File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Transforms/Scalar/LICM.cpp |
Warning: | line 1229, column 33 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===// | ||||||||
2 | // | ||||||||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | ||||||||
4 | // See https://llvm.org/LICENSE.txt for license information. | ||||||||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||||||||
6 | // | ||||||||
7 | //===----------------------------------------------------------------------===// | ||||||||
8 | // | ||||||||
9 | // This pass performs loop invariant code motion, attempting to remove as much | ||||||||
10 | // code from the body of a loop as possible. It does this by either hoisting | ||||||||
11 | // code into the preheader block, or by sinking code to the exit blocks if it is | ||||||||
12 | // safe. This pass also promotes must-aliased memory locations in the loop to | ||||||||
13 | // live in registers, thus hoisting and sinking "invariant" loads and stores. | ||||||||
14 | // | ||||||||
15 | // Hoisting operations out of loops is a canonicalization transform. It | ||||||||
16 | // enables and simplifies subsequent optimizations in the middle-end. | ||||||||
17 | // Rematerialization of hoisted instructions to reduce register pressure is the | ||||||||
18 | // responsibility of the back-end, which has more accurate information about | ||||||||
19 | // register pressure and also handles other optimizations than LICM that | ||||||||
20 | // increase live-ranges. | ||||||||
21 | // | ||||||||
22 | // This pass uses alias analysis for two purposes: | ||||||||
23 | // | ||||||||
24 | // 1. Moving loop invariant loads and calls out of loops. If we can determine | ||||||||
25 | // that a load or call inside of a loop never aliases anything stored to, | ||||||||
26 | // we can hoist it or sink it like any other instruction. | ||||||||
27 | // 2. Scalar Promotion of Memory - If there is a store instruction inside of | ||||||||
28 | // the loop, we try to move the store to happen AFTER the loop instead of | ||||||||
29 | // inside of the loop. This can only happen if a few conditions are true: | ||||||||
30 | // A. The pointer stored through is loop invariant | ||||||||
31 | // B. There are no stores or loads in the loop which _may_ alias the | ||||||||
32 | // pointer. There are no calls in the loop which mod/ref the pointer. | ||||||||
33 | // If these conditions are true, we can promote the loads and stores in the | ||||||||
34 | // loop of the pointer to use a temporary alloca'd variable. We then use | ||||||||
35 | // the SSAUpdater to construct the appropriate SSA form for the value. | ||||||||
36 | // | ||||||||
37 | //===----------------------------------------------------------------------===// | ||||||||
38 | |||||||||
39 | #include "llvm/Transforms/Scalar/LICM.h" | ||||||||
40 | #include "llvm/ADT/SetOperations.h" | ||||||||
41 | #include "llvm/ADT/Statistic.h" | ||||||||
42 | #include "llvm/Analysis/AliasAnalysis.h" | ||||||||
43 | #include "llvm/Analysis/AliasSetTracker.h" | ||||||||
44 | #include "llvm/Analysis/BasicAliasAnalysis.h" | ||||||||
45 | #include "llvm/Analysis/BlockFrequencyInfo.h" | ||||||||
46 | #include "llvm/Analysis/CaptureTracking.h" | ||||||||
47 | #include "llvm/Analysis/ConstantFolding.h" | ||||||||
48 | #include "llvm/Analysis/GlobalsModRef.h" | ||||||||
49 | #include "llvm/Analysis/GuardUtils.h" | ||||||||
50 | #include "llvm/Analysis/LazyBlockFrequencyInfo.h" | ||||||||
51 | #include "llvm/Analysis/Loads.h" | ||||||||
52 | #include "llvm/Analysis/LoopInfo.h" | ||||||||
53 | #include "llvm/Analysis/LoopIterator.h" | ||||||||
54 | #include "llvm/Analysis/LoopPass.h" | ||||||||
55 | #include "llvm/Analysis/MemoryBuiltins.h" | ||||||||
56 | #include "llvm/Analysis/MemorySSA.h" | ||||||||
57 | #include "llvm/Analysis/MemorySSAUpdater.h" | ||||||||
58 | #include "llvm/Analysis/MustExecute.h" | ||||||||
59 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | ||||||||
60 | #include "llvm/Analysis/ScalarEvolution.h" | ||||||||
61 | #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h" | ||||||||
62 | #include "llvm/Analysis/TargetLibraryInfo.h" | ||||||||
63 | #include "llvm/Analysis/ValueTracking.h" | ||||||||
64 | #include "llvm/IR/CFG.h" | ||||||||
65 | #include "llvm/IR/Constants.h" | ||||||||
66 | #include "llvm/IR/DataLayout.h" | ||||||||
67 | #include "llvm/IR/DebugInfoMetadata.h" | ||||||||
68 | #include "llvm/IR/DerivedTypes.h" | ||||||||
69 | #include "llvm/IR/Dominators.h" | ||||||||
70 | #include "llvm/IR/Instructions.h" | ||||||||
71 | #include "llvm/IR/IntrinsicInst.h" | ||||||||
72 | #include "llvm/IR/LLVMContext.h" | ||||||||
73 | #include "llvm/IR/Metadata.h" | ||||||||
74 | #include "llvm/IR/PatternMatch.h" | ||||||||
75 | #include "llvm/IR/PredIteratorCache.h" | ||||||||
76 | #include "llvm/InitializePasses.h" | ||||||||
77 | #include "llvm/Support/CommandLine.h" | ||||||||
78 | #include "llvm/Support/Debug.h" | ||||||||
79 | #include "llvm/Support/raw_ostream.h" | ||||||||
80 | #include "llvm/Transforms/Scalar.h" | ||||||||
81 | #include "llvm/Transforms/Scalar/LoopPassManager.h" | ||||||||
82 | #include "llvm/Transforms/Utils/AssumeBundleBuilder.h" | ||||||||
83 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | ||||||||
84 | #include "llvm/Transforms/Utils/Local.h" | ||||||||
85 | #include "llvm/Transforms/Utils/LoopUtils.h" | ||||||||
86 | #include "llvm/Transforms/Utils/SSAUpdater.h" | ||||||||
87 | #include <algorithm> | ||||||||
88 | #include <utility> | ||||||||
89 | using namespace llvm; | ||||||||
90 | |||||||||
91 | #define DEBUG_TYPE"licm" "licm" | ||||||||
92 | |||||||||
93 | STATISTIC(NumCreatedBlocks, "Number of blocks created")static llvm::Statistic NumCreatedBlocks = {"licm", "NumCreatedBlocks" , "Number of blocks created"}; | ||||||||
94 | STATISTIC(NumClonedBranches, "Number of branches cloned")static llvm::Statistic NumClonedBranches = {"licm", "NumClonedBranches" , "Number of branches cloned"}; | ||||||||
95 | STATISTIC(NumSunk, "Number of instructions sunk out of loop")static llvm::Statistic NumSunk = {"licm", "NumSunk", "Number of instructions sunk out of loop" }; | ||||||||
96 | STATISTIC(NumHoisted, "Number of instructions hoisted out of loop")static llvm::Statistic NumHoisted = {"licm", "NumHoisted", "Number of instructions hoisted out of loop" }; | ||||||||
97 | STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk")static llvm::Statistic NumMovedLoads = {"licm", "NumMovedLoads" , "Number of load insts hoisted or sunk"}; | ||||||||
98 | STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk")static llvm::Statistic NumMovedCalls = {"licm", "NumMovedCalls" , "Number of call insts hoisted or sunk"}; | ||||||||
99 | STATISTIC(NumPromoted, "Number of memory locations promoted to registers")static llvm::Statistic NumPromoted = {"licm", "NumPromoted", "Number of memory locations promoted to registers" }; | ||||||||
100 | |||||||||
101 | /// Memory promotion is enabled by default. | ||||||||
102 | static cl::opt<bool> | ||||||||
103 | DisablePromotion("disable-licm-promotion", cl::Hidden, cl::init(false), | ||||||||
104 | cl::desc("Disable memory promotion in LICM pass")); | ||||||||
105 | |||||||||
106 | static cl::opt<bool> ControlFlowHoisting( | ||||||||
107 | "licm-control-flow-hoisting", cl::Hidden, cl::init(false), | ||||||||
108 | cl::desc("Enable control flow (and PHI) hoisting in LICM")); | ||||||||
109 | |||||||||
110 | static cl::opt<unsigned> HoistSinkColdnessThreshold( | ||||||||
111 | "licm-coldness-threshold", cl::Hidden, cl::init(4), | ||||||||
112 | cl::desc("Relative coldness Threshold of hoisting/sinking destination " | ||||||||
113 | "block for LICM to be considered beneficial")); | ||||||||
114 | |||||||||
115 | static cl::opt<uint32_t> MaxNumUsesTraversed( | ||||||||
116 | "licm-max-num-uses-traversed", cl::Hidden, cl::init(8), | ||||||||
117 | cl::desc("Max num uses visited for identifying load " | ||||||||
118 | "invariance in loop using invariant start (default = 8)")); | ||||||||
119 | |||||||||
120 | // Default value of zero implies we use the regular alias set tracker mechanism | ||||||||
121 | // instead of the cross product using AA to identify aliasing of the memory | ||||||||
122 | // location we are interested in. | ||||||||
123 | static cl::opt<int> | ||||||||
124 | LICMN2Theshold("licm-n2-threshold", cl::Hidden, cl::init(0), | ||||||||
125 | cl::desc("How many instruction to cross product using AA")); | ||||||||
126 | |||||||||
127 | // Experimental option to allow imprecision in LICM in pathological cases, in | ||||||||
128 | // exchange for faster compile. This is to be removed if MemorySSA starts to | ||||||||
129 | // address the same issue. This flag applies only when LICM uses MemorySSA | ||||||||
130 | // instead on AliasSetTracker. LICM calls MemorySSAWalker's | ||||||||
131 | // getClobberingMemoryAccess, up to the value of the Cap, getting perfect | ||||||||
132 | // accuracy. Afterwards, LICM will call into MemorySSA's getDefiningAccess, | ||||||||
133 | // which may not be precise, since optimizeUses is capped. The result is | ||||||||
134 | // correct, but we may not get as "far up" as possible to get which access is | ||||||||
135 | // clobbering the one queried. | ||||||||
136 | cl::opt<unsigned> llvm::SetLicmMssaOptCap( | ||||||||
137 | "licm-mssa-optimization-cap", cl::init(100), cl::Hidden, | ||||||||
138 | cl::desc("Enable imprecision in LICM in pathological cases, in exchange " | ||||||||
139 | "for faster compile. Caps the MemorySSA clobbering calls.")); | ||||||||
140 | |||||||||
141 | // Experimentally, memory promotion carries less importance than sinking and | ||||||||
142 | // hoisting. Limit when we do promotion when using MemorySSA, in order to save | ||||||||
143 | // compile time. | ||||||||
144 | cl::opt<unsigned> llvm::SetLicmMssaNoAccForPromotionCap( | ||||||||
145 | "licm-mssa-max-acc-promotion", cl::init(250), cl::Hidden, | ||||||||
146 | cl::desc("[LICM & MemorySSA] When MSSA in LICM is disabled, this has no " | ||||||||
147 | "effect. When MSSA in LICM is enabled, then this is the maximum " | ||||||||
148 | "number of accesses allowed to be present in a loop in order to " | ||||||||
149 | "enable memory promotion.")); | ||||||||
150 | |||||||||
151 | static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI); | ||||||||
152 | static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop, | ||||||||
153 | const LoopSafetyInfo *SafetyInfo, | ||||||||
154 | TargetTransformInfo *TTI, bool &FreeInLoop); | ||||||||
155 | static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop, | ||||||||
156 | BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo, | ||||||||
157 | MemorySSAUpdater *MSSAU, ScalarEvolution *SE, | ||||||||
158 | OptimizationRemarkEmitter *ORE); | ||||||||
159 | static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT, | ||||||||
160 | BlockFrequencyInfo *BFI, const Loop *CurLoop, | ||||||||
161 | ICFLoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU, | ||||||||
162 | OptimizationRemarkEmitter *ORE); | ||||||||
163 | static bool isSafeToExecuteUnconditionally(Instruction &Inst, | ||||||||
164 | const DominatorTree *DT, | ||||||||
165 | const TargetLibraryInfo *TLI, | ||||||||
166 | const Loop *CurLoop, | ||||||||
167 | const LoopSafetyInfo *SafetyInfo, | ||||||||
168 | OptimizationRemarkEmitter *ORE, | ||||||||
169 | const Instruction *CtxI = nullptr); | ||||||||
170 | static bool pointerInvalidatedByLoop(MemoryLocation MemLoc, | ||||||||
171 | AliasSetTracker *CurAST, Loop *CurLoop, | ||||||||
172 | AAResults *AA); | ||||||||
173 | static bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU, | ||||||||
174 | Loop *CurLoop, Instruction &I, | ||||||||
175 | SinkAndHoistLICMFlags &Flags); | ||||||||
176 | static bool pointerInvalidatedByBlockWithMSSA(BasicBlock &BB, MemorySSA &MSSA, | ||||||||
177 | MemoryUse &MU); | ||||||||
178 | static Instruction *cloneInstructionInExitBlock( | ||||||||
179 | Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI, | ||||||||
180 | const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU); | ||||||||
181 | |||||||||
182 | static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo, | ||||||||
183 | AliasSetTracker *AST, MemorySSAUpdater *MSSAU); | ||||||||
184 | |||||||||
185 | static void moveInstructionBefore(Instruction &I, Instruction &Dest, | ||||||||
186 | ICFLoopSafetyInfo &SafetyInfo, | ||||||||
187 | MemorySSAUpdater *MSSAU, ScalarEvolution *SE); | ||||||||
188 | |||||||||
189 | static void foreachMemoryAccess(MemorySSA *MSSA, Loop *L, | ||||||||
190 | function_ref<void(Instruction *)> Fn); | ||||||||
191 | static SmallVector<SmallSetVector<Value *, 8>, 0> | ||||||||
192 | collectPromotionCandidates(MemorySSA *MSSA, AliasAnalysis *AA, Loop *L); | ||||||||
193 | |||||||||
194 | namespace { | ||||||||
195 | struct LoopInvariantCodeMotion { | ||||||||
196 | bool runOnLoop(Loop *L, AAResults *AA, LoopInfo *LI, DominatorTree *DT, | ||||||||
197 | BlockFrequencyInfo *BFI, TargetLibraryInfo *TLI, | ||||||||
198 | TargetTransformInfo *TTI, ScalarEvolution *SE, MemorySSA *MSSA, | ||||||||
199 | OptimizationRemarkEmitter *ORE, bool LoopNestMode = false); | ||||||||
200 | |||||||||
201 | LoopInvariantCodeMotion(unsigned LicmMssaOptCap, | ||||||||
202 | unsigned LicmMssaNoAccForPromotionCap) | ||||||||
203 | : LicmMssaOptCap(LicmMssaOptCap), | ||||||||
204 | LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap) {} | ||||||||
205 | |||||||||
206 | private: | ||||||||
207 | unsigned LicmMssaOptCap; | ||||||||
208 | unsigned LicmMssaNoAccForPromotionCap; | ||||||||
209 | |||||||||
210 | std::unique_ptr<AliasSetTracker> | ||||||||
211 | collectAliasInfoForLoop(Loop *L, LoopInfo *LI, AAResults *AA); | ||||||||
212 | }; | ||||||||
213 | |||||||||
214 | struct LegacyLICMPass : public LoopPass { | ||||||||
215 | static char ID; // Pass identification, replacement for typeid | ||||||||
216 | LegacyLICMPass( | ||||||||
217 | unsigned LicmMssaOptCap = SetLicmMssaOptCap, | ||||||||
218 | unsigned LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap) | ||||||||
219 | : LoopPass(ID), LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap) { | ||||||||
220 | initializeLegacyLICMPassPass(*PassRegistry::getPassRegistry()); | ||||||||
221 | } | ||||||||
222 | |||||||||
223 | bool runOnLoop(Loop *L, LPPassManager &LPM) override { | ||||||||
224 | if (skipLoop(L)) | ||||||||
225 | return false; | ||||||||
226 | |||||||||
227 | LLVM_DEBUG(dbgs() << "Perform LICM on Loop with header at block "do { } while (false) | ||||||||
228 | << L->getHeader()->getNameOrAsOperand() << "\n")do { } while (false); | ||||||||
229 | |||||||||
230 | auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>(); | ||||||||
231 | MemorySSA *MSSA = EnableMSSALoopDependency | ||||||||
232 | ? (&getAnalysis<MemorySSAWrapperPass>().getMSSA()) | ||||||||
233 | : nullptr; | ||||||||
234 | bool hasProfileData = L->getHeader()->getParent()->hasProfileData(); | ||||||||
235 | BlockFrequencyInfo *BFI = | ||||||||
236 | hasProfileData ? &getAnalysis<LazyBlockFrequencyInfoPass>().getBFI() | ||||||||
237 | : nullptr; | ||||||||
238 | // For the old PM, we can't use OptimizationRemarkEmitter as an analysis | ||||||||
239 | // pass. Function analyses need to be preserved across loop transformations | ||||||||
240 | // but ORE cannot be preserved (see comment before the pass definition). | ||||||||
241 | OptimizationRemarkEmitter ORE(L->getHeader()->getParent()); | ||||||||
242 | return LICM.runOnLoop( | ||||||||
243 | L, &getAnalysis<AAResultsWrapperPass>().getAAResults(), | ||||||||
244 | &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(), | ||||||||
245 | &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), BFI, | ||||||||
246 | &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI( | ||||||||
247 | *L->getHeader()->getParent()), | ||||||||
248 | &getAnalysis<TargetTransformInfoWrapperPass>().getTTI( | ||||||||
249 | *L->getHeader()->getParent()), | ||||||||
250 | SE ? &SE->getSE() : nullptr, MSSA, &ORE); | ||||||||
251 | } | ||||||||
252 | |||||||||
253 | /// This transformation requires natural loop information & requires that | ||||||||
254 | /// loop preheaders be inserted into the CFG... | ||||||||
255 | /// | ||||||||
256 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||||||
257 | AU.addPreserved<DominatorTreeWrapperPass>(); | ||||||||
258 | AU.addPreserved<LoopInfoWrapperPass>(); | ||||||||
259 | AU.addRequired<TargetLibraryInfoWrapperPass>(); | ||||||||
260 | if (EnableMSSALoopDependency) { | ||||||||
261 | AU.addRequired<MemorySSAWrapperPass>(); | ||||||||
262 | AU.addPreserved<MemorySSAWrapperPass>(); | ||||||||
263 | } | ||||||||
264 | AU.addRequired<TargetTransformInfoWrapperPass>(); | ||||||||
265 | getLoopAnalysisUsage(AU); | ||||||||
266 | LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU); | ||||||||
267 | AU.addPreserved<LazyBlockFrequencyInfoPass>(); | ||||||||
268 | AU.addPreserved<LazyBranchProbabilityInfoPass>(); | ||||||||
269 | } | ||||||||
270 | |||||||||
271 | private: | ||||||||
272 | LoopInvariantCodeMotion LICM; | ||||||||
273 | }; | ||||||||
274 | } // namespace | ||||||||
275 | |||||||||
276 | PreservedAnalyses LICMPass::run(Loop &L, LoopAnalysisManager &AM, | ||||||||
277 | LoopStandardAnalysisResults &AR, LPMUpdater &) { | ||||||||
278 | // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis | ||||||||
279 | // pass. Function analyses need to be preserved across loop transformations | ||||||||
280 | // but ORE cannot be preserved (see comment before the pass definition). | ||||||||
281 | OptimizationRemarkEmitter ORE(L.getHeader()->getParent()); | ||||||||
282 | |||||||||
283 | LoopInvariantCodeMotion LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap); | ||||||||
284 | if (!LICM.runOnLoop(&L, &AR.AA, &AR.LI, &AR.DT, AR.BFI, &AR.TLI, &AR.TTI, | ||||||||
285 | &AR.SE, AR.MSSA, &ORE)) | ||||||||
286 | return PreservedAnalyses::all(); | ||||||||
287 | |||||||||
288 | auto PA = getLoopPassPreservedAnalyses(); | ||||||||
289 | |||||||||
290 | PA.preserve<DominatorTreeAnalysis>(); | ||||||||
291 | PA.preserve<LoopAnalysis>(); | ||||||||
292 | if (AR.MSSA) | ||||||||
293 | PA.preserve<MemorySSAAnalysis>(); | ||||||||
294 | |||||||||
295 | return PA; | ||||||||
296 | } | ||||||||
297 | |||||||||
298 | PreservedAnalyses LNICMPass::run(LoopNest &LN, LoopAnalysisManager &AM, | ||||||||
299 | LoopStandardAnalysisResults &AR, | ||||||||
300 | LPMUpdater &) { | ||||||||
301 | // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis | ||||||||
302 | // pass. Function analyses need to be preserved across loop transformations | ||||||||
303 | // but ORE cannot be preserved (see comment before the pass definition). | ||||||||
304 | OptimizationRemarkEmitter ORE(LN.getParent()); | ||||||||
305 | |||||||||
306 | LoopInvariantCodeMotion LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap); | ||||||||
307 | |||||||||
308 | Loop &OutermostLoop = LN.getOutermostLoop(); | ||||||||
309 | bool Changed = LICM.runOnLoop(&OutermostLoop, &AR.AA, &AR.LI, &AR.DT, AR.BFI, | ||||||||
310 | &AR.TLI, &AR.TTI, &AR.SE, AR.MSSA, &ORE, true); | ||||||||
311 | |||||||||
312 | if (!Changed) | ||||||||
313 | return PreservedAnalyses::all(); | ||||||||
314 | |||||||||
315 | auto PA = getLoopPassPreservedAnalyses(); | ||||||||
316 | |||||||||
317 | PA.preserve<DominatorTreeAnalysis>(); | ||||||||
318 | PA.preserve<LoopAnalysis>(); | ||||||||
319 | if (AR.MSSA) | ||||||||
320 | PA.preserve<MemorySSAAnalysis>(); | ||||||||
321 | |||||||||
322 | return PA; | ||||||||
323 | } | ||||||||
324 | |||||||||
325 | char LegacyLICMPass::ID = 0; | ||||||||
326 | INITIALIZE_PASS_BEGIN(LegacyLICMPass, "licm", "Loop Invariant Code Motion",static void *initializeLegacyLICMPassPassOnce(PassRegistry & Registry) { | ||||||||
327 | false, false)static void *initializeLegacyLICMPassPassOnce(PassRegistry & Registry) { | ||||||||
328 | INITIALIZE_PASS_DEPENDENCY(LoopPass)initializeLoopPassPass(Registry); | ||||||||
329 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); | ||||||||
330 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | ||||||||
331 | INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)initializeMemorySSAWrapperPassPass(Registry); | ||||||||
332 | INITIALIZE_PASS_DEPENDENCY(LazyBFIPass)initializeLazyBFIPassPass(Registry); | ||||||||
333 | INITIALIZE_PASS_END(LegacyLICMPass, "licm", "Loop Invariant Code Motion", false,PassInfo *PI = new PassInfo( "Loop Invariant Code Motion", "licm" , &LegacyLICMPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <LegacyLICMPass>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeLegacyLICMPassPassFlag ; void llvm::initializeLegacyLICMPassPass(PassRegistry &Registry ) { llvm::call_once(InitializeLegacyLICMPassPassFlag, initializeLegacyLICMPassPassOnce , std::ref(Registry)); } | ||||||||
334 | false)PassInfo *PI = new PassInfo( "Loop Invariant Code Motion", "licm" , &LegacyLICMPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <LegacyLICMPass>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeLegacyLICMPassPassFlag ; void llvm::initializeLegacyLICMPassPass(PassRegistry &Registry ) { llvm::call_once(InitializeLegacyLICMPassPassFlag, initializeLegacyLICMPassPassOnce , std::ref(Registry)); } | ||||||||
335 | |||||||||
336 | Pass *llvm::createLICMPass() { return new LegacyLICMPass(); } | ||||||||
337 | Pass *llvm::createLICMPass(unsigned LicmMssaOptCap, | ||||||||
338 | unsigned LicmMssaNoAccForPromotionCap) { | ||||||||
339 | return new LegacyLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap); | ||||||||
340 | } | ||||||||
341 | |||||||||
342 | llvm::SinkAndHoistLICMFlags::SinkAndHoistLICMFlags(bool IsSink, Loop *L, | ||||||||
343 | MemorySSA *MSSA) | ||||||||
344 | : SinkAndHoistLICMFlags(SetLicmMssaOptCap, SetLicmMssaNoAccForPromotionCap, | ||||||||
345 | IsSink, L, MSSA) {} | ||||||||
346 | |||||||||
347 | llvm::SinkAndHoistLICMFlags::SinkAndHoistLICMFlags( | ||||||||
348 | unsigned LicmMssaOptCap, unsigned LicmMssaNoAccForPromotionCap, bool IsSink, | ||||||||
349 | Loop *L, MemorySSA *MSSA) | ||||||||
350 | : LicmMssaOptCap(LicmMssaOptCap), | ||||||||
351 | LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap), | ||||||||
352 | IsSink(IsSink) { | ||||||||
353 | assert(((L != nullptr) == (MSSA != nullptr)) &&((void)0) | ||||||||
354 | "Unexpected values for SinkAndHoistLICMFlags")((void)0); | ||||||||
355 | if (!MSSA) | ||||||||
356 | return; | ||||||||
357 | |||||||||
358 | unsigned AccessCapCount = 0; | ||||||||
359 | for (auto *BB : L->getBlocks()) | ||||||||
360 | if (const auto *Accesses = MSSA->getBlockAccesses(BB)) | ||||||||
361 | for (const auto &MA : *Accesses) { | ||||||||
362 | (void)MA; | ||||||||
363 | ++AccessCapCount; | ||||||||
364 | if (AccessCapCount > LicmMssaNoAccForPromotionCap) { | ||||||||
365 | NoOfMemAccTooLarge = true; | ||||||||
366 | return; | ||||||||
367 | } | ||||||||
368 | } | ||||||||
369 | } | ||||||||
370 | |||||||||
371 | /// Hoist expressions out of the specified loop. Note, alias info for inner | ||||||||
372 | /// loop is not preserved so it is not a good idea to run LICM multiple | ||||||||
373 | /// times on one loop. | ||||||||
374 | bool LoopInvariantCodeMotion::runOnLoop( | ||||||||
375 | Loop *L, AAResults *AA, LoopInfo *LI, DominatorTree *DT, | ||||||||
376 | BlockFrequencyInfo *BFI, TargetLibraryInfo *TLI, TargetTransformInfo *TTI, | ||||||||
377 | ScalarEvolution *SE, MemorySSA *MSSA, OptimizationRemarkEmitter *ORE, | ||||||||
378 | bool LoopNestMode) { | ||||||||
379 | bool Changed = false; | ||||||||
380 | |||||||||
381 | assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.")((void)0); | ||||||||
382 | |||||||||
383 | // If this loop has metadata indicating that LICM is not to be performed then | ||||||||
384 | // just exit. | ||||||||
385 | if (hasDisableLICMTransformsHint(L)) { | ||||||||
386 | return false; | ||||||||
387 | } | ||||||||
388 | |||||||||
389 | std::unique_ptr<AliasSetTracker> CurAST; | ||||||||
390 | std::unique_ptr<MemorySSAUpdater> MSSAU; | ||||||||
391 | std::unique_ptr<SinkAndHoistLICMFlags> Flags; | ||||||||
392 | |||||||||
393 | // Don't sink stores from loops with coroutine suspend instructions. | ||||||||
394 | // LICM would sink instructions into the default destination of | ||||||||
395 | // the coroutine switch. The default destination of the switch is to | ||||||||
396 | // handle the case where the coroutine is suspended, by which point the | ||||||||
397 | // coroutine frame may have been destroyed. No instruction can be sunk there. | ||||||||
398 | // FIXME: This would unfortunately hurt the performance of coroutines, however | ||||||||
399 | // there is currently no general solution for this. Similar issues could also | ||||||||
400 | // potentially happen in other passes where instructions are being moved | ||||||||
401 | // across that edge. | ||||||||
402 | bool HasCoroSuspendInst = llvm::any_of(L->getBlocks(), [](BasicBlock *BB) { | ||||||||
403 | return llvm::any_of(*BB, [](Instruction &I) { | ||||||||
404 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I); | ||||||||
405 | return II && II->getIntrinsicID() == Intrinsic::coro_suspend; | ||||||||
406 | }); | ||||||||
407 | }); | ||||||||
408 | |||||||||
409 | if (!MSSA) { | ||||||||
410 | LLVM_DEBUG(dbgs() << "LICM: Using Alias Set Tracker.\n")do { } while (false); | ||||||||
411 | CurAST = collectAliasInfoForLoop(L, LI, AA); | ||||||||
412 | Flags = std::make_unique<SinkAndHoistLICMFlags>( | ||||||||
413 | LicmMssaOptCap, LicmMssaNoAccForPromotionCap, /*IsSink=*/true); | ||||||||
414 | } else { | ||||||||
415 | LLVM_DEBUG(dbgs() << "LICM: Using MemorySSA.\n")do { } while (false); | ||||||||
416 | MSSAU = std::make_unique<MemorySSAUpdater>(MSSA); | ||||||||
417 | Flags = std::make_unique<SinkAndHoistLICMFlags>( | ||||||||
418 | LicmMssaOptCap, LicmMssaNoAccForPromotionCap, /*IsSink=*/true, L, MSSA); | ||||||||
419 | } | ||||||||
420 | |||||||||
421 | // Get the preheader block to move instructions into... | ||||||||
422 | BasicBlock *Preheader = L->getLoopPreheader(); | ||||||||
423 | |||||||||
424 | // Compute loop safety information. | ||||||||
425 | ICFLoopSafetyInfo SafetyInfo; | ||||||||
426 | SafetyInfo.computeLoopSafetyInfo(L); | ||||||||
427 | |||||||||
428 | // We want to visit all of the instructions in this loop... that are not parts | ||||||||
429 | // of our subloops (they have already had their invariants hoisted out of | ||||||||
430 | // their loop, into this loop, so there is no need to process the BODIES of | ||||||||
431 | // the subloops). | ||||||||
432 | // | ||||||||
433 | // Traverse the body of the loop in depth first order on the dominator tree so | ||||||||
434 | // that we are guaranteed to see definitions before we see uses. This allows | ||||||||
435 | // us to sink instructions in one pass, without iteration. After sinking | ||||||||
436 | // instructions, we perform another pass to hoist them out of the loop. | ||||||||
437 | if (L->hasDedicatedExits()) | ||||||||
438 | Changed |= | ||||||||
439 | sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, BFI, TLI, TTI, L, | ||||||||
440 | CurAST.get(), MSSAU.get(), &SafetyInfo, *Flags.get(), ORE); | ||||||||
441 | Flags->setIsSink(false); | ||||||||
442 | if (Preheader) | ||||||||
443 | Changed |= hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, BFI, TLI, L, | ||||||||
444 | CurAST.get(), MSSAU.get(), SE, &SafetyInfo, | ||||||||
445 | *Flags.get(), ORE, LoopNestMode); | ||||||||
446 | |||||||||
447 | // Now that all loop invariants have been removed from the loop, promote any | ||||||||
448 | // memory references to scalars that we can. | ||||||||
449 | // Don't sink stores from loops without dedicated block exits. Exits | ||||||||
450 | // containing indirect branches are not transformed by loop simplify, | ||||||||
451 | // make sure we catch that. An additional load may be generated in the | ||||||||
452 | // preheader for SSA updater, so also avoid sinking when no preheader | ||||||||
453 | // is available. | ||||||||
454 | if (!DisablePromotion && Preheader && L->hasDedicatedExits() && | ||||||||
455 | !Flags->tooManyMemoryAccesses() && !HasCoroSuspendInst) { | ||||||||
456 | // Figure out the loop exits and their insertion points | ||||||||
457 | SmallVector<BasicBlock *, 8> ExitBlocks; | ||||||||
458 | L->getUniqueExitBlocks(ExitBlocks); | ||||||||
459 | |||||||||
460 | // We can't insert into a catchswitch. | ||||||||
461 | bool HasCatchSwitch = llvm::any_of(ExitBlocks, [](BasicBlock *Exit) { | ||||||||
462 | return isa<CatchSwitchInst>(Exit->getTerminator()); | ||||||||
463 | }); | ||||||||
464 | |||||||||
465 | if (!HasCatchSwitch) { | ||||||||
466 | SmallVector<Instruction *, 8> InsertPts; | ||||||||
467 | SmallVector<MemoryAccess *, 8> MSSAInsertPts; | ||||||||
468 | InsertPts.reserve(ExitBlocks.size()); | ||||||||
469 | if (MSSAU) | ||||||||
470 | MSSAInsertPts.reserve(ExitBlocks.size()); | ||||||||
471 | for (BasicBlock *ExitBlock : ExitBlocks) { | ||||||||
472 | InsertPts.push_back(&*ExitBlock->getFirstInsertionPt()); | ||||||||
473 | if (MSSAU) | ||||||||
474 | MSSAInsertPts.push_back(nullptr); | ||||||||
475 | } | ||||||||
476 | |||||||||
477 | PredIteratorCache PIC; | ||||||||
478 | |||||||||
479 | bool Promoted = false; | ||||||||
480 | if (CurAST.get()) { | ||||||||
481 | // Loop over all of the alias sets in the tracker object. | ||||||||
482 | for (AliasSet &AS : *CurAST) { | ||||||||
483 | // We can promote this alias set if it has a store, if it is a "Must" | ||||||||
484 | // alias set, if the pointer is loop invariant, and if we are not | ||||||||
485 | // eliminating any volatile loads or stores. | ||||||||
486 | if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || | ||||||||
487 | !L->isLoopInvariant(AS.begin()->getValue())) | ||||||||
488 | continue; | ||||||||
489 | |||||||||
490 | assert(((void)0) | ||||||||
491 | !AS.empty() &&((void)0) | ||||||||
492 | "Must alias set should have at least one pointer element in it!")((void)0); | ||||||||
493 | |||||||||
494 | SmallSetVector<Value *, 8> PointerMustAliases; | ||||||||
495 | for (const auto &ASI : AS) | ||||||||
496 | PointerMustAliases.insert(ASI.getValue()); | ||||||||
497 | |||||||||
498 | Promoted |= promoteLoopAccessesToScalars( | ||||||||
499 | PointerMustAliases, ExitBlocks, InsertPts, MSSAInsertPts, PIC, LI, | ||||||||
500 | DT, TLI, L, CurAST.get(), MSSAU.get(), &SafetyInfo, ORE); | ||||||||
501 | } | ||||||||
502 | } else { | ||||||||
503 | // Promoting one set of accesses may make the pointers for another set | ||||||||
504 | // loop invariant, so run this in a loop (with the MaybePromotable set | ||||||||
505 | // decreasing in size over time). | ||||||||
506 | bool LocalPromoted; | ||||||||
507 | do { | ||||||||
508 | LocalPromoted = false; | ||||||||
509 | for (const SmallSetVector<Value *, 8> &PointerMustAliases : | ||||||||
510 | collectPromotionCandidates(MSSA, AA, L)) { | ||||||||
511 | LocalPromoted |= promoteLoopAccessesToScalars( | ||||||||
512 | PointerMustAliases, ExitBlocks, InsertPts, MSSAInsertPts, PIC, | ||||||||
513 | LI, DT, TLI, L, /*AST*/nullptr, MSSAU.get(), &SafetyInfo, ORE); | ||||||||
514 | } | ||||||||
515 | Promoted |= LocalPromoted; | ||||||||
516 | } while (LocalPromoted); | ||||||||
517 | } | ||||||||
518 | |||||||||
519 | // Once we have promoted values across the loop body we have to | ||||||||
520 | // recursively reform LCSSA as any nested loop may now have values defined | ||||||||
521 | // within the loop used in the outer loop. | ||||||||
522 | // FIXME: This is really heavy handed. It would be a bit better to use an | ||||||||
523 | // SSAUpdater strategy during promotion that was LCSSA aware and reformed | ||||||||
524 | // it as it went. | ||||||||
525 | if (Promoted) | ||||||||
526 | formLCSSARecursively(*L, *DT, LI, SE); | ||||||||
527 | |||||||||
528 | Changed |= Promoted; | ||||||||
529 | } | ||||||||
530 | } | ||||||||
531 | |||||||||
532 | // Check that neither this loop nor its parent have had LCSSA broken. LICM is | ||||||||
533 | // specifically moving instructions across the loop boundary and so it is | ||||||||
534 | // especially in need of sanity checking here. | ||||||||
535 | assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!")((void)0); | ||||||||
536 | assert((L->isOutermost() || L->getParentLoop()->isLCSSAForm(*DT)) &&((void)0) | ||||||||
537 | "Parent loop not left in LCSSA form after LICM!")((void)0); | ||||||||
538 | |||||||||
539 | if (MSSAU.get() && VerifyMemorySSA) | ||||||||
540 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||
541 | |||||||||
542 | if (Changed && SE) | ||||||||
543 | SE->forgetLoopDispositions(L); | ||||||||
544 | return Changed; | ||||||||
545 | } | ||||||||
546 | |||||||||
547 | /// Walk the specified region of the CFG (defined by all blocks dominated by | ||||||||
548 | /// the specified block, and that are in the current loop) in reverse depth | ||||||||
549 | /// first order w.r.t the DominatorTree. This allows us to visit uses before | ||||||||
550 | /// definitions, allowing us to sink a loop body in one pass without iteration. | ||||||||
551 | /// | ||||||||
552 | bool llvm::sinkRegion(DomTreeNode *N, AAResults *AA, LoopInfo *LI, | ||||||||
553 | DominatorTree *DT, BlockFrequencyInfo *BFI, | ||||||||
554 | TargetLibraryInfo *TLI, TargetTransformInfo *TTI, | ||||||||
555 | Loop *CurLoop, AliasSetTracker *CurAST, | ||||||||
556 | MemorySSAUpdater *MSSAU, ICFLoopSafetyInfo *SafetyInfo, | ||||||||
557 | SinkAndHoistLICMFlags &Flags, | ||||||||
558 | OptimizationRemarkEmitter *ORE) { | ||||||||
559 | |||||||||
560 | // Verify inputs. | ||||||||
561 | assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&((void)0) | ||||||||
562 | CurLoop != nullptr && SafetyInfo != nullptr &&((void)0) | ||||||||
563 | "Unexpected input to sinkRegion.")((void)0); | ||||||||
564 | assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&((void)0) | ||||||||
565 | "Either AliasSetTracker or MemorySSA should be initialized.")((void)0); | ||||||||
566 | |||||||||
567 | // We want to visit children before parents. We will enque all the parents | ||||||||
568 | // before their children in the worklist and process the worklist in reverse | ||||||||
569 | // order. | ||||||||
570 | SmallVector<DomTreeNode *, 16> Worklist = collectChildrenInLoop(N, CurLoop); | ||||||||
571 | |||||||||
572 | bool Changed = false; | ||||||||
573 | for (DomTreeNode *DTN : reverse(Worklist)) { | ||||||||
574 | BasicBlock *BB = DTN->getBlock(); | ||||||||
575 | // Only need to process the contents of this block if it is not part of a | ||||||||
576 | // subloop (which would already have been processed). | ||||||||
577 | if (inSubLoop(BB, CurLoop, LI)) | ||||||||
578 | continue; | ||||||||
579 | |||||||||
580 | for (BasicBlock::iterator II = BB->end(); II != BB->begin();) { | ||||||||
581 | Instruction &I = *--II; | ||||||||
582 | |||||||||
583 | // The instruction is not used in the loop if it is dead. In this case, | ||||||||
584 | // we just delete it instead of sinking it. | ||||||||
585 | if (isInstructionTriviallyDead(&I, TLI)) { | ||||||||
586 | LLVM_DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n')do { } while (false); | ||||||||
587 | salvageKnowledge(&I); | ||||||||
588 | salvageDebugInfo(I); | ||||||||
589 | ++II; | ||||||||
590 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | ||||||||
591 | Changed = true; | ||||||||
592 | continue; | ||||||||
593 | } | ||||||||
594 | |||||||||
595 | // Check to see if we can sink this instruction to the exit blocks | ||||||||
596 | // of the loop. We can do this if the all users of the instruction are | ||||||||
