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/Analysis/ValueTracking.h"
34 #include "llvm/Analysis/VectorUtils.h"
35 #include "llvm/CodeGen/Analysis.h"
36 #include "llvm/CodeGen/ISDOpcodes.h"
37 #include "llvm/CodeGen/SelectionDAGNodes.h"
38 #include "llvm/CodeGen/TargetLowering.h"
39 #include "llvm/CodeGen/TargetPassConfig.h"
40 #include "llvm/CodeGen/TargetSubtargetInfo.h"
41 #include "llvm/CodeGen/ValueTypes.h"
42 #include "llvm/Config/llvm-config.h"
43 #include "llvm/IR/Argument.h"
44 #include "llvm/IR/Attributes.h"
45 #include "llvm/IR/BasicBlock.h"
46 #include "llvm/IR/Constant.h"
47 #include "llvm/IR/Constants.h"
48 #include "llvm/IR/DataLayout.h"
49 #include "llvm/IR/DerivedTypes.h"
50 #include "llvm/IR/Dominators.h"
51 #include "llvm/IR/Function.h"
52 #include "llvm/IR/GetElementPtrTypeIterator.h"
53 #include "llvm/IR/GlobalValue.h"
54 #include "llvm/IR/GlobalVariable.h"
55 #include "llvm/IR/IRBuilder.h"
56 #include "llvm/IR/InlineAsm.h"
57 #include "llvm/IR/InstrTypes.h"
58 #include "llvm/IR/Instruction.h"
59 #include "llvm/IR/Instructions.h"
60 #include "llvm/IR/IntrinsicInst.h"
61 #include "llvm/IR/Intrinsics.h"
62 #include "llvm/IR/IntrinsicsAArch64.h"
63 #include "llvm/IR/LLVMContext.h"
64 #include "llvm/IR/MDBuilder.h"
65 #include "llvm/IR/Module.h"
66 #include "llvm/IR/Operator.h"
67 #include "llvm/IR/PatternMatch.h"
68 #include "llvm/IR/Statepoint.h"
69 #include "llvm/IR/Type.h"
70 #include "llvm/IR/Use.h"
71 #include "llvm/IR/User.h"
72 #include "llvm/IR/Value.h"
73 #include "llvm/IR/ValueHandle.h"
74 #include "llvm/IR/ValueMap.h"
75 #include "llvm/InitializePasses.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/Local.h"
92 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
93 #include "llvm/Transforms/Utils/SizeOpts.h"
94 #include <algorithm>
95 #include <cassert>
96 #include <cstdint>
97 #include <iterator>
98 #include <limits>
99 #include <memory>
100 #include <utility>
101 #include <vector>
102
103 using namespace llvm;
104 using namespace llvm::PatternMatch;
105
106 #define DEBUG_TYPE "codegenprepare"
107
108 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
109 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
110 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
111 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
112 "sunken Cmps");
113 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
114 "of sunken Casts");
115 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
116 "computations were sunk");
117 STATISTIC(NumMemoryInstsPhiCreated,
118 "Number of phis created when address "
119 "computations were sunk to memory instructions");
120 STATISTIC(NumMemoryInstsSelectCreated,
121 "Number of select created when address "
122 "computations were sunk to memory instructions");
123 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
124 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
125 STATISTIC(NumAndsAdded,
126 "Number of and mask instructions added to form ext loads");
127 STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
128 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
129 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
130 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
131 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
132
133 static cl::opt<bool> DisableBranchOpts(
134 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
135 cl::desc("Disable branch optimizations in CodeGenPrepare"));
136
137 static cl::opt<bool>
138 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
139 cl::desc("Disable GC optimizations in CodeGenPrepare"));
140
141 static cl::opt<bool> DisableSelectToBranch(
142 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
143 cl::desc("Disable select to branch conversion."));
144
145 static cl::opt<bool> AddrSinkUsingGEPs(
146 "addr-sink-using-gep", cl::Hidden, cl::init(true),
147 cl::desc("Address sinking in CGP using GEPs."));
148
149 static cl::opt<bool> EnableAndCmpSinking(
150 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
151 cl::desc("Enable sinkinig and/cmp into branches."));
152
153 static cl::opt<bool> DisableStoreExtract(
154 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
155 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
156
157 static cl::opt<bool> StressStoreExtract(
158 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
159 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
160
161 static cl::opt<bool> DisableExtLdPromotion(
162 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
163 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
164 "CodeGenPrepare"));
165
166 static cl::opt<bool> StressExtLdPromotion(
167 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
168 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
169 "optimization in CodeGenPrepare"));
170
171 static cl::opt<bool> DisablePreheaderProtect(
172 "disable-preheader-prot", cl::Hidden, cl::init(false),
173 cl::desc("Disable protection against removing loop preheaders"));
174
175 static cl::opt<bool> ProfileGuidedSectionPrefix(
176 "profile-guided-section-prefix", cl::Hidden, cl::init(true), cl::ZeroOrMore,
177 cl::desc("Use profile info to add section prefix for hot/cold functions"));
178
179 static cl::opt<bool> ProfileUnknownInSpecialSection(
180 "profile-unknown-in-special-section", cl::Hidden, cl::init(false),
181 cl::ZeroOrMore,
182 cl::desc("In profiling mode like sampleFDO, if a function doesn't have "
183 "profile, we cannot tell the function is cold for sure because "
184 "it may be a function newly added without ever being sampled. "
185 "With the flag enabled, compiler can put such profile unknown "
186 "functions into a special section, so runtime system can choose "
187 "to handle it in a different way than .text section, to save "
188 "RAM for example. "));
189
190 static cl::opt<unsigned> FreqRatioToSkipMerge(
191 "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2),
192 cl::desc("Skip merging empty blocks if (frequency of empty block) / "
193 "(frequency of destination block) is greater than this ratio"));
194
195 static cl::opt<bool> ForceSplitStore(
196 "force-split-store", cl::Hidden, cl::init(false),
197 cl::desc("Force store splitting no matter what the target query says."));
198
199 static cl::opt<bool>
200 EnableTypePromotionMerge("cgp-type-promotion-merge", cl::Hidden,
201 cl::desc("Enable merging of redundant sexts when one is dominating"
202 " the other."), cl::init(true));
203
204 static cl::opt<bool> DisableComplexAddrModes(
205 "disable-complex-addr-modes", cl::Hidden, cl::init(false),
206 cl::desc("Disables combining addressing modes with different parts "
207 "in optimizeMemoryInst."));
208
209 static cl::opt<bool>
210 AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(false),
211 cl::desc("Allow creation of Phis in Address sinking."));
212
213 static cl::opt<bool>
214 AddrSinkNewSelects("addr-sink-new-select", cl::Hidden, cl::init(true),
215 cl::desc("Allow creation of selects in Address sinking."));
216
217 static cl::opt<bool> AddrSinkCombineBaseReg(
218 "addr-sink-combine-base-reg", cl::Hidden, cl::init(true),
219 cl::desc("Allow combining of BaseReg field in Address sinking."));
220
221 static cl::opt<bool> AddrSinkCombineBaseGV(
222 "addr-sink-combine-base-gv", cl::Hidden, cl::init(true),
223 cl::desc("Allow combining of BaseGV field in Address sinking."));
224
225 static cl::opt<bool> AddrSinkCombineBaseOffs(
226 "addr-sink-combine-base-offs", cl::Hidden, cl::init(true),
227 cl::desc("Allow combining of BaseOffs field in Address sinking."));
228
229 static cl::opt<bool> AddrSinkCombineScaledReg(
230 "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(true),
231 cl::desc("Allow combining of ScaledReg field in Address sinking."));
232
233 static cl::opt<bool>
234 EnableGEPOffsetSplit("cgp-split-large-offset-gep", cl::Hidden,
235 cl::init(true),
236 cl::desc("Enable splitting large offset of GEP."));
237
238 static cl::opt<bool> EnableICMP_EQToICMP_ST(
239 "cgp-icmp-eq2icmp-st", cl::Hidden, cl::init(false),
240 cl::desc("Enable ICMP_EQ to ICMP_S(L|G)T conversion."));
241
242 static cl::opt<bool>
243 VerifyBFIUpdates("cgp-verify-bfi-updates", cl::Hidden, cl::init(false),
244 cl::desc("Enable BFI update verification for "
245 "CodeGenPrepare."));
246
247 static cl::opt<bool> OptimizePhiTypes(
248 "cgp-optimize-phi-types", cl::Hidden, cl::init(false),
249 cl::desc("Enable converting phi types in CodeGenPrepare"));
250
251 namespace {
252
253 enum ExtType {
254 ZeroExtension, // Zero extension has been seen.
255 SignExtension, // Sign extension has been seen.
256 BothExtension // This extension type is used if we saw sext after
257 // ZeroExtension had been set, or if we saw zext after
258 // SignExtension had been set. It makes the type
259 // information of a promoted instruction invalid.
260 };
261
262 using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
263 using TypeIsSExt = PointerIntPair<Type *, 2, ExtType>;
264 using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
265 using SExts = SmallVector<Instruction *, 16>;
266 using ValueToSExts = DenseMap<Value *, SExts>;
267
268 class TypePromotionTransaction;
269
270 class CodeGenPrepare : public FunctionPass {
271 const TargetMachine *TM = nullptr;
272 const TargetSubtargetInfo *SubtargetInfo;
273 const TargetLowering *TLI = nullptr;
274 const TargetRegisterInfo *TRI;
275 const TargetTransformInfo *TTI = nullptr;
276 const TargetLibraryInfo *TLInfo;
277 const LoopInfo *LI;
278 std::unique_ptr<BlockFrequencyInfo> BFI;
279 std::unique_ptr<BranchProbabilityInfo> BPI;
280 ProfileSummaryInfo *PSI;
281
282 /// As we scan instructions optimizing them, this is the next instruction
283 /// to optimize. Transforms that can invalidate this should update it.
284 BasicBlock::iterator CurInstIterator;
285
286 /// Keeps track of non-local addresses that have been sunk into a block.
287 /// This allows us to avoid inserting duplicate code for blocks with
288 /// multiple load/stores of the same address. The usage of WeakTrackingVH
289 /// enables SunkAddrs to be treated as a cache whose entries can be
290 /// invalidated if a sunken address computation has been erased.
291 ValueMap<Value*, WeakTrackingVH> SunkAddrs;
292
293 /// Keeps track of all instructions inserted for the current function.
294 SetOfInstrs InsertedInsts;
295
296 /// Keeps track of the type of the related instruction before their
297 /// promotion for the current function.
298 InstrToOrigTy PromotedInsts;
299
300 /// Keep track of instructions removed during promotion.
301 SetOfInstrs RemovedInsts;
302
303 /// Keep track of sext chains based on their initial value.
304 DenseMap<Value *, Instruction *> SeenChainsForSExt;
305
306 /// Keep track of GEPs accessing the same data structures such as structs or
307 /// arrays that are candidates to be split later because of their large
308 /// size.
309 MapVector<
310 AssertingVH<Value>,
311 SmallVector<std::pair<AssertingVH<GetElementPtrInst>, int64_t>, 32>>
312 LargeOffsetGEPMap;
313
314 /// Keep track of new GEP base after splitting the GEPs having large offset.
315 SmallSet<AssertingVH<Value>, 2> NewGEPBases;
316
317 /// Map serial numbers to Large offset GEPs.
318 DenseMap<AssertingVH<GetElementPtrInst>, int> LargeOffsetGEPID;
319
320 /// Keep track of SExt promoted.
321 ValueToSExts ValToSExtendedUses;
322
323 /// True if the function has the OptSize attribute.
324 bool OptSize;
325
326 /// DataLayout for the Function being processed.
327 const DataLayout *DL = nullptr;
328
329 /// Building the dominator tree can be expensive, so we only build it
330 /// lazily and update it when required.
331 std::unique_ptr<DominatorTree> DT;
332
333 public:
334 static char ID; // Pass identification, replacement for typeid
335
CodeGenPrepare()336 CodeGenPrepare() : FunctionPass(ID) {
337 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
338 }
339
340 bool runOnFunction(Function &F) override;
341
getPassName() const342 StringRef getPassName() const override { return "CodeGen Prepare"; }
343
getAnalysisUsage(AnalysisUsage & AU) const344 void getAnalysisUsage(AnalysisUsage &AU) const override {
345 // FIXME: When we can selectively preserve passes, preserve the domtree.
346 AU.addRequired<ProfileSummaryInfoWrapperPass>();
347 AU.addRequired<TargetLibraryInfoWrapperPass>();
348 AU.addRequired<TargetPassConfig>();
349 AU.addRequired<TargetTransformInfoWrapperPass>();
350 AU.addRequired<LoopInfoWrapperPass>();
351 }
352
353 private:
354 template <typename F>
resetIteratorIfInvalidatedWhileCalling(BasicBlock * BB,F f)355 void resetIteratorIfInvalidatedWhileCalling(BasicBlock *BB, F f) {
356 // Substituting can cause recursive simplifications, which can invalidate
357 // our iterator. Use a WeakTrackingVH to hold onto it in case this
358 // happens.
359 Value *CurValue = &*CurInstIterator;
360 WeakTrackingVH IterHandle(CurValue);
361
362 f();
363
364 // If the iterator instruction was recursively deleted, start over at the
365 // start of the block.
366 if (IterHandle != CurValue) {
367 CurInstIterator = BB->begin();
368 SunkAddrs.clear();
369 }
370 }
371
372 // Get the DominatorTree, building if necessary.
getDT(Function & F)373 DominatorTree &getDT(Function &F) {
374 if (!DT)
375 DT = std::make_unique<DominatorTree>(F);
376 return *DT;
377 }
378
379 void removeAllAssertingVHReferences(Value *V);
380 bool eliminateFallThrough(Function &F);
381 bool eliminateMostlyEmptyBlocks(Function &F);
382 BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
383 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
384 void eliminateMostlyEmptyBlock(BasicBlock *BB);
385 bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
386 bool isPreheader);
387 bool makeBitReverse(Instruction &I);
388 bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
389 bool optimizeInst(Instruction *I, bool &ModifiedDT);
390 bool optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
391 Type *AccessTy, unsigned AddrSpace);
392 bool optimizeGatherScatterInst(Instruction *MemoryInst, Value *Ptr);
393 bool optimizeInlineAsmInst(CallInst *CS);
394 bool optimizeCallInst(CallInst *CI, bool &ModifiedDT);
395 bool optimizeExt(Instruction *&I);
396 bool optimizeExtUses(Instruction *I);
397 bool optimizeLoadExt(LoadInst *Load);
398 bool optimizeShiftInst(BinaryOperator *BO);
399 bool optimizeFunnelShift(IntrinsicInst *Fsh);
400 bool optimizeSelectInst(SelectInst *SI);
401 bool optimizeShuffleVectorInst(ShuffleVectorInst *SVI);
402 bool optimizeSwitchInst(SwitchInst *SI);
403 bool optimizeExtractElementInst(Instruction *Inst);
404 bool dupRetToEnableTailCallOpts(BasicBlock *BB, bool &ModifiedDT);
405 bool fixupDbgValue(Instruction *I);
406 bool placeDbgValues(Function &F);
407 bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
408 LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
409 bool tryToPromoteExts(TypePromotionTransaction &TPT,
410 const SmallVectorImpl<Instruction *> &Exts,
411 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
412 unsigned CreatedInstsCost = 0);
413 bool mergeSExts(Function &F);
414 bool splitLargeGEPOffsets();
415 bool optimizePhiType(PHINode *Inst, SmallPtrSetImpl<PHINode *> &Visited,
416 SmallPtrSetImpl<Instruction *> &DeletedInstrs);
417 bool optimizePhiTypes(Function &F);
418 bool performAddressTypePromotion(
419 Instruction *&Inst,
420 bool AllowPromotionWithoutCommonHeader,
421 bool HasPromoted, TypePromotionTransaction &TPT,
422 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
423 bool splitBranchCondition(Function &F, bool &ModifiedDT);
424 bool simplifyOffsetableRelocate(GCStatepointInst &I);
425
426 bool tryToSinkFreeOperands(Instruction *I);
427 bool replaceMathCmpWithIntrinsic(BinaryOperator *BO, Value *Arg0,
428 Value *Arg1, CmpInst *Cmp,
429 Intrinsic::ID IID);
430 bool optimizeCmp(CmpInst *Cmp, bool &ModifiedDT);
431 bool combineToUSubWithOverflow(CmpInst *Cmp, bool &ModifiedDT);
432 bool combineToUAddWithOverflow(CmpInst *Cmp, bool &ModifiedDT);
433 void verifyBFIUpdates(Function &F);
434 };
435
436 } // end anonymous namespace
437
438 char CodeGenPrepare::ID = 0;
439
440 INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,
441 "Optimize for code generation", false, false)
INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)442 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
443 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
444 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
445 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
446 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
447 INITIALIZE_PASS_END(CodeGenPrepare, DEBUG_TYPE,
448 "Optimize for code generation", false, false)
449
450 FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); }
451
runOnFunction(Function & F)452 bool CodeGenPrepare::runOnFunction(Function &F) {
453 if (skipFunction(F))
454 return false;
455
456 DL = &F.getParent()->getDataLayout();
457
458 bool EverMadeChange = false;
459 // Clear per function information.
460 InsertedInsts.clear();
461 PromotedInsts.clear();
462
463 TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>();
464 SubtargetInfo = TM->getSubtargetImpl(F);
465 TLI = SubtargetInfo->getTargetLowering();
466 TRI = SubtargetInfo->getRegisterInfo();
467 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
468 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
469 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
470 BPI.reset(new BranchProbabilityInfo(F, *LI));
471 BFI.reset(new BlockFrequencyInfo(F, *BPI, *LI));
472 PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
473 OptSize = F.hasOptSize();
474 if (ProfileGuidedSectionPrefix) {
475 // The hot attribute overwrites profile count based hotness while profile
476 // counts based hotness overwrite the cold attribute.
477 // This is a conservative behabvior.
478 if (F.hasFnAttribute(Attribute::Hot) ||
479 PSI->isFunctionHotInCallGraph(&F, *BFI))
480 F.setSectionPrefix("hot");
481 // If PSI shows this function is not hot, we will placed the function
482 // into unlikely section if (1) PSI shows this is a cold function, or
483 // (2) the function has a attribute of cold.
484 else if (PSI->isFunctionColdInCallGraph(&F, *BFI) ||
485 F.hasFnAttribute(Attribute::Cold))
486 F.setSectionPrefix("unlikely");
487 else if (ProfileUnknownInSpecialSection && PSI->hasPartialSampleProfile() &&
488 PSI->isFunctionHotnessUnknown(F))
489 F.setSectionPrefix("unknown");
490 }
491
492 /// This optimization identifies DIV instructions that can be
493 /// profitably bypassed and carried out with a shorter, faster divide.
494 if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI->isSlowDivBypassed()) {
495 const DenseMap<unsigned int, unsigned int> &BypassWidths =
496 TLI->getBypassSlowDivWidths();
497 BasicBlock* BB = &*F.begin();
498 while (BB != nullptr) {
499 // bypassSlowDivision may create new BBs, but we don't want to reapply the
500 // optimization to those blocks.
501 BasicBlock* Next = BB->getNextNode();
502 // F.hasOptSize is already checked in the outer if statement.
503 if (!llvm::shouldOptimizeForSize(BB, PSI, BFI.get()))
504 EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
505 BB = Next;
506 }
507 }
508
509 // Eliminate blocks that contain only PHI nodes and an
510 // unconditional branch.
511 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
512
513 bool ModifiedDT = false;
514 if (!DisableBranchOpts)
515 EverMadeChange |= splitBranchCondition(F, ModifiedDT);
516
517 // Split some critical edges where one of the sources is an indirect branch,
518 // to help generate sane code for PHIs involving such edges.
519 EverMadeChange |= SplitIndirectBrCriticalEdges(F);
520
521 bool MadeChange = true;
522 while (MadeChange) {
523 MadeChange = false;
524 DT.reset();
525 for (Function::iterator I = F.begin(); I != F.end(); ) {
526 BasicBlock *BB = &*I++;
527 bool ModifiedDTOnIteration = false;
528 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
529
530 // Restart BB iteration if the dominator tree of the Function was changed
531 if (ModifiedDTOnIteration)
532 break;
533 }
534 if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
535 MadeChange |= mergeSExts(F);
536 if (!LargeOffsetGEPMap.empty())
537 MadeChange |= splitLargeGEPOffsets();
538 MadeChange |= optimizePhiTypes(F);
539
540 if (MadeChange)
541 eliminateFallThrough(F);
542
543 // Really free removed instructions during promotion.
544 for (Instruction *I : RemovedInsts)
545 I->deleteValue();
546
547 EverMadeChange |= MadeChange;
548 SeenChainsForSExt.clear();
549 ValToSExtendedUses.clear();
550 RemovedInsts.clear();
551 LargeOffsetGEPMap.clear();
552 LargeOffsetGEPID.clear();
553 }
554
555 NewGEPBases.clear();
556 SunkAddrs.clear();
557
558 if (!DisableBranchOpts) {
559 MadeChange = false;
560 // Use a set vector to get deterministic iteration order. The order the
561 // blocks are removed may affect whether or not PHI nodes in successors
562 // are removed.
563 SmallSetVector<BasicBlock*, 8> WorkList;
564 for (BasicBlock &BB : F) {
565 SmallVector<BasicBlock *, 2> Successors(successors(&BB));
566 MadeChange |= ConstantFoldTerminator(&BB, true);
567 if (!MadeChange) continue;
568
569 for (SmallVectorImpl<BasicBlock*>::iterator
570 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
571 if (pred_empty(*II))
572 WorkList.insert(*II);
573 }
574
575 // Delete the dead blocks and any of their dead successors.
576 MadeChange |= !WorkList.empty();
577 while (!WorkList.empty()) {
578 BasicBlock *BB = WorkList.pop_back_val();
579 SmallVector<BasicBlock*, 2> Successors(successors(BB));
580
581 DeleteDeadBlock(BB);
582
583 for (SmallVectorImpl<BasicBlock*>::iterator
584 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
585 if (pred_empty(*II))
586 WorkList.insert(*II);
587 }
588
589 // Merge pairs of basic blocks with unconditional branches, connected by
590 // a single edge.
591 if (EverMadeChange || MadeChange)
592 MadeChange |= eliminateFallThrough(F);
593
594 EverMadeChange |= MadeChange;
595 }
596
597 if (!DisableGCOpts) {
598 SmallVector<GCStatepointInst *, 2> Statepoints;
599 for (BasicBlock &BB : F)
600 for (Instruction &I : BB)
601 if (auto *SP = dyn_cast<GCStatepointInst>(&I))
602 Statepoints.push_back(SP);
603 for (auto &I : Statepoints)
604 EverMadeChange |= simplifyOffsetableRelocate(*I);
605 }
606
607 // Do this last to clean up use-before-def scenarios introduced by other
608 // preparatory transforms.
609 EverMadeChange |= placeDbgValues(F);
610
611 #ifndef NDEBUG
612 if (VerifyBFIUpdates)
613 verifyBFIUpdates(F);
614 #endif
615
616 return EverMadeChange;
617 }
618
619 /// An instruction is about to be deleted, so remove all references to it in our
620 /// GEP-tracking data strcutures.
removeAllAssertingVHReferences(Value * V)621 void CodeGenPrepare::removeAllAssertingVHReferences(Value *V) {
622 LargeOffsetGEPMap.erase(V);
623 NewGEPBases.erase(V);
624
625 auto GEP = dyn_cast<GetElementPtrInst>(V);
626 if (!GEP)
627 return;
628
629 LargeOffsetGEPID.erase(GEP);
630
631 auto VecI = LargeOffsetGEPMap.find(GEP->getPointerOperand());
632 if (VecI == LargeOffsetGEPMap.end())
633 return;
634
635 auto &GEPVector = VecI->second;
636 const auto &I =
637 llvm::find_if(GEPVector, [=](auto &Elt) { return Elt.first == GEP; });
638 if (I == GEPVector.end())
639 return;
640
641 GEPVector.erase(I);
642 if (GEPVector.empty())
643 LargeOffsetGEPMap.erase(VecI);
644 }
645
646 // Verify BFI has been updated correctly by recomputing BFI and comparing them.
verifyBFIUpdates(Function & F)647 void LLVM_ATTRIBUTE_UNUSED CodeGenPrepare::verifyBFIUpdates(Function &F) {
648 DominatorTree NewDT(F);
649 LoopInfo NewLI(NewDT);
650 BranchProbabilityInfo NewBPI(F, NewLI, TLInfo);
651 BlockFrequencyInfo NewBFI(F, NewBPI, NewLI);
652 NewBFI.verifyMatch(*BFI);
653 }
654
655 /// Merge basic blocks which are connected by a single edge, where one of the
656 /// basic blocks has a single successor pointing to the other basic block,
657 /// which has a single predecessor.
eliminateFallThrough(Function & F)658 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
659 bool Changed = false;
660 // Scan all of the blocks in the function, except for the entry block.
661 // Use a temporary array to avoid iterator being invalidated when
662 // deleting blocks.
663 SmallVector<WeakTrackingVH, 16> Blocks;
664 for (auto &Block : llvm::drop_begin(F))
665 Blocks.push_back(&Block);
666
667 SmallSet<WeakTrackingVH, 16> Preds;
668 for (auto &Block : Blocks) {
669 auto *BB = cast_or_null<BasicBlock>(Block);
670 if (!BB)
671 continue;
672 // If the destination block has a single pred, then this is a trivial
673 // edge, just collapse it.
674 BasicBlock *SinglePred = BB->getSinglePredecessor();
675
676 // Don't merge if BB's address is taken.
677 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
678
679 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
680 if (Term && !Term->isConditional()) {
681 Changed = true;
682 LLVM_DEBUG(dbgs() << "To merge:\n" << *BB << "\n\n\n");
683
684 // Merge BB into SinglePred and delete it.
685 MergeBlockIntoPredecessor(BB);
686 Preds.insert(SinglePred);
687 }
688 }
689
690 // (Repeatedly) merging blocks into their predecessors can create redundant
691 // debug intrinsics.
692 for (auto &Pred : Preds)
693 if (auto *BB = cast_or_null<BasicBlock>(Pred))
694 RemoveRedundantDbgInstrs(BB);
695
696 return Changed;
697 }
698
699 /// Find a destination block from BB if BB is mergeable empty block.
findDestBlockOfMergeableEmptyBlock(BasicBlock * BB)700 BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
701 // If this block doesn't end with an uncond branch, ignore it.
702 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
703 if (!BI || !BI->isUnconditional())
704 return nullptr;
705
706 // If the instruction before the branch (skipping debug info) isn't a phi
707 // node, then other stuff is happening here.
708 BasicBlock::iterator BBI = BI->getIterator();
709 if (BBI != BB->begin()) {
710 --BBI;
711 while (isa<DbgInfoIntrinsic>(BBI)) {
712 if (BBI == BB->begin())
713 break;
714 --BBI;
715 }
716 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
717 return nullptr;
718 }
719
720 // Do not break infinite loops.
721 BasicBlock *DestBB = BI->getSuccessor(0);
722 if (DestBB == BB)
723 return nullptr;
724
725 if (!canMergeBlocks(BB, DestBB))
726 DestBB = nullptr;
727
728 return DestBB;
729 }
730
731 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
732 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
733 /// edges in ways that are non-optimal for isel. Start by eliminating these
734 /// blocks so we can split them the way we want them.
eliminateMostlyEmptyBlocks(Function & F)735 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
736 SmallPtrSet<BasicBlock *, 16> Preheaders;
737 SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
738 while (!LoopList.empty()) {
739 Loop *L = LoopList.pop_back_val();
740 llvm::append_range(LoopList, *L);
741 if (BasicBlock *Preheader = L->getLoopPreheader())
742 Preheaders.insert(Preheader);
743 }
744
745 bool MadeChange = false;
746 // Copy blocks into a temporary array to avoid iterator invalidation issues
747 // as we remove them.
748 // Note that this intentionally skips the entry block.
749 SmallVector<WeakTrackingVH, 16> Blocks;
750 for (auto &Block : llvm::drop_begin(F))
751 Blocks.push_back(&Block);
752
753 for (auto &Block : Blocks) {
754 BasicBlock *BB = cast_or_null<BasicBlock>(Block);
755 if (!BB)
756 continue;
757 BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
758 if (!DestBB ||
759 !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB)))
760 continue;
761
762 eliminateMostlyEmptyBlock(BB);
763 MadeChange = true;
764 }
765 return MadeChange;
766 }
767
isMergingEmptyBlockProfitable(BasicBlock * BB,BasicBlock * DestBB,bool isPreheader)768 bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
769 BasicBlock *DestBB,
770 bool isPreheader) {
771 // Do not delete loop preheaders if doing so would create a critical edge.
772 // Loop preheaders can be good locations to spill registers. If the
773 // preheader is deleted and we create a critical edge, registers may be
774 // spilled in the loop body instead.
775 if (!DisablePreheaderProtect && isPreheader &&
776 !(BB->getSinglePredecessor() &&
777 BB->getSinglePredecessor()->getSingleSuccessor()))
778 return false;
779
780 // Skip merging if the block's successor is also a successor to any callbr
781 // that leads to this block.
782 // FIXME: Is this really needed? Is this a correctness issue?
783 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
784 if (auto *CBI = dyn_cast<CallBrInst>((*PI)->getTerminator()))
785 for (unsigned i = 0, e = CBI->getNumSuccessors(); i != e; ++i)
786 if (DestBB == CBI->getSuccessor(i))
787 return false;
788 }
789
790 // Try to skip merging if the unique predecessor of BB is terminated by a
791 // switch or indirect branch instruction, and BB is used as an incoming block
792 // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
793 // add COPY instructions in the predecessor of BB instead of BB (if it is not
794 // merged). Note that the critical edge created by merging such blocks wont be
795 // split in MachineSink because the jump table is not analyzable. By keeping
796 // such empty block (BB), ISel will place COPY instructions in BB, not in the
797 // predecessor of BB.
798 BasicBlock *Pred = BB->getUniquePredecessor();
799 if (!Pred ||
800 !(isa<SwitchInst>(Pred->getTerminator()) ||
801 isa<IndirectBrInst>(Pred->getTerminator())))
802 return true;
803
804 if (BB->getTerminator() != BB->getFirstNonPHIOrDbg())
805 return true;
806
807 // We use a simple cost heuristic which determine skipping merging is
808 // profitable if the cost of skipping merging is less than the cost of
809 // merging : Cost(skipping merging) < Cost(merging BB), where the
810 // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
811 // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
812 // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
813 // Freq(Pred) / Freq(BB) > 2.
814 // Note that if there are multiple empty blocks sharing the same incoming
815 // value for the PHIs in the DestBB, we consider them together. In such
816 // case, Cost(merging BB) will be the sum of their frequencies.
817
818 if (!isa<PHINode>(DestBB->begin()))
819 return true;
820
821 SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
822
823 // Find all other incoming blocks from which incoming values of all PHIs in
824 // DestBB are the same as the ones from BB.
825 for (pred_iterator PI = pred_begin(DestBB), E = pred_end(DestBB); PI != E;
826 ++PI) {
827 BasicBlock *DestBBPred = *PI;
828 if (DestBBPred == BB)
829 continue;
830
831 if (llvm::all_of(DestBB->phis(), [&](const PHINode &DestPN) {
832 return DestPN.getIncomingValueForBlock(BB) ==
833 DestPN.getIncomingValueForBlock(DestBBPred);
834 }))
835 SameIncomingValueBBs.insert(DestBBPred);
836 }
837
838 // See if all BB's incoming values are same as the value from Pred. In this
839 // case, no reason to skip merging because COPYs are expected to be place in
840 // Pred already.
841 if (SameIncomingValueBBs.count(Pred))
842 return true;
843
844 BlockFrequency PredFreq = BFI->getBlockFreq(Pred);
845 BlockFrequency BBFreq = BFI->getBlockFreq(BB);
846
847 for (auto *SameValueBB : SameIncomingValueBBs)
848 if (SameValueBB->getUniquePredecessor() == Pred &&
849 DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB))
850 BBFreq += BFI->getBlockFreq(SameValueBB);
851
852 return PredFreq.getFrequency() <=
853 BBFreq.getFrequency() * FreqRatioToSkipMerge;
854 }
855
856 /// Return true if we can merge BB into DestBB if there is a single
857 /// unconditional branch between them, and BB contains no other non-phi
858 /// instructions.
canMergeBlocks(const BasicBlock * BB,const BasicBlock * DestBB) const859 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
860 const BasicBlock *DestBB) const {
861 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
862 // the successor. If there are more complex condition (e.g. preheaders),
863 // don't mess around with them.
864 for (const PHINode &PN : BB->phis()) {
865 for (const User *U : PN.users()) {
866 const Instruction *UI = cast<Instruction>(U);
867 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
868 return false;
869 // If User is inside DestBB block and it is a PHINode then check
870 // incoming value. If incoming value is not from BB then this is
871 // a complex condition (e.g. preheaders) we want to avoid here.
872 if (UI->getParent() == DestBB) {
873 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
874 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
875 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
876 if (Insn && Insn->getParent() == BB &&
877 Insn->getParent() != UPN->getIncomingBlock(I))
878 return false;
879 }
880 }
881 }
882 }
883
884 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
885 // and DestBB may have conflicting incoming values for the block. If so, we
886 // can't merge the block.
887 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
888 if (!DestBBPN) return true; // no conflict.
889
890 // Collect the preds of BB.
891 SmallPtrSet<const BasicBlock*, 16> BBPreds;
892 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
893 // It is faster to get preds from a PHI than with pred_iterator.
894 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
895 BBPreds.insert(BBPN->getIncomingBlock(i));
896 } else {
897 BBPreds.insert(pred_begin(BB), pred_end(BB));
898 }
899
900 // Walk the preds of DestBB.
901 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
902 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
903 if (BBPreds.count(Pred)) { // Common predecessor?
904 for (const PHINode &PN : DestBB->phis()) {
905 const Value *V1 = PN.getIncomingValueForBlock(Pred);
906 const Value *V2 = PN.getIncomingValueForBlock(BB);
907
908 // If V2 is a phi node in BB, look up what the mapped value will be.
909 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
910 if (V2PN->getParent() == BB)
911 V2 = V2PN->getIncomingValueForBlock(Pred);
912
913 // If there is a conflict, bail out.
914 if (V1 != V2) return false;
915 }
916 }
917 }
918
919 return true;
920 }
921
922 /// Eliminate a basic block that has only phi's and an unconditional branch in
923 /// it.
eliminateMostlyEmptyBlock(BasicBlock * BB)924 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
925 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
926 BasicBlock *DestBB = BI->getSuccessor(0);
927
928 LLVM_DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"
929 << *BB << *DestBB);
930
931 // If the destination block has a single pred, then this is a trivial edge,
932 // just collapse it.
933 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
934 if (SinglePred != DestBB) {
935 assert(SinglePred == BB &&
936 "Single predecessor not the same as predecessor");
937 // Merge DestBB into SinglePred/BB and delete it.
938 MergeBlockIntoPredecessor(DestBB);
939 // Note: BB(=SinglePred) will not be deleted on this path.
940 // DestBB(=its single successor) is the one that was deleted.
941 LLVM_DEBUG(dbgs() << "AFTER:\n" << *SinglePred << "\n\n\n");
942 return;
943 }
944 }
945
946 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
947 // to handle the new incoming edges it is about to have.
948 for (PHINode &PN : DestBB->phis()) {
949 // Remove the incoming value for BB, and remember it.
950 Value *InVal = PN.removeIncomingValue(BB, false);
951
952 // Two options: either the InVal is a phi node defined in BB or it is some
953 // value that dominates BB.
954 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
955 if (InValPhi && InValPhi->getParent() == BB) {
956 // Add all of the input values of the input PHI as inputs of this phi.
957 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
958 PN.addIncoming(InValPhi->getIncomingValue(i),
959 InValPhi->getIncomingBlock(i));
960 } else {
961 // Otherwise, add one instance of the dominating value for each edge that
962 // we will be adding.
963 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
964 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
965 PN.addIncoming(InVal, BBPN->getIncomingBlock(i));
966 } else {
967 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
968 PN.addIncoming(InVal, *PI);
969 }
970 }
971 }
972
973 // The PHIs are now updated, change everything that refers to BB to use
974 // DestBB and remove BB.
975 BB->replaceAllUsesWith(DestBB);
976 BB->eraseFromParent();
977 ++NumBlocksElim;
978
979 LLVM_DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
980 }
981
982 // Computes a map of base pointer relocation instructions to corresponding
983 // derived pointer relocation instructions given a vector of all relocate calls
computeBaseDerivedRelocateMap(const SmallVectorImpl<GCRelocateInst * > & AllRelocateCalls,DenseMap<GCRelocateInst *,SmallVector<GCRelocateInst *,2>> & RelocateInstMap)984 static void computeBaseDerivedRelocateMap(
985 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
986 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
987 &RelocateInstMap) {
988 // Collect information in two maps: one primarily for locating the base object
989 // while filling the second map; the second map is the final structure holding
990 // a mapping between Base and corresponding Derived relocate calls
991 DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
992 for (auto *ThisRelocate : AllRelocateCalls) {
993 auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
994 ThisRelocate->getDerivedPtrIndex());
995 RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
996 }
997 for (auto &Item : RelocateIdxMap) {
998 std::pair<unsigned, unsigned> Key = Item.first;
999 if (Key.first == Key.second)
1000 // Base relocation: nothing to insert
1001 continue;
1002
1003 GCRelocateInst *I = Item.second;
1004 auto BaseKey = std::make_pair(Key.first, Key.first);
1005
1006 // We're iterating over RelocateIdxMap so we cannot modify it.
1007 auto MaybeBase = RelocateIdxMap.find(BaseKey);
1008 if (MaybeBase == RelocateIdxMap.end())
1009 // TODO: We might want to insert a new base object relocate and gep off
1010 // that, if there are enough derived object relocates.
1011 continue;
1012
1013 RelocateInstMap[MaybeBase->second].push_back(I);
1014 }
1015 }
1016
1017 // Accepts a GEP and extracts the operands into a vector provided they're all
1018 // small integer constants
getGEPSmallConstantIntOffsetV(GetElementPtrInst * GEP,SmallVectorImpl<Value * > & OffsetV)1019 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
1020 SmallVectorImpl<Value *> &OffsetV) {
1021 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
1022 // Only accept small constant integer operands
1023 auto *Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
1024 if (!Op || Op->getZExtValue() > 20)
1025 return false;
1026 }
1027
1028 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
1029 OffsetV.push_back(GEP->getOperand(i));
1030 return true;
1031 }
1032
1033 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
1034 // replace, computes a replacement, and affects it.
1035 static bool
simplifyRelocatesOffABase(GCRelocateInst * RelocatedBase,const SmallVectorImpl<GCRelocateInst * > & Targets)1036 simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
1037 const SmallVectorImpl<GCRelocateInst *> &Targets) {
1038 bool MadeChange = false;
1039 // We must ensure the relocation of derived pointer is defined after
1040 // relocation of base pointer. If we find a relocation corresponding to base
1041 // defined earlier than relocation of base then we move relocation of base
1042 // right before found relocation. We consider only relocation in the same
1043 // basic block as relocation of base. Relocations from other basic block will
1044 // be skipped by optimization and we do not care about them.
1045 for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
1046 &*R != RelocatedBase; ++R)
1047 if (auto *RI = dyn_cast<GCRelocateInst>(R))
1048 if (RI->getStatepoint() == RelocatedBase->getStatepoint())
1049 if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
1050 RelocatedBase->moveBefore(RI);
1051 break;
1052 }
1053
1054 for (GCRelocateInst *ToReplace : Targets) {
1055 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
1056 "Not relocating a derived object of the original base object");
1057 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
1058 // A duplicate relocate call. TODO: coalesce duplicates.
