File: | build/source/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp |
Warning: | line 474, column 21 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===-- AMDGPUPromoteAlloca.cpp - Promote Allocas -------------------------===// | ||||
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 eliminates allocas by either converting them into vectors or | ||||
10 | // by migrating them to local address space. | ||||
11 | // | ||||
12 | //===----------------------------------------------------------------------===// | ||||
13 | |||||
14 | #include "AMDGPU.h" | ||||
15 | #include "GCNSubtarget.h" | ||||
16 | #include "Utils/AMDGPUBaseInfo.h" | ||||
17 | #include "llvm/Analysis/CaptureTracking.h" | ||||
18 | #include "llvm/Analysis/ValueTracking.h" | ||||
19 | #include "llvm/CodeGen/TargetPassConfig.h" | ||||
20 | #include "llvm/IR/IRBuilder.h" | ||||
21 | #include "llvm/IR/IntrinsicInst.h" | ||||
22 | #include "llvm/IR/IntrinsicsAMDGPU.h" | ||||
23 | #include "llvm/IR/IntrinsicsR600.h" | ||||
24 | #include "llvm/Pass.h" | ||||
25 | #include "llvm/Target/TargetMachine.h" | ||||
26 | |||||
27 | #define DEBUG_TYPE"amdgpu-promote-alloca" "amdgpu-promote-alloca" | ||||
28 | |||||
29 | using namespace llvm; | ||||
30 | |||||
31 | namespace { | ||||
32 | |||||
33 | static cl::opt<bool> DisablePromoteAllocaToVector( | ||||
34 | "disable-promote-alloca-to-vector", | ||||
35 | cl::desc("Disable promote alloca to vector"), | ||||
36 | cl::init(false)); | ||||
37 | |||||
38 | static cl::opt<bool> DisablePromoteAllocaToLDS( | ||||
39 | "disable-promote-alloca-to-lds", | ||||
40 | cl::desc("Disable promote alloca to LDS"), | ||||
41 | cl::init(false)); | ||||
42 | |||||
43 | static cl::opt<unsigned> PromoteAllocaToVectorLimit( | ||||
44 | "amdgpu-promote-alloca-to-vector-limit", | ||||
45 | cl::desc("Maximum byte size to consider promote alloca to vector"), | ||||
46 | cl::init(0)); | ||||
47 | |||||
48 | // FIXME: This can create globals so should be a module pass. | ||||
49 | class AMDGPUPromoteAlloca : public FunctionPass { | ||||
50 | public: | ||||
51 | static char ID; | ||||
52 | |||||
53 | AMDGPUPromoteAlloca() : FunctionPass(ID) {} | ||||
54 | |||||
55 | bool runOnFunction(Function &F) override; | ||||
56 | |||||
57 | StringRef getPassName() const override { return "AMDGPU Promote Alloca"; } | ||||
58 | |||||
59 | bool handleAlloca(AllocaInst &I, bool SufficientLDS); | ||||
60 | |||||
61 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||
62 | AU.setPreservesCFG(); | ||||
63 | FunctionPass::getAnalysisUsage(AU); | ||||
64 | } | ||||
65 | }; | ||||
66 | |||||
67 | class AMDGPUPromoteAllocaImpl { | ||||
68 | private: | ||||
69 | const TargetMachine &TM; | ||||
70 | Module *Mod = nullptr; | ||||
71 | const DataLayout *DL = nullptr; | ||||
72 | |||||
73 | // FIXME: This should be per-kernel. | ||||
74 | uint32_t LocalMemLimit = 0; | ||||
75 | uint32_t CurrentLocalMemUsage = 0; | ||||
76 | unsigned MaxVGPRs; | ||||
77 | |||||
78 | bool IsAMDGCN = false; | ||||
79 | bool IsAMDHSA = false; | ||||
80 | |||||
81 | std::pair<Value *, Value *> getLocalSizeYZ(IRBuilder<> &Builder); | ||||
82 | Value *getWorkitemID(IRBuilder<> &Builder, unsigned N); | ||||
83 | |||||
84 | /// BaseAlloca is the alloca root the search started from. | ||||
85 | /// Val may be that alloca or a recursive user of it. | ||||
86 | bool collectUsesWithPtrTypes(Value *BaseAlloca, | ||||
87 | Value *Val, | ||||
88 | std::vector<Value*> &WorkList) const; | ||||
89 | |||||
90 | /// Val is a derived pointer from Alloca. OpIdx0/OpIdx1 are the operand | ||||
91 | /// indices to an instruction with 2 pointer inputs (e.g. select, icmp). | ||||
92 | /// Returns true if both operands are derived from the same alloca. Val should | ||||
93 | /// be the same value as one of the input operands of UseInst. | ||||
94 | bool binaryOpIsDerivedFromSameAlloca(Value *Alloca, Value *Val, | ||||
95 | Instruction *UseInst, | ||||
96 | int OpIdx0, int OpIdx1) const; | ||||
97 | |||||
98 | /// Check whether we have enough local memory for promotion. | ||||
99 | bool hasSufficientLocalMem(const Function &F); | ||||
100 | |||||
101 | bool handleAlloca(AllocaInst &I, bool SufficientLDS); | ||||
102 | |||||
103 | public: | ||||
104 | AMDGPUPromoteAllocaImpl(TargetMachine &TM) : TM(TM) {} | ||||
105 | bool run(Function &F); | ||||
106 | }; | ||||
107 | |||||
108 | class AMDGPUPromoteAllocaToVector : public FunctionPass { | ||||
109 | public: | ||||
110 | static char ID; | ||||
111 | |||||
112 | AMDGPUPromoteAllocaToVector() : FunctionPass(ID) {} | ||||
113 | |||||
114 | bool runOnFunction(Function &F) override; | ||||
115 | |||||
116 | StringRef getPassName() const override { | ||||
117 | return "AMDGPU Promote Alloca to vector"; | ||||
118 | } | ||||
119 | |||||
120 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||
121 | AU.setPreservesCFG(); | ||||
122 | FunctionPass::getAnalysisUsage(AU); | ||||
123 | } | ||||
124 | }; | ||||
125 | |||||
126 | } // end anonymous namespace | ||||
127 | |||||
128 | char AMDGPUPromoteAlloca::ID = 0; | ||||
129 | char AMDGPUPromoteAllocaToVector::ID = 0; | ||||
130 | |||||
131 | INITIALIZE_PASS_BEGIN(AMDGPUPromoteAlloca, DEBUG_TYPE,static void *initializeAMDGPUPromoteAllocaPassOnce(PassRegistry &Registry) { | ||||
132 | "AMDGPU promote alloca to vector or LDS", false, false)static void *initializeAMDGPUPromoteAllocaPassOnce(PassRegistry &Registry) { | ||||
133 | // Move LDS uses from functions to kernels before promote alloca for accurate | ||||
134 | // estimation of LDS available | ||||
135 | INITIALIZE_PASS_DEPENDENCY(AMDGPULowerModuleLDS)initializeAMDGPULowerModuleLDSPass(Registry); | ||||
136 | INITIALIZE_PASS_END(AMDGPUPromoteAlloca, DEBUG_TYPE,PassInfo *PI = new PassInfo( "AMDGPU promote alloca to vector or LDS" , "amdgpu-promote-alloca", &AMDGPUPromoteAlloca::ID, PassInfo ::NormalCtor_t(callDefaultCtor<AMDGPUPromoteAlloca>), false , false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeAMDGPUPromoteAllocaPassFlag; void llvm ::initializeAMDGPUPromoteAllocaPass(PassRegistry &Registry ) { llvm::call_once(InitializeAMDGPUPromoteAllocaPassFlag, initializeAMDGPUPromoteAllocaPassOnce , std::ref(Registry)); } | ||||
137 | "AMDGPU promote alloca to vector or LDS", false, false)PassInfo *PI = new PassInfo( "AMDGPU promote alloca to vector or LDS" , "amdgpu-promote-alloca", &AMDGPUPromoteAlloca::ID, PassInfo ::NormalCtor_t(callDefaultCtor<AMDGPUPromoteAlloca>), false , false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeAMDGPUPromoteAllocaPassFlag; void llvm ::initializeAMDGPUPromoteAllocaPass(PassRegistry &Registry ) { llvm::call_once(InitializeAMDGPUPromoteAllocaPassFlag, initializeAMDGPUPromoteAllocaPassOnce , std::ref(Registry)); } | ||||
138 | |||||
139 | INITIALIZE_PASS(AMDGPUPromoteAllocaToVector, DEBUG_TYPE "-to-vector",static void *initializeAMDGPUPromoteAllocaToVectorPassOnce(PassRegistry &Registry) { PassInfo *PI = new PassInfo( "AMDGPU promote alloca to vector" , "amdgpu-promote-alloca" "-to-vector", &AMDGPUPromoteAllocaToVector ::ID, PassInfo::NormalCtor_t(callDefaultCtor<AMDGPUPromoteAllocaToVector >), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeAMDGPUPromoteAllocaToVectorPassFlag ; void llvm::initializeAMDGPUPromoteAllocaToVectorPass(PassRegistry &Registry) { llvm::call_once(InitializeAMDGPUPromoteAllocaToVectorPassFlag , initializeAMDGPUPromoteAllocaToVectorPassOnce, std::ref(Registry )); } | ||||
140 | "AMDGPU promote alloca to vector", false, false)static void *initializeAMDGPUPromoteAllocaToVectorPassOnce(PassRegistry &Registry) { PassInfo *PI = new PassInfo( "AMDGPU promote alloca to vector" , "amdgpu-promote-alloca" "-to-vector", &AMDGPUPromoteAllocaToVector ::ID, PassInfo::NormalCtor_t(callDefaultCtor<AMDGPUPromoteAllocaToVector >), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeAMDGPUPromoteAllocaToVectorPassFlag ; void llvm::initializeAMDGPUPromoteAllocaToVectorPass(PassRegistry &Registry) { llvm::call_once(InitializeAMDGPUPromoteAllocaToVectorPassFlag , initializeAMDGPUPromoteAllocaToVectorPassOnce, std::ref(Registry )); } | ||||
141 | |||||
142 | char &llvm::AMDGPUPromoteAllocaID = AMDGPUPromoteAlloca::ID; | ||||
143 | char &llvm::AMDGPUPromoteAllocaToVectorID = AMDGPUPromoteAllocaToVector::ID; | ||||
144 | |||||
145 | bool AMDGPUPromoteAlloca::runOnFunction(Function &F) { | ||||
146 | if (skipFunction(F)) | ||||
147 | return false; | ||||
148 | |||||
149 | if (auto *TPC = getAnalysisIfAvailable<TargetPassConfig>()) { | ||||
150 | return AMDGPUPromoteAllocaImpl(TPC->getTM<TargetMachine>()).run(F); | ||||
151 | } | ||||
152 | return false; | ||||
153 | } | ||||
154 | |||||
155 | PreservedAnalyses AMDGPUPromoteAllocaPass::run(Function &F, | ||||
156 | FunctionAnalysisManager &AM) { | ||||
157 | bool Changed = AMDGPUPromoteAllocaImpl(TM).run(F); | ||||
158 | if (Changed) { | ||||
159 | PreservedAnalyses PA; | ||||
160 | PA.preserveSet<CFGAnalyses>(); | ||||
161 | return PA; | ||||
162 | } | ||||
163 | return PreservedAnalyses::all(); | ||||
164 | } | ||||
165 | |||||
166 | bool AMDGPUPromoteAllocaImpl::run(Function &F) { | ||||
167 | Mod = F.getParent(); | ||||
168 | DL = &Mod->getDataLayout(); | ||||
169 | |||||
170 | const Triple &TT = TM.getTargetTriple(); | ||||
171 | IsAMDGCN = TT.getArch() == Triple::amdgcn; | ||||
172 | IsAMDHSA = TT.getOS() == Triple::AMDHSA; | ||||
173 | |||||
174 | const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(TM, F); | ||||
175 | if (!ST.isPromoteAllocaEnabled()) | ||||
176 | return false; | ||||
177 | |||||
178 | if (IsAMDGCN) { | ||||
179 | const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F); | ||||
180 | MaxVGPRs = ST.getMaxNumVGPRs(ST.getWavesPerEU(F).first); | ||||
181 | // A non-entry function has only 32 caller preserved registers. | ||||
182 | // Do not promote alloca which will force spilling. | ||||
183 | if (!AMDGPU::isEntryFunctionCC(F.getCallingConv())) | ||||
184 | MaxVGPRs = std::min(MaxVGPRs, 32u); | ||||
185 | } else { | ||||
186 | MaxVGPRs = 128; | ||||
187 | } | ||||
188 | |||||
189 | bool SufficientLDS = hasSufficientLocalMem(F); | ||||
190 | bool Changed = false; | ||||
191 | BasicBlock &EntryBB = *F.begin(); | ||||
192 | |||||
193 | SmallVector<AllocaInst *, 16> Allocas; | ||||
194 | for (Instruction &I : EntryBB) { | ||||
195 | if (AllocaInst *AI = dyn_cast<AllocaInst>(&I)) | ||||
196 | Allocas.push_back(AI); | ||||
197 | } | ||||
198 | |||||
199 | for (AllocaInst *AI : Allocas) { | ||||
200 | if (handleAlloca(*AI, SufficientLDS)) | ||||
201 | Changed = true; | ||||
202 | } | ||||
203 | |||||
204 | return Changed; | ||||
205 | } | ||||
206 | |||||
207 | std::pair<Value *, Value *> | ||||
208 | AMDGPUPromoteAllocaImpl::getLocalSizeYZ(IRBuilder<> &Builder) { | ||||
209 | Function &F = *Builder.GetInsertBlock()->getParent(); | ||||
210 | const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(TM, F); | ||||
211 | |||||
212 | if (!IsAMDHSA) { | ||||
213 | Function *LocalSizeYFn | ||||
214 | = Intrinsic::getDeclaration(Mod, Intrinsic::r600_read_local_size_y); | ||||
215 | Function *LocalSizeZFn | ||||
216 | = Intrinsic::getDeclaration(Mod, Intrinsic::r600_read_local_size_z); | ||||
217 | |||||
218 | CallInst *LocalSizeY = Builder.CreateCall(LocalSizeYFn, {}); | ||||
219 | CallInst *LocalSizeZ = Builder.CreateCall(LocalSizeZFn, {}); | ||||
220 | |||||
221 | ST.makeLIDRangeMetadata(LocalSizeY); | ||||
222 | ST.makeLIDRangeMetadata(LocalSizeZ); | ||||
223 | |||||
224 | return std::pair(LocalSizeY, LocalSizeZ); | ||||
225 | } | ||||
226 | |||||
227 | // We must read the size out of the dispatch pointer. | ||||
228 | assert(IsAMDGCN)(static_cast <bool> (IsAMDGCN) ? void (0) : __assert_fail ("IsAMDGCN", "llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp" , 228, __extension__ __PRETTY_FUNCTION__)); | ||||
229 | |||||
230 | // We are indexing into this struct, and want to extract the workgroup_size_* | ||||
231 | // fields. | ||||
232 | // | ||||
233 | // typedef struct hsa_kernel_dispatch_packet_s { | ||||
234 | // uint16_t header; | ||||
235 | // uint16_t setup; | ||||
236 | // uint16_t workgroup_size_x ; | ||||
237 | // uint16_t workgroup_size_y; | ||||
238 | // uint16_t workgroup_size_z; | ||||
239 | // uint16_t reserved0; | ||||
240 | // uint32_t grid_size_x ; | ||||
241 | // uint32_t grid_size_y ; | ||||
242 | // uint32_t grid_size_z; | ||||
243 | // | ||||
244 | // uint32_t private_segment_size; | ||||
245 | // uint32_t group_segment_size; | ||||
246 | // uint64_t kernel_object; | ||||
247 | // | ||||
248 | // #ifdef HSA_LARGE_MODEL | ||||
249 | // void *kernarg_address; | ||||
250 | // #elif defined HSA_LITTLE_ENDIAN | ||||
251 | // void *kernarg_address; | ||||
252 | // uint32_t reserved1; | ||||
253 | // #else | ||||
254 | // uint32_t reserved1; | ||||
255 | // void *kernarg_address; | ||||
256 | // #endif | ||||
257 | // uint64_t reserved2; | ||||
258 | // hsa_signal_t completion_signal; // uint64_t wrapper | ||||
259 | // } hsa_kernel_dispatch_packet_t | ||||
260 | // | ||||
261 | Function *DispatchPtrFn | ||||
262 | = Intrinsic::getDeclaration(Mod, Intrinsic::amdgcn_dispatch_ptr); | ||||
263 | |||||
264 | CallInst *DispatchPtr = Builder.CreateCall(DispatchPtrFn, {}); | ||||
265 | DispatchPtr->addRetAttr(Attribute::NoAlias); | ||||
266 | DispatchPtr->addRetAttr(Attribute::NonNull); | ||||
267 | F.removeFnAttr("amdgpu-no-dispatch-ptr"); | ||||
268 | |||||
269 | // Size of the dispatch packet struct. | ||||
270 | DispatchPtr->addDereferenceableRetAttr(64); | ||||
271 | |||||
272 | Type *I32Ty = Type::getInt32Ty(Mod->getContext()); | ||||
273 | Value *CastDispatchPtr = Builder.CreateBitCast( | ||||
274 | DispatchPtr, PointerType::get(I32Ty, AMDGPUAS::CONSTANT_ADDRESS)); | ||||
275 | |||||
276 | // We could do a single 64-bit load here, but it's likely that the basic | ||||
277 | // 32-bit and extract sequence is already present, and it is probably easier | ||||
278 | // to CSE this. The loads should be mergeable later anyway. | ||||
279 | Value *GEPXY = Builder.CreateConstInBoundsGEP1_64(I32Ty, CastDispatchPtr, 1); | ||||
280 | LoadInst *LoadXY = Builder.CreateAlignedLoad(I32Ty, GEPXY, Align(4)); | ||||
281 | |||||
282 | Value *GEPZU = Builder.CreateConstInBoundsGEP1_64(I32Ty, CastDispatchPtr, 2); | ||||
283 | LoadInst *LoadZU = Builder.CreateAlignedLoad(I32Ty, GEPZU, Align(4)); | ||||
284 | |||||
285 | MDNode *MD = MDNode::get(Mod->getContext(), std::nullopt); | ||||
286 | LoadXY->setMetadata(LLVMContext::MD_invariant_load, MD); | ||||
287 | LoadZU->setMetadata(LLVMContext::MD_invariant_load, MD); | ||||
288 | ST.makeLIDRangeMetadata(LoadZU); | ||||
289 | |||||
290 | // Extract y component. Upper half of LoadZU should be zero already. | ||||
291 | Value *Y = Builder.CreateLShr(LoadXY, 16); | ||||
292 | |||||
293 | return std::pair(Y, LoadZU); | ||||
294 | } | ||||
295 | |||||
296 | Value *AMDGPUPromoteAllocaImpl::getWorkitemID(IRBuilder<> &Builder, | ||||
297 | unsigned N) { | ||||
298 | Function *F = Builder.GetInsertBlock()->getParent(); | ||||
299 | const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(TM, *F); | ||||
300 | Intrinsic::ID IntrID = Intrinsic::not_intrinsic; | ||||
301 | StringRef AttrName; | ||||
302 | |||||
303 | switch (N) { | ||||
304 | case 0: | ||||
305 | IntrID = IsAMDGCN ? (Intrinsic::ID)Intrinsic::amdgcn_workitem_id_x | ||||
306 | : (Intrinsic::ID)Intrinsic::r600_read_tidig_x; | ||||
307 | AttrName = "amdgpu-no-workitem-id-x"; | ||||
308 | break; | ||||
309 | case 1: | ||||
310 | IntrID = IsAMDGCN ? (Intrinsic::ID)Intrinsic::amdgcn_workitem_id_y | ||||
311 | : (Intrinsic::ID)Intrinsic::r600_read_tidig_y; | ||||
312 | AttrName = "amdgpu-no-workitem-id-y"; | ||||
313 | break; | ||||
314 | |||||
315 | case 2: | ||||
316 | IntrID = IsAMDGCN ? (Intrinsic::ID)Intrinsic::amdgcn_workitem_id_z | ||||
317 | : (Intrinsic::ID)Intrinsic::r600_read_tidig_z; | ||||
318 | AttrName = "amdgpu-no-workitem-id-z"; | ||||
319 | break; | ||||
320 | default: | ||||
321 | llvm_unreachable("invalid dimension")::llvm::llvm_unreachable_internal("invalid dimension", "llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp" , 321); | ||||
322 | } | ||||
323 | |||||
324 | Function *WorkitemIdFn = Intrinsic::getDeclaration(Mod, IntrID); | ||||
325 | CallInst *CI = Builder.CreateCall(WorkitemIdFn); | ||||
326 | ST.makeLIDRangeMetadata(CI); | ||||
327 | F->removeFnAttr(AttrName); | ||||
328 | |||||
329 | return CI; | ||||
330 | } | ||||
331 | |||||
332 | static FixedVectorType *arrayTypeToVecType(ArrayType *ArrayTy) { | ||||
333 | return FixedVectorType::get(ArrayTy->getElementType(), | ||||
334 | ArrayTy->getNumElements()); | ||||
335 | } | ||||
336 | |||||
337 | static Value * | ||||
338 | calculateVectorIndex(Value *Ptr, | ||||
339 | const std::map<GetElementPtrInst *, Value *> &GEPIdx) { | ||||
340 | auto *GEP = dyn_cast<GetElementPtrInst>(Ptr->stripPointerCasts()); | ||||
341 | if (!GEP) | ||||
342 | return ConstantInt::getNullValue(Type::getInt32Ty(Ptr->getContext())); | ||||
343 | |||||
344 | auto I = GEPIdx.find(GEP); | ||||
345 | assert(I != GEPIdx.end() && "Must have entry for GEP!")(static_cast <bool> (I != GEPIdx.end() && "Must have entry for GEP!" ) ? void (0) : __assert_fail ("I != GEPIdx.end() && \"Must have entry for GEP!\"" , "llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp", 345, __extension__ __PRETTY_FUNCTION__)); | ||||
346 | return I->second; | ||||
347 | } | ||||
348 | |||||
349 | static Value *GEPToVectorIndex(GetElementPtrInst *GEP, AllocaInst *Alloca, | ||||
350 | Type *VecElemTy, const DataLayout &DL) { | ||||
351 | // TODO: Extracting a "multiple of X" from a GEP might be a useful generic | ||||
352 | // helper. | ||||
353 | unsigned BW = DL.getIndexTypeSizeInBits(GEP->getType()); | ||||
354 | MapVector<Value *, APInt> VarOffsets; | ||||
355 | APInt ConstOffset(BW, 0); | ||||
356 | if (GEP->getPointerOperand()->stripPointerCasts() != Alloca || | ||||
357 | !GEP->collectOffset(DL, BW, VarOffsets, ConstOffset)) | ||||
358 | return nullptr; | ||||
359 | |||||
360 | unsigned VecElemSize = DL.getTypeAllocSize(VecElemTy); | ||||
361 | if (VarOffsets.size() > 1) | ||||
362 | return nullptr; | ||||
363 | |||||
364 | if (VarOffsets.size() == 1) { | ||||
365 | // Only handle cases where we don't need to insert extra arithmetic | ||||
366 | // instructions. | ||||
367 | const auto &VarOffset = VarOffsets.front(); | ||||
368 | if (!ConstOffset.isZero() || VarOffset.second != VecElemSize) | ||||
369 | return nullptr; | ||||
370 | return VarOffset.first; | ||||
371 | } | ||||
372 | |||||
373 | APInt Quot; | ||||
374 | uint64_t Rem; | ||||
375 | APInt::udivrem(ConstOffset, VecElemSize, Quot, Rem); | ||||
376 | if (Rem != 0) | ||||
377 | return nullptr; | ||||
378 | |||||
379 | return ConstantInt::get(GEP->getContext(), Quot); | ||||
380 | } | ||||
381 | |||||
382 | static bool tryPromoteAllocaToVector(AllocaInst *Alloca, const DataLayout &DL, | ||||
383 | unsigned MaxVGPRs) { | ||||
384 | |||||
385 | if (DisablePromoteAllocaToVector) { | ||||
| |||||
386 | LLVM_DEBUG(dbgs() << " Promotion alloca to vector is disabled\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << " Promotion alloca to vector is disabled\n" ; } } while (false); | ||||
387 | return false; | ||||
388 | } | ||||
389 | |||||
390 | Type *AllocaTy = Alloca->getAllocatedType(); | ||||
391 | auto *VectorTy = dyn_cast<FixedVectorType>(AllocaTy); | ||||
392 | if (auto *ArrayTy
| ||||
393 | if (VectorType::isValidElementType(ArrayTy->getElementType()) && | ||||
394 | ArrayTy->getNumElements() > 0) | ||||
395 | VectorTy = arrayTypeToVecType(ArrayTy); | ||||
396 | } | ||||
397 | |||||
398 | // Use up to 1/4 of available register budget for vectorization. | ||||
399 | unsigned Limit = PromoteAllocaToVectorLimit ? PromoteAllocaToVectorLimit * 8 | ||||
400 | : (MaxVGPRs * 32); | ||||
401 | |||||
402 | if (DL.getTypeSizeInBits(AllocaTy) * 4 > Limit) { | ||||
403 | LLVM_DEBUG(dbgs() << " Alloca too big for vectorization with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << " Alloca too big for vectorization with " << MaxVGPRs << " registers available\n"; } } while (false) | ||||
404 | << MaxVGPRs << " registers available\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << " Alloca too big for vectorization with " << MaxVGPRs << " registers available\n"; } } while (false); | ||||
405 | return false; | ||||
406 | } | ||||
407 | |||||
408 | LLVM_DEBUG(dbgs() << "Alloca candidate for vectorization\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << "Alloca candidate for vectorization\n" ; } } while (false); | ||||
409 | |||||
410 | // FIXME: There is no reason why we can't support larger arrays, we | ||||
411 | // are just being conservative for now. | ||||
412 | // FIXME: We also reject alloca's of the form [ 2 x [ 2 x i32 ]] or equivalent. Potentially these | ||||
413 | // could also be promoted but we don't currently handle this case | ||||
414 | if (!VectorTy || VectorTy->getNumElements() > 16 || | ||||
415 | VectorTy->getNumElements() < 2) { | ||||
416 | LLVM_DEBUG(dbgs() << " Cannot convert type to vector\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << " Cannot convert type to vector\n" ; } } while (false); | ||||
417 | return false; | ||||
418 | } | ||||
419 | |||||
420 | std::map<GetElementPtrInst*, Value*> GEPVectorIdx; | ||||
421 | SmallVector<Instruction *> WorkList; | ||||
422 | SmallVector<Use *, 8> Uses; | ||||
423 | for (Use &U : Alloca->uses()) | ||||
424 | Uses.push_back(&U); | ||||
425 | |||||
426 | Type *VecEltTy = VectorTy->getElementType(); | ||||
427 | while (!Uses.empty()) { | ||||
428 | Use *U = Uses.pop_back_val(); | ||||
429 | Instruction *Inst = dyn_cast<Instruction>(U->getUser()); | ||||
430 | |||||
431 | if (Value *Ptr
| ||||
432 | // This is a store of the pointer, not to the pointer. | ||||
433 | if (isa<StoreInst>(Inst) && | ||||
434 | U->getOperandNo() != StoreInst::getPointerOperandIndex()) | ||||
435 | return false; | ||||
436 | |||||
437 | Type *AccessTy = getLoadStoreType(Inst); | ||||
438 | Ptr = Ptr->stripPointerCasts(); | ||||
439 | |||||
440 | // Alloca already accessed as vector, leave alone. | ||||
441 | if (Ptr == Alloca && DL.getTypeStoreSize(Alloca->getAllocatedType()) == | ||||
442 | DL.getTypeStoreSize(AccessTy)) | ||||
443 | continue; | ||||
444 | |||||
445 | // Check that this is a simple access of a vector element. | ||||
446 | bool IsSimple = isa<LoadInst>(Inst) ? cast<LoadInst>(Inst)->isSimple() | ||||
447 | : cast<StoreInst>(Inst)->isSimple(); | ||||
448 | if (!IsSimple || | ||||
449 | !CastInst::isBitOrNoopPointerCastable(VecEltTy, AccessTy, DL)) | ||||
450 | return false; | ||||
451 | |||||
452 | WorkList.push_back(Inst); | ||||
453 | continue; | ||||
454 | } | ||||
455 | |||||
456 | if (isa<BitCastInst>(Inst)) { | ||||
457 | // Look through bitcasts. | ||||
458 | for (Use &U : Inst->uses()) | ||||
459 | Uses.push_back(&U); | ||||
460 | continue; | ||||
461 | } | ||||
462 | |||||
463 | if (auto *GEP = dyn_cast<GetElementPtrInst>(Inst)) { | ||||
464 | // If we can't compute a vector index from this GEP, then we can't | ||||
465 | // promote this alloca to vector. | ||||
466 | Value *Index = GEPToVectorIndex(GEP, Alloca, VecEltTy, DL); | ||||
467 | if (!Index) { | ||||
468 | LLVM_DEBUG(dbgs() << " Cannot compute vector index for GEP " << *GEPdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << " Cannot compute vector index for GEP " << *GEP << '\n'; } } while (false) | ||||
469 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << " Cannot compute vector index for GEP " << *GEP << '\n'; } } while (false); | ||||
470 | return false; | ||||
471 | } | ||||
472 | |||||
473 | GEPVectorIdx[GEP] = Index; | ||||
474 | for (Use &U : Inst->uses()) | ||||
| |||||
475 | Uses.push_back(&U); | ||||
476 | continue; | ||||
477 | } | ||||
478 | |||||
479 | // Ignore assume-like intrinsics and comparisons used in assumes. | ||||
480 | if (isAssumeLikeIntrinsic(Inst)) | ||||
481 | continue; | ||||
482 | |||||
483 | if (isa<ICmpInst>(Inst) && all_of(Inst->users(), [](User *U) { | ||||
484 | return isAssumeLikeIntrinsic(cast<Instruction>(U)); | ||||
485 | })) | ||||
486 | continue; | ||||
487 | |||||
488 | // Unknown user. | ||||
489 | return false; | ||||
490 | } | ||||
491 | |||||
492 | LLVM_DEBUG(dbgs() << " Converting alloca to vector " << *AllocaTy << " -> "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << " Converting alloca to vector " << *AllocaTy << " -> " << *VectorTy << '\n'; } } while (false) | ||||
493 | << *VectorTy << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << " Converting alloca to vector " << *AllocaTy << " -> " << *VectorTy << '\n'; } } while (false); | ||||
494 | |||||
495 | for (Instruction *Inst : WorkList) { | ||||
496 | IRBuilder<> Builder(Inst); | ||||
497 | switch (Inst->getOpcode()) { | ||||
498 | case Instruction::Load: { | ||||
499 | Value *Ptr = cast<LoadInst>(Inst)->getPointerOperand(); | ||||
500 | Value *Index = calculateVectorIndex(Ptr, GEPVectorIdx); | ||||
501 | Type *VecPtrTy = VectorTy->getPointerTo(Alloca->getAddressSpace()); | ||||
502 | Value *BitCast = Builder.CreateBitCast(Alloca, VecPtrTy); | ||||
503 | Value *VecValue = Builder.CreateLoad(VectorTy, BitCast); | ||||
504 | Value *ExtractElement = Builder.CreateExtractElement(VecValue, Index); | ||||
505 | if (Inst->getType() != VecEltTy) | ||||
506 | ExtractElement = Builder.CreateBitOrPointerCast(ExtractElement, Inst->getType()); | ||||
507 | Inst->replaceAllUsesWith(ExtractElement); | ||||
508 | Inst->eraseFromParent(); | ||||
509 | break; | ||||
510 | } | ||||
511 | case Instruction::Store: { | ||||
512 | StoreInst *SI = cast<StoreInst>(Inst); | ||||
513 | Value *Ptr = SI->getPointerOperand(); | ||||
514 | Value *Index = calculateVectorIndex(Ptr, GEPVectorIdx); | ||||
515 | Type *VecPtrTy = VectorTy->getPointerTo(Alloca->getAddressSpace()); | ||||
516 | Value *BitCast = Builder.CreateBitCast(Alloca, VecPtrTy); | ||||
517 | Value *VecValue = Builder.CreateLoad(VectorTy, BitCast); | ||||
518 | Value *Elt = SI->getValueOperand(); | ||||
519 | if (Elt->getType() != VecEltTy) | ||||
520 | Elt = Builder.CreateBitOrPointerCast(Elt, VecEltTy); | ||||
521 | Value *NewVecValue = Builder.CreateInsertElement(VecValue, Elt, Index); | ||||
522 | Builder.CreateStore(NewVecValue, BitCast); | ||||
523 | Inst->eraseFromParent(); | ||||
524 | break; | ||||
525 | } | ||||
526 | |||||
527 | default: | ||||
528 | llvm_unreachable("Inconsistency in instructions promotable to vector")::llvm::llvm_unreachable_internal("Inconsistency in instructions promotable to vector" , "llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp", 528); | ||||
529 | } | ||||
530 | } | ||||
531 | return true; | ||||
532 | } | ||||
533 | |||||
534 | static bool isCallPromotable(CallInst *CI) { | ||||
535 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI); | ||||
536 | if (!II) | ||||
537 | return false; | ||||
538 | |||||
539 | switch (II->getIntrinsicID()) { | ||||
540 | case Intrinsic::memcpy: | ||||
541 | case Intrinsic::memmove: | ||||
542 | case Intrinsic::memset: | ||||
543 | case Intrinsic::lifetime_start: | ||||
544 | case Intrinsic::lifetime_end: | ||||
545 | case Intrinsic::invariant_start: | ||||
546 | case Intrinsic::invariant_end: | ||||
547 | case Intrinsic::launder_invariant_group: | ||||
548 | case Intrinsic::strip_invariant_group: | ||||
549 | case Intrinsic::objectsize: | ||||
550 | return true; | ||||
551 | default: | ||||
552 | return false; | ||||
553 | } | ||||
554 | } | ||||
555 | |||||
556 | bool AMDGPUPromoteAllocaImpl::binaryOpIsDerivedFromSameAlloca( | ||||
557 | Value *BaseAlloca, Value *Val, Instruction *Inst, int OpIdx0, | ||||
558 | int OpIdx1) const { | ||||
559 | // Figure out which operand is the one we might not be promoting. | ||||
560 | Value *OtherOp = Inst->getOperand(OpIdx0); | ||||
561 | if (Val == OtherOp) | ||||
562 | OtherOp = Inst->getOperand(OpIdx1); | ||||
563 | |||||
564 | if (isa<ConstantPointerNull>(OtherOp)) | ||||
565 | return true; | ||||
566 | |||||
567 | Value *OtherObj = getUnderlyingObject(OtherOp); | ||||
568 | if (!isa<AllocaInst>(OtherObj)) | ||||
569 | return false; | ||||
570 | |||||
571 | // TODO: We should be able to replace undefs with the right pointer type. | ||||
572 | |||||
573 | // TODO: If we know the other base object is another promotable | ||||
574 | // alloca, not necessarily this alloca, we can do this. The | ||||
575 | // important part is both must have the same address space at | ||||
576 | // the end. | ||||
577 | if (OtherObj != BaseAlloca) { | ||||
578 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << "Found a binary instruction with another alloca object\n" ; } } while (false) | ||||
579 | dbgs() << "Found a binary instruction with another alloca object\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << "Found a binary instruction with another alloca object\n" ; } } while (false); | ||||
580 | return false; | ||||
581 | } | ||||
582 | |||||
583 | return true; | ||||
584 | } | ||||
585 | |||||
586 | bool AMDGPUPromoteAllocaImpl::collectUsesWithPtrTypes( | ||||
587 | Value *BaseAlloca, Value *Val, std::vector<Value *> &WorkList) const { | ||||
588 | |||||
589 | for (User *User : Val->users()) { | ||||
590 | if (is_contained(WorkList, User)) | ||||
591 | continue; | ||||
592 | |||||
593 | if (CallInst *CI = dyn_cast<CallInst>(User)) { | ||||
594 | if (!isCallPromotable(CI)) | ||||
595 | return false; | ||||
596 | |||||
597 | WorkList.push_back(User); | ||||
598 | continue; | ||||
599 | } | ||||
600 | |||||
601 | Instruction *UseInst = cast<Instruction>(User); | ||||
602 | if (UseInst->getOpcode() == Instruction::PtrToInt) | ||||
603 | return false; | ||||
604 | |||||
605 | if (LoadInst *LI = dyn_cast<LoadInst>(UseInst)) { | ||||
606 | if (LI->isVolatile()) | ||||
607 | return false; | ||||
608 | |||||
609 | continue; | ||||
610 | } | ||||
611 | |||||
612 | if (StoreInst *SI = dyn_cast<StoreInst>(UseInst)) { | ||||
613 | if (SI->isVolatile()) | ||||
614 | return false; | ||||
615 | |||||
616 | // Reject if the stored value is not the pointer operand. | ||||
617 | if (SI->getPointerOperand() != Val) | ||||
618 | return false; | ||||
619 | } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UseInst)) { | ||||
620 | if (RMW->isVolatile()) | ||||
621 | return false; | ||||
622 | } else if (AtomicCmpXchgInst *CAS = dyn_cast<AtomicCmpXchgInst>(UseInst)) { | ||||
623 | if (CAS->isVolatile()) | ||||
624 | return false; | ||||
625 | } | ||||
626 | |||||
627 | // Only promote a select if we know that the other select operand | ||||
628 | // is from another pointer that will also be promoted. | ||||
629 | if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) { | ||||
630 | if (!binaryOpIsDerivedFromSameAlloca(BaseAlloca, Val, ICmp, 0, 1)) | ||||
631 | return false; | ||||
632 | |||||
633 | // May need to rewrite constant operands. | ||||
634 | WorkList.push_back(ICmp); | ||||
635 | } | ||||
636 | |||||
637 | if (UseInst->getOpcode() == Instruction::AddrSpaceCast) { | ||||
638 | // Give up if the pointer may be captured. | ||||
639 | if (PointerMayBeCaptured(UseInst, true, true)) | ||||
640 | return false; | ||||
641 | // Don't collect the users of this. | ||||
642 | WorkList.push_back(User); | ||||
643 | continue; | ||||
644 | } | ||||
645 | |||||
646 | // Do not promote vector/aggregate type instructions. It is hard to track | ||||
647 | // their users. | ||||
648 | if (isa<InsertValueInst>(User) || isa<InsertElementInst>(User)) | ||||
649 | return false; | ||||
650 | |||||
651 | if (!User->getType()->isPointerTy()) | ||||
652 | continue; | ||||
653 | |||||
654 | if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(UseInst)) { | ||||
655 | // Be conservative if an address could be computed outside the bounds of | ||||
656 | // the alloca. | ||||
657 | if (!GEP->isInBounds()) | ||||
658 | return false; | ||||
659 | } | ||||
660 | |||||
661 | // Only promote a select if we know that the other select operand is from | ||||
662 | // another pointer that will also be promoted. | ||||
663 | if (SelectInst *SI = dyn_cast<SelectInst>(UseInst)) { | ||||
664 | if (!binaryOpIsDerivedFromSameAlloca(BaseAlloca, Val, SI, 1, 2)) | ||||
665 | return false; | ||||
666 | } | ||||
667 | |||||
668 | // Repeat for phis. | ||||
669 | if (PHINode *Phi = dyn_cast<PHINode>(UseInst)) { | ||||
670 | // TODO: Handle more complex cases. We should be able to replace loops | ||||
671 | // over arrays. | ||||
672 | switch (Phi->getNumIncomingValues()) { | ||||
673 | case 1: | ||||
674 | break; | ||||
675 | case 2: | ||||
676 | if (!binaryOpIsDerivedFromSameAlloca(BaseAlloca, Val, Phi, 0, 1)) | ||||
677 | return false; | ||||
678 | break; | ||||
679 | default: | ||||
680 | return false; | ||||
681 | } | ||||
682 | } | ||||
683 | |||||
684 | WorkList.push_back(User); | ||||
685 | if (!collectUsesWithPtrTypes(BaseAlloca, User, WorkList)) | ||||
686 | return false; | ||||
687 | } | ||||
688 | |||||
689 | return true; | ||||
690 | } | ||||
691 | |||||
692 | bool AMDGPUPromoteAllocaImpl::hasSufficientLocalMem(const Function &F) { | ||||
693 | |||||
694 | FunctionType *FTy = F.getFunctionType(); | ||||
695 | const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(TM, F); | ||||
696 | |||||
697 | // If the function has any arguments in the local address space, then it's | ||||
698 | // possible these arguments require the entire local memory space, so | ||||
699 | // we cannot use local memory in the pass. | ||||
700 | for (Type *ParamTy : FTy->params()) { | ||||
701 | PointerType *PtrTy = dyn_cast<PointerType>(ParamTy); | ||||
702 | if (PtrTy && PtrTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) { | ||||
703 | LocalMemLimit = 0; | ||||
704 | LLVM_DEBUG(dbgs() << "Function has local memory argument. Promoting to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << "Function has local memory argument. Promoting to " "local memory disabled.\n"; } } while (false) | ||||
705 | "local memory disabled.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << "Function has local memory argument. Promoting to " "local memory disabled.\n"; } } while (false); | ||||
706 | return false; | ||||
