1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/IR/CallSite.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/GetElementPtrTypeIterator.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/InlineAsm.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/MDBuilder.h"
35 #include "llvm/IR/PatternMatch.h"
36 #include "llvm/IR/Statepoint.h"
37 #include "llvm/IR/ValueHandle.h"
38 #include "llvm/IR/ValueMap.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Target/TargetLowering.h"
44 #include "llvm/Target/TargetSubtargetInfo.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/BuildLibCalls.h"
47 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
48 #include "llvm/Transforms/Utils/Local.h"
49 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
51 using namespace llvm::PatternMatch;
53 #define DEBUG_TYPE "codegenprepare"
55 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
56 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
57 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
58 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
60 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
62 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
63 "computations were sunk");
64 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
65 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
66 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
67 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
68 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
69 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
70 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
72 static cl::opt<bool> DisableBranchOpts(
73 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
74 cl::desc("Disable branch optimizations in CodeGenPrepare"));
77 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
78 cl::desc("Disable GC optimizations in CodeGenPrepare"));
80 static cl::opt<bool> DisableSelectToBranch(
81 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
82 cl::desc("Disable select to branch conversion."));
84 static cl::opt<bool> AddrSinkUsingGEPs(
85 "addr-sink-using-gep", cl::Hidden, cl::init(false),
86 cl::desc("Address sinking in CGP using GEPs."));
88 static cl::opt<bool> EnableAndCmpSinking(
89 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
90 cl::desc("Enable sinkinig and/cmp into branches."));
92 static cl::opt<bool> DisableStoreExtract(
93 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
94 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
96 static cl::opt<bool> StressStoreExtract(
97 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
98 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
100 static cl::opt<bool> DisableExtLdPromotion(
101 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
102 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
105 static cl::opt<bool> StressExtLdPromotion(
106 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
107 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
108 "optimization in CodeGenPrepare"));
111 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
115 TypeIsSExt(Type *Ty, bool IsSExt) : Ty(Ty), IsSExt(IsSExt) {}
117 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
118 class TypePromotionTransaction;
120 class CodeGenPrepare : public FunctionPass {
121 /// TLI - Keep a pointer of a TargetLowering to consult for determining
122 /// transformation profitability.
123 const TargetMachine *TM;
124 const TargetLowering *TLI;
125 const TargetTransformInfo *TTI;
126 const TargetLibraryInfo *TLInfo;
128 /// CurInstIterator - As we scan instructions optimizing them, this is the
129 /// next instruction to optimize. Xforms that can invalidate this should
131 BasicBlock::iterator CurInstIterator;
133 /// Keeps track of non-local addresses that have been sunk into a block.
134 /// This allows us to avoid inserting duplicate code for blocks with
135 /// multiple load/stores of the same address.
136 ValueMap<Value*, Value*> SunkAddrs;
138 /// Keeps track of all truncates inserted for the current function.
139 SetOfInstrs InsertedTruncsSet;
140 /// Keeps track of the type of the related instruction before their
141 /// promotion for the current function.
142 InstrToOrigTy PromotedInsts;
144 /// ModifiedDT - If CFG is modified in anyway.
147 /// OptSize - True if optimizing for size.
151 static char ID; // Pass identification, replacement for typeid
152 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
153 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr) {
154 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
156 bool runOnFunction(Function &F) override;
158 const char *getPassName() const override { return "CodeGen Prepare"; }
160 void getAnalysisUsage(AnalysisUsage &AU) const override {
161 AU.addPreserved<DominatorTreeWrapperPass>();
162 AU.addRequired<TargetLibraryInfoWrapperPass>();
163 AU.addRequired<TargetTransformInfoWrapperPass>();
167 bool EliminateFallThrough(Function &F);
168 bool EliminateMostlyEmptyBlocks(Function &F);
169 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
170 void EliminateMostlyEmptyBlock(BasicBlock *BB);
171 bool OptimizeBlock(BasicBlock &BB, bool& ModifiedDT);
172 bool OptimizeInst(Instruction *I, bool& ModifiedDT);
173 bool OptimizeMemoryInst(Instruction *I, Value *Addr,
174 Type *AccessTy, unsigned AS);
175 bool OptimizeInlineAsmInst(CallInst *CS);
176 bool OptimizeCallInst(CallInst *CI, bool& ModifiedDT);
177 bool MoveExtToFormExtLoad(Instruction *&I);
178 bool OptimizeExtUses(Instruction *I);
179 bool OptimizeSelectInst(SelectInst *SI);
180 bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
181 bool OptimizeExtractElementInst(Instruction *Inst);
182 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
183 bool PlaceDbgValues(Function &F);
184 bool sinkAndCmp(Function &F);
185 bool ExtLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
187 const SmallVectorImpl<Instruction *> &Exts,
188 unsigned CreatedInstCost);
189 bool splitBranchCondition(Function &F);
190 bool simplifyOffsetableRelocate(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 bool EverMadeChange = false;
207 // Clear per function information.
208 InsertedTruncsSet.clear();
209 PromotedInsts.clear();
213 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
214 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
215 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
216 OptSize = F.hasFnAttribute(Attribute::OptimizeForSize);
218 /// This optimization identifies DIV instructions that can be
219 /// profitably bypassed and carried out with a shorter, faster divide.
220 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
221 const DenseMap<unsigned int, unsigned int> &BypassWidths =
222 TLI->getBypassSlowDivWidths();
223 for (Function::iterator I = F.begin(); I != F.end(); I++)
224 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
227 // Eliminate blocks that contain only PHI nodes and an
228 // unconditional branch.
229 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
231 // llvm.dbg.value is far away from the value then iSel may not be able
232 // handle it properly. iSel will drop llvm.dbg.value if it can not
233 // find a node corresponding to the value.
234 EverMadeChange |= PlaceDbgValues(F);
236 // If there is a mask, compare against zero, and branch that can be combined
237 // into a single target instruction, push the mask and compare into branch
238 // users. Do this before OptimizeBlock -> OptimizeInst ->
239 // OptimizeCmpExpression, which perturbs the pattern being searched for.
240 if (!DisableBranchOpts) {
241 EverMadeChange |= sinkAndCmp(F);
242 EverMadeChange |= splitBranchCondition(F);
245 bool MadeChange = true;
248 for (Function::iterator I = F.begin(); I != F.end(); ) {
249 BasicBlock *BB = I++;
250 bool ModifiedDTOnIteration = false;
251 MadeChange |= OptimizeBlock(*BB, ModifiedDTOnIteration);
253 // Restart BB iteration if the dominator tree of the Function was changed
254 if (ModifiedDTOnIteration)
257 EverMadeChange |= MadeChange;
262 if (!DisableBranchOpts) {
264 SmallPtrSet<BasicBlock*, 8> WorkList;
265 for (BasicBlock &BB : F) {
266 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
267 MadeChange |= ConstantFoldTerminator(&BB, true);
268 if (!MadeChange) continue;
270 for (SmallVectorImpl<BasicBlock*>::iterator
271 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
272 if (pred_begin(*II) == pred_end(*II))
273 WorkList.insert(*II);
276 // Delete the dead blocks and any of their dead successors.
277 MadeChange |= !WorkList.empty();
278 while (!WorkList.empty()) {
279 BasicBlock *BB = *WorkList.begin();
281 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
285 for (SmallVectorImpl<BasicBlock*>::iterator
286 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
287 if (pred_begin(*II) == pred_end(*II))
288 WorkList.insert(*II);
291 // Merge pairs of basic blocks with unconditional branches, connected by
293 if (EverMadeChange || MadeChange)
294 MadeChange |= EliminateFallThrough(F);
296 EverMadeChange |= MadeChange;
299 if (!DisableGCOpts) {
300 SmallVector<Instruction *, 2> Statepoints;
301 for (BasicBlock &BB : F)
302 for (Instruction &I : BB)
304 Statepoints.push_back(&I);
305 for (auto &I : Statepoints)
306 EverMadeChange |= simplifyOffsetableRelocate(*I);
309 return EverMadeChange;
312 /// EliminateFallThrough - Merge basic blocks which are connected
313 /// by a single edge, where one of the basic blocks has a single successor
314 /// pointing to the other basic block, which has a single predecessor.
315 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
316 bool Changed = false;
317 // Scan all of the blocks in the function, except for the entry block.
318 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
319 BasicBlock *BB = I++;
320 // If the destination block has a single pred, then this is a trivial
321 // edge, just collapse it.
322 BasicBlock *SinglePred = BB->getSinglePredecessor();
324 // Don't merge if BB's address is taken.
325 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
327 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
328 if (Term && !Term->isConditional()) {
330 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
331 // Remember if SinglePred was the entry block of the function.
332 // If so, we will need to move BB back to the entry position.
333 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
334 MergeBasicBlockIntoOnlyPred(BB, nullptr);
336 if (isEntry && BB != &BB->getParent()->getEntryBlock())
337 BB->moveBefore(&BB->getParent()->getEntryBlock());
339 // We have erased a block. Update the iterator.
346 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
347 /// debug info directives, and an unconditional branch. Passes before isel
348 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
349 /// isel. Start by eliminating these blocks so we can split them the way we
351 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
352 bool MadeChange = false;
353 // Note that this intentionally skips the entry block.
354 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
355 BasicBlock *BB = I++;
357 // If this block doesn't end with an uncond branch, ignore it.
358 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
359 if (!BI || !BI->isUnconditional())
362 // If the instruction before the branch (skipping debug info) isn't a phi
363 // node, then other stuff is happening here.
364 BasicBlock::iterator BBI = BI;
365 if (BBI != BB->begin()) {
367 while (isa<DbgInfoIntrinsic>(BBI)) {
368 if (BBI == BB->begin())
372 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
376 // Do not break infinite loops.
377 BasicBlock *DestBB = BI->getSuccessor(0);
381 if (!CanMergeBlocks(BB, DestBB))
384 EliminateMostlyEmptyBlock(BB);
390 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
391 /// single uncond branch between them, and BB contains no other non-phi
393 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
394 const BasicBlock *DestBB) const {
395 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
396 // the successor. If there are more complex condition (e.g. preheaders),
397 // don't mess around with them.
398 BasicBlock::const_iterator BBI = BB->begin();
399 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
400 for (const User *U : PN->users()) {
401 const Instruction *UI = cast<Instruction>(U);
402 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
404 // If User is inside DestBB block and it is a PHINode then check
405 // incoming value. If incoming value is not from BB then this is
406 // a complex condition (e.g. preheaders) we want to avoid here.
407 if (UI->getParent() == DestBB) {
408 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
409 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
410 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
411 if (Insn && Insn->getParent() == BB &&
412 Insn->getParent() != UPN->getIncomingBlock(I))
419 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
420 // and DestBB may have conflicting incoming values for the block. If so, we
421 // can't merge the block.
422 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
423 if (!DestBBPN) return true; // no conflict.
425 // Collect the preds of BB.
426 SmallPtrSet<const BasicBlock*, 16> BBPreds;
427 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
428 // It is faster to get preds from a PHI than with pred_iterator.
429 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
430 BBPreds.insert(BBPN->getIncomingBlock(i));
432 BBPreds.insert(pred_begin(BB), pred_end(BB));
435 // Walk the preds of DestBB.
436 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
437 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
438 if (BBPreds.count(Pred)) { // Common predecessor?
439 BBI = DestBB->begin();
440 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
441 const Value *V1 = PN->getIncomingValueForBlock(Pred);
442 const Value *V2 = PN->getIncomingValueForBlock(BB);
444 // If V2 is a phi node in BB, look up what the mapped value will be.
445 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
446 if (V2PN->getParent() == BB)
447 V2 = V2PN->getIncomingValueForBlock(Pred);
449 // If there is a conflict, bail out.
450 if (V1 != V2) return false;
459 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
460 /// an unconditional branch in it.
461 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
462 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
463 BasicBlock *DestBB = BI->getSuccessor(0);
465 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
467 // If the destination block has a single pred, then this is a trivial edge,
469 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
470 if (SinglePred != DestBB) {
471 // Remember if SinglePred was the entry block of the function. If so, we
472 // will need to move BB back to the entry position.
473 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
474 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
476 if (isEntry && BB != &BB->getParent()->getEntryBlock())
477 BB->moveBefore(&BB->getParent()->getEntryBlock());
479 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
484 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
485 // to handle the new incoming edges it is about to have.
487 for (BasicBlock::iterator BBI = DestBB->begin();
488 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
489 // Remove the incoming value for BB, and remember it.
490 Value *InVal = PN->removeIncomingValue(BB, false);
492 // Two options: either the InVal is a phi node defined in BB or it is some
493 // value that dominates BB.
494 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
495 if (InValPhi && InValPhi->getParent() == BB) {
496 // Add all of the input values of the input PHI as inputs of this phi.
497 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
498 PN->addIncoming(InValPhi->getIncomingValue(i),
499 InValPhi->getIncomingBlock(i));
501 // Otherwise, add one instance of the dominating value for each edge that
502 // we will be adding.
503 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
504 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
505 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
507 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
508 PN->addIncoming(InVal, *PI);
513 // The PHIs are now updated, change everything that refers to BB to use
514 // DestBB and remove BB.
515 BB->replaceAllUsesWith(DestBB);
516 BB->eraseFromParent();
519 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
522 // Computes a map of base pointer relocation instructions to corresponding
523 // derived pointer relocation instructions given a vector of all relocate calls
524 static void computeBaseDerivedRelocateMap(
525 const SmallVectorImpl<User *> &AllRelocateCalls,
526 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
528 // Collect information in two maps: one primarily for locating the base object
529 // while filling the second map; the second map is the final structure holding
530 // a mapping between Base and corresponding Derived relocate calls
531 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
532 for (auto &U : AllRelocateCalls) {
533 GCRelocateOperands ThisRelocate(U);
534 IntrinsicInst *I = cast<IntrinsicInst>(U);
535 auto K = std::make_pair(ThisRelocate.getBasePtrIndex(),
536 ThisRelocate.getDerivedPtrIndex());
537 RelocateIdxMap.insert(std::make_pair(K, I));
539 for (auto &Item : RelocateIdxMap) {
540 std::pair<unsigned, unsigned> Key = Item.first;
541 if (Key.first == Key.second)
542 // Base relocation: nothing to insert
545 IntrinsicInst *I = Item.second;
546 auto BaseKey = std::make_pair(Key.first, Key.first);
548 // We're iterating over RelocateIdxMap so we cannot modify it.
549 auto MaybeBase = RelocateIdxMap.find(BaseKey);
550 if (MaybeBase == RelocateIdxMap.end())
551 // TODO: We might want to insert a new base object relocate and gep off
552 // that, if there are enough derived object relocates.
555 RelocateInstMap[MaybeBase->second].push_back(I);
559 // Accepts a GEP and extracts the operands into a vector provided they're all
560 // small integer constants
561 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
562 SmallVectorImpl<Value *> &OffsetV) {
563 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
564 // Only accept small constant integer operands
565 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
566 if (!Op || Op->getZExtValue() > 20)
570 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
571 OffsetV.push_back(GEP->getOperand(i));
575 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
576 // replace, computes a replacement, and affects it.
578 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
579 const SmallVectorImpl<IntrinsicInst *> &Targets) {
580 bool MadeChange = false;
581 for (auto &ToReplace : Targets) {
582 GCRelocateOperands MasterRelocate(RelocatedBase);
583 GCRelocateOperands ThisRelocate(ToReplace);
585 assert(ThisRelocate.getBasePtrIndex() == MasterRelocate.getBasePtrIndex() &&
586 "Not relocating a derived object of the original base object");
587 if (ThisRelocate.getBasePtrIndex() == ThisRelocate.getDerivedPtrIndex()) {
588 // A duplicate relocate call. TODO: coalesce duplicates.
592 Value *Base = ThisRelocate.getBasePtr();
593 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.getDerivedPtr());
594 if (!Derived || Derived->getPointerOperand() != Base)
597 SmallVector<Value *, 2> OffsetV;
598 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
601 // Create a Builder and replace the target callsite with a gep
602 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
604 // Insert after RelocatedBase
605 IRBuilder<> Builder(RelocatedBase->getNextNode());
606 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
608 // If gc_relocate does not match the actual type, cast it to the right type.
609 // In theory, there must be a bitcast after gc_relocate if the type does not
610 // match, and we should reuse it to get the derived pointer. But it could be
614 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
619 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
623 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
624 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
626 // In this case, we can not find the bitcast any more. So we insert a new bitcast
627 // no matter there is already one or not. In this way, we can handle all cases, and
628 // the extra bitcast should be optimized away in later passes.