1059 continue;
1060 }
1061
1062 if (RelocatedBase->getParent() != ToReplace->getParent()) {
1063 // Base and derived relocates are in different basic blocks.
1064 // In this case transform is only valid when base dominates derived
1065 // relocate. However it would be too expensive to check dominance
1066 // for each such relocate, so we skip the whole transformation.
1067 continue;
1068 }
1069
1070 Value *Base = ToReplace->getBasePtr();
1071 auto *Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
1072 if (!Derived || Derived->getPointerOperand() != Base)
1073 continue;
1074
1075 SmallVector<Value *, 2> OffsetV;
1076 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
1077 continue;
1078
1079 // Create a Builder and replace the target callsite with a gep
1080 assert(RelocatedBase->getNextNode() &&
1081 "Should always have one since it's not a terminator");
1082
1083 // Insert after RelocatedBase
1084 IRBuilder<> Builder(RelocatedBase->getNextNode());
1085 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1086
1087 // If gc_relocate does not match the actual type, cast it to the right type.
1088 // In theory, there must be a bitcast after gc_relocate if the type does not
1089 // match, and we should reuse it to get the derived pointer. But it could be
1090 // cases like this:
1091 // bb1:
1092 // ...
1093 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1094 // br label %merge
1095 //
1096 // bb2:
1097 // ...
1098 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1099 // br label %merge
1100 //
1101 // merge:
1102 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
1103 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
1104 //
1105 // In this case, we can not find the bitcast any more. So we insert a new bitcast
1106 // no matter there is already one or not. In this way, we can handle all cases, and
1107 // the extra bitcast should be optimized away in later passes.
1108 Value *ActualRelocatedBase = RelocatedBase;
1109 if (RelocatedBase->getType() != Base->getType()) {
1110 ActualRelocatedBase =
1111 Builder.CreateBitCast(RelocatedBase, Base->getType());
1112 }
1113 Value *Replacement = Builder.CreateGEP(
1114 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
1115 Replacement->takeName(ToReplace);
1116 // If the newly generated derived pointer's type does not match the original derived
1117 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
1118 Value *ActualReplacement = Replacement;
1119 if (Replacement->getType() != ToReplace->getType()) {
1120 ActualReplacement =
1121 Builder.CreateBitCast(Replacement, ToReplace->getType());
1122 }
1123 ToReplace->replaceAllUsesWith(ActualReplacement);
1124 ToReplace->eraseFromParent();
1125
1126 MadeChange = true;
1127 }
1128 return MadeChange;
1129 }
1130
1131 // Turns this:
1132 //
1133 // %base = ...
1134 // %ptr = gep %base + 15
1135 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1136 // %base' = relocate(%tok, i32 4, i32 4)
1137 // %ptr' = relocate(%tok, i32 4, i32 5)
1138 // %val = load %ptr'
1139 //
1140 // into this:
1141 //
1142 // %base = ...
1143 // %ptr = gep %base + 15
1144 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1145 // %base' = gc.relocate(%tok, i32 4, i32 4)
1146 // %ptr' = gep %base' + 15
1147 // %val = load %ptr'
simplifyOffsetableRelocate(GCStatepointInst & I)1148 bool CodeGenPrepare::simplifyOffsetableRelocate(GCStatepointInst &I) {
1149 bool MadeChange = false;
1150 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1151 for (auto *U : I.users())
1152 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1153 // Collect all the relocate calls associated with a statepoint
1154 AllRelocateCalls.push_back(Relocate);
1155
1156 // We need at least one base pointer relocation + one derived pointer
1157 // relocation to mangle
1158 if (AllRelocateCalls.size() < 2)
1159 return false;
1160
1161 // RelocateInstMap is a mapping from the base relocate instruction to the
1162 // corresponding derived relocate instructions
1163 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
1164 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1165 if (RelocateInstMap.empty())
1166 return false;
1167
1168 for (auto &Item : RelocateInstMap)
1169 // Item.first is the RelocatedBase to offset against
1170 // Item.second is the vector of Targets to replace
1171 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1172 return MadeChange;
1173 }
1174
1175 /// Sink the specified cast instruction into its user blocks.
SinkCast(CastInst * CI)1176 static bool SinkCast(CastInst *CI) {
1177 BasicBlock *DefBB = CI->getParent();
1178
1179 /// InsertedCasts - Only insert a cast in each block once.
1180 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1181
1182 bool MadeChange = false;
1183 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1184 UI != E; ) {
1185 Use &TheUse = UI.getUse();
1186 Instruction *User = cast<Instruction>(*UI);
1187
1188 // Figure out which BB this cast is used in. For PHI's this is the
1189 // appropriate predecessor block.
1190 BasicBlock *UserBB = User->getParent();
1191 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1192 UserBB = PN->getIncomingBlock(TheUse);
1193 }
1194
1195 // Preincrement use iterator so we don't invalidate it.
1196 ++UI;
1197
1198 // The first insertion point of a block containing an EH pad is after the
1199 // pad. If the pad is the user, we cannot sink the cast past the pad.
1200 if (User->isEHPad())
1201 continue;
1202
1203 // If the block selected to receive the cast is an EH pad that does not
1204 // allow non-PHI instructions before the terminator, we can't sink the
1205 // cast.
1206 if (UserBB->getTerminator()->isEHPad())
1207 continue;
1208
1209 // If this user is in the same block as the cast, don't change the cast.
1210 if (UserBB == DefBB) continue;
1211
1212 // If we have already inserted a cast into this block, use it.
1213 CastInst *&InsertedCast = InsertedCasts[UserBB];
1214
1215 if (!InsertedCast) {
1216 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1217 assert(InsertPt != UserBB->end());
1218 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1219 CI->getType(), "", &*InsertPt);
1220 InsertedCast->setDebugLoc(CI->getDebugLoc());
1221 }
1222
1223 // Replace a use of the cast with a use of the new cast.
1224 TheUse = InsertedCast;
1225 MadeChange = true;
1226 ++NumCastUses;
1227 }
1228
1229 // If we removed all uses, nuke the cast.
1230 if (CI->use_empty()) {
1231 salvageDebugInfo(*CI);
1232 CI->eraseFromParent();
1233 MadeChange = true;
1234 }
1235
1236 return MadeChange;
1237 }
1238
1239 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1240 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1241 /// reduce the number of virtual registers that must be created and coalesced.
1242 ///
1243 /// Return true if any changes are made.
OptimizeNoopCopyExpression(CastInst * CI,const TargetLowering & TLI,const DataLayout & DL)1244 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1245 const DataLayout &DL) {
1246 // Sink only "cheap" (or nop) address-space casts. This is a weaker condition
1247 // than sinking only nop casts, but is helpful on some platforms.
1248 if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) {
1249 if (!TLI.isFreeAddrSpaceCast(ASC->getSrcAddressSpace(),
1250 ASC->getDestAddressSpace()))
1251 return false;
1252 }
1253
1254 // If this is a noop copy,
1255 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1256 EVT DstVT = TLI.getValueType(DL, CI->getType());
1257
1258 // This is an fp<->int conversion?
1259 if (SrcVT.isInteger() != DstVT.isInteger())
1260 return false;
1261
1262 // If this is an extension, it will be a zero or sign extension, which
1263 // isn't a noop.
1264 if (SrcVT.bitsLT(DstVT)) return false;
1265
1266 // If these values will be promoted, find out what they will be promoted
1267 // to. This helps us consider truncates on PPC as noop copies when they
1268 // are.
1269 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1270 TargetLowering::TypePromoteInteger)
1271 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1272 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1273 TargetLowering::TypePromoteInteger)
1274 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1275
1276 // If, after promotion, these are the same types, this is a noop copy.
1277 if (SrcVT != DstVT)
1278 return false;
1279
1280 return SinkCast(CI);
1281 }
1282
replaceMathCmpWithIntrinsic(BinaryOperator * BO,Value * Arg0,Value * Arg1,CmpInst * Cmp,Intrinsic::ID IID)1283 bool CodeGenPrepare::replaceMathCmpWithIntrinsic(BinaryOperator *BO,
1284 Value *Arg0, Value *Arg1,
1285 CmpInst *Cmp,
1286 Intrinsic::ID IID) {
1287 if (BO->getParent() != Cmp->getParent()) {
1288 // We used to use a dominator tree here to allow multi-block optimization.
1289 // But that was problematic because:
1290 // 1. It could cause a perf regression by hoisting the math op into the
1291 // critical path.
1292 // 2. It could cause a perf regression by creating a value that was live
1293 // across multiple blocks and increasing register pressure.
1294 // 3. Use of a dominator tree could cause large compile-time regression.
1295 // This is because we recompute the DT on every change in the main CGP
1296 // run-loop. The recomputing is probably unnecessary in many cases, so if
1297 // that was fixed, using a DT here would be ok.
1298 return false;
1299 }
1300
1301 // We allow matching the canonical IR (add X, C) back to (usubo X, -C).
1302 if (BO->getOpcode() == Instruction::Add &&
1303 IID == Intrinsic::usub_with_overflow) {
1304 assert(isa<Constant>(Arg1) && "Unexpected input for usubo");
1305 Arg1 = ConstantExpr::getNeg(cast<Constant>(Arg1));
1306 }
1307
1308 // Insert at the first instruction of the pair.
1309 Instruction *InsertPt = nullptr;
1310 for (Instruction &Iter : *Cmp->getParent()) {
1311 // If BO is an XOR, it is not guaranteed that it comes after both inputs to
1312 // the overflow intrinsic are defined.
1313 if ((BO->getOpcode() != Instruction::Xor && &Iter == BO) || &Iter == Cmp) {
1314 InsertPt = &Iter;
1315 break;
1316 }
1317 }
1318 assert(InsertPt != nullptr && "Parent block did not contain cmp or binop");
1319
1320 IRBuilder<> Builder(InsertPt);
1321 Value *MathOV = Builder.CreateBinaryIntrinsic(IID, Arg0, Arg1);
1322 if (BO->getOpcode() != Instruction::Xor) {
1323 Value *Math = Builder.CreateExtractValue(MathOV, 0, "math");
1324 BO->replaceAllUsesWith(Math);
1325 } else
1326 assert(BO->hasOneUse() &&
1327 "Patterns with XOr should use the BO only in the compare");
1328 Value *OV = Builder.CreateExtractValue(MathOV, 1, "ov");
1329 Cmp->replaceAllUsesWith(OV);
1330 Cmp->eraseFromParent();
1331 BO->eraseFromParent();
1332 return true;
1333 }
1334
1335 /// Match special-case patterns that check for unsigned add overflow.
matchUAddWithOverflowConstantEdgeCases(CmpInst * Cmp,BinaryOperator * & Add)1336 static bool matchUAddWithOverflowConstantEdgeCases(CmpInst *Cmp,
1337 BinaryOperator *&Add) {
1338 // Add = add A, 1; Cmp = icmp eq A,-1 (overflow if A is max val)
1339 // Add = add A,-1; Cmp = icmp ne A, 0 (overflow if A is non-zero)
1340 Value *A = Cmp->getOperand(0), *B = Cmp->getOperand(1);
1341
1342 // We are not expecting non-canonical/degenerate code. Just bail out.
1343 if (isa<Constant>(A))
1344 return false;
1345
1346 ICmpInst::Predicate Pred = Cmp->getPredicate();
1347 if (Pred == ICmpInst::ICMP_EQ && match(B, m_AllOnes()))
1348 B = ConstantInt::get(B->getType(), 1);
1349 else if (Pred == ICmpInst::ICMP_NE && match(B, m_ZeroInt()))
1350 B = ConstantInt::get(B->getType(), -1);
1351 else
1352 return false;
1353
1354 // Check the users of the variable operand of the compare looking for an add
1355 // with the adjusted constant.
1356 for (User *U : A->users()) {
1357 if (match(U, m_Add(m_Specific(A), m_Specific(B)))) {
1358 Add = cast<BinaryOperator>(U);
1359 return true;
1360 }
1361 }
1362 return false;
1363 }
1364
1365 /// Try to combine the compare into a call to the llvm.uadd.with.overflow
1366 /// intrinsic. Return true if any changes were made.
combineToUAddWithOverflow(CmpInst * Cmp,bool & ModifiedDT)1367 bool CodeGenPrepare::combineToUAddWithOverflow(CmpInst *Cmp,
1368 bool &ModifiedDT) {
1369 Value *A, *B;
1370 BinaryOperator *Add;
1371 if (!match(Cmp, m_UAddWithOverflow(m_Value(A), m_Value(B), m_BinOp(Add)))) {
1372 if (!matchUAddWithOverflowConstantEdgeCases(Cmp, Add))
1373 return false;
1374 // Set A and B in case we match matchUAddWithOverflowConstantEdgeCases.
1375 A = Add->getOperand(0);
1376 B = Add->getOperand(1);
1377 }
1378
1379 if (!TLI->shouldFormOverflowOp(ISD::UADDO,
1380 TLI->getValueType(*DL, Add->getType()),
1381 Add->hasNUsesOrMore(2)))
1382 return false;
1383
1384 // We don't want to move around uses of condition values this late, so we
1385 // check if it is legal to create the call to the intrinsic in the basic
1386 // block containing the icmp.
1387 if (Add->getParent() != Cmp->getParent() && !Add->hasOneUse())
1388 return false;
1389
1390 if (!replaceMathCmpWithIntrinsic(Add, A, B, Cmp,
1391 Intrinsic::uadd_with_overflow))
1392 return false;
1393
1394 // Reset callers - do not crash by iterating over a dead instruction.
1395 ModifiedDT = true;
1396 return true;
1397 }
1398
combineToUSubWithOverflow(CmpInst * Cmp,bool & ModifiedDT)1399 bool CodeGenPrepare::combineToUSubWithOverflow(CmpInst *Cmp,
1400 bool &ModifiedDT) {
1401 // We are not expecting non-canonical/degenerate code. Just bail out.
1402 Value *A = Cmp->getOperand(0), *B = Cmp->getOperand(1);
1403 if (isa<Constant>(A) && isa<Constant>(B))
1404 return false;
1405
1406 // Convert (A u> B) to (A u< B) to simplify pattern matching.
1407 ICmpInst::Predicate Pred = Cmp->getPredicate();
1408 if (Pred == ICmpInst::ICMP_UGT) {
1409 std::swap(A, B);
1410 Pred = ICmpInst::ICMP_ULT;
1411 }
1412 // Convert special-case: (A == 0) is the same as (A u< 1).
1413 if (Pred == ICmpInst::ICMP_EQ && match(B, m_ZeroInt())) {
1414 B = ConstantInt::get(B->getType(), 1);
1415 Pred = ICmpInst::ICMP_ULT;
1416 }
1417 // Convert special-case: (A != 0) is the same as (0 u< A).
1418 if (Pred == ICmpInst::ICMP_NE && match(B, m_ZeroInt())) {
1419 std::swap(A, B);
1420 Pred = ICmpInst::ICMP_ULT;
1421 }
1422 if (Pred != ICmpInst::ICMP_ULT)
1423 return false;
1424
1425 // Walk the users of a variable operand of a compare looking for a subtract or
1426 // add with that same operand. Also match the 2nd operand of the compare to
1427 // the add/sub, but that may be a negated constant operand of an add.
1428 Value *CmpVariableOperand = isa<Constant>(A) ? B : A;
1429 BinaryOperator *Sub = nullptr;
1430 for (User *U : CmpVariableOperand->users()) {
1431 // A - B, A u< B --> usubo(A, B)
1432 if (match(U, m_Sub(m_Specific(A), m_Specific(B)))) {
1433 Sub = cast<BinaryOperator>(U);
1434 break;
1435 }
1436
1437 // A + (-C), A u< C (canonicalized form of (sub A, C))
1438 const APInt *CmpC, *AddC;
1439 if (match(U, m_Add(m_Specific(A), m_APInt(AddC))) &&
1440 match(B, m_APInt(CmpC)) && *AddC == -(*CmpC)) {
1441 Sub = cast<BinaryOperator>(U);
1442 break;
1443 }
1444 }
1445 if (!Sub)
1446 return false;
1447
1448 if (!TLI->shouldFormOverflowOp(ISD::USUBO,
1449 TLI->getValueType(*DL, Sub->getType()),
1450 Sub->hasNUsesOrMore(2)))
1451 return false;
1452
1453 if (!replaceMathCmpWithIntrinsic(Sub, Sub->getOperand(0), Sub->getOperand(1),
1454 Cmp, Intrinsic::usub_with_overflow))
1455 return false;
1456
1457 // Reset callers - do not crash by iterating over a dead instruction.
1458 ModifiedDT = true;
1459 return true;
1460 }
1461
1462 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1463 /// registers that must be created and coalesced. This is a clear win except on
1464 /// targets with multiple condition code registers (PowerPC), where it might
1465 /// lose; some adjustment may be wanted there.
1466 ///
1467 /// Return true if any changes are made.
sinkCmpExpression(CmpInst * Cmp,const TargetLowering & TLI)1468 static bool sinkCmpExpression(CmpInst *Cmp, const TargetLowering &TLI) {
1469 if (TLI.hasMultipleConditionRegisters())
1470 return false;
1471
1472 // Avoid sinking soft-FP comparisons, since this can move them into a loop.
1473 if (TLI.useSoftFloat() && isa<FCmpInst>(Cmp))
1474 return false;
1475
1476 // Only insert a cmp in each block once.
1477 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1478
1479 bool MadeChange = false;
1480 for (Value::user_iterator UI = Cmp->user_begin(), E = Cmp->user_end();
1481 UI != E; ) {
1482 Use &TheUse = UI.getUse();
1483 Instruction *User = cast<Instruction>(*UI);
1484
1485 // Preincrement use iterator so we don't invalidate it.
1486 ++UI;
1487
1488 // Don't bother for PHI nodes.
1489 if (isa<PHINode>(User))
1490 continue;
1491
1492 // Figure out which BB this cmp is used in.
1493 BasicBlock *UserBB = User->getParent();
1494 BasicBlock *DefBB = Cmp->getParent();
1495
1496 // If this user is in the same block as the cmp, don't change the cmp.
1497 if (UserBB == DefBB) continue;
1498
1499 // If we have already inserted a cmp into this block, use it.
1500 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1501
1502 if (!InsertedCmp) {
1503 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1504 assert(InsertPt != UserBB->end());
1505 InsertedCmp =
1506 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(),
1507 Cmp->getOperand(0), Cmp->getOperand(1), "",
1508 &*InsertPt);
1509 // Propagate the debug info.
1510 InsertedCmp->setDebugLoc(Cmp->getDebugLoc());
1511 }
1512
1513 // Replace a use of the cmp with a use of the new cmp.
1514 TheUse = InsertedCmp;
1515 MadeChange = true;
1516 ++NumCmpUses;
1517 }
1518
1519 // If we removed all uses, nuke the cmp.
1520 if (Cmp->use_empty()) {
1521 Cmp->eraseFromParent();
1522 MadeChange = true;
1523 }
1524
1525 return MadeChange;
1526 }
1527
1528 /// For pattern like:
1529 ///
1530 /// DomCond = icmp sgt/slt CmpOp0, CmpOp1 (might not be in DomBB)
1531 /// ...
1532 /// DomBB:
1533 /// ...
1534 /// br DomCond, TrueBB, CmpBB
1535 /// CmpBB: (with DomBB being the single predecessor)
1536 /// ...
1537 /// Cmp = icmp eq CmpOp0, CmpOp1
1538 /// ...
1539 ///
1540 /// It would use two comparison on targets that lowering of icmp sgt/slt is
1541 /// different from lowering of icmp eq (PowerPC). This function try to convert
1542 /// 'Cmp = icmp eq CmpOp0, CmpOp1' to ' Cmp = icmp slt/sgt CmpOp0, CmpOp1'.
1543 /// After that, DomCond and Cmp can use the same comparison so reduce one
1544 /// comparison.
1545 ///
1546 /// Return true if any changes are made.
foldICmpWithDominatingICmp(CmpInst * Cmp,const TargetLowering & TLI)1547 static bool foldICmpWithDominatingICmp(CmpInst *Cmp,
1548 const TargetLowering &TLI) {
1549 if (!EnableICMP_EQToICMP_ST && TLI.isEqualityCmpFoldedWithSignedCmp())
1550 return false;
1551
1552 ICmpInst::Predicate Pred = Cmp->getPredicate();
1553 if (Pred != ICmpInst::ICMP_EQ)
1554 return false;
1555
1556 // If icmp eq has users other than BranchInst and SelectInst, converting it to
1557 // icmp slt/sgt would introduce more redundant LLVM IR.
1558 for (User *U : Cmp->users()) {
1559 if (isa<BranchInst>(U))
1560 continue;
1561 if (isa<SelectInst>(U) && cast<SelectInst>(U)->getCondition() == Cmp)
1562 continue;
1563 return false;
1564 }
1565
1566 // This is a cheap/incomplete check for dominance - just match a single
1567 // predecessor with a conditional branch.
1568 BasicBlock *CmpBB = Cmp->getParent();
1569 BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1570 if (!DomBB)
1571 return false;
1572
1573 // We want to ensure that the only way control gets to the comparison of
1574 // interest is that a less/greater than comparison on the same operands is
1575 // false.
1576 Value *DomCond;
1577 BasicBlock *TrueBB, *FalseBB;
1578 if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1579 return false;
1580 if (CmpBB != FalseBB)
1581 return false;
1582
1583 Value *CmpOp0 = Cmp->getOperand(0), *CmpOp1 = Cmp->getOperand(1);
1584 ICmpInst::Predicate DomPred;
1585 if (!match(DomCond, m_ICmp(DomPred, m_Specific(CmpOp0), m_Specific(CmpOp1))))
1586 return false;
1587 if (DomPred != ICmpInst::ICMP_SGT && DomPred != ICmpInst::ICMP_SLT)
1588 return false;
1589
1590 // Convert the equality comparison to the opposite of the dominating
1591 // comparison and swap the direction for all branch/select users.
1592 // We have conceptually converted:
1593 // Res = (a < b) ? <LT_RES> : (a == b) ? <EQ_RES> : <GT_RES>;
1594 // to
1595 // Res = (a < b) ? <LT_RES> : (a > b) ? <GT_RES> : <EQ_RES>;
1596 // And similarly for branches.
1597 for (User *U : Cmp->users()) {
1598 if (auto *BI = dyn_cast<BranchInst>(U)) {
1599 assert(BI->isConditional() && "Must be conditional");
1600 BI->swapSuccessors();
1601 continue;
1602 }
1603 if (auto *SI = dyn_cast<SelectInst>(U)) {
1604 // Swap operands
1605 SI->swapValues();
1606 SI->swapProfMetadata();
1607 continue;
1608 }
1609 llvm_unreachable("Must be a branch or a select");
1610 }
1611 Cmp->setPredicate(CmpInst::getSwappedPredicate(DomPred));
1612 return true;
1613 }
1614
optimizeCmp(CmpInst * Cmp,bool & ModifiedDT)1615 bool CodeGenPrepare::optimizeCmp(CmpInst *Cmp, bool &ModifiedDT) {
1616 if (sinkCmpExpression(Cmp, *TLI))
1617 return true;
1618
1619 if (combineToUAddWithOverflow(Cmp, ModifiedDT))
1620 return true;
1621
1622 if (combineToUSubWithOverflow(Cmp, ModifiedDT))
1623 return true;
1624
1625 if (foldICmpWithDominatingICmp(Cmp, *TLI))
1626 return true;
1627
1628 return false;
1629 }
1630
1631 /// Duplicate and sink the given 'and' instruction into user blocks where it is
1632 /// used in a compare to allow isel to generate better code for targets where
1633 /// this operation can be combined.
1634 ///
1635 /// Return true if any changes are made.
sinkAndCmp0Expression(Instruction * AndI,const TargetLowering & TLI,SetOfInstrs & InsertedInsts)1636 static bool sinkAndCmp0Expression(Instruction *AndI,
1637 const TargetLowering &TLI,
1638 SetOfInstrs &InsertedInsts) {
1639 // Double-check that we're not trying to optimize an instruction that was
1640 // already optimized by some other part of this pass.
1641 assert(!InsertedInsts.count(AndI) &&
1642 "Attempting to optimize already optimized and instruction");
1643 (void) InsertedInsts;
1644
1645 // Nothing to do for single use in same basic block.
1646 if (AndI->hasOneUse() &&
1647 AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent())
1648 return false;
1649
1650 // Try to avoid cases where sinking/duplicating is likely to increase register
1651 // pressure.
1652 if (!isa<ConstantInt>(AndI->getOperand(0)) &&
1653 !isa<ConstantInt>(AndI->getOperand(1)) &&
1654 AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse())
1655 return false;
1656
1657 for (auto *U : AndI->users()) {
1658 Instruction *User = cast<Instruction>(U);
1659
1660 // Only sink 'and' feeding icmp with 0.
1661 if (!isa<ICmpInst>(User))
1662 return false;
1663
1664 auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1));
1665 if (!CmpC || !CmpC->isZero())
1666 return false;
1667 }
1668
1669 if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI))
1670 return false;
1671
1672 LLVM_DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n");
1673 LLVM_DEBUG(AndI->getParent()->dump());
1674
1675 // Push the 'and' into the same block as the icmp 0. There should only be
1676 // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
1677 // others, so we don't need to keep track of which BBs we insert into.
1678 for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
1679 UI != E; ) {
1680 Use &TheUse = UI.getUse();
1681 Instruction *User = cast<Instruction>(*UI);
1682
1683 // Preincrement use iterator so we don't invalidate it.
1684 ++UI;
1685
1686 LLVM_DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n");
1687
1688 // Keep the 'and' in the same place if the use is already in the same block.
1689 Instruction *InsertPt =
1690 User->getParent() == AndI->getParent() ? AndI : User;
1691 Instruction *InsertedAnd =
1692 BinaryOperator::Create(Instruction::And, AndI->getOperand(0),
1693 AndI->getOperand(1), "", InsertPt);
1694 // Propagate the debug info.
1695 InsertedAnd->setDebugLoc(AndI->getDebugLoc());
1696
1697 // Replace a use of the 'and' with a use of the new 'and'.
1698 TheUse = InsertedAnd;
1699 ++NumAndUses;
1700 LLVM_DEBUG(User->getParent()->dump());
1701 }
1702
1703 // We removed all uses, nuke the and.
1704 AndI->eraseFromParent();
1705 return true;
1706 }
1707
1708 /// Check if the candidates could be combined with a shift instruction, which
1709 /// includes:
1710 /// 1. Truncate instruction
1711 /// 2. And instruction and the imm is a mask of the low bits:
1712 /// imm & (imm+1) == 0
isExtractBitsCandidateUse(Instruction * User)1713 static bool isExtractBitsCandidateUse(Instruction *User) {
1714 if (!isa<TruncInst>(User)) {
1715 if (User->getOpcode() != Instruction::And ||
1716 !isa<ConstantInt>(User->getOperand(1)))
1717 return false;
1718
1719 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
1720
1721 if ((Cimm & (Cimm + 1)).getBoolValue())
1722 return false;
1723 }
1724 return true;
1725 }
1726
1727 /// Sink both shift and truncate instruction to the use of truncate's BB.
1728 static bool
SinkShiftAndTruncate(BinaryOperator * ShiftI,Instruction * User,ConstantInt * CI,DenseMap<BasicBlock *,BinaryOperator * > & InsertedShifts,const TargetLowering & TLI,const DataLayout & DL)1729 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
1730 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
1731 const TargetLowering &TLI, const DataLayout &DL) {
1732 BasicBlock *UserBB = User->getParent();
1733 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
1734 auto *TruncI = cast<TruncInst>(User);
1735 bool MadeChange = false;
1736
1737 for (Value::user_iterator TruncUI = TruncI->user_begin(),
1738 TruncE = TruncI->user_end();
1739 TruncUI != TruncE;) {
1740
1741 Use &TruncTheUse = TruncUI.getUse();
1742 Instruction *TruncUser = cast<Instruction>(*TruncUI);
1743 // Preincrement use iterator so we don't invalidate it.
1744
1745 ++TruncUI;
1746
1747 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
1748 if (!ISDOpcode)
1749 continue;
1750
1751 // If the use is actually a legal node, there will not be an
1752 // implicit truncate.
1753 // FIXME: always querying the result type is just an
1754 // approximation; some nodes' legality is determined by the
1755 // operand or other means. There's no good way to find out though.
1756 if (TLI.isOperationLegalOrCustom(
1757 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
1758 continue;
1759
1760 // Don't bother for PHI nodes.
1761 if (isa<PHINode>(TruncUser))
1762 continue;
1763
1764 BasicBlock *TruncUserBB = TruncUser->getParent();
1765
1766 if (UserBB == TruncUserBB)
1767 continue;
1768
1769 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
1770 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
1771
1772 if (!InsertedShift && !InsertedTrunc) {
1773 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
1774 assert(InsertPt != TruncUserBB->end());
1775 // Sink the shift
1776 if (ShiftI->getOpcode() == Instruction::AShr)
1777 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1778 "", &*InsertPt);
1779 else
1780 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1781 "", &*InsertPt);
1782 InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
1783
1784 // Sink the trunc
1785 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
1786 TruncInsertPt++;
1787 assert(TruncInsertPt != TruncUserBB->end());
1788
1789 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
1790 TruncI->getType(), "", &*TruncInsertPt);
1791 InsertedTrunc->setDebugLoc(TruncI->getDebugLoc());
1792
1793 MadeChange = true;
1794
1795 TruncTheUse = InsertedTrunc;
1796 }
1797 }
1798 return MadeChange;
1799 }
1800
1801 /// Sink the shift *right* instruction into user blocks if the uses could
1802 /// potentially be combined with this shift instruction and generate BitExtract
1803 /// instruction. It will only be applied if the architecture supports BitExtract
1804 /// instruction. Here is an example:
1805 /// BB1:
1806 /// %x.extract.shift = lshr i64 %arg1, 32
1807 /// BB2:
1808 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1809 /// ==>
1810 ///
1811 /// BB2:
1812 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1813 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1814 ///
1815 /// CodeGen will recognize the pattern in BB2 and generate BitExtract
1816 /// instruction.
1817 /// Return true if any changes are made.
OptimizeExtractBits(BinaryOperator * ShiftI,ConstantInt * CI,const TargetLowering & TLI,const DataLayout & DL)1818 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1819 const TargetLowering &TLI,
1820 const DataLayout &DL) {
1821 BasicBlock *DefBB = ShiftI->getParent();
1822
1823 /// Only insert instructions in each block once.
1824 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1825
1826 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1827
1828 bool MadeChange = false;
1829 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1830 UI != E;) {
1831 Use &TheUse = UI.getUse();
1832 Instruction *User = cast<Instruction>(*UI);
1833 // Preincrement use iterator so we don't invalidate it.
1834 ++UI;
1835
1836 // Don't bother for PHI nodes.
1837 if (isa<PHINode>(User))
1838 continue;
1839
1840 if (!isExtractBitsCandidateUse(User))
1841 continue;
1842
1843 BasicBlock *UserBB = User->getParent();
1844
1845 if (UserBB == DefBB) {
1846 // If the shift and truncate instruction are in the same BB. The use of
1847 // the truncate(TruncUse) may still introduce another truncate if not
1848 // legal. In this case, we would like to sink both shift and truncate
1849 // instruction to the BB of TruncUse.
1850 // for example:
1851 // BB1:
1852 // i64 shift.result = lshr i64 opnd, imm
1853 // trunc.result = trunc shift.result to i16
1854 //
1855 // BB2:
1856 // ----> We will have an implicit truncate here if the architecture does
1857 // not have i16 compare.
1858 // cmp i16 trunc.result, opnd2
1859 //
1860 if (isa<TruncInst>(User) && shiftIsLegal
1861 // If the type of the truncate is legal, no truncate will be
1862 // introduced in other basic blocks.
1863 &&
1864 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1865 MadeChange =
1866 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1867
1868 continue;
1869 }
1870 // If we have already inserted a shift into this block, use it.
1871 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1872
1873 if (!InsertedShift) {
1874 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1875 assert(InsertPt != UserBB->end());
1876
1877 if (ShiftI->getOpcode() == Instruction::AShr)
1878 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1879 "", &*InsertPt);
1880 else
1881 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1882 "", &*InsertPt);
1883 InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
1884
1885 MadeChange = true;
1886 }
1887
1888 // Replace a use of the shift with a use of the new shift.
1889 TheUse = InsertedShift;
1890 }
1891
1892 // If we removed all uses, or there are none, nuke the shift.
1893 if (ShiftI->use_empty()) {
1894 salvageDebugInfo(*ShiftI);
1895 ShiftI->eraseFromParent();
1896 MadeChange = true;
1897 }
1898
1899 return MadeChange;
1900 }
1901
1902 /// If counting leading or trailing zeros is an expensive operation and a zero
1903 /// input is defined, add a check for zero to avoid calling the intrinsic.
1904 ///
1905 /// We want to transform:
1906 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
1907 ///
1908 /// into:
1909 /// entry:
1910 /// %cmpz = icmp eq i64 %A, 0
1911 /// br i1 %cmpz, label %cond.end, label %cond.false
1912 /// cond.false:
1913 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
1914 /// br label %cond.end
1915 /// cond.end:
1916 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
1917 ///
1918 /// If the transform is performed, return true and set ModifiedDT to true.
despeculateCountZeros(IntrinsicInst * CountZeros,const TargetLowering * TLI,const DataLayout * DL,bool & ModifiedDT)1919 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
1920 const TargetLowering *TLI,
1921 const DataLayout *DL,
1922 bool &ModifiedDT) {
1923 // If a zero input is undefined, it doesn't make sense to despeculate that.
1924 if (match(CountZeros->getOperand(1), m_One()))
1925 return false;
1926
1927 // If it's cheap to speculate, there's nothing to do.
1928 auto IntrinsicID = CountZeros->getIntrinsicID();
1929 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
1930 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
1931 return false;
1932
1933 // Only handle legal scalar cases. Anything else requires too much work.
1934 Type *Ty = CountZeros->getType();
1935 unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
1936 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits())
1937 return false;
1938
1939 // The intrinsic will be sunk behind a compare against zero and branch.
1940 BasicBlock *StartBlock = CountZeros->getParent();
1941 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
1942
1943 // Create another block after the count zero intrinsic. A PHI will be added
1944 // in this block to select the result of the intrinsic or the bit-width
1945 // constant if the input to the intrinsic is zero.
1946 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
1947 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
1948
1949 // Set up a builder to create a compare, conditional branch, and PHI.
1950 IRBuilder<> Builder(CountZeros->getContext());
1951 Builder.SetInsertPoint(StartBlock->getTerminator());
1952 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
1953
1954 // Replace the unconditional branch that was created by the first split with
1955 // a compare against zero and a conditional branch.
1956 Value *Zero = Constant::getNullValue(Ty);
1957 Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
1958 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
1959 StartBlock->getTerminator()->eraseFromParent();
1960
1961 // Create a PHI in the end block to select either the output of the intrinsic
1962 // or the bit width of the operand.
1963 Builder.SetInsertPoint(&EndBlock->front());
1964 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
1965 CountZeros->replaceAllUsesWith(PN);
1966 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
1967 PN->addIncoming(BitWidth, StartBlock);
1968 PN->addIncoming(CountZeros, CallBlock);
1969
1970 // We are explicitly handling the zero case, so we can set the intrinsic's
1971 // undefined zero argument to 'true'. This will also prevent reprocessing the
1972 // intrinsic; we only despeculate when a zero input is defined.
1973 CountZeros->setArgOperand(1, Builder.getTrue());
1974 ModifiedDT = true;
1975 return true;
1976 }
1977
optimizeCallInst(CallInst * CI,bool & ModifiedDT)1978 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) {
1979 BasicBlock *BB = CI->getParent();
1980
1981 // Lower inline assembly if we can.
1982 // If we found an inline asm expession, and if the target knows how to
1983 // lower it to normal LLVM code, do so now.
1984 if (CI->isInlineAsm()) {
1985 if (TLI->ExpandInlineAsm(CI)) {
1986 // Avoid invalidating the iterator.
1987 CurInstIterator = BB->begin();
1988 // Avoid processing instructions out of order, which could cause
1989 // reuse before a value is defined.
1990 SunkAddrs.clear();
1991 return true;
1992 }
1993 // Sink address computing for memory operands into the block.
1994 if (optimizeInlineAsmInst(CI))
1995 return true;
1996 }
1997
1998 // Align the pointer arguments to this call if the target thinks it's a good
1999 // idea
2000 unsigned MinSize, PrefAlign;
2001 if (TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
2002 for (auto &Arg : CI->arg_operands()) {
2003 // We want to align both objects whose address is used directly and
2004 // objects whose address is used in casts and GEPs, though it only makes
2005 // sense for GEPs if the offset is a multiple of the desired alignment and
2006 // if size - offset meets the size threshold.
2007 if (!Arg->getType()->isPointerTy())
2008 continue;
2009 APInt Offset(DL->getIndexSizeInBits(
2010 cast<PointerType>(Arg->getType())->getAddressSpace()),
2011 0);
2012 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
2013 uint64_t Offset2 = Offset.getLimitedValue();
2014 if ((Offset2 & (PrefAlign-1)) != 0)
2015 continue;
2016 AllocaInst *AI;
2017 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
2018 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
2019 AI->setAlignment(Align(PrefAlign));
2020 // Global variables can only be aligned if they are defined in this
2021 // object (i.e. they are uniquely initialized in this object), and
2022 // over-aligning global variables that have an explicit section is
2023 // forbidden.
2024 GlobalVariable *GV;
2025 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
2026 GV->getPointerAlignment(*DL) < PrefAlign &&
2027 DL->getTypeAllocSize(GV->getValueType()) >=
2028 MinSize + Offset2)
2029 GV->setAlignment(MaybeAlign(PrefAlign));
2030 }
2031 // If this is a memcpy (or similar) then we may be able to improve the
2032 // alignment
2033 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
2034 Align DestAlign = getKnownAlignment(MI->getDest(), *DL);
2035 MaybeAlign MIDestAlign = MI->getDestAlign();
2036 if (!MIDestAlign || DestAlign > *MIDestAlign)
2037 MI->setDestAlignment(DestAlign);
2038 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
2039 MaybeAlign MTISrcAlign = MTI->getSourceAlign();
2040 Align SrcAlign = getKnownAlignment(MTI->getSource(), *DL);
2041 if (!MTISrcAlign || SrcAlign > *MTISrcAlign)
2042 MTI->setSourceAlignment(SrcAlign);
2043 }
2044 }
2045 }
2046
2047 // If we have a cold call site, try to sink addressing computation into the
2048 // cold block. This interacts with our handling for loads and stores to
2049 // ensure that we can fold all uses of a potential addressing computation
2050 // into their uses. TODO: generalize this to work over profiling data
2051 if (CI->hasFnAttr(Attribute::Cold) &&
2052 !OptSize && !llvm::shouldOptimizeForSize(BB, PSI, BFI.get()))
2053 for (auto &Arg : CI->arg_operands()) {
2054 if (!Arg->getType()->isPointerTy())
2055 continue;
2056 unsigned AS = Arg->getType()->getPointerAddressSpace();
2057 return optimizeMemoryInst(CI, Arg, Arg->getType(), AS);
2058 }
2059
2060 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
2061 if (II) {
2062 switch (II->getIntrinsicID()) {
2063 default: break;
2064 case Intrinsic::assume: {
2065 Value *Operand = II->getOperand(0);
2066 II->eraseFromParent();
2067 // Prune the operand, it's most likely dead.
2068 resetIteratorIfInvalidatedWhileCalling(BB, [&]() {
2069 RecursivelyDeleteTriviallyDeadInstructions(
2070 Operand, TLInfo, nullptr,
2071 [&](Value *V) { removeAllAssertingVHReferences(V); });
2072 });
2073 return true;
2074 }
2075
2076 case Intrinsic::experimental_widenable_condition: {
2077 // Give up on future widening oppurtunties so that we can fold away dead
2078 // paths and merge blocks before going into block-local instruction
2079 // selection.
