File: | lib/CodeGen/CodeGenPrepare.cpp |
Warning: | line 2236, column 18 Called C++ object pointer is null |
1 | //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===// | |||
2 | // | |||
3 | // The LLVM Compiler Infrastructure | |||
4 | // | |||
5 | // This file is distributed under the University of Illinois Open Source | |||
6 | // License. See LICENSE.TXT for details. | |||
7 | // | |||
8 | //===----------------------------------------------------------------------===// | |||
9 | // | |||
10 | // This pass munges the code in the input function to better prepare it for | |||
11 | // SelectionDAG-based code generation. This works around limitations in it's | |||
12 | // basic-block-at-a-time approach. It should eventually be removed. | |||
13 | // | |||
14 | //===----------------------------------------------------------------------===// | |||
15 | ||||
16 | #include "llvm/ADT/DenseMap.h" | |||
17 | #include "llvm/ADT/SetVector.h" | |||
18 | #include "llvm/ADT/SmallSet.h" | |||
19 | #include "llvm/ADT/Statistic.h" | |||
20 | #include "llvm/Analysis/BlockFrequencyInfo.h" | |||
21 | #include "llvm/Analysis/BranchProbabilityInfo.h" | |||
22 | #include "llvm/Analysis/CFG.h" | |||
23 | #include "llvm/Analysis/InstructionSimplify.h" | |||
24 | #include "llvm/Analysis/LoopInfo.h" | |||
25 | #include "llvm/Analysis/MemoryBuiltins.h" | |||
26 | #include "llvm/Analysis/ProfileSummaryInfo.h" | |||
27 | #include "llvm/Analysis/TargetLibraryInfo.h" | |||
28 | #include "llvm/Analysis/TargetTransformInfo.h" | |||
29 | #include "llvm/Analysis/ValueTracking.h" | |||
30 | #include "llvm/CodeGen/Analysis.h" | |||
31 | #include "llvm/CodeGen/Passes.h" | |||
32 | #include "llvm/CodeGen/TargetPassConfig.h" | |||
33 | #include "llvm/IR/CallSite.h" | |||
34 | #include "llvm/IR/Constants.h" | |||
35 | #include "llvm/IR/DataLayout.h" | |||
36 | #include "llvm/IR/DerivedTypes.h" | |||
37 | #include "llvm/IR/Dominators.h" | |||
38 | #include "llvm/IR/Function.h" | |||
39 | #include "llvm/IR/GetElementPtrTypeIterator.h" | |||
40 | #include "llvm/IR/IRBuilder.h" | |||
41 | #include "llvm/IR/InlineAsm.h" | |||
42 | #include "llvm/IR/Instructions.h" | |||
43 | #include "llvm/IR/IntrinsicInst.h" | |||
44 | #include "llvm/IR/MDBuilder.h" | |||
45 | #include "llvm/IR/PatternMatch.h" | |||
46 | #include "llvm/IR/Statepoint.h" | |||
47 | #include "llvm/IR/ValueHandle.h" | |||
48 | #include "llvm/IR/ValueMap.h" | |||
49 | #include "llvm/Pass.h" | |||
50 | #include "llvm/Support/BranchProbability.h" | |||
51 | #include "llvm/Support/CommandLine.h" | |||
52 | #include "llvm/Support/Debug.h" | |||
53 | #include "llvm/Support/raw_ostream.h" | |||
54 | #include "llvm/Target/TargetLowering.h" | |||
55 | #include "llvm/Target/TargetSubtargetInfo.h" | |||
56 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | |||
57 | #include "llvm/Transforms/Utils/BuildLibCalls.h" | |||
58 | #include "llvm/Transforms/Utils/BypassSlowDivision.h" | |||
59 | #include "llvm/Transforms/Utils/Cloning.h" | |||
60 | #include "llvm/Transforms/Utils/Local.h" | |||
61 | #include "llvm/Transforms/Utils/SimplifyLibCalls.h" | |||
62 | #include "llvm/Transforms/Utils/ValueMapper.h" | |||
63 | ||||
64 | using namespace llvm; | |||
65 | using namespace llvm::PatternMatch; | |||
66 | ||||
67 | #define DEBUG_TYPE"codegenprepare" "codegenprepare" | |||
68 | ||||
69 | STATISTIC(NumBlocksElim, "Number of blocks eliminated")static llvm::Statistic NumBlocksElim = {"codegenprepare", "NumBlocksElim" , "Number of blocks eliminated", {0}, false}; | |||
70 | STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated")static llvm::Statistic NumPHIsElim = {"codegenprepare", "NumPHIsElim" , "Number of trivial PHIs eliminated", {0}, false}; | |||
71 | STATISTIC(NumGEPsElim, "Number of GEPs converted to casts")static llvm::Statistic NumGEPsElim = {"codegenprepare", "NumGEPsElim" , "Number of GEPs converted to casts", {0}, false}; | |||
72 | STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "static llvm::Statistic NumCmpUses = {"codegenprepare", "NumCmpUses" , "Number of uses of Cmp expressions replaced with uses of " "sunken Cmps" , {0}, false} | |||
73 | "sunken Cmps")static llvm::Statistic NumCmpUses = {"codegenprepare", "NumCmpUses" , "Number of uses of Cmp expressions replaced with uses of " "sunken Cmps" , {0}, false}; | |||
74 | STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "static llvm::Statistic NumCastUses = {"codegenprepare", "NumCastUses" , "Number of uses of Cast expressions replaced with uses " "of sunken Casts" , {0}, false} | |||
75 | "of sunken Casts")static llvm::Statistic NumCastUses = {"codegenprepare", "NumCastUses" , "Number of uses of Cast expressions replaced with uses " "of sunken Casts" , {0}, false}; | |||
76 | STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "static llvm::Statistic NumMemoryInsts = {"codegenprepare", "NumMemoryInsts" , "Number of memory instructions whose address " "computations were sunk" , {0}, false} | |||
77 | "computations were sunk")static llvm::Statistic NumMemoryInsts = {"codegenprepare", "NumMemoryInsts" , "Number of memory instructions whose address " "computations were sunk" , {0}, false}; | |||
78 | STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads")static llvm::Statistic NumExtsMoved = {"codegenprepare", "NumExtsMoved" , "Number of [s|z]ext instructions combined with loads", {0}, false}; | |||
79 | STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized")static llvm::Statistic NumExtUses = {"codegenprepare", "NumExtUses" , "Number of uses of [s|z]ext instructions optimized", {0}, false }; | |||
80 | STATISTIC(NumAndsAdded,static llvm::Statistic NumAndsAdded = {"codegenprepare", "NumAndsAdded" , "Number of and mask instructions added to form ext loads", { 0}, false} | |||
81 | "Number of and mask instructions added to form ext loads")static llvm::Statistic NumAndsAdded = {"codegenprepare", "NumAndsAdded" , "Number of and mask instructions added to form ext loads", { 0}, false}; | |||
82 | STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized")static llvm::Statistic NumAndUses = {"codegenprepare", "NumAndUses" , "Number of uses of and mask instructions optimized", {0}, false }; | |||
83 | STATISTIC(NumRetsDup, "Number of return instructions duplicated")static llvm::Statistic NumRetsDup = {"codegenprepare", "NumRetsDup" , "Number of return instructions duplicated", {0}, false}; | |||
84 | STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved")static llvm::Statistic NumDbgValueMoved = {"codegenprepare", "NumDbgValueMoved" , "Number of debug value instructions moved", {0}, false}; | |||
85 | STATISTIC(NumSelectsExpanded, "Number of selects turned into branches")static llvm::Statistic NumSelectsExpanded = {"codegenprepare" , "NumSelectsExpanded", "Number of selects turned into branches" , {0}, false}; | |||
86 | STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed")static llvm::Statistic NumStoreExtractExposed = {"codegenprepare" , "NumStoreExtractExposed", "Number of store(extractelement) exposed" , {0}, false}; | |||
87 | ||||
88 | STATISTIC(NumMemCmpCalls, "Number of memcmp calls")static llvm::Statistic NumMemCmpCalls = {"codegenprepare", "NumMemCmpCalls" , "Number of memcmp calls", {0}, false}; | |||
89 | STATISTIC(NumMemCmpNotConstant, "Number of memcmp calls without constant size")static llvm::Statistic NumMemCmpNotConstant = {"codegenprepare" , "NumMemCmpNotConstant", "Number of memcmp calls without constant size" , {0}, false}; | |||
90 | STATISTIC(NumMemCmpGreaterThanMax,static llvm::Statistic NumMemCmpGreaterThanMax = {"codegenprepare" , "NumMemCmpGreaterThanMax", "Number of memcmp calls with size greater than max size" , {0}, false} | |||
91 | "Number of memcmp calls with size greater than max size")static llvm::Statistic NumMemCmpGreaterThanMax = {"codegenprepare" , "NumMemCmpGreaterThanMax", "Number of memcmp calls with size greater than max size" , {0}, false}; | |||
92 | STATISTIC(NumMemCmpInlined, "Number of inlined memcmp calls")static llvm::Statistic NumMemCmpInlined = {"codegenprepare", "NumMemCmpInlined" , "Number of inlined memcmp calls", {0}, false}; | |||
93 | ||||
94 | static cl::opt<bool> DisableBranchOpts( | |||
95 | "disable-cgp-branch-opts", cl::Hidden, cl::init(false), | |||
96 | cl::desc("Disable branch optimizations in CodeGenPrepare")); | |||
97 | ||||
98 | static cl::opt<bool> | |||
99 | DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false), | |||
100 | cl::desc("Disable GC optimizations in CodeGenPrepare")); | |||
101 | ||||
102 | static cl::opt<bool> DisableSelectToBranch( | |||
103 | "disable-cgp-select2branch", cl::Hidden, cl::init(false), | |||
104 | cl::desc("Disable select to branch conversion.")); | |||
105 | ||||
106 | static cl::opt<bool> AddrSinkUsingGEPs( | |||
107 | "addr-sink-using-gep", cl::Hidden, cl::init(true), | |||
108 | cl::desc("Address sinking in CGP using GEPs.")); | |||
109 | ||||
110 | static cl::opt<bool> EnableAndCmpSinking( | |||
111 | "enable-andcmp-sinking", cl::Hidden, cl::init(true), | |||
112 | cl::desc("Enable sinkinig and/cmp into branches.")); | |||
113 | ||||
114 | static cl::opt<bool> DisableStoreExtract( | |||
115 | "disable-cgp-store-extract", cl::Hidden, cl::init(false), | |||
116 | cl::desc("Disable store(extract) optimizations in CodeGenPrepare")); | |||
117 | ||||
118 | static cl::opt<bool> StressStoreExtract( | |||
119 | "stress-cgp-store-extract", cl::Hidden, cl::init(false), | |||
120 | cl::desc("Stress test store(extract) optimizations in CodeGenPrepare")); | |||
121 | ||||
122 | static cl::opt<bool> DisableExtLdPromotion( | |||
123 | "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false), | |||
124 | cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in " | |||
125 | "CodeGenPrepare")); | |||
126 | ||||
127 | static cl::opt<bool> StressExtLdPromotion( | |||
128 | "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false), | |||
129 | cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) " | |||
130 | "optimization in CodeGenPrepare")); | |||
131 | ||||
132 | static cl::opt<bool> DisablePreheaderProtect( | |||
133 | "disable-preheader-prot", cl::Hidden, cl::init(false), | |||
134 | cl::desc("Disable protection against removing loop preheaders")); | |||
135 | ||||
136 | static cl::opt<bool> ProfileGuidedSectionPrefix( | |||
137 | "profile-guided-section-prefix", cl::Hidden, cl::init(true), cl::ZeroOrMore, | |||
138 | cl::desc("Use profile info to add section prefix for hot/cold functions")); | |||
139 | ||||
140 | static cl::opt<unsigned> FreqRatioToSkipMerge( | |||
141 | "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2), | |||
142 | cl::desc("Skip merging empty blocks if (frequency of empty block) / " | |||
143 | "(frequency of destination block) is greater than this ratio")); | |||
144 | ||||
145 | static cl::opt<bool> ForceSplitStore( | |||
146 | "force-split-store", cl::Hidden, cl::init(false), | |||
147 | cl::desc("Force store splitting no matter what the target query says.")); | |||
148 | ||||
149 | static cl::opt<bool> | |||
150 | EnableTypePromotionMerge("cgp-type-promotion-merge", cl::Hidden, | |||
151 | cl::desc("Enable merging of redundant sexts when one is dominating" | |||
152 | " the other."), cl::init(true)); | |||
153 | ||||
154 | static cl::opt<unsigned> MemCmpNumLoadsPerBlock( | |||
155 | "memcmp-num-loads-per-block", cl::Hidden, cl::init(1), | |||
156 | cl::desc("The number of loads per basic block for inline expansion of " | |||
157 | "memcmp that is only being compared against zero.")); | |||
158 | ||||
159 | namespace { | |||
160 | typedef SmallPtrSet<Instruction *, 16> SetOfInstrs; | |||
161 | typedef PointerIntPair<Type *, 1, bool> TypeIsSExt; | |||
162 | typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy; | |||
163 | typedef SmallVector<Instruction *, 16> SExts; | |||
164 | typedef DenseMap<Value *, SExts> ValueToSExts; | |||
165 | class TypePromotionTransaction; | |||
166 | ||||
167 | class CodeGenPrepare : public FunctionPass { | |||
168 | const TargetMachine *TM; | |||
169 | const TargetSubtargetInfo *SubtargetInfo; | |||
170 | const TargetLowering *TLI; | |||
171 | const TargetRegisterInfo *TRI; | |||
172 | const TargetTransformInfo *TTI; | |||
173 | const TargetLibraryInfo *TLInfo; | |||
174 | const LoopInfo *LI; | |||
175 | std::unique_ptr<BlockFrequencyInfo> BFI; | |||
176 | std::unique_ptr<BranchProbabilityInfo> BPI; | |||
177 | ||||
178 | /// As we scan instructions optimizing them, this is the next instruction | |||
179 | /// to optimize. Transforms that can invalidate this should update it. | |||
180 | BasicBlock::iterator CurInstIterator; | |||
181 | ||||
182 | /// Keeps track of non-local addresses that have been sunk into a block. | |||
183 | /// This allows us to avoid inserting duplicate code for blocks with | |||
184 | /// multiple load/stores of the same address. | |||
185 | ValueMap<Value*, Value*> SunkAddrs; | |||
186 | ||||
187 | /// Keeps track of all instructions inserted for the current function. | |||
188 | SetOfInstrs InsertedInsts; | |||
189 | /// Keeps track of the type of the related instruction before their | |||
190 | /// promotion for the current function. | |||
191 | InstrToOrigTy PromotedInsts; | |||
192 | ||||
193 | /// Keep track of instructions removed during promotion. | |||
194 | SetOfInstrs RemovedInsts; | |||
195 | ||||
196 | /// Keep track of sext chains based on their initial value. | |||
197 | DenseMap<Value *, Instruction *> SeenChainsForSExt; | |||
198 | ||||
199 | /// Keep track of SExt promoted. | |||
200 | ValueToSExts ValToSExtendedUses; | |||
201 | ||||
202 | /// True if CFG is modified in any way. | |||
203 | bool ModifiedDT; | |||
204 | ||||
205 | /// True if optimizing for size. | |||
206 | bool OptSize; | |||
207 | ||||
208 | /// DataLayout for the Function being processed. | |||
209 | const DataLayout *DL; | |||
210 | ||||
211 | public: | |||
212 | static char ID; // Pass identification, replacement for typeid | |||
213 | CodeGenPrepare() | |||
214 | : FunctionPass(ID), TM(nullptr), TLI(nullptr), TTI(nullptr), | |||
215 | DL(nullptr) { | |||
216 | initializeCodeGenPreparePass(*PassRegistry::getPassRegistry()); | |||
217 | } | |||
218 | bool runOnFunction(Function &F) override; | |||
219 | ||||
220 | StringRef getPassName() const override { return "CodeGen Prepare"; } | |||
221 | ||||
222 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
223 | // FIXME: When we can selectively preserve passes, preserve the domtree. | |||
224 | AU.addRequired<ProfileSummaryInfoWrapperPass>(); | |||
225 | AU.addRequired<TargetLibraryInfoWrapperPass>(); | |||
226 | AU.addRequired<TargetTransformInfoWrapperPass>(); | |||
227 | AU.addRequired<LoopInfoWrapperPass>(); | |||
228 | } | |||
229 | ||||
230 | private: | |||
231 | bool eliminateFallThrough(Function &F); | |||
232 | bool eliminateMostlyEmptyBlocks(Function &F); | |||
233 | BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB); | |||
234 | bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; | |||
235 | void eliminateMostlyEmptyBlock(BasicBlock *BB); | |||
236 | bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB, | |||
237 | bool isPreheader); | |||
238 | bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT); | |||
239 | bool optimizeInst(Instruction *I, bool &ModifiedDT); | |||
240 | bool optimizeMemoryInst(Instruction *I, Value *Addr, | |||
241 | Type *AccessTy, unsigned AS); | |||
242 | bool optimizeInlineAsmInst(CallInst *CS); | |||
243 | bool optimizeCallInst(CallInst *CI, bool &ModifiedDT); | |||
244 | bool optimizeExt(Instruction *&I); | |||
245 | bool optimizeExtUses(Instruction *I); | |||
246 | bool optimizeLoadExt(LoadInst *I); | |||
247 | bool optimizeSelectInst(SelectInst *SI); | |||
248 | bool optimizeShuffleVectorInst(ShuffleVectorInst *SI); | |||
249 | bool optimizeSwitchInst(SwitchInst *CI); | |||
250 | bool optimizeExtractElementInst(Instruction *Inst); | |||
251 | bool dupRetToEnableTailCallOpts(BasicBlock *BB); | |||
252 | bool placeDbgValues(Function &F); | |||
253 | bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts, | |||
254 | LoadInst *&LI, Instruction *&Inst, bool HasPromoted); | |||
255 | bool tryToPromoteExts(TypePromotionTransaction &TPT, | |||
256 | const SmallVectorImpl<Instruction *> &Exts, | |||
257 | SmallVectorImpl<Instruction *> &ProfitablyMovedExts, | |||
258 | unsigned CreatedInstsCost = 0); | |||
259 | bool mergeSExts(Function &F); | |||
260 | bool performAddressTypePromotion( | |||
261 | Instruction *&Inst, | |||
262 | bool AllowPromotionWithoutCommonHeader, | |||
263 | bool HasPromoted, TypePromotionTransaction &TPT, | |||
264 | SmallVectorImpl<Instruction *> &SpeculativelyMovedExts); | |||
265 | bool splitBranchCondition(Function &F); | |||
266 | bool simplifyOffsetableRelocate(Instruction &I); | |||
267 | bool splitIndirectCriticalEdges(Function &F); | |||
268 | }; | |||
269 | } | |||
270 | ||||
271 | char CodeGenPrepare::ID = 0; | |||
272 | INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,static void *initializeCodeGenPreparePassOnce(PassRegistry & Registry) { | |||
273 | "Optimize for code generation", false, false)static void *initializeCodeGenPreparePassOnce(PassRegistry & Registry) { | |||
274 | INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)initializeProfileSummaryInfoWrapperPassPass(Registry); | |||
275 | INITIALIZE_PASS_END(CodeGenPrepare, DEBUG_TYPE,PassInfo *PI = new PassInfo( "Optimize for code generation", "codegenprepare" , &CodeGenPrepare::ID, PassInfo::NormalCtor_t(callDefaultCtor <CodeGenPrepare>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeCodeGenPreparePassFlag ; void llvm::initializeCodeGenPreparePass(PassRegistry &Registry ) { llvm::call_once(InitializeCodeGenPreparePassFlag, initializeCodeGenPreparePassOnce , std::ref(Registry)); } | |||
276 | "Optimize for code generation", false, false)PassInfo *PI = new PassInfo( "Optimize for code generation", "codegenprepare" , &CodeGenPrepare::ID, PassInfo::NormalCtor_t(callDefaultCtor <CodeGenPrepare>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeCodeGenPreparePassFlag ; void llvm::initializeCodeGenPreparePass(PassRegistry &Registry ) { llvm::call_once(InitializeCodeGenPreparePassFlag, initializeCodeGenPreparePassOnce , std::ref(Registry)); } | |||
277 | ||||
278 | FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); } | |||
279 | ||||
280 | bool CodeGenPrepare::runOnFunction(Function &F) { | |||
281 | if (skipFunction(F)) | |||
282 | return false; | |||
283 | ||||
284 | DL = &F.getParent()->getDataLayout(); | |||
285 | ||||
286 | bool EverMadeChange = false; | |||
287 | // Clear per function information. | |||
288 | InsertedInsts.clear(); | |||
289 | PromotedInsts.clear(); | |||
290 | BFI.reset(); | |||
291 | BPI.reset(); | |||
292 | ||||
293 | ModifiedDT = false; | |||
294 | if (auto *TPC = getAnalysisIfAvailable<TargetPassConfig>()) { | |||
295 | TM = &TPC->getTM<TargetMachine>(); | |||
296 | SubtargetInfo = TM->getSubtargetImpl(F); | |||
297 | TLI = SubtargetInfo->getTargetLowering(); | |||
298 | TRI = SubtargetInfo->getRegisterInfo(); | |||
299 | } | |||
300 | TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(); | |||
301 | TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | |||
302 | LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | |||
303 | OptSize = F.optForSize(); | |||
304 | ||||
305 | if (ProfileGuidedSectionPrefix) { | |||
306 | ProfileSummaryInfo *PSI = | |||
307 | getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); | |||
308 | if (PSI->isFunctionHotInCallGraph(&F)) | |||
309 | F.setSectionPrefix(".hot"); | |||
310 | else if (PSI->isFunctionColdInCallGraph(&F)) | |||
311 | F.setSectionPrefix(".unlikely"); | |||
312 | } | |||
313 | ||||
314 | /// This optimization identifies DIV instructions that can be | |||
315 | /// profitably bypassed and carried out with a shorter, faster divide. | |||
316 | if (!OptSize && TLI && TLI->isSlowDivBypassed()) { | |||
317 | const DenseMap<unsigned int, unsigned int> &BypassWidths = | |||
318 | TLI->getBypassSlowDivWidths(); | |||
319 | BasicBlock* BB = &*F.begin(); | |||
320 | while (BB != nullptr) { | |||
321 | // bypassSlowDivision may create new BBs, but we don't want to reapply the | |||
322 | // optimization to those blocks. | |||
323 | BasicBlock* Next = BB->getNextNode(); | |||
324 | EverMadeChange |= bypassSlowDivision(BB, BypassWidths); | |||
325 | BB = Next; | |||
326 | } | |||
327 | } | |||
328 | ||||
329 | // Eliminate blocks that contain only PHI nodes and an | |||
330 | // unconditional branch. | |||
331 | EverMadeChange |= eliminateMostlyEmptyBlocks(F); | |||
332 | ||||
333 | // llvm.dbg.value is far away from the value then iSel may not be able | |||
334 | // handle it properly. iSel will drop llvm.dbg.value if it can not | |||
335 | // find a node corresponding to the value. | |||
336 | EverMadeChange |= placeDbgValues(F); | |||
337 | ||||
338 | if (!DisableBranchOpts) | |||
339 | EverMadeChange |= splitBranchCondition(F); | |||
340 | ||||
341 | // Split some critical edges where one of the sources is an indirect branch, | |||
342 | // to help generate sane code for PHIs involving such edges. | |||
343 | EverMadeChange |= splitIndirectCriticalEdges(F); | |||
344 | ||||
345 | bool MadeChange = true; | |||
346 | while (MadeChange) { | |||
347 | MadeChange = false; | |||
348 | SeenChainsForSExt.clear(); | |||
349 | ValToSExtendedUses.clear(); | |||
350 | RemovedInsts.clear(); | |||
351 | for (Function::iterator I = F.begin(); I != F.end(); ) { | |||
352 | BasicBlock *BB = &*I++; | |||
353 | bool ModifiedDTOnIteration = false; | |||
354 | MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration); | |||
355 | ||||
356 | // Restart BB iteration if the dominator tree of the Function was changed | |||
357 | if (ModifiedDTOnIteration) | |||
358 | break; | |||
359 | } | |||
360 | if (EnableTypePromotionMerge && !ValToSExtendedUses.empty()) | |||
361 | MadeChange |= mergeSExts(F); | |||
362 | ||||
363 | // Really free removed instructions during promotion. | |||
364 | for (Instruction *I : RemovedInsts) | |||
365 | I->deleteValue(); | |||
366 | ||||
367 | EverMadeChange |= MadeChange; | |||
368 | } | |||
369 | ||||
370 | SunkAddrs.clear(); | |||
371 | ||||
372 | if (!DisableBranchOpts) { | |||
373 | MadeChange = false; | |||
374 | SmallPtrSet<BasicBlock*, 8> WorkList; | |||
375 | for (BasicBlock &BB : F) { | |||
376 | SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB)); | |||
377 | MadeChange |= ConstantFoldTerminator(&BB, true); | |||
378 | if (!MadeChange) continue; | |||
379 | ||||
380 | for (SmallVectorImpl<BasicBlock*>::iterator | |||
381 | II = Successors.begin(), IE = Successors.end(); II != IE; ++II) | |||
382 | if (pred_begin(*II) == pred_end(*II)) | |||
383 | WorkList.insert(*II); | |||
384 | } | |||
385 | ||||
386 | // Delete the dead blocks and any of their dead successors. | |||
387 | MadeChange |= !WorkList.empty(); | |||
388 | while (!WorkList.empty()) { | |||
389 | BasicBlock *BB = *WorkList.begin(); | |||
390 | WorkList.erase(BB); | |||
391 | SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB)); | |||
392 | ||||
393 | DeleteDeadBlock(BB); | |||
394 | ||||
395 | for (SmallVectorImpl<BasicBlock*>::iterator | |||
396 | II = Successors.begin(), IE = Successors.end(); II != IE; ++II) | |||
397 | if (pred_begin(*II) == pred_end(*II)) | |||
398 | WorkList.insert(*II); | |||
399 | } | |||
400 | ||||
401 | // Merge pairs of basic blocks with unconditional branches, connected by | |||
402 | // a single edge. | |||
403 | if (EverMadeChange || MadeChange) | |||
404 | MadeChange |= eliminateFallThrough(F); | |||
405 | ||||
406 | EverMadeChange |= MadeChange; | |||
407 | } | |||
408 | ||||
409 | if (!DisableGCOpts) { | |||
410 | SmallVector<Instruction *, 2> Statepoints; | |||
411 | for (BasicBlock &BB : F) | |||
412 | for (Instruction &I : BB) | |||
413 | if (isStatepoint(I)) | |||
414 | Statepoints.push_back(&I); | |||
415 | for (auto &I : Statepoints) | |||
416 | EverMadeChange |= simplifyOffsetableRelocate(*I); | |||
417 | } | |||
418 | ||||
419 | return EverMadeChange; | |||
420 | } | |||
421 | ||||
422 | /// Merge basic blocks which are connected by a single edge, where one of the | |||
423 | /// basic blocks has a single successor pointing to the other basic block, | |||
424 | /// which has a single predecessor. | |||
425 | bool CodeGenPrepare::eliminateFallThrough(Function &F) { | |||
426 | bool Changed = false; | |||
427 | // Scan all of the blocks in the function, except for the entry block. | |||
428 | for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) { | |||
429 | BasicBlock *BB = &*I++; | |||
430 | // If the destination block has a single pred, then this is a trivial | |||
431 | // edge, just collapse it. | |||
432 | BasicBlock *SinglePred = BB->getSinglePredecessor(); | |||
433 | ||||
434 | // Don't merge if BB's address is taken. | |||
435 | if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue; | |||
436 | ||||
437 | BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator()); | |||
438 | if (Term && !Term->isConditional()) { | |||
439 | Changed = true; | |||
440 | DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "To merge:\n"<< * SinglePred << "\n\n\n"; } } while (false); | |||
441 | // Remember if SinglePred was the entry block of the function. | |||
442 | // If so, we will need to move BB back to the entry position. | |||
443 | bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); | |||
444 | MergeBasicBlockIntoOnlyPred(BB, nullptr); | |||
445 | ||||
446 | if (isEntry && BB != &BB->getParent()->getEntryBlock()) | |||
447 | BB->moveBefore(&BB->getParent()->getEntryBlock()); | |||
448 | ||||
449 | // We have erased a block. Update the iterator. | |||
450 | I = BB->getIterator(); | |||
451 | } | |||
452 | } | |||
453 | return Changed; | |||
454 | } | |||
455 | ||||
456 | /// Find a destination block from BB if BB is mergeable empty block. | |||
457 | BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) { | |||
458 | // If this block doesn't end with an uncond branch, ignore it. | |||
459 | BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); | |||
460 | if (!BI || !BI->isUnconditional()) | |||
461 | return nullptr; | |||
462 | ||||
463 | // If the instruction before the branch (skipping debug info) isn't a phi | |||
464 | // node, then other stuff is happening here. | |||
465 | BasicBlock::iterator BBI = BI->getIterator(); | |||
466 | if (BBI != BB->begin()) { | |||
467 | --BBI; | |||
468 | while (isa<DbgInfoIntrinsic>(BBI)) { | |||
469 | if (BBI == BB->begin()) | |||
470 | break; | |||
471 | --BBI; | |||
472 | } | |||
473 | if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI)) | |||
474 | return nullptr; | |||
475 | } | |||
476 | ||||
477 | // Do not break infinite loops. | |||
478 | BasicBlock *DestBB = BI->getSuccessor(0); | |||
479 | if (DestBB == BB) | |||
480 | return nullptr; | |||
481 | ||||
482 | if (!canMergeBlocks(BB, DestBB)) | |||
483 | DestBB = nullptr; | |||
484 | ||||
485 | return DestBB; | |||
486 | } | |||
487 | ||||
488 | // Return the unique indirectbr predecessor of a block. This may return null | |||
489 | // even if such a predecessor exists, if it's not useful for splitting. | |||
490 | // If a predecessor is found, OtherPreds will contain all other (non-indirectbr) | |||
491 | // predecessors of BB. | |||
492 | static BasicBlock * | |||
493 | findIBRPredecessor(BasicBlock *BB, SmallVectorImpl<BasicBlock *> &OtherPreds) { | |||
494 | // If the block doesn't have any PHIs, we don't care about it, since there's | |||
495 | // no point in splitting it. | |||
496 | PHINode *PN = dyn_cast<PHINode>(BB->begin()); | |||
497 | if (!PN) | |||
498 | return nullptr; | |||
499 | ||||
500 | // Verify we have exactly one IBR predecessor. | |||
501 | // Conservatively bail out if one of the other predecessors is not a "regular" | |||
502 | // terminator (that is, not a switch or a br). | |||
503 | BasicBlock *IBB = nullptr; | |||
504 | for (unsigned Pred = 0, E = PN->getNumIncomingValues(); Pred != E; ++Pred) { | |||
505 | BasicBlock *PredBB = PN->getIncomingBlock(Pred); | |||
506 | TerminatorInst *PredTerm = PredBB->getTerminator(); | |||
507 | switch (PredTerm->getOpcode()) { | |||
508 | case Instruction::IndirectBr: | |||
509 | if (IBB) | |||
510 | return nullptr; | |||
511 | IBB = PredBB; | |||
512 | break; | |||
513 | case Instruction::Br: | |||
514 | case Instruction::Switch: | |||
515 | OtherPreds.push_back(PredBB); | |||
516 | continue; | |||
517 | default: | |||
518 | return nullptr; | |||
519 | } | |||
520 | } | |||
521 | ||||
522 | return IBB; | |||
523 | } | |||
524 | ||||
525 | // Split critical edges where the source of the edge is an indirectbr | |||
526 | // instruction. This isn't always possible, but we can handle some easy cases. | |||
527 | // This is useful because MI is unable to split such critical edges, | |||
528 | // which means it will not be able to sink instructions along those edges. | |||
529 | // This is especially painful for indirect branches with many successors, where | |||
530 | // we end up having to prepare all outgoing values in the origin block. | |||
531 | // | |||
532 | // Our normal algorithm for splitting critical edges requires us to update | |||
533 | // the outgoing edges of the edge origin block, but for an indirectbr this | |||
534 | // is hard, since it would require finding and updating the block addresses | |||
535 | // the indirect branch uses. But if a block only has a single indirectbr | |||
536 | // predecessor, with the others being regular branches, we can do it in a | |||
537 | // different way. | |||
538 | // Say we have A -> D, B -> D, I -> D where only I -> D is an indirectbr. | |||
539 | // We can split D into D0 and D1, where D0 contains only the PHIs from D, | |||
540 | // and D1 is the D block body. We can then duplicate D0 as D0A and D0B, and | |||
541 | // create the following structure: | |||
542 | // A -> D0A, B -> D0A, I -> D0B, D0A -> D1, D0B -> D1 | |||
543 | bool CodeGenPrepare::splitIndirectCriticalEdges(Function &F) { | |||
544 | // Check whether the function has any indirectbrs, and collect which blocks | |||
545 | // they may jump to. Since most functions don't have indirect branches, | |||
546 | // this lowers the common case's overhead to O(Blocks) instead of O(Edges). | |||
547 | SmallSetVector<BasicBlock *, 16> Targets; | |||
548 | for (auto &BB : F) { | |||
549 | auto *IBI = dyn_cast<IndirectBrInst>(BB.getTerminator()); | |||
550 | if (!IBI) | |||
551 | continue; | |||
552 | ||||
553 | for (unsigned Succ = 0, E = IBI->getNumSuccessors(); Succ != E; ++Succ) | |||
554 | Targets.insert(IBI->getSuccessor(Succ)); | |||
555 | } | |||
556 | ||||
557 | if (Targets.empty()) | |||
558 | return false; | |||
559 | ||||
560 | bool Changed = false; | |||
561 | for (BasicBlock *Target : Targets) { | |||
562 | SmallVector<BasicBlock *, 16> OtherPreds; | |||
563 | BasicBlock *IBRPred = findIBRPredecessor(Target, OtherPreds); | |||
564 | // If we did not found an indirectbr, or the indirectbr is the only | |||
565 | // incoming edge, this isn't the kind of edge we're looking for. | |||
566 | if (!IBRPred || OtherPreds.empty()) | |||
567 | continue; | |||
568 | ||||
569 | // Don't even think about ehpads/landingpads. | |||
570 | Instruction *FirstNonPHI = Target->getFirstNonPHI(); | |||
571 | if (FirstNonPHI->isEHPad() || Target->isLandingPad()) | |||
572 | continue; | |||
573 | ||||
574 | BasicBlock *BodyBlock = Target->splitBasicBlock(FirstNonPHI, ".split"); | |||
575 | // It's possible Target was its own successor through an indirectbr. | |||
576 | // In this case, the indirectbr now comes from BodyBlock. | |||
577 | if (IBRPred == Target) | |||
578 | IBRPred = BodyBlock; | |||
579 | ||||
580 | // At this point Target only has PHIs, and BodyBlock has the rest of the | |||
581 | // block's body. Create a copy of Target that will be used by the "direct" | |||
582 | // preds. | |||
583 | ValueToValueMapTy VMap; | |||
584 | BasicBlock *DirectSucc = CloneBasicBlock(Target, VMap, ".clone", &F); | |||
585 | ||||
586 | for (BasicBlock *Pred : OtherPreds) { | |||
587 | // If the target is a loop to itself, then the terminator of the split | |||
588 | // block needs to be updated. | |||
589 | if (Pred == Target) | |||
590 | BodyBlock->getTerminator()->replaceUsesOfWith(Target, DirectSucc); | |||
591 | else | |||
592 | Pred->getTerminator()->replaceUsesOfWith(Target, DirectSucc); | |||
593 | } | |||
594 | ||||
595 | // Ok, now fix up the PHIs. We know the two blocks only have PHIs, and that | |||
596 | // they are clones, so the number of PHIs are the same. | |||
597 | // (a) Remove the edge coming from IBRPred from the "Direct" PHI | |||
598 | // (b) Leave that as the only edge in the "Indirect" PHI. | |||
599 | // (c) Merge the two in the body block. | |||
600 | BasicBlock::iterator Indirect = Target->begin(), | |||
601 | End = Target->getFirstNonPHI()->getIterator(); | |||
602 | BasicBlock::iterator Direct = DirectSucc->begin(); | |||
603 | BasicBlock::iterator MergeInsert = BodyBlock->getFirstInsertionPt(); | |||
604 | ||||
605 | assert(&*End == Target->getTerminator() &&((&*End == Target->getTerminator() && "Block was expected to only contain PHIs" ) ? static_cast<void> (0) : __assert_fail ("&*End == Target->getTerminator() && \"Block was expected to only contain PHIs\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 606, __PRETTY_FUNCTION__)) | |||
606 | "Block was expected to only contain PHIs")((&*End == Target->getTerminator() && "Block was expected to only contain PHIs" ) ? static_cast<void> (0) : __assert_fail ("&*End == Target->getTerminator() && \"Block was expected to only contain PHIs\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 606, __PRETTY_FUNCTION__)); | |||
607 | ||||
608 | while (Indirect != End) { | |||
609 | PHINode *DirPHI = cast<PHINode>(Direct); | |||
610 | PHINode *IndPHI = cast<PHINode>(Indirect); | |||
611 | ||||
612 | // Now, clean up - the direct block shouldn't get the indirect value, | |||
613 | // and vice versa. | |||
614 | DirPHI->removeIncomingValue(IBRPred); | |||
615 | Direct++; | |||
616 | ||||
617 | // Advance the pointer here, to avoid invalidation issues when the old | |||
618 | // PHI is erased. | |||
619 | Indirect++; | |||
620 | ||||
621 | PHINode *NewIndPHI = PHINode::Create(IndPHI->getType(), 1, "ind", IndPHI); | |||
622 | NewIndPHI->addIncoming(IndPHI->getIncomingValueForBlock(IBRPred), | |||
623 | IBRPred); | |||
624 | ||||
625 | // Create a PHI in the body block, to merge the direct and indirect | |||
626 | // predecessors. | |||
627 | PHINode *MergePHI = | |||
628 | PHINode::Create(IndPHI->getType(), 2, "merge", &*MergeInsert); | |||
629 | MergePHI->addIncoming(NewIndPHI, Target); | |||
630 | MergePHI->addIncoming(DirPHI, DirectSucc); | |||
631 | ||||
632 | IndPHI->replaceAllUsesWith(MergePHI); | |||
633 | IndPHI->eraseFromParent(); | |||
634 | } | |||
635 | ||||
636 | Changed = true; | |||
637 | } | |||
638 | ||||
639 | return Changed; | |||
640 | } | |||
641 | ||||
642 | /// Eliminate blocks that contain only PHI nodes, debug info directives, and an | |||
643 | /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split | |||
644 | /// edges in ways that are non-optimal for isel. Start by eliminating these | |||
645 | /// blocks so we can split them the way we want them. | |||
646 | bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) { | |||
647 | SmallPtrSet<BasicBlock *, 16> Preheaders; | |||
648 | SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end()); | |||
649 | while (!LoopList.empty()) { | |||
650 | Loop *L = LoopList.pop_back_val(); | |||
651 | LoopList.insert(LoopList.end(), L->begin(), L->end()); | |||
652 | if (BasicBlock *Preheader = L->getLoopPreheader()) | |||
653 | Preheaders.insert(Preheader); | |||
654 | } | |||
655 | ||||
656 | bool MadeChange = false; | |||
657 | // Note that this intentionally skips the entry block. | |||
658 | for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) { | |||
659 | BasicBlock *BB = &*I++; | |||
660 | BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB); | |||
661 | if (!DestBB || | |||
662 | !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB))) | |||
663 | continue; | |||
664 | ||||
665 | eliminateMostlyEmptyBlock(BB); | |||
666 | MadeChange = true; | |||
667 | } | |||
668 | return MadeChange; | |||
669 | } | |||
670 | ||||
671 | bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB, | |||
672 | BasicBlock *DestBB, | |||
673 | bool isPreheader) { | |||
674 | // Do not delete loop preheaders if doing so would create a critical edge. | |||
675 | // Loop preheaders can be good locations to spill registers. If the | |||
676 | // preheader is deleted and we create a critical edge, registers may be | |||
677 | // spilled in the loop body instead. | |||
678 | if (!DisablePreheaderProtect && isPreheader && | |||
679 | !(BB->getSinglePredecessor() && | |||
680 | BB->getSinglePredecessor()->getSingleSuccessor())) | |||
681 | return false; | |||
682 | ||||
683 | // Try to skip merging if the unique predecessor of BB is terminated by a | |||
684 | // switch or indirect branch instruction, and BB is used as an incoming block | |||
685 | // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to | |||
686 | // add COPY instructions in the predecessor of BB instead of BB (if it is not | |||
687 | // merged). Note that the critical edge created by merging such blocks wont be | |||
688 | // split in MachineSink because the jump table is not analyzable. By keeping | |||
689 | // such empty block (BB), ISel will place COPY instructions in BB, not in the | |||
690 | // predecessor of BB. | |||
691 | BasicBlock *Pred = BB->getUniquePredecessor(); | |||
692 | if (!Pred || | |||
693 | !(isa<SwitchInst>(Pred->getTerminator()) || | |||
694 | isa<IndirectBrInst>(Pred->getTerminator()))) | |||
695 | return true; | |||
696 | ||||
697 | if (BB->getTerminator() != BB->getFirstNonPHI()) | |||
698 | return true; | |||
699 | ||||
700 | // We use a simple cost heuristic which determine skipping merging is | |||
701 | // profitable if the cost of skipping merging is less than the cost of | |||
702 | // merging : Cost(skipping merging) < Cost(merging BB), where the | |||
703 | // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and | |||
704 | // the Cost(merging BB) is Freq(Pred) * Cost(Copy). | |||
705 | // Assuming Cost(Copy) == Cost(Branch), we could simplify it to : | |||
706 | // Freq(Pred) / Freq(BB) > 2. | |||
707 | // Note that if there are multiple empty blocks sharing the same incoming | |||
708 | // value for the PHIs in the DestBB, we consider them together. In such | |||
709 | // case, Cost(merging BB) will be the sum of their frequencies. | |||
710 | ||||
711 | if (!isa<PHINode>(DestBB->begin())) | |||
712 | return true; | |||
713 | ||||
714 | SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs; | |||
715 | ||||
716 | // Find all other incoming blocks from which incoming values of all PHIs in | |||
717 | // DestBB are the same as the ones from BB. | |||
718 | for (pred_iterator PI = pred_begin(DestBB), E = pred_end(DestBB); PI != E; | |||
719 | ++PI) { | |||
720 | BasicBlock *DestBBPred = *PI; | |||
721 | if (DestBBPred == BB) | |||
722 | continue; | |||
723 | ||||
724 | bool HasAllSameValue = true; | |||
725 | BasicBlock::const_iterator DestBBI = DestBB->begin(); | |||
726 | while (const PHINode *DestPN = dyn_cast<PHINode>(DestBBI++)) { | |||
727 | if (DestPN->getIncomingValueForBlock(BB) != | |||
728 | DestPN->getIncomingValueForBlock(DestBBPred)) { | |||
729 | HasAllSameValue = false; | |||
730 | break; | |||
731 | } | |||
732 | } | |||
733 | if (HasAllSameValue) | |||
734 | SameIncomingValueBBs.insert(DestBBPred); | |||
735 | } | |||
736 | ||||
737 | // See if all BB's incoming values are same as the value from Pred. In this | |||
738 | // case, no reason to skip merging because COPYs are expected to be place in | |||
