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;
112 typedef PointerIntPair<Type *, 1, bool> TypeIsSExt;
113 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
114 class TypePromotionTransaction;
116 class CodeGenPrepare : public FunctionPass {
117 /// TLI - Keep a pointer of a TargetLowering to consult for determining
118 /// transformation profitability.
119 const TargetMachine *TM;
120 const TargetLowering *TLI;
121 const TargetTransformInfo *TTI;
122 const TargetLibraryInfo *TLInfo;
124 /// CurInstIterator - As we scan instructions optimizing them, this is the
125 /// next instruction to optimize. Xforms that can invalidate this should
127 BasicBlock::iterator CurInstIterator;
129 /// Keeps track of non-local addresses that have been sunk into a block.
130 /// This allows us to avoid inserting duplicate code for blocks with
131 /// multiple load/stores of the same address.
132 ValueMap<Value*, Value*> SunkAddrs;
134 /// Keeps track of all instructions inserted for the current function.
135 SetOfInstrs InsertedInsts;
136 /// Keeps track of the type of the related instruction before their
137 /// promotion for the current function.
138 InstrToOrigTy PromotedInsts;
140 /// ModifiedDT - If CFG is modified in anyway.
143 /// OptSize - True if optimizing for size.
146 /// DataLayout for the Function being processed.
147 const DataLayout *DL;
150 static char ID; // Pass identification, replacement for typeid
151 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
152 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr), DL(nullptr) {
153 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
155 bool runOnFunction(Function &F) override;
157 const char *getPassName() const override { return "CodeGen Prepare"; }
159 void getAnalysisUsage(AnalysisUsage &AU) const override {
160 AU.addPreserved<DominatorTreeWrapperPass>();
161 AU.addRequired<TargetLibraryInfoWrapperPass>();
162 AU.addRequired<TargetTransformInfoWrapperPass>();
166 bool EliminateFallThrough(Function &F);
167 bool EliminateMostlyEmptyBlocks(Function &F);
168 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
169 void EliminateMostlyEmptyBlock(BasicBlock *BB);
170 bool OptimizeBlock(BasicBlock &BB, bool& ModifiedDT);
171 bool OptimizeInst(Instruction *I, bool& ModifiedDT);
172 bool OptimizeMemoryInst(Instruction *I, Value *Addr,
173 Type *AccessTy, unsigned AS);
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);
190 void stripInvariantGroupMetadata(Instruction &I);
194 char CodeGenPrepare::ID = 0;
195 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
196 "Optimize for code generation", false, false)
198 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
199 return new CodeGenPrepare(TM);
202 bool CodeGenPrepare::runOnFunction(Function &F) {
203 if (skipOptnoneFunction(F))
206 DL = &F.getParent()->getDataLayout();
208 bool EverMadeChange = false;
209 // Clear per function information.
210 InsertedInsts.clear();
211 PromotedInsts.clear();
215 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
216 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
217 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
218 OptSize = F.optForSize();
220 /// This optimization identifies DIV instructions that can be
221 /// profitably bypassed and carried out with a shorter, faster divide.
222 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
223 const DenseMap<unsigned int, unsigned int> &BypassWidths =
224 TLI->getBypassSlowDivWidths();
225 for (Function::iterator I = F.begin(); I != F.end(); I++)
226 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
229 // Eliminate blocks that contain only PHI nodes and an
230 // unconditional branch.
231 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
233 // llvm.dbg.value is far away from the value then iSel may not be able
234 // handle it properly. iSel will drop llvm.dbg.value if it can not
235 // find a node corresponding to the value.
236 EverMadeChange |= PlaceDbgValues(F);
238 // If there is a mask, compare against zero, and branch that can be combined
239 // into a single target instruction, push the mask and compare into branch
240 // users. Do this before OptimizeBlock -> OptimizeInst ->
241 // OptimizeCmpExpression, which perturbs the pattern being searched for.
242 if (!DisableBranchOpts) {
243 EverMadeChange |= sinkAndCmp(F);
244 EverMadeChange |= splitBranchCondition(F);
247 bool MadeChange = true;
250 for (Function::iterator I = F.begin(); I != F.end(); ) {
251 BasicBlock *BB = I++;
252 bool ModifiedDTOnIteration = false;
253 MadeChange |= OptimizeBlock(*BB, ModifiedDTOnIteration);
255 // Restart BB iteration if the dominator tree of the Function was changed
256 if (ModifiedDTOnIteration)
259 EverMadeChange |= MadeChange;
264 if (!DisableBranchOpts) {
266 SmallPtrSet<BasicBlock*, 8> WorkList;
267 for (BasicBlock &BB : F) {
268 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
269 MadeChange |= ConstantFoldTerminator(&BB, true);
270 if (!MadeChange) continue;
272 for (SmallVectorImpl<BasicBlock*>::iterator
273 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
274 if (pred_begin(*II) == pred_end(*II))
275 WorkList.insert(*II);
278 // Delete the dead blocks and any of their dead successors.
279 MadeChange |= !WorkList.empty();
280 while (!WorkList.empty()) {
281 BasicBlock *BB = *WorkList.begin();
283 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
287 for (SmallVectorImpl<BasicBlock*>::iterator
288 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
289 if (pred_begin(*II) == pred_end(*II))
290 WorkList.insert(*II);
293 // Merge pairs of basic blocks with unconditional branches, connected by
295 if (EverMadeChange || MadeChange)
296 MadeChange |= EliminateFallThrough(F);
298 EverMadeChange |= MadeChange;
301 if (!DisableGCOpts) {
302 SmallVector<Instruction *, 2> Statepoints;
303 for (BasicBlock &BB : F)
304 for (Instruction &I : BB)
306 Statepoints.push_back(&I);
307 for (auto &I : Statepoints)
308 EverMadeChange |= simplifyOffsetableRelocate(*I);
311 return EverMadeChange;
314 /// EliminateFallThrough - Merge basic blocks which are connected
315 /// by a single edge, where one of the basic blocks has a single successor
316 /// pointing to the other basic block, which has a single predecessor.
317 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
318 bool Changed = false;
319 // Scan all of the blocks in the function, except for the entry block.
320 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
321 BasicBlock *BB = I++;
322 // If the destination block has a single pred, then this is a trivial
323 // edge, just collapse it.
324 BasicBlock *SinglePred = BB->getSinglePredecessor();
326 // Don't merge if BB's address is taken.
327 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
329 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
330 if (Term && !Term->isConditional()) {
332 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
333 // Remember if SinglePred was the entry block of the function.
334 // If so, we will need to move BB back to the entry position.
335 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
336 MergeBasicBlockIntoOnlyPred(BB, nullptr);
338 if (isEntry && BB != &BB->getParent()->getEntryBlock())
339 BB->moveBefore(&BB->getParent()->getEntryBlock());
341 // We have erased a block. Update the iterator.
348 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
349 /// debug info directives, and an unconditional branch. Passes before isel
350 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
351 /// isel. Start by eliminating these blocks so we can split them the way we
353 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
354 bool MadeChange = false;
355 // Note that this intentionally skips the entry block.
356 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
357 BasicBlock *BB = I++;
359 // If this block doesn't end with an uncond branch, ignore it.
360 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
361 if (!BI || !BI->isUnconditional())
364 // If the instruction before the branch (skipping debug info) isn't a phi
365 // node, then other stuff is happening here.
366 BasicBlock::iterator BBI = BI;
367 if (BBI != BB->begin()) {
369 while (isa<DbgInfoIntrinsic>(BBI)) {
370 if (BBI == BB->begin())
374 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
378 // Do not break infinite loops.
379 BasicBlock *DestBB = BI->getSuccessor(0);
383 if (!CanMergeBlocks(BB, DestBB))
386 EliminateMostlyEmptyBlock(BB);
392 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
393 /// single uncond branch between them, and BB contains no other non-phi
395 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
396 const BasicBlock *DestBB) const {
397 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
398 // the successor. If there are more complex condition (e.g. preheaders),
399 // don't mess around with them.
400 BasicBlock::const_iterator BBI = BB->begin();
401 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
402 for (const User *U : PN->users()) {
403 const Instruction *UI = cast<Instruction>(U);
404 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
406 // If User is inside DestBB block and it is a PHINode then check
407 // incoming value. If incoming value is not from BB then this is
408 // a complex condition (e.g. preheaders) we want to avoid here.
409 if (UI->getParent() == DestBB) {
410 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
411 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
412 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
413 if (Insn && Insn->getParent() == BB &&
414 Insn->getParent() != UPN->getIncomingBlock(I))
421 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
422 // and DestBB may have conflicting incoming values for the block. If so, we
423 // can't merge the block.
424 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
425 if (!DestBBPN) return true; // no conflict.
427 // Collect the preds of BB.
428 SmallPtrSet<const BasicBlock*, 16> BBPreds;
429 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
430 // It is faster to get preds from a PHI than with pred_iterator.
431 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
432 BBPreds.insert(BBPN->getIncomingBlock(i));
434 BBPreds.insert(pred_begin(BB), pred_end(BB));
437 // Walk the preds of DestBB.
438 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
439 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
440 if (BBPreds.count(Pred)) { // Common predecessor?
441 BBI = DestBB->begin();
442 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
443 const Value *V1 = PN->getIncomingValueForBlock(Pred);
444 const Value *V2 = PN->getIncomingValueForBlock(BB);
446 // If V2 is a phi node in BB, look up what the mapped value will be.
447 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
448 if (V2PN->getParent() == BB)
449 V2 = V2PN->getIncomingValueForBlock(Pred);
451 // If there is a conflict, bail out.
452 if (V1 != V2) return false;
461 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
462 /// an unconditional branch in it.
463 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
464 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
465 BasicBlock *DestBB = BI->getSuccessor(0);
467 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
469 // If the destination block has a single pred, then this is a trivial edge,
471 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
472 if (SinglePred != DestBB) {
473 // Remember if SinglePred was the entry block of the function. If so, we
474 // will need to move BB back to the entry position.
475 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
476 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
478 if (isEntry && BB != &BB->getParent()->getEntryBlock())
479 BB->moveBefore(&BB->getParent()->getEntryBlock());
481 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
486 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
487 // to handle the new incoming edges it is about to have.
489 for (BasicBlock::iterator BBI = DestBB->begin();
490 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
491 // Remove the incoming value for BB, and remember it.
492 Value *InVal = PN->removeIncomingValue(BB, false);
494 // Two options: either the InVal is a phi node defined in BB or it is some
495 // value that dominates BB.
496 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
497 if (InValPhi && InValPhi->getParent() == BB) {
498 // Add all of the input values of the input PHI as inputs of this phi.
499 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
500 PN->addIncoming(InValPhi->getIncomingValue(i),
501 InValPhi->getIncomingBlock(i));
503 // Otherwise, add one instance of the dominating value for each edge that
504 // we will be adding.
505 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
506 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
507 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
509 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
510 PN->addIncoming(InVal, *PI);
515 // The PHIs are now updated, change everything that refers to BB to use
516 // DestBB and remove BB.
517 BB->replaceAllUsesWith(DestBB);
518 BB->eraseFromParent();
521 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
524 // Computes a map of base pointer relocation instructions to corresponding
525 // derived pointer relocation instructions given a vector of all relocate calls
526 static void computeBaseDerivedRelocateMap(
527 const SmallVectorImpl<User *> &AllRelocateCalls,
528 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
530 // Collect information in two maps: one primarily for locating the base object
531 // while filling the second map; the second map is the final structure holding
532 // a mapping between Base and corresponding Derived relocate calls
533 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
534 for (auto &U : AllRelocateCalls) {
535 GCRelocateOperands ThisRelocate(U);
536 IntrinsicInst *I = cast<IntrinsicInst>(U);
537 auto K = std::make_pair(ThisRelocate.getBasePtrIndex(),
538 ThisRelocate.getDerivedPtrIndex());
539 RelocateIdxMap.insert(std::make_pair(K, I));
541 for (auto &Item : RelocateIdxMap) {
542 std::pair<unsigned, unsigned> Key = Item.first;
543 if (Key.first == Key.second)
544 // Base relocation: nothing to insert
547 IntrinsicInst *I = Item.second;
548 auto BaseKey = std::make_pair(Key.first, Key.first);
550 // We're iterating over RelocateIdxMap so we cannot modify it.
551 auto MaybeBase = RelocateIdxMap.find(BaseKey);
552 if (MaybeBase == RelocateIdxMap.end())
553 // TODO: We might want to insert a new base object relocate and gep off
554 // that, if there are enough derived object relocates.
557 RelocateInstMap[MaybeBase->second].push_back(I);
561 // Accepts a GEP and extracts the operands into a vector provided they're all
562 // small integer constants
563 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
564 SmallVectorImpl<Value *> &OffsetV) {
565 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
566 // Only accept small constant integer operands
567 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
568 if (!Op || Op->getZExtValue() > 20)
572 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
573 OffsetV.push_back(GEP->getOperand(i));
577 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
578 // replace, computes a replacement, and affects it.
580 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
581 const SmallVectorImpl<IntrinsicInst *> &Targets) {
582 bool MadeChange = false;
583 for (auto &ToReplace : Targets) {
584 GCRelocateOperands MasterRelocate(RelocatedBase);
585 GCRelocateOperands ThisRelocate(ToReplace);
587 assert(ThisRelocate.getBasePtrIndex() == MasterRelocate.getBasePtrIndex() &&
588 "Not relocating a derived object of the original base object");
589 if (ThisRelocate.getBasePtrIndex() == ThisRelocate.getDerivedPtrIndex()) {
590 // A duplicate relocate call. TODO: coalesce duplicates.
594 Value *Base = ThisRelocate.getBasePtr();
595 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.getDerivedPtr());
596 if (!Derived || Derived->getPointerOperand() != Base)
599 SmallVector<Value *, 2> OffsetV;
600 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
603 // Create a Builder and replace the target callsite with a gep
604 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
606 // Insert after RelocatedBase
607 IRBuilder<> Builder(RelocatedBase->getNextNode());
608 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
610 // If gc_relocate does not match the actual type, cast it to the right type.
611 // In theory, there must be a bitcast after gc_relocate if the type does not
612 // match, and we should reuse it to get the derived pointer. But it could be
616 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
621 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
625 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
626 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
628 // In this case, we can not find the bitcast any more. So we insert a new bitcast
629 // no matter there is already one or not. In this way, we can handle all cases, and
630 // the extra bitcast should be optimized away in later passes.
631 Instruction *ActualRelocatedBase = RelocatedBase;
632 if (RelocatedBase->getType() != Base->getType()) {
633 ActualRelocatedBase =
634 cast<Instruction>(Builder.CreateBitCast(RelocatedBase, Base->getType()));
636 Value *Replacement = Builder.CreateGEP(
637 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
638 Instruction *ReplacementInst = cast<Instruction>(Replacement);
639 Replacement->takeName(ToReplace);
640 // If the newly generated derived pointer's type does not match the original derived
641 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
642 Instruction *ActualReplacement = ReplacementInst;
643 if (ReplacementInst->getType() != ToReplace->getType()) {
645 cast<Instruction>(Builder.CreateBitCast(ReplacementInst, ToReplace->getType()));
647 ToReplace->replaceAllUsesWith(ActualReplacement);
648 ToReplace->eraseFromParent();
658 // %ptr = gep %base + 15
659 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
660 // %base' = relocate(%tok, i32 4, i32 4)
661 // %ptr' = relocate(%tok, i32 4, i32 5)
667 // %ptr = gep %base + 15
668 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
669 // %base' = gc.relocate(%tok, i32 4, i32 4)
670 // %ptr' = gep %base' + 15
672 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
673 bool MadeChange = false;
674 SmallVector<User *, 2> AllRelocateCalls;
676 for (auto *U : I.users())
677 if (isGCRelocate(dyn_cast<Instruction>(U)))
678 // Collect all the relocate calls associated with a statepoint
679 AllRelocateCalls.push_back(U);
681 // We need atleast one base pointer relocation + one derived pointer
682 // relocation to mangle
683 if (AllRelocateCalls.size() < 2)
686 // RelocateInstMap is a mapping from the base relocate instruction to the
687 // corresponding derived relocate instructions
688 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
689 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
690 if (RelocateInstMap.empty())
693 for (auto &Item : RelocateInstMap)
694 // Item.first is the RelocatedBase to offset against
695 // Item.second is the vector of Targets to replace
696 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
700 /// SinkCast - Sink the specified cast instruction into its user blocks
701 static bool SinkCast(CastInst *CI) {
702 BasicBlock *DefBB = CI->getParent();
704 /// InsertedCasts - Only insert a cast in each block once.
