1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/IR/CallSite.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/GetElementPtrTypeIterator.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/InlineAsm.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/MDBuilder.h"
35 #include "llvm/IR/PatternMatch.h"
36 #include "llvm/IR/Statepoint.h"
37 #include "llvm/IR/ValueHandle.h"
38 #include "llvm/IR/ValueMap.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Target/TargetLowering.h"
44 #include "llvm/Target/TargetSubtargetInfo.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/BuildLibCalls.h"
47 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
48 #include "llvm/Transforms/Utils/Local.h"
49 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
51 using namespace llvm::PatternMatch;
53 #define DEBUG_TYPE "codegenprepare"
55 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
56 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
57 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
58 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
60 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
62 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
63 "computations were sunk");
64 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
65 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
66 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
67 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
68 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
69 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
70 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
72 static cl::opt<bool> DisableBranchOpts(
73 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
74 cl::desc("Disable branch optimizations in CodeGenPrepare"));
77 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
78 cl::desc("Disable GC optimizations in CodeGenPrepare"));
80 static cl::opt<bool> DisableSelectToBranch(
81 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
82 cl::desc("Disable select to branch conversion."));
84 static cl::opt<bool> AddrSinkUsingGEPs(
85 "addr-sink-using-gep", cl::Hidden, cl::init(false),
86 cl::desc("Address sinking in CGP using GEPs."));
88 static cl::opt<bool> EnableAndCmpSinking(
89 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
90 cl::desc("Enable sinkinig and/cmp into branches."));
92 static cl::opt<bool> DisableStoreExtract(
93 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
94 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
96 static cl::opt<bool> StressStoreExtract(
97 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
98 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
100 static cl::opt<bool> DisableExtLdPromotion(
101 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
102 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
105 static cl::opt<bool> StressExtLdPromotion(
106 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
107 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
108 "optimization in CodeGenPrepare"));
111 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
115 TypeIsSExt(Type *Ty, bool IsSExt) : Ty(Ty), IsSExt(IsSExt) {}
117 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
118 class TypePromotionTransaction;
120 class CodeGenPrepare : public FunctionPass {
121 /// TLI - Keep a pointer of a TargetLowering to consult for determining
122 /// transformation profitability.
123 const TargetMachine *TM;
124 const TargetLowering *TLI;
125 const TargetTransformInfo *TTI;
126 const TargetLibraryInfo *TLInfo;
129 /// CurInstIterator - As we scan instructions optimizing them, this is the
130 /// next instruction to optimize. Xforms that can invalidate this should
132 BasicBlock::iterator CurInstIterator;
134 /// Keeps track of non-local addresses that have been sunk into a block.
135 /// This allows us to avoid inserting duplicate code for blocks with
136 /// multiple load/stores of the same address.
137 ValueMap<Value*, Value*> SunkAddrs;
139 /// Keeps track of all truncates inserted for the current function.
140 SetOfInstrs InsertedTruncsSet;
141 /// Keeps track of the type of the related instruction before their
142 /// promotion for the current function.
143 InstrToOrigTy PromotedInsts;
145 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
149 /// OptSize - True if optimizing for size.
153 static char ID; // Pass identification, replacement for typeid
154 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
155 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr) {
156 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
158 bool runOnFunction(Function &F) override;
160 const char *getPassName() const override { return "CodeGen Prepare"; }
162 void getAnalysisUsage(AnalysisUsage &AU) const override {
163 AU.addPreserved<DominatorTreeWrapperPass>();
164 AU.addRequired<TargetLibraryInfoWrapperPass>();
165 AU.addRequired<TargetTransformInfoWrapperPass>();
169 bool EliminateFallThrough(Function &F);
170 bool EliminateMostlyEmptyBlocks(Function &F);
171 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
172 void EliminateMostlyEmptyBlock(BasicBlock *BB);
173 bool OptimizeBlock(BasicBlock &BB, bool& ModifiedDT);
174 bool OptimizeInst(Instruction *I, bool& ModifiedDT);
175 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
176 bool OptimizeInlineAsmInst(CallInst *CS);
177 bool OptimizeCallInst(CallInst *CI, bool& ModifiedDT);
178 bool MoveExtToFormExtLoad(Instruction *&I);
179 bool OptimizeExtUses(Instruction *I);
180 bool OptimizeSelectInst(SelectInst *SI);
181 bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
182 bool OptimizeExtractElementInst(Instruction *Inst);
183 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
184 bool PlaceDbgValues(Function &F);
185 bool sinkAndCmp(Function &F);
186 bool ExtLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
188 const SmallVectorImpl<Instruction *> &Exts,
189 unsigned CreatedInstCost);
190 bool splitBranchCondition(Function &F);
191 bool simplifyOffsetableRelocate(Instruction &I);
195 char CodeGenPrepare::ID = 0;
196 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
197 "Optimize for code generation", false, false)
199 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
200 return new CodeGenPrepare(TM);
203 bool CodeGenPrepare::runOnFunction(Function &F) {
204 if (skipOptnoneFunction(F))
207 bool EverMadeChange = false;
208 // Clear per function information.
209 InsertedTruncsSet.clear();
210 PromotedInsts.clear();
214 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
215 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
216 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
217 DominatorTreeWrapperPass *DTWP =
218 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
219 DT = DTWP ? &DTWP->getDomTree() : nullptr;
220 OptSize = F.hasFnAttribute(Attribute::OptimizeForSize);
222 /// This optimization identifies DIV instructions that can be
223 /// profitably bypassed and carried out with a shorter, faster divide.
224 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
225 const DenseMap<unsigned int, unsigned int> &BypassWidths =
226 TLI->getBypassSlowDivWidths();
227 for (Function::iterator I = F.begin(); I != F.end(); I++)
228 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
231 // Eliminate blocks that contain only PHI nodes and an
232 // unconditional branch.
233 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
235 // llvm.dbg.value is far away from the value then iSel may not be able
236 // handle it properly. iSel will drop llvm.dbg.value if it can not
237 // find a node corresponding to the value.
238 EverMadeChange |= PlaceDbgValues(F);
240 // If there is a mask, compare against zero, and branch that can be combined
241 // into a single target instruction, push the mask and compare into branch
242 // users. Do this before OptimizeBlock -> OptimizeInst ->
243 // OptimizeCmpExpression, which perturbs the pattern being searched for.
244 if (!DisableBranchOpts) {
245 EverMadeChange |= sinkAndCmp(F);
246 EverMadeChange |= splitBranchCondition(F);
249 bool MadeChange = true;
252 for (Function::iterator I = F.begin(); I != F.end(); ) {
253 BasicBlock *BB = I++;
254 bool ModifiedDTOnIteration = false;
255 MadeChange |= OptimizeBlock(*BB, ModifiedDTOnIteration);
257 // Restart BB iteration if the dominator tree of the Function was changed
258 ModifiedDT |= ModifiedDTOnIteration;
259 if (ModifiedDTOnIteration)
262 EverMadeChange |= MadeChange;
267 if (!DisableBranchOpts) {
269 SmallPtrSet<BasicBlock*, 8> WorkList;
270 for (BasicBlock &BB : F) {
271 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
272 MadeChange |= ConstantFoldTerminator(&BB, true);
273 if (!MadeChange) continue;
275 for (SmallVectorImpl<BasicBlock*>::iterator
276 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
277 if (pred_begin(*II) == pred_end(*II))
278 WorkList.insert(*II);
281 // Delete the dead blocks and any of their dead successors.
282 MadeChange |= !WorkList.empty();
283 while (!WorkList.empty()) {
284 BasicBlock *BB = *WorkList.begin();
286 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
290 for (SmallVectorImpl<BasicBlock*>::iterator
291 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
292 if (pred_begin(*II) == pred_end(*II))
293 WorkList.insert(*II);
296 // Merge pairs of basic blocks with unconditional branches, connected by
298 if (EverMadeChange || MadeChange)
299 MadeChange |= EliminateFallThrough(F);
303 EverMadeChange |= MadeChange;
306 if (!DisableGCOpts) {
307 SmallVector<Instruction *, 2> Statepoints;
308 for (BasicBlock &BB : F)
309 for (Instruction &I : BB)
311 Statepoints.push_back(&I);
312 for (auto &I : Statepoints)
313 EverMadeChange |= simplifyOffsetableRelocate(*I);
316 if (ModifiedDT && DT)
319 return EverMadeChange;
322 /// EliminateFallThrough - Merge basic blocks which are connected
323 /// by a single edge, where one of the basic blocks has a single successor
324 /// pointing to the other basic block, which has a single predecessor.
325 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
326 bool Changed = false;
327 // Scan all of the blocks in the function, except for the entry block.
328 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
329 BasicBlock *BB = I++;
330 // If the destination block has a single pred, then this is a trivial
331 // edge, just collapse it.
332 BasicBlock *SinglePred = BB->getSinglePredecessor();
334 // Don't merge if BB's address is taken.
335 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
337 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
338 if (Term && !Term->isConditional()) {
340 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
341 // Remember if SinglePred was the entry block of the function.
342 // If so, we will need to move BB back to the entry position.
343 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
344 MergeBasicBlockIntoOnlyPred(BB, DT);
346 if (isEntry && BB != &BB->getParent()->getEntryBlock())
347 BB->moveBefore(&BB->getParent()->getEntryBlock());
349 // We have erased a block. Update the iterator.
356 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
357 /// debug info directives, and an unconditional branch. Passes before isel
358 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
359 /// isel. Start by eliminating these blocks so we can split them the way we
361 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
362 bool MadeChange = false;
363 // Note that this intentionally skips the entry block.
364 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
365 BasicBlock *BB = I++;
367 // If this block doesn't end with an uncond branch, ignore it.
368 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
369 if (!BI || !BI->isUnconditional())
372 // If the instruction before the branch (skipping debug info) isn't a phi
373 // node, then other stuff is happening here.
374 BasicBlock::iterator BBI = BI;
375 if (BBI != BB->begin()) {
377 while (isa<DbgInfoIntrinsic>(BBI)) {
378 if (BBI == BB->begin())
382 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
386 // Do not break infinite loops.
387 BasicBlock *DestBB = BI->getSuccessor(0);
391 if (!CanMergeBlocks(BB, DestBB))
394 EliminateMostlyEmptyBlock(BB);
400 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
401 /// single uncond branch between them, and BB contains no other non-phi
403 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
404 const BasicBlock *DestBB) const {
405 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
406 // the successor. If there are more complex condition (e.g. preheaders),
407 // don't mess around with them.
408 BasicBlock::const_iterator BBI = BB->begin();
409 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
410 for (const User *U : PN->users()) {
411 const Instruction *UI = cast<Instruction>(U);
412 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
414 // If User is inside DestBB block and it is a PHINode then check
415 // incoming value. If incoming value is not from BB then this is
416 // a complex condition (e.g. preheaders) we want to avoid here.
417 if (UI->getParent() == DestBB) {
418 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
419 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
420 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
421 if (Insn && Insn->getParent() == BB &&
422 Insn->getParent() != UPN->getIncomingBlock(I))
429 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
430 // and DestBB may have conflicting incoming values for the block. If so, we
431 // can't merge the block.
432 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
433 if (!DestBBPN) return true; // no conflict.
435 // Collect the preds of BB.
436 SmallPtrSet<const BasicBlock*, 16> BBPreds;
437 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
438 // It is faster to get preds from a PHI than with pred_iterator.
439 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
440 BBPreds.insert(BBPN->getIncomingBlock(i));
442 BBPreds.insert(pred_begin(BB), pred_end(BB));
445 // Walk the preds of DestBB.
446 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
447 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
448 if (BBPreds.count(Pred)) { // Common predecessor?
449 BBI = DestBB->begin();
450 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
451 const Value *V1 = PN->getIncomingValueForBlock(Pred);
452 const Value *V2 = PN->getIncomingValueForBlock(BB);
454 // If V2 is a phi node in BB, look up what the mapped value will be.
455 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
456 if (V2PN->getParent() == BB)
457 V2 = V2PN->getIncomingValueForBlock(Pred);
459 // If there is a conflict, bail out.
460 if (V1 != V2) return false;
469 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
470 /// an unconditional branch in it.
471 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
472 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
473 BasicBlock *DestBB = BI->getSuccessor(0);
475 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
477 // If the destination block has a single pred, then this is a trivial edge,
479 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
480 if (SinglePred != DestBB) {
481 // Remember if SinglePred was the entry block of the function. If so, we
482 // will need to move BB back to the entry position.
483 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
484 MergeBasicBlockIntoOnlyPred(DestBB, DT);
486 if (isEntry && BB != &BB->getParent()->getEntryBlock())
487 BB->moveBefore(&BB->getParent()->getEntryBlock());
489 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
494 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
495 // to handle the new incoming edges it is about to have.
497 for (BasicBlock::iterator BBI = DestBB->begin();
498 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
499 // Remove the incoming value for BB, and remember it.
500 Value *InVal = PN->removeIncomingValue(BB, false);
502 // Two options: either the InVal is a phi node defined in BB or it is some
503 // value that dominates BB.
504 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
505 if (InValPhi && InValPhi->getParent() == BB) {
506 // Add all of the input values of the input PHI as inputs of this phi.
507 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
508 PN->addIncoming(InValPhi->getIncomingValue(i),
509 InValPhi->getIncomingBlock(i));
511 // Otherwise, add one instance of the dominating value for each edge that
512 // we will be adding.
513 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
514 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
515 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
517 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
518 PN->addIncoming(InVal, *PI);
523 // The PHIs are now updated, change everything that refers to BB to use
524 // DestBB and remove BB.
525 BB->replaceAllUsesWith(DestBB);
526 if (DT && !ModifiedDT) {
527 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
528 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
529 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
530 DT->changeImmediateDominator(DestBB, NewIDom);
533 BB->eraseFromParent();
536 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
539 // Computes a map of base pointer relocation instructions to corresponding
540 // derived pointer relocation instructions given a vector of all relocate calls
541 static void computeBaseDerivedRelocateMap(
542 const SmallVectorImpl<User *> &AllRelocateCalls,
543 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
545 // Collect information in two maps: one primarily for locating the base object
546 // while filling the second map; the second map is the final structure holding
547 // a mapping between Base and corresponding Derived relocate calls
548 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
549 for (auto &U : AllRelocateCalls) {
550 GCRelocateOperands ThisRelocate(U);
551 IntrinsicInst *I = cast<IntrinsicInst>(U);
552 auto K = std::make_pair(ThisRelocate.basePtrIndex(),
553 ThisRelocate.derivedPtrIndex());
554 RelocateIdxMap.insert(std::make_pair(K, I));
556 for (auto &Item : RelocateIdxMap) {
557 std::pair<unsigned, unsigned> Key = Item.first;
558 if (Key.first == Key.second)
559 // Base relocation: nothing to insert
562 IntrinsicInst *I = Item.second;
563 auto BaseKey = std::make_pair(Key.first, Key.first);
565 // We're iterating over RelocateIdxMap so we cannot modify it.
566 auto MaybeBase = RelocateIdxMap.find(BaseKey);
567 if (MaybeBase == RelocateIdxMap.end())
568 // TODO: We might want to insert a new base object relocate and gep off
569 // that, if there are enough derived object relocates.
572 RelocateInstMap[MaybeBase->second].push_back(I);
576 // Accepts a GEP and extracts the operands into a vector provided they're all
577 // small integer constants
578 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
579 SmallVectorImpl<Value *> &OffsetV) {
580 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
581 // Only accept small constant integer operands
582 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
583 if (!Op || Op->getZExtValue() > 20)
587 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
588 OffsetV.push_back(GEP->getOperand(i));
592 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
593 // replace, computes a replacement, and affects it.
595 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
596 const SmallVectorImpl<IntrinsicInst *> &Targets) {
597 bool MadeChange = false;
598 for (auto &ToReplace : Targets) {
599 GCRelocateOperands MasterRelocate(RelocatedBase);
600 GCRelocateOperands ThisRelocate(ToReplace);
602 assert(ThisRelocate.basePtrIndex() == MasterRelocate.basePtrIndex() &&
603 "Not relocating a derived object of the original base object");
604 if (ThisRelocate.basePtrIndex() == ThisRelocate.derivedPtrIndex()) {
605 // A duplicate relocate call. TODO: coalesce duplicates.
