File: | lib/Transforms/Vectorize/SLPVectorizer.cpp |
Warning: | line 3062, column 27 Called C++ object pointer is null |
[?] Use j/k keys for keyboard navigation
1 | //===- SLPVectorizer.cpp - A bottom up SLP Vectorizer ---------------------===// | |||
2 | // | |||
3 | // The LLVM Compiler Infrastructure | |||
4 | // | |||
5 | // This file is distributed under the University of Illinois Open Source | |||
6 | // License. See LICENSE.TXT for details. | |||
7 | // | |||
8 | //===----------------------------------------------------------------------===// | |||
9 | // | |||
10 | // This pass implements the Bottom Up SLP vectorizer. It detects consecutive | |||
11 | // stores that can be put together into vector-stores. Next, it attempts to | |||
12 | // construct vectorizable tree using the use-def chains. If a profitable tree | |||
13 | // was found, the SLP vectorizer performs vectorization on the tree. | |||
14 | // | |||
15 | // The pass is inspired by the work described in the paper: | |||
16 | // "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks. | |||
17 | // | |||
18 | //===----------------------------------------------------------------------===// | |||
19 | ||||
20 | #include "llvm/Transforms/Vectorize/SLPVectorizer.h" | |||
21 | #include "llvm/ADT/ArrayRef.h" | |||
22 | #include "llvm/ADT/DenseMap.h" | |||
23 | #include "llvm/ADT/DenseSet.h" | |||
24 | #include "llvm/ADT/MapVector.h" | |||
25 | #include "llvm/ADT/None.h" | |||
26 | #include "llvm/ADT/Optional.h" | |||
27 | #include "llvm/ADT/PostOrderIterator.h" | |||
28 | #include "llvm/ADT/STLExtras.h" | |||
29 | #include "llvm/ADT/SetVector.h" | |||
30 | #include "llvm/ADT/SmallPtrSet.h" | |||
31 | #include "llvm/ADT/SmallSet.h" | |||
32 | #include "llvm/ADT/SmallVector.h" | |||
33 | #include "llvm/ADT/Statistic.h" | |||
34 | #include "llvm/ADT/iterator.h" | |||
35 | #include "llvm/ADT/iterator_range.h" | |||
36 | #include "llvm/Analysis/AliasAnalysis.h" | |||
37 | #include "llvm/Analysis/CodeMetrics.h" | |||
38 | #include "llvm/Analysis/DemandedBits.h" | |||
39 | #include "llvm/Analysis/GlobalsModRef.h" | |||
40 | #include "llvm/Analysis/LoopAccessAnalysis.h" | |||
41 | #include "llvm/Analysis/LoopInfo.h" | |||
42 | #include "llvm/Analysis/MemoryLocation.h" | |||
43 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | |||
44 | #include "llvm/Analysis/ScalarEvolution.h" | |||
45 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" | |||
46 | #include "llvm/Analysis/TargetLibraryInfo.h" | |||
47 | #include "llvm/Analysis/TargetTransformInfo.h" | |||
48 | #include "llvm/Analysis/ValueTracking.h" | |||
49 | #include "llvm/Analysis/VectorUtils.h" | |||
50 | #include "llvm/IR/Attributes.h" | |||
51 | #include "llvm/IR/BasicBlock.h" | |||
52 | #include "llvm/IR/Constant.h" | |||
53 | #include "llvm/IR/Constants.h" | |||
54 | #include "llvm/IR/DataLayout.h" | |||
55 | #include "llvm/IR/DebugLoc.h" | |||
56 | #include "llvm/IR/DerivedTypes.h" | |||
57 | #include "llvm/IR/Dominators.h" | |||
58 | #include "llvm/IR/Function.h" | |||
59 | #include "llvm/IR/IRBuilder.h" | |||
60 | #include "llvm/IR/InstrTypes.h" | |||
61 | #include "llvm/IR/Instruction.h" | |||
62 | #include "llvm/IR/Instructions.h" | |||
63 | #include "llvm/IR/IntrinsicInst.h" | |||
64 | #include "llvm/IR/Intrinsics.h" | |||
65 | #include "llvm/IR/Module.h" | |||
66 | #include "llvm/IR/NoFolder.h" | |||
67 | #include "llvm/IR/Operator.h" | |||
68 | #include "llvm/IR/PassManager.h" | |||
69 | #include "llvm/IR/PatternMatch.h" | |||
70 | #include "llvm/IR/Type.h" | |||
71 | #include "llvm/IR/Use.h" | |||
72 | #include "llvm/IR/User.h" | |||
73 | #include "llvm/IR/Value.h" | |||
74 | #include "llvm/IR/ValueHandle.h" | |||
75 | #include "llvm/IR/Verifier.h" | |||
76 | #include "llvm/Pass.h" | |||
77 | #include "llvm/Support/Casting.h" | |||
78 | #include "llvm/Support/CommandLine.h" | |||
79 | #include "llvm/Support/Compiler.h" | |||
80 | #include "llvm/Support/DOTGraphTraits.h" | |||
81 | #include "llvm/Support/Debug.h" | |||
82 | #include "llvm/Support/ErrorHandling.h" | |||
83 | #include "llvm/Support/GraphWriter.h" | |||
84 | #include "llvm/Support/KnownBits.h" | |||
85 | #include "llvm/Support/MathExtras.h" | |||
86 | #include "llvm/Support/raw_ostream.h" | |||
87 | #include "llvm/Transforms/Utils/LoopUtils.h" | |||
88 | #include "llvm/Transforms/Vectorize.h" | |||
89 | #include <algorithm> | |||
90 | #include <cassert> | |||
91 | #include <cstdint> | |||
92 | #include <iterator> | |||
93 | #include <memory> | |||
94 | #include <set> | |||
95 | #include <string> | |||
96 | #include <tuple> | |||
97 | #include <utility> | |||
98 | #include <vector> | |||
99 | ||||
100 | using namespace llvm; | |||
101 | using namespace llvm::PatternMatch; | |||
102 | using namespace slpvectorizer; | |||
103 | ||||
104 | #define SV_NAME"slp-vectorizer" "slp-vectorizer" | |||
105 | #define DEBUG_TYPE"SLP" "SLP" | |||
106 | ||||
107 | STATISTIC(NumVectorInstructions, "Number of vector instructions generated")static llvm::Statistic NumVectorInstructions = {"SLP", "NumVectorInstructions" , "Number of vector instructions generated", {0}, false}; | |||
108 | ||||
109 | static cl::opt<int> | |||
110 | SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden, | |||
111 | cl::desc("Only vectorize if you gain more than this " | |||
112 | "number ")); | |||
113 | ||||
114 | static cl::opt<bool> | |||
115 | ShouldVectorizeHor("slp-vectorize-hor", cl::init(true), cl::Hidden, | |||
116 | cl::desc("Attempt to vectorize horizontal reductions")); | |||
117 | ||||
118 | static cl::opt<bool> ShouldStartVectorizeHorAtStore( | |||
119 | "slp-vectorize-hor-store", cl::init(false), cl::Hidden, | |||
120 | cl::desc( | |||
121 | "Attempt to vectorize horizontal reductions feeding into a store")); | |||
122 | ||||
123 | static cl::opt<int> | |||
124 | MaxVectorRegSizeOption("slp-max-reg-size", cl::init(128), cl::Hidden, | |||
125 | cl::desc("Attempt to vectorize for this register size in bits")); | |||
126 | ||||
127 | /// Limits the size of scheduling regions in a block. | |||
128 | /// It avoid long compile times for _very_ large blocks where vector | |||
129 | /// instructions are spread over a wide range. | |||
130 | /// This limit is way higher than needed by real-world functions. | |||
131 | static cl::opt<int> | |||
132 | ScheduleRegionSizeBudget("slp-schedule-budget", cl::init(100000), cl::Hidden, | |||
133 | cl::desc("Limit the size of the SLP scheduling region per block")); | |||
134 | ||||
135 | static cl::opt<int> MinVectorRegSizeOption( | |||
136 | "slp-min-reg-size", cl::init(128), cl::Hidden, | |||
137 | cl::desc("Attempt to vectorize for this register size in bits")); | |||
138 | ||||
139 | static cl::opt<unsigned> RecursionMaxDepth( | |||
140 | "slp-recursion-max-depth", cl::init(12), cl::Hidden, | |||
141 | cl::desc("Limit the recursion depth when building a vectorizable tree")); | |||
142 | ||||
143 | static cl::opt<unsigned> MinTreeSize( | |||
144 | "slp-min-tree-size", cl::init(3), cl::Hidden, | |||
145 | cl::desc("Only vectorize small trees if they are fully vectorizable")); | |||
146 | ||||
147 | static cl::opt<bool> | |||
148 | ViewSLPTree("view-slp-tree", cl::Hidden, | |||
149 | cl::desc("Display the SLP trees with Graphviz")); | |||
150 | ||||
151 | // Limit the number of alias checks. The limit is chosen so that | |||
152 | // it has no negative effect on the llvm benchmarks. | |||
153 | static const unsigned AliasedCheckLimit = 10; | |||
154 | ||||
155 | // Another limit for the alias checks: The maximum distance between load/store | |||
156 | // instructions where alias checks are done. | |||
157 | // This limit is useful for very large basic blocks. | |||
158 | static const unsigned MaxMemDepDistance = 160; | |||
159 | ||||
160 | /// If the ScheduleRegionSizeBudget is exhausted, we allow small scheduling | |||
161 | /// regions to be handled. | |||
162 | static const int MinScheduleRegionSize = 16; | |||
163 | ||||
164 | /// \brief Predicate for the element types that the SLP vectorizer supports. | |||
165 | /// | |||
166 | /// The most important thing to filter here are types which are invalid in LLVM | |||
167 | /// vectors. We also filter target specific types which have absolutely no | |||
168 | /// meaningful vectorization path such as x86_fp80 and ppc_f128. This just | |||
169 | /// avoids spending time checking the cost model and realizing that they will | |||
170 | /// be inevitably scalarized. | |||
171 | static bool isValidElementType(Type *Ty) { | |||
172 | return VectorType::isValidElementType(Ty) && !Ty->isX86_FP80Ty() && | |||
173 | !Ty->isPPC_FP128Ty(); | |||
174 | } | |||
175 | ||||
176 | /// \returns true if all of the instructions in \p VL are in the same block or | |||
177 | /// false otherwise. | |||
178 | static bool allSameBlock(ArrayRef<Value *> VL) { | |||
179 | Instruction *I0 = dyn_cast<Instruction>(VL[0]); | |||
180 | if (!I0) | |||
181 | return false; | |||
182 | BasicBlock *BB = I0->getParent(); | |||
183 | for (int i = 1, e = VL.size(); i < e; i++) { | |||
184 | Instruction *I = dyn_cast<Instruction>(VL[i]); | |||
185 | if (!I) | |||
186 | return false; | |||
187 | ||||
188 | if (BB != I->getParent()) | |||
189 | return false; | |||
190 | } | |||
191 | return true; | |||
192 | } | |||
193 | ||||
194 | /// \returns True if all of the values in \p VL are constants. | |||
195 | static bool allConstant(ArrayRef<Value *> VL) { | |||
196 | for (Value *i : VL) | |||
197 | if (!isa<Constant>(i)) | |||
198 | return false; | |||
199 | return true; | |||
200 | } | |||
201 | ||||
202 | /// \returns True if all of the values in \p VL are identical. | |||
203 | static bool isSplat(ArrayRef<Value *> VL) { | |||
204 | for (unsigned i = 1, e = VL.size(); i < e; ++i) | |||
205 | if (VL[i] != VL[0]) | |||
206 | return false; | |||
207 | return true; | |||
208 | } | |||
209 | ||||
210 | /// Checks if the vector of instructions can be represented as a shuffle, like: | |||
211 | /// %x0 = extractelement <4 x i8> %x, i32 0 | |||
212 | /// %x3 = extractelement <4 x i8> %x, i32 3 | |||
213 | /// %y1 = extractelement <4 x i8> %y, i32 1 | |||
214 | /// %y2 = extractelement <4 x i8> %y, i32 2 | |||
215 | /// %x0x0 = mul i8 %x0, %x0 | |||
216 | /// %x3x3 = mul i8 %x3, %x3 | |||
217 | /// %y1y1 = mul i8 %y1, %y1 | |||
218 | /// %y2y2 = mul i8 %y2, %y2 | |||
219 | /// %ins1 = insertelement <4 x i8> undef, i8 %x0x0, i32 0 | |||
220 | /// %ins2 = insertelement <4 x i8> %ins1, i8 %x3x3, i32 1 | |||
221 | /// %ins3 = insertelement <4 x i8> %ins2, i8 %y1y1, i32 2 | |||
222 | /// %ins4 = insertelement <4 x i8> %ins3, i8 %y2y2, i32 3 | |||
223 | /// ret <4 x i8> %ins4 | |||
224 | /// can be transformed into: | |||
225 | /// %1 = shufflevector <4 x i8> %x, <4 x i8> %y, <4 x i32> <i32 0, i32 3, i32 5, | |||
226 | /// i32 6> | |||
227 | /// %2 = mul <4 x i8> %1, %1 | |||
228 | /// ret <4 x i8> %2 | |||
229 | /// We convert this initially to something like: | |||
230 | /// %x0 = extractelement <4 x i8> %x, i32 0 | |||
231 | /// %x3 = extractelement <4 x i8> %x, i32 3 | |||
232 | /// %y1 = extractelement <4 x i8> %y, i32 1 | |||
233 | /// %y2 = extractelement <4 x i8> %y, i32 2 | |||
234 | /// %1 = insertelement <4 x i8> undef, i8 %x0, i32 0 | |||
235 | /// %2 = insertelement <4 x i8> %1, i8 %x3, i32 1 | |||
236 | /// %3 = insertelement <4 x i8> %2, i8 %y1, i32 2 | |||
237 | /// %4 = insertelement <4 x i8> %3, i8 %y2, i32 3 | |||
238 | /// %5 = mul <4 x i8> %4, %4 | |||
239 | /// %6 = extractelement <4 x i8> %5, i32 0 | |||
240 | /// %ins1 = insertelement <4 x i8> undef, i8 %6, i32 0 | |||
241 | /// %7 = extractelement <4 x i8> %5, i32 1 | |||
242 | /// %ins2 = insertelement <4 x i8> %ins1, i8 %7, i32 1 | |||
243 | /// %8 = extractelement <4 x i8> %5, i32 2 | |||
244 | /// %ins3 = insertelement <4 x i8> %ins2, i8 %8, i32 2 | |||
245 | /// %9 = extractelement <4 x i8> %5, i32 3 | |||
246 | /// %ins4 = insertelement <4 x i8> %ins3, i8 %9, i32 3 | |||
247 | /// ret <4 x i8> %ins4 | |||
248 | /// InstCombiner transforms this into a shuffle and vector mul | |||
249 | static Optional<TargetTransformInfo::ShuffleKind> | |||
250 | isShuffle(ArrayRef<Value *> VL) { | |||
251 | auto *EI0 = cast<ExtractElementInst>(VL[0]); | |||
252 | unsigned Size = EI0->getVectorOperandType()->getVectorNumElements(); | |||
253 | Value *Vec1 = nullptr; | |||
254 | Value *Vec2 = nullptr; | |||
255 | enum ShuffleMode {Unknown, FirstAlternate, SecondAlternate, Permute}; | |||
256 | ShuffleMode CommonShuffleMode = Unknown; | |||
257 | for (unsigned I = 0, E = VL.size(); I < E; ++I) { | |||
258 | auto *EI = cast<ExtractElementInst>(VL[I]); | |||
259 | auto *Vec = EI->getVectorOperand(); | |||
260 | // All vector operands must have the same number of vector elements. | |||
261 | if (Vec->getType()->getVectorNumElements() != Size) | |||
262 | return None; | |||
263 | auto *Idx = dyn_cast<ConstantInt>(EI->getIndexOperand()); | |||
264 | if (!Idx) | |||
265 | return None; | |||
266 | // Undefined behavior if Idx is negative or >= Size. | |||
267 | if (Idx->getValue().uge(Size)) | |||
268 | continue; | |||
269 | unsigned IntIdx = Idx->getValue().getZExtValue(); | |||
270 | // We can extractelement from undef vector. | |||
271 | if (isa<UndefValue>(Vec)) | |||
272 | continue; | |||
273 | // For correct shuffling we have to have at most 2 different vector operands | |||
274 | // in all extractelement instructions. | |||
275 | if (Vec1 && Vec2 && Vec != Vec1 && Vec != Vec2) | |||
276 | return None; | |||
277 | if (CommonShuffleMode == Permute) | |||
278 | continue; | |||
279 | // If the extract index is not the same as the operation number, it is a | |||
280 | // permutation. | |||
281 | if (IntIdx != I) { | |||
282 | CommonShuffleMode = Permute; | |||
283 | continue; | |||
284 | } | |||
285 | // Check the shuffle mode for the current operation. | |||
286 | if (!Vec1) | |||
287 | Vec1 = Vec; | |||
288 | else if (Vec != Vec1) | |||
289 | Vec2 = Vec; | |||
290 | // Example: shufflevector A, B, <0,5,2,7> | |||
291 | // I is odd and IntIdx for A == I - FirstAlternate shuffle. | |||
292 | // I is even and IntIdx for B == I - FirstAlternate shuffle. | |||
293 | // Example: shufflevector A, B, <4,1,6,3> | |||
294 | // I is even and IntIdx for A == I - SecondAlternate shuffle. | |||
295 | // I is odd and IntIdx for B == I - SecondAlternate shuffle. | |||
296 | const bool IIsEven = I & 1; | |||
297 | const bool CurrVecIsA = Vec == Vec1; | |||
298 | const bool IIsOdd = !IIsEven; | |||
299 | const bool CurrVecIsB = !CurrVecIsA; | |||
300 | ShuffleMode CurrentShuffleMode = | |||
301 | ((IIsOdd && CurrVecIsA) || (IIsEven && CurrVecIsB)) ? FirstAlternate | |||
302 | : SecondAlternate; | |||
303 | // Common mode is not set or the same as the shuffle mode of the current | |||
304 | // operation - alternate. | |||
305 | if (CommonShuffleMode == Unknown) | |||
306 | CommonShuffleMode = CurrentShuffleMode; | |||
307 | // Common shuffle mode is not the same as the shuffle mode of the current | |||
308 | // operation - permutation. | |||
309 | if (CommonShuffleMode != CurrentShuffleMode) | |||
310 | CommonShuffleMode = Permute; | |||
311 | } | |||
312 | // If we're not crossing lanes in different vectors, consider it as blending. | |||
313 | if ((CommonShuffleMode == FirstAlternate || | |||
314 | CommonShuffleMode == SecondAlternate) && | |||
315 | Vec2) | |||
316 | return TargetTransformInfo::SK_Alternate; | |||
317 | // If Vec2 was never used, we have a permutation of a single vector, otherwise | |||
318 | // we have permutation of 2 vectors. | |||
319 | return Vec2 ? TargetTransformInfo::SK_PermuteTwoSrc | |||
320 | : TargetTransformInfo::SK_PermuteSingleSrc; | |||
321 | } | |||
322 | ||||
323 | ///\returns Opcode that can be clubbed with \p Op to create an alternate | |||
324 | /// sequence which can later be merged as a ShuffleVector instruction. | |||
325 | static unsigned getAltOpcode(unsigned Op) { | |||
326 | switch (Op) { | |||
327 | case Instruction::FAdd: | |||
328 | return Instruction::FSub; | |||
329 | case Instruction::FSub: | |||
330 | return Instruction::FAdd; | |||
331 | case Instruction::Add: | |||
332 | return Instruction::Sub; | |||
333 | case Instruction::Sub: | |||
334 | return Instruction::Add; | |||
335 | default: | |||
336 | return 0; | |||
337 | } | |||
338 | } | |||
339 | ||||
340 | static bool isOdd(unsigned Value) { | |||
341 | return Value & 1; | |||
342 | } | |||
343 | ||||
344 | static bool sameOpcodeOrAlt(unsigned Opcode, unsigned AltOpcode, | |||
345 | unsigned CheckedOpcode) { | |||
346 | return Opcode == CheckedOpcode || AltOpcode == CheckedOpcode; | |||
347 | } | |||
348 | ||||
349 | /// Chooses the correct key for scheduling data. If \p Op has the same (or | |||
350 | /// alternate) opcode as \p OpValue, the key is \p Op. Otherwise the key is \p | |||
351 | /// OpValue. | |||
352 | static Value *isOneOf(Value *OpValue, Value *Op) { | |||
353 | auto *I = dyn_cast<Instruction>(Op); | |||
354 | if (!I) | |||
355 | return OpValue; | |||
356 | auto *OpInst = cast<Instruction>(OpValue); | |||
357 | unsigned OpInstOpcode = OpInst->getOpcode(); | |||
358 | unsigned IOpcode = I->getOpcode(); | |||
359 | if (sameOpcodeOrAlt(OpInstOpcode, getAltOpcode(OpInstOpcode), IOpcode)) | |||
360 | return Op; | |||
361 | return OpValue; | |||
362 | } | |||
363 | ||||
364 | namespace { | |||
365 | ||||
366 | /// Contains data for the instructions going to be vectorized. | |||
367 | struct RawInstructionsData { | |||
368 | /// Main Opcode of the instructions going to be vectorized. | |||
369 | unsigned Opcode = 0; | |||
370 | ||||
371 | /// The list of instructions have some instructions with alternate opcodes. | |||
372 | bool HasAltOpcodes = false; | |||
373 | }; | |||
374 | ||||
375 | } // end anonymous namespace | |||
376 | ||||
377 | /// Checks the list of the vectorized instructions \p VL and returns info about | |||
378 | /// this list. | |||
379 | static RawInstructionsData getMainOpcode(ArrayRef<Value *> VL) { | |||
380 | auto *I0 = dyn_cast<Instruction>(VL[0]); | |||
381 | if (!I0) | |||
382 | return {}; | |||
383 | RawInstructionsData Res; | |||
384 | unsigned Opcode = I0->getOpcode(); | |||
385 | // Walk through the list of the vectorized instructions | |||
386 | // in order to check its structure described by RawInstructionsData. | |||
387 | for (unsigned Cnt = 0, E = VL.size(); Cnt != E; ++Cnt) { | |||
388 | auto *I = dyn_cast<Instruction>(VL[Cnt]); | |||
389 | if (!I) | |||
390 | return {}; | |||
391 | if (Opcode != I->getOpcode()) | |||
392 | Res.HasAltOpcodes = true; | |||
393 | } | |||
394 | Res.Opcode = Opcode; | |||
395 | return Res; | |||
396 | } | |||
397 | ||||
398 | namespace { | |||
399 | ||||
400 | /// Main data required for vectorization of instructions. | |||
401 | struct InstructionsState { | |||
402 | /// The very first instruction in the list with the main opcode. | |||
403 | Value *OpValue = nullptr; | |||
404 | ||||
405 | /// The main opcode for the list of instructions. | |||
406 | unsigned Opcode = 0; | |||
407 | ||||
408 | /// Some of the instructions in the list have alternate opcodes. | |||
409 | bool IsAltShuffle = false; | |||
410 | ||||
411 | InstructionsState() = default; | |||
412 | InstructionsState(Value *OpValue, unsigned Opcode, bool IsAltShuffle) | |||
413 | : OpValue(OpValue), Opcode(Opcode), IsAltShuffle(IsAltShuffle) {} | |||
414 | }; | |||
415 | ||||
416 | } // end anonymous namespace | |||
417 | ||||
418 | /// \returns analysis of the Instructions in \p VL described in | |||
419 | /// InstructionsState, the Opcode that we suppose the whole list | |||
420 | /// could be vectorized even if its structure is diverse. | |||
421 | static InstructionsState getSameOpcode(ArrayRef<Value *> VL) { | |||
422 | auto Res = getMainOpcode(VL); | |||
423 | unsigned Opcode = Res.Opcode; | |||
424 | if (!Res.HasAltOpcodes) | |||
425 | return InstructionsState(VL[0], Opcode, false); | |||
426 | auto *OpInst = cast<Instruction>(VL[0]); | |||
427 | unsigned AltOpcode = getAltOpcode(Opcode); | |||
428 | // Examine each element in the list instructions VL to determine | |||
429 | // if some operations there could be considered as an alternative | |||
430 | // (for example as subtraction relates to addition operation). | |||
431 | for (int Cnt = 0, E = VL.size(); Cnt < E; Cnt++) { | |||
432 | auto *I = cast<Instruction>(VL[Cnt]); | |||
433 | unsigned InstOpcode = I->getOpcode(); | |||
434 | if ((Res.HasAltOpcodes && | |||
435 | InstOpcode != (isOdd(Cnt) ? AltOpcode : Opcode)) || | |||
436 | (!Res.HasAltOpcodes && InstOpcode != Opcode)) { | |||
437 | return InstructionsState(OpInst, 0, false); | |||
438 | } | |||
439 | } | |||
440 | return InstructionsState(OpInst, Opcode, Res.HasAltOpcodes); | |||
441 | } | |||
442 | ||||
443 | /// \returns true if all of the values in \p VL have the same type or false | |||
444 | /// otherwise. | |||
445 | static bool allSameType(ArrayRef<Value *> VL) { | |||
446 | Type *Ty = VL[0]->getType(); | |||
447 | for (int i = 1, e = VL.size(); i < e; i++) | |||
448 | if (VL[i]->getType() != Ty) | |||
449 | return false; | |||
450 | ||||
451 | return true; | |||
452 | } | |||
453 | ||||
454 | /// \returns True if Extract{Value,Element} instruction extracts element Idx. | |||
455 | static bool matchExtractIndex(Instruction *E, unsigned Idx, unsigned Opcode) { | |||
456 | assert(Opcode == Instruction::ExtractElement ||(static_cast <bool> (Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) ? void (0) : __assert_fail ("Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 457, __extension__ __PRETTY_FUNCTION__)) | |||
457 | Opcode == Instruction::ExtractValue)(static_cast <bool> (Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) ? void (0) : __assert_fail ("Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 457, __extension__ __PRETTY_FUNCTION__)); | |||
458 | if (Opcode == Instruction::ExtractElement) { | |||
459 | ConstantInt *CI = dyn_cast<ConstantInt>(E->getOperand(1)); | |||
460 | return CI && CI->getZExtValue() == Idx; | |||
461 | } else { | |||
462 | ExtractValueInst *EI = cast<ExtractValueInst>(E); | |||
463 | return EI->getNumIndices() == 1 && *EI->idx_begin() == Idx; | |||
464 | } | |||
465 | } | |||
466 | ||||
467 | /// \returns True if in-tree use also needs extract. This refers to | |||
468 | /// possible scalar operand in vectorized instruction. | |||
469 | static bool InTreeUserNeedToExtract(Value *Scalar, Instruction *UserInst, | |||
470 | TargetLibraryInfo *TLI) { | |||
471 | unsigned Opcode = UserInst->getOpcode(); | |||
472 | switch (Opcode) { | |||
473 | case Instruction::Load: { | |||
474 | LoadInst *LI = cast<LoadInst>(UserInst); | |||
475 | return (LI->getPointerOperand() == Scalar); | |||
476 | } | |||
477 | case Instruction::Store: { | |||
478 | StoreInst *SI = cast<StoreInst>(UserInst); | |||
479 | return (SI->getPointerOperand() == Scalar); | |||
480 | } | |||
481 | case Instruction::Call: { | |||
482 | CallInst *CI = cast<CallInst>(UserInst); | |||
483 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
484 | if (hasVectorInstrinsicScalarOpd(ID, 1)) { | |||
485 | return (CI->getArgOperand(1) == Scalar); | |||
486 | } | |||
487 | LLVM_FALLTHROUGH[[clang::fallthrough]]; | |||
488 | } | |||
489 | default: | |||
490 | return false; | |||
491 | } | |||
492 | } | |||
493 | ||||
494 | /// \returns the AA location that is being access by the instruction. | |||
495 | static MemoryLocation getLocation(Instruction *I, AliasAnalysis *AA) { | |||
496 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) | |||
497 | return MemoryLocation::get(SI); | |||
498 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | |||
499 | return MemoryLocation::get(LI); | |||
500 | return MemoryLocation(); | |||
501 | } | |||
502 | ||||
503 | /// \returns True if the instruction is not a volatile or atomic load/store. | |||
504 | static bool isSimple(Instruction *I) { | |||
505 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | |||
506 | return LI->isSimple(); | |||
507 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) | |||
508 | return SI->isSimple(); | |||
509 | if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) | |||
510 | return !MI->isVolatile(); | |||
511 | return true; | |||
512 | } | |||
513 | ||||
514 | namespace llvm { | |||
515 | ||||
516 | namespace slpvectorizer { | |||
517 | ||||
518 | /// Bottom Up SLP Vectorizer. | |||
519 | class BoUpSLP { | |||
520 | public: | |||
521 | using ValueList = SmallVector<Value *, 8>; | |||
522 | using InstrList = SmallVector<Instruction *, 16>; | |||
523 | using ValueSet = SmallPtrSet<Value *, 16>; | |||
524 | using StoreList = SmallVector<StoreInst *, 8>; | |||
525 | using ExtraValueToDebugLocsMap = | |||
526 | MapVector<Value *, SmallVector<Instruction *, 2>>; | |||
527 | ||||
528 | BoUpSLP(Function *Func, ScalarEvolution *Se, TargetTransformInfo *Tti, | |||
529 | TargetLibraryInfo *TLi, AliasAnalysis *Aa, LoopInfo *Li, | |||
530 | DominatorTree *Dt, AssumptionCache *AC, DemandedBits *DB, | |||
531 | const DataLayout *DL, OptimizationRemarkEmitter *ORE) | |||
532 | : F(Func), SE(Se), TTI(Tti), TLI(TLi), AA(Aa), LI(Li), DT(Dt), AC(AC), | |||
533 | DB(DB), DL(DL), ORE(ORE), Builder(Se->getContext()) { | |||
534 | CodeMetrics::collectEphemeralValues(F, AC, EphValues); | |||
535 | // Use the vector register size specified by the target unless overridden | |||
536 | // by a command-line option. | |||
537 | // TODO: It would be better to limit the vectorization factor based on | |||
538 | // data type rather than just register size. For example, x86 AVX has | |||
539 | // 256-bit registers, but it does not support integer operations | |||
540 | // at that width (that requires AVX2). | |||
541 | if (MaxVectorRegSizeOption.getNumOccurrences()) | |||
542 | MaxVecRegSize = MaxVectorRegSizeOption; | |||
543 | else | |||
544 | MaxVecRegSize = TTI->getRegisterBitWidth(true); | |||
545 | ||||
546 | if (MinVectorRegSizeOption.getNumOccurrences()) | |||
547 | MinVecRegSize = MinVectorRegSizeOption; | |||
548 | else | |||
549 | MinVecRegSize = TTI->getMinVectorRegisterBitWidth(); | |||
550 | } | |||
551 | ||||
552 | /// \brief Vectorize the tree that starts with the elements in \p VL. | |||
553 | /// Returns the vectorized root. | |||
554 | Value *vectorizeTree(); | |||
555 | ||||
556 | /// Vectorize the tree but with the list of externally used values \p | |||
557 | /// ExternallyUsedValues. Values in this MapVector can be replaced but the | |||
558 | /// generated extractvalue instructions. | |||
559 | Value *vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues); | |||
560 | ||||
561 | /// \returns the cost incurred by unwanted spills and fills, caused by | |||
562 | /// holding live values over call sites. | |||
563 | int getSpillCost(); | |||
564 | ||||
565 | /// \returns the vectorization cost of the subtree that starts at \p VL. | |||
566 | /// A negative number means that this is profitable. | |||
567 | int getTreeCost(); | |||
568 | ||||
569 | /// Construct a vectorizable tree that starts at \p Roots, ignoring users for | |||
570 | /// the purpose of scheduling and extraction in the \p UserIgnoreLst. | |||
571 | void buildTree(ArrayRef<Value *> Roots, | |||
572 | ArrayRef<Value *> UserIgnoreLst = None); | |||
573 | ||||
574 | /// Construct a vectorizable tree that starts at \p Roots, ignoring users for | |||
575 | /// the purpose of scheduling and extraction in the \p UserIgnoreLst taking | |||
576 | /// into account (anf updating it, if required) list of externally used | |||
577 | /// values stored in \p ExternallyUsedValues. | |||
578 | void buildTree(ArrayRef<Value *> Roots, | |||
579 | ExtraValueToDebugLocsMap &ExternallyUsedValues, | |||
580 | ArrayRef<Value *> UserIgnoreLst = None); | |||
581 | ||||
582 | /// Clear the internal data structures that are created by 'buildTree'. | |||
583 | void deleteTree() { | |||
584 | VectorizableTree.clear(); | |||
585 | ScalarToTreeEntry.clear(); | |||
586 | MustGather.clear(); | |||
587 | ExternalUses.clear(); | |||
588 | NumLoadsWantToKeepOrder = 0; | |||
589 | NumLoadsWantToChangeOrder = 0; | |||
590 | for (auto &Iter : BlocksSchedules) { | |||
591 | BlockScheduling *BS = Iter.second.get(); | |||
592 | BS->clear(); | |||
593 | } | |||
594 | MinBWs.clear(); | |||
595 | } | |||
596 | ||||
597 | unsigned getTreeSize() const { return VectorizableTree.size(); } | |||
598 | ||||
599 | /// \brief Perform LICM and CSE on the newly generated gather sequences. | |||
600 | void optimizeGatherSequence(Function &F); | |||
601 | ||||
602 | /// \returns true if it is beneficial to reverse the vector order. | |||
603 | bool shouldReorder() const { | |||
604 | return NumLoadsWantToChangeOrder > NumLoadsWantToKeepOrder; | |||
605 | } | |||
606 | ||||
607 | /// \return The vector element size in bits to use when vectorizing the | |||
608 | /// expression tree ending at \p V. If V is a store, the size is the width of | |||
609 | /// the stored value. Otherwise, the size is the width of the largest loaded | |||
610 | /// value reaching V. This method is used by the vectorizer to calculate | |||
611 | /// vectorization factors. | |||
612 | unsigned getVectorElementSize(Value *V); | |||
613 | ||||
614 | /// Compute the minimum type sizes required to represent the entries in a | |||
615 | /// vectorizable tree. | |||
616 | void computeMinimumValueSizes(); | |||
617 | ||||
618 | // \returns maximum vector register size as set by TTI or overridden by cl::opt. | |||
619 | unsigned getMaxVecRegSize() const { | |||
620 | return MaxVecRegSize; | |||
621 | } | |||
622 | ||||
623 | // \returns minimum vector register size as set by cl::opt. | |||
624 | unsigned getMinVecRegSize() const { | |||
625 | return MinVecRegSize; | |||
626 | } | |||
627 | ||||
628 | /// \brief Check if ArrayType or StructType is isomorphic to some VectorType. | |||
629 | /// | |||
630 | /// \returns number of elements in vector if isomorphism exists, 0 otherwise. | |||
631 | unsigned canMapToVector(Type *T, const DataLayout &DL) const; | |||
632 | ||||
633 | /// \returns True if the VectorizableTree is both tiny and not fully | |||
634 | /// vectorizable. We do not vectorize such trees. | |||
635 | bool isTreeTinyAndNotFullyVectorizable(); | |||
636 | ||||
637 | OptimizationRemarkEmitter *getORE() { return ORE; } | |||
638 | ||||
639 | private: | |||
640 | struct TreeEntry; | |||
641 | ||||
642 | /// Checks if all users of \p I are the part of the vectorization tree. | |||
643 | bool areAllUsersVectorized(Instruction *I) const; | |||
644 | ||||
645 | /// \returns the cost of the vectorizable entry. | |||
646 | int getEntryCost(TreeEntry *E); | |||
647 | ||||
648 | /// This is the recursive part of buildTree. | |||
649 | void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth, int); | |||
650 | ||||
651 | /// \returns True if the ExtractElement/ExtractValue instructions in VL can | |||
652 | /// be vectorized to use the original vector (or aggregate "bitcast" to a vector). | |||
653 | bool canReuseExtract(ArrayRef<Value *> VL, Value *OpValue) const; | |||
654 | ||||
655 | /// Vectorize a single entry in the tree. | |||
656 | Value *vectorizeTree(TreeEntry *E); | |||
657 | ||||
658 | /// Vectorize a single entry in the tree, starting in \p VL. | |||
659 | Value *vectorizeTree(ArrayRef<Value *> VL); | |||
660 | ||||
661 | /// \returns the pointer to the vectorized value if \p VL is already | |||
662 | /// vectorized, or NULL. They may happen in cycles. | |||
663 | Value *alreadyVectorized(ArrayRef<Value *> VL, Value *OpValue) const; | |||
664 | ||||
665 | /// \returns the scalarization cost for this type. Scalarization in this | |||
666 | /// context means the creation of vectors from a group of scalars. | |||
667 | int getGatherCost(Type *Ty); | |||
668 | ||||
669 | /// \returns the scalarization cost for this list of values. Assuming that | |||
670 | /// this subtree gets vectorized, we may need to extract the values from the | |||
671 | /// roots. This method calculates the cost of extracting the values. | |||
672 | int getGatherCost(ArrayRef<Value *> VL); | |||
673 | ||||
674 | /// \brief Set the Builder insert point to one after the last instruction in | |||
675 | /// the bundle | |||
676 | void setInsertPointAfterBundle(ArrayRef<Value *> VL, Value *OpValue); | |||
677 | ||||
678 | /// \returns a vector from a collection of scalars in \p VL. | |||
679 | Value *Gather(ArrayRef<Value *> VL, VectorType *Ty); | |||
680 | ||||
681 | /// \returns whether the VectorizableTree is fully vectorizable and will | |||
682 | /// be beneficial even the tree height is tiny. | |||
683 | bool isFullyVectorizableTinyTree(); | |||
684 | ||||
685 | /// \reorder commutative operands in alt shuffle if they result in | |||
686 | /// vectorized code. | |||
687 | void reorderAltShuffleOperands(unsigned Opcode, ArrayRef<Value *> VL, | |||
688 | SmallVectorImpl<Value *> &Left, | |||
689 | SmallVectorImpl<Value *> &Right); | |||
690 | ||||
691 | /// \reorder commutative operands to get better probability of | |||
692 | /// generating vectorized code. | |||
693 | void reorderInputsAccordingToOpcode(unsigned Opcode, ArrayRef<Value *> VL, | |||
694 | SmallVectorImpl<Value *> &Left, | |||
695 | SmallVectorImpl<Value *> &Right); | |||
696 | struct TreeEntry { | |||
697 | TreeEntry(std::vector<TreeEntry> &Container) : Container(Container) {} | |||
698 | ||||
699 | /// \returns true if the scalars in VL are equal to this entry. | |||
700 | bool isSame(ArrayRef<Value *> VL) const { | |||
701 | assert(VL.size() == Scalars.size() && "Invalid size")(static_cast <bool> (VL.size() == Scalars.size() && "Invalid size") ? void (0) : __assert_fail ("VL.size() == Scalars.size() && \"Invalid size\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 701, __extension__ __PRETTY_FUNCTION__)); | |||
702 | return std::equal(VL.begin(), VL.end(), Scalars.begin()); | |||
703 | } | |||
704 | ||||
705 | /// A vector of scalars. | |||
706 | ValueList Scalars; | |||
707 | ||||
708 | /// The Scalars are vectorized into this value. It is initialized to Null. | |||
709 | Value *VectorizedValue = nullptr; | |||
710 | ||||
711 | /// Do we need to gather this sequence ? | |||
712 | bool NeedToGather = false; | |||
713 | ||||
714 | /// Points back to the VectorizableTree. | |||
715 | /// | |||
716 | /// Only used for Graphviz right now. Unfortunately GraphTrait::NodeRef has | |||
717 | /// to be a pointer and needs to be able to initialize the child iterator. | |||
718 | /// Thus we need a reference back to the container to translate the indices | |||
719 | /// to entries. | |||
720 | std::vector<TreeEntry> &Container; | |||
721 | ||||
722 | /// The TreeEntry index containing the user of this entry. We can actually | |||
723 | /// have multiple users so the data structure is not truly a tree. | |||
724 | SmallVector<int, 1> UserTreeIndices; | |||
725 | }; | |||
726 | ||||
727 | /// Create a new VectorizableTree entry. | |||
728 | TreeEntry *newTreeEntry(ArrayRef<Value *> VL, bool Vectorized, | |||
729 | int &UserTreeIdx) { | |||
730 | VectorizableTree.emplace_back(VectorizableTree); | |||
731 | int idx = VectorizableTree.size() - 1; | |||
732 | TreeEntry *Last = &VectorizableTree[idx]; | |||
733 | Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end()); | |||
734 | Last->NeedToGather = !Vectorized; | |||
735 | if (Vectorized) { | |||
736 | for (int i = 0, e = VL.size(); i != e; ++i) { | |||
737 | assert(!getTreeEntry(VL[i]) && "Scalar already in tree!")(static_cast <bool> (!getTreeEntry(VL[i]) && "Scalar already in tree!" ) ? void (0) : __assert_fail ("!getTreeEntry(VL[i]) && \"Scalar already in tree!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 737, __extension__ __PRETTY_FUNCTION__)); | |||
738 | ScalarToTreeEntry[VL[i]] = idx; | |||
739 | } | |||
740 | } else { | |||
741 | MustGather.insert(VL.begin(), VL.end()); | |||
742 | } | |||
743 | ||||
744 | if (UserTreeIdx >= 0) | |||
745 | Last->UserTreeIndices.push_back(UserTreeIdx); | |||
746 | UserTreeIdx = idx; | |||
747 | return Last; | |||
748 | } | |||
749 | ||||
750 | /// -- Vectorization State -- | |||
751 | /// Holds all of the tree entries. | |||
752 | std::vector<TreeEntry> VectorizableTree; | |||
753 | ||||
754 | TreeEntry *getTreeEntry(Value *V) { | |||
755 | auto I = ScalarToTreeEntry.find(V); | |||
756 | if (I != ScalarToTreeEntry.end()) | |||
757 | return &VectorizableTree[I->second]; | |||
758 | return nullptr; | |||
759 | } | |||
760 | ||||
761 | const TreeEntry *getTreeEntry(Value *V) const { | |||
762 | auto I = ScalarToTreeEntry.find(V); | |||
763 | if (I != ScalarToTreeEntry.end()) | |||
764 | return &VectorizableTree[I->second]; | |||
765 | return nullptr; | |||
766 | } | |||
767 | ||||
768 | /// Maps a specific scalar to its tree entry. | |||
769 | SmallDenseMap<Value*, int> ScalarToTreeEntry; | |||
770 | ||||
771 | /// A list of scalars that we found that we need to keep as scalars. | |||
772 | ValueSet MustGather; | |||
773 | ||||
774 | /// This POD struct describes one external user in the vectorized tree. | |||
775 | struct ExternalUser { | |||
776 | ExternalUser(Value *S, llvm::User *U, int L) | |||
777 | : Scalar(S), User(U), Lane(L) {} | |||
778 | ||||
779 | // Which scalar in our function. | |||
780 | Value *Scalar; | |||
781 | ||||
782 | // Which user that uses the scalar. | |||
783 | llvm::User *User; | |||
784 | ||||
785 | // Which lane does the scalar belong to. | |||
786 | int Lane; | |||
787 | }; | |||
788 | using UserList = SmallVector<ExternalUser, 16>; | |||
789 | ||||
790 | /// Checks if two instructions may access the same memory. | |||
791 | /// | |||
792 | /// \p Loc1 is the location of \p Inst1. It is passed explicitly because it | |||
793 | /// is invariant in the calling loop. | |||
794 | bool isAliased(const MemoryLocation &Loc1, Instruction *Inst1, | |||
795 | Instruction *Inst2) { | |||
796 | // First check if the result is already in the cache. | |||
797 | AliasCacheKey key = std::make_pair(Inst1, Inst2); | |||
798 | Optional<bool> &result = AliasCache[key]; | |||
799 | if (result.hasValue()) { | |||
800 | return result.getValue(); | |||
801 | } | |||
802 | MemoryLocation Loc2 = getLocation(Inst2, AA); | |||
803 | bool aliased = true; | |||
804 | if (Loc1.Ptr && Loc2.Ptr && isSimple(Inst1) && isSimple(Inst2)) { | |||
805 | // Do the alias check. | |||
806 | aliased = AA->alias(Loc1, Loc2); | |||
807 | } | |||
808 | // Store the result in the cache. | |||
809 | result = aliased; | |||
810 | return aliased; | |||
811 | } | |||
812 | ||||
813 | using AliasCacheKey = std::pair<Instruction *, Instruction *>; | |||
814 | ||||
815 | /// Cache for alias results. | |||
816 | /// TODO: consider moving this to the AliasAnalysis itself. | |||
817 | DenseMap<AliasCacheKey, Optional<bool>> AliasCache; | |||
818 | ||||
819 | /// Removes an instruction from its block and eventually deletes it. | |||
820 | /// It's like Instruction::eraseFromParent() except that the actual deletion | |||
821 | /// is delayed until BoUpSLP is destructed. | |||
822 | /// This is required to ensure that there are no incorrect collisions in the | |||
823 | /// AliasCache, which can happen if a new instruction is allocated at the | |||
824 | /// same address as a previously deleted instruction. | |||
825 | void eraseInstruction(Instruction *I) { | |||
826 | I->removeFromParent(); | |||
827 | I->dropAllReferences(); | |||
828 | DeletedInstructions.emplace_back(I); | |||
829 | } | |||
830 | ||||
831 | /// Temporary store for deleted instructions. Instructions will be deleted | |||
832 | /// eventually when the BoUpSLP is destructed. | |||
833 | SmallVector<unique_value, 8> DeletedInstructions; | |||
834 | ||||
835 | /// A list of values that need to extracted out of the tree. | |||
836 | /// This list holds pairs of (Internal Scalar : External User). External User | |||
837 | /// can be nullptr, it means that this Internal Scalar will be used later, | |||
838 | /// after vectorization. | |||
839 | UserList ExternalUses; | |||
840 | ||||
841 | /// Values used only by @llvm.assume calls. | |||
842 | SmallPtrSet<const Value *, 32> EphValues; | |||
843 | ||||
844 | /// Holds all of the instructions that we gathered. | |||
845 | SetVector<Instruction *> GatherSeq; | |||
846 | ||||
847 | /// A list of blocks that we are going to CSE. | |||
848 | SetVector<BasicBlock *> CSEBlocks; | |||
849 | ||||
850 | /// Contains all scheduling relevant data for an instruction. | |||
851 | /// A ScheduleData either represents a single instruction or a member of an | |||
852 | /// instruction bundle (= a group of instructions which is combined into a | |||
853 | /// vector instruction). | |||
854 | struct ScheduleData { | |||
855 | // The initial value for the dependency counters. It means that the | |||
856 | // dependencies are not calculated yet. | |||
857 | enum { InvalidDeps = -1 }; | |||
858 | ||||
859 | ScheduleData() = default; | |||
860 | ||||
861 | void init(int BlockSchedulingRegionID, Value *OpVal) { | |||
862 | FirstInBundle = this; | |||
863 | NextInBundle = nullptr; | |||
864 | NextLoadStore = nullptr; | |||
865 | IsScheduled = false; | |||
866 | SchedulingRegionID = BlockSchedulingRegionID; | |||
867 | UnscheduledDepsInBundle = UnscheduledDeps; | |||
868 | clearDependencies(); | |||
869 | OpValue = OpVal; | |||
870 | } | |||
871 | ||||
872 | /// Returns true if the dependency information has been calculated. | |||
873 | bool hasValidDependencies() const { return Dependencies != InvalidDeps; } | |||
874 | ||||
875 | /// Returns true for single instructions and for bundle representatives | |||
876 | /// (= the head of a bundle). | |||
877 | bool isSchedulingEntity() const { return FirstInBundle == this; } | |||
878 | ||||
879 | /// Returns true if it represents an instruction bundle and not only a | |||
880 | /// single instruction. | |||
881 | bool isPartOfBundle() const { | |||
882 | return NextInBundle != nullptr || FirstInBundle != this; | |||
883 | } | |||
884 | ||||
885 | /// Returns true if it is ready for scheduling, i.e. it has no more | |||
886 | /// unscheduled depending instructions/bundles. | |||
887 | bool isReady() const { | |||
888 | assert(isSchedulingEntity() &&(static_cast <bool> (isSchedulingEntity() && "can't consider non-scheduling entity for ready list" ) ? void (0) : __assert_fail ("isSchedulingEntity() && \"can't consider non-scheduling entity for ready list\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 889, __extension__ __PRETTY_FUNCTION__)) | |||
889 | "can't consider non-scheduling entity for ready list")(static_cast <bool> (isSchedulingEntity() && "can't consider non-scheduling entity for ready list" ) ? void (0) : __assert_fail ("isSchedulingEntity() && \"can't consider non-scheduling entity for ready list\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 889, __extension__ __PRETTY_FUNCTION__)); | |||
890 | return UnscheduledDepsInBundle == 0 && !IsScheduled; | |||
891 | } | |||
892 | ||||
893 | /// Modifies the number of unscheduled dependencies, also updating it for | |||
894 | /// the whole bundle. | |||
895 | int incrementUnscheduledDeps(int Incr) { | |||
896 | UnscheduledDeps += Incr; | |||
897 | return FirstInBundle->UnscheduledDepsInBundle += Incr; | |||
898 | } | |||
899 | ||||
900 | /// Sets the number of unscheduled dependencies to the number of | |||
901 | /// dependencies. | |||
902 | void resetUnscheduledDeps() { | |||
903 | incrementUnscheduledDeps(Dependencies - UnscheduledDeps); | |||
904 | } | |||
905 | ||||
906 | /// Clears all dependency information. | |||
907 | void clearDependencies() { | |||
908 | Dependencies = InvalidDeps; | |||
909 | resetUnscheduledDeps(); | |||
910 | MemoryDependencies.clear(); | |||
911 | } | |||
912 | ||||
913 | void dump(raw_ostream &os) const { | |||
914 | if (!isSchedulingEntity()) { | |||
915 | os << "/ " << *Inst; | |||
916 | } else if (NextInBundle) { | |||
917 | os << '[' << *Inst; | |||
918 | ScheduleData *SD = NextInBundle; | |||
919 | while (SD) { | |||
920 | os << ';' << *SD->Inst; | |||
921 | SD = SD->NextInBundle; | |||
922 | } | |||
923 | os << ']'; | |||
924 | } else { | |||
925 | os << *Inst; | |||
926 | } | |||
927 | } | |||
928 | ||||
929 | Instruction *Inst = nullptr; | |||
930 | ||||
931 | /// Points to the head in an instruction bundle (and always to this for | |||
932 | /// single instructions). | |||
933 | ScheduleData *FirstInBundle = nullptr; | |||
934 | ||||
935 | /// Single linked list of all instructions in a bundle. Null if it is a | |||
936 | /// single instruction. | |||
937 | ScheduleData *NextInBundle = nullptr; | |||
938 | ||||
939 | /// Single linked list of all memory instructions (e.g. load, store, call) | |||
940 | /// in the block - until the end of the scheduling region. | |||
941 | ScheduleData *NextLoadStore = nullptr; | |||
942 | ||||
943 | /// The dependent memory instructions. | |||
944 | /// This list is derived on demand in calculateDependencies(). | |||
945 | SmallVector<ScheduleData *, 4> MemoryDependencies; | |||
946 | ||||
947 | /// This ScheduleData is in the current scheduling region if this matches | |||
948 | /// the current SchedulingRegionID of BlockScheduling. | |||
949 | int SchedulingRegionID = 0; | |||
950 | ||||
951 | /// Used for getting a "good" final ordering of instructions. | |||
952 | int SchedulingPriority = 0; | |||
953 | ||||
954 | /// The number of dependencies. Constitutes of the number of users of the | |||
955 | /// instruction plus the number of dependent memory instructions (if any). | |||
956 | /// This value is calculated on demand. | |||
957 | /// If InvalidDeps, the number of dependencies is not calculated yet. | |||
958 | int Dependencies = InvalidDeps; | |||
959 | ||||
960 | /// The number of dependencies minus the number of dependencies of scheduled | |||
961 | /// instructions. As soon as this is zero, the instruction/bundle gets ready | |||
962 | /// for scheduling. | |||
963 | /// Note that this is negative as long as Dependencies is not calculated. | |||
964 | int UnscheduledDeps = InvalidDeps; | |||
965 | ||||
966 | /// The sum of UnscheduledDeps in a bundle. Equals to UnscheduledDeps for | |||
967 | /// single instructions. | |||
968 | int UnscheduledDepsInBundle = InvalidDeps; | |||
969 | ||||
970 | /// True if this instruction is scheduled (or considered as scheduled in the | |||
971 | /// dry-run). | |||
972 | bool IsScheduled = false; | |||
973 | ||||
974 | /// Opcode of the current instruction in the schedule data. | |||
975 | Value *OpValue = nullptr; | |||
976 | }; | |||
977 | ||||
978 | #ifndef NDEBUG | |||
979 | friend inline raw_ostream &operator<<(raw_ostream &os, | |||
980 | const BoUpSLP::ScheduleData &SD) { | |||
981 | SD.dump(os); | |||
982 | return os; | |||
983 | } | |||
984 | #endif | |||
985 | ||||
986 | friend struct GraphTraits<BoUpSLP *>; | |||
987 | friend struct DOTGraphTraits<BoUpSLP *>; | |||
988 | ||||
989 | /// Contains all scheduling data for a basic block. | |||
990 | struct BlockScheduling { | |||
991 | BlockScheduling(BasicBlock *BB) | |||
992 | : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize) {} | |||
993 | ||||
994 | void clear() { | |||
995 | ReadyInsts.clear(); | |||
996 | ScheduleStart = nullptr; | |||
997 | ScheduleEnd = nullptr; | |||
998 | FirstLoadStoreInRegion = nullptr; | |||
999 | LastLoadStoreInRegion = nullptr; | |||
1000 | ||||
1001 | // Reduce the maximum schedule region size by the size of the | |||
1002 | // previous scheduling run. | |||
1003 | ScheduleRegionSizeLimit -= ScheduleRegionSize; | |||
1004 | if (ScheduleRegionSizeLimit < MinScheduleRegionSize) | |||
1005 | ScheduleRegionSizeLimit = MinScheduleRegionSize; | |||
1006 | ScheduleRegionSize = 0; | |||
1007 | ||||
1008 | // Make a new scheduling region, i.e. all existing ScheduleData is not | |||
1009 | // in the new region yet. | |||
1010 | ++SchedulingRegionID; | |||
1011 | } | |||
1012 | ||||
1013 | ScheduleData *getScheduleData(Value *V) { | |||
1014 | ScheduleData *SD = ScheduleDataMap[V]; | |||
1015 | if (SD && SD->SchedulingRegionID == SchedulingRegionID) | |||
1016 | return SD; | |||
1017 | return nullptr; | |||
1018 | } | |||
1019 | ||||
1020 | ScheduleData *getScheduleData(Value *V, Value *Key) { | |||
1021 | if (V == Key) | |||
1022 | return getScheduleData(V); | |||
1023 | auto I = ExtraScheduleDataMap.find(V); | |||
1024 | if (I != ExtraScheduleDataMap.end()) { | |||
1025 | ScheduleData *SD = I->second[Key]; | |||
1026 | if (SD && SD->SchedulingRegionID == SchedulingRegionID) | |||
1027 | return SD; | |||
1028 | } | |||
1029 | return nullptr; | |||
1030 | } | |||
1031 | ||||
1032 | bool isInSchedulingRegion(ScheduleData *SD) { | |||
1033 | return SD->SchedulingRegionID == SchedulingRegionID; | |||
1034 | } | |||
1035 | ||||
1036 | /// Marks an instruction as scheduled and puts all dependent ready | |||
1037 | /// instructions into the ready-list. | |||
1038 | template <typename ReadyListType> | |||
1039 | void schedule(ScheduleData *SD, ReadyListType &ReadyList) { | |||
1040 | SD->IsScheduled = true; | |||
1041 | DEBUG(dbgs() << "SLP: schedule " << *SD << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: schedule " << *SD << "\n"; } } while (false); | |||
1042 | ||||
1043 | ScheduleData *BundleMember = SD; | |||
1044 | while (BundleMember) { | |||
1045 | if (BundleMember->Inst != BundleMember->OpValue) { | |||
1046 | BundleMember = BundleMember->NextInBundle; | |||
1047 | continue; | |||
1048 | } | |||
1049 | // Handle the def-use chain dependencies. | |||
1050 | for (Use &U : BundleMember->Inst->operands()) { | |||
1051 | auto *I = dyn_cast<Instruction>(U.get()); | |||
1052 | if (!I) | |||
1053 | continue; | |||
1054 | doForAllOpcodes(I, [&ReadyList](ScheduleData *OpDef) { | |||
1055 | if (OpDef && OpDef->hasValidDependencies() && | |||
1056 | OpDef->incrementUnscheduledDeps(-1) == 0) { | |||
1057 | // There are no more unscheduled dependencies after | |||
1058 | // decrementing, so we can put the dependent instruction | |||
1059 | // into the ready list. | |||
1060 | ScheduleData *DepBundle = OpDef->FirstInBundle; | |||
1061 | assert(!DepBundle->IsScheduled &&(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1062, __extension__ __PRETTY_FUNCTION__)) | |||
1062 | "already scheduled bundle gets ready")(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1062, __extension__ __PRETTY_FUNCTION__)); | |||
1063 | ReadyList.insert(DepBundle); | |||
1064 | DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"; } } while (false) | |||
1065 | << "SLP: gets ready (def): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"; } } while (false); | |||
1066 | } | |||
1067 | }); | |||
1068 | } | |||
1069 | // Handle the memory dependencies. | |||
1070 | for (ScheduleData *MemoryDepSD : BundleMember->MemoryDependencies) { | |||
1071 | if (MemoryDepSD->incrementUnscheduledDeps(-1) == 0) { | |||
1072 | // There are no more unscheduled dependencies after decrementing, | |||
1073 | // so we can put the dependent instruction into the ready list. | |||
1074 | ScheduleData *DepBundle = MemoryDepSD->FirstInBundle; | |||
1075 | assert(!DepBundle->IsScheduled &&(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1076, __extension__ __PRETTY_FUNCTION__)) | |||
1076 | "already scheduled bundle gets ready")(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1076, __extension__ __PRETTY_FUNCTION__)); | |||
1077 | ReadyList.insert(DepBundle); | |||
1078 | DEBUG(dbgs() << "SLP: gets ready (mem): " << *DepBundledo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"; } } while (false) | |||
1079 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"; } } while (false); | |||
1080 | } | |||
1081 | } | |||
1082 | BundleMember = BundleMember->NextInBundle; | |||
1083 | } | |||
1084 | } | |||
1085 | ||||
1086 | void doForAllOpcodes(Value *V, | |||
1087 | function_ref<void(ScheduleData *SD)> Action) { | |||
1088 | if (ScheduleData *SD = getScheduleData(V)) | |||
1089 | Action(SD); | |||
1090 | auto I = ExtraScheduleDataMap.find(V); | |||
1091 | if (I != ExtraScheduleDataMap.end()) | |||
1092 | for (auto &P : I->second) | |||
1093 | if (P.second->SchedulingRegionID == SchedulingRegionID) | |||
1094 | Action(P.second); | |||
1095 | } | |||
1096 | ||||
1097 | /// Put all instructions into the ReadyList which are ready for scheduling. | |||
1098 | template <typename ReadyListType> | |||
1099 | void initialFillReadyList(ReadyListType &ReadyList) { | |||
1100 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | |||
1101 | doForAllOpcodes(I, [&](ScheduleData *SD) { | |||
1102 | if (SD->isSchedulingEntity() && SD->isReady()) { | |||
1103 | ReadyList.insert(SD); | |||
1104 | DEBUG(dbgs() << "SLP: initially in ready list: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: initially in ready list: " << *I << "\n"; } } while (false); | |||
1105 | } | |||
1106 | }); | |||
1107 | } | |||
1108 | } | |||
1109 | ||||
1110 | /// Checks if a bundle of instructions can be scheduled, i.e. has no | |||
1111 | /// cyclic dependencies. This is only a dry-run, no instructions are | |||
1112 | /// actually moved at this stage. | |||
1113 | bool tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP, Value *OpValue); | |||
1114 | ||||
1115 | /// Un-bundles a group of instructions. | |||
1116 | void cancelScheduling(ArrayRef<Value *> VL, Value *OpValue); | |||
1117 | ||||
1118 | /// Allocates schedule data chunk. | |||
1119 | ScheduleData *allocateScheduleDataChunks(); | |||
1120 | ||||
1121 | /// Extends the scheduling region so that V is inside the region. | |||
1122 | /// \returns true if the region size is within the limit. | |||
1123 | bool extendSchedulingRegion(Value *V, Value *OpValue); | |||
1124 | ||||
1125 | /// Initialize the ScheduleData structures for new instructions in the | |||
1126 | /// scheduling region. | |||
1127 | void initScheduleData(Instruction *FromI, Instruction *ToI, | |||
1128 | ScheduleData *PrevLoadStore, | |||
1129 | ScheduleData *NextLoadStore); | |||
1130 | ||||
1131 | /// Updates the dependency information of a bundle and of all instructions/ | |||
1132 | /// bundles which depend on the original bundle. | |||
1133 | void calculateDependencies(ScheduleData *SD, bool InsertInReadyList, | |||
1134 | BoUpSLP *SLP); | |||
1135 | ||||
1136 | /// Sets all instruction in the scheduling region to un-scheduled. | |||
1137 | void resetSchedule(); | |||
1138 | ||||
1139 | BasicBlock *BB; | |||
1140 | ||||
1141 | /// Simple memory allocation for ScheduleData. | |||
1142 | std::vector<std::unique_ptr<ScheduleData[]>> ScheduleDataChunks; | |||
1143 | ||||
1144 | /// The size of a ScheduleData array in ScheduleDataChunks. | |||
1145 | int ChunkSize; | |||
1146 | ||||
1147 | /// The allocator position in the current chunk, which is the last entry | |||
1148 | /// of ScheduleDataChunks. | |||
1149 | int ChunkPos; | |||
1150 | ||||
1151 | /// Attaches ScheduleData to Instruction. | |||
1152 | /// Note that the mapping survives during all vectorization iterations, i.e. | |||
1153 | /// ScheduleData structures are recycled. | |||
1154 | DenseMap<Value *, ScheduleData *> ScheduleDataMap; | |||
1155 | ||||
1156 | /// Attaches ScheduleData to Instruction with the leading key. | |||
1157 | DenseMap<Value *, SmallDenseMap<Value *, ScheduleData *>> | |||
1158 | ExtraScheduleDataMap; | |||
1159 | ||||
1160 | struct ReadyList : SmallVector<ScheduleData *, 8> { | |||
1161 | void insert(ScheduleData *SD) { push_back(SD); } | |||
1162 | }; | |||
1163 | ||||
1164 | /// The ready-list for scheduling (only used for the dry-run). | |||
1165 | ReadyList ReadyInsts; | |||
1166 | ||||
1167 | /// The first instruction of the scheduling region. | |||
1168 | Instruction *ScheduleStart = nullptr; | |||
1169 | ||||
1170 | /// The first instruction _after_ the scheduling region. | |||
1171 | Instruction *ScheduleEnd = nullptr; | |||
1172 | ||||
1173 | /// The first memory accessing instruction in the scheduling region | |||
1174 | /// (can be null). | |||
1175 | ScheduleData *FirstLoadStoreInRegion = nullptr; | |||
1176 | ||||
1177 | /// The last memory accessing instruction in the scheduling region | |||
1178 | /// (can be null). | |||
1179 | ScheduleData *LastLoadStoreInRegion = nullptr; | |||
1180 | ||||
1181 | /// The current size of the scheduling region. | |||
1182 | int ScheduleRegionSize = 0; | |||
1183 | ||||
1184 | /// The maximum size allowed for the scheduling region. | |||
1185 | int ScheduleRegionSizeLimit = ScheduleRegionSizeBudget; | |||
1186 | ||||
1187 | /// The ID of the scheduling region. For a new vectorization iteration this | |||
1188 | /// is incremented which "removes" all ScheduleData from the region. | |||
1189 | // Make sure that the initial SchedulingRegionID is greater than the | |||
1190 | // initial SchedulingRegionID in ScheduleData (which is 0). | |||
1191 | int SchedulingRegionID = 1; | |||
1192 | }; | |||
1193 | ||||
1194 | /// Attaches the BlockScheduling structures to basic blocks. | |||
1195 | MapVector<BasicBlock *, std::unique_ptr<BlockScheduling>> BlocksSchedules; | |||
1196 | ||||
1197 | /// Performs the "real" scheduling. Done before vectorization is actually | |||
1198 | /// performed in a basic block. | |||
1199 | void scheduleBlock(BlockScheduling *BS); | |||
1200 | ||||
1201 | /// List of users to ignore during scheduling and that don't need extracting. | |||
1202 | ArrayRef<Value *> UserIgnoreList; | |||
1203 | ||||
1204 | // Number of load bundles that contain consecutive loads. | |||
1205 | int NumLoadsWantToKeepOrder = 0; | |||
1206 | ||||
1207 | // Number of load bundles that contain consecutive loads in reversed order. | |||
1208 | int NumLoadsWantToChangeOrder = 0; | |||
1209 | ||||
1210 | // Analysis and block reference. | |||
1211 | Function *F; | |||
1212 | ScalarEvolution *SE; | |||
1213 | TargetTransformInfo *TTI; | |||
1214 | TargetLibraryInfo *TLI; | |||
1215 | AliasAnalysis *AA; | |||
1216 | LoopInfo *LI; | |||
1217 | DominatorTree *DT; | |||
1218 | AssumptionCache *AC; | |||
1219 | DemandedBits *DB; | |||
1220 | const DataLayout *DL; | |||
1221 | OptimizationRemarkEmitter *ORE; | |||
1222 | ||||
1223 | unsigned MaxVecRegSize; // This is set by TTI or overridden by cl::opt. | |||
1224 | unsigned MinVecRegSize; // Set by cl::opt (default: 128). | |||
1225 | ||||
1226 | /// Instruction builder to construct the vectorized tree. | |||
1227 | IRBuilder<> Builder; | |||
1228 | ||||
1229 | /// A map of scalar integer values to the smallest bit width with which they | |||
1230 | /// can legally be represented. The values map to (width, signed) pairs, | |||
1231 | /// where "width" indicates the minimum bit width and "signed" is True if the | |||
1232 | /// value must be signed-extended, rather than zero-extended, back to its | |||
1233 | /// original width. | |||
1234 | MapVector<Value *, std::pair<uint64_t, bool>> MinBWs; | |||
1235 | }; | |||
1236 | ||||
1237 | } // end namespace slpvectorizer | |||
1238 | ||||
1239 | template <> struct GraphTraits<BoUpSLP *> { | |||
1240 | using TreeEntry = BoUpSLP::TreeEntry; | |||
1241 | ||||
1242 | /// NodeRef has to be a pointer per the GraphWriter. | |||
1243 | using NodeRef = TreeEntry *; | |||
1244 | ||||
1245 | /// \brief Add the VectorizableTree to the index iterator to be able to return | |||
1246 | /// TreeEntry pointers. | |||
1247 | struct ChildIteratorType | |||
1248 | : public iterator_adaptor_base<ChildIteratorType, | |||
1249 | SmallVector<int, 1>::iterator> { | |||
1250 | std::vector<TreeEntry> &VectorizableTree; | |||
1251 | ||||
1252 | ChildIteratorType(SmallVector<int, 1>::iterator W, | |||
1253 | std::vector<TreeEntry> &VT) | |||
1254 | : ChildIteratorType::iterator_adaptor_base(W), VectorizableTree(VT) {} | |||
1255 | ||||
1256 | NodeRef operator*() { return &VectorizableTree[*I]; } | |||
1257 | }; | |||
1258 | ||||
1259 | static NodeRef getEntryNode(BoUpSLP &R) { return &R.VectorizableTree[0]; } | |||
1260 | ||||
1261 | static ChildIteratorType child_begin(NodeRef N) { | |||
1262 | return {N->UserTreeIndices.begin(), N->Container}; | |||
1263 | } | |||
1264 | ||||
1265 | static ChildIteratorType child_end(NodeRef N) { | |||
1266 | return {N->UserTreeIndices.end(), N->Container}; | |||
1267 | } | |||
1268 | ||||
1269 | /// For the node iterator we just need to turn the TreeEntry iterator into a | |||
1270 | /// TreeEntry* iterator so that it dereferences to NodeRef. | |||
1271 | using nodes_iterator = pointer_iterator<std::vector<TreeEntry>::iterator>; | |||
1272 | ||||
1273 | static nodes_iterator nodes_begin(BoUpSLP *R) { | |||
1274 | return nodes_iterator(R->VectorizableTree.begin()); | |||
1275 | } | |||
1276 | ||||
1277 | static nodes_iterator nodes_end(BoUpSLP *R) { | |||
1278 | return nodes_iterator(R->VectorizableTree.end()); | |||
1279 | } | |||
1280 | ||||
1281 | static unsigned size(BoUpSLP *R) { return R->VectorizableTree.size(); } | |||
1282 | }; | |||
1283 | ||||
1284 | template <> struct DOTGraphTraits<BoUpSLP *> : public DefaultDOTGraphTraits { | |||
1285 | using TreeEntry = BoUpSLP::TreeEntry; | |||
1286 | ||||
1287 | DOTGraphTraits(bool isSimple = false) : DefaultDOTGraphTraits(isSimple) {} | |||
1288 | ||||
1289 | std::string getNodeLabel(const TreeEntry *Entry, const BoUpSLP *R) { | |||
1290 | std::string Str; | |||
1291 | raw_string_ostream OS(Str); | |||
1292 | if (isSplat(Entry->Scalars)) { | |||
1293 | OS << "<splat> " << *Entry->Scalars[0]; | |||
1294 | return Str; | |||
1295 | } | |||
1296 | for (auto V : Entry->Scalars) { | |||
1297 | OS << *V; | |||
1298 | if (std::any_of( | |||
1299 | R->ExternalUses.begin(), R->ExternalUses.end(), | |||
1300 | [&](const BoUpSLP::ExternalUser &EU) { return EU.Scalar == V; })) | |||
1301 | OS << " <extract>"; | |||
1302 | OS << "\n"; | |||
1303 | } | |||
1304 | return Str; | |||
1305 | } | |||
1306 | ||||
1307 | static std::string getNodeAttributes(const TreeEntry *Entry, | |||
1308 | const BoUpSLP *) { | |||
1309 | if (Entry->NeedToGather) | |||
1310 | return "color=red"; | |||
1311 | return ""; | |||
1312 | } | |||
1313 | }; | |||
1314 | ||||
1315 | } // end namespace llvm | |||
1316 | ||||
1317 | void BoUpSLP::buildTree(ArrayRef<Value *> Roots, | |||
1318 | ArrayRef<Value *> UserIgnoreLst) { | |||
1319 | ExtraValueToDebugLocsMap ExternallyUsedValues; | |||
1320 | buildTree(Roots, ExternallyUsedValues, UserIgnoreLst); | |||
1321 | } | |||
1322 | ||||
1323 | void BoUpSLP::buildTree(ArrayRef<Value *> Roots, | |||
1324 | ExtraValueToDebugLocsMap &ExternallyUsedValues, | |||
1325 | ArrayRef<Value *> UserIgnoreLst) { | |||
1326 | deleteTree(); | |||
1327 | UserIgnoreList = UserIgnoreLst; | |||
1328 | if (!allSameType(Roots)) | |||
1329 | return; | |||
1330 | buildTree_rec(Roots, 0, -1); | |||
1331 | ||||
1332 | // Collect the values that we need to extract from the tree. | |||
1333 | for (TreeEntry &EIdx : VectorizableTree) { | |||
1334 | TreeEntry *Entry = &EIdx; | |||
1335 | ||||
1336 | // No need to handle users of gathered values. | |||
1337 | if (Entry->NeedToGather) | |||
1338 | continue; | |||
1339 | ||||
1340 | // For each lane: | |||
1341 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { | |||
1342 | Value *Scalar = Entry->Scalars[Lane]; | |||
1343 | ||||
1344 | // Check if the scalar is externally used as an extra arg. | |||
1345 | auto ExtI = ExternallyUsedValues.find(Scalar); | |||
1346 | if (ExtI != ExternallyUsedValues.end()) { | |||
1347 | DEBUG(dbgs() << "SLP: Need to extract: Extra arg from lane " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to extract: Extra arg from lane " << Lane << " from " << *Scalar << ".\n" ; } } while (false) | |||
1348 | Lane << " from " << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to extract: Extra arg from lane " << Lane << " from " << *Scalar << ".\n" ; } } while (false); | |||
1349 | ExternalUses.emplace_back(Scalar, nullptr, Lane); | |||
1350 | continue; | |||
1351 | } | |||
1352 | for (User *U : Scalar->users()) { | |||
1353 | DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Checking user:" << *U << ".\n"; } } while (false); | |||
1354 | ||||
1355 | Instruction *UserInst = dyn_cast<Instruction>(U); | |||
1356 | if (!UserInst) | |||
1357 | continue; | |||
1358 | ||||
1359 | // Skip in-tree scalars that become vectors | |||
1360 | if (TreeEntry *UseEntry = getTreeEntry(U)) { | |||
1361 | Value *UseScalar = UseEntry->Scalars[0]; | |||
1362 | // Some in-tree scalars will remain as scalar in vectorized | |||
1363 | // instructions. If that is the case, the one in Lane 0 will | |||
1364 | // be used. | |||
1365 | if (UseScalar != U || | |||
1366 | !InTreeUserNeedToExtract(Scalar, UserInst, TLI)) { | |||
1367 | DEBUG(dbgs() << "SLP: \tInternal user will be removed:" << *Udo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tInternal user will be removed:" << *U << ".\n"; } } while (false) | |||
1368 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tInternal user will be removed:" << *U << ".\n"; } } while (false); | |||
1369 | assert(!UseEntry->NeedToGather && "Bad state")(static_cast <bool> (!UseEntry->NeedToGather && "Bad state") ? void (0) : __assert_fail ("!UseEntry->NeedToGather && \"Bad state\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1369, __extension__ __PRETTY_FUNCTION__)); | |||
1370 | continue; | |||
1371 | } | |||
1372 | } | |||
1373 | ||||
1374 | // Ignore users in the user ignore list. | |||
1375 | if (is_contained(UserIgnoreList, UserInst)) | |||
1376 | continue; | |||
1377 | ||||
1378 | DEBUG(dbgs() << "SLP: Need to extract:" << *U << " from lane " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to extract:" << * U << " from lane " << Lane << " from " << *Scalar << ".\n"; } } while (false) | |||
1379 | Lane << " from " << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to extract:" << * U << " from lane " << Lane << " from " << *Scalar << ".\n"; } } while (false); | |||
1380 | ExternalUses.push_back(ExternalUser(Scalar, U, Lane)); | |||
1381 | } | |||
1382 | } | |||
1383 | } | |||
1384 | } | |||
1385 | ||||
1386 | void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth, | |||
1387 | int UserTreeIdx) { | |||
1388 | assert((allConstant(VL) || allSameType(VL)) && "Invalid types!")(static_cast <bool> ((allConstant(VL) || allSameType(VL )) && "Invalid types!") ? void (0) : __assert_fail ("(allConstant(VL) || allSameType(VL)) && \"Invalid types!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1388, __extension__ __PRETTY_FUNCTION__)); | |||
1389 | ||||
1390 | InstructionsState S = getSameOpcode(VL); | |||
1391 | if (Depth == RecursionMaxDepth) { | |||
1392 | DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to max recursion depth.\n" ; } } while (false); | |||
1393 | newTreeEntry(VL, false, UserTreeIdx); | |||
1394 | return; | |||
1395 | } | |||
1396 | ||||
1397 | // Don't handle vectors. | |||
1398 | if (S.OpValue->getType()->isVectorTy()) { | |||
1399 | DEBUG(dbgs() << "SLP: Gathering due to vector type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to vector type.\n" ; } } while (false); | |||
1400 | newTreeEntry(VL, false, UserTreeIdx); | |||
1401 | return; | |||
1402 | } | |||
1403 | ||||
1404 | if (StoreInst *SI = dyn_cast<StoreInst>(S.OpValue)) | |||
1405 | if (SI->getValueOperand()->getType()->isVectorTy()) { | |||
1406 | DEBUG(dbgs() << "SLP: Gathering due to store vector type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to store vector type.\n" ; } } while (false); | |||
1407 | newTreeEntry(VL, false, UserTreeIdx); | |||
1408 | return; | |||
1409 | } | |||
1410 | ||||
1411 | // If all of the operands are identical or constant we have a simple solution. | |||
1412 | if (allConstant(VL) || isSplat(VL) || !allSameBlock(VL) || !S.Opcode) { | |||
1413 | DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O. \n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to C,S,B,O. \n" ; } } while (false); | |||
1414 | newTreeEntry(VL, false, UserTreeIdx); | |||
1415 | return; | |||
1416 | } | |||
1417 | ||||
1418 | // We now know that this is a vector of instructions of the same type from | |||
1419 | // the same block. | |||
1420 | ||||
1421 | // Don't vectorize ephemeral values. | |||
1422 | for (unsigned i = 0, e = VL.size(); i != e; ++i) { | |||
1423 | if (EphValues.count(VL[i])) { | |||
1424 | DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * VL[i] << ") is ephemeral.\n"; } } while (false) | |||
1425 | ") is ephemeral.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * VL[i] << ") is ephemeral.\n"; } } while (false); | |||
1426 | newTreeEntry(VL, false, UserTreeIdx); | |||
1427 | return; | |||
1428 | } | |||
1429 | } | |||
1430 | ||||
1431 | // Check if this is a duplicate of another entry. | |||
1432 | if (TreeEntry *E = getTreeEntry(S.OpValue)) { | |||
1433 | for (unsigned i = 0, e = VL.size(); i != e; ++i) { | |||
1434 | DEBUG(dbgs() << "SLP: \tChecking bundle: " << *VL[i] << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tChecking bundle: " << *VL[i] << ".\n"; } } while (false); | |||
1435 | if (E->Scalars[i] != VL[i]) { | |||
1436 | DEBUG(dbgs() << "SLP: Gathering due to partial overlap.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to partial overlap.\n" ; } } while (false); | |||
1437 | newTreeEntry(VL, false, UserTreeIdx); | |||
1438 | return; | |||
1439 | } | |||
1440 | } | |||
1441 | // Record the reuse of the tree node. FIXME, currently this is only used to | |||
1442 | // properly draw the graph rather than for the actual vectorization. | |||
1443 | E->UserTreeIndices.push_back(UserTreeIdx); | |||
1444 | DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *S.OpValue << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Perfect diamond merge at " << *S.OpValue << ".\n"; } } while (false); | |||
1445 | return; | |||
1446 | } | |||
1447 | ||||
1448 | // Check that none of the instructions in the bundle are already in the tree. | |||
1449 | for (unsigned i = 0, e = VL.size(); i != e; ++i) { | |||
1450 | auto *I = dyn_cast<Instruction>(VL[i]); | |||
1451 | if (!I) | |||
1452 | continue; | |||
1453 | if (getTreeEntry(I)) { | |||
1454 | DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * VL[i] << ") is already in tree.\n"; } } while (false) | |||
1455 | ") is already in tree.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * VL[i] << ") is already in tree.\n"; } } while (false); | |||
1456 | newTreeEntry(VL, false, UserTreeIdx); | |||
1457 | return; | |||
1458 | } | |||
1459 | } | |||
1460 | ||||
1461 | // If any of the scalars is marked as a value that needs to stay scalar, then | |||
1462 | // we need to gather the scalars. | |||
1463 | for (unsigned i = 0, e = VL.size(); i != e; ++i) { | |||
1464 | if (MustGather.count(VL[i])) { | |||
1465 | DEBUG(dbgs() << "SLP: Gathering due to gathered scalar.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to gathered scalar.\n" ; } } while (false); | |||
1466 | newTreeEntry(VL, false, UserTreeIdx); | |||
1467 | return; | |||
1468 | } | |||
1469 | } | |||
1470 | ||||
1471 | // Check that all of the users of the scalars that we want to vectorize are | |||
1472 | // schedulable. | |||
1473 | auto *VL0 = cast<Instruction>(S.OpValue); | |||
1474 | BasicBlock *BB = VL0->getParent(); | |||
1475 | ||||
1476 | if (!DT->isReachableFromEntry(BB)) { | |||
1477 | // Don't go into unreachable blocks. They may contain instructions with | |||
1478 | // dependency cycles which confuse the final scheduling. | |||
1479 | DEBUG(dbgs() << "SLP: bundle in unreachable block.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: bundle in unreachable block.\n" ; } } while (false); | |||
1480 | newTreeEntry(VL, false, UserTreeIdx); | |||
1481 | return; | |||
1482 | } | |||
1483 | ||||
1484 | // Check that every instruction appears once in this bundle. | |||
1485 | for (unsigned i = 0, e = VL.size(); i < e; ++i) | |||
1486 | for (unsigned j = i + 1; j < e; ++j) | |||
1487 | if (VL[i] == VL[j]) { | |||
1488 | DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Scalar used twice in bundle.\n" ; } } while (false); | |||
1489 | newTreeEntry(VL, false, UserTreeIdx); | |||
1490 | return; | |||
1491 | } | |||
1492 | ||||
1493 | auto &BSRef = BlocksSchedules[BB]; | |||
1494 | if (!BSRef) | |||
1495 | BSRef = llvm::make_unique<BlockScheduling>(BB); | |||
1496 | ||||
1497 | BlockScheduling &BS = *BSRef.get(); | |||
1498 | ||||
1499 | if (!BS.tryScheduleBundle(VL, this, S.OpValue)) { | |||
1500 | DEBUG(dbgs() << "SLP: We are not able to schedule this bundle!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: We are not able to schedule this bundle!\n" ; } } while (false); | |||
1501 | assert((!BS.getScheduleData(VL0) ||(static_cast <bool> ((!BS.getScheduleData(VL0) || !BS.getScheduleData (VL0)->isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure" ) ? void (0) : __assert_fail ("(!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1503, __extension__ __PRETTY_FUNCTION__)) | |||
1502 | !BS.getScheduleData(VL0)->isPartOfBundle()) &&(static_cast <bool> ((!BS.getScheduleData(VL0) || !BS.getScheduleData (VL0)->isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure" ) ? void (0) : __assert_fail ("(!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1503, __extension__ __PRETTY_FUNCTION__)) | |||
1503 | "tryScheduleBundle should cancelScheduling on failure")(static_cast <bool> ((!BS.getScheduleData(VL0) || !BS.getScheduleData (VL0)->isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure" ) ? void (0) : __assert_fail ("(!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1503, __extension__ __PRETTY_FUNCTION__)); | |||
1504 | newTreeEntry(VL, false, UserTreeIdx); | |||
1505 | return; | |||
1506 | } | |||
1507 | DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: We are able to schedule this bundle.\n" ; } } while (false); | |||
1508 | ||||
1509 | unsigned ShuffleOrOp = S.IsAltShuffle ? | |||
1510 | (unsigned) Instruction::ShuffleVector : S.Opcode; | |||
1511 | switch (ShuffleOrOp) { | |||
1512 | case Instruction::PHI: { | |||
1513 | PHINode *PH = dyn_cast<PHINode>(VL0); | |||
1514 | ||||
1515 | // Check for terminator values (e.g. invoke). | |||
1516 | for (unsigned j = 0; j < VL.size(); ++j) | |||
1517 | for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { | |||
1518 | TerminatorInst *Term = dyn_cast<TerminatorInst>( | |||
1519 | cast<PHINode>(VL[j])->getIncomingValueForBlock(PH->getIncomingBlock(i))); | |||
1520 | if (Term) { | |||
1521 | DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n" ; } } while (false); | |||
1522 | BS.cancelScheduling(VL, VL0); | |||
1523 | newTreeEntry(VL, false, UserTreeIdx); | |||
1524 | return; | |||
1525 | } | |||
1526 | } | |||
1527 | ||||
1528 | newTreeEntry(VL, true, UserTreeIdx); | |||
1529 | DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of PHINodes.\n" ; } } while (false); | |||
1530 | ||||
1531 | for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { | |||
1532 | ValueList Operands; | |||
1533 | // Prepare the operand vector. | |||
1534 | for (Value *j : VL) | |||
1535 | Operands.push_back(cast<PHINode>(j)->getIncomingValueForBlock( | |||
1536 | PH->getIncomingBlock(i))); | |||
1537 | ||||
1538 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1539 | } | |||
1540 | return; | |||
1541 | } | |||
1542 | case Instruction::ExtractValue: | |||
1543 | case Instruction::ExtractElement: { | |||
1544 | bool Reuse = canReuseExtract(VL, VL0); | |||
1545 | if (Reuse) { | |||