707 | } | ||||
708 | } | ||||
709 | |||||
710 | LocalMemLimit = ST.getLocalMemorySize(); | ||||
711 | if (LocalMemLimit == 0) | ||||
712 | return false; | ||||
713 | |||||
714 | SmallVector<const Constant *, 16> Stack; | ||||
715 | SmallPtrSet<const Constant *, 8> VisitedConstants; | ||||
716 | SmallPtrSet<const GlobalVariable *, 8> UsedLDS; | ||||
717 | |||||
718 | auto visitUsers = [&](const GlobalVariable *GV, const Constant *Val) -> bool { | ||||
719 | for (const User *U : Val->users()) { | ||||
720 | if (const Instruction *Use = dyn_cast<Instruction>(U)) { | ||||
721 | if (Use->getParent()->getParent() == &F) | ||||
722 | return true; | ||||
723 | } else { | ||||
724 | const Constant *C = cast<Constant>(U); | ||||
725 | if (VisitedConstants.insert(C).second) | ||||
726 | Stack.push_back(C); | ||||
727 | } | ||||
728 | } | ||||
729 | |||||
730 | return false; | ||||
731 | }; | ||||
732 | |||||
733 | for (GlobalVariable &GV : Mod->globals()) { | ||||
734 | if (GV.getAddressSpace() != AMDGPUAS::LOCAL_ADDRESS) | ||||
735 | continue; | ||||
736 | |||||
737 | if (visitUsers(&GV, &GV)) { | ||||
738 | UsedLDS.insert(&GV); | ||||
739 | Stack.clear(); | ||||
740 | continue; | ||||
741 | } | ||||
742 | |||||
743 | // For any ConstantExpr uses, we need to recursively search the users until | ||||
744 | // we see a function. | ||||
745 | while (!Stack.empty()) { | ||||
746 | const Constant *C = Stack.pop_back_val(); | ||||
747 | if (visitUsers(&GV, C)) { | ||||
748 | UsedLDS.insert(&GV); | ||||
749 | Stack.clear(); | ||||
750 | break; | ||||
751 | } | ||||
752 | } | ||||
753 | } | ||||
754 | |||||
755 | const DataLayout &DL = Mod->getDataLayout(); | ||||
756 | SmallVector<std::pair<uint64_t, Align>, 16> AllocatedSizes; | ||||
757 | AllocatedSizes.reserve(UsedLDS.size()); | ||||
758 | |||||
759 | for (const GlobalVariable *GV : UsedLDS) { | ||||
760 | Align Alignment = | ||||
761 | DL.getValueOrABITypeAlignment(GV->getAlign(), GV->getValueType()); | ||||
762 | uint64_t AllocSize = DL.getTypeAllocSize(GV->getValueType()); | ||||
763 | |||||
764 | // HIP uses an extern unsized array in local address space for dynamically | ||||
765 | // allocated shared memory. In that case, we have to disable the promotion. | ||||
766 | if (GV->hasExternalLinkage() && AllocSize == 0) { | ||||
767 | LocalMemLimit = 0; | ||||
768 | LLVM_DEBUG(dbgs() << "Function has a reference to externally allocated "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << "Function has a reference to externally allocated " "local memory. Promoting to local memory " "disabled.\n"; } } while (false) | ||||
769 | "local memory. Promoting to local memory "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << "Function has a reference to externally allocated " "local memory. Promoting to local memory " "disabled.\n"; } } while (false) | ||||
770 | "disabled.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << "Function has a reference to externally allocated " "local memory. Promoting to local memory " "disabled.\n"; } } while (false); | ||||
771 | return false; | ||||
772 | } | ||||
773 | |||||
774 | AllocatedSizes.emplace_back(AllocSize, Alignment); | ||||
775 | } | ||||
776 | |||||
777 | // Sort to try to estimate the worst case alignment padding | ||||
778 | // | ||||
779 | // FIXME: We should really do something to fix the addresses to a more optimal | ||||
780 | // value instead | ||||
781 | llvm::sort(AllocatedSizes, llvm::less_second()); | ||||
782 | |||||
783 | // Check how much local memory is being used by global objects | ||||
784 | CurrentLocalMemUsage = 0; | ||||
785 | |||||
786 | // FIXME: Try to account for padding here. The real padding and address is | ||||
787 | // currently determined from the inverse order of uses in the function when | ||||
788 | // legalizing, which could also potentially change. We try to estimate the | ||||
789 | // worst case here, but we probably should fix the addresses earlier. | ||||
790 | for (auto Alloc : AllocatedSizes) { | ||||
791 | CurrentLocalMemUsage = alignTo(CurrentLocalMemUsage, Alloc.second); | ||||
792 | CurrentLocalMemUsage += Alloc.first; | ||||
793 | } | ||||
794 | |||||
795 | unsigned MaxOccupancy = ST.getOccupancyWithLocalMemSize(CurrentLocalMemUsage, | ||||
796 | F); | ||||
797 | |||||
798 | // Restrict local memory usage so that we don't drastically reduce occupancy, | ||||
799 | // unless it is already significantly reduced. | ||||
800 | |||||
801 | // TODO: Have some sort of hint or other heuristics to guess occupancy based | ||||
802 | // on other factors.. | ||||
803 | unsigned OccupancyHint = ST.getWavesPerEU(F).second; | ||||
804 | if (OccupancyHint == 0) | ||||
805 | OccupancyHint = 7; | ||||
806 | |||||
807 | // Clamp to max value. | ||||
808 | OccupancyHint = std::min(OccupancyHint, ST.getMaxWavesPerEU()); | ||||
809 | |||||
810 | // Check the hint but ignore it if it's obviously wrong from the existing LDS | ||||
811 | // usage. | ||||
812 | MaxOccupancy = std::min(OccupancyHint, MaxOccupancy); | ||||
813 | |||||
814 | |||||
815 | // Round up to the next tier of usage. | ||||
816 | unsigned MaxSizeWithWaveCount | ||||
817 | = ST.getMaxLocalMemSizeWithWaveCount(MaxOccupancy, F); | ||||
818 | |||||
819 | // Program is possibly broken by using more local mem than available. | ||||
820 | if (CurrentLocalMemUsage > MaxSizeWithWaveCount) | ||||
821 | return false; | ||||
822 | |||||
823 | LocalMemLimit = MaxSizeWithWaveCount; | ||||
824 | |||||
825 | LLVM_DEBUG(dbgs() << F.getName() << " uses " << CurrentLocalMemUsagedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << F.getName() << " uses " << CurrentLocalMemUsage << " bytes of LDS\n" << " Rounding size to " << MaxSizeWithWaveCount << " with a maximum occupancy of " << MaxOccupancy << '\n' << " and " << (LocalMemLimit - CurrentLocalMemUsage ) << " available for promotion\n"; } } while (false) | ||||
826 | << " bytes of LDS\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << F.getName() << " uses " << CurrentLocalMemUsage << " bytes of LDS\n" << " Rounding size to " << MaxSizeWithWaveCount << " with a maximum occupancy of " << MaxOccupancy << '\n' << " and " << (LocalMemLimit - CurrentLocalMemUsage ) << " available for promotion\n"; } } while (false) | ||||
827 | << " Rounding size to " << MaxSizeWithWaveCountdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << F.getName() << " uses " << CurrentLocalMemUsage << " bytes of LDS\n" << " Rounding size to " << MaxSizeWithWaveCount << " with a maximum occupancy of " << MaxOccupancy << '\n' << " and " << (LocalMemLimit - CurrentLocalMemUsage ) << " available for promotion\n"; } } while (false) | ||||
828 | << " with a maximum occupancy of " << MaxOccupancy << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << F.getName() << " uses " << CurrentLocalMemUsage << " bytes of LDS\n" << " Rounding size to " << MaxSizeWithWaveCount << " with a maximum occupancy of " << MaxOccupancy << '\n' << " and " << (LocalMemLimit - CurrentLocalMemUsage ) << " available for promotion\n"; } } while (false) | ||||
829 | << " and " << (LocalMemLimit - CurrentLocalMemUsage)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << F.getName() << " uses " << CurrentLocalMemUsage << " bytes of LDS\n" << " Rounding size to " << MaxSizeWithWaveCount << " with a maximum occupancy of " << MaxOccupancy << '\n' << " and " << (LocalMemLimit - CurrentLocalMemUsage ) << " available for promotion\n"; } } while (false) | ||||
830 | << " available for promotion\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << F.getName() << " uses " << CurrentLocalMemUsage << " bytes of LDS\n" << " Rounding size to " << MaxSizeWithWaveCount << " with a maximum occupancy of " << MaxOccupancy << '\n' << " and " << (LocalMemLimit - CurrentLocalMemUsage ) << " available for promotion\n"; } } while (false); | ||||
831 | |||||
832 | return true; | ||||
833 | } | ||||
834 | |||||
835 | // FIXME: Should try to pick the most likely to be profitable allocas first. | ||||
836 | bool AMDGPUPromoteAllocaImpl::handleAlloca(AllocaInst &I, bool SufficientLDS) { | ||||
837 | // Array allocations are probably not worth handling, since an allocation of | ||||
838 | // the array type is the canonical form. | ||||
839 | if (!I.isStaticAlloca() || I.isArrayAllocation()) | ||||
840 | return false; | ||||
841 | |||||
842 | const DataLayout &DL = Mod->getDataLayout(); | ||||
843 | IRBuilder<> Builder(&I); | ||||
844 | |||||
845 | // First try to replace the alloca with a vector | ||||
846 | Type *AllocaTy = I.getAllocatedType(); | ||||
847 | |||||
848 | LLVM_DEBUG(dbgs() << "Trying to promote " << I << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << "Trying to promote " << I << '\n'; } } while (false); | ||||
849 | |||||
850 | if (tryPromoteAllocaToVector(&I, DL, MaxVGPRs)) | ||||
851 | return true; // Promoted to vector. | ||||
852 | |||||
853 | if (DisablePromoteAllocaToLDS) | ||||
854 | return false; | ||||
855 | |||||
856 | const Function &ContainingFunction = *I.getParent()->getParent(); | ||||
857 | CallingConv::ID CC = ContainingFunction.getCallingConv(); | ||||
858 | |||||
859 | // Don't promote the alloca to LDS for shader calling conventions as the work | ||||
860 | // item ID intrinsics are not supported for these calling conventions. | ||||
861 | // Furthermore not all LDS is available for some of the stages. | ||||
862 | switch (CC) { | ||||
863 | case CallingConv::AMDGPU_KERNEL: | ||||
864 | case CallingConv::SPIR_KERNEL: | ||||
865 | break; | ||||
866 | default: | ||||
867 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << " promote alloca to LDS not supported with calling convention.\n" ; } } while (false) | ||||
868 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << " promote alloca to LDS not supported with calling convention.\n" ; } } while (false) | ||||
869 | << " promote alloca to LDS not supported with calling convention.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << " promote alloca to LDS not supported with calling convention.\n" ; } } while (false); | ||||
870 | return false; | ||||
871 | } | ||||
872 | |||||
873 | // Not likely to have sufficient local memory for promotion. | ||||
874 | if (!SufficientLDS) | ||||
875 | return false; | ||||
876 | |||||
877 | const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(TM, ContainingFunction); | ||||
878 | unsigned WorkGroupSize = ST.getFlatWorkGroupSizes(ContainingFunction).second; | ||||
879 | |||||
880 | Align Alignment = | ||||
881 | DL.getValueOrABITypeAlignment(I.getAlign(), I.getAllocatedType()); | ||||
882 | |||||
883 | // FIXME: This computed padding is likely wrong since it depends on inverse | ||||
884 | // usage order. | ||||
885 | // | ||||
886 | // FIXME: It is also possible that if we're allowed to use all of the memory | ||||
887 | // could end up using more than the maximum due to alignment padding. | ||||
888 | |||||
889 | uint32_t NewSize = alignTo(CurrentLocalMemUsage, Alignment); | ||||
890 | uint32_t AllocSize = WorkGroupSize * DL.getTypeAllocSize(AllocaTy); | ||||
891 | NewSize += AllocSize; | ||||
892 | |||||
893 | if (NewSize > LocalMemLimit) { | ||||
894 | LLVM_DEBUG(dbgs() << " " << AllocSizedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << " " << AllocSize << " bytes of local memory not available to promote\n" ; } } while (false) | ||||
895 | << " bytes of local memory not available to promote\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << " " << AllocSize << " bytes of local memory not available to promote\n" ; } } while (false); | ||||
896 | return false; | ||||
897 | } | ||||
898 | |||||
899 | CurrentLocalMemUsage = NewSize; | ||||
900 | |||||
901 | std::vector<Value*> WorkList; | ||||
902 | |||||
903 | if (!collectUsesWithPtrTypes(&I, &I, WorkList)) { | ||||
904 | LLVM_DEBUG(dbgs() << " Do not know how to convert all uses\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << " Do not know how to convert all uses\n" ; } } while (false); | ||||
905 | return false; | ||||
906 | } | ||||
907 | |||||
908 | LLVM_DEBUG(dbgs() << "Promoting alloca to local memory\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << "Promoting alloca to local memory\n" ; } } while (false); | ||||
909 | |||||
910 | Function *F = I.getParent()->getParent(); | ||||
911 | |||||
912 | Type *GVTy = ArrayType::get(I.getAllocatedType(), WorkGroupSize); | ||||
913 | GlobalVariable *GV = new GlobalVariable( | ||||
914 | *Mod, GVTy, false, GlobalValue::InternalLinkage, PoisonValue::get(GVTy), | ||||
915 | Twine(F->getName()) + Twine('.') + I.getName(), nullptr, | ||||
916 | GlobalVariable::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS); | ||||
917 | GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); | ||||
918 | GV->setAlignment(I.getAlign()); | ||||
919 | |||||
920 | Value *TCntY, *TCntZ; | ||||
921 | |||||
922 | std::tie(TCntY, TCntZ) = getLocalSizeYZ(Builder); | ||||
923 | Value *TIdX = getWorkitemID(Builder, 0); | ||||
924 | Value *TIdY = getWorkitemID(Builder, 1); | ||||
925 | Value *TIdZ = getWorkitemID(Builder, 2); | ||||
926 | |||||
927 | Value *Tmp0 = Builder.CreateMul(TCntY, TCntZ, "", true, true); | ||||
928 | Tmp0 = Builder.CreateMul(Tmp0, TIdX); | ||||
929 | Value *Tmp1 = Builder.CreateMul(TIdY, TCntZ, "", true, true); | ||||
930 | Value *TID = Builder.CreateAdd(Tmp0, Tmp1); | ||||
931 | TID = Builder.CreateAdd(TID, TIdZ); | ||||
932 | |||||
933 | Value *Indices[] = { | ||||
934 | Constant::getNullValue(Type::getInt32Ty(Mod->getContext())), | ||||
935 | TID | ||||
936 | }; | ||||
937 | |||||
938 | Value *Offset = Builder.CreateInBoundsGEP(GVTy, GV, Indices); | ||||
939 | I.mutateType(Offset->getType()); | ||||
940 | I.replaceAllUsesWith(Offset); | ||||
941 | I.eraseFromParent(); | ||||
942 | |||||
943 | SmallVector<IntrinsicInst *> DeferredIntrs; | ||||
944 | |||||
945 | for (Value *V : WorkList) { | ||||
946 | CallInst *Call = dyn_cast<CallInst>(V); | ||||
947 | if (!Call) { | ||||
948 | if (ICmpInst *CI = dyn_cast<ICmpInst>(V)) { | ||||
949 | Value *Src0 = CI->getOperand(0); | ||||
950 | PointerType *NewTy = PointerType::getWithSamePointeeType( | ||||
951 | cast<PointerType>(Src0->getType()), AMDGPUAS::LOCAL_ADDRESS); | ||||
952 | |||||
953 | if (isa<ConstantPointerNull>(CI->getOperand(0))) | ||||
954 | CI->setOperand(0, ConstantPointerNull::get(NewTy)); | ||||
955 | |||||
956 | if (isa<ConstantPointerNull>(CI->getOperand(1))) | ||||
957 | CI->setOperand(1, ConstantPointerNull::get(NewTy)); | ||||
958 | |||||
959 | continue; | ||||
960 | } | ||||
961 | |||||
962 | // The operand's value should be corrected on its own and we don't want to | ||||
963 | // touch the users. | ||||
964 | if (isa<AddrSpaceCastInst>(V)) | ||||
965 | continue; | ||||
966 | |||||
967 | PointerType *NewTy = PointerType::getWithSamePointeeType( | ||||
968 | cast<PointerType>(V->getType()), AMDGPUAS::LOCAL_ADDRESS); | ||||
969 | |||||
970 | // FIXME: It doesn't really make sense to try to do this for all | ||||
971 | // instructions. | ||||
972 | V->mutateType(NewTy); | ||||
973 | |||||
974 | // Adjust the types of any constant operands. | ||||
975 | if (SelectInst *SI = dyn_cast<SelectInst>(V)) { | ||||
976 | if (isa<ConstantPointerNull>(SI->getOperand(1))) | ||||
977 | SI->setOperand(1, ConstantPointerNull::get(NewTy)); | ||||
978 | |||||
979 | if (isa<ConstantPointerNull>(SI->getOperand(2))) | ||||
980 | SI->setOperand(2, ConstantPointerNull::get(NewTy)); | ||||
981 | } else if (PHINode *Phi = dyn_cast<PHINode>(V)) { | ||||
982 | for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) { | ||||
983 | if (isa<ConstantPointerNull>(Phi->getIncomingValue(I))) | ||||
984 | Phi->setIncomingValue(I, ConstantPointerNull::get(NewTy)); | ||||
985 | } | ||||
986 | } | ||||
987 | |||||
988 | continue; | ||||
989 | } | ||||
990 | |||||
991 | IntrinsicInst *Intr = cast<IntrinsicInst>(Call); | ||||
992 | Builder.SetInsertPoint(Intr); | ||||
993 | switch (Intr->getIntrinsicID()) { | ||||
994 | case Intrinsic::lifetime_start: | ||||
995 | case Intrinsic::lifetime_end: | ||||
996 | // These intrinsics are for address space 0 only | ||||
997 | Intr->eraseFromParent(); | ||||
998 | continue; | ||||
999 | case Intrinsic::memcpy: | ||||
1000 | case Intrinsic::memmove: | ||||
1001 | // These have 2 pointer operands. In case if second pointer also needs | ||||
1002 | // to be replaced we defer processing of these intrinsics until all | ||||
1003 | // other values are processed. | ||||
1004 | DeferredIntrs.push_back(Intr); | ||||
1005 | continue; | ||||
1006 | case Intrinsic::memset: { | ||||
1007 | MemSetInst *MemSet = cast<MemSetInst>(Intr); | ||||
1008 | Builder.CreateMemSet( | ||||
1009 | MemSet->getRawDest(), MemSet->getValue(), MemSet->getLength(), | ||||
1010 | MaybeAlign(MemSet->getDestAlignment()), MemSet->isVolatile()); | ||||
1011 | Intr->eraseFromParent(); | ||||
1012 | continue; | ||||
1013 | } | ||||
1014 | case Intrinsic::invariant_start: | ||||
1015 | case Intrinsic::invariant_end: | ||||
1016 | case Intrinsic::launder_invariant_group: | ||||
1017 | case Intrinsic::strip_invariant_group: | ||||
1018 | Intr->eraseFromParent(); | ||||
1019 | // FIXME: I think the invariant marker should still theoretically apply, | ||||
1020 | // but the intrinsics need to be changed to accept pointers with any | ||||
1021 | // address space. | ||||
1022 | continue; | ||||
1023 | case Intrinsic::objectsize: { | ||||
1024 | Value *Src = Intr->getOperand(0); | ||||
1025 | Function *ObjectSize = Intrinsic::getDeclaration( | ||||
1026 | Mod, Intrinsic::objectsize, | ||||
1027 | {Intr->getType(), | ||||
1028 | PointerType::getWithSamePointeeType( | ||||
1029 | cast<PointerType>(Src->getType()), AMDGPUAS::LOCAL_ADDRESS)}); | ||||
1030 | |||||
1031 | CallInst *NewCall = Builder.CreateCall( | ||||
1032 | ObjectSize, | ||||
1033 | {Src, Intr->getOperand(1), Intr->getOperand(2), Intr->getOperand(3)}); | ||||
1034 | Intr->replaceAllUsesWith(NewCall); | ||||
1035 | Intr->eraseFromParent(); | ||||
1036 | continue; | ||||
1037 | } | ||||
1038 | default: | ||||
1039 | Intr->print(errs()); | ||||
1040 | llvm_unreachable("Don't know how to promote alloca intrinsic use.")::llvm::llvm_unreachable_internal("Don't know how to promote alloca intrinsic use." , "llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp", 1040); | ||||
1041 | } | ||||
1042 | } | ||||
1043 | |||||
1044 | for (IntrinsicInst *Intr : DeferredIntrs) { | ||||
1045 | Builder.SetInsertPoint(Intr); | ||||
1046 | Intrinsic::ID ID = Intr->getIntrinsicID(); | ||||
1047 | assert(ID == Intrinsic::memcpy || ID == Intrinsic::memmove)(static_cast <bool> (ID == Intrinsic::memcpy || ID == Intrinsic ::memmove) ? void (0) : __assert_fail ("ID == Intrinsic::memcpy || ID == Intrinsic::memmove" , "llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp", 1047, __extension__ __PRETTY_FUNCTION__)); | ||||
1048 | |||||
1049 | MemTransferInst *MI = cast<MemTransferInst>(Intr); | ||||
1050 | auto *B = | ||||
1051 | Builder.CreateMemTransferInst(ID, MI->getRawDest(), MI->getDestAlign(), | ||||
1052 | MI->getRawSource(), MI->getSourceAlign(), | ||||
1053 | MI->getLength(), MI->isVolatile()); | ||||
1054 | |||||
1055 | for (unsigned I = 0; I != 2; ++I) { | ||||
1056 | if (uint64_t Bytes = Intr->getParamDereferenceableBytes(I)) { | ||||
1057 | B->addDereferenceableParamAttr(I, Bytes); | ||||
1058 | } | ||||
1059 | } | ||||
1060 | |||||
1061 | Intr->eraseFromParent(); | ||||
1062 | } | ||||
1063 | |||||
1064 | return true; | ||||
1065 | } | ||||
1066 | |||||
1067 | bool handlePromoteAllocaToVector(AllocaInst &I, unsigned MaxVGPRs) { | ||||
1068 | // Array allocations are probably not worth handling, since an allocation of | ||||
1069 | // the array type is the canonical form. | ||||
1070 | if (!I.isStaticAlloca() || I.isArrayAllocation()) | ||||
1071 | return false; | ||||
1072 | |||||
1073 | LLVM_DEBUG(dbgs() << "Trying to promote " << I << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("amdgpu-promote-alloca")) { dbgs() << "Trying to promote " << I << '\n'; } } while (false); | ||||
1074 | |||||
1075 | Module *Mod = I.getParent()->getParent()->getParent(); | ||||
1076 | return tryPromoteAllocaToVector(&I, Mod->getDataLayout(), MaxVGPRs); | ||||
1077 | } | ||||
1078 | |||||
1079 | bool promoteAllocasToVector(Function &F, TargetMachine &TM) { | ||||
1080 | if (DisablePromoteAllocaToVector) | ||||
1081 | return false; | ||||
1082 | |||||
1083 | const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(TM, F); | ||||
1084 | if (!ST.isPromoteAllocaEnabled()) | ||||
1085 | return false; | ||||
1086 | |||||
1087 | unsigned MaxVGPRs; | ||||
1088 | if (TM.getTargetTriple().getArch() == Triple::amdgcn) { | ||||
1089 | const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F); | ||||
1090 | MaxVGPRs = ST.getMaxNumVGPRs(ST.getWavesPerEU(F).first); | ||||
1091 | // A non-entry function has only 32 caller preserved registers. | ||||
1092 | // Do not promote alloca which will force spilling. | ||||
1093 | if (!AMDGPU::isEntryFunctionCC(F.getCallingConv())) | ||||
1094 | MaxVGPRs = std::min(MaxVGPRs, 32u); | ||||
1095 | } else { | ||||
1096 | MaxVGPRs = 128; | ||||
1097 | } | ||||
1098 | |||||
1099 | bool Changed = false; | ||||
1100 | BasicBlock &EntryBB = *F.begin(); | ||||
1101 | |||||
1102 | SmallVector<AllocaInst *, 16> Allocas; | ||||
1103 | for (Instruction &I : EntryBB) { | ||||
1104 | if (AllocaInst *AI = dyn_cast<AllocaInst>(&I)) | ||||
1105 | Allocas.push_back(AI); | ||||
1106 | } | ||||
1107 | |||||
1108 | for (AllocaInst *AI : Allocas) { | ||||
1109 | if (handlePromoteAllocaToVector(*AI, MaxVGPRs)) | ||||
1110 | Changed = true; | ||||
1111 | } | ||||
1112 | |||||
1113 | return Changed; | ||||
1114 | } | ||||
1115 | |||||
1116 | bool AMDGPUPromoteAllocaToVector::runOnFunction(Function &F) { | ||||
1117 | if (skipFunction(F)) | ||||
1118 | return false; | ||||
1119 | if (auto *TPC = getAnalysisIfAvailable<TargetPassConfig>()) { | ||||
1120 | return promoteAllocasToVector(F, TPC->getTM<TargetMachine>()); | ||||
1121 | } | ||||
1122 | return false; | ||||
1123 | } | ||||
1124 | |||||
1125 | PreservedAnalyses | ||||
1126 | AMDGPUPromoteAllocaToVectorPass::run(Function &F, FunctionAnalysisManager &AM) { | ||||
1127 | bool Changed = promoteAllocasToVector(F, TM); | ||||
1128 | if (Changed) { | ||||
1129 | PreservedAnalyses PA; | ||||
1130 | PA.preserveSet<CFGAnalyses>(); | ||||
1131 | return PA; | ||||
1132 | } | ||||
1133 | return PreservedAnalyses::all(); | ||||
1134 | } | ||||
1135 | |||||
1136 | FunctionPass *llvm::createAMDGPUPromoteAlloca() { | ||||
1137 | return new AMDGPUPromoteAlloca(); | ||||
1138 | } | ||||
1139 | |||||
1140 | FunctionPass *llvm::createAMDGPUPromoteAllocaToVector() { | ||||
1141 | return new AMDGPUPromoteAllocaToVector(); | ||||
1142 | } |
1 | //===- llvm/Instructions.h - Instruction subclass definitions ---*- C++ -*-===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file exposes the class definitions of all of the subclasses of the |
10 | // Instruction class. This is meant to be an easy way to get access to all |
11 | // instruction subclasses. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #ifndef LLVM_IR_INSTRUCTIONS_H |
16 | #define LLVM_IR_INSTRUCTIONS_H |
17 | |
18 | #include "llvm/ADT/ArrayRef.h" |
19 | #include "llvm/ADT/Bitfields.h" |
20 | #include "llvm/ADT/MapVector.h" |
21 | #include "llvm/ADT/STLExtras.h" |
22 | #include "llvm/ADT/SmallVector.h" |
23 | #include "llvm/ADT/Twine.h" |
24 | #include "llvm/ADT/iterator.h" |
25 | #include "llvm/ADT/iterator_range.h" |
26 | #include "llvm/IR/CFG.h" |
27 | #include "llvm/IR/Constant.h" |
28 | #include "llvm/IR/DerivedTypes.h" |
29 | #include "llvm/IR/InstrTypes.h" |
30 | #include "llvm/IR/Instruction.h" |
31 | #include "llvm/IR/OperandTraits.h" |
32 | #include "llvm/IR/Use.h" |
33 | #include "llvm/IR/User.h" |
34 | #include "llvm/Support/AtomicOrdering.h" |
35 | #include "llvm/Support/ErrorHandling.h" |
36 | #include <cassert> |
37 | #include <cstddef> |
38 | #include <cstdint> |
39 | #include <iterator> |
40 | #include <optional> |
41 | |
42 | namespace llvm { |
43 | |
44 | class APFloat; |
45 | class APInt; |
46 | class BasicBlock; |
47 | class ConstantInt; |
48 | class DataLayout; |
49 | class StringRef; |
50 | class Type; |
51 | class Value; |
52 | |
53 | //===----------------------------------------------------------------------===// |
54 | // AllocaInst Class |
55 | //===----------------------------------------------------------------------===// |
56 | |
57 | /// an instruction to allocate memory on the stack |
58 | class AllocaInst : public UnaryInstruction { |
59 | Type *AllocatedType; |
60 | |
61 | using AlignmentField = AlignmentBitfieldElementT<0>; |
62 | using UsedWithInAllocaField = BoolBitfieldElementT<AlignmentField::NextBit>; |
63 | using SwiftErrorField = BoolBitfieldElementT<UsedWithInAllocaField::NextBit>; |
64 | static_assert(Bitfield::areContiguous<AlignmentField, UsedWithInAllocaField, |
65 | SwiftErrorField>(), |
66 | "Bitfields must be contiguous"); |
67 | |
68 | protected: |
69 | // Note: Instruction needs to be a friend here to call cloneImpl. |
70 | friend class Instruction; |
71 | |
72 | AllocaInst *cloneImpl() const; |
73 | |
74 | public: |
75 | explicit AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, |
76 | const Twine &Name, Instruction *InsertBefore); |
77 | AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, |
78 | const Twine &Name, BasicBlock *InsertAtEnd); |
79 | |
80 | AllocaInst(Type *Ty, unsigned AddrSpace, const Twine &Name, |
81 | Instruction *InsertBefore); |
82 | AllocaInst(Type *Ty, unsigned AddrSpace, |
83 | const Twine &Name, BasicBlock *InsertAtEnd); |
84 | |
85 | AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align, |
86 | const Twine &Name = "", Instruction *InsertBefore = nullptr); |
87 | AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align, |
88 | const Twine &Name, BasicBlock *InsertAtEnd); |
89 | |
90 | /// Return true if there is an allocation size parameter to the allocation |
91 | /// instruction that is not 1. |
92 | bool isArrayAllocation() const; |
93 | |
94 | /// Get the number of elements allocated. For a simple allocation of a single |
95 | /// element, this will return a constant 1 value. |
96 | const Value *getArraySize() const { return getOperand(0); } |
97 | Value *getArraySize() { return getOperand(0); } |
98 | |
99 | /// Overload to return most specific pointer type. |
100 | PointerType *getType() const { |
101 | return cast<PointerType>(Instruction::getType()); |
102 | } |
103 | |
104 | /// Return the address space for the allocation. |
105 | unsigned getAddressSpace() const { |
106 | return getType()->getAddressSpace(); |
107 | } |
108 | |
109 | /// Get allocation size in bits. Returns std::nullopt if size can't be |
110 | /// determined, e.g. in case of a VLA. |
111 | std::optional<TypeSize> getAllocationSizeInBits(const DataLayout &DL) const; |
112 | |
113 | /// Return the type that is being allocated by the instruction. |
114 | Type *getAllocatedType() const { return AllocatedType; } |
115 | /// for use only in special circumstances that need to generically |
116 | /// transform a whole instruction (eg: IR linking and vectorization). |
117 | void setAllocatedType(Type *Ty) { AllocatedType = Ty; } |
118 | |
119 | /// Return the alignment of the memory that is being allocated by the |
120 | /// instruction. |
121 | Align getAlign() const { |
122 | return Align(1ULL << getSubclassData<AlignmentField>()); |
123 | } |
124 | |
125 | void setAlignment(Align Align) { |
126 | setSubclassData<AlignmentField>(Log2(Align)); |
127 | } |
128 | |
129 | /// Return true if this alloca is in the entry block of the function and is a |
130 | /// constant size. If so, the code generator will fold it into the |
131 | /// prolog/epilog code, so it is basically free. |
132 | bool isStaticAlloca() const; |
133 | |
134 | /// Return true if this alloca is used as an inalloca argument to a call. Such |
135 | /// allocas are never considered static even if they are in the entry block. |
136 | bool isUsedWithInAlloca() const { |
137 | return getSubclassData<UsedWithInAllocaField>(); |
138 | } |
139 | |
140 | /// Specify whether this alloca is used to represent the arguments to a call. |
141 | void setUsedWithInAlloca(bool V) { |
142 | setSubclassData<UsedWithInAllocaField>(V); |
143 | } |
144 | |
145 | /// Return true if this alloca is used as a swifterror argument to a call. |
146 | bool isSwiftError() const { return getSubclassData<SwiftErrorField>(); } |
147 | /// Specify whether this alloca is used to represent a swifterror. |
148 | void setSwiftError(bool V) { setSubclassData<SwiftErrorField>(V); } |
149 | |
150 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
151 | static bool classof(const Instruction *I) { |
152 | return (I->getOpcode() == Instruction::Alloca); |
153 | } |
154 | static bool classof(const Value *V) { |
155 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
156 | } |
157 | |
158 | private: |
159 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
160 | // method so that subclasses cannot accidentally use it. |
161 | template <typename Bitfield> |
162 | void setSubclassData(typename Bitfield::Type Value) { |
163 | Instruction::setSubclassData<Bitfield>(Value); |
164 | } |
165 | }; |
166 | |
167 | //===----------------------------------------------------------------------===// |
168 | // LoadInst Class |
169 | //===----------------------------------------------------------------------===// |
170 | |
171 | /// An instruction for reading from memory. This uses the SubclassData field in |
172 | /// Value to store whether or not the load is volatile. |
173 | class LoadInst : public UnaryInstruction { |
174 | using VolatileField = BoolBitfieldElementT<0>; |
175 | using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>; |
176 | using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>; |
177 | static_assert( |
178 | Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(), |
179 | "Bitfields must be contiguous"); |
180 | |
181 | void AssertOK(); |
182 | |
183 | protected: |
184 | // Note: Instruction needs to be a friend here to call cloneImpl. |
185 | friend class Instruction; |
186 | |
187 | LoadInst *cloneImpl() const; |
188 | |
189 | public: |
190 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, |
191 | Instruction *InsertBefore); |
192 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, BasicBlock *InsertAtEnd); |
193 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
194 | Instruction *InsertBefore); |
195 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
196 | BasicBlock *InsertAtEnd); |
197 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
198 | Align Align, Instruction *InsertBefore = nullptr); |
199 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
200 | Align Align, BasicBlock *InsertAtEnd); |
201 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
202 | Align Align, AtomicOrdering Order, |
203 | SyncScope::ID SSID = SyncScope::System, |
204 | Instruction *InsertBefore = nullptr); |
205 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
206 | Align Align, AtomicOrdering Order, SyncScope::ID SSID, |
207 | BasicBlock *InsertAtEnd); |
208 | |
209 | /// Return true if this is a load from a volatile memory location. |
210 | bool isVolatile() const { return getSubclassData<VolatileField>(); } |
211 | |
212 | /// Specify whether this is a volatile load or not. |
213 | void setVolatile(bool V) { setSubclassData<VolatileField>(V); } |
214 | |
215 | /// Return the alignment of the access that is being performed. |
216 | Align getAlign() const { |
217 | return Align(1ULL << (getSubclassData<AlignmentField>())); |
218 | } |
219 | |
220 | void setAlignment(Align Align) { |
221 | setSubclassData<AlignmentField>(Log2(Align)); |
222 | } |
223 | |
224 | /// Returns the ordering constraint of this load instruction. |
225 | AtomicOrdering getOrdering() const { |
226 | return getSubclassData<OrderingField>(); |
227 | } |
228 | /// Sets the ordering constraint of this load instruction. May not be Release |
229 | /// or AcquireRelease. |
230 | void setOrdering(AtomicOrdering Ordering) { |
231 | setSubclassData<OrderingField>(Ordering); |
232 | } |
233 | |
234 | /// Returns the synchronization scope ID of this load instruction. |
235 | SyncScope::ID getSyncScopeID() const { |
236 | return SSID; |
237 | } |
238 | |
239 | /// Sets the synchronization scope ID of this load instruction. |
240 | void setSyncScopeID(SyncScope::ID SSID) { |
241 | this->SSID = SSID; |
242 | } |
243 | |
244 | /// Sets the ordering constraint and the synchronization scope ID of this load |
245 | /// instruction. |
246 | void setAtomic(AtomicOrdering Ordering, |
247 | SyncScope::ID SSID = SyncScope::System) { |
248 | setOrdering(Ordering); |
249 | setSyncScopeID(SSID); |
250 | } |
251 | |
252 | bool isSimple() const { return !isAtomic() && !isVolatile(); } |
253 | |
254 | bool isUnordered() const { |
255 | return (getOrdering() == AtomicOrdering::NotAtomic || |
256 | getOrdering() == AtomicOrdering::Unordered) && |
257 | !isVolatile(); |
258 | } |
259 | |
260 | Value *getPointerOperand() { return getOperand(0); } |
261 | const Value *getPointerOperand() const { return getOperand(0); } |
262 | static unsigned getPointerOperandIndex() { return 0U; } |
263 | Type *getPointerOperandType() const { return getPointerOperand()->getType(); } |
264 | |
265 | /// Returns the address space of the pointer operand. |
266 | unsigned getPointerAddressSpace() const { |
267 | return getPointerOperandType()->getPointerAddressSpace(); |
268 | } |
269 | |
270 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
271 | static bool classof(const Instruction *I) { |
272 | return I->getOpcode() == Instruction::Load; |
273 | } |
274 | static bool classof(const Value *V) { |
275 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
276 | } |
277 | |
278 | private: |
279 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
280 | // method so that subclasses cannot accidentally use it. |
281 | template <typename Bitfield> |
282 | void setSubclassData(typename Bitfield::Type Value) { |