705 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
707 bool MadeChange = false;
708 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
710 Use &TheUse = UI.getUse();
711 Instruction *User = cast<Instruction>(*UI);
713 // Figure out which BB this cast is used in. For PHI's this is the
714 // appropriate predecessor block.
715 BasicBlock *UserBB = User->getParent();
716 if (PHINode *PN = dyn_cast<PHINode>(User)) {
717 UserBB = PN->getIncomingBlock(TheUse);
720 // Preincrement use iterator so we don't invalidate it.
723 // If this user is in the same block as the cast, don't change the cast.
724 if (UserBB == DefBB) continue;
726 // If we have already inserted a cast into this block, use it.
727 CastInst *&InsertedCast = InsertedCasts[UserBB];
730 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
732 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
736 // Replace a use of the cast with a use of the new cast.
737 TheUse = InsertedCast;
742 // If we removed all uses, nuke the cast.
743 if (CI->use_empty()) {
744 CI->eraseFromParent();
751 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
752 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
753 /// sink it into user blocks to reduce the number of virtual
754 /// registers that must be created and coalesced.
756 /// Return true if any changes are made.
758 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
759 const DataLayout &DL) {
760 // If this is a noop copy,
761 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
762 EVT DstVT = TLI.getValueType(DL, CI->getType());
764 // This is an fp<->int conversion?
765 if (SrcVT.isInteger() != DstVT.isInteger())
768 // If this is an extension, it will be a zero or sign extension, which
770 if (SrcVT.bitsLT(DstVT)) return false;
772 // If these values will be promoted, find out what they will be promoted
773 // to. This helps us consider truncates on PPC as noop copies when they
775 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
776 TargetLowering::TypePromoteInteger)
777 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
778 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
779 TargetLowering::TypePromoteInteger)
780 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
782 // If, after promotion, these are the same types, this is a noop copy.
789 /// CombineUAddWithOverflow - try to combine CI into a call to the
790 /// llvm.uadd.with.overflow intrinsic if possible.
792 /// Return true if any changes were made.
793 static bool CombineUAddWithOverflow(CmpInst *CI) {
797 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
800 Type *Ty = AddI->getType();
801 if (!isa<IntegerType>(Ty))
804 // We don't want to move around uses of condition values this late, so we we
805 // check if it is legal to create the call to the intrinsic in the basic
806 // block containing the icmp:
808 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
812 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
814 if (AddI->hasOneUse())
815 assert(*AddI->user_begin() == CI && "expected!");
818 Module *M = CI->getParent()->getParent()->getParent();
819 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
821 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
823 auto *UAddWithOverflow =
824 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
825 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
827 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
829 CI->replaceAllUsesWith(Overflow);
830 AddI->replaceAllUsesWith(UAdd);
831 CI->eraseFromParent();
832 AddI->eraseFromParent();
836 /// SinkCmpExpression - Sink the given CmpInst into user blocks to reduce
837 /// the number of virtual registers that must be created and coalesced. This is
838 /// a clear win except on targets with multiple condition code registers
839 /// (PowerPC), where it might lose; some adjustment may be wanted there.
841 /// Return true if any changes are made.
842 static bool SinkCmpExpression(CmpInst *CI) {
843 BasicBlock *DefBB = CI->getParent();
845 /// InsertedCmp - Only insert a cmp in each block once.
846 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
848 bool MadeChange = false;
849 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
851 Use &TheUse = UI.getUse();
852 Instruction *User = cast<Instruction>(*UI);
854 // Preincrement use iterator so we don't invalidate it.
857 // Don't bother for PHI nodes.
858 if (isa<PHINode>(User))
861 // Figure out which BB this cmp is used in.
862 BasicBlock *UserBB = User->getParent();
864 // If this user is in the same block as the cmp, don't change the cmp.
865 if (UserBB == DefBB) continue;
867 // If we have already inserted a cmp into this block, use it.
868 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
871 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
873 CmpInst::Create(CI->getOpcode(),
874 CI->getPredicate(), CI->getOperand(0),
875 CI->getOperand(1), "", InsertPt);
878 // Replace a use of the cmp with a use of the new cmp.
879 TheUse = InsertedCmp;
884 // If we removed all uses, nuke the cmp.
885 if (CI->use_empty()) {
886 CI->eraseFromParent();
893 static bool OptimizeCmpExpression(CmpInst *CI) {
894 if (SinkCmpExpression(CI))
897 if (CombineUAddWithOverflow(CI))
903 /// isExtractBitsCandidateUse - Check if the candidates could
904 /// be combined with shift instruction, which includes:
905 /// 1. Truncate instruction
906 /// 2. And instruction and the imm is a mask of the low bits:
907 /// imm & (imm+1) == 0
908 static bool isExtractBitsCandidateUse(Instruction *User) {
909 if (!isa<TruncInst>(User)) {
910 if (User->getOpcode() != Instruction::And ||
911 !isa<ConstantInt>(User->getOperand(1)))
914 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
916 if ((Cimm & (Cimm + 1)).getBoolValue())
922 /// SinkShiftAndTruncate - sink both shift and truncate instruction
923 /// to the use of truncate's BB.
925 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
926 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
927 const TargetLowering &TLI, const DataLayout &DL) {
928 BasicBlock *UserBB = User->getParent();
929 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
930 TruncInst *TruncI = dyn_cast<TruncInst>(User);
931 bool MadeChange = false;
933 for (Value::user_iterator TruncUI = TruncI->user_begin(),
934 TruncE = TruncI->user_end();
935 TruncUI != TruncE;) {
937 Use &TruncTheUse = TruncUI.getUse();
938 Instruction *TruncUser = cast<Instruction>(*TruncUI);
939 // Preincrement use iterator so we don't invalidate it.
943 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
947 // If the use is actually a legal node, there will not be an
948 // implicit truncate.
949 // FIXME: always querying the result type is just an
950 // approximation; some nodes' legality is determined by the
951 // operand or other means. There's no good way to find out though.
952 if (TLI.isOperationLegalOrCustom(
953 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
956 // Don't bother for PHI nodes.
957 if (isa<PHINode>(TruncUser))
960 BasicBlock *TruncUserBB = TruncUser->getParent();
962 if (UserBB == TruncUserBB)
965 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
966 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
968 if (!InsertedShift && !InsertedTrunc) {
969 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
971 if (ShiftI->getOpcode() == Instruction::AShr)
973 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
976 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
979 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
982 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
983 TruncI->getType(), "", TruncInsertPt);
987 TruncTheUse = InsertedTrunc;
993 /// OptimizeExtractBits - sink the shift *right* instruction into user blocks if
994 /// the uses could potentially be combined with this shift instruction and
995 /// generate BitExtract instruction. It will only be applied if the architecture
996 /// supports BitExtract instruction. Here is an example:
998 /// %x.extract.shift = lshr i64 %arg1, 32
1000 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1004 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1005 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1007 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1009 /// Return true if any changes are made.
1010 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1011 const TargetLowering &TLI,
1012 const DataLayout &DL) {
1013 BasicBlock *DefBB = ShiftI->getParent();
1015 /// Only insert instructions in each block once.
1016 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1018 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1020 bool MadeChange = false;
1021 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1023 Use &TheUse = UI.getUse();
1024 Instruction *User = cast<Instruction>(*UI);
1025 // Preincrement use iterator so we don't invalidate it.
1028 // Don't bother for PHI nodes.
1029 if (isa<PHINode>(User))
1032 if (!isExtractBitsCandidateUse(User))
1035 BasicBlock *UserBB = User->getParent();
1037 if (UserBB == DefBB) {
1038 // If the shift and truncate instruction are in the same BB. The use of
1039 // the truncate(TruncUse) may still introduce another truncate if not
1040 // legal. In this case, we would like to sink both shift and truncate
1041 // instruction to the BB of TruncUse.
1044 // i64 shift.result = lshr i64 opnd, imm
1045 // trunc.result = trunc shift.result to i16
1048 // ----> We will have an implicit truncate here if the architecture does
1049 // not have i16 compare.
1050 // cmp i16 trunc.result, opnd2
1052 if (isa<TruncInst>(User) && shiftIsLegal
1053 // If the type of the truncate is legal, no trucate will be
1054 // introduced in other basic blocks.
1056 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1058 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1062 // If we have already inserted a shift into this block, use it.
1063 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1065 if (!InsertedShift) {
1066 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1068 if (ShiftI->getOpcode() == Instruction::AShr)
1070 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
1073 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
1078 // Replace a use of the shift with a use of the new shift.
1079 TheUse = InsertedShift;
1082 // If we removed all uses, nuke the shift.
1083 if (ShiftI->use_empty())
1084 ShiftI->eraseFromParent();
1089 // ScalarizeMaskedLoad() translates masked load intrinsic, like
1090 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1091 // <16 x i1> %mask, <16 x i32> %passthru)
1092 // to a chain of basic blocks, with loading element one-by-one if
1093 // the appropriate mask bit is set
1095 // %1 = bitcast i8* %addr to i32*
1096 // %2 = extractelement <16 x i1> %mask, i32 0
1097 // %3 = icmp eq i1 %2, true
1098 // br i1 %3, label %cond.load, label %else
1100 //cond.load: ; preds = %0
1101 // %4 = getelementptr i32* %1, i32 0
1102 // %5 = load i32* %4
1103 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1106 //else: ; preds = %0, %cond.load
1107 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1108 // %7 = extractelement <16 x i1> %mask, i32 1
1109 // %8 = icmp eq i1 %7, true
1110 // br i1 %8, label %cond.load1, label %else2
1112 //cond.load1: ; preds = %else
1113 // %9 = getelementptr i32* %1, i32 1
1114 // %10 = load i32* %9
1115 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1118 //else2: ; preds = %else, %cond.load1
1119 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1120 // %12 = extractelement <16 x i1> %mask, i32 2
1121 // %13 = icmp eq i1 %12, true
1122 // br i1 %13, label %cond.load4, label %else5
1124 static void ScalarizeMaskedLoad(CallInst *CI) {
1125 Value *Ptr = CI->getArgOperand(0);
1126 Value *Src0 = CI->getArgOperand(3);
1127 Value *Mask = CI->getArgOperand(2);
1128 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1129 Type *EltTy = VecType->getElementType();
1131 assert(VecType && "Unexpected return type of masked load intrinsic");
1133 IRBuilder<> Builder(CI->getContext());
1134 Instruction *InsertPt = CI;
1135 BasicBlock *IfBlock = CI->getParent();
1136 BasicBlock *CondBlock = nullptr;
1137 BasicBlock *PrevIfBlock = CI->getParent();
1138 Builder.SetInsertPoint(InsertPt);
1140 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1142 // Bitcast %addr fron i8* to EltTy*
1144 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1145 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1146 Value *UndefVal = UndefValue::get(VecType);
1148 // The result vector
1149 Value *VResult = UndefVal;
1151 PHINode *Phi = nullptr;
1152 Value *PrevPhi = UndefVal;
1154 unsigned VectorWidth = VecType->getNumElements();
1155 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1157 // Fill the "else" block, created in the previous iteration
1159 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1160 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1161 // %to_load = icmp eq i1 %mask_1, true
1162 // br i1 %to_load, label %cond.load, label %else
1165 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1166 Phi->addIncoming(VResult, CondBlock);
1167 Phi->addIncoming(PrevPhi, PrevIfBlock);
1172 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1173 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1174 ConstantInt::get(Predicate->getType(), 1));
1176 // Create "cond" block
1178 // %EltAddr = getelementptr i32* %1, i32 0
1179 // %Elt = load i32* %EltAddr
1180 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1182 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1183 Builder.SetInsertPoint(InsertPt);
1186 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1187 LoadInst* Load = Builder.CreateLoad(Gep, false);
1188 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1190 // Create "else" block, fill it in the next iteration
1191 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1192 Builder.SetInsertPoint(InsertPt);
1193 Instruction *OldBr = IfBlock->getTerminator();
1194 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1195 OldBr->eraseFromParent();
1196 PrevIfBlock = IfBlock;
1197 IfBlock = NewIfBlock;
1200 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1201 Phi->addIncoming(VResult, CondBlock);
1202 Phi->addIncoming(PrevPhi, PrevIfBlock);
1203 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1204 CI->replaceAllUsesWith(NewI);
1205 CI->eraseFromParent();
1208 // ScalarizeMaskedStore() translates masked store intrinsic, like
1209 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1211 // to a chain of basic blocks, that stores element one-by-one if
1212 // the appropriate mask bit is set
1214 // %1 = bitcast i8* %addr to i32*
1215 // %2 = extractelement <16 x i1> %mask, i32 0
1216 // %3 = icmp eq i1 %2, true
1217 // br i1 %3, label %cond.store, label %else
1219 // cond.store: ; preds = %0
1220 // %4 = extractelement <16 x i32> %val, i32 0
1221 // %5 = getelementptr i32* %1, i32 0
1222 // store i32 %4, i32* %5
1225 // else: ; preds = %0, %cond.store
1226 // %6 = extractelement <16 x i1> %mask, i32 1
1227 // %7 = icmp eq i1 %6, true
1228 // br i1 %7, label %cond.store1, label %else2
1230 // cond.store1: ; preds = %else
1231 // %8 = extractelement <16 x i32> %val, i32 1
1232 // %9 = getelementptr i32* %1, i32 1
1233 // store i32 %8, i32* %9
1236 static void ScalarizeMaskedStore(CallInst *CI) {
1237 Value *Ptr = CI->getArgOperand(1);
1238 Value *Src = CI->getArgOperand(0);
1239 Value *Mask = CI->getArgOperand(3);
1241 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1242 Type *EltTy = VecType->getElementType();
1244 assert(VecType && "Unexpected data type in masked store intrinsic");
1246 IRBuilder<> Builder(CI->getContext());
1247 Instruction *InsertPt = CI;
1248 BasicBlock *IfBlock = CI->getParent();
1249 Builder.SetInsertPoint(InsertPt);
1250 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1252 // Bitcast %addr fron i8* to EltTy*
1254 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1255 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1257 unsigned VectorWidth = VecType->getNumElements();
1258 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1260 // Fill the "else" block, created in the previous iteration
1262 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1263 // %to_store = icmp eq i1 %mask_1, true
1264 // br i1 %to_load, label %cond.store, label %else
1266 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1267 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1268 ConstantInt::get(Predicate->getType(), 1));
1270 // Create "cond" block
1272 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1273 // %EltAddr = getelementptr i32* %1, i32 0
1274 // %store i32 %OneElt, i32* %EltAddr
1276 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1277 Builder.SetInsertPoint(InsertPt);
1279 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1281 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1282 Builder.CreateStore(OneElt, Gep);
1284 // Create "else" block, fill it in the next iteration
1285 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1286 Builder.SetInsertPoint(InsertPt);
1287 Instruction *OldBr = IfBlock->getTerminator();
1288 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1289 OldBr->eraseFromParent();
1290 IfBlock = NewIfBlock;
1292 CI->eraseFromParent();
1295 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1296 BasicBlock *BB = CI->getParent();
1298 // Lower inline assembly if we can.
1299 // If we found an inline asm expession, and if the target knows how to
1300 // lower it to normal LLVM code, do so now.
1301 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1302 if (TLI->ExpandInlineAsm(CI)) {
1303 // Avoid invalidating the iterator.
1304 CurInstIterator = BB->begin();
1305 // Avoid processing instructions out of order, which could cause
1306 // reuse before a value is defined.
1310 // Sink address computing for memory operands into the block.
