1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass munges the code in the input function to better prepare it for
10 // SelectionDAG-based code generation. This works around limitations in it's
11 // basic-block-at-a-time approach. It should eventually be removed.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "llvm/ADT/APInt.h"
16 #include "llvm/ADT/ArrayRef.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/MapVector.h"
19 #include "llvm/ADT/PointerIntPair.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/Analysis/BlockFrequencyInfo.h"
25 #include "llvm/Analysis/BranchProbabilityInfo.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Analysis/InstructionSimplify.h"
28 #include "llvm/Analysis/LoopInfo.h"
29 #include "llvm/Analysis/MemoryBuiltins.h"
30 #include "llvm/Analysis/ProfileSummaryInfo.h"
31 #include "llvm/Analysis/TargetLibraryInfo.h"
32 #include "llvm/Analysis/TargetTransformInfo.h"
33 #include "llvm/Transforms/Utils/Local.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/Analysis/VectorUtils.h"
36 #include "llvm/CodeGen/Analysis.h"
37 #include "llvm/CodeGen/ISDOpcodes.h"
38 #include "llvm/CodeGen/SelectionDAGNodes.h"
39 #include "llvm/CodeGen/TargetLowering.h"
40 #include "llvm/CodeGen/TargetPassConfig.h"
41 #include "llvm/CodeGen/TargetSubtargetInfo.h"
42 #include "llvm/CodeGen/ValueTypes.h"
43 #include "llvm/Config/llvm-config.h"
44 #include "llvm/IR/Argument.h"
45 #include "llvm/IR/Attributes.h"
46 #include "llvm/IR/BasicBlock.h"
47 #include "llvm/IR/CallSite.h"
48 #include "llvm/IR/Constant.h"
49 #include "llvm/IR/Constants.h"
50 #include "llvm/IR/DataLayout.h"
51 #include "llvm/IR/DerivedTypes.h"
52 #include "llvm/IR/Dominators.h"
53 #include "llvm/IR/Function.h"
54 #include "llvm/IR/GetElementPtrTypeIterator.h"
55 #include "llvm/IR/GlobalValue.h"
56 #include "llvm/IR/GlobalVariable.h"
57 #include "llvm/IR/IRBuilder.h"
58 #include "llvm/IR/InlineAsm.h"
59 #include "llvm/IR/InstrTypes.h"
60 #include "llvm/IR/Instruction.h"
61 #include "llvm/IR/Instructions.h"
62 #include "llvm/IR/IntrinsicInst.h"
63 #include "llvm/IR/Intrinsics.h"
64 #include "llvm/IR/LLVMContext.h"
65 #include "llvm/IR/MDBuilder.h"
66 #include "llvm/IR/Module.h"
67 #include "llvm/IR/Operator.h"
68 #include "llvm/IR/PatternMatch.h"
69 #include "llvm/IR/Statepoint.h"
70 #include "llvm/IR/Type.h"
71 #include "llvm/IR/Use.h"
72 #include "llvm/IR/User.h"
73 #include "llvm/IR/Value.h"
74 #include "llvm/IR/ValueHandle.h"
75 #include "llvm/IR/ValueMap.h"
76 #include "llvm/Pass.h"
77 #include "llvm/Support/BlockFrequency.h"
78 #include "llvm/Support/BranchProbability.h"
79 #include "llvm/Support/Casting.h"
80 #include "llvm/Support/CommandLine.h"
81 #include "llvm/Support/Compiler.h"
82 #include "llvm/Support/Debug.h"
83 #include "llvm/Support/ErrorHandling.h"
84 #include "llvm/Support/MachineValueType.h"
85 #include "llvm/Support/MathExtras.h"
86 #include "llvm/Support/raw_ostream.h"
87 #include "llvm/Target/TargetMachine.h"
88 #include "llvm/Target/TargetOptions.h"
89 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
90 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
91 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
92 #include <algorithm>
93 #include <cassert>
94 #include <cstdint>
95 #include <iterator>
96 #include <limits>
97 #include <memory>
98 #include <utility>
99 #include <vector>
100
101 using namespace llvm;
102 using namespace llvm::PatternMatch;
103
104 #define DEBUG_TYPE "codegenprepare"
105
106 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
107 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
108 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
109 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
110 "sunken Cmps");
111 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
112 "of sunken Casts");
113 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
114 "computations were sunk");
115 STATISTIC(NumMemoryInstsPhiCreated,
116 "Number of phis created when address "
117 "computations were sunk to memory instructions");
118 STATISTIC(NumMemoryInstsSelectCreated,
119 "Number of select created when address "
120 "computations were sunk to memory instructions");
121 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
122 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
123 STATISTIC(NumAndsAdded,
124 "Number of and mask instructions added to form ext loads");
125 STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
126 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
127 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
128 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
129 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
130
131 static cl::opt<bool> DisableBranchOpts(
132 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
133 cl::desc("Disable branch optimizations in CodeGenPrepare"));
134
135 static cl::opt<bool>
136 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
137 cl::desc("Disable GC optimizations in CodeGenPrepare"));
138
139 static cl::opt<bool> DisableSelectToBranch(
140 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
141 cl::desc("Disable select to branch conversion."));
142
143 static cl::opt<bool> AddrSinkUsingGEPs(
144 "addr-sink-using-gep", cl::Hidden, cl::init(true),
145 cl::desc("Address sinking in CGP using GEPs."));
146
147 static cl::opt<bool> EnableAndCmpSinking(
148 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
149 cl::desc("Enable sinkinig and/cmp into branches."));
150
151 static cl::opt<bool> DisableStoreExtract(
152 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
153 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
154
155 static cl::opt<bool> StressStoreExtract(
156 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
157 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
158
159 static cl::opt<bool> DisableExtLdPromotion(
160 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
161 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
162 "CodeGenPrepare"));
163
164 static cl::opt<bool> StressExtLdPromotion(
165 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
166 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
167 "optimization in CodeGenPrepare"));
168
169 static cl::opt<bool> DisablePreheaderProtect(
170 "disable-preheader-prot", cl::Hidden, cl::init(false),
171 cl::desc("Disable protection against removing loop preheaders"));
172
173 static cl::opt<bool> ProfileGuidedSectionPrefix(
174 "profile-guided-section-prefix", cl::Hidden, cl::init(true), cl::ZeroOrMore,
175 cl::desc("Use profile info to add section prefix for hot/cold functions"));
176
177 static cl::opt<unsigned> FreqRatioToSkipMerge(
178 "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2),
179 cl::desc("Skip merging empty blocks if (frequency of empty block) / "
180 "(frequency of destination block) is greater than this ratio"));
181
182 static cl::opt<bool> ForceSplitStore(
183 "force-split-store", cl::Hidden, cl::init(false),
184 cl::desc("Force store splitting no matter what the target query says."));
185
186 static cl::opt<bool>
187 EnableTypePromotionMerge("cgp-type-promotion-merge", cl::Hidden,
188 cl::desc("Enable merging of redundant sexts when one is dominating"
189 " the other."), cl::init(true));
190
191 static cl::opt<bool> DisableComplexAddrModes(
192 "disable-complex-addr-modes", cl::Hidden, cl::init(false),
193 cl::desc("Disables combining addressing modes with different parts "
194 "in optimizeMemoryInst."));
195
196 static cl::opt<bool>
197 AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(false),
198 cl::desc("Allow creation of Phis in Address sinking."));
199
200 static cl::opt<bool>
201 AddrSinkNewSelects("addr-sink-new-select", cl::Hidden, cl::init(true),
202 cl::desc("Allow creation of selects in Address sinking."));
203
204 static cl::opt<bool> AddrSinkCombineBaseReg(
205 "addr-sink-combine-base-reg", cl::Hidden, cl::init(true),
206 cl::desc("Allow combining of BaseReg field in Address sinking."));
207
208 static cl::opt<bool> AddrSinkCombineBaseGV(
209 "addr-sink-combine-base-gv", cl::Hidden, cl::init(true),
210 cl::desc("Allow combining of BaseGV field in Address sinking."));
211
212 static cl::opt<bool> AddrSinkCombineBaseOffs(
213 "addr-sink-combine-base-offs", cl::Hidden, cl::init(true),
214 cl::desc("Allow combining of BaseOffs field in Address sinking."));
215
216 static cl::opt<bool> AddrSinkCombineScaledReg(
217 "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(true),
218 cl::desc("Allow combining of ScaledReg field in Address sinking."));
219
220 static cl::opt<bool>
221 EnableGEPOffsetSplit("cgp-split-large-offset-gep", cl::Hidden,
222 cl::init(true),
223 cl::desc("Enable splitting large offset of GEP."));
224
225 namespace {
226
227 enum ExtType {
228 ZeroExtension, // Zero extension has been seen.
229 SignExtension, // Sign extension has been seen.
230 BothExtension // This extension type is used if we saw sext after
231 // ZeroExtension had been set, or if we saw zext after
232 // SignExtension had been set. It makes the type
233 // information of a promoted instruction invalid.
234 };
235
236 using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
237 using TypeIsSExt = PointerIntPair<Type *, 2, ExtType>;
238 using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
239 using SExts = SmallVector<Instruction *, 16>;
240 using ValueToSExts = DenseMap<Value *, SExts>;
241
242 class TypePromotionTransaction;
243
244 class CodeGenPrepare : public FunctionPass {
245 const TargetMachine *TM = nullptr;
246 const TargetSubtargetInfo *SubtargetInfo;
247 const TargetLowering *TLI = nullptr;
248 const TargetRegisterInfo *TRI;
249 const TargetTransformInfo *TTI = nullptr;
250 const TargetLibraryInfo *TLInfo;
251 const LoopInfo *LI;
252 std::unique_ptr<BlockFrequencyInfo> BFI;
253 std::unique_ptr<BranchProbabilityInfo> BPI;
254
255 /// As we scan instructions optimizing them, this is the next instruction
256 /// to optimize. Transforms that can invalidate this should update it.
257 BasicBlock::iterator CurInstIterator;
258
259 /// Keeps track of non-local addresses that have been sunk into a block.
260 /// This allows us to avoid inserting duplicate code for blocks with
261 /// multiple load/stores of the same address. The usage of WeakTrackingVH
262 /// enables SunkAddrs to be treated as a cache whose entries can be
263 /// invalidated if a sunken address computation has been erased.
264 ValueMap<Value*, WeakTrackingVH> SunkAddrs;
265
266 /// Keeps track of all instructions inserted for the current function.
267 SetOfInstrs InsertedInsts;
268
269 /// Keeps track of the type of the related instruction before their
270 /// promotion for the current function.
271 InstrToOrigTy PromotedInsts;
272
273 /// Keep track of instructions removed during promotion.
274 SetOfInstrs RemovedInsts;
275
276 /// Keep track of sext chains based on their initial value.
277 DenseMap<Value *, Instruction *> SeenChainsForSExt;
278
279 /// Keep track of GEPs accessing the same data structures such as structs or
280 /// arrays that are candidates to be split later because of their large
281 /// size.
282 MapVector<
283 AssertingVH<Value>,
284 SmallVector<std::pair<AssertingVH<GetElementPtrInst>, int64_t>, 32>>
285 LargeOffsetGEPMap;
286
287 /// Keep track of new GEP base after splitting the GEPs having large offset.
288 SmallSet<AssertingVH<Value>, 2> NewGEPBases;
289
290 /// Map serial numbers to Large offset GEPs.
291 DenseMap<AssertingVH<GetElementPtrInst>, int> LargeOffsetGEPID;
292
293 /// Keep track of SExt promoted.
294 ValueToSExts ValToSExtendedUses;
295
296 /// True if optimizing for size.
297 bool OptSize;
298
299 /// DataLayout for the Function being processed.
300 const DataLayout *DL = nullptr;
301
302 /// Building the dominator tree can be expensive, so we only build it
303 /// lazily and update it when required.
304 std::unique_ptr<DominatorTree> DT;
305
306 public:
307 static char ID; // Pass identification, replacement for typeid
308
CodeGenPrepare()309 CodeGenPrepare() : FunctionPass(ID) {
310 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
311 }
312
313 bool runOnFunction(Function &F) override;
314
getPassName() const315 StringRef getPassName() const override { return "CodeGen Prepare"; }
316
getAnalysisUsage(AnalysisUsage & AU) const317 void getAnalysisUsage(AnalysisUsage &AU) const override {
318 // FIXME: When we can selectively preserve passes, preserve the domtree.
319 AU.addRequired<ProfileSummaryInfoWrapperPass>();
320 AU.addRequired<TargetLibraryInfoWrapperPass>();
321 AU.addRequired<TargetTransformInfoWrapperPass>();
322 AU.addRequired<LoopInfoWrapperPass>();
323 }
324
325 private:
326 template <typename F>
resetIteratorIfInvalidatedWhileCalling(BasicBlock * BB,F f)327 void resetIteratorIfInvalidatedWhileCalling(BasicBlock *BB, F f) {
328 // Substituting can cause recursive simplifications, which can invalidate
329 // our iterator. Use a WeakTrackingVH to hold onto it in case this
330 // happens.
331 Value *CurValue = &*CurInstIterator;
332 WeakTrackingVH IterHandle(CurValue);
333
334 f();
335
336 // If the iterator instruction was recursively deleted, start over at the
337 // start of the block.
338 if (IterHandle != CurValue) {
339 CurInstIterator = BB->begin();
340 SunkAddrs.clear();
341 }
342 }
343
344 // Get the DominatorTree, building if necessary.
getDT(Function & F)345 DominatorTree &getDT(Function &F) {
346 if (!DT)
347 DT = llvm::make_unique<DominatorTree>(F);
348 return *DT;
349 }
350
351 bool eliminateFallThrough(Function &F);
352 bool eliminateMostlyEmptyBlocks(Function &F);
353 BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
354 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
355 void eliminateMostlyEmptyBlock(BasicBlock *BB);
356 bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
357 bool isPreheader);
358 bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
359 bool optimizeInst(Instruction *I, bool &ModifiedDT);
360 bool optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
361 Type *AccessTy, unsigned AddrSpace);
362 bool optimizeInlineAsmInst(CallInst *CS);
363 bool optimizeCallInst(CallInst *CI, bool &ModifiedDT);
364 bool optimizeExt(Instruction *&I);
365 bool optimizeExtUses(Instruction *I);
366 bool optimizeLoadExt(LoadInst *Load);
367 bool optimizeShiftInst(BinaryOperator *BO);
368 bool optimizeSelectInst(SelectInst *SI);
369 bool optimizeShuffleVectorInst(ShuffleVectorInst *SVI);
370 bool optimizeSwitchInst(SwitchInst *SI);
371 bool optimizeExtractElementInst(Instruction *Inst);
372 bool dupRetToEnableTailCallOpts(BasicBlock *BB, bool &ModifiedDT);
373 bool placeDbgValues(Function &F);
374 bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
375 LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
376 bool tryToPromoteExts(TypePromotionTransaction &TPT,
377 const SmallVectorImpl<Instruction *> &Exts,
378 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
379 unsigned CreatedInstsCost = 0);
380 bool mergeSExts(Function &F);
381 bool splitLargeGEPOffsets();
382 bool performAddressTypePromotion(
383 Instruction *&Inst,
384 bool AllowPromotionWithoutCommonHeader,
385 bool HasPromoted, TypePromotionTransaction &TPT,
386 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
387 bool splitBranchCondition(Function &F, bool &ModifiedDT);
388 bool simplifyOffsetableRelocate(Instruction &I);
389
390 bool tryToSinkFreeOperands(Instruction *I);
391 bool replaceMathCmpWithIntrinsic(BinaryOperator *BO, CmpInst *Cmp,
392 Intrinsic::ID IID);
393 bool optimizeCmp(CmpInst *Cmp, bool &ModifiedDT);
394 bool combineToUSubWithOverflow(CmpInst *Cmp, bool &ModifiedDT);
395 bool combineToUAddWithOverflow(CmpInst *Cmp, bool &ModifiedDT);
396 };
397
398 } // end anonymous namespace
399
400 char CodeGenPrepare::ID = 0;
401
402 INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,
403 "Optimize for code generation", false, false)
INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)404 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
405 INITIALIZE_PASS_END(CodeGenPrepare, DEBUG_TYPE,
406 "Optimize for code generation", false, false)
407
408 FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); }
409
runOnFunction(Function & F)410 bool CodeGenPrepare::runOnFunction(Function &F) {
411 if (skipFunction(F))
412 return false;
413
414 DL = &F.getParent()->getDataLayout();
415
416 bool EverMadeChange = false;
417 // Clear per function information.
418 InsertedInsts.clear();
419 PromotedInsts.clear();
420
421 if (auto *TPC = getAnalysisIfAvailable<TargetPassConfig>()) {
422 TM = &TPC->getTM<TargetMachine>();
423 SubtargetInfo = TM->getSubtargetImpl(F);
424 TLI = SubtargetInfo->getTargetLowering();
425 TRI = SubtargetInfo->getRegisterInfo();
426 }
427 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
428 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
429 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
430 BPI.reset(new BranchProbabilityInfo(F, *LI));
431 BFI.reset(new BlockFrequencyInfo(F, *BPI, *LI));
432 OptSize = F.hasOptSize();
433
434 ProfileSummaryInfo *PSI =
435 &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
436 if (ProfileGuidedSectionPrefix) {
437 if (PSI->isFunctionHotInCallGraph(&F, *BFI))
438 F.setSectionPrefix(".hot");
439 else if (PSI->isFunctionColdInCallGraph(&F, *BFI))
440 F.setSectionPrefix(".unlikely");
441 }
442
443 /// This optimization identifies DIV instructions that can be
444 /// profitably bypassed and carried out with a shorter, faster divide.
445 if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI &&
446 TLI->isSlowDivBypassed()) {
447 const DenseMap<unsigned int, unsigned int> &BypassWidths =
448 TLI->getBypassSlowDivWidths();
449 BasicBlock* BB = &*F.begin();
450 while (BB != nullptr) {
451 // bypassSlowDivision may create new BBs, but we don't want to reapply the
452 // optimization to those blocks.
453 BasicBlock* Next = BB->getNextNode();
454 EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
455 BB = Next;
456 }
457 }
458
459 // Eliminate blocks that contain only PHI nodes and an
460 // unconditional branch.
461 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
462
463 bool ModifiedDT = false;
464 if (!DisableBranchOpts)
465 EverMadeChange |= splitBranchCondition(F, ModifiedDT);
466
467 // Split some critical edges where one of the sources is an indirect branch,
468 // to help generate sane code for PHIs involving such edges.
469 EverMadeChange |= SplitIndirectBrCriticalEdges(F);
470
471 bool MadeChange = true;
472 while (MadeChange) {
473 MadeChange = false;
474 DT.reset();
475 for (Function::iterator I = F.begin(); I != F.end(); ) {
476 BasicBlock *BB = &*I++;
477 bool ModifiedDTOnIteration = false;
478 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
479
480 // Restart BB iteration if the dominator tree of the Function was changed
481 if (ModifiedDTOnIteration)
482 break;
483 }
484 if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
485 MadeChange |= mergeSExts(F);
486 if (!LargeOffsetGEPMap.empty())
487 MadeChange |= splitLargeGEPOffsets();
488
489 // Really free removed instructions during promotion.
490 for (Instruction *I : RemovedInsts)
491 I->deleteValue();
492
493 EverMadeChange |= MadeChange;
494 SeenChainsForSExt.clear();
495 ValToSExtendedUses.clear();
496 RemovedInsts.clear();
497 LargeOffsetGEPMap.clear();
498 LargeOffsetGEPID.clear();
499 }
500
501 SunkAddrs.clear();
502
503 if (!DisableBranchOpts) {
504 MadeChange = false;
505 // Use a set vector to get deterministic iteration order. The order the
506 // blocks are removed may affect whether or not PHI nodes in successors
507 // are removed.
508 SmallSetVector<BasicBlock*, 8> WorkList;
509 for (BasicBlock &BB : F) {
510 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
511 MadeChange |= ConstantFoldTerminator(&BB, true);
512 if (!MadeChange) continue;
513
514 for (SmallVectorImpl<BasicBlock*>::iterator
515 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
516 if (pred_begin(*II) == pred_end(*II))
517 WorkList.insert(*II);
518 }
519
520 // Delete the dead blocks and any of their dead successors.
521 MadeChange |= !WorkList.empty();
522 while (!WorkList.empty()) {
523 BasicBlock *BB = WorkList.pop_back_val();
524 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
525
526 DeleteDeadBlock(BB);
527
528 for (SmallVectorImpl<BasicBlock*>::iterator
529 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
530 if (pred_begin(*II) == pred_end(*II))
531 WorkList.insert(*II);
532 }
533
534 // Merge pairs of basic blocks with unconditional branches, connected by
535 // a single edge.
536 if (EverMadeChange || MadeChange)
537 MadeChange |= eliminateFallThrough(F);
538
539 EverMadeChange |= MadeChange;
540 }
541
542 if (!DisableGCOpts) {
543 SmallVector<Instruction *, 2> Statepoints;
544 for (BasicBlock &BB : F)
545 for (Instruction &I : BB)
546 if (isStatepoint(I))
547 Statepoints.push_back(&I);
548 for (auto &I : Statepoints)
549 EverMadeChange |= simplifyOffsetableRelocate(*I);
550 }
551
552 // Do this last to clean up use-before-def scenarios introduced by other
553 // preparatory transforms.
554 EverMadeChange |= placeDbgValues(F);
555
556 return EverMadeChange;
557 }
558
559 /// Merge basic blocks which are connected by a single edge, where one of the
560 /// basic blocks has a single successor pointing to the other basic block,
561 /// which has a single predecessor.
eliminateFallThrough(Function & F)562 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
563 bool Changed = false;
564 // Scan all of the blocks in the function, except for the entry block.
565 // Use a temporary array to avoid iterator being invalidated when
566 // deleting blocks.
567 SmallVector<WeakTrackingVH, 16> Blocks;
568 for (auto &Block : llvm::make_range(std::next(F.begin()), F.end()))
569 Blocks.push_back(&Block);
570
571 for (auto &Block : Blocks) {
572 auto *BB = cast_or_null<BasicBlock>(Block);
573 if (!BB)
574 continue;
575 // If the destination block has a single pred, then this is a trivial
576 // edge, just collapse it.
577 BasicBlock *SinglePred = BB->getSinglePredecessor();
578
579 // Don't merge if BB's address is taken.
580 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
581
582 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
583 if (Term && !Term->isConditional()) {
584 Changed = true;
585 LLVM_DEBUG(dbgs() << "To merge:\n" << *BB << "\n\n\n");
586
587 // Merge BB into SinglePred and delete it.
588 MergeBlockIntoPredecessor(BB);
589 }
590 }
591 return Changed;
592 }
593
594 /// Find a destination block from BB if BB is mergeable empty block.
findDestBlockOfMergeableEmptyBlock(BasicBlock * BB)595 BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
596 // If this block doesn't end with an uncond branch, ignore it.
597 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
598 if (!BI || !BI->isUnconditional())
599 return nullptr;
600
601 // If the instruction before the branch (skipping debug info) isn't a phi
602 // node, then other stuff is happening here.
603 BasicBlock::iterator BBI = BI->getIterator();
604 if (BBI != BB->begin()) {
605 --BBI;
606 while (isa<DbgInfoIntrinsic>(BBI)) {
607 if (BBI == BB->begin())
608 break;
609 --BBI;
610 }
611 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
612 return nullptr;
613 }
614
615 // Do not break infinite loops.
616 BasicBlock *DestBB = BI->getSuccessor(0);
617 if (DestBB == BB)
618 return nullptr;
619
620 if (!canMergeBlocks(BB, DestBB))
621 DestBB = nullptr;
622
623 return DestBB;
624 }
625
626 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
627 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
628 /// edges in ways that are non-optimal for isel. Start by eliminating these
629 /// blocks so we can split them the way we want them.
eliminateMostlyEmptyBlocks(Function & F)630 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
631 SmallPtrSet<BasicBlock *, 16> Preheaders;
632 SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
633 while (!LoopList.empty()) {
634 Loop *L = LoopList.pop_back_val();
635 LoopList.insert(LoopList.end(), L->begin(), L->end());
636 if (BasicBlock *Preheader = L->getLoopPreheader())
637 Preheaders.insert(Preheader);
638 }
639
640 bool MadeChange = false;
641 // Copy blocks into a temporary array to avoid iterator invalidation issues
642 // as we remove them.
643 // Note that this intentionally skips the entry block.
644 SmallVector<WeakTrackingVH, 16> Blocks;
645 for (auto &Block : llvm::make_range(std::next(F.begin()), F.end()))
646 Blocks.push_back(&Block);
647
648 for (auto &Block : Blocks) {
649 BasicBlock *BB = cast_or_null<BasicBlock>(Block);
650 if (!BB)
651 continue;
652 BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
653 if (!DestBB ||
654 !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB)))
655 continue;
656
657 eliminateMostlyEmptyBlock(BB);
658 MadeChange = true;
659 }
660 return MadeChange;
661 }
662
isMergingEmptyBlockProfitable(BasicBlock * BB,BasicBlock * DestBB,bool isPreheader)663 bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
664 BasicBlock *DestBB,
665 bool isPreheader) {
666 // Do not delete loop preheaders if doing so would create a critical edge.
667 // Loop preheaders can be good locations to spill registers. If the
668 // preheader is deleted and we create a critical edge, registers may be
669 // spilled in the loop body instead.
670 if (!DisablePreheaderProtect && isPreheader &&
671 !(BB->getSinglePredecessor() &&
672 BB->getSinglePredecessor()->getSingleSuccessor()))
673 return false;
674
675 // Skip merging if the block's successor is also a successor to any callbr
676 // that leads to this block.
677 // FIXME: Is this really needed? Is this a correctness issue?
678 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
679 if (auto *CBI = dyn_cast<CallBrInst>((*PI)->getTerminator()))
680 for (unsigned i = 0, e = CBI->getNumSuccessors(); i != e; ++i)
681 if (DestBB == CBI->getSuccessor(i))
682 return false;
683 }
684
685 // Try to skip merging if the unique predecessor of BB is terminated by a
686 // switch or indirect branch instruction, and BB is used as an incoming block
687 // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
688 // add COPY instructions in the predecessor of BB instead of BB (if it is not
689 // merged). Note that the critical edge created by merging such blocks wont be
690 // split in MachineSink because the jump table is not analyzable. By keeping
691 // such empty block (BB), ISel will place COPY instructions in BB, not in the
692 // predecessor of BB.
693 BasicBlock *Pred = BB->getUniquePredecessor();
694 if (!Pred ||
695 !(isa<SwitchInst>(Pred->getTerminator()) ||
696 isa<IndirectBrInst>(Pred->getTerminator())))
697 return true;
698
699 if (BB->getTerminator() != BB->getFirstNonPHIOrDbg())
700 return true;
701
702 // We use a simple cost heuristic which determine skipping merging is
703 // profitable if the cost of skipping merging is less than the cost of
704 // merging : Cost(skipping merging) < Cost(merging BB), where the
705 // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
706 // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
707 // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
708 // Freq(Pred) / Freq(BB) > 2.
709 // Note that if there are multiple empty blocks sharing the same incoming
710 // value for the PHIs in the DestBB, we consider them together. In such
711 // case, Cost(merging BB) will be the sum of their frequencies.
712
713 if (!isa<PHINode>(DestBB->begin()))
714 return true;
715
716 SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
717
718 // Find all other incoming blocks from which incoming values of all PHIs in
719 // DestBB are the same as the ones from BB.
720 for (pred_iterator PI = pred_begin(DestBB), E = pred_end(DestBB); PI != E;
721 ++PI) {
722 BasicBlock *DestBBPred = *PI;
723 if (DestBBPred == BB)
724 continue;
725
726 if (llvm::all_of(DestBB->phis(), [&](const PHINode &DestPN) {
727 return DestPN.getIncomingValueForBlock(BB) ==
728 DestPN.getIncomingValueForBlock(DestBBPred);
729 }))
730 SameIncomingValueBBs.insert(DestBBPred);
731 }
732
733 // See if all BB's incoming values are same as the value from Pred. In this
734 // case, no reason to skip merging because COPYs are expected to be place in
735 // Pred already.
736 if (SameIncomingValueBBs.count(Pred))
737 return true;
738
739 BlockFrequency PredFreq = BFI->getBlockFreq(Pred);
740 BlockFrequency BBFreq = BFI->getBlockFreq(BB);
741
742 for (auto SameValueBB : SameIncomingValueBBs)
743 if (SameValueBB->getUniquePredecessor() == Pred &&
744 DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB))
745 BBFreq += BFI->getBlockFreq(SameValueBB);
746
747 return PredFreq.getFrequency() <=
748 BBFreq.getFrequency() * FreqRatioToSkipMerge;
749 }
750
751 /// Return true if we can merge BB into DestBB if there is a single
752 /// unconditional branch between them, and BB contains no other non-phi
753 /// instructions.
canMergeBlocks(const BasicBlock * BB,const BasicBlock * DestBB) const754 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
755 const BasicBlock *DestBB) const {
756 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
757 // the successor. If there are more complex condition (e.g. preheaders),
758 // don't mess around with them.
759 for (const PHINode &PN : BB->phis()) {
760 for (const User *U : PN.users()) {
761 const Instruction *UI = cast<Instruction>(U);
762 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
763 return false;
764 // If User is inside DestBB block and it is a PHINode then check
765 // incoming value. If incoming value is not from BB then this is
766 // a complex condition (e.g. preheaders) we want to avoid here.
767 if (UI->getParent() == DestBB) {
768 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
769 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
770 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
771 if (Insn && Insn->getParent() == BB &&
772 Insn->getParent() != UPN->getIncomingBlock(I))
773 return false;
774 }
775 }
776 }
777 }
778
779 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
780 // and DestBB may have conflicting incoming values for the block. If so, we
781 // can't merge the block.
782 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
783 if (!DestBBPN) return true; // no conflict.
784
785 // Collect the preds of BB.
786 SmallPtrSet<const BasicBlock*, 16> BBPreds;
787 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
788 // It is faster to get preds from a PHI than with pred_iterator.
789 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
790 BBPreds.insert(BBPN->getIncomingBlock(i));
791 } else {
792 BBPreds.insert(pred_begin(BB), pred_end(BB));
793 }
794
795 // Walk the preds of DestBB.
796 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
797 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
798 if (BBPreds.count(Pred)) { // Common predecessor?
799 for (const PHINode &PN : DestBB->phis()) {
800 const Value *V1 = PN.getIncomingValueForBlock(Pred);
801 const Value *V2 = PN.getIncomingValueForBlock(BB);
802
803 // If V2 is a phi node in BB, look up what the mapped value will be.
804 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
805 if (V2PN->getParent() == BB)
806 V2 = V2PN->getIncomingValueForBlock(Pred);
807
808 // If there is a conflict, bail out.
809 if (V1 != V2) return false;
810 }
811 }
812 }
813
814 return true;
815 }
816
817 /// Eliminate a basic block that has only phi's and an unconditional branch in
818 /// it.
eliminateMostlyEmptyBlock(BasicBlock * BB)819 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
820 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
821 BasicBlock *DestBB = BI->getSuccessor(0);
822
823 LLVM_DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"
824 << *BB << *DestBB);
825
826 // If the destination block has a single pred, then this is a trivial edge,
827 // just collapse it.
828 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
829 if (SinglePred != DestBB) {
830 assert(SinglePred == BB &&
831 "Single predecessor not the same as predecessor");
832 // Merge DestBB into SinglePred/BB and delete it.
833 MergeBlockIntoPredecessor(DestBB);
834 // Note: BB(=SinglePred) will not be deleted on this path.
835 // DestBB(=its single successor) is the one that was deleted.
836 LLVM_DEBUG(dbgs() << "AFTER:\n" << *SinglePred << "\n\n\n");
837 return;
838 }
839 }
840
841 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
842 // to handle the new incoming edges it is about to have.
843 for (PHINode &PN : DestBB->phis()) {
844 // Remove the incoming value for BB, and remember it.
845 Value *InVal = PN.removeIncomingValue(BB, false);
846
847 // Two options: either the InVal is a phi node defined in BB or it is some
848 // value that dominates BB.
849 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
850 if (InValPhi && InValPhi->getParent() == BB) {
851 // Add all of the input values of the input PHI as inputs of this phi.
852 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
853 PN.addIncoming(InValPhi->getIncomingValue(i),
854 InValPhi->getIncomingBlock(i));
855 } else {
856 // Otherwise, add one instance of the dominating value for each edge that
857 // we will be adding.
858 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
859 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
860 PN.addIncoming(InVal, BBPN->getIncomingBlock(i));
861 } else {
862 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
863 PN.addIncoming(InVal, *PI);
864 }
865 }
866 }
867
868 // The PHIs are now updated, change everything that refers to BB to use
869 // DestBB and remove BB.
870 BB->replaceAllUsesWith(DestBB);
871 BB->eraseFromParent();
872 ++NumBlocksElim;
873
874 LLVM_DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
875 }
876
877 // Computes a map of base pointer relocation instructions to corresponding
878 // derived pointer relocation instructions given a vector of all relocate calls
computeBaseDerivedRelocateMap(const SmallVectorImpl<GCRelocateInst * > & AllRelocateCalls,DenseMap<GCRelocateInst *,SmallVector<GCRelocateInst *,2>> & RelocateInstMap)879 static void computeBaseDerivedRelocateMap(
880 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
881 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
882 &RelocateInstMap) {
883 // Collect information in two maps: one primarily for locating the base object
884 // while filling the second map; the second map is the final structure holding
885 // a mapping between Base and corresponding Derived relocate calls
886 DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
887 for (auto *ThisRelocate : AllRelocateCalls) {
888 auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
889 ThisRelocate->getDerivedPtrIndex());
890 RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
891 }
892 for (auto &Item : RelocateIdxMap) {
893 std::pair<unsigned, unsigned> Key = Item.first;
894 if (Key.first == Key.second)
895 // Base relocation: nothing to insert
896 continue;
897
898 GCRelocateInst *I = Item.second;
899 auto BaseKey = std::make_pair(Key.first, Key.first);
900
901 // We're iterating over RelocateIdxMap so we cannot modify it.
902 auto MaybeBase = RelocateIdxMap.find(BaseKey);
903 if (MaybeBase == RelocateIdxMap.end())
904 // TODO: We might want to insert a new base object relocate and gep off
905 // that, if there are enough derived object relocates.
906 continue;
907
908 RelocateInstMap[MaybeBase->second].push_back(I);
909 }
910 }
911
912 // Accepts a GEP and extracts the operands into a vector provided they're all
913 // small integer constants
getGEPSmallConstantIntOffsetV(GetElementPtrInst * GEP,SmallVectorImpl<Value * > & OffsetV)914 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
915 SmallVectorImpl<Value *> &OffsetV) {
916 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
917 // Only accept small constant integer operands
918 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
919 if (!Op || Op->getZExtValue() > 20)
920 return false;
921 }
922
923 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
924 OffsetV.push_back(GEP->getOperand(i));
925 return true;
926 }
927
928 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
929 // replace, computes a replacement, and affects it.
930 static bool
simplifyRelocatesOffABase(GCRelocateInst * RelocatedBase,const SmallVectorImpl<GCRelocateInst * > & Targets)931 simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
932 const SmallVectorImpl<GCRelocateInst *> &Targets) {
933 bool MadeChange = false;
934 // We must ensure the relocation of derived pointer is defined after
935 // relocation of base pointer. If we find a relocation corresponding to base
936 // defined earlier than relocation of base then we move relocation of base
937 // right before found relocation. We consider only relocation in the same
938 // basic block as relocation of base. Relocations from other basic block will
939 // be skipped by optimization and we do not care about them.