283 | Instruction::setSubclassData<Bitfield>(Value); |
284 | } |
285 | |
286 | /// The synchronization scope ID of this load instruction. Not quite enough |
287 | /// room in SubClassData for everything, so synchronization scope ID gets its |
288 | /// own field. |
289 | SyncScope::ID SSID; |
290 | }; |
291 | |
292 | //===----------------------------------------------------------------------===// |
293 | // StoreInst Class |
294 | //===----------------------------------------------------------------------===// |
295 | |
296 | /// An instruction for storing to memory. |
297 | class StoreInst : public Instruction { |
298 | using VolatileField = BoolBitfieldElementT<0>; |
299 | using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>; |
300 | using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>; |
301 | static_assert( |
302 | Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(), |
303 | "Bitfields must be contiguous"); |
304 | |
305 | void AssertOK(); |
306 | |
307 | protected: |
308 | // Note: Instruction needs to be a friend here to call cloneImpl. |
309 | friend class Instruction; |
310 | |
311 | StoreInst *cloneImpl() const; |
312 | |
313 | public: |
314 | StoreInst(Value *Val, Value *Ptr, Instruction *InsertBefore); |
315 | StoreInst(Value *Val, Value *Ptr, BasicBlock *InsertAtEnd); |
316 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Instruction *InsertBefore); |
317 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, BasicBlock *InsertAtEnd); |
318 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
319 | Instruction *InsertBefore = nullptr); |
320 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
321 | BasicBlock *InsertAtEnd); |
322 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
323 | AtomicOrdering Order, SyncScope::ID SSID = SyncScope::System, |
324 | Instruction *InsertBefore = nullptr); |
325 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
326 | AtomicOrdering Order, SyncScope::ID SSID, BasicBlock *InsertAtEnd); |
327 | |
328 | // allocate space for exactly two operands |
329 | void *operator new(size_t S) { return User::operator new(S, 2); } |
330 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
331 | |
332 | /// Return true if this is a store to a volatile memory location. |
333 | bool isVolatile() const { return getSubclassData<VolatileField>(); } |
334 | |
335 | /// Specify whether this is a volatile store or not. |
336 | void setVolatile(bool V) { setSubclassData<VolatileField>(V); } |
337 | |
338 | /// Transparently provide more efficient getOperand methods. |
339 | 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; |
340 | |
341 | Align getAlign() const { |
342 | return Align(1ULL << (getSubclassData<AlignmentField>())); |
343 | } |
344 | |
345 | void setAlignment(Align Align) { |
346 | setSubclassData<AlignmentField>(Log2(Align)); |
347 | } |
348 | |
349 | /// Returns the ordering constraint of this store instruction. |
350 | AtomicOrdering getOrdering() const { |
351 | return getSubclassData<OrderingField>(); |
352 | } |
353 | |
354 | /// Sets the ordering constraint of this store instruction. May not be |
355 | /// Acquire or AcquireRelease. |
356 | void setOrdering(AtomicOrdering Ordering) { |
357 | setSubclassData<OrderingField>(Ordering); |
358 | } |
359 | |
360 | /// Returns the synchronization scope ID of this store instruction. |
361 | SyncScope::ID getSyncScopeID() const { |
362 | return SSID; |
363 | } |
364 | |
365 | /// Sets the synchronization scope ID of this store instruction. |
366 | void setSyncScopeID(SyncScope::ID SSID) { |
367 | this->SSID = SSID; |
368 | } |
369 | |
370 | /// Sets the ordering constraint and the synchronization scope ID of this |
371 | /// store instruction. |
372 | void setAtomic(AtomicOrdering Ordering, |
373 | SyncScope::ID SSID = SyncScope::System) { |
374 | setOrdering(Ordering); |
375 | setSyncScopeID(SSID); |
376 | } |
377 | |
378 | bool isSimple() const { return !isAtomic() && !isVolatile(); } |
379 | |
380 | bool isUnordered() const { |
381 | return (getOrdering() == AtomicOrdering::NotAtomic || |
382 | getOrdering() == AtomicOrdering::Unordered) && |
383 | !isVolatile(); |
384 | } |
385 | |
386 | Value *getValueOperand() { return getOperand(0); } |
387 | const Value *getValueOperand() const { return getOperand(0); } |
388 | |
389 | Value *getPointerOperand() { return getOperand(1); } |
390 | const Value *getPointerOperand() const { return getOperand(1); } |
391 | static unsigned getPointerOperandIndex() { return 1U; } |
392 | Type *getPointerOperandType() const { return getPointerOperand()->getType(); } |
393 | |
394 | /// Returns the address space of the pointer operand. |
395 | unsigned getPointerAddressSpace() const { |
396 | return getPointerOperandType()->getPointerAddressSpace(); |
397 | } |
398 | |
399 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
400 | static bool classof(const Instruction *I) { |
401 | return I->getOpcode() == Instruction::Store; |
402 | } |
403 | static bool classof(const Value *V) { |
404 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
405 | } |
406 | |
407 | private: |
408 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
409 | // method so that subclasses cannot accidentally use it. |
410 | template <typename Bitfield> |
411 | void setSubclassData(typename Bitfield::Type Value) { |
412 | Instruction::setSubclassData<Bitfield>(Value); |
413 | } |
414 | |
415 | /// The synchronization scope ID of this store instruction. Not quite enough |
416 | /// room in SubClassData for everything, so synchronization scope ID gets its |
417 | /// own field. |
418 | SyncScope::ID SSID; |
419 | }; |
420 | |
421 | template <> |
422 | struct OperandTraits<StoreInst> : public FixedNumOperandTraits<StoreInst, 2> { |
423 | }; |
424 | |
425 | 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 { (static_cast <bool> (i_nocapture < OperandTraits<StoreInst>::operands(this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<StoreInst>::operands(this) && \"getOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 425, __extension__ __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) { (static_cast <bool> (i_nocapture < OperandTraits<StoreInst>::operands(this) && "setOperand() out of range!" ) ? void (0) : __assert_fail ("i_nocapture < OperandTraits<StoreInst>::operands(this) && \"setOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 425, __extension__ __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); } |
426 | |
427 | //===----------------------------------------------------------------------===// |
428 | // FenceInst Class |
429 | //===----------------------------------------------------------------------===// |
430 | |
431 | /// An instruction for ordering other memory operations. |
432 | class FenceInst : public Instruction { |
433 | using OrderingField = AtomicOrderingBitfieldElementT<0>; |
434 | |
435 | void Init(AtomicOrdering Ordering, SyncScope::ID SSID); |
436 | |
437 | protected: |
438 | // Note: Instruction needs to be a friend here to call cloneImpl. |
439 | friend class Instruction; |
440 | |
441 | FenceInst *cloneImpl() const; |
442 | |
443 | public: |
444 | // Ordering may only be Acquire, Release, AcquireRelease, or |
445 | // SequentiallyConsistent. |
446 | FenceInst(LLVMContext &C, AtomicOrdering Ordering, |
447 | SyncScope::ID SSID = SyncScope::System, |
448 | Instruction *InsertBefore = nullptr); |
449 | FenceInst(LLVMContext &C, AtomicOrdering Ordering, SyncScope::ID SSID, |
450 | BasicBlock *InsertAtEnd); |
451 | |
452 | // allocate space for exactly zero operands |
453 | void *operator new(size_t S) { return User::operator new(S, 0); } |
454 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
455 | |
456 | /// Returns the ordering constraint of this fence instruction. |
457 | AtomicOrdering getOrdering() const { |
458 | return getSubclassData<OrderingField>(); |
459 | } |
460 | |
461 | /// Sets the ordering constraint of this fence instruction. May only be |
462 | /// Acquire, Release, AcquireRelease, or SequentiallyConsistent. |
463 | void setOrdering(AtomicOrdering Ordering) { |
464 | setSubclassData<OrderingField>(Ordering); |
465 | } |
466 | |
467 | /// Returns the synchronization scope ID of this fence instruction. |
468 | SyncScope::ID getSyncScopeID() const { |
469 | return SSID; |
470 | } |
471 | |
472 | /// Sets the synchronization scope ID of this fence instruction. |
473 | void setSyncScopeID(SyncScope::ID SSID) { |
474 | this->SSID = SSID; |
475 | } |
476 | |
477 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
478 | static bool classof(const Instruction *I) { |
479 | return I->getOpcode() == Instruction::Fence; |
480 | } |
481 | static bool classof(const Value *V) { |
482 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
483 | } |
484 | |
485 | private: |
486 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
487 | // method so that subclasses cannot accidentally use it. |
488 | template <typename Bitfield> |
489 | void setSubclassData(typename Bitfield::Type Value) { |
490 | Instruction::setSubclassData<Bitfield>(Value); |
491 | } |
492 | |
493 | /// The synchronization scope ID of this fence instruction. Not quite enough |
494 | /// room in SubClassData for everything, so synchronization scope ID gets its |
495 | /// own field. |
496 | SyncScope::ID SSID; |
497 | }; |
498 | |
499 | //===----------------------------------------------------------------------===// |
500 | // AtomicCmpXchgInst Class |
501 | //===----------------------------------------------------------------------===// |
502 | |
503 | /// An instruction that atomically checks whether a |
504 | /// specified value is in a memory location, and, if it is, stores a new value |
505 | /// there. The value returned by this instruction is a pair containing the |
506 | /// original value as first element, and an i1 indicating success (true) or |
507 | /// failure (false) as second element. |
508 | /// |
509 | class AtomicCmpXchgInst : public Instruction { |
510 | void Init(Value *Ptr, Value *Cmp, Value *NewVal, Align Align, |
511 | AtomicOrdering SuccessOrdering, AtomicOrdering FailureOrdering, |
512 | SyncScope::ID SSID); |
513 | |
514 | template <unsigned Offset> |
515 | using AtomicOrderingBitfieldElement = |
516 | typename Bitfield::Element<AtomicOrdering, Offset, 3, |
517 | AtomicOrdering::LAST>; |
518 | |
519 | protected: |
520 | // Note: Instruction needs to be a friend here to call cloneImpl. |
521 | friend class Instruction; |
522 | |
523 | AtomicCmpXchgInst *cloneImpl() const; |
524 | |
525 | public: |
526 | AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment, |
527 | AtomicOrdering SuccessOrdering, |
528 | AtomicOrdering FailureOrdering, SyncScope::ID SSID, |
529 | Instruction *InsertBefore = nullptr); |
530 | AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment, |
531 | AtomicOrdering SuccessOrdering, |
532 | AtomicOrdering FailureOrdering, SyncScope::ID SSID, |
533 | BasicBlock *InsertAtEnd); |
534 | |
535 | // allocate space for exactly three operands |
536 | void *operator new(size_t S) { return User::operator new(S, 3); } |
537 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
538 | |
539 | using VolatileField = BoolBitfieldElementT<0>; |
540 | using WeakField = BoolBitfieldElementT<VolatileField::NextBit>; |
541 | using SuccessOrderingField = |
542 | AtomicOrderingBitfieldElementT<WeakField::NextBit>; |
543 | using FailureOrderingField = |
544 | AtomicOrderingBitfieldElementT<SuccessOrderingField::NextBit>; |
545 | using AlignmentField = |
546 | AlignmentBitfieldElementT<FailureOrderingField::NextBit>; |
547 | static_assert( |
548 | Bitfield::areContiguous<VolatileField, WeakField, SuccessOrderingField, |
549 | FailureOrderingField, AlignmentField>(), |
550 | "Bitfields must be contiguous"); |
551 | |
552 | /// Return the alignment of the memory that is being allocated by the |
553 | /// instruction. |
554 | Align getAlign() const { |
555 | return Align(1ULL << getSubclassData<AlignmentField>()); |
556 | } |
557 | |
558 | void setAlignment(Align Align) { |
559 | setSubclassData<AlignmentField>(Log2(Align)); |
560 | } |
561 | |
562 | /// Return true if this is a cmpxchg from a volatile memory |
563 | /// location. |
564 | /// |
565 | bool isVolatile() const { return getSubclassData<VolatileField>(); } |
566 | |
567 | /// Specify whether this is a volatile cmpxchg. |
568 | /// |
569 | void setVolatile(bool V) { setSubclassData<VolatileField>(V); } |
570 | |
571 | /// Return true if this cmpxchg may spuriously fail. |
572 | bool isWeak() const { return getSubclassData<WeakField>(); } |
573 | |
574 | void setWeak(bool IsWeak) { setSubclassData<WeakField>(IsWeak); } |
575 | |
576 | /// Transparently provide more efficient getOperand methods. |
577 | 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; |
578 | |
579 | static bool isValidSuccessOrdering(AtomicOrdering Ordering) { |
580 | return Ordering != AtomicOrdering::NotAtomic && |
581 | Ordering != AtomicOrdering::Unordered; |
582 | } |
583 | |
584 | static bool isValidFailureOrdering(AtomicOrdering Ordering) { |
585 | return Ordering != AtomicOrdering::NotAtomic && |
586 | Ordering != AtomicOrdering::Unordered && |
587 | Ordering != AtomicOrdering::AcquireRelease && |
588 | Ordering != AtomicOrdering::Release; |
589 | } |
590 | |
591 | /// Returns the success ordering constraint of this cmpxchg instruction. |
592 | AtomicOrdering getSuccessOrdering() const { |
593 | return getSubclassData<SuccessOrderingField>(); |
594 | } |
595 | |
596 | /// Sets the success ordering constraint of this cmpxchg instruction. |
597 | void setSuccessOrdering(AtomicOrdering Ordering) { |
598 | assert(isValidSuccessOrdering(Ordering) &&(static_cast <bool> (isValidSuccessOrdering(Ordering) && "invalid CmpXchg success ordering") ? void (0) : __assert_fail ("isValidSuccessOrdering(Ordering) && \"invalid CmpXchg success ordering\"" , "llvm/include/llvm/IR/Instructions.h", 599, __extension__ __PRETTY_FUNCTION__ )) |
599 | "invalid CmpXchg success ordering")(static_cast <bool> (isValidSuccessOrdering(Ordering) && "invalid CmpXchg success ordering") ? void (0) : __assert_fail ("isValidSuccessOrdering(Ordering) && \"invalid CmpXchg success ordering\"" , "llvm/include/llvm/IR/Instructions.h", 599, __extension__ __PRETTY_FUNCTION__ )); |
600 | setSubclassData<SuccessOrderingField>(Ordering); |
601 | } |
602 | |
603 | /// Returns the failure ordering constraint of this cmpxchg instruction. |
604 | AtomicOrdering getFailureOrdering() const { |
605 | return getSubclassData<FailureOrderingField>(); |
606 | } |
607 | |
608 | /// Sets the failure ordering constraint of this cmpxchg instruction. |
609 | void setFailureOrdering(AtomicOrdering Ordering) { |
610 | assert(isValidFailureOrdering(Ordering) &&(static_cast <bool> (isValidFailureOrdering(Ordering) && "invalid CmpXchg failure ordering") ? void (0) : __assert_fail ("isValidFailureOrdering(Ordering) && \"invalid CmpXchg failure ordering\"" , "llvm/include/llvm/IR/Instructions.h", 611, __extension__ __PRETTY_FUNCTION__ )) |
611 | "invalid CmpXchg failure ordering")(static_cast <bool> (isValidFailureOrdering(Ordering) && "invalid CmpXchg failure ordering") ? void (0) : __assert_fail ("isValidFailureOrdering(Ordering) && \"invalid CmpXchg failure ordering\"" , "llvm/include/llvm/IR/Instructions.h", 611, __extension__ __PRETTY_FUNCTION__ )); |
612 | setSubclassData<FailureOrderingField>(Ordering); |
613 | } |
614 | |
615 | /// Returns a single ordering which is at least as strong as both the |
616 | /// success and failure orderings for this cmpxchg. |
617 | AtomicOrdering getMergedOrdering() const { |
618 | if (getFailureOrdering() == AtomicOrdering::SequentiallyConsistent) |
619 | return AtomicOrdering::SequentiallyConsistent; |
620 | if (getFailureOrdering() == AtomicOrdering::Acquire) { |
621 | if (getSuccessOrdering() == AtomicOrdering::Monotonic) |
622 | return AtomicOrdering::Acquire; |
623 | if (getSuccessOrdering() == AtomicOrdering::Release) |
624 | return AtomicOrdering::AcquireRelease; |
625 | } |
626 | return getSuccessOrdering(); |
627 | } |
628 | |
629 | /// Returns the synchronization scope ID of this cmpxchg instruction. |
630 | SyncScope::ID getSyncScopeID() const { |
631 | return SSID; |
632 | } |
633 | |
634 | /// Sets the synchronization scope ID of this cmpxchg instruction. |
635 | void setSyncScopeID(SyncScope::ID SSID) { |
636 | this->SSID = SSID; |
637 | } |
638 | |
639 | Value *getPointerOperand() { return getOperand(0); } |
640 | const Value *getPointerOperand() const { return getOperand(0); } |
641 | static unsigned getPointerOperandIndex() { return 0U; } |
642 | |
643 | Value *getCompareOperand() { return getOperand(1); } |
644 | const Value *getCompareOperand() const { return getOperand(1); } |
645 | |
646 | Value *getNewValOperand() { return getOperand(2); } |
647 | const Value *getNewValOperand() const { return getOperand(2); } |
648 | |
649 | /// Returns the address space of the pointer operand. |
650 | unsigned getPointerAddressSpace() const { |
651 | return getPointerOperand()->getType()->getPointerAddressSpace(); |
652 | } |
653 | |
654 | /// Returns the strongest permitted ordering on failure, given the |
655 | /// desired ordering on success. |
656 | /// |
657 | /// If the comparison in a cmpxchg operation fails, there is no atomic store |
658 | /// so release semantics cannot be provided. So this function drops explicit |
659 | /// Release requests from the AtomicOrdering. A SequentiallyConsistent |
660 | /// operation would remain SequentiallyConsistent. |
661 | static AtomicOrdering |
662 | getStrongestFailureOrdering(AtomicOrdering SuccessOrdering) { |
663 | switch (SuccessOrdering) { |
664 | default: |
665 | llvm_unreachable("invalid cmpxchg success ordering")::llvm::llvm_unreachable_internal("invalid cmpxchg success ordering" , "llvm/include/llvm/IR/Instructions.h", 665); |
666 | case AtomicOrdering::Release: |
667 | case AtomicOrdering::Monotonic: |
668 | return AtomicOrdering::Monotonic; |
669 | case AtomicOrdering::AcquireRelease: |
670 | case AtomicOrdering::Acquire: |
671 | return AtomicOrdering::Acquire; |
672 | case AtomicOrdering::SequentiallyConsistent: |
673 | return AtomicOrdering::SequentiallyConsistent; |
674 | } |
675 | } |
676 | |
677 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
678 | static bool classof(const Instruction *I) { |
679 | return I->getOpcode() == Instruction::AtomicCmpXchg; |
680 | } |
681 | static bool classof(const Value *V) { |
682 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
683 | } |
684 | |
685 | private: |
686 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
687 | // method so that subclasses cannot accidentally use it. |
688 | template <typename Bitfield> |
689 | void setSubclassData(typename Bitfield::Type Value) { |
690 | Instruction::setSubclassData<Bitfield>(Value); |
691 | } |
692 | |
693 | /// The synchronization scope ID of this cmpxchg instruction. Not quite |
694 | /// enough room in SubClassData for everything, so synchronization scope ID |
695 | /// gets its own field. |
696 | SyncScope::ID SSID; |
697 | }; |
698 | |
699 | template <> |
700 | struct OperandTraits<AtomicCmpXchgInst> : |
701 | public FixedNumOperandTraits<AtomicCmpXchgInst, 3> { |
702 | }; |
703 | |
704 | 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 { (static_cast <bool> (i_nocapture < OperandTraits<AtomicCmpXchgInst>::operands(this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicCmpXchgInst>::operands(this) && \"getOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 704, __extension__ __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) { (static_cast <bool> (i_nocapture < OperandTraits<AtomicCmpXchgInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicCmpXchgInst>::operands(this) && \"setOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 704, __extension__ __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 ); } |
705 | |
706 | //===----------------------------------------------------------------------===// |
707 | // AtomicRMWInst Class |
708 | //===----------------------------------------------------------------------===// |
709 | |
710 | /// an instruction that atomically reads a memory location, |
711 | /// combines it with another value, and then stores the result back. Returns |
712 | /// the old value. |
713 | /// |
714 | class AtomicRMWInst : public Instruction { |
715 | protected: |
716 | // Note: Instruction needs to be a friend here to call cloneImpl. |
717 | friend class Instruction; |
718 | |
719 | AtomicRMWInst *cloneImpl() const; |
720 | |
721 | public: |
722 | /// This enumeration lists the possible modifications atomicrmw can make. In |
723 | /// the descriptions, 'p' is the pointer to the instruction's memory location, |
724 | /// 'old' is the initial value of *p, and 'v' is the other value passed to the |
725 | /// instruction. These instructions always return 'old'. |
726 | enum BinOp : unsigned { |
727 | /// *p = v |
728 | Xchg, |
729 | /// *p = old + v |
730 | Add, |
731 | /// *p = old - v |
732 | Sub, |
733 | /// *p = old & v |
734 | And, |
735 | /// *p = ~(old & v) |
736 | Nand, |
737 | /// *p = old | v |
738 | Or, |
739 | /// *p = old ^ v |
740 | Xor, |
741 | /// *p = old >signed v ? old : v |
742 | Max, |
743 | /// *p = old <signed v ? old : v |
744 | Min, |
745 | /// *p = old >unsigned v ? old : v |
746 | UMax, |
747 | /// *p = old <unsigned v ? old : v |
748 | UMin, |
749 | |
750 | /// *p = old + v |
751 | FAdd, |
752 | |
753 | /// *p = old - v |
754 | FSub, |
755 | |
756 | /// *p = maxnum(old, v) |
757 | /// \p maxnum matches the behavior of \p llvm.maxnum.*. |
758 | FMax, |
759 | |
760 | /// *p = minnum(old, v) |
761 | /// \p minnum matches the behavior of \p llvm.minnum.*. |
762 | FMin, |
763 | |
764 | FIRST_BINOP = Xchg, |
765 | LAST_BINOP = FMin, |
766 | BAD_BINOP |
767 | }; |
768 | |
769 | private: |
770 | template <unsigned Offset> |
771 | using AtomicOrderingBitfieldElement = |
772 | typename Bitfield::Element<AtomicOrdering, Offset, 3, |
773 | AtomicOrdering::LAST>; |
774 | |
775 | template <unsigned Offset> |
776 | using BinOpBitfieldElement = |
777 | typename Bitfield::Element<BinOp, Offset, 4, BinOp::LAST_BINOP>; |
778 | |
779 | public: |
780 | AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment, |
781 | AtomicOrdering Ordering, SyncScope::ID SSID, |
782 | Instruction *InsertBefore = nullptr); |
783 | AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment, |
784 | AtomicOrdering Ordering, SyncScope::ID SSID, |
785 | BasicBlock *InsertAtEnd); |
786 | |
787 | // allocate space for exactly two operands |
788 | void *operator new(size_t S) { return User::operator new(S, 2); } |
789 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
790 | |
791 | using VolatileField = BoolBitfieldElementT<0>; |
792 | using AtomicOrderingField = |
793 | AtomicOrderingBitfieldElementT<VolatileField::NextBit>; |
794 | using OperationField = BinOpBitfieldElement<AtomicOrderingField::NextBit>; |
795 | using AlignmentField = AlignmentBitfieldElementT<OperationField::NextBit>; |
796 | static_assert(Bitfield::areContiguous<VolatileField, AtomicOrderingField, |
797 | OperationField, AlignmentField>(), |
798 | "Bitfields must be contiguous"); |
799 | |
800 | BinOp getOperation() const { return getSubclassData<OperationField>(); } |
801 | |
802 | static StringRef getOperationName(BinOp Op); |
803 | |
804 | static bool isFPOperation(BinOp Op) { |
805 | switch (Op) { |
806 | case AtomicRMWInst::FAdd: |
807 | case AtomicRMWInst::FSub: |
808 | case AtomicRMWInst::FMax: |
809 | case AtomicRMWInst::FMin: |
810 | return true; |
811 | default: |
812 | return false; |
813 | } |
814 | } |
815 | |
816 | void setOperation(BinOp Operation) { |
817 | setSubclassData<OperationField>(Operation); |
818 | } |
819 | |
820 | /// Return the alignment of the memory that is being allocated by the |
821 | /// instruction. |
822 | Align getAlign() const { |
823 | return Align(1ULL << getSubclassData<AlignmentField>()); |
824 | } |
825 | |
826 | void setAlignment(Align Align) { |
827 | setSubclassData<AlignmentField>(Log2(Align)); |
828 | } |
829 | |
830 | /// Return true if this is a RMW on a volatile memory location. |
831 | /// |
832 | bool isVolatile() const { return getSubclassData<VolatileField>(); } |
833 | |
834 | /// Specify whether this is a volatile RMW or not. |
835 | /// |
836 | void setVolatile(bool V) { setSubclassData<VolatileField>(V); } |
837 | |
838 | /// Transparently provide more efficient getOperand methods. |
839 | 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; |
840 | |
841 | /// Returns the ordering constraint of this rmw instruction. |
842 | AtomicOrdering getOrdering() const { |
843 | return getSubclassData<AtomicOrderingField>(); |
844 | } |
845 | |
846 | /// Sets the ordering constraint of this rmw instruction. |
847 | void setOrdering(AtomicOrdering Ordering) { |
848 | assert(Ordering != AtomicOrdering::NotAtomic &&(static_cast <bool> (Ordering != AtomicOrdering::NotAtomic && "atomicrmw instructions can only be atomic.") ? void (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"atomicrmw instructions can only be atomic.\"" , "llvm/include/llvm/IR/Instructions.h", 849, __extension__ __PRETTY_FUNCTION__ )) |
849 | "atomicrmw instructions can only be atomic.")(static_cast <bool> (Ordering != AtomicOrdering::NotAtomic && "atomicrmw instructions can only be atomic.") ? void (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"atomicrmw instructions can only be atomic.\"" , "llvm/include/llvm/IR/Instructions.h", 849, __extension__ __PRETTY_FUNCTION__ )); |
850 | assert(Ordering != AtomicOrdering::Unordered &&(static_cast <bool> (Ordering != AtomicOrdering::Unordered && "atomicrmw instructions cannot be unordered.") ? void (0) : __assert_fail ("Ordering != AtomicOrdering::Unordered && \"atomicrmw instructions cannot be unordered.\"" , "llvm/include/llvm/IR/Instructions.h", 851, __extension__ __PRETTY_FUNCTION__ )) |
851 | "atomicrmw instructions cannot be unordered.")(static_cast <bool> (Ordering != AtomicOrdering::Unordered && "atomicrmw instructions cannot be unordered.") ? void (0) : __assert_fail ("Ordering != AtomicOrdering::Unordered && \"atomicrmw instructions cannot be unordered.\"" , "llvm/include/llvm/IR/Instructions.h", 851, __extension__ __PRETTY_FUNCTION__ )); |
852 | setSubclassData<AtomicOrderingField>(Ordering); |
853 | } |
854 | |
855 | /// Returns the synchronization scope ID of this rmw instruction. |
856 | SyncScope::ID getSyncScopeID() const { |
857 | return SSID; |
858 | } |
859 | |
860 | /// Sets the synchronization scope ID of this rmw instruction. |
861 | void setSyncScopeID(SyncScope::ID SSID) { |
862 | this->SSID = SSID; |
863 | } |
864 | |
865 | Value *getPointerOperand() { return getOperand(0); } |
866 | const Value *getPointerOperand() const { return getOperand(0); } |
867 | static unsigned getPointerOperandIndex() { return 0U; } |
868 | |
869 | Value *getValOperand() { return getOperand(1); } |
870 | const Value *getValOperand() const { return getOperand(1); } |
871 | |
872 | /// Returns the address space of the pointer operand. |
873 | unsigned getPointerAddressSpace() const { |
874 | return getPointerOperand()->getType()->getPointerAddressSpace(); |
875 | } |
876 | |
877 | bool isFloatingPointOperation() const { |
878 | return isFPOperation(getOperation()); |
879 | } |
880 | |
881 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
882 | static bool classof(const Instruction *I) { |
883 | return I->getOpcode() == Instruction::AtomicRMW; |
884 | } |
885 | static bool classof(const Value *V) { |
886 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
887 | } |
888 | |
889 | private: |
890 | void Init(BinOp Operation, Value *Ptr, Value *Val, Align Align, |
891 | AtomicOrdering Ordering, SyncScope::ID SSID); |
892 | |
893 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
894 | // method so that subclasses cannot accidentally use it. |
895 | template <typename Bitfield> |
896 | void setSubclassData(typename Bitfield::Type Value) { |
897 | Instruction::setSubclassData<Bitfield>(Value); |
898 | } |
899 | |
900 | /// The synchronization scope ID of this rmw instruction. Not quite enough |
901 | /// room in SubClassData for everything, so synchronization scope ID gets its |
902 | /// own field. |
903 | SyncScope::ID SSID; |
904 | }; |
905 | |
906 | template <> |
907 | struct OperandTraits<AtomicRMWInst> |
908 | : public FixedNumOperandTraits<AtomicRMWInst,2> { |
909 | }; |
910 | |
911 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(AtomicRMWInst, Value)AtomicRMWInst::op_iterator AtomicRMWInst::op_begin() { return OperandTraits<AtomicRMWInst>::op_begin(this); } AtomicRMWInst ::const_op_iterator AtomicRMWInst::op_begin() const { return OperandTraits <AtomicRMWInst>::op_begin(const_cast<AtomicRMWInst*> (this)); } AtomicRMWInst::op_iterator AtomicRMWInst::op_end() { return OperandTraits<AtomicRMWInst>::op_end(this); } AtomicRMWInst::const_op_iterator AtomicRMWInst::op_end() const { return OperandTraits<AtomicRMWInst>::op_end(const_cast <AtomicRMWInst*>(this)); } Value *AtomicRMWInst::getOperand (unsigned i_nocapture) const { (static_cast <bool> (i_nocapture < OperandTraits<AtomicRMWInst>::operands(this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicRMWInst>::operands(this) && \"getOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 911, __extension__ __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) { (static_cast <bool> (i_nocapture < OperandTraits<AtomicRMWInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicRMWInst>::operands(this) && \"setOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 911, __extension__ __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 ); } |
912 | |
913 | //===----------------------------------------------------------------------===// |
914 | // GetElementPtrInst Class |
915 | //===----------------------------------------------------------------------===// |
916 | |
917 | // checkGEPType - Simple wrapper function to give a better assertion failure |
918 | // message on bad indexes for a gep instruction. |
919 | // |
920 | inline Type *checkGEPType(Type *Ty) { |
921 | assert(Ty && "Invalid GetElementPtrInst indices for type!")(static_cast <bool> (Ty && "Invalid GetElementPtrInst indices for type!" ) ? void (0) : __assert_fail ("Ty && \"Invalid GetElementPtrInst indices for type!\"" , "llvm/include/llvm/IR/Instructions.h", 921, __extension__ __PRETTY_FUNCTION__ )); |
922 | return Ty; |
923 | } |
924 | |
925 | /// an instruction for type-safe pointer arithmetic to |
926 | /// access elements of arrays and structs |
927 | /// |
928 | class GetElementPtrInst : public Instruction { |
929 | Type *SourceElementType; |
930 | Type *ResultElementType; |
931 | |
932 | GetElementPtrInst(const GetElementPtrInst &GEPI); |
933 | |
934 | /// Constructors - Create a getelementptr instruction with a base pointer an |
935 | /// list of indices. The first ctor can optionally insert before an existing |
936 | /// instruction, the second appends the new instruction to the specified |
937 | /// BasicBlock. |
938 | inline GetElementPtrInst(Type *PointeeType, Value *Ptr, |
939 | ArrayRef<Value *> IdxList, unsigned Values, |
940 | const Twine &NameStr, Instruction *InsertBefore); |
941 | inline GetElementPtrInst(Type *PointeeType, Value *Ptr, |
942 | ArrayRef<Value *> IdxList, unsigned Values, |
943 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
944 | |
945 | void init(Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr); |
946 | |
947 | protected: |
948 | // Note: Instruction needs to be a friend here to call cloneImpl. |
949 | friend class Instruction; |
950 | |
951 | GetElementPtrInst *cloneImpl() const; |
952 | |
953 | public: |
954 | static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr, |
955 | ArrayRef<Value *> IdxList, |
956 | const Twine &NameStr = "", |
957 | Instruction *InsertBefore = nullptr) { |
958 | unsigned Values = 1 + unsigned(IdxList.size()); |
959 | assert(PointeeType && "Must specify element type")(static_cast <bool> (PointeeType && "Must specify element type" ) ? void (0) : __assert_fail ("PointeeType && \"Must specify element type\"" , "llvm/include/llvm/IR/Instructions.h", 959, __extension__ __PRETTY_FUNCTION__ )); |
960 | assert(cast<PointerType>(Ptr->getType()->getScalarType())(static_cast <bool> (cast<PointerType>(Ptr->getType ()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType )) ? void (0) : __assert_fail ("cast<PointerType>(Ptr->getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType)" , "llvm/include/llvm/IR/Instructions.h", 961, __extension__ __PRETTY_FUNCTION__ )) |
961 | ->isOpaqueOrPointeeTypeMatches(PointeeType))(static_cast <bool> (cast<PointerType>(Ptr->getType ()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType )) ? void (0) : __assert_fail ("cast<PointerType>(Ptr->getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType)" , "llvm/include/llvm/IR/Instructions.h", 961, __extension__ __PRETTY_FUNCTION__ )); |
962 | return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values, |
963 | NameStr, InsertBefore); |
964 | } |
965 | |
966 | static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr, |
967 | ArrayRef<Value *> IdxList, |
968 | const Twine &NameStr, |
969 | BasicBlock *InsertAtEnd) { |
970 | unsigned Values = 1 + unsigned(IdxList.size()); |
971 | assert(PointeeType && "Must specify element type")(static_cast <bool> (PointeeType && "Must specify element type" ) ? void (0) : __assert_fail ("PointeeType && \"Must specify element type\"" , "llvm/include/llvm/IR/Instructions.h", 971, __extension__ __PRETTY_FUNCTION__ )); |
972 | assert(cast<PointerType>(Ptr->getType()->getScalarType())(static_cast <bool> (cast<PointerType>(Ptr->getType ()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType )) ? void (0) : __assert_fail ("cast<PointerType>(Ptr->getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType)" , "llvm/include/llvm/IR/Instructions.h", 973, __extension__ __PRETTY_FUNCTION__ )) |