1311 if (OptimizeInlineAsmInst(CI))
1315 // Align the pointer arguments to this call if the target thinks it's a good
1317 unsigned MinSize, PrefAlign;
1318 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1319 for (auto &Arg : CI->arg_operands()) {
1320 // We want to align both objects whose address is used directly and
1321 // objects whose address is used in casts and GEPs, though it only makes
1322 // sense for GEPs if the offset is a multiple of the desired alignment and
1323 // if size - offset meets the size threshold.
1324 if (!Arg->getType()->isPointerTy())
1326 APInt Offset(DL->getPointerSizeInBits(
1327 cast<PointerType>(Arg->getType())->getAddressSpace()),
1329 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
1330 uint64_t Offset2 = Offset.getLimitedValue();
1331 if ((Offset2 & (PrefAlign-1)) != 0)
1334 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1335 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1336 AI->setAlignment(PrefAlign);
1337 // Global variables can only be aligned if they are defined in this
1338 // object (i.e. they are uniquely initialized in this object), and
1339 // over-aligning global variables that have an explicit section is
1342 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->hasUniqueInitializer() &&
1343 !GV->hasSection() && GV->getAlignment() < PrefAlign &&
1344 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
1346 GV->setAlignment(PrefAlign);
1348 // If this is a memcpy (or similar) then we may be able to improve the
1350 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1351 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
1352 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1353 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
1354 if (Align > MI->getAlignment())
1355 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1359 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1361 switch (II->getIntrinsicID()) {
1363 case Intrinsic::objectsize: {
1364 // Lower all uses of llvm.objectsize.*
1365 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1366 Type *ReturnTy = CI->getType();
1367 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1369 // Substituting this can cause recursive simplifications, which can
1370 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1372 WeakVH IterHandle(CurInstIterator);
1374 replaceAndRecursivelySimplify(CI, RetVal,
1377 // If the iterator instruction was recursively deleted, start over at the
1378 // start of the block.
1379 if (IterHandle != CurInstIterator) {
1380 CurInstIterator = BB->begin();
1385 case Intrinsic::masked_load: {
1386 // Scalarize unsupported vector masked load
1387 if (!TTI->isLegalMaskedLoad(CI->getType(), 1)) {
1388 ScalarizeMaskedLoad(CI);
1394 case Intrinsic::masked_store: {
1395 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType(), 1)) {
1396 ScalarizeMaskedStore(CI);
1402 case Intrinsic::aarch64_stlxr:
1403 case Intrinsic::aarch64_stxr: {
1404 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1405 if (!ExtVal || !ExtVal->hasOneUse() ||
1406 ExtVal->getParent() == CI->getParent())
1408 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1409 ExtVal->moveBefore(CI);
1410 // Mark this instruction as "inserted by CGP", so that other
1411 // optimizations don't touch it.
1412 InsertedInsts.insert(ExtVal);
1415 case Intrinsic::invariant_group_barrier:
1416 II->replaceAllUsesWith(II->getArgOperand(0));
1417 II->eraseFromParent();
1422 // Unknown address space.
1423 // TODO: Target hook to pick which address space the intrinsic cares
1425 unsigned AddrSpace = ~0u;
1426 SmallVector<Value*, 2> PtrOps;
1428 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
1429 while (!PtrOps.empty())
1430 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
1435 // From here on out we're working with named functions.
1436 if (!CI->getCalledFunction()) return false;
1438 // Lower all default uses of _chk calls. This is very similar
1439 // to what InstCombineCalls does, but here we are only lowering calls
1440 // to fortified library functions (e.g. __memcpy_chk) that have the default
1441 // "don't know" as the objectsize. Anything else should be left alone.
1442 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1443 if (Value *V = Simplifier.optimizeCall(CI)) {
1444 CI->replaceAllUsesWith(V);
1445 CI->eraseFromParent();
1451 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
1452 /// instructions to the predecessor to enable tail call optimizations. The
1453 /// case it is currently looking for is:
1456 /// %tmp0 = tail call i32 @f0()
1457 /// br label %return
1459 /// %tmp1 = tail call i32 @f1()
1460 /// br label %return
1462 /// %tmp2 = tail call i32 @f2()
1463 /// br label %return
1465 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1473 /// %tmp0 = tail call i32 @f0()
1476 /// %tmp1 = tail call i32 @f1()
1479 /// %tmp2 = tail call i32 @f2()
1482 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
1486 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1490 PHINode *PN = nullptr;
1491 BitCastInst *BCI = nullptr;
1492 Value *V = RI->getReturnValue();
1494 BCI = dyn_cast<BitCastInst>(V);
1496 V = BCI->getOperand(0);
1498 PN = dyn_cast<PHINode>(V);
1503 if (PN && PN->getParent() != BB)
1506 // It's not safe to eliminate the sign / zero extension of the return value.
1507 // See llvm::isInTailCallPosition().
1508 const Function *F = BB->getParent();
1509 AttributeSet CallerAttrs = F->getAttributes();
1510 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1511 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1514 // Make sure there are no instructions between the PHI and return, or that the
1515 // return is the first instruction in the block.
1517 BasicBlock::iterator BI = BB->begin();
1518 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1520 // Also skip over the bitcast.
1525 BasicBlock::iterator BI = BB->begin();
1526 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1531 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1533 SmallVector<CallInst*, 4> TailCalls;
1535 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1536 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1537 // Make sure the phi value is indeed produced by the tail call.
1538 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1539 TLI->mayBeEmittedAsTailCall(CI))
1540 TailCalls.push_back(CI);
1543 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1544 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1545 if (!VisitedBBs.insert(*PI).second)
1548 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1549 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1550 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1551 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1555 CallInst *CI = dyn_cast<CallInst>(&*RI);
1556 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1557 TailCalls.push_back(CI);
1561 bool Changed = false;
1562 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1563 CallInst *CI = TailCalls[i];
1566 // Conservatively require the attributes of the call to match those of the
1567 // return. Ignore noalias because it doesn't affect the call sequence.
1568 AttributeSet CalleeAttrs = CS.getAttributes();
1569 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1570 removeAttribute(Attribute::NoAlias) !=
1571 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1572 removeAttribute(Attribute::NoAlias))
1575 // Make sure the call instruction is followed by an unconditional branch to
1576 // the return block.
1577 BasicBlock *CallBB = CI->getParent();
1578 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1579 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1582 // Duplicate the return into CallBB.
1583 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1584 ModifiedDT = Changed = true;
1588 // If we eliminated all predecessors of the block, delete the block now.
1589 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1590 BB->eraseFromParent();
1595 //===----------------------------------------------------------------------===//
1596 // Memory Optimization
1597 //===----------------------------------------------------------------------===//
1601 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1602 /// which holds actual Value*'s for register values.
1603 struct ExtAddrMode : public TargetLowering::AddrMode {
1606 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1607 void print(raw_ostream &OS) const;
1610 bool operator==(const ExtAddrMode& O) const {
1611 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1612 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1613 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1618 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1624 void ExtAddrMode::print(raw_ostream &OS) const {
1625 bool NeedPlus = false;
1628 OS << (NeedPlus ? " + " : "")
1630 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1635 OS << (NeedPlus ? " + " : "")
1641 OS << (NeedPlus ? " + " : "")
1643 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1647 OS << (NeedPlus ? " + " : "")
1649 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1655 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1656 void ExtAddrMode::dump() const {
1662 /// \brief This class provides transaction based operation on the IR.
1663 /// Every change made through this class is recorded in the internal state and
1664 /// can be undone (rollback) until commit is called.
1665 class TypePromotionTransaction {
1667 /// \brief This represents the common interface of the individual transaction.
1668 /// Each class implements the logic for doing one specific modification on
1669 /// the IR via the TypePromotionTransaction.
1670 class TypePromotionAction {
1672 /// The Instruction modified.
1676 /// \brief Constructor of the action.
1677 /// The constructor performs the related action on the IR.
1678 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1680 virtual ~TypePromotionAction() {}
1682 /// \brief Undo the modification done by this action.
1683 /// When this method is called, the IR must be in the same state as it was
1684 /// before this action was applied.
1685 /// \pre Undoing the action works if and only if the IR is in the exact same
1686 /// state as it was directly after this action was applied.
1687 virtual void undo() = 0;
1689 /// \brief Advocate every change made by this action.
1690 /// When the results on the IR of the action are to be kept, it is important
1691 /// to call this function, otherwise hidden information may be kept forever.
1692 virtual void commit() {
1693 // Nothing to be done, this action is not doing anything.
1697 /// \brief Utility to remember the position of an instruction.
1698 class InsertionHandler {
1699 /// Position of an instruction.
1700 /// Either an instruction:
1701 /// - Is the first in a basic block: BB is used.
1702 /// - Has a previous instructon: PrevInst is used.
1704 Instruction *PrevInst;
1707 /// Remember whether or not the instruction had a previous instruction.
1708 bool HasPrevInstruction;
1711 /// \brief Record the position of \p Inst.
1712 InsertionHandler(Instruction *Inst) {
1713 BasicBlock::iterator It = Inst;
1714 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1715 if (HasPrevInstruction)
1716 Point.PrevInst = --It;
1718 Point.BB = Inst->getParent();
1721 /// \brief Insert \p Inst at the recorded position.
1722 void insert(Instruction *Inst) {
1723 if (HasPrevInstruction) {
1724 if (Inst->getParent())
1725 Inst->removeFromParent();
1726 Inst->insertAfter(Point.PrevInst);
1728 Instruction *Position = Point.BB->getFirstInsertionPt();
1729 if (Inst->getParent())
1730 Inst->moveBefore(Position);
1732 Inst->insertBefore(Position);
1737 /// \brief Move an instruction before another.
1738 class InstructionMoveBefore : public TypePromotionAction {
1739 /// Original position of the instruction.
1740 InsertionHandler Position;
1743 /// \brief Move \p Inst before \p Before.
1744 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1745 : TypePromotionAction(Inst), Position(Inst) {
1746 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1747 Inst->moveBefore(Before);
1750 /// \brief Move the instruction back to its original position.
1751 void undo() override {
1752 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1753 Position.insert(Inst);
1757 /// \brief Set the operand of an instruction with a new value.
1758 class OperandSetter : public TypePromotionAction {
1759 /// Original operand of the instruction.
1761 /// Index of the modified instruction.
1765 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1766 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1767 : TypePromotionAction(Inst), Idx(Idx) {
1768 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1769 << "for:" << *Inst << "\n"
1770 << "with:" << *NewVal << "\n");
1771 Origin = Inst->getOperand(Idx);
1772 Inst->setOperand(Idx, NewVal);
1775 /// \brief Restore the original value of the instruction.
1776 void undo() override {
1777 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1778 << "for: " << *Inst << "\n"
1779 << "with: " << *Origin << "\n");
1780 Inst->setOperand(Idx, Origin);
1784 /// \brief Hide the operands of an instruction.
1785 /// Do as if this instruction was not using any of its operands.
1786 class OperandsHider : public TypePromotionAction {
1787 /// The list of original operands.
1788 SmallVector<Value *, 4> OriginalValues;
1791 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1792 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1793 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1794 unsigned NumOpnds = Inst->getNumOperands();
1795 OriginalValues.reserve(NumOpnds);
1796 for (unsigned It = 0; It < NumOpnds; ++It) {
1797 // Save the current operand.
1798 Value *Val = Inst->getOperand(It);
1799 OriginalValues.push_back(Val);
1801 // We could use OperandSetter here, but that would implied an overhead
1802 // that we are not willing to pay.
1803 Inst->setOperand(It, UndefValue::get(Val->getType()));
1807 /// \brief Restore the original list of uses.
1808 void undo() override {
1809 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1810 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1811 Inst->setOperand(It, OriginalValues[It]);
1815 /// \brief Build a truncate instruction.
1816 class TruncBuilder : public TypePromotionAction {
1819 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1821 /// trunc Opnd to Ty.
1822 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1823 IRBuilder<> Builder(Opnd);
1824 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
1825 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
1828 /// \brief Get the built value.
1829 Value *getBuiltValue() { return Val; }
1831 /// \brief Remove the built instruction.
1832 void undo() override {
1833 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
1834 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1835 IVal->eraseFromParent();
1839 /// \brief Build a sign extension instruction.
1840 class SExtBuilder : public TypePromotionAction {
1843 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1845 /// sext Opnd to Ty.
1846 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1847 : TypePromotionAction(InsertPt) {
1848 IRBuilder<> Builder(InsertPt);
1849 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
1850 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
1853 /// \brief Get the built value.
1854 Value *getBuiltValue() { return Val; }
1856 /// \brief Remove the built instruction.
1857 void undo() override {
1858 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
1859 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1860 IVal->eraseFromParent();
1864 /// \brief Build a zero extension instruction.
1865 class ZExtBuilder : public TypePromotionAction {
1868 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
1870 /// zext Opnd to Ty.
1871 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1872 : TypePromotionAction(InsertPt) {
1873 IRBuilder<> Builder(InsertPt);
1874 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
1875 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
1878 /// \brief Get the built value.
1879 Value *getBuiltValue() { return Val; }
1881 /// \brief Remove the built instruction.
1882 void undo() override {
1883 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
1884 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1885 IVal->eraseFromParent();
1889 /// \brief Mutate an instruction to another type.
1890 class TypeMutator : public TypePromotionAction {
1891 /// Record the original type.
1895 /// \brief Mutate the type of \p Inst into \p NewTy.
1896 TypeMutator(Instruction *Inst, Type *NewTy)
1897 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1898 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1900 Inst->mutateType(NewTy);
1903 /// \brief Mutate the instruction back to its original type.
1904 void undo() override {
1905 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1907 Inst->mutateType(OrigTy);
1911 /// \brief Replace the uses of an instruction by another instruction.
1912 class UsesReplacer : public TypePromotionAction {
1913 /// Helper structure to keep track of the replaced uses.
1914 struct InstructionAndIdx {
1915 /// The instruction using the instruction.
1917 /// The index where this instruction is used for Inst.
1919 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1920 : Inst(Inst), Idx(Idx) {}
1923 /// Keep track of the original uses (pair Instruction, Index).
1924 SmallVector<InstructionAndIdx, 4> OriginalUses;
1925 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1928 /// \brief Replace all the use of \p Inst by \p New.
1929 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1930 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1932 // Record the original uses.
1933 for (Use &U : Inst->uses()) {
1934 Instruction *UserI = cast<Instruction>(U.getUser());
1935 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1937 // Now, we can replace the uses.
1938 Inst->replaceAllUsesWith(New);
1941 /// \brief Reassign the original uses of Inst to Inst.
1942 void undo() override {
1943 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1944 for (use_iterator UseIt = OriginalUses.begin(),
1945 EndIt = OriginalUses.end();
1946 UseIt != EndIt; ++UseIt) {
1947 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1952 /// \brief Remove an instruction from the IR.
1953 class InstructionRemover : public TypePromotionAction {
1954 /// Original position of the instruction.
1955 InsertionHandler Inserter;
1956 /// Helper structure to hide all the link to the instruction. In other
1957 /// words, this helps to do as if the instruction was removed.
1958 OperandsHider Hider;
1959 /// Keep track of the uses replaced, if any.
1960 UsesReplacer *Replacer;
1963 /// \brief Remove all reference of \p Inst and optinally replace all its
1965 /// \pre If !Inst->use_empty(), then New != nullptr
1966 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1967 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1970 Replacer = new UsesReplacer(Inst, New);
1971 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1972 Inst->removeFromParent();
1975 ~InstructionRemover() override { delete Replacer; }
1977 /// \brief Really remove the instruction.
1978 void commit() override { delete Inst; }
1980 /// \brief Resurrect the instruction and reassign it to the proper uses if
1981 /// new value was provided when build this action.
1982 void undo() override {
1983 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1984 Inserter.insert(Inst);
1992 /// Restoration point.
1993 /// The restoration point is a pointer to an action instead of an iterator
1994 /// because the iterator may be invalidated but not the pointer.
1995 typedef const TypePromotionAction *ConstRestorationPt;
1996 /// Advocate every changes made in that transaction.
