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;
128 /// CurInstIterator - As we scan instructions optimizing them, this is the
129 /// next instruction to optimize. Xforms that can invalidate this should
131 BasicBlock::iterator CurInstIterator;
133 /// Keeps track of non-local addresses that have been sunk into a block.
134 /// This allows us to avoid inserting duplicate code for blocks with
135 /// multiple load/stores of the same address.
136 ValueMap<Value*, Value*> SunkAddrs;
138 /// Keeps track of all truncates inserted for the current function.
139 SetOfInstrs InsertedTruncsSet;
140 /// Keeps track of the type of the related instruction before their
141 /// promotion for the current function.
142 InstrToOrigTy PromotedInsts;
144 /// ModifiedDT - If CFG is modified in anyway.
147 /// OptSize - True if optimizing for size.
151 static char ID; // Pass identification, replacement for typeid
152 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
153 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr) {
154 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
156 bool runOnFunction(Function &F) override;
158 const char *getPassName() const override { return "CodeGen Prepare"; }
160 void getAnalysisUsage(AnalysisUsage &AU) const override {
161 AU.addPreserved<DominatorTreeWrapperPass>();
162 AU.addRequired<TargetLibraryInfoWrapperPass>();
163 AU.addRequired<TargetTransformInfoWrapperPass>();
167 bool EliminateFallThrough(Function &F);
168 bool EliminateMostlyEmptyBlocks(Function &F);
169 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
170 void EliminateMostlyEmptyBlock(BasicBlock *BB);
171 bool OptimizeBlock(BasicBlock &BB, bool& ModifiedDT);
172 bool OptimizeInst(Instruction *I, bool& ModifiedDT);
173 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
174 bool OptimizeInlineAsmInst(CallInst *CS);
175 bool OptimizeCallInst(CallInst *CI, bool& ModifiedDT);
176 bool MoveExtToFormExtLoad(Instruction *&I);
177 bool OptimizeExtUses(Instruction *I);
178 bool OptimizeSelectInst(SelectInst *SI);
179 bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
180 bool OptimizeExtractElementInst(Instruction *Inst);
181 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
182 bool PlaceDbgValues(Function &F);
183 bool sinkAndCmp(Function &F);
184 bool ExtLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
186 const SmallVectorImpl<Instruction *> &Exts,
187 unsigned CreatedInstCost);
188 bool splitBranchCondition(Function &F);
189 bool simplifyOffsetableRelocate(Instruction &I);
193 char CodeGenPrepare::ID = 0;
194 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
195 "Optimize for code generation", false, false)
197 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
198 return new CodeGenPrepare(TM);
201 bool CodeGenPrepare::runOnFunction(Function &F) {
202 if (skipOptnoneFunction(F))
205 bool EverMadeChange = false;
206 // Clear per function information.
207 InsertedTruncsSet.clear();
208 PromotedInsts.clear();
212 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
213 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
214 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
215 OptSize = F.hasFnAttribute(Attribute::OptimizeForSize);
217 /// This optimization identifies DIV instructions that can be
218 /// profitably bypassed and carried out with a shorter, faster divide.
219 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
220 const DenseMap<unsigned int, unsigned int> &BypassWidths =
221 TLI->getBypassSlowDivWidths();
222 for (Function::iterator I = F.begin(); I != F.end(); I++)
223 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
226 // Eliminate blocks that contain only PHI nodes and an
227 // unconditional branch.
228 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
230 // llvm.dbg.value is far away from the value then iSel may not be able
231 // handle it properly. iSel will drop llvm.dbg.value if it can not
232 // find a node corresponding to the value.
233 EverMadeChange |= PlaceDbgValues(F);
235 // If there is a mask, compare against zero, and branch that can be combined
236 // into a single target instruction, push the mask and compare into branch
237 // users. Do this before OptimizeBlock -> OptimizeInst ->
238 // OptimizeCmpExpression, which perturbs the pattern being searched for.
239 if (!DisableBranchOpts) {
240 EverMadeChange |= sinkAndCmp(F);
241 EverMadeChange |= splitBranchCondition(F);
244 bool MadeChange = true;
247 for (Function::iterator I = F.begin(); I != F.end(); ) {
248 BasicBlock *BB = I++;
249 bool ModifiedDTOnIteration = false;
250 MadeChange |= OptimizeBlock(*BB, ModifiedDTOnIteration);
252 // Restart BB iteration if the dominator tree of the Function was changed
253 if (ModifiedDTOnIteration)
256 EverMadeChange |= MadeChange;
261 if (!DisableBranchOpts) {
263 SmallPtrSet<BasicBlock*, 8> WorkList;
264 for (BasicBlock &BB : F) {
265 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
266 MadeChange |= ConstantFoldTerminator(&BB, true);
267 if (!MadeChange) continue;
269 for (SmallVectorImpl<BasicBlock*>::iterator
270 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
271 if (pred_begin(*II) == pred_end(*II))
272 WorkList.insert(*II);
275 // Delete the dead blocks and any of their dead successors.
276 MadeChange |= !WorkList.empty();
277 while (!WorkList.empty()) {
278 BasicBlock *BB = *WorkList.begin();
280 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
284 for (SmallVectorImpl<BasicBlock*>::iterator
285 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
286 if (pred_begin(*II) == pred_end(*II))
287 WorkList.insert(*II);
290 // Merge pairs of basic blocks with unconditional branches, connected by
292 if (EverMadeChange || MadeChange)
293 MadeChange |= EliminateFallThrough(F);
295 EverMadeChange |= MadeChange;
298 if (!DisableGCOpts) {
299 SmallVector<Instruction *, 2> Statepoints;
300 for (BasicBlock &BB : F)
301 for (Instruction &I : BB)
303 Statepoints.push_back(&I);
304 for (auto &I : Statepoints)
305 EverMadeChange |= simplifyOffsetableRelocate(*I);
308 return EverMadeChange;
311 /// EliminateFallThrough - Merge basic blocks which are connected
312 /// by a single edge, where one of the basic blocks has a single successor
313 /// pointing to the other basic block, which has a single predecessor.
314 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
315 bool Changed = false;
316 // Scan all of the blocks in the function, except for the entry block.
317 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
318 BasicBlock *BB = I++;
319 // If the destination block has a single pred, then this is a trivial
320 // edge, just collapse it.
321 BasicBlock *SinglePred = BB->getSinglePredecessor();
323 // Don't merge if BB's address is taken.
324 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
326 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
327 if (Term && !Term->isConditional()) {
329 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
330 // Remember if SinglePred was the entry block of the function.
331 // If so, we will need to move BB back to the entry position.
332 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
333 MergeBasicBlockIntoOnlyPred(BB, nullptr);
335 if (isEntry && BB != &BB->getParent()->getEntryBlock())
336 BB->moveBefore(&BB->getParent()->getEntryBlock());
338 // We have erased a block. Update the iterator.
345 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
346 /// debug info directives, and an unconditional branch. Passes before isel
347 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
348 /// isel. Start by eliminating these blocks so we can split them the way we
350 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
351 bool MadeChange = false;
352 // Note that this intentionally skips the entry block.
353 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
354 BasicBlock *BB = I++;
356 // If this block doesn't end with an uncond branch, ignore it.
357 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
358 if (!BI || !BI->isUnconditional())
361 // If the instruction before the branch (skipping debug info) isn't a phi
362 // node, then other stuff is happening here.
363 BasicBlock::iterator BBI = BI;
364 if (BBI != BB->begin()) {
366 while (isa<DbgInfoIntrinsic>(BBI)) {
367 if (BBI == BB->begin())
371 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
375 // Do not break infinite loops.
376 BasicBlock *DestBB = BI->getSuccessor(0);
380 if (!CanMergeBlocks(BB, DestBB))
383 EliminateMostlyEmptyBlock(BB);
389 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
390 /// single uncond branch between them, and BB contains no other non-phi
392 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
393 const BasicBlock *DestBB) const {
394 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
395 // the successor. If there are more complex condition (e.g. preheaders),
396 // don't mess around with them.
397 BasicBlock::const_iterator BBI = BB->begin();
398 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
399 for (const User *U : PN->users()) {
400 const Instruction *UI = cast<Instruction>(U);
401 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
403 // If User is inside DestBB block and it is a PHINode then check
404 // incoming value. If incoming value is not from BB then this is
405 // a complex condition (e.g. preheaders) we want to avoid here.
406 if (UI->getParent() == DestBB) {
407 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
408 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
409 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
410 if (Insn && Insn->getParent() == BB &&
411 Insn->getParent() != UPN->getIncomingBlock(I))
418 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
419 // and DestBB may have conflicting incoming values for the block. If so, we
420 // can't merge the block.
421 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
422 if (!DestBBPN) return true; // no conflict.
424 // Collect the preds of BB.
425 SmallPtrSet<const BasicBlock*, 16> BBPreds;
426 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
427 // It is faster to get preds from a PHI than with pred_iterator.
428 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
429 BBPreds.insert(BBPN->getIncomingBlock(i));
431 BBPreds.insert(pred_begin(BB), pred_end(BB));
434 // Walk the preds of DestBB.
435 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
436 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
437 if (BBPreds.count(Pred)) { // Common predecessor?
438 BBI = DestBB->begin();
439 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
440 const Value *V1 = PN->getIncomingValueForBlock(Pred);
441 const Value *V2 = PN->getIncomingValueForBlock(BB);
443 // If V2 is a phi node in BB, look up what the mapped value will be.
444 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
445 if (V2PN->getParent() == BB)
446 V2 = V2PN->getIncomingValueForBlock(Pred);
448 // If there is a conflict, bail out.
449 if (V1 != V2) return false;
458 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
459 /// an unconditional branch in it.
460 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
461 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
462 BasicBlock *DestBB = BI->getSuccessor(0);
464 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
466 // If the destination block has a single pred, then this is a trivial edge,
468 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
469 if (SinglePred != DestBB) {
470 // Remember if SinglePred was the entry block of the function. If so, we
471 // will need to move BB back to the entry position.
472 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
473 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
475 if (isEntry && BB != &BB->getParent()->getEntryBlock())
476 BB->moveBefore(&BB->getParent()->getEntryBlock());
478 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
483 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
484 // to handle the new incoming edges it is about to have.
486 for (BasicBlock::iterator BBI = DestBB->begin();
487 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
488 // Remove the incoming value for BB, and remember it.
489 Value *InVal = PN->removeIncomingValue(BB, false);
491 // Two options: either the InVal is a phi node defined in BB or it is some
492 // value that dominates BB.
493 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
494 if (InValPhi && InValPhi->getParent() == BB) {
495 // Add all of the input values of the input PHI as inputs of this phi.
496 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
497 PN->addIncoming(InValPhi->getIncomingValue(i),
498 InValPhi->getIncomingBlock(i));
500 // Otherwise, add one instance of the dominating value for each edge that
501 // we will be adding.
502 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
503 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
504 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
506 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
507 PN->addIncoming(InVal, *PI);
512 // The PHIs are now updated, change everything that refers to BB to use
513 // DestBB and remove BB.
514 BB->replaceAllUsesWith(DestBB);
515 BB->eraseFromParent();
518 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
521 // Computes a map of base pointer relocation instructions to corresponding
522 // derived pointer relocation instructions given a vector of all relocate calls
523 static void computeBaseDerivedRelocateMap(
524 const SmallVectorImpl<User *> &AllRelocateCalls,
525 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
527 // Collect information in two maps: one primarily for locating the base object
528 // while filling the second map; the second map is the final structure holding
529 // a mapping between Base and corresponding Derived relocate calls
530 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
531 for (auto &U : AllRelocateCalls) {
532 GCRelocateOperands ThisRelocate(U);
533 IntrinsicInst *I = cast<IntrinsicInst>(U);
534 auto K = std::make_pair(ThisRelocate.basePtrIndex(),
535 ThisRelocate.derivedPtrIndex());
536 RelocateIdxMap.insert(std::make_pair(K, I));
538 for (auto &Item : RelocateIdxMap) {
539 std::pair<unsigned, unsigned> Key = Item.first;
540 if (Key.first == Key.second)
541 // Base relocation: nothing to insert
544 IntrinsicInst *I = Item.second;
545 auto BaseKey = std::make_pair(Key.first, Key.first);
547 // We're iterating over RelocateIdxMap so we cannot modify it.
548 auto MaybeBase = RelocateIdxMap.find(BaseKey);
549 if (MaybeBase == RelocateIdxMap.end())
550 // TODO: We might want to insert a new base object relocate and gep off
551 // that, if there are enough derived object relocates.
554 RelocateInstMap[MaybeBase->second].push_back(I);
558 // Accepts a GEP and extracts the operands into a vector provided they're all
559 // small integer constants
560 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
561 SmallVectorImpl<Value *> &OffsetV) {
562 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
563 // Only accept small constant integer operands
564 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
565 if (!Op || Op->getZExtValue() > 20)
569 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
570 OffsetV.push_back(GEP->getOperand(i));
574 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
575 // replace, computes a replacement, and affects it.
577 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
578 const SmallVectorImpl<IntrinsicInst *> &Targets) {
579 bool MadeChange = false;
580 for (auto &ToReplace : Targets) {
581 GCRelocateOperands MasterRelocate(RelocatedBase);
582 GCRelocateOperands ThisRelocate(ToReplace);
584 assert(ThisRelocate.basePtrIndex() == MasterRelocate.basePtrIndex() &&
585 "Not relocating a derived object of the original base object");
586 if (ThisRelocate.basePtrIndex() == ThisRelocate.derivedPtrIndex()) {
587 // A duplicate relocate call. TODO: coalesce duplicates.
591 Value *Base = ThisRelocate.basePtr();
592 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.derivedPtr());
593 if (!Derived || Derived->getPointerOperand() != Base)
596 SmallVector<Value *, 2> OffsetV;
597 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
600 // Create a Builder and replace the target callsite with a gep
601 IRBuilder<> Builder(ToReplace);
602 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
603 Value *Replacement = Builder.CreateGEP(
604 Derived->getSourceElementType(), RelocatedBase, makeArrayRef(OffsetV));
605 Instruction *ReplacementInst = cast<Instruction>(Replacement);
606 ReplacementInst->removeFromParent();
607 ReplacementInst->insertAfter(RelocatedBase);
608 Replacement->takeName(ToReplace);
609 ToReplace->replaceAllUsesWith(Replacement);
610 ToReplace->eraseFromParent();
620 // %ptr = gep %base + 15
621 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
622 // %base' = relocate(%tok, i32 4, i32 4)
623 // %ptr' = relocate(%tok, i32 4, i32 5)
629 // %ptr = gep %base + 15
630 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
631 // %base' = gc.relocate(%tok, i32 4, i32 4)
632 // %ptr' = gep %base' + 15
634 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
635 bool MadeChange = false;
636 SmallVector<User *, 2> AllRelocateCalls;
638 for (auto *U : I.users())
639 if (isGCRelocate(dyn_cast<Instruction>(U)))
640 // Collect all the relocate calls associated with a statepoint
641 AllRelocateCalls.push_back(U);
643 // We need atleast one base pointer relocation + one derived pointer
644 // relocation to mangle
645 if (AllRelocateCalls.size() < 2)
648 // RelocateInstMap is a mapping from the base relocate instruction to the
649 // corresponding derived relocate instructions
650 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
651 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
652 if (RelocateInstMap.empty())
655 for (auto &Item : RelocateInstMap)
656 // Item.first is the RelocatedBase to offset against
657 // Item.second is the vector of Targets to replace
658 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
662 /// SinkCast - Sink the specified cast instruction into its user blocks
663 static bool SinkCast(CastInst *CI) {
664 BasicBlock *DefBB = CI->getParent();
666 /// InsertedCasts - Only insert a cast in each block once.