1546 | DEBUG(dbgs() << "SLP: Reusing extract sequence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Reusing extract sequence.\n" ; } } while (false); | |||
1547 | } else { | |||
1548 | BS.cancelScheduling(VL, VL0); | |||
1549 | } | |||
1550 | newTreeEntry(VL, Reuse, UserTreeIdx); | |||
1551 | return; | |||
1552 | } | |||
1553 | case Instruction::Load: { | |||
1554 | // Check that a vectorized load would load the same memory as a scalar | |||
1555 | // load. For example, we don't want to vectorize loads that are smaller | |||
1556 | // than 8-bit. Even though we have a packed struct {<i2, i2, i2, i2>} LLVM | |||
1557 | // treats loading/storing it as an i8 struct. If we vectorize loads/stores | |||
1558 | // from such a struct, we read/write packed bits disagreeing with the | |||
1559 | // unvectorized version. | |||
1560 | Type *ScalarTy = VL0->getType(); | |||
1561 | ||||
1562 | if (DL->getTypeSizeInBits(ScalarTy) != | |||
1563 | DL->getTypeAllocSizeInBits(ScalarTy)) { | |||
1564 | BS.cancelScheduling(VL, VL0); | |||
1565 | newTreeEntry(VL, false, UserTreeIdx); | |||
1566 | DEBUG(dbgs() << "SLP: Gathering loads of non-packed type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering loads of non-packed type.\n" ; } } while (false); | |||
1567 | return; | |||
1568 | } | |||
1569 | ||||
1570 | // Make sure all loads in the bundle are simple - we can't vectorize | |||
1571 | // atomic or volatile loads. | |||
1572 | for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) { | |||
1573 | LoadInst *L = cast<LoadInst>(VL[i]); | |||
1574 | if (!L->isSimple()) { | |||
1575 | BS.cancelScheduling(VL, VL0); | |||
1576 | newTreeEntry(VL, false, UserTreeIdx); | |||
1577 | DEBUG(dbgs() << "SLP: Gathering non-simple loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering non-simple loads.\n" ; } } while (false); | |||
1578 | return; | |||
1579 | } | |||
1580 | } | |||
1581 | ||||
1582 | // Check if the loads are consecutive, reversed, or neither. | |||
1583 | // TODO: What we really want is to sort the loads, but for now, check | |||
1584 | // the two likely directions. | |||
1585 | bool Consecutive = true; | |||
1586 | bool ReverseConsecutive = true; | |||
1587 | for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) { | |||
1588 | if (!isConsecutiveAccess(VL[i], VL[i + 1], *DL, *SE)) { | |||
1589 | Consecutive = false; | |||
1590 | break; | |||
1591 | } else { | |||
1592 | ReverseConsecutive = false; | |||
1593 | } | |||
1594 | } | |||
1595 | ||||
1596 | if (Consecutive) { | |||
1597 | ++NumLoadsWantToKeepOrder; | |||
1598 | newTreeEntry(VL, true, UserTreeIdx); | |||
1599 | DEBUG(dbgs() << "SLP: added a vector of loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of loads.\n"; } } while (false); | |||
1600 | return; | |||
1601 | } | |||
1602 | ||||
1603 | // If none of the load pairs were consecutive when checked in order, | |||
1604 | // check the reverse order. | |||
1605 | if (ReverseConsecutive) | |||
1606 | for (unsigned i = VL.size() - 1; i > 0; --i) | |||
1607 | if (!isConsecutiveAccess(VL[i], VL[i - 1], *DL, *SE)) { | |||
1608 | ReverseConsecutive = false; | |||
1609 | break; | |||
1610 | } | |||
1611 | ||||
1612 | BS.cancelScheduling(VL, VL0); | |||
1613 | newTreeEntry(VL, false, UserTreeIdx); | |||
1614 | ||||
1615 | if (ReverseConsecutive) { | |||
1616 | ++NumLoadsWantToChangeOrder; | |||
1617 | DEBUG(dbgs() << "SLP: Gathering reversed loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering reversed loads.\n" ; } } while (false); | |||
1618 | } else { | |||
1619 | DEBUG(dbgs() << "SLP: Gathering non-consecutive loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering non-consecutive loads.\n" ; } } while (false); | |||
1620 | } | |||
1621 | return; | |||
1622 | } | |||
1623 | case Instruction::ZExt: | |||
1624 | case Instruction::SExt: | |||
1625 | case Instruction::FPToUI: | |||
1626 | case Instruction::FPToSI: | |||
1627 | case Instruction::FPExt: | |||
1628 | case Instruction::PtrToInt: | |||
1629 | case Instruction::IntToPtr: | |||
1630 | case Instruction::SIToFP: | |||
1631 | case Instruction::UIToFP: | |||
1632 | case Instruction::Trunc: | |||
1633 | case Instruction::FPTrunc: | |||
1634 | case Instruction::BitCast: { | |||
1635 | Type *SrcTy = VL0->getOperand(0)->getType(); | |||
1636 | for (unsigned i = 0; i < VL.size(); ++i) { | |||
1637 | Type *Ty = cast<Instruction>(VL[i])->getOperand(0)->getType(); | |||
1638 | if (Ty != SrcTy || !isValidElementType(Ty)) { | |||
1639 | BS.cancelScheduling(VL, VL0); | |||
1640 | newTreeEntry(VL, false, UserTreeIdx); | |||
1641 | DEBUG(dbgs() << "SLP: Gathering casts with different src types.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering casts with different src types.\n" ; } } while (false); | |||
1642 | return; | |||
1643 | } | |||
1644 | } | |||
1645 | newTreeEntry(VL, true, UserTreeIdx); | |||
1646 | DEBUG(dbgs() << "SLP: added a vector of casts.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of casts.\n"; } } while (false); | |||
1647 | ||||
1648 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
1649 | ValueList Operands; | |||
1650 | // Prepare the operand vector. | |||
1651 | for (Value *j : VL) | |||
1652 | Operands.push_back(cast<Instruction>(j)->getOperand(i)); | |||
1653 | ||||
1654 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1655 | } | |||
1656 | return; | |||
1657 | } | |||
1658 | case Instruction::ICmp: | |||
1659 | case Instruction::FCmp: { | |||
1660 | // Check that all of the compares have the same predicate. | |||
1661 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); | |||
1662 | Type *ComparedTy = VL0->getOperand(0)->getType(); | |||
1663 | for (unsigned i = 1, e = VL.size(); i < e; ++i) { | |||
1664 | CmpInst *Cmp = cast<CmpInst>(VL[i]); | |||
1665 | if (Cmp->getPredicate() != P0 || | |||
1666 | Cmp->getOperand(0)->getType() != ComparedTy) { | |||
1667 | BS.cancelScheduling(VL, VL0); | |||
1668 | newTreeEntry(VL, false, UserTreeIdx); | |||
1669 | DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering cmp with different predicate.\n" ; } } while (false); | |||
1670 | return; | |||
1671 | } | |||
1672 | } | |||
1673 | ||||
1674 | newTreeEntry(VL, true, UserTreeIdx); | |||
1675 | DEBUG(dbgs() << "SLP: added a vector of compares.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of compares.\n" ; } } while (false); | |||
1676 | ||||
1677 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
1678 | ValueList Operands; | |||
1679 | // Prepare the operand vector. | |||
1680 | for (Value *j : VL) | |||
1681 | Operands.push_back(cast<Instruction>(j)->getOperand(i)); | |||
1682 | ||||
1683 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1684 | } | |||
1685 | return; | |||
1686 | } | |||
1687 | case Instruction::Select: | |||
1688 | case Instruction::Add: | |||
1689 | case Instruction::FAdd: | |||
1690 | case Instruction::Sub: | |||
1691 | case Instruction::FSub: | |||
1692 | case Instruction::Mul: | |||
1693 | case Instruction::FMul: | |||
1694 | case Instruction::UDiv: | |||
1695 | case Instruction::SDiv: | |||
1696 | case Instruction::FDiv: | |||
1697 | case Instruction::URem: | |||
1698 | case Instruction::SRem: | |||
1699 | case Instruction::FRem: | |||
1700 | case Instruction::Shl: | |||
1701 | case Instruction::LShr: | |||
1702 | case Instruction::AShr: | |||
1703 | case Instruction::And: | |||
1704 | case Instruction::Or: | |||
1705 | case Instruction::Xor: | |||
1706 | newTreeEntry(VL, true, UserTreeIdx); | |||
1707 | DEBUG(dbgs() << "SLP: added a vector of bin op.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of bin op.\n" ; } } while (false); | |||
1708 | ||||
1709 | // Sort operands of the instructions so that each side is more likely to | |||
1710 | // have the same opcode. | |||
1711 | if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) { | |||
1712 | ValueList Left, Right; | |||
1713 | reorderInputsAccordingToOpcode(S.Opcode, VL, Left, Right); | |||
1714 | buildTree_rec(Left, Depth + 1, UserTreeIdx); | |||
1715 | buildTree_rec(Right, Depth + 1, UserTreeIdx); | |||
1716 | return; | |||
1717 | } | |||
1718 | ||||
1719 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
1720 | ValueList Operands; | |||
1721 | // Prepare the operand vector. | |||
1722 | for (Value *j : VL) | |||
1723 | Operands.push_back(cast<Instruction>(j)->getOperand(i)); | |||
1724 | ||||
1725 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1726 | } | |||
1727 | return; | |||
1728 | ||||
1729 | case Instruction::GetElementPtr: { | |||
1730 | // We don't combine GEPs with complicated (nested) indexing. | |||
1731 | for (unsigned j = 0; j < VL.size(); ++j) { | |||
1732 | if (cast<Instruction>(VL[j])->getNumOperands() != 2) { | |||
1733 | DEBUG(dbgs() << "SLP: not-vectorizable GEP (nested indexes).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (nested indexes).\n" ; } } while (false); | |||
1734 | BS.cancelScheduling(VL, VL0); | |||
1735 | newTreeEntry(VL, false, UserTreeIdx); | |||
1736 | return; | |||
1737 | } | |||
1738 | } | |||
1739 | ||||
1740 | // We can't combine several GEPs into one vector if they operate on | |||
1741 | // different types. | |||
1742 | Type *Ty0 = VL0->getOperand(0)->getType(); | |||
1743 | for (unsigned j = 0; j < VL.size(); ++j) { | |||
1744 | Type *CurTy = cast<Instruction>(VL[j])->getOperand(0)->getType(); | |||
1745 | if (Ty0 != CurTy) { | |||
1746 | DEBUG(dbgs() << "SLP: not-vectorizable GEP (different types).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (different types).\n" ; } } while (false); | |||
1747 | BS.cancelScheduling(VL, VL0); | |||
1748 | newTreeEntry(VL, false, UserTreeIdx); | |||
1749 | return; | |||
1750 | } | |||
1751 | } | |||
1752 | ||||
1753 | // We don't combine GEPs with non-constant indexes. | |||
1754 | for (unsigned j = 0; j < VL.size(); ++j) { | |||
1755 | auto Op = cast<Instruction>(VL[j])->getOperand(1); | |||
1756 | if (!isa<ConstantInt>(Op)) { | |||
1757 | DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n" ; } } while (false) | |||
1758 | dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n" ; } } while (false); | |||
1759 | BS.cancelScheduling(VL, VL0); | |||
1760 | newTreeEntry(VL, false, UserTreeIdx); | |||
1761 | return; | |||
1762 | } | |||
1763 | } | |||
1764 | ||||
1765 | newTreeEntry(VL, true, UserTreeIdx); | |||
1766 | DEBUG(dbgs() << "SLP: added a vector of GEPs.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of GEPs.\n"; } } while (false); | |||
1767 | for (unsigned i = 0, e = 2; i < e; ++i) { | |||
1768 | ValueList Operands; | |||
1769 | // Prepare the operand vector. | |||
1770 | for (Value *j : VL) | |||
1771 | Operands.push_back(cast<Instruction>(j)->getOperand(i)); | |||
1772 | ||||
1773 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1774 | } | |||
1775 | return; | |||
1776 | } | |||
1777 | case Instruction::Store: { | |||
1778 | // Check if the stores are consecutive or of we need to swizzle them. | |||
1779 | for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) | |||
1780 | if (!isConsecutiveAccess(VL[i], VL[i + 1], *DL, *SE)) { | |||
1781 | BS.cancelScheduling(VL, VL0); | |||
1782 | newTreeEntry(VL, false, UserTreeIdx); | |||
1783 | DEBUG(dbgs() << "SLP: Non-consecutive store.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-consecutive store.\n"; } } while (false); | |||
1784 | return; | |||
1785 | } | |||
1786 | ||||
1787 | newTreeEntry(VL, true, UserTreeIdx); | |||
1788 | DEBUG(dbgs() << "SLP: added a vector of stores.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of stores.\n" ; } } while (false); | |||
1789 | ||||
1790 | ValueList Operands; | |||
1791 | for (Value *j : VL) | |||
1792 | Operands.push_back(cast<Instruction>(j)->getOperand(0)); | |||
1793 | ||||
1794 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1795 | return; | |||
1796 | } | |||
1797 | case Instruction::Call: { | |||
1798 | // Check if the calls are all to the same vectorizable intrinsic. | |||
1799 | CallInst *CI = cast<CallInst>(VL0); | |||
1800 | // Check if this is an Intrinsic call or something that can be | |||
1801 | // represented by an intrinsic call | |||
1802 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
1803 | if (!isTriviallyVectorizable(ID)) { | |||
1804 | BS.cancelScheduling(VL, VL0); | |||
1805 | newTreeEntry(VL, false, UserTreeIdx); | |||
1806 | DEBUG(dbgs() << "SLP: Non-vectorizable call.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-vectorizable call.\n"; } } while (false); | |||
1807 | return; | |||
1808 | } | |||
1809 | Function *Int = CI->getCalledFunction(); | |||
1810 | Value *A1I = nullptr; | |||
1811 | if (hasVectorInstrinsicScalarOpd(ID, 1)) | |||
1812 | A1I = CI->getArgOperand(1); | |||
1813 | for (unsigned i = 1, e = VL.size(); i != e; ++i) { | |||
1814 | CallInst *CI2 = dyn_cast<CallInst>(VL[i]); | |||
1815 | if (!CI2 || CI2->getCalledFunction() != Int || | |||
1816 | getVectorIntrinsicIDForCall(CI2, TLI) != ID || | |||
1817 | !CI->hasIdenticalOperandBundleSchema(*CI2)) { | |||
1818 | BS.cancelScheduling(VL, VL0); | |||
1819 | newTreeEntry(VL, false, UserTreeIdx); | |||
1820 | DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *VL[i]do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched calls:" << * CI << "!=" << *VL[i] << "\n"; } } while (false ) | |||
1821 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched calls:" << * CI << "!=" << *VL[i] << "\n"; } } while (false ); | |||
1822 | return; | |||
1823 | } | |||
1824 | // ctlz,cttz and powi are special intrinsics whose second argument | |||
1825 | // should be same in order for them to be vectorized. | |||
1826 | if (hasVectorInstrinsicScalarOpd(ID, 1)) { | |||
1827 | Value *A1J = CI2->getArgOperand(1); | |||
1828 | if (A1I != A1J) { | |||
1829 | BS.cancelScheduling(VL, VL0); | |||
1830 | newTreeEntry(VL, false, UserTreeIdx); | |||
1831 | DEBUG(dbgs() << "SLP: mismatched arguments in call:" << *CIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument "<< A1I<<"!=" << A1J << "\n"; } } while (false) | |||
1832 | << " argument "<< A1I<<"!=" << A1Jdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument "<< A1I<<"!=" << A1J << "\n"; } } while (false) | |||
1833 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument "<< A1I<<"!=" << A1J << "\n"; } } while (false); | |||
1834 | return; | |||
1835 | } | |||
1836 | } | |||
1837 | // Verify that the bundle operands are identical between the two calls. | |||
1838 | if (CI->hasOperandBundles() && | |||
1839 | !std::equal(CI->op_begin() + CI->getBundleOperandsStartIndex(), | |||
1840 | CI->op_begin() + CI->getBundleOperandsEndIndex(), | |||
1841 | CI2->op_begin() + CI2->getBundleOperandsStartIndex())) { | |||
1842 | BS.cancelScheduling(VL, VL0); | |||
1843 | newTreeEntry(VL, false, UserTreeIdx); | |||
1844 | DEBUG(dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!="do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!=" << *VL[i] << '\n'; } } while (false) | |||
1845 | << *VL[i] << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!=" << *VL[i] << '\n'; } } while (false); | |||
1846 | return; | |||
1847 | } | |||
1848 | } | |||
1849 | ||||
1850 | newTreeEntry(VL, true, UserTreeIdx); | |||
1851 | for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) { | |||
1852 | ValueList Operands; | |||
1853 | // Prepare the operand vector. | |||
1854 | for (Value *j : VL) { | |||
1855 | CallInst *CI2 = dyn_cast<CallInst>(j); | |||
1856 | Operands.push_back(CI2->getArgOperand(i)); | |||
1857 | } | |||
1858 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1859 | } | |||
1860 | return; | |||
1861 | } | |||
1862 | case Instruction::ShuffleVector: | |||
1863 | // If this is not an alternate sequence of opcode like add-sub | |||
1864 | // then do not vectorize this instruction. | |||
1865 | if (!S.IsAltShuffle) { | |||
1866 | BS.cancelScheduling(VL, VL0); | |||
1867 | newTreeEntry(VL, false, UserTreeIdx); | |||
1868 | DEBUG(dbgs() << "SLP: ShuffleVector are not vectorized.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: ShuffleVector are not vectorized.\n" ; } } while (false); | |||
1869 | return; | |||
1870 | } | |||
1871 | newTreeEntry(VL, true, UserTreeIdx); | |||
1872 | DEBUG(dbgs() << "SLP: added a ShuffleVector op.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a ShuffleVector op.\n" ; } } while (false); | |||
1873 | ||||
1874 | // Reorder operands if reordering would enable vectorization. | |||
1875 | if (isa<BinaryOperator>(VL0)) { | |||
1876 | ValueList Left, Right; | |||
1877 | reorderAltShuffleOperands(S.Opcode, VL, Left, Right); | |||
1878 | buildTree_rec(Left, Depth + 1, UserTreeIdx); | |||
1879 | buildTree_rec(Right, Depth + 1, UserTreeIdx); | |||
1880 | return; | |||
1881 | } | |||
1882 | ||||
1883 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | |||
1884 | ValueList Operands; | |||
1885 | // Prepare the operand vector. | |||
1886 | for (Value *j : VL) | |||
1887 | Operands.push_back(cast<Instruction>(j)->getOperand(i)); | |||
1888 | ||||
1889 | buildTree_rec(Operands, Depth + 1, UserTreeIdx); | |||
1890 | } | |||
1891 | return; | |||
1892 | ||||
1893 | default: | |||
1894 | BS.cancelScheduling(VL, VL0); | |||
1895 | newTreeEntry(VL, false, UserTreeIdx); | |||
1896 | DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering unknown instruction.\n" ; } } while (false); | |||
1897 | return; | |||
1898 | } | |||
1899 | } | |||
1900 | ||||
1901 | unsigned BoUpSLP::canMapToVector(Type *T, const DataLayout &DL) const { | |||
1902 | unsigned N; | |||
1903 | Type *EltTy; | |||
1904 | auto *ST = dyn_cast<StructType>(T); | |||
1905 | if (ST) { | |||
1906 | N = ST->getNumElements(); | |||
1907 | EltTy = *ST->element_begin(); | |||
1908 | } else { | |||
1909 | N = cast<ArrayType>(T)->getNumElements(); | |||
1910 | EltTy = cast<ArrayType>(T)->getElementType(); | |||
1911 | } | |||
1912 | if (!isValidElementType(EltTy)) | |||
1913 | return 0; | |||
1914 | uint64_t VTSize = DL.getTypeStoreSizeInBits(VectorType::get(EltTy, N)); | |||
1915 | if (VTSize < MinVecRegSize || VTSize > MaxVecRegSize || VTSize != DL.getTypeStoreSizeInBits(T)) | |||
1916 | return 0; | |||
1917 | if (ST) { | |||
1918 | // Check that struct is homogeneous. | |||
1919 | for (const auto *Ty : ST->elements()) | |||
1920 | if (Ty != EltTy) | |||
1921 | return 0; | |||
1922 | } | |||
1923 | return N; | |||
1924 | } | |||
1925 | ||||
1926 | bool BoUpSLP::canReuseExtract(ArrayRef<Value *> VL, Value *OpValue) const { | |||
1927 | Instruction *E0 = cast<Instruction>(OpValue); | |||
1928 | assert(E0->getOpcode() == Instruction::ExtractElement ||(static_cast <bool> (E0->getOpcode() == Instruction:: ExtractElement || E0->getOpcode() == Instruction::ExtractValue ) ? void (0) : __assert_fail ("E0->getOpcode() == Instruction::ExtractElement || E0->getOpcode() == Instruction::ExtractValue" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1929, __extension__ __PRETTY_FUNCTION__)) | |||
1929 | E0->getOpcode() == Instruction::ExtractValue)(static_cast <bool> (E0->getOpcode() == Instruction:: ExtractElement || E0->getOpcode() == Instruction::ExtractValue ) ? void (0) : __assert_fail ("E0->getOpcode() == Instruction::ExtractElement || E0->getOpcode() == Instruction::ExtractValue" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1929, __extension__ __PRETTY_FUNCTION__)); | |||
1930 | assert(E0->getOpcode() == getSameOpcode(VL).Opcode && "Invalid opcode")(static_cast <bool> (E0->getOpcode() == getSameOpcode (VL).Opcode && "Invalid opcode") ? void (0) : __assert_fail ("E0->getOpcode() == getSameOpcode(VL).Opcode && \"Invalid opcode\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1930, __extension__ __PRETTY_FUNCTION__)); | |||
1931 | // Check if all of the extracts come from the same vector and from the | |||
1932 | // correct offset. | |||
1933 | Value *Vec = E0->getOperand(0); | |||
1934 | ||||
1935 | // We have to extract from a vector/aggregate with the same number of elements. | |||
1936 | unsigned NElts; | |||
1937 | if (E0->getOpcode() == Instruction::ExtractValue) { | |||
1938 | const DataLayout &DL = E0->getModule()->getDataLayout(); | |||
1939 | NElts = canMapToVector(Vec->getType(), DL); | |||
1940 | if (!NElts) | |||
1941 | return false; | |||
1942 | // Check if load can be rewritten as load of vector. | |||
1943 | LoadInst *LI = dyn_cast<LoadInst>(Vec); | |||
1944 | if (!LI || !LI->isSimple() || !LI->hasNUses(VL.size())) | |||
1945 | return false; | |||
1946 | } else { | |||
1947 | NElts = Vec->getType()->getVectorNumElements(); | |||
1948 | } | |||
1949 | ||||
1950 | if (NElts != VL.size()) | |||
1951 | return false; | |||
1952 | ||||
1953 | // Check that all of the indices extract from the correct offset. | |||
1954 | for (unsigned I = 0, E = VL.size(); I < E; ++I) { | |||
1955 | Instruction *Inst = cast<Instruction>(VL[I]); | |||
1956 | if (!matchExtractIndex(Inst, I, Inst->getOpcode())) | |||
1957 | return false; | |||
1958 | if (Inst->getOperand(0) != Vec) | |||
1959 | return false; | |||
1960 | } | |||
1961 | ||||
1962 | return true; | |||
1963 | } | |||
1964 | ||||
1965 | bool BoUpSLP::areAllUsersVectorized(Instruction *I) const { | |||
1966 | return I->hasOneUse() || | |||
1967 | std::all_of(I->user_begin(), I->user_end(), [this](User *U) { | |||
1968 | return ScalarToTreeEntry.count(U) > 0; | |||
1969 | }); | |||
1970 | } | |||
1971 | ||||
1972 | int BoUpSLP::getEntryCost(TreeEntry *E) { | |||
1973 | ArrayRef<Value*> VL = E->Scalars; | |||
1974 | ||||
1975 | Type *ScalarTy = VL[0]->getType(); | |||
1976 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) | |||
1977 | ScalarTy = SI->getValueOperand()->getType(); | |||
1978 | else if (CmpInst *CI = dyn_cast<CmpInst>(VL[0])) | |||
1979 | ScalarTy = CI->getOperand(0)->getType(); | |||
1980 | VectorType *VecTy = VectorType::get(ScalarTy, VL.size()); | |||
1981 | ||||
1982 | // If we have computed a smaller type for the expression, update VecTy so | |||
1983 | // that the costs will be accurate. | |||
1984 | if (MinBWs.count(VL[0])) | |||
1985 | VecTy = VectorType::get( | |||
1986 | IntegerType::get(F->getContext(), MinBWs[VL[0]].first), VL.size()); | |||
1987 | ||||
1988 | if (E->NeedToGather) { | |||
1989 | if (allConstant(VL)) | |||
1990 | return 0; | |||
1991 | if (isSplat(VL)) { | |||
1992 | return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0); | |||
1993 | } | |||
1994 | if (getSameOpcode(VL).Opcode == Instruction::ExtractElement) { | |||
1995 | Optional<TargetTransformInfo::ShuffleKind> ShuffleKind = isShuffle(VL); | |||
1996 | if (ShuffleKind.hasValue()) { | |||
1997 | int Cost = TTI->getShuffleCost(ShuffleKind.getValue(), VecTy); | |||
1998 | for (auto *V : VL) { | |||
1999 | // If all users of instruction are going to be vectorized and this | |||
2000 | // instruction itself is not going to be vectorized, consider this | |||
2001 | // instruction as dead and remove its cost from the final cost of the | |||
2002 | // vectorized tree. | |||
2003 | if (areAllUsersVectorized(cast<Instruction>(V)) && | |||
2004 | !ScalarToTreeEntry.count(V)) { | |||
2005 | auto *IO = cast<ConstantInt>( | |||
2006 | cast<ExtractElementInst>(V)->getIndexOperand()); | |||
2007 | Cost -= TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, | |||
2008 | IO->getZExtValue()); | |||
2009 | } | |||
2010 | } | |||
2011 | return Cost; | |||
2012 | } | |||
2013 | } | |||
2014 | return getGatherCost(E->Scalars); | |||
2015 | } | |||
2016 | InstructionsState S = getSameOpcode(VL); | |||
2017 | assert(S.Opcode && allSameType(VL) && allSameBlock(VL) && "Invalid VL")(static_cast <bool> (S.Opcode && allSameType(VL ) && allSameBlock(VL) && "Invalid VL") ? void (0) : __assert_fail ("S.Opcode && allSameType(VL) && allSameBlock(VL) && \"Invalid VL\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2017, __extension__ __PRETTY_FUNCTION__)); | |||
2018 | Instruction *VL0 = cast<Instruction>(S.OpValue); | |||
2019 | unsigned ShuffleOrOp = S.IsAltShuffle ? | |||
2020 | (unsigned) Instruction::ShuffleVector : S.Opcode; | |||
2021 | switch (ShuffleOrOp) { | |||
2022 | case Instruction::PHI: | |||
2023 | return 0; | |||
2024 | ||||
2025 | case Instruction::ExtractValue: | |||
2026 | case Instruction::ExtractElement: | |||
2027 | if (canReuseExtract(VL, S.OpValue)) { | |||
2028 | int DeadCost = 0; | |||
2029 | for (unsigned i = 0, e = VL.size(); i < e; ++i) { | |||
2030 | Instruction *E = cast<Instruction>(VL[i]); | |||
2031 | // If all users are going to be vectorized, instruction can be | |||
2032 | // considered as dead. | |||
2033 | // The same, if have only one user, it will be vectorized for sure. | |||
2034 | if (areAllUsersVectorized(E)) | |||
2035 | // Take credit for instruction that will become dead. | |||
2036 | DeadCost += | |||
2037 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, i); | |||
2038 | } | |||
2039 | return -DeadCost; | |||
2040 | } | |||
2041 | return getGatherCost(VecTy); | |||
2042 | ||||
2043 | case Instruction::ZExt: | |||
2044 | case Instruction::SExt: | |||
2045 | case Instruction::FPToUI: | |||
2046 | case Instruction::FPToSI: | |||
2047 | case Instruction::FPExt: | |||
2048 | case Instruction::PtrToInt: | |||
2049 | case Instruction::IntToPtr: | |||
2050 | case Instruction::SIToFP: | |||
2051 | case Instruction::UIToFP: | |||
2052 | case Instruction::Trunc: | |||
2053 | case Instruction::FPTrunc: | |||
2054 | case Instruction::BitCast: { | |||
2055 | Type *SrcTy = VL0->getOperand(0)->getType(); | |||
2056 | ||||
2057 | // Calculate the cost of this instruction. | |||
2058 | int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(), | |||
2059 | VL0->getType(), SrcTy, VL0); | |||
2060 | ||||
2061 | VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size()); | |||
2062 | int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy, VL0); | |||
2063 | return VecCost - ScalarCost; | |||
2064 | } | |||
2065 | case Instruction::FCmp: | |||
2066 | case Instruction::ICmp: | |||
2067 | case Instruction::Select: { | |||
2068 | // Calculate the cost of this instruction. | |||
2069 | VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size()); | |||
2070 | int ScalarCost = VecTy->getNumElements() * | |||
2071 | TTI->getCmpSelInstrCost(S.Opcode, ScalarTy, Builder.getInt1Ty(), VL0); | |||
2072 | int VecCost = TTI->getCmpSelInstrCost(S.Opcode, VecTy, MaskTy, VL0); | |||
2073 | return VecCost - ScalarCost; | |||
2074 | } | |||
2075 | case Instruction::Add: | |||
2076 | case Instruction::FAdd: | |||
2077 | case Instruction::Sub: | |||
2078 | case Instruction::FSub: | |||
2079 | case Instruction::Mul: | |||
2080 | case Instruction::FMul: | |||
2081 | case Instruction::UDiv: | |||
2082 | case Instruction::SDiv: | |||
2083 | case Instruction::FDiv: | |||
2084 | case Instruction::URem: | |||
2085 | case Instruction::SRem: | |||
2086 | case Instruction::FRem: | |||
2087 | case Instruction::Shl: | |||
2088 | case Instruction::LShr: | |||
2089 | case Instruction::AShr: | |||
2090 | case Instruction::And: | |||
2091 | case Instruction::Or: | |||
2092 | case Instruction::Xor: { | |||
2093 | // Certain instructions can be cheaper to vectorize if they have a | |||
2094 | // constant second vector operand. | |||
2095 | TargetTransformInfo::OperandValueKind Op1VK = | |||
2096 | TargetTransformInfo::OK_AnyValue; | |||
2097 | TargetTransformInfo::OperandValueKind Op2VK = | |||
2098 | TargetTransformInfo::OK_UniformConstantValue; | |||
2099 | TargetTransformInfo::OperandValueProperties Op1VP = | |||
2100 | TargetTransformInfo::OP_None; | |||
2101 | TargetTransformInfo::OperandValueProperties Op2VP = | |||
2102 | TargetTransformInfo::OP_None; | |||
2103 | ||||
2104 | // If all operands are exactly the same ConstantInt then set the | |||
2105 | // operand kind to OK_UniformConstantValue. | |||
2106 | // If instead not all operands are constants, then set the operand kind | |||
2107 | // to OK_AnyValue. If all operands are constants but not the same, | |||
2108 | // then set the operand kind to OK_NonUniformConstantValue. | |||
2109 | ConstantInt *CInt = nullptr; | |||
2110 | for (unsigned i = 0; i < VL.size(); ++i) { | |||
2111 | const Instruction *I = cast<Instruction>(VL[i]); | |||
2112 | if (!isa<ConstantInt>(I->getOperand(1))) { | |||
2113 | Op2VK = TargetTransformInfo::OK_AnyValue; | |||
2114 | break; | |||
2115 | } | |||
2116 | if (i == 0) { | |||
2117 | CInt = cast<ConstantInt>(I->getOperand(1)); | |||
2118 | continue; | |||
2119 | } | |||
2120 | if (Op2VK == TargetTransformInfo::OK_UniformConstantValue && | |||
2121 | CInt != cast<ConstantInt>(I->getOperand(1))) | |||
2122 | Op2VK = TargetTransformInfo::OK_NonUniformConstantValue; | |||
2123 | } | |||
2124 | // FIXME: Currently cost of model modification for division by power of | |||
2125 | // 2 is handled for X86 and AArch64. Add support for other targets. | |||
2126 | if (Op2VK == TargetTransformInfo::OK_UniformConstantValue && CInt && | |||
2127 | CInt->getValue().isPowerOf2()) | |||
2128 | Op2VP = TargetTransformInfo::OP_PowerOf2; | |||
2129 | ||||
2130 | SmallVector<const Value *, 4> Operands(VL0->operand_values()); | |||
2131 | int ScalarCost = | |||
2132 | VecTy->getNumElements() * | |||
2133 | TTI->getArithmeticInstrCost(S.Opcode, ScalarTy, Op1VK, Op2VK, Op1VP, | |||
2134 | Op2VP, Operands); | |||
2135 | int VecCost = TTI->getArithmeticInstrCost(S.Opcode, VecTy, Op1VK, Op2VK, | |||
2136 | Op1VP, Op2VP, Operands); | |||
2137 | return VecCost - ScalarCost; | |||
2138 | } | |||
2139 | case Instruction::GetElementPtr: { | |||
2140 | TargetTransformInfo::OperandValueKind Op1VK = | |||
2141 | TargetTransformInfo::OK_AnyValue; | |||
2142 | TargetTransformInfo::OperandValueKind Op2VK = | |||
2143 | TargetTransformInfo::OK_UniformConstantValue; | |||
2144 | ||||
2145 | int ScalarCost = | |||
2146 | VecTy->getNumElements() * | |||
2147 | TTI->getArithmeticInstrCost(Instruction::Add, ScalarTy, Op1VK, Op2VK); | |||
2148 | int VecCost = | |||
2149 | TTI->getArithmeticInstrCost(Instruction::Add, VecTy, Op1VK, Op2VK); | |||
2150 | ||||
2151 | return VecCost - ScalarCost; | |||
2152 | } | |||
2153 | case Instruction::Load: { | |||
2154 | // Cost of wide load - cost of scalar loads. | |||
2155 | unsigned alignment = dyn_cast<LoadInst>(VL0)->getAlignment(); | |||
2156 | int ScalarLdCost = VecTy->getNumElements() * | |||
2157 | TTI->getMemoryOpCost(Instruction::Load, ScalarTy, alignment, 0, VL0); | |||
2158 | int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, | |||
2159 | VecTy, alignment, 0, VL0); | |||
2160 | return VecLdCost - ScalarLdCost; | |||
2161 | } | |||
2162 | case Instruction::Store: { | |||
2163 | // We know that we can merge the stores. Calculate the cost. | |||
2164 | unsigned alignment = dyn_cast<StoreInst>(VL0)->getAlignment(); | |||
2165 | int ScalarStCost = VecTy->getNumElements() * | |||
2166 | TTI->getMemoryOpCost(Instruction::Store, ScalarTy, alignment, 0, VL0); | |||
2167 | int VecStCost = TTI->getMemoryOpCost(Instruction::Store, | |||
2168 | VecTy, alignment, 0, VL0); | |||
2169 | return VecStCost - ScalarStCost; | |||
2170 | } | |||
2171 | case Instruction::Call: { | |||
2172 | CallInst *CI = cast<CallInst>(VL0); | |||
2173 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
2174 | ||||
2175 | // Calculate the cost of the scalar and vector calls. | |||
2176 | SmallVector<Type*, 4> ScalarTys; | |||
2177 | for (unsigned op = 0, opc = CI->getNumArgOperands(); op!= opc; ++op) | |||
2178 | ScalarTys.push_back(CI->getArgOperand(op)->getType()); | |||
2179 | ||||
2180 | FastMathFlags FMF; | |||
2181 | if (auto *FPMO = dyn_cast<FPMathOperator>(CI)) | |||
2182 | FMF = FPMO->getFastMathFlags(); | |||
2183 | ||||
2184 | int ScalarCallCost = VecTy->getNumElements() * | |||
2185 | TTI->getIntrinsicInstrCost(ID, ScalarTy, ScalarTys, FMF); | |||
2186 | ||||
2187 | SmallVector<Value *, 4> Args(CI->arg_operands()); | |||
2188 | int VecCallCost = TTI->getIntrinsicInstrCost(ID, CI->getType(), Args, FMF, | |||
2189 | VecTy->getNumElements()); | |||
2190 | ||||
2191 | DEBUG(dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false) | |||
2192 | << " (" << VecCallCost << "-" << ScalarCallCost << ")"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false) | |||
2193 | << " for " << *CI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false); | |||
2194 | ||||
2195 | return VecCallCost - ScalarCallCost; | |||
2196 | } | |||
2197 | case Instruction::ShuffleVector: { | |||
2198 | TargetTransformInfo::OperandValueKind Op1VK = | |||
2199 | TargetTransformInfo::OK_AnyValue; | |||
2200 | TargetTransformInfo::OperandValueKind Op2VK = | |||
2201 | TargetTransformInfo::OK_AnyValue; | |||
2202 | int ScalarCost = 0; | |||
2203 | int VecCost = 0; | |||
2204 | for (Value *i : VL) { | |||
2205 | Instruction *I = cast<Instruction>(i); | |||
2206 | if (!I) | |||
2207 | break; | |||
2208 | ScalarCost += | |||
2209 | TTI->getArithmeticInstrCost(I->getOpcode(), ScalarTy, Op1VK, Op2VK); | |||
2210 | } | |||
2211 | // VecCost is equal to sum of the cost of creating 2 vectors | |||
2212 | // and the cost of creating shuffle. | |||
2213 | Instruction *I0 = cast<Instruction>(VL[0]); | |||
2214 | VecCost = | |||
2215 | TTI->getArithmeticInstrCost(I0->getOpcode(), VecTy, Op1VK, Op2VK); | |||
2216 | Instruction *I1 = cast<Instruction>(VL[1]); | |||
2217 | VecCost += | |||
2218 | TTI->getArithmeticInstrCost(I1->getOpcode(), VecTy, Op1VK, Op2VK); | |||
2219 | VecCost += | |||
2220 | TTI->getShuffleCost(TargetTransformInfo::SK_Alternate, VecTy, 0); | |||
2221 | return VecCost - ScalarCost; | |||
2222 | } | |||
2223 | default: | |||
2224 | llvm_unreachable("Unknown instruction")::llvm::llvm_unreachable_internal("Unknown instruction", "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2224); | |||
2225 | } | |||
2226 | } | |||
2227 | ||||
2228 | bool BoUpSLP::isFullyVectorizableTinyTree() { | |||
2229 | DEBUG(dbgs() << "SLP: Check whether the tree with height " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Check whether the tree with height " << VectorizableTree.size() << " is fully vectorizable .\n" ; } } while (false) | |||
2230 | VectorizableTree.size() << " is fully vectorizable .\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Check whether the tree with height " << VectorizableTree.size() << " is fully vectorizable .\n" ; } } while (false); | |||
2231 | ||||
2232 | // We only handle trees of heights 1 and 2. | |||
2233 | if (VectorizableTree.size() == 1 && !VectorizableTree[0].NeedToGather) | |||
2234 | return true; | |||
2235 | ||||
2236 | if (VectorizableTree.size() != 2) | |||
2237 | return false; | |||
2238 | ||||
2239 | // Handle splat and all-constants stores. | |||
2240 | if (!VectorizableTree[0].NeedToGather && | |||
2241 | (allConstant(VectorizableTree[1].Scalars) || | |||
2242 | isSplat(VectorizableTree[1].Scalars))) | |||
2243 | return true; | |||
2244 | ||||
2245 | // Gathering cost would be too much for tiny trees. | |||
2246 | if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather) | |||
2247 | return false; | |||
2248 | ||||
2249 | return true; | |||
2250 | } | |||
2251 | ||||
2252 | bool BoUpSLP::isTreeTinyAndNotFullyVectorizable() { | |||
2253 | // We can vectorize the tree if its size is greater than or equal to the | |||
2254 | // minimum size specified by the MinTreeSize command line option. | |||
2255 | if (VectorizableTree.size() >= MinTreeSize) | |||
2256 | return false; | |||
2257 | ||||
2258 | // If we have a tiny tree (a tree whose size is less than MinTreeSize), we | |||
2259 | // can vectorize it if we can prove it fully vectorizable. | |||
2260 | if (isFullyVectorizableTinyTree()) | |||
2261 | return false; | |||
2262 | ||||
2263 | assert(VectorizableTree.empty()(static_cast <bool> (VectorizableTree.empty() ? ExternalUses .empty() : true && "We shouldn't have any external users" ) ? void (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2265, __extension__ __PRETTY_FUNCTION__)) | |||
2264 | ? ExternalUses.empty()(static_cast <bool> (VectorizableTree.empty() ? ExternalUses .empty() : true && "We shouldn't have any external users" ) ? void (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2265, __extension__ __PRETTY_FUNCTION__)) | |||
2265 | : true && "We shouldn't have any external users")(static_cast <bool> (VectorizableTree.empty() ? ExternalUses .empty() : true && "We shouldn't have any external users" ) ? void (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2265, __extension__ __PRETTY_FUNCTION__)); | |||
2266 | ||||
2267 | // Otherwise, we can't vectorize the tree. It is both tiny and not fully | |||
2268 | // vectorizable. | |||
2269 | return true; | |||
2270 | } | |||
2271 | ||||
2272 | int BoUpSLP::getSpillCost() { | |||
2273 | // Walk from the bottom of the tree to the top, tracking which values are | |||
2274 | // live. When we see a call instruction that is not part of our tree, | |||
2275 | // query TTI to see if there is a cost to keeping values live over it | |||
2276 | // (for example, if spills and fills are required). | |||
2277 | unsigned BundleWidth = VectorizableTree.front().Scalars.size(); | |||
2278 | int Cost = 0; | |||
2279 | ||||
2280 | SmallPtrSet<Instruction*, 4> LiveValues; | |||
2281 | Instruction *PrevInst = nullptr; | |||
2282 | ||||
2283 | for (const auto &N : VectorizableTree) { | |||
2284 | Instruction *Inst = dyn_cast<Instruction>(N.Scalars[0]); | |||
2285 | if (!Inst) | |||
2286 | continue; | |||
2287 | ||||
2288 | if (!PrevInst) { | |||
2289 | PrevInst = Inst; | |||
2290 | continue; | |||
2291 | } | |||
2292 | ||||
2293 | // Update LiveValues. | |||
2294 | LiveValues.erase(PrevInst); | |||
2295 | for (auto &J : PrevInst->operands()) { | |||
2296 | if (isa<Instruction>(&*J) && getTreeEntry(&*J)) | |||
2297 | LiveValues.insert(cast<Instruction>(&*J)); | |||
2298 | } | |||
2299 | ||||
2300 | DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: #LV: " << LiveValues.size (); for (auto *X : LiveValues) dbgs() << " " << X ->getName(); dbgs() << ", Looking at "; Inst->dump ();; } } while (false) | |||
2301 | dbgs() << "SLP: #LV: " << LiveValues.size();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: #LV: " << LiveValues.size (); for (auto *X : LiveValues) dbgs() << " " << X ->getName(); dbgs() << ", Looking at "; Inst->dump ();; } } while (false) | |||
2302 | for (auto *X : LiveValues)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: #LV: " << LiveValues.size (); for (auto *X : LiveValues) dbgs() << " " << X ->getName(); dbgs() << ", Looking at "; Inst->dump ();; } } while (false) | |||
2303 | dbgs() << " " << X->getName();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: #LV: " << LiveValues.size (); for (auto *X : LiveValues) dbgs() << " " << X ->getName(); dbgs() << ", Looking at "; Inst->dump ();; } } while (false) | |||
2304 | dbgs() << ", Looking at ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: #LV: " << LiveValues.size (); for (auto *X : LiveValues) dbgs() << " " << X ->getName(); dbgs() << ", Looking at "; Inst->dump ();; } } while (false) | |||
2305 | Inst->dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: #LV: " << LiveValues.size (); for (auto *X : LiveValues) dbgs() << " " << X ->getName(); dbgs() << ", Looking at "; Inst->dump ();; } } while (false) | |||
2306 | )do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: #LV: " << LiveValues.size (); for (auto *X : LiveValues) dbgs() << " " << X ->getName(); dbgs() << ", Looking at "; Inst->dump ();; } } while (false); | |||
2307 | ||||
2308 | // Now find the sequence of instructions between PrevInst and Inst. | |||
2309 | BasicBlock::reverse_iterator InstIt = ++Inst->getIterator().getReverse(), | |||
2310 | PrevInstIt = | |||
2311 | PrevInst->getIterator().getReverse(); | |||
2312 | while (InstIt != PrevInstIt) { | |||
2313 | if (PrevInstIt == PrevInst->getParent()->rend()) { | |||
2314 | PrevInstIt = Inst->getParent()->rbegin(); | |||
2315 | continue; | |||
2316 | } | |||
2317 | ||||
2318 | if (isa<CallInst>(&*PrevInstIt) && &*PrevInstIt != PrevInst) { | |||
2319 | SmallVector<Type*, 4> V; | |||
2320 | for (auto *II : LiveValues) | |||
2321 | V.push_back(VectorType::get(II->getType(), BundleWidth)); | |||
2322 | Cost += TTI->getCostOfKeepingLiveOverCall(V); | |||
2323 | } | |||
2324 | ||||
2325 | ++PrevInstIt; | |||
2326 | } | |||
2327 | ||||
2328 | PrevInst = Inst; | |||
2329 | } | |||
2330 | ||||
2331 | return Cost; | |||
2332 | } | |||
2333 | ||||
2334 | int BoUpSLP::getTreeCost() { | |||
2335 | int Cost = 0; | |||
2336 | DEBUG(dbgs() << "SLP: Calculating cost for tree of size " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Calculating cost for tree of size " << VectorizableTree.size() << ".\n"; } } while ( false) | |||
2337 | VectorizableTree.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Calculating cost for tree of size " << VectorizableTree.size() << ".\n"; } } while ( false); | |||
2338 | ||||
2339 | unsigned BundleWidth = VectorizableTree[0].Scalars.size(); | |||
2340 | ||||
2341 | for (TreeEntry &TE : VectorizableTree) { | |||
2342 | int C = getEntryCost(&TE); | |||
2343 | DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " << *TE.Scalars[0] << ".\n"; } } while (false) | |||
2344 | << *TE.Scalars[0] << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " << *TE.Scalars[0] << ".\n"; } } while (false); | |||
2345 | Cost += C; | |||
2346 | } | |||
2347 | ||||
2348 | SmallSet<Value *, 16> ExtractCostCalculated; | |||
2349 | int ExtractCost = 0; | |||
2350 | for (ExternalUser &EU : ExternalUses) { | |||
2351 | // We only add extract cost once for the same scalar. | |||
2352 | if (!ExtractCostCalculated.insert(EU.Scalar).second) | |||
2353 | continue; | |||
2354 | ||||
2355 | // Uses by ephemeral values are free (because the ephemeral value will be | |||
2356 | // removed prior to code generation, and so the extraction will be | |||
2357 | // removed as well). | |||
2358 | if (EphValues.count(EU.User)) | |||
2359 | continue; | |||
2360 | ||||
2361 | // If we plan to rewrite the tree in a smaller type, we will need to sign | |||
2362 | // extend the extracted value back to the original type. Here, we account | |||
2363 | // for the extract and the added cost of the sign extend if needed. | |||
2364 | auto *VecTy = VectorType::get(EU.Scalar->getType(), BundleWidth); | |||
2365 | auto *ScalarRoot = VectorizableTree[0].Scalars[0]; | |||
2366 | if (MinBWs.count(ScalarRoot)) { | |||
2367 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); | |||
2368 | auto Extend = | |||
2369 | MinBWs[ScalarRoot].second ? Instruction::SExt : Instruction::ZExt; | |||
2370 | VecTy = VectorType::get(MinTy, BundleWidth); | |||
2371 | ExtractCost += TTI->getExtractWithExtendCost(Extend, EU.Scalar->getType(), | |||
2372 | VecTy, EU.Lane); | |||
2373 | } else { | |||
2374 | ExtractCost += | |||
2375 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, EU.Lane); | |||
2376 | } | |||
2377 | } | |||
2378 | ||||
2379 | int SpillCost = getSpillCost(); | |||
2380 | Cost += SpillCost + ExtractCost; | |||
2381 | ||||
2382 | std::string Str; | |||
2383 | { | |||
2384 | raw_string_ostream OS(Str); | |||
2385 | OS << "SLP: Spill Cost = " << SpillCost << ".\n" | |||
2386 | << "SLP: Extract Cost = " << ExtractCost << ".\n" | |||
2387 | << "SLP: Total Cost = " << Cost << ".\n"; | |||
2388 | } | |||
2389 | DEBUG(dbgs() << Str)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << Str; } } while (false); | |||
2390 | ||||
2391 | if (ViewSLPTree) | |||
2392 | ViewGraph(this, "SLP" + F->getName(), false, Str); | |||
2393 | ||||
2394 | return Cost; | |||
2395 | } | |||
2396 | ||||
2397 | int BoUpSLP::getGatherCost(Type *Ty) { | |||
2398 | int Cost = 0; | |||
2399 | for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i) | |||
2400 | Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i); | |||
2401 | return Cost; | |||
2402 | } | |||
2403 | ||||
2404 | int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) { | |||
2405 | // Find the type of the operands in VL. | |||
2406 | Type *ScalarTy = VL[0]->getType(); | |||
2407 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) | |||
2408 | ScalarTy = SI->getValueOperand()->getType(); | |||
2409 | VectorType *VecTy = VectorType::get(ScalarTy, VL.size()); | |||
2410 | // Find the cost of inserting/extracting values from the vector. | |||
2411 | return getGatherCost(VecTy); | |||
2412 | } | |||
2413 | ||||
2414 | // Reorder commutative operations in alternate shuffle if the resulting vectors | |||
2415 | // are consecutive loads. This would allow us to vectorize the tree. | |||
2416 | // If we have something like- | |||
2417 | // load a[0] - load b[0] | |||
2418 | // load b[1] + load a[1] | |||
2419 | // load a[2] - load b[2] | |||
2420 | // load a[3] + load b[3] | |||
2421 | // Reordering the second load b[1] load a[1] would allow us to vectorize this | |||
2422 | // code. | |||
2423 | void BoUpSLP::reorderAltShuffleOperands(unsigned Opcode, ArrayRef<Value *> VL, | |||
2424 | SmallVectorImpl<Value *> &Left, | |||
2425 | SmallVectorImpl<Value *> &Right) { | |||
2426 | // Push left and right operands of binary operation into Left and Right | |||
2427 | unsigned AltOpcode = getAltOpcode(Opcode); | |||
2428 | (void)AltOpcode; | |||
2429 | for (Value *V : VL) { | |||
2430 | auto *I = cast<Instruction>(V); | |||
2431 | assert(sameOpcodeOrAlt(Opcode, AltOpcode, I->getOpcode()) &&(static_cast <bool> (sameOpcodeOrAlt(Opcode, AltOpcode, I->getOpcode()) && "Incorrect instruction in vector" ) ? void (0) : __assert_fail ("sameOpcodeOrAlt(Opcode, AltOpcode, I->getOpcode()) && \"Incorrect instruction in vector\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2432, __extension__ __PRETTY_FUNCTION__)) | |||
2432 | "Incorrect instruction in vector")(static_cast <bool> (sameOpcodeOrAlt(Opcode, AltOpcode, I->getOpcode()) && "Incorrect instruction in vector" ) ? void (0) : __assert_fail ("sameOpcodeOrAlt(Opcode, AltOpcode, I->getOpcode()) && \"Incorrect instruction in vector\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2432, __extension__ __PRETTY_FUNCTION__)); | |||
2433 | Left.push_back(I->getOperand(0)); | |||
2434 | Right.push_back(I->getOperand(1)); | |||
2435 | } | |||
2436 | ||||
2437 | // Reorder if we have a commutative operation and consecutive access | |||
2438 | // are on either side of the alternate instructions. | |||
2439 | for (unsigned j = 0; j < VL.size() - 1; ++j) { | |||
2440 | if (LoadInst *L = dyn_cast<LoadInst>(Left[j])) { | |||
2441 | if (LoadInst *L1 = dyn_cast<LoadInst>(Right[j + 1])) { | |||
2442 | Instruction *VL1 = cast<Instruction>(VL[j]); | |||
2443 | Instruction *VL2 = cast<Instruction>(VL[j + 1]); | |||
2444 | if (VL1->isCommutative() && isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2445 | std::swap(Left[j], Right[j]); | |||
2446 | continue; | |||
2447 | } else if (VL2->isCommutative() && | |||
2448 | isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2449 | std::swap(Left[j + 1], Right[j + 1]); | |||
2450 | continue; | |||
2451 | } | |||
2452 | // else unchanged | |||
2453 | } | |||
2454 | } | |||
2455 | if (LoadInst *L = dyn_cast<LoadInst>(Right[j])) { | |||
2456 | if (LoadInst *L1 = dyn_cast<LoadInst>(Left[j + 1])) { | |||
2457 | Instruction *VL1 = cast<Instruction>(VL[j]); | |||
2458 | Instruction *VL2 = cast<Instruction>(VL[j + 1]); | |||
2459 | if (VL1->isCommutative() && isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2460 | std::swap(Left[j], Right[j]); | |||
2461 | continue; | |||
2462 | } else if (VL2->isCommutative() && | |||
2463 | isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2464 | std::swap(Left[j + 1], Right[j + 1]); | |||
2465 | continue; | |||
2466 | } | |||
2467 | // else unchanged | |||
2468 | } | |||
2469 | } | |||
2470 | } | |||
2471 | } | |||
2472 | ||||
2473 | // Return true if I should be commuted before adding it's left and right | |||
2474 | // operands to the arrays Left and Right. | |||
2475 | // | |||
2476 | // The vectorizer is trying to either have all elements one side being | |||
2477 | // instruction with the same opcode to enable further vectorization, or having | |||
2478 | // a splat to lower the vectorizing cost. | |||
2479 | static bool shouldReorderOperands( | |||
2480 | int i, unsigned Opcode, Instruction &I, ArrayRef<Value *> Left, | |||
2481 | ArrayRef<Value *> Right, bool AllSameOpcodeLeft, bool AllSameOpcodeRight, | |||
2482 | bool SplatLeft, bool SplatRight, Value *&VLeft, Value *&VRight) { | |||
2483 | VLeft = I.getOperand(0); | |||
2484 | VRight = I.getOperand(1); | |||
2485 | // If we have "SplatRight", try to see if commuting is needed to preserve it. | |||
2486 | if (SplatRight) { | |||
2487 | if (VRight == Right[i - 1]) | |||
2488 | // Preserve SplatRight | |||
2489 | return false; | |||
2490 | if (VLeft == Right[i - 1]) { | |||
2491 | // Commuting would preserve SplatRight, but we don't want to break | |||
2492 | // SplatLeft either, i.e. preserve the original order if possible. | |||
2493 | // (FIXME: why do we care?) | |||
2494 | if (SplatLeft && VLeft == Left[i - 1]) | |||
2495 | return false; | |||
2496 | return true; | |||
2497 | } | |||
2498 | } | |||
2499 | // Symmetrically handle Right side. | |||
2500 | if (SplatLeft) { | |||
2501 | if (VLeft == Left[i - 1]) | |||
2502 | // Preserve SplatLeft | |||
2503 | return false; | |||
2504 | if (VRight == Left[i - 1]) | |||
2505 | return true; | |||
2506 | } | |||
2507 | ||||
2508 | Instruction *ILeft = dyn_cast<Instruction>(VLeft); | |||
2509 | Instruction *IRight = dyn_cast<Instruction>(VRight); | |||
2510 | ||||
2511 | // If we have "AllSameOpcodeRight", try to see if the left operands preserves | |||
2512 | // it and not the right, in this case we want to commute. | |||
2513 | if (AllSameOpcodeRight) { | |||
2514 | unsigned RightPrevOpcode = cast<Instruction>(Right[i - 1])->getOpcode(); | |||
2515 | if (IRight && RightPrevOpcode == IRight->getOpcode()) | |||
2516 | // Do not commute, a match on the right preserves AllSameOpcodeRight | |||
2517 | return false; | |||
2518 | if (ILeft && RightPrevOpcode == ILeft->getOpcode()) { | |||
2519 | // We have a match and may want to commute, but first check if there is | |||
2520 | // not also a match on the existing operands on the Left to preserve | |||
2521 | // AllSameOpcodeLeft, i.e. preserve the original order if possible. | |||
2522 | // (FIXME: why do we care?) | |||
2523 | if (AllSameOpcodeLeft && ILeft && | |||
2524 | cast<Instruction>(Left[i - 1])->getOpcode() == ILeft->getOpcode()) | |||
2525 | return false; | |||
2526 | return true; | |||
2527 | } | |||
2528 | } | |||
2529 | // Symmetrically handle Left side. | |||
2530 | if (AllSameOpcodeLeft) { | |||
2531 | unsigned LeftPrevOpcode = cast<Instruction>(Left[i - 1])->getOpcode(); | |||
2532 | if (ILeft && LeftPrevOpcode == ILeft->getOpcode()) | |||
2533 | return false; | |||
2534 | if (IRight && LeftPrevOpcode == IRight->getOpcode()) | |||
2535 | return true; | |||
2536 | } | |||
2537 | return false; | |||
2538 | } | |||
2539 | ||||
2540 | void BoUpSLP::reorderInputsAccordingToOpcode(unsigned Opcode, | |||
2541 | ArrayRef<Value *> VL, | |||
2542 | SmallVectorImpl<Value *> &Left, | |||
2543 | SmallVectorImpl<Value *> &Right) { | |||
2544 | if (!VL.empty()) { | |||
2545 | // Peel the first iteration out of the loop since there's nothing | |||
2546 | // interesting to do anyway and it simplifies the checks in the loop. | |||
2547 | auto *I = cast<Instruction>(VL[0]); | |||
2548 | Value *VLeft = I->getOperand(0); | |||
2549 | Value *VRight = I->getOperand(1); | |||
2550 | if (!isa<Instruction>(VRight) && isa<Instruction>(VLeft)) | |||
2551 | // Favor having instruction to the right. FIXME: why? | |||
2552 | std::swap(VLeft, VRight); | |||
2553 | Left.push_back(VLeft); | |||
2554 | Right.push_back(VRight); | |||
2555 | } | |||
2556 | ||||
2557 | // Keep track if we have instructions with all the same opcode on one side. | |||
2558 | bool AllSameOpcodeLeft = isa<Instruction>(Left[0]); | |||
2559 | bool AllSameOpcodeRight = isa<Instruction>(Right[0]); | |||
2560 | // Keep track if we have one side with all the same value (broadcast). | |||
2561 | bool SplatLeft = true; | |||
2562 | bool SplatRight = true; | |||
2563 | ||||
2564 | for (unsigned i = 1, e = VL.size(); i != e; ++i) { | |||
2565 | Instruction *I = cast<Instruction>(VL[i]); | |||
2566 | assert(((I->getOpcode() == Opcode && I->isCommutative()) ||(static_cast <bool> (((I->getOpcode() == Opcode && I->isCommutative()) || (I->getOpcode() != Opcode && Instruction::isCommutative(Opcode))) && "Can only process commutative instruction" ) ? void (0) : __assert_fail ("((I->getOpcode() == Opcode && I->isCommutative()) || (I->getOpcode() != Opcode && Instruction::isCommutative(Opcode))) && \"Can only process commutative instruction\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2568, __extension__ __PRETTY_FUNCTION__)) | |||
2567 | (I->getOpcode() != Opcode && Instruction::isCommutative(Opcode))) &&(static_cast <bool> (((I->getOpcode() == Opcode && I->isCommutative()) || (I->getOpcode() != Opcode && Instruction::isCommutative(Opcode))) && "Can only process commutative instruction" ) ? void (0) : __assert_fail ("((I->getOpcode() == Opcode && I->isCommutative()) || (I->getOpcode() != Opcode && Instruction::isCommutative(Opcode))) && \"Can only process commutative instruction\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2568, __extension__ __PRETTY_FUNCTION__)) | |||
2568 | "Can only process commutative instruction")(static_cast <bool> (((I->getOpcode() == Opcode && I->isCommutative()) || (I->getOpcode() != Opcode && Instruction::isCommutative(Opcode))) && "Can only process commutative instruction" ) ? void (0) : __assert_fail ("((I->getOpcode() == Opcode && I->isCommutative()) || (I->getOpcode() != Opcode && Instruction::isCommutative(Opcode))) && \"Can only process commutative instruction\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2568, __extension__ __PRETTY_FUNCTION__)); | |||
2569 | // Commute to favor either a splat or maximizing having the same opcodes on | |||
2570 | // one side. | |||
2571 | Value *VLeft; | |||
2572 | Value *VRight; | |||
2573 | if (shouldReorderOperands(i, Opcode, *I, Left, Right, AllSameOpcodeLeft, | |||
2574 | AllSameOpcodeRight, SplatLeft, SplatRight, VLeft, | |||
2575 | VRight)) { | |||
2576 | Left.push_back(VRight); | |||
2577 | Right.push_back(VLeft); | |||
2578 | } else { | |||
2579 | Left.push_back(VLeft); | |||
2580 | Right.push_back(VRight); | |||
2581 | } | |||
2582 | // Update Splat* and AllSameOpcode* after the insertion. | |||
2583 | SplatRight = SplatRight && (Right[i - 1] == Right[i]); | |||
2584 | SplatLeft = SplatLeft && (Left[i - 1] == Left[i]); | |||
2585 | AllSameOpcodeLeft = AllSameOpcodeLeft && isa<Instruction>(Left[i]) && | |||
2586 | (cast<Instruction>(Left[i - 1])->getOpcode() == | |||
2587 | cast<Instruction>(Left[i])->getOpcode()); | |||
2588 | AllSameOpcodeRight = AllSameOpcodeRight && isa<Instruction>(Right[i]) && | |||
2589 | (cast<Instruction>(Right[i - 1])->getOpcode() == | |||
2590 | cast<Instruction>(Right[i])->getOpcode()); | |||
2591 | } | |||
2592 | ||||
2593 | // If one operand end up being broadcast, return this operand order. | |||
2594 | if (SplatRight || SplatLeft) | |||
2595 | return; | |||
2596 | ||||
2597 | // Finally check if we can get longer vectorizable chain by reordering | |||
2598 | // without breaking the good operand order detected above. | |||
2599 | // E.g. If we have something like- | |||
2600 | // load a[0] load b[0] | |||
2601 | // load b[1] load a[1] | |||
2602 | // load a[2] load b[2] | |||
2603 | // load a[3] load b[3] | |||
2604 | // Reordering the second load b[1] load a[1] would allow us to vectorize | |||
2605 | // this code and we still retain AllSameOpcode property. | |||
2606 | // FIXME: This load reordering might break AllSameOpcode in some rare cases | |||
2607 | // such as- | |||
2608 | // add a[0],c[0] load b[0] | |||
2609 | // add a[1],c[2] load b[1] | |||
2610 | // b[2] load b[2] | |||
2611 | // add a[3],c[3] load b[3] | |||
2612 | for (unsigned j = 0; j < VL.size() - 1; ++j) { | |||
2613 | if (LoadInst *L = dyn_cast<LoadInst>(Left[j])) { | |||
2614 | if (LoadInst *L1 = dyn_cast<LoadInst>(Right[j + 1])) { | |||
2615 | if (isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2616 | std::swap(Left[j + 1], Right[j + 1]); | |||
2617 | continue; | |||
2618 | } | |||
2619 | } | |||
2620 | } | |||
2621 | if (LoadInst *L = dyn_cast<LoadInst>(Right[j])) { | |||
2622 | if (LoadInst *L1 = dyn_cast<LoadInst>(Left[j + 1])) { | |||
2623 | if (isConsecutiveAccess(L, L1, *DL, *SE)) { | |||
2624 | std::swap(Left[j + 1], Right[j + 1]); | |||
2625 | continue; | |||
2626 | } | |||
2627 | } | |||
2628 | } | |||
2629 | // else unchanged | |||
2630 | } | |||
2631 | } | |||
2632 | ||||
2633 | void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL, Value *OpValue) { | |||
2634 | // Get the basic block this bundle is in. All instructions in the bundle | |||
2635 | // should be in this block. | |||
2636 | auto *Front = cast<Instruction>(OpValue); | |||
2637 | auto *BB = Front->getParent(); | |||
2638 | const unsigned Opcode = cast<Instruction>(OpValue)->getOpcode(); | |||
2639 | const unsigned AltOpcode = getAltOpcode(Opcode); | |||
2640 | assert(llvm::all_of(make_range(VL.begin(), VL.end()), [=](Value *V) -> bool {(static_cast <bool> (llvm::all_of(make_range(VL.begin() , VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt (Opcode, AltOpcode, cast<Instruction>(V)->getOpcode( )) || cast<Instruction>(V)->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(make_range(VL.begin(), VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt(Opcode, AltOpcode, cast<Instruction>(V)->getOpcode()) || cast<Instruction>(V)->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2644, __extension__ __PRETTY_FUNCTION__)) | |||
2641 | return !sameOpcodeOrAlt(Opcode, AltOpcode,(static_cast <bool> (llvm::all_of(make_range(VL.begin() , VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt (Opcode, AltOpcode, cast<Instruction>(V)->getOpcode( )) || cast<Instruction>(V)->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(make_range(VL.begin(), VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt(Opcode, AltOpcode, cast<Instruction>(V)->getOpcode()) || cast<Instruction>(V)->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2644, __extension__ __PRETTY_FUNCTION__)) | |||
2642 | cast<Instruction>(V)->getOpcode()) ||(static_cast <bool> (llvm::all_of(make_range(VL.begin() , VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt (Opcode, AltOpcode, cast<Instruction>(V)->getOpcode( )) || cast<Instruction>(V)->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(make_range(VL.begin(), VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt(Opcode, AltOpcode, cast<Instruction>(V)->getOpcode()) || cast<Instruction>(V)->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2644, __extension__ __PRETTY_FUNCTION__)) | |||
2643 | cast<Instruction>(V)->getParent() == BB;(static_cast <bool> (llvm::all_of(make_range(VL.begin() , VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt (Opcode, AltOpcode, cast<Instruction>(V)->getOpcode( )) || cast<Instruction>(V)->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(make_range(VL.begin(), VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt(Opcode, AltOpcode, cast<Instruction>(V)->getOpcode()) || cast<Instruction>(V)->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2644, __extension__ __PRETTY_FUNCTION__)) | |||
2644 | }))(static_cast <bool> (llvm::all_of(make_range(VL.begin() , VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt (Opcode, AltOpcode, cast<Instruction>(V)->getOpcode( )) || cast<Instruction>(V)->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(make_range(VL.begin(), VL.end()), [=](Value *V) -> bool { return !sameOpcodeOrAlt(Opcode, AltOpcode, cast<Instruction>(V)->getOpcode()) || cast<Instruction>(V)->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2644, __extension__ __PRETTY_FUNCTION__)); | |||
2645 | ||||
2646 | // The last instruction in the bundle in program order. | |||
2647 | Instruction *LastInst = nullptr; | |||
2648 | ||||
2649 | // Find the last instruction. The common case should be that BB has been | |||
2650 | // scheduled, and the last instruction is VL.back(). So we start with | |||
2651 | // VL.back() and iterate over schedule data until we reach the end of the | |||
2652 | // bundle. The end of the bundle is marked by null ScheduleData. | |||
2653 | if (BlocksSchedules.count(BB)) { | |||
2654 | auto *Bundle = | |||
2655 | BlocksSchedules[BB]->getScheduleData(isOneOf(OpValue, VL.back())); | |||
2656 | if (Bundle && Bundle->isPartOfBundle()) | |||
2657 | for (; Bundle; Bundle = Bundle->NextInBundle) | |||
2658 | if (Bundle->OpValue == Bundle->Inst) | |||
2659 | LastInst = Bundle->Inst; | |||
2660 | } | |||
2661 | ||||
2662 | // LastInst can still be null at this point if there's either not an entry | |||
2663 | // for BB in BlocksSchedules or there's no ScheduleData available for | |||
2664 | // VL.back(). This can be the case if buildTree_rec aborts for various | |||
2665 | // reasons (e.g., the maximum recursion depth is reached, the maximum region | |||
2666 | // size is reached, etc.). ScheduleData is initialized in the scheduling | |||
2667 | // "dry-run". | |||
2668 | // | |||
2669 | // If this happens, we can still find the last instruction by brute force. We | |||
2670 | // iterate forwards from Front (inclusive) until we either see all | |||
2671 | // instructions in the bundle or reach the end of the block. If Front is the | |||
2672 | // last instruction in program order, LastInst will be set to Front, and we | |||
2673 | // will visit all the remaining instructions in the block. | |||
2674 | // | |||
2675 | // One of the reasons we exit early from buildTree_rec is to place an upper | |||
2676 | // bound on compile-time. Thus, taking an additional compile-time hit here is | |||
2677 | // not ideal. However, this should be exceedingly rare since it requires that | |||
2678 | // we both exit early from buildTree_rec and that the bundle be out-of-order | |||
2679 | // (causing us to iterate all the way to the end of the block). | |||
2680 | if (!LastInst) { | |||
2681 | SmallPtrSet<Value *, 16> Bundle(VL.begin(), VL.end()); | |||
2682 | for (auto &I : make_range(BasicBlock::iterator(Front), BB->end())) { | |||
2683 | if (Bundle.erase(&I) && sameOpcodeOrAlt(Opcode, AltOpcode, I.getOpcode())) | |||
2684 | LastInst = &I; | |||
2685 | if (Bundle.empty()) | |||
2686 | break; | |||
2687 | } | |||
2688 | } | |||
2689 | ||||
2690 | // Set the insertion point after the last instruction in the bundle. Set the | |||
2691 | // debug location to Front. | |||
2692 | Builder.SetInsertPoint(BB, ++LastInst->getIterator()); | |||
2693 | Builder.SetCurrentDebugLocation(Front->getDebugLoc()); | |||
2694 | } | |||
2695 | ||||
2696 | Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) { | |||
2697 | Value *Vec = UndefValue::get(Ty); | |||
2698 | // Generate the 'InsertElement' instruction. | |||
2699 | for (unsigned i = 0; i < Ty->getNumElements(); ++i) { | |||
2700 | Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i)); | |||
2701 | if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) { | |||
2702 | GatherSeq.insert(Insrt); | |||
2703 | CSEBlocks.insert(Insrt->getParent()); | |||
2704 | ||||
2705 | // Add to our 'need-to-extract' list. | |||
2706 | if (TreeEntry *E = getTreeEntry(VL[i])) { | |||
2707 | // Find which lane we need to extract. | |||
2708 | int FoundLane = -1; | |||
2709 | for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) { | |||
2710 | // Is this the lane of the scalar that we are looking for ? | |||
2711 | if (E->Scalars[Lane] == VL[i]) { | |||
2712 | FoundLane = Lane; | |||
2713 | break; | |||
2714 | } | |||
2715 | } | |||
2716 | assert(FoundLane >= 0 && "Could not find the correct lane")(static_cast <bool> (FoundLane >= 0 && "Could not find the correct lane" ) ? void (0) : __assert_fail ("FoundLane >= 0 && \"Could not find the correct lane\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2716, __extension__ __PRETTY_FUNCTION__)); | |||
2717 | ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane)); | |||
2718 | } | |||
2719 | } | |||
2720 | } | |||
2721 | ||||
2722 | return Vec; | |||
2723 | } | |||
2724 | ||||
2725 | Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL, Value *OpValue) const { | |||
2726 | if (const TreeEntry *En = getTreeEntry(OpValue)) { | |||
2727 | if (En->isSame(VL) && En->VectorizedValue) | |||
2728 | return En->VectorizedValue; | |||
2729 | } | |||
2730 | return nullptr; | |||
2731 | } | |||
2732 | ||||
2733 | Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) { | |||
2734 | InstructionsState S = getSameOpcode(VL); | |||
2735 | if (S.Opcode) { | |||
2736 | if (TreeEntry *E = getTreeEntry(S.OpValue)) { | |||
2737 | if (E->isSame(VL)) | |||
2738 | return vectorizeTree(E); | |||
2739 | } | |||
2740 | } | |||
2741 | ||||
2742 | Type *ScalarTy = S.OpValue->getType(); | |||
2743 | if (StoreInst *SI = dyn_cast<StoreInst>(S.OpValue)) | |||
2744 | ScalarTy = SI->getValueOperand()->getType(); | |||
2745 | VectorType *VecTy = VectorType::get(ScalarTy, VL.size()); | |||
2746 | ||||
2747 | return Gather(VL, VecTy); | |||
2748 | } | |||
2749 | ||||
2750 | Value *BoUpSLP::vectorizeTree(TreeEntry *E) { | |||
2751 | IRBuilder<>::InsertPointGuard Guard(Builder); | |||
2752 | ||||
2753 | if (E->VectorizedValue) { | |||
| ||||
2754 | DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n"; } } while (false); | |||
2755 | return E->VectorizedValue; | |||
2756 | } | |||
2757 | ||||
2758 | InstructionsState S = getSameOpcode(E->Scalars); | |||
2759 | Instruction *VL0 = cast<Instruction>(E->Scalars[0]); | |||
2760 | Type *ScalarTy = VL0->getType(); | |||
2761 | if (StoreInst *SI = dyn_cast<StoreInst>(VL0)) | |||
2762 | ScalarTy = SI->getValueOperand()->getType(); | |||
2763 | VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size()); | |||
2764 | ||||
2765 | if (E->NeedToGather) { | |||
2766 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
2767 | auto *V = Gather(E->Scalars, VecTy); | |||
2768 | E->VectorizedValue = V; | |||
2769 | return V; | |||
2770 | } | |||
2771 | ||||
2772 | unsigned ShuffleOrOp = S.IsAltShuffle ? | |||
2773 | (unsigned) Instruction::ShuffleVector : S.Opcode; | |||
2774 | switch (ShuffleOrOp) { | |||
2775 | case Instruction::PHI: { | |||
2776 | PHINode *PH = dyn_cast<PHINode>(VL0); | |||
2777 | Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI()); | |||
2778 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | |||
2779 | PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues()); | |||
2780 | E->VectorizedValue = NewPhi; | |||
2781 | ||||
2782 | // PHINodes may have multiple entries from the same block. We want to | |||
2783 | // visit every block once. | |||
2784 | SmallSet<BasicBlock*, 4> VisitedBBs; | |||
2785 | ||||
2786 | for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { | |||
2787 | ValueList Operands; | |||
2788 | BasicBlock *IBB = PH->getIncomingBlock(i); | |||
2789 | ||||
2790 | if (!VisitedBBs.insert(IBB).second) { | |||
2791 | NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB); | |||
2792 | continue; | |||
2793 | } | |||
2794 | ||||
2795 | // Prepare the operand vector. | |||
2796 | for (Value *V : E->Scalars) | |||
2797 | Operands.push_back(cast<PHINode>(V)->getIncomingValueForBlock(IBB)); | |||
2798 | ||||
2799 | Builder.SetInsertPoint(IBB->getTerminator()); | |||
2800 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | |||
2801 | Value *Vec = vectorizeTree(Operands); | |||
2802 | NewPhi->addIncoming(Vec, IBB); | |||
2803 | } | |||
2804 | ||||
2805 | assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&(static_cast <bool> (NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && "Invalid number of incoming values" ) ? void (0) : __assert_fail ("NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && \"Invalid number of incoming values\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2806, __extension__ __PRETTY_FUNCTION__)) | |||
2806 | "Invalid number of incoming values")(static_cast <bool> (NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && "Invalid number of incoming values" ) ? void (0) : __assert_fail ("NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && \"Invalid number of incoming values\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2806, __extension__ __PRETTY_FUNCTION__)); | |||
2807 | return NewPhi; | |||
2808 | } | |||
2809 | ||||
2810 | case Instruction::ExtractElement: { | |||
2811 | if (canReuseExtract(E->Scalars, VL0)) { | |||
2812 | Value *V = VL0->getOperand(0); | |||
2813 | E->VectorizedValue = V; | |||
2814 | return V; | |||
2815 | } | |||
2816 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
2817 | auto *V = Gather(E->Scalars, VecTy); | |||
2818 | E->VectorizedValue = V; | |||
2819 | return V; | |||
2820 | } | |||
2821 | case Instruction::ExtractValue: { | |||
2822 | if (canReuseExtract(E->Scalars, VL0)) { | |||
2823 | LoadInst *LI = cast<LoadInst>(VL0->getOperand(0)); | |||
2824 | Builder.SetInsertPoint(LI); | |||
2825 | PointerType *PtrTy = PointerType::get(VecTy, LI->getPointerAddressSpace()); | |||
2826 | Value *Ptr = Builder.CreateBitCast(LI->getOperand(0), PtrTy); | |||
2827 | LoadInst *V = Builder.CreateAlignedLoad(Ptr, LI->getAlignment()); | |||
2828 | E->VectorizedValue = V; | |||
2829 | return propagateMetadata(V, E->Scalars); | |||
2830 | } | |||
2831 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
2832 | auto *V = Gather(E->Scalars, VecTy); | |||
2833 | E->VectorizedValue = V; | |||
2834 | return V; | |||
2835 | } | |||
2836 | case Instruction::ZExt: | |||
2837 | case Instruction::SExt: | |||
2838 | case Instruction::FPToUI: | |||
2839 | case Instruction::FPToSI: | |||
2840 | case Instruction::FPExt: | |||
2841 | case Instruction::PtrToInt: | |||
2842 | case Instruction::IntToPtr: | |||
2843 | case Instruction::SIToFP: | |||
2844 | case Instruction::UIToFP: | |||
2845 | case Instruction::Trunc: | |||
2846 | case Instruction::FPTrunc: | |||
2847 | case Instruction::BitCast: { | |||
2848 | ValueList INVL; | |||
2849 | for (Value *V : E->Scalars) | |||
2850 | INVL.push_back(cast<Instruction>(V)->getOperand(0)); | |||
2851 | ||||
2852 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
2853 | ||||
2854 | Value *InVec = vectorizeTree(INVL); | |||
2855 | ||||
2856 | if (Value *V = alreadyVectorized(E->Scalars, VL0)) | |||
2857 | return V; | |||
2858 | ||||
2859 | CastInst *CI = dyn_cast<CastInst>(VL0); | |||
2860 | Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy); | |||
2861 | E->VectorizedValue = V; | |||
2862 | ++NumVectorInstructions; | |||
2863 | return V; | |||
2864 | } | |||
2865 | case Instruction::FCmp: | |||
2866 | case Instruction::ICmp: { | |||
2867 | ValueList LHSV, RHSV; | |||
2868 | for (Value *V : E->Scalars) { | |||
2869 | LHSV.push_back(cast<Instruction>(V)->getOperand(0)); | |||
2870 | RHSV.push_back(cast<Instruction>(V)->getOperand(1)); | |||
2871 | } | |||
2872 | ||||
2873 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
2874 | ||||
2875 | Value *L = vectorizeTree(LHSV); | |||
2876 | Value *R = vectorizeTree(RHSV); | |||
2877 | ||||
2878 | if (Value *V = alreadyVectorized(E->Scalars, VL0)) | |||
2879 | return V; | |||
2880 | ||||
2881 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); | |||
2882 | Value *V; | |||
2883 | if (S.Opcode == Instruction::FCmp) | |||
2884 | V = Builder.CreateFCmp(P0, L, R); | |||
2885 | else | |||
2886 | V = Builder.CreateICmp(P0, L, R); | |||
2887 | ||||
2888 | E->VectorizedValue = V; | |||
2889 | propagateIRFlags(E->VectorizedValue, E->Scalars, VL0); | |||
2890 | ++NumVectorInstructions; | |||
2891 | return V; | |||
2892 | } | |||
2893 | case Instruction::Select: { | |||
2894 | ValueList TrueVec, FalseVec, CondVec; | |||
2895 | for (Value *V : E->Scalars) { | |||
2896 | CondVec.push_back(cast<Instruction>(V)->getOperand(0)); | |||
2897 | TrueVec.push_back(cast<Instruction>(V)->getOperand(1)); | |||
2898 | FalseVec.push_back(cast<Instruction>(V)->getOperand(2)); | |||
2899 | } | |||
2900 | ||||
2901 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
2902 | ||||
2903 | Value *Cond = vectorizeTree(CondVec); | |||
2904 | Value *True = vectorizeTree(TrueVec); | |||
2905 | Value *False = vectorizeTree(FalseVec); | |||
2906 | ||||
2907 | if (Value *V = alreadyVectorized(E->Scalars, VL0)) | |||
2908 | return V; | |||
2909 | ||||
2910 | Value *V = Builder.CreateSelect(Cond, True, False); | |||
2911 | E->VectorizedValue = V; | |||
2912 | ++NumVectorInstructions; | |||
2913 | return V; | |||
2914 | } | |||
2915 | case Instruction::Add: | |||
2916 | case Instruction::FAdd: | |||
2917 | case Instruction::Sub: | |||
2918 | case Instruction::FSub: | |||
2919 | case Instruction::Mul: | |||
2920 | case Instruction::FMul: | |||
2921 | case Instruction::UDiv: | |||
2922 | case Instruction::SDiv: | |||
2923 | case Instruction::FDiv: | |||
2924 | case Instruction::URem: | |||
2925 | case Instruction::SRem: | |||
2926 | case Instruction::FRem: | |||
2927 | case Instruction::Shl: | |||
2928 | case Instruction::LShr: | |||
2929 | case Instruction::AShr: | |||
2930 | case Instruction::And: | |||
2931 | case Instruction::Or: | |||
2932 | case Instruction::Xor: { | |||
2933 | ValueList LHSVL, RHSVL; | |||
2934 | if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) | |||
2935 | reorderInputsAccordingToOpcode(S.Opcode, E->Scalars, LHSVL, | |||
2936 | RHSVL); | |||
2937 | else | |||
2938 | for (Value *V : E->Scalars) { | |||
2939 | auto *I = cast<Instruction>(V); | |||
2940 | LHSVL.push_back(I->getOperand(0)); | |||
2941 | RHSVL.push_back(I->getOperand(1)); | |||
2942 | } | |||
2943 | ||||
2944 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
2945 | ||||
2946 | Value *LHS = vectorizeTree(LHSVL); | |||
2947 | Value *RHS = vectorizeTree(RHSVL); | |||
2948 | ||||
2949 | if (Value *V = alreadyVectorized(E->Scalars, VL0)) | |||
2950 | return V; | |||
2951 | ||||
2952 | Value *V = Builder.CreateBinOp( | |||
2953 | static_cast<Instruction::BinaryOps>(S.Opcode), LHS, RHS); | |||
2954 | E->VectorizedValue = V; | |||
2955 | propagateIRFlags(E->VectorizedValue, E->Scalars, VL0); | |||
2956 | ++NumVectorInstructions; | |||
2957 | ||||
2958 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
2959 | return propagateMetadata(I, E->Scalars); | |||
2960 | ||||
2961 | return V; | |||
2962 | } | |||
2963 | case Instruction::Load: { | |||
2964 | // Loads are inserted at the head of the tree because we don't want to | |||
2965 | // sink them all the way down past store instructions. | |||
2966 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
2967 | ||||
2968 | LoadInst *LI = cast<LoadInst>(VL0); | |||
2969 | Type *ScalarLoadTy = LI->getType(); | |||
2970 | unsigned AS = LI->getPointerAddressSpace(); | |||
2971 | ||||
2972 | Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(), | |||
2973 | VecTy->getPointerTo(AS)); | |||
2974 | ||||
2975 | // The pointer operand uses an in-tree scalar so we add the new BitCast to | |||
2976 | // ExternalUses list to make sure that an extract will be generated in the | |||
2977 | // future. | |||
2978 | Value *PO = LI->getPointerOperand(); | |||
2979 | if (getTreeEntry(PO)) | |||
2980 | ExternalUses.push_back(ExternalUser(PO, cast<User>(VecPtr), 0)); | |||
2981 | ||||
2982 | unsigned Alignment = LI->getAlignment(); | |||
2983 | LI = Builder.CreateLoad(VecPtr); | |||
