File: | llvm/lib/Transforms/Scalar/LICM.cpp |
Warning: | line 1207, 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 | SmallVectorImpl<Instruction *> &MaybePromotable); | ||||||||
194 | |||||||||
195 | namespace { | ||||||||
196 | struct LoopInvariantCodeMotion { | ||||||||
197 | bool runOnLoop(Loop *L, AAResults *AA, LoopInfo *LI, DominatorTree *DT, | ||||||||
198 | BlockFrequencyInfo *BFI, TargetLibraryInfo *TLI, | ||||||||
199 | TargetTransformInfo *TTI, ScalarEvolution *SE, MemorySSA *MSSA, | ||||||||
200 | OptimizationRemarkEmitter *ORE); | ||||||||
201 | |||||||||
202 | LoopInvariantCodeMotion(unsigned LicmMssaOptCap, | ||||||||
203 | unsigned LicmMssaNoAccForPromotionCap) | ||||||||
204 | : LicmMssaOptCap(LicmMssaOptCap), | ||||||||
205 | LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap) {} | ||||||||
206 | |||||||||
207 | private: | ||||||||
208 | unsigned LicmMssaOptCap; | ||||||||
209 | unsigned LicmMssaNoAccForPromotionCap; | ||||||||
210 | |||||||||
211 | std::unique_ptr<AliasSetTracker> | ||||||||
212 | collectAliasInfoForLoop(Loop *L, LoopInfo *LI, AAResults *AA); | ||||||||
213 | }; | ||||||||
214 | |||||||||
215 | struct LegacyLICMPass : public LoopPass { | ||||||||
216 | static char ID; // Pass identification, replacement for typeid | ||||||||
217 | LegacyLICMPass( | ||||||||
218 | unsigned LicmMssaOptCap = SetLicmMssaOptCap, | ||||||||
219 | unsigned LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap) | ||||||||
220 | : LoopPass(ID), LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap) { | ||||||||
221 | initializeLegacyLICMPassPass(*PassRegistry::getPassRegistry()); | ||||||||
222 | } | ||||||||
223 | |||||||||
224 | bool runOnLoop(Loop *L, LPPassManager &LPM) override { | ||||||||
225 | if (skipLoop(L)) | ||||||||
226 | return false; | ||||||||
227 | |||||||||
228 | LLVM_DEBUG(dbgs() << "Perform LICM on Loop with header at block "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "Perform LICM on Loop with header at block " << L->getHeader()->getNameOrAsOperand() << "\n"; } } while (false) | ||||||||
229 | << L->getHeader()->getNameOrAsOperand() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "Perform LICM on Loop with header at block " << L->getHeader()->getNameOrAsOperand() << "\n"; } } while (false); | ||||||||
230 | |||||||||
231 | auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>(); | ||||||||
232 | MemorySSA *MSSA = EnableMSSALoopDependency | ||||||||
233 | ? (&getAnalysis<MemorySSAWrapperPass>().getMSSA()) | ||||||||
234 | : nullptr; | ||||||||
235 | bool hasProfileData = L->getHeader()->getParent()->hasProfileData(); | ||||||||
236 | BlockFrequencyInfo *BFI = | ||||||||
237 | hasProfileData ? &getAnalysis<LazyBlockFrequencyInfoPass>().getBFI() | ||||||||
238 | : nullptr; | ||||||||
239 | // For the old PM, we can't use OptimizationRemarkEmitter as an analysis | ||||||||
240 | // pass. Function analyses need to be preserved across loop transformations | ||||||||
241 | // but ORE cannot be preserved (see comment before the pass definition). | ||||||||
242 | OptimizationRemarkEmitter ORE(L->getHeader()->getParent()); | ||||||||
243 | return LICM.runOnLoop( | ||||||||
244 | L, &getAnalysis<AAResultsWrapperPass>().getAAResults(), | ||||||||
245 | &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(), | ||||||||
246 | &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), BFI, | ||||||||
247 | &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI( | ||||||||
248 | *L->getHeader()->getParent()), | ||||||||
249 | &getAnalysis<TargetTransformInfoWrapperPass>().getTTI( | ||||||||
250 | *L->getHeader()->getParent()), | ||||||||
251 | SE ? &SE->getSE() : nullptr, MSSA, &ORE); | ||||||||
252 | } | ||||||||
253 | |||||||||
254 | /// This transformation requires natural loop information & requires that | ||||||||
255 | /// loop preheaders be inserted into the CFG... | ||||||||
256 | /// | ||||||||
257 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||||||
258 | AU.addPreserved<DominatorTreeWrapperPass>(); | ||||||||
259 | AU.addPreserved<LoopInfoWrapperPass>(); | ||||||||
260 | AU.addRequired<TargetLibraryInfoWrapperPass>(); | ||||||||
261 | if (EnableMSSALoopDependency) { | ||||||||
262 | AU.addRequired<MemorySSAWrapperPass>(); | ||||||||
263 | AU.addPreserved<MemorySSAWrapperPass>(); | ||||||||
264 | } | ||||||||
265 | AU.addRequired<TargetTransformInfoWrapperPass>(); | ||||||||
266 | getLoopAnalysisUsage(AU); | ||||||||
267 | LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU); | ||||||||
268 | AU.addPreserved<LazyBlockFrequencyInfoPass>(); | ||||||||
269 | AU.addPreserved<LazyBranchProbabilityInfoPass>(); | ||||||||
270 | } | ||||||||
271 | |||||||||
272 | private: | ||||||||
273 | LoopInvariantCodeMotion LICM; | ||||||||
274 | }; | ||||||||
275 | } // namespace | ||||||||
276 | |||||||||
277 | PreservedAnalyses LICMPass::run(Loop &L, LoopAnalysisManager &AM, | ||||||||
278 | LoopStandardAnalysisResults &AR, LPMUpdater &) { | ||||||||
279 | // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis | ||||||||
280 | // pass. Function analyses need to be preserved across loop transformations | ||||||||
281 | // but ORE cannot be preserved (see comment before the pass definition). | ||||||||
282 | OptimizationRemarkEmitter ORE(L.getHeader()->getParent()); | ||||||||
283 | |||||||||
284 | LoopInvariantCodeMotion LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap); | ||||||||
285 | if (!LICM.runOnLoop(&L, &AR.AA, &AR.LI, &AR.DT, AR.BFI, &AR.TLI, &AR.TTI, | ||||||||
286 | &AR.SE, AR.MSSA, &ORE)) | ||||||||
287 | return PreservedAnalyses::all(); | ||||||||
288 | |||||||||
289 | auto PA = getLoopPassPreservedAnalyses(); | ||||||||
290 | |||||||||
291 | PA.preserve<DominatorTreeAnalysis>(); | ||||||||
292 | PA.preserve<LoopAnalysis>(); | ||||||||
293 | if (AR.MSSA) | ||||||||
294 | PA.preserve<MemorySSAAnalysis>(); | ||||||||
295 | |||||||||
296 | return PA; | ||||||||
297 | } | ||||||||
298 | |||||||||
299 | char LegacyLICMPass::ID = 0; | ||||||||
300 | INITIALIZE_PASS_BEGIN(LegacyLICMPass, "licm", "Loop Invariant Code Motion",static void *initializeLegacyLICMPassPassOnce(PassRegistry & Registry) { | ||||||||
301 | false, false)static void *initializeLegacyLICMPassPassOnce(PassRegistry & Registry) { | ||||||||
302 | INITIALIZE_PASS_DEPENDENCY(LoopPass)initializeLoopPassPass(Registry); | ||||||||
303 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); | ||||||||
304 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | ||||||||
305 | INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)initializeMemorySSAWrapperPassPass(Registry); | ||||||||
306 | INITIALIZE_PASS_DEPENDENCY(LazyBFIPass)initializeLazyBFIPassPass(Registry); | ||||||||
307 | 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)); } | ||||||||
308 | 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)); } | ||||||||
309 | |||||||||
310 | Pass *llvm::createLICMPass() { return new LegacyLICMPass(); } | ||||||||
311 | Pass *llvm::createLICMPass(unsigned LicmMssaOptCap, | ||||||||
312 | unsigned LicmMssaNoAccForPromotionCap) { | ||||||||
313 | return new LegacyLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap); | ||||||||
314 | } | ||||||||
315 | |||||||||
316 | llvm::SinkAndHoistLICMFlags::SinkAndHoistLICMFlags(bool IsSink, Loop *L, | ||||||||
317 | MemorySSA *MSSA) | ||||||||
318 | : SinkAndHoistLICMFlags(SetLicmMssaOptCap, SetLicmMssaNoAccForPromotionCap, | ||||||||
319 | IsSink, L, MSSA) {} | ||||||||
320 | |||||||||
321 | llvm::SinkAndHoistLICMFlags::SinkAndHoistLICMFlags( | ||||||||
322 | unsigned LicmMssaOptCap, unsigned LicmMssaNoAccForPromotionCap, bool IsSink, | ||||||||
323 | Loop *L, MemorySSA *MSSA) | ||||||||
324 | : LicmMssaOptCap(LicmMssaOptCap), | ||||||||
325 | LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap), | ||||||||
326 | IsSink(IsSink) { | ||||||||
327 | assert(((L != nullptr) == (MSSA != nullptr)) &&((((L != nullptr) == (MSSA != nullptr)) && "Unexpected values for SinkAndHoistLICMFlags" ) ? static_cast<void> (0) : __assert_fail ("((L != nullptr) == (MSSA != nullptr)) && \"Unexpected values for SinkAndHoistLICMFlags\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 328, __PRETTY_FUNCTION__)) | ||||||||
328 | "Unexpected values for SinkAndHoistLICMFlags")((((L != nullptr) == (MSSA != nullptr)) && "Unexpected values for SinkAndHoistLICMFlags" ) ? static_cast<void> (0) : __assert_fail ("((L != nullptr) == (MSSA != nullptr)) && \"Unexpected values for SinkAndHoistLICMFlags\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 328, __PRETTY_FUNCTION__)); | ||||||||
329 | if (!MSSA) | ||||||||
330 | return; | ||||||||
331 | |||||||||
332 | unsigned AccessCapCount = 0; | ||||||||
333 | for (auto *BB : L->getBlocks()) | ||||||||
334 | if (const auto *Accesses = MSSA->getBlockAccesses(BB)) | ||||||||
335 | for (const auto &MA : *Accesses) { | ||||||||
336 | (void)MA; | ||||||||
337 | ++AccessCapCount; | ||||||||
338 | if (AccessCapCount > LicmMssaNoAccForPromotionCap) { | ||||||||
339 | NoOfMemAccTooLarge = true; | ||||||||
340 | return; | ||||||||
341 | } | ||||||||
342 | } | ||||||||
343 | } | ||||||||
344 | |||||||||
345 | /// Hoist expressions out of the specified loop. Note, alias info for inner | ||||||||
346 | /// loop is not preserved so it is not a good idea to run LICM multiple | ||||||||
347 | /// times on one loop. | ||||||||
348 | bool LoopInvariantCodeMotion::runOnLoop( | ||||||||
349 | Loop *L, AAResults *AA, LoopInfo *LI, DominatorTree *DT, | ||||||||
350 | BlockFrequencyInfo *BFI, TargetLibraryInfo *TLI, TargetTransformInfo *TTI, | ||||||||
351 | ScalarEvolution *SE, MemorySSA *MSSA, OptimizationRemarkEmitter *ORE) { | ||||||||
352 | bool Changed = false; | ||||||||
353 | |||||||||
354 | assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.")((L->isLCSSAForm(*DT) && "Loop is not in LCSSA form." ) ? static_cast<void> (0) : __assert_fail ("L->isLCSSAForm(*DT) && \"Loop is not in LCSSA form.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 354, __PRETTY_FUNCTION__)); | ||||||||
355 | |||||||||
356 | // If this loop has metadata indicating that LICM is not to be performed then | ||||||||
357 | // just exit. | ||||||||
358 | if (hasDisableLICMTransformsHint(L)) { | ||||||||
359 | return false; | ||||||||
360 | } | ||||||||
361 | |||||||||
362 | std::unique_ptr<AliasSetTracker> CurAST; | ||||||||
363 | std::unique_ptr<MemorySSAUpdater> MSSAU; | ||||||||
364 | std::unique_ptr<SinkAndHoistLICMFlags> Flags; | ||||||||
365 | |||||||||
366 | // Don't sink stores from loops with coroutine suspend instructions. | ||||||||
367 | // LICM would sink instructions into the default destination of | ||||||||
368 | // the coroutine switch. The default destination of the switch is to | ||||||||
369 | // handle the case where the coroutine is suspended, by which point the | ||||||||
370 | // coroutine frame may have been destroyed. No instruction can be sunk there. | ||||||||
371 | // FIXME: This would unfortunately hurt the performance of coroutines, however | ||||||||
372 | // there is currently no general solution for this. Similar issues could also | ||||||||
373 | // potentially happen in other passes where instructions are being moved | ||||||||
374 | // across that edge. | ||||||||
375 | bool HasCoroSuspendInst = llvm::any_of(L->getBlocks(), [](BasicBlock *BB) { | ||||||||
376 | return llvm::any_of(*BB, [](Instruction &I) { | ||||||||
377 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I); | ||||||||
378 | return II && II->getIntrinsicID() == Intrinsic::coro_suspend; | ||||||||
379 | }); | ||||||||
380 | }); | ||||||||
381 | |||||||||
382 | if (!MSSA) { | ||||||||
383 | LLVM_DEBUG(dbgs() << "LICM: Using Alias Set Tracker.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM: Using Alias Set Tracker.\n" ; } } while (false); | ||||||||
384 | CurAST = collectAliasInfoForLoop(L, LI, AA); | ||||||||
385 | Flags = std::make_unique<SinkAndHoistLICMFlags>( | ||||||||
386 | LicmMssaOptCap, LicmMssaNoAccForPromotionCap, /*IsSink=*/true); | ||||||||
387 | } else { | ||||||||
388 | LLVM_DEBUG(dbgs() << "LICM: Using MemorySSA.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM: Using MemorySSA.\n"; } } while (false); | ||||||||
389 | MSSAU = std::make_unique<MemorySSAUpdater>(MSSA); | ||||||||
390 | Flags = std::make_unique<SinkAndHoistLICMFlags>( | ||||||||
391 | LicmMssaOptCap, LicmMssaNoAccForPromotionCap, /*IsSink=*/true, L, MSSA); | ||||||||
392 | } | ||||||||
393 | |||||||||
394 | // Get the preheader block to move instructions into... | ||||||||
395 | BasicBlock *Preheader = L->getLoopPreheader(); | ||||||||
396 | |||||||||
397 | // Compute loop safety information. | ||||||||
398 | ICFLoopSafetyInfo SafetyInfo; | ||||||||
399 | SafetyInfo.computeLoopSafetyInfo(L); | ||||||||
400 | |||||||||
401 | // We want to visit all of the instructions in this loop... that are not parts | ||||||||
402 | // of our subloops (they have already had their invariants hoisted out of | ||||||||
403 | // their loop, into this loop, so there is no need to process the BODIES of | ||||||||
404 | // the subloops). | ||||||||
405 | // | ||||||||
406 | // Traverse the body of the loop in depth first order on the dominator tree so | ||||||||
407 | // that we are guaranteed to see definitions before we see uses. This allows | ||||||||
408 | // us to sink instructions in one pass, without iteration. After sinking | ||||||||
409 | // instructions, we perform another pass to hoist them out of the loop. | ||||||||
410 | if (L->hasDedicatedExits()) | ||||||||
411 | Changed |= | ||||||||
412 | sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, BFI, TLI, TTI, L, | ||||||||
413 | CurAST.get(), MSSAU.get(), &SafetyInfo, *Flags.get(), ORE); | ||||||||
414 | Flags->setIsSink(false); | ||||||||
415 | if (Preheader) | ||||||||
416 | Changed |= hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, BFI, TLI, L, | ||||||||
417 | CurAST.get(), MSSAU.get(), SE, &SafetyInfo, | ||||||||
418 | *Flags.get(), ORE); | ||||||||
419 | |||||||||
420 | // Now that all loop invariants have been removed from the loop, promote any | ||||||||
421 | // memory references to scalars that we can. | ||||||||
422 | // Don't sink stores from loops without dedicated block exits. Exits | ||||||||
423 | // containing indirect branches are not transformed by loop simplify, | ||||||||
424 | // make sure we catch that. An additional load may be generated in the | ||||||||
425 | // preheader for SSA updater, so also avoid sinking when no preheader | ||||||||
426 | // is available. | ||||||||
427 | if (!DisablePromotion && Preheader && L->hasDedicatedExits() && | ||||||||
428 | !Flags->tooManyMemoryAccesses() && !HasCoroSuspendInst) { | ||||||||
429 | // Figure out the loop exits and their insertion points | ||||||||
430 | SmallVector<BasicBlock *, 8> ExitBlocks; | ||||||||
431 | L->getUniqueExitBlocks(ExitBlocks); | ||||||||
432 | |||||||||
433 | // We can't insert into a catchswitch. | ||||||||
434 | bool HasCatchSwitch = llvm::any_of(ExitBlocks, [](BasicBlock *Exit) { | ||||||||
435 | return isa<CatchSwitchInst>(Exit->getTerminator()); | ||||||||
436 | }); | ||||||||
437 | |||||||||
438 | if (!HasCatchSwitch) { | ||||||||
439 | SmallVector<Instruction *, 8> InsertPts; | ||||||||
440 | SmallVector<MemoryAccess *, 8> MSSAInsertPts; | ||||||||
441 | InsertPts.reserve(ExitBlocks.size()); | ||||||||
442 | if (MSSAU) | ||||||||
443 | MSSAInsertPts.reserve(ExitBlocks.size()); | ||||||||
444 | for (BasicBlock *ExitBlock : ExitBlocks) { | ||||||||
445 | InsertPts.push_back(&*ExitBlock->getFirstInsertionPt()); | ||||||||
446 | if (MSSAU) | ||||||||
447 | MSSAInsertPts.push_back(nullptr); | ||||||||
448 | } | ||||||||
449 | |||||||||
450 | PredIteratorCache PIC; | ||||||||
451 | |||||||||
452 | bool Promoted = false; | ||||||||
453 | if (CurAST.get()) { | ||||||||
454 | // Loop over all of the alias sets in the tracker object. | ||||||||
455 | for (AliasSet &AS : *CurAST) { | ||||||||
456 | // We can promote this alias set if it has a store, if it is a "Must" | ||||||||
457 | // alias set, if the pointer is loop invariant, and if we are not | ||||||||
458 | // eliminating any volatile loads or stores. | ||||||||
459 | if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || | ||||||||
460 | !L->isLoopInvariant(AS.begin()->getValue())) | ||||||||
461 | continue; | ||||||||
462 | |||||||||
463 | assert(((!AS.empty() && "Must alias set should have at least one pointer element in it!" ) ? static_cast<void> (0) : __assert_fail ("!AS.empty() && \"Must alias set should have at least one pointer element in it!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 465, __PRETTY_FUNCTION__)) | ||||||||
464 | !AS.empty() &&((!AS.empty() && "Must alias set should have at least one pointer element in it!" ) ? static_cast<void> (0) : __assert_fail ("!AS.empty() && \"Must alias set should have at least one pointer element in it!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 465, __PRETTY_FUNCTION__)) | ||||||||
465 | "Must alias set should have at least one pointer element in it!")((!AS.empty() && "Must alias set should have at least one pointer element in it!" ) ? static_cast<void> (0) : __assert_fail ("!AS.empty() && \"Must alias set should have at least one pointer element in it!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 465, __PRETTY_FUNCTION__)); | ||||||||
466 | |||||||||
467 | SmallSetVector<Value *, 8> PointerMustAliases; | ||||||||
468 | for (const auto &ASI : AS) | ||||||||
469 | PointerMustAliases.insert(ASI.getValue()); | ||||||||
470 | |||||||||
471 | Promoted |= promoteLoopAccessesToScalars( | ||||||||
472 | PointerMustAliases, ExitBlocks, InsertPts, MSSAInsertPts, PIC, LI, | ||||||||
473 | DT, TLI, L, CurAST.get(), MSSAU.get(), &SafetyInfo, ORE); | ||||||||
474 | } | ||||||||
475 | } else { | ||||||||
476 | SmallVector<Instruction *, 16> MaybePromotable; | ||||||||
477 | foreachMemoryAccess(MSSA, L, [&](Instruction *I) { | ||||||||
478 | MaybePromotable.push_back(I); | ||||||||
479 | }); | ||||||||
480 | |||||||||
481 | // Promoting one set of accesses may make the pointers for another set | ||||||||
482 | // loop invariant, so run this in a loop (with the MaybePromotable set | ||||||||
483 | // decreasing in size over time). | ||||||||
484 | bool LocalPromoted; | ||||||||
485 | do { | ||||||||
486 | LocalPromoted = false; | ||||||||
487 | for (const SmallSetVector<Value *, 8> &PointerMustAliases : | ||||||||
488 | collectPromotionCandidates(MSSA, AA, L, MaybePromotable)) { | ||||||||
489 | LocalPromoted |= promoteLoopAccessesToScalars( | ||||||||
490 | PointerMustAliases, ExitBlocks, InsertPts, MSSAInsertPts, PIC, | ||||||||
491 | LI, DT, TLI, L, /*AST*/nullptr, MSSAU.get(), &SafetyInfo, ORE); | ||||||||
492 | } | ||||||||
493 | Promoted |= LocalPromoted; | ||||||||
494 | } while (LocalPromoted); | ||||||||
495 | } | ||||||||
496 | |||||||||
497 | // Once we have promoted values across the loop body we have to | ||||||||
498 | // recursively reform LCSSA as any nested loop may now have values defined | ||||||||
499 | // within the loop used in the outer loop. | ||||||||
500 | // FIXME: This is really heavy handed. It would be a bit better to use an | ||||||||
501 | // SSAUpdater strategy during promotion that was LCSSA aware and reformed | ||||||||
502 | // it as it went. | ||||||||
503 | if (Promoted) | ||||||||
504 | formLCSSARecursively(*L, *DT, LI, SE); | ||||||||
505 | |||||||||
506 | Changed |= Promoted; | ||||||||
507 | } | ||||||||
508 | } | ||||||||
509 | |||||||||
510 | // Check that neither this loop nor its parent have had LCSSA broken. LICM is | ||||||||
511 | // specifically moving instructions across the loop boundary and so it is | ||||||||
512 | // especially in need of sanity checking here. | ||||||||
513 | assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!")((L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!" ) ? static_cast<void> (0) : __assert_fail ("L->isLCSSAForm(*DT) && \"Loop not left in LCSSA form after LICM!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 513, __PRETTY_FUNCTION__)); | ||||||||
514 | assert((L->isOutermost() || L->getParentLoop()->isLCSSAForm(*DT)) &&(((L->isOutermost() || L->getParentLoop()->isLCSSAForm (*DT)) && "Parent loop not left in LCSSA form after LICM!" ) ? static_cast<void> (0) : __assert_fail ("(L->isOutermost() || L->getParentLoop()->isLCSSAForm(*DT)) && \"Parent loop not left in LCSSA form after LICM!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 515, __PRETTY_FUNCTION__)) | ||||||||
515 | "Parent loop not left in LCSSA form after LICM!")(((L->isOutermost() || L->getParentLoop()->isLCSSAForm (*DT)) && "Parent loop not left in LCSSA form after LICM!" ) ? static_cast<void> (0) : __assert_fail ("(L->isOutermost() || L->getParentLoop()->isLCSSAForm(*DT)) && \"Parent loop not left in LCSSA form after LICM!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 515, __PRETTY_FUNCTION__)); | ||||||||
516 | |||||||||
517 | if (MSSAU.get() && VerifyMemorySSA) | ||||||||
518 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||
519 | |||||||||
520 | if (Changed && SE) | ||||||||
521 | SE->forgetLoopDispositions(L); | ||||||||
522 | return Changed; | ||||||||
523 | } | ||||||||
524 | |||||||||
525 | /// Walk the specified region of the CFG (defined by all blocks dominated by | ||||||||
526 | /// the specified block, and that are in the current loop) in reverse depth | ||||||||
527 | /// first order w.r.t the DominatorTree. This allows us to visit uses before | ||||||||
528 | /// definitions, allowing us to sink a loop body in one pass without iteration. | ||||||||
529 | /// | ||||||||
530 | bool llvm::sinkRegion(DomTreeNode *N, AAResults *AA, LoopInfo *LI, | ||||||||
531 | DominatorTree *DT, BlockFrequencyInfo *BFI, | ||||||||
532 | TargetLibraryInfo *TLI, TargetTransformInfo *TTI, | ||||||||
533 | Loop *CurLoop, AliasSetTracker *CurAST, | ||||||||
534 | MemorySSAUpdater *MSSAU, ICFLoopSafetyInfo *SafetyInfo, | ||||||||
535 | SinkAndHoistLICMFlags &Flags, | ||||||||
536 | OptimizationRemarkEmitter *ORE) { | ||||||||
537 | |||||||||
538 | // Verify inputs. | ||||||||
539 | assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&((N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && "Unexpected input to sinkRegion." ) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to sinkRegion.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 541, __PRETTY_FUNCTION__)) | ||||||||
540 | CurLoop != nullptr && SafetyInfo != nullptr &&((N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && "Unexpected input to sinkRegion." ) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to sinkRegion.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 541, __PRETTY_FUNCTION__)) | ||||||||
541 | "Unexpected input to sinkRegion.")((N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && "Unexpected input to sinkRegion." ) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to sinkRegion.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 541, __PRETTY_FUNCTION__)); | ||||||||
542 | assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&((((CurAST != nullptr) ^ (MSSAU != nullptr)) && "Either AliasSetTracker or MemorySSA should be initialized." ) ? static_cast<void> (0) : __assert_fail ("((CurAST != nullptr) ^ (MSSAU != nullptr)) && \"Either AliasSetTracker or MemorySSA should be initialized.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 543, __PRETTY_FUNCTION__)) | ||||||||
543 | "Either AliasSetTracker or MemorySSA should be initialized.")((((CurAST != nullptr) ^ (MSSAU != nullptr)) && "Either AliasSetTracker or MemorySSA should be initialized." ) ? static_cast<void> (0) : __assert_fail ("((CurAST != nullptr) ^ (MSSAU != nullptr)) && \"Either AliasSetTracker or MemorySSA should be initialized.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 543, __PRETTY_FUNCTION__)); | ||||||||
544 | |||||||||
545 | // We want to visit children before parents. We will enque all the parents | ||||||||
546 | // before their children in the worklist and process the worklist in reverse | ||||||||
547 | // order. | ||||||||
548 | SmallVector<DomTreeNode *, 16> Worklist = collectChildrenInLoop(N, CurLoop); | ||||||||
549 | |||||||||
550 | bool Changed = false; | ||||||||
551 | for (DomTreeNode *DTN : reverse(Worklist)) { | ||||||||
552 | BasicBlock *BB = DTN->getBlock(); | ||||||||
553 | // Only need to process the contents of this block if it is not part of a | ||||||||
554 | // subloop (which would already have been processed). | ||||||||
555 | if (inSubLoop(BB, CurLoop, LI)) | ||||||||
556 | continue; | ||||||||
557 | |||||||||
558 | for (BasicBlock::iterator II = BB->end(); II != BB->begin();) { | ||||||||
559 | Instruction &I = *--II; | ||||||||
560 | |||||||||
561 | // If the instruction is dead, we would try to sink it because it isn't | ||||||||
562 | // used in the loop, instead, just delete it. | ||||||||
563 | if (isInstructionTriviallyDead(&I, TLI)) { | ||||||||
564 | LLVM_DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM deleting dead inst: " << I << '\n'; } } while (false); | ||||||||
565 | salvageKnowledge(&I); | ||||||||
566 | salvageDebugInfo(I); | ||||||||
567 | ++II; | ||||||||
568 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | ||||||||
569 | Changed = true; | ||||||||
570 | continue; | ||||||||
571 | } | ||||||||
572 | |||||||||
573 | // Check to see if we can sink this instruction to the exit blocks | ||||||||
574 | // of the loop. We can do this if the all users of the instruction are | ||||||||
575 | // outside of the loop. In this case, it doesn't even matter if the | ||||||||
576 | // operands of the instruction are loop invariant. | ||||||||
577 | // | ||||||||
578 | bool FreeInLoop = false; | ||||||||
579 | if (!I.mayHaveSideEffects() && | ||||||||
580 | isNotUsedOrFreeInLoop(I, CurLoop, SafetyInfo, TTI, FreeInLoop) && | ||||||||
581 | canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags, | ||||||||
582 | ORE)) { | ||||||||
583 | if (sink(I, LI, DT, BFI, CurLoop, SafetyInfo, MSSAU, ORE)) { | ||||||||
584 | if (!FreeInLoop) { | ||||||||
585 | ++II; | ||||||||
586 | salvageDebugInfo(I); | ||||||||
587 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | ||||||||
588 | } | ||||||||
589 | Changed = true; | ||||||||
590 | } | ||||||||
591 | } | ||||||||
592 | } | ||||||||
593 | } | ||||||||
594 | if (MSSAU && VerifyMemorySSA) | ||||||||
595 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||
596 | return Changed; | ||||||||
597 | } | ||||||||
598 | |||||||||
599 | namespace { | ||||||||
600 | // This is a helper class for hoistRegion to make it able to hoist control flow | ||||||||
601 | // in order to be able to hoist phis. The way this works is that we initially | ||||||||
602 | // start hoisting to the loop preheader, and when we see a loop invariant branch | ||||||||
603 | // we make note of this. When we then come to hoist an instruction that's | ||||||||
604 | // conditional on such a branch we duplicate the branch and the relevant control | ||||||||
605 | // flow, then hoist the instruction into the block corresponding to its original | ||||||||
606 | // block in the duplicated control flow. | ||||||||
607 | class ControlFlowHoister { | ||||||||
608 | private: | ||||||||
609 | // Information about the loop we are hoisting from | ||||||||
610 | LoopInfo *LI; | ||||||||
611 | DominatorTree *DT; | ||||||||
612 | Loop *CurLoop; | ||||||||
613 | MemorySSAUpdater *MSSAU; | ||||||||
614 | |||||||||
615 | // A map of blocks in the loop to the block their instructions will be hoisted | ||||||||
616 | // to. | ||||||||
617 | DenseMap<BasicBlock *, BasicBlock *> HoistDestinationMap; | ||||||||
618 | |||||||||
619 | // The branches that we can hoist, mapped to the block that marks a | ||||||||
620 | // convergence point of their control flow. | ||||||||
621 | DenseMap<BranchInst *, BasicBlock *> HoistableBranches; | ||||||||
622 | |||||||||
623 | public: | ||||||||
624 | ControlFlowHoister(LoopInfo *LI, DominatorTree *DT, Loop *CurLoop, | ||||||||
625 | MemorySSAUpdater *MSSAU) | ||||||||
626 | : LI(LI), DT(DT), CurLoop(CurLoop), MSSAU(MSSAU) {} | ||||||||
627 | |||||||||
628 | void registerPossiblyHoistableBranch(BranchInst *BI) { | ||||||||
629 | // We can only hoist conditional branches with loop invariant operands. | ||||||||
630 | if (!ControlFlowHoisting || !BI->isConditional() || | ||||||||
631 | !CurLoop->hasLoopInvariantOperands(BI)) | ||||||||
632 | return; | ||||||||
633 | |||||||||
634 | // The branch destinations need to be in the loop, and we don't gain | ||||||||
635 | // anything by duplicating conditional branches with duplicate successors, | ||||||||
636 | // as it's essentially the same as an unconditional branch. | ||||||||
637 | BasicBlock *TrueDest = BI->getSuccessor(0); | ||||||||
638 | BasicBlock *FalseDest = BI->getSuccessor(1); | ||||||||
639 | if (!CurLoop->contains(TrueDest) || !CurLoop->contains(FalseDest) || | ||||||||
640 | TrueDest == FalseDest) | ||||||||
641 | return; | ||||||||
642 | |||||||||
643 | // We can hoist BI if one branch destination is the successor of the other, | ||||||||
644 | // or both have common successor which we check by seeing if the | ||||||||
645 | // intersection of their successors is non-empty. | ||||||||
646 | // TODO: This could be expanded to allowing branches where both ends | ||||||||
647 | // eventually converge to a single block. | ||||||||
648 | SmallPtrSet<BasicBlock *, 4> TrueDestSucc, FalseDestSucc; | ||||||||
649 | TrueDestSucc.insert(succ_begin(TrueDest), succ_end(TrueDest)); | ||||||||
650 | FalseDestSucc.insert(succ_begin(FalseDest), succ_end(FalseDest)); | ||||||||
651 | BasicBlock *CommonSucc = nullptr; | ||||||||
652 | if (TrueDestSucc.count(FalseDest)) { | ||||||||
653 | CommonSucc = FalseDest; | ||||||||
654 | } else if (FalseDestSucc.count(TrueDest)) { | ||||||||
655 | CommonSucc = TrueDest; | ||||||||
656 | } else { | ||||||||
657 | set_intersect(TrueDestSucc, FalseDestSucc); | ||||||||
658 | // If there's one common successor use that. | ||||||||
659 | if (TrueDestSucc.size() == 1) | ||||||||
660 | CommonSucc = *TrueDestSucc.begin(); | ||||||||
661 | // If there's more than one pick whichever appears first in the block list | ||||||||
662 | // (we can't use the value returned by TrueDestSucc.begin() as it's | ||||||||
663 | // unpredicatable which element gets returned). | ||||||||
664 | else if (!TrueDestSucc.empty()) { | ||||||||
665 | Function *F = TrueDest->getParent(); | ||||||||
666 | auto IsSucc = [&](BasicBlock &BB) { return TrueDestSucc.count(&BB); }; | ||||||||
667 | auto It = llvm::find_if(*F, IsSucc); | ||||||||
668 | assert(It != F->end() && "Could not find successor in function")((It != F->end() && "Could not find successor in function" ) ? static_cast<void> (0) : __assert_fail ("It != F->end() && \"Could not find successor in function\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 668, __PRETTY_FUNCTION__)); | ||||||||
669 | CommonSucc = &*It; | ||||||||
670 | } | ||||||||
671 | } | ||||||||
672 | // The common successor has to be dominated by the branch, as otherwise | ||||||||
673 | // there will be some other path to the successor that will not be | ||||||||
674 | // controlled by this branch so any phi we hoist would be controlled by the | ||||||||
675 | // wrong condition. This also takes care of avoiding hoisting of loop back | ||||||||
676 | // edges. | ||||||||
677 | // TODO: In some cases this could be relaxed if the successor is dominated | ||||||||
678 | // by another block that's been hoisted and we can guarantee that the | ||||||||
679 | // control flow has been replicated exactly. | ||||||||
680 | if (CommonSucc && DT->dominates(BI, CommonSucc)) | ||||||||
681 | HoistableBranches[BI] = CommonSucc; | ||||||||
682 | } | ||||||||
683 | |||||||||
684 | bool canHoistPHI(PHINode *PN) { | ||||||||
685 | // The phi must have loop invariant operands. | ||||||||
686 | if (!ControlFlowHoisting || !CurLoop->hasLoopInvariantOperands(PN)) | ||||||||
687 | return false; | ||||||||
688 | // We can hoist phis if the block they are in is the target of hoistable | ||||||||
689 | // branches which cover all of the predecessors of the block. | ||||||||
690 | SmallPtrSet<BasicBlock *, 8> PredecessorBlocks; | ||||||||
691 | BasicBlock *BB = PN->getParent(); | ||||||||
692 | for (BasicBlock *PredBB : predecessors(BB)) | ||||||||
693 | PredecessorBlocks.insert(PredBB); | ||||||||
694 | // If we have less predecessor blocks than predecessors then the phi will | ||||||||
695 | // have more than one incoming value for the same block which we can't | ||||||||
696 | // handle. | ||||||||
697 | // TODO: This could be handled be erasing some of the duplicate incoming | ||||||||
698 | // values. | ||||||||
699 | if (PredecessorBlocks.size() != pred_size(BB)) | ||||||||
700 | return false; | ||||||||
701 | for (auto &Pair : HoistableBranches) { | ||||||||
702 | if (Pair.second == BB) { | ||||||||
703 | // Which blocks are predecessors via this branch depends on if the | ||||||||
704 | // branch is triangle-like or diamond-like. | ||||||||
705 | if (Pair.first->getSuccessor(0) == BB) { | ||||||||
706 | PredecessorBlocks.erase(Pair.first->getParent()); | ||||||||
707 | PredecessorBlocks.erase(Pair.first->getSuccessor(1)); | ||||||||
708 | } else if (Pair.first->getSuccessor(1) == BB) { | ||||||||
709 | PredecessorBlocks.erase(Pair.first->getParent()); | ||||||||
710 | PredecessorBlocks.erase(Pair.first->getSuccessor(0)); | ||||||||
711 | } else { | ||||||||
712 | PredecessorBlocks.erase(Pair.first->getSuccessor(0)); | ||||||||
713 | PredecessorBlocks.erase(Pair.first->getSuccessor(1)); | ||||||||
714 | } | ||||||||
715 | } | ||||||||
716 | } | ||||||||
717 | // PredecessorBlocks will now be empty if for every predecessor of BB we | ||||||||
718 | // found a hoistable branch source. | ||||||||
719 | return PredecessorBlocks.empty(); | ||||||||
720 | } | ||||||||
721 | |||||||||
722 | BasicBlock *getOrCreateHoistedBlock(BasicBlock *BB) { | ||||||||
723 | if (!ControlFlowHoisting) | ||||||||
724 | return CurLoop->getLoopPreheader(); | ||||||||
725 | // If BB has already been hoisted, return that | ||||||||
726 | if (HoistDestinationMap.count(BB)) | ||||||||