1998 /// Undo all the changes made after the given point.
1999 void rollback(ConstRestorationPt Point);
2000 /// Get the current restoration point.
2001 ConstRestorationPt getRestorationPoint() const;
2003 /// \name API for IR modification with state keeping to support rollback.
2005 /// Same as Instruction::setOperand.
2006 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2007 /// Same as Instruction::eraseFromParent.
2008 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2009 /// Same as Value::replaceAllUsesWith.
2010 void replaceAllUsesWith(Instruction *Inst, Value *New);
2011 /// Same as Value::mutateType.
2012 void mutateType(Instruction *Inst, Type *NewTy);
2013 /// Same as IRBuilder::createTrunc.
2014 Value *createTrunc(Instruction *Opnd, Type *Ty);
2015 /// Same as IRBuilder::createSExt.
2016 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2017 /// Same as IRBuilder::createZExt.
2018 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2019 /// Same as Instruction::moveBefore.
2020 void moveBefore(Instruction *Inst, Instruction *Before);
2024 /// The ordered list of actions made so far.
2025 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
2026 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
2029 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2032 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
2035 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2038 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
2041 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2043 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2046 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2047 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2050 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2052 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2053 Value *Val = Ptr->getBuiltValue();
2054 Actions.push_back(std::move(Ptr));
2058 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2059 Value *Opnd, Type *Ty) {
2060 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2061 Value *Val = Ptr->getBuiltValue();
2062 Actions.push_back(std::move(Ptr));
2066 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2067 Value *Opnd, Type *Ty) {
2068 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2069 Value *Val = Ptr->getBuiltValue();
2070 Actions.push_back(std::move(Ptr));
2074 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2075 Instruction *Before) {
2077 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
2080 TypePromotionTransaction::ConstRestorationPt
2081 TypePromotionTransaction::getRestorationPoint() const {
2082 return !Actions.empty() ? Actions.back().get() : nullptr;
2085 void TypePromotionTransaction::commit() {
2086 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2092 void TypePromotionTransaction::rollback(
2093 TypePromotionTransaction::ConstRestorationPt Point) {
2094 while (!Actions.empty() && Point != Actions.back().get()) {
2095 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2100 /// \brief A helper class for matching addressing modes.
2102 /// This encapsulates the logic for matching the target-legal addressing modes.
2103 class AddressingModeMatcher {
2104 SmallVectorImpl<Instruction*> &AddrModeInsts;
2105 const TargetMachine &TM;
2106 const TargetLowering &TLI;
2107 const DataLayout &DL;
2109 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2110 /// the memory instruction that we're computing this address for.
2113 Instruction *MemoryInst;
2115 /// AddrMode - This is the addressing mode that we're building up. This is
2116 /// part of the return value of this addressing mode matching stuff.
2117 ExtAddrMode &AddrMode;
2119 /// The instructions inserted by other CodeGenPrepare optimizations.
2120 const SetOfInstrs &InsertedInsts;
2121 /// A map from the instructions to their type before promotion.
2122 InstrToOrigTy &PromotedInsts;
2123 /// The ongoing transaction where every action should be registered.
2124 TypePromotionTransaction &TPT;
2126 /// IgnoreProfitability - This is set to true when we should not do
2127 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
2128 /// always returns true.
2129 bool IgnoreProfitability;
2131 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2132 const TargetMachine &TM, Type *AT, unsigned AS,
2133 Instruction *MI, ExtAddrMode &AM,
2134 const SetOfInstrs &InsertedInsts,
2135 InstrToOrigTy &PromotedInsts,
2136 TypePromotionTransaction &TPT)
2137 : AddrModeInsts(AMI), TM(TM),
2138 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
2139 ->getTargetLowering()),
2140 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2141 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2142 PromotedInsts(PromotedInsts), TPT(TPT) {
2143 IgnoreProfitability = false;
2147 /// Match - Find the maximal addressing mode that a load/store of V can fold,
2148 /// give an access type of AccessTy. This returns a list of involved
2149 /// instructions in AddrModeInsts.
2150 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2152 /// \p PromotedInsts maps the instructions to their type before promotion.
2153 /// \p The ongoing transaction where every action should be registered.
2154 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2155 Instruction *MemoryInst,
2156 SmallVectorImpl<Instruction*> &AddrModeInsts,
2157 const TargetMachine &TM,
2158 const SetOfInstrs &InsertedInsts,
2159 InstrToOrigTy &PromotedInsts,
2160 TypePromotionTransaction &TPT) {
2163 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
2164 MemoryInst, Result, InsertedInsts,
2165 PromotedInsts, TPT).MatchAddr(V, 0);
2166 (void)Success; assert(Success && "Couldn't select *anything*?");
2170 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2171 bool MatchAddr(Value *V, unsigned Depth);
2172 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2173 bool *MovedAway = nullptr);
2174 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
2175 ExtAddrMode &AMBefore,
2176 ExtAddrMode &AMAfter);
2177 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2178 bool IsPromotionProfitable(unsigned NewCost, unsigned OldCost,
2179 Value *PromotedOperand) const;
2182 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
2183 /// Return true and update AddrMode if this addr mode is legal for the target,
2185 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
2187 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2188 // mode. Just process that directly.
2190 return MatchAddr(ScaleReg, Depth);
2192 // If the scale is 0, it takes nothing to add this.
2196 // If we already have a scale of this value, we can add to it, otherwise, we
2197 // need an available scale field.
2198 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2201 ExtAddrMode TestAddrMode = AddrMode;
2203 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2204 // [A+B + A*7] -> [B+A*8].
2205 TestAddrMode.Scale += Scale;
2206 TestAddrMode.ScaledReg = ScaleReg;
2208 // If the new address isn't legal, bail out.
2209 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
2212 // It was legal, so commit it.
2213 AddrMode = TestAddrMode;
2215 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2216 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2217 // X*Scale + C*Scale to addr mode.
2218 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2219 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2220 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2221 TestAddrMode.ScaledReg = AddLHS;
2222 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2224 // If this addressing mode is legal, commit it and remember that we folded
2225 // this instruction.
2226 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
2227 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2228 AddrMode = TestAddrMode;
2233 // Otherwise, not (x+c)*scale, just return what we have.
2237 /// MightBeFoldableInst - This is a little filter, which returns true if an
2238 /// addressing computation involving I might be folded into a load/store
2239 /// accessing it. This doesn't need to be perfect, but needs to accept at least
2240 /// the set of instructions that MatchOperationAddr can.
2241 static bool MightBeFoldableInst(Instruction *I) {
2242 switch (I->getOpcode()) {
2243 case Instruction::BitCast:
2244 case Instruction::AddrSpaceCast:
2245 // Don't touch identity bitcasts.
2246 if (I->getType() == I->getOperand(0)->getType())
2248 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2249 case Instruction::PtrToInt:
2250 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2252 case Instruction::IntToPtr:
2253 // We know the input is intptr_t, so this is foldable.
2255 case Instruction::Add:
2257 case Instruction::Mul:
2258 case Instruction::Shl:
2259 // Can only handle X*C and X << C.
2260 return isa<ConstantInt>(I->getOperand(1));
2261 case Instruction::GetElementPtr:
2268 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2269 /// \note \p Val is assumed to be the product of some type promotion.
2270 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2271 /// to be legal, as the non-promoted value would have had the same state.
2272 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
2273 const DataLayout &DL, Value *Val) {
2274 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2277 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2278 // If the ISDOpcode is undefined, it was undefined before the promotion.
2281 // Otherwise, check if the promoted instruction is legal or not.
2282 return TLI.isOperationLegalOrCustom(
2283 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
2286 /// \brief Hepler class to perform type promotion.
2287 class TypePromotionHelper {
2288 /// \brief Utility function to check whether or not a sign or zero extension
2289 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2290 /// either using the operands of \p Inst or promoting \p Inst.
2291 /// The type of the extension is defined by \p IsSExt.
2292 /// In other words, check if:
2293 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2294 /// #1 Promotion applies:
2295 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2296 /// #2 Operand reuses:
2297 /// ext opnd1 to ConsideredExtType.
2298 /// \p PromotedInsts maps the instructions to their type before promotion.
2299 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2300 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2302 /// \brief Utility function to determine if \p OpIdx should be promoted when
2303 /// promoting \p Inst.
2304 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2305 if (isa<SelectInst>(Inst) && OpIdx == 0)
2310 /// \brief Utility function to promote the operand of \p Ext when this
2311 /// operand is a promotable trunc or sext or zext.
2312 /// \p PromotedInsts maps the instructions to their type before promotion.
2313 /// \p CreatedInstsCost[out] contains the cost of all instructions
2314 /// created to promote the operand of Ext.
2315 /// Newly added extensions are inserted in \p Exts.
2316 /// Newly added truncates are inserted in \p Truncs.
2317 /// Should never be called directly.
2318 /// \return The promoted value which is used instead of Ext.
2319 static Value *promoteOperandForTruncAndAnyExt(
2320 Instruction *Ext, TypePromotionTransaction &TPT,
2321 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2322 SmallVectorImpl<Instruction *> *Exts,
2323 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
2325 /// \brief Utility function to promote the operand of \p Ext when this
2326 /// operand is promotable and is not a supported trunc or sext.
2327 /// \p PromotedInsts maps the instructions to their type before promotion.
2328 /// \p CreatedInstsCost[out] contains the cost of all the instructions
2329 /// created to promote the operand of Ext.
2330 /// Newly added extensions are inserted in \p Exts.
2331 /// Newly added truncates are inserted in \p Truncs.
2332 /// Should never be called directly.
2333 /// \return The promoted value which is used instead of Ext.
2334 static Value *promoteOperandForOther(Instruction *Ext,
2335 TypePromotionTransaction &TPT,
2336 InstrToOrigTy &PromotedInsts,
2337 unsigned &CreatedInstsCost,
2338 SmallVectorImpl<Instruction *> *Exts,
2339 SmallVectorImpl<Instruction *> *Truncs,
2340 const TargetLowering &TLI, bool IsSExt);
2342 /// \see promoteOperandForOther.
2343 static Value *signExtendOperandForOther(
2344 Instruction *Ext, TypePromotionTransaction &TPT,
2345 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2346 SmallVectorImpl<Instruction *> *Exts,
2347 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2348 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2349 Exts, Truncs, TLI, true);
2352 /// \see promoteOperandForOther.
2353 static Value *zeroExtendOperandForOther(
2354 Instruction *Ext, TypePromotionTransaction &TPT,
2355 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2356 SmallVectorImpl<Instruction *> *Exts,
2357 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2358 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2359 Exts, Truncs, TLI, false);
2363 /// Type for the utility function that promotes the operand of Ext.
2364 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2365 InstrToOrigTy &PromotedInsts,
2366 unsigned &CreatedInstsCost,
2367 SmallVectorImpl<Instruction *> *Exts,
2368 SmallVectorImpl<Instruction *> *Truncs,
2369 const TargetLowering &TLI);
2370 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2371 /// action to promote the operand of \p Ext instead of using Ext.
2372 /// \return NULL if no promotable action is possible with the current
2374 /// \p InsertedInsts keeps track of all the instructions inserted by the
2375 /// other CodeGenPrepare optimizations. This information is important
2376 /// because we do not want to promote these instructions as CodeGenPrepare
2377 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2378 /// \p PromotedInsts maps the instructions to their type before promotion.
2379 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
2380 const TargetLowering &TLI,
2381 const InstrToOrigTy &PromotedInsts);
2384 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2385 Type *ConsideredExtType,
2386 const InstrToOrigTy &PromotedInsts,
2388 // The promotion helper does not know how to deal with vector types yet.
2389 // To be able to fix that, we would need to fix the places where we
2390 // statically extend, e.g., constants and such.
2391 if (Inst->getType()->isVectorTy())
2394 // We can always get through zext.
2395 if (isa<ZExtInst>(Inst))
2398 // sext(sext) is ok too.
2399 if (IsSExt && isa<SExtInst>(Inst))
2402 // We can get through binary operator, if it is legal. In other words, the
2403 // binary operator must have a nuw or nsw flag.
2404 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2405 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2406 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2407 (IsSExt && BinOp->hasNoSignedWrap())))
2410 // Check if we can do the following simplification.
2411 // ext(trunc(opnd)) --> ext(opnd)
2412 if (!isa<TruncInst>(Inst))
2415 Value *OpndVal = Inst->getOperand(0);
2416 // Check if we can use this operand in the extension.
2417 // If the type is larger than the result type of the extension,
2419 if (!OpndVal->getType()->isIntegerTy() ||
2420 OpndVal->getType()->getIntegerBitWidth() >
2421 ConsideredExtType->getIntegerBitWidth())
2424 // If the operand of the truncate is not an instruction, we will not have
2425 // any information on the dropped bits.
2426 // (Actually we could for constant but it is not worth the extra logic).
2427 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2431 // Check if the source of the type is narrow enough.
2432 // I.e., check that trunc just drops extended bits of the same kind of
2434 // #1 get the type of the operand and check the kind of the extended bits.
2435 const Type *OpndType;
2436 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2437 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
2438 OpndType = It->second.getPointer();
2439 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2440 OpndType = Opnd->getOperand(0)->getType();
2444 // #2 check that the truncate just drop extended bits.
2445 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
2451 TypePromotionHelper::Action TypePromotionHelper::getAction(
2452 Instruction *Ext, const SetOfInstrs &InsertedInsts,
2453 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2454 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2455 "Unexpected instruction type");
2456 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2457 Type *ExtTy = Ext->getType();
2458 bool IsSExt = isa<SExtInst>(Ext);
2459 // If the operand of the extension is not an instruction, we cannot
2461 // If it, check we can get through.
2462 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2465 // Do not promote if the operand has been added by codegenprepare.
2466 // Otherwise, it means we are undoing an optimization that is likely to be
2467 // redone, thus causing potential infinite loop.
2468 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
2471 // SExt or Trunc instructions.
2472 // Return the related handler.
2473 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2474 isa<ZExtInst>(ExtOpnd))
2475 return promoteOperandForTruncAndAnyExt;
2477 // Regular instruction.
2478 // Abort early if we will have to insert non-free instructions.
2479 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2481 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2484 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2485 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2486 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2487 SmallVectorImpl<Instruction *> *Exts,
2488 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2489 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2490 // get through it and this method should not be called.
2491 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2492 Value *ExtVal = SExt;
2493 bool HasMergedNonFreeExt = false;
2494 if (isa<ZExtInst>(SExtOpnd)) {
2495 // Replace s|zext(zext(opnd))
2497 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
2499 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2500 TPT.replaceAllUsesWith(SExt, ZExt);
2501 TPT.eraseInstruction(SExt);
2504 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2506 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2508 CreatedInstsCost = 0;
2510 // Remove dead code.
2511 if (SExtOpnd->use_empty())
2512 TPT.eraseInstruction(SExtOpnd);
2514 // Check if the extension is still needed.
2515 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2516 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2519 Exts->push_back(ExtInst);
2520 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
2525 // At this point we have: ext ty opnd to ty.
2526 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2527 Value *NextVal = ExtInst->getOperand(0);
2528 TPT.eraseInstruction(ExtInst, NextVal);
2532 Value *TypePromotionHelper::promoteOperandForOther(
2533 Instruction *Ext, TypePromotionTransaction &TPT,
2534 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2535 SmallVectorImpl<Instruction *> *Exts,
2536 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
2538 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2539 // get through it and this method should not be called.
2540 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2541 CreatedInstsCost = 0;
2542 if (!ExtOpnd->hasOneUse()) {
2543 // ExtOpnd will be promoted.
2544 // All its uses, but Ext, will need to use a truncated value of the
2545 // promoted version.
2546 // Create the truncate now.
2547 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2548 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2549 ITrunc->removeFromParent();
2550 // Insert it just after the definition.
2551 ITrunc->insertAfter(ExtOpnd);
2553 Truncs->push_back(ITrunc);
2556 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2557 // Restore the operand of Ext (which has been replace by the previous call
2558 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2559 TPT.setOperand(Ext, 0, ExtOpnd);
2562 // Get through the Instruction:
2563 // 1. Update its type.