667 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
669 bool MadeChange = false;
670 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
672 Use &TheUse = UI.getUse();
673 Instruction *User = cast<Instruction>(*UI);
675 // Figure out which BB this cast is used in. For PHI's this is the
676 // appropriate predecessor block.
677 BasicBlock *UserBB = User->getParent();
678 if (PHINode *PN = dyn_cast<PHINode>(User)) {
679 UserBB = PN->getIncomingBlock(TheUse);
682 // Preincrement use iterator so we don't invalidate it.
685 // If this user is in the same block as the cast, don't change the cast.
686 if (UserBB == DefBB) continue;
688 // If we have already inserted a cast into this block, use it.
689 CastInst *&InsertedCast = InsertedCasts[UserBB];
692 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
694 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
699 // Replace a use of the cast with a use of the new cast.
700 TheUse = InsertedCast;
704 // If we removed all uses, nuke the cast.
705 if (CI->use_empty()) {
706 CI->eraseFromParent();
713 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
714 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
715 /// sink it into user blocks to reduce the number of virtual
716 /// registers that must be created and coalesced.
718 /// Return true if any changes are made.
720 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
721 // If this is a noop copy,
722 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
723 EVT DstVT = TLI.getValueType(CI->getType());
725 // This is an fp<->int conversion?
726 if (SrcVT.isInteger() != DstVT.isInteger())
729 // If this is an extension, it will be a zero or sign extension, which
731 if (SrcVT.bitsLT(DstVT)) return false;
733 // If these values will be promoted, find out what they will be promoted
734 // to. This helps us consider truncates on PPC as noop copies when they
736 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
737 TargetLowering::TypePromoteInteger)
738 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
739 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
740 TargetLowering::TypePromoteInteger)
741 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
743 // If, after promotion, these are the same types, this is a noop copy.
750 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
751 /// the number of virtual registers that must be created and coalesced. This is
752 /// a clear win except on targets with multiple condition code registers
753 /// (PowerPC), where it might lose; some adjustment may be wanted there.
755 /// Return true if any changes are made.
756 static bool OptimizeCmpExpression(CmpInst *CI) {
757 BasicBlock *DefBB = CI->getParent();
759 /// InsertedCmp - Only insert a cmp in each block once.
760 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
762 bool MadeChange = false;
763 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
765 Use &TheUse = UI.getUse();
766 Instruction *User = cast<Instruction>(*UI);
768 // Preincrement use iterator so we don't invalidate it.
771 // Don't bother for PHI nodes.
772 if (isa<PHINode>(User))
775 // Figure out which BB this cmp is used in.
776 BasicBlock *UserBB = User->getParent();
778 // If this user is in the same block as the cmp, don't change the cmp.
779 if (UserBB == DefBB) continue;
781 // If we have already inserted a cmp into this block, use it.
782 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
785 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
787 CmpInst::Create(CI->getOpcode(),
788 CI->getPredicate(), CI->getOperand(0),
789 CI->getOperand(1), "", InsertPt);
793 // Replace a use of the cmp with a use of the new cmp.
794 TheUse = InsertedCmp;
798 // If we removed all uses, nuke the cmp.
800 CI->eraseFromParent();
805 /// isExtractBitsCandidateUse - Check if the candidates could
806 /// be combined with shift instruction, which includes:
807 /// 1. Truncate instruction
808 /// 2. And instruction and the imm is a mask of the low bits:
809 /// imm & (imm+1) == 0
810 static bool isExtractBitsCandidateUse(Instruction *User) {
811 if (!isa<TruncInst>(User)) {
812 if (User->getOpcode() != Instruction::And ||
813 !isa<ConstantInt>(User->getOperand(1)))
816 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
818 if ((Cimm & (Cimm + 1)).getBoolValue())
824 /// SinkShiftAndTruncate - sink both shift and truncate instruction
825 /// to the use of truncate's BB.
827 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
828 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
829 const TargetLowering &TLI) {
830 BasicBlock *UserBB = User->getParent();
831 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
832 TruncInst *TruncI = dyn_cast<TruncInst>(User);
833 bool MadeChange = false;
835 for (Value::user_iterator TruncUI = TruncI->user_begin(),
836 TruncE = TruncI->user_end();
837 TruncUI != TruncE;) {
839 Use &TruncTheUse = TruncUI.getUse();
840 Instruction *TruncUser = cast<Instruction>(*TruncUI);
841 // Preincrement use iterator so we don't invalidate it.
845 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
849 // If the use is actually a legal node, there will not be an
850 // implicit truncate.
851 // FIXME: always querying the result type is just an
852 // approximation; some nodes' legality is determined by the
853 // operand or other means. There's no good way to find out though.
854 if (TLI.isOperationLegalOrCustom(
855 ISDOpcode, TLI.getValueType(TruncUser->getType(), true)))
858 // Don't bother for PHI nodes.
859 if (isa<PHINode>(TruncUser))
862 BasicBlock *TruncUserBB = TruncUser->getParent();
864 if (UserBB == TruncUserBB)
867 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
868 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
870 if (!InsertedShift && !InsertedTrunc) {
871 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
873 if (ShiftI->getOpcode() == Instruction::AShr)
875 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
878 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
881 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
884 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
885 TruncI->getType(), "", TruncInsertPt);
889 TruncTheUse = InsertedTrunc;
895 /// OptimizeExtractBits - sink the shift *right* instruction into user blocks if
896 /// the uses could potentially be combined with this shift instruction and
897 /// generate BitExtract instruction. It will only be applied if the architecture
898 /// supports BitExtract instruction. Here is an example:
900 /// %x.extract.shift = lshr i64 %arg1, 32
902 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
906 /// %x.extract.shift.1 = lshr i64 %arg1, 32
907 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
909 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
911 /// Return true if any changes are made.
912 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
913 const TargetLowering &TLI) {
914 BasicBlock *DefBB = ShiftI->getParent();
916 /// Only insert instructions in each block once.
917 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
919 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(ShiftI->getType()));
921 bool MadeChange = false;
922 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
924 Use &TheUse = UI.getUse();
925 Instruction *User = cast<Instruction>(*UI);
926 // Preincrement use iterator so we don't invalidate it.
929 // Don't bother for PHI nodes.
930 if (isa<PHINode>(User))
933 if (!isExtractBitsCandidateUse(User))
936 BasicBlock *UserBB = User->getParent();
938 if (UserBB == DefBB) {
939 // If the shift and truncate instruction are in the same BB. The use of
940 // the truncate(TruncUse) may still introduce another truncate if not
941 // legal. In this case, we would like to sink both shift and truncate
942 // instruction to the BB of TruncUse.
945 // i64 shift.result = lshr i64 opnd, imm
946 // trunc.result = trunc shift.result to i16
949 // ----> We will have an implicit truncate here if the architecture does
950 // not have i16 compare.
951 // cmp i16 trunc.result, opnd2
953 if (isa<TruncInst>(User) && shiftIsLegal
954 // If the type of the truncate is legal, no trucate will be
955 // introduced in other basic blocks.
956 && (!TLI.isTypeLegal(TLI.getValueType(User->getType()))))
958 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI);
962 // If we have already inserted a shift into this block, use it.
963 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
965 if (!InsertedShift) {
966 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
968 if (ShiftI->getOpcode() == Instruction::AShr)
970 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
973 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
978 // Replace a use of the shift with a use of the new shift.
979 TheUse = InsertedShift;
982 // If we removed all uses, nuke the shift.
983 if (ShiftI->use_empty())
984 ShiftI->eraseFromParent();
989 // ScalarizeMaskedLoad() translates masked load intrinsic, like
990 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
991 // <16 x i1> %mask, <16 x i32> %passthru)
992 // to a chain of basic blocks, whith loading element one-by-one if
993 // the appropriate mask bit is set
995 // %1 = bitcast i8* %addr to i32*
996 // %2 = extractelement <16 x i1> %mask, i32 0
997 // %3 = icmp eq i1 %2, true
998 // br i1 %3, label %cond.load, label %else
1000 //cond.load: ; preds = %0
1001 // %4 = getelementptr i32* %1, i32 0
1002 // %5 = load i32* %4
1003 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1006 //else: ; preds = %0, %cond.load
1007 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1008 // %7 = extractelement <16 x i1> %mask, i32 1
1009 // %8 = icmp eq i1 %7, true
1010 // br i1 %8, label %cond.load1, label %else2
1012 //cond.load1: ; preds = %else
1013 // %9 = getelementptr i32* %1, i32 1
1014 // %10 = load i32* %9
1015 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1018 //else2: ; preds = %else, %cond.load1
1019 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1020 // %12 = extractelement <16 x i1> %mask, i32 2
1021 // %13 = icmp eq i1 %12, true
1022 // br i1 %13, label %cond.load4, label %else5
1024 static void ScalarizeMaskedLoad(CallInst *CI) {
1025 Value *Ptr = CI->getArgOperand(0);
1026 Value *Src0 = CI->getArgOperand(3);
1027 Value *Mask = CI->getArgOperand(2);
1028 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1029 Type *EltTy = VecType->getElementType();
1031 assert(VecType && "Unexpected return type of masked load intrinsic");
1033 IRBuilder<> Builder(CI->getContext());
1034 Instruction *InsertPt = CI;
1035 BasicBlock *IfBlock = CI->getParent();
1036 BasicBlock *CondBlock = nullptr;
1037 BasicBlock *PrevIfBlock = CI->getParent();
1038 Builder.SetInsertPoint(InsertPt);
1040 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1042 // Bitcast %addr fron i8* to EltTy*
1044 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1045 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1046 Value *UndefVal = UndefValue::get(VecType);
1048 // The result vector
1049 Value *VResult = UndefVal;
1051 PHINode *Phi = nullptr;
1052 Value *PrevPhi = UndefVal;
1054 unsigned VectorWidth = VecType->getNumElements();
1055 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1057 // Fill the "else" block, created in the previous iteration
1059 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1060 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1061 // %to_load = icmp eq i1 %mask_1, true
1062 // br i1 %to_load, label %cond.load, label %else
1065 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1066 Phi->addIncoming(VResult, CondBlock);
1067 Phi->addIncoming(PrevPhi, PrevIfBlock);
1072 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1073 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1074 ConstantInt::get(Predicate->getType(), 1));
1076 // Create "cond" block
1078 // %EltAddr = getelementptr i32* %1, i32 0
1079 // %Elt = load i32* %EltAddr
1080 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1082 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1083 Builder.SetInsertPoint(InsertPt);
1086 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1087 LoadInst* Load = Builder.CreateLoad(Gep, false);
1088 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1090 // Create "else" block, fill it in the next iteration
1091 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1092 Builder.SetInsertPoint(InsertPt);
1093 Instruction *OldBr = IfBlock->getTerminator();
1094 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1095 OldBr->eraseFromParent();
1096 PrevIfBlock = IfBlock;
1097 IfBlock = NewIfBlock;
1100 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1101 Phi->addIncoming(VResult, CondBlock);
1102 Phi->addIncoming(PrevPhi, PrevIfBlock);
1103 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1104 CI->replaceAllUsesWith(NewI);
1105 CI->eraseFromParent();
1108 // ScalarizeMaskedStore() translates masked store intrinsic, like
1109 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1111 // to a chain of basic blocks, that stores element one-by-one if
1112 // the appropriate mask bit is set
1114 // %1 = bitcast i8* %addr to i32*
1115 // %2 = extractelement <16 x i1> %mask, i32 0
1116 // %3 = icmp eq i1 %2, true
1117 // br i1 %3, label %cond.store, label %else
1119 // cond.store: ; preds = %0
1120 // %4 = extractelement <16 x i32> %val, i32 0
1121 // %5 = getelementptr i32* %1, i32 0
1122 // store i32 %4, i32* %5
1125 // else: ; preds = %0, %cond.store
1126 // %6 = extractelement <16 x i1> %mask, i32 1
1127 // %7 = icmp eq i1 %6, true
1128 // br i1 %7, label %cond.store1, label %else2
1130 // cond.store1: ; preds = %else
1131 // %8 = extractelement <16 x i32> %val, i32 1
1132 // %9 = getelementptr i32* %1, i32 1
1133 // store i32 %8, i32* %9
1136 static void ScalarizeMaskedStore(CallInst *CI) {
1137 Value *Ptr = CI->getArgOperand(1);
1138 Value *Src = CI->getArgOperand(0);
1139 Value *Mask = CI->getArgOperand(3);
1141 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1142 Type *EltTy = VecType->getElementType();
1144 assert(VecType && "Unexpected data type in masked store intrinsic");
1146 IRBuilder<> Builder(CI->getContext());
1147 Instruction *InsertPt = CI;
1148 BasicBlock *IfBlock = CI->getParent();
1149 Builder.SetInsertPoint(InsertPt);
1150 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1152 // Bitcast %addr fron i8* to EltTy*
1154 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1155 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1157 unsigned VectorWidth = VecType->getNumElements();
1158 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1160 // Fill the "else" block, created in the previous iteration
1162 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1163 // %to_store = icmp eq i1 %mask_1, true
1164 // br i1 %to_load, label %cond.store, label %else
1166 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1167 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1168 ConstantInt::get(Predicate->getType(), 1));
1170 // Create "cond" block
1172 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1173 // %EltAddr = getelementptr i32* %1, i32 0
1174 // %store i32 %OneElt, i32* %EltAddr
1176 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1177 Builder.SetInsertPoint(InsertPt);
1179 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1181 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1182 Builder.CreateStore(OneElt, Gep);
1184 // Create "else" block, fill it in the next iteration
1185 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1186 Builder.SetInsertPoint(InsertPt);
1187 Instruction *OldBr = IfBlock->getTerminator();
1188 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1189 OldBr->eraseFromParent();
1190 IfBlock = NewIfBlock;
1192 CI->eraseFromParent();
1195 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1196 BasicBlock *BB = CI->getParent();
1198 // Lower inline assembly if we can.
1199 // If we found an inline asm expession, and if the target knows how to
1200 // lower it to normal LLVM code, do so now.
1201 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1202 if (TLI->ExpandInlineAsm(CI)) {
1203 // Avoid invalidating the iterator.
1204 CurInstIterator = BB->begin();
1205 // Avoid processing instructions out of order, which could cause
1206 // reuse before a value is defined.
1210 // Sink address computing for memory operands into the block.
1211 if (OptimizeInlineAsmInst(CI))
1215 const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
1217 // Align the pointer arguments to this call if the target thinks it's a good
1219 unsigned MinSize, PrefAlign;
1220 if (TLI && TD && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1221 for (auto &Arg : CI->arg_operands()) {
1222 // We want to align both objects whose address is used directly and
1223 // objects whose address is used in casts and GEPs, though it only makes
1224 // sense for GEPs if the offset is a multiple of the desired alignment and
1225 // if size - offset meets the size threshold.
1226 if (!Arg->getType()->isPointerTy())
1228 APInt Offset(TD->getPointerSizeInBits(
1229 cast<PointerType>(Arg->getType())->getAddressSpace()), 0);
1230 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*TD, Offset);
1231 uint64_t Offset2 = Offset.getLimitedValue();
1233 if ((Offset2 & (PrefAlign-1)) == 0 &&
1234 (AI = dyn_cast<AllocaInst>(Val)) &&
1235 AI->getAlignment() < PrefAlign &&
1236 TD->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1237 AI->setAlignment(PrefAlign);
1238 // TODO: Also align GlobalVariables
1240 // If this is a memcpy (or similar) then we may be able to improve the
1242 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1243 unsigned Align = getKnownAlignment(MI->getDest(), *TD);
1244 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1245 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *TD));
1246 if (Align > MI->getAlignment())
1247 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1251 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1253 switch (II->getIntrinsicID()) {
1255 case Intrinsic::objectsize: {
1256 // Lower all uses of llvm.objectsize.*
1257 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1258 Type *ReturnTy = CI->getType();
1259 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1261 // Substituting this can cause recursive simplifications, which can
1262 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1264 WeakVH IterHandle(CurInstIterator);
1266 replaceAndRecursivelySimplify(CI, RetVal,
1269 // If the iterator instruction was recursively deleted, start over at the
1270 // start of the block.