629 Instruction *ActualRelocatedBase = RelocatedBase;
630 if (RelocatedBase->getType() != Base->getType()) {
631 ActualRelocatedBase =
632 cast<Instruction>(Builder.CreateBitCast(RelocatedBase, Base->getType()));
634 Value *Replacement = Builder.CreateGEP(
635 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
636 Instruction *ReplacementInst = cast<Instruction>(Replacement);
637 Replacement->takeName(ToReplace);
638 // If the newly generated derived pointer's type does not match the original derived
639 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
640 Instruction *ActualReplacement = ReplacementInst;
641 if (ReplacementInst->getType() != ToReplace->getType()) {
643 cast<Instruction>(Builder.CreateBitCast(ReplacementInst, ToReplace->getType()));
645 ToReplace->replaceAllUsesWith(ActualReplacement);
646 ToReplace->eraseFromParent();
656 // %ptr = gep %base + 15
657 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
658 // %base' = relocate(%tok, i32 4, i32 4)
659 // %ptr' = relocate(%tok, i32 4, i32 5)
665 // %ptr = gep %base + 15
666 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
667 // %base' = gc.relocate(%tok, i32 4, i32 4)
668 // %ptr' = gep %base' + 15
670 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
671 bool MadeChange = false;
672 SmallVector<User *, 2> AllRelocateCalls;
674 for (auto *U : I.users())
675 if (isGCRelocate(dyn_cast<Instruction>(U)))
676 // Collect all the relocate calls associated with a statepoint
677 AllRelocateCalls.push_back(U);
679 // We need atleast one base pointer relocation + one derived pointer
680 // relocation to mangle
681 if (AllRelocateCalls.size() < 2)
684 // RelocateInstMap is a mapping from the base relocate instruction to the
685 // corresponding derived relocate instructions
686 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
687 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
688 if (RelocateInstMap.empty())
691 for (auto &Item : RelocateInstMap)
692 // Item.first is the RelocatedBase to offset against
693 // Item.second is the vector of Targets to replace
694 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
698 /// SinkCast - Sink the specified cast instruction into its user blocks
699 static bool SinkCast(CastInst *CI) {
700 BasicBlock *DefBB = CI->getParent();
702 /// InsertedCasts - Only insert a cast in each block once.
703 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
705 bool MadeChange = false;
706 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
708 Use &TheUse = UI.getUse();
709 Instruction *User = cast<Instruction>(*UI);
711 // Figure out which BB this cast is used in. For PHI's this is the
712 // appropriate predecessor block.
713 BasicBlock *UserBB = User->getParent();
714 if (PHINode *PN = dyn_cast<PHINode>(User)) {
715 UserBB = PN->getIncomingBlock(TheUse);
718 // Preincrement use iterator so we don't invalidate it.
721 // If this user is in the same block as the cast, don't change the cast.
722 if (UserBB == DefBB) continue;
724 // If we have already inserted a cast into this block, use it.
725 CastInst *&InsertedCast = InsertedCasts[UserBB];
728 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
730 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
734 // Replace a use of the cast with a use of the new cast.
735 TheUse = InsertedCast;
740 // If we removed all uses, nuke the cast.
741 if (CI->use_empty()) {
742 CI->eraseFromParent();
749 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
750 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
751 /// sink it into user blocks to reduce the number of virtual
752 /// registers that must be created and coalesced.
754 /// Return true if any changes are made.
756 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
757 // If this is a noop copy,
758 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
759 EVT DstVT = TLI.getValueType(CI->getType());
761 // This is an fp<->int conversion?
762 if (SrcVT.isInteger() != DstVT.isInteger())
765 // If this is an extension, it will be a zero or sign extension, which
767 if (SrcVT.bitsLT(DstVT)) return false;
769 // If these values will be promoted, find out what they will be promoted
770 // to. This helps us consider truncates on PPC as noop copies when they
772 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
773 TargetLowering::TypePromoteInteger)
774 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
775 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
776 TargetLowering::TypePromoteInteger)
777 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
779 // If, after promotion, these are the same types, this is a noop copy.
786 /// CombineUAddWithOverflow - try to combine CI into a call to the
787 /// llvm.uadd.with.overflow intrinsic if possible.
789 /// Return true if any changes were made.
790 static bool CombineUAddWithOverflow(CmpInst *CI) {
794 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
797 Type *Ty = AddI->getType();
798 if (!isa<IntegerType>(Ty))
801 // We don't want to move around uses of condition values this late, so we we
802 // check if it is legal to create the call to the intrinsic in the basic
803 // block containing the icmp:
805 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
809 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
811 if (AddI->hasOneUse())
812 assert(*AddI->user_begin() == CI && "expected!");
815 Module *M = CI->getParent()->getParent()->getParent();
816 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
818 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
820 auto *UAddWithOverflow =
821 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
822 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
824 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
826 CI->replaceAllUsesWith(Overflow);
827 AddI->replaceAllUsesWith(UAdd);
828 CI->eraseFromParent();
829 AddI->eraseFromParent();
833 /// SinkCmpExpression - Sink the given CmpInst into user blocks to reduce
834 /// the number of virtual registers that must be created and coalesced. This is
835 /// a clear win except on targets with multiple condition code registers
836 /// (PowerPC), where it might lose; some adjustment may be wanted there.
838 /// Return true if any changes are made.
839 static bool SinkCmpExpression(CmpInst *CI) {
840 BasicBlock *DefBB = CI->getParent();
842 /// InsertedCmp - Only insert a cmp in each block once.
843 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
845 bool MadeChange = false;
846 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
848 Use &TheUse = UI.getUse();
849 Instruction *User = cast<Instruction>(*UI);
851 // Preincrement use iterator so we don't invalidate it.
854 // Don't bother for PHI nodes.
855 if (isa<PHINode>(User))
858 // Figure out which BB this cmp is used in.
859 BasicBlock *UserBB = User->getParent();
861 // If this user is in the same block as the cmp, don't change the cmp.
862 if (UserBB == DefBB) continue;
864 // If we have already inserted a cmp into this block, use it.
865 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
868 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
870 CmpInst::Create(CI->getOpcode(),
871 CI->getPredicate(), CI->getOperand(0),
872 CI->getOperand(1), "", InsertPt);
875 // Replace a use of the cmp with a use of the new cmp.
876 TheUse = InsertedCmp;
881 // If we removed all uses, nuke the cmp.
882 if (CI->use_empty()) {
883 CI->eraseFromParent();
890 static bool OptimizeCmpExpression(CmpInst *CI) {
891 if (SinkCmpExpression(CI))
894 if (CombineUAddWithOverflow(CI))
900 /// isExtractBitsCandidateUse - Check if the candidates could
901 /// be combined with shift instruction, which includes:
902 /// 1. Truncate instruction
903 /// 2. And instruction and the imm is a mask of the low bits:
904 /// imm & (imm+1) == 0
905 static bool isExtractBitsCandidateUse(Instruction *User) {
906 if (!isa<TruncInst>(User)) {
907 if (User->getOpcode() != Instruction::And ||
908 !isa<ConstantInt>(User->getOperand(1)))
911 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
913 if ((Cimm & (Cimm + 1)).getBoolValue())
919 /// SinkShiftAndTruncate - sink both shift and truncate instruction
920 /// to the use of truncate's BB.
922 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
923 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
924 const TargetLowering &TLI) {
925 BasicBlock *UserBB = User->getParent();
926 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
927 TruncInst *TruncI = dyn_cast<TruncInst>(User);
928 bool MadeChange = false;
930 for (Value::user_iterator TruncUI = TruncI->user_begin(),
931 TruncE = TruncI->user_end();
932 TruncUI != TruncE;) {
934 Use &TruncTheUse = TruncUI.getUse();
935 Instruction *TruncUser = cast<Instruction>(*TruncUI);
936 // Preincrement use iterator so we don't invalidate it.
940 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
944 // If the use is actually a legal node, there will not be an
945 // implicit truncate.
946 // FIXME: always querying the result type is just an
947 // approximation; some nodes' legality is determined by the
948 // operand or other means. There's no good way to find out though.
949 if (TLI.isOperationLegalOrCustom(
950 ISDOpcode, TLI.getValueType(TruncUser->getType(), true)))
953 // Don't bother for PHI nodes.
954 if (isa<PHINode>(TruncUser))
957 BasicBlock *TruncUserBB = TruncUser->getParent();
959 if (UserBB == TruncUserBB)
962 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
963 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
965 if (!InsertedShift && !InsertedTrunc) {
966 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
968 if (ShiftI->getOpcode() == Instruction::AShr)
970 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
973 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
976 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
979 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
980 TruncI->getType(), "", TruncInsertPt);
984 TruncTheUse = InsertedTrunc;
990 /// OptimizeExtractBits - sink the shift *right* instruction into user blocks if
991 /// the uses could potentially be combined with this shift instruction and
992 /// generate BitExtract instruction. It will only be applied if the architecture
993 /// supports BitExtract instruction. Here is an example:
995 /// %x.extract.shift = lshr i64 %arg1, 32
997 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1001 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1002 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1004 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1006 /// Return true if any changes are made.
1007 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1008 const TargetLowering &TLI) {
1009 BasicBlock *DefBB = ShiftI->getParent();
1011 /// Only insert instructions in each block once.
1012 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1014 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(ShiftI->getType()));
1016 bool MadeChange = false;
1017 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1019 Use &TheUse = UI.getUse();
1020 Instruction *User = cast<Instruction>(*UI);
1021 // Preincrement use iterator so we don't invalidate it.
1024 // Don't bother for PHI nodes.
1025 if (isa<PHINode>(User))
1028 if (!isExtractBitsCandidateUse(User))
1031 BasicBlock *UserBB = User->getParent();
1033 if (UserBB == DefBB) {
1034 // If the shift and truncate instruction are in the same BB. The use of
1035 // the truncate(TruncUse) may still introduce another truncate if not
1036 // legal. In this case, we would like to sink both shift and truncate
1037 // instruction to the BB of TruncUse.
1040 // i64 shift.result = lshr i64 opnd, imm
1041 // trunc.result = trunc shift.result to i16
1044 // ----> We will have an implicit truncate here if the architecture does
1045 // not have i16 compare.
1046 // cmp i16 trunc.result, opnd2
1048 if (isa<TruncInst>(User) && shiftIsLegal
1049 // If the type of the truncate is legal, no trucate will be
1050 // introduced in other basic blocks.
1051 && (!TLI.isTypeLegal(TLI.getValueType(User->getType()))))
1053 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI);
1057 // If we have already inserted a shift into this block, use it.
1058 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1060 if (!InsertedShift) {
1061 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1063 if (ShiftI->getOpcode() == Instruction::AShr)
1065 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
1068 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
1073 // Replace a use of the shift with a use of the new shift.
1074 TheUse = InsertedShift;
1077 // If we removed all uses, nuke the shift.
1078 if (ShiftI->use_empty())
1079 ShiftI->eraseFromParent();
1084 // ScalarizeMaskedLoad() translates masked load intrinsic, like
1085 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1086 // <16 x i1> %mask, <16 x i32> %passthru)
1087 // to a chain of basic blocks, whith loading element one-by-one if
1088 // the appropriate mask bit is set
1090 // %1 = bitcast i8* %addr to i32*
1091 // %2 = extractelement <16 x i1> %mask, i32 0
1092 // %3 = icmp eq i1 %2, true
1093 // br i1 %3, label %cond.load, label %else
1095 //cond.load: ; preds = %0
1096 // %4 = getelementptr i32* %1, i32 0
1097 // %5 = load i32* %4
1098 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1101 //else: ; preds = %0, %cond.load
1102 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1103 // %7 = extractelement <16 x i1> %mask, i32 1
1104 // %8 = icmp eq i1 %7, true
1105 // br i1 %8, label %cond.load1, label %else2
1107 //cond.load1: ; preds = %else
1108 // %9 = getelementptr i32* %1, i32 1
1109 // %10 = load i32* %9
1110 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1113 //else2: ; preds = %else, %cond.load1
1114 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1115 // %12 = extractelement <16 x i1> %mask, i32 2
1116 // %13 = icmp eq i1 %12, true
1117 // br i1 %13, label %cond.load4, label %else5
1119 static void ScalarizeMaskedLoad(CallInst *CI) {
1120 Value *Ptr = CI->getArgOperand(0);
1121 Value *Src0 = CI->getArgOperand(3);
1122 Value *Mask = CI->getArgOperand(2);
1123 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1124 Type *EltTy = VecType->getElementType();
1126 assert(VecType && "Unexpected return type of masked load intrinsic");
1128 IRBuilder<> Builder(CI->getContext());
1129 Instruction *InsertPt = CI;
1130 BasicBlock *IfBlock = CI->getParent();
1131 BasicBlock *CondBlock = nullptr;
1132 BasicBlock *PrevIfBlock = CI->getParent();
1133 Builder.SetInsertPoint(InsertPt);
1135 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1137 // Bitcast %addr fron i8* to EltTy*
1139 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1140 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1141 Value *UndefVal = UndefValue::get(VecType);
1143 // The result vector
1144 Value *VResult = UndefVal;
1146 PHINode *Phi = nullptr;
1147 Value *PrevPhi = UndefVal;
1149 unsigned VectorWidth = VecType->getNumElements();
1150 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1152 // Fill the "else" block, created in the previous iteration
1154 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1155 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1156 // %to_load = icmp eq i1 %mask_1, true
1157 // br i1 %to_load, label %cond.load, label %else
1160 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1161 Phi->addIncoming(VResult, CondBlock);
1162 Phi->addIncoming(PrevPhi, PrevIfBlock);
1167 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1168 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1169 ConstantInt::get(Predicate->getType(), 1));
1171 // Create "cond" block
1173 // %EltAddr = getelementptr i32* %1, i32 0
1174 // %Elt = load i32* %EltAddr
1175 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1177 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1178 Builder.SetInsertPoint(InsertPt);
1181 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1182 LoadInst* Load = Builder.CreateLoad(Gep, false);
1183 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1185 // Create "else" block, fill it in the next iteration
1186 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1187 Builder.SetInsertPoint(InsertPt);
1188 Instruction *OldBr = IfBlock->getTerminator();
1189 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1190 OldBr->eraseFromParent();
1191 PrevIfBlock = IfBlock;
1192 IfBlock = NewIfBlock;
1195 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1196 Phi->addIncoming(VResult, CondBlock);
1197 Phi->addIncoming(PrevPhi, PrevIfBlock);
1198 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1199 CI->replaceAllUsesWith(NewI);
1200 CI->eraseFromParent();
1203 // ScalarizeMaskedStore() translates masked store intrinsic, like
1204 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1206 // to a chain of basic blocks, that stores element one-by-one if
1207 // the appropriate mask bit is set
1209 // %1 = bitcast i8* %addr to i32*
1210 // %2 = extractelement <16 x i1> %mask, i32 0
1211 // %3 = icmp eq i1 %2, true
1212 // br i1 %3, label %cond.store, label %else
1214 // cond.store: ; preds = %0
1215 // %4 = extractelement <16 x i32> %val, i32 0
1216 // %5 = getelementptr i32* %1, i32 0
1217 // store i32 %4, i32* %5
1220 // else: ; preds = %0, %cond.store
1221 // %6 = extractelement <16 x i1> %mask, i32 1
1222 // %7 = icmp eq i1 %6, true
1223 // br i1 %7, label %cond.store1, label %else2
1225 // cond.store1: ; preds = %else
1226 // %8 = extractelement <16 x i32> %val, i32 1
1227 // %9 = getelementptr i32* %1, i32 1
1228 // store i32 %8, i32* %9
1231 static void ScalarizeMaskedStore(CallInst *CI) {
1232 Value *Ptr = CI->getArgOperand(1);
1233 Value *Src = CI->getArgOperand(0);
1234 Value *Mask = CI->getArgOperand(3);
1236 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1237 Type *EltTy = VecType->getElementType();
1239 assert(VecType && "Unexpected data type in masked store intrinsic");
1241 IRBuilder<> Builder(CI->getContext());
1242 Instruction *InsertPt = CI;
1243 BasicBlock *IfBlock = CI->getParent();
1244 Builder.SetInsertPoint(InsertPt);
1245 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1247 // Bitcast %addr fron i8* to EltTy*
1249 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1250 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1252 unsigned VectorWidth = VecType->getNumElements();
1253 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1255 // Fill the "else" block, created in the previous iteration
1257 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1258 // %to_store = icmp eq i1 %mask_1, true
1259 // br i1 %to_load, label %cond.store, label %else
1261 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1262 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1263 ConstantInt::get(Predicate->getType(), 1));
1265 // Create "cond" block
1267 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1268 // %EltAddr = getelementptr i32* %1, i32 0
1269 // %store i32 %OneElt, i32* %EltAddr
1271 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1272 Builder.SetInsertPoint(InsertPt);
1274 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1276 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1277 Builder.CreateStore(OneElt, Gep);
1279 // Create "else" block, fill it in the next iteration
1280 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1281 Builder.SetInsertPoint(InsertPt);
1282 Instruction *OldBr = IfBlock->getTerminator();
1283 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1284 OldBr->eraseFromParent();
1285 IfBlock = NewIfBlock;
1287 CI->eraseFromParent();
1290 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1291 BasicBlock *BB = CI->getParent();
1293 // Lower inline assembly if we can.