739 | // Pred already. | |||
740 | if (SameIncomingValueBBs.count(Pred)) | |||
741 | return true; | |||
742 | ||||
743 | if (!BFI) { | |||
744 | Function &F = *BB->getParent(); | |||
745 | LoopInfo LI{DominatorTree(F)}; | |||
746 | BPI.reset(new BranchProbabilityInfo(F, LI)); | |||
747 | BFI.reset(new BlockFrequencyInfo(F, *BPI, LI)); | |||
748 | } | |||
749 | ||||
750 | BlockFrequency PredFreq = BFI->getBlockFreq(Pred); | |||
751 | BlockFrequency BBFreq = BFI->getBlockFreq(BB); | |||
752 | ||||
753 | for (auto SameValueBB : SameIncomingValueBBs) | |||
754 | if (SameValueBB->getUniquePredecessor() == Pred && | |||
755 | DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB)) | |||
756 | BBFreq += BFI->getBlockFreq(SameValueBB); | |||
757 | ||||
758 | return PredFreq.getFrequency() <= | |||
759 | BBFreq.getFrequency() * FreqRatioToSkipMerge; | |||
760 | } | |||
761 | ||||
762 | /// Return true if we can merge BB into DestBB if there is a single | |||
763 | /// unconditional branch between them, and BB contains no other non-phi | |||
764 | /// instructions. | |||
765 | bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB, | |||
766 | const BasicBlock *DestBB) const { | |||
767 | // We only want to eliminate blocks whose phi nodes are used by phi nodes in | |||
768 | // the successor. If there are more complex condition (e.g. preheaders), | |||
769 | // don't mess around with them. | |||
770 | BasicBlock::const_iterator BBI = BB->begin(); | |||
771 | while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { | |||
772 | for (const User *U : PN->users()) { | |||
773 | const Instruction *UI = cast<Instruction>(U); | |||
774 | if (UI->getParent() != DestBB || !isa<PHINode>(UI)) | |||
775 | return false; | |||
776 | // If User is inside DestBB block and it is a PHINode then check | |||
777 | // incoming value. If incoming value is not from BB then this is | |||
778 | // a complex condition (e.g. preheaders) we want to avoid here. | |||
779 | if (UI->getParent() == DestBB) { | |||
780 | if (const PHINode *UPN = dyn_cast<PHINode>(UI)) | |||
781 | for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) { | |||
782 | Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I)); | |||
783 | if (Insn && Insn->getParent() == BB && | |||
784 | Insn->getParent() != UPN->getIncomingBlock(I)) | |||
785 | return false; | |||
786 | } | |||
787 | } | |||
788 | } | |||
789 | } | |||
790 | ||||
791 | // If BB and DestBB contain any common predecessors, then the phi nodes in BB | |||
792 | // and DestBB may have conflicting incoming values for the block. If so, we | |||
793 | // can't merge the block. | |||
794 | const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin()); | |||
795 | if (!DestBBPN) return true; // no conflict. | |||
796 | ||||
797 | // Collect the preds of BB. | |||
798 | SmallPtrSet<const BasicBlock*, 16> BBPreds; | |||
799 | if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { | |||
800 | // It is faster to get preds from a PHI than with pred_iterator. | |||
801 | for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) | |||
802 | BBPreds.insert(BBPN->getIncomingBlock(i)); | |||
803 | } else { | |||
804 | BBPreds.insert(pred_begin(BB), pred_end(BB)); | |||
805 | } | |||
806 | ||||
807 | // Walk the preds of DestBB. | |||
808 | for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) { | |||
809 | BasicBlock *Pred = DestBBPN->getIncomingBlock(i); | |||
810 | if (BBPreds.count(Pred)) { // Common predecessor? | |||
811 | BBI = DestBB->begin(); | |||
812 | while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { | |||
813 | const Value *V1 = PN->getIncomingValueForBlock(Pred); | |||
814 | const Value *V2 = PN->getIncomingValueForBlock(BB); | |||
815 | ||||
816 | // If V2 is a phi node in BB, look up what the mapped value will be. | |||
817 | if (const PHINode *V2PN = dyn_cast<PHINode>(V2)) | |||
818 | if (V2PN->getParent() == BB) | |||
819 | V2 = V2PN->getIncomingValueForBlock(Pred); | |||
820 | ||||
821 | // If there is a conflict, bail out. | |||
822 | if (V1 != V2) return false; | |||
823 | } | |||
824 | } | |||
825 | } | |||
826 | ||||
827 | return true; | |||
828 | } | |||
829 | ||||
830 | ||||
831 | /// Eliminate a basic block that has only phi's and an unconditional branch in | |||
832 | /// it. | |||
833 | void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) { | |||
834 | BranchInst *BI = cast<BranchInst>(BB->getTerminator()); | |||
835 | BasicBlock *DestBB = BI->getSuccessor(0); | |||
836 | ||||
837 | DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB; } } while (false); | |||
838 | ||||
839 | // If the destination block has a single pred, then this is a trivial edge, | |||
840 | // just collapse it. | |||
841 | if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) { | |||
842 | if (SinglePred != DestBB) { | |||
843 | // Remember if SinglePred was the entry block of the function. If so, we | |||
844 | // will need to move BB back to the entry position. | |||
845 | bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); | |||
846 | MergeBasicBlockIntoOnlyPred(DestBB, nullptr); | |||
847 | ||||
848 | if (isEntry && BB != &BB->getParent()->getEntryBlock()) | |||
849 | BB->moveBefore(&BB->getParent()->getEntryBlock()); | |||
850 | ||||
851 | DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "AFTER:\n" << *DestBB << "\n\n\n"; } } while (false); | |||
852 | return; | |||
853 | } | |||
854 | } | |||
855 | ||||
856 | // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB | |||
857 | // to handle the new incoming edges it is about to have. | |||
858 | PHINode *PN; | |||
859 | for (BasicBlock::iterator BBI = DestBB->begin(); | |||
860 | (PN = dyn_cast<PHINode>(BBI)); ++BBI) { | |||
861 | // Remove the incoming value for BB, and remember it. | |||
862 | Value *InVal = PN->removeIncomingValue(BB, false); | |||
863 | ||||
864 | // Two options: either the InVal is a phi node defined in BB or it is some | |||
865 | // value that dominates BB. | |||
866 | PHINode *InValPhi = dyn_cast<PHINode>(InVal); | |||
867 | if (InValPhi && InValPhi->getParent() == BB) { | |||
868 | // Add all of the input values of the input PHI as inputs of this phi. | |||
869 | for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i) | |||
870 | PN->addIncoming(InValPhi->getIncomingValue(i), | |||
871 | InValPhi->getIncomingBlock(i)); | |||
872 | } else { | |||
873 | // Otherwise, add one instance of the dominating value for each edge that | |||
874 | // we will be adding. | |||
875 | if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { | |||
876 | for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) | |||
877 | PN->addIncoming(InVal, BBPN->getIncomingBlock(i)); | |||
878 | } else { | |||
879 | for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) | |||
880 | PN->addIncoming(InVal, *PI); | |||
881 | } | |||
882 | } | |||
883 | } | |||
884 | ||||
885 | // The PHIs are now updated, change everything that refers to BB to use | |||
886 | // DestBB and remove BB. | |||
887 | BB->replaceAllUsesWith(DestBB); | |||
888 | BB->eraseFromParent(); | |||
889 | ++NumBlocksElim; | |||
890 | ||||
891 | DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "AFTER:\n" << *DestBB << "\n\n\n"; } } while (false); | |||
892 | } | |||
893 | ||||
894 | // Computes a map of base pointer relocation instructions to corresponding | |||
895 | // derived pointer relocation instructions given a vector of all relocate calls | |||
896 | static void computeBaseDerivedRelocateMap( | |||
897 | const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls, | |||
898 | DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> | |||
899 | &RelocateInstMap) { | |||
900 | // Collect information in two maps: one primarily for locating the base object | |||
901 | // while filling the second map; the second map is the final structure holding | |||
902 | // a mapping between Base and corresponding Derived relocate calls | |||
903 | DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap; | |||
904 | for (auto *ThisRelocate : AllRelocateCalls) { | |||
905 | auto K = std::make_pair(ThisRelocate->getBasePtrIndex(), | |||
906 | ThisRelocate->getDerivedPtrIndex()); | |||
907 | RelocateIdxMap.insert(std::make_pair(K, ThisRelocate)); | |||
908 | } | |||
909 | for (auto &Item : RelocateIdxMap) { | |||
910 | std::pair<unsigned, unsigned> Key = Item.first; | |||
911 | if (Key.first == Key.second) | |||
912 | // Base relocation: nothing to insert | |||
913 | continue; | |||
914 | ||||
915 | GCRelocateInst *I = Item.second; | |||
916 | auto BaseKey = std::make_pair(Key.first, Key.first); | |||
917 | ||||
918 | // We're iterating over RelocateIdxMap so we cannot modify it. | |||
919 | auto MaybeBase = RelocateIdxMap.find(BaseKey); | |||
920 | if (MaybeBase == RelocateIdxMap.end()) | |||
921 | // TODO: We might want to insert a new base object relocate and gep off | |||
922 | // that, if there are enough derived object relocates. | |||
923 | continue; | |||
924 | ||||
925 | RelocateInstMap[MaybeBase->second].push_back(I); | |||
926 | } | |||
927 | } | |||
928 | ||||
929 | // Accepts a GEP and extracts the operands into a vector provided they're all | |||
930 | // small integer constants | |||
931 | static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP, | |||
932 | SmallVectorImpl<Value *> &OffsetV) { | |||
933 | for (unsigned i = 1; i < GEP->getNumOperands(); i++) { | |||
934 | // Only accept small constant integer operands | |||
935 | auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i)); | |||
936 | if (!Op || Op->getZExtValue() > 20) | |||
937 | return false; | |||
938 | } | |||
939 | ||||
940 | for (unsigned i = 1; i < GEP->getNumOperands(); i++) | |||
941 | OffsetV.push_back(GEP->getOperand(i)); | |||
942 | return true; | |||
943 | } | |||
944 | ||||
945 | // Takes a RelocatedBase (base pointer relocation instruction) and Targets to | |||
946 | // replace, computes a replacement, and affects it. | |||
947 | static bool | |||
948 | simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase, | |||
949 | const SmallVectorImpl<GCRelocateInst *> &Targets) { | |||
950 | bool MadeChange = false; | |||
951 | for (GCRelocateInst *ToReplace : Targets) { | |||
952 | assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&((ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex () && "Not relocating a derived object of the original base object" ) ? static_cast<void> (0) : __assert_fail ("ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() && \"Not relocating a derived object of the original base object\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 953, __PRETTY_FUNCTION__)) | |||
953 | "Not relocating a derived object of the original base object")((ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex () && "Not relocating a derived object of the original base object" ) ? static_cast<void> (0) : __assert_fail ("ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() && \"Not relocating a derived object of the original base object\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 953, __PRETTY_FUNCTION__)); | |||
954 | if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) { | |||
955 | // A duplicate relocate call. TODO: coalesce duplicates. | |||
956 | continue; | |||
957 | } | |||
958 | ||||
959 | if (RelocatedBase->getParent() != ToReplace->getParent()) { | |||
960 | // Base and derived relocates are in different basic blocks. | |||
961 | // In this case transform is only valid when base dominates derived | |||
962 | // relocate. However it would be too expensive to check dominance | |||
963 | // for each such relocate, so we skip the whole transformation. | |||
964 | continue; | |||
965 | } | |||
966 | ||||
967 | Value *Base = ToReplace->getBasePtr(); | |||
968 | auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr()); | |||
969 | if (!Derived || Derived->getPointerOperand() != Base) | |||
970 | continue; | |||
971 | ||||
972 | SmallVector<Value *, 2> OffsetV; | |||
973 | if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV)) | |||
974 | continue; | |||
975 | ||||
976 | // Create a Builder and replace the target callsite with a gep | |||
977 | assert(RelocatedBase->getNextNode() &&((RelocatedBase->getNextNode() && "Should always have one since it's not a terminator" ) ? static_cast<void> (0) : __assert_fail ("RelocatedBase->getNextNode() && \"Should always have one since it's not a terminator\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 978, __PRETTY_FUNCTION__)) | |||
978 | "Should always have one since it's not a terminator")((RelocatedBase->getNextNode() && "Should always have one since it's not a terminator" ) ? static_cast<void> (0) : __assert_fail ("RelocatedBase->getNextNode() && \"Should always have one since it's not a terminator\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 978, __PRETTY_FUNCTION__)); | |||
979 | ||||
980 | // Insert after RelocatedBase | |||
981 | IRBuilder<> Builder(RelocatedBase->getNextNode()); | |||
982 | Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc()); | |||
983 | ||||
984 | // If gc_relocate does not match the actual type, cast it to the right type. | |||
985 | // In theory, there must be a bitcast after gc_relocate if the type does not | |||
986 | // match, and we should reuse it to get the derived pointer. But it could be | |||
987 | // cases like this: | |||
988 | // bb1: | |||
989 | // ... | |||
990 | // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...) | |||
991 | // br label %merge | |||
992 | // | |||
993 | // bb2: | |||
994 | // ... | |||
995 | // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...) | |||
996 | // br label %merge | |||
997 | // | |||
998 | // merge: | |||
999 | // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ] | |||
1000 | // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)* | |||
1001 | // | |||
1002 | // In this case, we can not find the bitcast any more. So we insert a new bitcast | |||
1003 | // no matter there is already one or not. In this way, we can handle all cases, and | |||
1004 | // the extra bitcast should be optimized away in later passes. | |||
1005 | Value *ActualRelocatedBase = RelocatedBase; | |||
1006 | if (RelocatedBase->getType() != Base->getType()) { | |||
1007 | ActualRelocatedBase = | |||
1008 | Builder.CreateBitCast(RelocatedBase, Base->getType()); | |||
1009 | } | |||
1010 | Value *Replacement = Builder.CreateGEP( | |||
1011 | Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV)); | |||
1012 | Replacement->takeName(ToReplace); | |||
1013 | // If the newly generated derived pointer's type does not match the original derived | |||
1014 | // pointer's type, cast the new derived pointer to match it. Same reasoning as above. | |||
1015 | Value *ActualReplacement = Replacement; | |||
1016 | if (Replacement->getType() != ToReplace->getType()) { | |||
1017 | ActualReplacement = | |||
1018 | Builder.CreateBitCast(Replacement, ToReplace->getType()); | |||
1019 | } | |||
1020 | ToReplace->replaceAllUsesWith(ActualReplacement); | |||
1021 | ToReplace->eraseFromParent(); | |||
1022 | ||||
1023 | MadeChange = true; | |||
1024 | } | |||
1025 | return MadeChange; | |||
1026 | } | |||
1027 | ||||
1028 | // Turns this: | |||
1029 | // | |||
1030 | // %base = ... | |||
1031 | // %ptr = gep %base + 15 | |||
1032 | // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr) | |||
1033 | // %base' = relocate(%tok, i32 4, i32 4) | |||
1034 | // %ptr' = relocate(%tok, i32 4, i32 5) | |||
1035 | // %val = load %ptr' | |||
1036 | // | |||
1037 | // into this: | |||
1038 | // | |||
1039 | // %base = ... | |||
1040 | // %ptr = gep %base + 15 | |||
1041 | // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr) | |||
1042 | // %base' = gc.relocate(%tok, i32 4, i32 4) | |||
1043 | // %ptr' = gep %base' + 15 | |||
1044 | // %val = load %ptr' | |||
1045 | bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) { | |||
1046 | bool MadeChange = false; | |||
1047 | SmallVector<GCRelocateInst *, 2> AllRelocateCalls; | |||
1048 | ||||
1049 | for (auto *U : I.users()) | |||
1050 | if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U)) | |||
1051 | // Collect all the relocate calls associated with a statepoint | |||
1052 | AllRelocateCalls.push_back(Relocate); | |||
1053 | ||||
1054 | // We need atleast one base pointer relocation + one derived pointer | |||
1055 | // relocation to mangle | |||
1056 | if (AllRelocateCalls.size() < 2) | |||
1057 | return false; | |||
1058 | ||||
1059 | // RelocateInstMap is a mapping from the base relocate instruction to the | |||
1060 | // corresponding derived relocate instructions | |||
1061 | DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap; | |||
1062 | computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap); | |||
1063 | if (RelocateInstMap.empty()) | |||
1064 | return false; | |||
1065 | ||||
1066 | for (auto &Item : RelocateInstMap) | |||
1067 | // Item.first is the RelocatedBase to offset against | |||
1068 | // Item.second is the vector of Targets to replace | |||
1069 | MadeChange = simplifyRelocatesOffABase(Item.first, Item.second); | |||
1070 | return MadeChange; | |||
1071 | } | |||
1072 | ||||
1073 | /// SinkCast - Sink the specified cast instruction into its user blocks | |||
1074 | static bool SinkCast(CastInst *CI) { | |||
1075 | BasicBlock *DefBB = CI->getParent(); | |||
1076 | ||||
1077 | /// InsertedCasts - Only insert a cast in each block once. | |||
1078 | DenseMap<BasicBlock*, CastInst*> InsertedCasts; | |||
1079 | ||||
1080 | bool MadeChange = false; | |||
1081 | for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end(); | |||
1082 | UI != E; ) { | |||
1083 | Use &TheUse = UI.getUse(); | |||
1084 | Instruction *User = cast<Instruction>(*UI); | |||
1085 | ||||
1086 | // Figure out which BB this cast is used in. For PHI's this is the | |||
1087 | // appropriate predecessor block. | |||
1088 | BasicBlock *UserBB = User->getParent(); | |||
1089 | if (PHINode *PN = dyn_cast<PHINode>(User)) { | |||
1090 | UserBB = PN->getIncomingBlock(TheUse); | |||
1091 | } | |||
1092 | ||||
1093 | // Preincrement use iterator so we don't invalidate it. | |||
1094 | ++UI; | |||
1095 | ||||
1096 | // The first insertion point of a block containing an EH pad is after the | |||
1097 | // pad. If the pad is the user, we cannot sink the cast past the pad. | |||
1098 | if (User->isEHPad()) | |||
1099 | continue; | |||
1100 | ||||
1101 | // If the block selected to receive the cast is an EH pad that does not | |||
1102 | // allow non-PHI instructions before the terminator, we can't sink the | |||
1103 | // cast. | |||
1104 | if (UserBB->getTerminator()->isEHPad()) | |||
1105 | continue; | |||
1106 | ||||
1107 | // If this user is in the same block as the cast, don't change the cast. | |||
1108 | if (UserBB == DefBB) continue; | |||
1109 | ||||
1110 | // If we have already inserted a cast into this block, use it. | |||
1111 | CastInst *&InsertedCast = InsertedCasts[UserBB]; | |||
1112 | ||||
1113 | if (!InsertedCast) { | |||
1114 | BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); | |||
1115 | assert(InsertPt != UserBB->end())((InsertPt != UserBB->end()) ? static_cast<void> (0) : __assert_fail ("InsertPt != UserBB->end()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 1115, __PRETTY_FUNCTION__)); | |||
1116 | InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0), | |||
1117 | CI->getType(), "", &*InsertPt); | |||
1118 | } | |||
1119 | ||||
1120 | // Replace a use of the cast with a use of the new cast. | |||
1121 | TheUse = InsertedCast; | |||
1122 | MadeChange = true; | |||
1123 | ++NumCastUses; | |||
1124 | } | |||
1125 | ||||
1126 | // If we removed all uses, nuke the cast. | |||
1127 | if (CI->use_empty()) { | |||
1128 | CI->eraseFromParent(); | |||
1129 | MadeChange = true; | |||
1130 | } | |||
1131 | ||||
1132 | return MadeChange; | |||
1133 | } | |||
1134 | ||||
1135 | /// If the specified cast instruction is a noop copy (e.g. it's casting from | |||
1136 | /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to | |||
1137 | /// reduce the number of virtual registers that must be created and coalesced. | |||
1138 | /// | |||
1139 | /// Return true if any changes are made. | |||
1140 | /// | |||
1141 | static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI, | |||
1142 | const DataLayout &DL) { | |||
1143 | // Sink only "cheap" (or nop) address-space casts. This is a weaker condition | |||
1144 | // than sinking only nop casts, but is helpful on some platforms. | |||
1145 | if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) { | |||
1146 | if (!TLI.isCheapAddrSpaceCast(ASC->getSrcAddressSpace(), | |||
1147 | ASC->getDestAddressSpace())) | |||
1148 | return false; | |||
1149 | } | |||
1150 | ||||
1151 | // If this is a noop copy, | |||
1152 | EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType()); | |||
1153 | EVT DstVT = TLI.getValueType(DL, CI->getType()); | |||
1154 | ||||
1155 | // This is an fp<->int conversion? | |||
1156 | if (SrcVT.isInteger() != DstVT.isInteger()) | |||
1157 | return false; | |||
1158 | ||||
1159 | // If this is an extension, it will be a zero or sign extension, which | |||
1160 | // isn't a noop. | |||
1161 | if (SrcVT.bitsLT(DstVT)) return false; | |||
1162 | ||||
1163 | // If these values will be promoted, find out what they will be promoted | |||
1164 | // to. This helps us consider truncates on PPC as noop copies when they | |||
1165 | // are. | |||
1166 | if (TLI.getTypeAction(CI->getContext(), SrcVT) == | |||
1167 | TargetLowering::TypePromoteInteger) | |||
1168 | SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT); | |||
1169 | if (TLI.getTypeAction(CI->getContext(), DstVT) == | |||
1170 | TargetLowering::TypePromoteInteger) | |||
1171 | DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT); | |||
1172 | ||||
1173 | // If, after promotion, these are the same types, this is a noop copy. | |||
1174 | if (SrcVT != DstVT) | |||
1175 | return false; | |||
1176 | ||||
1177 | return SinkCast(CI); | |||
1178 | } | |||
1179 | ||||
1180 | /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if | |||
1181 | /// possible. | |||
1182 | /// | |||
1183 | /// Return true if any changes were made. | |||
1184 | static bool CombineUAddWithOverflow(CmpInst *CI) { | |||
1185 | Value *A, *B; | |||
1186 | Instruction *AddI; | |||
1187 | if (!match(CI, | |||
1188 | m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI)))) | |||
1189 | return false; | |||
1190 | ||||
1191 | Type *Ty = AddI->getType(); | |||
1192 | if (!isa<IntegerType>(Ty)) | |||
1193 | return false; | |||
1194 | ||||
1195 | // We don't want to move around uses of condition values this late, so we we | |||
1196 | // check if it is legal to create the call to the intrinsic in the basic | |||
1197 | // block containing the icmp: | |||
1198 | ||||
1199 | if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse()) | |||
1200 | return false; | |||
1201 | ||||
1202 | #ifndef NDEBUG | |||
1203 | // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption | |||
1204 | // for now: | |||
1205 | if (AddI->hasOneUse()) | |||
1206 | assert(*AddI->user_begin() == CI && "expected!")((*AddI->user_begin() == CI && "expected!") ? static_cast <void> (0) : __assert_fail ("*AddI->user_begin() == CI && \"expected!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 1206, __PRETTY_FUNCTION__)); | |||
1207 | #endif | |||
1208 | ||||
1209 | Module *M = CI->getModule(); | |||
1210 | Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty); | |||
1211 | ||||
1212 | auto *InsertPt = AddI->hasOneUse() ? CI : AddI; | |||
1213 | ||||
1214 | auto *UAddWithOverflow = | |||
1215 | CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt); | |||
1216 | auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt); | |||
1217 | auto *Overflow = | |||
1218 | ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt); | |||
1219 | ||||
1220 | CI->replaceAllUsesWith(Overflow); | |||
1221 | AddI->replaceAllUsesWith(UAdd); | |||
1222 | CI->eraseFromParent(); | |||
1223 | AddI->eraseFromParent(); | |||
1224 | return true; | |||
1225 | } | |||
1226 | ||||
1227 | /// Sink the given CmpInst into user blocks to reduce the number of virtual | |||
1228 | /// registers that must be created and coalesced. This is a clear win except on | |||
1229 | /// targets with multiple condition code registers (PowerPC), where it might | |||
1230 | /// lose; some adjustment may be wanted there. | |||
1231 | /// | |||
1232 | /// Return true if any changes are made. | |||
1233 | static bool SinkCmpExpression(CmpInst *CI, const TargetLowering *TLI) { | |||
1234 | BasicBlock *DefBB = CI->getParent(); | |||
1235 | ||||
1236 | // Avoid sinking soft-FP comparisons, since this can move them into a loop. | |||
1237 | if (TLI && TLI->useSoftFloat() && isa<FCmpInst>(CI)) | |||
1238 | return false; | |||
1239 | ||||
1240 | // Only insert a cmp in each block once. | |||
1241 | DenseMap<BasicBlock*, CmpInst*> InsertedCmps; | |||
1242 | ||||
1243 | bool MadeChange = false; | |||
1244 | for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end(); | |||
1245 | UI != E; ) { | |||
1246 | Use &TheUse = UI.getUse(); | |||
1247 | Instruction *User = cast<Instruction>(*UI); | |||
1248 | ||||
1249 | // Preincrement use iterator so we don't invalidate it. | |||
1250 | ++UI; | |||
1251 | ||||
1252 | // Don't bother for PHI nodes. | |||
1253 | if (isa<PHINode>(User)) | |||
1254 | continue; | |||
1255 | ||||
1256 | // Figure out which BB this cmp is used in. | |||
1257 | BasicBlock *UserBB = User->getParent(); | |||
1258 | ||||
1259 | // If this user is in the same block as the cmp, don't change the cmp. | |||
1260 | if (UserBB == DefBB) continue; | |||
1261 | ||||
1262 | // If we have already inserted a cmp into this block, use it. | |||
1263 | CmpInst *&InsertedCmp = InsertedCmps[UserBB]; | |||
1264 | ||||
1265 | if (!InsertedCmp) { | |||
1266 | BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); | |||
1267 | assert(InsertPt != UserBB->end())((InsertPt != UserBB->end()) ? static_cast<void> (0) : __assert_fail ("InsertPt != UserBB->end()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 1267, __PRETTY_FUNCTION__)); | |||
1268 | InsertedCmp = | |||
1269 | CmpInst::Create(CI->getOpcode(), CI->getPredicate(), | |||
1270 | CI->getOperand(0), CI->getOperand(1), "", &*InsertPt); | |||
1271 | // Propagate the debug info. | |||
1272 | InsertedCmp->setDebugLoc(CI->getDebugLoc()); | |||
1273 | } | |||
1274 | ||||
1275 | // Replace a use of the cmp with a use of the new cmp. | |||
1276 | TheUse = InsertedCmp; | |||
1277 | MadeChange = true; | |||
1278 | ++NumCmpUses; | |||
1279 | } | |||
1280 | ||||
1281 | // If we removed all uses, nuke the cmp. | |||
1282 | if (CI->use_empty()) { | |||
1283 | CI->eraseFromParent(); | |||
1284 | MadeChange = true; | |||
1285 | } | |||
1286 | ||||
1287 | return MadeChange; | |||
1288 | } | |||
1289 | ||||
1290 | static bool OptimizeCmpExpression(CmpInst *CI, const TargetLowering *TLI) { | |||
1291 | if (SinkCmpExpression(CI, TLI)) | |||
1292 | return true; | |||
1293 | ||||
1294 | if (CombineUAddWithOverflow(CI)) | |||
1295 | return true; | |||
1296 | ||||
1297 | return false; | |||
1298 | } | |||
1299 | ||||
1300 | /// Duplicate and sink the given 'and' instruction into user blocks where it is | |||
1301 | /// used in a compare to allow isel to generate better code for targets where | |||
1302 | /// this operation can be combined. | |||
1303 | /// | |||
1304 | /// Return true if any changes are made. | |||
1305 | static bool sinkAndCmp0Expression(Instruction *AndI, | |||
1306 | const TargetLowering &TLI, | |||
1307 | SetOfInstrs &InsertedInsts) { | |||
1308 | // Double-check that we're not trying to optimize an instruction that was | |||
1309 | // already optimized by some other part of this pass. | |||
1310 | assert(!InsertedInsts.count(AndI) &&((!InsertedInsts.count(AndI) && "Attempting to optimize already optimized and instruction" ) ? static_cast<void> (0) : __assert_fail ("!InsertedInsts.count(AndI) && \"Attempting to optimize already optimized and instruction\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 1311, __PRETTY_FUNCTION__)) | |||
1311 | "Attempting to optimize already optimized and instruction")((!InsertedInsts.count(AndI) && "Attempting to optimize already optimized and instruction" ) ? static_cast<void> (0) : __assert_fail ("!InsertedInsts.count(AndI) && \"Attempting to optimize already optimized and instruction\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 1311, __PRETTY_FUNCTION__)); | |||
1312 | (void) InsertedInsts; | |||
1313 | ||||
1314 | // Nothing to do for single use in same basic block. | |||
1315 | if (AndI->hasOneUse() && | |||
1316 | AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent()) | |||
1317 | return false; | |||
1318 | ||||
1319 | // Try to avoid cases where sinking/duplicating is likely to increase register | |||
1320 | // pressure. | |||
1321 | if (!isa<ConstantInt>(AndI->getOperand(0)) && | |||
1322 | !isa<ConstantInt>(AndI->getOperand(1)) && | |||
1323 | AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse()) | |||
1324 | return false; | |||
1325 | ||||
1326 | for (auto *U : AndI->users()) { | |||
1327 | Instruction *User = cast<Instruction>(U); | |||
1328 | ||||
1329 | // Only sink for and mask feeding icmp with 0. | |||
1330 | if (!isa<ICmpInst>(User)) | |||
1331 | return false; | |||
1332 | ||||
1333 | auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1)); | |||
1334 | if (!CmpC || !CmpC->isZero()) | |||
1335 | return false; | |||
1336 | } | |||
1337 | ||||
1338 | if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI)) | |||
1339 | return false; | |||
1340 | ||||
1341 | DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "found 'and' feeding only icmp 0;\n" ; } } while (false); | |||
1342 | DEBUG(AndI->getParent()->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { AndI->getParent()->dump(); } } while (false); | |||
1343 | ||||
1344 | // Push the 'and' into the same block as the icmp 0. There should only be | |||
1345 | // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any | |||
1346 | // others, so we don't need to keep track of which BBs we insert into. | |||
1347 | for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end(); | |||
1348 | UI != E; ) { | |||
1349 | Use &TheUse = UI.getUse(); | |||
1350 | Instruction *User = cast<Instruction>(*UI); | |||
1351 | ||||
1352 | // Preincrement use iterator so we don't invalidate it. | |||
1353 | ++UI; | |||
1354 | ||||
1355 | DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "sinking 'and' use: " << *User << "\n"; } } while (false); | |||
1356 | ||||
1357 | // Keep the 'and' in the same place if the use is already in the same block. | |||
1358 | Instruction *InsertPt = | |||
1359 | User->getParent() == AndI->getParent() ? AndI : User; | |||
1360 | Instruction *InsertedAnd = | |||
1361 | BinaryOperator::Create(Instruction::And, AndI->getOperand(0), | |||
1362 | AndI->getOperand(1), "", InsertPt); | |||
1363 | // Propagate the debug info. | |||
1364 | InsertedAnd->setDebugLoc(AndI->getDebugLoc()); | |||
1365 | ||||
1366 | // Replace a use of the 'and' with a use of the new 'and'. | |||
1367 | TheUse = InsertedAnd; | |||
1368 | ++NumAndUses; | |||
1369 | DEBUG(User->getParent()->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { User->getParent()->dump(); } } while (false); | |||
1370 | } | |||
1371 | ||||
1372 | // We removed all uses, nuke the and. | |||
1373 | AndI->eraseFromParent(); | |||
1374 | return true; | |||
1375 | } | |||
1376 | ||||
1377 | /// Check if the candidates could be combined with a shift instruction, which | |||
1378 | /// includes: | |||
1379 | /// 1. Truncate instruction | |||
1380 | /// 2. And instruction and the imm is a mask of the low bits: | |||
1381 | /// imm & (imm+1) == 0 | |||
1382 | static bool isExtractBitsCandidateUse(Instruction *User) { | |||
1383 | if (!isa<TruncInst>(User)) { | |||
1384 | if (User->getOpcode() != Instruction::And || | |||
1385 | !isa<ConstantInt>(User->getOperand(1))) | |||
1386 | return false; | |||
1387 | ||||
1388 | const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue(); | |||
1389 | ||||
1390 | if ((Cimm & (Cimm + 1)).getBoolValue()) | |||
1391 | return false; | |||
1392 | } | |||
1393 | return true; | |||
1394 | } | |||
1395 | ||||
1396 | /// Sink both shift and truncate instruction to the use of truncate's BB. | |||
1397 | static bool | |||
1398 | SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI, | |||
1399 | DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts, | |||
1400 | const TargetLowering &TLI, const DataLayout &DL) { | |||
1401 | BasicBlock *UserBB = User->getParent(); | |||
1402 | DenseMap<BasicBlock *, CastInst *> InsertedTruncs; | |||
1403 | TruncInst *TruncI = dyn_cast<TruncInst>(User); | |||
1404 | bool MadeChange = false; | |||
1405 | ||||
1406 | for (Value::user_iterator TruncUI = TruncI->user_begin(), | |||
1407 | TruncE = TruncI->user_end(); | |||
1408 | TruncUI != TruncE;) { | |||
1409 | ||||
1410 | Use &TruncTheUse = TruncUI.getUse(); | |||
1411 | Instruction *TruncUser = cast<Instruction>(*TruncUI); | |||
1412 | // Preincrement use iterator so we don't invalidate it. | |||
1413 | ||||
1414 | ++TruncUI; | |||
1415 | ||||
1416 | int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode()); | |||
1417 | if (!ISDOpcode) | |||
1418 | continue; | |||
1419 | ||||
1420 | // If the use is actually a legal node, there will not be an | |||
1421 | // implicit truncate. | |||
1422 | // FIXME: always querying the result type is just an | |||
1423 | // approximation; some nodes' legality is determined by the | |||
1424 | // operand or other means. There's no good way to find out though. | |||
1425 | if (TLI.isOperationLegalOrCustom( | |||
1426 | ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true))) | |||
1427 | continue; | |||
1428 | ||||
1429 | // Don't bother for PHI nodes. | |||
1430 | if (isa<PHINode>(TruncUser)) | |||
1431 | continue; | |||
1432 | ||||
1433 | BasicBlock *TruncUserBB = TruncUser->getParent(); | |||
1434 | ||||
1435 | if (UserBB == TruncUserBB) | |||
1436 | continue; | |||
1437 | ||||
1438 | BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB]; | |||
1439 | CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB]; | |||
1440 | ||||
1441 | if (!InsertedShift && !InsertedTrunc) { | |||
1442 | BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt(); | |||
1443 | assert(InsertPt != TruncUserBB->end())((InsertPt != TruncUserBB->end()) ? static_cast<void> (0) : __assert_fail ("InsertPt != TruncUserBB->end()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 1443, __PRETTY_FUNCTION__)); | |||
1444 | // Sink the shift | |||
1445 | if (ShiftI->getOpcode() == Instruction::AShr) | |||
1446 | InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, | |||
1447 | "", &*InsertPt); | |||
1448 | else | |||
1449 | InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, | |||
1450 | "", &*InsertPt); | |||
1451 | ||||
1452 | // Sink the trunc | |||
1453 | BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt(); | |||
1454 | TruncInsertPt++; | |||
1455 | assert(TruncInsertPt != TruncUserBB->end())((TruncInsertPt != TruncUserBB->end()) ? static_cast<void > (0) : __assert_fail ("TruncInsertPt != TruncUserBB->end()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 1455, __PRETTY_FUNCTION__)); | |||
1456 | ||||
1457 | InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift, | |||
1458 | TruncI->getType(), "", &*TruncInsertPt); | |||
1459 | ||||
1460 | MadeChange = true; | |||
1461 | ||||
1462 | TruncTheUse = InsertedTrunc; | |||
1463 | } | |||
1464 | } | |||
1465 | return MadeChange; | |||
1466 | } | |||
1467 | ||||
1468 | /// Sink the shift *right* instruction into user blocks if the uses could | |||
1469 | /// potentially be combined with this shift instruction and generate BitExtract | |||
1470 | /// instruction. It will only be applied if the architecture supports BitExtract | |||
1471 | /// instruction. Here is an example: | |||
1472 | /// BB1: | |||
1473 | /// %x.extract.shift = lshr i64 %arg1, 32 | |||
1474 | /// BB2: | |||
1475 | /// %x.extract.trunc = trunc i64 %x.extract.shift to i16 | |||
1476 | /// ==> | |||
1477 | /// | |||
1478 | /// BB2: | |||
1479 | /// %x.extract.shift.1 = lshr i64 %arg1, 32 | |||
1480 | /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16 | |||
1481 | /// | |||
1482 | /// CodeGen will recoginze the pattern in BB2 and generate BitExtract | |||
1483 | /// instruction. | |||
1484 | /// Return true if any changes are made. | |||
1485 | static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI, | |||
1486 | const TargetLowering &TLI, | |||
1487 | const DataLayout &DL) { | |||
1488 | BasicBlock *DefBB = ShiftI->getParent(); | |||
1489 | ||||
1490 | /// Only insert instructions in each block once. | |||
1491 | DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts; | |||
1492 | ||||
1493 | bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType())); | |||
1494 | ||||
1495 | bool MadeChange = false; | |||
1496 | for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end(); | |||
1497 | UI != E;) { | |||
1498 | Use &TheUse = UI.getUse(); | |||
1499 | Instruction *User = cast<Instruction>(*UI); | |||
1500 | // Preincrement use iterator so we don't invalidate it. | |||
1501 | ++UI; | |||
1502 | ||||
1503 | // Don't bother for PHI nodes. | |||
1504 | if (isa<PHINode>(User)) | |||
1505 | continue; | |||
1506 | ||||
1507 | if (!isExtractBitsCandidateUse(User)) | |||
1508 | continue; | |||
1509 | ||||
1510 | BasicBlock *UserBB = User->getParent(); | |||
1511 | ||||
1512 | if (UserBB == DefBB) { | |||
1513 | // If the shift and truncate instruction are in the same BB. The use of | |||
1514 | // the truncate(TruncUse) may still introduce another truncate if not | |||
1515 | // legal. In this case, we would like to sink both shift and truncate | |||
1516 | // instruction to the BB of TruncUse. | |||
1517 | // for example: | |||
1518 | // BB1: | |||
1519 | // i64 shift.result = lshr i64 opnd, imm | |||
1520 | // trunc.result = trunc shift.result to i16 | |||
1521 | // | |||
1522 | // BB2: | |||
1523 | // ----> We will have an implicit truncate here if the architecture does | |||
1524 | // not have i16 compare. | |||
1525 | // cmp i16 trunc.result, opnd2 | |||
1526 | // | |||
1527 | if (isa<TruncInst>(User) && shiftIsLegal | |||
1528 | // If the type of the truncate is legal, no trucate will be | |||
1529 | // introduced in other basic blocks. | |||
1530 | && | |||
1531 | (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType())))) | |||
1532 | MadeChange = | |||
1533 | SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL); | |||
1534 | ||||
1535 | continue; | |||
1536 | } | |||
1537 | // If we have already inserted a shift into this block, use it. | |||
1538 | BinaryOperator *&InsertedShift = InsertedShifts[UserBB]; | |||
1539 | ||||
1540 | if (!InsertedShift) { | |||
1541 | BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); | |||
1542 | assert(InsertPt != UserBB->end())((InsertPt != UserBB->end()) ? static_cast<void> (0) : __assert_fail ("InsertPt != UserBB->end()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 1542, __PRETTY_FUNCTION__)); | |||
1543 | ||||
1544 | if (ShiftI->getOpcode() == Instruction::AShr) | |||
1545 | InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, | |||
1546 | "", &*InsertPt); | |||
1547 | else | |||
1548 | InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, | |||
1549 | "", &*InsertPt); | |||
1550 | ||||
1551 | MadeChange = true; | |||
1552 | } | |||
1553 | ||||
1554 | // Replace a use of the shift with a use of the new shift. | |||
1555 | TheUse = InsertedShift; | |||
1556 | } | |||
1557 | ||||
1558 | // If we removed all uses, nuke the shift. | |||
1559 | if (ShiftI->use_empty()) | |||
1560 | ShiftI->eraseFromParent(); | |||
1561 | ||||
1562 | return MadeChange; | |||
1563 | } | |||
1564 | ||||
1565 | /// If counting leading or trailing zeros is an expensive operation and a zero | |||