940 for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
941 &*R != RelocatedBase; ++R)
942 if (auto RI = dyn_cast<GCRelocateInst>(R))
943 if (RI->getStatepoint() == RelocatedBase->getStatepoint())
944 if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
945 RelocatedBase->moveBefore(RI);
946 break;
947 }
948
949 for (GCRelocateInst *ToReplace : Targets) {
950 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
951 "Not relocating a derived object of the original base object");
952 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
953 // A duplicate relocate call. TODO: coalesce duplicates.
954 continue;
955 }
956
957 if (RelocatedBase->getParent() != ToReplace->getParent()) {
958 // Base and derived relocates are in different basic blocks.
959 // In this case transform is only valid when base dominates derived
960 // relocate. However it would be too expensive to check dominance
961 // for each such relocate, so we skip the whole transformation.
962 continue;
963 }
964
965 Value *Base = ToReplace->getBasePtr();
966 auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
967 if (!Derived || Derived->getPointerOperand() != Base)
968 continue;
969
970 SmallVector<Value *, 2> OffsetV;
971 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
972 continue;
973
974 // Create a Builder and replace the target callsite with a gep
975 assert(RelocatedBase->getNextNode() &&
976 "Should always have one since it's not a terminator");
977
978 // Insert after RelocatedBase
979 IRBuilder<> Builder(RelocatedBase->getNextNode());
980 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
981
982 // If gc_relocate does not match the actual type, cast it to the right type.
983 // In theory, there must be a bitcast after gc_relocate if the type does not
984 // match, and we should reuse it to get the derived pointer. But it could be
985 // cases like this:
986 // bb1:
987 // ...
988 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
989 // br label %merge
990 //
991 // bb2:
992 // ...
993 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
994 // br label %merge
995 //
996 // merge:
997 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
998 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
999 //
1000 // In this case, we can not find the bitcast any more. So we insert a new bitcast
1001 // no matter there is already one or not. In this way, we can handle all cases, and
1002 // the extra bitcast should be optimized away in later passes.
1003 Value *ActualRelocatedBase = RelocatedBase;
1004 if (RelocatedBase->getType() != Base->getType()) {
1005 ActualRelocatedBase =
1006 Builder.CreateBitCast(RelocatedBase, Base->getType());
1007 }
1008 Value *Replacement = Builder.CreateGEP(
1009 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
1010 Replacement->takeName(ToReplace);
1011 // If the newly generated derived pointer's type does not match the original derived
1012 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
1013 Value *ActualReplacement = Replacement;
1014 if (Replacement->getType() != ToReplace->getType()) {
1015 ActualReplacement =
1016 Builder.CreateBitCast(Replacement, ToReplace->getType());
1017 }
1018 ToReplace->replaceAllUsesWith(ActualReplacement);
1019 ToReplace->eraseFromParent();
1020
1021 MadeChange = true;
1022 }
1023 return MadeChange;
1024 }
1025
1026 // Turns this:
1027 //
1028 // %base = ...
1029 // %ptr = gep %base + 15
1030 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1031 // %base' = relocate(%tok, i32 4, i32 4)
1032 // %ptr' = relocate(%tok, i32 4, i32 5)
1033 // %val = load %ptr'
1034 //
1035 // into this:
1036 //
1037 // %base = ...
1038 // %ptr = gep %base + 15
1039 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1040 // %base' = gc.relocate(%tok, i32 4, i32 4)
1041 // %ptr' = gep %base' + 15
1042 // %val = load %ptr'
simplifyOffsetableRelocate(Instruction & I)1043 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
1044 bool MadeChange = false;
1045 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1046
1047 for (auto *U : I.users())
1048 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1049 // Collect all the relocate calls associated with a statepoint
1050 AllRelocateCalls.push_back(Relocate);
1051
1052 // We need atleast one base pointer relocation + one derived pointer
1053 // relocation to mangle
1054 if (AllRelocateCalls.size() < 2)
1055 return false;
1056
1057 // RelocateInstMap is a mapping from the base relocate instruction to the
1058 // corresponding derived relocate instructions
1059 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
1060 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1061 if (RelocateInstMap.empty())
1062 return false;
1063
1064 for (auto &Item : RelocateInstMap)
1065 // Item.first is the RelocatedBase to offset against
1066 // Item.second is the vector of Targets to replace
1067 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1068 return MadeChange;
1069 }
1070
1071 /// Sink the specified cast instruction into its user blocks.
SinkCast(CastInst * CI)1072 static bool SinkCast(CastInst *CI) {
1073 BasicBlock *DefBB = CI->getParent();
1074
1075 /// InsertedCasts - Only insert a cast in each block once.
1076 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1077
1078 bool MadeChange = false;
1079 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1080 UI != E; ) {
1081 Use &TheUse = UI.getUse();
1082 Instruction *User = cast<Instruction>(*UI);
1083
1084 // Figure out which BB this cast is used in. For PHI's this is the
1085 // appropriate predecessor block.
1086 BasicBlock *UserBB = User->getParent();
1087 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1088 UserBB = PN->getIncomingBlock(TheUse);
1089 }
1090
1091 // Preincrement use iterator so we don't invalidate it.
1092 ++UI;
1093
1094 // The first insertion point of a block containing an EH pad is after the
1095 // pad. If the pad is the user, we cannot sink the cast past the pad.
1096 if (User->isEHPad())
1097 continue;
1098
1099 // If the block selected to receive the cast is an EH pad that does not
1100 // allow non-PHI instructions before the terminator, we can't sink the
1101 // cast.
1102 if (UserBB->getTerminator()->isEHPad())
1103 continue;
1104
1105 // If this user is in the same block as the cast, don't change the cast.
1106 if (UserBB == DefBB) continue;
1107
1108 // If we have already inserted a cast into this block, use it.
1109 CastInst *&InsertedCast = InsertedCasts[UserBB];
1110
1111 if (!InsertedCast) {
1112 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1113 assert(InsertPt != UserBB->end());
1114 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1115 CI->getType(), "", &*InsertPt);
1116 InsertedCast->setDebugLoc(CI->getDebugLoc());
1117 }
1118
1119 // Replace a use of the cast with a use of the new cast.
1120 TheUse = InsertedCast;
1121 MadeChange = true;
1122 ++NumCastUses;
1123 }
1124
1125 // If we removed all uses, nuke the cast.
1126 if (CI->use_empty()) {
1127 salvageDebugInfo(*CI);
1128 CI->eraseFromParent();
1129 MadeChange = true;
1130 }
1131
1132 return MadeChange;
1133 }
1134
1135 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1136 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1137 /// reduce the number of virtual registers that must be created and coalesced.
1138 ///
1139 /// Return true if any changes are made.
OptimizeNoopCopyExpression(CastInst * CI,const TargetLowering & TLI,const DataLayout & DL)1140 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1141 const DataLayout &DL) {
1142 // Sink only "cheap" (or nop) address-space casts. This is a weaker condition
1143 // than sinking only nop casts, but is helpful on some platforms.
1144 if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) {
1145 if (!TLI.isFreeAddrSpaceCast(ASC->getSrcAddressSpace(),
1146 ASC->getDestAddressSpace()))
1147 return false;
1148 }
1149
1150 // If this is a noop copy,
1151 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1152 EVT DstVT = TLI.getValueType(DL, CI->getType());
1153
1154 // This is an fp<->int conversion?
1155 if (SrcVT.isInteger() != DstVT.isInteger())
1156 return false;
1157
1158 // If this is an extension, it will be a zero or sign extension, which
1159 // isn't a noop.
1160 if (SrcVT.bitsLT(DstVT)) return false;
1161
1162 // If these values will be promoted, find out what they will be promoted
1163 // to. This helps us consider truncates on PPC as noop copies when they
1164 // are.
1165 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1166 TargetLowering::TypePromoteInteger)
1167 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1168 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1169 TargetLowering::TypePromoteInteger)
1170 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1171
1172 // If, after promotion, these are the same types, this is a noop copy.
1173 if (SrcVT != DstVT)
1174 return false;
1175
1176 return SinkCast(CI);
1177 }
1178
replaceMathCmpWithIntrinsic(BinaryOperator * BO,CmpInst * Cmp,Intrinsic::ID IID)1179 bool CodeGenPrepare::replaceMathCmpWithIntrinsic(BinaryOperator *BO,
1180 CmpInst *Cmp,
1181 Intrinsic::ID IID) {
1182 if (BO->getParent() != Cmp->getParent()) {
1183 // We used to use a dominator tree here to allow multi-block optimization.
1184 // But that was problematic because:
1185 // 1. It could cause a perf regression by hoisting the math op into the
1186 // critical path.
1187 // 2. It could cause a perf regression by creating a value that was live
1188 // across multiple blocks and increasing register pressure.
1189 // 3. Use of a dominator tree could cause large compile-time regression.
1190 // This is because we recompute the DT on every change in the main CGP
1191 // run-loop. The recomputing is probably unnecessary in many cases, so if
1192 // that was fixed, using a DT here would be ok.
1193 return false;
1194 }
1195
1196 // We allow matching the canonical IR (add X, C) back to (usubo X, -C).
1197 Value *Arg0 = BO->getOperand(0);
1198 Value *Arg1 = BO->getOperand(1);
1199 if (BO->getOpcode() == Instruction::Add &&
1200 IID == Intrinsic::usub_with_overflow) {
1201 assert(isa<Constant>(Arg1) && "Unexpected input for usubo");
1202 Arg1 = ConstantExpr::getNeg(cast<Constant>(Arg1));
1203 }
1204
1205 // Insert at the first instruction of the pair.
1206 Instruction *InsertPt = nullptr;
1207 for (Instruction &Iter : *Cmp->getParent()) {
1208 if (&Iter == BO || &Iter == Cmp) {
1209 InsertPt = &Iter;
1210 break;
1211 }
1212 }
1213 assert(InsertPt != nullptr && "Parent block did not contain cmp or binop");
1214
1215 IRBuilder<> Builder(InsertPt);
1216 Value *MathOV = Builder.CreateBinaryIntrinsic(IID, Arg0, Arg1);
1217 Value *Math = Builder.CreateExtractValue(MathOV, 0, "math");
1218 Value *OV = Builder.CreateExtractValue(MathOV, 1, "ov");
1219 BO->replaceAllUsesWith(Math);
1220 Cmp->replaceAllUsesWith(OV);
1221 BO->eraseFromParent();
1222 Cmp->eraseFromParent();
1223 return true;
1224 }
1225
1226 /// Match special-case patterns that check for unsigned add overflow.
matchUAddWithOverflowConstantEdgeCases(CmpInst * Cmp,BinaryOperator * & Add)1227 static bool matchUAddWithOverflowConstantEdgeCases(CmpInst *Cmp,
1228 BinaryOperator *&Add) {
1229 // Add = add A, 1; Cmp = icmp eq A,-1 (overflow if A is max val)
1230 // Add = add A,-1; Cmp = icmp ne A, 0 (overflow if A is non-zero)
1231 Value *A = Cmp->getOperand(0), *B = Cmp->getOperand(1);
1232
1233 // We are not expecting non-canonical/degenerate code. Just bail out.
1234 if (isa<Constant>(A))
1235 return false;
1236
1237 ICmpInst::Predicate Pred = Cmp->getPredicate();
1238 if (Pred == ICmpInst::ICMP_EQ && match(B, m_AllOnes()))
1239 B = ConstantInt::get(B->getType(), 1);
1240 else if (Pred == ICmpInst::ICMP_NE && match(B, m_ZeroInt()))
1241 B = ConstantInt::get(B->getType(), -1);
1242 else
1243 return false;
1244
1245 // Check the users of the variable operand of the compare looking for an add
1246 // with the adjusted constant.
1247 for (User *U : A->users()) {
1248 if (match(U, m_Add(m_Specific(A), m_Specific(B)))) {
1249 Add = cast<BinaryOperator>(U);
1250 return true;
1251 }
1252 }
1253 return false;
1254 }
1255
1256 /// Try to combine the compare into a call to the llvm.uadd.with.overflow
1257 /// intrinsic. Return true if any changes were made.
combineToUAddWithOverflow(CmpInst * Cmp,bool & ModifiedDT)1258 bool CodeGenPrepare::combineToUAddWithOverflow(CmpInst *Cmp,
1259 bool &ModifiedDT) {
1260 Value *A, *B;
1261 BinaryOperator *Add;
1262 if (!match(Cmp, m_UAddWithOverflow(m_Value(A), m_Value(B), m_BinOp(Add))))
1263 if (!matchUAddWithOverflowConstantEdgeCases(Cmp, Add))
1264 return false;
1265
1266 if (!TLI->shouldFormOverflowOp(ISD::UADDO,
1267 TLI->getValueType(*DL, Add->getType())))
1268 return false;
1269
1270 // We don't want to move around uses of condition values this late, so we
1271 // check if it is legal to create the call to the intrinsic in the basic
1272 // block containing the icmp.
1273 if (Add->getParent() != Cmp->getParent() && !Add->hasOneUse())
1274 return false;
1275
1276 if (!replaceMathCmpWithIntrinsic(Add, Cmp, Intrinsic::uadd_with_overflow))
1277 return false;
1278
1279 // Reset callers - do not crash by iterating over a dead instruction.
1280 ModifiedDT = true;
1281 return true;
1282 }
1283
combineToUSubWithOverflow(CmpInst * Cmp,bool & ModifiedDT)1284 bool CodeGenPrepare::combineToUSubWithOverflow(CmpInst *Cmp,
1285 bool &ModifiedDT) {
1286 // We are not expecting non-canonical/degenerate code. Just bail out.
1287 Value *A = Cmp->getOperand(0), *B = Cmp->getOperand(1);
1288 if (isa<Constant>(A) && isa<Constant>(B))
1289 return false;
1290
1291 // Convert (A u> B) to (A u< B) to simplify pattern matching.
1292 ICmpInst::Predicate Pred = Cmp->getPredicate();
1293 if (Pred == ICmpInst::ICMP_UGT) {
1294 std::swap(A, B);
1295 Pred = ICmpInst::ICMP_ULT;
1296 }
1297 // Convert special-case: (A == 0) is the same as (A u< 1).
1298 if (Pred == ICmpInst::ICMP_EQ && match(B, m_ZeroInt())) {
1299 B = ConstantInt::get(B->getType(), 1);
1300 Pred = ICmpInst::ICMP_ULT;
1301 }
1302 // Convert special-case: (A != 0) is the same as (0 u< A).
1303 if (Pred == ICmpInst::ICMP_NE && match(B, m_ZeroInt())) {
1304 std::swap(A, B);
1305 Pred = ICmpInst::ICMP_ULT;
1306 }
1307 if (Pred != ICmpInst::ICMP_ULT)
1308 return false;
1309
1310 // Walk the users of a variable operand of a compare looking for a subtract or
1311 // add with that same operand. Also match the 2nd operand of the compare to
1312 // the add/sub, but that may be a negated constant operand of an add.
1313 Value *CmpVariableOperand = isa<Constant>(A) ? B : A;
1314 BinaryOperator *Sub = nullptr;
1315 for (User *U : CmpVariableOperand->users()) {
1316 // A - B, A u< B --> usubo(A, B)
1317 if (match(U, m_Sub(m_Specific(A), m_Specific(B)))) {
1318 Sub = cast<BinaryOperator>(U);
1319 break;
1320 }
1321
1322 // A + (-C), A u< C (canonicalized form of (sub A, C))
1323 const APInt *CmpC, *AddC;
1324 if (match(U, m_Add(m_Specific(A), m_APInt(AddC))) &&
1325 match(B, m_APInt(CmpC)) && *AddC == -(*CmpC)) {
1326 Sub = cast<BinaryOperator>(U);
1327 break;
1328 }
1329 }
1330 if (!Sub)
1331 return false;
1332
1333 if (!TLI->shouldFormOverflowOp(ISD::USUBO,
1334 TLI->getValueType(*DL, Sub->getType())))
1335 return false;
1336
1337 if (!replaceMathCmpWithIntrinsic(Sub, Cmp, Intrinsic::usub_with_overflow))
1338 return false;
1339
1340 // Reset callers - do not crash by iterating over a dead instruction.
1341 ModifiedDT = true;
1342 return true;
1343 }
1344
1345 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1346 /// registers that must be created and coalesced. This is a clear win except on
1347 /// targets with multiple condition code registers (PowerPC), where it might
1348 /// lose; some adjustment may be wanted there.
1349 ///
1350 /// Return true if any changes are made.
sinkCmpExpression(CmpInst * Cmp,const TargetLowering & TLI)1351 static bool sinkCmpExpression(CmpInst *Cmp, const TargetLowering &TLI) {
1352 if (TLI.hasMultipleConditionRegisters())
1353 return false;
1354
1355 // Avoid sinking soft-FP comparisons, since this can move them into a loop.
1356 if (TLI.useSoftFloat() && isa<FCmpInst>(Cmp))
1357 return false;
1358
1359 // Only insert a cmp in each block once.
1360 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1361
1362 bool MadeChange = false;
1363 for (Value::user_iterator UI = Cmp->user_begin(), E = Cmp->user_end();
1364 UI != E; ) {
1365 Use &TheUse = UI.getUse();
1366 Instruction *User = cast<Instruction>(*UI);
1367
1368 // Preincrement use iterator so we don't invalidate it.
1369 ++UI;
1370
1371 // Don't bother for PHI nodes.
1372 if (isa<PHINode>(User))
1373 continue;
1374
1375 // Figure out which BB this cmp is used in.
1376 BasicBlock *UserBB = User->getParent();
1377 BasicBlock *DefBB = Cmp->getParent();
1378
1379 // If this user is in the same block as the cmp, don't change the cmp.
1380 if (UserBB == DefBB) continue;
1381
1382 // If we have already inserted a cmp into this block, use it.
1383 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1384
1385 if (!InsertedCmp) {
1386 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1387 assert(InsertPt != UserBB->end());
1388 InsertedCmp =
1389 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(),
1390 Cmp->getOperand(0), Cmp->getOperand(1), "",
1391 &*InsertPt);
1392 // Propagate the debug info.
1393 InsertedCmp->setDebugLoc(Cmp->getDebugLoc());
1394 }
1395
1396 // Replace a use of the cmp with a use of the new cmp.
1397 TheUse = InsertedCmp;
1398 MadeChange = true;
1399 ++NumCmpUses;
1400 }
1401
1402 // If we removed all uses, nuke the cmp.
1403 if (Cmp->use_empty()) {
1404 Cmp->eraseFromParent();
1405 MadeChange = true;
1406 }
1407
1408 return MadeChange;
1409 }
1410
optimizeCmp(CmpInst * Cmp,bool & ModifiedDT)1411 bool CodeGenPrepare::optimizeCmp(CmpInst *Cmp, bool &ModifiedDT) {
1412 if (sinkCmpExpression(Cmp, *TLI))
1413 return true;
1414
1415 if (combineToUAddWithOverflow(Cmp, ModifiedDT))
1416 return true;
1417
1418 if (combineToUSubWithOverflow(Cmp, ModifiedDT))
1419 return true;
1420
1421 return false;
1422 }
1423
1424 /// Duplicate and sink the given 'and' instruction into user blocks where it is
1425 /// used in a compare to allow isel to generate better code for targets where
1426 /// this operation can be combined.
1427 ///
1428 /// Return true if any changes are made.
sinkAndCmp0Expression(Instruction * AndI,const TargetLowering & TLI,SetOfInstrs & InsertedInsts)1429 static bool sinkAndCmp0Expression(Instruction *AndI,
1430 const TargetLowering &TLI,
1431 SetOfInstrs &InsertedInsts) {
1432 // Double-check that we're not trying to optimize an instruction that was
1433 // already optimized by some other part of this pass.
1434 assert(!InsertedInsts.count(AndI) &&
1435 "Attempting to optimize already optimized and instruction");
1436 (void) InsertedInsts;
1437
1438 // Nothing to do for single use in same basic block.
1439 if (AndI->hasOneUse() &&
1440 AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent())
1441 return false;
1442
1443 // Try to avoid cases where sinking/duplicating is likely to increase register
1444 // pressure.
1445 if (!isa<ConstantInt>(AndI->getOperand(0)) &&
1446 !isa<ConstantInt>(AndI->getOperand(1)) &&
1447 AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse())
1448 return false;
1449
1450 for (auto *U : AndI->users()) {
1451 Instruction *User = cast<Instruction>(U);
1452
1453 // Only sink 'and' feeding icmp with 0.
1454 if (!isa<ICmpInst>(User))
1455 return false;
1456
1457 auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1));
1458 if (!CmpC || !CmpC->isZero())
1459 return false;
1460 }
1461
1462 if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI))
1463 return false;
1464
1465 LLVM_DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n");
1466 LLVM_DEBUG(AndI->getParent()->dump());
1467
1468 // Push the 'and' into the same block as the icmp 0. There should only be
1469 // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
1470 // others, so we don't need to keep track of which BBs we insert into.
1471 for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
1472 UI != E; ) {
1473 Use &TheUse = UI.getUse();
1474 Instruction *User = cast<Instruction>(*UI);
1475
1476 // Preincrement use iterator so we don't invalidate it.
1477 ++UI;
1478
1479 LLVM_DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n");
1480
1481 // Keep the 'and' in the same place if the use is already in the same block.
1482 Instruction *InsertPt =
1483 User->getParent() == AndI->getParent() ? AndI : User;
1484 Instruction *InsertedAnd =
1485 BinaryOperator::Create(Instruction::And, AndI->getOperand(0),
1486 AndI->getOperand(1), "", InsertPt);
1487 // Propagate the debug info.
1488 InsertedAnd->setDebugLoc(AndI->getDebugLoc());
1489
1490 // Replace a use of the 'and' with a use of the new 'and'.
1491 TheUse = InsertedAnd;
1492 ++NumAndUses;
1493 LLVM_DEBUG(User->getParent()->dump());
1494 }
1495
1496 // We removed all uses, nuke the and.
1497 AndI->eraseFromParent();
1498 return true;
1499 }
1500
1501 /// Check if the candidates could be combined with a shift instruction, which
1502 /// includes:
1503 /// 1. Truncate instruction
1504 /// 2. And instruction and the imm is a mask of the low bits:
1505 /// imm & (imm+1) == 0
isExtractBitsCandidateUse(Instruction * User)1506 static bool isExtractBitsCandidateUse(Instruction *User) {
1507 if (!isa<TruncInst>(User)) {
1508 if (User->getOpcode() != Instruction::And ||
1509 !isa<ConstantInt>(User->getOperand(1)))
1510 return false;
1511
1512 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
1513
1514 if ((Cimm & (Cimm + 1)).getBoolValue())
1515 return false;
1516 }
1517 return true;
1518 }
1519
1520 /// Sink both shift and truncate instruction to the use of truncate's BB.
1521 static bool
SinkShiftAndTruncate(BinaryOperator * ShiftI,Instruction * User,ConstantInt * CI,DenseMap<BasicBlock *,BinaryOperator * > & InsertedShifts,const TargetLowering & TLI,const DataLayout & DL)1522 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
1523 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
1524 const TargetLowering &TLI, const DataLayout &DL) {
1525 BasicBlock *UserBB = User->getParent();
1526 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
1527 TruncInst *TruncI = dyn_cast<TruncInst>(User);
1528 bool MadeChange = false;
1529
1530 for (Value::user_iterator TruncUI = TruncI->user_begin(),
1531 TruncE = TruncI->user_end();
1532 TruncUI != TruncE;) {
1533
1534 Use &TruncTheUse = TruncUI.getUse();
1535 Instruction *TruncUser = cast<Instruction>(*TruncUI);
1536 // Preincrement use iterator so we don't invalidate it.
1537
1538 ++TruncUI;
1539
1540 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
1541 if (!ISDOpcode)
1542 continue;
1543
1544 // If the use is actually a legal node, there will not be an
1545 // implicit truncate.
1546 // FIXME: always querying the result type is just an
1547 // approximation; some nodes' legality is determined by the
1548 // operand or other means. There's no good way to find out though.
1549 if (TLI.isOperationLegalOrCustom(
1550 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
1551 continue;
1552
1553 // Don't bother for PHI nodes.
1554 if (isa<PHINode>(TruncUser))
1555 continue;
1556
1557 BasicBlock *TruncUserBB = TruncUser->getParent();
1558
1559 if (UserBB == TruncUserBB)
1560 continue;
1561
1562 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
1563 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
1564
1565 if (!InsertedShift && !InsertedTrunc) {
1566 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
1567 assert(InsertPt != TruncUserBB->end());
1568 // Sink the shift
1569 if (ShiftI->getOpcode() == Instruction::AShr)
1570 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1571 "", &*InsertPt);
1572 else
1573 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1574 "", &*InsertPt);
1575 InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
1576
1577 // Sink the trunc
1578 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
1579 TruncInsertPt++;
1580 assert(TruncInsertPt != TruncUserBB->end());
1581
1582 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
1583 TruncI->getType(), "", &*TruncInsertPt);
1584 InsertedTrunc->setDebugLoc(TruncI->getDebugLoc());
1585
1586 MadeChange = true;
1587
1588 TruncTheUse = InsertedTrunc;
1589 }
1590 }
1591 return MadeChange;
1592 }
1593
1594 /// Sink the shift *right* instruction into user blocks if the uses could
1595 /// potentially be combined with this shift instruction and generate BitExtract
1596 /// instruction. It will only be applied if the architecture supports BitExtract
1597 /// instruction. Here is an example:
1598 /// BB1:
1599 /// %x.extract.shift = lshr i64 %arg1, 32
1600 /// BB2:
1601 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1602 /// ==>
1603 ///
1604 /// BB2:
1605 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1606 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1607 ///
1608 /// CodeGen will recognize the pattern in BB2 and generate BitExtract
1609 /// instruction.
1610 /// Return true if any changes are made.
OptimizeExtractBits(BinaryOperator * ShiftI,ConstantInt * CI,const TargetLowering & TLI,const DataLayout & DL)1611 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1612 const TargetLowering &TLI,
1613 const DataLayout &DL) {
1614 BasicBlock *DefBB = ShiftI->getParent();
1615
1616 /// Only insert instructions in each block once.
1617 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1618
1619 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1620
1621 bool MadeChange = false;
1622 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1623 UI != E;) {
1624 Use &TheUse = UI.getUse();
1625 Instruction *User = cast<Instruction>(*UI);
1626 // Preincrement use iterator so we don't invalidate it.
1627 ++UI;
1628
1629 // Don't bother for PHI nodes.
1630 if (isa<PHINode>(User))
1631 continue;
1632
1633 if (!isExtractBitsCandidateUse(User))
1634 continue;
1635
1636 BasicBlock *UserBB = User->getParent();
1637
1638 if (UserBB == DefBB) {
1639 // If the shift and truncate instruction are in the same BB. The use of
1640 // the truncate(TruncUse) may still introduce another truncate if not
1641 // legal. In this case, we would like to sink both shift and truncate
1642 // instruction to the BB of TruncUse.
1643 // for example:
1644 // BB1:
1645 // i64 shift.result = lshr i64 opnd, imm
1646 // trunc.result = trunc shift.result to i16
1647 //
1648 // BB2:
1649 // ----> We will have an implicit truncate here if the architecture does
1650 // not have i16 compare.
1651 // cmp i16 trunc.result, opnd2
1652 //
1653 if (isa<TruncInst>(User) && shiftIsLegal
1654 // If the type of the truncate is legal, no truncate will be
1655 // introduced in other basic blocks.
1656 &&
1657 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1658 MadeChange =
1659 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1660
1661 continue;
1662 }
1663 // If we have already inserted a shift into this block, use it.
1664 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1665
1666 if (!InsertedShift) {
1667 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1668 assert(InsertPt != UserBB->end());
1669
1670 if (ShiftI->getOpcode() == Instruction::AShr)
1671 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1672 "", &*InsertPt);
1673 else
1674 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1675 "", &*InsertPt);
1676 InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
1677
1678 MadeChange = true;
1679 }
1680
1681 // Replace a use of the shift with a use of the new shift.
1682 TheUse = InsertedShift;
1683 }
1684
1685 // If we removed all uses, or there are none, nuke the shift.
1686 if (ShiftI->use_empty()) {
1687 salvageDebugInfo(*ShiftI);
1688 ShiftI->eraseFromParent();
1689 MadeChange = true;
1690 }
1691
1692 return MadeChange;
1693 }
1694
1695 /// If counting leading or trailing zeros is an expensive operation and a zero
1696 /// input is defined, add a check for zero to avoid calling the intrinsic.
1697 ///
1698 /// We want to transform:
1699 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
1700 ///
1701 /// into:
1702 /// entry:
1703 /// %cmpz = icmp eq i64 %A, 0
1704 /// br i1 %cmpz, label %cond.end, label %cond.false
1705 /// cond.false:
1706 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
1707 /// br label %cond.end
1708 /// cond.end:
1709 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
1710 ///
1711 /// If the transform is performed, return true and set ModifiedDT to true.
despeculateCountZeros(IntrinsicInst * CountZeros,const TargetLowering * TLI,const DataLayout * DL,bool & ModifiedDT)1712 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
1713 const TargetLowering *TLI,
1714 const DataLayout *DL,
1715 bool &ModifiedDT) {
1716 if (!TLI || !DL)
1717 return false;
1718
1719 // If a zero input is undefined, it doesn't make sense to despeculate that.
1720 if (match(CountZeros->getOperand(1), m_One()))
1721 return false;
1722
1723 // If it's cheap to speculate, there's nothing to do.
1724 auto IntrinsicID = CountZeros->getIntrinsicID();
1725 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
1726 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
1727 return false;
1728
1729 // Only handle legal scalar cases. Anything else requires too much work.
1730 Type *Ty = CountZeros->getType();
1731 unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
1732 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits())
1733 return false;
1734
1735 // The intrinsic will be sunk behind a compare against zero and branch.
1736 BasicBlock *StartBlock = CountZeros->getParent();
1737 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
1738
1739 // Create another block after the count zero intrinsic. A PHI will be added
1740 // in this block to select the result of the intrinsic or the bit-width
1741 // constant if the input to the intrinsic is zero.
1742 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
1743 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
1744
1745 // Set up a builder to create a compare, conditional branch, and PHI.
1746 IRBuilder<> Builder(CountZeros->getContext());
1747 Builder.SetInsertPoint(StartBlock->getTerminator());
1748 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
1749
1750 // Replace the unconditional branch that was created by the first split with
1751 // a compare against zero and a conditional branch.
1752 Value *Zero = Constant::getNullValue(Ty);
1753 Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
1754 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
1755 StartBlock->getTerminator()->eraseFromParent();
1756
1757 // Create a PHI in the end block to select either the output of the intrinsic
1758 // or the bit width of the operand.
1759 Builder.SetInsertPoint(&EndBlock->front());
1760 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
1761 CountZeros->replaceAllUsesWith(PN);
1762 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
1763 PN->addIncoming(BitWidth, StartBlock);
1764 PN->addIncoming(CountZeros, CallBlock);
1765
1766 // We are explicitly handling the zero case, so we can set the intrinsic's
1767 // undefined zero argument to 'true'. This will also prevent reprocessing the
1768 // intrinsic; we only despeculate when a zero input is defined.
1769 CountZeros->setArgOperand(1, Builder.getTrue());
1770 ModifiedDT = true;
1771 return true;
1772 }
1773
optimizeCallInst(CallInst * CI,bool & ModifiedDT)1774 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) {
1775 BasicBlock *BB = CI->getParent();
1776
1777 // Lower inline assembly if we can.
1778 // If we found an inline asm expession, and if the target knows how to
1779 // lower it to normal LLVM code, do so now.
1780 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1781 if (TLI->ExpandInlineAsm(CI)) {
1782 // Avoid invalidating the iterator.
1783 CurInstIterator = BB->begin();
1784 // Avoid processing instructions out of order, which could cause
1785 // reuse before a value is defined.
1786 SunkAddrs.clear();
1787 return true;
1788 }
1789 // Sink address computing for memory operands into the block.
1790 if (optimizeInlineAsmInst(CI))
1791 return true;
1792 }
1793
1794 // Align the pointer arguments to this call if the target thinks it's a good
1795 // idea
1796 unsigned MinSize, PrefAlign;
1797 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1798 for (auto &Arg : CI->arg_operands()) {
1799 // We want to align both objects whose address is used directly and
1800 // objects whose address is used in casts and GEPs, though it only makes
1801 // sense for GEPs if the offset is a multiple of the desired alignment and
1802 // if size - offset meets the size threshold.
1803 if (!Arg->getType()->isPointerTy())
1804 continue;
1805 APInt Offset(DL->getIndexSizeInBits(
1806 cast<PointerType>(Arg->getType())->getAddressSpace()),
1807 0);
1808 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
1809 uint64_t Offset2 = Offset.getLimitedValue();
1810 if ((Offset2 & (PrefAlign-1)) != 0)
1811 continue;
1812 AllocaInst *AI;
1813 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1814 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1815 AI->setAlignment(PrefAlign);
1816 // Global variables can only be aligned if they are defined in this
1817 // object (i.e. they are uniquely initialized in this object), and
1818 // over-aligning global variables that have an explicit section is
1819 // forbidden.
1820 GlobalVariable *GV;
1821 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
1822 GV->getPointerAlignment(*DL) < PrefAlign &&
1823 DL->getTypeAllocSize(GV->getValueType()) >=
1824 MinSize + Offset2)
1825 GV->setAlignment(PrefAlign);
1826 }
1827 // If this is a memcpy (or similar) then we may be able to improve the
1828 // alignment
1829 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1830 unsigned DestAlign = getKnownAlignment(MI->getDest(), *DL);
1831 if (DestAlign > MI->getDestAlignment())
1832 MI->setDestAlignment(DestAlign);
1833 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
1834 unsigned SrcAlign = getKnownAlignment(MTI->getSource(), *DL);
1835 if (SrcAlign > MTI->getSourceAlignment())
1836 MTI->setSourceAlignment(SrcAlign);
1837 }
1838 }
1839 }
1840
1841 // If we have a cold call site, try to sink addressing computation into the
1842 // cold block. This interacts with our handling for loads and stores to
1843 // ensure that we can fold all uses of a potential addressing computation
1844 // into their uses. TODO: generalize this to work over profiling data
1845 if (!OptSize && CI->hasFnAttr(Attribute::Cold))
1846 for (auto &Arg : CI->arg_operands()) {
1847 if (!Arg->getType()->isPointerTy())
1848 continue;
1849 unsigned AS = Arg->getType()->getPointerAddressSpace();
1850 return optimizeMemoryInst(CI, Arg, Arg->getType(), AS);
1851 }
1852
1853 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1854 if (II) {
1855 switch (II->getIntrinsicID()) {
1856 default: break;
1857 case Intrinsic::experimental_widenable_condition: {
1858 // Give up on future widening oppurtunties so that we can fold away dead
1859 // paths and merge blocks before going into block-local instruction
1860 // selection.
1861 if (II->use_empty()) {
1862 II->eraseFromParent();
1863 return true;
1864 }
1865 Constant *RetVal = ConstantInt::getTrue(II->getContext());
1866 resetIteratorIfInvalidatedWhileCalling(BB, [&]() {
1867 replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
1868 });
1869 return true;
1870 }
1871 case Intrinsic::objectsize: {
1872 // Lower all uses of llvm.objectsize.*
1873 Value *RetVal =
1874 lowerObjectSizeCall(II, *DL, TLInfo, /*MustSucceed=*/true);
1875
1876 resetIteratorIfInvalidatedWhileCalling(BB, [&]() {
1877 replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
1878 });
1879 return true;
1880 }
1881 case Intrinsic::is_constant: {
1882 // If is_constant hasn't folded away yet, lower it to false now.