1294 // If we found an inline asm expession, and if the target knows how to
1295 // lower it to normal LLVM code, do so now.
1296 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1297 if (TLI->ExpandInlineAsm(CI)) {
1298 // Avoid invalidating the iterator.
1299 CurInstIterator = BB->begin();
1300 // Avoid processing instructions out of order, which could cause
1301 // reuse before a value is defined.
1305 // Sink address computing for memory operands into the block.
1306 if (OptimizeInlineAsmInst(CI))
1310 const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
1312 // Align the pointer arguments to this call if the target thinks it's a good
1314 unsigned MinSize, PrefAlign;
1315 if (TLI && TD && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1316 for (auto &Arg : CI->arg_operands()) {
1317 // We want to align both objects whose address is used directly and
1318 // objects whose address is used in casts and GEPs, though it only makes
1319 // sense for GEPs if the offset is a multiple of the desired alignment and
1320 // if size - offset meets the size threshold.
1321 if (!Arg->getType()->isPointerTy())
1323 APInt Offset(TD->getPointerSizeInBits(
1324 cast<PointerType>(Arg->getType())->getAddressSpace()), 0);
1325 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*TD, Offset);
1326 uint64_t Offset2 = Offset.getLimitedValue();
1327 if ((Offset2 & (PrefAlign-1)) != 0)
1330 if ((AI = dyn_cast<AllocaInst>(Val)) &&
1331 AI->getAlignment() < PrefAlign &&
1332 TD->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1333 AI->setAlignment(PrefAlign);
1334 // Global variables can only be aligned if they are defined in this
1335 // object (i.e. they are uniquely initialized in this object), and
1336 // over-aligning global variables that have an explicit section is
1339 if ((GV = dyn_cast<GlobalVariable>(Val)) &&
1340 GV->hasUniqueInitializer() &&
1341 !GV->hasSection() &&
1342 GV->getAlignment() < PrefAlign &&
1343 TD->getTypeAllocSize(
1344 GV->getType()->getElementType()) >= MinSize + Offset2)
1345 GV->setAlignment(PrefAlign);
1347 // If this is a memcpy (or similar) then we may be able to improve the
1349 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1350 unsigned Align = getKnownAlignment(MI->getDest(), *TD);
1351 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1352 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *TD));
1353 if (Align > MI->getAlignment())
1354 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1358 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1360 switch (II->getIntrinsicID()) {
1362 case Intrinsic::objectsize: {
1363 // Lower all uses of llvm.objectsize.*
1364 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1365 Type *ReturnTy = CI->getType();
1366 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1368 // Substituting this can cause recursive simplifications, which can
1369 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1371 WeakVH IterHandle(CurInstIterator);
1373 replaceAndRecursivelySimplify(CI, RetVal,
1376 // If the iterator instruction was recursively deleted, start over at the
1377 // start of the block.
1378 if (IterHandle != CurInstIterator) {
1379 CurInstIterator = BB->begin();
1384 case Intrinsic::masked_load: {
1385 // Scalarize unsupported vector masked load
1386 if (!TTI->isLegalMaskedLoad(CI->getType(), 1)) {
1387 ScalarizeMaskedLoad(CI);
1393 case Intrinsic::masked_store: {
1394 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType(), 1)) {
1395 ScalarizeMaskedStore(CI);
1401 case Intrinsic::aarch64_stlxr:
1402 case Intrinsic::aarch64_stxr: {
1403 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1404 if (!ExtVal || !ExtVal->hasOneUse() ||
1405 ExtVal->getParent() == CI->getParent())
1407 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1408 ExtVal->moveBefore(CI);
1414 // Unknown address space.
1415 // TODO: Target hook to pick which address space the intrinsic cares
1417 unsigned AddrSpace = ~0u;
1418 SmallVector<Value*, 2> PtrOps;
1420 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
1421 while (!PtrOps.empty())
1422 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
1427 // From here on out we're working with named functions.
1428 if (!CI->getCalledFunction()) return false;
1430 // Lower all default uses of _chk calls. This is very similar
1431 // to what InstCombineCalls does, but here we are only lowering calls
1432 // to fortified library functions (e.g. __memcpy_chk) that have the default
1433 // "don't know" as the objectsize. Anything else should be left alone.
1434 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1435 if (Value *V = Simplifier.optimizeCall(CI)) {
1436 CI->replaceAllUsesWith(V);
1437 CI->eraseFromParent();
1443 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
1444 /// instructions to the predecessor to enable tail call optimizations. The
1445 /// case it is currently looking for is:
1448 /// %tmp0 = tail call i32 @f0()
1449 /// br label %return
1451 /// %tmp1 = tail call i32 @f1()
1452 /// br label %return
1454 /// %tmp2 = tail call i32 @f2()
1455 /// br label %return
1457 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1465 /// %tmp0 = tail call i32 @f0()
1468 /// %tmp1 = tail call i32 @f1()
1471 /// %tmp2 = tail call i32 @f2()
1474 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
1478 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1482 PHINode *PN = nullptr;
1483 BitCastInst *BCI = nullptr;
1484 Value *V = RI->getReturnValue();
1486 BCI = dyn_cast<BitCastInst>(V);
1488 V = BCI->getOperand(0);
1490 PN = dyn_cast<PHINode>(V);
1495 if (PN && PN->getParent() != BB)
1498 // It's not safe to eliminate the sign / zero extension of the return value.
1499 // See llvm::isInTailCallPosition().
1500 const Function *F = BB->getParent();
1501 AttributeSet CallerAttrs = F->getAttributes();
1502 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1503 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1506 // Make sure there are no instructions between the PHI and return, or that the
1507 // return is the first instruction in the block.
1509 BasicBlock::iterator BI = BB->begin();
1510 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1512 // Also skip over the bitcast.
1517 BasicBlock::iterator BI = BB->begin();
1518 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1523 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1525 SmallVector<CallInst*, 4> TailCalls;
1527 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1528 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1529 // Make sure the phi value is indeed produced by the tail call.
1530 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1531 TLI->mayBeEmittedAsTailCall(CI))
1532 TailCalls.push_back(CI);
1535 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1536 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1537 if (!VisitedBBs.insert(*PI).second)
1540 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1541 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1542 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1543 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1547 CallInst *CI = dyn_cast<CallInst>(&*RI);
1548 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1549 TailCalls.push_back(CI);
1553 bool Changed = false;
1554 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1555 CallInst *CI = TailCalls[i];
1558 // Conservatively require the attributes of the call to match those of the
1559 // return. Ignore noalias because it doesn't affect the call sequence.
1560 AttributeSet CalleeAttrs = CS.getAttributes();
1561 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1562 removeAttribute(Attribute::NoAlias) !=
1563 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1564 removeAttribute(Attribute::NoAlias))
1567 // Make sure the call instruction is followed by an unconditional branch to
1568 // the return block.
1569 BasicBlock *CallBB = CI->getParent();
1570 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1571 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1574 // Duplicate the return into CallBB.
1575 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1576 ModifiedDT = Changed = true;
1580 // If we eliminated all predecessors of the block, delete the block now.
1581 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1582 BB->eraseFromParent();
1587 //===----------------------------------------------------------------------===//
1588 // Memory Optimization
1589 //===----------------------------------------------------------------------===//
1593 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1594 /// which holds actual Value*'s for register values.
1595 struct ExtAddrMode : public TargetLowering::AddrMode {
1598 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1599 void print(raw_ostream &OS) const;
1602 bool operator==(const ExtAddrMode& O) const {
1603 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1604 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1605 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1610 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1616 void ExtAddrMode::print(raw_ostream &OS) const {
1617 bool NeedPlus = false;
1620 OS << (NeedPlus ? " + " : "")
1622 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1627 OS << (NeedPlus ? " + " : "")
1633 OS << (NeedPlus ? " + " : "")
1635 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1639 OS << (NeedPlus ? " + " : "")
1641 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1647 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1648 void ExtAddrMode::dump() const {
1654 /// \brief This class provides transaction based operation on the IR.
1655 /// Every change made through this class is recorded in the internal state and
1656 /// can be undone (rollback) until commit is called.
1657 class TypePromotionTransaction {
1659 /// \brief This represents the common interface of the individual transaction.
1660 /// Each class implements the logic for doing one specific modification on
1661 /// the IR via the TypePromotionTransaction.
1662 class TypePromotionAction {
1664 /// The Instruction modified.
1668 /// \brief Constructor of the action.
1669 /// The constructor performs the related action on the IR.
1670 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1672 virtual ~TypePromotionAction() {}
1674 /// \brief Undo the modification done by this action.
1675 /// When this method is called, the IR must be in the same state as it was
1676 /// before this action was applied.
1677 /// \pre Undoing the action works if and only if the IR is in the exact same
1678 /// state as it was directly after this action was applied.
1679 virtual void undo() = 0;
1681 /// \brief Advocate every change made by this action.
1682 /// When the results on the IR of the action are to be kept, it is important
1683 /// to call this function, otherwise hidden information may be kept forever.
1684 virtual void commit() {
1685 // Nothing to be done, this action is not doing anything.
1689 /// \brief Utility to remember the position of an instruction.
1690 class InsertionHandler {
1691 /// Position of an instruction.
1692 /// Either an instruction:
1693 /// - Is the first in a basic block: BB is used.
1694 /// - Has a previous instructon: PrevInst is used.
1696 Instruction *PrevInst;
1699 /// Remember whether or not the instruction had a previous instruction.
1700 bool HasPrevInstruction;
1703 /// \brief Record the position of \p Inst.
1704 InsertionHandler(Instruction *Inst) {
1705 BasicBlock::iterator It = Inst;
1706 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1707 if (HasPrevInstruction)
1708 Point.PrevInst = --It;
1710 Point.BB = Inst->getParent();
1713 /// \brief Insert \p Inst at the recorded position.
1714 void insert(Instruction *Inst) {
1715 if (HasPrevInstruction) {
1716 if (Inst->getParent())
1717 Inst->removeFromParent();
1718 Inst->insertAfter(Point.PrevInst);
1720 Instruction *Position = Point.BB->getFirstInsertionPt();
1721 if (Inst->getParent())
1722 Inst->moveBefore(Position);
1724 Inst->insertBefore(Position);
1729 /// \brief Move an instruction before another.
1730 class InstructionMoveBefore : public TypePromotionAction {
1731 /// Original position of the instruction.
1732 InsertionHandler Position;
1735 /// \brief Move \p Inst before \p Before.
1736 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1737 : TypePromotionAction(Inst), Position(Inst) {
1738 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1739 Inst->moveBefore(Before);
1742 /// \brief Move the instruction back to its original position.
1743 void undo() override {
1744 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1745 Position.insert(Inst);
1749 /// \brief Set the operand of an instruction with a new value.
1750 class OperandSetter : public TypePromotionAction {
1751 /// Original operand of the instruction.
1753 /// Index of the modified instruction.
1757 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1758 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1759 : TypePromotionAction(Inst), Idx(Idx) {
1760 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1761 << "for:" << *Inst << "\n"
1762 << "with:" << *NewVal << "\n");
1763 Origin = Inst->getOperand(Idx);
1764 Inst->setOperand(Idx, NewVal);
1767 /// \brief Restore the original value of the instruction.
1768 void undo() override {
1769 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1770 << "for: " << *Inst << "\n"
1771 << "with: " << *Origin << "\n");
1772 Inst->setOperand(Idx, Origin);
1776 /// \brief Hide the operands of an instruction.
1777 /// Do as if this instruction was not using any of its operands.
1778 class OperandsHider : public TypePromotionAction {
1779 /// The list of original operands.
1780 SmallVector<Value *, 4> OriginalValues;
1783 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1784 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1785 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1786 unsigned NumOpnds = Inst->getNumOperands();
1787 OriginalValues.reserve(NumOpnds);
1788 for (unsigned It = 0; It < NumOpnds; ++It) {
1789 // Save the current operand.
1790 Value *Val = Inst->getOperand(It);
1791 OriginalValues.push_back(Val);
1793 // We could use OperandSetter here, but that would implied an overhead
1794 // that we are not willing to pay.
1795 Inst->setOperand(It, UndefValue::get(Val->getType()));
1799 /// \brief Restore the original list of uses.
1800 void undo() override {
1801 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1802 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1803 Inst->setOperand(It, OriginalValues[It]);
1807 /// \brief Build a truncate instruction.
1808 class TruncBuilder : public TypePromotionAction {
1811 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1813 /// trunc Opnd to Ty.
1814 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1815 IRBuilder<> Builder(Opnd);
1816 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
1817 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
1820 /// \brief Get the built value.
1821 Value *getBuiltValue() { return Val; }
1823 /// \brief Remove the built instruction.
1824 void undo() override {
1825 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
1826 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1827 IVal->eraseFromParent();
1831 /// \brief Build a sign extension instruction.
1832 class SExtBuilder : public TypePromotionAction {
1835 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1837 /// sext Opnd to Ty.
1838 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1839 : TypePromotionAction(InsertPt) {
1840 IRBuilder<> Builder(InsertPt);
1841 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
1842 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
1845 /// \brief Get the built value.
1846 Value *getBuiltValue() { return Val; }
1848 /// \brief Remove the built instruction.
1849 void undo() override {
1850 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
1851 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1852 IVal->eraseFromParent();
1856 /// \brief Build a zero extension instruction.
1857 class ZExtBuilder : public TypePromotionAction {
1860 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
1862 /// zext Opnd to Ty.
1863 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1864 : TypePromotionAction(InsertPt) {
1865 IRBuilder<> Builder(InsertPt);
1866 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
1867 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
1870 /// \brief Get the built value.
1871 Value *getBuiltValue() { return Val; }
1873 /// \brief Remove the built instruction.
1874 void undo() override {
1875 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
1876 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1877 IVal->eraseFromParent();
1881 /// \brief Mutate an instruction to another type.
1882 class TypeMutator : public TypePromotionAction {
1883 /// Record the original type.
1887 /// \brief Mutate the type of \p Inst into \p NewTy.
1888 TypeMutator(Instruction *Inst, Type *NewTy)
1889 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1890 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1892 Inst->mutateType(NewTy);
1895 /// \brief Mutate the instruction back to its original type.
1896 void undo() override {
1897 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1899 Inst->mutateType(OrigTy);
1903 /// \brief Replace the uses of an instruction by another instruction.
1904 class UsesReplacer : public TypePromotionAction {
1905 /// Helper structure to keep track of the replaced uses.
1906 struct InstructionAndIdx {
1907 /// The instruction using the instruction.
1909 /// The index where this instruction is used for Inst.
1911 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1912 : Inst(Inst), Idx(Idx) {}
1915 /// Keep track of the original uses (pair Instruction, Index).
1916 SmallVector<InstructionAndIdx, 4> OriginalUses;
1917 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1920 /// \brief Replace all the use of \p Inst by \p New.
1921 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1922 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1924 // Record the original uses.
1925 for (Use &U : Inst->uses()) {
1926 Instruction *UserI = cast<Instruction>(U.getUser());
1927 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1929 // Now, we can replace the uses.
1930 Inst->replaceAllUsesWith(New);
1933 /// \brief Reassign the original uses of Inst to Inst.
1934 void undo() override {
1935 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1936 for (use_iterator UseIt = OriginalUses.begin(),
1937 EndIt = OriginalUses.end();
1938 UseIt != EndIt; ++UseIt) {
1939 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1944 /// \brief Remove an instruction from the IR.
1945 class InstructionRemover : public TypePromotionAction {
1946 /// Original position of the instruction.