2080 if (II->use_empty()) {
2081 II->eraseFromParent();
2082 return true;
2083 }
2084 Constant *RetVal = ConstantInt::getTrue(II->getContext());
2085 resetIteratorIfInvalidatedWhileCalling(BB, [&]() {
2086 replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
2087 });
2088 return true;
2089 }
2090 case Intrinsic::objectsize:
2091 llvm_unreachable("llvm.objectsize.* should have been lowered already");
2092 case Intrinsic::is_constant:
2093 llvm_unreachable("llvm.is.constant.* should have been lowered already");
2094 case Intrinsic::aarch64_stlxr:
2095 case Intrinsic::aarch64_stxr: {
2096 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
2097 if (!ExtVal || !ExtVal->hasOneUse() ||
2098 ExtVal->getParent() == CI->getParent())
2099 return false;
2100 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
2101 ExtVal->moveBefore(CI);
2102 // Mark this instruction as "inserted by CGP", so that other
2103 // optimizations don't touch it.
2104 InsertedInsts.insert(ExtVal);
2105 return true;
2106 }
2107
2108 case Intrinsic::launder_invariant_group:
2109 case Intrinsic::strip_invariant_group: {
2110 Value *ArgVal = II->getArgOperand(0);
2111 auto it = LargeOffsetGEPMap.find(II);
2112 if (it != LargeOffsetGEPMap.end()) {
2113 // Merge entries in LargeOffsetGEPMap to reflect the RAUW.
2114 // Make sure not to have to deal with iterator invalidation
2115 // after possibly adding ArgVal to LargeOffsetGEPMap.
2116 auto GEPs = std::move(it->second);
2117 LargeOffsetGEPMap[ArgVal].append(GEPs.begin(), GEPs.end());
2118 LargeOffsetGEPMap.erase(II);
2119 }
2120
2121 II->replaceAllUsesWith(ArgVal);
2122 II->eraseFromParent();
2123 return true;
2124 }
2125 case Intrinsic::cttz:
2126 case Intrinsic::ctlz:
2127 // If counting zeros is expensive, try to avoid it.
2128 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
2129 case Intrinsic::fshl:
2130 case Intrinsic::fshr:
2131 return optimizeFunnelShift(II);
2132 case Intrinsic::dbg_value:
2133 return fixupDbgValue(II);
2134 case Intrinsic::vscale: {
2135 // If datalayout has no special restrictions on vector data layout,
2136 // replace `llvm.vscale` by an equivalent constant expression
2137 // to benefit from cheap constant propagation.
2138 Type *ScalableVectorTy =
2139 VectorType::get(Type::getInt8Ty(II->getContext()), 1, true);
2140 if (DL->getTypeAllocSize(ScalableVectorTy).getKnownMinSize() == 8) {
2141 auto *Null = Constant::getNullValue(ScalableVectorTy->getPointerTo());
2142 auto *One = ConstantInt::getSigned(II->getType(), 1);
2143 auto *CGep =
2144 ConstantExpr::getGetElementPtr(ScalableVectorTy, Null, One);
2145 II->replaceAllUsesWith(ConstantExpr::getPtrToInt(CGep, II->getType()));
2146 II->eraseFromParent();
2147 return true;
2148 }
2149 break;
2150 }
2151 case Intrinsic::masked_gather:
2152 return optimizeGatherScatterInst(II, II->getArgOperand(0));
2153 case Intrinsic::masked_scatter:
2154 return optimizeGatherScatterInst(II, II->getArgOperand(1));
2155 }
2156
2157 SmallVector<Value *, 2> PtrOps;
2158 Type *AccessTy;
2159 if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
2160 while (!PtrOps.empty()) {
2161 Value *PtrVal = PtrOps.pop_back_val();
2162 unsigned AS = PtrVal->getType()->getPointerAddressSpace();
2163 if (optimizeMemoryInst(II, PtrVal, AccessTy, AS))
2164 return true;
2165 }
2166 }
2167
2168 // From here on out we're working with named functions.
2169 if (!CI->getCalledFunction()) return false;
2170
2171 // Lower all default uses of _chk calls. This is very similar
2172 // to what InstCombineCalls does, but here we are only lowering calls
2173 // to fortified library functions (e.g. __memcpy_chk) that have the default
2174 // "don't know" as the objectsize. Anything else should be left alone.
2175 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
2176 IRBuilder<> Builder(CI);
2177 if (Value *V = Simplifier.optimizeCall(CI, Builder)) {
2178 CI->replaceAllUsesWith(V);
2179 CI->eraseFromParent();
2180 return true;
2181 }
2182
2183 return false;
2184 }
2185
2186 /// Look for opportunities to duplicate return instructions to the predecessor
2187 /// to enable tail call optimizations. The case it is currently looking for is:
2188 /// @code
2189 /// bb0:
2190 /// %tmp0 = tail call i32 @f0()
2191 /// br label %return
2192 /// bb1:
2193 /// %tmp1 = tail call i32 @f1()
2194 /// br label %return
2195 /// bb2:
2196 /// %tmp2 = tail call i32 @f2()
2197 /// br label %return
2198 /// return:
2199 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
2200 /// ret i32 %retval
2201 /// @endcode
2202 ///
2203 /// =>
2204 ///
2205 /// @code
2206 /// bb0:
2207 /// %tmp0 = tail call i32 @f0()
2208 /// ret i32 %tmp0
2209 /// bb1:
2210 /// %tmp1 = tail call i32 @f1()
2211 /// ret i32 %tmp1
2212 /// bb2:
2213 /// %tmp2 = tail call i32 @f2()
2214 /// ret i32 %tmp2
2215 /// @endcode
dupRetToEnableTailCallOpts(BasicBlock * BB,bool & ModifiedDT)2216 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB, bool &ModifiedDT) {
2217 ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator());
2218 if (!RetI)
2219 return false;
2220
2221 PHINode *PN = nullptr;
2222 ExtractValueInst *EVI = nullptr;
2223 BitCastInst *BCI = nullptr;
2224 Value *V = RetI->getReturnValue();
2225 if (V) {
2226 BCI = dyn_cast<BitCastInst>(V);
2227 if (BCI)
2228 V = BCI->getOperand(0);
2229
2230 EVI = dyn_cast<ExtractValueInst>(V);
2231 if (EVI) {
2232 V = EVI->getOperand(0);
2233 if (!llvm::all_of(EVI->indices(), [](unsigned idx) { return idx == 0; }))
2234 return false;
2235 }
2236
2237 PN = dyn_cast<PHINode>(V);
2238 if (!PN)
2239 return false;
2240 }
2241
2242 if (PN && PN->getParent() != BB)
2243 return false;
2244
2245 // Make sure there are no instructions between the PHI and return, or that the
2246 // return is the first instruction in the block.
2247 if (PN) {
2248 BasicBlock::iterator BI = BB->begin();
2249 // Skip over debug and the bitcast.
2250 do {
2251 ++BI;
2252 } while (isa<DbgInfoIntrinsic>(BI) || &*BI == BCI || &*BI == EVI ||
2253 isa<PseudoProbeInst>(BI));
2254 if (&*BI != RetI)
2255 return false;
2256 } else {
2257 if (BB->getFirstNonPHIOrDbg(true) != RetI)
2258 return false;
2259 }
2260
2261 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
2262 /// call.
2263 const Function *F = BB->getParent();
2264 SmallVector<BasicBlock*, 4> TailCallBBs;
2265 if (PN) {
2266 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
2267 // Look through bitcasts.
2268 Value *IncomingVal = PN->getIncomingValue(I)->stripPointerCasts();
2269 CallInst *CI = dyn_cast<CallInst>(IncomingVal);
2270 BasicBlock *PredBB = PN->getIncomingBlock(I);
2271 // Make sure the phi value is indeed produced by the tail call.
2272 if (CI && CI->hasOneUse() && CI->getParent() == PredBB &&
2273 TLI->mayBeEmittedAsTailCall(CI) &&
2274 attributesPermitTailCall(F, CI, RetI, *TLI))
2275 TailCallBBs.push_back(PredBB);
2276 }
2277 } else {
2278 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
2279 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
2280 if (!VisitedBBs.insert(*PI).second)
2281 continue;
2282 if (Instruction *I = (*PI)->rbegin()->getPrevNonDebugInstruction(true)) {
2283 CallInst *CI = dyn_cast<CallInst>(I);
2284 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) &&
2285 attributesPermitTailCall(F, CI, RetI, *TLI))
2286 TailCallBBs.push_back(*PI);
2287 }
2288 }
2289 }
2290
2291 bool Changed = false;
2292 for (auto const &TailCallBB : TailCallBBs) {
2293 // Make sure the call instruction is followed by an unconditional branch to
2294 // the return block.
2295 BranchInst *BI = dyn_cast<BranchInst>(TailCallBB->getTerminator());
2296 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
2297 continue;
2298
2299 // Duplicate the return into TailCallBB.
2300 (void)FoldReturnIntoUncondBranch(RetI, BB, TailCallBB);
2301 assert(!VerifyBFIUpdates ||
2302 BFI->getBlockFreq(BB) >= BFI->getBlockFreq(TailCallBB));
2303 BFI->setBlockFreq(
2304 BB,
2305 (BFI->getBlockFreq(BB) - BFI->getBlockFreq(TailCallBB)).getFrequency());
2306 ModifiedDT = Changed = true;
2307 ++NumRetsDup;
2308 }
2309
2310 // If we eliminated all predecessors of the block, delete the block now.
2311 if (Changed && !BB->hasAddressTaken() && pred_empty(BB))
2312 BB->eraseFromParent();
2313
2314 return Changed;
2315 }
2316
2317 //===----------------------------------------------------------------------===//
2318 // Memory Optimization
2319 //===----------------------------------------------------------------------===//
2320
2321 namespace {
2322
2323 /// This is an extended version of TargetLowering::AddrMode
2324 /// which holds actual Value*'s for register values.
2325 struct ExtAddrMode : public TargetLowering::AddrMode {
2326 Value *BaseReg = nullptr;
2327 Value *ScaledReg = nullptr;
2328 Value *OriginalValue = nullptr;
2329 bool InBounds = true;
2330
2331 enum FieldName {
2332 NoField = 0x00,
2333 BaseRegField = 0x01,
2334 BaseGVField = 0x02,
2335 BaseOffsField = 0x04,
2336 ScaledRegField = 0x08,
2337 ScaleField = 0x10,
2338 MultipleFields = 0xff
2339 };
2340
2341
2342 ExtAddrMode() = default;
2343
2344 void print(raw_ostream &OS) const;
2345 void dump() const;
2346
compare__anon7e9d5d500811::ExtAddrMode2347 FieldName compare(const ExtAddrMode &other) {
2348 // First check that the types are the same on each field, as differing types
2349 // is something we can't cope with later on.
2350 if (BaseReg && other.BaseReg &&
2351 BaseReg->getType() != other.BaseReg->getType())
2352 return MultipleFields;
2353 if (BaseGV && other.BaseGV &&
2354 BaseGV->getType() != other.BaseGV->getType())
2355 return MultipleFields;
2356 if (ScaledReg && other.ScaledReg &&
2357 ScaledReg->getType() != other.ScaledReg->getType())
2358 return MultipleFields;
2359
2360 // Conservatively reject 'inbounds' mismatches.
2361 if (InBounds != other.InBounds)
2362 return MultipleFields;
2363
2364 // Check each field to see if it differs.
2365 unsigned Result = NoField;
2366 if (BaseReg != other.BaseReg)
2367 Result |= BaseRegField;
2368 if (BaseGV != other.BaseGV)
2369 Result |= BaseGVField;
2370 if (BaseOffs != other.BaseOffs)
2371 Result |= BaseOffsField;
2372 if (ScaledReg != other.ScaledReg)
2373 Result |= ScaledRegField;
2374 // Don't count 0 as being a different scale, because that actually means
2375 // unscaled (which will already be counted by having no ScaledReg).
2376 if (Scale && other.Scale && Scale != other.Scale)
2377 Result |= ScaleField;
2378
2379 if (countPopulation(Result) > 1)
2380 return MultipleFields;
2381 else
2382 return static_cast<FieldName>(Result);
2383 }
2384
2385 // An AddrMode is trivial if it involves no calculation i.e. it is just a base
2386 // with no offset.
isTrivial__anon7e9d5d500811::ExtAddrMode2387 bool isTrivial() {
2388 // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
2389 // trivial if at most one of these terms is nonzero, except that BaseGV and
2390 // BaseReg both being zero actually means a null pointer value, which we
2391 // consider to be 'non-zero' here.
2392 return !BaseOffs && !Scale && !(BaseGV && BaseReg);
2393 }
2394
GetFieldAsValue__anon7e9d5d500811::ExtAddrMode2395 Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
2396 switch (Field) {
2397 default:
2398 return nullptr;
2399 case BaseRegField:
2400 return BaseReg;
2401 case BaseGVField:
2402 return BaseGV;
2403 case ScaledRegField:
2404 return ScaledReg;
2405 case BaseOffsField:
2406 return ConstantInt::get(IntPtrTy, BaseOffs);
2407 }
2408 }
2409
SetCombinedField__anon7e9d5d500811::ExtAddrMode2410 void SetCombinedField(FieldName Field, Value *V,
2411 const SmallVectorImpl<ExtAddrMode> &AddrModes) {
2412 switch (Field) {
2413 default:
2414 llvm_unreachable("Unhandled fields are expected to be rejected earlier");
2415 break;
2416 case ExtAddrMode::BaseRegField:
2417 BaseReg = V;
2418 break;
2419 case ExtAddrMode::BaseGVField:
2420 // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
2421 // in the BaseReg field.
2422 assert(BaseReg == nullptr);
2423 BaseReg = V;
2424 BaseGV = nullptr;
2425 break;
2426 case ExtAddrMode::ScaledRegField:
2427 ScaledReg = V;
2428 // If we have a mix of scaled and unscaled addrmodes then we want scale
2429 // to be the scale and not zero.
2430 if (!Scale)
2431 for (const ExtAddrMode &AM : AddrModes)
2432 if (AM.Scale) {
2433 Scale = AM.Scale;
2434 break;
2435 }
2436 break;
2437 case ExtAddrMode::BaseOffsField:
2438 // The offset is no longer a constant, so it goes in ScaledReg with a
2439 // scale of 1.
2440 assert(ScaledReg == nullptr);
2441 ScaledReg = V;
2442 Scale = 1;
2443 BaseOffs = 0;
2444 break;
2445 }
2446 }
2447 };
2448
2449 } // end anonymous namespace
2450
2451 #ifndef NDEBUG
operator <<(raw_ostream & OS,const ExtAddrMode & AM)2452 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2453 AM.print(OS);
2454 return OS;
2455 }
2456 #endif
2457
2458 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
print(raw_ostream & OS) const2459 void ExtAddrMode::print(raw_ostream &OS) const {
2460 bool NeedPlus = false;
2461 OS << "[";
2462 if (InBounds)
2463 OS << "inbounds ";
2464 if (BaseGV) {
2465 OS << (NeedPlus ? " + " : "")
2466 << "GV:";
2467 BaseGV->printAsOperand(OS, /*PrintType=*/false);
2468 NeedPlus = true;
2469 }
2470
2471 if (BaseOffs) {
2472 OS << (NeedPlus ? " + " : "")
2473 << BaseOffs;
2474 NeedPlus = true;
2475 }
2476
2477 if (BaseReg) {
2478 OS << (NeedPlus ? " + " : "")
2479 << "Base:";
2480 BaseReg->printAsOperand(OS, /*PrintType=*/false);
2481 NeedPlus = true;
2482 }
2483 if (Scale) {
2484 OS << (NeedPlus ? " + " : "")
2485 << Scale << "*";
2486 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2487 }
2488
2489 OS << ']';
2490 }
2491
dump() const2492 LLVM_DUMP_METHOD void ExtAddrMode::dump() const {
2493 print(dbgs());
2494 dbgs() << '\n';
2495 }
2496 #endif
2497
2498 namespace {
2499
2500 /// This class provides transaction based operation on the IR.
2501 /// Every change made through this class is recorded in the internal state and
2502 /// can be undone (rollback) until commit is called.
2503 /// CGP does not check if instructions could be speculatively executed when
2504 /// moved. Preserving the original location would pessimize the debugging
2505 /// experience, as well as negatively impact the quality of sample PGO.
2506 class TypePromotionTransaction {
2507 /// This represents the common interface of the individual transaction.
2508 /// Each class implements the logic for doing one specific modification on
2509 /// the IR via the TypePromotionTransaction.
2510 class TypePromotionAction {
2511 protected:
2512 /// The Instruction modified.
2513 Instruction *Inst;
2514
2515 public:
2516 /// Constructor of the action.
2517 /// The constructor performs the related action on the IR.
TypePromotionAction(Instruction * Inst)2518 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2519
2520 virtual ~TypePromotionAction() = default;
2521
2522 /// Undo the modification done by this action.
2523 /// When this method is called, the IR must be in the same state as it was
2524 /// before this action was applied.
2525 /// \pre Undoing the action works if and only if the IR is in the exact same
2526 /// state as it was directly after this action was applied.
2527 virtual void undo() = 0;
2528
2529 /// Advocate every change made by this action.
2530 /// When the results on the IR of the action are to be kept, it is important
2531 /// to call this function, otherwise hidden information may be kept forever.
commit()2532 virtual void commit() {
2533 // Nothing to be done, this action is not doing anything.
2534 }
2535 };
2536
2537 /// Utility to remember the position of an instruction.
2538 class InsertionHandler {
2539 /// Position of an instruction.
2540 /// Either an instruction:
2541 /// - Is the first in a basic block: BB is used.
2542 /// - Has a previous instruction: PrevInst is used.
2543 union {
2544 Instruction *PrevInst;
2545 BasicBlock *BB;
2546 } Point;
2547
2548 /// Remember whether or not the instruction had a previous instruction.
2549 bool HasPrevInstruction;
2550
2551 public:
2552 /// Record the position of \p Inst.
InsertionHandler(Instruction * Inst)2553 InsertionHandler(Instruction *Inst) {
2554 BasicBlock::iterator It = Inst->getIterator();
2555 HasPrevInstruction = (It != (Inst->getParent()->begin()));
2556 if (HasPrevInstruction)
2557 Point.PrevInst = &*--It;
2558 else
2559 Point.BB = Inst->getParent();
2560 }
2561
2562 /// Insert \p Inst at the recorded position.
insert(Instruction * Inst)2563 void insert(Instruction *Inst) {
2564 if (HasPrevInstruction) {
2565 if (Inst->getParent())
2566 Inst->removeFromParent();
2567 Inst->insertAfter(Point.PrevInst);
2568 } else {
2569 Instruction *Position = &*Point.BB->getFirstInsertionPt();
2570 if (Inst->getParent())
2571 Inst->moveBefore(Position);
2572 else
2573 Inst->insertBefore(Position);
2574 }
2575 }
2576 };
2577
2578 /// Move an instruction before another.
2579 class InstructionMoveBefore : public TypePromotionAction {
2580 /// Original position of the instruction.
2581 InsertionHandler Position;
2582
2583 public:
2584 /// Move \p Inst before \p Before.
InstructionMoveBefore(Instruction * Inst,Instruction * Before)2585 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2586 : TypePromotionAction(Inst), Position(Inst) {
2587 LLVM_DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before
2588 << "\n");
2589 Inst->moveBefore(Before);
2590 }
2591
2592 /// Move the instruction back to its original position.
undo()2593 void undo() override {
2594 LLVM_DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2595 Position.insert(Inst);
2596 }
2597 };
2598
2599 /// Set the operand of an instruction with a new value.
2600 class OperandSetter : public TypePromotionAction {
2601 /// Original operand of the instruction.
2602 Value *Origin;
2603
2604 /// Index of the modified instruction.
2605 unsigned Idx;
2606
2607 public:
2608 /// Set \p Idx operand of \p Inst with \p NewVal.
OperandSetter(Instruction * Inst,unsigned Idx,Value * NewVal)2609 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2610 : TypePromotionAction(Inst), Idx(Idx) {
2611 LLVM_DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2612 << "for:" << *Inst << "\n"
2613 << "with:" << *NewVal << "\n");
2614 Origin = Inst->getOperand(Idx);
2615 Inst->setOperand(Idx, NewVal);
2616 }
2617
2618 /// Restore the original value of the instruction.
undo()2619 void undo() override {
2620 LLVM_DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2621 << "for: " << *Inst << "\n"
2622 << "with: " << *Origin << "\n");
2623 Inst->setOperand(Idx, Origin);
2624 }
2625 };
2626
2627 /// Hide the operands of an instruction.
2628 /// Do as if this instruction was not using any of its operands.
2629 class OperandsHider : public TypePromotionAction {
2630 /// The list of original operands.
2631 SmallVector<Value *, 4> OriginalValues;
2632
2633 public:
2634 /// Remove \p Inst from the uses of the operands of \p Inst.
OperandsHider(Instruction * Inst)2635 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2636 LLVM_DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2637 unsigned NumOpnds = Inst->getNumOperands();
2638 OriginalValues.reserve(NumOpnds);
2639 for (unsigned It = 0; It < NumOpnds; ++It) {
2640 // Save the current operand.
2641 Value *Val = Inst->getOperand(It);
2642 OriginalValues.push_back(Val);
2643 // Set a dummy one.
2644 // We could use OperandSetter here, but that would imply an overhead
2645 // that we are not willing to pay.
2646 Inst->setOperand(It, UndefValue::get(Val->getType()));
2647 }
2648 }
2649
2650 /// Restore the original list of uses.
undo()2651 void undo() override {
2652 LLVM_DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2653 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2654 Inst->setOperand(It, OriginalValues[It]);
2655 }
2656 };
2657
2658 /// Build a truncate instruction.
2659 class TruncBuilder : public TypePromotionAction {
2660 Value *Val;
2661
2662 public:
2663 /// Build a truncate instruction of \p Opnd producing a \p Ty
2664 /// result.
2665 /// trunc Opnd to Ty.
TruncBuilder(Instruction * Opnd,Type * Ty)2666 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2667 IRBuilder<> Builder(Opnd);
2668 Builder.SetCurrentDebugLocation(DebugLoc());
2669 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2670 LLVM_DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2671 }
2672
2673 /// Get the built value.
getBuiltValue()2674 Value *getBuiltValue() { return Val; }
2675
2676 /// Remove the built instruction.
undo()2677 void undo() override {
2678 LLVM_DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2679 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2680 IVal->eraseFromParent();
2681 }
2682 };
2683
2684 /// Build a sign extension instruction.
2685 class SExtBuilder : public TypePromotionAction {
2686 Value *Val;
2687
2688 public:
2689 /// Build a sign extension instruction of \p Opnd producing a \p Ty
2690 /// result.
2691 /// sext Opnd to Ty.
SExtBuilder(Instruction * InsertPt,Value * Opnd,Type * Ty)2692 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2693 : TypePromotionAction(InsertPt) {
2694 IRBuilder<> Builder(InsertPt);
2695 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2696 LLVM_DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2697 }
2698
2699 /// Get the built value.
getBuiltValue()2700 Value *getBuiltValue() { return Val; }
2701
2702 /// Remove the built instruction.
undo()2703 void undo() override {
2704 LLVM_DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2705 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2706 IVal->eraseFromParent();
2707 }
2708 };
2709
2710 /// Build a zero extension instruction.
2711 class ZExtBuilder : public TypePromotionAction {
2712 Value *Val;
2713
2714 public:
2715 /// Build a zero extension instruction of \p Opnd producing a \p Ty
2716 /// result.
2717 /// zext Opnd to Ty.
ZExtBuilder(Instruction * InsertPt,Value * Opnd,Type * Ty)2718 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2719 : TypePromotionAction(InsertPt) {
2720 IRBuilder<> Builder(InsertPt);
2721 Builder.SetCurrentDebugLocation(DebugLoc());
2722 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2723 LLVM_DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2724 }
2725
2726 /// Get the built value.
getBuiltValue()2727 Value *getBuiltValue() { return Val; }
2728
2729 /// Remove the built instruction.
undo()2730 void undo() override {
2731 LLVM_DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2732 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2733 IVal->eraseFromParent();
2734 }
2735 };
2736
2737 /// Mutate an instruction to another type.
2738 class TypeMutator : public TypePromotionAction {
2739 /// Record the original type.
2740 Type *OrigTy;
2741
2742 public:
2743 /// Mutate the type of \p Inst into \p NewTy.
TypeMutator(Instruction * Inst,Type * NewTy)2744 TypeMutator(Instruction *Inst, Type *NewTy)
2745 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2746 LLVM_DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2747 << "\n");
2748 Inst->mutateType(NewTy);
2749 }
2750
2751 /// Mutate the instruction back to its original type.
undo()2752 void undo() override {
2753 LLVM_DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2754 << "\n");
2755 Inst->mutateType(OrigTy);
2756 }
2757 };
2758
2759 /// Replace the uses of an instruction by another instruction.
2760 class UsesReplacer : public TypePromotionAction {
2761 /// Helper structure to keep track of the replaced uses.
2762 struct InstructionAndIdx {
2763 /// The instruction using the instruction.
2764 Instruction *Inst;
2765
2766 /// The index where this instruction is used for Inst.
2767 unsigned Idx;
2768
InstructionAndIdx__anon7e9d5d500911::TypePromotionTransaction::UsesReplacer::InstructionAndIdx2769 InstructionAndIdx(Instruction *Inst, unsigned Idx)
2770 : Inst(Inst), Idx(Idx) {}
2771 };
2772
2773 /// Keep track of the original uses (pair Instruction, Index).
2774 SmallVector<InstructionAndIdx, 4> OriginalUses;
2775 /// Keep track of the debug users.
2776 SmallVector<DbgValueInst *, 1> DbgValues;
2777
2778 using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
2779
2780 public:
2781 /// Replace all the use of \p Inst by \p New.
UsesReplacer(Instruction * Inst,Value * New)2782 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2783 LLVM_DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2784 << "\n");
2785 // Record the original uses.
2786 for (Use &U : Inst->uses()) {
2787 Instruction *UserI = cast<Instruction>(U.getUser());
2788 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2789 }
2790 // Record the debug uses separately. They are not in the instruction's
2791 // use list, but they are replaced by RAUW.
2792 findDbgValues(DbgValues, Inst);
2793
2794 // Now, we can replace the uses.
2795 Inst->replaceAllUsesWith(New);
2796 }
2797
2798 /// Reassign the original uses of Inst to Inst.
undo()2799 void undo() override {
2800 LLVM_DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2801 for (use_iterator UseIt = OriginalUses.begin(),
2802 EndIt = OriginalUses.end();
2803 UseIt != EndIt; ++UseIt) {
2804 UseIt->Inst->setOperand(UseIt->Idx, Inst);
2805 }
2806 // RAUW has replaced all original uses with references to the new value,
2807 // including the debug uses. Since we are undoing the replacements,
2808 // the original debug uses must also be reinstated to maintain the
2809 // correctness and utility of debug value instructions.
2810 for (auto *DVI: DbgValues) {
2811 LLVMContext &Ctx = Inst->getType()->getContext();
2812 auto *MV = MetadataAsValue::get(Ctx, ValueAsMetadata::get(Inst));
2813 DVI->setOperand(0, MV);
2814 }
2815 }
2816 };
2817
2818 /// Remove an instruction from the IR.
2819 class InstructionRemover : public TypePromotionAction {
2820 /// Original position of the instruction.
2821 InsertionHandler Inserter;
2822
2823 /// Helper structure to hide all the link to the instruction. In other
2824 /// words, this helps to do as if the instruction was removed.
2825 OperandsHider Hider;
2826
2827 /// Keep track of the uses replaced, if any.
2828 UsesReplacer *Replacer = nullptr;
2829
2830 /// Keep track of instructions removed.
2831 SetOfInstrs &RemovedInsts;
2832
2833 public:
2834 /// Remove all reference of \p Inst and optionally replace all its
2835 /// uses with New.
2836 /// \p RemovedInsts Keep track of the instructions removed by this Action.
2837 /// \pre If !Inst->use_empty(), then New != nullptr
InstructionRemover(Instruction * Inst,SetOfInstrs & RemovedInsts,Value * New=nullptr)2838 InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
2839 Value *New = nullptr)
2840 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2841 RemovedInsts(RemovedInsts) {
2842 if (New)
2843 Replacer = new UsesReplacer(Inst, New);
2844 LLVM_DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2845 RemovedInsts.insert(Inst);
2846 /// The instructions removed here will be freed after completing
2847 /// optimizeBlock() for all blocks as we need to keep track of the
2848 /// removed instructions during promotion.
2849 Inst->removeFromParent();
2850 }
2851
~InstructionRemover()2852 ~InstructionRemover() override { delete Replacer; }
2853
2854 /// Resurrect the instruction and reassign it to the proper uses if
2855 /// new value was provided when build this action.
undo()2856 void undo() override {
2857 LLVM_DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2858 Inserter.insert(Inst);
2859 if (Replacer)
2860 Replacer->undo();
2861 Hider.undo();
2862 RemovedInsts.erase(Inst);
2863 }
2864 };
2865
2866 public:
2867 /// Restoration point.
2868 /// The restoration point is a pointer to an action instead of an iterator
2869 /// because the iterator may be invalidated but not the pointer.
2870 using ConstRestorationPt = const TypePromotionAction *;
2871
TypePromotionTransaction(SetOfInstrs & RemovedInsts)2872 TypePromotionTransaction(SetOfInstrs &RemovedInsts)
2873 : RemovedInsts(RemovedInsts) {}
2874
2875 /// Advocate every changes made in that transaction. Return true if any change
2876 /// happen.
2877 bool commit();
2878
2879 /// Undo all the changes made after the given point.
2880 void rollback(ConstRestorationPt Point);
2881
2882 /// Get the current restoration point.
2883 ConstRestorationPt getRestorationPoint() const;
2884
2885 /// \name API for IR modification with state keeping to support rollback.
2886 /// @{
2887 /// Same as Instruction::setOperand.
2888 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2889
2890 /// Same as Instruction::eraseFromParent.
2891 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2892
2893 /// Same as Value::replaceAllUsesWith.
2894 void replaceAllUsesWith(Instruction *Inst, Value *New);
2895
2896 /// Same as Value::mutateType.
2897 void mutateType(Instruction *Inst, Type *NewTy);
2898
2899 /// Same as IRBuilder::createTrunc.
2900 Value *createTrunc(Instruction *Opnd, Type *Ty);
2901
2902 /// Same as IRBuilder::createSExt.
2903 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2904
2905 /// Same as IRBuilder::createZExt.
2906 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2907
2908 /// Same as Instruction::moveBefore.
2909 void moveBefore(Instruction *Inst, Instruction *Before);
2910 /// @}
2911
2912 private:
2913 /// The ordered list of actions made so far.
2914 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
2915
2916 using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
2917
2918 SetOfInstrs &RemovedInsts;
2919 };
2920
2921 } // end anonymous namespace
2922
setOperand(Instruction * Inst,unsigned Idx,Value * NewVal)2923 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2924 Value *NewVal) {
2925 Actions.push_back(std::make_unique<TypePromotionTransaction::OperandSetter>(
2926 Inst, Idx, NewVal));
2927 }
2928
eraseInstruction(Instruction * Inst,Value * NewVal)2929 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2930 Value *NewVal) {
2931 Actions.push_back(
2932 std::make_unique<TypePromotionTransaction::InstructionRemover>(
2933 Inst, RemovedInsts, NewVal));
2934 }
2935
replaceAllUsesWith(Instruction * Inst,Value * New)2936 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2937 Value *New) {
2938 Actions.push_back(
2939 std::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2940 }
2941
mutateType(Instruction * Inst,Type * NewTy)2942 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2943 Actions.push_back(
2944 std::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2945 }
2946
createTrunc(Instruction * Opnd,Type * Ty)2947 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2948 Type *Ty) {
2949 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2950 Value *Val = Ptr->getBuiltValue();
2951 Actions.push_back(std::move(Ptr));
2952 return Val;
2953 }
2954
createSExt(Instruction * Inst,Value * Opnd,Type * Ty)2955 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2956 Value *Opnd, Type *Ty) {
2957 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2958 Value *Val = Ptr->getBuiltValue();
2959 Actions.push_back(std::move(Ptr));
2960 return Val;
2961 }
2962
createZExt(Instruction * Inst,Value * Opnd,Type * Ty)2963 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2964 Value *Opnd, Type *Ty) {
2965 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2966 Value *Val = Ptr->getBuiltValue();
2967 Actions.push_back(std::move(Ptr));
2968 return Val;
2969 }
2970
moveBefore(Instruction * Inst,Instruction * Before)2971 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2972 Instruction *Before) {
2973 Actions.push_back(
2974 std::make_unique<TypePromotionTransaction::InstructionMoveBefore>(
2975 Inst, Before));
2976 }
2977
2978 TypePromotionTransaction::ConstRestorationPt
getRestorationPoint() const2979 TypePromotionTransaction::getRestorationPoint() const {
2980 return !Actions.empty() ? Actions.back().get() : nullptr;
2981 }
2982
commit()2983 bool TypePromotionTransaction::commit() {
2984 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2985 ++It)
2986 (*It)->commit();
2987 bool Modified = !Actions.empty();
2988 Actions.clear();
2989 return Modified;
2990 }
2991
rollback(TypePromotionTransaction::ConstRestorationPt Point)2992 void TypePromotionTransaction::rollback(
2993 TypePromotionTransaction::ConstRestorationPt Point) {
2994 while (!Actions.empty() && Point != Actions.back().get()) {
2995 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2996 Curr->undo();
2997 }
2998 }
2999
3000 namespace {
3001
3002 /// A helper class for matching addressing modes.
3003 ///
3004 /// This encapsulates the logic for matching the target-legal addressing modes.
3005 class AddressingModeMatcher {
3006 SmallVectorImpl<Instruction*> &AddrModeInsts;
3007 const TargetLowering &TLI;
3008 const TargetRegisterInfo &TRI;
3009 const DataLayout &DL;
3010
3011 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
3012 /// the memory instruction that we're computing this address for.
3013 Type *AccessTy;
3014 unsigned AddrSpace;
3015 Instruction *MemoryInst;
3016
3017 /// This is the addressing mode that we're building up. This is
3018 /// part of the return value of this addressing mode matching stuff.
3019 ExtAddrMode &AddrMode;
3020
3021 /// The instructions inserted by other CodeGenPrepare optimizations.
3022 const SetOfInstrs &InsertedInsts;
3023
3024 /// A map from the instructions to their type before promotion.
3025 InstrToOrigTy &PromotedInsts;
3026
3027 /// The ongoing transaction where every action should be registered.
3028 TypePromotionTransaction &TPT;
3029
3030 // A GEP which has too large offset to be folded into the addressing mode.
3031 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP;
3032
3033 /// This is set to true when we should not do profitability checks.
3034 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
3035 bool IgnoreProfitability;
3036
3037 /// True if we are optimizing for size.
3038 bool OptSize;
3039
3040 ProfileSummaryInfo *PSI;
3041 BlockFrequencyInfo *BFI;
3042
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,bool OptSize,ProfileSummaryInfo * PSI,BlockFrequencyInfo * BFI)3043 AddressingModeMatcher(
3044 SmallVectorImpl<Instruction *> &AMI, const TargetLowering &TLI,
3045 const TargetRegisterInfo &TRI, Type *AT, unsigned AS, Instruction *MI,
3046 ExtAddrMode &AM, const SetOfInstrs &InsertedInsts,
3047 InstrToOrigTy &PromotedInsts, TypePromotionTransaction &TPT,
3048 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP,
3049 bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI)
3050 : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
3051 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
3052 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
3053 PromotedInsts(PromotedInsts), TPT(TPT), LargeOffsetGEP(LargeOffsetGEP),
3054 OptSize(OptSize), PSI(PSI), BFI(BFI) {
3055 IgnoreProfitability = false;
3056 }
3057
3058 public:
3059 /// Find the maximal addressing mode that a load/store of V can fold,
3060 /// give an access type of AccessTy. This returns a list of involved
3061 /// instructions in AddrModeInsts.
3062 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
3063 /// optimizations.
3064 /// \p PromotedInsts maps the instructions to their type before promotion.
3065 /// \p The ongoing transaction where every action should be registered.
3066 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,bool OptSize,ProfileSummaryInfo * PSI,BlockFrequencyInfo * BFI)3067 Match(Value *V, Type *AccessTy, unsigned AS, Instruction *MemoryInst,
3068 SmallVectorImpl<Instruction *> &AddrModeInsts,
3069 const TargetLowering &TLI, const TargetRegisterInfo &TRI,
3070 const SetOfInstrs &InsertedInsts, InstrToOrigTy &PromotedInsts,
3071 TypePromotionTransaction &TPT,
3072 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP,
3073 bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) {
3074 ExtAddrMode Result;
3075
3076 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI, AccessTy, AS,
3077 MemoryInst, Result, InsertedInsts,
3078 PromotedInsts, TPT, LargeOffsetGEP,
3079 OptSize, PSI, BFI)
3080 .matchAddr(V, 0);
3081 (void)Success; assert(Success && "Couldn't select *anything*?");
3082 return Result;
3083 }
3084
3085 private:
3086 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
3087 bool matchAddr(Value *Addr, unsigned Depth);
3088 bool matchOperationAddr(User *AddrInst, unsigned Opcode, unsigned Depth,
3089 bool *MovedAway = nullptr);
3090 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
3091 ExtAddrMode &AMBefore,
3092 ExtAddrMode &AMAfter);
3093 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
3094 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
3095 Value *PromotedOperand) const;
3096 };
3097
3098 class PhiNodeSet;
3099
3100 /// An iterator for PhiNodeSet.
3101 class PhiNodeSetIterator {
3102 PhiNodeSet * const Set;
3103 size_t CurrentIndex = 0;
3104
3105 public:
3106 /// The constructor. Start should point to either a valid element, or be equal
3107 /// to the size of the underlying SmallVector of the PhiNodeSet.
3108 PhiNodeSetIterator(PhiNodeSet * const Set, size_t Start);
3109 PHINode * operator*() const;
3110 PhiNodeSetIterator& operator++();
3111 bool operator==(const PhiNodeSetIterator &RHS) const;
3112 bool operator!=(const PhiNodeSetIterator &RHS) const;
3113 };
3114
3115 /// Keeps a set of PHINodes.
3116 ///
3117 /// This is a minimal set implementation for a specific use case:
3118 /// It is very fast when there are very few elements, but also provides good
3119 /// performance when there are many. It is similar to SmallPtrSet, but also
3120 /// provides iteration by insertion order, which is deterministic and stable
3121 /// across runs. It is also similar to SmallSetVector, but provides removing
3122 /// elements in O(1) time. This is achieved by not actually removing the element
3123 /// from the underlying vector, so comes at the cost of using more memory, but
3124 /// that is fine, since PhiNodeSets are used as short lived objects.
3125 class PhiNodeSet {
3126 friend class PhiNodeSetIterator;
3127
3128 using MapType = SmallDenseMap<PHINode *, size_t, 32>;
3129 using iterator = PhiNodeSetIterator;
3130
3131 /// Keeps the elements in the order of their insertion in the underlying
3132 /// vector. To achieve constant time removal, it never deletes any element.
3133 SmallVector<PHINode *, 32> NodeList;
3134
3135 /// Keeps the elements in the underlying set implementation. This (and not the
3136 /// NodeList defined above) is the source of truth on whether an element
3137 /// is actually in the collection.
3138 MapType NodeMap;
3139
3140 /// Points to the first valid (not deleted) element when the set is not empty
3141 /// and the value is not zero. Equals to the size of the underlying vector
3142 /// when the set is empty. When the value is 0, as in the beginning, the
3143 /// first element may or may not be valid.