727 | return HoistDestinationMap[BB]; | ||||||||
728 | |||||||||
729 | // Check if this block is conditional based on a pending branch | ||||||||
730 | auto HasBBAsSuccessor = | ||||||||
731 | [&](DenseMap<BranchInst *, BasicBlock *>::value_type &Pair) { | ||||||||
732 | return BB != Pair.second && (Pair.first->getSuccessor(0) == BB || | ||||||||
733 | Pair.first->getSuccessor(1) == BB); | ||||||||
734 | }; | ||||||||
735 | auto It = llvm::find_if(HoistableBranches, HasBBAsSuccessor); | ||||||||
736 | |||||||||
737 | // If not involved in a pending branch, hoist to preheader | ||||||||
738 | BasicBlock *InitialPreheader = CurLoop->getLoopPreheader(); | ||||||||
739 | if (It == HoistableBranches.end()) { | ||||||||
740 | LLVM_DEBUG(dbgs() << "LICM using "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM using " << InitialPreheader ->getNameOrAsOperand() << " as hoist destination for " << BB->getNameOrAsOperand() << "\n"; } } while (false) | ||||||||
741 | << InitialPreheader->getNameOrAsOperand()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM using " << InitialPreheader ->getNameOrAsOperand() << " as hoist destination for " << BB->getNameOrAsOperand() << "\n"; } } while (false) | ||||||||
742 | << " as hoist destination for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM using " << InitialPreheader ->getNameOrAsOperand() << " as hoist destination for " << BB->getNameOrAsOperand() << "\n"; } } while (false) | ||||||||
743 | << BB->getNameOrAsOperand() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM using " << InitialPreheader ->getNameOrAsOperand() << " as hoist destination for " << BB->getNameOrAsOperand() << "\n"; } } while (false); | ||||||||
744 | HoistDestinationMap[BB] = InitialPreheader; | ||||||||
745 | return InitialPreheader; | ||||||||
746 | } | ||||||||
747 | BranchInst *BI = It->first; | ||||||||
748 | assert(std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) ==((std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor ) == HoistableBranches.end() && "BB is expected to be the target of at most one branch" ) ? static_cast<void> (0) : __assert_fail ("std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) == HoistableBranches.end() && \"BB is expected to be the target of at most one branch\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 750, __PRETTY_FUNCTION__)) | ||||||||
749 | HoistableBranches.end() &&((std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor ) == HoistableBranches.end() && "BB is expected to be the target of at most one branch" ) ? static_cast<void> (0) : __assert_fail ("std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) == HoistableBranches.end() && \"BB is expected to be the target of at most one branch\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 750, __PRETTY_FUNCTION__)) | ||||||||
750 | "BB is expected to be the target of at most one branch")((std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor ) == HoistableBranches.end() && "BB is expected to be the target of at most one branch" ) ? static_cast<void> (0) : __assert_fail ("std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) == HoistableBranches.end() && \"BB is expected to be the target of at most one branch\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 750, __PRETTY_FUNCTION__)); | ||||||||
751 | |||||||||
752 | LLVMContext &C = BB->getContext(); | ||||||||
753 | BasicBlock *TrueDest = BI->getSuccessor(0); | ||||||||
754 | BasicBlock *FalseDest = BI->getSuccessor(1); | ||||||||
755 | BasicBlock *CommonSucc = HoistableBranches[BI]; | ||||||||
756 | BasicBlock *HoistTarget = getOrCreateHoistedBlock(BI->getParent()); | ||||||||
757 | |||||||||
758 | // Create hoisted versions of blocks that currently don't have them | ||||||||
759 | auto CreateHoistedBlock = [&](BasicBlock *Orig) { | ||||||||
760 | if (HoistDestinationMap.count(Orig)) | ||||||||
761 | return HoistDestinationMap[Orig]; | ||||||||
762 | BasicBlock *New = | ||||||||
763 | BasicBlock::Create(C, Orig->getName() + ".licm", Orig->getParent()); | ||||||||
764 | HoistDestinationMap[Orig] = New; | ||||||||
765 | DT->addNewBlock(New, HoistTarget); | ||||||||
766 | if (CurLoop->getParentLoop()) | ||||||||
767 | CurLoop->getParentLoop()->addBasicBlockToLoop(New, *LI); | ||||||||
768 | ++NumCreatedBlocks; | ||||||||
769 | LLVM_DEBUG(dbgs() << "LICM created " << New->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM created " << New-> getName() << " as hoist destination for " << Orig ->getName() << "\n"; } } while (false) | ||||||||
770 | << " as hoist destination for " << Orig->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM created " << New-> getName() << " as hoist destination for " << Orig ->getName() << "\n"; } } while (false) | ||||||||
771 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM created " << New-> getName() << " as hoist destination for " << Orig ->getName() << "\n"; } } while (false); | ||||||||
772 | return New; | ||||||||
773 | }; | ||||||||
774 | BasicBlock *HoistTrueDest = CreateHoistedBlock(TrueDest); | ||||||||
775 | BasicBlock *HoistFalseDest = CreateHoistedBlock(FalseDest); | ||||||||
776 | BasicBlock *HoistCommonSucc = CreateHoistedBlock(CommonSucc); | ||||||||
777 | |||||||||
778 | // Link up these blocks with branches. | ||||||||
779 | if (!HoistCommonSucc->getTerminator()) { | ||||||||
780 | // The new common successor we've generated will branch to whatever that | ||||||||
781 | // hoist target branched to. | ||||||||
782 | BasicBlock *TargetSucc = HoistTarget->getSingleSuccessor(); | ||||||||
783 | assert(TargetSucc && "Expected hoist target to have a single successor")((TargetSucc && "Expected hoist target to have a single successor" ) ? static_cast<void> (0) : __assert_fail ("TargetSucc && \"Expected hoist target to have a single successor\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 783, __PRETTY_FUNCTION__)); | ||||||||
784 | HoistCommonSucc->moveBefore(TargetSucc); | ||||||||
785 | BranchInst::Create(TargetSucc, HoistCommonSucc); | ||||||||
786 | } | ||||||||
787 | if (!HoistTrueDest->getTerminator()) { | ||||||||
788 | HoistTrueDest->moveBefore(HoistCommonSucc); | ||||||||
789 | BranchInst::Create(HoistCommonSucc, HoistTrueDest); | ||||||||
790 | } | ||||||||
791 | if (!HoistFalseDest->getTerminator()) { | ||||||||
792 | HoistFalseDest->moveBefore(HoistCommonSucc); | ||||||||
793 | BranchInst::Create(HoistCommonSucc, HoistFalseDest); | ||||||||
794 | } | ||||||||
795 | |||||||||
796 | // If BI is being cloned to what was originally the preheader then | ||||||||
797 | // HoistCommonSucc will now be the new preheader. | ||||||||
798 | if (HoistTarget == InitialPreheader) { | ||||||||
799 | // Phis in the loop header now need to use the new preheader. | ||||||||
800 | InitialPreheader->replaceSuccessorsPhiUsesWith(HoistCommonSucc); | ||||||||
801 | if (MSSAU) | ||||||||
802 | MSSAU->wireOldPredecessorsToNewImmediatePredecessor( | ||||||||
803 | HoistTarget->getSingleSuccessor(), HoistCommonSucc, {HoistTarget}); | ||||||||
804 | // The new preheader dominates the loop header. | ||||||||
805 | DomTreeNode *PreheaderNode = DT->getNode(HoistCommonSucc); | ||||||||
806 | DomTreeNode *HeaderNode = DT->getNode(CurLoop->getHeader()); | ||||||||
807 | DT->changeImmediateDominator(HeaderNode, PreheaderNode); | ||||||||
808 | // The preheader hoist destination is now the new preheader, with the | ||||||||
809 | // exception of the hoist destination of this branch. | ||||||||
810 | for (auto &Pair : HoistDestinationMap) | ||||||||
811 | if (Pair.second == InitialPreheader && Pair.first != BI->getParent()) | ||||||||
812 | Pair.second = HoistCommonSucc; | ||||||||
813 | } | ||||||||
814 | |||||||||
815 | // Now finally clone BI. | ||||||||
816 | ReplaceInstWithInst( | ||||||||
817 | HoistTarget->getTerminator(), | ||||||||
818 | BranchInst::Create(HoistTrueDest, HoistFalseDest, BI->getCondition())); | ||||||||
819 | ++NumClonedBranches; | ||||||||
820 | |||||||||
821 | assert(CurLoop->getLoopPreheader() &&((CurLoop->getLoopPreheader() && "Hoisting blocks should not have destroyed preheader" ) ? static_cast<void> (0) : __assert_fail ("CurLoop->getLoopPreheader() && \"Hoisting blocks should not have destroyed preheader\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 822, __PRETTY_FUNCTION__)) | ||||||||
822 | "Hoisting blocks should not have destroyed preheader")((CurLoop->getLoopPreheader() && "Hoisting blocks should not have destroyed preheader" ) ? static_cast<void> (0) : __assert_fail ("CurLoop->getLoopPreheader() && \"Hoisting blocks should not have destroyed preheader\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 822, __PRETTY_FUNCTION__)); | ||||||||
823 | return HoistDestinationMap[BB]; | ||||||||
824 | } | ||||||||
825 | }; | ||||||||
826 | } // namespace | ||||||||
827 | |||||||||
828 | // Hoisting/sinking instruction out of a loop isn't always beneficial. It's only | ||||||||
829 | // only worthwhile if the destination block is actually colder than current | ||||||||
830 | // block. | ||||||||
831 | static bool worthSinkOrHoistInst(Instruction &I, BasicBlock *DstBlock, | ||||||||
832 | OptimizationRemarkEmitter *ORE, | ||||||||
833 | BlockFrequencyInfo *BFI) { | ||||||||
834 | // Check block frequency only when runtime profile is available | ||||||||
835 | // to avoid pathological cases. With static profile, lean towards | ||||||||
836 | // hosting because it helps canonicalize the loop for vectorizer. | ||||||||
837 | if (!DstBlock->getParent()->hasProfileData()) | ||||||||
838 | return true; | ||||||||
839 | |||||||||
840 | if (!HoistSinkColdnessThreshold || !BFI) | ||||||||
841 | return true; | ||||||||
842 | |||||||||
843 | BasicBlock *SrcBlock = I.getParent(); | ||||||||
844 | if (BFI->getBlockFreq(DstBlock).getFrequency() / HoistSinkColdnessThreshold > | ||||||||
845 | BFI->getBlockFreq(SrcBlock).getFrequency()) { | ||||||||
846 | ORE->emit([&]() { | ||||||||
847 | return OptimizationRemarkMissed(DEBUG_TYPE"licm", "SinkHoistInst", &I) | ||||||||
848 | << "failed to sink or hoist instruction because containing block " | ||||||||
849 | "has lower frequency than destination block"; | ||||||||
850 | }); | ||||||||
851 | return false; | ||||||||
852 | } | ||||||||
853 | |||||||||
854 | return true; | ||||||||
855 | } | ||||||||
856 | |||||||||
857 | /// Walk the specified region of the CFG (defined by all blocks dominated by | ||||||||
858 | /// the specified block, and that are in the current loop) in depth first | ||||||||
859 | /// order w.r.t the DominatorTree. This allows us to visit definitions before | ||||||||
860 | /// uses, allowing us to hoist a loop body in one pass without iteration. | ||||||||
861 | /// | ||||||||
862 | bool llvm::hoistRegion(DomTreeNode *N, AAResults *AA, LoopInfo *LI, | ||||||||
863 | DominatorTree *DT, BlockFrequencyInfo *BFI, | ||||||||
864 | TargetLibraryInfo *TLI, Loop *CurLoop, | ||||||||
865 | AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU, | ||||||||
866 | ScalarEvolution *SE, ICFLoopSafetyInfo *SafetyInfo, | ||||||||
867 | SinkAndHoistLICMFlags &Flags, | ||||||||
868 | OptimizationRemarkEmitter *ORE) { | ||||||||
869 | // Verify inputs. | ||||||||
870 | assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&((N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && "Unexpected input to hoistRegion." ) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to hoistRegion.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 872, __PRETTY_FUNCTION__)) | ||||||||
871 | CurLoop != nullptr && SafetyInfo != nullptr &&((N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && "Unexpected input to hoistRegion." ) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to hoistRegion.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 872, __PRETTY_FUNCTION__)) | ||||||||
872 | "Unexpected input to hoistRegion.")((N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && "Unexpected input to hoistRegion." ) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to hoistRegion.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 872, __PRETTY_FUNCTION__)); | ||||||||
873 | assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&((((CurAST != nullptr) ^ (MSSAU != nullptr)) && "Either AliasSetTracker or MemorySSA should be initialized." ) ? static_cast<void> (0) : __assert_fail ("((CurAST != nullptr) ^ (MSSAU != nullptr)) && \"Either AliasSetTracker or MemorySSA should be initialized.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 874, __PRETTY_FUNCTION__)) | ||||||||
874 | "Either AliasSetTracker or MemorySSA should be initialized.")((((CurAST != nullptr) ^ (MSSAU != nullptr)) && "Either AliasSetTracker or MemorySSA should be initialized." ) ? static_cast<void> (0) : __assert_fail ("((CurAST != nullptr) ^ (MSSAU != nullptr)) && \"Either AliasSetTracker or MemorySSA should be initialized.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 874, __PRETTY_FUNCTION__)); | ||||||||
875 | |||||||||
876 | ControlFlowHoister CFH(LI, DT, CurLoop, MSSAU); | ||||||||
877 | |||||||||
878 | // Keep track of instructions that have been hoisted, as they may need to be | ||||||||
879 | // re-hoisted if they end up not dominating all of their uses. | ||||||||
880 | SmallVector<Instruction *, 16> HoistedInstructions; | ||||||||
881 | |||||||||
882 | // For PHI hoisting to work we need to hoist blocks before their successors. | ||||||||
883 | // We can do this by iterating through the blocks in the loop in reverse | ||||||||
884 | // post-order. | ||||||||
885 | LoopBlocksRPO Worklist(CurLoop); | ||||||||
886 | Worklist.perform(LI); | ||||||||
887 | bool Changed = false; | ||||||||
888 | for (BasicBlock *BB : Worklist) { | ||||||||
889 | // Only need to process the contents of this block if it is not part of a | ||||||||
890 | // subloop (which would already have been processed). | ||||||||
891 | if (inSubLoop(BB, CurLoop, LI)) | ||||||||
892 | continue; | ||||||||
893 | |||||||||
894 | for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) { | ||||||||
895 | Instruction &I = *II++; | ||||||||
896 | // Try constant folding this instruction. If all the operands are | ||||||||
897 | // constants, it is technically hoistable, but it would be better to | ||||||||
898 | // just fold it. | ||||||||
899 | if (Constant *C = ConstantFoldInstruction( | ||||||||
900 | &I, I.getModule()->getDataLayout(), TLI)) { | ||||||||
901 | LLVM_DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *Cdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n'; } } while (false) | ||||||||
902 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n'; } } while (false); | ||||||||
903 | if (CurAST) | ||||||||
904 | CurAST->copyValue(&I, C); | ||||||||
905 | // FIXME MSSA: Such replacements may make accesses unoptimized (D51960). | ||||||||
906 | I.replaceAllUsesWith(C); | ||||||||
907 | if (isInstructionTriviallyDead(&I, TLI)) | ||||||||
908 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | ||||||||
909 | Changed = true; | ||||||||
910 | continue; | ||||||||
911 | } | ||||||||
912 | |||||||||
913 | // Try hoisting the instruction out to the preheader. We can only do | ||||||||
914 | // this if all of the operands of the instruction are loop invariant and | ||||||||
915 | // if it is safe to hoist the instruction. We also check block frequency | ||||||||
916 | // to make sure instruction only gets hoisted into colder blocks. | ||||||||
917 | // TODO: It may be safe to hoist if we are hoisting to a conditional block | ||||||||
918 | // and we have accurately duplicated the control flow from the loop header | ||||||||
919 | // to that block. | ||||||||
920 | if (CurLoop->hasLoopInvariantOperands(&I) && | ||||||||
921 | canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags, | ||||||||
922 | ORE) && | ||||||||
923 | worthSinkOrHoistInst(I, CurLoop->getLoopPreheader(), ORE, BFI) && | ||||||||
924 | isSafeToExecuteUnconditionally( | ||||||||
925 | I, DT, TLI, CurLoop, SafetyInfo, ORE, | ||||||||
926 | CurLoop->getLoopPreheader()->getTerminator())) { | ||||||||
927 | hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo, | ||||||||
928 | MSSAU, SE, ORE); | ||||||||
929 | HoistedInstructions.push_back(&I); | ||||||||
930 | Changed = true; | ||||||||
931 | continue; | ||||||||
932 | } | ||||||||
933 | |||||||||
934 | // Attempt to remove floating point division out of the loop by | ||||||||
935 | // converting it to a reciprocal multiplication. | ||||||||
936 | if (I.getOpcode() == Instruction::FDiv && I.hasAllowReciprocal() && | ||||||||
937 | CurLoop->isLoopInvariant(I.getOperand(1))) { | ||||||||
938 | auto Divisor = I.getOperand(1); | ||||||||
939 | auto One = llvm::ConstantFP::get(Divisor->getType(), 1.0); | ||||||||
940 | auto ReciprocalDivisor = BinaryOperator::CreateFDiv(One, Divisor); | ||||||||
941 | ReciprocalDivisor->setFastMathFlags(I.getFastMathFlags()); | ||||||||
942 | SafetyInfo->insertInstructionTo(ReciprocalDivisor, I.getParent()); | ||||||||
943 | ReciprocalDivisor->insertBefore(&I); | ||||||||
944 | |||||||||
945 | auto Product = | ||||||||
946 | BinaryOperator::CreateFMul(I.getOperand(0), ReciprocalDivisor); | ||||||||
947 | Product->setFastMathFlags(I.getFastMathFlags()); | ||||||||
948 | SafetyInfo->insertInstructionTo(Product, I.getParent()); | ||||||||
949 | Product->insertAfter(&I); | ||||||||
950 | I.replaceAllUsesWith(Product); | ||||||||
951 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | ||||||||
952 | |||||||||
953 | hoist(*ReciprocalDivisor, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), | ||||||||
954 | SafetyInfo, MSSAU, SE, ORE); | ||||||||
955 | HoistedInstructions.push_back(ReciprocalDivisor); | ||||||||
956 | Changed = true; | ||||||||
957 | continue; | ||||||||
958 | } | ||||||||
959 | |||||||||
960 | auto IsInvariantStart = [&](Instruction &I) { | ||||||||
961 | using namespace PatternMatch; | ||||||||
962 | return I.use_empty() && | ||||||||
963 | match(&I, m_Intrinsic<Intrinsic::invariant_start>()); | ||||||||
964 | }; | ||||||||
965 | auto MustExecuteWithoutWritesBefore = [&](Instruction &I) { | ||||||||
966 | return SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop) && | ||||||||
967 | SafetyInfo->doesNotWriteMemoryBefore(I, CurLoop); | ||||||||
968 | }; | ||||||||
969 | if ((IsInvariantStart(I) || isGuard(&I)) && | ||||||||
970 | CurLoop->hasLoopInvariantOperands(&I) && | ||||||||
971 | MustExecuteWithoutWritesBefore(I)) { | ||||||||
972 | hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo, | ||||||||
973 | MSSAU, SE, ORE); | ||||||||
974 | HoistedInstructions.push_back(&I); | ||||||||
975 | Changed = true; | ||||||||
976 | continue; | ||||||||
977 | } | ||||||||
978 | |||||||||
979 | if (PHINode *PN = dyn_cast<PHINode>(&I)) { | ||||||||
980 | if (CFH.canHoistPHI(PN)) { | ||||||||
981 | // Redirect incoming blocks first to ensure that we create hoisted | ||||||||
982 | // versions of those blocks before we hoist the phi. | ||||||||
983 | for (unsigned int i = 0; i < PN->getNumIncomingValues(); ++i) | ||||||||
984 | PN->setIncomingBlock( | ||||||||
985 | i, CFH.getOrCreateHoistedBlock(PN->getIncomingBlock(i))); | ||||||||
986 | hoist(*PN, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo, | ||||||||
987 | MSSAU, SE, ORE); | ||||||||
988 | assert(DT->dominates(PN, BB) && "Conditional PHIs not expected")((DT->dominates(PN, BB) && "Conditional PHIs not expected" ) ? static_cast<void> (0) : __assert_fail ("DT->dominates(PN, BB) && \"Conditional PHIs not expected\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 988, __PRETTY_FUNCTION__)); | ||||||||
989 | Changed = true; | ||||||||
990 | continue; | ||||||||
991 | } | ||||||||
992 | } | ||||||||
993 | |||||||||
994 | // Remember possibly hoistable branches so we can actually hoist them | ||||||||
995 | // later if needed. | ||||||||
996 | if (BranchInst *BI = dyn_cast<BranchInst>(&I)) | ||||||||
997 | CFH.registerPossiblyHoistableBranch(BI); | ||||||||
998 | } | ||||||||
999 | } | ||||||||
1000 | |||||||||
1001 | // If we hoisted instructions to a conditional block they may not dominate | ||||||||
1002 | // their uses that weren't hoisted (such as phis where some operands are not | ||||||||
1003 | // loop invariant). If so make them unconditional by moving them to their | ||||||||
1004 | // immediate dominator. We iterate through the instructions in reverse order | ||||||||
1005 | // which ensures that when we rehoist an instruction we rehoist its operands, | ||||||||
1006 | // and also keep track of where in the block we are rehoisting to to make sure | ||||||||
1007 | // that we rehoist instructions before the instructions that use them. | ||||||||
1008 | Instruction *HoistPoint = nullptr; | ||||||||
1009 | if (ControlFlowHoisting) { | ||||||||
1010 | for (Instruction *I : reverse(HoistedInstructions)) { | ||||||||
1011 | if (!llvm::all_of(I->uses(), | ||||||||
1012 | [&](Use &U) { return DT->dominates(I, U); })) { | ||||||||
1013 | BasicBlock *Dominator = | ||||||||
1014 | DT->getNode(I->getParent())->getIDom()->getBlock(); | ||||||||
1015 | if (!HoistPoint || !DT->dominates(HoistPoint->getParent(), Dominator)) { | ||||||||
1016 | if (HoistPoint) | ||||||||
1017 | assert(DT->dominates(Dominator, HoistPoint->getParent()) &&((DT->dominates(Dominator, HoistPoint->getParent()) && "New hoist point expected to dominate old hoist point") ? static_cast <void> (0) : __assert_fail ("DT->dominates(Dominator, HoistPoint->getParent()) && \"New hoist point expected to dominate old hoist point\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1018, __PRETTY_FUNCTION__)) | ||||||||
1018 | "New hoist point expected to dominate old hoist point")((DT->dominates(Dominator, HoistPoint->getParent()) && "New hoist point expected to dominate old hoist point") ? static_cast <void> (0) : __assert_fail ("DT->dominates(Dominator, HoistPoint->getParent()) && \"New hoist point expected to dominate old hoist point\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1018, __PRETTY_FUNCTION__)); | ||||||||
1019 | HoistPoint = Dominator->getTerminator(); | ||||||||
1020 | } | ||||||||
1021 | LLVM_DEBUG(dbgs() << "LICM rehoisting to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM rehoisting to " << HoistPoint ->getParent()->getNameOrAsOperand() << ": " << *I << "\n"; } } while (false) | ||||||||
1022 | << HoistPoint->getParent()->getNameOrAsOperand()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM rehoisting to " << HoistPoint ->getParent()->getNameOrAsOperand() << ": " << *I << "\n"; } } while (false) | ||||||||
1023 | << ": " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM rehoisting to " << HoistPoint ->getParent()->getNameOrAsOperand() << ": " << *I << "\n"; } } while (false); | ||||||||
1024 | moveInstructionBefore(*I, *HoistPoint, *SafetyInfo, MSSAU, SE); | ||||||||
1025 | HoistPoint = I; | ||||||||
1026 | Changed = true; | ||||||||
1027 | } | ||||||||
1028 | } | ||||||||
1029 | } | ||||||||
1030 | if (MSSAU && VerifyMemorySSA) | ||||||||
1031 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||
1032 | |||||||||
1033 | // Now that we've finished hoisting make sure that LI and DT are still | ||||||||
1034 | // valid. | ||||||||
1035 | #ifdef EXPENSIVE_CHECKS | ||||||||
1036 | if (Changed) { | ||||||||
1037 | assert(DT->verify(DominatorTree::VerificationLevel::Fast) &&((DT->verify(DominatorTree::VerificationLevel::Fast) && "Dominator tree verification failed") ? static_cast<void> (0) : __assert_fail ("DT->verify(DominatorTree::VerificationLevel::Fast) && \"Dominator tree verification failed\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1038, __PRETTY_FUNCTION__)) | ||||||||
1038 | "Dominator tree verification failed")((DT->verify(DominatorTree::VerificationLevel::Fast) && "Dominator tree verification failed") ? static_cast<void> (0) : __assert_fail ("DT->verify(DominatorTree::VerificationLevel::Fast) && \"Dominator tree verification failed\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1038, __PRETTY_FUNCTION__)); | ||||||||
1039 | LI->verify(*DT); | ||||||||
1040 | } | ||||||||
1041 | #endif | ||||||||
1042 | |||||||||
1043 | return Changed; | ||||||||
1044 | } | ||||||||
1045 | |||||||||
1046 | // Return true if LI is invariant within scope of the loop. LI is invariant if | ||||||||
1047 | // CurLoop is dominated by an invariant.start representing the same memory | ||||||||
1048 | // location and size as the memory location LI loads from, and also the | ||||||||
1049 | // invariant.start has no uses. | ||||||||
1050 | static bool isLoadInvariantInLoop(LoadInst *LI, DominatorTree *DT, | ||||||||
1051 | Loop *CurLoop) { | ||||||||
1052 | Value *Addr = LI->getOperand(0); | ||||||||
1053 | const DataLayout &DL = LI->getModule()->getDataLayout(); | ||||||||
1054 | const TypeSize LocSizeInBits = DL.getTypeSizeInBits(LI->getType()); | ||||||||
1055 | |||||||||
1056 | // It is not currently possible for clang to generate an invariant.start | ||||||||
1057 | // intrinsic with scalable vector types because we don't support thread local | ||||||||
1058 | // sizeless types and we don't permit sizeless types in structs or classes. | ||||||||
1059 | // Furthermore, even if support is added for this in future the intrinsic | ||||||||
1060 | // itself is defined to have a size of -1 for variable sized objects. This | ||||||||
1061 | // makes it impossible to verify if the intrinsic envelops our region of | ||||||||
1062 | // interest. For example, both <vscale x 32 x i8> and <vscale x 16 x i8> | ||||||||
1063 | // types would have a -1 parameter, but the former is clearly double the size | ||||||||
1064 | // of the latter. | ||||||||
1065 | if (LocSizeInBits.isScalable()) | ||||||||
1066 | return false; | ||||||||
1067 | |||||||||
1068 | // if the type is i8 addrspace(x)*, we know this is the type of | ||||||||
1069 | // llvm.invariant.start operand | ||||||||
1070 | auto *PtrInt8Ty = PointerType::get(Type::getInt8Ty(LI->getContext()), | ||||||||
1071 | LI->getPointerAddressSpace()); | ||||||||
1072 | unsigned BitcastsVisited = 0; | ||||||||
1073 | // Look through bitcasts until we reach the i8* type (this is invariant.start | ||||||||
1074 | // operand type). | ||||||||
1075 | while (Addr->getType() != PtrInt8Ty) { | ||||||||
1076 | auto *BC = dyn_cast<BitCastInst>(Addr); | ||||||||
1077 | // Avoid traversing high number of bitcast uses. | ||||||||
1078 | if (++BitcastsVisited > MaxNumUsesTraversed || !BC) | ||||||||
1079 | return false; | ||||||||
1080 | Addr = BC->getOperand(0); | ||||||||
1081 | } | ||||||||
1082 | |||||||||
1083 | unsigned UsesVisited = 0; | ||||||||
1084 | // Traverse all uses of the load operand value, to see if invariant.start is | ||||||||
1085 | // one of the uses, and whether it dominates the load instruction. | ||||||||
1086 | for (auto *U : Addr->users()) { | ||||||||
1087 | // Avoid traversing for Load operand with high number of users. | ||||||||
1088 | if (++UsesVisited > MaxNumUsesTraversed) | ||||||||
1089 | return false; | ||||||||
1090 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(U); | ||||||||
1091 | // If there are escaping uses of invariant.start instruction, the load maybe | ||||||||
1092 | // non-invariant. | ||||||||
1093 | if (!II || II->getIntrinsicID() != Intrinsic::invariant_start || | ||||||||
1094 | !II->use_empty()) | ||||||||
1095 | continue; | ||||||||
1096 | ConstantInt *InvariantSize = cast<ConstantInt>(II->getArgOperand(0)); | ||||||||
1097 | // The intrinsic supports having a -1 argument for variable sized objects | ||||||||
1098 | // so we should check for that here. | ||||||||
1099 | if (InvariantSize->isNegative()) | ||||||||
1100 | continue; | ||||||||
1101 | uint64_t InvariantSizeInBits = InvariantSize->getSExtValue() * 8; | ||||||||
1102 | // Confirm the invariant.start location size contains the load operand size | ||||||||
1103 | // in bits. Also, the invariant.start should dominate the load, and we | ||||||||
1104 | // should not hoist the load out of a loop that contains this dominating | ||||||||
1105 | // invariant.start. | ||||||||
1106 | if (LocSizeInBits.getFixedSize() <= InvariantSizeInBits && | ||||||||
1107 | DT->properlyDominates(II->getParent(), CurLoop->getHeader())) | ||||||||
1108 | return true; | ||||||||
1109 | } | ||||||||
1110 | |||||||||
1111 | return false; | ||||||||
1112 | } | ||||||||
1113 | |||||||||
1114 | namespace { | ||||||||
1115 | /// Return true if-and-only-if we know how to (mechanically) both hoist and | ||||||||
1116 | /// sink a given instruction out of a loop. Does not address legality | ||||||||
1117 | /// concerns such as aliasing or speculation safety. | ||||||||
1118 | bool isHoistableAndSinkableInst(Instruction &I) { | ||||||||
1119 | // Only these instructions are hoistable/sinkable. | ||||||||
1120 | return (isa<LoadInst>(I) || isa<StoreInst>(I) || isa<CallInst>(I) || | ||||||||
1121 | isa<FenceInst>(I) || isa<CastInst>(I) || isa<UnaryOperator>(I) || | ||||||||
1122 | isa<BinaryOperator>(I) || isa<SelectInst>(I) || | ||||||||
1123 | isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || | ||||||||
1124 | isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || | ||||||||
1125 | isa<ShuffleVectorInst>(I) || isa<ExtractValueInst>(I) || | ||||||||
1126 | isa<InsertValueInst>(I) || isa<FreezeInst>(I)); | ||||||||
1127 | } | ||||||||
1128 | /// Return true if all of the alias sets within this AST are known not to | ||||||||
1129 | /// contain a Mod, or if MSSA knows thare are no MemoryDefs in the loop. | ||||||||
1130 | bool isReadOnly(AliasSetTracker *CurAST, const MemorySSAUpdater *MSSAU, | ||||||||
1131 | const Loop *L) { | ||||||||
1132 | if (CurAST) { | ||||||||
1133 | for (AliasSet &AS : *CurAST) { | ||||||||
1134 | if (!AS.isForwardingAliasSet() && AS.isMod()) { | ||||||||
1135 | return false; | ||||||||
1136 | } | ||||||||
1137 | } | ||||||||
1138 | return true; | ||||||||
1139 | } else { /*MSSAU*/ | ||||||||
1140 | for (auto *BB : L->getBlocks()) | ||||||||
1141 | if (MSSAU->getMemorySSA()->getBlockDefs(BB)) | ||||||||
1142 | return false; | ||||||||
1143 | return true; | ||||||||
1144 | } | ||||||||
1145 | } | ||||||||
1146 | |||||||||
1147 | /// Return true if I is the only Instruction with a MemoryAccess in L. | ||||||||
1148 | bool isOnlyMemoryAccess(const Instruction *I, const Loop *L, | ||||||||
1149 | const MemorySSAUpdater *MSSAU) { | ||||||||
1150 | for (auto *BB : L->getBlocks()) | ||||||||
1151 | if (auto *Accs = MSSAU->getMemorySSA()->getBlockAccesses(BB)) { | ||||||||
1152 | int NotAPhi = 0; | ||||||||
1153 | for (const auto &Acc : *Accs) { | ||||||||
1154 | if (isa<MemoryPhi>(&Acc)) | ||||||||
1155 | continue; | ||||||||
1156 | const auto *MUD = cast<MemoryUseOrDef>(&Acc); | ||||||||
1157 | if (MUD->getMemoryInst() != I || NotAPhi++ == 1) | ||||||||
1158 | return false; | ||||||||
1159 | } | ||||||||
1160 | } | ||||||||
1161 | return true; | ||||||||
1162 | } | ||||||||
1163 | } | ||||||||
1164 | |||||||||
1165 | bool llvm::canSinkOrHoistInst(Instruction &I, AAResults *AA, DominatorTree *DT, | ||||||||
1166 | Loop *CurLoop, AliasSetTracker *CurAST, | ||||||||
1167 | MemorySSAUpdater *MSSAU, | ||||||||
1168 | bool TargetExecutesOncePerLoop, | ||||||||
1169 | SinkAndHoistLICMFlags *Flags, | ||||||||
1170 | OptimizationRemarkEmitter *ORE) { | ||||||||
1171 | assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&((((CurAST != nullptr) ^ (MSSAU != nullptr)) && "Either AliasSetTracker or MemorySSA should be initialized." ) ? static_cast<void> (0) : __assert_fail ("((CurAST != nullptr) ^ (MSSAU != nullptr)) && \"Either AliasSetTracker or MemorySSA should be initialized.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1172, __PRETTY_FUNCTION__)) | ||||||||
| |||||||||
1172 | "Either AliasSetTracker or MemorySSA should be initialized.")((((CurAST != nullptr) ^ (MSSAU != nullptr)) && "Either AliasSetTracker or MemorySSA should be initialized." ) ? static_cast<void> (0) : __assert_fail ("((CurAST != nullptr) ^ (MSSAU != nullptr)) && \"Either AliasSetTracker or MemorySSA should be initialized.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1172, __PRETTY_FUNCTION__)); | ||||||||
1173 | |||||||||
1174 | // If we don't understand the instruction, bail early. | ||||||||
1175 | if (!isHoistableAndSinkableInst(I)) | ||||||||
1176 | return false; | ||||||||
1177 | |||||||||
1178 | MemorySSA *MSSA = MSSAU
| ||||||||
1179 | if (MSSA) | ||||||||
1180 | assert(Flags != nullptr && "Flags cannot be null.")((Flags != nullptr && "Flags cannot be null.") ? static_cast <void> (0) : __assert_fail ("Flags != nullptr && \"Flags cannot be null.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1180, __PRETTY_FUNCTION__)); | ||||||||
1181 | |||||||||
1182 | // Loads have extra constraints we have to verify before we can hoist them. | ||||||||
1183 | if (LoadInst *LI
| ||||||||
1184 | if (!LI->isUnordered()) | ||||||||
1185 | return false; // Don't sink/hoist volatile or ordered atomic loads! | ||||||||
1186 | |||||||||
1187 | // Loads from constant memory are always safe to move, even if they end up | ||||||||
1188 | // in the same alias set as something that ends up being modified. | ||||||||
1189 | if (AA->pointsToConstantMemory(LI->getOperand(0))) | ||||||||
1190 | return true; | ||||||||
1191 | if (LI->hasMetadata(LLVMContext::MD_invariant_load)) | ||||||||
1192 | return true; | ||||||||
1193 | |||||||||
1194 | if (LI->isAtomic() && !TargetExecutesOncePerLoop) | ||||||||
1195 | return false; // Don't risk duplicating unordered loads | ||||||||
1196 | |||||||||
1197 | // This checks for an invariant.start dominating the load. | ||||||||
1198 | if (isLoadInvariantInLoop(LI, DT, CurLoop)) | ||||||||
1199 | return true; | ||||||||
1200 | |||||||||
1201 | bool Invalidated; | ||||||||
1202 | if (CurAST
| ||||||||
1203 | Invalidated = pointerInvalidatedByLoop(MemoryLocation::get(LI), CurAST, | ||||||||
1204 | CurLoop, AA); | ||||||||
1205 | else | ||||||||
1206 | Invalidated = pointerInvalidatedByLoopWithMSSA( | ||||||||
1207 | MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(LI)), CurLoop, I, *Flags); | ||||||||
| |||||||||
1208 | // Check loop-invariant address because this may also be a sinkable load | ||||||||
1209 | // whose address is not necessarily loop-invariant. | ||||||||