1947 InsertionHandler Inserter;
1948 /// Helper structure to hide all the link to the instruction. In other
1949 /// words, this helps to do as if the instruction was removed.
1950 OperandsHider Hider;
1951 /// Keep track of the uses replaced, if any.
1952 UsesReplacer *Replacer;
1955 /// \brief Remove all reference of \p Inst and optinally replace all its
1957 /// \pre If !Inst->use_empty(), then New != nullptr
1958 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1959 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1962 Replacer = new UsesReplacer(Inst, New);
1963 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1964 Inst->removeFromParent();
1967 ~InstructionRemover() override { delete Replacer; }
1969 /// \brief Really remove the instruction.
1970 void commit() override { delete Inst; }
1972 /// \brief Resurrect the instruction and reassign it to the proper uses if
1973 /// new value was provided when build this action.
1974 void undo() override {
1975 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1976 Inserter.insert(Inst);
1984 /// Restoration point.
1985 /// The restoration point is a pointer to an action instead of an iterator
1986 /// because the iterator may be invalidated but not the pointer.
1987 typedef const TypePromotionAction *ConstRestorationPt;
1988 /// Advocate every changes made in that transaction.
1990 /// Undo all the changes made after the given point.
1991 void rollback(ConstRestorationPt Point);
1992 /// Get the current restoration point.
1993 ConstRestorationPt getRestorationPoint() const;
1995 /// \name API for IR modification with state keeping to support rollback.
1997 /// Same as Instruction::setOperand.
1998 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1999 /// Same as Instruction::eraseFromParent.
2000 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2001 /// Same as Value::replaceAllUsesWith.
2002 void replaceAllUsesWith(Instruction *Inst, Value *New);
2003 /// Same as Value::mutateType.
2004 void mutateType(Instruction *Inst, Type *NewTy);
2005 /// Same as IRBuilder::createTrunc.
2006 Value *createTrunc(Instruction *Opnd, Type *Ty);
2007 /// Same as IRBuilder::createSExt.
2008 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2009 /// Same as IRBuilder::createZExt.
2010 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2011 /// Same as Instruction::moveBefore.
2012 void moveBefore(Instruction *Inst, Instruction *Before);
2016 /// The ordered list of actions made so far.
2017 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
2018 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
2021 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2024 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
2027 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2030 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
2033 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2035 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2038 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2039 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2042 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2044 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2045 Value *Val = Ptr->getBuiltValue();
2046 Actions.push_back(std::move(Ptr));
2050 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2051 Value *Opnd, Type *Ty) {
2052 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2053 Value *Val = Ptr->getBuiltValue();
2054 Actions.push_back(std::move(Ptr));
2058 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2059 Value *Opnd, Type *Ty) {
2060 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2061 Value *Val = Ptr->getBuiltValue();
2062 Actions.push_back(std::move(Ptr));
2066 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2067 Instruction *Before) {
2069 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
2072 TypePromotionTransaction::ConstRestorationPt
2073 TypePromotionTransaction::getRestorationPoint() const {
2074 return !Actions.empty() ? Actions.back().get() : nullptr;
2077 void TypePromotionTransaction::commit() {
2078 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2084 void TypePromotionTransaction::rollback(
2085 TypePromotionTransaction::ConstRestorationPt Point) {
2086 while (!Actions.empty() && Point != Actions.back().get()) {
2087 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2092 /// \brief A helper class for matching addressing modes.
2094 /// This encapsulates the logic for matching the target-legal addressing modes.
2095 class AddressingModeMatcher {
2096 SmallVectorImpl<Instruction*> &AddrModeInsts;
2097 const TargetMachine &TM;
2098 const TargetLowering &TLI;
2100 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2101 /// the memory instruction that we're computing this address for.
2104 Instruction *MemoryInst;
2106 /// AddrMode - This is the addressing mode that we're building up. This is
2107 /// part of the return value of this addressing mode matching stuff.
2108 ExtAddrMode &AddrMode;
2110 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
2111 const SetOfInstrs &InsertedTruncs;
2112 /// A map from the instructions to their type before promotion.
2113 InstrToOrigTy &PromotedInsts;
2114 /// The ongoing transaction where every action should be registered.
2115 TypePromotionTransaction &TPT;
2117 /// IgnoreProfitability - This is set to true when we should not do
2118 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
2119 /// always returns true.
2120 bool IgnoreProfitability;
2122 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2123 const TargetMachine &TM, Type *AT, unsigned AS,
2124 Instruction *MI, ExtAddrMode &AM,
2125 const SetOfInstrs &InsertedTruncs,
2126 InstrToOrigTy &PromotedInsts,
2127 TypePromotionTransaction &TPT)
2128 : AddrModeInsts(AMI), TM(TM),
2129 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
2130 ->getTargetLowering()),
2131 AccessTy(AT), AddrSpace(AS), MemoryInst(MI), AddrMode(AM),
2132 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
2133 IgnoreProfitability = false;
2137 /// Match - Find the maximal addressing mode that a load/store of V can fold,
2138 /// give an access type of AccessTy. This returns a list of involved
2139 /// instructions in AddrModeInsts.
2140 /// \p InsertedTruncs The truncate instruction inserted by other
2143 /// \p PromotedInsts maps the instructions to their type before promotion.
2144 /// \p The ongoing transaction where every action should be registered.
2145 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2146 Instruction *MemoryInst,
2147 SmallVectorImpl<Instruction*> &AddrModeInsts,
2148 const TargetMachine &TM,
2149 const SetOfInstrs &InsertedTruncs,
2150 InstrToOrigTy &PromotedInsts,
2151 TypePromotionTransaction &TPT) {
2154 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
2155 MemoryInst, Result, InsertedTruncs,
2156 PromotedInsts, TPT).MatchAddr(V, 0);
2157 (void)Success; assert(Success && "Couldn't select *anything*?");
2161 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2162 bool MatchAddr(Value *V, unsigned Depth);
2163 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2164 bool *MovedAway = nullptr);
2165 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
2166 ExtAddrMode &AMBefore,
2167 ExtAddrMode &AMAfter);
2168 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2169 bool IsPromotionProfitable(unsigned NewCost, unsigned OldCost,
2170 Value *PromotedOperand) const;
2173 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
2174 /// Return true and update AddrMode if this addr mode is legal for the target,
2176 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
2178 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2179 // mode. Just process that directly.
2181 return MatchAddr(ScaleReg, Depth);
2183 // If the scale is 0, it takes nothing to add this.
2187 // If we already have a scale of this value, we can add to it, otherwise, we
2188 // need an available scale field.
2189 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2192 ExtAddrMode TestAddrMode = AddrMode;
2194 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2195 // [A+B + A*7] -> [B+A*8].
2196 TestAddrMode.Scale += Scale;
2197 TestAddrMode.ScaledReg = ScaleReg;
2199 // If the new address isn't legal, bail out.
2200 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy, AddrSpace))
2203 // It was legal, so commit it.
2204 AddrMode = TestAddrMode;
2206 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2207 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2208 // X*Scale + C*Scale to addr mode.
2209 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2210 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2211 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2212 TestAddrMode.ScaledReg = AddLHS;
2213 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2215 // If this addressing mode is legal, commit it and remember that we folded
2216 // this instruction.
2217 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy, AddrSpace)) {
2218 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2219 AddrMode = TestAddrMode;
2224 // Otherwise, not (x+c)*scale, just return what we have.
2228 /// MightBeFoldableInst - This is a little filter, which returns true if an
2229 /// addressing computation involving I might be folded into a load/store
2230 /// accessing it. This doesn't need to be perfect, but needs to accept at least
2231 /// the set of instructions that MatchOperationAddr can.
2232 static bool MightBeFoldableInst(Instruction *I) {
2233 switch (I->getOpcode()) {
2234 case Instruction::BitCast:
2235 case Instruction::AddrSpaceCast:
2236 // Don't touch identity bitcasts.
2237 if (I->getType() == I->getOperand(0)->getType())
2239 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2240 case Instruction::PtrToInt:
2241 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2243 case Instruction::IntToPtr:
2244 // We know the input is intptr_t, so this is foldable.
2246 case Instruction::Add:
2248 case Instruction::Mul:
2249 case Instruction::Shl:
2250 // Can only handle X*C and X << C.
2251 return isa<ConstantInt>(I->getOperand(1));
2252 case Instruction::GetElementPtr:
2259 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2260 /// \note \p Val is assumed to be the product of some type promotion.
2261 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2262 /// to be legal, as the non-promoted value would have had the same state.
2263 static bool isPromotedInstructionLegal(const TargetLowering &TLI, Value *Val) {
2264 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2267 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2268 // If the ISDOpcode is undefined, it was undefined before the promotion.
2271 // Otherwise, check if the promoted instruction is legal or not.
2272 return TLI.isOperationLegalOrCustom(
2273 ISDOpcode, TLI.getValueType(PromotedInst->getType()));
2276 /// \brief Hepler class to perform type promotion.
2277 class TypePromotionHelper {
2278 /// \brief Utility function to check whether or not a sign or zero extension
2279 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2280 /// either using the operands of \p Inst or promoting \p Inst.
2281 /// The type of the extension is defined by \p IsSExt.
2282 /// In other words, check if:
2283 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2284 /// #1 Promotion applies:
2285 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2286 /// #2 Operand reuses:
2287 /// ext opnd1 to ConsideredExtType.
2288 /// \p PromotedInsts maps the instructions to their type before promotion.
2289 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2290 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2292 /// \brief Utility function to determine if \p OpIdx should be promoted when
2293 /// promoting \p Inst.
2294 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2295 if (isa<SelectInst>(Inst) && OpIdx == 0)
2300 /// \brief Utility function to promote the operand of \p Ext when this
2301 /// operand is a promotable trunc or sext or zext.
2302 /// \p PromotedInsts maps the instructions to their type before promotion.
2303 /// \p CreatedInstsCost[out] contains the cost of all instructions
2304 /// created to promote the operand of Ext.
2305 /// Newly added extensions are inserted in \p Exts.
2306 /// Newly added truncates are inserted in \p Truncs.
2307 /// Should never be called directly.
2308 /// \return The promoted value which is used instead of Ext.
2309 static Value *promoteOperandForTruncAndAnyExt(
2310 Instruction *Ext, TypePromotionTransaction &TPT,
2311 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2312 SmallVectorImpl<Instruction *> *Exts,
2313 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
2315 /// \brief Utility function to promote the operand of \p Ext when this
2316 /// operand is promotable and is not a supported trunc or sext.
2317 /// \p PromotedInsts maps the instructions to their type before promotion.
2318 /// \p CreatedInstsCost[out] contains the cost of all the instructions
2319 /// created to promote the operand of Ext.
2320 /// Newly added extensions are inserted in \p Exts.
2321 /// Newly added truncates are inserted in \p Truncs.
2322 /// Should never be called directly.
2323 /// \return The promoted value which is used instead of Ext.
2324 static Value *promoteOperandForOther(Instruction *Ext,
2325 TypePromotionTransaction &TPT,
2326 InstrToOrigTy &PromotedInsts,
2327 unsigned &CreatedInstsCost,
2328 SmallVectorImpl<Instruction *> *Exts,
2329 SmallVectorImpl<Instruction *> *Truncs,
2330 const TargetLowering &TLI, bool IsSExt);
2332 /// \see promoteOperandForOther.
2333 static Value *signExtendOperandForOther(
2334 Instruction *Ext, TypePromotionTransaction &TPT,
2335 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2336 SmallVectorImpl<Instruction *> *Exts,
2337 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2338 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2339 Exts, Truncs, TLI, true);
2342 /// \see promoteOperandForOther.
2343 static Value *zeroExtendOperandForOther(
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, false);
2353 /// Type for the utility function that promotes the operand of Ext.
2354 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2355 InstrToOrigTy &PromotedInsts,
2356 unsigned &CreatedInstsCost,
2357 SmallVectorImpl<Instruction *> *Exts,
2358 SmallVectorImpl<Instruction *> *Truncs,
2359 const TargetLowering &TLI);
2360 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2361 /// action to promote the operand of \p Ext instead of using Ext.
2362 /// \return NULL if no promotable action is possible with the current
2364 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
2365 /// the others CodeGenPrepare optimizations. This information is important
2366 /// because we do not want to promote these instructions as CodeGenPrepare
2367 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2368 /// \p PromotedInsts maps the instructions to their type before promotion.
2369 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2370 const TargetLowering &TLI,
2371 const InstrToOrigTy &PromotedInsts);
2374 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2375 Type *ConsideredExtType,
2376 const InstrToOrigTy &PromotedInsts,
2378 // The promotion helper does not know how to deal with vector types yet.
2379 // To be able to fix that, we would need to fix the places where we
2380 // statically extend, e.g., constants and such.
2381 if (Inst->getType()->isVectorTy())
2384 // We can always get through zext.
2385 if (isa<ZExtInst>(Inst))
2388 // sext(sext) is ok too.
2389 if (IsSExt && isa<SExtInst>(Inst))
2392 // We can get through binary operator, if it is legal. In other words, the
2393 // binary operator must have a nuw or nsw flag.
2394 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2395 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2396 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2397 (IsSExt && BinOp->hasNoSignedWrap())))
2400 // Check if we can do the following simplification.
2401 // ext(trunc(opnd)) --> ext(opnd)
2402 if (!isa<TruncInst>(Inst))
2405 Value *OpndVal = Inst->getOperand(0);
2406 // Check if we can use this operand in the extension.
2407 // If the type is larger than the result type of the extension,
2409 if (!OpndVal->getType()->isIntegerTy() ||
2410 OpndVal->getType()->getIntegerBitWidth() >
2411 ConsideredExtType->getIntegerBitWidth())
2414 // If the operand of the truncate is not an instruction, we will not have
2415 // any information on the dropped bits.
2416 // (Actually we could for constant but it is not worth the extra logic).
2417 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2421 // Check if the source of the type is narrow enough.
2422 // I.e., check that trunc just drops extended bits of the same kind of
2424 // #1 get the type of the operand and check the kind of the extended bits.
2425 const Type *OpndType;
2426 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2427 if (It != PromotedInsts.end() && It->second.IsSExt == IsSExt)
2428 OpndType = It->second.Ty;
2429 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2430 OpndType = Opnd->getOperand(0)->getType();
2434 // #2 check that the truncate just drop extended bits.
2435 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
2441 TypePromotionHelper::Action TypePromotionHelper::getAction(
2442 Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2443 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2444 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2445 "Unexpected instruction type");
2446 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2447 Type *ExtTy = Ext->getType();
2448 bool IsSExt = isa<SExtInst>(Ext);
2449 // If the operand of the extension is not an instruction, we cannot
2451 // If it, check we can get through.
2452 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2455 // Do not promote if the operand has been added by codegenprepare.
2456 // Otherwise, it means we are undoing an optimization that is likely to be
2457 // redone, thus causing potential infinite loop.
2458 if (isa<TruncInst>(ExtOpnd) && InsertedTruncs.count(ExtOpnd))
2461 // SExt or Trunc instructions.
2462 // Return the related handler.
2463 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2464 isa<ZExtInst>(ExtOpnd))
2465 return promoteOperandForTruncAndAnyExt;
2467 // Regular instruction.
2468 // Abort early if we will have to insert non-free instructions.
2469 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2471 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2474 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2475 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2476 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2477 SmallVectorImpl<Instruction *> *Exts,
2478 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2479 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2480 // get through it and this method should not be called.
2481 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2482 Value *ExtVal = SExt;
2483 bool HasMergedNonFreeExt = false;
2484 if (isa<ZExtInst>(SExtOpnd)) {
2485 // Replace s|zext(zext(opnd))
2487 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
2489 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2490 TPT.replaceAllUsesWith(SExt, ZExt);
2491 TPT.eraseInstruction(SExt);
2494 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2496 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2498 CreatedInstsCost = 0;
2500 // Remove dead code.
2501 if (SExtOpnd->use_empty())
2502 TPT.eraseInstruction(SExtOpnd);
2504 // Check if the extension is still needed.
2505 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2506 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2509 Exts->push_back(ExtInst);
2510 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
2515 // At this point we have: ext ty opnd to ty.
2516 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2517 Value *NextVal = ExtInst->getOperand(0);
2518 TPT.eraseInstruction(ExtInst, NextVal);
2522 Value *TypePromotionHelper::promoteOperandForOther(
2523 Instruction *Ext, TypePromotionTransaction &TPT,
2524 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2525 SmallVectorImpl<Instruction *> *Exts,
2526 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
2528 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2529 // get through it and this method should not be called.