3144 size_t FirstValidElement = 0;
3145
3146 public:
3147 /// Inserts a new element to the collection.
3148 /// \returns true if the element is actually added, i.e. was not in the
3149 /// collection before the operation.
insert(PHINode * Ptr)3150 bool insert(PHINode *Ptr) {
3151 if (NodeMap.insert(std::make_pair(Ptr, NodeList.size())).second) {
3152 NodeList.push_back(Ptr);
3153 return true;
3154 }
3155 return false;
3156 }
3157
3158 /// Removes the element from the collection.
3159 /// \returns whether the element is actually removed, i.e. was in the
3160 /// collection before the operation.
erase(PHINode * Ptr)3161 bool erase(PHINode *Ptr) {
3162 if (NodeMap.erase(Ptr)) {
3163 SkipRemovedElements(FirstValidElement);
3164 return true;
3165 }
3166 return false;
3167 }
3168
3169 /// Removes all elements and clears the collection.
clear()3170 void clear() {
3171 NodeMap.clear();
3172 NodeList.clear();
3173 FirstValidElement = 0;
3174 }
3175
3176 /// \returns an iterator that will iterate the elements in the order of
3177 /// insertion.
begin()3178 iterator begin() {
3179 if (FirstValidElement == 0)
3180 SkipRemovedElements(FirstValidElement);
3181 return PhiNodeSetIterator(this, FirstValidElement);
3182 }
3183
3184 /// \returns an iterator that points to the end of the collection.
end()3185 iterator end() { return PhiNodeSetIterator(this, NodeList.size()); }
3186
3187 /// Returns the number of elements in the collection.
size() const3188 size_t size() const {
3189 return NodeMap.size();
3190 }
3191
3192 /// \returns 1 if the given element is in the collection, and 0 if otherwise.
count(PHINode * Ptr) const3193 size_t count(PHINode *Ptr) const {
3194 return NodeMap.count(Ptr);
3195 }
3196
3197 private:
3198 /// Updates the CurrentIndex so that it will point to a valid element.
3199 ///
3200 /// If the element of NodeList at CurrentIndex is valid, it does not
3201 /// change it. If there are no more valid elements, it updates CurrentIndex
3202 /// to point to the end of the NodeList.
SkipRemovedElements(size_t & CurrentIndex)3203 void SkipRemovedElements(size_t &CurrentIndex) {
3204 while (CurrentIndex < NodeList.size()) {
3205 auto it = NodeMap.find(NodeList[CurrentIndex]);
3206 // If the element has been deleted and added again later, NodeMap will
3207 // point to a different index, so CurrentIndex will still be invalid.
3208 if (it != NodeMap.end() && it->second == CurrentIndex)
3209 break;
3210 ++CurrentIndex;
3211 }
3212 }
3213 };
3214
PhiNodeSetIterator(PhiNodeSet * const Set,size_t Start)3215 PhiNodeSetIterator::PhiNodeSetIterator(PhiNodeSet *const Set, size_t Start)
3216 : Set(Set), CurrentIndex(Start) {}
3217
operator *() const3218 PHINode * PhiNodeSetIterator::operator*() const {
3219 assert(CurrentIndex < Set->NodeList.size() &&
3220 "PhiNodeSet access out of range");
3221 return Set->NodeList[CurrentIndex];
3222 }
3223
operator ++()3224 PhiNodeSetIterator& PhiNodeSetIterator::operator++() {
3225 assert(CurrentIndex < Set->NodeList.size() &&
3226 "PhiNodeSet access out of range");
3227 ++CurrentIndex;
3228 Set->SkipRemovedElements(CurrentIndex);
3229 return *this;
3230 }
3231
operator ==(const PhiNodeSetIterator & RHS) const3232 bool PhiNodeSetIterator::operator==(const PhiNodeSetIterator &RHS) const {
3233 return CurrentIndex == RHS.CurrentIndex;
3234 }
3235
operator !=(const PhiNodeSetIterator & RHS) const3236 bool PhiNodeSetIterator::operator!=(const PhiNodeSetIterator &RHS) const {
3237 return !((*this) == RHS);
3238 }
3239
3240 /// Keep track of simplification of Phi nodes.
3241 /// Accept the set of all phi nodes and erase phi node from this set
3242 /// if it is simplified.
3243 class SimplificationTracker {
3244 DenseMap<Value *, Value *> Storage;
3245 const SimplifyQuery &SQ;
3246 // Tracks newly created Phi nodes. The elements are iterated by insertion
3247 // order.
3248 PhiNodeSet AllPhiNodes;
3249 // Tracks newly created Select nodes.
3250 SmallPtrSet<SelectInst *, 32> AllSelectNodes;
3251
3252 public:
SimplificationTracker(const SimplifyQuery & sq)3253 SimplificationTracker(const SimplifyQuery &sq)
3254 : SQ(sq) {}
3255
Get(Value * V)3256 Value *Get(Value *V) {
3257 do {
3258 auto SV = Storage.find(V);
3259 if (SV == Storage.end())
3260 return V;
3261 V = SV->second;
3262 } while (true);
3263 }
3264
Simplify(Value * Val)3265 Value *Simplify(Value *Val) {
3266 SmallVector<Value *, 32> WorkList;
3267 SmallPtrSet<Value *, 32> Visited;
3268 WorkList.push_back(Val);
3269 while (!WorkList.empty()) {
3270 auto *P = WorkList.pop_back_val();
3271 if (!Visited.insert(P).second)
3272 continue;
3273 if (auto *PI = dyn_cast<Instruction>(P))
3274 if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) {
3275 for (auto *U : PI->users())
3276 WorkList.push_back(cast<Value>(U));
3277 Put(PI, V);
3278 PI->replaceAllUsesWith(V);
3279 if (auto *PHI = dyn_cast<PHINode>(PI))
3280 AllPhiNodes.erase(PHI);
3281 if (auto *Select = dyn_cast<SelectInst>(PI))
3282 AllSelectNodes.erase(Select);
3283 PI->eraseFromParent();
3284 }
3285 }
3286 return Get(Val);
3287 }
3288
Put(Value * From,Value * To)3289 void Put(Value *From, Value *To) {
3290 Storage.insert({ From, To });
3291 }
3292
ReplacePhi(PHINode * From,PHINode * To)3293 void ReplacePhi(PHINode *From, PHINode *To) {
3294 Value* OldReplacement = Get(From);
3295 while (OldReplacement != From) {
3296 From = To;
3297 To = dyn_cast<PHINode>(OldReplacement);
3298 OldReplacement = Get(From);
3299 }
3300 assert(To && Get(To) == To && "Replacement PHI node is already replaced.");
3301 Put(From, To);
3302 From->replaceAllUsesWith(To);
3303 AllPhiNodes.erase(From);
3304 From->eraseFromParent();
3305 }
3306
newPhiNodes()3307 PhiNodeSet& newPhiNodes() { return AllPhiNodes; }
3308
insertNewPhi(PHINode * PN)3309 void insertNewPhi(PHINode *PN) { AllPhiNodes.insert(PN); }
3310
insertNewSelect(SelectInst * SI)3311 void insertNewSelect(SelectInst *SI) { AllSelectNodes.insert(SI); }
3312
countNewPhiNodes() const3313 unsigned countNewPhiNodes() const { return AllPhiNodes.size(); }
3314
countNewSelectNodes() const3315 unsigned countNewSelectNodes() const { return AllSelectNodes.size(); }
3316
destroyNewNodes(Type * CommonType)3317 void destroyNewNodes(Type *CommonType) {
3318 // For safe erasing, replace the uses with dummy value first.
3319 auto *Dummy = UndefValue::get(CommonType);
3320 for (auto *I : AllPhiNodes) {
3321 I->replaceAllUsesWith(Dummy);
3322 I->eraseFromParent();
3323 }
3324 AllPhiNodes.clear();
3325 for (auto *I : AllSelectNodes) {
3326 I->replaceAllUsesWith(Dummy);
3327 I->eraseFromParent();
3328 }
3329 AllSelectNodes.clear();
3330 }
3331 };
3332
3333 /// A helper class for combining addressing modes.
3334 class AddressingModeCombiner {
3335 typedef DenseMap<Value *, Value *> FoldAddrToValueMapping;
3336 typedef std::pair<PHINode *, PHINode *> PHIPair;
3337
3338 private:
3339 /// The addressing modes we've collected.
3340 SmallVector<ExtAddrMode, 16> AddrModes;
3341
3342 /// The field in which the AddrModes differ, when we have more than one.
3343 ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
3344
3345 /// Are the AddrModes that we have all just equal to their original values?
3346 bool AllAddrModesTrivial = true;
3347
3348 /// Common Type for all different fields in addressing modes.
3349 Type *CommonType;
3350
3351 /// SimplifyQuery for simplifyInstruction utility.
3352 const SimplifyQuery &SQ;
3353
3354 /// Original Address.
3355 Value *Original;
3356
3357 public:
AddressingModeCombiner(const SimplifyQuery & _SQ,Value * OriginalValue)3358 AddressingModeCombiner(const SimplifyQuery &_SQ, Value *OriginalValue)
3359 : CommonType(nullptr), SQ(_SQ), Original(OriginalValue) {}
3360
3361 /// Get the combined AddrMode
getAddrMode() const3362 const ExtAddrMode &getAddrMode() const {
3363 return AddrModes[0];
3364 }
3365
3366 /// Add a new AddrMode if it's compatible with the AddrModes we already
3367 /// have.
3368 /// \return True iff we succeeded in doing so.
addNewAddrMode(ExtAddrMode & NewAddrMode)3369 bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
3370 // Take note of if we have any non-trivial AddrModes, as we need to detect
3371 // when all AddrModes are trivial as then we would introduce a phi or select
3372 // which just duplicates what's already there.
3373 AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
3374
3375 // If this is the first addrmode then everything is fine.
3376 if (AddrModes.empty()) {
3377 AddrModes.emplace_back(NewAddrMode);
3378 return true;
3379 }
3380
3381 // Figure out how different this is from the other address modes, which we
3382 // can do just by comparing against the first one given that we only care
3383 // about the cumulative difference.
3384 ExtAddrMode::FieldName ThisDifferentField =
3385 AddrModes[0].compare(NewAddrMode);
3386 if (DifferentField == ExtAddrMode::NoField)
3387 DifferentField = ThisDifferentField;
3388 else if (DifferentField != ThisDifferentField)
3389 DifferentField = ExtAddrMode::MultipleFields;
3390
3391 // If NewAddrMode differs in more than one dimension we cannot handle it.
3392 bool CanHandle = DifferentField != ExtAddrMode::MultipleFields;
3393
3394 // If Scale Field is different then we reject.
3395 CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField;
3396
3397 // We also must reject the case when base offset is different and
3398 // scale reg is not null, we cannot handle this case due to merge of
3399 // different offsets will be used as ScaleReg.
3400 CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField ||
3401 !NewAddrMode.ScaledReg);
3402
3403 // We also must reject the case when GV is different and BaseReg installed
3404 // due to we want to use base reg as a merge of GV values.
3405 CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField ||
3406 !NewAddrMode.HasBaseReg);
3407
3408 // Even if NewAddMode is the same we still need to collect it due to
3409 // original value is different. And later we will need all original values
3410 // as anchors during finding the common Phi node.
3411 if (CanHandle)
3412 AddrModes.emplace_back(NewAddrMode);
3413 else
3414 AddrModes.clear();
3415
3416 return CanHandle;
3417 }
3418
3419 /// Combine the addressing modes we've collected into a single
3420 /// addressing mode.
3421 /// \return True iff we successfully combined them or we only had one so
3422 /// didn't need to combine them anyway.
combineAddrModes()3423 bool combineAddrModes() {
3424 // If we have no AddrModes then they can't be combined.
3425 if (AddrModes.size() == 0)
3426 return false;
3427
3428 // A single AddrMode can trivially be combined.
3429 if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
3430 return true;
3431
3432 // If the AddrModes we collected are all just equal to the value they are
3433 // derived from then combining them wouldn't do anything useful.
3434 if (AllAddrModesTrivial)
3435 return false;
3436
3437 if (!addrModeCombiningAllowed())
3438 return false;
3439
3440 // Build a map between <original value, basic block where we saw it> to
3441 // value of base register.
3442 // Bail out if there is no common type.
3443 FoldAddrToValueMapping Map;
3444 if (!initializeMap(Map))
3445 return false;
3446
3447 Value *CommonValue = findCommon(Map);
3448 if (CommonValue)
3449 AddrModes[0].SetCombinedField(DifferentField, CommonValue, AddrModes);
3450 return CommonValue != nullptr;
3451 }
3452
3453 private:
3454 /// Initialize Map with anchor values. For address seen
3455 /// we set the value of different field saw in this address.
3456 /// At the same time we find a common type for different field we will
3457 /// use to create new Phi/Select nodes. Keep it in CommonType field.
3458 /// Return false if there is no common type found.
initializeMap(FoldAddrToValueMapping & Map)3459 bool initializeMap(FoldAddrToValueMapping &Map) {
3460 // Keep track of keys where the value is null. We will need to replace it
3461 // with constant null when we know the common type.
3462 SmallVector<Value *, 2> NullValue;
3463 Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
3464 for (auto &AM : AddrModes) {
3465 Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy);
3466 if (DV) {
3467 auto *Type = DV->getType();
3468 if (CommonType && CommonType != Type)
3469 return false;
3470 CommonType = Type;
3471 Map[AM.OriginalValue] = DV;
3472 } else {
3473 NullValue.push_back(AM.OriginalValue);
3474 }
3475 }
3476 assert(CommonType && "At least one non-null value must be!");
3477 for (auto *V : NullValue)
3478 Map[V] = Constant::getNullValue(CommonType);
3479 return true;
3480 }
3481
3482 /// We have mapping between value A and other value B where B was a field in
3483 /// addressing mode represented by A. Also we have an original value C
3484 /// representing an address we start with. Traversing from C through phi and
3485 /// selects we ended up with A's in a map. This utility function tries to find
3486 /// a value V which is a field in addressing mode C and traversing through phi
3487 /// nodes and selects we will end up in corresponded values B in a map.
3488 /// The utility will create a new Phi/Selects if needed.
3489 // The simple example looks as follows:
3490 // BB1:
3491 // p1 = b1 + 40
3492 // br cond BB2, BB3
3493 // BB2:
3494 // p2 = b2 + 40
3495 // br BB3
3496 // BB3:
3497 // p = phi [p1, BB1], [p2, BB2]
3498 // v = load p
3499 // Map is
3500 // p1 -> b1
3501 // p2 -> b2
3502 // Request is
3503 // p -> ?
3504 // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3.
findCommon(FoldAddrToValueMapping & Map)3505 Value *findCommon(FoldAddrToValueMapping &Map) {
3506 // Tracks the simplification of newly created phi nodes. The reason we use
3507 // this mapping is because we will add new created Phi nodes in AddrToBase.
3508 // Simplification of Phi nodes is recursive, so some Phi node may
3509 // be simplified after we added it to AddrToBase. In reality this
3510 // simplification is possible only if original phi/selects were not
3511 // simplified yet.
3512 // Using this mapping we can find the current value in AddrToBase.
3513 SimplificationTracker ST(SQ);
3514
3515 // First step, DFS to create PHI nodes for all intermediate blocks.
3516 // Also fill traverse order for the second step.
3517 SmallVector<Value *, 32> TraverseOrder;
3518 InsertPlaceholders(Map, TraverseOrder, ST);
3519
3520 // Second Step, fill new nodes by merged values and simplify if possible.
3521 FillPlaceholders(Map, TraverseOrder, ST);
3522
3523 if (!AddrSinkNewSelects && ST.countNewSelectNodes() > 0) {
3524 ST.destroyNewNodes(CommonType);
3525 return nullptr;
3526 }
3527
3528 // Now we'd like to match New Phi nodes to existed ones.
3529 unsigned PhiNotMatchedCount = 0;
3530 if (!MatchPhiSet(ST, AddrSinkNewPhis, PhiNotMatchedCount)) {
3531 ST.destroyNewNodes(CommonType);
3532 return nullptr;
3533 }
3534
3535 auto *Result = ST.Get(Map.find(Original)->second);
3536 if (Result) {
3537 NumMemoryInstsPhiCreated += ST.countNewPhiNodes() + PhiNotMatchedCount;
3538 NumMemoryInstsSelectCreated += ST.countNewSelectNodes();
3539 }
3540 return Result;
3541 }
3542
3543 /// Try to match PHI node to Candidate.
3544 /// Matcher tracks the matched Phi nodes.
MatchPhiNode(PHINode * PHI,PHINode * Candidate,SmallSetVector<PHIPair,8> & Matcher,PhiNodeSet & PhiNodesToMatch)3545 bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
3546 SmallSetVector<PHIPair, 8> &Matcher,
3547 PhiNodeSet &PhiNodesToMatch) {
3548 SmallVector<PHIPair, 8> WorkList;
3549 Matcher.insert({ PHI, Candidate });
3550 SmallSet<PHINode *, 8> MatchedPHIs;
3551 MatchedPHIs.insert(PHI);
3552 WorkList.push_back({ PHI, Candidate });
3553 SmallSet<PHIPair, 8> Visited;
3554 while (!WorkList.empty()) {
3555 auto Item = WorkList.pop_back_val();
3556 if (!Visited.insert(Item).second)
3557 continue;
3558 // We iterate over all incoming values to Phi to compare them.
3559 // If values are different and both of them Phi and the first one is a
3560 // Phi we added (subject to match) and both of them is in the same basic
3561 // block then we can match our pair if values match. So we state that
3562 // these values match and add it to work list to verify that.
3563 for (auto B : Item.first->blocks()) {
3564 Value *FirstValue = Item.first->getIncomingValueForBlock(B);
3565 Value *SecondValue = Item.second->getIncomingValueForBlock(B);
3566 if (FirstValue == SecondValue)
3567 continue;
3568
3569 PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue);
3570 PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue);
3571
3572 // One of them is not Phi or
3573 // The first one is not Phi node from the set we'd like to match or
3574 // Phi nodes from different basic blocks then
3575 // we will not be able to match.
3576 if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) ||
3577 FirstPhi->getParent() != SecondPhi->getParent())
3578 return false;
3579
3580 // If we already matched them then continue.
3581 if (Matcher.count({ FirstPhi, SecondPhi }))
3582 continue;
3583 // So the values are different and does not match. So we need them to
3584 // match. (But we register no more than one match per PHI node, so that
3585 // we won't later try to replace them twice.)
3586 if (MatchedPHIs.insert(FirstPhi).second)
3587 Matcher.insert({ FirstPhi, SecondPhi });
3588 // But me must check it.
3589 WorkList.push_back({ FirstPhi, SecondPhi });
3590 }
3591 }
3592 return true;
3593 }
3594
3595 /// For the given set of PHI nodes (in the SimplificationTracker) try
3596 /// to find their equivalents.
3597 /// Returns false if this matching fails and creation of new Phi is disabled.
MatchPhiSet(SimplificationTracker & ST,bool AllowNewPhiNodes,unsigned & PhiNotMatchedCount)3598 bool MatchPhiSet(SimplificationTracker &ST, bool AllowNewPhiNodes,
3599 unsigned &PhiNotMatchedCount) {
3600 // Matched and PhiNodesToMatch iterate their elements in a deterministic
3601 // order, so the replacements (ReplacePhi) are also done in a deterministic
3602 // order.
3603 SmallSetVector<PHIPair, 8> Matched;
3604 SmallPtrSet<PHINode *, 8> WillNotMatch;
3605 PhiNodeSet &PhiNodesToMatch = ST.newPhiNodes();
3606 while (PhiNodesToMatch.size()) {
3607 PHINode *PHI = *PhiNodesToMatch.begin();
3608
3609 // Add us, if no Phi nodes in the basic block we do not match.
3610 WillNotMatch.clear();
3611 WillNotMatch.insert(PHI);
3612
3613 // Traverse all Phis until we found equivalent or fail to do that.
3614 bool IsMatched = false;
3615 for (auto &P : PHI->getParent()->phis()) {
3616 if (&P == PHI)
3617 continue;
3618 if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch)))
3619 break;
3620 // If it does not match, collect all Phi nodes from matcher.
3621 // if we end up with no match, them all these Phi nodes will not match
3622 // later.
3623 for (auto M : Matched)
3624 WillNotMatch.insert(M.first);
3625 Matched.clear();
3626 }
3627 if (IsMatched) {
3628 // Replace all matched values and erase them.
3629 for (auto MV : Matched)
3630 ST.ReplacePhi(MV.first, MV.second);
3631 Matched.clear();
3632 continue;
3633 }
3634 // If we are not allowed to create new nodes then bail out.
3635 if (!AllowNewPhiNodes)
3636 return false;
3637 // Just remove all seen values in matcher. They will not match anything.
3638 PhiNotMatchedCount += WillNotMatch.size();
3639 for (auto *P : WillNotMatch)
3640 PhiNodesToMatch.erase(P);
3641 }
3642 return true;
3643 }
3644 /// Fill the placeholders with values from predecessors and simplify them.
FillPlaceholders(FoldAddrToValueMapping & Map,SmallVectorImpl<Value * > & TraverseOrder,SimplificationTracker & ST)3645 void FillPlaceholders(FoldAddrToValueMapping &Map,
3646 SmallVectorImpl<Value *> &TraverseOrder,
3647 SimplificationTracker &ST) {
3648 while (!TraverseOrder.empty()) {
3649 Value *Current = TraverseOrder.pop_back_val();
3650 assert(Map.find(Current) != Map.end() && "No node to fill!!!");
3651 Value *V = Map[Current];
3652
3653 if (SelectInst *Select = dyn_cast<SelectInst>(V)) {
3654 // CurrentValue also must be Select.
3655 auto *CurrentSelect = cast<SelectInst>(Current);
3656 auto *TrueValue = CurrentSelect->getTrueValue();
3657 assert(Map.find(TrueValue) != Map.end() && "No True Value!");
3658 Select->setTrueValue(ST.Get(Map[TrueValue]));
3659 auto *FalseValue = CurrentSelect->getFalseValue();
3660 assert(Map.find(FalseValue) != Map.end() && "No False Value!");
3661 Select->setFalseValue(ST.Get(Map[FalseValue]));
3662 } else {
3663 // Must be a Phi node then.
3664 auto *PHI = cast<PHINode>(V);
3665 // Fill the Phi node with values from predecessors.
3666 for (auto *B : predecessors(PHI->getParent())) {
3667 Value *PV = cast<PHINode>(Current)->getIncomingValueForBlock(B);
3668 assert(Map.find(PV) != Map.end() && "No predecessor Value!");
3669 PHI->addIncoming(ST.Get(Map[PV]), B);
3670 }
3671 }
3672 Map[Current] = ST.Simplify(V);
3673 }
3674 }
3675
3676 /// Starting from original value recursively iterates over def-use chain up to
3677 /// known ending values represented in a map. For each traversed phi/select
3678 /// inserts a placeholder Phi or Select.
3679 /// Reports all new created Phi/Select nodes by adding them to set.
3680 /// Also reports and order in what values have been traversed.
InsertPlaceholders(FoldAddrToValueMapping & Map,SmallVectorImpl<Value * > & TraverseOrder,SimplificationTracker & ST)3681 void InsertPlaceholders(FoldAddrToValueMapping &Map,
3682 SmallVectorImpl<Value *> &TraverseOrder,
3683 SimplificationTracker &ST) {
3684 SmallVector<Value *, 32> Worklist;
3685 assert((isa<PHINode>(Original) || isa<SelectInst>(Original)) &&
3686 "Address must be a Phi or Select node");
3687 auto *Dummy = UndefValue::get(CommonType);
3688 Worklist.push_back(Original);
3689 while (!Worklist.empty()) {
3690 Value *Current = Worklist.pop_back_val();
3691 // if it is already visited or it is an ending value then skip it.
3692 if (Map.find(Current) != Map.end())
3693 continue;
3694 TraverseOrder.push_back(Current);
3695
3696 // CurrentValue must be a Phi node or select. All others must be covered
3697 // by anchors.
3698 if (SelectInst *CurrentSelect = dyn_cast<SelectInst>(Current)) {
3699 // Is it OK to get metadata from OrigSelect?!
3700 // Create a Select placeholder with dummy value.
3701 SelectInst *Select = SelectInst::Create(
3702 CurrentSelect->getCondition(), Dummy, Dummy,
3703 CurrentSelect->getName(), CurrentSelect, CurrentSelect);
3704 Map[Current] = Select;
3705 ST.insertNewSelect(Select);
3706 // We are interested in True and False values.
3707 Worklist.push_back(CurrentSelect->getTrueValue());
3708 Worklist.push_back(CurrentSelect->getFalseValue());
3709 } else {
3710 // It must be a Phi node then.
3711 PHINode *CurrentPhi = cast<PHINode>(Current);
3712 unsigned PredCount = CurrentPhi->getNumIncomingValues();
3713 PHINode *PHI =
3714 PHINode::Create(CommonType, PredCount, "sunk_phi", CurrentPhi);
3715 Map[Current] = PHI;
3716 ST.insertNewPhi(PHI);
3717 append_range(Worklist, CurrentPhi->incoming_values());
3718 }
3719 }
3720 }
3721
addrModeCombiningAllowed()3722 bool addrModeCombiningAllowed() {
3723 if (DisableComplexAddrModes)
3724 return false;
3725 switch (DifferentField) {
3726 default:
3727 return false;
3728 case ExtAddrMode::BaseRegField:
3729 return AddrSinkCombineBaseReg;
3730 case ExtAddrMode::BaseGVField:
3731 return AddrSinkCombineBaseGV;
3732 case ExtAddrMode::BaseOffsField:
3733 return AddrSinkCombineBaseOffs;
3734 case ExtAddrMode::ScaledRegField:
3735 return AddrSinkCombineScaledReg;
3736 }
3737 }
3738 };
3739 } // end anonymous namespace
3740
3741 /// Try adding ScaleReg*Scale to the current addressing mode.
3742 /// Return true and update AddrMode if this addr mode is legal for the target,
3743 /// false if not.
matchScaledValue(Value * ScaleReg,int64_t Scale,unsigned Depth)3744 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3745 unsigned Depth) {
3746 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3747 // mode. Just process that directly.
3748 if (Scale == 1)
3749 return matchAddr(ScaleReg, Depth);
3750
3751 // If the scale is 0, it takes nothing to add this.
3752 if (Scale == 0)
3753 return true;
3754
3755 // If we already have a scale of this value, we can add to it, otherwise, we
3756 // need an available scale field.
3757 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3758 return false;
3759
3760 ExtAddrMode TestAddrMode = AddrMode;
3761
3762 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3763 // [A+B + A*7] -> [B+A*8].
3764 TestAddrMode.Scale += Scale;
3765 TestAddrMode.ScaledReg = ScaleReg;
3766
3767 // If the new address isn't legal, bail out.
3768 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3769 return false;
3770
3771 // It was legal, so commit it.
3772 AddrMode = TestAddrMode;
3773
3774 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3775 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3776 // X*Scale + C*Scale to addr mode.
3777 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3778 if (isa<Instruction>(ScaleReg) && // not a constant expr.
3779 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI))) &&
3780 CI->getValue().isSignedIntN(64)) {
3781 TestAddrMode.InBounds = false;
3782 TestAddrMode.ScaledReg = AddLHS;
3783 TestAddrMode.BaseOffs += CI->getSExtValue() * TestAddrMode.Scale;
3784
3785 // If this addressing mode is legal, commit it and remember that we folded
3786 // this instruction.
3787 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3788 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3789 AddrMode = TestAddrMode;
3790 return true;
3791 }
3792 }
3793
3794 // Otherwise, not (x+c)*scale, just return what we have.
3795 return true;
3796 }
3797
3798 /// This is a little filter, which returns true if an addressing computation
3799 /// involving I might be folded into a load/store accessing it.
3800 /// This doesn't need to be perfect, but needs to accept at least
3801 /// the set of instructions that MatchOperationAddr can.
MightBeFoldableInst(Instruction * I)3802 static bool MightBeFoldableInst(Instruction *I) {
3803 switch (I->getOpcode()) {
3804 case Instruction::BitCast:
3805 case Instruction::AddrSpaceCast:
3806 // Don't touch identity bitcasts.
3807 if (I->getType() == I->getOperand(0)->getType())
3808 return false;
3809 return I->getType()->isIntOrPtrTy();
3810 case Instruction::PtrToInt:
3811 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3812 return true;
3813 case Instruction::IntToPtr:
3814 // We know the input is intptr_t, so this is foldable.
3815 return true;
3816 case Instruction::Add:
3817 return true;
3818 case Instruction::Mul:
3819 case Instruction::Shl:
3820 // Can only handle X*C and X << C.
3821 return isa<ConstantInt>(I->getOperand(1));
3822 case Instruction::GetElementPtr:
3823 return true;
3824 default:
3825 return false;
3826 }
3827 }
3828
3829 /// Check whether or not \p Val is a legal instruction for \p TLI.
3830 /// \note \p Val is assumed to be the product of some type promotion.
3831 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3832 /// to be legal, as the non-promoted value would have had the same state.
isPromotedInstructionLegal(const TargetLowering & TLI,const DataLayout & DL,Value * Val)3833 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
3834 const DataLayout &DL, Value *Val) {
3835 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3836 if (!PromotedInst)
3837 return false;
3838 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3839 // If the ISDOpcode is undefined, it was undefined before the promotion.
3840 if (!ISDOpcode)
3841 return true;
3842 // Otherwise, check if the promoted instruction is legal or not.
3843 return TLI.isOperationLegalOrCustom(
3844 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3845 }
3846
3847 namespace {
3848
3849 /// Hepler class to perform type promotion.
3850 class TypePromotionHelper {
3851 /// Utility function to add a promoted instruction \p ExtOpnd to
3852 /// \p PromotedInsts and record the type of extension we have seen.
addPromotedInst(InstrToOrigTy & PromotedInsts,Instruction * ExtOpnd,bool IsSExt)3853 static void addPromotedInst(InstrToOrigTy &PromotedInsts,
3854 Instruction *ExtOpnd,
3855 bool IsSExt) {
3856 ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
3857 InstrToOrigTy::iterator It = PromotedInsts.find(ExtOpnd);
3858 if (It != PromotedInsts.end()) {
3859 // If the new extension is same as original, the information in
3860 // PromotedInsts[ExtOpnd] is still correct.
3861 if (It->second.getInt() == ExtTy)
3862 return;
3863
3864 // Now the new extension is different from old extension, we make
3865 // the type information invalid by setting extension type to
3866 // BothExtension.
3867 ExtTy = BothExtension;
3868 }
3869 PromotedInsts[ExtOpnd] = TypeIsSExt(ExtOpnd->getType(), ExtTy);
3870 }
3871
3872 /// Utility function to query the original type of instruction \p Opnd
3873 /// with a matched extension type. If the extension doesn't match, we
3874 /// cannot use the information we had on the original type.
3875 /// BothExtension doesn't match any extension type.
getOrigType(const InstrToOrigTy & PromotedInsts,Instruction * Opnd,bool IsSExt)3876 static const Type *getOrigType(const InstrToOrigTy &PromotedInsts,
3877 Instruction *Opnd,
3878 bool IsSExt) {
3879 ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
3880 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3881 if (It != PromotedInsts.end() && It->second.getInt() == ExtTy)
3882 return It->second.getPointer();
3883 return nullptr;
3884 }
3885
3886 /// Utility function to check whether or not a sign or zero extension
3887 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3888 /// either using the operands of \p Inst or promoting \p Inst.
3889 /// The type of the extension is defined by \p IsSExt.
3890 /// In other words, check if:
3891 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3892 /// #1 Promotion applies:
3893 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3894 /// #2 Operand reuses:
3895 /// ext opnd1 to ConsideredExtType.
3896 /// \p PromotedInsts maps the instructions to their type before promotion.
3897 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3898 const InstrToOrigTy &PromotedInsts, bool IsSExt);
3899
3900 /// Utility function to determine if \p OpIdx should be promoted when
3901 /// promoting \p Inst.
shouldExtOperand(const Instruction * Inst,int OpIdx)3902 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3903 return !(isa<SelectInst>(Inst) && OpIdx == 0);
3904 }
3905
3906 /// Utility function to promote the operand of \p Ext when this
3907 /// operand is a promotable trunc or sext or zext.
3908 /// \p PromotedInsts maps the instructions to their type before promotion.
3909 /// \p CreatedInstsCost[out] contains the cost of all instructions
3910 /// created to promote the operand of Ext.
3911 /// Newly added extensions are inserted in \p Exts.
3912 /// Newly added truncates are inserted in \p Truncs.
3913 /// Should never be called directly.
3914 /// \return The promoted value which is used instead of Ext.
3915 static Value *promoteOperandForTruncAndAnyExt(
3916 Instruction *Ext, TypePromotionTransaction &TPT,
3917 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3918 SmallVectorImpl<Instruction *> *Exts,
3919 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3920
3921 /// Utility function to promote the operand of \p Ext when this
3922 /// operand is promotable and is not a supported trunc or sext.
3923 /// \p PromotedInsts maps the instructions to their type before promotion.
3924 /// \p CreatedInstsCost[out] contains the cost of all the instructions
3925 /// created to promote the operand of Ext.
3926 /// Newly added extensions are inserted in \p Exts.
3927 /// Newly added truncates are inserted in \p Truncs.
3928 /// Should never be called directly.
3929 /// \return The promoted value which is used instead of Ext.
3930 static Value *promoteOperandForOther(Instruction *Ext,
3931 TypePromotionTransaction &TPT,
3932 InstrToOrigTy &PromotedInsts,
3933 unsigned &CreatedInstsCost,
3934 SmallVectorImpl<Instruction *> *Exts,
3935 SmallVectorImpl<Instruction *> *Truncs,
3936 const TargetLowering &TLI, bool IsSExt);
3937
3938 /// \see promoteOperandForOther.
signExtendOperandForOther(Instruction * Ext,TypePromotionTransaction & TPT,InstrToOrigTy & PromotedInsts,unsigned & CreatedInstsCost,SmallVectorImpl<Instruction * > * Exts,SmallVectorImpl<Instruction * > * Truncs,const TargetLowering & TLI)3939 static Value *signExtendOperandForOther(
3940 Instruction *Ext, TypePromotionTransaction &TPT,
3941 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3942 SmallVectorImpl<Instruction *> *Exts,
3943 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3944 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3945 Exts, Truncs, TLI, true);
3946 }
3947
3948 /// \see promoteOperandForOther.
zeroExtendOperandForOther(Instruction * Ext,TypePromotionTransaction & TPT,InstrToOrigTy & PromotedInsts,unsigned & CreatedInstsCost,SmallVectorImpl<Instruction * > * Exts,SmallVectorImpl<Instruction * > * Truncs,const TargetLowering & TLI)3949 static Value *zeroExtendOperandForOther(
3950 Instruction *Ext, TypePromotionTransaction &TPT,
3951 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3952 SmallVectorImpl<Instruction *> *Exts,
3953 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3954 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3955 Exts, Truncs, TLI, false);
3956 }
3957
3958 public:
3959 /// Type for the utility function that promotes the operand of Ext.
3960 using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
3961 InstrToOrigTy &PromotedInsts,
3962 unsigned &CreatedInstsCost,
3963 SmallVectorImpl<Instruction *> *Exts,
3964 SmallVectorImpl<Instruction *> *Truncs,
3965 const TargetLowering &TLI);
3966
3967 /// Given a sign/zero extend instruction \p Ext, return the appropriate
3968 /// action to promote the operand of \p Ext instead of using Ext.
3969 /// \return NULL if no promotable action is possible with the current
3970 /// sign extension.
3971 /// \p InsertedInsts keeps track of all the instructions inserted by the
3972 /// other CodeGenPrepare optimizations. This information is important
3973 /// because we do not want to promote these instructions as CodeGenPrepare
3974 /// will reinsert them later. Thus creating an infinite loop: create/remove.
3975 /// \p PromotedInsts maps the instructions to their type before promotion.
3976 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3977 const TargetLowering &TLI,
3978 const InstrToOrigTy &PromotedInsts);
3979 };
3980
3981 } // end anonymous namespace
3982
canGetThrough(const Instruction * Inst,Type * ConsideredExtType,const InstrToOrigTy & PromotedInsts,bool IsSExt)3983 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3984 Type *ConsideredExtType,
3985 const InstrToOrigTy &PromotedInsts,
3986 bool IsSExt) {
3987 // The promotion helper does not know how to deal with vector types yet.
3988 // To be able to fix that, we would need to fix the places where we
3989 // statically extend, e.g., constants and such.
3990 if (Inst->getType()->isVectorTy())
3991 return false;
3992
3993 // We can always get through zext.
3994 if (isa<ZExtInst>(Inst))
3995 return true;
3996
3997 // sext(sext) is ok too.
3998 if (IsSExt && isa<SExtInst>(Inst))
3999 return true;
4000
4001 // We can get through binary operator, if it is legal. In other words, the
4002 // binary operator must have a nuw or nsw flag.
4003 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
4004 if (isa_and_nonnull<OverflowingBinaryOperator>(BinOp) &&
4005 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
4006 (IsSExt && BinOp->hasNoSignedWrap())))
4007 return true;
4008
4009 // ext(and(opnd, cst)) --> and(ext(opnd), ext(cst))
4010 if ((Inst->getOpcode() == Instruction::And ||
4011 Inst->getOpcode() == Instruction::Or))
4012 return true;
4013
4014 // ext(xor(opnd, cst)) --> xor(ext(opnd), ext(cst))
4015 if (Inst->getOpcode() == Instruction::Xor) {
4016 const ConstantInt *Cst = dyn_cast<ConstantInt>(Inst->getOperand(1));
4017 // Make sure it is not a NOT.
4018 if (Cst && !Cst->getValue().isAllOnesValue())
4019 return true;
4020 }
4021
4022 // zext(shrl(opnd, cst)) --> shrl(zext(opnd), zext(cst))
4023 // It may change a poisoned value into a regular value, like
4024 // zext i32 (shrl i8 %val, 12) --> shrl i32 (zext i8 %val), 12
4025 // poisoned value regular value
4026 // It should be OK since undef covers valid value.
4027 if (Inst->getOpcode() == Instruction::LShr && !IsSExt)
4028 return true;
4029
4030 // and(ext(shl(opnd, cst)), cst) --> and(shl(ext(opnd), ext(cst)), cst)
4031 // It may change a poisoned value into a regular value, like
4032 // zext i32 (shl i8 %val, 12) --> shl i32 (zext i8 %val), 12
4033 // poisoned value regular value
4034 // It should be OK since undef covers valid value.
4035 if (Inst->getOpcode() == Instruction::Shl && Inst->hasOneUse()) {
4036 const auto *ExtInst = cast<const Instruction>(*Inst->user_begin());
4037 if (ExtInst->hasOneUse()) {
4038 const auto *AndInst = dyn_cast<const Instruction>(*ExtInst->user_begin());
4039 if (AndInst && AndInst->getOpcode() == Instruction::And) {
4040 const auto *Cst = dyn_cast<ConstantInt>(AndInst->getOperand(1));
4041 if (Cst &&
4042 Cst->getValue().isIntN(Inst->getType()->getIntegerBitWidth()))
4043 return true;
4044 }
4045 }
4046 }
4047
4048 // Check if we can do the following simplification.
4049 // ext(trunc(opnd)) --> ext(opnd)
4050 if (!isa<TruncInst>(Inst))
4051 return false;
4052
4053 Value *OpndVal = Inst->getOperand(0);
4054 // Check if we can use this operand in the extension.
4055 // If the type is larger than the result type of the extension, we cannot.
