]>
Commit | Line | Data |
---|---|---|
223e47cc LB |
1 | //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// |
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 transformation implements the well known scalar replacement of | |
11 | // aggregates transformation. This xform breaks up alloca instructions of | |
12 | // aggregate type (structure or array) into individual alloca instructions for | |
13 | // each member (if possible). Then, if possible, it transforms the individual | |
14 | // alloca instructions into nice clean scalar SSA form. | |
15 | // | |
16 | // This combines a simple SRoA algorithm with the Mem2Reg algorithm because they | |
17 | // often interact, especially for C++ programs. As such, iterating between | |
18 | // SRoA, then Mem2Reg until we run out of things to promote works well. | |
19 | // | |
20 | //===----------------------------------------------------------------------===// | |
21 | ||
22 | #define DEBUG_TYPE "scalarrepl" | |
23 | #include "llvm/Transforms/Scalar.h" | |
223e47cc LB |
24 | #include "llvm/ADT/SetVector.h" |
25 | #include "llvm/ADT/SmallVector.h" | |
26 | #include "llvm/ADT/Statistic.h" | |
27 | #include "llvm/Analysis/Dominators.h" | |
28 | #include "llvm/Analysis/Loads.h" | |
29 | #include "llvm/Analysis/ValueTracking.h" | |
970d7e83 LB |
30 | #include "llvm/DIBuilder.h" |
31 | #include "llvm/DebugInfo.h" | |
32 | #include "llvm/IR/Constants.h" | |
33 | #include "llvm/IR/DataLayout.h" | |
34 | #include "llvm/IR/DerivedTypes.h" | |
35 | #include "llvm/IR/Function.h" | |
36 | #include "llvm/IR/GlobalVariable.h" | |
37 | #include "llvm/IR/IRBuilder.h" | |
38 | #include "llvm/IR/Instructions.h" | |
39 | #include "llvm/IR/IntrinsicInst.h" | |
40 | #include "llvm/IR/LLVMContext.h" | |
41 | #include "llvm/IR/Module.h" | |
42 | #include "llvm/IR/Operator.h" | |
43 | #include "llvm/Pass.h" | |
223e47cc LB |
44 | #include "llvm/Support/CallSite.h" |
45 | #include "llvm/Support/Debug.h" | |
46 | #include "llvm/Support/ErrorHandling.h" | |
47 | #include "llvm/Support/GetElementPtrTypeIterator.h" | |
48 | #include "llvm/Support/MathExtras.h" | |
49 | #include "llvm/Support/raw_ostream.h" | |
223e47cc LB |
50 | #include "llvm/Transforms/Utils/Local.h" |
51 | #include "llvm/Transforms/Utils/PromoteMemToReg.h" | |
52 | #include "llvm/Transforms/Utils/SSAUpdater.h" | |
53 | using namespace llvm; | |
54 | ||
55 | STATISTIC(NumReplaced, "Number of allocas broken up"); | |
56 | STATISTIC(NumPromoted, "Number of allocas promoted"); | |
57 | STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); | |
58 | STATISTIC(NumConverted, "Number of aggregates converted to scalar"); | |
59 | ||
60 | namespace { | |
61 | struct SROA : public FunctionPass { | |
62 | SROA(int T, bool hasDT, char &ID, int ST, int AT, int SLT) | |
63 | : FunctionPass(ID), HasDomTree(hasDT) { | |
64 | if (T == -1) | |
65 | SRThreshold = 128; | |
66 | else | |
67 | SRThreshold = T; | |
68 | if (ST == -1) | |
69 | StructMemberThreshold = 32; | |
70 | else | |
71 | StructMemberThreshold = ST; | |
72 | if (AT == -1) | |
73 | ArrayElementThreshold = 8; | |
74 | else | |
75 | ArrayElementThreshold = AT; | |
76 | if (SLT == -1) | |
77 | // Do not limit the scalar integer load size if no threshold is given. | |
78 | ScalarLoadThreshold = -1; | |
79 | else | |
80 | ScalarLoadThreshold = SLT; | |
81 | } | |
82 | ||
83 | bool runOnFunction(Function &F); | |
84 | ||
85 | bool performScalarRepl(Function &F); | |
86 | bool performPromotion(Function &F); | |
87 | ||
88 | private: | |
89 | bool HasDomTree; | |
970d7e83 | 90 | DataLayout *TD; |
223e47cc LB |
91 | |
92 | /// DeadInsts - Keep track of instructions we have made dead, so that | |
93 | /// we can remove them after we are done working. | |
94 | SmallVector<Value*, 32> DeadInsts; | |
95 | ||
96 | /// AllocaInfo - When analyzing uses of an alloca instruction, this captures | |
97 | /// information about the uses. All these fields are initialized to false | |
98 | /// and set to true when something is learned. | |
99 | struct AllocaInfo { | |
100 | /// The alloca to promote. | |
101 | AllocaInst *AI; | |
102 | ||