1883 Constant *RetVal = ConstantInt::get(II->getType(), 0);
1884 resetIteratorIfInvalidatedWhileCalling(BB, [&]() {
1885 replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
1886 });
1887 return true;
1888 }
1889 case Intrinsic::aarch64_stlxr:
1890 case Intrinsic::aarch64_stxr: {
1891 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1892 if (!ExtVal || !ExtVal->hasOneUse() ||
1893 ExtVal->getParent() == CI->getParent())
1894 return false;
1895 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1896 ExtVal->moveBefore(CI);
1897 // Mark this instruction as "inserted by CGP", so that other
1898 // optimizations don't touch it.
1899 InsertedInsts.insert(ExtVal);
1900 return true;
1901 }
1902
1903 case Intrinsic::launder_invariant_group:
1904 case Intrinsic::strip_invariant_group: {
1905 Value *ArgVal = II->getArgOperand(0);
1906 auto it = LargeOffsetGEPMap.find(II);
1907 if (it != LargeOffsetGEPMap.end()) {
1908 // Merge entries in LargeOffsetGEPMap to reflect the RAUW.
1909 // Make sure not to have to deal with iterator invalidation
1910 // after possibly adding ArgVal to LargeOffsetGEPMap.
1911 auto GEPs = std::move(it->second);
1912 LargeOffsetGEPMap[ArgVal].append(GEPs.begin(), GEPs.end());
1913 LargeOffsetGEPMap.erase(II);
1914 }
1915
1916 II->replaceAllUsesWith(ArgVal);
1917 II->eraseFromParent();
1918 return true;
1919 }
1920 case Intrinsic::cttz:
1921 case Intrinsic::ctlz:
1922 // If counting zeros is expensive, try to avoid it.
1923 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
1924 }
1925
1926 if (TLI) {
1927 SmallVector<Value*, 2> PtrOps;
1928 Type *AccessTy;
1929 if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
1930 while (!PtrOps.empty()) {
1931 Value *PtrVal = PtrOps.pop_back_val();
1932 unsigned AS = PtrVal->getType()->getPointerAddressSpace();
1933 if (optimizeMemoryInst(II, PtrVal, AccessTy, AS))
1934 return true;
1935 }
1936 }
1937 }
1938
1939 // From here on out we're working with named functions.
1940 if (!CI->getCalledFunction()) return false;
1941
1942 // Lower all default uses of _chk calls. This is very similar
1943 // to what InstCombineCalls does, but here we are only lowering calls
1944 // to fortified library functions (e.g. __memcpy_chk) that have the default
1945 // "don't know" as the objectsize. Anything else should be left alone.
1946 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1947 if (Value *V = Simplifier.optimizeCall(CI)) {
1948 CI->replaceAllUsesWith(V);
1949 CI->eraseFromParent();
1950 return true;
1951 }
1952
1953 return false;
1954 }
1955
1956 /// Look for opportunities to duplicate return instructions to the predecessor
1957 /// to enable tail call optimizations. The case it is currently looking for is:
1958 /// @code
1959 /// bb0:
1960 /// %tmp0 = tail call i32 @f0()
1961 /// br label %return
1962 /// bb1:
1963 /// %tmp1 = tail call i32 @f1()
1964 /// br label %return
1965 /// bb2:
1966 /// %tmp2 = tail call i32 @f2()
1967 /// br label %return
1968 /// return:
1969 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1970 /// ret i32 %retval
1971 /// @endcode
1972 ///
1973 /// =>
1974 ///
1975 /// @code
1976 /// bb0:
1977 /// %tmp0 = tail call i32 @f0()
1978 /// ret i32 %tmp0
1979 /// bb1:
1980 /// %tmp1 = tail call i32 @f1()
1981 /// ret i32 %tmp1
1982 /// bb2:
1983 /// %tmp2 = tail call i32 @f2()
1984 /// ret i32 %tmp2
1985 /// @endcode
dupRetToEnableTailCallOpts(BasicBlock * BB,bool & ModifiedDT)1986 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB, bool &ModifiedDT) {
1987 if (!TLI)
1988 return false;
1989
1990 ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator());
1991 if (!RetI)
1992 return false;
1993
1994 PHINode *PN = nullptr;
1995 BitCastInst *BCI = nullptr;
1996 Value *V = RetI->getReturnValue();
1997 if (V) {
1998 BCI = dyn_cast<BitCastInst>(V);
1999 if (BCI)
2000 V = BCI->getOperand(0);
2001
2002 PN = dyn_cast<PHINode>(V);
2003 if (!PN)
2004 return false;
2005 }
2006
2007 if (PN && PN->getParent() != BB)
2008 return false;
2009
2010 // Make sure there are no instructions between the PHI and return, or that the
2011 // return is the first instruction in the block.
2012 if (PN) {
2013 BasicBlock::iterator BI = BB->begin();
2014 // Skip over debug and the bitcast.
2015 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI) || &*BI == BCI);
2016 if (&*BI != RetI)
2017 return false;
2018 } else {
2019 BasicBlock::iterator BI = BB->begin();
2020 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
2021 if (&*BI != RetI)
2022 return false;
2023 }
2024
2025 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
2026 /// call.
2027 const Function *F = BB->getParent();
2028 SmallVector<CallInst*, 4> TailCalls;
2029 if (PN) {
2030 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
2031 // Look through bitcasts.
2032 Value *IncomingVal = PN->getIncomingValue(I)->stripPointerCasts();
2033 CallInst *CI = dyn_cast<CallInst>(IncomingVal);
2034 // Make sure the phi value is indeed produced by the tail call.
2035 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
2036 TLI->mayBeEmittedAsTailCall(CI) &&
2037 attributesPermitTailCall(F, CI, RetI, *TLI))
2038 TailCalls.push_back(CI);
2039 }
2040 } else {
2041 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
2042 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
2043 if (!VisitedBBs.insert(*PI).second)
2044 continue;
2045
2046 BasicBlock::InstListType &InstList = (*PI)->getInstList();
2047 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
2048 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
2049 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
2050 if (RI == RE)
2051 continue;
2052
2053 CallInst *CI = dyn_cast<CallInst>(&*RI);
2054 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) &&
2055 attributesPermitTailCall(F, CI, RetI, *TLI))
2056 TailCalls.push_back(CI);
2057 }
2058 }
2059
2060 bool Changed = false;
2061 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
2062 CallInst *CI = TailCalls[i];
2063 CallSite CS(CI);
2064
2065 // Make sure the call instruction is followed by an unconditional branch to
2066 // the return block.
2067 BasicBlock *CallBB = CI->getParent();
2068 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
2069 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
2070 continue;
2071
2072 // Duplicate the return into CallBB.
2073 (void)FoldReturnIntoUncondBranch(RetI, BB, CallBB);
2074 ModifiedDT = Changed = true;
2075 ++NumRetsDup;
2076 }
2077
2078 // If we eliminated all predecessors of the block, delete the block now.
2079 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
2080 BB->eraseFromParent();
2081
2082 return Changed;
2083 }
2084
2085 //===----------------------------------------------------------------------===//
2086 // Memory Optimization
2087 //===----------------------------------------------------------------------===//
2088
2089 namespace {
2090
2091 /// This is an extended version of TargetLowering::AddrMode
2092 /// which holds actual Value*'s for register values.
2093 struct ExtAddrMode : public TargetLowering::AddrMode {
2094 Value *BaseReg = nullptr;
2095 Value *ScaledReg = nullptr;
2096 Value *OriginalValue = nullptr;
2097 bool InBounds = true;
2098
2099 enum FieldName {
2100 NoField = 0x00,
2101 BaseRegField = 0x01,
2102 BaseGVField = 0x02,
2103 BaseOffsField = 0x04,
2104 ScaledRegField = 0x08,
2105 ScaleField = 0x10,
2106 MultipleFields = 0xff
2107 };
2108
2109
2110 ExtAddrMode() = default;
2111
2112 void print(raw_ostream &OS) const;
2113 void dump() const;
2114
compare__anon57ad8bfe0611::ExtAddrMode2115 FieldName compare(const ExtAddrMode &other) {
2116 // First check that the types are the same on each field, as differing types
2117 // is something we can't cope with later on.
2118 if (BaseReg && other.BaseReg &&
2119 BaseReg->getType() != other.BaseReg->getType())
2120 return MultipleFields;
2121 if (BaseGV && other.BaseGV &&
2122 BaseGV->getType() != other.BaseGV->getType())
2123 return MultipleFields;
2124 if (ScaledReg && other.ScaledReg &&
2125 ScaledReg->getType() != other.ScaledReg->getType())
2126 return MultipleFields;
2127
2128 // Conservatively reject 'inbounds' mismatches.
2129 if (InBounds != other.InBounds)
2130 return MultipleFields;
2131
2132 // Check each field to see if it differs.
2133 unsigned Result = NoField;
2134 if (BaseReg != other.BaseReg)
2135 Result |= BaseRegField;
2136 if (BaseGV != other.BaseGV)
2137 Result |= BaseGVField;
2138 if (BaseOffs != other.BaseOffs)
2139 Result |= BaseOffsField;
2140 if (ScaledReg != other.ScaledReg)
2141 Result |= ScaledRegField;
2142 // Don't count 0 as being a different scale, because that actually means
2143 // unscaled (which will already be counted by having no ScaledReg).
2144 if (Scale && other.Scale && Scale != other.Scale)
2145 Result |= ScaleField;
2146
2147 if (countPopulation(Result) > 1)
2148 return MultipleFields;
2149 else
2150 return static_cast<FieldName>(Result);
2151 }
2152
2153 // An AddrMode is trivial if it involves no calculation i.e. it is just a base
2154 // with no offset.
isTrivial__anon57ad8bfe0611::ExtAddrMode2155 bool isTrivial() {
2156 // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
2157 // trivial if at most one of these terms is nonzero, except that BaseGV and
2158 // BaseReg both being zero actually means a null pointer value, which we
2159 // consider to be 'non-zero' here.
2160 return !BaseOffs && !Scale && !(BaseGV && BaseReg);
2161 }
2162
GetFieldAsValue__anon57ad8bfe0611::ExtAddrMode2163 Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
2164 switch (Field) {
2165 default:
2166 return nullptr;
2167 case BaseRegField:
2168 return BaseReg;
2169 case BaseGVField:
2170 return BaseGV;
2171 case ScaledRegField:
2172 return ScaledReg;
2173 case BaseOffsField:
2174 return ConstantInt::get(IntPtrTy, BaseOffs);
2175 }
2176 }
2177
SetCombinedField__anon57ad8bfe0611::ExtAddrMode2178 void SetCombinedField(FieldName Field, Value *V,
2179 const SmallVectorImpl<ExtAddrMode> &AddrModes) {
2180 switch (Field) {
2181 default:
2182 llvm_unreachable("Unhandled fields are expected to be rejected earlier");
2183 break;
2184 case ExtAddrMode::BaseRegField:
2185 BaseReg = V;
2186 break;
2187 case ExtAddrMode::BaseGVField:
2188 // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
2189 // in the BaseReg field.
2190 assert(BaseReg == nullptr);
2191 BaseReg = V;
2192 BaseGV = nullptr;
2193 break;
2194 case ExtAddrMode::ScaledRegField:
2195 ScaledReg = V;
2196 // If we have a mix of scaled and unscaled addrmodes then we want scale
2197 // to be the scale and not zero.
2198 if (!Scale)
2199 for (const ExtAddrMode &AM : AddrModes)
2200 if (AM.Scale) {
2201 Scale = AM.Scale;
2202 break;
2203 }
2204 break;
2205 case ExtAddrMode::BaseOffsField:
2206 // The offset is no longer a constant, so it goes in ScaledReg with a
2207 // scale of 1.
2208 assert(ScaledReg == nullptr);
2209 ScaledReg = V;
2210 Scale = 1;
2211 BaseOffs = 0;
2212 break;
2213 }
2214 }
2215 };
2216
2217 } // end anonymous namespace
2218
2219 #ifndef NDEBUG
operator <<(raw_ostream & OS,const ExtAddrMode & AM)2220 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2221 AM.print(OS);
2222 return OS;
2223 }
2224 #endif
2225
2226 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
print(raw_ostream & OS) const2227 void ExtAddrMode::print(raw_ostream &OS) const {
2228 bool NeedPlus = false;
2229 OS << "[";
2230 if (InBounds)
2231 OS << "inbounds ";
2232 if (BaseGV) {
2233 OS << (NeedPlus ? " + " : "")
2234 << "GV:";
2235 BaseGV->printAsOperand(OS, /*PrintType=*/false);
2236 NeedPlus = true;
2237 }
2238
2239 if (BaseOffs) {
2240 OS << (NeedPlus ? " + " : "")
2241 << BaseOffs;
2242 NeedPlus = true;
2243 }
2244
2245 if (BaseReg) {
2246 OS << (NeedPlus ? " + " : "")
2247 << "Base:";
2248 BaseReg->printAsOperand(OS, /*PrintType=*/false);
2249 NeedPlus = true;
2250 }
2251 if (Scale) {
2252 OS << (NeedPlus ? " + " : "")
2253 << Scale << "*";
2254 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2255 }
2256
2257 OS << ']';
2258 }
2259
dump() const2260 LLVM_DUMP_METHOD void ExtAddrMode::dump() const {
2261 print(dbgs());
2262 dbgs() << '\n';
2263 }
2264 #endif
2265
2266 namespace {
2267
2268 /// This class provides transaction based operation on the IR.
2269 /// Every change made through this class is recorded in the internal state and
2270 /// can be undone (rollback) until commit is called.
2271 class TypePromotionTransaction {
2272 /// This represents the common interface of the individual transaction.
2273 /// Each class implements the logic for doing one specific modification on
2274 /// the IR via the TypePromotionTransaction.
2275 class TypePromotionAction {
2276 protected:
2277 /// The Instruction modified.
2278 Instruction *Inst;
2279
2280 public:
2281 /// Constructor of the action.
2282 /// The constructor performs the related action on the IR.
TypePromotionAction(Instruction * Inst)2283 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2284
2285 virtual ~TypePromotionAction() = default;
2286
2287 /// Undo the modification done by this action.
2288 /// When this method is called, the IR must be in the same state as it was
2289 /// before this action was applied.
2290 /// \pre Undoing the action works if and only if the IR is in the exact same
2291 /// state as it was directly after this action was applied.
2292 virtual void undo() = 0;
2293
2294 /// Advocate every change made by this action.
2295 /// When the results on the IR of the action are to be kept, it is important
2296 /// to call this function, otherwise hidden information may be kept forever.
commit()2297 virtual void commit() {
2298 // Nothing to be done, this action is not doing anything.
2299 }
2300 };
2301
2302 /// Utility to remember the position of an instruction.
2303 class InsertionHandler {
2304 /// Position of an instruction.
2305 /// Either an instruction:
2306 /// - Is the first in a basic block: BB is used.
2307 /// - Has a previous instruction: PrevInst is used.
2308 union {
2309 Instruction *PrevInst;
2310 BasicBlock *BB;
2311 } Point;
2312
2313 /// Remember whether or not the instruction had a previous instruction.
2314 bool HasPrevInstruction;
2315
2316 public:
2317 /// Record the position of \p Inst.
InsertionHandler(Instruction * Inst)2318 InsertionHandler(Instruction *Inst) {
2319 BasicBlock::iterator It = Inst->getIterator();
2320 HasPrevInstruction = (It != (Inst->getParent()->begin()));
2321 if (HasPrevInstruction)
2322 Point.PrevInst = &*--It;
2323 else
2324 Point.BB = Inst->getParent();
2325 }
2326
2327 /// Insert \p Inst at the recorded position.
insert(Instruction * Inst)2328 void insert(Instruction *Inst) {
2329 if (HasPrevInstruction) {
2330 if (Inst->getParent())
2331 Inst->removeFromParent();
2332 Inst->insertAfter(Point.PrevInst);
2333 } else {
2334 Instruction *Position = &*Point.BB->getFirstInsertionPt();
2335 if (Inst->getParent())
2336 Inst->moveBefore(Position);
2337 else
2338 Inst->insertBefore(Position);
2339 }
2340 }
2341 };
2342
2343 /// Move an instruction before another.
2344 class InstructionMoveBefore : public TypePromotionAction {
2345 /// Original position of the instruction.
2346 InsertionHandler Position;
2347
2348 public:
2349 /// Move \p Inst before \p Before.
InstructionMoveBefore(Instruction * Inst,Instruction * Before)2350 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2351 : TypePromotionAction(Inst), Position(Inst) {
2352 LLVM_DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before
2353 << "\n");
2354 Inst->moveBefore(Before);
2355 }
2356
2357 /// Move the instruction back to its original position.
undo()2358 void undo() override {
2359 LLVM_DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2360 Position.insert(Inst);
2361 }
2362 };
2363
2364 /// Set the operand of an instruction with a new value.
2365 class OperandSetter : public TypePromotionAction {
2366 /// Original operand of the instruction.
2367 Value *Origin;
2368
2369 /// Index of the modified instruction.
2370 unsigned Idx;
2371
2372 public:
2373 /// Set \p Idx operand of \p Inst with \p NewVal.
OperandSetter(Instruction * Inst,unsigned Idx,Value * NewVal)2374 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2375 : TypePromotionAction(Inst), Idx(Idx) {
2376 LLVM_DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2377 << "for:" << *Inst << "\n"
2378 << "with:" << *NewVal << "\n");
2379 Origin = Inst->getOperand(Idx);
2380 Inst->setOperand(Idx, NewVal);
2381 }
2382
2383 /// Restore the original value of the instruction.
undo()2384 void undo() override {
2385 LLVM_DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2386 << "for: " << *Inst << "\n"
2387 << "with: " << *Origin << "\n");
2388 Inst->setOperand(Idx, Origin);
2389 }
2390 };
2391
2392 /// Hide the operands of an instruction.
2393 /// Do as if this instruction was not using any of its operands.
2394 class OperandsHider : public TypePromotionAction {
2395 /// The list of original operands.
2396 SmallVector<Value *, 4> OriginalValues;
2397
2398 public:
2399 /// Remove \p Inst from the uses of the operands of \p Inst.
OperandsHider(Instruction * Inst)2400 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2401 LLVM_DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2402 unsigned NumOpnds = Inst->getNumOperands();
2403 OriginalValues.reserve(NumOpnds);
2404 for (unsigned It = 0; It < NumOpnds; ++It) {
2405 // Save the current operand.
2406 Value *Val = Inst->getOperand(It);
2407 OriginalValues.push_back(Val);
2408 // Set a dummy one.
2409 // We could use OperandSetter here, but that would imply an overhead
2410 // that we are not willing to pay.
2411 Inst->setOperand(It, UndefValue::get(Val->getType()));
2412 }
2413 }
2414
2415 /// Restore the original list of uses.
undo()2416 void undo() override {
2417 LLVM_DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2418 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2419 Inst->setOperand(It, OriginalValues[It]);
2420 }
2421 };
2422
2423 /// Build a truncate instruction.
2424 class TruncBuilder : public TypePromotionAction {
2425 Value *Val;
2426
2427 public:
2428 /// Build a truncate instruction of \p Opnd producing a \p Ty
2429 /// result.
2430 /// trunc Opnd to Ty.
TruncBuilder(Instruction * Opnd,Type * Ty)2431 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2432 IRBuilder<> Builder(Opnd);
2433 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2434 LLVM_DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2435 }
2436
2437 /// Get the built value.
getBuiltValue()2438 Value *getBuiltValue() { return Val; }
2439
2440 /// Remove the built instruction.
undo()2441 void undo() override {
2442 LLVM_DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2443 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2444 IVal->eraseFromParent();
2445 }
2446 };
2447
2448 /// Build a sign extension instruction.
2449 class SExtBuilder : public TypePromotionAction {
2450 Value *Val;
2451
2452 public:
2453 /// Build a sign extension instruction of \p Opnd producing a \p Ty
2454 /// result.
2455 /// sext Opnd to Ty.
SExtBuilder(Instruction * InsertPt,Value * Opnd,Type * Ty)2456 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2457 : TypePromotionAction(InsertPt) {
2458 IRBuilder<> Builder(InsertPt);
2459 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2460 LLVM_DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2461 }
2462
2463 /// Get the built value.
getBuiltValue()2464 Value *getBuiltValue() { return Val; }
2465
2466 /// Remove the built instruction.
undo()2467 void undo() override {
2468 LLVM_DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2469 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2470 IVal->eraseFromParent();
2471 }
2472 };
2473
2474 /// Build a zero extension instruction.
2475 class ZExtBuilder : public TypePromotionAction {
2476 Value *Val;
2477
2478 public:
2479 /// Build a zero extension instruction of \p Opnd producing a \p Ty
2480 /// result.
2481 /// zext Opnd to Ty.
ZExtBuilder(Instruction * InsertPt,Value * Opnd,Type * Ty)2482 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2483 : TypePromotionAction(InsertPt) {
2484 IRBuilder<> Builder(InsertPt);
2485 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2486 LLVM_DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2487 }
2488
2489 /// Get the built value.
getBuiltValue()2490 Value *getBuiltValue() { return Val; }
2491
2492 /// Remove the built instruction.
undo()2493 void undo() override {
2494 LLVM_DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2495 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2496 IVal->eraseFromParent();
2497 }
2498 };
2499
2500 /// Mutate an instruction to another type.
2501 class TypeMutator : public TypePromotionAction {
2502 /// Record the original type.
2503 Type *OrigTy;
2504
2505 public:
2506 /// Mutate the type of \p Inst into \p NewTy.
TypeMutator(Instruction * Inst,Type * NewTy)2507 TypeMutator(Instruction *Inst, Type *NewTy)
2508 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2509 LLVM_DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2510 << "\n");
2511 Inst->mutateType(NewTy);
2512 }
2513
2514 /// Mutate the instruction back to its original type.
undo()2515 void undo() override {
2516 LLVM_DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2517 << "\n");
2518 Inst->mutateType(OrigTy);
2519 }
2520 };
2521
2522 /// Replace the uses of an instruction by another instruction.
2523 class UsesReplacer : public TypePromotionAction {
2524 /// Helper structure to keep track of the replaced uses.
2525 struct InstructionAndIdx {
2526 /// The instruction using the instruction.
2527 Instruction *Inst;
2528
2529 /// The index where this instruction is used for Inst.
2530 unsigned Idx;
2531
InstructionAndIdx__anon57ad8bfe0711::TypePromotionTransaction::UsesReplacer::InstructionAndIdx2532 InstructionAndIdx(Instruction *Inst, unsigned Idx)
2533 : Inst(Inst), Idx(Idx) {}
2534 };
2535
2536 /// Keep track of the original uses (pair Instruction, Index).
2537 SmallVector<InstructionAndIdx, 4> OriginalUses;
2538 /// Keep track of the debug users.
2539 SmallVector<DbgValueInst *, 1> DbgValues;
2540
2541 using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
2542
2543 public:
2544 /// Replace all the use of \p Inst by \p New.
UsesReplacer(Instruction * Inst,Value * New)2545 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2546 LLVM_DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2547 << "\n");
2548 // Record the original uses.
2549 for (Use &U : Inst->uses()) {
2550 Instruction *UserI = cast<Instruction>(U.getUser());
2551 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2552 }
2553 // Record the debug uses separately. They are not in the instruction's
2554 // use list, but they are replaced by RAUW.
2555 findDbgValues(DbgValues, Inst);
2556
2557 // Now, we can replace the uses.
2558 Inst->replaceAllUsesWith(New);
2559 }
2560
2561 /// Reassign the original uses of Inst to Inst.
undo()2562 void undo() override {
2563 LLVM_DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2564 for (use_iterator UseIt = OriginalUses.begin(),
2565 EndIt = OriginalUses.end();
2566 UseIt != EndIt; ++UseIt) {
2567 UseIt->Inst->setOperand(UseIt->Idx, Inst);
2568 }
2569 // RAUW has replaced all original uses with references to the new value,
2570 // including the debug uses. Since we are undoing the replacements,
2571 // the original debug uses must also be reinstated to maintain the
2572 // correctness and utility of debug value instructions.
2573 for (auto *DVI: DbgValues) {
2574 LLVMContext &Ctx = Inst->getType()->getContext();
2575 auto *MV = MetadataAsValue::get(Ctx, ValueAsMetadata::get(Inst));
2576 DVI->setOperand(0, MV);
2577 }
2578 }
2579 };
2580
2581 /// Remove an instruction from the IR.
2582 class InstructionRemover : public TypePromotionAction {
2583 /// Original position of the instruction.
2584 InsertionHandler Inserter;
2585
2586 /// Helper structure to hide all the link to the instruction. In other
2587 /// words, this helps to do as if the instruction was removed.
2588 OperandsHider Hider;
2589
2590 /// Keep track of the uses replaced, if any.
2591 UsesReplacer *Replacer = nullptr;
2592
2593 /// Keep track of instructions removed.
2594 SetOfInstrs &RemovedInsts;
2595
2596 public:
2597 /// Remove all reference of \p Inst and optionally replace all its
2598 /// uses with New.
2599 /// \p RemovedInsts Keep track of the instructions removed by this Action.
2600 /// \pre If !Inst->use_empty(), then New != nullptr
InstructionRemover(Instruction * Inst,SetOfInstrs & RemovedInsts,Value * New=nullptr)2601 InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
2602 Value *New = nullptr)
2603 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2604 RemovedInsts(RemovedInsts) {
2605 if (New)
2606 Replacer = new UsesReplacer(Inst, New);
2607 LLVM_DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2608 RemovedInsts.insert(Inst);
2609 /// The instructions removed here will be freed after completing
2610 /// optimizeBlock() for all blocks as we need to keep track of the
2611 /// removed instructions during promotion.
2612 Inst->removeFromParent();
2613 }
2614
~InstructionRemover()2615 ~InstructionRemover() override { delete Replacer; }
2616
2617 /// Resurrect the instruction and reassign it to the proper uses if
2618 /// new value was provided when build this action.
undo()2619 void undo() override {
2620 LLVM_DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2621 Inserter.insert(Inst);
2622 if (Replacer)
2623 Replacer->undo();
2624 Hider.undo();
2625 RemovedInsts.erase(Inst);
2626 }
2627 };
2628
2629 public:
2630 /// Restoration point.
2631 /// The restoration point is a pointer to an action instead of an iterator
2632 /// because the iterator may be invalidated but not the pointer.
2633 using ConstRestorationPt = const TypePromotionAction *;
2634
TypePromotionTransaction(SetOfInstrs & RemovedInsts)2635 TypePromotionTransaction(SetOfInstrs &RemovedInsts)
2636 : RemovedInsts(RemovedInsts) {}
2637
2638 /// Advocate every changes made in that transaction.
2639 void commit();
2640
2641 /// Undo all the changes made after the given point.
2642 void rollback(ConstRestorationPt Point);
2643
2644 /// Get the current restoration point.
2645 ConstRestorationPt getRestorationPoint() const;
2646
2647 /// \name API for IR modification with state keeping to support rollback.
2648 /// @{
2649 /// Same as Instruction::setOperand.
2650 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2651
2652 /// Same as Instruction::eraseFromParent.
2653 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2654
2655 /// Same as Value::replaceAllUsesWith.
2656 void replaceAllUsesWith(Instruction *Inst, Value *New);
2657
2658 /// Same as Value::mutateType.
2659 void mutateType(Instruction *Inst, Type *NewTy);
2660
2661 /// Same as IRBuilder::createTrunc.
2662 Value *createTrunc(Instruction *Opnd, Type *Ty);
2663
2664 /// Same as IRBuilder::createSExt.
2665 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2666
2667 /// Same as IRBuilder::createZExt.
2668 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2669
2670 /// Same as Instruction::moveBefore.
2671 void moveBefore(Instruction *Inst, Instruction *Before);
2672 /// @}
2673
2674 private:
2675 /// The ordered list of actions made so far.
2676 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
2677
2678 using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
2679
2680 SetOfInstrs &RemovedInsts;
2681 };
2682
2683 } // end anonymous namespace
2684
setOperand(Instruction * Inst,unsigned Idx,Value * NewVal)2685 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2686 Value *NewVal) {
2687 Actions.push_back(llvm::make_unique<TypePromotionTransaction::OperandSetter>(
2688 Inst, Idx, NewVal));
2689 }
2690
eraseInstruction(Instruction * Inst,Value * NewVal)2691 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2692 Value *NewVal) {
2693 Actions.push_back(
2694 llvm::make_unique<TypePromotionTransaction::InstructionRemover>(
2695 Inst, RemovedInsts, NewVal));
2696 }
2697
replaceAllUsesWith(Instruction * Inst,Value * New)2698 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2699 Value *New) {
2700 Actions.push_back(
2701 llvm::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2702 }
2703
mutateType(Instruction * Inst,Type * NewTy)2704 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2705 Actions.push_back(
2706 llvm::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2707 }
2708
createTrunc(Instruction * Opnd,Type * Ty)2709 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2710 Type *Ty) {
2711 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2712 Value *Val = Ptr->getBuiltValue();
2713 Actions.push_back(std::move(Ptr));
2714 return Val;
2715 }
2716
createSExt(Instruction * Inst,Value * Opnd,Type * Ty)2717 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2718 Value *Opnd, Type *Ty) {
2719 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2720 Value *Val = Ptr->getBuiltValue();
2721 Actions.push_back(std::move(Ptr));
2722 return Val;
2723 }
2724
createZExt(Instruction * Inst,Value * Opnd,Type * Ty)2725 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2726 Value *Opnd, Type *Ty) {
2727 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2728 Value *Val = Ptr->getBuiltValue();
2729 Actions.push_back(std::move(Ptr));
2730 return Val;
2731 }
2732
moveBefore(Instruction * Inst,Instruction * Before)2733 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2734 Instruction *Before) {
2735 Actions.push_back(
2736 llvm::make_unique<TypePromotionTransaction::InstructionMoveBefore>(
2737 Inst, Before));
2738 }
2739
2740 TypePromotionTransaction::ConstRestorationPt
getRestorationPoint() const2741 TypePromotionTransaction::getRestorationPoint() const {
2742 return !Actions.empty() ? Actions.back().get() : nullptr;
2743 }
2744
commit()2745 void TypePromotionTransaction::commit() {
2746 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2747 ++It)
2748 (*It)->commit();
2749 Actions.clear();
2750 }
2751
rollback(TypePromotionTransaction::ConstRestorationPt Point)2752 void TypePromotionTransaction::rollback(
2753 TypePromotionTransaction::ConstRestorationPt Point) {
2754 while (!Actions.empty() && Point != Actions.back().get()) {
2755 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2756 Curr->undo();
2757 }
2758 }
2759
2760 namespace {
2761
2762 /// A helper class for matching addressing modes.
2763 ///
2764 /// This encapsulates the logic for matching the target-legal addressing modes.
2765 class AddressingModeMatcher {
2766 SmallVectorImpl<Instruction*> &AddrModeInsts;
2767 const TargetLowering &TLI;
2768 const TargetRegisterInfo &TRI;
2769 const DataLayout &DL;
2770
2771 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2772 /// the memory instruction that we're computing this address for.
2773 Type *AccessTy;
2774 unsigned AddrSpace;
2775 Instruction *MemoryInst;
2776
2777 /// This is the addressing mode that we're building up. This is
2778 /// part of the return value of this addressing mode matching stuff.
2779 ExtAddrMode &AddrMode;
2780
2781 /// The instructions inserted by other CodeGenPrepare optimizations.
2782 const SetOfInstrs &InsertedInsts;
2783
2784 /// A map from the instructions to their type before promotion.
2785 InstrToOrigTy &PromotedInsts;
2786
2787 /// The ongoing transaction where every action should be registered.
2788 TypePromotionTransaction &TPT;
2789
2790 // A GEP which has too large offset to be folded into the addressing mode.
2791 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP;
2792
2793 /// This is set to true when we should not do profitability checks.
2794 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2795 bool IgnoreProfitability;
2796
AddressingModeMatcher(SmallVectorImpl<Instruction * > & AMI,const TargetLowering & TLI,const TargetRegisterInfo & TRI,Type * AT,unsigned AS,Instruction * MI,ExtAddrMode & AM,const SetOfInstrs & InsertedInsts,InstrToOrigTy & PromotedInsts,TypePromotionTransaction & TPT,std::pair<AssertingVH<GetElementPtrInst>,int64_t> & LargeOffsetGEP)2797 AddressingModeMatcher(
2798 SmallVectorImpl<Instruction *> &AMI, const TargetLowering &TLI,
2799 const TargetRegisterInfo &TRI, Type *AT, unsigned AS, Instruction *MI,
2800 ExtAddrMode &AM, const SetOfInstrs &InsertedInsts,
2801 InstrToOrigTy &PromotedInsts, TypePromotionTransaction &TPT,
2802 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP)
2803 : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
2804 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2805 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2806 PromotedInsts(PromotedInsts), TPT(TPT), LargeOffsetGEP(LargeOffsetGEP) {
2807 IgnoreProfitability = false;
2808 }
2809
2810 public:
2811 /// Find the maximal addressing mode that a load/store of V can fold,
2812 /// give an access type of AccessTy. This returns a list of involved
2813 /// instructions in AddrModeInsts.
2814 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2815 /// optimizations.
2816 /// \p PromotedInsts maps the instructions to their type before promotion.
2817 /// \p The ongoing transaction where every action should be registered.
2818 static ExtAddrMode
Match(Value * V,Type * AccessTy,unsigned AS,Instruction * MemoryInst,SmallVectorImpl<Instruction * > & AddrModeInsts,const TargetLowering & TLI,const TargetRegisterInfo & TRI,const SetOfInstrs & InsertedInsts,InstrToOrigTy & PromotedInsts,TypePromotionTransaction & TPT,std::pair<AssertingVH<GetElementPtrInst>,int64_t> & LargeOffsetGEP)2819 Match(Value *V, Type *AccessTy, unsigned AS, Instruction *MemoryInst,
2820 SmallVectorImpl<Instruction *> &AddrModeInsts,
2821 const TargetLowering &TLI, const TargetRegisterInfo &TRI,
2822 const SetOfInstrs &InsertedInsts, InstrToOrigTy &PromotedInsts,
2823 TypePromotionTransaction &TPT,
2824 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP) {
2825 ExtAddrMode Result;
2826
2827 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI, AccessTy, AS,
2828 MemoryInst, Result, InsertedInsts,
2829 PromotedInsts, TPT, LargeOffsetGEP)
2830 .matchAddr(V, 0);
2831 (void)Success; assert(Success && "Couldn't select *anything*?");
2832 return Result;
2833 }
2834
2835 private:
2836 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2837 bool matchAddr(Value *Addr, unsigned Depth);
2838 bool matchOperationAddr(User *AddrInst, unsigned Opcode, unsigned Depth,
2839 bool *MovedAway = nullptr);
2840 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2841 ExtAddrMode &AMBefore,
2842 ExtAddrMode &AMAfter);
2843 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2844 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2845 Value *PromotedOperand) const;
2846 };
2847
2848 class PhiNodeSet;
2849
2850 /// An iterator for PhiNodeSet.
2851 class PhiNodeSetIterator {
2852 PhiNodeSet * const Set;
2853 size_t CurrentIndex = 0;
2854
2855 public:
2856 /// The constructor. Start should point to either a valid element, or be equal
2857 /// to the size of the underlying SmallVector of the PhiNodeSet.
2858 PhiNodeSetIterator(PhiNodeSet * const Set, size_t Start);
2859 PHINode * operator*() const;
2860 PhiNodeSetIterator& operator++();
2861 bool operator==(const PhiNodeSetIterator &RHS) const;
2862 bool operator!=(const PhiNodeSetIterator &RHS) const;
2863 };
2864
2865 /// Keeps a set of PHINodes.
2866 ///
2867 /// This is a minimal set implementation for a specific use case:
2868 /// It is very fast when there are very few elements, but also provides good
2869 /// performance when there are many. It is similar to SmallPtrSet, but also
2870 /// provides iteration by insertion order, which is deterministic and stable
2871 /// across runs. It is also similar to SmallSetVector, but provides removing
2872 /// elements in O(1) time. This is achieved by not actually removing the element
2873 /// from the underlying vector, so comes at the cost of using more memory, but
2874 /// that is fine, since PhiNodeSets are used as short lived objects.