4056 if (!OpndVal->getType()->isIntegerTy() ||
4057 OpndVal->getType()->getIntegerBitWidth() >
4058 ConsideredExtType->getIntegerBitWidth())
4059 return false;
4060
4061 // If the operand of the truncate is not an instruction, we will not have
4062 // any information on the dropped bits.
4063 // (Actually we could for constant but it is not worth the extra logic).
4064 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
4065 if (!Opnd)
4066 return false;
4067
4068 // Check if the source of the type is narrow enough.
4069 // I.e., check that trunc just drops extended bits of the same kind of
4070 // the extension.
4071 // #1 get the type of the operand and check the kind of the extended bits.
4072 const Type *OpndType = getOrigType(PromotedInsts, Opnd, IsSExt);
4073 if (OpndType)
4074 ;
4075 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
4076 OpndType = Opnd->getOperand(0)->getType();
4077 else
4078 return false;
4079
4080 // #2 check that the truncate just drops extended bits.
4081 return Inst->getType()->getIntegerBitWidth() >=
4082 OpndType->getIntegerBitWidth();
4083 }
4084
getAction(Instruction * Ext,const SetOfInstrs & InsertedInsts,const TargetLowering & TLI,const InstrToOrigTy & PromotedInsts)4085 TypePromotionHelper::Action TypePromotionHelper::getAction(
4086 Instruction *Ext, const SetOfInstrs &InsertedInsts,
4087 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
4088 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
4089 "Unexpected instruction type");
4090 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
4091 Type *ExtTy = Ext->getType();
4092 bool IsSExt = isa<SExtInst>(Ext);
4093 // If the operand of the extension is not an instruction, we cannot
4094 // get through.
4095 // If it, check we can get through.
4096 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
4097 return nullptr;
4098
4099 // Do not promote if the operand has been added by codegenprepare.
4100 // Otherwise, it means we are undoing an optimization that is likely to be
4101 // redone, thus causing potential infinite loop.
4102 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
4103 return nullptr;
4104
4105 // SExt or Trunc instructions.
4106 // Return the related handler.
4107 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
4108 isa<ZExtInst>(ExtOpnd))
4109 return promoteOperandForTruncAndAnyExt;
4110
4111 // Regular instruction.
4112 // Abort early if we will have to insert non-free instructions.
4113 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
4114 return nullptr;
4115 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
4116 }
4117
promoteOperandForTruncAndAnyExt(Instruction * SExt,TypePromotionTransaction & TPT,InstrToOrigTy & PromotedInsts,unsigned & CreatedInstsCost,SmallVectorImpl<Instruction * > * Exts,SmallVectorImpl<Instruction * > * Truncs,const TargetLowering & TLI)4118 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
4119 Instruction *SExt, TypePromotionTransaction &TPT,
4120 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4121 SmallVectorImpl<Instruction *> *Exts,
4122 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4123 // By construction, the operand of SExt is an instruction. Otherwise we cannot
4124 // get through it and this method should not be called.
4125 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
4126 Value *ExtVal = SExt;
4127 bool HasMergedNonFreeExt = false;
4128 if (isa<ZExtInst>(SExtOpnd)) {
4129 // Replace s|zext(zext(opnd))
4130 // => zext(opnd).
4131 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
4132 Value *ZExt =
4133 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
4134 TPT.replaceAllUsesWith(SExt, ZExt);
4135 TPT.eraseInstruction(SExt);
4136 ExtVal = ZExt;
4137 } else {
4138 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
4139 // => z|sext(opnd).
4140 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
4141 }
4142 CreatedInstsCost = 0;
4143
4144 // Remove dead code.
4145 if (SExtOpnd->use_empty())
4146 TPT.eraseInstruction(SExtOpnd);
4147
4148 // Check if the extension is still needed.
4149 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
4150 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
4151 if (ExtInst) {
4152 if (Exts)
4153 Exts->push_back(ExtInst);
4154 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
4155 }
4156 return ExtVal;
4157 }
4158
4159 // At this point we have: ext ty opnd to ty.
4160 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
4161 Value *NextVal = ExtInst->getOperand(0);
4162 TPT.eraseInstruction(ExtInst, NextVal);
4163 return NextVal;
4164 }
4165
promoteOperandForOther(Instruction * Ext,TypePromotionTransaction & TPT,InstrToOrigTy & PromotedInsts,unsigned & CreatedInstsCost,SmallVectorImpl<Instruction * > * Exts,SmallVectorImpl<Instruction * > * Truncs,const TargetLowering & TLI,bool IsSExt)4166 Value *TypePromotionHelper::promoteOperandForOther(
4167 Instruction *Ext, TypePromotionTransaction &TPT,
4168 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4169 SmallVectorImpl<Instruction *> *Exts,
4170 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
4171 bool IsSExt) {
4172 // By construction, the operand of Ext is an instruction. Otherwise we cannot
4173 // get through it and this method should not be called.
4174 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
4175 CreatedInstsCost = 0;
4176 if (!ExtOpnd->hasOneUse()) {
4177 // ExtOpnd will be promoted.
4178 // All its uses, but Ext, will need to use a truncated value of the
4179 // promoted version.
4180 // Create the truncate now.
4181 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
4182 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
4183 // Insert it just after the definition.
4184 ITrunc->moveAfter(ExtOpnd);
4185 if (Truncs)
4186 Truncs->push_back(ITrunc);
4187 }
4188
4189 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
4190 // Restore the operand of Ext (which has been replaced by the previous call
4191 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
4192 TPT.setOperand(Ext, 0, ExtOpnd);
4193 }
4194
4195 // Get through the Instruction:
4196 // 1. Update its type.
4197 // 2. Replace the uses of Ext by Inst.
4198 // 3. Extend each operand that needs to be extended.
4199
4200 // Remember the original type of the instruction before promotion.
4201 // This is useful to know that the high bits are sign extended bits.
4202 addPromotedInst(PromotedInsts, ExtOpnd, IsSExt);
4203 // Step #1.
4204 TPT.mutateType(ExtOpnd, Ext->getType());
4205 // Step #2.
4206 TPT.replaceAllUsesWith(Ext, ExtOpnd);
4207 // Step #3.
4208 Instruction *ExtForOpnd = Ext;
4209
4210 LLVM_DEBUG(dbgs() << "Propagate Ext to operands\n");
4211 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
4212 ++OpIdx) {
4213 LLVM_DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
4214 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
4215 !shouldExtOperand(ExtOpnd, OpIdx)) {
4216 LLVM_DEBUG(dbgs() << "No need to propagate\n");
4217 continue;
4218 }
4219 // Check if we can statically extend the operand.
4220 Value *Opnd = ExtOpnd->getOperand(OpIdx);
4221 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
4222 LLVM_DEBUG(dbgs() << "Statically extend\n");
4223 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
4224 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
4225 : Cst->getValue().zext(BitWidth);
4226 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
4227 continue;
4228 }
4229 // UndefValue are typed, so we have to statically sign extend them.
4230 if (isa<UndefValue>(Opnd)) {
4231 LLVM_DEBUG(dbgs() << "Statically extend\n");
4232 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
4233 continue;
4234 }
4235
4236 // Otherwise we have to explicitly sign extend the operand.
4237 // Check if Ext was reused to extend an operand.
4238 if (!ExtForOpnd) {
4239 // If yes, create a new one.
4240 LLVM_DEBUG(dbgs() << "More operands to ext\n");
4241 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
4242 : TPT.createZExt(Ext, Opnd, Ext->getType());
4243 if (!isa<Instruction>(ValForExtOpnd)) {
4244 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
4245 continue;
4246 }
4247 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
4248 }
4249 if (Exts)
4250 Exts->push_back(ExtForOpnd);
4251 TPT.setOperand(ExtForOpnd, 0, Opnd);
4252
4253 // Move the sign extension before the insertion point.
4254 TPT.moveBefore(ExtForOpnd, ExtOpnd);
4255 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
4256 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
4257 // If more sext are required, new instructions will have to be created.
4258 ExtForOpnd = nullptr;
4259 }
4260 if (ExtForOpnd == Ext) {
4261 LLVM_DEBUG(dbgs() << "Extension is useless now\n");
4262 TPT.eraseInstruction(Ext);
4263 }
4264 return ExtOpnd;
4265 }
4266
4267 /// Check whether or not promoting an instruction to a wider type is profitable.
4268 /// \p NewCost gives the cost of extension instructions created by the
4269 /// promotion.
4270 /// \p OldCost gives the cost of extension instructions before the promotion
4271 /// plus the number of instructions that have been
4272 /// matched in the addressing mode the promotion.
4273 /// \p PromotedOperand is the value that has been promoted.
4274 /// \return True if the promotion is profitable, false otherwise.
isPromotionProfitable(unsigned NewCost,unsigned OldCost,Value * PromotedOperand) const4275 bool AddressingModeMatcher::isPromotionProfitable(
4276 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
4277 LLVM_DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost
4278 << '\n');
4279 // The cost of the new extensions is greater than the cost of the
4280 // old extension plus what we folded.
4281 // This is not profitable.
4282 if (NewCost > OldCost)
4283 return false;
4284 if (NewCost < OldCost)
4285 return true;
4286 // The promotion is neutral but it may help folding the sign extension in
4287 // loads for instance.
4288 // Check that we did not create an illegal instruction.
4289 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
4290 }
4291
4292 /// Given an instruction or constant expr, see if we can fold the operation
4293 /// into the addressing mode. If so, update the addressing mode and return
4294 /// true, otherwise return false without modifying AddrMode.
4295 /// If \p MovedAway is not NULL, it contains the information of whether or
4296 /// not AddrInst has to be folded into the addressing mode on success.
4297 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
4298 /// because it has been moved away.
4299 /// Thus AddrInst must not be added in the matched instructions.
4300 /// This state can happen when AddrInst is a sext, since it may be moved away.
4301 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
4302 /// not be referenced anymore.
matchOperationAddr(User * AddrInst,unsigned Opcode,unsigned Depth,bool * MovedAway)4303 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
4304 unsigned Depth,
4305 bool *MovedAway) {
4306 // Avoid exponential behavior on extremely deep expression trees.
4307 if (Depth >= 5) return false;
4308
4309 // By default, all matched instructions stay in place.
4310 if (MovedAway)
4311 *MovedAway = false;
4312
4313 switch (Opcode) {
4314 case Instruction::PtrToInt:
4315 // PtrToInt is always a noop, as we know that the int type is pointer sized.
4316 return matchAddr(AddrInst->getOperand(0), Depth);
4317 case Instruction::IntToPtr: {
4318 auto AS = AddrInst->getType()->getPointerAddressSpace();
4319 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
4320 // This inttoptr is a no-op if the integer type is pointer sized.
4321 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
4322 return matchAddr(AddrInst->getOperand(0), Depth);
4323 return false;
4324 }
4325 case Instruction::BitCast:
4326 // BitCast is always a noop, and we can handle it as long as it is
4327 // int->int or pointer->pointer (we don't want int<->fp or something).
4328 if (AddrInst->getOperand(0)->getType()->isIntOrPtrTy() &&
4329 // Don't touch identity bitcasts. These were probably put here by LSR,
4330 // and we don't want to mess around with them. Assume it knows what it
4331 // is doing.
4332 AddrInst->getOperand(0)->getType() != AddrInst->getType())
4333 return matchAddr(AddrInst->getOperand(0), Depth);
4334 return false;
4335 case Instruction::AddrSpaceCast: {
4336 unsigned SrcAS
4337 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
4338 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
4339 if (TLI.getTargetMachine().isNoopAddrSpaceCast(SrcAS, DestAS))
4340 return matchAddr(AddrInst->getOperand(0), Depth);
4341 return false;
4342 }
4343 case Instruction::Add: {
4344 // Check to see if we can merge in the RHS then the LHS. If so, we win.
4345 ExtAddrMode BackupAddrMode = AddrMode;
4346 unsigned OldSize = AddrModeInsts.size();
4347 // Start a transaction at this point.
4348 // The LHS may match but not the RHS.
4349 // Therefore, we need a higher level restoration point to undo partially
4350 // matched operation.
4351 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4352 TPT.getRestorationPoint();
4353
4354 AddrMode.InBounds = false;
4355 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
4356 matchAddr(AddrInst->getOperand(0), Depth+1))
4357 return true;
4358
4359 // Restore the old addr mode info.
4360 AddrMode = BackupAddrMode;
4361 AddrModeInsts.resize(OldSize);
4362 TPT.rollback(LastKnownGood);
4363
4364 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
4365 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
4366 matchAddr(AddrInst->getOperand(1), Depth+1))
4367 return true;
4368
4369 // Otherwise we definitely can't merge the ADD in.
4370 AddrMode = BackupAddrMode;
4371 AddrModeInsts.resize(OldSize);
4372 TPT.rollback(LastKnownGood);
4373 break;
4374 }
4375 //case Instruction::Or:
4376 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
4377 //break;
4378 case Instruction::Mul:
4379 case Instruction::Shl: {
4380 // Can only handle X*C and X << C.
4381 AddrMode.InBounds = false;
4382 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
4383 if (!RHS || RHS->getBitWidth() > 64)
4384 return false;
4385 int64_t Scale = RHS->getSExtValue();
4386 if (Opcode == Instruction::Shl)
4387 Scale = 1LL << Scale;
4388
4389 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
4390 }
4391 case Instruction::GetElementPtr: {
4392 // Scan the GEP. We check it if it contains constant offsets and at most
4393 // one variable offset.
4394 int VariableOperand = -1;
4395 unsigned VariableScale = 0;
4396
4397 int64_t ConstantOffset = 0;
4398 gep_type_iterator GTI = gep_type_begin(AddrInst);
4399 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
4400 if (StructType *STy = GTI.getStructTypeOrNull()) {
4401 const StructLayout *SL = DL.getStructLayout(STy);
4402 unsigned Idx =
4403 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
4404 ConstantOffset += SL->getElementOffset(Idx);
4405 } else {
4406 TypeSize TS = DL.getTypeAllocSize(GTI.getIndexedType());
4407 if (TS.isNonZero()) {
4408 // The optimisations below currently only work for fixed offsets.
4409 if (TS.isScalable())
4410 return false;
4411 int64_t TypeSize = TS.getFixedSize();
4412 if (ConstantInt *CI =
4413 dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
4414 const APInt &CVal = CI->getValue();
4415 if (CVal.getMinSignedBits() <= 64) {
4416 ConstantOffset += CVal.getSExtValue() * TypeSize;
4417 continue;
4418 }
4419 }
4420 // We only allow one variable index at the moment.
4421 if (VariableOperand != -1)
4422 return false;
4423
4424 // Remember the variable index.
4425 VariableOperand = i;
4426 VariableScale = TypeSize;
4427 }
4428 }
4429 }
4430
4431 // A common case is for the GEP to only do a constant offset. In this case,
4432 // just add it to the disp field and check validity.
4433 if (VariableOperand == -1) {
4434 AddrMode.BaseOffs += ConstantOffset;
4435 if (ConstantOffset == 0 ||
4436 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
4437 // Check to see if we can fold the base pointer in too.
4438 if (matchAddr(AddrInst->getOperand(0), Depth+1)) {
4439 if (!cast<GEPOperator>(AddrInst)->isInBounds())
4440 AddrMode.InBounds = false;
4441 return true;
4442 }
4443 } else if (EnableGEPOffsetSplit && isa<GetElementPtrInst>(AddrInst) &&
4444 TLI.shouldConsiderGEPOffsetSplit() && Depth == 0 &&
4445 ConstantOffset > 0) {
4446 // Record GEPs with non-zero offsets as candidates for splitting in the
4447 // event that the offset cannot fit into the r+i addressing mode.
4448 // Simple and common case that only one GEP is used in calculating the
4449 // address for the memory access.
4450 Value *Base = AddrInst->getOperand(0);
4451 auto *BaseI = dyn_cast<Instruction>(Base);
4452 auto *GEP = cast<GetElementPtrInst>(AddrInst);
4453 if (isa<Argument>(Base) || isa<GlobalValue>(Base) ||
4454 (BaseI && !isa<CastInst>(BaseI) &&
4455 !isa<GetElementPtrInst>(BaseI))) {
4456 // Make sure the parent block allows inserting non-PHI instructions
4457 // before the terminator.
4458 BasicBlock *Parent =
4459 BaseI ? BaseI->getParent() : &GEP->getFunction()->getEntryBlock();
4460 if (!Parent->getTerminator()->isEHPad())
4461 LargeOffsetGEP = std::make_pair(GEP, ConstantOffset);
4462 }
4463 }
4464 AddrMode.BaseOffs -= ConstantOffset;
4465 return false;
4466 }
4467
4468 // Save the valid addressing mode in case we can't match.
4469 ExtAddrMode BackupAddrMode = AddrMode;
4470 unsigned OldSize = AddrModeInsts.size();
4471
4472 // See if the scale and offset amount is valid for this target.
4473 AddrMode.BaseOffs += ConstantOffset;
4474 if (!cast<GEPOperator>(AddrInst)->isInBounds())
4475 AddrMode.InBounds = false;
4476
4477 // Match the base operand of the GEP.
4478 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
4479 // If it couldn't be matched, just stuff the value in a register.
4480 if (AddrMode.HasBaseReg) {
4481 AddrMode = BackupAddrMode;
4482 AddrModeInsts.resize(OldSize);
4483 return false;
4484 }
4485 AddrMode.HasBaseReg = true;
4486 AddrMode.BaseReg = AddrInst->getOperand(0);
4487 }
4488
4489 // Match the remaining variable portion of the GEP.
4490 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
4491 Depth)) {
4492 // If it couldn't be matched, try stuffing the base into a register
4493 // instead of matching it, and retrying the match of the scale.
4494 AddrMode = BackupAddrMode;
4495 AddrModeInsts.resize(OldSize);
4496 if (AddrMode.HasBaseReg)
4497 return false;
4498 AddrMode.HasBaseReg = true;
4499 AddrMode.BaseReg = AddrInst->getOperand(0);
4500 AddrMode.BaseOffs += ConstantOffset;
4501 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
4502 VariableScale, Depth)) {
4503 // If even that didn't work, bail.
4504 AddrMode = BackupAddrMode;
4505 AddrModeInsts.resize(OldSize);
4506 return false;
4507 }
4508 }
4509
4510 return true;
4511 }
4512 case Instruction::SExt:
4513 case Instruction::ZExt: {
4514 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
4515 if (!Ext)
4516 return false;
4517
4518 // Try to move this ext out of the way of the addressing mode.
4519 // Ask for a method for doing so.
4520 TypePromotionHelper::Action TPH =
4521 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
4522 if (!TPH)
4523 return false;
4524
4525 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4526 TPT.getRestorationPoint();
4527 unsigned CreatedInstsCost = 0;
4528 unsigned ExtCost = !TLI.isExtFree(Ext);
4529 Value *PromotedOperand =
4530 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
4531 // SExt has been moved away.
4532 // Thus either it will be rematched later in the recursive calls or it is
4533 // gone. Anyway, we must not fold it into the addressing mode at this point.
4534 // E.g.,
4535 // op = add opnd, 1
4536 // idx = ext op
4537 // addr = gep base, idx
4538 // is now:
4539 // promotedOpnd = ext opnd <- no match here
4540 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
4541 // addr = gep base, op <- match
4542 if (MovedAway)
4543 *MovedAway = true;
4544
4545 assert(PromotedOperand &&
4546 "TypePromotionHelper should have filtered out those cases");
4547
4548 ExtAddrMode BackupAddrMode = AddrMode;
4549 unsigned OldSize = AddrModeInsts.size();
4550
4551 if (!matchAddr(PromotedOperand, Depth) ||
4552 // The total of the new cost is equal to the cost of the created
4553 // instructions.
4554 // The total of the old cost is equal to the cost of the extension plus
4555 // what we have saved in the addressing mode.
4556 !isPromotionProfitable(CreatedInstsCost,
4557 ExtCost + (AddrModeInsts.size() - OldSize),
4558 PromotedOperand)) {
4559 AddrMode = BackupAddrMode;
4560 AddrModeInsts.resize(OldSize);
4561 LLVM_DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
4562 TPT.rollback(LastKnownGood);
4563 return false;
4564 }
4565 return true;
4566 }
4567 }
4568 return false;
4569 }
4570
4571 /// If we can, try to add the value of 'Addr' into the current addressing mode.
4572 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
4573 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
4574 /// for the target.
4575 ///
matchAddr(Value * Addr,unsigned Depth)4576 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
4577 // Start a transaction at this point that we will rollback if the matching
4578 // fails.
4579 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4580 TPT.getRestorationPoint();
4581 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
4582 if (CI->getValue().isSignedIntN(64)) {
4583 // Fold in immediates if legal for the target.
4584 AddrMode.BaseOffs += CI->getSExtValue();
4585 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4586 return true;
4587 AddrMode.BaseOffs -= CI->getSExtValue();
4588 }
4589 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
4590 // If this is a global variable, try to fold it into the addressing mode.
4591 if (!AddrMode.BaseGV) {
4592 AddrMode.BaseGV = GV;
4593 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4594 return true;
4595 AddrMode.BaseGV = nullptr;
4596 }
4597 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
4598 ExtAddrMode BackupAddrMode = AddrMode;
4599 unsigned OldSize = AddrModeInsts.size();
4600
4601 // Check to see if it is possible to fold this operation.
4602 bool MovedAway = false;
4603 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
4604 // This instruction may have been moved away. If so, there is nothing
4605 // to check here.
4606 if (MovedAway)
4607 return true;
4608 // Okay, it's possible to fold this. Check to see if it is actually
4609 // *profitable* to do so. We use a simple cost model to avoid increasing
4610 // register pressure too much.
4611 if (I->hasOneUse() ||
4612 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
4613 AddrModeInsts.push_back(I);
4614 return true;
4615 }
4616
4617 // It isn't profitable to do this, roll back.
4618 //cerr << "NOT FOLDING: " << *I;
4619 AddrMode = BackupAddrMode;
4620 AddrModeInsts.resize(OldSize);
4621 TPT.rollback(LastKnownGood);
4622 }
4623 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
4624 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
4625 return true;
4626 TPT.rollback(LastKnownGood);
4627 } else if (isa<ConstantPointerNull>(Addr)) {
4628 // Null pointer gets folded without affecting the addressing mode.
4629 return true;
4630 }
4631
4632 // Worse case, the target should support [reg] addressing modes. :)
4633 if (!AddrMode.HasBaseReg) {
4634 AddrMode.HasBaseReg = true;
4635 AddrMode.BaseReg = Addr;
4636 // Still check for legality in case the target supports [imm] but not [i+r].
4637 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4638 return true;
4639 AddrMode.HasBaseReg = false;
4640 AddrMode.BaseReg = nullptr;
4641 }
4642
4643 // If the base register is already taken, see if we can do [r+r].
4644 if (AddrMode.Scale == 0) {
4645 AddrMode.Scale = 1;
4646 AddrMode.ScaledReg = Addr;
4647 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4648 return true;
4649 AddrMode.Scale = 0;
4650 AddrMode.ScaledReg = nullptr;
4651 }
4652 // Couldn't match.
4653 TPT.rollback(LastKnownGood);
4654 return false;
4655 }
4656
4657 /// Check to see if all uses of OpVal by the specified inline asm call are due
4658 /// to memory operands. If so, return true, otherwise return false.
IsOperandAMemoryOperand(CallInst * CI,InlineAsm * IA,Value * OpVal,const TargetLowering & TLI,const TargetRegisterInfo & TRI)4659 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
4660 const TargetLowering &TLI,
4661 const TargetRegisterInfo &TRI) {
4662 const Function *F = CI->getFunction();
4663 TargetLowering::AsmOperandInfoVector TargetConstraints =
4664 TLI.ParseConstraints(F->getParent()->getDataLayout(), &TRI, *CI);
4665
4666 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4667 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4668
4669 // Compute the constraint code and ConstraintType to use.
4670 TLI.ComputeConstraintToUse(OpInfo, SDValue());
4671
4672 // If this asm operand is our Value*, and if it isn't an indirect memory
4673 // operand, we can't fold it!
4674 if (OpInfo.CallOperandVal == OpVal &&
4675 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
4676 !OpInfo.isIndirect))
4677 return false;
4678 }
4679
4680 return true;
4681 }
4682
4683 // Max number of memory uses to look at before aborting the search to conserve
4684 // compile time.
4685 static constexpr int MaxMemoryUsesToScan = 20;
4686
4687 /// Recursively walk all the uses of I until we find a memory use.
4688 /// If we find an obviously non-foldable instruction, return true.
4689 /// 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,bool OptSize,ProfileSummaryInfo * PSI,BlockFrequencyInfo * BFI,int SeenInsts=0)4690 static bool FindAllMemoryUses(
4691 Instruction *I,
4692 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
4693 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
4694 const TargetRegisterInfo &TRI, bool OptSize, ProfileSummaryInfo *PSI,
4695 BlockFrequencyInfo *BFI, int SeenInsts = 0) {
4696 // If we already considered this instruction, we're done.
4697 if (!ConsideredInsts.insert(I).second)
4698 return false;
4699
4700 // If this is an obviously unfoldable instruction, bail out.
4701 if (!MightBeFoldableInst(I))
4702 return true;
4703
4704 // Loop over all the uses, recursively processing them.
4705 for (Use &U : I->uses()) {
4706 // Conservatively return true if we're seeing a large number or a deep chain
4707 // of users. This avoids excessive compilation times in pathological cases.
4708 if (SeenInsts++ >= MaxMemoryUsesToScan)
4709 return true;
4710
4711 Instruction *UserI = cast<Instruction>(U.getUser());
4712 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4713 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
4714 continue;
4715 }
4716
4717 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4718 unsigned opNo = U.getOperandNo();
4719 if (opNo != StoreInst::getPointerOperandIndex())
4720 return true; // Storing addr, not into addr.
4721 MemoryUses.push_back(std::make_pair(SI, opNo));
4722 continue;
4723 }
4724
4725 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
4726 unsigned opNo = U.getOperandNo();
4727 if (opNo != AtomicRMWInst::getPointerOperandIndex())
4728 return true; // Storing addr, not into addr.
4729 MemoryUses.push_back(std::make_pair(RMW, opNo));
4730 continue;
4731 }
4732
4733 if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
4734 unsigned opNo = U.getOperandNo();
4735 if (opNo != AtomicCmpXchgInst::getPointerOperandIndex())
4736 return true; // Storing addr, not into addr.
4737 MemoryUses.push_back(std::make_pair(CmpX, opNo));
4738 continue;
4739 }
4740
4741 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4742 if (CI->hasFnAttr(Attribute::Cold)) {
4743 // If this is a cold call, we can sink the addressing calculation into
4744 // the cold path. See optimizeCallInst
4745 bool OptForSize = OptSize ||
4746 llvm::shouldOptimizeForSize(CI->getParent(), PSI, BFI);
4747 if (!OptForSize)
4748 continue;
4749 }
4750
4751 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledOperand());
4752 if (!IA) return true;
4753
4754 // If this is a memory operand, we're cool, otherwise bail out.
4755 if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI))
4756 return true;
4757 continue;
4758 }
4759
4760 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI, OptSize,
4761 PSI, BFI, SeenInsts))
4762 return true;
4763 }
4764
4765 return false;
4766 }
4767
4768 /// Return true if Val is already known to be live at the use site that we're
4769 /// folding it into. If so, there is no cost to include it in the addressing
4770 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4771 /// instruction already.
valueAlreadyLiveAtInst(Value * Val,Value * KnownLive1,Value * KnownLive2)4772 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4773 Value *KnownLive2) {
4774 // If Val is either of the known-live values, we know it is live!
4775 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4776 return true;
4777
4778 // All values other than instructions and arguments (e.g. constants) are live.
4779 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4780
4781 // If Val is a constant sized alloca in the entry block, it is live, this is
4782 // true because it is just a reference to the stack/frame pointer, which is
4783 // live for the whole function.
4784 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4785 if (AI->isStaticAlloca())
4786 return true;
4787
4788 // Check to see if this value is already used in the memory instruction's
4789 // block. If so, it's already live into the block at the very least, so we
4790 // can reasonably fold it.
4791 return Val->isUsedInBasicBlock(MemoryInst->getParent());
4792 }
4793
4794 /// It is possible for the addressing mode of the machine to fold the specified
4795 /// instruction into a load or store that ultimately uses it.
4796 /// However, the specified instruction has multiple uses.
4797 /// Given this, it may actually increase register pressure to fold it
4798 /// into the load. For example, consider this code:
4799 ///
4800 /// X = ...
4801 /// Y = X+1
4802 /// use(Y) -> nonload/store
4803 /// Z = Y+1
4804 /// load Z
4805 ///
4806 /// In this case, Y has multiple uses, and can be folded into the load of Z
4807 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4808 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4809 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4810 /// number of computations either.
4811 ///
4812 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4813 /// X was live across 'load Z' for other reasons, we actually *would* want to
4814 /// fold the addressing mode in the Z case. This would make Y die earlier.
4815 bool AddressingModeMatcher::
isProfitableToFoldIntoAddressingMode(Instruction * I,ExtAddrMode & AMBefore,ExtAddrMode & AMAfter)4816 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4817 ExtAddrMode &AMAfter) {
4818 if (IgnoreProfitability) return true;
4819
4820 // AMBefore is the addressing mode before this instruction was folded into it,
4821 // and AMAfter is the addressing mode after the instruction was folded. Get
4822 // the set of registers referenced by AMAfter and subtract out those
4823 // referenced by AMBefore: this is the set of values which folding in this
4824 // address extends the lifetime of.
4825 //
4826 // Note that there are only two potential values being referenced here,
4827 // BaseReg and ScaleReg (global addresses are always available, as are any
4828 // folded immediates).
4829 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4830
4831 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4832 // lifetime wasn't extended by adding this instruction.
4833 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4834 BaseReg = nullptr;
4835 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4836 ScaledReg = nullptr;
4837
4838 // If folding this instruction (and it's subexprs) didn't extend any live
4839 // ranges, we're ok with it.
4840 if (!BaseReg && !ScaledReg)
4841 return true;
4842
4843 // If all uses of this instruction can have the address mode sunk into them,
4844 // we can remove the addressing mode and effectively trade one live register
4845 // for another (at worst.) In this context, folding an addressing mode into
4846 // the use is just a particularly nice way of sinking it.
4847 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
4848 SmallPtrSet<Instruction*, 16> ConsideredInsts;
4849 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI, OptSize,
4850 PSI, BFI))
4851 return false; // Has a non-memory, non-foldable use!
4852
4853 // Now that we know that all uses of this instruction are part of a chain of
4854 // computation involving only operations that could theoretically be folded
4855 // into a memory use, loop over each of these memory operation uses and see
4856 // if they could *actually* fold the instruction. The assumption is that
4857 // addressing modes are cheap and that duplicating the computation involved
4858 // many times is worthwhile, even on a fastpath. For sinking candidates
4859 // (i.e. cold call sites), this serves as a way to prevent excessive code
4860 // growth since most architectures have some reasonable small and fast way to
4861 // compute an effective address. (i.e LEA on x86)
4862 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4863 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4864 Instruction *User = MemoryUses[i].first;
4865 unsigned OpNo = MemoryUses[i].second;
4866
4867 // Get the access type of this use. If the use isn't a pointer, we don't
4868 // know what it accesses.
4869 Value *Address = User->getOperand(OpNo);
4870 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4871 if (!AddrTy)
4872 return false;
4873 Type *AddressAccessTy = AddrTy->getElementType();
4874 unsigned AS = AddrTy->getAddressSpace();
4875
4876 // Do a match against the root of this address, ignoring profitability. This
4877 // will tell us if the addressing mode for the memory operation will
4878 // *actually* cover the shared instruction.
4879 ExtAddrMode Result;
4880 std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
4881 0);
4882 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4883 TPT.getRestorationPoint();
4884 AddressingModeMatcher Matcher(
4885 MatchedAddrModeInsts, TLI, TRI, AddressAccessTy, AS, MemoryInst, Result,
4886 InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP, OptSize, PSI, BFI);
4887 Matcher.IgnoreProfitability = true;
4888 bool Success = Matcher.matchAddr(Address, 0);
4889 (void)Success; assert(Success && "Couldn't select *anything*?");
4890
4891 // The match was to check the profitability, the changes made are not
4892 // part of the original matcher. Therefore, they should be dropped
4893 // otherwise the original matcher will not present the right state.
4894 TPT.rollback(LastKnownGood);
4895
4896 // If the match didn't cover I, then it won't be shared by it.
4897 if (!is_contained(MatchedAddrModeInsts, I))
4898 return false;
4899
4900 MatchedAddrModeInsts.clear();
4901 }
4902
4903 return true;
4904 }
4905
4906 /// Return true if the specified values are defined in a
4907 /// different basic block than BB.
IsNonLocalValue(Value * V,BasicBlock * BB)4908 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4909 if (Instruction *I = dyn_cast<Instruction>(V))
4910 return I->getParent() != BB;
4911 return false;
4912 }
4913
4914 /// Sink addressing mode computation immediate before MemoryInst if doing so
4915 /// can be done without increasing register pressure. The need for the
4916 /// register pressure constraint means this can end up being an all or nothing
4917 /// decision for all uses of the same addressing computation.
4918 ///
4919 /// Load and Store Instructions often have addressing modes that can do
4920 /// significant amounts of computation. As such, instruction selection will try
4921 /// to get the load or store to do as much computation as possible for the
4922 /// program. The problem is that isel can only see within a single block. As
4923 /// such, we sink as much legal addressing mode work into the block as possible.
4924 ///
4925 /// This method is used to optimize both load/store and inline asms with memory
4926 /// operands. It's also used to sink addressing computations feeding into cold
4927 /// call sites into their (cold) basic block.
4928 ///
4929 /// The motivation for handling sinking into cold blocks is that doing so can
4930 /// both enable other address mode sinking (by satisfying the register pressure
4931 /// constraint above), and reduce register pressure globally (by removing the
4932 /// addressing mode computation from the fast path entirely.).
optimizeMemoryInst(Instruction * MemoryInst,Value * Addr,Type * AccessTy,unsigned AddrSpace)4933 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4934 Type *AccessTy, unsigned AddrSpace) {
4935 Value *Repl = Addr;
4936
4937 // Try to collapse single-value PHI nodes. This is necessary to undo
4938 // unprofitable PRE transformations.
4939 SmallVector<Value*, 8> worklist;
4940 SmallPtrSet<Value*, 16> Visited;
4941 worklist.push_back(Addr);
4942
4943 // Use a worklist to iteratively look through PHI and select nodes, and
4944 // ensure that the addressing mode obtained from the non-PHI/select roots of
4945 // the graph are compatible.
4946 bool PhiOrSelectSeen = false;
4947 SmallVector<Instruction*, 16> AddrModeInsts;
4948 const SimplifyQuery SQ(*DL, TLInfo);
4949 AddressingModeCombiner AddrModes(SQ, Addr);
4950 TypePromotionTransaction TPT(RemovedInsts);
4951 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4952 TPT.getRestorationPoint();
4953 while (!worklist.empty()) {
4954 Value *V = worklist.back();
4955 worklist.pop_back();
4956
4957 // We allow traversing cyclic Phi nodes.
4958 // In case of success after this loop we ensure that traversing through
4959 // Phi nodes ends up with all cases to compute address of the form
4960 // BaseGV + Base + Scale * Index + Offset
4961 // where Scale and Offset are constans and BaseGV, Base and Index
4962 // are exactly the same Values in all cases.
4963 // It means that BaseGV, Scale and Offset dominate our memory instruction
4964 // and have the same value as they had in address computation represented
4965 // as Phi. So we can safely sink address computation to memory instruction.
4966 if (!Visited.insert(V).second)
4967 continue;
4968
4969 // For a PHI node, push all of its incoming values.
4970 if (PHINode *P = dyn_cast<PHINode>(V)) {
4971 append_range(worklist, P->incoming_values());
4972 PhiOrSelectSeen = true;
4973 continue;
4974 }
4975 // Similar for select.
4976 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
4977 worklist.push_back(SI->getFalseValue());
4978 worklist.push_back(SI->getTrueValue());
4979 PhiOrSelectSeen = true;
4980 continue;
4981 }
4982
4983 // For non-PHIs, determine the addressing mode being computed. Note that
4984 // the result may differ depending on what other uses our candidate
4985 // addressing instructions might have.
4986 AddrModeInsts.clear();
4987 std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
4988 0);
4989 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4990 V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *TRI,
4991 InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP, OptSize, PSI,
4992 BFI.get());
4993
4994 GetElementPtrInst *GEP = LargeOffsetGEP.first;
4995 if (GEP && !NewGEPBases.count(GEP)) {
4996 // If splitting the underlying data structure can reduce the offset of a
4997 // GEP, collect the GEP. Skip the GEPs that are the new bases of
4998 // previously split data structures.
4999 LargeOffsetGEPMap[GEP->getPointerOperand()].push_back(LargeOffsetGEP);
5000 if (LargeOffsetGEPID.find(GEP) == LargeOffsetGEPID.end())
5001 LargeOffsetGEPID[GEP] = LargeOffsetGEPID.size();
5002 }
5003
5004 NewAddrMode.OriginalValue = V;
5005 if (!AddrModes.addNewAddrMode(NewAddrMode))
5006 break;
5007 }
5008
5009 // Try to combine the AddrModes we've collected. If we couldn't collect any,
5010 // or we have multiple but either couldn't combine them or combining them
5011 // wouldn't do anything useful, bail out now.
5012 if (!AddrModes.combineAddrModes()) {
5013 TPT.rollback(LastKnownGood);
5014 return false;
5015 }
5016 bool Modified = TPT.commit();
5017
5018 // Get the combined AddrMode (or the only AddrMode, if we only had one).
5019 ExtAddrMode AddrMode = AddrModes.getAddrMode();
5020
5021 // If all the instructions matched are already in this BB, don't do anything.
5022 // If we saw a Phi node then it is not local definitely, and if we saw a select
5023 // then we want to push the address calculation past it even if it's already
5024 // in this BB.
5025 if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) {
5026 return IsNonLocalValue(V, MemoryInst->getParent());
5027 })) {
5028 LLVM_DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode
5029 << "\n");
5030 return Modified;
5031 }
5032
5033 // Insert this computation right after this user. Since our caller is
5034 // scanning from the top of the BB to the bottom, reuse of the expr are
5035 // guaranteed to happen later.
5036 IRBuilder<> Builder(MemoryInst);
5037
5038 // Now that we determined the addressing expression we want to use and know
5039 // that we have to sink it into this block. Check to see if we have already
5040 // done this for some other load/store instr in this block. If so, reuse
5041 // the computation. Before attempting reuse, check if the address is valid
5042 // as it may have been erased.
5043
5044 WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
5045
5046 Value * SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
5047 if (SunkAddr) {
5048 LLVM_DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode
5049 << " for " << *MemoryInst << "\n");
5050 if (SunkAddr->getType() != Addr->getType())
5051 SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
5052 } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() &&
5053 SubtargetInfo->addrSinkUsingGEPs())) {
5054 // By default, we use the GEP-based method when AA is used later. This
5055 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
5056 LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
5057 << " for " << *MemoryInst << "\n");
5058 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
5059 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
5060
5061 // First, find the pointer.
5062 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
5063 ResultPtr = AddrMode.BaseReg;
5064 AddrMode.BaseReg = nullptr;
5065 }
5066
5067 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
5068 // We can't add more than one pointer together, nor can we scale a
5069 // pointer (both of which seem meaningless).
5070 if (ResultPtr || AddrMode.Scale != 1)
5071 return Modified;
5072
5073 ResultPtr = AddrMode.ScaledReg;
5074 AddrMode.Scale = 0;
5075 }
5076
5077 // It is only safe to sign extend the BaseReg if we know that the math
5078 // required to create it did not overflow before we extend it. Since
5079 // the original IR value was tossed in favor of a constant back when
5080 // the AddrMode was created we need to bail out gracefully if widths
5081 // do not match instead of extending it.
