3349 /// \brief Generate a constant vector with \p Val with the same
3350 /// number of elements as the transition.
3351 /// \p UseSplat defines whether or not \p Val should be replicated
3352 /// accross the whole vector.
3353 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
3354 /// otherwise we generate a vector with as many undef as possible:
3355 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
3356 /// used at the index of the extract.
3357 Value *getConstantVector(Constant *Val, bool UseSplat) const {
3358 unsigned ExtractIdx = UINT_MAX;
3360 // If we cannot determine where the constant must be, we have to
3361 // use a splat constant.
3362 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
3363 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
3364 ExtractIdx = CstVal->getSExtValue();
3369 unsigned End = getTransitionType()->getVectorNumElements();
3371 return ConstantVector::getSplat(End, Val);
3373 SmallVector<Constant *, 4> ConstVec;
3374 UndefValue *UndefVal = UndefValue::get(Val->getType());
3375 for (unsigned Idx = 0; Idx != End; ++Idx) {
3376 if (Idx == ExtractIdx)
3377 ConstVec.push_back(Val);
3379 ConstVec.push_back(UndefVal);
3381 return ConstantVector::get(ConstVec);
3384 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
3385 /// in \p Use can trigger undefined behavior.
3386 static bool canCauseUndefinedBehavior(const Instruction *Use,
3387 unsigned OperandIdx) {
3388 // This is not safe to introduce undef when the operand is on
3389 // the right hand side of a division-like instruction.
3390 if (OperandIdx != 1)
3392 switch (Use->getOpcode()) {
3395 case Instruction::SDiv:
3396 case Instruction::UDiv:
3397 case Instruction::SRem:
3398 case Instruction::URem:
3400 case Instruction::FDiv:
3401 case Instruction::FRem:
3402 return !Use->hasNoNaNs();
3404 llvm_unreachable(nullptr);
3408 VectorPromoteHelper(const TargetLowering &TLI, const TargetTransformInfo &TTI,
3409 Instruction *Transition, unsigned CombineCost)
3410 : TLI(TLI), TTI(TTI), Transition(Transition),
3411 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
3412 assert(Transition && "Do not know how to promote null");
3415 /// \brief Check if we can promote \p ToBePromoted to \p Type.
3416 bool canPromote(const Instruction *ToBePromoted) const {
3417 // We could support CastInst too.
3418 return isa<BinaryOperator>(ToBePromoted);
3421 /// \brief Check if it is profitable to promote \p ToBePromoted
3422 /// by moving downward the transition through.
3423 bool shouldPromote(const Instruction *ToBePromoted) const {
3424 // Promote only if all the operands can be statically expanded.
3425 // Indeed, we do not want to introduce any new kind of transitions.
3426 for (const Use &U : ToBePromoted->operands()) {
3427 const Value *Val = U.get();
3428 if (Val == getEndOfTransition()) {
3429 // If the use is a division and the transition is on the rhs,
3430 // we cannot promote the operation, otherwise we may create a
3431 // division by zero.
3432 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
3436 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
3437 !isa<ConstantFP>(Val))
3440 // Check that the resulting operation is legal.
3441 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
3444 return StressStoreExtract ||
3445 TLI.isOperationLegalOrCustom(
3446 ISDOpcode, TLI.getValueType(getTransitionType(), true));
3449 /// \brief Check whether or not \p Use can be combined
3450 /// with the transition.
3451 /// I.e., is it possible to do Use(Transition) => AnotherUse?
3452 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
3454 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
3455 void enqueueForPromotion(Instruction *ToBePromoted) {
3456 InstsToBePromoted.push_back(ToBePromoted);
3459 /// \brief Set the instruction that will be combined with the transition.
3460 void recordCombineInstruction(Instruction *ToBeCombined) {
3461 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
3462 CombineInst = ToBeCombined;
3465 /// \brief Promote all the instructions enqueued for promotion if it is
3467 /// \return True if the promotion happened, false otherwise.
3469 // Check if there is something to promote.
3470 // Right now, if we do not have anything to combine with,
3471 // we assume the promotion is not profitable.