2875 class PhiNodeSet {
2876 friend class PhiNodeSetIterator;
2877
2878 using MapType = SmallDenseMap<PHINode *, size_t, 32>;
2879 using iterator = PhiNodeSetIterator;
2880
2881 /// Keeps the elements in the order of their insertion in the underlying
2882 /// vector. To achieve constant time removal, it never deletes any element.
2883 SmallVector<PHINode *, 32> NodeList;
2884
2885 /// Keeps the elements in the underlying set implementation. This (and not the
2886 /// NodeList defined above) is the source of truth on whether an element
2887 /// is actually in the collection.
2888 MapType NodeMap;
2889
2890 /// Points to the first valid (not deleted) element when the set is not empty
2891 /// and the value is not zero. Equals to the size of the underlying vector
2892 /// when the set is empty. When the value is 0, as in the beginning, the
2893 /// first element may or may not be valid.
2894 size_t FirstValidElement = 0;
2895
2896 public:
2897 /// Inserts a new element to the collection.
2898 /// \returns true if the element is actually added, i.e. was not in the
2899 /// collection before the operation.
insert(PHINode * Ptr)2900 bool insert(PHINode *Ptr) {
2901 if (NodeMap.insert(std::make_pair(Ptr, NodeList.size())).second) {
2902 NodeList.push_back(Ptr);
2903 return true;
2904 }
2905 return false;
2906 }
2907
2908 /// Removes the element from the collection.
2909 /// \returns whether the element is actually removed, i.e. was in the
2910 /// collection before the operation.
erase(PHINode * Ptr)2911 bool erase(PHINode *Ptr) {
2912 auto it = NodeMap.find(Ptr);
2913 if (it != NodeMap.end()) {
2914 NodeMap.erase(Ptr);
2915 SkipRemovedElements(FirstValidElement);
2916 return true;
2917 }
2918 return false;
2919 }
2920
2921 /// Removes all elements and clears the collection.
clear()2922 void clear() {
2923 NodeMap.clear();
2924 NodeList.clear();
2925 FirstValidElement = 0;
2926 }
2927
2928 /// \returns an iterator that will iterate the elements in the order of
2929 /// insertion.
begin()2930 iterator begin() {
2931 if (FirstValidElement == 0)
2932 SkipRemovedElements(FirstValidElement);
2933 return PhiNodeSetIterator(this, FirstValidElement);
2934 }
2935
2936 /// \returns an iterator that points to the end of the collection.
end()2937 iterator end() { return PhiNodeSetIterator(this, NodeList.size()); }
2938
2939 /// Returns the number of elements in the collection.
size() const2940 size_t size() const {
2941 return NodeMap.size();
2942 }
2943
2944 /// \returns 1 if the given element is in the collection, and 0 if otherwise.
count(PHINode * Ptr) const2945 size_t count(PHINode *Ptr) const {
2946 return NodeMap.count(Ptr);
2947 }
2948
2949 private:
2950 /// Updates the CurrentIndex so that it will point to a valid element.
2951 ///
2952 /// If the element of NodeList at CurrentIndex is valid, it does not
2953 /// change it. If there are no more valid elements, it updates CurrentIndex
2954 /// to point to the end of the NodeList.
SkipRemovedElements(size_t & CurrentIndex)2955 void SkipRemovedElements(size_t &CurrentIndex) {
2956 while (CurrentIndex < NodeList.size()) {
2957 auto it = NodeMap.find(NodeList[CurrentIndex]);
2958 // If the element has been deleted and added again later, NodeMap will
2959 // point to a different index, so CurrentIndex will still be invalid.
2960 if (it != NodeMap.end() && it->second == CurrentIndex)
2961 break;
2962 ++CurrentIndex;
2963 }
2964 }
2965 };
2966
PhiNodeSetIterator(PhiNodeSet * const Set,size_t Start)2967 PhiNodeSetIterator::PhiNodeSetIterator(PhiNodeSet *const Set, size_t Start)
2968 : Set(Set), CurrentIndex(Start) {}
2969
operator *() const2970 PHINode * PhiNodeSetIterator::operator*() const {
2971 assert(CurrentIndex < Set->NodeList.size() &&
2972 "PhiNodeSet access out of range");
2973 return Set->NodeList[CurrentIndex];
2974 }
2975
operator ++()2976 PhiNodeSetIterator& PhiNodeSetIterator::operator++() {
2977 assert(CurrentIndex < Set->NodeList.size() &&
2978 "PhiNodeSet access out of range");
2979 ++CurrentIndex;
2980 Set->SkipRemovedElements(CurrentIndex);
2981 return *this;
2982 }
2983
operator ==(const PhiNodeSetIterator & RHS) const2984 bool PhiNodeSetIterator::operator==(const PhiNodeSetIterator &RHS) const {
2985 return CurrentIndex == RHS.CurrentIndex;
2986 }
2987
operator !=(const PhiNodeSetIterator & RHS) const2988 bool PhiNodeSetIterator::operator!=(const PhiNodeSetIterator &RHS) const {
2989 return !((*this) == RHS);
2990 }
2991
2992 /// Keep track of simplification of Phi nodes.
2993 /// Accept the set of all phi nodes and erase phi node from this set
2994 /// if it is simplified.
2995 class SimplificationTracker {
2996 DenseMap<Value *, Value *> Storage;
2997 const SimplifyQuery &SQ;
2998 // Tracks newly created Phi nodes. The elements are iterated by insertion
2999 // order.
3000 PhiNodeSet AllPhiNodes;
3001 // Tracks newly created Select nodes.
3002 SmallPtrSet<SelectInst *, 32> AllSelectNodes;
3003
3004 public:
SimplificationTracker(const SimplifyQuery & sq)3005 SimplificationTracker(const SimplifyQuery &sq)
3006 : SQ(sq) {}
3007
Get(Value * V)3008 Value *Get(Value *V) {
3009 do {
3010 auto SV = Storage.find(V);
3011 if (SV == Storage.end())
3012 return V;
3013 V = SV->second;
3014 } while (true);
3015 }
3016
Simplify(Value * Val)3017 Value *Simplify(Value *Val) {
3018 SmallVector<Value *, 32> WorkList;
3019 SmallPtrSet<Value *, 32> Visited;
3020 WorkList.push_back(Val);
3021 while (!WorkList.empty()) {
3022 auto P = WorkList.pop_back_val();
3023 if (!Visited.insert(P).second)
3024 continue;
3025 if (auto *PI = dyn_cast<Instruction>(P))
3026 if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) {
3027 for (auto *U : PI->users())
3028 WorkList.push_back(cast<Value>(U));
3029 Put(PI, V);
3030 PI->replaceAllUsesWith(V);
3031 if (auto *PHI = dyn_cast<PHINode>(PI))
3032 AllPhiNodes.erase(PHI);
3033 if (auto *Select = dyn_cast<SelectInst>(PI))
3034 AllSelectNodes.erase(Select);
3035 PI->eraseFromParent();
3036 }
3037 }
3038 return Get(Val);
3039 }
3040
Put(Value * From,Value * To)3041 void Put(Value *From, Value *To) {
3042 Storage.insert({ From, To });
3043 }
3044
ReplacePhi(PHINode * From,PHINode * To)3045 void ReplacePhi(PHINode *From, PHINode *To) {
3046 Value* OldReplacement = Get(From);
3047 while (OldReplacement != From) {
3048 From = To;
3049 To = dyn_cast<PHINode>(OldReplacement);
3050 OldReplacement = Get(From);
3051 }
3052 assert(Get(To) == To && "Replacement PHI node is already replaced.");
3053 Put(From, To);
3054 From->replaceAllUsesWith(To);
3055 AllPhiNodes.erase(From);
3056 From->eraseFromParent();
3057 }
3058
newPhiNodes()3059 PhiNodeSet& newPhiNodes() { return AllPhiNodes; }
3060
insertNewPhi(PHINode * PN)3061 void insertNewPhi(PHINode *PN) { AllPhiNodes.insert(PN); }
3062
insertNewSelect(SelectInst * SI)3063 void insertNewSelect(SelectInst *SI) { AllSelectNodes.insert(SI); }
3064
countNewPhiNodes() const3065 unsigned countNewPhiNodes() const { return AllPhiNodes.size(); }
3066
countNewSelectNodes() const3067 unsigned countNewSelectNodes() const { return AllSelectNodes.size(); }
3068
destroyNewNodes(Type * CommonType)3069 void destroyNewNodes(Type *CommonType) {
3070 // For safe erasing, replace the uses with dummy value first.
3071 auto Dummy = UndefValue::get(CommonType);
3072 for (auto I : AllPhiNodes) {
3073 I->replaceAllUsesWith(Dummy);
3074 I->eraseFromParent();
3075 }
3076 AllPhiNodes.clear();
3077 for (auto I : AllSelectNodes) {
3078 I->replaceAllUsesWith(Dummy);
3079 I->eraseFromParent();
3080 }
3081 AllSelectNodes.clear();
3082 }
3083 };
3084
3085 /// A helper class for combining addressing modes.
3086 class AddressingModeCombiner {
3087 typedef DenseMap<Value *, Value *> FoldAddrToValueMapping;
3088 typedef std::pair<PHINode *, PHINode *> PHIPair;
3089
3090 private:
3091 /// The addressing modes we've collected.
3092 SmallVector<ExtAddrMode, 16> AddrModes;
3093
3094 /// The field in which the AddrModes differ, when we have more than one.
3095 ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
3096
3097 /// Are the AddrModes that we have all just equal to their original values?
3098 bool AllAddrModesTrivial = true;
3099
3100 /// Common Type for all different fields in addressing modes.
3101 Type *CommonType;
3102
3103 /// SimplifyQuery for simplifyInstruction utility.
3104 const SimplifyQuery &SQ;
3105
3106 /// Original Address.
3107 Value *Original;
3108
3109 public:
AddressingModeCombiner(const SimplifyQuery & _SQ,Value * OriginalValue)3110 AddressingModeCombiner(const SimplifyQuery &_SQ, Value *OriginalValue)
3111 : CommonType(nullptr), SQ(_SQ), Original(OriginalValue) {}
3112
3113 /// Get the combined AddrMode
getAddrMode() const3114 const ExtAddrMode &getAddrMode() const {
3115 return AddrModes[0];
3116 }
3117
3118 /// Add a new AddrMode if it's compatible with the AddrModes we already
3119 /// have.
3120 /// \return True iff we succeeded in doing so.
addNewAddrMode(ExtAddrMode & NewAddrMode)3121 bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
3122 // Take note of if we have any non-trivial AddrModes, as we need to detect
3123 // when all AddrModes are trivial as then we would introduce a phi or select
3124 // which just duplicates what's already there.
3125 AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
3126
3127 // If this is the first addrmode then everything is fine.
3128 if (AddrModes.empty()) {
3129 AddrModes.emplace_back(NewAddrMode);
3130 return true;
3131 }
3132
3133 // Figure out how different this is from the other address modes, which we
3134 // can do just by comparing against the first one given that we only care
3135 // about the cumulative difference.
3136 ExtAddrMode::FieldName ThisDifferentField =
3137 AddrModes[0].compare(NewAddrMode);
3138 if (DifferentField == ExtAddrMode::NoField)
3139 DifferentField = ThisDifferentField;
3140 else if (DifferentField != ThisDifferentField)
3141 DifferentField = ExtAddrMode::MultipleFields;
3142
3143 // If NewAddrMode differs in more than one dimension we cannot handle it.
3144 bool CanHandle = DifferentField != ExtAddrMode::MultipleFields;
3145
3146 // If Scale Field is different then we reject.
3147 CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField;
3148
3149 // We also must reject the case when base offset is different and
3150 // scale reg is not null, we cannot handle this case due to merge of
3151 // different offsets will be used as ScaleReg.
3152 CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField ||
3153 !NewAddrMode.ScaledReg);
3154
3155 // We also must reject the case when GV is different and BaseReg installed
3156 // due to we want to use base reg as a merge of GV values.
3157 CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField ||
3158 !NewAddrMode.HasBaseReg);
3159
3160 // Even if NewAddMode is the same we still need to collect it due to
3161 // original value is different. And later we will need all original values
3162 // as anchors during finding the common Phi node.
3163 if (CanHandle)
3164 AddrModes.emplace_back(NewAddrMode);
3165 else
3166 AddrModes.clear();
3167
3168 return CanHandle;
3169 }
3170
3171 /// Combine the addressing modes we've collected into a single
3172 /// addressing mode.
3173 /// \return True iff we successfully combined them or we only had one so
3174 /// didn't need to combine them anyway.
combineAddrModes()3175 bool combineAddrModes() {
3176 // If we have no AddrModes then they can't be combined.
3177 if (AddrModes.size() == 0)
3178 return false;
3179
3180 // A single AddrMode can trivially be combined.
3181 if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
3182 return true;
3183
3184 // If the AddrModes we collected are all just equal to the value they are
3185 // derived from then combining them wouldn't do anything useful.
3186 if (AllAddrModesTrivial)
3187 return false;
3188
3189 if (!addrModeCombiningAllowed())
3190 return false;
3191
3192 // Build a map between <original value, basic block where we saw it> to
3193 // value of base register.
3194 // Bail out if there is no common type.
3195 FoldAddrToValueMapping Map;
3196 if (!initializeMap(Map))
3197 return false;
3198
3199 Value *CommonValue = findCommon(Map);
3200 if (CommonValue)
3201 AddrModes[0].SetCombinedField(DifferentField, CommonValue, AddrModes);
3202 return CommonValue != nullptr;
3203 }
3204
3205 private:
3206 /// Initialize Map with anchor values. For address seen
3207 /// we set the value of different field saw in this address.
3208 /// At the same time we find a common type for different field we will
3209 /// use to create new Phi/Select nodes. Keep it in CommonType field.
3210 /// Return false if there is no common type found.
initializeMap(FoldAddrToValueMapping & Map)3211 bool initializeMap(FoldAddrToValueMapping &Map) {
3212 // Keep track of keys where the value is null. We will need to replace it
3213 // with constant null when we know the common type.
3214 SmallVector<Value *, 2> NullValue;
3215 Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
3216 for (auto &AM : AddrModes) {
3217 Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy);
3218 if (DV) {
3219 auto *Type = DV->getType();
3220 if (CommonType && CommonType != Type)
3221 return false;
3222 CommonType = Type;
3223 Map[AM.OriginalValue] = DV;
3224 } else {
3225 NullValue.push_back(AM.OriginalValue);
3226 }
3227 }
3228 assert(CommonType && "At least one non-null value must be!");
3229 for (auto *V : NullValue)
3230 Map[V] = Constant::getNullValue(CommonType);
3231 return true;
3232 }
3233
3234 /// We have mapping between value A and other value B where B was a field in
3235 /// addressing mode represented by A. Also we have an original value C
3236 /// representing an address we start with. Traversing from C through phi and
3237 /// selects we ended up with A's in a map. This utility function tries to find
3238 /// a value V which is a field in addressing mode C and traversing through phi
3239 /// nodes and selects we will end up in corresponded values B in a map.
3240 /// The utility will create a new Phi/Selects if needed.
3241 // The simple example looks as follows:
3242 // BB1:
3243 // p1 = b1 + 40
3244 // br cond BB2, BB3
3245 // BB2:
3246 // p2 = b2 + 40
3247 // br BB3
3248 // BB3:
3249 // p = phi [p1, BB1], [p2, BB2]
3250 // v = load p
3251 // Map is
3252 // p1 -> b1
3253 // p2 -> b2
3254 // Request is
3255 // p -> ?
3256 // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3.
findCommon(FoldAddrToValueMapping & Map)3257 Value *findCommon(FoldAddrToValueMapping &Map) {
3258 // Tracks the simplification of newly created phi nodes. The reason we use
3259 // this mapping is because we will add new created Phi nodes in AddrToBase.
3260 // Simplification of Phi nodes is recursive, so some Phi node may
3261 // be simplified after we added it to AddrToBase. In reality this
3262 // simplification is possible only if original phi/selects were not
3263 // simplified yet.
3264 // Using this mapping we can find the current value in AddrToBase.
3265 SimplificationTracker ST(SQ);
3266
3267 // First step, DFS to create PHI nodes for all intermediate blocks.
3268 // Also fill traverse order for the second step.
3269 SmallVector<Value *, 32> TraverseOrder;
3270 InsertPlaceholders(Map, TraverseOrder, ST);
3271
3272 // Second Step, fill new nodes by merged values and simplify if possible.
3273 FillPlaceholders(Map, TraverseOrder, ST);
3274
3275 if (!AddrSinkNewSelects && ST.countNewSelectNodes() > 0) {
3276 ST.destroyNewNodes(CommonType);
3277 return nullptr;
3278 }
3279
3280 // Now we'd like to match New Phi nodes to existed ones.
3281 unsigned PhiNotMatchedCount = 0;
3282 if (!MatchPhiSet(ST, AddrSinkNewPhis, PhiNotMatchedCount)) {
3283 ST.destroyNewNodes(CommonType);
3284 return nullptr;
3285 }
3286
3287 auto *Result = ST.Get(Map.find(Original)->second);
3288 if (Result) {
3289 NumMemoryInstsPhiCreated += ST.countNewPhiNodes() + PhiNotMatchedCount;
3290 NumMemoryInstsSelectCreated += ST.countNewSelectNodes();
3291 }
3292 return Result;
3293 }
3294
3295 /// Try to match PHI node to Candidate.
3296 /// Matcher tracks the matched Phi nodes.
MatchPhiNode(PHINode * PHI,PHINode * Candidate,SmallSetVector<PHIPair,8> & Matcher,PhiNodeSet & PhiNodesToMatch)3297 bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
3298 SmallSetVector<PHIPair, 8> &Matcher,
3299 PhiNodeSet &PhiNodesToMatch) {
3300 SmallVector<PHIPair, 8> WorkList;
3301 Matcher.insert({ PHI, Candidate });
3302 SmallSet<PHINode *, 8> MatchedPHIs;
3303 MatchedPHIs.insert(PHI);
3304 WorkList.push_back({ PHI, Candidate });
3305 SmallSet<PHIPair, 8> Visited;
3306 while (!WorkList.empty()) {
3307 auto Item = WorkList.pop_back_val();
3308 if (!Visited.insert(Item).second)
3309 continue;
3310 // We iterate over all incoming values to Phi to compare them.
3311 // If values are different and both of them Phi and the first one is a
3312 // Phi we added (subject to match) and both of them is in the same basic
3313 // block then we can match our pair if values match. So we state that
3314 // these values match and add it to work list to verify that.
3315 for (auto B : Item.first->blocks()) {
3316 Value *FirstValue = Item.first->getIncomingValueForBlock(B);
3317 Value *SecondValue = Item.second->getIncomingValueForBlock(B);
3318 if (FirstValue == SecondValue)
3319 continue;
3320
3321 PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue);
3322 PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue);
3323
3324 // One of them is not Phi or
3325 // The first one is not Phi node from the set we'd like to match or
3326 // Phi nodes from different basic blocks then
3327 // we will not be able to match.
3328 if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) ||
3329 FirstPhi->getParent() != SecondPhi->getParent())
3330 return false;
3331
3332 // If we already matched them then continue.
3333 if (Matcher.count({ FirstPhi, SecondPhi }))
3334 continue;
3335 // So the values are different and does not match. So we need them to
3336 // match. (But we register no more than one match per PHI node, so that
3337 // we won't later try to replace them twice.)
3338 if (!MatchedPHIs.insert(FirstPhi).second)
3339 Matcher.insert({ FirstPhi, SecondPhi });
3340 // But me must check it.
3341 WorkList.push_back({ FirstPhi, SecondPhi });
3342 }
3343 }
3344 return true;
3345 }
3346
3347 /// For the given set of PHI nodes (in the SimplificationTracker) try
3348 /// to find their equivalents.
3349 /// Returns false if this matching fails and creation of new Phi is disabled.
MatchPhiSet(SimplificationTracker & ST,bool AllowNewPhiNodes,unsigned & PhiNotMatchedCount)3350 bool MatchPhiSet(SimplificationTracker &ST, bool AllowNewPhiNodes,
3351 unsigned &PhiNotMatchedCount) {
3352 // Matched and PhiNodesToMatch iterate their elements in a deterministic
3353 // order, so the replacements (ReplacePhi) are also done in a deterministic
3354 // order.
3355 SmallSetVector<PHIPair, 8> Matched;
3356 SmallPtrSet<PHINode *, 8> WillNotMatch;
3357 PhiNodeSet &PhiNodesToMatch = ST.newPhiNodes();
3358 while (PhiNodesToMatch.size()) {
3359 PHINode *PHI = *PhiNodesToMatch.begin();
3360
3361 // Add us, if no Phi nodes in the basic block we do not match.
3362 WillNotMatch.clear();
3363 WillNotMatch.insert(PHI);
3364
3365 // Traverse all Phis until we found equivalent or fail to do that.
3366 bool IsMatched = false;
3367 for (auto &P : PHI->getParent()->phis()) {
3368 if (&P == PHI)
3369 continue;
3370 if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch)))
3371 break;
3372 // If it does not match, collect all Phi nodes from matcher.
3373 // if we end up with no match, them all these Phi nodes will not match
3374 // later.
3375 for (auto M : Matched)
3376 WillNotMatch.insert(M.first);
3377 Matched.clear();
3378 }
3379 if (IsMatched) {
3380 // Replace all matched values and erase them.
3381 for (auto MV : Matched)
3382 ST.ReplacePhi(MV.first, MV.second);
3383 Matched.clear();
3384 continue;
3385 }
3386 // If we are not allowed to create new nodes then bail out.
3387 if (!AllowNewPhiNodes)
3388 return false;
3389 // Just remove all seen values in matcher. They will not match anything.
3390 PhiNotMatchedCount += WillNotMatch.size();
3391 for (auto *P : WillNotMatch)
3392 PhiNodesToMatch.erase(P);
3393 }
3394 return true;
3395 }
3396 /// Fill the placeholders with values from predecessors and simplify them.
FillPlaceholders(FoldAddrToValueMapping & Map,SmallVectorImpl<Value * > & TraverseOrder,SimplificationTracker & ST)3397 void FillPlaceholders(FoldAddrToValueMapping &Map,
3398 SmallVectorImpl<Value *> &TraverseOrder,
3399 SimplificationTracker &ST) {
3400 while (!TraverseOrder.empty()) {
3401 Value *Current = TraverseOrder.pop_back_val();
3402 assert(Map.find(Current) != Map.end() && "No node to fill!!!");
3403 Value *V = Map[Current];
3404
3405 if (SelectInst *Select = dyn_cast<SelectInst>(V)) {
3406 // CurrentValue also must be Select.
3407 auto *CurrentSelect = cast<SelectInst>(Current);
3408 auto *TrueValue = CurrentSelect->getTrueValue();
3409 assert(Map.find(TrueValue) != Map.end() && "No True Value!");
3410 Select->setTrueValue(ST.Get(Map[TrueValue]));
3411 auto *FalseValue = CurrentSelect->getFalseValue();
3412 assert(Map.find(FalseValue) != Map.end() && "No False Value!");
3413 Select->setFalseValue(ST.Get(Map[FalseValue]));
3414 } else {
3415 // Must be a Phi node then.
3416 PHINode *PHI = cast<PHINode>(V);
3417 auto *CurrentPhi = dyn_cast<PHINode>(Current);
3418 // Fill the Phi node with values from predecessors.
3419 for (auto B : predecessors(PHI->getParent())) {
3420 Value *PV = CurrentPhi->getIncomingValueForBlock(B);
3421 assert(Map.find(PV) != Map.end() && "No predecessor Value!");
3422 PHI->addIncoming(ST.Get(Map[PV]), B);
3423 }
3424 }
3425 Map[Current] = ST.Simplify(V);
3426 }
3427 }
3428
3429 /// Starting from original value recursively iterates over def-use chain up to
3430 /// known ending values represented in a map. For each traversed phi/select
3431 /// inserts a placeholder Phi or Select.
3432 /// Reports all new created Phi/Select nodes by adding them to set.
3433 /// Also reports and order in what values have been traversed.
InsertPlaceholders(FoldAddrToValueMapping & Map,SmallVectorImpl<Value * > & TraverseOrder,SimplificationTracker & ST)3434 void InsertPlaceholders(FoldAddrToValueMapping &Map,
3435 SmallVectorImpl<Value *> &TraverseOrder,
3436 SimplificationTracker &ST) {
3437 SmallVector<Value *, 32> Worklist;
3438 assert((isa<PHINode>(Original) || isa<SelectInst>(Original)) &&
3439 "Address must be a Phi or Select node");
3440 auto *Dummy = UndefValue::get(CommonType);
3441 Worklist.push_back(Original);
3442 while (!Worklist.empty()) {
3443 Value *Current = Worklist.pop_back_val();
3444 // if it is already visited or it is an ending value then skip it.
3445 if (Map.find(Current) != Map.end())
3446 continue;
3447 TraverseOrder.push_back(Current);
3448
3449 // CurrentValue must be a Phi node or select. All others must be covered
3450 // by anchors.
3451 if (SelectInst *CurrentSelect = dyn_cast<SelectInst>(Current)) {
3452 // Is it OK to get metadata from OrigSelect?!
3453 // Create a Select placeholder with dummy value.
3454 SelectInst *Select = SelectInst::Create(
3455 CurrentSelect->getCondition(), Dummy, Dummy,
3456 CurrentSelect->getName(), CurrentSelect, CurrentSelect);
3457 Map[Current] = Select;
3458 ST.insertNewSelect(Select);
3459 // We are interested in True and False values.
3460 Worklist.push_back(CurrentSelect->getTrueValue());
3461 Worklist.push_back(CurrentSelect->getFalseValue());
3462 } else {
3463 // It must be a Phi node then.
3464 PHINode *CurrentPhi = cast<PHINode>(Current);
3465 unsigned PredCount = CurrentPhi->getNumIncomingValues();
3466 PHINode *PHI =
3467 PHINode::Create(CommonType, PredCount, "sunk_phi", CurrentPhi);
3468 Map[Current] = PHI;
3469 ST.insertNewPhi(PHI);
3470 for (Value *P : CurrentPhi->incoming_values())
3471 Worklist.push_back(P);
3472 }
3473 }
3474 }
3475
addrModeCombiningAllowed()3476 bool addrModeCombiningAllowed() {
3477 if (DisableComplexAddrModes)
3478 return false;
3479 switch (DifferentField) {
3480 default:
3481 return false;
3482 case ExtAddrMode::BaseRegField:
3483 return AddrSinkCombineBaseReg;
3484 case ExtAddrMode::BaseGVField:
3485 return AddrSinkCombineBaseGV;
3486 case ExtAddrMode::BaseOffsField:
3487 return AddrSinkCombineBaseOffs;
3488 case ExtAddrMode::ScaledRegField:
3489 return AddrSinkCombineScaledReg;
3490 }
3491 }
3492 };
3493 } // end anonymous namespace
3494
3495 /// Try adding ScaleReg*Scale to the current addressing mode.
3496 /// Return true and update AddrMode if this addr mode is legal for the target,
3497 /// false if not.
matchScaledValue(Value * ScaleReg,int64_t Scale,unsigned Depth)3498 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3499 unsigned Depth) {
3500 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3501 // mode. Just process that directly.
3502 if (Scale == 1)
3503 return matchAddr(ScaleReg, Depth);
3504
3505 // If the scale is 0, it takes nothing to add this.
3506 if (Scale == 0)
3507 return true;
3508
3509 // If we already have a scale of this value, we can add to it, otherwise, we
3510 // need an available scale field.
3511 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3512 return false;
3513
3514 ExtAddrMode TestAddrMode = AddrMode;
3515
3516 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3517 // [A+B + A*7] -> [B+A*8].
3518 TestAddrMode.Scale += Scale;
3519 TestAddrMode.ScaledReg = ScaleReg;
3520
3521 // If the new address isn't legal, bail out.
3522 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3523 return false;
3524
3525 // It was legal, so commit it.
3526 AddrMode = TestAddrMode;
3527
3528 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3529 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3530 // X*Scale + C*Scale to addr mode.
3531 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3532 if (isa<Instruction>(ScaleReg) && // not a constant expr.
3533 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3534 TestAddrMode.InBounds = false;
3535 TestAddrMode.ScaledReg = AddLHS;
3536 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3537
3538 // If this addressing mode is legal, commit it and remember that we folded
3539 // this instruction.
3540 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3541 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3542 AddrMode = TestAddrMode;
3543 return true;
3544 }
3545 }
3546
3547 // Otherwise, not (x+c)*scale, just return what we have.
3548 return true;
3549 }
3550
3551 /// This is a little filter, which returns true if an addressing computation
3552 /// involving I might be folded into a load/store accessing it.
3553 /// This doesn't need to be perfect, but needs to accept at least
3554 /// the set of instructions that MatchOperationAddr can.
MightBeFoldableInst(Instruction * I)3555 static bool MightBeFoldableInst(Instruction *I) {
3556 switch (I->getOpcode()) {
3557 case Instruction::BitCast:
3558 case Instruction::AddrSpaceCast:
3559 // Don't touch identity bitcasts.
3560 if (I->getType() == I->getOperand(0)->getType())
3561 return false;
3562 return I->getType()->isIntOrPtrTy();
3563 case Instruction::PtrToInt:
3564 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3565 return true;
3566 case Instruction::IntToPtr:
3567 // We know the input is intptr_t, so this is foldable.
3568 return true;
3569 case Instruction::Add:
3570 return true;
3571 case Instruction::Mul:
3572 case Instruction::Shl:
3573 // Can only handle X*C and X << C.
3574 return isa<ConstantInt>(I->getOperand(1));
3575 case Instruction::GetElementPtr:
3576 return true;
3577 default:
3578 return false;
3579 }
3580 }
3581
3582 /// Check whether or not \p Val is a legal instruction for \p TLI.
3583 /// \note \p Val is assumed to be the product of some type promotion.
3584 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3585 /// to be legal, as the non-promoted value would have had the same state.
isPromotedInstructionLegal(const TargetLowering & TLI,const DataLayout & DL,Value * Val)3586 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
3587 const DataLayout &DL, Value *Val) {
3588 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3589 if (!PromotedInst)
3590 return false;
3591 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3592 // If the ISDOpcode is undefined, it was undefined before the promotion.
3593 if (!ISDOpcode)
3594 return true;
3595 // Otherwise, check if the promoted instruction is legal or not.
3596 return TLI.isOperationLegalOrCustom(
3597 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3598 }
3599
3600 namespace {
3601
3602 /// Hepler class to perform type promotion.
3603 class TypePromotionHelper {
3604 /// Utility function to add a promoted instruction \p ExtOpnd to
3605 /// \p PromotedInsts and record the type of extension we have seen.
addPromotedInst(InstrToOrigTy & PromotedInsts,Instruction * ExtOpnd,bool IsSExt)3606 static void addPromotedInst(InstrToOrigTy &PromotedInsts,
3607 Instruction *ExtOpnd,
3608 bool IsSExt) {
3609 ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
3610 InstrToOrigTy::iterator It = PromotedInsts.find(ExtOpnd);
3611 if (It != PromotedInsts.end()) {
3612 // If the new extension is same as original, the information in
3613 // PromotedInsts[ExtOpnd] is still correct.
3614 if (It->second.getInt() == ExtTy)
3615 return;
3616
3617 // Now the new extension is different from old extension, we make
3618 // the type information invalid by setting extension type to
3619 // BothExtension.
3620 ExtTy = BothExtension;
3621 }
3622 PromotedInsts[ExtOpnd] = TypeIsSExt(ExtOpnd->getType(), ExtTy);
3623 }
3624
3625 /// Utility function to query the original type of instruction \p Opnd
3626 /// with a matched extension type. If the extension doesn't match, we
3627 /// cannot use the information we had on the original type.
3628 /// BothExtension doesn't match any extension type.
getOrigType(const InstrToOrigTy & PromotedInsts,Instruction * Opnd,bool IsSExt)3629 static const Type *getOrigType(const InstrToOrigTy &PromotedInsts,
3630 Instruction *Opnd,
3631 bool IsSExt) {
3632 ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
3633 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3634 if (It != PromotedInsts.end() && It->second.getInt() == ExtTy)
3635 return It->second.getPointer();
3636 return nullptr;
3637 }
3638
3639 /// Utility function to check whether or not a sign or zero extension
3640 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3641 /// either using the operands of \p Inst or promoting \p Inst.
3642 /// The type of the extension is defined by \p IsSExt.
3643 /// In other words, check if:
3644 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3645 /// #1 Promotion applies:
3646 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3647 /// #2 Operand reuses:
3648 /// ext opnd1 to ConsideredExtType.
3649 /// \p PromotedInsts maps the instructions to their type before promotion.
3650 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3651 const InstrToOrigTy &PromotedInsts, bool IsSExt);
3652
3653 /// Utility function to determine if \p OpIdx should be promoted when
3654 /// promoting \p Inst.
shouldExtOperand(const Instruction * Inst,int OpIdx)3655 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3656 return !(isa<SelectInst>(Inst) && OpIdx == 0);
3657 }
3658
3659 /// Utility function to promote the operand of \p Ext when this
3660 /// operand is a promotable trunc or sext or zext.
3661 /// \p PromotedInsts maps the instructions to their type before promotion.
3662 /// \p CreatedInstsCost[out] contains the cost of all instructions
3663 /// created to promote the operand of Ext.
3664 /// Newly added extensions are inserted in \p Exts.
3665 /// Newly added truncates are inserted in \p Truncs.
3666 /// Should never be called directly.
3667 /// \return The promoted value which is used instead of Ext.
3668 static Value *promoteOperandForTruncAndAnyExt(
3669 Instruction *Ext, TypePromotionTransaction &TPT,
3670 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3671 SmallVectorImpl<Instruction *> *Exts,
3672 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3673
3674 /// Utility function to promote the operand of \p Ext when this
3675 /// operand is promotable and is not a supported trunc or sext.
3676 /// \p PromotedInsts maps the instructions to their type before promotion.
3677 /// \p CreatedInstsCost[out] contains the cost of all the instructions
3678 /// created to promote the operand of Ext.
3679 /// Newly added extensions are inserted in \p Exts.
3680 /// Newly added truncates are inserted in \p Truncs.
3681 /// Should never be called directly.
3682 /// \return The promoted value which is used instead of Ext.
3683 static Value *promoteOperandForOther(Instruction *Ext,
3684 TypePromotionTransaction &TPT,
3685 InstrToOrigTy &PromotedInsts,
3686 unsigned &CreatedInstsCost,
3687 SmallVectorImpl<Instruction *> *Exts,
3688 SmallVectorImpl<Instruction *> *Truncs,
3689 const TargetLowering &TLI, bool IsSExt);
3690
3691 /// \see promoteOperandForOther.
signExtendOperandForOther(Instruction * Ext,TypePromotionTransaction & TPT,InstrToOrigTy & PromotedInsts,unsigned & CreatedInstsCost,SmallVectorImpl<Instruction * > * Exts,SmallVectorImpl<Instruction * > * Truncs,const TargetLowering & TLI)3692 static Value *signExtendOperandForOther(
3693 Instruction *Ext, TypePromotionTransaction &TPT,
3694 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3695 SmallVectorImpl<Instruction *> *Exts,
3696 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3697 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3698 Exts, Truncs, TLI, true);
3699 }
3700
3701 /// \see promoteOperandForOther.
zeroExtendOperandForOther(Instruction * Ext,TypePromotionTransaction & TPT,InstrToOrigTy & PromotedInsts,unsigned & CreatedInstsCost,SmallVectorImpl<Instruction * > * Exts,SmallVectorImpl<Instruction * > * Truncs,const TargetLowering & TLI)3702 static Value *zeroExtendOperandForOther(
3703 Instruction *Ext, TypePromotionTransaction &TPT,
3704 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3705 SmallVectorImpl<Instruction *> *Exts,
3706 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3707 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3708 Exts, Truncs, TLI, false);
3709 }
3710
3711 public:
3712 /// Type for the utility function that promotes the operand of Ext.
