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 /// CombineUAddWithOverflow - try to combine CI into a call to the
751 /// llvm.uadd.with.overflow intrinsic if possible.
753 /// Return true if any changes were made.
754 static bool CombineUAddWithOverflow(CmpInst *CI) {
758 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
761 Type *Ty = AddI->getType();
762 if (!isa<IntegerType>(Ty))
765 // We don't want to move around uses of condition values this late, so we we
766 // check if it is legal to create the call to the intrinsic in the basic
767 // block containing the icmp:
769 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
773 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
775 if (AddI->hasOneUse())
776 assert(*AddI->user_begin() == CI && "expected!");
779 Module *M = CI->getParent()->getParent()->getParent();
780 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
782 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
784 auto *UAddWithOverflow =
785 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
786 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
788 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
790 CI->replaceAllUsesWith(Overflow);
791 AddI->replaceAllUsesWith(UAdd);
792 CI->eraseFromParent();
793 AddI->eraseFromParent();
797 /// SinkCmpExpression - Sink the given CmpInst into user blocks to reduce
798 /// the number of virtual registers that must be created and coalesced. This is
799 /// a clear win except on targets with multiple condition code registers
800 /// (PowerPC), where it might lose; some adjustment may be wanted there.
802 /// Return true if any changes are made.
803 static bool SinkCmpExpression(CmpInst *CI) {
804 BasicBlock *DefBB = CI->getParent();
806 /// InsertedCmp - Only insert a cmp in each block once.
807 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
809 bool MadeChange = false;
810 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
812 Use &TheUse = UI.getUse();
813 Instruction *User = cast<Instruction>(*UI);
815 // Preincrement use iterator so we don't invalidate it.
818 // Don't bother for PHI nodes.
819 if (isa<PHINode>(User))
822 // Figure out which BB this cmp is used in.
823 BasicBlock *UserBB = User->getParent();
825 // If this user is in the same block as the cmp, don't change the cmp.
826 if (UserBB == DefBB) continue;
828 // If we have already inserted a cmp into this block, use it.
829 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
832 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
834 CmpInst::Create(CI->getOpcode(),
835 CI->getPredicate(), CI->getOperand(0),
836 CI->getOperand(1), "", InsertPt);
840 // Replace a use of the cmp with a use of the new cmp.
841 TheUse = InsertedCmp;
845 // If we removed all uses, nuke the cmp.
847 CI->eraseFromParent();
852 static bool OptimizeCmpExpression(CmpInst *CI) {
853 if (SinkCmpExpression(CI))
856 if (CombineUAddWithOverflow(CI))
862 /// isExtractBitsCandidateUse - Check if the candidates could
863 /// be combined with shift instruction, which includes:
864 /// 1. Truncate instruction
865 /// 2. And instruction and the imm is a mask of the low bits:
866 /// imm & (imm+1) == 0
867 static bool isExtractBitsCandidateUse(Instruction *User) {
868 if (!isa<TruncInst>(User)) {
869 if (User->getOpcode() != Instruction::And ||
870 !isa<ConstantInt>(User->getOperand(1)))
873 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
875 if ((Cimm & (Cimm + 1)).getBoolValue())
881 /// SinkShiftAndTruncate - sink both shift and truncate instruction
882 /// to the use of truncate's BB.
884 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
885 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
886 const TargetLowering &TLI) {
887 BasicBlock *UserBB = User->getParent();
888 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
889 TruncInst *TruncI = dyn_cast<TruncInst>(User);
890 bool MadeChange = false;
892 for (Value::user_iterator TruncUI = TruncI->user_begin(),
893 TruncE = TruncI->user_end();
894 TruncUI != TruncE;) {
896 Use &TruncTheUse = TruncUI.getUse();
897 Instruction *TruncUser = cast<Instruction>(*TruncUI);
898 // Preincrement use iterator so we don't invalidate it.
902 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
906 // If the use is actually a legal node, there will not be an
907 // implicit truncate.
908 // FIXME: always querying the result type is just an
909 // approximation; some nodes' legality is determined by the
910 // operand or other means. There's no good way to find out though.
911 if (TLI.isOperationLegalOrCustom(
912 ISDOpcode, TLI.getValueType(TruncUser->getType(), true)))
915 // Don't bother for PHI nodes.
916 if (isa<PHINode>(TruncUser))
919 BasicBlock *TruncUserBB = TruncUser->getParent();
921 if (UserBB == TruncUserBB)
924 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
925 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
927 if (!InsertedShift && !InsertedTrunc) {
928 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
930 if (ShiftI->getOpcode() == Instruction::AShr)
932 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
935 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
938 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
941 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
942 TruncI->getType(), "", TruncInsertPt);
946 TruncTheUse = InsertedTrunc;
952 /// OptimizeExtractBits - sink the shift *right* instruction into user blocks if
953 /// the uses could potentially be combined with this shift instruction and
954 /// generate BitExtract instruction. It will only be applied if the architecture
955 /// supports BitExtract instruction. Here is an example:
957 /// %x.extract.shift = lshr i64 %arg1, 32
959 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
963 /// %x.extract.shift.1 = lshr i64 %arg1, 32
964 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
966 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
968 /// Return true if any changes are made.
969 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
970 const TargetLowering &TLI) {
971 BasicBlock *DefBB = ShiftI->getParent();
973 /// Only insert instructions in each block once.
974 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
976 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(ShiftI->getType()));
978 bool MadeChange = false;
979 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
981 Use &TheUse = UI.getUse();
982 Instruction *User = cast<Instruction>(*UI);
983 // Preincrement use iterator so we don't invalidate it.
986 // Don't bother for PHI nodes.
987 if (isa<PHINode>(User))
990 if (!isExtractBitsCandidateUse(User))
993 BasicBlock *UserBB = User->getParent();
995 if (UserBB == DefBB) {
996 // If the shift and truncate instruction are in the same BB. The use of
997 // the truncate(TruncUse) may still introduce another truncate if not
998 // legal. In this case, we would like to sink both shift and truncate
999 // instruction to the BB of TruncUse.
1002 // i64 shift.result = lshr i64 opnd, imm
1003 // trunc.result = trunc shift.result to i16
1006 // ----> We will have an implicit truncate here if the architecture does
1007 // not have i16 compare.
1008 // cmp i16 trunc.result, opnd2
1010 if (isa<TruncInst>(User) && shiftIsLegal
1011 // If the type of the truncate is legal, no trucate will be
1012 // introduced in other basic blocks.
1013 && (!TLI.isTypeLegal(TLI.getValueType(User->getType()))))
1015 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI);
1019 // If we have already inserted a shift into this block, use it.
1020 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1022 if (!InsertedShift) {
1023 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1025 if (ShiftI->getOpcode() == Instruction::AShr)
1027 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
1030 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
1035 // Replace a use of the shift with a use of the new shift.
1036 TheUse = InsertedShift;
1039 // If we removed all uses, nuke the shift.
1040 if (ShiftI->use_empty())
1041 ShiftI->eraseFromParent();
1046 // ScalarizeMaskedLoad() translates masked load intrinsic, like
1047 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1048 // <16 x i1> %mask, <16 x i32> %passthru)
1049 // to a chain of basic blocks, whith loading element one-by-one if
1050 // the appropriate mask bit is set
1052 // %1 = bitcast i8* %addr to i32*
1053 // %2 = extractelement <16 x i1> %mask, i32 0
1054 // %3 = icmp eq i1 %2, true
1055 // br i1 %3, label %cond.load, label %else
1057 //cond.load: ; preds = %0
1058 // %4 = getelementptr i32* %1, i32 0
1059 // %5 = load i32* %4
1060 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1063 //else: ; preds = %0, %cond.load
1064 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1065 // %7 = extractelement <16 x i1> %mask, i32 1
1066 // %8 = icmp eq i1 %7, true
1067 // br i1 %8, label %cond.load1, label %else2
1069 //cond.load1: ; preds = %else
1070 // %9 = getelementptr i32* %1, i32 1
1071 // %10 = load i32* %9
1072 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1075 //else2: ; preds = %else, %cond.load1
1076 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1077 // %12 = extractelement <16 x i1> %mask, i32 2
1078 // %13 = icmp eq i1 %12, true
1079 // br i1 %13, label %cond.load4, label %else5
1081 static void ScalarizeMaskedLoad(CallInst *CI) {
1082 Value *Ptr = CI->getArgOperand(0);
1083 Value *Src0 = CI->getArgOperand(3);
1084 Value *Mask = CI->getArgOperand(2);
1085 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1086 Type *EltTy = VecType->getElementType();
1088 assert(VecType && "Unexpected return type of masked load intrinsic");
1090 IRBuilder<> Builder(CI->getContext());
1091 Instruction *InsertPt = CI;
1092 BasicBlock *IfBlock = CI->getParent();
1093 BasicBlock *CondBlock = nullptr;
1094 BasicBlock *PrevIfBlock = CI->getParent();
1095 Builder.SetInsertPoint(InsertPt);
1097 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1099 // Bitcast %addr fron i8* to EltTy*
1101 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1102 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1103 Value *UndefVal = UndefValue::get(VecType);
1105 // The result vector
1106 Value *VResult = UndefVal;
1108 PHINode *Phi = nullptr;
1109 Value *PrevPhi = UndefVal;
1111 unsigned VectorWidth = VecType->getNumElements();
1112 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1114 // Fill the "else" block, created in the previous iteration
1116 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1117 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1118 // %to_load = icmp eq i1 %mask_1, true
1119 // br i1 %to_load, label %cond.load, label %else
1122 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1123 Phi->addIncoming(VResult, CondBlock);
1124 Phi->addIncoming(PrevPhi, PrevIfBlock);
1129 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1130 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1131 ConstantInt::get(Predicate->getType(), 1));
1133 // Create "cond" block
1135 // %EltAddr = getelementptr i32* %1, i32 0
1136 // %Elt = load i32* %EltAddr
1137 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1139 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1140 Builder.SetInsertPoint(InsertPt);
1143 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1144 LoadInst* Load = Builder.CreateLoad(Gep, false);
1145 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1147 // Create "else" block, fill it in the next iteration
1148 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1149 Builder.SetInsertPoint(InsertPt);
1150 Instruction *OldBr = IfBlock->getTerminator();
1151 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1152 OldBr->eraseFromParent();
1153 PrevIfBlock = IfBlock;
1154 IfBlock = NewIfBlock;
1157 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1158 Phi->addIncoming(VResult, CondBlock);
1159 Phi->addIncoming(PrevPhi, PrevIfBlock);
1160 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1161 CI->replaceAllUsesWith(NewI);
1162 CI->eraseFromParent();
1165 // ScalarizeMaskedStore() translates masked store intrinsic, like
1166 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1168 // to a chain of basic blocks, that stores element one-by-one if
1169 // the appropriate mask bit is set
1171 // %1 = bitcast i8* %addr to i32*
1172 // %2 = extractelement <16 x i1> %mask, i32 0
1173 // %3 = icmp eq i1 %2, true
1174 // br i1 %3, label %cond.store, label %else
1176 // cond.store: ; preds = %0
1177 // %4 = extractelement <16 x i32> %val, i32 0
1178 // %5 = getelementptr i32* %1, i32 0
1179 // store i32 %4, i32* %5
1182 // else: ; preds = %0, %cond.store
1183 // %6 = extractelement <16 x i1> %mask, i32 1
1184 // %7 = icmp eq i1 %6, true
1185 // br i1 %7, label %cond.store1, label %else2
1187 // cond.store1: ; preds = %else
1188 // %8 = extractelement <16 x i32> %val, i32 1
1189 // %9 = getelementptr i32* %1, i32 1
1190 // store i32 %8, i32* %9
1193 static void ScalarizeMaskedStore(CallInst *CI) {
1194 Value *Ptr = CI->getArgOperand(1);
1195 Value *Src = CI->getArgOperand(0);
1196 Value *Mask = CI->getArgOperand(3);
1198 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1199 Type *EltTy = VecType->getElementType();
1201 assert(VecType && "Unexpected data type in masked store intrinsic");
1203 IRBuilder<> Builder(CI->getContext());
1204 Instruction *InsertPt = CI;
1205 BasicBlock *IfBlock = CI->getParent();
1206 Builder.SetInsertPoint(InsertPt);
1207 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1209 // Bitcast %addr fron i8* to EltTy*
1211 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1212 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1214 unsigned VectorWidth = VecType->getNumElements();
1215 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1217 // Fill the "else" block, created in the previous iteration
1219 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1220 // %to_store = icmp eq i1 %mask_1, true
1221 // br i1 %to_load, label %cond.store, label %else
1223 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1224 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1225 ConstantInt::get(Predicate->getType(), 1));
1227 // Create "cond" block
1229 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1230 // %EltAddr = getelementptr i32* %1, i32 0
1231 // %store i32 %OneElt, i32* %EltAddr
1233 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1234 Builder.SetInsertPoint(InsertPt);
1236 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1238 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1239 Builder.CreateStore(OneElt, Gep);
1241 // Create "else" block, fill it in the next iteration
1242 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1243 Builder.SetInsertPoint(InsertPt);
1244 Instruction *OldBr = IfBlock->getTerminator();
1245 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1246 OldBr->eraseFromParent();
1247 IfBlock = NewIfBlock;
1249 CI->eraseFromParent();
1252 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1253 BasicBlock *BB = CI->getParent();
1255 // Lower inline assembly if we can.
1256 // If we found an inline asm expession, and if the target knows how to
1257 // lower it to normal LLVM code, do so now.
1258 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1259 if (TLI->ExpandInlineAsm(CI)) {
1260 // Avoid invalidating the iterator.
1261 CurInstIterator = BB->begin();
1262 // Avoid processing instructions out of order, which could cause
1263 // reuse before a value is defined.