609 Value *Base = ThisRelocate.basePtr();
610 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.derivedPtr());
611 if (!Derived || Derived->getPointerOperand() != Base)
614 SmallVector<Value *, 2> OffsetV;
615 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
618 // Create a Builder and replace the target callsite with a gep
619 IRBuilder<> Builder(ToReplace);
620 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
622 Builder.CreateGEP(RelocatedBase, makeArrayRef(OffsetV));
623 Instruction *ReplacementInst = cast<Instruction>(Replacement);
624 ReplacementInst->removeFromParent();
625 ReplacementInst->insertAfter(RelocatedBase);
626 Replacement->takeName(ToReplace);
627 ToReplace->replaceAllUsesWith(Replacement);
628 ToReplace->eraseFromParent();
638 // %ptr = gep %base + 15
639 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
640 // %base' = relocate(%tok, i32 4, i32 4)
641 // %ptr' = relocate(%tok, i32 4, i32 5)
647 // %ptr = gep %base + 15
648 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
649 // %base' = gc.relocate(%tok, i32 4, i32 4)
650 // %ptr' = gep %base' + 15
652 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
653 bool MadeChange = false;
654 SmallVector<User *, 2> AllRelocateCalls;
656 for (auto *U : I.users())
657 if (isGCRelocate(dyn_cast<Instruction>(U)))
658 // Collect all the relocate calls associated with a statepoint
659 AllRelocateCalls.push_back(U);
661 // We need atleast one base pointer relocation + one derived pointer
662 // relocation to mangle
663 if (AllRelocateCalls.size() < 2)
666 // RelocateInstMap is a mapping from the base relocate instruction to the
667 // corresponding derived relocate instructions
668 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
669 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
670 if (RelocateInstMap.empty())
673 for (auto &Item : RelocateInstMap)
674 // Item.first is the RelocatedBase to offset against
675 // Item.second is the vector of Targets to replace
676 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
680 /// SinkCast - Sink the specified cast instruction into its user blocks
681 static bool SinkCast(CastInst *CI) {
682 BasicBlock *DefBB = CI->getParent();
684 /// InsertedCasts - Only insert a cast in each block once.
685 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
687 bool MadeChange = false;
688 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
690 Use &TheUse = UI.getUse();
691 Instruction *User = cast<Instruction>(*UI);
693 // Figure out which BB this cast is used in. For PHI's this is the
694 // appropriate predecessor block.
695 BasicBlock *UserBB = User->getParent();
696 if (PHINode *PN = dyn_cast<PHINode>(User)) {
697 UserBB = PN->getIncomingBlock(TheUse);
700 // Preincrement use iterator so we don't invalidate it.
703 // If this user is in the same block as the cast, don't change the cast.
704 if (UserBB == DefBB) continue;
706 // If we have already inserted a cast into this block, use it.
707 CastInst *&InsertedCast = InsertedCasts[UserBB];
710 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
712 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
717 // Replace a use of the cast with a use of the new cast.
718 TheUse = InsertedCast;
722 // If we removed all uses, nuke the cast.
723 if (CI->use_empty()) {
724 CI->eraseFromParent();
731 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
732 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
733 /// sink it into user blocks to reduce the number of virtual
734 /// registers that must be created and coalesced.
736 /// Return true if any changes are made.
738 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
739 // If this is a noop copy,
740 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
741 EVT DstVT = TLI.getValueType(CI->getType());
743 // This is an fp<->int conversion?
744 if (SrcVT.isInteger() != DstVT.isInteger())
747 // If this is an extension, it will be a zero or sign extension, which
749 if (SrcVT.bitsLT(DstVT)) return false;
751 // If these values will be promoted, find out what they will be promoted
752 // to. This helps us consider truncates on PPC as noop copies when they
754 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
755 TargetLowering::TypePromoteInteger)
756 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
757 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
758 TargetLowering::TypePromoteInteger)
759 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
761 // If, after promotion, these are the same types, this is a noop copy.
768 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
769 /// the number of virtual registers that must be created and coalesced. This is
770 /// a clear win except on targets with multiple condition code registers
771 /// (PowerPC), where it might lose; some adjustment may be wanted there.
773 /// Return true if any changes are made.
774 static bool OptimizeCmpExpression(CmpInst *CI) {
775 BasicBlock *DefBB = CI->getParent();
777 /// InsertedCmp - Only insert a cmp in each block once.
778 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
780 bool MadeChange = false;
781 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
783 Use &TheUse = UI.getUse();
784 Instruction *User = cast<Instruction>(*UI);
786 // Preincrement use iterator so we don't invalidate it.
789 // Don't bother for PHI nodes.
790 if (isa<PHINode>(User))
793 // Figure out which BB this cmp is used in.
794 BasicBlock *UserBB = User->getParent();
796 // If this user is in the same block as the cmp, don't change the cmp.
797 if (UserBB == DefBB) continue;
799 // If we have already inserted a cmp into this block, use it.
800 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
803 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
805 CmpInst::Create(CI->getOpcode(),
806 CI->getPredicate(), CI->getOperand(0),
807 CI->getOperand(1), "", InsertPt);
811 // Replace a use of the cmp with a use of the new cmp.
812 TheUse = InsertedCmp;
816 // If we removed all uses, nuke the cmp.
818 CI->eraseFromParent();
823 /// isExtractBitsCandidateUse - Check if the candidates could
824 /// be combined with shift instruction, which includes:
825 /// 1. Truncate instruction
826 /// 2. And instruction and the imm is a mask of the low bits:
827 /// imm & (imm+1) == 0
828 static bool isExtractBitsCandidateUse(Instruction *User) {
829 if (!isa<TruncInst>(User)) {
830 if (User->getOpcode() != Instruction::And ||
831 !isa<ConstantInt>(User->getOperand(1)))
834 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
836 if ((Cimm & (Cimm + 1)).getBoolValue())
842 /// SinkShiftAndTruncate - sink both shift and truncate instruction
843 /// to the use of truncate's BB.
845 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
846 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
847 const TargetLowering &TLI) {
848 BasicBlock *UserBB = User->getParent();
849 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
850 TruncInst *TruncI = dyn_cast<TruncInst>(User);
851 bool MadeChange = false;
853 for (Value::user_iterator TruncUI = TruncI->user_begin(),
854 TruncE = TruncI->user_end();
855 TruncUI != TruncE;) {
857 Use &TruncTheUse = TruncUI.getUse();
858 Instruction *TruncUser = cast<Instruction>(*TruncUI);
859 // Preincrement use iterator so we don't invalidate it.
863 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
867 // If the use is actually a legal node, there will not be an
868 // implicit truncate.
869 // FIXME: always querying the result type is just an
870 // approximation; some nodes' legality is determined by the
871 // operand or other means. There's no good way to find out though.
872 if (TLI.isOperationLegalOrCustom(
873 ISDOpcode, TLI.getValueType(TruncUser->getType(), true)))
876 // Don't bother for PHI nodes.
877 if (isa<PHINode>(TruncUser))
880 BasicBlock *TruncUserBB = TruncUser->getParent();
882 if (UserBB == TruncUserBB)
885 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
886 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
888 if (!InsertedShift && !InsertedTrunc) {
889 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
891 if (ShiftI->getOpcode() == Instruction::AShr)
893 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
896 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
899 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
902 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
903 TruncI->getType(), "", TruncInsertPt);
907 TruncTheUse = InsertedTrunc;
913 /// OptimizeExtractBits - sink the shift *right* instruction into user blocks if
914 /// the uses could potentially be combined with this shift instruction and
915 /// generate BitExtract instruction. It will only be applied if the architecture
916 /// supports BitExtract instruction. Here is an example:
918 /// %x.extract.shift = lshr i64 %arg1, 32
920 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
924 /// %x.extract.shift.1 = lshr i64 %arg1, 32
925 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
927 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
929 /// Return true if any changes are made.
930 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
931 const TargetLowering &TLI) {
932 BasicBlock *DefBB = ShiftI->getParent();
934 /// Only insert instructions in each block once.
935 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
937 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(ShiftI->getType()));
939 bool MadeChange = false;
940 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
942 Use &TheUse = UI.getUse();
943 Instruction *User = cast<Instruction>(*UI);
944 // Preincrement use iterator so we don't invalidate it.
947 // Don't bother for PHI nodes.
948 if (isa<PHINode>(User))
951 if (!isExtractBitsCandidateUse(User))
954 BasicBlock *UserBB = User->getParent();
956 if (UserBB == DefBB) {
957 // If the shift and truncate instruction are in the same BB. The use of
958 // the truncate(TruncUse) may still introduce another truncate if not
959 // legal. In this case, we would like to sink both shift and truncate
960 // instruction to the BB of TruncUse.
963 // i64 shift.result = lshr i64 opnd, imm
964 // trunc.result = trunc shift.result to i16
967 // ----> We will have an implicit truncate here if the architecture does
968 // not have i16 compare.
969 // cmp i16 trunc.result, opnd2
971 if (isa<TruncInst>(User) && shiftIsLegal
972 // If the type of the truncate is legal, no trucate will be
973 // introduced in other basic blocks.
974 && (!TLI.isTypeLegal(TLI.getValueType(User->getType()))))
976 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI);
980 // If we have already inserted a shift into this block, use it.
981 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
983 if (!InsertedShift) {
984 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
986 if (ShiftI->getOpcode() == Instruction::AShr)
988 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
991 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
996 // Replace a use of the shift with a use of the new shift.
997 TheUse = InsertedShift;
1000 // If we removed all uses, nuke the shift.
1001 if (ShiftI->use_empty())
1002 ShiftI->eraseFromParent();
1007 // ScalarizeMaskedLoad() translates masked load intrinsic, like
1008 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1009 // <16 x i1> %mask, <16 x i32> %passthru)
1010 // to a chain of basic blocks, whith loading element one-by-one if
1011 // the appropriate mask bit is set
1013 // %1 = bitcast i8* %addr to i32*
1014 // %2 = extractelement <16 x i1> %mask, i32 0
1015 // %3 = icmp eq i1 %2, true
1016 // br i1 %3, label %cond.load, label %else
1018 //cond.load: ; preds = %0
1019 // %4 = getelementptr i32* %1, i32 0
1020 // %5 = load i32* %4
1021 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1024 //else: ; preds = %0, %cond.load
1025 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1026 // %7 = extractelement <16 x i1> %mask, i32 1
1027 // %8 = icmp eq i1 %7, true
1028 // br i1 %8, label %cond.load1, label %else2
1030 //cond.load1: ; preds = %else
1031 // %9 = getelementptr i32* %1, i32 1
1032 // %10 = load i32* %9
1033 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1036 //else2: ; preds = %else, %cond.load1
1037 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1038 // %12 = extractelement <16 x i1> %mask, i32 2
1039 // %13 = icmp eq i1 %12, true
1040 // br i1 %13, label %cond.load4, label %else5
1042 static void ScalarizeMaskedLoad(CallInst *CI) {
1043 Value *Ptr = CI->getArgOperand(0);
1044 Value *Src0 = CI->getArgOperand(3);
1045 Value *Mask = CI->getArgOperand(2);
1046 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1047 Type *EltTy = VecType->getElementType();
1049 assert(VecType && "Unexpected return type of masked load intrinsic");
1051 IRBuilder<> Builder(CI->getContext());
1052 Instruction *InsertPt = CI;
1053 BasicBlock *IfBlock = CI->getParent();
1054 BasicBlock *CondBlock = nullptr;
1055 BasicBlock *PrevIfBlock = CI->getParent();
1056 Builder.SetInsertPoint(InsertPt);
1058 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1060 // Bitcast %addr fron i8* to EltTy*
1062 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1063 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1064 Value *UndefVal = UndefValue::get(VecType);
1066 // The result vector
1067 Value *VResult = UndefVal;
1069 PHINode *Phi = nullptr;
1070 Value *PrevPhi = UndefVal;
1072 unsigned VectorWidth = VecType->getNumElements();
1073 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1075 // Fill the "else" block, created in the previous iteration
1077 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1078 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1079 // %to_load = icmp eq i1 %mask_1, true
1080 // br i1 %to_load, label %cond.load, label %else
1083 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1084 Phi->addIncoming(VResult, CondBlock);
1085 Phi->addIncoming(PrevPhi, PrevIfBlock);
1090 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1091 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1092 ConstantInt::get(Predicate->getType(), 1));
1094 // Create "cond" block
1096 // %EltAddr = getelementptr i32* %1, i32 0
1097 // %Elt = load i32* %EltAddr
1098 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1100 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1101 Builder.SetInsertPoint(InsertPt);
1103 Value* Gep = Builder.CreateInBoundsGEP(FirstEltPtr, Builder.getInt32(Idx));
1104 LoadInst* Load = Builder.CreateLoad(Gep, false);
1105 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1107 // Create "else" block, fill it in the next iteration
1108 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1109 Builder.SetInsertPoint(InsertPt);
1110 Instruction *OldBr = IfBlock->getTerminator();
1111 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1112 OldBr->eraseFromParent();
1113 PrevIfBlock = IfBlock;
1114 IfBlock = NewIfBlock;
1117 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1118 Phi->addIncoming(VResult, CondBlock);
1119 Phi->addIncoming(PrevPhi, PrevIfBlock);
1120 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1121 CI->replaceAllUsesWith(NewI);
1122 CI->eraseFromParent();
1125 // ScalarizeMaskedStore() translates masked store intrinsic, like
1126 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1128 // to a chain of basic blocks, that stores element one-by-one if
1129 // the appropriate mask bit is set
1131 // %1 = bitcast i8* %addr to i32*
1132 // %2 = extractelement <16 x i1> %mask, i32 0
1133 // %3 = icmp eq i1 %2, true
1134 // br i1 %3, label %cond.store, label %else
1136 // cond.store: ; preds = %0
1137 // %4 = extractelement <16 x i32> %val, i32 0
1138 // %5 = getelementptr i32* %1, i32 0
1139 // store i32 %4, i32* %5
1142 // else: ; preds = %0, %cond.store
1143 // %6 = extractelement <16 x i1> %mask, i32 1
1144 // %7 = icmp eq i1 %6, true
1145 // br i1 %7, label %cond.store1, label %else2
1147 // cond.store1: ; preds = %else
1148 // %8 = extractelement <16 x i32> %val, i32 1
1149 // %9 = getelementptr i32* %1, i32 1
1150 // store i32 %8, i32* %9
1153 static void ScalarizeMaskedStore(CallInst *CI) {
1154 Value *Ptr = CI->getArgOperand(1);
1155 Value *Src = CI->getArgOperand(0);
1156 Value *Mask = CI->getArgOperand(3);
1158 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1159 Type *EltTy = VecType->getElementType();
1161 assert(VecType && "Unexpected data type in masked store intrinsic");
1163 IRBuilder<> Builder(CI->getContext());
1164 Instruction *InsertPt = CI;
1165 BasicBlock *IfBlock = CI->getParent();
1166 Builder.SetInsertPoint(InsertPt);
1167 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1169 // Bitcast %addr fron i8* to EltTy*
1171 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1172 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1174 unsigned VectorWidth = VecType->getNumElements();
1175 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1177 // Fill the "else" block, created in the previous iteration
1179 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1180 // %to_store = icmp eq i1 %mask_1, true
1181 // br i1 %to_load, label %cond.store, label %else
1183 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1184 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1185 ConstantInt::get(Predicate->getType(), 1));
1187 // Create "cond" block
1189 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1190 // %EltAddr = getelementptr i32* %1, i32 0
1191 // %store i32 %OneElt, i32* %EltAddr
1193 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1194 Builder.SetInsertPoint(InsertPt);
1196 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1197 Value* Gep = Builder.CreateInBoundsGEP(FirstEltPtr, Builder.getInt32(Idx));
1198 Builder.CreateStore(OneElt, Gep);
1200 // Create "else" block, fill it in the next iteration
1201 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1202 Builder.SetInsertPoint(InsertPt);
1203 Instruction *OldBr = IfBlock->getTerminator();
1204 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1205 OldBr->eraseFromParent();
1206 IfBlock = NewIfBlock;
1208 CI->eraseFromParent();
1211 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1212 BasicBlock *BB = CI->getParent();
1214 // Lower inline assembly if we can.