973 | ->isOpaqueOrPointeeTypeMatches(PointeeType))(static_cast <bool> (cast<PointerType>(Ptr->getType ()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType )) ? void (0) : __assert_fail ("cast<PointerType>(Ptr->getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType)" , "llvm/include/llvm/IR/Instructions.h", 973, __extension__ __PRETTY_FUNCTION__ )); |
974 | return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values, |
975 | NameStr, InsertAtEnd); |
976 | } |
977 | |
978 | /// Create an "inbounds" getelementptr. See the documentation for the |
979 | /// "inbounds" flag in LangRef.html for details. |
980 | static GetElementPtrInst * |
981 | CreateInBounds(Type *PointeeType, Value *Ptr, ArrayRef<Value *> IdxList, |
982 | const Twine &NameStr = "", |
983 | Instruction *InsertBefore = nullptr) { |
984 | GetElementPtrInst *GEP = |
985 | Create(PointeeType, Ptr, IdxList, NameStr, InsertBefore); |
986 | GEP->setIsInBounds(true); |
987 | return GEP; |
988 | } |
989 | |
990 | static GetElementPtrInst *CreateInBounds(Type *PointeeType, Value *Ptr, |
991 | ArrayRef<Value *> IdxList, |
992 | const Twine &NameStr, |
993 | BasicBlock *InsertAtEnd) { |
994 | GetElementPtrInst *GEP = |
995 | Create(PointeeType, Ptr, IdxList, NameStr, InsertAtEnd); |
996 | GEP->setIsInBounds(true); |
997 | return GEP; |
998 | } |
999 | |
1000 | /// Transparently provide more efficient getOperand methods. |
1001 | 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; |
1002 | |
1003 | Type *getSourceElementType() const { return SourceElementType; } |
1004 | |
1005 | void setSourceElementType(Type *Ty) { SourceElementType = Ty; } |
1006 | void setResultElementType(Type *Ty) { ResultElementType = Ty; } |
1007 | |
1008 | Type *getResultElementType() const { |
1009 | assert(cast<PointerType>(getType()->getScalarType())(static_cast <bool> (cast<PointerType>(getType()-> getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType )) ? void (0) : __assert_fail ("cast<PointerType>(getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType)" , "llvm/include/llvm/IR/Instructions.h", 1010, __extension__ __PRETTY_FUNCTION__ )) |
1010 | ->isOpaqueOrPointeeTypeMatches(ResultElementType))(static_cast <bool> (cast<PointerType>(getType()-> getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType )) ? void (0) : __assert_fail ("cast<PointerType>(getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType)" , "llvm/include/llvm/IR/Instructions.h", 1010, __extension__ __PRETTY_FUNCTION__ )); |
1011 | return ResultElementType; |
1012 | } |
1013 | |
1014 | /// Returns the address space of this instruction's pointer type. |
1015 | unsigned getAddressSpace() const { |
1016 | // Note that this is always the same as the pointer operand's address space |
1017 | // and that is cheaper to compute, so cheat here. |
1018 | return getPointerAddressSpace(); |
1019 | } |
1020 | |
1021 | /// Returns the result type of a getelementptr with the given source |
1022 | /// element type and indexes. |
1023 | /// |
1024 | /// Null is returned if the indices are invalid for the specified |
1025 | /// source element type. |
1026 | static Type *getIndexedType(Type *Ty, ArrayRef<Value *> IdxList); |
1027 | static Type *getIndexedType(Type *Ty, ArrayRef<Constant *> IdxList); |
1028 | static Type *getIndexedType(Type *Ty, ArrayRef<uint64_t> IdxList); |
1029 | |
1030 | /// Return the type of the element at the given index of an indexable |
1031 | /// type. This is equivalent to "getIndexedType(Agg, {Zero, Idx})". |
1032 | /// |
1033 | /// Returns null if the type can't be indexed, or the given index is not |
1034 | /// legal for the given type. |
1035 | static Type *getTypeAtIndex(Type *Ty, Value *Idx); |
1036 | static Type *getTypeAtIndex(Type *Ty, uint64_t Idx); |
1037 | |
1038 | inline op_iterator idx_begin() { return op_begin()+1; } |
1039 | inline const_op_iterator idx_begin() const { return op_begin()+1; } |
1040 | inline op_iterator idx_end() { return op_end(); } |
1041 | inline const_op_iterator idx_end() const { return op_end(); } |
1042 | |
1043 | inline iterator_range<op_iterator> indices() { |
1044 | return make_range(idx_begin(), idx_end()); |
1045 | } |
1046 | |
1047 | inline iterator_range<const_op_iterator> indices() const { |
1048 | return make_range(idx_begin(), idx_end()); |
1049 | } |
1050 | |
1051 | Value *getPointerOperand() { |
1052 | return getOperand(0); |
1053 | } |
1054 | const Value *getPointerOperand() const { |
1055 | return getOperand(0); |
1056 | } |
1057 | static unsigned getPointerOperandIndex() { |
1058 | return 0U; // get index for modifying correct operand. |
1059 | } |
1060 | |
1061 | /// Method to return the pointer operand as a |
1062 | /// PointerType. |
1063 | Type *getPointerOperandType() const { |
1064 | return getPointerOperand()->getType(); |
1065 | } |
1066 | |
1067 | /// Returns the address space of the pointer operand. |
1068 | unsigned getPointerAddressSpace() const { |
1069 | return getPointerOperandType()->getPointerAddressSpace(); |
1070 | } |
1071 | |
1072 | /// Returns the pointer type returned by the GEP |
1073 | /// instruction, which may be a vector of pointers. |
1074 | static Type *getGEPReturnType(Type *ElTy, Value *Ptr, |
1075 | ArrayRef<Value *> IdxList) { |
1076 | PointerType *OrigPtrTy = cast<PointerType>(Ptr->getType()->getScalarType()); |
1077 | unsigned AddrSpace = OrigPtrTy->getAddressSpace(); |
1078 | Type *ResultElemTy = checkGEPType(getIndexedType(ElTy, IdxList)); |
1079 | Type *PtrTy = OrigPtrTy->isOpaque() |
1080 | ? PointerType::get(OrigPtrTy->getContext(), AddrSpace) |
1081 | : PointerType::get(ResultElemTy, AddrSpace); |
1082 | // Vector GEP |
1083 | if (auto *PtrVTy = dyn_cast<VectorType>(Ptr->getType())) { |
1084 | ElementCount EltCount = PtrVTy->getElementCount(); |
1085 | return VectorType::get(PtrTy, EltCount); |
1086 | } |
1087 | for (Value *Index : IdxList) |
1088 | if (auto *IndexVTy = dyn_cast<VectorType>(Index->getType())) { |
1089 | ElementCount EltCount = IndexVTy->getElementCount(); |
1090 | return VectorType::get(PtrTy, EltCount); |
1091 | } |
1092 | // Scalar GEP |
1093 | return PtrTy; |
1094 | } |
1095 | |
1096 | unsigned getNumIndices() const { // Note: always non-negative |
1097 | return getNumOperands() - 1; |
1098 | } |
1099 | |
1100 | bool hasIndices() const { |
1101 | return getNumOperands() > 1; |
1102 | } |
1103 | |
1104 | /// Return true if all of the indices of this GEP are |
1105 | /// zeros. If so, the result pointer and the first operand have the same |
1106 | /// value, just potentially different types. |
1107 | bool hasAllZeroIndices() const; |
1108 | |
1109 | /// Return true if all of the indices of this GEP are |
1110 | /// constant integers. If so, the result pointer and the first operand have |
1111 | /// a constant offset between them. |
1112 | bool hasAllConstantIndices() const; |
1113 | |
1114 | /// Set or clear the inbounds flag on this GEP instruction. |
1115 | /// See LangRef.html for the meaning of inbounds on a getelementptr. |
1116 | void setIsInBounds(bool b = true); |
1117 | |
1118 | /// Determine whether the GEP has the inbounds flag. |
1119 | bool isInBounds() const; |
1120 | |
1121 | /// Accumulate the constant address offset of this GEP if possible. |
1122 | /// |
1123 | /// This routine accepts an APInt into which it will accumulate the constant |
1124 | /// offset of this GEP if the GEP is in fact constant. If the GEP is not |
1125 | /// all-constant, it returns false and the value of the offset APInt is |
1126 | /// undefined (it is *not* preserved!). The APInt passed into this routine |
1127 | /// must be at least as wide as the IntPtr type for the address space of |
1128 | /// the base GEP pointer. |
1129 | bool accumulateConstantOffset(const DataLayout &DL, APInt &Offset) const; |
1130 | bool collectOffset(const DataLayout &DL, unsigned BitWidth, |
1131 | MapVector<Value *, APInt> &VariableOffsets, |
1132 | APInt &ConstantOffset) const; |
1133 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1134 | static bool classof(const Instruction *I) { |
1135 | return (I->getOpcode() == Instruction::GetElementPtr); |
1136 | } |
1137 | static bool classof(const Value *V) { |
1138 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1139 | } |
1140 | }; |
1141 | |
1142 | template <> |
1143 | struct OperandTraits<GetElementPtrInst> : |
1144 | public VariadicOperandTraits<GetElementPtrInst, 1> { |
1145 | }; |
1146 | |
1147 | GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr, |
1148 | ArrayRef<Value *> IdxList, unsigned Values, |
1149 | const Twine &NameStr, |
1150 | Instruction *InsertBefore) |
1151 | : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr, |
1152 | OperandTraits<GetElementPtrInst>::op_end(this) - Values, |
1153 | Values, InsertBefore), |
1154 | SourceElementType(PointeeType), |
1155 | ResultElementType(getIndexedType(PointeeType, IdxList)) { |
1156 | assert(cast<PointerType>(getType()->getScalarType())(static_cast <bool> (cast<PointerType>(getType()-> getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType )) ? void (0) : __assert_fail ("cast<PointerType>(getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType)" , "llvm/include/llvm/IR/Instructions.h", 1157, __extension__ __PRETTY_FUNCTION__ )) |
1157 | ->isOpaqueOrPointeeTypeMatches(ResultElementType))(static_cast <bool> (cast<PointerType>(getType()-> getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType )) ? void (0) : __assert_fail ("cast<PointerType>(getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType)" , "llvm/include/llvm/IR/Instructions.h", 1157, __extension__ __PRETTY_FUNCTION__ )); |
1158 | init(Ptr, IdxList, NameStr); |
1159 | } |
1160 | |
1161 | GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr, |
1162 | ArrayRef<Value *> IdxList, unsigned Values, |
1163 | const Twine &NameStr, |
1164 | BasicBlock *InsertAtEnd) |
1165 | : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr, |
1166 | OperandTraits<GetElementPtrInst>::op_end(this) - Values, |
1167 | Values, InsertAtEnd), |
1168 | SourceElementType(PointeeType), |
1169 | ResultElementType(getIndexedType(PointeeType, IdxList)) { |
1170 | assert(cast<PointerType>(getType()->getScalarType())(static_cast <bool> (cast<PointerType>(getType()-> getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType )) ? void (0) : __assert_fail ("cast<PointerType>(getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType)" , "llvm/include/llvm/IR/Instructions.h", 1171, __extension__ __PRETTY_FUNCTION__ )) |
1171 | ->isOpaqueOrPointeeTypeMatches(ResultElementType))(static_cast <bool> (cast<PointerType>(getType()-> getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType )) ? void (0) : __assert_fail ("cast<PointerType>(getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType)" , "llvm/include/llvm/IR/Instructions.h", 1171, __extension__ __PRETTY_FUNCTION__ )); |
1172 | init(Ptr, IdxList, NameStr); |
1173 | } |
1174 | |
1175 | 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 { (static_cast <bool> (i_nocapture < OperandTraits<GetElementPtrInst>::operands(this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<GetElementPtrInst>::operands(this) && \"getOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 1175, __extension__ __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) { (static_cast <bool> (i_nocapture < OperandTraits<GetElementPtrInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<GetElementPtrInst>::operands(this) && \"setOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 1175, __extension__ __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 ); } |
1176 | |
1177 | //===----------------------------------------------------------------------===// |
1178 | // ICmpInst Class |
1179 | //===----------------------------------------------------------------------===// |
1180 | |
1181 | /// This instruction compares its operands according to the predicate given |
1182 | /// to the constructor. It only operates on integers or pointers. The operands |
1183 | /// must be identical types. |
1184 | /// Represent an integer comparison operator. |
1185 | class ICmpInst: public CmpInst { |
1186 | void AssertOK() { |
1187 | assert(isIntPredicate() &&(static_cast <bool> (isIntPredicate() && "Invalid ICmp predicate value" ) ? void (0) : __assert_fail ("isIntPredicate() && \"Invalid ICmp predicate value\"" , "llvm/include/llvm/IR/Instructions.h", 1188, __extension__ __PRETTY_FUNCTION__ )) |
1188 | "Invalid ICmp predicate value")(static_cast <bool> (isIntPredicate() && "Invalid ICmp predicate value" ) ? void (0) : __assert_fail ("isIntPredicate() && \"Invalid ICmp predicate value\"" , "llvm/include/llvm/IR/Instructions.h", 1188, __extension__ __PRETTY_FUNCTION__ )); |
1189 | assert(getOperand(0)->getType() == getOperand(1)->getType() &&(static_cast <bool> (getOperand(0)->getType() == getOperand (1)->getType() && "Both operands to ICmp instruction are not of the same type!" ) ? void (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to ICmp instruction are not of the same type!\"" , "llvm/include/llvm/IR/Instructions.h", 1190, __extension__ __PRETTY_FUNCTION__ )) |
1190 | "Both operands to ICmp instruction are not of the same type!")(static_cast <bool> (getOperand(0)->getType() == getOperand (1)->getType() && "Both operands to ICmp instruction are not of the same type!" ) ? void (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to ICmp instruction are not of the same type!\"" , "llvm/include/llvm/IR/Instructions.h", 1190, __extension__ __PRETTY_FUNCTION__ )); |
1191 | // Check that the operands are the right type |
1192 | assert((getOperand(0)->getType()->isIntOrIntVectorTy() ||(static_cast <bool> ((getOperand(0)->getType()->isIntOrIntVectorTy () || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && "Invalid operand types for ICmp instruction") ? void (0) : __assert_fail ("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\"" , "llvm/include/llvm/IR/Instructions.h", 1194, __extension__ __PRETTY_FUNCTION__ )) |
1193 | getOperand(0)->getType()->isPtrOrPtrVectorTy()) &&(static_cast <bool> ((getOperand(0)->getType()->isIntOrIntVectorTy () || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && "Invalid operand types for ICmp instruction") ? void (0) : __assert_fail ("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\"" , "llvm/include/llvm/IR/Instructions.h", 1194, __extension__ __PRETTY_FUNCTION__ )) |
1194 | "Invalid operand types for ICmp instruction")(static_cast <bool> ((getOperand(0)->getType()->isIntOrIntVectorTy () || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && "Invalid operand types for ICmp instruction") ? void (0) : __assert_fail ("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\"" , "llvm/include/llvm/IR/Instructions.h", 1194, __extension__ __PRETTY_FUNCTION__ )); |
1195 | } |
1196 | |
1197 | protected: |
1198 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1199 | friend class Instruction; |
1200 | |
1201 | /// Clone an identical ICmpInst |
1202 | ICmpInst *cloneImpl() const; |
1203 | |
1204 | public: |
1205 | /// Constructor with insert-before-instruction semantics. |
1206 | ICmpInst( |
1207 | Instruction *InsertBefore, ///< Where to insert |
1208 | Predicate pred, ///< The predicate to use for the comparison |
1209 | Value *LHS, ///< The left-hand-side of the expression |
1210 | Value *RHS, ///< The right-hand-side of the expression |
1211 | const Twine &NameStr = "" ///< Name of the instruction |
1212 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1213 | Instruction::ICmp, pred, LHS, RHS, NameStr, |
1214 | InsertBefore) { |
1215 | #ifndef NDEBUG |
1216 | AssertOK(); |
1217 | #endif |
1218 | } |
1219 | |
1220 | /// Constructor with insert-at-end semantics. |
1221 | ICmpInst( |
1222 | BasicBlock &InsertAtEnd, ///< Block to insert into. |
1223 | Predicate pred, ///< The predicate to use for the comparison |
1224 | Value *LHS, ///< The left-hand-side of the expression |
1225 | Value *RHS, ///< The right-hand-side of the expression |
1226 | const Twine &NameStr = "" ///< Name of the instruction |
1227 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1228 | Instruction::ICmp, pred, LHS, RHS, NameStr, |
1229 | &InsertAtEnd) { |
1230 | #ifndef NDEBUG |
1231 | AssertOK(); |
1232 | #endif |
1233 | } |
1234 | |
1235 | /// Constructor with no-insertion semantics |
1236 | ICmpInst( |
1237 | Predicate pred, ///< The predicate to use for the comparison |
1238 | Value *LHS, ///< The left-hand-side of the expression |
1239 | Value *RHS, ///< The right-hand-side of the expression |
1240 | const Twine &NameStr = "" ///< Name of the instruction |
1241 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1242 | Instruction::ICmp, pred, LHS, RHS, NameStr) { |
1243 | #ifndef NDEBUG |
1244 | AssertOK(); |
1245 | #endif |
1246 | } |
1247 | |
1248 | /// For example, EQ->EQ, SLE->SLE, UGT->SGT, etc. |
1249 | /// @returns the predicate that would be the result if the operand were |
1250 | /// regarded as signed. |
1251 | /// Return the signed version of the predicate |
1252 | Predicate getSignedPredicate() const { |
1253 | return getSignedPredicate(getPredicate()); |
1254 | } |
1255 | |
1256 | /// This is a static version that you can use without an instruction. |
1257 | /// Return the signed version of the predicate. |
1258 | static Predicate getSignedPredicate(Predicate pred); |
1259 | |
1260 | /// For example, EQ->EQ, SLE->ULE, UGT->UGT, etc. |
1261 | /// @returns the predicate that would be the result if the operand were |
1262 | /// regarded as unsigned. |
1263 | /// Return the unsigned version of the predicate |
1264 | Predicate getUnsignedPredicate() const { |
1265 | return getUnsignedPredicate(getPredicate()); |
1266 | } |
1267 | |
1268 | /// This is a static version that you can use without an instruction. |
1269 | /// Return the unsigned version of the predicate. |
1270 | static Predicate getUnsignedPredicate(Predicate pred); |
1271 | |
1272 | /// Return true if this predicate is either EQ or NE. This also |
1273 | /// tests for commutativity. |
1274 | static bool isEquality(Predicate P) { |
1275 | return P == ICMP_EQ || P == ICMP_NE; |
1276 | } |
1277 | |
1278 | /// Return true if this predicate is either EQ or NE. This also |
1279 | /// tests for commutativity. |
1280 | bool isEquality() const { |
1281 | return isEquality(getPredicate()); |
1282 | } |
1283 | |
1284 | /// @returns true if the predicate of this ICmpInst is commutative |
1285 | /// Determine if this relation is commutative. |
1286 | bool isCommutative() const { return isEquality(); } |
1287 | |
1288 | /// Return true if the predicate is relational (not EQ or NE). |
1289 | /// |
1290 | bool isRelational() const { |
1291 | return !isEquality(); |
1292 | } |
1293 | |
1294 | /// Return true if the predicate is relational (not EQ or NE). |
1295 | /// |
1296 | static bool isRelational(Predicate P) { |
1297 | return !isEquality(P); |
1298 | } |
1299 | |
1300 | /// Return true if the predicate is SGT or UGT. |
1301 | /// |
1302 | static bool isGT(Predicate P) { |
1303 | return P == ICMP_SGT || P == ICMP_UGT; |
1304 | } |
1305 | |
1306 | /// Return true if the predicate is SLT or ULT. |
1307 | /// |
1308 | static bool isLT(Predicate P) { |
1309 | return P == ICMP_SLT || P == ICMP_ULT; |
1310 | } |
1311 | |
1312 | /// Return true if the predicate is SGE or UGE. |
1313 | /// |
1314 | static bool isGE(Predicate P) { |
1315 | return P == ICMP_SGE || P == ICMP_UGE; |
1316 | } |
1317 | |
1318 | /// Return true if the predicate is SLE or ULE. |
1319 | /// |
1320 | static bool isLE(Predicate P) { |
1321 | return P == ICMP_SLE || P == ICMP_ULE; |
1322 | } |
1323 | |
1324 | /// Returns the sequence of all ICmp predicates. |
1325 | /// |
1326 | static auto predicates() { return ICmpPredicates(); } |
1327 | |
1328 | /// Exchange the two operands to this instruction in such a way that it does |
1329 | /// not modify the semantics of the instruction. The predicate value may be |
1330 | /// changed to retain the same result if the predicate is order dependent |
1331 | /// (e.g. ult). |
1332 | /// Swap operands and adjust predicate. |
1333 | void swapOperands() { |
1334 | setPredicate(getSwappedPredicate()); |
1335 | Op<0>().swap(Op<1>()); |
1336 | } |
1337 | |
1338 | /// Return result of `LHS Pred RHS` comparison. |
1339 | static bool compare(const APInt &LHS, const APInt &RHS, |
1340 | ICmpInst::Predicate Pred); |
1341 | |
1342 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1343 | static bool classof(const Instruction *I) { |
1344 | return I->getOpcode() == Instruction::ICmp; |
1345 | } |
1346 | static bool classof(const Value *V) { |
1347 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1348 | } |
1349 | }; |
1350 | |
1351 | //===----------------------------------------------------------------------===// |
1352 | // FCmpInst Class |
1353 | //===----------------------------------------------------------------------===// |
1354 | |
1355 | /// This instruction compares its operands according to the predicate given |
1356 | /// to the constructor. It only operates on floating point values or packed |
1357 | /// vectors of floating point values. The operands must be identical types. |
1358 | /// Represents a floating point comparison operator. |
1359 | class FCmpInst: public CmpInst { |
1360 | void AssertOK() { |
1361 | assert(isFPPredicate() && "Invalid FCmp predicate value")(static_cast <bool> (isFPPredicate() && "Invalid FCmp predicate value" ) ? void (0) : __assert_fail ("isFPPredicate() && \"Invalid FCmp predicate value\"" , "llvm/include/llvm/IR/Instructions.h", 1361, __extension__ __PRETTY_FUNCTION__ )); |
1362 | assert(getOperand(0)->getType() == getOperand(1)->getType() &&(static_cast <bool> (getOperand(0)->getType() == getOperand (1)->getType() && "Both operands to FCmp instruction are not of the same type!" ) ? void (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to FCmp instruction are not of the same type!\"" , "llvm/include/llvm/IR/Instructions.h", 1363, __extension__ __PRETTY_FUNCTION__ )) |
1363 | "Both operands to FCmp instruction are not of the same type!")(static_cast <bool> (getOperand(0)->getType() == getOperand (1)->getType() && "Both operands to FCmp instruction are not of the same type!" ) ? void (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to FCmp instruction are not of the same type!\"" , "llvm/include/llvm/IR/Instructions.h", 1363, __extension__ __PRETTY_FUNCTION__ )); |
1364 | // Check that the operands are the right type |
1365 | assert(getOperand(0)->getType()->isFPOrFPVectorTy() &&(static_cast <bool> (getOperand(0)->getType()->isFPOrFPVectorTy () && "Invalid operand types for FCmp instruction") ? void (0) : __assert_fail ("getOperand(0)->getType()->isFPOrFPVectorTy() && \"Invalid operand types for FCmp instruction\"" , "llvm/include/llvm/IR/Instructions.h", 1366, __extension__ __PRETTY_FUNCTION__ )) |
1366 | "Invalid operand types for FCmp instruction")(static_cast <bool> (getOperand(0)->getType()->isFPOrFPVectorTy () && "Invalid operand types for FCmp instruction") ? void (0) : __assert_fail ("getOperand(0)->getType()->isFPOrFPVectorTy() && \"Invalid operand types for FCmp instruction\"" , "llvm/include/llvm/IR/Instructions.h", 1366, __extension__ __PRETTY_FUNCTION__ )); |
1367 | } |
1368 | |
1369 | protected: |
1370 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1371 | friend class Instruction; |
1372 | |
1373 | /// Clone an identical FCmpInst |
1374 | FCmpInst *cloneImpl() const; |
1375 | |
1376 | public: |
1377 | /// Constructor with insert-before-instruction semantics. |
1378 | FCmpInst( |
1379 | Instruction *InsertBefore, ///< Where to insert |
1380 | Predicate pred, ///< The predicate to use for the comparison |
1381 | Value *LHS, ///< The left-hand-side of the expression |
1382 | Value *RHS, ///< The right-hand-side of the expression |
1383 | const Twine &NameStr = "" ///< Name of the instruction |
1384 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1385 | Instruction::FCmp, pred, LHS, RHS, NameStr, |
1386 | InsertBefore) { |
1387 | AssertOK(); |
1388 | } |
1389 | |
1390 | /// Constructor with insert-at-end semantics. |
1391 | FCmpInst( |
1392 | BasicBlock &InsertAtEnd, ///< Block to insert into. |
1393 | Predicate pred, ///< The predicate to use for the comparison |
1394 | Value *LHS, ///< The left-hand-side of the expression |
1395 | Value *RHS, ///< The right-hand-side of the expression |
1396 | const Twine &NameStr = "" ///< Name of the instruction |
1397 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1398 | Instruction::FCmp, pred, LHS, RHS, NameStr, |
1399 | &InsertAtEnd) { |
1400 | AssertOK(); |
1401 | } |
1402 | |
1403 | /// Constructor with no-insertion semantics |
1404 | FCmpInst( |
1405 | Predicate Pred, ///< The predicate to use for the comparison |
1406 | Value *LHS, ///< The left-hand-side of the expression |
1407 | Value *RHS, ///< The right-hand-side of the expression |
1408 | const Twine &NameStr = "", ///< Name of the instruction |
1409 | Instruction *FlagsSource = nullptr |
1410 | ) : CmpInst(makeCmpResultType(LHS->getType()), Instruction::FCmp, Pred, LHS, |
1411 | RHS, NameStr, nullptr, FlagsSource) { |
1412 | AssertOK(); |
1413 | } |
1414 | |
1415 | /// @returns true if the predicate of this instruction is EQ or NE. |
1416 | /// Determine if this is an equality predicate. |
1417 | static bool isEquality(Predicate Pred) { |
1418 | return Pred == FCMP_OEQ || Pred == FCMP_ONE || Pred == FCMP_UEQ || |
1419 | Pred == FCMP_UNE; |
1420 | } |
1421 | |
1422 | /// @returns true if the predicate of this instruction is EQ or NE. |
1423 | /// Determine if this is an equality predicate. |
1424 | bool isEquality() const { return isEquality(getPredicate()); } |
1425 | |
1426 | /// @returns true if the predicate of this instruction is commutative. |
1427 | /// Determine if this is a commutative predicate. |
1428 | bool isCommutative() const { |
1429 | return isEquality() || |
1430 | getPredicate() == FCMP_FALSE || |
1431 | getPredicate() == FCMP_TRUE || |
1432 | getPredicate() == FCMP_ORD || |
1433 | getPredicate() == FCMP_UNO; |
1434 | } |
1435 | |
1436 | /// @returns true if the predicate is relational (not EQ or NE). |
1437 | /// Determine if this a relational predicate. |
1438 | bool isRelational() const { return !isEquality(); } |
1439 | |
1440 | /// Exchange the two operands to this instruction in such a way that it does |
1441 | /// not modify the semantics of the instruction. The predicate value may be |
1442 | /// changed to retain the same result if the predicate is order dependent |
1443 | /// (e.g. ult). |
1444 | /// Swap operands and adjust predicate. |
1445 | void swapOperands() { |
1446 | setPredicate(getSwappedPredicate()); |
1447 | Op<0>().swap(Op<1>()); |
1448 | } |
1449 | |
1450 | /// Returns the sequence of all FCmp predicates. |
1451 | /// |
1452 | static auto predicates() { return FCmpPredicates(); } |
1453 | |
1454 | /// Return result of `LHS Pred RHS` comparison. |
1455 | static bool compare(const APFloat &LHS, const APFloat &RHS, |
1456 | FCmpInst::Predicate Pred); |
1457 | |
1458 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
1459 | static bool classof(const Instruction *I) { |
1460 | return I->getOpcode() == Instruction::FCmp; |
1461 | } |
1462 | static bool classof(const Value *V) { |
1463 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1464 | } |
1465 | }; |
1466 | |
1467 | //===----------------------------------------------------------------------===// |
1468 | /// This class represents a function call, abstracting a target |
1469 | /// machine's calling convention. This class uses low bit of the SubClassData |
1470 | /// field to indicate whether or not this is a tail call. The rest of the bits |
1471 | /// hold the calling convention of the call. |
1472 | /// |
1473 | class CallInst : public CallBase { |
1474 | CallInst(const CallInst &CI); |
1475 | |
1476 | /// Construct a CallInst given a range of arguments. |
1477 | /// Construct a CallInst from a range of arguments |
1478 | inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1479 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1480 | Instruction *InsertBefore); |
1481 | |
1482 | inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1483 | const Twine &NameStr, Instruction *InsertBefore) |
1484 | : CallInst(Ty, Func, Args, std::nullopt, NameStr, InsertBefore) {} |
1485 | |
1486 | /// Construct a CallInst given a range of arguments. |
1487 | /// Construct a CallInst from a range of arguments |
1488 | inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1489 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1490 | BasicBlock *InsertAtEnd); |
1491 | |
1492 | explicit CallInst(FunctionType *Ty, Value *F, const Twine &NameStr, |
1493 | Instruction *InsertBefore); |
1494 | |
1495 | CallInst(FunctionType *ty, Value *F, const Twine &NameStr, |
1496 | BasicBlock *InsertAtEnd); |
1497 | |
1498 | void init(FunctionType *FTy, Value *Func, ArrayRef<Value *> Args, |
1499 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr); |
1500 | void init(FunctionType *FTy, Value *Func, const Twine &NameStr); |
1501 | |
1502 | /// Compute the number of operands to allocate. |
1503 | static int ComputeNumOperands(int NumArgs, int NumBundleInputs = 0) { |
1504 | // We need one operand for the called function, plus the input operand |
1505 | // counts provided. |
1506 | return 1 + NumArgs + NumBundleInputs; |
1507 | } |
1508 | |
1509 | protected: |
1510 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1511 | friend class Instruction; |
1512 | |
1513 | CallInst *cloneImpl() const; |
1514 | |
1515 | public: |
1516 | static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr = "", |
1517 | Instruction *InsertBefore = nullptr) { |
1518 | return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertBefore); |
1519 | } |
1520 | |
1521 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1522 | const Twine &NameStr, |
1523 | Instruction *InsertBefore = nullptr) { |
1524 | return new (ComputeNumOperands(Args.size())) |
1525 | CallInst(Ty, Func, Args, std::nullopt, NameStr, InsertBefore); |
1526 | } |
1527 | |
1528 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1529 | ArrayRef<OperandBundleDef> Bundles = std::nullopt, |
1530 | const Twine &NameStr = "", |
1531 | Instruction *InsertBefore = nullptr) { |
1532 | const int NumOperands = |
1533 | ComputeNumOperands(Args.size(), CountBundleInputs(Bundles)); |
1534 | const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
1535 | |
1536 | return new (NumOperands, DescriptorBytes) |
1537 | CallInst(Ty, Func, Args, Bundles, NameStr, InsertBefore); |
1538 | } |
1539 | |
1540 | static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr, |
1541 | BasicBlock *InsertAtEnd) { |
1542 | return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertAtEnd); |
1543 | } |
1544 | |
1545 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1546 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
1547 | return new (ComputeNumOperands(Args.size())) |
1548 | CallInst(Ty, Func, Args, std::nullopt, NameStr, InsertAtEnd); |
1549 | } |
1550 | |
1551 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1552 | ArrayRef<OperandBundleDef> Bundles, |
1553 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
1554 | const int NumOperands = |
1555 | ComputeNumOperands(Args.size(), CountBundleInputs(Bundles)); |
1556 | const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
1557 | |
1558 | return new (NumOperands, DescriptorBytes) |
1559 | CallInst(Ty, Func, Args, Bundles, NameStr, InsertAtEnd); |
1560 | } |
1561 | |
1562 | static CallInst *Create(FunctionCallee Func, const Twine &NameStr = "", |
1563 | Instruction *InsertBefore = nullptr) { |
1564 | return Create(Func.getFunctionType(), Func.getCallee(), NameStr, |
1565 | InsertBefore); |
1566 | } |
1567 | |
1568 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
1569 | ArrayRef<OperandBundleDef> Bundles = std::nullopt, |
1570 | const Twine &NameStr = "", |
1571 | Instruction *InsertBefore = nullptr) { |
1572 | return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles, |
1573 | NameStr, InsertBefore); |
1574 | } |
1575 | |
1576 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
1577 | const Twine &NameStr, |
1578 | Instruction *InsertBefore = nullptr) { |
1579 | return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr, |
1580 | InsertBefore); |
1581 | } |
1582 | |
1583 | static CallInst *Create(FunctionCallee Func, const Twine &NameStr, |
1584 | BasicBlock *InsertAtEnd) { |
1585 | return Create(Func.getFunctionType(), Func.getCallee(), NameStr, |
1586 | InsertAtEnd); |
1587 | } |
1588 | |
1589 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
1590 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
1591 | return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr, |
1592 | InsertAtEnd); |
1593 | } |
1594 | |
1595 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
1596 | ArrayRef<OperandBundleDef> Bundles, |
1597 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
1598 | return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles, |
1599 | NameStr, InsertAtEnd); |
1600 | } |
1601 | |
1602 | /// Create a clone of \p CI with a different set of operand bundles and |
1603 | /// insert it before \p InsertPt. |
1604 | /// |
1605 | /// The returned call instruction is identical \p CI in every way except that |
1606 | /// the operand bundles for the new instruction are set to the operand bundles |
1607 | /// in \p Bundles. |
1608 | static CallInst *Create(CallInst *CI, ArrayRef<OperandBundleDef> Bundles, |
1609 | Instruction *InsertPt = nullptr); |
1610 | |
1611 | /// Generate the IR for a call to malloc: |
1612 | /// 1. Compute the malloc call's argument as the specified type's size, |
1613 | /// possibly multiplied by the array size if the array size is not |
1614 | /// constant 1. |
1615 | /// 2. Call malloc with that argument. |
1616 | /// 3. Bitcast the result of the malloc call to the specified type. |
1617 | static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy, |
1618 | Type *AllocTy, Value *AllocSize, |
1619 | Value *ArraySize = nullptr, |
1620 | Function *MallocF = nullptr, |
1621 | const Twine &Name = ""); |