1271 if (IterHandle != CurInstIterator) {
1272 CurInstIterator = BB->begin();
1277 case Intrinsic::masked_load: {
1278 // Scalarize unsupported vector masked load
1279 if (!TTI->isLegalMaskedLoad(CI->getType(), 1)) {
1280 ScalarizeMaskedLoad(CI);
1286 case Intrinsic::masked_store: {
1287 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType(), 1)) {
1288 ScalarizeMaskedStore(CI);
1297 SmallVector<Value*, 2> PtrOps;
1299 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
1300 while (!PtrOps.empty())
1301 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
1306 // From here on out we're working with named functions.
1307 if (!CI->getCalledFunction()) return false;
1309 // Lower all default uses of _chk calls. This is very similar
1310 // to what InstCombineCalls does, but here we are only lowering calls
1311 // to fortified library functions (e.g. __memcpy_chk) that have the default
1312 // "don't know" as the objectsize. Anything else should be left alone.
1313 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1314 if (Value *V = Simplifier.optimizeCall(CI)) {
1315 CI->replaceAllUsesWith(V);
1316 CI->eraseFromParent();
1322 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
1323 /// instructions to the predecessor to enable tail call optimizations. The
1324 /// case it is currently looking for is:
1327 /// %tmp0 = tail call i32 @f0()
1328 /// br label %return
1330 /// %tmp1 = tail call i32 @f1()
1331 /// br label %return
1333 /// %tmp2 = tail call i32 @f2()
1334 /// br label %return
1336 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1344 /// %tmp0 = tail call i32 @f0()
1347 /// %tmp1 = tail call i32 @f1()
1350 /// %tmp2 = tail call i32 @f2()
1353 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
1357 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1361 PHINode *PN = nullptr;
1362 BitCastInst *BCI = nullptr;
1363 Value *V = RI->getReturnValue();
1365 BCI = dyn_cast<BitCastInst>(V);
1367 V = BCI->getOperand(0);
1369 PN = dyn_cast<PHINode>(V);
1374 if (PN && PN->getParent() != BB)
1377 // It's not safe to eliminate the sign / zero extension of the return value.
1378 // See llvm::isInTailCallPosition().
1379 const Function *F = BB->getParent();
1380 AttributeSet CallerAttrs = F->getAttributes();
1381 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1382 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1385 // Make sure there are no instructions between the PHI and return, or that the
1386 // return is the first instruction in the block.
1388 BasicBlock::iterator BI = BB->begin();
1389 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1391 // Also skip over the bitcast.
1396 BasicBlock::iterator BI = BB->begin();
1397 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1402 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1404 SmallVector<CallInst*, 4> TailCalls;
1406 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1407 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1408 // Make sure the phi value is indeed produced by the tail call.
1409 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1410 TLI->mayBeEmittedAsTailCall(CI))
1411 TailCalls.push_back(CI);
1414 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1415 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1416 if (!VisitedBBs.insert(*PI).second)
1419 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1420 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1421 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1422 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1426 CallInst *CI = dyn_cast<CallInst>(&*RI);
1427 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1428 TailCalls.push_back(CI);
1432 bool Changed = false;
1433 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1434 CallInst *CI = TailCalls[i];
1437 // Conservatively require the attributes of the call to match those of the
1438 // return. Ignore noalias because it doesn't affect the call sequence.
1439 AttributeSet CalleeAttrs = CS.getAttributes();
1440 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1441 removeAttribute(Attribute::NoAlias) !=
1442 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1443 removeAttribute(Attribute::NoAlias))
1446 // Make sure the call instruction is followed by an unconditional branch to
1447 // the return block.
1448 BasicBlock *CallBB = CI->getParent();
1449 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1450 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1453 // Duplicate the return into CallBB.
1454 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1455 ModifiedDT = Changed = true;
1459 // If we eliminated all predecessors of the block, delete the block now.
1460 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1461 BB->eraseFromParent();
1466 //===----------------------------------------------------------------------===//
1467 // Memory Optimization
1468 //===----------------------------------------------------------------------===//
1472 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1473 /// which holds actual Value*'s for register values.
1474 struct ExtAddrMode : public TargetLowering::AddrMode {
1477 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1478 void print(raw_ostream &OS) const;
1481 bool operator==(const ExtAddrMode& O) const {
1482 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1483 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1484 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1489 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1495 void ExtAddrMode::print(raw_ostream &OS) const {
1496 bool NeedPlus = false;
1499 OS << (NeedPlus ? " + " : "")
1501 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1506 OS << (NeedPlus ? " + " : "")
1512 OS << (NeedPlus ? " + " : "")
1514 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1518 OS << (NeedPlus ? " + " : "")
1520 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1526 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1527 void ExtAddrMode::dump() const {
1533 /// \brief This class provides transaction based operation on the IR.
1534 /// Every change made through this class is recorded in the internal state and
1535 /// can be undone (rollback) until commit is called.
1536 class TypePromotionTransaction {
1538 /// \brief This represents the common interface of the individual transaction.
1539 /// Each class implements the logic for doing one specific modification on
1540 /// the IR via the TypePromotionTransaction.
1541 class TypePromotionAction {
1543 /// The Instruction modified.
1547 /// \brief Constructor of the action.
1548 /// The constructor performs the related action on the IR.
1549 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1551 virtual ~TypePromotionAction() {}
1553 /// \brief Undo the modification done by this action.
1554 /// When this method is called, the IR must be in the same state as it was
1555 /// before this action was applied.
1556 /// \pre Undoing the action works if and only if the IR is in the exact same
1557 /// state as it was directly after this action was applied.
1558 virtual void undo() = 0;
1560 /// \brief Advocate every change made by this action.
1561 /// When the results on the IR of the action are to be kept, it is important
1562 /// to call this function, otherwise hidden information may be kept forever.
1563 virtual void commit() {
1564 // Nothing to be done, this action is not doing anything.
1568 /// \brief Utility to remember the position of an instruction.
1569 class InsertionHandler {
1570 /// Position of an instruction.
1571 /// Either an instruction:
1572 /// - Is the first in a basic block: BB is used.
1573 /// - Has a previous instructon: PrevInst is used.
1575 Instruction *PrevInst;
1578 /// Remember whether or not the instruction had a previous instruction.
1579 bool HasPrevInstruction;
1582 /// \brief Record the position of \p Inst.
1583 InsertionHandler(Instruction *Inst) {
1584 BasicBlock::iterator It = Inst;
1585 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1586 if (HasPrevInstruction)
1587 Point.PrevInst = --It;
1589 Point.BB = Inst->getParent();
1592 /// \brief Insert \p Inst at the recorded position.
1593 void insert(Instruction *Inst) {
1594 if (HasPrevInstruction) {
1595 if (Inst->getParent())
1596 Inst->removeFromParent();
1597 Inst->insertAfter(Point.PrevInst);
1599 Instruction *Position = Point.BB->getFirstInsertionPt();
1600 if (Inst->getParent())
1601 Inst->moveBefore(Position);
1603 Inst->insertBefore(Position);
1608 /// \brief Move an instruction before another.
1609 class InstructionMoveBefore : public TypePromotionAction {
1610 /// Original position of the instruction.
1611 InsertionHandler Position;
1614 /// \brief Move \p Inst before \p Before.
1615 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1616 : TypePromotionAction(Inst), Position(Inst) {
1617 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1618 Inst->moveBefore(Before);
1621 /// \brief Move the instruction back to its original position.
1622 void undo() override {
1623 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1624 Position.insert(Inst);
1628 /// \brief Set the operand of an instruction with a new value.
1629 class OperandSetter : public TypePromotionAction {
1630 /// Original operand of the instruction.
1632 /// Index of the modified instruction.
1636 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1637 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1638 : TypePromotionAction(Inst), Idx(Idx) {
1639 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1640 << "for:" << *Inst << "\n"
1641 << "with:" << *NewVal << "\n");
1642 Origin = Inst->getOperand(Idx);
1643 Inst->setOperand(Idx, NewVal);
1646 /// \brief Restore the original value of the instruction.
1647 void undo() override {
1648 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1649 << "for: " << *Inst << "\n"
1650 << "with: " << *Origin << "\n");
1651 Inst->setOperand(Idx, Origin);
1655 /// \brief Hide the operands of an instruction.
1656 /// Do as if this instruction was not using any of its operands.
1657 class OperandsHider : public TypePromotionAction {
1658 /// The list of original operands.
1659 SmallVector<Value *, 4> OriginalValues;
1662 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1663 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1664 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1665 unsigned NumOpnds = Inst->getNumOperands();
1666 OriginalValues.reserve(NumOpnds);
1667 for (unsigned It = 0; It < NumOpnds; ++It) {
1668 // Save the current operand.
1669 Value *Val = Inst->getOperand(It);
1670 OriginalValues.push_back(Val);
1672 // We could use OperandSetter here, but that would implied an overhead
1673 // that we are not willing to pay.
1674 Inst->setOperand(It, UndefValue::get(Val->getType()));
1678 /// \brief Restore the original list of uses.
1679 void undo() override {
1680 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1681 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1682 Inst->setOperand(It, OriginalValues[It]);
1686 /// \brief Build a truncate instruction.
1687 class TruncBuilder : public TypePromotionAction {
1690 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1692 /// trunc Opnd to Ty.
1693 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1694 IRBuilder<> Builder(Opnd);
1695 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
1696 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
1699 /// \brief Get the built value.
1700 Value *getBuiltValue() { return Val; }
1702 /// \brief Remove the built instruction.
1703 void undo() override {
1704 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
1705 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1706 IVal->eraseFromParent();
1710 /// \brief Build a sign extension instruction.
1711 class SExtBuilder : public TypePromotionAction {
1714 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1716 /// sext Opnd to Ty.
1717 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1718 : TypePromotionAction(InsertPt) {
1719 IRBuilder<> Builder(InsertPt);
1720 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
1721 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
1724 /// \brief Get the built value.
1725 Value *getBuiltValue() { return Val; }
1727 /// \brief Remove the built instruction.
1728 void undo() override {
1729 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
1730 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1731 IVal->eraseFromParent();
1735 /// \brief Build a zero extension instruction.
1736 class ZExtBuilder : public TypePromotionAction {
1739 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
1741 /// zext Opnd to Ty.
1742 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1743 : TypePromotionAction(InsertPt) {
1744 IRBuilder<> Builder(InsertPt);
1745 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
1746 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
1749 /// \brief Get the built value.
1750 Value *getBuiltValue() { return Val; }
1752 /// \brief Remove the built instruction.
1753 void undo() override {
1754 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
1755 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1756 IVal->eraseFromParent();
1760 /// \brief Mutate an instruction to another type.
1761 class TypeMutator : public TypePromotionAction {
1762 /// Record the original type.
1766 /// \brief Mutate the type of \p Inst into \p NewTy.
1767 TypeMutator(Instruction *Inst, Type *NewTy)
1768 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1769 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1771 Inst->mutateType(NewTy);
1774 /// \brief Mutate the instruction back to its original type.
1775 void undo() override {
1776 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1778 Inst->mutateType(OrigTy);
1782 /// \brief Replace the uses of an instruction by another instruction.
1783 class UsesReplacer : public TypePromotionAction {
1784 /// Helper structure to keep track of the replaced uses.
1785 struct InstructionAndIdx {
1786 /// The instruction using the instruction.
1788 /// The index where this instruction is used for Inst.
1790 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1791 : Inst(Inst), Idx(Idx) {}
1794 /// Keep track of the original uses (pair Instruction, Index).
1795 SmallVector<InstructionAndIdx, 4> OriginalUses;
1796 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1799 /// \brief Replace all the use of \p Inst by \p New.
1800 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1801 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1803 // Record the original uses.
1804 for (Use &U : Inst->uses()) {
1805 Instruction *UserI = cast<Instruction>(U.getUser());
1806 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1808 // Now, we can replace the uses.
1809 Inst->replaceAllUsesWith(New);
1812 /// \brief Reassign the original uses of Inst to Inst.
1813 void undo() override {
1814 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1815 for (use_iterator UseIt = OriginalUses.begin(),
1816 EndIt = OriginalUses.end();
1817 UseIt != EndIt; ++UseIt) {
1818 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1823 /// \brief Remove an instruction from the IR.
1824 class InstructionRemover : public TypePromotionAction {
1825 /// Original position of the instruction.
1826 InsertionHandler Inserter;
1827 /// Helper structure to hide all the link to the instruction. In other
1828 /// words, this helps to do as if the instruction was removed.
1829 OperandsHider Hider;
1830 /// Keep track of the uses replaced, if any.
1831 UsesReplacer *Replacer;
1834 /// \brief Remove all reference of \p Inst and optinally replace all its
1836 /// \pre If !Inst->use_empty(), then New != nullptr
1837 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1838 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1841 Replacer = new UsesReplacer(Inst, New);
1842 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1843 Inst->removeFromParent();
1846 ~InstructionRemover() { delete Replacer; }
1848 /// \brief Really remove the instruction.
1849 void commit() override { delete Inst; }
1851 /// \brief Resurrect the instruction and reassign it to the proper uses if
1852 /// new value was provided when build this action.
1853 void undo() override {
1854 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1855 Inserter.insert(Inst);
1863 /// Restoration point.
1864 /// The restoration point is a pointer to an action instead of an iterator
1865 /// because the iterator may be invalidated but not the pointer.
1866 typedef const TypePromotionAction *ConstRestorationPt;
1867 /// Advocate every changes made in that transaction.
1869 /// Undo all the changes made after the given point.
1870 void rollback(ConstRestorationPt Point);
1871 /// Get the current restoration point.
1872 ConstRestorationPt getRestorationPoint() const;
1874 /// \name API for IR modification with state keeping to support rollback.
1876 /// Same as Instruction::setOperand.
1877 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1878 /// Same as Instruction::eraseFromParent.
1879 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
1880 /// Same as Value::replaceAllUsesWith.
1881 void replaceAllUsesWith(Instruction *Inst, Value *New);
1882 /// Same as Value::mutateType.
1883 void mutateType(Instruction *Inst, Type *NewTy);
1884 /// Same as IRBuilder::createTrunc.
1885 Value *createTrunc(Instruction *Opnd, Type *Ty);
1886 /// Same as IRBuilder::createSExt.
1887 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1888 /// Same as IRBuilder::createZExt.
1889 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
1890 /// Same as Instruction::moveBefore.
1891 void moveBefore(Instruction *Inst, Instruction *Before);
1895 /// The ordered list of actions made so far.
1896 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
1897 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
1900 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1903 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
1906 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1909 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
1912 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1914 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
1917 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1918 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
1921 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1923 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
1924 Value *Val = Ptr->getBuiltValue();
1925 Actions.push_back(std::move(Ptr));
1929 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
1930 Value *Opnd, Type *Ty) {
1931 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
1932 Value *Val = Ptr->getBuiltValue();
1933 Actions.push_back(std::move(Ptr));
1937 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
1938 Value *Opnd, Type *Ty) {
1939 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
1940 Value *Val = Ptr->getBuiltValue();
1941 Actions.push_back(std::move(Ptr));
1945 void TypePromotionTransaction::moveBefore(Instruction *Inst,
1946 Instruction *Before) {
1948 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
1951 TypePromotionTransaction::ConstRestorationPt
1952 TypePromotionTransaction::getRestorationPoint() const {
1953 return !Actions.empty() ? Actions.back().get() : nullptr;
1956 void TypePromotionTransaction::commit() {
1957 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
1963 void TypePromotionTransaction::rollback(
1964 TypePromotionTransaction::ConstRestorationPt Point) {
1965 while (!Actions.empty() && Point != Actions.back().get()) {
1966 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
1971 /// \brief A helper class for matching addressing modes.