2564 // 2. Replace the uses of Ext by Inst.
2565 // 3. Extend each operand that needs to be extended.
2567 // Remember the original type of the instruction before promotion.
2568 // This is useful to know that the high bits are sign extended bits.
2569 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2570 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2572 TPT.mutateType(ExtOpnd, Ext->getType());
2574 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2576 Instruction *ExtForOpnd = Ext;
2578 DEBUG(dbgs() << "Propagate Ext to operands\n");
2579 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2581 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2582 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2583 !shouldExtOperand(ExtOpnd, OpIdx)) {
2584 DEBUG(dbgs() << "No need to propagate\n");
2587 // Check if we can statically extend the operand.
2588 Value *Opnd = ExtOpnd->getOperand(OpIdx);
2589 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2590 DEBUG(dbgs() << "Statically extend\n");
2591 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2592 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
2593 : Cst->getValue().zext(BitWidth);
2594 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
2597 // UndefValue are typed, so we have to statically sign extend them.
2598 if (isa<UndefValue>(Opnd)) {
2599 DEBUG(dbgs() << "Statically extend\n");
2600 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2604 // Otherwise we have to explicity sign extend the operand.
2605 // Check if Ext was reused to extend an operand.
2607 // If yes, create a new one.
2608 DEBUG(dbgs() << "More operands to ext\n");
2609 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2610 : TPT.createZExt(Ext, Opnd, Ext->getType());
2611 if (!isa<Instruction>(ValForExtOpnd)) {
2612 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
2615 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
2618 Exts->push_back(ExtForOpnd);
2619 TPT.setOperand(ExtForOpnd, 0, Opnd);
2621 // Move the sign extension before the insertion point.
2622 TPT.moveBefore(ExtForOpnd, ExtOpnd);
2623 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2624 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
2625 // If more sext are required, new instructions will have to be created.
2626 ExtForOpnd = nullptr;
2628 if (ExtForOpnd == Ext) {
2629 DEBUG(dbgs() << "Extension is useless now\n");
2630 TPT.eraseInstruction(Ext);
2635 /// IsPromotionProfitable - Check whether or not promoting an instruction
2636 /// to a wider type was profitable.
2637 /// \p NewCost gives the cost of extension instructions created by the
2639 /// \p OldCost gives the cost of extension instructions before the promotion
2640 /// plus the number of instructions that have been
2641 /// matched in the addressing mode the promotion.
2642 /// \p PromotedOperand is the value that has been promoted.
2643 /// \return True if the promotion is profitable, false otherwise.
2644 bool AddressingModeMatcher::IsPromotionProfitable(
2645 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
2646 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
2647 // The cost of the new extensions is greater than the cost of the
2648 // old extension plus what we folded.
2649 // This is not profitable.
2650 if (NewCost > OldCost)
2652 if (NewCost < OldCost)
2654 // The promotion is neutral but it may help folding the sign extension in
2655 // loads for instance.
2656 // Check that we did not create an illegal instruction.
2657 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
2660 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
2661 /// fold the operation into the addressing mode. If so, update the addressing
2662 /// mode and return true, otherwise return false without modifying AddrMode.
2663 /// If \p MovedAway is not NULL, it contains the information of whether or
2664 /// not AddrInst has to be folded into the addressing mode on success.
2665 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2666 /// because it has been moved away.
2667 /// Thus AddrInst must not be added in the matched instructions.
2668 /// This state can happen when AddrInst is a sext, since it may be moved away.
2669 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2670 /// not be referenced anymore.
2671 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
2674 // Avoid exponential behavior on extremely deep expression trees.
2675 if (Depth >= 5) return false;
2677 // By default, all matched instructions stay in place.
2682 case Instruction::PtrToInt:
2683 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2684 return MatchAddr(AddrInst->getOperand(0), Depth);
2685 case Instruction::IntToPtr: {
2686 auto AS = AddrInst->getType()->getPointerAddressSpace();
2687 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
2688 // This inttoptr is a no-op if the integer type is pointer sized.
2689 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
2690 return MatchAddr(AddrInst->getOperand(0), Depth);
2693 case Instruction::BitCast:
2694 // BitCast is always a noop, and we can handle it as long as it is
2695 // int->int or pointer->pointer (we don't want int<->fp or something).
2696 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2697 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2698 // Don't touch identity bitcasts. These were probably put here by LSR,
2699 // and we don't want to mess around with them. Assume it knows what it
2701 AddrInst->getOperand(0)->getType() != AddrInst->getType())
2702 return MatchAddr(AddrInst->getOperand(0), Depth);
2704 case Instruction::AddrSpaceCast: {
2706 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
2707 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
2708 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
2709 return MatchAddr(AddrInst->getOperand(0), Depth);
2712 case Instruction::Add: {
2713 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2714 ExtAddrMode BackupAddrMode = AddrMode;
2715 unsigned OldSize = AddrModeInsts.size();
2716 // Start a transaction at this point.
2717 // The LHS may match but not the RHS.
2718 // Therefore, we need a higher level restoration point to undo partially
2719 // matched operation.
2720 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2721 TPT.getRestorationPoint();
2723 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2724 MatchAddr(AddrInst->getOperand(0), Depth+1))
2727 // Restore the old addr mode info.
2728 AddrMode = BackupAddrMode;
2729 AddrModeInsts.resize(OldSize);
2730 TPT.rollback(LastKnownGood);
2732 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2733 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2734 MatchAddr(AddrInst->getOperand(1), Depth+1))
2737 // Otherwise we definitely can't merge the ADD in.
2738 AddrMode = BackupAddrMode;
2739 AddrModeInsts.resize(OldSize);
2740 TPT.rollback(LastKnownGood);
2743 //case Instruction::Or:
2744 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2746 case Instruction::Mul:
2747 case Instruction::Shl: {
2748 // Can only handle X*C and X << C.
2749 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2752 int64_t Scale = RHS->getSExtValue();
2753 if (Opcode == Instruction::Shl)
2754 Scale = 1LL << Scale;
2756 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2758 case Instruction::GetElementPtr: {
2759 // Scan the GEP. We check it if it contains constant offsets and at most
2760 // one variable offset.
2761 int VariableOperand = -1;
2762 unsigned VariableScale = 0;
2764 int64_t ConstantOffset = 0;
2765 gep_type_iterator GTI = gep_type_begin(AddrInst);
2766 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2767 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2768 const StructLayout *SL = DL.getStructLayout(STy);
2770 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2771 ConstantOffset += SL->getElementOffset(Idx);
2773 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
2774 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2775 ConstantOffset += CI->getSExtValue()*TypeSize;
2776 } else if (TypeSize) { // Scales of zero don't do anything.
2777 // We only allow one variable index at the moment.
2778 if (VariableOperand != -1)
2781 // Remember the variable index.
2782 VariableOperand = i;
2783 VariableScale = TypeSize;
2788 // A common case is for the GEP to only do a constant offset. In this case,
2789 // just add it to the disp field and check validity.
2790 if (VariableOperand == -1) {
2791 AddrMode.BaseOffs += ConstantOffset;
2792 if (ConstantOffset == 0 ||
2793 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
2794 // Check to see if we can fold the base pointer in too.
2795 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2798 AddrMode.BaseOffs -= ConstantOffset;
2802 // Save the valid addressing mode in case we can't match.
2803 ExtAddrMode BackupAddrMode = AddrMode;
2804 unsigned OldSize = AddrModeInsts.size();
2806 // See if the scale and offset amount is valid for this target.
2807 AddrMode.BaseOffs += ConstantOffset;
2809 // Match the base operand of the GEP.
2810 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2811 // If it couldn't be matched, just stuff the value in a register.
2812 if (AddrMode.HasBaseReg) {
2813 AddrMode = BackupAddrMode;
2814 AddrModeInsts.resize(OldSize);
2817 AddrMode.HasBaseReg = true;
2818 AddrMode.BaseReg = AddrInst->getOperand(0);
2821 // Match the remaining variable portion of the GEP.
2822 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2824 // If it couldn't be matched, try stuffing the base into a register
2825 // instead of matching it, and retrying the match of the scale.
2826 AddrMode = BackupAddrMode;
2827 AddrModeInsts.resize(OldSize);
2828 if (AddrMode.HasBaseReg)
2830 AddrMode.HasBaseReg = true;
2831 AddrMode.BaseReg = AddrInst->getOperand(0);
2832 AddrMode.BaseOffs += ConstantOffset;
2833 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2834 VariableScale, Depth)) {
2835 // If even that didn't work, bail.
2836 AddrMode = BackupAddrMode;
2837 AddrModeInsts.resize(OldSize);
2844 case Instruction::SExt:
2845 case Instruction::ZExt: {
2846 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
2850 // Try to move this ext out of the way of the addressing mode.
2851 // Ask for a method for doing so.
2852 TypePromotionHelper::Action TPH =
2853 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
2857 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2858 TPT.getRestorationPoint();
2859 unsigned CreatedInstsCost = 0;
2860 unsigned ExtCost = !TLI.isExtFree(Ext);
2861 Value *PromotedOperand =
2862 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
2863 // SExt has been moved away.
2864 // Thus either it will be rematched later in the recursive calls or it is
2865 // gone. Anyway, we must not fold it into the addressing mode at this point.
2869 // addr = gep base, idx
2871 // promotedOpnd = ext opnd <- no match here
2872 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2873 // addr = gep base, op <- match
2877 assert(PromotedOperand &&
2878 "TypePromotionHelper should have filtered out those cases");
2880 ExtAddrMode BackupAddrMode = AddrMode;
2881 unsigned OldSize = AddrModeInsts.size();
2883 if (!MatchAddr(PromotedOperand, Depth) ||
2884 // The total of the new cost is equals to the cost of the created
2886 // The total of the old cost is equals to the cost of the extension plus
2887 // what we have saved in the addressing mode.
2888 !IsPromotionProfitable(CreatedInstsCost,
2889 ExtCost + (AddrModeInsts.size() - OldSize),
2891 AddrMode = BackupAddrMode;
2892 AddrModeInsts.resize(OldSize);
2893 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2894 TPT.rollback(LastKnownGood);
2903 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2904 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2905 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2906 /// or intptr_t for the target.
2908 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2909 // Start a transaction at this point that we will rollback if the matching
2911 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2912 TPT.getRestorationPoint();
2913 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2914 // Fold in immediates if legal for the target.
2915 AddrMode.BaseOffs += CI->getSExtValue();
2916 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
2918 AddrMode.BaseOffs -= CI->getSExtValue();
2919 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2920 // If this is a global variable, try to fold it into the addressing mode.
2921 if (!AddrMode.BaseGV) {
2922 AddrMode.BaseGV = GV;
2923 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
2925 AddrMode.BaseGV = nullptr;
2927 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2928 ExtAddrMode BackupAddrMode = AddrMode;
2929 unsigned OldSize = AddrModeInsts.size();
2931 // Check to see if it is possible to fold this operation.
2932 bool MovedAway = false;
2933 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2934 // This instruction may have been move away. If so, there is nothing
2938 // Okay, it's possible to fold this. Check to see if it is actually
2939 // *profitable* to do so. We use a simple cost model to avoid increasing
2940 // register pressure too much.
2941 if (I->hasOneUse() ||
2942 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2943 AddrModeInsts.push_back(I);
2947 // It isn't profitable to do this, roll back.
2948 //cerr << "NOT FOLDING: " << *I;
2949 AddrMode = BackupAddrMode;
2950 AddrModeInsts.resize(OldSize);
2951 TPT.rollback(LastKnownGood);
2953 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2954 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2956 TPT.rollback(LastKnownGood);
2957 } else if (isa<ConstantPointerNull>(Addr)) {
2958 // Null pointer gets folded without affecting the addressing mode.
2962 // Worse case, the target should support [reg] addressing modes. :)
2963 if (!AddrMode.HasBaseReg) {
2964 AddrMode.HasBaseReg = true;
2965 AddrMode.BaseReg = Addr;
2966 // Still check for legality in case the target supports [imm] but not [i+r].
2967 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
2969 AddrMode.HasBaseReg = false;
2970 AddrMode.BaseReg = nullptr;
2973 // If the base register is already taken, see if we can do [r+r].
2974 if (AddrMode.Scale == 0) {
2976 AddrMode.ScaledReg = Addr;
2977 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
2980 AddrMode.ScaledReg = nullptr;
2983 TPT.rollback(LastKnownGood);
2987 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2988 /// inline asm call are due to memory operands. If so, return true, otherwise
2990 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2991 const TargetMachine &TM) {
2992 const Function *F = CI->getParent()->getParent();
2993 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
2994 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
2995 TargetLowering::AsmOperandInfoVector TargetConstraints =
2996 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
2997 ImmutableCallSite(CI));
2998 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2999 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3001 // Compute the constraint code and ConstraintType to use.
3002 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3004 // If this asm operand is our Value*, and if it isn't an indirect memory
3005 // operand, we can't fold it!
3006 if (OpInfo.CallOperandVal == OpVal &&
3007 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
3008 !OpInfo.isIndirect))
3015 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
3016 /// memory use. If we find an obviously non-foldable instruction, return true.
3017 /// Add the ultimately found memory instructions to MemoryUses.
3018 static bool FindAllMemoryUses(
3020 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
3021 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
3022 // If we already considered this instruction, we're done.
3023 if (!ConsideredInsts.insert(I).second)
3026 // If this is an obviously unfoldable instruction, bail out.
3027 if (!MightBeFoldableInst(I))
3030 // Loop over all the uses, recursively processing them.
3031 for (Use &U : I->uses()) {
3032 Instruction *UserI = cast<Instruction>(U.getUser());
3034 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
3035 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
3039 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
3040 unsigned opNo = U.getOperandNo();
3041 if (opNo == 0) return true; // Storing addr, not into addr.
3042 MemoryUses.push_back(std::make_pair(SI, opNo));
3046 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
3047 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
3048 if (!IA) return true;
3050 // If this is a memory operand, we're cool, otherwise bail out.
3051 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
3056 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
3063 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
3064 /// the use site that we're folding it into. If so, there is no cost to
3065 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
3066 /// that we know are live at the instruction already.
3067 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
3068 Value *KnownLive2) {
3069 // If Val is either of the known-live values, we know it is live!
3070 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
3073 // All values other than instructions and arguments (e.g. constants) are live.
3074 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
3076 // If Val is a constant sized alloca in the entry block, it is live, this is
3077 // true because it is just a reference to the stack/frame pointer, which is
3078 // live for the whole function.
3079 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
3080 if (AI->isStaticAlloca())
3083 // Check to see if this value is already used in the memory instruction's
3084 // block. If so, it's already live into the block at the very least, so we
3085 // can reasonably fold it.
3086 return Val->isUsedInBasicBlock(MemoryInst->getParent());
3089 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
3090 /// mode of the machine to fold the specified instruction into a load or store
3091 /// that ultimately uses it. However, the specified instruction has multiple
3092 /// uses. Given this, it may actually increase register pressure to fold it
3093 /// into the load. For example, consider this code:
3097 /// use(Y) -> nonload/store
3101 /// In this case, Y has multiple uses, and can be folded into the load of Z
3102 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
3103 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
3104 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
3105 /// number of computations either.
3107 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
3108 /// X was live across 'load Z' for other reasons, we actually *would* want to
3109 /// fold the addressing mode in the Z case. This would make Y die earlier.
3110 bool AddressingModeMatcher::
3111 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
3112 ExtAddrMode &AMAfter) {
3113 if (IgnoreProfitability) return true;
3115 // AMBefore is the addressing mode before this instruction was folded into it,
3116 // and AMAfter is the addressing mode after the instruction was folded. Get
3117 // the set of registers referenced by AMAfter and subtract out those
3118 // referenced by AMBefore: this is the set of values which folding in this
3119 // address extends the lifetime of.
3121 // Note that there are only two potential values being referenced here,
3122 // BaseReg and ScaleReg (global addresses are always available, as are any
3123 // folded immediates).
3124 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
3126 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
3127 // lifetime wasn't extended by adding this instruction.