1622 | static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy, |
1623 | Type *AllocTy, Value *AllocSize, |
1624 | Value *ArraySize = nullptr, |
1625 | Function *MallocF = nullptr, |
1626 | const Twine &Name = ""); |
1627 | static Instruction * |
1628 | CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy, Type *AllocTy, |
1629 | Value *AllocSize, Value *ArraySize = nullptr, |
1630 | ArrayRef<OperandBundleDef> Bundles = std::nullopt, |
1631 | Function *MallocF = nullptr, const Twine &Name = ""); |
1632 | static Instruction * |
1633 | CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy, Type *AllocTy, |
1634 | Value *AllocSize, Value *ArraySize = nullptr, |
1635 | ArrayRef<OperandBundleDef> Bundles = std::nullopt, |
1636 | Function *MallocF = nullptr, const Twine &Name = ""); |
1637 | /// Generate the IR for a call to the builtin free function. |
1638 | static Instruction *CreateFree(Value *Source, Instruction *InsertBefore); |
1639 | static Instruction *CreateFree(Value *Source, BasicBlock *InsertAtEnd); |
1640 | static Instruction *CreateFree(Value *Source, |
1641 | ArrayRef<OperandBundleDef> Bundles, |
1642 | Instruction *InsertBefore); |
1643 | static Instruction *CreateFree(Value *Source, |
1644 | ArrayRef<OperandBundleDef> Bundles, |
1645 | BasicBlock *InsertAtEnd); |
1646 | |
1647 | // Note that 'musttail' implies 'tail'. |
1648 | enum TailCallKind : unsigned { |
1649 | TCK_None = 0, |
1650 | TCK_Tail = 1, |
1651 | TCK_MustTail = 2, |
1652 | TCK_NoTail = 3, |
1653 | TCK_LAST = TCK_NoTail |
1654 | }; |
1655 | |
1656 | using TailCallKindField = Bitfield::Element<TailCallKind, 0, 2, TCK_LAST>; |
1657 | static_assert( |
1658 | Bitfield::areContiguous<TailCallKindField, CallBase::CallingConvField>(), |
1659 | "Bitfields must be contiguous"); |
1660 | |
1661 | TailCallKind getTailCallKind() const { |
1662 | return getSubclassData<TailCallKindField>(); |
1663 | } |
1664 | |
1665 | bool isTailCall() const { |
1666 | TailCallKind Kind = getTailCallKind(); |
1667 | return Kind == TCK_Tail || Kind == TCK_MustTail; |
1668 | } |
1669 | |
1670 | bool isMustTailCall() const { return getTailCallKind() == TCK_MustTail; } |
1671 | |
1672 | bool isNoTailCall() const { return getTailCallKind() == TCK_NoTail; } |
1673 | |
1674 | void setTailCallKind(TailCallKind TCK) { |
1675 | setSubclassData<TailCallKindField>(TCK); |
1676 | } |
1677 | |
1678 | void setTailCall(bool IsTc = true) { |
1679 | setTailCallKind(IsTc ? TCK_Tail : TCK_None); |
1680 | } |
1681 | |
1682 | /// Return true if the call can return twice |
1683 | bool canReturnTwice() const { return hasFnAttr(Attribute::ReturnsTwice); } |
1684 | void setCanReturnTwice() { addFnAttr(Attribute::ReturnsTwice); } |
1685 | |
1686 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1687 | static bool classof(const Instruction *I) { |
1688 | return I->getOpcode() == Instruction::Call; |
1689 | } |
1690 | static bool classof(const Value *V) { |
1691 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1692 | } |
1693 | |
1694 | /// Updates profile metadata by scaling it by \p S / \p T. |
1695 | void updateProfWeight(uint64_t S, uint64_t T); |
1696 | |
1697 | private: |
1698 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
1699 | // method so that subclasses cannot accidentally use it. |
1700 | template <typename Bitfield> |
1701 | void setSubclassData(typename Bitfield::Type Value) { |
1702 | Instruction::setSubclassData<Bitfield>(Value); |
1703 | } |
1704 | }; |
1705 | |
1706 | CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1707 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1708 | BasicBlock *InsertAtEnd) |
1709 | : CallBase(Ty->getReturnType(), Instruction::Call, |
1710 | OperandTraits<CallBase>::op_end(this) - |
1711 | (Args.size() + CountBundleInputs(Bundles) + 1), |
1712 | unsigned(Args.size() + CountBundleInputs(Bundles) + 1), |
1713 | InsertAtEnd) { |
1714 | init(Ty, Func, Args, Bundles, NameStr); |
1715 | } |
1716 | |
1717 | CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1718 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1719 | Instruction *InsertBefore) |
1720 | : CallBase(Ty->getReturnType(), Instruction::Call, |
1721 | OperandTraits<CallBase>::op_end(this) - |
1722 | (Args.size() + CountBundleInputs(Bundles) + 1), |
1723 | unsigned(Args.size() + CountBundleInputs(Bundles) + 1), |
1724 | InsertBefore) { |
1725 | init(Ty, Func, Args, Bundles, NameStr); |
1726 | } |
1727 | |
1728 | //===----------------------------------------------------------------------===// |
1729 | // SelectInst Class |
1730 | //===----------------------------------------------------------------------===// |
1731 | |
1732 | /// This class represents the LLVM 'select' instruction. |
1733 | /// |
1734 | class SelectInst : public Instruction { |
1735 | SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr, |
1736 | Instruction *InsertBefore) |
1737 | : Instruction(S1->getType(), Instruction::Select, |
1738 | &Op<0>(), 3, InsertBefore) { |
1739 | init(C, S1, S2); |
1740 | setName(NameStr); |
1741 | } |
1742 | |
1743 | SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr, |
1744 | BasicBlock *InsertAtEnd) |
1745 | : Instruction(S1->getType(), Instruction::Select, |
1746 | &Op<0>(), 3, InsertAtEnd) { |
1747 | init(C, S1, S2); |
1748 | setName(NameStr); |
1749 | } |
1750 | |
1751 | void init(Value *C, Value *S1, Value *S2) { |
1752 | assert(!areInvalidOperands(C, S1, S2) && "Invalid operands for select")(static_cast <bool> (!areInvalidOperands(C, S1, S2) && "Invalid operands for select") ? void (0) : __assert_fail ("!areInvalidOperands(C, S1, S2) && \"Invalid operands for select\"" , "llvm/include/llvm/IR/Instructions.h", 1752, __extension__ __PRETTY_FUNCTION__ )); |
1753 | Op<0>() = C; |
1754 | Op<1>() = S1; |
1755 | Op<2>() = S2; |
1756 | } |
1757 | |
1758 | protected: |
1759 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1760 | friend class Instruction; |
1761 | |
1762 | SelectInst *cloneImpl() const; |
1763 | |
1764 | public: |
1765 | static SelectInst *Create(Value *C, Value *S1, Value *S2, |
1766 | const Twine &NameStr = "", |
1767 | Instruction *InsertBefore = nullptr, |
1768 | Instruction *MDFrom = nullptr) { |
1769 | SelectInst *Sel = new(3) SelectInst(C, S1, S2, NameStr, InsertBefore); |
1770 | if (MDFrom) |
1771 | Sel->copyMetadata(*MDFrom); |
1772 | return Sel; |
1773 | } |
1774 | |
1775 | static SelectInst *Create(Value *C, Value *S1, Value *S2, |
1776 | const Twine &NameStr, |
1777 | BasicBlock *InsertAtEnd) { |
1778 | return new(3) SelectInst(C, S1, S2, NameStr, InsertAtEnd); |
1779 | } |
1780 | |
1781 | const Value *getCondition() const { return Op<0>(); } |
1782 | const Value *getTrueValue() const { return Op<1>(); } |
1783 | const Value *getFalseValue() const { return Op<2>(); } |
1784 | Value *getCondition() { return Op<0>(); } |
1785 | Value *getTrueValue() { return Op<1>(); } |
1786 | Value *getFalseValue() { return Op<2>(); } |
1787 | |
1788 | void setCondition(Value *V) { Op<0>() = V; } |
1789 | void setTrueValue(Value *V) { Op<1>() = V; } |
1790 | void setFalseValue(Value *V) { Op<2>() = V; } |
1791 | |
1792 | /// Swap the true and false values of the select instruction. |
1793 | /// This doesn't swap prof metadata. |
1794 | void swapValues() { Op<1>().swap(Op<2>()); } |
1795 | |
1796 | /// Return a string if the specified operands are invalid |
1797 | /// for a select operation, otherwise return null. |
1798 | static const char *areInvalidOperands(Value *Cond, Value *True, Value *False); |
1799 | |
1800 | /// Transparently provide more efficient getOperand methods. |
1801 | 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; |
1802 | |
1803 | OtherOps getOpcode() const { |
1804 | return static_cast<OtherOps>(Instruction::getOpcode()); |
1805 | } |
1806 | |
1807 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1808 | static bool classof(const Instruction *I) { |
1809 | return I->getOpcode() == Instruction::Select; |
1810 | } |
1811 | static bool classof(const Value *V) { |
1812 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1813 | } |
1814 | }; |
1815 | |
1816 | template <> |
1817 | struct OperandTraits<SelectInst> : public FixedNumOperandTraits<SelectInst, 3> { |
1818 | }; |
1819 | |
1820 | 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 { (static_cast <bool > (i_nocapture < OperandTraits<SelectInst>::operands (this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<SelectInst>::operands(this) && \"getOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 1820, __extension__ __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) { (static_cast <bool> (i_nocapture < OperandTraits<SelectInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<SelectInst>::operands(this) && \"setOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 1820, __extension__ __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); } |
1821 | |
1822 | //===----------------------------------------------------------------------===// |
1823 | // VAArgInst Class |
1824 | //===----------------------------------------------------------------------===// |
1825 | |
1826 | /// This class represents the va_arg llvm instruction, which returns |
1827 | /// an argument of the specified type given a va_list and increments that list |
1828 | /// |
1829 | class VAArgInst : public UnaryInstruction { |
1830 | protected: |
1831 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1832 | friend class Instruction; |
1833 | |
1834 | VAArgInst *cloneImpl() const; |
1835 | |
1836 | public: |
1837 | VAArgInst(Value *List, Type *Ty, const Twine &NameStr = "", |
1838 | Instruction *InsertBefore = nullptr) |
1839 | : UnaryInstruction(Ty, VAArg, List, InsertBefore) { |
1840 | setName(NameStr); |
1841 | } |
1842 | |
1843 | VAArgInst(Value *List, Type *Ty, const Twine &NameStr, |
1844 | BasicBlock *InsertAtEnd) |
1845 | : UnaryInstruction(Ty, VAArg, List, InsertAtEnd) { |
1846 | setName(NameStr); |
1847 | } |
1848 | |
1849 | Value *getPointerOperand() { return getOperand(0); } |
1850 | const Value *getPointerOperand() const { return getOperand(0); } |
1851 | static unsigned getPointerOperandIndex() { return 0U; } |
1852 | |
1853 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1854 | static bool classof(const Instruction *I) { |
1855 | return I->getOpcode() == VAArg; |
1856 | } |
1857 | static bool classof(const Value *V) { |
1858 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1859 | } |
1860 | }; |
1861 | |
1862 | //===----------------------------------------------------------------------===// |
1863 | // ExtractElementInst Class |
1864 | //===----------------------------------------------------------------------===// |
1865 | |
1866 | /// This instruction extracts a single (scalar) |
1867 | /// element from a VectorType value |
1868 | /// |
1869 | class ExtractElementInst : public Instruction { |
1870 | ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr = "", |
1871 | Instruction *InsertBefore = nullptr); |
1872 | ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr, |
1873 | BasicBlock *InsertAtEnd); |
1874 | |
1875 | protected: |
1876 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1877 | friend class Instruction; |
1878 | |
1879 | ExtractElementInst *cloneImpl() const; |
1880 | |
1881 | public: |
1882 | static ExtractElementInst *Create(Value *Vec, Value *Idx, |
1883 | const Twine &NameStr = "", |
1884 | Instruction *InsertBefore = nullptr) { |
1885 | return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertBefore); |
1886 | } |
1887 | |
1888 | static ExtractElementInst *Create(Value *Vec, Value *Idx, |
1889 | const Twine &NameStr, |
1890 | BasicBlock *InsertAtEnd) { |
1891 | return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertAtEnd); |
1892 | } |
1893 | |
1894 | /// Return true if an extractelement instruction can be |
1895 | /// formed with the specified operands. |
1896 | static bool isValidOperands(const Value *Vec, const Value *Idx); |
1897 | |
1898 | Value *getVectorOperand() { return Op<0>(); } |
1899 | Value *getIndexOperand() { return Op<1>(); } |
1900 | const Value *getVectorOperand() const { return Op<0>(); } |
1901 | const Value *getIndexOperand() const { return Op<1>(); } |
1902 | |
1903 | VectorType *getVectorOperandType() const { |
1904 | return cast<VectorType>(getVectorOperand()->getType()); |
1905 | } |
1906 | |
1907 | /// Transparently provide more efficient getOperand methods. |
1908 | 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; |
1909 | |
1910 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1911 | static bool classof(const Instruction *I) { |
1912 | return I->getOpcode() == Instruction::ExtractElement; |
1913 | } |
1914 | static bool classof(const Value *V) { |
1915 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1916 | } |
1917 | }; |
1918 | |
1919 | template <> |
1920 | struct OperandTraits<ExtractElementInst> : |
1921 | public FixedNumOperandTraits<ExtractElementInst, 2> { |
1922 | }; |
1923 | |
1924 | 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 { (static_cast <bool> (i_nocapture < OperandTraits< ExtractElementInst>::operands(this) && "getOperand() out of range!" ) ? void (0) : __assert_fail ("i_nocapture < OperandTraits<ExtractElementInst>::operands(this) && \"getOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 1924, __extension__ __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) { (static_cast <bool> (i_nocapture < OperandTraits<ExtractElementInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<ExtractElementInst>::operands(this) && \"setOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 1924, __extension__ __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 ); } |
1925 | |
1926 | //===----------------------------------------------------------------------===// |
1927 | // InsertElementInst Class |
1928 | //===----------------------------------------------------------------------===// |
1929 | |
1930 | /// This instruction inserts a single (scalar) |
1931 | /// element into a VectorType value |
1932 | /// |
1933 | class InsertElementInst : public Instruction { |
1934 | InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, |
1935 | const Twine &NameStr = "", |
1936 | Instruction *InsertBefore = nullptr); |
1937 | InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, const Twine &NameStr, |
1938 | BasicBlock *InsertAtEnd); |
1939 | |
1940 | protected: |
1941 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1942 | friend class Instruction; |
1943 | |
1944 | InsertElementInst *cloneImpl() const; |
1945 | |
1946 | public: |
1947 | static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx, |
1948 | const Twine &NameStr = "", |
1949 | Instruction *InsertBefore = nullptr) { |
1950 | return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertBefore); |
1951 | } |
1952 | |
1953 | static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx, |
1954 | const Twine &NameStr, |
1955 | BasicBlock *InsertAtEnd) { |
1956 | return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertAtEnd); |
1957 | } |
1958 | |
1959 | /// Return true if an insertelement instruction can be |
1960 | /// formed with the specified operands. |
1961 | static bool isValidOperands(const Value *Vec, const Value *NewElt, |
1962 | const Value *Idx); |
1963 | |
1964 | /// Overload to return most specific vector type. |
1965 | /// |
1966 | VectorType *getType() const { |
1967 | return cast<VectorType>(Instruction::getType()); |
1968 | } |
1969 | |
1970 | /// Transparently provide more efficient getOperand methods. |
1971 | 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; |
1972 | |
1973 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1974 | static bool classof(const Instruction *I) { |
1975 | return I->getOpcode() == Instruction::InsertElement; |
1976 | } |
1977 | static bool classof(const Value *V) { |
1978 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1979 | } |
1980 | }; |
1981 | |
1982 | template <> |
1983 | struct OperandTraits<InsertElementInst> : |
1984 | public FixedNumOperandTraits<InsertElementInst, 3> { |
1985 | }; |
1986 | |
1987 | 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 { (static_cast <bool> (i_nocapture < OperandTraits<InsertElementInst>::operands(this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<InsertElementInst>::operands(this) && \"getOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 1987, __extension__ __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) { (static_cast <bool> (i_nocapture < OperandTraits<InsertElementInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<InsertElementInst>::operands(this) && \"setOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 1987, __extension__ __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 ); } |
1988 | |
1989 | //===----------------------------------------------------------------------===// |
1990 | // ShuffleVectorInst Class |
1991 | //===----------------------------------------------------------------------===// |
1992 | |
1993 | constexpr int UndefMaskElem = -1; |
1994 | |
1995 | /// This instruction constructs a fixed permutation of two |
1996 | /// input vectors. |
1997 | /// |
1998 | /// For each element of the result vector, the shuffle mask selects an element |
1999 | /// from one of the input vectors to copy to the result. Non-negative elements |
2000 | /// in the mask represent an index into the concatenated pair of input vectors. |
2001 | /// UndefMaskElem (-1) specifies that the result element is undefined. |
2002 | /// |
2003 | /// For scalable vectors, all the elements of the mask must be 0 or -1. This |
2004 | /// requirement may be relaxed in the future. |
2005 | class ShuffleVectorInst : public Instruction { |
2006 | SmallVector<int, 4> ShuffleMask; |
2007 | Constant *ShuffleMaskForBitcode; |
2008 | |
2009 | protected: |
2010 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2011 | friend class Instruction; |
2012 | |
2013 | ShuffleVectorInst *cloneImpl() const; |
2014 | |
2015 | public: |
2016 | ShuffleVectorInst(Value *V1, Value *Mask, const Twine &NameStr = "", |
2017 | Instruction *InsertBefore = nullptr); |
2018 | ShuffleVectorInst(Value *V1, Value *Mask, const Twine &NameStr, |
2019 | BasicBlock *InsertAtEnd); |
2020 | ShuffleVectorInst(Value *V1, ArrayRef<int> Mask, const Twine &NameStr = "", |
2021 | Instruction *InsertBefore = nullptr); |
2022 | ShuffleVectorInst(Value *V1, ArrayRef<int> Mask, const Twine &NameStr, |
2023 | BasicBlock *InsertAtEnd); |
2024 | ShuffleVectorInst(Value *V1, Value *V2, Value *Mask, |
2025 | const Twine &NameStr = "", |
2026 | Instruction *InsertBefor = nullptr); |
2027 | ShuffleVectorInst(Value *V1, Value *V2, Value *Mask, |
2028 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2029 | ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask, |
2030 | const Twine &NameStr = "", |
2031 | Instruction *InsertBefor = nullptr); |
2032 | ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask, |
2033 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2034 | |
2035 | void *operator new(size_t S) { return User::operator new(S, 2); } |
2036 | void operator delete(void *Ptr) { return User::operator delete(Ptr); } |
2037 | |
2038 | /// Swap the operands and adjust the mask to preserve the semantics |
2039 | /// of the instruction. |
2040 | void commute(); |
2041 | |
2042 | /// Return true if a shufflevector instruction can be |
2043 | /// formed with the specified operands. |
2044 | static bool isValidOperands(const Value *V1, const Value *V2, |
2045 | const Value *Mask); |
2046 | static bool isValidOperands(const Value *V1, const Value *V2, |
2047 | ArrayRef<int> Mask); |
2048 | |
2049 | /// Overload to return most specific vector type. |
2050 | /// |
2051 | VectorType *getType() const { |
2052 | return cast<VectorType>(Instruction::getType()); |
2053 | } |
2054 | |
2055 | /// Transparently provide more efficient getOperand methods. |
2056 | 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; |
2057 | |
2058 | /// Return the shuffle mask value of this instruction for the given element |
2059 | /// index. Return UndefMaskElem if the element is undef. |
2060 | int getMaskValue(unsigned Elt) const { return ShuffleMask[Elt]; } |
2061 | |
2062 | /// Convert the input shuffle mask operand to a vector of integers. Undefined |
2063 | /// elements of the mask are returned as UndefMaskElem. |
2064 | static void getShuffleMask(const Constant *Mask, |
2065 | SmallVectorImpl<int> &Result); |
2066 | |
2067 | /// Return the mask for this instruction as a vector of integers. Undefined |
2068 | /// elements of the mask are returned as UndefMaskElem. |
2069 | void getShuffleMask(SmallVectorImpl<int> &Result) const { |
2070 | Result.assign(ShuffleMask.begin(), ShuffleMask.end()); |
2071 | } |
2072 | |
2073 | /// Return the mask for this instruction, for use in bitcode. |
2074 | /// |
2075 | /// TODO: This is temporary until we decide a new bitcode encoding for |
2076 | /// shufflevector. |
2077 | Constant *getShuffleMaskForBitcode() const { return ShuffleMaskForBitcode; } |
2078 | |
2079 | static Constant *convertShuffleMaskForBitcode(ArrayRef<int> Mask, |
2080 | Type *ResultTy); |
2081 | |
2082 | void setShuffleMask(ArrayRef<int> Mask); |
2083 | |
2084 | ArrayRef<int> getShuffleMask() const { return ShuffleMask; } |
2085 | |
2086 | /// Return true if this shuffle returns a vector with a different number of |
2087 | /// elements than its source vectors. |
2088 | /// Examples: shufflevector <4 x n> A, <4 x n> B, <1,2,3> |
2089 | /// shufflevector <4 x n> A, <4 x n> B, <1,2,3,4,5> |
2090 | bool changesLength() const { |
2091 | unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType()) |
2092 | ->getElementCount() |
2093 | .getKnownMinValue(); |
2094 | unsigned NumMaskElts = ShuffleMask.size(); |
2095 | return NumSourceElts != NumMaskElts; |
2096 | } |
2097 | |
2098 | /// Return true if this shuffle returns a vector with a greater number of |
2099 | /// elements than its source vectors. |
2100 | /// Example: shufflevector <2 x n> A, <2 x n> B, <1,2,3> |
2101 | bool increasesLength() const { |
2102 | unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType()) |
2103 | ->getElementCount() |
2104 | .getKnownMinValue(); |
2105 | unsigned NumMaskElts = ShuffleMask.size(); |
2106 | return NumSourceElts < NumMaskElts; |
2107 | } |
2108 | |
2109 | /// Return true if this shuffle mask chooses elements from exactly one source |
2110 | /// vector. |
2111 | /// Example: <7,5,undef,7> |
2112 | /// This assumes that vector operands are the same length as the mask. |
2113 | static bool isSingleSourceMask(ArrayRef<int> Mask); |
2114 | static bool isSingleSourceMask(const Constant *Mask) { |
2115 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy( ) && "Shuffle needs vector constant.") ? void (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "llvm/include/llvm/IR/Instructions.h", 2115, __extension__ __PRETTY_FUNCTION__ )); |
2116 | SmallVector<int, 16> MaskAsInts; |
2117 | getShuffleMask(Mask, MaskAsInts); |
2118 | return isSingleSourceMask(MaskAsInts); |
2119 | } |
2120 | |
2121 | /// Return true if this shuffle chooses elements from exactly one source |
2122 | /// vector without changing the length of that vector. |
2123 | /// Example: shufflevector <4 x n> A, <4 x n> B, <3,0,undef,3> |
2124 | /// TODO: Optionally allow length-changing shuffles. |
2125 | bool isSingleSource() const { |
2126 | return !changesLength() && isSingleSourceMask(ShuffleMask); |
2127 | } |
2128 | |
2129 | /// Return true if this shuffle mask chooses elements from exactly one source |
2130 | /// vector without lane crossings. A shuffle using this mask is not |
2131 | /// necessarily a no-op because it may change the number of elements from its |
2132 | /// input vectors or it may provide demanded bits knowledge via undef lanes. |
2133 | /// Example: <undef,undef,2,3> |
2134 | static bool isIdentityMask(ArrayRef<int> Mask); |
2135 | static bool isIdentityMask(const Constant *Mask) { |
2136 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy( ) && "Shuffle needs vector constant.") ? void (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "llvm/include/llvm/IR/Instructions.h", 2136, __extension__ __PRETTY_FUNCTION__ )); |
2137 | |
2138 | // Not possible to express a shuffle mask for a scalable vector for this |
2139 | // case. |
2140 | if (isa<ScalableVectorType>(Mask->getType())) |
2141 | return false; |
2142 | |
2143 | SmallVector<int, 16> MaskAsInts; |
2144 | getShuffleMask(Mask, MaskAsInts); |
2145 | return isIdentityMask(MaskAsInts); |
2146 | } |
2147 | |
2148 | /// Return true if this shuffle chooses elements from exactly one source |
2149 | /// vector without lane crossings and does not change the number of elements |
2150 | /// from its input vectors. |
2151 | /// Example: shufflevector <4 x n> A, <4 x n> B, <4,undef,6,undef> |
2152 | bool isIdentity() const { |
2153 | // Not possible to express a shuffle mask for a scalable vector for this |
2154 | // case. |
2155 | if (isa<ScalableVectorType>(getType())) |
2156 | return false; |
2157 | |
2158 | return !changesLength() && isIdentityMask(ShuffleMask); |
2159 | } |
2160 | |
2161 | /// Return true if this shuffle lengthens exactly one source vector with |
2162 | /// undefs in the high elements. |
2163 | bool isIdentityWithPadding() const; |
2164 | |
2165 | /// Return true if this shuffle extracts the first N elements of exactly one |
2166 | /// source vector. |
2167 | bool isIdentityWithExtract() const; |
2168 | |
2169 | /// Return true if this shuffle concatenates its 2 source vectors. This |
2170 | /// returns false if either input is undefined. In that case, the shuffle is |
2171 | /// is better classified as an identity with padding operation. |
2172 | bool isConcat() const; |
2173 | |
2174 | /// Return true if this shuffle mask chooses elements from its source vectors |
2175 | /// without lane crossings. A shuffle using this mask would be |
2176 | /// equivalent to a vector select with a constant condition operand. |
2177 | /// Example: <4,1,6,undef> |
2178 | /// This returns false if the mask does not choose from both input vectors. |
2179 | /// In that case, the shuffle is better classified as an identity shuffle. |
2180 | /// This assumes that vector operands are the same length as the mask |
2181 | /// (a length-changing shuffle can never be equivalent to a vector select). |
2182 | static bool isSelectMask(ArrayRef<int> Mask); |
2183 | static bool isSelectMask(const Constant *Mask) { |
2184 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy( ) && "Shuffle needs vector constant.") ? void (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "llvm/include/llvm/IR/Instructions.h", 2184, __extension__ __PRETTY_FUNCTION__ )); |
2185 | SmallVector<int, 16> MaskAsInts; |
2186 | getShuffleMask(Mask, MaskAsInts); |
2187 | return isSelectMask(MaskAsInts); |
2188 | } |
2189 | |
2190 | /// Return true if this shuffle chooses elements from its source vectors |
2191 | /// without lane crossings and all operands have the same number of elements. |
2192 | /// In other words, this shuffle is equivalent to a vector select with a |
2193 | /// constant condition operand. |
2194 | /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,1,6,3> |
2195 | /// This returns false if the mask does not choose from both input vectors. |
2196 | /// In that case, the shuffle is better classified as an identity shuffle. |
2197 | /// TODO: Optionally allow length-changing shuffles. |
2198 | bool isSelect() const { |
2199 | return !changesLength() && isSelectMask(ShuffleMask); |
2200 | } |
2201 | |
2202 | /// Return true if this shuffle mask swaps the order of elements from exactly |
2203 | /// one source vector. |
2204 | /// Example: <7,6,undef,4> |
2205 | /// This assumes that vector operands are the same length as the mask. |
2206 | static bool isReverseMask(ArrayRef<int> Mask); |
2207 | static bool isReverseMask(const Constant *Mask) { |
2208 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy( ) && "Shuffle needs vector constant.") ? void (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "llvm/include/llvm/IR/Instructions.h", 2208, __extension__ __PRETTY_FUNCTION__ )); |
2209 | SmallVector<int, 16> MaskAsInts; |
2210 | getShuffleMask(Mask, MaskAsInts); |
2211 | return isReverseMask(MaskAsInts); |
2212 | } |
2213 | |
2214 | /// Return true if this shuffle swaps the order of elements from exactly |
2215 | /// one source vector. |
2216 | /// Example: shufflevector <4 x n> A, <4 x n> B, <3,undef,1,undef> |
2217 | /// TODO: Optionally allow length-changing shuffles. |
2218 | bool isReverse() const { |
2219 | return !changesLength() && isReverseMask(ShuffleMask); |
2220 | } |
2221 | |
2222 | /// Return true if this shuffle mask chooses all elements with the same value |
2223 | /// as the first element of exactly one source vector. |
2224 | /// Example: <4,undef,undef,4> |
2225 | /// This assumes that vector operands are the same length as the mask. |
2226 | static bool isZeroEltSplatMask(ArrayRef<int> Mask); |
2227 | static bool isZeroEltSplatMask(const Constant *Mask) { |
2228 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy( ) && "Shuffle needs vector constant.") ? void (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "llvm/include/llvm/IR/Instructions.h", 2228, __extension__ __PRETTY_FUNCTION__ )); |
2229 | SmallVector<int, 16> MaskAsInts; |
2230 | getShuffleMask(Mask, MaskAsInts); |
2231 | return isZeroEltSplatMask(MaskAsInts); |
2232 | } |
2233 | |
2234 | /// Return true if all elements of this shuffle are the same value as the |
2235 | /// first element of exactly one source vector without changing the length |
2236 | /// of that vector. |
2237 | /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,0,undef,0> |
2238 | /// TODO: Optionally allow length-changing shuffles. |
2239 | /// TODO: Optionally allow splats from other elements. |
2240 | bool isZeroEltSplat() const { |
2241 | return !changesLength() && isZeroEltSplatMask(ShuffleMask); |
2242 | } |
2243 | |
2244 | /// Return true if this shuffle mask is a transpose mask. |
2245 | /// Transpose vector masks transpose a 2xn matrix. They read corresponding |
2246 | /// even- or odd-numbered vector elements from two n-dimensional source |
2247 | /// vectors and write each result into consecutive elements of an |
2248 | /// n-dimensional destination vector. Two shuffles are necessary to complete |
2249 | /// the transpose, one for the even elements and another for the odd elements. |
2250 | /// This description closely follows how the TRN1 and TRN2 AArch64 |
2251 | /// instructions operate. |
2252 | /// |
2253 | /// For example, a simple 2x2 matrix can be transposed with: |
2254 | /// |
2255 | /// ; Original matrix |
2256 | /// m0 = < a, b > |
2257 | /// m1 = < c, d > |
2258 | /// |
2259 | /// ; Transposed matrix |
2260 | /// t0 = < a, c > = shufflevector m0, m1, < 0, 2 > |
2261 | /// t1 = < b, d > = shufflevector m0, m1, < 1, 3 > |
2262 | /// |
2263 | /// For matrices having greater than n columns, the resulting nx2 transposed |
2264 | /// matrix is stored in two result vectors such that one vector contains |
2265 | /// interleaved elements from all the even-numbered rows and the other vector |
2266 | /// contains interleaved elements from all the odd-numbered rows. For example, |
2267 | /// a 2x4 matrix can be transposed with: |
2268 | /// |
2269 | /// ; Original matrix |
2270 | /// m0 = < a, b, c, d > |
2271 | /// m1 = < e, f, g, h > |
2272 | /// |
2273 | /// ; Transposed matrix |
2274 | /// t0 = < a, e, c, g > = shufflevector m0, m1 < 0, 4, 2, 6 > |
2275 | /// t1 = < b, f, d, h > = shufflevector m0, m1 < 1, 5, 3, 7 > |
2276 | static bool isTransposeMask(ArrayRef<int> Mask); |
2277 | static bool isTransposeMask(const Constant *Mask) { |
2278 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy( ) && "Shuffle needs vector constant.") ? void (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "llvm/include/llvm/IR/Instructions.h", 2278, __extension__ __PRETTY_FUNCTION__ )); |
2279 | SmallVector<int, 16> MaskAsInts; |
2280 | getShuffleMask(Mask, MaskAsInts); |
2281 | return isTransposeMask(MaskAsInts); |
2282 | } |
2283 | |
2284 | /// Return true if this shuffle transposes the elements of its inputs without |
2285 | /// changing the length of the vectors. This operation may also be known as a |
2286 | /// merge or interleave. See the description for isTransposeMask() for the |
2287 | /// exact specification. |
2288 | /// Example: shufflevector <4 x n> A, <4 x n> B, <0,4,2,6> |
2289 | bool isTranspose() const { |
2290 | return !changesLength() && isTransposeMask(ShuffleMask); |
2291 | } |
2292 | |
2293 | /// Return true if this shuffle mask is a splice mask, concatenating the two |
2294 | /// inputs together and then extracts an original width vector starting from |
2295 | /// the splice index. |
2296 | /// Example: shufflevector <4 x n> A, <4 x n> B, <1,2,3,4> |
2297 | static bool isSpliceMask(ArrayRef<int> Mask, int &Index); |
2298 | static bool isSpliceMask(const Constant *Mask, int &Index) { |
2299 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy( ) && "Shuffle needs vector constant.") ? void (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "llvm/include/llvm/IR/Instructions.h", 2299, __extension__ __PRETTY_FUNCTION__ )); |
2300 | SmallVector<int, 16> MaskAsInts; |
2301 | getShuffleMask(Mask, MaskAsInts); |
2302 | return isSpliceMask(MaskAsInts, Index); |
2303 | } |
2304 | |
2305 | /// Return true if this shuffle splices two inputs without changing the length |
2306 | /// of the vectors. This operation concatenates the two inputs together and |
2307 | /// then extracts an original width vector starting from the splice index. |
2308 | /// Example: shufflevector <4 x n> A, <4 x n> B, <1,2,3,4> |
2309 | bool isSplice(int &Index) const { |
2310 | return !changesLength() && isSpliceMask(ShuffleMask, Index); |
2311 | } |
2312 | |
2313 | /// Return true if this shuffle mask is an extract subvector mask. |
2314 | /// A valid extract subvector mask returns a smaller vector from a single |
2315 | /// source operand. The base extraction index is returned as well. |