3713 using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
3714 InstrToOrigTy &PromotedInsts,
3715 unsigned &CreatedInstsCost,
3716 SmallVectorImpl<Instruction *> *Exts,
3717 SmallVectorImpl<Instruction *> *Truncs,
3718 const TargetLowering &TLI);
3719
3720 /// Given a sign/zero extend instruction \p Ext, return the appropriate
3721 /// action to promote the operand of \p Ext instead of using Ext.
3722 /// \return NULL if no promotable action is possible with the current
3723 /// sign extension.
3724 /// \p InsertedInsts keeps track of all the instructions inserted by the
3725 /// other CodeGenPrepare optimizations. This information is important
3726 /// because we do not want to promote these instructions as CodeGenPrepare
3727 /// will reinsert them later. Thus creating an infinite loop: create/remove.
3728 /// \p PromotedInsts maps the instructions to their type before promotion.
3729 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3730 const TargetLowering &TLI,
3731 const InstrToOrigTy &PromotedInsts);
3732 };
3733
3734 } // end anonymous namespace
3735
canGetThrough(const Instruction * Inst,Type * ConsideredExtType,const InstrToOrigTy & PromotedInsts,bool IsSExt)3736 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3737 Type *ConsideredExtType,
3738 const InstrToOrigTy &PromotedInsts,
3739 bool IsSExt) {
3740 // The promotion helper does not know how to deal with vector types yet.
3741 // To be able to fix that, we would need to fix the places where we
3742 // statically extend, e.g., constants and such.
3743 if (Inst->getType()->isVectorTy())
3744 return false;
3745
3746 // We can always get through zext.
3747 if (isa<ZExtInst>(Inst))
3748 return true;
3749
3750 // sext(sext) is ok too.
3751 if (IsSExt && isa<SExtInst>(Inst))
3752 return true;
3753
3754 // We can get through binary operator, if it is legal. In other words, the
3755 // binary operator must have a nuw or nsw flag.
3756 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3757 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3758 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3759 (IsSExt && BinOp->hasNoSignedWrap())))
3760 return true;
3761
3762 // ext(and(opnd, cst)) --> and(ext(opnd), ext(cst))
3763 if ((Inst->getOpcode() == Instruction::And ||
3764 Inst->getOpcode() == Instruction::Or))
3765 return true;
3766
3767 // ext(xor(opnd, cst)) --> xor(ext(opnd), ext(cst))
3768 if (Inst->getOpcode() == Instruction::Xor) {
3769 const ConstantInt *Cst = dyn_cast<ConstantInt>(Inst->getOperand(1));
3770 // Make sure it is not a NOT.
3771 if (Cst && !Cst->getValue().isAllOnesValue())
3772 return true;
3773 }
3774
3775 // zext(shrl(opnd, cst)) --> shrl(zext(opnd), zext(cst))
3776 // It may change a poisoned value into a regular value, like
3777 // zext i32 (shrl i8 %val, 12) --> shrl i32 (zext i8 %val), 12
3778 // poisoned value regular value
3779 // It should be OK since undef covers valid value.
3780 if (Inst->getOpcode() == Instruction::LShr && !IsSExt)
3781 return true;
3782
3783 // and(ext(shl(opnd, cst)), cst) --> and(shl(ext(opnd), ext(cst)), cst)
3784 // It may change a poisoned value into a regular value, like
3785 // zext i32 (shl i8 %val, 12) --> shl i32 (zext i8 %val), 12
3786 // poisoned value regular value
3787 // It should be OK since undef covers valid value.
3788 if (Inst->getOpcode() == Instruction::Shl && Inst->hasOneUse()) {
3789 const Instruction *ExtInst =
3790 dyn_cast<const Instruction>(*Inst->user_begin());
3791 if (ExtInst->hasOneUse()) {
3792 const Instruction *AndInst =
3793 dyn_cast<const Instruction>(*ExtInst->user_begin());
3794 if (AndInst && AndInst->getOpcode() == Instruction::And) {
3795 const ConstantInt *Cst = dyn_cast<ConstantInt>(AndInst->getOperand(1));
3796 if (Cst &&
3797 Cst->getValue().isIntN(Inst->getType()->getIntegerBitWidth()))
3798 return true;
3799 }
3800 }
3801 }
3802
3803 // Check if we can do the following simplification.
3804 // ext(trunc(opnd)) --> ext(opnd)
3805 if (!isa<TruncInst>(Inst))
3806 return false;
3807
3808 Value *OpndVal = Inst->getOperand(0);
3809 // Check if we can use this operand in the extension.
3810 // If the type is larger than the result type of the extension, we cannot.
3811 if (!OpndVal->getType()->isIntegerTy() ||
3812 OpndVal->getType()->getIntegerBitWidth() >
3813 ConsideredExtType->getIntegerBitWidth())
3814 return false;
3815
3816 // If the operand of the truncate is not an instruction, we will not have
3817 // any information on the dropped bits.
3818 // (Actually we could for constant but it is not worth the extra logic).
3819 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3820 if (!Opnd)
3821 return false;
3822
3823 // Check if the source of the type is narrow enough.
3824 // I.e., check that trunc just drops extended bits of the same kind of
3825 // the extension.
3826 // #1 get the type of the operand and check the kind of the extended bits.
3827 const Type *OpndType = getOrigType(PromotedInsts, Opnd, IsSExt);
3828 if (OpndType)
3829 ;
3830 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3831 OpndType = Opnd->getOperand(0)->getType();
3832 else
3833 return false;
3834
3835 // #2 check that the truncate just drops extended bits.
3836 return Inst->getType()->getIntegerBitWidth() >=
3837 OpndType->getIntegerBitWidth();
3838 }
3839
getAction(Instruction * Ext,const SetOfInstrs & InsertedInsts,const TargetLowering & TLI,const InstrToOrigTy & PromotedInsts)3840 TypePromotionHelper::Action TypePromotionHelper::getAction(
3841 Instruction *Ext, const SetOfInstrs &InsertedInsts,
3842 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3843 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
3844 "Unexpected instruction type");
3845 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3846 Type *ExtTy = Ext->getType();
3847 bool IsSExt = isa<SExtInst>(Ext);
3848 // If the operand of the extension is not an instruction, we cannot
3849 // get through.
3850 // If it, check we can get through.
3851 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3852 return nullptr;
3853
3854 // Do not promote if the operand has been added by codegenprepare.
3855 // Otherwise, it means we are undoing an optimization that is likely to be
3856 // redone, thus causing potential infinite loop.
3857 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3858 return nullptr;
3859
3860 // SExt or Trunc instructions.
3861 // Return the related handler.
3862 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3863 isa<ZExtInst>(ExtOpnd))
3864 return promoteOperandForTruncAndAnyExt;
3865
3866 // Regular instruction.
3867 // Abort early if we will have to insert non-free instructions.
3868 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3869 return nullptr;
3870 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3871 }
3872
promoteOperandForTruncAndAnyExt(Instruction * SExt,TypePromotionTransaction & TPT,InstrToOrigTy & PromotedInsts,unsigned & CreatedInstsCost,SmallVectorImpl<Instruction * > * Exts,SmallVectorImpl<Instruction * > * Truncs,const TargetLowering & TLI)3873 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3874 Instruction *SExt, TypePromotionTransaction &TPT,
3875 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3876 SmallVectorImpl<Instruction *> *Exts,
3877 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3878 // By construction, the operand of SExt is an instruction. Otherwise we cannot
3879 // get through it and this method should not be called.
3880 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
3881 Value *ExtVal = SExt;
3882 bool HasMergedNonFreeExt = false;
3883 if (isa<ZExtInst>(SExtOpnd)) {
3884 // Replace s|zext(zext(opnd))
3885 // => zext(opnd).
3886 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
3887 Value *ZExt =
3888 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
3889 TPT.replaceAllUsesWith(SExt, ZExt);
3890 TPT.eraseInstruction(SExt);
3891 ExtVal = ZExt;
3892 } else {
3893 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
3894 // => z|sext(opnd).
3895 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
3896 }
3897 CreatedInstsCost = 0;
3898
3899 // Remove dead code.
3900 if (SExtOpnd->use_empty())
3901 TPT.eraseInstruction(SExtOpnd);
3902
3903 // Check if the extension is still needed.
3904 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
3905 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
3906 if (ExtInst) {
3907 if (Exts)
3908 Exts->push_back(ExtInst);
3909 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
3910 }
3911 return ExtVal;
3912 }
3913
3914 // At this point we have: ext ty opnd to ty.
3915 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
3916 Value *NextVal = ExtInst->getOperand(0);
3917 TPT.eraseInstruction(ExtInst, NextVal);
3918 return NextVal;
3919 }
3920
promoteOperandForOther(Instruction * Ext,TypePromotionTransaction & TPT,InstrToOrigTy & PromotedInsts,unsigned & CreatedInstsCost,SmallVectorImpl<Instruction * > * Exts,SmallVectorImpl<Instruction * > * Truncs,const TargetLowering & TLI,bool IsSExt)3921 Value *TypePromotionHelper::promoteOperandForOther(
3922 Instruction *Ext, TypePromotionTransaction &TPT,
3923 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3924 SmallVectorImpl<Instruction *> *Exts,
3925 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
3926 bool IsSExt) {
3927 // By construction, the operand of Ext is an instruction. Otherwise we cannot
3928 // get through it and this method should not be called.
3929 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
3930 CreatedInstsCost = 0;
3931 if (!ExtOpnd->hasOneUse()) {
3932 // ExtOpnd will be promoted.
3933 // All its uses, but Ext, will need to use a truncated value of the
3934 // promoted version.
3935 // Create the truncate now.
3936 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
3937 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
3938 // Insert it just after the definition.
3939 ITrunc->moveAfter(ExtOpnd);
3940 if (Truncs)
3941 Truncs->push_back(ITrunc);
3942 }
3943
3944 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
3945 // Restore the operand of Ext (which has been replaced by the previous call
3946 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
3947 TPT.setOperand(Ext, 0, ExtOpnd);
3948 }
3949
3950 // Get through the Instruction:
3951 // 1. Update its type.
3952 // 2. Replace the uses of Ext by Inst.
3953 // 3. Extend each operand that needs to be extended.
3954
3955 // Remember the original type of the instruction before promotion.
3956 // This is useful to know that the high bits are sign extended bits.
3957 addPromotedInst(PromotedInsts, ExtOpnd, IsSExt);
3958 // Step #1.
3959 TPT.mutateType(ExtOpnd, Ext->getType());
3960 // Step #2.
3961 TPT.replaceAllUsesWith(Ext, ExtOpnd);
3962 // Step #3.
3963 Instruction *ExtForOpnd = Ext;
3964
3965 LLVM_DEBUG(dbgs() << "Propagate Ext to operands\n");
3966 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
3967 ++OpIdx) {
3968 LLVM_DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
3969 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
3970 !shouldExtOperand(ExtOpnd, OpIdx)) {
3971 LLVM_DEBUG(dbgs() << "No need to propagate\n");
3972 continue;
3973 }
3974 // Check if we can statically extend the operand.
3975 Value *Opnd = ExtOpnd->getOperand(OpIdx);
3976 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
3977 LLVM_DEBUG(dbgs() << "Statically extend\n");
3978 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
3979 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3980 : Cst->getValue().zext(BitWidth);
3981 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3982 continue;
3983 }
3984 // UndefValue are typed, so we have to statically sign extend them.
3985 if (isa<UndefValue>(Opnd)) {
3986 LLVM_DEBUG(dbgs() << "Statically extend\n");
3987 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3988 continue;
3989 }
3990
3991 // Otherwise we have to explicitly sign extend the operand.
3992 // Check if Ext was reused to extend an operand.
3993 if (!ExtForOpnd) {
3994 // If yes, create a new one.
3995 LLVM_DEBUG(dbgs() << "More operands to ext\n");
3996 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3997 : TPT.createZExt(Ext, Opnd, Ext->getType());
3998 if (!isa<Instruction>(ValForExtOpnd)) {
3999 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
4000 continue;
4001 }
4002 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
4003 }
4004 if (Exts)
4005 Exts->push_back(ExtForOpnd);
4006 TPT.setOperand(ExtForOpnd, 0, Opnd);
4007
4008 // Move the sign extension before the insertion point.
4009 TPT.moveBefore(ExtForOpnd, ExtOpnd);
4010 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
4011 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
4012 // If more sext are required, new instructions will have to be created.
4013 ExtForOpnd = nullptr;
4014 }
4015 if (ExtForOpnd == Ext) {
4016 LLVM_DEBUG(dbgs() << "Extension is useless now\n");
4017 TPT.eraseInstruction(Ext);
4018 }
4019 return ExtOpnd;
4020 }
4021
4022 /// Check whether or not promoting an instruction to a wider type is profitable.
4023 /// \p NewCost gives the cost of extension instructions created by the
4024 /// promotion.
4025 /// \p OldCost gives the cost of extension instructions before the promotion
4026 /// plus the number of instructions that have been
4027 /// matched in the addressing mode the promotion.
4028 /// \p PromotedOperand is the value that has been promoted.
4029 /// \return True if the promotion is profitable, false otherwise.
isPromotionProfitable(unsigned NewCost,unsigned OldCost,Value * PromotedOperand) const4030 bool AddressingModeMatcher::isPromotionProfitable(
4031 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
4032 LLVM_DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost
4033 << '\n');
4034 // The cost of the new extensions is greater than the cost of the
4035 // old extension plus what we folded.
4036 // This is not profitable.
4037 if (NewCost > OldCost)
4038 return false;
4039 if (NewCost < OldCost)
4040 return true;
4041 // The promotion is neutral but it may help folding the sign extension in
4042 // loads for instance.
4043 // Check that we did not create an illegal instruction.
4044 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
4045 }
4046
4047 /// Given an instruction or constant expr, see if we can fold the operation
4048 /// into the addressing mode. If so, update the addressing mode and return
4049 /// true, otherwise return false without modifying AddrMode.
4050 /// If \p MovedAway is not NULL, it contains the information of whether or
4051 /// not AddrInst has to be folded into the addressing mode on success.
4052 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
4053 /// because it has been moved away.
4054 /// Thus AddrInst must not be added in the matched instructions.
4055 /// This state can happen when AddrInst is a sext, since it may be moved away.
4056 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
4057 /// not be referenced anymore.
matchOperationAddr(User * AddrInst,unsigned Opcode,unsigned Depth,bool * MovedAway)4058 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
4059 unsigned Depth,
4060 bool *MovedAway) {
4061 // Avoid exponential behavior on extremely deep expression trees.
4062 if (Depth >= 5) return false;
4063
4064 // By default, all matched instructions stay in place.
4065 if (MovedAway)
4066 *MovedAway = false;
4067
4068 switch (Opcode) {
4069 case Instruction::PtrToInt:
4070 // PtrToInt is always a noop, as we know that the int type is pointer sized.
4071 return matchAddr(AddrInst->getOperand(0), Depth);
4072 case Instruction::IntToPtr: {
4073 auto AS = AddrInst->getType()->getPointerAddressSpace();
4074 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
4075 // This inttoptr is a no-op if the integer type is pointer sized.
4076 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
4077 return matchAddr(AddrInst->getOperand(0), Depth);
4078 return false;
4079 }
4080 case Instruction::BitCast:
4081 // BitCast is always a noop, and we can handle it as long as it is
4082 // int->int or pointer->pointer (we don't want int<->fp or something).
4083 if (AddrInst->getOperand(0)->getType()->isIntOrPtrTy() &&
4084 // Don't touch identity bitcasts. These were probably put here by LSR,
4085 // and we don't want to mess around with them. Assume it knows what it
4086 // is doing.
4087 AddrInst->getOperand(0)->getType() != AddrInst->getType())
4088 return matchAddr(AddrInst->getOperand(0), Depth);
4089 return false;
4090 case Instruction::AddrSpaceCast: {
4091 unsigned SrcAS
4092 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
4093 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
4094 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
4095 return matchAddr(AddrInst->getOperand(0), Depth);
4096 return false;
4097 }
4098 case Instruction::Add: {
4099 // Check to see if we can merge in the RHS then the LHS. If so, we win.
4100 ExtAddrMode BackupAddrMode = AddrMode;
4101 unsigned OldSize = AddrModeInsts.size();
4102 // Start a transaction at this point.
4103 // The LHS may match but not the RHS.
4104 // Therefore, we need a higher level restoration point to undo partially
4105 // matched operation.
4106 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4107 TPT.getRestorationPoint();
4108
4109 AddrMode.InBounds = false;
4110 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
4111 matchAddr(AddrInst->getOperand(0), Depth+1))
4112 return true;
4113
4114 // Restore the old addr mode info.
4115 AddrMode = BackupAddrMode;
4116 AddrModeInsts.resize(OldSize);
4117 TPT.rollback(LastKnownGood);
4118
4119 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
4120 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
4121 matchAddr(AddrInst->getOperand(1), Depth+1))
4122 return true;
4123
4124 // Otherwise we definitely can't merge the ADD in.
4125 AddrMode = BackupAddrMode;
4126 AddrModeInsts.resize(OldSize);
4127 TPT.rollback(LastKnownGood);
4128 break;
4129 }
4130 //case Instruction::Or:
4131 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
4132 //break;
4133 case Instruction::Mul:
4134 case Instruction::Shl: {
4135 // Can only handle X*C and X << C.
4136 AddrMode.InBounds = false;
4137 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
4138 if (!RHS || RHS->getBitWidth() > 64)
4139 return false;
4140 int64_t Scale = RHS->getSExtValue();
4141 if (Opcode == Instruction::Shl)
4142 Scale = 1LL << Scale;
4143
4144 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
4145 }
4146 case Instruction::GetElementPtr: {
4147 // Scan the GEP. We check it if it contains constant offsets and at most
4148 // one variable offset.
4149 int VariableOperand = -1;
4150 unsigned VariableScale = 0;
4151
4152 int64_t ConstantOffset = 0;
4153 gep_type_iterator GTI = gep_type_begin(AddrInst);
4154 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
4155 if (StructType *STy = GTI.getStructTypeOrNull()) {
4156 const StructLayout *SL = DL.getStructLayout(STy);
4157 unsigned Idx =
4158 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
4159 ConstantOffset += SL->getElementOffset(Idx);
4160 } else {
4161 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
4162 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
4163 const APInt &CVal = CI->getValue();
4164 if (CVal.getMinSignedBits() <= 64) {
4165 ConstantOffset += CVal.getSExtValue() * TypeSize;
4166 continue;
4167 }
4168 }
4169 if (TypeSize) { // Scales of zero don't do anything.
4170 // We only allow one variable index at the moment.
4171 if (VariableOperand != -1)
4172 return false;
4173
4174 // Remember the variable index.
4175 VariableOperand = i;
4176 VariableScale = TypeSize;
4177 }
4178 }
4179 }
4180
4181 // A common case is for the GEP to only do a constant offset. In this case,
4182 // just add it to the disp field and check validity.
4183 if (VariableOperand == -1) {
4184 AddrMode.BaseOffs += ConstantOffset;
4185 if (ConstantOffset == 0 ||
4186 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
4187 // Check to see if we can fold the base pointer in too.
4188 if (matchAddr(AddrInst->getOperand(0), Depth+1)) {
4189 if (!cast<GEPOperator>(AddrInst)->isInBounds())
4190 AddrMode.InBounds = false;
4191 return true;
4192 }
4193 } else if (EnableGEPOffsetSplit && isa<GetElementPtrInst>(AddrInst) &&
4194 TLI.shouldConsiderGEPOffsetSplit() && Depth == 0 &&
4195 ConstantOffset > 0) {
4196 // Record GEPs with non-zero offsets as candidates for splitting in the
4197 // event that the offset cannot fit into the r+i addressing mode.
4198 // Simple and common case that only one GEP is used in calculating the
4199 // address for the memory access.
4200 Value *Base = AddrInst->getOperand(0);
4201 auto *BaseI = dyn_cast<Instruction>(Base);
4202 auto *GEP = cast<GetElementPtrInst>(AddrInst);
4203 if (isa<Argument>(Base) || isa<GlobalValue>(Base) ||
4204 (BaseI && !isa<CastInst>(BaseI) &&
4205 !isa<GetElementPtrInst>(BaseI))) {
4206 // Make sure the parent block allows inserting non-PHI instructions
4207 // before the terminator.
4208 BasicBlock *Parent =
4209 BaseI ? BaseI->getParent() : &GEP->getFunction()->getEntryBlock();
4210 if (!Parent->getTerminator()->isEHPad())
4211 LargeOffsetGEP = std::make_pair(GEP, ConstantOffset);
4212 }
4213 }
4214 AddrMode.BaseOffs -= ConstantOffset;
4215 return false;
4216 }
4217
4218 // Save the valid addressing mode in case we can't match.
4219 ExtAddrMode BackupAddrMode = AddrMode;
4220 unsigned OldSize = AddrModeInsts.size();
4221
4222 // See if the scale and offset amount is valid for this target.
4223 AddrMode.BaseOffs += ConstantOffset;
4224 if (!cast<GEPOperator>(AddrInst)->isInBounds())
4225 AddrMode.InBounds = false;
4226
4227 // Match the base operand of the GEP.
4228 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
4229 // If it couldn't be matched, just stuff the value in a register.
4230 if (AddrMode.HasBaseReg) {
4231 AddrMode = BackupAddrMode;
4232 AddrModeInsts.resize(OldSize);
4233 return false;
4234 }
4235 AddrMode.HasBaseReg = true;
4236 AddrMode.BaseReg = AddrInst->getOperand(0);
4237 }
4238
4239 // Match the remaining variable portion of the GEP.
4240 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
4241 Depth)) {
4242 // If it couldn't be matched, try stuffing the base into a register
4243 // instead of matching it, and retrying the match of the scale.
4244 AddrMode = BackupAddrMode;
4245 AddrModeInsts.resize(OldSize);
4246 if (AddrMode.HasBaseReg)
4247 return false;
4248 AddrMode.HasBaseReg = true;
4249 AddrMode.BaseReg = AddrInst->getOperand(0);
4250 AddrMode.BaseOffs += ConstantOffset;
4251 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
4252 VariableScale, Depth)) {
4253 // If even that didn't work, bail.
4254 AddrMode = BackupAddrMode;
4255 AddrModeInsts.resize(OldSize);
4256 return false;
4257 }
4258 }
4259
4260 return true;
4261 }
4262 case Instruction::SExt:
4263 case Instruction::ZExt: {
4264 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
4265 if (!Ext)
4266 return false;
4267
4268 // Try to move this ext out of the way of the addressing mode.
4269 // Ask for a method for doing so.
4270 TypePromotionHelper::Action TPH =
4271 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
4272 if (!TPH)
4273 return false;
4274
4275 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4276 TPT.getRestorationPoint();
4277 unsigned CreatedInstsCost = 0;
4278 unsigned ExtCost = !TLI.isExtFree(Ext);
4279 Value *PromotedOperand =
4280 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
4281 // SExt has been moved away.
4282 // Thus either it will be rematched later in the recursive calls or it is
4283 // gone. Anyway, we must not fold it into the addressing mode at this point.
4284 // E.g.,
4285 // op = add opnd, 1
4286 // idx = ext op
4287 // addr = gep base, idx
4288 // is now:
4289 // promotedOpnd = ext opnd <- no match here
4290 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
4291 // addr = gep base, op <- match
4292 if (MovedAway)
4293 *MovedAway = true;
4294
4295 assert(PromotedOperand &&
4296 "TypePromotionHelper should have filtered out those cases");
4297
4298 ExtAddrMode BackupAddrMode = AddrMode;
4299 unsigned OldSize = AddrModeInsts.size();
4300
4301 if (!matchAddr(PromotedOperand, Depth) ||
4302 // The total of the new cost is equal to the cost of the created
4303 // instructions.
4304 // The total of the old cost is equal to the cost of the extension plus
4305 // what we have saved in the addressing mode.
4306 !isPromotionProfitable(CreatedInstsCost,
4307 ExtCost + (AddrModeInsts.size() - OldSize),
4308 PromotedOperand)) {
4309 AddrMode = BackupAddrMode;
4310 AddrModeInsts.resize(OldSize);
4311 LLVM_DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
4312 TPT.rollback(LastKnownGood);
4313 return false;
4314 }
4315 return true;
4316 }
4317 }
4318 return false;
4319 }
4320
4321 /// If we can, try to add the value of 'Addr' into the current addressing mode.
4322 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
4323 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
4324 /// for the target.
4325 ///
matchAddr(Value * Addr,unsigned Depth)4326 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
4327 // Start a transaction at this point that we will rollback if the matching
4328 // fails.
4329 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4330 TPT.getRestorationPoint();
4331 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
4332 // Fold in immediates if legal for the target.
4333 AddrMode.BaseOffs += CI->getSExtValue();
4334 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4335 return true;
4336 AddrMode.BaseOffs -= CI->getSExtValue();
4337 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
4338 // If this is a global variable, try to fold it into the addressing mode.
4339 if (!AddrMode.BaseGV) {
4340 AddrMode.BaseGV = GV;
4341 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4342 return true;
4343 AddrMode.BaseGV = nullptr;
4344 }
4345 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
4346 ExtAddrMode BackupAddrMode = AddrMode;
4347 unsigned OldSize = AddrModeInsts.size();
4348
4349 // Check to see if it is possible to fold this operation.
4350 bool MovedAway = false;
4351 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
4352 // This instruction may have been moved away. If so, there is nothing
4353 // to check here.
4354 if (MovedAway)
4355 return true;
4356 // Okay, it's possible to fold this. Check to see if it is actually
4357 // *profitable* to do so. We use a simple cost model to avoid increasing
4358 // register pressure too much.
4359 if (I->hasOneUse() ||
4360 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
4361 AddrModeInsts.push_back(I);
4362 return true;
4363 }
4364
4365 // It isn't profitable to do this, roll back.
4366 //cerr << "NOT FOLDING: " << *I;
4367 AddrMode = BackupAddrMode;
4368 AddrModeInsts.resize(OldSize);
4369 TPT.rollback(LastKnownGood);
4370 }
4371 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
4372 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
4373 return true;
4374 TPT.rollback(LastKnownGood);
4375 } else if (isa<ConstantPointerNull>(Addr)) {
4376 // Null pointer gets folded without affecting the addressing mode.
4377 return true;
4378 }
4379
4380 // Worse case, the target should support [reg] addressing modes. :)
4381 if (!AddrMode.HasBaseReg) {
4382 AddrMode.HasBaseReg = true;
4383 AddrMode.BaseReg = Addr;
4384 // Still check for legality in case the target supports [imm] but not [i+r].
4385 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4386 return true;
4387 AddrMode.HasBaseReg = false;
4388 AddrMode.BaseReg = nullptr;
4389 }
4390
4391 // If the base register is already taken, see if we can do [r+r].
4392 if (AddrMode.Scale == 0) {
4393 AddrMode.Scale = 1;
4394 AddrMode.ScaledReg = Addr;
4395 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4396 return true;
4397 AddrMode.Scale = 0;
4398 AddrMode.ScaledReg = nullptr;
4399 }
4400 // Couldn't match.
4401 TPT.rollback(LastKnownGood);
4402 return false;
4403 }
4404
4405 /// Check to see if all uses of OpVal by the specified inline asm call are due
4406 /// to memory operands. If so, return true, otherwise return false.
IsOperandAMemoryOperand(CallInst * CI,InlineAsm * IA,Value * OpVal,const TargetLowering & TLI,const TargetRegisterInfo & TRI)4407 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
4408 const TargetLowering &TLI,
4409 const TargetRegisterInfo &TRI) {
4410 const Function *F = CI->getFunction();
4411 TargetLowering::AsmOperandInfoVector TargetConstraints =
4412 TLI.ParseConstraints(F->getParent()->getDataLayout(), &TRI,
4413 ImmutableCallSite(CI));
4414
4415 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4416 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4417
4418 // Compute the constraint code and ConstraintType to use.
4419 TLI.ComputeConstraintToUse(OpInfo, SDValue());
4420
4421 // If this asm operand is our Value*, and if it isn't an indirect memory
4422 // operand, we can't fold it!
4423 if (OpInfo.CallOperandVal == OpVal &&
4424 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
4425 !OpInfo.isIndirect))
4426 return false;
4427 }
4428
4429 return true;
4430 }
4431
4432 // Max number of memory uses to look at before aborting the search to conserve
4433 // compile time.
4434 static constexpr int MaxMemoryUsesToScan = 20;
4435
4436 /// Recursively walk all the uses of I until we find a memory use.
4437 /// If we find an obviously non-foldable instruction, return true.
4438 /// Add the ultimately found memory instructions to MemoryUses.
FindAllMemoryUses(Instruction * I,SmallVectorImpl<std::pair<Instruction *,unsigned>> & MemoryUses,SmallPtrSetImpl<Instruction * > & ConsideredInsts,const TargetLowering & TLI,const TargetRegisterInfo & TRI,int SeenInsts=0)4439 static bool FindAllMemoryUses(
4440 Instruction *I,
4441 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
4442 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
4443 const TargetRegisterInfo &TRI, int SeenInsts = 0) {
4444 // If we already considered this instruction, we're done.
4445 if (!ConsideredInsts.insert(I).second)
4446 return false;
4447
4448 // If this is an obviously unfoldable instruction, bail out.
4449 if (!MightBeFoldableInst(I))
4450 return true;
4451
4452 const bool OptSize = I->getFunction()->hasOptSize();
4453
4454 // Loop over all the uses, recursively processing them.
4455 for (Use &U : I->uses()) {
4456 // Conservatively return true if we're seeing a large number or a deep chain
4457 // of users. This avoids excessive compilation times in pathological cases.
4458 if (SeenInsts++ >= MaxMemoryUsesToScan)
4459 return true;
4460
4461 Instruction *UserI = cast<Instruction>(U.getUser());
4462 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4463 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
4464 continue;
4465 }
4466
4467 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4468 unsigned opNo = U.getOperandNo();
4469 if (opNo != StoreInst::getPointerOperandIndex())
4470 return true; // Storing addr, not into addr.
4471 MemoryUses.push_back(std::make_pair(SI, opNo));
4472 continue;
4473 }
4474
4475 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
4476 unsigned opNo = U.getOperandNo();
4477 if (opNo != AtomicRMWInst::getPointerOperandIndex())
4478 return true; // Storing addr, not into addr.
4479 MemoryUses.push_back(std::make_pair(RMW, opNo));
4480 continue;
4481 }
4482
4483 if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
4484 unsigned opNo = U.getOperandNo();
4485 if (opNo != AtomicCmpXchgInst::getPointerOperandIndex())
4486 return true; // Storing addr, not into addr.
4487 MemoryUses.push_back(std::make_pair(CmpX, opNo));
4488 continue;
4489 }
4490
4491 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4492 // If this is a cold call, we can sink the addressing calculation into
4493 // the cold path. See optimizeCallInst
4494 if (!OptSize && CI->hasFnAttr(Attribute::Cold))
4495 continue;
4496
4497 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
4498 if (!IA) return true;
4499
4500 // If this is a memory operand, we're cool, otherwise bail out.
4501 if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI))
4502 return true;
4503 continue;
4504 }
4505
4506 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI,
4507 SeenInsts))
4508 return true;
4509 }
4510
4511 return false;
4512 }
4513
4514 /// Return true if Val is already known to be live at the use site that we're
4515 /// folding it into. If so, there is no cost to include it in the addressing
4516 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4517 /// instruction already.
valueAlreadyLiveAtInst(Value * Val,Value * KnownLive1,Value * KnownLive2)4518 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4519 Value *KnownLive2) {
4520 // If Val is either of the known-live values, we know it is live!
4521 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4522 return true;
4523
4524 // All values other than instructions and arguments (e.g. constants) are live.
4525 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4526
4527 // If Val is a constant sized alloca in the entry block, it is live, this is
4528 // true because it is just a reference to the stack/frame pointer, which is
4529 // live for the whole function.
4530 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4531 if (AI->isStaticAlloca())
4532 return true;
4533
4534 // Check to see if this value is already used in the memory instruction's
4535 // block. If so, it's already live into the block at the very least, so we
4536 // can reasonably fold it.
4537 return Val->isUsedInBasicBlock(MemoryInst->getParent());
4538 }
4539
4540 /// It is possible for the addressing mode of the machine to fold the specified
4541 /// instruction into a load or store that ultimately uses it.
4542 /// However, the specified instruction has multiple uses.
4543 /// Given this, it may actually increase register pressure to fold it
4544 /// into the load. For example, consider this code:
4545 ///
4546 /// X = ...
4547 /// Y = X+1
4548 /// use(Y) -> nonload/store
4549 /// Z = Y+1
4550 /// load Z
4551 ///
4552 /// In this case, Y has multiple uses, and can be folded into the load of Z
4553 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4554 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4555 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4556 /// number of computations either.
4557 ///
4558 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4559 /// X was live across 'load Z' for other reasons, we actually *would* want to
4560 /// fold the addressing mode in the Z case. This would make Y die earlier.
4561 bool AddressingModeMatcher::
isProfitableToFoldIntoAddressingMode(Instruction * I,ExtAddrMode & AMBefore,ExtAddrMode & AMAfter)4562 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4563 ExtAddrMode &AMAfter) {
4564 if (IgnoreProfitability) return true;
4565
4566 // AMBefore is the addressing mode before this instruction was folded into it,
4567 // and AMAfter is the addressing mode after the instruction was folded. Get
4568 // the set of registers referenced by AMAfter and subtract out those
4569 // referenced by AMBefore: this is the set of values which folding in this
4570 // address extends the lifetime of.
4571 //
4572 // Note that there are only two potential values being referenced here,
4573 // BaseReg and ScaleReg (global addresses are always available, as are any
4574 // folded immediates).
4575 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4576
4577 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4578 // lifetime wasn't extended by adding this instruction.
4579 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4580 BaseReg = nullptr;
4581 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4582 ScaledReg = nullptr;
4583
4584 // If folding this instruction (and it's subexprs) didn't extend any live
4585 // ranges, we're ok with it.
4586 if (!BaseReg && !ScaledReg)
4587 return true;
4588
4589 // If all uses of this instruction can have the address mode sunk into them,
4590 // we can remove the addressing mode and effectively trade one live register
4591 // for another (at worst.) In this context, folding an addressing mode into
4592 // the use is just a particularly nice way of sinking it.
4593 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
4594 SmallPtrSet<Instruction*, 16> ConsideredInsts;
4595 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI))
4596 return false; // Has a non-memory, non-foldable use!
4597
4598 // Now that we know that all uses of this instruction are part of a chain of
4599 // computation involving only operations that could theoretically be folded
4600 // into a memory use, loop over each of these memory operation uses and see
4601 // if they could *actually* fold the instruction. The assumption is that
4602 // addressing modes are cheap and that duplicating the computation involved
4603 // many times is worthwhile, even on a fastpath. For sinking candidates
4604 // (i.e. cold call sites), this serves as a way to prevent excessive code
4605 // growth since most architectures have some reasonable small and fast way to
4606 // compute an effective address. (i.e LEA on x86)
4607 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4608 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4609 Instruction *User = MemoryUses[i].first;
4610 unsigned OpNo = MemoryUses[i].second;
4611
4612 // Get the access type of this use. If the use isn't a pointer, we don't
4613 // know what it accesses.