1215 // If we found an inline asm expession, and if the target knows how to
1216 // lower it to normal LLVM code, do so now.
1217 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1218 if (TLI->ExpandInlineAsm(CI)) {
1219 // Avoid invalidating the iterator.
1220 CurInstIterator = BB->begin();
1221 // Avoid processing instructions out of order, which could cause
1222 // reuse before a value is defined.
1226 // Sink address computing for memory operands into the block.
1227 if (OptimizeInlineAsmInst(CI))
1231 const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
1233 // Align the pointer arguments to this call if the target thinks it's a good
1235 unsigned MinSize, PrefAlign;
1236 if (TLI && TD && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1237 for (auto &Arg : CI->arg_operands()) {
1238 // We want to align both objects whose address is used directly and
1239 // objects whose address is used in casts and GEPs, though it only makes
1240 // sense for GEPs if the offset is a multiple of the desired alignment and
1241 // if size - offset meets the size threshold.
1242 if (!Arg->getType()->isPointerTy())
1244 APInt Offset(TD->getPointerSizeInBits(
1245 cast<PointerType>(Arg->getType())->getAddressSpace()), 0);
1246 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*TD, Offset);
1247 uint64_t Offset2 = Offset.getLimitedValue();
1249 if ((Offset2 & (PrefAlign-1)) == 0 &&
1250 (AI = dyn_cast<AllocaInst>(Val)) &&
1251 AI->getAlignment() < PrefAlign &&
1252 TD->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1253 AI->setAlignment(PrefAlign);
1254 // TODO: Also align GlobalVariables
1256 // If this is a memcpy (or similar) then we may be able to improve the
1258 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1259 unsigned Align = getKnownAlignment(MI->getDest(), *TD);
1260 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1261 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *TD));
1262 if (Align > MI->getAlignment())
1263 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1267 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1269 switch (II->getIntrinsicID()) {
1271 case Intrinsic::objectsize: {
1272 // Lower all uses of llvm.objectsize.*
1273 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1274 Type *ReturnTy = CI->getType();
1275 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1277 // Substituting this can cause recursive simplifications, which can
1278 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1280 WeakVH IterHandle(CurInstIterator);
1282 replaceAndRecursivelySimplify(CI, RetVal,
1283 TLInfo, ModifiedDT ? nullptr : DT);
1285 // If the iterator instruction was recursively deleted, start over at the
1286 // start of the block.
1287 if (IterHandle != CurInstIterator) {
1288 CurInstIterator = BB->begin();
1293 case Intrinsic::masked_load: {
1294 // Scalarize unsupported vector masked load
1295 if (!TTI->isLegalMaskedLoad(CI->getType(), 1)) {
1296 ScalarizeMaskedLoad(CI);
1302 case Intrinsic::masked_store: {
1303 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType(), 1)) {
1304 ScalarizeMaskedStore(CI);
1313 SmallVector<Value*, 2> PtrOps;
1315 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
1316 while (!PtrOps.empty())
1317 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
1322 // From here on out we're working with named functions.
1323 if (!CI->getCalledFunction()) return false;
1325 // Lower all default uses of _chk calls. This is very similar
1326 // to what InstCombineCalls does, but here we are only lowering calls
1327 // to fortified library functions (e.g. __memcpy_chk) that have the default
1328 // "don't know" as the objectsize. Anything else should be left alone.
1329 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1330 if (Value *V = Simplifier.optimizeCall(CI)) {
1331 CI->replaceAllUsesWith(V);
1332 CI->eraseFromParent();
1338 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
1339 /// instructions to the predecessor to enable tail call optimizations. The
1340 /// case it is currently looking for is:
1343 /// %tmp0 = tail call i32 @f0()
1344 /// br label %return
1346 /// %tmp1 = tail call i32 @f1()
1347 /// br label %return
1349 /// %tmp2 = tail call i32 @f2()
1350 /// br label %return
1352 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1360 /// %tmp0 = tail call i32 @f0()
1363 /// %tmp1 = tail call i32 @f1()
1366 /// %tmp2 = tail call i32 @f2()
1369 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
1373 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1377 PHINode *PN = nullptr;
1378 BitCastInst *BCI = nullptr;
1379 Value *V = RI->getReturnValue();
1381 BCI = dyn_cast<BitCastInst>(V);
1383 V = BCI->getOperand(0);
1385 PN = dyn_cast<PHINode>(V);
1390 if (PN && PN->getParent() != BB)
1393 // It's not safe to eliminate the sign / zero extension of the return value.
1394 // See llvm::isInTailCallPosition().
1395 const Function *F = BB->getParent();
1396 AttributeSet CallerAttrs = F->getAttributes();
1397 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1398 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1401 // Make sure there are no instructions between the PHI and return, or that the
1402 // return is the first instruction in the block.
1404 BasicBlock::iterator BI = BB->begin();
1405 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1407 // Also skip over the bitcast.
1412 BasicBlock::iterator BI = BB->begin();
1413 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1418 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1420 SmallVector<CallInst*, 4> TailCalls;
1422 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1423 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1424 // Make sure the phi value is indeed produced by the tail call.
1425 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1426 TLI->mayBeEmittedAsTailCall(CI))
1427 TailCalls.push_back(CI);
1430 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1431 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1432 if (!VisitedBBs.insert(*PI).second)
1435 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1436 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1437 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1438 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1442 CallInst *CI = dyn_cast<CallInst>(&*RI);
1443 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1444 TailCalls.push_back(CI);
1448 bool Changed = false;
1449 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1450 CallInst *CI = TailCalls[i];
1453 // Conservatively require the attributes of the call to match those of the
1454 // return. Ignore noalias because it doesn't affect the call sequence.
1455 AttributeSet CalleeAttrs = CS.getAttributes();
1456 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1457 removeAttribute(Attribute::NoAlias) !=
1458 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1459 removeAttribute(Attribute::NoAlias))
1462 // Make sure the call instruction is followed by an unconditional branch to
1463 // the return block.
1464 BasicBlock *CallBB = CI->getParent();
1465 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1466 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1469 // Duplicate the return into CallBB.
1470 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1471 ModifiedDT = Changed = true;
1475 // If we eliminated all predecessors of the block, delete the block now.
1476 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1477 BB->eraseFromParent();
1482 //===----------------------------------------------------------------------===//
1483 // Memory Optimization
1484 //===----------------------------------------------------------------------===//
1488 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1489 /// which holds actual Value*'s for register values.
1490 struct ExtAddrMode : public TargetLowering::AddrMode {
1493 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1494 void print(raw_ostream &OS) const;
1497 bool operator==(const ExtAddrMode& O) const {
1498 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1499 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1500 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1505 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1511 void ExtAddrMode::print(raw_ostream &OS) const {
1512 bool NeedPlus = false;
1515 OS << (NeedPlus ? " + " : "")
1517 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1522 OS << (NeedPlus ? " + " : "")
1528 OS << (NeedPlus ? " + " : "")
1530 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1534 OS << (NeedPlus ? " + " : "")
1536 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1542 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1543 void ExtAddrMode::dump() const {
1549 /// \brief This class provides transaction based operation on the IR.
1550 /// Every change made through this class is recorded in the internal state and
1551 /// can be undone (rollback) until commit is called.
1552 class TypePromotionTransaction {
1554 /// \brief This represents the common interface of the individual transaction.
1555 /// Each class implements the logic for doing one specific modification on
1556 /// the IR via the TypePromotionTransaction.
1557 class TypePromotionAction {
1559 /// The Instruction modified.
1563 /// \brief Constructor of the action.
1564 /// The constructor performs the related action on the IR.
1565 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1567 virtual ~TypePromotionAction() {}
1569 /// \brief Undo the modification done by this action.
1570 /// When this method is called, the IR must be in the same state as it was
1571 /// before this action was applied.
1572 /// \pre Undoing the action works if and only if the IR is in the exact same
1573 /// state as it was directly after this action was applied.
1574 virtual void undo() = 0;
1576 /// \brief Advocate every change made by this action.
1577 /// When the results on the IR of the action are to be kept, it is important
1578 /// to call this function, otherwise hidden information may be kept forever.
1579 virtual void commit() {
1580 // Nothing to be done, this action is not doing anything.
1584 /// \brief Utility to remember the position of an instruction.
1585 class InsertionHandler {
1586 /// Position of an instruction.
1587 /// Either an instruction:
1588 /// - Is the first in a basic block: BB is used.
1589 /// - Has a previous instructon: PrevInst is used.
1591 Instruction *PrevInst;
1594 /// Remember whether or not the instruction had a previous instruction.
1595 bool HasPrevInstruction;
1598 /// \brief Record the position of \p Inst.
1599 InsertionHandler(Instruction *Inst) {
1600 BasicBlock::iterator It = Inst;
1601 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1602 if (HasPrevInstruction)
1603 Point.PrevInst = --It;
1605 Point.BB = Inst->getParent();
1608 /// \brief Insert \p Inst at the recorded position.
1609 void insert(Instruction *Inst) {
1610 if (HasPrevInstruction) {
1611 if (Inst->getParent())
1612 Inst->removeFromParent();
1613 Inst->insertAfter(Point.PrevInst);
1615 Instruction *Position = Point.BB->getFirstInsertionPt();
1616 if (Inst->getParent())
1617 Inst->moveBefore(Position);
1619 Inst->insertBefore(Position);
1624 /// \brief Move an instruction before another.
1625 class InstructionMoveBefore : public TypePromotionAction {
1626 /// Original position of the instruction.
1627 InsertionHandler Position;
1630 /// \brief Move \p Inst before \p Before.
1631 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1632 : TypePromotionAction(Inst), Position(Inst) {
1633 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1634 Inst->moveBefore(Before);
1637 /// \brief Move the instruction back to its original position.
1638 void undo() override {
1639 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1640 Position.insert(Inst);
1644 /// \brief Set the operand of an instruction with a new value.
1645 class OperandSetter : public TypePromotionAction {
1646 /// Original operand of the instruction.
1648 /// Index of the modified instruction.
1652 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1653 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1654 : TypePromotionAction(Inst), Idx(Idx) {
1655 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1656 << "for:" << *Inst << "\n"
1657 << "with:" << *NewVal << "\n");
1658 Origin = Inst->getOperand(Idx);
1659 Inst->setOperand(Idx, NewVal);
1662 /// \brief Restore the original value of the instruction.
1663 void undo() override {
1664 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1665 << "for: " << *Inst << "\n"
1666 << "with: " << *Origin << "\n");
1667 Inst->setOperand(Idx, Origin);
1671 /// \brief Hide the operands of an instruction.
1672 /// Do as if this instruction was not using any of its operands.
1673 class OperandsHider : public TypePromotionAction {
1674 /// The list of original operands.
1675 SmallVector<Value *, 4> OriginalValues;
1678 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1679 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1680 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1681 unsigned NumOpnds = Inst->getNumOperands();
1682 OriginalValues.reserve(NumOpnds);
1683 for (unsigned It = 0; It < NumOpnds; ++It) {
1684 // Save the current operand.
1685 Value *Val = Inst->getOperand(It);
1686 OriginalValues.push_back(Val);
1688 // We could use OperandSetter here, but that would implied an overhead
1689 // that we are not willing to pay.
1690 Inst->setOperand(It, UndefValue::get(Val->getType()));
1694 /// \brief Restore the original list of uses.
1695 void undo() override {
1696 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1697 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1698 Inst->setOperand(It, OriginalValues[It]);
1702 /// \brief Build a truncate instruction.
1703 class TruncBuilder : public TypePromotionAction {
1706 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1708 /// trunc Opnd to Ty.
1709 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1710 IRBuilder<> Builder(Opnd);
1711 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
1712 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
1715 /// \brief Get the built value.
1716 Value *getBuiltValue() { return Val; }
1718 /// \brief Remove the built instruction.
1719 void undo() override {
1720 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
1721 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1722 IVal->eraseFromParent();
1726 /// \brief Build a sign extension instruction.
1727 class SExtBuilder : public TypePromotionAction {
1730 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1732 /// sext Opnd to Ty.
1733 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1734 : TypePromotionAction(InsertPt) {
1735 IRBuilder<> Builder(InsertPt);
1736 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
1737 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
1740 /// \brief Get the built value.
1741 Value *getBuiltValue() { return Val; }
1743 /// \brief Remove the built instruction.
1744 void undo() override {
1745 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
1746 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1747 IVal->eraseFromParent();
1751 /// \brief Build a zero extension instruction.
1752 class ZExtBuilder : public TypePromotionAction {
1755 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
1757 /// zext Opnd to Ty.
1758 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1759 : TypePromotionAction(InsertPt) {
1760 IRBuilder<> Builder(InsertPt);
1761 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
1762 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
1765 /// \brief Get the built value.
1766 Value *getBuiltValue() { return Val; }
1768 /// \brief Remove the built instruction.
1769 void undo() override {
1770 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
1771 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1772 IVal->eraseFromParent();
1776 /// \brief Mutate an instruction to another type.
1777 class TypeMutator : public TypePromotionAction {
1778 /// Record the original type.
1782 /// \brief Mutate the type of \p Inst into \p NewTy.
1783 TypeMutator(Instruction *Inst, Type *NewTy)
1784 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1785 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1787 Inst->mutateType(NewTy);
1790 /// \brief Mutate the instruction back to its original type.
1791 void undo() override {
1792 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1794 Inst->mutateType(OrigTy);
1798 /// \brief Replace the uses of an instruction by another instruction.
1799 class UsesReplacer : public TypePromotionAction {
1800 /// Helper structure to keep track of the replaced uses.
1801 struct InstructionAndIdx {
1802 /// The instruction using the instruction.
1804 /// The index where this instruction is used for Inst.
1806 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1807 : Inst(Inst), Idx(Idx) {}
1810 /// Keep track of the original uses (pair Instruction, Index).
1811 SmallVector<InstructionAndIdx, 4> OriginalUses;
1812 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1815 /// \brief Replace all the use of \p Inst by \p New.
1816 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1817 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1819 // Record the original uses.
1820 for (Use &U : Inst->uses()) {
1821 Instruction *UserI = cast<Instruction>(U.getUser());
1822 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1824 // Now, we can replace the uses.
1825 Inst->replaceAllUsesWith(New);
1828 /// \brief Reassign the original uses of Inst to Inst.
1829 void undo() override {
1830 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1831 for (use_iterator UseIt = OriginalUses.begin(),
1832 EndIt = OriginalUses.end();
1833 UseIt != EndIt; ++UseIt) {
1834 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1839 /// \brief Remove an instruction from the IR.
1840 class InstructionRemover : public TypePromotionAction {
1841 /// Original position of the instruction.
1842 InsertionHandler Inserter;
1843 /// Helper structure to hide all the link to the instruction. In other
1844 /// words, this helps to do as if the instruction was removed.
1845 OperandsHider Hider;
1846 /// Keep track of the uses replaced, if any.
1847 UsesReplacer *Replacer;
1850 /// \brief Remove all reference of \p Inst and optinally replace all its
1852 /// \pre If !Inst->use_empty(), then New != nullptr
1853 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1854 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1857 Replacer = new UsesReplacer(Inst, New);
1858 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1859 Inst->removeFromParent();
1862 ~InstructionRemover() { delete Replacer; }
1864 /// \brief Really remove the instruction.
1865 void commit() override { delete Inst; }
1867 /// \brief Resurrect the instruction and reassign it to the proper uses if
1868 /// new value was provided when build this action.
1869 void undo() override {
1870 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1871 Inserter.insert(Inst);
1879 /// Restoration point.
1880 /// The restoration point is a pointer to an action instead of an iterator
1881 /// because the iterator may be invalidated but not the pointer.
1882 typedef const TypePromotionAction *ConstRestorationPt;
1883 /// Advocate every changes made in that transaction.
1885 /// Undo all the changes made after the given point.