2316 | static bool isExtractSubvectorMask(ArrayRef<int> Mask, int NumSrcElts, |
2317 | int &Index); |
2318 | static bool isExtractSubvectorMask(const Constant *Mask, int NumSrcElts, |
2319 | int &Index) { |
2320 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy( ) && "Shuffle needs vector constant.") ? void (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "llvm/include/llvm/IR/Instructions.h", 2320, __extension__ __PRETTY_FUNCTION__ )); |
2321 | // Not possible to express a shuffle mask for a scalable vector for this |
2322 | // case. |
2323 | if (isa<ScalableVectorType>(Mask->getType())) |
2324 | return false; |
2325 | SmallVector<int, 16> MaskAsInts; |
2326 | getShuffleMask(Mask, MaskAsInts); |
2327 | return isExtractSubvectorMask(MaskAsInts, NumSrcElts, Index); |
2328 | } |
2329 | |
2330 | /// Return true if this shuffle mask is an extract subvector mask. |
2331 | bool isExtractSubvectorMask(int &Index) const { |
2332 | // Not possible to express a shuffle mask for a scalable vector for this |
2333 | // case. |
2334 | if (isa<ScalableVectorType>(getType())) |
2335 | return false; |
2336 | |
2337 | int NumSrcElts = |
2338 | cast<FixedVectorType>(Op<0>()->getType())->getNumElements(); |
2339 | return isExtractSubvectorMask(ShuffleMask, NumSrcElts, Index); |
2340 | } |
2341 | |
2342 | /// Return true if this shuffle mask is an insert subvector mask. |
2343 | /// A valid insert subvector mask inserts the lowest elements of a second |
2344 | /// source operand into an in-place first source operand operand. |
2345 | /// Both the sub vector width and the insertion index is returned. |
2346 | static bool isInsertSubvectorMask(ArrayRef<int> Mask, int NumSrcElts, |
2347 | int &NumSubElts, int &Index); |
2348 | static bool isInsertSubvectorMask(const Constant *Mask, int NumSrcElts, |
2349 | int &NumSubElts, int &Index) { |
2350 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy( ) && "Shuffle needs vector constant.") ? void (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "llvm/include/llvm/IR/Instructions.h", 2350, __extension__ __PRETTY_FUNCTION__ )); |
2351 | // Not possible to express a shuffle mask for a scalable vector for this |
2352 | // case. |
2353 | if (isa<ScalableVectorType>(Mask->getType())) |
2354 | return false; |
2355 | SmallVector<int, 16> MaskAsInts; |
2356 | getShuffleMask(Mask, MaskAsInts); |
2357 | return isInsertSubvectorMask(MaskAsInts, NumSrcElts, NumSubElts, Index); |
2358 | } |
2359 | |
2360 | /// Return true if this shuffle mask is an insert subvector mask. |
2361 | bool isInsertSubvectorMask(int &NumSubElts, int &Index) const { |
2362 | // Not possible to express a shuffle mask for a scalable vector for this |
2363 | // case. |
2364 | if (isa<ScalableVectorType>(getType())) |
2365 | return false; |
2366 | |
2367 | int NumSrcElts = |
2368 | cast<FixedVectorType>(Op<0>()->getType())->getNumElements(); |
2369 | return isInsertSubvectorMask(ShuffleMask, NumSrcElts, NumSubElts, Index); |
2370 | } |
2371 | |
2372 | /// Return true if this shuffle mask replicates each of the \p VF elements |
2373 | /// in a vector \p ReplicationFactor times. |
2374 | /// For example, the mask for \p ReplicationFactor=3 and \p VF=4 is: |
2375 | /// <0,0,0,1,1,1,2,2,2,3,3,3> |
2376 | static bool isReplicationMask(ArrayRef<int> Mask, int &ReplicationFactor, |
2377 | int &VF); |
2378 | static bool isReplicationMask(const Constant *Mask, int &ReplicationFactor, |
2379 | int &VF) { |
2380 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy( ) && "Shuffle needs vector constant.") ? void (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\"" , "llvm/include/llvm/IR/Instructions.h", 2380, __extension__ __PRETTY_FUNCTION__ )); |
2381 | // Not possible to express a shuffle mask for a scalable vector for this |
2382 | // case. |
2383 | if (isa<ScalableVectorType>(Mask->getType())) |
2384 | return false; |
2385 | SmallVector<int, 16> MaskAsInts; |
2386 | getShuffleMask(Mask, MaskAsInts); |
2387 | return isReplicationMask(MaskAsInts, ReplicationFactor, VF); |
2388 | } |
2389 | |
2390 | /// Return true if this shuffle mask is a replication mask. |
2391 | bool isReplicationMask(int &ReplicationFactor, int &VF) const; |
2392 | |
2393 | /// Return true if this shuffle mask represents "clustered" mask of size VF, |
2394 | /// i.e. each index between [0..VF) is used exactly once in each submask of |
2395 | /// size VF. |
2396 | /// For example, the mask for \p VF=4 is: |
2397 | /// 0, 1, 2, 3, 3, 2, 0, 1 - "clustered", because each submask of size 4 |
2398 | /// (0,1,2,3 and 3,2,0,1) uses indices [0..VF) exactly one time. |
2399 | /// 0, 1, 2, 3, 3, 3, 1, 0 - not "clustered", because |
2400 | /// element 3 is used twice in the second submask |
2401 | /// (3,3,1,0) and index 2 is not used at all. |
2402 | static bool isOneUseSingleSourceMask(ArrayRef<int> Mask, int VF); |
2403 | |
2404 | /// Return true if this shuffle mask is a one-use-single-source("clustered") |
2405 | /// mask. |
2406 | bool isOneUseSingleSourceMask(int VF) const; |
2407 | |
2408 | /// Change values in a shuffle permute mask assuming the two vector operands |
2409 | /// of length InVecNumElts have swapped position. |
2410 | static void commuteShuffleMask(MutableArrayRef<int> Mask, |
2411 | unsigned InVecNumElts) { |
2412 | for (int &Idx : Mask) { |
2413 | if (Idx == -1) |
2414 | continue; |
2415 | Idx = Idx < (int)InVecNumElts ? Idx + InVecNumElts : Idx - InVecNumElts; |
2416 | assert(Idx >= 0 && Idx < (int)InVecNumElts * 2 &&(static_cast <bool> (Idx >= 0 && Idx < (int )InVecNumElts * 2 && "shufflevector mask index out of range" ) ? void (0) : __assert_fail ("Idx >= 0 && Idx < (int)InVecNumElts * 2 && \"shufflevector mask index out of range\"" , "llvm/include/llvm/IR/Instructions.h", 2417, __extension__ __PRETTY_FUNCTION__ )) |
2417 | "shufflevector mask index out of range")(static_cast <bool> (Idx >= 0 && Idx < (int )InVecNumElts * 2 && "shufflevector mask index out of range" ) ? void (0) : __assert_fail ("Idx >= 0 && Idx < (int)InVecNumElts * 2 && \"shufflevector mask index out of range\"" , "llvm/include/llvm/IR/Instructions.h", 2417, __extension__ __PRETTY_FUNCTION__ )); |
2418 | } |
2419 | } |
2420 | |
2421 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2422 | static bool classof(const Instruction *I) { |
2423 | return I->getOpcode() == Instruction::ShuffleVector; |
2424 | } |
2425 | static bool classof(const Value *V) { |
2426 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2427 | } |
2428 | }; |
2429 | |
2430 | template <> |
2431 | struct OperandTraits<ShuffleVectorInst> |
2432 | : public FixedNumOperandTraits<ShuffleVectorInst, 2> {}; |
2433 | |
2434 | 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 { (static_cast <bool> (i_nocapture < OperandTraits<ShuffleVectorInst>::operands(this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<ShuffleVectorInst>::operands(this) && \"getOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 2434, __extension__ __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) { (static_cast <bool> (i_nocapture < OperandTraits<ShuffleVectorInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<ShuffleVectorInst>::operands(this) && \"setOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 2434, __extension__ __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 ); } |
2435 | |
2436 | //===----------------------------------------------------------------------===// |
2437 | // ExtractValueInst Class |
2438 | //===----------------------------------------------------------------------===// |
2439 | |
2440 | /// This instruction extracts a struct member or array |
2441 | /// element value from an aggregate value. |
2442 | /// |
2443 | class ExtractValueInst : public UnaryInstruction { |
2444 | SmallVector<unsigned, 4> Indices; |
2445 | |
2446 | ExtractValueInst(const ExtractValueInst &EVI); |
2447 | |
2448 | /// Constructors - Create a extractvalue instruction with a base aggregate |
2449 | /// value and a list of indices. The first ctor can optionally insert before |
2450 | /// an existing instruction, the second appends the new instruction to the |
2451 | /// specified BasicBlock. |
2452 | inline ExtractValueInst(Value *Agg, |
2453 | ArrayRef<unsigned> Idxs, |
2454 | const Twine &NameStr, |
2455 | Instruction *InsertBefore); |
2456 | inline ExtractValueInst(Value *Agg, |
2457 | ArrayRef<unsigned> Idxs, |
2458 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2459 | |
2460 | void init(ArrayRef<unsigned> Idxs, const Twine &NameStr); |
2461 | |
2462 | protected: |
2463 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2464 | friend class Instruction; |
2465 | |
2466 | ExtractValueInst *cloneImpl() const; |
2467 | |
2468 | public: |
2469 | static ExtractValueInst *Create(Value *Agg, |
2470 | ArrayRef<unsigned> Idxs, |
2471 | const Twine &NameStr = "", |
2472 | Instruction *InsertBefore = nullptr) { |
2473 | return new |
2474 | ExtractValueInst(Agg, Idxs, NameStr, InsertBefore); |
2475 | } |
2476 | |
2477 | static ExtractValueInst *Create(Value *Agg, |
2478 | ArrayRef<unsigned> Idxs, |
2479 | const Twine &NameStr, |
2480 | BasicBlock *InsertAtEnd) { |
2481 | return new ExtractValueInst(Agg, Idxs, NameStr, InsertAtEnd); |
2482 | } |
2483 | |
2484 | /// Returns the type of the element that would be extracted |
2485 | /// with an extractvalue instruction with the specified parameters. |
2486 | /// |
2487 | /// Null is returned if the indices are invalid for the specified type. |
2488 | static Type *getIndexedType(Type *Agg, ArrayRef<unsigned> Idxs); |
2489 | |
2490 | using idx_iterator = const unsigned*; |
2491 | |
2492 | inline idx_iterator idx_begin() const { return Indices.begin(); } |
2493 | inline idx_iterator idx_end() const { return Indices.end(); } |
2494 | inline iterator_range<idx_iterator> indices() const { |
2495 | return make_range(idx_begin(), idx_end()); |
2496 | } |
2497 | |
2498 | Value *getAggregateOperand() { |
2499 | return getOperand(0); |
2500 | } |
2501 | const Value *getAggregateOperand() const { |
2502 | return getOperand(0); |
2503 | } |
2504 | static unsigned getAggregateOperandIndex() { |
2505 | return 0U; // get index for modifying correct operand |
2506 | } |
2507 | |
2508 | ArrayRef<unsigned> getIndices() const { |
2509 | return Indices; |
2510 | } |
2511 | |
2512 | unsigned getNumIndices() const { |
2513 | return (unsigned)Indices.size(); |
2514 | } |
2515 | |
2516 | bool hasIndices() const { |
2517 | return true; |
2518 | } |
2519 | |
2520 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2521 | static bool classof(const Instruction *I) { |
2522 | return I->getOpcode() == Instruction::ExtractValue; |
2523 | } |
2524 | static bool classof(const Value *V) { |
2525 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2526 | } |
2527 | }; |
2528 | |
2529 | ExtractValueInst::ExtractValueInst(Value *Agg, |
2530 | ArrayRef<unsigned> Idxs, |
2531 | const Twine &NameStr, |
2532 | Instruction *InsertBefore) |
2533 | : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)), |
2534 | ExtractValue, Agg, InsertBefore) { |
2535 | init(Idxs, NameStr); |
2536 | } |
2537 | |
2538 | ExtractValueInst::ExtractValueInst(Value *Agg, |
2539 | ArrayRef<unsigned> Idxs, |
2540 | const Twine &NameStr, |
2541 | BasicBlock *InsertAtEnd) |
2542 | : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)), |
2543 | ExtractValue, Agg, InsertAtEnd) { |
2544 | init(Idxs, NameStr); |
2545 | } |
2546 | |
2547 | //===----------------------------------------------------------------------===// |
2548 | // InsertValueInst Class |
2549 | //===----------------------------------------------------------------------===// |
2550 | |
2551 | /// This instruction inserts a struct field of array element |
2552 | /// value into an aggregate value. |
2553 | /// |
2554 | class InsertValueInst : public Instruction { |
2555 | SmallVector<unsigned, 4> Indices; |
2556 | |
2557 | InsertValueInst(const InsertValueInst &IVI); |
2558 | |
2559 | /// Constructors - Create a insertvalue instruction with a base aggregate |
2560 | /// value, a value to insert, and a list of indices. The first ctor can |
2561 | /// optionally insert before an existing instruction, the second appends |
2562 | /// the new instruction to the specified BasicBlock. |
2563 | inline InsertValueInst(Value *Agg, Value *Val, |
2564 | ArrayRef<unsigned> Idxs, |
2565 | const Twine &NameStr, |
2566 | Instruction *InsertBefore); |
2567 | inline InsertValueInst(Value *Agg, Value *Val, |
2568 | ArrayRef<unsigned> Idxs, |
2569 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2570 | |
2571 | /// Constructors - These two constructors are convenience methods because one |
2572 | /// and two index insertvalue instructions are so common. |
2573 | InsertValueInst(Value *Agg, Value *Val, unsigned Idx, |
2574 | const Twine &NameStr = "", |
2575 | Instruction *InsertBefore = nullptr); |
2576 | InsertValueInst(Value *Agg, Value *Val, unsigned Idx, const Twine &NameStr, |
2577 | BasicBlock *InsertAtEnd); |
2578 | |
2579 | void init(Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, |
2580 | const Twine &NameStr); |
2581 | |
2582 | protected: |
2583 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2584 | friend class Instruction; |
2585 | |
2586 | InsertValueInst *cloneImpl() const; |
2587 | |
2588 | public: |
2589 | // allocate space for exactly two operands |
2590 | void *operator new(size_t S) { return User::operator new(S, 2); } |
2591 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
2592 | |
2593 | static InsertValueInst *Create(Value *Agg, Value *Val, |
2594 | ArrayRef<unsigned> Idxs, |
2595 | const Twine &NameStr = "", |
2596 | Instruction *InsertBefore = nullptr) { |
2597 | return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertBefore); |
2598 | } |
2599 | |
2600 | static InsertValueInst *Create(Value *Agg, Value *Val, |
2601 | ArrayRef<unsigned> Idxs, |
2602 | const Twine &NameStr, |
2603 | BasicBlock *InsertAtEnd) { |
2604 | return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertAtEnd); |
2605 | } |
2606 | |
2607 | /// Transparently provide more efficient getOperand methods. |
2608 | 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; |
2609 | |
2610 | using idx_iterator = const unsigned*; |
2611 | |
2612 | inline idx_iterator idx_begin() const { return Indices.begin(); } |
2613 | inline idx_iterator idx_end() const { return Indices.end(); } |
2614 | inline iterator_range<idx_iterator> indices() const { |
2615 | return make_range(idx_begin(), idx_end()); |
2616 | } |
2617 | |
2618 | Value *getAggregateOperand() { |
2619 | return getOperand(0); |
2620 | } |
2621 | const Value *getAggregateOperand() const { |
2622 | return getOperand(0); |
2623 | } |
2624 | static unsigned getAggregateOperandIndex() { |
2625 | return 0U; // get index for modifying correct operand |
2626 | } |
2627 | |
2628 | Value *getInsertedValueOperand() { |
2629 | return getOperand(1); |
2630 | } |
2631 | const Value *getInsertedValueOperand() const { |
2632 | return getOperand(1); |
2633 | } |
2634 | static unsigned getInsertedValueOperandIndex() { |
2635 | return 1U; // get index for modifying correct operand |
2636 | } |
2637 | |
2638 | ArrayRef<unsigned> getIndices() const { |
2639 | return Indices; |
2640 | } |
2641 | |
2642 | unsigned getNumIndices() const { |
2643 | return (unsigned)Indices.size(); |
2644 | } |
2645 | |
2646 | bool hasIndices() const { |
2647 | return true; |
2648 | } |
2649 | |
2650 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2651 | static bool classof(const Instruction *I) { |
2652 | return I->getOpcode() == Instruction::InsertValue; |
2653 | } |
2654 | static bool classof(const Value *V) { |
2655 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2656 | } |
2657 | }; |
2658 | |
2659 | template <> |
2660 | struct OperandTraits<InsertValueInst> : |
2661 | public FixedNumOperandTraits<InsertValueInst, 2> { |
2662 | }; |
2663 | |
2664 | InsertValueInst::InsertValueInst(Value *Agg, |
2665 | Value *Val, |
2666 | ArrayRef<unsigned> Idxs, |
2667 | const Twine &NameStr, |
2668 | Instruction *InsertBefore) |
2669 | : Instruction(Agg->getType(), InsertValue, |
2670 | OperandTraits<InsertValueInst>::op_begin(this), |
2671 | 2, InsertBefore) { |
2672 | init(Agg, Val, Idxs, NameStr); |
2673 | } |
2674 | |
2675 | InsertValueInst::InsertValueInst(Value *Agg, |
2676 | Value *Val, |
2677 | ArrayRef<unsigned> Idxs, |
2678 | const Twine &NameStr, |
2679 | BasicBlock *InsertAtEnd) |
2680 | : Instruction(Agg->getType(), InsertValue, |
2681 | OperandTraits<InsertValueInst>::op_begin(this), |
2682 | 2, InsertAtEnd) { |
2683 | init(Agg, Val, Idxs, NameStr); |
2684 | } |
2685 | |
2686 | 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 { (static_cast <bool > (i_nocapture < OperandTraits<InsertValueInst>:: operands(this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<InsertValueInst>::operands(this) && \"getOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 2686, __extension__ __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) { (static_cast <bool> (i_nocapture < OperandTraits<InsertValueInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<InsertValueInst>::operands(this) && \"setOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 2686, __extension__ __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 ); } |
2687 | |
2688 | //===----------------------------------------------------------------------===// |
2689 | // PHINode Class |
2690 | //===----------------------------------------------------------------------===// |
2691 | |
2692 | // PHINode - The PHINode class is used to represent the magical mystical PHI |
2693 | // node, that can not exist in nature, but can be synthesized in a computer |
2694 | // scientist's overactive imagination. |
2695 | // |
2696 | class PHINode : public Instruction { |
2697 | /// The number of operands actually allocated. NumOperands is |
2698 | /// the number actually in use. |
2699 | unsigned ReservedSpace; |
2700 | |
2701 | PHINode(const PHINode &PN); |
2702 | |
2703 | explicit PHINode(Type *Ty, unsigned NumReservedValues, |
2704 | const Twine &NameStr = "", |
2705 | Instruction *InsertBefore = nullptr) |
2706 | : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertBefore), |
2707 | ReservedSpace(NumReservedValues) { |
2708 | assert(!Ty->isTokenTy() && "PHI nodes cannot have token type!")(static_cast <bool> (!Ty->isTokenTy() && "PHI nodes cannot have token type!" ) ? void (0) : __assert_fail ("!Ty->isTokenTy() && \"PHI nodes cannot have token type!\"" , "llvm/include/llvm/IR/Instructions.h", 2708, __extension__ __PRETTY_FUNCTION__ )); |
2709 | setName(NameStr); |
2710 | allocHungoffUses(ReservedSpace); |
2711 | } |
2712 | |
2713 | PHINode(Type *Ty, unsigned NumReservedValues, const Twine &NameStr, |
2714 | BasicBlock *InsertAtEnd) |
2715 | : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertAtEnd), |
2716 | ReservedSpace(NumReservedValues) { |
2717 | assert(!Ty->isTokenTy() && "PHI nodes cannot have token type!")(static_cast <bool> (!Ty->isTokenTy() && "PHI nodes cannot have token type!" ) ? void (0) : __assert_fail ("!Ty->isTokenTy() && \"PHI nodes cannot have token type!\"" , "llvm/include/llvm/IR/Instructions.h", 2717, __extension__ __PRETTY_FUNCTION__ )); |
2718 | setName(NameStr); |
2719 | allocHungoffUses(ReservedSpace); |
2720 | } |
2721 | |
2722 | protected: |
2723 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2724 | friend class Instruction; |
2725 | |
2726 | PHINode *cloneImpl() const; |
2727 | |
2728 | // allocHungoffUses - this is more complicated than the generic |
2729 | // User::allocHungoffUses, because we have to allocate Uses for the incoming |
2730 | // values and pointers to the incoming blocks, all in one allocation. |
2731 | void allocHungoffUses(unsigned N) { |
2732 | User::allocHungoffUses(N, /* IsPhi */ true); |
2733 | } |
2734 | |
2735 | public: |
2736 | /// Constructors - NumReservedValues is a hint for the number of incoming |
2737 | /// edges that this phi node will have (use 0 if you really have no idea). |
2738 | static PHINode *Create(Type *Ty, unsigned NumReservedValues, |
2739 | const Twine &NameStr = "", |
2740 | Instruction *InsertBefore = nullptr) { |
2741 | return new PHINode(Ty, NumReservedValues, NameStr, InsertBefore); |
2742 | } |
2743 | |
2744 | static PHINode *Create(Type *Ty, unsigned NumReservedValues, |
2745 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
2746 | return new PHINode(Ty, NumReservedValues, NameStr, InsertAtEnd); |
2747 | } |
2748 | |
2749 | /// Provide fast operand accessors |
2750 | 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; |
2751 | |
2752 | // Block iterator interface. This provides access to the list of incoming |
2753 | // basic blocks, which parallels the list of incoming values. |
2754 | // Please note that we are not providing non-const iterators for blocks to |
2755 | // force all updates go through an interface function. |
2756 | |
2757 | using block_iterator = BasicBlock **; |
2758 | using const_block_iterator = BasicBlock * const *; |
2759 | |
2760 | const_block_iterator block_begin() const { |
2761 | return reinterpret_cast<const_block_iterator>(op_begin() + ReservedSpace); |
2762 | } |
2763 | |
2764 | const_block_iterator block_end() const { |
2765 | return block_begin() + getNumOperands(); |
2766 | } |
2767 | |
2768 | iterator_range<const_block_iterator> blocks() const { |
2769 | return make_range(block_begin(), block_end()); |
2770 | } |
2771 | |
2772 | op_range incoming_values() { return operands(); } |
2773 | |
2774 | const_op_range incoming_values() const { return operands(); } |
2775 | |
2776 | /// Return the number of incoming edges |
2777 | /// |
2778 | unsigned getNumIncomingValues() const { return getNumOperands(); } |
2779 | |
2780 | /// Return incoming value number x |
2781 | /// |
2782 | Value *getIncomingValue(unsigned i) const { |
2783 | return getOperand(i); |
2784 | } |
2785 | void setIncomingValue(unsigned i, Value *V) { |
2786 | assert(V && "PHI node got a null value!")(static_cast <bool> (V && "PHI node got a null value!" ) ? void (0) : __assert_fail ("V && \"PHI node got a null value!\"" , "llvm/include/llvm/IR/Instructions.h", 2786, __extension__ __PRETTY_FUNCTION__ )); |
2787 | assert(getType() == V->getType() &&(static_cast <bool> (getType() == V->getType() && "All operands to PHI node must be the same type as the PHI node!" ) ? void (0) : __assert_fail ("getType() == V->getType() && \"All operands to PHI node must be the same type as the PHI node!\"" , "llvm/include/llvm/IR/Instructions.h", 2788, __extension__ __PRETTY_FUNCTION__ )) |
2788 | "All operands to PHI node must be the same type as the PHI node!")(static_cast <bool> (getType() == V->getType() && "All operands to PHI node must be the same type as the PHI node!" ) ? void (0) : __assert_fail ("getType() == V->getType() && \"All operands to PHI node must be the same type as the PHI node!\"" , "llvm/include/llvm/IR/Instructions.h", 2788, __extension__ __PRETTY_FUNCTION__ )); |
2789 | setOperand(i, V); |
2790 | } |
2791 | |
2792 | static unsigned getOperandNumForIncomingValue(unsigned i) { |
2793 | return i; |
2794 | } |
2795 | |
2796 | static unsigned getIncomingValueNumForOperand(unsigned i) { |
2797 | return i; |
2798 | } |
2799 | |
2800 | /// Return incoming basic block number @p i. |
2801 | /// |
2802 | BasicBlock *getIncomingBlock(unsigned i) const { |
2803 | return block_begin()[i]; |
2804 | } |
2805 | |
2806 | /// Return incoming basic block corresponding |
2807 | /// to an operand of the PHI. |
2808 | /// |
2809 | BasicBlock *getIncomingBlock(const Use &U) const { |
2810 | assert(this == U.getUser() && "Iterator doesn't point to PHI's Uses?")(static_cast <bool> (this == U.getUser() && "Iterator doesn't point to PHI's Uses?" ) ? void (0) : __assert_fail ("this == U.getUser() && \"Iterator doesn't point to PHI's Uses?\"" , "llvm/include/llvm/IR/Instructions.h", 2810, __extension__ __PRETTY_FUNCTION__ )); |
2811 | return getIncomingBlock(unsigned(&U - op_begin())); |
2812 | } |
2813 | |
2814 | /// Return incoming basic block corresponding |
2815 | /// to value use iterator. |
2816 | /// |
2817 | BasicBlock *getIncomingBlock(Value::const_user_iterator I) const { |
2818 | return getIncomingBlock(I.getUse()); |
2819 | } |
2820 | |
2821 | void setIncomingBlock(unsigned i, BasicBlock *BB) { |
2822 | const_cast<block_iterator>(block_begin())[i] = BB; |
2823 | } |
2824 | |
2825 | /// Copies the basic blocks from \p BBRange to the incoming basic block list |
2826 | /// of this PHINode, starting at \p ToIdx. |
2827 | void copyIncomingBlocks(iterator_range<const_block_iterator> BBRange, |
2828 | uint32_t ToIdx = 0) { |
2829 | copy(BBRange, const_cast<block_iterator>(block_begin()) + ToIdx); |
2830 | } |
2831 | |
2832 | /// Replace every incoming basic block \p Old to basic block \p New. |
2833 | void replaceIncomingBlockWith(const BasicBlock *Old, BasicBlock *New) { |
2834 | assert(New && Old && "PHI node got a null basic block!")(static_cast <bool> (New && Old && "PHI node got a null basic block!" ) ? void (0) : __assert_fail ("New && Old && \"PHI node got a null basic block!\"" , "llvm/include/llvm/IR/Instructions.h", 2834, __extension__ __PRETTY_FUNCTION__ )); |
2835 | for (unsigned Op = 0, NumOps = getNumOperands(); Op != NumOps; ++Op) |
2836 | if (getIncomingBlock(Op) == Old) |
2837 | setIncomingBlock(Op, New); |
2838 | } |
2839 | |
2840 | /// Add an incoming value to the end of the PHI list |
2841 | /// |
2842 | void addIncoming(Value *V, BasicBlock *BB) { |
2843 | if (getNumOperands() == ReservedSpace) |
2844 | growOperands(); // Get more space! |
2845 | // Initialize some new operands. |
2846 | setNumHungOffUseOperands(getNumOperands() + 1); |
2847 | setIncomingValue(getNumOperands() - 1, V); |
2848 | setIncomingBlock(getNumOperands() - 1, BB); |
2849 | } |
2850 | |
2851 | /// Remove an incoming value. This is useful if a |
2852 | /// predecessor basic block is deleted. The value removed is returned. |
2853 | /// |
2854 | /// If the last incoming value for a PHI node is removed (and DeletePHIIfEmpty |
2855 | /// is true), the PHI node is destroyed and any uses of it are replaced with |
2856 | /// dummy values. The only time there should be zero incoming values to a PHI |
2857 | /// node is when the block is dead, so this strategy is sound. |
2858 | /// |
2859 | Value *removeIncomingValue(unsigned Idx, bool DeletePHIIfEmpty = true); |
2860 | |
2861 | Value *removeIncomingValue(const BasicBlock *BB, bool DeletePHIIfEmpty=true) { |
2862 | int Idx = getBasicBlockIndex(BB); |
2863 | assert(Idx >= 0 && "Invalid basic block argument to remove!")(static_cast <bool> (Idx >= 0 && "Invalid basic block argument to remove!" ) ? void (0) : __assert_fail ("Idx >= 0 && \"Invalid basic block argument to remove!\"" , "llvm/include/llvm/IR/Instructions.h", 2863, __extension__ __PRETTY_FUNCTION__ )); |
2864 | return removeIncomingValue(Idx, DeletePHIIfEmpty); |
2865 | } |
2866 | |
2867 | /// Return the first index of the specified basic |
2868 | /// block in the value list for this PHI. Returns -1 if no instance. |
2869 | /// |
2870 | int getBasicBlockIndex(const BasicBlock *BB) const { |
2871 | for (unsigned i = 0, e = getNumOperands(); i != e; ++i) |
2872 | if (block_begin()[i] == BB) |
2873 | return i; |
2874 | return -1; |
2875 | } |
2876 | |
2877 | Value *getIncomingValueForBlock(const BasicBlock *BB) const { |
2878 | int Idx = getBasicBlockIndex(BB); |
2879 | assert(Idx >= 0 && "Invalid basic block argument!")(static_cast <bool> (Idx >= 0 && "Invalid basic block argument!" ) ? void (0) : __assert_fail ("Idx >= 0 && \"Invalid basic block argument!\"" , "llvm/include/llvm/IR/Instructions.h", 2879, __extension__ __PRETTY_FUNCTION__ )); |
2880 | return getIncomingValue(Idx); |
2881 | } |
2882 | |
2883 | /// Set every incoming value(s) for block \p BB to \p V. |
2884 | void setIncomingValueForBlock(const BasicBlock *BB, Value *V) { |
2885 | assert(BB && "PHI node got a null basic block!")(static_cast <bool> (BB && "PHI node got a null basic block!" ) ? void (0) : __assert_fail ("BB && \"PHI node got a null basic block!\"" , "llvm/include/llvm/IR/Instructions.h", 2885, __extension__ __PRETTY_FUNCTION__ )); |
2886 | bool Found = false; |
2887 | for (unsigned Op = 0, NumOps = getNumOperands(); Op != NumOps; ++Op) |
2888 | if (getIncomingBlock(Op) == BB) { |
2889 | Found = true; |
2890 | setIncomingValue(Op, V); |
2891 | } |
2892 | (void)Found; |
2893 | assert(Found && "Invalid basic block argument to set!")(static_cast <bool> (Found && "Invalid basic block argument to set!" ) ? void (0) : __assert_fail ("Found && \"Invalid basic block argument to set!\"" , "llvm/include/llvm/IR/Instructions.h", 2893, __extension__ __PRETTY_FUNCTION__ )); |
2894 | } |
2895 | |
2896 | /// If the specified PHI node always merges together the |
2897 | /// same value, return the value, otherwise return null. |
2898 | Value *hasConstantValue() const; |
2899 | |
2900 | /// Whether the specified PHI node always merges |
2901 | /// together the same value, assuming undefs are equal to a unique |
2902 | /// non-undef value. |
2903 | bool hasConstantOrUndefValue() const; |
2904 | |
2905 | /// If the PHI node is complete which means all of its parent's predecessors |
2906 | /// have incoming value in this PHI, return true, otherwise return false. |
2907 | bool isComplete() const { |
2908 | return llvm::all_of(predecessors(getParent()), |
2909 | [this](const BasicBlock *Pred) { |
2910 | return getBasicBlockIndex(Pred) >= 0; |
2911 | }); |
2912 | } |
2913 | |
2914 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
2915 | static bool classof(const Instruction *I) { |
2916 | return I->getOpcode() == Instruction::PHI; |
2917 | } |
2918 | static bool classof(const Value *V) { |
2919 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2920 | } |
2921 | |
2922 | private: |
2923 | void growOperands(); |
2924 | }; |
2925 | |
2926 | template <> |
2927 | struct OperandTraits<PHINode> : public HungoffOperandTraits<2> { |
2928 | }; |
2929 | |
2930 | 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 { (static_cast <bool> (i_nocapture < OperandTraits <PHINode>::operands(this) && "getOperand() out of range!" ) ? void (0) : __assert_fail ("i_nocapture < OperandTraits<PHINode>::operands(this) && \"getOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 2930, __extension__ __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) { (static_cast <bool> (i_nocapture < OperandTraits<PHINode>::operands(this) && "setOperand() out of range!" ) ? void (0) : __assert_fail ("i_nocapture < OperandTraits<PHINode>::operands(this) && \"setOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 2930, __extension__ __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); } |
2931 | |
2932 | //===----------------------------------------------------------------------===// |
2933 | // LandingPadInst Class |
2934 | //===----------------------------------------------------------------------===// |
2935 | |
2936 | //===--------------------------------------------------------------------------- |
2937 | /// The landingpad instruction holds all of the information |
2938 | /// necessary to generate correct exception handling. The landingpad instruction |
2939 | /// cannot be moved from the top of a landing pad block, which itself is |
2940 | /// accessible only from the 'unwind' edge of an invoke. This uses the |
2941 | /// SubclassData field in Value to store whether or not the landingpad is a |
2942 | /// cleanup. |
2943 | /// |
2944 | class LandingPadInst : public Instruction { |
2945 | using CleanupField = BoolBitfieldElementT<0>; |
2946 | |
2947 | /// The number of operands actually allocated. NumOperands is |
2948 | /// the number actually in use. |
2949 | unsigned ReservedSpace; |
2950 | |
2951 | LandingPadInst(const LandingPadInst &LP); |
2952 | |
2953 | public: |
2954 | enum ClauseType { Catch, Filter }; |
2955 | |
2956 | private: |
2957 | explicit LandingPadInst(Type *RetTy, unsigned NumReservedValues, |
2958 | const Twine &NameStr, Instruction *InsertBefore); |
2959 | explicit LandingPadInst(Type *RetTy, unsigned NumReservedValues, |
2960 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2961 | |
2962 | // Allocate space for exactly zero operands. |
2963 | void *operator new(size_t S) { return User::operator new(S); } |
2964 | |
2965 | void growOperands(unsigned Size); |
2966 | void init(unsigned NumReservedValues, const Twine &NameStr); |
2967 | |
2968 | protected: |
2969 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2970 | friend class Instruction; |
2971 | |
2972 | LandingPadInst *cloneImpl() const; |
2973 | |
2974 | public: |
2975 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
2976 | |
2977 | /// Constructors - NumReservedClauses is a hint for the number of incoming |
2978 | /// clauses that this landingpad will have (use 0 if you really have no idea). |
2979 | static LandingPadInst *Create(Type *RetTy, unsigned NumReservedClauses, |
2980 | const Twine &NameStr = "", |
2981 | Instruction *InsertBefore = nullptr); |
2982 | static LandingPadInst *Create(Type *RetTy, unsigned NumReservedClauses, |
2983 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2984 | |
2985 | /// Provide fast operand accessors |
2986 | 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; |
2987 | |
2988 | /// Return 'true' if this landingpad instruction is a |
2989 | /// cleanup. I.e., it should be run when unwinding even if its landing pad |
2990 | /// doesn't catch the exception. |
2991 | bool isCleanup() const { return getSubclassData<CleanupField>(); } |
2992 | |
2993 | /// Indicate that this landingpad instruction is a cleanup. |
2994 | void setCleanup(bool V) { setSubclassData<CleanupField>(V); } |
2995 | |
2996 | /// Add a catch or filter clause to the landing pad. |
2997 | void addClause(Constant *ClauseVal); |
2998 | |
2999 | /// Get the value of the clause at index Idx. Use isCatch/isFilter to |
3000 | /// determine what type of clause this is. |
3001 | Constant *getClause(unsigned Idx) const { |
3002 | return cast<Constant>(getOperandList()[Idx]); |
3003 | } |
3004 | |