103 | /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite | |
104 | /// looping and avoid redundant work. | |
105 | SmallPtrSet<PHINode*, 8> CheckedPHIs; | |
106 | ||
107 | /// isUnsafe - This is set to true if the alloca cannot be SROA'd. | |
108 | bool isUnsafe : 1; | |
109 | ||
110 | /// isMemCpySrc - This is true if this aggregate is memcpy'd from. | |
111 | bool isMemCpySrc : 1; | |
112 | ||
113 | /// isMemCpyDst - This is true if this aggregate is memcpy'd into. | |
114 | bool isMemCpyDst : 1; | |
115 | ||
116 | /// hasSubelementAccess - This is true if a subelement of the alloca is | |
117 | /// ever accessed, or false if the alloca is only accessed with mem | |
118 | /// intrinsics or load/store that only access the entire alloca at once. | |
119 | bool hasSubelementAccess : 1; | |
120 | ||
121 | /// hasALoadOrStore - This is true if there are any loads or stores to it. | |
122 | /// The alloca may just be accessed with memcpy, for example, which would | |
123 | /// not set this. | |
124 | bool hasALoadOrStore : 1; | |
125 | ||
126 | explicit AllocaInfo(AllocaInst *ai) | |
127 | : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), | |
128 | hasSubelementAccess(false), hasALoadOrStore(false) {} | |
129 | }; | |
130 | ||
131 | /// SRThreshold - The maximum alloca size to considered for SROA. | |
132 | unsigned SRThreshold; | |
133 | ||
134 | /// StructMemberThreshold - The maximum number of members a struct can | |
135 | /// contain to be considered for SROA. | |
136 | unsigned StructMemberThreshold; | |
137 | ||
138 | /// ArrayElementThreshold - The maximum number of elements an array can | |
139 | /// have to be considered for SROA. | |
140 | unsigned ArrayElementThreshold; | |
141 | ||
142 | /// ScalarLoadThreshold - The maximum size in bits of scalars to load when | |
143 | /// converting to scalar | |
144 | unsigned ScalarLoadThreshold; | |
145 | ||
146 | void MarkUnsafe(AllocaInfo &I, Instruction *User) { | |
147 | I.isUnsafe = true; | |
148 | DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n'); | |
149 | } | |
150 | ||
151 | bool isSafeAllocaToScalarRepl(AllocaInst *AI); | |
152 | ||
153 | void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info); | |
154 | void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset, | |
155 | AllocaInfo &Info); | |
156 | void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info); | |
157 | void isSafeMemAccess(uint64_t Offset, uint64_t MemSize, | |
158 | Type *MemOpType, bool isStore, AllocaInfo &Info, | |
159 | Instruction *TheAccess, bool AllowWholeAccess); | |
160 | bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size); | |
161 | uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset, | |
162 | Type *&IdxTy); | |
163 | ||
164 | void DoScalarReplacement(AllocaInst *AI, | |
165 | std::vector<AllocaInst*> &WorkList); | |
166 | void DeleteDeadInstructions(); | |
167 | ||
168 | void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, | |
169 | SmallVector<AllocaInst*, 32> &NewElts); | |
170 | void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, | |
171 | SmallVector<AllocaInst*, 32> &NewElts); | |
172 | void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, | |
173 | SmallVector<AllocaInst*, 32> &NewElts); | |
174 | void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, | |
175 | uint64_t Offset, | |
176 | SmallVector<AllocaInst*, 32> &NewElts); | |
177 | void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, | |
178 | AllocaInst *AI, | |
179 | SmallVector<AllocaInst*, 32> &NewElts); | |
180 | void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, | |
181 | SmallVector<AllocaInst*, 32> &NewElts); | |
182 | void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, | |
183 | SmallVector<AllocaInst*, 32> &NewElts); | |
184 | bool ShouldAttemptScalarRepl(AllocaInst *AI); | |
185 | }; | |
186 | ||
187 | // SROA_DT - SROA that uses DominatorTree. | |
188 | struct SROA_DT : public SROA { | |
189 | static char ID; | |
190 | public: | |
191 | SROA_DT(int T = -1, int ST = -1, int AT = -1, int SLT = -1) : | |
192 | SROA(T, true, ID, ST, AT, SLT) { | |
193 | initializeSROA_DTPass(*PassRegistry::getPassRegistry()); | |
194 | } | |