3128 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3130 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3131 ScaledReg = nullptr;
3133 // If folding this instruction (and it's subexprs) didn't extend any live
3134 // ranges, we're ok with it.
3135 if (!BaseReg && !ScaledReg)
3138 // If all uses of this instruction are ultimately load/store/inlineasm's,
3139 // check to see if their addressing modes will include this instruction. If
3140 // so, we can fold it into all uses, so it doesn't matter if it has multiple
3142 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
3143 SmallPtrSet<Instruction*, 16> ConsideredInsts;
3144 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
3145 return false; // Has a non-memory, non-foldable use!
3147 // Now that we know that all uses of this instruction are part of a chain of
3148 // computation involving only operations that could theoretically be folded
3149 // into a memory use, loop over each of these uses and see if they could
3150 // *actually* fold the instruction.
3151 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
3152 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
3153 Instruction *User = MemoryUses[i].first;
3154 unsigned OpNo = MemoryUses[i].second;
3156 // Get the access type of this use. If the use isn't a pointer, we don't
3157 // know what it accesses.
3158 Value *Address = User->getOperand(OpNo);
3159 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
3162 Type *AddressAccessTy = AddrTy->getElementType();
3163 unsigned AS = AddrTy->getAddressSpace();
3165 // Do a match against the root of this address, ignoring profitability. This
3166 // will tell us if the addressing mode for the memory operation will
3167 // *actually* cover the shared instruction.
3169 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3170 TPT.getRestorationPoint();
3171 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
3172 MemoryInst, Result, InsertedInsts,
3173 PromotedInsts, TPT);
3174 Matcher.IgnoreProfitability = true;
3175 bool Success = Matcher.MatchAddr(Address, 0);
3176 (void)Success; assert(Success && "Couldn't select *anything*?");
3178 // The match was to check the profitability, the changes made are not
3179 // part of the original matcher. Therefore, they should be dropped
3180 // otherwise the original matcher will not present the right state.
3181 TPT.rollback(LastKnownGood);
3183 // If the match didn't cover I, then it won't be shared by it.
3184 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3185 I) == MatchedAddrModeInsts.end())
3188 MatchedAddrModeInsts.clear();
3194 } // end anonymous namespace
3196 /// IsNonLocalValue - Return true if the specified values are defined in a
3197 /// different basic block than BB.
3198 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3199 if (Instruction *I = dyn_cast<Instruction>(V))
3200 return I->getParent() != BB;
3204 /// OptimizeMemoryInst - Load and Store Instructions often have
3205 /// addressing modes that can do significant amounts of computation. As such,
3206 /// instruction selection will try to get the load or store to do as much
3207 /// computation as possible for the program. The problem is that isel can only
3208 /// see within a single block. As such, we sink as much legal addressing mode
3209 /// stuff into the block as possible.
3211 /// This method is used to optimize both load/store and inline asms with memory
3213 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3214 Type *AccessTy, unsigned AddrSpace) {
3217 // Try to collapse single-value PHI nodes. This is necessary to undo
3218 // unprofitable PRE transformations.
3219 SmallVector<Value*, 8> worklist;
3220 SmallPtrSet<Value*, 16> Visited;
3221 worklist.push_back(Addr);
3223 // Use a worklist to iteratively look through PHI nodes, and ensure that
3224 // the addressing mode obtained from the non-PHI roots of the graph
3226 Value *Consensus = nullptr;
3227 unsigned NumUsesConsensus = 0;
3228 bool IsNumUsesConsensusValid = false;
3229 SmallVector<Instruction*, 16> AddrModeInsts;
3230 ExtAddrMode AddrMode;
3231 TypePromotionTransaction TPT;
3232 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3233 TPT.getRestorationPoint();
3234 while (!worklist.empty()) {
3235 Value *V = worklist.back();
3236 worklist.pop_back();
3238 // Break use-def graph loops.
3239 if (!Visited.insert(V).second) {
3240 Consensus = nullptr;
3244 // For a PHI node, push all of its incoming values.
3245 if (PHINode *P = dyn_cast<PHINode>(V)) {
3246 for (Value *IncValue : P->incoming_values())
3247 worklist.push_back(IncValue);
3251 // For non-PHIs, determine the addressing mode being computed.
3252 SmallVector<Instruction*, 16> NewAddrModeInsts;
3253 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3254 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
3255 InsertedInsts, PromotedInsts, TPT);
3257 // This check is broken into two cases with very similar code to avoid using
3258 // getNumUses() as much as possible. Some values have a lot of uses, so
3259 // calling getNumUses() unconditionally caused a significant compile-time
3263 AddrMode = NewAddrMode;
3264 AddrModeInsts = NewAddrModeInsts;
3266 } else if (NewAddrMode == AddrMode) {
3267 if (!IsNumUsesConsensusValid) {
3268 NumUsesConsensus = Consensus->getNumUses();
3269 IsNumUsesConsensusValid = true;
3272 // Ensure that the obtained addressing mode is equivalent to that obtained
3273 // for all other roots of the PHI traversal. Also, when choosing one
3274 // such root as representative, select the one with the most uses in order
3275 // to keep the cost modeling heuristics in AddressingModeMatcher
3277 unsigned NumUses = V->getNumUses();
3278 if (NumUses > NumUsesConsensus) {
3280 NumUsesConsensus = NumUses;
3281 AddrModeInsts = NewAddrModeInsts;
3286 Consensus = nullptr;
3290 // If the addressing mode couldn't be determined, or if multiple different
3291 // ones were determined, bail out now.
3293 TPT.rollback(LastKnownGood);
3298 // Check to see if any of the instructions supersumed by this addr mode are
3299 // non-local to I's BB.
3300 bool AnyNonLocal = false;
3301 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3302 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3308 // If all the instructions matched are already in this BB, don't do anything.
3310 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3314 // Insert this computation right after this user. Since our caller is
3315 // scanning from the top of the BB to the bottom, reuse of the expr are
3316 // guaranteed to happen later.
3317 IRBuilder<> Builder(MemoryInst);
3319 // Now that we determined the addressing expression we want to use and know
3320 // that we have to sink it into this block. Check to see if we have already
3321 // done this for some other load/store instr in this block. If so, reuse the
3323 Value *&SunkAddr = SunkAddrs[Addr];
3325 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3326 << *MemoryInst << "\n");
3327 if (SunkAddr->getType() != Addr->getType())
3328 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3329 } else if (AddrSinkUsingGEPs ||
3330 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3331 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3333 // By default, we use the GEP-based method when AA is used later. This
3334 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3335 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3336 << *MemoryInst << "\n");
3337 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3338 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3340 // First, find the pointer.
3341 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3342 ResultPtr = AddrMode.BaseReg;
3343 AddrMode.BaseReg = nullptr;
3346 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3347 // We can't add more than one pointer together, nor can we scale a
3348 // pointer (both of which seem meaningless).
3349 if (ResultPtr || AddrMode.Scale != 1)
3352 ResultPtr = AddrMode.ScaledReg;
3356 if (AddrMode.BaseGV) {
3360 ResultPtr = AddrMode.BaseGV;
3363 // If the real base value actually came from an inttoptr, then the matcher
3364 // will look through it and provide only the integer value. In that case,
3366 if (!ResultPtr && AddrMode.BaseReg) {
3368 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3369 AddrMode.BaseReg = nullptr;
3370 } else if (!ResultPtr && AddrMode.Scale == 1) {
3372 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3377 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3378 SunkAddr = Constant::getNullValue(Addr->getType());
3379 } else if (!ResultPtr) {
3383 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3384 Type *I8Ty = Builder.getInt8Ty();
3386 // Start with the base register. Do this first so that subsequent address
3387 // matching finds it last, which will prevent it from trying to match it
3388 // as the scaled value in case it happens to be a mul. That would be
3389 // problematic if we've sunk a different mul for the scale, because then
3390 // we'd end up sinking both muls.
3391 if (AddrMode.BaseReg) {
3392 Value *V = AddrMode.BaseReg;
3393 if (V->getType() != IntPtrTy)
3394 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3399 // Add the scale value.
3400 if (AddrMode.Scale) {
3401 Value *V = AddrMode.ScaledReg;
3402 if (V->getType() == IntPtrTy) {
3404 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3405 cast<IntegerType>(V->getType())->getBitWidth()) {
3406 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3408 // It is only safe to sign extend the BaseReg if we know that the math
3409 // required to create it did not overflow before we extend it. Since
3410 // the original IR value was tossed in favor of a constant back when
3411 // the AddrMode was created we need to bail out gracefully if widths
3412 // do not match instead of extending it.
3413 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3414 if (I && (ResultIndex != AddrMode.BaseReg))
3415 I->eraseFromParent();
3419 if (AddrMode.Scale != 1)
3420 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3423 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3428 // Add in the Base Offset if present.
3429 if (AddrMode.BaseOffs) {
3430 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3432 // We need to add this separately from the scale above to help with
3433 // SDAG consecutive load/store merging.
3434 if (ResultPtr->getType() != I8PtrTy)
3435 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3436 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3443 SunkAddr = ResultPtr;
3445 if (ResultPtr->getType() != I8PtrTy)
3446 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3447 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3450 if (SunkAddr->getType() != Addr->getType())
3451 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3454 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3455 << *MemoryInst << "\n");
3456 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3457 Value *Result = nullptr;
3459 // Start with the base register. Do this first so that subsequent address
3460 // matching finds it last, which will prevent it from trying to match it
3461 // as the scaled value in case it happens to be a mul. That would be
3462 // problematic if we've sunk a different mul for the scale, because then
3463 // we'd end up sinking both muls.
3464 if (AddrMode.BaseReg) {
3465 Value *V = AddrMode.BaseReg;
3466 if (V->getType()->isPointerTy())
3467 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3468 if (V->getType() != IntPtrTy)
3469 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3473 // Add the scale value.
3474 if (AddrMode.Scale) {
3475 Value *V = AddrMode.ScaledReg;
3476 if (V->getType() == IntPtrTy) {
3478 } else if (V->getType()->isPointerTy()) {
3479 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3480 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3481 cast<IntegerType>(V->getType())->getBitWidth()) {
3482 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3484 // It is only safe to sign extend the BaseReg if we know that the math
3485 // required to create it did not overflow before we extend it. Since
3486 // the original IR value was tossed in favor of a constant back when
3487 // the AddrMode was created we need to bail out gracefully if widths
3488 // do not match instead of extending it.
3489 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3490 if (I && (Result != AddrMode.BaseReg))
3491 I->eraseFromParent();
3494 if (AddrMode.Scale != 1)
3495 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3498 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3503 // Add in the BaseGV if present.
3504 if (AddrMode.BaseGV) {
3505 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3507 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3512 // Add in the Base Offset if present.
3513 if (AddrMode.BaseOffs) {
3514 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3516 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3522 SunkAddr = Constant::getNullValue(Addr->getType());
3524 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3527 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3529 // If we have no uses, recursively delete the value and all dead instructions
3531 if (Repl->use_empty()) {
3532 // This can cause recursive deletion, which can invalidate our iterator.
3533 // Use a WeakVH to hold onto it in case this happens.
3534 WeakVH IterHandle(CurInstIterator);
3535 BasicBlock *BB = CurInstIterator->getParent();
3537 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3539 if (IterHandle != CurInstIterator) {
3540 // If the iterator instruction was recursively deleted, start over at the
3541 // start of the block.
3542 CurInstIterator = BB->begin();
3550 /// OptimizeInlineAsmInst - If there are any memory operands, use
3551 /// OptimizeMemoryInst to sink their address computing into the block when
3552 /// possible / profitable.
3553 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
3554 bool MadeChange = false;
3556 const TargetRegisterInfo *TRI =
3557 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3558 TargetLowering::AsmOperandInfoVector TargetConstraints =
3559 TLI->ParseConstraints(*DL, TRI, CS);
3561 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3562 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3564 // Compute the constraint code and ConstraintType to use.
3565 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3567 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3568 OpInfo.isIndirect) {
3569 Value *OpVal = CS->getArgOperand(ArgNo++);
3570 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
3571 } else if (OpInfo.Type == InlineAsm::isInput)
3578 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3579 /// sign extensions.
3580 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3581 assert(!Inst->use_empty() && "Input must have at least one use");
3582 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3583 bool IsSExt = isa<SExtInst>(FirstUser);
3584 Type *ExtTy = FirstUser->getType();
3585 for (const User *U : Inst->users()) {
3586 const Instruction *UI = cast<Instruction>(U);
3587 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3589 Type *CurTy = UI->getType();
3590 // Same input and output types: Same instruction after CSE.
3594 // If IsSExt is true, we are in this situation:
3596 // b = sext ty1 a to ty2
3597 // c = sext ty1 a to ty3
3598 // Assuming ty2 is shorter than ty3, this could be turned into:
3600 // b = sext ty1 a to ty2
3601 // c = sext ty2 b to ty3
3602 // However, the last sext is not free.
3606 // This is a ZExt, maybe this is free to extend from one type to another.
3607 // In that case, we would not account for a different use.
3610 if (ExtTy->getScalarType()->getIntegerBitWidth() >
3611 CurTy->getScalarType()->getIntegerBitWidth()) {
3619 if (!TLI.isZExtFree(NarrowTy, LargeTy))
3622 // All uses are the same or can be derived from one another for free.
3626 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
3627 /// load instruction.
3628 /// If an ext(load) can be formed, it is returned via \p LI for the load
3629 /// and \p Inst for the extension.
3630 /// Otherwise LI == nullptr and Inst == nullptr.
3631 /// When some promotion happened, \p TPT contains the proper state to
3634 /// \return true when promoting was necessary to expose the ext(load)
3635 /// opportunity, false otherwise.
3639 /// %ld = load i32* %addr
3640 /// %add = add nuw i32 %ld, 4
3641 /// %zext = zext i32 %add to i64
3645 /// %ld = load i32* %addr
3646 /// %zext = zext i32 %ld to i64
3647 /// %add = add nuw i64 %zext, 4
3649 /// Thanks to the promotion, we can match zext(load i32*) to i64.
3650 bool CodeGenPrepare::ExtLdPromotion(TypePromotionTransaction &TPT,
3651 LoadInst *&LI, Instruction *&Inst,
3652 const SmallVectorImpl<Instruction *> &Exts,
3653 unsigned CreatedInstsCost = 0) {
3654 // Iterate over all the extensions to see if one form an ext(load).
3655 for (auto I : Exts) {
3656 // Check if we directly have ext(load).
3657 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
3659 // No promotion happened here.
3662 // Check whether or not we want to do any promotion.
3663 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
3665 // Get the action to perform the promotion.
3666 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
3667 I, InsertedInsts, *TLI, PromotedInsts);
3668 // Check if we can promote.
3671 // Save the current state.
3672 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3673 TPT.getRestorationPoint();
3674 SmallVector<Instruction *, 4> NewExts;
3675 unsigned NewCreatedInstsCost = 0;
3676 unsigned ExtCost = !TLI->isExtFree(I);
3678 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
3679 &NewExts, nullptr, *TLI);
3680 assert(PromotedVal &&
3681 "TypePromotionHelper should have filtered out those cases");
3683 // We would be able to merge only one extension in a load.
3684 // Therefore, if we have more than 1 new extension we heuristically
3685 // cut this search path, because it means we degrade the code quality.
3686 // With exactly 2, the transformation is neutral, because we will merge
3687 // one extension but leave one. However, we optimistically keep going,
3688 // because the new extension may be removed too.
3689 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
3690 TotalCreatedInstsCost -= ExtCost;
3691 if (!StressExtLdPromotion &&
3692 (TotalCreatedInstsCost > 1 ||
3693 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
3694 // The promotion is not profitable, rollback to the previous state.
3695 TPT.rollback(LastKnownGood);
3698 // The promotion is profitable.
3699 // Check if it exposes an ext(load).
3700 (void)ExtLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
3701 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
3702 // If we have created a new extension, i.e., now we have two
3703 // extensions. We must make sure one of them is merged with
3704 // the load, otherwise we may degrade the code quality.
3705 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
3706 // Promotion happened.