5082 //
5083 // (See below for code to add the scale.)
5084 if (AddrMode.Scale) {
5085 Type *ScaledRegTy = AddrMode.ScaledReg->getType();
5086 if (cast<IntegerType>(IntPtrTy)->getBitWidth() >
5087 cast<IntegerType>(ScaledRegTy)->getBitWidth())
5088 return Modified;
5089 }
5090
5091 if (AddrMode.BaseGV) {
5092 if (ResultPtr)
5093 return Modified;
5094
5095 ResultPtr = AddrMode.BaseGV;
5096 }
5097
5098 // If the real base value actually came from an inttoptr, then the matcher
5099 // will look through it and provide only the integer value. In that case,
5100 // use it here.
5101 if (!DL->isNonIntegralPointerType(Addr->getType())) {
5102 if (!ResultPtr && AddrMode.BaseReg) {
5103 ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(),
5104 "sunkaddr");
5105 AddrMode.BaseReg = nullptr;
5106 } else if (!ResultPtr && AddrMode.Scale == 1) {
5107 ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(),
5108 "sunkaddr");
5109 AddrMode.Scale = 0;
5110 }
5111 }
5112
5113 if (!ResultPtr &&
5114 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
5115 SunkAddr = Constant::getNullValue(Addr->getType());
5116 } else if (!ResultPtr) {
5117 return Modified;
5118 } else {
5119 Type *I8PtrTy =
5120 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
5121 Type *I8Ty = Builder.getInt8Ty();
5122
5123 // Start with the base register. Do this first so that subsequent address
5124 // matching finds it last, which will prevent it from trying to match it
5125 // as the scaled value in case it happens to be a mul. That would be
5126 // problematic if we've sunk a different mul for the scale, because then
5127 // we'd end up sinking both muls.
5128 if (AddrMode.BaseReg) {
5129 Value *V = AddrMode.BaseReg;
5130 if (V->getType() != IntPtrTy)
5131 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
5132
5133 ResultIndex = V;
5134 }
5135
5136 // Add the scale value.
5137 if (AddrMode.Scale) {
5138 Value *V = AddrMode.ScaledReg;
5139 if (V->getType() == IntPtrTy) {
5140 // done.
5141 } else {
5142 assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <
5143 cast<IntegerType>(V->getType())->getBitWidth() &&
5144 "We can't transform if ScaledReg is too narrow");
5145 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
5146 }
5147
5148 if (AddrMode.Scale != 1)
5149 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
5150 "sunkaddr");
5151 if (ResultIndex)
5152 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
5153 else
5154 ResultIndex = V;
5155 }
5156
5157 // Add in the Base Offset if present.
5158 if (AddrMode.BaseOffs) {
5159 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
5160 if (ResultIndex) {
5161 // We need to add this separately from the scale above to help with
5162 // SDAG consecutive load/store merging.
5163 if (ResultPtr->getType() != I8PtrTy)
5164 ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
5165 ResultPtr =
5166 AddrMode.InBounds
5167 ? Builder.CreateInBoundsGEP(I8Ty, ResultPtr, ResultIndex,
5168 "sunkaddr")
5169 : Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
5170 }
5171
5172 ResultIndex = V;
5173 }
5174
5175 if (!ResultIndex) {
5176 SunkAddr = ResultPtr;
5177 } else {
5178 if (ResultPtr->getType() != I8PtrTy)
5179 ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
5180 SunkAddr =
5181 AddrMode.InBounds
5182 ? Builder.CreateInBoundsGEP(I8Ty, ResultPtr, ResultIndex,
5183 "sunkaddr")
5184 : Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
5185 }
5186
5187 if (SunkAddr->getType() != Addr->getType())
5188 SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
5189 }
5190 } else {
5191 // We'd require a ptrtoint/inttoptr down the line, which we can't do for
5192 // non-integral pointers, so in that case bail out now.
5193 Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
5194 Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
5195 PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy);
5196 PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy);
5197 if (DL->isNonIntegralPointerType(Addr->getType()) ||
5198 (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) ||
5199 (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) ||
5200 (AddrMode.BaseGV &&
5201 DL->isNonIntegralPointerType(AddrMode.BaseGV->getType())))
5202 return Modified;
5203
5204 LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
5205 << " for " << *MemoryInst << "\n");
5206 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
5207 Value *Result = nullptr;
5208
5209 // Start with the base register. Do this first so that subsequent address
5210 // matching finds it last, which will prevent it from trying to match it
5211 // as the scaled value in case it happens to be a mul. That would be
5212 // problematic if we've sunk a different mul for the scale, because then
5213 // we'd end up sinking both muls.
5214 if (AddrMode.BaseReg) {
5215 Value *V = AddrMode.BaseReg;
5216 if (V->getType()->isPointerTy())
5217 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
5218 if (V->getType() != IntPtrTy)
5219 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
5220 Result = V;
5221 }
5222
5223 // Add the scale value.
5224 if (AddrMode.Scale) {
5225 Value *V = AddrMode.ScaledReg;
5226 if (V->getType() == IntPtrTy) {
5227 // done.
5228 } else if (V->getType()->isPointerTy()) {
5229 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
5230 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
5231 cast<IntegerType>(V->getType())->getBitWidth()) {
5232 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
5233 } else {
5234 // It is only safe to sign extend the BaseReg if we know that the math
5235 // required to create it did not overflow before we extend it. Since
5236 // the original IR value was tossed in favor of a constant back when
5237 // the AddrMode was created we need to bail out gracefully if widths
5238 // do not match instead of extending it.
5239 Instruction *I = dyn_cast_or_null<Instruction>(Result);
5240 if (I && (Result != AddrMode.BaseReg))
5241 I->eraseFromParent();
5242 return Modified;
5243 }
5244 if (AddrMode.Scale != 1)
5245 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
5246 "sunkaddr");
5247 if (Result)
5248 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5249 else
5250 Result = V;
5251 }
5252
5253 // Add in the BaseGV if present.
5254 if (AddrMode.BaseGV) {
5255 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
5256 if (Result)
5257 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5258 else
5259 Result = V;
5260 }
5261
5262 // Add in the Base Offset if present.
5263 if (AddrMode.BaseOffs) {
5264 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
5265 if (Result)
5266 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5267 else
5268 Result = V;
5269 }
5270
5271 if (!Result)
5272 SunkAddr = Constant::getNullValue(Addr->getType());
5273 else
5274 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
5275 }
5276
5277 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
5278 // Store the newly computed address into the cache. In the case we reused a
5279 // value, this should be idempotent.
5280 SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
5281
5282 // If we have no uses, recursively delete the value and all dead instructions
5283 // using it.
5284 if (Repl->use_empty()) {
5285 resetIteratorIfInvalidatedWhileCalling(CurInstIterator->getParent(), [&]() {
5286 RecursivelyDeleteTriviallyDeadInstructions(
5287 Repl, TLInfo, nullptr,
5288 [&](Value *V) { removeAllAssertingVHReferences(V); });
5289 });
5290 }
5291 ++NumMemoryInsts;
5292 return true;
5293 }
5294
5295 /// Rewrite GEP input to gather/scatter to enable SelectionDAGBuilder to find
5296 /// a uniform base to use for ISD::MGATHER/MSCATTER. SelectionDAGBuilder can
5297 /// only handle a 2 operand GEP in the same basic block or a splat constant
5298 /// vector. The 2 operands to the GEP must have a scalar pointer and a vector
5299 /// index.
5300 ///
5301 /// If the existing GEP has a vector base pointer that is splat, we can look
5302 /// through the splat to find the scalar pointer. If we can't find a scalar
5303 /// pointer there's nothing we can do.
5304 ///
5305 /// If we have a GEP with more than 2 indices where the middle indices are all
5306 /// zeroes, we can replace it with 2 GEPs where the second has 2 operands.
5307 ///
5308 /// If the final index isn't a vector or is a splat, we can emit a scalar GEP
5309 /// followed by a GEP with an all zeroes vector index. This will enable
5310 /// SelectionDAGBuilder to use the scalar GEP as the uniform base and have a
5311 /// zero index.
optimizeGatherScatterInst(Instruction * MemoryInst,Value * Ptr)5312 bool CodeGenPrepare::optimizeGatherScatterInst(Instruction *MemoryInst,
5313 Value *Ptr) {
5314 Value *NewAddr;
5315
5316 if (const auto *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
5317 // Don't optimize GEPs that don't have indices.
5318 if (!GEP->hasIndices())
5319 return false;
5320
5321 // If the GEP and the gather/scatter aren't in the same BB, don't optimize.
5322 // FIXME: We should support this by sinking the GEP.
5323 if (MemoryInst->getParent() != GEP->getParent())
5324 return false;
5325
5326 SmallVector<Value *, 2> Ops(GEP->operands());
5327
5328 bool RewriteGEP = false;
5329
5330 if (Ops[0]->getType()->isVectorTy()) {
5331 Ops[0] = getSplatValue(Ops[0]);
5332 if (!Ops[0])
5333 return false;
5334 RewriteGEP = true;
5335 }
5336
5337 unsigned FinalIndex = Ops.size() - 1;
5338
5339 // Ensure all but the last index is 0.
5340 // FIXME: This isn't strictly required. All that's required is that they are
5341 // all scalars or splats.
5342 for (unsigned i = 1; i < FinalIndex; ++i) {
5343 auto *C = dyn_cast<Constant>(Ops[i]);
5344 if (!C)
5345 return false;
5346 if (isa<VectorType>(C->getType()))
5347 C = C->getSplatValue();
5348 auto *CI = dyn_cast_or_null<ConstantInt>(C);
5349 if (!CI || !CI->isZero())
5350 return false;
5351 // Scalarize the index if needed.
5352 Ops[i] = CI;
5353 }
5354
5355 // Try to scalarize the final index.
5356 if (Ops[FinalIndex]->getType()->isVectorTy()) {
5357 if (Value *V = getSplatValue(Ops[FinalIndex])) {
5358 auto *C = dyn_cast<ConstantInt>(V);
5359 // Don't scalarize all zeros vector.
5360 if (!C || !C->isZero()) {
5361 Ops[FinalIndex] = V;
5362 RewriteGEP = true;
5363 }
5364 }
5365 }
5366
5367 // If we made any changes or the we have extra operands, we need to generate
5368 // new instructions.
5369 if (!RewriteGEP && Ops.size() == 2)
5370 return false;
5371
5372 auto NumElts = cast<VectorType>(Ptr->getType())->getElementCount();
5373
5374 IRBuilder<> Builder(MemoryInst);
5375
5376 Type *ScalarIndexTy = DL->getIndexType(Ops[0]->getType()->getScalarType());
5377
5378 // If the final index isn't a vector, emit a scalar GEP containing all ops
5379 // and a vector GEP with all zeroes final index.
5380 if (!Ops[FinalIndex]->getType()->isVectorTy()) {
5381 NewAddr = Builder.CreateGEP(Ops[0], makeArrayRef(Ops).drop_front());
5382 auto *IndexTy = VectorType::get(ScalarIndexTy, NumElts);
5383 NewAddr = Builder.CreateGEP(NewAddr, Constant::getNullValue(IndexTy));
5384 } else {
5385 Value *Base = Ops[0];
5386 Value *Index = Ops[FinalIndex];
5387
5388 // Create a scalar GEP if there are more than 2 operands.
5389 if (Ops.size() != 2) {
5390 // Replace the last index with 0.
5391 Ops[FinalIndex] = Constant::getNullValue(ScalarIndexTy);
5392 Base = Builder.CreateGEP(Base, makeArrayRef(Ops).drop_front());
5393 }
5394
5395 // Now create the GEP with scalar pointer and vector index.
5396 NewAddr = Builder.CreateGEP(Base, Index);
5397 }
5398 } else if (!isa<Constant>(Ptr)) {
5399 // Not a GEP, maybe its a splat and we can create a GEP to enable
5400 // SelectionDAGBuilder to use it as a uniform base.
5401 Value *V = getSplatValue(Ptr);
5402 if (!V)
5403 return false;
5404
5405 auto NumElts = cast<VectorType>(Ptr->getType())->getElementCount();
5406
5407 IRBuilder<> Builder(MemoryInst);
5408
5409 // Emit a vector GEP with a scalar pointer and all 0s vector index.
5410 Type *ScalarIndexTy = DL->getIndexType(V->getType()->getScalarType());
5411 auto *IndexTy = VectorType::get(ScalarIndexTy, NumElts);
5412 NewAddr = Builder.CreateGEP(V, Constant::getNullValue(IndexTy));
5413 } else {
5414 // Constant, SelectionDAGBuilder knows to check if its a splat.
5415 return false;
5416 }
5417
5418 MemoryInst->replaceUsesOfWith(Ptr, NewAddr);
5419
5420 // If we have no uses, recursively delete the value and all dead instructions
5421 // using it.
5422 if (Ptr->use_empty())
5423 RecursivelyDeleteTriviallyDeadInstructions(
5424 Ptr, TLInfo, nullptr,
5425 [&](Value *V) { removeAllAssertingVHReferences(V); });
5426
5427 return true;
5428 }
5429
5430 /// If there are any memory operands, use OptimizeMemoryInst to sink their
5431 /// address computing into the block when possible / profitable.
optimizeInlineAsmInst(CallInst * CS)5432 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
5433 bool MadeChange = false;
5434
5435 const TargetRegisterInfo *TRI =
5436 TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
5437 TargetLowering::AsmOperandInfoVector TargetConstraints =
5438 TLI->ParseConstraints(*DL, TRI, *CS);
5439 unsigned ArgNo = 0;
5440 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5441 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
5442
5443 // Compute the constraint code and ConstraintType to use.
5444 TLI->ComputeConstraintToUse(OpInfo, SDValue());
5445
5446 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5447 OpInfo.isIndirect) {
5448 Value *OpVal = CS->getArgOperand(ArgNo++);
5449 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
5450 } else if (OpInfo.Type == InlineAsm::isInput)
5451 ArgNo++;
5452 }
5453
5454 return MadeChange;
5455 }
5456
5457 /// Check if all the uses of \p Val are equivalent (or free) zero or
5458 /// sign extensions.
hasSameExtUse(Value * Val,const TargetLowering & TLI)5459 static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
5460 assert(!Val->use_empty() && "Input must have at least one use");
5461 const Instruction *FirstUser = cast<Instruction>(*Val->user_begin());
5462 bool IsSExt = isa<SExtInst>(FirstUser);
5463 Type *ExtTy = FirstUser->getType();
5464 for (const User *U : Val->users()) {
5465 const Instruction *UI = cast<Instruction>(U);
5466 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
5467 return false;
5468 Type *CurTy = UI->getType();
5469 // Same input and output types: Same instruction after CSE.
5470 if (CurTy == ExtTy)
5471 continue;
5472
5473 // If IsSExt is true, we are in this situation:
5474 // a = Val
5475 // b = sext ty1 a to ty2
5476 // c = sext ty1 a to ty3
5477 // Assuming ty2 is shorter than ty3, this could be turned into:
5478 // a = Val
5479 // b = sext ty1 a to ty2
5480 // c = sext ty2 b to ty3
5481 // However, the last sext is not free.
5482 if (IsSExt)
5483 return false;
5484
5485 // This is a ZExt, maybe this is free to extend from one type to another.
5486 // In that case, we would not account for a different use.
5487 Type *NarrowTy;
5488 Type *LargeTy;
5489 if (ExtTy->getScalarType()->getIntegerBitWidth() >
5490 CurTy->getScalarType()->getIntegerBitWidth()) {
5491 NarrowTy = CurTy;
5492 LargeTy = ExtTy;
5493 } else {
5494 NarrowTy = ExtTy;
5495 LargeTy = CurTy;
5496 }
5497
5498 if (!TLI.isZExtFree(NarrowTy, LargeTy))
5499 return false;
5500 }
5501 // All uses are the same or can be derived from one another for free.
5502 return true;
5503 }
5504
5505 /// Try to speculatively promote extensions in \p Exts and continue
5506 /// promoting through newly promoted operands recursively as far as doing so is
5507 /// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
5508 /// When some promotion happened, \p TPT contains the proper state to revert
5509 /// them.
5510 ///
5511 /// \return true if some promotion happened, false otherwise.
tryToPromoteExts(TypePromotionTransaction & TPT,const SmallVectorImpl<Instruction * > & Exts,SmallVectorImpl<Instruction * > & ProfitablyMovedExts,unsigned CreatedInstsCost)5512 bool CodeGenPrepare::tryToPromoteExts(
5513 TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
5514 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
5515 unsigned CreatedInstsCost) {
5516 bool Promoted = false;
5517
5518 // Iterate over all the extensions to try to promote them.
5519 for (auto *I : Exts) {
5520 // Early check if we directly have ext(load).
5521 if (isa<LoadInst>(I->getOperand(0))) {
5522 ProfitablyMovedExts.push_back(I);
5523 continue;
5524 }
5525
5526 // Check whether or not we want to do any promotion. The reason we have
5527 // this check inside the for loop is to catch the case where an extension
5528 // is directly fed by a load because in such case the extension can be moved
5529 // up without any promotion on its operands.
5530 if (!TLI->enableExtLdPromotion() || DisableExtLdPromotion)
5531 return false;
5532
5533 // Get the action to perform the promotion.
5534 TypePromotionHelper::Action TPH =
5535 TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
5536 // Check if we can promote.
5537 if (!TPH) {
5538 // Save the current extension as we cannot move up through its operand.
5539 ProfitablyMovedExts.push_back(I);
5540 continue;
5541 }
5542
5543 // Save the current state.
5544 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5545 TPT.getRestorationPoint();
5546 SmallVector<Instruction *, 4> NewExts;
5547 unsigned NewCreatedInstsCost = 0;
5548 unsigned ExtCost = !TLI->isExtFree(I);
5549 // Promote.
5550 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
5551 &NewExts, nullptr, *TLI);
5552 assert(PromotedVal &&
5553 "TypePromotionHelper should have filtered out those cases");
5554
5555 // We would be able to merge only one extension in a load.
5556 // Therefore, if we have more than 1 new extension we heuristically
5557 // cut this search path, because it means we degrade the code quality.
5558 // With exactly 2, the transformation is neutral, because we will merge
5559 // one extension but leave one. However, we optimistically keep going,
5560 // because the new extension may be removed too.
5561 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
5562 // FIXME: It would be possible to propagate a negative value instead of
5563 // conservatively ceiling it to 0.
5564 TotalCreatedInstsCost =
5565 std::max((long long)0, (TotalCreatedInstsCost - ExtCost));
5566 if (!StressExtLdPromotion &&
5567 (TotalCreatedInstsCost > 1 ||
5568 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
5569 // This promotion is not profitable, rollback to the previous state, and
5570 // save the current extension in ProfitablyMovedExts as the latest
5571 // speculative promotion turned out to be unprofitable.
5572 TPT.rollback(LastKnownGood);
5573 ProfitablyMovedExts.push_back(I);
5574 continue;
5575 }
5576 // Continue promoting NewExts as far as doing so is profitable.
5577 SmallVector<Instruction *, 2> NewlyMovedExts;
5578 (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost);
5579 bool NewPromoted = false;
5580 for (auto *ExtInst : NewlyMovedExts) {
5581 Instruction *MovedExt = cast<Instruction>(ExtInst);
5582 Value *ExtOperand = MovedExt->getOperand(0);
5583 // If we have reached to a load, we need this extra profitability check
5584 // as it could potentially be merged into an ext(load).
5585 if (isa<LoadInst>(ExtOperand) &&
5586 !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
5587 (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI))))
5588 continue;
5589
5590 ProfitablyMovedExts.push_back(MovedExt);
5591 NewPromoted = true;
5592 }
5593
5594 // If none of speculative promotions for NewExts is profitable, rollback
5595 // and save the current extension (I) as the last profitable extension.
5596 if (!NewPromoted) {
5597 TPT.rollback(LastKnownGood);
5598 ProfitablyMovedExts.push_back(I);
5599 continue;
5600 }
5601 // The promotion is profitable.
5602 Promoted = true;
5603 }
5604 return Promoted;
5605 }
5606
5607 /// Merging redundant sexts when one is dominating the other.
mergeSExts(Function & F)5608 bool CodeGenPrepare::mergeSExts(Function &F) {
5609 bool Changed = false;
5610 for (auto &Entry : ValToSExtendedUses) {
5611 SExts &Insts = Entry.second;
5612 SExts CurPts;
5613 for (Instruction *Inst : Insts) {
5614 if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) ||
5615 Inst->getOperand(0) != Entry.first)
5616 continue;
5617 bool inserted = false;
5618 for (auto &Pt : CurPts) {
5619 if (getDT(F).dominates(Inst, Pt)) {
5620 Pt->replaceAllUsesWith(Inst);
5621 RemovedInsts.insert(Pt);
5622 Pt->removeFromParent();
5623 Pt = Inst;
5624 inserted = true;
5625 Changed = true;
5626 break;
5627 }
5628 if (!getDT(F).dominates(Pt, Inst))
5629 // Give up if we need to merge in a common dominator as the
5630 // experiments show it is not profitable.
5631 continue;
5632 Inst->replaceAllUsesWith(Pt);
5633 RemovedInsts.insert(Inst);
5634 Inst->removeFromParent();
5635 inserted = true;
5636 Changed = true;
5637 break;
5638 }
5639 if (!inserted)
5640 CurPts.push_back(Inst);
5641 }
5642 }
5643 return Changed;
5644 }
5645
5646 // Splitting large data structures so that the GEPs accessing them can have
5647 // smaller offsets so that they can be sunk to the same blocks as their users.
5648 // For example, a large struct starting from %base is split into two parts
5649 // where the second part starts from %new_base.
5650 //
5651 // Before:
5652 // BB0:
5653 // %base =
5654 //
5655 // BB1:
5656 // %gep0 = gep %base, off0
5657 // %gep1 = gep %base, off1
5658 // %gep2 = gep %base, off2
5659 //
5660 // BB2:
5661 // %load1 = load %gep0
5662 // %load2 = load %gep1
5663 // %load3 = load %gep2
5664 //
5665 // After:
5666 // BB0:
5667 // %base =
5668 // %new_base = gep %base, off0
5669 //
5670 // BB1:
5671 // %new_gep0 = %new_base
5672 // %new_gep1 = gep %new_base, off1 - off0
5673 // %new_gep2 = gep %new_base, off2 - off0
5674 //
5675 // BB2:
5676 // %load1 = load i32, i32* %new_gep0
5677 // %load2 = load i32, i32* %new_gep1
5678 // %load3 = load i32, i32* %new_gep2
5679 //
5680 // %new_gep1 and %new_gep2 can be sunk to BB2 now after the splitting because
5681 // their offsets are smaller enough to fit into the addressing mode.
splitLargeGEPOffsets()5682 bool CodeGenPrepare::splitLargeGEPOffsets() {
5683 bool Changed = false;
5684 for (auto &Entry : LargeOffsetGEPMap) {
5685 Value *OldBase = Entry.first;
5686 SmallVectorImpl<std::pair<AssertingVH<GetElementPtrInst>, int64_t>>
5687 &LargeOffsetGEPs = Entry.second;
5688 auto compareGEPOffset =
5689 [&](const std::pair<GetElementPtrInst *, int64_t> &LHS,
5690 const std::pair<GetElementPtrInst *, int64_t> &RHS) {
5691 if (LHS.first == RHS.first)
5692 return false;
5693 if (LHS.second != RHS.second)
5694 return LHS.second < RHS.second;
5695 return LargeOffsetGEPID[LHS.first] < LargeOffsetGEPID[RHS.first];
5696 };
5697 // Sorting all the GEPs of the same data structures based on the offsets.
5698 llvm::sort(LargeOffsetGEPs, compareGEPOffset);
5699 LargeOffsetGEPs.erase(
5700 std::unique(LargeOffsetGEPs.begin(), LargeOffsetGEPs.end()),
5701 LargeOffsetGEPs.end());
5702 // Skip if all the GEPs have the same offsets.
5703 if (LargeOffsetGEPs.front().second == LargeOffsetGEPs.back().second)
5704 continue;
5705 GetElementPtrInst *BaseGEP = LargeOffsetGEPs.begin()->first;
5706 int64_t BaseOffset = LargeOffsetGEPs.begin()->second;
5707 Value *NewBaseGEP = nullptr;
5708
5709 auto *LargeOffsetGEP = LargeOffsetGEPs.begin();
5710 while (LargeOffsetGEP != LargeOffsetGEPs.end()) {
5711 GetElementPtrInst *GEP = LargeOffsetGEP->first;
5712 int64_t Offset = LargeOffsetGEP->second;
5713 if (Offset != BaseOffset) {
5714 TargetLowering::AddrMode AddrMode;
5715 AddrMode.BaseOffs = Offset - BaseOffset;
5716 // The result type of the GEP might not be the type of the memory
5717 // access.
5718 if (!TLI->isLegalAddressingMode(*DL, AddrMode,
5719 GEP->getResultElementType(),
5720 GEP->getAddressSpace())) {
5721 // We need to create a new base if the offset to the current base is
5722 // too large to fit into the addressing mode. So, a very large struct
5723 // may be split into several parts.
5724 BaseGEP = GEP;
5725 BaseOffset = Offset;
5726 NewBaseGEP = nullptr;
5727 }
5728 }
5729
5730 // Generate a new GEP to replace the current one.
5731 LLVMContext &Ctx = GEP->getContext();
5732 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
5733 Type *I8PtrTy =
5734 Type::getInt8PtrTy(Ctx, GEP->getType()->getPointerAddressSpace());
5735 Type *I8Ty = Type::getInt8Ty(Ctx);
5736
5737 if (!NewBaseGEP) {
5738 // Create a new base if we don't have one yet. Find the insertion
5739 // pointer for the new base first.
5740 BasicBlock::iterator NewBaseInsertPt;
5741 BasicBlock *NewBaseInsertBB;
5742 if (auto *BaseI = dyn_cast<Instruction>(OldBase)) {
5743 // If the base of the struct is an instruction, the new base will be
5744 // inserted close to it.
5745 NewBaseInsertBB = BaseI->getParent();
5746 if (isa<PHINode>(BaseI))
5747 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5748 else if (InvokeInst *Invoke = dyn_cast<InvokeInst>(BaseI)) {
5749 NewBaseInsertBB =
5750 SplitEdge(NewBaseInsertBB, Invoke->getNormalDest());
5751 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5752 } else
5753 NewBaseInsertPt = std::next(BaseI->getIterator());
5754 } else {
5755 // If the current base is an argument or global value, the new base
5756 // will be inserted to the entry block.
5757 NewBaseInsertBB = &BaseGEP->getFunction()->getEntryBlock();
5758 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5759 }
5760 IRBuilder<> NewBaseBuilder(NewBaseInsertBB, NewBaseInsertPt);
5761 // Create a new base.
5762 Value *BaseIndex = ConstantInt::get(IntPtrTy, BaseOffset);
5763 NewBaseGEP = OldBase;
5764 if (NewBaseGEP->getType() != I8PtrTy)
5765 NewBaseGEP = NewBaseBuilder.CreatePointerCast(NewBaseGEP, I8PtrTy);
5766 NewBaseGEP =
5767 NewBaseBuilder.CreateGEP(I8Ty, NewBaseGEP, BaseIndex, "splitgep");
5768 NewGEPBases.insert(NewBaseGEP);
5769 }
5770
5771 IRBuilder<> Builder(GEP);
5772 Value *NewGEP = NewBaseGEP;
5773 if (Offset == BaseOffset) {
5774 if (GEP->getType() != I8PtrTy)
5775 NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
5776 } else {
5777 // Calculate the new offset for the new GEP.
5778 Value *Index = ConstantInt::get(IntPtrTy, Offset - BaseOffset);
5779 NewGEP = Builder.CreateGEP(I8Ty, NewBaseGEP, Index);
5780
5781 if (GEP->getType() != I8PtrTy)
5782 NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
5783 }
5784 GEP->replaceAllUsesWith(NewGEP);
5785 LargeOffsetGEPID.erase(GEP);
5786 LargeOffsetGEP = LargeOffsetGEPs.erase(LargeOffsetGEP);
5787 GEP->eraseFromParent();
5788 Changed = true;
5789 }
5790 }
5791 return Changed;
5792 }
5793
optimizePhiType(PHINode * I,SmallPtrSetImpl<PHINode * > & Visited,SmallPtrSetImpl<Instruction * > & DeletedInstrs)5794 bool CodeGenPrepare::optimizePhiType(
5795 PHINode *I, SmallPtrSetImpl<PHINode *> &Visited,
5796 SmallPtrSetImpl<Instruction *> &DeletedInstrs) {
5797 // We are looking for a collection on interconnected phi nodes that together
5798 // only use loads/bitcasts and are used by stores/bitcasts, and the bitcasts
5799 // are of the same type. Convert the whole set of nodes to the type of the
5800 // bitcast.
5801 Type *PhiTy = I->getType();
5802 Type *ConvertTy = nullptr;
5803 if (Visited.count(I) ||
5804 (!I->getType()->isIntegerTy() && !I->getType()->isFloatingPointTy()))
5805 return false;
5806
5807 SmallVector<Instruction *, 4> Worklist;
5808 Worklist.push_back(cast<Instruction>(I));
5809 SmallPtrSet<PHINode *, 4> PhiNodes;
5810 PhiNodes.insert(I);
5811 Visited.insert(I);
5812 SmallPtrSet<Instruction *, 4> Defs;
5813 SmallPtrSet<Instruction *, 4> Uses;
5814 // This works by adding extra bitcasts between load/stores and removing
5815 // existing bicasts. If we have a phi(bitcast(load)) or a store(bitcast(phi))
5816 // we can get in the situation where we remove a bitcast in one iteration
5817 // just to add it again in the next. We need to ensure that at least one
5818 // bitcast we remove are anchored to something that will not change back.
5819 bool AnyAnchored = false;
5820
5821 while (!Worklist.empty()) {
5822 Instruction *II = Worklist.pop_back_val();
5823
5824 if (auto *Phi = dyn_cast<PHINode>(II)) {
5825 // Handle Defs, which might also be PHI's
5826 for (Value *V : Phi->incoming_values()) {
5827 if (auto *OpPhi = dyn_cast<PHINode>(V)) {
5828 if (!PhiNodes.count(OpPhi)) {
5829 if (Visited.count(OpPhi))
5830 return false;
5831 PhiNodes.insert(OpPhi);
5832 Visited.insert(OpPhi);
5833 Worklist.push_back(OpPhi);
5834 }
5835 } else if (auto *OpLoad = dyn_cast<LoadInst>(V)) {
5836 if (!OpLoad->isSimple())
5837 return false;
5838 if (!Defs.count(OpLoad)) {
5839 Defs.insert(OpLoad);
5840 Worklist.push_back(OpLoad);
5841 }
5842 } else if (auto *OpEx = dyn_cast<ExtractElementInst>(V)) {
5843 if (!Defs.count(OpEx)) {
5844 Defs.insert(OpEx);
5845 Worklist.push_back(OpEx);
5846 }
5847 } else if (auto *OpBC = dyn_cast<BitCastInst>(V)) {
5848 if (!ConvertTy)
5849 ConvertTy = OpBC->getOperand(0)->getType();
5850 if (OpBC->getOperand(0)->getType() != ConvertTy)
5851 return false;
5852 if (!Defs.count(OpBC)) {
5853 Defs.insert(OpBC);
5854 Worklist.push_back(OpBC);
5855 AnyAnchored |= !isa<LoadInst>(OpBC->getOperand(0)) &&
5856 !isa<ExtractElementInst>(OpBC->getOperand(0));
5857 }
5858 } else if (!isa<UndefValue>(V)) {
5859 return false;
5860 }
5861 }
5862 }
5863
5864 // Handle uses which might also be phi's
5865 for (User *V : II->users()) {
5866 if (auto *OpPhi = dyn_cast<PHINode>(V)) {
5867 if (!PhiNodes.count(OpPhi)) {
5868 if (Visited.count(OpPhi))
5869 return false;
5870 PhiNodes.insert(OpPhi);
5871 Visited.insert(OpPhi);
5872 Worklist.push_back(OpPhi);
5873 }
5874 } else if (auto *OpStore = dyn_cast<StoreInst>(V)) {
5875 if (!OpStore->isSimple() || OpStore->getOperand(0) != II)
5876 return false;
5877 Uses.insert(OpStore);
5878 } else if (auto *OpBC = dyn_cast<BitCastInst>(V)) {
5879 if (!ConvertTy)
5880 ConvertTy = OpBC->getType();
5881 if (OpBC->getType() != ConvertTy)
5882 return false;
5883 Uses.insert(OpBC);
5884 AnyAnchored |=
5885 any_of(OpBC->users(), [](User *U) { return !isa<StoreInst>(U); });
5886 } else {
5887 return false;
5888 }
5889 }
5890 }
5891
5892 if (!ConvertTy || !AnyAnchored || !TLI->shouldConvertPhiType(PhiTy, ConvertTy))
5893 return false;
5894
5895 LLVM_DEBUG(dbgs() << "Converting " << *I << "\n and connected nodes to "
5896 << *ConvertTy << "\n");
5897
5898 // Create all the new phi nodes of the new type, and bitcast any loads to the
5899 // correct type.
5900 ValueToValueMap ValMap;
5901 ValMap[UndefValue::get(PhiTy)] = UndefValue::get(ConvertTy);
5902 for (Instruction *D : Defs) {
5903 if (isa<BitCastInst>(D)) {
5904 ValMap[D] = D->getOperand(0);
5905 DeletedInstrs.insert(D);
5906 } else {
5907 ValMap[D] =
5908 new BitCastInst(D, ConvertTy, D->getName() + ".bc", D->getNextNode());
5909 }
5910 }
5911 for (PHINode *Phi : PhiNodes)
5912 ValMap[Phi] = PHINode::Create(ConvertTy, Phi->getNumIncomingValues(),
5913 Phi->getName() + ".tc", Phi);
5914 // Pipe together all the PhiNodes.
5915 for (PHINode *Phi : PhiNodes) {
5916 PHINode *NewPhi = cast<PHINode>(ValMap[Phi]);
5917 for (int i = 0, e = Phi->getNumIncomingValues(); i < e; i++)
5918 NewPhi->addIncoming(ValMap[Phi->getIncomingValue(i)],
5919 Phi->getIncomingBlock(i));
5920 Visited.insert(NewPhi);
5921 }
5922 // And finally pipe up the stores and bitcasts
5923 for (Instruction *U : Uses) {
5924 if (isa<BitCastInst>(U)) {
5925 DeletedInstrs.insert(U);
5926 U->replaceAllUsesWith(ValMap[U->getOperand(0)]);
5927 } else {
5928 U->setOperand(0,
5929 new BitCastInst(ValMap[U->getOperand(0)], PhiTy, "bc", U));
5930 }
5931 }
5932
5933 // Save the removed phis to be deleted later.
5934 for (PHINode *Phi : PhiNodes)
5935 DeletedInstrs.insert(Phi);
5936 return true;
5937 }
5938
optimizePhiTypes(Function & F)5939 bool CodeGenPrepare::optimizePhiTypes(Function &F) {
5940 if (!OptimizePhiTypes)
5941 return false;
5942
5943 bool Changed = false;
5944 SmallPtrSet<PHINode *, 4> Visited;
5945 SmallPtrSet<Instruction *, 4> DeletedInstrs;
5946
5947 // Attempt to optimize all the phis in the functions to the correct type.
5948 for (auto &BB : F)
5949 for (auto &Phi : BB.phis())
5950 Changed |= optimizePhiType(&Phi, Visited, DeletedInstrs);
5951
5952 // Remove any old phi's that have been converted.
5953 for (auto *I : DeletedInstrs) {
5954 I->replaceAllUsesWith(UndefValue::get(I->getType()));
5955 I->eraseFromParent();
5956 }
5957
5958 return Changed;
5959 }
5960
5961 /// Return true, if an ext(load) can be formed from an extension in
5962 /// \p MovedExts.
canFormExtLd(const SmallVectorImpl<Instruction * > & MovedExts,LoadInst * & LI,Instruction * & Inst,bool HasPromoted)5963 bool CodeGenPrepare::canFormExtLd(
5964 const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
5965 Instruction *&Inst, bool HasPromoted) {
5966 for (auto *MovedExtInst : MovedExts) {
5967 if (isa<LoadInst>(MovedExtInst->getOperand(0))) {
5968 LI = cast<LoadInst>(MovedExtInst->getOperand(0));
5969 Inst = MovedExtInst;
5970 break;
5971 }
5972 }
5973 if (!LI)
5974 return false;
5975
5976 // If they're already in the same block, there's nothing to do.
5977 // Make the cheap checks first if we did not promote.
5978 // If we promoted, we need to check if it is indeed profitable.
5979 if (!HasPromoted && LI->getParent() == Inst->getParent())
5980 return false;
5981
5982 return TLI->isExtLoad(LI, Inst, *DL);
5983 }
5984
5985 /// Move a zext or sext fed by a load into the same basic block as the load,
5986 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
5987 /// extend into the load.
5988 ///
5989 /// E.g.,
5990 /// \code
5991 /// %ld = load i32* %addr
5992 /// %add = add nuw i32 %ld, 4
5993 /// %zext = zext i32 %add to i64
5994 // \endcode
5995 /// =>
5996 /// \code
5997 /// %ld = load i32* %addr
5998 /// %zext = zext i32 %ld to i64
5999 /// %add = add nuw i64 %zext, 4
6000 /// \encode
6001 /// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
6002 /// allow us to match zext(load i32*) to i64.
6003 ///
6004 /// Also, try to promote the computations used to obtain a sign extended
6005 /// value used into memory accesses.
6006 /// E.g.,
6007 /// \code
6008 /// a = add nsw i32 b, 3
6009 /// d = sext i32 a to i64
6010 /// e = getelementptr ..., i64 d
6011 /// \endcode
6012 /// =>
6013 /// \code
6014 /// f = sext i32 b to i64
6015 /// a = add nsw i64 f, 3
6016 /// e = getelementptr ..., i64 a
6017 /// \endcode
6018 ///
6019 /// \p Inst[in/out] the extension may be modified during the process if some
6020 /// promotions apply.
optimizeExt(Instruction * & Inst)6021 bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
6022 bool AllowPromotionWithoutCommonHeader = false;
6023 /// See if it is an interesting sext operations for the address type
6024 /// promotion before trying to promote it, e.g., the ones with the right
6025 /// type and used in memory accesses.
6026 bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
6027 *Inst, AllowPromotionWithoutCommonHeader);
6028 TypePromotionTransaction TPT(RemovedInsts);
6029 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
6030 TPT.getRestorationPoint();
6031 SmallVector<Instruction *, 1> Exts;
6032 SmallVector<Instruction *, 2> SpeculativelyMovedExts;
6033 Exts.push_back(Inst);
6034
6035 bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts);
6036
6037 // Look for a load being extended.
6038 LoadInst *LI = nullptr;
6039 Instruction *ExtFedByLoad;
6040
6041 // Try to promote a chain of computation if it allows to form an extended
6042 // load.
6043 if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) {
6044 assert(LI && ExtFedByLoad && "Expect a valid load and extension");
6045 TPT.commit();
6046 // Move the extend into the same block as the load.
6047 ExtFedByLoad->moveAfter(LI);
6048 ++NumExtsMoved;
6049 Inst = ExtFedByLoad;
6050 return true;
6051 }
6052
6053 // Continue promoting SExts if known as considerable depending on targets.
6054 if (ATPConsiderable &&
6055 performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
6056 HasPromoted, TPT, SpeculativelyMovedExts))
6057 return true;
6058
6059 TPT.rollback(LastKnownGood);
6060 return false;
6061 }
6062
6063 // Perform address type promotion if doing so is profitable.