195 | ||
196 | // getAnalysisUsage - This pass does not require any passes, but we know it | |
197 | // will not alter the CFG, so say so. | |
198 | virtual void getAnalysisUsage(AnalysisUsage &AU) const { | |
199 | AU.addRequired<DominatorTree>(); | |
200 | AU.setPreservesCFG(); | |
201 | } | |
202 | }; | |
203 | ||
204 | // SROA_SSAUp - SROA that uses SSAUpdater. | |
205 | struct SROA_SSAUp : public SROA { | |
206 | static char ID; | |
207 | public: | |
208 | SROA_SSAUp(int T = -1, int ST = -1, int AT = -1, int SLT = -1) : | |
209 | SROA(T, false, ID, ST, AT, SLT) { | |
210 | initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); | |
211 | } | |
212 | ||
213 | // getAnalysisUsage - This pass does not require any passes, but we know it | |
214 | // will not alter the CFG, so say so. | |
215 | virtual void getAnalysisUsage(AnalysisUsage &AU) const { | |
216 | AU.setPreservesCFG(); | |
217 | } | |
218 | }; | |
219 | ||
220 | } | |
221 | ||
222 | char SROA_DT::ID = 0; | |
223 | char SROA_SSAUp::ID = 0; | |
224 | ||
225 | INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", | |
226 | "Scalar Replacement of Aggregates (DT)", false, false) | |
227 | INITIALIZE_PASS_DEPENDENCY(DominatorTree) | |
228 | INITIALIZE_PASS_END(SROA_DT, "scalarrepl", | |
229 | "Scalar Replacement of Aggregates (DT)", false, false) | |
230 | ||
231 | INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", | |
232 | "Scalar Replacement of Aggregates (SSAUp)", false, false) | |
233 | INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", | |
234 | "Scalar Replacement of Aggregates (SSAUp)", false, false) | |
235 | ||
236 | // Public interface to the ScalarReplAggregates pass | |
237 | FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, | |
238 | bool UseDomTree, | |
239 | int StructMemberThreshold, | |
240 | int ArrayElementThreshold, | |
241 | int ScalarLoadThreshold) { | |
242 | if (UseDomTree) | |
243 | return new SROA_DT(Threshold, StructMemberThreshold, ArrayElementThreshold, | |
244 | ScalarLoadThreshold); | |
245 | return new SROA_SSAUp(Threshold, StructMemberThreshold, | |
246 | ArrayElementThreshold, ScalarLoadThreshold); | |
247 | } | |
248 | ||
249 | ||
250 | //===----------------------------------------------------------------------===// | |
251 | // Convert To Scalar Optimization. | |
252 | //===----------------------------------------------------------------------===// | |
253 | ||
254 | namespace { | |
255 | /// ConvertToScalarInfo - This class implements the "Convert To Scalar" | |
256 | /// optimization, which scans the uses of an alloca and determines if it can | |
257 | /// rewrite it in terms of a single new alloca that can be mem2reg'd. | |
258 | class ConvertToScalarInfo { | |
259 | /// AllocaSize - The size of the alloca being considered in bytes. | |
260 | unsigned AllocaSize; | |
970d7e83 | 261 | const DataLayout &TD; |
223e47cc LB |
262 | unsigned ScalarLoadThreshold; |
263 | ||
264 | /// IsNotTrivial - This is set to true if there is some access to the object | |
265 | /// which means that mem2reg can't promote it. | |
266 | bool IsNotTrivial; | |
267 | ||
268 | /// ScalarKind - Tracks the kind of alloca being considered for promotion, | |
269 | /// computed based on the uses of the alloca rather than the LLVM type system. | |
270 | enum { | |
271 | Unknown, | |
272 | ||
273 | // Accesses via GEPs that are consistent with element access of a vector | |
274 | // type. This will not be converted into a vector unless there is a later | |
275 | // access using an actual vector type. | |
276 | ImplicitVector, | |
277 | ||
278 | // Accesses via vector operations and GEPs that are consistent with the | |
279 | // layout of a vector type. | |
280 | Vector, | |
281 | ||
282 | // An integer bag-of-bits with bitwise operations for insertion and | |
283 | // extraction. Any combination of types can be converted into this kind | |
284 | // of scalar. | |
285 | Integer | |
286 | } ScalarKind; | |
287 | ||
288 | /// VectorTy - This tracks the type that we should promote the vector to if | |
289 | /// it is possible to turn it into a vector. This starts out null, and if it | |