3005 | /// Return 'true' if the clause and index Idx is a catch clause. |
3006 | bool isCatch(unsigned Idx) const { |
3007 | return !isa<ArrayType>(getOperandList()[Idx]->getType()); |
3008 | } |
3009 | |
3010 | /// Return 'true' if the clause and index Idx is a filter clause. |
3011 | bool isFilter(unsigned Idx) const { |
3012 | return isa<ArrayType>(getOperandList()[Idx]->getType()); |
3013 | } |
3014 | |
3015 | /// Get the number of clauses for this landing pad. |
3016 | unsigned getNumClauses() const { return getNumOperands(); } |
3017 | |
3018 | /// Grow the size of the operand list to accommodate the new |
3019 | /// number of clauses. |
3020 | void reserveClauses(unsigned Size) { growOperands(Size); } |
3021 | |
3022 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
3023 | static bool classof(const Instruction *I) { |
3024 | return I->getOpcode() == Instruction::LandingPad; |
3025 | } |
3026 | static bool classof(const Value *V) { |
3027 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
3028 | } |
3029 | }; |
3030 | |
3031 | template <> |
3032 | struct OperandTraits<LandingPadInst> : public HungoffOperandTraits<1> { |
3033 | }; |
3034 | |
3035 | 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 { (static_cast <bool > (i_nocapture < OperandTraits<LandingPadInst>::operands (this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<LandingPadInst>::operands(this) && \"getOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 3035, __extension__ __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) { (static_cast <bool> (i_nocapture < OperandTraits<LandingPadInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<LandingPadInst>::operands(this) && \"setOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 3035, __extension__ __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 ); } |
3036 | |
3037 | //===----------------------------------------------------------------------===// |
3038 | // ReturnInst Class |
3039 | //===----------------------------------------------------------------------===// |
3040 | |
3041 | //===--------------------------------------------------------------------------- |
3042 | /// Return a value (possibly void), from a function. Execution |
3043 | /// does not continue in this function any longer. |
3044 | /// |
3045 | class ReturnInst : public Instruction { |
3046 | ReturnInst(const ReturnInst &RI); |
3047 | |
3048 | private: |
3049 | // ReturnInst constructors: |
3050 | // ReturnInst() - 'ret void' instruction |
3051 | // ReturnInst( null) - 'ret void' instruction |
3052 | // ReturnInst(Value* X) - 'ret X' instruction |
3053 | // ReturnInst( null, Inst *I) - 'ret void' instruction, insert before I |
3054 | // ReturnInst(Value* X, Inst *I) - 'ret X' instruction, insert before I |
3055 | // ReturnInst( null, BB *B) - 'ret void' instruction, insert @ end of B |
3056 | // ReturnInst(Value* X, BB *B) - 'ret X' instruction, insert @ end of B |
3057 | // |
3058 | // NOTE: If the Value* passed is of type void then the constructor behaves as |
3059 | // if it was passed NULL. |
3060 | explicit ReturnInst(LLVMContext &C, Value *retVal = nullptr, |
3061 | Instruction *InsertBefore = nullptr); |
3062 | ReturnInst(LLVMContext &C, Value *retVal, BasicBlock *InsertAtEnd); |
3063 | explicit ReturnInst(LLVMContext &C, BasicBlock *InsertAtEnd); |
3064 | |
3065 | protected: |
3066 | // Note: Instruction needs to be a friend here to call cloneImpl. |
3067 | friend class Instruction; |
3068 | |
3069 | ReturnInst *cloneImpl() const; |
3070 | |
3071 | public: |
3072 | static ReturnInst* Create(LLVMContext &C, Value *retVal = nullptr, |
3073 | Instruction *InsertBefore = nullptr) { |
3074 | return new(!!retVal) ReturnInst(C, retVal, InsertBefore); |
3075 | } |
3076 | |
3077 | static ReturnInst* Create(LLVMContext &C, Value *retVal, |
3078 | BasicBlock *InsertAtEnd) { |
3079 | return new(!!retVal) ReturnInst(C, retVal, InsertAtEnd); |
3080 | } |
3081 | |
3082 | static ReturnInst* Create(LLVMContext &C, BasicBlock *InsertAtEnd) { |
3083 | return new(0) ReturnInst(C, InsertAtEnd); |
3084 | } |
3085 | |
3086 | /// Provide fast operand accessors |
3087 | 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; |
3088 | |
3089 | /// Convenience accessor. Returns null if there is no return value. |
3090 | Value *getReturnValue() const { |
3091 | return getNumOperands() != 0 ? getOperand(0) : nullptr; |
3092 | } |
3093 | |
3094 | unsigned getNumSuccessors() const { return 0; } |
3095 | |
3096 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
3097 | static bool classof(const Instruction *I) { |
3098 | return (I->getOpcode() == Instruction::Ret); |
3099 | } |
3100 | static bool classof(const Value *V) { |
3101 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
3102 | } |
3103 | |
3104 | private: |
3105 | BasicBlock *getSuccessor(unsigned idx) const { |
3106 | llvm_unreachable("ReturnInst has no successors!")::llvm::llvm_unreachable_internal("ReturnInst has no successors!" , "llvm/include/llvm/IR/Instructions.h", 3106); |
3107 | } |
3108 | |
3109 | void setSuccessor(unsigned idx, BasicBlock *B) { |
3110 | llvm_unreachable("ReturnInst has no successors!")::llvm::llvm_unreachable_internal("ReturnInst has no successors!" , "llvm/include/llvm/IR/Instructions.h", 3110); |
3111 | } |
3112 | }; |
3113 | |
3114 | template <> |
3115 | struct OperandTraits<ReturnInst> : public VariadicOperandTraits<ReturnInst> { |
3116 | }; |
3117 | |
3118 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ReturnInst, Value)ReturnInst::op_iterator ReturnInst::op_begin() { return OperandTraits <ReturnInst>::op_begin(this); } ReturnInst::const_op_iterator ReturnInst::op_begin() const { return OperandTraits<ReturnInst >::op_begin(const_cast<ReturnInst*>(this)); } ReturnInst ::op_iterator ReturnInst::op_end() { return OperandTraits< ReturnInst>::op_end(this); } ReturnInst::const_op_iterator ReturnInst::op_end() const { return OperandTraits<ReturnInst >::op_end(const_cast<ReturnInst*>(this)); } Value *ReturnInst ::getOperand(unsigned i_nocapture) const { (static_cast <bool > (i_nocapture < OperandTraits<ReturnInst>::operands (this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<ReturnInst>::operands(this) && \"getOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 3118, __extension__ __PRETTY_FUNCTION__ )); return cast_or_null<Value>( OperandTraits<ReturnInst >::op_begin(const_cast<ReturnInst*>(this))[i_nocapture ].get()); } void ReturnInst::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (static_cast <bool> (i_nocapture < OperandTraits<ReturnInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<ReturnInst>::operands(this) && \"setOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 3118, __extension__ __PRETTY_FUNCTION__ )); OperandTraits<ReturnInst>::op_begin(this)[i_nocapture ] = Val_nocapture; } unsigned ReturnInst::getNumOperands() const { return OperandTraits<ReturnInst>::operands(this); } template <int Idx_nocapture> Use &ReturnInst::Op() { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &ReturnInst::Op() const { return this->OpFrom<Idx_nocapture>(this); } |
3119 | |
3120 | //===----------------------------------------------------------------------===// |
3121 | // BranchInst Class |
3122 | //===----------------------------------------------------------------------===// |
3123 | |
3124 | //===--------------------------------------------------------------------------- |
3125 | /// Conditional or Unconditional Branch instruction. |
3126 | /// |
3127 | class BranchInst : public Instruction { |
3128 | /// Ops list - Branches are strange. The operands are ordered: |
3129 | /// [Cond, FalseDest,] TrueDest. This makes some accessors faster because |
3130 | /// they don't have to check for cond/uncond branchness. These are mostly |
3131 | /// accessed relative from op_end(). |
3132 | BranchInst(const BranchInst &BI); |
3133 | // BranchInst constructors (where {B, T, F} are blocks, and C is a condition): |
3134 | // BranchInst(BB *B) - 'br B' |
3135 | // BranchInst(BB* T, BB *F, Value *C) - 'br C, T, F' |
3136 | // BranchInst(BB* B, Inst *I) - 'br B' insert before I |
3137 | // BranchInst(BB* T, BB *F, Value *C, Inst *I) - 'br C, T, F', insert before I |
3138 | // BranchInst(BB* B, BB *I) - 'br B' insert at end |
3139 | // BranchInst(BB* T, BB *F, Value *C, BB *I) - 'br C, T, F', insert at end |
3140 | explicit BranchInst(BasicBlock *IfTrue, Instruction *InsertBefore = nullptr); |
3141 | BranchInst(BasicBlock *IfTrue, BasicBlock *IfFalse, Value *Cond, |
3142 | Instruction *InsertBefore = nullptr); |
3143 | BranchInst(BasicBlock *IfTrue, BasicBlock *InsertAtEnd); |
3144 | BranchInst(BasicBlock *IfTrue, BasicBlock *IfFalse, Value *Cond, |
3145 | BasicBlock *InsertAtEnd); |
3146 | |
3147 | void AssertOK(); |
3148 | |
3149 | protected: |
3150 | // Note: Instruction needs to be a friend here to call cloneImpl. |
3151 | friend class Instruction; |
3152 | |
3153 | BranchInst *cloneImpl() const; |
3154 | |
3155 | public: |
3156 | /// Iterator type that casts an operand to a basic block. |
3157 | /// |
3158 | /// This only makes sense because the successors are stored as adjacent |
3159 | /// operands for branch instructions. |
3160 | struct succ_op_iterator |
3161 | : iterator_adaptor_base<succ_op_iterator, value_op_iterator, |
3162 | std::random_access_iterator_tag, BasicBlock *, |
3163 | ptrdiff_t, BasicBlock *, BasicBlock *> { |
3164 | explicit succ_op_iterator(value_op_iterator I) : iterator_adaptor_base(I) {} |
3165 | |
3166 | BasicBlock *operator*() const { return cast<BasicBlock>(*I); } |
3167 | BasicBlock *operator->() const { return operator*(); } |
3168 | }; |
3169 | |
3170 | /// The const version of `succ_op_iterator`. |
3171 | struct const_succ_op_iterator |
3172 | : iterator_adaptor_base<const_succ_op_iterator, const_value_op_iterator, |
3173 | std::random_access_iterator_tag, |
3174 | const BasicBlock *, ptrdiff_t, const BasicBlock *, |
3175 | const BasicBlock *> { |
3176 | explicit const_succ_op_iterator(const_value_op_iterator I) |
3177 | : iterator_adaptor_base(I) {} |
3178 | |
3179 | const BasicBlock *operator*() const { return cast<BasicBlock>(*I); } |
3180 | const BasicBlock *operator->() const { return operator*(); } |
3181 | }; |
3182 | |
3183 | static BranchInst *Create(BasicBlock *IfTrue, |
3184 | Instruction *InsertBefore = nullptr) { |
3185 | return new(1) BranchInst(IfTrue, InsertBefore); |
3186 | } |
3187 | |
3188 | static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *IfFalse, |
3189 | Value *Cond, Instruction *InsertBefore = nullptr) { |
3190 | return new(3) BranchInst(IfTrue, IfFalse, Cond, InsertBefore); |
3191 | } |
3192 | |
3193 | static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *InsertAtEnd) { |
3194 | return new(1) BranchInst(IfTrue, InsertAtEnd); |
3195 | } |
3196 | |
3197 | static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *IfFalse, |
3198 | Value *Cond, BasicBlock *InsertAtEnd) { |
3199 | return new(3) BranchInst(IfTrue, IfFalse, Cond, InsertAtEnd); |
3200 | } |
3201 | |
3202 | /// Transparently provide more efficient getOperand methods. |
3203 | 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; |
3204 | |
3205 | bool isUnconditional() const { return getNumOperands() == 1; } |
3206 | bool isConditional() const { return getNumOperands() == 3; } |
3207 | |
3208 | Value *getCondition() const { |
3209 | assert(isConditional() && "Cannot get condition of an uncond branch!")(static_cast <bool> (isConditional() && "Cannot get condition of an uncond branch!" ) ? void (0) : __assert_fail ("isConditional() && \"Cannot get condition of an uncond branch!\"" , "llvm/include/llvm/IR/Instructions.h", 3209, __extension__ __PRETTY_FUNCTION__ )); |
3210 | return Op<-3>(); |
3211 | } |
3212 | |
3213 | void setCondition(Value *V) { |
3214 | assert(isConditional() && "Cannot set condition of unconditional branch!")(static_cast <bool> (isConditional() && "Cannot set condition of unconditional branch!" ) ? void (0) : __assert_fail ("isConditional() && \"Cannot set condition of unconditional branch!\"" , "llvm/include/llvm/IR/Instructions.h", 3214, __extension__ __PRETTY_FUNCTION__ )); |
3215 | Op<-3>() = V; |
3216 | } |
3217 | |
3218 | unsigned getNumSuccessors() const { return 1+isConditional(); } |
3219 | |
3220 | BasicBlock *getSuccessor(unsigned i) const { |
3221 | assert(i < getNumSuccessors() && "Successor # out of range for Branch!")(static_cast <bool> (i < getNumSuccessors() && "Successor # out of range for Branch!") ? void (0) : __assert_fail ("i < getNumSuccessors() && \"Successor # out of range for Branch!\"" , "llvm/include/llvm/IR/Instructions.h", 3221, __extension__ __PRETTY_FUNCTION__ )); |
3222 | return cast_or_null<BasicBlock>((&Op<-1>() - i)->get()); |
3223 | } |
3224 | |
3225 | void setSuccessor(unsigned idx, BasicBlock *NewSucc) { |
3226 | assert(idx < getNumSuccessors() && "Successor # out of range for Branch!")(static_cast <bool> (idx < getNumSuccessors() && "Successor # out of range for Branch!") ? void (0) : __assert_fail ("idx < getNumSuccessors() && \"Successor # out of range for Branch!\"" , "llvm/include/llvm/IR/Instructions.h", 3226, __extension__ __PRETTY_FUNCTION__ )); |
3227 | *(&Op<-1>() - idx) = NewSucc; |
3228 | } |
3229 | |
3230 | /// Swap the successors of this branch instruction. |
3231 | /// |
3232 | /// Swaps the successors of the branch instruction. This also swaps any |
3233 | /// branch weight metadata associated with the instruction so that it |
3234 | /// continues to map correctly to each operand. |
3235 | void swapSuccessors(); |
3236 | |
3237 | iterator_range<succ_op_iterator> successors() { |
3238 | return make_range( |
3239 | succ_op_iterator(std::next(value_op_begin(), isConditional() ? 1 : 0)), |
3240 | succ_op_iterator(value_op_end())); |
3241 | } |
3242 | |
3243 | iterator_range<const_succ_op_iterator> successors() const { |
3244 | return make_range(const_succ_op_iterator( |
3245 | std::next(value_op_begin(), isConditional() ? 1 : 0)), |
3246 | const_succ_op_iterator(value_op_end())); |
3247 | } |
3248 | |
3249 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
3250 | static bool classof(const Instruction *I) { |
3251 | return (I->getOpcode() == Instruction::Br); |
3252 | } |
3253 | static bool classof(const Value *V) { |
3254 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
3255 | } |
3256 | }; |
3257 | |
3258 | template <> |
3259 | struct OperandTraits<BranchInst> : public VariadicOperandTraits<BranchInst, 1> { |
3260 | }; |
3261 | |
3262 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(BranchInst, Value)BranchInst::op_iterator BranchInst::op_begin() { return OperandTraits <BranchInst>::op_begin(this); } BranchInst::const_op_iterator BranchInst::op_begin() const { return OperandTraits<BranchInst >::op_begin(const_cast<BranchInst*>(this)); } BranchInst ::op_iterator BranchInst::op_end() { return OperandTraits< BranchInst>::op_end(this); } BranchInst::const_op_iterator BranchInst::op_end() const { return OperandTraits<BranchInst >::op_end(const_cast<BranchInst*>(this)); } Value *BranchInst ::getOperand(unsigned i_nocapture) const { (static_cast <bool > (i_nocapture < OperandTraits<BranchInst>::operands (this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<BranchInst>::operands(this) && \"getOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 3262, __extension__ __PRETTY_FUNCTION__ )); return cast_or_null<Value>( OperandTraits<BranchInst >::op_begin(const_cast<BranchInst*>(this))[i_nocapture ].get()); } void BranchInst::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (static_cast <bool> (i_nocapture < OperandTraits<BranchInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<BranchInst>::operands(this) && \"setOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 3262, __extension__ __PRETTY_FUNCTION__ )); OperandTraits<BranchInst>::op_begin(this)[i_nocapture ] = Val_nocapture; } unsigned BranchInst::getNumOperands() const { return OperandTraits<BranchInst>::operands(this); } template <int Idx_nocapture> Use &BranchInst::Op() { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &BranchInst::Op() const { return this->OpFrom<Idx_nocapture>(this); } |
3263 | |
3264 | //===----------------------------------------------------------------------===// |
3265 | // SwitchInst Class |
3266 | //===----------------------------------------------------------------------===// |
3267 | |
3268 | //===--------------------------------------------------------------------------- |
3269 | /// Multiway switch |
3270 | /// |
3271 | class SwitchInst : public Instruction { |
3272 | unsigned ReservedSpace; |
3273 | |
3274 | // Operand[0] = Value to switch on |
3275 | // Operand[1] = Default basic block destination |
3276 | // Operand[2n ] = Value to match |
3277 | // Operand[2n+1] = BasicBlock to go to on match |
3278 | SwitchInst(const SwitchInst &SI); |
3279 | |
3280 | /// Create a new switch instruction, specifying a value to switch on and a |
3281 | /// default destination. The number of additional cases can be specified here |
3282 | /// to make memory allocation more efficient. This constructor can also |
3283 | /// auto-insert before another instruction. |
3284 | SwitchInst(Value *Value, BasicBlock *Default, unsigned NumCases, |
3285 | Instruction *InsertBefore); |
3286 | |
3287 | /// Create a new switch instruction, specifying a value to switch on and a |
3288 | /// default destination. The number of additional cases can be specified here |
3289 | /// to make memory allocation more efficient. This constructor also |
3290 | /// auto-inserts at the end of the specified BasicBlock. |
3291 | SwitchInst(Value *Value, BasicBlock *Default, unsigned NumCases, |
3292 | BasicBlock *InsertAtEnd); |
3293 | |
3294 | // allocate space for exactly zero operands |
3295 | void *operator new(size_t S) { return User::operator new(S); } |
3296 | |
3297 | void init(Value *Value, BasicBlock *Default, unsigned NumReserved); |
3298 | void growOperands(); |
3299 | |
3300 | protected: |
3301 | // Note: Instruction needs to be a friend here to call cloneImpl. |
3302 | friend class Instruction; |
3303 | |
3304 | SwitchInst *cloneImpl() const; |
3305 | |
3306 | public: |
3307 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
3308 | |
3309 | // -2 |
3310 | static const unsigned DefaultPseudoIndex = static_cast<unsigned>(~0L-1); |
3311 | |
3312 | template <typename CaseHandleT> class CaseIteratorImpl; |
3313 | |
3314 | /// A handle to a particular switch case. It exposes a convenient interface |
3315 | /// to both the case value and the successor block. |
3316 | /// |
3317 | /// We define this as a template and instantiate it to form both a const and |
3318 | /// non-const handle. |
3319 | template <typename SwitchInstT, typename ConstantIntT, typename BasicBlockT> |
3320 | class CaseHandleImpl { |
3321 | // Directly befriend both const and non-const iterators. |
3322 | friend class SwitchInst::CaseIteratorImpl< |
3323 | CaseHandleImpl<SwitchInstT, ConstantIntT, BasicBlockT>>; |
3324 | |
3325 | protected: |
3326 | // Expose the switch type we're parameterized with to the iterator. |
3327 | using SwitchInstType = SwitchInstT; |
3328 | |
3329 | SwitchInstT *SI; |
3330 | ptrdiff_t Index; |
3331 | |
3332 | CaseHandleImpl() = default; |
3333 | CaseHandleImpl(SwitchInstT *SI, ptrdiff_t Index) : SI(SI), Index(Index) {} |
3334 | |
3335 | public: |
3336 | /// Resolves case value for current case. |
3337 | ConstantIntT *getCaseValue() const { |
3338 | assert((unsigned)Index < SI->getNumCases() &&(static_cast <bool> ((unsigned)Index < SI->getNumCases () && "Index out the number of cases.") ? void (0) : __assert_fail ("(unsigned)Index < SI->getNumCases() && \"Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3339, __extension__ __PRETTY_FUNCTION__ )) |
3339 | "Index out the number of cases.")(static_cast <bool> ((unsigned)Index < SI->getNumCases () && "Index out the number of cases.") ? void (0) : __assert_fail ("(unsigned)Index < SI->getNumCases() && \"Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3339, __extension__ __PRETTY_FUNCTION__ )); |
3340 | return reinterpret_cast<ConstantIntT *>(SI->getOperand(2 + Index * 2)); |
3341 | } |
3342 | |
3343 | /// Resolves successor for current case. |
3344 | BasicBlockT *getCaseSuccessor() const { |
3345 | assert(((unsigned)Index < SI->getNumCases() ||(static_cast <bool> (((unsigned)Index < SI->getNumCases () || (unsigned)Index == DefaultPseudoIndex) && "Index out the number of cases." ) ? void (0) : __assert_fail ("((unsigned)Index < SI->getNumCases() || (unsigned)Index == DefaultPseudoIndex) && \"Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3347, __extension__ __PRETTY_FUNCTION__ )) |
3346 | (unsigned)Index == DefaultPseudoIndex) &&(static_cast <bool> (((unsigned)Index < SI->getNumCases () || (unsigned)Index == DefaultPseudoIndex) && "Index out the number of cases." ) ? void (0) : __assert_fail ("((unsigned)Index < SI->getNumCases() || (unsigned)Index == DefaultPseudoIndex) && \"Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3347, __extension__ __PRETTY_FUNCTION__ )) |
3347 | "Index out the number of cases.")(static_cast <bool> (((unsigned)Index < SI->getNumCases () || (unsigned)Index == DefaultPseudoIndex) && "Index out the number of cases." ) ? void (0) : __assert_fail ("((unsigned)Index < SI->getNumCases() || (unsigned)Index == DefaultPseudoIndex) && \"Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3347, __extension__ __PRETTY_FUNCTION__ )); |
3348 | return SI->getSuccessor(getSuccessorIndex()); |
3349 | } |
3350 | |
3351 | /// Returns number of current case. |
3352 | unsigned getCaseIndex() const { return Index; } |
3353 | |
3354 | /// Returns successor index for current case successor. |
3355 | unsigned getSuccessorIndex() const { |
3356 | assert(((unsigned)Index == DefaultPseudoIndex ||(static_cast <bool> (((unsigned)Index == DefaultPseudoIndex || (unsigned)Index < SI->getNumCases()) && "Index out the number of cases." ) ? void (0) : __assert_fail ("((unsigned)Index == DefaultPseudoIndex || (unsigned)Index < SI->getNumCases()) && \"Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3358, __extension__ __PRETTY_FUNCTION__ )) |
3357 | (unsigned)Index < SI->getNumCases()) &&(static_cast <bool> (((unsigned)Index == DefaultPseudoIndex || (unsigned)Index < SI->getNumCases()) && "Index out the number of cases." ) ? void (0) : __assert_fail ("((unsigned)Index == DefaultPseudoIndex || (unsigned)Index < SI->getNumCases()) && \"Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3358, __extension__ __PRETTY_FUNCTION__ )) |
3358 | "Index out the number of cases.")(static_cast <bool> (((unsigned)Index == DefaultPseudoIndex || (unsigned)Index < SI->getNumCases()) && "Index out the number of cases." ) ? void (0) : __assert_fail ("((unsigned)Index == DefaultPseudoIndex || (unsigned)Index < SI->getNumCases()) && \"Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3358, __extension__ __PRETTY_FUNCTION__ )); |
3359 | return (unsigned)Index != DefaultPseudoIndex ? Index + 1 : 0; |
3360 | } |
3361 | |
3362 | bool operator==(const CaseHandleImpl &RHS) const { |
3363 | assert(SI == RHS.SI && "Incompatible operators.")(static_cast <bool> (SI == RHS.SI && "Incompatible operators." ) ? void (0) : __assert_fail ("SI == RHS.SI && \"Incompatible operators.\"" , "llvm/include/llvm/IR/Instructions.h", 3363, __extension__ __PRETTY_FUNCTION__ )); |
3364 | return Index == RHS.Index; |
3365 | } |
3366 | }; |
3367 | |
3368 | using ConstCaseHandle = |
3369 | CaseHandleImpl<const SwitchInst, const ConstantInt, const BasicBlock>; |
3370 | |
3371 | class CaseHandle |
3372 | : public CaseHandleImpl<SwitchInst, ConstantInt, BasicBlock> { |
3373 | friend class SwitchInst::CaseIteratorImpl<CaseHandle>; |
3374 | |
3375 | public: |
3376 | CaseHandle(SwitchInst *SI, ptrdiff_t Index) : CaseHandleImpl(SI, Index) {} |
3377 | |
3378 | /// Sets the new value for current case. |
3379 | void setValue(ConstantInt *V) const { |
3380 | assert((unsigned)Index < SI->getNumCases() &&(static_cast <bool> ((unsigned)Index < SI->getNumCases () && "Index out the number of cases.") ? void (0) : __assert_fail ("(unsigned)Index < SI->getNumCases() && \"Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3381, __extension__ __PRETTY_FUNCTION__ )) |
3381 | "Index out the number of cases.")(static_cast <bool> ((unsigned)Index < SI->getNumCases () && "Index out the number of cases.") ? void (0) : __assert_fail ("(unsigned)Index < SI->getNumCases() && \"Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3381, __extension__ __PRETTY_FUNCTION__ )); |
3382 | SI->setOperand(2 + Index*2, reinterpret_cast<Value*>(V)); |
3383 | } |
3384 | |
3385 | /// Sets the new successor for current case. |
3386 | void setSuccessor(BasicBlock *S) const { |
3387 | SI->setSuccessor(getSuccessorIndex(), S); |
3388 | } |
3389 | }; |
3390 | |
3391 | template <typename CaseHandleT> |
3392 | class CaseIteratorImpl |
3393 | : public iterator_facade_base<CaseIteratorImpl<CaseHandleT>, |
3394 | std::random_access_iterator_tag, |
3395 | const CaseHandleT> { |
3396 | using SwitchInstT = typename CaseHandleT::SwitchInstType; |
3397 | |
3398 | CaseHandleT Case; |
3399 | |
3400 | public: |
3401 | /// Default constructed iterator is in an invalid state until assigned to |
3402 | /// a case for a particular switch. |
3403 | CaseIteratorImpl() = default; |
3404 | |
3405 | /// Initializes case iterator for given SwitchInst and for given |
3406 | /// case number. |
3407 | CaseIteratorImpl(SwitchInstT *SI, unsigned CaseNum) : Case(SI, CaseNum) {} |
3408 | |
3409 | /// Initializes case iterator for given SwitchInst and for given |
3410 | /// successor index. |
3411 | static CaseIteratorImpl fromSuccessorIndex(SwitchInstT *SI, |
3412 | unsigned SuccessorIndex) { |
3413 | assert(SuccessorIndex < SI->getNumSuccessors() &&(static_cast <bool> (SuccessorIndex < SI->getNumSuccessors () && "Successor index # out of range!") ? void (0) : __assert_fail ("SuccessorIndex < SI->getNumSuccessors() && \"Successor index # out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 3414, __extension__ __PRETTY_FUNCTION__ )) |
3414 | "Successor index # out of range!")(static_cast <bool> (SuccessorIndex < SI->getNumSuccessors () && "Successor index # out of range!") ? void (0) : __assert_fail ("SuccessorIndex < SI->getNumSuccessors() && \"Successor index # out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 3414, __extension__ __PRETTY_FUNCTION__ )); |
3415 | return SuccessorIndex != 0 ? CaseIteratorImpl(SI, SuccessorIndex - 1) |
3416 | : CaseIteratorImpl(SI, DefaultPseudoIndex); |
3417 | } |
3418 | |
3419 | /// Support converting to the const variant. This will be a no-op for const |
3420 | /// variant. |
3421 | operator CaseIteratorImpl<ConstCaseHandle>() const { |
3422 | return CaseIteratorImpl<ConstCaseHandle>(Case.SI, Case.Index); |
3423 | } |
3424 | |
3425 | CaseIteratorImpl &operator+=(ptrdiff_t N) { |
3426 | // Check index correctness after addition. |
3427 | // Note: Index == getNumCases() means end(). |
3428 | assert(Case.Index + N >= 0 &&(static_cast <bool> (Case.Index + N >= 0 && ( unsigned)(Case.Index + N) <= Case.SI->getNumCases() && "Case.Index out the number of cases.") ? void (0) : __assert_fail ("Case.Index + N >= 0 && (unsigned)(Case.Index + N) <= Case.SI->getNumCases() && \"Case.Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3430, __extension__ __PRETTY_FUNCTION__ )) |
3429 | (unsigned)(Case.Index + N) <= Case.SI->getNumCases() &&(static_cast <bool> (Case.Index + N >= 0 && ( unsigned)(Case.Index + N) <= Case.SI->getNumCases() && "Case.Index out the number of cases.") ? void (0) : __assert_fail ("Case.Index + N >= 0 && (unsigned)(Case.Index + N) <= Case.SI->getNumCases() && \"Case.Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3430, __extension__ __PRETTY_FUNCTION__ )) |
3430 | "Case.Index out the number of cases.")(static_cast <bool> (Case.Index + N >= 0 && ( unsigned)(Case.Index + N) <= Case.SI->getNumCases() && "Case.Index out the number of cases.") ? void (0) : __assert_fail ("Case.Index + N >= 0 && (unsigned)(Case.Index + N) <= Case.SI->getNumCases() && \"Case.Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3430, __extension__ __PRETTY_FUNCTION__ )); |
3431 | Case.Index += N; |
3432 | return *this; |
3433 | } |
3434 | CaseIteratorImpl &operator-=(ptrdiff_t N) { |
3435 | // Check index correctness after subtraction. |
3436 | // Note: Case.Index == getNumCases() means end(). |
3437 | assert(Case.Index - N >= 0 &&(static_cast <bool> (Case.Index - N >= 0 && ( unsigned)(Case.Index - N) <= Case.SI->getNumCases() && "Case.Index out the number of cases.") ? void (0) : __assert_fail ("Case.Index - N >= 0 && (unsigned)(Case.Index - N) <= Case.SI->getNumCases() && \"Case.Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3439, __extension__ __PRETTY_FUNCTION__ )) |
3438 | (unsigned)(Case.Index - N) <= Case.SI->getNumCases() &&(static_cast <bool> (Case.Index - N >= 0 && ( unsigned)(Case.Index - N) <= Case.SI->getNumCases() && "Case.Index out the number of cases.") ? void (0) : __assert_fail ("Case.Index - N >= 0 && (unsigned)(Case.Index - N) <= Case.SI->getNumCases() && \"Case.Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3439, __extension__ __PRETTY_FUNCTION__ )) |
3439 | "Case.Index out the number of cases.")(static_cast <bool> (Case.Index - N >= 0 && ( unsigned)(Case.Index - N) <= Case.SI->getNumCases() && "Case.Index out the number of cases.") ? void (0) : __assert_fail ("Case.Index - N >= 0 && (unsigned)(Case.Index - N) <= Case.SI->getNumCases() && \"Case.Index out the number of cases.\"" , "llvm/include/llvm/IR/Instructions.h", 3439, __extension__ __PRETTY_FUNCTION__ )); |
3440 | Case.Index -= N; |
3441 | return *this; |
3442 | } |
3443 | ptrdiff_t operator-(const CaseIteratorImpl &RHS) const { |
3444 | assert(Case.SI == RHS.Case.SI && "Incompatible operators.")(static_cast <bool> (Case.SI == RHS.Case.SI && "Incompatible operators." ) ? void (0) : __assert_fail ("Case.SI == RHS.Case.SI && \"Incompatible operators.\"" , "llvm/include/llvm/IR/Instructions.h", 3444, __extension__ __PRETTY_FUNCTION__ )); |
3445 | return Case.Index - RHS.Case.Index; |
3446 | } |
3447 | bool operator==(const CaseIteratorImpl &RHS) const { |
3448 | return Case == RHS.Case; |
3449 | } |
3450 | bool operator<(const CaseIteratorImpl &RHS) const { |
3451 | assert(Case.SI == RHS.Case.SI && "Incompatible operators.")(static_cast <bool> (Case.SI == RHS.Case.SI && "Incompatible operators." ) ? void (0) : __assert_fail ("Case.SI == RHS.Case.SI && \"Incompatible operators.\"" , "llvm/include/llvm/IR/Instructions.h", 3451, __extension__ __PRETTY_FUNCTION__ )); |
3452 | return Case.Index < RHS.Case.Index; |
3453 | } |
3454 | const CaseHandleT &operator*() const { return Case; } |
3455 | }; |
3456 | |
3457 | using CaseIt = CaseIteratorImpl<CaseHandle>; |
3458 | using ConstCaseIt = CaseIteratorImpl<ConstCaseHandle>; |
3459 | |
3460 | static SwitchInst *Create(Value *Value, BasicBlock *Default, |
3461 | unsigned NumCases, |
3462 | Instruction *InsertBefore = nullptr) { |
3463 | return new SwitchInst(Value, Default, NumCases, InsertBefore); |
3464 | } |
3465 | |
3466 | static SwitchInst *Create(Value *Value, BasicBlock *Default, |
3467 | unsigned NumCases, BasicBlock *InsertAtEnd) { |
3468 | return new SwitchInst(Value, Default, NumCases, InsertAtEnd); |
3469 | } |
3470 | |
3471 | /// Provide fast operand accessors |
3472 | 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; |
3473 | |
3474 | // Accessor Methods for Switch stmt |
3475 | Value *getCondition() const { return getOperand(0); } |
3476 | void setCondition(Value *V) { setOperand(0, V); } |
3477 | |
3478 | BasicBlock *getDefaultDest() const { |
3479 | return cast<BasicBlock>(getOperand(1)); |
3480 | } |
3481 | |
3482 | void setDefaultDest(BasicBlock *DefaultCase) { |
3483 | setOperand(1, reinterpret_cast<Value*>(DefaultCase)); |
3484 | } |
3485 | |
3486 | /// Return the number of 'cases' in this switch instruction, excluding the |
3487 | /// default case. |
3488 | unsigned getNumCases() const { |
3489 | return getNumOperands()/2 - 1; |
3490 | } |
3491 | |
3492 | /// Returns a read/write iterator that points to the first case in the |
3493 | /// SwitchInst. |
3494 | CaseIt case_begin() { |
3495 | return CaseIt(this, 0); |
3496 | } |
3497 | |
3498 | /// Returns a read-only iterator that points to the first case in the |
3499 | /// SwitchInst. |
3500 | ConstCaseIt case_begin() const { |
3501 | return ConstCaseIt(this, 0); |
3502 | } |
3503 | |
3504 | /// Returns a read/write iterator that points one past the last in the |
3505 | /// SwitchInst. |
3506 | CaseIt case_end() { |
3507 | return CaseIt(this, getNumCases()); |
3508 | } |
3509 | |
3510 | /// Returns a read-only iterator that points one past the last in the |
3511 | /// SwitchInst. |
3512 | ConstCaseIt case_end() const { |
3513 | return ConstCaseIt(this, getNumCases()); |
3514 | } |
3515 | |
3516 | /// Iteration adapter for range-for loops. |
3517 | iterator_range<CaseIt> cases() { |
3518 | return make_range(case_begin(), case_end()); |
3519 | } |
3520 | |
3521 | /// Constant iteration adapter for range-for loops. |
3522 | iterator_range<ConstCaseIt> cases() const { |
3523 | return make_range(case_begin(), case_end()); |
3524 | } |
3525 | |
3526 | /// Returns an iterator that points to the default case. |
3527 | /// Note: this iterator allows to resolve successor only. Attempt |
3528 | /// to resolve case value causes an assertion. |
3529 | /// Also note, that increment and decrement also causes an assertion and |
3530 | /// makes iterator invalid. |
3531 | CaseIt case_default() { |
3532 | return CaseIt(this, DefaultPseudoIndex); |
3533 | } |
3534 | ConstCaseIt case_default() const { |
3535 | return ConstCaseIt(this, DefaultPseudoIndex); |
3536 | } |
3537 | |
3538 | /// Search all of the case values for the specified constant. If it is |
3539 | /// explicitly handled, return the case iterator of it, otherwise return |
3540 | /// default case iterator to indicate that it is handled by the default |
3541 | /// handler. |
3542 | CaseIt findCaseValue(const ConstantInt *C) { |
3543 | return CaseIt( |