3708 // If this does not help to expose an ext(load) then, rollback.
3709 TPT.rollback(LastKnownGood);
3711 // None of the extension can form an ext(load).
3717 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
3718 /// basic block as the load, unless conditions are unfavorable. This allows
3719 /// SelectionDAG to fold the extend into the load.
3720 /// \p I[in/out] the extension may be modified during the process if some
3721 /// promotions apply.
3723 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *&I) {
3724 // Try to promote a chain of computation if it allows to form
3725 // an extended load.
3726 TypePromotionTransaction TPT;
3727 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3728 TPT.getRestorationPoint();
3729 SmallVector<Instruction *, 1> Exts;
3731 // Look for a load being extended.
3732 LoadInst *LI = nullptr;
3733 Instruction *OldExt = I;
3734 bool HasPromoted = ExtLdPromotion(TPT, LI, I, Exts);
3736 assert(!HasPromoted && !LI && "If we did not match any load instruction "
3737 "the code must remain the same");
3742 // If they're already in the same block, there's nothing to do.
3743 // Make the cheap checks first if we did not promote.
3744 // If we promoted, we need to check if it is indeed profitable.
3745 if (!HasPromoted && LI->getParent() == I->getParent())
3748 EVT VT = TLI->getValueType(*DL, I->getType());
3749 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
3751 // If the load has other users and the truncate is not free, this probably
3752 // isn't worthwhile.
3753 if (!LI->hasOneUse() && TLI &&
3754 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
3755 !TLI->isTruncateFree(I->getType(), LI->getType())) {
3757 TPT.rollback(LastKnownGood);
3761 // Check whether the target supports casts folded into loads.
3763 if (isa<ZExtInst>(I))
3764 LType = ISD::ZEXTLOAD;
3766 assert(isa<SExtInst>(I) && "Unexpected ext type!");
3767 LType = ISD::SEXTLOAD;
3769 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
3771 TPT.rollback(LastKnownGood);
3775 // Move the extend into the same block as the load, so that SelectionDAG
3778 I->removeFromParent();
3784 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
3785 BasicBlock *DefBB = I->getParent();
3787 // If the result of a {s|z}ext and its source are both live out, rewrite all
3788 // other uses of the source with result of extension.
3789 Value *Src = I->getOperand(0);
3790 if (Src->hasOneUse())
3793 // Only do this xform if truncating is free.
3794 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
3797 // Only safe to perform the optimization if the source is also defined in
3799 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
3802 bool DefIsLiveOut = false;
3803 for (User *U : I->users()) {
3804 Instruction *UI = cast<Instruction>(U);
3806 // Figure out which BB this ext is used in.
3807 BasicBlock *UserBB = UI->getParent();
3808 if (UserBB == DefBB) continue;
3809 DefIsLiveOut = true;
3815 // Make sure none of the uses are PHI nodes.
3816 for (User *U : Src->users()) {
3817 Instruction *UI = cast<Instruction>(U);
3818 BasicBlock *UserBB = UI->getParent();
3819 if (UserBB == DefBB) continue;
3820 // Be conservative. We don't want this xform to end up introducing
3821 // reloads just before load / store instructions.
3822 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
3826 // InsertedTruncs - Only insert one trunc in each block once.
3827 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
3829 bool MadeChange = false;
3830 for (Use &U : Src->uses()) {
3831 Instruction *User = cast<Instruction>(U.getUser());
3833 // Figure out which BB this ext is used in.
3834 BasicBlock *UserBB = User->getParent();
3835 if (UserBB == DefBB) continue;
3837 // Both src and def are live in this block. Rewrite the use.
3838 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
3840 if (!InsertedTrunc) {
3841 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3842 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
3843 InsertedInsts.insert(InsertedTrunc);
3846 // Replace a use of the {s|z}ext source with a use of the result.
3855 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
3856 /// turned into an explicit branch.
3857 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
3858 // FIXME: This should use the same heuristics as IfConversion to determine
3859 // whether a select is better represented as a branch. This requires that
3860 // branch probability metadata is preserved for the select, which is not the
3863 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3865 // If the branch is predicted right, an out of order CPU can avoid blocking on
3866 // the compare. Emit cmovs on compares with a memory operand as branches to
3867 // avoid stalls on the load from memory. If the compare has more than one use
3868 // there's probably another cmov or setcc around so it's not worth emitting a
3873 Value *CmpOp0 = Cmp->getOperand(0);
3874 Value *CmpOp1 = Cmp->getOperand(1);
3876 // We check that the memory operand has one use to avoid uses of the loaded
3877 // value directly after the compare, making branches unprofitable.
3878 return Cmp->hasOneUse() &&
3879 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3880 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
3884 /// If we have a SelectInst that will likely profit from branch prediction,
3885 /// turn it into a branch.
3886 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
3887 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3889 // Can we convert the 'select' to CF ?
3890 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3893 TargetLowering::SelectSupportKind SelectKind;
3895 SelectKind = TargetLowering::VectorMaskSelect;
3896 else if (SI->getType()->isVectorTy())
3897 SelectKind = TargetLowering::ScalarCondVectorVal;
3899 SelectKind = TargetLowering::ScalarValSelect;
3901 // Do we have efficient codegen support for this kind of 'selects' ?
3902 if (TLI->isSelectSupported(SelectKind)) {
3903 // We have efficient codegen support for the select instruction.
3904 // Check if it is profitable to keep this 'select'.
3905 if (!TLI->isPredictableSelectExpensive() ||
3906 !isFormingBranchFromSelectProfitable(SI))
3912 // First, we split the block containing the select into 2 blocks.
3913 BasicBlock *StartBlock = SI->getParent();
3914 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3915 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3917 // Create a new block serving as the landing pad for the branch.
3918 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3919 NextBlock->getParent(), NextBlock);
3921 // Move the unconditional branch from the block with the select in it into our
3922 // landing pad block.
3923 StartBlock->getTerminator()->eraseFromParent();
3924 BranchInst::Create(NextBlock, SmallBlock);
3926 // Insert the real conditional branch based on the original condition.
3927 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3929 // The select itself is replaced with a PHI Node.
3930 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3932 PN->addIncoming(SI->getTrueValue(), StartBlock);
3933 PN->addIncoming(SI->getFalseValue(), SmallBlock);
3934 SI->replaceAllUsesWith(PN);
3935 SI->eraseFromParent();
3937 // Instruct OptimizeBlock to skip to the next block.
3938 CurInstIterator = StartBlock->end();
3939 ++NumSelectsExpanded;
3943 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3944 SmallVector<int, 16> Mask(SVI->getShuffleMask());
3946 for (unsigned i = 0; i < Mask.size(); ++i) {
3947 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3949 SplatElem = Mask[i];
3955 /// Some targets have expensive vector shifts if the lanes aren't all the same
3956 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3957 /// it's often worth sinking a shufflevector splat down to its use so that
3958 /// codegen can spot all lanes are identical.
3959 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3960 BasicBlock *DefBB = SVI->getParent();
3962 // Only do this xform if variable vector shifts are particularly expensive.
3963 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3966 // We only expect better codegen by sinking a shuffle if we can recognise a
3968 if (!isBroadcastShuffle(SVI))
3971 // InsertedShuffles - Only insert a shuffle in each block once.
3972 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3974 bool MadeChange = false;
3975 for (User *U : SVI->users()) {
3976 Instruction *UI = cast<Instruction>(U);
3978 // Figure out which BB this ext is used in.
3979 BasicBlock *UserBB = UI->getParent();
3980 if (UserBB == DefBB) continue;
3982 // For now only apply this when the splat is used by a shift instruction.
3983 if (!UI->isShift()) continue;
3985 // Everything checks out, sink the shuffle if the user's block doesn't
3986 // already have a copy.
3987 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3989 if (!InsertedShuffle) {
3990 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3991 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3993 SVI->getOperand(2), "", InsertPt);
3996 UI->replaceUsesOfWith(SVI, InsertedShuffle);
4000 // If we removed all uses, nuke the shuffle.
4001 if (SVI->use_empty()) {
4002 SVI->eraseFromParent();
4010 /// \brief Helper class to promote a scalar operation to a vector one.
4011 /// This class is used to move downward extractelement transition.
4013 /// a = vector_op <2 x i32>
4014 /// b = extractelement <2 x i32> a, i32 0
4019 /// a = vector_op <2 x i32>
4020 /// c = vector_op a (equivalent to scalar_op on the related lane)
4021 /// * d = extractelement <2 x i32> c, i32 0
4023 /// Assuming both extractelement and store can be combine, we get rid of the
4025 class VectorPromoteHelper {
4026 /// DataLayout associated with the current module.
4027 const DataLayout &DL;
4029 /// Used to perform some checks on the legality of vector operations.
4030 const TargetLowering &TLI;
4032 /// Used to estimated the cost of the promoted chain.
4033 const TargetTransformInfo &TTI;
4035 /// The transition being moved downwards.
4036 Instruction *Transition;
4037 /// The sequence of instructions to be promoted.
4038 SmallVector<Instruction *, 4> InstsToBePromoted;
4039 /// Cost of combining a store and an extract.
4040 unsigned StoreExtractCombineCost;
4041 /// Instruction that will be combined with the transition.
4042 Instruction *CombineInst;
4044 /// \brief The instruction that represents the current end of the transition.
4045 /// Since we are faking the promotion until we reach the end of the chain
4046 /// of computation, we need a way to get the current end of the transition.
4047 Instruction *getEndOfTransition() const {
4048 if (InstsToBePromoted.empty())
4050 return InstsToBePromoted.back();
4053 /// \brief Return the index of the original value in the transition.
4054 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
4055 /// c, is at index 0.
4056 unsigned getTransitionOriginalValueIdx() const {
4057 assert(isa<ExtractElementInst>(Transition) &&
4058 "Other kind of transitions are not supported yet");
4062 /// \brief Return the index of the index in the transition.
4063 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
4065 unsigned getTransitionIdx() const {
4066 assert(isa<ExtractElementInst>(Transition) &&
4067 "Other kind of transitions are not supported yet");
4071 /// \brief Get the type of the transition.
4072 /// This is the type of the original value.
4073 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
4074 /// transition is <2 x i32>.
4075 Type *getTransitionType() const {
4076 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
4079 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
4080 /// I.e., we have the following sequence:
4081 /// Def = Transition <ty1> a to <ty2>
4082 /// b = ToBePromoted <ty2> Def, ...
4084 /// b = ToBePromoted <ty1> a, ...
4085 /// Def = Transition <ty1> ToBePromoted to <ty2>
4086 void promoteImpl(Instruction *ToBePromoted);
4088 /// \brief Check whether or not it is profitable to promote all the
4089 /// instructions enqueued to be promoted.
4090 bool isProfitableToPromote() {
4091 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
4092 unsigned Index = isa<ConstantInt>(ValIdx)
4093 ? cast<ConstantInt>(ValIdx)->getZExtValue()
4095 Type *PromotedType = getTransitionType();
4097 StoreInst *ST = cast<StoreInst>(CombineInst);
4098 unsigned AS = ST->getPointerAddressSpace();
4099 unsigned Align = ST->getAlignment();
4100 // Check if this store is supported.
4101 if (!TLI.allowsMisalignedMemoryAccesses(
4102 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
4104 // If this is not supported, there is no way we can combine
4105 // the extract with the store.
4109 // The scalar chain of computation has to pay for the transition
4110 // scalar to vector.
4111 // The vector chain has to account for the combining cost.
4112 uint64_t ScalarCost =
4113 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
4114 uint64_t VectorCost = StoreExtractCombineCost;
4115 for (const auto &Inst : InstsToBePromoted) {
4116 // Compute the cost.
4117 // By construction, all instructions being promoted are arithmetic ones.
4118 // Moreover, one argument is a constant that can be viewed as a splat
4120 Value *Arg0 = Inst->getOperand(0);
4121 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
4122 isa<ConstantFP>(Arg0);
4123 TargetTransformInfo::OperandValueKind Arg0OVK =
4124 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4125 : TargetTransformInfo::OK_AnyValue;
4126 TargetTransformInfo::OperandValueKind Arg1OVK =
4127 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4128 : TargetTransformInfo::OK_AnyValue;
4129 ScalarCost += TTI.getArithmeticInstrCost(
4130 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
4131 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
4134 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
4135 << ScalarCost << "\nVector: " << VectorCost << '\n');
4136 return ScalarCost > VectorCost;
4139 /// \brief Generate a constant vector with \p Val with the same
4140 /// number of elements as the transition.
4141 /// \p UseSplat defines whether or not \p Val should be replicated
4142 /// across the whole vector.
4143 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
4144 /// otherwise we generate a vector with as many undef as possible:
4145 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
4146 /// used at the index of the extract.
4147 Value *getConstantVector(Constant *Val, bool UseSplat) const {
4148 unsigned ExtractIdx = UINT_MAX;
4150 // If we cannot determine where the constant must be, we have to
4151 // use a splat constant.
4152 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
4153 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
4154 ExtractIdx = CstVal->getSExtValue();
4159 unsigned End = getTransitionType()->getVectorNumElements();
4161 return ConstantVector::getSplat(End, Val);
4163 SmallVector<Constant *, 4> ConstVec;
4164 UndefValue *UndefVal = UndefValue::get(Val->getType());
4165 for (unsigned Idx = 0; Idx != End; ++Idx) {
4166 if (Idx == ExtractIdx)
4167 ConstVec.push_back(Val);
4169 ConstVec.push_back(UndefVal);
4171 return ConstantVector::get(ConstVec);
4174 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
4175 /// in \p Use can trigger undefined behavior.
4176 static bool canCauseUndefinedBehavior(const Instruction *Use,
4177 unsigned OperandIdx) {
4178 // This is not safe to introduce undef when the operand is on
4179 // the right hand side of a division-like instruction.
4180 if (OperandIdx != 1)
4182 switch (Use->getOpcode()) {
4185 case Instruction::SDiv:
4186 case Instruction::UDiv:
4187 case Instruction::SRem:
4188 case Instruction::URem:
4190 case Instruction::FDiv:
4191 case Instruction::FRem:
4192 return !Use->hasNoNaNs();
4194 llvm_unreachable(nullptr);
4198 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
4199 const TargetTransformInfo &TTI, Instruction *Transition,
4200 unsigned CombineCost)
4201 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
4202 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4203 assert(Transition && "Do not know how to promote null");
4206 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4207 bool canPromote(const Instruction *ToBePromoted) const {
4208 // We could support CastInst too.
4209 return isa<BinaryOperator>(ToBePromoted);
4212 /// \brief Check if it is profitable to promote \p ToBePromoted
4213 /// by moving downward the transition through.
4214 bool shouldPromote(const Instruction *ToBePromoted) const {
4215 // Promote only if all the operands can be statically expanded.
4216 // Indeed, we do not want to introduce any new kind of transitions.
4217 for (const Use &U : ToBePromoted->operands()) {
4218 const Value *Val = U.get();
4219 if (Val == getEndOfTransition()) {
4220 // If the use is a division and the transition is on the rhs,
4221 // we cannot promote the operation, otherwise we may create a
4222 // division by zero.
4223 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4227 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4228 !isa<ConstantFP>(Val))
4231 // Check that the resulting operation is legal.
4232 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4235 return StressStoreExtract ||
4236 TLI.isOperationLegalOrCustom(
4237 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
4240 /// \brief Check whether or not \p Use can be combined
4241 /// with the transition.
4242 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4243 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4245 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4246 void enqueueForPromotion(Instruction *ToBePromoted) {
4247 InstsToBePromoted.push_back(ToBePromoted);
4250 /// \brief Set the instruction that will be combined with the transition.
4251 void recordCombineInstruction(Instruction *ToBeCombined) {
4252 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4253 CombineInst = ToBeCombined;
4256 /// \brief Promote all the instructions enqueued for promotion if it is
4258 /// \return True if the promotion happened, false otherwise.
4260 // Check if there is something to promote.
4261 // Right now, if we do not have anything to combine with,
4262 // we assume the promotion is not profitable.