3472 if (InstsToBePromoted.empty() || !CombineInst)
3476 if (!StressStoreExtract && !isProfitableToPromote())
3480 for (auto &ToBePromoted : InstsToBePromoted)
3481 promoteImpl(ToBePromoted);
3482 InstsToBePromoted.clear();
3486 } // End of anonymous namespace.
3488 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
3489 // At this point, we know that all the operands of ToBePromoted but Def
3490 // can be statically promoted.
3491 // For Def, we need to use its parameter in ToBePromoted:
3492 // b = ToBePromoted ty1 a
3493 // Def = Transition ty1 b to ty2
3494 // Move the transition down.
3495 // 1. Replace all uses of the promoted operation by the transition.
3496 // = ... b => = ... Def.
3497 assert(ToBePromoted->getType() == Transition->getType() &&
3498 "The type of the result of the transition does not match "
3500 ToBePromoted->replaceAllUsesWith(Transition);
3501 // 2. Update the type of the uses.
3502 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
3503 Type *TransitionTy = getTransitionType();
3504 ToBePromoted->mutateType(TransitionTy);
3505 // 3. Update all the operands of the promoted operation with promoted
3507 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
3508 for (Use &U : ToBePromoted->operands()) {
3509 Value *Val = U.get();
3510 Value *NewVal = nullptr;
3511 if (Val == Transition)
3512 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
3513 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
3514 isa<ConstantFP>(Val)) {
3515 // Use a splat constant if it is not safe to use undef.
3516 NewVal = getConstantVector(
3517 cast<Constant>(Val),
3518 isa<UndefValue>(Val) ||
3519 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
3521 assert(0 && "Did you modified shouldPromote and forgot to update this?");
3522 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
3524 Transition->removeFromParent();
3525 Transition->insertAfter(ToBePromoted);
3526 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
3529 /// Some targets can do store(extractelement) with one instruction.
3530 /// Try to push the extractelement towards the stores when the target
3531 /// has this feature and this is profitable.
3532 bool CodeGenPrepare::OptimizeExtractElementInst(Instruction *Inst) {
3533 unsigned CombineCost = UINT_MAX;
3534 if (DisableStoreExtract || !TLI ||
3535 (!StressStoreExtract &&
3536 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
3537 Inst->getOperand(1), CombineCost)))
3540 // At this point we know that Inst is a vector to scalar transition.
3541 // Try to move it down the def-use chain, until:
3542 // - We can combine the transition with its single use
3543 // => we got rid of the transition.
3544 // - We escape the current basic block
3545 // => we would need to check that we are moving it at a cheaper place and
3546 // we do not do that for now.
3547 BasicBlock *Parent = Inst->getParent();
3548 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
3549 VectorPromoteHelper VPH(*TLI, *TTI, Inst, CombineCost);
3550 // If the transition has more than one use, assume this is not going to be
3552 while (Inst->hasOneUse()) {
3553 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
3554 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
3556 if (ToBePromoted->getParent() != Parent) {
3557 DEBUG(dbgs() << "Instruction to promote is in a different block ("
3558 << ToBePromoted->getParent()->getName()
3559 << ") than the transition (" << Parent->getName() << ").\n");
3563 if (VPH.canCombine(ToBePromoted)) {
3564 DEBUG(dbgs() << "Assume " << *Inst << '\n'
3565 << "will be combined with: " << *ToBePromoted << '\n');
3566 VPH.recordCombineInstruction(ToBePromoted);
3567 bool Changed = VPH.promote();
3568 NumStoreExtractExposed += Changed;
3572 DEBUG(dbgs() << "Try promoting.\n");
3573 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
3576 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
3578 VPH.enqueueForPromotion(ToBePromoted);
3579 Inst = ToBePromoted;
3584 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
3585 if (PHINode *P = dyn_cast<PHINode>(I)) {
3586 // It is possible for very late stage optimizations (such as SimplifyCFG)
3587 // to introduce PHI nodes too late to be cleaned up. If we detect such a
3588 // trivial PHI, go ahead and zap it here.