597 | // outside of the loop. In this case, it doesn't even matter if the | ||||||||
598 | // operands of the instruction are loop invariant. | ||||||||
599 | // | ||||||||
600 | bool FreeInLoop = false; | ||||||||
601 | if (!I.mayHaveSideEffects() && | ||||||||
602 | isNotUsedOrFreeInLoop(I, CurLoop, SafetyInfo, TTI, FreeInLoop) && | ||||||||
603 | canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags, | ||||||||
604 | ORE)) { | ||||||||
605 | if (sink(I, LI, DT, BFI, CurLoop, SafetyInfo, MSSAU, ORE)) { | ||||||||
606 | if (!FreeInLoop) { | ||||||||
607 | ++II; | ||||||||
608 | salvageDebugInfo(I); | ||||||||
609 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | ||||||||
610 | } | ||||||||
611 | Changed = true; | ||||||||
612 | } | ||||||||
613 | } | ||||||||
614 | } | ||||||||
615 | } | ||||||||
616 | if (MSSAU && VerifyMemorySSA) | ||||||||
617 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||
618 | return Changed; | ||||||||
619 | } | ||||||||
620 | |||||||||
621 | namespace { | ||||||||
622 | // This is a helper class for hoistRegion to make it able to hoist control flow | ||||||||
623 | // in order to be able to hoist phis. The way this works is that we initially | ||||||||
624 | // start hoisting to the loop preheader, and when we see a loop invariant branch | ||||||||
625 | // we make note of this. When we then come to hoist an instruction that's | ||||||||
626 | // conditional on such a branch we duplicate the branch and the relevant control | ||||||||
627 | // flow, then hoist the instruction into the block corresponding to its original | ||||||||
628 | // block in the duplicated control flow. | ||||||||
629 | class ControlFlowHoister { | ||||||||
630 | private: | ||||||||
631 | // Information about the loop we are hoisting from | ||||||||
632 | LoopInfo *LI; | ||||||||
633 | DominatorTree *DT; | ||||||||
634 | Loop *CurLoop; | ||||||||
635 | MemorySSAUpdater *MSSAU; | ||||||||
636 | |||||||||
637 | // A map of blocks in the loop to the block their instructions will be hoisted | ||||||||
638 | // to. | ||||||||
639 | DenseMap<BasicBlock *, BasicBlock *> HoistDestinationMap; | ||||||||
640 | |||||||||
641 | // The branches that we can hoist, mapped to the block that marks a | ||||||||
642 | // convergence point of their control flow. | ||||||||
643 | DenseMap<BranchInst *, BasicBlock *> HoistableBranches; | ||||||||
644 | |||||||||
645 | public: | ||||||||
646 | ControlFlowHoister(LoopInfo *LI, DominatorTree *DT, Loop *CurLoop, | ||||||||
647 | MemorySSAUpdater *MSSAU) | ||||||||
648 | : LI(LI), DT(DT), CurLoop(CurLoop), MSSAU(MSSAU) {} | ||||||||
649 | |||||||||
650 | void registerPossiblyHoistableBranch(BranchInst *BI) { | ||||||||
651 | // We can only hoist conditional branches with loop invariant operands. | ||||||||
652 | if (!ControlFlowHoisting || !BI->isConditional() || | ||||||||
653 | !CurLoop->hasLoopInvariantOperands(BI)) | ||||||||
654 | return; | ||||||||
655 | |||||||||
656 | // The branch destinations need to be in the loop, and we don't gain | ||||||||
657 | // anything by duplicating conditional branches with duplicate successors, | ||||||||
658 | // as it's essentially the same as an unconditional branch. | ||||||||
659 | BasicBlock *TrueDest = BI->getSuccessor(0); | ||||||||
660 | BasicBlock *FalseDest = BI->getSuccessor(1); | ||||||||
661 | if (!CurLoop->contains(TrueDest) || !CurLoop->contains(FalseDest) || | ||||||||
662 | TrueDest == FalseDest) | ||||||||
663 | return; | ||||||||
664 | |||||||||
665 | // We can hoist BI if one branch destination is the successor of the other, | ||||||||
666 | // or both have common successor which we check by seeing if the | ||||||||
667 | // intersection of their successors is non-empty. | ||||||||
668 | // TODO: This could be expanded to allowing branches where both ends | ||||||||
669 | // eventually converge to a single block. | ||||||||
670 | SmallPtrSet<BasicBlock *, 4> TrueDestSucc, FalseDestSucc; | ||||||||
671 | TrueDestSucc.insert(succ_begin(TrueDest), succ_end(TrueDest)); | ||||||||
672 | FalseDestSucc.insert(succ_begin(FalseDest), succ_end(FalseDest)); | ||||||||
673 | BasicBlock *CommonSucc = nullptr; | ||||||||
674 | if (TrueDestSucc.count(FalseDest)) { | ||||||||
675 | CommonSucc = FalseDest; | ||||||||
676 | } else if (FalseDestSucc.count(TrueDest)) { | ||||||||
677 | CommonSucc = TrueDest; | ||||||||
678 | } else { | ||||||||
679 | set_intersect(TrueDestSucc, FalseDestSucc); | ||||||||
680 | // If there's one common successor use that. | ||||||||
681 | if (TrueDestSucc.size() == 1) | ||||||||
682 | CommonSucc = *TrueDestSucc.begin(); | ||||||||
683 | // If there's more than one pick whichever appears first in the block list | ||||||||
684 | // (we can't use the value returned by TrueDestSucc.begin() as it's | ||||||||
685 | // unpredicatable which element gets returned). | ||||||||
686 | else if (!TrueDestSucc.empty()) { | ||||||||
687 | Function *F = TrueDest->getParent(); | ||||||||
688 | auto IsSucc = [&](BasicBlock &BB) { return TrueDestSucc.count(&BB); }; | ||||||||
689 | auto It = llvm::find_if(*F, IsSucc); | ||||||||
690 | assert(It != F->end() && "Could not find successor in function")((void)0); | ||||||||
691 | CommonSucc = &*It; | ||||||||
692 | } | ||||||||
693 | } | ||||||||
694 | // The common successor has to be dominated by the branch, as otherwise | ||||||||
695 | // there will be some other path to the successor that will not be | ||||||||
696 | // controlled by this branch so any phi we hoist would be controlled by the | ||||||||
697 | // wrong condition. This also takes care of avoiding hoisting of loop back | ||||||||
698 | // edges. | ||||||||
699 | // TODO: In some cases this could be relaxed if the successor is dominated | ||||||||
700 | // by another block that's been hoisted and we can guarantee that the | ||||||||
701 | // control flow has been replicated exactly. | ||||||||
702 | if (CommonSucc && DT->dominates(BI, CommonSucc)) | ||||||||
703 | HoistableBranches[BI] = CommonSucc; | ||||||||
704 | } | ||||||||
705 | |||||||||
706 | bool canHoistPHI(PHINode *PN) { | ||||||||
707 | // The phi must have loop invariant operands. | ||||||||
708 | if (!ControlFlowHoisting || !CurLoop->hasLoopInvariantOperands(PN)) | ||||||||
709 | return false; | ||||||||
710 | // We can hoist phis if the block they are in is the target of hoistable | ||||||||
711 | // branches which cover all of the predecessors of the block. | ||||||||
712 | SmallPtrSet<BasicBlock *, 8> PredecessorBlocks; | ||||||||
713 | BasicBlock *BB = PN->getParent(); | ||||||||
714 | for (BasicBlock *PredBB : predecessors(BB)) | ||||||||
715 | PredecessorBlocks.insert(PredBB); | ||||||||
716 | // If we have less predecessor blocks than predecessors then the phi will | ||||||||
717 | // have more than one incoming value for the same block which we can't | ||||||||
718 | // handle. | ||||||||
719 | // TODO: This could be handled be erasing some of the duplicate incoming | ||||||||
720 | // values. | ||||||||
721 | if (PredecessorBlocks.size() != pred_size(BB)) | ||||||||
722 | return false; | ||||||||
723 | for (auto &Pair : HoistableBranches) { | ||||||||
724 | if (Pair.second == BB) { | ||||||||
725 | // Which blocks are predecessors via this branch depends on if the | ||||||||
726 | // branch is triangle-like or diamond-like. | ||||||||
727 | if (Pair.first->getSuccessor(0) == BB) { | ||||||||
728 | PredecessorBlocks.erase(Pair.first->getParent()); | ||||||||
729 | PredecessorBlocks.erase(Pair.first->getSuccessor(1)); | ||||||||
730 | } else if (Pair.first->getSuccessor(1) == BB) { | ||||||||
731 | PredecessorBlocks.erase(Pair.first->getParent()); | ||||||||
732 | PredecessorBlocks.erase(Pair.first->getSuccessor(0)); | ||||||||
733 | } else { | ||||||||
734 | PredecessorBlocks.erase(Pair.first->getSuccessor(0)); | ||||||||
735 | PredecessorBlocks.erase(Pair.first->getSuccessor(1)); | ||||||||
736 | } | ||||||||
737 | } | ||||||||
738 | } | ||||||||
739 | // PredecessorBlocks will now be empty if for every predecessor of BB we | ||||||||
740 | // found a hoistable branch source. | ||||||||
741 | return PredecessorBlocks.empty(); | ||||||||
742 | } | ||||||||
743 | |||||||||
744 | BasicBlock *getOrCreateHoistedBlock(BasicBlock *BB) { | ||||||||
745 | if (!ControlFlowHoisting) | ||||||||
746 | return CurLoop->getLoopPreheader(); | ||||||||
747 | // If BB has already been hoisted, return that | ||||||||
748 | if (HoistDestinationMap.count(BB)) | ||||||||
749 | return HoistDestinationMap[BB]; | ||||||||
750 | |||||||||
751 | // Check if this block is conditional based on a pending branch | ||||||||
752 | auto HasBBAsSuccessor = | ||||||||
753 | [&](DenseMap<BranchInst *, BasicBlock *>::value_type &Pair) { | ||||||||
754 | return BB != Pair.second && (Pair.first->getSuccessor(0) == BB || | ||||||||
755 | Pair.first->getSuccessor(1) == BB); | ||||||||
756 | }; | ||||||||
757 | auto It = llvm::find_if(HoistableBranches, HasBBAsSuccessor); | ||||||||
758 | |||||||||
759 | // If not involved in a pending branch, hoist to preheader | ||||||||
760 | BasicBlock *InitialPreheader = CurLoop->getLoopPreheader(); | ||||||||
761 | if (It == HoistableBranches.end()) { | ||||||||
762 | LLVM_DEBUG(dbgs() << "LICM using "do { } while (false) | ||||||||
763 | << InitialPreheader->getNameOrAsOperand()do { } while (false) | ||||||||
764 | << " as hoist destination for "do { } while (false) | ||||||||
765 | << BB->getNameOrAsOperand() << "\n")do { } while (false); | ||||||||
766 | HoistDestinationMap[BB] = InitialPreheader; | ||||||||
767 | return InitialPreheader; | ||||||||
768 | } | ||||||||
769 | BranchInst *BI = It->first; | ||||||||
770 | assert(std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) ==((void)0) | ||||||||
771 | HoistableBranches.end() &&((void)0) | ||||||||
772 | "BB is expected to be the target of at most one branch")((void)0); | ||||||||
773 | |||||||||
774 | LLVMContext &C = BB->getContext(); | ||||||||
775 | BasicBlock *TrueDest = BI->getSuccessor(0); | ||||||||
776 | BasicBlock *FalseDest = BI->getSuccessor(1); | ||||||||
777 | BasicBlock *CommonSucc = HoistableBranches[BI]; | ||||||||
778 | BasicBlock *HoistTarget = getOrCreateHoistedBlock(BI->getParent()); | ||||||||
779 | |||||||||
780 | // Create hoisted versions of blocks that currently don't have them | ||||||||
781 | auto CreateHoistedBlock = [&](BasicBlock *Orig) { | ||||||||
782 | if (HoistDestinationMap.count(Orig)) | ||||||||
783 | return HoistDestinationMap[Orig]; | ||||||||
784 | BasicBlock *New = | ||||||||
785 | BasicBlock::Create(C, Orig->getName() + ".licm", Orig->getParent()); | ||||||||
786 | HoistDestinationMap[Orig] = New; | ||||||||
787 | DT->addNewBlock(New, HoistTarget); | ||||||||
788 | if (CurLoop->getParentLoop()) | ||||||||
789 | CurLoop->getParentLoop()->addBasicBlockToLoop(New, *LI); | ||||||||
790 | ++NumCreatedBlocks; | ||||||||
791 | LLVM_DEBUG(dbgs() << "LICM created " << New->getName()do { } while (false) | ||||||||
792 | << " as hoist destination for " << Orig->getName()do { } while (false) | ||||||||
793 | << "\n")do { } while (false); | ||||||||
794 | return New; | ||||||||
795 | }; | ||||||||
796 | BasicBlock *HoistTrueDest = CreateHoistedBlock(TrueDest); | ||||||||
797 | BasicBlock *HoistFalseDest = CreateHoistedBlock(FalseDest); | ||||||||
798 | BasicBlock *HoistCommonSucc = CreateHoistedBlock(CommonSucc); | ||||||||
799 | |||||||||
800 | // Link up these blocks with branches. | ||||||||
801 | if (!HoistCommonSucc->getTerminator()) { | ||||||||
802 | // The new common successor we've generated will branch to whatever that | ||||||||
803 | // hoist target branched to. | ||||||||
804 | BasicBlock *TargetSucc = HoistTarget->getSingleSuccessor(); | ||||||||
805 | assert(TargetSucc && "Expected hoist target to have a single successor")((void)0); | ||||||||
806 | HoistCommonSucc->moveBefore(TargetSucc); | ||||||||
807 | BranchInst::Create(TargetSucc, HoistCommonSucc); | ||||||||
808 | } | ||||||||
809 | if (!HoistTrueDest->getTerminator()) { | ||||||||
810 | HoistTrueDest->moveBefore(HoistCommonSucc); | ||||||||
811 | BranchInst::Create(HoistCommonSucc, HoistTrueDest); | ||||||||
812 | } | ||||||||
813 | if (!HoistFalseDest->getTerminator()) { | ||||||||
814 | HoistFalseDest->moveBefore(HoistCommonSucc); | ||||||||
815 | BranchInst::Create(HoistCommonSucc, HoistFalseDest); | ||||||||
816 | } | ||||||||
817 | |||||||||
818 | // If BI is being cloned to what was originally the preheader then | ||||||||
819 | // HoistCommonSucc will now be the new preheader. | ||||||||
820 | if (HoistTarget == InitialPreheader) { | ||||||||
821 | // Phis in the loop header now need to use the new preheader. | ||||||||
822 | InitialPreheader->replaceSuccessorsPhiUsesWith(HoistCommonSucc); | ||||||||
823 | if (MSSAU) | ||||||||
824 | MSSAU->wireOldPredecessorsToNewImmediatePredecessor( | ||||||||
825 | HoistTarget->getSingleSuccessor(), HoistCommonSucc, {HoistTarget}); | ||||||||
826 | // The new preheader dominates the loop header. | ||||||||
827 | DomTreeNode *PreheaderNode = DT->getNode(HoistCommonSucc); | ||||||||
828 | DomTreeNode *HeaderNode = DT->getNode(CurLoop->getHeader()); | ||||||||
829 | DT->changeImmediateDominator(HeaderNode, PreheaderNode); | ||||||||
830 | // The preheader hoist destination is now the new preheader, with the | ||||||||
831 | // exception of the hoist destination of this branch. | ||||||||
832 | for (auto &Pair : HoistDestinationMap) | ||||||||
833 | if (Pair.second == InitialPreheader && Pair.first != BI->getParent()) | ||||||||
834 | Pair.second = HoistCommonSucc; | ||||||||
835 | } | ||||||||
836 | |||||||||
837 | // Now finally clone BI. | ||||||||
838 | ReplaceInstWithInst( | ||||||||
839 | HoistTarget->getTerminator(), | ||||||||
840 | BranchInst::Create(HoistTrueDest, HoistFalseDest, BI->getCondition())); | ||||||||
841 | ++NumClonedBranches; | ||||||||
842 | |||||||||
843 | assert(CurLoop->getLoopPreheader() &&((void)0) | ||||||||
844 | "Hoisting blocks should not have destroyed preheader")((void)0); | ||||||||
845 | return HoistDestinationMap[BB]; | ||||||||
846 | } | ||||||||
847 | }; | ||||||||
848 | } // namespace | ||||||||
849 | |||||||||
850 | // Hoisting/sinking instruction out of a loop isn't always beneficial. It's only | ||||||||
851 | // only worthwhile if the destination block is actually colder than current | ||||||||
852 | // block. | ||||||||
853 | static bool worthSinkOrHoistInst(Instruction &I, BasicBlock *DstBlock, | ||||||||
854 | OptimizationRemarkEmitter *ORE, | ||||||||
855 | BlockFrequencyInfo *BFI) { | ||||||||
856 | // Check block frequency only when runtime profile is available | ||||||||
857 | // to avoid pathological cases. With static profile, lean towards | ||||||||
858 | // hosting because it helps canonicalize the loop for vectorizer. | ||||||||
859 | if (!DstBlock->getParent()->hasProfileData()) | ||||||||
860 | return true; | ||||||||
861 | |||||||||
862 | if (!HoistSinkColdnessThreshold || !BFI) | ||||||||
863 | return true; | ||||||||
864 | |||||||||
865 | BasicBlock *SrcBlock = I.getParent(); | ||||||||
866 | if (BFI->getBlockFreq(DstBlock).getFrequency() / HoistSinkColdnessThreshold > | ||||||||
867 | BFI->getBlockFreq(SrcBlock).getFrequency()) { | ||||||||
868 | ORE->emit([&]() { | ||||||||
869 | return OptimizationRemarkMissed(DEBUG_TYPE"licm", "SinkHoistInst", &I) | ||||||||
870 | << "failed to sink or hoist instruction because containing block " | ||||||||
871 | "has lower frequency than destination block"; | ||||||||
872 | }); | ||||||||
873 | return false; | ||||||||
874 | } | ||||||||
875 | |||||||||
876 | return true; | ||||||||
877 | } | ||||||||
878 | |||||||||
879 | /// Walk the specified region of the CFG (defined by all blocks dominated by | ||||||||
880 | /// the specified block, and that are in the current loop) in depth first | ||||||||
881 | /// order w.r.t the DominatorTree. This allows us to visit definitions before | ||||||||
882 | /// uses, allowing us to hoist a loop body in one pass without iteration. | ||||||||
883 | /// | ||||||||
884 | bool llvm::hoistRegion(DomTreeNode *N, AAResults *AA, LoopInfo *LI, | ||||||||
885 | DominatorTree *DT, BlockFrequencyInfo *BFI, | ||||||||
886 | TargetLibraryInfo *TLI, Loop *CurLoop, | ||||||||
887 | AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU, | ||||||||
888 | ScalarEvolution *SE, ICFLoopSafetyInfo *SafetyInfo, | ||||||||
889 | SinkAndHoistLICMFlags &Flags, | ||||||||
890 | OptimizationRemarkEmitter *ORE, bool LoopNestMode) { | ||||||||
891 | // Verify inputs. | ||||||||
892 | assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&((void)0) | ||||||||
893 | CurLoop != nullptr && SafetyInfo != nullptr &&((void)0) | ||||||||
894 | "Unexpected input to hoistRegion.")((void)0); | ||||||||
895 | assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&((void)0) | ||||||||
896 | "Either AliasSetTracker or MemorySSA should be initialized.")((void)0); | ||||||||
897 | |||||||||
898 | ControlFlowHoister CFH(LI, DT, CurLoop, MSSAU); | ||||||||
899 | |||||||||
900 | // Keep track of instructions that have been hoisted, as they may need to be | ||||||||
901 | // re-hoisted if they end up not dominating all of their uses. | ||||||||
902 | SmallVector<Instruction *, 16> HoistedInstructions; | ||||||||
903 | |||||||||
904 | // For PHI hoisting to work we need to hoist blocks before their successors. | ||||||||
905 | // We can do this by iterating through the blocks in the loop in reverse | ||||||||
906 | // post-order. | ||||||||
907 | LoopBlocksRPO Worklist(CurLoop); | ||||||||
908 | Worklist.perform(LI); | ||||||||
909 | bool Changed = false; | ||||||||
910 | for (BasicBlock *BB : Worklist) { | ||||||||
911 | // Only need to process the contents of this block if it is not part of a | ||||||||
912 | // subloop (which would already have been processed). | ||||||||
913 | if (!LoopNestMode && inSubLoop(BB, CurLoop, LI)) | ||||||||
914 | continue; | ||||||||
915 | |||||||||
916 | for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) { | ||||||||
917 | Instruction &I = *II++; | ||||||||
918 | // Try constant folding this instruction. If all the operands are | ||||||||
919 | // constants, it is technically hoistable, but it would be better to | ||||||||
920 | // just fold it. | ||||||||
921 | if (Constant *C = ConstantFoldInstruction( | ||||||||
922 | &I, I.getModule()->getDataLayout(), TLI)) { | ||||||||
923 | LLVM_DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *Cdo { } while (false) | ||||||||
924 | << '\n')do { } while (false); | ||||||||
925 | if (CurAST) | ||||||||
926 | CurAST->copyValue(&I, C); | ||||||||
927 | // FIXME MSSA: Such replacements may make accesses unoptimized (D51960). | ||||||||
928 | I.replaceAllUsesWith(C); | ||||||||
929 | if (isInstructionTriviallyDead(&I, TLI)) | ||||||||
930 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | ||||||||
931 | Changed = true; | ||||||||
932 | continue; | ||||||||
933 | } | ||||||||
934 | |||||||||
935 | // Try hoisting the instruction out to the preheader. We can only do | ||||||||
936 | // this if all of the operands of the instruction are loop invariant and | ||||||||
937 | // if it is safe to hoist the instruction. We also check block frequency | ||||||||
938 | // to make sure instruction only gets hoisted into colder blocks. | ||||||||
939 | // TODO: It may be safe to hoist if we are hoisting to a conditional block | ||||||||
940 | // and we have accurately duplicated the control flow from the loop header | ||||||||
941 | // to that block. | ||||||||
942 | if (CurLoop->hasLoopInvariantOperands(&I) && | ||||||||
943 | canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags, | ||||||||
944 | ORE) && | ||||||||
945 | worthSinkOrHoistInst(I, CurLoop->getLoopPreheader(), ORE, BFI) && | ||||||||
946 | isSafeToExecuteUnconditionally( | ||||||||
947 | I, DT, TLI, CurLoop, SafetyInfo, ORE, | ||||||||
948 | CurLoop->getLoopPreheader()->getTerminator())) { | ||||||||
949 | hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo, | ||||||||
950 | MSSAU, SE, ORE); | ||||||||
951 | HoistedInstructions.push_back(&I); | ||||||||
952 | Changed = true; | ||||||||
953 | continue; | ||||||||
954 | } | ||||||||
955 | |||||||||
956 | // Attempt to remove floating point division out of the loop by | ||||||||
957 | // converting it to a reciprocal multiplication. | ||||||||
958 | if (I.getOpcode() == Instruction::FDiv && I.hasAllowReciprocal() && | ||||||||
959 | CurLoop->isLoopInvariant(I.getOperand(1))) { | ||||||||
960 | auto Divisor = I.getOperand(1); | ||||||||
961 | auto One = llvm::ConstantFP::get(Divisor->getType(), 1.0); | ||||||||
962 | auto ReciprocalDivisor = BinaryOperator::CreateFDiv(One, Divisor); | ||||||||
963 | ReciprocalDivisor->setFastMathFlags(I.getFastMathFlags()); | ||||||||
964 | SafetyInfo->insertInstructionTo(ReciprocalDivisor, I.getParent()); | ||||||||
965 | ReciprocalDivisor->insertBefore(&I); | ||||||||
966 | |||||||||
967 | auto Product = | ||||||||
968 | BinaryOperator::CreateFMul(I.getOperand(0), ReciprocalDivisor); | ||||||||
969 | Product->setFastMathFlags(I.getFastMathFlags()); | ||||||||
970 | SafetyInfo->insertInstructionTo(Product, I.getParent()); | ||||||||
971 | Product->insertAfter(&I); | ||||||||
972 | I.replaceAllUsesWith(Product); | ||||||||
973 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | ||||||||
974 | |||||||||
975 | hoist(*ReciprocalDivisor, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), | ||||||||
976 | SafetyInfo, MSSAU, SE, ORE); | ||||||||
977 | HoistedInstructions.push_back(ReciprocalDivisor); | ||||||||
978 | Changed = true; | ||||||||
979 | continue; | ||||||||
980 | } | ||||||||
981 | |||||||||
982 | auto IsInvariantStart = [&](Instruction &I) { | ||||||||
983 | using namespace PatternMatch; | ||||||||
984 | return I.use_empty() && | ||||||||
985 | match(&I, m_Intrinsic<Intrinsic::invariant_start>()); | ||||||||
986 | }; | ||||||||
987 | auto MustExecuteWithoutWritesBefore = [&](Instruction &I) { | ||||||||
988 | return SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop) && | ||||||||
989 | SafetyInfo->doesNotWriteMemoryBefore(I, CurLoop); | ||||||||
990 | }; | ||||||||
991 | if ((IsInvariantStart(I) || isGuard(&I)) && | ||||||||
992 | CurLoop->hasLoopInvariantOperands(&I) && | ||||||||
993 | MustExecuteWithoutWritesBefore(I)) { | ||||||||
994 | hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo, | ||||||||
995 | MSSAU, SE, ORE); | ||||||||
996 | HoistedInstructions.push_back(&I); | ||||||||
997 | Changed = true; | ||||||||
998 | continue; | ||||||||
999 | } | ||||||||
1000 | |||||||||
1001 | if (PHINode *PN = dyn_cast<PHINode>(&I)) { | ||||||||
1002 | if (CFH.canHoistPHI(PN)) { | ||||||||
1003 | // Redirect incoming blocks first to ensure that we create hoisted | ||||||||
1004 | // versions of those blocks before we hoist the phi. | ||||||||
1005 | for (unsigned int i = 0; i < PN->getNumIncomingValues(); ++i) | ||||||||
1006 | PN->setIncomingBlock( | ||||||||
1007 | i, CFH.getOrCreateHoistedBlock(PN->getIncomingBlock(i))); | ||||||||
1008 | hoist(*PN, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo, | ||||||||
1009 | MSSAU, SE, ORE); | ||||||||
1010 | assert(DT->dominates(PN, BB) && "Conditional PHIs not expected")((void)0); | ||||||||
1011 | Changed = true; | ||||||||
1012 | continue; | ||||||||
1013 | } | ||||||||
1014 | } | ||||||||
1015 | |||||||||
1016 | // Remember possibly hoistable branches so we can actually hoist them | ||||||||
1017 | // later if needed. | ||||||||
1018 | if (BranchInst *BI = dyn_cast<BranchInst>(&I)) | ||||||||
1019 | CFH.registerPossiblyHoistableBranch(BI); | ||||||||
1020 | } | ||||||||
1021 | } | ||||||||
1022 | |||||||||
1023 | // If we hoisted instructions to a conditional block they may not dominate | ||||||||
1024 | // their uses that weren't hoisted (such as phis where some operands are not | ||||||||
1025 | // loop invariant). If so make them unconditional by moving them to their | ||||||||
1026 | // immediate dominator. We iterate through the instructions in reverse order | ||||||||
1027 | // which ensures that when we rehoist an instruction we rehoist its operands, | ||||||||
1028 | // and also keep track of where in the block we are rehoisting to to make sure | ||||||||
1029 | // that we rehoist instructions before the instructions that use them. | ||||||||
1030 | Instruction *HoistPoint = nullptr; | ||||||||
1031 | if (ControlFlowHoisting) { | ||||||||
1032 | for (Instruction *I : reverse(HoistedInstructions)) { | ||||||||
1033 | if (!llvm::all_of(I->uses(), | ||||||||
1034 | [&](Use &U) { return DT->dominates(I, U); })) { | ||||||||
1035 | BasicBlock *Dominator = | ||||||||
1036 | DT->getNode(I->getParent())->getIDom()->getBlock(); | ||||||||
1037 | if (!HoistPoint || !DT->dominates(HoistPoint->getParent(), Dominator)) { | ||||||||
1038 | if (HoistPoint) | ||||||||
1039 | assert(DT->dominates(Dominator, HoistPoint->getParent()) &&((void)0) | ||||||||
1040 | "New hoist point expected to dominate old hoist point")((void)0); | ||||||||
1041 | HoistPoint = Dominator->getTerminator(); | ||||||||
1042 | } | ||||||||
1043 | LLVM_DEBUG(dbgs() << "LICM rehoisting to "do { } while (false) | ||||||||
1044 | << HoistPoint->getParent()->getNameOrAsOperand()do { } while (false) | ||||||||
1045 | << ": " << *I << "\n")do { } while (false); | ||||||||
1046 | moveInstructionBefore(*I, *HoistPoint, *SafetyInfo, MSSAU, SE); | ||||||||
1047 | HoistPoint = I; | ||||||||
1048 | Changed = true; | ||||||||
1049 | } | ||||||||
1050 | } | ||||||||
1051 | } | ||||||||
1052 | if (MSSAU && VerifyMemorySSA) | ||||||||
1053 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||
1054 | |||||||||
1055 | // Now that we've finished hoisting make sure that LI and DT are still | ||||||||
1056 | // valid. | ||||||||
1057 | #ifdef EXPENSIVE_CHECKS | ||||||||
1058 | if (Changed) { | ||||||||
1059 | assert(DT->verify(DominatorTree::VerificationLevel::Fast) &&((void)0) | ||||||||
1060 | "Dominator tree verification failed")((void)0); | ||||||||
1061 | LI->verify(*DT); | ||||||||
1062 | } | ||||||||
1063 | #endif | ||||||||
1064 | |||||||||
1065 | return Changed; | ||||||||
1066 | } | ||||||||
1067 | |||||||||
1068 | // Return true if LI is invariant within scope of the loop. LI is invariant if | ||||||||
1069 | // CurLoop is dominated by an invariant.start representing the same memory | ||||||||
1070 | // location and size as the memory location LI loads from, and also the | ||||||||
1071 | // invariant.start has no uses. | ||||||||
1072 | static bool isLoadInvariantInLoop(LoadInst *LI, DominatorTree *DT, | ||||||||
1073 | Loop *CurLoop) { | ||||||||
1074 | Value *Addr = LI->getOperand(0); | ||||||||
1075 | const DataLayout &DL = LI->getModule()->getDataLayout(); | ||||||||
1076 | const TypeSize LocSizeInBits = DL.getTypeSizeInBits(LI->getType()); | ||||||||
1077 | |||||||||
1078 | // It is not currently possible for clang to generate an invariant.start | ||||||||
1079 | // intrinsic with scalable vector types because we don't support thread local | ||||||||
1080 | // sizeless types and we don't permit sizeless types in structs or classes. | ||||||||
1081 | // Furthermore, even if support is added for this in future the intrinsic | ||||||||
1082 | // itself is defined to have a size of -1 for variable sized objects. This | ||||||||
1083 | // makes it impossible to verify if the intrinsic envelops our region of | ||||||||
1084 | // interest. For example, both <vscale x 32 x i8> and <vscale x 16 x i8> | ||||||||
1085 | // types would have a -1 parameter, but the former is clearly double the size | ||||||||
1086 | // of the latter. | ||||||||
1087 | if (LocSizeInBits.isScalable()) | ||||||||
1088 | return false; | ||||||||
1089 | |||||||||
1090 | // if the type is i8 addrspace(x)*, we know this is the type of | ||||||||
1091 | // llvm.invariant.start operand | ||||||||
1092 | auto *PtrInt8Ty = PointerType::get(Type::getInt8Ty(LI->getContext()), | ||||||||
1093 | LI->getPointerAddressSpace()); | ||||||||
1094 | unsigned BitcastsVisited = 0; | ||||||||
1095 | // Look through bitcasts until we reach the i8* type (this is invariant.start | ||||||||
1096 | // operand type). | ||||||||
1097 | while (Addr->getType() != PtrInt8Ty) { | ||||||||
1098 | auto *BC = dyn_cast<BitCastInst>(Addr); | ||||||||
1099 | // Avoid traversing high number of bitcast uses. | ||||||||
1100 | if (++BitcastsVisited > MaxNumUsesTraversed || !BC) | ||||||||
1101 | return false; | ||||||||
1102 | Addr = BC->getOperand(0); | ||||||||
1103 | } | ||||||||
1104 | |||||||||
1105 | unsigned UsesVisited = 0; | ||||||||
1106 | // Traverse all uses of the load operand value, to see if invariant.start is | ||||||||
1107 | // one of the uses, and whether it dominates the load instruction. | ||||||||
1108 | for (auto *U : Addr->users()) { | ||||||||
1109 | // Avoid traversing for Load operand with high number of users. | ||||||||
1110 | if (++UsesVisited > MaxNumUsesTraversed) | ||||||||
1111 | return false; | ||||||||
1112 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(U); | ||||||||
1113 | // If there are escaping uses of invariant.start instruction, the load maybe | ||||||||
1114 | // non-invariant. | ||||||||
1115 | if (!II || II->getIntrinsicID() != Intrinsic::invariant_start || | ||||||||
1116 | !II->use_empty()) | ||||||||
1117 | continue; | ||||||||
1118 | ConstantInt *InvariantSize = cast<ConstantInt>(II->getArgOperand(0)); | ||||||||
1119 | // The intrinsic supports having a -1 argument for variable sized objects | ||||||||
1120 | // so we should check for that here. | ||||||||
1121 | if (InvariantSize->isNegative()) | ||||||||
1122 | continue; | ||||||||
1123 | uint64_t InvariantSizeInBits = InvariantSize->getSExtValue() * 8; | ||||||||
1124 | // Confirm the invariant.start location size contains the load operand size | ||||||||
1125 | // in bits. Also, the invariant.start should dominate the load, and we | ||||||||
1126 | // should not hoist the load out of a loop that contains this dominating | ||||||||
1127 | // invariant.start. | ||||||||
1128 | if (LocSizeInBits.getFixedSize() <= InvariantSizeInBits && | ||||||||
1129 | DT->properlyDominates(II->getParent(), CurLoop->getHeader())) | ||||||||
1130 | return true; | ||||||||
1131 | } | ||||||||
1132 | |||||||||
1133 | return false; | ||||||||
1134 | } | ||||||||
1135 | |||||||||
1136 | namespace { | ||||||||
1137 | /// Return true if-and-only-if we know how to (mechanically) both hoist and | ||||||||
1138 | /// sink a given instruction out of a loop. Does not address legality | ||||||||
1139 | /// concerns such as aliasing or speculation safety. | ||||||||
1140 | bool isHoistableAndSinkableInst(Instruction &I) { | ||||||||
1141 | // Only these instructions are hoistable/sinkable. | ||||||||
1142 | return (isa<LoadInst>(I) || isa<StoreInst>(I) || isa<CallInst>(I) || | ||||||||
1143 | isa<FenceInst>(I) || isa<CastInst>(I) || isa<UnaryOperator>(I) || | ||||||||
1144 | isa<BinaryOperator>(I) || isa<SelectInst>(I) || | ||||||||
1145 | isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || | ||||||||
1146 | isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || | ||||||||
1147 | isa<ShuffleVectorInst>(I) || isa<ExtractValueInst>(I) || | ||||||||
1148 | isa<InsertValueInst>(I) || isa<FreezeInst>(I)); | ||||||||
1149 | } | ||||||||
1150 | /// Return true if all of the alias sets within this AST are known not to | ||||||||
1151 | /// contain a Mod, or if MSSA knows there are no MemoryDefs in the loop. | ||||||||
1152 | bool isReadOnly(AliasSetTracker *CurAST, const MemorySSAUpdater *MSSAU, | ||||||||
1153 | const Loop *L) { | ||||||||
1154 | if (CurAST) { | ||||||||
1155 | for (AliasSet &AS : *CurAST) { | ||||||||
1156 | if (!AS.isForwardingAliasSet() && AS.isMod()) { | ||||||||
1157 | return false; | ||||||||
1158 | } | ||||||||
1159 | } | ||||||||
1160 | return true; | ||||||||
1161 | } else { /*MSSAU*/ | ||||||||
1162 | for (auto *BB : L->getBlocks()) | ||||||||
1163 | if (MSSAU->getMemorySSA()->getBlockDefs(BB)) | ||||||||
1164 | return false; | ||||||||
1165 | return true; | ||||||||
1166 | } | ||||||||
1167 | } | ||||||||
1168 | |||||||||
1169 | /// Return true if I is the only Instruction with a MemoryAccess in L. | ||||||||
1170 | bool isOnlyMemoryAccess(const Instruction *I, const Loop *L, | ||||||||
1171 | const MemorySSAUpdater *MSSAU) { | ||||||||
1172 | for (auto *BB : L->getBlocks()) | ||||||||
1173 | if (auto *Accs = MSSAU->getMemorySSA()->getBlockAccesses(BB)) { | ||||||||
1174 | int NotAPhi = 0; | ||||||||
1175 | for (const auto &Acc : *Accs) { | ||||||||
1176 | if (isa<MemoryPhi>(&Acc)) | ||||||||
1177 | continue; | ||||||||
1178 | const auto *MUD = cast<MemoryUseOrDef>(&Acc); | ||||||||
1179 | if (MUD->getMemoryInst() != I || NotAPhi++ == 1) | ||||||||
1180 | return false; | ||||||||
1181 | } | ||||||||
1182 | } | ||||||||
1183 | return true; | ||||||||
1184 | } | ||||||||
1185 | } | ||||||||
1186 | |||||||||
1187 | bool llvm::canSinkOrHoistInst(Instruction &I, AAResults *AA, DominatorTree *DT, | ||||||||
1188 | Loop *CurLoop, AliasSetTracker *CurAST, | ||||||||
1189 | MemorySSAUpdater *MSSAU, | ||||||||
1190 | bool TargetExecutesOncePerLoop, | ||||||||
1191 | SinkAndHoistLICMFlags *Flags, | ||||||||
1192 | OptimizationRemarkEmitter *ORE) { | ||||||||
1193 | assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&((void)0) | ||||||||
1194 | "Either AliasSetTracker or MemorySSA should be initialized.")((void)0); | ||||||||
1195 | |||||||||
1196 | // If we don't understand the instruction, bail early. | ||||||||
1197 | if (!isHoistableAndSinkableInst(I)) | ||||||||
| |||||||||
1198 | return false; | ||||||||
1199 | |||||||||
1200 | MemorySSA *MSSA = MSSAU ? MSSAU->getMemorySSA() : nullptr; | ||||||||
1201 | if (MSSA
| ||||||||
1202 | assert(Flags != nullptr && "Flags cannot be null.")((void)0); | ||||||||
1203 | |||||||||
1204 | // Loads have extra constraints we have to verify before we can hoist them. | ||||||||
1205 | if (LoadInst *LI
| ||||||||
1206 | if (!LI->isUnordered()) | ||||||||
1207 | return false; // Don't sink/hoist volatile or ordered atomic loads! | ||||||||
1208 | |||||||||
1209 | // Loads from constant memory are always safe to move, even if they end up | ||||||||
1210 | // in the same alias set as something that ends up being modified. | ||||||||
1211 | if (AA->pointsToConstantMemory(LI->getOperand(0))) | ||||||||
1212 | return true; | ||||||||
1213 | if (LI->hasMetadata(LLVMContext::MD_invariant_load)) | ||||||||
1214 | return true; | ||||||||
1215 | |||||||||
1216 | if (LI->isAtomic() && !TargetExecutesOncePerLoop) | ||||||||
1217 | return false; // Don't risk duplicating unordered loads | ||||||||
1218 | |||||||||
1219 | // This checks for an invariant.start dominating the load. | ||||||||
1220 | if (isLoadInvariantInLoop(LI, DT, CurLoop)) | ||||||||
1221 | return true; | ||||||||
1222 | |||||||||
1223 | bool Invalidated; | ||||||||
1224 | if (CurAST) | ||||||||
1225 | Invalidated = pointerInvalidatedByLoop(MemoryLocation::get(LI), CurAST, | ||||||||
1226 | CurLoop, AA); | ||||||||
1227 | else | ||||||||
1228 | Invalidated = pointerInvalidatedByLoopWithMSSA( | ||||||||
1229 | MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(LI)), CurLoop, I, *Flags); | ||||||||
| |||||||||
1230 | // Check loop-invariant address because this may also be a sinkable load | ||||||||
1231 | // whose address is not necessarily loop-invariant. | ||||||||
1232 | if (ORE && Invalidated && CurLoop->isLoopInvariant(LI->getPointerOperand())) | ||||||||
1233 | ORE->emit([&]() { | ||||||||
1234 | return OptimizationRemarkMissed( | ||||||||
1235 | DEBUG_TYPE"licm", "LoadWithLoopInvariantAddressInvalidated", LI) | ||||||||