1566 | /// input is defined, add a check for zero to avoid calling the intrinsic. | |||
1567 | /// | |||
1568 | /// We want to transform: | |||
1569 | /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false) | |||
1570 | /// | |||
1571 | /// into: | |||
1572 | /// entry: | |||
1573 | /// %cmpz = icmp eq i64 %A, 0 | |||
1574 | /// br i1 %cmpz, label %cond.end, label %cond.false | |||
1575 | /// cond.false: | |||
1576 | /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true) | |||
1577 | /// br label %cond.end | |||
1578 | /// cond.end: | |||
1579 | /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ] | |||
1580 | /// | |||
1581 | /// If the transform is performed, return true and set ModifiedDT to true. | |||
1582 | static bool despeculateCountZeros(IntrinsicInst *CountZeros, | |||
1583 | const TargetLowering *TLI, | |||
1584 | const DataLayout *DL, | |||
1585 | bool &ModifiedDT) { | |||
1586 | if (!TLI || !DL) | |||
1587 | return false; | |||
1588 | ||||
1589 | // If a zero input is undefined, it doesn't make sense to despeculate that. | |||
1590 | if (match(CountZeros->getOperand(1), m_One())) | |||
1591 | return false; | |||
1592 | ||||
1593 | // If it's cheap to speculate, there's nothing to do. | |||
1594 | auto IntrinsicID = CountZeros->getIntrinsicID(); | |||
1595 | if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) || | |||
1596 | (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz())) | |||
1597 | return false; | |||
1598 | ||||
1599 | // Only handle legal scalar cases. Anything else requires too much work. | |||
1600 | Type *Ty = CountZeros->getType(); | |||
1601 | unsigned SizeInBits = Ty->getPrimitiveSizeInBits(); | |||
1602 | if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits()) | |||
1603 | return false; | |||
1604 | ||||
1605 | // The intrinsic will be sunk behind a compare against zero and branch. | |||
1606 | BasicBlock *StartBlock = CountZeros->getParent(); | |||
1607 | BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false"); | |||
1608 | ||||
1609 | // Create another block after the count zero intrinsic. A PHI will be added | |||
1610 | // in this block to select the result of the intrinsic or the bit-width | |||
1611 | // constant if the input to the intrinsic is zero. | |||
1612 | BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros)); | |||
1613 | BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end"); | |||
1614 | ||||
1615 | // Set up a builder to create a compare, conditional branch, and PHI. | |||
1616 | IRBuilder<> Builder(CountZeros->getContext()); | |||
1617 | Builder.SetInsertPoint(StartBlock->getTerminator()); | |||
1618 | Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc()); | |||
1619 | ||||
1620 | // Replace the unconditional branch that was created by the first split with | |||
1621 | // a compare against zero and a conditional branch. | |||
1622 | Value *Zero = Constant::getNullValue(Ty); | |||
1623 | Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz"); | |||
1624 | Builder.CreateCondBr(Cmp, EndBlock, CallBlock); | |||
1625 | StartBlock->getTerminator()->eraseFromParent(); | |||
1626 | ||||
1627 | // Create a PHI in the end block to select either the output of the intrinsic | |||
1628 | // or the bit width of the operand. | |||
1629 | Builder.SetInsertPoint(&EndBlock->front()); | |||
1630 | PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz"); | |||
1631 | CountZeros->replaceAllUsesWith(PN); | |||
1632 | Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits)); | |||
1633 | PN->addIncoming(BitWidth, StartBlock); | |||
1634 | PN->addIncoming(CountZeros, CallBlock); | |||
1635 | ||||
1636 | // We are explicitly handling the zero case, so we can set the intrinsic's | |||
1637 | // undefined zero argument to 'true'. This will also prevent reprocessing the | |||
1638 | // intrinsic; we only despeculate when a zero input is defined. | |||
1639 | CountZeros->setArgOperand(1, Builder.getTrue()); | |||
1640 | ModifiedDT = true; | |||
1641 | return true; | |||
1642 | } | |||
1643 | ||||
1644 | // This class provides helper functions to expand a memcmp library call into an | |||
1645 | // inline expansion. | |||
1646 | class MemCmpExpansion { | |||
1647 | struct ResultBlock { | |||
1648 | BasicBlock *BB; | |||
1649 | PHINode *PhiSrc1; | |||
1650 | PHINode *PhiSrc2; | |||
1651 | ResultBlock(); | |||
1652 | }; | |||
1653 | ||||
1654 | CallInst *CI; | |||
1655 | ResultBlock ResBlock; | |||
1656 | unsigned MaxLoadSize; | |||
1657 | unsigned NumBlocks; | |||
1658 | unsigned NumBlocksNonOneByte; | |||
1659 | unsigned NumLoadsPerBlock; | |||
1660 | std::vector<BasicBlock *> LoadCmpBlocks; | |||
1661 | BasicBlock *EndBlock; | |||
1662 | PHINode *PhiRes; | |||
1663 | bool IsUsedForZeroCmp; | |||
1664 | const DataLayout &DL; | |||
1665 | IRBuilder<> Builder; | |||
1666 | ||||
1667 | unsigned calculateNumBlocks(unsigned Size); | |||
1668 | void createLoadCmpBlocks(); | |||
1669 | void createResultBlock(); | |||
1670 | void setupResultBlockPHINodes(); | |||
1671 | void setupEndBlockPHINodes(); | |||
1672 | void emitLoadCompareBlock(unsigned Index, unsigned LoadSize, | |||
1673 | unsigned GEPIndex); | |||
1674 | Value *getCompareLoadPairs(unsigned Index, unsigned Size, | |||
1675 | unsigned &NumBytesProcessed); | |||
1676 | void emitLoadCompareBlockMultipleLoads(unsigned Index, unsigned Size, | |||
1677 | unsigned &NumBytesProcessed); | |||
1678 | void emitLoadCompareByteBlock(unsigned Index, unsigned GEPIndex); | |||
1679 | void emitMemCmpResultBlock(); | |||
1680 | Value *getMemCmpExpansionZeroCase(unsigned Size); | |||
1681 | Value *getMemCmpEqZeroOneBlock(unsigned Size); | |||
1682 | unsigned getLoadSize(unsigned Size); | |||
1683 | unsigned getNumLoads(unsigned Size); | |||
1684 | ||||
1685 | public: | |||
1686 | MemCmpExpansion(CallInst *CI, uint64_t Size, unsigned MaxLoadSize, | |||
1687 | unsigned NumLoadsPerBlock, const DataLayout &DL); | |||
1688 | Value *getMemCmpExpansion(uint64_t Size); | |||
1689 | }; | |||
1690 | ||||
1691 | MemCmpExpansion::ResultBlock::ResultBlock() | |||
1692 | : BB(nullptr), PhiSrc1(nullptr), PhiSrc2(nullptr) {} | |||
1693 | ||||
1694 | // Initialize the basic block structure required for expansion of memcmp call | |||
1695 | // with given maximum load size and memcmp size parameter. | |||
1696 | // This structure includes: | |||
1697 | // 1. A list of load compare blocks - LoadCmpBlocks. | |||
1698 | // 2. An EndBlock, split from original instruction point, which is the block to | |||
1699 | // return from. | |||
1700 | // 3. ResultBlock, block to branch to for early exit when a | |||
1701 | // LoadCmpBlock finds a difference. | |||
1702 | MemCmpExpansion::MemCmpExpansion(CallInst *CI, uint64_t Size, | |||
1703 | unsigned MaxLoadSize, unsigned LoadsPerBlock, | |||
1704 | const DataLayout &TheDataLayout) | |||
1705 | : CI(CI), MaxLoadSize(MaxLoadSize), NumLoadsPerBlock(LoadsPerBlock), | |||
1706 | DL(TheDataLayout), Builder(CI) { | |||
1707 | ||||
1708 | // A memcmp with zero-comparison with only one block of load and compare does | |||
1709 | // not need to set up any extra blocks. This case could be handled in the DAG, | |||
1710 | // but since we have all of the machinery to flexibly expand any memcpy here, | |||
1711 | // we choose to handle this case too to avoid fragmented lowering. | |||
1712 | IsUsedForZeroCmp = isOnlyUsedInZeroEqualityComparison(CI); | |||
1713 | NumBlocks = calculateNumBlocks(Size); | |||
1714 | if (!IsUsedForZeroCmp || NumBlocks != 1) { | |||
1715 | BasicBlock *StartBlock = CI->getParent(); | |||
1716 | EndBlock = StartBlock->splitBasicBlock(CI, "endblock"); | |||
1717 | setupEndBlockPHINodes(); | |||
1718 | createResultBlock(); | |||
1719 | ||||
1720 | // If return value of memcmp is not used in a zero equality, we need to | |||
1721 | // calculate which source was larger. The calculation requires the | |||
1722 | // two loaded source values of each load compare block. | |||
1723 | // These will be saved in the phi nodes created by setupResultBlockPHINodes. | |||
1724 | if (!IsUsedForZeroCmp) | |||
1725 | setupResultBlockPHINodes(); | |||
1726 | ||||
1727 | // Create the number of required load compare basic blocks. | |||
1728 | createLoadCmpBlocks(); | |||
1729 | ||||
1730 | // Update the terminator added by splitBasicBlock to branch to the first | |||
1731 | // LoadCmpBlock. | |||
1732 | StartBlock->getTerminator()->setSuccessor(0, LoadCmpBlocks[0]); | |||
1733 | } | |||
1734 | ||||
1735 | Builder.SetCurrentDebugLocation(CI->getDebugLoc()); | |||
1736 | } | |||
1737 | ||||
1738 | void MemCmpExpansion::createLoadCmpBlocks() { | |||
1739 | for (unsigned i = 0; i < NumBlocks; i++) { | |||
1740 | BasicBlock *BB = BasicBlock::Create(CI->getContext(), "loadbb", | |||
1741 | EndBlock->getParent(), EndBlock); | |||
1742 | LoadCmpBlocks.push_back(BB); | |||
1743 | } | |||
1744 | } | |||
1745 | ||||
1746 | void MemCmpExpansion::createResultBlock() { | |||
1747 | ResBlock.BB = BasicBlock::Create(CI->getContext(), "res_block", | |||
1748 | EndBlock->getParent(), EndBlock); | |||
1749 | } | |||
1750 | ||||
1751 | // This function creates the IR instructions for loading and comparing 1 byte. | |||
1752 | // It loads 1 byte from each source of the memcmp parameters with the given | |||
1753 | // GEPIndex. It then subtracts the two loaded values and adds this result to the | |||
1754 | // final phi node for selecting the memcmp result. | |||
1755 | void MemCmpExpansion::emitLoadCompareByteBlock(unsigned Index, | |||
1756 | unsigned GEPIndex) { | |||
1757 | Value *Source1 = CI->getArgOperand(0); | |||
1758 | Value *Source2 = CI->getArgOperand(1); | |||
1759 | ||||
1760 | Builder.SetInsertPoint(LoadCmpBlocks[Index]); | |||
1761 | Type *LoadSizeType = Type::getInt8Ty(CI->getContext()); | |||
1762 | // Cast source to LoadSizeType*. | |||
1763 | if (Source1->getType() != LoadSizeType) | |||
1764 | Source1 = Builder.CreateBitCast(Source1, LoadSizeType->getPointerTo()); | |||
1765 | if (Source2->getType() != LoadSizeType) | |||
1766 | Source2 = Builder.CreateBitCast(Source2, LoadSizeType->getPointerTo()); | |||
1767 | ||||
1768 | // Get the base address using the GEPIndex. | |||
1769 | if (GEPIndex != 0) { | |||
1770 | Source1 = Builder.CreateGEP(LoadSizeType, Source1, | |||
1771 | ConstantInt::get(LoadSizeType, GEPIndex)); | |||
1772 | Source2 = Builder.CreateGEP(LoadSizeType, Source2, | |||
1773 | ConstantInt::get(LoadSizeType, GEPIndex)); | |||
1774 | } | |||
1775 | ||||
1776 | Value *LoadSrc1 = Builder.CreateLoad(LoadSizeType, Source1); | |||
1777 | Value *LoadSrc2 = Builder.CreateLoad(LoadSizeType, Source2); | |||
1778 | ||||
1779 | LoadSrc1 = Builder.CreateZExt(LoadSrc1, Type::getInt32Ty(CI->getContext())); | |||
1780 | LoadSrc2 = Builder.CreateZExt(LoadSrc2, Type::getInt32Ty(CI->getContext())); | |||
1781 | Value *Diff = Builder.CreateSub(LoadSrc1, LoadSrc2); | |||
1782 | ||||
1783 | PhiRes->addIncoming(Diff, LoadCmpBlocks[Index]); | |||
1784 | ||||
1785 | if (Index < (LoadCmpBlocks.size() - 1)) { | |||
1786 | // Early exit branch if difference found to EndBlock. Otherwise, continue to | |||
1787 | // next LoadCmpBlock, | |||
1788 | Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_NE, Diff, | |||
1789 | ConstantInt::get(Diff->getType(), 0)); | |||
1790 | BranchInst *CmpBr = | |||
1791 | BranchInst::Create(EndBlock, LoadCmpBlocks[Index + 1], Cmp); | |||
1792 | Builder.Insert(CmpBr); | |||
1793 | } else { | |||
1794 | // The last block has an unconditional branch to EndBlock. | |||
1795 | BranchInst *CmpBr = BranchInst::Create(EndBlock); | |||
1796 | Builder.Insert(CmpBr); | |||
1797 | } | |||
1798 | } | |||
1799 | ||||
1800 | unsigned MemCmpExpansion::getNumLoads(unsigned Size) { | |||
1801 | return (Size / MaxLoadSize) + countPopulation(Size % MaxLoadSize); | |||
1802 | } | |||
1803 | ||||
1804 | unsigned MemCmpExpansion::getLoadSize(unsigned Size) { | |||
1805 | return MinAlign(PowerOf2Floor(Size), MaxLoadSize); | |||
1806 | } | |||
1807 | ||||
1808 | /// Generate an equality comparison for one or more pairs of loaded values. | |||
1809 | /// This is used in the case where the memcmp() call is compared equal or not | |||
1810 | /// equal to zero. | |||
1811 | Value *MemCmpExpansion::getCompareLoadPairs(unsigned Index, unsigned Size, | |||
1812 | unsigned &NumBytesProcessed) { | |||
1813 | std::vector<Value *> XorList, OrList; | |||
1814 | Value *Diff; | |||
1815 | ||||
1816 | unsigned RemainingBytes = Size - NumBytesProcessed; | |||
1817 | unsigned NumLoadsRemaining = getNumLoads(RemainingBytes); | |||
1818 | unsigned NumLoads = std::min(NumLoadsRemaining, NumLoadsPerBlock); | |||
1819 | ||||
1820 | // For a single-block expansion, start inserting before the memcmp call. | |||
1821 | if (LoadCmpBlocks.empty()) | |||
1822 | Builder.SetInsertPoint(CI); | |||
1823 | else | |||
1824 | Builder.SetInsertPoint(LoadCmpBlocks[Index]); | |||
1825 | ||||
1826 | Value *Cmp = nullptr; | |||
1827 | for (unsigned i = 0; i < NumLoads; ++i) { | |||
1828 | unsigned LoadSize = getLoadSize(RemainingBytes); | |||
1829 | unsigned GEPIndex = NumBytesProcessed / LoadSize; | |||
1830 | NumBytesProcessed += LoadSize; | |||
1831 | RemainingBytes -= LoadSize; | |||
1832 | ||||
1833 | Type *LoadSizeType = IntegerType::get(CI->getContext(), LoadSize * 8); | |||
1834 | Type *MaxLoadType = IntegerType::get(CI->getContext(), MaxLoadSize * 8); | |||
1835 | assert(LoadSize <= MaxLoadSize && "Unexpected load type")((LoadSize <= MaxLoadSize && "Unexpected load type" ) ? static_cast<void> (0) : __assert_fail ("LoadSize <= MaxLoadSize && \"Unexpected load type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 1835, __PRETTY_FUNCTION__)); | |||
1836 | ||||
1837 | Value *Source1 = CI->getArgOperand(0); | |||
1838 | Value *Source2 = CI->getArgOperand(1); | |||
1839 | ||||
1840 | // Cast source to LoadSizeType*. | |||
1841 | if (Source1->getType() != LoadSizeType) | |||
1842 | Source1 = Builder.CreateBitCast(Source1, LoadSizeType->getPointerTo()); | |||
1843 | if (Source2->getType() != LoadSizeType) | |||
1844 | Source2 = Builder.CreateBitCast(Source2, LoadSizeType->getPointerTo()); | |||
1845 | ||||
1846 | // Get the base address using the GEPIndex. | |||
1847 | if (GEPIndex != 0) { | |||
1848 | Source1 = Builder.CreateGEP(LoadSizeType, Source1, | |||
1849 | ConstantInt::get(LoadSizeType, GEPIndex)); | |||
1850 | Source2 = Builder.CreateGEP(LoadSizeType, Source2, | |||
1851 | ConstantInt::get(LoadSizeType, GEPIndex)); | |||
1852 | } | |||
1853 | ||||
1854 | // Get a constant or load a value for each source address. | |||
1855 | Value *LoadSrc1 = nullptr; | |||
1856 | if (auto *Source1C = dyn_cast<Constant>(Source1)) | |||
1857 | LoadSrc1 = ConstantFoldLoadFromConstPtr(Source1C, LoadSizeType, DL); | |||
1858 | if (!LoadSrc1) | |||
1859 | LoadSrc1 = Builder.CreateLoad(LoadSizeType, Source1); | |||
1860 | ||||
1861 | Value *LoadSrc2 = nullptr; | |||
1862 | if (auto *Source2C = dyn_cast<Constant>(Source2)) | |||
1863 | LoadSrc2 = ConstantFoldLoadFromConstPtr(Source2C, LoadSizeType, DL); | |||
1864 | if (!LoadSrc2) | |||
1865 | LoadSrc2 = Builder.CreateLoad(LoadSizeType, Source2); | |||
1866 | ||||
1867 | if (NumLoads != 1) { | |||
1868 | if (LoadSizeType != MaxLoadType) { | |||
1869 | LoadSrc1 = Builder.CreateZExt(LoadSrc1, MaxLoadType); | |||
1870 | LoadSrc2 = Builder.CreateZExt(LoadSrc2, MaxLoadType); | |||
1871 | } | |||
1872 | // If we have multiple loads per block, we need to generate a composite | |||
1873 | // comparison using xor+or. | |||
1874 | Diff = Builder.CreateXor(LoadSrc1, LoadSrc2); | |||
1875 | Diff = Builder.CreateZExt(Diff, MaxLoadType); | |||
1876 | XorList.push_back(Diff); | |||
1877 | } else { | |||
1878 | // If there's only one load per block, we just compare the loaded values. | |||
1879 | Cmp = Builder.CreateICmpNE(LoadSrc1, LoadSrc2); | |||
1880 | } | |||
1881 | } | |||
1882 | ||||
1883 | auto pairWiseOr = [&](std::vector<Value *> &InList) -> std::vector<Value *> { | |||
1884 | std::vector<Value *> OutList; | |||
1885 | for (unsigned i = 0; i < InList.size() - 1; i = i + 2) { | |||
1886 | Value *Or = Builder.CreateOr(InList[i], InList[i + 1]); | |||
1887 | OutList.push_back(Or); | |||
1888 | } | |||
1889 | if (InList.size() % 2 != 0) | |||
1890 | OutList.push_back(InList.back()); | |||
1891 | return OutList; | |||
1892 | }; | |||
1893 | ||||
1894 | if (!Cmp) { | |||
1895 | // Pairwise OR the XOR results. | |||
1896 | OrList = pairWiseOr(XorList); | |||
1897 | ||||
1898 | // Pairwise OR the OR results until one result left. | |||
1899 | while (OrList.size() != 1) { | |||
1900 | OrList = pairWiseOr(OrList); | |||
1901 | } | |||
1902 | Cmp = Builder.CreateICmpNE(OrList[0], ConstantInt::get(Diff->getType(), 0)); | |||
1903 | } | |||
1904 | ||||
1905 | return Cmp; | |||
1906 | } | |||
1907 | ||||
1908 | void MemCmpExpansion::emitLoadCompareBlockMultipleLoads( | |||
1909 | unsigned Index, unsigned Size, unsigned &NumBytesProcessed) { | |||
1910 | Value *Cmp = getCompareLoadPairs(Index, Size, NumBytesProcessed); | |||
1911 | ||||
1912 | BasicBlock *NextBB = (Index == (LoadCmpBlocks.size() - 1)) | |||
1913 | ? EndBlock | |||
1914 | : LoadCmpBlocks[Index + 1]; | |||
1915 | // Early exit branch if difference found to ResultBlock. Otherwise, | |||
1916 | // continue to next LoadCmpBlock or EndBlock. | |||
1917 | BranchInst *CmpBr = BranchInst::Create(ResBlock.BB, NextBB, Cmp); | |||
1918 | Builder.Insert(CmpBr); | |||
1919 | ||||
1920 | // Add a phi edge for the last LoadCmpBlock to Endblock with a value of 0 | |||
1921 | // since early exit to ResultBlock was not taken (no difference was found in | |||
1922 | // any of the bytes). | |||
1923 | if (Index == LoadCmpBlocks.size() - 1) { | |||
1924 | Value *Zero = ConstantInt::get(Type::getInt32Ty(CI->getContext()), 0); | |||
1925 | PhiRes->addIncoming(Zero, LoadCmpBlocks[Index]); | |||
1926 | } | |||
1927 | } | |||
1928 | ||||
1929 | // This function creates the IR intructions for loading and comparing using the | |||
1930 | // given LoadSize. It loads the number of bytes specified by LoadSize from each | |||
1931 | // source of the memcmp parameters. It then does a subtract to see if there was | |||
1932 | // a difference in the loaded values. If a difference is found, it branches | |||
1933 | // with an early exit to the ResultBlock for calculating which source was | |||
1934 | // larger. Otherwise, it falls through to the either the next LoadCmpBlock or | |||
1935 | // the EndBlock if this is the last LoadCmpBlock. Loading 1 byte is handled with | |||
1936 | // a special case through emitLoadCompareByteBlock. The special handling can | |||
1937 | // simply subtract the loaded values and add it to the result phi node. | |||
1938 | void MemCmpExpansion::emitLoadCompareBlock(unsigned Index, unsigned LoadSize, | |||
1939 | unsigned GEPIndex) { | |||
1940 | if (LoadSize == 1) { | |||
1941 | MemCmpExpansion::emitLoadCompareByteBlock(Index, GEPIndex); | |||
1942 | return; | |||
1943 | } | |||
1944 | ||||
1945 | Type *LoadSizeType = IntegerType::get(CI->getContext(), LoadSize * 8); | |||
1946 | Type *MaxLoadType = IntegerType::get(CI->getContext(), MaxLoadSize * 8); | |||
1947 | assert(LoadSize <= MaxLoadSize && "Unexpected load type")((LoadSize <= MaxLoadSize && "Unexpected load type" ) ? static_cast<void> (0) : __assert_fail ("LoadSize <= MaxLoadSize && \"Unexpected load type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 1947, __PRETTY_FUNCTION__)); | |||
1948 | ||||
1949 | Value *Source1 = CI->getArgOperand(0); | |||
1950 | Value *Source2 = CI->getArgOperand(1); | |||
1951 | ||||
1952 | Builder.SetInsertPoint(LoadCmpBlocks[Index]); | |||
1953 | // Cast source to LoadSizeType*. | |||
1954 | if (Source1->getType() != LoadSizeType) | |||
1955 | Source1 = Builder.CreateBitCast(Source1, LoadSizeType->getPointerTo()); | |||
1956 | if (Source2->getType() != LoadSizeType) | |||
1957 | Source2 = Builder.CreateBitCast(Source2, LoadSizeType->getPointerTo()); | |||
1958 | ||||
1959 | // Get the base address using the GEPIndex. | |||
1960 | if (GEPIndex != 0) { | |||
1961 | Source1 = Builder.CreateGEP(LoadSizeType, Source1, | |||
1962 | ConstantInt::get(LoadSizeType, GEPIndex)); | |||
1963 | Source2 = Builder.CreateGEP(LoadSizeType, Source2, | |||
1964 | ConstantInt::get(LoadSizeType, GEPIndex)); | |||
1965 | } | |||
1966 | ||||
1967 | // Load LoadSizeType from the base address. | |||
1968 | Value *LoadSrc1 = Builder.CreateLoad(LoadSizeType, Source1); | |||
1969 | Value *LoadSrc2 = Builder.CreateLoad(LoadSizeType, Source2); | |||
1970 | ||||
1971 | if (DL.isLittleEndian()) { | |||
1972 | Function *Bswap = Intrinsic::getDeclaration(CI->getModule(), | |||
1973 | Intrinsic::bswap, LoadSizeType); | |||
1974 | LoadSrc1 = Builder.CreateCall(Bswap, LoadSrc1); | |||
1975 | LoadSrc2 = Builder.CreateCall(Bswap, LoadSrc2); | |||
1976 | } | |||
1977 | ||||
1978 | if (LoadSizeType != MaxLoadType) { | |||
1979 | LoadSrc1 = Builder.CreateZExt(LoadSrc1, MaxLoadType); | |||
1980 | LoadSrc2 = Builder.CreateZExt(LoadSrc2, MaxLoadType); | |||
1981 | } | |||
1982 | ||||
1983 | // Add the loaded values to the phi nodes for calculating memcmp result only | |||
1984 | // if result is not used in a zero equality. | |||
1985 | if (!IsUsedForZeroCmp) { | |||
1986 | ResBlock.PhiSrc1->addIncoming(LoadSrc1, LoadCmpBlocks[Index]); | |||
1987 | ResBlock.PhiSrc2->addIncoming(LoadSrc2, LoadCmpBlocks[Index]); | |||
1988 | } | |||
1989 | ||||
1990 | Value *Diff = Builder.CreateSub(LoadSrc1, LoadSrc2); | |||
1991 | ||||
1992 | Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_NE, Diff, | |||
1993 | ConstantInt::get(Diff->getType(), 0)); | |||
1994 | BasicBlock *NextBB = (Index == (LoadCmpBlocks.size() - 1)) | |||
1995 | ? EndBlock | |||
1996 | : LoadCmpBlocks[Index + 1]; | |||
1997 | // Early exit branch if difference found to ResultBlock. Otherwise, continue | |||
1998 | // to next LoadCmpBlock or EndBlock. | |||
1999 | BranchInst *CmpBr = BranchInst::Create(ResBlock.BB, NextBB, Cmp); | |||
2000 | Builder.Insert(CmpBr); | |||
2001 | ||||
2002 | // Add a phi edge for the last LoadCmpBlock to Endblock with a value of 0 | |||
2003 | // since early exit to ResultBlock was not taken (no difference was found in | |||
2004 | // any of the bytes). | |||
2005 | if (Index == LoadCmpBlocks.size() - 1) { | |||
2006 | Value *Zero = ConstantInt::get(Type::getInt32Ty(CI->getContext()), 0); | |||
2007 | PhiRes->addIncoming(Zero, LoadCmpBlocks[Index]); | |||
2008 | } | |||
2009 | } | |||
2010 | ||||
2011 | // This function populates the ResultBlock with a sequence to calculate the | |||
2012 | // memcmp result. It compares the two loaded source values and returns -1 if | |||
2013 | // src1 < src2 and 1 if src1 > src2. | |||
2014 | void MemCmpExpansion::emitMemCmpResultBlock() { | |||
2015 | // Special case: if memcmp result is used in a zero equality, result does not | |||
2016 | // need to be calculated and can simply return 1. | |||
2017 | if (IsUsedForZeroCmp) { | |||
2018 | BasicBlock::iterator InsertPt = ResBlock.BB->getFirstInsertionPt(); | |||
2019 | Builder.SetInsertPoint(ResBlock.BB, InsertPt); | |||
2020 | Value *Res = ConstantInt::get(Type::getInt32Ty(CI->getContext()), 1); | |||
2021 | PhiRes->addIncoming(Res, ResBlock.BB); | |||
2022 | BranchInst *NewBr = BranchInst::Create(EndBlock); | |||
2023 | Builder.Insert(NewBr); | |||
2024 | return; | |||
2025 | } | |||
2026 | BasicBlock::iterator InsertPt = ResBlock.BB->getFirstInsertionPt(); | |||
2027 | Builder.SetInsertPoint(ResBlock.BB, InsertPt); | |||
2028 | ||||
2029 | Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_ULT, ResBlock.PhiSrc1, | |||
2030 | ResBlock.PhiSrc2); | |||
2031 | ||||
2032 | Value *Res = | |||
2033 | Builder.CreateSelect(Cmp, ConstantInt::get(Builder.getInt32Ty(), -1), | |||
2034 | ConstantInt::get(Builder.getInt32Ty(), 1)); | |||
2035 | ||||
2036 | BranchInst *NewBr = BranchInst::Create(EndBlock); | |||
2037 | Builder.Insert(NewBr); | |||
2038 | PhiRes->addIncoming(Res, ResBlock.BB); | |||
2039 | } | |||
2040 | ||||
2041 | unsigned MemCmpExpansion::calculateNumBlocks(unsigned Size) { | |||
2042 | unsigned NumBlocks = 0; | |||
2043 | bool HaveOneByteLoad = false; | |||
2044 | unsigned RemainingSize = Size; | |||
2045 | unsigned LoadSize = MaxLoadSize; | |||
2046 | while (RemainingSize) { | |||
2047 | if (LoadSize == 1) | |||
2048 | HaveOneByteLoad = true; | |||
2049 | NumBlocks += RemainingSize / LoadSize; | |||
2050 | RemainingSize = RemainingSize % LoadSize; | |||
2051 | LoadSize = LoadSize / 2; | |||
2052 | } | |||
2053 | NumBlocksNonOneByte = HaveOneByteLoad ? (NumBlocks - 1) : NumBlocks; | |||
2054 | ||||
2055 | if (IsUsedForZeroCmp) | |||
2056 | NumBlocks = NumBlocks / NumLoadsPerBlock + | |||
2057 | (NumBlocks % NumLoadsPerBlock != 0 ? 1 : 0); | |||
2058 | ||||
2059 | return NumBlocks; | |||
2060 | } | |||
2061 | ||||
2062 | void MemCmpExpansion::setupResultBlockPHINodes() { | |||
2063 | Type *MaxLoadType = IntegerType::get(CI->getContext(), MaxLoadSize * 8); | |||
2064 | Builder.SetInsertPoint(ResBlock.BB); | |||
2065 | ResBlock.PhiSrc1 = | |||
2066 | Builder.CreatePHI(MaxLoadType, NumBlocksNonOneByte, "phi.src1"); | |||
2067 | ResBlock.PhiSrc2 = | |||
2068 | Builder.CreatePHI(MaxLoadType, NumBlocksNonOneByte, "phi.src2"); | |||
2069 | } | |||
2070 | ||||
2071 | void MemCmpExpansion::setupEndBlockPHINodes() { | |||
2072 | Builder.SetInsertPoint(&EndBlock->front()); | |||
2073 | PhiRes = Builder.CreatePHI(Type::getInt32Ty(CI->getContext()), 2, "phi.res"); | |||
2074 | } | |||
2075 | ||||
2076 | Value *MemCmpExpansion::getMemCmpExpansionZeroCase(unsigned Size) { | |||
2077 | unsigned NumBytesProcessed = 0; | |||
2078 | // This loop populates each of the LoadCmpBlocks with the IR sequence to | |||
2079 | // handle multiple loads per block. | |||
2080 | for (unsigned i = 0; i < NumBlocks; ++i) | |||
2081 | emitLoadCompareBlockMultipleLoads(i, Size, NumBytesProcessed); | |||
2082 | ||||
2083 | emitMemCmpResultBlock(); | |||
2084 | return PhiRes; | |||
2085 | } | |||
2086 | ||||
2087 | /// A memcmp expansion that compares equality with 0 and only has one block of | |||
2088 | /// load and compare can bypass the compare, branch, and phi IR that is required | |||
2089 | /// in the general case. | |||
2090 | Value *MemCmpExpansion::getMemCmpEqZeroOneBlock(unsigned Size) { | |||
2091 | unsigned NumBytesProcessed = 0; | |||
2092 | Value *Cmp = getCompareLoadPairs(0, Size, NumBytesProcessed); | |||
2093 | return Builder.CreateZExt(Cmp, Type::getInt32Ty(CI->getContext())); | |||
2094 | } | |||
2095 | ||||
2096 | // This function expands the memcmp call into an inline expansion and returns | |||
2097 | // the memcmp result. | |||
2098 | Value *MemCmpExpansion::getMemCmpExpansion(uint64_t Size) { | |||
2099 | if (IsUsedForZeroCmp) | |||
2100 | return NumBlocks == 1 ? getMemCmpEqZeroOneBlock(Size) : | |||
2101 | getMemCmpExpansionZeroCase(Size); | |||
2102 | ||||
2103 | // This loop calls emitLoadCompareBlock for comparing Size bytes of the two | |||
2104 | // memcmp sources. It starts with loading using the maximum load size set by | |||
2105 | // the target. It processes any remaining bytes using a load size which is the | |||
2106 | // next smallest power of 2. | |||
2107 | unsigned LoadSize = MaxLoadSize; | |||
2108 | unsigned NumBytesToBeProcessed = Size; | |||
2109 | unsigned Index = 0; | |||
2110 | while (NumBytesToBeProcessed) { | |||
2111 | // Calculate how many blocks we can create with the current load size. | |||
2112 | unsigned NumBlocks = NumBytesToBeProcessed / LoadSize; | |||
2113 | unsigned GEPIndex = (Size - NumBytesToBeProcessed) / LoadSize; | |||
2114 | NumBytesToBeProcessed = NumBytesToBeProcessed % LoadSize; | |||
2115 | ||||
2116 | // For each NumBlocks, populate the instruction sequence for loading and | |||
2117 | // comparing LoadSize bytes. | |||
2118 | while (NumBlocks--) { | |||
2119 | emitLoadCompareBlock(Index, LoadSize, GEPIndex); | |||
2120 | Index++; | |||
2121 | GEPIndex++; | |||
2122 | } | |||
2123 | // Get the next LoadSize to use. | |||
2124 | LoadSize = LoadSize / 2; | |||
2125 | } | |||
2126 | ||||
2127 | emitMemCmpResultBlock(); | |||
2128 | return PhiRes; | |||
2129 | } | |||
2130 | ||||
2131 | // This function checks to see if an expansion of memcmp can be generated. | |||
2132 | // It checks for constant compare size that is less than the max inline size. | |||
2133 | // If an expansion cannot occur, returns false to leave as a library call. | |||
2134 | // Otherwise, the library call is replaced with a new IR instruction sequence. | |||
2135 | /// We want to transform: | |||
2136 | /// %call = call signext i32 @memcmp(i8* %0, i8* %1, i64 15) | |||
2137 | /// To: | |||
2138 | /// loadbb: | |||
2139 | /// %0 = bitcast i32* %buffer2 to i8* | |||
2140 | /// %1 = bitcast i32* %buffer1 to i8* | |||
2141 | /// %2 = bitcast i8* %1 to i64* | |||
2142 | /// %3 = bitcast i8* %0 to i64* | |||
2143 | /// %4 = load i64, i64* %2 | |||
2144 | /// %5 = load i64, i64* %3 | |||
2145 | /// %6 = call i64 @llvm.bswap.i64(i64 %4) | |||
2146 | /// %7 = call i64 @llvm.bswap.i64(i64 %5) | |||
2147 | /// %8 = sub i64 %6, %7 | |||
2148 | /// %9 = icmp ne i64 %8, 0 | |||
2149 | /// br i1 %9, label %res_block, label %loadbb1 | |||
2150 | /// res_block: ; preds = %loadbb2, | |||
2151 | /// %loadbb1, %loadbb | |||
2152 | /// %phi.src1 = phi i64 [ %6, %loadbb ], [ %22, %loadbb1 ], [ %36, %loadbb2 ] | |||
2153 | /// %phi.src2 = phi i64 [ %7, %loadbb ], [ %23, %loadbb1 ], [ %37, %loadbb2 ] | |||
2154 | /// %10 = icmp ult i64 %phi.src1, %phi.src2 | |||
2155 | /// %11 = select i1 %10, i32 -1, i32 1 | |||
2156 | /// br label %endblock | |||
2157 | /// loadbb1: ; preds = %loadbb | |||
2158 | /// %12 = bitcast i32* %buffer2 to i8* | |||
2159 | /// %13 = bitcast i32* %buffer1 to i8* | |||
2160 | /// %14 = bitcast i8* %13 to i32* | |||
2161 | /// %15 = bitcast i8* %12 to i32* | |||
2162 | /// %16 = getelementptr i32, i32* %14, i32 2 | |||
2163 | /// %17 = getelementptr i32, i32* %15, i32 2 | |||
2164 | /// %18 = load i32, i32* %16 | |||
2165 | /// %19 = load i32, i32* %17 | |||
2166 | /// %20 = call i32 @llvm.bswap.i32(i32 %18) | |||
2167 | /// %21 = call i32 @llvm.bswap.i32(i32 %19) | |||
2168 | /// %22 = zext i32 %20 to i64 | |||
2169 | /// %23 = zext i32 %21 to i64 | |||
2170 | /// %24 = sub i64 %22, %23 | |||
2171 | /// %25 = icmp ne i64 %24, 0 | |||
2172 | /// br i1 %25, label %res_block, label %loadbb2 | |||
2173 | /// loadbb2: ; preds = %loadbb1 | |||
2174 | /// %26 = bitcast i32* %buffer2 to i8* | |||
2175 | /// %27 = bitcast i32* %buffer1 to i8* | |||
2176 | /// %28 = bitcast i8* %27 to i16* | |||
2177 | /// %29 = bitcast i8* %26 to i16* | |||
2178 | /// %30 = getelementptr i16, i16* %28, i16 6 | |||
2179 | /// %31 = getelementptr i16, i16* %29, i16 6 | |||
2180 | /// %32 = load i16, i16* %30 | |||
2181 | /// %33 = load i16, i16* %31 | |||
2182 | /// %34 = call i16 @llvm.bswap.i16(i16 %32) | |||
2183 | /// %35 = call i16 @llvm.bswap.i16(i16 %33) | |||
2184 | /// %36 = zext i16 %34 to i64 | |||
2185 | /// %37 = zext i16 %35 to i64 | |||
2186 | /// %38 = sub i64 %36, %37 | |||
2187 | /// %39 = icmp ne i64 %38, 0 | |||
2188 | /// br i1 %39, label %res_block, label %loadbb3 | |||
2189 | /// loadbb3: ; preds = %loadbb2 | |||
2190 | /// %40 = bitcast i32* %buffer2 to i8* | |||
2191 | /// %41 = bitcast i32* %buffer1 to i8* | |||
2192 | /// %42 = getelementptr i8, i8* %41, i8 14 | |||
2193 | /// %43 = getelementptr i8, i8* %40, i8 14 | |||
2194 | /// %44 = load i8, i8* %42 | |||
2195 | /// %45 = load i8, i8* %43 | |||
2196 | /// %46 = zext i8 %44 to i32 | |||
2197 | /// %47 = zext i8 %45 to i32 | |||
2198 | /// %48 = sub i32 %46, %47 | |||
2199 | /// br label %endblock | |||
2200 | /// endblock: ; preds = %res_block, | |||
2201 | /// %loadbb3 | |||
2202 | /// %phi.res = phi i32 [ %48, %loadbb3 ], [ %11, %res_block ] | |||
2203 | /// ret i32 %phi.res | |||
2204 | static bool expandMemCmp(CallInst *CI, const TargetTransformInfo *TTI, | |||
2205 | const TargetLowering *TLI, const DataLayout *DL) { | |||
2206 | NumMemCmpCalls++; | |||
2207 | ||||
2208 | // TTI call to check if target would like to expand memcmp. Also, get the | |||
2209 | // MaxLoadSize. | |||
2210 | unsigned MaxLoadSize; | |||
2211 | if (!TTI->expandMemCmp(CI, MaxLoadSize)) | |||
2212 | return false; | |||
2213 | ||||
2214 | // Early exit from expansion if -Oz. | |||
2215 | if (CI->getFunction()->optForMinSize()) | |||
2216 | return false; | |||
2217 | ||||
2218 | // Early exit from expansion if size is not a constant. | |||
2219 | ConstantInt *SizeCast = dyn_cast<ConstantInt>(CI->getArgOperand(2)); | |||
2220 | if (!SizeCast) { | |||
2221 | NumMemCmpNotConstant++; | |||
2222 | return false; | |||
2223 | } | |||
2224 | ||||
2225 | // Early exit from expansion if size greater than max bytes to load. | |||
2226 | uint64_t SizeVal = SizeCast->getZExtValue(); | |||
2227 | unsigned NumLoads = 0; | |||
2228 | unsigned RemainingSize = SizeVal; | |||
2229 | unsigned LoadSize = MaxLoadSize; | |||
2230 | while (RemainingSize) { | |||
2231 | NumLoads += RemainingSize / LoadSize; | |||
2232 | RemainingSize = RemainingSize % LoadSize; | |||
2233 | LoadSize = LoadSize / 2; | |||
2234 | } | |||
2235 | ||||
2236 | if (NumLoads > TLI->getMaxExpandSizeMemcmp(CI->getFunction()->optForSize())) { | |||
| ||||
2237 | NumMemCmpGreaterThanMax++; | |||
2238 | return false; | |||
2239 | } | |||
2240 | ||||
2241 | NumMemCmpInlined++; | |||
2242 | ||||
2243 | // MemCmpHelper object creates and sets up basic blocks required for | |||
2244 | // expanding memcmp with size SizeVal. | |||
2245 | unsigned NumLoadsPerBlock = MemCmpNumLoadsPerBlock; | |||
2246 | MemCmpExpansion MemCmpHelper(CI, SizeVal, MaxLoadSize, NumLoadsPerBlock, *DL); | |||
2247 | ||||
2248 | Value *Res = MemCmpHelper.getMemCmpExpansion(SizeVal); | |||
2249 | ||||
2250 | // Replace call with result of expansion and erase call. | |||
2251 | CI->replaceAllUsesWith(Res); | |||
2252 | CI->eraseFromParent(); | |||
2253 | ||||
2254 | return true; | |||
2255 | } | |||
2256 | ||||
2257 | bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) { | |||
2258 | BasicBlock *BB = CI->getParent(); | |||
2259 | ||||
2260 | // Lower inline assembly if we can. | |||
2261 | // If we found an inline asm expession, and if the target knows how to | |||
2262 | // lower it to normal LLVM code, do so now. | |||
2263 | if (TLI && isa<InlineAsm>(CI->getCalledValue())) { | |||
| ||||
2264 | if (TLI->ExpandInlineAsm(CI)) { | |||
2265 | // Avoid invalidating the iterator. | |||
2266 | CurInstIterator = BB->begin(); | |||
2267 | // Avoid processing instructions out of order, which could cause | |||
2268 | // reuse before a value is defined. | |||
2269 | SunkAddrs.clear(); | |||
2270 | return true; | |||
2271 | } | |||
2272 | // Sink address computing for memory operands into the block. | |||
2273 | if (optimizeInlineAsmInst(CI)) | |||
2274 | return true; | |||
2275 | } | |||
2276 | ||||
2277 | // Align the pointer arguments to this call if the target thinks it's a good | |||
2278 | // idea | |||
2279 | unsigned MinSize, PrefAlign; | |||
2280 | if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) { | |||
2281 | for (auto &Arg : CI->arg_operands()) { | |||
2282 | // We want to align both objects whose address is used directly and | |||
2283 | // objects whose address is used in casts and GEPs, though it only makes | |||
2284 | // sense for GEPs if the offset is a multiple of the desired alignment and | |||
2285 | // if size - offset meets the size threshold. | |||
2286 | if (!Arg->getType()->isPointerTy()) | |||
2287 | continue; | |||
2288 | APInt Offset(DL->getPointerSizeInBits( | |||
2289 | cast<PointerType>(Arg->getType())->getAddressSpace()), | |||
2290 | 0); | |||
2291 | Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset); | |||
2292 | uint64_t Offset2 = Offset.getLimitedValue(); | |||
2293 | if ((Offset2 & (PrefAlign-1)) != 0) | |||
2294 | continue; | |||
2295 | AllocaInst *AI; | |||
2296 | if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign && | |||
2297 | DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2) | |||
2298 | AI->setAlignment(PrefAlign); | |||
2299 | // Global variables can only be aligned if they are defined in this | |||
2300 | // object (i.e. they are uniquely initialized in this object), and | |||
2301 | // over-aligning global variables that have an explicit section is | |||
2302 | // forbidden. | |||
2303 | GlobalVariable *GV; | |||
2304 | if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() && | |||
2305 | GV->getPointerAlignment(*DL) < PrefAlign && | |||
2306 | DL->getTypeAllocSize(GV->getValueType()) >= | |||
2307 | MinSize + Offset2) | |||
2308 | GV->setAlignment(PrefAlign); | |||
2309 | } | |||
2310 | // If this is a memcpy (or similar) then we may be able to improve the | |||
2311 | // alignment | |||
2312 | if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) { | |||
2313 | unsigned Align = getKnownAlignment(MI->getDest(), *DL); | |||
2314 | if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) | |||
2315 | Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL)); | |||
2316 | if (Align > MI->getAlignment()) | |||
2317 | MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align)); | |||
2318 | } | |||
2319 | } | |||
2320 | ||||
2321 | // If we have a cold call site, try to sink addressing computation into the | |||
2322 | // cold block. This interacts with our handling for loads and stores to | |||
2323 | // ensure that we can fold all uses of a potential addressing computation | |||
2324 | // into their uses. TODO: generalize this to work over profiling data | |||
2325 | if (!OptSize && CI->hasFnAttr(Attribute::Cold)) | |||
2326 | for (auto &Arg : CI->arg_operands()) { | |||
2327 | if (!Arg->getType()->isPointerTy()) | |||
2328 | continue; | |||
2329 | unsigned AS = Arg->getType()->getPointerAddressSpace(); | |||
2330 | return optimizeMemoryInst(CI, Arg, Arg->getType(), AS); | |||
2331 | } | |||
2332 | ||||
2333 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI); | |||
2334 | if (II) { | |||
2335 | switch (II->getIntrinsicID()) { | |||
2336 | default: break; | |||
2337 | case Intrinsic::objectsize: { | |||
2338 | // Lower all uses of llvm.objectsize.* | |||
2339 | ConstantInt *RetVal = | |||
2340 | lowerObjectSizeCall(II, *DL, TLInfo, /*MustSucceed=*/true); | |||
2341 | // Substituting this can cause recursive simplifications, which can | |||
2342 | // invalidate our iterator. Use a WeakTrackingVH to hold onto it in case | |||
2343 | // this | |||
2344 | // happens. | |||
2345 | Value *CurValue = &*CurInstIterator; | |||
2346 | WeakTrackingVH IterHandle(CurValue); | |||
2347 | ||||
2348 | replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr); | |||
2349 | ||||
2350 | // If the iterator instruction was recursively deleted, start over at the | |||
2351 | // start of the block. | |||
2352 | if (IterHandle != CurValue) { | |||
2353 | CurInstIterator = BB->begin(); | |||
2354 | SunkAddrs.clear(); | |||
2355 | } | |||
2356 | return true; | |||
2357 | } | |||
2358 | case Intrinsic::aarch64_stlxr: | |||
2359 | case Intrinsic::aarch64_stxr: { | |||
2360 | ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0)); | |||
2361 | if (!ExtVal || !ExtVal->hasOneUse() || | |||
2362 | ExtVal->getParent() == CI->getParent()) | |||
2363 | return false; | |||
2364 | // Sink a zext feeding stlxr/stxr before it, so it can be folded into it. | |||
2365 | ExtVal->moveBefore(CI); | |||
2366 | // Mark this instruction as "inserted by CGP", so that other | |||
2367 | // optimizations don't touch it. | |||
2368 | InsertedInsts.insert(ExtVal); | |||
2369 | return true; | |||
2370 | } | |||
2371 | case Intrinsic::invariant_group_barrier: | |||
2372 | II->replaceAllUsesWith(II->getArgOperand(0)); | |||
2373 | II->eraseFromParent(); | |||
2374 | return true; | |||
2375 | ||||
2376 | case Intrinsic::cttz: | |||
2377 | case Intrinsic::ctlz: | |||