2530 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2531 CreatedInstsCost = 0;
2532 if (!ExtOpnd->hasOneUse()) {
2533 // ExtOpnd will be promoted.
2534 // All its uses, but Ext, will need to use a truncated value of the
2535 // promoted version.
2536 // Create the truncate now.
2537 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2538 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2539 ITrunc->removeFromParent();
2540 // Insert it just after the definition.
2541 ITrunc->insertAfter(ExtOpnd);
2543 Truncs->push_back(ITrunc);
2546 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2547 // Restore the operand of Ext (which has been replace by the previous call
2548 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2549 TPT.setOperand(Ext, 0, ExtOpnd);
2552 // Get through the Instruction:
2553 // 1. Update its type.
2554 // 2. Replace the uses of Ext by Inst.
2555 // 3. Extend each operand that needs to be extended.
2557 // Remember the original type of the instruction before promotion.
2558 // This is useful to know that the high bits are sign extended bits.
2559 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2560 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2562 TPT.mutateType(ExtOpnd, Ext->getType());
2564 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2566 Instruction *ExtForOpnd = Ext;
2568 DEBUG(dbgs() << "Propagate Ext to operands\n");
2569 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2571 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2572 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2573 !shouldExtOperand(ExtOpnd, OpIdx)) {
2574 DEBUG(dbgs() << "No need to propagate\n");
2577 // Check if we can statically extend the operand.
2578 Value *Opnd = ExtOpnd->getOperand(OpIdx);
2579 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2580 DEBUG(dbgs() << "Statically extend\n");
2581 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2582 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
2583 : Cst->getValue().zext(BitWidth);
2584 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
2587 // UndefValue are typed, so we have to statically sign extend them.
2588 if (isa<UndefValue>(Opnd)) {
2589 DEBUG(dbgs() << "Statically extend\n");
2590 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2594 // Otherwise we have to explicity sign extend the operand.
2595 // Check if Ext was reused to extend an operand.
2597 // If yes, create a new one.
2598 DEBUG(dbgs() << "More operands to ext\n");
2599 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2600 : TPT.createZExt(Ext, Opnd, Ext->getType());
2601 if (!isa<Instruction>(ValForExtOpnd)) {
2602 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
2605 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
2608 Exts->push_back(ExtForOpnd);
2609 TPT.setOperand(ExtForOpnd, 0, Opnd);
2611 // Move the sign extension before the insertion point.
2612 TPT.moveBefore(ExtForOpnd, ExtOpnd);
2613 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2614 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
2615 // If more sext are required, new instructions will have to be created.
2616 ExtForOpnd = nullptr;
2618 if (ExtForOpnd == Ext) {
2619 DEBUG(dbgs() << "Extension is useless now\n");
2620 TPT.eraseInstruction(Ext);
2625 /// IsPromotionProfitable - Check whether or not promoting an instruction
2626 /// to a wider type was profitable.
2627 /// \p NewCost gives the cost of extension instructions created by the
2629 /// \p OldCost gives the cost of extension instructions before the promotion
2630 /// plus the number of instructions that have been
2631 /// matched in the addressing mode the promotion.
2632 /// \p PromotedOperand is the value that has been promoted.
2633 /// \return True if the promotion is profitable, false otherwise.
2634 bool AddressingModeMatcher::IsPromotionProfitable(
2635 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
2636 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
2637 // The cost of the new extensions is greater than the cost of the
2638 // old extension plus what we folded.
2639 // This is not profitable.
2640 if (NewCost > OldCost)
2642 if (NewCost < OldCost)
2644 // The promotion is neutral but it may help folding the sign extension in
2645 // loads for instance.
2646 // Check that we did not create an illegal instruction.
2647 return isPromotedInstructionLegal(TLI, PromotedOperand);
2650 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
2651 /// fold the operation into the addressing mode. If so, update the addressing
2652 /// mode and return true, otherwise return false without modifying AddrMode.
2653 /// If \p MovedAway is not NULL, it contains the information of whether or
2654 /// not AddrInst has to be folded into the addressing mode on success.
2655 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2656 /// because it has been moved away.
2657 /// Thus AddrInst must not be added in the matched instructions.
2658 /// This state can happen when AddrInst is a sext, since it may be moved away.
2659 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2660 /// not be referenced anymore.
2661 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
2664 // Avoid exponential behavior on extremely deep expression trees.
2665 if (Depth >= 5) return false;
2667 // By default, all matched instructions stay in place.
2672 case Instruction::PtrToInt:
2673 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2674 return MatchAddr(AddrInst->getOperand(0), Depth);
2675 case Instruction::IntToPtr:
2676 // This inttoptr is a no-op if the integer type is pointer sized.
2677 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
2678 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
2679 return MatchAddr(AddrInst->getOperand(0), Depth);
2681 case Instruction::BitCast:
2682 // BitCast is always a noop, and we can handle it as long as it is
2683 // int->int or pointer->pointer (we don't want int<->fp or something).
2684 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2685 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2686 // Don't touch identity bitcasts. These were probably put here by LSR,
2687 // and we don't want to mess around with them. Assume it knows what it
2689 AddrInst->getOperand(0)->getType() != AddrInst->getType())
2690 return MatchAddr(AddrInst->getOperand(0), Depth);
2692 case Instruction::AddrSpaceCast: {
2694 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
2695 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
2696 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
2697 return MatchAddr(AddrInst->getOperand(0), Depth);
2700 case Instruction::Add: {
2701 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2702 ExtAddrMode BackupAddrMode = AddrMode;
2703 unsigned OldSize = AddrModeInsts.size();
2704 // Start a transaction at this point.
2705 // The LHS may match but not the RHS.
2706 // Therefore, we need a higher level restoration point to undo partially
2707 // matched operation.
2708 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2709 TPT.getRestorationPoint();
2711 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2712 MatchAddr(AddrInst->getOperand(0), Depth+1))
2715 // Restore the old addr mode info.
2716 AddrMode = BackupAddrMode;
2717 AddrModeInsts.resize(OldSize);
2718 TPT.rollback(LastKnownGood);
2720 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2721 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2722 MatchAddr(AddrInst->getOperand(1), Depth+1))
2725 // Otherwise we definitely can't merge the ADD in.
2726 AddrMode = BackupAddrMode;
2727 AddrModeInsts.resize(OldSize);
2728 TPT.rollback(LastKnownGood);
2731 //case Instruction::Or:
2732 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2734 case Instruction::Mul:
2735 case Instruction::Shl: {
2736 // Can only handle X*C and X << C.
2737 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2740 int64_t Scale = RHS->getSExtValue();
2741 if (Opcode == Instruction::Shl)
2742 Scale = 1LL << Scale;
2744 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2746 case Instruction::GetElementPtr: {
2747 // Scan the GEP. We check it if it contains constant offsets and at most
2748 // one variable offset.
2749 int VariableOperand = -1;
2750 unsigned VariableScale = 0;
2752 int64_t ConstantOffset = 0;
2753 const DataLayout *TD = TLI.getDataLayout();
2754 gep_type_iterator GTI = gep_type_begin(AddrInst);
2755 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2756 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2757 const StructLayout *SL = TD->getStructLayout(STy);
2759 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2760 ConstantOffset += SL->getElementOffset(Idx);
2762 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
2763 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2764 ConstantOffset += CI->getSExtValue()*TypeSize;
2765 } else if (TypeSize) { // Scales of zero don't do anything.
2766 // We only allow one variable index at the moment.
2767 if (VariableOperand != -1)
2770 // Remember the variable index.
2771 VariableOperand = i;
2772 VariableScale = TypeSize;
2777 // A common case is for the GEP to only do a constant offset. In this case,
2778 // just add it to the disp field and check validity.
2779 if (VariableOperand == -1) {
2780 AddrMode.BaseOffs += ConstantOffset;
2781 if (ConstantOffset == 0 ||
2782 TLI.isLegalAddressingMode(AddrMode, AccessTy, AddrSpace)) {
2783 // Check to see if we can fold the base pointer in too.
2784 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2787 AddrMode.BaseOffs -= ConstantOffset;
2791 // Save the valid addressing mode in case we can't match.
2792 ExtAddrMode BackupAddrMode = AddrMode;
2793 unsigned OldSize = AddrModeInsts.size();
2795 // See if the scale and offset amount is valid for this target.
2796 AddrMode.BaseOffs += ConstantOffset;
2798 // Match the base operand of the GEP.
2799 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2800 // If it couldn't be matched, just stuff the value in a register.
2801 if (AddrMode.HasBaseReg) {
2802 AddrMode = BackupAddrMode;
2803 AddrModeInsts.resize(OldSize);
2806 AddrMode.HasBaseReg = true;
2807 AddrMode.BaseReg = AddrInst->getOperand(0);
2810 // Match the remaining variable portion of the GEP.
2811 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2813 // If it couldn't be matched, try stuffing the base into a register
2814 // instead of matching it, and retrying the match of the scale.
2815 AddrMode = BackupAddrMode;
2816 AddrModeInsts.resize(OldSize);
2817 if (AddrMode.HasBaseReg)
2819 AddrMode.HasBaseReg = true;
2820 AddrMode.BaseReg = AddrInst->getOperand(0);
2821 AddrMode.BaseOffs += ConstantOffset;
2822 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2823 VariableScale, Depth)) {
2824 // If even that didn't work, bail.
2825 AddrMode = BackupAddrMode;
2826 AddrModeInsts.resize(OldSize);
2833 case Instruction::SExt:
2834 case Instruction::ZExt: {
2835 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
2839 // Try to move this ext out of the way of the addressing mode.
2840 // Ask for a method for doing so.
2841 TypePromotionHelper::Action TPH =
2842 TypePromotionHelper::getAction(Ext, InsertedTruncs, TLI, PromotedInsts);
2846 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2847 TPT.getRestorationPoint();
2848 unsigned CreatedInstsCost = 0;
2849 unsigned ExtCost = !TLI.isExtFree(Ext);
2850 Value *PromotedOperand =
2851 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
2852 // SExt has been moved away.
2853 // Thus either it will be rematched later in the recursive calls or it is
2854 // gone. Anyway, we must not fold it into the addressing mode at this point.
2858 // addr = gep base, idx
2860 // promotedOpnd = ext opnd <- no match here
2861 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2862 // addr = gep base, op <- match
2866 assert(PromotedOperand &&
2867 "TypePromotionHelper should have filtered out those cases");
2869 ExtAddrMode BackupAddrMode = AddrMode;
2870 unsigned OldSize = AddrModeInsts.size();
2872 if (!MatchAddr(PromotedOperand, Depth) ||
2873 // The total of the new cost is equals to the cost of the created
2875 // The total of the old cost is equals to the cost of the extension plus
2876 // what we have saved in the addressing mode.
2877 !IsPromotionProfitable(CreatedInstsCost,
2878 ExtCost + (AddrModeInsts.size() - OldSize),
2880 AddrMode = BackupAddrMode;
2881 AddrModeInsts.resize(OldSize);
2882 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2883 TPT.rollback(LastKnownGood);
2892 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2893 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2894 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2895 /// or intptr_t for the target.
2897 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2898 // Start a transaction at this point that we will rollback if the matching
2900 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2901 TPT.getRestorationPoint();
2902 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2903 // Fold in immediates if legal for the target.
2904 AddrMode.BaseOffs += CI->getSExtValue();
2905 if (TLI.isLegalAddressingMode(AddrMode, AccessTy, AddrSpace))
2907 AddrMode.BaseOffs -= CI->getSExtValue();
2908 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2909 // If this is a global variable, try to fold it into the addressing mode.
2910 if (!AddrMode.BaseGV) {
2911 AddrMode.BaseGV = GV;
2912 if (TLI.isLegalAddressingMode(AddrMode, AccessTy, AddrSpace))
2914 AddrMode.BaseGV = nullptr;
2916 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2917 ExtAddrMode BackupAddrMode = AddrMode;
2918 unsigned OldSize = AddrModeInsts.size();
2920 // Check to see if it is possible to fold this operation.
2921 bool MovedAway = false;
2922 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2923 // This instruction may have been move away. If so, there is nothing
2927 // Okay, it's possible to fold this. Check to see if it is actually
2928 // *profitable* to do so. We use a simple cost model to avoid increasing
2929 // register pressure too much.
2930 if (I->hasOneUse() ||
2931 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2932 AddrModeInsts.push_back(I);
2936 // It isn't profitable to do this, roll back.
2937 //cerr << "NOT FOLDING: " << *I;
2938 AddrMode = BackupAddrMode;
2939 AddrModeInsts.resize(OldSize);
2940 TPT.rollback(LastKnownGood);
2942 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2943 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2945 TPT.rollback(LastKnownGood);
2946 } else if (isa<ConstantPointerNull>(Addr)) {
2947 // Null pointer gets folded without affecting the addressing mode.
2951 // Worse case, the target should support [reg] addressing modes. :)
2952 if (!AddrMode.HasBaseReg) {
2953 AddrMode.HasBaseReg = true;
2954 AddrMode.BaseReg = Addr;
2955 // Still check for legality in case the target supports [imm] but not [i+r].
2956 if (TLI.isLegalAddressingMode(AddrMode, AccessTy, AddrSpace))
2958 AddrMode.HasBaseReg = false;
2959 AddrMode.BaseReg = nullptr;
2962 // If the base register is already taken, see if we can do [r+r].
2963 if (AddrMode.Scale == 0) {
2965 AddrMode.ScaledReg = Addr;
2966 if (TLI.isLegalAddressingMode(AddrMode, AccessTy, AddrSpace))
2969 AddrMode.ScaledReg = nullptr;
2972 TPT.rollback(LastKnownGood);
2976 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2977 /// inline asm call are due to memory operands. If so, return true, otherwise
2979 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2980 const TargetMachine &TM) {
2981 const Function *F = CI->getParent()->getParent();
2982 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
2983 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
2984 TargetLowering::AsmOperandInfoVector TargetConstraints =
2985 TLI->ParseConstraints(TRI, ImmutableCallSite(CI));
2986 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2987 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2989 // Compute the constraint code and ConstraintType to use.
2990 TLI->ComputeConstraintToUse(OpInfo, SDValue());
2992 // If this asm operand is our Value*, and if it isn't an indirect memory
2993 // operand, we can't fold it!
2994 if (OpInfo.CallOperandVal == OpVal &&
2995 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2996 !OpInfo.isIndirect))
3003 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
3004 /// memory use. If we find an obviously non-foldable instruction, return true.
3005 /// Add the ultimately found memory instructions to MemoryUses.
3006 static bool FindAllMemoryUses(
3008 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
3009 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
3010 // If we already considered this instruction, we're done.
3011 if (!ConsideredInsts.insert(I).second)
3014 // If this is an obviously unfoldable instruction, bail out.
3015 if (!MightBeFoldableInst(I))
3018 // Loop over all the uses, recursively processing them.
3019 for (Use &U : I->uses()) {
3020 Instruction *UserI = cast<Instruction>(U.getUser());
3022 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
3023 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
3027 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
3028 unsigned opNo = U.getOperandNo();
3029 if (opNo == 0) return true; // Storing addr, not into addr.
3030 MemoryUses.push_back(std::make_pair(SI, opNo));
3034 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
3035 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
3036 if (!IA) return true;
3038 // If this is a memory operand, we're cool, otherwise bail out.
3039 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
3044 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
3051 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
3052 /// the use site that we're folding it into. If so, there is no cost to
3053 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
3054 /// that we know are live at the instruction already.
3055 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
3056 Value *KnownLive2) {
3057 // If Val is either of the known-live values, we know it is live!
3058 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
3061 // All values other than instructions and arguments (e.g. constants) are live.
3062 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
3064 // If Val is a constant sized alloca in the entry block, it is live, this is
3065 // true because it is just a reference to the stack/frame pointer, which is
3066 // live for the whole function.
3067 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
3068 if (AI->isStaticAlloca())
3071 // Check to see if this value is already used in the memory instruction's
3072 // block. If so, it's already live into the block at the very least, so we
3073 // can reasonably fold it.
3074 return Val->isUsedInBasicBlock(MemoryInst->getParent());
3077 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
3078 /// mode of the machine to fold the specified instruction into a load or store
3079 /// that ultimately uses it. However, the specified instruction has multiple
3080 /// uses. Given this, it may actually increase register pressure to fold it
3081 /// into the load. For example, consider this code:
3085 /// use(Y) -> nonload/store
3089 /// In this case, Y has multiple uses, and can be folded into the load of Z
3090 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
3091 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
3092 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
3093 /// number of computations either.