1267 // Sink address computing for memory operands into the block.
1268 if (OptimizeInlineAsmInst(CI))
1272 const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
1274 // Align the pointer arguments to this call if the target thinks it's a good
1276 unsigned MinSize, PrefAlign;
1277 if (TLI && TD && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1278 for (auto &Arg : CI->arg_operands()) {
1279 // We want to align both objects whose address is used directly and
1280 // objects whose address is used in casts and GEPs, though it only makes
1281 // sense for GEPs if the offset is a multiple of the desired alignment and
1282 // if size - offset meets the size threshold.
1283 if (!Arg->getType()->isPointerTy())
1285 APInt Offset(TD->getPointerSizeInBits(
1286 cast<PointerType>(Arg->getType())->getAddressSpace()), 0);
1287 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*TD, Offset);
1288 uint64_t Offset2 = Offset.getLimitedValue();
1290 if ((Offset2 & (PrefAlign-1)) == 0 &&
1291 (AI = dyn_cast<AllocaInst>(Val)) &&
1292 AI->getAlignment() < PrefAlign &&
1293 TD->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1294 AI->setAlignment(PrefAlign);
1295 // TODO: Also align GlobalVariables
1297 // If this is a memcpy (or similar) then we may be able to improve the
1299 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1300 unsigned Align = getKnownAlignment(MI->getDest(), *TD);
1301 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1302 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *TD));
1303 if (Align > MI->getAlignment())
1304 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1308 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1310 switch (II->getIntrinsicID()) {
1312 case Intrinsic::objectsize: {
1313 // Lower all uses of llvm.objectsize.*
1314 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1315 Type *ReturnTy = CI->getType();
1316 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1318 // Substituting this can cause recursive simplifications, which can
1319 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1321 WeakVH IterHandle(CurInstIterator);
1323 replaceAndRecursivelySimplify(CI, RetVal,
1326 // If the iterator instruction was recursively deleted, start over at the
1327 // start of the block.
1328 if (IterHandle != CurInstIterator) {
1329 CurInstIterator = BB->begin();
1334 case Intrinsic::masked_load: {
1335 // Scalarize unsupported vector masked load
1336 if (!TTI->isLegalMaskedLoad(CI->getType(), 1)) {
1337 ScalarizeMaskedLoad(CI);
1343 case Intrinsic::masked_store: {
1344 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType(), 1)) {
1345 ScalarizeMaskedStore(CI);
1354 SmallVector<Value*, 2> PtrOps;
1356 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
1357 while (!PtrOps.empty())
1358 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
1363 // From here on out we're working with named functions.
1364 if (!CI->getCalledFunction()) return false;
1366 // Lower all default uses of _chk calls. This is very similar
1367 // to what InstCombineCalls does, but here we are only lowering calls
1368 // to fortified library functions (e.g. __memcpy_chk) that have the default
1369 // "don't know" as the objectsize. Anything else should be left alone.
1370 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1371 if (Value *V = Simplifier.optimizeCall(CI)) {
1372 CI->replaceAllUsesWith(V);
1373 CI->eraseFromParent();
1379 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
1380 /// instructions to the predecessor to enable tail call optimizations. The
1381 /// case it is currently looking for is:
1384 /// %tmp0 = tail call i32 @f0()
1385 /// br label %return
1387 /// %tmp1 = tail call i32 @f1()
1388 /// br label %return
1390 /// %tmp2 = tail call i32 @f2()
1391 /// br label %return
1393 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1401 /// %tmp0 = tail call i32 @f0()
1404 /// %tmp1 = tail call i32 @f1()
1407 /// %tmp2 = tail call i32 @f2()
1410 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
1414 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1418 PHINode *PN = nullptr;
1419 BitCastInst *BCI = nullptr;
1420 Value *V = RI->getReturnValue();
1422 BCI = dyn_cast<BitCastInst>(V);
1424 V = BCI->getOperand(0);
1426 PN = dyn_cast<PHINode>(V);
1431 if (PN && PN->getParent() != BB)
1434 // It's not safe to eliminate the sign / zero extension of the return value.
1435 // See llvm::isInTailCallPosition().
1436 const Function *F = BB->getParent();
1437 AttributeSet CallerAttrs = F->getAttributes();
1438 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1439 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1442 // Make sure there are no instructions between the PHI and return, or that the
1443 // return is the first instruction in the block.
1445 BasicBlock::iterator BI = BB->begin();
1446 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1448 // Also skip over the bitcast.
1453 BasicBlock::iterator BI = BB->begin();
1454 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1459 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1461 SmallVector<CallInst*, 4> TailCalls;
1463 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1464 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1465 // Make sure the phi value is indeed produced by the tail call.
1466 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1467 TLI->mayBeEmittedAsTailCall(CI))
1468 TailCalls.push_back(CI);
1471 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1472 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1473 if (!VisitedBBs.insert(*PI).second)
1476 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1477 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1478 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1479 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1483 CallInst *CI = dyn_cast<CallInst>(&*RI);
1484 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1485 TailCalls.push_back(CI);
1489 bool Changed = false;
1490 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1491 CallInst *CI = TailCalls[i];
1494 // Conservatively require the attributes of the call to match those of the
1495 // return. Ignore noalias because it doesn't affect the call sequence.
1496 AttributeSet CalleeAttrs = CS.getAttributes();
1497 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1498 removeAttribute(Attribute::NoAlias) !=
1499 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1500 removeAttribute(Attribute::NoAlias))
1503 // Make sure the call instruction is followed by an unconditional branch to
1504 // the return block.
1505 BasicBlock *CallBB = CI->getParent();
1506 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1507 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1510 // Duplicate the return into CallBB.
1511 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1512 ModifiedDT = Changed = true;
1516 // If we eliminated all predecessors of the block, delete the block now.
1517 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1518 BB->eraseFromParent();
1523 //===----------------------------------------------------------------------===//
1524 // Memory Optimization
1525 //===----------------------------------------------------------------------===//
1529 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1530 /// which holds actual Value*'s for register values.
1531 struct ExtAddrMode : public TargetLowering::AddrMode {
1534 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1535 void print(raw_ostream &OS) const;
1538 bool operator==(const ExtAddrMode& O) const {
1539 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1540 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1541 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1546 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1552 void ExtAddrMode::print(raw_ostream &OS) const {
1553 bool NeedPlus = false;
1556 OS << (NeedPlus ? " + " : "")
1558 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1563 OS << (NeedPlus ? " + " : "")
1569 OS << (NeedPlus ? " + " : "")
1571 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1575 OS << (NeedPlus ? " + " : "")
1577 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1583 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1584 void ExtAddrMode::dump() const {
1590 /// \brief This class provides transaction based operation on the IR.
1591 /// Every change made through this class is recorded in the internal state and
1592 /// can be undone (rollback) until commit is called.
1593 class TypePromotionTransaction {
1595 /// \brief This represents the common interface of the individual transaction.
1596 /// Each class implements the logic for doing one specific modification on
1597 /// the IR via the TypePromotionTransaction.
1598 class TypePromotionAction {
1600 /// The Instruction modified.
1604 /// \brief Constructor of the action.
1605 /// The constructor performs the related action on the IR.
1606 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1608 virtual ~TypePromotionAction() {}
1610 /// \brief Undo the modification done by this action.
1611 /// When this method is called, the IR must be in the same state as it was
1612 /// before this action was applied.
1613 /// \pre Undoing the action works if and only if the IR is in the exact same
1614 /// state as it was directly after this action was applied.
1615 virtual void undo() = 0;
1617 /// \brief Advocate every change made by this action.
1618 /// When the results on the IR of the action are to be kept, it is important
1619 /// to call this function, otherwise hidden information may be kept forever.
1620 virtual void commit() {
1621 // Nothing to be done, this action is not doing anything.
1625 /// \brief Utility to remember the position of an instruction.
1626 class InsertionHandler {
1627 /// Position of an instruction.
1628 /// Either an instruction:
1629 /// - Is the first in a basic block: BB is used.
1630 /// - Has a previous instructon: PrevInst is used.
1632 Instruction *PrevInst;
1635 /// Remember whether or not the instruction had a previous instruction.
1636 bool HasPrevInstruction;
1639 /// \brief Record the position of \p Inst.
1640 InsertionHandler(Instruction *Inst) {
1641 BasicBlock::iterator It = Inst;
1642 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1643 if (HasPrevInstruction)
1644 Point.PrevInst = --It;
1646 Point.BB = Inst->getParent();
1649 /// \brief Insert \p Inst at the recorded position.
1650 void insert(Instruction *Inst) {
1651 if (HasPrevInstruction) {
1652 if (Inst->getParent())
1653 Inst->removeFromParent();
1654 Inst->insertAfter(Point.PrevInst);
1656 Instruction *Position = Point.BB->getFirstInsertionPt();
1657 if (Inst->getParent())
1658 Inst->moveBefore(Position);
1660 Inst->insertBefore(Position);
1665 /// \brief Move an instruction before another.
1666 class InstructionMoveBefore : public TypePromotionAction {
1667 /// Original position of the instruction.
1668 InsertionHandler Position;
1671 /// \brief Move \p Inst before \p Before.
1672 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1673 : TypePromotionAction(Inst), Position(Inst) {
1674 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1675 Inst->moveBefore(Before);
1678 /// \brief Move the instruction back to its original position.
1679 void undo() override {
1680 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1681 Position.insert(Inst);
1685 /// \brief Set the operand of an instruction with a new value.
1686 class OperandSetter : public TypePromotionAction {
1687 /// Original operand of the instruction.
1689 /// Index of the modified instruction.
1693 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1694 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1695 : TypePromotionAction(Inst), Idx(Idx) {
1696 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1697 << "for:" << *Inst << "\n"
1698 << "with:" << *NewVal << "\n");
1699 Origin = Inst->getOperand(Idx);
1700 Inst->setOperand(Idx, NewVal);
1703 /// \brief Restore the original value of the instruction.
1704 void undo() override {
1705 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1706 << "for: " << *Inst << "\n"
1707 << "with: " << *Origin << "\n");
1708 Inst->setOperand(Idx, Origin);
1712 /// \brief Hide the operands of an instruction.
1713 /// Do as if this instruction was not using any of its operands.
1714 class OperandsHider : public TypePromotionAction {
1715 /// The list of original operands.
1716 SmallVector<Value *, 4> OriginalValues;
1719 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1720 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1721 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1722 unsigned NumOpnds = Inst->getNumOperands();
1723 OriginalValues.reserve(NumOpnds);
1724 for (unsigned It = 0; It < NumOpnds; ++It) {
1725 // Save the current operand.
1726 Value *Val = Inst->getOperand(It);
1727 OriginalValues.push_back(Val);
1729 // We could use OperandSetter here, but that would implied an overhead
1730 // that we are not willing to pay.
1731 Inst->setOperand(It, UndefValue::get(Val->getType()));
1735 /// \brief Restore the original list of uses.
1736 void undo() override {
1737 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1738 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1739 Inst->setOperand(It, OriginalValues[It]);
1743 /// \brief Build a truncate instruction.
1744 class TruncBuilder : public TypePromotionAction {
1747 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1749 /// trunc Opnd to Ty.
1750 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1751 IRBuilder<> Builder(Opnd);
1752 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
1753 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
1756 /// \brief Get the built value.
1757 Value *getBuiltValue() { return Val; }
1759 /// \brief Remove the built instruction.
1760 void undo() override {
1761 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
1762 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1763 IVal->eraseFromParent();
1767 /// \brief Build a sign extension instruction.
1768 class SExtBuilder : public TypePromotionAction {
1771 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1773 /// sext Opnd to Ty.
1774 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1775 : TypePromotionAction(InsertPt) {
1776 IRBuilder<> Builder(InsertPt);
1777 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
1778 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
1781 /// \brief Get the built value.
1782 Value *getBuiltValue() { return Val; }
1784 /// \brief Remove the built instruction.
1785 void undo() override {
1786 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
1787 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1788 IVal->eraseFromParent();
1792 /// \brief Build a zero extension instruction.
1793 class ZExtBuilder : public TypePromotionAction {
1796 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
1798 /// zext Opnd to Ty.
1799 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1800 : TypePromotionAction(InsertPt) {
1801 IRBuilder<> Builder(InsertPt);
1802 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
1803 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
1806 /// \brief Get the built value.
1807 Value *getBuiltValue() { return Val; }
1809 /// \brief Remove the built instruction.
1810 void undo() override {
1811 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
1812 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1813 IVal->eraseFromParent();
1817 /// \brief Mutate an instruction to another type.
1818 class TypeMutator : public TypePromotionAction {
1819 /// Record the original type.
1823 /// \brief Mutate the type of \p Inst into \p NewTy.
1824 TypeMutator(Instruction *Inst, Type *NewTy)
1825 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1826 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1828 Inst->mutateType(NewTy);
1831 /// \brief Mutate the instruction back to its original type.
1832 void undo() override {
1833 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1835 Inst->mutateType(OrigTy);
1839 /// \brief Replace the uses of an instruction by another instruction.
1840 class UsesReplacer : public TypePromotionAction {
1841 /// Helper structure to keep track of the replaced uses.
1842 struct InstructionAndIdx {
1843 /// The instruction using the instruction.
1845 /// The index where this instruction is used for Inst.
1847 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1848 : Inst(Inst), Idx(Idx) {}
1851 /// Keep track of the original uses (pair Instruction, Index).
1852 SmallVector<InstructionAndIdx, 4> OriginalUses;
1853 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1856 /// \brief Replace all the use of \p Inst by \p New.
1857 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1858 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1860 // Record the original uses.
1861 for (Use &U : Inst->uses()) {
1862 Instruction *UserI = cast<Instruction>(U.getUser());
1863 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1865 // Now, we can replace the uses.