2378 | // If counting zeros is expensive, try to avoid it. | |||
2379 | return despeculateCountZeros(II, TLI, DL, ModifiedDT); | |||
2380 | } | |||
2381 | ||||
2382 | if (TLI) { | |||
2383 | SmallVector<Value*, 2> PtrOps; | |||
2384 | Type *AccessTy; | |||
2385 | if (TLI->getAddrModeArguments(II, PtrOps, AccessTy)) | |||
2386 | while (!PtrOps.empty()) { | |||
2387 | Value *PtrVal = PtrOps.pop_back_val(); | |||
2388 | unsigned AS = PtrVal->getType()->getPointerAddressSpace(); | |||
2389 | if (optimizeMemoryInst(II, PtrVal, AccessTy, AS)) | |||
2390 | return true; | |||
2391 | } | |||
2392 | } | |||
2393 | } | |||
2394 | ||||
2395 | // From here on out we're working with named functions. | |||
2396 | if (!CI->getCalledFunction()) return false; | |||
2397 | ||||
2398 | // Lower all default uses of _chk calls. This is very similar | |||
2399 | // to what InstCombineCalls does, but here we are only lowering calls | |||
2400 | // to fortified library functions (e.g. __memcpy_chk) that have the default | |||
2401 | // "don't know" as the objectsize. Anything else should be left alone. | |||
2402 | FortifiedLibCallSimplifier Simplifier(TLInfo, true); | |||
2403 | if (Value *V = Simplifier.optimizeCall(CI)) { | |||
2404 | CI->replaceAllUsesWith(V); | |||
2405 | CI->eraseFromParent(); | |||
2406 | return true; | |||
2407 | } | |||
2408 | ||||
2409 | LibFunc Func; | |||
2410 | if (TLInfo->getLibFunc(ImmutableCallSite(CI), Func) && | |||
2411 | Func == LibFunc_memcmp && expandMemCmp(CI, TTI, TLI, DL)) { | |||
2412 | ModifiedDT = true; | |||
2413 | return true; | |||
2414 | } | |||
2415 | return false; | |||
2416 | } | |||
2417 | ||||
2418 | /// Look for opportunities to duplicate return instructions to the predecessor | |||
2419 | /// to enable tail call optimizations. The case it is currently looking for is: | |||
2420 | /// @code | |||
2421 | /// bb0: | |||
2422 | /// %tmp0 = tail call i32 @f0() | |||
2423 | /// br label %return | |||
2424 | /// bb1: | |||
2425 | /// %tmp1 = tail call i32 @f1() | |||
2426 | /// br label %return | |||
2427 | /// bb2: | |||
2428 | /// %tmp2 = tail call i32 @f2() | |||
2429 | /// br label %return | |||
2430 | /// return: | |||
2431 | /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ] | |||
2432 | /// ret i32 %retval | |||
2433 | /// @endcode | |||
2434 | /// | |||
2435 | /// => | |||
2436 | /// | |||
2437 | /// @code | |||
2438 | /// bb0: | |||
2439 | /// %tmp0 = tail call i32 @f0() | |||
2440 | /// ret i32 %tmp0 | |||
2441 | /// bb1: | |||
2442 | /// %tmp1 = tail call i32 @f1() | |||
2443 | /// ret i32 %tmp1 | |||
2444 | /// bb2: | |||
2445 | /// %tmp2 = tail call i32 @f2() | |||
2446 | /// ret i32 %tmp2 | |||
2447 | /// @endcode | |||
2448 | bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) { | |||
2449 | if (!TLI) | |||
2450 | return false; | |||
2451 | ||||
2452 | ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator()); | |||
2453 | if (!RetI) | |||
2454 | return false; | |||
2455 | ||||
2456 | PHINode *PN = nullptr; | |||
2457 | BitCastInst *BCI = nullptr; | |||
2458 | Value *V = RetI->getReturnValue(); | |||
2459 | if (V) { | |||
2460 | BCI = dyn_cast<BitCastInst>(V); | |||
2461 | if (BCI) | |||
2462 | V = BCI->getOperand(0); | |||
2463 | ||||
2464 | PN = dyn_cast<PHINode>(V); | |||
2465 | if (!PN) | |||
2466 | return false; | |||
2467 | } | |||
2468 | ||||
2469 | if (PN && PN->getParent() != BB) | |||
2470 | return false; | |||
2471 | ||||
2472 | // Make sure there are no instructions between the PHI and return, or that the | |||
2473 | // return is the first instruction in the block. | |||
2474 | if (PN) { | |||
2475 | BasicBlock::iterator BI = BB->begin(); | |||
2476 | do { ++BI; } while (isa<DbgInfoIntrinsic>(BI)); | |||
2477 | if (&*BI == BCI) | |||
2478 | // Also skip over the bitcast. | |||
2479 | ++BI; | |||
2480 | if (&*BI != RetI) | |||
2481 | return false; | |||
2482 | } else { | |||
2483 | BasicBlock::iterator BI = BB->begin(); | |||
2484 | while (isa<DbgInfoIntrinsic>(BI)) ++BI; | |||
2485 | if (&*BI != RetI) | |||
2486 | return false; | |||
2487 | } | |||
2488 | ||||
2489 | /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail | |||
2490 | /// call. | |||
2491 | const Function *F = BB->getParent(); | |||
2492 | SmallVector<CallInst*, 4> TailCalls; | |||
2493 | if (PN) { | |||
2494 | for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) { | |||
2495 | CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I)); | |||
2496 | // Make sure the phi value is indeed produced by the tail call. | |||
2497 | if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) && | |||
2498 | TLI->mayBeEmittedAsTailCall(CI) && | |||
2499 | attributesPermitTailCall(F, CI, RetI, *TLI)) | |||
2500 | TailCalls.push_back(CI); | |||
2501 | } | |||
2502 | } else { | |||
2503 | SmallPtrSet<BasicBlock*, 4> VisitedBBs; | |||
2504 | for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) { | |||
2505 | if (!VisitedBBs.insert(*PI).second) | |||
2506 | continue; | |||
2507 | ||||
2508 | BasicBlock::InstListType &InstList = (*PI)->getInstList(); | |||
2509 | BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin(); | |||
2510 | BasicBlock::InstListType::reverse_iterator RE = InstList.rend(); | |||
2511 | do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI)); | |||
2512 | if (RI == RE) | |||
2513 | continue; | |||
2514 | ||||
2515 | CallInst *CI = dyn_cast<CallInst>(&*RI); | |||
2516 | if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) && | |||
2517 | attributesPermitTailCall(F, CI, RetI, *TLI)) | |||
2518 | TailCalls.push_back(CI); | |||
2519 | } | |||
2520 | } | |||
2521 | ||||
2522 | bool Changed = false; | |||
2523 | for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) { | |||
2524 | CallInst *CI = TailCalls[i]; | |||
2525 | CallSite CS(CI); | |||
2526 | ||||
2527 | // Conservatively require the attributes of the call to match those of the | |||
2528 | // return. Ignore noalias because it doesn't affect the call sequence. | |||
2529 | AttributeList CalleeAttrs = CS.getAttributes(); | |||
2530 | if (AttrBuilder(CalleeAttrs, AttributeList::ReturnIndex) | |||
2531 | .removeAttribute(Attribute::NoAlias) != | |||
2532 | AttrBuilder(CalleeAttrs, AttributeList::ReturnIndex) | |||
2533 | .removeAttribute(Attribute::NoAlias)) | |||
2534 | continue; | |||
2535 | ||||
2536 | // Make sure the call instruction is followed by an unconditional branch to | |||
2537 | // the return block. | |||
2538 | BasicBlock *CallBB = CI->getParent(); | |||
2539 | BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator()); | |||
2540 | if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB) | |||
2541 | continue; | |||
2542 | ||||
2543 | // Duplicate the return into CallBB. | |||
2544 | (void)FoldReturnIntoUncondBranch(RetI, BB, CallBB); | |||
2545 | ModifiedDT = Changed = true; | |||
2546 | ++NumRetsDup; | |||
2547 | } | |||
2548 | ||||
2549 | // If we eliminated all predecessors of the block, delete the block now. | |||
2550 | if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB)) | |||
2551 | BB->eraseFromParent(); | |||
2552 | ||||
2553 | return Changed; | |||
2554 | } | |||
2555 | ||||
2556 | //===----------------------------------------------------------------------===// | |||
2557 | // Memory Optimization | |||
2558 | //===----------------------------------------------------------------------===// | |||
2559 | ||||
2560 | namespace { | |||
2561 | ||||
2562 | /// This is an extended version of TargetLowering::AddrMode | |||
2563 | /// which holds actual Value*'s for register values. | |||
2564 | struct ExtAddrMode : public TargetLowering::AddrMode { | |||
2565 | Value *BaseReg; | |||
2566 | Value *ScaledReg; | |||
2567 | ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {} | |||
2568 | void print(raw_ostream &OS) const; | |||
2569 | void dump() const; | |||
2570 | ||||
2571 | bool operator==(const ExtAddrMode& O) const { | |||
2572 | return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) && | |||
2573 | (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) && | |||
2574 | (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale); | |||
2575 | } | |||
2576 | }; | |||
2577 | ||||
2578 | #ifndef NDEBUG | |||
2579 | static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) { | |||
2580 | AM.print(OS); | |||
2581 | return OS; | |||
2582 | } | |||
2583 | #endif | |||
2584 | ||||
2585 | void ExtAddrMode::print(raw_ostream &OS) const { | |||
2586 | bool NeedPlus = false; | |||
2587 | OS << "["; | |||
2588 | if (BaseGV) { | |||
2589 | OS << (NeedPlus ? " + " : "") | |||
2590 | << "GV:"; | |||
2591 | BaseGV->printAsOperand(OS, /*PrintType=*/false); | |||
2592 | NeedPlus = true; | |||
2593 | } | |||
2594 | ||||
2595 | if (BaseOffs) { | |||
2596 | OS << (NeedPlus ? " + " : "") | |||
2597 | << BaseOffs; | |||
2598 | NeedPlus = true; | |||
2599 | } | |||
2600 | ||||
2601 | if (BaseReg) { | |||
2602 | OS << (NeedPlus ? " + " : "") | |||
2603 | << "Base:"; | |||
2604 | BaseReg->printAsOperand(OS, /*PrintType=*/false); | |||
2605 | NeedPlus = true; | |||
2606 | } | |||
2607 | if (Scale) { | |||
2608 | OS << (NeedPlus ? " + " : "") | |||
2609 | << Scale << "*"; | |||
2610 | ScaledReg->printAsOperand(OS, /*PrintType=*/false); | |||
2611 | } | |||
2612 | ||||
2613 | OS << ']'; | |||
2614 | } | |||
2615 | ||||
2616 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | |||
2617 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void ExtAddrMode::dump() const { | |||
2618 | print(dbgs()); | |||
2619 | dbgs() << '\n'; | |||
2620 | } | |||
2621 | #endif | |||
2622 | ||||
2623 | /// \brief This class provides transaction based operation on the IR. | |||
2624 | /// Every change made through this class is recorded in the internal state and | |||
2625 | /// can be undone (rollback) until commit is called. | |||
2626 | class TypePromotionTransaction { | |||
2627 | ||||
2628 | /// \brief This represents the common interface of the individual transaction. | |||
2629 | /// Each class implements the logic for doing one specific modification on | |||
2630 | /// the IR via the TypePromotionTransaction. | |||
2631 | class TypePromotionAction { | |||
2632 | protected: | |||
2633 | /// The Instruction modified. | |||
2634 | Instruction *Inst; | |||
2635 | ||||
2636 | public: | |||
2637 | /// \brief Constructor of the action. | |||
2638 | /// The constructor performs the related action on the IR. | |||
2639 | TypePromotionAction(Instruction *Inst) : Inst(Inst) {} | |||
2640 | ||||
2641 | virtual ~TypePromotionAction() {} | |||
2642 | ||||
2643 | /// \brief Undo the modification done by this action. | |||
2644 | /// When this method is called, the IR must be in the same state as it was | |||
2645 | /// before this action was applied. | |||
2646 | /// \pre Undoing the action works if and only if the IR is in the exact same | |||
2647 | /// state as it was directly after this action was applied. | |||
2648 | virtual void undo() = 0; | |||
2649 | ||||
2650 | /// \brief Advocate every change made by this action. | |||
2651 | /// When the results on the IR of the action are to be kept, it is important | |||
2652 | /// to call this function, otherwise hidden information may be kept forever. | |||
2653 | virtual void commit() { | |||
2654 | // Nothing to be done, this action is not doing anything. | |||
2655 | } | |||
2656 | }; | |||
2657 | ||||
2658 | /// \brief Utility to remember the position of an instruction. | |||
2659 | class InsertionHandler { | |||
2660 | /// Position of an instruction. | |||
2661 | /// Either an instruction: | |||
2662 | /// - Is the first in a basic block: BB is used. | |||
2663 | /// - Has a previous instructon: PrevInst is used. | |||
2664 | union { | |||
2665 | Instruction *PrevInst; | |||
2666 | BasicBlock *BB; | |||
2667 | } Point; | |||
2668 | /// Remember whether or not the instruction had a previous instruction. | |||
2669 | bool HasPrevInstruction; | |||
2670 | ||||
2671 | public: | |||
2672 | /// \brief Record the position of \p Inst. | |||
2673 | InsertionHandler(Instruction *Inst) { | |||
2674 | BasicBlock::iterator It = Inst->getIterator(); | |||
2675 | HasPrevInstruction = (It != (Inst->getParent()->begin())); | |||
2676 | if (HasPrevInstruction) | |||
2677 | Point.PrevInst = &*--It; | |||
2678 | else | |||
2679 | Point.BB = Inst->getParent(); | |||
2680 | } | |||
2681 | ||||
2682 | /// \brief Insert \p Inst at the recorded position. | |||
2683 | void insert(Instruction *Inst) { | |||
2684 | if (HasPrevInstruction) { | |||
2685 | if (Inst->getParent()) | |||
2686 | Inst->removeFromParent(); | |||
2687 | Inst->insertAfter(Point.PrevInst); | |||
2688 | } else { | |||
2689 | Instruction *Position = &*Point.BB->getFirstInsertionPt(); | |||
2690 | if (Inst->getParent()) | |||
2691 | Inst->moveBefore(Position); | |||
2692 | else | |||
2693 | Inst->insertBefore(Position); | |||
2694 | } | |||
2695 | } | |||
2696 | }; | |||
2697 | ||||
2698 | /// \brief Move an instruction before another. | |||
2699 | class InstructionMoveBefore : public TypePromotionAction { | |||
2700 | /// Original position of the instruction. | |||
2701 | InsertionHandler Position; | |||
2702 | ||||
2703 | public: | |||
2704 | /// \brief Move \p Inst before \p Before. | |||
2705 | InstructionMoveBefore(Instruction *Inst, Instruction *Before) | |||
2706 | : TypePromotionAction(Inst), Position(Inst) { | |||
2707 | DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: move: " << * Inst << "\nbefore: " << *Before << "\n"; } } while (false); | |||
2708 | Inst->moveBefore(Before); | |||
2709 | } | |||
2710 | ||||
2711 | /// \brief Move the instruction back to its original position. | |||
2712 | void undo() override { | |||
2713 | DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: moveBefore: " << *Inst << "\n"; } } while (false); | |||
2714 | Position.insert(Inst); | |||
2715 | } | |||
2716 | }; | |||
2717 | ||||
2718 | /// \brief Set the operand of an instruction with a new value. | |||
2719 | class OperandSetter : public TypePromotionAction { | |||
2720 | /// Original operand of the instruction. | |||
2721 | Value *Origin; | |||
2722 | /// Index of the modified instruction. | |||
2723 | unsigned Idx; | |||
2724 | ||||
2725 | public: | |||
2726 | /// \brief Set \p Idx operand of \p Inst with \p NewVal. | |||
2727 | OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal) | |||
2728 | : TypePromotionAction(Inst), Idx(Idx) { | |||
2729 | DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: setOperand: " << Idx << "\n" << "for:" << *Inst << "\n" << "with:" << *NewVal << "\n"; } } while ( false) | |||
2730 | << "for:" << *Inst << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: setOperand: " << Idx << "\n" << "for:" << *Inst << "\n" << "with:" << *NewVal << "\n"; } } while ( false) | |||
2731 | << "with:" << *NewVal << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: setOperand: " << Idx << "\n" << "for:" << *Inst << "\n" << "with:" << *NewVal << "\n"; } } while ( false); | |||
2732 | Origin = Inst->getOperand(Idx); | |||
2733 | Inst->setOperand(Idx, NewVal); | |||
2734 | } | |||
2735 | ||||
2736 | /// \brief Restore the original value of the instruction. | |||
2737 | void undo() override { | |||
2738 | DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: setOperand:" << Idx << "\n" << "for: " << *Inst << "\n" << "with: " << *Origin << "\n"; } } while ( false) | |||
2739 | << "for: " << *Inst << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: setOperand:" << Idx << "\n" << "for: " << *Inst << "\n" << "with: " << *Origin << "\n"; } } while ( false) | |||
2740 | << "with: " << *Origin << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: setOperand:" << Idx << "\n" << "for: " << *Inst << "\n" << "with: " << *Origin << "\n"; } } while ( false); | |||
2741 | Inst->setOperand(Idx, Origin); | |||
2742 | } | |||
2743 | }; | |||
2744 | ||||
2745 | /// \brief Hide the operands of an instruction. | |||
2746 | /// Do as if this instruction was not using any of its operands. | |||
2747 | class OperandsHider : public TypePromotionAction { | |||
2748 | /// The list of original operands. | |||
2749 | SmallVector<Value *, 4> OriginalValues; | |||
2750 | ||||
2751 | public: | |||
2752 | /// \brief Remove \p Inst from the uses of the operands of \p Inst. | |||
2753 | OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) { | |||
2754 | DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: OperandsHider: " << *Inst << "\n"; } } while (false); | |||
2755 | unsigned NumOpnds = Inst->getNumOperands(); | |||
2756 | OriginalValues.reserve(NumOpnds); | |||
2757 | for (unsigned It = 0; It < NumOpnds; ++It) { | |||
2758 | // Save the current operand. | |||
2759 | Value *Val = Inst->getOperand(It); | |||
2760 | OriginalValues.push_back(Val); | |||
2761 | // Set a dummy one. | |||
2762 | // We could use OperandSetter here, but that would imply an overhead | |||
2763 | // that we are not willing to pay. | |||
2764 | Inst->setOperand(It, UndefValue::get(Val->getType())); | |||
2765 | } | |||
2766 | } | |||
2767 | ||||
2768 | /// \brief Restore the original list of uses. | |||
2769 | void undo() override { | |||
2770 | DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: OperandsHider: " << *Inst << "\n"; } } while (false); | |||
2771 | for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It) | |||
2772 | Inst->setOperand(It, OriginalValues[It]); | |||
2773 | } | |||
2774 | }; | |||
2775 | ||||
2776 | /// \brief Build a truncate instruction. | |||
2777 | class TruncBuilder : public TypePromotionAction { | |||
2778 | Value *Val; | |||
2779 | public: | |||
2780 | /// \brief Build a truncate instruction of \p Opnd producing a \p Ty | |||
2781 | /// result. | |||
2782 | /// trunc Opnd to Ty. | |||
2783 | TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) { | |||
2784 | IRBuilder<> Builder(Opnd); | |||
2785 | Val = Builder.CreateTrunc(Opnd, Ty, "promoted"); | |||
2786 | DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: TruncBuilder: " << *Val << "\n"; } } while (false); | |||
2787 | } | |||
2788 | ||||
2789 | /// \brief Get the built value. | |||
2790 | Value *getBuiltValue() { return Val; } | |||
2791 | ||||
2792 | /// \brief Remove the built instruction. | |||
2793 | void undo() override { | |||
2794 | DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: TruncBuilder: " << *Val << "\n"; } } while (false); | |||
2795 | if (Instruction *IVal = dyn_cast<Instruction>(Val)) | |||
2796 | IVal->eraseFromParent(); | |||
2797 | } | |||
2798 | }; | |||
2799 | ||||
2800 | /// \brief Build a sign extension instruction. | |||
2801 | class SExtBuilder : public TypePromotionAction { | |||
2802 | Value *Val; | |||
2803 | public: | |||
2804 | /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty | |||
2805 | /// result. | |||
2806 | /// sext Opnd to Ty. | |||
2807 | SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty) | |||
2808 | : TypePromotionAction(InsertPt) { | |||
2809 | IRBuilder<> Builder(InsertPt); | |||
2810 | Val = Builder.CreateSExt(Opnd, Ty, "promoted"); | |||
2811 | DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: SExtBuilder: " << *Val << "\n"; } } while (false); | |||
2812 | } | |||
2813 | ||||
2814 | /// \brief Get the built value. | |||
2815 | Value *getBuiltValue() { return Val; } | |||
2816 | ||||
2817 | /// \brief Remove the built instruction. | |||
2818 | void undo() override { | |||
2819 | DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: SExtBuilder: " << *Val << "\n"; } } while (false); | |||
2820 | if (Instruction *IVal = dyn_cast<Instruction>(Val)) | |||
2821 | IVal->eraseFromParent(); | |||
2822 | } | |||
2823 | }; | |||
2824 | ||||
2825 | /// \brief Build a zero extension instruction. | |||
2826 | class ZExtBuilder : public TypePromotionAction { | |||
2827 | Value *Val; | |||
2828 | public: | |||
2829 | /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty | |||
2830 | /// result. | |||
2831 | /// zext Opnd to Ty. | |||
2832 | ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty) | |||
2833 | : TypePromotionAction(InsertPt) { | |||
2834 | IRBuilder<> Builder(InsertPt); | |||
2835 | Val = Builder.CreateZExt(Opnd, Ty, "promoted"); | |||
2836 | DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: ZExtBuilder: " << *Val << "\n"; } } while (false); | |||
2837 | } | |||
2838 | ||||
2839 | /// \brief Get the built value. | |||
2840 | Value *getBuiltValue() { return Val; } | |||
2841 | ||||
2842 | /// \brief Remove the built instruction. | |||
2843 | void undo() override { | |||
2844 | DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: ZExtBuilder: " << *Val << "\n"; } } while (false); | |||
2845 | if (Instruction *IVal = dyn_cast<Instruction>(Val)) | |||
2846 | IVal->eraseFromParent(); | |||
2847 | } | |||
2848 | }; | |||
2849 | ||||
2850 | /// \brief Mutate an instruction to another type. | |||
2851 | class TypeMutator : public TypePromotionAction { | |||
2852 | /// Record the original type. | |||
2853 | Type *OrigTy; | |||
2854 | ||||
2855 | public: | |||
2856 | /// \brief Mutate the type of \p Inst into \p NewTy. | |||
2857 | TypeMutator(Instruction *Inst, Type *NewTy) | |||
2858 | : TypePromotionAction(Inst), OrigTy(Inst->getType()) { | |||
2859 | DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTydo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy << "\n"; } } while (false) | |||
2860 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy << "\n"; } } while (false); | |||
2861 | Inst->mutateType(NewTy); | |||
2862 | } | |||
2863 | ||||
2864 | /// \brief Mutate the instruction back to its original type. | |||
2865 | void undo() override { | |||
2866 | DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTydo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy << "\n"; } } while (false) | |||
2867 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy << "\n"; } } while (false); | |||
2868 | Inst->mutateType(OrigTy); | |||
2869 | } | |||
2870 | }; | |||
2871 | ||||
2872 | /// \brief Replace the uses of an instruction by another instruction. | |||
2873 | class UsesReplacer : public TypePromotionAction { | |||
2874 | /// Helper structure to keep track of the replaced uses. | |||
2875 | struct InstructionAndIdx { | |||
2876 | /// The instruction using the instruction. | |||
2877 | Instruction *Inst; | |||
2878 | /// The index where this instruction is used for Inst. | |||
2879 | unsigned Idx; | |||
2880 | InstructionAndIdx(Instruction *Inst, unsigned Idx) | |||
2881 | : Inst(Inst), Idx(Idx) {} | |||
2882 | }; | |||
2883 | ||||
2884 | /// Keep track of the original uses (pair Instruction, Index). | |||
2885 | SmallVector<InstructionAndIdx, 4> OriginalUses; | |||
2886 | typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator; | |||
2887 | ||||
2888 | public: | |||
2889 | /// \brief Replace all the use of \p Inst by \p New. | |||
2890 | UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) { | |||
2891 | DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *Newdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New << "\n"; } } while (false) | |||
2892 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New << "\n"; } } while (false); | |||
2893 | // Record the original uses. | |||
2894 | for (Use &U : Inst->uses()) { | |||
2895 | Instruction *UserI = cast<Instruction>(U.getUser()); | |||
2896 | OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo())); | |||
2897 | } | |||
2898 | // Now, we can replace the uses. | |||
2899 | Inst->replaceAllUsesWith(New); | |||
2900 | } | |||
2901 | ||||
2902 | /// \brief Reassign the original uses of Inst to Inst. | |||
2903 | void undo() override { | |||
2904 | DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: UsersReplacer: " << *Inst << "\n"; } } while (false); | |||
2905 | for (use_iterator UseIt = OriginalUses.begin(), | |||
2906 | EndIt = OriginalUses.end(); | |||
2907 | UseIt != EndIt; ++UseIt) { | |||
2908 | UseIt->Inst->setOperand(UseIt->Idx, Inst); | |||
2909 | } | |||
2910 | } | |||
2911 | }; | |||
2912 | ||||
2913 | /// \brief Remove an instruction from the IR. | |||
2914 | class InstructionRemover : public TypePromotionAction { | |||
2915 | /// Original position of the instruction. | |||
2916 | InsertionHandler Inserter; | |||
2917 | /// Helper structure to hide all the link to the instruction. In other | |||
2918 | /// words, this helps to do as if the instruction was removed. | |||
2919 | OperandsHider Hider; | |||
2920 | /// Keep track of the uses replaced, if any. | |||
2921 | UsesReplacer *Replacer; | |||
2922 | /// Keep track of instructions removed. | |||
2923 | SetOfInstrs &RemovedInsts; | |||
2924 | ||||
2925 | public: | |||
2926 | /// \brief Remove all reference of \p Inst and optinally replace all its | |||
2927 | /// uses with New. | |||
2928 | /// \p RemovedInsts Keep track of the instructions removed by this Action. | |||
2929 | /// \pre If !Inst->use_empty(), then New != nullptr | |||
2930 | InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts, | |||
2931 | Value *New = nullptr) | |||
2932 | : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst), | |||
2933 | Replacer(nullptr), RemovedInsts(RemovedInsts) { | |||
2934 | if (New) | |||
2935 | Replacer = new UsesReplacer(Inst, New); | |||
2936 | DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Do: InstructionRemover: " << *Inst << "\n"; } } while (false); | |||
2937 | RemovedInsts.insert(Inst); | |||
2938 | /// The instructions removed here will be freed after completing | |||
2939 | /// optimizeBlock() for all blocks as we need to keep track of the | |||
2940 | /// removed instructions during promotion. | |||
2941 | Inst->removeFromParent(); | |||
2942 | } | |||
2943 | ||||
2944 | ~InstructionRemover() override { delete Replacer; } | |||
2945 | ||||
2946 | /// \brief Resurrect the instruction and reassign it to the proper uses if | |||
2947 | /// new value was provided when build this action. | |||
2948 | void undo() override { | |||
2949 | DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Undo: InstructionRemover: " << *Inst << "\n"; } } while (false); | |||
2950 | Inserter.insert(Inst); | |||
2951 | if (Replacer) | |||
2952 | Replacer->undo(); | |||
2953 | Hider.undo(); | |||
2954 | RemovedInsts.erase(Inst); | |||
2955 | } | |||
2956 | }; | |||
2957 | ||||
2958 | public: | |||
2959 | /// Restoration point. | |||
2960 | /// The restoration point is a pointer to an action instead of an iterator | |||
2961 | /// because the iterator may be invalidated but not the pointer. | |||
2962 | typedef const TypePromotionAction *ConstRestorationPt; | |||
2963 | ||||
2964 | TypePromotionTransaction(SetOfInstrs &RemovedInsts) | |||
2965 | : RemovedInsts(RemovedInsts) {} | |||
2966 | ||||
2967 | /// Advocate every changes made in that transaction. | |||
2968 | void commit(); | |||
2969 | /// Undo all the changes made after the given point. | |||
2970 | void rollback(ConstRestorationPt Point); | |||
2971 | /// Get the current restoration point. | |||
2972 | ConstRestorationPt getRestorationPoint() const; | |||
2973 | ||||
2974 | /// \name API for IR modification with state keeping to support rollback. | |||
2975 | /// @{ | |||
2976 | /// Same as Instruction::setOperand. | |||
2977 | void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal); | |||
2978 | /// Same as Instruction::eraseFromParent. | |||
2979 | void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr); | |||
2980 | /// Same as Value::replaceAllUsesWith. | |||
2981 | void replaceAllUsesWith(Instruction *Inst, Value *New); | |||
2982 | /// Same as Value::mutateType. | |||
2983 | void mutateType(Instruction *Inst, Type *NewTy); | |||
2984 | /// Same as IRBuilder::createTrunc. | |||
2985 | Value *createTrunc(Instruction *Opnd, Type *Ty); | |||
2986 | /// Same as IRBuilder::createSExt. | |||
2987 | Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty); | |||
2988 | /// Same as IRBuilder::createZExt. | |||
2989 | Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty); | |||
2990 | /// Same as Instruction::moveBefore. | |||
2991 | void moveBefore(Instruction *Inst, Instruction *Before); | |||
2992 | /// @} | |||
2993 | ||||
2994 | private: | |||
2995 | /// The ordered list of actions made so far. | |||
2996 | SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions; | |||
2997 | typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt; | |||
2998 | SetOfInstrs &RemovedInsts; | |||
2999 | }; | |||
3000 | ||||
3001 | void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx, | |||
3002 | Value *NewVal) { | |||
3003 | Actions.push_back( | |||
3004 | make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal)); | |||
3005 | } | |||
3006 | ||||
3007 | void TypePromotionTransaction::eraseInstruction(Instruction *Inst, | |||
3008 | Value *NewVal) { | |||
3009 | Actions.push_back( | |||
3010 | make_unique<TypePromotionTransaction::InstructionRemover>(Inst, | |||
3011 | RemovedInsts, NewVal)); | |||
3012 | } | |||
3013 | ||||
3014 | void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst, | |||
3015 | Value *New) { | |||
3016 | Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New)); | |||
3017 | } | |||
3018 | ||||
3019 | void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) { | |||
3020 | Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy)); | |||
3021 | } | |||
3022 | ||||
3023 | Value *TypePromotionTransaction::createTrunc(Instruction *Opnd, | |||
3024 | Type *Ty) { | |||
3025 | std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty)); | |||
3026 | Value *Val = Ptr->getBuiltValue(); | |||
3027 | Actions.push_back(std::move(Ptr)); | |||
3028 | return Val; | |||
3029 | } | |||
3030 | ||||
3031 | Value *TypePromotionTransaction::createSExt(Instruction *Inst, | |||
3032 | Value *Opnd, Type *Ty) { | |||
3033 | std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty)); | |||
3034 | Value *Val = Ptr->getBuiltValue(); | |||
3035 | Actions.push_back(std::move(Ptr)); | |||
3036 | return Val; | |||
3037 | } | |||
3038 | ||||
3039 | Value *TypePromotionTransaction::createZExt(Instruction *Inst, | |||
3040 | Value *Opnd, Type *Ty) { | |||
3041 | std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty)); | |||
3042 | Value *Val = Ptr->getBuiltValue(); | |||
3043 | Actions.push_back(std::move(Ptr)); | |||
3044 | return Val; | |||
3045 | } | |||
3046 | ||||
3047 | void TypePromotionTransaction::moveBefore(Instruction *Inst, | |||
3048 | Instruction *Before) { | |||
3049 | Actions.push_back( | |||
3050 | make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before)); | |||
3051 | } | |||
3052 | ||||
3053 | TypePromotionTransaction::ConstRestorationPt | |||
3054 | TypePromotionTransaction::getRestorationPoint() const { | |||
3055 | return !Actions.empty() ? Actions.back().get() : nullptr; | |||
3056 | } | |||
3057 | ||||
3058 | void TypePromotionTransaction::commit() { | |||
3059 | for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt; | |||
3060 | ++It) | |||
3061 | (*It)->commit(); | |||
3062 | Actions.clear(); | |||
3063 | } | |||
3064 | ||||
3065 | void TypePromotionTransaction::rollback( | |||
3066 | TypePromotionTransaction::ConstRestorationPt Point) { | |||
3067 | while (!Actions.empty() && Point != Actions.back().get()) { | |||
3068 | std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val(); | |||
3069 | Curr->undo(); | |||
3070 | } | |||
3071 | } | |||
3072 | ||||
3073 | /// \brief A helper class for matching addressing modes. | |||
3074 | /// | |||
3075 | /// This encapsulates the logic for matching the target-legal addressing modes. | |||
3076 | class AddressingModeMatcher { | |||
3077 | SmallVectorImpl<Instruction*> &AddrModeInsts; | |||
3078 | const TargetLowering &TLI; | |||
3079 | const TargetRegisterInfo &TRI; | |||
3080 | const DataLayout &DL; | |||
3081 | ||||
3082 | /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and | |||
3083 | /// the memory instruction that we're computing this address for. | |||
3084 | Type *AccessTy; | |||
3085 | unsigned AddrSpace; | |||
3086 | Instruction *MemoryInst; | |||
3087 | ||||
3088 | /// This is the addressing mode that we're building up. This is | |||
3089 | /// part of the return value of this addressing mode matching stuff. | |||
3090 | ExtAddrMode &AddrMode; | |||
3091 | ||||
3092 | /// The instructions inserted by other CodeGenPrepare optimizations. | |||
3093 | const SetOfInstrs &InsertedInsts; | |||
3094 | /// A map from the instructions to their type before promotion. | |||
3095 | InstrToOrigTy &PromotedInsts; | |||
3096 | /// The ongoing transaction where every action should be registered. | |||
3097 | TypePromotionTransaction &TPT; | |||
3098 | ||||
3099 | /// This is set to true when we should not do profitability checks. | |||
3100 | /// When true, IsProfitableToFoldIntoAddressingMode always returns true. | |||
3101 | bool IgnoreProfitability; | |||
3102 | ||||
3103 | AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI, | |||
3104 | const TargetLowering &TLI, | |||
3105 | const TargetRegisterInfo &TRI, | |||
3106 | Type *AT, unsigned AS, | |||
3107 | Instruction *MI, ExtAddrMode &AM, | |||
3108 | const SetOfInstrs &InsertedInsts, | |||
3109 | InstrToOrigTy &PromotedInsts, | |||
3110 | TypePromotionTransaction &TPT) | |||
3111 | : AddrModeInsts(AMI), TLI(TLI), TRI(TRI), | |||
3112 | DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS), | |||
3113 | MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts), | |||
3114 | PromotedInsts(PromotedInsts), TPT(TPT) { | |||
3115 | IgnoreProfitability = false; | |||
3116 | } | |||
3117 | public: | |||
3118 | ||||
3119 | /// Find the maximal addressing mode that a load/store of V can fold, | |||
3120 | /// give an access type of AccessTy. This returns a list of involved | |||
3121 | /// instructions in AddrModeInsts. | |||
3122 | /// \p InsertedInsts The instructions inserted by other CodeGenPrepare | |||
3123 | /// optimizations. | |||
3124 | /// \p PromotedInsts maps the instructions to their type before promotion. | |||
3125 | /// \p The ongoing transaction where every action should be registered. | |||
3126 | static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS, | |||
3127 | Instruction *MemoryInst, | |||
3128 | SmallVectorImpl<Instruction*> &AddrModeInsts, | |||
3129 | const TargetLowering &TLI, | |||
3130 | const TargetRegisterInfo &TRI, | |||
3131 | const SetOfInstrs &InsertedInsts, | |||
3132 | InstrToOrigTy &PromotedInsts, | |||
3133 | TypePromotionTransaction &TPT) { | |||
3134 | ExtAddrMode Result; | |||
3135 | ||||
3136 | bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI, | |||
3137 | AccessTy, AS, | |||
3138 | MemoryInst, Result, InsertedInsts, | |||
3139 | PromotedInsts, TPT).matchAddr(V, 0); | |||
3140 | (void)Success; assert(Success && "Couldn't select *anything*?")((Success && "Couldn't select *anything*?") ? static_cast <void> (0) : __assert_fail ("Success && \"Couldn't select *anything*?\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 3140, __PRETTY_FUNCTION__)); | |||
3141 | return Result; | |||
3142 | } | |||
3143 | private: | |||
3144 | bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth); | |||
3145 | bool matchAddr(Value *V, unsigned Depth); | |||
3146 | bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth, | |||
3147 | bool *MovedAway = nullptr); | |||
3148 | bool isProfitableToFoldIntoAddressingMode(Instruction *I, | |||
3149 | ExtAddrMode &AMBefore, | |||
3150 | ExtAddrMode &AMAfter); | |||
3151 | bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2); | |||
3152 | bool isPromotionProfitable(unsigned NewCost, unsigned OldCost, | |||
3153 | Value *PromotedOperand) const; | |||
3154 | }; | |||
3155 | ||||
3156 | /// Try adding ScaleReg*Scale to the current addressing mode. | |||
3157 | /// Return true and update AddrMode if this addr mode is legal for the target, | |||
3158 | /// false if not. | |||
3159 | bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale, | |||
3160 | unsigned Depth) { | |||
3161 | // If Scale is 1, then this is the same as adding ScaleReg to the addressing | |||
3162 | // mode. Just process that directly. | |||
3163 | if (Scale == 1) | |||
3164 | return matchAddr(ScaleReg, Depth); | |||
3165 | ||||
3166 | // If the scale is 0, it takes nothing to add this. | |||
3167 | if (Scale == 0) | |||
3168 | return true; | |||
3169 | ||||
3170 | // If we already have a scale of this value, we can add to it, otherwise, we | |||
3171 | // need an available scale field. | |||
3172 | if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg) | |||
3173 | return false; | |||
3174 | ||||
3175 | ExtAddrMode TestAddrMode = AddrMode; | |||
3176 | ||||
3177 | // Add scale to turn X*4+X*3 -> X*7. This could also do things like | |||
3178 | // [A+B + A*7] -> [B+A*8]. | |||
3179 | TestAddrMode.Scale += Scale; | |||
3180 | TestAddrMode.ScaledReg = ScaleReg; | |||
3181 | ||||
3182 | // If the new address isn't legal, bail out. | |||
3183 | if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) | |||
3184 | return false; | |||
3185 | ||||
3186 | // It was legal, so commit it. | |||
3187 | AddrMode = TestAddrMode; | |||
3188 | ||||
3189 | // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now | |||
3190 | // to see if ScaleReg is actually X+C. If so, we can turn this into adding | |||
3191 | // X*Scale + C*Scale to addr mode. | |||
3192 | ConstantInt *CI = nullptr; Value *AddLHS = nullptr; | |||
3193 | if (isa<Instruction>(ScaleReg) && // not a constant expr. | |||
3194 | match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) { | |||
3195 | TestAddrMode.ScaledReg = AddLHS; | |||
3196 | TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale; | |||
3197 | ||||
3198 | // If this addressing mode is legal, commit it and remember that we folded | |||
3199 | // this instruction. | |||
3200 | if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) { | |||
3201 | AddrModeInsts.push_back(cast<Instruction>(ScaleReg)); | |||
3202 | AddrMode = TestAddrMode; | |||
3203 | return true; | |||
3204 | } | |||
3205 | } | |||
3206 | ||||
3207 | // Otherwise, not (x+c)*scale, just return what we have. | |||
3208 | return true; | |||
3209 | } | |||
3210 | ||||
3211 | /// This is a little filter, which returns true if an addressing computation | |||
3212 | /// involving I might be folded into a load/store accessing it. | |||
3213 | /// This doesn't need to be perfect, but needs to accept at least | |||
3214 | /// the set of instructions that MatchOperationAddr can. | |||
3215 | static bool MightBeFoldableInst(Instruction *I) { | |||
3216 | switch (I->getOpcode()) { | |||
3217 | case Instruction::BitCast: | |||
3218 | case Instruction::AddrSpaceCast: | |||
3219 | // Don't touch identity bitcasts. | |||
3220 | if (I->getType() == I->getOperand(0)->getType()) | |||
3221 | return false; | |||
3222 | return I->getType()->isPointerTy() || I->getType()->isIntegerTy(); | |||
3223 | case Instruction::PtrToInt: | |||
3224 | // PtrToInt is always a noop, as we know that the int type is pointer sized. | |||
3225 | return true; | |||
3226 | case Instruction::IntToPtr: | |||
3227 | // We know the input is intptr_t, so this is foldable. | |||
3228 | return true; | |||
3229 | case Instruction::Add: | |||
3230 | return true; | |||
3231 | case Instruction::Mul: | |||
3232 | case Instruction::Shl: | |||
3233 | // Can only handle X*C and X << C. | |||
3234 | return isa<ConstantInt>(I->getOperand(1)); | |||
3235 | case Instruction::GetElementPtr: | |||
3236 | return true; | |||
3237 | default: | |||
3238 | return false; | |||
3239 | } | |||
3240 | } | |||
3241 | ||||
3242 | /// \brief Check whether or not \p Val is a legal instruction for \p TLI. | |||
3243 | /// \note \p Val is assumed to be the product of some type promotion. | |||
3244 | /// Therefore if \p Val has an undefined state in \p TLI, this is assumed | |||
3245 | /// to be legal, as the non-promoted value would have had the same state. | |||
3246 | static bool isPromotedInstructionLegal(const TargetLowering &TLI, | |||
3247 | const DataLayout &DL, Value *Val) { | |||