4614 Value *Address = User->getOperand(OpNo);
4615 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4616 if (!AddrTy)
4617 return false;
4618 Type *AddressAccessTy = AddrTy->getElementType();
4619 unsigned AS = AddrTy->getAddressSpace();
4620
4621 // Do a match against the root of this address, ignoring profitability. This
4622 // will tell us if the addressing mode for the memory operation will
4623 // *actually* cover the shared instruction.
4624 ExtAddrMode Result;
4625 std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
4626 0);
4627 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4628 TPT.getRestorationPoint();
4629 AddressingModeMatcher Matcher(
4630 MatchedAddrModeInsts, TLI, TRI, AddressAccessTy, AS, MemoryInst, Result,
4631 InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP);
4632 Matcher.IgnoreProfitability = true;
4633 bool Success = Matcher.matchAddr(Address, 0);
4634 (void)Success; assert(Success && "Couldn't select *anything*?");
4635
4636 // The match was to check the profitability, the changes made are not
4637 // part of the original matcher. Therefore, they should be dropped
4638 // otherwise the original matcher will not present the right state.
4639 TPT.rollback(LastKnownGood);
4640
4641 // If the match didn't cover I, then it won't be shared by it.
4642 if (!is_contained(MatchedAddrModeInsts, I))
4643 return false;
4644
4645 MatchedAddrModeInsts.clear();
4646 }
4647
4648 return true;
4649 }
4650
4651 /// Return true if the specified values are defined in a
4652 /// different basic block than BB.
IsNonLocalValue(Value * V,BasicBlock * BB)4653 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4654 if (Instruction *I = dyn_cast<Instruction>(V))
4655 return I->getParent() != BB;
4656 return false;
4657 }
4658
4659 /// Sink addressing mode computation immediate before MemoryInst if doing so
4660 /// can be done without increasing register pressure. The need for the
4661 /// register pressure constraint means this can end up being an all or nothing
4662 /// decision for all uses of the same addressing computation.
4663 ///
4664 /// Load and Store Instructions often have addressing modes that can do
4665 /// significant amounts of computation. As such, instruction selection will try
4666 /// to get the load or store to do as much computation as possible for the
4667 /// program. The problem is that isel can only see within a single block. As
4668 /// such, we sink as much legal addressing mode work into the block as possible.
4669 ///
4670 /// This method is used to optimize both load/store and inline asms with memory
4671 /// operands. It's also used to sink addressing computations feeding into cold
4672 /// call sites into their (cold) basic block.
4673 ///
4674 /// The motivation for handling sinking into cold blocks is that doing so can
4675 /// both enable other address mode sinking (by satisfying the register pressure
4676 /// constraint above), and reduce register pressure globally (by removing the
4677 /// addressing mode computation from the fast path entirely.).
optimizeMemoryInst(Instruction * MemoryInst,Value * Addr,Type * AccessTy,unsigned AddrSpace)4678 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4679 Type *AccessTy, unsigned AddrSpace) {
4680 Value *Repl = Addr;
4681
4682 // Try to collapse single-value PHI nodes. This is necessary to undo
4683 // unprofitable PRE transformations.
4684 SmallVector<Value*, 8> worklist;
4685 SmallPtrSet<Value*, 16> Visited;
4686 worklist.push_back(Addr);
4687
4688 // Use a worklist to iteratively look through PHI and select nodes, and
4689 // ensure that the addressing mode obtained from the non-PHI/select roots of
4690 // the graph are compatible.
4691 bool PhiOrSelectSeen = false;
4692 SmallVector<Instruction*, 16> AddrModeInsts;
4693 const SimplifyQuery SQ(*DL, TLInfo);
4694 AddressingModeCombiner AddrModes(SQ, Addr);
4695 TypePromotionTransaction TPT(RemovedInsts);
4696 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4697 TPT.getRestorationPoint();
4698 while (!worklist.empty()) {
4699 Value *V = worklist.back();
4700 worklist.pop_back();
4701
4702 // We allow traversing cyclic Phi nodes.
4703 // In case of success after this loop we ensure that traversing through
4704 // Phi nodes ends up with all cases to compute address of the form
4705 // BaseGV + Base + Scale * Index + Offset
4706 // where Scale and Offset are constans and BaseGV, Base and Index
4707 // are exactly the same Values in all cases.
4708 // It means that BaseGV, Scale and Offset dominate our memory instruction
4709 // and have the same value as they had in address computation represented
4710 // as Phi. So we can safely sink address computation to memory instruction.
4711 if (!Visited.insert(V).second)
4712 continue;
4713
4714 // For a PHI node, push all of its incoming values.
4715 if (PHINode *P = dyn_cast<PHINode>(V)) {
4716 for (Value *IncValue : P->incoming_values())
4717 worklist.push_back(IncValue);
4718 PhiOrSelectSeen = true;
4719 continue;
4720 }
4721 // Similar for select.
4722 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
4723 worklist.push_back(SI->getFalseValue());
4724 worklist.push_back(SI->getTrueValue());
4725 PhiOrSelectSeen = true;
4726 continue;
4727 }
4728
4729 // For non-PHIs, determine the addressing mode being computed. Note that
4730 // the result may differ depending on what other uses our candidate
4731 // addressing instructions might have.
4732 AddrModeInsts.clear();
4733 std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
4734 0);
4735 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4736 V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *TRI,
4737 InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP);
4738
4739 GetElementPtrInst *GEP = LargeOffsetGEP.first;
4740 if (GEP && !NewGEPBases.count(GEP)) {
4741 // If splitting the underlying data structure can reduce the offset of a
4742 // GEP, collect the GEP. Skip the GEPs that are the new bases of
4743 // previously split data structures.
4744 LargeOffsetGEPMap[GEP->getPointerOperand()].push_back(LargeOffsetGEP);
4745 if (LargeOffsetGEPID.find(GEP) == LargeOffsetGEPID.end())
4746 LargeOffsetGEPID[GEP] = LargeOffsetGEPID.size();
4747 }
4748
4749 NewAddrMode.OriginalValue = V;
4750 if (!AddrModes.addNewAddrMode(NewAddrMode))
4751 break;
4752 }
4753
4754 // Try to combine the AddrModes we've collected. If we couldn't collect any,
4755 // or we have multiple but either couldn't combine them or combining them
4756 // wouldn't do anything useful, bail out now.
4757 if (!AddrModes.combineAddrModes()) {
4758 TPT.rollback(LastKnownGood);
4759 return false;
4760 }
4761 TPT.commit();
4762
4763 // Get the combined AddrMode (or the only AddrMode, if we only had one).
4764 ExtAddrMode AddrMode = AddrModes.getAddrMode();
4765
4766 // If all the instructions matched are already in this BB, don't do anything.
4767 // If we saw a Phi node then it is not local definitely, and if we saw a select
4768 // then we want to push the address calculation past it even if it's already
4769 // in this BB.
4770 if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) {
4771 return IsNonLocalValue(V, MemoryInst->getParent());
4772 })) {
4773 LLVM_DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode
4774 << "\n");
4775 return false;
4776 }
4777
4778 // Insert this computation right after this user. Since our caller is
4779 // scanning from the top of the BB to the bottom, reuse of the expr are
4780 // guaranteed to happen later.
4781 IRBuilder<> Builder(MemoryInst);
4782
4783 // Now that we determined the addressing expression we want to use and know
4784 // that we have to sink it into this block. Check to see if we have already
4785 // done this for some other load/store instr in this block. If so, reuse
4786 // the computation. Before attempting reuse, check if the address is valid
4787 // as it may have been erased.
4788
4789 WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
4790
4791 Value * SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
4792 if (SunkAddr) {
4793 LLVM_DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode
4794 << " for " << *MemoryInst << "\n");
4795 if (SunkAddr->getType() != Addr->getType())
4796 SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4797 } else if (AddrSinkUsingGEPs ||
4798 (!AddrSinkUsingGEPs.getNumOccurrences() && TM && TTI->useAA())) {
4799 // By default, we use the GEP-based method when AA is used later. This
4800 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4801 LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
4802 << " for " << *MemoryInst << "\n");
4803 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4804 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4805
4806 // First, find the pointer.
4807 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4808 ResultPtr = AddrMode.BaseReg;
4809 AddrMode.BaseReg = nullptr;
4810 }
4811
4812 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4813 // We can't add more than one pointer together, nor can we scale a
4814 // pointer (both of which seem meaningless).
4815 if (ResultPtr || AddrMode.Scale != 1)
4816 return false;
4817
4818 ResultPtr = AddrMode.ScaledReg;
4819 AddrMode.Scale = 0;
4820 }
4821
4822 // It is only safe to sign extend the BaseReg if we know that the math
4823 // required to create it did not overflow before we extend it. Since
4824 // the original IR value was tossed in favor of a constant back when
4825 // the AddrMode was created we need to bail out gracefully if widths
4826 // do not match instead of extending it.
4827 //
4828 // (See below for code to add the scale.)
4829 if (AddrMode.Scale) {
4830 Type *ScaledRegTy = AddrMode.ScaledReg->getType();
4831 if (cast<IntegerType>(IntPtrTy)->getBitWidth() >
4832 cast<IntegerType>(ScaledRegTy)->getBitWidth())
4833 return false;
4834 }
4835
4836 if (AddrMode.BaseGV) {
4837 if (ResultPtr)
4838 return false;
4839
4840 ResultPtr = AddrMode.BaseGV;
4841 }
4842
4843 // If the real base value actually came from an inttoptr, then the matcher
4844 // will look through it and provide only the integer value. In that case,
4845 // use it here.
4846 if (!DL->isNonIntegralPointerType(Addr->getType())) {
4847 if (!ResultPtr && AddrMode.BaseReg) {
4848 ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(),
4849 "sunkaddr");
4850 AddrMode.BaseReg = nullptr;
4851 } else if (!ResultPtr && AddrMode.Scale == 1) {
4852 ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(),
4853 "sunkaddr");
4854 AddrMode.Scale = 0;
4855 }
4856 }
4857
4858 if (!ResultPtr &&
4859 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4860 SunkAddr = Constant::getNullValue(Addr->getType());
4861 } else if (!ResultPtr) {
4862 return false;
4863 } else {
4864 Type *I8PtrTy =
4865 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4866 Type *I8Ty = Builder.getInt8Ty();
4867
4868 // Start with the base register. Do this first so that subsequent address
4869 // matching finds it last, which will prevent it from trying to match it
4870 // as the scaled value in case it happens to be a mul. That would be
4871 // problematic if we've sunk a different mul for the scale, because then
4872 // we'd end up sinking both muls.
4873 if (AddrMode.BaseReg) {
4874 Value *V = AddrMode.BaseReg;
4875 if (V->getType() != IntPtrTy)
4876 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4877
4878 ResultIndex = V;
4879 }
4880
4881 // Add the scale value.
4882 if (AddrMode.Scale) {
4883 Value *V = AddrMode.ScaledReg;
4884 if (V->getType() == IntPtrTy) {
4885 // done.
4886 } else {
4887 assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <
4888 cast<IntegerType>(V->getType())->getBitWidth() &&
4889 "We can't transform if ScaledReg is too narrow");
4890 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4891 }
4892
4893 if (AddrMode.Scale != 1)
4894 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4895 "sunkaddr");
4896 if (ResultIndex)
4897 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4898 else
4899 ResultIndex = V;
4900 }
4901
4902 // Add in the Base Offset if present.
4903 if (AddrMode.BaseOffs) {
4904 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4905 if (ResultIndex) {
4906 // We need to add this separately from the scale above to help with
4907 // SDAG consecutive load/store merging.
4908 if (ResultPtr->getType() != I8PtrTy)
4909 ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4910 ResultPtr =
4911 AddrMode.InBounds
4912 ? Builder.CreateInBoundsGEP(I8Ty, ResultPtr, ResultIndex,
4913 "sunkaddr")
4914 : Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4915 }
4916
4917 ResultIndex = V;
4918 }
4919
4920 if (!ResultIndex) {
4921 SunkAddr = ResultPtr;
4922 } else {
4923 if (ResultPtr->getType() != I8PtrTy)
4924 ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4925 SunkAddr =
4926 AddrMode.InBounds
4927 ? Builder.CreateInBoundsGEP(I8Ty, ResultPtr, ResultIndex,
4928 "sunkaddr")
4929 : Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4930 }
4931
4932 if (SunkAddr->getType() != Addr->getType())
4933 SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4934 }
4935 } else {
4936 // We'd require a ptrtoint/inttoptr down the line, which we can't do for
4937 // non-integral pointers, so in that case bail out now.
4938 Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
4939 Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
4940 PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy);
4941 PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy);
4942 if (DL->isNonIntegralPointerType(Addr->getType()) ||
4943 (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) ||
4944 (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) ||
4945 (AddrMode.BaseGV &&
4946 DL->isNonIntegralPointerType(AddrMode.BaseGV->getType())))
4947 return false;
4948
4949 LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
4950 << " for " << *MemoryInst << "\n");
4951 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4952 Value *Result = nullptr;
4953
4954 // Start with the base register. Do this first so that subsequent address
4955 // matching finds it last, which will prevent it from trying to match it
4956 // as the scaled value in case it happens to be a mul. That would be
4957 // problematic if we've sunk a different mul for the scale, because then
4958 // we'd end up sinking both muls.
4959 if (AddrMode.BaseReg) {
4960 Value *V = AddrMode.BaseReg;
4961 if (V->getType()->isPointerTy())
4962 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4963 if (V->getType() != IntPtrTy)
4964 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4965 Result = V;
4966 }
4967
4968 // Add the scale value.
4969 if (AddrMode.Scale) {
4970 Value *V = AddrMode.ScaledReg;
4971 if (V->getType() == IntPtrTy) {
4972 // done.
4973 } else if (V->getType()->isPointerTy()) {
4974 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4975 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4976 cast<IntegerType>(V->getType())->getBitWidth()) {
4977 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4978 } else {
4979 // It is only safe to sign extend the BaseReg if we know that the math
4980 // required to create it did not overflow before we extend it. Since
4981 // the original IR value was tossed in favor of a constant back when
4982 // the AddrMode was created we need to bail out gracefully if widths
4983 // do not match instead of extending it.
4984 Instruction *I = dyn_cast_or_null<Instruction>(Result);
4985 if (I && (Result != AddrMode.BaseReg))
4986 I->eraseFromParent();
4987 return false;
4988 }
4989 if (AddrMode.Scale != 1)
4990 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4991 "sunkaddr");
4992 if (Result)
4993 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4994 else
4995 Result = V;
4996 }
4997
4998 // Add in the BaseGV if present.
4999 if (AddrMode.BaseGV) {
5000 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
5001 if (Result)
5002 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5003 else
5004 Result = V;
5005 }
5006
5007 // Add in the Base Offset if present.
5008 if (AddrMode.BaseOffs) {
5009 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
5010 if (Result)
5011 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5012 else
5013 Result = V;
5014 }
5015
5016 if (!Result)
5017 SunkAddr = Constant::getNullValue(Addr->getType());
5018 else
5019 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
5020 }
5021
5022 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
5023 // Store the newly computed address into the cache. In the case we reused a
5024 // value, this should be idempotent.
5025 SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
5026
5027 // If we have no uses, recursively delete the value and all dead instructions
5028 // using it.
5029 if (Repl->use_empty()) {
5030 // This can cause recursive deletion, which can invalidate our iterator.
5031 // Use a WeakTrackingVH to hold onto it in case this happens.
5032 Value *CurValue = &*CurInstIterator;
5033 WeakTrackingVH IterHandle(CurValue);
5034 BasicBlock *BB = CurInstIterator->getParent();
5035
5036 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
5037
5038 if (IterHandle != CurValue) {
5039 // If the iterator instruction was recursively deleted, start over at the
5040 // start of the block.
5041 CurInstIterator = BB->begin();
5042 SunkAddrs.clear();
5043 }
5044 }
5045 ++NumMemoryInsts;
5046 return true;
5047 }
5048
5049 /// If there are any memory operands, use OptimizeMemoryInst to sink their
5050 /// address computing into the block when possible / profitable.
optimizeInlineAsmInst(CallInst * CS)5051 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
5052 bool MadeChange = false;
5053
5054 const TargetRegisterInfo *TRI =
5055 TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
5056 TargetLowering::AsmOperandInfoVector TargetConstraints =
5057 TLI->ParseConstraints(*DL, TRI, CS);
5058 unsigned ArgNo = 0;
5059 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5060 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
5061
5062 // Compute the constraint code and ConstraintType to use.
5063 TLI->ComputeConstraintToUse(OpInfo, SDValue());
5064
5065 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5066 OpInfo.isIndirect) {
5067 Value *OpVal = CS->getArgOperand(ArgNo++);
5068 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
5069 } else if (OpInfo.Type == InlineAsm::isInput)
5070 ArgNo++;
5071 }
5072
5073 return MadeChange;
5074 }
5075
5076 /// Check if all the uses of \p Val are equivalent (or free) zero or
5077 /// sign extensions.
hasSameExtUse(Value * Val,const TargetLowering & TLI)5078 static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
5079 assert(!Val->use_empty() && "Input must have at least one use");
5080 const Instruction *FirstUser = cast<Instruction>(*Val->user_begin());
5081 bool IsSExt = isa<SExtInst>(FirstUser);
5082 Type *ExtTy = FirstUser->getType();
5083 for (const User *U : Val->users()) {
5084 const Instruction *UI = cast<Instruction>(U);
5085 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
5086 return false;
5087 Type *CurTy = UI->getType();
5088 // Same input and output types: Same instruction after CSE.
5089 if (CurTy == ExtTy)
5090 continue;
5091
5092 // If IsSExt is true, we are in this situation:
5093 // a = Val
5094 // b = sext ty1 a to ty2
5095 // c = sext ty1 a to ty3
5096 // Assuming ty2 is shorter than ty3, this could be turned into:
5097 // a = Val
5098 // b = sext ty1 a to ty2
5099 // c = sext ty2 b to ty3
5100 // However, the last sext is not free.
5101 if (IsSExt)
5102 return false;
5103
5104 // This is a ZExt, maybe this is free to extend from one type to another.
5105 // In that case, we would not account for a different use.
5106 Type *NarrowTy;
5107 Type *LargeTy;
5108 if (ExtTy->getScalarType()->getIntegerBitWidth() >
5109 CurTy->getScalarType()->getIntegerBitWidth()) {
5110 NarrowTy = CurTy;
5111 LargeTy = ExtTy;
5112 } else {
5113 NarrowTy = ExtTy;
5114 LargeTy = CurTy;
5115 }
5116
5117 if (!TLI.isZExtFree(NarrowTy, LargeTy))
5118 return false;
5119 }
5120 // All uses are the same or can be derived from one another for free.
5121 return true;
5122 }
5123
5124 /// Try to speculatively promote extensions in \p Exts and continue
5125 /// promoting through newly promoted operands recursively as far as doing so is
5126 /// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
5127 /// When some promotion happened, \p TPT contains the proper state to revert
5128 /// them.
5129 ///
5130 /// \return true if some promotion happened, false otherwise.
tryToPromoteExts(TypePromotionTransaction & TPT,const SmallVectorImpl<Instruction * > & Exts,SmallVectorImpl<Instruction * > & ProfitablyMovedExts,unsigned CreatedInstsCost)5131 bool CodeGenPrepare::tryToPromoteExts(
5132 TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
5133 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
5134 unsigned CreatedInstsCost) {
5135 bool Promoted = false;
5136
5137 // Iterate over all the extensions to try to promote them.
5138 for (auto I : Exts) {
5139 // Early check if we directly have ext(load).
5140 if (isa<LoadInst>(I->getOperand(0))) {
5141 ProfitablyMovedExts.push_back(I);
5142 continue;
5143 }
5144
5145 // Check whether or not we want to do any promotion. The reason we have
5146 // this check inside the for loop is to catch the case where an extension
5147 // is directly fed by a load because in such case the extension can be moved
5148 // up without any promotion on its operands.
5149 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
5150 return false;
5151
5152 // Get the action to perform the promotion.
5153 TypePromotionHelper::Action TPH =
5154 TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
5155 // Check if we can promote.
5156 if (!TPH) {
5157 // Save the current extension as we cannot move up through its operand.
5158 ProfitablyMovedExts.push_back(I);
5159 continue;
5160 }
5161
5162 // Save the current state.
5163 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5164 TPT.getRestorationPoint();
5165 SmallVector<Instruction *, 4> NewExts;
5166 unsigned NewCreatedInstsCost = 0;
5167 unsigned ExtCost = !TLI->isExtFree(I);
5168 // Promote.
5169 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
5170 &NewExts, nullptr, *TLI);
5171 assert(PromotedVal &&
5172 "TypePromotionHelper should have filtered out those cases");
5173
5174 // We would be able to merge only one extension in a load.
5175 // Therefore, if we have more than 1 new extension we heuristically
5176 // cut this search path, because it means we degrade the code quality.
5177 // With exactly 2, the transformation is neutral, because we will merge
5178 // one extension but leave one. However, we optimistically keep going,
5179 // because the new extension may be removed too.
5180 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
5181 // FIXME: It would be possible to propagate a negative value instead of
5182 // conservatively ceiling it to 0.
5183 TotalCreatedInstsCost =
5184 std::max((long long)0, (TotalCreatedInstsCost - ExtCost));
5185 if (!StressExtLdPromotion &&
5186 (TotalCreatedInstsCost > 1 ||
5187 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
5188 // This promotion is not profitable, rollback to the previous state, and
5189 // save the current extension in ProfitablyMovedExts as the latest
5190 // speculative promotion turned out to be unprofitable.
5191 TPT.rollback(LastKnownGood);
5192 ProfitablyMovedExts.push_back(I);
5193 continue;
5194 }
5195 // Continue promoting NewExts as far as doing so is profitable.
5196 SmallVector<Instruction *, 2> NewlyMovedExts;
5197 (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost);
5198 bool NewPromoted = false;
5199 for (auto ExtInst : NewlyMovedExts) {
5200 Instruction *MovedExt = cast<Instruction>(ExtInst);
5201 Value *ExtOperand = MovedExt->getOperand(0);
5202 // If we have reached to a load, we need this extra profitability check
5203 // as it could potentially be merged into an ext(load).
5204 if (isa<LoadInst>(ExtOperand) &&
5205 !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
5206 (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI))))
5207 continue;
5208
5209 ProfitablyMovedExts.push_back(MovedExt);
5210 NewPromoted = true;
5211 }
5212
5213 // If none of speculative promotions for NewExts is profitable, rollback
5214 // and save the current extension (I) as the last profitable extension.
5215 if (!NewPromoted) {
5216 TPT.rollback(LastKnownGood);
5217 ProfitablyMovedExts.push_back(I);
5218 continue;
5219 }
5220 // The promotion is profitable.
5221 Promoted = true;
5222 }
5223 return Promoted;
5224 }
5225
5226 /// Merging redundant sexts when one is dominating the other.
mergeSExts(Function & F)5227 bool CodeGenPrepare::mergeSExts(Function &F) {
5228 bool Changed = false;
5229 for (auto &Entry : ValToSExtendedUses) {
5230 SExts &Insts = Entry.second;
5231 SExts CurPts;
5232 for (Instruction *Inst : Insts) {
5233 if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) ||
5234 Inst->getOperand(0) != Entry.first)
5235 continue;
5236 bool inserted = false;
5237 for (auto &Pt : CurPts) {
5238 if (getDT(F).dominates(Inst, Pt)) {
5239 Pt->replaceAllUsesWith(Inst);
5240 RemovedInsts.insert(Pt);
5241 Pt->removeFromParent();
5242 Pt = Inst;
5243 inserted = true;
5244 Changed = true;
5245 break;
5246 }
5247 if (!getDT(F).dominates(Pt, Inst))
5248 // Give up if we need to merge in a common dominator as the
5249 // experiments show it is not profitable.
5250 continue;
5251 Inst->replaceAllUsesWith(Pt);
5252 RemovedInsts.insert(Inst);
5253 Inst->removeFromParent();
5254 inserted = true;
5255 Changed = true;
5256 break;
5257 }
5258 if (!inserted)
5259 CurPts.push_back(Inst);
5260 }
5261 }
5262 return Changed;
5263 }
5264
5265 // Spliting large data structures so that the GEPs accessing them can have
5266 // smaller offsets so that they can be sunk to the same blocks as their users.
5267 // For example, a large struct starting from %base is splitted into two parts
5268 // where the second part starts from %new_base.
5269 //
5270 // Before:
5271 // BB0:
5272 // %base =
5273 //
5274 // BB1:
5275 // %gep0 = gep %base, off0
5276 // %gep1 = gep %base, off1
5277 // %gep2 = gep %base, off2
5278 //
5279 // BB2:
5280 // %load1 = load %gep0
5281 // %load2 = load %gep1
5282 // %load3 = load %gep2
5283 //
5284 // After:
5285 // BB0:
5286 // %base =
5287 // %new_base = gep %base, off0
5288 //
5289 // BB1:
5290 // %new_gep0 = %new_base
5291 // %new_gep1 = gep %new_base, off1 - off0
5292 // %new_gep2 = gep %new_base, off2 - off0
5293 //
5294 // BB2:
5295 // %load1 = load i32, i32* %new_gep0
5296 // %load2 = load i32, i32* %new_gep1
5297 // %load3 = load i32, i32* %new_gep2
5298 //
5299 // %new_gep1 and %new_gep2 can be sunk to BB2 now after the splitting because
5300 // their offsets are smaller enough to fit into the addressing mode.
splitLargeGEPOffsets()5301 bool CodeGenPrepare::splitLargeGEPOffsets() {
5302 bool Changed = false;
5303 for (auto &Entry : LargeOffsetGEPMap) {
5304 Value *OldBase = Entry.first;
5305 SmallVectorImpl<std::pair<AssertingVH<GetElementPtrInst>, int64_t>>
5306 &LargeOffsetGEPs = Entry.second;
5307 auto compareGEPOffset =
5308 [&](const std::pair<GetElementPtrInst *, int64_t> &LHS,
5309 const std::pair<GetElementPtrInst *, int64_t> &RHS) {
5310 if (LHS.first == RHS.first)
5311 return false;
5312 if (LHS.second != RHS.second)
5313 return LHS.second < RHS.second;
5314 return LargeOffsetGEPID[LHS.first] < LargeOffsetGEPID[RHS.first];
5315 };
5316 // Sorting all the GEPs of the same data structures based on the offsets.
5317 llvm::sort(LargeOffsetGEPs, compareGEPOffset);
5318 LargeOffsetGEPs.erase(
5319 std::unique(LargeOffsetGEPs.begin(), LargeOffsetGEPs.end()),
5320 LargeOffsetGEPs.end());
5321 // Skip if all the GEPs have the same offsets.
5322 if (LargeOffsetGEPs.front().second == LargeOffsetGEPs.back().second)
5323 continue;
5324 GetElementPtrInst *BaseGEP = LargeOffsetGEPs.begin()->first;
5325 int64_t BaseOffset = LargeOffsetGEPs.begin()->second;
5326 Value *NewBaseGEP = nullptr;
5327
5328 auto LargeOffsetGEP = LargeOffsetGEPs.begin();
5329 while (LargeOffsetGEP != LargeOffsetGEPs.end()) {
5330 GetElementPtrInst *GEP = LargeOffsetGEP->first;
5331 int64_t Offset = LargeOffsetGEP->second;
5332 if (Offset != BaseOffset) {
5333 TargetLowering::AddrMode AddrMode;
5334 AddrMode.BaseOffs = Offset - BaseOffset;
5335 // The result type of the GEP might not be the type of the memory
5336 // access.
5337 if (!TLI->isLegalAddressingMode(*DL, AddrMode,
5338 GEP->getResultElementType(),
5339 GEP->getAddressSpace())) {
5340 // We need to create a new base if the offset to the current base is
5341 // too large to fit into the addressing mode. So, a very large struct
5342 // may be splitted into several parts.
5343 BaseGEP = GEP;
5344 BaseOffset = Offset;
5345 NewBaseGEP = nullptr;
5346 }
5347 }
5348
5349 // Generate a new GEP to replace the current one.
5350 LLVMContext &Ctx = GEP->getContext();
5351 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
5352 Type *I8PtrTy =
5353 Type::getInt8PtrTy(Ctx, GEP->getType()->getPointerAddressSpace());
5354 Type *I8Ty = Type::getInt8Ty(Ctx);
5355
5356 if (!NewBaseGEP) {
5357 // Create a new base if we don't have one yet. Find the insertion
5358 // pointer for the new base first.
5359 BasicBlock::iterator NewBaseInsertPt;
5360 BasicBlock *NewBaseInsertBB;
5361 if (auto *BaseI = dyn_cast<Instruction>(OldBase)) {
5362 // If the base of the struct is an instruction, the new base will be
5363 // inserted close to it.
5364 NewBaseInsertBB = BaseI->getParent();
5365 if (isa<PHINode>(BaseI))
5366 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5367 else if (InvokeInst *Invoke = dyn_cast<InvokeInst>(BaseI)) {
5368 NewBaseInsertBB =
5369 SplitEdge(NewBaseInsertBB, Invoke->getNormalDest());
5370 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5371 } else
5372 NewBaseInsertPt = std::next(BaseI->getIterator());
5373 } else {
5374 // If the current base is an argument or global value, the new base
5375 // will be inserted to the entry block.
5376 NewBaseInsertBB = &BaseGEP->getFunction()->getEntryBlock();
5377 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5378 }
5379 IRBuilder<> NewBaseBuilder(NewBaseInsertBB, NewBaseInsertPt);
5380 // Create a new base.
5381 Value *BaseIndex = ConstantInt::get(IntPtrTy, BaseOffset);
5382 NewBaseGEP = OldBase;
5383 if (NewBaseGEP->getType() != I8PtrTy)
5384 NewBaseGEP = NewBaseBuilder.CreatePointerCast(NewBaseGEP, I8PtrTy);
5385 NewBaseGEP =
5386 NewBaseBuilder.CreateGEP(I8Ty, NewBaseGEP, BaseIndex, "splitgep");
5387 NewGEPBases.insert(NewBaseGEP);
5388 }
5389
5390 IRBuilder<> Builder(GEP);
5391 Value *NewGEP = NewBaseGEP;
5392 if (Offset == BaseOffset) {
5393 if (GEP->getType() != I8PtrTy)
5394 NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
5395 } else {
5396 // Calculate the new offset for the new GEP.
5397 Value *Index = ConstantInt::get(IntPtrTy, Offset - BaseOffset);
5398 NewGEP = Builder.CreateGEP(I8Ty, NewBaseGEP, Index);
5399
5400 if (GEP->getType() != I8PtrTy)
5401 NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
5402 }
5403 GEP->replaceAllUsesWith(NewGEP);
5404 LargeOffsetGEPID.erase(GEP);
5405 LargeOffsetGEP = LargeOffsetGEPs.erase(LargeOffsetGEP);
5406 GEP->eraseFromParent();
5407 Changed = true;
5408 }
5409 }
5410 return Changed;
5411 }
5412
5413 /// Return true, if an ext(load) can be formed from an extension in
5414 /// \p MovedExts.
canFormExtLd(const SmallVectorImpl<Instruction * > & MovedExts,LoadInst * & LI,Instruction * & Inst,bool HasPromoted)5415 bool CodeGenPrepare::canFormExtLd(
5416 const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
5417 Instruction *&Inst, bool HasPromoted) {
5418 for (auto *MovedExtInst : MovedExts) {
5419 if (isa<LoadInst>(MovedExtInst->getOperand(0))) {
5420 LI = cast<LoadInst>(MovedExtInst->getOperand(0));
5421 Inst = MovedExtInst;
5422 break;
5423 }
5424 }
5425 if (!LI)
5426 return false;
5427
5428 // If they're already in the same block, there's nothing to do.
5429 // Make the cheap checks first if we did not promote.
5430 // If we promoted, we need to check if it is indeed profitable.
5431 if (!HasPromoted && LI->getParent() == Inst->getParent())
5432 return false;
5433
5434 return TLI->isExtLoad(LI, Inst, *DL);
5435 }
5436
5437 /// Move a zext or sext fed by a load into the same basic block as the load,
5438 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
5439 /// extend into the load.
5440 ///
5441 /// E.g.,
5442 /// \code
5443 /// %ld = load i32* %addr
5444 /// %add = add nuw i32 %ld, 4
5445 /// %zext = zext i32 %add to i64
5446 // \endcode
5447 /// =>
5448 /// \code
5449 /// %ld = load i32* %addr
5450 /// %zext = zext i32 %ld to i64
5451 /// %add = add nuw i64 %zext, 4
5452 /// \encode
5453 /// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
5454 /// allow us to match zext(load i32*) to i64.
5455 ///
5456 /// Also, try to promote the computations used to obtain a sign extended
5457 /// value used into memory accesses.
5458 /// E.g.,
5459 /// \code
5460 /// a = add nsw i32 b, 3
5461 /// d = sext i32 a to i64
5462 /// e = getelementptr ..., i64 d
5463 /// \endcode
5464 /// =>
5465 /// \code
5466 /// f = sext i32 b to i64
5467 /// a = add nsw i64 f, 3
5468 /// e = getelementptr ..., i64 a
5469 /// \endcode
5470 ///
5471 /// \p Inst[in/out] the extension may be modified during the process if some
5472 /// promotions apply.
optimizeExt(Instruction * & Inst)5473 bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
5474 // ExtLoad formation and address type promotion infrastructure requires TLI to
5475 // be effective.
5476 if (!TLI)
5477 return false;
5478
5479 bool AllowPromotionWithoutCommonHeader = false;
5480 /// See if it is an interesting sext operations for the address type
5481 /// promotion before trying to promote it, e.g., the ones with the right
5482 /// type and used in memory accesses.
5483 bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
5484 *Inst, AllowPromotionWithoutCommonHeader);
5485 TypePromotionTransaction TPT(RemovedInsts);
5486 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5487 TPT.getRestorationPoint();
5488 SmallVector<Instruction *, 1> Exts;
5489 SmallVector<Instruction *, 2> SpeculativelyMovedExts;
5490 Exts.push_back(Inst);
5491
5492 bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts);
5493
5494 // Look for a load being extended.
5495 LoadInst *LI = nullptr;
5496 Instruction *ExtFedByLoad;
5497
5498 // Try to promote a chain of computation if it allows to form an extended
5499 // load.
5500 if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) {
5501 assert(LI && ExtFedByLoad && "Expect a valid load and extension");
5502 TPT.commit();
5503 // Move the extend into the same block as the load
5504 ExtFedByLoad->moveAfter(LI);
5505 // CGP does not check if the zext would be speculatively executed when moved
5506 // to the same basic block as the load. Preserving its original location
5507 // would pessimize the debugging experience, as well as negatively impact
5508 // the quality of sample pgo. We don't want to use "line 0" as that has a
5509 // size cost in the line-table section and logically the zext can be seen as
5510 // part of the load. Therefore we conservatively reuse the same debug
5511 // location for the load and the zext.
5512 ExtFedByLoad->setDebugLoc(LI->getDebugLoc());
5513 ++NumExtsMoved;
5514 Inst = ExtFedByLoad;
5515 return true;
5516 }
5517
5518 // Continue promoting SExts if known as considerable depending on targets.
5519 if (ATPConsiderable &&
5520 performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
5521 HasPromoted, TPT, SpeculativelyMovedExts))
5522 return true;
5523
5524 TPT.rollback(LastKnownGood);
5525 return false;
5526 }
5527
5528 // Perform address type promotion if doing so is profitable.