3589 if (Value *V = SimplifyInstruction(P, TLI ? TLI->getDataLayout() : nullptr,
3591 P->replaceAllUsesWith(V);
3592 P->eraseFromParent();
3599 if (CastInst *CI = dyn_cast<CastInst>(I)) {
3600 // If the source of the cast is a constant, then this should have
3601 // already been constant folded. The only reason NOT to constant fold
3602 // it is if something (e.g. LSR) was careful to place the constant
3603 // evaluation in a block other than then one that uses it (e.g. to hoist
3604 // the address of globals out of a loop). If this is the case, we don't
3605 // want to forward-subst the cast.
3606 if (isa<Constant>(CI->getOperand(0)))
3609 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
3612 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
3613 /// Sink a zext or sext into its user blocks if the target type doesn't
3614 /// fit in one register
3615 if (TLI && TLI->getTypeAction(CI->getContext(),
3616 TLI->getValueType(CI->getType())) ==
3617 TargetLowering::TypeExpandInteger) {
3618 return SinkCast(CI);
3620 bool MadeChange = MoveExtToFormExtLoad(I);
3621 return MadeChange | OptimizeExtUses(I);
3627 if (CmpInst *CI = dyn_cast<CmpInst>(I))
3628 if (!TLI || !TLI->hasMultipleConditionRegisters())
3629 return OptimizeCmpExpression(CI);
3631 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3633 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
3637 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3639 return OptimizeMemoryInst(I, SI->getOperand(1),
3640 SI->getOperand(0)->getType());
3644 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
3646 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
3647 BinOp->getOpcode() == Instruction::LShr)) {
3648 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
3649 if (TLI && CI && TLI->hasExtractBitsInsn())
3650 return OptimizeExtractBits(BinOp, CI, *TLI);
3655 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
3656 if (GEPI->hasAllZeroIndices()) {
3657 /// The GEP operand must be a pointer, so must its result -> BitCast
3658 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
3659 GEPI->getName(), GEPI);
3660 GEPI->replaceAllUsesWith(NC);
3661 GEPI->eraseFromParent();
3669 if (CallInst *CI = dyn_cast<CallInst>(I))
3670 return OptimizeCallInst(CI);
3672 if (SelectInst *SI = dyn_cast<SelectInst>(I))
3673 return OptimizeSelectInst(SI);
3675 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
3676 return OptimizeShuffleVectorInst(SVI);
3678 if (isa<ExtractElementInst>(I))
3679 return OptimizeExtractElementInst(I);
3684 // In this pass we look for GEP and cast instructions that are used
3685 // across basic blocks and rewrite them to improve basic-block-at-a-time
3687 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
3689 bool MadeChange = false;
3691 CurInstIterator = BB.begin();
3692 while (CurInstIterator != BB.end())
3693 MadeChange |= OptimizeInst(CurInstIterator++);
3695 MadeChange |= DupRetToEnableTailCallOpts(&BB);
3700 // llvm.dbg.value is far away from the value then iSel may not be able
3701 // handle it properly. iSel will drop llvm.dbg.value if it can not
3702 // find a node corresponding to the value.
3703 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
3704 bool MadeChange = false;
3705 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
3706 Instruction *PrevNonDbgInst = nullptr;
3707 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
3708 Instruction *Insn = BI; ++BI;
3709 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
3710 // Leave dbg.values that refer to an alloca alone. These
3711 // instrinsics describe the address of a variable (= the alloca)
3712 // being taken. They should not be moved next to the alloca
3713 // (and to the beginning of the scope), but rather stay close to
3714 // where said address is used.
3715 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
3716 PrevNonDbgInst = Insn;
3720 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
3721 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
3722 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
3723 DVI->removeFromParent();
3724 if (isa<PHINode>(VI))
3725 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
3727 DVI->insertAfter(VI);
3736 // If there is a sequence that branches based on comparing a single bit
3737 // against zero that can be combined into a single instruction, and the
3738 // target supports folding these into a single instruction, sink the
3739 // mask and compare into the branch uses. Do this before OptimizeBlock ->
3740 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
3742 bool CodeGenPrepare::sinkAndCmp(Function &F) {
3743 if (!EnableAndCmpSinking)
3745 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
3747 bool MadeChange = false;
3748 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
3749 BasicBlock *BB = I++;
3751 // Does this BB end with the following?