2984 | if (!Alignment) { | |||
2985 | Alignment = DL->getABITypeAlignment(ScalarLoadTy); | |||
2986 | } | |||
2987 | LI->setAlignment(Alignment); | |||
2988 | E->VectorizedValue = LI; | |||
2989 | ++NumVectorInstructions; | |||
2990 | return propagateMetadata(LI, E->Scalars); | |||
2991 | } | |||
2992 | case Instruction::Store: { | |||
2993 | StoreInst *SI = cast<StoreInst>(VL0); | |||
2994 | unsigned Alignment = SI->getAlignment(); | |||
2995 | unsigned AS = SI->getPointerAddressSpace(); | |||
2996 | ||||
2997 | ValueList ScalarStoreValues; | |||
2998 | for (Value *V : E->Scalars) | |||
2999 | ScalarStoreValues.push_back(cast<StoreInst>(V)->getValueOperand()); | |||
3000 | ||||
3001 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3002 | ||||
3003 | Value *VecValue = vectorizeTree(ScalarStoreValues); | |||
3004 | Value *ScalarPtr = SI->getPointerOperand(); | |||
3005 | Value *VecPtr = Builder.CreateBitCast(ScalarPtr, VecTy->getPointerTo(AS)); | |||
3006 | StoreInst *S = Builder.CreateStore(VecValue, VecPtr); | |||
3007 | ||||
3008 | // The pointer operand uses an in-tree scalar, so add the new BitCast to | |||
3009 | // ExternalUses to make sure that an extract will be generated in the | |||
3010 | // future. | |||
3011 | if (getTreeEntry(ScalarPtr)) | |||
3012 | ExternalUses.push_back(ExternalUser(ScalarPtr, cast<User>(VecPtr), 0)); | |||
3013 | ||||
3014 | if (!Alignment) | |||
3015 | Alignment = DL->getABITypeAlignment(SI->getValueOperand()->getType()); | |||
3016 | ||||
3017 | S->setAlignment(Alignment); | |||
3018 | E->VectorizedValue = S; | |||
3019 | ++NumVectorInstructions; | |||
3020 | return propagateMetadata(S, E->Scalars); | |||
3021 | } | |||
3022 | case Instruction::GetElementPtr: { | |||
3023 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3024 | ||||
3025 | ValueList Op0VL; | |||
3026 | for (Value *V : E->Scalars) | |||
3027 | Op0VL.push_back(cast<GetElementPtrInst>(V)->getOperand(0)); | |||
3028 | ||||
3029 | Value *Op0 = vectorizeTree(Op0VL); | |||
3030 | ||||
3031 | std::vector<Value *> OpVecs; | |||
3032 | for (int j = 1, e = cast<GetElementPtrInst>(VL0)->getNumOperands(); j < e; | |||
3033 | ++j) { | |||
3034 | ValueList OpVL; | |||
3035 | for (Value *V : E->Scalars) | |||
3036 | OpVL.push_back(cast<GetElementPtrInst>(V)->getOperand(j)); | |||
3037 | ||||
3038 | Value *OpVec = vectorizeTree(OpVL); | |||
3039 | OpVecs.push_back(OpVec); | |||
3040 | } | |||
3041 | ||||
3042 | Value *V = Builder.CreateGEP( | |||
3043 | cast<GetElementPtrInst>(VL0)->getSourceElementType(), Op0, OpVecs); | |||
3044 | E->VectorizedValue = V; | |||
3045 | ++NumVectorInstructions; | |||
3046 | ||||
3047 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
3048 | return propagateMetadata(I, E->Scalars); | |||
3049 | ||||
3050 | return V; | |||
3051 | } | |||
3052 | case Instruction::Call: { | |||
3053 | CallInst *CI = cast<CallInst>(VL0); | |||
3054 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3055 | Function *FI; | |||
3056 | Intrinsic::ID IID = Intrinsic::not_intrinsic; | |||
3057 | Value *ScalarArg = nullptr; | |||
3058 | if (CI && (FI = CI->getCalledFunction())) { | |||
3059 | IID = FI->getIntrinsicID(); | |||
3060 | } | |||
3061 | std::vector<Value *> OpVecs; | |||
3062 | for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) { | |||
| ||||
3063 | ValueList OpVL; | |||
3064 | // ctlz,cttz and powi are special intrinsics whose second argument is | |||
3065 | // a scalar. This argument should not be vectorized. | |||
3066 | if (hasVectorInstrinsicScalarOpd(IID, 1) && j == 1) { | |||
3067 | CallInst *CEI = cast<CallInst>(VL0); | |||
3068 | ScalarArg = CEI->getArgOperand(j); | |||
3069 | OpVecs.push_back(CEI->getArgOperand(j)); | |||
3070 | continue; | |||
3071 | } | |||
3072 | for (Value *V : E->Scalars) { | |||
3073 | CallInst *CEI = cast<CallInst>(V); | |||
3074 | OpVL.push_back(CEI->getArgOperand(j)); | |||
3075 | } | |||
3076 | ||||
3077 | Value *OpVec = vectorizeTree(OpVL); | |||
3078 | DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n"; } } while (false); | |||
3079 | OpVecs.push_back(OpVec); | |||
3080 | } | |||
3081 | ||||
3082 | Module *M = F->getParent(); | |||
3083 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | |||
3084 | Type *Tys[] = { VectorType::get(CI->getType(), E->Scalars.size()) }; | |||
3085 | Function *CF = Intrinsic::getDeclaration(M, ID, Tys); | |||
3086 | SmallVector<OperandBundleDef, 1> OpBundles; | |||
3087 | CI->getOperandBundlesAsDefs(OpBundles); | |||
3088 | Value *V = Builder.CreateCall(CF, OpVecs, OpBundles); | |||
3089 | ||||
3090 | // The scalar argument uses an in-tree scalar so we add the new vectorized | |||
3091 | // call to ExternalUses list to make sure that an extract will be | |||
3092 | // generated in the future. | |||
3093 | if (ScalarArg && getTreeEntry(ScalarArg)) | |||
3094 | ExternalUses.push_back(ExternalUser(ScalarArg, cast<User>(V), 0)); | |||
3095 | ||||
3096 | E->VectorizedValue = V; | |||
3097 | propagateIRFlags(E->VectorizedValue, E->Scalars, VL0); | |||
3098 | ++NumVectorInstructions; | |||
3099 | return V; | |||
3100 | } | |||
3101 | case Instruction::ShuffleVector: { | |||
3102 | ValueList LHSVL, RHSVL; | |||
3103 | assert(Instruction::isBinaryOp(S.Opcode) &&(static_cast <bool> (Instruction::isBinaryOp(S.Opcode) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("Instruction::isBinaryOp(S.Opcode) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3104, __extension__ __PRETTY_FUNCTION__)) | |||
3104 | "Invalid Shuffle Vector Operand")(static_cast <bool> (Instruction::isBinaryOp(S.Opcode) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("Instruction::isBinaryOp(S.Opcode) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3104, __extension__ __PRETTY_FUNCTION__)); | |||
3105 | reorderAltShuffleOperands(S.Opcode, E->Scalars, LHSVL, RHSVL); | |||
3106 | setInsertPointAfterBundle(E->Scalars, VL0); | |||
3107 | ||||
3108 | Value *LHS = vectorizeTree(LHSVL); | |||
3109 | Value *RHS = vectorizeTree(RHSVL); | |||
3110 | ||||
3111 | if (Value *V = alreadyVectorized(E->Scalars, VL0)) | |||
3112 | return V; | |||
3113 | ||||
3114 | // Create a vector of LHS op1 RHS | |||
3115 | Value *V0 = Builder.CreateBinOp( | |||
3116 | static_cast<Instruction::BinaryOps>(S.Opcode), LHS, RHS); | |||
3117 | ||||
3118 | unsigned AltOpcode = getAltOpcode(S.Opcode); | |||
3119 | // Create a vector of LHS op2 RHS | |||
3120 | Value *V1 = Builder.CreateBinOp( | |||
3121 | static_cast<Instruction::BinaryOps>(AltOpcode), LHS, RHS); | |||
3122 | ||||
3123 | // Create shuffle to take alternate operations from the vector. | |||
3124 | // Also, gather up odd and even scalar ops to propagate IR flags to | |||
3125 | // each vector operation. | |||
3126 | ValueList OddScalars, EvenScalars; | |||
3127 | unsigned e = E->Scalars.size(); | |||
3128 | SmallVector<Constant *, 8> Mask(e); | |||
3129 | for (unsigned i = 0; i < e; ++i) { | |||
3130 | if (isOdd(i)) { | |||
3131 | Mask[i] = Builder.getInt32(e + i); | |||
3132 | OddScalars.push_back(E->Scalars[i]); | |||
3133 | } else { | |||
3134 | Mask[i] = Builder.getInt32(i); | |||
3135 | EvenScalars.push_back(E->Scalars[i]); | |||
3136 | } | |||
3137 | } | |||
3138 | ||||
3139 | Value *ShuffleMask = ConstantVector::get(Mask); | |||
3140 | propagateIRFlags(V0, EvenScalars); | |||
3141 | propagateIRFlags(V1, OddScalars); | |||
3142 | ||||
3143 | Value *V = Builder.CreateShuffleVector(V0, V1, ShuffleMask); | |||
3144 | E->VectorizedValue = V; | |||
3145 | ++NumVectorInstructions; | |||
3146 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
3147 | return propagateMetadata(I, E->Scalars); | |||
3148 | ||||
3149 | return V; | |||
3150 | } | |||
3151 | default: | |||
3152 | llvm_unreachable("unknown inst")::llvm::llvm_unreachable_internal("unknown inst", "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3152); | |||
3153 | } | |||
3154 | return nullptr; | |||
3155 | } | |||
3156 | ||||
3157 | Value *BoUpSLP::vectorizeTree() { | |||
3158 | ExtraValueToDebugLocsMap ExternallyUsedValues; | |||
3159 | return vectorizeTree(ExternallyUsedValues); | |||
3160 | } | |||
3161 | ||||
3162 | Value * | |||
3163 | BoUpSLP::vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues) { | |||
3164 | // All blocks must be scheduled before any instructions are inserted. | |||
3165 | for (auto &BSIter : BlocksSchedules) { | |||
3166 | scheduleBlock(BSIter.second.get()); | |||
3167 | } | |||
3168 | ||||
3169 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | |||
3170 | auto *VectorRoot = vectorizeTree(&VectorizableTree[0]); | |||
3171 | ||||
3172 | // If the vectorized tree can be rewritten in a smaller type, we truncate the | |||
3173 | // vectorized root. InstCombine will then rewrite the entire expression. We | |||
3174 | // sign extend the extracted values below. | |||
3175 | auto *ScalarRoot = VectorizableTree[0].Scalars[0]; | |||
3176 | if (MinBWs.count(ScalarRoot)) { | |||
3177 | if (auto *I = dyn_cast<Instruction>(VectorRoot)) | |||
3178 | Builder.SetInsertPoint(&*++BasicBlock::iterator(I)); | |||
3179 | auto BundleWidth = VectorizableTree[0].Scalars.size(); | |||
3180 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); | |||
3181 | auto *VecTy = VectorType::get(MinTy, BundleWidth); | |||
3182 | auto *Trunc = Builder.CreateTrunc(VectorRoot, VecTy); | |||
3183 | VectorizableTree[0].VectorizedValue = Trunc; | |||
3184 | } | |||
3185 | ||||
3186 | DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Extracting " << ExternalUses .size() << " values .\n"; } } while (false); | |||
3187 | ||||
3188 | // If necessary, sign-extend or zero-extend ScalarRoot to the larger type | |||
3189 | // specified by ScalarType. | |||
3190 | auto extend = [&](Value *ScalarRoot, Value *Ex, Type *ScalarType) { | |||
3191 | if (!MinBWs.count(ScalarRoot)) | |||
3192 | return Ex; | |||
3193 | if (MinBWs[ScalarRoot].second) | |||
3194 | return Builder.CreateSExt(Ex, ScalarType); | |||
3195 | return Builder.CreateZExt(Ex, ScalarType); | |||
3196 | }; | |||
3197 | ||||
3198 | // Extract all of the elements with the external uses. | |||
3199 | for (const auto &ExternalUse : ExternalUses) { | |||
3200 | Value *Scalar = ExternalUse.Scalar; | |||
3201 | llvm::User *User = ExternalUse.User; | |||
3202 | ||||
3203 | // Skip users that we already RAUW. This happens when one instruction | |||
3204 | // has multiple uses of the same value. | |||
3205 | if (User && !is_contained(Scalar->users(), User)) | |||
3206 | continue; | |||
3207 | TreeEntry *E = getTreeEntry(Scalar); | |||
3208 | assert(E && "Invalid scalar")(static_cast <bool> (E && "Invalid scalar") ? void (0) : __assert_fail ("E && \"Invalid scalar\"", "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3208, __extension__ __PRETTY_FUNCTION__)); | |||
3209 | assert(!E->NeedToGather && "Extracting from a gather list")(static_cast <bool> (!E->NeedToGather && "Extracting from a gather list" ) ? void (0) : __assert_fail ("!E->NeedToGather && \"Extracting from a gather list\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3209, __extension__ __PRETTY_FUNCTION__)); | |||
3210 | ||||
3211 | Value *Vec = E->VectorizedValue; | |||
3212 | assert(Vec && "Can't find vectorizable value")(static_cast <bool> (Vec && "Can't find vectorizable value" ) ? void (0) : __assert_fail ("Vec && \"Can't find vectorizable value\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3212, __extension__ __PRETTY_FUNCTION__)); | |||
3213 | ||||
3214 | Value *Lane = Builder.getInt32(ExternalUse.Lane); | |||
3215 | // If User == nullptr, the Scalar is used as extra arg. Generate | |||
3216 | // ExtractElement instruction and update the record for this scalar in | |||
3217 | // ExternallyUsedValues. | |||
3218 | if (!User) { | |||
3219 | assert(ExternallyUsedValues.count(Scalar) &&(static_cast <bool> (ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? void (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3221, __extension__ __PRETTY_FUNCTION__)) | |||
3220 | "Scalar with nullptr as an external user must be registered in "(static_cast <bool> (ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? void (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3221, __extension__ __PRETTY_FUNCTION__)) | |||
3221 | "ExternallyUsedValues map")(static_cast <bool> (ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? void (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3221, __extension__ __PRETTY_FUNCTION__)); | |||
3222 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { | |||
3223 | Builder.SetInsertPoint(VecI->getParent(), | |||
3224 | std::next(VecI->getIterator())); | |||
3225 | } else { | |||
3226 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | |||
3227 | } | |||
3228 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | |||
3229 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | |||
3230 | CSEBlocks.insert(cast<Instruction>(Scalar)->getParent()); | |||
3231 | auto &Locs = ExternallyUsedValues[Scalar]; | |||
3232 | ExternallyUsedValues.insert({Ex, Locs}); | |||
3233 | ExternallyUsedValues.erase(Scalar); | |||
3234 | continue; | |||
3235 | } | |||
3236 | ||||
3237 | // Generate extracts for out-of-tree users. | |||
3238 | // Find the insertion point for the extractelement lane. | |||
3239 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { | |||
3240 | if (PHINode *PH = dyn_cast<PHINode>(User)) { | |||
3241 | for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) { | |||
3242 | if (PH->getIncomingValue(i) == Scalar) { | |||
3243 | TerminatorInst *IncomingTerminator = | |||
3244 | PH->getIncomingBlock(i)->getTerminator(); | |||
3245 | if (isa<CatchSwitchInst>(IncomingTerminator)) { | |||
3246 | Builder.SetInsertPoint(VecI->getParent(), | |||
3247 | std::next(VecI->getIterator())); | |||
3248 | } else { | |||
3249 | Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator()); | |||
3250 | } | |||
3251 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | |||
3252 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | |||
3253 | CSEBlocks.insert(PH->getIncomingBlock(i)); | |||
3254 | PH->setOperand(i, Ex); | |||
3255 | } | |||
3256 | } | |||
3257 | } else { | |||
3258 | Builder.SetInsertPoint(cast<Instruction>(User)); | |||
3259 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | |||
3260 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | |||
3261 | CSEBlocks.insert(cast<Instruction>(User)->getParent()); | |||
3262 | User->replaceUsesOfWith(Scalar, Ex); | |||
3263 | } | |||
3264 | } else { | |||
3265 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | |||
3266 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | |||
3267 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | |||
3268 | CSEBlocks.insert(&F->getEntryBlock()); | |||
3269 | User->replaceUsesOfWith(Scalar, Ex); | |||
3270 | } | |||
3271 | ||||
3272 | DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Replaced:" << *User << ".\n"; } } while (false); | |||
3273 | } | |||
3274 | ||||
3275 | // For each vectorized value: | |||
3276 | for (TreeEntry &EIdx : VectorizableTree) { | |||
3277 | TreeEntry *Entry = &EIdx; | |||
3278 | ||||
3279 | // No need to handle users of gathered values. | |||
3280 | if (Entry->NeedToGather) | |||
3281 | continue; | |||
3282 | ||||
3283 | assert(Entry->VectorizedValue && "Can't find vectorizable value")(static_cast <bool> (Entry->VectorizedValue && "Can't find vectorizable value") ? void (0) : __assert_fail ( "Entry->VectorizedValue && \"Can't find vectorizable value\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3283, __extension__ __PRETTY_FUNCTION__)); | |||
3284 | ||||
3285 | // For each lane: | |||
3286 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { | |||
3287 | Value *Scalar = Entry->Scalars[Lane]; | |||
3288 | ||||
3289 | Type *Ty = Scalar->getType(); | |||
3290 | if (!Ty->isVoidTy()) { | |||
3291 | #ifndef NDEBUG | |||
3292 | for (User *U : Scalar->users()) { | |||
3293 | DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tvalidating user:" << *U << ".\n"; } } while (false); | |||
3294 | ||||
3295 | // It is legal to replace users in the ignorelist by undef. | |||
3296 | assert((getTreeEntry(U) || is_contained(UserIgnoreList, U)) &&(static_cast <bool> ((getTreeEntry(U) || is_contained(UserIgnoreList , U)) && "Replacing out-of-tree value with undef") ? void (0) : __assert_fail ("(getTreeEntry(U) || is_contained(UserIgnoreList, U)) && \"Replacing out-of-tree value with undef\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3297, __extension__ __PRETTY_FUNCTION__)) | |||
3297 | "Replacing out-of-tree value with undef")(static_cast <bool> ((getTreeEntry(U) || is_contained(UserIgnoreList , U)) && "Replacing out-of-tree value with undef") ? void (0) : __assert_fail ("(getTreeEntry(U) || is_contained(UserIgnoreList, U)) && \"Replacing out-of-tree value with undef\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3297, __extension__ __PRETTY_FUNCTION__)); | |||
3298 | } | |||
3299 | #endif | |||
3300 | Value *Undef = UndefValue::get(Ty); | |||
3301 | Scalar->replaceAllUsesWith(Undef); | |||
3302 | } | |||
3303 | DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tErasing scalar:" << * Scalar << ".\n"; } } while (false); | |||
3304 | eraseInstruction(cast<Instruction>(Scalar)); | |||
3305 | } | |||
3306 | } | |||
3307 | ||||
3308 | Builder.ClearInsertionPoint(); | |||
3309 | ||||
3310 | return VectorizableTree[0].VectorizedValue; | |||
3311 | } | |||
3312 | ||||
3313 | void BoUpSLP::optimizeGatherSequence(Function &F) { | |||
3314 | DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherSeq .size() << " gather sequences instructions.\n"; } } while (false) | |||
3315 | << " gather sequences instructions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherSeq .size() << " gather sequences instructions.\n"; } } while (false); | |||
3316 | // LICM InsertElementInst sequences. | |||
3317 | for (Instruction *it : GatherSeq) { | |||
3318 | InsertElementInst *Insert = dyn_cast<InsertElementInst>(it); | |||
3319 | ||||
3320 | if (!Insert) | |||
3321 | continue; | |||
3322 | ||||
3323 | // Check if this block is inside a loop. | |||
3324 | Loop *L = LI->getLoopFor(Insert->getParent()); | |||
3325 | if (!L) | |||
3326 | continue; | |||
3327 | ||||
3328 | // Check if it has a preheader. | |||
3329 | BasicBlock *PreHeader = L->getLoopPreheader(); | |||
3330 | if (!PreHeader) | |||
3331 | continue; | |||
3332 | ||||
3333 | // If the vector or the element that we insert into it are | |||
3334 | // instructions that are defined in this basic block then we can't | |||
3335 | // hoist this instruction. | |||
3336 | Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0)); | |||
3337 | Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1)); | |||
3338 | if (CurrVec && L->contains(CurrVec)) | |||
3339 | continue; | |||
3340 | if (NewElem && L->contains(NewElem)) | |||
3341 | continue; | |||
3342 | ||||
3343 | // We can hoist this instruction. Move it to the pre-header. | |||
3344 | Insert->moveBefore(PreHeader->getTerminator()); | |||
3345 | } | |||
3346 | ||||
3347 | // Perform O(N^2) search over the gather sequences and merge identical | |||
3348 | // instructions. TODO: We can further optimize this scan if we split the | |||
3349 | // instructions into different buckets based on the insert lane. | |||
3350 | SmallVector<Instruction *, 16> Visited; | |||
3351 | ReversePostOrderTraversal<Function *> RPOT(&F); | |||
3352 | for (auto BB : RPOT) { | |||
3353 | // Traverse CSEBlocks by RPOT order. | |||
3354 | if (!CSEBlocks.count(BB)) | |||
3355 | continue; | |||
3356 | ||||
3357 | // For all instructions in blocks containing gather sequences: | |||
3358 | for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) { | |||
3359 | Instruction *In = &*it++; | |||
3360 | if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In)) | |||
3361 | continue; | |||
3362 | ||||
3363 | // Check if we can replace this instruction with any of the | |||
3364 | // visited instructions. | |||
3365 | for (Instruction *v : Visited) { | |||
3366 | if (In->isIdenticalTo(v) && | |||
3367 | DT->dominates(v->getParent(), In->getParent())) { | |||
3368 | In->replaceAllUsesWith(v); | |||
3369 | eraseInstruction(In); | |||
3370 | In = nullptr; | |||
3371 | break; | |||
3372 | } | |||
3373 | } | |||
3374 | if (In) { | |||
3375 | assert(!is_contained(Visited, In))(static_cast <bool> (!is_contained(Visited, In)) ? void (0) : __assert_fail ("!is_contained(Visited, In)", "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3375, __extension__ __PRETTY_FUNCTION__)); | |||
3376 | Visited.push_back(In); | |||
3377 | } | |||
3378 | } | |||
3379 | } | |||
3380 | CSEBlocks.clear(); | |||
3381 | GatherSeq.clear(); | |||
3382 | } | |||
3383 | ||||
3384 | // Groups the instructions to a bundle (which is then a single scheduling entity) | |||
3385 | // and schedules instructions until the bundle gets ready. | |||
3386 | bool BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef<Value *> VL, | |||
3387 | BoUpSLP *SLP, Value *OpValue) { | |||
3388 | if (isa<PHINode>(OpValue)) | |||
3389 | return true; | |||
3390 | ||||
3391 | // Initialize the instruction bundle. | |||
3392 | Instruction *OldScheduleEnd = ScheduleEnd; | |||
3393 | ScheduleData *PrevInBundle = nullptr; | |||
3394 | ScheduleData *Bundle = nullptr; | |||
3395 | bool ReSchedule = false; | |||
3396 | DEBUG(dbgs() << "SLP: bundle: " << *OpValue << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: bundle: " << *OpValue << "\n"; } } while (false); | |||
3397 | ||||
3398 | // Make sure that the scheduling region contains all | |||
3399 | // instructions of the bundle. | |||
3400 | for (Value *V : VL) { | |||
3401 | if (!extendSchedulingRegion(V, OpValue)) | |||
3402 | return false; | |||
3403 | } | |||
3404 | ||||
3405 | for (Value *V : VL) { | |||
3406 | ScheduleData *BundleMember = getScheduleData(V); | |||
3407 | assert(BundleMember &&(static_cast <bool> (BundleMember && "no ScheduleData for bundle member (maybe not in same basic block)" ) ? void (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member (maybe not in same basic block)\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3408, __extension__ __PRETTY_FUNCTION__)) | |||
3408 | "no ScheduleData for bundle member (maybe not in same basic block)")(static_cast <bool> (BundleMember && "no ScheduleData for bundle member (maybe not in same basic block)" ) ? void (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member (maybe not in same basic block)\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3408, __extension__ __PRETTY_FUNCTION__)); | |||
3409 | if (BundleMember->IsScheduled) { | |||
3410 | // A bundle member was scheduled as single instruction before and now | |||
3411 | // needs to be scheduled as part of the bundle. We just get rid of the | |||
3412 | // existing schedule. | |||
3413 | DEBUG(dbgs() << "SLP: reset schedule because " << *BundleMemberdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: reset schedule because " << *BundleMember << " was already scheduled\n"; } } while (false) | |||
3414 | << " was already scheduled\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: reset schedule because " << *BundleMember << " was already scheduled\n"; } } while (false); | |||
3415 | ReSchedule = true; | |||
3416 | } | |||
3417 | assert(BundleMember->isSchedulingEntity() &&(static_cast <bool> (BundleMember->isSchedulingEntity () && "bundle member already part of other bundle") ? void (0) : __assert_fail ("BundleMember->isSchedulingEntity() && \"bundle member already part of other bundle\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3418, __extension__ __PRETTY_FUNCTION__)) | |||
3418 | "bundle member already part of other bundle")(static_cast <bool> (BundleMember->isSchedulingEntity () && "bundle member already part of other bundle") ? void (0) : __assert_fail ("BundleMember->isSchedulingEntity() && \"bundle member already part of other bundle\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3418, __extension__ __PRETTY_FUNCTION__)); | |||
3419 | if (PrevInBundle) { | |||
3420 | PrevInBundle->NextInBundle = BundleMember; | |||
3421 | } else { | |||
3422 | Bundle = BundleMember; | |||
3423 | } | |||
3424 | BundleMember->UnscheduledDepsInBundle = 0; | |||
3425 | Bundle->UnscheduledDepsInBundle += BundleMember->UnscheduledDeps; | |||
3426 | ||||
3427 | // Group the instructions to a bundle. | |||
3428 | BundleMember->FirstInBundle = Bundle; | |||
3429 | PrevInBundle = BundleMember; | |||
3430 | } | |||
3431 | if (ScheduleEnd != OldScheduleEnd) { | |||
3432 | // The scheduling region got new instructions at the lower end (or it is a | |||
3433 | // new region for the first bundle). This makes it necessary to | |||
3434 | // recalculate all dependencies. | |||
3435 | // It is seldom that this needs to be done a second time after adding the | |||
3436 | // initial bundle to the region. | |||
3437 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | |||
3438 | doForAllOpcodes(I, [](ScheduleData *SD) { | |||
3439 | SD->clearDependencies(); | |||
3440 | }); | |||
3441 | } | |||
3442 | ReSchedule = true; | |||
3443 | } | |||
3444 | if (ReSchedule) { | |||
3445 | resetSchedule(); | |||
3446 | initialFillReadyList(ReadyInsts); | |||
3447 | } | |||
3448 | ||||
3449 | DEBUG(dbgs() << "SLP: try schedule bundle " << *Bundle << " in block "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: try schedule bundle " << *Bundle << " in block " << BB->getName() << "\n"; } } while (false) | |||
3450 | << BB->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: try schedule bundle " << *Bundle << " in block " << BB->getName() << "\n"; } } while (false); | |||
3451 | ||||
3452 | calculateDependencies(Bundle, true, SLP); | |||
3453 | ||||
3454 | // Now try to schedule the new bundle. As soon as the bundle is "ready" it | |||
3455 | // means that there are no cyclic dependencies and we can schedule it. | |||
3456 | // Note that's important that we don't "schedule" the bundle yet (see | |||
3457 | // cancelScheduling). | |||
3458 | while (!Bundle->isReady() && !ReadyInsts.empty()) { | |||
3459 | ||||
3460 | ScheduleData *pickedSD = ReadyInsts.back(); | |||
3461 | ReadyInsts.pop_back(); | |||
3462 | ||||
3463 | if (pickedSD->isSchedulingEntity() && pickedSD->isReady()) { | |||
3464 | schedule(pickedSD, ReadyInsts); | |||
3465 | } | |||
3466 | } | |||
3467 | if (!Bundle->isReady()) { | |||
3468 | cancelScheduling(VL, OpValue); | |||
3469 | return false; | |||
3470 | } | |||
3471 | return true; | |||
3472 | } | |||
3473 | ||||
3474 | void BoUpSLP::BlockScheduling::cancelScheduling(ArrayRef<Value *> VL, | |||
3475 | Value *OpValue) { | |||
3476 | if (isa<PHINode>(OpValue)) | |||
3477 | return; | |||
3478 | ||||
3479 | ScheduleData *Bundle = getScheduleData(OpValue); | |||
3480 | DEBUG(dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n"; } } while (false); | |||
3481 | assert(!Bundle->IsScheduled &&(static_cast <bool> (!Bundle->IsScheduled && "Can't cancel bundle which is already scheduled") ? void (0) : __assert_fail ("!Bundle->IsScheduled && \"Can't cancel bundle which is already scheduled\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3482, __extension__ __PRETTY_FUNCTION__)) | |||
3482 | "Can't cancel bundle which is already scheduled")(static_cast <bool> (!Bundle->IsScheduled && "Can't cancel bundle which is already scheduled") ? void (0) : __assert_fail ("!Bundle->IsScheduled && \"Can't cancel bundle which is already scheduled\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3482, __extension__ __PRETTY_FUNCTION__)); | |||
3483 | assert(Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() &&(static_cast <bool> (Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && "tried to unbundle something which is not a bundle" ) ? void (0) : __assert_fail ("Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && \"tried to unbundle something which is not a bundle\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3484, __extension__ __PRETTY_FUNCTION__)) | |||
3484 | "tried to unbundle something which is not a bundle")(static_cast <bool> (Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && "tried to unbundle something which is not a bundle" ) ? void (0) : __assert_fail ("Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && \"tried to unbundle something which is not a bundle\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3484, __extension__ __PRETTY_FUNCTION__)); | |||
3485 | ||||
3486 | // Un-bundle: make single instructions out of the bundle. | |||
3487 | ScheduleData *BundleMember = Bundle; | |||
3488 | while (BundleMember) { | |||
3489 | assert(BundleMember->FirstInBundle == Bundle && "corrupt bundle links")(static_cast <bool> (BundleMember->FirstInBundle == Bundle && "corrupt bundle links") ? void (0) : __assert_fail ("BundleMember->FirstInBundle == Bundle && \"corrupt bundle links\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3489, __extension__ __PRETTY_FUNCTION__)); | |||
3490 | BundleMember->FirstInBundle = BundleMember; | |||
3491 | ScheduleData *Next = BundleMember->NextInBundle; | |||
3492 | BundleMember->NextInBundle = nullptr; | |||
3493 | BundleMember->UnscheduledDepsInBundle = BundleMember->UnscheduledDeps; | |||
3494 | if (BundleMember->UnscheduledDepsInBundle == 0) { | |||
3495 | ReadyInsts.insert(BundleMember); | |||
3496 | } | |||
3497 | BundleMember = Next; | |||
3498 | } | |||
3499 | } | |||
3500 | ||||
3501 | BoUpSLP::ScheduleData *BoUpSLP::BlockScheduling::allocateScheduleDataChunks() { | |||
3502 | // Allocate a new ScheduleData for the instruction. | |||
3503 | if (ChunkPos >= ChunkSize) { | |||
3504 | ScheduleDataChunks.push_back(llvm::make_unique<ScheduleData[]>(ChunkSize)); | |||
3505 | ChunkPos = 0; | |||
3506 | } | |||
3507 | return &(ScheduleDataChunks.back()[ChunkPos++]); | |||
3508 | } | |||
3509 | ||||
3510 | bool BoUpSLP::BlockScheduling::extendSchedulingRegion(Value *V, | |||
3511 | Value *OpValue) { | |||
3512 | if (getScheduleData(V, isOneOf(OpValue, V))) | |||
3513 | return true; | |||
3514 | Instruction *I = dyn_cast<Instruction>(V); | |||
3515 | assert(I && "bundle member must be an instruction")(static_cast <bool> (I && "bundle member must be an instruction" ) ? void (0) : __assert_fail ("I && \"bundle member must be an instruction\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3515, __extension__ __PRETTY_FUNCTION__)); | |||
3516 | assert(!isa<PHINode>(I) && "phi nodes don't need to be scheduled")(static_cast <bool> (!isa<PHINode>(I) && "phi nodes don't need to be scheduled" ) ? void (0) : __assert_fail ("!isa<PHINode>(I) && \"phi nodes don't need to be scheduled\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3516, __extension__ __PRETTY_FUNCTION__)); | |||
3517 | auto &&CheckSheduleForI = [this, OpValue](Instruction *I) -> bool { | |||
3518 | ScheduleData *ISD = getScheduleData(I); | |||
3519 | if (!ISD) | |||
3520 | return false; | |||
3521 | assert(isInSchedulingRegion(ISD) &&(static_cast <bool> (isInSchedulingRegion(ISD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(ISD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3522, __extension__ __PRETTY_FUNCTION__)) | |||
3522 | "ScheduleData not in scheduling region")(static_cast <bool> (isInSchedulingRegion(ISD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(ISD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3522, __extension__ __PRETTY_FUNCTION__)); | |||
3523 | ScheduleData *SD = allocateScheduleDataChunks(); | |||
3524 | SD->Inst = I; | |||
3525 | SD->init(SchedulingRegionID, OpValue); | |||
3526 | ExtraScheduleDataMap[I][OpValue] = SD; | |||
3527 | return true; | |||
3528 | }; | |||
3529 | if (CheckSheduleForI(I)) | |||
3530 | return true; | |||
3531 | if (!ScheduleStart) { | |||
3532 | // It's the first instruction in the new region. | |||
3533 | initScheduleData(I, I->getNextNode(), nullptr, nullptr); | |||
3534 | ScheduleStart = I; | |||
3535 | ScheduleEnd = I->getNextNode(); | |||
3536 | if (isOneOf(OpValue, I) != I) | |||
3537 | CheckSheduleForI(I); | |||
3538 | assert(ScheduleEnd && "tried to vectorize a TerminatorInst?")(static_cast <bool> (ScheduleEnd && "tried to vectorize a TerminatorInst?" ) ? void (0) : __assert_fail ("ScheduleEnd && \"tried to vectorize a TerminatorInst?\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3538, __extension__ __PRETTY_FUNCTION__)); | |||
3539 | DEBUG(dbgs() << "SLP: initialize schedule region to " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: initialize schedule region to " << *I << "\n"; } } while (false); | |||
3540 | return true; | |||
3541 | } | |||
3542 | // Search up and down at the same time, because we don't know if the new | |||
3543 | // instruction is above or below the existing scheduling region. | |||
3544 | BasicBlock::reverse_iterator UpIter = | |||
3545 | ++ScheduleStart->getIterator().getReverse(); | |||
3546 | BasicBlock::reverse_iterator UpperEnd = BB->rend(); | |||
3547 | BasicBlock::iterator DownIter = ScheduleEnd->getIterator(); | |||
3548 | BasicBlock::iterator LowerEnd = BB->end(); | |||
3549 | while (true) { | |||
3550 | if (++ScheduleRegionSize > ScheduleRegionSizeLimit) { | |||
3551 | DEBUG(dbgs() << "SLP: exceeded schedule region size limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: exceeded schedule region size limit\n" ; } } while (false); | |||
3552 | return false; | |||
3553 | } | |||
3554 | ||||
3555 | if (UpIter != UpperEnd) { | |||
3556 | if (&*UpIter == I) { | |||
3557 | initScheduleData(I, ScheduleStart, nullptr, FirstLoadStoreInRegion); | |||
3558 | ScheduleStart = I; | |||
3559 | if (isOneOf(OpValue, I) != I) | |||
3560 | CheckSheduleForI(I); | |||
3561 | DEBUG(dbgs() << "SLP: extend schedule region start to " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: extend schedule region start to " << *I << "\n"; } } while (false); | |||
3562 | return true; | |||
3563 | } | |||
3564 | UpIter++; | |||
3565 | } | |||
3566 | if (DownIter != LowerEnd) { | |||
3567 | if (&*DownIter == I) { | |||
3568 | initScheduleData(ScheduleEnd, I->getNextNode(), LastLoadStoreInRegion, | |||
3569 | nullptr); | |||
3570 | ScheduleEnd = I->getNextNode(); | |||
3571 | if (isOneOf(OpValue, I) != I) | |||
3572 | CheckSheduleForI(I); | |||
3573 | assert(ScheduleEnd && "tried to vectorize a TerminatorInst?")(static_cast <bool> (ScheduleEnd && "tried to vectorize a TerminatorInst?" ) ? void (0) : __assert_fail ("ScheduleEnd && \"tried to vectorize a TerminatorInst?\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3573, __extension__ __PRETTY_FUNCTION__)); | |||
3574 | DEBUG(dbgs() << "SLP: extend schedule region end to " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: extend schedule region end to " << *I << "\n"; } } while (false); | |||
3575 | return true; | |||
3576 | } | |||
3577 | DownIter++; | |||
3578 | } | |||
3579 | assert((UpIter != UpperEnd || DownIter != LowerEnd) &&(static_cast <bool> ((UpIter != UpperEnd || DownIter != LowerEnd) && "instruction not found in block") ? void (0) : __assert_fail ("(UpIter != UpperEnd || DownIter != LowerEnd) && \"instruction not found in block\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3580, __extension__ __PRETTY_FUNCTION__)) | |||
3580 | "instruction not found in block")(static_cast <bool> ((UpIter != UpperEnd || DownIter != LowerEnd) && "instruction not found in block") ? void (0) : __assert_fail ("(UpIter != UpperEnd || DownIter != LowerEnd) && \"instruction not found in block\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3580, __extension__ __PRETTY_FUNCTION__)); | |||
3581 | } | |||
3582 | return true; | |||
3583 | } | |||
3584 | ||||
3585 | void BoUpSLP::BlockScheduling::initScheduleData(Instruction *FromI, | |||
3586 | Instruction *ToI, | |||
3587 | ScheduleData *PrevLoadStore, | |||
3588 | ScheduleData *NextLoadStore) { | |||
3589 | ScheduleData *CurrentLoadStore = PrevLoadStore; | |||
3590 | for (Instruction *I = FromI; I != ToI; I = I->getNextNode()) { | |||
3591 | ScheduleData *SD = ScheduleDataMap[I]; | |||
3592 | if (!SD) { | |||
3593 | SD = allocateScheduleDataChunks(); | |||
3594 | ScheduleDataMap[I] = SD; | |||
3595 | SD->Inst = I; | |||
3596 | } | |||
3597 | assert(!isInSchedulingRegion(SD) &&(static_cast <bool> (!isInSchedulingRegion(SD) && "new ScheduleData already in scheduling region") ? void (0) : __assert_fail ("!isInSchedulingRegion(SD) && \"new ScheduleData already in scheduling region\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3598, __extension__ __PRETTY_FUNCTION__)) | |||
3598 | "new ScheduleData already in scheduling region")(static_cast <bool> (!isInSchedulingRegion(SD) && "new ScheduleData already in scheduling region") ? void (0) : __assert_fail ("!isInSchedulingRegion(SD) && \"new ScheduleData already in scheduling region\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3598, __extension__ __PRETTY_FUNCTION__)); | |||
3599 | SD->init(SchedulingRegionID, I); | |||
3600 | ||||
3601 | if (I->mayReadOrWriteMemory() && | |||
3602 | (!isa<IntrinsicInst>(I) || | |||
3603 | cast<IntrinsicInst>(I)->getIntrinsicID() != Intrinsic::sideeffect)) { | |||
3604 | // Update the linked list of memory accessing instructions. | |||
3605 | if (CurrentLoadStore) { | |||
3606 | CurrentLoadStore->NextLoadStore = SD; | |||
3607 | } else { | |||
3608 | FirstLoadStoreInRegion = SD; | |||
3609 | } | |||
3610 | CurrentLoadStore = SD; | |||
3611 | } | |||
3612 | } | |||
3613 | if (NextLoadStore) { | |||
3614 | if (CurrentLoadStore) | |||
3615 | CurrentLoadStore->NextLoadStore = NextLoadStore; | |||
3616 | } else { | |||
3617 | LastLoadStoreInRegion = CurrentLoadStore; | |||
3618 | } | |||
3619 | } | |||
3620 | ||||
3621 | void BoUpSLP::BlockScheduling::calculateDependencies(ScheduleData *SD, | |||
3622 | bool InsertInReadyList, | |||
3623 | BoUpSLP *SLP) { | |||
3624 | assert(SD->isSchedulingEntity())(static_cast <bool> (SD->isSchedulingEntity()) ? void (0) : __assert_fail ("SD->isSchedulingEntity()", "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3624, __extension__ __PRETTY_FUNCTION__)); | |||
3625 | ||||
3626 | SmallVector<ScheduleData *, 10> WorkList; | |||
3627 | WorkList.push_back(SD); | |||
3628 | ||||
3629 | while (!WorkList.empty()) { | |||
3630 | ScheduleData *SD = WorkList.back(); | |||
3631 | WorkList.pop_back(); | |||
3632 | ||||
3633 | ScheduleData *BundleMember = SD; | |||
3634 | while (BundleMember) { | |||
3635 | assert(isInSchedulingRegion(BundleMember))(static_cast <bool> (isInSchedulingRegion(BundleMember) ) ? void (0) : __assert_fail ("isInSchedulingRegion(BundleMember)" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3635, __extension__ __PRETTY_FUNCTION__)); | |||
3636 | if (!BundleMember->hasValidDependencies()) { | |||
3637 | ||||
3638 | DEBUG(dbgs() << "SLP: update deps of " << *BundleMember << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: update deps of " << *BundleMember << "\n"; } } while (false); | |||
3639 | BundleMember->Dependencies = 0; | |||
3640 | BundleMember->resetUnscheduledDeps(); | |||
3641 | ||||
3642 | // Handle def-use chain dependencies. | |||
3643 | if (BundleMember->OpValue != BundleMember->Inst) { | |||
3644 | ScheduleData *UseSD = getScheduleData(BundleMember->Inst); | |||
3645 | if (UseSD && isInSchedulingRegion(UseSD->FirstInBundle)) { | |||
3646 | BundleMember->Dependencies++; | |||
3647 | ScheduleData *DestBundle = UseSD->FirstInBundle; | |||
3648 | if (!DestBundle->IsScheduled) | |||
3649 | BundleMember->incrementUnscheduledDeps(1); | |||
3650 | if (!DestBundle->hasValidDependencies()) | |||
3651 | WorkList.push_back(DestBundle); | |||
3652 | } | |||
3653 | } else { | |||
3654 | for (User *U : BundleMember->Inst->users()) { | |||
3655 | if (isa<Instruction>(U)) { | |||
3656 | ScheduleData *UseSD = getScheduleData(U); | |||
3657 | if (UseSD && isInSchedulingRegion(UseSD->FirstInBundle)) { | |||
3658 | BundleMember->Dependencies++; | |||
3659 | ScheduleData *DestBundle = UseSD->FirstInBundle; | |||
3660 | if (!DestBundle->IsScheduled) | |||
3661 | BundleMember->incrementUnscheduledDeps(1); | |||
3662 | if (!DestBundle->hasValidDependencies()) | |||
3663 | WorkList.push_back(DestBundle); | |||
3664 | } | |||
3665 | } else { | |||
3666 | // I'm not sure if this can ever happen. But we need to be safe. | |||
3667 | // This lets the instruction/bundle never be scheduled and | |||
3668 | // eventually disable vectorization. | |||
3669 | BundleMember->Dependencies++; | |||
3670 | BundleMember->incrementUnscheduledDeps(1); | |||
3671 | } | |||
3672 | } | |||
3673 | } | |||
3674 | ||||
3675 | // Handle the memory dependencies. | |||
3676 | ScheduleData *DepDest = BundleMember->NextLoadStore; | |||
3677 | if (DepDest) { | |||
3678 | Instruction *SrcInst = BundleMember->Inst; | |||
3679 | MemoryLocation SrcLoc = getLocation(SrcInst, SLP->AA); | |||
3680 | bool SrcMayWrite = BundleMember->Inst->mayWriteToMemory(); | |||
3681 | unsigned numAliased = 0; | |||
3682 | unsigned DistToSrc = 1; | |||
3683 | ||||
3684 | while (DepDest) { | |||
3685 | assert(isInSchedulingRegion(DepDest))(static_cast <bool> (isInSchedulingRegion(DepDest)) ? void (0) : __assert_fail ("isInSchedulingRegion(DepDest)", "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3685, __extension__ __PRETTY_FUNCTION__)); | |||
3686 | ||||
3687 | // We have two limits to reduce the complexity: | |||
3688 | // 1) AliasedCheckLimit: It's a small limit to reduce calls to | |||
3689 | // SLP->isAliased (which is the expensive part in this loop). | |||
3690 | // 2) MaxMemDepDistance: It's for very large blocks and it aborts | |||
3691 | // the whole loop (even if the loop is fast, it's quadratic). | |||
3692 | // It's important for the loop break condition (see below) to | |||
3693 | // check this limit even between two read-only instructions. | |||
3694 | if (DistToSrc >= MaxMemDepDistance || | |||
3695 | ((SrcMayWrite || DepDest->Inst->mayWriteToMemory()) && | |||
3696 | (numAliased >= AliasedCheckLimit || | |||
3697 | SLP->isAliased(SrcLoc, SrcInst, DepDest->Inst)))) { | |||
3698 | ||||
3699 | // We increment the counter only if the locations are aliased | |||
3700 | // (instead of counting all alias checks). This gives a better | |||
3701 | // balance between reduced runtime and accurate dependencies. | |||
3702 | numAliased++; | |||
3703 | ||||
3704 | DepDest->MemoryDependencies.push_back(BundleMember); | |||
3705 | BundleMember->Dependencies++; | |||
3706 | ScheduleData *DestBundle = DepDest->FirstInBundle; | |||
3707 | if (!DestBundle->IsScheduled) { | |||
3708 | BundleMember->incrementUnscheduledDeps(1); | |||
3709 | } | |||
3710 | if (!DestBundle->hasValidDependencies()) { | |||
3711 | WorkList.push_back(DestBundle); | |||
3712 | } | |||
3713 | } | |||
3714 | DepDest = DepDest->NextLoadStore; | |||
3715 | ||||
3716 | // Example, explaining the loop break condition: Let's assume our | |||
3717 | // starting instruction is i0 and MaxMemDepDistance = 3. | |||
3718 | // | |||
3719 | // +--------v--v--v | |||
3720 | // i0,i1,i2,i3,i4,i5,i6,i7,i8 | |||
3721 | // +--------^--^--^ | |||
3722 | // | |||
3723 | // MaxMemDepDistance let us stop alias-checking at i3 and we add | |||
3724 | // dependencies from i0 to i3,i4,.. (even if they are not aliased). | |||
3725 | // Previously we already added dependencies from i3 to i6,i7,i8 | |||
3726 | // (because of MaxMemDepDistance). As we added a dependency from | |||
3727 | // i0 to i3, we have transitive dependencies from i0 to i6,i7,i8 | |||
3728 | // and we can abort this loop at i6. | |||
3729 | if (DistToSrc >= 2 * MaxMemDepDistance) | |||
3730 | break; | |||
3731 | DistToSrc++; | |||
3732 | } | |||
3733 | } | |||
3734 | } | |||
3735 | BundleMember = BundleMember->NextInBundle; | |||
3736 | } | |||
3737 | if (InsertInReadyList && SD->isReady()) { | |||
3738 | ReadyInsts.push_back(SD); | |||
3739 | DEBUG(dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n"; } } while (false); | |||
3740 | } | |||
3741 | } | |||
3742 | } | |||
3743 | ||||
3744 | void BoUpSLP::BlockScheduling::resetSchedule() { | |||
3745 | assert(ScheduleStart &&(static_cast <bool> (ScheduleStart && "tried to reset schedule on block which has not been scheduled" ) ? void (0) : __assert_fail ("ScheduleStart && \"tried to reset schedule on block which has not been scheduled\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3746, __extension__ __PRETTY_FUNCTION__)) | |||
3746 | "tried to reset schedule on block which has not been scheduled")(static_cast <bool> (ScheduleStart && "tried to reset schedule on block which has not been scheduled" ) ? void (0) : __assert_fail ("ScheduleStart && \"tried to reset schedule on block which has not been scheduled\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3746, __extension__ __PRETTY_FUNCTION__)); | |||
3747 | for (Instruction *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | |||
3748 | doForAllOpcodes(I, [&](ScheduleData *SD) { | |||
3749 | assert(isInSchedulingRegion(SD) &&(static_cast <bool> (isInSchedulingRegion(SD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(SD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3750, __extension__ __PRETTY_FUNCTION__)) | |||
3750 | "ScheduleData not in scheduling region")(static_cast <bool> (isInSchedulingRegion(SD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(SD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3750, __extension__ __PRETTY_FUNCTION__)); | |||
3751 | SD->IsScheduled = false; | |||
3752 | SD->resetUnscheduledDeps(); | |||
3753 | }); | |||
3754 | } | |||
3755 | ReadyInsts.clear(); | |||
3756 | } | |||
3757 | ||||
3758 | void BoUpSLP::scheduleBlock(BlockScheduling *BS) { | |||
3759 | if (!BS->ScheduleStart) | |||
3760 | return; | |||
3761 | ||||
3762 | DEBUG(dbgs() << "SLP: schedule block " << BS->BB->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: schedule block " << BS ->BB->getName() << "\n"; } } while (false); | |||
3763 | ||||
3764 | BS->resetSchedule(); | |||
3765 | ||||
3766 | // For the real scheduling we use a more sophisticated ready-list: it is | |||
3767 | // sorted by the original instruction location. This lets the final schedule | |||
3768 | // be as close as possible to the original instruction order. | |||
3769 | struct ScheduleDataCompare { | |||
3770 | bool operator()(ScheduleData *SD1, ScheduleData *SD2) const { | |||
3771 | return SD2->SchedulingPriority < SD1->SchedulingPriority; | |||
3772 | } | |||
3773 | }; | |||
3774 | std::set<ScheduleData *, ScheduleDataCompare> ReadyInsts; | |||
3775 | ||||
3776 | // Ensure that all dependency data is updated and fill the ready-list with | |||
3777 | // initial instructions. | |||
3778 | int Idx = 0; | |||
3779 | int NumToSchedule = 0; | |||
3780 | for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd; | |||
3781 | I = I->getNextNode()) { | |||
3782 | BS->doForAllOpcodes(I, [this, &Idx, &NumToSchedule, BS](ScheduleData *SD) { | |||
3783 | assert(SD->isPartOfBundle() ==(static_cast <bool> (SD->isPartOfBundle() == (getTreeEntry (SD->Inst) != nullptr) && "scheduler and vectorizer bundle mismatch" ) ? void (0) : __assert_fail ("SD->isPartOfBundle() == (getTreeEntry(SD->Inst) != nullptr) && \"scheduler and vectorizer bundle mismatch\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3785, __extension__ __PRETTY_FUNCTION__)) | |||
3784 | (getTreeEntry(SD->Inst) != nullptr) &&(static_cast <bool> (SD->isPartOfBundle() == (getTreeEntry (SD->Inst) != nullptr) && "scheduler and vectorizer bundle mismatch" ) ? void (0) : __assert_fail ("SD->isPartOfBundle() == (getTreeEntry(SD->Inst) != nullptr) && \"scheduler and vectorizer bundle mismatch\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3785, __extension__ __PRETTY_FUNCTION__)) | |||
3785 | "scheduler and vectorizer bundle mismatch")(static_cast <bool> (SD->isPartOfBundle() == (getTreeEntry (SD->Inst) != nullptr) && "scheduler and vectorizer bundle mismatch" ) ? void (0) : __assert_fail ("SD->isPartOfBundle() == (getTreeEntry(SD->Inst) != nullptr) && \"scheduler and vectorizer bundle mismatch\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3785, __extension__ __PRETTY_FUNCTION__)); | |||
3786 | SD->FirstInBundle->SchedulingPriority = Idx++; | |||
3787 | if (SD->isSchedulingEntity()) { | |||
3788 | BS->calculateDependencies(SD, false, this); | |||
3789 | NumToSchedule++; | |||
3790 | } | |||
3791 | }); | |||
3792 | } | |||
3793 | BS->initialFillReadyList(ReadyInsts); | |||
3794 | ||||
3795 | Instruction *LastScheduledInst = BS->ScheduleEnd; | |||
3796 | ||||
3797 | // Do the "real" scheduling. | |||
3798 | while (!ReadyInsts.empty()) { | |||
3799 | ScheduleData *picked = *ReadyInsts.begin(); | |||
3800 | ReadyInsts.erase(ReadyInsts.begin()); | |||
3801 | ||||
3802 | // Move the scheduled instruction(s) to their dedicated places, if not | |||
3803 | // there yet. | |||
3804 | ScheduleData *BundleMember = picked; | |||
3805 | while (BundleMember) { | |||
3806 | Instruction *pickedInst = BundleMember->Inst; | |||
3807 | if (LastScheduledInst->getNextNode() != pickedInst) { | |||
3808 | BS->BB->getInstList().remove(pickedInst); | |||
3809 | BS->BB->getInstList().insert(LastScheduledInst->getIterator(), | |||
3810 | pickedInst); | |||
3811 | } | |||
3812 | LastScheduledInst = pickedInst; | |||
3813 | BundleMember = BundleMember->NextInBundle; | |||
3814 | } | |||
3815 | ||||
3816 | BS->schedule(picked, ReadyInsts); | |||
3817 | NumToSchedule--; | |||
3818 | } | |||
3819 | assert(NumToSchedule == 0 && "could not schedule all instructions")(static_cast <bool> (NumToSchedule == 0 && "could not schedule all instructions" ) ? void (0) : __assert_fail ("NumToSchedule == 0 && \"could not schedule all instructions\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3819, __extension__ __PRETTY_FUNCTION__)); | |||
3820 | ||||
3821 | // Avoid duplicate scheduling of the block. | |||
3822 | BS->ScheduleStart = nullptr; | |||
3823 | } | |||
3824 | ||||
3825 | unsigned BoUpSLP::getVectorElementSize(Value *V) { | |||
3826 | // If V is a store, just return the width of the stored value without | |||
3827 | // traversing the expression tree. This is the common case. | |||
3828 | if (auto *Store = dyn_cast<StoreInst>(V)) | |||
3829 | return DL->getTypeSizeInBits(Store->getValueOperand()->getType()); | |||
3830 | ||||
3831 | // If V is not a store, we can traverse the expression tree to find loads | |||
3832 | // that feed it. The type of the loaded value may indicate a more suitable | |||
3833 | // width than V's type. We want to base the vector element size on the width | |||
3834 | // of memory operations where possible. | |||
3835 | SmallVector<Instruction *, 16> Worklist; | |||
3836 | SmallPtrSet<Instruction *, 16> Visited; | |||
3837 | if (auto *I = dyn_cast<Instruction>(V)) | |||
3838 | Worklist.push_back(I); | |||
3839 | ||||
3840 | // Traverse the expression tree in bottom-up order looking for loads. If we | |||
3841 | // encounter an instruciton we don't yet handle, we give up. | |||
3842 | auto MaxWidth = 0u; | |||
3843 | auto FoundUnknownInst = false; | |||
3844 | while (!Worklist.empty() && !FoundUnknownInst) { | |||
3845 | auto *I = Worklist.pop_back_val(); | |||
3846 | Visited.insert(I); | |||
3847 | ||||
3848 | // We should only be looking at scalar instructions here. If the current | |||
3849 | // instruction has a vector type, give up. | |||
3850 | auto *Ty = I->getType(); | |||
3851 | if (isa<VectorType>(Ty)) | |||
3852 | FoundUnknownInst = true; | |||
3853 | ||||
3854 | // If the current instruction is a load, update MaxWidth to reflect the | |||
3855 | // width of the loaded value. | |||
3856 | else if (isa<LoadInst>(I)) | |||
3857 | MaxWidth = std::max<unsigned>(MaxWidth, DL->getTypeSizeInBits(Ty)); | |||
3858 | ||||
3859 | // Otherwise, we need to visit the operands of the instruction. We only | |||
3860 | // handle the interesting cases from buildTree here. If an operand is an | |||
3861 | // instruction we haven't yet visited, we add it to the worklist. | |||
3862 | else if (isa<PHINode>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I) || | |||
3863 | isa<CmpInst>(I) || isa<SelectInst>(I) || isa<BinaryOperator>(I)) { | |||
3864 | for (Use &U : I->operands()) | |||
3865 | if (auto *J = dyn_cast<Instruction>(U.get())) | |||
3866 | if (!Visited.count(J)) | |||
3867 | Worklist.push_back(J); | |||
3868 | } | |||
3869 | ||||
3870 | // If we don't yet handle the instruction, give up. | |||
3871 | else | |||
3872 | FoundUnknownInst = true; | |||
3873 | } | |||
3874 | ||||
3875 | // If we didn't encounter a memory access in the expression tree, or if we | |||
3876 | // gave up for some reason, just return the width of V. | |||
3877 | if (!MaxWidth || FoundUnknownInst) | |||
3878 | return DL->getTypeSizeInBits(V->getType()); | |||
3879 | ||||
3880 | // Otherwise, return the maximum width we found. | |||
3881 | return MaxWidth; | |||
3882 | } | |||
3883 | ||||
3884 | // Determine if a value V in a vectorizable expression Expr can be demoted to a | |||
3885 | // smaller type with a truncation. We collect the values that will be demoted | |||
3886 | // in ToDemote and additional roots that require investigating in Roots. | |||
3887 | static bool collectValuesToDemote(Value *V, SmallPtrSetImpl<Value *> &Expr, | |||
3888 | SmallVectorImpl<Value *> &ToDemote, | |||
3889 | SmallVectorImpl<Value *> &Roots) { | |||
3890 | // We can always demote constants. | |||
3891 | if (isa<Constant>(V)) { | |||
3892 | ToDemote.push_back(V); | |||
3893 | return true; | |||
3894 | } | |||
3895 | ||||
3896 | // If the value is not an instruction in the expression with only one use, it | |||
3897 | // cannot be demoted. | |||
3898 | auto *I = dyn_cast<Instruction>(V); | |||
3899 | if (!I || !I->hasOneUse() || !Expr.count(I)) | |||
3900 | return false; | |||
3901 | ||||
3902 | switch (I->getOpcode()) { | |||
3903 | ||||
3904 | // We can always demote truncations and extensions. Since truncations can | |||
3905 | // seed additional demotion, we save the truncated value. | |||
3906 | case Instruction::Trunc: | |||
3907 | Roots.push_back(I->getOperand(0)); | |||
3908 | break; | |||
3909 | case Instruction::ZExt: | |||
3910 | case Instruction::SExt: | |||
3911 | break; | |||
3912 | ||||
3913 | // We can demote certain binary operations if we can demote both of their | |||
3914 | // operands. | |||
3915 | case Instruction::Add: | |||
3916 | case Instruction::Sub: | |||
3917 | case Instruction::Mul: | |||
3918 | case Instruction::And: | |||
3919 | case Instruction::Or: | |||
3920 | case Instruction::Xor: | |||
3921 | if (!collectValuesToDemote(I->getOperand(0), Expr, ToDemote, Roots) || | |||
3922 | !collectValuesToDemote(I->getOperand(1), Expr, ToDemote, Roots)) | |||
3923 | return false; | |||
3924 | break; | |||
3925 | ||||
3926 | // We can demote selects if we can demote their true and false values. | |||
3927 | case Instruction::Select: { | |||
3928 | SelectInst *SI = cast<SelectInst>(I); | |||
3929 | if (!collectValuesToDemote(SI->getTrueValue(), Expr, ToDemote, Roots) || | |||
3930 | !collectValuesToDemote(SI->getFalseValue(), Expr, ToDemote, Roots)) | |||
3931 | return false; | |||
3932 | break; | |||
3933 | } | |||
3934 | ||||
3935 | // We can demote phis if we can demote all their incoming operands. Note that | |||
3936 | // we don't need to worry about cycles since we ensure single use above. | |||
3937 | case Instruction::PHI: { | |||
3938 | PHINode *PN = cast<PHINode>(I); | |||
3939 | for (Value *IncValue : PN->incoming_values()) | |||
3940 | if (!collectValuesToDemote(IncValue, Expr, ToDemote, Roots)) | |||
3941 | return false; | |||
3942 | break; | |||
3943 | } | |||
3944 | ||||
3945 | // Otherwise, conservatively give up. | |||
3946 | default: | |||
3947 | return false; | |||
3948 | } | |||
3949 | ||||
3950 | // Record the value that we can demote. | |||
3951 | ToDemote.push_back(V); | |||
3952 | return true; | |||
3953 | } | |||
3954 | ||||
3955 | void BoUpSLP::computeMinimumValueSizes() { | |||
3956 | // If there are no external uses, the expression tree must be rooted by a | |||
3957 | // store. We can't demote in-memory values, so there is nothing to do here. | |||
3958 | if (ExternalUses.empty()) | |||
3959 | return; | |||
3960 | ||||
3961 | // We only attempt to truncate integer expressions. | |||
3962 | auto &TreeRoot = VectorizableTree[0].Scalars; | |||
3963 | auto *TreeRootIT = dyn_cast<IntegerType>(TreeRoot[0]->getType()); | |||
3964 | if (!TreeRootIT) | |||
3965 | return; | |||
3966 | ||||
3967 | // If the expression is not rooted by a store, these roots should have | |||
3968 | // external uses. We will rely on InstCombine to rewrite the expression in | |||
3969 | // the narrower type. However, InstCombine only rewrites single-use values. | |||
3970 | // This means that if a tree entry other than a root is used externally, it | |||
3971 | // must have multiple uses and InstCombine will not rewrite it. The code | |||
3972 | // below ensures that only the roots are used externally. | |||
3973 | SmallPtrSet<Value *, 32> Expr(TreeRoot.begin(), TreeRoot.end()); | |||
3974 | for (auto &EU : ExternalUses) | |||
3975 | if (!Expr.erase(EU.Scalar)) | |||
3976 | return; | |||
3977 | if (!Expr.empty()) | |||
3978 | return; | |||
3979 | ||||
3980 | // Collect the scalar values of the vectorizable expression. We will use this | |||
3981 | // context to determine which values can be demoted. If we see a truncation, | |||
3982 | // we mark it as seeding another demotion. | |||
3983 | for (auto &Entry : VectorizableTree) | |||
3984 | Expr.insert(Entry.Scalars.begin(), Entry.Scalars.end()); | |||
3985 | ||||
3986 | // Ensure the roots of the vectorizable tree don't form a cycle. They must | |||
3987 | // have a single external user that is not in the vectorizable tree. | |||
3988 | for (auto *Root : TreeRoot) | |||
3989 | if (!Root->hasOneUse() || Expr.count(*Root->user_begin())) | |||
3990 | return; | |||
3991 | ||||
3992 | // Conservatively determine if we can actually truncate the roots of the | |||
3993 | // expression. Collect the values that can be demoted in ToDemote and | |||
3994 | // additional roots that require investigating in Roots. | |||
3995 | SmallVector<Value *, 32> ToDemote; | |||
3996 | SmallVector<Value *, 4> Roots; | |||
3997 | for (auto *Root : TreeRoot) | |||
3998 | if (!collectValuesToDemote(Root, Expr, ToDemote, Roots)) | |||
3999 | return; | |||
4000 | ||||
4001 | // The maximum bit width required to represent all the values that can be | |||
4002 | // demoted without loss of precision. It would be safe to truncate the roots | |||
4003 | // of the expression to this width. | |||
4004 | auto MaxBitWidth = 8u; | |||
4005 | ||||
4006 | // We first check if all the bits of the roots are demanded. If they're not, | |||
4007 | // we can truncate the roots to this narrower type. | |||
4008 | for (auto *Root : TreeRoot) { | |||
4009 | auto Mask = DB->getDemandedBits(cast<Instruction>(Root)); | |||
4010 | MaxBitWidth = std::max<unsigned>( | |||
4011 | Mask.getBitWidth() - Mask.countLeadingZeros(), MaxBitWidth); | |||
4012 | } | |||
4013 | ||||
4014 | // True if the roots can be zero-extended back to their original type, rather | |||
4015 | // than sign-extended. We know that if the leading bits are not demanded, we | |||
4016 | // can safely zero-extend. So we initialize IsKnownPositive to True. | |||
4017 | bool IsKnownPositive = true; | |||
4018 | ||||
4019 | // If all the bits of the roots are demanded, we can try a little harder to | |||
4020 | // compute a narrower type. This can happen, for example, if the roots are | |||
4021 | // getelementptr indices. InstCombine promotes these indices to the pointer | |||
4022 | // width. Thus, all their bits are technically demanded even though the | |||
4023 | // address computation might be vectorized in a smaller type. | |||
4024 | // | |||
4025 | // We start by looking at each entry that can be demoted. We compute the | |||
4026 | // maximum bit width required to store the scalar by using ValueTracking to | |||
4027 | // compute the number of high-order bits we can truncate. | |||
4028 | if (MaxBitWidth == DL->getTypeSizeInBits(TreeRoot[0]->getType())) { | |||
4029 | MaxBitWidth = 8u; | |||
4030 | ||||
4031 | // Determine if the sign bit of all the roots is known to be zero. If not, | |||
4032 | // IsKnownPositive is set to False. | |||
4033 | IsKnownPositive = llvm::all_of(TreeRoot, [&](Value *R) { | |||
4034 | KnownBits Known = computeKnownBits(R, *DL); | |||
4035 | return Known.isNonNegative(); | |||
4036 | }); | |||
4037 | ||||
4038 | // Determine the maximum number of bits required to store the scalar | |||
4039 | // values. | |||
4040 | for (auto *Scalar : ToDemote) { | |||
4041 | auto NumSignBits = ComputeNumSignBits(Scalar, *DL, 0, AC, nullptr, DT); | |||
4042 | auto NumTypeBits = DL->getTypeSizeInBits(Scalar->getType()); | |||
4043 | MaxBitWidth = std::max<unsigned>(NumTypeBits - NumSignBits, MaxBitWidth); | |||
4044 | } | |||
4045 | ||||
4046 | // If we can't prove that the sign bit is zero, we must add one to the | |||
4047 | // maximum bit width to account for the unknown sign bit. This preserves | |||
4048 | // the existing sign bit so we can safely sign-extend the root back to the | |||
4049 | // original type. Otherwise, if we know the sign bit is zero, we will | |||
4050 | // zero-extend the root instead. | |||
4051 | // | |||
4052 | // FIXME: This is somewhat suboptimal, as there will be cases where adding | |||
4053 | // one to the maximum bit width will yield a larger-than-necessary | |||
4054 | // type. In general, we need to add an extra bit only if we can't | |||
4055 | // prove that the upper bit of the original type is equal to the | |||
4056 | // upper bit of the proposed smaller type. If these two bits are the | |||
4057 | // same (either zero or one) we know that sign-extending from the | |||
4058 | // smaller type will result in the same value. Here, since we can't | |||
4059 | // yet prove this, we are just making the proposed smaller type | |||
4060 | // larger to ensure correctness. | |||
4061 | if (!IsKnownPositive) | |||
4062 | ++MaxBitWidth; | |||
4063 | } | |||
4064 | ||||
4065 | // Round MaxBitWidth up to the next power-of-two. | |||
4066 | if (!isPowerOf2_64(MaxBitWidth)) | |||
4067 | MaxBitWidth = NextPowerOf2(MaxBitWidth); | |||
4068 | ||||
4069 | // If the maximum bit width we compute is less than the with of the roots' | |||
4070 | // type, we can proceed with the narrowing. Otherwise, do nothing. | |||
4071 | if (MaxBitWidth >= TreeRootIT->getBitWidth()) | |||
4072 | return; | |||
4073 | ||||
4074 | // If we can truncate the root, we must collect additional values that might | |||
4075 | // be demoted as a result. That is, those seeded by truncations we will | |||
4076 | // modify. | |||
4077 | while (!Roots.empty()) | |||
4078 | collectValuesToDemote(Roots.pop_back_val(), Expr, ToDemote, Roots); | |||
4079 | ||||
4080 | // Finally, map the values we can demote to the maximum bit with we computed. | |||
4081 | for (auto *Scalar : ToDemote) | |||
4082 | MinBWs[Scalar] = std::make_pair(MaxBitWidth, !IsKnownPositive); | |||
4083 | } | |||
4084 | ||||
4085 | namespace { | |||
4086 | ||||
4087 | /// The SLPVectorizer Pass. | |||
4088 | struct SLPVectorizer : public FunctionPass { | |||
4089 | SLPVectorizerPass Impl; | |||
4090 | ||||
4091 | /// Pass identification, replacement for typeid | |||
4092 | static char ID; | |||
4093 | ||||
4094 | explicit SLPVectorizer() : FunctionPass(ID) { | |||
4095 | initializeSLPVectorizerPass(*PassRegistry::getPassRegistry()); | |||
4096 | } | |||
4097 | ||||
4098 | bool doInitialization(Module &M) override { | |||
4099 | return false; | |||
4100 | } | |||
4101 | ||||
4102 | bool runOnFunction(Function &F) override { | |||
4103 | if (skipFunction(F)) | |||
4104 | return false; | |||
4105 | ||||
4106 | auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); | |||
4107 | auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | |||
4108 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); | |||
4109 | auto *TLI = TLIP ? &TLIP->getTLI() : nullptr; | |||
4110 | auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); | |||
4111 | auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | |||
4112 | auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | |||
4113 | auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | |||
4114 | auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits(); | |||
4115 | auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); | |||
4116 | ||||
4117 | return Impl.runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); | |||
4118 | } | |||
4119 | ||||
4120 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
4121 | FunctionPass::getAnalysisUsage(AU); | |||
4122 | AU.addRequired<AssumptionCacheTracker>(); | |||
4123 | AU.addRequired<ScalarEvolutionWrapperPass>(); | |||
4124 | AU.addRequired<AAResultsWrapperPass>(); | |||
4125 | AU.addRequired<TargetTransformInfoWrapperPass>(); | |||
4126 | AU.addRequired<LoopInfoWrapperPass>(); | |||
4127 | AU.addRequired<DominatorTreeWrapperPass>(); | |||
4128 | AU.addRequired<DemandedBitsWrapperPass>(); | |||
4129 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); | |||
4130 | AU.addPreserved<LoopInfoWrapperPass>(); | |||
4131 | AU.addPreserved<DominatorTreeWrapperPass>(); | |||
4132 | AU.addPreserved<AAResultsWrapperPass>(); | |||
4133 | AU.addPreserved<GlobalsAAWrapperPass>(); | |||
4134 | AU.setPreservesCFG(); | |||
4135 | } | |||
4136 | }; | |||
4137 | ||||
4138 | } // end anonymous namespace | |||
4139 | ||||
4140 | PreservedAnalyses SLPVectorizerPass::run(Function &F, FunctionAnalysisManager &AM) { | |||
4141 | auto *SE = &AM.getResult<ScalarEvolutionAnalysis>(F); | |||
4142 | auto *TTI = &AM.getResult<TargetIRAnalysis>(F); | |||
4143 | auto *TLI = AM.getCachedResult<TargetLibraryAnalysis>(F); | |||
4144 | auto *AA = &AM.getResult<AAManager>(F); | |||
4145 | auto *LI = &AM.getResult<LoopAnalysis>(F); | |||
4146 | auto *DT = &AM.getResult<DominatorTreeAnalysis>(F); | |||
4147 | auto *AC = &AM.getResult<AssumptionAnalysis>(F); | |||
4148 | auto *DB = &AM.getResult<DemandedBitsAnalysis>(F); | |||
4149 | auto *ORE = &AM.getResult<OptimizationRemarkEmitterAnalysis>(F); | |||
4150 | ||||
4151 | bool Changed = runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); | |||
4152 | if (!Changed) | |||
4153 | return PreservedAnalyses::all(); | |||
4154 | ||||
4155 | PreservedAnalyses PA; | |||
4156 | PA.preserveSet<CFGAnalyses>(); | |||
4157 | PA.preserve<AAManager>(); | |||
4158 | PA.preserve<GlobalsAA>(); | |||
4159 | return PA; | |||
4160 | } | |||
4161 | ||||
4162 | bool SLPVectorizerPass::runImpl(Function &F, ScalarEvolution *SE_, | |||
4163 | TargetTransformInfo *TTI_, | |||
4164 | TargetLibraryInfo *TLI_, AliasAnalysis *AA_, | |||
4165 | LoopInfo *LI_, DominatorTree *DT_, | |||
4166 | AssumptionCache *AC_, DemandedBits *DB_, | |||
4167 | OptimizationRemarkEmitter *ORE_) { | |||
4168 | SE = SE_; | |||
4169 | TTI = TTI_; | |||
4170 | TLI = TLI_; | |||
4171 | AA = AA_; | |||
4172 | LI = LI_; | |||
4173 | DT = DT_; | |||
4174 | AC = AC_; | |||
4175 | DB = DB_; | |||
4176 | DL = &F.getParent()->getDataLayout(); | |||
4177 | ||||
4178 | Stores.clear(); | |||
4179 | GEPs.clear(); | |||
4180 | bool Changed = false; | |||
4181 | ||||
4182 | // If the target claims to have no vector registers don't attempt | |||
4183 | // vectorization. | |||
4184 | if (!TTI->getNumberOfRegisters(true)) | |||
4185 | return false; | |||
4186 | ||||
4187 | // Don't vectorize when the attribute NoImplicitFloat is used. | |||
4188 | if (F.hasFnAttribute(Attribute::NoImplicitFloat)) | |||
4189 | return false; | |||
4190 | ||||
4191 | DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n"; } } while (false); | |||
4192 | ||||
4193 | // Use the bottom up slp vectorizer to construct chains that start with | |||
4194 | // store instructions. | |||
4195 | BoUpSLP R(&F, SE, TTI, TLI, AA, LI, DT, AC, DB, DL, ORE_); | |||
4196 | ||||
4197 | // A general note: the vectorizer must use BoUpSLP::eraseInstruction() to | |||
4198 | // delete instructions. | |||
4199 | ||||
4200 | // Scan the blocks in the function in post order. | |||
4201 | for (auto BB : post_order(&F.getEntryBlock())) { | |||
4202 | collectSeedInstructions(BB); | |||
4203 | ||||
4204 | // Vectorize trees that end at stores. | |||
4205 | if (!Stores.empty()) { | |||
4206 | DEBUG(dbgs() << "SLP: Found stores for " << Stores.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found stores for " << Stores .size() << " underlying objects.\n"; } } while (false) | |||
4207 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found stores for " << Stores .size() << " underlying objects.\n"; } } while (false); | |||
4208 | Changed |= vectorizeStoreChains(R); | |||
4209 | } | |||
4210 | ||||
4211 | // Vectorize trees that end at reductions. | |||
4212 | Changed |= vectorizeChainsInBlock(BB, R); | |||
4213 | ||||
4214 | // Vectorize the index computations of getelementptr instructions. This | |||
4215 | // is primarily intended to catch gather-like idioms ending at | |||
4216 | // non-consecutive loads. | |||
4217 | if (!GEPs.empty()) { | |||
4218 | DEBUG(dbgs() << "SLP: Found GEPs for " << GEPs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found GEPs for " << GEPs .size() << " underlying objects.\n"; } } while (false) | |||
4219 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found GEPs for " << GEPs .size() << " underlying objects.\n"; } } while (false); | |||
4220 | Changed |= vectorizeGEPIndices(BB, R); | |||
4221 | } | |||
4222 | } | |||
4223 | ||||
4224 | if (Changed) { | |||
4225 | R.optimizeGatherSequence(F); | |||
4226 | DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: vectorized \"" << F.getName () << "\"\n"; } } while (false); | |||
4227 | DEBUG(verifyFunction(F))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { verifyFunction(F); } } while (false); | |||
4228 | } | |||
4229 | return Changed; | |||
4230 | } | |||
4231 | ||||
4232 | /// \brief Check that the Values in the slice in VL array are still existent in | |||
4233 | /// the WeakTrackingVH array. | |||
4234 | /// Vectorization of part of the VL array may cause later values in the VL array | |||
4235 | /// to become invalid. We track when this has happened in the WeakTrackingVH | |||
4236 | /// array. | |||
4237 | static bool hasValueBeenRAUWed(ArrayRef<Value *> VL, | |||
4238 | ArrayRef<WeakTrackingVH> VH, unsigned SliceBegin, | |||
4239 | unsigned SliceSize) { | |||
4240 | VL = VL.slice(SliceBegin, SliceSize); | |||
4241 | VH = VH.slice(SliceBegin, SliceSize); | |||
4242 | return !std::equal(VL.begin(), VL.end(), VH.begin()); | |||
4243 | } | |||
4244 | ||||
4245 | bool SLPVectorizerPass::vectorizeStoreChain(ArrayRef<Value *> Chain, BoUpSLP &R, | |||
4246 | unsigned VecRegSize) { | |||
4247 | unsigned ChainLen = Chain.size(); | |||
4248 | DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLendo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << ChainLen << "\n"; } } while (false) | |||
4249 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << ChainLen << "\n"; } } while (false); | |||
4250 | unsigned Sz = R.getVectorElementSize(Chain[0]); | |||
4251 | unsigned VF = VecRegSize / Sz; | |||
4252 | ||||
4253 | if (!isPowerOf2_32(Sz) || VF < 2) | |||
4254 | return false; | |||
4255 | ||||
4256 | // Keep track of values that were deleted by vectorizing in the loop below. | |||
4257 | SmallVector<WeakTrackingVH, 8> TrackValues(Chain.begin(), Chain.end()); | |||
4258 | ||||
4259 | bool Changed = false; | |||
4260 | // Look for profitable vectorizable trees at all offsets, starting at zero. | |||
4261 | for (unsigned i = 0, e = ChainLen; i < e; ++i) { | |||
4262 | if (i + VF > e) | |||
4263 | break; | |||
4264 | ||||
4265 | // Check that a previous iteration of this loop did not delete the Value. | |||
4266 | if (hasValueBeenRAUWed(Chain, TrackValues, i, VF)) | |||
4267 | continue; | |||
4268 | ||||
4269 | DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i << "\n"; } } while (false ) | |||
4270 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i << "\n"; } } while (false ); | |||
4271 | ArrayRef<Value *> Operands = Chain.slice(i, VF); | |||
4272 | ||||
4273 | R.buildTree(Operands); | |||
4274 | if (R.isTreeTinyAndNotFullyVectorizable()) | |||
4275 | continue; | |||
4276 | ||||
4277 | R.computeMinimumValueSizes(); | |||
4278 | ||||
4279 | int Cost = R.getTreeCost(); | |||
4280 | ||||
4281 | DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n"; } } while (false); | |||
4282 | if (Cost < -SLPCostThreshold) { | |||
4283 | DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n"; } } while (false); | |||
4284 | ||||
4285 | using namespace ore; | |||
4286 | ||||
4287 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "StoresVectorized", | |||
4288 | cast<StoreInst>(Chain[i])) | |||
4289 | << "Stores SLP vectorized with cost " << NV("Cost", Cost) | |||
4290 | << " and with tree size " | |||
4291 | << NV("TreeSize", R.getTreeSize())); | |||
4292 | ||||
4293 | R.vectorizeTree(); | |||
4294 | ||||
4295 | // Move to the next bundle. | |||
4296 | i += VF - 1; | |||
4297 | Changed = true; | |||
4298 | } | |||
4299 | } | |||
4300 | ||||
4301 | return Changed; | |||
4302 | } | |||
4303 | ||||
4304 | bool SLPVectorizerPass::vectorizeStores(ArrayRef<StoreInst *> Stores, | |||
4305 | BoUpSLP &R) { | |||
4306 | SetVector<StoreInst *> Heads; | |||
4307 | SmallDenseSet<StoreInst *> Tails; | |||
4308 | SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain; | |||
4309 | ||||
4310 | // We may run into multiple chains that merge into a single chain. We mark the | |||
4311 | // stores that we vectorized so that we don't visit the same store twice. | |||
4312 | BoUpSLP::ValueSet VectorizedStores; | |||
4313 | bool Changed = false; | |||
4314 | ||||
4315 | // Do a quadratic search on all of the given stores in reverse order and find | |||
4316 | // all of the pairs of stores that follow each other. | |||
4317 | SmallVector<unsigned, 16> IndexQueue; | |||
4318 | unsigned E = Stores.size(); | |||
4319 | IndexQueue.resize(E - 1); | |||
4320 | for (unsigned I = E; I > 0; --I) { | |||
4321 | unsigned Idx = I - 1; | |||
4322 | // If a store has multiple consecutive store candidates, search Stores | |||
4323 | // array according to the sequence: Idx-1, Idx+1, Idx-2, Idx+2, ... | |||
4324 | // This is because usually pairing with immediate succeeding or preceding | |||
4325 | // candidate create the best chance to find slp vectorization opportunity. | |||
4326 | unsigned Offset = 1; | |||
4327 | unsigned Cnt = 0; | |||
4328 | for (unsigned J = 0; J < E - 1; ++J, ++Offset) { | |||
4329 | if (Idx >= Offset) { | |||
4330 | IndexQueue[Cnt] = Idx - Offset; | |||
4331 | ++Cnt; | |||
4332 | } | |||
4333 | if (Idx + Offset < E) { | |||
4334 | IndexQueue[Cnt] = Idx + Offset; | |||
4335 | ++Cnt; | |||
4336 | } | |||
4337 | } | |||
4338 | ||||
4339 | for (auto K : IndexQueue) { | |||
4340 | if (isConsecutiveAccess(Stores[K], Stores[Idx], *DL, *SE)) { | |||
4341 | Tails.insert(Stores[Idx]); | |||
4342 | Heads.insert(Stores[K]); | |||
4343 | ConsecutiveChain[Stores[K]] = Stores[Idx]; | |||
4344 | break; | |||
4345 | } | |||
4346 | } | |||
4347 | } | |||
4348 | ||||
4349 | // For stores that start but don't end a link in the chain: | |||
4350 | for (auto *SI : llvm::reverse(Heads)) { | |||
4351 | if (Tails.count(SI)) | |||
4352 | continue; | |||
4353 | ||||