6064 // If AllowPromotionWithoutCommonHeader == false, we should find other sext
6065 // instructions that sign extended the same initial value. However, if
6066 // AllowPromotionWithoutCommonHeader == true, we expect promoting the
6067 // extension is just profitable.
performAddressTypePromotion(Instruction * & Inst,bool AllowPromotionWithoutCommonHeader,bool HasPromoted,TypePromotionTransaction & TPT,SmallVectorImpl<Instruction * > & SpeculativelyMovedExts)6068 bool CodeGenPrepare::performAddressTypePromotion(
6069 Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
6070 bool HasPromoted, TypePromotionTransaction &TPT,
6071 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
6072 bool Promoted = false;
6073 SmallPtrSet<Instruction *, 1> UnhandledExts;
6074 bool AllSeenFirst = true;
6075 for (auto *I : SpeculativelyMovedExts) {
6076 Value *HeadOfChain = I->getOperand(0);
6077 DenseMap<Value *, Instruction *>::iterator AlreadySeen =
6078 SeenChainsForSExt.find(HeadOfChain);
6079 // If there is an unhandled SExt which has the same header, try to promote
6080 // it as well.
6081 if (AlreadySeen != SeenChainsForSExt.end()) {
6082 if (AlreadySeen->second != nullptr)
6083 UnhandledExts.insert(AlreadySeen->second);
6084 AllSeenFirst = false;
6085 }
6086 }
6087
6088 if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
6089 SpeculativelyMovedExts.size() == 1)) {
6090 TPT.commit();
6091 if (HasPromoted)
6092 Promoted = true;
6093 for (auto *I : SpeculativelyMovedExts) {
6094 Value *HeadOfChain = I->getOperand(0);
6095 SeenChainsForSExt[HeadOfChain] = nullptr;
6096 ValToSExtendedUses[HeadOfChain].push_back(I);
6097 }
6098 // Update Inst as promotion happen.
6099 Inst = SpeculativelyMovedExts.pop_back_val();
6100 } else {
6101 // This is the first chain visited from the header, keep the current chain
6102 // as unhandled. Defer to promote this until we encounter another SExt
6103 // chain derived from the same header.
6104 for (auto *I : SpeculativelyMovedExts) {
6105 Value *HeadOfChain = I->getOperand(0);
6106 SeenChainsForSExt[HeadOfChain] = Inst;
6107 }
6108 return false;
6109 }
6110
6111 if (!AllSeenFirst && !UnhandledExts.empty())
6112 for (auto *VisitedSExt : UnhandledExts) {
6113 if (RemovedInsts.count(VisitedSExt))
6114 continue;
6115 TypePromotionTransaction TPT(RemovedInsts);
6116 SmallVector<Instruction *, 1> Exts;
6117 SmallVector<Instruction *, 2> Chains;
6118 Exts.push_back(VisitedSExt);
6119 bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains);
6120 TPT.commit();
6121 if (HasPromoted)
6122 Promoted = true;
6123 for (auto *I : Chains) {
6124 Value *HeadOfChain = I->getOperand(0);
6125 // Mark this as handled.
6126 SeenChainsForSExt[HeadOfChain] = nullptr;
6127 ValToSExtendedUses[HeadOfChain].push_back(I);
6128 }
6129 }
6130 return Promoted;
6131 }
6132
optimizeExtUses(Instruction * I)6133 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
6134 BasicBlock *DefBB = I->getParent();
6135
6136 // If the result of a {s|z}ext and its source are both live out, rewrite all
6137 // other uses of the source with result of extension.
6138 Value *Src = I->getOperand(0);
6139 if (Src->hasOneUse())
6140 return false;
6141
6142 // Only do this xform if truncating is free.
6143 if (!TLI->isTruncateFree(I->getType(), Src->getType()))
6144 return false;
6145
6146 // Only safe to perform the optimization if the source is also defined in
6147 // this block.
6148 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
6149 return false;
6150
6151 bool DefIsLiveOut = false;
6152 for (User *U : I->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 DefIsLiveOut = true;
6159 break;
6160 }
6161 if (!DefIsLiveOut)
6162 return false;
6163
6164 // Make sure none of the uses are PHI nodes.
6165 for (User *U : Src->users()) {
6166 Instruction *UI = cast<Instruction>(U);
6167 BasicBlock *UserBB = UI->getParent();
6168 if (UserBB == DefBB) continue;
6169 // Be conservative. We don't want this xform to end up introducing
6170 // reloads just before load / store instructions.
6171 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
6172 return false;
6173 }
6174
6175 // InsertedTruncs - Only insert one trunc in each block once.
6176 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
6177
6178 bool MadeChange = false;
6179 for (Use &U : Src->uses()) {
6180 Instruction *User = cast<Instruction>(U.getUser());
6181
6182 // Figure out which BB this ext is used in.
6183 BasicBlock *UserBB = User->getParent();
6184 if (UserBB == DefBB) continue;
6185
6186 // Both src and def are live in this block. Rewrite the use.
6187 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
6188
6189 if (!InsertedTrunc) {
6190 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
6191 assert(InsertPt != UserBB->end());
6192 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
6193 InsertedInsts.insert(InsertedTrunc);
6194 }
6195
6196 // Replace a use of the {s|z}ext source with a use of the result.
6197 U = InsertedTrunc;
6198 ++NumExtUses;
6199 MadeChange = true;
6200 }
6201
6202 return MadeChange;
6203 }
6204
6205 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
6206 // just after the load if the target can fold this into one extload instruction,
6207 // with the hope of eliminating some of the other later "and" instructions using
6208 // the loaded value. "and"s that are made trivially redundant by the insertion
6209 // of the new "and" are removed by this function, while others (e.g. those whose
6210 // path from the load goes through a phi) are left for isel to potentially
6211 // remove.
6212 //
6213 // For example:
6214 //
6215 // b0:
6216 // x = load i32
6217 // ...
6218 // b1:
6219 // y = and x, 0xff
6220 // z = use y
6221 //
6222 // becomes:
6223 //
6224 // b0:
6225 // x = load i32
6226 // x' = and x, 0xff
6227 // ...
6228 // b1:
6229 // z = use x'
6230 //
6231 // whereas:
6232 //
6233 // b0:
6234 // x1 = load i32
6235 // ...
6236 // b1:
6237 // x2 = load i32
6238 // ...
6239 // b2:
6240 // x = phi x1, x2
6241 // y = and x, 0xff
6242 //
6243 // becomes (after a call to optimizeLoadExt for each load):
6244 //
6245 // b0:
6246 // x1 = load i32
6247 // x1' = and x1, 0xff
6248 // ...
6249 // b1:
6250 // x2 = load i32
6251 // x2' = and x2, 0xff
6252 // ...
6253 // b2:
6254 // x = phi x1', x2'
6255 // y = and x, 0xff
optimizeLoadExt(LoadInst * Load)6256 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
6257 if (!Load->isSimple() || !Load->getType()->isIntOrPtrTy())
6258 return false;
6259
6260 // Skip loads we've already transformed.
6261 if (Load->hasOneUse() &&
6262 InsertedInsts.count(cast<Instruction>(*Load->user_begin())))
6263 return false;
6264
6265 // Look at all uses of Load, looking through phis, to determine how many bits
6266 // of the loaded value are needed.
6267 SmallVector<Instruction *, 8> WorkList;
6268 SmallPtrSet<Instruction *, 16> Visited;
6269 SmallVector<Instruction *, 8> AndsToMaybeRemove;
6270 for (auto *U : Load->users())
6271 WorkList.push_back(cast<Instruction>(U));
6272
6273 EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
6274 unsigned BitWidth = LoadResultVT.getSizeInBits();
6275 APInt DemandBits(BitWidth, 0);
6276 APInt WidestAndBits(BitWidth, 0);
6277
6278 while (!WorkList.empty()) {
6279 Instruction *I = WorkList.back();
6280 WorkList.pop_back();
6281
6282 // Break use-def graph loops.
6283 if (!Visited.insert(I).second)
6284 continue;
6285
6286 // For a PHI node, push all of its users.
6287 if (auto *Phi = dyn_cast<PHINode>(I)) {
6288 for (auto *U : Phi->users())
6289 WorkList.push_back(cast<Instruction>(U));
6290 continue;
6291 }
6292
6293 switch (I->getOpcode()) {
6294 case Instruction::And: {
6295 auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
6296 if (!AndC)
6297 return false;
6298 APInt AndBits = AndC->getValue();
6299 DemandBits |= AndBits;
6300 // Keep track of the widest and mask we see.
6301 if (AndBits.ugt(WidestAndBits))
6302 WidestAndBits = AndBits;
6303 if (AndBits == WidestAndBits && I->getOperand(0) == Load)
6304 AndsToMaybeRemove.push_back(I);
6305 break;
6306 }
6307
6308 case Instruction::Shl: {
6309 auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
6310 if (!ShlC)
6311 return false;
6312 uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
6313 DemandBits.setLowBits(BitWidth - ShiftAmt);
6314 break;
6315 }
6316
6317 case Instruction::Trunc: {
6318 EVT TruncVT = TLI->getValueType(*DL, I->getType());
6319 unsigned TruncBitWidth = TruncVT.getSizeInBits();
6320 DemandBits.setLowBits(TruncBitWidth);
6321 break;
6322 }
6323
6324 default:
6325 return false;
6326 }
6327 }
6328
6329 uint32_t ActiveBits = DemandBits.getActiveBits();
6330 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
6331 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
6332 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
6333 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
6334 // followed by an AND.
6335 // TODO: Look into removing this restriction by fixing backends to either
6336 // return false for isLoadExtLegal for i1 or have them select this pattern to
6337 // a single instruction.
6338 //
6339 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
6340 // mask, since these are the only ands that will be removed by isel.
6341 if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) ||
6342 WidestAndBits != DemandBits)
6343 return false;
6344
6345 LLVMContext &Ctx = Load->getType()->getContext();
6346 Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
6347 EVT TruncVT = TLI->getValueType(*DL, TruncTy);
6348
6349 // Reject cases that won't be matched as extloads.
6350 if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
6351 !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
6352 return false;
6353
6354 IRBuilder<> Builder(Load->getNextNode());
6355 auto *NewAnd = cast<Instruction>(
6356 Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
6357 // Mark this instruction as "inserted by CGP", so that other
6358 // optimizations don't touch it.
6359 InsertedInsts.insert(NewAnd);
6360
6361 // Replace all uses of load with new and (except for the use of load in the
6362 // new and itself).
6363 Load->replaceAllUsesWith(NewAnd);
6364 NewAnd->setOperand(0, Load);
6365
6366 // Remove any and instructions that are now redundant.
6367 for (auto *And : AndsToMaybeRemove)
6368 // Check that the and mask is the same as the one we decided to put on the
6369 // new and.
6370 if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
6371 And->replaceAllUsesWith(NewAnd);
6372 if (&*CurInstIterator == And)
6373 CurInstIterator = std::next(And->getIterator());
6374 And->eraseFromParent();
6375 ++NumAndUses;
6376 }
6377
6378 ++NumAndsAdded;
6379 return true;
6380 }
6381
6382 /// Check if V (an operand of a select instruction) is an expensive instruction
6383 /// that is only used once.
sinkSelectOperand(const TargetTransformInfo * TTI,Value * V)6384 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
6385 auto *I = dyn_cast<Instruction>(V);
6386 // If it's safe to speculatively execute, then it should not have side
6387 // effects; therefore, it's safe to sink and possibly *not* execute.
6388 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
6389 TTI->getUserCost(I, TargetTransformInfo::TCK_SizeAndLatency) >=
6390 TargetTransformInfo::TCC_Expensive;
6391 }
6392
6393 /// Returns true if a SelectInst should be turned into an explicit branch.
isFormingBranchFromSelectProfitable(const TargetTransformInfo * TTI,const TargetLowering * TLI,SelectInst * SI)6394 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
6395 const TargetLowering *TLI,
6396 SelectInst *SI) {
6397 // If even a predictable select is cheap, then a branch can't be cheaper.
6398 if (!TLI->isPredictableSelectExpensive())
6399 return false;
6400
6401 // FIXME: This should use the same heuristics as IfConversion to determine
6402 // whether a select is better represented as a branch.
6403
6404 // If metadata tells us that the select condition is obviously predictable,
6405 // then we want to replace the select with a branch.
6406 uint64_t TrueWeight, FalseWeight;
6407 if (SI->extractProfMetadata(TrueWeight, FalseWeight)) {
6408 uint64_t Max = std::max(TrueWeight, FalseWeight);
6409 uint64_t Sum = TrueWeight + FalseWeight;
6410 if (Sum != 0) {
6411 auto Probability = BranchProbability::getBranchProbability(Max, Sum);
6412 if (Probability > TLI->getPredictableBranchThreshold())
6413 return true;
6414 }
6415 }
6416
6417 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
6418
6419 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
6420 // comparison condition. If the compare has more than one use, there's
6421 // probably another cmov or setcc around, so it's not worth emitting a branch.
6422 if (!Cmp || !Cmp->hasOneUse())
6423 return false;
6424
6425 // If either operand of the select is expensive and only needed on one side
6426 // of the select, we should form a branch.
6427 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
6428 sinkSelectOperand(TTI, SI->getFalseValue()))
6429 return true;
6430
6431 return false;
6432 }
6433
6434 /// If \p isTrue is true, return the true value of \p SI, otherwise return
6435 /// false value of \p SI. If the true/false value of \p SI is defined by any
6436 /// select instructions in \p Selects, look through the defining select
6437 /// instruction until the true/false value is not defined in \p Selects.
getTrueOrFalseValue(SelectInst * SI,bool isTrue,const SmallPtrSet<const Instruction *,2> & Selects)6438 static Value *getTrueOrFalseValue(
6439 SelectInst *SI, bool isTrue,
6440 const SmallPtrSet<const Instruction *, 2> &Selects) {
6441 Value *V = nullptr;
6442
6443 for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI);
6444 DefSI = dyn_cast<SelectInst>(V)) {
6445 assert(DefSI->getCondition() == SI->getCondition() &&
6446 "The condition of DefSI does not match with SI");
6447 V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
6448 }
6449
6450 assert(V && "Failed to get select true/false value");
6451 return V;
6452 }
6453
optimizeShiftInst(BinaryOperator * Shift)6454 bool CodeGenPrepare::optimizeShiftInst(BinaryOperator *Shift) {
6455 assert(Shift->isShift() && "Expected a shift");
6456
6457 // If this is (1) a vector shift, (2) shifts by scalars are cheaper than
6458 // general vector shifts, and (3) the shift amount is a select-of-splatted
6459 // values, hoist the shifts before the select:
6460 // shift Op0, (select Cond, TVal, FVal) -->
6461 // select Cond, (shift Op0, TVal), (shift Op0, FVal)
6462 //
6463 // This is inverting a generic IR transform when we know that the cost of a
6464 // general vector shift is more than the cost of 2 shift-by-scalars.
6465 // We can't do this effectively in SDAG because we may not be able to
6466 // determine if the select operands are splats from within a basic block.
6467 Type *Ty = Shift->getType();
6468 if (!Ty->isVectorTy() || !TLI->isVectorShiftByScalarCheap(Ty))
6469 return false;
6470 Value *Cond, *TVal, *FVal;
6471 if (!match(Shift->getOperand(1),
6472 m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))))
6473 return false;
6474 if (!isSplatValue(TVal) || !isSplatValue(FVal))
6475 return false;
6476
6477 IRBuilder<> Builder(Shift);
6478 BinaryOperator::BinaryOps Opcode = Shift->getOpcode();
6479 Value *NewTVal = Builder.CreateBinOp(Opcode, Shift->getOperand(0), TVal);
6480 Value *NewFVal = Builder.CreateBinOp(Opcode, Shift->getOperand(0), FVal);
6481 Value *NewSel = Builder.CreateSelect(Cond, NewTVal, NewFVal);
6482 Shift->replaceAllUsesWith(NewSel);
6483 Shift->eraseFromParent();
6484 return true;
6485 }
6486
optimizeFunnelShift(IntrinsicInst * Fsh)6487 bool CodeGenPrepare::optimizeFunnelShift(IntrinsicInst *Fsh) {
6488 Intrinsic::ID Opcode = Fsh->getIntrinsicID();
6489 assert((Opcode == Intrinsic::fshl || Opcode == Intrinsic::fshr) &&
6490 "Expected a funnel shift");
6491
6492 // If this is (1) a vector funnel shift, (2) shifts by scalars are cheaper
6493 // than general vector shifts, and (3) the shift amount is select-of-splatted
6494 // values, hoist the funnel shifts before the select:
6495 // fsh Op0, Op1, (select Cond, TVal, FVal) -->
6496 // select Cond, (fsh Op0, Op1, TVal), (fsh Op0, Op1, FVal)
6497 //
6498 // This is inverting a generic IR transform when we know that the cost of a
6499 // general vector shift is more than the cost of 2 shift-by-scalars.
6500 // We can't do this effectively in SDAG because we may not be able to
6501 // determine if the select operands are splats from within a basic block.
6502 Type *Ty = Fsh->getType();
6503 if (!Ty->isVectorTy() || !TLI->isVectorShiftByScalarCheap(Ty))
6504 return false;
6505 Value *Cond, *TVal, *FVal;
6506 if (!match(Fsh->getOperand(2),
6507 m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))))
6508 return false;
6509 if (!isSplatValue(TVal) || !isSplatValue(FVal))
6510 return false;
6511
6512 IRBuilder<> Builder(Fsh);
6513 Value *X = Fsh->getOperand(0), *Y = Fsh->getOperand(1);
6514 Value *NewTVal = Builder.CreateIntrinsic(Opcode, Ty, { X, Y, TVal });
6515 Value *NewFVal = Builder.CreateIntrinsic(Opcode, Ty, { X, Y, FVal });
6516 Value *NewSel = Builder.CreateSelect(Cond, NewTVal, NewFVal);
6517 Fsh->replaceAllUsesWith(NewSel);
6518 Fsh->eraseFromParent();
6519 return true;
6520 }
6521
6522 /// If we have a SelectInst that will likely profit from branch prediction,
6523 /// turn it into a branch.
optimizeSelectInst(SelectInst * SI)6524 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
6525 if (DisableSelectToBranch)
6526 return false;
6527
6528 // Find all consecutive select instructions that share the same condition.
6529 SmallVector<SelectInst *, 2> ASI;
6530 ASI.push_back(SI);
6531 for (BasicBlock::iterator It = ++BasicBlock::iterator(SI);
6532 It != SI->getParent()->end(); ++It) {
6533 SelectInst *I = dyn_cast<SelectInst>(&*It);
6534 if (I && SI->getCondition() == I->getCondition()) {
6535 ASI.push_back(I);
6536 } else {
6537 break;
6538 }
6539 }
6540
6541 SelectInst *LastSI = ASI.back();
6542 // Increment the current iterator to skip all the rest of select instructions
6543 // because they will be either "not lowered" or "all lowered" to branch.
6544 CurInstIterator = std::next(LastSI->getIterator());
6545
6546 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
6547
6548 // Can we convert the 'select' to CF ?
6549 if (VectorCond || SI->getMetadata(LLVMContext::MD_unpredictable))
6550 return false;
6551
6552 TargetLowering::SelectSupportKind SelectKind;
6553 if (VectorCond)
6554 SelectKind = TargetLowering::VectorMaskSelect;
6555 else if (SI->getType()->isVectorTy())
6556 SelectKind = TargetLowering::ScalarCondVectorVal;
6557 else
6558 SelectKind = TargetLowering::ScalarValSelect;
6559
6560 if (TLI->isSelectSupported(SelectKind) &&
6561 (!isFormingBranchFromSelectProfitable(TTI, TLI, SI) || OptSize ||
6562 llvm::shouldOptimizeForSize(SI->getParent(), PSI, BFI.get())))
6563 return false;
6564
6565 // The DominatorTree needs to be rebuilt by any consumers after this
6566 // transformation. We simply reset here rather than setting the ModifiedDT
6567 // flag to avoid restarting the function walk in runOnFunction for each
6568 // select optimized.
6569 DT.reset();
6570
6571 // Transform a sequence like this:
6572 // start:
6573 // %cmp = cmp uge i32 %a, %b
6574 // %sel = select i1 %cmp, i32 %c, i32 %d
6575 //
6576 // Into:
6577 // start:
6578 // %cmp = cmp uge i32 %a, %b
6579 // %cmp.frozen = freeze %cmp
6580 // br i1 %cmp.frozen, label %select.true, label %select.false
6581 // select.true:
6582 // br label %select.end
6583 // select.false:
6584 // br label %select.end
6585 // select.end:
6586 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
6587 //
6588 // %cmp should be frozen, otherwise it may introduce undefined behavior.
6589 // In addition, we may sink instructions that produce %c or %d from
6590 // the entry block into the destination(s) of the new branch.
6591 // If the true or false blocks do not contain a sunken instruction, that
6592 // block and its branch may be optimized away. In that case, one side of the
6593 // first branch will point directly to select.end, and the corresponding PHI
6594 // predecessor block will be the start block.
6595
6596 // First, we split the block containing the select into 2 blocks.
6597 BasicBlock *StartBlock = SI->getParent();
6598 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
6599 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
6600 BFI->setBlockFreq(EndBlock, BFI->getBlockFreq(StartBlock).getFrequency());
6601
6602 // Delete the unconditional branch that was just created by the split.
6603 StartBlock->getTerminator()->eraseFromParent();
6604
6605 // These are the new basic blocks for the conditional branch.
6606 // At least one will become an actual new basic block.
6607 BasicBlock *TrueBlock = nullptr;
6608 BasicBlock *FalseBlock = nullptr;
6609 BranchInst *TrueBranch = nullptr;
6610 BranchInst *FalseBranch = nullptr;
6611
6612 // Sink expensive instructions into the conditional blocks to avoid executing
6613 // them speculatively.
6614 for (SelectInst *SI : ASI) {
6615 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
6616 if (TrueBlock == nullptr) {
6617 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
6618 EndBlock->getParent(), EndBlock);
6619 TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
6620 TrueBranch->setDebugLoc(SI->getDebugLoc());
6621 }
6622 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
6623 TrueInst->moveBefore(TrueBranch);
6624 }
6625 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
6626 if (FalseBlock == nullptr) {
6627 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
6628 EndBlock->getParent(), EndBlock);
6629 FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
6630 FalseBranch->setDebugLoc(SI->getDebugLoc());
6631 }
6632 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
6633 FalseInst->moveBefore(FalseBranch);
6634 }
6635 }
6636
6637 // If there was nothing to sink, then arbitrarily choose the 'false' side
6638 // for a new input value to the PHI.
6639 if (TrueBlock == FalseBlock) {
6640 assert(TrueBlock == nullptr &&
6641 "Unexpected basic block transform while optimizing select");
6642
6643 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
6644 EndBlock->getParent(), EndBlock);
6645 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
6646 FalseBranch->setDebugLoc(SI->getDebugLoc());
6647 }
6648
6649 // Insert the real conditional branch based on the original condition.
6650 // If we did not create a new block for one of the 'true' or 'false' paths
6651 // of the condition, it means that side of the branch goes to the end block
6652 // directly and the path originates from the start block from the point of
6653 // view of the new PHI.
6654 BasicBlock *TT, *FT;
6655 if (TrueBlock == nullptr) {
6656 TT = EndBlock;
6657 FT = FalseBlock;
6658 TrueBlock = StartBlock;
6659 } else if (FalseBlock == nullptr) {
6660 TT = TrueBlock;
6661 FT = EndBlock;
6662 FalseBlock = StartBlock;
6663 } else {
6664 TT = TrueBlock;
6665 FT = FalseBlock;
6666 }
6667 IRBuilder<> IB(SI);
6668 auto *CondFr = IB.CreateFreeze(SI->getCondition(), SI->getName() + ".frozen");
6669 IB.CreateCondBr(CondFr, TT, FT, SI);
6670
6671 SmallPtrSet<const Instruction *, 2> INS;
6672 INS.insert(ASI.begin(), ASI.end());
6673 // Use reverse iterator because later select may use the value of the
6674 // earlier select, and we need to propagate value through earlier select
6675 // to get the PHI operand.
6676 for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) {
6677 SelectInst *SI = *It;
6678 // The select itself is replaced with a PHI Node.
6679 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
6680 PN->takeName(SI);
6681 PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock);
6682 PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock);
6683 PN->setDebugLoc(SI->getDebugLoc());
6684
6685 SI->replaceAllUsesWith(PN);
6686 SI->eraseFromParent();
6687 INS.erase(SI);
6688 ++NumSelectsExpanded;
6689 }
6690
6691 // Instruct OptimizeBlock to skip to the next block.
6692 CurInstIterator = StartBlock->end();
6693 return true;
6694 }
6695
6696 /// Some targets only accept certain types for splat inputs. For example a VDUP
6697 /// in MVE takes a GPR (integer) register, and the instruction that incorporate
6698 /// a VDUP (such as a VADD qd, qm, rm) also require a gpr register.
optimizeShuffleVectorInst(ShuffleVectorInst * SVI)6699 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
6700 // Accept shuf(insertelem(undef/poison, val, 0), undef/poison, <0,0,..>) only
6701 if (!match(SVI, m_Shuffle(m_InsertElt(m_Undef(), m_Value(), m_ZeroInt()),
6702 m_Undef(), m_ZeroMask())))
6703 return false;
6704 Type *NewType = TLI->shouldConvertSplatType(SVI);
6705 if (!NewType)
6706 return false;
6707
6708 auto *SVIVecType = cast<FixedVectorType>(SVI->getType());
6709 assert(!NewType->isVectorTy() && "Expected a scalar type!");
6710 assert(NewType->getScalarSizeInBits() == SVIVecType->getScalarSizeInBits() &&
6711 "Expected a type of the same size!");
6712 auto *NewVecType =
6713 FixedVectorType::get(NewType, SVIVecType->getNumElements());
6714
6715 // Create a bitcast (shuffle (insert (bitcast(..))))
6716 IRBuilder<> Builder(SVI->getContext());
6717 Builder.SetInsertPoint(SVI);
6718 Value *BC1 = Builder.CreateBitCast(
6719 cast<Instruction>(SVI->getOperand(0))->getOperand(1), NewType);
6720 Value *Shuffle = Builder.CreateVectorSplat(NewVecType->getNumElements(), BC1);
6721 Value *BC2 = Builder.CreateBitCast(Shuffle, SVIVecType);
6722
6723 SVI->replaceAllUsesWith(BC2);
6724 RecursivelyDeleteTriviallyDeadInstructions(
6725 SVI, TLInfo, nullptr, [&](Value *V) { removeAllAssertingVHReferences(V); });
6726
6727 // Also hoist the bitcast up to its operand if it they are not in the same
6728 // block.
6729 if (auto *BCI = dyn_cast<Instruction>(BC1))
6730 if (auto *Op = dyn_cast<Instruction>(BCI->getOperand(0)))
6731 if (BCI->getParent() != Op->getParent() && !isa<PHINode>(Op) &&
6732 !Op->isTerminator() && !Op->isEHPad())
6733 BCI->moveAfter(Op);
6734
6735 return true;
6736 }
6737
tryToSinkFreeOperands(Instruction * I)6738 bool CodeGenPrepare::tryToSinkFreeOperands(Instruction *I) {
6739 // If the operands of I can be folded into a target instruction together with
6740 // I, duplicate and sink them.
6741 SmallVector<Use *, 4> OpsToSink;
6742 if (!TLI->shouldSinkOperands(I, OpsToSink))
6743 return false;
6744
6745 // OpsToSink can contain multiple uses in a use chain (e.g.
6746 // (%u1 with %u1 = shufflevector), (%u2 with %u2 = zext %u1)). The dominating
6747 // uses must come first, so we process the ops in reverse order so as to not
6748 // create invalid IR.
6749 BasicBlock *TargetBB = I->getParent();
6750 bool Changed = false;
6751 SmallVector<Use *, 4> ToReplace;
6752 for (Use *U : reverse(OpsToSink)) {
6753 auto *UI = cast<Instruction>(U->get());
6754 if (UI->getParent() == TargetBB || isa<PHINode>(UI))
6755 continue;
6756 ToReplace.push_back(U);
6757 }
6758
6759 SetVector<Instruction *> MaybeDead;
6760 DenseMap<Instruction *, Instruction *> NewInstructions;
6761 Instruction *InsertPoint = I;
6762 for (Use *U : ToReplace) {
6763 auto *UI = cast<Instruction>(U->get());
6764 Instruction *NI = UI->clone();
6765 NewInstructions[UI] = NI;
6766 MaybeDead.insert(UI);
6767 LLVM_DEBUG(dbgs() << "Sinking " << *UI << " to user " << *I << "\n");
6768 NI->insertBefore(InsertPoint);
6769 InsertPoint = NI;
6770 InsertedInsts.insert(NI);
6771
6772 // Update the use for the new instruction, making sure that we update the
6773 // sunk instruction uses, if it is part of a chain that has already been
6774 // sunk.
6775 Instruction *OldI = cast<Instruction>(U->getUser());
6776 if (NewInstructions.count(OldI))
6777 NewInstructions[OldI]->setOperand(U->getOperandNo(), NI);
6778 else
6779 U->set(NI);
6780 Changed = true;
6781 }
6782
6783 // Remove instructions that are dead after sinking.
6784 for (auto *I : MaybeDead) {
6785 if (!I->hasNUsesOrMore(1)) {
6786 LLVM_DEBUG(dbgs() << "Removing dead instruction: " << *I << "\n");
6787 I->eraseFromParent();
6788 }
6789 }
6790
6791 return Changed;
6792 }
6793
optimizeSwitchInst(SwitchInst * SI)6794 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
6795 Value *Cond = SI->getCondition();
6796 Type *OldType = Cond->getType();
6797 LLVMContext &Context = Cond->getContext();
6798 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
6799 unsigned RegWidth = RegType.getSizeInBits();
6800
6801 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
6802 return false;
6803
6804 // If the register width is greater than the type width, expand the condition
6805 // of the switch instruction and each case constant to the width of the
6806 // register. By widening the type of the switch condition, subsequent
6807 // comparisons (for case comparisons) will not need to be extended to the
6808 // preferred register width, so we will potentially eliminate N-1 extends,
6809 // where N is the number of cases in the switch.
6810 auto *NewType = Type::getIntNTy(Context, RegWidth);
6811
6812 // Zero-extend the switch condition and case constants unless the switch
6813 // condition is a function argument that is already being sign-extended.
6814 // In that case, we can avoid an unnecessary mask/extension by sign-extending
6815 // everything instead.
6816 Instruction::CastOps ExtType = Instruction::ZExt;
6817 if (auto *Arg = dyn_cast<Argument>(Cond))
6818 if (Arg->hasSExtAttr())
6819 ExtType = Instruction::SExt;
6820
6821 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
6822 ExtInst->insertBefore(SI);
6823 ExtInst->setDebugLoc(SI->getDebugLoc());
6824 SI->setCondition(ExtInst);
6825 for (auto Case : SI->cases()) {
6826 APInt NarrowConst = Case.getCaseValue()->getValue();
6827 APInt WideConst = (ExtType == Instruction::ZExt) ?
6828 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
6829 Case.setValue(ConstantInt::get(Context, WideConst));
6830 }
6831
6832 return true;
6833 }
6834
6835
6836 namespace {
6837
6838 /// Helper class to promote a scalar operation to a vector one.
6839 /// This class is used to move downward extractelement transition.
6840 /// E.g.,
6841 /// a = vector_op <2 x i32>
6842 /// b = extractelement <2 x i32> a, i32 0
6843 /// c = scalar_op b
6844 /// store c
6845 ///
6846 /// =>
6847 /// a = vector_op <2 x i32>
6848 /// c = vector_op a (equivalent to scalar_op on the related lane)
6849 /// * d = extractelement <2 x i32> c, i32 0
6850 /// * store d
6851 /// Assuming both extractelement and store can be combine, we get rid of the
6852 /// transition.
6853 class VectorPromoteHelper {
6854 /// DataLayout associated with the current module.
6855 const DataLayout &DL;
6856
6857 /// Used to perform some checks on the legality of vector operations.
6858 const TargetLowering &TLI;
6859
6860 /// Used to estimated the cost of the promoted chain.
6861 const TargetTransformInfo &TTI;
6862
6863 /// The transition being moved downwards.
6864 Instruction *Transition;
6865
6866 /// The sequence of instructions to be promoted.
6867 SmallVector<Instruction *, 4> InstsToBePromoted;
6868
6869 /// Cost of combining a store and an extract.
6870 unsigned StoreExtractCombineCost;
6871
6872 /// Instruction that will be combined with the transition.
6873 Instruction *CombineInst = nullptr;
6874
6875 /// The instruction that represents the current end of the transition.
6876 /// Since we are faking the promotion until we reach the end of the chain
6877 /// of computation, we need a way to get the current end of the transition.
getEndOfTransition() const6878 Instruction *getEndOfTransition() const {
6879 if (InstsToBePromoted.empty())
6880 return Transition;
6881 return InstsToBePromoted.back();
6882 }
6883
6884 /// Return the index of the original value in the transition.
6885 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
6886 /// c, is at index 0.
getTransitionOriginalValueIdx() const6887 unsigned getTransitionOriginalValueIdx() const {
6888 assert(isa<ExtractElementInst>(Transition) &&
6889 "Other kind of transitions are not supported yet");
6890 return 0;
6891 }
6892
6893 /// Return the index of the index in the transition.
6894 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
6895 /// is at index 1.
getTransitionIdx() const6896 unsigned getTransitionIdx() const {
6897 assert(isa<ExtractElementInst>(Transition) &&
6898 "Other kind of transitions are not supported yet");
6899 return 1;
6900 }
6901
6902 /// Get the type of the transition.
6903 /// This is the type of the original value.
6904 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
6905 /// transition is <2 x i32>.
getTransitionType() const6906 Type *getTransitionType() const {
6907 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
6908 }
6909
6910 /// Promote \p ToBePromoted by moving \p Def downward through.
6911 /// I.e., we have the following sequence:
6912 /// Def = Transition <ty1> a to <ty2>
6913 /// b = ToBePromoted <ty2> Def, ...
6914 /// =>
6915 /// b = ToBePromoted <ty1> a, ...
6916 /// Def = Transition <ty1> ToBePromoted to <ty2>
6917 void promoteImpl(Instruction *ToBePromoted);
6918
6919 /// Check whether or not it is profitable to promote all the
6920 /// instructions enqueued to be promoted.
isProfitableToPromote()6921 bool isProfitableToPromote() {
6922 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
6923 unsigned Index = isa<ConstantInt>(ValIdx)
6924 ? cast<ConstantInt>(ValIdx)->getZExtValue()
6925 : -1;
6926 Type *PromotedType = getTransitionType();
6927
6928 StoreInst *ST = cast<StoreInst>(CombineInst);
6929 unsigned AS = ST->getPointerAddressSpace();
6930 unsigned Align = ST->getAlignment();
6931 // Check if this store is supported.
6932 if (!TLI.allowsMisalignedMemoryAccesses(
6933 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
6934 Align)) {
6935 // If this is not supported, there is no way we can combine
6936 // the extract with the store.
6937 return false;
6938 }
6939
6940 // The scalar chain of computation has to pay for the transition
6941 // scalar to vector.
6942 // The vector chain has to account for the combining cost.
6943 uint64_t ScalarCost =
6944 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
6945 uint64_t VectorCost = StoreExtractCombineCost;
6946 enum TargetTransformInfo::TargetCostKind CostKind =
6947 TargetTransformInfo::TCK_RecipThroughput;
6948 for (const auto &Inst : InstsToBePromoted) {
6949 // Compute the cost.
6950 // By construction, all instructions being promoted are arithmetic ones.
6951 // Moreover, one argument is a constant that can be viewed as a splat
6952 // constant.
6953 Value *Arg0 = Inst->getOperand(0);
6954 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
6955 isa<ConstantFP>(Arg0);
6956 TargetTransformInfo::OperandValueKind Arg0OVK =
6957 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
6958 : TargetTransformInfo::OK_AnyValue;
6959 TargetTransformInfo::OperandValueKind Arg1OVK =
6960 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
6961 : TargetTransformInfo::OK_AnyValue;
6962 ScalarCost += TTI.getArithmeticInstrCost(
6963 Inst->getOpcode(), Inst->getType(), CostKind, Arg0OVK, Arg1OVK);
6964 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
6965 CostKind,
6966 Arg0OVK, Arg1OVK);
6967 }
6968 LLVM_DEBUG(
6969 dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
6970 << ScalarCost << "\nVector: " << VectorCost << '\n');
6971 return ScalarCost > VectorCost;
6972 }
6973
6974 /// Generate a constant vector with \p Val with the same
6975 /// number of elements as the transition.
6976 /// \p UseSplat defines whether or not \p Val should be replicated
6977 /// across the whole vector.
6978 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
6979 /// otherwise we generate a vector with as many undef as possible:
6980 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
6981 /// used at the index of the extract.
getConstantVector(Constant * Val,bool UseSplat) const6982 Value *getConstantVector(Constant *Val, bool UseSplat) const {
6983 unsigned ExtractIdx = std::numeric_limits<unsigned>::max();
6984 if (!UseSplat) {
6985 // If we cannot determine where the constant must be, we have to
6986 // use a splat constant.
6987 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
6988 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
6989 ExtractIdx = CstVal->getSExtValue();
6990 else
6991 UseSplat = true;
6992 }
6993
6994 ElementCount EC = cast<VectorType>(getTransitionType())->getElementCount();
6995 if (UseSplat)
6996 return ConstantVector::getSplat(EC, Val);
6997
6998 if (!EC.isScalable()) {
6999 SmallVector<Constant *, 4> ConstVec;
7000 UndefValue *UndefVal = UndefValue::get(Val->getType());
7001 for (unsigned Idx = 0; Idx != EC.getKnownMinValue(); ++Idx) {
7002 if (Idx == ExtractIdx)
7003 ConstVec.push_back(Val);
7004 else
7005 ConstVec.push_back(UndefVal);
7006 }
7007 return ConstantVector::get(ConstVec);
7008 } else
7009 llvm_unreachable(
7010 "Generate scalable vector for non-splat is unimplemented");
7011 }
7012
7013 /// Check if promoting to a vector type an operand at \p OperandIdx
7014 /// in \p Use can trigger undefined behavior.
canCauseUndefinedBehavior(const Instruction * Use,unsigned OperandIdx)7015 static bool canCauseUndefinedBehavior(const Instruction *Use,
7016 unsigned OperandIdx) {
7017 // This is not safe to introduce undef when the operand is on
7018 // the right hand side of a division-like instruction.
7019 if (OperandIdx != 1)
7020 return false;
7021 switch (Use->getOpcode()) {
7022 default:
7023 return false;
7024 case Instruction::SDiv:
7025 case Instruction::UDiv:
7026 case Instruction::SRem:
7027 case Instruction::URem:
7028 return true;
7029 case Instruction::FDiv:
7030 case Instruction::FRem:
7031 return !Use->hasNoNaNs();
7032 }
7033 llvm_unreachable(nullptr);
7034 }
7035
7036 public:
VectorPromoteHelper(const DataLayout & DL,const TargetLowering & TLI,const TargetTransformInfo & TTI,Instruction * Transition,unsigned CombineCost)7037 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
7038 const TargetTransformInfo &TTI, Instruction *Transition,
7039 unsigned CombineCost)
7040 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
7041 StoreExtractCombineCost(CombineCost) {
7042 assert(Transition && "Do not know how to promote null");
7043 }
7044
7045 /// Check if we can promote \p ToBePromoted to \p Type.
canPromote(const Instruction * ToBePromoted) const7046 bool canPromote(const Instruction *ToBePromoted) const {
7047 // We could support CastInst too.
7048 return isa<BinaryOperator>(ToBePromoted);
7049 }
7050
7051 /// Check if it is profitable to promote \p ToBePromoted
7052 /// by moving downward the transition through.
shouldPromote(const Instruction * ToBePromoted) const7053 bool shouldPromote(const Instruction *ToBePromoted) const {
7054 // Promote only if all the operands can be statically expanded.