1973 /// This encapsulates the logic for matching the target-legal addressing modes.
1974 class AddressingModeMatcher {
1975 SmallVectorImpl<Instruction*> &AddrModeInsts;
1976 const TargetMachine &TM;
1977 const TargetLowering &TLI;
1979 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
1980 /// the memory instruction that we're computing this address for.
1982 Instruction *MemoryInst;
1984 /// AddrMode - This is the addressing mode that we're building up. This is
1985 /// part of the return value of this addressing mode matching stuff.
1986 ExtAddrMode &AddrMode;
1988 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
1989 const SetOfInstrs &InsertedTruncs;
1990 /// A map from the instructions to their type before promotion.
1991 InstrToOrigTy &PromotedInsts;
1992 /// The ongoing transaction where every action should be registered.
1993 TypePromotionTransaction &TPT;
1995 /// IgnoreProfitability - This is set to true when we should not do
1996 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
1997 /// always returns true.
1998 bool IgnoreProfitability;
2000 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2001 const TargetMachine &TM, Type *AT, Instruction *MI,
2002 ExtAddrMode &AM, const SetOfInstrs &InsertedTruncs,
2003 InstrToOrigTy &PromotedInsts,
2004 TypePromotionTransaction &TPT)
2005 : AddrModeInsts(AMI), TM(TM),
2006 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
2007 ->getTargetLowering()),
2008 AccessTy(AT), MemoryInst(MI), AddrMode(AM),
2009 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
2010 IgnoreProfitability = false;
2014 /// Match - Find the maximal addressing mode that a load/store of V can fold,
2015 /// give an access type of AccessTy. This returns a list of involved
2016 /// instructions in AddrModeInsts.
2017 /// \p InsertedTruncs The truncate instruction inserted by other
2020 /// \p PromotedInsts maps the instructions to their type before promotion.
2021 /// \p The ongoing transaction where every action should be registered.
2022 static ExtAddrMode Match(Value *V, Type *AccessTy,
2023 Instruction *MemoryInst,
2024 SmallVectorImpl<Instruction*> &AddrModeInsts,
2025 const TargetMachine &TM,
2026 const SetOfInstrs &InsertedTruncs,
2027 InstrToOrigTy &PromotedInsts,
2028 TypePromotionTransaction &TPT) {
2031 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy,
2032 MemoryInst, Result, InsertedTruncs,
2033 PromotedInsts, TPT).MatchAddr(V, 0);
2034 (void)Success; assert(Success && "Couldn't select *anything*?");
2038 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2039 bool MatchAddr(Value *V, unsigned Depth);
2040 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2041 bool *MovedAway = nullptr);
2042 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
2043 ExtAddrMode &AMBefore,
2044 ExtAddrMode &AMAfter);
2045 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2046 bool IsPromotionProfitable(unsigned NewCost, unsigned OldCost,
2047 Value *PromotedOperand) const;
2050 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
2051 /// Return true and update AddrMode if this addr mode is legal for the target,
2053 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
2055 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2056 // mode. Just process that directly.
2058 return MatchAddr(ScaleReg, Depth);
2060 // If the scale is 0, it takes nothing to add this.
2064 // If we already have a scale of this value, we can add to it, otherwise, we
2065 // need an available scale field.
2066 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2069 ExtAddrMode TestAddrMode = AddrMode;
2071 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2072 // [A+B + A*7] -> [B+A*8].
2073 TestAddrMode.Scale += Scale;
2074 TestAddrMode.ScaledReg = ScaleReg;
2076 // If the new address isn't legal, bail out.
2077 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
2080 // It was legal, so commit it.
2081 AddrMode = TestAddrMode;
2083 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2084 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2085 // X*Scale + C*Scale to addr mode.
2086 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2087 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2088 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2089 TestAddrMode.ScaledReg = AddLHS;
2090 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2092 // If this addressing mode is legal, commit it and remember that we folded
2093 // this instruction.
2094 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
2095 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2096 AddrMode = TestAddrMode;
2101 // Otherwise, not (x+c)*scale, just return what we have.
2105 /// MightBeFoldableInst - This is a little filter, which returns true if an
2106 /// addressing computation involving I might be folded into a load/store
2107 /// accessing it. This doesn't need to be perfect, but needs to accept at least
2108 /// the set of instructions that MatchOperationAddr can.
2109 static bool MightBeFoldableInst(Instruction *I) {
2110 switch (I->getOpcode()) {
2111 case Instruction::BitCast:
2112 case Instruction::AddrSpaceCast:
2113 // Don't touch identity bitcasts.
2114 if (I->getType() == I->getOperand(0)->getType())
2116 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2117 case Instruction::PtrToInt:
2118 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2120 case Instruction::IntToPtr:
2121 // We know the input is intptr_t, so this is foldable.
2123 case Instruction::Add:
2125 case Instruction::Mul:
2126 case Instruction::Shl:
2127 // Can only handle X*C and X << C.
2128 return isa<ConstantInt>(I->getOperand(1));
2129 case Instruction::GetElementPtr:
2136 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2137 /// \note \p Val is assumed to be the product of some type promotion.
2138 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2139 /// to be legal, as the non-promoted value would have had the same state.
2140 static bool isPromotedInstructionLegal(const TargetLowering &TLI, Value *Val) {
2141 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2144 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2145 // If the ISDOpcode is undefined, it was undefined before the promotion.
2148 // Otherwise, check if the promoted instruction is legal or not.
2149 return TLI.isOperationLegalOrCustom(
2150 ISDOpcode, TLI.getValueType(PromotedInst->getType()));
2153 /// \brief Hepler class to perform type promotion.
2154 class TypePromotionHelper {
2155 /// \brief Utility function to check whether or not a sign or zero extension
2156 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2157 /// either using the operands of \p Inst or promoting \p Inst.
2158 /// The type of the extension is defined by \p IsSExt.
2159 /// In other words, check if:
2160 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2161 /// #1 Promotion applies:
2162 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2163 /// #2 Operand reuses:
2164 /// ext opnd1 to ConsideredExtType.
2165 /// \p PromotedInsts maps the instructions to their type before promotion.
2166 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2167 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2169 /// \brief Utility function to determine if \p OpIdx should be promoted when
2170 /// promoting \p Inst.
2171 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2172 if (isa<SelectInst>(Inst) && OpIdx == 0)
2177 /// \brief Utility function to promote the operand of \p Ext when this
2178 /// operand is a promotable trunc or sext or zext.
2179 /// \p PromotedInsts maps the instructions to their type before promotion.
2180 /// \p CreatedInstsCost[out] contains the cost of all instructions
2181 /// created to promote the operand of Ext.
2182 /// Newly added extensions are inserted in \p Exts.
2183 /// Newly added truncates are inserted in \p Truncs.
2184 /// Should never be called directly.
2185 /// \return The promoted value which is used instead of Ext.
2186 static Value *promoteOperandForTruncAndAnyExt(
2187 Instruction *Ext, TypePromotionTransaction &TPT,
2188 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2189 SmallVectorImpl<Instruction *> *Exts,
2190 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
2192 /// \brief Utility function to promote the operand of \p Ext when this
2193 /// operand is promotable and is not a supported trunc or sext.
2194 /// \p PromotedInsts maps the instructions to their type before promotion.
2195 /// \p CreatedInstsCost[out] contains the cost of all the instructions
2196 /// created to promote the operand of Ext.
2197 /// Newly added extensions are inserted in \p Exts.
2198 /// Newly added truncates are inserted in \p Truncs.
2199 /// Should never be called directly.
2200 /// \return The promoted value which is used instead of Ext.
2201 static Value *promoteOperandForOther(Instruction *Ext,
2202 TypePromotionTransaction &TPT,
2203 InstrToOrigTy &PromotedInsts,
2204 unsigned &CreatedInstsCost,
2205 SmallVectorImpl<Instruction *> *Exts,
2206 SmallVectorImpl<Instruction *> *Truncs,
2207 const TargetLowering &TLI, bool IsSExt);
2209 /// \see promoteOperandForOther.
2210 static Value *signExtendOperandForOther(
2211 Instruction *Ext, TypePromotionTransaction &TPT,
2212 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2213 SmallVectorImpl<Instruction *> *Exts,
2214 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2215 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2216 Exts, Truncs, TLI, true);
2219 /// \see promoteOperandForOther.
2220 static Value *zeroExtendOperandForOther(
2221 Instruction *Ext, TypePromotionTransaction &TPT,
2222 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2223 SmallVectorImpl<Instruction *> *Exts,
2224 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2225 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2226 Exts, Truncs, TLI, false);
2230 /// Type for the utility function that promotes the operand of Ext.
2231 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2232 InstrToOrigTy &PromotedInsts,
2233 unsigned &CreatedInstsCost,
2234 SmallVectorImpl<Instruction *> *Exts,
2235 SmallVectorImpl<Instruction *> *Truncs,
2236 const TargetLowering &TLI);
2237 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2238 /// action to promote the operand of \p Ext instead of using Ext.
2239 /// \return NULL if no promotable action is possible with the current
2241 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
2242 /// the others CodeGenPrepare optimizations. This information is important
2243 /// because we do not want to promote these instructions as CodeGenPrepare
2244 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2245 /// \p PromotedInsts maps the instructions to their type before promotion.
2246 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2247 const TargetLowering &TLI,
2248 const InstrToOrigTy &PromotedInsts);
2251 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2252 Type *ConsideredExtType,
2253 const InstrToOrigTy &PromotedInsts,
2255 // The promotion helper does not know how to deal with vector types yet.
2256 // To be able to fix that, we would need to fix the places where we
2257 // statically extend, e.g., constants and such.
2258 if (Inst->getType()->isVectorTy())
2261 // We can always get through zext.
2262 if (isa<ZExtInst>(Inst))
2265 // sext(sext) is ok too.
2266 if (IsSExt && isa<SExtInst>(Inst))
2269 // We can get through binary operator, if it is legal. In other words, the
2270 // binary operator must have a nuw or nsw flag.
2271 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2272 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2273 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2274 (IsSExt && BinOp->hasNoSignedWrap())))
2277 // Check if we can do the following simplification.
2278 // ext(trunc(opnd)) --> ext(opnd)
2279 if (!isa<TruncInst>(Inst))
2282 Value *OpndVal = Inst->getOperand(0);
2283 // Check if we can use this operand in the extension.
2284 // If the type is larger than the result type of the extension,
2286 if (!OpndVal->getType()->isIntegerTy() ||
2287 OpndVal->getType()->getIntegerBitWidth() >
2288 ConsideredExtType->getIntegerBitWidth())
2291 // If the operand of the truncate is not an instruction, we will not have
2292 // any information on the dropped bits.
2293 // (Actually we could for constant but it is not worth the extra logic).
2294 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2298 // Check if the source of the type is narrow enough.
2299 // I.e., check that trunc just drops extended bits of the same kind of
2301 // #1 get the type of the operand and check the kind of the extended bits.
2302 const Type *OpndType;
2303 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2304 if (It != PromotedInsts.end() && It->second.IsSExt == IsSExt)
2305 OpndType = It->second.Ty;
2306 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2307 OpndType = Opnd->getOperand(0)->getType();
2311 // #2 check that the truncate just drop extended bits.
2312 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
2318 TypePromotionHelper::Action TypePromotionHelper::getAction(
2319 Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2320 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2321 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2322 "Unexpected instruction type");
2323 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2324 Type *ExtTy = Ext->getType();
2325 bool IsSExt = isa<SExtInst>(Ext);
2326 // If the operand of the extension is not an instruction, we cannot
2328 // If it, check we can get through.
2329 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2332 // Do not promote if the operand has been added by codegenprepare.
2333 // Otherwise, it means we are undoing an optimization that is likely to be
2334 // redone, thus causing potential infinite loop.
2335 if (isa<TruncInst>(ExtOpnd) && InsertedTruncs.count(ExtOpnd))
2338 // SExt or Trunc instructions.
2339 // Return the related handler.
2340 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2341 isa<ZExtInst>(ExtOpnd))
2342 return promoteOperandForTruncAndAnyExt;
2344 // Regular instruction.
2345 // Abort early if we will have to insert non-free instructions.
2346 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2348 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2351 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2352 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2353 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2354 SmallVectorImpl<Instruction *> *Exts,
2355 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2356 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2357 // get through it and this method should not be called.
2358 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2359 Value *ExtVal = SExt;
2360 bool HasMergedNonFreeExt = false;
2361 if (isa<ZExtInst>(SExtOpnd)) {
2362 // Replace s|zext(zext(opnd))
2364 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
2366 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2367 TPT.replaceAllUsesWith(SExt, ZExt);
2368 TPT.eraseInstruction(SExt);
2371 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2373 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2375 CreatedInstsCost = 0;
2377 // Remove dead code.
2378 if (SExtOpnd->use_empty())
2379 TPT.eraseInstruction(SExtOpnd);
2381 // Check if the extension is still needed.
2382 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2383 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2386 Exts->push_back(ExtInst);
2387 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
2392 // At this point we have: ext ty opnd to ty.
2393 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2394 Value *NextVal = ExtInst->getOperand(0);
2395 TPT.eraseInstruction(ExtInst, NextVal);
2399 Value *TypePromotionHelper::promoteOperandForOther(
2400 Instruction *Ext, TypePromotionTransaction &TPT,
2401 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2402 SmallVectorImpl<Instruction *> *Exts,
2403 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
2405 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2406 // get through it and this method should not be called.
2407 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2408 CreatedInstsCost = 0;
2409 if (!ExtOpnd->hasOneUse()) {
2410 // ExtOpnd will be promoted.
2411 // All its uses, but Ext, will need to use a truncated value of the
2412 // promoted version.
2413 // Create the truncate now.
2414 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2415 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2416 ITrunc->removeFromParent();
2417 // Insert it just after the definition.
2418 ITrunc->insertAfter(ExtOpnd);
2420 Truncs->push_back(ITrunc);
2423 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2424 // Restore the operand of Ext (which has been replace by the previous call
2425 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2426 TPT.setOperand(Ext, 0, ExtOpnd);
2429 // Get through the Instruction:
2430 // 1. Update its type.
2431 // 2. Replace the uses of Ext by Inst.
2432 // 3. Extend each operand that needs to be extended.
2434 // Remember the original type of the instruction before promotion.
2435 // This is useful to know that the high bits are sign extended bits.
2436 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2437 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2439 TPT.mutateType(ExtOpnd, Ext->getType());
2441 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2443 Instruction *ExtForOpnd = Ext;
2445 DEBUG(dbgs() << "Propagate Ext to operands\n");
2446 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2448 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2449 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2450 !shouldExtOperand(ExtOpnd, OpIdx)) {
2451 DEBUG(dbgs() << "No need to propagate\n");
2454 // Check if we can statically extend the operand.
2455 Value *Opnd = ExtOpnd->getOperand(OpIdx);
2456 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2457 DEBUG(dbgs() << "Statically extend\n");
2458 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2459 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
2460 : Cst->getValue().zext(BitWidth);
2461 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
2464 // UndefValue are typed, so we have to statically sign extend them.
2465 if (isa<UndefValue>(Opnd)) {
2466 DEBUG(dbgs() << "Statically extend\n");
2467 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2471 // Otherwise we have to explicity sign extend the operand.
2472 // Check if Ext was reused to extend an operand.
2474 // If yes, create a new one.