1236 | << "failed to move load with loop-invariant address " | ||||||||
1237 | "because the loop may invalidate its value"; | ||||||||
1238 | }); | ||||||||
1239 | |||||||||
1240 | return !Invalidated; | ||||||||
1241 | } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { | ||||||||
1242 | // Don't sink or hoist dbg info; it's legal, but not useful. | ||||||||
1243 | if (isa<DbgInfoIntrinsic>(I)) | ||||||||
1244 | return false; | ||||||||
1245 | |||||||||
1246 | // Don't sink calls which can throw. | ||||||||
1247 | if (CI->mayThrow()) | ||||||||
1248 | return false; | ||||||||
1249 | |||||||||
1250 | // Convergent attribute has been used on operations that involve | ||||||||
1251 | // inter-thread communication which results are implicitly affected by the | ||||||||
1252 | // enclosing control flows. It is not safe to hoist or sink such operations | ||||||||
1253 | // across control flow. | ||||||||
1254 | if (CI->isConvergent()) | ||||||||
1255 | return false; | ||||||||
1256 | |||||||||
1257 | using namespace PatternMatch; | ||||||||
1258 | if (match(CI, m_Intrinsic<Intrinsic::assume>())) | ||||||||
1259 | // Assumes don't actually alias anything or throw | ||||||||
1260 | return true; | ||||||||
1261 | |||||||||
1262 | if (match(CI, m_Intrinsic<Intrinsic::experimental_widenable_condition>())) | ||||||||
1263 | // Widenable conditions don't actually alias anything or throw | ||||||||
1264 | return true; | ||||||||
1265 | |||||||||
1266 | // Handle simple cases by querying alias analysis. | ||||||||
1267 | FunctionModRefBehavior Behavior = AA->getModRefBehavior(CI); | ||||||||
1268 | if (Behavior == FMRB_DoesNotAccessMemory) | ||||||||
1269 | return true; | ||||||||
1270 | if (AAResults::onlyReadsMemory(Behavior)) { | ||||||||
1271 | // A readonly argmemonly function only reads from memory pointed to by | ||||||||
1272 | // it's arguments with arbitrary offsets. If we can prove there are no | ||||||||
1273 | // writes to this memory in the loop, we can hoist or sink. | ||||||||
1274 | if (AAResults::onlyAccessesArgPointees(Behavior)) { | ||||||||
1275 | // TODO: expand to writeable arguments | ||||||||
1276 | for (Value *Op : CI->arg_operands()) | ||||||||
1277 | if (Op->getType()->isPointerTy()) { | ||||||||
1278 | bool Invalidated; | ||||||||
1279 | if (CurAST) | ||||||||
1280 | Invalidated = pointerInvalidatedByLoop( | ||||||||
1281 | MemoryLocation::getBeforeOrAfter(Op), CurAST, CurLoop, AA); | ||||||||
1282 | else | ||||||||
1283 | Invalidated = pointerInvalidatedByLoopWithMSSA( | ||||||||
1284 | MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(CI)), CurLoop, I, | ||||||||
1285 | *Flags); | ||||||||
1286 | if (Invalidated) | ||||||||
1287 | return false; | ||||||||
1288 | } | ||||||||
1289 | return true; | ||||||||
1290 | } | ||||||||
1291 | |||||||||
1292 | // If this call only reads from memory and there are no writes to memory | ||||||||
1293 | // in the loop, we can hoist or sink the call as appropriate. | ||||||||
1294 | if (isReadOnly(CurAST, MSSAU, CurLoop)) | ||||||||
1295 | return true; | ||||||||
1296 | } | ||||||||
1297 | |||||||||
1298 | // FIXME: This should use mod/ref information to see if we can hoist or | ||||||||
1299 | // sink the call. | ||||||||
1300 | |||||||||
1301 | return false; | ||||||||
1302 | } else if (auto *FI = dyn_cast<FenceInst>(&I)) { | ||||||||
1303 | // Fences alias (most) everything to provide ordering. For the moment, | ||||||||
1304 | // just give up if there are any other memory operations in the loop. | ||||||||
1305 | if (CurAST) { | ||||||||
1306 | auto Begin = CurAST->begin(); | ||||||||
1307 | assert(Begin != CurAST->end() && "must contain FI")((void)0); | ||||||||
1308 | if (std::next(Begin) != CurAST->end()) | ||||||||
1309 | // constant memory for instance, TODO: handle better | ||||||||
1310 | return false; | ||||||||
1311 | auto *UniqueI = Begin->getUniqueInstruction(); | ||||||||
1312 | if (!UniqueI) | ||||||||
1313 | // other memory op, give up | ||||||||
1314 | return false; | ||||||||
1315 | (void)FI; // suppress unused variable warning | ||||||||
1316 | assert(UniqueI == FI && "AS must contain FI")((void)0); | ||||||||
1317 | return true; | ||||||||
1318 | } else // MSSAU | ||||||||
1319 | return isOnlyMemoryAccess(FI, CurLoop, MSSAU); | ||||||||
1320 | } else if (auto *SI = dyn_cast<StoreInst>(&I)) { | ||||||||
1321 | if (!SI->isUnordered()) | ||||||||
1322 | return false; // Don't sink/hoist volatile or ordered atomic store! | ||||||||
1323 | |||||||||
1324 | // We can only hoist a store that we can prove writes a value which is not | ||||||||
1325 | // read or overwritten within the loop. For those cases, we fallback to | ||||||||
1326 | // load store promotion instead. TODO: We can extend this to cases where | ||||||||
1327 | // there is exactly one write to the location and that write dominates an | ||||||||
1328 | // arbitrary number of reads in the loop. | ||||||||
1329 | if (CurAST) { | ||||||||
1330 | auto &AS = CurAST->getAliasSetFor(MemoryLocation::get(SI)); | ||||||||
1331 | |||||||||
1332 | if (AS.isRef() || !AS.isMustAlias()) | ||||||||
1333 | // Quick exit test, handled by the full path below as well. | ||||||||
1334 | return false; | ||||||||
1335 | auto *UniqueI = AS.getUniqueInstruction(); | ||||||||
1336 | if (!UniqueI) | ||||||||
1337 | // other memory op, give up | ||||||||
1338 | return false; | ||||||||
1339 | assert(UniqueI == SI && "AS must contain SI")((void)0); | ||||||||
1340 | return true; | ||||||||
1341 | } else { // MSSAU | ||||||||
1342 | if (isOnlyMemoryAccess(SI, CurLoop, MSSAU)) | ||||||||
1343 | return true; | ||||||||
1344 | // If there are more accesses than the Promotion cap or no "quota" to | ||||||||
1345 | // check clobber, then give up as we're not walking a list that long. | ||||||||
1346 | if (Flags->tooManyMemoryAccesses() || Flags->tooManyClobberingCalls()) | ||||||||
1347 | return false; | ||||||||
1348 | // If there are interfering Uses (i.e. their defining access is in the | ||||||||
1349 | // loop), or ordered loads (stored as Defs!), don't move this store. | ||||||||
1350 | // Could do better here, but this is conservatively correct. | ||||||||
1351 | // TODO: Cache set of Uses on the first walk in runOnLoop, update when | ||||||||
1352 | // moving accesses. Can also extend to dominating uses. | ||||||||
1353 | auto *SIMD = MSSA->getMemoryAccess(SI); | ||||||||
1354 | for (auto *BB : CurLoop->getBlocks()) | ||||||||
1355 | if (auto *Accesses = MSSA->getBlockAccesses(BB)) { | ||||||||
1356 | for (const auto &MA : *Accesses) | ||||||||
1357 | if (const auto *MU = dyn_cast<MemoryUse>(&MA)) { | ||||||||
1358 | auto *MD = MU->getDefiningAccess(); | ||||||||
1359 | if (!MSSA->isLiveOnEntryDef(MD) && | ||||||||
1360 | CurLoop->contains(MD->getBlock())) | ||||||||
1361 | return false; | ||||||||
1362 | // Disable hoisting past potentially interfering loads. Optimized | ||||||||
1363 | // Uses may point to an access outside the loop, as getClobbering | ||||||||
1364 | // checks the previous iteration when walking the backedge. | ||||||||
1365 | // FIXME: More precise: no Uses that alias SI. | ||||||||
1366 | if (!Flags->getIsSink() && !MSSA->dominates(SIMD, MU)) | ||||||||
1367 | return false; | ||||||||
1368 | } else if (const auto *MD = dyn_cast<MemoryDef>(&MA)) { | ||||||||
1369 | if (auto *LI = dyn_cast<LoadInst>(MD->getMemoryInst())) { | ||||||||
1370 | (void)LI; // Silence warning. | ||||||||
1371 | assert(!LI->isUnordered() && "Expected unordered load")((void)0); | ||||||||
1372 | return false; | ||||||||
1373 | } | ||||||||
1374 | // Any call, while it may not be clobbering SI, it may be a use. | ||||||||
1375 | if (auto *CI = dyn_cast<CallInst>(MD->getMemoryInst())) { | ||||||||
1376 | // Check if the call may read from the memory location written | ||||||||
1377 | // to by SI. Check CI's attributes and arguments; the number of | ||||||||
1378 | // such checks performed is limited above by NoOfMemAccTooLarge. | ||||||||
1379 | ModRefInfo MRI = AA->getModRefInfo(CI, MemoryLocation::get(SI)); | ||||||||
1380 | if (isModOrRefSet(MRI)) | ||||||||
1381 | return false; | ||||||||
1382 | } | ||||||||
1383 | } | ||||||||
1384 | } | ||||||||
1385 | auto *Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(SI); | ||||||||
1386 | Flags->incrementClobberingCalls(); | ||||||||
1387 | // If there are no clobbering Defs in the loop, store is safe to hoist. | ||||||||
1388 | return MSSA->isLiveOnEntryDef(Source) || | ||||||||
1389 | !CurLoop->contains(Source->getBlock()); | ||||||||
1390 | } | ||||||||
1391 | } | ||||||||
1392 | |||||||||
1393 | assert(!I.mayReadOrWriteMemory() && "unhandled aliasing")((void)0); | ||||||||
1394 | |||||||||
1395 | // We've established mechanical ability and aliasing, it's up to the caller | ||||||||
1396 | // to check fault safety | ||||||||
1397 | return true; | ||||||||
1398 | } | ||||||||
1399 | |||||||||
1400 | /// Returns true if a PHINode is a trivially replaceable with an | ||||||||
1401 | /// Instruction. | ||||||||
1402 | /// This is true when all incoming values are that instruction. | ||||||||
1403 | /// This pattern occurs most often with LCSSA PHI nodes. | ||||||||
1404 | /// | ||||||||
1405 | static bool isTriviallyReplaceablePHI(const PHINode &PN, const Instruction &I) { | ||||||||
1406 | for (const Value *IncValue : PN.incoming_values()) | ||||||||
1407 | if (IncValue != &I) | ||||||||
1408 | return false; | ||||||||
1409 | |||||||||
1410 | return true; | ||||||||
1411 | } | ||||||||
1412 | |||||||||
1413 | /// Return true if the instruction is free in the loop. | ||||||||
1414 | static bool isFreeInLoop(const Instruction &I, const Loop *CurLoop, | ||||||||
1415 | const TargetTransformInfo *TTI) { | ||||||||
1416 | |||||||||
1417 | if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I)) { | ||||||||
1418 | if (TTI->getUserCost(GEP, TargetTransformInfo::TCK_SizeAndLatency) != | ||||||||
1419 | TargetTransformInfo::TCC_Free) | ||||||||
1420 | return false; | ||||||||
1421 | // For a GEP, we cannot simply use getUserCost because currently it | ||||||||
1422 | // optimistically assume that a GEP will fold into addressing mode | ||||||||
1423 | // regardless of its users. | ||||||||
1424 | const BasicBlock *BB = GEP->getParent(); | ||||||||
1425 | for (const User *U : GEP->users()) { | ||||||||
1426 | const Instruction *UI = cast<Instruction>(U); | ||||||||
1427 | if (CurLoop->contains(UI) && | ||||||||
1428 | (BB != UI->getParent() || | ||||||||
1429 | (!isa<StoreInst>(UI) && !isa<LoadInst>(UI)))) | ||||||||
1430 | return false; | ||||||||
1431 | } | ||||||||
1432 | return true; | ||||||||
1433 | } else | ||||||||
1434 | return TTI->getUserCost(&I, TargetTransformInfo::TCK_SizeAndLatency) == | ||||||||
1435 | TargetTransformInfo::TCC_Free; | ||||||||
1436 | } | ||||||||
1437 | |||||||||
1438 | /// Return true if the only users of this instruction are outside of | ||||||||
1439 | /// the loop. If this is true, we can sink the instruction to the exit | ||||||||
1440 | /// blocks of the loop. | ||||||||
1441 | /// | ||||||||
1442 | /// We also return true if the instruction could be folded away in lowering. | ||||||||
1443 | /// (e.g., a GEP can be folded into a load as an addressing mode in the loop). | ||||||||
1444 | static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop, | ||||||||
1445 | const LoopSafetyInfo *SafetyInfo, | ||||||||
1446 | TargetTransformInfo *TTI, bool &FreeInLoop) { | ||||||||
1447 | const auto &BlockColors = SafetyInfo->getBlockColors(); | ||||||||
1448 | bool IsFree = isFreeInLoop(I, CurLoop, TTI); | ||||||||
1449 | for (const User *U : I.users()) { | ||||||||
1450 | const Instruction *UI = cast<Instruction>(U); | ||||||||
1451 | if (const PHINode *PN = dyn_cast<PHINode>(UI)) { | ||||||||
1452 | const BasicBlock *BB = PN->getParent(); | ||||||||
1453 | // We cannot sink uses in catchswitches. | ||||||||
1454 | if (isa<CatchSwitchInst>(BB->getTerminator())) | ||||||||
1455 | return false; | ||||||||
1456 | |||||||||
1457 | // We need to sink a callsite to a unique funclet. Avoid sinking if the | ||||||||
1458 | // phi use is too muddled. | ||||||||
1459 | if (isa<CallInst>(I)) | ||||||||
1460 | if (!BlockColors.empty() && | ||||||||
1461 | BlockColors.find(const_cast<BasicBlock *>(BB))->second.size() != 1) | ||||||||
1462 | return false; | ||||||||
1463 | } | ||||||||
1464 | |||||||||
1465 | if (CurLoop->contains(UI)) { | ||||||||
1466 | if (IsFree) { | ||||||||
1467 | FreeInLoop = true; | ||||||||
1468 | continue; | ||||||||
1469 | } | ||||||||
1470 | return false; | ||||||||
1471 | } | ||||||||
1472 | } | ||||||||
1473 | return true; | ||||||||
1474 | } | ||||||||
1475 | |||||||||
1476 | static Instruction *cloneInstructionInExitBlock( | ||||||||
1477 | Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI, | ||||||||
1478 | const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU) { | ||||||||
1479 | Instruction *New; | ||||||||
1480 | if (auto *CI = dyn_cast<CallInst>(&I)) { | ||||||||
1481 | const auto &BlockColors = SafetyInfo->getBlockColors(); | ||||||||
1482 | |||||||||
1483 | // Sinking call-sites need to be handled differently from other | ||||||||
1484 | // instructions. The cloned call-site needs a funclet bundle operand | ||||||||
1485 | // appropriate for its location in the CFG. | ||||||||
1486 | SmallVector<OperandBundleDef, 1> OpBundles; | ||||||||
1487 | for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles(); | ||||||||
1488 | BundleIdx != BundleEnd; ++BundleIdx) { | ||||||||
1489 | OperandBundleUse Bundle = CI->getOperandBundleAt(BundleIdx); | ||||||||
1490 | if (Bundle.getTagID() == LLVMContext::OB_funclet) | ||||||||
1491 | continue; | ||||||||
1492 | |||||||||
1493 | OpBundles.emplace_back(Bundle); | ||||||||
1494 | } | ||||||||
1495 | |||||||||
1496 | if (!BlockColors.empty()) { | ||||||||
1497 | const ColorVector &CV = BlockColors.find(&ExitBlock)->second; | ||||||||
1498 | assert(CV.size() == 1 && "non-unique color for exit block!")((void)0); | ||||||||
1499 | BasicBlock *BBColor = CV.front(); | ||||||||
1500 | Instruction *EHPad = BBColor->getFirstNonPHI(); | ||||||||
1501 | if (EHPad->isEHPad()) | ||||||||
1502 | OpBundles.emplace_back("funclet", EHPad); | ||||||||
1503 | } | ||||||||
1504 | |||||||||
1505 | New = CallInst::Create(CI, OpBundles); | ||||||||
1506 | } else { | ||||||||
1507 | New = I.clone(); | ||||||||
1508 | } | ||||||||
1509 | |||||||||
1510 | ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New); | ||||||||
1511 | if (!I.getName().empty()) | ||||||||
1512 | New->setName(I.getName() + ".le"); | ||||||||
1513 | |||||||||
1514 | if (MSSAU && MSSAU->getMemorySSA()->getMemoryAccess(&I)) { | ||||||||
1515 | // Create a new MemoryAccess and let MemorySSA set its defining access. | ||||||||
1516 | MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( | ||||||||
1517 | New, nullptr, New->getParent(), MemorySSA::Beginning); | ||||||||
1518 | if (NewMemAcc) { | ||||||||
1519 | if (auto *MemDef = dyn_cast<MemoryDef>(NewMemAcc)) | ||||||||
1520 | MSSAU->insertDef(MemDef, /*RenameUses=*/true); | ||||||||
1521 | else { | ||||||||
1522 | auto *MemUse = cast<MemoryUse>(NewMemAcc); | ||||||||
1523 | MSSAU->insertUse(MemUse, /*RenameUses=*/true); | ||||||||
1524 | } | ||||||||
1525 | } | ||||||||
1526 | } | ||||||||
1527 | |||||||||
1528 | // Build LCSSA PHI nodes for any in-loop operands (if legal). Note that | ||||||||
1529 | // this is particularly cheap because we can rip off the PHI node that we're | ||||||||
1530 | // replacing for the number and blocks of the predecessors. | ||||||||
1531 | // OPT: If this shows up in a profile, we can instead finish sinking all | ||||||||
1532 | // invariant instructions, and then walk their operands to re-establish | ||||||||
1533 | // LCSSA. That will eliminate creating PHI nodes just to nuke them when | ||||||||
1534 | // sinking bottom-up. | ||||||||
1535 | for (Use &Op : New->operands()) | ||||||||
1536 | if (LI->wouldBeOutOfLoopUseRequiringLCSSA(Op.get(), PN.getParent())) { | ||||||||
1537 | auto *OInst = cast<Instruction>(Op.get()); | ||||||||
1538 | PHINode *OpPN = | ||||||||
1539 | PHINode::Create(OInst->getType(), PN.getNumIncomingValues(), | ||||||||
1540 | OInst->getName() + ".lcssa", &ExitBlock.front()); | ||||||||
1541 | for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) | ||||||||
1542 | OpPN->addIncoming(OInst, PN.getIncomingBlock(i)); | ||||||||
1543 | Op = OpPN; | ||||||||
1544 | } | ||||||||
1545 | return New; | ||||||||
1546 | } | ||||||||
1547 | |||||||||
1548 | static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo, | ||||||||
1549 | AliasSetTracker *AST, MemorySSAUpdater *MSSAU) { | ||||||||
1550 | if (AST) | ||||||||
1551 | AST->deleteValue(&I); | ||||||||
1552 | if (MSSAU) | ||||||||
1553 | MSSAU->removeMemoryAccess(&I); | ||||||||
1554 | SafetyInfo.removeInstruction(&I); | ||||||||
1555 | I.eraseFromParent(); | ||||||||
1556 | } | ||||||||
1557 | |||||||||
1558 | static void moveInstructionBefore(Instruction &I, Instruction &Dest, | ||||||||
1559 | ICFLoopSafetyInfo &SafetyInfo, | ||||||||
1560 | MemorySSAUpdater *MSSAU, | ||||||||
1561 | ScalarEvolution *SE) { | ||||||||
1562 | SafetyInfo.removeInstruction(&I); | ||||||||
1563 | SafetyInfo.insertInstructionTo(&I, Dest.getParent()); | ||||||||
1564 | I.moveBefore(&Dest); | ||||||||
1565 | if (MSSAU) | ||||||||
1566 | if (MemoryUseOrDef *OldMemAcc = cast_or_null<MemoryUseOrDef>( | ||||||||
1567 | MSSAU->getMemorySSA()->getMemoryAccess(&I))) | ||||||||
1568 | MSSAU->moveToPlace(OldMemAcc, Dest.getParent(), | ||||||||
1569 | MemorySSA::BeforeTerminator); | ||||||||
1570 | if (SE) | ||||||||
1571 | SE->forgetValue(&I); | ||||||||
1572 | } | ||||||||
1573 | |||||||||
1574 | static Instruction *sinkThroughTriviallyReplaceablePHI( | ||||||||
1575 | PHINode *TPN, Instruction *I, LoopInfo *LI, | ||||||||
1576 | SmallDenseMap<BasicBlock *, Instruction *, 32> &SunkCopies, | ||||||||
1577 | const LoopSafetyInfo *SafetyInfo, const Loop *CurLoop, | ||||||||
1578 | MemorySSAUpdater *MSSAU) { | ||||||||
1579 | assert(isTriviallyReplaceablePHI(*TPN, *I) &&((void)0) | ||||||||
1580 | "Expect only trivially replaceable PHI")((void)0); | ||||||||
1581 | BasicBlock *ExitBlock = TPN->getParent(); | ||||||||
1582 | Instruction *New; | ||||||||
1583 | auto It = SunkCopies.find(ExitBlock); | ||||||||
1584 | if (It != SunkCopies.end()) | ||||||||
1585 | New = It->second; | ||||||||
1586 | else | ||||||||
1587 | New = SunkCopies[ExitBlock] = cloneInstructionInExitBlock( | ||||||||
1588 | *I, *ExitBlock, *TPN, LI, SafetyInfo, MSSAU); | ||||||||
1589 | return New; | ||||||||
1590 | } | ||||||||
1591 | |||||||||
1592 | static bool canSplitPredecessors(PHINode *PN, LoopSafetyInfo *SafetyInfo) { | ||||||||
1593 | BasicBlock *BB = PN->getParent(); | ||||||||
1594 | if (!BB->canSplitPredecessors()) | ||||||||
1595 | return false; | ||||||||
1596 | // It's not impossible to split EHPad blocks, but if BlockColors already exist | ||||||||
1597 | // it require updating BlockColors for all offspring blocks accordingly. By | ||||||||
1598 | // skipping such corner case, we can make updating BlockColors after splitting | ||||||||
1599 | // predecessor fairly simple. | ||||||||
1600 | if (!SafetyInfo->getBlockColors().empty() && BB->getFirstNonPHI()->isEHPad()) | ||||||||
1601 | return false; | ||||||||
1602 | for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { | ||||||||
1603 | BasicBlock *BBPred = *PI; | ||||||||
1604 | if (isa<IndirectBrInst>(BBPred->getTerminator()) || | ||||||||
1605 | isa<CallBrInst>(BBPred->getTerminator())) | ||||||||
1606 | return false; | ||||||||
1607 | } | ||||||||
1608 | return true; | ||||||||
1609 | } | ||||||||
1610 | |||||||||
1611 | static void splitPredecessorsOfLoopExit(PHINode *PN, DominatorTree *DT, | ||||||||
1612 | LoopInfo *LI, const Loop *CurLoop, | ||||||||
1613 | LoopSafetyInfo *SafetyInfo, | ||||||||
1614 | MemorySSAUpdater *MSSAU) { | ||||||||
1615 | #ifndef NDEBUG1 | ||||||||
1616 | SmallVector<BasicBlock *, 32> ExitBlocks; | ||||||||
1617 | CurLoop->getUniqueExitBlocks(ExitBlocks); | ||||||||
1618 | SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(), | ||||||||
1619 | ExitBlocks.end()); | ||||||||
1620 | #endif | ||||||||
1621 | BasicBlock *ExitBB = PN->getParent(); | ||||||||
1622 | assert(ExitBlockSet.count(ExitBB) && "Expect the PHI is in an exit block.")((void)0); | ||||||||
1623 | |||||||||
1624 | // Split predecessors of the loop exit to make instructions in the loop are | ||||||||
1625 | // exposed to exit blocks through trivially replaceable PHIs while keeping the | ||||||||
1626 | // loop in the canonical form where each predecessor of each exit block should | ||||||||
1627 | // be contained within the loop. For example, this will convert the loop below | ||||||||
1628 | // from | ||||||||
1629 | // | ||||||||
1630 | // LB1: | ||||||||
1631 | // %v1 = | ||||||||
1632 | // br %LE, %LB2 | ||||||||
1633 | // LB2: | ||||||||
1634 | // %v2 = | ||||||||
1635 | // br %LE, %LB1 | ||||||||
1636 | // LE: | ||||||||
1637 | // %p = phi [%v1, %LB1], [%v2, %LB2] <-- non-trivially replaceable | ||||||||
1638 | // | ||||||||
1639 | // to | ||||||||
1640 | // | ||||||||
1641 | // LB1: | ||||||||
1642 | // %v1 = | ||||||||
1643 | // br %LE.split, %LB2 | ||||||||
1644 | // LB2: | ||||||||
1645 | // %v2 = | ||||||||
1646 | // br %LE.split2, %LB1 | ||||||||
1647 | // LE.split: | ||||||||
1648 | // %p1 = phi [%v1, %LB1] <-- trivially replaceable | ||||||||
1649 | // br %LE | ||||||||
1650 | // LE.split2: | ||||||||
1651 | // %p2 = phi [%v2, %LB2] <-- trivially replaceable | ||||||||
1652 | // br %LE | ||||||||
1653 | // LE: | ||||||||
1654 | // %p = phi [%p1, %LE.split], [%p2, %LE.split2] | ||||||||
1655 | // | ||||||||
1656 | const auto &BlockColors = SafetyInfo->getBlockColors(); | ||||||||
1657 | SmallSetVector<BasicBlock *, 8> PredBBs(pred_begin(ExitBB), pred_end(ExitBB)); | ||||||||
1658 | while (!PredBBs.empty()) { | ||||||||
1659 | BasicBlock *PredBB = *PredBBs.begin(); | ||||||||
1660 | assert(CurLoop->contains(PredBB) &&((void)0) | ||||||||
1661 | "Expect all predecessors are in the loop")((void)0); | ||||||||
1662 | if (PN->getBasicBlockIndex(PredBB) >= 0) { | ||||||||
1663 | BasicBlock *NewPred = SplitBlockPredecessors( | ||||||||
1664 | ExitBB, PredBB, ".split.loop.exit", DT, LI, MSSAU, true); | ||||||||
1665 | // Since we do not allow splitting EH-block with BlockColors in | ||||||||
1666 | // canSplitPredecessors(), we can simply assign predecessor's color to | ||||||||
1667 | // the new block. | ||||||||
1668 | if (!BlockColors.empty()) | ||||||||
1669 | // Grab a reference to the ColorVector to be inserted before getting the | ||||||||
1670 | // reference to the vector we are copying because inserting the new | ||||||||
1671 | // element in BlockColors might cause the map to be reallocated. | ||||||||
1672 | SafetyInfo->copyColors(NewPred, PredBB); | ||||||||
1673 | } | ||||||||
1674 | PredBBs.remove(PredBB); | ||||||||
1675 | } | ||||||||
1676 | } | ||||||||
1677 | |||||||||
1678 | /// When an instruction is found to only be used outside of the loop, this | ||||||||
1679 | /// function moves it to the exit blocks and patches up SSA form as needed. | ||||||||
1680 | /// This method is guaranteed to remove the original instruction from its | ||||||||
1681 | /// position, and may either delete it or move it to outside of the loop. | ||||||||
1682 | /// | ||||||||
1683 | static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT, | ||||||||
1684 | BlockFrequencyInfo *BFI, const Loop *CurLoop, | ||||||||
1685 | ICFLoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU, | ||||||||
1686 | OptimizationRemarkEmitter *ORE) { | ||||||||
1687 | bool Changed = false; | ||||||||
1688 | LLVM_DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n")do { } while (false); | ||||||||
1689 | |||||||||
1690 | // Iterate over users to be ready for actual sinking. Replace users via | ||||||||
1691 | // unreachable blocks with undef and make all user PHIs trivially replaceable. | ||||||||
1692 | SmallPtrSet<Instruction *, 8> VisitedUsers; | ||||||||
1693 | for (Value::user_iterator UI = I.user_begin(), UE = I.user_end(); UI != UE;) { | ||||||||
1694 | auto *User = cast<Instruction>(*UI); | ||||||||
1695 | Use &U = UI.getUse(); | ||||||||
1696 | ++UI; | ||||||||
1697 | |||||||||
1698 | if (VisitedUsers.count(User) || CurLoop->contains(User)) | ||||||||
1699 | continue; | ||||||||
1700 | |||||||||
1701 | if (!DT->isReachableFromEntry(User->getParent())) { | ||||||||
1702 | U = UndefValue::get(I.getType()); | ||||||||
1703 | Changed = true; | ||||||||
1704 | continue; | ||||||||
1705 | } | ||||||||
1706 | |||||||||
1707 | // The user must be a PHI node. | ||||||||
1708 | PHINode *PN = cast<PHINode>(User); | ||||||||
1709 | |||||||||
1710 | // Surprisingly, instructions can be used outside of loops without any | ||||||||
1711 | // exits. This can only happen in PHI nodes if the incoming block is | ||||||||
1712 | // unreachable. | ||||||||
1713 | BasicBlock *BB = PN->getIncomingBlock(U); | ||||||||
1714 | if (!DT->isReachableFromEntry(BB)) { | ||||||||
1715 | U = UndefValue::get(I.getType()); | ||||||||
1716 | Changed = true; | ||||||||
1717 | continue; | ||||||||
1718 | } | ||||||||
1719 | |||||||||
1720 | VisitedUsers.insert(PN); | ||||||||
1721 | if (isTriviallyReplaceablePHI(*PN, I)) | ||||||||
1722 | continue; | ||||||||
1723 | |||||||||
1724 | if (!canSplitPredecessors(PN, SafetyInfo)) | ||||||||
1725 | return Changed; | ||||||||
1726 | |||||||||
1727 | // Split predecessors of the PHI so that we can make users trivially | ||||||||
1728 | // replaceable. | ||||||||
1729 | splitPredecessorsOfLoopExit(PN, DT, LI, CurLoop, SafetyInfo, MSSAU); | ||||||||
1730 | |||||||||
1731 | // Should rebuild the iterators, as they may be invalidated by | ||||||||
1732 | // splitPredecessorsOfLoopExit(). | ||||||||
1733 | UI = I.user_begin(); | ||||||||
1734 | UE = I.user_end(); | ||||||||
1735 | } | ||||||||
1736 | |||||||||
1737 | if (VisitedUsers.empty()) | ||||||||
1738 | return Changed; | ||||||||
1739 | |||||||||
1740 | ORE->emit([&]() { | ||||||||
1741 | return OptimizationRemark(DEBUG_TYPE"licm", "InstSunk", &I) | ||||||||
1742 | << "sinking " << ore::NV("Inst", &I); | ||||||||
1743 | }); | ||||||||
1744 | if (isa<LoadInst>(I)) | ||||||||
1745 | ++NumMovedLoads; | ||||||||
1746 | else if (isa<CallInst>(I)) | ||||||||
1747 | ++NumMovedCalls; | ||||||||
1748 | ++NumSunk; | ||||||||
1749 | |||||||||
1750 | #ifndef NDEBUG1 | ||||||||
1751 | SmallVector<BasicBlock *, 32> ExitBlocks; | ||||||||
1752 | CurLoop->getUniqueExitBlocks(ExitBlocks); | ||||||||
1753 | SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(), | ||||||||
1754 | ExitBlocks.end()); | ||||||||
1755 | #endif | ||||||||
1756 | |||||||||
1757 | // Clones of this instruction. Don't create more than one per exit block! | ||||||||
1758 | SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies; | ||||||||
1759 | |||||||||
1760 | // If this instruction is only used outside of the loop, then all users are | ||||||||
1761 | // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of | ||||||||
1762 | // the instruction. | ||||||||
1763 | // First check if I is worth sinking for all uses. Sink only when it is worth | ||||||||
1764 | // across all uses. | ||||||||
1765 | SmallSetVector<User*, 8> Users(I.user_begin(), I.user_end()); | ||||||||
1766 | SmallVector<PHINode *, 8> ExitPNs; | ||||||||
1767 | for (auto *UI : Users) { | ||||||||
1768 | auto *User = cast<Instruction>(UI); | ||||||||
1769 | |||||||||
1770 | if (CurLoop->contains(User)) | ||||||||
1771 | continue; | ||||||||
1772 | |||||||||
1773 | PHINode *PN = cast<PHINode>(User); | ||||||||
1774 | assert(ExitBlockSet.count(PN->getParent()) &&((void)0) | ||||||||
1775 | "The LCSSA PHI is not in an exit block!")((void)0); | ||||||||
1776 | if (!worthSinkOrHoistInst(I, PN->getParent(), ORE, BFI)) { | ||||||||
1777 | return Changed; | ||||||||
1778 | } | ||||||||
1779 | |||||||||
1780 | ExitPNs.push_back(PN); | ||||||||
1781 | } | ||||||||
1782 | |||||||||
1783 | for (auto *PN : ExitPNs) { | ||||||||
1784 | |||||||||
1785 | // The PHI must be trivially replaceable. | ||||||||
1786 | Instruction *New = sinkThroughTriviallyReplaceablePHI( | ||||||||
1787 | PN, &I, LI, SunkCopies, SafetyInfo, CurLoop, MSSAU); | ||||||||
1788 | PN->replaceAllUsesWith(New); | ||||||||
1789 | eraseInstruction(*PN, *SafetyInfo, nullptr, nullptr); | ||||||||
1790 | Changed = true; | ||||||||
1791 | } | ||||||||
1792 | return Changed; | ||||||||
1793 | } | ||||||||
1794 | |||||||||
1795 | /// When an instruction is found to only use loop invariant operands that | ||||||||
1796 | /// is safe to hoist, this instruction is called to do the dirty work. | ||||||||
1797 | /// | ||||||||
1798 | static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop, | ||||||||
1799 | BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo, | ||||||||
1800 | MemorySSAUpdater *MSSAU, ScalarEvolution *SE, | ||||||||
1801 | OptimizationRemarkEmitter *ORE) { | ||||||||
1802 | LLVM_DEBUG(dbgs() << "LICM hoisting to " << Dest->getNameOrAsOperand() << ": "do { } while (false) | ||||||||
1803 | << I << "\n")do { } while (false); | ||||||||
1804 | ORE->emit([&]() { | ||||||||
1805 | return OptimizationRemark(DEBUG_TYPE"licm", "Hoisted", &I) << "hoisting " | ||||||||
1806 | << ore::NV("Inst", &I); | ||||||||
1807 | }); | ||||||||
1808 | |||||||||
1809 | // Metadata can be dependent on conditions we are hoisting above. | ||||||||
1810 | // Conservatively strip all metadata on the instruction unless we were | ||||||||
1811 | // guaranteed to execute I if we entered the loop, in which case the metadata | ||||||||
1812 | // is valid in the loop preheader. | ||||||||
1813 | // Similarly, If I is a call and it is not guaranteed to execute in the loop, | ||||||||
1814 | // then moving to the preheader means we should strip attributes on the call | ||||||||
1815 | // that can cause UB since we may be hoisting above conditions that allowed | ||||||||
1816 | // inferring those attributes. They may not be valid at the preheader. | ||||||||
1817 | if ((I.hasMetadataOtherThanDebugLoc() || isa<CallInst>(I)) && | ||||||||
1818 | // The check on hasMetadataOtherThanDebugLoc is to prevent us from burning | ||||||||
1819 | // time in isGuaranteedToExecute if we don't actually have anything to | ||||||||
1820 | // drop. It is a compile time optimization, not required for correctness. | ||||||||
1821 | !SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop)) | ||||||||
1822 | I.dropUndefImplyingAttrsAndUnknownMetadata(); | ||||||||
1823 | |||||||||
1824 | if (isa<PHINode>(I)) | ||||||||
1825 | // Move the new node to the end of the phi list in the destination block. | ||||||||
1826 | moveInstructionBefore(I, *Dest->getFirstNonPHI(), *SafetyInfo, MSSAU, SE); | ||||||||
1827 | else | ||||||||
1828 | // Move the new node to the destination block, before its terminator. | ||||||||
1829 | moveInstructionBefore(I, *Dest->getTerminator(), *SafetyInfo, MSSAU, SE); | ||||||||
1830 | |||||||||
1831 | I.updateLocationAfterHoist(); | ||||||||
1832 | |||||||||
1833 | if (isa<LoadInst>(I)) | ||||||||
1834 | ++NumMovedLoads; | ||||||||
1835 | else if (isa<CallInst>(I)) | ||||||||
1836 | ++NumMovedCalls; | ||||||||
1837 | ++NumHoisted; | ||||||||
1838 | } | ||||||||
1839 | |||||||||
1840 | /// Only sink or hoist an instruction if it is not a trapping instruction, | ||||||||
1841 | /// or if the instruction is known not to trap when moved to the preheader. | ||||||||
1842 | /// or if it is a trapping instruction and is guaranteed to execute. | ||||||||
1843 | static bool isSafeToExecuteUnconditionally(Instruction &Inst, | ||||||||
1844 | const DominatorTree *DT, | ||||||||
1845 | const TargetLibraryInfo *TLI, | ||||||||
1846 | const Loop *CurLoop, | ||||||||
1847 | const LoopSafetyInfo *SafetyInfo, | ||||||||
1848 | OptimizationRemarkEmitter *ORE, | ||||||||
1849 | const Instruction *CtxI) { | ||||||||
1850 | if (isSafeToSpeculativelyExecute(&Inst, CtxI, DT, TLI)) | ||||||||
1851 | return true; | ||||||||
1852 | |||||||||
1853 | bool GuaranteedToExecute = | ||||||||
1854 | SafetyInfo->isGuaranteedToExecute(Inst, DT, CurLoop); | ||||||||
1855 | |||||||||
1856 | if (!GuaranteedToExecute) { | ||||||||
1857 | auto *LI = dyn_cast<LoadInst>(&Inst); | ||||||||
1858 | if (LI && CurLoop->isLoopInvariant(LI->getPointerOperand())) | ||||||||
1859 | ORE->emit([&]() { | ||||||||
1860 | return OptimizationRemarkMissed( | ||||||||
1861 | DEBUG_TYPE"licm", "LoadWithLoopInvariantAddressCondExecuted", LI) | ||||||||
1862 | << "failed to hoist load with loop-invariant address " | ||||||||
1863 | "because load is conditionally executed"; | ||||||||
1864 | }); | ||||||||
1865 | } | ||||||||
1866 | |||||||||
1867 | return GuaranteedToExecute; | ||||||||
1868 | } | ||||||||
1869 | |||||||||
1870 | namespace { | ||||||||
1871 | class LoopPromoter : public LoadAndStorePromoter { | ||||||||
1872 | Value *SomePtr; // Designated pointer to store to. | ||||||||
1873 | const SmallSetVector<Value *, 8> &PointerMustAliases; | ||||||||
1874 | SmallVectorImpl<BasicBlock *> &LoopExitBlocks; | ||||||||
1875 | SmallVectorImpl<Instruction *> &LoopInsertPts; | ||||||||
1876 | SmallVectorImpl<MemoryAccess *> &MSSAInsertPts; | ||||||||
1877 | PredIteratorCache &PredCache; | ||||||||
1878 | AliasSetTracker *AST; | ||||||||
1879 | MemorySSAUpdater *MSSAU; | ||||||||
1880 | LoopInfo &LI; | ||||||||
1881 | DebugLoc DL; | ||||||||
1882 | int Alignment; | ||||||||
1883 | bool UnorderedAtomic; | ||||||||
1884 | AAMDNodes AATags; | ||||||||
1885 | ICFLoopSafetyInfo &SafetyInfo; | ||||||||
1886 | |||||||||
1887 | // We're about to add a use of V in a loop exit block. Insert an LCSSA phi | ||||||||
1888 | // (if legal) if doing so would add an out-of-loop use to an instruction | ||||||||
1889 | // defined in-loop. | ||||||||
1890 | Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const { | ||||||||
1891 | if (!LI.wouldBeOutOfLoopUseRequiringLCSSA(V, BB)) | ||||||||
1892 | return V; | ||||||||
1893 | |||||||||
1894 | Instruction *I = cast<Instruction>(V); | ||||||||
1895 | // We need to create an LCSSA PHI node for the incoming value and | ||||||||
1896 | // store that. | ||||||||
1897 | PHINode *PN = PHINode::Create(I->getType(), PredCache.size(BB), | ||||||||
1898 | I->getName() + ".lcssa", &BB->front()); | ||||||||
1899 | for (BasicBlock *Pred : PredCache.get(BB)) | ||||||||
1900 | PN->addIncoming(I, Pred); | ||||||||
1901 | return PN; | ||||||||
1902 | } | ||||||||
1903 | |||||||||
1904 | public: | ||||||||
1905 | LoopPromoter(Value *SP, ArrayRef<const Instruction *> Insts, SSAUpdater &S, | ||||||||
1906 | const SmallSetVector<Value *, 8> &PMA, | ||||||||
1907 | SmallVectorImpl<BasicBlock *> &LEB, | ||||||||
1908 | SmallVectorImpl<Instruction *> &LIP, | ||||||||
1909 | SmallVectorImpl<MemoryAccess *> &MSSAIP, PredIteratorCache &PIC, | ||||||||
1910 | AliasSetTracker *ast, MemorySSAUpdater *MSSAU, LoopInfo &li, | ||||||||
1911 | DebugLoc dl, int alignment, bool UnorderedAtomic, | ||||||||
1912 | const AAMDNodes &AATags, ICFLoopSafetyInfo &SafetyInfo) | ||||||||
1913 | : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA), | ||||||||
1914 | LoopExitBlocks(LEB), LoopInsertPts(LIP), MSSAInsertPts(MSSAIP), | ||||||||
1915 | PredCache(PIC), AST(ast), MSSAU(MSSAU), LI(li), DL(std::move(dl)), | ||||||||
1916 | Alignment(alignment), UnorderedAtomic(UnorderedAtomic), AATags(AATags), | ||||||||
1917 | SafetyInfo(SafetyInfo) {} | ||||||||
1918 | |||||||||
1919 | bool isInstInList(Instruction *I, | ||||||||
1920 | const SmallVectorImpl<Instruction *> &) const override { | ||||||||