4354 | // We found a store instr that starts a chain. Now follow the chain and try | |||
4355 | // to vectorize it. | |||
4356 | BoUpSLP::ValueList Operands; | |||
4357 | StoreInst *I = SI; | |||
4358 | // Collect the chain into a list. | |||
4359 | while ((Tails.count(I) || Heads.count(I)) && !VectorizedStores.count(I)) { | |||
4360 | Operands.push_back(I); | |||
4361 | // Move to the next value in the chain. | |||
4362 | I = ConsecutiveChain[I]; | |||
4363 | } | |||
4364 | ||||
4365 | // FIXME: Is division-by-2 the correct step? Should we assert that the | |||
4366 | // register size is a power-of-2? | |||
4367 | for (unsigned Size = R.getMaxVecRegSize(); Size >= R.getMinVecRegSize(); | |||
4368 | Size /= 2) { | |||
4369 | if (vectorizeStoreChain(Operands, R, Size)) { | |||
4370 | // Mark the vectorized stores so that we don't vectorize them again. | |||
4371 | VectorizedStores.insert(Operands.begin(), Operands.end()); | |||
4372 | Changed = true; | |||
4373 | break; | |||
4374 | } | |||
4375 | } | |||
4376 | } | |||
4377 | ||||
4378 | return Changed; | |||
4379 | } | |||
4380 | ||||
4381 | void SLPVectorizerPass::collectSeedInstructions(BasicBlock *BB) { | |||
4382 | // Initialize the collections. We will make a single pass over the block. | |||
4383 | Stores.clear(); | |||
4384 | GEPs.clear(); | |||
4385 | ||||
4386 | // Visit the store and getelementptr instructions in BB and organize them in | |||
4387 | // Stores and GEPs according to the underlying objects of their pointer | |||
4388 | // operands. | |||
4389 | for (Instruction &I : *BB) { | |||
4390 | // Ignore store instructions that are volatile or have a pointer operand | |||
4391 | // that doesn't point to a scalar type. | |||
4392 | if (auto *SI = dyn_cast<StoreInst>(&I)) { | |||
4393 | if (!SI->isSimple()) | |||
4394 | continue; | |||
4395 | if (!isValidElementType(SI->getValueOperand()->getType())) | |||
4396 | continue; | |||
4397 | Stores[GetUnderlyingObject(SI->getPointerOperand(), *DL)].push_back(SI); | |||
4398 | } | |||
4399 | ||||
4400 | // Ignore getelementptr instructions that have more than one index, a | |||
4401 | // constant index, or a pointer operand that doesn't point to a scalar | |||
4402 | // type. | |||
4403 | else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) { | |||
4404 | auto Idx = GEP->idx_begin()->get(); | |||
4405 | if (GEP->getNumIndices() > 1 || isa<Constant>(Idx)) | |||
4406 | continue; | |||
4407 | if (!isValidElementType(Idx->getType())) | |||
4408 | continue; | |||
4409 | if (GEP->getType()->isVectorTy()) | |||
4410 | continue; | |||
4411 | GEPs[GetUnderlyingObject(GEP->getPointerOperand(), *DL)].push_back(GEP); | |||
4412 | } | |||
4413 | } | |||
4414 | } | |||
4415 | ||||
4416 | bool SLPVectorizerPass::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) { | |||
4417 | if (!A || !B) | |||
4418 | return false; | |||
4419 | Value *VL[] = { A, B }; | |||
4420 | return tryToVectorizeList(VL, R, None, true); | |||
4421 | } | |||
4422 | ||||
4423 | bool SLPVectorizerPass::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R, | |||
4424 | ArrayRef<Value *> BuildVector, | |||
4425 | bool AllowReorder, | |||
4426 | bool NeedExtraction) { | |||
4427 | if (VL.size() < 2) | |||
4428 | return false; | |||
4429 | ||||
4430 | DEBUG(dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size() << ".\n"; } } while (false) | |||
4431 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size() << ".\n"; } } while (false); | |||
4432 | ||||
4433 | // Check that all of the parts are scalar instructions of the same type. | |||
4434 | Instruction *I0 = dyn_cast<Instruction>(VL[0]); | |||
4435 | if (!I0) | |||
4436 | return false; | |||
4437 | ||||
4438 | unsigned Opcode0 = I0->getOpcode(); | |||
4439 | ||||
4440 | unsigned Sz = R.getVectorElementSize(I0); | |||
4441 | unsigned MinVF = std::max(2U, R.getMinVecRegSize() / Sz); | |||
4442 | unsigned MaxVF = std::max<unsigned>(PowerOf2Floor(VL.size()), MinVF); | |||
4443 | if (MaxVF < 2) { | |||
4444 | R.getORE()->emit([&]() { | |||
4445 | return OptimizationRemarkMissed( | |||
4446 | SV_NAME"slp-vectorizer", "SmallVF", I0) | |||
4447 | << "Cannot SLP vectorize list: vectorization factor " | |||
4448 | << "less than 2 is not supported"; | |||
4449 | }); | |||
4450 | return false; | |||
4451 | } | |||
4452 | ||||
4453 | for (Value *V : VL) { | |||
4454 | Type *Ty = V->getType(); | |||
4455 | if (!isValidElementType(Ty)) { | |||
4456 | // NOTE: the following will give user internal llvm type name, which may not be useful | |||
4457 | R.getORE()->emit([&]() { | |||
4458 | std::string type_str; | |||
4459 | llvm::raw_string_ostream rso(type_str); | |||
4460 | Ty->print(rso); | |||
4461 | return OptimizationRemarkMissed( | |||
4462 | SV_NAME"slp-vectorizer", "UnsupportedType", I0) | |||
4463 | << "Cannot SLP vectorize list: type " | |||
4464 | << rso.str() + " is unsupported by vectorizer"; | |||
4465 | }); | |||
4466 | return false; | |||
4467 | } | |||
4468 | Instruction *Inst = dyn_cast<Instruction>(V); | |||
4469 | ||||
4470 | if (!Inst) | |||
4471 | return false; | |||
4472 | if (Inst->getOpcode() != Opcode0) { | |||
4473 | R.getORE()->emit([&]() { | |||
4474 | return OptimizationRemarkMissed( | |||
4475 | SV_NAME"slp-vectorizer", "InequableTypes", I0) | |||
4476 | << "Cannot SLP vectorize list: not all of the " | |||
4477 | << "parts of scalar instructions are of the same type: " | |||
4478 | << ore::NV("Instruction1Opcode", I0) << " and " | |||
4479 | << ore::NV("Instruction2Opcode", Inst); | |||
4480 | }); | |||
4481 | return false; | |||
4482 | } | |||
4483 | } | |||
4484 | ||||
4485 | bool Changed = false; | |||
4486 | bool CandidateFound = false; | |||
4487 | int MinCost = SLPCostThreshold; | |||
4488 | ||||
4489 | // Keep track of values that were deleted by vectorizing in the loop below. | |||
4490 | SmallVector<WeakTrackingVH, 8> TrackValues(VL.begin(), VL.end()); | |||
4491 | ||||
4492 | unsigned NextInst = 0, MaxInst = VL.size(); | |||
4493 | for (unsigned VF = MaxVF; NextInst + 1 < MaxInst && VF >= MinVF; | |||
4494 | VF /= 2) { | |||
4495 | // No actual vectorization should happen, if number of parts is the same as | |||
4496 | // provided vectorization factor (i.e. the scalar type is used for vector | |||
4497 | // code during codegen). | |||
4498 | auto *VecTy = VectorType::get(VL[0]->getType(), VF); | |||
4499 | if (TTI->getNumberOfParts(VecTy) == VF) | |||
4500 | continue; | |||
4501 | for (unsigned I = NextInst; I < MaxInst; ++I) { | |||
4502 | unsigned OpsWidth = 0; | |||
4503 | ||||
4504 | if (I + VF > MaxInst) | |||
4505 | OpsWidth = MaxInst - I; | |||
4506 | else | |||
4507 | OpsWidth = VF; | |||
4508 | ||||
4509 | if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2) | |||
4510 | break; | |||
4511 | ||||
4512 | // Check that a previous iteration of this loop did not delete the Value. | |||
4513 | if (hasValueBeenRAUWed(VL, TrackValues, I, OpsWidth)) | |||
4514 | continue; | |||
4515 | ||||
4516 | DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false) | |||
4517 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false); | |||
4518 | ArrayRef<Value *> Ops = VL.slice(I, OpsWidth); | |||
4519 | ||||
4520 | ArrayRef<Value *> EmptyArray; | |||
4521 | ArrayRef<Value *> BuildVectorSlice; | |||
4522 | if (!BuildVector.empty()) | |||
4523 | BuildVectorSlice = BuildVector.slice(I, OpsWidth); | |||
4524 | ||||
4525 | R.buildTree(Ops, NeedExtraction ? EmptyArray : BuildVectorSlice); | |||
4526 | // TODO: check if we can allow reordering for more cases. | |||
4527 | if (AllowReorder && R.shouldReorder()) { | |||
4528 | // Conceptually, there is nothing actually preventing us from trying to | |||
4529 | // reorder a larger list. In fact, we do exactly this when vectorizing | |||
4530 | // reductions. However, at this point, we only expect to get here when | |||
4531 | // there are exactly two operations. | |||
4532 | assert(Ops.size() == 2)(static_cast <bool> (Ops.size() == 2) ? void (0) : __assert_fail ("Ops.size() == 2", "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4532, __extension__ __PRETTY_FUNCTION__)); | |||
4533 | assert(BuildVectorSlice.empty())(static_cast <bool> (BuildVectorSlice.empty()) ? void ( 0) : __assert_fail ("BuildVectorSlice.empty()", "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4533, __extension__ __PRETTY_FUNCTION__)); | |||
4534 | Value *ReorderedOps[] = {Ops[1], Ops[0]}; | |||
4535 | R.buildTree(ReorderedOps, None); | |||
4536 | } | |||
4537 | if (R.isTreeTinyAndNotFullyVectorizable()) | |||
4538 | continue; | |||
4539 | ||||
4540 | R.computeMinimumValueSizes(); | |||
4541 | int Cost = R.getTreeCost(); | |||
4542 | CandidateFound = true; | |||
4543 | MinCost = std::min(MinCost, Cost); | |||
4544 | ||||
4545 | if (Cost < -SLPCostThreshold) { | |||
4546 | DEBUG(dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n"; } } while (false); | |||
4547 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "VectorizedList", | |||
4548 | cast<Instruction>(Ops[0])) | |||
4549 | << "SLP vectorized with cost " << ore::NV("Cost", Cost) | |||
4550 | << " and with tree size " | |||
4551 | << ore::NV("TreeSize", R.getTreeSize())); | |||
4552 | ||||
4553 | Value *VectorizedRoot = R.vectorizeTree(); | |||
4554 | ||||
4555 | // Reconstruct the build vector by extracting the vectorized root. This | |||
4556 | // way we handle the case where some elements of the vector are | |||
4557 | // undefined. | |||
4558 | // (return (inserelt <4 xi32> (insertelt undef (opd0) 0) (opd1) 2)) | |||
4559 | if (!BuildVectorSlice.empty()) { | |||
4560 | // The insert point is the last build vector instruction. The | |||
4561 | // vectorized root will precede it. This guarantees that we get an | |||
4562 | // instruction. The vectorized tree could have been constant folded. | |||
4563 | Instruction *InsertAfter = cast<Instruction>(BuildVectorSlice.back()); | |||
4564 | unsigned VecIdx = 0; | |||
4565 | for (auto &V : BuildVectorSlice) { | |||
4566 | IRBuilder<NoFolder> Builder(InsertAfter->getParent(), | |||
4567 | ++BasicBlock::iterator(InsertAfter)); | |||
4568 | Instruction *I = cast<Instruction>(V); | |||
4569 | assert(isa<InsertElementInst>(I) || isa<InsertValueInst>(I))(static_cast <bool> (isa<InsertElementInst>(I) || isa<InsertValueInst>(I)) ? void (0) : __assert_fail ("isa<InsertElementInst>(I) || isa<InsertValueInst>(I)" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4569, __extension__ __PRETTY_FUNCTION__)); | |||
4570 | Instruction *Extract = | |||
4571 | cast<Instruction>(Builder.CreateExtractElement( | |||
4572 | VectorizedRoot, Builder.getInt32(VecIdx++))); | |||
4573 | I->setOperand(1, Extract); | |||
4574 | I->moveAfter(Extract); | |||
4575 | InsertAfter = I; | |||
4576 | } | |||
4577 | } | |||
4578 | // Move to the next bundle. | |||
4579 | I += VF - 1; | |||
4580 | NextInst = I + 1; | |||
4581 | Changed = true; | |||
4582 | } | |||
4583 | } | |||
4584 | } | |||
4585 | ||||
4586 | if (!Changed && CandidateFound) { | |||
4587 | R.getORE()->emit([&]() { | |||
4588 | return OptimizationRemarkMissed( | |||
4589 | SV_NAME"slp-vectorizer", "NotBeneficial", I0) | |||
4590 | << "List vectorization was possible but not beneficial with cost " | |||
4591 | << ore::NV("Cost", MinCost) << " >= " | |||
4592 | << ore::NV("Treshold", -SLPCostThreshold); | |||
4593 | }); | |||
4594 | } else if (!Changed) { | |||
4595 | R.getORE()->emit([&]() { | |||
4596 | return OptimizationRemarkMissed( | |||
4597 | SV_NAME"slp-vectorizer", "NotPossible", I0) | |||
4598 | << "Cannot SLP vectorize list: vectorization was impossible" | |||
4599 | << " with available vectorization factors"; | |||
4600 | }); | |||
4601 | } | |||
4602 | return Changed; | |||
4603 | } | |||
4604 | ||||
4605 | bool SLPVectorizerPass::tryToVectorize(Instruction *I, BoUpSLP &R) { | |||
4606 | if (!I) | |||
4607 | return false; | |||
4608 | ||||
4609 | if (!isa<BinaryOperator>(I) && !isa<CmpInst>(I)) | |||
4610 | return false; | |||
4611 | ||||
4612 | Value *P = I->getParent(); | |||
4613 | ||||
4614 | // Vectorize in current basic block only. | |||
4615 | auto *Op0 = dyn_cast<Instruction>(I->getOperand(0)); | |||
4616 | auto *Op1 = dyn_cast<Instruction>(I->getOperand(1)); | |||
4617 | if (!Op0 || !Op1 || Op0->getParent() != P || Op1->getParent() != P) | |||
4618 | return false; | |||
4619 | ||||
4620 | // Try to vectorize V. | |||
4621 | if (tryToVectorizePair(Op0, Op1, R)) | |||
4622 | return true; | |||
4623 | ||||
4624 | auto *A = dyn_cast<BinaryOperator>(Op0); | |||
4625 | auto *B = dyn_cast<BinaryOperator>(Op1); | |||
4626 | // Try to skip B. | |||
4627 | if (B && B->hasOneUse()) { | |||
4628 | auto *B0 = dyn_cast<BinaryOperator>(B->getOperand(0)); | |||
4629 | auto *B1 = dyn_cast<BinaryOperator>(B->getOperand(1)); | |||
4630 | if (B0 && B0->getParent() == P && tryToVectorizePair(A, B0, R)) | |||
4631 | return true; | |||
4632 | if (B1 && B1->getParent() == P && tryToVectorizePair(A, B1, R)) | |||
4633 | return true; | |||
4634 | } | |||
4635 | ||||
4636 | // Try to skip A. | |||
4637 | if (A && A->hasOneUse()) { | |||
4638 | auto *A0 = dyn_cast<BinaryOperator>(A->getOperand(0)); | |||
4639 | auto *A1 = dyn_cast<BinaryOperator>(A->getOperand(1)); | |||
4640 | if (A0 && A0->getParent() == P && tryToVectorizePair(A0, B, R)) | |||
4641 | return true; | |||
4642 | if (A1 && A1->getParent() == P && tryToVectorizePair(A1, B, R)) | |||
4643 | return true; | |||
4644 | } | |||
4645 | return false; | |||
4646 | } | |||
4647 | ||||
4648 | /// \brief Generate a shuffle mask to be used in a reduction tree. | |||
4649 | /// | |||
4650 | /// \param VecLen The length of the vector to be reduced. | |||
4651 | /// \param NumEltsToRdx The number of elements that should be reduced in the | |||
4652 | /// vector. | |||
4653 | /// \param IsPairwise Whether the reduction is a pairwise or splitting | |||
4654 | /// reduction. A pairwise reduction will generate a mask of | |||
4655 | /// <0,2,...> or <1,3,..> while a splitting reduction will generate | |||
4656 | /// <2,3, undef,undef> for a vector of 4 and NumElts = 2. | |||
4657 | /// \param IsLeft True will generate a mask of even elements, odd otherwise. | |||
4658 | static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx, | |||
4659 | bool IsPairwise, bool IsLeft, | |||
4660 | IRBuilder<> &Builder) { | |||
4661 | assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask")(static_cast <bool> ((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask") ? void (0) : __assert_fail ("(IsPairwise || !IsLeft) && \"Don't support a <0,1,undef,...> mask\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4661, __extension__ __PRETTY_FUNCTION__)); | |||
4662 | ||||
4663 | SmallVector<Constant *, 32> ShuffleMask( | |||
4664 | VecLen, UndefValue::get(Builder.getInt32Ty())); | |||
4665 | ||||
4666 | if (IsPairwise) | |||
4667 | // Build a mask of 0, 2, ... (left) or 1, 3, ... (right). | |||
4668 | for (unsigned i = 0; i != NumEltsToRdx; ++i) | |||
4669 | ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft); | |||
4670 | else | |||
4671 | // Move the upper half of the vector to the lower half. | |||
4672 | for (unsigned i = 0; i != NumEltsToRdx; ++i) | |||
4673 | ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i); | |||
4674 | ||||
4675 | return ConstantVector::get(ShuffleMask); | |||
4676 | } | |||
4677 | ||||
4678 | namespace { | |||
4679 | ||||
4680 | /// Model horizontal reductions. | |||
4681 | /// | |||
4682 | /// A horizontal reduction is a tree of reduction operations (currently add and | |||
4683 | /// fadd) that has operations that can be put into a vector as its leaf. | |||
4684 | /// For example, this tree: | |||
4685 | /// | |||
4686 | /// mul mul mul mul | |||
4687 | /// \ / \ / | |||
4688 | /// + + | |||
4689 | /// \ / | |||
4690 | /// + | |||
4691 | /// This tree has "mul" as its reduced values and "+" as its reduction | |||
4692 | /// operations. A reduction might be feeding into a store or a binary operation | |||
4693 | /// feeding a phi. | |||
4694 | /// ... | |||
4695 | /// \ / | |||
4696 | /// + | |||
4697 | /// | | |||
4698 | /// phi += | |||
4699 | /// | |||
4700 | /// Or: | |||
4701 | /// ... | |||
4702 | /// \ / | |||
4703 | /// + | |||
4704 | /// | | |||
4705 | /// *p = | |||
4706 | /// | |||
4707 | class HorizontalReduction { | |||
4708 | using ReductionOpsType = SmallVector<Value *, 16>; | |||
4709 | using ReductionOpsListType = SmallVector<ReductionOpsType, 2>; | |||
4710 | ReductionOpsListType ReductionOps; | |||
4711 | SmallVector<Value *, 32> ReducedVals; | |||
4712 | // Use map vector to make stable output. | |||
4713 | MapVector<Instruction *, Value *> ExtraArgs; | |||
4714 | ||||
4715 | /// Kind of the reduction data. | |||
4716 | enum ReductionKind { | |||
4717 | RK_None, /// Not a reduction. | |||
4718 | RK_Arithmetic, /// Binary reduction data. | |||
4719 | RK_Min, /// Minimum reduction data. | |||
4720 | RK_UMin, /// Unsigned minimum reduction data. | |||
4721 | RK_Max, /// Maximum reduction data. | |||
4722 | RK_UMax, /// Unsigned maximum reduction data. | |||
4723 | }; | |||
4724 | ||||
4725 | /// Contains info about operation, like its opcode, left and right operands. | |||
4726 | class OperationData { | |||
4727 | /// Opcode of the instruction. | |||
4728 | unsigned Opcode = 0; | |||
4729 | ||||
4730 | /// Left operand of the reduction operation. | |||
4731 | Value *LHS = nullptr; | |||
4732 | ||||
4733 | /// Right operand of the reduction operation. | |||
4734 | Value *RHS = nullptr; | |||
4735 | ||||
4736 | /// Kind of the reduction operation. | |||
4737 | ReductionKind Kind = RK_None; | |||
4738 | ||||
4739 | /// True if float point min/max reduction has no NaNs. | |||
4740 | bool NoNaN = false; | |||
4741 | ||||
4742 | /// Checks if the reduction operation can be vectorized. | |||
4743 | bool isVectorizable() const { | |||
4744 | return LHS && RHS && | |||
4745 | // We currently only support adds && min/max reductions. | |||
4746 | ((Kind == RK_Arithmetic && | |||
4747 | (Opcode == Instruction::Add || Opcode == Instruction::FAdd)) || | |||
4748 | ((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && | |||
4749 | (Kind == RK_Min || Kind == RK_Max)) || | |||
4750 | (Opcode == Instruction::ICmp && | |||
4751 | (Kind == RK_UMin || Kind == RK_UMax))); | |||
4752 | } | |||
4753 | ||||
4754 | /// Creates reduction operation with the current opcode. | |||
4755 | Value *createOp(IRBuilder<> &Builder, const Twine &Name) const { | |||
4756 | assert(isVectorizable() &&(static_cast <bool> (isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? void (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4757, __extension__ __PRETTY_FUNCTION__)) | |||
4757 | "Expected add|fadd or min/max reduction operation.")(static_cast <bool> (isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? void (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4757, __extension__ __PRETTY_FUNCTION__)); | |||
4758 | Value *Cmp; | |||
4759 | switch (Kind) { | |||
4760 | case RK_Arithmetic: | |||
4761 | return Builder.CreateBinOp((Instruction::BinaryOps)Opcode, LHS, RHS, | |||
4762 | Name); | |||
4763 | case RK_Min: | |||
4764 | Cmp = Opcode == Instruction::ICmp ? Builder.CreateICmpSLT(LHS, RHS) | |||
4765 | : Builder.CreateFCmpOLT(LHS, RHS); | |||
4766 | break; | |||
4767 | case RK_Max: | |||
4768 | Cmp = Opcode == Instruction::ICmp ? Builder.CreateICmpSGT(LHS, RHS) | |||
4769 | : Builder.CreateFCmpOGT(LHS, RHS); | |||
4770 | break; | |||
4771 | case RK_UMin: | |||
4772 | assert(Opcode == Instruction::ICmp && "Expected integer types.")(static_cast <bool> (Opcode == Instruction::ICmp && "Expected integer types.") ? void (0) : __assert_fail ("Opcode == Instruction::ICmp && \"Expected integer types.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4772, __extension__ __PRETTY_FUNCTION__)); | |||
4773 | Cmp = Builder.CreateICmpULT(LHS, RHS); | |||
4774 | break; | |||
4775 | case RK_UMax: | |||
4776 | assert(Opcode == Instruction::ICmp && "Expected integer types.")(static_cast <bool> (Opcode == Instruction::ICmp && "Expected integer types.") ? void (0) : __assert_fail ("Opcode == Instruction::ICmp && \"Expected integer types.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4776, __extension__ __PRETTY_FUNCTION__)); | |||
4777 | Cmp = Builder.CreateICmpUGT(LHS, RHS); | |||
4778 | break; | |||
4779 | case RK_None: | |||
4780 | llvm_unreachable("Unknown reduction operation.")::llvm::llvm_unreachable_internal("Unknown reduction operation." , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4780); | |||
4781 | } | |||
4782 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | |||
4783 | } | |||
4784 | ||||
4785 | public: | |||
4786 | explicit OperationData() = default; | |||
4787 | ||||
4788 | /// Construction for reduced values. They are identified by opcode only and | |||
4789 | /// don't have associated LHS/RHS values. | |||
4790 | explicit OperationData(Value *V) { | |||
4791 | if (auto *I = dyn_cast<Instruction>(V)) | |||
4792 | Opcode = I->getOpcode(); | |||
4793 | } | |||
4794 | ||||
4795 | /// Constructor for reduction operations with opcode and its left and | |||
4796 | /// right operands. | |||
4797 | OperationData(unsigned Opcode, Value *LHS, Value *RHS, ReductionKind Kind, | |||
4798 | bool NoNaN = false) | |||
4799 | : Opcode(Opcode), LHS(LHS), RHS(RHS), Kind(Kind), NoNaN(NoNaN) { | |||
4800 | assert(Kind != RK_None && "One of the reduction operations is expected.")(static_cast <bool> (Kind != RK_None && "One of the reduction operations is expected." ) ? void (0) : __assert_fail ("Kind != RK_None && \"One of the reduction operations is expected.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4800, __extension__ __PRETTY_FUNCTION__)); | |||
4801 | } | |||
4802 | ||||
4803 | explicit operator bool() const { return Opcode; } | |||
4804 | ||||
4805 | /// Get the index of the first operand. | |||
4806 | unsigned getFirstOperandIndex() const { | |||
4807 | assert(!!*this && "The opcode is not set.")(static_cast <bool> (!!*this && "The opcode is not set." ) ? void (0) : __assert_fail ("!!*this && \"The opcode is not set.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4807, __extension__ __PRETTY_FUNCTION__)); | |||
4808 | switch (Kind) { | |||
4809 | case RK_Min: | |||
4810 | case RK_UMin: | |||
4811 | case RK_Max: | |||
4812 | case RK_UMax: | |||
4813 | return 1; | |||
4814 | case RK_Arithmetic: | |||
4815 | case RK_None: | |||
4816 | break; | |||
4817 | } | |||
4818 | return 0; | |||
4819 | } | |||
4820 | ||||
4821 | /// Total number of operands in the reduction operation. | |||
4822 | unsigned getNumberOfOperands() const { | |||
4823 | assert(Kind != RK_None && !!*this && LHS && RHS &&(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4824, __extension__ __PRETTY_FUNCTION__)) | |||
4824 | "Expected reduction operation.")(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4824, __extension__ __PRETTY_FUNCTION__)); | |||
4825 | switch (Kind) { | |||
4826 | case RK_Arithmetic: | |||
4827 | return 2; | |||
4828 | case RK_Min: | |||
4829 | case RK_UMin: | |||
4830 | case RK_Max: | |||
4831 | case RK_UMax: | |||
4832 | return 3; | |||
4833 | case RK_None: | |||
4834 | break; | |||
4835 | } | |||
4836 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4836); | |||
4837 | } | |||
4838 | ||||
4839 | /// Checks if the operation has the same parent as \p P. | |||
4840 | bool hasSameParent(Instruction *I, Value *P, bool IsRedOp) const { | |||
4841 | assert(Kind != RK_None && !!*this && LHS && RHS &&(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4842, __extension__ __PRETTY_FUNCTION__)) | |||
4842 | "Expected reduction operation.")(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4842, __extension__ __PRETTY_FUNCTION__)); | |||
4843 | if (!IsRedOp) | |||
4844 | return I->getParent() == P; | |||
4845 | switch (Kind) { | |||
4846 | case RK_Arithmetic: | |||
4847 | // Arithmetic reduction operation must be used once only. | |||
4848 | return I->getParent() == P; | |||
4849 | case RK_Min: | |||
4850 | case RK_UMin: | |||
4851 | case RK_Max: | |||
4852 | case RK_UMax: { | |||
4853 | // SelectInst must be used twice while the condition op must have single | |||
4854 | // use only. | |||
4855 | auto *Cmp = cast<Instruction>(cast<SelectInst>(I)->getCondition()); | |||
4856 | return I->getParent() == P && Cmp && Cmp->getParent() == P; | |||
4857 | } | |||
4858 | case RK_None: | |||
4859 | break; | |||
4860 | } | |||
4861 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4861); | |||
4862 | } | |||
4863 | /// Expected number of uses for reduction operations/reduced values. | |||
4864 | bool hasRequiredNumberOfUses(Instruction *I, bool IsReductionOp) const { | |||
4865 | assert(Kind != RK_None && !!*this && LHS && RHS &&(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4866, __extension__ __PRETTY_FUNCTION__)) | |||
4866 | "Expected reduction operation.")(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4866, __extension__ __PRETTY_FUNCTION__)); | |||
4867 | switch (Kind) { | |||
4868 | case RK_Arithmetic: | |||
4869 | return I->hasOneUse(); | |||
4870 | case RK_Min: | |||
4871 | case RK_UMin: | |||
4872 | case RK_Max: | |||
4873 | case RK_UMax: | |||
4874 | return I->hasNUses(2) && | |||
4875 | (!IsReductionOp || | |||
4876 | cast<SelectInst>(I)->getCondition()->hasOneUse()); | |||
4877 | case RK_None: | |||
4878 | break; | |||
4879 | } | |||
4880 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4880); | |||
4881 | } | |||
4882 | ||||
4883 | /// Initializes the list of reduction operations. | |||
4884 | void initReductionOps(ReductionOpsListType &ReductionOps) { | |||
4885 | assert(Kind != RK_None && !!*this && LHS && RHS &&(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4886, __extension__ __PRETTY_FUNCTION__)) | |||
4886 | "Expected reduction operation.")(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4886, __extension__ __PRETTY_FUNCTION__)); | |||
4887 | switch (Kind) { | |||
4888 | case RK_Arithmetic: | |||
4889 | ReductionOps.assign(1, ReductionOpsType()); | |||
4890 | break; | |||
4891 | case RK_Min: | |||
4892 | case RK_UMin: | |||
4893 | case RK_Max: | |||
4894 | case RK_UMax: | |||
4895 | ReductionOps.assign(2, ReductionOpsType()); | |||
4896 | break; | |||
4897 | case RK_None: | |||
4898 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4898); | |||
4899 | } | |||
4900 | } | |||
4901 | /// Add all reduction operations for the reduction instruction \p I. | |||
4902 | void addReductionOps(Instruction *I, ReductionOpsListType &ReductionOps) { | |||
4903 | assert(Kind != RK_None && !!*this && LHS && RHS &&(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4904, __extension__ __PRETTY_FUNCTION__)) | |||
4904 | "Expected reduction operation.")(static_cast <bool> (Kind != RK_None && !!*this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && !!*this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4904, __extension__ __PRETTY_FUNCTION__)); | |||
4905 | switch (Kind) { | |||
4906 | case RK_Arithmetic: | |||
4907 | ReductionOps[0].emplace_back(I); | |||
4908 | break; | |||
4909 | case RK_Min: | |||
4910 | case RK_UMin: | |||
4911 | case RK_Max: | |||
4912 | case RK_UMax: | |||
4913 | ReductionOps[0].emplace_back(cast<SelectInst>(I)->getCondition()); | |||
4914 | ReductionOps[1].emplace_back(I); | |||
4915 | break; | |||
4916 | case RK_None: | |||
4917 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4917); | |||
4918 | } | |||
4919 | } | |||
4920 | ||||
4921 | /// Checks if instruction is associative and can be vectorized. | |||
4922 | bool isAssociative(Instruction *I) const { | |||
4923 | assert(Kind != RK_None && *this && LHS && RHS &&(static_cast <bool> (Kind != RK_None && *this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && *this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4924, __extension__ __PRETTY_FUNCTION__)) | |||
4924 | "Expected reduction operation.")(static_cast <bool> (Kind != RK_None && *this && LHS && RHS && "Expected reduction operation." ) ? void (0) : __assert_fail ("Kind != RK_None && *this && LHS && RHS && \"Expected reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4924, __extension__ __PRETTY_FUNCTION__)); | |||
4925 | switch (Kind) { | |||
4926 | case RK_Arithmetic: | |||
4927 | return I->isAssociative(); | |||
4928 | case RK_Min: | |||
4929 | case RK_Max: | |||
4930 | return Opcode == Instruction::ICmp || | |||
4931 | cast<Instruction>(I->getOperand(0))->isFast(); | |||
4932 | case RK_UMin: | |||
4933 | case RK_UMax: | |||
4934 | assert(Opcode == Instruction::ICmp &&(static_cast <bool> (Opcode == Instruction::ICmp && "Only integer compare operation is expected.") ? void (0) : __assert_fail ("Opcode == Instruction::ICmp && \"Only integer compare operation is expected.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4935, __extension__ __PRETTY_FUNCTION__)) | |||
4935 | "Only integer compare operation is expected.")(static_cast <bool> (Opcode == Instruction::ICmp && "Only integer compare operation is expected.") ? void (0) : __assert_fail ("Opcode == Instruction::ICmp && \"Only integer compare operation is expected.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4935, __extension__ __PRETTY_FUNCTION__)); | |||
4936 | return true; | |||
4937 | case RK_None: | |||
4938 | break; | |||
4939 | } | |||
4940 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4940); | |||
4941 | } | |||
4942 | ||||
4943 | /// Checks if the reduction operation can be vectorized. | |||
4944 | bool isVectorizable(Instruction *I) const { | |||
4945 | return isVectorizable() && isAssociative(I); | |||
4946 | } | |||
4947 | ||||
4948 | /// Checks if two operation data are both a reduction op or both a reduced | |||
4949 | /// value. | |||
4950 | bool operator==(const OperationData &OD) { | |||
4951 | assert(((Kind != OD.Kind) || ((!LHS == !OD.LHS) && (!RHS == !OD.RHS))) &&(static_cast <bool> (((Kind != OD.Kind) || ((!LHS == !OD .LHS) && (!RHS == !OD.RHS))) && "One of the comparing operations is incorrect." ) ? void (0) : __assert_fail ("((Kind != OD.Kind) || ((!LHS == !OD.LHS) && (!RHS == !OD.RHS))) && \"One of the comparing operations is incorrect.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4952, __extension__ __PRETTY_FUNCTION__)) | |||
4952 | "One of the comparing operations is incorrect.")(static_cast <bool> (((Kind != OD.Kind) || ((!LHS == !OD .LHS) && (!RHS == !OD.RHS))) && "One of the comparing operations is incorrect." ) ? void (0) : __assert_fail ("((Kind != OD.Kind) || ((!LHS == !OD.LHS) && (!RHS == !OD.RHS))) && \"One of the comparing operations is incorrect.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4952, __extension__ __PRETTY_FUNCTION__)); | |||
4953 | return this == &OD || (Kind == OD.Kind && Opcode == OD.Opcode); | |||
4954 | } | |||
4955 | bool operator!=(const OperationData &OD) { return !(*this == OD); } | |||
4956 | void clear() { | |||
4957 | Opcode = 0; | |||
4958 | LHS = nullptr; | |||
4959 | RHS = nullptr; | |||
4960 | Kind = RK_None; | |||
4961 | NoNaN = false; | |||
4962 | } | |||
4963 | ||||
4964 | /// Get the opcode of the reduction operation. | |||
4965 | unsigned getOpcode() const { | |||
4966 | assert(isVectorizable() && "Expected vectorizable operation.")(static_cast <bool> (isVectorizable() && "Expected vectorizable operation." ) ? void (0) : __assert_fail ("isVectorizable() && \"Expected vectorizable operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4966, __extension__ __PRETTY_FUNCTION__)); | |||
4967 | return Opcode; | |||
4968 | } | |||
4969 | ||||
4970 | /// Get kind of reduction data. | |||
4971 | ReductionKind getKind() const { return Kind; } | |||
4972 | Value *getLHS() const { return LHS; } | |||
4973 | Value *getRHS() const { return RHS; } | |||
4974 | Type *getConditionType() const { | |||
4975 | switch (Kind) { | |||
4976 | case RK_Arithmetic: | |||
4977 | return nullptr; | |||
4978 | case RK_Min: | |||
4979 | case RK_Max: | |||
4980 | case RK_UMin: | |||
4981 | case RK_UMax: | |||
4982 | return CmpInst::makeCmpResultType(LHS->getType()); | |||
4983 | case RK_None: | |||
4984 | break; | |||
4985 | } | |||
4986 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4986); | |||
4987 | } | |||
4988 | ||||
4989 | /// Creates reduction operation with the current opcode with the IR flags | |||
4990 | /// from \p ReductionOps. | |||
4991 | Value *createOp(IRBuilder<> &Builder, const Twine &Name, | |||
4992 | const ReductionOpsListType &ReductionOps) const { | |||
4993 | assert(isVectorizable() &&(static_cast <bool> (isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? void (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4994, __extension__ __PRETTY_FUNCTION__)) | |||
4994 | "Expected add|fadd or min/max reduction operation.")(static_cast <bool> (isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? void (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4994, __extension__ __PRETTY_FUNCTION__)); | |||
4995 | auto *Op = createOp(Builder, Name); | |||
4996 | switch (Kind) { | |||
4997 | case RK_Arithmetic: | |||
4998 | propagateIRFlags(Op, ReductionOps[0]); | |||
4999 | return Op; | |||
5000 | case RK_Min: | |||
5001 | case RK_Max: | |||
5002 | case RK_UMin: | |||
5003 | case RK_UMax: | |||
5004 | if (auto *SI = dyn_cast<SelectInst>(Op)) | |||
5005 | propagateIRFlags(SI->getCondition(), ReductionOps[0]); | |||
5006 | propagateIRFlags(Op, ReductionOps[1]); | |||
5007 | return Op; | |||
5008 | case RK_None: | |||
5009 | break; | |||
5010 | } | |||
5011 | llvm_unreachable("Unknown reduction operation.")::llvm::llvm_unreachable_internal("Unknown reduction operation." , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5011); | |||
5012 | } | |||
5013 | /// Creates reduction operation with the current opcode with the IR flags | |||
5014 | /// from \p I. | |||
5015 | Value *createOp(IRBuilder<> &Builder, const Twine &Name, | |||
5016 | Instruction *I) const { | |||
5017 | assert(isVectorizable() &&(static_cast <bool> (isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? void (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5018, __extension__ __PRETTY_FUNCTION__)) | |||
5018 | "Expected add|fadd or min/max reduction operation.")(static_cast <bool> (isVectorizable() && "Expected add|fadd or min/max reduction operation." ) ? void (0) : __assert_fail ("isVectorizable() && \"Expected add|fadd or min/max reduction operation.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5018, __extension__ __PRETTY_FUNCTION__)); | |||
5019 | auto *Op = createOp(Builder, Name); | |||
5020 | switch (Kind) { | |||
5021 | case RK_Arithmetic: | |||
5022 | propagateIRFlags(Op, I); | |||
5023 | return Op; | |||
5024 | case RK_Min: | |||
5025 | case RK_Max: | |||
5026 | case RK_UMin: | |||
5027 | case RK_UMax: | |||
5028 | if (auto *SI = dyn_cast<SelectInst>(Op)) { | |||
5029 | propagateIRFlags(SI->getCondition(), | |||
5030 | cast<SelectInst>(I)->getCondition()); | |||
5031 | } | |||
5032 | propagateIRFlags(Op, I); | |||
5033 | return Op; | |||
5034 | case RK_None: | |||
5035 | break; | |||
5036 | } | |||
5037 | llvm_unreachable("Unknown reduction operation.")::llvm::llvm_unreachable_internal("Unknown reduction operation." , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5037); | |||
5038 | } | |||
5039 | ||||
5040 | TargetTransformInfo::ReductionFlags getFlags() const { | |||
5041 | TargetTransformInfo::ReductionFlags Flags; | |||
5042 | Flags.NoNaN = NoNaN; | |||
5043 | switch (Kind) { | |||
5044 | case RK_Arithmetic: | |||
5045 | break; | |||
5046 | case RK_Min: | |||
5047 | Flags.IsSigned = Opcode == Instruction::ICmp; | |||
5048 | Flags.IsMaxOp = false; | |||
5049 | break; | |||
5050 | case RK_Max: | |||
5051 | Flags.IsSigned = Opcode == Instruction::ICmp; | |||
5052 | Flags.IsMaxOp = true; | |||
5053 | break; | |||
5054 | case RK_UMin: | |||
5055 | Flags.IsSigned = false; | |||
5056 | Flags.IsMaxOp = false; | |||
5057 | break; | |||
5058 | case RK_UMax: | |||
5059 | Flags.IsSigned = false; | |||
5060 | Flags.IsMaxOp = true; | |||
5061 | break; | |||
5062 | case RK_None: | |||
5063 | llvm_unreachable("Reduction kind is not set")::llvm::llvm_unreachable_internal("Reduction kind is not set" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5063); | |||
5064 | } | |||
5065 | return Flags; | |||
5066 | } | |||
5067 | }; | |||
5068 | ||||
5069 | Instruction *ReductionRoot = nullptr; | |||
5070 | ||||
5071 | /// The operation data of the reduction operation. | |||
5072 | OperationData ReductionData; | |||
5073 | ||||
5074 | /// The operation data of the values we perform a reduction on. | |||
5075 | OperationData ReducedValueData; | |||
5076 | ||||
5077 | /// Should we model this reduction as a pairwise reduction tree or a tree that | |||
5078 | /// splits the vector in halves and adds those halves. | |||
5079 | bool IsPairwiseReduction = false; | |||
5080 | ||||
5081 | /// Checks if the ParentStackElem.first should be marked as a reduction | |||
5082 | /// operation with an extra argument or as extra argument itself. | |||
5083 | void markExtraArg(std::pair<Instruction *, unsigned> &ParentStackElem, | |||
5084 | Value *ExtraArg) { | |||
5085 | if (ExtraArgs.count(ParentStackElem.first)) { | |||
5086 | ExtraArgs[ParentStackElem.first] = nullptr; | |||
5087 | // We ran into something like: | |||
5088 | // ParentStackElem.first = ExtraArgs[ParentStackElem.first] + ExtraArg. | |||
5089 | // The whole ParentStackElem.first should be considered as an extra value | |||
5090 | // in this case. | |||
5091 | // Do not perform analysis of remaining operands of ParentStackElem.first | |||
5092 | // instruction, this whole instruction is an extra argument. | |||
5093 | ParentStackElem.second = ParentStackElem.first->getNumOperands(); | |||
5094 | } else { | |||
5095 | // We ran into something like: | |||
5096 | // ParentStackElem.first += ... + ExtraArg + ... | |||
5097 | ExtraArgs[ParentStackElem.first] = ExtraArg; | |||
5098 | } | |||
5099 | } | |||
5100 | ||||
5101 | static OperationData getOperationData(Value *V) { | |||
5102 | if (!V) | |||
5103 | return OperationData(); | |||
5104 | ||||
5105 | Value *LHS; | |||
5106 | Value *RHS; | |||
5107 | if (m_BinOp(m_Value(LHS), m_Value(RHS)).match(V)) { | |||
5108 | return OperationData(cast<BinaryOperator>(V)->getOpcode(), LHS, RHS, | |||
5109 | RK_Arithmetic); | |||
5110 | } | |||
5111 | if (auto *Select = dyn_cast<SelectInst>(V)) { | |||
5112 | // Look for a min/max pattern. | |||
5113 | if (m_UMin(m_Value(LHS), m_Value(RHS)).match(Select)) { | |||
5114 | return OperationData(Instruction::ICmp, LHS, RHS, RK_UMin); | |||
5115 | } else if (m_SMin(m_Value(LHS), m_Value(RHS)).match(Select)) { | |||
5116 | return OperationData(Instruction::ICmp, LHS, RHS, RK_Min); | |||
5117 | } else if (m_OrdFMin(m_Value(LHS), m_Value(RHS)).match(Select) || | |||
5118 | m_UnordFMin(m_Value(LHS), m_Value(RHS)).match(Select)) { | |||
5119 | return OperationData( | |||
5120 | Instruction::FCmp, LHS, RHS, RK_Min, | |||
5121 | cast<Instruction>(Select->getCondition())->hasNoNaNs()); | |||
5122 | } else if (m_UMax(m_Value(LHS), m_Value(RHS)).match(Select)) { | |||
5123 | return OperationData(Instruction::ICmp, LHS, RHS, RK_UMax); | |||
5124 | } else if (m_SMax(m_Value(LHS), m_Value(RHS)).match(Select)) { | |||
5125 | return OperationData(Instruction::ICmp, LHS, RHS, RK_Max); | |||
5126 | } else if (m_OrdFMax(m_Value(LHS), m_Value(RHS)).match(Select) || | |||
5127 | m_UnordFMax(m_Value(LHS), m_Value(RHS)).match(Select)) { | |||
5128 | return OperationData( | |||
5129 | Instruction::FCmp, LHS, RHS, RK_Max, | |||
5130 | cast<Instruction>(Select->getCondition())->hasNoNaNs()); | |||
5131 | } | |||
5132 | } | |||
5133 | return OperationData(V); | |||
5134 | } | |||
5135 | ||||
5136 | public: | |||
5137 | HorizontalReduction() = default; | |||
5138 | ||||
5139 | /// \brief Try to find a reduction tree. | |||
5140 | bool matchAssociativeReduction(PHINode *Phi, Instruction *B) { | |||
5141 | assert((!Phi || is_contained(Phi->operands(), B)) &&(static_cast <bool> ((!Phi || is_contained(Phi->operands (), B)) && "Thi phi needs to use the binary operator" ) ? void (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), B)) && \"Thi phi needs to use the binary operator\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5142, __extension__ __PRETTY_FUNCTION__)) | |||
5142 | "Thi phi needs to use the binary operator")(static_cast <bool> ((!Phi || is_contained(Phi->operands (), B)) && "Thi phi needs to use the binary operator" ) ? void (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), B)) && \"Thi phi needs to use the binary operator\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5142, __extension__ __PRETTY_FUNCTION__)); | |||
5143 | ||||
5144 | ReductionData = getOperationData(B); | |||
5145 | ||||
5146 | // We could have a initial reductions that is not an add. | |||
5147 | // r *= v1 + v2 + v3 + v4 | |||
5148 | // In such a case start looking for a tree rooted in the first '+'. | |||
5149 | if (Phi) { | |||
5150 | if (ReductionData.getLHS() == Phi) { | |||
5151 | Phi = nullptr; | |||
5152 | B = dyn_cast<Instruction>(ReductionData.getRHS()); | |||
5153 | ReductionData = getOperationData(B); | |||
5154 | } else if (ReductionData.getRHS() == Phi) { | |||
5155 | Phi = nullptr; | |||
5156 | B = dyn_cast<Instruction>(ReductionData.getLHS()); | |||
5157 | ReductionData = getOperationData(B); | |||
5158 | } | |||
5159 | } | |||
5160 | ||||
5161 | if (!ReductionData.isVectorizable(B)) | |||
5162 | return false; | |||
5163 | ||||
5164 | Type *Ty = B->getType(); | |||
5165 | if (!isValidElementType(Ty)) | |||
5166 | return false; | |||
5167 | ||||
5168 | ReducedValueData.clear(); | |||
5169 | ReductionRoot = B; | |||
5170 | ||||
5171 | // Post order traverse the reduction tree starting at B. We only handle true | |||
5172 | // trees containing only binary operators. | |||
5173 | SmallVector<std::pair<Instruction *, unsigned>, 32> Stack; | |||
5174 | Stack.push_back(std::make_pair(B, ReductionData.getFirstOperandIndex())); | |||
5175 | ReductionData.initReductionOps(ReductionOps); | |||
5176 | while (!Stack.empty()) { | |||
5177 | Instruction *TreeN = Stack.back().first; | |||
5178 | unsigned EdgeToVist = Stack.back().second++; | |||
5179 | OperationData OpData = getOperationData(TreeN); | |||
5180 | bool IsReducedValue = OpData != ReductionData; | |||
5181 | ||||
5182 | // Postorder vist. | |||
5183 | if (IsReducedValue || EdgeToVist == OpData.getNumberOfOperands()) { | |||
5184 | if (IsReducedValue) | |||
5185 | ReducedVals.push_back(TreeN); | |||
5186 | else { | |||
5187 | auto I = ExtraArgs.find(TreeN); | |||
5188 | if (I != ExtraArgs.end() && !I->second) { | |||
5189 | // Check if TreeN is an extra argument of its parent operation. | |||
5190 | if (Stack.size() <= 1) { | |||
5191 | // TreeN can't be an extra argument as it is a root reduction | |||
5192 | // operation. | |||
5193 | return false; | |||
5194 | } | |||
5195 | // Yes, TreeN is an extra argument, do not add it to a list of | |||
5196 | // reduction operations. | |||
5197 | // Stack[Stack.size() - 2] always points to the parent operation. | |||
5198 | markExtraArg(Stack[Stack.size() - 2], TreeN); | |||
5199 | ExtraArgs.erase(TreeN); | |||
5200 | } else | |||
5201 | ReductionData.addReductionOps(TreeN, ReductionOps); | |||
5202 | } | |||
5203 | // Retract. | |||
5204 | Stack.pop_back(); | |||
5205 | continue; | |||
5206 | } | |||
5207 | ||||
5208 | // Visit left or right. | |||
5209 | Value *NextV = TreeN->getOperand(EdgeToVist); | |||
5210 | if (NextV != Phi) { | |||
5211 | auto *I = dyn_cast<Instruction>(NextV); | |||
5212 | OpData = getOperationData(I); | |||
5213 | // Continue analysis if the next operand is a reduction operation or | |||
5214 | // (possibly) a reduced value. If the reduced value opcode is not set, | |||
5215 | // the first met operation != reduction operation is considered as the | |||
5216 | // reduced value class. | |||
5217 | if (I && (!ReducedValueData || OpData == ReducedValueData || | |||
5218 | OpData == ReductionData)) { | |||
5219 | const bool IsReductionOperation = OpData == ReductionData; | |||
5220 | // Only handle trees in the current basic block. | |||
5221 | if (!ReductionData.hasSameParent(I, B->getParent(), | |||
5222 | IsReductionOperation)) { | |||
5223 | // I is an extra argument for TreeN (its parent operation). | |||
5224 | markExtraArg(Stack.back(), I); | |||
5225 | continue; | |||
5226 | } | |||
5227 | ||||
5228 | // Each tree node needs to have minimal number of users except for the | |||
5229 | // ultimate reduction. | |||
5230 | if (!ReductionData.hasRequiredNumberOfUses(I, | |||
5231 | OpData == ReductionData) && | |||
5232 | I != B) { | |||
5233 | // I is an extra argument for TreeN (its parent operation). | |||
5234 | markExtraArg(Stack.back(), I); | |||
5235 | continue; | |||
5236 | } | |||
5237 | ||||
5238 | if (IsReductionOperation) { | |||
5239 | // We need to be able to reassociate the reduction operations. | |||
5240 | if (!OpData.isAssociative(I)) { | |||
5241 | // I is an extra argument for TreeN (its parent operation). | |||
5242 | markExtraArg(Stack.back(), I); | |||
5243 | continue; | |||
5244 | } | |||
5245 | } else if (ReducedValueData && | |||
5246 | ReducedValueData != OpData) { | |||
5247 | // Make sure that the opcodes of the operations that we are going to | |||
5248 | // reduce match. | |||
5249 | // I is an extra argument for TreeN (its parent operation). | |||
5250 | markExtraArg(Stack.back(), I); | |||
5251 | continue; | |||
5252 | } else if (!ReducedValueData) | |||
5253 | ReducedValueData = OpData; | |||
5254 | ||||
5255 | Stack.push_back(std::make_pair(I, OpData.getFirstOperandIndex())); | |||
5256 | continue; | |||
5257 | } | |||
5258 | } | |||
5259 | // NextV is an extra argument for TreeN (its parent operation). | |||
5260 | markExtraArg(Stack.back(), NextV); | |||
5261 | } | |||
5262 | return true; | |||
5263 | } | |||
5264 | ||||
5265 | /// \brief Attempt to vectorize the tree found by | |||
5266 | /// matchAssociativeReduction. | |||
5267 | bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) { | |||
5268 | if (ReducedVals.empty()) | |||
5269 | return false; | |||
5270 | ||||
5271 | // If there is a sufficient number of reduction values, reduce | |||
5272 | // to a nearby power-of-2. Can safely generate oversized | |||
5273 | // vectors and rely on the backend to split them to legal sizes. | |||
5274 | unsigned NumReducedVals = ReducedVals.size(); | |||
5275 | if (NumReducedVals < 4) | |||
5276 | return false; | |||
5277 | ||||
5278 | unsigned ReduxWidth = PowerOf2Floor(NumReducedVals); | |||
5279 | ||||
5280 | Value *VectorizedTree = nullptr; | |||
5281 | IRBuilder<> Builder(ReductionRoot); | |||
5282 | FastMathFlags Unsafe; | |||
5283 | Unsafe.setFast(); | |||
5284 | Builder.setFastMathFlags(Unsafe); | |||
5285 | unsigned i = 0; | |||
5286 | ||||
5287 | BoUpSLP::ExtraValueToDebugLocsMap ExternallyUsedValues; | |||
5288 | // The same extra argument may be used several time, so log each attempt | |||
5289 | // to use it. | |||
5290 | for (auto &Pair : ExtraArgs) | |||
5291 | ExternallyUsedValues[Pair.second].push_back(Pair.first); | |||
5292 | SmallVector<Value *, 16> IgnoreList; | |||
5293 | for (auto &V : ReductionOps) | |||
5294 | IgnoreList.append(V.begin(), V.end()); | |||
5295 | while (i < NumReducedVals - ReduxWidth + 1 && ReduxWidth > 2) { | |||
5296 | auto VL = makeArrayRef(&ReducedVals[i], ReduxWidth); | |||
5297 | V.buildTree(VL, ExternallyUsedValues, IgnoreList); | |||
5298 | if (V.shouldReorder()) { | |||
5299 | SmallVector<Value *, 8> Reversed(VL.rbegin(), VL.rend()); | |||
5300 | V.buildTree(Reversed, ExternallyUsedValues, IgnoreList); | |||
5301 | } | |||
5302 | if (V.isTreeTinyAndNotFullyVectorizable()) | |||
5303 | break; | |||
5304 | ||||
5305 | V.computeMinimumValueSizes(); | |||
5306 | ||||
5307 | // Estimate cost. | |||
5308 | int Cost = | |||
5309 | V.getTreeCost() + getReductionCost(TTI, ReducedVals[i], ReduxWidth); | |||
5310 | if (Cost >= -SLPCostThreshold) { | |||
5311 | V.getORE()->emit([&]() { | |||
5312 | return OptimizationRemarkMissed( | |||
5313 | SV_NAME"slp-vectorizer", "HorSLPNotBeneficial", cast<Instruction>(VL[0])) | |||
5314 | << "Vectorizing horizontal reduction is possible" | |||
5315 | << "but not beneficial with cost " | |||
5316 | << ore::NV("Cost", Cost) << " and threshold " | |||
5317 | << ore::NV("Threshold", -SLPCostThreshold); | |||
5318 | }); | |||
5319 | break; | |||
5320 | } | |||
5321 | ||||
5322 | DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Costdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost << ". (HorRdx)\n"; } } while (false) | |||
5323 | << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost << ". (HorRdx)\n"; } } while (false); | |||
5324 | V.getORE()->emit([&]() { | |||
5325 | return OptimizationRemark( | |||
5326 | SV_NAME"slp-vectorizer", "VectorizedHorizontalReduction", cast<Instruction>(VL[0])) | |||
5327 | << "Vectorized horizontal reduction with cost " | |||
5328 | << ore::NV("Cost", Cost) << " and with tree size " | |||
5329 | << ore::NV("TreeSize", V.getTreeSize()); | |||
5330 | }); | |||
5331 | ||||
5332 | // Vectorize a tree. | |||
5333 | DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc(); | |||
5334 | Value *VectorizedRoot = V.vectorizeTree(ExternallyUsedValues); | |||
5335 | ||||
5336 | // Emit a reduction. | |||
5337 | Value *ReducedSubTree = | |||
5338 | emitReduction(VectorizedRoot, Builder, ReduxWidth, TTI); | |||
5339 | if (VectorizedTree) { | |||
5340 | Builder.SetCurrentDebugLocation(Loc); | |||
5341 | OperationData VectReductionData(ReductionData.getOpcode(), | |||
5342 | VectorizedTree, ReducedSubTree, | |||
5343 | ReductionData.getKind()); | |||
5344 | VectorizedTree = | |||
5345 | VectReductionData.createOp(Builder, "op.rdx", ReductionOps); | |||
5346 | } else | |||
5347 | VectorizedTree = ReducedSubTree; | |||
5348 | i += ReduxWidth; | |||
5349 | ReduxWidth = PowerOf2Floor(NumReducedVals - i); | |||
5350 | } | |||
5351 | ||||
5352 | if (VectorizedTree) { | |||
5353 | // Finish the reduction. | |||
5354 | for (; i < NumReducedVals; ++i) { | |||
5355 | auto *I = cast<Instruction>(ReducedVals[i]); | |||
5356 | Builder.SetCurrentDebugLocation(I->getDebugLoc()); | |||
5357 | OperationData VectReductionData(ReductionData.getOpcode(), | |||
5358 | VectorizedTree, I, | |||
5359 | ReductionData.getKind()); | |||
5360 | VectorizedTree = VectReductionData.createOp(Builder, "", ReductionOps); | |||
5361 | } | |||
5362 | for (auto &Pair : ExternallyUsedValues) { | |||
5363 | assert(!Pair.second.empty() &&(static_cast <bool> (!Pair.second.empty() && "At least one DebugLoc must be inserted" ) ? void (0) : __assert_fail ("!Pair.second.empty() && \"At least one DebugLoc must be inserted\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5364, __extension__ __PRETTY_FUNCTION__)) | |||
5364 | "At least one DebugLoc must be inserted")(static_cast <bool> (!Pair.second.empty() && "At least one DebugLoc must be inserted" ) ? void (0) : __assert_fail ("!Pair.second.empty() && \"At least one DebugLoc must be inserted\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5364, __extension__ __PRETTY_FUNCTION__)); | |||
5365 | // Add each externally used value to the final reduction. | |||
5366 | for (auto *I : Pair.second) { | |||
5367 | Builder.SetCurrentDebugLocation(I->getDebugLoc()); | |||
5368 | OperationData VectReductionData(ReductionData.getOpcode(), | |||
5369 | VectorizedTree, Pair.first, | |||
5370 | ReductionData.getKind()); | |||
5371 | VectorizedTree = VectReductionData.createOp(Builder, "op.extra", I); | |||
5372 | } | |||
5373 | } | |||
5374 | // Update users. | |||
5375 | ReductionRoot->replaceAllUsesWith(VectorizedTree); | |||
5376 | } | |||
5377 | return VectorizedTree != nullptr; | |||
5378 | } | |||
5379 | ||||
5380 | unsigned numReductionValues() const { | |||
5381 | return ReducedVals.size(); | |||
5382 | } | |||
5383 | ||||
5384 | private: | |||
5385 | /// \brief Calculate the cost of a reduction. | |||
5386 | int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal, | |||
5387 | unsigned ReduxWidth) { | |||
5388 | Type *ScalarTy = FirstReducedVal->getType(); | |||
5389 | Type *VecTy = VectorType::get(ScalarTy, ReduxWidth); | |||
5390 | ||||
5391 | int PairwiseRdxCost; | |||
5392 | int SplittingRdxCost; | |||
5393 | switch (ReductionData.getKind()) { | |||
5394 | case RK_Arithmetic: | |||
5395 | PairwiseRdxCost = | |||
5396 | TTI->getArithmeticReductionCost(ReductionData.getOpcode(), VecTy, | |||
5397 | /*IsPairwiseForm=*/true); | |||
5398 | SplittingRdxCost = | |||
5399 | TTI->getArithmeticReductionCost(ReductionData.getOpcode(), VecTy, | |||
5400 | /*IsPairwiseForm=*/false); | |||
5401 | break; | |||
5402 | case RK_Min: | |||
5403 | case RK_Max: | |||
5404 | case RK_UMin: | |||
5405 | case RK_UMax: { | |||
5406 | Type *VecCondTy = CmpInst::makeCmpResultType(VecTy); | |||
5407 | bool IsUnsigned = ReductionData.getKind() == RK_UMin || | |||
5408 | ReductionData.getKind() == RK_UMax; | |||
5409 | PairwiseRdxCost = | |||
5410 | TTI->getMinMaxReductionCost(VecTy, VecCondTy, | |||
5411 | /*IsPairwiseForm=*/true, IsUnsigned); | |||
5412 | SplittingRdxCost = | |||
5413 | TTI->getMinMaxReductionCost(VecTy, VecCondTy, | |||
5414 | /*IsPairwiseForm=*/false, IsUnsigned); | |||
5415 | break; | |||
5416 | } | |||
5417 | case RK_None: | |||
5418 | llvm_unreachable("Expected arithmetic or min/max reduction operation")::llvm::llvm_unreachable_internal("Expected arithmetic or min/max reduction operation" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5418); | |||
5419 | } | |||
5420 | ||||
5421 | IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost; | |||
5422 | int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost; | |||
5423 | ||||
5424 | int ScalarReduxCost; | |||
5425 | switch (ReductionData.getKind()) { | |||
5426 | case RK_Arithmetic: | |||
5427 | ScalarReduxCost = | |||
5428 | TTI->getArithmeticInstrCost(ReductionData.getOpcode(), ScalarTy); | |||
5429 | break; | |||
5430 | case RK_Min: | |||
5431 | case RK_Max: | |||
5432 | case RK_UMin: | |||
5433 | case RK_UMax: | |||
5434 | ScalarReduxCost = | |||
5435 | TTI->getCmpSelInstrCost(ReductionData.getOpcode(), ScalarTy) + | |||
5436 | TTI->getCmpSelInstrCost(Instruction::Select, ScalarTy, | |||
5437 | CmpInst::makeCmpResultType(ScalarTy)); | |||
5438 | break; | |||
5439 | case RK_None: | |||
5440 | llvm_unreachable("Expected arithmetic or min/max reduction operation")::llvm::llvm_unreachable_internal("Expected arithmetic or min/max reduction operation" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5440); | |||
5441 | } | |||
5442 | ScalarReduxCost *= (ReduxWidth - 1); | |||
5443 | ||||
5444 | DEBUG(dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost << " for reduction that starts with " << *FirstReducedVal << " (It is a " << (IsPairwiseReduction ? "pairwise" : "splitting") << " reduction)\n"; } } while (false) | |||
5445 | << " for reduction that starts with " << *FirstReducedValdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost << " for reduction that starts with " << *FirstReducedVal << " (It is a " << (IsPairwiseReduction ? "pairwise" : "splitting") << " reduction)\n"; } } while (false) | |||
5446 | << " (It is a "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost << " for reduction that starts with " << *FirstReducedVal << " (It is a " << (IsPairwiseReduction ? "pairwise" : "splitting") << " reduction)\n"; } } while (false) | |||
5447 | << (IsPairwiseReduction ? "pairwise" : "splitting")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost << " for reduction that starts with " << *FirstReducedVal << " (It is a " << (IsPairwiseReduction ? "pairwise" : "splitting") << " reduction)\n"; } } while (false) | |||
5448 | << " reduction)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost << " for reduction that starts with " << *FirstReducedVal << " (It is a " << (IsPairwiseReduction ? "pairwise" : "splitting") << " reduction)\n"; } } while (false); | |||
5449 | ||||
5450 | return VecReduxCost - ScalarReduxCost; | |||
5451 | } | |||
5452 | ||||
5453 | /// \brief Emit a horizontal reduction of the vectorized value. | |||
5454 | Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder, | |||
5455 | unsigned ReduxWidth, const TargetTransformInfo *TTI) { | |||
5456 | assert(VectorizedValue && "Need to have a vectorized tree node")(static_cast <bool> (VectorizedValue && "Need to have a vectorized tree node" ) ? void (0) : __assert_fail ("VectorizedValue && \"Need to have a vectorized tree node\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5456, __extension__ __PRETTY_FUNCTION__)); | |||
5457 | assert(isPowerOf2_32(ReduxWidth) &&(static_cast <bool> (isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now") ? void (0) : __assert_fail ("isPowerOf2_32(ReduxWidth) && \"We only handle power-of-two reductions for now\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5458, __extension__ __PRETTY_FUNCTION__)) | |||
5458 | "We only handle power-of-two reductions for now")(static_cast <bool> (isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now") ? void (0) : __assert_fail ("isPowerOf2_32(ReduxWidth) && \"We only handle power-of-two reductions for now\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5458, __extension__ __PRETTY_FUNCTION__)); | |||
5459 | ||||
5460 | if (!IsPairwiseReduction) | |||
5461 | return createSimpleTargetReduction( | |||
5462 | Builder, TTI, ReductionData.getOpcode(), VectorizedValue, | |||
5463 | ReductionData.getFlags(), ReductionOps.back()); | |||
5464 | ||||
5465 | Value *TmpVec = VectorizedValue; | |||
5466 | for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) { | |||
5467 | Value *LeftMask = | |||
5468 | createRdxShuffleMask(ReduxWidth, i, true, true, Builder); | |||
5469 | Value *RightMask = | |||
5470 | createRdxShuffleMask(ReduxWidth, i, true, false, Builder); | |||
5471 | ||||
5472 | Value *LeftShuf = Builder.CreateShuffleVector( | |||
5473 | TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l"); | |||
5474 | Value *RightShuf = Builder.CreateShuffleVector( | |||
5475 | TmpVec, UndefValue::get(TmpVec->getType()), (RightMask), | |||
5476 | "rdx.shuf.r"); | |||
5477 | OperationData VectReductionData(ReductionData.getOpcode(), LeftShuf, | |||
5478 | RightShuf, ReductionData.getKind()); | |||
5479 | TmpVec = VectReductionData.createOp(Builder, "op.rdx", ReductionOps); | |||
5480 | } | |||
5481 | ||||
5482 | // The result is in the first element of the vector. | |||
5483 | return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0)); | |||
5484 | } | |||
5485 | }; | |||
5486 | ||||
5487 | } // end anonymous namespace | |||
5488 | ||||
5489 | /// \brief Recognize construction of vectors like | |||
5490 | /// %ra = insertelement <4 x float> undef, float %s0, i32 0 | |||
5491 | /// %rb = insertelement <4 x float> %ra, float %s1, i32 1 | |||
5492 | /// %rc = insertelement <4 x float> %rb, float %s2, i32 2 | |||
5493 | /// %rd = insertelement <4 x float> %rc, float %s3, i32 3 | |||
5494 | /// starting from the last insertelement instruction. | |||
5495 | /// | |||
5496 | /// Returns true if it matches | |||
5497 | static bool findBuildVector(InsertElementInst *LastInsertElem, | |||
5498 | SmallVectorImpl<Value *> &BuildVector, | |||
5499 | SmallVectorImpl<Value *> &BuildVectorOpds) { | |||
5500 | Value *V = nullptr; | |||
5501 | do { | |||
5502 | BuildVector.push_back(LastInsertElem); | |||
5503 | BuildVectorOpds.push_back(LastInsertElem->getOperand(1)); | |||
5504 | V = LastInsertElem->getOperand(0); | |||
5505 | if (isa<UndefValue>(V)) | |||
5506 | break; | |||
5507 | LastInsertElem = dyn_cast<InsertElementInst>(V); | |||
5508 | if (!LastInsertElem || !LastInsertElem->hasOneUse()) | |||
5509 | return false; | |||
5510 | } while (true); | |||
5511 | std::reverse(BuildVector.begin(), BuildVector.end()); | |||
5512 | std::reverse(BuildVectorOpds.begin(), BuildVectorOpds.end()); | |||
5513 | return true; | |||
5514 | } | |||
5515 | ||||
5516 | /// \brief Like findBuildVector, but looks for construction of aggregate. | |||
5517 | /// | |||
5518 | /// \return true if it matches. | |||
5519 | static bool findBuildAggregate(InsertValueInst *IV, | |||
5520 | SmallVectorImpl<Value *> &BuildVector, | |||
5521 | SmallVectorImpl<Value *> &BuildVectorOpds) { | |||
5522 | Value *V; | |||
5523 | do { | |||
5524 | BuildVector.push_back(IV); | |||
5525 | BuildVectorOpds.push_back(IV->getInsertedValueOperand()); | |||
5526 | V = IV->getAggregateOperand(); | |||
5527 | if (isa<UndefValue>(V)) | |||
5528 | break; | |||
5529 | IV = dyn_cast<InsertValueInst>(V); | |||
5530 | if (!IV || !IV->hasOneUse()) | |||
5531 | return false; | |||
5532 | } while (true); | |||
5533 | std::reverse(BuildVector.begin(), BuildVector.end()); | |||
5534 | std::reverse(BuildVectorOpds.begin(), BuildVectorOpds.end()); | |||
5535 | return true; | |||
5536 | } | |||
5537 | ||||
5538 | static bool PhiTypeSorterFunc(Value *V, Value *V2) { | |||
5539 | return V->getType() < V2->getType(); | |||
5540 | } | |||
5541 | ||||
5542 | /// \brief Try and get a reduction value from a phi node. | |||
5543 | /// | |||
5544 | /// Given a phi node \p P in a block \p ParentBB, consider possible reductions | |||
5545 | /// if they come from either \p ParentBB or a containing loop latch. | |||
5546 | /// | |||
5547 | /// \returns A candidate reduction value if possible, or \code nullptr \endcode | |||
5548 | /// if not possible. | |||
5549 | static Value *getReductionValue(const DominatorTree *DT, PHINode *P, | |||
5550 | BasicBlock *ParentBB, LoopInfo *LI) { | |||
5551 | // There are situations where the reduction value is not dominated by the | |||
5552 | // reduction phi. Vectorizing such cases has been reported to cause | |||
5553 | // miscompiles. See PR25787. | |||
5554 | auto DominatedReduxValue = [&](Value *R) { | |||
5555 | return ( | |||
5556 | dyn_cast<Instruction>(R) && | |||
5557 | DT->dominates(P->getParent(), dyn_cast<Instruction>(R)->getParent())); | |||
5558 | }; | |||
5559 | ||||
5560 | Value *Rdx = nullptr; | |||
5561 | ||||
5562 | // Return the incoming value if it comes from the same BB as the phi node. | |||
5563 | if (P->getIncomingBlock(0) == ParentBB) { | |||
5564 | Rdx = P->getIncomingValue(0); | |||
5565 | } else if (P->getIncomingBlock(1) == ParentBB) { | |||
5566 | Rdx = P->getIncomingValue(1); | |||
5567 | } | |||
5568 | ||||
5569 | if (Rdx && DominatedReduxValue(Rdx)) | |||
5570 | return Rdx; | |||
5571 | ||||
5572 | // Otherwise, check whether we have a loop latch to look at. | |||
5573 | Loop *BBL = LI->getLoopFor(ParentBB); | |||
5574 | if (!BBL) | |||
5575 | return nullptr; | |||
5576 | BasicBlock *BBLatch = BBL->getLoopLatch(); | |||
5577 | if (!BBLatch) | |||
5578 | return nullptr; | |||
5579 | ||||
5580 | // There is a loop latch, return the incoming value if it comes from | |||
5581 | // that. This reduction pattern occasionally turns up. | |||
5582 | if (P->getIncomingBlock(0) == BBLatch) { | |||
5583 | Rdx = P->getIncomingValue(0); | |||
5584 | } else if (P->getIncomingBlock(1) == BBLatch) { | |||
5585 | Rdx = P->getIncomingValue(1); | |||
5586 | } | |||
5587 | ||||
5588 | if (Rdx && DominatedReduxValue(Rdx)) | |||
5589 | return Rdx; | |||
5590 | ||||
5591 | return nullptr; | |||
5592 | } | |||
5593 | ||||
5594 | /// Attempt to reduce a horizontal reduction. | |||
5595 | /// If it is legal to match a horizontal reduction feeding the phi node \a P | |||
5596 | /// with reduction operators \a Root (or one of its operands) in a basic block | |||
5597 | /// \a BB, then check if it can be done. If horizontal reduction is not found | |||
5598 | /// and root instruction is a binary operation, vectorization of the operands is | |||
5599 | /// attempted. | |||
5600 | /// \returns true if a horizontal reduction was matched and reduced or operands | |||
5601 | /// of one of the binary instruction were vectorized. | |||
5602 | /// \returns false if a horizontal reduction was not matched (or not possible) | |||
5603 | /// or no vectorization of any binary operation feeding \a Root instruction was | |||
5604 | /// performed. | |||
5605 | static bool tryToVectorizeHorReductionOrInstOperands( | |||
5606 | PHINode *P, Instruction *Root, BasicBlock *BB, BoUpSLP &R, | |||
5607 | TargetTransformInfo *TTI, | |||
5608 | const function_ref<bool(Instruction *, BoUpSLP &)> Vectorize) { | |||
5609 | if (!ShouldVectorizeHor) | |||
5610 | return false; | |||
5611 | ||||
5612 | if (!Root) | |||
5613 | return false; | |||
5614 | ||||
5615 | if (Root->getParent() != BB || isa<PHINode>(Root)) | |||
5616 | return false; | |||
5617 | // Start analysis starting from Root instruction. If horizontal reduction is | |||
5618 | // found, try to vectorize it. If it is not a horizontal reduction or | |||
5619 | // vectorization is not possible or not effective, and currently analyzed | |||
5620 | // instruction is a binary operation, try to vectorize the operands, using | |||
5621 | // pre-order DFS traversal order. If the operands were not vectorized, repeat | |||
5622 | // the same procedure considering each operand as a possible root of the | |||
5623 | // horizontal reduction. | |||
5624 | // Interrupt the process if the Root instruction itself was vectorized or all | |||
5625 | // sub-trees not higher that RecursionMaxDepth were analyzed/vectorized. | |||
5626 | SmallVector<std::pair<WeakTrackingVH, unsigned>, 8> Stack(1, {Root, 0}); | |||
5627 | SmallSet<Value *, 8> VisitedInstrs; | |||
5628 | bool Res = false; | |||
5629 | while (!Stack.empty()) { | |||
5630 | Value *V; | |||
5631 | unsigned Level; | |||
5632 | std::tie(V, Level) = Stack.pop_back_val(); | |||
5633 | if (!V) | |||
5634 | continue; | |||
5635 | auto *Inst = dyn_cast<Instruction>(V); | |||
5636 | if (!Inst) | |||
5637 | continue; | |||
5638 | auto *BI = dyn_cast<BinaryOperator>(Inst); | |||
5639 | auto *SI = dyn_cast<SelectInst>(Inst); | |||
5640 | if (BI || SI) { | |||
5641 | HorizontalReduction HorRdx; | |||
5642 | if (HorRdx.matchAssociativeReduction(P, Inst)) { | |||
5643 | if (HorRdx.tryToReduce(R, TTI)) { | |||
5644 | Res = true; | |||
5645 | // Set P to nullptr to avoid re-analysis of phi node in | |||
5646 | // matchAssociativeReduction function unless this is the root node. | |||
5647 | P = nullptr; | |||
5648 | continue; | |||
5649 | } | |||
5650 | } | |||
5651 | if (P && BI) { | |||
5652 | Inst = dyn_cast<Instruction>(BI->getOperand(0)); | |||
5653 | if (Inst == P) | |||
5654 | Inst = dyn_cast<Instruction>(BI->getOperand(1)); | |||
5655 | if (!Inst) { | |||
5656 | // Set P to nullptr to avoid re-analysis of phi node in | |||
5657 | // matchAssociativeReduction function unless this is the root node. | |||
5658 | P = nullptr; | |||
5659 | continue; | |||
5660 | } | |||
5661 | } | |||
5662 | } | |||
5663 | // Set P to nullptr to avoid re-analysis of phi node in | |||
5664 | // matchAssociativeReduction function unless this is the root node. | |||
5665 | P = nullptr; | |||
5666 | if (Vectorize(Inst, R)) { | |||
5667 | Res = true; | |||
5668 | continue; | |||
5669 | } | |||
5670 | ||||
5671 | // Try to vectorize operands. | |||
5672 | // Continue analysis for the instruction from the same basic block only to | |||
5673 | // save compile time. | |||
5674 | if (++Level < RecursionMaxDepth) | |||
5675 | for (auto *Op : Inst->operand_values()) | |||
5676 | if (VisitedInstrs.insert(Op).second) | |||
5677 | if (auto *I = dyn_cast<Instruction>(Op)) | |||
5678 | if (!isa<PHINode>(I) && I->getParent() == BB) | |||
5679 | Stack.emplace_back(Op, Level); | |||
5680 | } | |||
5681 | return Res; | |||
5682 | } | |||
5683 | ||||
5684 | bool SLPVectorizerPass::vectorizeRootInstruction(PHINode *P, Value *V, | |||
5685 | BasicBlock *BB, BoUpSLP &R, | |||
5686 | TargetTransformInfo *TTI) { | |||
5687 | if (!V) | |||
5688 | return false; | |||
5689 | auto *I = dyn_cast<Instruction>(V); | |||
5690 | if (!I) | |||
5691 | return false; | |||
5692 | ||||
5693 | if (!isa<BinaryOperator>(I)) | |||
5694 | P = nullptr; | |||
5695 | // Try to match and vectorize a horizontal reduction. | |||
5696 | auto &&ExtraVectorization = [this](Instruction *I, BoUpSLP &R) -> bool { | |||
5697 | return tryToVectorize(I, R); | |||
5698 | }; | |||
5699 | return tryToVectorizeHorReductionOrInstOperands(P, I, BB, R, TTI, | |||
5700 | ExtraVectorization); | |||
5701 | } | |||
5702 | ||||
5703 | bool SLPVectorizerPass::vectorizeInsertValueInst(InsertValueInst *IVI, | |||
5704 | BasicBlock *BB, BoUpSLP &R) { | |||
5705 | const DataLayout &DL = BB->getModule()->getDataLayout(); | |||
5706 | if (!R.canMapToVector(IVI->getType(), DL)) | |||
5707 | return false; | |||
5708 | ||||
5709 | SmallVector<Value *, 16> BuildVector; | |||
5710 | SmallVector<Value *, 16> BuildVectorOpds; | |||
5711 | if (!findBuildAggregate(IVI, BuildVector, BuildVectorOpds)) | |||
5712 | return false; | |||
5713 | ||||
5714 | DEBUG(dbgs() << "SLP: array mappable to vector: " << *IVI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: array mappable to vector: " << *IVI << "\n"; } } while (false); | |||
5715 | // Aggregate value is unlikely to be processed in vector register, we need to | |||
5716 | // extract scalars into scalar registers, so NeedExtraction is set true. | |||
5717 | return tryToVectorizeList(BuildVectorOpds, R, BuildVector, false, true); | |||
5718 | } | |||
5719 | ||||
5720 | bool SLPVectorizerPass::vectorizeInsertElementInst(InsertElementInst *IEI, | |||
5721 | BasicBlock *BB, BoUpSLP &R) { | |||
5722 | SmallVector<Value *, 16> BuildVector; | |||
5723 | SmallVector<Value *, 16> BuildVectorOpds; | |||
5724 | if (!findBuildVector(IEI, BuildVector, BuildVectorOpds)) | |||
5725 | return false; | |||
5726 | ||||
5727 | // Vectorize starting with the build vector operands ignoring the BuildVector | |||
5728 | // instructions for the purpose of scheduling and user extraction. | |||
5729 | return tryToVectorizeList(BuildVectorOpds, R, BuildVector); | |||
5730 | } | |||
5731 | ||||
5732 | bool SLPVectorizerPass::vectorizeCmpInst(CmpInst *CI, BasicBlock *BB, | |||
5733 | BoUpSLP &R) { | |||
5734 | if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) | |||
5735 | return true; | |||
5736 | ||||
5737 | bool OpsChanged = false; | |||
5738 | for (int Idx = 0; Idx < 2; ++Idx) { | |||
5739 | OpsChanged |= | |||
5740 | vectorizeRootInstruction(nullptr, CI->getOperand(Idx), BB, R, TTI); | |||
5741 | } | |||
5742 | return OpsChanged; | |||
5743 | } | |||
5744 | ||||
5745 | bool SLPVectorizerPass::vectorizeSimpleInstructions( | |||
5746 | SmallVectorImpl<WeakVH> &Instructions, BasicBlock *BB, BoUpSLP &R) { | |||
5747 | bool OpsChanged = false; | |||
5748 | for (auto &VH : reverse(Instructions)) { | |||
5749 | auto *I = dyn_cast_or_null<Instruction>(VH); | |||
5750 | if (!I) | |||
5751 | continue; | |||
5752 | if (auto *LastInsertValue = dyn_cast<InsertValueInst>(I)) | |||
5753 | OpsChanged |= vectorizeInsertValueInst(LastInsertValue, BB, R); | |||
5754 | else if (auto *LastInsertElem = dyn_cast<InsertElementInst>(I)) | |||
5755 | OpsChanged |= vectorizeInsertElementInst(LastInsertElem, BB, R); | |||
5756 | else if (auto *CI = dyn_cast<CmpInst>(I)) | |||
5757 | OpsChanged |= vectorizeCmpInst(CI, BB, R); | |||
5758 | } | |||
5759 | Instructions.clear(); | |||
5760 | return OpsChanged; | |||
5761 | } | |||
5762 | ||||
5763 | bool SLPVectorizerPass::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { | |||
5764 | bool Changed = false; | |||
5765 | SmallVector<Value *, 4> Incoming; | |||
5766 | SmallSet<Value *, 16> VisitedInstrs; | |||
5767 | ||||
5768 | bool HaveVectorizedPhiNodes = true; | |||
5769 | while (HaveVectorizedPhiNodes) { | |||
5770 | HaveVectorizedPhiNodes = false; | |||
5771 | ||||
5772 | // Collect the incoming values from the PHIs. | |||
5773 | Incoming.clear(); | |||
5774 | for (Instruction &I : *BB) { | |||
5775 | PHINode *P = dyn_cast<PHINode>(&I); | |||
5776 | if (!P) | |||
5777 | break; | |||
5778 | ||||
5779 | if (!VisitedInstrs.count(P)) | |||
5780 | Incoming.push_back(P); | |||
5781 | } | |||
5782 | ||||
5783 | // Sort by type. | |||
5784 | std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc); | |||
5785 | ||||
5786 | // Try to vectorize elements base on their type. | |||
5787 | for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(), | |||
5788 | E = Incoming.end(); | |||
5789 | IncIt != E;) { | |||
5790 | ||||
5791 | // Look for the next elements with the same type. | |||
5792 | SmallVector<Value *, 4>::iterator SameTypeIt = IncIt; | |||
5793 | while (SameTypeIt != E && | |||
5794 | (*SameTypeIt)->getType() == (*IncIt)->getType()) { | |||
5795 | VisitedInstrs.insert(*SameTypeIt); | |||
5796 | ++SameTypeIt; | |||
5797 | } | |||
5798 | ||||
5799 | // Try to vectorize them. | |||
5800 | unsigned NumElts = (SameTypeIt - IncIt); | |||
5801 | DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n"; } } while (false); | |||
5802 | // The order in which the phi nodes appear in the program does not matter. | |||
5803 | // So allow tryToVectorizeList to reorder them if it is beneficial. This | |||
5804 | // is done when there are exactly two elements since tryToVectorizeList | |||
5805 | // asserts that there are only two values when AllowReorder is true. | |||
5806 | bool AllowReorder = NumElts == 2; | |||
5807 | if (NumElts > 1 && tryToVectorizeList(makeArrayRef(IncIt, NumElts), R, | |||
5808 | None, AllowReorder)) { | |||
5809 | // Success start over because instructions might have been changed. | |||
5810 | HaveVectorizedPhiNodes = true; | |||
5811 | Changed = true; | |||
5812 | break; | |||
5813 | } | |||
5814 | ||||
5815 | // Start over at the next instruction of a different type (or the end). | |||
5816 | IncIt = SameTypeIt; | |||
5817 | } | |||
5818 | } | |||
5819 | ||||
5820 | VisitedInstrs.clear(); | |||
5821 | ||||
5822 | SmallVector<WeakVH, 8> PostProcessInstructions; | |||
5823 | SmallDenseSet<Instruction *, 4> KeyNodes; | |||
5824 | for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) { | |||
5825 | // We may go through BB multiple times so skip the one we have checked. | |||
5826 | if (!VisitedInstrs.insert(&*it).second) { | |||
5827 | if (it->use_empty() && KeyNodes.count(&*it) > 0 && | |||
5828 | vectorizeSimpleInstructions(PostProcessInstructions, BB, R)) { | |||
5829 | // We would like to start over since some instructions are deleted | |||
5830 | // and the iterator may become invalid value. | |||
5831 | Changed = true; | |||
5832 | it = BB->begin(); | |||
5833 | e = BB->end(); | |||
5834 | } | |||
5835 | continue; | |||
5836 | } | |||
5837 | ||||
5838 | if (isa<DbgInfoIntrinsic>(it)) | |||
5839 | continue; | |||
5840 | ||||
5841 | // Try to vectorize reductions that use PHINodes. | |||
5842 | if (PHINode *P = dyn_cast<PHINode>(it)) { | |||
5843 | // Check that the PHI is a reduction PHI. | |||
5844 | if (P->getNumIncomingValues() != 2) | |||
5845 | return Changed; | |||
5846 | ||||
5847 | // Try to match and vectorize a horizontal reduction. | |||
5848 | if (vectorizeRootInstruction(P, getReductionValue(DT, P, BB, LI), BB, R, | |||
5849 | TTI)) { | |||
5850 | Changed = true; | |||
5851 | it = BB->begin(); | |||
5852 | e = BB->end(); | |||
5853 | continue; | |||
5854 | } | |||
5855 | continue; | |||
5856 | } | |||
5857 | ||||
5858 | // Ran into an instruction without users, like terminator, or function call | |||