2475 DEBUG(dbgs() << "More operands to ext\n");
2476 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2477 : TPT.createZExt(Ext, Opnd, Ext->getType());
2478 if (!isa<Instruction>(ValForExtOpnd)) {
2479 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
2482 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
2485 Exts->push_back(ExtForOpnd);
2486 TPT.setOperand(ExtForOpnd, 0, Opnd);
2488 // Move the sign extension before the insertion point.
2489 TPT.moveBefore(ExtForOpnd, ExtOpnd);
2490 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2491 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
2492 // If more sext are required, new instructions will have to be created.
2493 ExtForOpnd = nullptr;
2495 if (ExtForOpnd == Ext) {
2496 DEBUG(dbgs() << "Extension is useless now\n");
2497 TPT.eraseInstruction(Ext);
2502 /// IsPromotionProfitable - Check whether or not promoting an instruction
2503 /// to a wider type was profitable.
2504 /// \p NewCost gives the cost of extension instructions created by the
2506 /// \p OldCost gives the cost of extension instructions before the promotion
2507 /// plus the number of instructions that have been
2508 /// matched in the addressing mode the promotion.
2509 /// \p PromotedOperand is the value that has been promoted.
2510 /// \return True if the promotion is profitable, false otherwise.
2511 bool AddressingModeMatcher::IsPromotionProfitable(
2512 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
2513 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
2514 // The cost of the new extensions is greater than the cost of the
2515 // old extension plus what we folded.
2516 // This is not profitable.
2517 if (NewCost > OldCost)
2519 if (NewCost < OldCost)
2521 // The promotion is neutral but it may help folding the sign extension in
2522 // loads for instance.
2523 // Check that we did not create an illegal instruction.
2524 return isPromotedInstructionLegal(TLI, PromotedOperand);
2527 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
2528 /// fold the operation into the addressing mode. If so, update the addressing
2529 /// mode and return true, otherwise return false without modifying AddrMode.
2530 /// If \p MovedAway is not NULL, it contains the information of whether or
2531 /// not AddrInst has to be folded into the addressing mode on success.
2532 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2533 /// because it has been moved away.
2534 /// Thus AddrInst must not be added in the matched instructions.
2535 /// This state can happen when AddrInst is a sext, since it may be moved away.
2536 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2537 /// not be referenced anymore.
2538 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
2541 // Avoid exponential behavior on extremely deep expression trees.
2542 if (Depth >= 5) return false;
2544 // By default, all matched instructions stay in place.
2549 case Instruction::PtrToInt:
2550 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2551 return MatchAddr(AddrInst->getOperand(0), Depth);
2552 case Instruction::IntToPtr:
2553 // This inttoptr is a no-op if the integer type is pointer sized.
2554 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
2555 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
2556 return MatchAddr(AddrInst->getOperand(0), Depth);
2558 case Instruction::BitCast:
2559 case Instruction::AddrSpaceCast:
2560 // BitCast is always a noop, and we can handle it as long as it is
2561 // int->int or pointer->pointer (we don't want int<->fp or something).
2562 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2563 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2564 // Don't touch identity bitcasts. These were probably put here by LSR,
2565 // and we don't want to mess around with them. Assume it knows what it
2567 AddrInst->getOperand(0)->getType() != AddrInst->getType())
2568 return MatchAddr(AddrInst->getOperand(0), Depth);
2570 case Instruction::Add: {
2571 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2572 ExtAddrMode BackupAddrMode = AddrMode;
2573 unsigned OldSize = AddrModeInsts.size();
2574 // Start a transaction at this point.
2575 // The LHS may match but not the RHS.
2576 // Therefore, we need a higher level restoration point to undo partially
2577 // matched operation.
2578 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2579 TPT.getRestorationPoint();
2581 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2582 MatchAddr(AddrInst->getOperand(0), Depth+1))
2585 // Restore the old addr mode info.
2586 AddrMode = BackupAddrMode;
2587 AddrModeInsts.resize(OldSize);
2588 TPT.rollback(LastKnownGood);
2590 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2591 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2592 MatchAddr(AddrInst->getOperand(1), Depth+1))
2595 // Otherwise we definitely can't merge the ADD in.
2596 AddrMode = BackupAddrMode;
2597 AddrModeInsts.resize(OldSize);
2598 TPT.rollback(LastKnownGood);
2601 //case Instruction::Or:
2602 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2604 case Instruction::Mul:
2605 case Instruction::Shl: {
2606 // Can only handle X*C and X << C.
2607 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2610 int64_t Scale = RHS->getSExtValue();
2611 if (Opcode == Instruction::Shl)
2612 Scale = 1LL << Scale;
2614 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2616 case Instruction::GetElementPtr: {
2617 // Scan the GEP. We check it if it contains constant offsets and at most
2618 // one variable offset.
2619 int VariableOperand = -1;
2620 unsigned VariableScale = 0;
2622 int64_t ConstantOffset = 0;
2623 const DataLayout *TD = TLI.getDataLayout();
2624 gep_type_iterator GTI = gep_type_begin(AddrInst);
2625 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2626 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2627 const StructLayout *SL = TD->getStructLayout(STy);
2629 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2630 ConstantOffset += SL->getElementOffset(Idx);
2632 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
2633 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2634 ConstantOffset += CI->getSExtValue()*TypeSize;
2635 } else if (TypeSize) { // Scales of zero don't do anything.
2636 // We only allow one variable index at the moment.
2637 if (VariableOperand != -1)
2640 // Remember the variable index.
2641 VariableOperand = i;
2642 VariableScale = TypeSize;
2647 // A common case is for the GEP to only do a constant offset. In this case,
2648 // just add it to the disp field and check validity.
2649 if (VariableOperand == -1) {
2650 AddrMode.BaseOffs += ConstantOffset;
2651 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
2652 // Check to see if we can fold the base pointer in too.
2653 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2656 AddrMode.BaseOffs -= ConstantOffset;
2660 // Save the valid addressing mode in case we can't match.
2661 ExtAddrMode BackupAddrMode = AddrMode;
2662 unsigned OldSize = AddrModeInsts.size();
2664 // See if the scale and offset amount is valid for this target.
2665 AddrMode.BaseOffs += ConstantOffset;
2667 // Match the base operand of the GEP.
2668 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2669 // If it couldn't be matched, just stuff the value in a register.
2670 if (AddrMode.HasBaseReg) {
2671 AddrMode = BackupAddrMode;
2672 AddrModeInsts.resize(OldSize);
2675 AddrMode.HasBaseReg = true;
2676 AddrMode.BaseReg = AddrInst->getOperand(0);
2679 // Match the remaining variable portion of the GEP.
2680 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2682 // If it couldn't be matched, try stuffing the base into a register
2683 // instead of matching it, and retrying the match of the scale.
2684 AddrMode = BackupAddrMode;
2685 AddrModeInsts.resize(OldSize);
2686 if (AddrMode.HasBaseReg)
2688 AddrMode.HasBaseReg = true;
2689 AddrMode.BaseReg = AddrInst->getOperand(0);
2690 AddrMode.BaseOffs += ConstantOffset;
2691 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2692 VariableScale, Depth)) {
2693 // If even that didn't work, bail.
2694 AddrMode = BackupAddrMode;
2695 AddrModeInsts.resize(OldSize);
2702 case Instruction::SExt:
2703 case Instruction::ZExt: {
2704 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
2708 // Try to move this ext out of the way of the addressing mode.
2709 // Ask for a method for doing so.
2710 TypePromotionHelper::Action TPH =
2711 TypePromotionHelper::getAction(Ext, InsertedTruncs, TLI, PromotedInsts);
2715 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2716 TPT.getRestorationPoint();
2717 unsigned CreatedInstsCost = 0;
2718 unsigned ExtCost = !TLI.isExtFree(Ext);
2719 Value *PromotedOperand =
2720 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
2721 // SExt has been moved away.
2722 // Thus either it will be rematched later in the recursive calls or it is
2723 // gone. Anyway, we must not fold it into the addressing mode at this point.
2727 // addr = gep base, idx
2729 // promotedOpnd = ext opnd <- no match here
2730 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2731 // addr = gep base, op <- match
2735 assert(PromotedOperand &&
2736 "TypePromotionHelper should have filtered out those cases");
2738 ExtAddrMode BackupAddrMode = AddrMode;
2739 unsigned OldSize = AddrModeInsts.size();
2741 if (!MatchAddr(PromotedOperand, Depth) ||
2742 // The total of the new cost is equals to the cost of the created
2744 // The total of the old cost is equals to the cost of the extension plus
2745 // what we have saved in the addressing mode.
2746 !IsPromotionProfitable(CreatedInstsCost,
2747 ExtCost + (AddrModeInsts.size() - OldSize),
2749 AddrMode = BackupAddrMode;
2750 AddrModeInsts.resize(OldSize);
2751 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2752 TPT.rollback(LastKnownGood);
2761 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2762 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2763 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2764 /// or intptr_t for the target.
2766 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2767 // Start a transaction at this point that we will rollback if the matching
2769 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2770 TPT.getRestorationPoint();
2771 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2772 // Fold in immediates if legal for the target.
2773 AddrMode.BaseOffs += CI->getSExtValue();
2774 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2776 AddrMode.BaseOffs -= CI->getSExtValue();
2777 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2778 // If this is a global variable, try to fold it into the addressing mode.
2779 if (!AddrMode.BaseGV) {
2780 AddrMode.BaseGV = GV;
2781 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2783 AddrMode.BaseGV = nullptr;
2785 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2786 ExtAddrMode BackupAddrMode = AddrMode;
2787 unsigned OldSize = AddrModeInsts.size();
2789 // Check to see if it is possible to fold this operation.
2790 bool MovedAway = false;
2791 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2792 // This instruction may have been move away. If so, there is nothing
2796 // Okay, it's possible to fold this. Check to see if it is actually
2797 // *profitable* to do so. We use a simple cost model to avoid increasing
2798 // register pressure too much.
2799 if (I->hasOneUse() ||
2800 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2801 AddrModeInsts.push_back(I);
2805 // It isn't profitable to do this, roll back.
2806 //cerr << "NOT FOLDING: " << *I;
2807 AddrMode = BackupAddrMode;
2808 AddrModeInsts.resize(OldSize);
2809 TPT.rollback(LastKnownGood);
2811 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2812 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2814 TPT.rollback(LastKnownGood);
2815 } else if (isa<ConstantPointerNull>(Addr)) {
2816 // Null pointer gets folded without affecting the addressing mode.
2820 // Worse case, the target should support [reg] addressing modes. :)
2821 if (!AddrMode.HasBaseReg) {
2822 AddrMode.HasBaseReg = true;
2823 AddrMode.BaseReg = Addr;
2824 // Still check for legality in case the target supports [imm] but not [i+r].
2825 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2827 AddrMode.HasBaseReg = false;
2828 AddrMode.BaseReg = nullptr;
2831 // If the base register is already taken, see if we can do [r+r].
2832 if (AddrMode.Scale == 0) {
2834 AddrMode.ScaledReg = Addr;
2835 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2838 AddrMode.ScaledReg = nullptr;
2841 TPT.rollback(LastKnownGood);
2845 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2846 /// inline asm call are due to memory operands. If so, return true, otherwise
2848 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2849 const TargetMachine &TM) {
2850 const Function *F = CI->getParent()->getParent();
2851 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
2852 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
2853 TargetLowering::AsmOperandInfoVector TargetConstraints =
2854 TLI->ParseConstraints(TRI, ImmutableCallSite(CI));
2855 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2856 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2858 // Compute the constraint code and ConstraintType to use.
2859 TLI->ComputeConstraintToUse(OpInfo, SDValue());
2861 // If this asm operand is our Value*, and if it isn't an indirect memory
2862 // operand, we can't fold it!
2863 if (OpInfo.CallOperandVal == OpVal &&
2864 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2865 !OpInfo.isIndirect))
2872 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2873 /// memory use. If we find an obviously non-foldable instruction, return true.
2874 /// Add the ultimately found memory instructions to MemoryUses.
2875 static bool FindAllMemoryUses(
2877 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
2878 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
2879 // If we already considered this instruction, we're done.
2880 if (!ConsideredInsts.insert(I).second)
2883 // If this is an obviously unfoldable instruction, bail out.
2884 if (!MightBeFoldableInst(I))
2887 // Loop over all the uses, recursively processing them.
2888 for (Use &U : I->uses()) {
2889 Instruction *UserI = cast<Instruction>(U.getUser());
2891 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2892 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2896 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2897 unsigned opNo = U.getOperandNo();
2898 if (opNo == 0) return true; // Storing addr, not into addr.
2899 MemoryUses.push_back(std::make_pair(SI, opNo));
2903 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2904 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2905 if (!IA) return true;
2907 // If this is a memory operand, we're cool, otherwise bail out.
2908 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
2913 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
2920 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2921 /// the use site that we're folding it into. If so, there is no cost to
2922 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2923 /// that we know are live at the instruction already.
2924 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2925 Value *KnownLive2) {
2926 // If Val is either of the known-live values, we know it is live!
2927 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
2930 // All values other than instructions and arguments (e.g. constants) are live.
2931 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2933 // If Val is a constant sized alloca in the entry block, it is live, this is
2934 // true because it is just a reference to the stack/frame pointer, which is
2935 // live for the whole function.
2936 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2937 if (AI->isStaticAlloca())
2940 // Check to see if this value is already used in the memory instruction's
2941 // block. If so, it's already live into the block at the very least, so we
2942 // can reasonably fold it.
2943 return Val->isUsedInBasicBlock(MemoryInst->getParent());
2946 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
2947 /// mode of the machine to fold the specified instruction into a load or store
2948 /// that ultimately uses it. However, the specified instruction has multiple
2949 /// uses. Given this, it may actually increase register pressure to fold it
2950 /// into the load. For example, consider this code:
2954 /// use(Y) -> nonload/store
2958 /// In this case, Y has multiple uses, and can be folded into the load of Z
2959 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
2960 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
2961 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
2962 /// number of computations either.
2964 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
2965 /// X was live across 'load Z' for other reasons, we actually *would* want to
2966 /// fold the addressing mode in the Z case. This would make Y die earlier.
2967 bool AddressingModeMatcher::
2968 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
2969 ExtAddrMode &AMAfter) {
2970 if (IgnoreProfitability) return true;
2972 // AMBefore is the addressing mode before this instruction was folded into it,
2973 // and AMAfter is the addressing mode after the instruction was folded. Get
2974 // the set of registers referenced by AMAfter and subtract out those
2975 // referenced by AMBefore: this is the set of values which folding in this
2976 // address extends the lifetime of.
2978 // Note that there are only two potential values being referenced here,
2979 // BaseReg and ScaleReg (global addresses are always available, as are any
2980 // folded immediates).
2981 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
2983 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
2984 // lifetime wasn't extended by adding this instruction.
2985 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2987 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2988 ScaledReg = nullptr;
2990 // If folding this instruction (and it's subexprs) didn't extend any live
2991 // ranges, we're ok with it.
2992 if (!BaseReg && !ScaledReg)
2995 // If all uses of this instruction are ultimately load/store/inlineasm's,
2996 // check to see if their addressing modes will include this instruction. If
2997 // so, we can fold it into all uses, so it doesn't matter if it has multiple
2999 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
3000 SmallPtrSet<Instruction*, 16> ConsideredInsts;
3001 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
3002 return false; // Has a non-memory, non-foldable use!
3004 // Now that we know that all uses of this instruction are part of a chain of
3005 // computation involving only operations that could theoretically be folded
3006 // into a memory use, loop over each of these uses and see if they could
3007 // *actually* fold the instruction.
3008 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
3009 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
3010 Instruction *User = MemoryUses[i].first;
3011 unsigned OpNo = MemoryUses[i].second;
3013 // Get the access type of this use. If the use isn't a pointer, we don't
3014 // know what it accesses.