290 | /// isn't possible to turn into a vector type, it gets set to VoidTy. | |
291 | VectorType *VectorTy; | |
292 | ||
293 | /// HadNonMemTransferAccess - True if there is at least one access to the | |
294 | /// alloca that is not a MemTransferInst. We don't want to turn structs into | |
295 | /// large integers unless there is some potential for optimization. | |
296 | bool HadNonMemTransferAccess; | |
297 | ||
298 | /// HadDynamicAccess - True if some element of this alloca was dynamic. | |
299 | /// We don't yet have support for turning a dynamic access into a large | |
300 | /// integer. | |
301 | bool HadDynamicAccess; | |
302 | ||
303 | public: | |
970d7e83 | 304 | explicit ConvertToScalarInfo(unsigned Size, const DataLayout &td, |
223e47cc LB |
305 | unsigned SLT) |
306 | : AllocaSize(Size), TD(td), ScalarLoadThreshold(SLT), IsNotTrivial(false), | |
307 | ScalarKind(Unknown), VectorTy(0), HadNonMemTransferAccess(false), | |
308 | HadDynamicAccess(false) { } | |
309 | ||
310 | AllocaInst *TryConvert(AllocaInst *AI); | |
311 | ||
312 | private: | |
313 | bool CanConvertToScalar(Value *V, uint64_t Offset, Value* NonConstantIdx); | |
314 | void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset); | |
315 | bool MergeInVectorType(VectorType *VInTy, uint64_t Offset); | |
316 | void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset, | |
317 | Value *NonConstantIdx); | |
318 | ||
319 | Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType, | |
320 | uint64_t Offset, Value* NonConstantIdx, | |
321 | IRBuilder<> &Builder); | |
322 | Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, | |
323 | uint64_t Offset, Value* NonConstantIdx, | |
324 | IRBuilder<> &Builder); | |
325 | }; | |
326 | } // end anonymous namespace. | |
327 | ||
328 | ||
329 | /// TryConvert - Analyze the specified alloca, and if it is safe to do so, | |
330 | /// rewrite it to be a new alloca which is mem2reg'able. This returns the new | |
331 | /// alloca if possible or null if not. | |
332 | AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { | |
333 | // If we can't convert this scalar, or if mem2reg can trivially do it, bail | |
334 | // out. | |
335 | if (!CanConvertToScalar(AI, 0, 0) || !IsNotTrivial) | |
336 | return 0; | |
337 | ||
338 | // If an alloca has only memset / memcpy uses, it may still have an Unknown | |
339 | // ScalarKind. Treat it as an Integer below. | |
340 | if (ScalarKind == Unknown) | |
341 | ScalarKind = Integer; | |
342 | ||
343 | if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8) | |
344 | ScalarKind = Integer; | |
345 | ||
346 | // If we were able to find a vector type that can handle this with | |
347 | // insert/extract elements, and if there was at least one use that had | |
348 | // a vector type, promote this to a vector. We don't want to promote | |
349 | // random stuff that doesn't use vectors (e.g. <9 x double>) because then | |
350 | // we just get a lot of insert/extracts. If at least one vector is | |
351 | // involved, then we probably really do have a union of vector/array. | |
352 | Type *NewTy; | |
353 | if (ScalarKind == Vector) { | |
354 | assert(VectorTy && "Missing type for vector scalar."); | |
355 | DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " | |
356 | << *VectorTy << '\n'); | |
357 | NewTy = VectorTy; // Use the vector type. | |
358 | } else { | |
359 | unsigned BitWidth = AllocaSize * 8; | |
360 | ||
361 | // Do not convert to scalar integer if the alloca size exceeds the | |
362 | // scalar load threshold. | |
363 | if (BitWidth > ScalarLoadThreshold) | |
364 | return 0; | |
365 | ||
366 | if ((ScalarKind == ImplicitVector || ScalarKind == Integer) && | |
367 | !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth)) | |
368 | return 0; | |
369 | // Dynamic accesses on integers aren't yet supported. They need us to shift | |
370 | // by a dynamic amount which could be difficult to work out as we might not | |
371 | // know whether to use a left or right shift. | |
372 | if (ScalarKind == Integer && HadDynamicAccess) | |
373 | return 0; | |
374 | ||
375 | DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); | |
376 | // Create and insert the integer alloca. | |