3095 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
3096 /// X was live across 'load Z' for other reasons, we actually *would* want to
3097 /// fold the addressing mode in the Z case. This would make Y die earlier.
3098 bool AddressingModeMatcher::
3099 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
3100 ExtAddrMode &AMAfter) {
3101 if (IgnoreProfitability) return true;
3103 // AMBefore is the addressing mode before this instruction was folded into it,
3104 // and AMAfter is the addressing mode after the instruction was folded. Get
3105 // the set of registers referenced by AMAfter and subtract out those
3106 // referenced by AMBefore: this is the set of values which folding in this
3107 // address extends the lifetime of.
3109 // Note that there are only two potential values being referenced here,
3110 // BaseReg and ScaleReg (global addresses are always available, as are any
3111 // folded immediates).
3112 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
3114 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
3115 // lifetime wasn't extended by adding this instruction.
3116 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3118 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3119 ScaledReg = nullptr;
3121 // If folding this instruction (and it's subexprs) didn't extend any live
3122 // ranges, we're ok with it.
3123 if (!BaseReg && !ScaledReg)
3126 // If all uses of this instruction are ultimately load/store/inlineasm's,
3127 // check to see if their addressing modes will include this instruction. If
3128 // so, we can fold it into all uses, so it doesn't matter if it has multiple
3130 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
3131 SmallPtrSet<Instruction*, 16> ConsideredInsts;
3132 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
3133 return false; // Has a non-memory, non-foldable use!
3135 // Now that we know that all uses of this instruction are part of a chain of
3136 // computation involving only operations that could theoretically be folded
3137 // into a memory use, loop over each of these uses and see if they could
3138 // *actually* fold the instruction.
3139 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
3140 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
3141 Instruction *User = MemoryUses[i].first;
3142 unsigned OpNo = MemoryUses[i].second;
3144 // Get the access type of this use. If the use isn't a pointer, we don't
3145 // know what it accesses.
3146 Value *Address = User->getOperand(OpNo);
3147 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
3150 Type *AddressAccessTy = AddrTy->getElementType();
3151 unsigned AS = AddrTy->getAddressSpace();
3153 // Do a match against the root of this address, ignoring profitability. This
3154 // will tell us if the addressing mode for the memory operation will
3155 // *actually* cover the shared instruction.
3157 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3158 TPT.getRestorationPoint();
3159 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
3160 MemoryInst, Result, InsertedTruncs,
3161 PromotedInsts, TPT);
3162 Matcher.IgnoreProfitability = true;
3163 bool Success = Matcher.MatchAddr(Address, 0);
3164 (void)Success; assert(Success && "Couldn't select *anything*?");
3166 // The match was to check the profitability, the changes made are not
3167 // part of the original matcher. Therefore, they should be dropped
3168 // otherwise the original matcher will not present the right state.
3169 TPT.rollback(LastKnownGood);
3171 // If the match didn't cover I, then it won't be shared by it.
3172 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3173 I) == MatchedAddrModeInsts.end())
3176 MatchedAddrModeInsts.clear();
3182 } // end anonymous namespace
3184 /// IsNonLocalValue - Return true if the specified values are defined in a
3185 /// different basic block than BB.
3186 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3187 if (Instruction *I = dyn_cast<Instruction>(V))
3188 return I->getParent() != BB;
3192 /// OptimizeMemoryInst - Load and Store Instructions often have
3193 /// addressing modes that can do significant amounts of computation. As such,
3194 /// instruction selection will try to get the load or store to do as much
3195 /// computation as possible for the program. The problem is that isel can only
3196 /// see within a single block. As such, we sink as much legal addressing mode
3197 /// stuff into the block as possible.
3199 /// This method is used to optimize both load/store and inline asms with memory
3201 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3202 Type *AccessTy, unsigned AddrSpace) {
3205 // Try to collapse single-value PHI nodes. This is necessary to undo
3206 // unprofitable PRE transformations.
3207 SmallVector<Value*, 8> worklist;
3208 SmallPtrSet<Value*, 16> Visited;
3209 worklist.push_back(Addr);
3211 // Use a worklist to iteratively look through PHI nodes, and ensure that
3212 // the addressing mode obtained from the non-PHI roots of the graph
3214 Value *Consensus = nullptr;
3215 unsigned NumUsesConsensus = 0;
3216 bool IsNumUsesConsensusValid = false;
3217 SmallVector<Instruction*, 16> AddrModeInsts;
3218 ExtAddrMode AddrMode;
3219 TypePromotionTransaction TPT;
3220 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3221 TPT.getRestorationPoint();
3222 while (!worklist.empty()) {
3223 Value *V = worklist.back();
3224 worklist.pop_back();
3226 // Break use-def graph loops.
3227 if (!Visited.insert(V).second) {
3228 Consensus = nullptr;
3232 // For a PHI node, push all of its incoming values.
3233 if (PHINode *P = dyn_cast<PHINode>(V)) {
3234 for (Value *IncValue : P->incoming_values())
3235 worklist.push_back(IncValue);
3239 // For non-PHIs, determine the addressing mode being computed.
3240 SmallVector<Instruction*, 16> NewAddrModeInsts;
3241 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3242 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
3243 InsertedTruncsSet, PromotedInsts, TPT);
3245 // This check is broken into two cases with very similar code to avoid using
3246 // getNumUses() as much as possible. Some values have a lot of uses, so
3247 // calling getNumUses() unconditionally caused a significant compile-time
3251 AddrMode = NewAddrMode;
3252 AddrModeInsts = NewAddrModeInsts;
3254 } else if (NewAddrMode == AddrMode) {
3255 if (!IsNumUsesConsensusValid) {
3256 NumUsesConsensus = Consensus->getNumUses();
3257 IsNumUsesConsensusValid = true;
3260 // Ensure that the obtained addressing mode is equivalent to that obtained
3261 // for all other roots of the PHI traversal. Also, when choosing one
3262 // such root as representative, select the one with the most uses in order
3263 // to keep the cost modeling heuristics in AddressingModeMatcher
3265 unsigned NumUses = V->getNumUses();
3266 if (NumUses > NumUsesConsensus) {
3268 NumUsesConsensus = NumUses;
3269 AddrModeInsts = NewAddrModeInsts;
3274 Consensus = nullptr;
3278 // If the addressing mode couldn't be determined, or if multiple different
3279 // ones were determined, bail out now.
3281 TPT.rollback(LastKnownGood);
3286 // Check to see if any of the instructions supersumed by this addr mode are
3287 // non-local to I's BB.
3288 bool AnyNonLocal = false;
3289 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3290 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3296 // If all the instructions matched are already in this BB, don't do anything.
3298 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3302 // Insert this computation right after this user. Since our caller is
3303 // scanning from the top of the BB to the bottom, reuse of the expr are
3304 // guaranteed to happen later.
3305 IRBuilder<> Builder(MemoryInst);
3307 // Now that we determined the addressing expression we want to use and know
3308 // that we have to sink it into this block. Check to see if we have already
3309 // done this for some other load/store instr in this block. If so, reuse the
3311 Value *&SunkAddr = SunkAddrs[Addr];
3313 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3314 << *MemoryInst << "\n");
3315 if (SunkAddr->getType() != Addr->getType())
3316 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3317 } else if (AddrSinkUsingGEPs ||
3318 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3319 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3321 // By default, we use the GEP-based method when AA is used later. This
3322 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3323 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3324 << *MemoryInst << "\n");
3325 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3326 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3328 // First, find the pointer.
3329 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3330 ResultPtr = AddrMode.BaseReg;
3331 AddrMode.BaseReg = nullptr;
3334 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3335 // We can't add more than one pointer together, nor can we scale a
3336 // pointer (both of which seem meaningless).
3337 if (ResultPtr || AddrMode.Scale != 1)
3340 ResultPtr = AddrMode.ScaledReg;
3344 if (AddrMode.BaseGV) {
3348 ResultPtr = AddrMode.BaseGV;
3351 // If the real base value actually came from an inttoptr, then the matcher
3352 // will look through it and provide only the integer value. In that case,
3354 if (!ResultPtr && AddrMode.BaseReg) {
3356 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3357 AddrMode.BaseReg = nullptr;
3358 } else if (!ResultPtr && AddrMode.Scale == 1) {
3360 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3365 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3366 SunkAddr = Constant::getNullValue(Addr->getType());
3367 } else if (!ResultPtr) {
3371 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3372 Type *I8Ty = Builder.getInt8Ty();
3374 // Start with the base register. Do this first so that subsequent address
3375 // matching finds it last, which will prevent it from trying to match it
3376 // as the scaled value in case it happens to be a mul. That would be
3377 // problematic if we've sunk a different mul for the scale, because then
3378 // we'd end up sinking both muls.
3379 if (AddrMode.BaseReg) {
3380 Value *V = AddrMode.BaseReg;
3381 if (V->getType() != IntPtrTy)
3382 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3387 // Add the scale value.
3388 if (AddrMode.Scale) {
3389 Value *V = AddrMode.ScaledReg;
3390 if (V->getType() == IntPtrTy) {
3392 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3393 cast<IntegerType>(V->getType())->getBitWidth()) {
3394 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3396 // It is only safe to sign extend the BaseReg if we know that the math
3397 // required to create it did not overflow before we extend it. Since
3398 // the original IR value was tossed in favor of a constant back when
3399 // the AddrMode was created we need to bail out gracefully if widths
3400 // do not match instead of extending it.
3401 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3402 if (I && (ResultIndex != AddrMode.BaseReg))
3403 I->eraseFromParent();
3407 if (AddrMode.Scale != 1)
3408 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3411 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3416 // Add in the Base Offset if present.
3417 if (AddrMode.BaseOffs) {
3418 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3420 // We need to add this separately from the scale above to help with
3421 // SDAG consecutive load/store merging.
3422 if (ResultPtr->getType() != I8PtrTy)
3423 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3424 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3431 SunkAddr = ResultPtr;
3433 if (ResultPtr->getType() != I8PtrTy)
3434 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3435 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3438 if (SunkAddr->getType() != Addr->getType())
3439 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3442 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3443 << *MemoryInst << "\n");
3444 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3445 Value *Result = nullptr;
3447 // Start with the base register. Do this first so that subsequent address
3448 // matching finds it last, which will prevent it from trying to match it
3449 // as the scaled value in case it happens to be a mul. That would be
3450 // problematic if we've sunk a different mul for the scale, because then
3451 // we'd end up sinking both muls.
3452 if (AddrMode.BaseReg) {
3453 Value *V = AddrMode.BaseReg;
3454 if (V->getType()->isPointerTy())
3455 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3456 if (V->getType() != IntPtrTy)
3457 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3461 // Add the scale value.
3462 if (AddrMode.Scale) {
3463 Value *V = AddrMode.ScaledReg;
3464 if (V->getType() == IntPtrTy) {
3466 } else if (V->getType()->isPointerTy()) {
3467 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3468 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3469 cast<IntegerType>(V->getType())->getBitWidth()) {
3470 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3472 // It is only safe to sign extend the BaseReg if we know that the math
3473 // required to create it did not overflow before we extend it. Since
3474 // the original IR value was tossed in favor of a constant back when
3475 // the AddrMode was created we need to bail out gracefully if widths
3476 // do not match instead of extending it.
3477 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3478 if (I && (Result != AddrMode.BaseReg))
3479 I->eraseFromParent();
3482 if (AddrMode.Scale != 1)
3483 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3486 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3491 // Add in the BaseGV if present.
3492 if (AddrMode.BaseGV) {
3493 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3495 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3500 // Add in the Base Offset if present.
3501 if (AddrMode.BaseOffs) {
3502 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3504 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3510 SunkAddr = Constant::getNullValue(Addr->getType());
3512 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3515 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3517 // If we have no uses, recursively delete the value and all dead instructions
3519 if (Repl->use_empty()) {
3520 // This can cause recursive deletion, which can invalidate our iterator.
3521 // Use a WeakVH to hold onto it in case this happens.
3522 WeakVH IterHandle(CurInstIterator);
3523 BasicBlock *BB = CurInstIterator->getParent();
3525 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3527 if (IterHandle != CurInstIterator) {
3528 // If the iterator instruction was recursively deleted, start over at the
3529 // start of the block.
3530 CurInstIterator = BB->begin();
3538 /// OptimizeInlineAsmInst - If there are any memory operands, use
3539 /// OptimizeMemoryInst to sink their address computing into the block when
3540 /// possible / profitable.
3541 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
3542 bool MadeChange = false;
3544 const TargetRegisterInfo *TRI =
3545 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3546 TargetLowering::AsmOperandInfoVector
3547 TargetConstraints = TLI->ParseConstraints(TRI, CS);
3549 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3550 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3552 // Compute the constraint code and ConstraintType to use.
3553 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3555 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3556 OpInfo.isIndirect) {
3557 Value *OpVal = CS->getArgOperand(ArgNo++);
3558 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
3559 } else if (OpInfo.Type == InlineAsm::isInput)
3566 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3567 /// sign extensions.
3568 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3569 assert(!Inst->use_empty() && "Input must have at least one use");
3570 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3571 bool IsSExt = isa<SExtInst>(FirstUser);
3572 Type *ExtTy = FirstUser->getType();
3573 for (const User *U : Inst->users()) {
3574 const Instruction *UI = cast<Instruction>(U);
3575 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3577 Type *CurTy = UI->getType();
3578 // Same input and output types: Same instruction after CSE.
3582 // If IsSExt is true, we are in this situation:
3584 // b = sext ty1 a to ty2
3585 // c = sext ty1 a to ty3
3586 // Assuming ty2 is shorter than ty3, this could be turned into:
3588 // b = sext ty1 a to ty2
3589 // c = sext ty2 b to ty3
3590 // However, the last sext is not free.
3594 // This is a ZExt, maybe this is free to extend from one type to another.
3595 // In that case, we would not account for a different use.
3598 if (ExtTy->getScalarType()->getIntegerBitWidth() >
3599 CurTy->getScalarType()->getIntegerBitWidth()) {
3607 if (!TLI.isZExtFree(NarrowTy, LargeTy))
3610 // All uses are the same or can be derived from one another for free.
3614 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
3615 /// load instruction.
3616 /// If an ext(load) can be formed, it is returned via \p LI for the load
3617 /// and \p Inst for the extension.
3618 /// Otherwise LI == nullptr and Inst == nullptr.
3619 /// When some promotion happened, \p TPT contains the proper state to
3622 /// \return true when promoting was necessary to expose the ext(load)
3623 /// opportunity, false otherwise.
3627 /// %ld = load i32* %addr
3628 /// %add = add nuw i32 %ld, 4
3629 /// %zext = zext i32 %add to i64
3633 /// %ld = load i32* %addr
3634 /// %zext = zext i32 %ld to i64
3635 /// %add = add nuw i64 %zext, 4
3637 /// Thanks to the promotion, we can match zext(load i32*) to i64.
3638 bool CodeGenPrepare::ExtLdPromotion(TypePromotionTransaction &TPT,
3639 LoadInst *&LI, Instruction *&Inst,
3640 const SmallVectorImpl<Instruction *> &Exts,
3641 unsigned CreatedInstsCost = 0) {
3642 // Iterate over all the extensions to see if one form an ext(load).
3643 for (auto I : Exts) {
3644 // Check if we directly have ext(load).
3645 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
3647 // No promotion happened here.
3650 // Check whether or not we want to do any promotion.
3651 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
3653 // Get the action to perform the promotion.
3654 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
3655 I, InsertedTruncsSet, *TLI, PromotedInsts);
3656 // Check if we can promote.
3659 // Save the current state.
3660 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3661 TPT.getRestorationPoint();
3662 SmallVector<Instruction *, 4> NewExts;
3663 unsigned NewCreatedInstsCost = 0;
3664 unsigned ExtCost = !TLI->isExtFree(I);
3666 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
3667 &NewExts, nullptr, *TLI);
3668 assert(PromotedVal &&
3669 "TypePromotionHelper should have filtered out those cases");
3671 // We would be able to merge only one extension in a load.
3672 // Therefore, if we have more than 1 new extension we heuristically
3673 // cut this search path, because it means we degrade the code quality.
3674 // With exactly 2, the transformation is neutral, because we will merge
3675 // one extension but leave one. However, we optimistically keep going,
3676 // because the new extension may be removed too.