1886 void rollback(ConstRestorationPt Point);
1887 /// Get the current restoration point.
1888 ConstRestorationPt getRestorationPoint() const;
1890 /// \name API for IR modification with state keeping to support rollback.
1892 /// Same as Instruction::setOperand.
1893 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1894 /// Same as Instruction::eraseFromParent.
1895 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
1896 /// Same as Value::replaceAllUsesWith.
1897 void replaceAllUsesWith(Instruction *Inst, Value *New);
1898 /// Same as Value::mutateType.
1899 void mutateType(Instruction *Inst, Type *NewTy);
1900 /// Same as IRBuilder::createTrunc.
1901 Value *createTrunc(Instruction *Opnd, Type *Ty);
1902 /// Same as IRBuilder::createSExt.
1903 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1904 /// Same as IRBuilder::createZExt.
1905 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
1906 /// Same as Instruction::moveBefore.
1907 void moveBefore(Instruction *Inst, Instruction *Before);
1911 /// The ordered list of actions made so far.
1912 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
1913 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
1916 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1919 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
1922 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1925 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
1928 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1930 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
1933 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1934 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
1937 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1939 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
1940 Value *Val = Ptr->getBuiltValue();
1941 Actions.push_back(std::move(Ptr));
1945 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
1946 Value *Opnd, Type *Ty) {
1947 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
1948 Value *Val = Ptr->getBuiltValue();
1949 Actions.push_back(std::move(Ptr));
1953 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
1954 Value *Opnd, Type *Ty) {
1955 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
1956 Value *Val = Ptr->getBuiltValue();
1957 Actions.push_back(std::move(Ptr));
1961 void TypePromotionTransaction::moveBefore(Instruction *Inst,
1962 Instruction *Before) {
1964 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
1967 TypePromotionTransaction::ConstRestorationPt
1968 TypePromotionTransaction::getRestorationPoint() const {
1969 return !Actions.empty() ? Actions.back().get() : nullptr;
1972 void TypePromotionTransaction::commit() {
1973 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
1979 void TypePromotionTransaction::rollback(
1980 TypePromotionTransaction::ConstRestorationPt Point) {
1981 while (!Actions.empty() && Point != Actions.back().get()) {
1982 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
1987 /// \brief A helper class for matching addressing modes.
1989 /// This encapsulates the logic for matching the target-legal addressing modes.
1990 class AddressingModeMatcher {
1991 SmallVectorImpl<Instruction*> &AddrModeInsts;
1992 const TargetMachine &TM;
1993 const TargetLowering &TLI;
1995 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
1996 /// the memory instruction that we're computing this address for.
1998 Instruction *MemoryInst;
2000 /// AddrMode - This is the addressing mode that we're building up. This is
2001 /// part of the return value of this addressing mode matching stuff.
2002 ExtAddrMode &AddrMode;
2004 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
2005 const SetOfInstrs &InsertedTruncs;
2006 /// A map from the instructions to their type before promotion.
2007 InstrToOrigTy &PromotedInsts;
2008 /// The ongoing transaction where every action should be registered.
2009 TypePromotionTransaction &TPT;
2011 /// IgnoreProfitability - This is set to true when we should not do
2012 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
2013 /// always returns true.
2014 bool IgnoreProfitability;
2016 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2017 const TargetMachine &TM, Type *AT, Instruction *MI,
2018 ExtAddrMode &AM, const SetOfInstrs &InsertedTruncs,
2019 InstrToOrigTy &PromotedInsts,
2020 TypePromotionTransaction &TPT)
2021 : AddrModeInsts(AMI), TM(TM),
2022 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
2023 ->getTargetLowering()),
2024 AccessTy(AT), MemoryInst(MI), AddrMode(AM),
2025 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
2026 IgnoreProfitability = false;
2030 /// Match - Find the maximal addressing mode that a load/store of V can fold,
2031 /// give an access type of AccessTy. This returns a list of involved
2032 /// instructions in AddrModeInsts.
2033 /// \p InsertedTruncs The truncate instruction inserted by other
2036 /// \p PromotedInsts maps the instructions to their type before promotion.
2037 /// \p The ongoing transaction where every action should be registered.
2038 static ExtAddrMode Match(Value *V, Type *AccessTy,
2039 Instruction *MemoryInst,
2040 SmallVectorImpl<Instruction*> &AddrModeInsts,
2041 const TargetMachine &TM,
2042 const SetOfInstrs &InsertedTruncs,
2043 InstrToOrigTy &PromotedInsts,
2044 TypePromotionTransaction &TPT) {
2047 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy,
2048 MemoryInst, Result, InsertedTruncs,
2049 PromotedInsts, TPT).MatchAddr(V, 0);
2050 (void)Success; assert(Success && "Couldn't select *anything*?");
2054 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2055 bool MatchAddr(Value *V, unsigned Depth);
2056 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2057 bool *MovedAway = nullptr);
2058 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
2059 ExtAddrMode &AMBefore,
2060 ExtAddrMode &AMAfter);
2061 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2062 bool IsPromotionProfitable(unsigned NewCost, unsigned OldCost,
2063 Value *PromotedOperand) const;
2066 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
2067 /// Return true and update AddrMode if this addr mode is legal for the target,
2069 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
2071 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2072 // mode. Just process that directly.
2074 return MatchAddr(ScaleReg, Depth);
2076 // If the scale is 0, it takes nothing to add this.
2080 // If we already have a scale of this value, we can add to it, otherwise, we
2081 // need an available scale field.
2082 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2085 ExtAddrMode TestAddrMode = AddrMode;
2087 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2088 // [A+B + A*7] -> [B+A*8].
2089 TestAddrMode.Scale += Scale;
2090 TestAddrMode.ScaledReg = ScaleReg;
2092 // If the new address isn't legal, bail out.
2093 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
2096 // It was legal, so commit it.
2097 AddrMode = TestAddrMode;
2099 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2100 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2101 // X*Scale + C*Scale to addr mode.
2102 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2103 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2104 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2105 TestAddrMode.ScaledReg = AddLHS;
2106 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2108 // If this addressing mode is legal, commit it and remember that we folded
2109 // this instruction.
2110 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
2111 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2112 AddrMode = TestAddrMode;
2117 // Otherwise, not (x+c)*scale, just return what we have.
2121 /// MightBeFoldableInst - This is a little filter, which returns true if an
2122 /// addressing computation involving I might be folded into a load/store
2123 /// accessing it. This doesn't need to be perfect, but needs to accept at least
2124 /// the set of instructions that MatchOperationAddr can.
2125 static bool MightBeFoldableInst(Instruction *I) {
2126 switch (I->getOpcode()) {
2127 case Instruction::BitCast:
2128 case Instruction::AddrSpaceCast:
2129 // Don't touch identity bitcasts.
2130 if (I->getType() == I->getOperand(0)->getType())
2132 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2133 case Instruction::PtrToInt:
2134 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2136 case Instruction::IntToPtr:
2137 // We know the input is intptr_t, so this is foldable.
2139 case Instruction::Add:
2141 case Instruction::Mul:
2142 case Instruction::Shl:
2143 // Can only handle X*C and X << C.
2144 return isa<ConstantInt>(I->getOperand(1));
2145 case Instruction::GetElementPtr:
2152 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2153 /// \note \p Val is assumed to be the product of some type promotion.
2154 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2155 /// to be legal, as the non-promoted value would have had the same state.
2156 static bool isPromotedInstructionLegal(const TargetLowering &TLI, Value *Val) {
2157 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2160 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2161 // If the ISDOpcode is undefined, it was undefined before the promotion.
2164 // Otherwise, check if the promoted instruction is legal or not.
2165 return TLI.isOperationLegalOrCustom(
2166 ISDOpcode, TLI.getValueType(PromotedInst->getType()));
2169 /// \brief Hepler class to perform type promotion.
2170 class TypePromotionHelper {
2171 /// \brief Utility function to check whether or not a sign or zero extension
2172 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2173 /// either using the operands of \p Inst or promoting \p Inst.
2174 /// The type of the extension is defined by \p IsSExt.
2175 /// In other words, check if:
2176 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2177 /// #1 Promotion applies:
2178 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2179 /// #2 Operand reuses:
2180 /// ext opnd1 to ConsideredExtType.
2181 /// \p PromotedInsts maps the instructions to their type before promotion.
2182 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2183 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2185 /// \brief Utility function to determine if \p OpIdx should be promoted when
2186 /// promoting \p Inst.
2187 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2188 if (isa<SelectInst>(Inst) && OpIdx == 0)
2193 /// \brief Utility function to promote the operand of \p Ext when this
2194 /// operand is a promotable trunc or sext or zext.
2195 /// \p PromotedInsts maps the instructions to their type before promotion.
2196 /// \p CreatedInstsCost[out] contains the cost of all instructions
2197 /// created to promote the operand of Ext.
2198 /// Newly added extensions are inserted in \p Exts.
2199 /// Newly added truncates are inserted in \p Truncs.
2200 /// Should never be called directly.
2201 /// \return The promoted value which is used instead of Ext.
2202 static Value *promoteOperandForTruncAndAnyExt(
2203 Instruction *Ext, TypePromotionTransaction &TPT,
2204 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2205 SmallVectorImpl<Instruction *> *Exts,
2206 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
2208 /// \brief Utility function to promote the operand of \p Ext when this
2209 /// operand is promotable and is not a supported trunc or sext.
2210 /// \p PromotedInsts maps the instructions to their type before promotion.
2211 /// \p CreatedInstsCost[out] contains the cost of all the instructions
2212 /// created to promote the operand of Ext.
2213 /// Newly added extensions are inserted in \p Exts.
2214 /// Newly added truncates are inserted in \p Truncs.
2215 /// Should never be called directly.
2216 /// \return The promoted value which is used instead of Ext.
2217 static Value *promoteOperandForOther(Instruction *Ext,
2218 TypePromotionTransaction &TPT,
2219 InstrToOrigTy &PromotedInsts,
2220 unsigned &CreatedInstsCost,
2221 SmallVectorImpl<Instruction *> *Exts,
2222 SmallVectorImpl<Instruction *> *Truncs,
2223 const TargetLowering &TLI, bool IsSExt);
2225 /// \see promoteOperandForOther.
2226 static Value *signExtendOperandForOther(
2227 Instruction *Ext, TypePromotionTransaction &TPT,
2228 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2229 SmallVectorImpl<Instruction *> *Exts,
2230 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2231 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2232 Exts, Truncs, TLI, true);
2235 /// \see promoteOperandForOther.
2236 static Value *zeroExtendOperandForOther(
2237 Instruction *Ext, TypePromotionTransaction &TPT,
2238 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2239 SmallVectorImpl<Instruction *> *Exts,
2240 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2241 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2242 Exts, Truncs, TLI, false);
2246 /// Type for the utility function that promotes the operand of Ext.
2247 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2248 InstrToOrigTy &PromotedInsts,
2249 unsigned &CreatedInstsCost,
2250 SmallVectorImpl<Instruction *> *Exts,
2251 SmallVectorImpl<Instruction *> *Truncs,
2252 const TargetLowering &TLI);
2253 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2254 /// action to promote the operand of \p Ext instead of using Ext.
2255 /// \return NULL if no promotable action is possible with the current
2257 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
2258 /// the others CodeGenPrepare optimizations. This information is important
2259 /// because we do not want to promote these instructions as CodeGenPrepare
2260 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2261 /// \p PromotedInsts maps the instructions to their type before promotion.
2262 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2263 const TargetLowering &TLI,
2264 const InstrToOrigTy &PromotedInsts);
2267 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2268 Type *ConsideredExtType,
2269 const InstrToOrigTy &PromotedInsts,
2271 // The promotion helper does not know how to deal with vector types yet.
2272 // To be able to fix that, we would need to fix the places where we
2273 // statically extend, e.g., constants and such.
2274 if (Inst->getType()->isVectorTy())
2277 // We can always get through zext.
2278 if (isa<ZExtInst>(Inst))
2281 // sext(sext) is ok too.
2282 if (IsSExt && isa<SExtInst>(Inst))
2285 // We can get through binary operator, if it is legal. In other words, the
2286 // binary operator must have a nuw or nsw flag.
2287 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2288 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2289 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2290 (IsSExt && BinOp->hasNoSignedWrap())))
2293 // Check if we can do the following simplification.
2294 // ext(trunc(opnd)) --> ext(opnd)
2295 if (!isa<TruncInst>(Inst))
2298 Value *OpndVal = Inst->getOperand(0);
2299 // Check if we can use this operand in the extension.
2300 // If the type is larger than the result type of the extension,
2302 if (!OpndVal->getType()->isIntegerTy() ||
2303 OpndVal->getType()->getIntegerBitWidth() >
2304 ConsideredExtType->getIntegerBitWidth())
2307 // If the operand of the truncate is not an instruction, we will not have
2308 // any information on the dropped bits.
2309 // (Actually we could for constant but it is not worth the extra logic).
2310 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2314 // Check if the source of the type is narrow enough.
2315 // I.e., check that trunc just drops extended bits of the same kind of
2317 // #1 get the type of the operand and check the kind of the extended bits.
2318 const Type *OpndType;
2319 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2320 if (It != PromotedInsts.end() && It->second.IsSExt == IsSExt)
2321 OpndType = It->second.Ty;
2322 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2323 OpndType = Opnd->getOperand(0)->getType();
2327 // #2 check that the truncate just drop extended bits.
2328 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
2334 TypePromotionHelper::Action TypePromotionHelper::getAction(
2335 Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2336 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2337 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2338 "Unexpected instruction type");
2339 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2340 Type *ExtTy = Ext->getType();
2341 bool IsSExt = isa<SExtInst>(Ext);
2342 // If the operand of the extension is not an instruction, we cannot
2344 // If it, check we can get through.
2345 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2348 // Do not promote if the operand has been added by codegenprepare.
2349 // Otherwise, it means we are undoing an optimization that is likely to be
2350 // redone, thus causing potential infinite loop.
2351 if (isa<TruncInst>(ExtOpnd) && InsertedTruncs.count(ExtOpnd))
2354 // SExt or Trunc instructions.
2355 // Return the related handler.
2356 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2357 isa<ZExtInst>(ExtOpnd))
2358 return promoteOperandForTruncAndAnyExt;
2360 // Regular instruction.
2361 // Abort early if we will have to insert non-free instructions.
2362 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2364 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2367 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2368 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2369 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2370 SmallVectorImpl<Instruction *> *Exts,
2371 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2372 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2373 // get through it and this method should not be called.
2374 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2375 Value *ExtVal = SExt;
2376 bool HasMergedNonFreeExt = false;
2377 if (isa<ZExtInst>(SExtOpnd)) {
2378 // Replace s|zext(zext(opnd))
2380 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
2382 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2383 TPT.replaceAllUsesWith(SExt, ZExt);
2384 TPT.eraseInstruction(SExt);
2387 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2389 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2391 CreatedInstsCost = 0;
2393 // Remove dead code.
2394 if (SExtOpnd->use_empty())
2395 TPT.eraseInstruction(SExtOpnd);
2397 // Check if the extension is still needed.
2398 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2399 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2402 Exts->push_back(ExtInst);
2403 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
2408 // At this point we have: ext ty opnd to ty.
2409 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2410 Value *NextVal = ExtInst->getOperand(0);
2411 TPT.eraseInstruction(ExtInst, NextVal);
2415 Value *TypePromotionHelper::promoteOperandForOther(
2416 Instruction *Ext, TypePromotionTransaction &TPT,
2417 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2418 SmallVectorImpl<Instruction *> *Exts,
2419 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
2421 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2422 // get through it and this method should not be called.
2423 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2424 CreatedInstsCost = 0;
2425 if (!ExtOpnd->hasOneUse()) {
2426 // ExtOpnd will be promoted.
2427 // All its uses, but Ext, will need to use a truncated value of the
2428 // promoted version.
2429 // Create the truncate now.
2430 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2431 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2432 ITrunc->removeFromParent();
2433 // Insert it just after the definition.
2434 ITrunc->insertAfter(ExtOpnd);
2436 Truncs->push_back(ITrunc);
2439 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2440 // Restore the operand of Ext (which has been replace by the previous call
2441 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2442 TPT.setOperand(Ext, 0, ExtOpnd);
2445 // Get through the Instruction:
2446 // 1. Update its type.