1210 | if (ORE && Invalidated && CurLoop->isLoopInvariant(LI->getPointerOperand())) | ||||||||
1211 | ORE->emit([&]() { | ||||||||
1212 | return OptimizationRemarkMissed( | ||||||||
1213 | DEBUG_TYPE"licm", "LoadWithLoopInvariantAddressInvalidated", LI) | ||||||||
1214 | << "failed to move load with loop-invariant address " | ||||||||
1215 | "because the loop may invalidate its value"; | ||||||||
1216 | }); | ||||||||
1217 | |||||||||
1218 | return !Invalidated; | ||||||||
1219 | } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { | ||||||||
1220 | // Don't sink or hoist dbg info; it's legal, but not useful. | ||||||||
1221 | if (isa<DbgInfoIntrinsic>(I)) | ||||||||
1222 | return false; | ||||||||
1223 | |||||||||
1224 | // Don't sink calls which can throw. | ||||||||
1225 | if (CI->mayThrow()) | ||||||||
1226 | return false; | ||||||||
1227 | |||||||||
1228 | // Convergent attribute has been used on operations that involve | ||||||||
1229 | // inter-thread communication which results are implicitly affected by the | ||||||||
1230 | // enclosing control flows. It is not safe to hoist or sink such operations | ||||||||
1231 | // across control flow. | ||||||||
1232 | if (CI->isConvergent()) | ||||||||
1233 | return false; | ||||||||
1234 | |||||||||
1235 | using namespace PatternMatch; | ||||||||
1236 | if (match(CI, m_Intrinsic<Intrinsic::assume>())) | ||||||||
1237 | // Assumes don't actually alias anything or throw | ||||||||
1238 | return true; | ||||||||
1239 | |||||||||
1240 | if (match(CI, m_Intrinsic<Intrinsic::experimental_widenable_condition>())) | ||||||||
1241 | // Widenable conditions don't actually alias anything or throw | ||||||||
1242 | return true; | ||||||||
1243 | |||||||||
1244 | // Handle simple cases by querying alias analysis. | ||||||||
1245 | FunctionModRefBehavior Behavior = AA->getModRefBehavior(CI); | ||||||||
1246 | if (Behavior == FMRB_DoesNotAccessMemory) | ||||||||
1247 | return true; | ||||||||
1248 | if (AAResults::onlyReadsMemory(Behavior)) { | ||||||||
1249 | // A readonly argmemonly function only reads from memory pointed to by | ||||||||
1250 | // it's arguments with arbitrary offsets. If we can prove there are no | ||||||||
1251 | // writes to this memory in the loop, we can hoist or sink. | ||||||||
1252 | if (AAResults::onlyAccessesArgPointees(Behavior)) { | ||||||||
1253 | // TODO: expand to writeable arguments | ||||||||
1254 | for (Value *Op : CI->arg_operands()) | ||||||||
1255 | if (Op->getType()->isPointerTy()) { | ||||||||
1256 | bool Invalidated; | ||||||||
1257 | if (CurAST) | ||||||||
1258 | Invalidated = pointerInvalidatedByLoop( | ||||||||
1259 | MemoryLocation::getBeforeOrAfter(Op), CurAST, CurLoop, AA); | ||||||||
1260 | else | ||||||||
1261 | Invalidated = pointerInvalidatedByLoopWithMSSA( | ||||||||
1262 | MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(CI)), CurLoop, I, | ||||||||
1263 | *Flags); | ||||||||
1264 | if (Invalidated) | ||||||||
1265 | return false; | ||||||||
1266 | } | ||||||||
1267 | return true; | ||||||||
1268 | } | ||||||||
1269 | |||||||||
1270 | // If this call only reads from memory and there are no writes to memory | ||||||||
1271 | // in the loop, we can hoist or sink the call as appropriate. | ||||||||
1272 | if (isReadOnly(CurAST, MSSAU, CurLoop)) | ||||||||
1273 | return true; | ||||||||
1274 | } | ||||||||
1275 | |||||||||
1276 | // FIXME: This should use mod/ref information to see if we can hoist or | ||||||||
1277 | // sink the call. | ||||||||
1278 | |||||||||
1279 | return false; | ||||||||
1280 | } else if (auto *FI = dyn_cast<FenceInst>(&I)) { | ||||||||
1281 | // Fences alias (most) everything to provide ordering. For the moment, | ||||||||
1282 | // just give up if there are any other memory operations in the loop. | ||||||||
1283 | if (CurAST) { | ||||||||
1284 | auto Begin = CurAST->begin(); | ||||||||
1285 | assert(Begin != CurAST->end() && "must contain FI")((Begin != CurAST->end() && "must contain FI") ? static_cast <void> (0) : __assert_fail ("Begin != CurAST->end() && \"must contain FI\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1285, __PRETTY_FUNCTION__)); | ||||||||
1286 | if (std::next(Begin) != CurAST->end()) | ||||||||
1287 | // constant memory for instance, TODO: handle better | ||||||||
1288 | return false; | ||||||||
1289 | auto *UniqueI = Begin->getUniqueInstruction(); | ||||||||
1290 | if (!UniqueI) | ||||||||
1291 | // other memory op, give up | ||||||||
1292 | return false; | ||||||||
1293 | (void)FI; // suppress unused variable warning | ||||||||
1294 | assert(UniqueI == FI && "AS must contain FI")((UniqueI == FI && "AS must contain FI") ? static_cast <void> (0) : __assert_fail ("UniqueI == FI && \"AS must contain FI\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1294, __PRETTY_FUNCTION__)); | ||||||||
1295 | return true; | ||||||||
1296 | } else // MSSAU | ||||||||
1297 | return isOnlyMemoryAccess(FI, CurLoop, MSSAU); | ||||||||
1298 | } else if (auto *SI = dyn_cast<StoreInst>(&I)) { | ||||||||
1299 | if (!SI->isUnordered()) | ||||||||
1300 | return false; // Don't sink/hoist volatile or ordered atomic store! | ||||||||
1301 | |||||||||
1302 | // We can only hoist a store that we can prove writes a value which is not | ||||||||
1303 | // read or overwritten within the loop. For those cases, we fallback to | ||||||||
1304 | // load store promotion instead. TODO: We can extend this to cases where | ||||||||
1305 | // there is exactly one write to the location and that write dominates an | ||||||||
1306 | // arbitrary number of reads in the loop. | ||||||||
1307 | if (CurAST) { | ||||||||
1308 | auto &AS = CurAST->getAliasSetFor(MemoryLocation::get(SI)); | ||||||||
1309 | |||||||||
1310 | if (AS.isRef() || !AS.isMustAlias()) | ||||||||
1311 | // Quick exit test, handled by the full path below as well. | ||||||||
1312 | return false; | ||||||||
1313 | auto *UniqueI = AS.getUniqueInstruction(); | ||||||||
1314 | if (!UniqueI) | ||||||||
1315 | // other memory op, give up | ||||||||
1316 | return false; | ||||||||
1317 | assert(UniqueI == SI && "AS must contain SI")((UniqueI == SI && "AS must contain SI") ? static_cast <void> (0) : __assert_fail ("UniqueI == SI && \"AS must contain SI\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1317, __PRETTY_FUNCTION__)); | ||||||||
1318 | return true; | ||||||||
1319 | } else { // MSSAU | ||||||||
1320 | if (isOnlyMemoryAccess(SI, CurLoop, MSSAU)) | ||||||||
1321 | return true; | ||||||||
1322 | // If there are more accesses than the Promotion cap or no "quota" to | ||||||||
1323 | // check clobber, then give up as we're not walking a list that long. | ||||||||
1324 | if (Flags->tooManyMemoryAccesses() || Flags->tooManyClobberingCalls()) | ||||||||
1325 | return false; | ||||||||
1326 | // If there are interfering Uses (i.e. their defining access is in the | ||||||||
1327 | // loop), or ordered loads (stored as Defs!), don't move this store. | ||||||||
1328 | // Could do better here, but this is conservatively correct. | ||||||||
1329 | // TODO: Cache set of Uses on the first walk in runOnLoop, update when | ||||||||
1330 | // moving accesses. Can also extend to dominating uses. | ||||||||
1331 | auto *SIMD = MSSA->getMemoryAccess(SI); | ||||||||
1332 | for (auto *BB : CurLoop->getBlocks()) | ||||||||
1333 | if (auto *Accesses = MSSA->getBlockAccesses(BB)) { | ||||||||
1334 | for (const auto &MA : *Accesses) | ||||||||
1335 | if (const auto *MU = dyn_cast<MemoryUse>(&MA)) { | ||||||||
1336 | auto *MD = MU->getDefiningAccess(); | ||||||||
1337 | if (!MSSA->isLiveOnEntryDef(MD) && | ||||||||
1338 | CurLoop->contains(MD->getBlock())) | ||||||||
1339 | return false; | ||||||||
1340 | // Disable hoisting past potentially interfering loads. Optimized | ||||||||
1341 | // Uses may point to an access outside the loop, as getClobbering | ||||||||
1342 | // checks the previous iteration when walking the backedge. | ||||||||
1343 | // FIXME: More precise: no Uses that alias SI. | ||||||||
1344 | if (!Flags->getIsSink() && !MSSA->dominates(SIMD, MU)) | ||||||||
1345 | return false; | ||||||||
1346 | } else if (const auto *MD = dyn_cast<MemoryDef>(&MA)) { | ||||||||
1347 | if (auto *LI = dyn_cast<LoadInst>(MD->getMemoryInst())) { | ||||||||
1348 | (void)LI; // Silence warning. | ||||||||
1349 | assert(!LI->isUnordered() && "Expected unordered load")((!LI->isUnordered() && "Expected unordered load") ? static_cast<void> (0) : __assert_fail ("!LI->isUnordered() && \"Expected unordered load\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1349, __PRETTY_FUNCTION__)); | ||||||||
1350 | return false; | ||||||||
1351 | } | ||||||||
1352 | // Any call, while it may not be clobbering SI, it may be a use. | ||||||||
1353 | if (auto *CI = dyn_cast<CallInst>(MD->getMemoryInst())) { | ||||||||
1354 | // Check if the call may read from the memory locattion written | ||||||||
1355 | // to by SI. Check CI's attributes and arguments; the number of | ||||||||
1356 | // such checks performed is limited above by NoOfMemAccTooLarge. | ||||||||
1357 | ModRefInfo MRI = AA->getModRefInfo(CI, MemoryLocation::get(SI)); | ||||||||
1358 | if (isModOrRefSet(MRI)) | ||||||||
1359 | return false; | ||||||||
1360 | } | ||||||||
1361 | } | ||||||||
1362 | } | ||||||||
1363 | auto *Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(SI); | ||||||||
1364 | Flags->incrementClobberingCalls(); | ||||||||
1365 | // If there are no clobbering Defs in the loop, store is safe to hoist. | ||||||||
1366 | return MSSA->isLiveOnEntryDef(Source) || | ||||||||
1367 | !CurLoop->contains(Source->getBlock()); | ||||||||
1368 | } | ||||||||
1369 | } | ||||||||
1370 | |||||||||
1371 | assert(!I.mayReadOrWriteMemory() && "unhandled aliasing")((!I.mayReadOrWriteMemory() && "unhandled aliasing") ? static_cast<void> (0) : __assert_fail ("!I.mayReadOrWriteMemory() && \"unhandled aliasing\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1371, __PRETTY_FUNCTION__)); | ||||||||
1372 | |||||||||
1373 | // We've established mechanical ability and aliasing, it's up to the caller | ||||||||
1374 | // to check fault safety | ||||||||
1375 | return true; | ||||||||
1376 | } | ||||||||
1377 | |||||||||
1378 | /// Returns true if a PHINode is a trivially replaceable with an | ||||||||
1379 | /// Instruction. | ||||||||
1380 | /// This is true when all incoming values are that instruction. | ||||||||
1381 | /// This pattern occurs most often with LCSSA PHI nodes. | ||||||||
1382 | /// | ||||||||
1383 | static bool isTriviallyReplaceablePHI(const PHINode &PN, const Instruction &I) { | ||||||||
1384 | for (const Value *IncValue : PN.incoming_values()) | ||||||||
1385 | if (IncValue != &I) | ||||||||
1386 | return false; | ||||||||
1387 | |||||||||
1388 | return true; | ||||||||
1389 | } | ||||||||
1390 | |||||||||
1391 | /// Return true if the instruction is free in the loop. | ||||||||
1392 | static bool isFreeInLoop(const Instruction &I, const Loop *CurLoop, | ||||||||
1393 | const TargetTransformInfo *TTI) { | ||||||||
1394 | |||||||||
1395 | if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I)) { | ||||||||
1396 | if (TTI->getUserCost(GEP, TargetTransformInfo::TCK_SizeAndLatency) != | ||||||||
1397 | TargetTransformInfo::TCC_Free) | ||||||||
1398 | return false; | ||||||||
1399 | // For a GEP, we cannot simply use getUserCost because currently it | ||||||||
1400 | // optimistically assume that a GEP will fold into addressing mode | ||||||||
1401 | // regardless of its users. | ||||||||
1402 | const BasicBlock *BB = GEP->getParent(); | ||||||||
1403 | for (const User *U : GEP->users()) { | ||||||||
1404 | const Instruction *UI = cast<Instruction>(U); | ||||||||
1405 | if (CurLoop->contains(UI) && | ||||||||
1406 | (BB != UI->getParent() || | ||||||||
1407 | (!isa<StoreInst>(UI) && !isa<LoadInst>(UI)))) | ||||||||
1408 | return false; | ||||||||
1409 | } | ||||||||
1410 | return true; | ||||||||
1411 | } else | ||||||||
1412 | return TTI->getUserCost(&I, TargetTransformInfo::TCK_SizeAndLatency) == | ||||||||
1413 | TargetTransformInfo::TCC_Free; | ||||||||
1414 | } | ||||||||
1415 | |||||||||
1416 | /// Return true if the only users of this instruction are outside of | ||||||||
1417 | /// the loop. If this is true, we can sink the instruction to the exit | ||||||||
1418 | /// blocks of the loop. | ||||||||
1419 | /// | ||||||||
1420 | /// We also return true if the instruction could be folded away in lowering. | ||||||||
1421 | /// (e.g., a GEP can be folded into a load as an addressing mode in the loop). | ||||||||
1422 | static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop, | ||||||||
1423 | const LoopSafetyInfo *SafetyInfo, | ||||||||
1424 | TargetTransformInfo *TTI, bool &FreeInLoop) { | ||||||||
1425 | const auto &BlockColors = SafetyInfo->getBlockColors(); | ||||||||
1426 | bool IsFree = isFreeInLoop(I, CurLoop, TTI); | ||||||||
1427 | for (const User *U : I.users()) { | ||||||||
1428 | const Instruction *UI = cast<Instruction>(U); | ||||||||
1429 | if (const PHINode *PN = dyn_cast<PHINode>(UI)) { | ||||||||
1430 | const BasicBlock *BB = PN->getParent(); | ||||||||
1431 | // We cannot sink uses in catchswitches. | ||||||||
1432 | if (isa<CatchSwitchInst>(BB->getTerminator())) | ||||||||
1433 | return false; | ||||||||
1434 | |||||||||
1435 | // We need to sink a callsite to a unique funclet. Avoid sinking if the | ||||||||
1436 | // phi use is too muddled. | ||||||||
1437 | if (isa<CallInst>(I)) | ||||||||
1438 | if (!BlockColors.empty() && | ||||||||
1439 | BlockColors.find(const_cast<BasicBlock *>(BB))->second.size() != 1) | ||||||||
1440 | return false; | ||||||||
1441 | } | ||||||||
1442 | |||||||||
1443 | if (CurLoop->contains(UI)) { | ||||||||
1444 | if (IsFree) { | ||||||||
1445 | FreeInLoop = true; | ||||||||
1446 | continue; | ||||||||
1447 | } | ||||||||
1448 | return false; | ||||||||
1449 | } | ||||||||
1450 | } | ||||||||
1451 | return true; | ||||||||
1452 | } | ||||||||
1453 | |||||||||
1454 | static Instruction *cloneInstructionInExitBlock( | ||||||||
1455 | Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI, | ||||||||
1456 | const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU) { | ||||||||
1457 | Instruction *New; | ||||||||
1458 | if (auto *CI = dyn_cast<CallInst>(&I)) { | ||||||||
1459 | const auto &BlockColors = SafetyInfo->getBlockColors(); | ||||||||
1460 | |||||||||
1461 | // Sinking call-sites need to be handled differently from other | ||||||||
1462 | // instructions. The cloned call-site needs a funclet bundle operand | ||||||||
1463 | // appropriate for its location in the CFG. | ||||||||
1464 | SmallVector<OperandBundleDef, 1> OpBundles; | ||||||||
1465 | for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles(); | ||||||||
1466 | BundleIdx != BundleEnd; ++BundleIdx) { | ||||||||
1467 | OperandBundleUse Bundle = CI->getOperandBundleAt(BundleIdx); | ||||||||
1468 | if (Bundle.getTagID() == LLVMContext::OB_funclet) | ||||||||
1469 | continue; | ||||||||
1470 | |||||||||
1471 | OpBundles.emplace_back(Bundle); | ||||||||
1472 | } | ||||||||
1473 | |||||||||
1474 | if (!BlockColors.empty()) { | ||||||||
1475 | const ColorVector &CV = BlockColors.find(&ExitBlock)->second; | ||||||||
1476 | assert(CV.size() == 1 && "non-unique color for exit block!")((CV.size() == 1 && "non-unique color for exit block!" ) ? static_cast<void> (0) : __assert_fail ("CV.size() == 1 && \"non-unique color for exit block!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1476, __PRETTY_FUNCTION__)); | ||||||||
1477 | BasicBlock *BBColor = CV.front(); | ||||||||
1478 | Instruction *EHPad = BBColor->getFirstNonPHI(); | ||||||||
1479 | if (EHPad->isEHPad()) | ||||||||
1480 | OpBundles.emplace_back("funclet", EHPad); | ||||||||
1481 | } | ||||||||
1482 | |||||||||
1483 | New = CallInst::Create(CI, OpBundles); | ||||||||
1484 | } else { | ||||||||
1485 | New = I.clone(); | ||||||||
1486 | } | ||||||||
1487 | |||||||||
1488 | ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New); | ||||||||
1489 | if (!I.getName().empty()) | ||||||||
1490 | New->setName(I.getName() + ".le"); | ||||||||
1491 | |||||||||
1492 | if (MSSAU && MSSAU->getMemorySSA()->getMemoryAccess(&I)) { | ||||||||
1493 | // Create a new MemoryAccess and let MemorySSA set its defining access. | ||||||||
1494 | MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( | ||||||||
1495 | New, nullptr, New->getParent(), MemorySSA::Beginning); | ||||||||
1496 | if (NewMemAcc) { | ||||||||
1497 | if (auto *MemDef = dyn_cast<MemoryDef>(NewMemAcc)) | ||||||||
1498 | MSSAU->insertDef(MemDef, /*RenameUses=*/true); | ||||||||
1499 | else { | ||||||||
1500 | auto *MemUse = cast<MemoryUse>(NewMemAcc); | ||||||||
1501 | MSSAU->insertUse(MemUse, /*RenameUses=*/true); | ||||||||
1502 | } | ||||||||
1503 | } | ||||||||
1504 | } | ||||||||
1505 | |||||||||
1506 | // Build LCSSA PHI nodes for any in-loop operands (if legal). Note that | ||||||||
1507 | // this is particularly cheap because we can rip off the PHI node that we're | ||||||||
1508 | // replacing for the number and blocks of the predecessors. | ||||||||
1509 | // OPT: If this shows up in a profile, we can instead finish sinking all | ||||||||
1510 | // invariant instructions, and then walk their operands to re-establish | ||||||||
1511 | // LCSSA. That will eliminate creating PHI nodes just to nuke them when | ||||||||
1512 | // sinking bottom-up. | ||||||||
1513 | for (Use &Op : New->operands()) | ||||||||
1514 | if (LI->wouldBeOutOfLoopUseRequiringLCSSA(Op.get(), PN.getParent())) { | ||||||||
1515 | auto *OInst = cast<Instruction>(Op.get()); | ||||||||
1516 | PHINode *OpPN = | ||||||||
1517 | PHINode::Create(OInst->getType(), PN.getNumIncomingValues(), | ||||||||
1518 | OInst->getName() + ".lcssa", &ExitBlock.front()); | ||||||||
1519 | for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) | ||||||||
1520 | OpPN->addIncoming(OInst, PN.getIncomingBlock(i)); | ||||||||
1521 | Op = OpPN; | ||||||||
1522 | } | ||||||||
1523 | return New; | ||||||||
1524 | } | ||||||||
1525 | |||||||||
1526 | static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo, | ||||||||
1527 | AliasSetTracker *AST, MemorySSAUpdater *MSSAU) { | ||||||||
1528 | if (AST) | ||||||||
1529 | AST->deleteValue(&I); | ||||||||
1530 | if (MSSAU) | ||||||||
1531 | MSSAU->removeMemoryAccess(&I); | ||||||||
1532 | SafetyInfo.removeInstruction(&I); | ||||||||
1533 | I.eraseFromParent(); | ||||||||
1534 | } | ||||||||
1535 | |||||||||
1536 | static void moveInstructionBefore(Instruction &I, Instruction &Dest, | ||||||||
1537 | ICFLoopSafetyInfo &SafetyInfo, | ||||||||
1538 | MemorySSAUpdater *MSSAU, | ||||||||
1539 | ScalarEvolution *SE) { | ||||||||
1540 | SafetyInfo.removeInstruction(&I); | ||||||||
1541 | SafetyInfo.insertInstructionTo(&I, Dest.getParent()); | ||||||||
1542 | I.moveBefore(&Dest); | ||||||||
1543 | if (MSSAU) | ||||||||
1544 | if (MemoryUseOrDef *OldMemAcc = cast_or_null<MemoryUseOrDef>( | ||||||||
1545 | MSSAU->getMemorySSA()->getMemoryAccess(&I))) | ||||||||
1546 | MSSAU->moveToPlace(OldMemAcc, Dest.getParent(), | ||||||||
1547 | MemorySSA::BeforeTerminator); | ||||||||
1548 | if (SE) | ||||||||
1549 | SE->forgetValue(&I); | ||||||||
1550 | } | ||||||||
1551 | |||||||||
1552 | static Instruction *sinkThroughTriviallyReplaceablePHI( | ||||||||
1553 | PHINode *TPN, Instruction *I, LoopInfo *LI, | ||||||||
1554 | SmallDenseMap<BasicBlock *, Instruction *, 32> &SunkCopies, | ||||||||
1555 | const LoopSafetyInfo *SafetyInfo, const Loop *CurLoop, | ||||||||
1556 | MemorySSAUpdater *MSSAU) { | ||||||||
1557 | assert(isTriviallyReplaceablePHI(*TPN, *I) &&((isTriviallyReplaceablePHI(*TPN, *I) && "Expect only trivially replaceable PHI" ) ? static_cast<void> (0) : __assert_fail ("isTriviallyReplaceablePHI(*TPN, *I) && \"Expect only trivially replaceable PHI\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1558, __PRETTY_FUNCTION__)) | ||||||||
1558 | "Expect only trivially replaceable PHI")((isTriviallyReplaceablePHI(*TPN, *I) && "Expect only trivially replaceable PHI" ) ? static_cast<void> (0) : __assert_fail ("isTriviallyReplaceablePHI(*TPN, *I) && \"Expect only trivially replaceable PHI\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1558, __PRETTY_FUNCTION__)); | ||||||||
1559 | BasicBlock *ExitBlock = TPN->getParent(); | ||||||||
1560 | Instruction *New; | ||||||||
1561 | auto It = SunkCopies.find(ExitBlock); | ||||||||
1562 | if (It != SunkCopies.end()) | ||||||||
1563 | New = It->second; | ||||||||
1564 | else | ||||||||
1565 | New = SunkCopies[ExitBlock] = cloneInstructionInExitBlock( | ||||||||
1566 | *I, *ExitBlock, *TPN, LI, SafetyInfo, MSSAU); | ||||||||
1567 | return New; | ||||||||
1568 | } | ||||||||
1569 | |||||||||
1570 | static bool canSplitPredecessors(PHINode *PN, LoopSafetyInfo *SafetyInfo) { | ||||||||
1571 | BasicBlock *BB = PN->getParent(); | ||||||||
1572 | if (!BB->canSplitPredecessors()) | ||||||||
1573 | return false; | ||||||||
1574 | // It's not impossible to split EHPad blocks, but if BlockColors already exist | ||||||||
1575 | // it require updating BlockColors for all offspring blocks accordingly. By | ||||||||
1576 | // skipping such corner case, we can make updating BlockColors after splitting | ||||||||
1577 | // predecessor fairly simple. | ||||||||
1578 | if (!SafetyInfo->getBlockColors().empty() && BB->getFirstNonPHI()->isEHPad()) | ||||||||
1579 | return false; | ||||||||
1580 | for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { | ||||||||
1581 | BasicBlock *BBPred = *PI; | ||||||||
1582 | if (isa<IndirectBrInst>(BBPred->getTerminator()) || | ||||||||
1583 | isa<CallBrInst>(BBPred->getTerminator())) | ||||||||
1584 | return false; | ||||||||
1585 | } | ||||||||
1586 | return true; | ||||||||
1587 | } | ||||||||
1588 | |||||||||
1589 | static void splitPredecessorsOfLoopExit(PHINode *PN, DominatorTree *DT, | ||||||||
1590 | LoopInfo *LI, const Loop *CurLoop, | ||||||||
1591 | LoopSafetyInfo *SafetyInfo, | ||||||||
1592 | MemorySSAUpdater *MSSAU) { | ||||||||
1593 | #ifndef NDEBUG | ||||||||
1594 | SmallVector<BasicBlock *, 32> ExitBlocks; | ||||||||
1595 | CurLoop->getUniqueExitBlocks(ExitBlocks); | ||||||||
1596 | SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(), | ||||||||
1597 | ExitBlocks.end()); | ||||||||
1598 | #endif | ||||||||
1599 | BasicBlock *ExitBB = PN->getParent(); | ||||||||
1600 | assert(ExitBlockSet.count(ExitBB) && "Expect the PHI is in an exit block.")((ExitBlockSet.count(ExitBB) && "Expect the PHI is in an exit block." ) ? static_cast<void> (0) : __assert_fail ("ExitBlockSet.count(ExitBB) && \"Expect the PHI is in an exit block.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1600, __PRETTY_FUNCTION__)); | ||||||||
1601 | |||||||||
1602 | // Split predecessors of the loop exit to make instructions in the loop are | ||||||||
1603 | // exposed to exit blocks through trivially replaceable PHIs while keeping the | ||||||||
1604 | // loop in the canonical form where each predecessor of each exit block should | ||||||||
1605 | // be contained within the loop. For example, this will convert the loop below | ||||||||
1606 | // from | ||||||||
1607 | // | ||||||||
1608 | // LB1: | ||||||||
1609 | // %v1 = | ||||||||
1610 | // br %LE, %LB2 | ||||||||
1611 | // LB2: | ||||||||
1612 | // %v2 = | ||||||||
1613 | // br %LE, %LB1 | ||||||||
1614 | // LE: | ||||||||
1615 | // %p = phi [%v1, %LB1], [%v2, %LB2] <-- non-trivially replaceable | ||||||||
1616 | // | ||||||||
1617 | // to | ||||||||
1618 | // | ||||||||
1619 | // LB1: | ||||||||
1620 | // %v1 = | ||||||||
1621 | // br %LE.split, %LB2 | ||||||||
1622 | // LB2: | ||||||||
1623 | // %v2 = | ||||||||
1624 | // br %LE.split2, %LB1 | ||||||||
1625 | // LE.split: | ||||||||
1626 | // %p1 = phi [%v1, %LB1] <-- trivially replaceable | ||||||||
1627 | // br %LE | ||||||||
1628 | // LE.split2: | ||||||||
1629 | // %p2 = phi [%v2, %LB2] <-- trivially replaceable | ||||||||
1630 | // br %LE | ||||||||
1631 | // LE: | ||||||||
1632 | // %p = phi [%p1, %LE.split], [%p2, %LE.split2] | ||||||||
1633 | // | ||||||||
1634 | const auto &BlockColors = SafetyInfo->getBlockColors(); | ||||||||
1635 | SmallSetVector<BasicBlock *, 8> PredBBs(pred_begin(ExitBB), pred_end(ExitBB)); | ||||||||
1636 | while (!PredBBs.empty()) { | ||||||||
1637 | BasicBlock *PredBB = *PredBBs.begin(); | ||||||||
1638 | assert(CurLoop->contains(PredBB) &&((CurLoop->contains(PredBB) && "Expect all predecessors are in the loop" ) ? static_cast<void> (0) : __assert_fail ("CurLoop->contains(PredBB) && \"Expect all predecessors are in the loop\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1639, __PRETTY_FUNCTION__)) | ||||||||
1639 | "Expect all predecessors are in the loop")((CurLoop->contains(PredBB) && "Expect all predecessors are in the loop" ) ? static_cast<void> (0) : __assert_fail ("CurLoop->contains(PredBB) && \"Expect all predecessors are in the loop\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1639, __PRETTY_FUNCTION__)); | ||||||||
1640 | if (PN->getBasicBlockIndex(PredBB) >= 0) { | ||||||||
1641 | BasicBlock *NewPred = SplitBlockPredecessors( | ||||||||
1642 | ExitBB, PredBB, ".split.loop.exit", DT, LI, MSSAU, true); | ||||||||
1643 | // Since we do not allow splitting EH-block with BlockColors in | ||||||||
1644 | // canSplitPredecessors(), we can simply assign predecessor's color to | ||||||||
1645 | // the new block. | ||||||||
1646 | if (!BlockColors.empty()) | ||||||||
1647 | // Grab a reference to the ColorVector to be inserted before getting the | ||||||||
1648 | // reference to the vector we are copying because inserting the new | ||||||||
1649 | // element in BlockColors might cause the map to be reallocated. | ||||||||
1650 | SafetyInfo->copyColors(NewPred, PredBB); | ||||||||
1651 | } | ||||||||
1652 | PredBBs.remove(PredBB); | ||||||||
1653 | } | ||||||||
1654 | } | ||||||||
1655 | |||||||||
1656 | /// When an instruction is found to only be used outside of the loop, this | ||||||||
1657 | /// function moves it to the exit blocks and patches up SSA form as needed. | ||||||||
1658 | /// This method is guaranteed to remove the original instruction from its | ||||||||
1659 | /// position, and may either delete it or move it to outside of the loop. | ||||||||
1660 | /// | ||||||||
1661 | static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT, | ||||||||
1662 | BlockFrequencyInfo *BFI, const Loop *CurLoop, | ||||||||
1663 | ICFLoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU, | ||||||||
1664 | OptimizationRemarkEmitter *ORE) { | ||||||||
1665 | bool Changed = false; | ||||||||
1666 | LLVM_DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM sinking instruction: " << I << "\n"; } } while (false); | ||||||||
1667 | |||||||||
1668 | // Iterate over users to be ready for actual sinking. Replace users via | ||||||||
1669 | // unreachable blocks with undef and make all user PHIs trivially replaceable. | ||||||||
1670 | SmallPtrSet<Instruction *, 8> VisitedUsers; | ||||||||
1671 | for (Value::user_iterator UI = I.user_begin(), UE = I.user_end(); UI != UE;) { | ||||||||
1672 | auto *User = cast<Instruction>(*UI); | ||||||||
1673 | Use &U = UI.getUse(); | ||||||||
1674 | ++UI; | ||||||||
1675 | |||||||||
1676 | if (VisitedUsers.count(User) || CurLoop->contains(User)) | ||||||||
1677 | continue; | ||||||||
1678 | |||||||||
1679 | if (!DT->isReachableFromEntry(User->getParent())) { | ||||||||
1680 | U = UndefValue::get(I.getType()); | ||||||||
1681 | Changed = true; | ||||||||
1682 | continue; | ||||||||
1683 | } | ||||||||
1684 | |||||||||
1685 | // The user must be a PHI node. | ||||||||
1686 | PHINode *PN = cast<PHINode>(User); | ||||||||
1687 | |||||||||
1688 | // Surprisingly, instructions can be used outside of loops without any | ||||||||
1689 | // exits. This can only happen in PHI nodes if the incoming block is | ||||||||
1690 | // unreachable. | ||||||||
1691 | BasicBlock *BB = PN->getIncomingBlock(U); | ||||||||
1692 | if (!DT->isReachableFromEntry(BB)) { | ||||||||
1693 | U = UndefValue::get(I.getType()); | ||||||||
1694 | Changed = true; | ||||||||
1695 | continue; | ||||||||
1696 | } | ||||||||
1697 | |||||||||
1698 | VisitedUsers.insert(PN); | ||||||||
1699 | if (isTriviallyReplaceablePHI(*PN, I)) | ||||||||
1700 | continue; | ||||||||
1701 | |||||||||
1702 | if (!canSplitPredecessors(PN, SafetyInfo)) | ||||||||
1703 | return Changed; | ||||||||
1704 | |||||||||
1705 | // Split predecessors of the PHI so that we can make users trivially | ||||||||
1706 | // replaceable. | ||||||||
1707 | splitPredecessorsOfLoopExit(PN, DT, LI, CurLoop, SafetyInfo, MSSAU); | ||||||||
1708 | |||||||||
1709 | // Should rebuild the iterators, as they may be invalidated by | ||||||||
1710 | // splitPredecessorsOfLoopExit(). | ||||||||
1711 | UI = I.user_begin(); | ||||||||
1712 | UE = I.user_end(); | ||||||||
1713 | } | ||||||||
1714 | |||||||||
1715 | if (VisitedUsers.empty()) | ||||||||
1716 | return Changed; | ||||||||
1717 | |||||||||
1718 | ORE->emit([&]() { | ||||||||
1719 | return OptimizationRemark(DEBUG_TYPE"licm", "InstSunk", &I) | ||||||||
1720 | << "sinking " << ore::NV("Inst", &I); | ||||||||
1721 | }); | ||||||||
1722 | if (isa<LoadInst>(I)) | ||||||||
1723 | ++NumMovedLoads; | ||||||||
1724 | else if (isa<CallInst>(I)) | ||||||||
1725 | ++NumMovedCalls; | ||||||||
1726 | ++NumSunk; | ||||||||
1727 | |||||||||
1728 | #ifndef NDEBUG | ||||||||
1729 | SmallVector<BasicBlock *, 32> ExitBlocks; | ||||||||
1730 | CurLoop->getUniqueExitBlocks(ExitBlocks); | ||||||||
1731 | SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(), | ||||||||
1732 | ExitBlocks.end()); | ||||||||
1733 | #endif | ||||||||
1734 | |||||||||
1735 | // Clones of this instruction. Don't create more than one per exit block! | ||||||||
1736 | SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies; | ||||||||
1737 | |||||||||
1738 | // If this instruction is only used outside of the loop, then all users are | ||||||||
1739 | // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of | ||||||||
1740 | // the instruction. | ||||||||
1741 | // First check if I is worth sinking for all uses. Sink only when it is worth | ||||||||
1742 | // across all uses. | ||||||||
1743 | SmallSetVector<User*, 8> Users(I.user_begin(), I.user_end()); | ||||||||
1744 | SmallVector<PHINode *, 8> ExitPNs; | ||||||||
1745 | for (auto *UI : Users) { | ||||||||
1746 | auto *User = cast<Instruction>(UI); | ||||||||
1747 | |||||||||
1748 | if (CurLoop->contains(User)) | ||||||||
1749 | continue; | ||||||||
1750 | |||||||||
1751 | PHINode *PN = cast<PHINode>(User); | ||||||||
1752 | assert(ExitBlockSet.count(PN->getParent()) &&((ExitBlockSet.count(PN->getParent()) && "The LCSSA PHI is not in an exit block!" ) ? static_cast<void> (0) : __assert_fail ("ExitBlockSet.count(PN->getParent()) && \"The LCSSA PHI is not in an exit block!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1753, __PRETTY_FUNCTION__)) | ||||||||
1753 | "The LCSSA PHI is not in an exit block!")((ExitBlockSet.count(PN->getParent()) && "The LCSSA PHI is not in an exit block!" ) ? static_cast<void> (0) : __assert_fail ("ExitBlockSet.count(PN->getParent()) && \"The LCSSA PHI is not in an exit block!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1753, __PRETTY_FUNCTION__)); | ||||||||
1754 | if (!worthSinkOrHoistInst(I, PN->getParent(), ORE, BFI)) { | ||||||||
1755 | return Changed; | ||||||||
1756 | } | ||||||||
1757 | |||||||||
1758 | ExitPNs.push_back(PN); | ||||||||
1759 | } | ||||||||
1760 | |||||||||
1761 | for (auto *PN : ExitPNs) { | ||||||||
1762 | |||||||||
1763 | // The PHI must be trivially replaceable. | ||||||||
1764 | Instruction *New = sinkThroughTriviallyReplaceablePHI( | ||||||||
1765 | PN, &I, LI, SunkCopies, SafetyInfo, CurLoop, MSSAU); | ||||||||
1766 | PN->replaceAllUsesWith(New); | ||||||||
1767 | eraseInstruction(*PN, *SafetyInfo, nullptr, nullptr); | ||||||||
1768 | Changed = true; | ||||||||
1769 | } | ||||||||
1770 | return Changed; | ||||||||
1771 | } | ||||||||
1772 | |||||||||
1773 | /// When an instruction is found to only use loop invariant operands that | ||||||||
1774 | /// is safe to hoist, this instruction is called to do the dirty work. | ||||||||
1775 | /// | ||||||||
1776 | static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop, | ||||||||
1777 | BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo, | ||||||||
1778 | MemorySSAUpdater *MSSAU, ScalarEvolution *SE, | ||||||||
1779 | OptimizationRemarkEmitter *ORE) { | ||||||||
1780 | LLVM_DEBUG(dbgs() << "LICM hoisting to " << Dest->getNameOrAsOperand() << ": "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM hoisting to " << Dest ->getNameOrAsOperand() << ": " << I << "\n" ; } } while (false) | ||||||||
1781 | << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM hoisting to " << Dest ->getNameOrAsOperand() << ": " << I << "\n" ; } } while (false); | ||||||||
1782 | ORE->emit([&]() { | ||||||||
1783 | return OptimizationRemark(DEBUG_TYPE"licm", "Hoisted", &I) << "hoisting " | ||||||||
1784 | << ore::NV("Inst", &I); | ||||||||
1785 | }); | ||||||||
1786 | |||||||||
1787 | // Metadata can be dependent on conditions we are hoisting above. | ||||||||
1788 | // Conservatively strip all metadata on the instruction unless we were | ||||||||
1789 | // guaranteed to execute I if we entered the loop, in which case the metadata | ||||||||
1790 | // is valid in the loop preheader. | ||||||||
1791 | if (I.hasMetadataOtherThanDebugLoc() && | ||||||||
1792 | // The check on hasMetadataOtherThanDebugLoc is to prevent us from burning | ||||||||
1793 | // time in isGuaranteedToExecute if we don't actually have anything to | ||||||||
1794 | // drop. It is a compile time optimization, not required for correctness. | ||||||||
1795 | !SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop)) | ||||||||
1796 | I.dropUnknownNonDebugMetadata(); | ||||||||
1797 | |||||||||
1798 | if (isa<PHINode>(I)) | ||||||||
1799 | // Move the new node to the end of the phi list in the destination block. | ||||||||
1800 | moveInstructionBefore(I, *Dest->getFirstNonPHI(), *SafetyInfo, MSSAU, SE); | ||||||||
1801 | else | ||||||||
1802 | // Move the new node to the destination block, before its terminator. | ||||||||