1866 Inst->replaceAllUsesWith(New);
1869 /// \brief Reassign the original uses of Inst to Inst.
1870 void undo() override {
1871 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1872 for (use_iterator UseIt = OriginalUses.begin(),
1873 EndIt = OriginalUses.end();
1874 UseIt != EndIt; ++UseIt) {
1875 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1880 /// \brief Remove an instruction from the IR.
1881 class InstructionRemover : public TypePromotionAction {
1882 /// Original position of the instruction.
1883 InsertionHandler Inserter;
1884 /// Helper structure to hide all the link to the instruction. In other
1885 /// words, this helps to do as if the instruction was removed.
1886 OperandsHider Hider;
1887 /// Keep track of the uses replaced, if any.
1888 UsesReplacer *Replacer;
1891 /// \brief Remove all reference of \p Inst and optinally replace all its
1893 /// \pre If !Inst->use_empty(), then New != nullptr
1894 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1895 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1898 Replacer = new UsesReplacer(Inst, New);
1899 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1900 Inst->removeFromParent();
1903 ~InstructionRemover() { delete Replacer; }
1905 /// \brief Really remove the instruction.
1906 void commit() override { delete Inst; }
1908 /// \brief Resurrect the instruction and reassign it to the proper uses if
1909 /// new value was provided when build this action.
1910 void undo() override {
1911 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1912 Inserter.insert(Inst);
1920 /// Restoration point.
1921 /// The restoration point is a pointer to an action instead of an iterator
1922 /// because the iterator may be invalidated but not the pointer.
1923 typedef const TypePromotionAction *ConstRestorationPt;
1924 /// Advocate every changes made in that transaction.
1926 /// Undo all the changes made after the given point.
1927 void rollback(ConstRestorationPt Point);
1928 /// Get the current restoration point.
1929 ConstRestorationPt getRestorationPoint() const;
1931 /// \name API for IR modification with state keeping to support rollback.
1933 /// Same as Instruction::setOperand.
1934 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1935 /// Same as Instruction::eraseFromParent.
1936 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
1937 /// Same as Value::replaceAllUsesWith.
1938 void replaceAllUsesWith(Instruction *Inst, Value *New);
1939 /// Same as Value::mutateType.
1940 void mutateType(Instruction *Inst, Type *NewTy);
1941 /// Same as IRBuilder::createTrunc.
1942 Value *createTrunc(Instruction *Opnd, Type *Ty);
1943 /// Same as IRBuilder::createSExt.
1944 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1945 /// Same as IRBuilder::createZExt.
1946 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
1947 /// Same as Instruction::moveBefore.
1948 void moveBefore(Instruction *Inst, Instruction *Before);
1952 /// The ordered list of actions made so far.
1953 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
1954 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
1957 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1960 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
1963 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1966 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
1969 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1971 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
1974 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1975 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
1978 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1980 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
1981 Value *Val = Ptr->getBuiltValue();
1982 Actions.push_back(std::move(Ptr));
1986 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
1987 Value *Opnd, Type *Ty) {
1988 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
1989 Value *Val = Ptr->getBuiltValue();
1990 Actions.push_back(std::move(Ptr));
1994 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
1995 Value *Opnd, Type *Ty) {
1996 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
1997 Value *Val = Ptr->getBuiltValue();
1998 Actions.push_back(std::move(Ptr));
2002 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2003 Instruction *Before) {
2005 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
2008 TypePromotionTransaction::ConstRestorationPt
2009 TypePromotionTransaction::getRestorationPoint() const {
2010 return !Actions.empty() ? Actions.back().get() : nullptr;
2013 void TypePromotionTransaction::commit() {
2014 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2020 void TypePromotionTransaction::rollback(
2021 TypePromotionTransaction::ConstRestorationPt Point) {
2022 while (!Actions.empty() && Point != Actions.back().get()) {
2023 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2028 /// \brief A helper class for matching addressing modes.
2030 /// This encapsulates the logic for matching the target-legal addressing modes.
2031 class AddressingModeMatcher {
2032 SmallVectorImpl<Instruction*> &AddrModeInsts;
2033 const TargetMachine &TM;
2034 const TargetLowering &TLI;
2036 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2037 /// the memory instruction that we're computing this address for.
2039 Instruction *MemoryInst;
2041 /// AddrMode - This is the addressing mode that we're building up. This is
2042 /// part of the return value of this addressing mode matching stuff.
2043 ExtAddrMode &AddrMode;
2045 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
2046 const SetOfInstrs &InsertedTruncs;
2047 /// A map from the instructions to their type before promotion.
2048 InstrToOrigTy &PromotedInsts;
2049 /// The ongoing transaction where every action should be registered.
2050 TypePromotionTransaction &TPT;
2052 /// IgnoreProfitability - This is set to true when we should not do
2053 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
2054 /// always returns true.
2055 bool IgnoreProfitability;
2057 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2058 const TargetMachine &TM, Type *AT, Instruction *MI,
2059 ExtAddrMode &AM, const SetOfInstrs &InsertedTruncs,
2060 InstrToOrigTy &PromotedInsts,
2061 TypePromotionTransaction &TPT)
2062 : AddrModeInsts(AMI), TM(TM),
2063 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
2064 ->getTargetLowering()),
2065 AccessTy(AT), MemoryInst(MI), AddrMode(AM),
2066 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
2067 IgnoreProfitability = false;
2071 /// Match - Find the maximal addressing mode that a load/store of V can fold,
2072 /// give an access type of AccessTy. This returns a list of involved
2073 /// instructions in AddrModeInsts.
2074 /// \p InsertedTruncs The truncate instruction inserted by other
2077 /// \p PromotedInsts maps the instructions to their type before promotion.
2078 /// \p The ongoing transaction where every action should be registered.
2079 static ExtAddrMode Match(Value *V, Type *AccessTy,
2080 Instruction *MemoryInst,
2081 SmallVectorImpl<Instruction*> &AddrModeInsts,
2082 const TargetMachine &TM,
2083 const SetOfInstrs &InsertedTruncs,
2084 InstrToOrigTy &PromotedInsts,
2085 TypePromotionTransaction &TPT) {
2088 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy,
2089 MemoryInst, Result, InsertedTruncs,
2090 PromotedInsts, TPT).MatchAddr(V, 0);
2091 (void)Success; assert(Success && "Couldn't select *anything*?");
2095 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2096 bool MatchAddr(Value *V, unsigned Depth);
2097 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2098 bool *MovedAway = nullptr);
2099 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
2100 ExtAddrMode &AMBefore,
2101 ExtAddrMode &AMAfter);
2102 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2103 bool IsPromotionProfitable(unsigned NewCost, unsigned OldCost,
2104 Value *PromotedOperand) const;
2107 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
2108 /// Return true and update AddrMode if this addr mode is legal for the target,
2110 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
2112 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2113 // mode. Just process that directly.
2115 return MatchAddr(ScaleReg, Depth);
2117 // If the scale is 0, it takes nothing to add this.
2121 // If we already have a scale of this value, we can add to it, otherwise, we
2122 // need an available scale field.
2123 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2126 ExtAddrMode TestAddrMode = AddrMode;
2128 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2129 // [A+B + A*7] -> [B+A*8].
2130 TestAddrMode.Scale += Scale;
2131 TestAddrMode.ScaledReg = ScaleReg;
2133 // If the new address isn't legal, bail out.
2134 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
2137 // It was legal, so commit it.
2138 AddrMode = TestAddrMode;
2140 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2141 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2142 // X*Scale + C*Scale to addr mode.
2143 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2144 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2145 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2146 TestAddrMode.ScaledReg = AddLHS;
2147 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2149 // If this addressing mode is legal, commit it and remember that we folded
2150 // this instruction.
2151 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
2152 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2153 AddrMode = TestAddrMode;
2158 // Otherwise, not (x+c)*scale, just return what we have.
2162 /// MightBeFoldableInst - This is a little filter, which returns true if an
2163 /// addressing computation involving I might be folded into a load/store
2164 /// accessing it. This doesn't need to be perfect, but needs to accept at least
2165 /// the set of instructions that MatchOperationAddr can.
2166 static bool MightBeFoldableInst(Instruction *I) {
2167 switch (I->getOpcode()) {
2168 case Instruction::BitCast:
2169 case Instruction::AddrSpaceCast:
2170 // Don't touch identity bitcasts.
2171 if (I->getType() == I->getOperand(0)->getType())
2173 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2174 case Instruction::PtrToInt:
2175 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2177 case Instruction::IntToPtr:
2178 // We know the input is intptr_t, so this is foldable.
2180 case Instruction::Add:
2182 case Instruction::Mul:
2183 case Instruction::Shl:
2184 // Can only handle X*C and X << C.
2185 return isa<ConstantInt>(I->getOperand(1));
2186 case Instruction::GetElementPtr:
2193 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2194 /// \note \p Val is assumed to be the product of some type promotion.
2195 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2196 /// to be legal, as the non-promoted value would have had the same state.
2197 static bool isPromotedInstructionLegal(const TargetLowering &TLI, Value *Val) {
2198 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2201 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2202 // If the ISDOpcode is undefined, it was undefined before the promotion.
2205 // Otherwise, check if the promoted instruction is legal or not.
2206 return TLI.isOperationLegalOrCustom(
2207 ISDOpcode, TLI.getValueType(PromotedInst->getType()));
2210 /// \brief Hepler class to perform type promotion.
2211 class TypePromotionHelper {
2212 /// \brief Utility function to check whether or not a sign or zero extension
2213 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2214 /// either using the operands of \p Inst or promoting \p Inst.
2215 /// The type of the extension is defined by \p IsSExt.
2216 /// In other words, check if:
2217 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2218 /// #1 Promotion applies:
2219 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2220 /// #2 Operand reuses:
2221 /// ext opnd1 to ConsideredExtType.
2222 /// \p PromotedInsts maps the instructions to their type before promotion.
2223 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2224 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2226 /// \brief Utility function to determine if \p OpIdx should be promoted when
2227 /// promoting \p Inst.
2228 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2229 if (isa<SelectInst>(Inst) && OpIdx == 0)
2234 /// \brief Utility function to promote the operand of \p Ext when this
2235 /// operand is a promotable trunc or sext or zext.
2236 /// \p PromotedInsts maps the instructions to their type before promotion.
2237 /// \p CreatedInstsCost[out] contains the cost of all instructions
2238 /// created to promote the operand of Ext.
2239 /// Newly added extensions are inserted in \p Exts.
2240 /// Newly added truncates are inserted in \p Truncs.
2241 /// Should never be called directly.
2242 /// \return The promoted value which is used instead of Ext.
2243 static Value *promoteOperandForTruncAndAnyExt(
2244 Instruction *Ext, TypePromotionTransaction &TPT,
2245 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2246 SmallVectorImpl<Instruction *> *Exts,
2247 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
2249 /// \brief Utility function to promote the operand of \p Ext when this
2250 /// operand is promotable and is not a supported trunc or sext.
2251 /// \p PromotedInsts maps the instructions to their type before promotion.
2252 /// \p CreatedInstsCost[out] contains the cost of all the instructions
2253 /// created to promote the operand of Ext.
2254 /// Newly added extensions are inserted in \p Exts.
2255 /// Newly added truncates are inserted in \p Truncs.
2256 /// Should never be called directly.
2257 /// \return The promoted value which is used instead of Ext.
2258 static Value *promoteOperandForOther(Instruction *Ext,
2259 TypePromotionTransaction &TPT,
2260 InstrToOrigTy &PromotedInsts,
2261 unsigned &CreatedInstsCost,
2262 SmallVectorImpl<Instruction *> *Exts,
2263 SmallVectorImpl<Instruction *> *Truncs,
2264 const TargetLowering &TLI, bool IsSExt);
2266 /// \see promoteOperandForOther.
2267 static Value *signExtendOperandForOther(
2268 Instruction *Ext, TypePromotionTransaction &TPT,
2269 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2270 SmallVectorImpl<Instruction *> *Exts,
2271 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2272 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2273 Exts, Truncs, TLI, true);
2276 /// \see promoteOperandForOther.
2277 static Value *zeroExtendOperandForOther(
2278 Instruction *Ext, TypePromotionTransaction &TPT,
2279 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2280 SmallVectorImpl<Instruction *> *Exts,
2281 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2282 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2283 Exts, Truncs, TLI, false);
2287 /// Type for the utility function that promotes the operand of Ext.
2288 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2289 InstrToOrigTy &PromotedInsts,
2290 unsigned &CreatedInstsCost,
2291 SmallVectorImpl<Instruction *> *Exts,
2292 SmallVectorImpl<Instruction *> *Truncs,
2293 const TargetLowering &TLI);
2294 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2295 /// action to promote the operand of \p Ext instead of using Ext.
2296 /// \return NULL if no promotable action is possible with the current
2298 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
2299 /// the others CodeGenPrepare optimizations. This information is important
2300 /// because we do not want to promote these instructions as CodeGenPrepare
2301 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2302 /// \p PromotedInsts maps the instructions to their type before promotion.
2303 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2304 const TargetLowering &TLI,
2305 const InstrToOrigTy &PromotedInsts);
2308 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2309 Type *ConsideredExtType,
2310 const InstrToOrigTy &PromotedInsts,
2312 // The promotion helper does not know how to deal with vector types yet.
2313 // To be able to fix that, we would need to fix the places where we
2314 // statically extend, e.g., constants and such.
2315 if (Inst->getType()->isVectorTy())
2318 // We can always get through zext.
2319 if (isa<ZExtInst>(Inst))
2322 // sext(sext) is ok too.
2323 if (IsSExt && isa<SExtInst>(Inst))
2326 // We can get through binary operator, if it is legal. In other words, the
2327 // binary operator must have a nuw or nsw flag.