1921 | Value *Ptr; | ||||||||
1922 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | ||||||||
1923 | Ptr = LI->getOperand(0); | ||||||||
1924 | else | ||||||||
1925 | Ptr = cast<StoreInst>(I)->getPointerOperand(); | ||||||||
1926 | return PointerMustAliases.count(Ptr); | ||||||||
1927 | } | ||||||||
1928 | |||||||||
1929 | void doExtraRewritesBeforeFinalDeletion() override { | ||||||||
1930 | // Insert stores after in the loop exit blocks. Each exit block gets a | ||||||||
1931 | // store of the live-out values that feed them. Since we've already told | ||||||||
1932 | // the SSA updater about the defs in the loop and the preheader | ||||||||
1933 | // definition, it is all set and we can start using it. | ||||||||
1934 | for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) { | ||||||||
1935 | BasicBlock *ExitBlock = LoopExitBlocks[i]; | ||||||||
1936 | Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock); | ||||||||
1937 | LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock); | ||||||||
1938 | Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock); | ||||||||
1939 | Instruction *InsertPos = LoopInsertPts[i]; | ||||||||
1940 | StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos); | ||||||||
1941 | if (UnorderedAtomic) | ||||||||
1942 | NewSI->setOrdering(AtomicOrdering::Unordered); | ||||||||
1943 | NewSI->setAlignment(Align(Alignment)); | ||||||||
1944 | NewSI->setDebugLoc(DL); | ||||||||
1945 | if (AATags) | ||||||||
1946 | NewSI->setAAMetadata(AATags); | ||||||||
1947 | |||||||||
1948 | if (MSSAU) { | ||||||||
1949 | MemoryAccess *MSSAInsertPoint = MSSAInsertPts[i]; | ||||||||
1950 | MemoryAccess *NewMemAcc; | ||||||||
1951 | if (!MSSAInsertPoint) { | ||||||||
1952 | NewMemAcc = MSSAU->createMemoryAccessInBB( | ||||||||
1953 | NewSI, nullptr, NewSI->getParent(), MemorySSA::Beginning); | ||||||||
1954 | } else { | ||||||||
1955 | NewMemAcc = | ||||||||
1956 | MSSAU->createMemoryAccessAfter(NewSI, nullptr, MSSAInsertPoint); | ||||||||
1957 | } | ||||||||
1958 | MSSAInsertPts[i] = NewMemAcc; | ||||||||
1959 | MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true); | ||||||||
1960 | // FIXME: true for safety, false may still be correct. | ||||||||
1961 | } | ||||||||
1962 | } | ||||||||
1963 | } | ||||||||
1964 | |||||||||
1965 | void replaceLoadWithValue(LoadInst *LI, Value *V) const override { | ||||||||
1966 | // Update alias analysis. | ||||||||
1967 | if (AST) | ||||||||
1968 | AST->copyValue(LI, V); | ||||||||
1969 | } | ||||||||
1970 | void instructionDeleted(Instruction *I) const override { | ||||||||
1971 | SafetyInfo.removeInstruction(I); | ||||||||
1972 | if (AST) | ||||||||
1973 | AST->deleteValue(I); | ||||||||
1974 | if (MSSAU) | ||||||||
1975 | MSSAU->removeMemoryAccess(I); | ||||||||
1976 | } | ||||||||
1977 | }; | ||||||||
1978 | |||||||||
1979 | bool isNotCapturedBeforeOrInLoop(const Value *V, const Loop *L, | ||||||||
1980 | DominatorTree *DT) { | ||||||||
1981 | // We can perform the captured-before check against any instruction in the | ||||||||
1982 | // loop header, as the loop header is reachable from any instruction inside | ||||||||
1983 | // the loop. | ||||||||
1984 | // TODO: ReturnCaptures=true shouldn't be necessary here. | ||||||||
1985 | return !PointerMayBeCapturedBefore(V, /* ReturnCaptures */ true, | ||||||||
1986 | /* StoreCaptures */ true, | ||||||||
1987 | L->getHeader()->getTerminator(), DT); | ||||||||
1988 | } | ||||||||
1989 | |||||||||
1990 | /// Return true iff we can prove that a caller of this function can not inspect | ||||||||
1991 | /// the contents of the provided object in a well defined program. | ||||||||
1992 | bool isKnownNonEscaping(Value *Object, const Loop *L, | ||||||||
1993 | const TargetLibraryInfo *TLI, DominatorTree *DT) { | ||||||||
1994 | if (isa<AllocaInst>(Object)) | ||||||||
1995 | // Since the alloca goes out of scope, we know the caller can't retain a | ||||||||
1996 | // reference to it and be well defined. Thus, we don't need to check for | ||||||||
1997 | // capture. | ||||||||
1998 | return true; | ||||||||
1999 | |||||||||
2000 | // For all other objects we need to know that the caller can't possibly | ||||||||
2001 | // have gotten a reference to the object. There are two components of | ||||||||
2002 | // that: | ||||||||
2003 | // 1) Object can't be escaped by this function. This is what | ||||||||
2004 | // PointerMayBeCaptured checks. | ||||||||
2005 | // 2) Object can't have been captured at definition site. For this, we | ||||||||
2006 | // need to know the return value is noalias. At the moment, we use a | ||||||||
2007 | // weaker condition and handle only AllocLikeFunctions (which are | ||||||||
2008 | // known to be noalias). TODO | ||||||||
2009 | return isAllocLikeFn(Object, TLI) && | ||||||||
2010 | isNotCapturedBeforeOrInLoop(Object, L, DT); | ||||||||
2011 | } | ||||||||
2012 | |||||||||
2013 | } // namespace | ||||||||
2014 | |||||||||
2015 | /// Try to promote memory values to scalars by sinking stores out of the | ||||||||
2016 | /// loop and moving loads to before the loop. We do this by looping over | ||||||||
2017 | /// the stores in the loop, looking for stores to Must pointers which are | ||||||||
2018 | /// loop invariant. | ||||||||
2019 | /// | ||||||||
2020 | bool llvm::promoteLoopAccessesToScalars( | ||||||||
2021 | const SmallSetVector<Value *, 8> &PointerMustAliases, | ||||||||
2022 | SmallVectorImpl<BasicBlock *> &ExitBlocks, | ||||||||
2023 | SmallVectorImpl<Instruction *> &InsertPts, | ||||||||
2024 | SmallVectorImpl<MemoryAccess *> &MSSAInsertPts, PredIteratorCache &PIC, | ||||||||
2025 | LoopInfo *LI, DominatorTree *DT, const TargetLibraryInfo *TLI, | ||||||||
2026 | Loop *CurLoop, AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU, | ||||||||
2027 | ICFLoopSafetyInfo *SafetyInfo, OptimizationRemarkEmitter *ORE) { | ||||||||
2028 | // Verify inputs. | ||||||||
2029 | assert(LI != nullptr && DT != nullptr && CurLoop != nullptr &&((void)0) | ||||||||
2030 | SafetyInfo != nullptr &&((void)0) | ||||||||
2031 | "Unexpected Input to promoteLoopAccessesToScalars")((void)0); | ||||||||
2032 | |||||||||
2033 | Value *SomePtr = *PointerMustAliases.begin(); | ||||||||
2034 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); | ||||||||
2035 | |||||||||
2036 | // It is not safe to promote a load/store from the loop if the load/store is | ||||||||
2037 | // conditional. For example, turning: | ||||||||
2038 | // | ||||||||
2039 | // for () { if (c) *P += 1; } | ||||||||
2040 | // | ||||||||
2041 | // into: | ||||||||
2042 | // | ||||||||
2043 | // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; | ||||||||
2044 | // | ||||||||
2045 | // is not safe, because *P may only be valid to access if 'c' is true. | ||||||||
2046 | // | ||||||||
2047 | // The safety property divides into two parts: | ||||||||
2048 | // p1) The memory may not be dereferenceable on entry to the loop. In this | ||||||||
2049 | // case, we can't insert the required load in the preheader. | ||||||||
2050 | // p2) The memory model does not allow us to insert a store along any dynamic | ||||||||
2051 | // path which did not originally have one. | ||||||||
2052 | // | ||||||||
2053 | // If at least one store is guaranteed to execute, both properties are | ||||||||
2054 | // satisfied, and promotion is legal. | ||||||||
2055 | // | ||||||||
2056 | // This, however, is not a necessary condition. Even if no store/load is | ||||||||
2057 | // guaranteed to execute, we can still establish these properties. | ||||||||
2058 | // We can establish (p1) by proving that hoisting the load into the preheader | ||||||||
2059 | // is safe (i.e. proving dereferenceability on all paths through the loop). We | ||||||||
2060 | // can use any access within the alias set to prove dereferenceability, | ||||||||
2061 | // since they're all must alias. | ||||||||
2062 | // | ||||||||
2063 | // There are two ways establish (p2): | ||||||||
2064 | // a) Prove the location is thread-local. In this case the memory model | ||||||||
2065 | // requirement does not apply, and stores are safe to insert. | ||||||||
2066 | // b) Prove a store dominates every exit block. In this case, if an exit | ||||||||
2067 | // blocks is reached, the original dynamic path would have taken us through | ||||||||
2068 | // the store, so inserting a store into the exit block is safe. Note that this | ||||||||
2069 | // is different from the store being guaranteed to execute. For instance, | ||||||||
2070 | // if an exception is thrown on the first iteration of the loop, the original | ||||||||
2071 | // store is never executed, but the exit blocks are not executed either. | ||||||||
2072 | |||||||||
2073 | bool DereferenceableInPH = false; | ||||||||
2074 | bool SafeToInsertStore = false; | ||||||||
2075 | |||||||||
2076 | SmallVector<Instruction *, 64> LoopUses; | ||||||||
2077 | |||||||||
2078 | // We start with an alignment of one and try to find instructions that allow | ||||||||
2079 | // us to prove better alignment. | ||||||||
2080 | Align Alignment; | ||||||||
2081 | // Keep track of which types of access we see | ||||||||
2082 | bool SawUnorderedAtomic = false; | ||||||||
2083 | bool SawNotAtomic = false; | ||||||||
2084 | AAMDNodes AATags; | ||||||||
2085 | |||||||||
2086 | const DataLayout &MDL = Preheader->getModule()->getDataLayout(); | ||||||||
2087 | |||||||||
2088 | bool IsKnownThreadLocalObject = false; | ||||||||
2089 | if (SafetyInfo->anyBlockMayThrow()) { | ||||||||
2090 | // If a loop can throw, we have to insert a store along each unwind edge. | ||||||||
2091 | // That said, we can't actually make the unwind edge explicit. Therefore, | ||||||||
2092 | // we have to prove that the store is dead along the unwind edge. We do | ||||||||
2093 | // this by proving that the caller can't have a reference to the object | ||||||||
2094 | // after return and thus can't possibly load from the object. | ||||||||
2095 | Value *Object = getUnderlyingObject(SomePtr); | ||||||||
2096 | if (!isKnownNonEscaping(Object, CurLoop, TLI, DT)) | ||||||||
2097 | return false; | ||||||||
2098 | // Subtlety: Alloca's aren't visible to callers, but *are* potentially | ||||||||
2099 | // visible to other threads if captured and used during their lifetimes. | ||||||||
2100 | IsKnownThreadLocalObject = !isa<AllocaInst>(Object); | ||||||||
2101 | } | ||||||||
2102 | |||||||||
2103 | // Check that all of the pointers in the alias set have the same type. We | ||||||||
2104 | // cannot (yet) promote a memory location that is loaded and stored in | ||||||||
2105 | // different sizes. While we are at it, collect alignment and AA info. | ||||||||
2106 | for (Value *ASIV : PointerMustAliases) { | ||||||||
2107 | // Check that all of the pointers in the alias set have the same type. We | ||||||||
2108 | // cannot (yet) promote a memory location that is loaded and stored in | ||||||||
2109 | // different sizes. | ||||||||
2110 | if (SomePtr->getType() != ASIV->getType()) | ||||||||
2111 | return false; | ||||||||
2112 | |||||||||
2113 | for (User *U : ASIV->users()) { | ||||||||
2114 | // Ignore instructions that are outside the loop. | ||||||||
2115 | Instruction *UI = dyn_cast<Instruction>(U); | ||||||||
2116 | if (!UI || !CurLoop->contains(UI)) | ||||||||
2117 | continue; | ||||||||
2118 | |||||||||
2119 | // If there is an non-load/store instruction in the loop, we can't promote | ||||||||
2120 | // it. | ||||||||
2121 | if (LoadInst *Load = dyn_cast<LoadInst>(UI)) { | ||||||||
2122 | if (!Load->isUnordered()) | ||||||||
2123 | return false; | ||||||||
2124 | |||||||||
2125 | SawUnorderedAtomic |= Load->isAtomic(); | ||||||||
2126 | SawNotAtomic |= !Load->isAtomic(); | ||||||||
2127 | |||||||||
2128 | Align InstAlignment = Load->getAlign(); | ||||||||
2129 | |||||||||
2130 | // Note that proving a load safe to speculate requires proving | ||||||||
2131 | // sufficient alignment at the target location. Proving it guaranteed | ||||||||
2132 | // to execute does as well. Thus we can increase our guaranteed | ||||||||
2133 | // alignment as well. | ||||||||
2134 | if (!DereferenceableInPH || (InstAlignment > Alignment)) | ||||||||
2135 | if (isSafeToExecuteUnconditionally(*Load, DT, TLI, CurLoop, | ||||||||
2136 | SafetyInfo, ORE, | ||||||||
2137 | Preheader->getTerminator())) { | ||||||||
2138 | DereferenceableInPH = true; | ||||||||
2139 | Alignment = std::max(Alignment, InstAlignment); | ||||||||
2140 | } | ||||||||
2141 | } else if (const StoreInst *Store = dyn_cast<StoreInst>(UI)) { | ||||||||
2142 | // Stores *of* the pointer are not interesting, only stores *to* the | ||||||||
2143 | // pointer. | ||||||||
2144 | if (UI->getOperand(1) != ASIV) | ||||||||
2145 | continue; | ||||||||
2146 | if (!Store->isUnordered()) | ||||||||
2147 | return false; | ||||||||
2148 | |||||||||
2149 | SawUnorderedAtomic |= Store->isAtomic(); | ||||||||
2150 | SawNotAtomic |= !Store->isAtomic(); | ||||||||
2151 | |||||||||
2152 | // If the store is guaranteed to execute, both properties are satisfied. | ||||||||
2153 | // We may want to check if a store is guaranteed to execute even if we | ||||||||
2154 | // already know that promotion is safe, since it may have higher | ||||||||
2155 | // alignment than any other guaranteed stores, in which case we can | ||||||||
2156 | // raise the alignment on the promoted store. | ||||||||
2157 | Align InstAlignment = Store->getAlign(); | ||||||||
2158 | |||||||||
2159 | if (!DereferenceableInPH || !SafeToInsertStore || | ||||||||
2160 | (InstAlignment > Alignment)) { | ||||||||
2161 | if (SafetyInfo->isGuaranteedToExecute(*UI, DT, CurLoop)) { | ||||||||
2162 | DereferenceableInPH = true; | ||||||||
2163 | SafeToInsertStore = true; | ||||||||
2164 | Alignment = std::max(Alignment, InstAlignment); | ||||||||
2165 | } | ||||||||
2166 | } | ||||||||
2167 | |||||||||
2168 | // If a store dominates all exit blocks, it is safe to sink. | ||||||||
2169 | // As explained above, if an exit block was executed, a dominating | ||||||||
2170 | // store must have been executed at least once, so we are not | ||||||||
2171 | // introducing stores on paths that did not have them. | ||||||||
2172 | // Note that this only looks at explicit exit blocks. If we ever | ||||||||
2173 | // start sinking stores into unwind edges (see above), this will break. | ||||||||
2174 | if (!SafeToInsertStore) | ||||||||
2175 | SafeToInsertStore = llvm::all_of(ExitBlocks, [&](BasicBlock *Exit) { | ||||||||
2176 | return DT->dominates(Store->getParent(), Exit); | ||||||||
2177 | }); | ||||||||
2178 | |||||||||
2179 | // If the store is not guaranteed to execute, we may still get | ||||||||
2180 | // deref info through it. | ||||||||
2181 | if (!DereferenceableInPH) { | ||||||||
2182 | DereferenceableInPH = isDereferenceableAndAlignedPointer( | ||||||||
2183 | Store->getPointerOperand(), Store->getValueOperand()->getType(), | ||||||||
2184 | Store->getAlign(), MDL, Preheader->getTerminator(), DT, TLI); | ||||||||
2185 | } | ||||||||
2186 | } else | ||||||||
2187 | return false; // Not a load or store. | ||||||||
2188 | |||||||||
2189 | // Merge the AA tags. | ||||||||
2190 | if (LoopUses.empty()) { | ||||||||
2191 | // On the first load/store, just take its AA tags. | ||||||||
2192 | UI->getAAMetadata(AATags); | ||||||||
2193 | } else if (AATags) { | ||||||||
2194 | UI->getAAMetadata(AATags, /* Merge = */ true); | ||||||||
2195 | } | ||||||||
2196 | |||||||||
2197 | LoopUses.push_back(UI); | ||||||||
2198 | } | ||||||||
2199 | } | ||||||||
2200 | |||||||||
2201 | // If we found both an unordered atomic instruction and a non-atomic memory | ||||||||
2202 | // access, bail. We can't blindly promote non-atomic to atomic since we | ||||||||
2203 | // might not be able to lower the result. We can't downgrade since that | ||||||||
2204 | // would violate memory model. Also, align 0 is an error for atomics. | ||||||||
2205 | if (SawUnorderedAtomic && SawNotAtomic) | ||||||||
2206 | return false; | ||||||||
2207 | |||||||||
2208 | // If we're inserting an atomic load in the preheader, we must be able to | ||||||||
2209 | // lower it. We're only guaranteed to be able to lower naturally aligned | ||||||||
2210 | // atomics. | ||||||||
2211 | auto *SomePtrElemType = SomePtr->getType()->getPointerElementType(); | ||||||||
2212 | if (SawUnorderedAtomic && | ||||||||
2213 | Alignment < MDL.getTypeStoreSize(SomePtrElemType)) | ||||||||
2214 | return false; | ||||||||
2215 | |||||||||
2216 | // If we couldn't prove we can hoist the load, bail. | ||||||||
2217 | if (!DereferenceableInPH) | ||||||||
2218 | return false; | ||||||||
2219 | |||||||||
2220 | // We know we can hoist the load, but don't have a guaranteed store. | ||||||||
2221 | // Check whether the location is thread-local. If it is, then we can insert | ||||||||
2222 | // stores along paths which originally didn't have them without violating the | ||||||||
2223 | // memory model. | ||||||||
2224 | if (!SafeToInsertStore) { | ||||||||
2225 | if (IsKnownThreadLocalObject) | ||||||||
2226 | SafeToInsertStore = true; | ||||||||
2227 | else { | ||||||||
2228 | Value *Object = getUnderlyingObject(SomePtr); | ||||||||
2229 | SafeToInsertStore = | ||||||||
2230 | (isAllocLikeFn(Object, TLI) || isa<AllocaInst>(Object)) && | ||||||||
2231 | isNotCapturedBeforeOrInLoop(Object, CurLoop, DT); | ||||||||
2232 | } | ||||||||
2233 | } | ||||||||
2234 | |||||||||
2235 | // If we've still failed to prove we can sink the store, give up. | ||||||||
2236 | if (!SafeToInsertStore) | ||||||||
2237 | return false; | ||||||||
2238 | |||||||||
2239 | // Otherwise, this is safe to promote, lets do it! | ||||||||
2240 | LLVM_DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " << *SomePtrdo { } while (false) | ||||||||
2241 | << '\n')do { } while (false); | ||||||||
2242 | ORE->emit([&]() { | ||||||||
2243 | return OptimizationRemark(DEBUG_TYPE"licm", "PromoteLoopAccessesToScalar", | ||||||||
2244 | LoopUses[0]) | ||||||||
2245 | << "Moving accesses to memory location out of the loop"; | ||||||||
2246 | }); | ||||||||
2247 | ++NumPromoted; | ||||||||
2248 | |||||||||
2249 | // Look at all the loop uses, and try to merge their locations. | ||||||||
2250 | std::vector<const DILocation *> LoopUsesLocs; | ||||||||
2251 | for (auto U : LoopUses) | ||||||||
2252 | LoopUsesLocs.push_back(U->getDebugLoc().get()); | ||||||||
2253 | auto DL = DebugLoc(DILocation::getMergedLocations(LoopUsesLocs)); | ||||||||
2254 | |||||||||
2255 | // We use the SSAUpdater interface to insert phi nodes as required. | ||||||||
2256 | SmallVector<PHINode *, 16> NewPHIs; | ||||||||
2257 | SSAUpdater SSA(&NewPHIs); | ||||||||
2258 | LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks, | ||||||||
2259 | InsertPts, MSSAInsertPts, PIC, CurAST, MSSAU, *LI, DL, | ||||||||
2260 | Alignment.value(), SawUnorderedAtomic, AATags, | ||||||||
2261 | *SafetyInfo); | ||||||||
2262 | |||||||||
2263 | // Set up the preheader to have a definition of the value. It is the live-out | ||||||||
2264 | // value from the preheader that uses in the loop will use. | ||||||||
2265 | LoadInst *PreheaderLoad = new LoadInst( | ||||||||
2266 | SomePtr->getType()->getPointerElementType(), SomePtr, | ||||||||
2267 | SomePtr->getName() + ".promoted", Preheader->getTerminator()); | ||||||||
2268 | if (SawUnorderedAtomic) | ||||||||
2269 | PreheaderLoad->setOrdering(AtomicOrdering::Unordered); | ||||||||
2270 | PreheaderLoad->setAlignment(Alignment); | ||||||||
2271 | PreheaderLoad->setDebugLoc(DebugLoc()); | ||||||||
2272 | if (AATags) | ||||||||
2273 | PreheaderLoad->setAAMetadata(AATags); | ||||||||
2274 | SSA.AddAvailableValue(Preheader, PreheaderLoad); | ||||||||
2275 | |||||||||
2276 | if (MSSAU) { | ||||||||
2277 | MemoryAccess *PreheaderLoadMemoryAccess = MSSAU->createMemoryAccessInBB( | ||||||||
2278 | PreheaderLoad, nullptr, PreheaderLoad->getParent(), MemorySSA::End); | ||||||||
2279 | MemoryUse *NewMemUse = cast<MemoryUse>(PreheaderLoadMemoryAccess); | ||||||||
2280 | MSSAU->insertUse(NewMemUse, /*RenameUses=*/true); | ||||||||
2281 | } | ||||||||
2282 | |||||||||
2283 | if (MSSAU && VerifyMemorySSA) | ||||||||
2284 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||
2285 | // Rewrite all the loads in the loop and remember all the definitions from | ||||||||
2286 | // stores in the loop. | ||||||||
2287 | Promoter.run(LoopUses); | ||||||||
2288 | |||||||||
2289 | if (MSSAU && VerifyMemorySSA) | ||||||||
2290 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||
2291 | // If the SSAUpdater didn't use the load in the preheader, just zap it now. | ||||||||
2292 | if (PreheaderLoad->use_empty()) | ||||||||
2293 | eraseInstruction(*PreheaderLoad, *SafetyInfo, CurAST, MSSAU); | ||||||||
2294 | |||||||||
2295 | return true; | ||||||||
2296 | } | ||||||||
2297 | |||||||||
2298 | static void foreachMemoryAccess(MemorySSA *MSSA, Loop *L, | ||||||||
2299 | function_ref<void(Instruction *)> Fn) { | ||||||||
2300 | for (const BasicBlock *BB : L->blocks()) | ||||||||
2301 | if (const auto *Accesses = MSSA->getBlockAccesses(BB)) | ||||||||
2302 | for (const auto &Access : *Accesses) | ||||||||
2303 | if (const auto *MUD = dyn_cast<MemoryUseOrDef>(&Access)) | ||||||||
2304 | Fn(MUD->getMemoryInst()); | ||||||||
2305 | } | ||||||||
2306 | |||||||||
2307 | static SmallVector<SmallSetVector<Value *, 8>, 0> | ||||||||
2308 | collectPromotionCandidates(MemorySSA *MSSA, AliasAnalysis *AA, Loop *L) { | ||||||||
2309 | AliasSetTracker AST(*AA); | ||||||||
2310 | |||||||||
2311 | auto IsPotentiallyPromotable = [L](const Instruction *I) { | ||||||||
2312 | if (const auto *SI = dyn_cast<StoreInst>(I)) | ||||||||
2313 | return L->isLoopInvariant(SI->getPointerOperand()); | ||||||||
2314 | if (const auto *LI = dyn_cast<LoadInst>(I)) | ||||||||
2315 | return L->isLoopInvariant(LI->getPointerOperand()); | ||||||||
2316 | return false; | ||||||||
2317 | }; | ||||||||
2318 | |||||||||
2319 | // Populate AST with potentially promotable accesses and remove them from | ||||||||
2320 | // MaybePromotable, so they will not be checked again on the next iteration. | ||||||||
2321 | SmallPtrSet<Value *, 16> AttemptingPromotion; | ||||||||
2322 | foreachMemoryAccess(MSSA, L, [&](Instruction *I) { | ||||||||
2323 | if (IsPotentiallyPromotable(I)) { | ||||||||
2324 | AttemptingPromotion.insert(I); | ||||||||
2325 | AST.add(I); | ||||||||
2326 | } | ||||||||
2327 | }); | ||||||||
2328 | |||||||||
2329 | // We're only interested in must-alias sets that contain a mod. | ||||||||
2330 | SmallVector<const AliasSet *, 8> Sets; | ||||||||
2331 | for (AliasSet &AS : AST) | ||||||||
2332 | if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias()) | ||||||||
2333 | Sets.push_back(&AS); | ||||||||
2334 | |||||||||
2335 | if (Sets.empty()) | ||||||||
2336 | return {}; // Nothing to promote... | ||||||||
2337 | |||||||||
2338 | // Discard any sets for which there is an aliasing non-promotable access. | ||||||||
2339 | foreachMemoryAccess(MSSA, L, [&](Instruction *I) { | ||||||||
2340 | if (AttemptingPromotion.contains(I)) | ||||||||
2341 | return; | ||||||||
2342 | |||||||||
2343 | llvm::erase_if(Sets, [&](const AliasSet *AS) { | ||||||||
2344 | return AS->aliasesUnknownInst(I, *AA); | ||||||||
2345 | }); | ||||||||
2346 | }); | ||||||||
2347 | |||||||||
2348 | SmallVector<SmallSetVector<Value *, 8>, 0> Result; | ||||||||
2349 | for (const AliasSet *Set : Sets) { | ||||||||
2350 | SmallSetVector<Value *, 8> PointerMustAliases; | ||||||||
2351 | for (const auto &ASI : *Set) | ||||||||
2352 | PointerMustAliases.insert(ASI.getValue()); | ||||||||
2353 | Result.push_back(std::move(PointerMustAliases)); | ||||||||
2354 | } | ||||||||
2355 | |||||||||
2356 | return Result; | ||||||||
2357 | } | ||||||||
2358 | |||||||||
2359 | /// Returns an owning pointer to an alias set which incorporates aliasing info | ||||||||
2360 | /// from L and all subloops of L. | ||||||||
2361 | std::unique_ptr<AliasSetTracker> | ||||||||
2362 | LoopInvariantCodeMotion::collectAliasInfoForLoop(Loop *L, LoopInfo *LI, | ||||||||
2363 | AAResults *AA) { | ||||||||
2364 | auto CurAST = std::make_unique<AliasSetTracker>(*AA); | ||||||||
2365 | |||||||||
2366 | // Add everything from all the sub loops. | ||||||||
2367 | for (Loop *InnerL : L->getSubLoops()) | ||||||||
2368 | for (BasicBlock *BB : InnerL->blocks()) | ||||||||
2369 | CurAST->add(*BB); | ||||||||
2370 | |||||||||
2371 | // And merge in this loop (without anything from inner loops). | ||||||||
2372 | for (BasicBlock *BB : L->blocks()) | ||||||||
2373 | if (LI->getLoopFor(BB) == L) | ||||||||
2374 | CurAST->add(*BB); | ||||||||
2375 | |||||||||
2376 | return CurAST; | ||||||||
2377 | } | ||||||||
2378 | |||||||||
2379 | static bool pointerInvalidatedByLoop(MemoryLocation MemLoc, | ||||||||
2380 | AliasSetTracker *CurAST, Loop *CurLoop, | ||||||||
2381 | AAResults *AA) { | ||||||||
2382 | // First check to see if any of the basic blocks in CurLoop invalidate *V. | ||||||||
2383 | bool isInvalidatedAccordingToAST = CurAST->getAliasSetFor(MemLoc).isMod(); | ||||||||
2384 | |||||||||
2385 | if (!isInvalidatedAccordingToAST || !LICMN2Theshold) | ||||||||
2386 | return isInvalidatedAccordingToAST; | ||||||||
2387 | |||||||||
2388 | // Check with a diagnostic analysis if we can refine the information above. | ||||||||
2389 | // This is to identify the limitations of using the AST. | ||||||||
2390 | // The alias set mechanism used by LICM has a major weakness in that it | ||||||||
2391 | // combines all things which may alias into a single set *before* asking | ||||||||
2392 | // modref questions. As a result, a single readonly call within a loop will | ||||||||
2393 | // collapse all loads and stores into a single alias set and report | ||||||||
2394 | // invalidation if the loop contains any store. For example, readonly calls | ||||||||
2395 | // with deopt states have this form and create a general alias set with all | ||||||||
2396 | // loads and stores. In order to get any LICM in loops containing possible | ||||||||
2397 | // deopt states we need a more precise invalidation of checking the mod ref | ||||||||
2398 | // info of each instruction within the loop and LI. This has a complexity of | ||||||||
2399 | // O(N^2), so currently, it is used only as a diagnostic tool since the | ||||||||
2400 | // default value of LICMN2Threshold is zero. | ||||||||
2401 | |||||||||
2402 | // Don't look at nested loops. | ||||||||
2403 | if (CurLoop->begin() != CurLoop->end()) | ||||||||
2404 | return true; | ||||||||
2405 | |||||||||
2406 | int N = 0; | ||||||||
2407 | for (BasicBlock *BB : CurLoop->getBlocks()) | ||||||||
2408 | for (Instruction &I : *BB) { | ||||||||
2409 | if (N >= LICMN2Theshold) { | ||||||||
2410 | LLVM_DEBUG(dbgs() << "Alasing N2 threshold exhausted for "do { } while (false) | ||||||||
2411 | << *(MemLoc.Ptr) << "\n")do { } while (false); | ||||||||
2412 | return true; | ||||||||
2413 | } | ||||||||
2414 | N++; | ||||||||
2415 | auto Res = AA->getModRefInfo(&I, MemLoc); | ||||||||
2416 | if (isModSet(Res)) { | ||||||||
2417 | LLVM_DEBUG(dbgs() << "Aliasing failed on " << I << " for "do { } while (false) | ||||||||
2418 | << *(MemLoc.Ptr) << "\n")do { } while (false); | ||||||||
2419 | return true; | ||||||||
2420 | } | ||||||||
2421 | } | ||||||||
2422 | LLVM_DEBUG(dbgs() << "Aliasing okay for " << *(MemLoc.Ptr) << "\n")do { } while (false); | ||||||||
2423 | return false; | ||||||||
2424 | } | ||||||||
2425 | |||||||||
2426 | bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU, | ||||||||
2427 | Loop *CurLoop, Instruction &I, | ||||||||
2428 | SinkAndHoistLICMFlags &Flags) { | ||||||||
2429 | // For hoisting, use the walker to determine safety | ||||||||
2430 | if (!Flags.getIsSink()) { | ||||||||
2431 | MemoryAccess *Source; | ||||||||
2432 | // See declaration of SetLicmMssaOptCap for usage details. | ||||||||
2433 | if (Flags.tooManyClobberingCalls()) | ||||||||
2434 | Source = MU->getDefiningAccess(); | ||||||||
2435 | else { | ||||||||
2436 | Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(MU); | ||||||||
2437 | Flags.incrementClobberingCalls(); | ||||||||
2438 | } | ||||||||
2439 | return !MSSA->isLiveOnEntryDef(Source) && | ||||||||
2440 | CurLoop->contains(Source->getBlock()); | ||||||||
2441 | } | ||||||||
2442 | |||||||||
2443 | // For sinking, we'd need to check all Defs below this use. The getClobbering | ||||||||
2444 | // call will look on the backedge of the loop, but will check aliasing with | ||||||||
2445 | // the instructions on the previous iteration. | ||||||||
2446 | // For example: | ||||||||
2447 | // for (i ... ) | ||||||||
2448 | // load a[i] ( Use (LoE) | ||||||||
2449 | // store a[i] ( 1 = Def (2), with 2 = Phi for the loop. | ||||||||
2450 | // i++; | ||||||||
2451 | // The load sees no clobbering inside the loop, as the backedge alias check | ||||||||
2452 | // does phi translation, and will check aliasing against store a[i-1]. | ||||||||
2453 | // However sinking the load outside the loop, below the store is incorrect. | ||||||||
2454 | |||||||||
2455 | // For now, only sink if there are no Defs in the loop, and the existing ones | ||||||||
2456 | // precede the use and are in the same block. | ||||||||
2457 | // FIXME: Increase precision: Safe to sink if Use post dominates the Def; | ||||||||
2458 | // needs PostDominatorTreeAnalysis. | ||||||||
2459 | // FIXME: More precise: no Defs that alias this Use. | ||||||||
2460 | if (Flags.tooManyMemoryAccesses()) | ||||||||
2461 | return true; | ||||||||
2462 | for (auto *BB : CurLoop->getBlocks()) | ||||||||
2463 | if (pointerInvalidatedByBlockWithMSSA(*BB, *MSSA, *MU)) | ||||||||
2464 | return true; | ||||||||
2465 | // When sinking, the source block may not be part of the loop so check it. | ||||||||
2466 | if (!CurLoop->contains(&I)) | ||||||||
2467 | return pointerInvalidatedByBlockWithMSSA(*I.getParent(), *MSSA, *MU); | ||||||||
2468 | |||||||||
2469 | return false; | ||||||||
2470 | } | ||||||||
2471 | |||||||||
2472 | bool pointerInvalidatedByBlockWithMSSA(BasicBlock &BB, MemorySSA &MSSA, | ||||||||
2473 | MemoryUse &MU) { | ||||||||
2474 | if (const auto *Accesses = MSSA.getBlockDefs(&BB)) | ||||||||
2475 | for (const auto &MA : *Accesses) | ||||||||
2476 | if (const auto *MD = dyn_cast<MemoryDef>(&MA)) | ||||||||
2477 | if (MU.getBlock() != MD->getBlock() || !MSSA.locallyDominates(MD, &MU)) | ||||||||
2478 | return true; | ||||||||
2479 | return false; | ||||||||
2480 | } | ||||||||
2481 | |||||||||
2482 | /// Little predicate that returns true if the specified basic block is in | ||||||||
2483 | /// a subloop of the current one, not the current one itself. | ||||||||
2484 | /// | ||||||||
2485 | static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) { | ||||||||
2486 | assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop")((void)0); | ||||||||
2487 | return LI->getLoopFor(BB) != CurLoop; | ||||||||
2488 | } |
1 | //===- llvm/Instructions.h - Instruction subclass definitions ---*- C++ -*-===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file exposes the class definitions of all of the subclasses of the |
10 | // Instruction class. This is meant to be an easy way to get access to all |
11 | // instruction subclasses. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #ifndef LLVM_IR_INSTRUCTIONS_H |
16 | #define LLVM_IR_INSTRUCTIONS_H |
17 | |
18 | #include "llvm/ADT/ArrayRef.h" |
19 | #include "llvm/ADT/Bitfields.h" |
20 | #include "llvm/ADT/MapVector.h" |
21 | #include "llvm/ADT/None.h" |
22 | #include "llvm/ADT/STLExtras.h" |
23 | #include "llvm/ADT/SmallVector.h" |
24 | #include "llvm/ADT/StringRef.h" |
25 | #include "llvm/ADT/Twine.h" |
26 | #include "llvm/ADT/iterator.h" |
27 | #include "llvm/ADT/iterator_range.h" |
28 | #include "llvm/IR/Attributes.h" |
29 | #include "llvm/IR/BasicBlock.h" |
30 | #include "llvm/IR/CallingConv.h" |
31 | #include "llvm/IR/CFG.h" |
32 | #include "llvm/IR/Constant.h" |
33 | #include "llvm/IR/DerivedTypes.h" |
34 | #include "llvm/IR/Function.h" |
35 | #include "llvm/IR/InstrTypes.h" |
36 | #include "llvm/IR/Instruction.h" |
37 | #include "llvm/IR/OperandTraits.h" |
38 | #include "llvm/IR/Type.h" |
39 | #include "llvm/IR/Use.h" |
40 | #include "llvm/IR/User.h" |
41 | #include "llvm/IR/Value.h" |
42 | #include "llvm/Support/AtomicOrdering.h" |
43 | #include "llvm/Support/Casting.h" |
44 | #include "llvm/Support/ErrorHandling.h" |
45 | #include <cassert> |
46 | #include <cstddef> |
47 | #include <cstdint> |
48 | #include <iterator> |
49 | |
50 | namespace llvm { |
51 | |
52 | class APInt; |
53 | class ConstantInt; |
54 | class DataLayout; |
55 | class LLVMContext; |
56 | |
57 | //===----------------------------------------------------------------------===// |
58 | // AllocaInst Class |
59 | //===----------------------------------------------------------------------===// |
60 | |
61 | /// an instruction to allocate memory on the stack |
62 | class AllocaInst : public UnaryInstruction { |
63 | Type *AllocatedType; |
64 | |
65 | using AlignmentField = AlignmentBitfieldElementT<0>; |
66 | using UsedWithInAllocaField = BoolBitfieldElementT<AlignmentField::NextBit>; |
67 | using SwiftErrorField = BoolBitfieldElementT<UsedWithInAllocaField::NextBit>; |
68 | static_assert(Bitfield::areContiguous<AlignmentField, UsedWithInAllocaField, |
69 | SwiftErrorField>(), |
70 | "Bitfields must be contiguous"); |
71 | |
72 | protected: |
73 | // Note: Instruction needs to be a friend here to call cloneImpl. |
74 | friend class Instruction; |
75 | |
76 | AllocaInst *cloneImpl() const; |
77 | |
78 | public: |
79 | explicit AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, |
80 | const Twine &Name, Instruction *InsertBefore); |
81 | AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, |
82 | const Twine &Name, BasicBlock *InsertAtEnd); |
83 | |
84 | AllocaInst(Type *Ty, unsigned AddrSpace, const Twine &Name, |
85 | Instruction *InsertBefore); |
86 | AllocaInst(Type *Ty, unsigned AddrSpace, |
87 | const Twine &Name, BasicBlock *InsertAtEnd); |
88 | |
89 | AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align, |
90 | const Twine &Name = "", Instruction *InsertBefore = nullptr); |
91 | AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align, |
92 | const Twine &Name, BasicBlock *InsertAtEnd); |
93 | |
94 | /// Return true if there is an allocation size parameter to the allocation |
95 | /// instruction that is not 1. |
96 | bool isArrayAllocation() const; |
97 | |
98 | /// Get the number of elements allocated. For a simple allocation of a single |
99 | /// element, this will return a constant 1 value. |
100 | const Value *getArraySize() const { return getOperand(0); } |
101 | Value *getArraySize() { return getOperand(0); } |
102 | |
103 | /// Overload to return most specific pointer type. |
104 | PointerType *getType() const { |
105 | return cast<PointerType>(Instruction::getType()); |
106 | } |
107 | |
108 | /// Get allocation size in bits. Returns None if size can't be determined, |
109 | /// e.g. in case of a VLA. |
110 | Optional<TypeSize> getAllocationSizeInBits(const DataLayout &DL) const; |
111 | |
112 | /// Return the type that is being allocated by the instruction. |
113 | Type *getAllocatedType() const { return AllocatedType; } |
114 | /// for use only in special circumstances that need to generically |
115 | /// transform a whole instruction (eg: IR linking and vectorization). |
116 | void setAllocatedType(Type *Ty) { AllocatedType = Ty; } |
117 | |
118 | /// Return the alignment of the memory that is being allocated by the |
119 | /// instruction. |
120 | Align getAlign() const { |
121 | return Align(1ULL << getSubclassData<AlignmentField>()); |