3752 // %andVal = and %val, #single-bit-set
3753 // %icmpVal = icmp %andResult, 0
3754 // br i1 %cmpVal label %dest1, label %dest2"
3755 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
3756 if (!Brcc || !Brcc->isConditional())
3758 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
3759 if (!Cmp || Cmp->getParent() != BB)
3761 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
3762 if (!Zero || !Zero->isZero())
3764 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
3765 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
3767 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
3768 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
3770 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
3772 // Push the "and; icmp" for any users that are conditional branches.
3773 // Since there can only be one branch use per BB, we don't need to keep
3774 // track of which BBs we insert into.
3775 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
3779 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
3781 if (!BrccUser || !BrccUser->isConditional())
3783 BasicBlock *UserBB = BrccUser->getParent();
3784 if (UserBB == BB) continue;
3785 DEBUG(dbgs() << "found Brcc use\n");
3787 // Sink the "and; icmp" to use.
3789 BinaryOperator *NewAnd =
3790 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
3793 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
3797 DEBUG(BrccUser->getParent()->dump());
3803 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
3804 /// success, or returns false if no or invalid metadata was found.
3805 static bool extractBranchMetadata(BranchInst *BI,
3806 uint64_t &ProbTrue, uint64_t &ProbFalse) {
3807 assert(BI->isConditional() &&
3808 "Looking for probabilities on unconditional branch?");
3809 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
3810 if (!ProfileData || ProfileData->getNumOperands() != 3)
3813 const auto *CITrue =
3814 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
3815 const auto *CIFalse =
3816 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
3817 if (!CITrue || !CIFalse)
3820 ProbTrue = CITrue->getValue().getZExtValue();
3821 ProbFalse = CIFalse->getValue().getZExtValue();
3826 /// \brief Scale down both weights to fit into uint32_t.
3827 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
3828 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
3829 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
3830 NewTrue = NewTrue / Scale;
3831 NewFalse = NewFalse / Scale;
3834 /// \brief Some targets prefer to split a conditional branch like:
3836 /// %0 = icmp ne i32 %a, 0
3837 /// %1 = icmp ne i32 %b, 0
3838 /// %or.cond = or i1 %0, %1
3839 /// br i1 %or.cond, label %TrueBB, label %FalseBB
3841 /// into multiple branch instructions like:
3844 /// %0 = icmp ne i32 %a, 0
3845 /// br i1 %0, label %TrueBB, label %bb2
3847 /// %1 = icmp ne i32 %b, 0
3848 /// br i1 %1, label %TrueBB, label %FalseBB
3850 /// This usually allows instruction selection to do even further optimizations
3851 /// and combine the compare with the branch instruction. Currently this is
3852 /// applied for targets which have "cheap" jump instructions.
3854 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
3856 bool CodeGenPrepare::splitBranchCondition(Function &F) {
3857 if (!TM || TM->Options.EnableFastISel != true ||
3858 !TLI || TLI->isJumpExpensive())
3861 bool MadeChange = false;
3862 for (auto &BB : F) {
3863 // Does this BB end with the following?
3864 // %cond1 = icmp|fcmp|binary instruction ...
3865 // %cond2 = icmp|fcmp|binary instruction ...
3866 // %cond.or = or|and i1 %cond1, cond2
3867 // br i1 %cond.or label %dest1, label %dest2"
3868 BinaryOperator *LogicOp;
3869 BasicBlock *TBB, *FBB;
3870 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
3874 Value *Cond1, *Cond2;
3875 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
3876 m_OneUse(m_Value(Cond2)))))
3877 Opc = Instruction::And;
3878 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
3879 m_OneUse(m_Value(Cond2)))))
3880 Opc = Instruction::Or;
3884 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
3885 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
3888 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
3891 auto *InsertBefore = std::next(Function::iterator(BB))
3892 .getNodePtrUnchecked();
3893 auto TmpBB = BasicBlock::Create(BB.getContext(),
3894 BB.getName() + ".cond.split",
3895 BB.getParent(), InsertBefore);
3897 // Update original basic block by using the first condition directly by the
3898 // branch instruction and removing the no longer needed and/or instruction.