4263 if (InstsToBePromoted.empty() || !CombineInst)
4267 if (!StressStoreExtract && !isProfitableToPromote())
4271 for (auto &ToBePromoted : InstsToBePromoted)
4272 promoteImpl(ToBePromoted);
4273 InstsToBePromoted.clear();
4277 } // End of anonymous namespace.
4279 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4280 // At this point, we know that all the operands of ToBePromoted but Def
4281 // can be statically promoted.
4282 // For Def, we need to use its parameter in ToBePromoted:
4283 // b = ToBePromoted ty1 a
4284 // Def = Transition ty1 b to ty2
4285 // Move the transition down.
4286 // 1. Replace all uses of the promoted operation by the transition.
4287 // = ... b => = ... Def.
4288 assert(ToBePromoted->getType() == Transition->getType() &&
4289 "The type of the result of the transition does not match "
4291 ToBePromoted->replaceAllUsesWith(Transition);
4292 // 2. Update the type of the uses.
4293 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
4294 Type *TransitionTy = getTransitionType();
4295 ToBePromoted->mutateType(TransitionTy);
4296 // 3. Update all the operands of the promoted operation with promoted
4298 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
4299 for (Use &U : ToBePromoted->operands()) {
4300 Value *Val = U.get();
4301 Value *NewVal = nullptr;
4302 if (Val == Transition)
4303 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
4304 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
4305 isa<ConstantFP>(Val)) {
4306 // Use a splat constant if it is not safe to use undef.
4307 NewVal = getConstantVector(
4308 cast<Constant>(Val),
4309 isa<UndefValue>(Val) ||
4310 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
4312 llvm_unreachable("Did you modified shouldPromote and forgot to update "
4314 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
4316 Transition->removeFromParent();
4317 Transition->insertAfter(ToBePromoted);
4318 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
4321 /// Some targets can do store(extractelement) with one instruction.
4322 /// Try to push the extractelement towards the stores when the target
4323 /// has this feature and this is profitable.
4324 bool CodeGenPrepare::OptimizeExtractElementInst(Instruction *Inst) {
4325 unsigned CombineCost = UINT_MAX;
4326 if (DisableStoreExtract || !TLI ||
4327 (!StressStoreExtract &&
4328 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4329 Inst->getOperand(1), CombineCost)))
4332 // At this point we know that Inst is a vector to scalar transition.
4333 // Try to move it down the def-use chain, until:
4334 // - We can combine the transition with its single use
4335 // => we got rid of the transition.
4336 // - We escape the current basic block
4337 // => we would need to check that we are moving it at a cheaper place and
4338 // we do not do that for now.
4339 BasicBlock *Parent = Inst->getParent();
4340 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4341 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
4342 // If the transition has more than one use, assume this is not going to be
4344 while (Inst->hasOneUse()) {
4345 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4346 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4348 if (ToBePromoted->getParent() != Parent) {
4349 DEBUG(dbgs() << "Instruction to promote is in a different block ("
4350 << ToBePromoted->getParent()->getName()
4351 << ") than the transition (" << Parent->getName() << ").\n");
4355 if (VPH.canCombine(ToBePromoted)) {
4356 DEBUG(dbgs() << "Assume " << *Inst << '\n'
4357 << "will be combined with: " << *ToBePromoted << '\n');
4358 VPH.recordCombineInstruction(ToBePromoted);
4359 bool Changed = VPH.promote();
4360 NumStoreExtractExposed += Changed;
4364 DEBUG(dbgs() << "Try promoting.\n");
4365 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4368 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4370 VPH.enqueueForPromotion(ToBePromoted);
4371 Inst = ToBePromoted;
4376 bool CodeGenPrepare::OptimizeInst(Instruction *I, bool& ModifiedDT) {
4377 // Bail out if we inserted the instruction to prevent optimizations from
4378 // stepping on each other's toes.
4379 if (InsertedInsts.count(I))
4382 if (PHINode *P = dyn_cast<PHINode>(I)) {
4383 // It is possible for very late stage optimizations (such as SimplifyCFG)
4384 // to introduce PHI nodes too late to be cleaned up. If we detect such a
4385 // trivial PHI, go ahead and zap it here.
4386 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
4387 P->replaceAllUsesWith(V);
4388 P->eraseFromParent();
4395 if (CastInst *CI = dyn_cast<CastInst>(I)) {
4396 // If the source of the cast is a constant, then this should have
4397 // already been constant folded. The only reason NOT to constant fold
4398 // it is if something (e.g. LSR) was careful to place the constant
4399 // evaluation in a block other than then one that uses it (e.g. to hoist
4400 // the address of globals out of a loop). If this is the case, we don't
4401 // want to forward-subst the cast.
4402 if (isa<Constant>(CI->getOperand(0)))
4405 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
4408 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4409 /// Sink a zext or sext into its user blocks if the target type doesn't
4410 /// fit in one register
4412 TLI->getTypeAction(CI->getContext(),
4413 TLI->getValueType(*DL, CI->getType())) ==
4414 TargetLowering::TypeExpandInteger) {
4415 return SinkCast(CI);
4417 bool MadeChange = MoveExtToFormExtLoad(I);
4418 return MadeChange | OptimizeExtUses(I);
4424 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4425 if (!TLI || !TLI->hasMultipleConditionRegisters())
4426 return OptimizeCmpExpression(CI);
4428 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4429 stripInvariantGroupMetadata(*LI);
4431 unsigned AS = LI->getPointerAddressSpace();
4432 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
4437 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4438 stripInvariantGroupMetadata(*SI);
4440 unsigned AS = SI->getPointerAddressSpace();
4441 return OptimizeMemoryInst(I, SI->getOperand(1),
4442 SI->getOperand(0)->getType(), AS);
4447 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4449 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4450 BinOp->getOpcode() == Instruction::LShr)) {
4451 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4452 if (TLI && CI && TLI->hasExtractBitsInsn())
4453 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
4458 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4459 if (GEPI->hasAllZeroIndices()) {
4460 /// The GEP operand must be a pointer, so must its result -> BitCast
4461 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4462 GEPI->getName(), GEPI);
4463 GEPI->replaceAllUsesWith(NC);
4464 GEPI->eraseFromParent();
4466 OptimizeInst(NC, ModifiedDT);
4472 if (CallInst *CI = dyn_cast<CallInst>(I))
4473 return OptimizeCallInst(CI, ModifiedDT);
4475 if (SelectInst *SI = dyn_cast<SelectInst>(I))
4476 return OptimizeSelectInst(SI);
4478 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
4479 return OptimizeShuffleVectorInst(SVI);
4481 if (isa<ExtractElementInst>(I))
4482 return OptimizeExtractElementInst(I);
4487 // In this pass we look for GEP and cast instructions that are used
4488 // across basic blocks and rewrite them to improve basic-block-at-a-time
4490 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
4492 bool MadeChange = false;
4494 CurInstIterator = BB.begin();
4495 while (CurInstIterator != BB.end()) {
4496 MadeChange |= OptimizeInst(CurInstIterator++, ModifiedDT);
4500 MadeChange |= DupRetToEnableTailCallOpts(&BB);
4505 // llvm.dbg.value is far away from the value then iSel may not be able
4506 // handle it properly. iSel will drop llvm.dbg.value if it can not
4507 // find a node corresponding to the value.
4508 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
4509 bool MadeChange = false;
4510 for (BasicBlock &BB : F) {
4511 Instruction *PrevNonDbgInst = nullptr;
4512 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
4513 Instruction *Insn = BI++;
4514 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
4515 // Leave dbg.values that refer to an alloca alone. These
4516 // instrinsics describe the address of a variable (= the alloca)
4517 // being taken. They should not be moved next to the alloca
4518 // (and to the beginning of the scope), but rather stay close to
4519 // where said address is used.
4520 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
4521 PrevNonDbgInst = Insn;
4525 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
4526 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
4527 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
4528 DVI->removeFromParent();
4529 if (isa<PHINode>(VI))
4530 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
4532 DVI->insertAfter(VI);
4541 // If there is a sequence that branches based on comparing a single bit
4542 // against zero that can be combined into a single instruction, and the
4543 // target supports folding these into a single instruction, sink the
4544 // mask and compare into the branch uses. Do this before OptimizeBlock ->
4545 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
4547 bool CodeGenPrepare::sinkAndCmp(Function &F) {
4548 if (!EnableAndCmpSinking)
4550 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
4552 bool MadeChange = false;
4553 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
4554 BasicBlock *BB = I++;
4556 // Does this BB end with the following?
4557 // %andVal = and %val, #single-bit-set
4558 // %icmpVal = icmp %andResult, 0
4559 // br i1 %cmpVal label %dest1, label %dest2"
4560 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
4561 if (!Brcc || !Brcc->isConditional())
4563 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
4564 if (!Cmp || Cmp->getParent() != BB)
4566 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
4567 if (!Zero || !Zero->isZero())
4569 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
4570 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
4572 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
4573 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
4575 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
4577 // Push the "and; icmp" for any users that are conditional branches.
4578 // Since there can only be one branch use per BB, we don't need to keep
4579 // track of which BBs we insert into.
4580 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
4584 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
4586 if (!BrccUser || !BrccUser->isConditional())
4588 BasicBlock *UserBB = BrccUser->getParent();
4589 if (UserBB == BB) continue;
4590 DEBUG(dbgs() << "found Brcc use\n");
4592 // Sink the "and; icmp" to use.
4594 BinaryOperator *NewAnd =
4595 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
4598 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
4602 DEBUG(BrccUser->getParent()->dump());
4608 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
4609 /// success, or returns false if no or invalid metadata was found.
4610 static bool extractBranchMetadata(BranchInst *BI,
4611 uint64_t &ProbTrue, uint64_t &ProbFalse) {
4612 assert(BI->isConditional() &&
4613 "Looking for probabilities on unconditional branch?");
4614 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
4615 if (!ProfileData || ProfileData->getNumOperands() != 3)
4618 const auto *CITrue =
4619 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
4620 const auto *CIFalse =
4621 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
4622 if (!CITrue || !CIFalse)
4625 ProbTrue = CITrue->getValue().getZExtValue();
4626 ProbFalse = CIFalse->getValue().getZExtValue();
4631 /// \brief Scale down both weights to fit into uint32_t.
4632 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
4633 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
4634 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
4635 NewTrue = NewTrue / Scale;
4636 NewFalse = NewFalse / Scale;
4639 /// \brief Some targets prefer to split a conditional branch like:
4641 /// %0 = icmp ne i32 %a, 0
4642 /// %1 = icmp ne i32 %b, 0
4643 /// %or.cond = or i1 %0, %1
4644 /// br i1 %or.cond, label %TrueBB, label %FalseBB
4646 /// into multiple branch instructions like:
4649 /// %0 = icmp ne i32 %a, 0
4650 /// br i1 %0, label %TrueBB, label %bb2
4652 /// %1 = icmp ne i32 %b, 0
4653 /// br i1 %1, label %TrueBB, label %FalseBB
4655 /// This usually allows instruction selection to do even further optimizations
4656 /// and combine the compare with the branch instruction. Currently this is
4657 /// applied for targets which have "cheap" jump instructions.
4659 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
4661 bool CodeGenPrepare::splitBranchCondition(Function &F) {
4662 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
4665 bool MadeChange = false;
4666 for (auto &BB : F) {
4667 // Does this BB end with the following?
4668 // %cond1 = icmp|fcmp|binary instruction ...
4669 // %cond2 = icmp|fcmp|binary instruction ...
4670 // %cond.or = or|and i1 %cond1, cond2
4671 // br i1 %cond.or label %dest1, label %dest2"
4672 BinaryOperator *LogicOp;
4673 BasicBlock *TBB, *FBB;
4674 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
4677 auto *Br1 = cast<BranchInst>(BB.getTerminator());
4678 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
4682 Value *Cond1, *Cond2;
4683 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
4684 m_OneUse(m_Value(Cond2)))))
4685 Opc = Instruction::And;
4686 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
4687 m_OneUse(m_Value(Cond2)))))
4688 Opc = Instruction::Or;
4692 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
4693 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
4696 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
4699 auto *InsertBefore = std::next(Function::iterator(BB))
4700 .getNodePtrUnchecked();
4701 auto TmpBB = BasicBlock::Create(BB.getContext(),
4702 BB.getName() + ".cond.split",
4703 BB.getParent(), InsertBefore);
4705 // Update original basic block by using the first condition directly by the
4706 // branch instruction and removing the no longer needed and/or instruction.
4707 Br1->setCondition(Cond1);
4708 LogicOp->eraseFromParent();
4710 // Depending on the conditon we have to either replace the true or the false
4711 // successor of the original branch instruction.
4712 if (Opc == Instruction::And)
4713 Br1->setSuccessor(0, TmpBB);
4715 Br1->setSuccessor(1, TmpBB);
4717 // Fill in the new basic block.
4718 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
4719 if (auto *I = dyn_cast<Instruction>(Cond2)) {
4720 I->removeFromParent();
4721 I->insertBefore(Br2);
4724 // Update PHI nodes in both successors. The original BB needs to be
4725 // replaced in one succesor's PHI nodes, because the branch comes now from
4726 // the newly generated BB (NewBB). In the other successor we need to add one
4727 // incoming edge to the PHI nodes, because both branch instructions target
4728 // now the same successor. Depending on the original branch condition
4729 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
4730 // we perfrom the correct update for the PHI nodes.
4731 // This doesn't change the successor order of the just created branch
4732 // instruction (or any other instruction).
4733 if (Opc == Instruction::Or)
4734 std::swap(TBB, FBB);
4736 // Replace the old BB with the new BB.
4737 for (auto &I : *TBB) {
4738 PHINode *PN = dyn_cast<PHINode>(&I);
4742 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
4743 PN->setIncomingBlock(i, TmpBB);
4746 // Add another incoming edge form the new BB.
4747 for (auto &I : *FBB) {
4748 PHINode *PN = dyn_cast<PHINode>(&I);
4751 auto *Val = PN->getIncomingValueForBlock(&BB);
4752 PN->addIncoming(Val, TmpBB);
4755 // Update the branch weights (from SelectionDAGBuilder::
4756 // FindMergedConditions).
4757 if (Opc == Instruction::Or) {
4758 // Codegen X | Y as:
4767 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
4768 // The requirement is that
4769 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
4770 // = TrueProb for orignal BB.
4771 // Assuming the orignal weights are A and B, one choice is to set BB1's
4772 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
4774 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
4775 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
4776 // TmpBB, but the math is more complicated.
4777 uint64_t TrueWeight, FalseWeight;
4778 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4779 uint64_t NewTrueWeight = TrueWeight;
4780 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
4781 scaleWeights(NewTrueWeight, NewFalseWeight);
4782 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4783 .createBranchWeights(TrueWeight, FalseWeight));
4785 NewTrueWeight = TrueWeight;
4786 NewFalseWeight = 2 * FalseWeight;
4787 scaleWeights(NewTrueWeight, NewFalseWeight);
4788 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4789 .createBranchWeights(TrueWeight, FalseWeight));
4792 // Codegen X & Y as:
4800 // This requires creation of TmpBB after CurBB.
4802 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
4803 // The requirement is that
4804 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
4805 // = FalseProb for orignal BB.
4806 // Assuming the orignal weights are A and B, one choice is to set BB1's
4807 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
4809 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
4810 uint64_t TrueWeight, FalseWeight;
4811 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4812 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
4813 uint64_t NewFalseWeight = FalseWeight;
4814 scaleWeights(NewTrueWeight, NewFalseWeight);
4815 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4816 .createBranchWeights(TrueWeight, FalseWeight));
4818 NewTrueWeight = 2 * TrueWeight;
4819 NewFalseWeight = FalseWeight;
4820 scaleWeights(NewTrueWeight, NewFalseWeight);
4821 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4822 .createBranchWeights(TrueWeight, FalseWeight));
4826 // Note: No point in getting fancy here, since the DT info is never
4827 // available to CodeGenPrepare.
4832 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
4838 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
4839 if (auto *InvariantMD = I.getMetadata("invariant.group"))
4840 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());