122 | } |
123 | |
124 | void setAlignment(Align Align) { |
125 | setSubclassData<AlignmentField>(Log2(Align)); |
126 | } |
127 | |
128 | // FIXME: Remove this one transition to Align is over. |
129 | unsigned getAlignment() const { return getAlign().value(); } |
130 | |
131 | /// Return true if this alloca is in the entry block of the function and is a |
132 | /// constant size. If so, the code generator will fold it into the |
133 | /// prolog/epilog code, so it is basically free. |
134 | bool isStaticAlloca() const; |
135 | |
136 | /// Return true if this alloca is used as an inalloca argument to a call. Such |
137 | /// allocas are never considered static even if they are in the entry block. |
138 | bool isUsedWithInAlloca() const { |
139 | return getSubclassData<UsedWithInAllocaField>(); |
140 | } |
141 | |
142 | /// Specify whether this alloca is used to represent the arguments to a call. |
143 | void setUsedWithInAlloca(bool V) { |
144 | setSubclassData<UsedWithInAllocaField>(V); |
145 | } |
146 | |
147 | /// Return true if this alloca is used as a swifterror argument to a call. |
148 | bool isSwiftError() const { return getSubclassData<SwiftErrorField>(); } |
149 | /// Specify whether this alloca is used to represent a swifterror. |
150 | void setSwiftError(bool V) { setSubclassData<SwiftErrorField>(V); } |
151 | |
152 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
153 | static bool classof(const Instruction *I) { |
154 | return (I->getOpcode() == Instruction::Alloca); |
155 | } |
156 | static bool classof(const Value *V) { |
157 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
158 | } |
159 | |
160 | private: |
161 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
162 | // method so that subclasses cannot accidentally use it. |
163 | template <typename Bitfield> |
164 | void setSubclassData(typename Bitfield::Type Value) { |
165 | Instruction::setSubclassData<Bitfield>(Value); |
166 | } |
167 | }; |
168 | |
169 | //===----------------------------------------------------------------------===// |
170 | // LoadInst Class |
171 | //===----------------------------------------------------------------------===// |
172 | |
173 | /// An instruction for reading from memory. This uses the SubclassData field in |
174 | /// Value to store whether or not the load is volatile. |
175 | class LoadInst : public UnaryInstruction { |
176 | using VolatileField = BoolBitfieldElementT<0>; |
177 | using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>; |
178 | using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>; |
179 | static_assert( |
180 | Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(), |
181 | "Bitfields must be contiguous"); |
182 | |
183 | void AssertOK(); |
184 | |
185 | protected: |
186 | // Note: Instruction needs to be a friend here to call cloneImpl. |
187 | friend class Instruction; |
188 | |
189 | LoadInst *cloneImpl() const; |
190 | |
191 | public: |
192 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, |
193 | Instruction *InsertBefore); |
194 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, BasicBlock *InsertAtEnd); |
195 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
196 | Instruction *InsertBefore); |
197 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
198 | BasicBlock *InsertAtEnd); |
199 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
200 | Align Align, Instruction *InsertBefore = nullptr); |
201 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
202 | Align Align, BasicBlock *InsertAtEnd); |
203 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
204 | Align Align, AtomicOrdering Order, |
205 | SyncScope::ID SSID = SyncScope::System, |
206 | Instruction *InsertBefore = nullptr); |
207 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
208 | Align Align, AtomicOrdering Order, SyncScope::ID SSID, |
209 | BasicBlock *InsertAtEnd); |
210 | |
211 | /// Return true if this is a load from a volatile memory location. |
212 | bool isVolatile() const { return getSubclassData<VolatileField>(); } |
213 | |
214 | /// Specify whether this is a volatile load or not. |
215 | void setVolatile(bool V) { setSubclassData<VolatileField>(V); } |
216 | |
217 | /// Return the alignment of the access that is being performed. |
218 | /// FIXME: Remove this function once transition to Align is over. |
219 | /// Use getAlign() instead. |
220 | unsigned getAlignment() const { return getAlign().value(); } |
221 | |
222 | /// Return the alignment of the access that is being performed. |
223 | Align getAlign() const { |
224 | return Align(1ULL << (getSubclassData<AlignmentField>())); |
225 | } |
226 | |
227 | void setAlignment(Align Align) { |
228 | setSubclassData<AlignmentField>(Log2(Align)); |
229 | } |
230 | |
231 | /// Returns the ordering constraint of this load instruction. |
232 | AtomicOrdering getOrdering() const { |
233 | return getSubclassData<OrderingField>(); |
234 | } |
235 | /// Sets the ordering constraint of this load instruction. May not be Release |
236 | /// or AcquireRelease. |
237 | void setOrdering(AtomicOrdering Ordering) { |
238 | setSubclassData<OrderingField>(Ordering); |
239 | } |
240 | |
241 | /// Returns the synchronization scope ID of this load instruction. |
242 | SyncScope::ID getSyncScopeID() const { |
243 | return SSID; |
244 | } |
245 | |
246 | /// Sets the synchronization scope ID of this load instruction. |
247 | void setSyncScopeID(SyncScope::ID SSID) { |
248 | this->SSID = SSID; |
249 | } |
250 | |
251 | /// Sets the ordering constraint and the synchronization scope ID of this load |
252 | /// instruction. |
253 | void setAtomic(AtomicOrdering Ordering, |
254 | SyncScope::ID SSID = SyncScope::System) { |
255 | setOrdering(Ordering); |
256 | setSyncScopeID(SSID); |
257 | } |
258 | |
259 | bool isSimple() const { return !isAtomic() && !isVolatile(); } |
260 | |
261 | bool isUnordered() const { |
262 | return (getOrdering() == AtomicOrdering::NotAtomic || |
263 | getOrdering() == AtomicOrdering::Unordered) && |
264 | !isVolatile(); |
265 | } |
266 | |
267 | Value *getPointerOperand() { return getOperand(0); } |
268 | const Value *getPointerOperand() const { return getOperand(0); } |
269 | static unsigned getPointerOperandIndex() { return 0U; } |
270 | Type *getPointerOperandType() const { return getPointerOperand()->getType(); } |
271 | |
272 | /// Returns the address space of the pointer operand. |
273 | unsigned getPointerAddressSpace() const { |
274 | return getPointerOperandType()->getPointerAddressSpace(); |
275 | } |
276 | |
277 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
278 | static bool classof(const Instruction *I) { |
279 | return I->getOpcode() == Instruction::Load; |
280 | } |
281 | static bool classof(const Value *V) { |
282 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
283 | } |
284 | |
285 | private: |
286 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
287 | // method so that subclasses cannot accidentally use it. |
288 | template <typename Bitfield> |
289 | void setSubclassData(typename Bitfield::Type Value) { |
290 | Instruction::setSubclassData<Bitfield>(Value); |
291 | } |
292 | |
293 | /// The synchronization scope ID of this load instruction. Not quite enough |
294 | /// room in SubClassData for everything, so synchronization scope ID gets its |
295 | /// own field. |
296 | SyncScope::ID SSID; |
297 | }; |
298 | |
299 | //===----------------------------------------------------------------------===// |
300 | // StoreInst Class |
301 | //===----------------------------------------------------------------------===// |
302 | |
303 | /// An instruction for storing to memory. |
304 | class StoreInst : public Instruction { |
305 | using VolatileField = BoolBitfieldElementT<0>; |
306 | using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>; |
307 | using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>; |
308 | static_assert( |
309 | Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(), |
310 | "Bitfields must be contiguous"); |
311 | |
312 | void AssertOK(); |
313 | |
314 | protected: |
315 | // Note: Instruction needs to be a friend here to call cloneImpl. |
316 | friend class Instruction; |
317 | |
318 | StoreInst *cloneImpl() const; |
319 | |
320 | public: |
321 | StoreInst(Value *Val, Value *Ptr, Instruction *InsertBefore); |
322 | StoreInst(Value *Val, Value *Ptr, BasicBlock *InsertAtEnd); |
323 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Instruction *InsertBefore); |
324 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, BasicBlock *InsertAtEnd); |
325 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
326 | Instruction *InsertBefore = nullptr); |
327 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
328 | BasicBlock *InsertAtEnd); |
329 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
330 | AtomicOrdering Order, SyncScope::ID SSID = SyncScope::System, |
331 | Instruction *InsertBefore = nullptr); |
332 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
333 | AtomicOrdering Order, SyncScope::ID SSID, BasicBlock *InsertAtEnd); |
334 | |
335 | // allocate space for exactly two operands |
336 | void *operator new(size_t S) { return User::operator new(S, 2); } |
337 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
338 | |
339 | /// Return true if this is a store to a volatile memory location. |
340 | bool isVolatile() const { return getSubclassData<VolatileField>(); } |
341 | |
342 | /// Specify whether this is a volatile store or not. |
343 | void setVolatile(bool V) { setSubclassData<VolatileField>(V); } |
344 | |
345 | /// Transparently provide more efficient getOperand methods. |
346 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
347 | |
348 | /// Return the alignment of the access that is being performed |
349 | /// FIXME: Remove this function once transition to Align is over. |
350 | /// Use getAlign() instead. |
351 | unsigned getAlignment() const { return getAlign().value(); } |
352 | |
353 | Align getAlign() const { |
354 | return Align(1ULL << (getSubclassData<AlignmentField>())); |
355 | } |
356 | |
357 | void setAlignment(Align Align) { |
358 | setSubclassData<AlignmentField>(Log2(Align)); |
359 | } |
360 | |
361 | /// Returns the ordering constraint of this store instruction. |
362 | AtomicOrdering getOrdering() const { |
363 | return getSubclassData<OrderingField>(); |
364 | } |
365 | |
366 | /// Sets the ordering constraint of this store instruction. May not be |
367 | /// Acquire or AcquireRelease. |
368 | void setOrdering(AtomicOrdering Ordering) { |
369 | setSubclassData<OrderingField>(Ordering); |
370 | } |
371 | |
372 | /// Returns the synchronization scope ID of this store instruction. |
373 | SyncScope::ID getSyncScopeID() const { |
374 | return SSID; |
375 | } |
376 | |
377 | /// Sets the synchronization scope ID of this store instruction. |
378 | void setSyncScopeID(SyncScope::ID SSID) { |
379 | this->SSID = SSID; |
380 | } |
381 | |
382 | /// Sets the ordering constraint and the synchronization scope ID of this |
383 | /// store instruction. |
384 | void setAtomic(AtomicOrdering Ordering, |
385 | SyncScope::ID SSID = SyncScope::System) { |
386 | setOrdering(Ordering); |
387 | setSyncScopeID(SSID); |
388 | } |
389 | |
390 | bool isSimple() const { return !isAtomic() && !isVolatile(); } |
391 | |
392 | bool isUnordered() const { |
393 | return (getOrdering() == AtomicOrdering::NotAtomic || |
394 | getOrdering() == AtomicOrdering::Unordered) && |
395 | !isVolatile(); |
396 | } |
397 | |
398 | Value *getValueOperand() { return getOperand(0); } |
399 | const Value *getValueOperand() const { return getOperand(0); } |
400 | |
401 | Value *getPointerOperand() { return getOperand(1); } |
402 | const Value *getPointerOperand() const { return getOperand(1); } |
403 | static unsigned getPointerOperandIndex() { return 1U; } |
404 | Type *getPointerOperandType() const { return getPointerOperand()->getType(); } |
405 | |
406 | /// Returns the address space of the pointer operand. |
407 | unsigned getPointerAddressSpace() const { |
408 | return getPointerOperandType()->getPointerAddressSpace(); |
409 | } |
410 | |
411 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
412 | static bool classof(const Instruction *I) { |
413 | return I->getOpcode() == Instruction::Store; |
414 | } |
415 | static bool classof(const Value *V) { |
416 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
417 | } |
418 | |
419 | private: |
420 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
421 | // method so that subclasses cannot accidentally use it. |
422 | template <typename Bitfield> |
423 | void setSubclassData(typename Bitfield::Type Value) { |
424 | Instruction::setSubclassData<Bitfield>(Value); |
425 | } |
426 | |
427 | /// The synchronization scope ID of this store instruction. Not quite enough |
428 | /// room in SubClassData for everything, so synchronization scope ID gets its |
429 | /// own field. |
430 | SyncScope::ID SSID; |
431 | }; |
432 | |
433 | template <> |
434 | struct OperandTraits<StoreInst> : public FixedNumOperandTraits<StoreInst, 2> { |
435 | }; |
436 | |
437 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(StoreInst, Value)StoreInst::op_iterator StoreInst::op_begin() { return OperandTraits <StoreInst>::op_begin(this); } StoreInst::const_op_iterator StoreInst::op_begin() const { return OperandTraits<StoreInst >::op_begin(const_cast<StoreInst*>(this)); } StoreInst ::op_iterator StoreInst::op_end() { return OperandTraits<StoreInst >::op_end(this); } StoreInst::const_op_iterator StoreInst:: op_end() const { return OperandTraits<StoreInst>::op_end (const_cast<StoreInst*>(this)); } Value *StoreInst::getOperand (unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<StoreInst>::op_begin(const_cast <StoreInst*>(this))[i_nocapture].get()); } void StoreInst ::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (( void)0); OperandTraits<StoreInst>::op_begin(this)[i_nocapture ] = Val_nocapture; } unsigned StoreInst::getNumOperands() const { return OperandTraits<StoreInst>::operands(this); } template <int Idx_nocapture> Use &StoreInst::Op() { return this ->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture > const Use &StoreInst::Op() const { return this->OpFrom <Idx_nocapture>(this); } |
438 | |
439 | //===----------------------------------------------------------------------===// |
440 | // FenceInst Class |
441 | //===----------------------------------------------------------------------===// |
442 | |
443 | /// An instruction for ordering other memory operations. |
444 | class FenceInst : public Instruction { |
445 | using OrderingField = AtomicOrderingBitfieldElementT<0>; |
446 | |
447 | void Init(AtomicOrdering Ordering, SyncScope::ID SSID); |
448 | |
449 | protected: |
450 | // Note: Instruction needs to be a friend here to call cloneImpl. |
451 | friend class Instruction; |
452 | |
453 | FenceInst *cloneImpl() const; |
454 | |
455 | public: |
456 | // Ordering may only be Acquire, Release, AcquireRelease, or |
457 | // SequentiallyConsistent. |
458 | FenceInst(LLVMContext &C, AtomicOrdering Ordering, |
459 | SyncScope::ID SSID = SyncScope::System, |
460 | Instruction *InsertBefore = nullptr); |
461 | FenceInst(LLVMContext &C, AtomicOrdering Ordering, SyncScope::ID SSID, |
462 | BasicBlock *InsertAtEnd); |
463 | |
464 | // allocate space for exactly zero operands |
465 | void *operator new(size_t S) { return User::operator new(S, 0); } |
466 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
467 | |
468 | /// Returns the ordering constraint of this fence instruction. |
469 | AtomicOrdering getOrdering() const { |
470 | return getSubclassData<OrderingField>(); |
471 | } |
472 | |
473 | /// Sets the ordering constraint of this fence instruction. May only be |
474 | /// Acquire, Release, AcquireRelease, or SequentiallyConsistent. |
475 | void setOrdering(AtomicOrdering Ordering) { |
476 | setSubclassData<OrderingField>(Ordering); |
477 | } |
478 | |
479 | /// Returns the synchronization scope ID of this fence instruction. |
480 | SyncScope::ID getSyncScopeID() const { |
481 | return SSID; |
482 | } |
483 | |
484 | /// Sets the synchronization scope ID of this fence instruction. |
485 | void setSyncScopeID(SyncScope::ID SSID) { |
486 | this->SSID = SSID; |
487 | } |
488 | |
489 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
490 | static bool classof(const Instruction *I) { |
491 | return I->getOpcode() == Instruction::Fence; |
492 | } |
493 | static bool classof(const Value *V) { |
494 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
495 | } |
496 | |
497 | private: |
498 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
499 | // method so that subclasses cannot accidentally use it. |
500 | template <typename Bitfield> |
501 | void setSubclassData(typename Bitfield::Type Value) { |
502 | Instruction::setSubclassData<Bitfield>(Value); |
503 | } |
504 | |
505 | /// The synchronization scope ID of this fence instruction. Not quite enough |
506 | /// room in SubClassData for everything, so synchronization scope ID gets its |
507 | /// own field. |
508 | SyncScope::ID SSID; |
509 | }; |
510 | |
511 | //===----------------------------------------------------------------------===// |
512 | // AtomicCmpXchgInst Class |
513 | //===----------------------------------------------------------------------===// |
514 | |
515 | /// An instruction that atomically checks whether a |
516 | /// specified value is in a memory location, and, if it is, stores a new value |
517 | /// there. The value returned by this instruction is a pair containing the |
518 | /// original value as first element, and an i1 indicating success (true) or |
519 | /// failure (false) as second element. |
520 | /// |
521 | class AtomicCmpXchgInst : public Instruction { |
522 | void Init(Value *Ptr, Value *Cmp, Value *NewVal, Align Align, |
523 | AtomicOrdering SuccessOrdering, AtomicOrdering FailureOrdering, |
524 | SyncScope::ID SSID); |
525 | |
526 | template <unsigned Offset> |
527 | using AtomicOrderingBitfieldElement = |
528 | typename Bitfield::Element<AtomicOrdering, Offset, 3, |
529 | AtomicOrdering::LAST>; |
530 | |
531 | protected: |
532 | // Note: Instruction needs to be a friend here to call cloneImpl. |
533 | friend class Instruction; |
534 | |
535 | AtomicCmpXchgInst *cloneImpl() const; |
536 | |
537 | public: |
538 | AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment, |
539 | AtomicOrdering SuccessOrdering, |
540 | AtomicOrdering FailureOrdering, SyncScope::ID SSID, |
541 | Instruction *InsertBefore = nullptr); |
542 | AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment, |
543 | AtomicOrdering SuccessOrdering, |
544 | AtomicOrdering FailureOrdering, SyncScope::ID SSID, |
545 | BasicBlock *InsertAtEnd); |
546 | |
547 | // allocate space for exactly three operands |
548 | void *operator new(size_t S) { return User::operator new(S, 3); } |
549 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
550 | |
551 | using VolatileField = BoolBitfieldElementT<0>; |
552 | using WeakField = BoolBitfieldElementT<VolatileField::NextBit>; |
553 | using SuccessOrderingField = |
554 | AtomicOrderingBitfieldElementT<WeakField::NextBit>; |
555 | using FailureOrderingField = |
556 | AtomicOrderingBitfieldElementT<SuccessOrderingField::NextBit>; |
557 | using AlignmentField = |
558 | AlignmentBitfieldElementT<FailureOrderingField::NextBit>; |
559 | static_assert( |
560 | Bitfield::areContiguous<VolatileField, WeakField, SuccessOrderingField, |
561 | FailureOrderingField, AlignmentField>(), |
562 | "Bitfields must be contiguous"); |
563 | |
564 | /// Return the alignment of the memory that is being allocated by the |
565 | /// instruction. |
566 | Align getAlign() const { |
567 | return Align(1ULL << getSubclassData<AlignmentField>()); |
568 | } |
569 | |
570 | void setAlignment(Align Align) { |
571 | setSubclassData<AlignmentField>(Log2(Align)); |
572 | } |
573 | |
574 | /// Return true if this is a cmpxchg from a volatile memory |
575 | /// location. |
576 | /// |
577 | bool isVolatile() const { return getSubclassData<VolatileField>(); } |
578 | |
579 | /// Specify whether this is a volatile cmpxchg. |
580 | /// |
581 | void setVolatile(bool V) { setSubclassData<VolatileField>(V); } |
582 | |
583 | /// Return true if this cmpxchg may spuriously fail. |
584 | bool isWeak() const { return getSubclassData<WeakField>(); } |
585 | |
586 | void setWeak(bool IsWeak) { setSubclassData<WeakField>(IsWeak); } |
587 | |
588 | /// Transparently provide more efficient getOperand methods. |
589 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
590 | |
591 | static bool isValidSuccessOrdering(AtomicOrdering Ordering) { |
592 | return Ordering != AtomicOrdering::NotAtomic && |
593 | Ordering != AtomicOrdering::Unordered; |
594 | } |
595 | |
596 | static bool isValidFailureOrdering(AtomicOrdering Ordering) { |
597 | return Ordering != AtomicOrdering::NotAtomic && |
598 | Ordering != AtomicOrdering::Unordered && |
599 | Ordering != AtomicOrdering::AcquireRelease && |
600 | Ordering != AtomicOrdering::Release; |
601 | } |
602 | |
603 | /// Returns the success ordering constraint of this cmpxchg instruction. |
604 | AtomicOrdering getSuccessOrdering() const { |
605 | return getSubclassData<SuccessOrderingField>(); |
606 | } |
607 | |
608 | /// Sets the success ordering constraint of this cmpxchg instruction. |
609 | void setSuccessOrdering(AtomicOrdering Ordering) { |
610 | assert(isValidSuccessOrdering(Ordering) &&((void)0) |
611 | "invalid CmpXchg success ordering")((void)0); |
612 | setSubclassData<SuccessOrderingField>(Ordering); |
613 | } |
614 | |
615 | /// Returns the failure ordering constraint of this cmpxchg instruction. |
616 | AtomicOrdering getFailureOrdering() const { |
617 | return getSubclassData<FailureOrderingField>(); |
618 | } |
619 | |
620 | /// Sets the failure ordering constraint of this cmpxchg instruction. |
621 | void setFailureOrdering(AtomicOrdering Ordering) { |
622 | assert(isValidFailureOrdering(Ordering) &&((void)0) |
623 | "invalid CmpXchg failure ordering")((void)0); |
624 | setSubclassData<FailureOrderingField>(Ordering); |
625 | } |
626 | |
627 | /// Returns a single ordering which is at least as strong as both the |
628 | /// success and failure orderings for this cmpxchg. |
629 | AtomicOrdering getMergedOrdering() const { |
630 | if (getFailureOrdering() == AtomicOrdering::SequentiallyConsistent) |
631 | return AtomicOrdering::SequentiallyConsistent; |
632 | if (getFailureOrdering() == AtomicOrdering::Acquire) { |
633 | if (getSuccessOrdering() == AtomicOrdering::Monotonic) |
634 | return AtomicOrdering::Acquire; |
635 | if (getSuccessOrdering() == AtomicOrdering::Release) |
636 | return AtomicOrdering::AcquireRelease; |
637 | } |
638 | return getSuccessOrdering(); |
639 | } |
640 | |
641 | /// Returns the synchronization scope ID of this cmpxchg instruction. |
642 | SyncScope::ID getSyncScopeID() const { |
643 | return SSID; |
644 | } |
645 | |
646 | /// Sets the synchronization scope ID of this cmpxchg instruction. |
647 | void setSyncScopeID(SyncScope::ID SSID) { |
648 | this->SSID = SSID; |
649 | } |
650 | |
651 | Value *getPointerOperand() { return getOperand(0); } |
652 | const Value *getPointerOperand() const { return getOperand(0); } |
653 | static unsigned getPointerOperandIndex() { return 0U; } |
654 | |
655 | Value *getCompareOperand() { return getOperand(1); } |
656 | const Value *getCompareOperand() const { return getOperand(1); } |
657 | |
658 | Value *getNewValOperand() { return getOperand(2); } |
659 | const Value *getNewValOperand() const { return getOperand(2); } |
660 | |
661 | /// Returns the address space of the pointer operand. |
662 | unsigned getPointerAddressSpace() const { |
663 | return getPointerOperand()->getType()->getPointerAddressSpace(); |
664 | } |
665 | |
666 | /// Returns the strongest permitted ordering on failure, given the |
667 | /// desired ordering on success. |
668 | /// |
669 | /// If the comparison in a cmpxchg operation fails, there is no atomic store |
670 | /// so release semantics cannot be provided. So this function drops explicit |
671 | /// Release requests from the AtomicOrdering. A SequentiallyConsistent |
672 | /// operation would remain SequentiallyConsistent. |
673 | static AtomicOrdering |
674 | getStrongestFailureOrdering(AtomicOrdering SuccessOrdering) { |
675 | switch (SuccessOrdering) { |
676 | default: |
677 | llvm_unreachable("invalid cmpxchg success ordering")__builtin_unreachable(); |
678 | case AtomicOrdering::Release: |
679 | case AtomicOrdering::Monotonic: |
680 | return AtomicOrdering::Monotonic; |
681 | case AtomicOrdering::AcquireRelease: |
682 | case AtomicOrdering::Acquire: |
683 | return AtomicOrdering::Acquire; |
684 | case AtomicOrdering::SequentiallyConsistent: |
685 | return AtomicOrdering::SequentiallyConsistent; |
686 | } |
687 | } |
688 | |
689 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
690 | static bool classof(const Instruction *I) { |
691 | return I->getOpcode() == Instruction::AtomicCmpXchg; |
692 | } |
693 | static bool classof(const Value *V) { |
694 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
695 | } |
696 | |
697 | private: |
698 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
699 | // method so that subclasses cannot accidentally use it. |
700 | template <typename Bitfield> |
701 | void setSubclassData(typename Bitfield::Type Value) { |
702 | Instruction::setSubclassData<Bitfield>(Value); |
703 | } |
704 | |
705 | /// The synchronization scope ID of this cmpxchg instruction. Not quite |
706 | /// enough room in SubClassData for everything, so synchronization scope ID |
707 | /// gets its own field. |
708 | SyncScope::ID SSID; |
709 | }; |
710 | |
711 | template <> |
712 | struct OperandTraits<AtomicCmpXchgInst> : |
713 | public FixedNumOperandTraits<AtomicCmpXchgInst, 3> { |
714 | }; |
715 | |
716 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(AtomicCmpXchgInst, Value)AtomicCmpXchgInst::op_iterator AtomicCmpXchgInst::op_begin() { return OperandTraits<AtomicCmpXchgInst>::op_begin(this ); } AtomicCmpXchgInst::const_op_iterator AtomicCmpXchgInst:: op_begin() const { return OperandTraits<AtomicCmpXchgInst> ::op_begin(const_cast<AtomicCmpXchgInst*>(this)); } AtomicCmpXchgInst ::op_iterator AtomicCmpXchgInst::op_end() { return OperandTraits <AtomicCmpXchgInst>::op_end(this); } AtomicCmpXchgInst:: const_op_iterator AtomicCmpXchgInst::op_end() const { return OperandTraits <AtomicCmpXchgInst>::op_end(const_cast<AtomicCmpXchgInst *>(this)); } Value *AtomicCmpXchgInst::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null<Value >( OperandTraits<AtomicCmpXchgInst>::op_begin(const_cast <AtomicCmpXchgInst*>(this))[i_nocapture].get()); } void AtomicCmpXchgInst::setOperand(unsigned i_nocapture, Value *Val_nocapture ) { ((void)0); OperandTraits<AtomicCmpXchgInst>::op_begin (this)[i_nocapture] = Val_nocapture; } unsigned AtomicCmpXchgInst ::getNumOperands() const { return OperandTraits<AtomicCmpXchgInst >::operands(this); } template <int Idx_nocapture> Use &AtomicCmpXchgInst::Op() { return this->OpFrom<Idx_nocapture >(this); } template <int Idx_nocapture> const Use & AtomicCmpXchgInst::Op() const { return this->OpFrom<Idx_nocapture >(this); } |
717 | |
718 | //===----------------------------------------------------------------------===// |
719 | // AtomicRMWInst Class |
720 | //===----------------------------------------------------------------------===// |
721 | |
722 | /// an instruction that atomically reads a memory location, |
723 | /// combines it with another value, and then stores the result back. Returns |
724 | /// the old value. |
725 | /// |
726 | class AtomicRMWInst : public Instruction { |
727 | protected: |
728 | // Note: Instruction needs to be a friend here to call cloneImpl. |
729 | friend class Instruction; |
730 | |
731 | AtomicRMWInst *cloneImpl() const; |
732 | |
733 | public: |
734 | /// This enumeration lists the possible modifications atomicrmw can make. In |
735 | /// the descriptions, 'p' is the pointer to the instruction's memory location, |
736 | /// 'old' is the initial value of *p, and 'v' is the other value passed to the |
737 | /// instruction. These instructions always return 'old'. |
738 | enum BinOp : unsigned { |
739 | /// *p = v |
740 | Xchg, |
741 | /// *p = old + v |
742 | Add, |
743 | /// *p = old - v |
744 | Sub, |
745 | /// *p = old & v |
746 | And, |
747 | /// *p = ~(old & v) |
748 | Nand, |
749 | /// *p = old | v |
750 | Or, |
751 | /// *p = old ^ v |
752 | Xor, |
753 | /// *p = old >signed v ? old : v |
754 | Max, |
755 | /// *p = old <signed v ? old : v |
756 | Min, |
757 | /// *p = old >unsigned v ? old : v |
758 | UMax, |
759 | /// *p = old <unsigned v ? old : v |
760 | UMin, |
761 | |
762 | /// *p = old + v |
763 | FAdd, |
764 | |
765 | /// *p = old - v |
766 | FSub, |
767 | |
768 | FIRST_BINOP = Xchg, |
769 | LAST_BINOP = FSub, |
770 | BAD_BINOP |
771 | }; |
772 | |
773 | private: |
774 | template <unsigned Offset> |
775 | using AtomicOrderingBitfieldElement = |
776 | typename Bitfield::Element<AtomicOrdering, Offset, 3, |
777 | AtomicOrdering::LAST>; |
778 | |
779 | template <unsigned Offset> |
780 | using BinOpBitfieldElement = |
781 | typename Bitfield::Element<BinOp, Offset, 4, BinOp::LAST_BINOP>; |
782 | |
783 | public: |
784 | AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment, |
785 | AtomicOrdering Ordering, SyncScope::ID SSID, |
786 | Instruction *InsertBefore = nullptr); |
787 | AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment, |
788 | AtomicOrdering Ordering, SyncScope::ID SSID, |
789 | BasicBlock *InsertAtEnd); |
790 | |
791 | // allocate space for exactly two operands |
792 | void *operator new(size_t S) { return User::operator new(S, 2); } |
793 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
794 | |
795 | using VolatileField = BoolBitfieldElementT<0>; |
796 | using AtomicOrderingField = |
797 | AtomicOrderingBitfieldElementT<VolatileField::NextBit>; |
798 | using OperationField = BinOpBitfieldElement<AtomicOrderingField::NextBit>; |
799 | using AlignmentField = AlignmentBitfieldElementT<OperationField::NextBit>; |
800 | static_assert(Bitfield::areContiguous<VolatileField, AtomicOrderingField, |
801 | OperationField, AlignmentField>(), |
802 | "Bitfields must be contiguous"); |
803 | |
804 | BinOp getOperation() const { return getSubclassData<OperationField>(); } |
805 | |
806 | static StringRef getOperationName(BinOp Op); |
807 | |
808 | static bool isFPOperation(BinOp Op) { |
809 | switch (Op) { |
810 | case AtomicRMWInst::FAdd: |
811 | case AtomicRMWInst::FSub: |
812 | return true; |
813 | default: |
814 | return false; |
815 | } |
816 | } |
817 | |
818 | void setOperation(BinOp Operation) { |
819 | setSubclassData<OperationField>(Operation); |
820 | } |
821 | |
822 | /// Return the alignment of the memory that is being allocated by the |
823 | /// instruction. |
824 | Align getAlign() const { |
825 | return Align(1ULL << getSubclassData<AlignmentField>()); |
826 | } |
827 | |
828 | void setAlignment(Align Align) { |
829 | setSubclassData<AlignmentField>(Log2(Align)); |
830 | } |
831 | |
832 | /// Return true if this is a RMW on a volatile memory location. |
833 | /// |
834 | bool isVolatile() const { return getSubclassData<VolatileField>(); } |
835 | |
836 | /// Specify whether this is a volatile RMW or not. |
837 | /// |
838 | void setVolatile(bool V) { setSubclassData<VolatileField>(V); } |
839 | |
840 | /// Transparently provide more efficient getOperand methods. |
841 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
842 | |
843 | /// Returns the ordering constraint of this rmw instruction. |
844 | AtomicOrdering getOrdering() const { |
845 | return getSubclassData<AtomicOrderingField>(); |
846 | } |
847 | |
848 | /// Sets the ordering constraint of this rmw instruction. |
849 | void setOrdering(AtomicOrdering Ordering) { |
850 | assert(Ordering != AtomicOrdering::NotAtomic &&((void)0) |
851 | "atomicrmw instructions can only be atomic.")((void)0); |
852 | setSubclassData<AtomicOrderingField>(Ordering); |
853 | } |
854 | |
855 | /// Returns the synchronization scope ID of this rmw instruction. |
856 | SyncScope::ID getSyncScopeID() const { |
857 | return SSID; |
858 | } |
859 | |
860 | /// Sets the synchronization scope ID of this rmw instruction. |
861 | void setSyncScopeID(SyncScope::ID SSID) { |
862 | this->SSID = SSID; |
863 | } |
864 | |
865 | Value *getPointerOperand() { return getOperand(0); } |
866 | const Value *getPointerOperand() const { return getOperand(0); } |
867 | static unsigned getPointerOperandIndex() { return 0U; } |
868 | |
869 | Value *getValOperand() { return getOperand(1); } |
870 | const Value *getValOperand() const { return getOperand(1); } |
871 | |
872 | /// Returns the address space of the pointer operand. |
873 | unsigned getPointerAddressSpace() const { |
874 | return getPointerOperand()->getType()->getPointerAddressSpace(); |
875 | } |
876 | |
877 | bool isFloatingPointOperation() const { |
878 | return isFPOperation(getOperation()); |
879 | } |
880 | |
881 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
882 | static bool classof(const Instruction *I) { |
883 | return I->getOpcode() == Instruction::AtomicRMW; |
884 | } |
885 | static bool classof(const Value *V) { |
886 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
887 | } |
888 | |
889 | private: |
890 | void Init(BinOp Operation, Value *Ptr, Value *Val, Align Align, |
891 | AtomicOrdering Ordering, SyncScope::ID SSID); |
892 | |
893 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
894 | // method so that subclasses cannot accidentally use it. |
895 | template <typename Bitfield> |
896 | void setSubclassData(typename Bitfield::Type Value) { |
897 | Instruction::setSubclassData<Bitfield>(Value); |
898 | } |
899 | |
900 | /// The synchronization scope ID of this rmw instruction. Not quite enough |
901 | /// room in SubClassData for everything, so synchronization scope ID gets its |
902 | /// own field. |
903 | SyncScope::ID SSID; |
904 | }; |
905 | |
906 | template <> |
907 | struct OperandTraits<AtomicRMWInst> |
908 | : public FixedNumOperandTraits<AtomicRMWInst,2> { |
909 | }; |
910 | |
911 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(AtomicRMWInst, Value)AtomicRMWInst::op_iterator AtomicRMWInst::op_begin() { return OperandTraits<AtomicRMWInst>::op_begin(this); } AtomicRMWInst ::const_op_iterator AtomicRMWInst::op_begin() const { return OperandTraits <AtomicRMWInst>::op_begin(const_cast<AtomicRMWInst*> (this)); } AtomicRMWInst::op_iterator AtomicRMWInst::op_end() { return OperandTraits<AtomicRMWInst>::op_end(this); } AtomicRMWInst::const_op_iterator AtomicRMWInst::op_end() const { return OperandTraits<AtomicRMWInst>::op_end(const_cast <AtomicRMWInst*>(this)); } Value *AtomicRMWInst::getOperand (unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<AtomicRMWInst>::op_begin(const_cast <AtomicRMWInst*>(this))[i_nocapture].get()); } void AtomicRMWInst ::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (( void)0); OperandTraits<AtomicRMWInst>::op_begin(this)[i_nocapture ] = Val_nocapture; } unsigned AtomicRMWInst::getNumOperands() const { return OperandTraits<AtomicRMWInst>::operands( this); } template <int Idx_nocapture> Use &AtomicRMWInst ::Op() { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &AtomicRMWInst ::Op() const { return this->OpFrom<Idx_nocapture>(this ); } |
912 | |
913 | //===----------------------------------------------------------------------===// |
914 | // GetElementPtrInst Class |
915 | //===----------------------------------------------------------------------===// |
916 | |