377 | NewTy = IntegerType::get(AI->getContext(), BitWidth); | |
378 | } | |
379 | AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); | |
380 | ConvertUsesToScalar(AI, NewAI, 0, 0); | |
381 | return NewAI; | |
382 | } | |
383 | ||
384 | /// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type | |
385 | /// (VectorTy) so far at the offset specified by Offset (which is specified in | |
386 | /// bytes). | |
387 | /// | |
388 | /// There are two cases we handle here: | |
389 | /// 1) A union of vector types of the same size and potentially its elements. | |
390 | /// Here we turn element accesses into insert/extract element operations. | |
391 | /// This promotes a <4 x float> with a store of float to the third element | |
392 | /// into a <4 x float> that uses insert element. | |
393 | /// 2) A fully general blob of memory, which we turn into some (potentially | |
394 | /// large) integer type with extract and insert operations where the loads | |
395 | /// and stores would mutate the memory. We mark this by setting VectorTy | |
396 | /// to VoidTy. | |
397 | void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In, | |
398 | uint64_t Offset) { | |
399 | // If we already decided to turn this into a blob of integer memory, there is | |
400 | // nothing to be done. | |
401 | if (ScalarKind == Integer) | |
402 | return; | |
403 | ||
404 | // If this could be contributing to a vector, analyze it. | |
405 | ||
406 | // If the In type is a vector that is the same size as the alloca, see if it | |
407 | // matches the existing VecTy. | |
408 | if (VectorType *VInTy = dyn_cast<VectorType>(In)) { | |
409 | if (MergeInVectorType(VInTy, Offset)) | |
410 | return; | |
411 | } else if (In->isFloatTy() || In->isDoubleTy() || | |
412 | (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && | |
413 | isPowerOf2_32(In->getPrimitiveSizeInBits()))) { | |
414 | // Full width accesses can be ignored, because they can always be turned | |
415 | // into bitcasts. | |
416 | unsigned EltSize = In->getPrimitiveSizeInBits()/8; | |
417 | if (EltSize == AllocaSize) | |
418 | return; | |
419 | ||
420 | // If we're accessing something that could be an element of a vector, see | |
421 | // if the implied vector agrees with what we already have and if Offset is | |
422 | // compatible with it. | |
423 | if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && | |
424 | (!VectorTy || EltSize == VectorTy->getElementType() | |
425 | ->getPrimitiveSizeInBits()/8)) { | |
426 | if (!VectorTy) { | |
427 | ScalarKind = ImplicitVector; | |
428 | VectorTy = VectorType::get(In, AllocaSize/EltSize); | |
429 | } | |
430 | return; | |
431 | } | |
432 | } | |
433 | ||
434 | // Otherwise, we have a case that we can't handle with an optimized vector | |
435 | // form. We can still turn this into a large integer. | |
436 | ScalarKind = Integer; | |
437 | } | |
438 | ||
439 | /// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore, | |
440 | /// returning true if the type was successfully merged and false otherwise. | |
441 | bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy, | |
442 | uint64_t Offset) { | |
443 | if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { | |
444 | // If we're storing/loading a vector of the right size, allow it as a | |
445 | // vector. If this the first vector we see, remember the type so that | |
446 | // we know the element size. If this is a subsequent access, ignore it | |
447 | // even if it is a differing type but the same size. Worst case we can | |
448 | // bitcast the resultant vectors. | |
449 | if (!VectorTy) | |
450 | VectorTy = VInTy; | |
451 | ScalarKind = Vector; | |
452 | return true; | |
453 | } | |
454 | ||
455 | return false; | |
456 | } | |
457 | ||
458 | /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all | |
459 | /// its accesses to a single vector type, return true and set VecTy to | |
460 | /// the new type. If we could convert the alloca into a single promotable | |
461 | /// integer, return true but set VecTy to VoidTy. Further, if the use is not a | |
462 | /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset | |
463 | /// is the current offset from the base of the alloca being analyzed. | |
464 | /// | |
465 | /// If we see at least one access to the value that is as a vector type, set the | |
466 | /// SawVec flag. | |
467 | bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset, | |
468 | Value* NonConstantIdx) { | |
469 | for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { | |
470 | Instruction *User = cast<Instruction>(*UI); | |
471 | ||
472 | if (LoadInst *LI = dyn_cast<LoadInst>(User)) { | |
473 | // Don't break volatile loads. | |
474 | if (!LI->isSimple()) | |
475 | return false; | |
476 | // Don't touch MMX operations. | |
477 | if (LI->getType()->isX86_MMXTy()) | |
478 | return false; | |
479 | HadNonMemTransferAccess = true; | |
480 | MergeInTypeForLoadOrStore(LI->getType(), Offset); | |
481 | continue; | |
482 | } | |
483 | ||
484 | if (StoreInst *SI = dyn_cast<StoreInst>(User)) { | |
485 | // Storing the pointer, not into the value? | |
486 | if (SI->getOperand(0) == V || !SI->isSimple()) return false; | |
487 | // Don't touch MMX operations. | |
488 | if (SI->getOperand(0)->getType()->isX86_MMXTy()) | |
489 | return false; | |
490 | HadNonMemTransferAccess = true; | |
491 | MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset); | |
492 | continue; | |
493 | } | |
494 | ||
495 | if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { | |
496 | if (!onlyUsedByLifetimeMarkers(BCI)) | |
497 | IsNotTrivial = true; // Can't be mem2reg'd. | |
498 | if (!CanConvertToScalar(BCI, Offset, NonConstantIdx)) | |
499 | return false; | |
500 | continue; | |
501 | } | |
502 | ||
503 | if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { | |
504 | // If this is a GEP with a variable indices, we can't handle it. | |
505 | PointerType* PtrTy = dyn_cast<PointerType>(GEP->getPointerOperandType()); | |
506 | if (!PtrTy) | |
507 | return false; | |
508 | ||
509 | // Compute the offset that this GEP adds to the pointer. | |
510 | SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); | |
511 | Value *GEPNonConstantIdx = 0; | |
512 | if (!GEP->hasAllConstantIndices()) { | |
513 | if (!isa<VectorType>(PtrTy->getElementType())) | |
514 | return false; | |
515 | if (NonConstantIdx) | |
516 | return false; | |
517 | GEPNonConstantIdx = Indices.pop_back_val(); | |
518 | if (!GEPNonConstantIdx->getType()->isIntegerTy(32)) | |
519 | return false; | |
520 | HadDynamicAccess = true; | |
521 | } else | |
522 | GEPNonConstantIdx = NonConstantIdx; | |
523 | uint64_t GEPOffset = TD.getIndexedOffset(PtrTy, | |
524 | Indices); | |
525 | // See if all uses can be converted. | |
526 | if (!CanConvertToScalar(GEP, Offset+GEPOffset, GEPNonConstantIdx)) | |
527 | return false; | |
528 | IsNotTrivial = true; // Can't be mem2reg'd. | |
529 | HadNonMemTransferAccess = true; | |
530 | continue; | |
531 | } | |
532 | ||
533 | // If this is a constant sized memset of a constant value (e.g. 0) we can | |
534 | // handle it. | |
535 | if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { | |
536 | // Store to dynamic index. | |
537 | if (NonConstantIdx) | |
538 | return false; | |
539 | // Store of constant value. | |
540 | if (!isa<ConstantInt>(MSI->getValue())) | |
541 | return false; | |
542 | ||
543 | // Store of constant size. | |
544 | ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength()); | |
545 | if (!Len) | |
546 | return false; | |
547 | ||
548 | // If the size differs from the alloca, we can only convert the alloca to | |
549 | // an integer bag-of-bits. | |
550 | // FIXME: This should handle all of the cases that are currently accepted | |
551 | // as vector element insertions. | |
552 | if (Len->getZExtValue() != AllocaSize || Offset != 0) | |
553 | ScalarKind = Integer; | |
554 | ||
555 | IsNotTrivial = true; // Can't be mem2reg'd. | |
556 | HadNonMemTransferAccess = true; | |
557 | continue; | |
558 | } | |
559 | ||
560 | // If this is a memcpy or memmove into or out of the whole allocation, we | |
561 | // can handle it like a load or store of the scalar type. | |
562 | if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { | |
563 | // Store to dynamic index. | |
564 | if (NonConstantIdx) | |