2328 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2329 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2330 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2331 (IsSExt && BinOp->hasNoSignedWrap())))
2334 // Check if we can do the following simplification.
2335 // ext(trunc(opnd)) --> ext(opnd)
2336 if (!isa<TruncInst>(Inst))
2339 Value *OpndVal = Inst->getOperand(0);
2340 // Check if we can use this operand in the extension.
2341 // If the type is larger than the result type of the extension,
2343 if (!OpndVal->getType()->isIntegerTy() ||
2344 OpndVal->getType()->getIntegerBitWidth() >
2345 ConsideredExtType->getIntegerBitWidth())
2348 // If the operand of the truncate is not an instruction, we will not have
2349 // any information on the dropped bits.
2350 // (Actually we could for constant but it is not worth the extra logic).
2351 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2355 // Check if the source of the type is narrow enough.
2356 // I.e., check that trunc just drops extended bits of the same kind of
2358 // #1 get the type of the operand and check the kind of the extended bits.
2359 const Type *OpndType;
2360 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2361 if (It != PromotedInsts.end() && It->second.IsSExt == IsSExt)
2362 OpndType = It->second.Ty;
2363 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2364 OpndType = Opnd->getOperand(0)->getType();
2368 // #2 check that the truncate just drop extended bits.
2369 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
2375 TypePromotionHelper::Action TypePromotionHelper::getAction(
2376 Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2377 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2378 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2379 "Unexpected instruction type");
2380 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2381 Type *ExtTy = Ext->getType();
2382 bool IsSExt = isa<SExtInst>(Ext);
2383 // If the operand of the extension is not an instruction, we cannot
2385 // If it, check we can get through.
2386 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2389 // Do not promote if the operand has been added by codegenprepare.
2390 // Otherwise, it means we are undoing an optimization that is likely to be
2391 // redone, thus causing potential infinite loop.
2392 if (isa<TruncInst>(ExtOpnd) && InsertedTruncs.count(ExtOpnd))
2395 // SExt or Trunc instructions.
2396 // Return the related handler.
2397 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2398 isa<ZExtInst>(ExtOpnd))
2399 return promoteOperandForTruncAndAnyExt;
2401 // Regular instruction.
2402 // Abort early if we will have to insert non-free instructions.
2403 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2405 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2408 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2409 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2410 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2411 SmallVectorImpl<Instruction *> *Exts,
2412 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2413 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2414 // get through it and this method should not be called.
2415 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2416 Value *ExtVal = SExt;
2417 bool HasMergedNonFreeExt = false;
2418 if (isa<ZExtInst>(SExtOpnd)) {
2419 // Replace s|zext(zext(opnd))
2421 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
2423 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2424 TPT.replaceAllUsesWith(SExt, ZExt);
2425 TPT.eraseInstruction(SExt);
2428 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2430 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2432 CreatedInstsCost = 0;
2434 // Remove dead code.
2435 if (SExtOpnd->use_empty())
2436 TPT.eraseInstruction(SExtOpnd);
2438 // Check if the extension is still needed.
2439 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2440 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2443 Exts->push_back(ExtInst);
2444 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
2449 // At this point we have: ext ty opnd to ty.
2450 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2451 Value *NextVal = ExtInst->getOperand(0);
2452 TPT.eraseInstruction(ExtInst, NextVal);
2456 Value *TypePromotionHelper::promoteOperandForOther(
2457 Instruction *Ext, TypePromotionTransaction &TPT,
2458 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2459 SmallVectorImpl<Instruction *> *Exts,
2460 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
2462 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2463 // get through it and this method should not be called.
2464 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2465 CreatedInstsCost = 0;
2466 if (!ExtOpnd->hasOneUse()) {
2467 // ExtOpnd will be promoted.
2468 // All its uses, but Ext, will need to use a truncated value of the
2469 // promoted version.
2470 // Create the truncate now.
2471 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2472 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2473 ITrunc->removeFromParent();
2474 // Insert it just after the definition.
2475 ITrunc->insertAfter(ExtOpnd);
2477 Truncs->push_back(ITrunc);
2480 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2481 // Restore the operand of Ext (which has been replace by the previous call
2482 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2483 TPT.setOperand(Ext, 0, ExtOpnd);
2486 // Get through the Instruction:
2487 // 1. Update its type.
2488 // 2. Replace the uses of Ext by Inst.
2489 // 3. Extend each operand that needs to be extended.
2491 // Remember the original type of the instruction before promotion.
2492 // This is useful to know that the high bits are sign extended bits.
2493 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2494 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2496 TPT.mutateType(ExtOpnd, Ext->getType());
2498 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2500 Instruction *ExtForOpnd = Ext;
2502 DEBUG(dbgs() << "Propagate Ext to operands\n");
2503 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2505 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2506 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2507 !shouldExtOperand(ExtOpnd, OpIdx)) {
2508 DEBUG(dbgs() << "No need to propagate\n");
2511 // Check if we can statically extend the operand.
2512 Value *Opnd = ExtOpnd->getOperand(OpIdx);
2513 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2514 DEBUG(dbgs() << "Statically extend\n");
2515 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2516 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
2517 : Cst->getValue().zext(BitWidth);
2518 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
2521 // UndefValue are typed, so we have to statically sign extend them.
2522 if (isa<UndefValue>(Opnd)) {
2523 DEBUG(dbgs() << "Statically extend\n");
2524 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2528 // Otherwise we have to explicity sign extend the operand.
2529 // Check if Ext was reused to extend an operand.
2531 // If yes, create a new one.
2532 DEBUG(dbgs() << "More operands to ext\n");
2533 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2534 : TPT.createZExt(Ext, Opnd, Ext->getType());
2535 if (!isa<Instruction>(ValForExtOpnd)) {
2536 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
2539 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
2542 Exts->push_back(ExtForOpnd);
2543 TPT.setOperand(ExtForOpnd, 0, Opnd);
2545 // Move the sign extension before the insertion point.
2546 TPT.moveBefore(ExtForOpnd, ExtOpnd);
2547 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2548 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
2549 // If more sext are required, new instructions will have to be created.
2550 ExtForOpnd = nullptr;
2552 if (ExtForOpnd == Ext) {
2553 DEBUG(dbgs() << "Extension is useless now\n");
2554 TPT.eraseInstruction(Ext);
2559 /// IsPromotionProfitable - Check whether or not promoting an instruction
2560 /// to a wider type was profitable.
2561 /// \p NewCost gives the cost of extension instructions created by the
2563 /// \p OldCost gives the cost of extension instructions before the promotion
2564 /// plus the number of instructions that have been
2565 /// matched in the addressing mode the promotion.
2566 /// \p PromotedOperand is the value that has been promoted.
2567 /// \return True if the promotion is profitable, false otherwise.
2568 bool AddressingModeMatcher::IsPromotionProfitable(
2569 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
2570 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
2571 // The cost of the new extensions is greater than the cost of the
2572 // old extension plus what we folded.
2573 // This is not profitable.
2574 if (NewCost > OldCost)
2576 if (NewCost < OldCost)
2578 // The promotion is neutral but it may help folding the sign extension in
2579 // loads for instance.
2580 // Check that we did not create an illegal instruction.
2581 return isPromotedInstructionLegal(TLI, PromotedOperand);
2584 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
2585 /// fold the operation into the addressing mode. If so, update the addressing
2586 /// mode and return true, otherwise return false without modifying AddrMode.
2587 /// If \p MovedAway is not NULL, it contains the information of whether or
2588 /// not AddrInst has to be folded into the addressing mode on success.
2589 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2590 /// because it has been moved away.
2591 /// Thus AddrInst must not be added in the matched instructions.
2592 /// This state can happen when AddrInst is a sext, since it may be moved away.
2593 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2594 /// not be referenced anymore.
2595 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
2598 // Avoid exponential behavior on extremely deep expression trees.
2599 if (Depth >= 5) return false;
2601 // By default, all matched instructions stay in place.
2606 case Instruction::PtrToInt:
2607 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2608 return MatchAddr(AddrInst->getOperand(0), Depth);
2609 case Instruction::IntToPtr:
2610 // This inttoptr is a no-op if the integer type is pointer sized.
2611 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
2612 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
2613 return MatchAddr(AddrInst->getOperand(0), Depth);
2615 case Instruction::BitCast:
2616 case Instruction::AddrSpaceCast:
2617 // BitCast is always a noop, and we can handle it as long as it is
2618 // int->int or pointer->pointer (we don't want int<->fp or something).
2619 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2620 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2621 // Don't touch identity bitcasts. These were probably put here by LSR,
2622 // and we don't want to mess around with them. Assume it knows what it
2624 AddrInst->getOperand(0)->getType() != AddrInst->getType())
2625 return MatchAddr(AddrInst->getOperand(0), Depth);
2627 case Instruction::Add: {
2628 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2629 ExtAddrMode BackupAddrMode = AddrMode;
2630 unsigned OldSize = AddrModeInsts.size();
2631 // Start a transaction at this point.
2632 // The LHS may match but not the RHS.
2633 // Therefore, we need a higher level restoration point to undo partially
2634 // matched operation.
2635 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2636 TPT.getRestorationPoint();
2638 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2639 MatchAddr(AddrInst->getOperand(0), Depth+1))
2642 // Restore the old addr mode info.
2643 AddrMode = BackupAddrMode;
2644 AddrModeInsts.resize(OldSize);
2645 TPT.rollback(LastKnownGood);
2647 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2648 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2649 MatchAddr(AddrInst->getOperand(1), Depth+1))
2652 // Otherwise we definitely can't merge the ADD in.
2653 AddrMode = BackupAddrMode;
2654 AddrModeInsts.resize(OldSize);
2655 TPT.rollback(LastKnownGood);
2658 //case Instruction::Or:
2659 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2661 case Instruction::Mul:
2662 case Instruction::Shl: {
2663 // Can only handle X*C and X << C.
2664 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2667 int64_t Scale = RHS->getSExtValue();
2668 if (Opcode == Instruction::Shl)
2669 Scale = 1LL << Scale;
2671 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2673 case Instruction::GetElementPtr: {
2674 // Scan the GEP. We check it if it contains constant offsets and at most
2675 // one variable offset.
2676 int VariableOperand = -1;
2677 unsigned VariableScale = 0;
2679 int64_t ConstantOffset = 0;
2680 const DataLayout *TD = TLI.getDataLayout();
2681 gep_type_iterator GTI = gep_type_begin(AddrInst);
2682 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2683 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2684 const StructLayout *SL = TD->getStructLayout(STy);
2686 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2687 ConstantOffset += SL->getElementOffset(Idx);
2689 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
2690 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2691 ConstantOffset += CI->getSExtValue()*TypeSize;
2692 } else if (TypeSize) { // Scales of zero don't do anything.
2693 // We only allow one variable index at the moment.
2694 if (VariableOperand != -1)
2697 // Remember the variable index.
2698 VariableOperand = i;
2699 VariableScale = TypeSize;
2704 // A common case is for the GEP to only do a constant offset. In this case,
2705 // just add it to the disp field and check validity.
2706 if (VariableOperand == -1) {
2707 AddrMode.BaseOffs += ConstantOffset;
2708 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
2709 // Check to see if we can fold the base pointer in too.
2710 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2713 AddrMode.BaseOffs -= ConstantOffset;
2717 // Save the valid addressing mode in case we can't match.
2718 ExtAddrMode BackupAddrMode = AddrMode;
2719 unsigned OldSize = AddrModeInsts.size();
2721 // See if the scale and offset amount is valid for this target.
2722 AddrMode.BaseOffs += ConstantOffset;
2724 // Match the base operand of the GEP.
2725 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2726 // If it couldn't be matched, just stuff the value in a register.
2727 if (AddrMode.HasBaseReg) {
2728 AddrMode = BackupAddrMode;
2729 AddrModeInsts.resize(OldSize);
2732 AddrMode.HasBaseReg = true;
2733 AddrMode.BaseReg = AddrInst->getOperand(0);
2736 // Match the remaining variable portion of the GEP.
2737 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2739 // If it couldn't be matched, try stuffing the base into a register
2740 // instead of matching it, and retrying the match of the scale.
2741 AddrMode = BackupAddrMode;
2742 AddrModeInsts.resize(OldSize);
2743 if (AddrMode.HasBaseReg)
2745 AddrMode.HasBaseReg = true;
2746 AddrMode.BaseReg = AddrInst->getOperand(0);
2747 AddrMode.BaseOffs += ConstantOffset;
2748 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2749 VariableScale, Depth)) {
2750 // If even that didn't work, bail.
2751 AddrMode = BackupAddrMode;
2752 AddrModeInsts.resize(OldSize);
2759 case Instruction::SExt:
2760 case Instruction::ZExt: {
2761 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
2765 // Try to move this ext out of the way of the addressing mode.
2766 // Ask for a method for doing so.
2767 TypePromotionHelper::Action TPH =
2768 TypePromotionHelper::getAction(Ext, InsertedTruncs, TLI, PromotedInsts);
2772 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2773 TPT.getRestorationPoint();
2774 unsigned CreatedInstsCost = 0;
2775 unsigned ExtCost = !TLI.isExtFree(Ext);
2776 Value *PromotedOperand =
2777 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
2778 // SExt has been moved away.
2779 // Thus either it will be rematched later in the recursive calls or it is
2780 // gone. Anyway, we must not fold it into the addressing mode at this point.
2784 // addr = gep base, idx
2786 // promotedOpnd = ext opnd <- no match here
2787 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2788 // addr = gep base, op <- match
2792 assert(PromotedOperand &&
2793 "TypePromotionHelper should have filtered out those cases");
2795 ExtAddrMode BackupAddrMode = AddrMode;
2796 unsigned OldSize = AddrModeInsts.size();
2798 if (!MatchAddr(PromotedOperand, Depth) ||
2799 // The total of the new cost is equals to the cost of the created
2801 // The total of the old cost is equals to the cost of the extension plus
2802 // what we have saved in the addressing mode.