917 | // checkGEPType - Simple wrapper function to give a better assertion failure |
918 | // message on bad indexes for a gep instruction. |
919 | // |
920 | inline Type *checkGEPType(Type *Ty) { |
921 | assert(Ty && "Invalid GetElementPtrInst indices for type!")((void)0); |
922 | return Ty; |
923 | } |
924 | |
925 | /// an instruction for type-safe pointer arithmetic to |
926 | /// access elements of arrays and structs |
927 | /// |
928 | class GetElementPtrInst : public Instruction { |
929 | Type *SourceElementType; |
930 | Type *ResultElementType; |
931 | |
932 | GetElementPtrInst(const GetElementPtrInst &GEPI); |
933 | |
934 | /// Constructors - Create a getelementptr instruction with a base pointer an |
935 | /// list of indices. The first ctor can optionally insert before an existing |
936 | /// instruction, the second appends the new instruction to the specified |
937 | /// BasicBlock. |
938 | inline GetElementPtrInst(Type *PointeeType, Value *Ptr, |
939 | ArrayRef<Value *> IdxList, unsigned Values, |
940 | const Twine &NameStr, Instruction *InsertBefore); |
941 | inline GetElementPtrInst(Type *PointeeType, Value *Ptr, |
942 | ArrayRef<Value *> IdxList, unsigned Values, |
943 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
944 | |
945 | void init(Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr); |
946 | |
947 | protected: |
948 | // Note: Instruction needs to be a friend here to call cloneImpl. |
949 | friend class Instruction; |
950 | |
951 | GetElementPtrInst *cloneImpl() const; |
952 | |
953 | public: |
954 | static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr, |
955 | ArrayRef<Value *> IdxList, |
956 | const Twine &NameStr = "", |
957 | Instruction *InsertBefore = nullptr) { |
958 | unsigned Values = 1 + unsigned(IdxList.size()); |
959 | assert(PointeeType && "Must specify element type")((void)0); |
960 | assert(cast<PointerType>(Ptr->getType()->getScalarType())((void)0) |
961 | ->isOpaqueOrPointeeTypeMatches(PointeeType))((void)0); |
962 | return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values, |
963 | NameStr, InsertBefore); |
964 | } |
965 | |
966 | static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr, |
967 | ArrayRef<Value *> IdxList, |
968 | const Twine &NameStr, |
969 | BasicBlock *InsertAtEnd) { |
970 | unsigned Values = 1 + unsigned(IdxList.size()); |
971 | assert(PointeeType && "Must specify element type")((void)0); |
972 | assert(cast<PointerType>(Ptr->getType()->getScalarType())((void)0) |
973 | ->isOpaqueOrPointeeTypeMatches(PointeeType))((void)0); |
974 | return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values, |
975 | NameStr, InsertAtEnd); |
976 | } |
977 | |
978 | LLVM_ATTRIBUTE_DEPRECATED(static GetElementPtrInst *CreateInBounds([[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr = "", Instruction *InsertBefore = nullptr) |
979 | Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr = "",[[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr = "", Instruction *InsertBefore = nullptr) |
980 | Instruction *InsertBefore = nullptr),[[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr = "", Instruction *InsertBefore = nullptr) |
981 | "Use the version with explicit element type instead")[[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr = "", Instruction *InsertBefore = nullptr) { |
982 | return CreateInBounds( |
983 | Ptr->getType()->getScalarType()->getPointerElementType(), Ptr, IdxList, |
984 | NameStr, InsertBefore); |
985 | } |
986 | |
987 | /// Create an "inbounds" getelementptr. See the documentation for the |
988 | /// "inbounds" flag in LangRef.html for details. |
989 | static GetElementPtrInst * |
990 | CreateInBounds(Type *PointeeType, Value *Ptr, ArrayRef<Value *> IdxList, |
991 | const Twine &NameStr = "", |
992 | Instruction *InsertBefore = nullptr) { |
993 | GetElementPtrInst *GEP = |
994 | Create(PointeeType, Ptr, IdxList, NameStr, InsertBefore); |
995 | GEP->setIsInBounds(true); |
996 | return GEP; |
997 | } |
998 | |
999 | LLVM_ATTRIBUTE_DEPRECATED(static GetElementPtrInst *CreateInBounds([[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr, BasicBlock *InsertAtEnd) |
1000 | Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr,[[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr, BasicBlock *InsertAtEnd) |
1001 | BasicBlock *InsertAtEnd),[[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr, BasicBlock *InsertAtEnd) |
1002 | "Use the version with explicit element type instead")[[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr, BasicBlock *InsertAtEnd) { |
1003 | return CreateInBounds( |
1004 | Ptr->getType()->getScalarType()->getPointerElementType(), Ptr, IdxList, |
1005 | NameStr, InsertAtEnd); |
1006 | } |
1007 | |
1008 | static GetElementPtrInst *CreateInBounds(Type *PointeeType, Value *Ptr, |
1009 | ArrayRef<Value *> IdxList, |
1010 | const Twine &NameStr, |
1011 | BasicBlock *InsertAtEnd) { |
1012 | GetElementPtrInst *GEP = |
1013 | Create(PointeeType, Ptr, IdxList, NameStr, InsertAtEnd); |
1014 | GEP->setIsInBounds(true); |
1015 | return GEP; |
1016 | } |
1017 | |
1018 | /// Transparently provide more efficient getOperand methods. |
1019 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
1020 | |
1021 | Type *getSourceElementType() const { return SourceElementType; } |
1022 | |
1023 | void setSourceElementType(Type *Ty) { SourceElementType = Ty; } |
1024 | void setResultElementType(Type *Ty) { ResultElementType = Ty; } |
1025 | |
1026 | Type *getResultElementType() const { |
1027 | assert(cast<PointerType>(getType()->getScalarType())((void)0) |
1028 | ->isOpaqueOrPointeeTypeMatches(ResultElementType))((void)0); |
1029 | return ResultElementType; |
1030 | } |
1031 | |
1032 | /// Returns the address space of this instruction's pointer type. |
1033 | unsigned getAddressSpace() const { |
1034 | // Note that this is always the same as the pointer operand's address space |
1035 | // and that is cheaper to compute, so cheat here. |
1036 | return getPointerAddressSpace(); |
1037 | } |
1038 | |
1039 | /// Returns the result type of a getelementptr with the given source |
1040 | /// element type and indexes. |
1041 | /// |
1042 | /// Null is returned if the indices are invalid for the specified |
1043 | /// source element type. |
1044 | static Type *getIndexedType(Type *Ty, ArrayRef<Value *> IdxList); |
1045 | static Type *getIndexedType(Type *Ty, ArrayRef<Constant *> IdxList); |
1046 | static Type *getIndexedType(Type *Ty, ArrayRef<uint64_t> IdxList); |
1047 | |
1048 | /// Return the type of the element at the given index of an indexable |
1049 | /// type. This is equivalent to "getIndexedType(Agg, {Zero, Idx})". |
1050 | /// |
1051 | /// Returns null if the type can't be indexed, or the given index is not |
1052 | /// legal for the given type. |
1053 | static Type *getTypeAtIndex(Type *Ty, Value *Idx); |
1054 | static Type *getTypeAtIndex(Type *Ty, uint64_t Idx); |
1055 | |
1056 | inline op_iterator idx_begin() { return op_begin()+1; } |
1057 | inline const_op_iterator idx_begin() const { return op_begin()+1; } |
1058 | inline op_iterator idx_end() { return op_end(); } |
1059 | inline const_op_iterator idx_end() const { return op_end(); } |
1060 | |
1061 | inline iterator_range<op_iterator> indices() { |
1062 | return make_range(idx_begin(), idx_end()); |
1063 | } |
1064 | |
1065 | inline iterator_range<const_op_iterator> indices() const { |
1066 | return make_range(idx_begin(), idx_end()); |
1067 | } |
1068 | |
1069 | Value *getPointerOperand() { |
1070 | return getOperand(0); |
1071 | } |
1072 | const Value *getPointerOperand() const { |
1073 | return getOperand(0); |
1074 | } |
1075 | static unsigned getPointerOperandIndex() { |
1076 | return 0U; // get index for modifying correct operand. |
1077 | } |
1078 | |
1079 | /// Method to return the pointer operand as a |
1080 | /// PointerType. |
1081 | Type *getPointerOperandType() const { |
1082 | return getPointerOperand()->getType(); |
1083 | } |
1084 | |
1085 | /// Returns the address space of the pointer operand. |
1086 | unsigned getPointerAddressSpace() const { |
1087 | return getPointerOperandType()->getPointerAddressSpace(); |
1088 | } |
1089 | |
1090 | /// Returns the pointer type returned by the GEP |
1091 | /// instruction, which may be a vector of pointers. |
1092 | static Type *getGEPReturnType(Type *ElTy, Value *Ptr, |
1093 | ArrayRef<Value *> IdxList) { |
1094 | PointerType *OrigPtrTy = cast<PointerType>(Ptr->getType()->getScalarType()); |
1095 | unsigned AddrSpace = OrigPtrTy->getAddressSpace(); |
1096 | Type *ResultElemTy = checkGEPType(getIndexedType(ElTy, IdxList)); |
1097 | Type *PtrTy = OrigPtrTy->isOpaque() |
1098 | ? PointerType::get(OrigPtrTy->getContext(), AddrSpace) |
1099 | : PointerType::get(ResultElemTy, AddrSpace); |
1100 | // Vector GEP |
1101 | if (auto *PtrVTy = dyn_cast<VectorType>(Ptr->getType())) { |
1102 | ElementCount EltCount = PtrVTy->getElementCount(); |
1103 | return VectorType::get(PtrTy, EltCount); |
1104 | } |
1105 | for (Value *Index : IdxList) |
1106 | if (auto *IndexVTy = dyn_cast<VectorType>(Index->getType())) { |
1107 | ElementCount EltCount = IndexVTy->getElementCount(); |
1108 | return VectorType::get(PtrTy, EltCount); |
1109 | } |
1110 | // Scalar GEP |
1111 | return PtrTy; |
1112 | } |
1113 | |
1114 | unsigned getNumIndices() const { // Note: always non-negative |
1115 | return getNumOperands() - 1; |
1116 | } |
1117 | |
1118 | bool hasIndices() const { |
1119 | return getNumOperands() > 1; |
1120 | } |
1121 | |
1122 | /// Return true if all of the indices of this GEP are |
1123 | /// zeros. If so, the result pointer and the first operand have the same |
1124 | /// value, just potentially different types. |
1125 | bool hasAllZeroIndices() const; |
1126 | |
1127 | /// Return true if all of the indices of this GEP are |
1128 | /// constant integers. If so, the result pointer and the first operand have |
1129 | /// a constant offset between them. |
1130 | bool hasAllConstantIndices() const; |
1131 | |
1132 | /// Set or clear the inbounds flag on this GEP instruction. |
1133 | /// See LangRef.html for the meaning of inbounds on a getelementptr. |
1134 | void setIsInBounds(bool b = true); |
1135 | |
1136 | /// Determine whether the GEP has the inbounds flag. |
1137 | bool isInBounds() const; |
1138 | |
1139 | /// Accumulate the constant address offset of this GEP if possible. |
1140 | /// |
1141 | /// This routine accepts an APInt into which it will accumulate the constant |
1142 | /// offset of this GEP if the GEP is in fact constant. If the GEP is not |
1143 | /// all-constant, it returns false and the value of the offset APInt is |
1144 | /// undefined (it is *not* preserved!). The APInt passed into this routine |
1145 | /// must be at least as wide as the IntPtr type for the address space of |
1146 | /// the base GEP pointer. |
1147 | bool accumulateConstantOffset(const DataLayout &DL, APInt &Offset) const; |
1148 | bool collectOffset(const DataLayout &DL, unsigned BitWidth, |
1149 | MapVector<Value *, APInt> &VariableOffsets, |
1150 | APInt &ConstantOffset) const; |
1151 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1152 | static bool classof(const Instruction *I) { |
1153 | return (I->getOpcode() == Instruction::GetElementPtr); |
1154 | } |
1155 | static bool classof(const Value *V) { |
1156 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1157 | } |
1158 | }; |
1159 | |
1160 | template <> |
1161 | struct OperandTraits<GetElementPtrInst> : |
1162 | public VariadicOperandTraits<GetElementPtrInst, 1> { |
1163 | }; |
1164 | |
1165 | GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr, |
1166 | ArrayRef<Value *> IdxList, unsigned Values, |
1167 | const Twine &NameStr, |
1168 | Instruction *InsertBefore) |
1169 | : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr, |
1170 | OperandTraits<GetElementPtrInst>::op_end(this) - Values, |
1171 | Values, InsertBefore), |
1172 | SourceElementType(PointeeType), |
1173 | ResultElementType(getIndexedType(PointeeType, IdxList)) { |
1174 | assert(cast<PointerType>(getType()->getScalarType())((void)0) |
1175 | ->isOpaqueOrPointeeTypeMatches(ResultElementType))((void)0); |
1176 | init(Ptr, IdxList, NameStr); |
1177 | } |
1178 | |
1179 | GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr, |
1180 | ArrayRef<Value *> IdxList, unsigned Values, |
1181 | const Twine &NameStr, |
1182 | BasicBlock *InsertAtEnd) |
1183 | : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr, |
1184 | OperandTraits<GetElementPtrInst>::op_end(this) - Values, |
1185 | Values, InsertAtEnd), |
1186 | SourceElementType(PointeeType), |
1187 | ResultElementType(getIndexedType(PointeeType, IdxList)) { |
1188 | assert(cast<PointerType>(getType()->getScalarType())((void)0) |
1189 | ->isOpaqueOrPointeeTypeMatches(ResultElementType))((void)0); |
1190 | init(Ptr, IdxList, NameStr); |
1191 | } |
1192 | |
1193 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(GetElementPtrInst, Value)GetElementPtrInst::op_iterator GetElementPtrInst::op_begin() { return OperandTraits<GetElementPtrInst>::op_begin(this ); } GetElementPtrInst::const_op_iterator GetElementPtrInst:: op_begin() const { return OperandTraits<GetElementPtrInst> ::op_begin(const_cast<GetElementPtrInst*>(this)); } GetElementPtrInst ::op_iterator GetElementPtrInst::op_end() { return OperandTraits <GetElementPtrInst>::op_end(this); } GetElementPtrInst:: const_op_iterator GetElementPtrInst::op_end() const { return OperandTraits <GetElementPtrInst>::op_end(const_cast<GetElementPtrInst *>(this)); } Value *GetElementPtrInst::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null<Value >( OperandTraits<GetElementPtrInst>::op_begin(const_cast <GetElementPtrInst*>(this))[i_nocapture].get()); } void GetElementPtrInst::setOperand(unsigned i_nocapture, Value *Val_nocapture ) { ((void)0); OperandTraits<GetElementPtrInst>::op_begin (this)[i_nocapture] = Val_nocapture; } unsigned GetElementPtrInst ::getNumOperands() const { return OperandTraits<GetElementPtrInst >::operands(this); } template <int Idx_nocapture> Use &GetElementPtrInst::Op() { return this->OpFrom<Idx_nocapture >(this); } template <int Idx_nocapture> const Use & GetElementPtrInst::Op() const { return this->OpFrom<Idx_nocapture >(this); } |
1194 | |
1195 | //===----------------------------------------------------------------------===// |
1196 | // ICmpInst Class |
1197 | //===----------------------------------------------------------------------===// |
1198 | |
1199 | /// This instruction compares its operands according to the predicate given |
1200 | /// to the constructor. It only operates on integers or pointers. The operands |
1201 | /// must be identical types. |
1202 | /// Represent an integer comparison operator. |
1203 | class ICmpInst: public CmpInst { |
1204 | void AssertOK() { |
1205 | assert(isIntPredicate() &&((void)0) |
1206 | "Invalid ICmp predicate value")((void)0); |
1207 | assert(getOperand(0)->getType() == getOperand(1)->getType() &&((void)0) |
1208 | "Both operands to ICmp instruction are not of the same type!")((void)0); |
1209 | // Check that the operands are the right type |
1210 | assert((getOperand(0)->getType()->isIntOrIntVectorTy() ||((void)0) |
1211 | getOperand(0)->getType()->isPtrOrPtrVectorTy()) &&((void)0) |
1212 | "Invalid operand types for ICmp instruction")((void)0); |
1213 | } |
1214 | |
1215 | protected: |
1216 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1217 | friend class Instruction; |
1218 | |
1219 | /// Clone an identical ICmpInst |
1220 | ICmpInst *cloneImpl() const; |
1221 | |
1222 | public: |
1223 | /// Constructor with insert-before-instruction semantics. |
1224 | ICmpInst( |
1225 | Instruction *InsertBefore, ///< Where to insert |
1226 | Predicate pred, ///< The predicate to use for the comparison |
1227 | Value *LHS, ///< The left-hand-side of the expression |
1228 | Value *RHS, ///< The right-hand-side of the expression |
1229 | const Twine &NameStr = "" ///< Name of the instruction |
1230 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1231 | Instruction::ICmp, pred, LHS, RHS, NameStr, |
1232 | InsertBefore) { |
1233 | #ifndef NDEBUG1 |
1234 | AssertOK(); |
1235 | #endif |
1236 | } |
1237 | |
1238 | /// Constructor with insert-at-end semantics. |
1239 | ICmpInst( |
1240 | BasicBlock &InsertAtEnd, ///< Block to insert into. |
1241 | Predicate pred, ///< The predicate to use for the comparison |
1242 | Value *LHS, ///< The left-hand-side of the expression |
1243 | Value *RHS, ///< The right-hand-side of the expression |
1244 | const Twine &NameStr = "" ///< Name of the instruction |
1245 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1246 | Instruction::ICmp, pred, LHS, RHS, NameStr, |
1247 | &InsertAtEnd) { |
1248 | #ifndef NDEBUG1 |
1249 | AssertOK(); |
1250 | #endif |
1251 | } |
1252 | |
1253 | /// Constructor with no-insertion semantics |
1254 | ICmpInst( |
1255 | Predicate pred, ///< The predicate to use for the comparison |
1256 | Value *LHS, ///< The left-hand-side of the expression |
1257 | Value *RHS, ///< The right-hand-side of the expression |
1258 | const Twine &NameStr = "" ///< Name of the instruction |
1259 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1260 | Instruction::ICmp, pred, LHS, RHS, NameStr) { |
1261 | #ifndef NDEBUG1 |
1262 | AssertOK(); |
1263 | #endif |
1264 | } |
1265 | |
1266 | /// For example, EQ->EQ, SLE->SLE, UGT->SGT, etc. |
1267 | /// @returns the predicate that would be the result if the operand were |
1268 | /// regarded as signed. |
1269 | /// Return the signed version of the predicate |
1270 | Predicate getSignedPredicate() const { |
1271 | return getSignedPredicate(getPredicate()); |
1272 | } |
1273 | |
1274 | /// This is a static version that you can use without an instruction. |
1275 | /// Return the signed version of the predicate. |
1276 | static Predicate getSignedPredicate(Predicate pred); |
1277 | |
1278 | /// For example, EQ->EQ, SLE->ULE, UGT->UGT, etc. |
1279 | /// @returns the predicate that would be the result if the operand were |
1280 | /// regarded as unsigned. |
1281 | /// Return the unsigned version of the predicate |
1282 | Predicate getUnsignedPredicate() const { |
1283 | return getUnsignedPredicate(getPredicate()); |
1284 | } |
1285 | |
1286 | /// This is a static version that you can use without an instruction. |
1287 | /// Return the unsigned version of the predicate. |
1288 | static Predicate getUnsignedPredicate(Predicate pred); |
1289 | |
1290 | /// Return true if this predicate is either EQ or NE. This also |
1291 | /// tests for commutativity. |
1292 | static bool isEquality(Predicate P) { |
1293 | return P == ICMP_EQ || P == ICMP_NE; |
1294 | } |
1295 | |
1296 | /// Return true if this predicate is either EQ or NE. This also |
1297 | /// tests for commutativity. |
1298 | bool isEquality() const { |
1299 | return isEquality(getPredicate()); |
1300 | } |
1301 | |
1302 | /// @returns true if the predicate of this ICmpInst is commutative |
1303 | /// Determine if this relation is commutative. |
1304 | bool isCommutative() const { return isEquality(); } |
1305 | |
1306 | /// Return true if the predicate is relational (not EQ or NE). |
1307 | /// |
1308 | bool isRelational() const { |
1309 | return !isEquality(); |
1310 | } |
1311 | |
1312 | /// Return true if the predicate is relational (not EQ or NE). |
1313 | /// |
1314 | static bool isRelational(Predicate P) { |
1315 | return !isEquality(P); |
1316 | } |
1317 | |
1318 | /// Return true if the predicate is SGT or UGT. |
1319 | /// |
1320 | static bool isGT(Predicate P) { |
1321 | return P == ICMP_SGT || P == ICMP_UGT; |
1322 | } |
1323 | |
1324 | /// Return true if the predicate is SLT or ULT. |
1325 | /// |
1326 | static bool isLT(Predicate P) { |
1327 | return P == ICMP_SLT || P == ICMP_ULT; |
1328 | } |
1329 | |
1330 | /// Return true if the predicate is SGE or UGE. |
1331 | /// |
1332 | static bool isGE(Predicate P) { |
1333 | return P == ICMP_SGE || P == ICMP_UGE; |
1334 | } |
1335 | |
1336 | /// Return true if the predicate is SLE or ULE. |
1337 | /// |
1338 | static bool isLE(Predicate P) { |
1339 | return P == ICMP_SLE || P == ICMP_ULE; |
1340 | } |
1341 | |
1342 | /// Exchange the two operands to this instruction in such a way that it does |
1343 | /// not modify the semantics of the instruction. The predicate value may be |
1344 | /// changed to retain the same result if the predicate is order dependent |
1345 | /// (e.g. ult). |
1346 | /// Swap operands and adjust predicate. |
1347 | void swapOperands() { |
1348 | setPredicate(getSwappedPredicate()); |
1349 | Op<0>().swap(Op<1>()); |
1350 | } |
1351 | |
1352 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1353 | static bool classof(const Instruction *I) { |
1354 | return I->getOpcode() == Instruction::ICmp; |
1355 | } |
1356 | static bool classof(const Value *V) { |
1357 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1358 | } |
1359 | }; |
1360 | |
1361 | //===----------------------------------------------------------------------===// |
1362 | // FCmpInst Class |
1363 | //===----------------------------------------------------------------------===// |
1364 | |
1365 | /// This instruction compares its operands according to the predicate given |
1366 | /// to the constructor. It only operates on floating point values or packed |
1367 | /// vectors of floating point values. The operands must be identical types. |
1368 | /// Represents a floating point comparison operator. |
1369 | class FCmpInst: public CmpInst { |
1370 | void AssertOK() { |
1371 | assert(isFPPredicate() && "Invalid FCmp predicate value")((void)0); |
1372 | assert(getOperand(0)->getType() == getOperand(1)->getType() &&((void)0) |
1373 | "Both operands to FCmp instruction are not of the same type!")((void)0); |
1374 | // Check that the operands are the right type |
1375 | assert(getOperand(0)->getType()->isFPOrFPVectorTy() &&((void)0) |
1376 | "Invalid operand types for FCmp instruction")((void)0); |
1377 | } |
1378 | |
1379 | protected: |
1380 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1381 | friend class Instruction; |
1382 | |
1383 | /// Clone an identical FCmpInst |
1384 | FCmpInst *cloneImpl() const; |
1385 | |
1386 | public: |
1387 | /// Constructor with insert-before-instruction semantics. |
1388 | FCmpInst( |
1389 | Instruction *InsertBefore, ///< Where to insert |
1390 | Predicate pred, ///< The predicate to use for the comparison |
1391 | Value *LHS, ///< The left-hand-side of the expression |
1392 | Value *RHS, ///< The right-hand-side of the expression |
1393 | const Twine &NameStr = "" ///< Name of the instruction |
1394 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1395 | Instruction::FCmp, pred, LHS, RHS, NameStr, |
1396 | InsertBefore) { |
1397 | AssertOK(); |
1398 | } |
1399 | |
1400 | /// Constructor with insert-at-end semantics. |
1401 | FCmpInst( |
1402 | BasicBlock &InsertAtEnd, ///< Block to insert into. |
1403 | Predicate pred, ///< The predicate to use for the comparison |
1404 | Value *LHS, ///< The left-hand-side of the expression |
1405 | Value *RHS, ///< The right-hand-side of the expression |
1406 | const Twine &NameStr = "" ///< Name of the instruction |
1407 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1408 | Instruction::FCmp, pred, LHS, RHS, NameStr, |
1409 | &InsertAtEnd) { |
1410 | AssertOK(); |
1411 | } |
1412 | |
1413 | /// Constructor with no-insertion semantics |
1414 | FCmpInst( |
1415 | Predicate Pred, ///< The predicate to use for the comparison |
1416 | Value *LHS, ///< The left-hand-side of the expression |
1417 | Value *RHS, ///< The right-hand-side of the expression |
1418 | const Twine &NameStr = "", ///< Name of the instruction |
1419 | Instruction *FlagsSource = nullptr |
1420 | ) : CmpInst(makeCmpResultType(LHS->getType()), Instruction::FCmp, Pred, LHS, |
1421 | RHS, NameStr, nullptr, FlagsSource) { |
1422 | AssertOK(); |
1423 | } |
1424 | |
1425 | /// @returns true if the predicate of this instruction is EQ or NE. |
1426 | /// Determine if this is an equality predicate. |
1427 | static bool isEquality(Predicate Pred) { |
1428 | return Pred == FCMP_OEQ || Pred == FCMP_ONE || Pred == FCMP_UEQ || |
1429 | Pred == FCMP_UNE; |
1430 | } |
1431 | |
1432 | /// @returns true if the predicate of this instruction is EQ or NE. |
1433 | /// Determine if this is an equality predicate. |
1434 | bool isEquality() const { return isEquality(getPredicate()); } |
1435 | |
1436 | /// @returns true if the predicate of this instruction is commutative. |
1437 | /// Determine if this is a commutative predicate. |
1438 | bool isCommutative() const { |
1439 | return isEquality() || |
1440 | getPredicate() == FCMP_FALSE || |
1441 | getPredicate() == FCMP_TRUE || |
1442 | getPredicate() == FCMP_ORD || |
1443 | getPredicate() == FCMP_UNO; |
1444 | } |
1445 | |
1446 | /// @returns true if the predicate is relational (not EQ or NE). |
1447 | /// Determine if this a relational predicate. |
1448 | bool isRelational() const { return !isEquality(); } |
1449 | |
1450 | /// Exchange the two operands to this instruction in such a way that it does |
1451 | /// not modify the semantics of the instruction. The predicate value may be |
1452 | /// changed to retain the same result if the predicate is order dependent |
1453 | /// (e.g. ult). |
1454 | /// Swap operands and adjust predicate. |
1455 | void swapOperands() { |
1456 | setPredicate(getSwappedPredicate()); |
1457 | Op<0>().swap(Op<1>()); |
1458 | } |
1459 | |
1460 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
1461 | static bool classof(const Instruction *I) { |
1462 | return I->getOpcode() == Instruction::FCmp; |
1463 | } |
1464 | static bool classof(const Value *V) { |
1465 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1466 | } |
1467 | }; |
1468 | |
1469 | //===----------------------------------------------------------------------===// |
1470 | /// This class represents a function call, abstracting a target |
1471 | /// machine's calling convention. This class uses low bit of the SubClassData |
1472 | /// field to indicate whether or not this is a tail call. The rest of the bits |
1473 | /// hold the calling convention of the call. |
1474 | /// |
1475 | class CallInst : public CallBase { |
1476 | CallInst(const CallInst &CI); |
1477 | |
1478 | /// Construct a CallInst given a range of arguments. |
1479 | /// Construct a CallInst from a range of arguments |
1480 | inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1481 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1482 | Instruction *InsertBefore); |
1483 | |
1484 | inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1485 | const Twine &NameStr, Instruction *InsertBefore) |
1486 | : CallInst(Ty, Func, Args, None, NameStr, InsertBefore) {} |
1487 | |
1488 | /// Construct a CallInst given a range of arguments. |
1489 | /// Construct a CallInst from a range of arguments |
1490 | inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1491 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1492 | BasicBlock *InsertAtEnd); |
1493 | |
1494 | explicit CallInst(FunctionType *Ty, Value *F, const Twine &NameStr, |
1495 | Instruction *InsertBefore); |
1496 | |
1497 | CallInst(FunctionType *ty, Value *F, const Twine &NameStr, |
1498 | BasicBlock *InsertAtEnd); |
1499 | |
1500 | void init(FunctionType *FTy, Value *Func, ArrayRef<Value *> Args, |
1501 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr); |
1502 | void init(FunctionType *FTy, Value *Func, const Twine &NameStr); |
1503 | |
1504 | /// Compute the number of operands to allocate. |
1505 | static int ComputeNumOperands(int NumArgs, int NumBundleInputs = 0) { |
1506 | // We need one operand for the called function, plus the input operand |
1507 | // counts provided. |
1508 | return 1 + NumArgs + NumBundleInputs; |
1509 | } |
1510 | |
1511 | protected: |
1512 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1513 | friend class Instruction; |
1514 | |
1515 | CallInst *cloneImpl() const; |
1516 | |
1517 | public: |
1518 | static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr = "", |
1519 | Instruction *InsertBefore = nullptr) { |
1520 | return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertBefore); |
1521 | } |
1522 | |
1523 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1524 | const Twine &NameStr, |
1525 | Instruction *InsertBefore = nullptr) { |
1526 | return new (ComputeNumOperands(Args.size())) |
1527 | CallInst(Ty, Func, Args, None, NameStr, InsertBefore); |
1528 | } |
1529 | |
1530 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1531 | ArrayRef<OperandBundleDef> Bundles = None, |
1532 | const Twine &NameStr = "", |
1533 | Instruction *InsertBefore = nullptr) { |
1534 | const int NumOperands = |
1535 | ComputeNumOperands(Args.size(), CountBundleInputs(Bundles)); |
1536 | const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
1537 | |
1538 | return new (NumOperands, DescriptorBytes) |
1539 | CallInst(Ty, Func, Args, Bundles, NameStr, InsertBefore); |
1540 | } |
1541 | |
1542 | static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr, |
1543 | BasicBlock *InsertAtEnd) { |
1544 | return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertAtEnd); |
1545 | } |
1546 | |
1547 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1548 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
1549 | return new (ComputeNumOperands(Args.size())) |
1550 | CallInst(Ty, Func, Args, None, NameStr, InsertAtEnd); |
1551 | } |
1552 | |
1553 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1554 | ArrayRef<OperandBundleDef> Bundles, |
1555 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
1556 | const int NumOperands = |
1557 | ComputeNumOperands(Args.size(), CountBundleInputs(Bundles)); |
1558 | const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
1559 | |
1560 | return new (NumOperands, DescriptorBytes) |
1561 | CallInst(Ty, Func, Args, Bundles, NameStr, InsertAtEnd); |
1562 | } |
1563 | |
1564 | static CallInst *Create(FunctionCallee Func, const Twine &NameStr = "", |
1565 | Instruction *InsertBefore = nullptr) { |
1566 | return Create(Func.getFunctionType(), Func.getCallee(), NameStr, |
1567 | InsertBefore); |
1568 | } |
1569 | |
1570 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
1571 | ArrayRef<OperandBundleDef> Bundles = None, |
1572 | const Twine &NameStr = "", |
1573 | Instruction *InsertBefore = nullptr) { |
1574 | return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles, |
1575 | NameStr, InsertBefore); |
1576 | } |
1577 | |
1578 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
1579 | const Twine &NameStr, |
1580 | Instruction *InsertBefore = nullptr) { |
1581 | return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr, |
1582 | InsertBefore); |
1583 | } |
1584 | |
1585 | static CallInst *Create(FunctionCallee Func, const Twine &NameStr, |
1586 | BasicBlock *InsertAtEnd) { |
1587 | return Create(Func.getFunctionType(), Func.getCallee(), NameStr, |
1588 | InsertAtEnd); |
1589 | } |
1590 | |
1591 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
1592 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
1593 | return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr, |
1594 | InsertAtEnd); |
1595 | } |
1596 | |
1597 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
1598 | ArrayRef<OperandBundleDef> Bundles, |
1599 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
1600 | return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles, |
1601 | NameStr, InsertAtEnd); |
1602 | } |
1603 | |
1604 | /// Create a clone of \p CI with a different set of operand bundles and |
1605 | /// insert it before \p InsertPt. |
1606 | /// |
1607 | /// The returned call instruction is identical \p CI in every way except that |
1608 | /// the operand bundles for the new instruction are set to the operand bundles |
1609 | /// in \p Bundles. |
1610 | static CallInst *Create(CallInst *CI, ArrayRef<OperandBundleDef> Bundles, |
1611 | Instruction *InsertPt = nullptr); |
1612 | |
1613 | /// Generate the IR for a call to malloc: |
1614 | /// 1. Compute the malloc call's argument as the specified type's size, |
1615 | /// possibly multiplied by the array size if the array size is not |
1616 | /// constant 1. |
1617 | /// 2. Call malloc with that argument. |
1618 | /// 3. Bitcast the result of the malloc call to the specified type. |
1619 | static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy, |
1620 | Type *AllocTy, Value *AllocSize, |
1621 | Value *ArraySize = nullptr, |
1622 | Function *MallocF = nullptr, |
1623 | const Twine &Name = ""); |
1624 | static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy, |
1625 | Type *AllocTy, Value *AllocSize, |
1626 | Value *ArraySize = nullptr, |
1627 | Function *MallocF = nullptr, |
1628 | const Twine &Name = ""); |
1629 | static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy, |
1630 | Type *AllocTy, Value *AllocSize, |
1631 | Value *ArraySize = nullptr, |
1632 | ArrayRef<OperandBundleDef> Bundles = None, |
1633 | Function *MallocF = nullptr, |
1634 | const Twine &Name = ""); |
1635 | static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy, |
1636 | Type *AllocTy, Value *AllocSize, |
1637 | Value *ArraySize = nullptr, |
1638 | ArrayRef<OperandBundleDef> Bundles = None, |
1639 | Function *MallocF = nullptr, |
1640 | const Twine &Name = ""); |
1641 | /// Generate the IR for a call to the builtin free function. |
1642 | static Instruction *CreateFree(Value *Source, Instruction *InsertBefore); |
1643 | static Instruction *CreateFree(Value *Source, BasicBlock *InsertAtEnd); |
1644 | static Instruction *CreateFree(Value *Source, |
1645 | ArrayRef<OperandBundleDef> Bundles, |
1646 | Instruction *InsertBefore); |
1647 | static Instruction *CreateFree(Value *Source, |
1648 | ArrayRef<OperandBundleDef> Bundles, |
1649 | BasicBlock *InsertAtEnd); |
1650 | |
1651 | // Note that 'musttail' implies 'tail'. |
1652 | enum TailCallKind : unsigned { |
1653 | TCK_None = 0, |
1654 | TCK_Tail = 1, |
1655 | TCK_MustTail = 2, |
1656 | TCK_NoTail = 3, |
1657 | TCK_LAST = TCK_NoTail |
1658 | }; |
1659 | |
1660 | using TailCallKindField = Bitfield::Element<TailCallKind, 0, 2, TCK_LAST>; |
1661 | static_assert( |
1662 | Bitfield::areContiguous<TailCallKindField, CallBase::CallingConvField>(), |
1663 | "Bitfields must be contiguous"); |
1664 | |
1665 | TailCallKind getTailCallKind() const { |
1666 | return getSubclassData<TailCallKindField>(); |
1667 | } |
1668 | |
1669 | bool isTailCall() const { |
1670 | TailCallKind Kind = getTailCallKind(); |
1671 | return Kind == TCK_Tail || Kind == TCK_MustTail; |
1672 | } |
1673 | |
1674 | bool isMustTailCall() const { return getTailCallKind() == TCK_MustTail; } |
1675 | |
1676 | bool isNoTailCall() const { return getTailCallKind() == TCK_NoTail; } |
1677 | |
1678 | void setTailCallKind(TailCallKind TCK) { |
1679 | setSubclassData<TailCallKindField>(TCK); |
1680 | } |
1681 | |
1682 | void setTailCall(bool IsTc = true) { |
1683 | setTailCallKind(IsTc ? TCK_Tail : TCK_None); |
1684 | } |
1685 | |
1686 | /// Return true if the call can return twice |
1687 | bool canReturnTwice() const { return hasFnAttr(Attribute::ReturnsTwice); } |
1688 | void setCanReturnTwice() { |
1689 | addAttribute(AttributeList::FunctionIndex, Attribute::ReturnsTwice); |
1690 | } |
1691 | |
1692 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1693 | static bool classof(const Instruction *I) { |
1694 | return I->getOpcode() == Instruction::Call; |
1695 | } |
1696 | static bool classof(const Value *V) { |
1697 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1698 | } |
1699 | |
1700 | /// Updates profile metadata by scaling it by \p S / \p T. |
1701 | void updateProfWeight(uint64_t S, uint64_t T); |
1702 | |
1703 | private: |