565 | return false; | |
566 | ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); | |
567 | if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) | |
568 | return false; | |
569 | ||
570 | IsNotTrivial = true; // Can't be mem2reg'd. | |
571 | continue; | |
572 | } | |
573 | ||
574 | // If this is a lifetime intrinsic, we can handle it. | |
575 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { | |
576 | if (II->getIntrinsicID() == Intrinsic::lifetime_start || | |
577 | II->getIntrinsicID() == Intrinsic::lifetime_end) { | |
578 | continue; | |
579 | } | |
580 | } | |
581 | ||
582 | // Otherwise, we cannot handle this! | |
583 | return false; | |
584 | } | |
585 | ||
586 | return true; | |
587 | } | |
588 | ||
589 | /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca | |
590 | /// directly. This happens when we are converting an "integer union" to a | |
591 | /// single integer scalar, or when we are converting a "vector union" to a | |
592 | /// vector with insert/extractelement instructions. | |
593 | /// | |
594 | /// Offset is an offset from the original alloca, in bits that need to be | |
595 | /// shifted to the right. By the end of this, there should be no uses of Ptr. | |
596 | void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, | |
597 | uint64_t Offset, | |
598 | Value* NonConstantIdx) { | |
599 | while (!Ptr->use_empty()) { | |
600 | Instruction *User = cast<Instruction>(Ptr->use_back()); | |
601 | ||
602 | if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { | |
603 | ConvertUsesToScalar(CI, NewAI, Offset, NonConstantIdx); | |
604 | CI->eraseFromParent(); | |
605 | continue; | |
606 | } | |
607 | ||
608 | if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { | |
609 | // Compute the offset that this GEP adds to the pointer. | |
610 | SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); | |
611 | Value* GEPNonConstantIdx = 0; | |
612 | if (!GEP->hasAllConstantIndices()) { | |
613 | assert(!NonConstantIdx && | |
614 | "Dynamic GEP reading from dynamic GEP unsupported"); | |
615 | GEPNonConstantIdx = Indices.pop_back_val(); | |
616 | } else | |
617 | GEPNonConstantIdx = NonConstantIdx; | |
618 | uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), | |
619 | Indices); | |
620 | ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8, GEPNonConstantIdx); | |
621 | GEP->eraseFromParent(); | |
622 | continue; | |
623 | } | |
624 | ||
625 | IRBuilder<> Builder(User); | |
626 | ||
627 | if (LoadInst *LI = dyn_cast<LoadInst>(User)) { | |
628 | // The load is a bit extract from NewAI shifted right by Offset bits. | |
629 | Value *LoadedVal = Builder.CreateLoad(NewAI); | |
630 | Value *NewLoadVal | |
631 | = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, | |
632 | NonConstantIdx, Builder); | |
633 | LI->replaceAllUsesWith(NewLoadVal); | |
634 | LI->eraseFromParent(); | |
635 | continue; | |
636 | } | |
637 | ||
638 | if (StoreInst *SI = dyn_cast<StoreInst>(User)) { | |
639 | assert(SI->getOperand(0) != Ptr && "Consistency error!"); | |
640 | Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); | |
641 | Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, | |
642 | NonConstantIdx, Builder); | |
643 | Builder.CreateStore(New, NewAI); | |
644 | SI->eraseFromParent(); | |
645 | ||
646 | // If the load we just inserted is now dead, then the inserted store | |
647 | // overwrote the entire thing. | |
648 | if (Old->use_empty()) | |
649 | Old->eraseFromParent(); | |
650 | continue; | |
651 | } | |
652 | ||
653 | // If this is a constant sized memset of a constant value (e.g. 0) we can | |
654 | // transform it into a store of the expanded constant value. | |
655 | if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { | |
656 | assert(MSI->getRawDest() == Ptr && "Consistency error!"); | |
657 | assert(!NonConstantIdx && "Cannot replace dynamic memset with insert"); | |
658 | int64_t SNumBytes = cast<ConstantInt>(MSI->getLength())->getSExtValue(); | |
659 | if (SNumBytes > 0 && (SNumBytes >> 32) == 0) { | |
660 | unsigned NumBytes = static_cast<unsigned>(SNumBytes); | |
661 | unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); | |
662 | ||