2803 !IsPromotionProfitable(CreatedInstsCost,
2804 ExtCost + (AddrModeInsts.size() - OldSize),
2806 AddrMode = BackupAddrMode;
2807 AddrModeInsts.resize(OldSize);
2808 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2809 TPT.rollback(LastKnownGood);
2818 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2819 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2820 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2821 /// or intptr_t for the target.
2823 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2824 // Start a transaction at this point that we will rollback if the matching
2826 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2827 TPT.getRestorationPoint();
2828 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2829 // Fold in immediates if legal for the target.
2830 AddrMode.BaseOffs += CI->getSExtValue();
2831 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2833 AddrMode.BaseOffs -= CI->getSExtValue();
2834 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2835 // If this is a global variable, try to fold it into the addressing mode.
2836 if (!AddrMode.BaseGV) {
2837 AddrMode.BaseGV = GV;
2838 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2840 AddrMode.BaseGV = nullptr;
2842 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2843 ExtAddrMode BackupAddrMode = AddrMode;
2844 unsigned OldSize = AddrModeInsts.size();
2846 // Check to see if it is possible to fold this operation.
2847 bool MovedAway = false;
2848 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2849 // This instruction may have been move away. If so, there is nothing
2853 // Okay, it's possible to fold this. Check to see if it is actually
2854 // *profitable* to do so. We use a simple cost model to avoid increasing
2855 // register pressure too much.
2856 if (I->hasOneUse() ||
2857 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2858 AddrModeInsts.push_back(I);
2862 // It isn't profitable to do this, roll back.
2863 //cerr << "NOT FOLDING: " << *I;
2864 AddrMode = BackupAddrMode;
2865 AddrModeInsts.resize(OldSize);
2866 TPT.rollback(LastKnownGood);
2868 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2869 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2871 TPT.rollback(LastKnownGood);
2872 } else if (isa<ConstantPointerNull>(Addr)) {
2873 // Null pointer gets folded without affecting the addressing mode.
2877 // Worse case, the target should support [reg] addressing modes. :)
2878 if (!AddrMode.HasBaseReg) {
2879 AddrMode.HasBaseReg = true;
2880 AddrMode.BaseReg = Addr;
2881 // Still check for legality in case the target supports [imm] but not [i+r].
2882 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2884 AddrMode.HasBaseReg = false;
2885 AddrMode.BaseReg = nullptr;
2888 // If the base register is already taken, see if we can do [r+r].
2889 if (AddrMode.Scale == 0) {
2891 AddrMode.ScaledReg = Addr;
2892 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2895 AddrMode.ScaledReg = nullptr;
2898 TPT.rollback(LastKnownGood);
2902 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2903 /// inline asm call are due to memory operands. If so, return true, otherwise
2905 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2906 const TargetMachine &TM) {
2907 const Function *F = CI->getParent()->getParent();
2908 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
2909 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
2910 TargetLowering::AsmOperandInfoVector TargetConstraints =
2911 TLI->ParseConstraints(TRI, ImmutableCallSite(CI));
2912 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2913 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2915 // Compute the constraint code and ConstraintType to use.
2916 TLI->ComputeConstraintToUse(OpInfo, SDValue());
2918 // If this asm operand is our Value*, and if it isn't an indirect memory
2919 // operand, we can't fold it!
2920 if (OpInfo.CallOperandVal == OpVal &&
2921 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2922 !OpInfo.isIndirect))
2929 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2930 /// memory use. If we find an obviously non-foldable instruction, return true.
2931 /// Add the ultimately found memory instructions to MemoryUses.
2932 static bool FindAllMemoryUses(
2934 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
2935 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
2936 // If we already considered this instruction, we're done.
2937 if (!ConsideredInsts.insert(I).second)
2940 // If this is an obviously unfoldable instruction, bail out.
2941 if (!MightBeFoldableInst(I))
2944 // Loop over all the uses, recursively processing them.
2945 for (Use &U : I->uses()) {
2946 Instruction *UserI = cast<Instruction>(U.getUser());
2948 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2949 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2953 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2954 unsigned opNo = U.getOperandNo();
2955 if (opNo == 0) return true; // Storing addr, not into addr.
2956 MemoryUses.push_back(std::make_pair(SI, opNo));
2960 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2961 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2962 if (!IA) return true;
2964 // If this is a memory operand, we're cool, otherwise bail out.
2965 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
2970 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
2977 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2978 /// the use site that we're folding it into. If so, there is no cost to
2979 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2980 /// that we know are live at the instruction already.
2981 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2982 Value *KnownLive2) {
2983 // If Val is either of the known-live values, we know it is live!
2984 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
2987 // All values other than instructions and arguments (e.g. constants) are live.
2988 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2990 // If Val is a constant sized alloca in the entry block, it is live, this is
2991 // true because it is just a reference to the stack/frame pointer, which is
2992 // live for the whole function.
2993 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2994 if (AI->isStaticAlloca())
2997 // Check to see if this value is already used in the memory instruction's
2998 // block. If so, it's already live into the block at the very least, so we
2999 // can reasonably fold it.
3000 return Val->isUsedInBasicBlock(MemoryInst->getParent());
3003 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
3004 /// mode of the machine to fold the specified instruction into a load or store
3005 /// that ultimately uses it. However, the specified instruction has multiple
3006 /// uses. Given this, it may actually increase register pressure to fold it
3007 /// into the load. For example, consider this code:
3011 /// use(Y) -> nonload/store
3015 /// In this case, Y has multiple uses, and can be folded into the load of Z
3016 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
3017 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
3018 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
3019 /// number of computations either.
3021 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
3022 /// X was live across 'load Z' for other reasons, we actually *would* want to
3023 /// fold the addressing mode in the Z case. This would make Y die earlier.
3024 bool AddressingModeMatcher::
3025 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
3026 ExtAddrMode &AMAfter) {
3027 if (IgnoreProfitability) return true;
3029 // AMBefore is the addressing mode before this instruction was folded into it,
3030 // and AMAfter is the addressing mode after the instruction was folded. Get
3031 // the set of registers referenced by AMAfter and subtract out those
3032 // referenced by AMBefore: this is the set of values which folding in this
3033 // address extends the lifetime of.
3035 // Note that there are only two potential values being referenced here,
3036 // BaseReg and ScaleReg (global addresses are always available, as are any
3037 // folded immediates).
3038 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
3040 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
3041 // lifetime wasn't extended by adding this instruction.
3042 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3044 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3045 ScaledReg = nullptr;
3047 // If folding this instruction (and it's subexprs) didn't extend any live
3048 // ranges, we're ok with it.
3049 if (!BaseReg && !ScaledReg)
3052 // If all uses of this instruction are ultimately load/store/inlineasm's,
3053 // check to see if their addressing modes will include this instruction. If
3054 // so, we can fold it into all uses, so it doesn't matter if it has multiple
3056 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
3057 SmallPtrSet<Instruction*, 16> ConsideredInsts;
3058 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
3059 return false; // Has a non-memory, non-foldable use!
3061 // Now that we know that all uses of this instruction are part of a chain of
3062 // computation involving only operations that could theoretically be folded
3063 // into a memory use, loop over each of these uses and see if they could
3064 // *actually* fold the instruction.
3065 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
3066 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
3067 Instruction *User = MemoryUses[i].first;
3068 unsigned OpNo = MemoryUses[i].second;
3070 // Get the access type of this use. If the use isn't a pointer, we don't
3071 // know what it accesses.
3072 Value *Address = User->getOperand(OpNo);
3073 if (!Address->getType()->isPointerTy())
3075 Type *AddressAccessTy = Address->getType()->getPointerElementType();
3077 // Do a match against the root of this address, ignoring profitability. This
3078 // will tell us if the addressing mode for the memory operation will
3079 // *actually* cover the shared instruction.
3081 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3082 TPT.getRestorationPoint();
3083 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy,
3084 MemoryInst, Result, InsertedTruncs,
3085 PromotedInsts, TPT);
3086 Matcher.IgnoreProfitability = true;
3087 bool Success = Matcher.MatchAddr(Address, 0);
3088 (void)Success; assert(Success && "Couldn't select *anything*?");
3090 // The match was to check the profitability, the changes made are not
3091 // part of the original matcher. Therefore, they should be dropped
3092 // otherwise the original matcher will not present the right state.
3093 TPT.rollback(LastKnownGood);
3095 // If the match didn't cover I, then it won't be shared by it.
3096 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3097 I) == MatchedAddrModeInsts.end())
3100 MatchedAddrModeInsts.clear();
3106 } // end anonymous namespace
3108 /// IsNonLocalValue - Return true if the specified values are defined in a
3109 /// different basic block than BB.
3110 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3111 if (Instruction *I = dyn_cast<Instruction>(V))
3112 return I->getParent() != BB;
3116 /// OptimizeMemoryInst - Load and Store Instructions often have
3117 /// addressing modes that can do significant amounts of computation. As such,
3118 /// instruction selection will try to get the load or store to do as much
3119 /// computation as possible for the program. The problem is that isel can only
3120 /// see within a single block. As such, we sink as much legal addressing mode
3121 /// stuff into the block as possible.
3123 /// This method is used to optimize both load/store and inline asms with memory
3125 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3129 // Try to collapse single-value PHI nodes. This is necessary to undo
3130 // unprofitable PRE transformations.
3131 SmallVector<Value*, 8> worklist;
3132 SmallPtrSet<Value*, 16> Visited;
3133 worklist.push_back(Addr);
3135 // Use a worklist to iteratively look through PHI nodes, and ensure that
3136 // the addressing mode obtained from the non-PHI roots of the graph
3138 Value *Consensus = nullptr;
3139 unsigned NumUsesConsensus = 0;
3140 bool IsNumUsesConsensusValid = false;
3141 SmallVector<Instruction*, 16> AddrModeInsts;
3142 ExtAddrMode AddrMode;
3143 TypePromotionTransaction TPT;
3144 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3145 TPT.getRestorationPoint();
3146 while (!worklist.empty()) {
3147 Value *V = worklist.back();
3148 worklist.pop_back();
3150 // Break use-def graph loops.
3151 if (!Visited.insert(V).second) {
3152 Consensus = nullptr;
3156 // For a PHI node, push all of its incoming values.
3157 if (PHINode *P = dyn_cast<PHINode>(V)) {
3158 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
3159 worklist.push_back(P->getIncomingValue(i));
3163 // For non-PHIs, determine the addressing mode being computed.
3164 SmallVector<Instruction*, 16> NewAddrModeInsts;
3165 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3166 V, AccessTy, MemoryInst, NewAddrModeInsts, *TM, InsertedTruncsSet,
3167 PromotedInsts, TPT);
3169 // This check is broken into two cases with very similar code to avoid using
3170 // getNumUses() as much as possible. Some values have a lot of uses, so
3171 // calling getNumUses() unconditionally caused a significant compile-time
3175 AddrMode = NewAddrMode;
3176 AddrModeInsts = NewAddrModeInsts;
3178 } else if (NewAddrMode == AddrMode) {
3179 if (!IsNumUsesConsensusValid) {
3180 NumUsesConsensus = Consensus->getNumUses();
3181 IsNumUsesConsensusValid = true;
3184 // Ensure that the obtained addressing mode is equivalent to that obtained
3185 // for all other roots of the PHI traversal. Also, when choosing one
3186 // such root as representative, select the one with the most uses in order
3187 // to keep the cost modeling heuristics in AddressingModeMatcher
3189 unsigned NumUses = V->getNumUses();
3190 if (NumUses > NumUsesConsensus) {
3192 NumUsesConsensus = NumUses;
3193 AddrModeInsts = NewAddrModeInsts;
3198 Consensus = nullptr;
3202 // If the addressing mode couldn't be determined, or if multiple different
3203 // ones were determined, bail out now.
3205 TPT.rollback(LastKnownGood);
3210 // Check to see if any of the instructions supersumed by this addr mode are
3211 // non-local to I's BB.
3212 bool AnyNonLocal = false;
3213 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3214 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3220 // If all the instructions matched are already in this BB, don't do anything.
3222 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3226 // Insert this computation right after this user. Since our caller is
3227 // scanning from the top of the BB to the bottom, reuse of the expr are
3228 // guaranteed to happen later.
3229 IRBuilder<> Builder(MemoryInst);
3231 // Now that we determined the addressing expression we want to use and know
3232 // that we have to sink it into this block. Check to see if we have already
3233 // done this for some other load/store instr in this block. If so, reuse the
3235 Value *&SunkAddr = SunkAddrs[Addr];
3237 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3238 << *MemoryInst << "\n");
3239 if (SunkAddr->getType() != Addr->getType())
3240 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3241 } else if (AddrSinkUsingGEPs ||
3242 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3243 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3245 // By default, we use the GEP-based method when AA is used later. This
3246 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3247 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3248 << *MemoryInst << "\n");
3249 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3250 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3252 // First, find the pointer.
3253 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3254 ResultPtr = AddrMode.BaseReg;
3255 AddrMode.BaseReg = nullptr;
3258 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3259 // We can't add more than one pointer together, nor can we scale a
3260 // pointer (both of which seem meaningless).