5529 // If AllowPromotionWithoutCommonHeader == false, we should find other sext
5530 // instructions that sign extended the same initial value. However, if
5531 // AllowPromotionWithoutCommonHeader == true, we expect promoting the
5532 // extension is just profitable.
performAddressTypePromotion(Instruction * & Inst,bool AllowPromotionWithoutCommonHeader,bool HasPromoted,TypePromotionTransaction & TPT,SmallVectorImpl<Instruction * > & SpeculativelyMovedExts)5533 bool CodeGenPrepare::performAddressTypePromotion(
5534 Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
5535 bool HasPromoted, TypePromotionTransaction &TPT,
5536 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
5537 bool Promoted = false;
5538 SmallPtrSet<Instruction *, 1> UnhandledExts;
5539 bool AllSeenFirst = true;
5540 for (auto I : SpeculativelyMovedExts) {
5541 Value *HeadOfChain = I->getOperand(0);
5542 DenseMap<Value *, Instruction *>::iterator AlreadySeen =
5543 SeenChainsForSExt.find(HeadOfChain);
5544 // If there is an unhandled SExt which has the same header, try to promote
5545 // it as well.
5546 if (AlreadySeen != SeenChainsForSExt.end()) {
5547 if (AlreadySeen->second != nullptr)
5548 UnhandledExts.insert(AlreadySeen->second);
5549 AllSeenFirst = false;
5550 }
5551 }
5552
5553 if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
5554 SpeculativelyMovedExts.size() == 1)) {
5555 TPT.commit();
5556 if (HasPromoted)
5557 Promoted = true;
5558 for (auto I : SpeculativelyMovedExts) {
5559 Value *HeadOfChain = I->getOperand(0);
5560 SeenChainsForSExt[HeadOfChain] = nullptr;
5561 ValToSExtendedUses[HeadOfChain].push_back(I);
5562 }
5563 // Update Inst as promotion happen.
5564 Inst = SpeculativelyMovedExts.pop_back_val();
5565 } else {
5566 // This is the first chain visited from the header, keep the current chain
5567 // as unhandled. Defer to promote this until we encounter another SExt
5568 // chain derived from the same header.
5569 for (auto I : SpeculativelyMovedExts) {
5570 Value *HeadOfChain = I->getOperand(0);
5571 SeenChainsForSExt[HeadOfChain] = Inst;
5572 }
5573 return false;
5574 }
5575
5576 if (!AllSeenFirst && !UnhandledExts.empty())
5577 for (auto VisitedSExt : UnhandledExts) {
5578 if (RemovedInsts.count(VisitedSExt))
5579 continue;
5580 TypePromotionTransaction TPT(RemovedInsts);
5581 SmallVector<Instruction *, 1> Exts;
5582 SmallVector<Instruction *, 2> Chains;
5583 Exts.push_back(VisitedSExt);
5584 bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains);
5585 TPT.commit();
5586 if (HasPromoted)
5587 Promoted = true;
5588 for (auto I : Chains) {
5589 Value *HeadOfChain = I->getOperand(0);
5590 // Mark this as handled.
5591 SeenChainsForSExt[HeadOfChain] = nullptr;
5592 ValToSExtendedUses[HeadOfChain].push_back(I);
5593 }
5594 }
5595 return Promoted;
5596 }
5597
optimizeExtUses(Instruction * I)5598 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
5599 BasicBlock *DefBB = I->getParent();
5600
5601 // If the result of a {s|z}ext and its source are both live out, rewrite all
5602 // other uses of the source with result of extension.
5603 Value *Src = I->getOperand(0);
5604 if (Src->hasOneUse())
5605 return false;
5606
5607 // Only do this xform if truncating is free.
5608 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
5609 return false;
5610
5611 // Only safe to perform the optimization if the source is also defined in
5612 // this block.
5613 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
5614 return false;
5615
5616 bool DefIsLiveOut = false;
5617 for (User *U : I->users()) {
5618 Instruction *UI = cast<Instruction>(U);
5619
5620 // Figure out which BB this ext is used in.
5621 BasicBlock *UserBB = UI->getParent();
5622 if (UserBB == DefBB) continue;
5623 DefIsLiveOut = true;
5624 break;
5625 }
5626 if (!DefIsLiveOut)
5627 return false;
5628
5629 // Make sure none of the uses are PHI nodes.
5630 for (User *U : Src->users()) {
5631 Instruction *UI = cast<Instruction>(U);
5632 BasicBlock *UserBB = UI->getParent();
5633 if (UserBB == DefBB) continue;
5634 // Be conservative. We don't want this xform to end up introducing
5635 // reloads just before load / store instructions.
5636 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
5637 return false;
5638 }
5639
5640 // InsertedTruncs - Only insert one trunc in each block once.
5641 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
5642
5643 bool MadeChange = false;
5644 for (Use &U : Src->uses()) {
5645 Instruction *User = cast<Instruction>(U.getUser());
5646
5647 // Figure out which BB this ext is used in.
5648 BasicBlock *UserBB = User->getParent();
5649 if (UserBB == DefBB) continue;
5650
5651 // Both src and def are live in this block. Rewrite the use.
5652 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
5653
5654 if (!InsertedTrunc) {
5655 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5656 assert(InsertPt != UserBB->end());
5657 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5658 InsertedInsts.insert(InsertedTrunc);
5659 }
5660
5661 // Replace a use of the {s|z}ext source with a use of the result.
5662 U = InsertedTrunc;
5663 ++NumExtUses;
5664 MadeChange = true;
5665 }
5666
5667 return MadeChange;
5668 }
5669
5670 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
5671 // just after the load if the target can fold this into one extload instruction,
5672 // with the hope of eliminating some of the other later "and" instructions using
5673 // the loaded value. "and"s that are made trivially redundant by the insertion
5674 // of the new "and" are removed by this function, while others (e.g. those whose
5675 // path from the load goes through a phi) are left for isel to potentially
5676 // remove.
5677 //
5678 // For example:
5679 //
5680 // b0:
5681 // x = load i32
5682 // ...
5683 // b1:
5684 // y = and x, 0xff
5685 // z = use y
5686 //
5687 // becomes:
5688 //
5689 // b0:
5690 // x = load i32
5691 // x' = and x, 0xff
5692 // ...
5693 // b1:
5694 // z = use x'
5695 //
5696 // whereas:
5697 //
5698 // b0:
5699 // x1 = load i32
5700 // ...
5701 // b1:
5702 // x2 = load i32
5703 // ...
5704 // b2:
5705 // x = phi x1, x2
5706 // y = and x, 0xff
5707 //
5708 // becomes (after a call to optimizeLoadExt for each load):
5709 //
5710 // b0:
5711 // x1 = load i32
5712 // x1' = and x1, 0xff
5713 // ...
5714 // b1:
5715 // x2 = load i32
5716 // x2' = and x2, 0xff
5717 // ...
5718 // b2:
5719 // x = phi x1', x2'
5720 // y = and x, 0xff
optimizeLoadExt(LoadInst * Load)5721 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5722 if (!Load->isSimple() || !Load->getType()->isIntOrPtrTy())
5723 return false;
5724
5725 // Skip loads we've already transformed.
5726 if (Load->hasOneUse() &&
5727 InsertedInsts.count(cast<Instruction>(*Load->user_begin())))
5728 return false;
5729
5730 // Look at all uses of Load, looking through phis, to determine how many bits
5731 // of the loaded value are needed.
5732 SmallVector<Instruction *, 8> WorkList;
5733 SmallPtrSet<Instruction *, 16> Visited;
5734 SmallVector<Instruction *, 8> AndsToMaybeRemove;
5735 for (auto *U : Load->users())
5736 WorkList.push_back(cast<Instruction>(U));
5737
5738 EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5739 unsigned BitWidth = LoadResultVT.getSizeInBits();
5740 APInt DemandBits(BitWidth, 0);
5741 APInt WidestAndBits(BitWidth, 0);
5742
5743 while (!WorkList.empty()) {
5744 Instruction *I = WorkList.back();
5745 WorkList.pop_back();
5746
5747 // Break use-def graph loops.
5748 if (!Visited.insert(I).second)
5749 continue;
5750
5751 // For a PHI node, push all of its users.
5752 if (auto *Phi = dyn_cast<PHINode>(I)) {
5753 for (auto *U : Phi->users())
5754 WorkList.push_back(cast<Instruction>(U));
5755 continue;
5756 }
5757
5758 switch (I->getOpcode()) {
5759 case Instruction::And: {
5760 auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5761 if (!AndC)
5762 return false;
5763 APInt AndBits = AndC->getValue();
5764 DemandBits |= AndBits;
5765 // Keep track of the widest and mask we see.
5766 if (AndBits.ugt(WidestAndBits))
5767 WidestAndBits = AndBits;
5768 if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5769 AndsToMaybeRemove.push_back(I);
5770 break;
5771 }
5772
5773 case Instruction::Shl: {
5774 auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5775 if (!ShlC)
5776 return false;
5777 uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5778 DemandBits.setLowBits(BitWidth - ShiftAmt);
5779 break;
5780 }
5781
5782 case Instruction::Trunc: {
5783 EVT TruncVT = TLI->getValueType(*DL, I->getType());
5784 unsigned TruncBitWidth = TruncVT.getSizeInBits();
5785 DemandBits.setLowBits(TruncBitWidth);
5786 break;
5787 }
5788
5789 default:
5790 return false;
5791 }
5792 }
5793
5794 uint32_t ActiveBits = DemandBits.getActiveBits();
5795 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5796 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5797 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5798 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5799 // followed by an AND.
5800 // TODO: Look into removing this restriction by fixing backends to either
5801 // return false for isLoadExtLegal for i1 or have them select this pattern to
5802 // a single instruction.
5803 //
5804 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5805 // mask, since these are the only ands that will be removed by isel.
5806 if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) ||
5807 WidestAndBits != DemandBits)
5808 return false;
5809
5810 LLVMContext &Ctx = Load->getType()->getContext();
5811 Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5812 EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5813
5814 // Reject cases that won't be matched as extloads.
5815 if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5816 !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5817 return false;
5818
5819 IRBuilder<> Builder(Load->getNextNode());
5820 auto *NewAnd = dyn_cast<Instruction>(
5821 Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5822 // Mark this instruction as "inserted by CGP", so that other
5823 // optimizations don't touch it.
5824 InsertedInsts.insert(NewAnd);
5825
5826 // Replace all uses of load with new and (except for the use of load in the
5827 // new and itself).
5828 Load->replaceAllUsesWith(NewAnd);
5829 NewAnd->setOperand(0, Load);
5830
5831 // Remove any and instructions that are now redundant.
5832 for (auto *And : AndsToMaybeRemove)
5833 // Check that the and mask is the same as the one we decided to put on the
5834 // new and.
5835 if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5836 And->replaceAllUsesWith(NewAnd);
5837 if (&*CurInstIterator == And)
5838 CurInstIterator = std::next(And->getIterator());
5839 And->eraseFromParent();
5840 ++NumAndUses;
5841 }
5842
5843 ++NumAndsAdded;
5844 return true;
5845 }
5846
5847 /// Check if V (an operand of a select instruction) is an expensive instruction
5848 /// that is only used once.
sinkSelectOperand(const TargetTransformInfo * TTI,Value * V)5849 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5850 auto *I = dyn_cast<Instruction>(V);
5851 // If it's safe to speculatively execute, then it should not have side
5852 // effects; therefore, it's safe to sink and possibly *not* execute.
5853 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5854 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
5855 }
5856
5857 /// Returns true if a SelectInst should be turned into an explicit branch.
isFormingBranchFromSelectProfitable(const TargetTransformInfo * TTI,const TargetLowering * TLI,SelectInst * SI)5858 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
5859 const TargetLowering *TLI,
5860 SelectInst *SI) {
5861 // If even a predictable select is cheap, then a branch can't be cheaper.
5862 if (!TLI->isPredictableSelectExpensive())
5863 return false;
5864
5865 // FIXME: This should use the same heuristics as IfConversion to determine
5866 // whether a select is better represented as a branch.
5867
5868 // If metadata tells us that the select condition is obviously predictable,
5869 // then we want to replace the select with a branch.
5870 uint64_t TrueWeight, FalseWeight;
5871 if (SI->extractProfMetadata(TrueWeight, FalseWeight)) {
5872 uint64_t Max = std::max(TrueWeight, FalseWeight);
5873 uint64_t Sum = TrueWeight + FalseWeight;
5874 if (Sum != 0) {
5875 auto Probability = BranchProbability::getBranchProbability(Max, Sum);
5876 if (Probability > TLI->getPredictableBranchThreshold())
5877 return true;
5878 }
5879 }
5880
5881 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5882
5883 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5884 // comparison condition. If the compare has more than one use, there's
5885 // probably another cmov or setcc around, so it's not worth emitting a branch.
5886 if (!Cmp || !Cmp->hasOneUse())
5887 return false;
5888
5889 // If either operand of the select is expensive and only needed on one side
5890 // of the select, we should form a branch.
5891 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5892 sinkSelectOperand(TTI, SI->getFalseValue()))
5893 return true;
5894
5895 return false;
5896 }
5897
5898 /// If \p isTrue is true, return the true value of \p SI, otherwise return
5899 /// false value of \p SI. If the true/false value of \p SI is defined by any
5900 /// select instructions in \p Selects, look through the defining select
5901 /// instruction until the true/false value is not defined in \p Selects.
getTrueOrFalseValue(SelectInst * SI,bool isTrue,const SmallPtrSet<const Instruction *,2> & Selects)5902 static Value *getTrueOrFalseValue(
5903 SelectInst *SI, bool isTrue,
5904 const SmallPtrSet<const Instruction *, 2> &Selects) {
5905 Value *V = nullptr;
5906
5907 for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI);
5908 DefSI = dyn_cast<SelectInst>(V)) {
5909 assert(DefSI->getCondition() == SI->getCondition() &&
5910 "The condition of DefSI does not match with SI");
5911 V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
5912 }
5913
5914 assert(V && "Failed to get select true/false value");
5915 return V;
5916 }
5917
optimizeShiftInst(BinaryOperator * Shift)5918 bool CodeGenPrepare::optimizeShiftInst(BinaryOperator *Shift) {
5919 assert(Shift->isShift() && "Expected a shift");
5920
5921 // If this is (1) a vector shift, (2) shifts by scalars are cheaper than
5922 // general vector shifts, and (3) the shift amount is a select-of-splatted
5923 // values, hoist the shifts before the select:
5924 // shift Op0, (select Cond, TVal, FVal) -->
5925 // select Cond, (shift Op0, TVal), (shift Op0, FVal)
5926 //
5927 // This is inverting a generic IR transform when we know that the cost of a
5928 // general vector shift is more than the cost of 2 shift-by-scalars.
5929 // We can't do this effectively in SDAG because we may not be able to
5930 // determine if the select operands are splats from within a basic block.
5931 Type *Ty = Shift->getType();
5932 if (!Ty->isVectorTy() || !TLI->isVectorShiftByScalarCheap(Ty))
5933 return false;
5934 Value *Cond, *TVal, *FVal;
5935 if (!match(Shift->getOperand(1),
5936 m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))))
5937 return false;
5938 if (!isSplatValue(TVal) || !isSplatValue(FVal))
5939 return false;
5940
5941 IRBuilder<> Builder(Shift);
5942 BinaryOperator::BinaryOps Opcode = Shift->getOpcode();
5943 Value *NewTVal = Builder.CreateBinOp(Opcode, Shift->getOperand(0), TVal);
5944 Value *NewFVal = Builder.CreateBinOp(Opcode, Shift->getOperand(0), FVal);
5945 Value *NewSel = Builder.CreateSelect(Cond, NewTVal, NewFVal);
5946 Shift->replaceAllUsesWith(NewSel);
5947 Shift->eraseFromParent();
5948 return true;
5949 }
5950
5951 /// If we have a SelectInst that will likely profit from branch prediction,
5952 /// turn it into a branch.
optimizeSelectInst(SelectInst * SI)5953 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5954 // If branch conversion isn't desirable, exit early.
5955 if (DisableSelectToBranch || OptSize || !TLI)
5956 return false;
5957
5958 // Find all consecutive select instructions that share the same condition.
5959 SmallVector<SelectInst *, 2> ASI;
5960 ASI.push_back(SI);
5961 for (BasicBlock::iterator It = ++BasicBlock::iterator(SI);
5962 It != SI->getParent()->end(); ++It) {
5963 SelectInst *I = dyn_cast<SelectInst>(&*It);
5964 if (I && SI->getCondition() == I->getCondition()) {
5965 ASI.push_back(I);
5966 } else {
5967 break;
5968 }
5969 }
5970
5971 SelectInst *LastSI = ASI.back();
5972 // Increment the current iterator to skip all the rest of select instructions
5973 // because they will be either "not lowered" or "all lowered" to branch.
5974 CurInstIterator = std::next(LastSI->getIterator());
5975
5976 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
5977
5978 // Can we convert the 'select' to CF ?
5979 if (VectorCond || SI->getMetadata(LLVMContext::MD_unpredictable))
5980 return false;
5981
5982 TargetLowering::SelectSupportKind SelectKind;
5983 if (VectorCond)
5984 SelectKind = TargetLowering::VectorMaskSelect;
5985 else if (SI->getType()->isVectorTy())
5986 SelectKind = TargetLowering::ScalarCondVectorVal;
5987 else
5988 SelectKind = TargetLowering::ScalarValSelect;
5989
5990 if (TLI->isSelectSupported(SelectKind) &&
5991 !isFormingBranchFromSelectProfitable(TTI, TLI, SI))
5992 return false;
5993
5994 // The DominatorTree needs to be rebuilt by any consumers after this
5995 // transformation. We simply reset here rather than setting the ModifiedDT
5996 // flag to avoid restarting the function walk in runOnFunction for each
5997 // select optimized.
5998 DT.reset();
5999
6000 // Transform a sequence like this:
6001 // start:
6002 // %cmp = cmp uge i32 %a, %b
6003 // %sel = select i1 %cmp, i32 %c, i32 %d
6004 //
6005 // Into:
6006 // start:
6007 // %cmp = cmp uge i32 %a, %b
6008 // br i1 %cmp, label %select.true, label %select.false
6009 // select.true:
6010 // br label %select.end
6011 // select.false:
6012 // br label %select.end
6013 // select.end:
6014 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
6015 //
6016 // In addition, we may sink instructions that produce %c or %d from
6017 // the entry block into the destination(s) of the new branch.
6018 // If the true or false blocks do not contain a sunken instruction, that
6019 // block and its branch may be optimized away. In that case, one side of the
6020 // first branch will point directly to select.end, and the corresponding PHI
6021 // predecessor block will be the start block.
6022
6023 // First, we split the block containing the select into 2 blocks.
6024 BasicBlock *StartBlock = SI->getParent();
6025 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
6026 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
6027
6028 // Delete the unconditional branch that was just created by the split.
6029 StartBlock->getTerminator()->eraseFromParent();
6030
6031 // These are the new basic blocks for the conditional branch.
6032 // At least one will become an actual new basic block.
6033 BasicBlock *TrueBlock = nullptr;
6034 BasicBlock *FalseBlock = nullptr;
6035 BranchInst *TrueBranch = nullptr;
6036 BranchInst *FalseBranch = nullptr;
6037
6038 // Sink expensive instructions into the conditional blocks to avoid executing
6039 // them speculatively.
6040 for (SelectInst *SI : ASI) {
6041 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
6042 if (TrueBlock == nullptr) {
6043 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
6044 EndBlock->getParent(), EndBlock);
6045 TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
6046 TrueBranch->setDebugLoc(SI->getDebugLoc());
6047 }
6048 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
6049 TrueInst->moveBefore(TrueBranch);
6050 }
6051 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
6052 if (FalseBlock == nullptr) {
6053 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
6054 EndBlock->getParent(), EndBlock);
6055 FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
6056 FalseBranch->setDebugLoc(SI->getDebugLoc());
6057 }
6058 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
6059 FalseInst->moveBefore(FalseBranch);
6060 }
6061 }
6062
6063 // If there was nothing to sink, then arbitrarily choose the 'false' side
6064 // for a new input value to the PHI.
6065 if (TrueBlock == FalseBlock) {
6066 assert(TrueBlock == nullptr &&
6067 "Unexpected basic block transform while optimizing select");
6068
6069 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
6070 EndBlock->getParent(), EndBlock);
6071 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
6072 FalseBranch->setDebugLoc(SI->getDebugLoc());
6073 }
6074
6075 // Insert the real conditional branch based on the original condition.
6076 // If we did not create a new block for one of the 'true' or 'false' paths
6077 // of the condition, it means that side of the branch goes to the end block
6078 // directly and the path originates from the start block from the point of
6079 // view of the new PHI.
6080 BasicBlock *TT, *FT;
6081 if (TrueBlock == nullptr) {
6082 TT = EndBlock;
6083 FT = FalseBlock;
6084 TrueBlock = StartBlock;
6085 } else if (FalseBlock == nullptr) {
6086 TT = TrueBlock;
6087 FT = EndBlock;
6088 FalseBlock = StartBlock;
6089 } else {
6090 TT = TrueBlock;
6091 FT = FalseBlock;
6092 }
6093 IRBuilder<>(SI).CreateCondBr(SI->getCondition(), TT, FT, SI);
6094
6095 SmallPtrSet<const Instruction *, 2> INS;
6096 INS.insert(ASI.begin(), ASI.end());
6097 // Use reverse iterator because later select may use the value of the
6098 // earlier select, and we need to propagate value through earlier select
6099 // to get the PHI operand.
6100 for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) {
6101 SelectInst *SI = *It;
6102 // The select itself is replaced with a PHI Node.
6103 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
6104 PN->takeName(SI);
6105 PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock);
6106 PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock);
6107 PN->setDebugLoc(SI->getDebugLoc());
6108
6109 SI->replaceAllUsesWith(PN);
6110 SI->eraseFromParent();
6111 INS.erase(SI);
6112 ++NumSelectsExpanded;
6113 }
6114
6115 // Instruct OptimizeBlock to skip to the next block.
6116 CurInstIterator = StartBlock->end();
6117 return true;
6118 }
6119
isBroadcastShuffle(ShuffleVectorInst * SVI)6120 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
6121 SmallVector<int, 16> Mask(SVI->getShuffleMask());
6122 int SplatElem = -1;
6123 for (unsigned i = 0; i < Mask.size(); ++i) {
6124 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
6125 return false;
6126 SplatElem = Mask[i];
6127 }
6128
6129 return true;
6130 }
6131
6132 /// Some targets have expensive vector shifts if the lanes aren't all the same
6133 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
6134 /// it's often worth sinking a shufflevector splat down to its use so that
6135 /// codegen can spot all lanes are identical.
optimizeShuffleVectorInst(ShuffleVectorInst * SVI)6136 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
6137 BasicBlock *DefBB = SVI->getParent();
6138
6139 // Only do this xform if variable vector shifts are particularly expensive.
6140 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
6141 return false;
6142
6143 // We only expect better codegen by sinking a shuffle if we can recognise a
6144 // constant splat.
6145 if (!isBroadcastShuffle(SVI))
6146 return false;
6147
6148 // InsertedShuffles - Only insert a shuffle in each block once.
6149 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
6150
6151 bool MadeChange = false;
6152 for (User *U : SVI->users()) {
6153 Instruction *UI = cast<Instruction>(U);
6154
6155 // Figure out which BB this ext is used in.
6156 BasicBlock *UserBB = UI->getParent();
6157 if (UserBB == DefBB) continue;
6158
6159 // For now only apply this when the splat is used by a shift instruction.
6160 if (!UI->isShift()) continue;
6161
6162 // Everything checks out, sink the shuffle if the user's block doesn't
6163 // already have a copy.
6164 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
6165
6166 if (!InsertedShuffle) {
6167 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
6168 assert(InsertPt != UserBB->end());
6169 InsertedShuffle =
6170 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
6171 SVI->getOperand(2), "", &*InsertPt);
6172 InsertedShuffle->setDebugLoc(SVI->getDebugLoc());
6173 }
6174
6175 UI->replaceUsesOfWith(SVI, InsertedShuffle);
6176 MadeChange = true;
6177 }
6178
6179 // If we removed all uses, nuke the shuffle.
6180 if (SVI->use_empty()) {
6181 SVI->eraseFromParent();
6182 MadeChange = true;
6183 }
6184
6185 return MadeChange;
6186 }
6187
tryToSinkFreeOperands(Instruction * I)6188 bool CodeGenPrepare::tryToSinkFreeOperands(Instruction *I) {
6189 // If the operands of I can be folded into a target instruction together with
6190 // I, duplicate and sink them.
6191 SmallVector<Use *, 4> OpsToSink;
6192 if (!TLI || !TLI->shouldSinkOperands(I, OpsToSink))
6193 return false;
6194
6195 // OpsToSink can contain multiple uses in a use chain (e.g.
6196 // (%u1 with %u1 = shufflevector), (%u2 with %u2 = zext %u1)). The dominating
6197 // uses must come first, which means they are sunk first, temporarily creating
6198 // invalid IR. This will be fixed once their dominated users are sunk and
6199 // updated.
6200 BasicBlock *TargetBB = I->getParent();
6201 bool Changed = false;
6202 SmallVector<Use *, 4> ToReplace;
6203 for (Use *U : OpsToSink) {
6204 auto *UI = cast<Instruction>(U->get());
6205 if (UI->getParent() == TargetBB || isa<PHINode>(UI))
6206 continue;
6207 ToReplace.push_back(U);
6208 }
6209
6210 SmallPtrSet<Instruction *, 4> MaybeDead;
6211 for (Use *U : ToReplace) {
6212 auto *UI = cast<Instruction>(U->get());
6213 Instruction *NI = UI->clone();
6214 MaybeDead.insert(UI);
6215 LLVM_DEBUG(dbgs() << "Sinking " << *UI << " to user " << *I << "\n");
6216 NI->insertBefore(I);
6217 InsertedInsts.insert(NI);
6218 U->set(NI);
6219 Changed = true;
6220 }
6221
6222 // Remove instructions that are dead after sinking.
6223 for (auto *I : MaybeDead)
6224 if (!I->hasNUsesOrMore(1))
6225 I->eraseFromParent();
6226
6227 return Changed;
6228 }
6229
optimizeSwitchInst(SwitchInst * SI)6230 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
6231 if (!TLI || !DL)
6232 return false;
6233
6234 Value *Cond = SI->getCondition();
6235 Type *OldType = Cond->getType();
6236 LLVMContext &Context = Cond->getContext();
6237 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
6238 unsigned RegWidth = RegType.getSizeInBits();
6239
6240 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
6241 return false;
6242
6243 // If the register width is greater than the type width, expand the condition
6244 // of the switch instruction and each case constant to the width of the
6245 // register. By widening the type of the switch condition, subsequent
6246 // comparisons (for case comparisons) will not need to be extended to the
6247 // preferred register width, so we will potentially eliminate N-1 extends,
6248 // where N is the number of cases in the switch.
6249 auto *NewType = Type::getIntNTy(Context, RegWidth);
6250
6251 // Zero-extend the switch condition and case constants unless the switch
6252 // condition is a function argument that is already being sign-extended.
6253 // In that case, we can avoid an unnecessary mask/extension by sign-extending
6254 // everything instead.
6255 Instruction::CastOps ExtType = Instruction::ZExt;
6256 if (auto *Arg = dyn_cast<Argument>(Cond))
6257 if (Arg->hasSExtAttr())
6258 ExtType = Instruction::SExt;
6259
6260 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
6261 ExtInst->insertBefore(SI);
6262 ExtInst->setDebugLoc(SI->getDebugLoc());
6263 SI->setCondition(ExtInst);
6264 for (auto Case : SI->cases()) {
6265 APInt NarrowConst = Case.getCaseValue()->getValue();
6266 APInt WideConst = (ExtType == Instruction::ZExt) ?
6267 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
6268 Case.setValue(ConstantInt::get(Context, WideConst));
6269 }
6270
6271 return true;
6272 }
6273
6274
6275 namespace {
6276
6277 /// Helper class to promote a scalar operation to a vector one.
6278 /// This class is used to move downward extractelement transition.
6279 /// E.g.,
6280 /// a = vector_op <2 x i32>
6281 /// b = extractelement <2 x i32> a, i32 0
6282 /// c = scalar_op b
6283 /// store c
6284 ///
6285 /// =>
6286 /// a = vector_op <2 x i32>
6287 /// c = vector_op a (equivalent to scalar_op on the related lane)
6288 /// * d = extractelement <2 x i32> c, i32 0
6289 /// * store d
6290 /// Assuming both extractelement and store can be combine, we get rid of the
6291 /// transition.
6292 class VectorPromoteHelper {
6293 /// DataLayout associated with the current module.
6294 const DataLayout &DL;
6295
6296 /// Used to perform some checks on the legality of vector operations.
6297 const TargetLowering &TLI;
6298
6299 /// Used to estimated the cost of the promoted chain.
6300 const TargetTransformInfo &TTI;
6301
6302 /// The transition being moved downwards.
6303 Instruction *Transition;
6304
6305 /// The sequence of instructions to be promoted.
6306 SmallVector<Instruction *, 4> InstsToBePromoted;
6307
6308 /// Cost of combining a store and an extract.
6309 unsigned StoreExtractCombineCost;
6310
6311 /// Instruction that will be combined with the transition.
6312 Instruction *CombineInst = nullptr;
6313
6314 /// The instruction that represents the current end of the transition.
6315 /// Since we are faking the promotion until we reach the end of the chain
6316 /// of computation, we need a way to get the current end of the transition.
getEndOfTransition() const6317 Instruction *getEndOfTransition() const {
6318 if (InstsToBePromoted.empty())
6319 return Transition;
6320 return InstsToBePromoted.back();
6321 }
6322
6323 /// Return the index of the original value in the transition.
6324 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
6325 /// c, is at index 0.
getTransitionOriginalValueIdx() const6326 unsigned getTransitionOriginalValueIdx() const {
6327 assert(isa<ExtractElementInst>(Transition) &&
6328 "Other kind of transitions are not supported yet");
6329 return 0;
6330 }
6331
6332 /// Return the index of the index in the transition.
6333 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
6334 /// is at index 1.
getTransitionIdx() const6335 unsigned getTransitionIdx() const {
6336 assert(isa<ExtractElementInst>(Transition) &&
6337 "Other kind of transitions are not supported yet");
6338 return 1;
6339 }
6340
6341 /// Get the type of the transition.
6342 /// This is the type of the original value.
6343 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
6344 /// transition is <2 x i32>.
getTransitionType() const6345 Type *getTransitionType() const {
6346 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
6347 }
6348
6349 /// Promote \p ToBePromoted by moving \p Def downward through.
6350 /// I.e., we have the following sequence:
6351 /// Def = Transition <ty1> a to <ty2>
6352 /// b = ToBePromoted <ty2> Def, ...
6353 /// =>
6354 /// b = ToBePromoted <ty1> a, ...
6355 /// Def = Transition <ty1> ToBePromoted to <ty2>
6356 void promoteImpl(Instruction *ToBePromoted);
6357
6358 /// Check whether or not it is profitable to promote all the
6359 /// instructions enqueued to be promoted.
isProfitableToPromote()6360 bool isProfitableToPromote() {
6361 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
6362 unsigned Index = isa<ConstantInt>(ValIdx)
6363 ? cast<ConstantInt>(ValIdx)->getZExtValue()
6364 : -1;
6365 Type *PromotedType = getTransitionType();
6366
6367 StoreInst *ST = cast<StoreInst>(CombineInst);
6368 unsigned AS = ST->getPointerAddressSpace();
6369 unsigned Align = ST->getAlignment();
6370 // Check if this store is supported.
6371 if (!TLI.allowsMisalignedMemoryAccesses(
6372 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
6373 Align)) {
6374 // If this is not supported, there is no way we can combine
6375 // the extract with the store.
6376 return false;
6377 }
6378
6379 // The scalar chain of computation has to pay for the transition
6380 // scalar to vector.
6381 // The vector chain has to account for the combining cost.
6382 uint64_t ScalarCost =
6383 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
6384 uint64_t VectorCost = StoreExtractCombineCost;
6385 for (const auto &Inst : InstsToBePromoted) {
6386 // Compute the cost.
6387 // By construction, all instructions being promoted are arithmetic ones.
6388 // Moreover, one argument is a constant that can be viewed as a splat
6389 // constant.
6390 Value *Arg0 = Inst->getOperand(0);
6391 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
6392 isa<ConstantFP>(Arg0);
6393 TargetTransformInfo::OperandValueKind Arg0OVK =
6394 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
6395 : TargetTransformInfo::OK_AnyValue;
6396 TargetTransformInfo::OperandValueKind Arg1OVK =
6397 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
6398 : TargetTransformInfo::OK_AnyValue;
6399 ScalarCost += TTI.getArithmeticInstrCost(
6400 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
6401 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
6402 Arg0OVK, Arg1OVK);
6403 }
6404 LLVM_DEBUG(
6405 dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
6406 << ScalarCost << "\nVector: " << VectorCost << '\n');
6407 return ScalarCost > VectorCost;
6408 }
6409
6410 /// Generate a constant vector with \p Val with the same
6411 /// number of elements as the transition.
6412 /// \p UseSplat defines whether or not \p Val should be replicated
6413 /// across the whole vector.
6414 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
6415 /// otherwise we generate a vector with as many undef as possible:
6416 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
6417 /// used at the index of the extract.
getConstantVector(Constant * Val,bool UseSplat) const6418 Value *getConstantVector(Constant *Val, bool UseSplat) const {
6419 unsigned ExtractIdx = std::numeric_limits<unsigned>::max();
6420 if (!UseSplat) {
6421 // If we cannot determine where the constant must be, we have to
6422 // use a splat constant.
6423 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
6424 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
6425 ExtractIdx = CstVal->getSExtValue();
6426 else
6427 UseSplat = true;
6428 }
6429
6430 unsigned End = getTransitionType()->getVectorNumElements();
6431 if (UseSplat)
6432 return ConstantVector::getSplat(End, Val);
6433
6434 SmallVector<Constant *, 4> ConstVec;
6435 UndefValue *UndefVal = UndefValue::get(Val->getType());
6436 for (unsigned Idx = 0; Idx != End; ++Idx) {
6437 if (Idx == ExtractIdx)
6438 ConstVec.push_back(Val);
6439 else
6440 ConstVec.push_back(UndefVal);
6441 }
6442 return ConstantVector::get(ConstVec);
6443 }
6444
6445 /// Check if promoting to a vector type an operand at \p OperandIdx
6446 /// in \p Use can trigger undefined behavior.
canCauseUndefinedBehavior(const Instruction * Use,unsigned OperandIdx)6447 static bool canCauseUndefinedBehavior(const Instruction *Use,
6448 unsigned OperandIdx) {
6449 // This is not safe to introduce undef when the operand is on
6450 // the right hand side of a division-like instruction.
6451 if (OperandIdx != 1)
6452 return false;
6453 switch (Use->getOpcode()) {
6454 default:
6455 return false;
6456 case Instruction::SDiv:
6457 case Instruction::UDiv:
6458 case Instruction::SRem:
6459 case Instruction::URem:
6460 return true;
6461 case Instruction::FDiv:
6462 case Instruction::FRem:
6463 return !Use->hasNoNaNs();
6464 }
6465 llvm_unreachable(nullptr);
6466 }
6467
6468 public:
VectorPromoteHelper(const DataLayout & DL,const TargetLowering & TLI,const TargetTransformInfo & TTI,Instruction * Transition,unsigned CombineCost)6469 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
6470 const TargetTransformInfo &TTI, Instruction *Transition,
6471 unsigned CombineCost)
6472 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
6473 StoreExtractCombineCost(CombineCost) {
6474 assert(Transition && "Do not know how to promote null");
6475 }
6476
6477 /// Check if we can promote \p ToBePromoted to \p Type.
canPromote(const Instruction * ToBePromoted) const6478 bool canPromote(const Instruction *ToBePromoted) const {
6479 // We could support CastInst too.