3248 | Instruction *PromotedInst = dyn_cast<Instruction>(Val); | |||
3249 | if (!PromotedInst) | |||
3250 | return false; | |||
3251 | int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode()); | |||
3252 | // If the ISDOpcode is undefined, it was undefined before the promotion. | |||
3253 | if (!ISDOpcode) | |||
3254 | return true; | |||
3255 | // Otherwise, check if the promoted instruction is legal or not. | |||
3256 | return TLI.isOperationLegalOrCustom( | |||
3257 | ISDOpcode, TLI.getValueType(DL, PromotedInst->getType())); | |||
3258 | } | |||
3259 | ||||
3260 | /// \brief Hepler class to perform type promotion. | |||
3261 | class TypePromotionHelper { | |||
3262 | /// \brief Utility function to check whether or not a sign or zero extension | |||
3263 | /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by | |||
3264 | /// either using the operands of \p Inst or promoting \p Inst. | |||
3265 | /// The type of the extension is defined by \p IsSExt. | |||
3266 | /// In other words, check if: | |||
3267 | /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType. | |||
3268 | /// #1 Promotion applies: | |||
3269 | /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...). | |||
3270 | /// #2 Operand reuses: | |||
3271 | /// ext opnd1 to ConsideredExtType. | |||
3272 | /// \p PromotedInsts maps the instructions to their type before promotion. | |||
3273 | static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType, | |||
3274 | const InstrToOrigTy &PromotedInsts, bool IsSExt); | |||
3275 | ||||
3276 | /// \brief Utility function to determine if \p OpIdx should be promoted when | |||
3277 | /// promoting \p Inst. | |||
3278 | static bool shouldExtOperand(const Instruction *Inst, int OpIdx) { | |||
3279 | return !(isa<SelectInst>(Inst) && OpIdx == 0); | |||
3280 | } | |||
3281 | ||||
3282 | /// \brief Utility function to promote the operand of \p Ext when this | |||
3283 | /// operand is a promotable trunc or sext or zext. | |||
3284 | /// \p PromotedInsts maps the instructions to their type before promotion. | |||
3285 | /// \p CreatedInstsCost[out] contains the cost of all instructions | |||
3286 | /// created to promote the operand of Ext. | |||
3287 | /// Newly added extensions are inserted in \p Exts. | |||
3288 | /// Newly added truncates are inserted in \p Truncs. | |||
3289 | /// Should never be called directly. | |||
3290 | /// \return The promoted value which is used instead of Ext. | |||
3291 | static Value *promoteOperandForTruncAndAnyExt( | |||
3292 | Instruction *Ext, TypePromotionTransaction &TPT, | |||
3293 | InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, | |||
3294 | SmallVectorImpl<Instruction *> *Exts, | |||
3295 | SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI); | |||
3296 | ||||
3297 | /// \brief Utility function to promote the operand of \p Ext when this | |||
3298 | /// operand is promotable and is not a supported trunc or sext. | |||
3299 | /// \p PromotedInsts maps the instructions to their type before promotion. | |||
3300 | /// \p CreatedInstsCost[out] contains the cost of all the instructions | |||
3301 | /// created to promote the operand of Ext. | |||
3302 | /// Newly added extensions are inserted in \p Exts. | |||
3303 | /// Newly added truncates are inserted in \p Truncs. | |||
3304 | /// Should never be called directly. | |||
3305 | /// \return The promoted value which is used instead of Ext. | |||
3306 | static Value *promoteOperandForOther(Instruction *Ext, | |||
3307 | TypePromotionTransaction &TPT, | |||
3308 | InstrToOrigTy &PromotedInsts, | |||
3309 | unsigned &CreatedInstsCost, | |||
3310 | SmallVectorImpl<Instruction *> *Exts, | |||
3311 | SmallVectorImpl<Instruction *> *Truncs, | |||
3312 | const TargetLowering &TLI, bool IsSExt); | |||
3313 | ||||
3314 | /// \see promoteOperandForOther. | |||
3315 | static Value *signExtendOperandForOther( | |||
3316 | Instruction *Ext, TypePromotionTransaction &TPT, | |||
3317 | InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, | |||
3318 | SmallVectorImpl<Instruction *> *Exts, | |||
3319 | SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) { | |||
3320 | return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost, | |||
3321 | Exts, Truncs, TLI, true); | |||
3322 | } | |||
3323 | ||||
3324 | /// \see promoteOperandForOther. | |||
3325 | static Value *zeroExtendOperandForOther( | |||
3326 | Instruction *Ext, TypePromotionTransaction &TPT, | |||
3327 | InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, | |||
3328 | SmallVectorImpl<Instruction *> *Exts, | |||
3329 | SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) { | |||
3330 | return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost, | |||
3331 | Exts, Truncs, TLI, false); | |||
3332 | } | |||
3333 | ||||
3334 | public: | |||
3335 | /// Type for the utility function that promotes the operand of Ext. | |||
3336 | typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT, | |||
3337 | InstrToOrigTy &PromotedInsts, | |||
3338 | unsigned &CreatedInstsCost, | |||
3339 | SmallVectorImpl<Instruction *> *Exts, | |||
3340 | SmallVectorImpl<Instruction *> *Truncs, | |||
3341 | const TargetLowering &TLI); | |||
3342 | /// \brief Given a sign/zero extend instruction \p Ext, return the approriate | |||
3343 | /// action to promote the operand of \p Ext instead of using Ext. | |||
3344 | /// \return NULL if no promotable action is possible with the current | |||
3345 | /// sign extension. | |||
3346 | /// \p InsertedInsts keeps track of all the instructions inserted by the | |||
3347 | /// other CodeGenPrepare optimizations. This information is important | |||
3348 | /// because we do not want to promote these instructions as CodeGenPrepare | |||
3349 | /// will reinsert them later. Thus creating an infinite loop: create/remove. | |||
3350 | /// \p PromotedInsts maps the instructions to their type before promotion. | |||
3351 | static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts, | |||
3352 | const TargetLowering &TLI, | |||
3353 | const InstrToOrigTy &PromotedInsts); | |||
3354 | }; | |||
3355 | ||||
3356 | bool TypePromotionHelper::canGetThrough(const Instruction *Inst, | |||
3357 | Type *ConsideredExtType, | |||
3358 | const InstrToOrigTy &PromotedInsts, | |||
3359 | bool IsSExt) { | |||
3360 | // The promotion helper does not know how to deal with vector types yet. | |||
3361 | // To be able to fix that, we would need to fix the places where we | |||
3362 | // statically extend, e.g., constants and such. | |||
3363 | if (Inst->getType()->isVectorTy()) | |||
3364 | return false; | |||
3365 | ||||
3366 | // We can always get through zext. | |||
3367 | if (isa<ZExtInst>(Inst)) | |||
3368 | return true; | |||
3369 | ||||
3370 | // sext(sext) is ok too. | |||
3371 | if (IsSExt && isa<SExtInst>(Inst)) | |||
3372 | return true; | |||
3373 | ||||
3374 | // We can get through binary operator, if it is legal. In other words, the | |||
3375 | // binary operator must have a nuw or nsw flag. | |||
3376 | const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst); | |||
3377 | if (BinOp && isa<OverflowingBinaryOperator>(BinOp) && | |||
3378 | ((!IsSExt && BinOp->hasNoUnsignedWrap()) || | |||
3379 | (IsSExt && BinOp->hasNoSignedWrap()))) | |||
3380 | return true; | |||
3381 | ||||
3382 | // Check if we can do the following simplification. | |||
3383 | // ext(trunc(opnd)) --> ext(opnd) | |||
3384 | if (!isa<TruncInst>(Inst)) | |||
3385 | return false; | |||
3386 | ||||
3387 | Value *OpndVal = Inst->getOperand(0); | |||
3388 | // Check if we can use this operand in the extension. | |||
3389 | // If the type is larger than the result type of the extension, we cannot. | |||
3390 | if (!OpndVal->getType()->isIntegerTy() || | |||
3391 | OpndVal->getType()->getIntegerBitWidth() > | |||
3392 | ConsideredExtType->getIntegerBitWidth()) | |||
3393 | return false; | |||
3394 | ||||
3395 | // If the operand of the truncate is not an instruction, we will not have | |||
3396 | // any information on the dropped bits. | |||
3397 | // (Actually we could for constant but it is not worth the extra logic). | |||
3398 | Instruction *Opnd = dyn_cast<Instruction>(OpndVal); | |||
3399 | if (!Opnd) | |||
3400 | return false; | |||
3401 | ||||
3402 | // Check if the source of the type is narrow enough. | |||
3403 | // I.e., check that trunc just drops extended bits of the same kind of | |||
3404 | // the extension. | |||
3405 | // #1 get the type of the operand and check the kind of the extended bits. | |||
3406 | const Type *OpndType; | |||
3407 | InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd); | |||
3408 | if (It != PromotedInsts.end() && It->second.getInt() == IsSExt) | |||
3409 | OpndType = It->second.getPointer(); | |||
3410 | else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd))) | |||
3411 | OpndType = Opnd->getOperand(0)->getType(); | |||
3412 | else | |||
3413 | return false; | |||
3414 | ||||
3415 | // #2 check that the truncate just drops extended bits. | |||
3416 | return Inst->getType()->getIntegerBitWidth() >= | |||
3417 | OpndType->getIntegerBitWidth(); | |||
3418 | } | |||
3419 | ||||
3420 | TypePromotionHelper::Action TypePromotionHelper::getAction( | |||
3421 | Instruction *Ext, const SetOfInstrs &InsertedInsts, | |||
3422 | const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) { | |||
3423 | assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&(((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && "Unexpected instruction type") ? static_cast<void> (0) : __assert_fail ("(isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && \"Unexpected instruction type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 3424, __PRETTY_FUNCTION__)) | |||
3424 | "Unexpected instruction type")(((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && "Unexpected instruction type") ? static_cast<void> (0) : __assert_fail ("(isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && \"Unexpected instruction type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 3424, __PRETTY_FUNCTION__)); | |||
3425 | Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0)); | |||
3426 | Type *ExtTy = Ext->getType(); | |||
3427 | bool IsSExt = isa<SExtInst>(Ext); | |||
3428 | // If the operand of the extension is not an instruction, we cannot | |||
3429 | // get through. | |||
3430 | // If it, check we can get through. | |||
3431 | if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt)) | |||
3432 | return nullptr; | |||
3433 | ||||
3434 | // Do not promote if the operand has been added by codegenprepare. | |||
3435 | // Otherwise, it means we are undoing an optimization that is likely to be | |||
3436 | // redone, thus causing potential infinite loop. | |||
3437 | if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd)) | |||
3438 | return nullptr; | |||
3439 | ||||
3440 | // SExt or Trunc instructions. | |||
3441 | // Return the related handler. | |||
3442 | if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) || | |||
3443 | isa<ZExtInst>(ExtOpnd)) | |||
3444 | return promoteOperandForTruncAndAnyExt; | |||
3445 | ||||
3446 | // Regular instruction. | |||
3447 | // Abort early if we will have to insert non-free instructions. | |||
3448 | if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType())) | |||
3449 | return nullptr; | |||
3450 | return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther; | |||
3451 | } | |||
3452 | ||||
3453 | Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt( | |||
3454 | llvm::Instruction *SExt, TypePromotionTransaction &TPT, | |||
3455 | InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, | |||
3456 | SmallVectorImpl<Instruction *> *Exts, | |||
3457 | SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) { | |||
3458 | // By construction, the operand of SExt is an instruction. Otherwise we cannot | |||
3459 | // get through it and this method should not be called. | |||
3460 | Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0)); | |||
3461 | Value *ExtVal = SExt; | |||
3462 | bool HasMergedNonFreeExt = false; | |||
3463 | if (isa<ZExtInst>(SExtOpnd)) { | |||
3464 | // Replace s|zext(zext(opnd)) | |||
3465 | // => zext(opnd). | |||
3466 | HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd); | |||
3467 | Value *ZExt = | |||
3468 | TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType()); | |||
3469 | TPT.replaceAllUsesWith(SExt, ZExt); | |||
3470 | TPT.eraseInstruction(SExt); | |||
3471 | ExtVal = ZExt; | |||
3472 | } else { | |||
3473 | // Replace z|sext(trunc(opnd)) or sext(sext(opnd)) | |||
3474 | // => z|sext(opnd). | |||
3475 | TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0)); | |||
3476 | } | |||
3477 | CreatedInstsCost = 0; | |||
3478 | ||||
3479 | // Remove dead code. | |||
3480 | if (SExtOpnd->use_empty()) | |||
3481 | TPT.eraseInstruction(SExtOpnd); | |||
3482 | ||||
3483 | // Check if the extension is still needed. | |||
3484 | Instruction *ExtInst = dyn_cast<Instruction>(ExtVal); | |||
3485 | if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) { | |||
3486 | if (ExtInst) { | |||
3487 | if (Exts) | |||
3488 | Exts->push_back(ExtInst); | |||
3489 | CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt; | |||
3490 | } | |||
3491 | return ExtVal; | |||
3492 | } | |||
3493 | ||||
3494 | // At this point we have: ext ty opnd to ty. | |||
3495 | // Reassign the uses of ExtInst to the opnd and remove ExtInst. | |||
3496 | Value *NextVal = ExtInst->getOperand(0); | |||
3497 | TPT.eraseInstruction(ExtInst, NextVal); | |||
3498 | return NextVal; | |||
3499 | } | |||
3500 | ||||
3501 | Value *TypePromotionHelper::promoteOperandForOther( | |||
3502 | Instruction *Ext, TypePromotionTransaction &TPT, | |||
3503 | InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, | |||
3504 | SmallVectorImpl<Instruction *> *Exts, | |||
3505 | SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI, | |||
3506 | bool IsSExt) { | |||
3507 | // By construction, the operand of Ext is an instruction. Otherwise we cannot | |||
3508 | // get through it and this method should not be called. | |||
3509 | Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0)); | |||
3510 | CreatedInstsCost = 0; | |||
3511 | if (!ExtOpnd->hasOneUse()) { | |||
3512 | // ExtOpnd will be promoted. | |||
3513 | // All its uses, but Ext, will need to use a truncated value of the | |||
3514 | // promoted version. | |||
3515 | // Create the truncate now. | |||
3516 | Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType()); | |||
3517 | if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) { | |||
3518 | ITrunc->removeFromParent(); | |||
3519 | // Insert it just after the definition. | |||
3520 | ITrunc->insertAfter(ExtOpnd); | |||
3521 | if (Truncs) | |||
3522 | Truncs->push_back(ITrunc); | |||
3523 | } | |||
3524 | ||||
3525 | TPT.replaceAllUsesWith(ExtOpnd, Trunc); | |||
3526 | // Restore the operand of Ext (which has been replaced by the previous call | |||
3527 | // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext. | |||
3528 | TPT.setOperand(Ext, 0, ExtOpnd); | |||
3529 | } | |||
3530 | ||||
3531 | // Get through the Instruction: | |||
3532 | // 1. Update its type. | |||
3533 | // 2. Replace the uses of Ext by Inst. | |||
3534 | // 3. Extend each operand that needs to be extended. | |||
3535 | ||||
3536 | // Remember the original type of the instruction before promotion. | |||
3537 | // This is useful to know that the high bits are sign extended bits. | |||
3538 | PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>( | |||
3539 | ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt))); | |||
3540 | // Step #1. | |||
3541 | TPT.mutateType(ExtOpnd, Ext->getType()); | |||
3542 | // Step #2. | |||
3543 | TPT.replaceAllUsesWith(Ext, ExtOpnd); | |||
3544 | // Step #3. | |||
3545 | Instruction *ExtForOpnd = Ext; | |||
3546 | ||||
3547 | DEBUG(dbgs() << "Propagate Ext to operands\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Propagate Ext to operands\n" ; } } while (false); | |||
3548 | for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx; | |||
3549 | ++OpIdx) { | |||
3550 | DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Operand:\n" << * (ExtOpnd->getOperand(OpIdx)) << '\n'; } } while (false ); | |||
3551 | if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() || | |||
3552 | !shouldExtOperand(ExtOpnd, OpIdx)) { | |||
3553 | DEBUG(dbgs() << "No need to propagate\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "No need to propagate\n" ; } } while (false); | |||
3554 | continue; | |||
3555 | } | |||
3556 | // Check if we can statically extend the operand. | |||
3557 | Value *Opnd = ExtOpnd->getOperand(OpIdx); | |||
3558 | if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) { | |||
3559 | DEBUG(dbgs() << "Statically extend\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Statically extend\n"; } } while (false); | |||
3560 | unsigned BitWidth = Ext->getType()->getIntegerBitWidth(); | |||
3561 | APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth) | |||
3562 | : Cst->getValue().zext(BitWidth); | |||
3563 | TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal)); | |||
3564 | continue; | |||
3565 | } | |||
3566 | // UndefValue are typed, so we have to statically sign extend them. | |||
3567 | if (isa<UndefValue>(Opnd)) { | |||
3568 | DEBUG(dbgs() << "Statically extend\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Statically extend\n"; } } while (false); | |||
3569 | TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType())); | |||
3570 | continue; | |||
3571 | } | |||
3572 | ||||
3573 | // Otherwise we have to explicity sign extend the operand. | |||
3574 | // Check if Ext was reused to extend an operand. | |||
3575 | if (!ExtForOpnd) { | |||
3576 | // If yes, create a new one. | |||
3577 | DEBUG(dbgs() << "More operands to ext\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "More operands to ext\n" ; } } while (false); | |||
3578 | Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType()) | |||
3579 | : TPT.createZExt(Ext, Opnd, Ext->getType()); | |||
3580 | if (!isa<Instruction>(ValForExtOpnd)) { | |||
3581 | TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd); | |||
3582 | continue; | |||
3583 | } | |||
3584 | ExtForOpnd = cast<Instruction>(ValForExtOpnd); | |||
3585 | } | |||
3586 | if (Exts) | |||
3587 | Exts->push_back(ExtForOpnd); | |||
3588 | TPT.setOperand(ExtForOpnd, 0, Opnd); | |||
3589 | ||||
3590 | // Move the sign extension before the insertion point. | |||
3591 | TPT.moveBefore(ExtForOpnd, ExtOpnd); | |||
3592 | TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd); | |||
3593 | CreatedInstsCost += !TLI.isExtFree(ExtForOpnd); | |||
3594 | // If more sext are required, new instructions will have to be created. | |||
3595 | ExtForOpnd = nullptr; | |||
3596 | } | |||
3597 | if (ExtForOpnd == Ext) { | |||
3598 | DEBUG(dbgs() << "Extension is useless now\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Extension is useless now\n" ; } } while (false); | |||
3599 | TPT.eraseInstruction(Ext); | |||
3600 | } | |||
3601 | return ExtOpnd; | |||
3602 | } | |||
3603 | ||||
3604 | /// Check whether or not promoting an instruction to a wider type is profitable. | |||
3605 | /// \p NewCost gives the cost of extension instructions created by the | |||
3606 | /// promotion. | |||
3607 | /// \p OldCost gives the cost of extension instructions before the promotion | |||
3608 | /// plus the number of instructions that have been | |||
3609 | /// matched in the addressing mode the promotion. | |||
3610 | /// \p PromotedOperand is the value that has been promoted. | |||
3611 | /// \return True if the promotion is profitable, false otherwise. | |||
3612 | bool AddressingModeMatcher::isPromotionProfitable( | |||
3613 | unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const { | |||
3614 | DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n'; } } while (false); | |||
3615 | // The cost of the new extensions is greater than the cost of the | |||
3616 | // old extension plus what we folded. | |||
3617 | // This is not profitable. | |||
3618 | if (NewCost > OldCost) | |||
3619 | return false; | |||
3620 | if (NewCost < OldCost) | |||
3621 | return true; | |||
3622 | // The promotion is neutral but it may help folding the sign extension in | |||
3623 | // loads for instance. | |||
3624 | // Check that we did not create an illegal instruction. | |||
3625 | return isPromotedInstructionLegal(TLI, DL, PromotedOperand); | |||
3626 | } | |||
3627 | ||||
3628 | /// Given an instruction or constant expr, see if we can fold the operation | |||
3629 | /// into the addressing mode. If so, update the addressing mode and return | |||
3630 | /// true, otherwise return false without modifying AddrMode. | |||
3631 | /// If \p MovedAway is not NULL, it contains the information of whether or | |||
3632 | /// not AddrInst has to be folded into the addressing mode on success. | |||
3633 | /// If \p MovedAway == true, \p AddrInst will not be part of the addressing | |||
3634 | /// because it has been moved away. | |||
3635 | /// Thus AddrInst must not be added in the matched instructions. | |||
3636 | /// This state can happen when AddrInst is a sext, since it may be moved away. | |||
3637 | /// Therefore, AddrInst may not be valid when MovedAway is true and it must | |||
3638 | /// not be referenced anymore. | |||
3639 | bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode, | |||
3640 | unsigned Depth, | |||
3641 | bool *MovedAway) { | |||
3642 | // Avoid exponential behavior on extremely deep expression trees. | |||
3643 | if (Depth >= 5) return false; | |||
3644 | ||||
3645 | // By default, all matched instructions stay in place. | |||
3646 | if (MovedAway) | |||
3647 | *MovedAway = false; | |||
3648 | ||||
3649 | switch (Opcode) { | |||
3650 | case Instruction::PtrToInt: | |||
3651 | // PtrToInt is always a noop, as we know that the int type is pointer sized. | |||
3652 | return matchAddr(AddrInst->getOperand(0), Depth); | |||
3653 | case Instruction::IntToPtr: { | |||
3654 | auto AS = AddrInst->getType()->getPointerAddressSpace(); | |||
3655 | auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS)); | |||
3656 | // This inttoptr is a no-op if the integer type is pointer sized. | |||
3657 | if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy) | |||
3658 | return matchAddr(AddrInst->getOperand(0), Depth); | |||
3659 | return false; | |||
3660 | } | |||
3661 | case Instruction::BitCast: | |||
3662 | // BitCast is always a noop, and we can handle it as long as it is | |||
3663 | // int->int or pointer->pointer (we don't want int<->fp or something). | |||
3664 | if ((AddrInst->getOperand(0)->getType()->isPointerTy() || | |||
3665 | AddrInst->getOperand(0)->getType()->isIntegerTy()) && | |||
3666 | // Don't touch identity bitcasts. These were probably put here by LSR, | |||
3667 | // and we don't want to mess around with them. Assume it knows what it | |||
3668 | // is doing. | |||
3669 | AddrInst->getOperand(0)->getType() != AddrInst->getType()) | |||
3670 | return matchAddr(AddrInst->getOperand(0), Depth); | |||
3671 | return false; | |||
3672 | case Instruction::AddrSpaceCast: { | |||
3673 | unsigned SrcAS | |||
3674 | = AddrInst->getOperand(0)->getType()->getPointerAddressSpace(); | |||
3675 | unsigned DestAS = AddrInst->getType()->getPointerAddressSpace(); | |||
3676 | if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS)) | |||
3677 | return matchAddr(AddrInst->getOperand(0), Depth); | |||
3678 | return false; | |||
3679 | } | |||
3680 | case Instruction::Add: { | |||
3681 | // Check to see if we can merge in the RHS then the LHS. If so, we win. | |||
3682 | ExtAddrMode BackupAddrMode = AddrMode; | |||
3683 | unsigned OldSize = AddrModeInsts.size(); | |||
3684 | // Start a transaction at this point. | |||
3685 | // The LHS may match but not the RHS. | |||
3686 | // Therefore, we need a higher level restoration point to undo partially | |||
3687 | // matched operation. | |||
3688 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | |||
3689 | TPT.getRestorationPoint(); | |||
3690 | ||||
3691 | if (matchAddr(AddrInst->getOperand(1), Depth+1) && | |||
3692 | matchAddr(AddrInst->getOperand(0), Depth+1)) | |||
3693 | return true; | |||
3694 | ||||
3695 | // Restore the old addr mode info. | |||
3696 | AddrMode = BackupAddrMode; | |||
3697 | AddrModeInsts.resize(OldSize); | |||
3698 | TPT.rollback(LastKnownGood); | |||
3699 | ||||
3700 | // Otherwise this was over-aggressive. Try merging in the LHS then the RHS. | |||
3701 | if (matchAddr(AddrInst->getOperand(0), Depth+1) && | |||
3702 | matchAddr(AddrInst->getOperand(1), Depth+1)) | |||
3703 | return true; | |||
3704 | ||||
3705 | // Otherwise we definitely can't merge the ADD in. | |||
3706 | AddrMode = BackupAddrMode; | |||
3707 | AddrModeInsts.resize(OldSize); | |||
3708 | TPT.rollback(LastKnownGood); | |||
3709 | break; | |||
3710 | } | |||
3711 | //case Instruction::Or: | |||
3712 | // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD. | |||
3713 | //break; | |||
3714 | case Instruction::Mul: | |||
3715 | case Instruction::Shl: { | |||
3716 | // Can only handle X*C and X << C. | |||
3717 | ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1)); | |||
3718 | if (!RHS) | |||
3719 | return false; | |||
3720 | int64_t Scale = RHS->getSExtValue(); | |||
3721 | if (Opcode == Instruction::Shl) | |||
3722 | Scale = 1LL << Scale; | |||
3723 | ||||
3724 | return matchScaledValue(AddrInst->getOperand(0), Scale, Depth); | |||
3725 | } | |||
3726 | case Instruction::GetElementPtr: { | |||
3727 | // Scan the GEP. We check it if it contains constant offsets and at most | |||
3728 | // one variable offset. | |||
3729 | int VariableOperand = -1; | |||
3730 | unsigned VariableScale = 0; | |||
3731 | ||||
3732 | int64_t ConstantOffset = 0; | |||
3733 | gep_type_iterator GTI = gep_type_begin(AddrInst); | |||
3734 | for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) { | |||
3735 | if (StructType *STy = GTI.getStructTypeOrNull()) { | |||
3736 | const StructLayout *SL = DL.getStructLayout(STy); | |||
3737 | unsigned Idx = | |||
3738 | cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue(); | |||
3739 | ConstantOffset += SL->getElementOffset(Idx); | |||
3740 | } else { | |||
3741 | uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType()); | |||
3742 | if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) { | |||
3743 | ConstantOffset += CI->getSExtValue()*TypeSize; | |||
3744 | } else if (TypeSize) { // Scales of zero don't do anything. | |||
3745 | // We only allow one variable index at the moment. | |||
3746 | if (VariableOperand != -1) | |||
3747 | return false; | |||
3748 | ||||
3749 | // Remember the variable index. | |||
3750 | VariableOperand = i; | |||
3751 | VariableScale = TypeSize; | |||
3752 | } | |||
3753 | } | |||
3754 | } | |||
3755 | ||||
3756 | // A common case is for the GEP to only do a constant offset. In this case, | |||
3757 | // just add it to the disp field and check validity. | |||
3758 | if (VariableOperand == -1) { | |||
3759 | AddrMode.BaseOffs += ConstantOffset; | |||
3760 | if (ConstantOffset == 0 || | |||
3761 | TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) { | |||
3762 | // Check to see if we can fold the base pointer in too. | |||
3763 | if (matchAddr(AddrInst->getOperand(0), Depth+1)) | |||
3764 | return true; | |||
3765 | } | |||
3766 | AddrMode.BaseOffs -= ConstantOffset; | |||
3767 | return false; | |||
3768 | } | |||
3769 | ||||
3770 | // Save the valid addressing mode in case we can't match. | |||
3771 | ExtAddrMode BackupAddrMode = AddrMode; | |||
3772 | unsigned OldSize = AddrModeInsts.size(); | |||
3773 | ||||
3774 | // See if the scale and offset amount is valid for this target. | |||
3775 | AddrMode.BaseOffs += ConstantOffset; | |||
3776 | ||||
3777 | // Match the base operand of the GEP. | |||
3778 | if (!matchAddr(AddrInst->getOperand(0), Depth+1)) { | |||
3779 | // If it couldn't be matched, just stuff the value in a register. | |||
3780 | if (AddrMode.HasBaseReg) { | |||
3781 | AddrMode = BackupAddrMode; | |||
3782 | AddrModeInsts.resize(OldSize); | |||
3783 | return false; | |||
3784 | } | |||
3785 | AddrMode.HasBaseReg = true; | |||
3786 | AddrMode.BaseReg = AddrInst->getOperand(0); | |||
3787 | } | |||
3788 | ||||
3789 | // Match the remaining variable portion of the GEP. | |||
3790 | if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale, | |||
3791 | Depth)) { | |||
3792 | // If it couldn't be matched, try stuffing the base into a register | |||
3793 | // instead of matching it, and retrying the match of the scale. | |||
3794 | AddrMode = BackupAddrMode; | |||
3795 | AddrModeInsts.resize(OldSize); | |||
3796 | if (AddrMode.HasBaseReg) | |||
3797 | return false; | |||
3798 | AddrMode.HasBaseReg = true; | |||
3799 | AddrMode.BaseReg = AddrInst->getOperand(0); | |||
3800 | AddrMode.BaseOffs += ConstantOffset; | |||
3801 | if (!matchScaledValue(AddrInst->getOperand(VariableOperand), | |||
3802 | VariableScale, Depth)) { | |||
3803 | // If even that didn't work, bail. | |||
3804 | AddrMode = BackupAddrMode; | |||
3805 | AddrModeInsts.resize(OldSize); | |||
3806 | return false; | |||
3807 | } | |||
3808 | } | |||
3809 | ||||
3810 | return true; | |||
3811 | } | |||
3812 | case Instruction::SExt: | |||
3813 | case Instruction::ZExt: { | |||
3814 | Instruction *Ext = dyn_cast<Instruction>(AddrInst); | |||
3815 | if (!Ext) | |||
3816 | return false; | |||
3817 | ||||
3818 | // Try to move this ext out of the way of the addressing mode. | |||
3819 | // Ask for a method for doing so. | |||
3820 | TypePromotionHelper::Action TPH = | |||
3821 | TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts); | |||
3822 | if (!TPH) | |||
3823 | return false; | |||
3824 | ||||
3825 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | |||
3826 | TPT.getRestorationPoint(); | |||
3827 | unsigned CreatedInstsCost = 0; | |||
3828 | unsigned ExtCost = !TLI.isExtFree(Ext); | |||
3829 | Value *PromotedOperand = | |||
3830 | TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI); | |||
3831 | // SExt has been moved away. | |||
3832 | // Thus either it will be rematched later in the recursive calls or it is | |||
3833 | // gone. Anyway, we must not fold it into the addressing mode at this point. | |||
3834 | // E.g., | |||
3835 | // op = add opnd, 1 | |||
3836 | // idx = ext op | |||
3837 | // addr = gep base, idx | |||
3838 | // is now: | |||
3839 | // promotedOpnd = ext opnd <- no match here | |||
3840 | // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls) | |||
3841 | // addr = gep base, op <- match | |||
3842 | if (MovedAway) | |||
3843 | *MovedAway = true; | |||
3844 | ||||
3845 | assert(PromotedOperand &&((PromotedOperand && "TypePromotionHelper should have filtered out those cases" ) ? static_cast<void> (0) : __assert_fail ("PromotedOperand && \"TypePromotionHelper should have filtered out those cases\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 3846, __PRETTY_FUNCTION__)) | |||
3846 | "TypePromotionHelper should have filtered out those cases")((PromotedOperand && "TypePromotionHelper should have filtered out those cases" ) ? static_cast<void> (0) : __assert_fail ("PromotedOperand && \"TypePromotionHelper should have filtered out those cases\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 3846, __PRETTY_FUNCTION__)); | |||
3847 | ||||
3848 | ExtAddrMode BackupAddrMode = AddrMode; | |||
3849 | unsigned OldSize = AddrModeInsts.size(); | |||
3850 | ||||
3851 | if (!matchAddr(PromotedOperand, Depth) || | |||
3852 | // The total of the new cost is equal to the cost of the created | |||
3853 | // instructions. | |||
3854 | // The total of the old cost is equal to the cost of the extension plus | |||
3855 | // what we have saved in the addressing mode. | |||
3856 | !isPromotionProfitable(CreatedInstsCost, | |||
3857 | ExtCost + (AddrModeInsts.size() - OldSize), | |||
3858 | PromotedOperand)) { | |||
3859 | AddrMode = BackupAddrMode; | |||
3860 | AddrModeInsts.resize(OldSize); | |||
3861 | DEBUG(dbgs() << "Sign extension does not pay off: rollback\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Sign extension does not pay off: rollback\n" ; } } while (false); | |||
3862 | TPT.rollback(LastKnownGood); | |||
3863 | return false; | |||
3864 | } | |||
3865 | return true; | |||
3866 | } | |||
3867 | } | |||
3868 | return false; | |||
3869 | } | |||
3870 | ||||
3871 | /// If we can, try to add the value of 'Addr' into the current addressing mode. | |||
3872 | /// If Addr can't be added to AddrMode this returns false and leaves AddrMode | |||
3873 | /// unmodified. This assumes that Addr is either a pointer type or intptr_t | |||
3874 | /// for the target. | |||
3875 | /// | |||
3876 | bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) { | |||
3877 | // Start a transaction at this point that we will rollback if the matching | |||
3878 | // fails. | |||
3879 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | |||
3880 | TPT.getRestorationPoint(); | |||
3881 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) { | |||
3882 | // Fold in immediates if legal for the target. | |||
3883 | AddrMode.BaseOffs += CI->getSExtValue(); | |||
3884 | if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) | |||
3885 | return true; | |||
3886 | AddrMode.BaseOffs -= CI->getSExtValue(); | |||
3887 | } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) { | |||
3888 | // If this is a global variable, try to fold it into the addressing mode. | |||
3889 | if (!AddrMode.BaseGV) { | |||
3890 | AddrMode.BaseGV = GV; | |||
3891 | if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) | |||
3892 | return true; | |||
3893 | AddrMode.BaseGV = nullptr; | |||
3894 | } | |||
3895 | } else if (Instruction *I = dyn_cast<Instruction>(Addr)) { | |||
3896 | ExtAddrMode BackupAddrMode = AddrMode; | |||
3897 | unsigned OldSize = AddrModeInsts.size(); | |||
3898 | ||||
3899 | // Check to see if it is possible to fold this operation. | |||
3900 | bool MovedAway = false; | |||
3901 | if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) { | |||
3902 | // This instruction may have been moved away. If so, there is nothing | |||
3903 | // to check here. | |||
3904 | if (MovedAway) | |||
3905 | return true; | |||
3906 | // Okay, it's possible to fold this. Check to see if it is actually | |||
3907 | // *profitable* to do so. We use a simple cost model to avoid increasing | |||
3908 | // register pressure too much. | |||
3909 | if (I->hasOneUse() || | |||
3910 | isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) { | |||
3911 | AddrModeInsts.push_back(I); | |||
3912 | return true; | |||
3913 | } | |||
3914 | ||||
3915 | // It isn't profitable to do this, roll back. | |||
3916 | //cerr << "NOT FOLDING: " << *I; | |||
3917 | AddrMode = BackupAddrMode; | |||
3918 | AddrModeInsts.resize(OldSize); | |||
3919 | TPT.rollback(LastKnownGood); | |||
3920 | } | |||
3921 | } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) { | |||
3922 | if (matchOperationAddr(CE, CE->getOpcode(), Depth)) | |||
3923 | return true; | |||
3924 | TPT.rollback(LastKnownGood); | |||
3925 | } else if (isa<ConstantPointerNull>(Addr)) { | |||
3926 | // Null pointer gets folded without affecting the addressing mode. | |||
3927 | return true; | |||
3928 | } | |||
3929 | ||||
3930 | // Worse case, the target should support [reg] addressing modes. :) | |||
3931 | if (!AddrMode.HasBaseReg) { | |||
3932 | AddrMode.HasBaseReg = true; | |||
3933 | AddrMode.BaseReg = Addr; | |||
3934 | // Still check for legality in case the target supports [imm] but not [i+r]. | |||
3935 | if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) | |||
3936 | return true; | |||
3937 | AddrMode.HasBaseReg = false; | |||
3938 | AddrMode.BaseReg = nullptr; | |||
3939 | } | |||
3940 | ||||
3941 | // If the base register is already taken, see if we can do [r+r]. | |||
3942 | if (AddrMode.Scale == 0) { | |||
3943 | AddrMode.Scale = 1; | |||
3944 | AddrMode.ScaledReg = Addr; | |||
3945 | if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) | |||
3946 | return true; | |||
3947 | AddrMode.Scale = 0; | |||
3948 | AddrMode.ScaledReg = nullptr; | |||
3949 | } | |||
3950 | // Couldn't match. | |||
3951 | TPT.rollback(LastKnownGood); | |||
3952 | return false; | |||
3953 | } | |||
3954 | ||||
3955 | /// Check to see if all uses of OpVal by the specified inline asm call are due | |||
3956 | /// to memory operands. If so, return true, otherwise return false. | |||
3957 | static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal, | |||
3958 | const TargetLowering &TLI, | |||
3959 | const TargetRegisterInfo &TRI) { | |||
3960 | const Function *F = CI->getFunction(); | |||
3961 | TargetLowering::AsmOperandInfoVector TargetConstraints = | |||
3962 | TLI.ParseConstraints(F->getParent()->getDataLayout(), &TRI, | |||
3963 | ImmutableCallSite(CI)); | |||
3964 | ||||
3965 | for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) { | |||
3966 | TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i]; | |||
3967 | ||||
3968 | // Compute the constraint code and ConstraintType to use. | |||
3969 | TLI.ComputeConstraintToUse(OpInfo, SDValue()); | |||
3970 | ||||
3971 | // If this asm operand is our Value*, and if it isn't an indirect memory | |||
3972 | // operand, we can't fold it! | |||
3973 | if (OpInfo.CallOperandVal == OpVal && | |||
3974 | (OpInfo.ConstraintType != TargetLowering::C_Memory || | |||
3975 | !OpInfo.isIndirect)) | |||
3976 | return false; | |||
3977 | } | |||
3978 | ||||
3979 | return true; | |||
3980 | } | |||
3981 | ||||
3982 | /// Recursively walk all the uses of I until we find a memory use. | |||
3983 | /// If we find an obviously non-foldable instruction, return true. | |||
3984 | /// Add the ultimately found memory instructions to MemoryUses. | |||
3985 | static bool FindAllMemoryUses( | |||
3986 | Instruction *I, | |||
3987 | SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses, | |||
3988 | SmallPtrSetImpl<Instruction *> &ConsideredInsts, | |||
3989 | const TargetLowering &TLI, const TargetRegisterInfo &TRI) { | |||
3990 | // If we already considered this instruction, we're done. | |||
3991 | if (!ConsideredInsts.insert(I).second) | |||
3992 | return false; | |||
3993 | ||||
3994 | // If this is an obviously unfoldable instruction, bail out. | |||
3995 | if (!MightBeFoldableInst(I)) | |||
3996 | return true; | |||
3997 | ||||
3998 | const bool OptSize = I->getFunction()->optForSize(); | |||
3999 | ||||
4000 | // Loop over all the uses, recursively processing them. | |||
4001 | for (Use &U : I->uses()) { | |||
4002 | Instruction *UserI = cast<Instruction>(U.getUser()); | |||
4003 | ||||
4004 | if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) { | |||
4005 | MemoryUses.push_back(std::make_pair(LI, U.getOperandNo())); | |||
4006 | continue; | |||
4007 | } | |||
4008 | ||||
4009 | if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) { | |||
4010 | unsigned opNo = U.getOperandNo(); | |||
4011 | if (opNo != StoreInst::getPointerOperandIndex()) | |||
4012 | return true; // Storing addr, not into addr. | |||
4013 | MemoryUses.push_back(std::make_pair(SI, opNo)); | |||
4014 | continue; | |||
4015 | } | |||
4016 | ||||
4017 | if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) { | |||
4018 | unsigned opNo = U.getOperandNo(); | |||
4019 | if (opNo != AtomicRMWInst::getPointerOperandIndex()) | |||
4020 | return true; // Storing addr, not into addr. | |||
4021 | MemoryUses.push_back(std::make_pair(RMW, opNo)); | |||
4022 | continue; | |||
4023 | } | |||
4024 | ||||
4025 | if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) { | |||
4026 | unsigned opNo = U.getOperandNo(); | |||
4027 | if (opNo != AtomicCmpXchgInst::getPointerOperandIndex()) | |||
4028 | return true; // Storing addr, not into addr. | |||
4029 | MemoryUses.push_back(std::make_pair(CmpX, opNo)); | |||
4030 | continue; | |||
4031 | } | |||
4032 | ||||