3677 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
3678 TotalCreatedInstsCost -= ExtCost;
3679 if (!StressExtLdPromotion &&
3680 (TotalCreatedInstsCost > 1 ||
3681 !isPromotedInstructionLegal(*TLI, PromotedVal))) {
3682 // The promotion is not profitable, rollback to the previous state.
3683 TPT.rollback(LastKnownGood);
3686 // The promotion is profitable.
3687 // Check if it exposes an ext(load).
3688 (void)ExtLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
3689 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
3690 // If we have created a new extension, i.e., now we have two
3691 // extensions. We must make sure one of them is merged with
3692 // the load, otherwise we may degrade the code quality.
3693 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
3694 // Promotion happened.
3696 // If this does not help to expose an ext(load) then, rollback.
3697 TPT.rollback(LastKnownGood);
3699 // None of the extension can form an ext(load).
3705 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
3706 /// basic block as the load, unless conditions are unfavorable. This allows
3707 /// SelectionDAG to fold the extend into the load.
3708 /// \p I[in/out] the extension may be modified during the process if some
3709 /// promotions apply.
3711 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *&I) {
3712 // Try to promote a chain of computation if it allows to form
3713 // an extended load.
3714 TypePromotionTransaction TPT;
3715 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3716 TPT.getRestorationPoint();
3717 SmallVector<Instruction *, 1> Exts;
3719 // Look for a load being extended.
3720 LoadInst *LI = nullptr;
3721 Instruction *OldExt = I;
3722 bool HasPromoted = ExtLdPromotion(TPT, LI, I, Exts);
3724 assert(!HasPromoted && !LI && "If we did not match any load instruction "
3725 "the code must remain the same");
3730 // If they're already in the same block, there's nothing to do.
3731 // Make the cheap checks first if we did not promote.
3732 // If we promoted, we need to check if it is indeed profitable.
3733 if (!HasPromoted && LI->getParent() == I->getParent())
3736 EVT VT = TLI->getValueType(I->getType());
3737 EVT LoadVT = TLI->getValueType(LI->getType());
3739 // If the load has other users and the truncate is not free, this probably
3740 // isn't worthwhile.
3741 if (!LI->hasOneUse() && TLI &&
3742 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
3743 !TLI->isTruncateFree(I->getType(), LI->getType())) {
3745 TPT.rollback(LastKnownGood);
3749 // Check whether the target supports casts folded into loads.
3751 if (isa<ZExtInst>(I))
3752 LType = ISD::ZEXTLOAD;
3754 assert(isa<SExtInst>(I) && "Unexpected ext type!");
3755 LType = ISD::SEXTLOAD;
3757 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
3759 TPT.rollback(LastKnownGood);
3763 // Move the extend into the same block as the load, so that SelectionDAG
3766 I->removeFromParent();
3772 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
3773 BasicBlock *DefBB = I->getParent();
3775 // If the result of a {s|z}ext and its source are both live out, rewrite all
3776 // other uses of the source with result of extension.
3777 Value *Src = I->getOperand(0);
3778 if (Src->hasOneUse())
3781 // Only do this xform if truncating is free.
3782 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
3785 // Only safe to perform the optimization if the source is also defined in
3787 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
3790 bool DefIsLiveOut = false;
3791 for (User *U : I->users()) {
3792 Instruction *UI = cast<Instruction>(U);
3794 // Figure out which BB this ext is used in.
3795 BasicBlock *UserBB = UI->getParent();
3796 if (UserBB == DefBB) continue;
3797 DefIsLiveOut = true;
3803 // Make sure none of the uses are PHI nodes.
3804 for (User *U : Src->users()) {
3805 Instruction *UI = cast<Instruction>(U);
3806 BasicBlock *UserBB = UI->getParent();
3807 if (UserBB == DefBB) continue;
3808 // Be conservative. We don't want this xform to end up introducing
3809 // reloads just before load / store instructions.
3810 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
3814 // InsertedTruncs - Only insert one trunc in each block once.
3815 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
3817 bool MadeChange = false;
3818 for (Use &U : Src->uses()) {
3819 Instruction *User = cast<Instruction>(U.getUser());
3821 // Figure out which BB this ext is used in.
3822 BasicBlock *UserBB = User->getParent();
3823 if (UserBB == DefBB) continue;
3825 // Both src and def are live in this block. Rewrite the use.
3826 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
3828 if (!InsertedTrunc) {
3829 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3830 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
3831 InsertedTruncsSet.insert(InsertedTrunc);
3834 // Replace a use of the {s|z}ext source with a use of the result.
3843 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
3844 /// turned into an explicit branch.
3845 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
3846 // FIXME: This should use the same heuristics as IfConversion to determine
3847 // whether a select is better represented as a branch. This requires that
3848 // branch probability metadata is preserved for the select, which is not the
3851 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3853 // If the branch is predicted right, an out of order CPU can avoid blocking on
3854 // the compare. Emit cmovs on compares with a memory operand as branches to
3855 // avoid stalls on the load from memory. If the compare has more than one use
3856 // there's probably another cmov or setcc around so it's not worth emitting a
3861 Value *CmpOp0 = Cmp->getOperand(0);
3862 Value *CmpOp1 = Cmp->getOperand(1);
3864 // We check that the memory operand has one use to avoid uses of the loaded
3865 // value directly after the compare, making branches unprofitable.
3866 return Cmp->hasOneUse() &&
3867 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3868 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
3872 /// If we have a SelectInst that will likely profit from branch prediction,
3873 /// turn it into a branch.
3874 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
3875 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3877 // Can we convert the 'select' to CF ?
3878 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3881 TargetLowering::SelectSupportKind SelectKind;
3883 SelectKind = TargetLowering::VectorMaskSelect;
3884 else if (SI->getType()->isVectorTy())
3885 SelectKind = TargetLowering::ScalarCondVectorVal;
3887 SelectKind = TargetLowering::ScalarValSelect;
3889 // Do we have efficient codegen support for this kind of 'selects' ?
3890 if (TLI->isSelectSupported(SelectKind)) {
3891 // We have efficient codegen support for the select instruction.
3892 // Check if it is profitable to keep this 'select'.
3893 if (!TLI->isPredictableSelectExpensive() ||
3894 !isFormingBranchFromSelectProfitable(SI))
3900 // First, we split the block containing the select into 2 blocks.
3901 BasicBlock *StartBlock = SI->getParent();
3902 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3903 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3905 // Create a new block serving as the landing pad for the branch.
3906 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3907 NextBlock->getParent(), NextBlock);
3909 // Move the unconditional branch from the block with the select in it into our
3910 // landing pad block.
3911 StartBlock->getTerminator()->eraseFromParent();
3912 BranchInst::Create(NextBlock, SmallBlock);
3914 // Insert the real conditional branch based on the original condition.
3915 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3917 // The select itself is replaced with a PHI Node.
3918 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3920 PN->addIncoming(SI->getTrueValue(), StartBlock);
3921 PN->addIncoming(SI->getFalseValue(), SmallBlock);
3922 SI->replaceAllUsesWith(PN);
3923 SI->eraseFromParent();
3925 // Instruct OptimizeBlock to skip to the next block.
3926 CurInstIterator = StartBlock->end();
3927 ++NumSelectsExpanded;
3931 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3932 SmallVector<int, 16> Mask(SVI->getShuffleMask());
3934 for (unsigned i = 0; i < Mask.size(); ++i) {
3935 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3937 SplatElem = Mask[i];
3943 /// Some targets have expensive vector shifts if the lanes aren't all the same
3944 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3945 /// it's often worth sinking a shufflevector splat down to its use so that
3946 /// codegen can spot all lanes are identical.
3947 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3948 BasicBlock *DefBB = SVI->getParent();
3950 // Only do this xform if variable vector shifts are particularly expensive.
3951 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3954 // We only expect better codegen by sinking a shuffle if we can recognise a
3956 if (!isBroadcastShuffle(SVI))
3959 // InsertedShuffles - Only insert a shuffle in each block once.
3960 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3962 bool MadeChange = false;
3963 for (User *U : SVI->users()) {
3964 Instruction *UI = cast<Instruction>(U);
3966 // Figure out which BB this ext is used in.
3967 BasicBlock *UserBB = UI->getParent();
3968 if (UserBB == DefBB) continue;
3970 // For now only apply this when the splat is used by a shift instruction.
3971 if (!UI->isShift()) continue;
3973 // Everything checks out, sink the shuffle if the user's block doesn't
3974 // already have a copy.
3975 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3977 if (!InsertedShuffle) {
3978 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3979 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3981 SVI->getOperand(2), "", InsertPt);
3984 UI->replaceUsesOfWith(SVI, InsertedShuffle);
3988 // If we removed all uses, nuke the shuffle.
3989 if (SVI->use_empty()) {
3990 SVI->eraseFromParent();
3998 /// \brief Helper class to promote a scalar operation to a vector one.
3999 /// This class is used to move downward extractelement transition.
4001 /// a = vector_op <2 x i32>
4002 /// b = extractelement <2 x i32> a, i32 0
4007 /// a = vector_op <2 x i32>
4008 /// c = vector_op a (equivalent to scalar_op on the related lane)
4009 /// * d = extractelement <2 x i32> c, i32 0
4011 /// Assuming both extractelement and store can be combine, we get rid of the
4013 class VectorPromoteHelper {
4014 /// Used to perform some checks on the legality of vector operations.
4015 const TargetLowering &TLI;
4017 /// Used to estimated the cost of the promoted chain.
4018 const TargetTransformInfo &TTI;
4020 /// The transition being moved downwards.
4021 Instruction *Transition;
4022 /// The sequence of instructions to be promoted.
4023 SmallVector<Instruction *, 4> InstsToBePromoted;
4024 /// Cost of combining a store and an extract.
4025 unsigned StoreExtractCombineCost;
4026 /// Instruction that will be combined with the transition.
4027 Instruction *CombineInst;
4029 /// \brief The instruction that represents the current end of the transition.
4030 /// Since we are faking the promotion until we reach the end of the chain
4031 /// of computation, we need a way to get the current end of the transition.
4032 Instruction *getEndOfTransition() const {
4033 if (InstsToBePromoted.empty())
4035 return InstsToBePromoted.back();
4038 /// \brief Return the index of the original value in the transition.
4039 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
4040 /// c, is at index 0.
4041 unsigned getTransitionOriginalValueIdx() const {
4042 assert(isa<ExtractElementInst>(Transition) &&
4043 "Other kind of transitions are not supported yet");
4047 /// \brief Return the index of the index in the transition.
4048 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
4050 unsigned getTransitionIdx() const {
4051 assert(isa<ExtractElementInst>(Transition) &&
4052 "Other kind of transitions are not supported yet");
4056 /// \brief Get the type of the transition.
4057 /// This is the type of the original value.
4058 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
4059 /// transition is <2 x i32>.
4060 Type *getTransitionType() const {
4061 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
4064 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
4065 /// I.e., we have the following sequence:
4066 /// Def = Transition <ty1> a to <ty2>
4067 /// b = ToBePromoted <ty2> Def, ...
4069 /// b = ToBePromoted <ty1> a, ...
4070 /// Def = Transition <ty1> ToBePromoted to <ty2>
4071 void promoteImpl(Instruction *ToBePromoted);
4073 /// \brief Check whether or not it is profitable to promote all the
4074 /// instructions enqueued to be promoted.
4075 bool isProfitableToPromote() {
4076 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
4077 unsigned Index = isa<ConstantInt>(ValIdx)
4078 ? cast<ConstantInt>(ValIdx)->getZExtValue()
4080 Type *PromotedType = getTransitionType();
4082 StoreInst *ST = cast<StoreInst>(CombineInst);
4083 unsigned AS = ST->getPointerAddressSpace();
4084 unsigned Align = ST->getAlignment();
4085 // Check if this store is supported.
4086 if (!TLI.allowsMisalignedMemoryAccesses(
4087 TLI.getValueType(ST->getValueOperand()->getType()), AS, Align)) {
4088 // If this is not supported, there is no way we can combine
4089 // the extract with the store.
4093 // The scalar chain of computation has to pay for the transition
4094 // scalar to vector.
4095 // The vector chain has to account for the combining cost.
4096 uint64_t ScalarCost =
4097 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
4098 uint64_t VectorCost = StoreExtractCombineCost;
4099 for (const auto &Inst : InstsToBePromoted) {
4100 // Compute the cost.
4101 // By construction, all instructions being promoted are arithmetic ones.
4102 // Moreover, one argument is a constant that can be viewed as a splat
4104 Value *Arg0 = Inst->getOperand(0);
4105 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
4106 isa<ConstantFP>(Arg0);
4107 TargetTransformInfo::OperandValueKind Arg0OVK =
4108 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4109 : TargetTransformInfo::OK_AnyValue;
4110 TargetTransformInfo::OperandValueKind Arg1OVK =
4111 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4112 : TargetTransformInfo::OK_AnyValue;
4113 ScalarCost += TTI.getArithmeticInstrCost(
4114 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
4115 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
4118 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
4119 << ScalarCost << "\nVector: " << VectorCost << '\n');
4120 return ScalarCost > VectorCost;
4123 /// \brief Generate a constant vector with \p Val with the same
4124 /// number of elements as the transition.
4125 /// \p UseSplat defines whether or not \p Val should be replicated
4126 /// accross the whole vector.
4127 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
4128 /// otherwise we generate a vector with as many undef as possible:
4129 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
4130 /// used at the index of the extract.
4131 Value *getConstantVector(Constant *Val, bool UseSplat) const {
4132 unsigned ExtractIdx = UINT_MAX;
4134 // If we cannot determine where the constant must be, we have to
4135 // use a splat constant.
4136 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
4137 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
4138 ExtractIdx = CstVal->getSExtValue();
4143 unsigned End = getTransitionType()->getVectorNumElements();
4145 return ConstantVector::getSplat(End, Val);
4147 SmallVector<Constant *, 4> ConstVec;
4148 UndefValue *UndefVal = UndefValue::get(Val->getType());
4149 for (unsigned Idx = 0; Idx != End; ++Idx) {
4150 if (Idx == ExtractIdx)
4151 ConstVec.push_back(Val);
4153 ConstVec.push_back(UndefVal);
4155 return ConstantVector::get(ConstVec);
4158 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
4159 /// in \p Use can trigger undefined behavior.
4160 static bool canCauseUndefinedBehavior(const Instruction *Use,
4161 unsigned OperandIdx) {
4162 // This is not safe to introduce undef when the operand is on
4163 // the right hand side of a division-like instruction.
4164 if (OperandIdx != 1)
4166 switch (Use->getOpcode()) {
4169 case Instruction::SDiv:
4170 case Instruction::UDiv:
4171 case Instruction::SRem:
4172 case Instruction::URem:
4174 case Instruction::FDiv:
4175 case Instruction::FRem:
4176 return !Use->hasNoNaNs();
4178 llvm_unreachable(nullptr);
4182 VectorPromoteHelper(const TargetLowering &TLI, const TargetTransformInfo &TTI,
4183 Instruction *Transition, unsigned CombineCost)
4184 : TLI(TLI), TTI(TTI), Transition(Transition),
4185 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4186 assert(Transition && "Do not know how to promote null");
4189 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4190 bool canPromote(const Instruction *ToBePromoted) const {
4191 // We could support CastInst too.
4192 return isa<BinaryOperator>(ToBePromoted);
4195 /// \brief Check if it is profitable to promote \p ToBePromoted
4196 /// by moving downward the transition through.
4197 bool shouldPromote(const Instruction *ToBePromoted) const {
4198 // Promote only if all the operands can be statically expanded.
4199 // Indeed, we do not want to introduce any new kind of transitions.
4200 for (const Use &U : ToBePromoted->operands()) {
4201 const Value *Val = U.get();
4202 if (Val == getEndOfTransition()) {
4203 // If the use is a division and the transition is on the rhs,
4204 // we cannot promote the operation, otherwise we may create a
4205 // division by zero.
4206 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4210 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4211 !isa<ConstantFP>(Val))
4214 // Check that the resulting operation is legal.
4215 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4218 return StressStoreExtract ||
4219 TLI.isOperationLegalOrCustom(
4220 ISDOpcode, TLI.getValueType(getTransitionType(), true));
4223 /// \brief Check whether or not \p Use can be combined
4224 /// with the transition.
4225 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4226 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4228 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4229 void enqueueForPromotion(Instruction *ToBePromoted) {
4230 InstsToBePromoted.push_back(ToBePromoted);
4233 /// \brief Set the instruction that will be combined with the transition.