5859 | // with ignored return value, store. Ignore unused instructions (basing on | |||
5860 | // instruction type, except for CallInst and InvokeInst). | |||
5861 | if (it->use_empty() && (it->getType()->isVoidTy() || isa<CallInst>(it) || | |||
5862 | isa<InvokeInst>(it))) { | |||
5863 | KeyNodes.insert(&*it); | |||
5864 | bool OpsChanged = false; | |||
5865 | if (ShouldStartVectorizeHorAtStore || !isa<StoreInst>(it)) { | |||
5866 | for (auto *V : it->operand_values()) { | |||
5867 | // Try to match and vectorize a horizontal reduction. | |||
5868 | OpsChanged |= vectorizeRootInstruction(nullptr, V, BB, R, TTI); | |||
5869 | } | |||
5870 | } | |||
5871 | // Start vectorization of post-process list of instructions from the | |||
5872 | // top-tree instructions to try to vectorize as many instructions as | |||
5873 | // possible. | |||
5874 | OpsChanged |= vectorizeSimpleInstructions(PostProcessInstructions, BB, R); | |||
5875 | if (OpsChanged) { | |||
5876 | // We would like to start over since some instructions are deleted | |||
5877 | // and the iterator may become invalid value. | |||
5878 | Changed = true; | |||
5879 | it = BB->begin(); | |||
5880 | e = BB->end(); | |||
5881 | continue; | |||
5882 | } | |||
5883 | } | |||
5884 | ||||
5885 | if (isa<InsertElementInst>(it) || isa<CmpInst>(it) || | |||
5886 | isa<InsertValueInst>(it)) | |||
5887 | PostProcessInstructions.push_back(&*it); | |||
5888 | ||||
5889 | } | |||
5890 | ||||
5891 | return Changed; | |||
5892 | } | |||
5893 | ||||
5894 | bool SLPVectorizerPass::vectorizeGEPIndices(BasicBlock *BB, BoUpSLP &R) { | |||
5895 | auto Changed = false; | |||
5896 | for (auto &Entry : GEPs) { | |||
5897 | // If the getelementptr list has fewer than two elements, there's nothing | |||
5898 | // to do. | |||
5899 | if (Entry.second.size() < 2) | |||
5900 | continue; | |||
5901 | ||||
5902 | DEBUG(dbgs() << "SLP: Analyzing a getelementptr list of length "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a getelementptr list of length " << Entry.second.size() << ".\n"; } } while (false ) | |||
5903 | << Entry.second.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a getelementptr list of length " << Entry.second.size() << ".\n"; } } while (false ); | |||
5904 | ||||
5905 | // We process the getelementptr list in chunks of 16 (like we do for | |||
5906 | // stores) to minimize compile-time. | |||
5907 | for (unsigned BI = 0, BE = Entry.second.size(); BI < BE; BI += 16) { | |||
5908 | auto Len = std::min<unsigned>(BE - BI, 16); | |||
5909 | auto GEPList = makeArrayRef(&Entry.second[BI], Len); | |||
5910 | ||||
5911 | // Initialize a set a candidate getelementptrs. Note that we use a | |||
5912 | // SetVector here to preserve program order. If the index computations | |||
5913 | // are vectorizable and begin with loads, we want to minimize the chance | |||
5914 | // of having to reorder them later. | |||
5915 | SetVector<Value *> Candidates(GEPList.begin(), GEPList.end()); | |||
5916 | ||||
5917 | // Some of the candidates may have already been vectorized after we | |||
5918 | // initially collected them. If so, the WeakTrackingVHs will have | |||
5919 | // nullified the | |||
5920 | // values, so remove them from the set of candidates. | |||
5921 | Candidates.remove(nullptr); | |||
5922 | ||||
5923 | // Remove from the set of candidates all pairs of getelementptrs with | |||
5924 | // constant differences. Such getelementptrs are likely not good | |||
5925 | // candidates for vectorization in a bottom-up phase since one can be | |||
5926 | // computed from the other. We also ensure all candidate getelementptr | |||
5927 | // indices are unique. | |||
5928 | for (int I = 0, E = GEPList.size(); I < E && Candidates.size() > 1; ++I) { | |||
5929 | auto *GEPI = cast<GetElementPtrInst>(GEPList[I]); | |||
5930 | if (!Candidates.count(GEPI)) | |||
5931 | continue; | |||
5932 | auto *SCEVI = SE->getSCEV(GEPList[I]); | |||
5933 | for (int J = I + 1; J < E && Candidates.size() > 1; ++J) { | |||
5934 | auto *GEPJ = cast<GetElementPtrInst>(GEPList[J]); | |||
5935 | auto *SCEVJ = SE->getSCEV(GEPList[J]); | |||
5936 | if (isa<SCEVConstant>(SE->getMinusSCEV(SCEVI, SCEVJ))) { | |||
5937 | Candidates.remove(GEPList[I]); | |||
5938 | Candidates.remove(GEPList[J]); | |||
5939 | } else if (GEPI->idx_begin()->get() == GEPJ->idx_begin()->get()) { | |||
5940 | Candidates.remove(GEPList[J]); | |||
5941 | } | |||
5942 | } | |||
5943 | } | |||
5944 | ||||
5945 | // We break out of the above computation as soon as we know there are | |||
5946 | // fewer than two candidates remaining. | |||
5947 | if (Candidates.size() < 2) | |||
5948 | continue; | |||
5949 | ||||
5950 | // Add the single, non-constant index of each candidate to the bundle. We | |||
5951 | // ensured the indices met these constraints when we originally collected | |||
5952 | // the getelementptrs. | |||
5953 | SmallVector<Value *, 16> Bundle(Candidates.size()); | |||
5954 | auto BundleIndex = 0u; | |||
5955 | for (auto *V : Candidates) { | |||
5956 | auto *GEP = cast<GetElementPtrInst>(V); | |||
5957 | auto *GEPIdx = GEP->idx_begin()->get(); | |||
5958 | assert(GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx))(static_cast <bool> (GEP->getNumIndices() == 1 || !isa <Constant>(GEPIdx)) ? void (0) : __assert_fail ("GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx)" , "/build/llvm-toolchain-snapshot-6.0~svn321639/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5958, __extension__ __PRETTY_FUNCTION__)); | |||
5959 | Bundle[BundleIndex++] = GEPIdx; | |||
5960 | } | |||
5961 | ||||
5962 | // Try and vectorize the indices. We are currently only interested in | |||
5963 | // gather-like cases of the form: | |||
5964 | // | |||
5965 | // ... = g[a[0] - b[0]] + g[a[1] - b[1]] + ... | |||
5966 | // | |||
5967 | // where the loads of "a", the loads of "b", and the subtractions can be | |||
5968 | // performed in parallel. It's likely that detecting this pattern in a | |||
5969 | // bottom-up phase will be simpler and less costly than building a | |||
5970 | // full-blown top-down phase beginning at the consecutive loads. | |||
5971 | Changed |= tryToVectorizeList(Bundle, R); | |||
5972 | } | |||
5973 | } | |||
5974 | return Changed; | |||
5975 | } | |||
5976 | ||||
5977 | bool SLPVectorizerPass::vectorizeStoreChains(BoUpSLP &R) { | |||
5978 | bool Changed = false; | |||
5979 | // Attempt to sort and vectorize each of the store-groups. | |||
5980 | for (StoreListMap::iterator it = Stores.begin(), e = Stores.end(); it != e; | |||
5981 | ++it) { | |||
5982 | if (it->second.size() < 2) | |||
5983 | continue; | |||
5984 | ||||
5985 | DEBUG(dbgs() << "SLP: Analyzing a store chain of length "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << it->second.size() << ".\n"; } } while (false ) | |||
5986 | << it->second.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << it->second.size() << ".\n"; } } while (false ); | |||
5987 | ||||
5988 | // Process the stores in chunks of 16. | |||
5989 | // TODO: The limit of 16 inhibits greater vectorization factors. | |||
5990 | // For example, AVX2 supports v32i8. Increasing this limit, however, | |||
5991 | // may cause a significant compile-time increase. | |||
5992 | for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) { | |||
5993 | unsigned Len = std::min<unsigned>(CE - CI, 16); | |||
5994 | Changed |= vectorizeStores(makeArrayRef(&it->second[CI], Len), R); | |||
5995 | } | |||
5996 | } | |||
5997 | return Changed; | |||
5998 | } | |||
5999 | ||||
6000 | char SLPVectorizer::ID = 0; | |||
6001 | ||||
6002 | static const char lv_name[] = "SLP Vectorizer"; | |||
6003 | ||||
6004 | INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)static void *initializeSLPVectorizerPassOnce(PassRegistry & Registry) { | |||
6005 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); | |||
6006 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | |||
6007 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | |||
6008 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); | |||
6009 | INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry); | |||
6010 | INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)initializeDemandedBitsWrapperPassPass(Registry); | |||
6011 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); | |||
6012 | INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)PassInfo *PI = new PassInfo( lv_name, "slp-vectorizer", & SLPVectorizer::ID, PassInfo::NormalCtor_t(callDefaultCtor< SLPVectorizer>), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeSLPVectorizerPassFlag ; void llvm::initializeSLPVectorizerPass(PassRegistry &Registry ) { llvm::call_once(InitializeSLPVectorizerPassFlag, initializeSLPVectorizerPassOnce , std::ref(Registry)); } | |||
6013 | ||||
6014 | Pass *llvm::createSLPVectorizerPass() { return new SLPVectorizer(); } |
1 | //===- llvm/Support/Casting.h - Allow flexible, checked, casts --*- C++ -*-===// |
2 | // |
3 | // The LLVM Compiler Infrastructure |
4 | // |
5 | // This file is distributed under the University of Illinois Open Source |
6 | // License. See LICENSE.TXT for details. |
7 | // |
8 | //===----------------------------------------------------------------------===// |
9 | // |
10 | // This file defines the isa<X>(), cast<X>(), dyn_cast<X>(), cast_or_null<X>(), |
11 | // and dyn_cast_or_null<X>() templates. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #ifndef LLVM_SUPPORT_CASTING_H |
16 | #define LLVM_SUPPORT_CASTING_H |
17 | |
18 | #include "llvm/Support/Compiler.h" |
19 | #include "llvm/Support/type_traits.h" |
20 | #include <cassert> |
21 | #include <memory> |
22 | #include <type_traits> |
23 | |
24 | namespace llvm { |
25 | |
26 | //===----------------------------------------------------------------------===// |
27 | // isa<x> Support Templates |
28 | //===----------------------------------------------------------------------===// |
29 | |
30 | // Define a template that can be specialized by smart pointers to reflect the |
31 | // fact that they are automatically dereferenced, and are not involved with the |
32 | // template selection process... the default implementation is a noop. |
33 | // |
34 | template<typename From> struct simplify_type { |
35 | using SimpleType = From; // The real type this represents... |
36 | |
37 | // An accessor to get the real value... |
38 | static SimpleType &getSimplifiedValue(From &Val) { return Val; } |
39 | }; |
40 | |
41 | template<typename From> struct simplify_type<const From> { |
42 | using NonConstSimpleType = typename simplify_type<From>::SimpleType; |
43 | using SimpleType = |
44 | typename add_const_past_pointer<NonConstSimpleType>::type; |
45 | using RetType = |
46 | typename add_lvalue_reference_if_not_pointer<SimpleType>::type; |
47 | |
48 | static RetType getSimplifiedValue(const From& Val) { |
49 | return simplify_type<From>::getSimplifiedValue(const_cast<From&>(Val)); |
50 | } |
51 | }; |
52 | |
53 | // The core of the implementation of isa<X> is here; To and From should be |
54 | // the names of classes. This template can be specialized to customize the |
55 | // implementation of isa<> without rewriting it from scratch. |
56 | template <typename To, typename From, typename Enabler = void> |
57 | struct isa_impl { |
58 | static inline bool doit(const From &Val) { |
59 | return To::classof(&Val); |
60 | } |
61 | }; |
62 | |
63 | /// \brief Always allow upcasts, and perform no dynamic check for them. |
64 | template <typename To, typename From> |
65 | struct isa_impl< |
66 | To, From, typename std::enable_if<std::is_base_of<To, From>::value>::type> { |
67 | static inline bool doit(const From &) { return true; } |
68 | }; |
69 | |
70 | template <typename To, typename From> struct isa_impl_cl { |
71 | static inline bool doit(const From &Val) { |
72 | return isa_impl<To, From>::doit(Val); |
73 | } |
74 | }; |
75 | |
76 | template <typename To, typename From> struct isa_impl_cl<To, const From> { |
77 | static inline bool doit(const From &Val) { |
78 | return isa_impl<To, From>::doit(Val); |
79 | } |
80 | }; |
81 | |
82 | template <typename To, typename From> |
83 | struct isa_impl_cl<To, const std::unique_ptr<From>> { |
84 | static inline bool doit(const std::unique_ptr<From> &Val) { |
85 | assert(Val && "isa<> used on a null pointer")(static_cast <bool> (Val && "isa<> used on a null pointer" ) ? void (0) : __assert_fail ("Val && \"isa<> used on a null pointer\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/Support/Casting.h" , 85, __extension__ __PRETTY_FUNCTION__)); |
86 | return isa_impl_cl<To, From>::doit(*Val); |
87 | } |
88 | }; |
89 | |
90 | template <typename To, typename From> struct isa_impl_cl<To, From*> { |
91 | static inline bool doit(const From *Val) { |
92 | assert(Val && "isa<> used on a null pointer")(static_cast <bool> (Val && "isa<> used on a null pointer" ) ? void (0) : __assert_fail ("Val && \"isa<> used on a null pointer\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/Support/Casting.h" , 92, __extension__ __PRETTY_FUNCTION__)); |
93 | return isa_impl<To, From>::doit(*Val); |
94 | } |
95 | }; |
96 | |
97 | template <typename To, typename From> struct isa_impl_cl<To, From*const> { |
98 | static inline bool doit(const From *Val) { |
99 | assert(Val && "isa<> used on a null pointer")(static_cast <bool> (Val && "isa<> used on a null pointer" ) ? void (0) : __assert_fail ("Val && \"isa<> used on a null pointer\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/Support/Casting.h" , 99, __extension__ __PRETTY_FUNCTION__)); |
100 | return isa_impl<To, From>::doit(*Val); |
101 | } |
102 | }; |
103 | |
104 | template <typename To, typename From> struct isa_impl_cl<To, const From*> { |
105 | static inline bool doit(const From *Val) { |
106 | assert(Val && "isa<> used on a null pointer")(static_cast <bool> (Val && "isa<> used on a null pointer" ) ? void (0) : __assert_fail ("Val && \"isa<> used on a null pointer\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/Support/Casting.h" , 106, __extension__ __PRETTY_FUNCTION__)); |
107 | return isa_impl<To, From>::doit(*Val); |
108 | } |
109 | }; |
110 | |
111 | template <typename To, typename From> struct isa_impl_cl<To, const From*const> { |
112 | static inline bool doit(const From *Val) { |
113 | assert(Val && "isa<> used on a null pointer")(static_cast <bool> (Val && "isa<> used on a null pointer" ) ? void (0) : __assert_fail ("Val && \"isa<> used on a null pointer\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/Support/Casting.h" , 113, __extension__ __PRETTY_FUNCTION__)); |
114 | return isa_impl<To, From>::doit(*Val); |
115 | } |
116 | }; |
117 | |
118 | template<typename To, typename From, typename SimpleFrom> |
119 | struct isa_impl_wrap { |
120 | // When From != SimplifiedType, we can simplify the type some more by using |
121 | // the simplify_type template. |
122 | static bool doit(const From &Val) { |
123 | return isa_impl_wrap<To, SimpleFrom, |
124 | typename simplify_type<SimpleFrom>::SimpleType>::doit( |
125 | simplify_type<const From>::getSimplifiedValue(Val)); |
126 | } |
127 | }; |
128 | |
129 | template<typename To, typename FromTy> |
130 | struct isa_impl_wrap<To, FromTy, FromTy> { |
131 | // When From == SimpleType, we are as simple as we are going to get. |
132 | static bool doit(const FromTy &Val) { |
133 | return isa_impl_cl<To,FromTy>::doit(Val); |
134 | } |
135 | }; |
136 | |
137 | // isa<X> - Return true if the parameter to the template is an instance of the |
138 | // template type argument. Used like this: |
139 | // |
140 | // if (isa<Type>(myVal)) { ... } |
141 | // |
142 | template <class X, class Y> LLVM_NODISCARD[[clang::warn_unused_result]] inline bool isa(const Y &Val) { |
143 | return isa_impl_wrap<X, const Y, |
144 | typename simplify_type<const Y>::SimpleType>::doit(Val); |
145 | } |
146 | |
147 | //===----------------------------------------------------------------------===// |
148 | // cast<x> Support Templates |
149 | //===----------------------------------------------------------------------===// |
150 | |
151 | template<class To, class From> struct cast_retty; |
152 | |
153 | // Calculate what type the 'cast' function should return, based on a requested |
154 | // type of To and a source type of From. |
155 | template<class To, class From> struct cast_retty_impl { |
156 | using ret_type = To &; // Normal case, return Ty& |
157 | }; |
158 | template<class To, class From> struct cast_retty_impl<To, const From> { |
159 | using ret_type = const To &; // Normal case, return Ty& |
160 | }; |
161 | |
162 | template<class To, class From> struct cast_retty_impl<To, From*> { |
163 | using ret_type = To *; // Pointer arg case, return Ty* |
164 | }; |
165 | |
166 | template<class To, class From> struct cast_retty_impl<To, const From*> { |
167 | using ret_type = const To *; // Constant pointer arg case, return const Ty* |
168 | }; |
169 | |
170 | template<class To, class From> struct cast_retty_impl<To, const From*const> { |
171 | using ret_type = const To *; // Constant pointer arg case, return const Ty* |
172 | }; |
173 | |
174 | template <class To, class From> |
175 | struct cast_retty_impl<To, std::unique_ptr<From>> { |
176 | private: |
177 | using PointerType = typename cast_retty_impl<To, From *>::ret_type; |
178 | using ResultType = typename std::remove_pointer<PointerType>::type; |
179 | |
180 | public: |
181 | using ret_type = std::unique_ptr<ResultType>; |
182 | }; |
183 | |
184 | template<class To, class From, class SimpleFrom> |
185 | struct cast_retty_wrap { |
186 | // When the simplified type and the from type are not the same, use the type |
187 | // simplifier to reduce the type, then reuse cast_retty_impl to get the |
188 | // resultant type. |
189 | using ret_type = typename cast_retty<To, SimpleFrom>::ret_type; |
190 | }; |
191 | |
192 | template<class To, class FromTy> |
193 | struct cast_retty_wrap<To, FromTy, FromTy> { |
194 | // When the simplified type is equal to the from type, use it directly. |
195 | using ret_type = typename cast_retty_impl<To,FromTy>::ret_type; |
196 | }; |
197 | |
198 | template<class To, class From> |
199 | struct cast_retty { |
200 | using ret_type = typename cast_retty_wrap< |
201 | To, From, typename simplify_type<From>::SimpleType>::ret_type; |
202 | }; |
203 | |
204 | // Ensure the non-simple values are converted using the simplify_type template |
205 | // that may be specialized by smart pointers... |
206 | // |
207 | template<class To, class From, class SimpleFrom> struct cast_convert_val { |
208 | // This is not a simple type, use the template to simplify it... |
209 | static typename cast_retty<To, From>::ret_type doit(From &Val) { |
210 | return cast_convert_val<To, SimpleFrom, |
211 | typename simplify_type<SimpleFrom>::SimpleType>::doit( |
212 | simplify_type<From>::getSimplifiedValue(Val)); |
213 | } |
214 | }; |
215 | |
216 | template<class To, class FromTy> struct cast_convert_val<To,FromTy,FromTy> { |
217 | // This _is_ a simple type, just cast it. |
218 | static typename cast_retty<To, FromTy>::ret_type doit(const FromTy &Val) { |
219 | typename cast_retty<To, FromTy>::ret_type Res2 |
220 | = (typename cast_retty<To, FromTy>::ret_type)const_cast<FromTy&>(Val); |
221 | return Res2; |
222 | } |
223 | }; |
224 | |
225 | template <class X> struct is_simple_type { |
226 | static const bool value = |
227 | std::is_same<X, typename simplify_type<X>::SimpleType>::value; |
228 | }; |
229 | |
230 | // cast<X> - Return the argument parameter cast to the specified type. This |
231 | // casting operator asserts that the type is correct, so it does not return null |
232 | // on failure. It does not allow a null argument (use cast_or_null for that). |
233 | // It is typically used like this: |
234 | // |
235 | // cast<Instruction>(myVal)->getParent() |
236 | // |
237 | template <class X, class Y> |
238 | inline typename std::enable_if<!is_simple_type<Y>::value, |
239 | typename cast_retty<X, const Y>::ret_type>::type |
240 | cast(const Y &Val) { |
241 | assert(isa<X>(Val) && "cast<Ty>() argument of incompatible type!")(static_cast <bool> (isa<X>(Val) && "cast<Ty>() argument of incompatible type!" ) ? void (0) : __assert_fail ("isa<X>(Val) && \"cast<Ty>() argument of incompatible type!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/Support/Casting.h" , 241, __extension__ __PRETTY_FUNCTION__)); |
242 | return cast_convert_val< |
243 | X, const Y, typename simplify_type<const Y>::SimpleType>::doit(Val); |
244 | } |
245 | |
246 | template <class X, class Y> |
247 | inline typename cast_retty<X, Y>::ret_type cast(Y &Val) { |
248 | assert(isa<X>(Val) && "cast<Ty>() argument of incompatible type!")(static_cast <bool> (isa<X>(Val) && "cast<Ty>() argument of incompatible type!" ) ? void (0) : __assert_fail ("isa<X>(Val) && \"cast<Ty>() argument of incompatible type!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/Support/Casting.h" , 248, __extension__ __PRETTY_FUNCTION__)); |
249 | return cast_convert_val<X, Y, |
250 | typename simplify_type<Y>::SimpleType>::doit(Val); |
251 | } |
252 | |
253 | template <class X, class Y> |
254 | inline typename cast_retty<X, Y *>::ret_type cast(Y *Val) { |
255 | assert(isa<X>(Val) && "cast<Ty>() argument of incompatible type!")(static_cast <bool> (isa<X>(Val) && "cast<Ty>() argument of incompatible type!" ) ? void (0) : __assert_fail ("isa<X>(Val) && \"cast<Ty>() argument of incompatible type!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/Support/Casting.h" , 255, __extension__ __PRETTY_FUNCTION__)); |
256 | return cast_convert_val<X, Y*, |
257 | typename simplify_type<Y*>::SimpleType>::doit(Val); |
258 | } |
259 | |
260 | template <class X, class Y> |
261 | inline typename cast_retty<X, std::unique_ptr<Y>>::ret_type |
262 | cast(std::unique_ptr<Y> &&Val) { |
263 | assert(isa<X>(Val.get()) && "cast<Ty>() argument of incompatible type!")(static_cast <bool> (isa<X>(Val.get()) && "cast<Ty>() argument of incompatible type!") ? void (0 ) : __assert_fail ("isa<X>(Val.get()) && \"cast<Ty>() argument of incompatible type!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/Support/Casting.h" , 263, __extension__ __PRETTY_FUNCTION__)); |
264 | using ret_type = typename cast_retty<X, std::unique_ptr<Y>>::ret_type; |
265 | return ret_type( |
266 | cast_convert_val<X, Y *, typename simplify_type<Y *>::SimpleType>::doit( |
267 | Val.release())); |
268 | } |
269 | |
270 | // cast_or_null<X> - Functionally identical to cast, except that a null value is |
271 | // accepted. |
272 | // |
273 | template <class X, class Y> |
274 | LLVM_NODISCARD[[clang::warn_unused_result]] inline |
275 | typename std::enable_if<!is_simple_type<Y>::value, |
276 | typename cast_retty<X, const Y>::ret_type>::type |
277 | cast_or_null(const Y &Val) { |
278 | if (!Val) |
279 | return nullptr; |
280 | assert(isa<X>(Val) && "cast_or_null<Ty>() argument of incompatible type!")(static_cast <bool> (isa<X>(Val) && "cast_or_null<Ty>() argument of incompatible type!" ) ? void (0) : __assert_fail ("isa<X>(Val) && \"cast_or_null<Ty>() argument of incompatible type!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/Support/Casting.h" , 280, __extension__ __PRETTY_FUNCTION__)); |
281 | return cast<X>(Val); |
282 | } |
283 | |
284 | template <class X, class Y> |
285 | LLVM_NODISCARD[[clang::warn_unused_result]] inline |
286 | typename std::enable_if<!is_simple_type<Y>::value, |
287 | typename cast_retty<X, Y>::ret_type>::type |
288 | cast_or_null(Y &Val) { |
289 | if (!Val) |
290 | return nullptr; |
291 | assert(isa<X>(Val) && "cast_or_null<Ty>() argument of incompatible type!")(static_cast <bool> (isa<X>(Val) && "cast_or_null<Ty>() argument of incompatible type!" ) ? void (0) : __assert_fail ("isa<X>(Val) && \"cast_or_null<Ty>() argument of incompatible type!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/Support/Casting.h" , 291, __extension__ __PRETTY_FUNCTION__)); |
292 | return cast<X>(Val); |
293 | } |
294 | |
295 | template <class X, class Y> |
296 | LLVM_NODISCARD[[clang::warn_unused_result]] inline typename cast_retty<X, Y *>::ret_type |
297 | cast_or_null(Y *Val) { |
298 | if (!Val) return nullptr; |
299 | assert(isa<X>(Val) && "cast_or_null<Ty>() argument of incompatible type!")(static_cast <bool> (isa<X>(Val) && "cast_or_null<Ty>() argument of incompatible type!" ) ? void (0) : __assert_fail ("isa<X>(Val) && \"cast_or_null<Ty>() argument of incompatible type!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/Support/Casting.h" , 299, __extension__ __PRETTY_FUNCTION__)); |
300 | return cast<X>(Val); |
301 | } |
302 | |
303 | template <class X, class Y> |
304 | inline typename cast_retty<X, std::unique_ptr<Y>>::ret_type |
305 | cast_or_null(std::unique_ptr<Y> &&Val) { |
306 | if (!Val) |
307 | return nullptr; |
308 | return cast<X>(std::move(Val)); |
309 | } |
310 | |
311 | // dyn_cast<X> - Return the argument parameter cast to the specified type. This |
312 | // casting operator returns null if the argument is of the wrong type, so it can |
313 | // be used to test for a type as well as cast if successful. This should be |
314 | // used in the context of an if statement like this: |
315 | // |
316 | // if (const Instruction *I = dyn_cast<Instruction>(myVal)) { ... } |
317 | // |
318 | |
319 | template <class X, class Y> |
320 | LLVM_NODISCARD[[clang::warn_unused_result]] inline |
321 | typename std::enable_if<!is_simple_type<Y>::value, |
322 | typename cast_retty<X, const Y>::ret_type>::type |
323 | dyn_cast(const Y &Val) { |
324 | return isa<X>(Val) ? cast<X>(Val) : nullptr; |
325 | } |
326 | |
327 | template <class X, class Y> |
328 | LLVM_NODISCARD[[clang::warn_unused_result]] inline typename cast_retty<X, Y>::ret_type dyn_cast(Y &Val) { |
329 | return isa<X>(Val) ? cast<X>(Val) : nullptr; |
330 | } |
331 | |
332 | template <class X, class Y> |
333 | LLVM_NODISCARD[[clang::warn_unused_result]] inline typename cast_retty<X, Y *>::ret_type dyn_cast(Y *Val) { |
334 | return isa<X>(Val) ? cast<X>(Val) : nullptr; |
335 | } |
336 | |
337 | // dyn_cast_or_null<X> - Functionally identical to dyn_cast, except that a null |
338 | // value is accepted. |
339 | // |
340 | template <class X, class Y> |
341 | LLVM_NODISCARD[[clang::warn_unused_result]] inline |
342 | typename std::enable_if<!is_simple_type<Y>::value, |
343 | typename cast_retty<X, const Y>::ret_type>::type |
344 | dyn_cast_or_null(const Y &Val) { |
345 | return (Val && isa<X>(Val)) ? cast<X>(Val) : nullptr; |
346 | } |
347 | |
348 | template <class X, class Y> |
349 | LLVM_NODISCARD[[clang::warn_unused_result]] inline |
350 | typename std::enable_if<!is_simple_type<Y>::value, |
351 | typename cast_retty<X, Y>::ret_type>::type |
352 | dyn_cast_or_null(Y &Val) { |
353 | return (Val && isa<X>(Val)) ? cast<X>(Val) : nullptr; |
354 | } |
355 | |
356 | template <class X, class Y> |
357 | LLVM_NODISCARD[[clang::warn_unused_result]] inline typename cast_retty<X, Y *>::ret_type |
358 | dyn_cast_or_null(Y *Val) { |
359 | return (Val && isa<X>(Val)) ? cast<X>(Val) : nullptr; |
360 | } |
361 | |
362 | // unique_dyn_cast<X> - Given a unique_ptr<Y>, try to return a unique_ptr<X>, |
363 | // taking ownership of the input pointer iff isa<X>(Val) is true. If the |
364 | // cast is successful, From refers to nullptr on exit and the casted value |
365 | // is returned. If the cast is unsuccessful, the function returns nullptr |
366 | // and From is unchanged. |
367 | template <class X, class Y> |
368 | LLVM_NODISCARD[[clang::warn_unused_result]] inline auto unique_dyn_cast(std::unique_ptr<Y> &Val) |
369 | -> decltype(cast<X>(Val)) { |
370 | if (!isa<X>(Val)) |
371 | return nullptr; |
372 | return cast<X>(std::move(Val)); |
373 | } |
374 | |
375 | template <class X, class Y> |
376 | LLVM_NODISCARD[[clang::warn_unused_result]] inline auto unique_dyn_cast(std::unique_ptr<Y> &&Val) |
377 | -> decltype(cast<X>(Val)) { |
378 | return unique_dyn_cast<X, Y>(Val); |
379 | } |
380 | |
381 | // dyn_cast_or_null<X> - Functionally identical to unique_dyn_cast, except that |
382 | // a null value is accepted. |
383 | template <class X, class Y> |
384 | LLVM_NODISCARD[[clang::warn_unused_result]] inline auto unique_dyn_cast_or_null(std::unique_ptr<Y> &Val) |
385 | -> decltype(cast<X>(Val)) { |
386 | if (!Val) |
387 | return nullptr; |
388 | return unique_dyn_cast<X, Y>(Val); |
389 | } |
390 | |
391 | template <class X, class Y> |
392 | LLVM_NODISCARD[[clang::warn_unused_result]] inline auto unique_dyn_cast_or_null(std::unique_ptr<Y> &&Val) |
393 | -> decltype(cast<X>(Val)) { |
394 | return unique_dyn_cast_or_null<X, Y>(Val); |
395 | } |
396 | |
397 | } // end namespace llvm |
398 | |
399 | #endif // LLVM_SUPPORT_CASTING_H |
1 | //===- llvm/Instructions.h - Instruction subclass definitions ---*- C++ -*-===// |
2 | // |
3 | // The LLVM Compiler Infrastructure |
4 | // |
5 | // This file is distributed under the University of Illinois Open Source |
6 | // License. See LICENSE.TXT for details. |
7 | // |
8 | //===----------------------------------------------------------------------===// |
9 | // |
10 | // This file exposes the class definitions of all of the subclasses of the |
11 | // Instruction class. This is meant to be an easy way to get access to all |
12 | // instruction subclasses. |
13 | // |
14 | //===----------------------------------------------------------------------===// |
15 | |
16 | #ifndef LLVM_IR_INSTRUCTIONS_H |
17 | #define LLVM_IR_INSTRUCTIONS_H |
18 | |
19 | #include "llvm/ADT/ArrayRef.h" |
20 | #include "llvm/ADT/None.h" |
21 | #include "llvm/ADT/STLExtras.h" |
22 | #include "llvm/ADT/SmallVector.h" |
23 | #include "llvm/ADT/StringRef.h" |
24 | #include "llvm/ADT/Twine.h" |
25 | #include "llvm/ADT/iterator.h" |
26 | #include "llvm/ADT/iterator_range.h" |
27 | #include "llvm/IR/Attributes.h" |
28 | #include "llvm/IR/BasicBlock.h" |
29 | #include "llvm/IR/CallingConv.h" |
30 | #include "llvm/IR/Constant.h" |
31 | #include "llvm/IR/DerivedTypes.h" |
32 | #include "llvm/IR/Function.h" |
33 | #include "llvm/IR/InstrTypes.h" |
34 | #include "llvm/IR/Instruction.h" |
35 | #include "llvm/IR/OperandTraits.h" |
36 | #include "llvm/IR/Type.h" |
37 | #include "llvm/IR/Use.h" |
38 | #include "llvm/IR/User.h" |
39 | #include "llvm/IR/Value.h" |
40 | #include "llvm/Support/AtomicOrdering.h" |
41 | #include "llvm/Support/Casting.h" |
42 | #include "llvm/Support/ErrorHandling.h" |
43 | #include <cassert> |
44 | #include <cstddef> |
45 | #include <cstdint> |
46 | #include <iterator> |
47 | |
48 | namespace llvm { |
49 | |
50 | class APInt; |
51 | class ConstantInt; |
52 | class DataLayout; |
53 | class LLVMContext; |
54 | |
55 | //===----------------------------------------------------------------------===// |
56 | // AllocaInst Class |
57 | //===----------------------------------------------------------------------===// |
58 | |
59 | /// an instruction to allocate memory on the stack |
60 | class AllocaInst : public UnaryInstruction { |
61 | Type *AllocatedType; |
62 | |
63 | protected: |
64 | // Note: Instruction needs to be a friend here to call cloneImpl. |
65 | friend class Instruction; |
66 | |
67 | AllocaInst *cloneImpl() const; |
68 | |
69 | public: |
70 | explicit AllocaInst(Type *Ty, unsigned AddrSpace, |
71 | Value *ArraySize = nullptr, |
72 | const Twine &Name = "", |
73 | Instruction *InsertBefore = nullptr); |
74 | AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, |
75 | const Twine &Name, BasicBlock *InsertAtEnd); |
76 | |
77 | AllocaInst(Type *Ty, unsigned AddrSpace, |
78 | const Twine &Name, Instruction *InsertBefore = nullptr); |
79 | AllocaInst(Type *Ty, unsigned AddrSpace, |
80 | const Twine &Name, BasicBlock *InsertAtEnd); |
81 | |
82 | AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, unsigned Align, |
83 | const Twine &Name = "", Instruction *InsertBefore = nullptr); |
84 | AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, unsigned Align, |
85 | const Twine &Name, BasicBlock *InsertAtEnd); |
86 | |
87 | /// Return true if there is an allocation size parameter to the allocation |
88 | /// instruction that is not 1. |
89 | bool isArrayAllocation() const; |
90 | |
91 | /// Get the number of elements allocated. For a simple allocation of a single |
92 | /// element, this will return a constant 1 value. |
93 | const Value *getArraySize() const { return getOperand(0); } |
94 | Value *getArraySize() { return getOperand(0); } |
95 | |
96 | /// Overload to return most specific pointer type. |
97 | PointerType *getType() const { |
98 | return cast<PointerType>(Instruction::getType()); |
99 | } |
100 | |
101 | /// Return the type that is being allocated by the instruction. |
102 | Type *getAllocatedType() const { return AllocatedType; } |
103 | /// for use only in special circumstances that need to generically |
104 | /// transform a whole instruction (eg: IR linking and vectorization). |
105 | void setAllocatedType(Type *Ty) { AllocatedType = Ty; } |
106 | |
107 | /// Return the alignment of the memory that is being allocated by the |
108 | /// instruction. |
109 | unsigned getAlignment() const { |
110 | return (1u << (getSubclassDataFromInstruction() & 31)) >> 1; |
111 | } |
112 | void setAlignment(unsigned Align); |
113 | |
114 | /// Return true if this alloca is in the entry block of the function and is a |
115 | /// constant size. If so, the code generator will fold it into the |
116 | /// prolog/epilog code, so it is basically free. |
117 | bool isStaticAlloca() const; |
118 | |
119 | /// Return true if this alloca is used as an inalloca argument to a call. Such |
120 | /// allocas are never considered static even if they are in the entry block. |
121 | bool isUsedWithInAlloca() const { |
122 | return getSubclassDataFromInstruction() & 32; |
123 | } |
124 | |
125 | /// Specify whether this alloca is used to represent the arguments to a call. |
126 | void setUsedWithInAlloca(bool V) { |
127 | setInstructionSubclassData((getSubclassDataFromInstruction() & ~32) | |
128 | (V ? 32 : 0)); |
129 | } |
130 | |
131 | /// Return true if this alloca is used as a swifterror argument to a call. |
132 | bool isSwiftError() const { |
133 | return getSubclassDataFromInstruction() & 64; |
134 | } |
135 | |
136 | /// Specify whether this alloca is used to represent a swifterror. |
137 | void setSwiftError(bool V) { |
138 | setInstructionSubclassData((getSubclassDataFromInstruction() & ~64) | |
139 | (V ? 64 : 0)); |
140 | } |
141 | |
142 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
143 | static bool classof(const Instruction *I) { |
144 | return (I->getOpcode() == Instruction::Alloca); |
145 | } |
146 | static bool classof(const Value *V) { |
147 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
148 | } |
149 | |
150 | private: |
151 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
152 | // method so that subclasses cannot accidentally use it. |
153 | void setInstructionSubclassData(unsigned short D) { |
154 | Instruction::setInstructionSubclassData(D); |
155 | } |
156 | }; |
157 | |
158 | //===----------------------------------------------------------------------===// |
159 | // LoadInst Class |
160 | //===----------------------------------------------------------------------===// |
161 | |
162 | /// An instruction for reading from memory. This uses the SubclassData field in |
163 | /// Value to store whether or not the load is volatile. |
164 | class LoadInst : public UnaryInstruction { |
165 | void AssertOK(); |
166 | |
167 | protected: |
168 | // Note: Instruction needs to be a friend here to call cloneImpl. |
169 | friend class Instruction; |
170 | |
171 | LoadInst *cloneImpl() const; |
172 | |
173 | public: |
174 | LoadInst(Value *Ptr, const Twine &NameStr, Instruction *InsertBefore); |
175 | LoadInst(Value *Ptr, const Twine &NameStr, BasicBlock *InsertAtEnd); |
176 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile = false, |
177 | Instruction *InsertBefore = nullptr); |
178 | LoadInst(Value *Ptr, const Twine &NameStr, bool isVolatile = false, |
179 | Instruction *InsertBefore = nullptr) |
180 | : LoadInst(cast<PointerType>(Ptr->getType())->getElementType(), Ptr, |
181 | NameStr, isVolatile, InsertBefore) {} |
182 | LoadInst(Value *Ptr, const Twine &NameStr, bool isVolatile, |
183 | BasicBlock *InsertAtEnd); |
184 | LoadInst(Value *Ptr, const Twine &NameStr, bool isVolatile, unsigned Align, |
185 | Instruction *InsertBefore = nullptr) |
186 | : LoadInst(cast<PointerType>(Ptr->getType())->getElementType(), Ptr, |
187 | NameStr, isVolatile, Align, InsertBefore) {} |
188 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
189 | unsigned Align, Instruction *InsertBefore = nullptr); |
190 | LoadInst(Value *Ptr, const Twine &NameStr, bool isVolatile, |
191 | unsigned Align, BasicBlock *InsertAtEnd); |
192 | LoadInst(Value *Ptr, const Twine &NameStr, bool isVolatile, unsigned Align, |
193 | AtomicOrdering Order, SyncScope::ID SSID = SyncScope::System, |
194 | Instruction *InsertBefore = nullptr) |
195 | : LoadInst(cast<PointerType>(Ptr->getType())->getElementType(), Ptr, |
196 | NameStr, isVolatile, Align, Order, SSID, InsertBefore) {} |
197 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
198 | unsigned Align, AtomicOrdering Order, |
199 | SyncScope::ID SSID = SyncScope::System, |
200 | Instruction *InsertBefore = nullptr); |
201 | LoadInst(Value *Ptr, const Twine &NameStr, bool isVolatile, |
202 | unsigned Align, AtomicOrdering Order, SyncScope::ID SSID, |
203 | BasicBlock *InsertAtEnd); |
204 | LoadInst(Value *Ptr, const char *NameStr, Instruction *InsertBefore); |
205 | LoadInst(Value *Ptr, const char *NameStr, BasicBlock *InsertAtEnd); |
206 | LoadInst(Type *Ty, Value *Ptr, const char *NameStr = nullptr, |
207 | bool isVolatile = false, Instruction *InsertBefore = nullptr); |
208 | explicit LoadInst(Value *Ptr, const char *NameStr = nullptr, |
209 | bool isVolatile = false, |
210 | Instruction *InsertBefore = nullptr) |
211 | : LoadInst(cast<PointerType>(Ptr->getType())->getElementType(), Ptr, |
212 | NameStr, isVolatile, InsertBefore) {} |
213 | LoadInst(Value *Ptr, const char *NameStr, bool isVolatile, |
214 | BasicBlock *InsertAtEnd); |
215 | |
216 | /// Return true if this is a load from a volatile memory location. |
217 | bool isVolatile() const { return getSubclassDataFromInstruction() & 1; } |
218 | |
219 | /// Specify whether this is a volatile load or not. |
220 | void setVolatile(bool V) { |
221 | setInstructionSubclassData((getSubclassDataFromInstruction() & ~1) | |
222 | (V ? 1 : 0)); |
223 | } |
224 | |
225 | /// Return the alignment of the access that is being performed. |
226 | unsigned getAlignment() const { |
227 | return (1 << ((getSubclassDataFromInstruction() >> 1) & 31)) >> 1; |
228 | } |
229 | |
230 | void setAlignment(unsigned Align); |
231 | |
232 | /// Returns the ordering constraint of this load instruction. |
233 | AtomicOrdering getOrdering() const { |
234 | return AtomicOrdering((getSubclassDataFromInstruction() >> 7) & 7); |
235 | } |
236 | |
237 | /// Sets the ordering constraint of this load instruction. May not be Release |
238 | /// or AcquireRelease. |
239 | void setOrdering(AtomicOrdering Ordering) { |
240 | setInstructionSubclassData((getSubclassDataFromInstruction() & ~(7 << 7)) | |
241 | ((unsigned)Ordering << 7)); |
242 | } |
243 | |
244 | /// Returns the synchronization scope ID of this load instruction. |
245 | SyncScope::ID getSyncScopeID() const { |
246 | return SSID; |
247 | } |
248 | |
249 | /// Sets the synchronization scope ID of this load instruction. |
250 | void setSyncScopeID(SyncScope::ID SSID) { |
251 | this->SSID = SSID; |
252 | } |
253 | |
254 | /// Sets the ordering constraint and the synchronization scope ID of this load |
255 | /// instruction. |
256 | void setAtomic(AtomicOrdering Ordering, |
257 | SyncScope::ID SSID = SyncScope::System) { |
258 | setOrdering(Ordering); |
259 | setSyncScopeID(SSID); |
260 | } |
261 | |
262 | bool isSimple() const { return !isAtomic() && !isVolatile(); } |
263 | |
264 | bool isUnordered() const { |
265 | return (getOrdering() == AtomicOrdering::NotAtomic || |
266 | getOrdering() == AtomicOrdering::Unordered) && |
267 | !isVolatile(); |
268 | } |
269 | |
270 | Value *getPointerOperand() { return getOperand(0); } |
271 | const Value *getPointerOperand() const { return getOperand(0); } |
272 | static unsigned getPointerOperandIndex() { return 0U; } |
273 | Type *getPointerOperandType() const { return getPointerOperand()->getType(); } |
274 | |
275 | /// Returns the address space of the pointer operand. |
276 | unsigned getPointerAddressSpace() const { |
277 | return getPointerOperandType()->getPointerAddressSpace(); |
278 | } |
279 | |
280 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
281 | static bool classof(const Instruction *I) { |
282 | return I->getOpcode() == Instruction::Load; |
283 | } |
284 | static bool classof(const Value *V) { |
285 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
286 | } |
287 | |
288 | private: |
289 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
290 | // method so that subclasses cannot accidentally use it. |
291 | void setInstructionSubclassData(unsigned short D) { |
292 | Instruction::setInstructionSubclassData(D); |
293 | } |
294 | |
295 | /// The synchronization scope ID of this load instruction. Not quite enough |
296 | /// room in SubClassData for everything, so synchronization scope ID gets its |
297 | /// own field. |
298 | SyncScope::ID SSID; |
299 | }; |
300 | |
301 | //===----------------------------------------------------------------------===// |
302 | // StoreInst Class |
303 | //===----------------------------------------------------------------------===// |
304 | |
305 | /// An instruction for storing to memory. |
306 | class StoreInst : public Instruction { |
307 | void AssertOK(); |
308 | |
309 | protected: |
310 | // Note: Instruction needs to be a friend here to call cloneImpl. |
311 | friend class Instruction; |
312 | |
313 | StoreInst *cloneImpl() const; |
314 | |
315 | public: |
316 | StoreInst(Value *Val, Value *Ptr, Instruction *InsertBefore); |
317 | StoreInst(Value *Val, Value *Ptr, BasicBlock *InsertAtEnd); |
318 | StoreInst(Value *Val, Value *Ptr, bool isVolatile = false, |
319 | Instruction *InsertBefore = nullptr); |
320 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, BasicBlock *InsertAtEnd); |
321 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, |
322 | unsigned Align, Instruction *InsertBefore = nullptr); |
323 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, |
324 | unsigned Align, BasicBlock *InsertAtEnd); |
325 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, |
326 | unsigned Align, AtomicOrdering Order, |
327 | SyncScope::ID SSID = SyncScope::System, |
328 | Instruction *InsertBefore = nullptr); |
329 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, |
330 | unsigned Align, AtomicOrdering Order, SyncScope::ID SSID, |
331 | BasicBlock *InsertAtEnd); |
332 | |
333 | // allocate space for exactly two operands |
334 | void *operator new(size_t s) { |
335 | return User::operator new(s, 2); |
336 | } |
337 | |
338 | /// Return true if this is a store to a volatile memory location. |
339 | bool isVolatile() const { return getSubclassDataFromInstruction() & 1; } |
340 | |
341 | /// Specify whether this is a volatile store or not. |
342 | void setVolatile(bool V) { |
343 | setInstructionSubclassData((getSubclassDataFromInstruction() & ~1) | |
344 | (V ? 1 : 0)); |
345 | } |
346 | |
347 | /// Transparently provide more efficient getOperand methods. |
348 | 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; |
349 | |
350 | /// Return the alignment of the access that is being performed |
351 | unsigned getAlignment() const { |
352 | return (1 << ((getSubclassDataFromInstruction() >> 1) & 31)) >> 1; |
353 | } |
354 | |
355 | void setAlignment(unsigned Align); |
356 | |
357 | /// Returns the ordering constraint of this store instruction. |
358 | AtomicOrdering getOrdering() const { |
359 | return AtomicOrdering((getSubclassDataFromInstruction() >> 7) & 7); |
360 | } |
361 | |
362 | /// Sets the ordering constraint of this store instruction. May not be |
363 | /// Acquire or AcquireRelease. |
364 | void setOrdering(AtomicOrdering Ordering) { |
365 | setInstructionSubclassData((getSubclassDataFromInstruction() & ~(7 << 7)) | |
366 | ((unsigned)Ordering << 7)); |
367 | } |
368 | |
369 | /// Returns the synchronization scope ID of this store instruction. |
370 | SyncScope::ID getSyncScopeID() const { |
371 | return SSID; |
372 | } |
373 | |
374 | /// Sets the synchronization scope ID of this store instruction. |
375 | void setSyncScopeID(SyncScope::ID SSID) { |
376 | this->SSID = SSID; |
377 | } |
378 | |
379 | /// Sets the ordering constraint and the synchronization scope ID of this |
380 | /// store instruction. |
381 | void setAtomic(AtomicOrdering Ordering, |
382 | SyncScope::ID SSID = SyncScope::System) { |
383 | setOrdering(Ordering); |
384 | setSyncScopeID(SSID); |
385 | } |
386 | |
387 | bool isSimple() const { return !isAtomic() && !isVolatile(); } |
388 | |
389 | bool isUnordered() const { |
390 | return (getOrdering() == AtomicOrdering::NotAtomic || |
391 | getOrdering() == AtomicOrdering::Unordered) && |
392 | !isVolatile(); |
393 | } |
394 | |
395 | Value *getValueOperand() { return getOperand(0); } |
396 | const Value *getValueOperand() const { return getOperand(0); } |
397 | |
398 | Value *getPointerOperand() { return getOperand(1); } |
399 | const Value *getPointerOperand() const { return getOperand(1); } |
400 | static unsigned getPointerOperandIndex() { return 1U; } |
401 | Type *getPointerOperandType() const { return getPointerOperand()->getType(); } |
402 | |
403 | /// Returns the address space of the pointer operand. |
404 | unsigned getPointerAddressSpace() const { |
405 | return getPointerOperandType()->getPointerAddressSpace(); |
406 | } |
407 | |
408 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
409 | static bool classof(const Instruction *I) { |
410 | return I->getOpcode() == Instruction::Store; |
411 | } |
412 | static bool classof(const Value *V) { |
413 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
414 | } |
415 | |
416 | private: |
417 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
418 | // method so that subclasses cannot accidentally use it. |
419 | void setInstructionSubclassData(unsigned short D) { |
420 | Instruction::setInstructionSubclassData(D); |
421 | } |
422 | |
423 | /// The synchronization scope ID of this store instruction. Not quite enough |
424 | /// room in SubClassData for everything, so synchronization scope ID gets its |
425 | /// own field. |
426 | SyncScope::ID SSID; |
427 | }; |
428 | |
429 | template <> |
430 | struct OperandTraits<StoreInst> : public FixedNumOperandTraits<StoreInst, 2> { |
431 | }; |
432 | |
433 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 433, __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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 433, __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); } |
434 | |
435 | //===----------------------------------------------------------------------===// |
436 | // FenceInst Class |
437 | //===----------------------------------------------------------------------===// |
438 | |
439 | /// An instruction for ordering other memory operations. |
440 | class FenceInst : public Instruction { |
441 | void Init(AtomicOrdering Ordering, SyncScope::ID SSID); |
442 | |
443 | protected: |
444 | // Note: Instruction needs to be a friend here to call cloneImpl. |
445 | friend class Instruction; |
446 | |
447 | FenceInst *cloneImpl() const; |
448 | |
449 | public: |
450 | // Ordering may only be Acquire, Release, AcquireRelease, or |
451 | // SequentiallyConsistent. |
452 | FenceInst(LLVMContext &C, AtomicOrdering Ordering, |
453 | SyncScope::ID SSID = SyncScope::System, |
454 | Instruction *InsertBefore = nullptr); |
455 | FenceInst(LLVMContext &C, AtomicOrdering Ordering, SyncScope::ID SSID, |
456 | BasicBlock *InsertAtEnd); |
457 | |
458 | // allocate space for exactly zero operands |
459 | void *operator new(size_t s) { |
460 | return User::operator new(s, 0); |
461 | } |
462 | |
463 | /// Returns the ordering constraint of this fence instruction. |
464 | AtomicOrdering getOrdering() const { |
465 | return AtomicOrdering(getSubclassDataFromInstruction() >> 1); |
466 | } |
467 | |
468 | /// Sets the ordering constraint of this fence instruction. May only be |
469 | /// Acquire, Release, AcquireRelease, or SequentiallyConsistent. |
470 | void setOrdering(AtomicOrdering Ordering) { |
471 | setInstructionSubclassData((getSubclassDataFromInstruction() & 1) | |
472 | ((unsigned)Ordering << 1)); |
473 | } |
474 | |
475 | /// Returns the synchronization scope ID of this fence instruction. |
476 | SyncScope::ID getSyncScopeID() const { |
477 | return SSID; |
478 | } |
479 | |
480 | /// Sets the synchronization scope ID of this fence instruction. |
481 | void setSyncScopeID(SyncScope::ID SSID) { |
482 | this->SSID = SSID; |
483 | } |
484 | |
485 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
486 | static bool classof(const Instruction *I) { |
487 | return I->getOpcode() == Instruction::Fence; |
488 | } |
489 | static bool classof(const Value *V) { |
490 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
491 | } |
492 | |
493 | private: |
494 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
495 | // method so that subclasses cannot accidentally use it. |
496 | void setInstructionSubclassData(unsigned short D) { |
497 | Instruction::setInstructionSubclassData(D); |
498 | } |
499 | |
500 | /// The synchronization scope ID of this fence instruction. Not quite enough |
501 | /// room in SubClassData for everything, so synchronization scope ID gets its |
502 | /// own field. |
503 | SyncScope::ID SSID; |
504 | }; |
505 | |
506 | //===----------------------------------------------------------------------===// |
507 | // AtomicCmpXchgInst Class |
508 | //===----------------------------------------------------------------------===// |
509 | |
510 | /// an instruction that atomically checks whether a |
511 | /// specified value is in a memory location, and, if it is, stores a new value |
512 | /// there. Returns the value that was loaded. |
513 | /// |
514 | class AtomicCmpXchgInst : public Instruction { |
515 | void Init(Value *Ptr, Value *Cmp, Value *NewVal, |
516 | AtomicOrdering SuccessOrdering, AtomicOrdering FailureOrdering, |
517 | SyncScope::ID SSID); |
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, |
527 | AtomicOrdering SuccessOrdering, |
528 | AtomicOrdering FailureOrdering, |
529 | SyncScope::ID SSID, Instruction *InsertBefore = nullptr); |
530 | AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, |
531 | AtomicOrdering SuccessOrdering, |
532 | AtomicOrdering FailureOrdering, |
533 | SyncScope::ID SSID, BasicBlock *InsertAtEnd); |
534 | |
535 | // allocate space for exactly three operands |
536 | void *operator new(size_t s) { |
537 | return User::operator new(s, 3); |
538 | } |
539 | |
540 | /// Return true if this is a cmpxchg from a volatile memory |
541 | /// location. |
542 | /// |
543 | bool isVolatile() const { |
544 | return getSubclassDataFromInstruction() & 1; |
545 | } |
546 | |
547 | /// Specify whether this is a volatile cmpxchg. |
548 | /// |
549 | void setVolatile(bool V) { |
550 | setInstructionSubclassData((getSubclassDataFromInstruction() & ~1) | |
551 | (unsigned)V); |
552 | } |
553 | |
554 | /// Return true if this cmpxchg may spuriously fail. |
555 | bool isWeak() const { |
556 | return getSubclassDataFromInstruction() & 0x100; |
557 | } |
558 | |
559 | void setWeak(bool IsWeak) { |
560 | setInstructionSubclassData((getSubclassDataFromInstruction() & ~0x100) | |
561 | (IsWeak << 8)); |
562 | } |
563 | |
564 | /// Transparently provide more efficient getOperand methods. |
565 | 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; |
566 | |
567 | /// Returns the success ordering constraint of this cmpxchg instruction. |
568 | AtomicOrdering getSuccessOrdering() const { |
569 | return AtomicOrdering((getSubclassDataFromInstruction() >> 2) & 7); |
570 | } |
571 | |
572 | /// Sets the success ordering constraint of this cmpxchg instruction. |
573 | void setSuccessOrdering(AtomicOrdering Ordering) { |
574 | assert(Ordering != AtomicOrdering::NotAtomic &&(static_cast <bool> (Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic.") ? void (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 575, __extension__ __PRETTY_FUNCTION__)) |
575 | "CmpXchg instructions can only be atomic.")(static_cast <bool> (Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic.") ? void (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 575, __extension__ __PRETTY_FUNCTION__)); |
576 | setInstructionSubclassData((getSubclassDataFromInstruction() & ~0x1c) | |
577 | ((unsigned)Ordering << 2)); |
578 | } |
579 | |
580 | /// Returns the failure ordering constraint of this cmpxchg instruction. |
581 | AtomicOrdering getFailureOrdering() const { |
582 | return AtomicOrdering((getSubclassDataFromInstruction() >> 5) & 7); |
583 | } |
584 | |
585 | /// Sets the failure ordering constraint of this cmpxchg instruction. |
586 | void setFailureOrdering(AtomicOrdering Ordering) { |
587 | assert(Ordering != AtomicOrdering::NotAtomic &&(static_cast <bool> (Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic.") ? void (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 588, __extension__ __PRETTY_FUNCTION__)) |
588 | "CmpXchg instructions can only be atomic.")(static_cast <bool> (Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic.") ? void (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 588, __extension__ __PRETTY_FUNCTION__)); |
589 | setInstructionSubclassData((getSubclassDataFromInstruction() & ~0xe0) | |
590 | ((unsigned)Ordering << 5)); |
591 | } |
592 | |
593 | /// Returns the synchronization scope ID of this cmpxchg instruction. |
594 | SyncScope::ID getSyncScopeID() const { |
595 | return SSID; |
596 | } |
597 | |
598 | /// Sets the synchronization scope ID of this cmpxchg instruction. |
599 | void setSyncScopeID(SyncScope::ID SSID) { |
600 | this->SSID = SSID; |
601 | } |
602 | |
603 | Value *getPointerOperand() { return getOperand(0); } |
604 | const Value *getPointerOperand() const { return getOperand(0); } |
605 | static unsigned getPointerOperandIndex() { return 0U; } |
606 | |
607 | Value *getCompareOperand() { return getOperand(1); } |
608 | const Value *getCompareOperand() const { return getOperand(1); } |
609 | |
610 | Value *getNewValOperand() { return getOperand(2); } |
611 | const Value *getNewValOperand() const { return getOperand(2); } |
612 | |
613 | /// Returns the address space of the pointer operand. |
614 | unsigned getPointerAddressSpace() const { |
615 | return getPointerOperand()->getType()->getPointerAddressSpace(); |
616 | } |
617 | |
618 | /// Returns the strongest permitted ordering on failure, given the |
619 | /// desired ordering on success. |
620 | /// |
621 | /// If the comparison in a cmpxchg operation fails, there is no atomic store |
622 | /// so release semantics cannot be provided. So this function drops explicit |
623 | /// Release requests from the AtomicOrdering. A SequentiallyConsistent |
624 | /// operation would remain SequentiallyConsistent. |
625 | static AtomicOrdering |
626 | getStrongestFailureOrdering(AtomicOrdering SuccessOrdering) { |
627 | switch (SuccessOrdering) { |
628 | default: |
629 | llvm_unreachable("invalid cmpxchg success ordering")::llvm::llvm_unreachable_internal("invalid cmpxchg success ordering" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 629); |
630 | case AtomicOrdering::Release: |
631 | case AtomicOrdering::Monotonic: |
632 | return AtomicOrdering::Monotonic; |
633 | case AtomicOrdering::AcquireRelease: |
634 | case AtomicOrdering::Acquire: |
635 | return AtomicOrdering::Acquire; |
636 | case AtomicOrdering::SequentiallyConsistent: |
637 | return AtomicOrdering::SequentiallyConsistent; |
638 | } |
639 | } |
640 | |
641 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
642 | static bool classof(const Instruction *I) { |
643 | return I->getOpcode() == Instruction::AtomicCmpXchg; |
644 | } |
645 | static bool classof(const Value *V) { |
646 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
647 | } |
648 | |
649 | private: |
650 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
651 | // method so that subclasses cannot accidentally use it. |
652 | void setInstructionSubclassData(unsigned short D) { |
653 | Instruction::setInstructionSubclassData(D); |
654 | } |
655 | |
656 | /// The synchronization scope ID of this cmpxchg instruction. Not quite |
657 | /// enough room in SubClassData for everything, so synchronization scope ID |
658 | /// gets its own field. |
659 | SyncScope::ID SSID; |
660 | }; |
661 | |
662 | template <> |
663 | struct OperandTraits<AtomicCmpXchgInst> : |
664 | public FixedNumOperandTraits<AtomicCmpXchgInst, 3> { |
665 | }; |
666 | |
667 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 667, __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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 667, __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 ); } |
668 | |
669 | //===----------------------------------------------------------------------===// |
670 | // AtomicRMWInst Class |
671 | //===----------------------------------------------------------------------===// |
672 | |
673 | /// an instruction that atomically reads a memory location, |
674 | /// combines it with another value, and then stores the result back. Returns |
675 | /// the old value. |
676 | /// |
677 | class AtomicRMWInst : public Instruction { |
678 | protected: |
679 | // Note: Instruction needs to be a friend here to call cloneImpl. |
680 | friend class Instruction; |
681 | |
682 | AtomicRMWInst *cloneImpl() const; |
683 | |
684 | public: |
685 | /// This enumeration lists the possible modifications atomicrmw can make. In |
686 | /// the descriptions, 'p' is the pointer to the instruction's memory location, |
687 | /// 'old' is the initial value of *p, and 'v' is the other value passed to the |
688 | /// instruction. These instructions always return 'old'. |
689 | enum BinOp { |
690 | /// *p = v |
691 | Xchg, |
692 | /// *p = old + v |
693 | Add, |
694 | /// *p = old - v |
695 | Sub, |
696 | /// *p = old & v |
697 | And, |
698 | /// *p = ~(old & v) |
699 | Nand, |
700 | /// *p = old | v |
701 | Or, |
702 | /// *p = old ^ v |
703 | Xor, |
704 | /// *p = old >signed v ? old : v |
705 | Max, |
706 | /// *p = old <signed v ? old : v |
707 | Min, |
708 | /// *p = old >unsigned v ? old : v |
709 | UMax, |
710 | /// *p = old <unsigned v ? old : v |
711 | UMin, |
712 | |
713 | FIRST_BINOP = Xchg, |
714 | LAST_BINOP = UMin, |
715 | BAD_BINOP |
716 | }; |
717 | |
718 | AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, |
719 | AtomicOrdering Ordering, SyncScope::ID SSID, |
720 | Instruction *InsertBefore = nullptr); |
721 | AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, |
722 | AtomicOrdering Ordering, SyncScope::ID SSID, |
723 | BasicBlock *InsertAtEnd); |
724 | |
725 | // allocate space for exactly two operands |
726 | void *operator new(size_t s) { |
727 | return User::operator new(s, 2); |
728 | } |
729 | |
730 | BinOp getOperation() const { |
731 | return static_cast<BinOp>(getSubclassDataFromInstruction() >> 5); |
732 | } |
733 | |
734 | void setOperation(BinOp Operation) { |
735 | unsigned short SubclassData = getSubclassDataFromInstruction(); |
736 | setInstructionSubclassData((SubclassData & 31) | |
737 | (Operation << 5)); |
738 | } |
739 | |
740 | /// Return true if this is a RMW on a volatile memory location. |
741 | /// |
742 | bool isVolatile() const { |
743 | return getSubclassDataFromInstruction() & 1; |
744 | } |
745 | |
746 | /// Specify whether this is a volatile RMW or not. |
747 | /// |
748 | void setVolatile(bool V) { |
749 | setInstructionSubclassData((getSubclassDataFromInstruction() & ~1) | |
750 | (unsigned)V); |
751 | } |
752 | |
753 | /// Transparently provide more efficient getOperand methods. |
754 | 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; |
755 | |
756 | /// Returns the ordering constraint of this rmw instruction. |
757 | AtomicOrdering getOrdering() const { |
758 | return AtomicOrdering((getSubclassDataFromInstruction() >> 2) & 7); |
759 | } |
760 | |
761 | /// Sets the ordering constraint of this rmw instruction. |
762 | void setOrdering(AtomicOrdering Ordering) { |
763 | 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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 764, __extension__ __PRETTY_FUNCTION__)) |
764 | "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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 764, __extension__ __PRETTY_FUNCTION__)); |
765 | setInstructionSubclassData((getSubclassDataFromInstruction() & ~(7 << 2)) | |
766 | ((unsigned)Ordering << 2)); |
767 | } |
768 | |
769 | /// Returns the synchronization scope ID of this rmw instruction. |
770 | SyncScope::ID getSyncScopeID() const { |
771 | return SSID; |
772 | } |
773 | |
774 | /// Sets the synchronization scope ID of this rmw instruction. |
775 | void setSyncScopeID(SyncScope::ID SSID) { |
776 | this->SSID = SSID; |
777 | } |
778 | |
779 | Value *getPointerOperand() { return getOperand(0); } |
780 | const Value *getPointerOperand() const { return getOperand(0); } |
781 | static unsigned getPointerOperandIndex() { return 0U; } |
782 | |
783 | Value *getValOperand() { return getOperand(1); } |
784 | const Value *getValOperand() const { return getOperand(1); } |
785 | |
786 | /// Returns the address space of the pointer operand. |
787 | unsigned getPointerAddressSpace() const { |
788 | return getPointerOperand()->getType()->getPointerAddressSpace(); |
789 | } |
790 | |
791 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
792 | static bool classof(const Instruction *I) { |
793 | return I->getOpcode() == Instruction::AtomicRMW; |
794 | } |
795 | static bool classof(const Value *V) { |
796 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
797 | } |
798 | |
799 | private: |
800 | void Init(BinOp Operation, Value *Ptr, Value *Val, |
801 | AtomicOrdering Ordering, SyncScope::ID SSID); |
802 | |
803 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
804 | // method so that subclasses cannot accidentally use it. |
805 | void setInstructionSubclassData(unsigned short D) { |
806 | Instruction::setInstructionSubclassData(D); |
807 | } |
808 | |
809 | /// The synchronization scope ID of this rmw instruction. Not quite enough |
810 | /// room in SubClassData for everything, so synchronization scope ID gets its |
811 | /// own field. |
812 | SyncScope::ID SSID; |
813 | }; |
814 | |
815 | template <> |
816 | struct OperandTraits<AtomicRMWInst> |
817 | : public FixedNumOperandTraits<AtomicRMWInst,2> { |
818 | }; |
819 | |
820 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 820, __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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 820, __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); } |
821 | |
822 | //===----------------------------------------------------------------------===// |
823 | // GetElementPtrInst Class |
824 | //===----------------------------------------------------------------------===// |
825 | |
826 | // checkGEPType - Simple wrapper function to give a better assertion failure |
827 | // message on bad indexes for a gep instruction. |
828 | // |
829 | inline Type *checkGEPType(Type *Ty) { |
830 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 830, __extension__ __PRETTY_FUNCTION__)); |
831 | return Ty; |
832 | } |
833 | |
834 | /// an instruction for type-safe pointer arithmetic to |
835 | /// access elements of arrays and structs |
836 | /// |
837 | class GetElementPtrInst : public Instruction { |
838 | Type *SourceElementType; |
839 | Type *ResultElementType; |
840 | |
841 | GetElementPtrInst(const GetElementPtrInst &GEPI); |
842 | |
843 | /// Constructors - Create a getelementptr instruction with a base pointer an |
844 | /// list of indices. The first ctor can optionally insert before an existing |
845 | /// instruction, the second appends the new instruction to the specified |
846 | /// BasicBlock. |
847 | inline GetElementPtrInst(Type *PointeeType, Value *Ptr, |
848 | ArrayRef<Value *> IdxList, unsigned Values, |
849 | const Twine &NameStr, Instruction *InsertBefore); |
850 | inline GetElementPtrInst(Type *PointeeType, Value *Ptr, |
851 | ArrayRef<Value *> IdxList, unsigned Values, |
852 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
853 | |
854 | void init(Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr); |
855 | |
856 | protected: |
857 | // Note: Instruction needs to be a friend here to call cloneImpl. |
858 | friend class Instruction; |
859 | |
860 | GetElementPtrInst *cloneImpl() const; |
861 | |
862 | public: |
863 | static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr, |
864 | ArrayRef<Value *> IdxList, |
865 | const Twine &NameStr = "", |
866 | Instruction *InsertBefore = nullptr) { |
867 | unsigned Values = 1 + unsigned(IdxList.size()); |
868 | if (!PointeeType) |
869 | PointeeType = |
870 | cast<PointerType>(Ptr->getType()->getScalarType())->getElementType(); |
871 | else |
872 | assert((static_cast <bool> (PointeeType == cast<PointerType >(Ptr->getType()->getScalarType())->getElementType ()) ? void (0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 874, __extension__ __PRETTY_FUNCTION__)) |
873 | PointeeType ==(static_cast <bool> (PointeeType == cast<PointerType >(Ptr->getType()->getScalarType())->getElementType ()) ? void (0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 874, __extension__ __PRETTY_FUNCTION__)) |
874 | cast<PointerType>(Ptr->getType()->getScalarType())->getElementType())(static_cast <bool> (PointeeType == cast<PointerType >(Ptr->getType()->getScalarType())->getElementType ()) ? void (0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 874, __extension__ __PRETTY_FUNCTION__)); |
875 | return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values, |
876 | NameStr, InsertBefore); |
877 | } |
878 | |
879 | static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr, |
880 | ArrayRef<Value *> IdxList, |
881 | const Twine &NameStr, |
882 | BasicBlock *InsertAtEnd) { |
883 | unsigned Values = 1 + unsigned(IdxList.size()); |
884 | if (!PointeeType) |
885 | PointeeType = |
886 | cast<PointerType>(Ptr->getType()->getScalarType())->getElementType(); |
887 | else |
888 | assert((static_cast <bool> (PointeeType == cast<PointerType >(Ptr->getType()->getScalarType())->getElementType ()) ? void (0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 890, __extension__ __PRETTY_FUNCTION__)) |
889 | PointeeType ==(static_cast <bool> (PointeeType == cast<PointerType >(Ptr->getType()->getScalarType())->getElementType ()) ? void (0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 890, __extension__ __PRETTY_FUNCTION__)) |
890 | cast<PointerType>(Ptr->getType()->getScalarType())->getElementType())(static_cast <bool> (PointeeType == cast<PointerType >(Ptr->getType()->getScalarType())->getElementType ()) ? void (0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 890, __extension__ __PRETTY_FUNCTION__)); |
891 | return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values, |
892 | NameStr, InsertAtEnd); |
893 | } |
894 | |
895 | /// Create an "inbounds" getelementptr. See the documentation for the |
896 | /// "inbounds" flag in LangRef.html for details. |
897 | static GetElementPtrInst *CreateInBounds(Value *Ptr, |
898 | ArrayRef<Value *> IdxList, |
899 | const Twine &NameStr = "", |
900 | Instruction *InsertBefore = nullptr){ |
901 | return CreateInBounds(nullptr, Ptr, IdxList, NameStr, InsertBefore); |
902 | } |
903 | |
904 | static GetElementPtrInst * |
905 | CreateInBounds(Type *PointeeType, Value *Ptr, ArrayRef<Value *> IdxList, |
906 | const Twine &NameStr = "", |
907 | Instruction *InsertBefore = nullptr) { |
908 | GetElementPtrInst *GEP = |
909 | Create(PointeeType, Ptr, IdxList, NameStr, InsertBefore); |
910 | GEP->setIsInBounds(true); |
911 | return GEP; |
912 | } |
913 | |
914 | static GetElementPtrInst *CreateInBounds(Value *Ptr, |
915 | ArrayRef<Value *> IdxList, |
916 | const Twine &NameStr, |
917 | BasicBlock *InsertAtEnd) { |
918 | return CreateInBounds(nullptr, Ptr, IdxList, NameStr, InsertAtEnd); |
919 | } |
920 | |
921 | static GetElementPtrInst *CreateInBounds(Type *PointeeType, Value *Ptr, |
922 | ArrayRef<Value *> IdxList, |
923 | const Twine &NameStr, |
924 | BasicBlock *InsertAtEnd) { |
925 | GetElementPtrInst *GEP = |
926 | Create(PointeeType, Ptr, IdxList, NameStr, InsertAtEnd); |
927 | GEP->setIsInBounds(true); |
928 | return GEP; |
929 | } |
930 | |
931 | /// Transparently provide more efficient getOperand methods. |
932 | 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; |
933 | |
934 | Type *getSourceElementType() const { return SourceElementType; } |
935 | |
936 | void setSourceElementType(Type *Ty) { SourceElementType = Ty; } |
937 | void setResultElementType(Type *Ty) { ResultElementType = Ty; } |
938 | |
939 | Type *getResultElementType() const { |
940 | assert(ResultElementType ==(static_cast <bool> (ResultElementType == cast<PointerType >(getType()->getScalarType())->getElementType()) ? void (0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 941, __extension__ __PRETTY_FUNCTION__)) |
941 | cast<PointerType>(getType()->getScalarType())->getElementType())(static_cast <bool> (ResultElementType == cast<PointerType >(getType()->getScalarType())->getElementType()) ? void (0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 941, __extension__ __PRETTY_FUNCTION__)); |
942 | return ResultElementType; |
943 | } |
944 | |
945 | /// Returns the address space of this instruction's pointer type. |
946 | unsigned getAddressSpace() const { |
947 | // Note that this is always the same as the pointer operand's address space |
948 | // and that is cheaper to compute, so cheat here. |
949 | return getPointerAddressSpace(); |
950 | } |
951 | |
952 | /// Returns the type of the element that would be loaded with |
953 | /// a load instruction with the specified parameters. |
954 | /// |
955 | /// Null is returned if the indices are invalid for the specified |
956 | /// pointer type. |
957 | /// |
958 | static Type *getIndexedType(Type *Ty, ArrayRef<Value *> IdxList); |
959 | static Type *getIndexedType(Type *Ty, ArrayRef<Constant *> IdxList); |
960 | static Type *getIndexedType(Type *Ty, ArrayRef<uint64_t> IdxList); |
961 | |
962 | inline op_iterator idx_begin() { return op_begin()+1; } |
963 | inline const_op_iterator idx_begin() const { return op_begin()+1; } |
964 | inline op_iterator idx_end() { return op_end(); } |
965 | inline const_op_iterator idx_end() const { return op_end(); } |
966 | |
967 | inline iterator_range<op_iterator> indices() { |
968 | return make_range(idx_begin(), idx_end()); |
969 | } |
970 | |
971 | inline iterator_range<const_op_iterator> indices() const { |
972 | return make_range(idx_begin(), idx_end()); |
973 | } |
974 | |
975 | Value *getPointerOperand() { |
976 | return getOperand(0); |
977 | } |
978 | const Value *getPointerOperand() const { |
979 | return getOperand(0); |
980 | } |
981 | static unsigned getPointerOperandIndex() { |
982 | return 0U; // get index for modifying correct operand. |
983 | } |
984 | |
985 | /// Method to return the pointer operand as a |
986 | /// PointerType. |
987 | Type *getPointerOperandType() const { |
988 | return getPointerOperand()->getType(); |
989 | } |
990 | |
991 | /// Returns the address space of the pointer operand. |
992 | unsigned getPointerAddressSpace() const { |
993 | return getPointerOperandType()->getPointerAddressSpace(); |
994 | } |
995 | |
996 | /// Returns the pointer type returned by the GEP |
997 | /// instruction, which may be a vector of pointers. |
998 | static Type *getGEPReturnType(Value *Ptr, ArrayRef<Value *> IdxList) { |
999 | return getGEPReturnType( |
1000 | cast<PointerType>(Ptr->getType()->getScalarType())->getElementType(), |
1001 | Ptr, IdxList); |
1002 | } |
1003 | static Type *getGEPReturnType(Type *ElTy, Value *Ptr, |
1004 | ArrayRef<Value *> IdxList) { |
1005 | Type *PtrTy = PointerType::get(checkGEPType(getIndexedType(ElTy, IdxList)), |
1006 | Ptr->getType()->getPointerAddressSpace()); |
1007 | // Vector GEP |
1008 | if (Ptr->getType()->isVectorTy()) { |
1009 | unsigned NumElem = Ptr->getType()->getVectorNumElements(); |
1010 | return VectorType::get(PtrTy, NumElem); |
1011 | } |
1012 | for (Value *Index : IdxList) |
1013 | if (Index->getType()->isVectorTy()) { |
1014 | unsigned NumElem = Index->getType()->getVectorNumElements(); |
1015 | return VectorType::get(PtrTy, NumElem); |
1016 | } |
1017 | // Scalar GEP |
1018 | return PtrTy; |
1019 | } |
1020 | |
1021 | unsigned getNumIndices() const { // Note: always non-negative |
1022 | return getNumOperands() - 1; |
1023 | } |
1024 | |
1025 | bool hasIndices() const { |
1026 | return getNumOperands() > 1; |
1027 | } |
1028 | |
1029 | /// Return true if all of the indices of this GEP are |
1030 | /// zeros. If so, the result pointer and the first operand have the same |
1031 | /// value, just potentially different types. |
1032 | bool hasAllZeroIndices() const; |
1033 | |
1034 | /// Return true if all of the indices of this GEP are |
1035 | /// constant integers. If so, the result pointer and the first operand have |
1036 | /// a constant offset between them. |
1037 | bool hasAllConstantIndices() const; |
1038 | |
1039 | /// Set or clear the inbounds flag on this GEP instruction. |
1040 | /// See LangRef.html for the meaning of inbounds on a getelementptr. |
1041 | void setIsInBounds(bool b = true); |
1042 | |
1043 | /// Determine whether the GEP has the inbounds flag. |
1044 | bool isInBounds() const; |
1045 | |
1046 | /// Accumulate the constant address offset of this GEP if possible. |
1047 | /// |
1048 | /// This routine accepts an APInt into which it will accumulate the constant |
1049 | /// offset of this GEP if the GEP is in fact constant. If the GEP is not |
1050 | /// all-constant, it returns false and the value of the offset APInt is |
1051 | /// undefined (it is *not* preserved!). The APInt passed into this routine |
1052 | /// must be at least as wide as the IntPtr type for the address space of |
1053 | /// the base GEP pointer. |
1054 | bool accumulateConstantOffset(const DataLayout &DL, APInt &Offset) const; |
1055 | |
1056 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1057 | static bool classof(const Instruction *I) { |
1058 | return (I->getOpcode() == Instruction::GetElementPtr); |
1059 | } |
1060 | static bool classof(const Value *V) { |
1061 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1062 | } |
1063 | }; |
1064 | |
1065 | template <> |
1066 | struct OperandTraits<GetElementPtrInst> : |
1067 | public VariadicOperandTraits<GetElementPtrInst, 1> { |
1068 | }; |
1069 | |
1070 | GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr, |
1071 | ArrayRef<Value *> IdxList, unsigned Values, |
1072 | const Twine &NameStr, |
1073 | Instruction *InsertBefore) |
1074 | : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr, |
1075 | OperandTraits<GetElementPtrInst>::op_end(this) - Values, |
1076 | Values, InsertBefore), |