3261 if (ResultPtr || AddrMode.Scale != 1)
3264 ResultPtr = AddrMode.ScaledReg;
3268 if (AddrMode.BaseGV) {
3272 ResultPtr = AddrMode.BaseGV;
3275 // If the real base value actually came from an inttoptr, then the matcher
3276 // will look through it and provide only the integer value. In that case,
3278 if (!ResultPtr && AddrMode.BaseReg) {
3280 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3281 AddrMode.BaseReg = nullptr;
3282 } else if (!ResultPtr && AddrMode.Scale == 1) {
3284 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3289 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3290 SunkAddr = Constant::getNullValue(Addr->getType());
3291 } else if (!ResultPtr) {
3295 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3296 Type *I8Ty = Builder.getInt8Ty();
3298 // Start with the base register. Do this first so that subsequent address
3299 // matching finds it last, which will prevent it from trying to match it
3300 // as the scaled value in case it happens to be a mul. That would be
3301 // problematic if we've sunk a different mul for the scale, because then
3302 // we'd end up sinking both muls.
3303 if (AddrMode.BaseReg) {
3304 Value *V = AddrMode.BaseReg;
3305 if (V->getType() != IntPtrTy)
3306 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3311 // Add the scale value.
3312 if (AddrMode.Scale) {
3313 Value *V = AddrMode.ScaledReg;
3314 if (V->getType() == IntPtrTy) {
3316 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3317 cast<IntegerType>(V->getType())->getBitWidth()) {
3318 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3320 // It is only safe to sign extend the BaseReg if we know that the math
3321 // required to create it did not overflow before we extend it. Since
3322 // the original IR value was tossed in favor of a constant back when
3323 // the AddrMode was created we need to bail out gracefully if widths
3324 // do not match instead of extending it.
3325 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3326 if (I && (ResultIndex != AddrMode.BaseReg))
3327 I->eraseFromParent();
3331 if (AddrMode.Scale != 1)
3332 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3335 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3340 // Add in the Base Offset if present.
3341 if (AddrMode.BaseOffs) {
3342 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3344 // We need to add this separately from the scale above to help with
3345 // SDAG consecutive load/store merging.
3346 if (ResultPtr->getType() != I8PtrTy)
3347 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3348 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3355 SunkAddr = ResultPtr;
3357 if (ResultPtr->getType() != I8PtrTy)
3358 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3359 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3362 if (SunkAddr->getType() != Addr->getType())
3363 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3366 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3367 << *MemoryInst << "\n");
3368 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3369 Value *Result = nullptr;
3371 // Start with the base register. Do this first so that subsequent address
3372 // matching finds it last, which will prevent it from trying to match it
3373 // as the scaled value in case it happens to be a mul. That would be
3374 // problematic if we've sunk a different mul for the scale, because then
3375 // we'd end up sinking both muls.
3376 if (AddrMode.BaseReg) {
3377 Value *V = AddrMode.BaseReg;
3378 if (V->getType()->isPointerTy())
3379 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3380 if (V->getType() != IntPtrTy)
3381 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3385 // Add the scale value.
3386 if (AddrMode.Scale) {
3387 Value *V = AddrMode.ScaledReg;
3388 if (V->getType() == IntPtrTy) {
3390 } else if (V->getType()->isPointerTy()) {
3391 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3392 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3393 cast<IntegerType>(V->getType())->getBitWidth()) {
3394 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3396 // It is only safe to sign extend the BaseReg if we know that the math
3397 // required to create it did not overflow before we extend it. Since
3398 // the original IR value was tossed in favor of a constant back when
3399 // the AddrMode was created we need to bail out gracefully if widths
3400 // do not match instead of extending it.
3401 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3402 if (I && (Result != AddrMode.BaseReg))
3403 I->eraseFromParent();
3406 if (AddrMode.Scale != 1)
3407 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3410 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3415 // Add in the BaseGV if present.
3416 if (AddrMode.BaseGV) {
3417 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3419 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3424 // Add in the Base Offset if present.
3425 if (AddrMode.BaseOffs) {
3426 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3428 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3434 SunkAddr = Constant::getNullValue(Addr->getType());
3436 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3439 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3441 // If we have no uses, recursively delete the value and all dead instructions
3443 if (Repl->use_empty()) {
3444 // This can cause recursive deletion, which can invalidate our iterator.
3445 // Use a WeakVH to hold onto it in case this happens.
3446 WeakVH IterHandle(CurInstIterator);
3447 BasicBlock *BB = CurInstIterator->getParent();
3449 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3451 if (IterHandle != CurInstIterator) {
3452 // If the iterator instruction was recursively deleted, start over at the
3453 // start of the block.
3454 CurInstIterator = BB->begin();
3462 /// OptimizeInlineAsmInst - If there are any memory operands, use
3463 /// OptimizeMemoryInst to sink their address computing into the block when
3464 /// possible / profitable.
3465 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
3466 bool MadeChange = false;
3468 const TargetRegisterInfo *TRI =
3469 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3470 TargetLowering::AsmOperandInfoVector
3471 TargetConstraints = TLI->ParseConstraints(TRI, CS);
3473 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3474 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3476 // Compute the constraint code and ConstraintType to use.
3477 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3479 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3480 OpInfo.isIndirect) {
3481 Value *OpVal = CS->getArgOperand(ArgNo++);
3482 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
3483 } else if (OpInfo.Type == InlineAsm::isInput)
3490 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3491 /// sign extensions.
3492 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3493 assert(!Inst->use_empty() && "Input must have at least one use");
3494 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3495 bool IsSExt = isa<SExtInst>(FirstUser);
3496 Type *ExtTy = FirstUser->getType();
3497 for (const User *U : Inst->users()) {
3498 const Instruction *UI = cast<Instruction>(U);
3499 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3501 Type *CurTy = UI->getType();
3502 // Same input and output types: Same instruction after CSE.
3506 // If IsSExt is true, we are in this situation:
3508 // b = sext ty1 a to ty2
3509 // c = sext ty1 a to ty3
3510 // Assuming ty2 is shorter than ty3, this could be turned into:
3512 // b = sext ty1 a to ty2
3513 // c = sext ty2 b to ty3
3514 // However, the last sext is not free.
3518 // This is a ZExt, maybe this is free to extend from one type to another.
3519 // In that case, we would not account for a different use.
3522 if (ExtTy->getScalarType()->getIntegerBitWidth() >
3523 CurTy->getScalarType()->getIntegerBitWidth()) {
3531 if (!TLI.isZExtFree(NarrowTy, LargeTy))
3534 // All uses are the same or can be derived from one another for free.
3538 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
3539 /// load instruction.
3540 /// If an ext(load) can be formed, it is returned via \p LI for the load
3541 /// and \p Inst for the extension.
3542 /// Otherwise LI == nullptr and Inst == nullptr.
3543 /// When some promotion happened, \p TPT contains the proper state to
3546 /// \return true when promoting was necessary to expose the ext(load)
3547 /// opportunity, false otherwise.
3551 /// %ld = load i32* %addr
3552 /// %add = add nuw i32 %ld, 4
3553 /// %zext = zext i32 %add to i64
3557 /// %ld = load i32* %addr
3558 /// %zext = zext i32 %ld to i64
3559 /// %add = add nuw i64 %zext, 4
3561 /// Thanks to the promotion, we can match zext(load i32*) to i64.
3562 bool CodeGenPrepare::ExtLdPromotion(TypePromotionTransaction &TPT,
3563 LoadInst *&LI, Instruction *&Inst,
3564 const SmallVectorImpl<Instruction *> &Exts,
3565 unsigned CreatedInstsCost = 0) {
3566 // Iterate over all the extensions to see if one form an ext(load).
3567 for (auto I : Exts) {
3568 // Check if we directly have ext(load).
3569 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
3571 // No promotion happened here.
3574 // Check whether or not we want to do any promotion.
3575 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
3577 // Get the action to perform the promotion.
3578 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
3579 I, InsertedTruncsSet, *TLI, PromotedInsts);
3580 // Check if we can promote.
3583 // Save the current state.
3584 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3585 TPT.getRestorationPoint();
3586 SmallVector<Instruction *, 4> NewExts;
3587 unsigned NewCreatedInstsCost = 0;
3588 unsigned ExtCost = !TLI->isExtFree(I);
3590 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
3591 &NewExts, nullptr, *TLI);
3592 assert(PromotedVal &&
3593 "TypePromotionHelper should have filtered out those cases");
3595 // We would be able to merge only one extension in a load.
3596 // Therefore, if we have more than 1 new extension we heuristically
3597 // cut this search path, because it means we degrade the code quality.
3598 // With exactly 2, the transformation is neutral, because we will merge
3599 // one extension but leave one. However, we optimistically keep going,
3600 // because the new extension may be removed too.
3601 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
3602 TotalCreatedInstsCost -= ExtCost;
3603 if (!StressExtLdPromotion &&
3604 (TotalCreatedInstsCost > 1 ||
3605 !isPromotedInstructionLegal(*TLI, PromotedVal))) {
3606 // The promotion is not profitable, rollback to the previous state.
3607 TPT.rollback(LastKnownGood);
3610 // The promotion is profitable.
3611 // Check if it exposes an ext(load).
3612 (void)ExtLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
3613 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
3614 // If we have created a new extension, i.e., now we have two
3615 // extensions. We must make sure one of them is merged with
3616 // the load, otherwise we may degrade the code quality.
3617 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
3618 // Promotion happened.
3620 // If this does not help to expose an ext(load) then, rollback.
3621 TPT.rollback(LastKnownGood);
3623 // None of the extension can form an ext(load).
3629 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
3630 /// basic block as the load, unless conditions are unfavorable. This allows
3631 /// SelectionDAG to fold the extend into the load.
3632 /// \p I[in/out] the extension may be modified during the process if some
3633 /// promotions apply.
3635 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *&I) {
3636 // Try to promote a chain of computation if it allows to form
3637 // an extended load.
3638 TypePromotionTransaction TPT;
3639 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3640 TPT.getRestorationPoint();
3641 SmallVector<Instruction *, 1> Exts;
3643 // Look for a load being extended.
3644 LoadInst *LI = nullptr;
3645 Instruction *OldExt = I;
3646 bool HasPromoted = ExtLdPromotion(TPT, LI, I, Exts);
3648 assert(!HasPromoted && !LI && "If we did not match any load instruction "
3649 "the code must remain the same");
3654 // If they're already in the same block, there's nothing to do.
3655 // Make the cheap checks first if we did not promote.
3656 // If we promoted, we need to check if it is indeed profitable.
3657 if (!HasPromoted && LI->getParent() == I->getParent())
3660 EVT VT = TLI->getValueType(I->getType());
3661 EVT LoadVT = TLI->getValueType(LI->getType());
3663 // If the load has other users and the truncate is not free, this probably
3664 // isn't worthwhile.
3665 if (!LI->hasOneUse() && TLI &&
3666 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
3667 !TLI->isTruncateFree(I->getType(), LI->getType())) {
3669 TPT.rollback(LastKnownGood);
3673 // Check whether the target supports casts folded into loads.
3675 if (isa<ZExtInst>(I))
3676 LType = ISD::ZEXTLOAD;
3678 assert(isa<SExtInst>(I) && "Unexpected ext type!");
3679 LType = ISD::SEXTLOAD;
3681 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
3683 TPT.rollback(LastKnownGood);
3687 // Move the extend into the same block as the load, so that SelectionDAG
3690 I->removeFromParent();
3696 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
3697 BasicBlock *DefBB = I->getParent();
3699 // If the result of a {s|z}ext and its source are both live out, rewrite all
3700 // other uses of the source with result of extension.
3701 Value *Src = I->getOperand(0);
3702 if (Src->hasOneUse())
3705 // Only do this xform if truncating is free.
3706 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
3709 // Only safe to perform the optimization if the source is also defined in
3711 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
3714 bool DefIsLiveOut = false;
3715 for (User *U : I->users()) {
3716 Instruction *UI = cast<Instruction>(U);
3718 // Figure out which BB this ext is used in.
3719 BasicBlock *UserBB = UI->getParent();
3720 if (UserBB == DefBB) continue;
3721 DefIsLiveOut = true;
3727 // Make sure none of the uses are PHI nodes.
3728 for (User *U : Src->users()) {
3729 Instruction *UI = cast<Instruction>(U);
3730 BasicBlock *UserBB = UI->getParent();
3731 if (UserBB == DefBB) continue;
3732 // Be conservative. We don't want this xform to end up introducing
3733 // reloads just before load / store instructions.
3734 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
3738 // InsertedTruncs - Only insert one trunc in each block once.
3739 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
3741 bool MadeChange = false;
3742 for (Use &U : Src->uses()) {
3743 Instruction *User = cast<Instruction>(U.getUser());
3745 // Figure out which BB this ext is used in.
3746 BasicBlock *UserBB = User->getParent();
3747 if (UserBB == DefBB) continue;
3749 // Both src and def are live in this block. Rewrite the use.
3750 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
3752 if (!InsertedTrunc) {
3753 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3754 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
3755 InsertedTruncsSet.insert(InsertedTrunc);
3758 // Replace a use of the {s|z}ext source with a use of the result.
3767 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
3768 /// turned into an explicit branch.
3769 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
3770 // FIXME: This should use the same heuristics as IfConversion to determine
3771 // whether a select is better represented as a branch. This requires that
3772 // branch probability metadata is preserved for the select, which is not the
3775 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3777 // If the branch is predicted right, an out of order CPU can avoid blocking on
3778 // the compare. Emit cmovs on compares with a memory operand as branches to
3779 // avoid stalls on the load from memory. If the compare has more than one use
3780 // there's probably another cmov or setcc around so it's not worth emitting a
3785 Value *CmpOp0 = Cmp->getOperand(0);
3786 Value *CmpOp1 = Cmp->getOperand(1);
3788 // We check that the memory operand has one use to avoid uses of the loaded
3789 // value directly after the compare, making branches unprofitable.
3790 return Cmp->hasOneUse() &&
3791 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3792 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
3796 /// If we have a SelectInst that will likely profit from branch prediction,
3797 /// turn it into a branch.