3015 Value *Address = User->getOperand(OpNo);
3016 if (!Address->getType()->isPointerTy())
3018 Type *AddressAccessTy = Address->getType()->getPointerElementType();
3020 // Do a match against the root of this address, ignoring profitability. This
3021 // will tell us if the addressing mode for the memory operation will
3022 // *actually* cover the shared instruction.
3024 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3025 TPT.getRestorationPoint();
3026 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy,
3027 MemoryInst, Result, InsertedTruncs,
3028 PromotedInsts, TPT);
3029 Matcher.IgnoreProfitability = true;
3030 bool Success = Matcher.MatchAddr(Address, 0);
3031 (void)Success; assert(Success && "Couldn't select *anything*?");
3033 // The match was to check the profitability, the changes made are not
3034 // part of the original matcher. Therefore, they should be dropped
3035 // otherwise the original matcher will not present the right state.
3036 TPT.rollback(LastKnownGood);
3038 // If the match didn't cover I, then it won't be shared by it.
3039 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3040 I) == MatchedAddrModeInsts.end())
3043 MatchedAddrModeInsts.clear();
3049 } // end anonymous namespace
3051 /// IsNonLocalValue - Return true if the specified values are defined in a
3052 /// different basic block than BB.
3053 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3054 if (Instruction *I = dyn_cast<Instruction>(V))
3055 return I->getParent() != BB;
3059 /// OptimizeMemoryInst - Load and Store Instructions often have
3060 /// addressing modes that can do significant amounts of computation. As such,
3061 /// instruction selection will try to get the load or store to do as much
3062 /// computation as possible for the program. The problem is that isel can only
3063 /// see within a single block. As such, we sink as much legal addressing mode
3064 /// stuff into the block as possible.
3066 /// This method is used to optimize both load/store and inline asms with memory
3068 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3072 // Try to collapse single-value PHI nodes. This is necessary to undo
3073 // unprofitable PRE transformations.
3074 SmallVector<Value*, 8> worklist;
3075 SmallPtrSet<Value*, 16> Visited;
3076 worklist.push_back(Addr);
3078 // Use a worklist to iteratively look through PHI nodes, and ensure that
3079 // the addressing mode obtained from the non-PHI roots of the graph
3081 Value *Consensus = nullptr;
3082 unsigned NumUsesConsensus = 0;
3083 bool IsNumUsesConsensusValid = false;
3084 SmallVector<Instruction*, 16> AddrModeInsts;
3085 ExtAddrMode AddrMode;
3086 TypePromotionTransaction TPT;
3087 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3088 TPT.getRestorationPoint();
3089 while (!worklist.empty()) {
3090 Value *V = worklist.back();
3091 worklist.pop_back();
3093 // Break use-def graph loops.
3094 if (!Visited.insert(V).second) {
3095 Consensus = nullptr;
3099 // For a PHI node, push all of its incoming values.
3100 if (PHINode *P = dyn_cast<PHINode>(V)) {
3101 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
3102 worklist.push_back(P->getIncomingValue(i));
3106 // For non-PHIs, determine the addressing mode being computed.
3107 SmallVector<Instruction*, 16> NewAddrModeInsts;
3108 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3109 V, AccessTy, MemoryInst, NewAddrModeInsts, *TM, InsertedTruncsSet,
3110 PromotedInsts, TPT);
3112 // This check is broken into two cases with very similar code to avoid using
3113 // getNumUses() as much as possible. Some values have a lot of uses, so
3114 // calling getNumUses() unconditionally caused a significant compile-time
3118 AddrMode = NewAddrMode;
3119 AddrModeInsts = NewAddrModeInsts;
3121 } else if (NewAddrMode == AddrMode) {
3122 if (!IsNumUsesConsensusValid) {
3123 NumUsesConsensus = Consensus->getNumUses();
3124 IsNumUsesConsensusValid = true;
3127 // Ensure that the obtained addressing mode is equivalent to that obtained
3128 // for all other roots of the PHI traversal. Also, when choosing one
3129 // such root as representative, select the one with the most uses in order
3130 // to keep the cost modeling heuristics in AddressingModeMatcher
3132 unsigned NumUses = V->getNumUses();
3133 if (NumUses > NumUsesConsensus) {
3135 NumUsesConsensus = NumUses;
3136 AddrModeInsts = NewAddrModeInsts;
3141 Consensus = nullptr;
3145 // If the addressing mode couldn't be determined, or if multiple different
3146 // ones were determined, bail out now.
3148 TPT.rollback(LastKnownGood);
3153 // Check to see if any of the instructions supersumed by this addr mode are
3154 // non-local to I's BB.
3155 bool AnyNonLocal = false;
3156 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3157 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3163 // If all the instructions matched are already in this BB, don't do anything.
3165 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3169 // Insert this computation right after this user. Since our caller is
3170 // scanning from the top of the BB to the bottom, reuse of the expr are
3171 // guaranteed to happen later.
3172 IRBuilder<> Builder(MemoryInst);
3174 // Now that we determined the addressing expression we want to use and know
3175 // that we have to sink it into this block. Check to see if we have already
3176 // done this for some other load/store instr in this block. If so, reuse the
3178 Value *&SunkAddr = SunkAddrs[Addr];
3180 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3181 << *MemoryInst << "\n");
3182 if (SunkAddr->getType() != Addr->getType())
3183 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3184 } else if (AddrSinkUsingGEPs ||
3185 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3186 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3188 // By default, we use the GEP-based method when AA is used later. This
3189 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3190 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3191 << *MemoryInst << "\n");
3192 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3193 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3195 // First, find the pointer.
3196 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3197 ResultPtr = AddrMode.BaseReg;
3198 AddrMode.BaseReg = nullptr;
3201 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3202 // We can't add more than one pointer together, nor can we scale a
3203 // pointer (both of which seem meaningless).
3204 if (ResultPtr || AddrMode.Scale != 1)
3207 ResultPtr = AddrMode.ScaledReg;
3211 if (AddrMode.BaseGV) {
3215 ResultPtr = AddrMode.BaseGV;
3218 // If the real base value actually came from an inttoptr, then the matcher
3219 // will look through it and provide only the integer value. In that case,
3221 if (!ResultPtr && AddrMode.BaseReg) {
3223 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3224 AddrMode.BaseReg = nullptr;
3225 } else if (!ResultPtr && AddrMode.Scale == 1) {
3227 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3232 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3233 SunkAddr = Constant::getNullValue(Addr->getType());
3234 } else if (!ResultPtr) {
3238 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3239 Type *I8Ty = Builder.getInt8Ty();
3241 // Start with the base register. Do this first so that subsequent address
3242 // matching finds it last, which will prevent it from trying to match it
3243 // as the scaled value in case it happens to be a mul. That would be
3244 // problematic if we've sunk a different mul for the scale, because then
3245 // we'd end up sinking both muls.
3246 if (AddrMode.BaseReg) {
3247 Value *V = AddrMode.BaseReg;
3248 if (V->getType() != IntPtrTy)
3249 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3254 // Add the scale value.
3255 if (AddrMode.Scale) {
3256 Value *V = AddrMode.ScaledReg;
3257 if (V->getType() == IntPtrTy) {
3259 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3260 cast<IntegerType>(V->getType())->getBitWidth()) {
3261 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3263 // It is only safe to sign extend the BaseReg if we know that the math
3264 // required to create it did not overflow before we extend it. Since
3265 // the original IR value was tossed in favor of a constant back when
3266 // the AddrMode was created we need to bail out gracefully if widths
3267 // do not match instead of extending it.
3268 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3269 if (I && (ResultIndex != AddrMode.BaseReg))
3270 I->eraseFromParent();
3274 if (AddrMode.Scale != 1)
3275 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3278 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3283 // Add in the Base Offset if present.
3284 if (AddrMode.BaseOffs) {
3285 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3287 // We need to add this separately from the scale above to help with
3288 // SDAG consecutive load/store merging.
3289 if (ResultPtr->getType() != I8PtrTy)
3290 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3291 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3298 SunkAddr = ResultPtr;
3300 if (ResultPtr->getType() != I8PtrTy)
3301 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3302 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3305 if (SunkAddr->getType() != Addr->getType())
3306 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3309 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3310 << *MemoryInst << "\n");
3311 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3312 Value *Result = nullptr;
3314 // Start with the base register. Do this first so that subsequent address
3315 // matching finds it last, which will prevent it from trying to match it
3316 // as the scaled value in case it happens to be a mul. That would be
3317 // problematic if we've sunk a different mul for the scale, because then
3318 // we'd end up sinking both muls.
3319 if (AddrMode.BaseReg) {
3320 Value *V = AddrMode.BaseReg;
3321 if (V->getType()->isPointerTy())
3322 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3323 if (V->getType() != IntPtrTy)
3324 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3328 // Add the scale value.
3329 if (AddrMode.Scale) {
3330 Value *V = AddrMode.ScaledReg;
3331 if (V->getType() == IntPtrTy) {
3333 } else if (V->getType()->isPointerTy()) {
3334 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3335 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3336 cast<IntegerType>(V->getType())->getBitWidth()) {
3337 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3339 // It is only safe to sign extend the BaseReg if we know that the math
3340 // required to create it did not overflow before we extend it. Since
3341 // the original IR value was tossed in favor of a constant back when
3342 // the AddrMode was created we need to bail out gracefully if widths
3343 // do not match instead of extending it.
3344 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3345 if (I && (Result != AddrMode.BaseReg))
3346 I->eraseFromParent();
3349 if (AddrMode.Scale != 1)
3350 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3353 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3358 // Add in the BaseGV if present.
3359 if (AddrMode.BaseGV) {
3360 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3362 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3367 // Add in the Base Offset if present.
3368 if (AddrMode.BaseOffs) {
3369 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3371 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3377 SunkAddr = Constant::getNullValue(Addr->getType());
3379 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3382 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3384 // If we have no uses, recursively delete the value and all dead instructions
3386 if (Repl->use_empty()) {
3387 // This can cause recursive deletion, which can invalidate our iterator.
3388 // Use a WeakVH to hold onto it in case this happens.
3389 WeakVH IterHandle(CurInstIterator);
3390 BasicBlock *BB = CurInstIterator->getParent();
3392 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3394 if (IterHandle != CurInstIterator) {
3395 // If the iterator instruction was recursively deleted, start over at the
3396 // start of the block.
3397 CurInstIterator = BB->begin();
3405 /// OptimizeInlineAsmInst - If there are any memory operands, use
3406 /// OptimizeMemoryInst to sink their address computing into the block when
3407 /// possible / profitable.
3408 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
3409 bool MadeChange = false;
3411 const TargetRegisterInfo *TRI =
3412 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3413 TargetLowering::AsmOperandInfoVector
3414 TargetConstraints = TLI->ParseConstraints(TRI, CS);
3416 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3417 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3419 // Compute the constraint code and ConstraintType to use.
3420 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3422 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3423 OpInfo.isIndirect) {
3424 Value *OpVal = CS->getArgOperand(ArgNo++);
3425 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
3426 } else if (OpInfo.Type == InlineAsm::isInput)
3433 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3434 /// sign extensions.
3435 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3436 assert(!Inst->use_empty() && "Input must have at least one use");
3437 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3438 bool IsSExt = isa<SExtInst>(FirstUser);
3439 Type *ExtTy = FirstUser->getType();
3440 for (const User *U : Inst->users()) {
3441 const Instruction *UI = cast<Instruction>(U);
3442 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3444 Type *CurTy = UI->getType();
3445 // Same input and output types: Same instruction after CSE.
3449 // If IsSExt is true, we are in this situation:
3451 // b = sext ty1 a to ty2
3452 // c = sext ty1 a to ty3
3453 // Assuming ty2 is shorter than ty3, this could be turned into:
3455 // b = sext ty1 a to ty2
3456 // c = sext ty2 b to ty3
3457 // However, the last sext is not free.
3461 // This is a ZExt, maybe this is free to extend from one type to another.
3462 // In that case, we would not account for a different use.
3465 if (ExtTy->getScalarType()->getIntegerBitWidth() >
3466 CurTy->getScalarType()->getIntegerBitWidth()) {
3474 if (!TLI.isZExtFree(NarrowTy, LargeTy))
3477 // All uses are the same or can be derived from one another for free.
3481 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
3482 /// load instruction.
3483 /// If an ext(load) can be formed, it is returned via \p LI for the load
3484 /// and \p Inst for the extension.
3485 /// Otherwise LI == nullptr and Inst == nullptr.
3486 /// When some promotion happened, \p TPT contains the proper state to
3489 /// \return true when promoting was necessary to expose the ext(load)
3490 /// opportunity, false otherwise.
3494 /// %ld = load i32* %addr
3495 /// %add = add nuw i32 %ld, 4
3496 /// %zext = zext i32 %add to i64
3500 /// %ld = load i32* %addr
3501 /// %zext = zext i32 %ld to i64
3502 /// %add = add nuw i64 %zext, 4
3504 /// Thanks to the promotion, we can match zext(load i32*) to i64.
3505 bool CodeGenPrepare::ExtLdPromotion(TypePromotionTransaction &TPT,
3506 LoadInst *&LI, Instruction *&Inst,
3507 const SmallVectorImpl<Instruction *> &Exts,
3508 unsigned CreatedInstsCost = 0) {
3509 // Iterate over all the extensions to see if one form an ext(load).
3510 for (auto I : Exts) {
3511 // Check if we directly have ext(load).
3512 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
3514 // No promotion happened here.
3517 // Check whether or not we want to do any promotion.
3518 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
3520 // Get the action to perform the promotion.
3521 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
3522 I, InsertedTruncsSet, *TLI, PromotedInsts);
3523 // Check if we can promote.
3526 // Save the current state.
3527 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3528 TPT.getRestorationPoint();
3529 SmallVector<Instruction *, 4> NewExts;
3530 unsigned NewCreatedInstsCost = 0;
3531 unsigned ExtCost = !TLI->isExtFree(I);
3533 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
3534 &NewExts, nullptr, *TLI);
3535 assert(PromotedVal &&
3536 "TypePromotionHelper should have filtered out those cases");
3538 // We would be able to merge only one extension in a load.
3539 // Therefore, if we have more than 1 new extension we heuristically
3540 // cut this search path, because it means we degrade the code quality.
3541 // With exactly 2, the transformation is neutral, because we will merge
3542 // one extension but leave one. However, we optimistically keep going,
3543 // because the new extension may be removed too.
3544 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
3545 TotalCreatedInstsCost -= ExtCost;
3546 if (!StressExtLdPromotion &&
3547 (TotalCreatedInstsCost > 1 ||
3548 !isPromotedInstructionLegal(*TLI, PromotedVal))) {
3549 // The promotion is not profitable, rollback to the previous state.
3550 TPT.rollback(LastKnownGood);
3553 // The promotion is profitable.
3554 // Check if it exposes an ext(load).
3555 (void)ExtLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
3556 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
3557 // If we have created a new extension, i.e., now we have two
3558 // extensions. We must make sure one of them is merged with
3559 // the load, otherwise we may degrade the code quality.
3560 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
3561 // Promotion happened.
3563 // If this does not help to expose an ext(load) then, rollback.
3564 TPT.rollback(LastKnownGood);
3566 // None of the extension can form an ext(load).
3572 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
3573 /// basic block as the load, unless conditions are unfavorable. This allows
3574 /// SelectionDAG to fold the extend into the load.
3575 /// \p I[in/out] the extension may be modified during the process if some
3576 /// promotions apply.
3578 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *&I) {
3579 // Try to promote a chain of computation if it allows to form
3580 // an extended load.
3581 TypePromotionTransaction TPT;
3582 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3583 TPT.getRestorationPoint();
3584 SmallVector<Instruction *, 1> Exts;
3586 // Look for a load being extended.