1077 | SourceElementType(PointeeType), |
1078 | ResultElementType(getIndexedType(PointeeType, IdxList)) { |
1079 | assert(ResultElementType ==(static_cast <bool> (ResultElementType == cast<PointerType >(getType()->getScalarType())->getElementType()) ? void (0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1080, __extension__ __PRETTY_FUNCTION__)) |
1080 | cast<PointerType>(getType()->getScalarType())->getElementType())(static_cast <bool> (ResultElementType == cast<PointerType >(getType()->getScalarType())->getElementType()) ? void (0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1080, __extension__ __PRETTY_FUNCTION__)); |
1081 | init(Ptr, IdxList, NameStr); |
1082 | } |
1083 | |
1084 | GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr, |
1085 | ArrayRef<Value *> IdxList, unsigned Values, |
1086 | const Twine &NameStr, |
1087 | BasicBlock *InsertAtEnd) |
1088 | : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr, |
1089 | OperandTraits<GetElementPtrInst>::op_end(this) - Values, |
1090 | Values, InsertAtEnd), |
1091 | SourceElementType(PointeeType), |
1092 | ResultElementType(getIndexedType(PointeeType, IdxList)) { |
1093 | assert(ResultElementType ==(static_cast <bool> (ResultElementType == cast<PointerType >(getType()->getScalarType())->getElementType()) ? void (0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1094, __extension__ __PRETTY_FUNCTION__)) |
1094 | cast<PointerType>(getType()->getScalarType())->getElementType())(static_cast <bool> (ResultElementType == cast<PointerType >(getType()->getScalarType())->getElementType()) ? void (0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1094, __extension__ __PRETTY_FUNCTION__)); |
1095 | init(Ptr, IdxList, NameStr); |
1096 | } |
1097 | |
1098 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1098, __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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1098, __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 ); } |
1099 | |
1100 | //===----------------------------------------------------------------------===// |
1101 | // ICmpInst Class |
1102 | //===----------------------------------------------------------------------===// |
1103 | |
1104 | /// This instruction compares its operands according to the predicate given |
1105 | /// to the constructor. It only operates on integers or pointers. The operands |
1106 | /// must be identical types. |
1107 | /// Represent an integer comparison operator. |
1108 | class ICmpInst: public CmpInst { |
1109 | void AssertOK() { |
1110 | assert(isIntPredicate() &&(static_cast <bool> (isIntPredicate() && "Invalid ICmp predicate value" ) ? void (0) : __assert_fail ("isIntPredicate() && \"Invalid ICmp predicate value\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1111, __extension__ __PRETTY_FUNCTION__)) |
1111 | "Invalid ICmp predicate value")(static_cast <bool> (isIntPredicate() && "Invalid ICmp predicate value" ) ? void (0) : __assert_fail ("isIntPredicate() && \"Invalid ICmp predicate value\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1111, __extension__ __PRETTY_FUNCTION__)); |
1112 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1113, __extension__ __PRETTY_FUNCTION__)) |
1113 | "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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1113, __extension__ __PRETTY_FUNCTION__)); |
1114 | // Check that the operands are the right type |
1115 | 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\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1117, __extension__ __PRETTY_FUNCTION__)) |
1116 | 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\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1117, __extension__ __PRETTY_FUNCTION__)) |
1117 | "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\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1117, __extension__ __PRETTY_FUNCTION__)); |
1118 | } |
1119 | |
1120 | protected: |
1121 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1122 | friend class Instruction; |
1123 | |
1124 | /// Clone an identical ICmpInst |
1125 | ICmpInst *cloneImpl() const; |
1126 | |
1127 | public: |
1128 | /// Constructor with insert-before-instruction semantics. |
1129 | ICmpInst( |
1130 | Instruction *InsertBefore, ///< Where to insert |
1131 | Predicate pred, ///< The predicate to use for the comparison |
1132 | Value *LHS, ///< The left-hand-side of the expression |
1133 | Value *RHS, ///< The right-hand-side of the expression |
1134 | const Twine &NameStr = "" ///< Name of the instruction |
1135 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1136 | Instruction::ICmp, pred, LHS, RHS, NameStr, |
1137 | InsertBefore) { |
1138 | #ifndef NDEBUG |
1139 | AssertOK(); |
1140 | #endif |
1141 | } |
1142 | |
1143 | /// Constructor with insert-at-end semantics. |
1144 | ICmpInst( |
1145 | BasicBlock &InsertAtEnd, ///< Block to insert into. |
1146 | Predicate pred, ///< The predicate to use for the comparison |
1147 | Value *LHS, ///< The left-hand-side of the expression |
1148 | Value *RHS, ///< The right-hand-side of the expression |
1149 | const Twine &NameStr = "" ///< Name of the instruction |
1150 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1151 | Instruction::ICmp, pred, LHS, RHS, NameStr, |
1152 | &InsertAtEnd) { |
1153 | #ifndef NDEBUG |
1154 | AssertOK(); |
1155 | #endif |
1156 | } |
1157 | |
1158 | /// Constructor with no-insertion semantics |
1159 | ICmpInst( |
1160 | Predicate pred, ///< The predicate to use for the comparison |
1161 | Value *LHS, ///< The left-hand-side of the expression |
1162 | Value *RHS, ///< The right-hand-side of the expression |
1163 | const Twine &NameStr = "" ///< Name of the instruction |
1164 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1165 | Instruction::ICmp, pred, LHS, RHS, NameStr) { |
1166 | #ifndef NDEBUG |
1167 | AssertOK(); |
1168 | #endif |
1169 | } |
1170 | |
1171 | /// For example, EQ->EQ, SLE->SLE, UGT->SGT, etc. |
1172 | /// @returns the predicate that would be the result if the operand were |
1173 | /// regarded as signed. |
1174 | /// Return the signed version of the predicate |
1175 | Predicate getSignedPredicate() const { |
1176 | return getSignedPredicate(getPredicate()); |
1177 | } |
1178 | |
1179 | /// This is a static version that you can use without an instruction. |
1180 | /// Return the signed version of the predicate. |
1181 | static Predicate getSignedPredicate(Predicate pred); |
1182 | |
1183 | /// For example, EQ->EQ, SLE->ULE, UGT->UGT, etc. |
1184 | /// @returns the predicate that would be the result if the operand were |
1185 | /// regarded as unsigned. |
1186 | /// Return the unsigned version of the predicate |
1187 | Predicate getUnsignedPredicate() const { |
1188 | return getUnsignedPredicate(getPredicate()); |
1189 | } |
1190 | |
1191 | /// This is a static version that you can use without an instruction. |
1192 | /// Return the unsigned version of the predicate. |
1193 | static Predicate getUnsignedPredicate(Predicate pred); |
1194 | |
1195 | /// Return true if this predicate is either EQ or NE. This also |
1196 | /// tests for commutativity. |
1197 | static bool isEquality(Predicate P) { |
1198 | return P == ICMP_EQ || P == ICMP_NE; |
1199 | } |
1200 | |
1201 | /// Return true if this predicate is either EQ or NE. This also |
1202 | /// tests for commutativity. |
1203 | bool isEquality() const { |
1204 | return isEquality(getPredicate()); |
1205 | } |
1206 | |
1207 | /// @returns true if the predicate of this ICmpInst is commutative |
1208 | /// Determine if this relation is commutative. |
1209 | bool isCommutative() const { return isEquality(); } |
1210 | |
1211 | /// Return true if the predicate is relational (not EQ or NE). |
1212 | /// |
1213 | bool isRelational() const { |
1214 | return !isEquality(); |
1215 | } |
1216 | |
1217 | /// Return true if the predicate is relational (not EQ or NE). |
1218 | /// |
1219 | static bool isRelational(Predicate P) { |
1220 | return !isEquality(P); |
1221 | } |
1222 | |
1223 | /// Exchange the two operands to this instruction in such a way that it does |
1224 | /// not modify the semantics of the instruction. The predicate value may be |
1225 | /// changed to retain the same result if the predicate is order dependent |
1226 | /// (e.g. ult). |
1227 | /// Swap operands and adjust predicate. |
1228 | void swapOperands() { |
1229 | setPredicate(getSwappedPredicate()); |
1230 | Op<0>().swap(Op<1>()); |
1231 | } |
1232 | |
1233 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1234 | static bool classof(const Instruction *I) { |
1235 | return I->getOpcode() == Instruction::ICmp; |
1236 | } |
1237 | static bool classof(const Value *V) { |
1238 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1239 | } |
1240 | }; |
1241 | |
1242 | //===----------------------------------------------------------------------===// |
1243 | // FCmpInst Class |
1244 | //===----------------------------------------------------------------------===// |
1245 | |
1246 | /// This instruction compares its operands according to the predicate given |
1247 | /// to the constructor. It only operates on floating point values or packed |
1248 | /// vectors of floating point values. The operands must be identical types. |
1249 | /// Represents a floating point comparison operator. |
1250 | class FCmpInst: public CmpInst { |
1251 | void AssertOK() { |
1252 | assert(isFPPredicate() && "Invalid FCmp predicate value")(static_cast <bool> (isFPPredicate() && "Invalid FCmp predicate value" ) ? void (0) : __assert_fail ("isFPPredicate() && \"Invalid FCmp predicate value\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1252, __extension__ __PRETTY_FUNCTION__)); |
1253 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1254, __extension__ __PRETTY_FUNCTION__)) |
1254 | "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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1254, __extension__ __PRETTY_FUNCTION__)); |
1255 | // Check that the operands are the right type |
1256 | 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\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1257, __extension__ __PRETTY_FUNCTION__)) |
1257 | "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\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1257, __extension__ __PRETTY_FUNCTION__)); |
1258 | } |
1259 | |
1260 | protected: |
1261 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1262 | friend class Instruction; |
1263 | |
1264 | /// Clone an identical FCmpInst |
1265 | FCmpInst *cloneImpl() const; |
1266 | |
1267 | public: |
1268 | /// Constructor with insert-before-instruction semantics. |
1269 | FCmpInst( |
1270 | Instruction *InsertBefore, ///< Where to insert |
1271 | Predicate pred, ///< The predicate to use for the comparison |
1272 | Value *LHS, ///< The left-hand-side of the expression |
1273 | Value *RHS, ///< The right-hand-side of the expression |
1274 | const Twine &NameStr = "" ///< Name of the instruction |
1275 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1276 | Instruction::FCmp, pred, LHS, RHS, NameStr, |
1277 | InsertBefore) { |
1278 | AssertOK(); |
1279 | } |
1280 | |
1281 | /// Constructor with insert-at-end semantics. |
1282 | FCmpInst( |
1283 | BasicBlock &InsertAtEnd, ///< Block to insert into. |
1284 | Predicate pred, ///< The predicate to use for the comparison |
1285 | Value *LHS, ///< The left-hand-side of the expression |
1286 | Value *RHS, ///< The right-hand-side of the expression |
1287 | const Twine &NameStr = "" ///< Name of the instruction |
1288 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1289 | Instruction::FCmp, pred, LHS, RHS, NameStr, |
1290 | &InsertAtEnd) { |
1291 | AssertOK(); |
1292 | } |
1293 | |
1294 | /// Constructor with no-insertion semantics |
1295 | FCmpInst( |
1296 | Predicate pred, ///< The predicate to use for the comparison |
1297 | Value *LHS, ///< The left-hand-side of the expression |
1298 | Value *RHS, ///< The right-hand-side of the expression |
1299 | const Twine &NameStr = "" ///< Name of the instruction |
1300 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
1301 | Instruction::FCmp, pred, LHS, RHS, NameStr) { |
1302 | AssertOK(); |
1303 | } |
1304 | |
1305 | /// @returns true if the predicate of this instruction is EQ or NE. |
1306 | /// Determine if this is an equality predicate. |
1307 | static bool isEquality(Predicate Pred) { |
1308 | return Pred == FCMP_OEQ || Pred == FCMP_ONE || Pred == FCMP_UEQ || |
1309 | Pred == FCMP_UNE; |
1310 | } |
1311 | |
1312 | /// @returns true if the predicate of this instruction is EQ or NE. |
1313 | /// Determine if this is an equality predicate. |
1314 | bool isEquality() const { return isEquality(getPredicate()); } |
1315 | |
1316 | /// @returns true if the predicate of this instruction is commutative. |
1317 | /// Determine if this is a commutative predicate. |
1318 | bool isCommutative() const { |
1319 | return isEquality() || |
1320 | getPredicate() == FCMP_FALSE || |
1321 | getPredicate() == FCMP_TRUE || |
1322 | getPredicate() == FCMP_ORD || |
1323 | getPredicate() == FCMP_UNO; |
1324 | } |
1325 | |
1326 | /// @returns true if the predicate is relational (not EQ or NE). |
1327 | /// Determine if this a relational predicate. |
1328 | bool isRelational() const { return !isEquality(); } |
1329 | |
1330 | /// Exchange the two operands to this instruction in such a way that it does |
1331 | /// not modify the semantics of the instruction. The predicate value may be |
1332 | /// changed to retain the same result if the predicate is order dependent |
1333 | /// (e.g. ult). |
1334 | /// Swap operands and adjust predicate. |
1335 | void swapOperands() { |
1336 | setPredicate(getSwappedPredicate()); |
1337 | Op<0>().swap(Op<1>()); |
1338 | } |
1339 | |
1340 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
1341 | static bool classof(const Instruction *I) { |
1342 | return I->getOpcode() == Instruction::FCmp; |
1343 | } |
1344 | static bool classof(const Value *V) { |
1345 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1346 | } |
1347 | }; |
1348 | |
1349 | //===----------------------------------------------------------------------===// |
1350 | /// This class represents a function call, abstracting a target |
1351 | /// machine's calling convention. This class uses low bit of the SubClassData |
1352 | /// field to indicate whether or not this is a tail call. The rest of the bits |
1353 | /// hold the calling convention of the call. |
1354 | /// |
1355 | class CallInst : public Instruction, |
1356 | public OperandBundleUser<CallInst, User::op_iterator> { |
1357 | friend class OperandBundleUser<CallInst, User::op_iterator>; |
1358 | |
1359 | AttributeList Attrs; ///< parameter attributes for call |
1360 | FunctionType *FTy; |
1361 | |
1362 | CallInst(const CallInst &CI); |
1363 | |
1364 | /// Construct a CallInst given a range of arguments. |
1365 | /// Construct a CallInst from a range of arguments |
1366 | inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1367 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1368 | Instruction *InsertBefore); |
1369 | |
1370 | inline CallInst(Value *Func, ArrayRef<Value *> Args, |
1371 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1372 | Instruction *InsertBefore) |
1373 | : CallInst(cast<FunctionType>( |
1374 | cast<PointerType>(Func->getType())->getElementType()), |
1375 | Func, Args, Bundles, NameStr, InsertBefore) {} |
1376 | |
1377 | inline CallInst(Value *Func, ArrayRef<Value *> Args, const Twine &NameStr, |
1378 | Instruction *InsertBefore) |
1379 | : CallInst(Func, Args, None, NameStr, InsertBefore) {} |
1380 | |
1381 | /// Construct a CallInst given a range of arguments. |
1382 | /// Construct a CallInst from a range of arguments |
1383 | inline CallInst(Value *Func, ArrayRef<Value *> Args, |
1384 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1385 | BasicBlock *InsertAtEnd); |
1386 | |
1387 | explicit CallInst(Value *F, const Twine &NameStr, |
1388 | Instruction *InsertBefore); |
1389 | |
1390 | CallInst(Value *F, const Twine &NameStr, BasicBlock *InsertAtEnd); |
1391 | |
1392 | void init(Value *Func, ArrayRef<Value *> Args, |
1393 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr) { |
1394 | init(cast<FunctionType>( |
1395 | cast<PointerType>(Func->getType())->getElementType()), |
1396 | Func, Args, Bundles, NameStr); |
1397 | } |
1398 | void init(FunctionType *FTy, Value *Func, ArrayRef<Value *> Args, |
1399 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr); |
1400 | void init(Value *Func, const Twine &NameStr); |
1401 | |
1402 | bool hasDescriptor() const { return HasDescriptor; } |
1403 | |
1404 | protected: |
1405 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1406 | friend class Instruction; |
1407 | |
1408 | CallInst *cloneImpl() const; |
1409 | |
1410 | public: |
1411 | static CallInst *Create(Value *Func, ArrayRef<Value *> Args, |
1412 | ArrayRef<OperandBundleDef> Bundles = None, |
1413 | const Twine &NameStr = "", |
1414 | Instruction *InsertBefore = nullptr) { |
1415 | return Create(cast<FunctionType>( |
1416 | cast<PointerType>(Func->getType())->getElementType()), |
1417 | Func, Args, Bundles, NameStr, InsertBefore); |
1418 | } |
1419 | |
1420 | static CallInst *Create(Value *Func, ArrayRef<Value *> Args, |
1421 | const Twine &NameStr, |
1422 | Instruction *InsertBefore = nullptr) { |
1423 | return Create(cast<FunctionType>( |
1424 | cast<PointerType>(Func->getType())->getElementType()), |
1425 | Func, Args, None, NameStr, InsertBefore); |
1426 | } |
1427 | |
1428 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1429 | const Twine &NameStr, |
1430 | Instruction *InsertBefore = nullptr) { |
1431 | return new (unsigned(Args.size() + 1)) |
1432 | CallInst(Ty, Func, Args, None, NameStr, InsertBefore); |
1433 | } |
1434 | |
1435 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1436 | ArrayRef<OperandBundleDef> Bundles = None, |
1437 | const Twine &NameStr = "", |
1438 | Instruction *InsertBefore = nullptr) { |
1439 | const unsigned TotalOps = |
1440 | unsigned(Args.size()) + CountBundleInputs(Bundles) + 1; |
1441 | const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
1442 | |
1443 | return new (TotalOps, DescriptorBytes) |
1444 | CallInst(Ty, Func, Args, Bundles, NameStr, InsertBefore); |
1445 | } |
1446 | |
1447 | static CallInst *Create(Value *Func, ArrayRef<Value *> Args, |
1448 | ArrayRef<OperandBundleDef> Bundles, |
1449 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
1450 | const unsigned TotalOps = |
1451 | unsigned(Args.size()) + CountBundleInputs(Bundles) + 1; |
1452 | const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
1453 | |
1454 | return new (TotalOps, DescriptorBytes) |
1455 | CallInst(Func, Args, Bundles, NameStr, InsertAtEnd); |
1456 | } |
1457 | |
1458 | static CallInst *Create(Value *Func, ArrayRef<Value *> Args, |
1459 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
1460 | return new (unsigned(Args.size() + 1)) |
1461 | CallInst(Func, Args, None, NameStr, InsertAtEnd); |
1462 | } |
1463 | |
1464 | static CallInst *Create(Value *F, const Twine &NameStr = "", |
1465 | Instruction *InsertBefore = nullptr) { |
1466 | return new(1) CallInst(F, NameStr, InsertBefore); |
1467 | } |
1468 | |
1469 | static CallInst *Create(Value *F, const Twine &NameStr, |
1470 | BasicBlock *InsertAtEnd) { |
1471 | return new(1) CallInst(F, NameStr, InsertAtEnd); |
1472 | } |
1473 | |
1474 | /// Create a clone of \p CI with a different set of operand bundles and |
1475 | /// insert it before \p InsertPt. |
1476 | /// |
1477 | /// The returned call instruction is identical \p CI in every way except that |
1478 | /// the operand bundles for the new instruction are set to the operand bundles |
1479 | /// in \p Bundles. |
1480 | static CallInst *Create(CallInst *CI, ArrayRef<OperandBundleDef> Bundles, |
1481 | Instruction *InsertPt = nullptr); |
1482 | |
1483 | /// Generate the IR for a call to malloc: |
1484 | /// 1. Compute the malloc call's argument as the specified type's size, |
1485 | /// possibly multiplied by the array size if the array size is not |
1486 | /// constant 1. |
1487 | /// 2. Call malloc with that argument. |
1488 | /// 3. Bitcast the result of the malloc call to the specified type. |
1489 | static Instruction *CreateMalloc(Instruction *InsertBefore, |
1490 | Type *IntPtrTy, Type *AllocTy, |
1491 | Value *AllocSize, Value *ArraySize = nullptr, |
1492 | Function* MallocF = nullptr, |
1493 | const Twine &Name = ""); |
1494 | static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, |
1495 | Type *IntPtrTy, Type *AllocTy, |
1496 | Value *AllocSize, Value *ArraySize = nullptr, |
1497 | Function* MallocF = nullptr, |
1498 | const Twine &Name = ""); |
1499 | static Instruction *CreateMalloc(Instruction *InsertBefore, |
1500 | Type *IntPtrTy, Type *AllocTy, |
1501 | Value *AllocSize, Value *ArraySize = nullptr, |
1502 | ArrayRef<OperandBundleDef> Bundles = None, |
1503 | Function* MallocF = nullptr, |
1504 | const Twine &Name = ""); |
1505 | static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, |
1506 | Type *IntPtrTy, Type *AllocTy, |
1507 | Value *AllocSize, Value *ArraySize = nullptr, |
1508 | ArrayRef<OperandBundleDef> Bundles = None, |
1509 | Function* MallocF = nullptr, |
1510 | const Twine &Name = ""); |
1511 | /// Generate the IR for a call to the builtin free function. |
1512 | static Instruction *CreateFree(Value *Source, |
1513 | Instruction *InsertBefore); |
1514 | static Instruction *CreateFree(Value *Source, |
1515 | BasicBlock *InsertAtEnd); |
1516 | static Instruction *CreateFree(Value *Source, |
1517 | ArrayRef<OperandBundleDef> Bundles, |
1518 | Instruction *InsertBefore); |
1519 | static Instruction *CreateFree(Value *Source, |
1520 | ArrayRef<OperandBundleDef> Bundles, |
1521 | BasicBlock *InsertAtEnd); |
1522 | |
1523 | FunctionType *getFunctionType() const { return FTy; } |
1524 | |
1525 | void mutateFunctionType(FunctionType *FTy) { |
1526 | mutateType(FTy->getReturnType()); |
1527 | this->FTy = FTy; |
1528 | } |
1529 | |
1530 | // Note that 'musttail' implies 'tail'. |
1531 | enum TailCallKind { TCK_None = 0, TCK_Tail = 1, TCK_MustTail = 2, |
1532 | TCK_NoTail = 3 }; |
1533 | TailCallKind getTailCallKind() const { |
1534 | return TailCallKind(getSubclassDataFromInstruction() & 3); |
1535 | } |
1536 | |
1537 | bool isTailCall() const { |
1538 | unsigned Kind = getSubclassDataFromInstruction() & 3; |
1539 | return Kind == TCK_Tail || Kind == TCK_MustTail; |
1540 | } |
1541 | |
1542 | bool isMustTailCall() const { |
1543 | return (getSubclassDataFromInstruction() & 3) == TCK_MustTail; |
1544 | } |
1545 | |
1546 | bool isNoTailCall() const { |
1547 | return (getSubclassDataFromInstruction() & 3) == TCK_NoTail; |
1548 | } |
1549 | |
1550 | void setTailCall(bool isTC = true) { |
1551 | setInstructionSubclassData((getSubclassDataFromInstruction() & ~3) | |
1552 | unsigned(isTC ? TCK_Tail : TCK_None)); |
1553 | } |
1554 | |
1555 | void setTailCallKind(TailCallKind TCK) { |
1556 | setInstructionSubclassData((getSubclassDataFromInstruction() & ~3) | |
1557 | unsigned(TCK)); |
1558 | } |
1559 | |
1560 | /// Provide fast operand accessors |
1561 | 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; |
1562 | |
1563 | /// Return the number of call arguments. |
1564 | /// |
1565 | unsigned getNumArgOperands() const { |
1566 | return getNumOperands() - getNumTotalBundleOperands() - 1; |
1567 | } |
1568 | |
1569 | /// getArgOperand/setArgOperand - Return/set the i-th call argument. |
1570 | /// |
1571 | Value *getArgOperand(unsigned i) const { |
1572 | assert(i < getNumArgOperands() && "Out of bounds!")(static_cast <bool> (i < getNumArgOperands() && "Out of bounds!") ? void (0) : __assert_fail ("i < getNumArgOperands() && \"Out of bounds!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1572, __extension__ __PRETTY_FUNCTION__)); |
1573 | return getOperand(i); |
1574 | } |
1575 | void setArgOperand(unsigned i, Value *v) { |
1576 | assert(i < getNumArgOperands() && "Out of bounds!")(static_cast <bool> (i < getNumArgOperands() && "Out of bounds!") ? void (0) : __assert_fail ("i < getNumArgOperands() && \"Out of bounds!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1576, __extension__ __PRETTY_FUNCTION__)); |
1577 | setOperand(i, v); |
1578 | } |
1579 | |
1580 | /// Return the iterator pointing to the beginning of the argument list. |
1581 | op_iterator arg_begin() { return op_begin(); } |
1582 | |
1583 | /// Return the iterator pointing to the end of the argument list. |
1584 | op_iterator arg_end() { |
1585 | // [ call args ], [ operand bundles ], callee |
1586 | return op_end() - getNumTotalBundleOperands() - 1; |
1587 | } |
1588 | |
1589 | /// Iteration adapter for range-for loops. |
1590 | iterator_range<op_iterator> arg_operands() { |
1591 | return make_range(arg_begin(), arg_end()); |
1592 | } |
1593 | |
1594 | /// Return the iterator pointing to the beginning of the argument list. |
1595 | const_op_iterator arg_begin() const { return op_begin(); } |
1596 | |
1597 | /// Return the iterator pointing to the end of the argument list. |
1598 | const_op_iterator arg_end() const { |
1599 | // [ call args ], [ operand bundles ], callee |
1600 | return op_end() - getNumTotalBundleOperands() - 1; |
1601 | } |
1602 | |
1603 | /// Iteration adapter for range-for loops. |
1604 | iterator_range<const_op_iterator> arg_operands() const { |
1605 | return make_range(arg_begin(), arg_end()); |
1606 | } |
1607 | |
1608 | /// Wrappers for getting the \c Use of a call argument. |
1609 | const Use &getArgOperandUse(unsigned i) const { |
1610 | assert(i < getNumArgOperands() && "Out of bounds!")(static_cast <bool> (i < getNumArgOperands() && "Out of bounds!") ? void (0) : __assert_fail ("i < getNumArgOperands() && \"Out of bounds!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1610, __extension__ __PRETTY_FUNCTION__)); |
1611 | return getOperandUse(i); |
1612 | } |
1613 | Use &getArgOperandUse(unsigned i) { |
1614 | assert(i < getNumArgOperands() && "Out of bounds!")(static_cast <bool> (i < getNumArgOperands() && "Out of bounds!") ? void (0) : __assert_fail ("i < getNumArgOperands() && \"Out of bounds!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1614, __extension__ __PRETTY_FUNCTION__)); |
1615 | return getOperandUse(i); |
1616 | } |
1617 | |
1618 | /// If one of the arguments has the 'returned' attribute, return its |
1619 | /// operand value. Otherwise, return nullptr. |
1620 | Value *getReturnedArgOperand() const; |
1621 | |
1622 | /// getCallingConv/setCallingConv - Get or set the calling convention of this |
1623 | /// function call. |
1624 | CallingConv::ID getCallingConv() const { |
1625 | return static_cast<CallingConv::ID>(getSubclassDataFromInstruction() >> 2); |
1626 | } |
1627 | void setCallingConv(CallingConv::ID CC) { |
1628 | auto ID = static_cast<unsigned>(CC); |
1629 | assert(!(ID & ~CallingConv::MaxID) && "Unsupported calling convention")(static_cast <bool> (!(ID & ~CallingConv::MaxID) && "Unsupported calling convention") ? void (0) : __assert_fail ("!(ID & ~CallingConv::MaxID) && \"Unsupported calling convention\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1629, __extension__ __PRETTY_FUNCTION__)); |
1630 | setInstructionSubclassData((getSubclassDataFromInstruction() & 3) | |
1631 | (ID << 2)); |
1632 | } |
1633 | |
1634 | /// Return the parameter attributes for this call. |
1635 | /// |
1636 | AttributeList getAttributes() const { return Attrs; } |
1637 | |
1638 | /// Set the parameter attributes for this call. |
1639 | /// |
1640 | void setAttributes(AttributeList A) { Attrs = A; } |
1641 | |
1642 | /// adds the attribute to the list of attributes. |
1643 | void addAttribute(unsigned i, Attribute::AttrKind Kind); |
1644 | |
1645 | /// adds the attribute to the list of attributes. |
1646 | void addAttribute(unsigned i, Attribute Attr); |
1647 | |
1648 | /// Adds the attribute to the indicated argument |
1649 | void addParamAttr(unsigned ArgNo, Attribute::AttrKind Kind); |
1650 | |
1651 | /// Adds the attribute to the indicated argument |
1652 | void addParamAttr(unsigned ArgNo, Attribute Attr); |
1653 | |
1654 | /// removes the attribute from the list of attributes. |
1655 | void removeAttribute(unsigned i, Attribute::AttrKind Kind); |
1656 | |
1657 | /// removes the attribute from the list of attributes. |
1658 | void removeAttribute(unsigned i, StringRef Kind); |
1659 | |
1660 | /// Removes the attribute from the given argument |
1661 | void removeParamAttr(unsigned ArgNo, Attribute::AttrKind Kind); |
1662 | |
1663 | /// Removes the attribute from the given argument |
1664 | void removeParamAttr(unsigned ArgNo, StringRef Kind); |
1665 | |
1666 | /// adds the dereferenceable attribute to the list of attributes. |
1667 | void addDereferenceableAttr(unsigned i, uint64_t Bytes); |
1668 | |
1669 | /// adds the dereferenceable_or_null attribute to the list of |
1670 | /// attributes. |
1671 | void addDereferenceableOrNullAttr(unsigned i, uint64_t Bytes); |
1672 | |
1673 | /// Determine whether this call has the given attribute. |
1674 | bool hasFnAttr(Attribute::AttrKind Kind) const { |
1675 | assert(Kind != Attribute::NoBuiltin &&(static_cast <bool> (Kind != Attribute::NoBuiltin && "Use CallInst::isNoBuiltin() to check for Attribute::NoBuiltin" ) ? void (0) : __assert_fail ("Kind != Attribute::NoBuiltin && \"Use CallInst::isNoBuiltin() to check for Attribute::NoBuiltin\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1676, __extension__ __PRETTY_FUNCTION__)) |
1676 | "Use CallInst::isNoBuiltin() to check for Attribute::NoBuiltin")(static_cast <bool> (Kind != Attribute::NoBuiltin && "Use CallInst::isNoBuiltin() to check for Attribute::NoBuiltin" ) ? void (0) : __assert_fail ("Kind != Attribute::NoBuiltin && \"Use CallInst::isNoBuiltin() to check for Attribute::NoBuiltin\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1676, __extension__ __PRETTY_FUNCTION__)); |
1677 | return hasFnAttrImpl(Kind); |
1678 | } |
1679 | |
1680 | /// Determine whether this call has the given attribute. |
1681 | bool hasFnAttr(StringRef Kind) const { |
1682 | return hasFnAttrImpl(Kind); |
1683 | } |
1684 | |
1685 | /// Determine whether the return value has the given attribute. |
1686 | bool hasRetAttr(Attribute::AttrKind Kind) const; |
1687 | |
1688 | /// Determine whether the argument or parameter has the given attribute. |
1689 | bool paramHasAttr(unsigned ArgNo, Attribute::AttrKind Kind) const; |
1690 | |
1691 | /// Get the attribute of a given kind at a position. |
1692 | Attribute getAttribute(unsigned i, Attribute::AttrKind Kind) const { |
1693 | return getAttributes().getAttribute(i, Kind); |
1694 | } |
1695 | |
1696 | /// Get the attribute of a given kind at a position. |
1697 | Attribute getAttribute(unsigned i, StringRef Kind) const { |
1698 | return getAttributes().getAttribute(i, Kind); |
1699 | } |
1700 | |
1701 | /// Get the attribute of a given kind from a given arg |
1702 | Attribute getParamAttr(unsigned ArgNo, Attribute::AttrKind Kind) const { |
1703 | assert(ArgNo < getNumArgOperands() && "Out of bounds")(static_cast <bool> (ArgNo < getNumArgOperands() && "Out of bounds") ? void (0) : __assert_fail ("ArgNo < getNumArgOperands() && \"Out of bounds\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1703, __extension__ __PRETTY_FUNCTION__)); |
1704 | return getAttributes().getParamAttr(ArgNo, Kind); |
1705 | } |
1706 | |
1707 | /// Get the attribute of a given kind from a given arg |
1708 | Attribute getParamAttr(unsigned ArgNo, StringRef Kind) const { |
1709 | assert(ArgNo < getNumArgOperands() && "Out of bounds")(static_cast <bool> (ArgNo < getNumArgOperands() && "Out of bounds") ? void (0) : __assert_fail ("ArgNo < getNumArgOperands() && \"Out of bounds\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1709, __extension__ __PRETTY_FUNCTION__)); |
1710 | return getAttributes().getParamAttr(ArgNo, Kind); |
1711 | } |
1712 | |
1713 | /// Return true if the data operand at index \p i has the attribute \p |
1714 | /// A. |
1715 | /// |
1716 | /// Data operands include call arguments and values used in operand bundles, |
1717 | /// but does not include the callee operand. This routine dispatches to the |
1718 | /// underlying AttributeList or the OperandBundleUser as appropriate. |
1719 | /// |
1720 | /// The index \p i is interpreted as |
1721 | /// |
1722 | /// \p i == Attribute::ReturnIndex -> the return value |
1723 | /// \p i in [1, arg_size + 1) -> argument number (\p i - 1) |
1724 | /// \p i in [arg_size + 1, data_operand_size + 1) -> bundle operand at index |
1725 | /// (\p i - 1) in the operand list. |
1726 | bool dataOperandHasImpliedAttr(unsigned i, Attribute::AttrKind Kind) const; |
1727 | |
1728 | /// Extract the alignment of the return value. |
1729 | unsigned getRetAlignment() const { return Attrs.getRetAlignment(); } |
1730 | |
1731 | /// Extract the alignment for a call or parameter (0=unknown). |
1732 | unsigned getParamAlignment(unsigned ArgNo) const { |
1733 | return Attrs.getParamAlignment(ArgNo); |
1734 | } |
1735 | |
1736 | /// Extract the number of dereferenceable bytes for a call or |
1737 | /// parameter (0=unknown). |
1738 | uint64_t getDereferenceableBytes(unsigned i) const { |
1739 | return Attrs.getDereferenceableBytes(i); |
1740 | } |
1741 | |
1742 | /// Extract the number of dereferenceable_or_null bytes for a call or |
1743 | /// parameter (0=unknown). |
1744 | uint64_t getDereferenceableOrNullBytes(unsigned i) const { |
1745 | return Attrs.getDereferenceableOrNullBytes(i); |
1746 | } |
1747 | |
1748 | /// @brief Determine if the return value is marked with NoAlias attribute. |
1749 | bool returnDoesNotAlias() const { |
1750 | return Attrs.hasAttribute(AttributeList::ReturnIndex, Attribute::NoAlias); |
1751 | } |
1752 | |
1753 | /// Return true if the call should not be treated as a call to a |
1754 | /// builtin. |
1755 | bool isNoBuiltin() const { |
1756 | return hasFnAttrImpl(Attribute::NoBuiltin) && |
1757 | !hasFnAttrImpl(Attribute::Builtin); |
1758 | } |
1759 | |
1760 | /// Determine if the call requires strict floating point semantics. |
1761 | bool isStrictFP() const { return hasFnAttr(Attribute::StrictFP); } |
1762 | |
1763 | /// Return true if the call should not be inlined. |
1764 | bool isNoInline() const { return hasFnAttr(Attribute::NoInline); } |
1765 | void setIsNoInline() { |
1766 | addAttribute(AttributeList::FunctionIndex, Attribute::NoInline); |
1767 | } |
1768 | |
1769 | /// Return true if the call can return twice |
1770 | bool canReturnTwice() const { |
1771 | return hasFnAttr(Attribute::ReturnsTwice); |
1772 | } |
1773 | void setCanReturnTwice() { |
1774 | addAttribute(AttributeList::FunctionIndex, Attribute::ReturnsTwice); |
1775 | } |
1776 | |
1777 | /// Determine if the call does not access memory. |
1778 | bool doesNotAccessMemory() const { |
1779 | return hasFnAttr(Attribute::ReadNone); |
1780 | } |
1781 | void setDoesNotAccessMemory() { |
1782 | addAttribute(AttributeList::FunctionIndex, Attribute::ReadNone); |
1783 | } |
1784 | |
1785 | /// Determine if the call does not access or only reads memory. |
1786 | bool onlyReadsMemory() const { |
1787 | return doesNotAccessMemory() || hasFnAttr(Attribute::ReadOnly); |
1788 | } |
1789 | void setOnlyReadsMemory() { |
1790 | addAttribute(AttributeList::FunctionIndex, Attribute::ReadOnly); |
1791 | } |
1792 | |
1793 | /// Determine if the call does not access or only writes memory. |
1794 | bool doesNotReadMemory() const { |
1795 | return doesNotAccessMemory() || hasFnAttr(Attribute::WriteOnly); |
1796 | } |
1797 | void setDoesNotReadMemory() { |
1798 | addAttribute(AttributeList::FunctionIndex, Attribute::WriteOnly); |
1799 | } |
1800 | |
1801 | /// @brief Determine if the call can access memmory only using pointers based |
1802 | /// on its arguments. |
1803 | bool onlyAccessesArgMemory() const { |
1804 | return hasFnAttr(Attribute::ArgMemOnly); |
1805 | } |
1806 | void setOnlyAccessesArgMemory() { |
1807 | addAttribute(AttributeList::FunctionIndex, Attribute::ArgMemOnly); |
1808 | } |
1809 | |
1810 | /// @brief Determine if the function may only access memory that is |
1811 | /// inaccessible from the IR. |
1812 | bool onlyAccessesInaccessibleMemory() const { |
1813 | return hasFnAttr(Attribute::InaccessibleMemOnly); |
1814 | } |
1815 | void setOnlyAccessesInaccessibleMemory() { |
1816 | addAttribute(AttributeList::FunctionIndex, Attribute::InaccessibleMemOnly); |
1817 | } |
1818 | |
1819 | /// @brief Determine if the function may only access memory that is |
1820 | /// either inaccessible from the IR or pointed to by its arguments. |
1821 | bool onlyAccessesInaccessibleMemOrArgMem() const { |
1822 | return hasFnAttr(Attribute::InaccessibleMemOrArgMemOnly); |
1823 | } |
1824 | void setOnlyAccessesInaccessibleMemOrArgMem() { |
1825 | addAttribute(AttributeList::FunctionIndex, Attribute::InaccessibleMemOrArgMemOnly); |
1826 | } |
1827 | |
1828 | /// Determine if the call cannot return. |
1829 | bool doesNotReturn() const { return hasFnAttr(Attribute::NoReturn); } |
1830 | void setDoesNotReturn() { |
1831 | addAttribute(AttributeList::FunctionIndex, Attribute::NoReturn); |
1832 | } |
1833 | |
1834 | /// Determine if the call cannot unwind. |
1835 | bool doesNotThrow() const { return hasFnAttr(Attribute::NoUnwind); } |
1836 | void setDoesNotThrow() { |
1837 | addAttribute(AttributeList::FunctionIndex, Attribute::NoUnwind); |
1838 | } |
1839 | |
1840 | /// Determine if the call cannot be duplicated. |
1841 | bool cannotDuplicate() const {return hasFnAttr(Attribute::NoDuplicate); } |
1842 | void setCannotDuplicate() { |
1843 | addAttribute(AttributeList::FunctionIndex, Attribute::NoDuplicate); |
1844 | } |
1845 | |
1846 | /// Determine if the call is convergent |
1847 | bool isConvergent() const { return hasFnAttr(Attribute::Convergent); } |
1848 | void setConvergent() { |
1849 | addAttribute(AttributeList::FunctionIndex, Attribute::Convergent); |
1850 | } |
1851 | void setNotConvergent() { |
1852 | removeAttribute(AttributeList::FunctionIndex, Attribute::Convergent); |
1853 | } |
1854 | |
1855 | /// Determine if the call returns a structure through first |
1856 | /// pointer argument. |
1857 | bool hasStructRetAttr() const { |
1858 | if (getNumArgOperands() == 0) |
1859 | return false; |
1860 | |
1861 | // Be friendly and also check the callee. |
1862 | return paramHasAttr(0, Attribute::StructRet); |
1863 | } |
1864 | |
1865 | /// Determine if any call argument is an aggregate passed by value. |
1866 | bool hasByValArgument() const { |
1867 | return Attrs.hasAttrSomewhere(Attribute::ByVal); |
1868 | } |
1869 | |
1870 | /// Return the function called, or null if this is an |
1871 | /// indirect function invocation. |
1872 | /// |
1873 | Function *getCalledFunction() const { |
1874 | return dyn_cast<Function>(Op<-1>()); |
1875 | } |
1876 | |
1877 | /// Get a pointer to the function that is invoked by this |
1878 | /// instruction. |
1879 | const Value *getCalledValue() const { return Op<-1>(); } |
1880 | Value *getCalledValue() { return Op<-1>(); } |
1881 | |
1882 | /// Set the function called. |
1883 | void setCalledFunction(Value* Fn) { |
1884 | setCalledFunction( |
1885 | cast<FunctionType>(cast<PointerType>(Fn->getType())->getElementType()), |
1886 | Fn); |
1887 | } |
1888 | void setCalledFunction(FunctionType *FTy, Value *Fn) { |
1889 | this->FTy = FTy; |
1890 | assert(FTy == cast<FunctionType>((static_cast <bool> (FTy == cast<FunctionType>( cast <PointerType>(Fn->getType())->getElementType())) ? void (0) : __assert_fail ("FTy == cast<FunctionType>( cast<PointerType>(Fn->getType())->getElementType())" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1891, __extension__ __PRETTY_FUNCTION__)) |
1891 | cast<PointerType>(Fn->getType())->getElementType()))(static_cast <bool> (FTy == cast<FunctionType>( cast <PointerType>(Fn->getType())->getElementType())) ? void (0) : __assert_fail ("FTy == cast<FunctionType>( cast<PointerType>(Fn->getType())->getElementType())" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1891, __extension__ __PRETTY_FUNCTION__)); |
1892 | Op<-1>() = Fn; |
1893 | } |
1894 | |
1895 | /// Check if this call is an inline asm statement. |
1896 | bool isInlineAsm() const { |
1897 | return isa<InlineAsm>(Op<-1>()); |
1898 | } |
1899 | |
1900 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
1901 | static bool classof(const Instruction *I) { |
1902 | return I->getOpcode() == Instruction::Call; |
1903 | } |
1904 | static bool classof(const Value *V) { |
1905 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1906 | } |
1907 | |
1908 | private: |
1909 | template <typename AttrKind> bool hasFnAttrImpl(AttrKind Kind) const { |
1910 | if (Attrs.hasAttribute(AttributeList::FunctionIndex, Kind)) |
1911 | return true; |
1912 | |
1913 | // Operand bundles override attributes on the called function, but don't |
1914 | // override attributes directly present on the call instruction. |
1915 | if (isFnAttrDisallowedByOpBundle(Kind)) |
1916 | return false; |
1917 | |
1918 | if (const Function *F = getCalledFunction()) |
1919 | return F->getAttributes().hasAttribute(AttributeList::FunctionIndex, |
1920 | Kind); |
1921 | return false; |
1922 | } |
1923 | |
1924 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
1925 | // method so that subclasses cannot accidentally use it. |
1926 | void setInstructionSubclassData(unsigned short D) { |
1927 | Instruction::setInstructionSubclassData(D); |
1928 | } |
1929 | }; |
1930 | |
1931 | template <> |
1932 | struct OperandTraits<CallInst> : public VariadicOperandTraits<CallInst, 1> { |
1933 | }; |
1934 | |
1935 | CallInst::CallInst(Value *Func, ArrayRef<Value *> Args, |
1936 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1937 | BasicBlock *InsertAtEnd) |
1938 | : Instruction( |
1939 | cast<FunctionType>(cast<PointerType>(Func->getType()) |
1940 | ->getElementType())->getReturnType(), |
1941 | Instruction::Call, OperandTraits<CallInst>::op_end(this) - |
1942 | (Args.size() + CountBundleInputs(Bundles) + 1), |
1943 | unsigned(Args.size() + CountBundleInputs(Bundles) + 1), InsertAtEnd) { |
1944 | init(Func, Args, Bundles, NameStr); |
1945 | } |
1946 | |
1947 | CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
1948 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
1949 | Instruction *InsertBefore) |
1950 | : Instruction(Ty->getReturnType(), Instruction::Call, |
1951 | OperandTraits<CallInst>::op_end(this) - |
1952 | (Args.size() + CountBundleInputs(Bundles) + 1), |
1953 | unsigned(Args.size() + CountBundleInputs(Bundles) + 1), |
1954 | InsertBefore) { |
1955 | init(Ty, Func, Args, Bundles, NameStr); |
1956 | } |
1957 | |
1958 | // Note: if you get compile errors about private methods then |
1959 | // please update your code to use the high-level operand |
1960 | // interfaces. See line 943 above. |
1961 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(CallInst, Value)CallInst::op_iterator CallInst::op_begin() { return OperandTraits <CallInst>::op_begin(this); } CallInst::const_op_iterator CallInst::op_begin() const { return OperandTraits<CallInst >::op_begin(const_cast<CallInst*>(this)); } CallInst ::op_iterator CallInst::op_end() { return OperandTraits<CallInst >::op_end(this); } CallInst::const_op_iterator CallInst::op_end () const { return OperandTraits<CallInst>::op_end(const_cast <CallInst*>(this)); } Value *CallInst::getOperand(unsigned i_nocapture) const { (static_cast <bool> (i_nocapture < OperandTraits<CallInst>::operands(this) && "getOperand() out of range!" ) ? void (0) : __assert_fail ("i_nocapture < OperandTraits<CallInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1961, __extension__ __PRETTY_FUNCTION__)); return cast_or_null <Value>( OperandTraits<CallInst>::op_begin(const_cast <CallInst*>(this))[i_nocapture].get()); } void CallInst ::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (static_cast <bool> (i_nocapture < OperandTraits<CallInst> ::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<CallInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1961, __extension__ __PRETTY_FUNCTION__)); OperandTraits< CallInst>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned CallInst::getNumOperands() const { return OperandTraits<CallInst >::operands(this); } template <int Idx_nocapture> Use &CallInst::Op() { return this->OpFrom<Idx_nocapture >(this); } template <int Idx_nocapture> const Use & CallInst::Op() const { return this->OpFrom<Idx_nocapture >(this); } |
1962 | |
1963 | //===----------------------------------------------------------------------===// |
1964 | // SelectInst Class |
1965 | //===----------------------------------------------------------------------===// |
1966 | |
1967 | /// This class represents the LLVM 'select' instruction. |
1968 | /// |
1969 | class SelectInst : public Instruction { |
1970 | SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr, |
1971 | Instruction *InsertBefore) |
1972 | : Instruction(S1->getType(), Instruction::Select, |
1973 | &Op<0>(), 3, InsertBefore) { |
1974 | init(C, S1, S2); |
1975 | setName(NameStr); |
1976 | } |
1977 | |
1978 | SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr, |
1979 | BasicBlock *InsertAtEnd) |
1980 | : Instruction(S1->getType(), Instruction::Select, |
1981 | &Op<0>(), 3, InsertAtEnd) { |
1982 | init(C, S1, S2); |
1983 | setName(NameStr); |
1984 | } |
1985 | |
1986 | void init(Value *C, Value *S1, Value *S2) { |
1987 | 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\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 1987, __extension__ __PRETTY_FUNCTION__)); |
1988 | Op<0>() = C; |
1989 | Op<1>() = S1; |
1990 | Op<2>() = S2; |
1991 | } |
1992 | |
1993 | protected: |
1994 | // Note: Instruction needs to be a friend here to call cloneImpl. |
1995 | friend class Instruction; |
1996 | |
1997 | SelectInst *cloneImpl() const; |
1998 | |
1999 | public: |
2000 | static SelectInst *Create(Value *C, Value *S1, Value *S2, |
2001 | const Twine &NameStr = "", |
2002 | Instruction *InsertBefore = nullptr, |
2003 | Instruction *MDFrom = nullptr) { |
2004 | SelectInst *Sel = new(3) SelectInst(C, S1, S2, NameStr, InsertBefore); |
2005 | if (MDFrom) |
2006 | Sel->copyMetadata(*MDFrom); |
2007 | return Sel; |
2008 | } |
2009 | |
2010 | static SelectInst *Create(Value *C, Value *S1, Value *S2, |
2011 | const Twine &NameStr, |
2012 | BasicBlock *InsertAtEnd) { |
2013 | return new(3) SelectInst(C, S1, S2, NameStr, InsertAtEnd); |
2014 | } |
2015 | |
2016 | const Value *getCondition() const { return Op<0>(); } |
2017 | const Value *getTrueValue() const { return Op<1>(); } |
2018 | const Value *getFalseValue() const { return Op<2>(); } |
2019 | Value *getCondition() { return Op<0>(); } |
2020 | Value *getTrueValue() { return Op<1>(); } |
2021 | Value *getFalseValue() { return Op<2>(); } |
2022 | |
2023 | void setCondition(Value *V) { Op<0>() = V; } |
2024 | void setTrueValue(Value *V) { Op<1>() = V; } |
2025 | void setFalseValue(Value *V) { Op<2>() = V; } |
2026 | |
2027 | /// Return a string if the specified operands are invalid |
2028 | /// for a select operation, otherwise return null. |
2029 | static const char *areInvalidOperands(Value *Cond, Value *True, Value *False); |
2030 | |
2031 | /// Transparently provide more efficient getOperand methods. |
2032 | 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; |
2033 | |
2034 | OtherOps getOpcode() const { |
2035 | return static_cast<OtherOps>(Instruction::getOpcode()); |
2036 | } |
2037 | |
2038 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2039 | static bool classof(const Instruction *I) { |
2040 | return I->getOpcode() == Instruction::Select; |
2041 | } |
2042 | static bool classof(const Value *V) { |
2043 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2044 | } |
2045 | }; |
2046 | |
2047 | template <> |
2048 | struct OperandTraits<SelectInst> : public FixedNumOperandTraits<SelectInst, 3> { |
2049 | }; |
2050 | |
2051 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2051, __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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2051, __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); } |
2052 | |
2053 | //===----------------------------------------------------------------------===// |
2054 | // VAArgInst Class |
2055 | //===----------------------------------------------------------------------===// |
2056 | |
2057 | /// This class represents the va_arg llvm instruction, which returns |
2058 | /// an argument of the specified type given a va_list and increments that list |
2059 | /// |
2060 | class VAArgInst : public UnaryInstruction { |
2061 | protected: |
2062 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2063 | friend class Instruction; |
2064 | |
2065 | VAArgInst *cloneImpl() const; |
2066 | |
2067 | public: |
2068 | VAArgInst(Value *List, Type *Ty, const Twine &NameStr = "", |
2069 | Instruction *InsertBefore = nullptr) |
2070 | : UnaryInstruction(Ty, VAArg, List, InsertBefore) { |
2071 | setName(NameStr); |
2072 | } |
2073 | |
2074 | VAArgInst(Value *List, Type *Ty, const Twine &NameStr, |
2075 | BasicBlock *InsertAtEnd) |
2076 | : UnaryInstruction(Ty, VAArg, List, InsertAtEnd) { |
2077 | setName(NameStr); |
2078 | } |
2079 | |
2080 | Value *getPointerOperand() { return getOperand(0); } |
2081 | const Value *getPointerOperand() const { return getOperand(0); } |
2082 | static unsigned getPointerOperandIndex() { return 0U; } |
2083 | |
2084 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2085 | static bool classof(const Instruction *I) { |
2086 | return I->getOpcode() == VAArg; |
2087 | } |
2088 | static bool classof(const Value *V) { |
2089 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2090 | } |
2091 | }; |
2092 | |
2093 | //===----------------------------------------------------------------------===// |
2094 | // ExtractElementInst Class |
2095 | //===----------------------------------------------------------------------===// |
2096 | |
2097 | /// This instruction extracts a single (scalar) |
2098 | /// element from a VectorType value |
2099 | /// |
2100 | class ExtractElementInst : public Instruction { |
2101 | ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr = "", |
2102 | Instruction *InsertBefore = nullptr); |
2103 | ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr, |
2104 | BasicBlock *InsertAtEnd); |
2105 | |
2106 | protected: |
2107 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2108 | friend class Instruction; |
2109 | |
2110 | ExtractElementInst *cloneImpl() const; |
2111 | |
2112 | public: |
2113 | static ExtractElementInst *Create(Value *Vec, Value *Idx, |
2114 | const Twine &NameStr = "", |
2115 | Instruction *InsertBefore = nullptr) { |
2116 | return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertBefore); |
2117 | } |
2118 | |
2119 | static ExtractElementInst *Create(Value *Vec, Value *Idx, |
2120 | const Twine &NameStr, |
2121 | BasicBlock *InsertAtEnd) { |
2122 | return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertAtEnd); |
2123 | } |
2124 | |
2125 | /// Return true if an extractelement instruction can be |
2126 | /// formed with the specified operands. |
2127 | static bool isValidOperands(const Value *Vec, const Value *Idx); |
2128 | |
2129 | Value *getVectorOperand() { return Op<0>(); } |
2130 | Value *getIndexOperand() { return Op<1>(); } |
2131 | const Value *getVectorOperand() const { return Op<0>(); } |
2132 | const Value *getIndexOperand() const { return Op<1>(); } |
2133 | |
2134 | VectorType *getVectorOperandType() const { |
2135 | return cast<VectorType>(getVectorOperand()->getType()); |
2136 | } |
2137 | |
2138 | /// Transparently provide more efficient getOperand methods. |
2139 | 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; |
2140 | |
2141 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2142 | static bool classof(const Instruction *I) { |
2143 | return I->getOpcode() == Instruction::ExtractElement; |
2144 | } |
2145 | static bool classof(const Value *V) { |
2146 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2147 | } |
2148 | }; |
2149 | |
2150 | template <> |
2151 | struct OperandTraits<ExtractElementInst> : |
2152 | public FixedNumOperandTraits<ExtractElementInst, 2> { |
2153 | }; |
2154 | |
2155 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2155, __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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2155, __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 ); } |
2156 | |
2157 | //===----------------------------------------------------------------------===// |
2158 | // InsertElementInst Class |
2159 | //===----------------------------------------------------------------------===// |
2160 | |
2161 | /// This instruction inserts a single (scalar) |
2162 | /// element into a VectorType value |
2163 | /// |
2164 | class InsertElementInst : public Instruction { |
2165 | InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, |
2166 | const Twine &NameStr = "", |
2167 | Instruction *InsertBefore = nullptr); |
2168 | InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, const Twine &NameStr, |
2169 | BasicBlock *InsertAtEnd); |
2170 | |
2171 | protected: |
2172 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2173 | friend class Instruction; |
2174 | |
2175 | InsertElementInst *cloneImpl() const; |
2176 | |
2177 | public: |
2178 | static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx, |
2179 | const Twine &NameStr = "", |
2180 | Instruction *InsertBefore = nullptr) { |
2181 | return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertBefore); |
2182 | } |
2183 | |
2184 | static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx, |
2185 | const Twine &NameStr, |
2186 | BasicBlock *InsertAtEnd) { |
2187 | return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertAtEnd); |
2188 | } |
2189 | |
2190 | /// Return true if an insertelement instruction can be |
2191 | /// formed with the specified operands. |
2192 | static bool isValidOperands(const Value *Vec, const Value *NewElt, |
2193 | const Value *Idx); |
2194 | |
2195 | /// Overload to return most specific vector type. |
2196 | /// |
2197 | VectorType *getType() const { |
2198 | return cast<VectorType>(Instruction::getType()); |
2199 | } |
2200 | |
2201 | /// Transparently provide more efficient getOperand methods. |
2202 | 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; |
2203 | |
2204 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2205 | static bool classof(const Instruction *I) { |
2206 | return I->getOpcode() == Instruction::InsertElement; |
2207 | } |
2208 | static bool classof(const Value *V) { |
2209 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2210 | } |
2211 | }; |
2212 | |
2213 | template <> |
2214 | struct OperandTraits<InsertElementInst> : |
2215 | public FixedNumOperandTraits<InsertElementInst, 3> { |
2216 | }; |
2217 | |
2218 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2218, __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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2218, __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 ); } |
2219 | |
2220 | //===----------------------------------------------------------------------===// |
2221 | // ShuffleVectorInst Class |
2222 | //===----------------------------------------------------------------------===// |
2223 | |
2224 | /// This instruction constructs a fixed permutation of two |
2225 | /// input vectors. |
2226 | /// |
2227 | class ShuffleVectorInst : public Instruction { |
2228 | protected: |
2229 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2230 | friend class Instruction; |
2231 | |
2232 | ShuffleVectorInst *cloneImpl() const; |
2233 | |
2234 | public: |
2235 | ShuffleVectorInst(Value *V1, Value *V2, Value *Mask, |
2236 | const Twine &NameStr = "", |
2237 | Instruction *InsertBefor = nullptr); |
2238 | ShuffleVectorInst(Value *V1, Value *V2, Value *Mask, |
2239 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2240 | |
2241 | // allocate space for exactly three operands |
2242 | void *operator new(size_t s) { |
2243 | return User::operator new(s, 3); |
2244 | } |
2245 | |
2246 | /// Return true if a shufflevector instruction can be |
2247 | /// formed with the specified operands. |
2248 | static bool isValidOperands(const Value *V1, const Value *V2, |
2249 | const Value *Mask); |
2250 | |
2251 | /// Overload to return most specific vector type. |
2252 | /// |
2253 | VectorType *getType() const { |
2254 | return cast<VectorType>(Instruction::getType()); |
2255 | } |
2256 | |
2257 | /// Transparently provide more efficient getOperand methods. |
2258 | 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; |
2259 | |
2260 | Constant *getMask() const { |
2261 | return cast<Constant>(getOperand(2)); |
2262 | } |
2263 | |
2264 | /// Return the shuffle mask value for the specified element of the mask. |
2265 | /// Return -1 if the element is undef. |
2266 | static int getMaskValue(Constant *Mask, unsigned Elt); |
2267 | |
2268 | /// Return the shuffle mask value of this instruction for the given element |
2269 | /// index. Return -1 if the element is undef. |
2270 | int getMaskValue(unsigned Elt) const { |
2271 | return getMaskValue(getMask(), Elt); |
2272 | } |
2273 | |
2274 | /// Convert the input shuffle mask operand to a vector of integers. Undefined |
2275 | /// elements of the mask are returned as -1. |
2276 | static void getShuffleMask(Constant *Mask, SmallVectorImpl<int> &Result); |
2277 | |
2278 | /// Return the mask for this instruction as a vector of integers. Undefined |
2279 | /// elements of the mask are returned as -1. |
2280 | void getShuffleMask(SmallVectorImpl<int> &Result) const { |
2281 | return getShuffleMask(getMask(), Result); |
2282 | } |
2283 | |
2284 | SmallVector<int, 16> getShuffleMask() const { |
2285 | SmallVector<int, 16> Mask; |
2286 | getShuffleMask(Mask); |
2287 | return Mask; |
2288 | } |
2289 | |
2290 | /// Change values in a shuffle permute mask assuming the two vector operands |
2291 | /// of length InVecNumElts have swapped position. |
2292 | static void commuteShuffleMask(MutableArrayRef<int> Mask, |
2293 | unsigned InVecNumElts) { |
2294 | for (int &Idx : Mask) { |
2295 | if (Idx == -1) |
2296 | continue; |
2297 | Idx = Idx < (int)InVecNumElts ? Idx + InVecNumElts : Idx - InVecNumElts; |
2298 | 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\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2299, __extension__ __PRETTY_FUNCTION__)) |
2299 | "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\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2299, __extension__ __PRETTY_FUNCTION__)); |
2300 | } |
2301 | } |
2302 | |
2303 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2304 | static bool classof(const Instruction *I) { |
2305 | return I->getOpcode() == Instruction::ShuffleVector; |
2306 | } |
2307 | static bool classof(const Value *V) { |
2308 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2309 | } |
2310 | }; |
2311 | |
2312 | template <> |
2313 | struct OperandTraits<ShuffleVectorInst> : |
2314 | public FixedNumOperandTraits<ShuffleVectorInst, 3> { |
2315 | }; |
2316 | |
2317 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2317, __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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2317, __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 ); } |
2318 | |
2319 | //===----------------------------------------------------------------------===// |
2320 | // ExtractValueInst Class |
2321 | //===----------------------------------------------------------------------===// |
2322 | |
2323 | /// This instruction extracts a struct member or array |
2324 | /// element value from an aggregate value. |
2325 | /// |
2326 | class ExtractValueInst : public UnaryInstruction { |
2327 | SmallVector<unsigned, 4> Indices; |
2328 | |
2329 | ExtractValueInst(const ExtractValueInst &EVI); |
2330 | |
2331 | /// Constructors - Create a extractvalue instruction with a base aggregate |
2332 | /// value and a list of indices. The first ctor can optionally insert before |
2333 | /// an existing instruction, the second appends the new instruction to the |
2334 | /// specified BasicBlock. |
2335 | inline ExtractValueInst(Value *Agg, |
2336 | ArrayRef<unsigned> Idxs, |
2337 | const Twine &NameStr, |
2338 | Instruction *InsertBefore); |
2339 | inline ExtractValueInst(Value *Agg, |
2340 | ArrayRef<unsigned> Idxs, |
2341 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2342 | |
2343 | void init(ArrayRef<unsigned> Idxs, const Twine &NameStr); |
2344 | |
2345 | protected: |
2346 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2347 | friend class Instruction; |
2348 | |
2349 | ExtractValueInst *cloneImpl() const; |
2350 | |
2351 | public: |
2352 | static ExtractValueInst *Create(Value *Agg, |
2353 | ArrayRef<unsigned> Idxs, |
2354 | const Twine &NameStr = "", |
2355 | Instruction *InsertBefore = nullptr) { |
2356 | return new |
2357 | ExtractValueInst(Agg, Idxs, NameStr, InsertBefore); |
2358 | } |
2359 | |
2360 | static ExtractValueInst *Create(Value *Agg, |
2361 | ArrayRef<unsigned> Idxs, |
2362 | const Twine &NameStr, |
2363 | BasicBlock *InsertAtEnd) { |
2364 | return new ExtractValueInst(Agg, Idxs, NameStr, InsertAtEnd); |
2365 | } |
2366 | |
2367 | /// Returns the type of the element that would be extracted |
2368 | /// with an extractvalue instruction with the specified parameters. |
2369 | /// |
2370 | /// Null is returned if the indices are invalid for the specified type. |
2371 | static Type *getIndexedType(Type *Agg, ArrayRef<unsigned> Idxs); |
2372 | |
2373 | using idx_iterator = const unsigned*; |
2374 | |
2375 | inline idx_iterator idx_begin() const { return Indices.begin(); } |
2376 | inline idx_iterator idx_end() const { return Indices.end(); } |
2377 | inline iterator_range<idx_iterator> indices() const { |
2378 | return make_range(idx_begin(), idx_end()); |
2379 | } |
2380 | |
2381 | Value *getAggregateOperand() { |
2382 | return getOperand(0); |
2383 | } |
2384 | const Value *getAggregateOperand() const { |
2385 | return getOperand(0); |
2386 | } |
2387 | static unsigned getAggregateOperandIndex() { |
2388 | return 0U; // get index for modifying correct operand |
2389 | } |
2390 | |
2391 | ArrayRef<unsigned> getIndices() const { |
2392 | return Indices; |
2393 | } |
2394 | |
2395 | unsigned getNumIndices() const { |
2396 | return (unsigned)Indices.size(); |
2397 | } |
2398 | |
2399 | bool hasIndices() const { |
2400 | return true; |
2401 | } |
2402 | |
2403 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2404 | static bool classof(const Instruction *I) { |
2405 | return I->getOpcode() == Instruction::ExtractValue; |
2406 | } |
2407 | static bool classof(const Value *V) { |
2408 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2409 | } |
2410 | }; |
2411 | |
2412 | ExtractValueInst::ExtractValueInst(Value *Agg, |
2413 | ArrayRef<unsigned> Idxs, |
2414 | const Twine &NameStr, |
2415 | Instruction *InsertBefore) |
2416 | : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)), |
2417 | ExtractValue, Agg, InsertBefore) { |
2418 | init(Idxs, NameStr); |
2419 | } |
2420 | |
2421 | ExtractValueInst::ExtractValueInst(Value *Agg, |
2422 | ArrayRef<unsigned> Idxs, |
2423 | const Twine &NameStr, |
2424 | BasicBlock *InsertAtEnd) |
2425 | : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)), |
2426 | ExtractValue, Agg, InsertAtEnd) { |
2427 | init(Idxs, NameStr); |
2428 | } |
2429 | |
2430 | //===----------------------------------------------------------------------===// |
2431 | // InsertValueInst Class |
2432 | //===----------------------------------------------------------------------===// |
2433 | |
2434 | /// This instruction inserts a struct field of array element |
2435 | /// value into an aggregate value. |
2436 | /// |
2437 | class InsertValueInst : public Instruction { |
2438 | SmallVector<unsigned, 4> Indices; |
2439 | |
2440 | InsertValueInst(const InsertValueInst &IVI); |
2441 | |
2442 | /// Constructors - Create a insertvalue instruction with a base aggregate |
2443 | /// value, a value to insert, and a list of indices. The first ctor can |
2444 | /// optionally insert before an existing instruction, the second appends |
2445 | /// the new instruction to the specified BasicBlock. |
2446 | inline InsertValueInst(Value *Agg, Value *Val, |
2447 | ArrayRef<unsigned> Idxs, |
2448 | const Twine &NameStr, |
2449 | Instruction *InsertBefore); |
2450 | inline InsertValueInst(Value *Agg, Value *Val, |
2451 | ArrayRef<unsigned> Idxs, |
2452 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2453 | |
2454 | /// Constructors - These two constructors are convenience methods because one |
2455 | /// and two index insertvalue instructions are so common. |
2456 | InsertValueInst(Value *Agg, Value *Val, unsigned Idx, |
2457 | const Twine &NameStr = "", |
2458 | Instruction *InsertBefore = nullptr); |
2459 | InsertValueInst(Value *Agg, Value *Val, unsigned Idx, const Twine &NameStr, |
2460 | BasicBlock *InsertAtEnd); |
2461 | |
2462 | void init(Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, |
2463 | const Twine &NameStr); |
2464 | |
2465 | protected: |
2466 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2467 | friend class Instruction; |
2468 | |
2469 | InsertValueInst *cloneImpl() const; |
2470 | |
2471 | public: |
2472 | // allocate space for exactly two operands |
2473 | void *operator new(size_t s) { |
2474 | return User::operator new(s, 2); |
2475 | } |
2476 | |
2477 | static InsertValueInst *Create(Value *Agg, Value *Val, |
2478 | ArrayRef<unsigned> Idxs, |
2479 | const Twine &NameStr = "", |
2480 | Instruction *InsertBefore = nullptr) { |
2481 | return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertBefore); |
2482 | } |
2483 | |
2484 | static InsertValueInst *Create(Value *Agg, Value *Val, |
2485 | ArrayRef<unsigned> Idxs, |
2486 | const Twine &NameStr, |
2487 | BasicBlock *InsertAtEnd) { |
2488 | return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertAtEnd); |
2489 | } |
2490 | |
2491 | /// Transparently provide more efficient getOperand methods. |
2492 | 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; |
2493 | |
2494 | using idx_iterator = const unsigned*; |
2495 | |
2496 | inline idx_iterator idx_begin() const { return Indices.begin(); } |
2497 | inline idx_iterator idx_end() const { return Indices.end(); } |
2498 | inline iterator_range<idx_iterator> indices() const { |
2499 | return make_range(idx_begin(), idx_end()); |
2500 | } |
2501 | |
2502 | Value *getAggregateOperand() { |
2503 | return getOperand(0); |
2504 | } |
2505 | const Value *getAggregateOperand() const { |
2506 | return getOperand(0); |
2507 | } |
2508 | static unsigned getAggregateOperandIndex() { |
2509 | return 0U; // get index for modifying correct operand |
2510 | } |
2511 | |
2512 | Value *getInsertedValueOperand() { |
2513 | return getOperand(1); |
2514 | } |
2515 | const Value *getInsertedValueOperand() const { |
2516 | return getOperand(1); |
2517 | } |
2518 | static unsigned getInsertedValueOperandIndex() { |
2519 | return 1U; // get index for modifying correct operand |
2520 | } |
2521 | |
2522 | ArrayRef<unsigned> getIndices() const { |
2523 | return Indices; |
2524 | } |
2525 | |
2526 | unsigned getNumIndices() const { |
2527 | return (unsigned)Indices.size(); |
2528 | } |
2529 | |
2530 | bool hasIndices() const { |
2531 | return true; |
2532 | } |
2533 | |
2534 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2535 | static bool classof(const Instruction *I) { |
2536 | return I->getOpcode() == Instruction::InsertValue; |
2537 | } |
2538 | static bool classof(const Value *V) { |
2539 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2540 | } |
2541 | }; |
2542 | |
2543 | template <> |
2544 | struct OperandTraits<InsertValueInst> : |
2545 | public FixedNumOperandTraits<InsertValueInst, 2> { |
2546 | }; |
2547 | |
2548 | InsertValueInst::InsertValueInst(Value *Agg, |
2549 | Value *Val, |
2550 | ArrayRef<unsigned> Idxs, |
2551 | const Twine &NameStr, |
2552 | Instruction *InsertBefore) |
2553 | : Instruction(Agg->getType(), InsertValue, |
2554 | OperandTraits<InsertValueInst>::op_begin(this), |
2555 | 2, InsertBefore) { |
2556 | init(Agg, Val, Idxs, NameStr); |
2557 | } |
2558 | |
2559 | InsertValueInst::InsertValueInst(Value *Agg, |
2560 | Value *Val, |
2561 | ArrayRef<unsigned> Idxs, |
2562 | const Twine &NameStr, |
2563 | BasicBlock *InsertAtEnd) |
2564 | : Instruction(Agg->getType(), InsertValue, |
2565 | OperandTraits<InsertValueInst>::op_begin(this), |
2566 | 2, InsertAtEnd) { |
2567 | init(Agg, Val, Idxs, NameStr); |
2568 | } |
2569 | |
2570 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2570, __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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2570, __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); } |
2571 | |
2572 | //===----------------------------------------------------------------------===// |
2573 | // PHINode Class |
2574 | //===----------------------------------------------------------------------===// |
2575 | |
2576 | // PHINode - The PHINode class is used to represent the magical mystical PHI |
2577 | // node, that can not exist in nature, but can be synthesized in a computer |
2578 | // scientist's overactive imagination. |
2579 | // |
2580 | class PHINode : public Instruction { |
2581 | /// The number of operands actually allocated. NumOperands is |
2582 | /// the number actually in use. |
2583 | unsigned ReservedSpace; |
2584 | |
2585 | PHINode(const PHINode &PN); |
2586 | |
2587 | explicit PHINode(Type *Ty, unsigned NumReservedValues, |
2588 | const Twine &NameStr = "", |
2589 | Instruction *InsertBefore = nullptr) |
2590 | : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertBefore), |
2591 | ReservedSpace(NumReservedValues) { |
2592 | setName(NameStr); |
2593 | allocHungoffUses(ReservedSpace); |
2594 | } |
2595 | |
2596 | PHINode(Type *Ty, unsigned NumReservedValues, const Twine &NameStr, |
2597 | BasicBlock *InsertAtEnd) |
2598 | : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertAtEnd), |
2599 | ReservedSpace(NumReservedValues) { |
2600 | setName(NameStr); |
2601 | allocHungoffUses(ReservedSpace); |
2602 | } |
2603 | |
2604 | protected: |
2605 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2606 | friend class Instruction; |
2607 | |
2608 | PHINode *cloneImpl() const; |
2609 | |
2610 | // allocHungoffUses - this is more complicated than the generic |
2611 | // User::allocHungoffUses, because we have to allocate Uses for the incoming |
2612 | // values and pointers to the incoming blocks, all in one allocation. |
2613 | void allocHungoffUses(unsigned N) { |
2614 | User::allocHungoffUses(N, /* IsPhi */ true); |
2615 | } |
2616 | |
2617 | public: |
2618 | /// Constructors - NumReservedValues is a hint for the number of incoming |
2619 | /// edges that this phi node will have (use 0 if you really have no idea). |
2620 | static PHINode *Create(Type *Ty, unsigned NumReservedValues, |
2621 | const Twine &NameStr = "", |
2622 | Instruction *InsertBefore = nullptr) { |
2623 | return new PHINode(Ty, NumReservedValues, NameStr, InsertBefore); |
2624 | } |
2625 | |
2626 | static PHINode *Create(Type *Ty, unsigned NumReservedValues, |
2627 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
2628 | return new PHINode(Ty, NumReservedValues, NameStr, InsertAtEnd); |
2629 | } |
2630 | |
2631 | /// Provide fast operand accessors |
2632 | 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; |
2633 | |
2634 | // Block iterator interface. This provides access to the list of incoming |
2635 | // basic blocks, which parallels the list of incoming values. |
2636 | |
2637 | using block_iterator = BasicBlock **; |
2638 | using const_block_iterator = BasicBlock * const *; |
2639 | |
2640 | block_iterator block_begin() { |
2641 | Use::UserRef *ref = |
2642 | reinterpret_cast<Use::UserRef*>(op_begin() + ReservedSpace); |
2643 | return reinterpret_cast<block_iterator>(ref + 1); |
2644 | } |
2645 | |
2646 | const_block_iterator block_begin() const { |
2647 | const Use::UserRef *ref = |
2648 | reinterpret_cast<const Use::UserRef*>(op_begin() + ReservedSpace); |
2649 | return reinterpret_cast<const_block_iterator>(ref + 1); |
2650 | } |
2651 | |
2652 | block_iterator block_end() { |
2653 | return block_begin() + getNumOperands(); |
2654 | } |
2655 | |
2656 | const_block_iterator block_end() const { |
2657 | return block_begin() + getNumOperands(); |
2658 | } |
2659 | |
2660 | iterator_range<block_iterator> blocks() { |
2661 | return make_range(block_begin(), block_end()); |
2662 | } |
2663 | |
2664 | iterator_range<const_block_iterator> blocks() const { |
2665 | return make_range(block_begin(), block_end()); |
2666 | } |
2667 | |
2668 | op_range incoming_values() { return operands(); } |
2669 | |
2670 | const_op_range incoming_values() const { return operands(); } |
2671 | |
2672 | /// Return the number of incoming edges |
2673 | /// |
2674 | unsigned getNumIncomingValues() const { return getNumOperands(); } |
2675 | |
2676 | /// Return incoming value number x |
2677 | /// |
2678 | Value *getIncomingValue(unsigned i) const { |
2679 | return getOperand(i); |
2680 | } |
2681 | void setIncomingValue(unsigned i, Value *V) { |
2682 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2682, __extension__ __PRETTY_FUNCTION__)); |
2683 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2684, __extension__ __PRETTY_FUNCTION__)) |
2684 | "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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2684, __extension__ __PRETTY_FUNCTION__)); |
2685 | setOperand(i, V); |
2686 | } |
2687 | |
2688 | static unsigned getOperandNumForIncomingValue(unsigned i) { |
2689 | return i; |
2690 | } |
2691 | |
2692 | static unsigned getIncomingValueNumForOperand(unsigned i) { |
2693 | return i; |
2694 | } |
2695 | |
2696 | /// Return incoming basic block number @p i. |
2697 | /// |
2698 | BasicBlock *getIncomingBlock(unsigned i) const { |
2699 | return block_begin()[i]; |
2700 | } |
2701 | |
2702 | /// Return incoming basic block corresponding |
2703 | /// to an operand of the PHI. |
2704 | /// |
2705 | BasicBlock *getIncomingBlock(const Use &U) const { |
2706 | 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?\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2706, __extension__ __PRETTY_FUNCTION__)); |
2707 | return getIncomingBlock(unsigned(&U - op_begin())); |
2708 | } |
2709 | |
2710 | /// Return incoming basic block corresponding |
2711 | /// to value use iterator. |
2712 | /// |
2713 | BasicBlock *getIncomingBlock(Value::const_user_iterator I) const { |
2714 | return getIncomingBlock(I.getUse()); |
2715 | } |
2716 | |
2717 | void setIncomingBlock(unsigned i, BasicBlock *BB) { |
2718 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2718, __extension__ __PRETTY_FUNCTION__)); |
2719 | block_begin()[i] = BB; |
2720 | } |
2721 | |
2722 | /// Add an incoming value to the end of the PHI list |
2723 | /// |
2724 | void addIncoming(Value *V, BasicBlock *BB) { |
2725 | if (getNumOperands() == ReservedSpace) |
2726 | growOperands(); // Get more space! |
2727 | // Initialize some new operands. |
2728 | setNumHungOffUseOperands(getNumOperands() + 1); |