3798 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
3799 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3801 // Can we convert the 'select' to CF ?
3802 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3805 TargetLowering::SelectSupportKind SelectKind;
3807 SelectKind = TargetLowering::VectorMaskSelect;
3808 else if (SI->getType()->isVectorTy())
3809 SelectKind = TargetLowering::ScalarCondVectorVal;
3811 SelectKind = TargetLowering::ScalarValSelect;
3813 // Do we have efficient codegen support for this kind of 'selects' ?
3814 if (TLI->isSelectSupported(SelectKind)) {
3815 // We have efficient codegen support for the select instruction.
3816 // Check if it is profitable to keep this 'select'.
3817 if (!TLI->isPredictableSelectExpensive() ||
3818 !isFormingBranchFromSelectProfitable(SI))
3824 // First, we split the block containing the select into 2 blocks.
3825 BasicBlock *StartBlock = SI->getParent();
3826 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3827 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3829 // Create a new block serving as the landing pad for the branch.
3830 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3831 NextBlock->getParent(), NextBlock);
3833 // Move the unconditional branch from the block with the select in it into our
3834 // landing pad block.
3835 StartBlock->getTerminator()->eraseFromParent();
3836 BranchInst::Create(NextBlock, SmallBlock);
3838 // Insert the real conditional branch based on the original condition.
3839 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3841 // The select itself is replaced with a PHI Node.
3842 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3844 PN->addIncoming(SI->getTrueValue(), StartBlock);
3845 PN->addIncoming(SI->getFalseValue(), SmallBlock);
3846 SI->replaceAllUsesWith(PN);
3847 SI->eraseFromParent();
3849 // Instruct OptimizeBlock to skip to the next block.
3850 CurInstIterator = StartBlock->end();
3851 ++NumSelectsExpanded;
3855 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3856 SmallVector<int, 16> Mask(SVI->getShuffleMask());
3858 for (unsigned i = 0; i < Mask.size(); ++i) {
3859 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3861 SplatElem = Mask[i];
3867 /// Some targets have expensive vector shifts if the lanes aren't all the same
3868 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3869 /// it's often worth sinking a shufflevector splat down to its use so that
3870 /// codegen can spot all lanes are identical.
3871 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3872 BasicBlock *DefBB = SVI->getParent();
3874 // Only do this xform if variable vector shifts are particularly expensive.
3875 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3878 // We only expect better codegen by sinking a shuffle if we can recognise a
3880 if (!isBroadcastShuffle(SVI))
3883 // InsertedShuffles - Only insert a shuffle in each block once.
3884 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3886 bool MadeChange = false;
3887 for (User *U : SVI->users()) {
3888 Instruction *UI = cast<Instruction>(U);
3890 // Figure out which BB this ext is used in.
3891 BasicBlock *UserBB = UI->getParent();
3892 if (UserBB == DefBB) continue;
3894 // For now only apply this when the splat is used by a shift instruction.
3895 if (!UI->isShift()) continue;
3897 // Everything checks out, sink the shuffle if the user's block doesn't
3898 // already have a copy.
3899 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3901 if (!InsertedShuffle) {
3902 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3903 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3905 SVI->getOperand(2), "", InsertPt);
3908 UI->replaceUsesOfWith(SVI, InsertedShuffle);
3912 // If we removed all uses, nuke the shuffle.
3913 if (SVI->use_empty()) {
3914 SVI->eraseFromParent();
3922 /// \brief Helper class to promote a scalar operation to a vector one.
3923 /// This class is used to move downward extractelement transition.
3925 /// a = vector_op <2 x i32>
3926 /// b = extractelement <2 x i32> a, i32 0
3931 /// a = vector_op <2 x i32>
3932 /// c = vector_op a (equivalent to scalar_op on the related lane)
3933 /// * d = extractelement <2 x i32> c, i32 0
3935 /// Assuming both extractelement and store can be combine, we get rid of the
3937 class VectorPromoteHelper {
3938 /// Used to perform some checks on the legality of vector operations.
3939 const TargetLowering &TLI;
3941 /// Used to estimated the cost of the promoted chain.
3942 const TargetTransformInfo &TTI;
3944 /// The transition being moved downwards.
3945 Instruction *Transition;
3946 /// The sequence of instructions to be promoted.
3947 SmallVector<Instruction *, 4> InstsToBePromoted;
3948 /// Cost of combining a store and an extract.
3949 unsigned StoreExtractCombineCost;
3950 /// Instruction that will be combined with the transition.
3951 Instruction *CombineInst;
3953 /// \brief The instruction that represents the current end of the transition.
3954 /// Since we are faking the promotion until we reach the end of the chain
3955 /// of computation, we need a way to get the current end of the transition.
3956 Instruction *getEndOfTransition() const {
3957 if (InstsToBePromoted.empty())
3959 return InstsToBePromoted.back();
3962 /// \brief Return the index of the original value in the transition.
3963 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
3964 /// c, is at index 0.
3965 unsigned getTransitionOriginalValueIdx() const {
3966 assert(isa<ExtractElementInst>(Transition) &&
3967 "Other kind of transitions are not supported yet");
3971 /// \brief Return the index of the index in the transition.
3972 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
3974 unsigned getTransitionIdx() const {
3975 assert(isa<ExtractElementInst>(Transition) &&
3976 "Other kind of transitions are not supported yet");
3980 /// \brief Get the type of the transition.
3981 /// This is the type of the original value.
3982 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
3983 /// transition is <2 x i32>.
3984 Type *getTransitionType() const {
3985 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
3988 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
3989 /// I.e., we have the following sequence:
3990 /// Def = Transition <ty1> a to <ty2>
3991 /// b = ToBePromoted <ty2> Def, ...
3993 /// b = ToBePromoted <ty1> a, ...
3994 /// Def = Transition <ty1> ToBePromoted to <ty2>
3995 void promoteImpl(Instruction *ToBePromoted);
3997 /// \brief Check whether or not it is profitable to promote all the
3998 /// instructions enqueued to be promoted.
3999 bool isProfitableToPromote() {
4000 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
4001 unsigned Index = isa<ConstantInt>(ValIdx)
4002 ? cast<ConstantInt>(ValIdx)->getZExtValue()
4004 Type *PromotedType = getTransitionType();
4006 StoreInst *ST = cast<StoreInst>(CombineInst);
4007 unsigned AS = ST->getPointerAddressSpace();
4008 unsigned Align = ST->getAlignment();
4009 // Check if this store is supported.
4010 if (!TLI.allowsMisalignedMemoryAccesses(
4011 TLI.getValueType(ST->getValueOperand()->getType()), AS, Align)) {
4012 // If this is not supported, there is no way we can combine
4013 // the extract with the store.
4017 // The scalar chain of computation has to pay for the transition
4018 // scalar to vector.
4019 // The vector chain has to account for the combining cost.
4020 uint64_t ScalarCost =
4021 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
4022 uint64_t VectorCost = StoreExtractCombineCost;
4023 for (const auto &Inst : InstsToBePromoted) {
4024 // Compute the cost.
4025 // By construction, all instructions being promoted are arithmetic ones.
4026 // Moreover, one argument is a constant that can be viewed as a splat
4028 Value *Arg0 = Inst->getOperand(0);
4029 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
4030 isa<ConstantFP>(Arg0);
4031 TargetTransformInfo::OperandValueKind Arg0OVK =
4032 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4033 : TargetTransformInfo::OK_AnyValue;
4034 TargetTransformInfo::OperandValueKind Arg1OVK =
4035 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4036 : TargetTransformInfo::OK_AnyValue;
4037 ScalarCost += TTI.getArithmeticInstrCost(
4038 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
4039 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
4042 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
4043 << ScalarCost << "\nVector: " << VectorCost << '\n');
4044 return ScalarCost > VectorCost;
4047 /// \brief Generate a constant vector with \p Val with the same
4048 /// number of elements as the transition.
4049 /// \p UseSplat defines whether or not \p Val should be replicated
4050 /// accross the whole vector.
4051 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
4052 /// otherwise we generate a vector with as many undef as possible:
4053 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
4054 /// used at the index of the extract.
4055 Value *getConstantVector(Constant *Val, bool UseSplat) const {
4056 unsigned ExtractIdx = UINT_MAX;
4058 // If we cannot determine where the constant must be, we have to
4059 // use a splat constant.
4060 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
4061 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
4062 ExtractIdx = CstVal->getSExtValue();
4067 unsigned End = getTransitionType()->getVectorNumElements();
4069 return ConstantVector::getSplat(End, Val);
4071 SmallVector<Constant *, 4> ConstVec;
4072 UndefValue *UndefVal = UndefValue::get(Val->getType());
4073 for (unsigned Idx = 0; Idx != End; ++Idx) {
4074 if (Idx == ExtractIdx)
4075 ConstVec.push_back(Val);
4077 ConstVec.push_back(UndefVal);
4079 return ConstantVector::get(ConstVec);
4082 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
4083 /// in \p Use can trigger undefined behavior.
4084 static bool canCauseUndefinedBehavior(const Instruction *Use,
4085 unsigned OperandIdx) {
4086 // This is not safe to introduce undef when the operand is on
4087 // the right hand side of a division-like instruction.
4088 if (OperandIdx != 1)
4090 switch (Use->getOpcode()) {
4093 case Instruction::SDiv:
4094 case Instruction::UDiv:
4095 case Instruction::SRem:
4096 case Instruction::URem:
4098 case Instruction::FDiv:
4099 case Instruction::FRem:
4100 return !Use->hasNoNaNs();
4102 llvm_unreachable(nullptr);
4106 VectorPromoteHelper(const TargetLowering &TLI, const TargetTransformInfo &TTI,
4107 Instruction *Transition, unsigned CombineCost)
4108 : TLI(TLI), TTI(TTI), Transition(Transition),
4109 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4110 assert(Transition && "Do not know how to promote null");
4113 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4114 bool canPromote(const Instruction *ToBePromoted) const {
4115 // We could support CastInst too.
4116 return isa<BinaryOperator>(ToBePromoted);
4119 /// \brief Check if it is profitable to promote \p ToBePromoted
4120 /// by moving downward the transition through.
4121 bool shouldPromote(const Instruction *ToBePromoted) const {
4122 // Promote only if all the operands can be statically expanded.
4123 // Indeed, we do not want to introduce any new kind of transitions.
4124 for (const Use &U : ToBePromoted->operands()) {
4125 const Value *Val = U.get();
4126 if (Val == getEndOfTransition()) {
4127 // If the use is a division and the transition is on the rhs,
4128 // we cannot promote the operation, otherwise we may create a
4129 // division by zero.
4130 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4134 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4135 !isa<ConstantFP>(Val))
4138 // Check that the resulting operation is legal.
4139 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4142 return StressStoreExtract ||
4143 TLI.isOperationLegalOrCustom(
4144 ISDOpcode, TLI.getValueType(getTransitionType(), true));
4147 /// \brief Check whether or not \p Use can be combined
4148 /// with the transition.
4149 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4150 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4152 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4153 void enqueueForPromotion(Instruction *ToBePromoted) {
4154 InstsToBePromoted.push_back(ToBePromoted);
4157 /// \brief Set the instruction that will be combined with the transition.
4158 void recordCombineInstruction(Instruction *ToBeCombined) {
4159 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4160 CombineInst = ToBeCombined;
4163 /// \brief Promote all the instructions enqueued for promotion if it is
4165 /// \return True if the promotion happened, false otherwise.
4167 // Check if there is something to promote.
4168 // Right now, if we do not have anything to combine with,
4169 // we assume the promotion is not profitable.
4170 if (InstsToBePromoted.empty() || !CombineInst)
4174 if (!StressStoreExtract && !isProfitableToPromote())
4178 for (auto &ToBePromoted : InstsToBePromoted)
4179 promoteImpl(ToBePromoted);
4180 InstsToBePromoted.clear();
4184 } // End of anonymous namespace.
4186 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4187 // At this point, we know that all the operands of ToBePromoted but Def
4188 // can be statically promoted.
4189 // For Def, we need to use its parameter in ToBePromoted:
4190 // b = ToBePromoted ty1 a
4191 // Def = Transition ty1 b to ty2
4192 // Move the transition down.
4193 // 1. Replace all uses of the promoted operation by the transition.
4194 // = ... b => = ... Def.
4195 assert(ToBePromoted->getType() == Transition->getType() &&
4196 "The type of the result of the transition does not match "
4198 ToBePromoted->replaceAllUsesWith(Transition);
4199 // 2. Update the type of the uses.
4200 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
4201 Type *TransitionTy = getTransitionType();
4202 ToBePromoted->mutateType(TransitionTy);
4203 // 3. Update all the operands of the promoted operation with promoted
4205 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
4206 for (Use &U : ToBePromoted->operands()) {
4207 Value *Val = U.get();
4208 Value *NewVal = nullptr;
4209 if (Val == Transition)
4210 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
4211 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
4212 isa<ConstantFP>(Val)) {
4213 // Use a splat constant if it is not safe to use undef.
4214 NewVal = getConstantVector(
4215 cast<Constant>(Val),
4216 isa<UndefValue>(Val) ||
4217 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
4219 llvm_unreachable("Did you modified shouldPromote and forgot to update "
4221 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
4223 Transition->removeFromParent();
4224 Transition->insertAfter(ToBePromoted);
4225 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
4228 /// Some targets can do store(extractelement) with one instruction.
4229 /// Try to push the extractelement towards the stores when the target
4230 /// has this feature and this is profitable.
4231 bool CodeGenPrepare::OptimizeExtractElementInst(Instruction *Inst) {
4232 unsigned CombineCost = UINT_MAX;
4233 if (DisableStoreExtract || !TLI ||
4234 (!StressStoreExtract &&
4235 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4236 Inst->getOperand(1), CombineCost)))
4239 // At this point we know that Inst is a vector to scalar transition.