3899 auto *Br1 = cast<BranchInst>(BB.getTerminator());
3900 Br1->setCondition(Cond1);
3901 LogicOp->eraseFromParent();
3903 // Depending on the conditon we have to either replace the true or the false
3904 // successor of the original branch instruction.
3905 if (Opc == Instruction::And)
3906 Br1->setSuccessor(0, TmpBB);
3908 Br1->setSuccessor(1, TmpBB);
3910 // Fill in the new basic block.
3911 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
3912 if (auto *I = dyn_cast<Instruction>(Cond2)) {
3913 I->removeFromParent();
3914 I->insertBefore(Br2);
3917 // Update PHI nodes in both successors. The original BB needs to be
3918 // replaced in one succesor's PHI nodes, because the branch comes now from
3919 // the newly generated BB (NewBB). In the other successor we need to add one
3920 // incoming edge to the PHI nodes, because both branch instructions target
3921 // now the same successor. Depending on the original branch condition
3922 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
3923 // we perfrom the correct update for the PHI nodes.
3924 // This doesn't change the successor order of the just created branch
3925 // instruction (or any other instruction).
3926 if (Opc == Instruction::Or)
3927 std::swap(TBB, FBB);
3929 // Replace the old BB with the new BB.
3930 for (auto &I : *TBB) {
3931 PHINode *PN = dyn_cast<PHINode>(&I);
3935 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
3936 PN->setIncomingBlock(i, TmpBB);
3939 // Add another incoming edge form the new BB.
3940 for (auto &I : *FBB) {
3941 PHINode *PN = dyn_cast<PHINode>(&I);
3944 auto *Val = PN->getIncomingValueForBlock(&BB);
3945 PN->addIncoming(Val, TmpBB);
3948 // Update the branch weights (from SelectionDAGBuilder::
3949 // FindMergedConditions).
3950 if (Opc == Instruction::Or) {
3951 // Codegen X | Y as:
3960 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
3961 // The requirement is that
3962 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
3963 // = TrueProb for orignal BB.
3964 // Assuming the orignal weights are A and B, one choice is to set BB1's
3965 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
3967 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
3968 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
3969 // TmpBB, but the math is more complicated.
3970 uint64_t TrueWeight, FalseWeight;
3971 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
3972 uint64_t NewTrueWeight = TrueWeight;
3973 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
3974 scaleWeights(NewTrueWeight, NewFalseWeight);
3975 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
3976 .createBranchWeights(TrueWeight, FalseWeight));
3978 NewTrueWeight = TrueWeight;
3979 NewFalseWeight = 2 * FalseWeight;
3980 scaleWeights(NewTrueWeight, NewFalseWeight);
3981 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
3982 .createBranchWeights(TrueWeight, FalseWeight));
3985 // Codegen X & Y as:
3993 // This requires creation of TmpBB after CurBB.
3995 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
3996 // The requirement is that
3997 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
3998 // = FalseProb for orignal BB.
3999 // Assuming the orignal weights are A and B, one choice is to set BB1's
4000 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
4002 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
4003 uint64_t TrueWeight, FalseWeight;
4004 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4005 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
4006 uint64_t NewFalseWeight = FalseWeight;
4007 scaleWeights(NewTrueWeight, NewFalseWeight);
4008 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4009 .createBranchWeights(TrueWeight, FalseWeight));
4011 NewTrueWeight = 2 * TrueWeight;
4012 NewFalseWeight = FalseWeight;
4013 scaleWeights(NewTrueWeight, NewFalseWeight);
4014 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4015 .createBranchWeights(TrueWeight, FalseWeight));
4019 // Request DOM Tree update.
4020 // Note: No point in getting fancy here, since the DT info is never
4021 // available to CodeGenPrepare and the existing update code is broken
4027 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();