2729 | setIncomingValue(getNumOperands() - 1, V); |
2730 | setIncomingBlock(getNumOperands() - 1, BB); |
2731 | } |
2732 | |
2733 | /// Remove an incoming value. This is useful if a |
2734 | /// predecessor basic block is deleted. The value removed is returned. |
2735 | /// |
2736 | /// If the last incoming value for a PHI node is removed (and DeletePHIIfEmpty |
2737 | /// is true), the PHI node is destroyed and any uses of it are replaced with |
2738 | /// dummy values. The only time there should be zero incoming values to a PHI |
2739 | /// node is when the block is dead, so this strategy is sound. |
2740 | /// |
2741 | Value *removeIncomingValue(unsigned Idx, bool DeletePHIIfEmpty = true); |
2742 | |
2743 | Value *removeIncomingValue(const BasicBlock *BB, bool DeletePHIIfEmpty=true) { |
2744 | int Idx = getBasicBlockIndex(BB); |
2745 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2745, __extension__ __PRETTY_FUNCTION__)); |
2746 | return removeIncomingValue(Idx, DeletePHIIfEmpty); |
2747 | } |
2748 | |
2749 | /// Return the first index of the specified basic |
2750 | /// block in the value list for this PHI. Returns -1 if no instance. |
2751 | /// |
2752 | int getBasicBlockIndex(const BasicBlock *BB) const { |
2753 | for (unsigned i = 0, e = getNumOperands(); i != e; ++i) |
2754 | if (block_begin()[i] == BB) |
2755 | return i; |
2756 | return -1; |
2757 | } |
2758 | |
2759 | Value *getIncomingValueForBlock(const BasicBlock *BB) const { |
2760 | int Idx = getBasicBlockIndex(BB); |
2761 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2761, __extension__ __PRETTY_FUNCTION__)); |
2762 | return getIncomingValue(Idx); |
2763 | } |
2764 | |
2765 | /// If the specified PHI node always merges together the |
2766 | /// same value, return the value, otherwise return null. |
2767 | Value *hasConstantValue() const; |
2768 | |
2769 | /// Whether the specified PHI node always merges |
2770 | /// together the same value, assuming undefs are equal to a unique |
2771 | /// non-undef value. |
2772 | bool hasConstantOrUndefValue() const; |
2773 | |
2774 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
2775 | static bool classof(const Instruction *I) { |
2776 | return I->getOpcode() == Instruction::PHI; |
2777 | } |
2778 | static bool classof(const Value *V) { |
2779 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2780 | } |
2781 | |
2782 | private: |
2783 | void growOperands(); |
2784 | }; |
2785 | |
2786 | template <> |
2787 | struct OperandTraits<PHINode> : public HungoffOperandTraits<2> { |
2788 | }; |
2789 | |
2790 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2790, __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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2790, __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); } |
2791 | |
2792 | //===----------------------------------------------------------------------===// |
2793 | // LandingPadInst Class |
2794 | //===----------------------------------------------------------------------===// |
2795 | |
2796 | //===--------------------------------------------------------------------------- |
2797 | /// The landingpad instruction holds all of the information |
2798 | /// necessary to generate correct exception handling. The landingpad instruction |
2799 | /// cannot be moved from the top of a landing pad block, which itself is |
2800 | /// accessible only from the 'unwind' edge of an invoke. This uses the |
2801 | /// SubclassData field in Value to store whether or not the landingpad is a |
2802 | /// cleanup. |
2803 | /// |
2804 | class LandingPadInst : public Instruction { |
2805 | /// The number of operands actually allocated. NumOperands is |
2806 | /// the number actually in use. |
2807 | unsigned ReservedSpace; |
2808 | |
2809 | LandingPadInst(const LandingPadInst &LP); |
2810 | |
2811 | public: |
2812 | enum ClauseType { Catch, Filter }; |
2813 | |
2814 | private: |
2815 | explicit LandingPadInst(Type *RetTy, unsigned NumReservedValues, |
2816 | const Twine &NameStr, Instruction *InsertBefore); |
2817 | explicit LandingPadInst(Type *RetTy, unsigned NumReservedValues, |
2818 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2819 | |
2820 | // Allocate space for exactly zero operands. |
2821 | void *operator new(size_t s) { |
2822 | return User::operator new(s); |
2823 | } |
2824 | |
2825 | void growOperands(unsigned Size); |
2826 | void init(unsigned NumReservedValues, const Twine &NameStr); |
2827 | |
2828 | protected: |
2829 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2830 | friend class Instruction; |
2831 | |
2832 | LandingPadInst *cloneImpl() const; |
2833 | |
2834 | public: |
2835 | /// Constructors - NumReservedClauses is a hint for the number of incoming |
2836 | /// clauses that this landingpad will have (use 0 if you really have no idea). |
2837 | static LandingPadInst *Create(Type *RetTy, unsigned NumReservedClauses, |
2838 | const Twine &NameStr = "", |
2839 | Instruction *InsertBefore = nullptr); |
2840 | static LandingPadInst *Create(Type *RetTy, unsigned NumReservedClauses, |
2841 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2842 | |
2843 | /// Provide fast operand accessors |
2844 | 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; |
2845 | |
2846 | /// Return 'true' if this landingpad instruction is a |
2847 | /// cleanup. I.e., it should be run when unwinding even if its landing pad |
2848 | /// doesn't catch the exception. |
2849 | bool isCleanup() const { return getSubclassDataFromInstruction() & 1; } |
2850 | |
2851 | /// Indicate that this landingpad instruction is a cleanup. |
2852 | void setCleanup(bool V) { |
2853 | setInstructionSubclassData((getSubclassDataFromInstruction() & ~1) | |
2854 | (V ? 1 : 0)); |
2855 | } |
2856 | |
2857 | /// Add a catch or filter clause to the landing pad. |
2858 | void addClause(Constant *ClauseVal); |
2859 | |
2860 | /// Get the value of the clause at index Idx. Use isCatch/isFilter to |
2861 | /// determine what type of clause this is. |
2862 | Constant *getClause(unsigned Idx) const { |
2863 | return cast<Constant>(getOperandList()[Idx]); |
2864 | } |
2865 | |
2866 | /// Return 'true' if the clause and index Idx is a catch clause. |
2867 | bool isCatch(unsigned Idx) const { |
2868 | return !isa<ArrayType>(getOperandList()[Idx]->getType()); |
2869 | } |
2870 | |
2871 | /// Return 'true' if the clause and index Idx is a filter clause. |
2872 | bool isFilter(unsigned Idx) const { |
2873 | return isa<ArrayType>(getOperandList()[Idx]->getType()); |
2874 | } |
2875 | |
2876 | /// Get the number of clauses for this landing pad. |
2877 | unsigned getNumClauses() const { return getNumOperands(); } |
2878 | |
2879 | /// Grow the size of the operand list to accommodate the new |
2880 | /// number of clauses. |
2881 | void reserveClauses(unsigned Size) { growOperands(Size); } |
2882 | |
2883 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2884 | static bool classof(const Instruction *I) { |
2885 | return I->getOpcode() == Instruction::LandingPad; |
2886 | } |
2887 | static bool classof(const Value *V) { |
2888 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2889 | } |
2890 | }; |
2891 | |
2892 | template <> |
2893 | struct OperandTraits<LandingPadInst> : public HungoffOperandTraits<1> { |
2894 | }; |
2895 | |
2896 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2896, __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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2896, __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); } |
2897 | |
2898 | //===----------------------------------------------------------------------===// |
2899 | // ReturnInst Class |
2900 | //===----------------------------------------------------------------------===// |
2901 | |
2902 | //===--------------------------------------------------------------------------- |
2903 | /// Return a value (possibly void), from a function. Execution |
2904 | /// does not continue in this function any longer. |
2905 | /// |
2906 | class ReturnInst : public TerminatorInst { |
2907 | ReturnInst(const ReturnInst &RI); |
2908 | |
2909 | private: |
2910 | // ReturnInst constructors: |
2911 | // ReturnInst() - 'ret void' instruction |
2912 | // ReturnInst( null) - 'ret void' instruction |
2913 | // ReturnInst(Value* X) - 'ret X' instruction |
2914 | // ReturnInst( null, Inst *I) - 'ret void' instruction, insert before I |
2915 | // ReturnInst(Value* X, Inst *I) - 'ret X' instruction, insert before I |
2916 | // ReturnInst( null, BB *B) - 'ret void' instruction, insert @ end of B |
2917 | // ReturnInst(Value* X, BB *B) - 'ret X' instruction, insert @ end of B |
2918 | // |
2919 | // NOTE: If the Value* passed is of type void then the constructor behaves as |
2920 | // if it was passed NULL. |
2921 | explicit ReturnInst(LLVMContext &C, Value *retVal = nullptr, |
2922 | Instruction *InsertBefore = nullptr); |
2923 | ReturnInst(LLVMContext &C, Value *retVal, BasicBlock *InsertAtEnd); |
2924 | explicit ReturnInst(LLVMContext &C, BasicBlock *InsertAtEnd); |
2925 | |
2926 | protected: |
2927 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2928 | friend class Instruction; |
2929 | |
2930 | ReturnInst *cloneImpl() const; |
2931 | |
2932 | public: |
2933 | static ReturnInst* Create(LLVMContext &C, Value *retVal = nullptr, |
2934 | Instruction *InsertBefore = nullptr) { |
2935 | return new(!!retVal) ReturnInst(C, retVal, InsertBefore); |
2936 | } |
2937 | |
2938 | static ReturnInst* Create(LLVMContext &C, Value *retVal, |
2939 | BasicBlock *InsertAtEnd) { |
2940 | return new(!!retVal) ReturnInst(C, retVal, InsertAtEnd); |
2941 | } |
2942 | |
2943 | static ReturnInst* Create(LLVMContext &C, BasicBlock *InsertAtEnd) { |
2944 | return new(0) ReturnInst(C, InsertAtEnd); |
2945 | } |
2946 | |
2947 | /// Provide fast operand accessors |
2948 | 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; |
2949 | |
2950 | /// Convenience accessor. Returns null if there is no return value. |
2951 | Value *getReturnValue() const { |
2952 | return getNumOperands() != 0 ? getOperand(0) : nullptr; |
2953 | } |
2954 | |
2955 | unsigned getNumSuccessors() const { return 0; } |
2956 | |
2957 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2958 | static bool classof(const Instruction *I) { |
2959 | return (I->getOpcode() == Instruction::Ret); |
2960 | } |
2961 | static bool classof(const Value *V) { |
2962 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2963 | } |
2964 | |
2965 | private: |
2966 | friend TerminatorInst; |
2967 | |
2968 | BasicBlock *getSuccessor(unsigned idx) const { |
2969 | llvm_unreachable("ReturnInst has no successors!")::llvm::llvm_unreachable_internal("ReturnInst has no successors!" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2969); |
2970 | } |
2971 | |
2972 | void setSuccessor(unsigned idx, BasicBlock *B) { |
2973 | llvm_unreachable("ReturnInst has no successors!")::llvm::llvm_unreachable_internal("ReturnInst has no successors!" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2973); |
2974 | } |
2975 | }; |
2976 | |
2977 | template <> |
2978 | struct OperandTraits<ReturnInst> : public VariadicOperandTraits<ReturnInst> { |
2979 | }; |
2980 | |
2981 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2981, __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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 2981, __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); } |
2982 | |
2983 | //===----------------------------------------------------------------------===// |
2984 | // BranchInst Class |
2985 | //===----------------------------------------------------------------------===// |
2986 | |
2987 | //===--------------------------------------------------------------------------- |
2988 | /// Conditional or Unconditional Branch instruction. |
2989 | /// |
2990 | class BranchInst : public TerminatorInst { |
2991 | /// Ops list - Branches are strange. The operands are ordered: |
2992 | /// [Cond, FalseDest,] TrueDest. This makes some accessors faster because |
2993 | /// they don't have to check for cond/uncond branchness. These are mostly |
2994 | /// accessed relative from op_end(). |
2995 | BranchInst(const BranchInst &BI); |
2996 | // BranchInst constructors (where {B, T, F} are blocks, and C is a condition): |
2997 | // BranchInst(BB *B) - 'br B' |
2998 | // BranchInst(BB* T, BB *F, Value *C) - 'br C, T, F' |
2999 | // BranchInst(BB* B, Inst *I) - 'br B' insert before I |
3000 | // BranchInst(BB* T, BB *F, Value *C, Inst *I) - 'br C, T, F', insert before I |
3001 | // BranchInst(BB* B, BB *I) - 'br B' insert at end |
3002 | // BranchInst(BB* T, BB *F, Value *C, BB *I) - 'br C, T, F', insert at end |
3003 | explicit BranchInst(BasicBlock *IfTrue, Instruction *InsertBefore = nullptr); |
3004 | BranchInst(BasicBlock *IfTrue, BasicBlock *IfFalse, Value *Cond, |
3005 | Instruction *InsertBefore = nullptr); |
3006 | BranchInst(BasicBlock *IfTrue, BasicBlock *InsertAtEnd); |
3007 | BranchInst(BasicBlock *IfTrue, BasicBlock *IfFalse, Value *Cond, |
3008 | BasicBlock *InsertAtEnd); |
3009 | |
3010 | void AssertOK(); |
3011 | |
3012 | protected: |
3013 | // Note: Instruction needs to be a friend here to call cloneImpl. |
3014 | friend class Instruction; |
3015 | |
3016 | BranchInst *cloneImpl() const; |
3017 | |
3018 | public: |
3019 | static BranchInst *Create(BasicBlock *IfTrue, |
3020 | Instruction *InsertBefore = nullptr) { |
3021 | return new(1) BranchInst(IfTrue, InsertBefore); |
3022 | } |
3023 | |
3024 | static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *IfFalse, |
3025 | Value *Cond, Instruction *InsertBefore = nullptr) { |
3026 | return new(3) BranchInst(IfTrue, IfFalse, Cond, InsertBefore); |
3027 | } |
3028 | |
3029 | static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *InsertAtEnd) { |
3030 | return new(1) BranchInst(IfTrue, InsertAtEnd); |
3031 | } |
3032 | |
3033 | static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *IfFalse, |
3034 | Value *Cond, BasicBlock *InsertAtEnd) { |
3035 | return new(3) BranchInst(IfTrue, IfFalse, Cond, InsertAtEnd); |
3036 | } |
3037 | |
3038 | /// Transparently provide more efficient getOperand methods. |
3039 | 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; |
3040 | |
3041 | bool isUnconditional() const { return getNumOperands() == 1; } |
3042 | bool isConditional() const { return getNumOperands() == 3; } |
3043 | |
3044 | Value *getCondition() const { |
3045 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3045, __extension__ __PRETTY_FUNCTION__)); |
3046 | return Op<-3>(); |
3047 | } |
3048 | |
3049 | void setCondition(Value *V) { |
3050 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3050, __extension__ __PRETTY_FUNCTION__)); |
3051 | Op<-3>() = V; |
3052 | } |
3053 | |
3054 | unsigned getNumSuccessors() const { return 1+isConditional(); } |
3055 | |
3056 | BasicBlock *getSuccessor(unsigned i) const { |
3057 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3057, __extension__ __PRETTY_FUNCTION__)); |
3058 | return cast_or_null<BasicBlock>((&Op<-1>() - i)->get()); |
3059 | } |
3060 | |
3061 | void setSuccessor(unsigned idx, BasicBlock *NewSucc) { |
3062 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3062, __extension__ __PRETTY_FUNCTION__)); |
3063 | *(&Op<-1>() - idx) = NewSucc; |
3064 | } |
3065 | |
3066 | /// Swap the successors of this branch instruction. |
3067 | /// |
3068 | /// Swaps the successors of the branch instruction. This also swaps any |
3069 | /// branch weight metadata associated with the instruction so that it |
3070 | /// continues to map correctly to each operand. |
3071 | void swapSuccessors(); |
3072 | |
3073 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
3074 | static bool classof(const Instruction *I) { |
3075 | return (I->getOpcode() == Instruction::Br); |
3076 | } |
3077 | static bool classof(const Value *V) { |
3078 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
3079 | } |
3080 | }; |
3081 | |
3082 | template <> |
3083 | struct OperandTraits<BranchInst> : public VariadicOperandTraits<BranchInst, 1> { |
3084 | }; |
3085 | |
3086 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3086, __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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3086, __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); } |
3087 | |
3088 | //===----------------------------------------------------------------------===// |
3089 | // SwitchInst Class |
3090 | //===----------------------------------------------------------------------===// |
3091 | |
3092 | //===--------------------------------------------------------------------------- |
3093 | /// Multiway switch |
3094 | /// |
3095 | class SwitchInst : public TerminatorInst { |
3096 | unsigned ReservedSpace; |
3097 | |
3098 | // Operand[0] = Value to switch on |
3099 | // Operand[1] = Default basic block destination |
3100 | // Operand[2n ] = Value to match |
3101 | // Operand[2n+1] = BasicBlock to go to on match |
3102 | SwitchInst(const SwitchInst &SI); |
3103 | |
3104 | /// Create a new switch instruction, specifying a value to switch on and a |
3105 | /// default destination. The number of additional cases can be specified here |
3106 | /// to make memory allocation more efficient. This constructor can also |
3107 | /// auto-insert before another instruction. |
3108 | SwitchInst(Value *Value, BasicBlock *Default, unsigned NumCases, |
3109 | Instruction *InsertBefore); |
3110 | |
3111 | /// Create a new switch instruction, specifying a value to switch on and a |
3112 | /// default destination. The number of additional cases can be specified here |
3113 | /// to make memory allocation more efficient. This constructor also |
3114 | /// auto-inserts at the end of the specified BasicBlock. |
3115 | SwitchInst(Value *Value, BasicBlock *Default, unsigned NumCases, |
3116 | BasicBlock *InsertAtEnd); |
3117 | |
3118 | // allocate space for exactly zero operands |
3119 | void *operator new(size_t s) { |
3120 | return User::operator new(s); |
3121 | } |
3122 | |
3123 | void init(Value *Value, BasicBlock *Default, unsigned NumReserved); |
3124 | void growOperands(); |
3125 | |
3126 | protected: |
3127 | // Note: Instruction needs to be a friend here to call cloneImpl. |
3128 | friend class Instruction; |
3129 | |
3130 | SwitchInst *cloneImpl() const; |
3131 | |
3132 | public: |
3133 | // -2 |
3134 | static const unsigned DefaultPseudoIndex = static_cast<unsigned>(~0L-1); |
3135 | |
3136 | template <typename CaseHandleT> class CaseIteratorImpl; |
3137 | |
3138 | /// A handle to a particular switch case. It exposes a convenient interface |
3139 | /// to both the case value and the successor block. |
3140 | /// |
3141 | /// We define this as a template and instantiate it to form both a const and |
3142 | /// non-const handle. |
3143 | template <typename SwitchInstT, typename ConstantIntT, typename BasicBlockT> |
3144 | class CaseHandleImpl { |
3145 | // Directly befriend both const and non-const iterators. |
3146 | friend class SwitchInst::CaseIteratorImpl< |
3147 | CaseHandleImpl<SwitchInstT, ConstantIntT, BasicBlockT>>; |
3148 | |
3149 | protected: |
3150 | // Expose the switch type we're parameterized with to the iterator. |
3151 | using SwitchInstType = SwitchInstT; |
3152 | |
3153 | SwitchInstT *SI; |
3154 | ptrdiff_t Index; |
3155 | |
3156 | CaseHandleImpl() = default; |
3157 | CaseHandleImpl(SwitchInstT *SI, ptrdiff_t Index) : SI(SI), Index(Index) {} |
3158 | |
3159 | public: |
3160 | /// Resolves case value for current case. |
3161 | ConstantIntT *getCaseValue() const { |
3162 | 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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3163, __extension__ __PRETTY_FUNCTION__)) |
3163 | "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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3163, __extension__ __PRETTY_FUNCTION__)); |
3164 | return reinterpret_cast<ConstantIntT *>(SI->getOperand(2 + Index * 2)); |
3165 | } |
3166 | |
3167 | /// Resolves successor for current case. |
3168 | BasicBlockT *getCaseSuccessor() const { |
3169 | 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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3171, __extension__ __PRETTY_FUNCTION__)) |
3170 | (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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3171, __extension__ __PRETTY_FUNCTION__)) |
3171 | "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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3171, __extension__ __PRETTY_FUNCTION__)); |
3172 | return SI->getSuccessor(getSuccessorIndex()); |
3173 | } |
3174 | |
3175 | /// Returns number of current case. |
3176 | unsigned getCaseIndex() const { return Index; } |
3177 | |
3178 | /// Returns TerminatorInst's successor index for current case successor. |
3179 | unsigned getSuccessorIndex() const { |
3180 | 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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3182, __extension__ __PRETTY_FUNCTION__)) |
3181 | (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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3182, __extension__ __PRETTY_FUNCTION__)) |
3182 | "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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3182, __extension__ __PRETTY_FUNCTION__)); |
3183 | return (unsigned)Index != DefaultPseudoIndex ? Index + 1 : 0; |
3184 | } |
3185 | |
3186 | bool operator==(const CaseHandleImpl &RHS) const { |
3187 | assert(SI == RHS.SI && "Incompatible operators.")(static_cast <bool> (SI == RHS.SI && "Incompatible operators." ) ? void (0) : __assert_fail ("SI == RHS.SI && \"Incompatible operators.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3187, __extension__ __PRETTY_FUNCTION__)); |
3188 | return Index == RHS.Index; |
3189 | } |
3190 | }; |
3191 | |
3192 | using ConstCaseHandle = |
3193 | CaseHandleImpl<const SwitchInst, const ConstantInt, const BasicBlock>; |
3194 | |
3195 | class CaseHandle |
3196 | : public CaseHandleImpl<SwitchInst, ConstantInt, BasicBlock> { |
3197 | friend class SwitchInst::CaseIteratorImpl<CaseHandle>; |
3198 | |
3199 | public: |
3200 | CaseHandle(SwitchInst *SI, ptrdiff_t Index) : CaseHandleImpl(SI, Index) {} |
3201 | |
3202 | /// Sets the new value for current case. |
3203 | void setValue(ConstantInt *V) { |
3204 | 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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3205, __extension__ __PRETTY_FUNCTION__)) |
3205 | "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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3205, __extension__ __PRETTY_FUNCTION__)); |
3206 | SI->setOperand(2 + Index*2, reinterpret_cast<Value*>(V)); |
3207 | } |
3208 | |
3209 | /// Sets the new successor for current case. |
3210 | void setSuccessor(BasicBlock *S) { |
3211 | SI->setSuccessor(getSuccessorIndex(), S); |
3212 | } |
3213 | }; |
3214 | |
3215 | template <typename CaseHandleT> |
3216 | class CaseIteratorImpl |
3217 | : public iterator_facade_base<CaseIteratorImpl<CaseHandleT>, |
3218 | std::random_access_iterator_tag, |
3219 | CaseHandleT> { |
3220 | using SwitchInstT = typename CaseHandleT::SwitchInstType; |
3221 | |
3222 | CaseHandleT Case; |
3223 | |
3224 | public: |
3225 | /// Default constructed iterator is in an invalid state until assigned to |
3226 | /// a case for a particular switch. |
3227 | CaseIteratorImpl() = default; |
3228 | |
3229 | /// Initializes case iterator for given SwitchInst and for given |
3230 | /// case number. |
3231 | CaseIteratorImpl(SwitchInstT *SI, unsigned CaseNum) : Case(SI, CaseNum) {} |
3232 | |
3233 | /// Initializes case iterator for given SwitchInst and for given |
3234 | /// TerminatorInst's successor index. |
3235 | static CaseIteratorImpl fromSuccessorIndex(SwitchInstT *SI, |
3236 | unsigned SuccessorIndex) { |
3237 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3238, __extension__ __PRETTY_FUNCTION__)) |
3238 | "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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3238, __extension__ __PRETTY_FUNCTION__)); |
3239 | return SuccessorIndex != 0 ? CaseIteratorImpl(SI, SuccessorIndex - 1) |
3240 | : CaseIteratorImpl(SI, DefaultPseudoIndex); |
3241 | } |
3242 | |
3243 | /// Support converting to the const variant. This will be a no-op for const |
3244 | /// variant. |
3245 | operator CaseIteratorImpl<ConstCaseHandle>() const { |
3246 | return CaseIteratorImpl<ConstCaseHandle>(Case.SI, Case.Index); |
3247 | } |
3248 | |
3249 | CaseIteratorImpl &operator+=(ptrdiff_t N) { |
3250 | // Check index correctness after addition. |
3251 | // Note: Index == getNumCases() means end(). |
3252 | 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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3254, __extension__ __PRETTY_FUNCTION__)) |
3253 | (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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3254, __extension__ __PRETTY_FUNCTION__)) |
3254 | "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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3254, __extension__ __PRETTY_FUNCTION__)); |
3255 | Case.Index += N; |
3256 | return *this; |
3257 | } |
3258 | CaseIteratorImpl &operator-=(ptrdiff_t N) { |
3259 | // Check index correctness after subtraction. |
3260 | // Note: Case.Index == getNumCases() means end(). |
3261 | 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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3263, __extension__ __PRETTY_FUNCTION__)) |
3262 | (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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3263, __extension__ __PRETTY_FUNCTION__)) |
3263 | "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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3263, __extension__ __PRETTY_FUNCTION__)); |
3264 | Case.Index -= N; |
3265 | return *this; |
3266 | } |
3267 | ptrdiff_t operator-(const CaseIteratorImpl &RHS) const { |
3268 | 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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3268, __extension__ __PRETTY_FUNCTION__)); |
3269 | return Case.Index - RHS.Case.Index; |
3270 | } |
3271 | bool operator==(const CaseIteratorImpl &RHS) const { |
3272 | return Case == RHS.Case; |
3273 | } |
3274 | bool operator<(const CaseIteratorImpl &RHS) const { |
3275 | 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.\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3275, __extension__ __PRETTY_FUNCTION__)); |
3276 | return Case.Index < RHS.Case.Index; |
3277 | } |
3278 | CaseHandleT &operator*() { return Case; } |
3279 | const CaseHandleT &operator*() const { return Case; } |
3280 | }; |
3281 | |
3282 | using CaseIt = CaseIteratorImpl<CaseHandle>; |
3283 | using ConstCaseIt = CaseIteratorImpl<ConstCaseHandle>; |
3284 | |
3285 | static SwitchInst *Create(Value *Value, BasicBlock *Default, |
3286 | unsigned NumCases, |
3287 | Instruction *InsertBefore = nullptr) { |
3288 | return new SwitchInst(Value, Default, NumCases, InsertBefore); |
3289 | } |
3290 | |
3291 | static SwitchInst *Create(Value *Value, BasicBlock *Default, |
3292 | unsigned NumCases, BasicBlock *InsertAtEnd) { |
3293 | return new SwitchInst(Value, Default, NumCases, InsertAtEnd); |
3294 | } |
3295 | |
3296 | /// Provide fast operand accessors |
3297 | 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; |
3298 | |
3299 | // Accessor Methods for Switch stmt |
3300 | Value *getCondition() const { return getOperand(0); } |
3301 | void setCondition(Value *V) { setOperand(0, V); } |
3302 | |
3303 | BasicBlock *getDefaultDest() const { |
3304 | return cast<BasicBlock>(getOperand(1)); |
3305 | } |
3306 | |
3307 | void setDefaultDest(BasicBlock *DefaultCase) { |
3308 | setOperand(1, reinterpret_cast<Value*>(DefaultCase)); |
3309 | } |
3310 | |
3311 | /// Return the number of 'cases' in this switch instruction, excluding the |
3312 | /// default case. |
3313 | unsigned getNumCases() const { |
3314 | return getNumOperands()/2 - 1; |
3315 | } |
3316 | |
3317 | /// Returns a read/write iterator that points to the first case in the |
3318 | /// SwitchInst. |
3319 | CaseIt case_begin() { |
3320 | return CaseIt(this, 0); |
3321 | } |
3322 | |
3323 | /// Returns a read-only iterator that points to the first case in the |
3324 | /// SwitchInst. |
3325 | ConstCaseIt case_begin() const { |
3326 | return ConstCaseIt(this, 0); |
3327 | } |
3328 | |
3329 | /// Returns a read/write iterator that points one past the last in the |
3330 | /// SwitchInst. |
3331 | CaseIt case_end() { |
3332 | return CaseIt(this, getNumCases()); |
3333 | } |
3334 | |
3335 | /// Returns a read-only iterator that points one past the last in the |
3336 | /// SwitchInst. |
3337 | ConstCaseIt case_end() const { |
3338 | return ConstCaseIt(this, getNumCases()); |
3339 | } |
3340 | |
3341 | /// Iteration adapter for range-for loops. |
3342 | iterator_range<CaseIt> cases() { |
3343 | return make_range(case_begin(), case_end()); |
3344 | } |
3345 | |
3346 | /// Constant iteration adapter for range-for loops. |
3347 | iterator_range<ConstCaseIt> cases() const { |
3348 | return make_range(case_begin(), case_end()); |
3349 | } |
3350 | |
3351 | /// Returns an iterator that points to the default case. |
3352 | /// Note: this iterator allows to resolve successor only. Attempt |
3353 | /// to resolve case value causes an assertion. |
3354 | /// Also note, that increment and decrement also causes an assertion and |
3355 | /// makes iterator invalid. |
3356 | CaseIt case_default() { |
3357 | return CaseIt(this, DefaultPseudoIndex); |
3358 | } |
3359 | ConstCaseIt case_default() const { |
3360 | return ConstCaseIt(this, DefaultPseudoIndex); |
3361 | } |
3362 | |
3363 | /// Search all of the case values for the specified constant. If it is |
3364 | /// explicitly handled, return the case iterator of it, otherwise return |
3365 | /// default case iterator to indicate that it is handled by the default |
3366 | /// handler. |
3367 | CaseIt findCaseValue(const ConstantInt *C) { |
3368 | CaseIt I = llvm::find_if( |
3369 | cases(), [C](CaseHandle &Case) { return Case.getCaseValue() == C; }); |
3370 | if (I != case_end()) |
3371 | return I; |
3372 | |
3373 | return case_default(); |
3374 | } |
3375 | ConstCaseIt findCaseValue(const ConstantInt *C) const { |
3376 | ConstCaseIt I = llvm::find_if(cases(), [C](ConstCaseHandle &Case) { |
3377 | return Case.getCaseValue() == C; |
3378 | }); |
3379 | if (I != case_end()) |
3380 | return I; |
3381 | |
3382 | return case_default(); |
3383 | } |
3384 | |
3385 | /// Finds the unique case value for a given successor. Returns null if the |
3386 | /// successor is not found, not unique, or is the default case. |
3387 | ConstantInt *findCaseDest(BasicBlock *BB) { |
3388 | if (BB == getDefaultDest()) |
3389 | return nullptr; |
3390 | |
3391 | ConstantInt *CI = nullptr; |
3392 | for (auto Case : cases()) { |
3393 | if (Case.getCaseSuccessor() != BB) |
3394 | continue; |
3395 | |
3396 | if (CI) |
3397 | return nullptr; // Multiple cases lead to BB. |
3398 | |
3399 | CI = Case.getCaseValue(); |
3400 | } |
3401 | |
3402 | return CI; |
3403 | } |
3404 | |
3405 | /// Add an entry to the switch instruction. |
3406 | /// Note: |
3407 | /// This action invalidates case_end(). Old case_end() iterator will |
3408 | /// point to the added case. |
3409 | void addCase(ConstantInt *OnVal, BasicBlock *Dest); |
3410 | |
3411 | /// This method removes the specified case and its successor from the switch |
3412 | /// instruction. Note that this operation may reorder the remaining cases at |
3413 | /// index idx and above. |
3414 | /// Note: |
3415 | /// This action invalidates iterators for all cases following the one removed, |
3416 | /// including the case_end() iterator. It returns an iterator for the next |
3417 | /// case. |
3418 | CaseIt removeCase(CaseIt I); |
3419 | |
3420 | unsigned getNumSuccessors() const { return getNumOperands()/2; } |
3421 | BasicBlock *getSuccessor(unsigned idx) const { |
3422 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3422, __extension__ __PRETTY_FUNCTION__)); |
3423 | return cast<BasicBlock>(getOperand(idx*2+1)); |
3424 | } |
3425 | void setSuccessor(unsigned idx, BasicBlock *NewSucc) { |
3426 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3426, __extension__ __PRETTY_FUNCTION__)); |
3427 | setOperand(idx * 2 + 1, NewSucc); |
3428 | } |
3429 | |
3430 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
3431 | static bool classof(const Instruction *I) { |
3432 | return I->getOpcode() == Instruction::Switch; |
3433 | } |
3434 | static bool classof(const Value *V) { |
3435 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
3436 | } |
3437 | }; |
3438 | |
3439 | template <> |
3440 | struct OperandTraits<SwitchInst> : public HungoffOperandTraits<2> { |
3441 | }; |
3442 | |
3443 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3443, __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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3443, __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); } |
3444 | |
3445 | //===----------------------------------------------------------------------===// |
3446 | // IndirectBrInst Class |
3447 | //===----------------------------------------------------------------------===// |
3448 | |
3449 | //===--------------------------------------------------------------------------- |
3450 | /// Indirect Branch Instruction. |
3451 | /// |
3452 | class IndirectBrInst : public TerminatorInst { |
3453 | unsigned ReservedSpace; |
3454 | |
3455 | // Operand[0] = Address to jump to |
3456 | // Operand[n+1] = n-th destination |
3457 | IndirectBrInst(const IndirectBrInst &IBI); |
3458 | |
3459 | /// Create a new indirectbr instruction, specifying an |
3460 | /// Address to jump to. The number of expected destinations can be specified |
3461 | /// here to make memory allocation more efficient. This constructor can also |
3462 | /// autoinsert before another instruction. |
3463 | IndirectBrInst(Value *Address, unsigned NumDests, Instruction *InsertBefore); |
3464 | |
3465 | /// Create a new indirectbr instruction, specifying an |
3466 | /// Address to jump to. The number of expected destinations can be specified |
3467 | /// here to make memory allocation more efficient. This constructor also |
3468 | /// autoinserts at the end of the specified BasicBlock. |
3469 | IndirectBrInst(Value *Address, unsigned NumDests, BasicBlock *InsertAtEnd); |
3470 | |
3471 | // allocate space for exactly zero operands |
3472 | void *operator new(size_t s) { |
3473 | return User::operator new(s); |
3474 | } |
3475 | |
3476 | void init(Value *Address, unsigned NumDests); |
3477 | void growOperands(); |
3478 | |
3479 | protected: |
3480 | // Note: Instruction needs to be a friend here to call cloneImpl. |
3481 | friend class Instruction; |
3482 | |
3483 | IndirectBrInst *cloneImpl() const; |
3484 | |
3485 | public: |
3486 | static IndirectBrInst *Create(Value *Address, unsigned NumDests, |
3487 | Instruction *InsertBefore = nullptr) { |
3488 | return new IndirectBrInst(Address, NumDests, InsertBefore); |
3489 | } |
3490 | |
3491 | static IndirectBrInst *Create(Value *Address, unsigned NumDests, |
3492 | BasicBlock *InsertAtEnd) { |
3493 | return new IndirectBrInst(Address, NumDests, InsertAtEnd); |
3494 | } |
3495 | |
3496 | /// Provide fast operand accessors. |
3497 | 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; |
3498 | |
3499 | // Accessor Methods for IndirectBrInst instruction. |
3500 | Value *getAddress() { return getOperand(0); } |
3501 | const Value *getAddress() const { return getOperand(0); } |
3502 | void setAddress(Value *V) { setOperand(0, V); } |
3503 | |
3504 | /// return the number of possible destinations in this |
3505 | /// indirectbr instruction. |
3506 | unsigned getNumDestinations() const { return getNumOperands()-1; } |
3507 | |
3508 | /// Return the specified destination. |
3509 | BasicBlock *getDestination(unsigned i) { return getSuccessor(i); } |
3510 | const BasicBlock *getDestination(unsigned i) const { return getSuccessor(i); } |
3511 | |
3512 | /// Add a destination. |
3513 | /// |
3514 | void addDestination(BasicBlock *Dest); |
3515 | |
3516 | /// This method removes the specified successor from the |
3517 | /// indirectbr instruction. |
3518 | void removeDestination(unsigned i); |
3519 | |
3520 | unsigned getNumSuccessors() const { return getNumOperands()-1; } |
3521 | BasicBlock *getSuccessor(unsigned i) const { |
3522 | return cast<BasicBlock>(getOperand(i+1)); |
3523 | } |
3524 | void setSuccessor(unsigned i, BasicBlock *NewSucc) { |
3525 | setOperand(i + 1, NewSucc); |
3526 | } |
3527 | |
3528 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
3529 | static bool classof(const Instruction *I) { |
3530 | return I->getOpcode() == Instruction::IndirectBr; |
3531 | } |
3532 | static bool classof(const Value *V) { |
3533 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
3534 | } |
3535 | }; |
3536 | |
3537 | template <> |
3538 | struct OperandTraits<IndirectBrInst> : public HungoffOperandTraits<1> { |
3539 | }; |
3540 | |
3541 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3541, __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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3541, __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); } |
3542 | |
3543 | //===----------------------------------------------------------------------===// |
3544 | // InvokeInst Class |
3545 | //===----------------------------------------------------------------------===// |
3546 | |
3547 | /// Invoke instruction. The SubclassData field is used to hold the |
3548 | /// calling convention of the call. |
3549 | /// |
3550 | class InvokeInst : public TerminatorInst, |
3551 | public OperandBundleUser<InvokeInst, User::op_iterator> { |
3552 | friend class OperandBundleUser<InvokeInst, User::op_iterator>; |
3553 | |
3554 | AttributeList Attrs; |
3555 | FunctionType *FTy; |
3556 | |
3557 | InvokeInst(const InvokeInst &BI); |
3558 | |
3559 | /// Construct an InvokeInst given a range of arguments. |
3560 | /// |
3561 | /// Construct an InvokeInst from a range of arguments |
3562 | inline InvokeInst(Value *Func, BasicBlock *IfNormal, BasicBlock *IfException, |
3563 | ArrayRef<Value *> Args, ArrayRef<OperandBundleDef> Bundles, |
3564 | unsigned Values, const Twine &NameStr, |
3565 | Instruction *InsertBefore) |
3566 | : InvokeInst(cast<FunctionType>( |
3567 | cast<PointerType>(Func->getType())->getElementType()), |
3568 | Func, IfNormal, IfException, Args, Bundles, Values, NameStr, |
3569 | InsertBefore) {} |
3570 | |
3571 | inline InvokeInst(FunctionType *Ty, Value *Func, BasicBlock *IfNormal, |
3572 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3573 | ArrayRef<OperandBundleDef> Bundles, unsigned Values, |
3574 | const Twine &NameStr, Instruction *InsertBefore); |
3575 | /// Construct an InvokeInst given a range of arguments. |
3576 | /// |
3577 | /// Construct an InvokeInst from a range of arguments |
3578 | inline InvokeInst(Value *Func, BasicBlock *IfNormal, BasicBlock *IfException, |
3579 | ArrayRef<Value *> Args, ArrayRef<OperandBundleDef> Bundles, |
3580 | unsigned Values, const Twine &NameStr, |
3581 | BasicBlock *InsertAtEnd); |
3582 | |
3583 | bool hasDescriptor() const { return HasDescriptor; } |
3584 | |
3585 | void init(Value *Func, BasicBlock *IfNormal, BasicBlock *IfException, |
3586 | ArrayRef<Value *> Args, ArrayRef<OperandBundleDef> Bundles, |
3587 | const Twine &NameStr) { |
3588 | init(cast<FunctionType>( |
3589 | cast<PointerType>(Func->getType())->getElementType()), |
3590 | Func, IfNormal, IfException, Args, Bundles, NameStr); |
3591 | } |
3592 | |
3593 | void init(FunctionType *FTy, Value *Func, BasicBlock *IfNormal, |
3594 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3595 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr); |
3596 | |
3597 | protected: |
3598 | // Note: Instruction needs to be a friend here to call cloneImpl. |
3599 | friend class Instruction; |
3600 | |
3601 | InvokeInst *cloneImpl() const; |
3602 | |
3603 | public: |
3604 | static InvokeInst *Create(Value *Func, BasicBlock *IfNormal, |
3605 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3606 | const Twine &NameStr, |
3607 | Instruction *InsertBefore = nullptr) { |
3608 | return Create(cast<FunctionType>( |
3609 | cast<PointerType>(Func->getType())->getElementType()), |
3610 | Func, IfNormal, IfException, Args, None, NameStr, |
3611 | InsertBefore); |
3612 | } |
3613 | |
3614 | static InvokeInst *Create(Value *Func, BasicBlock *IfNormal, |
3615 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3616 | ArrayRef<OperandBundleDef> Bundles = None, |
3617 | const Twine &NameStr = "", |
3618 | Instruction *InsertBefore = nullptr) { |
3619 | return Create(cast<FunctionType>( |
3620 | cast<PointerType>(Func->getType())->getElementType()), |
3621 | Func, IfNormal, IfException, Args, Bundles, NameStr, |
3622 | InsertBefore); |
3623 | } |
3624 | |
3625 | static InvokeInst *Create(FunctionType *Ty, Value *Func, BasicBlock *IfNormal, |
3626 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3627 | const Twine &NameStr, |
3628 | Instruction *InsertBefore = nullptr) { |
3629 | unsigned Values = unsigned(Args.size()) + 3; |
3630 | return new (Values) InvokeInst(Ty, Func, IfNormal, IfException, Args, None, |
3631 | Values, NameStr, InsertBefore); |
3632 | } |
3633 | |
3634 | static InvokeInst *Create(FunctionType *Ty, Value *Func, BasicBlock *IfNormal, |
3635 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3636 | ArrayRef<OperandBundleDef> Bundles = None, |
3637 | const Twine &NameStr = "", |
3638 | Instruction *InsertBefore = nullptr) { |
3639 | unsigned Values = unsigned(Args.size()) + CountBundleInputs(Bundles) + 3; |
3640 | unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
3641 | |
3642 | return new (Values, DescriptorBytes) |
3643 | InvokeInst(Ty, Func, IfNormal, IfException, Args, Bundles, Values, |
3644 | NameStr, InsertBefore); |
3645 | } |
3646 | |
3647 | static InvokeInst *Create(Value *Func, |
3648 | BasicBlock *IfNormal, BasicBlock *IfException, |
3649 | ArrayRef<Value *> Args, const Twine &NameStr, |
3650 | BasicBlock *InsertAtEnd) { |
3651 | unsigned Values = unsigned(Args.size()) + 3; |
3652 | return new (Values) InvokeInst(Func, IfNormal, IfException, Args, None, |
3653 | Values, NameStr, InsertAtEnd); |
3654 | } |
3655 | |
3656 | static InvokeInst *Create(Value *Func, BasicBlock *IfNormal, |
3657 | BasicBlock *IfException, ArrayRef<Value *> Args, |
3658 | ArrayRef<OperandBundleDef> Bundles, |
3659 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
3660 | unsigned Values = unsigned(Args.size()) + CountBundleInputs(Bundles) + 3; |
3661 | unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
3662 | |
3663 | return new (Values, DescriptorBytes) |
3664 | InvokeInst(Func, IfNormal, IfException, Args, Bundles, Values, NameStr, |
3665 | InsertAtEnd); |
3666 | } |
3667 | |
3668 | /// Create a clone of \p II with a different set of operand bundles and |
3669 | /// insert it before \p InsertPt. |
3670 | /// |
3671 | /// The returned invoke instruction is identical to \p II in every way except |
3672 | /// that the operand bundles for the new instruction are set to the operand |
3673 | /// bundles in \p Bundles. |
3674 | static InvokeInst *Create(InvokeInst *II, ArrayRef<OperandBundleDef> Bundles, |
3675 | Instruction *InsertPt = nullptr); |
3676 | |
3677 | /// Provide fast operand accessors |
3678 | 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; |
3679 | |
3680 | FunctionType *getFunctionType() const { return FTy; } |
3681 | |
3682 | void mutateFunctionType(FunctionType *FTy) { |
3683 | mutateType(FTy->getReturnType()); |
3684 | this->FTy = FTy; |
3685 | } |
3686 | |
3687 | /// Return the number of invoke arguments. |
3688 | /// |
3689 | unsigned getNumArgOperands() const { |
3690 | return getNumOperands() - getNumTotalBundleOperands() - 3; |
3691 | } |
3692 | |
3693 | /// getArgOperand/setArgOperand - Return/set the i-th invoke argument. |
3694 | /// |
3695 | Value *getArgOperand(unsigned i) const { |
3696 | assert(i < getNumArgOperands() && "Out of bounds!")(static_cast <bool> (i < getNumArgOperands() && "Out of bounds!") ? void (0) : __assert_fail ("i < getNumArgOperands() && \"Out of bounds!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3696, __extension__ __PRETTY_FUNCTION__)); |
3697 | return getOperand(i); |
3698 | } |
3699 | void setArgOperand(unsigned i, Value *v) { |
3700 | assert(i < getNumArgOperands() && "Out of bounds!")(static_cast <bool> (i < getNumArgOperands() && "Out of bounds!") ? void (0) : __assert_fail ("i < getNumArgOperands() && \"Out of bounds!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3700, __extension__ __PRETTY_FUNCTION__)); |
3701 | setOperand(i, v); |
3702 | } |
3703 | |
3704 | /// Return the iterator pointing to the beginning of the argument list. |
3705 | op_iterator arg_begin() { return op_begin(); } |
3706 | |
3707 | /// Return the iterator pointing to the end of the argument list. |
3708 | op_iterator arg_end() { |
3709 | // [ invoke args ], [ operand bundles ], normal dest, unwind dest, callee |
3710 | return op_end() - getNumTotalBundleOperands() - 3; |
3711 | } |
3712 | |
3713 | /// Iteration adapter for range-for loops. |
3714 | iterator_range<op_iterator> arg_operands() { |
3715 | return make_range(arg_begin(), arg_end()); |
3716 | } |
3717 | |
3718 | /// Return the iterator pointing to the beginning of the argument list. |
3719 | const_op_iterator arg_begin() const { return op_begin(); } |
3720 | |
3721 | /// Return the iterator pointing to the end of the argument list. |
3722 | const_op_iterator arg_end() const { |
3723 | // [ invoke args ], [ operand bundles ], normal dest, unwind dest, callee |
3724 | return op_end() - getNumTotalBundleOperands() - 3; |
3725 | } |
3726 | |
3727 | /// Iteration adapter for range-for loops. |
3728 | iterator_range<const_op_iterator> arg_operands() const { |
3729 | return make_range(arg_begin(), arg_end()); |
3730 | } |
3731 | |
3732 | /// Wrappers for getting the \c Use of a invoke argument. |
3733 | const Use &getArgOperandUse(unsigned i) const { |
3734 | assert(i < getNumArgOperands() && "Out of bounds!")(static_cast <bool> (i < getNumArgOperands() && "Out of bounds!") ? void (0) : __assert_fail ("i < getNumArgOperands() && \"Out of bounds!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3734, __extension__ __PRETTY_FUNCTION__)); |
3735 | return getOperandUse(i); |
3736 | } |
3737 | Use &getArgOperandUse(unsigned i) { |
3738 | assert(i < getNumArgOperands() && "Out of bounds!")(static_cast <bool> (i < getNumArgOperands() && "Out of bounds!") ? void (0) : __assert_fail ("i < getNumArgOperands() && \"Out of bounds!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3738, __extension__ __PRETTY_FUNCTION__)); |
3739 | return getOperandUse(i); |
3740 | } |
3741 | |
3742 | /// If one of the arguments has the 'returned' attribute, return its |
3743 | /// operand value. Otherwise, return nullptr. |
3744 | Value *getReturnedArgOperand() const; |
3745 | |
3746 | /// getCallingConv/setCallingConv - Get or set the calling convention of this |
3747 | /// function call. |
3748 | CallingConv::ID getCallingConv() const { |
3749 | return static_cast<CallingConv::ID>(getSubclassDataFromInstruction()); |
3750 | } |
3751 | void setCallingConv(CallingConv::ID CC) { |
3752 | auto ID = static_cast<unsigned>(CC); |
3753 | assert(!(ID & ~CallingConv::MaxID) && "Unsupported calling convention")(static_cast <bool> (!(ID & ~CallingConv::MaxID) && "Unsupported calling convention") ? void (0) : __assert_fail ("!(ID & ~CallingConv::MaxID) && \"Unsupported calling convention\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3753, __extension__ __PRETTY_FUNCTION__)); |
3754 | setInstructionSubclassData(ID); |
3755 | } |
3756 | |
3757 | /// Return the parameter attributes for this invoke. |
3758 | /// |
3759 | AttributeList getAttributes() const { return Attrs; } |
3760 | |
3761 | /// Set the parameter attributes for this invoke. |
3762 | /// |
3763 | void setAttributes(AttributeList A) { Attrs = A; } |
3764 | |
3765 | /// adds the attribute to the list of attributes. |
3766 | void addAttribute(unsigned i, Attribute::AttrKind Kind); |
3767 | |
3768 | /// adds the attribute to the list of attributes. |
3769 | void addAttribute(unsigned i, Attribute Attr); |
3770 | |
3771 | /// Adds the attribute to the indicated argument |
3772 | void addParamAttr(unsigned ArgNo, Attribute::AttrKind Kind); |
3773 | |
3774 | /// removes the attribute from the list of attributes. |
3775 | void removeAttribute(unsigned i, Attribute::AttrKind Kind); |
3776 | |
3777 | /// removes the attribute from the list of attributes. |
3778 | void removeAttribute(unsigned i, StringRef Kind); |
3779 | |
3780 | /// Removes the attribute from the given argument |
3781 | void removeParamAttr(unsigned ArgNo, Attribute::AttrKind Kind); |
3782 | |
3783 | /// adds the dereferenceable attribute to the list of attributes. |
3784 | void addDereferenceableAttr(unsigned i, uint64_t Bytes); |
3785 | |
3786 | /// adds the dereferenceable_or_null attribute to the list of |
3787 | /// attributes. |
3788 | void addDereferenceableOrNullAttr(unsigned i, uint64_t Bytes); |
3789 | |
3790 | /// Determine whether this call has the given attribute. |
3791 | bool hasFnAttr(Attribute::AttrKind Kind) const { |
3792 | assert(Kind != Attribute::NoBuiltin &&(static_cast <bool> (Kind != Attribute::NoBuiltin && "Use CallInst::isNoBuiltin() to check for Attribute::NoBuiltin" ) ? void (0) : __assert_fail ("Kind != Attribute::NoBuiltin && \"Use CallInst::isNoBuiltin() to check for Attribute::NoBuiltin\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3793, __extension__ __PRETTY_FUNCTION__)) |
3793 | "Use CallInst::isNoBuiltin() to check for Attribute::NoBuiltin")(static_cast <bool> (Kind != Attribute::NoBuiltin && "Use CallInst::isNoBuiltin() to check for Attribute::NoBuiltin" ) ? void (0) : __assert_fail ("Kind != Attribute::NoBuiltin && \"Use CallInst::isNoBuiltin() to check for Attribute::NoBuiltin\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3793, __extension__ __PRETTY_FUNCTION__)); |
3794 | return hasFnAttrImpl(Kind); |
3795 | } |
3796 | |
3797 | /// Determine whether this call has the given attribute. |
3798 | bool hasFnAttr(StringRef Kind) const { |
3799 | return hasFnAttrImpl(Kind); |
3800 | } |
3801 | |
3802 | /// Determine whether the return value has the given attribute. |
3803 | bool hasRetAttr(Attribute::AttrKind Kind) const; |
3804 | |
3805 | /// Determine whether the argument or parameter has the given attribute. |
3806 | bool paramHasAttr(unsigned ArgNo, Attribute::AttrKind Kind) const; |
3807 | |
3808 | /// Get the attribute of a given kind at a position. |
3809 | Attribute getAttribute(unsigned i, Attribute::AttrKind Kind) const { |
3810 | return getAttributes().getAttribute(i, Kind); |
3811 | } |
3812 | |
3813 | /// Get the attribute of a given kind at a position. |
3814 | Attribute getAttribute(unsigned i, StringRef Kind) const { |
3815 | return getAttributes().getAttribute(i, Kind); |
3816 | } |
3817 | |
3818 | /// Return true if the data operand at index \p i has the attribute \p |
3819 | /// A. |
3820 | /// |
3821 | /// Data operands include invoke arguments and values used in operand bundles, |
3822 | /// but does not include the invokee operand, or the two successor blocks. |
3823 | /// This routine dispatches to the underlying AttributeList or the |
3824 | /// OperandBundleUser as appropriate. |
3825 | /// |
3826 | /// The index \p i is interpreted as |
3827 | /// |
3828 | /// \p i == Attribute::ReturnIndex -> the return value |
3829 | /// \p i in [1, arg_size + 1) -> argument number (\p i - 1) |
3830 | /// \p i in [arg_size + 1, data_operand_size + 1) -> bundle operand at index |
3831 | /// (\p i - 1) in the operand list. |
3832 | bool dataOperandHasImpliedAttr(unsigned i, Attribute::AttrKind Kind) const; |
3833 | |
3834 | /// Extract the alignment of the return value. |
3835 | unsigned getRetAlignment() const { return Attrs.getRetAlignment(); } |
3836 | |
3837 | /// Extract the alignment for a call or parameter (0=unknown). |
3838 | unsigned getParamAlignment(unsigned ArgNo) const { |
3839 | return Attrs.getParamAlignment(ArgNo); |
3840 | } |
3841 | |
3842 | /// Extract the number of dereferenceable bytes for a call or |
3843 | /// parameter (0=unknown). |
3844 | uint64_t getDereferenceableBytes(unsigned i) const { |
3845 | return Attrs.getDereferenceableBytes(i); |
3846 | } |
3847 | |
3848 | /// Extract the number of dereferenceable_or_null bytes for a call or |
3849 | /// parameter (0=unknown). |
3850 | uint64_t getDereferenceableOrNullBytes(unsigned i) const { |
3851 | return Attrs.getDereferenceableOrNullBytes(i); |
3852 | } |
3853 | |
3854 | /// @brief Determine if the return value is marked with NoAlias attribute. |
3855 | bool returnDoesNotAlias() const { |
3856 | return Attrs.hasAttribute(AttributeList::ReturnIndex, Attribute::NoAlias); |
3857 | } |
3858 | |
3859 | /// Return true if the call should not be treated as a call to a |
3860 | /// builtin. |
3861 | bool isNoBuiltin() const { |
3862 | // We assert in hasFnAttr if one passes in Attribute::NoBuiltin, so we have |
3863 | // to check it by hand. |
3864 | return hasFnAttrImpl(Attribute::NoBuiltin) && |
3865 | !hasFnAttrImpl(Attribute::Builtin); |
3866 | } |
3867 | |
3868 | /// Determine if the call requires strict floating point semantics. |
3869 | bool isStrictFP() const { return hasFnAttr(Attribute::StrictFP); } |
3870 | |
3871 | /// Return true if the call should not be inlined. |
3872 | bool isNoInline() const { return hasFnAttr(Attribute::NoInline); } |
3873 | void setIsNoInline() { |
3874 | addAttribute(AttributeList::FunctionIndex, Attribute::NoInline); |
3875 | } |
3876 | |
3877 | /// Determine if the call does not access memory. |
3878 | bool doesNotAccessMemory() const { |
3879 | return hasFnAttr(Attribute::ReadNone); |
3880 | } |
3881 | void setDoesNotAccessMemory() { |
3882 | addAttribute(AttributeList::FunctionIndex, Attribute::ReadNone); |
3883 | } |
3884 | |
3885 | /// Determine if the call does not access or only reads memory. |
3886 | bool onlyReadsMemory() const { |
3887 | return doesNotAccessMemory() || hasFnAttr(Attribute::ReadOnly); |
3888 | } |
3889 | void setOnlyReadsMemory() { |
3890 | addAttribute(AttributeList::FunctionIndex, Attribute::ReadOnly); |
3891 | } |
3892 | |
3893 | /// Determine if the call does not access or only writes memory. |
3894 | bool doesNotReadMemory() const { |
3895 | return doesNotAccessMemory() || hasFnAttr(Attribute::WriteOnly); |
3896 | } |
3897 | void setDoesNotReadMemory() { |
3898 | addAttribute(AttributeList::FunctionIndex, Attribute::WriteOnly); |
3899 | } |
3900 | |
3901 | /// @brief Determine if the call access memmory only using it's pointer |
3902 | /// arguments. |
3903 | bool onlyAccessesArgMemory() const { |
3904 | return hasFnAttr(Attribute::ArgMemOnly); |
3905 | } |
3906 | void setOnlyAccessesArgMemory() { |
3907 | addAttribute(AttributeList::FunctionIndex, Attribute::ArgMemOnly); |
3908 | } |
3909 | |
3910 | /// @brief Determine if the function may only access memory that is |
3911 | /// inaccessible from the IR. |
3912 | bool onlyAccessesInaccessibleMemory() const { |
3913 | return hasFnAttr(Attribute::InaccessibleMemOnly); |
3914 | } |
3915 | void setOnlyAccessesInaccessibleMemory() { |
3916 | addAttribute(AttributeList::FunctionIndex, Attribute::InaccessibleMemOnly); |
3917 | } |
3918 | |
3919 | /// @brief Determine if the function may only access memory that is |
3920 | /// either inaccessible from the IR or pointed to by its arguments. |
3921 | bool onlyAccessesInaccessibleMemOrArgMem() const { |
3922 | return hasFnAttr(Attribute::InaccessibleMemOrArgMemOnly); |
3923 | } |
3924 | void setOnlyAccessesInaccessibleMemOrArgMem() { |
3925 | addAttribute(AttributeList::FunctionIndex, Attribute::InaccessibleMemOrArgMemOnly); |
3926 | } |
3927 | |
3928 | /// Determine if the call cannot return. |
3929 | bool doesNotReturn() const { return hasFnAttr(Attribute::NoReturn); } |
3930 | void setDoesNotReturn() { |
3931 | addAttribute(AttributeList::FunctionIndex, Attribute::NoReturn); |
3932 | } |
3933 | |
3934 | /// Determine if the call cannot unwind. |
3935 | bool doesNotThrow() const { return hasFnAttr(Attribute::NoUnwind); } |
3936 | void setDoesNotThrow() { |
3937 | addAttribute(AttributeList::FunctionIndex, Attribute::NoUnwind); |
3938 | } |
3939 | |
3940 | /// Determine if the invoke cannot be duplicated. |
3941 | bool cannotDuplicate() const {return hasFnAttr(Attribute::NoDuplicate); } |
3942 | void setCannotDuplicate() { |
3943 | addAttribute(AttributeList::FunctionIndex, Attribute::NoDuplicate); |
3944 | } |
3945 | |
3946 | /// Determine if the invoke is convergent |
3947 | bool isConvergent() const { return hasFnAttr(Attribute::Convergent); } |
3948 | void setConvergent() { |
3949 | addAttribute(AttributeList::FunctionIndex, Attribute::Convergent); |
3950 | } |
3951 | void setNotConvergent() { |
3952 | removeAttribute(AttributeList::FunctionIndex, Attribute::Convergent); |
3953 | } |
3954 | |
3955 | /// Determine if the call returns a structure through first |
3956 | /// pointer argument. |
3957 | bool hasStructRetAttr() const { |
3958 | if (getNumArgOperands() == 0) |
3959 | return false; |
3960 | |
3961 | // Be friendly and also check the callee. |
3962 | return paramHasAttr(0, Attribute::StructRet); |
3963 | } |
3964 | |
3965 | /// Determine if any call argument is an aggregate passed by value. |
3966 | bool hasByValArgument() const { |
3967 | return Attrs.hasAttrSomewhere(Attribute::ByVal); |
3968 | } |
3969 | |
3970 | /// Return the function called, or null if this is an |
3971 | /// indirect function invocation. |
3972 | /// |
3973 | Function *getCalledFunction() const { |
3974 | return dyn_cast<Function>(Op<-3>()); |
3975 | } |
3976 | |
3977 | /// Get a pointer to the function that is invoked by this |
3978 | /// instruction |
3979 | const Value *getCalledValue() const { return Op<-3>(); } |
3980 | Value *getCalledValue() { return Op<-3>(); } |
3981 | |
3982 | /// Set the function called. |
3983 | void setCalledFunction(Value* Fn) { |
3984 | setCalledFunction( |
3985 | cast<FunctionType>(cast<PointerType>(Fn->getType())->getElementType()), |
3986 | Fn); |
3987 | } |
3988 | void setCalledFunction(FunctionType *FTy, Value *Fn) { |
3989 | this->FTy = FTy; |
3990 | assert(FTy == cast<FunctionType>((static_cast <bool> (FTy == cast<FunctionType>( cast <PointerType>(Fn->getType())->getElementType())) ? void (0) : __assert_fail ("FTy == cast<FunctionType>( cast<PointerType>(Fn->getType())->getElementType())" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3991, __extension__ __PRETTY_FUNCTION__)) |
3991 | cast<PointerType>(Fn->getType())->getElementType()))(static_cast <bool> (FTy == cast<FunctionType>( cast <PointerType>(Fn->getType())->getElementType())) ? void (0) : __assert_fail ("FTy == cast<FunctionType>( cast<PointerType>(Fn->getType())->getElementType())" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 3991, __extension__ __PRETTY_FUNCTION__)); |
3992 | Op<-3>() = Fn; |
3993 | } |
3994 | |
3995 | // get*Dest - Return the destination basic blocks... |
3996 | BasicBlock *getNormalDest() const { |
3997 | return cast<BasicBlock>(Op<-2>()); |
3998 | } |
3999 | BasicBlock *getUnwindDest() const { |
4000 | return cast<BasicBlock>(Op<-1>()); |
4001 | } |
4002 | void setNormalDest(BasicBlock *B) { |
4003 | Op<-2>() = reinterpret_cast<Value*>(B); |
4004 | } |
4005 | void setUnwindDest(BasicBlock *B) { |
4006 | Op<-1>() = reinterpret_cast<Value*>(B); |
4007 | } |
4008 | |
4009 | /// Get the landingpad instruction from the landing pad |
4010 | /// block (the unwind destination). |
4011 | LandingPadInst *getLandingPadInst() const; |
4012 | |
4013 | BasicBlock *getSuccessor(unsigned i) const { |
4014 | 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!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4014, __extension__ __PRETTY_FUNCTION__)); |
4015 | return i == 0 ? getNormalDest() : getUnwindDest(); |
4016 | } |
4017 | |
4018 | void setSuccessor(unsigned idx, BasicBlock *NewSucc) { |
4019 | assert(idx < 2 && "Successor # out of range for invoke!")(static_cast <bool> (idx < 2 && "Successor # out of range for invoke!" ) ? void (0) : __assert_fail ("idx < 2 && \"Successor # out of range for invoke!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4019, __extension__ __PRETTY_FUNCTION__)); |
4020 | *(&Op<-2>() + idx) = reinterpret_cast<Value*>(NewSucc); |
4021 | } |
4022 | |
4023 | unsigned getNumSuccessors() const { return 2; } |
4024 | |
4025 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
4026 | static bool classof(const Instruction *I) { |
4027 | return (I->getOpcode() == Instruction::Invoke); |
4028 | } |
4029 | static bool classof(const Value *V) { |
4030 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4031 | } |
4032 | |
4033 | private: |
4034 | template <typename AttrKind> bool hasFnAttrImpl(AttrKind Kind) const { |
4035 | if (Attrs.hasAttribute(AttributeList::FunctionIndex, Kind)) |
4036 | return true; |
4037 | |
4038 | // Operand bundles override attributes on the called function, but don't |
4039 | // override attributes directly present on the invoke instruction. |
4040 | if (isFnAttrDisallowedByOpBundle(Kind)) |
4041 | return false; |
4042 | |
4043 | if (const Function *F = getCalledFunction()) |
4044 | return F->getAttributes().hasAttribute(AttributeList::FunctionIndex, |
4045 | Kind); |
4046 | return false; |
4047 | } |
4048 | |
4049 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
4050 | // method so that subclasses cannot accidentally use it. |
4051 | void setInstructionSubclassData(unsigned short D) { |
4052 | Instruction::setInstructionSubclassData(D); |
4053 | } |
4054 | }; |
4055 | |
4056 | template <> |
4057 | struct OperandTraits<InvokeInst> : public VariadicOperandTraits<InvokeInst, 3> { |
4058 | }; |
4059 | |
4060 | InvokeInst::InvokeInst(FunctionType *Ty, Value *Func, BasicBlock *IfNormal, |
4061 | BasicBlock *IfException, ArrayRef<Value *> Args, |
4062 | ArrayRef<OperandBundleDef> Bundles, unsigned Values, |
4063 | const Twine &NameStr, Instruction *InsertBefore) |
4064 | : TerminatorInst(Ty->getReturnType(), Instruction::Invoke, |
4065 | OperandTraits<InvokeInst>::op_end(this) - Values, Values, |
4066 | InsertBefore) { |
4067 | init(Ty, Func, IfNormal, IfException, Args, Bundles, NameStr); |
4068 | } |
4069 | |
4070 | InvokeInst::InvokeInst(Value *Func, BasicBlock *IfNormal, |
4071 | BasicBlock *IfException, ArrayRef<Value *> Args, |
4072 | ArrayRef<OperandBundleDef> Bundles, unsigned Values, |
4073 | const Twine &NameStr, BasicBlock *InsertAtEnd) |
4074 | : TerminatorInst( |
4075 | cast<FunctionType>(cast<PointerType>(Func->getType()) |
4076 | ->getElementType())->getReturnType(), |
4077 | Instruction::Invoke, OperandTraits<InvokeInst>::op_end(this) - Values, |
4078 | Values, InsertAtEnd) { |
4079 | init(Func, IfNormal, IfException, Args, Bundles, NameStr); |
4080 | } |
4081 | |
4082 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(InvokeInst, Value)InvokeInst::op_iterator InvokeInst::op_begin() { return OperandTraits <InvokeInst>::op_begin(this); } InvokeInst::const_op_iterator InvokeInst::op_begin() const { return OperandTraits<InvokeInst >::op_begin(const_cast<InvokeInst*>(this)); } InvokeInst ::op_iterator InvokeInst::op_end() { return OperandTraits< InvokeInst>::op_end(this); } InvokeInst::const_op_iterator InvokeInst::op_end() const { return OperandTraits<InvokeInst >::op_end(const_cast<InvokeInst*>(this)); } Value *InvokeInst ::getOperand(unsigned i_nocapture) const { (static_cast <bool > (i_nocapture < OperandTraits<InvokeInst>::operands (this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<InvokeInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4082, __extension__ __PRETTY_FUNCTION__)); return cast_or_null <Value>( OperandTraits<InvokeInst>::op_begin(const_cast <InvokeInst*>(this))[i_nocapture].get()); } void InvokeInst ::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (static_cast <bool> (i_nocapture < OperandTraits<InvokeInst> ::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<InvokeInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4082, __extension__ __PRETTY_FUNCTION__)); OperandTraits< InvokeInst>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned InvokeInst::getNumOperands() const { return OperandTraits <InvokeInst>::operands(this); } template <int Idx_nocapture > Use &InvokeInst::Op() { return this->OpFrom<Idx_nocapture >(this); } template <int Idx_nocapture> const Use & InvokeInst::Op() const { return this->OpFrom<Idx_nocapture >(this); } |
4083 | |
4084 | //===----------------------------------------------------------------------===// |
4085 | // ResumeInst Class |
4086 | //===----------------------------------------------------------------------===// |
4087 | |
4088 | //===--------------------------------------------------------------------------- |
4089 | /// Resume the propagation of an exception. |
4090 | /// |
4091 | class ResumeInst : public TerminatorInst { |
4092 | ResumeInst(const ResumeInst &RI); |
4093 | |
4094 | explicit ResumeInst(Value *Exn, Instruction *InsertBefore=nullptr); |
4095 | ResumeInst(Value *Exn, BasicBlock *InsertAtEnd); |
4096 | |
4097 | protected: |
4098 | // Note: Instruction needs to be a friend here to call cloneImpl. |
4099 | friend class Instruction; |
4100 | |
4101 | ResumeInst *cloneImpl() const; |
4102 | |
4103 | public: |
4104 | static ResumeInst *Create(Value *Exn, Instruction *InsertBefore = nullptr) { |
4105 | return new(1) ResumeInst(Exn, InsertBefore); |
4106 | } |
4107 | |
4108 | static ResumeInst *Create(Value *Exn, BasicBlock *InsertAtEnd) { |
4109 | return new(1) ResumeInst(Exn, InsertAtEnd); |
4110 | } |
4111 | |
4112 | /// Provide fast operand accessors |
4113 | 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; |
4114 | |
4115 | /// Convenience accessor. |
4116 | Value *getValue() const { return Op<0>(); } |
4117 | |
4118 | unsigned getNumSuccessors() const { return 0; } |
4119 | |
4120 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
4121 | static bool classof(const Instruction *I) { |
4122 | return I->getOpcode() == Instruction::Resume; |
4123 | } |
4124 | static bool classof(const Value *V) { |
4125 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4126 | } |
4127 | |
4128 | private: |
4129 | friend TerminatorInst; |
4130 | |
4131 | BasicBlock *getSuccessor(unsigned idx) const { |
4132 | llvm_unreachable("ResumeInst has no successors!")::llvm::llvm_unreachable_internal("ResumeInst has no successors!" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4132); |
4133 | } |
4134 | |
4135 | void setSuccessor(unsigned idx, BasicBlock *NewSucc) { |
4136 | llvm_unreachable("ResumeInst has no successors!")::llvm::llvm_unreachable_internal("ResumeInst has no successors!" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4136); |
4137 | } |
4138 | }; |
4139 | |
4140 | template <> |
4141 | struct OperandTraits<ResumeInst> : |
4142 | public FixedNumOperandTraits<ResumeInst, 1> { |
4143 | }; |
4144 | |
4145 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ResumeInst, Value)ResumeInst::op_iterator ResumeInst::op_begin() { return OperandTraits <ResumeInst>::op_begin(this); } ResumeInst::const_op_iterator ResumeInst::op_begin() const { return OperandTraits<ResumeInst >::op_begin(const_cast<ResumeInst*>(this)); } ResumeInst ::op_iterator ResumeInst::op_end() { return OperandTraits< ResumeInst>::op_end(this); } ResumeInst::const_op_iterator ResumeInst::op_end() const { return OperandTraits<ResumeInst >::op_end(const_cast<ResumeInst*>(this)); } Value *ResumeInst ::getOperand(unsigned i_nocapture) const { (static_cast <bool > (i_nocapture < OperandTraits<ResumeInst>::operands (this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<ResumeInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4145, __extension__ __PRETTY_FUNCTION__)); return cast_or_null <Value>( OperandTraits<ResumeInst>::op_begin(const_cast <ResumeInst*>(this))[i_nocapture].get()); } void ResumeInst ::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (static_cast <bool> (i_nocapture < OperandTraits<ResumeInst> ::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<ResumeInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4145, __extension__ __PRETTY_FUNCTION__)); OperandTraits< ResumeInst>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned ResumeInst::getNumOperands() const { return OperandTraits <ResumeInst>::operands(this); } template <int Idx_nocapture > Use &ResumeInst::Op() { return this->OpFrom<Idx_nocapture >(this); } template <int Idx_nocapture> const Use & ResumeInst::Op() const { return this->OpFrom<Idx_nocapture >(this); } |
4146 | |
4147 | //===----------------------------------------------------------------------===// |
4148 | // CatchSwitchInst Class |
4149 | //===----------------------------------------------------------------------===// |
4150 | class CatchSwitchInst : public TerminatorInst { |
4151 | /// The number of operands actually allocated. NumOperands is |
4152 | /// the number actually in use. |
4153 | unsigned ReservedSpace; |
4154 | |
4155 | // Operand[0] = Outer scope |
4156 | // Operand[1] = Unwind block destination |
4157 | // Operand[n] = BasicBlock to go to on match |
4158 | CatchSwitchInst(const CatchSwitchInst &CSI); |
4159 | |
4160 | /// Create a new switch instruction, specifying a |
4161 | /// default destination. The number of additional handlers can be specified |
4162 | /// here to make memory allocation more efficient. |
4163 | /// This constructor can also autoinsert before another instruction. |
4164 | CatchSwitchInst(Value *ParentPad, BasicBlock *UnwindDest, |
4165 | unsigned NumHandlers, const Twine &NameStr, |
4166 | Instruction *InsertBefore); |
4167 | |
4168 | /// Create a new switch instruction, specifying a |
4169 | /// default destination. The number of additional handlers can be specified |
4170 | /// here to make memory allocation more efficient. |
4171 | /// This constructor also autoinserts at the end of the specified BasicBlock. |
4172 | CatchSwitchInst(Value *ParentPad, BasicBlock *UnwindDest, |
4173 | unsigned NumHandlers, const Twine &NameStr, |
4174 | BasicBlock *InsertAtEnd); |
4175 | |
4176 | // allocate space for exactly zero operands |
4177 | void *operator new(size_t s) { return User::operator new(s); } |
4178 | |
4179 | void init(Value *ParentPad, BasicBlock *UnwindDest, unsigned NumReserved); |
4180 | void growOperands(unsigned Size); |
4181 | |
4182 | protected: |
4183 | // Note: Instruction needs to be a friend here to call cloneImpl. |
4184 | friend class Instruction; |
4185 | |
4186 | CatchSwitchInst *cloneImpl() const; |
4187 | |
4188 | public: |
4189 | static CatchSwitchInst *Create(Value *ParentPad, BasicBlock *UnwindDest, |
4190 | unsigned NumHandlers, |
4191 | const Twine &NameStr = "", |
4192 | Instruction *InsertBefore = nullptr) { |
4193 | return new CatchSwitchInst(ParentPad, UnwindDest, NumHandlers, NameStr, |
4194 | InsertBefore); |
4195 | } |
4196 | |
4197 | static CatchSwitchInst *Create(Value *ParentPad, BasicBlock *UnwindDest, |
4198 | unsigned NumHandlers, const Twine &NameStr, |
4199 | BasicBlock *InsertAtEnd) { |
4200 | return new CatchSwitchInst(ParentPad, UnwindDest, NumHandlers, NameStr, |
4201 | InsertAtEnd); |
4202 | } |
4203 | |
4204 | /// Provide fast operand accessors |
4205 | 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; |
4206 | |
4207 | // Accessor Methods for CatchSwitch stmt |
4208 | Value *getParentPad() const { return getOperand(0); } |
4209 | void setParentPad(Value *ParentPad) { setOperand(0, ParentPad); } |
4210 | |
4211 | // Accessor Methods for CatchSwitch stmt |
4212 | bool hasUnwindDest() const { return getSubclassDataFromInstruction() & 1; } |
4213 | bool unwindsToCaller() const { return !hasUnwindDest(); } |
4214 | BasicBlock *getUnwindDest() const { |
4215 | if (hasUnwindDest()) |
4216 | return cast<BasicBlock>(getOperand(1)); |
4217 | return nullptr; |
4218 | } |
4219 | void setUnwindDest(BasicBlock *UnwindDest) { |
4220 | assert(UnwindDest)(static_cast <bool> (UnwindDest) ? void (0) : __assert_fail ("UnwindDest", "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4220, __extension__ __PRETTY_FUNCTION__)); |
4221 | assert(hasUnwindDest())(static_cast <bool> (hasUnwindDest()) ? void (0) : __assert_fail ("hasUnwindDest()", "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4221, __extension__ __PRETTY_FUNCTION__)); |
4222 | setOperand(1, UnwindDest); |
4223 | } |
4224 | |
4225 | /// return the number of 'handlers' in this catchswitch |
4226 | /// instruction, except the default handler |
4227 | unsigned getNumHandlers() const { |
4228 | if (hasUnwindDest()) |
4229 | return getNumOperands() - 2; |
4230 | return getNumOperands() - 1; |
4231 | } |
4232 | |
4233 | private: |
4234 | static BasicBlock *handler_helper(Value *V) { return cast<BasicBlock>(V); } |
4235 | static const BasicBlock *handler_helper(const Value *V) { |
4236 | return cast<BasicBlock>(V); |
4237 | } |
4238 | |
4239 | public: |
4240 | using DerefFnTy = BasicBlock *(*)(Value *); |
4241 | using handler_iterator = mapped_iterator<op_iterator, DerefFnTy>; |
4242 | using handler_range = iterator_range<handler_iterator>; |
4243 | using ConstDerefFnTy = const BasicBlock *(*)(const Value *); |
4244 | using const_handler_iterator = |
4245 | mapped_iterator<const_op_iterator, ConstDerefFnTy>; |
4246 | using const_handler_range = iterator_range<const_handler_iterator>; |
4247 | |
4248 | /// Returns an iterator that points to the first handler in CatchSwitchInst. |
4249 | handler_iterator handler_begin() { |
4250 | op_iterator It = op_begin() + 1; |
4251 | if (hasUnwindDest()) |
4252 | ++It; |
4253 | return handler_iterator(It, DerefFnTy(handler_helper)); |
4254 | } |
4255 | |
4256 | /// Returns an iterator that points to the first handler in the |
4257 | /// CatchSwitchInst. |
4258 | const_handler_iterator handler_begin() const { |
4259 | const_op_iterator It = op_begin() + 1; |
4260 | if (hasUnwindDest()) |
4261 | ++It; |
4262 | return const_handler_iterator(It, ConstDerefFnTy(handler_helper)); |
4263 | } |
4264 | |
4265 | /// Returns a read-only iterator that points one past the last |
4266 | /// handler in the CatchSwitchInst. |
4267 | handler_iterator handler_end() { |
4268 | return handler_iterator(op_end(), DerefFnTy(handler_helper)); |
4269 | } |
4270 | |
4271 | /// Returns an iterator that points one past the last handler in the |
4272 | /// CatchSwitchInst. |
4273 | const_handler_iterator handler_end() const { |
4274 | return const_handler_iterator(op_end(), ConstDerefFnTy(handler_helper)); |
4275 | } |
4276 | |
4277 | /// iteration adapter for range-for loops. |
4278 | handler_range handlers() { |
4279 | return make_range(handler_begin(), handler_end()); |
4280 | } |
4281 | |
4282 | /// iteration adapter for range-for loops. |
4283 | const_handler_range handlers() const { |
4284 | return make_range(handler_begin(), handler_end()); |
4285 | } |
4286 | |
4287 | /// Add an entry to the switch instruction... |
4288 | /// Note: |