7055 // Indeed, we do not want to introduce any new kind of transitions.
7056 for (const Use &U : ToBePromoted->operands()) {
7057 const Value *Val = U.get();
7058 if (Val == getEndOfTransition()) {
7059 // If the use is a division and the transition is on the rhs,
7060 // we cannot promote the operation, otherwise we may create a
7061 // division by zero.
7062 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
7063 return false;
7064 continue;
7065 }
7066 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
7067 !isa<ConstantFP>(Val))
7068 return false;
7069 }
7070 // Check that the resulting operation is legal.
7071 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
7072 if (!ISDOpcode)
7073 return false;
7074 return StressStoreExtract ||
7075 TLI.isOperationLegalOrCustom(
7076 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
7077 }
7078
7079 /// Check whether or not \p Use can be combined
7080 /// with the transition.
7081 /// I.e., is it possible to do Use(Transition) => AnotherUse?
canCombine(const Instruction * Use)7082 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
7083
7084 /// Record \p ToBePromoted as part of the chain to be promoted.
enqueueForPromotion(Instruction * ToBePromoted)7085 void enqueueForPromotion(Instruction *ToBePromoted) {
7086 InstsToBePromoted.push_back(ToBePromoted);
7087 }
7088
7089 /// Set the instruction that will be combined with the transition.
recordCombineInstruction(Instruction * ToBeCombined)7090 void recordCombineInstruction(Instruction *ToBeCombined) {
7091 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
7092 CombineInst = ToBeCombined;
7093 }
7094
7095 /// Promote all the instructions enqueued for promotion if it is
7096 /// is profitable.
7097 /// \return True if the promotion happened, false otherwise.
promote()7098 bool promote() {
7099 // Check if there is something to promote.
7100 // Right now, if we do not have anything to combine with,
7101 // we assume the promotion is not profitable.
7102 if (InstsToBePromoted.empty() || !CombineInst)
7103 return false;
7104
7105 // Check cost.
7106 if (!StressStoreExtract && !isProfitableToPromote())
7107 return false;
7108
7109 // Promote.
7110 for (auto &ToBePromoted : InstsToBePromoted)
7111 promoteImpl(ToBePromoted);
7112 InstsToBePromoted.clear();
7113 return true;
7114 }
7115 };
7116
7117 } // end anonymous namespace
7118
promoteImpl(Instruction * ToBePromoted)7119 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
7120 // At this point, we know that all the operands of ToBePromoted but Def
7121 // can be statically promoted.
7122 // For Def, we need to use its parameter in ToBePromoted:
7123 // b = ToBePromoted ty1 a
7124 // Def = Transition ty1 b to ty2
7125 // Move the transition down.
7126 // 1. Replace all uses of the promoted operation by the transition.
7127 // = ... b => = ... Def.
7128 assert(ToBePromoted->getType() == Transition->getType() &&
7129 "The type of the result of the transition does not match "
7130 "the final type");
7131 ToBePromoted->replaceAllUsesWith(Transition);
7132 // 2. Update the type of the uses.
7133 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
7134 Type *TransitionTy = getTransitionType();
7135 ToBePromoted->mutateType(TransitionTy);
7136 // 3. Update all the operands of the promoted operation with promoted
7137 // operands.
7138 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
7139 for (Use &U : ToBePromoted->operands()) {
7140 Value *Val = U.get();
7141 Value *NewVal = nullptr;
7142 if (Val == Transition)
7143 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
7144 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
7145 isa<ConstantFP>(Val)) {
7146 // Use a splat constant if it is not safe to use undef.
7147 NewVal = getConstantVector(
7148 cast<Constant>(Val),
7149 isa<UndefValue>(Val) ||
7150 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
7151 } else
7152 llvm_unreachable("Did you modified shouldPromote and forgot to update "
7153 "this?");
7154 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
7155 }
7156 Transition->moveAfter(ToBePromoted);
7157 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
7158 }
7159
7160 /// Some targets can do store(extractelement) with one instruction.
7161 /// Try to push the extractelement towards the stores when the target
7162 /// has this feature and this is profitable.
optimizeExtractElementInst(Instruction * Inst)7163 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
7164 unsigned CombineCost = std::numeric_limits<unsigned>::max();
7165 if (DisableStoreExtract ||
7166 (!StressStoreExtract &&
7167 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
7168 Inst->getOperand(1), CombineCost)))
7169 return false;
7170
7171 // At this point we know that Inst is a vector to scalar transition.
7172 // Try to move it down the def-use chain, until:
7173 // - We can combine the transition with its single use
7174 // => we got rid of the transition.
7175 // - We escape the current basic block
7176 // => we would need to check that we are moving it at a cheaper place and
7177 // we do not do that for now.
7178 BasicBlock *Parent = Inst->getParent();
7179 LLVM_DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
7180 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
7181 // If the transition has more than one use, assume this is not going to be
7182 // beneficial.
7183 while (Inst->hasOneUse()) {
7184 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
7185 LLVM_DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
7186
7187 if (ToBePromoted->getParent() != Parent) {
7188 LLVM_DEBUG(dbgs() << "Instruction to promote is in a different block ("
7189 << ToBePromoted->getParent()->getName()
7190 << ") than the transition (" << Parent->getName()
7191 << ").\n");
7192 return false;
7193 }
7194
7195 if (VPH.canCombine(ToBePromoted)) {
7196 LLVM_DEBUG(dbgs() << "Assume " << *Inst << '\n'
7197 << "will be combined with: " << *ToBePromoted << '\n');
7198 VPH.recordCombineInstruction(ToBePromoted);
7199 bool Changed = VPH.promote();
7200 NumStoreExtractExposed += Changed;
7201 return Changed;
7202 }
7203
7204 LLVM_DEBUG(dbgs() << "Try promoting.\n");
7205 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
7206 return false;
7207
7208 LLVM_DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
7209
7210 VPH.enqueueForPromotion(ToBePromoted);
7211 Inst = ToBePromoted;
7212 }
7213 return false;
7214 }
7215
7216 /// For the instruction sequence of store below, F and I values
7217 /// are bundled together as an i64 value before being stored into memory.
7218 /// Sometimes it is more efficient to generate separate stores for F and I,
7219 /// which can remove the bitwise instructions or sink them to colder places.
7220 ///
7221 /// (store (or (zext (bitcast F to i32) to i64),
7222 /// (shl (zext I to i64), 32)), addr) -->
7223 /// (store F, addr) and (store I, addr+4)
7224 ///
7225 /// Similarly, splitting for other merged store can also be beneficial, like:
7226 /// For pair of {i32, i32}, i64 store --> two i32 stores.
7227 /// For pair of {i32, i16}, i64 store --> two i32 stores.
7228 /// For pair of {i16, i16}, i32 store --> two i16 stores.
7229 /// For pair of {i16, i8}, i32 store --> two i16 stores.
7230 /// For pair of {i8, i8}, i16 store --> two i8 stores.
7231 ///
7232 /// We allow each target to determine specifically which kind of splitting is
7233 /// supported.
7234 ///
7235 /// The store patterns are commonly seen from the simple code snippet below
7236 /// if only std::make_pair(...) is sroa transformed before inlined into hoo.
7237 /// void goo(const std::pair<int, float> &);
7238 /// hoo() {
7239 /// ...
7240 /// goo(std::make_pair(tmp, ftmp));
7241 /// ...
7242 /// }
7243 ///
7244 /// Although we already have similar splitting in DAG Combine, we duplicate
7245 /// it in CodeGenPrepare to catch the case in which pattern is across
7246 /// multiple BBs. The logic in DAG Combine is kept to catch case generated
7247 /// during code expansion.
splitMergedValStore(StoreInst & SI,const DataLayout & DL,const TargetLowering & TLI)7248 static bool splitMergedValStore(StoreInst &SI, const DataLayout &DL,
7249 const TargetLowering &TLI) {
7250 // Handle simple but common cases only.
7251 Type *StoreType = SI.getValueOperand()->getType();
7252
7253 // The code below assumes shifting a value by <number of bits>,
7254 // whereas scalable vectors would have to be shifted by
7255 // <2log(vscale) + number of bits> in order to store the
7256 // low/high parts. Bailing out for now.
7257 if (isa<ScalableVectorType>(StoreType))
7258 return false;
7259
7260 if (!DL.typeSizeEqualsStoreSize(StoreType) ||
7261 DL.getTypeSizeInBits(StoreType) == 0)
7262 return false;
7263
7264 unsigned HalfValBitSize = DL.getTypeSizeInBits(StoreType) / 2;
7265 Type *SplitStoreType = Type::getIntNTy(SI.getContext(), HalfValBitSize);
7266 if (!DL.typeSizeEqualsStoreSize(SplitStoreType))
7267 return false;
7268
7269 // Don't split the store if it is volatile.
7270 if (SI.isVolatile())
7271 return false;
7272
7273 // Match the following patterns:
7274 // (store (or (zext LValue to i64),
7275 // (shl (zext HValue to i64), 32)), HalfValBitSize)
7276 // or
7277 // (store (or (shl (zext HValue to i64), 32)), HalfValBitSize)
7278 // (zext LValue to i64),
7279 // Expect both operands of OR and the first operand of SHL have only
7280 // one use.
7281 Value *LValue, *HValue;
7282 if (!match(SI.getValueOperand(),
7283 m_c_Or(m_OneUse(m_ZExt(m_Value(LValue))),
7284 m_OneUse(m_Shl(m_OneUse(m_ZExt(m_Value(HValue))),
7285 m_SpecificInt(HalfValBitSize))))))
7286 return false;
7287
7288 // Check LValue and HValue are int with size less or equal than 32.
7289 if (!LValue->getType()->isIntegerTy() ||
7290 DL.getTypeSizeInBits(LValue->getType()) > HalfValBitSize ||
7291 !HValue->getType()->isIntegerTy() ||
7292 DL.getTypeSizeInBits(HValue->getType()) > HalfValBitSize)
7293 return false;
7294
7295 // If LValue/HValue is a bitcast instruction, use the EVT before bitcast
7296 // as the input of target query.
7297 auto *LBC = dyn_cast<BitCastInst>(LValue);
7298 auto *HBC = dyn_cast<BitCastInst>(HValue);
7299 EVT LowTy = LBC ? EVT::getEVT(LBC->getOperand(0)->getType())
7300 : EVT::getEVT(LValue->getType());
7301 EVT HighTy = HBC ? EVT::getEVT(HBC->getOperand(0)->getType())
7302 : EVT::getEVT(HValue->getType());
7303 if (!ForceSplitStore && !TLI.isMultiStoresCheaperThanBitsMerge(LowTy, HighTy))
7304 return false;
7305
7306 // Start to split store.
7307 IRBuilder<> Builder(SI.getContext());
7308 Builder.SetInsertPoint(&SI);
7309
7310 // If LValue/HValue is a bitcast in another BB, create a new one in current
7311 // BB so it may be merged with the splitted stores by dag combiner.
7312 if (LBC && LBC->getParent() != SI.getParent())
7313 LValue = Builder.CreateBitCast(LBC->getOperand(0), LBC->getType());
7314 if (HBC && HBC->getParent() != SI.getParent())
7315 HValue = Builder.CreateBitCast(HBC->getOperand(0), HBC->getType());
7316
7317 bool IsLE = SI.getModule()->getDataLayout().isLittleEndian();
7318 auto CreateSplitStore = [&](Value *V, bool Upper) {
7319 V = Builder.CreateZExtOrBitCast(V, SplitStoreType);
7320 Value *Addr = Builder.CreateBitCast(
7321 SI.getOperand(1),
7322 SplitStoreType->getPointerTo(SI.getPointerAddressSpace()));
7323 Align Alignment = SI.getAlign();
7324 const bool IsOffsetStore = (IsLE && Upper) || (!IsLE && !Upper);
7325 if (IsOffsetStore) {
7326 Addr = Builder.CreateGEP(
7327 SplitStoreType, Addr,
7328 ConstantInt::get(Type::getInt32Ty(SI.getContext()), 1));
7329
7330 // When splitting the store in half, naturally one half will retain the
7331 // alignment of the original wider store, regardless of whether it was
7332 // over-aligned or not, while the other will require adjustment.
7333 Alignment = commonAlignment(Alignment, HalfValBitSize / 8);
7334 }
7335 Builder.CreateAlignedStore(V, Addr, Alignment);
7336 };
7337
7338 CreateSplitStore(LValue, false);
7339 CreateSplitStore(HValue, true);
7340
7341 // Delete the old store.
7342 SI.eraseFromParent();
7343 return true;
7344 }
7345
7346 // Return true if the GEP has two operands, the first operand is of a sequential
7347 // type, and the second operand is a constant.
GEPSequentialConstIndexed(GetElementPtrInst * GEP)7348 static bool GEPSequentialConstIndexed(GetElementPtrInst *GEP) {
7349 gep_type_iterator I = gep_type_begin(*GEP);
7350 return GEP->getNumOperands() == 2 &&
7351 I.isSequential() &&
7352 isa<ConstantInt>(GEP->getOperand(1));
7353 }
7354
7355 // Try unmerging GEPs to reduce liveness interference (register pressure) across
7356 // IndirectBr edges. Since IndirectBr edges tend to touch on many blocks,
7357 // reducing liveness interference across those edges benefits global register
7358 // allocation. Currently handles only certain cases.
7359 //
7360 // For example, unmerge %GEPI and %UGEPI as below.
7361 //
7362 // ---------- BEFORE ----------
7363 // SrcBlock:
7364 // ...
7365 // %GEPIOp = ...
7366 // ...
7367 // %GEPI = gep %GEPIOp, Idx
7368 // ...
7369 // indirectbr ... [ label %DstB0, label %DstB1, ... label %DstBi ... ]
7370 // (* %GEPI is alive on the indirectbr edges due to other uses ahead)
7371 // (* %GEPIOp is alive on the indirectbr edges only because of it's used by
7372 // %UGEPI)
7373 //
7374 // DstB0: ... (there may be a gep similar to %UGEPI to be unmerged)
7375 // DstB1: ... (there may be a gep similar to %UGEPI to be unmerged)
7376 // ...
7377 //
7378 // DstBi:
7379 // ...
7380 // %UGEPI = gep %GEPIOp, UIdx
7381 // ...
7382 // ---------------------------
7383 //
7384 // ---------- AFTER ----------
7385 // SrcBlock:
7386 // ... (same as above)
7387 // (* %GEPI is still alive on the indirectbr edges)
7388 // (* %GEPIOp is no longer alive on the indirectbr edges as a result of the
7389 // unmerging)
7390 // ...
7391 //
7392 // DstBi:
7393 // ...
7394 // %UGEPI = gep %GEPI, (UIdx-Idx)
7395 // ...
7396 // ---------------------------
7397 //
7398 // The register pressure on the IndirectBr edges is reduced because %GEPIOp is
7399 // no longer alive on them.
7400 //
7401 // We try to unmerge GEPs here in CodGenPrepare, as opposed to limiting merging
7402 // of GEPs in the first place in InstCombiner::visitGetElementPtrInst() so as
7403 // not to disable further simplications and optimizations as a result of GEP
7404 // merging.
7405 //
7406 // Note this unmerging may increase the length of the data flow critical path
7407 // (the path from %GEPIOp to %UGEPI would go through %GEPI), which is a tradeoff
7408 // between the register pressure and the length of data-flow critical
7409 // path. Restricting this to the uncommon IndirectBr case would minimize the
7410 // impact of potentially longer critical path, if any, and the impact on compile
7411 // time.
tryUnmergingGEPsAcrossIndirectBr(GetElementPtrInst * GEPI,const TargetTransformInfo * TTI)7412 static bool tryUnmergingGEPsAcrossIndirectBr(GetElementPtrInst *GEPI,
7413 const TargetTransformInfo *TTI) {
7414 BasicBlock *SrcBlock = GEPI->getParent();
7415 // Check that SrcBlock ends with an IndirectBr. If not, give up. The common
7416 // (non-IndirectBr) cases exit early here.
7417 if (!isa<IndirectBrInst>(SrcBlock->getTerminator()))
7418 return false;
7419 // Check that GEPI is a simple gep with a single constant index.
7420 if (!GEPSequentialConstIndexed(GEPI))
7421 return false;
7422 ConstantInt *GEPIIdx = cast<ConstantInt>(GEPI->getOperand(1));
7423 // Check that GEPI is a cheap one.
7424 if (TTI->getIntImmCost(GEPIIdx->getValue(), GEPIIdx->getType(),
7425 TargetTransformInfo::TCK_SizeAndLatency)
7426 > TargetTransformInfo::TCC_Basic)
7427 return false;
7428 Value *GEPIOp = GEPI->getOperand(0);
7429 // Check that GEPIOp is an instruction that's also defined in SrcBlock.
7430 if (!isa<Instruction>(GEPIOp))
7431 return false;
7432 auto *GEPIOpI = cast<Instruction>(GEPIOp);
7433 if (GEPIOpI->getParent() != SrcBlock)
7434 return false;
7435 // Check that GEP is used outside the block, meaning it's alive on the
7436 // IndirectBr edge(s).
7437 if (find_if(GEPI->users(), [&](User *Usr) {
7438 if (auto *I = dyn_cast<Instruction>(Usr)) {
7439 if (I->getParent() != SrcBlock) {
7440 return true;
7441 }
7442 }
7443 return false;
7444 }) == GEPI->users().end())
7445 return false;
7446 // The second elements of the GEP chains to be unmerged.
7447 std::vector<GetElementPtrInst *> UGEPIs;
7448 // Check each user of GEPIOp to check if unmerging would make GEPIOp not alive
7449 // on IndirectBr edges.
7450 for (User *Usr : GEPIOp->users()) {
7451 if (Usr == GEPI) continue;
7452 // Check if Usr is an Instruction. If not, give up.
7453 if (!isa<Instruction>(Usr))
7454 return false;
7455 auto *UI = cast<Instruction>(Usr);
7456 // Check if Usr in the same block as GEPIOp, which is fine, skip.
7457 if (UI->getParent() == SrcBlock)
7458 continue;
7459 // Check if Usr is a GEP. If not, give up.
7460 if (!isa<GetElementPtrInst>(Usr))
7461 return false;
7462 auto *UGEPI = cast<GetElementPtrInst>(Usr);
7463 // Check if UGEPI is a simple gep with a single constant index and GEPIOp is
7464 // the pointer operand to it. If so, record it in the vector. If not, give
7465 // up.
7466 if (!GEPSequentialConstIndexed(UGEPI))
7467 return false;
7468 if (UGEPI->getOperand(0) != GEPIOp)
7469 return false;
7470 if (GEPIIdx->getType() !=
7471 cast<ConstantInt>(UGEPI->getOperand(1))->getType())
7472 return false;
7473 ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
7474 if (TTI->getIntImmCost(UGEPIIdx->getValue(), UGEPIIdx->getType(),
7475 TargetTransformInfo::TCK_SizeAndLatency)
7476 > TargetTransformInfo::TCC_Basic)
7477 return false;
7478 UGEPIs.push_back(UGEPI);
7479 }
7480 if (UGEPIs.size() == 0)
7481 return false;
7482 // Check the materializing cost of (Uidx-Idx).
7483 for (GetElementPtrInst *UGEPI : UGEPIs) {
7484 ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
7485 APInt NewIdx = UGEPIIdx->getValue() - GEPIIdx->getValue();
7486 unsigned ImmCost =
7487 TTI->getIntImmCost(NewIdx, GEPIIdx->getType(),
7488 TargetTransformInfo::TCK_SizeAndLatency);
7489 if (ImmCost > TargetTransformInfo::TCC_Basic)
7490 return false;
7491 }
7492 // Now unmerge between GEPI and UGEPIs.
7493 for (GetElementPtrInst *UGEPI : UGEPIs) {
7494 UGEPI->setOperand(0, GEPI);
7495 ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
7496 Constant *NewUGEPIIdx =
7497 ConstantInt::get(GEPIIdx->getType(),
7498 UGEPIIdx->getValue() - GEPIIdx->getValue());
7499 UGEPI->setOperand(1, NewUGEPIIdx);
7500 // If GEPI is not inbounds but UGEPI is inbounds, change UGEPI to not
7501 // inbounds to avoid UB.
7502 if (!GEPI->isInBounds()) {
7503 UGEPI->setIsInBounds(false);
7504 }
7505 }
7506 // After unmerging, verify that GEPIOp is actually only used in SrcBlock (not
7507 // alive on IndirectBr edges).
7508 assert(find_if(GEPIOp->users(), [&](User *Usr) {
7509 return cast<Instruction>(Usr)->getParent() != SrcBlock;
7510 }) == GEPIOp->users().end() && "GEPIOp is used outside SrcBlock");
7511 return true;
7512 }
7513
optimizeInst(Instruction * I,bool & ModifiedDT)7514 bool CodeGenPrepare::optimizeInst(Instruction *I, bool &ModifiedDT) {
7515 // Bail out if we inserted the instruction to prevent optimizations from
7516 // stepping on each other's toes.
7517 if (InsertedInsts.count(I))
7518 return false;
7519
7520 // TODO: Move into the switch on opcode below here.
7521 if (PHINode *P = dyn_cast<PHINode>(I)) {
7522 // It is possible for very late stage optimizations (such as SimplifyCFG)
7523 // to introduce PHI nodes too late to be cleaned up. If we detect such a
7524 // trivial PHI, go ahead and zap it here.
7525 if (Value *V = SimplifyInstruction(P, {*DL, TLInfo})) {
7526 LargeOffsetGEPMap.erase(P);
7527 P->replaceAllUsesWith(V);
7528 P->eraseFromParent();
7529 ++NumPHIsElim;
7530 return true;
7531 }
7532 return false;
7533 }
7534
7535 if (CastInst *CI = dyn_cast<CastInst>(I)) {
7536 // If the source of the cast is a constant, then this should have
7537 // already been constant folded. The only reason NOT to constant fold
7538 // it is if something (e.g. LSR) was careful to place the constant
7539 // evaluation in a block other than then one that uses it (e.g. to hoist
7540 // the address of globals out of a loop). If this is the case, we don't
7541 // want to forward-subst the cast.
7542 if (isa<Constant>(CI->getOperand(0)))
7543 return false;
7544
7545 if (OptimizeNoopCopyExpression(CI, *TLI, *DL))
7546 return true;
7547
7548 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7549 /// Sink a zext or sext into its user blocks if the target type doesn't
7550 /// fit in one register
7551 if (TLI->getTypeAction(CI->getContext(),
7552 TLI->getValueType(*DL, CI->getType())) ==
7553 TargetLowering::TypeExpandInteger) {
7554 return SinkCast(CI);
7555 } else {
7556 bool MadeChange = optimizeExt(I);
7557 return MadeChange | optimizeExtUses(I);
7558 }
7559 }
7560 return false;
7561 }
7562
7563 if (auto *Cmp = dyn_cast<CmpInst>(I))
7564 if (optimizeCmp(Cmp, ModifiedDT))
7565 return true;
7566
7567 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7568 LI->setMetadata(LLVMContext::MD_invariant_group, nullptr);
7569 bool Modified = optimizeLoadExt(LI);
7570 unsigned AS = LI->getPointerAddressSpace();
7571 Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
7572 return Modified;
7573 }
7574
7575 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
7576 if (splitMergedValStore(*SI, *DL, *TLI))
7577 return true;
7578 SI->setMetadata(LLVMContext::MD_invariant_group, nullptr);
7579 unsigned AS = SI->getPointerAddressSpace();
7580 return optimizeMemoryInst(I, SI->getOperand(1),
7581 SI->getOperand(0)->getType(), AS);
7582 }
7583
7584 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
7585 unsigned AS = RMW->getPointerAddressSpace();
7586 return optimizeMemoryInst(I, RMW->getPointerOperand(),
7587 RMW->getType(), AS);
7588 }
7589
7590 if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(I)) {
7591 unsigned AS = CmpX->getPointerAddressSpace();
7592 return optimizeMemoryInst(I, CmpX->getPointerOperand(),
7593 CmpX->getCompareOperand()->getType(), AS);
7594 }
7595
7596 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
7597
7598 if (BinOp && (BinOp->getOpcode() == Instruction::And) && EnableAndCmpSinking)
7599 return sinkAndCmp0Expression(BinOp, *TLI, InsertedInsts);
7600
7601 // TODO: Move this into the switch on opcode - it handles shifts already.
7602 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
7603 BinOp->getOpcode() == Instruction::LShr)) {
7604 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
7605 if (CI && TLI->hasExtractBitsInsn())
7606 if (OptimizeExtractBits(BinOp, CI, *TLI, *DL))
7607 return true;
7608 }
7609
7610 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
7611 if (GEPI->hasAllZeroIndices()) {
7612 /// The GEP operand must be a pointer, so must its result -> BitCast
7613 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
7614 GEPI->getName(), GEPI);
7615 NC->setDebugLoc(GEPI->getDebugLoc());
7616 GEPI->replaceAllUsesWith(NC);
7617 GEPI->eraseFromParent();
7618 ++NumGEPsElim;
7619 optimizeInst(NC, ModifiedDT);
7620 return true;
7621 }
7622 if (tryUnmergingGEPsAcrossIndirectBr(GEPI, TTI)) {
7623 return true;
7624 }
7625 return false;
7626 }
7627
7628 if (FreezeInst *FI = dyn_cast<FreezeInst>(I)) {
7629 // freeze(icmp a, const)) -> icmp (freeze a), const
7630 // This helps generate efficient conditional jumps.
7631 Instruction *CmpI = nullptr;
7632 if (ICmpInst *II = dyn_cast<ICmpInst>(FI->getOperand(0)))
7633 CmpI = II;
7634 else if (FCmpInst *F = dyn_cast<FCmpInst>(FI->getOperand(0)))
7635 CmpI = F->getFastMathFlags().none() ? F : nullptr;
7636
7637 if (CmpI && CmpI->hasOneUse()) {
7638 auto Op0 = CmpI->getOperand(0), Op1 = CmpI->getOperand(1);
7639 bool Const0 = isa<ConstantInt>(Op0) || isa<ConstantFP>(Op0) ||
7640 isa<ConstantPointerNull>(Op0);
7641 bool Const1 = isa<ConstantInt>(Op1) || isa<ConstantFP>(Op1) ||
7642 isa<ConstantPointerNull>(Op1);
7643 if (Const0 || Const1) {
7644 if (!Const0 || !Const1) {
7645 auto *F = new FreezeInst(Const0 ? Op1 : Op0, "", CmpI);
7646 F->takeName(FI);
7647 CmpI->setOperand(Const0 ? 1 : 0, F);
7648 }
7649 FI->replaceAllUsesWith(CmpI);
7650 FI->eraseFromParent();
7651 return true;
7652 }
7653 }
7654 return false;
7655 }
7656
7657 if (tryToSinkFreeOperands(I))
7658 return true;
7659
7660 switch (I->getOpcode()) {
7661 case Instruction::Shl:
7662 case Instruction::LShr:
7663 case Instruction::AShr:
7664 return optimizeShiftInst(cast<BinaryOperator>(I));
7665 case Instruction::Call:
7666 return optimizeCallInst(cast<CallInst>(I), ModifiedDT);
7667 case Instruction::Select:
7668 return optimizeSelectInst(cast<SelectInst>(I));
7669 case Instruction::ShuffleVector:
7670 return optimizeShuffleVectorInst(cast<ShuffleVectorInst>(I));
7671 case Instruction::Switch:
7672 return optimizeSwitchInst(cast<SwitchInst>(I));
7673 case Instruction::ExtractElement:
7674 return optimizeExtractElementInst(cast<ExtractElementInst>(I));
7675 }
7676
7677 return false;
7678 }
7679
7680 /// Given an OR instruction, check to see if this is a bitreverse
7681 /// idiom. If so, insert the new intrinsic and return true.
makeBitReverse(Instruction & I)7682 bool CodeGenPrepare::makeBitReverse(Instruction &I) {
7683 if (!I.getType()->isIntegerTy() ||
7684 !TLI->isOperationLegalOrCustom(ISD::BITREVERSE,
7685 TLI->getValueType(*DL, I.getType(), true)))
7686 return false;
7687
7688 SmallVector<Instruction*, 4> Insts;
7689 if (!recognizeBSwapOrBitReverseIdiom(&I, false, true, Insts))
7690 return false;
7691 Instruction *LastInst = Insts.back();
7692 I.replaceAllUsesWith(LastInst);
7693 RecursivelyDeleteTriviallyDeadInstructions(
7694 &I, TLInfo, nullptr, [&](Value *V) { removeAllAssertingVHReferences(V); });
7695 return true;
7696 }
7697
7698 // In this pass we look for GEP and cast instructions that are used
7699 // across basic blocks and rewrite them to improve basic-block-at-a-time
7700 // selection.
optimizeBlock(BasicBlock & BB,bool & ModifiedDT)7701 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool &ModifiedDT) {
7702 SunkAddrs.clear();
7703 bool MadeChange = false;
7704
7705 CurInstIterator = BB.begin();
7706 while (CurInstIterator != BB.end()) {
7707 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
7708 if (ModifiedDT)
7709 return true;
7710 }
7711
7712 bool MadeBitReverse = true;
7713 while (MadeBitReverse) {
7714 MadeBitReverse = false;
7715 for (auto &I : reverse(BB)) {
7716 if (makeBitReverse(I)) {
7717 MadeBitReverse = MadeChange = true;
7718 break;
7719 }
7720 }
7721 }
7722 MadeChange |= dupRetToEnableTailCallOpts(&BB, ModifiedDT);
7723
7724 return MadeChange;
7725 }
7726
7727 // Some CGP optimizations may move or alter what's computed in a block. Check
7728 // whether a dbg.value intrinsic could be pointed at a more appropriate operand.
fixupDbgValue(Instruction * I)7729 bool CodeGenPrepare::fixupDbgValue(Instruction *I) {
7730 assert(isa<DbgValueInst>(I));
7731 DbgValueInst &DVI = *cast<DbgValueInst>(I);
7732
7733 // Does this dbg.value refer to a sunk address calculation?
7734 Value *Location = DVI.getVariableLocation();
7735 WeakTrackingVH SunkAddrVH = SunkAddrs[Location];
7736 Value *SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
7737 if (SunkAddr) {
7738 // Point dbg.value at locally computed address, which should give the best
7739 // opportunity to be accurately lowered. This update may change the type of
7740 // pointer being referred to; however this makes no difference to debugging
7741 // information, and we can't generate bitcasts that may affect codegen.
7742 DVI.setOperand(0, MetadataAsValue::get(DVI.getContext(),
7743 ValueAsMetadata::get(SunkAddr)));
7744 return true;
7745 }
7746 return false;
7747 }
7748
7749 // A llvm.dbg.value may be using a value before its definition, due to
7750 // optimizations in this pass and others. Scan for such dbg.values, and rescue
7751 // them by moving the dbg.value to immediately after the value definition.
7752 // FIXME: Ideally this should never be necessary, and this has the potential
7753 // to re-order dbg.value intrinsics.
placeDbgValues(Function & F)7754 bool CodeGenPrepare::placeDbgValues(Function &F) {
7755 bool MadeChange = false;
7756 DominatorTree DT(F);
7757
7758 for (BasicBlock &BB : F) {
7759 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
7760 Instruction *Insn = &*BI++;
7761 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
7762 if (!DVI)
7763 continue;
7764
7765 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
7766
7767 if (!VI || VI->isTerminator())
7768 continue;
7769
7770 // If VI is a phi in a block with an EHPad terminator, we can't insert
7771 // after it.
7772 if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
7773 continue;
7774
7775 // If the defining instruction dominates the dbg.value, we do not need
7776 // to move the dbg.value.
7777 if (DT.dominates(VI, DVI))
7778 continue;
7779
7780 LLVM_DEBUG(dbgs() << "Moving Debug Value before :\n"
7781 << *DVI << ' ' << *VI);
7782 DVI->removeFromParent();
7783 if (isa<PHINode>(VI))
7784 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
7785 else
7786 DVI->insertAfter(VI);
7787 MadeChange = true;
7788 ++NumDbgValueMoved;
7789 }
7790 }
7791 return MadeChange;
7792 }
7793
7794 /// Scale down both weights to fit into uint32_t.
scaleWeights(uint64_t & NewTrue,uint64_t & NewFalse)7795 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
7796 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
7797 uint32_t Scale = (NewMax / std::numeric_limits<uint32_t>::max()) + 1;
7798 NewTrue = NewTrue / Scale;
7799 NewFalse = NewFalse / Scale;
7800 }
7801
7802 /// Some targets prefer to split a conditional branch like:
7803 /// \code
7804 /// %0 = icmp ne i32 %a, 0
7805 /// %1 = icmp ne i32 %b, 0
7806 /// %or.cond = or i1 %0, %1
7807 /// br i1 %or.cond, label %TrueBB, label %FalseBB
7808 /// \endcode
7809 /// into multiple branch instructions like:
7810 /// \code
7811 /// bb1:
7812 /// %0 = icmp ne i32 %a, 0
7813 /// br i1 %0, label %TrueBB, label %bb2
7814 /// bb2:
7815 /// %1 = icmp ne i32 %b, 0
7816 /// br i1 %1, label %TrueBB, label %FalseBB
7817 /// \endcode
7818 /// This usually allows instruction selection to do even further optimizations
7819 /// and combine the compare with the branch instruction. Currently this is
7820 /// applied for targets which have "cheap" jump instructions.
7821 ///
7822 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
7823 ///
splitBranchCondition(Function & F,bool & ModifiedDT)7824 bool CodeGenPrepare::splitBranchCondition(Function &F, bool &ModifiedDT) {
7825 if (!TM->Options.EnableFastISel || TLI->isJumpExpensive())
7826 return false;
7827
7828 bool MadeChange = false;
7829 for (auto &BB : F) {
7830 // Does this BB end with the following?
7831 // %cond1 = icmp|fcmp|binary instruction ...
7832 // %cond2 = icmp|fcmp|binary instruction ...
7833 // %cond.or = or|and i1 %cond1, cond2
7834 // br i1 %cond.or label %dest1, label %dest2"
7835 Instruction *LogicOp;
7836 BasicBlock *TBB, *FBB;
7837 if (!match(BB.getTerminator(),
7838 m_Br(m_OneUse(m_Instruction(LogicOp)), TBB, FBB)))
7839 continue;
7840
7841 auto *Br1 = cast<BranchInst>(BB.getTerminator());
7842 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
7843 continue;
7844
7845 // The merging of mostly empty BB can cause a degenerate branch.
7846 if (TBB == FBB)
7847 continue;
7848
7849 unsigned Opc;
7850 Value *Cond1, *Cond2;
7851 if (match(LogicOp,
7852 m_LogicalAnd(m_OneUse(m_Value(Cond1)), m_OneUse(m_Value(Cond2)))))
7853 Opc = Instruction::And;
7854 else if (match(LogicOp, m_LogicalOr(m_OneUse(m_Value(Cond1)),
7855 m_OneUse(m_Value(Cond2)))))
7856 Opc = Instruction::Or;
7857 else
7858 continue;
7859
7860 auto IsGoodCond = [](Value *Cond) {
7861 return match(
7862 Cond,
7863 m_CombineOr(m_Cmp(), m_CombineOr(m_LogicalAnd(m_Value(), m_Value()),
7864 m_LogicalOr(m_Value(), m_Value()))));
7865 };
7866 if (!IsGoodCond(Cond1) || !IsGoodCond(Cond2))
7867 continue;
7868
7869 LLVM_DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
7870
7871 // Create a new BB.
7872 auto *TmpBB =
7873 BasicBlock::Create(BB.getContext(), BB.getName() + ".cond.split",
7874 BB.getParent(), BB.getNextNode());
7875
7876 // Update original basic block by using the first condition directly by the
7877 // branch instruction and removing the no longer needed and/or instruction.
7878 Br1->setCondition(Cond1);
7879 LogicOp->eraseFromParent();
7880
7881 // Depending on the condition we have to either replace the true or the
7882 // false successor of the original branch instruction.
7883 if (Opc == Instruction::And)
7884 Br1->setSuccessor(0, TmpBB);
7885 else
7886 Br1->setSuccessor(1, TmpBB);
7887
7888 // Fill in the new basic block.
7889 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
7890 if (auto *I = dyn_cast<Instruction>(Cond2)) {
7891 I->removeFromParent();
7892 I->insertBefore(Br2);
7893 }
7894
7895 // Update PHI nodes in both successors. The original BB needs to be
7896 // replaced in one successor's PHI nodes, because the branch comes now from
7897 // the newly generated BB (NewBB). In the other successor we need to add one
7898 // incoming edge to the PHI nodes, because both branch instructions target
7899 // now the same successor. Depending on the original branch condition
7900 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
7901 // we perform the correct update for the PHI nodes.
7902 // This doesn't change the successor order of the just created branch
7903 // instruction (or any other instruction).
7904 if (Opc == Instruction::Or)
7905 std::swap(TBB, FBB);
7906
7907 // Replace the old BB with the new BB.
7908 TBB->replacePhiUsesWith(&BB, TmpBB);
7909
7910 // Add another incoming edge form the new BB.
7911 for (PHINode &PN : FBB->phis()) {
7912 auto *Val = PN.getIncomingValueForBlock(&BB);
7913 PN.addIncoming(Val, TmpBB);
7914 }
7915
7916 // Update the branch weights (from SelectionDAGBuilder::
7917 // FindMergedConditions).
7918 if (Opc == Instruction::Or) {
7919 // Codegen X | Y as:
7920 // BB1:
7921 // jmp_if_X TBB
7922 // jmp TmpBB
7923 // TmpBB:
7924 // jmp_if_Y TBB
7925 // jmp FBB
7926 //
7927
7928 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
7929 // The requirement is that
7930 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
7931 // = TrueProb for original BB.
7932 // Assuming the original weights are A and B, one choice is to set BB1's
7933 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
7934 // assumes that
7935 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
7936 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
7937 // TmpBB, but the math is more complicated.
7938 uint64_t TrueWeight, FalseWeight;
7939 if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) {
7940 uint64_t NewTrueWeight = TrueWeight;
7941 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
7942 scaleWeights(NewTrueWeight, NewFalseWeight);
7943 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
7944 .createBranchWeights(TrueWeight, FalseWeight));
7945
7946 NewTrueWeight = TrueWeight;
7947 NewFalseWeight = 2 * FalseWeight;
7948 scaleWeights(NewTrueWeight, NewFalseWeight);
7949 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
7950 .createBranchWeights(TrueWeight, FalseWeight));
7951 }
7952 } else {
7953 // Codegen X & Y as:
7954 // BB1:
7955 // jmp_if_X TmpBB
7956 // jmp FBB
7957 // TmpBB:
7958 // jmp_if_Y TBB
7959 // jmp FBB
7960 //
7961 // This requires creation of TmpBB after CurBB.
7962
7963 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
7964 // The requirement is that
7965 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
7966 // = FalseProb for original BB.
7967 // Assuming the original weights are A and B, one choice is to set BB1's
7968 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
7969 // assumes that
7970 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
7971 uint64_t TrueWeight, FalseWeight;
7972 if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) {
7973 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
7974 uint64_t NewFalseWeight = FalseWeight;
7975 scaleWeights(NewTrueWeight, NewFalseWeight);
7976 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
7977 .createBranchWeights(TrueWeight, FalseWeight));
7978
7979 NewTrueWeight = 2 * TrueWeight;
7980 NewFalseWeight = FalseWeight;
7981 scaleWeights(NewTrueWeight, NewFalseWeight);
7982 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
7983 .createBranchWeights(TrueWeight, FalseWeight));
7984 }
7985 }
7986
7987 ModifiedDT = true;
7988 MadeChange = true;
7989
7990 LLVM_DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
7991 TmpBB->dump());
7992 }
7993 return MadeChange;
7994 }
7995