2447 // 2. Replace the uses of Ext by Inst.
2448 // 3. Extend each operand that needs to be extended.
2450 // Remember the original type of the instruction before promotion.
2451 // This is useful to know that the high bits are sign extended bits.
2452 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2453 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2455 TPT.mutateType(ExtOpnd, Ext->getType());
2457 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2459 Instruction *ExtForOpnd = Ext;
2461 DEBUG(dbgs() << "Propagate Ext to operands\n");
2462 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2464 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2465 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2466 !shouldExtOperand(ExtOpnd, OpIdx)) {
2467 DEBUG(dbgs() << "No need to propagate\n");
2470 // Check if we can statically extend the operand.
2471 Value *Opnd = ExtOpnd->getOperand(OpIdx);
2472 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2473 DEBUG(dbgs() << "Statically extend\n");
2474 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2475 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
2476 : Cst->getValue().zext(BitWidth);
2477 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
2480 // UndefValue are typed, so we have to statically sign extend them.
2481 if (isa<UndefValue>(Opnd)) {
2482 DEBUG(dbgs() << "Statically extend\n");
2483 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2487 // Otherwise we have to explicity sign extend the operand.
2488 // Check if Ext was reused to extend an operand.
2490 // If yes, create a new one.
2491 DEBUG(dbgs() << "More operands to ext\n");
2492 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2493 : TPT.createZExt(Ext, Opnd, Ext->getType());
2494 if (!isa<Instruction>(ValForExtOpnd)) {
2495 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
2498 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
2501 Exts->push_back(ExtForOpnd);
2502 TPT.setOperand(ExtForOpnd, 0, Opnd);
2504 // Move the sign extension before the insertion point.
2505 TPT.moveBefore(ExtForOpnd, ExtOpnd);
2506 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2507 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
2508 // If more sext are required, new instructions will have to be created.
2509 ExtForOpnd = nullptr;
2511 if (ExtForOpnd == Ext) {
2512 DEBUG(dbgs() << "Extension is useless now\n");
2513 TPT.eraseInstruction(Ext);
2518 /// IsPromotionProfitable - Check whether or not promoting an instruction
2519 /// to a wider type was profitable.
2520 /// \p NewCost gives the cost of extension instructions created by the
2522 /// \p OldCost gives the cost of extension instructions before the promotion
2523 /// plus the number of instructions that have been
2524 /// matched in the addressing mode the promotion.
2525 /// \p PromotedOperand is the value that has been promoted.
2526 /// \return True if the promotion is profitable, false otherwise.
2527 bool AddressingModeMatcher::IsPromotionProfitable(
2528 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
2529 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
2530 // The cost of the new extensions is greater than the cost of the
2531 // old extension plus what we folded.
2532 // This is not profitable.
2533 if (NewCost > OldCost)
2535 if (NewCost < OldCost)
2537 // The promotion is neutral but it may help folding the sign extension in
2538 // loads for instance.
2539 // Check that we did not create an illegal instruction.
2540 return isPromotedInstructionLegal(TLI, PromotedOperand);
2543 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
2544 /// fold the operation into the addressing mode. If so, update the addressing
2545 /// mode and return true, otherwise return false without modifying AddrMode.
2546 /// If \p MovedAway is not NULL, it contains the information of whether or
2547 /// not AddrInst has to be folded into the addressing mode on success.
2548 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2549 /// because it has been moved away.
2550 /// Thus AddrInst must not be added in the matched instructions.
2551 /// This state can happen when AddrInst is a sext, since it may be moved away.
2552 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2553 /// not be referenced anymore.
2554 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
2557 // Avoid exponential behavior on extremely deep expression trees.
2558 if (Depth >= 5) return false;
2560 // By default, all matched instructions stay in place.
2565 case Instruction::PtrToInt:
2566 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2567 return MatchAddr(AddrInst->getOperand(0), Depth);
2568 case Instruction::IntToPtr:
2569 // This inttoptr is a no-op if the integer type is pointer sized.
2570 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
2571 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
2572 return MatchAddr(AddrInst->getOperand(0), Depth);
2574 case Instruction::BitCast:
2575 case Instruction::AddrSpaceCast:
2576 // BitCast is always a noop, and we can handle it as long as it is
2577 // int->int or pointer->pointer (we don't want int<->fp or something).
2578 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2579 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2580 // Don't touch identity bitcasts. These were probably put here by LSR,
2581 // and we don't want to mess around with them. Assume it knows what it
2583 AddrInst->getOperand(0)->getType() != AddrInst->getType())
2584 return MatchAddr(AddrInst->getOperand(0), Depth);
2586 case Instruction::Add: {
2587 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2588 ExtAddrMode BackupAddrMode = AddrMode;
2589 unsigned OldSize = AddrModeInsts.size();
2590 // Start a transaction at this point.
2591 // The LHS may match but not the RHS.
2592 // Therefore, we need a higher level restoration point to undo partially
2593 // matched operation.
2594 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2595 TPT.getRestorationPoint();
2597 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2598 MatchAddr(AddrInst->getOperand(0), Depth+1))
2601 // Restore the old addr mode info.
2602 AddrMode = BackupAddrMode;
2603 AddrModeInsts.resize(OldSize);
2604 TPT.rollback(LastKnownGood);
2606 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2607 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2608 MatchAddr(AddrInst->getOperand(1), Depth+1))
2611 // Otherwise we definitely can't merge the ADD in.
2612 AddrMode = BackupAddrMode;
2613 AddrModeInsts.resize(OldSize);
2614 TPT.rollback(LastKnownGood);
2617 //case Instruction::Or:
2618 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2620 case Instruction::Mul:
2621 case Instruction::Shl: {
2622 // Can only handle X*C and X << C.
2623 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2626 int64_t Scale = RHS->getSExtValue();
2627 if (Opcode == Instruction::Shl)
2628 Scale = 1LL << Scale;
2630 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2632 case Instruction::GetElementPtr: {
2633 // Scan the GEP. We check it if it contains constant offsets and at most
2634 // one variable offset.
2635 int VariableOperand = -1;
2636 unsigned VariableScale = 0;
2638 int64_t ConstantOffset = 0;
2639 const DataLayout *TD = TLI.getDataLayout();
2640 gep_type_iterator GTI = gep_type_begin(AddrInst);
2641 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2642 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2643 const StructLayout *SL = TD->getStructLayout(STy);
2645 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2646 ConstantOffset += SL->getElementOffset(Idx);
2648 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
2649 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2650 ConstantOffset += CI->getSExtValue()*TypeSize;
2651 } else if (TypeSize) { // Scales of zero don't do anything.
2652 // We only allow one variable index at the moment.
2653 if (VariableOperand != -1)
2656 // Remember the variable index.
2657 VariableOperand = i;
2658 VariableScale = TypeSize;
2663 // A common case is for the GEP to only do a constant offset. In this case,
2664 // just add it to the disp field and check validity.
2665 if (VariableOperand == -1) {
2666 AddrMode.BaseOffs += ConstantOffset;
2667 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
2668 // Check to see if we can fold the base pointer in too.
2669 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2672 AddrMode.BaseOffs -= ConstantOffset;
2676 // Save the valid addressing mode in case we can't match.
2677 ExtAddrMode BackupAddrMode = AddrMode;
2678 unsigned OldSize = AddrModeInsts.size();
2680 // See if the scale and offset amount is valid for this target.
2681 AddrMode.BaseOffs += ConstantOffset;
2683 // Match the base operand of the GEP.
2684 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2685 // If it couldn't be matched, just stuff the value in a register.
2686 if (AddrMode.HasBaseReg) {
2687 AddrMode = BackupAddrMode;
2688 AddrModeInsts.resize(OldSize);
2691 AddrMode.HasBaseReg = true;
2692 AddrMode.BaseReg = AddrInst->getOperand(0);
2695 // Match the remaining variable portion of the GEP.
2696 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2698 // If it couldn't be matched, try stuffing the base into a register
2699 // instead of matching it, and retrying the match of the scale.
2700 AddrMode = BackupAddrMode;
2701 AddrModeInsts.resize(OldSize);
2702 if (AddrMode.HasBaseReg)
2704 AddrMode.HasBaseReg = true;
2705 AddrMode.BaseReg = AddrInst->getOperand(0);
2706 AddrMode.BaseOffs += ConstantOffset;
2707 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2708 VariableScale, Depth)) {
2709 // If even that didn't work, bail.
2710 AddrMode = BackupAddrMode;
2711 AddrModeInsts.resize(OldSize);
2718 case Instruction::SExt:
2719 case Instruction::ZExt: {
2720 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
2724 // Try to move this ext out of the way of the addressing mode.
2725 // Ask for a method for doing so.
2726 TypePromotionHelper::Action TPH =
2727 TypePromotionHelper::getAction(Ext, InsertedTruncs, TLI, PromotedInsts);
2731 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2732 TPT.getRestorationPoint();
2733 unsigned CreatedInstsCost = 0;
2734 unsigned ExtCost = !TLI.isExtFree(Ext);
2735 Value *PromotedOperand =
2736 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
2737 // SExt has been moved away.
2738 // Thus either it will be rematched later in the recursive calls or it is
2739 // gone. Anyway, we must not fold it into the addressing mode at this point.
2743 // addr = gep base, idx
2745 // promotedOpnd = ext opnd <- no match here
2746 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2747 // addr = gep base, op <- match
2751 assert(PromotedOperand &&
2752 "TypePromotionHelper should have filtered out those cases");
2754 ExtAddrMode BackupAddrMode = AddrMode;
2755 unsigned OldSize = AddrModeInsts.size();
2757 if (!MatchAddr(PromotedOperand, Depth) ||
2758 // The total of the new cost is equals to the cost of the created
2760 // The total of the old cost is equals to the cost of the extension plus
2761 // what we have saved in the addressing mode.
2762 !IsPromotionProfitable(CreatedInstsCost,
2763 ExtCost + (AddrModeInsts.size() - OldSize),
2765 AddrMode = BackupAddrMode;
2766 AddrModeInsts.resize(OldSize);
2767 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2768 TPT.rollback(LastKnownGood);
2777 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2778 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2779 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2780 /// or intptr_t for the target.
2782 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2783 // Start a transaction at this point that we will rollback if the matching
2785 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2786 TPT.getRestorationPoint();
2787 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2788 // Fold in immediates if legal for the target.
2789 AddrMode.BaseOffs += CI->getSExtValue();
2790 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2792 AddrMode.BaseOffs -= CI->getSExtValue();
2793 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2794 // If this is a global variable, try to fold it into the addressing mode.
2795 if (!AddrMode.BaseGV) {
2796 AddrMode.BaseGV = GV;
2797 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2799 AddrMode.BaseGV = nullptr;
2801 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2802 ExtAddrMode BackupAddrMode = AddrMode;
2803 unsigned OldSize = AddrModeInsts.size();
2805 // Check to see if it is possible to fold this operation.
2806 bool MovedAway = false;
2807 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2808 // This instruction may have been move away. If so, there is nothing
2812 // Okay, it's possible to fold this. Check to see if it is actually
2813 // *profitable* to do so. We use a simple cost model to avoid increasing
2814 // register pressure too much.
2815 if (I->hasOneUse() ||
2816 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2817 AddrModeInsts.push_back(I);
2821 // It isn't profitable to do this, roll back.
2822 //cerr << "NOT FOLDING: " << *I;
2823 AddrMode = BackupAddrMode;
2824 AddrModeInsts.resize(OldSize);
2825 TPT.rollback(LastKnownGood);
2827 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2828 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2830 TPT.rollback(LastKnownGood);
2831 } else if (isa<ConstantPointerNull>(Addr)) {
2832 // Null pointer gets folded without affecting the addressing mode.
2836 // Worse case, the target should support [reg] addressing modes. :)
2837 if (!AddrMode.HasBaseReg) {
2838 AddrMode.HasBaseReg = true;
2839 AddrMode.BaseReg = Addr;
2840 // Still check for legality in case the target supports [imm] but not [i+r].
2841 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2843 AddrMode.HasBaseReg = false;
2844 AddrMode.BaseReg = nullptr;
2847 // If the base register is already taken, see if we can do [r+r].
2848 if (AddrMode.Scale == 0) {
2850 AddrMode.ScaledReg = Addr;
2851 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2854 AddrMode.ScaledReg = nullptr;
2857 TPT.rollback(LastKnownGood);
2861 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2862 /// inline asm call are due to memory operands. If so, return true, otherwise
2864 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2865 const TargetMachine &TM) {
2866 const Function *F = CI->getParent()->getParent();
2867 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
2868 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
2869 TargetLowering::AsmOperandInfoVector TargetConstraints =
2870 TLI->ParseConstraints(TRI, ImmutableCallSite(CI));
2871 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2872 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2874 // Compute the constraint code and ConstraintType to use.
2875 TLI->ComputeConstraintToUse(OpInfo, SDValue());
2877 // If this asm operand is our Value*, and if it isn't an indirect memory
2878 // operand, we can't fold it!
2879 if (OpInfo.CallOperandVal == OpVal &&
2880 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2881 !OpInfo.isIndirect))
2888 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2889 /// memory use. If we find an obviously non-foldable instruction, return true.
2890 /// Add the ultimately found memory instructions to MemoryUses.
2891 static bool FindAllMemoryUses(
2893 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
2894 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
2895 // If we already considered this instruction, we're done.
2896 if (!ConsideredInsts.insert(I).second)
2899 // If this is an obviously unfoldable instruction, bail out.
2900 if (!MightBeFoldableInst(I))
2903 // Loop over all the uses, recursively processing them.
2904 for (Use &U : I->uses()) {
2905 Instruction *UserI = cast<Instruction>(U.getUser());
2907 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2908 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2912 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2913 unsigned opNo = U.getOperandNo();
2914 if (opNo == 0) return true; // Storing addr, not into addr.
2915 MemoryUses.push_back(std::make_pair(SI, opNo));
2919 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2920 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2921 if (!IA) return true;
2923 // If this is a memory operand, we're cool, otherwise bail out.
2924 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
2929 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
2936 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2937 /// the use site that we're folding it into. If so, there is no cost to
2938 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2939 /// that we know are live at the instruction already.
2940 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2941 Value *KnownLive2) {
2942 // If Val is either of the known-live values, we know it is live!
2943 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
2946 // All values other than instructions and arguments (e.g. constants) are live.
2947 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2949 // If Val is a constant sized alloca in the entry block, it is live, this is
2950 // true because it is just a reference to the stack/frame pointer, which is
2951 // live for the whole function.
2952 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2953 if (AI->isStaticAlloca())
2956 // Check to see if this value is already used in the memory instruction's
2957 // block. If so, it's already live into the block at the very least, so we
2958 // can reasonably fold it.
2959 return Val->isUsedInBasicBlock(MemoryInst->getParent());
2962 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
2963 /// mode of the machine to fold the specified instruction into a load or store
2964 /// that ultimately uses it. However, the specified instruction has multiple
2965 /// uses. Given this, it may actually increase register pressure to fold it
2966 /// into the load. For example, consider this code:
2970 /// use(Y) -> nonload/store
2974 /// In this case, Y has multiple uses, and can be folded into the load of Z
2975 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
2976 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
2977 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
2978 /// number of computations either.
2980 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
2981 /// X was live across 'load Z' for other reasons, we actually *would* want to
2982 /// fold the addressing mode in the Z case. This would make Y die earlier.
2983 bool AddressingModeMatcher::
2984 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
2985 ExtAddrMode &AMAfter) {
2986 if (IgnoreProfitability) return true;
2988 // AMBefore is the addressing mode before this instruction was folded into it,
2989 // and AMAfter is the addressing mode after the instruction was folded. Get
2990 // the set of registers referenced by AMAfter and subtract out those
2991 // referenced by AMBefore: this is the set of values which folding in this
2992 // address extends the lifetime of.
2994 // Note that there are only two potential values being referenced here,
2995 // BaseReg and ScaleReg (global addresses are always available, as are any
2996 // folded immediates).