3544 | this, |
3545 | const_cast<const SwitchInst *>(this)->findCaseValue(C)->getCaseIndex()); |
3546 | } |
3547 | ConstCaseIt findCaseValue(const ConstantInt *C) const { |
3548 | ConstCaseIt I = llvm::find_if(cases(), [C](const ConstCaseHandle &Case) { |
3549 | return Case.getCaseValue() == C; |
3550 | }); |
3551 | if (I != case_end()) |
3552 | return I; |
3553 | |
3554 | return case_default(); |
3555 | } |
3556 | |
3557 | /// Finds the unique case value for a given successor. Returns null if the |
3558 | /// successor is not found, not unique, or is the default case. |
3559 | ConstantInt *findCaseDest(BasicBlock *BB) { |
3560 | if (BB == getDefaultDest()) |
3561 | return nullptr; |
3562 | |
3563 | ConstantInt *CI = nullptr; |
3564 | for (auto Case : cases()) { |
3565 | if (Case.getCaseSuccessor() != BB) |
3566 | continue; |
3567 | |
3568 | if (CI) |
3569 | return nullptr; // Multiple cases lead to BB. |
3570 | |
3571 | CI = Case.getCaseValue(); |
3572 | } |
3573 | |
3574 | return CI; |
3575 | } |
3576 | |
3577 | /// Add an entry to the switch instruction. |
3578 | /// Note: |
3579 | /// This action invalidates case_end(). Old case_end() iterator will |
3580 | /// point to the added case. |
3581 | void addCase(ConstantInt *OnVal, BasicBlock *Dest); |
3582 | |
3583 | /// This method removes the specified case and its successor from the switch |
3584 | /// instruction. Note that this operation may reorder the remaining cases at |
3585 | /// index idx and above. |
3586 | /// Note: |
3587 | /// This action invalidates iterators for all cases following the one removed, |
3588 | /// including the case_end() iterator. It returns an iterator for the next |
3589 | /// case. |
3590 | CaseIt removeCase(CaseIt I); |
3591 | |
3592 | unsigned getNumSuccessors() const { return getNumOperands()/2; } |
3593 | BasicBlock *getSuccessor(unsigned idx) const { |
3594 | assert(idx < getNumSuccessors() &&"Successor idx out of range for switch!")(static_cast <bool> (idx < getNumSuccessors() && "Successor idx out of range for switch!") ? void (0) : __assert_fail ("idx < getNumSuccessors() &&\"Successor idx out of range for switch!\"" , "llvm/include/llvm/IR/Instructions.h", 3594, __extension__ __PRETTY_FUNCTION__ )); |
3595 | return cast<BasicBlock>(getOperand(idx*2+1)); |
3596 | } |
3597 | void setSuccessor(unsigned idx, BasicBlock *NewSucc) { |
3598 | assert(idx < getNumSuccessors() && "Successor # out of range for switch!")(static_cast <bool> (idx < getNumSuccessors() && "Successor # out of range for switch!") ? void (0) : __assert_fail ("idx < getNumSuccessors() && \"Successor # out of range for switch!\"" , "llvm/include/llvm/IR/Instructions.h", 3598, __extension__ __PRETTY_FUNCTION__ )); |
3599 | setOperand(idx * 2 + 1, NewSucc); |
3600 | } |
3601 | |
3602 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
3603 | static bool classof(const Instruction *I) { |
3604 | return I->getOpcode() == Instruction::Switch; |
3605 | } |
3606 | static bool classof(const Value *V) { |
3607 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
3608 | } |
3609 | }; |
3610 | |
3611 | /// A wrapper class to simplify modification of SwitchInst cases along with |
3612 | /// their prof branch_weights metadata. |
3613 | class SwitchInstProfUpdateWrapper { |
3614 | SwitchInst &SI; |
3615 | std::optional<SmallVector<uint32_t, 8>> Weights; |
3616 | bool Changed = false; |
3617 | |
3618 | protected: |
3619 | static MDNode *getProfBranchWeightsMD(const SwitchInst &SI); |
3620 | |
3621 | MDNode *buildProfBranchWeightsMD(); |
3622 | |
3623 | void init(); |
3624 | |
3625 | public: |
3626 | using CaseWeightOpt = std::optional<uint32_t>; |
3627 | SwitchInst *operator->() { return &SI; } |
3628 | SwitchInst &operator*() { return SI; } |
3629 | operator SwitchInst *() { return &SI; } |
3630 | |
3631 | SwitchInstProfUpdateWrapper(SwitchInst &SI) : SI(SI) { init(); } |
3632 | |
3633 | ~SwitchInstProfUpdateWrapper() { |
3634 | if (Changed) |
3635 | SI.setMetadata(LLVMContext::MD_prof, buildProfBranchWeightsMD()); |
3636 | } |
3637 | |
3638 | /// Delegate the call to the underlying SwitchInst::removeCase() and remove |
3639 | /// correspondent branch weight. |
3640 | SwitchInst::CaseIt removeCase(SwitchInst::CaseIt I); |
3641 | |
3642 | /// Delegate the call to the underlying SwitchInst::addCase() and set the |
3643 | /// specified branch weight for the added case. |
3644 | void addCase(ConstantInt *OnVal, BasicBlock *Dest, CaseWeightOpt W); |
3645 | |
3646 | /// Delegate the call to the underlying SwitchInst::eraseFromParent() and mark |
3647 | /// this object to not touch the underlying SwitchInst in destructor. |
3648 | SymbolTableList<Instruction>::iterator eraseFromParent(); |
3649 | |
3650 | void setSuccessorWeight(unsigned idx, CaseWeightOpt W); |
3651 | CaseWeightOpt getSuccessorWeight(unsigned idx); |
3652 | |
3653 | static CaseWeightOpt getSuccessorWeight(const SwitchInst &SI, unsigned idx); |
3654 | }; |
3655 | |
3656 | template <> |
3657 | struct OperandTraits<SwitchInst> : public HungoffOperandTraits<2> { |
3658 | }; |
3659 | |
3660 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(SwitchInst, Value)SwitchInst::op_iterator SwitchInst::op_begin() { return OperandTraits <SwitchInst>::op_begin(this); } SwitchInst::const_op_iterator SwitchInst::op_begin() const { return OperandTraits<SwitchInst >::op_begin(const_cast<SwitchInst*>(this)); } SwitchInst ::op_iterator SwitchInst::op_end() { return OperandTraits< SwitchInst>::op_end(this); } SwitchInst::const_op_iterator SwitchInst::op_end() const { return OperandTraits<SwitchInst >::op_end(const_cast<SwitchInst*>(this)); } Value *SwitchInst ::getOperand(unsigned i_nocapture) const { (static_cast <bool > (i_nocapture < OperandTraits<SwitchInst>::operands (this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<SwitchInst>::operands(this) && \"getOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 3660, __extension__ __PRETTY_FUNCTION__ )); return cast_or_null<Value>( OperandTraits<SwitchInst >::op_begin(const_cast<SwitchInst*>(this))[i_nocapture ].get()); } void SwitchInst::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (static_cast <bool> (i_nocapture < OperandTraits<SwitchInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<SwitchInst>::operands(this) && \"setOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 3660, __extension__ __PRETTY_FUNCTION__ )); OperandTraits<SwitchInst>::op_begin(this)[i_nocapture ] = Val_nocapture; } unsigned SwitchInst::getNumOperands() const { return OperandTraits<SwitchInst>::operands(this); } template <int Idx_nocapture> Use &SwitchInst::Op() { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &SwitchInst::Op() const { return this->OpFrom<Idx_nocapture>(this); } |
3661 | |
3662 | //===----------------------------------------------------------------------===// |
3663 | // IndirectBrInst Class |
3664 | //===----------------------------------------------------------------------===// |
3665 | |
3666 | //===--------------------------------------------------------------------------- |
3667 | /// Indirect Branch Instruction. |
3668 | /// |
3669 | class IndirectBrInst : public Instruction { |
3670 | unsigned ReservedSpace; |
3671 | |
3672 | // Operand[0] = Address to jump to |
3673 | // Operand[n+1] = n-th destination |
3674 | IndirectBrInst(const IndirectBrInst &IBI); |
3675 | |
3676 | /// Create a new indirectbr instruction, specifying an |
3677 | /// Address to jump to. The number of expected destinations can be specified |
3678 | /// here to make memory allocation more efficient. This constructor can also |
3679 | /// autoinsert before another instruction. |
3680 | IndirectBrInst(Value *Address, unsigned NumDests, Instruction *InsertBefore); |
3681 | |
3682 | /// Create a new indirectbr instruction, specifying an |
3683 | /// Address to jump to. The number of expected destinations can be specified |
3684 | /// here to make memory allocation more efficient. This constructor also |
3685 | /// autoinserts at the end of the specified BasicBlock. |
3686 | IndirectBrInst(Value *Address, unsigned NumDests, BasicBlock *InsertAtEnd); |
3687 | |
3688 | // allocate space for exactly zero operands |
3689 | void *operator new(size_t S) { return User::operator new(S); } |
3690 | |
3691 | void init(Value *Address, unsigned NumDests); |
3692 | void growOperands(); |
3693 | |
3694 | protected: |
3695 | // Note: Instruction needs to be a friend here to call cloneImpl. |
3696 | friend class Instruction; |
3697 | |
3698 | IndirectBrInst *cloneImpl() const; |
3699 | |
3700 | public: |
3701 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
3702 | |
3703 | /// Iterator type that casts an operand to a basic block. |
3704 | /// |
3705 | /// This only makes sense because the successors are stored as adjacent |
3706 | /// operands for indirectbr instructions. |
3707 | struct succ_op_iterator |
3708 | : iterator_adaptor_base<succ_op_iterator, value_op_iterator, |
3709 | std::random_access_iterator_tag, BasicBlock *, |
3710 | ptrdiff_t, BasicBlock *, BasicBlock *> { |
3711 | explicit succ_op_iterator(value_op_iterator I) : iterator_adaptor_base(I) {} |
3712 | |
3713 | BasicBlock *operator*() const { return cast<BasicBlock>(*I); } |
3714 | BasicBlock *operator->() const { return operator*(); } |
3715 | }; |
3716 | |
3717 | /// The const version of `succ_op_iterator`. |
3718 | struct const_succ_op_iterator |
3719 | : iterator_adaptor_base<const_succ_op_iterator, const_value_op_iterator, |
3720 | std::random_access_iterator_tag, |
3721 | const BasicBlock *, ptrdiff_t, const BasicBlock *, |
3722 | const BasicBlock *> { |
3723 | explicit const_succ_op_iterator(const_value_op_iterator I) |
3724 | : iterator_adaptor_base(I) {} |
3725 | |
3726 | const BasicBlock *operator*() const { return cast<BasicBlock>(*I); } |
3727 | const BasicBlock *operator->() const { return operator*(); } |
3728 | }; |
3729 | |
3730 | static IndirectBrInst *Create(Value *Address, unsigned NumDests, |
3731 | Instruction *InsertBefore = nullptr) { |
3732 | return new IndirectBrInst(Address, NumDests, InsertBefore); |
3733 | } |
3734 | |
3735 | static IndirectBrInst *Create(Value *Address, unsigned NumDests, |
3736 | BasicBlock *InsertAtEnd) { |
3737 | return new IndirectBrInst(Address, NumDests, InsertAtEnd); |
3738 | } |
3739 | |
3740 | /// Provide fast operand accessors. |
3741 | 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; |
3742 | |
3743 | // Accessor Methods for IndirectBrInst instruction. |
3744 | Value *getAddress() { return getOperand(0); } |
3745 | const Value *getAddress() const { return getOperand(0); } |
3746 | void setAddress(Value *V) { setOperand(0, V); } |
3747 | |
3748 | /// return the number of possible destinations in this |
3749 | /// indirectbr instruction. |
3750 | unsigned getNumDestinations() const { return getNumOperands()-1; } |
3751 | |
3752 | /// Return the specified destination. |
3753 | BasicBlock *getDestination(unsigned i) { return getSuccessor(i); } |
3754 | const BasicBlock *getDestination(unsigned i) const { return getSuccessor(i); } |
3755 | |
3756 | /// Add a destination. |
3757 | /// |
3758 | void addDestination(BasicBlock *Dest); |
3759 | |
3760 | /// This method removes the specified successor from the |
3761 | /// indirectbr instruction. |
3762 | void removeDestination(unsigned i); |
3763 | |
3764 | unsigned getNumSuccessors() const { return getNumOperands()-1; } |
3765 | BasicBlock *getSuccessor(unsigned i) const { |
3766 | return cast<BasicBlock>(getOperand(i+1)); |
3767 | } |
3768 | void setSuccessor(unsigned i, BasicBlock *NewSucc) { |
3769 | setOperand(i + 1, NewSucc); |
3770 | } |
3771 | |
3772 | iterator_range<succ_op_iterator> successors() { |
3773 | return make_range(succ_op_iterator(std::next(value_op_begin())), |
3774 | succ_op_iterator(value_op_end())); |
3775 | } |
3776 | |
3777 | iterator_range<const_succ_op_iterator> successors() const { |
3778 | return make_range(const_succ_op_iterator(std::next(value_op_begin())), |
3779 | const_succ_op_iterator(value_op_end())); |
3780 | } |
3781 | |
3782 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
3783 | static bool classof(const Instruction *I) { |
3784 | return I->getOpcode() == Instruction::IndirectBr; |
3785 | } |
3786 | static bool classof(const Value *V) { |
3787 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
3788 | } |
3789 | }; |
3790 | |
3791 | template <> |
3792 | struct OperandTraits<IndirectBrInst> : public HungoffOperandTraits<1> { |
3793 | }; |
3794 | |
3795 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(IndirectBrInst, Value)IndirectBrInst::op_iterator IndirectBrInst::op_begin() { return OperandTraits<IndirectBrInst>::op_begin(this); } IndirectBrInst ::const_op_iterator IndirectBrInst::op_begin() const { return OperandTraits<IndirectBrInst>::op_begin(const_cast< IndirectBrInst*>(this)); } IndirectBrInst::op_iterator IndirectBrInst ::op_end() { return OperandTraits<IndirectBrInst>::op_end (this); } IndirectBrInst::const_op_iterator IndirectBrInst::op_end () const { return OperandTraits<IndirectBrInst>::op_end (const_cast<IndirectBrInst*>(this)); } Value *IndirectBrInst ::getOperand(unsigned i_nocapture) const { (static_cast <bool > (i_nocapture < OperandTraits<IndirectBrInst>::operands (this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<IndirectBrInst>::operands(this) && \"getOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 3795, __extension__ __PRETTY_FUNCTION__ )); return cast_or_null<Value>( OperandTraits<IndirectBrInst >::op_begin(const_cast<IndirectBrInst*>(this))[i_nocapture ].get()); } void IndirectBrInst::setOperand(unsigned i_nocapture , Value *Val_nocapture) { (static_cast <bool> (i_nocapture < OperandTraits<IndirectBrInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<IndirectBrInst>::operands(this) && \"setOperand() out of range!\"" , "llvm/include/llvm/IR/Instructions.h", 3795, __extension__ __PRETTY_FUNCTION__ )); OperandTraits<IndirectBrInst>::op_begin(this)[i_nocapture ] = Val_nocapture; } unsigned IndirectBrInst::getNumOperands( ) const { return OperandTraits<IndirectBrInst>::operands (this); } template <int Idx_nocapture> Use &IndirectBrInst ::Op() { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &IndirectBrInst ::Op() const { return this->OpFrom<Idx_nocapture>(this ); } |
3796 | |
3797 | //===----------------------------------------------------------------------===// |
3798 | // InvokeInst Class |
3799 | //===----------------------------------------------------------------------===// |
3800 | |
3801 | /// Invoke instruction. The SubclassData field is used to hold the |
3802 | /// calling convention of the call. |
3803 | /// |
3804 | class InvokeInst : public CallBase { |
3805 | /// The number of operands for this call beyond the called function, |
3806 | /// arguments, and operand bundles. |
3807 | static constexpr int NumExtraOperands = 2; |
3808 | |
3809 | /// The index from the end of the operand array to the normal destination. |
3810 | static constexpr int NormalDestOpEndIdx = -3; |
3811 | |
3812 | /// The index from the end of the operand array to the unwind destination. |
3813 | static constexpr int UnwindDestOpEndIdx = -2; |
3814 | |
3815 | InvokeInst(const InvokeInst &BI); |
3816 | |
3817 | /// Construct an InvokeInst given a range of arguments. |
3818 | /// |
3819 | /// Construct an InvokeInst from a range of arguments |
3820 | inline InvokeInst(FunctionType *Ty, Value *Func, BasicBlock *IfNormal, |
3821 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3822 | ArrayRef<OperandBundleDef> Bundles, int NumOperands, |
3823 | const Twine &NameStr, Instruction *InsertBefore); |
3824 | |
3825 | inline InvokeInst(FunctionType *Ty, Value *Func, BasicBlock *IfNormal, |
3826 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3827 | ArrayRef<OperandBundleDef> Bundles, int NumOperands, |
3828 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
3829 | |
3830 | void init(FunctionType *Ty, Value *Func, BasicBlock *IfNormal, |
3831 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3832 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr); |
3833 | |
3834 | /// Compute the number of operands to allocate. |
3835 | static int ComputeNumOperands(int NumArgs, int NumBundleInputs = 0) { |
3836 | // We need one operand for the called function, plus our extra operands and |
3837 | // the input operand counts provided. |
3838 | return 1 + NumExtraOperands + NumArgs + NumBundleInputs; |
3839 | } |
3840 | |
3841 | protected: |
3842 | // Note: Instruction needs to be a friend here to call cloneImpl. |
3843 | friend class Instruction; |
3844 | |
3845 | InvokeInst *cloneImpl() const; |
3846 | |
3847 | public: |
3848 | static InvokeInst *Create(FunctionType *Ty, Value *Func, BasicBlock *IfNormal, |
3849 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3850 | const Twine &NameStr, |
3851 | Instruction *InsertBefore = nullptr) { |
3852 | int NumOperands = ComputeNumOperands(Args.size()); |
3853 | return new (NumOperands) |
3854 | InvokeInst(Ty, Func, IfNormal, IfException, Args, std::nullopt, |
3855 | NumOperands, NameStr, InsertBefore); |
3856 | } |
3857 | |
3858 | static InvokeInst *Create(FunctionType *Ty, Value *Func, BasicBlock *IfNormal, |
3859 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3860 | ArrayRef<OperandBundleDef> Bundles = std::nullopt, |
3861 | const Twine &NameStr = "", |
3862 | Instruction *InsertBefore = nullptr) { |
3863 | int NumOperands = |
3864 | ComputeNumOperands(Args.size(), CountBundleInputs(Bundles)); |
3865 | unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
3866 | |
3867 | return new (NumOperands, DescriptorBytes) |
3868 | InvokeInst(Ty, Func, IfNormal, IfException, Args, Bundles, NumOperands, |
3869 | NameStr, InsertBefore); |
3870 | } |
3871 | |
3872 | static InvokeInst *Create(FunctionType *Ty, Value *Func, BasicBlock *IfNormal, |
3873 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3874 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
3875 | int NumOperands = ComputeNumOperands(Args.size()); |
3876 | return new (NumOperands) |
3877 | InvokeInst(Ty, Func, IfNormal, IfException, Args, std::nullopt, |
3878 | NumOperands, NameStr, InsertAtEnd); |
3879 | } |
3880 | |
3881 | static InvokeInst *Create(FunctionType *Ty, Value *Func, BasicBlock *IfNormal, |
3882 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3883 | ArrayRef<OperandBundleDef> Bundles, |
3884 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
3885 | int NumOperands = |
3886 | ComputeNumOperands(Args.size(), CountBundleInputs(Bundles)); |
3887 | unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
3888 | |
3889 | return new (NumOperands, DescriptorBytes) |
3890 | InvokeInst(Ty, Func, IfNormal, IfException, Args, Bundles, NumOperands, |
3891 | NameStr, InsertAtEnd); |
3892 | } |
3893 | |
3894 | static InvokeInst *Create(FunctionCallee Func, BasicBlock *IfNormal, |
3895 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3896 | const Twine &NameStr, |
3897 | Instruction *InsertBefore = nullptr) { |
3898 | return Create(Func.getFunctionType(), Func.getCallee(), IfNormal, |
3899 | IfException, Args, std::nullopt, NameStr, InsertBefore); |
3900 | } |
3901 | |
3902 | static InvokeInst *Create(FunctionCallee Func, BasicBlock *IfNormal, |
3903 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3904 | ArrayRef<OperandBundleDef> Bundles = std::nullopt, |
3905 | const Twine &NameStr = "", |
3906 | Instruction *InsertBefore = nullptr) { |
3907 | return Create(Func.getFunctionType(), Func.getCallee(), IfNormal, |
3908 | IfException, Args, Bundles, NameStr, InsertBefore); |
3909 | } |
3910 | |
3911 | static InvokeInst *Create(FunctionCallee Func, BasicBlock *IfNormal, |
3912 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3913 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
3914 | return Create(Func.getFunctionType(), Func.getCallee(), IfNormal, |
3915 | IfException, Args, NameStr, InsertAtEnd); |
3916 | } |
3917 | |
3918 | static InvokeInst *Create(FunctionCallee Func, BasicBlock *IfNormal, |
3919 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3920 | ArrayRef<OperandBundleDef> Bundles, |
3921 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
3922 | return Create(Func.getFunctionType(), Func.getCallee(), IfNormal, |
3923 | IfException, Args, Bundles, NameStr, InsertAtEnd); |
3924 | } |
3925 | |
3926 | /// Create a clone of \p II with a different set of operand bundles and |
3927 | /// insert it before \p InsertPt. |
3928 | /// |
3929 | /// The returned invoke instruction is identical to \p II in every way except |
3930 | /// that the operand bundles for the new instruction are set to the operand |
3931 | /// bundles in \p Bundles. |
3932 | static InvokeInst *Create(InvokeInst *II, ArrayRef<OperandBundleDef> Bundles, |
3933 | Instruction *InsertPt = nullptr); |
3934 | |
3935 | // get*Dest - Return the destination basic blocks... |
3936 | BasicBlock *getNormalDest() const { |
3937 | return cast<BasicBlock>(Op<NormalDestOpEndIdx>()); |
3938 | } |
3939 | BasicBlock *getUnwindDest() const { |
3940 | return cast<BasicBlock>(Op<UnwindDestOpEndIdx>()); |
3941 | } |
3942 | void setNormalDest(BasicBlock *B) { |
3943 | Op<NormalDestOpEndIdx>() = reinterpret_cast<Value *>(B); |
3944 | } |
3945 | void setUnwindDest(BasicBlock *B) { |
3946 | Op<UnwindDestOpEndIdx>() = reinterpret_cast<Value *>(B); |
3947 | } |
3948 | |
3949 | /// Get the landingpad instruction from the landing pad |
3950 | /// block (the unwind destination). |
3951 | LandingPadInst *getLandingPadInst() const; |
3952 | |
3953 | BasicBlock *getSuccessor(unsigned i) const { |
3954 | assert(i < 2 && "Successor # out of range for invoke!")(static_cast <bool> (i < 2 && "Successor # out of range for invoke!" ) ? void (0) : __assert_fail ("i < 2 && \"Successor # out of range for invoke!\"" , "llvm/include/llvm/IR/Instructions.h", 3954, __extension__ __PRETTY_FUNCTION__ )); |
3955 | return i == 0 ? getNormalDest() : getUnwindDest(); |
3956 | } |
3957 | |
3958 | void setSuccessor(unsigned i, BasicBlock *NewSucc) { |
3959 | assert(i < 2 && "Successor # out of range for invoke!")(static_cast <bool> (i < 2 && "Successor # out of range for invoke!" ) ? void (0) : __assert_fail ("i < 2 && \"Successor # out of range for invoke!\"" , "llvm/include/llvm/IR/Instructions.h", 3959, __extension__ __PRETTY_FUNCTION__ )); |
3960 | if (i == 0) |
3961 | setNormalDest(NewSucc); |
3962 | else |
3963 | setUnwindDest(NewSucc); |
3964 | } |
3965 | |
3966 | unsigned getNumSuccessors() const { return 2; } |
3967 | |
3968 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
3969 | static bool classof(const Instruction *I) { |
3970 | return (I->getOpcode() == Instruction::Invoke); |
3971 | } |
3972 | static bool classof(const Value *V) { |
3973 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
3974 | } |
3975 | |
3976 | private: |
3977 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
3978 | // method so that subclasses cannot accidentally use it. |
3979 | template <typename Bitfield> |
3980 | void setSubclassData(typename Bitfield::Type Value) { |
3981 | Instruction::setSubclassData<Bitfield>(Value); |
3982 | } |
3983 | }; |
3984 | |
3985 | InvokeInst::InvokeInst(FunctionType *Ty, Value *Func, BasicBlock *IfNormal, |
3986 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3987 | ArrayRef<OperandBundleDef> Bundles, int NumOperands, |
3988 | const Twine &NameStr, Instruction *InsertBefore) |
3989 | : CallBase(Ty->getReturnType(), Instruction::Invoke, |
3990 | OperandTraits<CallBase>::op_end(this) - NumOperands, NumOperands, |
3991 | InsertBefore) { |
3992 | init(Ty, Func, IfNormal, IfException, Args, Bundles, NameStr); |
3993 | } |
3994 | |
3995 | InvokeInst::InvokeInst(FunctionType *Ty, Value *Func, BasicBlock *IfNormal, |
3996 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3997 | ArrayRef<OperandBundleDef> Bundles, int NumOperands, |
3998 | const Twine &NameStr, BasicBlock *InsertAtEnd) |
3999 | : CallBase(Ty->getReturnType(), Instruction::Invoke, |
4000 | OperandTraits<CallBase>::op_end(this) - NumOperands, NumOperands, |
4001 | InsertAtEnd) { |
4002 | init(Ty, Func, IfNormal, IfException, Args, Bundles, NameStr); |
4003 | } |
4004 | |
4005 | //===----------------------------------------------------------------------===// |
4006 | // CallBrInst Class |
4007 | //===----------------------------------------------------------------------===// |
4008 | |
4009 | /// CallBr instruction, tracking function calls that may not return control but |
4010 | /// instead transfer it to a third location. The SubclassData field is used to |
4011 | /// hold the calling convention of the call. |
4012 | /// |
4013 | class CallBrInst : public CallBase { |
4014 | |
4015 | unsigned NumIndirectDests; |
4016 | |
4017 | CallBrInst(const CallBrInst &BI); |
4018 | |
4019 | /// Construct a CallBrInst given a range of arguments. |
4020 | /// |
4021 | /// Construct a CallBrInst from a range of arguments |
4022 | inline CallBrInst(FunctionType *Ty, Value *Func, BasicBlock *DefaultDest, |
4023 | ArrayRef<BasicBlock *> IndirectDests, |
4024 | ArrayRef<Value *> Args, |
4025 | ArrayRef<OperandBundleDef> Bundles, int NumOperands, |
4026 | const Twine &NameStr, Instruction *InsertBefore); |
4027 | |
4028 | inline CallBrInst(FunctionType *Ty, Value *Func, BasicBlock *DefaultDest, |
4029 | ArrayRef<BasicBlock *> IndirectDests, |
4030 | ArrayRef<Value *> Args, |
4031 | ArrayRef<OperandBundleDef> Bundles, int NumOperands, |
4032 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
4033 | |
4034 | void init(FunctionType *FTy, Value *Func, BasicBlock *DefaultDest, |
4035 | ArrayRef<BasicBlock *> IndirectDests, ArrayRef<Value *> Args, |
4036 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr); |
4037 | |
4038 | /// Compute the number of operands to allocate. |
4039 | static int ComputeNumOperands(int NumArgs, int NumIndirectDests, |
4040 | int NumBundleInputs = 0) { |
4041 | // We need one operand for the called function, plus our extra operands and |
4042 | // the input operand counts provided. |
4043 | return 2 + NumIndirectDests + NumArgs + NumBundleInputs; |
4044 | } |
4045 | |
4046 | protected: |
4047 | // Note: Instruction needs to be a friend here to call cloneImpl. |
4048 | friend class Instruction; |
4049 | |
4050 | CallBrInst *cloneImpl() const; |
4051 | |
4052 | public: |
4053 | static CallBrInst *Create(FunctionType *Ty, Value *Func, |
4054 | BasicBlock *DefaultDest, |
4055 | ArrayRef<BasicBlock *> IndirectDests, |
4056 | ArrayRef<Value *> Args, const Twine &NameStr, |
4057 | Instruction *InsertBefore = nullptr) { |
4058 | int NumOperands = ComputeNumOperands(Args.size(), IndirectDests.size()); |
4059 | return new (NumOperands) |
4060 | CallBrInst(Ty, Func, DefaultDest, IndirectDests, Args, std::nullopt, |
4061 | NumOperands, NameStr, InsertBefore); |
4062 | } |
4063 | |
4064 | static CallBrInst * |
4065 | Create(FunctionType *Ty, Value *Func, BasicBlock *DefaultDest, |
4066 | ArrayRef<BasicBlock *> IndirectDests, ArrayRef<Value *> Args, |
4067 | ArrayRef<OperandBundleDef> Bundles = std::nullopt, |
4068 | const Twine &NameStr = "", Instruction *InsertBefore = nullptr) { |
4069 | int NumOperands = ComputeNumOperands(Args.size(), IndirectDests.size(), |
4070 | CountBundleInputs(Bundles)); |
4071 | unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
4072 | |
4073 | return new (NumOperands, DescriptorBytes) |
4074 | CallBrInst(Ty, Func, DefaultDest, IndirectDests, Args, Bundles, |
4075 | NumOperands, NameStr, InsertBefore); |
4076 | } |
4077 | |
4078 | static CallBrInst *Create(FunctionType *Ty, Value *Func, |
4079 | BasicBlock *DefaultDest, |
4080 | ArrayRef<BasicBlock *> IndirectDests, |
4081 | ArrayRef<Value *> Args, const Twine &NameStr, |
4082 | BasicBlock *InsertAtEnd) { |
4083 | int NumOperands = ComputeNumOperands(Args.size(), IndirectDests.size()); |
4084 | return new (NumOperands) |
4085 | CallBrInst(Ty, Func, DefaultDest, IndirectDests, Args, std::nullopt, |
4086 | NumOperands, NameStr, InsertAtEnd); |
4087 | } |
4088 | |
4089 | static CallBrInst *Create(FunctionType *Ty, Value *Func, |
4090 | BasicBlock *DefaultDest, |
4091 | ArrayRef<BasicBlock *> IndirectDests, |
4092 | ArrayRef<Value *> Args, |
4093 | ArrayRef<OperandBundleDef> Bundles, |
4094 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
4095 | int NumOperands = ComputeNumOperands(Args.size(), IndirectDests.size(), |
4096 | CountBundleInputs(Bundles)); |
4097 | unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
4098 | |
4099 | return new (NumOperands, DescriptorBytes) |
4100 | CallBrInst(Ty, Func, DefaultDest, IndirectDests, Args, Bundles, |
4101 | NumOperands, NameStr, InsertAtEnd); |
4102 | } |
4103 | |
4104 | static CallBrInst *Create(FunctionCallee Func, BasicBlock *DefaultDest, |
4105 | ArrayRef<BasicBlock *> IndirectDests, |
4106 | ArrayRef<Value *> Args, const Twine &NameStr, |
4107 | Instruction *InsertBefore = nullptr) { |
4108 | return Create(Func.getFunctionType(), Func.getCallee(), DefaultDest, |
4109 | IndirectDests, Args, NameStr, InsertBefore); |
4110 | } |
4111 | |
4112 | static CallBrInst *Create(FunctionCallee Func, BasicBlock *DefaultDest, |
4113 | ArrayRef<BasicBlock *> IndirectDests, |
4114 | ArrayRef<Value *> Args, |
4115 | ArrayRef<OperandBundleDef> Bundles = std::nullopt, |
4116 | const Twine &NameStr = "", |
4117 | Instruction *InsertBefore = nullptr) { |
4118 | return Create(Func.getFunctionType(), Func.getCallee(), DefaultDest, |
4119 | IndirectDests, Args, Bundles, NameStr, InsertBefore); |
4120 | } |
4121 | |
4122 | static CallBrInst *Create(FunctionCallee Func, BasicBlock *DefaultDest, |
4123 | ArrayRef<BasicBlock *> IndirectDests, |
4124 | ArrayRef<Value *> Args, const Twine &NameStr, |
4125 | BasicBlock *InsertAtEnd) { |
4126 | return Create(Func.getFunctionType(), Func.getCallee(), DefaultDest, |
4127 | IndirectDests, Args, NameStr, InsertAtEnd); |
4128 | } |
4129 | |
4130 | static CallBrInst *Create(FunctionCallee Func, |
4131 | BasicBlock *DefaultDest, |
4132 | ArrayRef<BasicBlock *> IndirectDests, |
4133 | ArrayRef<Value *> Args, |
4134 | ArrayRef<OperandBundleDef> Bundles, |
4135 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
4136 | return Create(Func.getFunctionType(), Func.getCallee(), DefaultDest, |
4137 | IndirectDests, Args, Bundles, NameStr, InsertAtEnd); |
4138 | } |
4139 | |
4140 | /// Create a clone of \p CBI with a different set of operand bundles and |
4141 | /// insert it before \p InsertPt. |
4142 | /// |
4143 | /// The returned callbr instruction is identical to \p CBI in every way |
4144 | /// except that the operand bundles for the new instruction are set to the |
4145 | /// operand bundles in \p Bundles. |
4146 | static CallBrInst *Create(CallBrInst *CBI, |
4147 | ArrayRef<OperandBundleDef> Bundles, |
4148 | Instruction *InsertPt = nullptr); |
4149 | |
4150 | /// Return the number of callbr indirect dest labels. |
4151 | /// |
4152 | unsigned getNumIndirectDests() const { return NumIndirectDests; } |
4153 | |
4154 | /// getIndirectDestLabel - Return the i-th indirect dest label. |
4155 | /// |
4156 | Value *getIndirectDestLabel(unsigned i) const { |
4157 | assert(i < getNumIndirectDests() && "Out of bounds!")(static_cast <bool> (i < getNumIndirectDests() && "Out of bounds!") ? void (0) : __assert_fail ("i < getNumIndirectDests() && \"Out of bounds!\"" , "llvm/include/llvm/IR/Instructions.h", 4157, __extension__ __PRETTY_FUNCTION__ )); |
4158 | return getOperand(i + arg_size() + getNumTotalBundleOperands() + 1); |
4159 | } |
4160 | |
4161 | Value *getIndirectDestLabelUse(unsigned i) const { |
4162 | assert(i < getNumIndirectDests() && "Out of bounds!")(static_cast <bool> (i < getNumIndirectDests() && "Out of bounds!") ? void (0) : __assert_fail ("i < getNumIndirectDests() && \"Out of bounds!\"" , "llvm/include/llvm/IR/Instructions.h", 4162, __extension__ __PRETTY_FUNCTION__ )); |
4163 | return getOperandUse(i + arg_size() + getNumTotalBundleOperands() + 1); |
4164 | } |
4165 | |
4166 | // Return the destination basic blocks... |
4167 | BasicBlock *getDefaultDest() const { |
4168 | return cast<BasicBlock>(*(&Op<-1>() - getNumIndirectDests() - 1)); |
4169 | } |
4170 | BasicBlock *getIndirectDest(unsigned i) const { |
4171 | return cast_or_null<BasicBlock>(*(&Op<-1>() - getNumIndirectDests() + i)); |
4172 | } |
4173 | SmallVector<BasicBlock *, 16> getIndirectDests() const { |
4174 | SmallVector<BasicBlock *, 16> IndirectDests; |
4175 | for (unsigned i = 0, e = getNumIndirectDests(); i < e; ++i) |
4176 | IndirectDests.push_back(getIndirectDest(i)); |
4177 | return IndirectDests; |
4178 | } |
4179 | void setDefaultDest(BasicBlock *B) { |
4180 | *(&Op<-1>() - getNumIndirectDests() - 1) = reinterpret_cast<Value *>(B); |
4181 | } |
4182 | void setIndirectDest(unsigned i, BasicBlock *B) { |
4183 | *(&Op<-1>() - getNumIndirectDests() + i) = reinterpret_cast<Value *>(B); |
4184 | } |
4185 | |
4186 | BasicBlock *getSuccessor(unsigned i) const { |
4187 | assert(i < getNumSuccessors() + 1 &&(static_cast <bool> (i < getNumSuccessors() + 1 && "Successor # out of range for callbr!") ? void (0) : __assert_fail ("i < getNumSuccessors() + 1 && \"Successor # out of range for callbr!\"" , "llvm/include/llvm/IR/Instructions.h", 4188, __extension__ __PRETTY_FUNCTION__ )) |
4188 | "Successor # out of range for callbr!")(static_cast <bool> (i < getNumSuccessors() + 1 && "Successor # out of range for callbr!") ? void (0) : __assert_fail ("i < getNumSuccessors() + 1 && \"Successor # out of range for callbr!\"" , "llvm/include/llvm/IR/Instructions.h", 4188, __extension__ __PRETTY_FUNCTION__ )); |
4189 | return i == 0 ? getDefaultDest() : getIndirectDest(i - 1); |
4190 | } |
4191 | |
4192 | void setSuccessor(unsigned i, BasicBlock *NewSucc) { |
4193 | assert(i < getNumIndirectDests() + 1 &&(static_cast <bool> (i < getNumIndirectDests() + 1 && "Successor # out of range for callbr!") ? void (0) : __assert_fail ("i < getNumIndirectDests() + 1 && \"Successor # out of range for callbr!\"" , "llvm/include/llvm/IR/Instructions.h", 4194, __extension__ __PRETTY_FUNCTION__ )) |
4194 | "Successor # out of range for callbr!") |