4234 void recordCombineInstruction(Instruction *ToBeCombined) {
4235 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4236 CombineInst = ToBeCombined;
4239 /// \brief Promote all the instructions enqueued for promotion if it is
4241 /// \return True if the promotion happened, false otherwise.
4243 // Check if there is something to promote.
4244 // Right now, if we do not have anything to combine with,
4245 // we assume the promotion is not profitable.
4246 if (InstsToBePromoted.empty() || !CombineInst)
4250 if (!StressStoreExtract && !isProfitableToPromote())
4254 for (auto &ToBePromoted : InstsToBePromoted)
4255 promoteImpl(ToBePromoted);
4256 InstsToBePromoted.clear();
4260 } // End of anonymous namespace.
4262 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4263 // At this point, we know that all the operands of ToBePromoted but Def
4264 // can be statically promoted.
4265 // For Def, we need to use its parameter in ToBePromoted:
4266 // b = ToBePromoted ty1 a
4267 // Def = Transition ty1 b to ty2
4268 // Move the transition down.
4269 // 1. Replace all uses of the promoted operation by the transition.
4270 // = ... b => = ... Def.
4271 assert(ToBePromoted->getType() == Transition->getType() &&
4272 "The type of the result of the transition does not match "
4274 ToBePromoted->replaceAllUsesWith(Transition);
4275 // 2. Update the type of the uses.
4276 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
4277 Type *TransitionTy = getTransitionType();
4278 ToBePromoted->mutateType(TransitionTy);
4279 // 3. Update all the operands of the promoted operation with promoted
4281 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
4282 for (Use &U : ToBePromoted->operands()) {
4283 Value *Val = U.get();
4284 Value *NewVal = nullptr;
4285 if (Val == Transition)
4286 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
4287 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
4288 isa<ConstantFP>(Val)) {
4289 // Use a splat constant if it is not safe to use undef.
4290 NewVal = getConstantVector(
4291 cast<Constant>(Val),
4292 isa<UndefValue>(Val) ||
4293 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
4295 llvm_unreachable("Did you modified shouldPromote and forgot to update "
4297 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
4299 Transition->removeFromParent();
4300 Transition->insertAfter(ToBePromoted);
4301 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
4304 /// Some targets can do store(extractelement) with one instruction.
4305 /// Try to push the extractelement towards the stores when the target
4306 /// has this feature and this is profitable.
4307 bool CodeGenPrepare::OptimizeExtractElementInst(Instruction *Inst) {
4308 unsigned CombineCost = UINT_MAX;
4309 if (DisableStoreExtract || !TLI ||
4310 (!StressStoreExtract &&
4311 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4312 Inst->getOperand(1), CombineCost)))
4315 // At this point we know that Inst is a vector to scalar transition.
4316 // Try to move it down the def-use chain, until:
4317 // - We can combine the transition with its single use
4318 // => we got rid of the transition.
4319 // - We escape the current basic block
4320 // => we would need to check that we are moving it at a cheaper place and
4321 // we do not do that for now.
4322 BasicBlock *Parent = Inst->getParent();
4323 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4324 VectorPromoteHelper VPH(*TLI, *TTI, Inst, CombineCost);
4325 // If the transition has more than one use, assume this is not going to be
4327 while (Inst->hasOneUse()) {
4328 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4329 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4331 if (ToBePromoted->getParent() != Parent) {
4332 DEBUG(dbgs() << "Instruction to promote is in a different block ("
4333 << ToBePromoted->getParent()->getName()
4334 << ") than the transition (" << Parent->getName() << ").\n");
4338 if (VPH.canCombine(ToBePromoted)) {
4339 DEBUG(dbgs() << "Assume " << *Inst << '\n'
4340 << "will be combined with: " << *ToBePromoted << '\n');
4341 VPH.recordCombineInstruction(ToBePromoted);
4342 bool Changed = VPH.promote();
4343 NumStoreExtractExposed += Changed;
4347 DEBUG(dbgs() << "Try promoting.\n");
4348 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4351 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4353 VPH.enqueueForPromotion(ToBePromoted);
4354 Inst = ToBePromoted;
4359 bool CodeGenPrepare::OptimizeInst(Instruction *I, bool& ModifiedDT) {
4360 if (PHINode *P = dyn_cast<PHINode>(I)) {
4361 // It is possible for very late stage optimizations (such as SimplifyCFG)
4362 // to introduce PHI nodes too late to be cleaned up. If we detect such a
4363 // trivial PHI, go ahead and zap it here.
4364 const DataLayout &DL = I->getModule()->getDataLayout();
4365 if (Value *V = SimplifyInstruction(P, DL, TLInfo, nullptr)) {
4366 P->replaceAllUsesWith(V);
4367 P->eraseFromParent();
4374 if (CastInst *CI = dyn_cast<CastInst>(I)) {
4375 // If the source of the cast is a constant, then this should have
4376 // already been constant folded. The only reason NOT to constant fold
4377 // it is if something (e.g. LSR) was careful to place the constant
4378 // evaluation in a block other than then one that uses it (e.g. to hoist
4379 // the address of globals out of a loop). If this is the case, we don't
4380 // want to forward-subst the cast.
4381 if (isa<Constant>(CI->getOperand(0)))
4384 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
4387 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4388 /// Sink a zext or sext into its user blocks if the target type doesn't
4389 /// fit in one register
4390 if (TLI && TLI->getTypeAction(CI->getContext(),
4391 TLI->getValueType(CI->getType())) ==
4392 TargetLowering::TypeExpandInteger) {
4393 return SinkCast(CI);
4395 bool MadeChange = MoveExtToFormExtLoad(I);
4396 return MadeChange | OptimizeExtUses(I);
4402 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4403 if (!TLI || !TLI->hasMultipleConditionRegisters())
4404 return OptimizeCmpExpression(CI);
4406 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4408 unsigned AS = LI->getPointerAddressSpace();
4409 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
4414 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4416 unsigned AS = SI->getPointerAddressSpace();
4417 return OptimizeMemoryInst(I, SI->getOperand(1),
4418 SI->getOperand(0)->getType(), AS);
4423 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4425 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4426 BinOp->getOpcode() == Instruction::LShr)) {
4427 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4428 if (TLI && CI && TLI->hasExtractBitsInsn())
4429 return OptimizeExtractBits(BinOp, CI, *TLI);
4434 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4435 if (GEPI->hasAllZeroIndices()) {
4436 /// The GEP operand must be a pointer, so must its result -> BitCast
4437 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4438 GEPI->getName(), GEPI);
4439 GEPI->replaceAllUsesWith(NC);
4440 GEPI->eraseFromParent();
4442 OptimizeInst(NC, ModifiedDT);
4448 if (CallInst *CI = dyn_cast<CallInst>(I))
4449 return OptimizeCallInst(CI, ModifiedDT);
4451 if (SelectInst *SI = dyn_cast<SelectInst>(I))
4452 return OptimizeSelectInst(SI);
4454 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
4455 return OptimizeShuffleVectorInst(SVI);
4457 if (isa<ExtractElementInst>(I))
4458 return OptimizeExtractElementInst(I);
4463 // In this pass we look for GEP and cast instructions that are used
4464 // across basic blocks and rewrite them to improve basic-block-at-a-time
4466 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
4468 bool MadeChange = false;
4470 CurInstIterator = BB.begin();
4471 while (CurInstIterator != BB.end()) {
4472 MadeChange |= OptimizeInst(CurInstIterator++, ModifiedDT);
4476 MadeChange |= DupRetToEnableTailCallOpts(&BB);
4481 // llvm.dbg.value is far away from the value then iSel may not be able
4482 // handle it properly. iSel will drop llvm.dbg.value if it can not
4483 // find a node corresponding to the value.
4484 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
4485 bool MadeChange = false;
4486 for (BasicBlock &BB : F) {
4487 Instruction *PrevNonDbgInst = nullptr;
4488 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
4489 Instruction *Insn = BI++;
4490 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
4491 // Leave dbg.values that refer to an alloca alone. These
4492 // instrinsics describe the address of a variable (= the alloca)
4493 // being taken. They should not be moved next to the alloca
4494 // (and to the beginning of the scope), but rather stay close to
4495 // where said address is used.
4496 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
4497 PrevNonDbgInst = Insn;
4501 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
4502 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
4503 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
4504 DVI->removeFromParent();
4505 if (isa<PHINode>(VI))
4506 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
4508 DVI->insertAfter(VI);
4517 // If there is a sequence that branches based on comparing a single bit
4518 // against zero that can be combined into a single instruction, and the
4519 // target supports folding these into a single instruction, sink the
4520 // mask and compare into the branch uses. Do this before OptimizeBlock ->
4521 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
4523 bool CodeGenPrepare::sinkAndCmp(Function &F) {
4524 if (!EnableAndCmpSinking)
4526 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
4528 bool MadeChange = false;
4529 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
4530 BasicBlock *BB = I++;
4532 // Does this BB end with the following?
4533 // %andVal = and %val, #single-bit-set
4534 // %icmpVal = icmp %andResult, 0
4535 // br i1 %cmpVal label %dest1, label %dest2"
4536 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
4537 if (!Brcc || !Brcc->isConditional())
4539 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
4540 if (!Cmp || Cmp->getParent() != BB)
4542 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
4543 if (!Zero || !Zero->isZero())
4545 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
4546 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
4548 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
4549 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
4551 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
4553 // Push the "and; icmp" for any users that are conditional branches.
4554 // Since there can only be one branch use per BB, we don't need to keep
4555 // track of which BBs we insert into.
4556 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
4560 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
4562 if (!BrccUser || !BrccUser->isConditional())
4564 BasicBlock *UserBB = BrccUser->getParent();
4565 if (UserBB == BB) continue;
4566 DEBUG(dbgs() << "found Brcc use\n");
4568 // Sink the "and; icmp" to use.
4570 BinaryOperator *NewAnd =
4571 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
4574 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
4578 DEBUG(BrccUser->getParent()->dump());
4584 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
4585 /// success, or returns false if no or invalid metadata was found.
4586 static bool extractBranchMetadata(BranchInst *BI,
4587 uint64_t &ProbTrue, uint64_t &ProbFalse) {
4588 assert(BI->isConditional() &&
4589 "Looking for probabilities on unconditional branch?");
4590 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
4591 if (!ProfileData || ProfileData->getNumOperands() != 3)
4594 const auto *CITrue =
4595 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
4596 const auto *CIFalse =
4597 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
4598 if (!CITrue || !CIFalse)
4601 ProbTrue = CITrue->getValue().getZExtValue();
4602 ProbFalse = CIFalse->getValue().getZExtValue();
4607 /// \brief Scale down both weights to fit into uint32_t.
4608 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
4609 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
4610 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
4611 NewTrue = NewTrue / Scale;
4612 NewFalse = NewFalse / Scale;
4615 /// \brief Some targets prefer to split a conditional branch like:
4617 /// %0 = icmp ne i32 %a, 0
4618 /// %1 = icmp ne i32 %b, 0
4619 /// %or.cond = or i1 %0, %1
4620 /// br i1 %or.cond, label %TrueBB, label %FalseBB
4622 /// into multiple branch instructions like:
4625 /// %0 = icmp ne i32 %a, 0
4626 /// br i1 %0, label %TrueBB, label %bb2
4628 /// %1 = icmp ne i32 %b, 0
4629 /// br i1 %1, label %TrueBB, label %FalseBB
4631 /// This usually allows instruction selection to do even further optimizations
4632 /// and combine the compare with the branch instruction. Currently this is
4633 /// applied for targets which have "cheap" jump instructions.
4635 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
4637 bool CodeGenPrepare::splitBranchCondition(Function &F) {
4638 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
4641 bool MadeChange = false;
4642 for (auto &BB : F) {
4643 // Does this BB end with the following?
4644 // %cond1 = icmp|fcmp|binary instruction ...
4645 // %cond2 = icmp|fcmp|binary instruction ...
4646 // %cond.or = or|and i1 %cond1, cond2
4647 // br i1 %cond.or label %dest1, label %dest2"
4648 BinaryOperator *LogicOp;
4649 BasicBlock *TBB, *FBB;
4650 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
4654 Value *Cond1, *Cond2;
4655 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
4656 m_OneUse(m_Value(Cond2)))))
4657 Opc = Instruction::And;
4658 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
4659 m_OneUse(m_Value(Cond2)))))
4660 Opc = Instruction::Or;
4664 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
4665 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
4668 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
4671 auto *InsertBefore = std::next(Function::iterator(BB))
4672 .getNodePtrUnchecked();
4673 auto TmpBB = BasicBlock::Create(BB.getContext(),
4674 BB.getName() + ".cond.split",
4675 BB.getParent(), InsertBefore);
4677 // Update original basic block by using the first condition directly by the
4678 // branch instruction and removing the no longer needed and/or instruction.
4679 auto *Br1 = cast<BranchInst>(BB.getTerminator());
4680 Br1->setCondition(Cond1);
4681 LogicOp->eraseFromParent();
4683 // Depending on the conditon we have to either replace the true or the false
4684 // successor of the original branch instruction.
4685 if (Opc == Instruction::And)
4686 Br1->setSuccessor(0, TmpBB);
4688 Br1->setSuccessor(1, TmpBB);
4690 // Fill in the new basic block.
4691 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
4692 if (auto *I = dyn_cast<Instruction>(Cond2)) {
4693 I->removeFromParent();
4694 I->insertBefore(Br2);
4697 // Update PHI nodes in both successors. The original BB needs to be
4698 // replaced in one succesor's PHI nodes, because the branch comes now from
4699 // the newly generated BB (NewBB). In the other successor we need to add one
4700 // incoming edge to the PHI nodes, because both branch instructions target
4701 // now the same successor. Depending on the original branch condition
4702 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
4703 // we perfrom the correct update for the PHI nodes.
4704 // This doesn't change the successor order of the just created branch
4705 // instruction (or any other instruction).
4706 if (Opc == Instruction::Or)
4707 std::swap(TBB, FBB);
4709 // Replace the old BB with the new BB.
4710 for (auto &I : *TBB) {
4711 PHINode *PN = dyn_cast<PHINode>(&I);
4715 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
4716 PN->setIncomingBlock(i, TmpBB);
4719 // Add another incoming edge form the new BB.
4720 for (auto &I : *FBB) {
4721 PHINode *PN = dyn_cast<PHINode>(&I);
4724 auto *Val = PN->getIncomingValueForBlock(&BB);
4725 PN->addIncoming(Val, TmpBB);
4728 // Update the branch weights (from SelectionDAGBuilder::
4729 // FindMergedConditions).
4730 if (Opc == Instruction::Or) {
4731 // Codegen X | Y as:
4740 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
4741 // The requirement is that
4742 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
4743 // = TrueProb for orignal BB.
4744 // Assuming the orignal weights are A and B, one choice is to set BB1's
4745 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
4747 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
4748 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
4749 // TmpBB, but the math is more complicated.
4750 uint64_t TrueWeight, FalseWeight;
4751 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4752 uint64_t NewTrueWeight = TrueWeight;
4753 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
4754 scaleWeights(NewTrueWeight, NewFalseWeight);
4755 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4756 .createBranchWeights(TrueWeight, FalseWeight));
4758 NewTrueWeight = TrueWeight;
4759 NewFalseWeight = 2 * FalseWeight;
4760 scaleWeights(NewTrueWeight, NewFalseWeight);
4761 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4762 .createBranchWeights(TrueWeight, FalseWeight));
4765 // Codegen X & Y as:
4773 // This requires creation of TmpBB after CurBB.
4775 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
4776 // The requirement is that
4777 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
4778 // = FalseProb for orignal BB.
4779 // Assuming the orignal weights are A and B, one choice is to set BB1's
4780 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
4782 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
4783 uint64_t TrueWeight, FalseWeight;
4784 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4785 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
4786 uint64_t NewFalseWeight = FalseWeight;
4787 scaleWeights(NewTrueWeight, NewFalseWeight);
4788 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4789 .createBranchWeights(TrueWeight, FalseWeight));
4791 NewTrueWeight = 2 * TrueWeight;
4792 NewFalseWeight = FalseWeight;
4793 scaleWeights(NewTrueWeight, NewFalseWeight);
4794 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4795 .createBranchWeights(TrueWeight, FalseWeight));
4799 // Note: No point in getting fancy here, since the DT info is never
4800 // available to CodeGenPrepare.
4805 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();