3587 LoadInst *LI = nullptr;
3588 Instruction *OldExt = I;
3589 bool HasPromoted = ExtLdPromotion(TPT, LI, I, Exts);
3591 assert(!HasPromoted && !LI && "If we did not match any load instruction "
3592 "the code must remain the same");
3597 // If they're already in the same block, there's nothing to do.
3598 // Make the cheap checks first if we did not promote.
3599 // If we promoted, we need to check if it is indeed profitable.
3600 if (!HasPromoted && LI->getParent() == I->getParent())
3603 EVT VT = TLI->getValueType(I->getType());
3604 EVT LoadVT = TLI->getValueType(LI->getType());
3606 // If the load has other users and the truncate is not free, this probably
3607 // isn't worthwhile.
3608 if (!LI->hasOneUse() && TLI &&
3609 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
3610 !TLI->isTruncateFree(I->getType(), LI->getType())) {
3612 TPT.rollback(LastKnownGood);
3616 // Check whether the target supports casts folded into loads.
3618 if (isa<ZExtInst>(I))
3619 LType = ISD::ZEXTLOAD;
3621 assert(isa<SExtInst>(I) && "Unexpected ext type!");
3622 LType = ISD::SEXTLOAD;
3624 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
3626 TPT.rollback(LastKnownGood);
3630 // Move the extend into the same block as the load, so that SelectionDAG
3633 I->removeFromParent();
3639 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
3640 BasicBlock *DefBB = I->getParent();
3642 // If the result of a {s|z}ext and its source are both live out, rewrite all
3643 // other uses of the source with result of extension.
3644 Value *Src = I->getOperand(0);
3645 if (Src->hasOneUse())
3648 // Only do this xform if truncating is free.
3649 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
3652 // Only safe to perform the optimization if the source is also defined in
3654 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
3657 bool DefIsLiveOut = false;
3658 for (User *U : I->users()) {
3659 Instruction *UI = cast<Instruction>(U);
3661 // Figure out which BB this ext is used in.
3662 BasicBlock *UserBB = UI->getParent();
3663 if (UserBB == DefBB) continue;
3664 DefIsLiveOut = true;
3670 // Make sure none of the uses are PHI nodes.
3671 for (User *U : Src->users()) {
3672 Instruction *UI = cast<Instruction>(U);
3673 BasicBlock *UserBB = UI->getParent();
3674 if (UserBB == DefBB) continue;
3675 // Be conservative. We don't want this xform to end up introducing
3676 // reloads just before load / store instructions.
3677 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
3681 // InsertedTruncs - Only insert one trunc in each block once.
3682 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
3684 bool MadeChange = false;
3685 for (Use &U : Src->uses()) {
3686 Instruction *User = cast<Instruction>(U.getUser());
3688 // Figure out which BB this ext is used in.
3689 BasicBlock *UserBB = User->getParent();
3690 if (UserBB == DefBB) continue;
3692 // Both src and def are live in this block. Rewrite the use.
3693 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
3695 if (!InsertedTrunc) {
3696 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3697 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
3698 InsertedTruncsSet.insert(InsertedTrunc);
3701 // Replace a use of the {s|z}ext source with a use of the result.
3710 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
3711 /// turned into an explicit branch.
3712 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
3713 // FIXME: This should use the same heuristics as IfConversion to determine
3714 // whether a select is better represented as a branch. This requires that
3715 // branch probability metadata is preserved for the select, which is not the
3718 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3720 // If the branch is predicted right, an out of order CPU can avoid blocking on
3721 // the compare. Emit cmovs on compares with a memory operand as branches to
3722 // avoid stalls on the load from memory. If the compare has more than one use
3723 // there's probably another cmov or setcc around so it's not worth emitting a
3728 Value *CmpOp0 = Cmp->getOperand(0);
3729 Value *CmpOp1 = Cmp->getOperand(1);
3731 // We check that the memory operand has one use to avoid uses of the loaded
3732 // value directly after the compare, making branches unprofitable.
3733 return Cmp->hasOneUse() &&
3734 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3735 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
3739 /// If we have a SelectInst that will likely profit from branch prediction,
3740 /// turn it into a branch.
3741 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
3742 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3744 // Can we convert the 'select' to CF ?
3745 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3748 TargetLowering::SelectSupportKind SelectKind;
3750 SelectKind = TargetLowering::VectorMaskSelect;
3751 else if (SI->getType()->isVectorTy())
3752 SelectKind = TargetLowering::ScalarCondVectorVal;
3754 SelectKind = TargetLowering::ScalarValSelect;
3756 // Do we have efficient codegen support for this kind of 'selects' ?
3757 if (TLI->isSelectSupported(SelectKind)) {
3758 // We have efficient codegen support for the select instruction.
3759 // Check if it is profitable to keep this 'select'.
3760 if (!TLI->isPredictableSelectExpensive() ||
3761 !isFormingBranchFromSelectProfitable(SI))
3767 // First, we split the block containing the select into 2 blocks.
3768 BasicBlock *StartBlock = SI->getParent();
3769 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3770 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3772 // Create a new block serving as the landing pad for the branch.
3773 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3774 NextBlock->getParent(), NextBlock);
3776 // Move the unconditional branch from the block with the select in it into our
3777 // landing pad block.
3778 StartBlock->getTerminator()->eraseFromParent();
3779 BranchInst::Create(NextBlock, SmallBlock);
3781 // Insert the real conditional branch based on the original condition.
3782 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3784 // The select itself is replaced with a PHI Node.
3785 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3787 PN->addIncoming(SI->getTrueValue(), StartBlock);
3788 PN->addIncoming(SI->getFalseValue(), SmallBlock);
3789 SI->replaceAllUsesWith(PN);
3790 SI->eraseFromParent();
3792 // Instruct OptimizeBlock to skip to the next block.
3793 CurInstIterator = StartBlock->end();
3794 ++NumSelectsExpanded;
3798 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3799 SmallVector<int, 16> Mask(SVI->getShuffleMask());
3801 for (unsigned i = 0; i < Mask.size(); ++i) {
3802 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3804 SplatElem = Mask[i];
3810 /// Some targets have expensive vector shifts if the lanes aren't all the same
3811 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3812 /// it's often worth sinking a shufflevector splat down to its use so that
3813 /// codegen can spot all lanes are identical.
3814 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3815 BasicBlock *DefBB = SVI->getParent();
3817 // Only do this xform if variable vector shifts are particularly expensive.
3818 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3821 // We only expect better codegen by sinking a shuffle if we can recognise a
3823 if (!isBroadcastShuffle(SVI))
3826 // InsertedShuffles - Only insert a shuffle in each block once.
3827 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3829 bool MadeChange = false;
3830 for (User *U : SVI->users()) {
3831 Instruction *UI = cast<Instruction>(U);
3833 // Figure out which BB this ext is used in.
3834 BasicBlock *UserBB = UI->getParent();
3835 if (UserBB == DefBB) continue;
3837 // For now only apply this when the splat is used by a shift instruction.
3838 if (!UI->isShift()) continue;
3840 // Everything checks out, sink the shuffle if the user's block doesn't
3841 // already have a copy.
3842 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3844 if (!InsertedShuffle) {
3845 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3846 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3848 SVI->getOperand(2), "", InsertPt);
3851 UI->replaceUsesOfWith(SVI, InsertedShuffle);
3855 // If we removed all uses, nuke the shuffle.
3856 if (SVI->use_empty()) {
3857 SVI->eraseFromParent();
3865 /// \brief Helper class to promote a scalar operation to a vector one.
3866 /// This class is used to move downward extractelement transition.
3868 /// a = vector_op <2 x i32>
3869 /// b = extractelement <2 x i32> a, i32 0
3874 /// a = vector_op <2 x i32>
3875 /// c = vector_op a (equivalent to scalar_op on the related lane)
3876 /// * d = extractelement <2 x i32> c, i32 0
3878 /// Assuming both extractelement and store can be combine, we get rid of the
3880 class VectorPromoteHelper {
3881 /// Used to perform some checks on the legality of vector operations.
3882 const TargetLowering &TLI;
3884 /// Used to estimated the cost of the promoted chain.
3885 const TargetTransformInfo &TTI;
3887 /// The transition being moved downwards.
3888 Instruction *Transition;
3889 /// The sequence of instructions to be promoted.
3890 SmallVector<Instruction *, 4> InstsToBePromoted;
3891 /// Cost of combining a store and an extract.
3892 unsigned StoreExtractCombineCost;
3893 /// Instruction that will be combined with the transition.
3894 Instruction *CombineInst;
3896 /// \brief The instruction that represents the current end of the transition.
3897 /// Since we are faking the promotion until we reach the end of the chain
3898 /// of computation, we need a way to get the current end of the transition.
3899 Instruction *getEndOfTransition() const {
3900 if (InstsToBePromoted.empty())
3902 return InstsToBePromoted.back();
3905 /// \brief Return the index of the original value in the transition.
3906 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
3907 /// c, is at index 0.
3908 unsigned getTransitionOriginalValueIdx() const {
3909 assert(isa<ExtractElementInst>(Transition) &&
3910 "Other kind of transitions are not supported yet");
3914 /// \brief Return the index of the index in the transition.
3915 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
3917 unsigned getTransitionIdx() const {
3918 assert(isa<ExtractElementInst>(Transition) &&
3919 "Other kind of transitions are not supported yet");
3923 /// \brief Get the type of the transition.
3924 /// This is the type of the original value.
3925 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
3926 /// transition is <2 x i32>.
3927 Type *getTransitionType() const {
3928 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
3931 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
3932 /// I.e., we have the following sequence:
3933 /// Def = Transition <ty1> a to <ty2>
3934 /// b = ToBePromoted <ty2> Def, ...
3936 /// b = ToBePromoted <ty1> a, ...
3937 /// Def = Transition <ty1> ToBePromoted to <ty2>
3938 void promoteImpl(Instruction *ToBePromoted);
3940 /// \brief Check whether or not it is profitable to promote all the
3941 /// instructions enqueued to be promoted.
3942 bool isProfitableToPromote() {
3943 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
3944 unsigned Index = isa<ConstantInt>(ValIdx)
3945 ? cast<ConstantInt>(ValIdx)->getZExtValue()
3947 Type *PromotedType = getTransitionType();
3949 StoreInst *ST = cast<StoreInst>(CombineInst);
3950 unsigned AS = ST->getPointerAddressSpace();
3951 unsigned Align = ST->getAlignment();
3952 // Check if this store is supported.
3953 if (!TLI.allowsMisalignedMemoryAccesses(
3954 TLI.getValueType(ST->getValueOperand()->getType()), AS, Align)) {
3955 // If this is not supported, there is no way we can combine
3956 // the extract with the store.
3960 // The scalar chain of computation has to pay for the transition
3961 // scalar to vector.
3962 // The vector chain has to account for the combining cost.
3963 uint64_t ScalarCost =
3964 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
3965 uint64_t VectorCost = StoreExtractCombineCost;
3966 for (const auto &Inst : InstsToBePromoted) {
3967 // Compute the cost.
3968 // By construction, all instructions being promoted are arithmetic ones.
3969 // Moreover, one argument is a constant that can be viewed as a splat
3971 Value *Arg0 = Inst->getOperand(0);
3972 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
3973 isa<ConstantFP>(Arg0);
3974 TargetTransformInfo::OperandValueKind Arg0OVK =
3975 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3976 : TargetTransformInfo::OK_AnyValue;
3977 TargetTransformInfo::OperandValueKind Arg1OVK =
3978 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3979 : TargetTransformInfo::OK_AnyValue;
3980 ScalarCost += TTI.getArithmeticInstrCost(
3981 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
3982 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
3985 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
3986 << ScalarCost << "\nVector: " << VectorCost << '\n');
3987 return ScalarCost > VectorCost;
3990 /// \brief Generate a constant vector with \p Val with the same
3991 /// number of elements as the transition.
3992 /// \p UseSplat defines whether or not \p Val should be replicated
3993 /// accross the whole vector.
3994 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
3995 /// otherwise we generate a vector with as many undef as possible:
3996 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
3997 /// used at the index of the extract.
3998 Value *getConstantVector(Constant *Val, bool UseSplat) const {
3999 unsigned ExtractIdx = UINT_MAX;
4001 // If we cannot determine where the constant must be, we have to
4002 // use a splat constant.
4003 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
4004 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
4005 ExtractIdx = CstVal->getSExtValue();
4010 unsigned End = getTransitionType()->getVectorNumElements();
4012 return ConstantVector::getSplat(End, Val);
4014 SmallVector<Constant *, 4> ConstVec;
4015 UndefValue *UndefVal = UndefValue::get(Val->getType());
4016 for (unsigned Idx = 0; Idx != End; ++Idx) {
4017 if (Idx == ExtractIdx)
4018 ConstVec.push_back(Val);
4020 ConstVec.push_back(UndefVal);
4022 return ConstantVector::get(ConstVec);
4025 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
4026 /// in \p Use can trigger undefined behavior.
4027 static bool canCauseUndefinedBehavior(const Instruction *Use,
4028 unsigned OperandIdx) {
4029 // This is not safe to introduce undef when the operand is on
4030 // the right hand side of a division-like instruction.
4031 if (OperandIdx != 1)
4033 switch (Use->getOpcode()) {
4036 case Instruction::SDiv:
4037 case Instruction::UDiv:
4038 case Instruction::SRem:
4039 case Instruction::URem:
4041 case Instruction::FDiv:
4042 case Instruction::FRem:
4043 return !Use->hasNoNaNs();
4045 llvm_unreachable(nullptr);
4049 VectorPromoteHelper(const TargetLowering &TLI, const TargetTransformInfo &TTI,
4050 Instruction *Transition, unsigned CombineCost)
4051 : TLI(TLI), TTI(TTI), Transition(Transition),
4052 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4053 assert(Transition && "Do not know how to promote null");
4056 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4057 bool canPromote(const Instruction *ToBePromoted) const {
4058 // We could support CastInst too.
4059 return isa<BinaryOperator>(ToBePromoted);
4062 /// \brief Check if it is profitable to promote \p ToBePromoted
4063 /// by moving downward the transition through.
4064 bool shouldPromote(const Instruction *ToBePromoted) const {
4065 // Promote only if all the operands can be statically expanded.
4066 // Indeed, we do not want to introduce any new kind of transitions.
4067 for (const Use &U : ToBePromoted->operands()) {
4068 const Value *Val = U.get();
4069 if (Val == getEndOfTransition()) {
4070 // If the use is a division and the transition is on the rhs,
4071 // we cannot promote the operation, otherwise we may create a
4072 // division by zero.
4073 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4077 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4078 !isa<ConstantFP>(Val))
4081 // Check that the resulting operation is legal.
4082 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4085 return StressStoreExtract ||
4086 TLI.isOperationLegalOrCustom(
4087 ISDOpcode, TLI.getValueType(getTransitionType(), true));
4090 /// \brief Check whether or not \p Use can be combined
4091 /// with the transition.
4092 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4093 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4095 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4096 void enqueueForPromotion(Instruction *ToBePromoted) {
4097 InstsToBePromoted.push_back(ToBePromoted);
4100 /// \brief Set the instruction that will be combined with the transition.
4101 void recordCombineInstruction(Instruction *ToBeCombined) {
4102 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4103 CombineInst = ToBeCombined;
4106 /// \brief Promote all the instructions enqueued for promotion if it is
4108 /// \return True if the promotion happened, false otherwise.
4110 // Check if there is something to promote.
4111 // Right now, if we do not have anything to combine with,
4112 // we assume the promotion is not profitable.