4240 // Try to move it down the def-use chain, until:
4241 // - We can combine the transition with its single use
4242 // => we got rid of the transition.
4243 // - We escape the current basic block
4244 // => we would need to check that we are moving it at a cheaper place and
4245 // we do not do that for now.
4246 BasicBlock *Parent = Inst->getParent();
4247 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4248 VectorPromoteHelper VPH(*TLI, *TTI, Inst, CombineCost);
4249 // If the transition has more than one use, assume this is not going to be
4251 while (Inst->hasOneUse()) {
4252 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4253 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4255 if (ToBePromoted->getParent() != Parent) {
4256 DEBUG(dbgs() << "Instruction to promote is in a different block ("
4257 << ToBePromoted->getParent()->getName()
4258 << ") than the transition (" << Parent->getName() << ").\n");
4262 if (VPH.canCombine(ToBePromoted)) {
4263 DEBUG(dbgs() << "Assume " << *Inst << '\n'
4264 << "will be combined with: " << *ToBePromoted << '\n');
4265 VPH.recordCombineInstruction(ToBePromoted);
4266 bool Changed = VPH.promote();
4267 NumStoreExtractExposed += Changed;
4271 DEBUG(dbgs() << "Try promoting.\n");
4272 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4275 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4277 VPH.enqueueForPromotion(ToBePromoted);
4278 Inst = ToBePromoted;
4283 bool CodeGenPrepare::OptimizeInst(Instruction *I, bool& ModifiedDT) {
4284 if (PHINode *P = dyn_cast<PHINode>(I)) {
4285 // It is possible for very late stage optimizations (such as SimplifyCFG)
4286 // to introduce PHI nodes too late to be cleaned up. If we detect such a
4287 // trivial PHI, go ahead and zap it here.
4288 const DataLayout &DL = I->getModule()->getDataLayout();
4289 if (Value *V = SimplifyInstruction(P, DL, TLInfo, nullptr)) {
4290 P->replaceAllUsesWith(V);
4291 P->eraseFromParent();
4298 if (CastInst *CI = dyn_cast<CastInst>(I)) {
4299 // If the source of the cast is a constant, then this should have
4300 // already been constant folded. The only reason NOT to constant fold
4301 // it is if something (e.g. LSR) was careful to place the constant
4302 // evaluation in a block other than then one that uses it (e.g. to hoist
4303 // the address of globals out of a loop). If this is the case, we don't
4304 // want to forward-subst the cast.
4305 if (isa<Constant>(CI->getOperand(0)))
4308 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
4311 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4312 /// Sink a zext or sext into its user blocks if the target type doesn't
4313 /// fit in one register
4314 if (TLI && TLI->getTypeAction(CI->getContext(),
4315 TLI->getValueType(CI->getType())) ==
4316 TargetLowering::TypeExpandInteger) {
4317 return SinkCast(CI);
4319 bool MadeChange = MoveExtToFormExtLoad(I);
4320 return MadeChange | OptimizeExtUses(I);
4326 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4327 if (!TLI || !TLI->hasMultipleConditionRegisters())
4328 return OptimizeCmpExpression(CI);
4330 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4332 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
4336 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4338 return OptimizeMemoryInst(I, SI->getOperand(1),
4339 SI->getOperand(0)->getType());
4343 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4345 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4346 BinOp->getOpcode() == Instruction::LShr)) {
4347 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4348 if (TLI && CI && TLI->hasExtractBitsInsn())
4349 return OptimizeExtractBits(BinOp, CI, *TLI);
4354 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4355 if (GEPI->hasAllZeroIndices()) {
4356 /// The GEP operand must be a pointer, so must its result -> BitCast
4357 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4358 GEPI->getName(), GEPI);
4359 GEPI->replaceAllUsesWith(NC);
4360 GEPI->eraseFromParent();
4362 OptimizeInst(NC, ModifiedDT);
4368 if (CallInst *CI = dyn_cast<CallInst>(I))
4369 return OptimizeCallInst(CI, ModifiedDT);
4371 if (SelectInst *SI = dyn_cast<SelectInst>(I))
4372 return OptimizeSelectInst(SI);
4374 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
4375 return OptimizeShuffleVectorInst(SVI);
4377 if (isa<ExtractElementInst>(I))
4378 return OptimizeExtractElementInst(I);
4383 // In this pass we look for GEP and cast instructions that are used
4384 // across basic blocks and rewrite them to improve basic-block-at-a-time
4386 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
4388 bool MadeChange = false;
4390 CurInstIterator = BB.begin();
4391 while (CurInstIterator != BB.end()) {
4392 MadeChange |= OptimizeInst(CurInstIterator++, ModifiedDT);
4396 MadeChange |= DupRetToEnableTailCallOpts(&BB);
4401 // llvm.dbg.value is far away from the value then iSel may not be able
4402 // handle it properly. iSel will drop llvm.dbg.value if it can not
4403 // find a node corresponding to the value.
4404 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
4405 bool MadeChange = false;
4406 for (BasicBlock &BB : F) {
4407 Instruction *PrevNonDbgInst = nullptr;
4408 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
4409 Instruction *Insn = BI++;
4410 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
4411 // Leave dbg.values that refer to an alloca alone. These
4412 // instrinsics describe the address of a variable (= the alloca)
4413 // being taken. They should not be moved next to the alloca
4414 // (and to the beginning of the scope), but rather stay close to
4415 // where said address is used.
4416 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
4417 PrevNonDbgInst = Insn;
4421 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
4422 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
4423 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
4424 DVI->removeFromParent();
4425 if (isa<PHINode>(VI))
4426 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
4428 DVI->insertAfter(VI);
4437 // If there is a sequence that branches based on comparing a single bit
4438 // against zero that can be combined into a single instruction, and the
4439 // target supports folding these into a single instruction, sink the
4440 // mask and compare into the branch uses. Do this before OptimizeBlock ->
4441 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
4443 bool CodeGenPrepare::sinkAndCmp(Function &F) {
4444 if (!EnableAndCmpSinking)
4446 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
4448 bool MadeChange = false;
4449 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
4450 BasicBlock *BB = I++;
4452 // Does this BB end with the following?
4453 // %andVal = and %val, #single-bit-set
4454 // %icmpVal = icmp %andResult, 0
4455 // br i1 %cmpVal label %dest1, label %dest2"
4456 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
4457 if (!Brcc || !Brcc->isConditional())
4459 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
4460 if (!Cmp || Cmp->getParent() != BB)
4462 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
4463 if (!Zero || !Zero->isZero())
4465 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
4466 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
4468 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
4469 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
4471 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
4473 // Push the "and; icmp" for any users that are conditional branches.
4474 // Since there can only be one branch use per BB, we don't need to keep
4475 // track of which BBs we insert into.
4476 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
4480 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
4482 if (!BrccUser || !BrccUser->isConditional())
4484 BasicBlock *UserBB = BrccUser->getParent();
4485 if (UserBB == BB) continue;
4486 DEBUG(dbgs() << "found Brcc use\n");
4488 // Sink the "and; icmp" to use.
4490 BinaryOperator *NewAnd =
4491 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
4494 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
4498 DEBUG(BrccUser->getParent()->dump());
4504 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
4505 /// success, or returns false if no or invalid metadata was found.
4506 static bool extractBranchMetadata(BranchInst *BI,
4507 uint64_t &ProbTrue, uint64_t &ProbFalse) {
4508 assert(BI->isConditional() &&
4509 "Looking for probabilities on unconditional branch?");
4510 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
4511 if (!ProfileData || ProfileData->getNumOperands() != 3)
4514 const auto *CITrue =
4515 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
4516 const auto *CIFalse =
4517 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
4518 if (!CITrue || !CIFalse)
4521 ProbTrue = CITrue->getValue().getZExtValue();
4522 ProbFalse = CIFalse->getValue().getZExtValue();
4527 /// \brief Scale down both weights to fit into uint32_t.
4528 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
4529 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
4530 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
4531 NewTrue = NewTrue / Scale;
4532 NewFalse = NewFalse / Scale;
4535 /// \brief Some targets prefer to split a conditional branch like:
4537 /// %0 = icmp ne i32 %a, 0
4538 /// %1 = icmp ne i32 %b, 0
4539 /// %or.cond = or i1 %0, %1
4540 /// br i1 %or.cond, label %TrueBB, label %FalseBB
4542 /// into multiple branch instructions like:
4545 /// %0 = icmp ne i32 %a, 0
4546 /// br i1 %0, label %TrueBB, label %bb2
4548 /// %1 = icmp ne i32 %b, 0
4549 /// br i1 %1, label %TrueBB, label %FalseBB
4551 /// This usually allows instruction selection to do even further optimizations
4552 /// and combine the compare with the branch instruction. Currently this is
4553 /// applied for targets which have "cheap" jump instructions.
4555 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
4557 bool CodeGenPrepare::splitBranchCondition(Function &F) {
4558 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
4561 bool MadeChange = false;
4562 for (auto &BB : F) {
4563 // Does this BB end with the following?
4564 // %cond1 = icmp|fcmp|binary instruction ...
4565 // %cond2 = icmp|fcmp|binary instruction ...
4566 // %cond.or = or|and i1 %cond1, cond2
4567 // br i1 %cond.or label %dest1, label %dest2"
4568 BinaryOperator *LogicOp;
4569 BasicBlock *TBB, *FBB;
4570 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
4574 Value *Cond1, *Cond2;
4575 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
4576 m_OneUse(m_Value(Cond2)))))
4577 Opc = Instruction::And;
4578 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
4579 m_OneUse(m_Value(Cond2)))))
4580 Opc = Instruction::Or;
4584 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
4585 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
4588 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
4591 auto *InsertBefore = std::next(Function::iterator(BB))
4592 .getNodePtrUnchecked();
4593 auto TmpBB = BasicBlock::Create(BB.getContext(),
4594 BB.getName() + ".cond.split",
4595 BB.getParent(), InsertBefore);
4597 // Update original basic block by using the first condition directly by the
4598 // branch instruction and removing the no longer needed and/or instruction.
4599 auto *Br1 = cast<BranchInst>(BB.getTerminator());
4600 Br1->setCondition(Cond1);
4601 LogicOp->eraseFromParent();
4603 // Depending on the conditon we have to either replace the true or the false
4604 // successor of the original branch instruction.
4605 if (Opc == Instruction::And)
4606 Br1->setSuccessor(0, TmpBB);
4608 Br1->setSuccessor(1, TmpBB);
4610 // Fill in the new basic block.
4611 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
4612 if (auto *I = dyn_cast<Instruction>(Cond2)) {
4613 I->removeFromParent();
4614 I->insertBefore(Br2);
4617 // Update PHI nodes in both successors. The original BB needs to be
4618 // replaced in one succesor's PHI nodes, because the branch comes now from
4619 // the newly generated BB (NewBB). In the other successor we need to add one
4620 // incoming edge to the PHI nodes, because both branch instructions target
4621 // now the same successor. Depending on the original branch condition
4622 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
4623 // we perfrom the correct update for the PHI nodes.
4624 // This doesn't change the successor order of the just created branch
4625 // instruction (or any other instruction).
4626 if (Opc == Instruction::Or)
4627 std::swap(TBB, FBB);
4629 // Replace the old BB with the new BB.
4630 for (auto &I : *TBB) {
4631 PHINode *PN = dyn_cast<PHINode>(&I);
4635 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
4636 PN->setIncomingBlock(i, TmpBB);
4639 // Add another incoming edge form the new BB.
4640 for (auto &I : *FBB) {
4641 PHINode *PN = dyn_cast<PHINode>(&I);
4644 auto *Val = PN->getIncomingValueForBlock(&BB);
4645 PN->addIncoming(Val, TmpBB);
4648 // Update the branch weights (from SelectionDAGBuilder::
4649 // FindMergedConditions).
4650 if (Opc == Instruction::Or) {
4651 // Codegen X | Y as:
4660 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
4661 // The requirement is that
4662 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
4663 // = TrueProb for orignal BB.
4664 // Assuming the orignal weights are A and B, one choice is to set BB1's
4665 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
4667 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
4668 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
4669 // TmpBB, but the math is more complicated.
4670 uint64_t TrueWeight, FalseWeight;
4671 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4672 uint64_t NewTrueWeight = TrueWeight;
4673 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
4674 scaleWeights(NewTrueWeight, NewFalseWeight);
4675 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4676 .createBranchWeights(TrueWeight, FalseWeight));
4678 NewTrueWeight = TrueWeight;
4679 NewFalseWeight = 2 * FalseWeight;
4680 scaleWeights(NewTrueWeight, NewFalseWeight);
4681 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4682 .createBranchWeights(TrueWeight, FalseWeight));
4685 // Codegen X & Y as:
4693 // This requires creation of TmpBB after CurBB.
4695 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
4696 // The requirement is that
4697 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
4698 // = FalseProb for orignal BB.
4699 // Assuming the orignal weights are A and B, one choice is to set BB1's
4700 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
4702 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
4703 uint64_t TrueWeight, FalseWeight;
4704 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4705 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
4706 uint64_t NewFalseWeight = FalseWeight;
4707 scaleWeights(NewTrueWeight, NewFalseWeight);
4708 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4709 .createBranchWeights(TrueWeight, FalseWeight));
4711 NewTrueWeight = 2 * TrueWeight;
4712 NewFalseWeight = FalseWeight;
4713 scaleWeights(NewTrueWeight, NewFalseWeight);
4714 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4715 .createBranchWeights(TrueWeight, FalseWeight));
4719 // Note: No point in getting fancy here, since the DT info is never
4720 // available to CodeGenPrepare.
4725 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();