4033 | if (CallInst *CI = dyn_cast<CallInst>(UserI)) { | |||
4034 | // If this is a cold call, we can sink the addressing calculation into | |||
4035 | // the cold path. See optimizeCallInst | |||
4036 | if (!OptSize && CI->hasFnAttr(Attribute::Cold)) | |||
4037 | continue; | |||
4038 | ||||
4039 | InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue()); | |||
4040 | if (!IA) return true; | |||
4041 | ||||
4042 | // If this is a memory operand, we're cool, otherwise bail out. | |||
4043 | if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI)) | |||
4044 | return true; | |||
4045 | continue; | |||
4046 | } | |||
4047 | ||||
4048 | if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI)) | |||
4049 | return true; | |||
4050 | } | |||
4051 | ||||
4052 | return false; | |||
4053 | } | |||
4054 | ||||
4055 | /// Return true if Val is already known to be live at the use site that we're | |||
4056 | /// folding it into. If so, there is no cost to include it in the addressing | |||
4057 | /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the | |||
4058 | /// instruction already. | |||
4059 | bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1, | |||
4060 | Value *KnownLive2) { | |||
4061 | // If Val is either of the known-live values, we know it is live! | |||
4062 | if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2) | |||
4063 | return true; | |||
4064 | ||||
4065 | // All values other than instructions and arguments (e.g. constants) are live. | |||
4066 | if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true; | |||
4067 | ||||
4068 | // If Val is a constant sized alloca in the entry block, it is live, this is | |||
4069 | // true because it is just a reference to the stack/frame pointer, which is | |||
4070 | // live for the whole function. | |||
4071 | if (AllocaInst *AI = dyn_cast<AllocaInst>(Val)) | |||
4072 | if (AI->isStaticAlloca()) | |||
4073 | return true; | |||
4074 | ||||
4075 | // Check to see if this value is already used in the memory instruction's | |||
4076 | // block. If so, it's already live into the block at the very least, so we | |||
4077 | // can reasonably fold it. | |||
4078 | return Val->isUsedInBasicBlock(MemoryInst->getParent()); | |||
4079 | } | |||
4080 | ||||
4081 | /// It is possible for the addressing mode of the machine to fold the specified | |||
4082 | /// instruction into a load or store that ultimately uses it. | |||
4083 | /// However, the specified instruction has multiple uses. | |||
4084 | /// Given this, it may actually increase register pressure to fold it | |||
4085 | /// into the load. For example, consider this code: | |||
4086 | /// | |||
4087 | /// X = ... | |||
4088 | /// Y = X+1 | |||
4089 | /// use(Y) -> nonload/store | |||
4090 | /// Z = Y+1 | |||
4091 | /// load Z | |||
4092 | /// | |||
4093 | /// In this case, Y has multiple uses, and can be folded into the load of Z | |||
4094 | /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to | |||
4095 | /// be live at the use(Y) line. If we don't fold Y into load Z, we use one | |||
4096 | /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the | |||
4097 | /// number of computations either. | |||
4098 | /// | |||
4099 | /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If | |||
4100 | /// X was live across 'load Z' for other reasons, we actually *would* want to | |||
4101 | /// fold the addressing mode in the Z case. This would make Y die earlier. | |||
4102 | bool AddressingModeMatcher:: | |||
4103 | isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore, | |||
4104 | ExtAddrMode &AMAfter) { | |||
4105 | if (IgnoreProfitability) return true; | |||
4106 | ||||
4107 | // AMBefore is the addressing mode before this instruction was folded into it, | |||
4108 | // and AMAfter is the addressing mode after the instruction was folded. Get | |||
4109 | // the set of registers referenced by AMAfter and subtract out those | |||
4110 | // referenced by AMBefore: this is the set of values which folding in this | |||
4111 | // address extends the lifetime of. | |||
4112 | // | |||
4113 | // Note that there are only two potential values being referenced here, | |||
4114 | // BaseReg and ScaleReg (global addresses are always available, as are any | |||
4115 | // folded immediates). | |||
4116 | Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg; | |||
4117 | ||||
4118 | // If the BaseReg or ScaledReg was referenced by the previous addrmode, their | |||
4119 | // lifetime wasn't extended by adding this instruction. | |||
4120 | if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg)) | |||
4121 | BaseReg = nullptr; | |||
4122 | if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg)) | |||
4123 | ScaledReg = nullptr; | |||
4124 | ||||
4125 | // If folding this instruction (and it's subexprs) didn't extend any live | |||
4126 | // ranges, we're ok with it. | |||
4127 | if (!BaseReg && !ScaledReg) | |||
4128 | return true; | |||
4129 | ||||
4130 | // If all uses of this instruction can have the address mode sunk into them, | |||
4131 | // we can remove the addressing mode and effectively trade one live register | |||
4132 | // for another (at worst.) In this context, folding an addressing mode into | |||
4133 | // the use is just a particularly nice way of sinking it. | |||
4134 | SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses; | |||
4135 | SmallPtrSet<Instruction*, 16> ConsideredInsts; | |||
4136 | if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI)) | |||
4137 | return false; // Has a non-memory, non-foldable use! | |||
4138 | ||||
4139 | // Now that we know that all uses of this instruction are part of a chain of | |||
4140 | // computation involving only operations that could theoretically be folded | |||
4141 | // into a memory use, loop over each of these memory operation uses and see | |||
4142 | // if they could *actually* fold the instruction. The assumption is that | |||
4143 | // addressing modes are cheap and that duplicating the computation involved | |||
4144 | // many times is worthwhile, even on a fastpath. For sinking candidates | |||
4145 | // (i.e. cold call sites), this serves as a way to prevent excessive code | |||
4146 | // growth since most architectures have some reasonable small and fast way to | |||
4147 | // compute an effective address. (i.e LEA on x86) | |||
4148 | SmallVector<Instruction*, 32> MatchedAddrModeInsts; | |||
4149 | for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) { | |||
4150 | Instruction *User = MemoryUses[i].first; | |||
4151 | unsigned OpNo = MemoryUses[i].second; | |||
4152 | ||||
4153 | // Get the access type of this use. If the use isn't a pointer, we don't | |||
4154 | // know what it accesses. | |||
4155 | Value *Address = User->getOperand(OpNo); | |||
4156 | PointerType *AddrTy = dyn_cast<PointerType>(Address->getType()); | |||
4157 | if (!AddrTy) | |||
4158 | return false; | |||
4159 | Type *AddressAccessTy = AddrTy->getElementType(); | |||
4160 | unsigned AS = AddrTy->getAddressSpace(); | |||
4161 | ||||
4162 | // Do a match against the root of this address, ignoring profitability. This | |||
4163 | // will tell us if the addressing mode for the memory operation will | |||
4164 | // *actually* cover the shared instruction. | |||
4165 | ExtAddrMode Result; | |||
4166 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | |||
4167 | TPT.getRestorationPoint(); | |||
4168 | AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, TRI, | |||
4169 | AddressAccessTy, AS, | |||
4170 | MemoryInst, Result, InsertedInsts, | |||
4171 | PromotedInsts, TPT); | |||
4172 | Matcher.IgnoreProfitability = true; | |||
4173 | bool Success = Matcher.matchAddr(Address, 0); | |||
4174 | (void)Success; assert(Success && "Couldn't select *anything*?")((Success && "Couldn't select *anything*?") ? static_cast <void> (0) : __assert_fail ("Success && \"Couldn't select *anything*?\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 4174, __PRETTY_FUNCTION__)); | |||
4175 | ||||
4176 | // The match was to check the profitability, the changes made are not | |||
4177 | // part of the original matcher. Therefore, they should be dropped | |||
4178 | // otherwise the original matcher will not present the right state. | |||
4179 | TPT.rollback(LastKnownGood); | |||
4180 | ||||
4181 | // If the match didn't cover I, then it won't be shared by it. | |||
4182 | if (!is_contained(MatchedAddrModeInsts, I)) | |||
4183 | return false; | |||
4184 | ||||
4185 | MatchedAddrModeInsts.clear(); | |||
4186 | } | |||
4187 | ||||
4188 | return true; | |||
4189 | } | |||
4190 | ||||
4191 | } // end anonymous namespace | |||
4192 | ||||
4193 | /// Return true if the specified values are defined in a | |||
4194 | /// different basic block than BB. | |||
4195 | static bool IsNonLocalValue(Value *V, BasicBlock *BB) { | |||
4196 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
4197 | return I->getParent() != BB; | |||
4198 | return false; | |||
4199 | } | |||
4200 | ||||
4201 | /// Sink addressing mode computation immediate before MemoryInst if doing so | |||
4202 | /// can be done without increasing register pressure. The need for the | |||
4203 | /// register pressure constraint means this can end up being an all or nothing | |||
4204 | /// decision for all uses of the same addressing computation. | |||
4205 | /// | |||
4206 | /// Load and Store Instructions often have addressing modes that can do | |||
4207 | /// significant amounts of computation. As such, instruction selection will try | |||
4208 | /// to get the load or store to do as much computation as possible for the | |||
4209 | /// program. The problem is that isel can only see within a single block. As | |||
4210 | /// such, we sink as much legal addressing mode work into the block as possible. | |||
4211 | /// | |||
4212 | /// This method is used to optimize both load/store and inline asms with memory | |||
4213 | /// operands. It's also used to sink addressing computations feeding into cold | |||
4214 | /// call sites into their (cold) basic block. | |||
4215 | /// | |||
4216 | /// The motivation for handling sinking into cold blocks is that doing so can | |||
4217 | /// both enable other address mode sinking (by satisfying the register pressure | |||
4218 | /// constraint above), and reduce register pressure globally (by removing the | |||
4219 | /// addressing mode computation from the fast path entirely.). | |||
4220 | bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr, | |||
4221 | Type *AccessTy, unsigned AddrSpace) { | |||
4222 | Value *Repl = Addr; | |||
4223 | ||||
4224 | // Try to collapse single-value PHI nodes. This is necessary to undo | |||
4225 | // unprofitable PRE transformations. | |||
4226 | SmallVector<Value*, 8> worklist; | |||
4227 | SmallPtrSet<Value*, 16> Visited; | |||
4228 | worklist.push_back(Addr); | |||
4229 | ||||
4230 | // Use a worklist to iteratively look through PHI nodes, and ensure that | |||
4231 | // the addressing mode obtained from the non-PHI roots of the graph | |||
4232 | // are equivalent. | |||
4233 | Value *Consensus = nullptr; | |||
4234 | unsigned NumUsesConsensus = 0; | |||
4235 | bool IsNumUsesConsensusValid = false; | |||
4236 | SmallVector<Instruction*, 16> AddrModeInsts; | |||
4237 | ExtAddrMode AddrMode; | |||
4238 | TypePromotionTransaction TPT(RemovedInsts); | |||
4239 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | |||
4240 | TPT.getRestorationPoint(); | |||
4241 | while (!worklist.empty()) { | |||
4242 | Value *V = worklist.back(); | |||
4243 | worklist.pop_back(); | |||
4244 | ||||
4245 | // Break use-def graph loops. | |||
4246 | if (!Visited.insert(V).second) { | |||
4247 | Consensus = nullptr; | |||
4248 | break; | |||
4249 | } | |||
4250 | ||||
4251 | // For a PHI node, push all of its incoming values. | |||
4252 | if (PHINode *P = dyn_cast<PHINode>(V)) { | |||
4253 | for (Value *IncValue : P->incoming_values()) | |||
4254 | worklist.push_back(IncValue); | |||
4255 | continue; | |||
4256 | } | |||
4257 | ||||
4258 | // For non-PHIs, determine the addressing mode being computed. Note that | |||
4259 | // the result may differ depending on what other uses our candidate | |||
4260 | // addressing instructions might have. | |||
4261 | SmallVector<Instruction*, 16> NewAddrModeInsts; | |||
4262 | ExtAddrMode NewAddrMode = AddressingModeMatcher::Match( | |||
4263 | V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TLI, *TRI, | |||
4264 | InsertedInsts, PromotedInsts, TPT); | |||
4265 | ||||
4266 | // This check is broken into two cases with very similar code to avoid using | |||
4267 | // getNumUses() as much as possible. Some values have a lot of uses, so | |||
4268 | // calling getNumUses() unconditionally caused a significant compile-time | |||
4269 | // regression. | |||
4270 | if (!Consensus) { | |||
4271 | Consensus = V; | |||
4272 | AddrMode = NewAddrMode; | |||
4273 | AddrModeInsts = NewAddrModeInsts; | |||
4274 | continue; | |||
4275 | } else if (NewAddrMode == AddrMode) { | |||
4276 | if (!IsNumUsesConsensusValid) { | |||
4277 | NumUsesConsensus = Consensus->getNumUses(); | |||
4278 | IsNumUsesConsensusValid = true; | |||
4279 | } | |||
4280 | ||||
4281 | // Ensure that the obtained addressing mode is equivalent to that obtained | |||
4282 | // for all other roots of the PHI traversal. Also, when choosing one | |||
4283 | // such root as representative, select the one with the most uses in order | |||
4284 | // to keep the cost modeling heuristics in AddressingModeMatcher | |||
4285 | // applicable. | |||
4286 | unsigned NumUses = V->getNumUses(); | |||
4287 | if (NumUses > NumUsesConsensus) { | |||
4288 | Consensus = V; | |||
4289 | NumUsesConsensus = NumUses; | |||
4290 | AddrModeInsts = NewAddrModeInsts; | |||
4291 | } | |||
4292 | continue; | |||
4293 | } | |||
4294 | ||||
4295 | Consensus = nullptr; | |||
4296 | break; | |||
4297 | } | |||
4298 | ||||
4299 | // If the addressing mode couldn't be determined, or if multiple different | |||
4300 | // ones were determined, bail out now. | |||
4301 | if (!Consensus) { | |||
4302 | TPT.rollback(LastKnownGood); | |||
4303 | return false; | |||
4304 | } | |||
4305 | TPT.commit(); | |||
4306 | ||||
4307 | // If all the instructions matched are already in this BB, don't do anything. | |||
4308 | if (none_of(AddrModeInsts, [&](Value *V) { | |||
4309 | return IsNonLocalValue(V, MemoryInst->getParent()); | |||
4310 | })) { | |||
4311 | DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n"; } } while (false); | |||
4312 | return false; | |||
4313 | } | |||
4314 | ||||
4315 | // Insert this computation right after this user. Since our caller is | |||
4316 | // scanning from the top of the BB to the bottom, reuse of the expr are | |||
4317 | // guaranteed to happen later. | |||
4318 | IRBuilder<> Builder(MemoryInst); | |||
4319 | ||||
4320 | // Now that we determined the addressing expression we want to use and know | |||
4321 | // that we have to sink it into this block. Check to see if we have already | |||
4322 | // done this for some other load/store instr in this block. If so, reuse the | |||
4323 | // computation. | |||
4324 | Value *&SunkAddr = SunkAddrs[Addr]; | |||
4325 | if (SunkAddr) { | |||
4326 | DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for " << *MemoryInst << "\n"; } } while (false) | |||
4327 | << *MemoryInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for " << *MemoryInst << "\n"; } } while (false); | |||
4328 | if (SunkAddr->getType() != Addr->getType()) | |||
4329 | SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType()); | |||
4330 | } else if (AddrSinkUsingGEPs || | |||
4331 | (!AddrSinkUsingGEPs.getNumOccurrences() && TM && | |||
4332 | SubtargetInfo->useAA())) { | |||
4333 | // By default, we use the GEP-based method when AA is used later. This | |||
4334 | // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities. | |||
4335 | DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " << *MemoryInst << "\n"; } } while (false) | |||
4336 | << *MemoryInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " << *MemoryInst << "\n"; } } while (false); | |||
4337 | Type *IntPtrTy = DL->getIntPtrType(Addr->getType()); | |||
4338 | Value *ResultPtr = nullptr, *ResultIndex = nullptr; | |||
4339 | ||||
4340 | // First, find the pointer. | |||
4341 | if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) { | |||
4342 | ResultPtr = AddrMode.BaseReg; | |||
4343 | AddrMode.BaseReg = nullptr; | |||
4344 | } | |||
4345 | ||||
4346 | if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) { | |||
4347 | // We can't add more than one pointer together, nor can we scale a | |||
4348 | // pointer (both of which seem meaningless). | |||
4349 | if (ResultPtr || AddrMode.Scale != 1) | |||
4350 | return false; | |||
4351 | ||||
4352 | ResultPtr = AddrMode.ScaledReg; | |||
4353 | AddrMode.Scale = 0; | |||
4354 | } | |||
4355 | ||||
4356 | if (AddrMode.BaseGV) { | |||
4357 | if (ResultPtr) | |||
4358 | return false; | |||
4359 | ||||
4360 | ResultPtr = AddrMode.BaseGV; | |||
4361 | } | |||
4362 | ||||
4363 | // If the real base value actually came from an inttoptr, then the matcher | |||
4364 | // will look through it and provide only the integer value. In that case, | |||
4365 | // use it here. | |||
4366 | if (!ResultPtr && AddrMode.BaseReg) { | |||
4367 | ResultPtr = | |||
4368 | Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr"); | |||
4369 | AddrMode.BaseReg = nullptr; | |||
4370 | } else if (!ResultPtr && AddrMode.Scale == 1) { | |||
4371 | ResultPtr = | |||
4372 | Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr"); | |||
4373 | AddrMode.Scale = 0; | |||
4374 | } | |||
4375 | ||||
4376 | if (!ResultPtr && | |||
4377 | !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) { | |||
4378 | SunkAddr = Constant::getNullValue(Addr->getType()); | |||
4379 | } else if (!ResultPtr) { | |||
4380 | return false; | |||
4381 | } else { | |||
4382 | Type *I8PtrTy = | |||
4383 | Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace()); | |||
4384 | Type *I8Ty = Builder.getInt8Ty(); | |||
4385 | ||||
4386 | // Start with the base register. Do this first so that subsequent address | |||
4387 | // matching finds it last, which will prevent it from trying to match it | |||
4388 | // as the scaled value in case it happens to be a mul. That would be | |||
4389 | // problematic if we've sunk a different mul for the scale, because then | |||
4390 | // we'd end up sinking both muls. | |||
4391 | if (AddrMode.BaseReg) { | |||
4392 | Value *V = AddrMode.BaseReg; | |||
4393 | if (V->getType() != IntPtrTy) | |||
4394 | V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr"); | |||
4395 | ||||
4396 | ResultIndex = V; | |||
4397 | } | |||
4398 | ||||
4399 | // Add the scale value. | |||
4400 | if (AddrMode.Scale) { | |||
4401 | Value *V = AddrMode.ScaledReg; | |||
4402 | if (V->getType() == IntPtrTy) { | |||
4403 | // done. | |||
4404 | } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < | |||
4405 | cast<IntegerType>(V->getType())->getBitWidth()) { | |||
4406 | V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr"); | |||
4407 | } else { | |||
4408 | // It is only safe to sign extend the BaseReg if we know that the math | |||
4409 | // required to create it did not overflow before we extend it. Since | |||
4410 | // the original IR value was tossed in favor of a constant back when | |||
4411 | // the AddrMode was created we need to bail out gracefully if widths | |||
4412 | // do not match instead of extending it. | |||
4413 | Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex); | |||
4414 | if (I && (ResultIndex != AddrMode.BaseReg)) | |||
4415 | I->eraseFromParent(); | |||
4416 | return false; | |||
4417 | } | |||
4418 | ||||
4419 | if (AddrMode.Scale != 1) | |||
4420 | V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale), | |||
4421 | "sunkaddr"); | |||
4422 | if (ResultIndex) | |||
4423 | ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr"); | |||
4424 | else | |||
4425 | ResultIndex = V; | |||
4426 | } | |||
4427 | ||||
4428 | // Add in the Base Offset if present. | |||
4429 | if (AddrMode.BaseOffs) { | |||
4430 | Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); | |||
4431 | if (ResultIndex) { | |||
4432 | // We need to add this separately from the scale above to help with | |||
4433 | // SDAG consecutive load/store merging. | |||
4434 | if (ResultPtr->getType() != I8PtrTy) | |||
4435 | ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy); | |||
4436 | ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr"); | |||
4437 | } | |||
4438 | ||||
4439 | ResultIndex = V; | |||
4440 | } | |||
4441 | ||||
4442 | if (!ResultIndex) { | |||
4443 | SunkAddr = ResultPtr; | |||
4444 | } else { | |||
4445 | if (ResultPtr->getType() != I8PtrTy) | |||
4446 | ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy); | |||
4447 | SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr"); | |||
4448 | } | |||
4449 | ||||
4450 | if (SunkAddr->getType() != Addr->getType()) | |||
4451 | SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType()); | |||
4452 | } | |||
4453 | } else { | |||
4454 | DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " << *MemoryInst << "\n"; } } while (false) | |||
4455 | << *MemoryInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " << *MemoryInst << "\n"; } } while (false); | |||
4456 | Type *IntPtrTy = DL->getIntPtrType(Addr->getType()); | |||
4457 | Value *Result = nullptr; | |||
4458 | ||||
4459 | // Start with the base register. Do this first so that subsequent address | |||
4460 | // matching finds it last, which will prevent it from trying to match it | |||
4461 | // as the scaled value in case it happens to be a mul. That would be | |||
4462 | // problematic if we've sunk a different mul for the scale, because then | |||
4463 | // we'd end up sinking both muls. | |||
4464 | if (AddrMode.BaseReg) { | |||
4465 | Value *V = AddrMode.BaseReg; | |||
4466 | if (V->getType()->isPointerTy()) | |||
4467 | V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr"); | |||
4468 | if (V->getType() != IntPtrTy) | |||
4469 | V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr"); | |||
4470 | Result = V; | |||
4471 | } | |||
4472 | ||||
4473 | // Add the scale value. | |||
4474 | if (AddrMode.Scale) { | |||
4475 | Value *V = AddrMode.ScaledReg; | |||
4476 | if (V->getType() == IntPtrTy) { | |||
4477 | // done. | |||
4478 | } else if (V->getType()->isPointerTy()) { | |||
4479 | V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr"); | |||
4480 | } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < | |||
4481 | cast<IntegerType>(V->getType())->getBitWidth()) { | |||
4482 | V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr"); | |||
4483 | } else { | |||
4484 | // It is only safe to sign extend the BaseReg if we know that the math | |||
4485 | // required to create it did not overflow before we extend it. Since | |||
4486 | // the original IR value was tossed in favor of a constant back when | |||
4487 | // the AddrMode was created we need to bail out gracefully if widths | |||
4488 | // do not match instead of extending it. | |||
4489 | Instruction *I = dyn_cast_or_null<Instruction>(Result); | |||
4490 | if (I && (Result != AddrMode.BaseReg)) | |||
4491 | I->eraseFromParent(); | |||
4492 | return false; | |||
4493 | } | |||
4494 | if (AddrMode.Scale != 1) | |||
4495 | V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale), | |||
4496 | "sunkaddr"); | |||
4497 | if (Result) | |||
4498 | Result = Builder.CreateAdd(Result, V, "sunkaddr"); | |||
4499 | else | |||
4500 | Result = V; | |||
4501 | } | |||
4502 | ||||
4503 | // Add in the BaseGV if present. | |||
4504 | if (AddrMode.BaseGV) { | |||
4505 | Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr"); | |||
4506 | if (Result) | |||
4507 | Result = Builder.CreateAdd(Result, V, "sunkaddr"); | |||
4508 | else | |||
4509 | Result = V; | |||
4510 | } | |||
4511 | ||||
4512 | // Add in the Base Offset if present. | |||
4513 | if (AddrMode.BaseOffs) { | |||
4514 | Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); | |||
4515 | if (Result) | |||
4516 | Result = Builder.CreateAdd(Result, V, "sunkaddr"); | |||
4517 | else | |||
4518 | Result = V; | |||
4519 | } | |||
4520 | ||||
4521 | if (!Result) | |||
4522 | SunkAddr = Constant::getNullValue(Addr->getType()); | |||
4523 | else | |||
4524 | SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr"); | |||
4525 | } | |||
4526 | ||||
4527 | MemoryInst->replaceUsesOfWith(Repl, SunkAddr); | |||
4528 | ||||
4529 | // If we have no uses, recursively delete the value and all dead instructions | |||
4530 | // using it. | |||
4531 | if (Repl->use_empty()) { | |||
4532 | // This can cause recursive deletion, which can invalidate our iterator. | |||
4533 | // Use a WeakTrackingVH to hold onto it in case this happens. | |||
4534 | Value *CurValue = &*CurInstIterator; | |||
4535 | WeakTrackingVH IterHandle(CurValue); | |||
4536 | BasicBlock *BB = CurInstIterator->getParent(); | |||
4537 | ||||
4538 | RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo); | |||
4539 | ||||
4540 | if (IterHandle != CurValue) { | |||
4541 | // If the iterator instruction was recursively deleted, start over at the | |||
4542 | // start of the block. | |||
4543 | CurInstIterator = BB->begin(); | |||
4544 | SunkAddrs.clear(); | |||
4545 | } | |||
4546 | } | |||
4547 | ++NumMemoryInsts; | |||
4548 | return true; | |||
4549 | } | |||
4550 | ||||
4551 | /// If there are any memory operands, use OptimizeMemoryInst to sink their | |||
4552 | /// address computing into the block when possible / profitable. | |||
4553 | bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) { | |||
4554 | bool MadeChange = false; | |||
4555 | ||||
4556 | const TargetRegisterInfo *TRI = | |||
4557 | TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo(); | |||
4558 | TargetLowering::AsmOperandInfoVector TargetConstraints = | |||
4559 | TLI->ParseConstraints(*DL, TRI, CS); | |||
4560 | unsigned ArgNo = 0; | |||
4561 | for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) { | |||
4562 | TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i]; | |||
4563 | ||||
4564 | // Compute the constraint code and ConstraintType to use. | |||
4565 | TLI->ComputeConstraintToUse(OpInfo, SDValue()); | |||
4566 | ||||
4567 | if (OpInfo.ConstraintType == TargetLowering::C_Memory && | |||
4568 | OpInfo.isIndirect) { | |||
4569 | Value *OpVal = CS->getArgOperand(ArgNo++); | |||
4570 | MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u); | |||
4571 | } else if (OpInfo.Type == InlineAsm::isInput) | |||
4572 | ArgNo++; | |||
4573 | } | |||
4574 | ||||
4575 | return MadeChange; | |||
4576 | } | |||
4577 | ||||
4578 | /// \brief Check if all the uses of \p Val are equivalent (or free) zero or | |||
4579 | /// sign extensions. | |||
4580 | static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) { | |||
4581 | assert(!Val->use_empty() && "Input must have at least one use")((!Val->use_empty() && "Input must have at least one use" ) ? static_cast<void> (0) : __assert_fail ("!Val->use_empty() && \"Input must have at least one use\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 4581, __PRETTY_FUNCTION__)); | |||
4582 | const Instruction *FirstUser = cast<Instruction>(*Val->user_begin()); | |||
4583 | bool IsSExt = isa<SExtInst>(FirstUser); | |||
4584 | Type *ExtTy = FirstUser->getType(); | |||
4585 | for (const User *U : Val->users()) { | |||
4586 | const Instruction *UI = cast<Instruction>(U); | |||
4587 | if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI))) | |||
4588 | return false; | |||
4589 | Type *CurTy = UI->getType(); | |||
4590 | // Same input and output types: Same instruction after CSE. | |||
4591 | if (CurTy == ExtTy) | |||
4592 | continue; | |||
4593 | ||||
4594 | // If IsSExt is true, we are in this situation: | |||
4595 | // a = Val | |||
4596 | // b = sext ty1 a to ty2 | |||
4597 | // c = sext ty1 a to ty3 | |||
4598 | // Assuming ty2 is shorter than ty3, this could be turned into: | |||
4599 | // a = Val | |||
4600 | // b = sext ty1 a to ty2 | |||
4601 | // c = sext ty2 b to ty3 | |||
4602 | // However, the last sext is not free. | |||
4603 | if (IsSExt) | |||
4604 | return false; | |||
4605 | ||||
4606 | // This is a ZExt, maybe this is free to extend from one type to another. | |||
4607 | // In that case, we would not account for a different use. | |||
4608 | Type *NarrowTy; | |||
4609 | Type *LargeTy; | |||
4610 | if (ExtTy->getScalarType()->getIntegerBitWidth() > | |||
4611 | CurTy->getScalarType()->getIntegerBitWidth()) { | |||
4612 | NarrowTy = CurTy; | |||
4613 | LargeTy = ExtTy; | |||
4614 | } else { | |||
4615 | NarrowTy = ExtTy; | |||
4616 | LargeTy = CurTy; | |||
4617 | } | |||
4618 | ||||
4619 | if (!TLI.isZExtFree(NarrowTy, LargeTy)) | |||
4620 | return false; | |||
4621 | } | |||
4622 | // All uses are the same or can be derived from one another for free. | |||
4623 | return true; | |||
4624 | } | |||
4625 | ||||
4626 | /// \brief Try to speculatively promote extensions in \p Exts and continue | |||
4627 | /// promoting through newly promoted operands recursively as far as doing so is | |||
4628 | /// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts. | |||
4629 | /// When some promotion happened, \p TPT contains the proper state to revert | |||
4630 | /// them. | |||
4631 | /// | |||
4632 | /// \return true if some promotion happened, false otherwise. | |||
4633 | bool CodeGenPrepare::tryToPromoteExts( | |||
4634 | TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts, | |||
4635 | SmallVectorImpl<Instruction *> &ProfitablyMovedExts, | |||
4636 | unsigned CreatedInstsCost) { | |||
4637 | bool Promoted = false; | |||
4638 | ||||
4639 | // Iterate over all the extensions to try to promote them. | |||
4640 | for (auto I : Exts) { | |||
4641 | // Early check if we directly have ext(load). | |||
4642 | if (isa<LoadInst>(I->getOperand(0))) { | |||
4643 | ProfitablyMovedExts.push_back(I); | |||
4644 | continue; | |||
4645 | } | |||
4646 | ||||
4647 | // Check whether or not we want to do any promotion. The reason we have | |||
4648 | // this check inside the for loop is to catch the case where an extension | |||
4649 | // is directly fed by a load because in such case the extension can be moved | |||
4650 | // up without any promotion on its operands. | |||
4651 | if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion) | |||
4652 | return false; | |||
4653 | ||||
4654 | // Get the action to perform the promotion. | |||
4655 | TypePromotionHelper::Action TPH = | |||
4656 | TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts); | |||
4657 | // Check if we can promote. | |||
4658 | if (!TPH) { | |||
4659 | // Save the current extension as we cannot move up through its operand. | |||
4660 | ProfitablyMovedExts.push_back(I); | |||
4661 | continue; | |||
4662 | } | |||
4663 | ||||
4664 | // Save the current state. | |||
4665 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | |||
4666 | TPT.getRestorationPoint(); | |||
4667 | SmallVector<Instruction *, 4> NewExts; | |||
4668 | unsigned NewCreatedInstsCost = 0; | |||
4669 | unsigned ExtCost = !TLI->isExtFree(I); | |||
4670 | // Promote. | |||
4671 | Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost, | |||
4672 | &NewExts, nullptr, *TLI); | |||
4673 | assert(PromotedVal &&((PromotedVal && "TypePromotionHelper should have filtered out those cases" ) ? static_cast<void> (0) : __assert_fail ("PromotedVal && \"TypePromotionHelper should have filtered out those cases\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 4674, __PRETTY_FUNCTION__)) | |||
4674 | "TypePromotionHelper should have filtered out those cases")((PromotedVal && "TypePromotionHelper should have filtered out those cases" ) ? static_cast<void> (0) : __assert_fail ("PromotedVal && \"TypePromotionHelper should have filtered out those cases\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 4674, __PRETTY_FUNCTION__)); | |||
4675 | ||||
4676 | // We would be able to merge only one extension in a load. | |||
4677 | // Therefore, if we have more than 1 new extension we heuristically | |||
4678 | // cut this search path, because it means we degrade the code quality. | |||
4679 | // With exactly 2, the transformation is neutral, because we will merge | |||
4680 | // one extension but leave one. However, we optimistically keep going, | |||
4681 | // because the new extension may be removed too. | |||
4682 | long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost; | |||
4683 | // FIXME: It would be possible to propagate a negative value instead of | |||
4684 | // conservatively ceiling it to 0. | |||
4685 | TotalCreatedInstsCost = | |||
4686 | std::max((long long)0, (TotalCreatedInstsCost - ExtCost)); | |||
4687 | if (!StressExtLdPromotion && | |||
4688 | (TotalCreatedInstsCost > 1 || | |||
4689 | !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) { | |||
4690 | // This promotion is not profitable, rollback to the previous state, and | |||
4691 | // save the current extension in ProfitablyMovedExts as the latest | |||
4692 | // speculative promotion turned out to be unprofitable. | |||
4693 | TPT.rollback(LastKnownGood); | |||
4694 | ProfitablyMovedExts.push_back(I); | |||
4695 | continue; | |||
4696 | } | |||
4697 | // Continue promoting NewExts as far as doing so is profitable. | |||
4698 | SmallVector<Instruction *, 2> NewlyMovedExts; | |||
4699 | (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost); | |||
4700 | bool NewPromoted = false; | |||
4701 | for (auto ExtInst : NewlyMovedExts) { | |||
4702 | Instruction *MovedExt = cast<Instruction>(ExtInst); | |||
4703 | Value *ExtOperand = MovedExt->getOperand(0); | |||
4704 | // If we have reached to a load, we need this extra profitability check | |||
4705 | // as it could potentially be merged into an ext(load). | |||
4706 | if (isa<LoadInst>(ExtOperand) && | |||
4707 | !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost || | |||
4708 | (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI)))) | |||
4709 | continue; | |||
4710 | ||||
4711 | ProfitablyMovedExts.push_back(MovedExt); | |||
4712 | NewPromoted = true; | |||
4713 | } | |||
4714 | ||||
4715 | // If none of speculative promotions for NewExts is profitable, rollback | |||
4716 | // and save the current extension (I) as the last profitable extension. | |||
4717 | if (!NewPromoted) { | |||
4718 | TPT.rollback(LastKnownGood); | |||
4719 | ProfitablyMovedExts.push_back(I); | |||
4720 | continue; | |||
4721 | } | |||
4722 | // The promotion is profitable. | |||
4723 | Promoted = true; | |||
4724 | } | |||
4725 | return Promoted; | |||
4726 | } | |||
4727 | ||||
4728 | /// Merging redundant sexts when one is dominating the other. | |||
4729 | bool CodeGenPrepare::mergeSExts(Function &F) { | |||
4730 | DominatorTree DT(F); | |||
4731 | bool Changed = false; | |||
4732 | for (auto &Entry : ValToSExtendedUses) { | |||
4733 | SExts &Insts = Entry.second; | |||
4734 | SExts CurPts; | |||
4735 | for (Instruction *Inst : Insts) { | |||
4736 | if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) || | |||
4737 | Inst->getOperand(0) != Entry.first) | |||
4738 | continue; | |||
4739 | bool inserted = false; | |||
4740 | for (auto &Pt : CurPts) { | |||
4741 | if (DT.dominates(Inst, Pt)) { | |||
4742 | Pt->replaceAllUsesWith(Inst); | |||
4743 | RemovedInsts.insert(Pt); | |||
4744 | Pt->removeFromParent(); | |||
4745 | Pt = Inst; | |||
4746 | inserted = true; | |||
4747 | Changed = true; | |||
4748 | break; | |||
4749 | } | |||
4750 | if (!DT.dominates(Pt, Inst)) | |||
4751 | // Give up if we need to merge in a common dominator as the | |||
4752 | // expermients show it is not profitable. | |||
4753 | continue; | |||
4754 | Inst->replaceAllUsesWith(Pt); | |||
4755 | RemovedInsts.insert(Inst); | |||
4756 | Inst->removeFromParent(); | |||
4757 | inserted = true; | |||
4758 | Changed = true; | |||
4759 | break; | |||
4760 | } | |||
4761 | if (!inserted) | |||
4762 | CurPts.push_back(Inst); | |||
4763 | } | |||
4764 | } | |||
4765 | return Changed; | |||
4766 | } | |||
4767 | ||||
4768 | /// Return true, if an ext(load) can be formed from an extension in | |||
4769 | /// \p MovedExts. | |||
4770 | bool CodeGenPrepare::canFormExtLd( | |||
4771 | const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI, | |||
4772 | Instruction *&Inst, bool HasPromoted) { | |||
4773 | for (auto *MovedExtInst : MovedExts) { | |||
4774 | if (isa<LoadInst>(MovedExtInst->getOperand(0))) { | |||
4775 | LI = cast<LoadInst>(MovedExtInst->getOperand(0)); | |||
4776 | Inst = MovedExtInst; | |||
4777 | break; | |||
4778 | } | |||
4779 | } | |||
4780 | if (!LI) | |||
4781 | return false; | |||
4782 | ||||
4783 | // If they're already in the same block, there's nothing to do. | |||
4784 | // Make the cheap checks first if we did not promote. | |||
4785 | // If we promoted, we need to check if it is indeed profitable. | |||
4786 | if (!HasPromoted && LI->getParent() == Inst->getParent()) | |||
4787 | return false; | |||
4788 | ||||
4789 | EVT VT = TLI->getValueType(*DL, Inst->getType()); | |||
4790 | EVT LoadVT = TLI->getValueType(*DL, LI->getType()); | |||
4791 | ||||
4792 | // If the load has other users and the truncate is not free, this probably | |||
4793 | // isn't worthwhile. | |||
4794 | if (!LI->hasOneUse() && (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) && | |||
4795 | !TLI->isTruncateFree(Inst->getType(), LI->getType())) | |||
4796 | return false; | |||
4797 | ||||
4798 | // Check whether the target supports casts folded into loads. | |||
4799 | unsigned LType; | |||
4800 | if (isa<ZExtInst>(Inst)) | |||
4801 | LType = ISD::ZEXTLOAD; | |||
4802 | else { | |||
4803 | assert(isa<SExtInst>(Inst) && "Unexpected ext type!")((isa<SExtInst>(Inst) && "Unexpected ext type!" ) ? static_cast<void> (0) : __assert_fail ("isa<SExtInst>(Inst) && \"Unexpected ext type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 4803, __PRETTY_FUNCTION__)); | |||
4804 | LType = ISD::SEXTLOAD; | |||
4805 | } | |||
4806 | ||||
4807 | return TLI->isLoadExtLegal(LType, VT, LoadVT); | |||
4808 | } | |||
4809 | ||||
4810 | /// Move a zext or sext fed by a load into the same basic block as the load, | |||
4811 | /// unless conditions are unfavorable. This allows SelectionDAG to fold the | |||
4812 | /// extend into the load. | |||
4813 | /// | |||
4814 | /// E.g., | |||
4815 | /// \code | |||
4816 | /// %ld = load i32* %addr | |||
4817 | /// %add = add nuw i32 %ld, 4 | |||
4818 | /// %zext = zext i32 %add to i64 | |||
4819 | // \endcode | |||
4820 | /// => | |||
4821 | /// \code | |||
4822 | /// %ld = load i32* %addr | |||
4823 | /// %zext = zext i32 %ld to i64 | |||
4824 | /// %add = add nuw i64 %zext, 4 | |||
4825 | /// \encode | |||
4826 | /// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which | |||
4827 | /// allow us to match zext(load i32*) to i64. | |||
4828 | /// | |||
4829 | /// Also, try to promote the computations used to obtain a sign extended | |||
4830 | /// value used into memory accesses. | |||
4831 | /// E.g., | |||
4832 | /// \code | |||
4833 | /// a = add nsw i32 b, 3 | |||
4834 | /// d = sext i32 a to i64 | |||
4835 | /// e = getelementptr ..., i64 d | |||
4836 | /// \endcode | |||
4837 | /// => | |||
4838 | /// \code | |||
4839 | /// f = sext i32 b to i64 | |||
4840 | /// a = add nsw i64 f, 3 | |||
4841 | /// e = getelementptr ..., i64 a | |||
4842 | /// \endcode | |||
4843 | /// | |||
4844 | /// \p Inst[in/out] the extension may be modified during the process if some | |||
4845 | /// promotions apply. | |||
4846 | bool CodeGenPrepare::optimizeExt(Instruction *&Inst) { | |||
4847 | // ExtLoad formation and address type promotion infrastructure requires TLI to | |||
4848 | // be effective. | |||
4849 | if (!TLI) | |||
4850 | return false; | |||
4851 | ||||
4852 | bool AllowPromotionWithoutCommonHeader = false; | |||
4853 | /// See if it is an interesting sext operations for the address type | |||