1803 | moveInstructionBefore(I, *Dest->getTerminator(), *SafetyInfo, MSSAU, SE); | ||||||||
1804 | |||||||||
1805 | I.updateLocationAfterHoist(); | ||||||||
1806 | |||||||||
1807 | if (isa<LoadInst>(I)) | ||||||||
1808 | ++NumMovedLoads; | ||||||||
1809 | else if (isa<CallInst>(I)) | ||||||||
1810 | ++NumMovedCalls; | ||||||||
1811 | ++NumHoisted; | ||||||||
1812 | } | ||||||||
1813 | |||||||||
1814 | /// Only sink or hoist an instruction if it is not a trapping instruction, | ||||||||
1815 | /// or if the instruction is known not to trap when moved to the preheader. | ||||||||
1816 | /// or if it is a trapping instruction and is guaranteed to execute. | ||||||||
1817 | static bool isSafeToExecuteUnconditionally(Instruction &Inst, | ||||||||
1818 | const DominatorTree *DT, | ||||||||
1819 | const TargetLibraryInfo *TLI, | ||||||||
1820 | const Loop *CurLoop, | ||||||||
1821 | const LoopSafetyInfo *SafetyInfo, | ||||||||
1822 | OptimizationRemarkEmitter *ORE, | ||||||||
1823 | const Instruction *CtxI) { | ||||||||
1824 | if (isSafeToSpeculativelyExecute(&Inst, CtxI, DT, TLI)) | ||||||||
1825 | return true; | ||||||||
1826 | |||||||||
1827 | bool GuaranteedToExecute = | ||||||||
1828 | SafetyInfo->isGuaranteedToExecute(Inst, DT, CurLoop); | ||||||||
1829 | |||||||||
1830 | if (!GuaranteedToExecute) { | ||||||||
1831 | auto *LI = dyn_cast<LoadInst>(&Inst); | ||||||||
1832 | if (LI && CurLoop->isLoopInvariant(LI->getPointerOperand())) | ||||||||
1833 | ORE->emit([&]() { | ||||||||
1834 | return OptimizationRemarkMissed( | ||||||||
1835 | DEBUG_TYPE"licm", "LoadWithLoopInvariantAddressCondExecuted", LI) | ||||||||
1836 | << "failed to hoist load with loop-invariant address " | ||||||||
1837 | "because load is conditionally executed"; | ||||||||
1838 | }); | ||||||||
1839 | } | ||||||||
1840 | |||||||||
1841 | return GuaranteedToExecute; | ||||||||
1842 | } | ||||||||
1843 | |||||||||
1844 | namespace { | ||||||||
1845 | class LoopPromoter : public LoadAndStorePromoter { | ||||||||
1846 | Value *SomePtr; // Designated pointer to store to. | ||||||||
1847 | const SmallSetVector<Value *, 8> &PointerMustAliases; | ||||||||
1848 | SmallVectorImpl<BasicBlock *> &LoopExitBlocks; | ||||||||
1849 | SmallVectorImpl<Instruction *> &LoopInsertPts; | ||||||||
1850 | SmallVectorImpl<MemoryAccess *> &MSSAInsertPts; | ||||||||
1851 | PredIteratorCache &PredCache; | ||||||||
1852 | AliasSetTracker *AST; | ||||||||
1853 | MemorySSAUpdater *MSSAU; | ||||||||
1854 | LoopInfo &LI; | ||||||||
1855 | DebugLoc DL; | ||||||||
1856 | int Alignment; | ||||||||
1857 | bool UnorderedAtomic; | ||||||||
1858 | AAMDNodes AATags; | ||||||||
1859 | ICFLoopSafetyInfo &SafetyInfo; | ||||||||
1860 | |||||||||
1861 | // We're about to add a use of V in a loop exit block. Insert an LCSSA phi | ||||||||
1862 | // (if legal) if doing so would add an out-of-loop use to an instruction | ||||||||
1863 | // defined in-loop. | ||||||||
1864 | Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const { | ||||||||
1865 | if (!LI.wouldBeOutOfLoopUseRequiringLCSSA(V, BB)) | ||||||||
1866 | return V; | ||||||||
1867 | |||||||||
1868 | Instruction *I = cast<Instruction>(V); | ||||||||
1869 | // We need to create an LCSSA PHI node for the incoming value and | ||||||||
1870 | // store that. | ||||||||
1871 | PHINode *PN = PHINode::Create(I->getType(), PredCache.size(BB), | ||||||||
1872 | I->getName() + ".lcssa", &BB->front()); | ||||||||
1873 | for (BasicBlock *Pred : PredCache.get(BB)) | ||||||||
1874 | PN->addIncoming(I, Pred); | ||||||||
1875 | return PN; | ||||||||
1876 | } | ||||||||
1877 | |||||||||
1878 | public: | ||||||||
1879 | LoopPromoter(Value *SP, ArrayRef<const Instruction *> Insts, SSAUpdater &S, | ||||||||
1880 | const SmallSetVector<Value *, 8> &PMA, | ||||||||
1881 | SmallVectorImpl<BasicBlock *> &LEB, | ||||||||
1882 | SmallVectorImpl<Instruction *> &LIP, | ||||||||
1883 | SmallVectorImpl<MemoryAccess *> &MSSAIP, PredIteratorCache &PIC, | ||||||||
1884 | AliasSetTracker *ast, MemorySSAUpdater *MSSAU, LoopInfo &li, | ||||||||
1885 | DebugLoc dl, int alignment, bool UnorderedAtomic, | ||||||||
1886 | const AAMDNodes &AATags, ICFLoopSafetyInfo &SafetyInfo) | ||||||||
1887 | : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA), | ||||||||
1888 | LoopExitBlocks(LEB), LoopInsertPts(LIP), MSSAInsertPts(MSSAIP), | ||||||||
1889 | PredCache(PIC), AST(ast), MSSAU(MSSAU), LI(li), DL(std::move(dl)), | ||||||||
1890 | Alignment(alignment), UnorderedAtomic(UnorderedAtomic), AATags(AATags), | ||||||||
1891 | SafetyInfo(SafetyInfo) {} | ||||||||
1892 | |||||||||
1893 | bool isInstInList(Instruction *I, | ||||||||
1894 | const SmallVectorImpl<Instruction *> &) const override { | ||||||||
1895 | Value *Ptr; | ||||||||
1896 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | ||||||||
1897 | Ptr = LI->getOperand(0); | ||||||||
1898 | else | ||||||||
1899 | Ptr = cast<StoreInst>(I)->getPointerOperand(); | ||||||||
1900 | return PointerMustAliases.count(Ptr); | ||||||||
1901 | } | ||||||||
1902 | |||||||||
1903 | void doExtraRewritesBeforeFinalDeletion() override { | ||||||||
1904 | // Insert stores after in the loop exit blocks. Each exit block gets a | ||||||||
1905 | // store of the live-out values that feed them. Since we've already told | ||||||||
1906 | // the SSA updater about the defs in the loop and the preheader | ||||||||
1907 | // definition, it is all set and we can start using it. | ||||||||
1908 | for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) { | ||||||||
1909 | BasicBlock *ExitBlock = LoopExitBlocks[i]; | ||||||||
1910 | Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock); | ||||||||
1911 | LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock); | ||||||||
1912 | Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock); | ||||||||
1913 | Instruction *InsertPos = LoopInsertPts[i]; | ||||||||
1914 | StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos); | ||||||||
1915 | if (UnorderedAtomic) | ||||||||
1916 | NewSI->setOrdering(AtomicOrdering::Unordered); | ||||||||
1917 | NewSI->setAlignment(Align(Alignment)); | ||||||||
1918 | NewSI->setDebugLoc(DL); | ||||||||
1919 | if (AATags) | ||||||||
1920 | NewSI->setAAMetadata(AATags); | ||||||||
1921 | |||||||||
1922 | if (MSSAU) { | ||||||||
1923 | MemoryAccess *MSSAInsertPoint = MSSAInsertPts[i]; | ||||||||
1924 | MemoryAccess *NewMemAcc; | ||||||||
1925 | if (!MSSAInsertPoint) { | ||||||||
1926 | NewMemAcc = MSSAU->createMemoryAccessInBB( | ||||||||
1927 | NewSI, nullptr, NewSI->getParent(), MemorySSA::Beginning); | ||||||||
1928 | } else { | ||||||||
1929 | NewMemAcc = | ||||||||
1930 | MSSAU->createMemoryAccessAfter(NewSI, nullptr, MSSAInsertPoint); | ||||||||
1931 | } | ||||||||
1932 | MSSAInsertPts[i] = NewMemAcc; | ||||||||
1933 | MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true); | ||||||||
1934 | // FIXME: true for safety, false may still be correct. | ||||||||
1935 | } | ||||||||
1936 | } | ||||||||
1937 | } | ||||||||
1938 | |||||||||
1939 | void replaceLoadWithValue(LoadInst *LI, Value *V) const override { | ||||||||
1940 | // Update alias analysis. | ||||||||
1941 | if (AST) | ||||||||
1942 | AST->copyValue(LI, V); | ||||||||
1943 | } | ||||||||
1944 | void instructionDeleted(Instruction *I) const override { | ||||||||
1945 | SafetyInfo.removeInstruction(I); | ||||||||
1946 | if (AST) | ||||||||
1947 | AST->deleteValue(I); | ||||||||
1948 | if (MSSAU) | ||||||||
1949 | MSSAU->removeMemoryAccess(I); | ||||||||
1950 | } | ||||||||
1951 | }; | ||||||||
1952 | |||||||||
1953 | |||||||||
1954 | /// Return true iff we can prove that a caller of this function can not inspect | ||||||||
1955 | /// the contents of the provided object in a well defined program. | ||||||||
1956 | bool isKnownNonEscaping(Value *Object, const TargetLibraryInfo *TLI) { | ||||||||
1957 | if (isa<AllocaInst>(Object)) | ||||||||
1958 | // Since the alloca goes out of scope, we know the caller can't retain a | ||||||||
1959 | // reference to it and be well defined. Thus, we don't need to check for | ||||||||
1960 | // capture. | ||||||||
1961 | return true; | ||||||||
1962 | |||||||||
1963 | // For all other objects we need to know that the caller can't possibly | ||||||||
1964 | // have gotten a reference to the object. There are two components of | ||||||||
1965 | // that: | ||||||||
1966 | // 1) Object can't be escaped by this function. This is what | ||||||||
1967 | // PointerMayBeCaptured checks. | ||||||||
1968 | // 2) Object can't have been captured at definition site. For this, we | ||||||||
1969 | // need to know the return value is noalias. At the moment, we use a | ||||||||
1970 | // weaker condition and handle only AllocLikeFunctions (which are | ||||||||
1971 | // known to be noalias). TODO | ||||||||
1972 | return isAllocLikeFn(Object, TLI) && | ||||||||
1973 | !PointerMayBeCaptured(Object, true, true); | ||||||||
1974 | } | ||||||||
1975 | |||||||||
1976 | } // namespace | ||||||||
1977 | |||||||||
1978 | /// Try to promote memory values to scalars by sinking stores out of the | ||||||||
1979 | /// loop and moving loads to before the loop. We do this by looping over | ||||||||
1980 | /// the stores in the loop, looking for stores to Must pointers which are | ||||||||
1981 | /// loop invariant. | ||||||||
1982 | /// | ||||||||
1983 | bool llvm::promoteLoopAccessesToScalars( | ||||||||
1984 | const SmallSetVector<Value *, 8> &PointerMustAliases, | ||||||||
1985 | SmallVectorImpl<BasicBlock *> &ExitBlocks, | ||||||||
1986 | SmallVectorImpl<Instruction *> &InsertPts, | ||||||||
1987 | SmallVectorImpl<MemoryAccess *> &MSSAInsertPts, PredIteratorCache &PIC, | ||||||||
1988 | LoopInfo *LI, DominatorTree *DT, const TargetLibraryInfo *TLI, | ||||||||
1989 | Loop *CurLoop, AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU, | ||||||||
1990 | ICFLoopSafetyInfo *SafetyInfo, OptimizationRemarkEmitter *ORE) { | ||||||||
1991 | // Verify inputs. | ||||||||
1992 | assert(LI != nullptr && DT != nullptr && CurLoop != nullptr &&((LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && "Unexpected Input to promoteLoopAccessesToScalars" ) ? static_cast<void> (0) : __assert_fail ("LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected Input to promoteLoopAccessesToScalars\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1994, __PRETTY_FUNCTION__)) | ||||||||
1993 | SafetyInfo != nullptr &&((LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && "Unexpected Input to promoteLoopAccessesToScalars" ) ? static_cast<void> (0) : __assert_fail ("LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected Input to promoteLoopAccessesToScalars\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1994, __PRETTY_FUNCTION__)) | ||||||||
1994 | "Unexpected Input to promoteLoopAccessesToScalars")((LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && "Unexpected Input to promoteLoopAccessesToScalars" ) ? static_cast<void> (0) : __assert_fail ("LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected Input to promoteLoopAccessesToScalars\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 1994, __PRETTY_FUNCTION__)); | ||||||||
1995 | |||||||||
1996 | Value *SomePtr = *PointerMustAliases.begin(); | ||||||||
1997 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); | ||||||||
1998 | |||||||||
1999 | // It is not safe to promote a load/store from the loop if the load/store is | ||||||||
2000 | // conditional. For example, turning: | ||||||||
2001 | // | ||||||||
2002 | // for () { if (c) *P += 1; } | ||||||||
2003 | // | ||||||||
2004 | // into: | ||||||||
2005 | // | ||||||||
2006 | // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; | ||||||||
2007 | // | ||||||||
2008 | // is not safe, because *P may only be valid to access if 'c' is true. | ||||||||
2009 | // | ||||||||
2010 | // The safety property divides into two parts: | ||||||||
2011 | // p1) The memory may not be dereferenceable on entry to the loop. In this | ||||||||
2012 | // case, we can't insert the required load in the preheader. | ||||||||
2013 | // p2) The memory model does not allow us to insert a store along any dynamic | ||||||||
2014 | // path which did not originally have one. | ||||||||
2015 | // | ||||||||
2016 | // If at least one store is guaranteed to execute, both properties are | ||||||||
2017 | // satisfied, and promotion is legal. | ||||||||
2018 | // | ||||||||
2019 | // This, however, is not a necessary condition. Even if no store/load is | ||||||||
2020 | // guaranteed to execute, we can still establish these properties. | ||||||||
2021 | // We can establish (p1) by proving that hoisting the load into the preheader | ||||||||
2022 | // is safe (i.e. proving dereferenceability on all paths through the loop). We | ||||||||
2023 | // can use any access within the alias set to prove dereferenceability, | ||||||||
2024 | // since they're all must alias. | ||||||||
2025 | // | ||||||||
2026 | // There are two ways establish (p2): | ||||||||
2027 | // a) Prove the location is thread-local. In this case the memory model | ||||||||
2028 | // requirement does not apply, and stores are safe to insert. | ||||||||
2029 | // b) Prove a store dominates every exit block. In this case, if an exit | ||||||||
2030 | // blocks is reached, the original dynamic path would have taken us through | ||||||||
2031 | // the store, so inserting a store into the exit block is safe. Note that this | ||||||||
2032 | // is different from the store being guaranteed to execute. For instance, | ||||||||
2033 | // if an exception is thrown on the first iteration of the loop, the original | ||||||||
2034 | // store is never executed, but the exit blocks are not executed either. | ||||||||
2035 | |||||||||
2036 | bool DereferenceableInPH = false; | ||||||||
2037 | bool SafeToInsertStore = false; | ||||||||
2038 | |||||||||
2039 | SmallVector<Instruction *, 64> LoopUses; | ||||||||
2040 | |||||||||
2041 | // We start with an alignment of one and try to find instructions that allow | ||||||||
2042 | // us to prove better alignment. | ||||||||
2043 | Align Alignment; | ||||||||
2044 | // Keep track of which types of access we see | ||||||||
2045 | bool SawUnorderedAtomic = false; | ||||||||
2046 | bool SawNotAtomic = false; | ||||||||
2047 | AAMDNodes AATags; | ||||||||
2048 | |||||||||
2049 | const DataLayout &MDL = Preheader->getModule()->getDataLayout(); | ||||||||
2050 | |||||||||
2051 | bool IsKnownThreadLocalObject = false; | ||||||||
2052 | if (SafetyInfo->anyBlockMayThrow()) { | ||||||||
2053 | // If a loop can throw, we have to insert a store along each unwind edge. | ||||||||
2054 | // That said, we can't actually make the unwind edge explicit. Therefore, | ||||||||
2055 | // we have to prove that the store is dead along the unwind edge. We do | ||||||||
2056 | // this by proving that the caller can't have a reference to the object | ||||||||
2057 | // after return and thus can't possibly load from the object. | ||||||||
2058 | Value *Object = getUnderlyingObject(SomePtr); | ||||||||
2059 | if (!isKnownNonEscaping(Object, TLI)) | ||||||||
2060 | return false; | ||||||||
2061 | // Subtlety: Alloca's aren't visible to callers, but *are* potentially | ||||||||
2062 | // visible to other threads if captured and used during their lifetimes. | ||||||||
2063 | IsKnownThreadLocalObject = !isa<AllocaInst>(Object); | ||||||||
2064 | } | ||||||||
2065 | |||||||||
2066 | // Check that all of the pointers in the alias set have the same type. We | ||||||||
2067 | // cannot (yet) promote a memory location that is loaded and stored in | ||||||||
2068 | // different sizes. While we are at it, collect alignment and AA info. | ||||||||
2069 | for (Value *ASIV : PointerMustAliases) { | ||||||||
2070 | // Check that all of the pointers in the alias set have the same type. We | ||||||||
2071 | // cannot (yet) promote a memory location that is loaded and stored in | ||||||||
2072 | // different sizes. | ||||||||
2073 | if (SomePtr->getType() != ASIV->getType()) | ||||||||
2074 | return false; | ||||||||
2075 | |||||||||
2076 | for (User *U : ASIV->users()) { | ||||||||
2077 | // Ignore instructions that are outside the loop. | ||||||||
2078 | Instruction *UI = dyn_cast<Instruction>(U); | ||||||||
2079 | if (!UI || !CurLoop->contains(UI)) | ||||||||
2080 | continue; | ||||||||
2081 | |||||||||
2082 | // If there is an non-load/store instruction in the loop, we can't promote | ||||||||
2083 | // it. | ||||||||
2084 | if (LoadInst *Load = dyn_cast<LoadInst>(UI)) { | ||||||||
2085 | if (!Load->isUnordered()) | ||||||||
2086 | return false; | ||||||||
2087 | |||||||||
2088 | SawUnorderedAtomic |= Load->isAtomic(); | ||||||||
2089 | SawNotAtomic |= !Load->isAtomic(); | ||||||||
2090 | |||||||||
2091 | Align InstAlignment = Load->getAlign(); | ||||||||
2092 | |||||||||
2093 | // Note that proving a load safe to speculate requires proving | ||||||||
2094 | // sufficient alignment at the target location. Proving it guaranteed | ||||||||
2095 | // to execute does as well. Thus we can increase our guaranteed | ||||||||
2096 | // alignment as well. | ||||||||
2097 | if (!DereferenceableInPH || (InstAlignment > Alignment)) | ||||||||
2098 | if (isSafeToExecuteUnconditionally(*Load, DT, TLI, CurLoop, | ||||||||
2099 | SafetyInfo, ORE, | ||||||||
2100 | Preheader->getTerminator())) { | ||||||||
2101 | DereferenceableInPH = true; | ||||||||
2102 | Alignment = std::max(Alignment, InstAlignment); | ||||||||
2103 | } | ||||||||
2104 | } else if (const StoreInst *Store = dyn_cast<StoreInst>(UI)) { | ||||||||
2105 | // Stores *of* the pointer are not interesting, only stores *to* the | ||||||||
2106 | // pointer. | ||||||||
2107 | if (UI->getOperand(1) != ASIV) | ||||||||
2108 | continue; | ||||||||
2109 | if (!Store->isUnordered()) | ||||||||
2110 | return false; | ||||||||
2111 | |||||||||
2112 | SawUnorderedAtomic |= Store->isAtomic(); | ||||||||
2113 | SawNotAtomic |= !Store->isAtomic(); | ||||||||
2114 | |||||||||
2115 | // If the store is guaranteed to execute, both properties are satisfied. | ||||||||
2116 | // We may want to check if a store is guaranteed to execute even if we | ||||||||
2117 | // already know that promotion is safe, since it may have higher | ||||||||
2118 | // alignment than any other guaranteed stores, in which case we can | ||||||||
2119 | // raise the alignment on the promoted store. | ||||||||
2120 | Align InstAlignment = Store->getAlign(); | ||||||||
2121 | |||||||||
2122 | if (!DereferenceableInPH || !SafeToInsertStore || | ||||||||
2123 | (InstAlignment > Alignment)) { | ||||||||
2124 | if (SafetyInfo->isGuaranteedToExecute(*UI, DT, CurLoop)) { | ||||||||
2125 | DereferenceableInPH = true; | ||||||||
2126 | SafeToInsertStore = true; | ||||||||
2127 | Alignment = std::max(Alignment, InstAlignment); | ||||||||
2128 | } | ||||||||
2129 | } | ||||||||
2130 | |||||||||
2131 | // If a store dominates all exit blocks, it is safe to sink. | ||||||||
2132 | // As explained above, if an exit block was executed, a dominating | ||||||||
2133 | // store must have been executed at least once, so we are not | ||||||||
2134 | // introducing stores on paths that did not have them. | ||||||||
2135 | // Note that this only looks at explicit exit blocks. If we ever | ||||||||
2136 | // start sinking stores into unwind edges (see above), this will break. | ||||||||
2137 | if (!SafeToInsertStore) | ||||||||
2138 | SafeToInsertStore = llvm::all_of(ExitBlocks, [&](BasicBlock *Exit) { | ||||||||
2139 | return DT->dominates(Store->getParent(), Exit); | ||||||||
2140 | }); | ||||||||
2141 | |||||||||
2142 | // If the store is not guaranteed to execute, we may still get | ||||||||
2143 | // deref info through it. | ||||||||
2144 | if (!DereferenceableInPH) { | ||||||||
2145 | DereferenceableInPH = isDereferenceableAndAlignedPointer( | ||||||||
2146 | Store->getPointerOperand(), Store->getValueOperand()->getType(), | ||||||||
2147 | Store->getAlign(), MDL, Preheader->getTerminator(), DT, TLI); | ||||||||
2148 | } | ||||||||
2149 | } else | ||||||||
2150 | return false; // Not a load or store. | ||||||||
2151 | |||||||||
2152 | // Merge the AA tags. | ||||||||
2153 | if (LoopUses.empty()) { | ||||||||
2154 | // On the first load/store, just take its AA tags. | ||||||||
2155 | UI->getAAMetadata(AATags); | ||||||||
2156 | } else if (AATags) { | ||||||||
2157 | UI->getAAMetadata(AATags, /* Merge = */ true); | ||||||||
2158 | } | ||||||||
2159 | |||||||||
2160 | LoopUses.push_back(UI); | ||||||||
2161 | } | ||||||||
2162 | } | ||||||||
2163 | |||||||||
2164 | // If we found both an unordered atomic instruction and a non-atomic memory | ||||||||
2165 | // access, bail. We can't blindly promote non-atomic to atomic since we | ||||||||
2166 | // might not be able to lower the result. We can't downgrade since that | ||||||||
2167 | // would violate memory model. Also, align 0 is an error for atomics. | ||||||||
2168 | if (SawUnorderedAtomic && SawNotAtomic) | ||||||||
2169 | return false; | ||||||||
2170 | |||||||||
2171 | // If we're inserting an atomic load in the preheader, we must be able to | ||||||||
2172 | // lower it. We're only guaranteed to be able to lower naturally aligned | ||||||||
2173 | // atomics. | ||||||||
2174 | auto *SomePtrElemType = SomePtr->getType()->getPointerElementType(); | ||||||||
2175 | if (SawUnorderedAtomic && | ||||||||
2176 | Alignment < MDL.getTypeStoreSize(SomePtrElemType)) | ||||||||
2177 | return false; | ||||||||
2178 | |||||||||
2179 | // If we couldn't prove we can hoist the load, bail. | ||||||||
2180 | if (!DereferenceableInPH) | ||||||||
2181 | return false; | ||||||||
2182 | |||||||||
2183 | // We know we can hoist the load, but don't have a guaranteed store. | ||||||||
2184 | // Check whether the location is thread-local. If it is, then we can insert | ||||||||
2185 | // stores along paths which originally didn't have them without violating the | ||||||||
2186 | // memory model. | ||||||||
2187 | if (!SafeToInsertStore) { | ||||||||
2188 | if (IsKnownThreadLocalObject) | ||||||||
2189 | SafeToInsertStore = true; | ||||||||
2190 | else { | ||||||||
2191 | Value *Object = getUnderlyingObject(SomePtr); | ||||||||
2192 | SafeToInsertStore = | ||||||||
2193 | (isAllocLikeFn(Object, TLI) || isa<AllocaInst>(Object)) && | ||||||||
2194 | !PointerMayBeCaptured(Object, true, true); | ||||||||
2195 | } | ||||||||
2196 | } | ||||||||
2197 | |||||||||
2198 | // If we've still failed to prove we can sink the store, give up. | ||||||||
2199 | if (!SafeToInsertStore) | ||||||||
2200 | return false; | ||||||||
2201 | |||||||||
2202 | // Otherwise, this is safe to promote, lets do it! | ||||||||
2203 | LLVM_DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " << *SomePtrdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM: Promoting value stored to in loop: " << *SomePtr << '\n'; } } while (false) | ||||||||
2204 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "LICM: Promoting value stored to in loop: " << *SomePtr << '\n'; } } while (false); | ||||||||
2205 | ORE->emit([&]() { | ||||||||
2206 | return OptimizationRemark(DEBUG_TYPE"licm", "PromoteLoopAccessesToScalar", | ||||||||
2207 | LoopUses[0]) | ||||||||
2208 | << "Moving accesses to memory location out of the loop"; | ||||||||
2209 | }); | ||||||||
2210 | ++NumPromoted; | ||||||||
2211 | |||||||||
2212 | // Look at all the loop uses, and try to merge their locations. | ||||||||
2213 | std::vector<const DILocation *> LoopUsesLocs; | ||||||||
2214 | for (auto U : LoopUses) | ||||||||
2215 | LoopUsesLocs.push_back(U->getDebugLoc().get()); | ||||||||
2216 | auto DL = DebugLoc(DILocation::getMergedLocations(LoopUsesLocs)); | ||||||||
2217 | |||||||||
2218 | // We use the SSAUpdater interface to insert phi nodes as required. | ||||||||
2219 | SmallVector<PHINode *, 16> NewPHIs; | ||||||||
2220 | SSAUpdater SSA(&NewPHIs); | ||||||||
2221 | LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks, | ||||||||
2222 | InsertPts, MSSAInsertPts, PIC, CurAST, MSSAU, *LI, DL, | ||||||||
2223 | Alignment.value(), SawUnorderedAtomic, AATags, | ||||||||
2224 | *SafetyInfo); | ||||||||
2225 | |||||||||
2226 | // Set up the preheader to have a definition of the value. It is the live-out | ||||||||
2227 | // value from the preheader that uses in the loop will use. | ||||||||
2228 | LoadInst *PreheaderLoad = new LoadInst( | ||||||||
2229 | SomePtr->getType()->getPointerElementType(), SomePtr, | ||||||||
2230 | SomePtr->getName() + ".promoted", Preheader->getTerminator()); | ||||||||
2231 | if (SawUnorderedAtomic) | ||||||||
2232 | PreheaderLoad->setOrdering(AtomicOrdering::Unordered); | ||||||||
2233 | PreheaderLoad->setAlignment(Alignment); | ||||||||
2234 | PreheaderLoad->setDebugLoc(DebugLoc()); | ||||||||
2235 | if (AATags) | ||||||||
2236 | PreheaderLoad->setAAMetadata(AATags); | ||||||||
2237 | SSA.AddAvailableValue(Preheader, PreheaderLoad); | ||||||||
2238 | |||||||||
2239 | if (MSSAU) { | ||||||||
2240 | MemoryAccess *PreheaderLoadMemoryAccess = MSSAU->createMemoryAccessInBB( | ||||||||
2241 | PreheaderLoad, nullptr, PreheaderLoad->getParent(), MemorySSA::End); | ||||||||
2242 | MemoryUse *NewMemUse = cast<MemoryUse>(PreheaderLoadMemoryAccess); | ||||||||
2243 | MSSAU->insertUse(NewMemUse, /*RenameUses=*/true); | ||||||||
2244 | } | ||||||||
2245 | |||||||||
2246 | if (MSSAU && VerifyMemorySSA) | ||||||||
2247 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||
2248 | // Rewrite all the loads in the loop and remember all the definitions from | ||||||||
2249 | // stores in the loop. | ||||||||
2250 | Promoter.run(LoopUses); | ||||||||
2251 | |||||||||
2252 | if (MSSAU && VerifyMemorySSA) | ||||||||
2253 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||
2254 | // If the SSAUpdater didn't use the load in the preheader, just zap it now. | ||||||||
2255 | if (PreheaderLoad->use_empty()) | ||||||||
2256 | eraseInstruction(*PreheaderLoad, *SafetyInfo, CurAST, MSSAU); | ||||||||
2257 | |||||||||
2258 | return true; | ||||||||
2259 | } | ||||||||
2260 | |||||||||
2261 | static void foreachMemoryAccess(MemorySSA *MSSA, Loop *L, | ||||||||
2262 | function_ref<void(Instruction *)> Fn) { | ||||||||
2263 | for (const BasicBlock *BB : L->blocks()) | ||||||||
2264 | if (const auto *Accesses = MSSA->getBlockAccesses(BB)) | ||||||||
2265 | for (const auto &Access : *Accesses) | ||||||||
2266 | if (const auto *MUD = dyn_cast<MemoryUseOrDef>(&Access)) | ||||||||
2267 | Fn(MUD->getMemoryInst()); | ||||||||
2268 | } | ||||||||
2269 | |||||||||
2270 | static SmallVector<SmallSetVector<Value *, 8>, 0> | ||||||||
2271 | collectPromotionCandidates(MemorySSA *MSSA, AliasAnalysis *AA, Loop *L, | ||||||||
2272 | SmallVectorImpl<Instruction *> &MaybePromotable) { | ||||||||
2273 | AliasSetTracker AST(*AA); | ||||||||
2274 | |||||||||
2275 | auto IsPotentiallyPromotable = [L](const Instruction *I) { | ||||||||
2276 | if (const auto *SI = dyn_cast<StoreInst>(I)) | ||||||||
2277 | return L->isLoopInvariant(SI->getPointerOperand()); | ||||||||
2278 | if (const auto *LI = dyn_cast<LoadInst>(I)) | ||||||||
2279 | return L->isLoopInvariant(LI->getPointerOperand()); | ||||||||
2280 | return false; | ||||||||
2281 | }; | ||||||||
2282 | |||||||||
2283 | // Populate AST with potentially promotable accesses and remove them from | ||||||||
2284 | // MaybePromotable, so they will not be checked again on the next iteration. | ||||||||
2285 | SmallPtrSet<Value *, 16> AttemptingPromotion; | ||||||||
2286 | llvm::erase_if(MaybePromotable, [&](Instruction *I) { | ||||||||
2287 | if (IsPotentiallyPromotable(I)) { | ||||||||
2288 | AttemptingPromotion.insert(I); | ||||||||
2289 | AST.add(I); | ||||||||
2290 | return true; | ||||||||
2291 | } | ||||||||
2292 | return false; | ||||||||
2293 | }); | ||||||||
2294 | |||||||||
2295 | // We're only interested in must-alias sets that contain a mod. | ||||||||
2296 | SmallVector<const AliasSet *, 8> Sets; | ||||||||
2297 | for (AliasSet &AS : AST) | ||||||||
2298 | if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias()) | ||||||||
2299 | Sets.push_back(&AS); | ||||||||
2300 | |||||||||
2301 | if (Sets.empty()) | ||||||||
2302 | return {}; // Nothing to promote... | ||||||||
2303 | |||||||||
2304 | // Discard any sets for which there is an aliasing non-promotable access. | ||||||||
2305 | foreachMemoryAccess(MSSA, L, [&](Instruction *I) { | ||||||||
2306 | if (AttemptingPromotion.contains(I)) | ||||||||
2307 | return; | ||||||||
2308 | |||||||||
2309 | llvm::erase_if(Sets, [&](const AliasSet *AS) { | ||||||||
2310 | return AS->aliasesUnknownInst(I, *AA); | ||||||||
2311 | }); | ||||||||
2312 | }); | ||||||||
2313 | |||||||||
2314 | SmallVector<SmallSetVector<Value *, 8>, 0> Result; | ||||||||
2315 | for (const AliasSet *Set : Sets) { | ||||||||
2316 | SmallSetVector<Value *, 8> PointerMustAliases; | ||||||||
2317 | for (const auto &ASI : *Set) | ||||||||
2318 | PointerMustAliases.insert(ASI.getValue()); | ||||||||
2319 | Result.push_back(std::move(PointerMustAliases)); | ||||||||
2320 | } | ||||||||
2321 | |||||||||
2322 | return Result; | ||||||||
2323 | } | ||||||||
2324 | |||||||||
2325 | /// Returns an owning pointer to an alias set which incorporates aliasing info | ||||||||
2326 | /// from L and all subloops of L. | ||||||||
2327 | std::unique_ptr<AliasSetTracker> | ||||||||
2328 | LoopInvariantCodeMotion::collectAliasInfoForLoop(Loop *L, LoopInfo *LI, | ||||||||
2329 | AAResults *AA) { | ||||||||
2330 | auto CurAST = std::make_unique<AliasSetTracker>(*AA); | ||||||||
2331 | |||||||||
2332 | // Add everything from all the sub loops. | ||||||||
2333 | for (Loop *InnerL : L->getSubLoops()) | ||||||||
2334 | for (BasicBlock *BB : InnerL->blocks()) | ||||||||
2335 | CurAST->add(*BB); | ||||||||
2336 | |||||||||
2337 | // And merge in this loop (without anything from inner loops). | ||||||||
2338 | for (BasicBlock *BB : L->blocks()) | ||||||||
2339 | if (LI->getLoopFor(BB) == L) | ||||||||
2340 | CurAST->add(*BB); | ||||||||
2341 | |||||||||
2342 | return CurAST; | ||||||||
2343 | } | ||||||||
2344 | |||||||||
2345 | static bool pointerInvalidatedByLoop(MemoryLocation MemLoc, | ||||||||
2346 | AliasSetTracker *CurAST, Loop *CurLoop, | ||||||||
2347 | AAResults *AA) { | ||||||||
2348 | // First check to see if any of the basic blocks in CurLoop invalidate *V. | ||||||||
2349 | bool isInvalidatedAccordingToAST = CurAST->getAliasSetFor(MemLoc).isMod(); | ||||||||
2350 | |||||||||
2351 | if (!isInvalidatedAccordingToAST || !LICMN2Theshold) | ||||||||
2352 | return isInvalidatedAccordingToAST; | ||||||||
2353 | |||||||||
2354 | // Check with a diagnostic analysis if we can refine the information above. | ||||||||
2355 | // This is to identify the limitations of using the AST. | ||||||||
2356 | // The alias set mechanism used by LICM has a major weakness in that it | ||||||||
2357 | // combines all things which may alias into a single set *before* asking | ||||||||
2358 | // modref questions. As a result, a single readonly call within a loop will | ||||||||
2359 | // collapse all loads and stores into a single alias set and report | ||||||||
2360 | // invalidation if the loop contains any store. For example, readonly calls | ||||||||
2361 | // with deopt states have this form and create a general alias set with all | ||||||||
2362 | // loads and stores. In order to get any LICM in loops containing possible | ||||||||
2363 | // deopt states we need a more precise invalidation of checking the mod ref | ||||||||
2364 | // info of each instruction within the loop and LI. This has a complexity of | ||||||||
2365 | // O(N^2), so currently, it is used only as a diagnostic tool since the | ||||||||
2366 | // default value of LICMN2Threshold is zero. | ||||||||
2367 | |||||||||
2368 | // Don't look at nested loops. | ||||||||
2369 | if (CurLoop->begin() != CurLoop->end()) | ||||||||
2370 | return true; | ||||||||
2371 | |||||||||
2372 | int N = 0; | ||||||||
2373 | for (BasicBlock *BB : CurLoop->getBlocks()) | ||||||||
2374 | for (Instruction &I : *BB) { | ||||||||
2375 | if (N >= LICMN2Theshold) { | ||||||||
2376 | LLVM_DEBUG(dbgs() << "Alasing N2 threshold exhausted for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "Alasing N2 threshold exhausted for " << *(MemLoc.Ptr) << "\n"; } } while (false) | ||||||||
2377 | << *(MemLoc.Ptr) << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "Alasing N2 threshold exhausted for " << *(MemLoc.Ptr) << "\n"; } } while (false); | ||||||||
2378 | return true; | ||||||||
2379 | } | ||||||||
2380 | N++; | ||||||||
2381 | auto Res = AA->getModRefInfo(&I, MemLoc); | ||||||||
2382 | if (isModSet(Res)) { | ||||||||
2383 | LLVM_DEBUG(dbgs() << "Aliasing failed on " << I << " for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "Aliasing failed on " << I << " for " << *(MemLoc.Ptr) << "\n"; } } while (false ) | ||||||||
2384 | << *(MemLoc.Ptr) << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "Aliasing failed on " << I << " for " << *(MemLoc.Ptr) << "\n"; } } while (false ); | ||||||||
2385 | return true; | ||||||||
2386 | } | ||||||||
2387 | } | ||||||||
2388 | LLVM_DEBUG(dbgs() << "Aliasing okay for " << *(MemLoc.Ptr) << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("licm")) { dbgs() << "Aliasing okay for " << *(MemLoc .Ptr) << "\n"; } } while (false); | ||||||||
2389 | return false; | ||||||||
2390 | } | ||||||||
2391 | |||||||||
2392 | bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU, | ||||||||
2393 | Loop *CurLoop, Instruction &I, | ||||||||
2394 | SinkAndHoistLICMFlags &Flags) { | ||||||||
2395 | // For hoisting, use the walker to determine safety | ||||||||
2396 | if (!Flags.getIsSink()) { | ||||||||
2397 | MemoryAccess *Source; | ||||||||
2398 | // See declaration of SetLicmMssaOptCap for usage details. | ||||||||
2399 | if (Flags.tooManyClobberingCalls()) | ||||||||
2400 | Source = MU->getDefiningAccess(); | ||||||||
2401 | else { | ||||||||
2402 | Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(MU); | ||||||||
2403 | Flags.incrementClobberingCalls(); | ||||||||
2404 | } | ||||||||
2405 | return !MSSA->isLiveOnEntryDef(Source) && | ||||||||
2406 | CurLoop->contains(Source->getBlock()); | ||||||||
2407 | } | ||||||||
2408 | |||||||||