1704 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
1705 | // method so that subclasses cannot accidentally use it. |
1706 | template <typename Bitfield> |
1707 | void setSubclassData(typename Bitfield::Type Value) { |
1708 | Instruction::setSubclassData<Bitfield>(Value); |
1709 | } |
1710 | }; |
1711 | |
1712 | CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1713 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1714 | BasicBlock *InsertAtEnd) |
1715 | : CallBase(Ty->getReturnType(), Instruction::Call, |
1716 | OperandTraits<CallBase>::op_end(this) - |
1717 | (Args.size() + CountBundleInputs(Bundles) + 1), |
1718 | unsigned(Args.size() + CountBundleInputs(Bundles) + 1), |
1719 | InsertAtEnd) { |
1720 | init(Ty, Func, Args, Bundles, NameStr); |
1721 | } |
1722 | |
1723 | CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1724 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1725 | Instruction *InsertBefore) |
1726 | : CallBase(Ty->getReturnType(), Instruction::Call, |
1727 | OperandTraits<CallBase>::op_end(this) - |
1728 | (Args.size() + CountBundleInputs(Bundles) + 1), |
1729 | unsigned(Args.size() + CountBundleInputs(Bundles) + 1), |
1730 | InsertBefore) { |
1731 | init(Ty, Func, Args, Bundles, NameStr); |
1732 | } |
1733 | |
1734 | //===----------------------------------------------------------------------===// |
1735 | // SelectInst Class |
1736 | //===----------------------------------------------------------------------===// |
1737 | |
1738 | /// This class represents the LLVM 'select' instruction. |
1739 | /// |
1740 | class SelectInst : public Instruction { |
1741 | SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr, |
1742 | Instruction *InsertBefore) |
1743 | : Instruction(S1->getType(), Instruction::Select, |
1744 | &Op<0>(), 3, InsertBefore) { |
1745 | init(C, S1, S2); |
1746 | setName(NameStr); |
1747 | } |
1748 | |
1749 | SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr, |
1750 | BasicBlock *InsertAtEnd) |
1751 | : Instruction(S1->getType(), Instruction::Select, |
1752 | &Op<0>(), 3, InsertAtEnd) { |
1753 | init(C, S1, S2); |
1754 | setName(NameStr); |
1755 | } |
1756 | |
1757 | void init(Value *C, Value *S1, Value *S2) { |
1758 | assert(!areInvalidOperands(C, S1, S2) && "Invalid operands for select")((void)0); |
1759 | Op<0>() = C; |
1760 | Op<1>() = S1; |
1761 | Op<2>() = S2; |
1762 | } |
1763 | |
1764 | protected: |
1765 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1766 | friend class Instruction; |
1767 | |
1768 | SelectInst *cloneImpl() const; |
1769 | |
1770 | public: |
1771 | static SelectInst *Create(Value *C, Value *S1, Value *S2, |
1772 | const Twine &NameStr = "", |
1773 | Instruction *InsertBefore = nullptr, |
1774 | Instruction *MDFrom = nullptr) { |
1775 | SelectInst *Sel = new(3) SelectInst(C, S1, S2, NameStr, InsertBefore); |
1776 | if (MDFrom) |
1777 | Sel->copyMetadata(*MDFrom); |
1778 | return Sel; |
1779 | } |
1780 | |
1781 | static SelectInst *Create(Value *C, Value *S1, Value *S2, |
1782 | const Twine &NameStr, |
1783 | BasicBlock *InsertAtEnd) { |
1784 | return new(3) SelectInst(C, S1, S2, NameStr, InsertAtEnd); |
1785 | } |
1786 | |
1787 | const Value *getCondition() const { return Op<0>(); } |
1788 | const Value *getTrueValue() const { return Op<1>(); } |
1789 | const Value *getFalseValue() const { return Op<2>(); } |
1790 | Value *getCondition() { return Op<0>(); } |
1791 | Value *getTrueValue() { return Op<1>(); } |
1792 | Value *getFalseValue() { return Op<2>(); } |
1793 | |
1794 | void setCondition(Value *V) { Op<0>() = V; } |
1795 | void setTrueValue(Value *V) { Op<1>() = V; } |
1796 | void setFalseValue(Value *V) { Op<2>() = V; } |
1797 | |
1798 | /// Swap the true and false values of the select instruction. |
1799 | /// This doesn't swap prof metadata. |
1800 | void swapValues() { Op<1>().swap(Op<2>()); } |
1801 | |
1802 | /// Return a string if the specified operands are invalid |
1803 | /// for a select operation, otherwise return null. |
1804 | static const char *areInvalidOperands(Value *Cond, Value *True, Value *False); |
1805 | |
1806 | /// Transparently provide more efficient getOperand methods. |
1807 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
1808 | |
1809 | OtherOps getOpcode() const { |
1810 | return static_cast<OtherOps>(Instruction::getOpcode()); |
1811 | } |
1812 | |
1813 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1814 | static bool classof(const Instruction *I) { |
1815 | return I->getOpcode() == Instruction::Select; |
1816 | } |
1817 | static bool classof(const Value *V) { |
1818 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1819 | } |
1820 | }; |
1821 | |
1822 | template <> |
1823 | struct OperandTraits<SelectInst> : public FixedNumOperandTraits<SelectInst, 3> { |
1824 | }; |
1825 | |
1826 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(SelectInst, Value)SelectInst::op_iterator SelectInst::op_begin() { return OperandTraits <SelectInst>::op_begin(this); } SelectInst::const_op_iterator SelectInst::op_begin() const { return OperandTraits<SelectInst >::op_begin(const_cast<SelectInst*>(this)); } SelectInst ::op_iterator SelectInst::op_end() { return OperandTraits< SelectInst>::op_end(this); } SelectInst::const_op_iterator SelectInst::op_end() const { return OperandTraits<SelectInst >::op_end(const_cast<SelectInst*>(this)); } Value *SelectInst ::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<SelectInst>::op_begin(const_cast <SelectInst*>(this))[i_nocapture].get()); } void SelectInst ::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (( void)0); OperandTraits<SelectInst>::op_begin(this)[i_nocapture ] = Val_nocapture; } unsigned SelectInst::getNumOperands() const { return OperandTraits<SelectInst>::operands(this); } template <int Idx_nocapture> Use &SelectInst::Op() { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &SelectInst::Op() const { return this->OpFrom<Idx_nocapture>(this); } |
1827 | |
1828 | //===----------------------------------------------------------------------===// |
1829 | // VAArgInst Class |
1830 | //===----------------------------------------------------------------------===// |
1831 | |
1832 | /// This class represents the va_arg llvm instruction, which returns |
1833 | /// an argument of the specified type given a va_list and increments that list |
1834 | /// |
1835 | class VAArgInst : public UnaryInstruction { |
1836 | protected: |
1837 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1838 | friend class Instruction; |
1839 | |
1840 | VAArgInst *cloneImpl() const; |
1841 | |
1842 | public: |
1843 | VAArgInst(Value *List, Type *Ty, const Twine &NameStr = "", |
1844 | Instruction *InsertBefore = nullptr) |
1845 | : UnaryInstruction(Ty, VAArg, List, InsertBefore) { |
1846 | setName(NameStr); |
1847 | } |
1848 | |
1849 | VAArgInst(Value *List, Type *Ty, const Twine &NameStr, |
1850 | BasicBlock *InsertAtEnd) |
1851 | : UnaryInstruction(Ty, VAArg, List, InsertAtEnd) { |
1852 | setName(NameStr); |
1853 | } |
1854 | |
1855 | Value *getPointerOperand() { return getOperand(0); } |
1856 | const Value *getPointerOperand() const { return getOperand(0); } |
1857 | static unsigned getPointerOperandIndex() { return 0U; } |
1858 | |
1859 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1860 | static bool classof(const Instruction *I) { |
1861 | return I->getOpcode() == VAArg; |
1862 | } |
1863 | static bool classof(const Value *V) { |
1864 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1865 | } |
1866 | }; |
1867 | |
1868 | //===----------------------------------------------------------------------===// |
1869 | // ExtractElementInst Class |
1870 | //===----------------------------------------------------------------------===// |
1871 | |
1872 | /// This instruction extracts a single (scalar) |
1873 | /// element from a VectorType value |
1874 | /// |
1875 | class ExtractElementInst : public Instruction { |
1876 | ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr = "", |
1877 | Instruction *InsertBefore = nullptr); |
1878 | ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr, |
1879 | BasicBlock *InsertAtEnd); |
1880 | |
1881 | protected: |
1882 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1883 | friend class Instruction; |
1884 | |
1885 | ExtractElementInst *cloneImpl() const; |
1886 | |
1887 | public: |
1888 | static ExtractElementInst *Create(Value *Vec, Value *Idx, |
1889 | const Twine &NameStr = "", |
1890 | Instruction *InsertBefore = nullptr) { |
1891 | return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertBefore); |
1892 | } |
1893 | |
1894 | static ExtractElementInst *Create(Value *Vec, Value *Idx, |
1895 | const Twine &NameStr, |
1896 | BasicBlock *InsertAtEnd) { |
1897 | return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertAtEnd); |
1898 | } |
1899 | |
1900 | /// Return true if an extractelement instruction can be |
1901 | /// formed with the specified operands. |
1902 | static bool isValidOperands(const Value *Vec, const Value *Idx); |
1903 | |
1904 | Value *getVectorOperand() { return Op<0>(); } |
1905 | Value *getIndexOperand() { return Op<1>(); } |
1906 | const Value *getVectorOperand() const { return Op<0>(); } |
1907 | const Value *getIndexOperand() const { return Op<1>(); } |
1908 | |
1909 | VectorType *getVectorOperandType() const { |
1910 | return cast<VectorType>(getVectorOperand()->getType()); |
1911 | } |
1912 | |
1913 | /// Transparently provide more efficient getOperand methods. |
1914 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
1915 | |
1916 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1917 | static bool classof(const Instruction *I) { |
1918 | return I->getOpcode() == Instruction::ExtractElement; |
1919 | } |
1920 | static bool classof(const Value *V) { |
1921 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1922 | } |
1923 | }; |
1924 | |
1925 | template <> |
1926 | struct OperandTraits<ExtractElementInst> : |
1927 | public FixedNumOperandTraits<ExtractElementInst, 2> { |
1928 | }; |
1929 | |
1930 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ExtractElementInst, Value)ExtractElementInst::op_iterator ExtractElementInst::op_begin( ) { return OperandTraits<ExtractElementInst>::op_begin( this); } ExtractElementInst::const_op_iterator ExtractElementInst ::op_begin() const { return OperandTraits<ExtractElementInst >::op_begin(const_cast<ExtractElementInst*>(this)); } ExtractElementInst::op_iterator ExtractElementInst::op_end() { return OperandTraits<ExtractElementInst>::op_end(this ); } ExtractElementInst::const_op_iterator ExtractElementInst ::op_end() const { return OperandTraits<ExtractElementInst >::op_end(const_cast<ExtractElementInst*>(this)); } Value *ExtractElementInst::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null<Value>( OperandTraits< ExtractElementInst>::op_begin(const_cast<ExtractElementInst *>(this))[i_nocapture].get()); } void ExtractElementInst:: setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((void )0); OperandTraits<ExtractElementInst>::op_begin(this)[ i_nocapture] = Val_nocapture; } unsigned ExtractElementInst:: getNumOperands() const { return OperandTraits<ExtractElementInst >::operands(this); } template <int Idx_nocapture> Use &ExtractElementInst::Op() { return this->OpFrom<Idx_nocapture >(this); } template <int Idx_nocapture> const Use & ExtractElementInst::Op() const { return this->OpFrom<Idx_nocapture >(this); } |
1931 | |
1932 | //===----------------------------------------------------------------------===// |
1933 | // InsertElementInst Class |
1934 | //===----------------------------------------------------------------------===// |
1935 | |
1936 | /// This instruction inserts a single (scalar) |
1937 | /// element into a VectorType value |
1938 | /// |
1939 | class InsertElementInst : public Instruction { |
1940 | InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, |
1941 | const Twine &NameStr = "", |
1942 | Instruction *InsertBefore = nullptr); |
1943 | InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, const Twine &NameStr, |
1944 | BasicBlock *InsertAtEnd); |
1945 | |
1946 | protected: |
1947 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1948 | friend class Instruction; |
1949 | |
1950 | InsertElementInst *cloneImpl() const; |
1951 | |
1952 | public: |
1953 | static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx, |
1954 | const Twine &NameStr = "", |
1955 | Instruction *InsertBefore = nullptr) { |
1956 | return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertBefore); |
1957 | } |
1958 | |
1959 | static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx, |
1960 | const Twine &NameStr, |
1961 | BasicBlock *InsertAtEnd) { |
1962 | return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertAtEnd); |
1963 | } |
1964 | |
1965 | /// Return true if an insertelement instruction can be |
1966 | /// formed with the specified operands. |
1967 | static bool isValidOperands(const Value *Vec, const Value *NewElt, |
1968 | const Value *Idx); |
1969 | |
1970 | /// Overload to return most specific vector type. |
1971 | /// |
1972 | VectorType *getType() const { |
1973 | return cast<VectorType>(Instruction::getType()); |
1974 | } |
1975 | |
1976 | /// Transparently provide more efficient getOperand methods. |
1977 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
1978 | |
1979 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1980 | static bool classof(const Instruction *I) { |
1981 | return I->getOpcode() == Instruction::InsertElement; |
1982 | } |
1983 | static bool classof(const Value *V) { |
1984 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1985 | } |
1986 | }; |
1987 | |
1988 | template <> |
1989 | struct OperandTraits<InsertElementInst> : |
1990 | public FixedNumOperandTraits<InsertElementInst, 3> { |
1991 | }; |
1992 | |
1993 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(InsertElementInst, Value)InsertElementInst::op_iterator InsertElementInst::op_begin() { return OperandTraits<InsertElementInst>::op_begin(this ); } InsertElementInst::const_op_iterator InsertElementInst:: op_begin() const { return OperandTraits<InsertElementInst> ::op_begin(const_cast<InsertElementInst*>(this)); } InsertElementInst ::op_iterator InsertElementInst::op_end() { return OperandTraits <InsertElementInst>::op_end(this); } InsertElementInst:: const_op_iterator InsertElementInst::op_end() const { return OperandTraits <InsertElementInst>::op_end(const_cast<InsertElementInst *>(this)); } Value *InsertElementInst::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null<Value >( OperandTraits<InsertElementInst>::op_begin(const_cast <InsertElementInst*>(this))[i_nocapture].get()); } void InsertElementInst::setOperand(unsigned i_nocapture, Value *Val_nocapture ) { ((void)0); OperandTraits<InsertElementInst>::op_begin (this)[i_nocapture] = Val_nocapture; } unsigned InsertElementInst ::getNumOperands() const { return OperandTraits<InsertElementInst >::operands(this); } template <int Idx_nocapture> Use &InsertElementInst::Op() { return this->OpFrom<Idx_nocapture >(this); } template <int Idx_nocapture> const Use & InsertElementInst::Op() const { return this->OpFrom<Idx_nocapture >(this); } |
1994 | |
1995 | //===----------------------------------------------------------------------===// |
1996 | // ShuffleVectorInst Class |
1997 | //===----------------------------------------------------------------------===// |
1998 | |
1999 | constexpr int UndefMaskElem = -1; |
2000 | |
2001 | /// This instruction constructs a fixed permutation of two |
2002 | /// input vectors. |
2003 | /// |
2004 | /// For each element of the result vector, the shuffle mask selects an element |
2005 | /// from one of the input vectors to copy to the result. Non-negative elements |
2006 | /// in the mask represent an index into the concatenated pair of input vectors. |
2007 | /// UndefMaskElem (-1) specifies that the result element is undefined. |
2008 | /// |
2009 | /// For scalable vectors, all the elements of the mask must be 0 or -1. This |
2010 | /// requirement may be relaxed in the future. |
2011 | class ShuffleVectorInst : public Instruction { |
2012 | SmallVector<int, 4> ShuffleMask; |
2013 | Constant *ShuffleMaskForBitcode; |
2014 | |
2015 | protected: |
2016 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2017 | friend class Instruction; |
2018 | |
2019 | ShuffleVectorInst *cloneImpl() const; |
2020 | |
2021 | public: |
2022 | ShuffleVectorInst(Value *V1, Value *V2, Value *Mask, |
2023 | const Twine &NameStr = "", |
2024 | Instruction *InsertBefor = nullptr); |
2025 | ShuffleVectorInst(Value *V1, Value *V2, Value *Mask, |
2026 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2027 | ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask, |
2028 | const Twine &NameStr = "", |
2029 | Instruction *InsertBefor = nullptr); |
2030 | ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask, |
2031 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2032 | |
2033 | void *operator new(size_t S) { return User::operator new(S, 2); } |
2034 | void operator delete(void *Ptr) { return User::operator delete(Ptr); } |
2035 | |
2036 | /// Swap the operands and adjust the mask to preserve the semantics |
2037 | /// of the instruction. |
2038 | void commute(); |
2039 | |
2040 | /// Return true if a shufflevector instruction can be |
2041 | /// formed with the specified operands. |
2042 | static bool isValidOperands(const Value *V1, const Value *V2, |
2043 | const Value *Mask); |
2044 | static bool isValidOperands(const Value *V1, const Value *V2, |
2045 | ArrayRef<int> Mask); |
2046 | |
2047 | /// Overload to return most specific vector type. |
2048 | /// |
2049 | VectorType *getType() const { |
2050 | return cast<VectorType>(Instruction::getType()); |
2051 | } |
2052 | |
2053 | /// Transparently provide more efficient getOperand methods. |
2054 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
2055 | |
2056 | /// Return the shuffle mask value of this instruction for the given element |
2057 | /// index. Return UndefMaskElem if the element is undef. |
2058 | int getMaskValue(unsigned Elt) const { return ShuffleMask[Elt]; } |
2059 | |
2060 | /// Convert the input shuffle mask operand to a vector of integers. Undefined |
2061 | /// elements of the mask are returned as UndefMaskElem. |
2062 | static void getShuffleMask(const Constant *Mask, |
2063 | SmallVectorImpl<int> &Result); |
2064 | |
2065 | /// Return the mask for this instruction as a vector of integers. Undefined |
2066 | /// elements of the mask are returned as UndefMaskElem. |
2067 | void getShuffleMask(SmallVectorImpl<int> &Result) const { |
2068 | Result.assign(ShuffleMask.begin(), ShuffleMask.end()); |
2069 | } |
2070 | |
2071 | /// Return the mask for this instruction, for use in bitcode. |
2072 | /// |
2073 | /// TODO: This is temporary until we decide a new bitcode encoding for |
2074 | /// shufflevector. |
2075 | Constant *getShuffleMaskForBitcode() const { return ShuffleMaskForBitcode; } |
2076 | |
2077 | static Constant *convertShuffleMaskForBitcode(ArrayRef<int> Mask, |
2078 | Type *ResultTy); |
2079 | |
2080 | void setShuffleMask(ArrayRef<int> Mask); |
2081 | |
2082 | ArrayRef<int> getShuffleMask() const { return ShuffleMask; } |
2083 | |
2084 | /// Return true if this shuffle returns a vector with a different number of |
2085 | /// elements than its source vectors. |
2086 | /// Examples: shufflevector <4 x n> A, <4 x n> B, <1,2,3> |
2087 | /// shufflevector <4 x n> A, <4 x n> B, <1,2,3,4,5> |
2088 | bool changesLength() const { |
2089 | unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType()) |
2090 | ->getElementCount() |
2091 | .getKnownMinValue(); |
2092 | unsigned NumMaskElts = ShuffleMask.size(); |
2093 | return NumSourceElts != NumMaskElts; |
2094 | } |
2095 | |
2096 | /// Return true if this shuffle returns a vector with a greater number of |
2097 | /// elements than its source vectors. |
2098 | /// Example: shufflevector <2 x n> A, <2 x n> B, <1,2,3> |
2099 | bool increasesLength() const { |
2100 | unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType()) |
2101 | ->getElementCount() |
2102 | .getKnownMinValue(); |
2103 | unsigned NumMaskElts = ShuffleMask.size(); |
2104 | return NumSourceElts < NumMaskElts; |
2105 | } |
2106 | |
2107 | /// Return true if this shuffle mask chooses elements from exactly one source |
2108 | /// vector. |
2109 | /// Example: <7,5,undef,7> |
2110 | /// This assumes that vector operands are the same length as the mask. |
2111 | static bool isSingleSourceMask(ArrayRef<int> Mask); |
2112 | static bool isSingleSourceMask(const Constant *Mask) { |
2113 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((void)0); |
2114 | SmallVector<int, 16> MaskAsInts; |
2115 | getShuffleMask(Mask, MaskAsInts); |
2116 | return isSingleSourceMask(MaskAsInts); |
2117 | } |
2118 | |
2119 | /// Return true if this shuffle chooses elements from exactly one source |
2120 | /// vector without changing the length of that vector. |
2121 | /// Example: shufflevector <4 x n> A, <4 x n> B, <3,0,undef,3> |
2122 | /// TODO: Optionally allow length-changing shuffles. |
2123 | bool isSingleSource() const { |
2124 | return !changesLength() && isSingleSourceMask(ShuffleMask); |
2125 | } |
2126 | |
2127 | /// Return true if this shuffle mask chooses elements from exactly one source |
2128 | /// vector without lane crossings. A shuffle using this mask is not |
2129 | /// necessarily a no-op because it may change the number of elements from its |
2130 | /// input vectors or it may provide demanded bits knowledge via undef lanes. |
2131 | /// Example: <undef,undef,2,3> |
2132 | static bool isIdentityMask(ArrayRef<int> Mask); |
2133 | static bool isIdentityMask(const Constant *Mask) { |
2134 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((void)0); |
2135 | SmallVector<int, 16> MaskAsInts; |
2136 | getShuffleMask(Mask, MaskAsInts); |
2137 | return isIdentityMask(MaskAsInts); |
2138 | } |
2139 | |
2140 | /// Return true if this shuffle chooses elements from exactly one source |
2141 | /// vector without lane crossings and does not change the number of elements |
2142 | /// from its input vectors. |
2143 | /// Example: shufflevector <4 x n> A, <4 x n> B, <4,undef,6,undef> |
2144 | bool isIdentity() const { |
2145 | return !changesLength() && isIdentityMask(ShuffleMask); |
2146 | } |
2147 | |
2148 | /// Return true if this shuffle lengthens exactly one source vector with |
2149 | /// undefs in the high elements. |
2150 | bool isIdentityWithPadding() const; |
2151 | |
2152 | /// Return true if this shuffle extracts the first N elements of exactly one |
2153 | /// source vector. |
2154 | bool isIdentityWithExtract() const; |
2155 | |
2156 | /// Return true if this shuffle concatenates its 2 source vectors. This |
2157 | /// returns false if either input is undefined. In that case, the shuffle is |
2158 | /// is better classified as an identity with padding operation. |
2159 | bool isConcat() const; |
2160 | |
2161 | /// Return true if this shuffle mask chooses elements from its source vectors |
2162 | /// without lane crossings. A shuffle using this mask would be |
2163 | /// equivalent to a vector select with a constant condition operand. |
2164 | /// Example: <4,1,6,undef> |
2165 | /// This returns false if the mask does not choose from both input vectors. |
2166 | /// In that case, the shuffle is better classified as an identity shuffle. |
2167 | /// This assumes that vector operands are the same length as the mask |
2168 | /// (a length-changing shuffle can never be equivalent to a vector select). |
2169 | static bool isSelectMask(ArrayRef<int> Mask); |
2170 | static bool isSelectMask(const Constant *Mask) { |
2171 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((void)0); |
2172 | SmallVector<int, 16> MaskAsInts; |
2173 | getShuffleMask(Mask, MaskAsInts); |
2174 | return isSelectMask(MaskAsInts); |
2175 | } |
2176 | |
2177 | /// Return true if this shuffle chooses elements from its source vectors |
2178 | /// without lane crossings and all operands have the same number of elements. |
2179 | /// In other words, this shuffle is equivalent to a vector select with a |
2180 | /// constant condition operand. |
2181 | /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,1,6,3> |
2182 | /// This returns false if the mask does not choose from both input vectors. |
2183 | /// In that case, the shuffle is better classified as an identity shuffle. |
2184 | /// TODO: Optionally allow length-changing shuffles. |
2185 | bool isSelect() const { |
2186 | return !changesLength() && isSelectMask(ShuffleMask); |
2187 | } |
2188 | |
2189 | /// Return true if this shuffle mask swaps the order of elements from exactly |
2190 | /// one source vector. |
2191 | /// Example: <7,6,undef,4> |
2192 | /// This assumes that vector operands are the same length as the mask. |
2193 | static bool isReverseMask(ArrayRef<int> Mask); |
2194 | static bool isReverseMask(const Constant *Mask) { |
2195 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((void)0); |
2196 | SmallVector<int, 16> MaskAsInts; |
2197 | getShuffleMask(Mask, MaskAsInts); |
2198 | return isReverseMask(MaskAsInts); |
2199 | } |
2200 | |
2201 | /// Return true if this shuffle swaps the order of elements from exactly |
2202 | /// one source vector. |
2203 | /// Example: shufflevector <4 x n> A, <4 x n> B, <3,undef,1,undef> |
2204 | /// TODO: Optionally allow length-changing shuffles. |
2205 | bool isReverse() const { |
2206 | return !changesLength() && isReverseMask(ShuffleMask); |
2207 | } |
2208 | |
2209 | /// Return true if this shuffle mask chooses all elements with the same value |
2210 | /// as the first element of exactly one source vector. |
2211 | /// Example: <4,undef,undef,4> |
2212 | /// This assumes that vector operands are the same length as the mask. |
2213 | static bool isZeroEltSplatMask(ArrayRef<int> Mask); |
2214 | static bool isZeroEltSplatMask(const Constant *Mask) { |
2215 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((void)0); |
2216 | SmallVector<int, 16> MaskAsInts; |
2217 | getShuffleMask(Mask, MaskAsInts); |
2218 | return isZeroEltSplatMask(MaskAsInts); |
2219 | } |
2220 | |
2221 | /// Return true if all elements of this shuffle are the same value as the |
2222 | /// first element of exactly one source vector without changing the length |
2223 | /// of that vector. |
2224 | /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,0,undef,0> |
2225 | /// TODO: Optionally allow length-changing shuffles. |
2226 | /// TODO: Optionally allow splats from other elements. |
2227 | bool isZeroEltSplat() const { |
2228 | return !changesLength() && isZeroEltSplatMask(ShuffleMask); |
2229 | } |
2230 | |
2231 | /// Return true if this shuffle mask is a transpose mask. |
2232 | /// Transpose vector masks transpose a 2xn matrix. They read corresponding |
2233 | /// even- or odd-numbered vector elements from two n-dimensional source |
2234 | /// vectors and write each result into consecutive elements of an |
2235 | /// n-dimensional destination vector. Two shuffles are necessary to complete |
2236 | /// the transpose, one for the even elements and another for the odd elements. |
2237 | /// This description closely follows how the TRN1 and TRN2 AArch64 |
2238 | /// instructions operate. |
2239 | /// |
2240 | /// For example, a simple 2x2 matrix can be transposed with: |
2241 | /// |
2242 | /// ; Original matrix |
2243 | /// m0 = < a, b > |
2244 | /// m1 = < c, d > |
2245 | /// |
2246 | /// ; Transposed matrix |
2247 | /// t0 = < a, c > = shufflevector m0, m1, < 0, 2 > |
2248 | /// t1 = < b, d > = shufflevector m0, m1, < 1, 3 > |
2249 | /// |
2250 | /// For matrices having greater than n columns, the resulting nx2 transposed |
2251 | /// matrix is stored in two result vectors such that one vector contains |
2252 | /// interleaved elements from all the even-numbered rows and the other vector |
2253 | /// contains interleaved elements from all the odd-numbered rows. For example, |
2254 | /// a 2x4 matrix can be transposed with: |
2255 | /// |
2256 | /// ; Original matrix |
2257 | /// m0 = < a, b, c, d > |
2258 | /// m1 = < e, f, g, h > |
2259 | /// |
2260 | /// ; Transposed matrix |
2261 | /// t0 = < a, e, c, g > = shufflevector m0, m1 < 0, 4, 2, 6 > |
2262 | /// t1 = < b, f, d, h > = shufflevector m0, m1 < 1, 5, 3, 7 > |
2263 | static bool isTransposeMask(ArrayRef<int> Mask); |
2264 | static bool isTransposeMask(const Constant *Mask) { |
2265 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((void)0); |
2266 | SmallVector<int, 16> MaskAsInts; |
2267 | getShuffleMask(Mask, MaskAsInts); |
2268 | return isTransposeMask(MaskAsInts); |
2269 | } |
2270 | |
2271 | /// Return true if this shuffle transposes the elements of its inputs without |
2272 | /// changing the length of the vectors. This operation may also be known as a |
2273 | /// merge or interleave. See the description for isTransposeMask() for the |
2274 | /// exact specification. |
2275 | /// Example: shufflevector <4 x n> A, <4 x n> B, <0,4,2,6> |
2276 | bool isTranspose() const { |
2277 | return !changesLength() && isTransposeMask(ShuffleMask); |
2278 | } |
2279 | |
2280 | /// Return true if this shuffle mask is an extract subvector mask. |
2281 | /// A valid extract subvector mask returns a smaller vector from a single |
2282 | /// source operand. The base extraction index is returned as well. |
2283 | static bool isExtractSubvectorMask(ArrayRef<int> Mask, int NumSrcElts, |
2284 | int &Index); |
2285 | static bool isExtractSubvectorMask(const Constant *Mask, int NumSrcElts, |
2286 | int &Index) { |
2287 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((void)0); |
2288 | // Not possible to express a shuffle mask for a scalable vector for this |
2289 | // case. |
2290 | if (isa<ScalableVectorType>(Mask->getType())) |
2291 | return false; |
2292 | SmallVector<int, 16> MaskAsInts; |
2293 | getShuffleMask(Mask, MaskAsInts); |
2294 | return isExtractSubvectorMask(MaskAsInts, NumSrcElts, Index); |
2295 | } |
2296 | |
2297 | /// Return true if this shuffle mask is an extract subvector mask. |
2298 | bool isExtractSubvectorMask(int &Index) const { |
2299 | // Not possible to express a shuffle mask for a scalable vector for this |
2300 | // case. |
2301 | if (isa<ScalableVectorType>(getType())) |
2302 | return false; |
2303 | |
2304 | int NumSrcElts = |
2305 | cast<FixedVectorType>(Op<0>()->getType())->getNumElements(); |
2306 | return isExtractSubvectorMask(ShuffleMask, NumSrcElts, Index); |
2307 | } |
2308 | |
2309 | /// Change values in a shuffle permute mask assuming the two vector operands |
2310 | /// of length InVecNumElts have swapped position. |
2311 | static void commuteShuffleMask(MutableArrayRef<int> Mask, |
2312 | unsigned InVecNumElts) { |
2313 | for (int &Idx : Mask) { |
2314 | if (Idx == -1) |
2315 | continue; |
2316 | Idx = Idx < (int)InVecNumElts ? Idx + InVecNumElts : Idx - InVecNumElts; |
2317 | assert(Idx >= 0 && Idx < (int)InVecNumElts * 2 &&((void)0) |
2318 | "shufflevector mask index out of range")((void)0); |
2319 | } |
2320 | } |
2321 | |
2322 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2323 | static bool classof(const Instruction *I) { |
2324 | return I->getOpcode() == Instruction::ShuffleVector; |
2325 | } |
2326 | static bool classof(const Value *V) { |
2327 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2328 | } |
2329 | }; |
2330 | |
2331 | template <> |
2332 | struct OperandTraits<ShuffleVectorInst> |
2333 | : public FixedNumOperandTraits<ShuffleVectorInst, 2> {}; |
2334 | |
2335 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ShuffleVectorInst, Value)ShuffleVectorInst::op_iterator ShuffleVectorInst::op_begin() { return OperandTraits<ShuffleVectorInst>::op_begin(this ); } ShuffleVectorInst::const_op_iterator ShuffleVectorInst:: op_begin() const { return OperandTraits<ShuffleVectorInst> ::op_begin(const_cast<ShuffleVectorInst*>(this)); } ShuffleVectorInst ::op_iterator ShuffleVectorInst::op_end() { return OperandTraits <ShuffleVectorInst>::op_end(this); } ShuffleVectorInst:: const_op_iterator ShuffleVectorInst::op_end() const { return OperandTraits <ShuffleVectorInst>::op_end(const_cast<ShuffleVectorInst *>(this)); } Value *ShuffleVectorInst::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null<Value >( OperandTraits<ShuffleVectorInst>::op_begin(const_cast <ShuffleVectorInst*>(this))[i_nocapture].get()); } void ShuffleVectorInst::setOperand(unsigned i_nocapture, Value *Val_nocapture ) { ((void)0); OperandTraits<ShuffleVectorInst>::op_begin (this)[i_nocapture] = Val_nocapture; } unsigned ShuffleVectorInst ::getNumOperands() const { return OperandTraits<ShuffleVectorInst >::operands(this); } template <int Idx_nocapture> Use &ShuffleVectorInst::Op() { return this->OpFrom<Idx_nocapture >(this); } template <int Idx_nocapture> const Use & ShuffleVectorInst::Op() const { return this->OpFrom<Idx_nocapture >(this); } |
2336 | |
2337 | //===----------------------------------------------------------------------===// |
2338 | // ExtractValueInst Class |
2339 | //===----------------------------------------------------------------------===// |
2340 | |
2341 | /// This instruction extracts a struct member or array |
2342 | /// element value from an aggregate value. |
2343 | /// |
2344 | class ExtractValueInst : public UnaryInstruction { |
2345 | SmallVector<unsigned, 4> Indices; |
2346 | |
2347 | ExtractValueInst(const ExtractValueInst &EVI); |
2348 | |
2349 | /// Constructors - Create a extractvalue instruction with a base aggregate |
2350 | /// value and a list of indices. The first ctor can optionally insert before |
2351 | /// an existing instruction, the second appends the new instruction to the |
2352 | /// specified BasicBlock. |
2353 | inline ExtractValueInst(Value *Agg, |
2354 | ArrayRef<unsigned> Idxs, |
2355 | const Twine &NameStr, |
2356 | Instruction *InsertBefore); |
2357 | inline ExtractValueInst(Value *Agg, |
2358 | ArrayRef<unsigned> Idxs, |
2359 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2360 | |
2361 | void init(ArrayRef<unsigned> Idxs, const Twine &NameStr); |
2362 | |
2363 | protected: |
2364 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2365 | friend class Instruction; |
2366 | |
2367 | ExtractValueInst *cloneImpl() const; |
2368 | |
2369 | public: |
2370 | static ExtractValueInst *Create(Value *Agg, |
2371 | ArrayRef<unsigned> Idxs, |
2372 | const Twine &NameStr = "", |
2373 | Instruction *InsertBefore = nullptr) { |
2374 | return new |
2375 | ExtractValueInst(Agg, Idxs, NameStr, InsertBefore); |
2376 | } |
2377 | |
2378 | static ExtractValueInst *Create(Value *Agg, |
2379 | ArrayRef<unsigned> Idxs, |
2380 | const Twine &NameStr, |
2381 | BasicBlock *InsertAtEnd) { |
2382 | return new ExtractValueInst(Agg, Idxs, NameStr, InsertAtEnd); |
2383 | } |
2384 | |
2385 | /// Returns the type of the element that would be extracted |
2386 | /// with an extractvalue instruction with the specified parameters. |
2387 | /// |
2388 | /// Null is returned if the indices are invalid for the specified type. |
2389 | static Type *getIndexedType(Type *Agg, ArrayRef<unsigned> Idxs); |
2390 | |
2391 | using idx_iterator = const unsigned*; |
2392 | |
2393 | inline idx_iterator idx_begin() const { return Indices.begin(); } |
2394 | inline idx_iterator idx_end() const { return Indices.end(); } |
2395 | inline iterator_range<idx_iterator> indices() const { |
2396 | return make_range(idx_begin(), idx_end()); |
2397 | } |
2398 | |
2399 | Value *getAggregateOperand() { |
2400 | return getOperand(0); |
2401 | } |
2402 | const Value *getAggregateOperand() const { |
2403 | return getOperand(0); |
2404 | } |
2405 | static unsigned getAggregateOperandIndex() { |