4113 if (InstsToBePromoted.empty() || !CombineInst)
4117 if (!StressStoreExtract && !isProfitableToPromote())
4121 for (auto &ToBePromoted : InstsToBePromoted)
4122 promoteImpl(ToBePromoted);
4123 InstsToBePromoted.clear();
4127 } // End of anonymous namespace.
4129 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4130 // At this point, we know that all the operands of ToBePromoted but Def
4131 // can be statically promoted.
4132 // For Def, we need to use its parameter in ToBePromoted:
4133 // b = ToBePromoted ty1 a
4134 // Def = Transition ty1 b to ty2
4135 // Move the transition down.
4136 // 1. Replace all uses of the promoted operation by the transition.
4137 // = ... b => = ... Def.
4138 assert(ToBePromoted->getType() == Transition->getType() &&
4139 "The type of the result of the transition does not match "
4141 ToBePromoted->replaceAllUsesWith(Transition);
4142 // 2. Update the type of the uses.
4143 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
4144 Type *TransitionTy = getTransitionType();
4145 ToBePromoted->mutateType(TransitionTy);
4146 // 3. Update all the operands of the promoted operation with promoted
4148 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
4149 for (Use &U : ToBePromoted->operands()) {
4150 Value *Val = U.get();
4151 Value *NewVal = nullptr;
4152 if (Val == Transition)
4153 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
4154 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
4155 isa<ConstantFP>(Val)) {
4156 // Use a splat constant if it is not safe to use undef.
4157 NewVal = getConstantVector(
4158 cast<Constant>(Val),
4159 isa<UndefValue>(Val) ||
4160 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
4162 llvm_unreachable("Did you modified shouldPromote and forgot to update "
4164 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
4166 Transition->removeFromParent();
4167 Transition->insertAfter(ToBePromoted);
4168 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
4171 /// Some targets can do store(extractelement) with one instruction.
4172 /// Try to push the extractelement towards the stores when the target
4173 /// has this feature and this is profitable.
4174 bool CodeGenPrepare::OptimizeExtractElementInst(Instruction *Inst) {
4175 unsigned CombineCost = UINT_MAX;
4176 if (DisableStoreExtract || !TLI ||
4177 (!StressStoreExtract &&
4178 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4179 Inst->getOperand(1), CombineCost)))
4182 // At this point we know that Inst is a vector to scalar transition.
4183 // Try to move it down the def-use chain, until:
4184 // - We can combine the transition with its single use
4185 // => we got rid of the transition.
4186 // - We escape the current basic block
4187 // => we would need to check that we are moving it at a cheaper place and
4188 // we do not do that for now.
4189 BasicBlock *Parent = Inst->getParent();
4190 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4191 VectorPromoteHelper VPH(*TLI, *TTI, Inst, CombineCost);
4192 // If the transition has more than one use, assume this is not going to be
4194 while (Inst->hasOneUse()) {
4195 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4196 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4198 if (ToBePromoted->getParent() != Parent) {
4199 DEBUG(dbgs() << "Instruction to promote is in a different block ("
4200 << ToBePromoted->getParent()->getName()
4201 << ") than the transition (" << Parent->getName() << ").\n");
4205 if (VPH.canCombine(ToBePromoted)) {
4206 DEBUG(dbgs() << "Assume " << *Inst << '\n'
4207 << "will be combined with: " << *ToBePromoted << '\n');
4208 VPH.recordCombineInstruction(ToBePromoted);
4209 bool Changed = VPH.promote();
4210 NumStoreExtractExposed += Changed;
4214 DEBUG(dbgs() << "Try promoting.\n");
4215 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4218 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4220 VPH.enqueueForPromotion(ToBePromoted);
4221 Inst = ToBePromoted;
4226 bool CodeGenPrepare::OptimizeInst(Instruction *I, bool& ModifiedDT) {
4227 if (PHINode *P = dyn_cast<PHINode>(I)) {
4228 // It is possible for very late stage optimizations (such as SimplifyCFG)
4229 // to introduce PHI nodes too late to be cleaned up. If we detect such a
4230 // trivial PHI, go ahead and zap it here.
4231 const DataLayout &DL = I->getModule()->getDataLayout();
4232 if (Value *V = SimplifyInstruction(P, DL, TLInfo, nullptr)) {
4233 P->replaceAllUsesWith(V);
4234 P->eraseFromParent();
4241 if (CastInst *CI = dyn_cast<CastInst>(I)) {
4242 // If the source of the cast is a constant, then this should have
4243 // already been constant folded. The only reason NOT to constant fold
4244 // it is if something (e.g. LSR) was careful to place the constant
4245 // evaluation in a block other than then one that uses it (e.g. to hoist
4246 // the address of globals out of a loop). If this is the case, we don't
4247 // want to forward-subst the cast.
4248 if (isa<Constant>(CI->getOperand(0)))
4251 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
4254 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4255 /// Sink a zext or sext into its user blocks if the target type doesn't
4256 /// fit in one register
4257 if (TLI && TLI->getTypeAction(CI->getContext(),
4258 TLI->getValueType(CI->getType())) ==
4259 TargetLowering::TypeExpandInteger) {
4260 return SinkCast(CI);
4262 bool MadeChange = MoveExtToFormExtLoad(I);
4263 return MadeChange | OptimizeExtUses(I);
4269 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4270 if (!TLI || !TLI->hasMultipleConditionRegisters())
4271 return OptimizeCmpExpression(CI);
4273 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4275 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
4279 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4281 return OptimizeMemoryInst(I, SI->getOperand(1),
4282 SI->getOperand(0)->getType());
4286 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4288 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4289 BinOp->getOpcode() == Instruction::LShr)) {
4290 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4291 if (TLI && CI && TLI->hasExtractBitsInsn())
4292 return OptimizeExtractBits(BinOp, CI, *TLI);
4297 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4298 if (GEPI->hasAllZeroIndices()) {
4299 /// The GEP operand must be a pointer, so must its result -> BitCast
4300 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4301 GEPI->getName(), GEPI);
4302 GEPI->replaceAllUsesWith(NC);
4303 GEPI->eraseFromParent();
4305 OptimizeInst(NC, ModifiedDT);
4311 if (CallInst *CI = dyn_cast<CallInst>(I))
4312 return OptimizeCallInst(CI, ModifiedDT);
4314 if (SelectInst *SI = dyn_cast<SelectInst>(I))
4315 return OptimizeSelectInst(SI);
4317 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
4318 return OptimizeShuffleVectorInst(SVI);
4320 if (isa<ExtractElementInst>(I))
4321 return OptimizeExtractElementInst(I);
4326 // In this pass we look for GEP and cast instructions that are used
4327 // across basic blocks and rewrite them to improve basic-block-at-a-time
4329 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
4331 bool MadeChange = false;
4333 CurInstIterator = BB.begin();
4334 while (CurInstIterator != BB.end()) {
4335 MadeChange |= OptimizeInst(CurInstIterator++, ModifiedDT);
4339 MadeChange |= DupRetToEnableTailCallOpts(&BB);
4344 // llvm.dbg.value is far away from the value then iSel may not be able
4345 // handle it properly. iSel will drop llvm.dbg.value if it can not
4346 // find a node corresponding to the value.
4347 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
4348 bool MadeChange = false;
4349 for (BasicBlock &BB : F) {
4350 Instruction *PrevNonDbgInst = nullptr;
4351 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
4352 Instruction *Insn = BI++;
4353 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
4354 // Leave dbg.values that refer to an alloca alone. These
4355 // instrinsics describe the address of a variable (= the alloca)
4356 // being taken. They should not be moved next to the alloca
4357 // (and to the beginning of the scope), but rather stay close to
4358 // where said address is used.
4359 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
4360 PrevNonDbgInst = Insn;
4364 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
4365 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
4366 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
4367 DVI->removeFromParent();
4368 if (isa<PHINode>(VI))
4369 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
4371 DVI->insertAfter(VI);
4380 // If there is a sequence that branches based on comparing a single bit
4381 // against zero that can be combined into a single instruction, and the
4382 // target supports folding these into a single instruction, sink the
4383 // mask and compare into the branch uses. Do this before OptimizeBlock ->
4384 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
4386 bool CodeGenPrepare::sinkAndCmp(Function &F) {
4387 if (!EnableAndCmpSinking)
4389 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
4391 bool MadeChange = false;
4392 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
4393 BasicBlock *BB = I++;
4395 // Does this BB end with the following?
4396 // %andVal = and %val, #single-bit-set
4397 // %icmpVal = icmp %andResult, 0
4398 // br i1 %cmpVal label %dest1, label %dest2"
4399 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
4400 if (!Brcc || !Brcc->isConditional())
4402 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
4403 if (!Cmp || Cmp->getParent() != BB)
4405 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
4406 if (!Zero || !Zero->isZero())
4408 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
4409 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
4411 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
4412 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
4414 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
4416 // Push the "and; icmp" for any users that are conditional branches.
4417 // Since there can only be one branch use per BB, we don't need to keep
4418 // track of which BBs we insert into.
4419 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
4423 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
4425 if (!BrccUser || !BrccUser->isConditional())
4427 BasicBlock *UserBB = BrccUser->getParent();
4428 if (UserBB == BB) continue;
4429 DEBUG(dbgs() << "found Brcc use\n");
4431 // Sink the "and; icmp" to use.
4433 BinaryOperator *NewAnd =
4434 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
4437 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
4441 DEBUG(BrccUser->getParent()->dump());
4447 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
4448 /// success, or returns false if no or invalid metadata was found.
4449 static bool extractBranchMetadata(BranchInst *BI,
4450 uint64_t &ProbTrue, uint64_t &ProbFalse) {
4451 assert(BI->isConditional() &&
4452 "Looking for probabilities on unconditional branch?");
4453 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
4454 if (!ProfileData || ProfileData->getNumOperands() != 3)
4457 const auto *CITrue =
4458 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
4459 const auto *CIFalse =
4460 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
4461 if (!CITrue || !CIFalse)
4464 ProbTrue = CITrue->getValue().getZExtValue();
4465 ProbFalse = CIFalse->getValue().getZExtValue();
4470 /// \brief Scale down both weights to fit into uint32_t.
4471 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
4472 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
4473 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
4474 NewTrue = NewTrue / Scale;
4475 NewFalse = NewFalse / Scale;
4478 /// \brief Some targets prefer to split a conditional branch like:
4480 /// %0 = icmp ne i32 %a, 0
4481 /// %1 = icmp ne i32 %b, 0
4482 /// %or.cond = or i1 %0, %1
4483 /// br i1 %or.cond, label %TrueBB, label %FalseBB
4485 /// into multiple branch instructions like:
4488 /// %0 = icmp ne i32 %a, 0
4489 /// br i1 %0, label %TrueBB, label %bb2
4491 /// %1 = icmp ne i32 %b, 0
4492 /// br i1 %1, label %TrueBB, label %FalseBB
4494 /// This usually allows instruction selection to do even further optimizations
4495 /// and combine the compare with the branch instruction. Currently this is
4496 /// applied for targets which have "cheap" jump instructions.
4498 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
4500 bool CodeGenPrepare::splitBranchCondition(Function &F) {
4501 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
4504 bool MadeChange = false;
4505 for (auto &BB : F) {
4506 // Does this BB end with the following?
4507 // %cond1 = icmp|fcmp|binary instruction ...
4508 // %cond2 = icmp|fcmp|binary instruction ...
4509 // %cond.or = or|and i1 %cond1, cond2
4510 // br i1 %cond.or label %dest1, label %dest2"
4511 BinaryOperator *LogicOp;
4512 BasicBlock *TBB, *FBB;
4513 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
4517 Value *Cond1, *Cond2;
4518 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
4519 m_OneUse(m_Value(Cond2)))))
4520 Opc = Instruction::And;
4521 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
4522 m_OneUse(m_Value(Cond2)))))
4523 Opc = Instruction::Or;
4527 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
4528 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
4531 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
4534 auto *InsertBefore = std::next(Function::iterator(BB))
4535 .getNodePtrUnchecked();
4536 auto TmpBB = BasicBlock::Create(BB.getContext(),
4537 BB.getName() + ".cond.split",
4538 BB.getParent(), InsertBefore);
4540 // Update original basic block by using the first condition directly by the
4541 // branch instruction and removing the no longer needed and/or instruction.
4542 auto *Br1 = cast<BranchInst>(BB.getTerminator());
4543 Br1->setCondition(Cond1);
4544 LogicOp->eraseFromParent();
4546 // Depending on the conditon we have to either replace the true or the false
4547 // successor of the original branch instruction.
4548 if (Opc == Instruction::And)
4549 Br1->setSuccessor(0, TmpBB);
4551 Br1->setSuccessor(1, TmpBB);
4553 // Fill in the new basic block.
4554 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
4555 if (auto *I = dyn_cast<Instruction>(Cond2)) {
4556 I->removeFromParent();
4557 I->insertBefore(Br2);
4560 // Update PHI nodes in both successors. The original BB needs to be
4561 // replaced in one succesor's PHI nodes, because the branch comes now from
4562 // the newly generated BB (NewBB). In the other successor we need to add one
4563 // incoming edge to the PHI nodes, because both branch instructions target
4564 // now the same successor. Depending on the original branch condition
4565 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
4566 // we perfrom the correct update for the PHI nodes.
4567 // This doesn't change the successor order of the just created branch
4568 // instruction (or any other instruction).
4569 if (Opc == Instruction::Or)
4570 std::swap(TBB, FBB);
4572 // Replace the old BB with the new BB.
4573 for (auto &I : *TBB) {
4574 PHINode *PN = dyn_cast<PHINode>(&I);
4578 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
4579 PN->setIncomingBlock(i, TmpBB);
4582 // Add another incoming edge form the new BB.
4583 for (auto &I : *FBB) {
4584 PHINode *PN = dyn_cast<PHINode>(&I);
4587 auto *Val = PN->getIncomingValueForBlock(&BB);
4588 PN->addIncoming(Val, TmpBB);
4591 // Update the branch weights (from SelectionDAGBuilder::
4592 // FindMergedConditions).
4593 if (Opc == Instruction::Or) {
4594 // Codegen X | Y as:
4603 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
4604 // The requirement is that
4605 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
4606 // = TrueProb for orignal BB.
4607 // Assuming the orignal weights are A and B, one choice is to set BB1's
4608 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
4610 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
4611 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
4612 // TmpBB, but the math is more complicated.
4613 uint64_t TrueWeight, FalseWeight;
4614 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4615 uint64_t NewTrueWeight = TrueWeight;
4616 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
4617 scaleWeights(NewTrueWeight, NewFalseWeight);
4618 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4619 .createBranchWeights(TrueWeight, FalseWeight));
4621 NewTrueWeight = TrueWeight;
4622 NewFalseWeight = 2 * FalseWeight;
4623 scaleWeights(NewTrueWeight, NewFalseWeight);
4624 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4625 .createBranchWeights(TrueWeight, FalseWeight));
4628 // Codegen X & Y as:
4636 // This requires creation of TmpBB after CurBB.
4638 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
4639 // The requirement is that
4640 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
4641 // = FalseProb for orignal BB.
4642 // Assuming the orignal weights are A and B, one choice is to set BB1's
4643 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
4645 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
4646 uint64_t TrueWeight, FalseWeight;
4647 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4648 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
4649 uint64_t NewFalseWeight = FalseWeight;
4650 scaleWeights(NewTrueWeight, NewFalseWeight);
4651 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4652 .createBranchWeights(TrueWeight, FalseWeight));
4654 NewTrueWeight = 2 * TrueWeight;
4655 NewFalseWeight = FalseWeight;
4656 scaleWeights(NewTrueWeight, NewFalseWeight);
4657 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4658 .createBranchWeights(TrueWeight, FalseWeight));
4662 // Note: No point in getting fancy here, since the DT info is never
4663 // available to CodeGenPrepare.
4668 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();