2409 | // For sinking, we'd need to check all Defs below this use. The getClobbering | ||||||||
2410 | // call will look on the backedge of the loop, but will check aliasing with | ||||||||
2411 | // the instructions on the previous iteration. | ||||||||
2412 | // For example: | ||||||||
2413 | // for (i ... ) | ||||||||
2414 | // load a[i] ( Use (LoE) | ||||||||
2415 | // store a[i] ( 1 = Def (2), with 2 = Phi for the loop. | ||||||||
2416 | // i++; | ||||||||
2417 | // The load sees no clobbering inside the loop, as the backedge alias check | ||||||||
2418 | // does phi translation, and will check aliasing against store a[i-1]. | ||||||||
2419 | // However sinking the load outside the loop, below the store is incorrect. | ||||||||
2420 | |||||||||
2421 | // For now, only sink if there are no Defs in the loop, and the existing ones | ||||||||
2422 | // precede the use and are in the same block. | ||||||||
2423 | // FIXME: Increase precision: Safe to sink if Use post dominates the Def; | ||||||||
2424 | // needs PostDominatorTreeAnalysis. | ||||||||
2425 | // FIXME: More precise: no Defs that alias this Use. | ||||||||
2426 | if (Flags.tooManyMemoryAccesses()) | ||||||||
2427 | return true; | ||||||||
2428 | for (auto *BB : CurLoop->getBlocks()) | ||||||||
2429 | if (pointerInvalidatedByBlockWithMSSA(*BB, *MSSA, *MU)) | ||||||||
2430 | return true; | ||||||||
2431 | // When sinking, the source block may not be part of the loop so check it. | ||||||||
2432 | if (!CurLoop->contains(&I)) | ||||||||
2433 | return pointerInvalidatedByBlockWithMSSA(*I.getParent(), *MSSA, *MU); | ||||||||
2434 | |||||||||
2435 | return false; | ||||||||
2436 | } | ||||||||
2437 | |||||||||
2438 | bool pointerInvalidatedByBlockWithMSSA(BasicBlock &BB, MemorySSA &MSSA, | ||||||||
2439 | MemoryUse &MU) { | ||||||||
2440 | if (const auto *Accesses = MSSA.getBlockDefs(&BB)) | ||||||||
2441 | for (const auto &MA : *Accesses) | ||||||||
2442 | if (const auto *MD = dyn_cast<MemoryDef>(&MA)) | ||||||||
2443 | if (MU.getBlock() != MD->getBlock() || !MSSA.locallyDominates(MD, &MU)) | ||||||||
2444 | return true; | ||||||||
2445 | return false; | ||||||||
2446 | } | ||||||||
2447 | |||||||||
2448 | /// Little predicate that returns true if the specified basic block is in | ||||||||
2449 | /// a subloop of the current one, not the current one itself. | ||||||||
2450 | /// | ||||||||
2451 | static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) { | ||||||||
2452 | assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop")((CurLoop->contains(BB) && "Only valid if BB is IN the loop" ) ? static_cast<void> (0) : __assert_fail ("CurLoop->contains(BB) && \"Only valid if BB is IN the loop\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/LICM.cpp" , 2452, __PRETTY_FUNCTION__)); | ||||||||
2453 | return LI->getLoopFor(BB) != CurLoop; | ||||||||
2454 | } |
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/None.h" |
21 | #include "llvm/ADT/STLExtras.h" |
22 | #include "llvm/ADT/SmallVector.h" |
23 | #include "llvm/ADT/StringRef.h" |
24 | #include "llvm/ADT/Twine.h" |
25 | #include "llvm/ADT/iterator.h" |
26 | #include "llvm/ADT/iterator_range.h" |
27 | #include "llvm/IR/Attributes.h" |
28 | #include "llvm/IR/BasicBlock.h" |
29 | #include "llvm/IR/CallingConv.h" |
30 | #include "llvm/IR/CFG.h" |
31 | #include "llvm/IR/Constant.h" |
32 | #include "llvm/IR/DerivedTypes.h" |
33 | #include "llvm/IR/Function.h" |
34 | #include "llvm/IR/InstrTypes.h" |
35 | #include "llvm/IR/Instruction.h" |
36 | #include "llvm/IR/OperandTraits.h" |
37 | #include "llvm/IR/Type.h" |
38 | #include "llvm/IR/Use.h" |
39 | #include "llvm/IR/User.h" |
40 | #include "llvm/IR/Value.h" |
41 | #include "llvm/Support/AtomicOrdering.h" |
42 | #include "llvm/Support/Casting.h" |
43 | #include "llvm/Support/ErrorHandling.h" |
44 | #include <cassert> |
45 | #include <cstddef> |
46 | #include <cstdint> |
47 | #include <iterator> |
48 | |
49 | namespace llvm { |
50 | |
51 | class APInt; |
52 | class ConstantInt; |
53 | class DataLayout; |
54 | class LLVMContext; |
55 | |
56 | //===----------------------------------------------------------------------===// |
57 | // AllocaInst Class |
58 | //===----------------------------------------------------------------------===// |
59 | |
60 | /// an instruction to allocate memory on the stack |
61 | class AllocaInst : public UnaryInstruction { |
62 | Type *AllocatedType; |
63 | |
64 | using AlignmentField = AlignmentBitfieldElementT<0>; |
65 | using UsedWithInAllocaField = BoolBitfieldElementT<AlignmentField::NextBit>; |
66 | using SwiftErrorField = BoolBitfieldElementT<UsedWithInAllocaField::NextBit>; |
67 | static_assert(Bitfield::areContiguous<AlignmentField, UsedWithInAllocaField, |
68 | SwiftErrorField>(), |
69 | "Bitfields must be contiguous"); |
70 | |
71 | protected: |
72 | // Note: Instruction needs to be a friend here to call cloneImpl. |
73 | friend class Instruction; |
74 | |
75 | AllocaInst *cloneImpl() const; |
76 | |
77 | public: |
78 | explicit AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, |
79 | const Twine &Name, Instruction *InsertBefore); |
80 | AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, |
81 | const Twine &Name, BasicBlock *InsertAtEnd); |
82 | |
83 | AllocaInst(Type *Ty, unsigned AddrSpace, const Twine &Name, |
84 | Instruction *InsertBefore); |
85 | AllocaInst(Type *Ty, unsigned AddrSpace, |
86 | const Twine &Name, BasicBlock *InsertAtEnd); |
87 | |
88 | AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align, |
89 | const Twine &Name = "", Instruction *InsertBefore = nullptr); |
90 | AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align, |
91 | const Twine &Name, BasicBlock *InsertAtEnd); |
92 | |
93 | /// Return true if there is an allocation size parameter to the allocation |
94 | /// instruction that is not 1. |
95 | bool isArrayAllocation() const; |
96 | |
97 | /// Get the number of elements allocated. For a simple allocation of a single |
98 | /// element, this will return a constant 1 value. |
99 | const Value *getArraySize() const { return getOperand(0); } |
100 | Value *getArraySize() { return getOperand(0); } |
101 | |
102 | /// Overload to return most specific pointer type. |
103 | PointerType *getType() const { |
104 | return cast<PointerType>(Instruction::getType()); |
105 | } |
106 | |
107 | /// Get allocation size in bits. Returns None if size can't be determined, |
108 | /// e.g. in case of a VLA. |
109 | Optional<TypeSize> getAllocationSizeInBits(const DataLayout &DL) const; |
110 | |
111 | /// Return the type that is being allocated by the instruction. |
112 | Type *getAllocatedType() const { return AllocatedType; } |
113 | /// for use only in special circumstances that need to generically |
114 | /// transform a whole instruction (eg: IR linking and vectorization). |
115 | void setAllocatedType(Type *Ty) { AllocatedType = Ty; } |
116 | |
117 | /// Return the alignment of the memory that is being allocated by the |
118 | /// instruction. |
119 | Align getAlign() const { |
120 | return Align(1ULL << getSubclassData<AlignmentField>()); |
121 | } |
122 | |
123 | void setAlignment(Align Align) { |
124 | setSubclassData<AlignmentField>(Log2(Align)); |
125 | } |
126 | |
127 | // FIXME: Remove this one transition to Align is over. |
128 | unsigned getAlignment() const { return getAlign().value(); } |
129 | |
130 | /// Return true if this alloca is in the entry block of the function and is a |
131 | /// constant size. If so, the code generator will fold it into the |
132 | /// prolog/epilog code, so it is basically free. |
133 | bool isStaticAlloca() const; |
134 | |
135 | /// Return true if this alloca is used as an inalloca argument to a call. Such |
136 | /// allocas are never considered static even if they are in the entry block. |
137 | bool isUsedWithInAlloca() const { |
138 | return getSubclassData<UsedWithInAllocaField>(); |
139 | } |
140 | |
141 | /// Specify whether this alloca is used to represent the arguments to a call. |
142 | void setUsedWithInAlloca(bool V) { |
143 | setSubclassData<UsedWithInAllocaField>(V); |
144 | } |
145 | |
146 | /// Return true if this alloca is used as a swifterror argument to a call. |
147 | bool isSwiftError() const { return getSubclassData<SwiftErrorField>(); } |
148 | /// Specify whether this alloca is used to represent a swifterror. |
149 | void setSwiftError(bool V) { setSubclassData<SwiftErrorField>(V); } |
150 | |
151 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
152 | static bool classof(const Instruction *I) { |
153 | return (I->getOpcode() == Instruction::Alloca); |
154 | } |
155 | static bool classof(const Value *V) { |
156 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
157 | } |
158 | |
159 | private: |
160 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
161 | // method so that subclasses cannot accidentally use it. |
162 | template <typename Bitfield> |
163 | void setSubclassData(typename Bitfield::Type Value) { |
164 | Instruction::setSubclassData<Bitfield>(Value); |
165 | } |
166 | }; |
167 | |
168 | //===----------------------------------------------------------------------===// |
169 | // LoadInst Class |
170 | //===----------------------------------------------------------------------===// |
171 | |
172 | /// An instruction for reading from memory. This uses the SubclassData field in |
173 | /// Value to store whether or not the load is volatile. |
174 | class LoadInst : public UnaryInstruction { |
175 | using VolatileField = BoolBitfieldElementT<0>; |
176 | using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>; |
177 | using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>; |
178 | static_assert( |
179 | Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(), |
180 | "Bitfields must be contiguous"); |
181 | |
182 | void AssertOK(); |
183 | |
184 | protected: |
185 | // Note: Instruction needs to be a friend here to call cloneImpl. |
186 | friend class Instruction; |
187 | |
188 | LoadInst *cloneImpl() const; |
189 | |
190 | public: |
191 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, |
192 | Instruction *InsertBefore); |
193 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, BasicBlock *InsertAtEnd); |
194 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
195 | Instruction *InsertBefore); |
196 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
197 | BasicBlock *InsertAtEnd); |
198 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
199 | Align Align, Instruction *InsertBefore = nullptr); |
200 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
201 | Align Align, BasicBlock *InsertAtEnd); |
202 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
203 | Align Align, AtomicOrdering Order, |
204 | SyncScope::ID SSID = SyncScope::System, |
205 | Instruction *InsertBefore = nullptr); |
206 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
207 | Align Align, AtomicOrdering Order, SyncScope::ID SSID, |
208 | BasicBlock *InsertAtEnd); |
209 | |
210 | /// Return true if this is a load from a volatile memory location. |
211 | bool isVolatile() const { return getSubclassData<VolatileField>(); } |
212 | |
213 | /// Specify whether this is a volatile load or not. |
214 | void setVolatile(bool V) { setSubclassData<VolatileField>(V); } |
215 | |
216 | /// Return the alignment of the access that is being performed. |
217 | /// FIXME: Remove this function once transition to Align is over. |
218 | /// Use getAlign() instead. |
219 | unsigned getAlignment() const { return getAlign().value(); } |
220 | |
221 | /// Return the alignment of the access that is being performed. |
222 | Align getAlign() const { |
223 | return Align(1ULL << (getSubclassData<AlignmentField>())); |
224 | } |
225 | |
226 | void setAlignment(Align Align) { |
227 | setSubclassData<AlignmentField>(Log2(Align)); |
228 | } |
229 | |
230 | /// Returns the ordering constraint of this load instruction. |
231 | AtomicOrdering getOrdering() const { |
232 | return getSubclassData<OrderingField>(); |
233 | } |
234 | /// Sets the ordering constraint of this load instruction. May not be Release |
235 | /// or AcquireRelease. |
236 | void setOrdering(AtomicOrdering Ordering) { |
237 | setSubclassData<OrderingField>(Ordering); |
238 | } |
239 | |
240 | /// Returns the synchronization scope ID of this load instruction. |
241 | SyncScope::ID getSyncScopeID() const { |
242 | return SSID; |
243 | } |
244 | |
245 | /// Sets the synchronization scope ID of this load instruction. |
246 | void setSyncScopeID(SyncScope::ID SSID) { |
247 | this->SSID = SSID; |
248 | } |
249 | |
250 | /// Sets the ordering constraint and the synchronization scope ID of this load |
251 | /// instruction. |
252 | void setAtomic(AtomicOrdering Ordering, |
253 | SyncScope::ID SSID = SyncScope::System) { |
254 | setOrdering(Ordering); |
255 | setSyncScopeID(SSID); |
256 | } |
257 | |
258 | bool isSimple() const { return !isAtomic() && !isVolatile(); } |
259 | |
260 | bool isUnordered() const { |
261 | return (getOrdering() == AtomicOrdering::NotAtomic || |
262 | getOrdering() == AtomicOrdering::Unordered) && |
263 | !isVolatile(); |
264 | } |
265 | |
266 | Value *getPointerOperand() { return getOperand(0); } |
267 | const Value *getPointerOperand() const { return getOperand(0); } |
268 | static unsigned getPointerOperandIndex() { return 0U; } |
269 | Type *getPointerOperandType() const { return getPointerOperand()->getType(); } |
270 | |
271 | /// Returns the address space of the pointer operand. |
272 | unsigned getPointerAddressSpace() const { |
273 | return getPointerOperandType()->getPointerAddressSpace(); |
274 | } |
275 | |
276 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
277 | static bool classof(const Instruction *I) { |
278 | return I->getOpcode() == Instruction::Load; |
279 | } |
280 | static bool classof(const Value *V) { |
281 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
282 | } |
283 | |
284 | private: |
285 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
286 | // method so that subclasses cannot accidentally use it. |
287 | template <typename Bitfield> |
288 | void setSubclassData(typename Bitfield::Type Value) { |
289 | Instruction::setSubclassData<Bitfield>(Value); |
290 | } |
291 | |
292 | /// The synchronization scope ID of this load instruction. Not quite enough |
293 | /// room in SubClassData for everything, so synchronization scope ID gets its |
294 | /// own field. |
295 | SyncScope::ID SSID; |
296 | }; |
297 | |
298 | //===----------------------------------------------------------------------===// |
299 | // StoreInst Class |
300 | //===----------------------------------------------------------------------===// |
301 | |
302 | /// An instruction for storing to memory. |
303 | class StoreInst : public Instruction { |
304 | using VolatileField = BoolBitfieldElementT<0>; |
305 | using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>; |
306 | using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>; |
307 | static_assert( |
308 | Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(), |
309 | "Bitfields must be contiguous"); |
310 | |
311 | void AssertOK(); |
312 | |
313 | protected: |
314 | // Note: Instruction needs to be a friend here to call cloneImpl. |
315 | friend class Instruction; |
316 | |
317 | StoreInst *cloneImpl() const; |
318 | |
319 | public: |
320 | StoreInst(Value *Val, Value *Ptr, Instruction *InsertBefore); |
321 | StoreInst(Value *Val, Value *Ptr, BasicBlock *InsertAtEnd); |
322 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Instruction *InsertBefore); |
323 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, BasicBlock *InsertAtEnd); |
324 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
325 | Instruction *InsertBefore = nullptr); |
326 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
327 | BasicBlock *InsertAtEnd); |
328 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
329 | AtomicOrdering Order, SyncScope::ID SSID = SyncScope::System, |
330 | Instruction *InsertBefore = nullptr); |
331 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
332 | AtomicOrdering Order, SyncScope::ID SSID, BasicBlock *InsertAtEnd); |
333 | |
334 | // allocate space for exactly two operands |
335 | void *operator new(size_t s) { |
336 | return User::operator new(s, 2); |
337 | } |
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 { ((i_nocapture < OperandTraits <StoreInst>::operands(this) && "getOperand() out of range!" ) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<StoreInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 437, __PRETTY_FUNCTION__)); 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) { ((i_nocapture < OperandTraits<StoreInst>::operands(this) && "setOperand() out of range!" ) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<StoreInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 437, __PRETTY_FUNCTION__)); 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) { |
466 | return User::operator new(s, 0); |
467 | } |
468 | |
469 | /// Returns the ordering constraint of this fence instruction. |
470 | AtomicOrdering getOrdering() const { |
471 | return getSubclassData<OrderingField>(); |
472 | } |
473 | |
474 | /// Sets the ordering constraint of this fence instruction. May only be |
475 | /// Acquire, Release, AcquireRelease, or SequentiallyConsistent. |
476 | void setOrdering(AtomicOrdering Ordering) { |
477 | setSubclassData<OrderingField>(Ordering); |
478 | } |
479 | |
480 | /// Returns the synchronization scope ID of this fence instruction. |
481 | SyncScope::ID getSyncScopeID() const { |
482 | return SSID; |
483 | } |
484 | |
485 | /// Sets the synchronization scope ID of this fence instruction. |
486 | void setSyncScopeID(SyncScope::ID SSID) { |
487 | this->SSID = SSID; |
488 | } |
489 | |
490 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
491 | static bool classof(const Instruction *I) { |
492 | return I->getOpcode() == Instruction::Fence; |
493 | } |
494 | static bool classof(const Value *V) { |
495 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
496 | } |
497 | |
498 | private: |
499 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
500 | // method so that subclasses cannot accidentally use it. |
501 | template <typename Bitfield> |
502 | void setSubclassData(typename Bitfield::Type Value) { |
503 | Instruction::setSubclassData<Bitfield>(Value); |
504 | } |
505 | |
506 | /// The synchronization scope ID of this fence instruction. Not quite enough |
507 | /// room in SubClassData for everything, so synchronization scope ID gets its |
508 | /// own field. |
509 | SyncScope::ID SSID; |
510 | }; |
511 | |
512 | //===----------------------------------------------------------------------===// |
513 | // AtomicCmpXchgInst Class |
514 | //===----------------------------------------------------------------------===// |
515 | |
516 | /// An instruction that atomically checks whether a |
517 | /// specified value is in a memory location, and, if it is, stores a new value |
518 | /// there. The value returned by this instruction is a pair containing the |
519 | /// original value as first element, and an i1 indicating success (true) or |
520 | /// failure (false) as second element. |
521 | /// |
522 | class AtomicCmpXchgInst : public Instruction { |
523 | void Init(Value *Ptr, Value *Cmp, Value *NewVal, Align Align, |
524 | AtomicOrdering SuccessOrdering, AtomicOrdering FailureOrdering, |
525 | SyncScope::ID SSID); |
526 | |
527 | template <unsigned Offset> |
528 | using AtomicOrderingBitfieldElement = |
529 | typename Bitfield::Element<AtomicOrdering, Offset, 3, |
530 | AtomicOrdering::LAST>; |
531 | |
532 | protected: |
533 | // Note: Instruction needs to be a friend here to call cloneImpl. |
534 | friend class Instruction; |
535 | |
536 | AtomicCmpXchgInst *cloneImpl() const; |
537 | |
538 | public: |
539 | AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment, |
540 | AtomicOrdering SuccessOrdering, |
541 | AtomicOrdering FailureOrdering, SyncScope::ID SSID, |
542 | Instruction *InsertBefore = nullptr); |
543 | AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment, |
544 | AtomicOrdering SuccessOrdering, |
545 | AtomicOrdering FailureOrdering, SyncScope::ID SSID, |
546 | BasicBlock *InsertAtEnd); |
547 | |
548 | // allocate space for exactly three operands |
549 | void *operator new(size_t s) { |
550 | return User::operator new(s, 3); |
551 | } |
552 | |
553 | using VolatileField = BoolBitfieldElementT<0>; |
554 | using WeakField = BoolBitfieldElementT<VolatileField::NextBit>; |
555 | using SuccessOrderingField = |
556 | AtomicOrderingBitfieldElementT<WeakField::NextBit>; |
557 | using FailureOrderingField = |
558 | AtomicOrderingBitfieldElementT<SuccessOrderingField::NextBit>; |
559 | using AlignmentField = |
560 | AlignmentBitfieldElementT<FailureOrderingField::NextBit>; |
561 | static_assert( |
562 | Bitfield::areContiguous<VolatileField, WeakField, SuccessOrderingField, |
563 | FailureOrderingField, AlignmentField>(), |
564 | "Bitfields must be contiguous"); |
565 | |
566 | /// Return the alignment of the memory that is being allocated by the |
567 | /// instruction. |
568 | Align getAlign() const { |
569 | return Align(1ULL << getSubclassData<AlignmentField>()); |
570 | } |
571 | |
572 | void setAlignment(Align Align) { |
573 | setSubclassData<AlignmentField>(Log2(Align)); |
574 | } |
575 | |
576 | /// Return true if this is a cmpxchg from a volatile memory |
577 | /// location. |
578 | /// |
579 | bool isVolatile() const { return getSubclassData<VolatileField>(); } |
580 | |
581 | /// Specify whether this is a volatile cmpxchg. |
582 | /// |
583 | void setVolatile(bool V) { setSubclassData<VolatileField>(V); } |
584 | |
585 | /// Return true if this cmpxchg may spuriously fail. |
586 | bool isWeak() const { return getSubclassData<WeakField>(); } |
587 | |
588 | void setWeak(bool IsWeak) { setSubclassData<WeakField>(IsWeak); } |
589 | |
590 | /// Transparently provide more efficient getOperand methods. |
591 | 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; |
592 | |
593 | /// Returns the success ordering constraint of this cmpxchg instruction. |
594 | AtomicOrdering getSuccessOrdering() const { |
595 | return getSubclassData<SuccessOrderingField>(); |
596 | } |
597 | |
598 | /// Sets the success ordering constraint of this cmpxchg instruction. |
599 | void setSuccessOrdering(AtomicOrdering Ordering) { |
600 | assert(Ordering != AtomicOrdering::NotAtomic &&((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic." ) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 601, __PRETTY_FUNCTION__)) |
601 | "CmpXchg instructions can only be atomic.")((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic." ) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 601, __PRETTY_FUNCTION__)); |
602 | setSubclassData<SuccessOrderingField>(Ordering); |
603 | } |
604 | |
605 | /// Returns the failure ordering constraint of this cmpxchg instruction. |
606 | AtomicOrdering getFailureOrdering() const { |
607 | return getSubclassData<FailureOrderingField>(); |
608 | } |
609 | |
610 | /// Sets the failure ordering constraint of this cmpxchg instruction. |
611 | void setFailureOrdering(AtomicOrdering Ordering) { |
612 | assert(Ordering != AtomicOrdering::NotAtomic &&((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic." ) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 613, __PRETTY_FUNCTION__)) |
613 | "CmpXchg instructions can only be atomic.")((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic." ) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 613, __PRETTY_FUNCTION__)); |
614 | setSubclassData<FailureOrderingField>(Ordering); |
615 | } |
616 | |
617 | /// Returns the synchronization scope ID of this cmpxchg instruction. |
618 | SyncScope::ID getSyncScopeID() const { |
619 | return SSID; |
620 | } |
621 | |
622 | /// Sets the synchronization scope ID of this cmpxchg instruction. |
623 | void setSyncScopeID(SyncScope::ID SSID) { |
624 | this->SSID = SSID; |
625 | } |
626 | |
627 | Value *getPointerOperand() { return getOperand(0); } |
628 | const Value *getPointerOperand() const { return getOperand(0); } |
629 | static unsigned getPointerOperandIndex() { return 0U; } |
630 | |
631 | Value *getCompareOperand() { return getOperand(1); } |
632 | const Value *getCompareOperand() const { return getOperand(1); } |
633 | |
634 | Value *getNewValOperand() { return getOperand(2); } |
635 | const Value *getNewValOperand() const { return getOperand(2); } |
636 | |
637 | /// Returns the address space of the pointer operand. |
638 | unsigned getPointerAddressSpace() const { |
639 | return getPointerOperand()->getType()->getPointerAddressSpace(); |
640 | } |
641 | |
642 | /// Returns the strongest permitted ordering on failure, given the |
643 | /// desired ordering on success. |
644 | /// |
645 | /// If the comparison in a cmpxchg operation fails, there is no atomic store |
646 | /// so release semantics cannot be provided. So this function drops explicit |
647 | /// Release requests from the AtomicOrdering. A SequentiallyConsistent |
648 | /// operation would remain SequentiallyConsistent. |
649 | static AtomicOrdering |
650 | getStrongestFailureOrdering(AtomicOrdering SuccessOrdering) { |
651 | switch (SuccessOrdering) { |
652 | default: |
653 | llvm_unreachable("invalid cmpxchg success ordering")::llvm::llvm_unreachable_internal("invalid cmpxchg success ordering" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 653); |
654 | case AtomicOrdering::Release: |
655 | case AtomicOrdering::Monotonic: |
656 | return AtomicOrdering::Monotonic; |
657 | case AtomicOrdering::AcquireRelease: |
658 | case AtomicOrdering::Acquire: |
659 | return AtomicOrdering::Acquire; |
660 | case AtomicOrdering::SequentiallyConsistent: |
661 | return AtomicOrdering::SequentiallyConsistent; |
662 | } |
663 | } |
664 | |
665 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
666 | static bool classof(const Instruction *I) { |
667 | return I->getOpcode() == Instruction::AtomicCmpXchg; |
668 | } |
669 | static bool classof(const Value *V) { |
670 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
671 | } |
672 | |
673 | private: |
674 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
675 | // method so that subclasses cannot accidentally use it. |
676 | template <typename Bitfield> |
677 | void setSubclassData(typename Bitfield::Type Value) { |
678 | Instruction::setSubclassData<Bitfield>(Value); |
679 | } |
680 | |
681 | /// The synchronization scope ID of this cmpxchg instruction. Not quite |
682 | /// enough room in SubClassData for everything, so synchronization scope ID |
683 | /// gets its own field. |
684 | SyncScope::ID SSID; |
685 | }; |
686 | |
687 | template <> |
688 | struct OperandTraits<AtomicCmpXchgInst> : |
689 | public FixedNumOperandTraits<AtomicCmpXchgInst, 3> { |
690 | }; |
691 | |
692 | 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 { ((i_nocapture < OperandTraits<AtomicCmpXchgInst >::operands(this) && "getOperand() out of range!") ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicCmpXchgInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 692, __PRETTY_FUNCTION__)); 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 ) { ((i_nocapture < OperandTraits<AtomicCmpXchgInst> ::operands(this) && "setOperand() out of range!") ? static_cast <void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicCmpXchgInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 692, __PRETTY_FUNCTION__)); 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); } |
693 | |
694 | //===----------------------------------------------------------------------===// |
695 | // AtomicRMWInst Class |
696 | //===----------------------------------------------------------------------===// |
697 | |
698 | /// an instruction that atomically reads a memory location, |
699 | /// combines it with another value, and then stores the result back. Returns |
700 | /// the old value. |
701 | /// |
702 | class AtomicRMWInst : public Instruction { |
703 | protected: |
704 | // Note: Instruction needs to be a friend here to call cloneImpl. |
705 | friend class Instruction; |
706 | |
707 | AtomicRMWInst *cloneImpl() const; |
708 | |
709 | public: |
710 | /// This enumeration lists the possible modifications atomicrmw can make. In |
711 | /// the descriptions, 'p' is the pointer to the instruction's memory location, |
712 | /// 'old' is the initial value of *p, and 'v' is the other value passed to the |
713 | /// instruction. These instructions always return 'old'. |
714 | enum BinOp : unsigned { |
715 | /// *p = v |
716 | Xchg, |
717 | /// *p = old + v |
718 | Add, |
719 | /// *p = old - v |
720 | Sub, |
721 | /// *p = old & v |
722 | And, |
723 | /// *p = ~(old & v) |
724 | Nand, |
725 | /// *p = old | v |
726 | Or, |
727 | /// *p = old ^ v |
728 | Xor, |
729 | /// *p = old >signed v ? old : v |
730 | Max, |
731 | /// *p = old <signed v ? old : v |
732 | Min, |
733 | /// *p = old >unsigned v ? old : v |
734 | UMax, |
735 | /// *p = old <unsigned v ? old : v |
736 | UMin, |
737 | |
738 | /// *p = old + v |
739 | FAdd, |
740 | |
741 | /// *p = old - v |
742 | FSub, |
743 | |
744 | FIRST_BINOP = Xchg, |
745 | LAST_BINOP = FSub, |
746 | BAD_BINOP |
747 | }; |
748 | |
749 | private: |
750 | template <unsigned Offset> |
751 | using AtomicOrderingBitfieldElement = |
752 | typename Bitfield::Element<AtomicOrdering, Offset, 3, |
753 | AtomicOrdering::LAST>; |
754 | |
755 | template <unsigned Offset> |
756 | using BinOpBitfieldElement = |
757 | typename Bitfield::Element<BinOp, Offset, 4, BinOp::LAST_BINOP>; |
758 | |
759 | public: |
760 | AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment, |
761 | AtomicOrdering Ordering, SyncScope::ID SSID, |
762 | Instruction *InsertBefore = nullptr); |
763 | AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment, |
764 | AtomicOrdering Ordering, SyncScope::ID SSID, |
765 | BasicBlock *InsertAtEnd); |
766 | |
767 | // allocate space for exactly two operands |
768 | void *operator new(size_t s) { |
769 | return User::operator new(s, 2); |
770 | } |
771 | |
772 | using VolatileField = BoolBitfieldElementT<0>; |
773 | using AtomicOrderingField = |
774 | AtomicOrderingBitfieldElementT<VolatileField::NextBit>; |
775 | using OperationField = BinOpBitfieldElement<AtomicOrderingField::NextBit>; |
776 | using AlignmentField = AlignmentBitfieldElementT<OperationField::NextBit>; |
777 | static_assert(Bitfield::areContiguous<VolatileField, AtomicOrderingField, |
778 | OperationField, AlignmentField>(), |
779 | "Bitfields must be contiguous"); |
780 | |
781 | BinOp getOperation() const { return getSubclassData<OperationField>(); } |
782 | |
783 | static StringRef getOperationName(BinOp Op); |
784 | |
785 | static bool isFPOperation(BinOp Op) { |
786 | switch (Op) { |
787 | case AtomicRMWInst::FAdd: |
788 | case AtomicRMWInst::FSub: |
789 | return true; |
790 | default: |
791 | return false; |
792 | } |
793 | } |
794 | |
795 | void setOperation(BinOp Operation) { |
796 | setSubclassData<OperationField>(Operation); |
797 | } |
798 | |
799 | /// Return the alignment of the memory that is being allocated by the |
800 | /// instruction. |
801 | Align getAlign() const { |
802 | return Align(1ULL << getSubclassData<AlignmentField>()); |
803 | } |
804 | |
805 | void setAlignment(Align Align) { |
806 | setSubclassData<AlignmentField>(Log2(Align)); |
807 | } |
808 | |
809 | /// Return true if this is a RMW on a volatile memory location. |
810 | /// |
811 | bool isVolatile() const { return getSubclassData<VolatileField>(); } |
812 | |
813 | /// Specify whether this is a volatile RMW or not. |
814 | /// |
815 | void setVolatile(bool V) { setSubclassData<VolatileField>(V); } |
816 | |
817 | /// Transparently provide more efficient getOperand methods. |
818 | 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; |
819 | |
820 | /// Returns the ordering constraint of this rmw instruction. |
821 | AtomicOrdering getOrdering() const { |
822 | return getSubclassData<AtomicOrderingField>(); |
823 | } |
824 | |
825 | /// Sets the ordering constraint of this rmw instruction. |
826 | void setOrdering(AtomicOrdering Ordering) { |
827 | assert(Ordering != AtomicOrdering::NotAtomic &&((Ordering != AtomicOrdering::NotAtomic && "atomicrmw instructions can only be atomic." ) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"atomicrmw instructions can only be atomic.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 828, __PRETTY_FUNCTION__)) |
828 | "atomicrmw instructions can only be atomic.")((Ordering != AtomicOrdering::NotAtomic && "atomicrmw instructions can only be atomic." ) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"atomicrmw instructions can only be atomic.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 828, __PRETTY_FUNCTION__)); |
829 | setSubclassData<AtomicOrderingField>(Ordering); |
830 | } |
831 | |
832 | /// Returns the synchronization scope ID of this rmw instruction. |
833 | SyncScope::ID getSyncScopeID() const { |
834 | return SSID; |
835 | } |
836 | |
837 | /// Sets the synchronization scope ID of this rmw instruction. |
838 | void setSyncScopeID(SyncScope::ID SSID) { |
839 | this->SSID = SSID; |
840 | } |
841 | |
842 | Value *getPointerOperand() { return getOperand(0); } |
843 | const Value *getPointerOperand() const { return getOperand(0); } |
844 | static unsigned getPointerOperandIndex() { return 0U; } |
845 | |
846 | Value *getValOperand() { return getOperand(1); } |
847 | const Value *getValOperand() const { return getOperand(1); } |
848 | |
849 | /// Returns the address space of the pointer operand. |
850 | unsigned getPointerAddressSpace() const { |
851 | return getPointerOperand()->getType()->getPointerAddressSpace(); |
852 | } |
853 | |
854 | bool isFloatingPointOperation() const { |
855 | return isFPOperation(getOperation()); |
856 | } |
857 | |
858 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
859 | static bool classof(const Instruction *I) { |
860 | return I->getOpcode() == Instruction::AtomicRMW; |
861 | } |
862 | static bool classof(const Value *V) { |
863 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
864 | } |
865 | |
866 | private: |
867 | void Init(BinOp Operation, Value *Ptr, Value *Val, Align Align, |