2997 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
2999 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
3000 // lifetime wasn't extended by adding this instruction.
3001 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3003 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3004 ScaledReg = nullptr;
3006 // If folding this instruction (and it's subexprs) didn't extend any live
3007 // ranges, we're ok with it.
3008 if (!BaseReg && !ScaledReg)
3011 // If all uses of this instruction are ultimately load/store/inlineasm's,
3012 // check to see if their addressing modes will include this instruction. If
3013 // so, we can fold it into all uses, so it doesn't matter if it has multiple
3015 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
3016 SmallPtrSet<Instruction*, 16> ConsideredInsts;
3017 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
3018 return false; // Has a non-memory, non-foldable use!
3020 // Now that we know that all uses of this instruction are part of a chain of
3021 // computation involving only operations that could theoretically be folded
3022 // into a memory use, loop over each of these uses and see if they could
3023 // *actually* fold the instruction.
3024 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
3025 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
3026 Instruction *User = MemoryUses[i].first;
3027 unsigned OpNo = MemoryUses[i].second;
3029 // Get the access type of this use. If the use isn't a pointer, we don't
3030 // know what it accesses.
3031 Value *Address = User->getOperand(OpNo);
3032 if (!Address->getType()->isPointerTy())
3034 Type *AddressAccessTy = Address->getType()->getPointerElementType();
3036 // Do a match against the root of this address, ignoring profitability. This
3037 // will tell us if the addressing mode for the memory operation will
3038 // *actually* cover the shared instruction.
3040 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3041 TPT.getRestorationPoint();
3042 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy,
3043 MemoryInst, Result, InsertedTruncs,
3044 PromotedInsts, TPT);
3045 Matcher.IgnoreProfitability = true;
3046 bool Success = Matcher.MatchAddr(Address, 0);
3047 (void)Success; assert(Success && "Couldn't select *anything*?");
3049 // The match was to check the profitability, the changes made are not
3050 // part of the original matcher. Therefore, they should be dropped
3051 // otherwise the original matcher will not present the right state.
3052 TPT.rollback(LastKnownGood);
3054 // If the match didn't cover I, then it won't be shared by it.
3055 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3056 I) == MatchedAddrModeInsts.end())
3059 MatchedAddrModeInsts.clear();
3065 } // end anonymous namespace
3067 /// IsNonLocalValue - Return true if the specified values are defined in a
3068 /// different basic block than BB.
3069 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3070 if (Instruction *I = dyn_cast<Instruction>(V))
3071 return I->getParent() != BB;
3075 /// OptimizeMemoryInst - Load and Store Instructions often have
3076 /// addressing modes that can do significant amounts of computation. As such,
3077 /// instruction selection will try to get the load or store to do as much
3078 /// computation as possible for the program. The problem is that isel can only
3079 /// see within a single block. As such, we sink as much legal addressing mode
3080 /// stuff into the block as possible.
3082 /// This method is used to optimize both load/store and inline asms with memory
3084 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3088 // Try to collapse single-value PHI nodes. This is necessary to undo
3089 // unprofitable PRE transformations.
3090 SmallVector<Value*, 8> worklist;
3091 SmallPtrSet<Value*, 16> Visited;
3092 worklist.push_back(Addr);
3094 // Use a worklist to iteratively look through PHI nodes, and ensure that
3095 // the addressing mode obtained from the non-PHI roots of the graph
3097 Value *Consensus = nullptr;
3098 unsigned NumUsesConsensus = 0;
3099 bool IsNumUsesConsensusValid = false;
3100 SmallVector<Instruction*, 16> AddrModeInsts;
3101 ExtAddrMode AddrMode;
3102 TypePromotionTransaction TPT;
3103 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3104 TPT.getRestorationPoint();
3105 while (!worklist.empty()) {
3106 Value *V = worklist.back();
3107 worklist.pop_back();
3109 // Break use-def graph loops.
3110 if (!Visited.insert(V).second) {
3111 Consensus = nullptr;
3115 // For a PHI node, push all of its incoming values.
3116 if (PHINode *P = dyn_cast<PHINode>(V)) {
3117 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
3118 worklist.push_back(P->getIncomingValue(i));
3122 // For non-PHIs, determine the addressing mode being computed.
3123 SmallVector<Instruction*, 16> NewAddrModeInsts;
3124 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3125 V, AccessTy, MemoryInst, NewAddrModeInsts, *TM, InsertedTruncsSet,
3126 PromotedInsts, TPT);
3128 // This check is broken into two cases with very similar code to avoid using
3129 // getNumUses() as much as possible. Some values have a lot of uses, so
3130 // calling getNumUses() unconditionally caused a significant compile-time
3134 AddrMode = NewAddrMode;
3135 AddrModeInsts = NewAddrModeInsts;
3137 } else if (NewAddrMode == AddrMode) {
3138 if (!IsNumUsesConsensusValid) {
3139 NumUsesConsensus = Consensus->getNumUses();
3140 IsNumUsesConsensusValid = true;
3143 // Ensure that the obtained addressing mode is equivalent to that obtained
3144 // for all other roots of the PHI traversal. Also, when choosing one
3145 // such root as representative, select the one with the most uses in order
3146 // to keep the cost modeling heuristics in AddressingModeMatcher
3148 unsigned NumUses = V->getNumUses();
3149 if (NumUses > NumUsesConsensus) {
3151 NumUsesConsensus = NumUses;
3152 AddrModeInsts = NewAddrModeInsts;
3157 Consensus = nullptr;
3161 // If the addressing mode couldn't be determined, or if multiple different
3162 // ones were determined, bail out now.
3164 TPT.rollback(LastKnownGood);
3169 // Check to see if any of the instructions supersumed by this addr mode are
3170 // non-local to I's BB.
3171 bool AnyNonLocal = false;
3172 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3173 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3179 // If all the instructions matched are already in this BB, don't do anything.
3181 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3185 // Insert this computation right after this user. Since our caller is
3186 // scanning from the top of the BB to the bottom, reuse of the expr are
3187 // guaranteed to happen later.
3188 IRBuilder<> Builder(MemoryInst);
3190 // Now that we determined the addressing expression we want to use and know
3191 // that we have to sink it into this block. Check to see if we have already
3192 // done this for some other load/store instr in this block. If so, reuse the
3194 Value *&SunkAddr = SunkAddrs[Addr];
3196 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3197 << *MemoryInst << "\n");
3198 if (SunkAddr->getType() != Addr->getType())
3199 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3200 } else if (AddrSinkUsingGEPs ||
3201 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3202 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3204 // By default, we use the GEP-based method when AA is used later. This
3205 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3206 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3207 << *MemoryInst << "\n");
3208 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3209 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3211 // First, find the pointer.
3212 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3213 ResultPtr = AddrMode.BaseReg;
3214 AddrMode.BaseReg = nullptr;
3217 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3218 // We can't add more than one pointer together, nor can we scale a
3219 // pointer (both of which seem meaningless).
3220 if (ResultPtr || AddrMode.Scale != 1)
3223 ResultPtr = AddrMode.ScaledReg;
3227 if (AddrMode.BaseGV) {
3231 ResultPtr = AddrMode.BaseGV;
3234 // If the real base value actually came from an inttoptr, then the matcher
3235 // will look through it and provide only the integer value. In that case,
3237 if (!ResultPtr && AddrMode.BaseReg) {
3239 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3240 AddrMode.BaseReg = nullptr;
3241 } else if (!ResultPtr && AddrMode.Scale == 1) {
3243 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3248 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3249 SunkAddr = Constant::getNullValue(Addr->getType());
3250 } else if (!ResultPtr) {
3254 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3256 // Start with the base register. Do this first so that subsequent address
3257 // matching finds it last, which will prevent it from trying to match it
3258 // as the scaled value in case it happens to be a mul. That would be
3259 // problematic if we've sunk a different mul for the scale, because then
3260 // we'd end up sinking both muls.
3261 if (AddrMode.BaseReg) {
3262 Value *V = AddrMode.BaseReg;
3263 if (V->getType() != IntPtrTy)
3264 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3269 // Add the scale value.
3270 if (AddrMode.Scale) {
3271 Value *V = AddrMode.ScaledReg;
3272 if (V->getType() == IntPtrTy) {
3274 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3275 cast<IntegerType>(V->getType())->getBitWidth()) {
3276 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3278 // It is only safe to sign extend the BaseReg if we know that the math
3279 // required to create it did not overflow before we extend it. Since
3280 // the original IR value was tossed in favor of a constant back when
3281 // the AddrMode was created we need to bail out gracefully if widths
3282 // do not match instead of extending it.
3283 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3284 if (I && (ResultIndex != AddrMode.BaseReg))
3285 I->eraseFromParent();
3289 if (AddrMode.Scale != 1)
3290 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3293 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3298 // Add in the Base Offset if present.
3299 if (AddrMode.BaseOffs) {
3300 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3302 // We need to add this separately from the scale above to help with
3303 // SDAG consecutive load/store merging.
3304 if (ResultPtr->getType() != I8PtrTy)
3305 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3306 ResultPtr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
3313 SunkAddr = ResultPtr;
3315 if (ResultPtr->getType() != I8PtrTy)
3316 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3317 SunkAddr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
3320 if (SunkAddr->getType() != Addr->getType())
3321 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3324 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3325 << *MemoryInst << "\n");
3326 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3327 Value *Result = nullptr;
3329 // Start with the base register. Do this first so that subsequent address
3330 // matching finds it last, which will prevent it from trying to match it
3331 // as the scaled value in case it happens to be a mul. That would be
3332 // problematic if we've sunk a different mul for the scale, because then
3333 // we'd end up sinking both muls.
3334 if (AddrMode.BaseReg) {
3335 Value *V = AddrMode.BaseReg;
3336 if (V->getType()->isPointerTy())
3337 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3338 if (V->getType() != IntPtrTy)
3339 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3343 // Add the scale value.
3344 if (AddrMode.Scale) {
3345 Value *V = AddrMode.ScaledReg;
3346 if (V->getType() == IntPtrTy) {
3348 } else if (V->getType()->isPointerTy()) {
3349 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3350 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3351 cast<IntegerType>(V->getType())->getBitWidth()) {
3352 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3354 // It is only safe to sign extend the BaseReg if we know that the math
3355 // required to create it did not overflow before we extend it. Since
3356 // the original IR value was tossed in favor of a constant back when
3357 // the AddrMode was created we need to bail out gracefully if widths
3358 // do not match instead of extending it.
3359 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3360 if (I && (Result != AddrMode.BaseReg))
3361 I->eraseFromParent();
3364 if (AddrMode.Scale != 1)
3365 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3368 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3373 // Add in the BaseGV if present.
3374 if (AddrMode.BaseGV) {
3375 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3377 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3382 // Add in the Base Offset if present.
3383 if (AddrMode.BaseOffs) {
3384 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3386 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3392 SunkAddr = Constant::getNullValue(Addr->getType());
3394 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3397 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3399 // If we have no uses, recursively delete the value and all dead instructions
3401 if (Repl->use_empty()) {
3402 // This can cause recursive deletion, which can invalidate our iterator.
3403 // Use a WeakVH to hold onto it in case this happens.
3404 WeakVH IterHandle(CurInstIterator);
3405 BasicBlock *BB = CurInstIterator->getParent();
3407 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3409 if (IterHandle != CurInstIterator) {
3410 // If the iterator instruction was recursively deleted, start over at the
3411 // start of the block.
3412 CurInstIterator = BB->begin();
3420 /// OptimizeInlineAsmInst - If there are any memory operands, use
3421 /// OptimizeMemoryInst to sink their address computing into the block when
3422 /// possible / profitable.
3423 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
3424 bool MadeChange = false;
3426 const TargetRegisterInfo *TRI =
3427 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3428 TargetLowering::AsmOperandInfoVector
3429 TargetConstraints = TLI->ParseConstraints(TRI, CS);
3431 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3432 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3434 // Compute the constraint code and ConstraintType to use.
3435 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3437 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3438 OpInfo.isIndirect) {
3439 Value *OpVal = CS->getArgOperand(ArgNo++);
3440 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
3441 } else if (OpInfo.Type == InlineAsm::isInput)
3448 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3449 /// sign extensions.
3450 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3451 assert(!Inst->use_empty() && "Input must have at least one use");
3452 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3453 bool IsSExt = isa<SExtInst>(FirstUser);
3454 Type *ExtTy = FirstUser->getType();
3455 for (const User *U : Inst->users()) {
3456 const Instruction *UI = cast<Instruction>(U);
3457 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3459 Type *CurTy = UI->getType();
3460 // Same input and output types: Same instruction after CSE.
3464 // If IsSExt is true, we are in this situation:
3466 // b = sext ty1 a to ty2
3467 // c = sext ty1 a to ty3
3468 // Assuming ty2 is shorter than ty3, this could be turned into:
3470 // b = sext ty1 a to ty2
3471 // c = sext ty2 b to ty3
3472 // However, the last sext is not free.
3476 // This is a ZExt, maybe this is free to extend from one type to another.
3477 // In that case, we would not account for a different use.
3480 if (ExtTy->getScalarType()->getIntegerBitWidth() >
3481 CurTy->getScalarType()->getIntegerBitWidth()) {
3489 if (!TLI.isZExtFree(NarrowTy, LargeTy))
3492 // All uses are the same or can be derived from one another for free.
3496 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
3497 /// load instruction.
3498 /// If an ext(load) can be formed, it is returned via \p LI for the load
3499 /// and \p Inst for the extension.
3500 /// Otherwise LI == nullptr and Inst == nullptr.
3501 /// When some promotion happened, \p TPT contains the proper state to
3504 /// \return true when promoting was necessary to expose the ext(load)
3505 /// opportunity, false otherwise.
3509 /// %ld = load i32* %addr
3510 /// %add = add nuw i32 %ld, 4
3511 /// %zext = zext i32 %add to i64
3515 /// %ld = load i32* %addr
3516 /// %zext = zext i32 %ld to i64
3517 /// %add = add nuw i64 %zext, 4
3519 /// Thanks to the promotion, we can match zext(load i32*) to i64.
3520 bool CodeGenPrepare::ExtLdPromotion(TypePromotionTransaction &TPT,
3521 LoadInst *&LI, Instruction *&Inst,
3522 const SmallVectorImpl<Instruction *> &Exts,
3523 unsigned CreatedInstsCost = 0) {
3524 // Iterate over all the extensions to see if one form an ext(load).
3525 for (auto I : Exts) {
3526 // Check if we directly have ext(load).
3527 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
3529 // No promotion happened here.
3532 // Check whether or not we want to do any promotion.
3533 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
3535 // Get the action to perform the promotion.
3536 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
3537 I, InsertedTruncsSet, *TLI, PromotedInsts);
3538 // Check if we can promote.
3541 // Save the current state.
3542 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3543 TPT.getRestorationPoint();
3544 SmallVector<Instruction *, 4> NewExts;
3545 unsigned NewCreatedInstsCost = 0;
3546 unsigned ExtCost = !TLI->isExtFree(I);
3548 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
3549 &NewExts, nullptr, *TLI);
3550 assert(PromotedVal &&
3551 "TypePromotionHelper should have filtered out those cases");
3553 // We would be able to merge only one extension in a load.
3554 // Therefore, if we have more than 1 new extension we heuristically
3555 // cut this search path, because it means we degrade the code quality.
3556 // With exactly 2, the transformation is neutral, because we will merge
3557 // one extension but leave one. However, we optimistically keep going,
3558 // because the new extension may be removed too.
3559 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
3560 TotalCreatedInstsCost -= ExtCost;
3561 if (!StressExtLdPromotion &&
3562 (TotalCreatedInstsCost > 1 ||
3563 !isPromotedInstructionLegal(*TLI, PromotedVal))) {
3564 // The promotion is not profitable, rollback to the previous state.
3565 TPT.rollback(LastKnownGood);
3568 // The promotion is profitable.
3569 // Check if it exposes an ext(load).
3570 (void)ExtLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
3571 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
3572 // If we have created a new extension, i.e., now we have two
3573 // extensions. We must make sure one of them is merged with
3574 // the load, otherwise we may degrade the code quality.