2406 | return 0U; // get index for modifying correct operand |
2407 | } |
2408 | |
2409 | ArrayRef<unsigned> getIndices() const { |
2410 | return Indices; |
2411 | } |
2412 | |
2413 | unsigned getNumIndices() const { |
2414 | return (unsigned)Indices.size(); |
2415 | } |
2416 | |
2417 | bool hasIndices() const { |
2418 | return true; |
2419 | } |
2420 | |
2421 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2422 | static bool classof(const Instruction *I) { |
2423 | return I->getOpcode() == Instruction::ExtractValue; |
2424 | } |
2425 | static bool classof(const Value *V) { |
2426 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2427 | } |
2428 | }; |
2429 | |
2430 | ExtractValueInst::ExtractValueInst(Value *Agg, |
2431 | ArrayRef<unsigned> Idxs, |
2432 | const Twine &NameStr, |
2433 | Instruction *InsertBefore) |
2434 | : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)), |
2435 | ExtractValue, Agg, InsertBefore) { |
2436 | init(Idxs, NameStr); |
2437 | } |
2438 | |
2439 | ExtractValueInst::ExtractValueInst(Value *Agg, |
2440 | ArrayRef<unsigned> Idxs, |
2441 | const Twine &NameStr, |
2442 | BasicBlock *InsertAtEnd) |
2443 | : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)), |
2444 | ExtractValue, Agg, InsertAtEnd) { |
2445 | init(Idxs, NameStr); |
2446 | } |
2447 | |
2448 | //===----------------------------------------------------------------------===// |
2449 | // InsertValueInst Class |
2450 | //===----------------------------------------------------------------------===// |
2451 | |
2452 | /// This instruction inserts a struct field of array element |
2453 | /// value into an aggregate value. |
2454 | /// |
2455 | class InsertValueInst : public Instruction { |
2456 | SmallVector<unsigned, 4> Indices; |
2457 | |
2458 | InsertValueInst(const InsertValueInst &IVI); |
2459 | |
2460 | /// Constructors - Create a insertvalue instruction with a base aggregate |
2461 | /// value, a value to insert, and a list of indices. The first ctor can |
2462 | /// optionally insert before an existing instruction, the second appends |
2463 | /// the new instruction to the specified BasicBlock. |
2464 | inline InsertValueInst(Value *Agg, Value *Val, |
2465 | ArrayRef<unsigned> Idxs, |
2466 | const Twine &NameStr, |
2467 | Instruction *InsertBefore); |
2468 | inline InsertValueInst(Value *Agg, Value *Val, |
2469 | ArrayRef<unsigned> Idxs, |
2470 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2471 | |
2472 | /// Constructors - These two constructors are convenience methods because one |
2473 | /// and two index insertvalue instructions are so common. |
2474 | InsertValueInst(Value *Agg, Value *Val, unsigned Idx, |
2475 | const Twine &NameStr = "", |
2476 | Instruction *InsertBefore = nullptr); |
2477 | InsertValueInst(Value *Agg, Value *Val, unsigned Idx, const Twine &NameStr, |
2478 | BasicBlock *InsertAtEnd); |
2479 | |
2480 | void init(Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, |
2481 | const Twine &NameStr); |
2482 | |
2483 | protected: |
2484 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2485 | friend class Instruction; |
2486 | |
2487 | InsertValueInst *cloneImpl() const; |
2488 | |
2489 | public: |
2490 | // allocate space for exactly two operands |
2491 | void *operator new(size_t S) { return User::operator new(S, 2); } |
2492 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
2493 | |
2494 | static InsertValueInst *Create(Value *Agg, Value *Val, |
2495 | ArrayRef<unsigned> Idxs, |
2496 | const Twine &NameStr = "", |
2497 | Instruction *InsertBefore = nullptr) { |
2498 | return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertBefore); |
2499 | } |
2500 | |
2501 | static InsertValueInst *Create(Value *Agg, Value *Val, |
2502 | ArrayRef<unsigned> Idxs, |
2503 | const Twine &NameStr, |
2504 | BasicBlock *InsertAtEnd) { |
2505 | return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertAtEnd); |
2506 | } |
2507 | |
2508 | /// Transparently provide more efficient getOperand methods. |
2509 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
2510 | |
2511 | using idx_iterator = const unsigned*; |
2512 | |
2513 | inline idx_iterator idx_begin() const { return Indices.begin(); } |
2514 | inline idx_iterator idx_end() const { return Indices.end(); } |
2515 | inline iterator_range<idx_iterator> indices() const { |
2516 | return make_range(idx_begin(), idx_end()); |
2517 | } |
2518 | |
2519 | Value *getAggregateOperand() { |
2520 | return getOperand(0); |
2521 | } |
2522 | const Value *getAggregateOperand() const { |
2523 | return getOperand(0); |
2524 | } |
2525 | static unsigned getAggregateOperandIndex() { |
2526 | return 0U; // get index for modifying correct operand |
2527 | } |
2528 | |
2529 | Value *getInsertedValueOperand() { |
2530 | return getOperand(1); |
2531 | } |
2532 | const Value *getInsertedValueOperand() const { |
2533 | return getOperand(1); |
2534 | } |
2535 | static unsigned getInsertedValueOperandIndex() { |
2536 | return 1U; // get index for modifying correct operand |
2537 | } |
2538 | |
2539 | ArrayRef<unsigned> getIndices() const { |
2540 | return Indices; |
2541 | } |
2542 | |
2543 | unsigned getNumIndices() const { |
2544 | return (unsigned)Indices.size(); |
2545 | } |
2546 | |
2547 | bool hasIndices() const { |
2548 | return true; |
2549 | } |
2550 | |
2551 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2552 | static bool classof(const Instruction *I) { |
2553 | return I->getOpcode() == Instruction::InsertValue; |
2554 | } |
2555 | static bool classof(const Value *V) { |
2556 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2557 | } |
2558 | }; |
2559 | |
2560 | template <> |
2561 | struct OperandTraits<InsertValueInst> : |
2562 | public FixedNumOperandTraits<InsertValueInst, 2> { |
2563 | }; |
2564 | |
2565 | InsertValueInst::InsertValueInst(Value *Agg, |
2566 | Value *Val, |
2567 | ArrayRef<unsigned> Idxs, |
2568 | const Twine &NameStr, |
2569 | Instruction *InsertBefore) |
2570 | : Instruction(Agg->getType(), InsertValue, |
2571 | OperandTraits<InsertValueInst>::op_begin(this), |
2572 | 2, InsertBefore) { |
2573 | init(Agg, Val, Idxs, NameStr); |
2574 | } |
2575 | |
2576 | InsertValueInst::InsertValueInst(Value *Agg, |
2577 | Value *Val, |
2578 | ArrayRef<unsigned> Idxs, |
2579 | const Twine &NameStr, |
2580 | BasicBlock *InsertAtEnd) |
2581 | : Instruction(Agg->getType(), InsertValue, |
2582 | OperandTraits<InsertValueInst>::op_begin(this), |
2583 | 2, InsertAtEnd) { |
2584 | init(Agg, Val, Idxs, NameStr); |
2585 | } |
2586 | |
2587 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(InsertValueInst, Value)InsertValueInst::op_iterator InsertValueInst::op_begin() { return OperandTraits<InsertValueInst>::op_begin(this); } InsertValueInst ::const_op_iterator InsertValueInst::op_begin() const { return OperandTraits<InsertValueInst>::op_begin(const_cast< InsertValueInst*>(this)); } InsertValueInst::op_iterator InsertValueInst ::op_end() { return OperandTraits<InsertValueInst>::op_end (this); } InsertValueInst::const_op_iterator InsertValueInst:: op_end() const { return OperandTraits<InsertValueInst>:: op_end(const_cast<InsertValueInst*>(this)); } Value *InsertValueInst ::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<InsertValueInst>::op_begin (const_cast<InsertValueInst*>(this))[i_nocapture].get() ); } void InsertValueInst::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((void)0); OperandTraits<InsertValueInst >::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned InsertValueInst::getNumOperands() const { return OperandTraits <InsertValueInst>::operands(this); } template <int Idx_nocapture > Use &InsertValueInst::Op() { return this->OpFrom< Idx_nocapture>(this); } template <int Idx_nocapture> const Use &InsertValueInst::Op() const { return this-> OpFrom<Idx_nocapture>(this); } |
2588 | |
2589 | //===----------------------------------------------------------------------===// |
2590 | // PHINode Class |
2591 | //===----------------------------------------------------------------------===// |
2592 | |
2593 | // PHINode - The PHINode class is used to represent the magical mystical PHI |
2594 | // node, that can not exist in nature, but can be synthesized in a computer |
2595 | // scientist's overactive imagination. |
2596 | // |
2597 | class PHINode : public Instruction { |
2598 | /// The number of operands actually allocated. NumOperands is |
2599 | /// the number actually in use. |
2600 | unsigned ReservedSpace; |
2601 | |
2602 | PHINode(const PHINode &PN); |
2603 | |
2604 | explicit PHINode(Type *Ty, unsigned NumReservedValues, |
2605 | const Twine &NameStr = "", |
2606 | Instruction *InsertBefore = nullptr) |
2607 | : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertBefore), |
2608 | ReservedSpace(NumReservedValues) { |
2609 | assert(!Ty->isTokenTy() && "PHI nodes cannot have token type!")((void)0); |
2610 | setName(NameStr); |
2611 | allocHungoffUses(ReservedSpace); |
2612 | } |
2613 | |
2614 | PHINode(Type *Ty, unsigned NumReservedValues, const Twine &NameStr, |
2615 | BasicBlock *InsertAtEnd) |
2616 | : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertAtEnd), |
2617 | ReservedSpace(NumReservedValues) { |
2618 | assert(!Ty->isTokenTy() && "PHI nodes cannot have token type!")((void)0); |
2619 | setName(NameStr); |
2620 | allocHungoffUses(ReservedSpace); |
2621 | } |
2622 | |
2623 | protected: |
2624 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2625 | friend class Instruction; |
2626 | |
2627 | PHINode *cloneImpl() const; |
2628 | |
2629 | // allocHungoffUses - this is more complicated than the generic |
2630 | // User::allocHungoffUses, because we have to allocate Uses for the incoming |
2631 | // values and pointers to the incoming blocks, all in one allocation. |
2632 | void allocHungoffUses(unsigned N) { |
2633 | User::allocHungoffUses(N, /* IsPhi */ true); |
2634 | } |
2635 | |
2636 | public: |
2637 | /// Constructors - NumReservedValues is a hint for the number of incoming |
2638 | /// edges that this phi node will have (use 0 if you really have no idea). |
2639 | static PHINode *Create(Type *Ty, unsigned NumReservedValues, |
2640 | const Twine &NameStr = "", |
2641 | Instruction *InsertBefore = nullptr) { |
2642 | return new PHINode(Ty, NumReservedValues, NameStr, InsertBefore); |
2643 | } |
2644 | |
2645 | static PHINode *Create(Type *Ty, unsigned NumReservedValues, |
2646 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
2647 | return new PHINode(Ty, NumReservedValues, NameStr, InsertAtEnd); |
2648 | } |
2649 | |
2650 | /// Provide fast operand accessors |
2651 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
2652 | |
2653 | // Block iterator interface. This provides access to the list of incoming |
2654 | // basic blocks, which parallels the list of incoming values. |
2655 | |
2656 | using block_iterator = BasicBlock **; |
2657 | using const_block_iterator = BasicBlock * const *; |
2658 | |
2659 | block_iterator block_begin() { |
2660 | return reinterpret_cast<block_iterator>(op_begin() + ReservedSpace); |
2661 | } |
2662 | |
2663 | const_block_iterator block_begin() const { |
2664 | return reinterpret_cast<const_block_iterator>(op_begin() + ReservedSpace); |
2665 | } |
2666 | |
2667 | block_iterator block_end() { |
2668 | return block_begin() + getNumOperands(); |
2669 | } |
2670 | |
2671 | const_block_iterator block_end() const { |
2672 | return block_begin() + getNumOperands(); |
2673 | } |
2674 | |
2675 | iterator_range<block_iterator> blocks() { |
2676 | return make_range(block_begin(), block_end()); |
2677 | } |
2678 | |
2679 | iterator_range<const_block_iterator> blocks() const { |
2680 | return make_range(block_begin(), block_end()); |
2681 | } |
2682 | |
2683 | op_range incoming_values() { return operands(); } |
2684 | |
2685 | const_op_range incoming_values() const { return operands(); } |
2686 | |
2687 | /// Return the number of incoming edges |
2688 | /// |
2689 | unsigned getNumIncomingValues() const { return getNumOperands(); } |
2690 | |
2691 | /// Return incoming value number x |
2692 | /// |
2693 | Value *getIncomingValue(unsigned i) const { |
2694 | return getOperand(i); |
2695 | } |
2696 | void setIncomingValue(unsigned i, Value *V) { |
2697 | assert(V && "PHI node got a null value!")((void)0); |
2698 | assert(getType() == V->getType() &&((void)0) |
2699 | "All operands to PHI node must be the same type as the PHI node!")((void)0); |
2700 | setOperand(i, V); |
2701 | } |
2702 | |
2703 | static unsigned getOperandNumForIncomingValue(unsigned i) { |
2704 | return i; |
2705 | } |
2706 | |
2707 | static unsigned getIncomingValueNumForOperand(unsigned i) { |
2708 | return i; |
2709 | } |
2710 | |
2711 | /// Return incoming basic block number @p i. |
2712 | /// |
2713 | BasicBlock *getIncomingBlock(unsigned i) const { |
2714 | return block_begin()[i]; |
2715 | } |
2716 | |
2717 | /// Return incoming basic block corresponding |
2718 | /// to an operand of the PHI. |
2719 | /// |
2720 | BasicBlock *getIncomingBlock(const Use &U) const { |
2721 | assert(this == U.getUser() && "Iterator doesn't point to PHI's Uses?")((void)0); |
2722 | return getIncomingBlock(unsigned(&U - op_begin())); |
2723 | } |
2724 | |
2725 | /// Return incoming basic block corresponding |
2726 | /// to value use iterator. |
2727 | /// |
2728 | BasicBlock *getIncomingBlock(Value::const_user_iterator I) const { |
2729 | return getIncomingBlock(I.getUse()); |
2730 | } |
2731 | |
2732 | void setIncomingBlock(unsigned i, BasicBlock *BB) { |
2733 | assert(BB && "PHI node got a null basic block!")((void)0); |
2734 | block_begin()[i] = BB; |
2735 | } |
2736 | |
2737 | /// Replace every incoming basic block \p Old to basic block \p New. |
2738 | void replaceIncomingBlockWith(const BasicBlock *Old, BasicBlock *New) { |
2739 | assert(New && Old && "PHI node got a null basic block!")((void)0); |
2740 | for (unsigned Op = 0, NumOps = getNumOperands(); Op != NumOps; ++Op) |
2741 | if (getIncomingBlock(Op) == Old) |
2742 | setIncomingBlock(Op, New); |
2743 | } |
2744 | |
2745 | /// Add an incoming value to the end of the PHI list |
2746 | /// |
2747 | void addIncoming(Value *V, BasicBlock *BB) { |
2748 | if (getNumOperands() == ReservedSpace) |
2749 | growOperands(); // Get more space! |
2750 | // Initialize some new operands. |
2751 | setNumHungOffUseOperands(getNumOperands() + 1); |
2752 | setIncomingValue(getNumOperands() - 1, V); |
2753 | setIncomingBlock(getNumOperands() - 1, BB); |
2754 | } |
2755 | |
2756 | /// Remove an incoming value. This is useful if a |
2757 | /// predecessor basic block is deleted. The value removed is returned. |
2758 | /// |
2759 | /// If the last incoming value for a PHI node is removed (and DeletePHIIfEmpty |
2760 | /// is true), the PHI node is destroyed and any uses of it are replaced with |
2761 | /// dummy values. The only time there should be zero incoming values to a PHI |
2762 | /// node is when the block is dead, so this strategy is sound. |
2763 | /// |
2764 | Value *removeIncomingValue(unsigned Idx, bool DeletePHIIfEmpty = true); |
2765 | |
2766 | Value *removeIncomingValue(const BasicBlock *BB, bool DeletePHIIfEmpty=true) { |
2767 | int Idx = getBasicBlockIndex(BB); |
2768 | assert(Idx >= 0 && "Invalid basic block argument to remove!")((void)0); |
2769 | return removeIncomingValue(Idx, DeletePHIIfEmpty); |
2770 | } |
2771 | |
2772 | /// Return the first index of the specified basic |
2773 | /// block in the value list for this PHI. Returns -1 if no instance. |
2774 | /// |
2775 | int getBasicBlockIndex(const BasicBlock *BB) const { |
2776 | for (unsigned i = 0, e = getNumOperands(); i != e; ++i) |
2777 | if (block_begin()[i] == BB) |
2778 | return i; |
2779 | return -1; |
2780 | } |
2781 | |
2782 | Value *getIncomingValueForBlock(const BasicBlock *BB) const { |
2783 | int Idx = getBasicBlockIndex(BB); |
2784 | assert(Idx >= 0 && "Invalid basic block argument!")((void)0); |
2785 | return getIncomingValue(Idx); |
2786 | } |
2787 | |
2788 | /// Set every incoming value(s) for block \p BB to \p V. |
2789 | void setIncomingValueForBlock(const BasicBlock *BB, Value *V) { |
2790 | assert(BB && "PHI node got a null basic block!")((void)0); |
2791 | bool Found = false; |
2792 | for (unsigned Op = 0, NumOps = getNumOperands(); Op != NumOps; ++Op) |
2793 | if (getIncomingBlock(Op) == BB) { |
2794 | Found = true; |
2795 | setIncomingValue(Op, V); |
2796 | } |
2797 | (void)Found; |
2798 | assert(Found && "Invalid basic block argument to set!")((void)0); |
2799 | } |
2800 | |
2801 | /// If the specified PHI node always merges together the |
2802 | /// same value, return the value, otherwise return null. |
2803 | Value *hasConstantValue() const; |
2804 | |
2805 | /// Whether the specified PHI node always merges |
2806 | /// together the same value, assuming undefs are equal to a unique |
2807 | /// non-undef value. |
2808 | bool hasConstantOrUndefValue() const; |
2809 | |
2810 | /// If the PHI node is complete which means all of its parent's predecessors |
2811 | /// have incoming value in this PHI, return true, otherwise return false. |
2812 | bool isComplete() const { |
2813 | return llvm::all_of(predecessors(getParent()), |
2814 | [this](const BasicBlock *Pred) { |
2815 | return getBasicBlockIndex(Pred) >= 0; |
2816 | }); |
2817 | } |
2818 | |
2819 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
2820 | static bool classof(const Instruction *I) { |
2821 | return I->getOpcode() == Instruction::PHI; |
2822 | } |
2823 | static bool classof(const Value *V) { |
2824 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2825 | } |
2826 | |
2827 | private: |
2828 | void growOperands(); |
2829 | }; |
2830 | |
2831 | template <> |
2832 | struct OperandTraits<PHINode> : public HungoffOperandTraits<2> { |
2833 | }; |
2834 | |
2835 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(PHINode, Value)PHINode::op_iterator PHINode::op_begin() { return OperandTraits <PHINode>::op_begin(this); } PHINode::const_op_iterator PHINode::op_begin() const { return OperandTraits<PHINode> ::op_begin(const_cast<PHINode*>(this)); } PHINode::op_iterator PHINode::op_end() { return OperandTraits<PHINode>::op_end (this); } PHINode::const_op_iterator PHINode::op_end() const { return OperandTraits<PHINode>::op_end(const_cast<PHINode *>(this)); } Value *PHINode::getOperand(unsigned i_nocapture ) const { ((void)0); return cast_or_null<Value>( OperandTraits <PHINode>::op_begin(const_cast<PHINode*>(this))[i_nocapture ].get()); } void PHINode::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((void)0); OperandTraits<PHINode>::op_begin (this)[i_nocapture] = Val_nocapture; } unsigned PHINode::getNumOperands () const { return OperandTraits<PHINode>::operands(this ); } template <int Idx_nocapture> Use &PHINode::Op( ) { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &PHINode::Op() const { return this->OpFrom<Idx_nocapture>(this); } |
2836 | |
2837 | //===----------------------------------------------------------------------===// |
2838 | // LandingPadInst Class |
2839 | //===----------------------------------------------------------------------===// |
2840 | |
2841 | //===--------------------------------------------------------------------------- |
2842 | /// The landingpad instruction holds all of the information |
2843 | /// necessary to generate correct exception handling. The landingpad instruction |
2844 | /// cannot be moved from the top of a landing pad block, which itself is |
2845 | /// accessible only from the 'unwind' edge of an invoke. This uses the |
2846 | /// SubclassData field in Value to store whether or not the landingpad is a |
2847 | /// cleanup. |
2848 | /// |
2849 | class LandingPadInst : public Instruction { |
2850 | using CleanupField = BoolBitfieldElementT<0>; |
2851 | |
2852 | /// The number of operands actually allocated. NumOperands is |
2853 | /// the number actually in use. |
2854 | unsigned ReservedSpace; |
2855 | |
2856 | LandingPadInst(const LandingPadInst &LP); |
2857 | |
2858 | public: |
2859 | enum ClauseType { Catch, Filter }; |
2860 | |
2861 | private: |
2862 | explicit LandingPadInst(Type *RetTy, unsigned NumReservedValues, |
2863 | const Twine &NameStr, Instruction *InsertBefore); |
2864 | explicit LandingPadInst(Type *RetTy, unsigned NumReservedValues, |
2865 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2866 | |
2867 | // Allocate space for exactly zero operands. |
2868 | void *operator new(size_t S) { return User::operator new(S); } |
2869 | |
2870 | void growOperands(unsigned Size); |
2871 | void init(unsigned NumReservedValues, const Twine &NameStr); |
2872 | |
2873 | protected: |
2874 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2875 | friend class Instruction; |
2876 | |
2877 | LandingPadInst *cloneImpl() const; |
2878 | |
2879 | public: |
2880 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
2881 | |
2882 | /// Constructors - NumReservedClauses is a hint for the number of incoming |
2883 | /// clauses that this landingpad will have (use 0 if you really have no idea). |
2884 | static LandingPadInst *Create(Type *RetTy, unsigned NumReservedClauses, |
2885 | const Twine &NameStr = "", |
2886 | Instruction *InsertBefore = nullptr); |
2887 | static LandingPadInst *Create(Type *RetTy, unsigned NumReservedClauses, |
2888 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2889 | |
2890 | /// Provide fast operand accessors |
2891 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
2892 | |
2893 | /// Return 'true' if this landingpad instruction is a |
2894 | /// cleanup. I.e., it should be run when unwinding even if its landing pad |
2895 | /// doesn't catch the exception. |
2896 | bool isCleanup() const { return getSubclassData<CleanupField>(); } |
2897 | |
2898 | /// Indicate that this landingpad instruction is a cleanup. |
2899 | void setCleanup(bool V) { setSubclassData<CleanupField>(V); } |
2900 | |
2901 | /// Add a catch or filter clause to the landing pad. |
2902 | void addClause(Constant *ClauseVal); |
2903 | |
2904 | /// Get the value of the clause at index Idx. Use isCatch/isFilter to |
2905 | /// determine what type of clause this is. |
2906 | Constant *getClause(unsigned Idx) const { |
2907 | return cast<Constant>(getOperandList()[Idx]); |
2908 | } |
2909 | |
2910 | /// Return 'true' if the clause and index Idx is a catch clause. |
2911 | bool isCatch(unsigned Idx) const { |
2912 | return !isa<ArrayType>(getOperandList()[Idx]->getType()); |
2913 | } |
2914 | |
2915 | /// Return 'true' if the clause and index Idx is a filter clause. |
2916 | bool isFilter(unsigned Idx) const { |
2917 | return isa<ArrayType>(getOperandList()[Idx]->getType()); |
2918 | } |
2919 | |
2920 | /// Get the number of clauses for this landing pad. |
2921 | unsigned getNumClauses() const { return getNumOperands(); } |
2922 | |
2923 | /// Grow the size of the operand list to accommodate the new |
2924 | /// number of clauses. |
2925 | void reserveClauses(unsigned Size) { growOperands(Size); } |
2926 | |
2927 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2928 | static bool classof(const Instruction *I) { |
2929 | return I->getOpcode() == Instruction::LandingPad; |
2930 | } |
2931 | static bool classof(const Value *V) { |
2932 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2933 | } |
2934 | }; |
2935 | |
2936 | template <> |
2937 | struct OperandTraits<LandingPadInst> : public HungoffOperandTraits<1> { |
2938 | }; |
2939 | |
2940 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(LandingPadInst, Value)LandingPadInst::op_iterator LandingPadInst::op_begin() { return OperandTraits<LandingPadInst>::op_begin(this); } LandingPadInst ::const_op_iterator LandingPadInst::op_begin() const { return OperandTraits<LandingPadInst>::op_begin(const_cast< LandingPadInst*>(this)); } LandingPadInst::op_iterator LandingPadInst ::op_end() { return OperandTraits<LandingPadInst>::op_end (this); } LandingPadInst::const_op_iterator LandingPadInst::op_end () const { return OperandTraits<LandingPadInst>::op_end (const_cast<LandingPadInst*>(this)); } Value *LandingPadInst ::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<LandingPadInst>::op_begin( const_cast<LandingPadInst*>(this))[i_nocapture].get()); } void LandingPadInst::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((void)0); OperandTraits<LandingPadInst >::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned LandingPadInst::getNumOperands() const { return OperandTraits <LandingPadInst>::operands(this); } template <int Idx_nocapture > Use &LandingPadInst::Op() { return this->OpFrom< Idx_nocapture>(this); } template <int Idx_nocapture> const Use &LandingPadInst::Op() const { return this-> OpFrom<Idx_nocapture>(this); } |
2941 | |
2942 | //===----------------------------------------------------------------------===// |
2943 | // ReturnInst Class |
2944 | //===----------------------------------------------------------------------===// |
2945 | |
2946 | //===--------------------------------------------------------------------------- |
2947 | /// Return a value (possibly void), from a function. Execution |
2948 | /// does not continue in this function any longer. |
2949 | /// |
2950 | class ReturnInst : public Instruction { |
2951 | ReturnInst(const ReturnInst &RI); |
2952 | |
2953 | private: |
2954 | // ReturnInst constructors: |
2955 | // ReturnInst() - 'ret void' instruction |
2956 | // ReturnInst( null) - 'ret void' instruction |
2957 | // ReturnInst(Value* X) - 'ret X' instruction |
2958 | // ReturnInst( null, Inst *I) - 'ret void' instruction, insert before I |
2959 | // ReturnInst(Value* X, Inst *I) - 'ret X' instruction, insert before I |
2960 | // ReturnInst( null, BB *B) - 'ret void' instruction, insert @ end of B |
2961 | // ReturnInst(Value* X, BB *B) - 'ret X' instruction, insert @ end of B |
2962 | // |
2963 | // NOTE: If the Value* passed is of type void then the constructor behaves as |
2964 | // if it was passed NULL. |
2965 | explicit ReturnInst(LLVMContext &C, Value *retVal = nullptr, |
2966 | Instruction *InsertBefore = nullptr); |
2967 | ReturnInst(LLVMContext &C, Value *retVal, BasicBlock *InsertAtEnd); |
2968 | explicit ReturnInst(LLVMContext &C, BasicBlock *InsertAtEnd); |
2969 | |
2970 | protected: |
2971 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2972 | friend class Instruction; |
2973 | |
2974 | ReturnInst *cloneImpl() const; |
2975 | |
2976 | public: |
2977 | static ReturnInst* Create(LLVMContext &C, Value *retVal = nullptr, |
2978 | Instruction *InsertBefore = nullptr) { |
2979 | return new(!!retVal) ReturnInst(C, retVal, InsertBefore); |
2980 | } |
2981 | |
2982 | static ReturnInst* Create(LLVMContext &C, Value *retVal, |
2983 | BasicBlock *InsertAtEnd) { |
2984 | return new(!!retVal) ReturnInst(C, retVal, InsertAtEnd); |
2985 | } |
2986 | |
2987 | static ReturnInst* Create(LLVMContext &C, BasicBlock *InsertAtEnd) { |
2988 | return new(0) ReturnInst(C, InsertAtEnd); |
2989 | } |
2990 | |
2991 | /// Provide fast operand accessors |
2992 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
2993 | |
2994 | /// Convenience accessor. Returns null if there is no return value. |
2995 | Value *getReturnValue() const { |
2996 | return getNumOperands() != 0 ? getOperand(0) : nullptr; |
2997 | } |
2998 | |
2999 | unsigned getNumSuccessors() const { return 0; } |
3000 | |
3001 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
3002 | static bool classof(const Instruction *I) { |
3003 | return (I->getOpcode() == Instruction::Ret); |
3004 | } |
3005 | static bool classof(const Value *V) { |
3006 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
3007 | } |
3008 | |
3009 | private: |
3010 | BasicBlock *getSuccessor(unsigned idx) const { |
3011 | llvm_unreachable("ReturnInst has no successors!")__builtin_unreachable(); |
3012 | } |
3013 | |
3014 | void setSuccessor(unsigned idx, BasicBlock *B) { |
3015 | llvm_unreachable("ReturnInst has no successors!")__builtin_unreachable(); |
3016 | } |
3017 | }; |
3018 | |
3019 | template <> |
3020 | struct OperandTraits<ReturnInst> : public VariadicOperandTraits<ReturnInst> { |
3021 | }; |
3022 | |
3023 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ReturnInst, Value)ReturnInst::op_iterator ReturnInst::op_begin() { return OperandTraits <ReturnInst>::op_begin(this); } ReturnInst::const_op_iterator ReturnInst::op_begin() const { return OperandTraits<ReturnInst >::op_begin(const_cast<ReturnInst*>(this)); } ReturnInst ::op_iterator ReturnInst::op_end() { return OperandTraits< ReturnInst>::op_end(this); } ReturnInst::const_op_iterator ReturnInst::op_end() const { return OperandTraits<ReturnInst >::op_end(const_cast<ReturnInst*>(this)); } Value *ReturnInst ::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<ReturnInst>::op_begin(const_cast <ReturnInst*>(this))[i_nocapture].get()); } void ReturnInst ::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (( void)0); OperandTraits<ReturnInst>::op_begin(this)[i_nocapture ] = Val_nocapture; } unsigned ReturnInst::getNumOperands() const { return OperandTraits<ReturnInst>::operands(this); } template <int Idx_nocapture> Use &ReturnInst::Op() { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &ReturnInst::Op() const { return this->OpFrom<Idx_nocapture>(this); } |
3024 | |
3025 | //===----------------------------------------------------------------------===// |
3026 | // BranchInst Class |
3027 | //===----------------------------------------------------------------------===// |
3028 | |
3029 | //===--------------------------------------------------------------------------- |
3030 | /// Conditional or Unconditional Branch instruction. |
3031 | /// |
3032 | class BranchInst : public Instruction { |
3033 | /// Ops list - Branches are strange. The operands are ordered: |
3034 | /// [Cond, FalseDest,] TrueDest. This makes some accessors faster because |
3035 | /// they don't have to check for cond/uncond branchness. These are mostly |
3036 | /// accessed relative from op_end(). |
3037 | BranchInst(const BranchInst &BI); |
3038 | // BranchInst constructors (where {B, T, F} are blocks, and C is a condition): |
3039 | // BranchInst(BB *B) - 'br B' |
3040 | // BranchInst(BB* T, BB *F, Value *C) - 'br C, T, F' |
3041 | // BranchInst(BB* B, Inst *I) - 'br B' insert before I |
3042 | // BranchInst(BB* T, BB *F, Value *C, Inst *I) - 'br C, T, F', insert before I |
3043 | // BranchInst(BB* B, BB *I) - 'br B' insert at end |
3044 | // BranchInst(BB* T, BB *F, Value *C, BB *I) - 'br C, T, F', insert at end |
3045 | explicit BranchInst(BasicBlock *IfTrue, Instruction *InsertBefore = nullptr); |
3046 | BranchInst(BasicBlock *IfTrue, BasicBlock *IfFalse, Value *Cond, |
3047 | Instruction *InsertBefore = nullptr); |
3048 | BranchInst(BasicBlock *IfTrue, BasicBlock *InsertAtEnd); |
3049 | BranchInst(BasicBlock *IfTrue, BasicBlock *IfFalse, Value *Cond, |
3050 | BasicBlock *InsertAtEnd); |
3051 | |
3052 | void AssertOK(); |
3053 | |
3054 | protected: |
3055 | // Note: Instruction needs to be a friend here to call cloneImpl. |
3056 | friend class Instruction; |
3057 | |
3058 | BranchInst *cloneImpl() const; |
3059 | |
3060 | public: |
3061 | /// Iterator type that casts an operand to a basic block. |
3062 | /// |
3063 | /// This only makes sense because the successors are stored as adjacent |
3064 | /// operands for branch instructions. |
3065 | struct succ_op_iterator |
3066 | : iterator_adaptor_base<succ_op_iterator, value_op_iterator, |
3067 | std::random_access_iterator_tag, BasicBlock *, |
3068 | ptrdiff_t, BasicBlock *, BasicBlock *> { |
3069 | explicit succ_op_iterator(value_op_iterator I) : iterator_adaptor_base(I) {} |
3070 | |
3071 | BasicBlock *operator*() const { return cast<BasicBlock>(*I); } |
3072 | BasicBlock *operator->() const { return operator*(); } |
3073 | }; |
3074 | |
3075 | /// The const version of `succ_op_iterator`. |
3076 | struct const_succ_op_iterator |
3077 | : iterator_adaptor_base<const_succ_op_iterator, const_value_op_iterator, |
3078 | std::random_access_iterator_tag, |
3079 | const BasicBlock *, ptrdiff_t, const BasicBlock *, |
3080 | const BasicBlock *> { |
3081 | explicit const_succ_op_iterator(const_value_op_iterator I) |
3082 | : iterator_adaptor_base(I) {} |
3083 | |
3084 | const BasicBlock *operator*() const { return cast<BasicBlock>(*I); } |
3085 | const BasicBlock *operator->() const { return operator*(); } |
3086 | }; |
3087 | |
3088 | static BranchInst *Create(BasicBlock *IfTrue, |
3089 | Instruction *InsertBefore = nullptr) { |
3090 | return new(1) BranchInst(IfTrue, InsertBefore); |
3091 | } |
3092 | |
3093 | static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *IfFalse, |
3094 | Value *Cond, Instruction *InsertBefore = nullptr) { |
3095 | return new(3) BranchInst(IfTrue, IfFalse, Cond, InsertBefore); |
3096 | } |
3097 | |
3098 | static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *InsertAtEnd) { |
3099 | return new(1) BranchInst(IfTrue, InsertAtEnd); |
3100 | } |
3101 | |
3102 | static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *IfFalse, |
3103 | Value *Cond, BasicBlock *InsertAtEnd) { |
3104 | return new(3) BranchInst(IfTrue, IfFalse, Cond, InsertAtEnd); |
3105 | } |
3106 | |
3107 | /// Transparently provide more efficient getOperand methods. |
3108 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
3109 | |
3110 | bool isUnconditional() const { return getNumOperands() == 1; } |
3111 | bool isConditional() const { return getNumOperands() == 3; } |
3112 | |
3113 | Value *getCondition() const { |
3114 | assert(isConditional() && "Cannot get condition of an uncond branch!")((void)0); |
3115 | return Op<-3>(); |
3116 | } |
3117 | |
3118 | void setCondition(Value *V) { |
3119 | assert(isConditional() && "Cannot set condition of unconditional branch!")((void)0); |
3120 | Op<-3>() = V; |
3121 | } |
3122 | |
3123 | unsigned getNumSuccessors() const { return 1+isConditional(); } |
3124 | |
3125 | BasicBlock *getSuccessor(unsigned i) const { |
3126 | assert(i < getNumSuccessors() && "Successor # out of range for Branch!")((void)0); |
3127 | return cast_or_null<BasicBlock>((&Op<-1>() - i)->get()); |
3128 | } |
3129 | |
3130 | void setSuccessor(unsigned idx, BasicBlock *NewSucc) { |
3131 | assert(idx < getNumSuccessors() && "Successor # out of range for Branch!")((void)0); |
3132 | *(&Op<-1>() - idx) = NewSucc; |
3133 | } |
3134 | |
3135 | /// Swap the successors of this branch instruction. |
3136 | /// |
3137 | /// Swaps the successors of the branch instruction. This also swaps any |
3138 | /// branch weight metadata associated with the instruction so that it |
3139 | /// continues to map correctly to each operand. |
3140 | void swapSuccessors(); |
3141 | |
3142 | iterator_range<succ_op_iterator> successors() { |
3143 | return make_range( |
3144 | succ_op_iterator(std::next(value_op_begin(), isConditional() ? 1 : 0)), |
3145 | succ_op_iterator(value_op_end())); |
3146 | } |
3147 | |
3148 | iterator_range<const_succ_op_iterator> successors() const { |
3149 | return make_range(const_succ_op_iterator( |
3150 | std::next(value_op_begin(), isConditional() ? 1 : 0)), |
3151 | const_succ_op_iterator(value_op_end())); |
3152 | } |
3153 | |
3154 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
3155 | static bool classof(const Instruction *I) { |
3156 | return (I->getOpcode() == Instruction::Br); |
3157 | } |
3158 | static bool classof(const Value *V) { |
3159 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
3160 | } |
3161 | }; |
3162 | |
3163 | template <> |
3164 | struct OperandTraits<BranchInst> : public VariadicOperandTraits<BranchInst, 1> { |
3165 | }; |
3166 | |
3167 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(BranchInst, Value)BranchInst::op_iterator BranchInst::op_begin() { return OperandTraits <BranchIn |