868 | AtomicOrdering Ordering, SyncScope::ID SSID); |
869 | |
870 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
871 | // method so that subclasses cannot accidentally use it. |
872 | template <typename Bitfield> |
873 | void setSubclassData(typename Bitfield::Type Value) { |
874 | Instruction::setSubclassData<Bitfield>(Value); |
875 | } |
876 | |
877 | /// The synchronization scope ID of this rmw instruction. Not quite enough |
878 | /// room in SubClassData for everything, so synchronization scope ID gets its |
879 | /// own field. |
880 | SyncScope::ID SSID; |
881 | }; |
882 | |
883 | template <> |
884 | struct OperandTraits<AtomicRMWInst> |
885 | : public FixedNumOperandTraits<AtomicRMWInst,2> { |
886 | }; |
887 | |
888 | 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 { ((i_nocapture < OperandTraits <AtomicRMWInst>::operands(this) && "getOperand() out of range!" ) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicRMWInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 888, __PRETTY_FUNCTION__)); 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) { (( i_nocapture < OperandTraits<AtomicRMWInst>::operands (this) && "setOperand() out of range!") ? static_cast <void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicRMWInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 888, __PRETTY_FUNCTION__)); 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); } |
889 | |
890 | //===----------------------------------------------------------------------===// |
891 | // GetElementPtrInst Class |
892 | //===----------------------------------------------------------------------===// |
893 | |
894 | // checkGEPType - Simple wrapper function to give a better assertion failure |
895 | // message on bad indexes for a gep instruction. |
896 | // |
897 | inline Type *checkGEPType(Type *Ty) { |
898 | assert(Ty && "Invalid GetElementPtrInst indices for type!")((Ty && "Invalid GetElementPtrInst indices for type!" ) ? static_cast<void> (0) : __assert_fail ("Ty && \"Invalid GetElementPtrInst indices for type!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 898, __PRETTY_FUNCTION__)); |
899 | return Ty; |
900 | } |
901 | |
902 | /// an instruction for type-safe pointer arithmetic to |
903 | /// access elements of arrays and structs |
904 | /// |
905 | class GetElementPtrInst : public Instruction { |
906 | Type *SourceElementType; |
907 | Type *ResultElementType; |
908 | |
909 | GetElementPtrInst(const GetElementPtrInst &GEPI); |
910 | |
911 | /// Constructors - Create a getelementptr instruction with a base pointer an |
912 | /// list of indices. The first ctor can optionally insert before an existing |
913 | /// instruction, the second appends the new instruction to the specified |
914 | /// BasicBlock. |
915 | inline GetElementPtrInst(Type *PointeeType, Value *Ptr, |
916 | ArrayRef<Value *> IdxList, unsigned Values, |
917 | const Twine &NameStr, Instruction *InsertBefore); |
918 | inline GetElementPtrInst(Type *PointeeType, Value *Ptr, |
919 | ArrayRef<Value *> IdxList, unsigned Values, |
920 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
921 | |
922 | void init(Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr); |
923 | |
924 | protected: |
925 | // Note: Instruction needs to be a friend here to call cloneImpl. |
926 | friend class Instruction; |
927 | |
928 | GetElementPtrInst *cloneImpl() const; |
929 | |
930 | public: |
931 | static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr, |
932 | ArrayRef<Value *> IdxList, |
933 | const Twine &NameStr = "", |
934 | Instruction *InsertBefore = nullptr) { |
935 | unsigned Values = 1 + unsigned(IdxList.size()); |
936 | if (!PointeeType) |
937 | PointeeType = |
938 | cast<PointerType>(Ptr->getType()->getScalarType())->getElementType(); |
939 | else |
940 | assert(((PointeeType == cast<PointerType>(Ptr->getType()-> getScalarType())->getElementType()) ? static_cast<void> (0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 942, __PRETTY_FUNCTION__)) |
941 | PointeeType ==((PointeeType == cast<PointerType>(Ptr->getType()-> getScalarType())->getElementType()) ? static_cast<void> (0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 942, __PRETTY_FUNCTION__)) |
942 | cast<PointerType>(Ptr->getType()->getScalarType())->getElementType())((PointeeType == cast<PointerType>(Ptr->getType()-> getScalarType())->getElementType()) ? static_cast<void> (0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 942, __PRETTY_FUNCTION__)); |
943 | return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values, |
944 | NameStr, InsertBefore); |
945 | } |
946 | |
947 | static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr, |
948 | ArrayRef<Value *> IdxList, |
949 | const Twine &NameStr, |
950 | BasicBlock *InsertAtEnd) { |
951 | unsigned Values = 1 + unsigned(IdxList.size()); |
952 | if (!PointeeType) |
953 | PointeeType = |
954 | cast<PointerType>(Ptr->getType()->getScalarType())->getElementType(); |
955 | else |
956 | assert(((PointeeType == cast<PointerType>(Ptr->getType()-> getScalarType())->getElementType()) ? static_cast<void> (0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 958, __PRETTY_FUNCTION__)) |
957 | PointeeType ==((PointeeType == cast<PointerType>(Ptr->getType()-> getScalarType())->getElementType()) ? static_cast<void> (0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 958, __PRETTY_FUNCTION__)) |
958 | cast<PointerType>(Ptr->getType()->getScalarType())->getElementType())((PointeeType == cast<PointerType>(Ptr->getType()-> getScalarType())->getElementType()) ? static_cast<void> (0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 958, __PRETTY_FUNCTION__)); |
959 | return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values, |
960 | NameStr, InsertAtEnd); |
961 | } |
962 | |
963 | /// Create an "inbounds" getelementptr. See the documentation for the |
964 | /// "inbounds" flag in LangRef.html for details. |
965 | static GetElementPtrInst *CreateInBounds(Value *Ptr, |
966 | ArrayRef<Value *> IdxList, |
967 | const Twine &NameStr = "", |
968 | Instruction *InsertBefore = nullptr){ |
969 | return CreateInBounds(nullptr, Ptr, IdxList, NameStr, InsertBefore); |
970 | } |
971 | |
972 | static GetElementPtrInst * |
973 | CreateInBounds(Type *PointeeType, Value *Ptr, ArrayRef<Value *> IdxList, |
974 | const Twine &NameStr = "", |
975 | Instruction *InsertBefore = nullptr) { |
976 | GetElementPtrInst *GEP = |
977 | Create(PointeeType, Ptr, IdxList, NameStr, InsertBefore); |
978 | GEP->setIsInBounds(true); |
979 | return GEP; |
980 | } |
981 | |
982 | static GetElementPtrInst *CreateInBounds(Value *Ptr, |
983 | ArrayRef<Value *> IdxList, |
984 | const Twine &NameStr, |
985 | BasicBlock *InsertAtEnd) { |
986 | return CreateInBounds(nullptr, Ptr, IdxList, NameStr, InsertAtEnd); |
987 | } |
988 | |
989 | static GetElementPtrInst *CreateInBounds(Type *PointeeType, Value *Ptr, |
990 | ArrayRef<Value *> IdxList, |
991 | const Twine &NameStr, |
992 | BasicBlock *InsertAtEnd) { |
993 | GetElementPtrInst *GEP = |
994 | Create(PointeeType, Ptr, IdxList, NameStr, InsertAtEnd); |
995 | GEP->setIsInBounds(true); |
996 | return GEP; |
997 | } |
998 | |
999 | /// Transparently provide more efficient getOperand methods. |
1000 | 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; |
1001 | |
1002 | Type *getSourceElementType() const { return SourceElementType; } |
1003 | |
1004 | void setSourceElementType(Type *Ty) { SourceElementType = Ty; } |
1005 | void setResultElementType(Type *Ty) { ResultElementType = Ty; } |
1006 | |
1007 | Type *getResultElementType() const { |
1008 | assert(ResultElementType ==((ResultElementType == cast<PointerType>(getType()-> getScalarType())->getElementType()) ? static_cast<void> (0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1009, __PRETTY_FUNCTION__)) |
1009 | cast<PointerType>(getType()->getScalarType())->getElementType())((ResultElementType == cast<PointerType>(getType()-> getScalarType())->getElementType()) ? static_cast<void> (0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1009, __PRETTY_FUNCTION__)); |
1010 | return ResultElementType; |
1011 | } |
1012 | |
1013 | /// Returns the address space of this instruction's pointer type. |
1014 | unsigned getAddressSpace() const { |
1015 | // Note that this is always the same as the pointer operand's address space |
1016 | // and that is cheaper to compute, so cheat here. |
1017 | return getPointerAddressSpace(); |
1018 | } |
1019 | |
1020 | /// Returns the result type of a getelementptr with the given source |
1021 | /// element type and indexes. |
1022 | /// |
1023 | /// Null is returned if the indices are invalid for the specified |
1024 | /// source element type. |
1025 | static Type *getIndexedType(Type *Ty, ArrayRef<Value *> IdxList); |
1026 | static Type *getIndexedType(Type *Ty, ArrayRef<Constant *> IdxList); |
1027 | static Type *getIndexedType(Type *Ty, ArrayRef<uint64_t> IdxList); |
1028 | |
1029 | /// Return the type of the element at the given index of an indexable |
1030 | /// type. This is equivalent to "getIndexedType(Agg, {Zero, Idx})". |
1031 | /// |
1032 | /// Returns null if the type can't be indexed, or the given index is not |
1033 | /// legal for the given type. |
1034 | static Type *getTypeAtIndex(Type *Ty, Value *Idx); |
1035 | static Type *getTypeAtIndex(Type *Ty, uint64_t Idx); |
1036 | |
1037 | inline op_iterator idx_begin() { return op_begin()+1; } |
1038 | inline const_op_iterator idx_begin() const { return op_begin()+1; } |
1039 | inline op_iterator idx_end() { return op_end(); } |
1040 | inline const_op_iterator idx_end() const { return op_end(); } |
1041 | |
1042 | inline iterator_range<op_iterator> indices() { |
1043 | return make_range(idx_begin(), idx_end()); |
1044 | } |
1045 | |
1046 | inline iterator_range<const_op_iterator> indices() const { |
1047 | return make_range(idx_begin(), idx_end()); |
1048 | } |
1049 | |
1050 | Value *getPointerOperand() { |
1051 | return getOperand(0); |
1052 | } |
1053 | const Value *getPointerOperand() const { |
1054 | return getOperand(0); |
1055 | } |
1056 | static unsigned getPointerOperandIndex() { |
1057 | return 0U; // get index for modifying correct operand. |
1058 | } |
1059 | |
1060 | /// Method to return the pointer operand as a |
1061 | /// PointerType. |
1062 | Type *getPointerOperandType() const { |
1063 | return getPointerOperand()->getType(); |
1064 | } |
1065 | |
1066 | /// Returns the address space of the pointer operand. |
1067 | unsigned getPointerAddressSpace() const { |
1068 | return getPointerOperandType()->getPointerAddressSpace(); |
1069 | } |
1070 | |
1071 | /// Returns the pointer type returned by the GEP |
1072 | /// instruction, which may be a vector of pointers. |
1073 | static Type *getGEPReturnType(Type *ElTy, Value *Ptr, |
1074 | ArrayRef<Value *> IdxList) { |
1075 | Type *PtrTy = PointerType::get(checkGEPType(getIndexedType(ElTy, IdxList)), |
1076 | Ptr->getType()->getPointerAddressSpace()); |
1077 | // Vector GEP |
1078 | if (auto *PtrVTy = dyn_cast<VectorType>(Ptr->getType())) { |
1079 | ElementCount EltCount = PtrVTy->getElementCount(); |
1080 | return VectorType::get(PtrTy, EltCount); |
1081 | } |
1082 | for (Value *Index : IdxList) |
1083 | if (auto *IndexVTy = dyn_cast<VectorType>(Index->getType())) { |
1084 | ElementCount EltCount = IndexVTy->getElementCount(); |
1085 | return VectorType::get(PtrTy, EltCount); |
1086 | } |
1087 | // Scalar GEP |
1088 | return PtrTy; |
1089 | } |
1090 | |
1091 | unsigned getNumIndices() const { // Note: always non-negative |
1092 | return getNumOperands() - 1; |
1093 | } |
1094 | |
1095 | bool hasIndices() const { |
1096 | return getNumOperands() > 1; |
1097 | } |
1098 | |
1099 | /// Return true if all of the indices of this GEP are |
1100 | /// zeros. If so, the result pointer and the first operand have the same |
1101 | /// value, just potentially different types. |
1102 | bool hasAllZeroIndices() const; |
1103 | |
1104 | /// Return true if all of the indices of this GEP are |
1105 | /// constant integers. If so, the result pointer and the first operand have |
1106 | /// a constant offset between them. |
1107 | bool hasAllConstantIndices() const; |
1108 | |
1109 | /// Set or clear the inbounds flag on this GEP instruction. |
1110 | /// See LangRef.html for the meaning of inbounds on a getelementptr. |
1111 | void setIsInBounds(bool b = true); |
1112 | |
1113 | /// Determine whether the GEP has the inbounds flag. |
1114 | bool isInBounds() const; |
1115 | |
1116 | /// Accumulate the constant address offset of this GEP if possible. |
1117 | /// |
1118 | /// This routine accepts an APInt into which it will accumulate the constant |
1119 | /// offset of this GEP if the GEP is in fact constant. If the GEP is not |
1120 | /// all-constant, it returns false and the value of the offset APInt is |
1121 | /// undefined (it is *not* preserved!). The APInt passed into this routine |
1122 | /// must be at least as wide as the IntPtr type for the address space of |
1123 | /// the base GEP pointer. |
1124 | bool accumulateConstantOffset(const DataLayout &DL, APInt &Offset) const; |
1125 | |
1126 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1127 | static bool classof(const Instruction *I) { |
1128 | return (I->getOpcode() == Instruction::GetElementPtr); |
1129 | } |
1130 | static bool classof(const Value *V) { |
1131 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1132 | } |
1133 | }; |
1134 | |
1135 | template <> |
1136 | struct OperandTraits<GetElementPtrInst> : |
1137 | public VariadicOperandTraits<GetElementPtrInst, 1> { |
1138 | }; |
1139 | |
1140 | GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr, |
1141 | ArrayRef<Value *> IdxList, unsigned Values, |
1142 | const Twine &NameStr, |
1143 | Instruction *InsertBefore) |
1144 | : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr, |
1145 | OperandTraits<GetElementPtrInst>::op_end(this) - Values, |
1146 | Values, InsertBefore), |
1147 | SourceElementType(PointeeType), |
1148 | ResultElementType(getIndexedType(PointeeType, IdxList)) { |
1149 | assert(ResultElementType ==((ResultElementType == cast<PointerType>(getType()-> getScalarType())->getElementType()) ? static_cast<void> (0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1150, __PRETTY_FUNCTION__)) |
1150 | cast<PointerType>(getType()->getScalarType())->getElementType())((ResultElementType == cast<PointerType>(getType()-> getScalarType())->getElementType()) ? static_cast<void> (0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1150, __PRETTY_FUNCTION__)); |
1151 | init(Ptr, IdxList, NameStr); |
1152 | } |
1153 | |
1154 | GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr, |
1155 | ArrayRef<Value *> IdxList, unsigned Values, |
1156 | const Twine &NameStr, |
1157 | BasicBlock *InsertAtEnd) |
1158 | : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr, |
1159 | OperandTraits<GetElementPtrInst>::op_end(this) - Values, |
1160 | Values, InsertAtEnd), |
1161 | SourceElementType(PointeeType), |
1162 | ResultElementType(getIndexedType(PointeeType, IdxList)) { |
1163 | assert(ResultElementType ==((ResultElementType == cast<PointerType>(getType()-> getScalarType())->getElementType()) ? static_cast<void> (0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1164, __PRETTY_FUNCTION__)) |
1164 | cast<PointerType>(getType()->getScalarType())->getElementType())((ResultElementType == cast<PointerType>(getType()-> getScalarType())->getElementType()) ? static_cast<void> (0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1164, __PRETTY_FUNCTION__)); |
1165 | init(Ptr, IdxList, NameStr); |
1166 | } |
1167 | |
1168 | 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 { ((i_nocapture < OperandTraits<GetElementPtrInst >::operands(this) && "getOperand() out of range!") ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<GetElementPtrInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1168, __PRETTY_FUNCTION__)); 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 ) { ((i_nocapture < OperandTraits<GetElementPtrInst> ::operands(this) && "setOperand() out of range!") ? static_cast <void> (0) : __assert_fail ("i_nocapture < OperandTraits<GetElementPtrInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1168, __PRETTY_FUNCTION__)); 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); } |
1169 | |
1170 | //===----------------------------------------------------------------------===// |
1171 | // ICmpInst Class |
1172 | //===----------------------------------------------------------------------===// |
1173 | |
1174 | /// This instruction compares its operands according to the predicate given |
1175 | /// to the constructor. It only operates on integers or pointers. The operands |
1176 | /// must be identical types. |
1177 | /// Represent an integer comparison operator. |
1178 | class ICmpInst: public CmpInst { |
1179 | void AssertOK() { |
1180 | assert(isIntPredicate() &&((isIntPredicate() && "Invalid ICmp predicate value") ? static_cast<void> (0) : __assert_fail ("isIntPredicate() && \"Invalid ICmp predicate value\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1181, __PRETTY_FUNCTION__)) |
1181 | "Invalid ICmp predicate value")((isIntPredicate() && "Invalid ICmp predicate value") ? static_cast<void> (0) : __assert_fail ("isIntPredicate() && \"Invalid ICmp predicate value\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1181, __PRETTY_FUNCTION__)); |
1182 | assert(getOperand(0)->getType() == getOperand(1)->getType() &&((getOperand(0)->getType() == getOperand(1)->getType() && "Both operands to ICmp instruction are not of the same type!" ) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to ICmp instruction are not of the same type!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1183, __PRETTY_FUNCTION__)) |
1183 | "Both operands to ICmp instruction are not of the same type!")((getOperand(0)->getType() == getOperand(1)->getType() && "Both operands to ICmp instruction are not of the same type!" ) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to ICmp instruction are not of the same type!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1183, __PRETTY_FUNCTION__)); |
1184 | // Check that the operands are the right type |
1185 | assert((getOperand(0)->getType()->isIntOrIntVectorTy() ||(((getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand (0)->getType()->isPtrOrPtrVectorTy()) && "Invalid operand types for ICmp instruction" ) ? static_cast<void> (0) : __assert_fail ("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1187, __PRETTY_FUNCTION__)) |
1186 | getOperand(0)->getType()->isPtrOrPtrVectorTy()) &&(((getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand (0)->getType()->isPtrOrPtrVectorTy()) && "Invalid operand types for ICmp instruction" ) ? static_cast<void> (0) : __assert_fail ("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1187, __PRETTY_FUNCTION__)) |
1187 | "Invalid operand types for ICmp instruction")(((getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand (0)->getType()->isPtrOrPtrVectorTy()) && "Invalid operand types for ICmp instruction" ) ? static_cast<void> (0) : __assert_fail ("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1187, __PRETTY_FUNCTION__)); |
1188 | } |
1189 | |
1190 | protected: |
1191 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1192 | friend class Instruction; |
1193 | |
1194 | /// Clone an identical ICmpInst |
1195 | ICmpInst *cloneImpl() const; |
1196 | |
1197 | public: |
1198 | /// Constructor with insert-before-instruction semantics. |
1199 | ICmpInst( |
1200 | Instruction *InsertBefore, ///< Where to insert |
1201 | Predicate pred, ///< The predicate to use for the comparison |
1202 | Value *LHS, ///< The left-hand-side of the expression |
1203 | Value *RHS, ///< The right-hand-side of the expression |
1204 | const Twine &NameStr = "" ///< Name of the instruction |
1205 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1206 | Instruction::ICmp, pred, LHS, RHS, NameStr, |
1207 | InsertBefore) { |
1208 | #ifndef NDEBUG |
1209 | AssertOK(); |
1210 | #endif |
1211 | } |
1212 | |
1213 | /// Constructor with insert-at-end semantics. |
1214 | ICmpInst( |
1215 | BasicBlock &InsertAtEnd, ///< Block to insert into. |
1216 | Predicate pred, ///< The predicate to use for the comparison |
1217 | Value *LHS, ///< The left-hand-side of the expression |
1218 | Value *RHS, ///< The right-hand-side of the expression |
1219 | const Twine &NameStr = "" ///< Name of the instruction |
1220 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1221 | Instruction::ICmp, pred, LHS, RHS, NameStr, |
1222 | &InsertAtEnd) { |
1223 | #ifndef NDEBUG |
1224 | AssertOK(); |
1225 | #endif |
1226 | } |
1227 | |
1228 | /// Constructor with no-insertion semantics |
1229 | ICmpInst( |
1230 | Predicate pred, ///< The predicate to use for the comparison |
1231 | Value *LHS, ///< The left-hand-side of the expression |
1232 | Value *RHS, ///< The right-hand-side of the expression |
1233 | const Twine &NameStr = "" ///< Name of the instruction |
1234 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1235 | Instruction::ICmp, pred, LHS, RHS, NameStr) { |
1236 | #ifndef NDEBUG |
1237 | AssertOK(); |
1238 | #endif |
1239 | } |
1240 | |
1241 | /// For example, EQ->EQ, SLE->SLE, UGT->SGT, etc. |
1242 | /// @returns the predicate that would be the result if the operand were |
1243 | /// regarded as signed. |
1244 | /// Return the signed version of the predicate |
1245 | Predicate getSignedPredicate() const { |
1246 | return getSignedPredicate(getPredicate()); |
1247 | } |
1248 | |
1249 | /// This is a static version that you can use without an instruction. |
1250 | /// Return the signed version of the predicate. |
1251 | static Predicate getSignedPredicate(Predicate pred); |
1252 | |
1253 | /// For example, EQ->EQ, SLE->ULE, UGT->UGT, etc. |
1254 | /// @returns the predicate that would be the result if the operand were |
1255 | /// regarded as unsigned. |
1256 | /// Return the unsigned version of the predicate |
1257 | Predicate getUnsignedPredicate() const { |
1258 | return getUnsignedPredicate(getPredicate()); |
1259 | } |
1260 | |
1261 | /// This is a static version that you can use without an instruction. |
1262 | /// Return the unsigned version of the predicate. |
1263 | static Predicate getUnsignedPredicate(Predicate pred); |
1264 | |
1265 | /// Return true if this predicate is either EQ or NE. This also |
1266 | /// tests for commutativity. |
1267 | static bool isEquality(Predicate P) { |
1268 | return P == ICMP_EQ || P == ICMP_NE; |
1269 | } |
1270 | |
1271 | /// Return true if this predicate is either EQ or NE. This also |
1272 | /// tests for commutativity. |
1273 | bool isEquality() const { |
1274 | return isEquality(getPredicate()); |
1275 | } |
1276 | |
1277 | /// @returns true if the predicate of this ICmpInst is commutative |
1278 | /// Determine if this relation is commutative. |
1279 | bool isCommutative() const { return isEquality(); } |
1280 | |
1281 | /// Return true if the predicate is relational (not EQ or NE). |
1282 | /// |
1283 | bool isRelational() const { |
1284 | return !isEquality(); |
1285 | } |
1286 | |
1287 | /// Return true if the predicate is relational (not EQ or NE). |
1288 | /// |
1289 | static bool isRelational(Predicate P) { |
1290 | return !isEquality(P); |
1291 | } |
1292 | |
1293 | /// Return true if the predicate is SGT or UGT. |
1294 | /// |
1295 | static bool isGT(Predicate P) { |
1296 | return P == ICMP_SGT || P == ICMP_UGT; |
1297 | } |
1298 | |
1299 | /// Return true if the predicate is SLT or ULT. |
1300 | /// |
1301 | static bool isLT(Predicate P) { |
1302 | return P == ICMP_SLT || P == ICMP_ULT; |
1303 | } |
1304 | |
1305 | /// Return true if the predicate is SGE or UGE. |
1306 | /// |
1307 | static bool isGE(Predicate P) { |
1308 | return P == ICMP_SGE || P == ICMP_UGE; |
1309 | } |
1310 | |
1311 | /// Return true if the predicate is SLE or ULE. |
1312 | /// |
1313 | static bool isLE(Predicate P) { |
1314 | return P == ICMP_SLE || P == ICMP_ULE; |
1315 | } |
1316 | |
1317 | /// Exchange the two operands to this instruction in such a way that it does |
1318 | /// not modify the semantics of the instruction. The predicate value may be |
1319 | /// changed to retain the same result if the predicate is order dependent |
1320 | /// (e.g. ult). |
1321 | /// Swap operands and adjust predicate. |
1322 | void swapOperands() { |
1323 | setPredicate(getSwappedPredicate()); |
1324 | Op<0>().swap(Op<1>()); |
1325 | } |
1326 | |
1327 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1328 | static bool classof(const Instruction *I) { |
1329 | return I->getOpcode() == Instruction::ICmp; |
1330 | } |
1331 | static bool classof(const Value *V) { |
1332 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1333 | } |
1334 | }; |
1335 | |
1336 | //===----------------------------------------------------------------------===// |
1337 | // FCmpInst Class |
1338 | //===----------------------------------------------------------------------===// |
1339 | |
1340 | /// This instruction compares its operands according to the predicate given |
1341 | /// to the constructor. It only operates on floating point values or packed |
1342 | /// vectors of floating point values. The operands must be identical types. |
1343 | /// Represents a floating point comparison operator. |
1344 | class FCmpInst: public CmpInst { |
1345 | void AssertOK() { |
1346 | assert(isFPPredicate() && "Invalid FCmp predicate value")((isFPPredicate() && "Invalid FCmp predicate value") ? static_cast<void> (0) : __assert_fail ("isFPPredicate() && \"Invalid FCmp predicate value\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1346, __PRETTY_FUNCTION__)); |
1347 | assert(getOperand(0)->getType() == getOperand(1)->getType() &&((getOperand(0)->getType() == getOperand(1)->getType() && "Both operands to FCmp instruction are not of the same type!" ) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to FCmp instruction are not of the same type!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1348, __PRETTY_FUNCTION__)) |
1348 | "Both operands to FCmp instruction are not of the same type!")((getOperand(0)->getType() == getOperand(1)->getType() && "Both operands to FCmp instruction are not of the same type!" ) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to FCmp instruction are not of the same type!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1348, __PRETTY_FUNCTION__)); |
1349 | // Check that the operands are the right type |
1350 | assert(getOperand(0)->getType()->isFPOrFPVectorTy() &&((getOperand(0)->getType()->isFPOrFPVectorTy() && "Invalid operand types for FCmp instruction") ? static_cast< void> (0) : __assert_fail ("getOperand(0)->getType()->isFPOrFPVectorTy() && \"Invalid operand types for FCmp instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1351, __PRETTY_FUNCTION__)) |
1351 | "Invalid operand types for FCmp instruction")((getOperand(0)->getType()->isFPOrFPVectorTy() && "Invalid operand types for FCmp instruction") ? static_cast< void> (0) : __assert_fail ("getOperand(0)->getType()->isFPOrFPVectorTy() && \"Invalid operand types for FCmp instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1351, __PRETTY_FUNCTION__)); |
1352 | } |
1353 | |
1354 | protected: |
1355 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1356 | friend class Instruction; |
1357 | |
1358 | /// Clone an identical FCmpInst |
1359 | FCmpInst *cloneImpl() const; |
1360 | |
1361 | public: |
1362 | /// Constructor with insert-before-instruction semantics. |
1363 | FCmpInst( |
1364 | Instruction *InsertBefore, ///< Where to insert |
1365 | Predicate pred, ///< The predicate to use for the comparison |
1366 | Value *LHS, ///< The left-hand-side of the expression |
1367 | Value *RHS, ///< The right-hand-side of the expression |
1368 | const Twine &NameStr = "" ///< Name of the instruction |
1369 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1370 | Instruction::FCmp, pred, LHS, RHS, NameStr, |
1371 | InsertBefore) { |
1372 | AssertOK(); |
1373 | } |
1374 | |
1375 | /// Constructor with insert-at-end semantics. |
1376 | FCmpInst( |
1377 | BasicBlock &InsertAtEnd, ///< Block to insert into. |
1378 | Predicate pred, ///< The predicate to use for the comparison |
1379 | Value *LHS, ///< The left-hand-side of the expression |
1380 | Value *RHS, ///< The right-hand-side of the expression |
1381 | const Twine &NameStr = "" ///< Name of the instruction |
1382 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1383 | Instruction::FCmp, pred, LHS, RHS, NameStr, |
1384 | &InsertAtEnd) { |
1385 | AssertOK(); |
1386 | } |
1387 | |
1388 | /// Constructor with no-insertion semantics |
1389 | FCmpInst( |
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 | Instruction *FlagsSource = nullptr |
1395 | ) : CmpInst(makeCmpResultType(LHS->getType()), Instruction::FCmp, Pred, LHS, |
1396 | RHS, NameStr, nullptr, FlagsSource) { |
1397 | AssertOK(); |
1398 | } |
1399 | |
1400 | /// @returns true if the predicate of this instruction is EQ or NE. |
1401 | /// Determine if this is an equality predicate. |
1402 | static bool isEquality(Predicate Pred) { |
1403 | return Pred == FCMP_OEQ || Pred == FCMP_ONE || Pred == FCMP_UEQ || |
1404 | Pred == FCMP_UNE; |
1405 | } |
1406 | |
1407 | /// @returns true if the predicate of this instruction is EQ or NE. |
1408 | /// Determine if this is an equality predicate. |
1409 | bool isEquality() const { return isEquality(getPredicate()); } |
1410 | |
1411 | /// @returns true if the predicate of this instruction is commutative. |
1412 | /// Determine if this is a commutative predicate. |
1413 | bool isCommutative() const { |
1414 | return isEquality() || |
1415 | getPredicate() == FCMP_FALSE || |
1416 | getPredicate() == FCMP_TRUE || |
1417 | getPredicate() == FCMP_ORD || |
1418 | getPredicate() == FCMP_UNO; |
1419 | } |
1420 | |
1421 | /// @returns true if the predicate is relational (not EQ or NE). |
1422 | /// Determine if this a relational predicate. |
1423 | bool isRelational() const { return !isEquality(); } |
1424 | |
1425 | /// Exchange the two operands to this instruction in such a way that it does |
1426 | /// not modify the semantics of the instruction. The predicate value may be |
1427 | /// changed to retain the same result if the predicate is order dependent |
1428 | /// (e.g. ult). |
1429 | /// Swap operands and adjust predicate. |
1430 | void swapOperands() { |
1431 | setPredicate(getSwappedPredicate()); |
1432 | Op<0>().swap(Op<1>()); |
1433 | } |
1434 | |
1435 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
1436 | static bool classof(const Instruction *I) { |
1437 | return I->getOpcode() == Instruction::FCmp; |
1438 | } |
1439 | static bool classof(const Value *V) { |
1440 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1441 | } |
1442 | }; |
1443 | |
1444 | //===----------------------------------------------------------------------===// |
1445 | /// This class represents a function call, abstracting a target |
1446 | /// machine's calling convention. This class uses low bit of the SubClassData |
1447 | /// field to indicate whether or not this is a tail call. The rest of the bits |
1448 | /// hold the calling convention of the call. |
1449 | /// |
1450 | class CallInst : public CallBase { |
1451 | CallInst(const CallInst &CI); |
1452 | |
1453 | /// Construct a CallInst given a range of arguments. |
1454 | /// Construct a CallInst from a range of arguments |
1455 | inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1456 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1457 | Instruction *InsertBefore); |
1458 | |
1459 | inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1460 | const Twine &NameStr, Instruction *InsertBefore) |
1461 | : CallInst(Ty, Func, Args, None, NameStr, InsertBefore) {} |
1462 | |
1463 | /// Construct a CallInst given a range of arguments. |
1464 | /// Construct a CallInst from a range of arguments |
1465 | inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1466 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1467 | BasicBlock *InsertAtEnd); |
1468 | |
1469 | explicit CallInst(FunctionType *Ty, Value *F, const Twine &NameStr, |
1470 | Instruction *InsertBefore); |
1471 | |
1472 | CallInst(FunctionType *ty, Value *F, const Twine &NameStr, |
1473 | BasicBlock *InsertAtEnd); |
1474 | |
1475 | void init(FunctionType *FTy, Value *Func, ArrayRef<Value *> Args, |
1476 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr); |
1477 | void init(FunctionType *FTy, Value *Func, const Twine &NameStr); |
1478 | |
1479 | /// Compute the number of operands to allocate. |
1480 | static int ComputeNumOperands(int NumArgs, int NumBundleInputs = 0) { |
1481 | // We need one operand for the called function, plus the input operand |
1482 | // counts provided. |
1483 | return 1 + NumArgs + NumBundleInputs; |
1484 | } |
1485 | |
1486 | protected: |
1487 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1488 | friend class Instruction; |
1489 | |
1490 | CallInst *cloneImpl() const; |
1491 | |
1492 | public: |
1493 | static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr = "", |
1494 | Instruction *InsertBefore = nullptr) { |
1495 | return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertBefore); |
1496 | } |
1497 | |
1498 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1499 | const Twine &NameStr, |
1500 | Instruction *InsertBefore = nullptr) { |
1501 | return new (ComputeNumOperands(Args.size())) |
1502 | CallInst(Ty, Func, Args, None, NameStr, InsertBefore); |
1503 | } |
1504 | |
1505 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1506 | ArrayRef<OperandBundleDef> Bundles = None, |
1507 | const Twine &NameStr = "", |
1508 | Instruction *InsertBefore = nullptr) { |
1509 | const int NumOperands = |
1510 | ComputeNumOperands(Args.size(), CountBundleInputs(Bundles)); |
1511 | const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
1512 | |
1513 | return new (NumOperands, DescriptorBytes) |
1514 | CallInst(Ty, Func, Args, Bundles, NameStr, InsertBefore); |
1515 | } |
1516 | |
1517 | static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr, |
1518 | BasicBlock *InsertAtEnd) { |
1519 | return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertAtEnd); |
1520 | } |