3575 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
3576 // Promotion happened.
3578 // If this does not help to expose an ext(load) then, rollback.
3579 TPT.rollback(LastKnownGood);
3581 // None of the extension can form an ext(load).
3587 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
3588 /// basic block as the load, unless conditions are unfavorable. This allows
3589 /// SelectionDAG to fold the extend into the load.
3590 /// \p I[in/out] the extension may be modified during the process if some
3591 /// promotions apply.
3593 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *&I) {
3594 // Try to promote a chain of computation if it allows to form
3595 // an extended load.
3596 TypePromotionTransaction TPT;
3597 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3598 TPT.getRestorationPoint();
3599 SmallVector<Instruction *, 1> Exts;
3601 // Look for a load being extended.
3602 LoadInst *LI = nullptr;
3603 Instruction *OldExt = I;
3604 bool HasPromoted = ExtLdPromotion(TPT, LI, I, Exts);
3606 assert(!HasPromoted && !LI && "If we did not match any load instruction "
3607 "the code must remain the same");
3612 // If they're already in the same block, there's nothing to do.
3613 // Make the cheap checks first if we did not promote.
3614 // If we promoted, we need to check if it is indeed profitable.
3615 if (!HasPromoted && LI->getParent() == I->getParent())
3618 EVT VT = TLI->getValueType(I->getType());
3619 EVT LoadVT = TLI->getValueType(LI->getType());
3621 // If the load has other users and the truncate is not free, this probably
3622 // isn't worthwhile.
3623 if (!LI->hasOneUse() && TLI &&
3624 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
3625 !TLI->isTruncateFree(I->getType(), LI->getType())) {
3627 TPT.rollback(LastKnownGood);
3631 // Check whether the target supports casts folded into loads.
3633 if (isa<ZExtInst>(I))
3634 LType = ISD::ZEXTLOAD;
3636 assert(isa<SExtInst>(I) && "Unexpected ext type!");
3637 LType = ISD::SEXTLOAD;
3639 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
3641 TPT.rollback(LastKnownGood);
3645 // Move the extend into the same block as the load, so that SelectionDAG
3648 I->removeFromParent();
3654 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
3655 BasicBlock *DefBB = I->getParent();
3657 // If the result of a {s|z}ext and its source are both live out, rewrite all
3658 // other uses of the source with result of extension.
3659 Value *Src = I->getOperand(0);
3660 if (Src->hasOneUse())
3663 // Only do this xform if truncating is free.
3664 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
3667 // Only safe to perform the optimization if the source is also defined in
3669 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
3672 bool DefIsLiveOut = false;
3673 for (User *U : I->users()) {
3674 Instruction *UI = cast<Instruction>(U);
3676 // Figure out which BB this ext is used in.
3677 BasicBlock *UserBB = UI->getParent();
3678 if (UserBB == DefBB) continue;
3679 DefIsLiveOut = true;
3685 // Make sure none of the uses are PHI nodes.
3686 for (User *U : Src->users()) {
3687 Instruction *UI = cast<Instruction>(U);
3688 BasicBlock *UserBB = UI->getParent();
3689 if (UserBB == DefBB) continue;
3690 // Be conservative. We don't want this xform to end up introducing
3691 // reloads just before load / store instructions.
3692 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
3696 // InsertedTruncs - Only insert one trunc in each block once.
3697 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
3699 bool MadeChange = false;
3700 for (Use &U : Src->uses()) {
3701 Instruction *User = cast<Instruction>(U.getUser());
3703 // Figure out which BB this ext is used in.
3704 BasicBlock *UserBB = User->getParent();
3705 if (UserBB == DefBB) continue;
3707 // Both src and def are live in this block. Rewrite the use.
3708 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
3710 if (!InsertedTrunc) {
3711 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3712 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
3713 InsertedTruncsSet.insert(InsertedTrunc);
3716 // Replace a use of the {s|z}ext source with a use of the result.
3725 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
3726 /// turned into an explicit branch.
3727 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
3728 // FIXME: This should use the same heuristics as IfConversion to determine
3729 // whether a select is better represented as a branch. This requires that
3730 // branch probability metadata is preserved for the select, which is not the
3733 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3735 // If the branch is predicted right, an out of order CPU can avoid blocking on
3736 // the compare. Emit cmovs on compares with a memory operand as branches to
3737 // avoid stalls on the load from memory. If the compare has more than one use
3738 // there's probably another cmov or setcc around so it's not worth emitting a
3743 Value *CmpOp0 = Cmp->getOperand(0);
3744 Value *CmpOp1 = Cmp->getOperand(1);
3746 // We check that the memory operand has one use to avoid uses of the loaded
3747 // value directly after the compare, making branches unprofitable.
3748 return Cmp->hasOneUse() &&
3749 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3750 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
3754 /// If we have a SelectInst that will likely profit from branch prediction,
3755 /// turn it into a branch.
3756 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
3757 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3759 // Can we convert the 'select' to CF ?
3760 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3763 TargetLowering::SelectSupportKind SelectKind;
3765 SelectKind = TargetLowering::VectorMaskSelect;
3766 else if (SI->getType()->isVectorTy())
3767 SelectKind = TargetLowering::ScalarCondVectorVal;
3769 SelectKind = TargetLowering::ScalarValSelect;
3771 // Do we have efficient codegen support for this kind of 'selects' ?
3772 if (TLI->isSelectSupported(SelectKind)) {
3773 // We have efficient codegen support for the select instruction.
3774 // Check if it is profitable to keep this 'select'.
3775 if (!TLI->isPredictableSelectExpensive() ||
3776 !isFormingBranchFromSelectProfitable(SI))
3782 // First, we split the block containing the select into 2 blocks.
3783 BasicBlock *StartBlock = SI->getParent();
3784 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3785 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3787 // Create a new block serving as the landing pad for the branch.
3788 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3789 NextBlock->getParent(), NextBlock);
3791 // Move the unconditional branch from the block with the select in it into our
3792 // landing pad block.
3793 StartBlock->getTerminator()->eraseFromParent();
3794 BranchInst::Create(NextBlock, SmallBlock);
3796 // Insert the real conditional branch based on the original condition.
3797 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3799 // The select itself is replaced with a PHI Node.
3800 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3802 PN->addIncoming(SI->getTrueValue(), StartBlock);
3803 PN->addIncoming(SI->getFalseValue(), SmallBlock);
3804 SI->replaceAllUsesWith(PN);
3805 SI->eraseFromParent();
3807 // Instruct OptimizeBlock to skip to the next block.
3808 CurInstIterator = StartBlock->end();
3809 ++NumSelectsExpanded;
3813 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3814 SmallVector<int, 16> Mask(SVI->getShuffleMask());
3816 for (unsigned i = 0; i < Mask.size(); ++i) {
3817 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3819 SplatElem = Mask[i];
3825 /// Some targets have expensive vector shifts if the lanes aren't all the same
3826 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3827 /// it's often worth sinking a shufflevector splat down to its use so that
3828 /// codegen can spot all lanes are identical.
3829 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3830 BasicBlock *DefBB = SVI->getParent();
3832 // Only do this xform if variable vector shifts are particularly expensive.
3833 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3836 // We only expect better codegen by sinking a shuffle if we can recognise a
3838 if (!isBroadcastShuffle(SVI))
3841 // InsertedShuffles - Only insert a shuffle in each block once.
3842 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3844 bool MadeChange = false;
3845 for (User *U : SVI->users()) {
3846 Instruction *UI = cast<Instruction>(U);
3848 // Figure out which BB this ext is used in.
3849 BasicBlock *UserBB = UI->getParent();
3850 if (UserBB == DefBB) continue;
3852 // For now only apply this when the splat is used by a shift instruction.
3853 if (!UI->isShift()) continue;
3855 // Everything checks out, sink the shuffle if the user's block doesn't
3856 // already have a copy.
3857 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3859 if (!InsertedShuffle) {
3860 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3861 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3863 SVI->getOperand(2), "", InsertPt);
3866 UI->replaceUsesOfWith(SVI, InsertedShuffle);
3870 // If we removed all uses, nuke the shuffle.
3871 if (SVI->use_empty()) {
3872 SVI->eraseFromParent();
3880 /// \brief Helper class to promote a scalar operation to a vector one.
3881 /// This class is used to move downward extractelement transition.
3883 /// a = vector_op <2 x i32>
3884 /// b = extractelement <2 x i32> a, i32 0
3889 /// a = vector_op <2 x i32>
3890 /// c = vector_op a (equivalent to scalar_op on the related lane)
3891 /// * d = extractelement <2 x i32> c, i32 0
3893 /// Assuming both extractelement and store can be combine, we get rid of the
3895 class VectorPromoteHelper {
3896 /// Used to perform some checks on the legality of vector operations.
3897 const TargetLowering &TLI;
3899 /// Used to estimated the cost of the promoted chain.
3900 const TargetTransformInfo &TTI;
3902 /// The transition being moved downwards.
3903 Instruction *Transition;
3904 /// The sequence of instructions to be promoted.
3905 SmallVector<Instruction *, 4> InstsToBePromoted;
3906 /// Cost of combining a store and an extract.
3907 unsigned StoreExtractCombineCost;
3908 /// Instruction that will be combined with the transition.
3909 Instruction *CombineInst;
3911 /// \brief The instruction that represents the current end of the transition.
3912 /// Since we are faking the promotion until we reach the end of the chain
3913 /// of computation, we need a way to get the current end of the transition.
3914 Instruction *getEndOfTransition() const {
3915 if (InstsToBePromoted.empty())
3917 return InstsToBePromoted.back();
3920 /// \brief Return the index of the original value in the transition.
3921 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
3922 /// c, is at index 0.
3923 unsigned getTransitionOriginalValueIdx() const {
3924 assert(isa<ExtractElementInst>(Transition) &&
3925 "Other kind of transitions are not supported yet");
3929 /// \brief Return the index of the index in the transition.
3930 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
3932 unsigned getTransitionIdx() const {
3933 assert(isa<ExtractElementInst>(Transition) &&
3934 "Other kind of transitions are not supported yet");
3938 /// \brief Get the type of the transition.
3939 /// This is the type of the original value.
3940 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
3941 /// transition is <2 x i32>.
3942 Type *getTransitionType() const {
3943 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
3946 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
3947 /// I.e., we have the following sequence:
3948 /// Def = Transition <ty1> a to <ty2>
3949 /// b = ToBePromoted <ty2> Def, ...
3951 /// b = ToBePromoted <ty1> a, ...
3952 /// Def = Transition <ty1> ToBePromoted to <ty2>
3953 void promoteImpl(Instruction *ToBePromoted);
3955 /// \brief Check whether or not it is profitable to promote all the
3956 /// instructions enqueued to be promoted.
3957 bool isProfitableToPromote() {
3958 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
3959 unsigned Index = isa<ConstantInt>(ValIdx)
3960 ? cast<ConstantInt>(ValIdx)->getZExtValue()
3962 Type *PromotedType = getTransitionType();
3964 StoreInst *ST = cast<StoreInst>(CombineInst);
3965 unsigned AS = ST->getPointerAddressSpace();
3966 unsigned Align = ST->getAlignment();
3967 // Check if this store is supported.
3968 if (!TLI.allowsMisalignedMemoryAccesses(
3969 TLI.getValueType(ST->getValueOperand()->getType()), AS, Align)) {
3970 // If this is not supported, there is no way we can combine
3971 // the extract with the store.
3975 // The scalar chain of computation has to pay for the transition
3976 // scalar to vector.
3977 // The vector chain has to account for the combining cost.
3978 uint64_t ScalarCost =
3979 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
3980 uint64_t VectorCost = StoreExtractCombineCost;
3981 for (const auto &Inst : InstsToBePromoted) {
3982 // Compute the cost.
3983 // By construction, all instructions being promoted are arithmetic ones.
3984 // Moreover, one argument is a constant that can be viewed as a splat
3986 Value *Arg0 = Inst->getOperand(0);
3987 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
3988 isa<ConstantFP>(Arg0);
3989 TargetTransformInfo::OperandValueKind Arg0OVK =
3990 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3991 : TargetTransformInfo::OK_AnyValue;
3992 TargetTransformInfo::OperandValueKind Arg1OVK =
3993 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3994 : TargetTransformInfo::OK_AnyValue;
3995 ScalarCost += TTI.getArithmeticInstrCost(
3996 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
3997 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
4000 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
4001 << ScalarCost << "\nVector: " << VectorCost << '\n');
4002 return ScalarCost > VectorCost;
4005 /// \brief Generate a constant vector with \p Val with the same
4006 /// number of elements as the transition.
4007 /// \p UseSplat defines whether or not \p Val should be replicated
4008 /// accross the whole vector.
4009 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
4010 /// otherwise we generate a vector with as many undef as possible:
4011 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
4012 /// used at the index of the extract.
4013 Value *getConstantVector(Constant *Val, bool UseSplat) const {
4014 unsigned ExtractIdx = UINT_MAX;
4016 // If we cannot determine where the constant must be, we have to
4017 // use a splat constant.
4018 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
4019 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
4020 ExtractIdx = CstVal->getSExtValue();
4025 unsigned End = getTransitionType()->getVectorNumElements();
4027 return ConstantVector::getSplat(End, Val);
4029 SmallVector<Constant *, 4> ConstVec;
4030 UndefValue *UndefVal = UndefValue::get(Val->getType());
4031 for (unsigned Idx = 0; Idx != End; ++Idx) {
4032 if (Idx == ExtractIdx)
4033 ConstVec.push_back(Val);
4035 ConstVec.push_back(UndefVal);
4037 return ConstantVector::get(ConstVec);
4040 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
4041 /// in \p Use can trigger undefined behavior.
4042 static bool canCauseUndefinedBehavior(const Instruction *Use,
4043 unsigned OperandIdx) {
4044 // This is not safe to introduce undef when the operand is on
4045 // the right hand side of a division-like instruction.
4046 if (OperandIdx != 1)
4048 switch (Use->getOpcode()) {
4051 case Instruction::SDiv:
4052 case Instruction::UDiv:
4053 case Instruction::SRem:
4054 case Instruction::URem:
4056 case Instruction::FDiv:
4057 case Instruction::FRem:
4058 return !Use->hasNoNaNs();
4060 llvm_unreachable(nullptr);
4064 VectorPromoteHelper(const TargetLowering &TLI, const TargetTransformInfo &TTI,
4065 Instruction *Transition, unsigned CombineCost)
4066 : TLI(TLI), TTI(TTI), Transition(Transition),
4067 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4068 assert(Transition && "Do not know how to promote null");
4071 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4072 bool canPromote(const Instruction *ToBePromoted) const {
4073 // We could support CastInst too.
4074 return isa<BinaryOperator>(ToBePromoted);
4077 /// \brief Check if it is profitable to promote \p ToBePromoted
4078 /// by moving downward the transition through.
4079 bool shouldPromote(const Instruction *ToBePromoted) const {
4080 // Promote only if all the operands can be statically expanded.
4081 // Indeed, we do not want to introduce any new kind of transitions.
4082 for (const Use &U : ToBePromoted->operands()) {
4083 const Value *Val = U.get();
4084 if (Val == getEndOfTransition()) {
4085 // If the use is a division and the transition is on the rhs,
4086 // we cannot promote the operation, otherwise we may create a
4087 // division by zero.
4088 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4092 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4093 !isa<ConstantFP>(Val))
4096 // Check that the resulting operation is legal.
4097 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4100 return StressStoreExtract ||
4101 TLI.isOperationLegalOrCustom(
4102 ISDOpcode, TLI.getValueType(getTransitionType(), true));
4105 /// \brief Check whether or not \p Use can be combined
4106 /// with the transition.
4107 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4108 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4110 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4111 void enqueueForPromotion(Instruction *ToBePromoted) {
4112 InstsToBePromoted.push_back(ToBePromoted);
4115 /// \brief Set the instruction that will be combined with the transition.
4116 void recordCombineInstruction(Instruction *ToBeCombined) {
4117 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4118 CombineInst = ToBeCombined;
4121 /// \brief Promote all the instructions enqueued for promotion if it is
4123 /// \return True if the promotion happened, false otherwise.