4289 | /// This action invalidates handler_end(). Old handler_end() iterator will |
4290 | /// point to the added handler. |
4291 | void addHandler(BasicBlock *Dest); |
4292 | |
4293 | void removeHandler(handler_iterator HI); |
4294 | |
4295 | unsigned getNumSuccessors() const { return getNumOperands() - 1; } |
4296 | BasicBlock *getSuccessor(unsigned Idx) const { |
4297 | assert(Idx < getNumSuccessors() &&(static_cast <bool> (Idx < getNumSuccessors() && "Successor # out of range for catchswitch!") ? void (0) : __assert_fail ("Idx < getNumSuccessors() && \"Successor # out of range for catchswitch!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4298, __extension__ __PRETTY_FUNCTION__)) |
4298 | "Successor # out of range for catchswitch!")(static_cast <bool> (Idx < getNumSuccessors() && "Successor # out of range for catchswitch!") ? void (0) : __assert_fail ("Idx < getNumSuccessors() && \"Successor # out of range for catchswitch!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4298, __extension__ __PRETTY_FUNCTION__)); |
4299 | return cast<BasicBlock>(getOperand(Idx + 1)); |
4300 | } |
4301 | void setSuccessor(unsigned Idx, BasicBlock *NewSucc) { |
4302 | assert(Idx < getNumSuccessors() &&(static_cast <bool> (Idx < getNumSuccessors() && "Successor # out of range for catchswitch!") ? void (0) : __assert_fail ("Idx < getNumSuccessors() && \"Successor # out of range for catchswitch!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4303, __extension__ __PRETTY_FUNCTION__)) |
4303 | "Successor # out of range for catchswitch!")(static_cast <bool> (Idx < getNumSuccessors() && "Successor # out of range for catchswitch!") ? void (0) : __assert_fail ("Idx < getNumSuccessors() && \"Successor # out of range for catchswitch!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4303, __extension__ __PRETTY_FUNCTION__)); |
4304 | setOperand(Idx + 1, NewSucc); |
4305 | } |
4306 | |
4307 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
4308 | static bool classof(const Instruction *I) { |
4309 | return I->getOpcode() == Instruction::CatchSwitch; |
4310 | } |
4311 | static bool classof(const Value *V) { |
4312 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4313 | } |
4314 | }; |
4315 | |
4316 | template <> |
4317 | struct OperandTraits<CatchSwitchInst> : public HungoffOperandTraits<2> {}; |
4318 | |
4319 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(CatchSwitchInst, Value)CatchSwitchInst::op_iterator CatchSwitchInst::op_begin() { return OperandTraits<CatchSwitchInst>::op_begin(this); } CatchSwitchInst ::const_op_iterator CatchSwitchInst::op_begin() const { return OperandTraits<CatchSwitchInst>::op_begin(const_cast< CatchSwitchInst*>(this)); } CatchSwitchInst::op_iterator CatchSwitchInst ::op_end() { return OperandTraits<CatchSwitchInst>::op_end (this); } CatchSwitchInst::const_op_iterator CatchSwitchInst:: op_end() const { return OperandTraits<CatchSwitchInst>:: op_end(const_cast<CatchSwitchInst*>(this)); } Value *CatchSwitchInst ::getOperand(unsigned i_nocapture) const { (static_cast <bool > (i_nocapture < OperandTraits<CatchSwitchInst>:: operands(this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<CatchSwitchInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4319, __extension__ __PRETTY_FUNCTION__)); return cast_or_null <Value>( OperandTraits<CatchSwitchInst>::op_begin (const_cast<CatchSwitchInst*>(this))[i_nocapture].get() ); } void CatchSwitchInst::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (static_cast <bool> (i_nocapture < OperandTraits<CatchSwitchInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<CatchSwitchInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4319, __extension__ __PRETTY_FUNCTION__)); OperandTraits< CatchSwitchInst>::op_begin(this)[i_nocapture] = Val_nocapture ; } unsigned CatchSwitchInst::getNumOperands() const { return OperandTraits<CatchSwitchInst>::operands(this); } template <int Idx_nocapture> Use &CatchSwitchInst::Op() { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &CatchSwitchInst::Op() const { return this->OpFrom<Idx_nocapture>(this); } |
4320 | |
4321 | //===----------------------------------------------------------------------===// |
4322 | // CleanupPadInst Class |
4323 | //===----------------------------------------------------------------------===// |
4324 | class CleanupPadInst : public FuncletPadInst { |
4325 | private: |
4326 | explicit CleanupPadInst(Value *ParentPad, ArrayRef<Value *> Args, |
4327 | unsigned Values, const Twine &NameStr, |
4328 | Instruction *InsertBefore) |
4329 | : FuncletPadInst(Instruction::CleanupPad, ParentPad, Args, Values, |
4330 | NameStr, InsertBefore) {} |
4331 | explicit CleanupPadInst(Value *ParentPad, ArrayRef<Value *> Args, |
4332 | unsigned Values, const Twine &NameStr, |
4333 | BasicBlock *InsertAtEnd) |
4334 | : FuncletPadInst(Instruction::CleanupPad, ParentPad, Args, Values, |
4335 | NameStr, InsertAtEnd) {} |
4336 | |
4337 | public: |
4338 | static CleanupPadInst *Create(Value *ParentPad, ArrayRef<Value *> Args = None, |
4339 | const Twine &NameStr = "", |
4340 | Instruction *InsertBefore = nullptr) { |
4341 | unsigned Values = 1 + Args.size(); |
4342 | return new (Values) |
4343 | CleanupPadInst(ParentPad, Args, Values, NameStr, InsertBefore); |
4344 | } |
4345 | |
4346 | static CleanupPadInst *Create(Value *ParentPad, ArrayRef<Value *> Args, |
4347 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
4348 | unsigned Values = 1 + Args.size(); |
4349 | return new (Values) |
4350 | CleanupPadInst(ParentPad, Args, Values, NameStr, InsertAtEnd); |
4351 | } |
4352 | |
4353 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
4354 | static bool classof(const Instruction *I) { |
4355 | return I->getOpcode() == Instruction::CleanupPad; |
4356 | } |
4357 | static bool classof(const Value *V) { |
4358 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4359 | } |
4360 | }; |
4361 | |
4362 | //===----------------------------------------------------------------------===// |
4363 | // CatchPadInst Class |
4364 | //===----------------------------------------------------------------------===// |
4365 | class CatchPadInst : public FuncletPadInst { |
4366 | private: |
4367 | explicit CatchPadInst(Value *CatchSwitch, ArrayRef<Value *> Args, |
4368 | unsigned Values, const Twine &NameStr, |
4369 | Instruction *InsertBefore) |
4370 | : FuncletPadInst(Instruction::CatchPad, CatchSwitch, Args, Values, |
4371 | NameStr, InsertBefore) {} |
4372 | explicit CatchPadInst(Value *CatchSwitch, ArrayRef<Value *> Args, |
4373 | unsigned Values, const Twine &NameStr, |
4374 | BasicBlock *InsertAtEnd) |
4375 | : FuncletPadInst(Instruction::CatchPad, CatchSwitch, Args, Values, |
4376 | NameStr, InsertAtEnd) {} |
4377 | |
4378 | public: |
4379 | static CatchPadInst *Create(Value *CatchSwitch, ArrayRef<Value *> Args, |
4380 | const Twine &NameStr = "", |
4381 | Instruction *InsertBefore = nullptr) { |
4382 | unsigned Values = 1 + Args.size(); |
4383 | return new (Values) |
4384 | CatchPadInst(CatchSwitch, Args, Values, NameStr, InsertBefore); |
4385 | } |
4386 | |
4387 | static CatchPadInst *Create(Value *CatchSwitch, ArrayRef<Value *> Args, |
4388 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
4389 | unsigned Values = 1 + Args.size(); |
4390 | return new (Values) |
4391 | CatchPadInst(CatchSwitch, Args, Values, NameStr, InsertAtEnd); |
4392 | } |
4393 | |
4394 | /// Convenience accessors |
4395 | CatchSwitchInst *getCatchSwitch() const { |
4396 | return cast<CatchSwitchInst>(Op<-1>()); |
4397 | } |
4398 | void setCatchSwitch(Value *CatchSwitch) { |
4399 | assert(CatchSwitch)(static_cast <bool> (CatchSwitch) ? void (0) : __assert_fail ("CatchSwitch", "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4399, __extension__ __PRETTY_FUNCTION__)); |
4400 | Op<-1>() = CatchSwitch; |
4401 | } |
4402 | |
4403 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
4404 | static bool classof(const Instruction *I) { |
4405 | return I->getOpcode() == Instruction::CatchPad; |
4406 | } |
4407 | static bool classof(const Value *V) { |
4408 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4409 | } |
4410 | }; |
4411 | |
4412 | //===----------------------------------------------------------------------===// |
4413 | // CatchReturnInst Class |
4414 | //===----------------------------------------------------------------------===// |
4415 | |
4416 | class CatchReturnInst : public TerminatorInst { |
4417 | CatchReturnInst(const CatchReturnInst &RI); |
4418 | CatchReturnInst(Value *CatchPad, BasicBlock *BB, Instruction *InsertBefore); |
4419 | CatchReturnInst(Value *CatchPad, BasicBlock *BB, BasicBlock *InsertAtEnd); |
4420 | |
4421 | void init(Value *CatchPad, BasicBlock *BB); |
4422 | |
4423 | protected: |
4424 | // Note: Instruction needs to be a friend here to call cloneImpl. |
4425 | friend class Instruction; |
4426 | |
4427 | CatchReturnInst *cloneImpl() const; |
4428 | |
4429 | public: |
4430 | static CatchReturnInst *Create(Value *CatchPad, BasicBlock *BB, |
4431 | Instruction *InsertBefore = nullptr) { |
4432 | assert(CatchPad)(static_cast <bool> (CatchPad) ? void (0) : __assert_fail ("CatchPad", "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4432, __extension__ __PRETTY_FUNCTION__)); |
4433 | assert(BB)(static_cast <bool> (BB) ? void (0) : __assert_fail ("BB" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4433, __extension__ __PRETTY_FUNCTION__)); |
4434 | return new (2) CatchReturnInst(CatchPad, BB, InsertBefore); |
4435 | } |
4436 | |
4437 | static CatchReturnInst *Create(Value *CatchPad, BasicBlock *BB, |
4438 | BasicBlock *InsertAtEnd) { |
4439 | assert(CatchPad)(static_cast <bool> (CatchPad) ? void (0) : __assert_fail ("CatchPad", "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4439, __extension__ __PRETTY_FUNCTION__)); |
4440 | assert(BB)(static_cast <bool> (BB) ? void (0) : __assert_fail ("BB" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4440, __extension__ __PRETTY_FUNCTION__)); |
4441 | return new (2) CatchReturnInst(CatchPad, BB, InsertAtEnd); |
4442 | } |
4443 | |
4444 | /// Provide fast operand accessors |
4445 | 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; |
4446 | |
4447 | /// Convenience accessors. |
4448 | CatchPadInst *getCatchPad() const { return cast<CatchPadInst>(Op<0>()); } |
4449 | void setCatchPad(CatchPadInst *CatchPad) { |
4450 | assert(CatchPad)(static_cast <bool> (CatchPad) ? void (0) : __assert_fail ("CatchPad", "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4450, __extension__ __PRETTY_FUNCTION__)); |
4451 | Op<0>() = CatchPad; |
4452 | } |
4453 | |
4454 | BasicBlock *getSuccessor() const { return cast<BasicBlock>(Op<1>()); } |
4455 | void setSuccessor(BasicBlock *NewSucc) { |
4456 | assert(NewSucc)(static_cast <bool> (NewSucc) ? void (0) : __assert_fail ("NewSucc", "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4456, __extension__ __PRETTY_FUNCTION__)); |
4457 | Op<1>() = NewSucc; |
4458 | } |
4459 | unsigned getNumSuccessors() const { return 1; } |
4460 | |
4461 | /// Get the parentPad of this catchret's catchpad's catchswitch. |
4462 | /// The successor block is implicitly a member of this funclet. |
4463 | Value *getCatchSwitchParentPad() const { |
4464 | return getCatchPad()->getCatchSwitch()->getParentPad(); |
4465 | } |
4466 | |
4467 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
4468 | static bool classof(const Instruction *I) { |
4469 | return (I->getOpcode() == Instruction::CatchRet); |
4470 | } |
4471 | static bool classof(const Value *V) { |
4472 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4473 | } |
4474 | |
4475 | private: |
4476 | friend TerminatorInst; |
4477 | |
4478 | BasicBlock *getSuccessor(unsigned Idx) const { |
4479 | assert(Idx < getNumSuccessors() && "Successor # out of range for catchret!")(static_cast <bool> (Idx < getNumSuccessors() && "Successor # out of range for catchret!") ? void (0) : __assert_fail ("Idx < getNumSuccessors() && \"Successor # out of range for catchret!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4479, __extension__ __PRETTY_FUNCTION__)); |
4480 | return getSuccessor(); |
4481 | } |
4482 | |
4483 | void setSuccessor(unsigned Idx, BasicBlock *B) { |
4484 | assert(Idx < getNumSuccessors() && "Successor # out of range for catchret!")(static_cast <bool> (Idx < getNumSuccessors() && "Successor # out of range for catchret!") ? void (0) : __assert_fail ("Idx < getNumSuccessors() && \"Successor # out of range for catchret!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4484, __extension__ __PRETTY_FUNCTION__)); |
4485 | setSuccessor(B); |
4486 | } |
4487 | }; |
4488 | |
4489 | template <> |
4490 | struct OperandTraits<CatchReturnInst> |
4491 | : public FixedNumOperandTraits<CatchReturnInst, 2> {}; |
4492 | |
4493 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(CatchReturnInst, Value)CatchReturnInst::op_iterator CatchReturnInst::op_begin() { return OperandTraits<CatchReturnInst>::op_begin(this); } CatchReturnInst ::const_op_iterator CatchReturnInst::op_begin() const { return OperandTraits<CatchReturnInst>::op_begin(const_cast< CatchReturnInst*>(this)); } CatchReturnInst::op_iterator CatchReturnInst ::op_end() { return OperandTraits<CatchReturnInst>::op_end (this); } CatchReturnInst::const_op_iterator CatchReturnInst:: op_end() const { return OperandTraits<CatchReturnInst>:: op_end(const_cast<CatchReturnInst*>(this)); } Value *CatchReturnInst ::getOperand(unsigned i_nocapture) const { (static_cast <bool > (i_nocapture < OperandTraits<CatchReturnInst>:: operands(this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<CatchReturnInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4493, __extension__ __PRETTY_FUNCTION__)); return cast_or_null <Value>( OperandTraits<CatchReturnInst>::op_begin (const_cast<CatchReturnInst*>(this))[i_nocapture].get() ); } void CatchReturnInst::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (static_cast <bool> (i_nocapture < OperandTraits<CatchReturnInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<CatchReturnInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4493, __extension__ __PRETTY_FUNCTION__)); OperandTraits< CatchReturnInst>::op_begin(this)[i_nocapture] = Val_nocapture ; } unsigned CatchReturnInst::getNumOperands() const { return OperandTraits<CatchReturnInst>::operands(this); } template <int Idx_nocapture> Use &CatchReturnInst::Op() { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &CatchReturnInst::Op() const { return this->OpFrom<Idx_nocapture>(this); } |
4494 | |
4495 | //===----------------------------------------------------------------------===// |
4496 | // CleanupReturnInst Class |
4497 | //===----------------------------------------------------------------------===// |
4498 | |
4499 | class CleanupReturnInst : public TerminatorInst { |
4500 | private: |
4501 | CleanupReturnInst(const CleanupReturnInst &RI); |
4502 | CleanupReturnInst(Value *CleanupPad, BasicBlock *UnwindBB, unsigned Values, |
4503 | Instruction *InsertBefore = nullptr); |
4504 | CleanupReturnInst(Value *CleanupPad, BasicBlock *UnwindBB, unsigned Values, |
4505 | BasicBlock *InsertAtEnd); |
4506 | |
4507 | void init(Value *CleanupPad, BasicBlock *UnwindBB); |
4508 | |
4509 | protected: |
4510 | // Note: Instruction needs to be a friend here to call cloneImpl. |
4511 | friend class Instruction; |
4512 | |
4513 | CleanupReturnInst *cloneImpl() const; |
4514 | |
4515 | public: |
4516 | static CleanupReturnInst *Create(Value *CleanupPad, |
4517 | BasicBlock *UnwindBB = nullptr, |
4518 | Instruction *InsertBefore = nullptr) { |
4519 | assert(CleanupPad)(static_cast <bool> (CleanupPad) ? void (0) : __assert_fail ("CleanupPad", "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4519, __extension__ __PRETTY_FUNCTION__)); |
4520 | unsigned Values = 1; |
4521 | if (UnwindBB) |
4522 | ++Values; |
4523 | return new (Values) |
4524 | CleanupReturnInst(CleanupPad, UnwindBB, Values, InsertBefore); |
4525 | } |
4526 | |
4527 | static CleanupReturnInst *Create(Value *CleanupPad, BasicBlock *UnwindBB, |
4528 | BasicBlock *InsertAtEnd) { |
4529 | assert(CleanupPad)(static_cast <bool> (CleanupPad) ? void (0) : __assert_fail ("CleanupPad", "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4529, __extension__ __PRETTY_FUNCTION__)); |
4530 | unsigned Values = 1; |
4531 | if (UnwindBB) |
4532 | ++Values; |
4533 | return new (Values) |
4534 | CleanupReturnInst(CleanupPad, UnwindBB, Values, InsertAtEnd); |
4535 | } |
4536 | |
4537 | /// Provide fast operand accessors |
4538 | 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; |
4539 | |
4540 | bool hasUnwindDest() const { return getSubclassDataFromInstruction() & 1; } |
4541 | bool unwindsToCaller() const { return !hasUnwindDest(); } |
4542 | |
4543 | /// Convenience accessor. |
4544 | CleanupPadInst *getCleanupPad() const { |
4545 | return cast<CleanupPadInst>(Op<0>()); |
4546 | } |
4547 | void setCleanupPad(CleanupPadInst *CleanupPad) { |
4548 | assert(CleanupPad)(static_cast <bool> (CleanupPad) ? void (0) : __assert_fail ("CleanupPad", "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4548, __extension__ __PRETTY_FUNCTION__)); |
4549 | Op<0>() = CleanupPad; |
4550 | } |
4551 | |
4552 | unsigned getNumSuccessors() const { return hasUnwindDest() ? 1 : 0; } |
4553 | |
4554 | BasicBlock *getUnwindDest() const { |
4555 | return hasUnwindDest() ? cast<BasicBlock>(Op<1>()) : nullptr; |
4556 | } |
4557 | void setUnwindDest(BasicBlock *NewDest) { |
4558 | assert(NewDest)(static_cast <bool> (NewDest) ? void (0) : __assert_fail ("NewDest", "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4558, __extension__ __PRETTY_FUNCTION__)); |
4559 | assert(hasUnwindDest())(static_cast <bool> (hasUnwindDest()) ? void (0) : __assert_fail ("hasUnwindDest()", "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4559, __extension__ __PRETTY_FUNCTION__)); |
4560 | Op<1>() = NewDest; |
4561 | } |
4562 | |
4563 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
4564 | static bool classof(const Instruction *I) { |
4565 | return (I->getOpcode() == Instruction::CleanupRet); |
4566 | } |
4567 | static bool classof(const Value *V) { |
4568 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4569 | } |
4570 | |
4571 | private: |
4572 | friend TerminatorInst; |
4573 | |
4574 | BasicBlock *getSuccessor(unsigned Idx) const { |
4575 | assert(Idx == 0)(static_cast <bool> (Idx == 0) ? void (0) : __assert_fail ("Idx == 0", "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4575, __extension__ __PRETTY_FUNCTION__)); |
4576 | return getUnwindDest(); |
4577 | } |
4578 | |
4579 | void setSuccessor(unsigned Idx, BasicBlock *B) { |
4580 | assert(Idx == 0)(static_cast <bool> (Idx == 0) ? void (0) : __assert_fail ("Idx == 0", "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4580, __extension__ __PRETTY_FUNCTION__)); |
4581 | setUnwindDest(B); |
4582 | } |
4583 | |
4584 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
4585 | // method so that subclasses cannot accidentally use it. |
4586 | void setInstructionSubclassData(unsigned short D) { |
4587 | Instruction::setInstructionSubclassData(D); |
4588 | } |
4589 | }; |
4590 | |
4591 | template <> |
4592 | struct OperandTraits<CleanupReturnInst> |
4593 | : public VariadicOperandTraits<CleanupReturnInst, /*MINARITY=*/1> {}; |
4594 | |
4595 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(CleanupReturnInst, Value)CleanupReturnInst::op_iterator CleanupReturnInst::op_begin() { return OperandTraits<CleanupReturnInst>::op_begin(this ); } CleanupReturnInst::const_op_iterator CleanupReturnInst:: op_begin() const { return OperandTraits<CleanupReturnInst> ::op_begin(const_cast<CleanupReturnInst*>(this)); } CleanupReturnInst ::op_iterator CleanupReturnInst::op_end() { return OperandTraits <CleanupReturnInst>::op_end(this); } CleanupReturnInst:: const_op_iterator CleanupReturnInst::op_end() const { return OperandTraits <CleanupReturnInst>::op_end(const_cast<CleanupReturnInst *>(this)); } Value *CleanupReturnInst::getOperand(unsigned i_nocapture) const { (static_cast <bool> (i_nocapture < OperandTraits<CleanupReturnInst>::operands(this) && "getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<CleanupReturnInst>::operands(this) && \"getOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4595, __extension__ __PRETTY_FUNCTION__)); return cast_or_null <Value>( OperandTraits<CleanupReturnInst>::op_begin (const_cast<CleanupReturnInst*>(this))[i_nocapture].get ()); } void CleanupReturnInst::setOperand(unsigned i_nocapture , Value *Val_nocapture) { (static_cast <bool> (i_nocapture < OperandTraits<CleanupReturnInst>::operands(this) && "setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<CleanupReturnInst>::operands(this) && \"setOperand() out of range!\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4595, __extension__ __PRETTY_FUNCTION__)); OperandTraits< CleanupReturnInst>::op_begin(this)[i_nocapture] = Val_nocapture ; } unsigned CleanupReturnInst::getNumOperands() const { return OperandTraits<CleanupReturnInst>::operands(this); } template <int Idx_nocapture> Use &CleanupReturnInst::Op() { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &CleanupReturnInst:: Op() const { return this->OpFrom<Idx_nocapture>(this ); } |
4596 | |
4597 | //===----------------------------------------------------------------------===// |
4598 | // UnreachableInst Class |
4599 | //===----------------------------------------------------------------------===// |
4600 | |
4601 | //===--------------------------------------------------------------------------- |
4602 | /// This function has undefined behavior. In particular, the |
4603 | /// presence of this instruction indicates some higher level knowledge that the |
4604 | /// end of the block cannot be reached. |
4605 | /// |
4606 | class UnreachableInst : public TerminatorInst { |
4607 | protected: |
4608 | // Note: Instruction needs to be a friend here to call cloneImpl. |
4609 | friend class Instruction; |
4610 | |
4611 | UnreachableInst *cloneImpl() const; |
4612 | |
4613 | public: |
4614 | explicit UnreachableInst(LLVMContext &C, Instruction *InsertBefore = nullptr); |
4615 | explicit UnreachableInst(LLVMContext &C, BasicBlock *InsertAtEnd); |
4616 | |
4617 | // allocate space for exactly zero operands |
4618 | void *operator new(size_t s) { |
4619 | return User::operator new(s, 0); |
4620 | } |
4621 | |
4622 | unsigned getNumSuccessors() const { return 0; } |
4623 | |
4624 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
4625 | static bool classof(const Instruction *I) { |
4626 | return I->getOpcode() == Instruction::Unreachable; |
4627 | } |
4628 | static bool classof(const Value *V) { |
4629 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4630 | } |
4631 | |
4632 | private: |
4633 | friend TerminatorInst; |
4634 | |
4635 | BasicBlock *getSuccessor(unsigned idx) const { |
4636 | llvm_unreachable("UnreachableInst has no successors!")::llvm::llvm_unreachable_internal("UnreachableInst has no successors!" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4636); |
4637 | } |
4638 | |
4639 | void setSuccessor(unsigned idx, BasicBlock *B) { |
4640 | llvm_unreachable("UnreachableInst has no successors!")::llvm::llvm_unreachable_internal("UnreachableInst has no successors!" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instructions.h" , 4640); |
4641 | } |
4642 | }; |
4643 | |
4644 | //===----------------------------------------------------------------------===// |
4645 | // TruncInst Class |
4646 | //===----------------------------------------------------------------------===// |
4647 | |
4648 | /// This class represents a truncation of integer types. |
4649 | class TruncInst : public CastInst { |
4650 | protected: |
4651 | // Note: Instruction needs to be a friend here to call cloneImpl. |
4652 | friend class Instruction; |
4653 | |
4654 | /// Clone an identical TruncInst |
4655 | TruncInst *cloneImpl() const; |
4656 | |
4657 | public: |
4658 | /// Constructor with insert-before-instruction semantics |
4659 | TruncInst( |
4660 | Value *S, ///< The value to be truncated |
4661 | Type *Ty, ///< The (smaller) type to truncate to |
4662 | const Twine &NameStr = "", ///< A name for the new instruction |
4663 | Instruction *InsertBefore = nullptr ///< Where to insert the new instruction |
4664 | ); |
4665 | |
4666 | /// Constructor with insert-at-end-of-block semantics |
4667 | TruncInst( |
4668 | Value *S, ///< The value to be truncated |
4669 | Type *Ty, ///< The (smaller) type to truncate to |
4670 | const Twine &NameStr, ///< A name for the new instruction |
4671 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
4672 | ); |
4673 | |
4674 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
4675 | static bool classof(const Instruction *I) { |
4676 | return I->getOpcode() == Trunc; |
4677 | } |
4678 | static bool classof(const Value *V) { |
4679 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4680 | } |
4681 | }; |
4682 | |
4683 | //===----------------------------------------------------------------------===// |
4684 | // ZExtInst Class |
4685 | //===----------------------------------------------------------------------===// |
4686 | |
4687 | /// This class represents zero extension of integer types. |
4688 | class ZExtInst : public CastInst { |
4689 | protected: |
4690 | // Note: Instruction needs to be a friend here to call cloneImpl. |
4691 | friend class Instruction; |
4692 | |
4693 | /// Clone an identical ZExtInst |
4694 | ZExtInst *cloneImpl() const; |
4695 | |
4696 | public: |
4697 | /// Constructor with insert-before-instruction semantics |
4698 | ZExtInst( |
4699 | Value *S, ///< The value to be zero extended |
4700 | Type *Ty, ///< The type to zero extend to |
4701 | const Twine &NameStr = "", ///< A name for the new instruction |
4702 | Instruction *InsertBefore = nullptr ///< Where to insert the new instruction |
4703 | ); |
4704 | |
4705 | /// Constructor with insert-at-end semantics. |
4706 | ZExtInst( |
4707 | Value *S, ///< The value to be zero extended |
4708 | Type *Ty, ///< The type to zero extend to |
4709 | const Twine &NameStr, ///< A name for the new instruction |
4710 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
4711 | ); |
4712 | |
4713 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
4714 | static bool classof(const Instruction *I) { |
4715 | return I->getOpcode() == ZExt; |
4716 | } |
4717 | static bool classof(const Value *V) { |
4718 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4719 | } |
4720 | }; |
4721 | |
4722 | //===----------------------------------------------------------------------===// |
4723 | // SExtInst Class |
4724 | //===----------------------------------------------------------------------===// |
4725 | |
4726 | /// This class represents a sign extension of integer types. |
4727 | class SExtInst : public CastInst { |
4728 | protected: |
4729 | // Note: Instruction needs to be a friend here to call cloneImpl. |
4730 | friend class Instruction; |
4731 | |
4732 | /// Clone an identical SExtInst |
4733 | SExtInst *cloneImpl() const; |
4734 | |
4735 | public: |
4736 | /// Constructor with insert-before-instruction semantics |
4737 | SExtInst( |
4738 | Value *S, ///< The value to be sign extended |
4739 | Type *Ty, ///< The type to sign extend to |
4740 | const Twine &NameStr = "", ///< A name for the new instruction |
4741 | Instruction *InsertBefore = nullptr ///< Where to insert the new instruction |
4742 | ); |
4743 | |
4744 | /// Constructor with insert-at-end-of-block semantics |
4745 | SExtInst( |
4746 | Value *S, ///< The value to be sign extended |
4747 | Type *Ty, ///< The type to sign extend to |
4748 | const Twine &NameStr, ///< A name for the new instruction |
4749 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
4750 | ); |
4751 | |
4752 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
4753 | static bool classof(const Instruction *I) { |
4754 | return I->getOpcode() == SExt; |
4755 | } |
4756 | static bool classof(const Value *V) { |
4757 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4758 | } |
4759 | }; |
4760 | |
4761 | //===----------------------------------------------------------------------===// |
4762 | // FPTruncInst Class |
4763 | //===----------------------------------------------------------------------===// |
4764 | |
4765 | /// This class represents a truncation of floating point types. |
4766 | class FPTruncInst : public CastInst { |
4767 | protected: |
4768 | // Note: Instruction needs to be a friend here to call cloneImpl. |
4769 | friend class Instruction; |
4770 | |
4771 | /// Clone an identical FPTruncInst |
4772 | FPTruncInst *cloneImpl() const; |
4773 | |
4774 | public: |
4775 | /// Constructor with insert-before-instruction semantics |
4776 | FPTruncInst( |
4777 | Value *S, ///< The value to be truncated |
4778 | Type *Ty, ///< The type to truncate to |
4779 | const Twine &NameStr = "", ///< A name for the new instruction |
4780 | Instruction *InsertBefore = nullptr ///< Where to insert the new instruction |
4781 | ); |
4782 | |
4783 | /// Constructor with insert-before-instruction semantics |
4784 | FPTruncInst( |
4785 | Value *S, ///< The value to be truncated |
4786 | Type *Ty, ///< The type to truncate to |
4787 | const Twine &NameStr, ///< A name for the new instruction |
4788 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
4789 | ); |
4790 | |
4791 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
4792 | static bool classof(const Instruction *I) { |
4793 | return I->getOpcode() == FPTrunc; |
4794 | } |
4795 | static bool classof(const Value *V) { |
4796 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4797 | } |
4798 | }; |
4799 | |
4800 | //===----------------------------------------------------------------------===// |
4801 | // FPExtInst Class |
4802 | //===----------------------------------------------------------------------===// |
4803 | |
4804 | /// This class represents an extension of floating point types. |
4805 | class FPExtInst : public CastInst { |
4806 | protected: |
4807 | // Note: Instruction needs to be a friend here to call cloneImpl. |
4808 | friend class Instruction; |
4809 | |
4810 | /// Clone an identical FPExtInst |
4811 | FPExtInst *cloneImpl() const; |
4812 | |
4813 | public: |
4814 | /// Constructor with insert-before-instruction semantics |
4815 | FPExtInst( |
4816 | Value *S, ///< The value to be extended |
4817 | Type *Ty, ///< The type to extend to |
4818 | const Twine &NameStr = "", ///< A name for the new instruction |
4819 | Instruction *InsertBefore = nullptr ///< Where to insert the new instruction |
4820 | ); |
4821 | |
4822 | /// Constructor with insert-at-end-of-block semantics |
4823 | FPExtInst( |
4824 | Value *S, ///< The value to be extended |
4825 | Type *Ty, ///< The type to extend to |
4826 | const Twine &NameStr, ///< A name for the new instruction |
4827 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
4828 | ); |
4829 | |
4830 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
4831 | static bool classof(const Instruction *I) { |
4832 | return I->getOpcode() == FPExt; |
4833 | } |
4834 | static bool classof(const Value *V) { |
4835 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4836 | } |
4837 | }; |
4838 | |
4839 | //===----------------------------------------------------------------------===// |
4840 | // UIToFPInst Class |
4841 | //===----------------------------------------------------------------------===// |
4842 | |
4843 | /// This class represents a cast unsigned integer to floating point. |
4844 | class UIToFPInst : public CastInst { |
4845 | protected: |
4846 | // Note: Instruction needs to be a friend here to call cloneImpl. |
4847 | friend class Instruction; |
4848 | |
4849 | /// Clone an identical UIToFPInst |
4850 | UIToFPInst *cloneImpl() const; |
4851 | |
4852 | public: |
4853 | /// Constructor with insert-before-instruction semantics |
4854 | UIToFPInst( |
4855 | Value *S, ///< The value to be converted |
4856 | Type *Ty, ///< The type to convert to |
4857 | const Twine &NameStr = "", ///< A name for the new instruction |
4858 | Instruction *InsertBefore = nullptr ///< Where to insert the new instruction |
4859 | ); |
4860 | |
4861 | /// Constructor with insert-at-end-of-block semantics |
4862 | UIToFPInst( |
4863 | Value *S, ///< The value to be converted |
4864 | Type *Ty, ///< The type to convert to |
4865 | const Twine &NameStr, ///< A name for the new instruction |
4866 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
4867 | ); |
4868 | |
4869 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
4870 | static bool classof(const Instruction *I) { |
4871 | return I->getOpcode() == UIToFP; |
4872 | } |
4873 | static bool classof(const Value *V) { |
4874 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4875 | } |
4876 | }; |
4877 | |
4878 | //===----------------------------------------------------------------------===// |
4879 | // SIToFPInst Class |
4880 | //===----------------------------------------------------------------------===// |
4881 | |
4882 | /// This class represents a cast from signed integer to floating point. |
4883 | class SIToFPInst : public CastInst { |
4884 | protected: |
4885 | // Note: Instruction needs to be a friend here to call cloneImpl. |
4886 | friend class Instruction; |
4887 | |
4888 | /// Clone an identical SIToFPInst |
4889 | SIToFPInst *cloneImpl() const; |
4890 | |
4891 | public: |
4892 | /// Constructor with insert-before-instruction semantics |
4893 | SIToFPInst( |
4894 | Value *S, ///< The value to be converted |
4895 | Type *Ty, ///< The type to convert to |
4896 | const Twine &NameStr = "", ///< A name for the new instruction |
4897 | Instruction *InsertBefore = nullptr ///< Where to insert the new instruction |
4898 | ); |
4899 | |
4900 | /// Constructor with insert-at-end-of-block semantics |
4901 | SIToFPInst( |
4902 | Value *S, ///< The value to be converted |
4903 | Type *Ty, ///< The type to convert to |
4904 | const Twine &NameStr, ///< A name for the new instruction |
4905 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
4906 | ); |
4907 | |
4908 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
4909 | static bool classof(const Instruction *I) { |
4910 | return I->getOpcode() == SIToFP; |
4911 | } |
4912 | static bool classof(const Value *V) { |
4913 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4914 | } |
4915 | }; |
4916 | |
4917 | //===----------------------------------------------------------------------===// |
4918 | // FPToUIInst Class |
4919 | //===----------------------------------------------------------------------===// |
4920 | |
4921 | /// This class represents a cast from floating point to unsigned integer |
4922 | class FPToUIInst : public CastInst { |
4923 | protected: |
4924 | // Note: Instruction needs to be a friend here to call cloneImpl. |
4925 | friend class Instruction; |
4926 | |
4927 | /// Clone an identical FPToUIInst |
4928 | FPToUIInst *cloneImpl() const; |
4929 | |
4930 | public: |
4931 | /// Constructor with insert-before-instruction semantics |
4932 | FPToUIInst( |
4933 | Value *S, ///< The value to be converted |
4934 | Type *Ty, ///< The type to convert to |
4935 | const Twine &NameStr = "", ///< A name for the new instruction |
4936 | Instruction *InsertBefore = nullptr ///< Where to insert the new instruction |
4937 | ); |
4938 | |
4939 | /// Constructor with insert-at-end-of-block semantics |
4940 | FPToUIInst( |
4941 | Value *S, ///< The value to be converted |
4942 | Type *Ty, ///< The type to convert to |
4943 | const Twine &NameStr, ///< A name for the new instruction |
4944 | BasicBlock *InsertAtEnd ///< Where to insert the new instruction |
4945 | ); |
4946 | |
4947 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
4948 | static bool classof(const Instruction *I) { |
4949 | return I->getOpcode() == FPToUI; |
4950 | } |
4951 | static bool classof(const Value *V) { |
4952 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4953 | } |
4954 | }; |
4955 | |
4956 | //===----------------------------------------------------------------------===// |
4957 | // FPToSIInst Class |
4958 | //===----------------------------------------------------------------------===// |
4959 | |
4960 | /// This class represents a cast from floating point to signed integer. |
4961 | class FPToSIInst : public CastInst { |
4962 | protected: |
4963 | // Note: Instruction needs to be a friend here to call cloneImpl. |
4964 | friend class Instruction; |
4965 | |
4966 | /// Clone an identical FPToSIInst |
4967 | FPToSIInst *cloneImpl() const; |
4968 | |
4969 | public: |
4970 | /// Constructor with insert-before-instruction semantics |
4971 | FPToSIInst( |
4972 | Value *S, ///< The value to be converted |
4973 | Type *Ty, ///< The type to convert to |
4974 | const Twine &NameStr = "", ///< A name for the new instruction |
4975 | Instruction *InsertBefore = nullptr ///< Where to insert the new instruction |
4976 | ); |
4977 | |
4978 | /// Constructor with insert-at-end-of-block semantics |
4979 | FPToSIInst( |
4980 | Value *S, ///< The value to be converted |
4981 | Type *Ty, ///< The type to convert to |
4982 | const Twine &NameStr, ///< A name for the new instruction |
4983 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
4984 | ); |
4985 | |
4986 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
4987 | static bool classof(const Instruction *I) { |
4988 | return I->getOpcode() == FPToSI; |
4989 | } |
4990 | static bool classof(const Value *V) { |
4991 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
4992 | } |
4993 | }; |
4994 | |
4995 | //===----------------------------------------------------------------------===// |
4996 | // IntToPtrInst Class |
4997 | //===----------------------------------------------------------------------===// |
4998 | |
4999 | /// This class represents a cast from an integer to a pointer. |
5000 | class IntToPtrInst : public CastInst { |
5001 | public: |
5002 | // Note: Instruction needs to be a friend here to call cloneImpl. |
5003 | friend class Instruction; |
5004 | |
5005 | /// Constructor with insert-before-instruction semantics |
5006 | IntToPtrInst( |
5007 | Value *S, ///< The value to be converted |
5008 | Type *Ty, ///< The type to convert to |
5009 | const Twine &NameStr = "", ///< A name for the new instruction |
5010 | Instruction *InsertBefore = nullptr ///< Where to insert the new instruction |
5011 | ); |
5012 | |
5013 | /// Constructor with insert-at-end-of-block semantics |
5014 | IntToPtrInst( |
5015 | Value *S, ///< The value to be converted |
5016 | Type *Ty, ///< The type to convert to |
5017 | const Twine &NameStr, ///< A name for the new instruction |
5018 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
5019 | ); |
5020 | |
5021 | /// Clone an identical IntToPtrInst. |
5022 | IntToPtrInst *cloneImpl() const; |
5023 | |
5024 | /// Returns the address space of this instruction's pointer type. |
5025 | unsigned getAddressSpace() const { |
5026 | return getType()->getPointerAddressSpace(); |
5027 | } |
5028 | |
5029 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
5030 | static bool classof(const Instruction *I) { |
5031 | return I->getOpcode() == IntToPtr; |
5032 | } |
5033 | static bool classof(const Value *V) { |
5034 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
5035 | } |
5036 | }; |
5037 | |
5038 | //===----------------------------------------------------------------------===// |
5039 | // PtrToIntInst Class |
5040 | //===----------------------------------------------------------------------===// |
5041 | |
5042 | /// This class represents a cast from a pointer to an integer. |
5043 | class PtrToIntInst : public CastInst { |
5044 | protected: |
5045 | // Note: Instruction needs to be a friend here to call cloneImpl. |
5046 | friend class Instruction; |
5047 | |
5048 | /// Clone an identical PtrToIntInst. |
5049 | PtrToIntInst *cloneImpl() const; |
5050 | |
5051 | public: |
5052 | /// Constructor with insert-before-instruction semantics |
5053 | PtrToIntInst( |
5054 | Value *S, ///< The value to be converted |
5055 | Type *Ty, ///< The type to convert to |
5056 | const Twine &NameStr = "", ///< A name for the new instruction |
5057 | Instruction *InsertBefore = nullptr ///< Where to insert the new instruction |
5058 | ); |
5059 | |
5060 | /// Constructor with insert-at-end-of-block semantics |
5061 | PtrToIntInst( |
5062 | Value *S, ///< The value to be converted |
5063 | Type *Ty, ///< The type to convert to |
5064 | const Twine &NameStr, ///< A name for the new instruction |
5065 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
5066 | ); |
5067 | |
5068 | /// Gets the pointer operand. |
5069 | Value *getPointerOperand() { return getOperand(0); } |
5070 | /// Gets the pointer operand. |
5071 | const Value *getPointerOperand() const { return getOperand(0); } |
5072 | /// Gets the operand index of the pointer operand. |
5073 | static unsigned getPointerOperandIndex() { return 0U; } |
5074 | |
5075 | /// Returns the address space of the pointer operand. |
5076 | unsigned getPointerAddressSpace() const { |
5077 | return getPointerOperand()->getType()->getPointerAddressSpace(); |
5078 | } |
5079 | |
5080 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
5081 | static bool classof(const Instruction *I) { |
5082 | return I->getOpcode() == PtrToInt; |
5083 | } |
5084 | static bool classof(const Value *V) { |
5085 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
5086 | } |
5087 | }; |
5088 | |
5089 | //===----------------------------------------------------------------------===// |
5090 | // BitCastInst Class |
5091 | //===----------------------------------------------------------------------===// |
5092 | |
5093 | /// This class represents a no-op cast from one type to another. |
5094 | class BitCastInst : public CastInst { |
5095 | protected: |
5096 | // Note: Instruction needs to be a friend here to call cloneImpl. |
5097 | friend class Instruction; |
5098 | |
5099 | /// Clone an identical BitCastInst. |
5100 | BitCastInst *cloneImpl() const; |
5101 | |
5102 | public: |
5103 | /// Constructor with insert-before-instruction semantics |
5104 | BitCastInst( |
5105 | Value *S, ///< The value to be casted |
5106 | Type *Ty, ///< The type to casted to |
5107 | const Twine &NameStr = "", ///< A name for the new instruction |
5108 | Instruction *InsertBefore = nullptr ///< Where to insert the new instruction |
5109 | ); |
5110 | |
5111 | /// Constructor with insert-at-end-of-block semantics |
5112 | BitCastInst( |
5113 | Value *S, ///< The value to be casted |
5114 | Type *Ty, ///< The type to casted to |
5115 | const Twine &NameStr, ///< A name for the new instruction |
5116 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
5117 | ); |
5118 | |
5119 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
5120 | static bool classof(const Instruction *I) { |
5121 | return I->getOpcode() == BitCast; |
5122 | } |
5123 | static bool classof(const Value *V) { |
5124 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
5125 | } |
5126 | }; |
5127 | |
5128 | //===----------------------------------------------------------------------===// |
5129 | // AddrSpaceCastInst Class |
5130 | //===----------------------------------------------------------------------===// |
5131 | |
5132 | /// This class represents a conversion between pointers from one address space |
5133 | /// to another. |
5134 | class AddrSpaceCastInst : public CastInst { |
5135 | protected: |
5136 | // Note: Instruction needs to be a friend here to call cloneImpl. |
5137 | friend class Instruction; |
5138 | |
5139 | /// Clone an identical AddrSpaceCastInst. |
5140 | AddrSpaceCastInst *cloneImpl() const; |
5141 | |
5142 | public: |
5143 | /// Constructor with insert-before-instruction semantics |
5144 | AddrSpaceCastInst( |
5145 | Value *S, ///< The value to be casted |
5146 | Type *Ty, ///< The type to casted to |
5147 | const Twine &NameStr = "", ///< A name for the new instruction |
5148 | Instruction *InsertBefore = nullptr ///< Where to insert the new instruction |
5149 | ); |
5150 | |
5151 | /// Constructor with insert-at-end-of-block semantics |
5152 | AddrSpaceCastInst( |
5153 | Value *S, ///< The value to be casted |
5154 | Type *Ty, ///< The type to casted to |
5155 | const Twine &NameStr, ///< A name for the new instruction |
5156 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
5157 | ); |
5158 | |
5159 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
5160 | static bool classof(const Instruction *I) { |
5161 | return I->getOpcode() == AddrSpaceCast; |
5162 | } |
5163 | static bool classof(const Value *V) { |
5164 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
5165 | } |
5166 | |
5167 | /// Gets the pointer operand. |
5168 | Value *getPointerOperand() { |
5169 | return getOperand(0); |
5170 | } |
5171 | |
5172 | /// Gets the pointer operand. |
5173 | const Value *getPointerOperand() const { |
5174 | return getOperand(0); |
5175 | } |
5176 | |
5177 | /// Gets the operand index of the pointer operand. |
5178 | static unsigned getPointerOperandIndex() { |
5179 | return 0U; |
5180 | } |
5181 | |
5182 | /// Returns the address space of the pointer operand. |
5183 | unsigned getSrcAddressSpace() const { |
5184 | return getPointerOperand()->getType()->getPointerAddressSpace(); |
5185 | } |
5186 | |
5187 | /// Returns the address space of the result. |
5188 | unsigned getDestAddressSpace() const { |
5189 | return getType()->getPointerAddressSpace(); |
5190 | } |
5191 | }; |
5192 | |
5193 | } // end namespace llvm |
5194 | |
5195 | #endif // LLVM_IR_INSTRUCTIONS_H |
1 | //===-- llvm/Instruction.h - Instruction class definition -------*- C++ -*-===// |
2 | // |
3 | // The LLVM Compiler Infrastructure |
4 | // |
5 | // This file is distributed under the University of Illinois Open Source |
6 | // License. See LICENSE.TXT for details. |
7 | // |
8 | //===----------------------------------------------------------------------===// |
9 | // |
10 | // This file contains the declaration of the Instruction class, which is the |
11 | // base class for all of the LLVM instructions. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #ifndef LLVM_IR_INSTRUCTION_H |
16 | #define LLVM_IR_INSTRUCTION_H |
17 | |
18 | #include "llvm/ADT/ArrayRef.h" |
19 | #include "llvm/ADT/None.h" |
20 | #include "llvm/ADT/StringRef.h" |
21 | #include "llvm/ADT/ilist_node.h" |
22 | #include "llvm/IR/DebugLoc.h" |
23 | #include "llvm/IR/SymbolTableListTraits.h" |
24 | #include "llvm/IR/User.h" |
25 | #include "llvm/IR/Value.h" |
26 | #include "llvm/Support/Casting.h" |
27 | #include <algorithm> |
28 | #include <cassert> |
29 | #include <cstdint> |
30 | #include <utility> |
31 | |
32 | namespace llvm { |
33 | |
34 | class BasicBlock; |
35 | class FastMathFlags; |
36 | class MDNode; |
37 | class Module; |
38 | struct AAMDNodes; |
39 | |
40 | template <> struct ilist_alloc_traits<Instruction> { |
41 | static inline void deleteNode(Instruction *V); |
42 | }; |
43 | |
44 | class Instruction : public User, |
45 | public ilist_node_with_parent<Instruction, BasicBlock> { |
46 | BasicBlock *Parent; |
47 | DebugLoc DbgLoc; // 'dbg' Metadata cache. |
48 | |
49 | enum { |
50 | /// This is a bit stored in the SubClassData field which indicates whether |
51 | /// this instruction has metadata attached to it or not. |
52 | HasMetadataBit = 1 << 15 |
53 | }; |
54 | |
55 | protected: |
56 | ~Instruction(); // Use deleteValue() to delete a generic Instruction. |
57 | |
58 | public: |
59 | Instruction(const Instruction &) = delete; |
60 | Instruction &operator=(const Instruction &) = delete; |
61 | |
62 | /// Specialize the methods defined in Value, as we know that an instruction |
63 | /// can only be used by other instructions. |
64 | Instruction *user_back() { return cast<Instruction>(*user_begin());} |
65 | const Instruction *user_back() const { return cast<Instruction>(*user_begin());} |
66 | |
67 | inline const BasicBlock *getParent() const { return Parent; } |
68 | inline BasicBlock *getParent() { return Parent; } |
69 | |
70 | /// Return the module owning the function this instruction belongs to |
71 | /// or nullptr it the function does not have a module. |
72 | /// |
73 | /// Note: this is undefined behavior if the instruction does not have a |
74 | /// parent, or the parent basic block does not have a parent function. |
75 | const Module *getModule() const; |
76 | Module *getModule() { |
77 | return const_cast<Module *>( |
78 | static_cast<const Instruction *>(this)->getModule()); |
79 | } |
80 | |
81 | /// Return the function this instruction belongs to. |
82 | /// |
83 | /// Note: it is undefined behavior to call this on an instruction not |
84 | /// currently inserted into a function. |
85 | const Function *getFunction() const; |
86 | Function *getFunction() { |
87 | return const_cast<Function *>( |
88 | static_cast<const Instruction *>(this)->getFunction()); |
89 | } |
90 | |
91 | /// This method unlinks 'this' from the containing basic block, but does not |
92 | /// delete it. |
93 | void removeFromParent(); |
94 | |
95 | /// This method unlinks 'this' from the containing basic block and deletes it. |
96 | /// |
97 | /// \returns an iterator pointing to the element after the erased one |
98 | SymbolTableList<Instruction>::iterator eraseFromParent(); |
99 | |
100 | /// Insert an unlinked instruction into a basic block immediately before |
101 | /// the specified instruction. |
102 | void insertBefore(Instruction *InsertPos); |
103 | |
104 | /// Insert an unlinked instruction into a basic block immediately after the |
105 | /// specified instruction. |
106 | void insertAfter(Instruction *InsertPos); |
107 | |
108 | /// Unlink this instruction from its current basic block and insert it into |
109 | /// the basic block that MovePos lives in, right before MovePos. |
110 | void moveBefore(Instruction *MovePos); |
111 | |
112 | /// Unlink this instruction and insert into BB before I. |
113 | /// |
114 | /// \pre I is a valid iterator into BB. |
115 | void moveBefore(BasicBlock &BB, SymbolTableList<Instruction>::iterator I); |
116 | |
117 | /// Unlink this instruction from its current basic block and insert it into |
118 | /// the basic block that MovePos lives in, right after MovePos. |
119 | void moveAfter(Instruction *MovePos); |
120 | |
121 | //===--------------------------------------------------------------------===// |
122 | // Subclass classification. |
123 | //===--------------------------------------------------------------------===// |
124 | |
125 | /// Returns a member of one of the enums like Instruction::Add. |
126 | unsigned getOpcode() const { return getValueID() - InstructionVal; } |
127 | |
128 | const char *getOpcodeName() const { return getOpcodeName(getOpcode()); } |
129 | bool isTerminator() const { return isTerminator(getOpcode()); } |
130 | bool isBinaryOp() const { return isBinaryOp(getOpcode()); } |
131 | bool isShift() { return isShift(getOpcode()); } |
132 | bool isCast() const { return isCast(getOpcode()); } |
133 | bool isFuncletPad() const { return isFuncletPad(getOpcode()); } |
134 | |
135 | static const char* getOpcodeName(unsigned OpCode); |
136 | |
137 | static inline bool isTerminator(unsigned OpCode) { |
138 | return OpCode >= TermOpsBegin && OpCode < TermOpsEnd; |
139 | } |
140 | |
141 | static inline bool isBinaryOp(unsigned Opcode) { |
142 | return Opcode >= BinaryOpsBegin && Opcode < BinaryOpsEnd; |
143 | } |
144 | |
145 | /// Determine if the Opcode is one of the shift instructions. |
146 | static inline bool isShift(unsigned Opcode) { |
147 | return Opcode >= Shl && Opcode <= AShr; |
148 | } |
149 | |
150 | /// Return true if this is a logical shift left or a logical shift right. |
151 | inline bool isLogicalShift() const { |
152 | return getOpcode() == Shl || getOpcode() == LShr; |
153 | } |
154 | |
155 | /// Return true if this is an arithmetic shift right. |
156 | inline bool isArithmeticShift() const { |
157 | return getOpcode() == AShr; |
158 | } |
159 | |
160 | /// Determine if the Opcode is and/or/xor. |
161 | static inline bool isBitwiseLogicOp(unsigned Opcode) { |
162 | return Opcode == And || Opcode == Or || Opcode == Xor; |
163 | } |
164 | |
165 | /// Return true if this is and/or/xor. |
166 | inline bool isBitwiseLogicOp() const { |
167 | return isBitwiseLogicOp(getOpcode()); |
168 | } |
169 | |
170 | /// Determine if the OpCode is one of the CastInst instructions. |
171 | static inline bool isCast(unsigned OpCode) { |
172 | return OpCode >= CastOpsBegin && OpCode < CastOpsEnd; |
173 | } |
174 | |
175 | /// Determine if the OpCode is one of the FuncletPadInst instructions. |
176 | static inline bool isFuncletPad(unsigned OpCode) { |
177 | return OpCode >= FuncletPadOpsBegin && OpCode < FuncletPadOpsEnd; |
178 | } |
179 | |
180 | //===--------------------------------------------------------------------===// |
181 | // Metadata manipulation. |
182 | //===--------------------------------------------------------------------===// |
183 | |
184 | /// Return true if this instruction has any metadata attached to it. |
185 | bool hasMetadata() const { return DbgLoc || hasMetadataHashEntry(); } |
186 | |
187 | /// Return true if this instruction has metadata attached to it other than a |
188 | /// debug location. |
189 | bool hasMetadataOtherThanDebugLoc() const { |
190 | return hasMetadataHashEntry(); |
191 | } |
192 | |
193 | /// Get the metadata of given kind attached to this Instruction. |
194 | /// If the metadata is not found then return null. |
195 | MDNode *getMetadata(unsigned KindID) const { |
196 | if (!hasMetadata()) return nullptr; |
197 | return getMetadataImpl(KindID); |
198 | } |
199 | |
200 | /// Get the metadata of given kind attached to this Instruction. |
201 | /// If the metadata is not found then return null. |
202 | MDNode *getMetadata(StringRef Kind) const { |
203 | if (!hasMetadata()) return nullptr; |
204 | return getMetadataImpl(Kind); |
205 | } |
206 | |
207 | /// Get all metadata attached to this Instruction. The first element of each |
208 | /// pair returned is the KindID, the second element is the metadata value. |
209 | /// This list is returned sorted by the KindID. |
210 | void |
211 | getAllMetadata(SmallVectorImpl<std::pair<unsigned, MDNode *>> &MDs) const { |
212 | if (hasMetadata()) |
213 | getAllMetadataImpl(MDs); |
214 | } |
215 | |
216 | /// This does the same thing as getAllMetadata, except that it filters out the |
217 | /// debug location. |
218 | void getAllMetadataOtherThanDebugLoc( |
219 | SmallVectorImpl<std::pair<unsigned, MDNode *>> &MDs) const { |
220 | if (hasMetadataOtherThanDebugLoc()) |
221 | getAllMetadataOtherThanDebugLocImpl(MDs); |
222 | } |
223 | |
224 | /// Fills the AAMDNodes structure with AA metadata from this instruction. |
225 | /// When Merge is true, the existing AA metadata is merged with that from this |
226 | /// instruction providing the most-general result. |
227 | void getAAMetadata(AAMDNodes &N, bool Merge = false) const; |
228 | |
229 | /// Set the metadata of the specified kind to the specified node. This updates |
230 | /// or replaces metadata if already present, or removes it if Node is null. |
231 | void setMetadata(unsigned KindID, MDNode *Node); |
232 | void setMetadata(StringRef Kind, MDNode *Node); |
233 | |
234 | /// Copy metadata from \p SrcInst to this instruction. \p WL, if not empty, |
235 | /// specifies the list of meta data that needs to be copied. If \p WL is |
236 | /// empty, all meta data will be copied. |
237 | void copyMetadata(const Instruction &SrcInst, |
238 | ArrayRef<unsigned> WL = ArrayRef<unsigned>()); |
239 | |
240 | /// If the instruction has "branch_weights" MD_prof metadata and the MDNode |
241 | /// has three operands (including name string), swap the order of the |
242 | /// metadata. |
243 | void swapProfMetadata(); |
244 | |
245 | /// Drop all unknown metadata except for debug locations. |
246 | /// @{ |
247 | /// Passes are required to drop metadata they don't understand. This is a |
248 | /// convenience method for passes to do so. |
249 | void dropUnknownNonDebugMetadata(ArrayRef<unsigned> KnownIDs); |
250 | void dropUnknownNonDebugMetadata() { |
251 | return dropUnknownNonDebugMetadata(None); |
252 | } |
253 | void dropUnknownNonDebugMetadata(unsigned ID1) { |
254 | return dropUnknownNonDebugMetadata(makeArrayRef(ID1)); |
255 | } |
256 | void dropUnknownNonDebugMetadata(unsigned ID1, unsigned ID2) { |
257 | unsigned IDs[] = {ID1, ID2}; |
258 | return dropUnknownNonDebugMetadata(IDs); |
259 | } |
260 | /// @} |
261 | |
262 | /// Sets the metadata on this instruction from the AAMDNodes structure. |
263 | void setAAMetadata(const AAMDNodes &N); |
264 | |
265 | /// Retrieve the raw weight values of a conditional branch or select. |
266 | /// Returns true on success with profile weights filled in. |
267 | /// Returns false if no metadata or invalid metadata was found. |
268 | bool extractProfMetadata(uint64_t &TrueVal, uint64_t &FalseVal) const; |
269 | |
270 | /// Retrieve total raw weight values of a branch. |
271 | /// Returns true on success with profile total weights filled in. |
272 | /// Returns false if no metadata was found. |
273 | bool extractProfTotalWeight(uint64_t &TotalVal) const; |
274 | |
275 | /// Updates branch_weights metadata by scaling it by \p S / \p T. |
276 | void updateProfWeight(uint64_t S, uint64_t T); |
277 | |
278 | /// Sets the branch_weights metadata to \p W for CallInst. |
279 | void setProfWeight(uint64_t W); |
280 | |
281 | /// Set the debug location information for this instruction. |
282 | void setDebugLoc(DebugLoc Loc) { DbgLoc = std::move(Loc); } |
283 | |
284 | /// Return the debug location for this node as a DebugLoc. |
285 | const DebugLoc &getDebugLoc() const { return DbgLoc; } |
286 | |
287 | /// Set or clear the nsw flag on this instruction, which must be an operator |
288 | /// which supports this flag. See LangRef.html for the meaning of this flag. |
289 | void setHasNoUnsignedWrap(bool b = true); |
290 | |
291 | /// Set or clear the nsw flag on this instruction, which must be an operator |
292 | /// which supports this flag. See LangRef.html for the meaning of this flag. |
293 | void setHasNoSignedWrap(bool b = true); |
294 | |
295 | /// Set or clear the exact flag on this instruction, which must be an operator |
296 | /// which supports this flag. See LangRef.html for the meaning of this flag. |
297 | void setIsExact(bool b = true); |
298 | |
299 | /// Determine whether the no unsigned wrap flag is set. |
300 | bool hasNoUnsignedWrap() const; |
301 | |
302 | /// Determine whether the no signed wrap flag is set. |
303 | bool hasNoSignedWrap() const; |
304 | |
305 | /// Drops flags that may cause this instruction to evaluate to poison despite |
306 | /// having non-poison inputs. |
307 | void dropPoisonGeneratingFlags(); |
308 | |
309 | /// Determine whether the exact flag is set. |
310 | bool isExact() const; |
311 | |
312 | /// Set or clear all fast-math-flags on this instruction, which must be an |
313 | /// operator which supports this flag. See LangRef.html for the meaning of |
314 | /// this flag. |
315 | void setFast(bool B); |
316 | |
317 | /// Set or clear the reassociation flag on this instruction, which must be |
318 | /// an operator which supports this flag. See LangRef.html for the meaning of |
319 | /// this flag. |
320 | void setHasAllowReassoc(bool B); |
321 | |
322 | /// Set or clear the no-nans flag on this instruction, which must be an |
323 | /// operator which supports this flag. See LangRef.html for the meaning of |
324 | /// this flag. |
325 | void setHasNoNaNs(bool B); |
326 | |
327 | /// Set or clear the no-infs flag on this instruction, which must be an |
328 | /// operator which supports this flag. See LangRef.html for the meaning of |
329 | /// this flag. |
330 | void setHasNoInfs(bool B); |
331 | |
332 | /// Set or clear the no-signed-zeros flag on this instruction, which must be |
333 | /// an operator which supports this flag. See LangRef.html for the meaning of |
334 | /// this flag. |
335 | void setHasNoSignedZeros(bool B); |
336 | |
337 | /// Set or clear the allow-reciprocal flag on this instruction, which must be |
338 | /// an operator which supports this flag. See LangRef.html for the meaning of |
339 | /// this flag. |
340 | void setHasAllowReciprocal(bool B); |
341 | |
342 | /// Set or clear the approximate-math-functions flag on this instruction, |
343 | /// which must be an operator which supports this flag. See LangRef.html for |
344 | /// the meaning of this flag. |
345 | void setHasApproxFunc(bool B); |
346 | |
347 | /// Convenience function for setting multiple fast-math flags on this |
348 | /// instruction, which must be an operator which supports these flags. See |
349 | /// LangRef.html for the meaning of these flags. |
350 | void setFastMathFlags(FastMathFlags FMF); |
351 | |
352 | /// Convenience function for transferring all fast-math flag values to this |
353 | /// instruction, which must be an operator which supports these flags. See |
354 | /// LangRef.html for the meaning of these flags. |
355 | void copyFastMathFlags(FastMathFlags FMF); |
356 | |
357 | /// Determine whether all fast-math-flags are set. |
358 | bool isFast() const; |
359 | |
360 | /// Determine whether the allow-reassociation flag is set. |
361 | bool hasAllowReassoc() const; |
362 | |
363 | /// Determine whether the no-NaNs flag is set. |
364 | bool hasNoNaNs() const; |
365 | |
366 | /// Determine whether the no-infs flag is set. |
367 | bool hasNoInfs() const; |
368 | |
369 | /// Determine whether the no-signed-zeros flag is set. |
370 | bool hasNoSignedZeros() const; |
371 | |
372 | /// Determine whether the allow-reciprocal flag is set. |
373 | bool hasAllowReciprocal() const; |
374 | |
375 | /// Determine whether the allow-contract flag is set. |
376 | bool hasAllowContract() const; |
377 | |
378 | /// Determine whether the approximate-math-functions flag is set. |
379 | bool hasApproxFunc() const; |
380 | |
381 | /// Convenience function for getting all the fast-math flags, which must be an |
382 | /// operator which supports these flags. See LangRef.html for the meaning of |
383 | /// these flags. |
384 | FastMathFlags getFastMathFlags() const; |
385 | |
386 | /// Copy I's fast-math flags |
387 | void copyFastMathFlags(const Instruction *I); |
388 | |
389 | /// Convenience method to copy supported exact, fast-math, and (optionally) |
390 | /// wrapping flags from V to this instruction. |
391 | void copyIRFlags(const Value *V, bool IncludeWrapFlags = true); |
392 | |
393 | /// Logical 'and' of any supported wrapping, exact, and fast-math flags of |
394 | /// V and this instruction. |
395 | void andIRFlags(const Value *V); |
396 | |
397 | /// Merge 2 debug locations and apply it to the Instruction. If the |
398 | /// instruction is a CallIns, we need to traverse the inline chain to find |
399 | /// the common scope. This is not efficient for N-way merging as each time |
400 | /// you merge 2 iterations, you need to rebuild the hashmap to find the |
401 | /// common scope. However, we still choose this API because: |
402 | /// 1) Simplicity: it takes 2 locations instead of a list of locations. |
403 | /// 2) In worst case, it increases the complexity from O(N*I) to |
404 | /// O(2*N*I), where N is # of Instructions to merge, and I is the |
405 | /// maximum level of inline stack. So it is still linear. |
406 | /// 3) Merging of call instructions should be extremely rare in real |
407 | /// applications, thus the N-way merging should be in code path. |
408 | /// The DebugLoc attached to this instruction will be overwritten by the |
409 | /// merged DebugLoc. |
410 | void applyMergedLocation(const DILocation *LocA, const DILocation *LocB); |
411 | |
412 | private: |
413 | /// Return true if we have an entry in the on-the-side metadata hash. |
414 | bool hasMetadataHashEntry() const { |
415 | return (getSubclassDataFromValue() & HasMetadataBit) != 0; |
416 | } |
417 | |
418 | // These are all implemented in Metadata.cpp. |
419 | MDNode *getMetadataImpl(unsigned KindID) const; |
420 | MDNode *getMetadataImpl(StringRef Kind) const; |
421 | void |
422 | getAllMetadataImpl(SmallVectorImpl<std::pair<unsigned, MDNode *>> &) const; |
423 | void getAllMetadataOtherThanDebugLocImpl( |
424 | SmallVectorImpl<std::pair<unsigned, MDNode *>> &) const; |
425 | /// Clear all hashtable-based metadata from this instruction. |
426 | void clearMetadataHashEntries(); |
427 | |
428 | public: |
429 | //===--------------------------------------------------------------------===// |
430 | // Predicates and helper methods. |
431 | //===--------------------------------------------------------------------===// |
432 | |
433 | /// Return true if the instruction is associative: |
434 | /// |
435 | /// Associative operators satisfy: x op (y op z) === (x op y) op z |
436 | /// |
437 | /// In LLVM, the Add, Mul, And, Or, and Xor operators are associative. |
438 | /// |
439 | bool isAssociative() const LLVM_READONLY__attribute__((__pure__)); |
440 | static bool isAssociative(unsigned Opcode) { |
441 | return Opcode == And || Opcode == Or || Opcode == Xor || |
442 | Opcode == Add || Opcode == Mul; |
443 | } |
444 | |
445 | /// Return true if the instruction is commutative: |
446 | /// |
447 | /// Commutative operators satisfy: (x op y) === (y op x) |
448 | /// |
449 | /// In LLVM, these are the commutative operators, plus SetEQ and SetNE, when |
450 | /// applied to any type. |
451 | /// |
452 | bool isCommutative() const { return isCommutative(getOpcode()); } |
453 | static bool isCommutative(unsigned Opcode) { |
454 | switch (Opcode) { |
455 | case Add: case FAdd: |
456 | case Mul: case FMul: |
457 | case And: case Or: case Xor: |
458 | return true; |
459 | default: |
460 | return false; |
461 | } |
462 | } |
463 | |
464 | /// Return true if the instruction is idempotent: |
465 | /// |
466 | /// Idempotent operators satisfy: x op x === x |
467 | /// |
468 | /// In LLVM, the And and Or operators are idempotent. |
469 | /// |
470 | bool isIdempotent() const { return isIdempotent(getOpcode()); } |
471 | static bool isIdempotent(unsigned Opcode) { |
472 | return Opcode == And || Opcode == Or; |
473 | } |
474 | |
475 | /// Return true if the instruction is nilpotent: |
476 | /// |
477 | /// Nilpotent operators satisfy: x op x === Id, |
478 | /// |
479 | /// where Id is the identity for the operator, i.e. a constant such that |
480 | /// x op Id === x and Id op x === x for all x. |
481 | /// |
482 | /// In LLVM, the Xor operator is nilpotent. |
483 | /// |
484 | bool isNilpotent() const { return isNilpotent(getOpcode()); } |
485 | static bool isNilpotent(unsigned Opcode) { |
486 | return Opcode == Xor; |
487 | } |
488 | |
489 | /// Return true if this instruction may modify memory. |
490 | bool mayWriteToMemory() const; |
491 | |
492 | /// Return true if this instruction may read memory. |
493 | bool mayReadFromMemory() const; |
494 | |
495 | /// Return true if this instruction may read or write memory. |
496 | bool mayReadOrWriteMemory() const { |
497 | return mayReadFromMemory() || mayWriteToMemory(); |
498 | } |
499 | |
500 | /// Return true if this instruction has an AtomicOrdering of unordered or |
501 | /// higher. |
502 | bool isAtomic() const; |
503 | |
504 | /// Return true if this atomic instruction loads from memory. |
505 | bool hasAtomicLoad() const; |
506 | |
507 | /// Return true if this atomic instruction stores to memory. |
508 | bool hasAtomicStore() const; |
509 | |
510 | /// Return true if this instruction may throw an exception. |
511 | bool mayThrow() const; |
512 | |
513 | /// Return true if this instruction behaves like a memory fence: it can load |
514 | /// or store to memory location without being given a memory location. |
515 | bool isFenceLike() const { |
516 | switch (getOpcode()) { |
517 | default: |
518 | return false; |
519 | // This list should be kept in sync with the list in mayWriteToMemory for |
520 | // all opcodes which don't have a memory location. |
521 | case Instruction::Fence: |
522 | case Instruction::CatchPad: |
523 | case Instruction::CatchRet: |
524 | case Instruction::Call: |
525 | case Instruction::Invoke: |
526 | return true; |
527 | } |
528 | } |
529 | |
530 | /// Return true if the instruction may have side effects. |
531 | /// |
532 | /// Note that this does not consider malloc and alloca to have side |
533 | /// effects because the newly allocated memory is completely invisible to |
534 | /// instructions which don't use the returned value. For cases where this |
535 | /// matters, isSafeToSpeculativelyExecute may be more appropriate. |
536 | bool mayHaveSideEffects() const { return mayWriteToMemory() || mayThrow(); } |
537 | |
538 | /// Return true if the instruction is a variety of EH-block. |
539 | bool isEHPad() const { |
540 | switch (getOpcode()) { |
541 | case Instruction::CatchSwitch: |
542 | case Instruction::CatchPad: |
543 | case Instruction::CleanupPad: |
544 | case Instruction::LandingPad: |
545 | return true; |
546 | default: |
547 | return false; |
548 | } |
549 | } |
550 | |
551 | /// Create a copy of 'this' instruction that is identical in all ways except |
552 | /// the following: |
553 | /// * The instruction has no parent |
554 | /// * The instruction has no name |
555 | /// |
556 | Instruction *clone() const; |
557 | |
558 | /// Return true if the specified instruction is exactly identical to the |
559 | /// current one. This means that all operands match and any extra information |
560 | /// (e.g. load is volatile) agree. |
561 | bool isIdenticalTo(const Instruction *I) const; |
562 | |
563 | /// This is like isIdenticalTo, except that it ignores the |
564 | /// SubclassOptionalData flags, which may specify conditions under which the |
565 | /// instruction's result is undefined. |
566 | bool isIdenticalToWhenDefined(const Instruction *I) const; |
567 | |
568 | /// When checking for operation equivalence (using isSameOperationAs) it is |
569 | /// sometimes useful to ignore certain attributes. |
570 | enum OperationEquivalenceFlags { |
571 | /// Check for equivalence ignoring load/store alignment. |
572 | CompareIgnoringAlignment = 1<<0, |
573 | /// Check for equivalence treating a type and a vector of that type |
574 | /// as equivalent. |
575 | CompareUsingScalarTypes = 1<<1 |
576 | }; |
577 | |
578 | /// This function determines if the specified instruction executes the same |
579 | /// operation as the current one. This means that the opcodes, type, operand |
580 | /// types and any other factors affecting the operation must be the same. This |
581 | /// is similar to isIdenticalTo except the operands themselves don't have to |
582 | /// be identical. |
583 | /// @returns true if the specified instruction is the same operation as |
584 | /// the current one. |
585 | /// @brief Determine if one instruction is the same operation as another. |
586 | bool isSameOperationAs(const Instruction *I, unsigned flags = 0) const; |
587 | |
588 | /// Return true if there are any uses of this instruction in blocks other than |
589 | /// the specified block. Note that PHI nodes are considered to evaluate their |
590 | /// operands in the corresponding predecessor block. |
591 | bool isUsedOutsideOfBlock(const BasicBlock *BB) const; |
592 | |
593 | |
594 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
595 | static bool classof(const Value *V) { |
596 | return V->getValueID() >= Value::InstructionVal; |
597 | } |
598 | |
599 | //---------------------------------------------------------------------- |
600 | // Exported enumerations. |
601 | // |
602 | enum TermOps { // These terminate basic blocks |
603 | #define FIRST_TERM_INST(N) TermOpsBegin = N, |
604 | #define HANDLE_TERM_INST(N, OPC, CLASS) OPC = N, |
605 | #define LAST_TERM_INST(N) TermOpsEnd = N+1 |
606 | #include "llvm/IR/Instruction.def" |
607 | }; |
608 | |
609 | enum BinaryOps { |
610 | #define FIRST_BINARY_INST(N) BinaryOpsBegin = N, |
611 | #define HANDLE_BINARY_INST(N, OPC, CLASS) OPC = N, |
612 | #define LAST_BINARY_INST(N) BinaryOpsEnd = N+1 |
613 | #include "llvm/IR/Instruction.def" |
614 | }; |
615 | |
616 | enum MemoryOps { |
617 | #define FIRST_MEMORY_INST(N) MemoryOpsBegin = N, |
618 | #define HANDLE_MEMORY_INST(N, OPC, CLASS) OPC = N, |
619 | #define LAST_MEMORY_INST(N) MemoryOpsEnd = N+1 |
620 | #include "llvm/IR/Instruction.def" |
621 | }; |
622 | |
623 | enum CastOps { |
624 | #define FIRST_CAST_INST(N) CastOpsBegin = N, |
625 | #define HANDLE_CAST_INST(N, OPC, CLASS) OPC = N, |
626 | #define LAST_CAST_INST(N) CastOpsEnd = N+1 |
627 | #include "llvm/IR/Instruction.def" |
628 | }; |
629 | |
630 | enum FuncletPadOps { |
631 | #define FIRST_FUNCLETPAD_INST(N) FuncletPadOpsBegin = N, |
632 | #define HANDLE_FUNCLETPAD_INST(N, OPC, CLASS) OPC = N, |
633 | #define LAST_FUNCLETPAD_INST(N) FuncletPadOpsEnd = N+1 |
634 | #include "llvm/IR/Instruction.def" |
635 | }; |
636 | |
637 | enum OtherOps { |
638 | #define FIRST_OTHER_INST(N) OtherOpsBegin = N, |
639 | #define HANDLE_OTHER_INST(N, OPC, CLASS) OPC = N, |
640 | #define LAST_OTHER_INST(N) OtherOpsEnd = N+1 |
641 | #include "llvm/IR/Instruction.def" |
642 | }; |
643 | |
644 | private: |
645 | friend class SymbolTableListTraits<Instruction>; |
646 | |
647 | // Shadow Value::setValueSubclassData with a private forwarding method so that |
648 | // subclasses cannot accidentally use it. |
649 | void setValueSubclassData(unsigned short D) { |
650 | Value::setValueSubclassData(D); |
651 | } |
652 | |
653 | unsigned short getSubclassDataFromValue() const { |
654 | return Value::getSubclassDataFromValue(); |
655 | } |
656 | |
657 | void setHasMetadataHashEntry(bool V) { |
658 | setValueSubclassData((getSubclassDataFromValue() & ~HasMetadataBit) | |
659 | (V ? HasMetadataBit : 0)); |
660 | } |
661 | |
662 | void setParent(BasicBlock *P); |
663 | |
664 | protected: |
665 | // Instruction subclasses can stick up to 15 bits of stuff into the |
666 | // SubclassData field of instruction with these members. |
667 | |
668 | // Verify that only the low 15 bits are used. |
669 | void setInstructionSubclassData(unsigned short D) { |
670 | assert((D & HasMetadataBit) == 0 && "Out of range value put into field")(static_cast <bool> ((D & HasMetadataBit) == 0 && "Out of range value put into field") ? void (0) : __assert_fail ("(D & HasMetadataBit) == 0 && \"Out of range value put into field\"" , "/build/llvm-toolchain-snapshot-6.0~svn321639/include/llvm/IR/Instruction.h" , 670, __extension__ __PRETTY_FUNCTION__)); |
671 | setValueSubclassData((getSubclassDataFromValue() & HasMetadataBit) | D); |
672 | } |
673 | |
674 | unsigned getSubclassDataFromInstruction() const { |
675 | return getSubclassDataFromValue() & ~HasMetadataBit; |
676 | } |
677 | |
678 | Instruction(Type *Ty, unsigned iType, Use *Ops, unsigned NumOps, |
679 | Instruction *InsertBefore = nullptr); |
680 | Instruction(Type *Ty, unsigned iType, Use *Ops, unsigned NumOps, |
681 | BasicBlock *InsertAtEnd); |
682 | |
683 | private: |
684 | /// Create a copy of this instruction. |
685 | Instruction *cloneImpl() const; |
686 | }; |
687 | |
688 | inline void ilist_alloc_traits<Instruction>::deleteNode(Instruction *V) { |
689 | V->deleteValue(); |
690 | } |
691 | |
692 | } // end namespace llvm |
693 | |
694 | #endif // LLVM_IR_INSTRUCTION_H |