1521 | |
1522 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1523 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
1524 | return new (ComputeNumOperands(Args.size())) |
1525 | CallInst(Ty, Func, Args, None, NameStr, InsertAtEnd); |
1526 | } |
1527 | |
1528 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1529 | ArrayRef<OperandBundleDef> Bundles, |
1530 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
1531 | const int NumOperands = |
1532 | ComputeNumOperands(Args.size(), CountBundleInputs(Bundles)); |
1533 | const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
1534 | |
1535 | return new (NumOperands, DescriptorBytes) |
1536 | CallInst(Ty, Func, Args, Bundles, NameStr, InsertAtEnd); |
1537 | } |
1538 | |
1539 | static CallInst *Create(FunctionCallee Func, const Twine &NameStr = "", |
1540 | Instruction *InsertBefore = nullptr) { |
1541 | return Create(Func.getFunctionType(), Func.getCallee(), NameStr, |
1542 | InsertBefore); |
1543 | } |
1544 | |
1545 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
1546 | ArrayRef<OperandBundleDef> Bundles = None, |
1547 | const Twine &NameStr = "", |
1548 | Instruction *InsertBefore = nullptr) { |
1549 | return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles, |
1550 | NameStr, InsertBefore); |
1551 | } |
1552 | |
1553 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
1554 | const Twine &NameStr, |
1555 | Instruction *InsertBefore = nullptr) { |
1556 | return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr, |
1557 | InsertBefore); |
1558 | } |
1559 | |
1560 | static CallInst *Create(FunctionCallee Func, const Twine &NameStr, |
1561 | BasicBlock *InsertAtEnd) { |
1562 | return Create(Func.getFunctionType(), Func.getCallee(), NameStr, |
1563 | InsertAtEnd); |
1564 | } |
1565 | |
1566 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
1567 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
1568 | return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr, |
1569 | InsertAtEnd); |
1570 | } |
1571 | |
1572 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
1573 | ArrayRef<OperandBundleDef> Bundles, |
1574 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
1575 | return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles, |
1576 | NameStr, InsertAtEnd); |
1577 | } |
1578 | |
1579 | /// Create a clone of \p CI with a different set of operand bundles and |
1580 | /// insert it before \p InsertPt. |
1581 | /// |
1582 | /// The returned call instruction is identical \p CI in every way except that |
1583 | /// the operand bundles for the new instruction are set to the operand bundles |
1584 | /// in \p Bundles. |
1585 | static CallInst *Create(CallInst *CI, ArrayRef<OperandBundleDef> Bundles, |
1586 | Instruction *InsertPt = nullptr); |
1587 | |
1588 | /// Generate the IR for a call to malloc: |
1589 | /// 1. Compute the malloc call's argument as the specified type's size, |
1590 | /// possibly multiplied by the array size if the array size is not |
1591 | /// constant 1. |
1592 | /// 2. Call malloc with that argument. |
1593 | /// 3. Bitcast the result of the malloc call to the specified type. |
1594 | static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy, |
1595 | Type *AllocTy, Value *AllocSize, |
1596 | Value *ArraySize = nullptr, |
1597 | Function *MallocF = nullptr, |
1598 | const Twine &Name = ""); |
1599 | static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy, |
1600 | Type *AllocTy, Value *AllocSize, |
1601 | Value *ArraySize = nullptr, |
1602 | Function *MallocF = nullptr, |
1603 | const Twine &Name = ""); |
1604 | static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy, |
1605 | Type *AllocTy, Value *AllocSize, |
1606 | Value *ArraySize = nullptr, |
1607 | ArrayRef<OperandBundleDef> Bundles = None, |
1608 | Function *MallocF = nullptr, |
1609 | const Twine &Name = ""); |
1610 | static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy, |
1611 | Type *AllocTy, Value *AllocSize, |
1612 | Value *ArraySize = nullptr, |
1613 | ArrayRef<OperandBundleDef> Bundles = None, |
1614 | Function *MallocF = nullptr, |
1615 | const Twine &Name = ""); |
1616 | /// Generate the IR for a call to the builtin free function. |
1617 | static Instruction *CreateFree(Value *Source, Instruction *InsertBefore); |
1618 | static Instruction *CreateFree(Value *Source, BasicBlock *InsertAtEnd); |
1619 | static Instruction *CreateFree(Value *Source, |
1620 | ArrayRef<OperandBundleDef> Bundles, |
1621 | Instruction *InsertBefore); |
1622 | static Instruction *CreateFree(Value *Source, |
1623 | ArrayRef<OperandBundleDef> Bundles, |
1624 | BasicBlock *InsertAtEnd); |
1625 | |
1626 | // Note that 'musttail' implies 'tail'. |
1627 | enum TailCallKind : unsigned { |
1628 | TCK_None = 0, |
1629 | TCK_Tail = 1, |
1630 | TCK_MustTail = 2, |
1631 | TCK_NoTail = 3, |
1632 | TCK_LAST = TCK_NoTail |
1633 | }; |
1634 | |
1635 | using TailCallKindField = Bitfield::Element<TailCallKind, 0, 2, TCK_LAST>; |
1636 | static_assert( |
1637 | Bitfield::areContiguous<TailCallKindField, CallBase::CallingConvField>(), |
1638 | "Bitfields must be contiguous"); |
1639 | |
1640 | TailCallKind getTailCallKind() const { |
1641 | return getSubclassData<TailCallKindField>(); |
1642 | } |
1643 | |
1644 | bool isTailCall() const { |
1645 | TailCallKind Kind = getTailCallKind(); |
1646 | return Kind == TCK_Tail || Kind == TCK_MustTail; |
1647 | } |
1648 | |
1649 | bool isMustTailCall() const { return getTailCallKind() == TCK_MustTail; } |
1650 | |
1651 | bool isNoTailCall() const { return getTailCallKind() == TCK_NoTail; } |
1652 | |
1653 | void setTailCallKind(TailCallKind TCK) { |
1654 | setSubclassData<TailCallKindField>(TCK); |
1655 | } |
1656 | |
1657 | void setTailCall(bool IsTc = true) { |
1658 | setTailCallKind(IsTc ? TCK_Tail : TCK_None); |
1659 | } |
1660 | |
1661 | /// Return true if the call can return twice |
1662 | bool canReturnTwice() const { return hasFnAttr(Attribute::ReturnsTwice); } |
1663 | void setCanReturnTwice() { |
1664 | addAttribute(AttributeList::FunctionIndex, Attribute::ReturnsTwice); |
1665 | } |
1666 | |
1667 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1668 | static bool classof(const Instruction *I) { |
1669 | return I->getOpcode() == Instruction::Call; |
1670 | } |
1671 | static bool classof(const Value *V) { |
1672 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1673 | } |
1674 | |
1675 | /// Updates profile metadata by scaling it by \p S / \p T. |
1676 | void updateProfWeight(uint64_t S, uint64_t T); |
1677 | |
1678 | private: |
1679 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
1680 | // method so that subclasses cannot accidentally use it. |
1681 | template <typename Bitfield> |
1682 | void setSubclassData(typename Bitfield::Type Value) { |
1683 | Instruction::setSubclassData<Bitfield>(Value); |
1684 | } |
1685 | }; |
1686 | |
1687 | CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1688 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1689 | BasicBlock *InsertAtEnd) |
1690 | : CallBase(Ty->getReturnType(), Instruction::Call, |
1691 | OperandTraits<CallBase>::op_end(this) - |
1692 | (Args.size() + CountBundleInputs(Bundles) + 1), |
1693 | unsigned(Args.size() + CountBundleInputs(Bundles) + 1), |
1694 | InsertAtEnd) { |
1695 | init(Ty, Func, Args, Bundles, NameStr); |
1696 | } |
1697 | |
1698 | CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1699 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1700 | Instruction *InsertBefore) |
1701 | : CallBase(Ty->getReturnType(), Instruction::Call, |
1702 | OperandTraits<CallBase>::op_end(this) - |
1703 | (Args.size() + CountBundleInputs(Bundles) + 1), |
1704 | unsigned(Args.size() + CountBundleInputs(Bundles) + 1), |
1705 | InsertBefore) { |
1706 | init(Ty, Func, Args, Bundles, NameStr); |
1707 | } |
1708 | |
1709 | //===----------------------------------------------------------------------===// |
1710 | // SelectInst Class |
1711 | //===----------------------------------------------------------------------===// |
1712 | |
1713 | /// This class represents the LLVM 'select' instruction. |
1714 | /// |
1715 | class SelectInst : public Instruction { |
1716 | SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr, |
1717 | Instruction *InsertBefore) |
1718 | : Instruction(S1->getType(), Instruction::Select, |
1719 | &Op<0>(), 3, InsertBefore) { |
1720 | init(C, S1, S2); |
1721 | setName(NameStr); |
1722 | } |
1723 | |
1724 | SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr, |
1725 | BasicBlock *InsertAtEnd) |
1726 | : Instruction(S1->getType(), Instruction::Select, |
1727 | &Op<0>(), 3, InsertAtEnd) { |
1728 | init(C, S1, S2); |
1729 | setName(NameStr); |
1730 | } |
1731 | |
1732 | void init(Value *C, Value *S1, Value *S2) { |
1733 | assert(!areInvalidOperands(C, S1, S2) && "Invalid operands for select")((!areInvalidOperands(C, S1, S2) && "Invalid operands for select" ) ? static_cast<void> (0) : __assert_fail ("!areInvalidOperands(C, S1, S2) && \"Invalid operands for select\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1733, __PRETTY_FUNCTION__)); |
1734 | Op<0>() = C; |
1735 | Op<1>() = S1; |
1736 | Op<2>() = S2; |
1737 | } |
1738 | |
1739 | protected: |
1740 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1741 | friend class Instruction; |
1742 | |
1743 | SelectInst *cloneImpl() const; |
1744 | |
1745 | public: |
1746 | static SelectInst *Create(Value *C, Value *S1, Value *S2, |
1747 | const Twine &NameStr = "", |
1748 | Instruction *InsertBefore = nullptr, |
1749 | Instruction *MDFrom = nullptr) { |
1750 | SelectInst *Sel = new(3) SelectInst(C, S1, S2, NameStr, InsertBefore); |
1751 | if (MDFrom) |
1752 | Sel->copyMetadata(*MDFrom); |
1753 | return Sel; |
1754 | } |
1755 | |
1756 | static SelectInst *Create(Value *C, Value *S1, Value *S2, |
1757 | const Twine &NameStr, |
1758 | BasicBlock *InsertAtEnd) { |
1759 | return new(3) SelectInst(C, S1, S2, NameStr, InsertAtEnd); |
1760 | } |
1761 | |
1762 | const Value *getCondition() const { return Op<0>(); } |
1763 | const Value *getTrueValue() const { return Op<1>(); } |
1764 | const Value *getFalseValue() const { return Op<2>(); } |
1765 | Value *getCondition() { return Op<0>(); } |
1766 | Value *getTrueValue() { return Op<1>(); } |
1767 | Value *getFalseValue() { return Op<2>(); } |
1768 | |
1769 | void setCondition(Value *V) { Op<0>() = V; } |
1770 | void setTrueValue(Value *V) { Op<1>() = V; } |
1771 | void setFalseValue(Value *V) { Op<2>() = V; } |
1772 | |
1773 | /// Swap the true and false values of the select instruction. |
1774 | /// This doesn't swap prof metadata. |
1775 | void swapValues() { Op<1>().swap(Op<2>()); } |
1776 | |
1777 | /// Return a string if the specified operands are invalid |
1778 | /// for a select operation, otherwise return null. |
1779 | static const char *areInvalidOperands(Value *Cond, Value *True, Value *False); |
1780 | |
1781 | /// Transparently provide more efficient getOperand methods. |
1782 | 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; |
1783 | |
1784 | OtherOps getOpcode() const { |
1785 | return static_cast<OtherOps>(Instruction::getOpcode()); |
1786 | } |
1787 | |
1788 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1789 | static bool classof(const Instruction *I) { |
1790 | return I->getOpcode() == Instruction::Select; |
1791 | } |
1792 | static bool classof(const Value *V) { |
1793 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1794 | } |
1795 | }; |
1796 | |
1797 | template <> |
1798 | struct OperandTraits<SelectInst> : public FixedNumOperandTraits<SelectInst, 3> { |
1799 | }; |
1800 | |
1801 | 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 { ((i_nocapture < OperandTraits<SelectInst>::operands(this) && "getOperand() out of range!" ) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<SelectInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1801, __PRETTY_FUNCTION__)); 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) { ((i_nocapture < OperandTraits<SelectInst>::operands(this) && "setOperand() out of range!" ) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<SelectInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1801, __PRETTY_FUNCTION__)); 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); } |
1802 | |
1803 | //===----------------------------------------------------------------------===// |
1804 | // VAArgInst Class |
1805 | //===----------------------------------------------------------------------===// |
1806 | |
1807 | /// This class represents the va_arg llvm instruction, which returns |
1808 | /// an argument of the specified type given a va_list and increments that list |
1809 | /// |
1810 | class VAArgInst : public UnaryInstruction { |
1811 | protected: |
1812 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1813 | friend class Instruction; |
1814 | |
1815 | VAArgInst *cloneImpl() const; |
1816 | |
1817 | public: |
1818 | VAArgInst(Value *List, Type *Ty, const Twine &NameStr = "", |
1819 | Instruction *InsertBefore = nullptr) |
1820 | : UnaryInstruction(Ty, VAArg, List, InsertBefore) { |
1821 | setName(NameStr); |
1822 | } |
1823 | |
1824 | VAArgInst(Value *List, Type *Ty, const Twine &NameStr, |
1825 | BasicBlock *InsertAtEnd) |
1826 | : UnaryInstruction(Ty, VAArg, List, InsertAtEnd) { |
1827 | setName(NameStr); |
1828 | } |
1829 | |
1830 | Value *getPointerOperand() { return getOperand(0); } |
1831 | const Value *getPointerOperand() const { return getOperand(0); } |
1832 | static unsigned getPointerOperandIndex() { return 0U; } |
1833 | |
1834 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1835 | static bool classof(const Instruction *I) { |
1836 | return I->getOpcode() == VAArg; |
1837 | } |
1838 | static bool classof(const Value *V) { |
1839 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1840 | } |
1841 | }; |
1842 | |
1843 | //===----------------------------------------------------------------------===// |
1844 | // ExtractElementInst Class |
1845 | //===----------------------------------------------------------------------===// |
1846 | |
1847 | /// This instruction extracts a single (scalar) |
1848 | /// element from a VectorType value |
1849 | /// |
1850 | class ExtractElementInst : public Instruction { |
1851 | ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr = "", |
1852 | Instruction *InsertBefore = nullptr); |
1853 | ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr, |
1854 | BasicBlock *InsertAtEnd); |
1855 | |
1856 | protected: |
1857 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1858 | friend class Instruction; |
1859 | |
1860 | ExtractElementInst *cloneImpl() const; |
1861 | |
1862 | public: |
1863 | static ExtractElementInst *Create(Value *Vec, Value *Idx, |
1864 | const Twine &NameStr = "", |
1865 | Instruction *InsertBefore = nullptr) { |
1866 | return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertBefore); |
1867 | } |
1868 | |
1869 | static ExtractElementInst *Create(Value *Vec, Value *Idx, |
1870 | const Twine &NameStr, |
1871 | BasicBlock *InsertAtEnd) { |
1872 | return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertAtEnd); |
1873 | } |
1874 | |
1875 | /// Return true if an extractelement instruction can be |
1876 | /// formed with the specified operands. |
1877 | static bool isValidOperands(const Value *Vec, const Value *Idx); |
1878 | |
1879 | Value *getVectorOperand() { return Op<0>(); } |
1880 | Value *getIndexOperand() { return Op<1>(); } |
1881 | const Value *getVectorOperand() const { return Op<0>(); } |
1882 | const Value *getIndexOperand() const { return Op<1>(); } |
1883 | |
1884 | VectorType *getVectorOperandType() const { |
1885 | return cast<VectorType>(getVectorOperand()->getType()); |
1886 | } |
1887 | |
1888 | /// Transparently provide more efficient getOperand methods. |
1889 | 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; |
1890 | |
1891 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1892 | static bool classof(const Instruction *I) { |
1893 | return I->getOpcode() == Instruction::ExtractElement; |
1894 | } |
1895 | static bool classof(const Value *V) { |
1896 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1897 | } |
1898 | }; |
1899 | |
1900 | template <> |
1901 | struct OperandTraits<ExtractElementInst> : |
1902 | public FixedNumOperandTraits<ExtractElementInst, 2> { |
1903 | }; |
1904 | |
1905 | 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 { ((i_nocapture < OperandTraits<ExtractElementInst>:: operands(this) && "getOperand() out of range!") ? static_cast <void> (0) : __assert_fail ("i_nocapture < OperandTraits<ExtractElementInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1905, __PRETTY_FUNCTION__)); 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) { ((i_nocapture < OperandTraits<ExtractElementInst >::operands(this) && "setOperand() out of range!") ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ExtractElementInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1905, __PRETTY_FUNCTION__)); 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); } |
1906 | |
1907 | //===----------------------------------------------------------------------===// |
1908 | // InsertElementInst Class |
1909 | //===----------------------------------------------------------------------===// |
1910 | |
1911 | /// This instruction inserts a single (scalar) |
1912 | /// element into a VectorType value |
1913 | /// |
1914 | class InsertElementInst : public Instruction { |
1915 | InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, |
1916 | const Twine &NameStr = "", |
1917 | Instruction *InsertBefore = nullptr); |
1918 | InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, const Twine &NameStr, |
1919 | BasicBlock *InsertAtEnd); |
1920 | |
1921 | protected: |
1922 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1923 | friend class Instruction; |
1924 | |
1925 | InsertElementInst *cloneImpl() const; |
1926 | |
1927 | public: |
1928 | static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx, |
1929 | const Twine &NameStr = "", |
1930 | Instruction *InsertBefore = nullptr) { |
1931 | return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertBefore); |
1932 | } |
1933 | |
1934 | static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx, |
1935 | const Twine &NameStr, |
1936 | BasicBlock *InsertAtEnd) { |
1937 | return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertAtEnd); |
1938 | } |
1939 | |
1940 | /// Return true if an insertelement instruction can be |
1941 | /// formed with the specified operands. |
1942 | static bool isValidOperands(const Value *Vec, const Value *NewElt, |
1943 | const Value *Idx); |
1944 | |
1945 | /// Overload to return most specific vector type. |
1946 | /// |
1947 | VectorType *getType() const { |
1948 | return cast<VectorType>(Instruction::getType()); |
1949 | } |
1950 | |
1951 | /// Transparently provide more efficient getOperand methods. |
1952 | 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; |
1953 | |
1954 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1955 | static bool classof(const Instruction *I) { |
1956 | return I->getOpcode() == Instruction::InsertElement; |
1957 | } |
1958 | static bool classof(const Value *V) { |
1959 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1960 | } |
1961 | }; |
1962 | |
1963 | template <> |
1964 | struct OperandTraits<InsertElementInst> : |
1965 | public FixedNumOperandTraits<InsertElementInst, 3> { |
1966 | }; |
1967 | |
1968 | 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 { ((i_nocapture < OperandTraits<InsertElementInst >::operands(this) && "getOperand() out of range!") ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<InsertElementInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1968, __PRETTY_FUNCTION__)); 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 ) { ((i_nocapture < OperandTraits<InsertElementInst> ::operands(this) && "setOperand() out of range!") ? static_cast <void> (0) : __assert_fail ("i_nocapture < OperandTraits<InsertElementInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 1968, __PRETTY_FUNCTION__)); 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); } |
1969 | |
1970 | //===----------------------------------------------------------------------===// |
1971 | // ShuffleVectorInst Class |
1972 | //===----------------------------------------------------------------------===// |
1973 | |
1974 | constexpr int UndefMaskElem = -1; |
1975 | |
1976 | /// This instruction constructs a fixed permutation of two |
1977 | /// input vectors. |
1978 | /// |
1979 | /// For each element of the result vector, the shuffle mask selects an element |
1980 | /// from one of the input vectors to copy to the result. Non-negative elements |
1981 | /// in the mask represent an index into the concatenated pair of input vectors. |
1982 | /// UndefMaskElem (-1) specifies that the result element is undefined. |
1983 | /// |
1984 | /// For scalable vectors, all the elements of the mask must be 0 or -1. This |
1985 | /// requirement may be relaxed in the future. |
1986 | class ShuffleVectorInst : public Instruction { |
1987 | SmallVector<int, 4> ShuffleMask; |
1988 | Constant *ShuffleMaskForBitcode; |
1989 | |
1990 | protected: |
1991 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1992 | friend class Instruction; |
1993 | |
1994 | ShuffleVectorInst *cloneImpl() const; |
1995 | |
1996 | public: |
1997 | ShuffleVectorInst(Value *V1, Value *V2, Value *Mask, |
1998 | const Twine &NameStr = "", |
1999 | Instruction *InsertBefor = nullptr); |
2000 | ShuffleVectorInst(Value *V1, Value *V2, Value *Mask, |
2001 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2002 | ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask, |
2003 | const Twine &NameStr = "", |
2004 | Instruction *InsertBefor = nullptr); |
2005 | ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask, |
2006 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2007 | |
2008 | void *operator new(size_t s) { return User::operator new(s, 2); } |
2009 | |
2010 | /// Swap the operands and adjust the mask to preserve the semantics |
2011 | /// of the instruction. |
2012 | void commute(); |
2013 | |
2014 | /// Return true if a shufflevector instruction can be |
2015 | /// formed with the specified operands. |
2016 | static bool isValidOperands(const Value *V1, const Value *V2, |
2017 | const Value *Mask); |
2018 | static bool isValidOperands(const Value *V1, const Value *V2, |
2019 | ArrayRef<int> Mask); |
2020 | |
2021 | /// Overload to return most specific vector type. |
2022 | /// |
2023 | VectorType *getType() const { |
2024 | return cast<VectorType>(Instruction::getType()); |
2025 | } |
2026 | |
2027 | /// Transparently provide more efficient getOperand methods. |
2028 | 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; |
2029 | |
2030 | /// Return the shuffle mask value of this instruction for the given element |
2031 | /// index. Return UndefMaskElem if the element is undef. |
2032 | int getMaskValue(unsigned Elt) const { return ShuffleMask[Elt]; } |
2033 | |
2034 | /// Convert the input shuffle mask operand to a vector of integers. Undefined |
2035 | /// elements of the mask are returned as UndefMaskElem. |
2036 | static void getShuffleMask(const Constant *Mask, |
2037 | SmallVectorImpl<int> &Result); |
2038 | |
2039 | /// Return the mask for this instruction as a vector of integers. Undefined |
2040 | /// elements of the mask are returned as UndefMaskElem. |
2041 | void getShuffleMask(SmallVectorImpl<int> &Result) const { |
2042 | Result.assign(ShuffleMask.begin(), ShuffleMask.end()); |
2043 | } |
2044 | |
2045 | /// Return the mask for this instruction, for use in bitcode. |
2046 | /// |
2047 | /// TODO: This is temporary until we decide a new bitcode encoding for |
2048 | /// shufflevector. |
2049 | Constant *getShuffleMaskForBitcode() const { return ShuffleMaskForBitcode; } |
2050 | |
2051 | static Constant *convertShuffleMaskForBitcode(ArrayRef<int> Mask, |
2052 | Type *ResultTy); |
2053 | |
2054 | void setShuffleMask(ArrayRef<int> Mask); |
2055 | |
2056 | ArrayRef<int> getShuffleMask() const { return ShuffleMask; } |
2057 | |
2058 | /// Return true if this shuffle returns a vector with a different number of |
2059 | /// elements than its source vectors. |
2060 | /// Examples: shufflevector <4 x n> A, <4 x n> B, <1,2,3> |
2061 | /// shufflevector <4 x n> A, <4 x n> B, <1,2,3,4,5> |
2062 | bool changesLength() const { |
2063 | unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType()) |
2064 | ->getElementCount() |
2065 | .getKnownMinValue(); |
2066 | unsigned NumMaskElts = ShuffleMask.size(); |
2067 | return NumSourceElts != NumMaskElts; |
2068 | } |
2069 | |
2070 | /// Return true if this shuffle returns a vector with a greater number of |
2071 | /// elements than its source vectors. |
2072 | /// Example: shufflevector <2 x n> A, <2 x n> B, <1,2,3> |
2073 | bool increasesLength() const { |
2074 | unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType()) |
2075 | ->getElementCount() |
2076 | .getKnownMinValue(); |
2077 | unsigned NumMaskElts = ShuffleMask.size(); |
2078 | return NumSourceElts < NumMaskElts; |
2079 | } |
2080 | |
2081 | /// Return true if this shuffle mask chooses elements from exactly one source |
2082 | /// vector. |
2083 | /// Example: <7,5,undef,7> |
2084 | /// This assumes that vector operands are the same length as the mask. |
2085 | static bool isSingleSourceMask(ArrayRef<int> Mask); |
2086 | static bool isSingleSourceMask(const Constant *Mask) { |
2087 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant." ) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2087, __PRETTY_FUNCTION__)); |
2088 | SmallVector<int, 16> MaskAsInts; |
2089 | getShuffleMask(Mask, MaskAsInts); |
2090 | return isSingleSourceMask(MaskAsInts); |
2091 | } |
2092 | |
2093 | /// Return true if this shuffle chooses elements from exactly one source |
2094 | /// vector without changing the length of that vector. |
2095 | /// Example: shufflevector <4 x n> A, <4 x n> B, <3,0,undef,3> |
2096 | /// TODO: Optionally allow length-changing shuffles. |
2097 | bool isSingleSource() const { |
2098 | return !changesLength() && isSingleSourceMask(ShuffleMask); |
2099 | } |
2100 | |
2101 | /// Return true if this shuffle mask chooses elements from exactly one source |
2102 | /// vector without lane crossings. A shuffle using this mask is not |
2103 | /// necessarily a no-op because it may change the number of elements from its |
2104 | /// input vectors or it may provide demanded bits knowledge via undef lanes. |
2105 | /// Example: <undef,undef,2,3> |
2106 | static bool isIdentityMask(ArrayRef<int> Mask); |
2107 | static bool isIdentityMask(const Constant *Mask) { |
2108 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant." ) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2108, __PRETTY_FUNCTION__)); |
2109 | SmallVector<int, 16> MaskAsInts; |
2110 | getShuffleMask(Mask, MaskAsInts); |
2111 | return isIdentityMask(MaskAsInts); |
2112 | } |
2113 | |
2114 | /// Return true if this shuffle chooses elements from exactly one source |
2115 | /// vector without lane crossings and does not change the number of elements |
2116 | /// from its input vectors. |
2117 | /// Example: shufflevector <4 x n> A, <4 x n> B, <4,undef,6,undef> |
2118 | bool isIdentity() const { |
2119 | return !changesLength() && isIdentityMask(ShuffleMask); |
2120 | } |
2121 | |
2122 | /// Return true if this shuffle lengthens exactly one source vector with |
2123 | /// undefs in the high elements. |
2124 | bool isIdentityWithPadding() const; |
2125 | |
2126 | /// Return true if this shuffle extracts the first N elements of exactly one |
2127 | /// source vector. |
2128 | bool isIdentityWithExtract() const; |
2129 | |
2130 | /// Return true if this shuffle concatenates its 2 source vectors. This |
2131 | /// returns false if either input is undefined. In that case, the shuffle is |
2132 | /// is better classified as an identity with padding operation. |
2133 | bool isConcat() const; |
2134 | |
2135 | /// Return true if this shuffle mask chooses elements from its source vectors |
2136 | /// without lane crossings. A shuffle using this mask would be |
2137 | /// equivalent to a vector select with a constant condition operand. |
2138 | /// Example: <4,1,6,undef> |
2139 | /// This returns false if the mask does not choose from both input vectors. |
2140 | /// In that case, the shuffle is better classified as an identity shuffle. |
2141 | /// This assumes that vector operands are the same length as the mask |
2142 | /// (a length-changing shuffle can never be equivalent to a vector select). |
2143 | static bool isSelectMask(ArrayRef<int> Mask); |
2144 | static bool isSelectMask(const Constant *Mask) { |
2145 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant." ) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2145, __PRETTY_FUNCTION__)); |
2146 | SmallVector<int, 16> MaskAsInts; |
2147 | getShuffleMask(Mask, MaskAsInts); |
2148 | return isSelectMask(MaskAsInts); |
2149 | } |
2150 | |
2151 | /// Return true if this shuffle chooses elements from its source vectors |
2152 | /// without lane crossings and all operands have the same number of elements. |
2153 | /// In other words, this shuffle is equivalent to a vector select with a |
2154 | /// constant condition operand. |
2155 | /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,1,6,3> |
2156 | /// This returns false if the mask does not choose from both input vectors. |
2157 | /// In that case, the shuffle is better classified as an identity shuffle. |
2158 | /// TODO: Optionally allow length-changing shuffles. |
2159 | bool isSelect() const { |
2160 | return !changesLength() && isSelectMask(ShuffleMask); |
2161 | } |
2162 | |
2163 | /// Return true if this shuffle mask swaps the order of elements from exactly |
2164 | /// one source vector. |
2165 | /// Example: <7,6,undef,4> |
2166 | /// This assumes that vector operands are the same length as the mask. |
2167 | static bool isReverseMask(ArrayRef<int> Mask); |
2168 | static bool isReverseMask(const Constant *Mask) { |
2169 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant." ) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2169, __PRETTY_FUNCTION__)); |
2170 | SmallVector<int, 16> MaskAsInts; |
2171 | getShuffleMask(Mask, MaskAsInts); |
2172 | return isReverseMask(MaskAsInts); |
2173 | } |
2174 | |
2175 | /// Return true if this shuffle swaps the order of elements from exactly |
2176 | /// one source vector. |
2177 | /// Example: shufflevector <4 x n> A, <4 x n> B, <3,undef,1,undef> |
2178 | /// TODO: Optionally allow length-changing shuffles. |
2179 | bool isReverse() const { |
2180 | return !changesLength() && isReverseMask(ShuffleMask); |
2181 | } |
2182 | |
2183 | /// Return true if this shuffle mask chooses all elements with the same value |
2184 | /// as the first element of exactly one source vector. |
2185 | /// Example: <4,undef,undef,4> |
2186 | /// This assumes that vector operands are the same length as the mask. |
2187 | static bool isZeroEltSplatMask(ArrayRef<int> Mask); |
2188 | static bool isZeroEltSplatMask(const Constant *Mask) { |
2189 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant." ) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2189, __PRETTY_FUNCTION__)); |
2190 | SmallVector<int, 16> MaskAsInts; |
2191 | getShuffleMask(Mask, MaskAsInts); |
2192 | return isZeroEltSplatMask(MaskAsInts); |
2193 | } |
2194 | |
2195 | /// Return true if all elements of this shuffle are the same value as the |
2196 | /// first element of exactly one source vector without changing the length |
2197 | /// of that vector. |
2198 | /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,0,undef,0> |
2199 | /// TODO: Optionally allow length-changing shuffles. |
2200 | /// TODO: Optionally allow splats from other elements. |
2201 | bool isZeroEltSplat() const { |
2202 | return !changesLength() && isZeroEltSplatMask(ShuffleMask); |
2203 | } |
2204 | |
2205 | /// Return true if this shuffle mask is a transpose mask. |
2206 | /// Transpose vector masks transpose a 2xn matrix. They read corresponding |
2207 | /// even- or odd-numbered vector elements from two n-dimensional source |
2208 | /// vectors and write each result into consecutive elements of an |
2209 | /// n-dimensional destination vector. Two shuffles are necessary to complete |
2210 | /// the transpose, one for the even elements and another for the odd elements. |
2211 | /// This description closely follows how the TRN1 and TRN2 AArch64 |
2212 | /// instructions operate. |
2213 | /// |
2214 | /// For example, a simple 2x2 matrix can be transposed with: |
2215 | /// |
2216 | /// ; Original matrix |
2217 | /// m0 = < a, b > |
2218 | /// m1 = < c, d > |
2219 | /// |
2220 | /// ; Transposed matrix |
2221 | /// t0 = < a, c > = shufflevector m0, m1, < 0, 2 > |
2222 | /// t1 = < b, d > = shufflevector m0, m1, < 1, 3 > |
2223 | /// |
2224 | /// For matrices having greater than n columns, the resulting nx2 transposed |
2225 | /// matrix is stored in two result vectors such that one vector contains |
2226 | /// interleaved elements from all the even-numbered rows and the other vector |
2227 | /// contains interleaved elements from all the odd-numbered rows. For example, |
2228 | /// a 2x4 matrix can be transposed with: |
2229 | /// |
2230 | /// ; Original matrix |
2231 | /// m0 = < a, b, c, d > |
2232 | /// m1 = < e, f, g, h > |
2233 | /// |
2234 | /// ; Transposed matrix |
2235 | /// t0 = < a, e, c, g > = shufflevector m0, m1 < 0, 4, 2, 6 > |
2236 | /// t1 = < b, f, d, h > = shufflevector m0, m1 < 1, 5, 3, 7 > |
2237 | static bool isTransposeMask(ArrayRef<int> Mask); |
2238 | static bool isTransposeMask(const Constant *Mask) { |
2239 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant." ) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2239, __PRETTY_FUNCTION__)); |
2240 | SmallVector<int, 16> MaskAsInts; |
2241 | getShuffleMask(Mask, MaskAsInts); |
2242 | return isTransposeMask(MaskAsInts); |
2243 | } |
2244 | |
2245 | /// Return true if this shuffle transposes the elements of its inputs without |
2246 | /// changing the length of the vectors. This operation may also be known as a |