4854 | /// promotion before trying to promote it, e.g., the ones with the right | |||
4855 | /// type and used in memory accesses. | |||
4856 | bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion( | |||
4857 | *Inst, AllowPromotionWithoutCommonHeader); | |||
4858 | TypePromotionTransaction TPT(RemovedInsts); | |||
4859 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | |||
4860 | TPT.getRestorationPoint(); | |||
4861 | SmallVector<Instruction *, 1> Exts; | |||
4862 | SmallVector<Instruction *, 2> SpeculativelyMovedExts; | |||
4863 | Exts.push_back(Inst); | |||
4864 | ||||
4865 | bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts); | |||
4866 | ||||
4867 | // Look for a load being extended. | |||
4868 | LoadInst *LI = nullptr; | |||
4869 | Instruction *ExtFedByLoad; | |||
4870 | ||||
4871 | // Try to promote a chain of computation if it allows to form an extended | |||
4872 | // load. | |||
4873 | if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) { | |||
4874 | assert(LI && ExtFedByLoad && "Expect a valid load and extension")((LI && ExtFedByLoad && "Expect a valid load and extension" ) ? static_cast<void> (0) : __assert_fail ("LI && ExtFedByLoad && \"Expect a valid load and extension\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 4874, __PRETTY_FUNCTION__)); | |||
4875 | TPT.commit(); | |||
4876 | // Move the extend into the same block as the load | |||
4877 | ExtFedByLoad->removeFromParent(); | |||
4878 | ExtFedByLoad->insertAfter(LI); | |||
4879 | // CGP does not check if the zext would be speculatively executed when moved | |||
4880 | // to the same basic block as the load. Preserving its original location | |||
4881 | // would pessimize the debugging experience, as well as negatively impact | |||
4882 | // the quality of sample pgo. We don't want to use "line 0" as that has a | |||
4883 | // size cost in the line-table section and logically the zext can be seen as | |||
4884 | // part of the load. Therefore we conservatively reuse the same debug | |||
4885 | // location for the load and the zext. | |||
4886 | ExtFedByLoad->setDebugLoc(LI->getDebugLoc()); | |||
4887 | ++NumExtsMoved; | |||
4888 | Inst = ExtFedByLoad; | |||
4889 | return true; | |||
4890 | } | |||
4891 | ||||
4892 | // Continue promoting SExts if known as considerable depending on targets. | |||
4893 | if (ATPConsiderable && | |||
4894 | performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader, | |||
4895 | HasPromoted, TPT, SpeculativelyMovedExts)) | |||
4896 | return true; | |||
4897 | ||||
4898 | TPT.rollback(LastKnownGood); | |||
4899 | return false; | |||
4900 | } | |||
4901 | ||||
4902 | // Perform address type promotion if doing so is profitable. | |||
4903 | // If AllowPromotionWithoutCommonHeader == false, we should find other sext | |||
4904 | // instructions that sign extended the same initial value. However, if | |||
4905 | // AllowPromotionWithoutCommonHeader == true, we expect promoting the | |||
4906 | // extension is just profitable. | |||
4907 | bool CodeGenPrepare::performAddressTypePromotion( | |||
4908 | Instruction *&Inst, bool AllowPromotionWithoutCommonHeader, | |||
4909 | bool HasPromoted, TypePromotionTransaction &TPT, | |||
4910 | SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) { | |||
4911 | bool Promoted = false; | |||
4912 | SmallPtrSet<Instruction *, 1> UnhandledExts; | |||
4913 | bool AllSeenFirst = true; | |||
4914 | for (auto I : SpeculativelyMovedExts) { | |||
4915 | Value *HeadOfChain = I->getOperand(0); | |||
4916 | DenseMap<Value *, Instruction *>::iterator AlreadySeen = | |||
4917 | SeenChainsForSExt.find(HeadOfChain); | |||
4918 | // If there is an unhandled SExt which has the same header, try to promote | |||
4919 | // it as well. | |||
4920 | if (AlreadySeen != SeenChainsForSExt.end()) { | |||
4921 | if (AlreadySeen->second != nullptr) | |||
4922 | UnhandledExts.insert(AlreadySeen->second); | |||
4923 | AllSeenFirst = false; | |||
4924 | } | |||
4925 | } | |||
4926 | ||||
4927 | if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader && | |||
4928 | SpeculativelyMovedExts.size() == 1)) { | |||
4929 | TPT.commit(); | |||
4930 | if (HasPromoted) | |||
4931 | Promoted = true; | |||
4932 | for (auto I : SpeculativelyMovedExts) { | |||
4933 | Value *HeadOfChain = I->getOperand(0); | |||
4934 | SeenChainsForSExt[HeadOfChain] = nullptr; | |||
4935 | ValToSExtendedUses[HeadOfChain].push_back(I); | |||
4936 | } | |||
4937 | // Update Inst as promotion happen. | |||
4938 | Inst = SpeculativelyMovedExts.pop_back_val(); | |||
4939 | } else { | |||
4940 | // This is the first chain visited from the header, keep the current chain | |||
4941 | // as unhandled. Defer to promote this until we encounter another SExt | |||
4942 | // chain derived from the same header. | |||
4943 | for (auto I : SpeculativelyMovedExts) { | |||
4944 | Value *HeadOfChain = I->getOperand(0); | |||
4945 | SeenChainsForSExt[HeadOfChain] = Inst; | |||
4946 | } | |||
4947 | return false; | |||
4948 | } | |||
4949 | ||||
4950 | if (!AllSeenFirst && !UnhandledExts.empty()) | |||
4951 | for (auto VisitedSExt : UnhandledExts) { | |||
4952 | if (RemovedInsts.count(VisitedSExt)) | |||
4953 | continue; | |||
4954 | TypePromotionTransaction TPT(RemovedInsts); | |||
4955 | SmallVector<Instruction *, 1> Exts; | |||
4956 | SmallVector<Instruction *, 2> Chains; | |||
4957 | Exts.push_back(VisitedSExt); | |||
4958 | bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains); | |||
4959 | TPT.commit(); | |||
4960 | if (HasPromoted) | |||
4961 | Promoted = true; | |||
4962 | for (auto I : Chains) { | |||
4963 | Value *HeadOfChain = I->getOperand(0); | |||
4964 | // Mark this as handled. | |||
4965 | SeenChainsForSExt[HeadOfChain] = nullptr; | |||
4966 | ValToSExtendedUses[HeadOfChain].push_back(I); | |||
4967 | } | |||
4968 | } | |||
4969 | return Promoted; | |||
4970 | } | |||
4971 | ||||
4972 | bool CodeGenPrepare::optimizeExtUses(Instruction *I) { | |||
4973 | BasicBlock *DefBB = I->getParent(); | |||
4974 | ||||
4975 | // If the result of a {s|z}ext and its source are both live out, rewrite all | |||
4976 | // other uses of the source with result of extension. | |||
4977 | Value *Src = I->getOperand(0); | |||
4978 | if (Src->hasOneUse()) | |||
4979 | return false; | |||
4980 | ||||
4981 | // Only do this xform if truncating is free. | |||
4982 | if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType())) | |||
4983 | return false; | |||
4984 | ||||
4985 | // Only safe to perform the optimization if the source is also defined in | |||
4986 | // this block. | |||
4987 | if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent()) | |||
4988 | return false; | |||
4989 | ||||
4990 | bool DefIsLiveOut = false; | |||
4991 | for (User *U : I->users()) { | |||
4992 | Instruction *UI = cast<Instruction>(U); | |||
4993 | ||||
4994 | // Figure out which BB this ext is used in. | |||
4995 | BasicBlock *UserBB = UI->getParent(); | |||
4996 | if (UserBB == DefBB) continue; | |||
4997 | DefIsLiveOut = true; | |||
4998 | break; | |||
4999 | } | |||
5000 | if (!DefIsLiveOut) | |||
5001 | return false; | |||
5002 | ||||
5003 | // Make sure none of the uses are PHI nodes. | |||
5004 | for (User *U : Src->users()) { | |||
5005 | Instruction *UI = cast<Instruction>(U); | |||
5006 | BasicBlock *UserBB = UI->getParent(); | |||
5007 | if (UserBB == DefBB) continue; | |||
5008 | // Be conservative. We don't want this xform to end up introducing | |||
5009 | // reloads just before load / store instructions. | |||
5010 | if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI)) | |||
5011 | return false; | |||
5012 | } | |||
5013 | ||||
5014 | // InsertedTruncs - Only insert one trunc in each block once. | |||
5015 | DenseMap<BasicBlock*, Instruction*> InsertedTruncs; | |||
5016 | ||||
5017 | bool MadeChange = false; | |||
5018 | for (Use &U : Src->uses()) { | |||
5019 | Instruction *User = cast<Instruction>(U.getUser()); | |||
5020 | ||||
5021 | // Figure out which BB this ext is used in. | |||
5022 | BasicBlock *UserBB = User->getParent(); | |||
5023 | if (UserBB == DefBB) continue; | |||
5024 | ||||
5025 | // Both src and def are live in this block. Rewrite the use. | |||
5026 | Instruction *&InsertedTrunc = InsertedTruncs[UserBB]; | |||
5027 | ||||
5028 | if (!InsertedTrunc) { | |||
5029 | BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); | |||
5030 | assert(InsertPt != UserBB->end())((InsertPt != UserBB->end()) ? static_cast<void> (0) : __assert_fail ("InsertPt != UserBB->end()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5030, __PRETTY_FUNCTION__)); | |||
5031 | InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt); | |||
5032 | InsertedInsts.insert(InsertedTrunc); | |||
5033 | } | |||
5034 | ||||
5035 | // Replace a use of the {s|z}ext source with a use of the result. | |||
5036 | U = InsertedTrunc; | |||
5037 | ++NumExtUses; | |||
5038 | MadeChange = true; | |||
5039 | } | |||
5040 | ||||
5041 | return MadeChange; | |||
5042 | } | |||
5043 | ||||
5044 | // Find loads whose uses only use some of the loaded value's bits. Add an "and" | |||
5045 | // just after the load if the target can fold this into one extload instruction, | |||
5046 | // with the hope of eliminating some of the other later "and" instructions using | |||
5047 | // the loaded value. "and"s that are made trivially redundant by the insertion | |||
5048 | // of the new "and" are removed by this function, while others (e.g. those whose | |||
5049 | // path from the load goes through a phi) are left for isel to potentially | |||
5050 | // remove. | |||
5051 | // | |||
5052 | // For example: | |||
5053 | // | |||
5054 | // b0: | |||
5055 | // x = load i32 | |||
5056 | // ... | |||
5057 | // b1: | |||
5058 | // y = and x, 0xff | |||
5059 | // z = use y | |||
5060 | // | |||
5061 | // becomes: | |||
5062 | // | |||
5063 | // b0: | |||
5064 | // x = load i32 | |||
5065 | // x' = and x, 0xff | |||
5066 | // ... | |||
5067 | // b1: | |||
5068 | // z = use x' | |||
5069 | // | |||
5070 | // whereas: | |||
5071 | // | |||
5072 | // b0: | |||
5073 | // x1 = load i32 | |||
5074 | // ... | |||
5075 | // b1: | |||
5076 | // x2 = load i32 | |||
5077 | // ... | |||
5078 | // b2: | |||
5079 | // x = phi x1, x2 | |||
5080 | // y = and x, 0xff | |||
5081 | // | |||
5082 | // becomes (after a call to optimizeLoadExt for each load): | |||
5083 | // | |||
5084 | // b0: | |||
5085 | // x1 = load i32 | |||
5086 | // x1' = and x1, 0xff | |||
5087 | // ... | |||
5088 | // b1: | |||
5089 | // x2 = load i32 | |||
5090 | // x2' = and x2, 0xff | |||
5091 | // ... | |||
5092 | // b2: | |||
5093 | // x = phi x1', x2' | |||
5094 | // y = and x, 0xff | |||
5095 | // | |||
5096 | ||||
5097 | bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) { | |||
5098 | ||||
5099 | if (!Load->isSimple() || | |||
5100 | !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy())) | |||
5101 | return false; | |||
5102 | ||||
5103 | // Skip loads we've already transformed. | |||
5104 | if (Load->hasOneUse() && | |||
5105 | InsertedInsts.count(cast<Instruction>(*Load->user_begin()))) | |||
5106 | return false; | |||
5107 | ||||
5108 | // Look at all uses of Load, looking through phis, to determine how many bits | |||
5109 | // of the loaded value are needed. | |||
5110 | SmallVector<Instruction *, 8> WorkList; | |||
5111 | SmallPtrSet<Instruction *, 16> Visited; | |||
5112 | SmallVector<Instruction *, 8> AndsToMaybeRemove; | |||
5113 | for (auto *U : Load->users()) | |||
5114 | WorkList.push_back(cast<Instruction>(U)); | |||
5115 | ||||
5116 | EVT LoadResultVT = TLI->getValueType(*DL, Load->getType()); | |||
5117 | unsigned BitWidth = LoadResultVT.getSizeInBits(); | |||
5118 | APInt DemandBits(BitWidth, 0); | |||
5119 | APInt WidestAndBits(BitWidth, 0); | |||
5120 | ||||
5121 | while (!WorkList.empty()) { | |||
5122 | Instruction *I = WorkList.back(); | |||
5123 | WorkList.pop_back(); | |||
5124 | ||||
5125 | // Break use-def graph loops. | |||
5126 | if (!Visited.insert(I).second) | |||
5127 | continue; | |||
5128 | ||||
5129 | // For a PHI node, push all of its users. | |||
5130 | if (auto *Phi = dyn_cast<PHINode>(I)) { | |||
5131 | for (auto *U : Phi->users()) | |||
5132 | WorkList.push_back(cast<Instruction>(U)); | |||
5133 | continue; | |||
5134 | } | |||
5135 | ||||
5136 | switch (I->getOpcode()) { | |||
5137 | case llvm::Instruction::And: { | |||
5138 | auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1)); | |||
5139 | if (!AndC) | |||
5140 | return false; | |||
5141 | APInt AndBits = AndC->getValue(); | |||
5142 | DemandBits |= AndBits; | |||
5143 | // Keep track of the widest and mask we see. | |||
5144 | if (AndBits.ugt(WidestAndBits)) | |||
5145 | WidestAndBits = AndBits; | |||
5146 | if (AndBits == WidestAndBits && I->getOperand(0) == Load) | |||
5147 | AndsToMaybeRemove.push_back(I); | |||
5148 | break; | |||
5149 | } | |||
5150 | ||||
5151 | case llvm::Instruction::Shl: { | |||
5152 | auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1)); | |||
5153 | if (!ShlC) | |||
5154 | return false; | |||
5155 | uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1); | |||
5156 | DemandBits.setLowBits(BitWidth - ShiftAmt); | |||
5157 | break; | |||
5158 | } | |||
5159 | ||||
5160 | case llvm::Instruction::Trunc: { | |||
5161 | EVT TruncVT = TLI->getValueType(*DL, I->getType()); | |||
5162 | unsigned TruncBitWidth = TruncVT.getSizeInBits(); | |||
5163 | DemandBits.setLowBits(TruncBitWidth); | |||
5164 | break; | |||
5165 | } | |||
5166 | ||||
5167 | default: | |||
5168 | return false; | |||
5169 | } | |||
5170 | } | |||
5171 | ||||
5172 | uint32_t ActiveBits = DemandBits.getActiveBits(); | |||
5173 | // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the | |||
5174 | // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example, | |||
5175 | // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but | |||
5176 | // (and (load x) 1) is not matched as a single instruction, rather as a LDR | |||
5177 | // followed by an AND. | |||
5178 | // TODO: Look into removing this restriction by fixing backends to either | |||
5179 | // return false for isLoadExtLegal for i1 or have them select this pattern to | |||
5180 | // a single instruction. | |||
5181 | // | |||
5182 | // Also avoid hoisting if we didn't see any ands with the exact DemandBits | |||
5183 | // mask, since these are the only ands that will be removed by isel. | |||
5184 | if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) || | |||
5185 | WidestAndBits != DemandBits) | |||
5186 | return false; | |||
5187 | ||||
5188 | LLVMContext &Ctx = Load->getType()->getContext(); | |||
5189 | Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits); | |||
5190 | EVT TruncVT = TLI->getValueType(*DL, TruncTy); | |||
5191 | ||||
5192 | // Reject cases that won't be matched as extloads. | |||
5193 | if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() || | |||
5194 | !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT)) | |||
5195 | return false; | |||
5196 | ||||
5197 | IRBuilder<> Builder(Load->getNextNode()); | |||
5198 | auto *NewAnd = dyn_cast<Instruction>( | |||
5199 | Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits))); | |||
5200 | // Mark this instruction as "inserted by CGP", so that other | |||
5201 | // optimizations don't touch it. | |||
5202 | InsertedInsts.insert(NewAnd); | |||
5203 | ||||
5204 | // Replace all uses of load with new and (except for the use of load in the | |||
5205 | // new and itself). | |||
5206 | Load->replaceAllUsesWith(NewAnd); | |||
5207 | NewAnd->setOperand(0, Load); | |||
5208 | ||||
5209 | // Remove any and instructions that are now redundant. | |||
5210 | for (auto *And : AndsToMaybeRemove) | |||
5211 | // Check that the and mask is the same as the one we decided to put on the | |||
5212 | // new and. | |||
5213 | if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) { | |||
5214 | And->replaceAllUsesWith(NewAnd); | |||
5215 | if (&*CurInstIterator == And) | |||
5216 | CurInstIterator = std::next(And->getIterator()); | |||
5217 | And->eraseFromParent(); | |||
5218 | ++NumAndUses; | |||
5219 | } | |||
5220 | ||||
5221 | ++NumAndsAdded; | |||
5222 | return true; | |||
5223 | } | |||
5224 | ||||
5225 | /// Check if V (an operand of a select instruction) is an expensive instruction | |||
5226 | /// that is only used once. | |||
5227 | static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) { | |||
5228 | auto *I = dyn_cast<Instruction>(V); | |||
5229 | // If it's safe to speculatively execute, then it should not have side | |||
5230 | // effects; therefore, it's safe to sink and possibly *not* execute. | |||
5231 | return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) && | |||
5232 | TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive; | |||
5233 | } | |||
5234 | ||||
5235 | /// Returns true if a SelectInst should be turned into an explicit branch. | |||
5236 | static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI, | |||
5237 | const TargetLowering *TLI, | |||
5238 | SelectInst *SI) { | |||
5239 | // If even a predictable select is cheap, then a branch can't be cheaper. | |||
5240 | if (!TLI->isPredictableSelectExpensive()) | |||
5241 | return false; | |||
5242 | ||||
5243 | // FIXME: This should use the same heuristics as IfConversion to determine | |||
5244 | // whether a select is better represented as a branch. | |||
5245 | ||||
5246 | // If metadata tells us that the select condition is obviously predictable, | |||
5247 | // then we want to replace the select with a branch. | |||
5248 | uint64_t TrueWeight, FalseWeight; | |||
5249 | if (SI->extractProfMetadata(TrueWeight, FalseWeight)) { | |||
5250 | uint64_t Max = std::max(TrueWeight, FalseWeight); | |||
5251 | uint64_t Sum = TrueWeight + FalseWeight; | |||
5252 | if (Sum != 0) { | |||
5253 | auto Probability = BranchProbability::getBranchProbability(Max, Sum); | |||
5254 | if (Probability > TLI->getPredictableBranchThreshold()) | |||
5255 | return true; | |||
5256 | } | |||
5257 | } | |||
5258 | ||||
5259 | CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition()); | |||
5260 | ||||
5261 | // If a branch is predictable, an out-of-order CPU can avoid blocking on its | |||
5262 | // comparison condition. If the compare has more than one use, there's | |||
5263 | // probably another cmov or setcc around, so it's not worth emitting a branch. | |||
5264 | if (!Cmp || !Cmp->hasOneUse()) | |||
5265 | return false; | |||
5266 | ||||
5267 | // If either operand of the select is expensive and only needed on one side | |||
5268 | // of the select, we should form a branch. | |||
5269 | if (sinkSelectOperand(TTI, SI->getTrueValue()) || | |||
5270 | sinkSelectOperand(TTI, SI->getFalseValue())) | |||
5271 | return true; | |||
5272 | ||||
5273 | return false; | |||
5274 | } | |||
5275 | ||||
5276 | /// If \p isTrue is true, return the true value of \p SI, otherwise return | |||
5277 | /// false value of \p SI. If the true/false value of \p SI is defined by any | |||
5278 | /// select instructions in \p Selects, look through the defining select | |||
5279 | /// instruction until the true/false value is not defined in \p Selects. | |||
5280 | static Value *getTrueOrFalseValue( | |||
5281 | SelectInst *SI, bool isTrue, | |||
5282 | const SmallPtrSet<const Instruction *, 2> &Selects) { | |||
5283 | Value *V; | |||
5284 | ||||
5285 | for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI); | |||
5286 | DefSI = dyn_cast<SelectInst>(V)) { | |||
5287 | assert(DefSI->getCondition() == SI->getCondition() &&((DefSI->getCondition() == SI->getCondition() && "The condition of DefSI does not match with SI") ? static_cast <void> (0) : __assert_fail ("DefSI->getCondition() == SI->getCondition() && \"The condition of DefSI does not match with SI\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5288, __PRETTY_FUNCTION__)) | |||
5288 | "The condition of DefSI does not match with SI")((DefSI->getCondition() == SI->getCondition() && "The condition of DefSI does not match with SI") ? static_cast <void> (0) : __assert_fail ("DefSI->getCondition() == SI->getCondition() && \"The condition of DefSI does not match with SI\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5288, __PRETTY_FUNCTION__)); | |||
5289 | V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue()); | |||
5290 | } | |||
5291 | return V; | |||
5292 | } | |||
5293 | ||||
5294 | /// If we have a SelectInst that will likely profit from branch prediction, | |||
5295 | /// turn it into a branch. | |||
5296 | bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) { | |||
5297 | // Find all consecutive select instructions that share the same condition. | |||
5298 | SmallVector<SelectInst *, 2> ASI; | |||
5299 | ASI.push_back(SI); | |||
5300 | for (BasicBlock::iterator It = ++BasicBlock::iterator(SI); | |||
5301 | It != SI->getParent()->end(); ++It) { | |||
5302 | SelectInst *I = dyn_cast<SelectInst>(&*It); | |||
5303 | if (I && SI->getCondition() == I->getCondition()) { | |||
5304 | ASI.push_back(I); | |||
5305 | } else { | |||
5306 | break; | |||
5307 | } | |||
5308 | } | |||
5309 | ||||
5310 | SelectInst *LastSI = ASI.back(); | |||
5311 | // Increment the current iterator to skip all the rest of select instructions | |||
5312 | // because they will be either "not lowered" or "all lowered" to branch. | |||
5313 | CurInstIterator = std::next(LastSI->getIterator()); | |||
5314 | ||||
5315 | bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1); | |||
5316 | ||||
5317 | // Can we convert the 'select' to CF ? | |||
5318 | if (DisableSelectToBranch || OptSize || !TLI || VectorCond || | |||
5319 | SI->getMetadata(LLVMContext::MD_unpredictable)) | |||
5320 | return false; | |||
5321 | ||||
5322 | TargetLowering::SelectSupportKind SelectKind; | |||
5323 | if (VectorCond) | |||
5324 | SelectKind = TargetLowering::VectorMaskSelect; | |||
5325 | else if (SI->getType()->isVectorTy()) | |||
5326 | SelectKind = TargetLowering::ScalarCondVectorVal; | |||
5327 | else | |||
5328 | SelectKind = TargetLowering::ScalarValSelect; | |||
5329 | ||||
5330 | if (TLI->isSelectSupported(SelectKind) && | |||
5331 | !isFormingBranchFromSelectProfitable(TTI, TLI, SI)) | |||
5332 | return false; | |||
5333 | ||||
5334 | ModifiedDT = true; | |||
5335 | ||||
5336 | // Transform a sequence like this: | |||
5337 | // start: | |||
5338 | // %cmp = cmp uge i32 %a, %b | |||
5339 | // %sel = select i1 %cmp, i32 %c, i32 %d | |||
5340 | // | |||
5341 | // Into: | |||
5342 | // start: | |||
5343 | // %cmp = cmp uge i32 %a, %b | |||
5344 | // br i1 %cmp, label %select.true, label %select.false | |||
5345 | // select.true: | |||
5346 | // br label %select.end | |||
5347 | // select.false: | |||
5348 | // br label %select.end | |||
5349 | // select.end: | |||
5350 | // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ] | |||
5351 | // | |||
5352 | // In addition, we may sink instructions that produce %c or %d from | |||
5353 | // the entry block into the destination(s) of the new branch. | |||
5354 | // If the true or false blocks do not contain a sunken instruction, that | |||
5355 | // block and its branch may be optimized away. In that case, one side of the | |||
5356 | // first branch will point directly to select.end, and the corresponding PHI | |||
5357 | // predecessor block will be the start block. | |||
5358 | ||||
5359 | // First, we split the block containing the select into 2 blocks. | |||
5360 | BasicBlock *StartBlock = SI->getParent(); | |||
5361 | BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI)); | |||
5362 | BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end"); | |||
5363 | ||||
5364 | // Delete the unconditional branch that was just created by the split. | |||
5365 | StartBlock->getTerminator()->eraseFromParent(); | |||
5366 | ||||
5367 | // These are the new basic blocks for the conditional branch. | |||
5368 | // At least one will become an actual new basic block. | |||
5369 | BasicBlock *TrueBlock = nullptr; | |||
5370 | BasicBlock *FalseBlock = nullptr; | |||
5371 | BranchInst *TrueBranch = nullptr; | |||
5372 | BranchInst *FalseBranch = nullptr; | |||
5373 | ||||
5374 | // Sink expensive instructions into the conditional blocks to avoid executing | |||
5375 | // them speculatively. | |||
5376 | for (SelectInst *SI : ASI) { | |||
5377 | if (sinkSelectOperand(TTI, SI->getTrueValue())) { | |||
5378 | if (TrueBlock == nullptr) { | |||
5379 | TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink", | |||
5380 | EndBlock->getParent(), EndBlock); | |||
5381 | TrueBranch = BranchInst::Create(EndBlock, TrueBlock); | |||
5382 | } | |||
5383 | auto *TrueInst = cast<Instruction>(SI->getTrueValue()); | |||
5384 | TrueInst->moveBefore(TrueBranch); | |||
5385 | } | |||
5386 | if (sinkSelectOperand(TTI, SI->getFalseValue())) { | |||
5387 | if (FalseBlock == nullptr) { | |||
5388 | FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink", | |||
5389 | EndBlock->getParent(), EndBlock); | |||
5390 | FalseBranch = BranchInst::Create(EndBlock, FalseBlock); | |||
5391 | } | |||
5392 | auto *FalseInst = cast<Instruction>(SI->getFalseValue()); | |||
5393 | FalseInst->moveBefore(FalseBranch); | |||
5394 | } | |||
5395 | } | |||
5396 | ||||
5397 | // If there was nothing to sink, then arbitrarily choose the 'false' side | |||
5398 | // for a new input value to the PHI. | |||
5399 | if (TrueBlock == FalseBlock) { | |||
5400 | assert(TrueBlock == nullptr &&((TrueBlock == nullptr && "Unexpected basic block transform while optimizing select" ) ? static_cast<void> (0) : __assert_fail ("TrueBlock == nullptr && \"Unexpected basic block transform while optimizing select\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5401, __PRETTY_FUNCTION__)) | |||
5401 | "Unexpected basic block transform while optimizing select")((TrueBlock == nullptr && "Unexpected basic block transform while optimizing select" ) ? static_cast<void> (0) : __assert_fail ("TrueBlock == nullptr && \"Unexpected basic block transform while optimizing select\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5401, __PRETTY_FUNCTION__)); | |||
5402 | ||||
5403 | FalseBlock = BasicBlock::Create(SI->getContext(), "select.false", | |||
5404 | EndBlock->getParent(), EndBlock); | |||
5405 | BranchInst::Create(EndBlock, FalseBlock); | |||
5406 | } | |||
5407 | ||||
5408 | // Insert the real conditional branch based on the original condition. | |||
5409 | // If we did not create a new block for one of the 'true' or 'false' paths | |||
5410 | // of the condition, it means that side of the branch goes to the end block | |||
5411 | // directly and the path originates from the start block from the point of | |||
5412 | // view of the new PHI. | |||
5413 | BasicBlock *TT, *FT; | |||
5414 | if (TrueBlock == nullptr) { | |||
5415 | TT = EndBlock; | |||
5416 | FT = FalseBlock; | |||
5417 | TrueBlock = StartBlock; | |||
5418 | } else if (FalseBlock == nullptr) { | |||
5419 | TT = TrueBlock; | |||
5420 | FT = EndBlock; | |||
5421 | FalseBlock = StartBlock; | |||
5422 | } else { | |||
5423 | TT = TrueBlock; | |||
5424 | FT = FalseBlock; | |||
5425 | } | |||
5426 | IRBuilder<>(SI).CreateCondBr(SI->getCondition(), TT, FT, SI); | |||
5427 | ||||
5428 | SmallPtrSet<const Instruction *, 2> INS; | |||
5429 | INS.insert(ASI.begin(), ASI.end()); | |||
5430 | // Use reverse iterator because later select may use the value of the | |||
5431 | // earlier select, and we need to propagate value through earlier select | |||
5432 | // to get the PHI operand. | |||
5433 | for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) { | |||
5434 | SelectInst *SI = *It; | |||
5435 | // The select itself is replaced with a PHI Node. | |||
5436 | PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front()); | |||
5437 | PN->takeName(SI); | |||
5438 | PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock); | |||
5439 | PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock); | |||
5440 | ||||
5441 | SI->replaceAllUsesWith(PN); | |||
5442 | SI->eraseFromParent(); | |||
5443 | INS.erase(SI); | |||
5444 | ++NumSelectsExpanded; | |||
5445 | } | |||
5446 | ||||
5447 | // Instruct OptimizeBlock to skip to the next block. | |||
5448 | CurInstIterator = StartBlock->end(); | |||
5449 | return true; | |||
5450 | } | |||
5451 | ||||
5452 | static bool isBroadcastShuffle(ShuffleVectorInst *SVI) { | |||
5453 | SmallVector<int, 16> Mask(SVI->getShuffleMask()); | |||
5454 | int SplatElem = -1; | |||
5455 | for (unsigned i = 0; i < Mask.size(); ++i) { | |||
5456 | if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem) | |||
5457 | return false; | |||
5458 | SplatElem = Mask[i]; | |||
5459 | } | |||
5460 | ||||
5461 | return true; | |||
5462 | } | |||
5463 | ||||
5464 | /// Some targets have expensive vector shifts if the lanes aren't all the same | |||
5465 | /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases | |||
5466 | /// it's often worth sinking a shufflevector splat down to its use so that | |||
5467 | /// codegen can spot all lanes are identical. | |||
5468 | bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) { | |||
5469 | BasicBlock *DefBB = SVI->getParent(); | |||
5470 | ||||
5471 | // Only do this xform if variable vector shifts are particularly expensive. | |||
5472 | if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType())) | |||
5473 | return false; | |||
5474 | ||||
5475 | // We only expect better codegen by sinking a shuffle if we can recognise a | |||
5476 | // constant splat. | |||
5477 | if (!isBroadcastShuffle(SVI)) | |||
5478 | return false; | |||
5479 | ||||
5480 | // InsertedShuffles - Only insert a shuffle in each block once. | |||
5481 | DenseMap<BasicBlock*, Instruction*> InsertedShuffles; | |||
5482 | ||||
5483 | bool MadeChange = false; | |||
5484 | for (User *U : SVI->users()) { | |||
5485 | Instruction *UI = cast<Instruction>(U); | |||
5486 | ||||
5487 | // Figure out which BB this ext is used in. | |||
5488 | BasicBlock *UserBB = UI->getParent(); | |||
5489 | if (UserBB == DefBB) continue; | |||
5490 | ||||
5491 | // For now only apply this when the splat is used by a shift instruction. | |||
5492 | if (!UI->isShift()) continue; | |||
5493 | ||||
5494 | // Everything checks out, sink the shuffle if the user's block doesn't | |||
5495 | // already have a copy. | |||
5496 | Instruction *&InsertedShuffle = InsertedShuffles[UserBB]; | |||
5497 | ||||
5498 | if (!InsertedShuffle) { | |||
5499 | BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); | |||
5500 | assert(InsertPt != UserBB->end())((InsertPt != UserBB->end()) ? static_cast<void> (0) : __assert_fail ("InsertPt != UserBB->end()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5500, __PRETTY_FUNCTION__)); | |||
5501 | InsertedShuffle = | |||
5502 | new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1), | |||
5503 | SVI->getOperand(2), "", &*InsertPt); | |||
5504 | } | |||
5505 | ||||
5506 | UI->replaceUsesOfWith(SVI, InsertedShuffle); | |||
5507 | MadeChange = true; | |||
5508 | } | |||
5509 | ||||
5510 | // If we removed all uses, nuke the shuffle. | |||
5511 | if (SVI->use_empty()) { | |||
5512 | SVI->eraseFromParent(); | |||
5513 | MadeChange = true; | |||
5514 | } | |||
5515 | ||||
5516 | return MadeChange; | |||
5517 | } | |||
5518 | ||||
5519 | bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) { | |||
5520 | if (!TLI || !DL) | |||
5521 | return false; | |||
5522 | ||||
5523 | Value *Cond = SI->getCondition(); | |||
5524 | Type *OldType = Cond->getType(); | |||
5525 | LLVMContext &Context = Cond->getContext(); | |||
5526 | MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType)); | |||
5527 | unsigned RegWidth = RegType.getSizeInBits(); | |||
5528 | ||||
5529 | if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth()) | |||
5530 | return false; | |||
5531 | ||||
5532 | // If the register width is greater than the type width, expand the condition | |||
5533 | // of the switch instruction and each case constant to the width of the | |||
5534 | // register. By widening the type of the switch condition, subsequent | |||
5535 | // comparisons (for case comparisons) will not need to be extended to the | |||
5536 | // preferred register width, so we will potentially eliminate N-1 extends, | |||
5537 | // where N is the number of cases in the switch. | |||
5538 | auto *NewType = Type::getIntNTy(Context, RegWidth); | |||
5539 | ||||
5540 | // Zero-extend the switch condition and case constants unless the switch | |||
5541 | // condition is a function argument that is already being sign-extended. | |||
5542 | // In that case, we can avoid an unnecessary mask/extension by sign-extending | |||
5543 | // everything instead. | |||
5544 | Instruction::CastOps ExtType = Instruction::ZExt; | |||
5545 | if (auto *Arg = dyn_cast<Argument>(Cond)) | |||
5546 | if (Arg->hasSExtAttr()) | |||
5547 | ExtType = Instruction::SExt; | |||
5548 | ||||
5549 | auto *ExtInst = CastInst::Create(ExtType, Cond, NewType); | |||
5550 | ExtInst->insertBefore(SI); | |||
5551 | SI->setCondition(ExtInst); | |||
5552 | for (auto Case : SI->cases()) { | |||
5553 | APInt NarrowConst = Case.getCaseValue()->getValue(); | |||
5554 | APInt WideConst = (ExtType == Instruction::ZExt) ? | |||
5555 | NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth); | |||
5556 | Case.setValue(ConstantInt::get(Context, WideConst)); | |||
5557 | } | |||
5558 | ||||
5559 | return true; | |||
5560 | } | |||
5561 | ||||
5562 | ||||
5563 | namespace { | |||
5564 | /// \brief Helper class to promote a scalar operation to a vector one. | |||
5565 | /// This class is used to move downward extractelement transition. | |||
5566 | /// E.g., | |||
5567 | /// a = vector_op <2 x i32> | |||
5568 | /// b = extractelement <2 x i32> a, i32 0 | |||
5569 | /// c = scalar_op b | |||
5570 | /// store c | |||
5571 | /// | |||
5572 | /// => | |||
5573 | /// a = vector_op <2 x i32> | |||
5574 | /// c = vector_op a (equivalent to scalar_op on the related lane) | |||
5575 | /// * d = extractelement <2 x i32> c, i32 0 | |||
5576 | /// * store d | |||
5577 | /// Assuming both extractelement and store can be combine, we get rid of the | |||
5578 | /// transition. | |||
5579 | class VectorPromoteHelper { | |||
5580 | /// DataLayout associated with the current module. | |||
5581 | const DataLayout &DL; | |||
5582 | ||||
5583 | /// Used to perform some checks on the legality of vector operations. | |||
5584 | const TargetLowering &TLI; | |||
5585 | ||||
5586 | /// Used to estimated the cost of the promoted chain. | |||
5587 | const TargetTransformInfo &TTI; | |||
5588 | ||||
5589 | /// The transition being moved downwards. | |||
5590 | Instruction *Transition; | |||
5591 | /// The sequence of instructions to be promoted. | |||
5592 | SmallVector<Instruction *, 4> InstsToBePromoted; | |||
5593 | /// Cost of combining a store and an extract. | |||
5594 | unsigned StoreExtractCombineCost; | |||
5595 | /// Instruction that will be combined with the transition. | |||
5596 | Instruction *CombineInst; | |||
5597 | ||||
5598 | /// \brief The instruction that represents the current end of the transition. | |||
5599 | /// Since we are faking the promotion until we reach the end of the chain | |||
5600 | /// of computation, we need a way to get the current end of the transition. | |||
5601 | Instruction *getEndOfTransition() const { | |||
5602 | if (InstsToBePromoted.empty()) | |||
5603 | return Transition; | |||
5604 | return InstsToBePromoted.back(); | |||
5605 | } | |||
5606 | ||||
5607 | /// \brief Return the index of the original value in the transition. | |||
5608 | /// E.g., for "extractelement <2 x i32> c, i32 1" the original value, | |||
5609 | /// c, is at index 0. | |||
5610 | unsigned getTransitionOriginalValueIdx() const { | |||
5611 | assert(isa<ExtractElementInst>(Transition) &&((isa<ExtractElementInst>(Transition) && "Other kind of transitions are not supported yet" ) ? static_cast<void> (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5612, __PRETTY_FUNCTION__)) | |||
5612 | "Other kind of transitions are not supported yet")((isa<ExtractElementInst>(Transition) && "Other kind of transitions are not supported yet" ) ? static_cast<void> (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5612, __PRETTY_FUNCTION__)); | |||
5613 | return 0; | |||
5614 | } | |||
5615 | ||||
5616 | /// \brief Return the index of the index in the transition. | |||
5617 | /// E.g., for "extractelement <2 x i32> c, i32 0" the index | |||
5618 | /// is at index 1. | |||
5619 | unsigned getTransitionIdx() const { | |||
5620 | assert(isa<ExtractElementInst>(Transition) &&((isa<ExtractElementInst>(Transition) && "Other kind of transitions are not supported yet" ) ? static_cast<void> (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5621, __PRETTY_FUNCTION__)) | |||
5621 | "Other kind of transitions are not supported yet")((isa<ExtractElementInst>(Transition) && "Other kind of transitions are not supported yet" ) ? static_cast<void> (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5621, __PRETTY_FUNCTION__)); | |||
5622 | return 1; | |||
5623 | } | |||
5624 | ||||
5625 | /// \brief Get the type of the transition. | |||
5626 | /// This is the type of the original value. | |||
5627 | /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the | |||
5628 | /// transition is <2 x i32>. | |||
5629 | Type *getTransitionType() const { | |||
5630 | return Transition->getOperand(getTransitionOriginalValueIdx())->getType(); | |||
5631 | } | |||
5632 | ||||
5633 | /// \brief Promote \p ToBePromoted by moving \p Def downward through. | |||
5634 | /// I.e., we have the following sequence: | |||
5635 | /// Def = Transition <ty1> a to <ty2> | |||
5636 | /// b = ToBePromoted <ty2> Def, ... | |||
5637 | /// => | |||
5638 | /// b = ToBePromoted <ty1> a, ... | |||
5639 | /// Def = Transition <ty1> ToBePromoted to <ty2> | |||
5640 | void promoteImpl(Instruction *ToBePromoted); | |||
5641 | ||||
5642 | /// \brief Check whether or not it is profitable to promote all the | |||
5643 | /// instructions enqueued to be promoted. | |||
5644 | bool isProfitableToPromote() { | |||
5645 | Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx()); | |||
5646 | unsigned Index = isa<ConstantInt>(ValIdx) | |||
5647 | ? cast<ConstantInt>(ValIdx)->getZExtValue() | |||
5648 | : -1; | |||
5649 | Type *PromotedType = getTransitionType(); | |||
5650 | ||||
5651 | StoreInst *ST = cast<StoreInst>(CombineInst); | |||
5652 | unsigned AS = ST->getPointerAddressSpace(); | |||
5653 | unsigned Align = ST->getAlignment(); | |||
5654 | // Check if this store is supported. | |||
5655 | if (!TLI.allowsMisalignedMemoryAccesses( | |||
5656 | TLI.getValueType(DL, ST->getValueOperand()->getType()), AS, | |||
5657 | Align)) { | |||
5658 | // If this is not supported, there is no way we can combine | |||
5659 | // the extract with the store. | |||
5660 | return false; | |||
5661 | } | |||
5662 | ||||
5663 | // The scalar chain of computation has to pay for the transition | |||
5664 | // scalar to vector. | |||
5665 | // The vector chain has to account for the combining cost. | |||
5666 | uint64_t ScalarCost = | |||
5667 | TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index); | |||
5668 | uint64_t VectorCost = StoreExtractCombineCost; | |||
5669 | for (const auto &Inst : InstsToBePromoted) { | |||
5670 | // Compute the cost. | |||
5671 | // By construction, all instructions being promoted are arithmetic ones. | |||
5672 | // Moreover, one argument is a constant that can be viewed as a splat | |||
5673 | // constant. | |||
5674 | Value *Arg0 = Inst->getOperand(0); | |||
5675 | bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) || | |||
5676 | isa<ConstantFP>(Arg0); | |||
5677 | TargetTransformInfo::OperandValueKind Arg0OVK = | |||
5678 | IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue | |||
5679 | : TargetTransformInfo::OK_AnyValue; | |||
5680 | TargetTransformInfo::OperandValueKind Arg1OVK = | |||