4125 // Check if there is something to promote.
4126 // Right now, if we do not have anything to combine with,
4127 // we assume the promotion is not profitable.
4128 if (InstsToBePromoted.empty() || !CombineInst)
4132 if (!StressStoreExtract && !isProfitableToPromote())
4136 for (auto &ToBePromoted : InstsToBePromoted)
4137 promoteImpl(ToBePromoted);
4138 InstsToBePromoted.clear();
4142 } // End of anonymous namespace.
4144 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4145 // At this point, we know that all the operands of ToBePromoted but Def
4146 // can be statically promoted.
4147 // For Def, we need to use its parameter in ToBePromoted:
4148 // b = ToBePromoted ty1 a
4149 // Def = Transition ty1 b to ty2
4150 // Move the transition down.
4151 // 1. Replace all uses of the promoted operation by the transition.
4152 // = ... b => = ... Def.
4153 assert(ToBePromoted->getType() == Transition->getType() &&
4154 "The type of the result of the transition does not match "
4156 ToBePromoted->replaceAllUsesWith(Transition);
4157 // 2. Update the type of the uses.
4158 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
4159 Type *TransitionTy = getTransitionType();
4160 ToBePromoted->mutateType(TransitionTy);
4161 // 3. Update all the operands of the promoted operation with promoted
4163 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
4164 for (Use &U : ToBePromoted->operands()) {
4165 Value *Val = U.get();
4166 Value *NewVal = nullptr;
4167 if (Val == Transition)
4168 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
4169 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
4170 isa<ConstantFP>(Val)) {
4171 // Use a splat constant if it is not safe to use undef.
4172 NewVal = getConstantVector(
4173 cast<Constant>(Val),
4174 isa<UndefValue>(Val) ||
4175 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
4177 llvm_unreachable("Did you modified shouldPromote and forgot to update "
4179 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
4181 Transition->removeFromParent();
4182 Transition->insertAfter(ToBePromoted);
4183 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
4186 /// Some targets can do store(extractelement) with one instruction.
4187 /// Try to push the extractelement towards the stores when the target
4188 /// has this feature and this is profitable.
4189 bool CodeGenPrepare::OptimizeExtractElementInst(Instruction *Inst) {
4190 unsigned CombineCost = UINT_MAX;
4191 if (DisableStoreExtract || !TLI ||
4192 (!StressStoreExtract &&
4193 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4194 Inst->getOperand(1), CombineCost)))
4197 // At this point we know that Inst is a vector to scalar transition.
4198 // Try to move it down the def-use chain, until:
4199 // - We can combine the transition with its single use
4200 // => we got rid of the transition.
4201 // - We escape the current basic block
4202 // => we would need to check that we are moving it at a cheaper place and
4203 // we do not do that for now.
4204 BasicBlock *Parent = Inst->getParent();
4205 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4206 VectorPromoteHelper VPH(*TLI, *TTI, Inst, CombineCost);
4207 // If the transition has more than one use, assume this is not going to be
4209 while (Inst->hasOneUse()) {
4210 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4211 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4213 if (ToBePromoted->getParent() != Parent) {
4214 DEBUG(dbgs() << "Instruction to promote is in a different block ("
4215 << ToBePromoted->getParent()->getName()
4216 << ") than the transition (" << Parent->getName() << ").\n");
4220 if (VPH.canCombine(ToBePromoted)) {
4221 DEBUG(dbgs() << "Assume " << *Inst << '\n'
4222 << "will be combined with: " << *ToBePromoted << '\n');
4223 VPH.recordCombineInstruction(ToBePromoted);
4224 bool Changed = VPH.promote();
4225 NumStoreExtractExposed += Changed;
4229 DEBUG(dbgs() << "Try promoting.\n");
4230 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4233 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4235 VPH.enqueueForPromotion(ToBePromoted);
4236 Inst = ToBePromoted;
4241 bool CodeGenPrepare::OptimizeInst(Instruction *I, bool& ModifiedDT) {
4242 if (PHINode *P = dyn_cast<PHINode>(I)) {
4243 // It is possible for very late stage optimizations (such as SimplifyCFG)
4244 // to introduce PHI nodes too late to be cleaned up. If we detect such a
4245 // trivial PHI, go ahead and zap it here.
4246 const DataLayout &DL = I->getModule()->getDataLayout();
4247 if (Value *V = SimplifyInstruction(P, DL, TLInfo, DT)) {
4248 P->replaceAllUsesWith(V);
4249 P->eraseFromParent();
4256 if (CastInst *CI = dyn_cast<CastInst>(I)) {
4257 // If the source of the cast is a constant, then this should have
4258 // already been constant folded. The only reason NOT to constant fold
4259 // it is if something (e.g. LSR) was careful to place the constant
4260 // evaluation in a block other than then one that uses it (e.g. to hoist
4261 // the address of globals out of a loop). If this is the case, we don't
4262 // want to forward-subst the cast.
4263 if (isa<Constant>(CI->getOperand(0)))
4266 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
4269 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4270 /// Sink a zext or sext into its user blocks if the target type doesn't
4271 /// fit in one register
4272 if (TLI && TLI->getTypeAction(CI->getContext(),
4273 TLI->getValueType(CI->getType())) ==
4274 TargetLowering::TypeExpandInteger) {
4275 return SinkCast(CI);
4277 bool MadeChange = MoveExtToFormExtLoad(I);
4278 return MadeChange | OptimizeExtUses(I);
4284 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4285 if (!TLI || !TLI->hasMultipleConditionRegisters())
4286 return OptimizeCmpExpression(CI);
4288 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4290 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
4294 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4296 return OptimizeMemoryInst(I, SI->getOperand(1),
4297 SI->getOperand(0)->getType());
4301 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4303 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4304 BinOp->getOpcode() == Instruction::LShr)) {
4305 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4306 if (TLI && CI && TLI->hasExtractBitsInsn())
4307 return OptimizeExtractBits(BinOp, CI, *TLI);
4312 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4313 if (GEPI->hasAllZeroIndices()) {
4314 /// The GEP operand must be a pointer, so must its result -> BitCast
4315 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4316 GEPI->getName(), GEPI);
4317 GEPI->replaceAllUsesWith(NC);
4318 GEPI->eraseFromParent();
4320 OptimizeInst(NC, ModifiedDT);
4326 if (CallInst *CI = dyn_cast<CallInst>(I))
4327 return OptimizeCallInst(CI, ModifiedDT);
4329 if (SelectInst *SI = dyn_cast<SelectInst>(I))
4330 return OptimizeSelectInst(SI);
4332 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
4333 return OptimizeShuffleVectorInst(SVI);
4335 if (isa<ExtractElementInst>(I))
4336 return OptimizeExtractElementInst(I);
4341 // In this pass we look for GEP and cast instructions that are used
4342 // across basic blocks and rewrite them to improve basic-block-at-a-time
4344 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
4346 bool MadeChange = false;
4348 CurInstIterator = BB.begin();
4349 while (CurInstIterator != BB.end()) {
4350 MadeChange |= OptimizeInst(CurInstIterator++, ModifiedDT);
4354 MadeChange |= DupRetToEnableTailCallOpts(&BB);
4359 // llvm.dbg.value is far away from the value then iSel may not be able
4360 // handle it properly. iSel will drop llvm.dbg.value if it can not
4361 // find a node corresponding to the value.
4362 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
4363 bool MadeChange = false;
4364 for (BasicBlock &BB : F) {
4365 Instruction *PrevNonDbgInst = nullptr;
4366 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
4367 Instruction *Insn = BI++;
4368 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
4369 // Leave dbg.values that refer to an alloca alone. These
4370 // instrinsics describe the address of a variable (= the alloca)
4371 // being taken. They should not be moved next to the alloca
4372 // (and to the beginning of the scope), but rather stay close to
4373 // where said address is used.
4374 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
4375 PrevNonDbgInst = Insn;
4379 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
4380 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
4381 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
4382 DVI->removeFromParent();
4383 if (isa<PHINode>(VI))
4384 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
4386 DVI->insertAfter(VI);
4395 // If there is a sequence that branches based on comparing a single bit
4396 // against zero that can be combined into a single instruction, and the
4397 // target supports folding these into a single instruction, sink the
4398 // mask and compare into the branch uses. Do this before OptimizeBlock ->
4399 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
4401 bool CodeGenPrepare::sinkAndCmp(Function &F) {
4402 if (!EnableAndCmpSinking)
4404 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
4406 bool MadeChange = false;
4407 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
4408 BasicBlock *BB = I++;
4410 // Does this BB end with the following?
4411 // %andVal = and %val, #single-bit-set
4412 // %icmpVal = icmp %andResult, 0
4413 // br i1 %cmpVal label %dest1, label %dest2"
4414 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
4415 if (!Brcc || !Brcc->isConditional())
4417 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
4418 if (!Cmp || Cmp->getParent() != BB)
4420 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
4421 if (!Zero || !Zero->isZero())
4423 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
4424 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
4426 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
4427 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
4429 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
4431 // Push the "and; icmp" for any users that are conditional branches.
4432 // Since there can only be one branch use per BB, we don't need to keep
4433 // track of which BBs we insert into.
4434 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
4438 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
4440 if (!BrccUser || !BrccUser->isConditional())
4442 BasicBlock *UserBB = BrccUser->getParent();
4443 if (UserBB == BB) continue;
4444 DEBUG(dbgs() << "found Brcc use\n");
4446 // Sink the "and; icmp" to use.
4448 BinaryOperator *NewAnd =
4449 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
4452 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
4456 DEBUG(BrccUser->getParent()->dump());
4462 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
4463 /// success, or returns false if no or invalid metadata was found.
4464 static bool extractBranchMetadata(BranchInst *BI,
4465 uint64_t &ProbTrue, uint64_t &ProbFalse) {
4466 assert(BI->isConditional() &&
4467 "Looking for probabilities on unconditional branch?");
4468 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
4469 if (!ProfileData || ProfileData->getNumOperands() != 3)
4472 const auto *CITrue =
4473 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
4474 const auto *CIFalse =
4475 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
4476 if (!CITrue || !CIFalse)
4479 ProbTrue = CITrue->getValue().getZExtValue();
4480 ProbFalse = CIFalse->getValue().getZExtValue();
4485 /// \brief Scale down both weights to fit into uint32_t.
4486 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
4487 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
4488 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
4489 NewTrue = NewTrue / Scale;
4490 NewFalse = NewFalse / Scale;
4493 /// \brief Some targets prefer to split a conditional branch like:
4495 /// %0 = icmp ne i32 %a, 0
4496 /// %1 = icmp ne i32 %b, 0
4497 /// %or.cond = or i1 %0, %1
4498 /// br i1 %or.cond, label %TrueBB, label %FalseBB
4500 /// into multiple branch instructions like:
4503 /// %0 = icmp ne i32 %a, 0
4504 /// br i1 %0, label %TrueBB, label %bb2
4506 /// %1 = icmp ne i32 %b, 0
4507 /// br i1 %1, label %TrueBB, label %FalseBB
4509 /// This usually allows instruction selection to do even further optimizations
4510 /// and combine the compare with the branch instruction. Currently this is
4511 /// applied for targets which have "cheap" jump instructions.
4513 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
4515 bool CodeGenPrepare::splitBranchCondition(Function &F) {
4516 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
4519 bool MadeChange = false;
4520 for (auto &BB : F) {
4521 // Does this BB end with the following?
4522 // %cond1 = icmp|fcmp|binary instruction ...
4523 // %cond2 = icmp|fcmp|binary instruction ...
4524 // %cond.or = or|and i1 %cond1, cond2
4525 // br i1 %cond.or label %dest1, label %dest2"
4526 BinaryOperator *LogicOp;
4527 BasicBlock *TBB, *FBB;
4528 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
4532 Value *Cond1, *Cond2;
4533 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
4534 m_OneUse(m_Value(Cond2)))))
4535 Opc = Instruction::And;
4536 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
4537 m_OneUse(m_Value(Cond2)))))
4538 Opc = Instruction::Or;
4542 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
4543 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
4546 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
4549 auto *InsertBefore = std::next(Function::iterator(BB))
4550 .getNodePtrUnchecked();
4551 auto TmpBB = BasicBlock::Create(BB.getContext(),
4552 BB.getName() + ".cond.split",
4553 BB.getParent(), InsertBefore);
4555 // Update original basic block by using the first condition directly by the
4556 // branch instruction and removing the no longer needed and/or instruction.
4557 auto *Br1 = cast<BranchInst>(BB.getTerminator());
4558 Br1->setCondition(Cond1);
4559 LogicOp->eraseFromParent();
4561 // Depending on the conditon we have to either replace the true or the false
4562 // successor of the original branch instruction.
4563 if (Opc == Instruction::And)
4564 Br1->setSuccessor(0, TmpBB);
4566 Br1->setSuccessor(1, TmpBB);
4568 // Fill in the new basic block.
4569 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
4570 if (auto *I = dyn_cast<Instruction>(Cond2)) {
4571 I->removeFromParent();
4572 I->insertBefore(Br2);
4575 // Update PHI nodes in both successors. The original BB needs to be
4576 // replaced in one succesor's PHI nodes, because the branch comes now from
4577 // the newly generated BB (NewBB). In the other successor we need to add one
4578 // incoming edge to the PHI nodes, because both branch instructions target
4579 // now the same successor. Depending on the original branch condition
4580 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
4581 // we perfrom the correct update for the PHI nodes.
4582 // This doesn't change the successor order of the just created branch
4583 // instruction (or any other instruction).
4584 if (Opc == Instruction::Or)
4585 std::swap(TBB, FBB);
4587 // Replace the old BB with the new BB.
4588 for (auto &I : *TBB) {
4589 PHINode *PN = dyn_cast<PHINode>(&I);
4593 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
4594 PN->setIncomingBlock(i, TmpBB);
4597 // Add another incoming edge form the new BB.
4598 for (auto &I : *FBB) {
4599 PHINode *PN = dyn_cast<PHINode>(&I);
4602 auto *Val = PN->getIncomingValueForBlock(&BB);
4603 PN->addIncoming(Val, TmpBB);
4606 // Update the branch weights (from SelectionDAGBuilder::
4607 // FindMergedConditions).
4608 if (Opc == Instruction::Or) {
4609 // Codegen X | Y as:
4618 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
4619 // The requirement is that
4620 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
4621 // = TrueProb for orignal BB.
4622 // Assuming the orignal weights are A and B, one choice is to set BB1's
4623 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
4625 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
4626 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
4627 // TmpBB, but the math is more complicated.
4628 uint64_t TrueWeight, FalseWeight;
4629 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4630 uint64_t NewTrueWeight = TrueWeight;
4631 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
4632 scaleWeights(NewTrueWeight, NewFalseWeight);
4633 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4634 .createBranchWeights(TrueWeight, FalseWeight));
4636 NewTrueWeight = TrueWeight;
4637 NewFalseWeight = 2 * FalseWeight;
4638 scaleWeights(NewTrueWeight, NewFalseWeight);
4639 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4640 .createBranchWeights(TrueWeight, FalseWeight));
4643 // Codegen X & Y as:
4651 // This requires creation of TmpBB after CurBB.
4653 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
4654 // The requirement is that
4655 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
4656 // = FalseProb for orignal BB.
4657 // Assuming the orignal weights are A and B, one choice is to set BB1's
4658 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
4660 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
4661 uint64_t TrueWeight, FalseWeight;
4662 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4663 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
4664 uint64_t NewFalseWeight = FalseWeight;
4665 scaleWeights(NewTrueWeight, NewFalseWeight);
4666 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4667 .createBranchWeights(TrueWeight, FalseWeight));
4669 NewTrueWeight = 2 * TrueWeight;
4670 NewFalseWeight = FalseWeight;
4671 scaleWeights(NewTrueWeight, NewFalseWeight);
4672 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4673 .createBranchWeights(TrueWeight, FalseWeight));
4677 // Request DOM Tree update.
4678 // Note: No point in getting fancy here, since the DT info is never
4679 // available to CodeGenPrepare and the existing update code is broken
4685 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();