2247 | /// merge or interleave. See the description for isTransposeMask() for the |
2248 | /// exact specification. |
2249 | /// Example: shufflevector <4 x n> A, <4 x n> B, <0,4,2,6> |
2250 | bool isTranspose() const { |
2251 | return !changesLength() && isTransposeMask(ShuffleMask); |
2252 | } |
2253 | |
2254 | /// Return true if this shuffle mask is an extract subvector mask. |
2255 | /// A valid extract subvector mask returns a smaller vector from a single |
2256 | /// source operand. The base extraction index is returned as well. |
2257 | static bool isExtractSubvectorMask(ArrayRef<int> Mask, int NumSrcElts, |
2258 | int &Index); |
2259 | static bool isExtractSubvectorMask(const Constant *Mask, int NumSrcElts, |
2260 | int &Index) { |
2261 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant." ) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2261, __PRETTY_FUNCTION__)); |
2262 | // Not possible to express a shuffle mask for a scalable vector for this |
2263 | // case. |
2264 | if (isa<ScalableVectorType>(Mask->getType())) |
2265 | return false; |
2266 | SmallVector<int, 16> MaskAsInts; |
2267 | getShuffleMask(Mask, MaskAsInts); |
2268 | return isExtractSubvectorMask(MaskAsInts, NumSrcElts, Index); |
2269 | } |
2270 | |
2271 | /// Return true if this shuffle mask is an extract subvector mask. |
2272 | bool isExtractSubvectorMask(int &Index) const { |
2273 | // Not possible to express a shuffle mask for a scalable vector for this |
2274 | // case. |
2275 | if (isa<ScalableVectorType>(getType())) |
2276 | return false; |
2277 | |
2278 | int NumSrcElts = |
2279 | cast<FixedVectorType>(Op<0>()->getType())->getNumElements(); |
2280 | return isExtractSubvectorMask(ShuffleMask, NumSrcElts, Index); |
2281 | } |
2282 | |
2283 | /// Change values in a shuffle permute mask assuming the two vector operands |
2284 | /// of length InVecNumElts have swapped position. |
2285 | static void commuteShuffleMask(MutableArrayRef<int> Mask, |
2286 | unsigned InVecNumElts) { |
2287 | for (int &Idx : Mask) { |
2288 | if (Idx == -1) |
2289 | continue; |
2290 | Idx = Idx < (int)InVecNumElts ? Idx + InVecNumElts : Idx - InVecNumElts; |
2291 | assert(Idx >= 0 && Idx < (int)InVecNumElts * 2 &&((Idx >= 0 && Idx < (int)InVecNumElts * 2 && "shufflevector mask index out of range") ? static_cast<void > (0) : __assert_fail ("Idx >= 0 && Idx < (int)InVecNumElts * 2 && \"shufflevector mask index out of range\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2292, __PRETTY_FUNCTION__)) |
2292 | "shufflevector mask index out of range")((Idx >= 0 && Idx < (int)InVecNumElts * 2 && "shufflevector mask index out of range") ? static_cast<void > (0) : __assert_fail ("Idx >= 0 && Idx < (int)InVecNumElts * 2 && \"shufflevector mask index out of range\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2292, __PRETTY_FUNCTION__)); |
2293 | } |
2294 | } |
2295 | |
2296 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2297 | static bool classof(const Instruction *I) { |
2298 | return I->getOpcode() == Instruction::ShuffleVector; |
2299 | } |
2300 | static bool classof(const Value *V) { |
2301 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2302 | } |
2303 | }; |
2304 | |
2305 | template <> |
2306 | struct OperandTraits<ShuffleVectorInst> |
2307 | : public FixedNumOperandTraits<ShuffleVectorInst, 2> {}; |
2308 | |
2309 | 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 { ((i_nocapture < OperandTraits<ShuffleVectorInst >::operands(this) && "getOperand() out of range!") ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ShuffleVectorInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2309, __PRETTY_FUNCTION__)); 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 ) { ((i_nocapture < OperandTraits<ShuffleVectorInst> ::operands(this) && "setOperand() out of range!") ? static_cast <void> (0) : __assert_fail ("i_nocapture < OperandTraits<ShuffleVectorInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2309, __PRETTY_FUNCTION__)); 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); } |
2310 | |
2311 | //===----------------------------------------------------------------------===// |
2312 | // ExtractValueInst Class |
2313 | //===----------------------------------------------------------------------===// |
2314 | |
2315 | /// This instruction extracts a struct member or array |
2316 | /// element value from an aggregate value. |
2317 | /// |
2318 | class ExtractValueInst : public UnaryInstruction { |
2319 | SmallVector<unsigned, 4> Indices; |
2320 | |
2321 | ExtractValueInst(const ExtractValueInst &EVI); |
2322 | |
2323 | /// Constructors - Create a extractvalue instruction with a base aggregate |
2324 | /// value and a list of indices. The first ctor can optionally insert before |
2325 | /// an existing instruction, the second appends the new instruction to the |
2326 | /// specified BasicBlock. |
2327 | inline ExtractValueInst(Value *Agg, |
2328 | ArrayRef<unsigned> Idxs, |
2329 | const Twine &NameStr, |
2330 | Instruction *InsertBefore); |
2331 | inline ExtractValueInst(Value *Agg, |
2332 | ArrayRef<unsigned> Idxs, |
2333 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2334 | |
2335 | void init(ArrayRef<unsigned> Idxs, const Twine &NameStr); |
2336 | |
2337 | protected: |
2338 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2339 | friend class Instruction; |
2340 | |
2341 | ExtractValueInst *cloneImpl() const; |
2342 | |
2343 | public: |
2344 | static ExtractValueInst *Create(Value *Agg, |
2345 | ArrayRef<unsigned> Idxs, |
2346 | const Twine &NameStr = "", |
2347 | Instruction *InsertBefore = nullptr) { |
2348 | return new |
2349 | ExtractValueInst(Agg, Idxs, NameStr, InsertBefore); |
2350 | } |
2351 | |
2352 | static ExtractValueInst *Create(Value *Agg, |
2353 | ArrayRef<unsigned> Idxs, |
2354 | const Twine &NameStr, |
2355 | BasicBlock *InsertAtEnd) { |
2356 | return new ExtractValueInst(Agg, Idxs, NameStr, InsertAtEnd); |
2357 | } |
2358 | |
2359 | /// Returns the type of the element that would be extracted |
2360 | /// with an extractvalue instruction with the specified parameters. |
2361 | /// |
2362 | /// Null is returned if the indices are invalid for the specified type. |
2363 | static Type *getIndexedType(Type *Agg, ArrayRef<unsigned> Idxs); |
2364 | |
2365 | using idx_iterator = const unsigned*; |
2366 | |
2367 | inline idx_iterator idx_begin() const { return Indices.begin(); } |
2368 | inline idx_iterator idx_end() const { return Indices.end(); } |
2369 | inline iterator_range<idx_iterator> indices() const { |
2370 | return make_range(idx_begin(), idx_end()); |
2371 | } |
2372 | |
2373 | Value *getAggregateOperand() { |
2374 | return getOperand(0); |
2375 | } |
2376 | const Value *getAggregateOperand() const { |
2377 | return getOperand(0); |
2378 | } |
2379 | static unsigned getAggregateOperandIndex() { |
2380 | return 0U; // get index for modifying correct operand |
2381 | } |
2382 | |
2383 | ArrayRef<unsigned> getIndices() const { |
2384 | return Indices; |
2385 | } |
2386 | |
2387 | unsigned getNumIndices() const { |
2388 | return (unsigned)Indices.size(); |
2389 | } |
2390 | |
2391 | bool hasIndices() const { |
2392 | return true; |
2393 | } |
2394 | |
2395 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2396 | static bool classof(const Instruction *I) { |
2397 | return I->getOpcode() == Instruction::ExtractValue; |
2398 | } |
2399 | static bool classof(const Value *V) { |
2400 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2401 | } |
2402 | }; |
2403 | |
2404 | ExtractValueInst::ExtractValueInst(Value *Agg, |
2405 | ArrayRef<unsigned> Idxs, |
2406 | const Twine &NameStr, |
2407 | Instruction *InsertBefore) |
2408 | : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)), |
2409 | ExtractValue, Agg, InsertBefore) { |
2410 | init(Idxs, NameStr); |
2411 | } |
2412 | |
2413 | ExtractValueInst::ExtractValueInst(Value *Agg, |
2414 | ArrayRef<unsigned> Idxs, |
2415 | const Twine &NameStr, |
2416 | BasicBlock *InsertAtEnd) |
2417 | : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)), |
2418 | ExtractValue, Agg, InsertAtEnd) { |
2419 | init(Idxs, NameStr); |
2420 | } |
2421 | |
2422 | //===----------------------------------------------------------------------===// |
2423 | // InsertValueInst Class |
2424 | //===----------------------------------------------------------------------===// |
2425 | |
2426 | /// This instruction inserts a struct field of array element |
2427 | /// value into an aggregate value. |
2428 | /// |
2429 | class InsertValueInst : public Instruction { |
2430 | SmallVector<unsigned, 4> Indices; |
2431 | |
2432 | InsertValueInst(const InsertValueInst &IVI); |
2433 | |
2434 | /// Constructors - Create a insertvalue instruction with a base aggregate |
2435 | /// value, a value to insert, and a list of indices. The first ctor can |
2436 | /// optionally insert before an existing instruction, the second appends |
2437 | /// the new instruction to the specified BasicBlock. |
2438 | inline InsertValueInst(Value *Agg, Value *Val, |
2439 | ArrayRef<unsigned> Idxs, |
2440 | const Twine &NameStr, |
2441 | Instruction *InsertBefore); |
2442 | inline InsertValueInst(Value *Agg, Value *Val, |
2443 | ArrayRef<unsigned> Idxs, |
2444 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2445 | |
2446 | /// Constructors - These two constructors are convenience methods because one |
2447 | /// and two index insertvalue instructions are so common. |
2448 | InsertValueInst(Value *Agg, Value *Val, unsigned Idx, |
2449 | const Twine &NameStr = "", |
2450 | Instruction *InsertBefore = nullptr); |
2451 | InsertValueInst(Value *Agg, Value *Val, unsigned Idx, const Twine &NameStr, |
2452 | BasicBlock *InsertAtEnd); |
2453 | |
2454 | void init(Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, |
2455 | const Twine &NameStr); |
2456 | |
2457 | protected: |
2458 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2459 | friend class Instruction; |
2460 | |
2461 | InsertValueInst *cloneImpl() const; |
2462 | |
2463 | public: |
2464 | // allocate space for exactly two operands |
2465 | void *operator new(size_t s) { |
2466 | return User::operator new(s, 2); |
2467 | } |
2468 | |
2469 | static InsertValueInst *Create(Value *Agg, Value *Val, |
2470 | ArrayRef<unsigned> Idxs, |
2471 | const Twine &NameStr = "", |
2472 | Instruction *InsertBefore = nullptr) { |
2473 | return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertBefore); |
2474 | } |
2475 | |
2476 | static InsertValueInst *Create(Value *Agg, Value *Val, |
2477 | ArrayRef<unsigned> Idxs, |
2478 | const Twine &NameStr, |
2479 | BasicBlock *InsertAtEnd) { |
2480 | return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertAtEnd); |
2481 | } |
2482 | |
2483 | /// Transparently provide more efficient getOperand methods. |
2484 | 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; |
2485 | |
2486 | using idx_iterator = const unsigned*; |
2487 | |
2488 | inline idx_iterator idx_begin() const { return Indices.begin(); } |
2489 | inline idx_iterator idx_end() const { return Indices.end(); } |
2490 | inline iterator_range<idx_iterator> indices() const { |
2491 | return make_range(idx_begin(), idx_end()); |
2492 | } |
2493 | |
2494 | Value *getAggregateOperand() { |
2495 | return getOperand(0); |
2496 | } |
2497 | const Value *getAggregateOperand() const { |
2498 | return getOperand(0); |
2499 | } |
2500 | static unsigned getAggregateOperandIndex() { |
2501 | return 0U; // get index for modifying correct operand |
2502 | } |
2503 | |
2504 | Value *getInsertedValueOperand() { |
2505 | return getOperand(1); |
2506 | } |
2507 | const Value *getInsertedValueOperand() const { |
2508 | return getOperand(1); |
2509 | } |
2510 | static unsigned getInsertedValueOperandIndex() { |
2511 | return 1U; // get index for modifying correct operand |
2512 | } |
2513 | |
2514 | ArrayRef<unsigned> getIndices() const { |
2515 | return Indices; |
2516 | } |
2517 | |
2518 | unsigned getNumIndices() const { |
2519 | return (unsigned)Indices.size(); |
2520 | } |
2521 | |
2522 | bool hasIndices() const { |
2523 | return true; |
2524 | } |
2525 | |
2526 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2527 | static bool classof(const Instruction *I) { |
2528 | return I->getOpcode() == Instruction::InsertValue; |
2529 | } |
2530 | static bool classof(const Value *V) { |
2531 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2532 | } |
2533 | }; |
2534 | |
2535 | template <> |
2536 | struct OperandTraits<InsertValueInst> : |
2537 | public FixedNumOperandTraits<InsertValueInst, 2> { |
2538 | }; |
2539 | |
2540 | InsertValueInst::InsertValueInst(Value *Agg, |
2541 | Value *Val, |
2542 | ArrayRef<unsigned> Idxs, |
2543 | const Twine &NameStr, |
2544 | Instruction *InsertBefore) |
2545 | : Instruction(Agg->getType(), InsertValue, |
2546 | OperandTraits<InsertValueInst>::op_begin(this), |
2547 | 2, InsertBefore) { |
2548 | init(Agg, Val, Idxs, NameStr); |
2549 | } |
2550 | |
2551 | InsertValueInst::InsertValueInst(Value *Agg, |
2552 | Value *Val, |
2553 | ArrayRef<unsigned> Idxs, |
2554 | const Twine &NameStr, |
2555 | BasicBlock *InsertAtEnd) |
2556 | : Instruction(Agg->getType(), InsertValue, |
2557 | OperandTraits<InsertValueInst>::op_begin(this), |
2558 | 2, InsertAtEnd) { |
2559 | init(Agg, Val, Idxs, NameStr); |
2560 | } |
2561 | |
2562 | 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 { ((i_nocapture < OperandTraits<InsertValueInst>::operands(this) && "getOperand() out of range!") ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<InsertValueInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2562, __PRETTY_FUNCTION__)); 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) { (( i_nocapture < OperandTraits<InsertValueInst>::operands (this) && "setOperand() out of range!") ? static_cast <void> (0) : __assert_fail ("i_nocapture < OperandTraits<InsertValueInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2562, __PRETTY_FUNCTION__)); 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); } |
2563 | |
2564 | //===----------------------------------------------------------------------===// |
2565 | // PHINode Class |
2566 | //===----------------------------------------------------------------------===// |
2567 | |
2568 | // PHINode - The PHINode class is used to represent the magical mystical PHI |
2569 | // node, that can not exist in nature, but can be synthesized in a computer |
2570 | // scientist's overactive imagination. |
2571 | // |
2572 | class PHINode : public Instruction { |
2573 | /// The number of operands actually allocated. NumOperands is |
2574 | /// the number actually in use. |
2575 | unsigned ReservedSpace; |
2576 | |
2577 | PHINode(const PHINode &PN); |
2578 | |
2579 | explicit PHINode(Type *Ty, unsigned NumReservedValues, |
2580 | const Twine &NameStr = "", |
2581 | Instruction *InsertBefore = nullptr) |
2582 | : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertBefore), |
2583 | ReservedSpace(NumReservedValues) { |
2584 | setName(NameStr); |
2585 | allocHungoffUses(ReservedSpace); |
2586 | } |
2587 | |
2588 | PHINode(Type *Ty, unsigned NumReservedValues, const Twine &NameStr, |
2589 | BasicBlock *InsertAtEnd) |
2590 | : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertAtEnd), |
2591 | ReservedSpace(NumReservedValues) { |
2592 | setName(NameStr); |
2593 | allocHungoffUses(ReservedSpace); |
2594 | } |
2595 | |
2596 | protected: |
2597 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2598 | friend class Instruction; |
2599 | |
2600 | PHINode *cloneImpl() const; |
2601 | |
2602 | // allocHungoffUses - this is more complicated than the generic |
2603 | // User::allocHungoffUses, because we have to allocate Uses for the incoming |
2604 | // values and pointers to the incoming blocks, all in one allocation. |
2605 | void allocHungoffUses(unsigned N) { |
2606 | User::allocHungoffUses(N, /* IsPhi */ true); |
2607 | } |
2608 | |
2609 | public: |
2610 | /// Constructors - NumReservedValues is a hint for the number of incoming |
2611 | /// edges that this phi node will have (use 0 if you really have no idea). |
2612 | static PHINode *Create(Type *Ty, unsigned NumReservedValues, |
2613 | const Twine &NameStr = "", |
2614 | Instruction *InsertBefore = nullptr) { |
2615 | return new PHINode(Ty, NumReservedValues, NameStr, InsertBefore); |
2616 | } |
2617 | |
2618 | static PHINode *Create(Type *Ty, unsigned NumReservedValues, |
2619 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
2620 | return new PHINode(Ty, NumReservedValues, NameStr, InsertAtEnd); |
2621 | } |
2622 | |
2623 | /// Provide fast operand accessors |
2624 | 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; |
2625 | |
2626 | // Block iterator interface. This provides access to the list of incoming |
2627 | // basic blocks, which parallels the list of incoming values. |
2628 | |
2629 | using block_iterator = BasicBlock **; |
2630 | using const_block_iterator = BasicBlock * const *; |
2631 | |
2632 | block_iterator block_begin() { |
2633 | return reinterpret_cast<block_iterator>(op_begin() + ReservedSpace); |
2634 | } |
2635 | |
2636 | const_block_iterator block_begin() const { |
2637 | return reinterpret_cast<const_block_iterator>(op_begin() + ReservedSpace); |
2638 | } |
2639 | |
2640 | block_iterator block_end() { |
2641 | return block_begin() + getNumOperands(); |
2642 | } |
2643 | |
2644 | const_block_iterator block_end() const { |
2645 | return block_begin() + getNumOperands(); |
2646 | } |
2647 | |
2648 | iterator_range<block_iterator> blocks() { |
2649 | return make_range(block_begin(), block_end()); |
2650 | } |
2651 | |
2652 | iterator_range<const_block_iterator> blocks() const { |
2653 | return make_range(block_begin(), block_end()); |
2654 | } |
2655 | |
2656 | op_range incoming_values() { return operands(); } |
2657 | |
2658 | const_op_range incoming_values() const { return operands(); } |
2659 | |
2660 | /// Return the number of incoming edges |
2661 | /// |
2662 | unsigned getNumIncomingValues() const { return getNumOperands(); } |
2663 | |
2664 | /// Return incoming value number x |
2665 | /// |
2666 | Value *getIncomingValue(unsigned i) const { |
2667 | return getOperand(i); |
2668 | } |
2669 | void setIncomingValue(unsigned i, Value *V) { |
2670 | assert(V && "PHI node got a null value!")((V && "PHI node got a null value!") ? static_cast< void> (0) : __assert_fail ("V && \"PHI node got a null value!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2670, __PRETTY_FUNCTION__)); |
2671 | assert(getType() == V->getType() &&((getType() == V->getType() && "All operands to PHI node must be the same type as the PHI node!" ) ? static_cast<void> (0) : __assert_fail ("getType() == V->getType() && \"All operands to PHI node must be the same type as the PHI node!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2672, __PRETTY_FUNCTION__)) |
2672 | "All operands to PHI node must be the same type as the PHI node!")((getType() == V->getType() && "All operands to PHI node must be the same type as the PHI node!" ) ? static_cast<void> (0) : __assert_fail ("getType() == V->getType() && \"All operands to PHI node must be the same type as the PHI node!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2672, __PRETTY_FUNCTION__)); |
2673 | setOperand(i, V); |
2674 | } |
2675 | |
2676 | static unsigned getOperandNumForIncomingValue(unsigned i) { |
2677 | return i; |
2678 | } |
2679 | |
2680 | static unsigned getIncomingValueNumForOperand(unsigned i) { |
2681 | return i; |
2682 | } |
2683 | |
2684 | /// Return incoming basic block number @p i. |
2685 | /// |
2686 | BasicBlock *getIncomingBlock(unsigned i) const { |
2687 | return block_begin()[i]; |
2688 | } |
2689 | |
2690 | /// Return incoming basic block corresponding |
2691 | /// to an operand of the PHI. |
2692 | /// |
2693 | BasicBlock *getIncomingBlock(const Use &U) const { |
2694 | assert(this == U.getUser() && "Iterator doesn't point to PHI's Uses?")((this == U.getUser() && "Iterator doesn't point to PHI's Uses?" ) ? static_cast<void> (0) : __assert_fail ("this == U.getUser() && \"Iterator doesn't point to PHI's Uses?\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2694, __PRETTY_FUNCTION__)); |
2695 | return getIncomingBlock(unsigned(&U - op_begin())); |
2696 | } |
2697 | |
2698 | /// Return incoming basic block corresponding |
2699 | /// to value use iterator. |
2700 | /// |
2701 | BasicBlock *getIncomingBlock(Value::const_user_iterator I) const { |
2702 | return getIncomingBlock(I.getUse()); |
2703 | } |
2704 | |
2705 | void setIncomingBlock(unsigned i, BasicBlock *BB) { |
2706 | assert(BB && "PHI node got a null basic block!")((BB && "PHI node got a null basic block!") ? static_cast <void> (0) : __assert_fail ("BB && \"PHI node got a null basic block!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2706, __PRETTY_FUNCTION__)); |
2707 | block_begin()[i] = BB; |
2708 | } |
2709 | |
2710 | /// Replace every incoming basic block \p Old to basic block \p New. |
2711 | void replaceIncomingBlockWith(const BasicBlock *Old, BasicBlock *New) { |
2712 | assert(New && Old && "PHI node got a null basic block!")((New && Old && "PHI node got a null basic block!" ) ? static_cast<void> (0) : __assert_fail ("New && Old && \"PHI node got a null basic block!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2712, __PRETTY_FUNCTION__)); |
2713 | for (unsigned Op = 0, NumOps = getNumOperands(); Op != NumOps; ++Op) |
2714 | if (getIncomingBlock(Op) == Old) |
2715 | setIncomingBlock(Op, New); |
2716 | } |
2717 | |
2718 | /// Add an incoming value to the end of the PHI list |
2719 | /// |
2720 | void addIncoming(Value *V, BasicBlock *BB) { |
2721 | if (getNumOperands() == ReservedSpace) |
2722 | growOperands(); // Get more space! |
2723 | // Initialize some new operands. |
2724 | setNumHungOffUseOperands(getNumOperands() + 1); |
2725 | setIncomingValue(getNumOperands() - 1, V); |
2726 | setIncomingBlock(getNumOperands() - 1, BB); |
2727 | } |
2728 | |
2729 | /// Remove an incoming value. This is useful if a |
2730 | /// predecessor basic block is deleted. The value removed is returned. |
2731 | /// |
2732 | /// If the last incoming value for a PHI node is removed (and DeletePHIIfEmpty |
2733 | /// is true), the PHI node is destroyed and any uses of it are replaced with |
2734 | /// dummy values. The only time there should be zero incoming values to a PHI |
2735 | /// node is when the block is dead, so this strategy is sound. |
2736 | /// |
2737 | Value *removeIncomingValue(unsigned Idx, bool DeletePHIIfEmpty = true); |
2738 | |
2739 | Value *removeIncomingValue(const BasicBlock *BB, bool DeletePHIIfEmpty=true) { |
2740 | int Idx = getBasicBlockIndex(BB); |
2741 | assert(Idx >= 0 && "Invalid basic block argument to remove!")((Idx >= 0 && "Invalid basic block argument to remove!" ) ? static_cast<void> (0) : __assert_fail ("Idx >= 0 && \"Invalid basic block argument to remove!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2741, __PRETTY_FUNCTION__)); |
2742 | return removeIncomingValue(Idx, DeletePHIIfEmpty); |
2743 | } |
2744 | |
2745 | /// Return the first index of the specified basic |
2746 | /// block in the value list for this PHI. Returns -1 if no instance. |
2747 | /// |
2748 | int getBasicBlockIndex(const BasicBlock *BB) const { |
2749 | for (unsigned i = 0, e = getNumOperands(); i != e; ++i) |
2750 | if (block_begin()[i] == BB) |
2751 | return i; |
2752 | return -1; |
2753 | } |
2754 | |
2755 | Value *getIncomingValueForBlock(const BasicBlock *BB) const { |
2756 | int Idx = getBasicBlockIndex(BB); |
2757 | assert(Idx >= 0 && "Invalid basic block argument!")((Idx >= 0 && "Invalid basic block argument!") ? static_cast <void> (0) : __assert_fail ("Idx >= 0 && \"Invalid basic block argument!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2757, __PRETTY_FUNCTION__)); |
2758 | return getIncomingValue(Idx); |
2759 | } |
2760 | |
2761 | /// Set every incoming value(s) for block \p BB to \p V. |
2762 | void setIncomingValueForBlock(const BasicBlock *BB, Value *V) { |
2763 | assert(BB && "PHI node got a null basic block!")((BB && "PHI node got a null basic block!") ? static_cast <void> (0) : __assert_fail ("BB && \"PHI node got a null basic block!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2763, __PRETTY_FUNCTION__)); |
2764 | bool Found = false; |
2765 | for (unsigned Op = 0, NumOps = getNumOperands(); Op != NumOps; ++Op) |
2766 | if (getIncomingBlock(Op) == BB) { |
2767 | Found = true; |
2768 | setIncomingValue(Op, V); |
2769 | } |
2770 | (void)Found; |
2771 | assert(Found && "Invalid basic block argument to set!")((Found && "Invalid basic block argument to set!") ? static_cast <void> (0) : __assert_fail ("Found && \"Invalid basic block argument to set!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2771, __PRETTY_FUNCTION__)); |
2772 | } |
2773 | |
2774 | /// If the specified PHI node always merges together the |
2775 | /// same value, return the value, otherwise return null. |
2776 | Value *hasConstantValue() const; |
2777 | |
2778 | /// Whether the specified PHI node always merges |
2779 | /// together the same value, assuming undefs are equal to a unique |
2780 | /// non-undef value. |
2781 | bool hasConstantOrUndefValue() const; |
2782 | |
2783 | /// If the PHI node is complete which means all of its parent's predecessors |
2784 | /// have incoming value in this PHI, return true, otherwise return false. |
2785 | bool isComplete() const { |
2786 | return llvm::all_of(predecessors(getParent()), |
2787 | [this](const BasicBlock *Pred) { |
2788 | return getBasicBlockIndex(Pred) >= 0; |
2789 | }); |
2790 | } |
2791 | |
2792 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
2793 | static bool classof(const Instruction *I) { |
2794 | return I->getOpcode() == Instruction::PHI; |
2795 | } |
2796 | static bool classof(const Value *V) { |
2797 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2798 | } |
2799 | |
2800 | private: |
2801 | void growOperands(); |
2802 | }; |
2803 | |
2804 | template <> |
2805 | struct OperandTraits<PHINode> : public HungoffOperandTraits<2> { |
2806 | }; |
2807 | |
2808 | 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 { ((i_nocapture < OperandTraits<PHINode>::operands (this) && "getOperand() out of range!") ? static_cast <void> (0) : __assert_fail ("i_nocapture < OperandTraits<PHINode>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2808, __PRETTY_FUNCTION__)); 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) { ((i_nocapture < OperandTraits<PHINode>::operands(this) && "setOperand() out of range!" ) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<PHINode>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2808, __PRETTY_FUNCTION__)); 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 ); } |
2809 | |
2810 | //===----------------------------------------------------------------------===// |
2811 | // LandingPadInst Class |
2812 | //===----------------------------------------------------------------------===// |
2813 | |
2814 | //===--------------------------------------------------------------------------- |
2815 | /// The landingpad instruction holds all of the information |
2816 | /// necessary to generate correct exception handling. The landingpad instruction |
2817 | /// cannot be moved from the top of a landing pad block, which itself is |
2818 | /// accessible only from the 'unwind' edge of an invoke. This uses the |
2819 | /// SubclassData field in Value to store whether or not the landingpad is a |
2820 | /// cleanup. |
2821 | /// |
2822 | class LandingPadInst : public Instruction { |
2823 | using CleanupField = BoolBitfieldElementT<0>; |
2824 | |
2825 | /// The number of operands actually allocated. NumOperands is |
2826 | /// the number actually in use. |
2827 | unsigned ReservedSpace; |
2828 | |
2829 | LandingPadInst(const LandingPadInst &LP); |
2830 | |
2831 | public: |
2832 | enum ClauseType { Catch, Filter }; |
2833 | |
2834 | private: |
2835 | explicit LandingPadInst(Type *RetTy, unsigned NumReservedValues, |
2836 | const Twine &NameStr, Instruction *InsertBefore); |
2837 | explicit LandingPadInst(Type *RetTy, unsigned NumReservedValues, |
2838 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2839 | |
2840 | // Allocate space for exactly zero operands. |
2841 | void *operator new(size_t s) { |
2842 | return User::operator new(s); |
2843 | } |
2844 | |
2845 | void growOperands(unsigned Size); |
2846 | void init(unsigned NumReservedValues, const Twine &NameStr); |
2847 | |
2848 | protected: |
2849 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2850 | friend class Instruction; |
2851 | |
2852 | LandingPadInst *cloneImpl() const; |
2853 | |
2854 | public: |
2855 | /// Constructors - NumReservedClauses is a hint for the number of incoming |
2856 | /// clauses that this landingpad will have (use 0 if you really have no idea). |
2857 | static LandingPadInst *Create(Type *RetTy, unsigned NumReservedClauses, |
2858 | const Twine &NameStr = "", |
2859 | Instruction *InsertBefore = nullptr); |
2860 | static LandingPadInst *Create(Type *RetTy, unsigned NumReservedClauses, |
2861 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2862 | |
2863 | /// Provide fast operand accessors |
2864 | 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; |
2865 | |
2866 | /// Return 'true' if this landingpad instruction is a |
2867 | /// cleanup. I.e., it should be run when unwinding even if its landing pad |
2868 | /// doesn't catch the exception. |
2869 | bool isCleanup() const { return getSubclassData<CleanupField>(); } |
2870 | |
2871 | /// Indicate that this landingpad instruction is a cleanup. |
2872 | void setCleanup(bool V) { setSubclassData<CleanupField>(V); } |
2873 | |
2874 | /// Add a catch or filter clause to the landing pad. |
2875 | void addClause(Constant *ClauseVal); |
2876 | |
2877 | /// Get the value of the clause at index Idx. Use isCatch/isFilter to |
2878 | /// determine what type of clause this is. |
2879 | Constant *getClause(unsigned Idx) const { |
2880 | return cast<Constant>(getOperandList()[Idx]); |
2881 | } |
2882 | |
2883 | /// Return 'true' if the clause and index Idx is a catch clause. |
2884 | bool isCatch(unsigned Idx) const { |
2885 | return !isa<ArrayType>(getOperandList()[Idx]->getType()); |
2886 | } |
2887 | |
2888 | /// Return 'true' if the clause and index Idx is a filter clause. |
2889 | bool isFilter(unsigned Idx) const { |
2890 | return isa<ArrayType>(getOperandList()[Idx]->getType()); |
2891 | } |
2892 | |
2893 | /// Get the number of clauses for this landing pad. |
2894 | unsigned getNumClauses() const { return getNumOperands(); } |
2895 | |
2896 | /// Grow the size of the operand list to accommodate the new |
2897 | /// number of clauses. |
2898 | void reserveClauses(unsigned Size) { growOperands(Size); } |
2899 | |
2900 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2901 | static bool classof(const Instruction *I) { |
2902 | return I->getOpcode() == Instruction::LandingPad; |
2903 | } |
2904 | static bool classof(const Value *V) { |
2905 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2906 | } |
2907 | }; |
2908 | |
2909 | template <> |
2910 | struct OperandTraits<LandingPadInst> : public HungoffOperandTraits<1> { |
2911 | }; |
2912 | |
2913 | 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 { ((i_nocapture < OperandTraits<LandingPadInst>::operands(this) && "getOperand() out of range!") ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<LandingPadInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2913, __PRETTY_FUNCTION__)); 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) { (( i_nocapture < OperandTraits<LandingPadInst>::operands (this) && "setOperand() out of range!") ? static_cast <void> (0) : __assert_fail ("i_nocapture < OperandTraits<LandingPadInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h" , 2913, __PRETTY_FUNCTION__)); 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); } |
2914 | |
2915 | //===----------------------------------------------------------------------===// |
2916 | // ReturnInst Class |
2917 | //===----------------------------------------------------------------------===// |
2918 | |
2919 | //===--------------------------------------------------------------------------- |
2920 | /// Return a value (possibly void), from a function. Execution |
2921 | /// does not continue in this function any longer. |
2922 | /// |
2923 | class ReturnInst : public Instruction { |
2924 | ReturnInst(const ReturnInst &RI); |
2925 | |
2926 | private: |
2927 | // ReturnInst constructors: |
2928 | // ReturnInst() - 'ret void' instruction |
2929 | // ReturnInst( null) - 'ret void' instruction |
2930 | // ReturnInst(Value* X) - 'ret X' instruction |
2931 | // ReturnInst( null, Inst *I) - 'ret void' instruction, insert before I |
2932 | // ReturnInst(Value* X, Inst *I) - 'ret X' instruction, insert before I |
2933 | // ReturnInst( null, BB *B) - 'ret void' instruction, insert @ end of B |
2934 | // ReturnInst(Value* X, BB *B) - 'ret X' instruction, insert @ end of B |
2935 | // |
2936 | // NOTE: If the Value* passed is of type void then the constructor behaves as |
2937 | // if it was passed NULL. |
2938 | explicit ReturnInst(LLVMContext &C, Value *retVal = nullptr, |
2939 | Instruction *InsertBefore = nullptr); |
2940 | ReturnInst(LLVMContext &C, Value *retVal, BasicBlock *InsertAtEnd); |
2941 | explicit ReturnInst(LLVMContext &C, BasicBlock *InsertAtEnd); |
2942 | |
2943 | protected: |
2944 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2945 | friend class Instruction; |
2946 | |
2947 | ReturnInst *cloneImpl() const; |
2948 | |
2949 | public: |
2950 | static ReturnInst* Create(LLVMContext &C, Value *retVal = nullptr, |
2951 | Instruction *InsertBefore = nullptr) { |
2952 | return new(!!retVal) ReturnInst(C, retVal, InsertBefore); |
2953 | } |
2954 | |
2955 | static ReturnInst* Create(LLVMContext &C, Value *retVal, |
2956 | BasicBlock *InsertAtEnd) { |
2957 | return new(!!retVal) ReturnInst(C, retVal, InsertAtEnd); |
2958 | } |
2959 | |
2960 | static ReturnInst* Create(LLVMContext &C, BasicBlock *InsertAtEnd) { |
2961 | return new(0) ReturnInst(C, InsertAtEnd); |
2962 | } |
2963 | |
2964 | /// Provide fast operand accessors |
2965 | 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; |