5681 | !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue | |||
5682 | : TargetTransformInfo::OK_AnyValue; | |||
5683 | ScalarCost += TTI.getArithmeticInstrCost( | |||
5684 | Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK); | |||
5685 | VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType, | |||
5686 | Arg0OVK, Arg1OVK); | |||
5687 | } | |||
5688 | DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Estimated cost of computation to be promoted:\nScalar: " << ScalarCost << "\nVector: " << VectorCost << '\n'; } } while (false) | |||
5689 | << ScalarCost << "\nVector: " << VectorCost << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Estimated cost of computation to be promoted:\nScalar: " << ScalarCost << "\nVector: " << VectorCost << '\n'; } } while (false); | |||
5690 | return ScalarCost > VectorCost; | |||
5691 | } | |||
5692 | ||||
5693 | /// \brief Generate a constant vector with \p Val with the same | |||
5694 | /// number of elements as the transition. | |||
5695 | /// \p UseSplat defines whether or not \p Val should be replicated | |||
5696 | /// across the whole vector. | |||
5697 | /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>, | |||
5698 | /// otherwise we generate a vector with as many undef as possible: | |||
5699 | /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only | |||
5700 | /// used at the index of the extract. | |||
5701 | Value *getConstantVector(Constant *Val, bool UseSplat) const { | |||
5702 | unsigned ExtractIdx = UINT_MAX(2147483647 *2U +1U); | |||
5703 | if (!UseSplat) { | |||
5704 | // If we cannot determine where the constant must be, we have to | |||
5705 | // use a splat constant. | |||
5706 | Value *ValExtractIdx = Transition->getOperand(getTransitionIdx()); | |||
5707 | if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx)) | |||
5708 | ExtractIdx = CstVal->getSExtValue(); | |||
5709 | else | |||
5710 | UseSplat = true; | |||
5711 | } | |||
5712 | ||||
5713 | unsigned End = getTransitionType()->getVectorNumElements(); | |||
5714 | if (UseSplat) | |||
5715 | return ConstantVector::getSplat(End, Val); | |||
5716 | ||||
5717 | SmallVector<Constant *, 4> ConstVec; | |||
5718 | UndefValue *UndefVal = UndefValue::get(Val->getType()); | |||
5719 | for (unsigned Idx = 0; Idx != End; ++Idx) { | |||
5720 | if (Idx == ExtractIdx) | |||
5721 | ConstVec.push_back(Val); | |||
5722 | else | |||
5723 | ConstVec.push_back(UndefVal); | |||
5724 | } | |||
5725 | return ConstantVector::get(ConstVec); | |||
5726 | } | |||
5727 | ||||
5728 | /// \brief Check if promoting to a vector type an operand at \p OperandIdx | |||
5729 | /// in \p Use can trigger undefined behavior. | |||
5730 | static bool canCauseUndefinedBehavior(const Instruction *Use, | |||
5731 | unsigned OperandIdx) { | |||
5732 | // This is not safe to introduce undef when the operand is on | |||
5733 | // the right hand side of a division-like instruction. | |||
5734 | if (OperandIdx != 1) | |||
5735 | return false; | |||
5736 | switch (Use->getOpcode()) { | |||
5737 | default: | |||
5738 | return false; | |||
5739 | case Instruction::SDiv: | |||
5740 | case Instruction::UDiv: | |||
5741 | case Instruction::SRem: | |||
5742 | case Instruction::URem: | |||
5743 | return true; | |||
5744 | case Instruction::FDiv: | |||
5745 | case Instruction::FRem: | |||
5746 | return !Use->hasNoNaNs(); | |||
5747 | } | |||
5748 | llvm_unreachable(nullptr)::llvm::llvm_unreachable_internal(nullptr, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5748); | |||
5749 | } | |||
5750 | ||||
5751 | public: | |||
5752 | VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI, | |||
5753 | const TargetTransformInfo &TTI, Instruction *Transition, | |||
5754 | unsigned CombineCost) | |||
5755 | : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition), | |||
5756 | StoreExtractCombineCost(CombineCost), CombineInst(nullptr) { | |||
5757 | assert(Transition && "Do not know how to promote null")((Transition && "Do not know how to promote null") ? static_cast <void> (0) : __assert_fail ("Transition && \"Do not know how to promote null\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5757, __PRETTY_FUNCTION__)); | |||
5758 | } | |||
5759 | ||||
5760 | /// \brief Check if we can promote \p ToBePromoted to \p Type. | |||
5761 | bool canPromote(const Instruction *ToBePromoted) const { | |||
5762 | // We could support CastInst too. | |||
5763 | return isa<BinaryOperator>(ToBePromoted); | |||
5764 | } | |||
5765 | ||||
5766 | /// \brief Check if it is profitable to promote \p ToBePromoted | |||
5767 | /// by moving downward the transition through. | |||
5768 | bool shouldPromote(const Instruction *ToBePromoted) const { | |||
5769 | // Promote only if all the operands can be statically expanded. | |||
5770 | // Indeed, we do not want to introduce any new kind of transitions. | |||
5771 | for (const Use &U : ToBePromoted->operands()) { | |||
5772 | const Value *Val = U.get(); | |||
5773 | if (Val == getEndOfTransition()) { | |||
5774 | // If the use is a division and the transition is on the rhs, | |||
5775 | // we cannot promote the operation, otherwise we may create a | |||
5776 | // division by zero. | |||
5777 | if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo())) | |||
5778 | return false; | |||
5779 | continue; | |||
5780 | } | |||
5781 | if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) && | |||
5782 | !isa<ConstantFP>(Val)) | |||
5783 | return false; | |||
5784 | } | |||
5785 | // Check that the resulting operation is legal. | |||
5786 | int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode()); | |||
5787 | if (!ISDOpcode) | |||
5788 | return false; | |||
5789 | return StressStoreExtract || | |||
5790 | TLI.isOperationLegalOrCustom( | |||
5791 | ISDOpcode, TLI.getValueType(DL, getTransitionType(), true)); | |||
5792 | } | |||
5793 | ||||
5794 | /// \brief Check whether or not \p Use can be combined | |||
5795 | /// with the transition. | |||
5796 | /// I.e., is it possible to do Use(Transition) => AnotherUse? | |||
5797 | bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); } | |||
5798 | ||||
5799 | /// \brief Record \p ToBePromoted as part of the chain to be promoted. | |||
5800 | void enqueueForPromotion(Instruction *ToBePromoted) { | |||
5801 | InstsToBePromoted.push_back(ToBePromoted); | |||
5802 | } | |||
5803 | ||||
5804 | /// \brief Set the instruction that will be combined with the transition. | |||
5805 | void recordCombineInstruction(Instruction *ToBeCombined) { | |||
5806 | assert(canCombine(ToBeCombined) && "Unsupported instruction to combine")((canCombine(ToBeCombined) && "Unsupported instruction to combine" ) ? static_cast<void> (0) : __assert_fail ("canCombine(ToBeCombined) && \"Unsupported instruction to combine\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5806, __PRETTY_FUNCTION__)); | |||
5807 | CombineInst = ToBeCombined; | |||
5808 | } | |||
5809 | ||||
5810 | /// \brief Promote all the instructions enqueued for promotion if it is | |||
5811 | /// is profitable. | |||
5812 | /// \return True if the promotion happened, false otherwise. | |||
5813 | bool promote() { | |||
5814 | // Check if there is something to promote. | |||
5815 | // Right now, if we do not have anything to combine with, | |||
5816 | // we assume the promotion is not profitable. | |||
5817 | if (InstsToBePromoted.empty() || !CombineInst) | |||
5818 | return false; | |||
5819 | ||||
5820 | // Check cost. | |||
5821 | if (!StressStoreExtract && !isProfitableToPromote()) | |||
5822 | return false; | |||
5823 | ||||
5824 | // Promote. | |||
5825 | for (auto &ToBePromoted : InstsToBePromoted) | |||
5826 | promoteImpl(ToBePromoted); | |||
5827 | InstsToBePromoted.clear(); | |||
5828 | return true; | |||
5829 | } | |||
5830 | }; | |||
5831 | } // End of anonymous namespace. | |||
5832 | ||||
5833 | void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) { | |||
5834 | // At this point, we know that all the operands of ToBePromoted but Def | |||
5835 | // can be statically promoted. | |||
5836 | // For Def, we need to use its parameter in ToBePromoted: | |||
5837 | // b = ToBePromoted ty1 a | |||
5838 | // Def = Transition ty1 b to ty2 | |||
5839 | // Move the transition down. | |||
5840 | // 1. Replace all uses of the promoted operation by the transition. | |||
5841 | // = ... b => = ... Def. | |||
5842 | assert(ToBePromoted->getType() == Transition->getType() &&((ToBePromoted->getType() == Transition->getType() && "The type of the result of the transition does not match " "the final type" ) ? static_cast<void> (0) : __assert_fail ("ToBePromoted->getType() == Transition->getType() && \"The type of the result of the transition does not match \" \"the final type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5844, __PRETTY_FUNCTION__)) | |||
5843 | "The type of the result of the transition does not match "((ToBePromoted->getType() == Transition->getType() && "The type of the result of the transition does not match " "the final type" ) ? static_cast<void> (0) : __assert_fail ("ToBePromoted->getType() == Transition->getType() && \"The type of the result of the transition does not match \" \"the final type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5844, __PRETTY_FUNCTION__)) | |||
5844 | "the final type")((ToBePromoted->getType() == Transition->getType() && "The type of the result of the transition does not match " "the final type" ) ? static_cast<void> (0) : __assert_fail ("ToBePromoted->getType() == Transition->getType() && \"The type of the result of the transition does not match \" \"the final type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5844, __PRETTY_FUNCTION__)); | |||
5845 | ToBePromoted->replaceAllUsesWith(Transition); | |||
5846 | // 2. Update the type of the uses. | |||
5847 | // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def. | |||
5848 | Type *TransitionTy = getTransitionType(); | |||
5849 | ToBePromoted->mutateType(TransitionTy); | |||
5850 | // 3. Update all the operands of the promoted operation with promoted | |||
5851 | // operands. | |||
5852 | // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a. | |||
5853 | for (Use &U : ToBePromoted->operands()) { | |||
5854 | Value *Val = U.get(); | |||
5855 | Value *NewVal = nullptr; | |||
5856 | if (Val == Transition) | |||
5857 | NewVal = Transition->getOperand(getTransitionOriginalValueIdx()); | |||
5858 | else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) || | |||
5859 | isa<ConstantFP>(Val)) { | |||
5860 | // Use a splat constant if it is not safe to use undef. | |||
5861 | NewVal = getConstantVector( | |||
5862 | cast<Constant>(Val), | |||
5863 | isa<UndefValue>(Val) || | |||
5864 | canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo())); | |||
5865 | } else | |||
5866 | llvm_unreachable("Did you modified shouldPromote and forgot to update "::llvm::llvm_unreachable_internal("Did you modified shouldPromote and forgot to update " "this?", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5867) | |||
5867 | "this?")::llvm::llvm_unreachable_internal("Did you modified shouldPromote and forgot to update " "this?", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/CodeGen/CodeGenPrepare.cpp" , 5867); | |||
5868 | ToBePromoted->setOperand(U.getOperandNo(), NewVal); | |||
5869 | } | |||
5870 | Transition->removeFromParent(); | |||
5871 | Transition->insertAfter(ToBePromoted); | |||
5872 | Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted); | |||
5873 | } | |||
5874 | ||||
5875 | /// Some targets can do store(extractelement) with one instruction. | |||
5876 | /// Try to push the extractelement towards the stores when the target | |||
5877 | /// has this feature and this is profitable. | |||
5878 | bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) { | |||
5879 | unsigned CombineCost = UINT_MAX(2147483647 *2U +1U); | |||
5880 | if (DisableStoreExtract || !TLI || | |||
5881 | (!StressStoreExtract && | |||
5882 | !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(), | |||
5883 | Inst->getOperand(1), CombineCost))) | |||
5884 | return false; | |||
5885 | ||||
5886 | // At this point we know that Inst is a vector to scalar transition. | |||
5887 | // Try to move it down the def-use chain, until: | |||
5888 | // - We can combine the transition with its single use | |||
5889 | // => we got rid of the transition. | |||
5890 | // - We escape the current basic block | |||
5891 | // => we would need to check that we are moving it at a cheaper place and | |||
5892 | // we do not do that for now. | |||
5893 | BasicBlock *Parent = Inst->getParent(); | |||
5894 | DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Found an interesting transition: " << *Inst << '\n'; } } while (false); | |||
5895 | VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost); | |||
5896 | // If the transition has more than one use, assume this is not going to be | |||
5897 | // beneficial. | |||
5898 | while (Inst->hasOneUse()) { | |||
5899 | Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin()); | |||
5900 | DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Use: " << *ToBePromoted << '\n'; } } while (false); | |||
5901 | ||||
5902 | if (ToBePromoted->getParent() != Parent) { | |||
5903 | DEBUG(dbgs() << "Instruction to promote is in a different block ("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Instruction to promote is in a different block (" << ToBePromoted->getParent()->getName() << ") than the transition (" << Parent->getName() << ").\n"; } } while (false) | |||
5904 | << ToBePromoted->getParent()->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Instruction to promote is in a different block (" << ToBePromoted->getParent()->getName() << ") than the transition (" << Parent->getName() << ").\n"; } } while (false) | |||
5905 | << ") than the transition (" << Parent->getName() << ").\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Instruction to promote is in a different block (" << ToBePromoted->getParent()->getName() << ") than the transition (" << Parent->getName() << ").\n"; } } while (false); | |||
5906 | return false; | |||
5907 | } | |||
5908 | ||||
5909 | if (VPH.canCombine(ToBePromoted)) { | |||
5910 | DEBUG(dbgs() << "Assume " << *Inst << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Assume " << *Inst << '\n' << "will be combined with: " << *ToBePromoted << '\n'; } } while (false) | |||
5911 | << "will be combined with: " << *ToBePromoted << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Assume " << *Inst << '\n' << "will be combined with: " << *ToBePromoted << '\n'; } } while (false); | |||
5912 | VPH.recordCombineInstruction(ToBePromoted); | |||
5913 | bool Changed = VPH.promote(); | |||
5914 | NumStoreExtractExposed += Changed; | |||
5915 | return Changed; | |||
5916 | } | |||
5917 | ||||
5918 | DEBUG(dbgs() << "Try promoting.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Try promoting.\n"; } } while (false); | |||
5919 | if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted)) | |||
5920 | return false; | |||
5921 | ||||
5922 | DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Promoting is possible... Enqueue for promotion!\n" ; } } while (false); | |||
5923 | ||||
5924 | VPH.enqueueForPromotion(ToBePromoted); | |||
5925 | Inst = ToBePromoted; | |||
5926 | } | |||
5927 | return false; | |||
5928 | } | |||
5929 | ||||
5930 | /// For the instruction sequence of store below, F and I values | |||
5931 | /// are bundled together as an i64 value before being stored into memory. | |||
5932 | /// Sometimes it is more efficent to generate separate stores for F and I, | |||
5933 | /// which can remove the bitwise instructions or sink them to colder places. | |||
5934 | /// | |||
5935 | /// (store (or (zext (bitcast F to i32) to i64), | |||
5936 | /// (shl (zext I to i64), 32)), addr) --> | |||
5937 | /// (store F, addr) and (store I, addr+4) | |||
5938 | /// | |||
5939 | /// Similarly, splitting for other merged store can also be beneficial, like: | |||
5940 | /// For pair of {i32, i32}, i64 store --> two i32 stores. | |||
5941 | /// For pair of {i32, i16}, i64 store --> two i32 stores. | |||
5942 | /// For pair of {i16, i16}, i32 store --> two i16 stores. | |||
5943 | /// For pair of {i16, i8}, i32 store --> two i16 stores. | |||
5944 | /// For pair of {i8, i8}, i16 store --> two i8 stores. | |||
5945 | /// | |||
5946 | /// We allow each target to determine specifically which kind of splitting is | |||
5947 | /// supported. | |||
5948 | /// | |||
5949 | /// The store patterns are commonly seen from the simple code snippet below | |||
5950 | /// if only std::make_pair(...) is sroa transformed before inlined into hoo. | |||
5951 | /// void goo(const std::pair<int, float> &); | |||
5952 | /// hoo() { | |||
5953 | /// ... | |||
5954 | /// goo(std::make_pair(tmp, ftmp)); | |||
5955 | /// ... | |||
5956 | /// } | |||
5957 | /// | |||
5958 | /// Although we already have similar splitting in DAG Combine, we duplicate | |||
5959 | /// it in CodeGenPrepare to catch the case in which pattern is across | |||
5960 | /// multiple BBs. The logic in DAG Combine is kept to catch case generated | |||
5961 | /// during code expansion. | |||
5962 | static bool splitMergedValStore(StoreInst &SI, const DataLayout &DL, | |||
5963 | const TargetLowering &TLI) { | |||
5964 | // Handle simple but common cases only. | |||
5965 | Type *StoreType = SI.getValueOperand()->getType(); | |||
5966 | if (DL.getTypeStoreSizeInBits(StoreType) != DL.getTypeSizeInBits(StoreType) || | |||
5967 | DL.getTypeSizeInBits(StoreType) == 0) | |||
5968 | return false; | |||
5969 | ||||
5970 | unsigned HalfValBitSize = DL.getTypeSizeInBits(StoreType) / 2; | |||
5971 | Type *SplitStoreType = Type::getIntNTy(SI.getContext(), HalfValBitSize); | |||
5972 | if (DL.getTypeStoreSizeInBits(SplitStoreType) != | |||
5973 | DL.getTypeSizeInBits(SplitStoreType)) | |||
5974 | return false; | |||
5975 | ||||
5976 | // Match the following patterns: | |||
5977 | // (store (or (zext LValue to i64), | |||
5978 | // (shl (zext HValue to i64), 32)), HalfValBitSize) | |||
5979 | // or | |||
5980 | // (store (or (shl (zext HValue to i64), 32)), HalfValBitSize) | |||
5981 | // (zext LValue to i64), | |||
5982 | // Expect both operands of OR and the first operand of SHL have only | |||
5983 | // one use. | |||
5984 | Value *LValue, *HValue; | |||
5985 | if (!match(SI.getValueOperand(), | |||
5986 | m_c_Or(m_OneUse(m_ZExt(m_Value(LValue))), | |||
5987 | m_OneUse(m_Shl(m_OneUse(m_ZExt(m_Value(HValue))), | |||
5988 | m_SpecificInt(HalfValBitSize)))))) | |||
5989 | return false; | |||
5990 | ||||
5991 | // Check LValue and HValue are int with size less or equal than 32. | |||
5992 | if (!LValue->getType()->isIntegerTy() || | |||
5993 | DL.getTypeSizeInBits(LValue->getType()) > HalfValBitSize || | |||
5994 | !HValue->getType()->isIntegerTy() || | |||
5995 | DL.getTypeSizeInBits(HValue->getType()) > HalfValBitSize) | |||
5996 | return false; | |||
5997 | ||||
5998 | // If LValue/HValue is a bitcast instruction, use the EVT before bitcast | |||
5999 | // as the input of target query. | |||
6000 | auto *LBC = dyn_cast<BitCastInst>(LValue); | |||
6001 | auto *HBC = dyn_cast<BitCastInst>(HValue); | |||
6002 | EVT LowTy = LBC ? EVT::getEVT(LBC->getOperand(0)->getType()) | |||
6003 | : EVT::getEVT(LValue->getType()); | |||
6004 | EVT HighTy = HBC ? EVT::getEVT(HBC->getOperand(0)->getType()) | |||
6005 | : EVT::getEVT(HValue->getType()); | |||
6006 | if (!ForceSplitStore && !TLI.isMultiStoresCheaperThanBitsMerge(LowTy, HighTy)) | |||
6007 | return false; | |||
6008 | ||||
6009 | // Start to split store. | |||
6010 | IRBuilder<> Builder(SI.getContext()); | |||
6011 | Builder.SetInsertPoint(&SI); | |||
6012 | ||||
6013 | // If LValue/HValue is a bitcast in another BB, create a new one in current | |||
6014 | // BB so it may be merged with the splitted stores by dag combiner. | |||
6015 | if (LBC && LBC->getParent() != SI.getParent()) | |||
6016 | LValue = Builder.CreateBitCast(LBC->getOperand(0), LBC->getType()); | |||
6017 | if (HBC && HBC->getParent() != SI.getParent()) | |||
6018 | HValue = Builder.CreateBitCast(HBC->getOperand(0), HBC->getType()); | |||
6019 | ||||
6020 | auto CreateSplitStore = [&](Value *V, bool Upper) { | |||
6021 | V = Builder.CreateZExtOrBitCast(V, SplitStoreType); | |||
6022 | Value *Addr = Builder.CreateBitCast( | |||
6023 | SI.getOperand(1), | |||
6024 | SplitStoreType->getPointerTo(SI.getPointerAddressSpace())); | |||
6025 | if (Upper) | |||
6026 | Addr = Builder.CreateGEP( | |||
6027 | SplitStoreType, Addr, | |||
6028 | ConstantInt::get(Type::getInt32Ty(SI.getContext()), 1)); | |||
6029 | Builder.CreateAlignedStore( | |||
6030 | V, Addr, Upper ? SI.getAlignment() / 2 : SI.getAlignment()); | |||
6031 | }; | |||
6032 | ||||
6033 | CreateSplitStore(LValue, false); | |||
6034 | CreateSplitStore(HValue, true); | |||
6035 | ||||
6036 | // Delete the old store. | |||
6037 | SI.eraseFromParent(); | |||
6038 | return true; | |||
6039 | } | |||
6040 | ||||
6041 | bool CodeGenPrepare::optimizeInst(Instruction *I, bool &ModifiedDT) { | |||
6042 | // Bail out if we inserted the instruction to prevent optimizations from | |||
6043 | // stepping on each other's toes. | |||
6044 | if (InsertedInsts.count(I)) | |||
6045 | return false; | |||
6046 | ||||
6047 | if (PHINode *P = dyn_cast<PHINode>(I)) { | |||
6048 | // It is possible for very late stage optimizations (such as SimplifyCFG) | |||
6049 | // to introduce PHI nodes too late to be cleaned up. If we detect such a | |||
6050 | // trivial PHI, go ahead and zap it here. | |||
6051 | if (Value *V = SimplifyInstruction(P, {*DL, TLInfo})) { | |||
6052 | P->replaceAllUsesWith(V); | |||
6053 | P->eraseFromParent(); | |||
6054 | ++NumPHIsElim; | |||
6055 | return true; | |||
6056 | } | |||
6057 | return false; | |||
6058 | } | |||
6059 | ||||
6060 | if (CastInst *CI = dyn_cast<CastInst>(I)) { | |||
6061 | // If the source of the cast is a constant, then this should have | |||
6062 | // already been constant folded. The only reason NOT to constant fold | |||
6063 | // it is if something (e.g. LSR) was careful to place the constant | |||
6064 | // evaluation in a block other than then one that uses it (e.g. to hoist | |||
6065 | // the address of globals out of a loop). If this is the case, we don't | |||
6066 | // want to forward-subst the cast. | |||
6067 | if (isa<Constant>(CI->getOperand(0))) | |||
6068 | return false; | |||
6069 | ||||
6070 | if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL)) | |||
6071 | return true; | |||
6072 | ||||
6073 | if (isa<ZExtInst>(I) || isa<SExtInst>(I)) { | |||
6074 | /// Sink a zext or sext into its user blocks if the target type doesn't | |||
6075 | /// fit in one register | |||
6076 | if (TLI && | |||
6077 | TLI->getTypeAction(CI->getContext(), | |||
6078 | TLI->getValueType(*DL, CI->getType())) == | |||
6079 | TargetLowering::TypeExpandInteger) { | |||
6080 | return SinkCast(CI); | |||
6081 | } else { | |||
6082 | bool MadeChange = optimizeExt(I); | |||
6083 | return MadeChange | optimizeExtUses(I); | |||
6084 | } | |||
6085 | } | |||
6086 | return false; | |||
6087 | } | |||
6088 | ||||
6089 | if (CmpInst *CI = dyn_cast<CmpInst>(I)) | |||
6090 | if (!TLI || !TLI->hasMultipleConditionRegisters()) | |||
6091 | return OptimizeCmpExpression(CI, TLI); | |||
6092 | ||||
6093 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) { | |||
6094 | LI->setMetadata(LLVMContext::MD_invariant_group, nullptr); | |||
6095 | if (TLI) { | |||
6096 | bool Modified = optimizeLoadExt(LI); | |||
6097 | unsigned AS = LI->getPointerAddressSpace(); | |||
6098 | Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS); | |||
6099 | return Modified; | |||
6100 | } | |||
6101 | return false; | |||
6102 | } | |||
6103 | ||||
6104 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) { | |||
6105 | if (TLI && splitMergedValStore(*SI, *DL, *TLI)) | |||
6106 | return true; | |||
6107 | SI->setMetadata(LLVMContext::MD_invariant_group, nullptr); | |||
6108 | if (TLI) { | |||
6109 | unsigned AS = SI->getPointerAddressSpace(); | |||
6110 | return optimizeMemoryInst(I, SI->getOperand(1), | |||
6111 | SI->getOperand(0)->getType(), AS); | |||
6112 | } | |||
6113 | return false; | |||
6114 | } | |||
6115 | ||||
6116 | if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) { | |||
6117 | unsigned AS = RMW->getPointerAddressSpace(); | |||
6118 | return optimizeMemoryInst(I, RMW->getPointerOperand(), | |||
6119 | RMW->getType(), AS); | |||
6120 | } | |||
6121 | ||||
6122 | if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(I)) { | |||
6123 | unsigned AS = CmpX->getPointerAddressSpace(); | |||
6124 | return optimizeMemoryInst(I, CmpX->getPointerOperand(), | |||
6125 | CmpX->getCompareOperand()->getType(), AS); | |||
6126 | } | |||
6127 | ||||
6128 | BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I); | |||
6129 | ||||
6130 | if (BinOp && (BinOp->getOpcode() == Instruction::And) && | |||
6131 | EnableAndCmpSinking && TLI) | |||
6132 | return sinkAndCmp0Expression(BinOp, *TLI, InsertedInsts); | |||
6133 | ||||
6134 | if (BinOp && (BinOp->getOpcode() == Instruction::AShr || | |||
6135 | BinOp->getOpcode() == Instruction::LShr)) { | |||
6136 | ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1)); | |||
6137 | if (TLI && CI && TLI->hasExtractBitsInsn()) | |||
6138 | return OptimizeExtractBits(BinOp, CI, *TLI, *DL); | |||
6139 | ||||
6140 | return false; | |||
6141 | } | |||
6142 | ||||
6143 | if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { | |||
6144 | if (GEPI->hasAllZeroIndices()) { | |||
6145 | /// The GEP operand must be a pointer, so must its result -> BitCast | |||
6146 | Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), | |||
6147 | GEPI->getName(), GEPI); | |||
6148 | GEPI->replaceAllUsesWith(NC); | |||
6149 | GEPI->eraseFromParent(); | |||
6150 | ++NumGEPsElim; | |||
6151 | optimizeInst(NC, ModifiedDT); | |||
6152 | return true; | |||
6153 | } | |||
6154 | return false; | |||
6155 | } | |||
6156 | ||||
6157 | if (CallInst *CI = dyn_cast<CallInst>(I)) | |||
6158 | return optimizeCallInst(CI, ModifiedDT); | |||
6159 | ||||
6160 | if (SelectInst *SI = dyn_cast<SelectInst>(I)) | |||
6161 | return optimizeSelectInst(SI); | |||
6162 | ||||
6163 | if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) | |||
6164 | return optimizeShuffleVectorInst(SVI); | |||
6165 | ||||
6166 | if (auto *Switch = dyn_cast<SwitchInst>(I)) | |||
6167 | return optimizeSwitchInst(Switch); | |||
6168 | ||||
6169 | if (isa<ExtractElementInst>(I)) | |||
6170 | return optimizeExtractElementInst(I); | |||
6171 | ||||
6172 | return false; | |||
6173 | } | |||
6174 | ||||
6175 | /// Given an OR instruction, check to see if this is a bitreverse | |||
6176 | /// idiom. If so, insert the new intrinsic and return true. | |||
6177 | static bool makeBitReverse(Instruction &I, const DataLayout &DL, | |||
6178 | const TargetLowering &TLI) { | |||
6179 | if (!I.getType()->isIntegerTy() || | |||
6180 | !TLI.isOperationLegalOrCustom(ISD::BITREVERSE, | |||
6181 | TLI.getValueType(DL, I.getType(), true))) | |||
6182 | return false; | |||
6183 | ||||
6184 | SmallVector<Instruction*, 4> Insts; | |||
6185 | if (!recognizeBSwapOrBitReverseIdiom(&I, false, true, Insts)) | |||
6186 | return false; | |||
6187 | Instruction *LastInst = Insts.back(); | |||
6188 | I.replaceAllUsesWith(LastInst); | |||
6189 | RecursivelyDeleteTriviallyDeadInstructions(&I); | |||
6190 | return true; | |||
6191 | } | |||
6192 | ||||
6193 | // In this pass we look for GEP and cast instructions that are used | |||
6194 | // across basic blocks and rewrite them to improve basic-block-at-a-time | |||
6195 | // selection. | |||
6196 | bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool &ModifiedDT) { | |||
6197 | SunkAddrs.clear(); | |||
6198 | bool MadeChange = false; | |||
6199 | ||||
6200 | CurInstIterator = BB.begin(); | |||
6201 | while (CurInstIterator != BB.end()) { | |||
6202 | MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT); | |||
6203 | if (ModifiedDT) | |||
6204 | return true; | |||
6205 | } | |||
6206 | ||||
6207 | bool MadeBitReverse = true; | |||
6208 | while (TLI && MadeBitReverse) { | |||
6209 | MadeBitReverse = false; | |||
6210 | for (auto &I : reverse(BB)) { | |||
6211 | if (makeBitReverse(I, *DL, *TLI)) { | |||
6212 | MadeBitReverse = MadeChange = true; | |||
6213 | ModifiedDT = true; | |||
6214 | break; | |||
6215 | } | |||
6216 | } | |||
6217 | } | |||
6218 | MadeChange |= dupRetToEnableTailCallOpts(&BB); | |||
6219 | ||||
6220 | return MadeChange; | |||
6221 | } | |||
6222 | ||||
6223 | // llvm.dbg.value is far away from the value then iSel may not be able | |||
6224 | // handle it properly. iSel will drop llvm.dbg.value if it can not | |||
6225 | // find a node corresponding to the value. | |||
6226 | bool CodeGenPrepare::placeDbgValues(Function &F) { | |||
6227 | bool MadeChange = false; | |||
6228 | for (BasicBlock &BB : F) { | |||
6229 | Instruction *PrevNonDbgInst = nullptr; | |||
6230 | for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) { | |||
6231 | Instruction *Insn = &*BI++; | |||
6232 | DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn); | |||
6233 | // Leave dbg.values that refer to an alloca alone. These | |||
6234 | // instrinsics describe the address of a variable (= the alloca) | |||
6235 | // being taken. They should not be moved next to the alloca | |||
6236 | // (and to the beginning of the scope), but rather stay close to | |||
6237 | // where said address is used. | |||
6238 | if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) { | |||
6239 | PrevNonDbgInst = Insn; | |||
6240 | continue; | |||
6241 | } | |||
6242 | ||||
6243 | Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue()); | |||
6244 | if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) { | |||
6245 | // If VI is a phi in a block with an EHPad terminator, we can't insert | |||
6246 | // after it. | |||
6247 | if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad()) | |||
6248 | continue; | |||
6249 | DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI; } } while (false); | |||
6250 | DVI->removeFromParent(); | |||
6251 | if (isa<PHINode>(VI)) | |||
6252 | DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt()); | |||
6253 | else | |||
6254 | DVI->insertAfter(VI); | |||
6255 | MadeChange = true; | |||
6256 | ++NumDbgValueMoved; | |||
6257 | } | |||
6258 | } | |||
6259 | } | |||
6260 | return MadeChange; | |||
6261 | } | |||
6262 | ||||
6263 | /// \brief Scale down both weights to fit into uint32_t. | |||
6264 | static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) { | |||
6265 | uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse; | |||
6266 | uint32_t Scale = (NewMax / UINT32_MAX(4294967295U)) + 1; | |||
6267 | NewTrue = NewTrue / Scale; | |||
6268 | NewFalse = NewFalse / Scale; | |||
6269 | } | |||
6270 | ||||
6271 | /// \brief Some targets prefer to split a conditional branch like: | |||
6272 | /// \code | |||
6273 | /// %0 = icmp ne i32 %a, 0 | |||
6274 | /// %1 = icmp ne i32 %b, 0 | |||
6275 | /// %or.cond = or i1 %0, %1 | |||
6276 | /// br i1 %or.cond, label %TrueBB, label %FalseBB | |||
6277 | /// \endcode | |||
6278 | /// into multiple branch instructions like: | |||
6279 | /// \code | |||
6280 | /// bb1: | |||
6281 | /// %0 = icmp ne i32 %a, 0 | |||
6282 | /// br i1 %0, label %TrueBB, label %bb2 | |||
6283 | /// bb2: | |||
6284 | /// %1 = icmp ne i32 %b, 0 | |||
6285 | /// br i1 %1, label %TrueBB, label %FalseBB | |||
6286 | /// \endcode | |||
6287 | /// This usually allows instruction selection to do even further optimizations | |||
6288 | /// and combine the compare with the branch instruction. Currently this is | |||
6289 | /// applied for targets which have "cheap" jump instructions. | |||
6290 | /// | |||
6291 | /// FIXME: Remove the (equivalent?) implementation in SelectionDAG. | |||
6292 | /// | |||
6293 | bool CodeGenPrepare::splitBranchCondition(Function &F) { | |||
6294 | if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive()) | |||
6295 | return false; | |||
6296 | ||||
6297 | bool MadeChange = false; | |||
6298 | for (auto &BB : F) { | |||
6299 | // Does this BB end with the following? | |||
6300 | // %cond1 = icmp|fcmp|binary instruction ... | |||
6301 | // %cond2 = icmp|fcmp|binary instruction ... | |||
6302 | // %cond.or = or|and i1 %cond1, cond2 | |||
6303 | // br i1 %cond.or label %dest1, label %dest2" | |||
6304 | BinaryOperator *LogicOp; | |||
6305 | BasicBlock *TBB, *FBB; | |||
6306 | if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB))) | |||
6307 | continue; | |||
6308 | ||||
6309 | auto *Br1 = cast<BranchInst>(BB.getTerminator()); | |||
6310 | if (Br1->getMetadata(LLVMContext::MD_unpredictable)) | |||
6311 | continue; | |||
6312 | ||||
6313 | unsigned Opc; | |||
6314 | Value *Cond1, *Cond2; | |||
6315 | if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)), | |||
6316 | m_OneUse(m_Value(Cond2))))) | |||
6317 | Opc = Instruction::And; | |||
6318 | else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)), | |||
6319 | m_OneUse(m_Value(Cond2))))) | |||
6320 | Opc = Instruction::Or; | |||
6321 | else | |||
6322 | continue; | |||
6323 | ||||
6324 | if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) || | |||
6325 | !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) ) | |||
6326 | continue; | |||
6327 | ||||
6328 | DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "Before branch condition splitting\n" ; BB.dump(); } } while (false); | |||
6329 | ||||
6330 | // Create a new BB. | |||
6331 | auto TmpBB = | |||
6332 | BasicBlock::Create(BB.getContext(), BB.getName() + ".cond.split", | |||
6333 | BB.getParent(), BB.getNextNode()); | |||
6334 | ||||
6335 | // Update original basic block by using the first condition directly by the | |||
6336 | // branch instruction and removing the no longer needed and/or instruction. | |||
6337 | Br1->setCondition(Cond1); | |||
6338 | LogicOp->eraseFromParent(); | |||
6339 | ||||
6340 | // Depending on the conditon we have to either replace the true or the false | |||
6341 | // successor of the original branch instruction. | |||
6342 | if (Opc == Instruction::And) | |||
6343 | Br1->setSuccessor(0, TmpBB); | |||
6344 | else | |||
6345 | Br1->setSuccessor(1, TmpBB); | |||
6346 | ||||
6347 | // Fill in the new basic block. | |||
6348 | auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB); | |||
6349 | if (auto *I = dyn_cast<Instruction>(Cond2)) { | |||
6350 | I->removeFromParent(); | |||
6351 | I->insertBefore(Br2); | |||
6352 | } | |||
6353 | ||||
6354 | // Update PHI nodes in both successors. The original BB needs to be | |||
6355 | // replaced in one successor's PHI nodes, because the branch comes now from | |||
6356 | // the newly generated BB (NewBB). In the other successor we need to add one | |||
6357 | // incoming edge to the PHI nodes, because both branch instructions target | |||
6358 | // now the same successor. Depending on the original branch condition | |||
6359 | // (and/or) we have to swap the successors (TrueDest, FalseDest), so that | |||
6360 | // we perform the correct update for the PHI nodes. | |||
6361 | // This doesn't change the successor order of the just created branch | |||
6362 | // instruction (or any other instruction). | |||
6363 | if (Opc == Instruction::Or) | |||
6364 | std::swap(TBB, FBB); | |||
6365 | ||||
6366 | // Replace the old BB with the new BB. | |||
6367 | for (auto &I : *TBB) { | |||
6368 | PHINode *PN = dyn_cast<PHINode>(&I); | |||
6369 | if (!PN) | |||
6370 | break; | |||
6371 | int i; | |||
6372 | while ((i = PN->getBasicBlockIndex(&BB)) >= 0) | |||
6373 | PN->setIncomingBlock(i, TmpBB); | |||
6374 | } | |||
6375 | ||||
6376 | // Add another incoming edge form the new BB. | |||
6377 | for (auto &I : *FBB) { | |||
6378 | PHINode *PN = dyn_cast<PHINode>(&I); | |||
6379 | if (!PN) | |||
6380 | break; | |||
6381 | auto *Val = PN->getIncomingValueForBlock(&BB); | |||
6382 | PN->addIncoming(Val, TmpBB); | |||
6383 | } | |||
6384 | ||||
6385 | // Update the branch weights (from SelectionDAGBuilder:: | |||
6386 | // FindMergedConditions). | |||
6387 | if (Opc == Instruction::Or) { | |||
6388 | // Codegen X | Y as: | |||
6389 | // BB1: | |||
6390 | // jmp_if_X TBB | |||
6391 | // jmp TmpBB | |||
6392 | // TmpBB: | |||
6393 | // jmp_if_Y TBB | |||
6394 | // jmp FBB | |||
6395 | // | |||
6396 | ||||
6397 | // We have flexibility in setting Prob for BB1 and Prob for NewBB. | |||
6398 | // The requirement is that | |||
6399 | // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB) | |||
6400 | // = TrueProb for orignal BB. | |||
6401 | // Assuming the orignal weights are A and B, one choice is to set BB1's | |||
6402 | // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice | |||
6403 | // assumes that | |||
6404 | // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB. | |||
6405 | // Another choice is to assume TrueProb for BB1 equals to TrueProb for | |||
6406 | // TmpBB, but the math is more complicated. | |||
6407 | uint64_t TrueWeight, FalseWeight; | |||
6408 | if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) { | |||
6409 | uint64_t NewTrueWeight = TrueWeight; | |||
6410 | uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight; | |||
6411 | scaleWeights(NewTrueWeight, NewFalseWeight); | |||
6412 | Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext()) | |||
6413 | .createBranchWeights(TrueWeight, FalseWeight)); | |||
6414 | ||||
6415 | NewTrueWeight = TrueWeight; | |||
6416 | NewFalseWeight = 2 * FalseWeight; | |||
6417 | scaleWeights(NewTrueWeight, NewFalseWeight); | |||
6418 | Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext()) | |||
6419 | .createBranchWeights(TrueWeight, FalseWeight)); | |||
6420 | } | |||
6421 | } else { | |||
6422 | // Codegen X & Y as: | |||
6423 | // BB1: | |||
6424 | // jmp_if_X TmpBB | |||
6425 | // jmp FBB | |||
6426 | // TmpBB: | |||
6427 | // jmp_if_Y TBB | |||
6428 | // jmp FBB | |||
6429 | // | |||
6430 | // This requires creation of TmpBB after CurBB. | |||
6431 | ||||
6432 | // We have flexibility in setting Prob for BB1 and Prob for TmpBB. | |||
6433 | // The requirement is that | |||
6434 | // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB) | |||
6435 | // = FalseProb for orignal BB. | |||
6436 | // Assuming the orignal weights are A and B, one choice is to set BB1's | |||
6437 | // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice | |||
6438 | // assumes that | |||
6439 | // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB. | |||
6440 | uint64_t TrueWeight, FalseWeight; | |||
6441 | if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) { | |||
6442 | uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight; | |||
6443 | uint64_t NewFalseWeight = FalseWeight; | |||
6444 | scaleWeights(NewTrueWeight, NewFalseWeight); | |||
6445 | Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext()) | |||
6446 | .createBranchWeights(TrueWeight, FalseWeight)); | |||
6447 | ||||
6448 | NewTrueWeight = 2 * TrueWeight; | |||
6449 | NewFalseWeight = FalseWeight; | |||
6450 | scaleWeights(NewTrueWeight, NewFalseWeight); | |||
6451 | Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext()) | |||
6452 | .createBranchWeights(TrueWeight, FalseWeight)); | |||
6453 | } | |||
6454 | } | |||
6455 | ||||
6456 | // Note: No point in getting fancy here, since the DT info is never | |||
6457 | // available to CodeGenPrepare. | |||
6458 | ModifiedDT = true; | |||
6459 | ||||
6460 | MadeChange = true; | |||
6461 | ||||
6462 | DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "After branch condition splitting\n" ; BB.dump(); TmpBB->dump(); } } while (false) | |||
6463 | TmpBB->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("codegenprepare")) { dbgs() << "After branch condition splitting\n" ; BB.dump(); TmpBB->dump(); } } while (false); | |||
6464 | } | |||
6465 | return MadeChange; | |||
6466 | } |