6480 return isa<BinaryOperator>(ToBePromoted);
6481 }
6482
6483 /// Check if it is profitable to promote \p ToBePromoted
6484 /// by moving downward the transition through.
shouldPromote(const Instruction * ToBePromoted) const6485 bool shouldPromote(const Instruction *ToBePromoted) const {
6486 // Promote only if all the operands can be statically expanded.
6487 // Indeed, we do not want to introduce any new kind of transitions.
6488 for (const Use &U : ToBePromoted->operands()) {
6489 const Value *Val = U.get();
6490 if (Val == getEndOfTransition()) {
6491 // If the use is a division and the transition is on the rhs,
6492 // we cannot promote the operation, otherwise we may create a
6493 // division by zero.
6494 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
6495 return false;
6496 continue;
6497 }
6498 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
6499 !isa<ConstantFP>(Val))
6500 return false;
6501 }
6502 // Check that the resulting operation is legal.
6503 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
6504 if (!ISDOpcode)
6505 return false;
6506 return StressStoreExtract ||
6507 TLI.isOperationLegalOrCustom(
6508 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
6509 }
6510
6511 /// Check whether or not \p Use can be combined
6512 /// with the transition.
6513 /// I.e., is it possible to do Use(Transition) => AnotherUse?
canCombine(const Instruction * Use)6514 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
6515
6516 /// Record \p ToBePromoted as part of the chain to be promoted.
enqueueForPromotion(Instruction * ToBePromoted)6517 void enqueueForPromotion(Instruction *ToBePromoted) {
6518 InstsToBePromoted.push_back(ToBePromoted);
6519 }
6520
6521 /// Set the instruction that will be combined with the transition.
recordCombineInstruction(Instruction * ToBeCombined)6522 void recordCombineInstruction(Instruction *ToBeCombined) {
6523 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
6524 CombineInst = ToBeCombined;
6525 }
6526
6527 /// Promote all the instructions enqueued for promotion if it is
6528 /// is profitable.
6529 /// \return True if the promotion happened, false otherwise.
promote()6530 bool promote() {
6531 // Check if there is something to promote.
6532 // Right now, if we do not have anything to combine with,
6533 // we assume the promotion is not profitable.
6534 if (InstsToBePromoted.empty() || !CombineInst)
6535 return false;
6536
6537 // Check cost.
6538 if (!StressStoreExtract && !isProfitableToPromote())
6539 return false;
6540
6541 // Promote.
6542 for (auto &ToBePromoted : InstsToBePromoted)
6543 promoteImpl(ToBePromoted);
6544 InstsToBePromoted.clear();
6545 return true;
6546 }
6547 };
6548
6549 } // end anonymous namespace
6550
promoteImpl(Instruction * ToBePromoted)6551 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
6552 // At this point, we know that all the operands of ToBePromoted but Def
6553 // can be statically promoted.
6554 // For Def, we need to use its parameter in ToBePromoted:
6555 // b = ToBePromoted ty1 a
6556 // Def = Transition ty1 b to ty2
6557 // Move the transition down.
6558 // 1. Replace all uses of the promoted operation by the transition.
6559 // = ... b => = ... Def.
6560 assert(ToBePromoted->getType() == Transition->getType() &&
6561 "The type of the result of the transition does not match "
6562 "the final type");
6563 ToBePromoted->replaceAllUsesWith(Transition);
6564 // 2. Update the type of the uses.
6565 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
6566 Type *TransitionTy = getTransitionType();
6567 ToBePromoted->mutateType(TransitionTy);
6568 // 3. Update all the operands of the promoted operation with promoted
6569 // operands.
6570 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
6571 for (Use &U : ToBePromoted->operands()) {
6572 Value *Val = U.get();
6573 Value *NewVal = nullptr;
6574 if (Val == Transition)
6575 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
6576 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
6577 isa<ConstantFP>(Val)) {
6578 // Use a splat constant if it is not safe to use undef.
6579 NewVal = getConstantVector(
6580 cast<Constant>(Val),
6581 isa<UndefValue>(Val) ||
6582 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
6583 } else
6584 llvm_unreachable("Did you modified shouldPromote and forgot to update "
6585 "this?");
6586 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
6587 }
6588 Transition->moveAfter(ToBePromoted);
6589 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
6590 }
6591
6592 /// Some targets can do store(extractelement) with one instruction.
6593 /// Try to push the extractelement towards the stores when the target
6594 /// has this feature and this is profitable.
optimizeExtractElementInst(Instruction * Inst)6595 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
6596 unsigned CombineCost = std::numeric_limits<unsigned>::max();
6597 if (DisableStoreExtract || !TLI ||
6598 (!StressStoreExtract &&
6599 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
6600 Inst->getOperand(1), CombineCost)))
6601 return false;
6602
6603 // At this point we know that Inst is a vector to scalar transition.
6604 // Try to move it down the def-use chain, until:
6605 // - We can combine the transition with its single use
6606 // => we got rid of the transition.
6607 // - We escape the current basic block
6608 // => we would need to check that we are moving it at a cheaper place and
6609 // we do not do that for now.
6610 BasicBlock *Parent = Inst->getParent();
6611 LLVM_DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
6612 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
6613 // If the transition has more than one use, assume this is not going to be
6614 // beneficial.
6615 while (Inst->hasOneUse()) {
6616 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
6617 LLVM_DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
6618
6619 if (ToBePromoted->getParent() != Parent) {
6620 LLVM_DEBUG(dbgs() << "Instruction to promote is in a different block ("
6621 << ToBePromoted->getParent()->getName()
6622 << ") than the transition (" << Parent->getName()
6623 << ").\n");
6624 return false;
6625 }
6626
6627 if (VPH.canCombine(ToBePromoted)) {
6628 LLVM_DEBUG(dbgs() << "Assume " << *Inst << '\n'
6629 << "will be combined with: " << *ToBePromoted << '\n');
6630 VPH.recordCombineInstruction(ToBePromoted);
6631 bool Changed = VPH.promote();
6632 NumStoreExtractExposed += Changed;
6633 return Changed;
6634 }
6635
6636 LLVM_DEBUG(dbgs() << "Try promoting.\n");
6637 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
6638 return false;
6639
6640 LLVM_DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
6641
6642 VPH.enqueueForPromotion(ToBePromoted);
6643 Inst = ToBePromoted;
6644 }
6645 return false;
6646 }
6647
6648 /// For the instruction sequence of store below, F and I values
6649 /// are bundled together as an i64 value before being stored into memory.
6650 /// Sometimes it is more efficient to generate separate stores for F and I,
6651 /// which can remove the bitwise instructions or sink them to colder places.
6652 ///
6653 /// (store (or (zext (bitcast F to i32) to i64),
6654 /// (shl (zext I to i64), 32)), addr) -->
6655 /// (store F, addr) and (store I, addr+4)
6656 ///
6657 /// Similarly, splitting for other merged store can also be beneficial, like:
6658 /// For pair of {i32, i32}, i64 store --> two i32 stores.
6659 /// For pair of {i32, i16}, i64 store --> two i32 stores.
6660 /// For pair of {i16, i16}, i32 store --> two i16 stores.
6661 /// For pair of {i16, i8}, i32 store --> two i16 stores.
6662 /// For pair of {i8, i8}, i16 store --> two i8 stores.
6663 ///
6664 /// We allow each target to determine specifically which kind of splitting is
6665 /// supported.
6666 ///
6667 /// The store patterns are commonly seen from the simple code snippet below
6668 /// if only std::make_pair(...) is sroa transformed before inlined into hoo.
6669 /// void goo(const std::pair<int, float> &);
6670 /// hoo() {
6671 /// ...
6672 /// goo(std::make_pair(tmp, ftmp));
6673 /// ...
6674 /// }
6675 ///
6676 /// Although we already have similar splitting in DAG Combine, we duplicate
6677 /// it in CodeGenPrepare to catch the case in which pattern is across
6678 /// multiple BBs. The logic in DAG Combine is kept to catch case generated
6679 /// during code expansion.
splitMergedValStore(StoreInst & SI,const DataLayout & DL,const TargetLowering & TLI)6680 static bool splitMergedValStore(StoreInst &SI, const DataLayout &DL,
6681 const TargetLowering &TLI) {
6682 // Handle simple but common cases only.
6683 Type *StoreType = SI.getValueOperand()->getType();
6684 if (!DL.typeSizeEqualsStoreSize(StoreType) ||
6685 DL.getTypeSizeInBits(StoreType) == 0)
6686 return false;
6687
6688 unsigned HalfValBitSize = DL.getTypeSizeInBits(StoreType) / 2;
6689 Type *SplitStoreType = Type::getIntNTy(SI.getContext(), HalfValBitSize);
6690 if (!DL.typeSizeEqualsStoreSize(SplitStoreType))
6691 return false;
6692
6693 // Don't split the store if it is volatile.
6694 if (SI.isVolatile())
6695 return false;
6696
6697 // Match the following patterns:
6698 // (store (or (zext LValue to i64),
6699 // (shl (zext HValue to i64), 32)), HalfValBitSize)
6700 // or
6701 // (store (or (shl (zext HValue to i64), 32)), HalfValBitSize)
6702 // (zext LValue to i64),
6703 // Expect both operands of OR and the first operand of SHL have only
6704 // one use.
6705 Value *LValue, *HValue;
6706 if (!match(SI.getValueOperand(),
6707 m_c_Or(m_OneUse(m_ZExt(m_Value(LValue))),
6708 m_OneUse(m_Shl(m_OneUse(m_ZExt(m_Value(HValue))),
6709 m_SpecificInt(HalfValBitSize))))))
6710 return false;
6711
6712 // Check LValue and HValue are int with size less or equal than 32.
6713 if (!LValue->getType()->isIntegerTy() ||
6714 DL.getTypeSizeInBits(LValue->getType()) > HalfValBitSize ||
6715 !HValue->getType()->isIntegerTy() ||
6716 DL.getTypeSizeInBits(HValue->getType()) > HalfValBitSize)
6717 return false;
6718
6719 // If LValue/HValue is a bitcast instruction, use the EVT before bitcast
6720 // as the input of target query.
6721 auto *LBC = dyn_cast<BitCastInst>(LValue);
6722 auto *HBC = dyn_cast<BitCastInst>(HValue);
6723 EVT LowTy = LBC ? EVT::getEVT(LBC->getOperand(0)->getType())
6724 : EVT::getEVT(LValue->getType());
6725 EVT HighTy = HBC ? EVT::getEVT(HBC->getOperand(0)->getType())
6726 : EVT::getEVT(HValue->getType());
6727 if (!ForceSplitStore && !TLI.isMultiStoresCheaperThanBitsMerge(LowTy, HighTy))
6728 return false;
6729
6730 // Start to split store.
6731 IRBuilder<> Builder(SI.getContext());
6732 Builder.SetInsertPoint(&SI);
6733
6734 // If LValue/HValue is a bitcast in another BB, create a new one in current
6735 // BB so it may be merged with the splitted stores by dag combiner.
6736 if (LBC && LBC->getParent() != SI.getParent())
6737 LValue = Builder.CreateBitCast(LBC->getOperand(0), LBC->getType());
6738 if (HBC && HBC->getParent() != SI.getParent())
6739 HValue = Builder.CreateBitCast(HBC->getOperand(0), HBC->getType());
6740
6741 bool IsLE = SI.getModule()->getDataLayout().isLittleEndian();
6742 auto CreateSplitStore = [&](Value *V, bool Upper) {
6743 V = Builder.CreateZExtOrBitCast(V, SplitStoreType);
6744 Value *Addr = Builder.CreateBitCast(
6745 SI.getOperand(1),
6746 SplitStoreType->getPointerTo(SI.getPointerAddressSpace()));
6747 if ((IsLE && Upper) || (!IsLE && !Upper))
6748 Addr = Builder.CreateGEP(
6749 SplitStoreType, Addr,
6750 ConstantInt::get(Type::getInt32Ty(SI.getContext()), 1));
6751 Builder.CreateAlignedStore(
6752 V, Addr, Upper ? SI.getAlignment() / 2 : SI.getAlignment());
6753 };
6754
6755 CreateSplitStore(LValue, false);
6756 CreateSplitStore(HValue, true);
6757
6758 // Delete the old store.
6759 SI.eraseFromParent();
6760 return true;
6761 }
6762
6763 // Return true if the GEP has two operands, the first operand is of a sequential
6764 // type, and the second operand is a constant.
GEPSequentialConstIndexed(GetElementPtrInst * GEP)6765 static bool GEPSequentialConstIndexed(GetElementPtrInst *GEP) {
6766 gep_type_iterator I = gep_type_begin(*GEP);
6767 return GEP->getNumOperands() == 2 &&
6768 I.isSequential() &&
6769 isa<ConstantInt>(GEP->getOperand(1));
6770 }
6771
6772 // Try unmerging GEPs to reduce liveness interference (register pressure) across
6773 // IndirectBr edges. Since IndirectBr edges tend to touch on many blocks,
6774 // reducing liveness interference across those edges benefits global register
6775 // allocation. Currently handles only certain cases.
6776 //
6777 // For example, unmerge %GEPI and %UGEPI as below.
6778 //
6779 // ---------- BEFORE ----------
6780 // SrcBlock:
6781 // ...
6782 // %GEPIOp = ...
6783 // ...
6784 // %GEPI = gep %GEPIOp, Idx
6785 // ...
6786 // indirectbr ... [ label %DstB0, label %DstB1, ... label %DstBi ... ]
6787 // (* %GEPI is alive on the indirectbr edges due to other uses ahead)
6788 // (* %GEPIOp is alive on the indirectbr edges only because of it's used by
6789 // %UGEPI)
6790 //
6791 // DstB0: ... (there may be a gep similar to %UGEPI to be unmerged)
6792 // DstB1: ... (there may be a gep similar to %UGEPI to be unmerged)
6793 // ...
6794 //
6795 // DstBi:
6796 // ...
6797 // %UGEPI = gep %GEPIOp, UIdx
6798 // ...
6799 // ---------------------------
6800 //
6801 // ---------- AFTER ----------
6802 // SrcBlock:
6803 // ... (same as above)
6804 // (* %GEPI is still alive on the indirectbr edges)
6805 // (* %GEPIOp is no longer alive on the indirectbr edges as a result of the
6806 // unmerging)
6807 // ...
6808 //
6809 // DstBi:
6810 // ...
6811 // %UGEPI = gep %GEPI, (UIdx-Idx)
6812 // ...
6813 // ---------------------------
6814 //
6815 // The register pressure on the IndirectBr edges is reduced because %GEPIOp is
6816 // no longer alive on them.
6817 //
6818 // We try to unmerge GEPs here in CodGenPrepare, as opposed to limiting merging
6819 // of GEPs in the first place in InstCombiner::visitGetElementPtrInst() so as
6820 // not to disable further simplications and optimizations as a result of GEP
6821 // merging.
6822 //
6823 // Note this unmerging may increase the length of the data flow critical path
6824 // (the path from %GEPIOp to %UGEPI would go through %GEPI), which is a tradeoff
6825 // between the register pressure and the length of data-flow critical
6826 // path. Restricting this to the uncommon IndirectBr case would minimize the
6827 // impact of potentially longer critical path, if any, and the impact on compile
6828 // time.
tryUnmergingGEPsAcrossIndirectBr(GetElementPtrInst * GEPI,const TargetTransformInfo * TTI)6829 static bool tryUnmergingGEPsAcrossIndirectBr(GetElementPtrInst *GEPI,
6830 const TargetTransformInfo *TTI) {
6831 BasicBlock *SrcBlock = GEPI->getParent();
6832 // Check that SrcBlock ends with an IndirectBr. If not, give up. The common
6833 // (non-IndirectBr) cases exit early here.
6834 if (!isa<IndirectBrInst>(SrcBlock->getTerminator()))
6835 return false;
6836 // Check that GEPI is a simple gep with a single constant index.
6837 if (!GEPSequentialConstIndexed(GEPI))
6838 return false;
6839 ConstantInt *GEPIIdx = cast<ConstantInt>(GEPI->getOperand(1));
6840 // Check that GEPI is a cheap one.
6841 if (TTI->getIntImmCost(GEPIIdx->getValue(), GEPIIdx->getType())
6842 > TargetTransformInfo::TCC_Basic)
6843 return false;
6844 Value *GEPIOp = GEPI->getOperand(0);
6845 // Check that GEPIOp is an instruction that's also defined in SrcBlock.
6846 if (!isa<Instruction>(GEPIOp))
6847 return false;
6848 auto *GEPIOpI = cast<Instruction>(GEPIOp);
6849 if (GEPIOpI->getParent() != SrcBlock)
6850 return false;
6851 // Check that GEP is used outside the block, meaning it's alive on the
6852 // IndirectBr edge(s).
6853 if (find_if(GEPI->users(), [&](User *Usr) {
6854 if (auto *I = dyn_cast<Instruction>(Usr)) {
6855 if (I->getParent() != SrcBlock) {
6856 return true;
6857 }
6858 }
6859 return false;
6860 }) == GEPI->users().end())
6861 return false;
6862 // The second elements of the GEP chains to be unmerged.
6863 std::vector<GetElementPtrInst *> UGEPIs;
6864 // Check each user of GEPIOp to check if unmerging would make GEPIOp not alive
6865 // on IndirectBr edges.
6866 for (User *Usr : GEPIOp->users()) {
6867 if (Usr == GEPI) continue;
6868 // Check if Usr is an Instruction. If not, give up.
6869 if (!isa<Instruction>(Usr))
6870 return false;
6871 auto *UI = cast<Instruction>(Usr);
6872 // Check if Usr in the same block as GEPIOp, which is fine, skip.
6873 if (UI->getParent() == SrcBlock)
6874 continue;
6875 // Check if Usr is a GEP. If not, give up.
6876 if (!isa<GetElementPtrInst>(Usr))
6877 return false;
6878 auto *UGEPI = cast<GetElementPtrInst>(Usr);
6879 // Check if UGEPI is a simple gep with a single constant index and GEPIOp is
6880 // the pointer operand to it. If so, record it in the vector. If not, give
6881 // up.
6882 if (!GEPSequentialConstIndexed(UGEPI))
6883 return false;
6884 if (UGEPI->getOperand(0) != GEPIOp)
6885 return false;
6886 if (GEPIIdx->getType() !=
6887 cast<ConstantInt>(UGEPI->getOperand(1))->getType())
6888 return false;
6889 ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
6890 if (TTI->getIntImmCost(UGEPIIdx->getValue(), UGEPIIdx->getType())
6891 > TargetTransformInfo::TCC_Basic)
6892 return false;
6893 UGEPIs.push_back(UGEPI);
6894 }
6895 if (UGEPIs.size() == 0)
6896 return false;
6897 // Check the materializing cost of (Uidx-Idx).
6898 for (GetElementPtrInst *UGEPI : UGEPIs) {
6899 ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
6900 APInt NewIdx = UGEPIIdx->getValue() - GEPIIdx->getValue();
6901 unsigned ImmCost = TTI->getIntImmCost(NewIdx, GEPIIdx->getType());
6902 if (ImmCost > TargetTransformInfo::TCC_Basic)
6903 return false;
6904 }
6905 // Now unmerge between GEPI and UGEPIs.
6906 for (GetElementPtrInst *UGEPI : UGEPIs) {
6907 UGEPI->setOperand(0, GEPI);
6908 ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
6909 Constant *NewUGEPIIdx =
6910 ConstantInt::get(GEPIIdx->getType(),
6911 UGEPIIdx->getValue() - GEPIIdx->getValue());
6912 UGEPI->setOperand(1, NewUGEPIIdx);
6913 // If GEPI is not inbounds but UGEPI is inbounds, change UGEPI to not
6914 // inbounds to avoid UB.
6915 if (!GEPI->isInBounds()) {
6916 UGEPI->setIsInBounds(false);
6917 }
6918 }
6919 // After unmerging, verify that GEPIOp is actually only used in SrcBlock (not
6920 // alive on IndirectBr edges).
6921 assert(find_if(GEPIOp->users(), [&](User *Usr) {
6922 return cast<Instruction>(Usr)->getParent() != SrcBlock;
6923 }) == GEPIOp->users().end() && "GEPIOp is used outside SrcBlock");
6924 return true;
6925 }
6926
optimizeInst(Instruction * I,bool & ModifiedDT)6927 bool CodeGenPrepare::optimizeInst(Instruction *I, bool &ModifiedDT) {
6928 // Bail out if we inserted the instruction to prevent optimizations from
6929 // stepping on each other's toes.
6930 if (InsertedInsts.count(I))
6931 return false;
6932
6933 // TODO: Move into the switch on opcode below here.
6934 if (PHINode *P = dyn_cast<PHINode>(I)) {
6935 // It is possible for very late stage optimizations (such as SimplifyCFG)
6936 // to introduce PHI nodes too late to be cleaned up. If we detect such a
6937 // trivial PHI, go ahead and zap it here.
6938 if (Value *V = SimplifyInstruction(P, {*DL, TLInfo})) {
6939 LargeOffsetGEPMap.erase(P);
6940 P->replaceAllUsesWith(V);
6941 P->eraseFromParent();
6942 ++NumPHIsElim;
6943 return true;
6944 }
6945 return false;
6946 }
6947
6948 if (CastInst *CI = dyn_cast<CastInst>(I)) {
6949 // If the source of the cast is a constant, then this should have
6950 // already been constant folded. The only reason NOT to constant fold
6951 // it is if something (e.g. LSR) was careful to place the constant
6952 // evaluation in a block other than then one that uses it (e.g. to hoist
6953 // the address of globals out of a loop). If this is the case, we don't
6954 // want to forward-subst the cast.
6955 if (isa<Constant>(CI->getOperand(0)))
6956 return false;
6957
6958 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
6959 return true;
6960
6961 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6962 /// Sink a zext or sext into its user blocks if the target type doesn't
6963 /// fit in one register
6964 if (TLI &&
6965 TLI->getTypeAction(CI->getContext(),
6966 TLI->getValueType(*DL, CI->getType())) ==
6967 TargetLowering::TypeExpandInteger) {
6968 return SinkCast(CI);
6969 } else {
6970 bool MadeChange = optimizeExt(I);
6971 return MadeChange | optimizeExtUses(I);
6972 }
6973 }
6974 return false;
6975 }
6976
6977 if (auto *Cmp = dyn_cast<CmpInst>(I))
6978 if (TLI && optimizeCmp(Cmp, ModifiedDT))
6979 return true;
6980
6981 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
6982 LI->setMetadata(LLVMContext::MD_invariant_group, nullptr);
6983 if (TLI) {
6984 bool Modified = optimizeLoadExt(LI);
6985 unsigned AS = LI->getPointerAddressSpace();
6986 Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
6987 return Modified;
6988 }
6989 return false;
6990 }
6991
6992 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
6993 if (TLI && splitMergedValStore(*SI, *DL, *TLI))
6994 return true;
6995 SI->setMetadata(LLVMContext::MD_invariant_group, nullptr);
6996 if (TLI) {
6997 unsigned AS = SI->getPointerAddressSpace();
6998 return optimizeMemoryInst(I, SI->getOperand(1),
6999 SI->getOperand(0)->getType(), AS);
7000 }
7001 return false;
7002 }
7003
7004 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
7005 unsigned AS = RMW->getPointerAddressSpace();
7006 return optimizeMemoryInst(I, RMW->getPointerOperand(),
7007 RMW->getType(), AS);
7008 }
7009
7010 if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(I)) {
7011 unsigned AS = CmpX->getPointerAddressSpace();
7012 return optimizeMemoryInst(I, CmpX->getPointerOperand(),
7013 CmpX->getCompareOperand()->getType(), AS);
7014 }
7015
7016 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
7017
7018 if (BinOp && (BinOp->getOpcode() == Instruction::And) &&
7019 EnableAndCmpSinking && TLI)
7020 return sinkAndCmp0Expression(BinOp, *TLI, InsertedInsts);
7021
7022 // TODO: Move this into the switch on opcode - it handles shifts already.
7023 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
7024 BinOp->getOpcode() == Instruction::LShr)) {
7025 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
7026 if (TLI && CI && TLI->hasExtractBitsInsn())
7027 if (OptimizeExtractBits(BinOp, CI, *TLI, *DL))
7028 return true;
7029 }
7030
7031 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
7032 if (GEPI->hasAllZeroIndices()) {
7033 /// The GEP operand must be a pointer, so must its result -> BitCast
7034 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
7035 GEPI->getName(), GEPI);
7036 NC->setDebugLoc(GEPI->getDebugLoc());
7037 GEPI->replaceAllUsesWith(NC);
7038 GEPI->eraseFromParent();
7039 ++NumGEPsElim;
7040 optimizeInst(NC, ModifiedDT);
7041 return true;
7042 }
7043 if (tryUnmergingGEPsAcrossIndirectBr(GEPI, TTI)) {
7044 return true;
7045 }
7046 return false;
7047 }
7048
7049 if (tryToSinkFreeOperands(I))
7050 return true;
7051
7052 switch (I->getOpcode()) {
7053 case Instruction::Shl:
7054 case Instruction::LShr:
7055 case Instruction::AShr:
7056 return optimizeShiftInst(cast<BinaryOperator>(I));
7057 case Instruction::Call:
7058 return optimizeCallInst(cast<CallInst>(I), ModifiedDT);
7059 case Instruction::Select:
7060 return optimizeSelectInst(cast<SelectInst>(I));
7061 case Instruction::ShuffleVector:
7062 return optimizeShuffleVectorInst(cast<ShuffleVectorInst>(I));
7063 case Instruction::Switch:
7064 return optimizeSwitchInst(cast<SwitchInst>(I));
7065 case Instruction::ExtractElement:
7066 return optimizeExtractElementInst(cast<ExtractElementInst>(I));
7067 }
7068
7069 return false;
7070 }
7071
7072 /// Given an OR instruction, check to see if this is a bitreverse
7073 /// idiom. If so, insert the new intrinsic and return true.
makeBitReverse(Instruction & I,const DataLayout & DL,const TargetLowering & TLI)7074 static bool makeBitReverse(Instruction &I, const DataLayout &DL,
7075 const TargetLowering &TLI) {
7076 if (!I.getType()->isIntegerTy() ||
7077 !TLI.isOperationLegalOrCustom(ISD::BITREVERSE,
7078 TLI.getValueType(DL, I.getType(), true)))
7079 return false;
7080
7081 SmallVector<Instruction*, 4> Insts;
7082 if (!recognizeBSwapOrBitReverseIdiom(&I, false, true, Insts))
7083 return false;
7084 Instruction *LastInst = Insts.back();
7085 I.replaceAllUsesWith(LastInst);
7086 RecursivelyDeleteTriviallyDeadInstructions(&I);
7087 return true;
7088 }
7089
7090 // In this pass we look for GEP and cast instructions that are used
7091 // across basic blocks and rewrite them to improve basic-block-at-a-time
7092 // selection.
optimizeBlock(BasicBlock & BB,bool & ModifiedDT)7093 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool &ModifiedDT) {
7094 SunkAddrs.clear();
7095 bool MadeChange = false;
7096
7097 CurInstIterator = BB.begin();
7098 while (CurInstIterator != BB.end()) {
7099 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
7100 if (ModifiedDT)
7101 return true;
7102 }
7103
7104 bool MadeBitReverse = true;
7105 while (TLI && MadeBitReverse) {
7106 MadeBitReverse = false;
7107 for (auto &I : reverse(BB)) {
7108 if (makeBitReverse(I, *DL, *TLI)) {
7109 MadeBitReverse = MadeChange = true;
7110 ModifiedDT = true;
7111 break;
7112 }
7113 }
7114 }
7115 MadeChange |= dupRetToEnableTailCallOpts(&BB, ModifiedDT);
7116
7117 return MadeChange;
7118 }
7119
7120 // llvm.dbg.value is far away from the value then iSel may not be able
7121 // handle it properly. iSel will drop llvm.dbg.value if it can not
7122 // find a node corresponding to the value.
placeDbgValues(Function & F)7123 bool CodeGenPrepare::placeDbgValues(Function &F) {
7124 bool MadeChange = false;
7125 for (BasicBlock &BB : F) {
7126 Instruction *PrevNonDbgInst = nullptr;
7127 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
7128 Instruction *Insn = &*BI++;
7129 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
7130 // Leave dbg.values that refer to an alloca alone. These
7131 // intrinsics describe the address of a variable (= the alloca)
7132 // being taken. They should not be moved next to the alloca
7133 // (and to the beginning of the scope), but rather stay close to
7134 // where said address is used.
7135 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
7136 PrevNonDbgInst = Insn;
7137 continue;
7138 }
7139
7140 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
7141 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
7142 // If VI is a phi in a block with an EHPad terminator, we can't insert
7143 // after it.
7144 if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
7145 continue;
7146 LLVM_DEBUG(dbgs() << "Moving Debug Value before :\n"
7147 << *DVI << ' ' << *VI);
7148 DVI->removeFromParent();
7149 if (isa<PHINode>(VI))
7150 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
7151 else
7152 DVI->insertAfter(VI);
7153 MadeChange = true;
7154 ++NumDbgValueMoved;
7155 }
7156 }
7157 }
7158 return MadeChange;
7159 }
7160
7161 /// Scale down both weights to fit into uint32_t.
scaleWeights(uint64_t & NewTrue,uint64_t & NewFalse)7162 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
7163 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
7164 uint32_t Scale = (NewMax / std::numeric_limits<uint32_t>::max()) + 1;
7165 NewTrue = NewTrue / Scale;
7166 NewFalse = NewFalse / Scale;
7167 }
7168
7169 /// Some targets prefer to split a conditional branch like:
7170 /// \code
7171 /// %0 = icmp ne i32 %a, 0
7172 /// %1 = icmp ne i32 %b, 0
7173 /// %or.cond = or i1 %0, %1
7174 /// br i1 %or.cond, label %TrueBB, label %FalseBB
7175 /// \endcode
7176 /// into multiple branch instructions like:
7177 /// \code
7178 /// bb1:
7179 /// %0 = icmp ne i32 %a, 0
7180 /// br i1 %0, label %TrueBB, label %bb2
7181 /// bb2:
7182 /// %1 = icmp ne i32 %b, 0
7183 /// br i1 %1, label %TrueBB, label %FalseBB
7184 /// \endcode
7185 /// This usually allows instruction selection to do even further optimizations
7186 /// and combine the compare with the branch instruction. Currently this is
7187 /// applied for targets which have "cheap" jump instructions.
7188 ///
7189 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
7190 ///
splitBranchCondition(Function & F,bool & ModifiedDT)7191 bool CodeGenPrepare::splitBranchCondition(Function &F, bool &ModifiedDT) {
7192 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
7193 return false;
7194
7195 bool MadeChange = false;
7196 for (auto &BB : F) {
7197 // Does this BB end with the following?
7198 // %cond1 = icmp|fcmp|binary instruction ...
7199 // %cond2 = icmp|fcmp|binary instruction ...
7200 // %cond.or = or|and i1 %cond1, cond2
7201 // br i1 %cond.or label %dest1, label %dest2"
7202 BinaryOperator *LogicOp;
7203 BasicBlock *TBB, *FBB;
7204 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
7205 continue;
7206
7207 auto *Br1 = cast<BranchInst>(BB.getTerminator());
7208 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
7209 continue;
7210
7211 unsigned Opc;
7212 Value *Cond1, *Cond2;
7213 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
7214 m_OneUse(m_Value(Cond2)))))
7215 Opc = Instruction::And;
7216 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
7217 m_OneUse(m_Value(Cond2)))))
7218 Opc = Instruction::Or;
7219 else
7220 continue;
7221
7222 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
7223 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
7224 continue;
7225
7226 LLVM_DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
7227
7228 // Create a new BB.
7229 auto TmpBB =
7230 BasicBlock::Create(BB.getContext(), BB.getName() + ".cond.split",
7231 BB.getParent(), BB.getNextNode());
7232
7233 // Update original basic block by using the first condition directly by the
7234 // branch instruction and removing the no longer needed and/or instruction.
7235 Br1->setCondition(Cond1);
7236 LogicOp->eraseFromParent();
7237
7238 // Depending on the condition we have to either replace the true or the
7239 // false successor of the original branch instruction.
7240 if (Opc == Instruction::And)
7241 Br1->setSuccessor(0, TmpBB);
7242 else
7243 Br1->setSuccessor(1, TmpBB);
7244
7245 // Fill in the new basic block.
7246 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
7247 if (auto *I = dyn_cast<Instruction>(Cond2)) {
7248 I->removeFromParent();
7249 I->insertBefore(Br2);
7250 }
7251
7252 // Update PHI nodes in both successors. The original BB needs to be
7253 // replaced in one successor's PHI nodes, because the branch comes now from
7254 // the newly generated BB (NewBB). In the other successor we need to add one
7255 // incoming edge to the PHI nodes, because both branch instructions target
7256 // now the same successor. Depending on the original branch condition
7257 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
7258 // we perform the correct update for the PHI nodes.
7259 // This doesn't change the successor order of the just created branch
7260 // instruction (or any other instruction).
7261 if (Opc == Instruction::Or)
7262 std::swap(TBB, FBB);
7263
7264 // Replace the old BB with the new BB.
7265 TBB->replacePhiUsesWith(&BB, TmpBB);
7266
7267 // Add another incoming edge form the new BB.
7268 for (PHINode &PN : FBB->phis()) {
7269 auto *Val = PN.getIncomingValueForBlock(&BB);
7270 PN.addIncoming(Val, TmpBB);
7271 }
7272
7273 // Update the branch weights (from SelectionDAGBuilder::
7274 // FindMergedConditions).
7275 if (Opc == Instruction::Or) {
7276 // Codegen X | Y as:
7277 // BB1:
7278 // jmp_if_X TBB
7279 // jmp TmpBB
7280 // TmpBB:
7281 // jmp_if_Y TBB
7282 // jmp FBB
7283 //
7284
7285 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
7286 // The requirement is that
7287 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
7288 // = TrueProb for original BB.
7289 // Assuming the original weights are A and B, one choice is to set BB1's
7290 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
7291 // assumes that
7292 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
7293 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
7294 // TmpBB, but the math is more complicated.
7295 uint64_t TrueWeight, FalseWeight;
7296 if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) {
7297 uint64_t NewTrueWeight = TrueWeight;
7298 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
7299 scaleWeights(NewTrueWeight, NewFalseWeight);
7300 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
7301 .createBranchWeights(TrueWeight, FalseWeight));
7302
7303 NewTrueWeight = TrueWeight;
7304 NewFalseWeight = 2 * FalseWeight;
7305 scaleWeights(NewTrueWeight, NewFalseWeight);
7306 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
7307 .createBranchWeights(TrueWeight, FalseWeight));
7308 }
7309 } else {
7310 // Codegen X & Y as:
7311 // BB1:
7312 // jmp_if_X TmpBB
7313 // jmp FBB
7314 // TmpBB:
7315 // jmp_if_Y TBB
7316 // jmp FBB
7317 //
7318 // This requires creation of TmpBB after CurBB.
7319
7320 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
7321 // The requirement is that
7322 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
7323 // = FalseProb for original BB.
7324 // Assuming the original weights are A and B, one choice is to set BB1's
7325 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
7326 // assumes that
7327 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
7328 uint64_t TrueWeight, FalseWeight;
7329 if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) {
7330 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
7331 uint64_t NewFalseWeight = FalseWeight;
7332 scaleWeights(NewTrueWeight, NewFalseWeight);
7333 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
7334 .createBranchWeights(TrueWeight, FalseWeight));
7335
7336 NewTrueWeight = 2 * TrueWeight;
7337 NewFalseWeight = FalseWeight;
7338 scaleWeights(NewTrueWeight, NewFalseWeight);
7339 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
7340 .createBranchWeights(TrueWeight, FalseWeight));
7341 }
7342 }
7343
7344 ModifiedDT = true;
7345 MadeChange = true;
7346
7347 LLVM_DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
7348 TmpBB->dump());
7349 }
7350 return MadeChange;
7351 }
7352