663 | // Compute the value replicated the right number of times. | |
664 | APInt APVal(NumBytes*8, Val); | |
665 | ||
666 | // Splat the value if non-zero. | |
667 | if (Val) | |
668 | for (unsigned i = 1; i != NumBytes; ++i) | |
669 | APVal |= APVal << 8; | |
670 | ||
671 | Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); | |
672 | Value *New = ConvertScalar_InsertValue( | |
673 | ConstantInt::get(User->getContext(), APVal), | |
674 | Old, Offset, 0, Builder); | |
675 | Builder.CreateStore(New, NewAI); | |
676 | ||
677 | // If the load we just inserted is now dead, then the memset overwrote | |
678 | // the entire thing. | |
679 | if (Old->use_empty()) | |
680 | Old->eraseFromParent(); | |
681 | } | |
682 | MSI->eraseFromParent(); | |
683 | continue; | |
684 | } | |
685 | ||
686 | // If this is a memcpy or memmove into or out of the whole allocation, we | |
687 | // can handle it like a load or store of the scalar type. | |
688 | if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { | |
689 | assert(Offset == 0 && "must be store to start of alloca"); | |
690 | assert(!NonConstantIdx && "Cannot replace dynamic transfer with insert"); | |
691 | ||
692 | // If the source and destination are both to the same alloca, then this is | |
693 | // a noop copy-to-self, just delete it. Otherwise, emit a load and store | |
694 | // as appropriate. | |
695 | AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); | |
696 | ||
697 | if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { | |
698 | // Dest must be OrigAI, change this to be a load from the original | |
699 | // pointer (bitcasted), then a store to our new alloca. | |
700 | assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); | |
701 | Value *SrcPtr = MTI->getSource(); | |
702 | PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); | |
703 | PointerType* AIPTy = cast<PointerType>(NewAI->getType()); | |
704 | if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { | |
705 | AIPTy = PointerType::get(AIPTy->getElementType(), | |
706 | SPTy->getAddressSpace()); | |
707 | } | |
708 | SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); | |
709 | ||
710 | LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); | |
711 | SrcVal->setAlignment(MTI->getAlignment()); | |
712 | Builder.CreateStore(SrcVal, NewAI); | |
713 | } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { | |
714 | // Src must be OrigAI, change this to be a load from NewAI then a store | |
715 | // through the original dest pointer (bitcasted). | |
716 | assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); | |
717 | LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); | |
718 | ||
719 | PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); | |
720 | PointerType* AIPTy = cast<PointerType>(NewAI->getType()); | |
721 | if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { | |
722 | AIPTy = PointerType::get(AIPTy->getElementType(), | |
723 | DPTy->getAddressSpace()); | |
724 | } | |
725 | Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); | |
726 | ||
727 | StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); | |
728 | NewStore->setAlignment(MTI->getAlignment()); | |
729 | } else { | |
730 | // Noop transfer. Src == Dst | |
731 | } | |
732 | ||
733 | MTI->eraseFromParent(); | |
734 | continue; | |
735 | } | |
736 | ||
737 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { | |
738 | if (II->getIntrinsicID() == Intrinsic::lifetime_start || | |
739 | II->getIntrinsicID() == Intrinsic::lifetime_end) { | |
740 | // There's no need to preserve these, as the resulting alloca will be | |
741 | // converted to a register anyways. | |
742 | II->eraseFromParent(); | |
743 | continue; | |
744 | } | |
745 | } | |
746 | ||
747 | llvm_unreachable("Unsupported operation!"); | |
748 | } | |
749 | } | |
750 | ||
751 | /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer | |
752 | /// or vector value FromVal, extracting the bits from the offset specified by | |
753 | /// Offset. This returns the value, which is of type ToType. | |
754 | /// | |
755 | /// This happens when we are converting an "integer union" to a single | |
756 | /// integer scalar, or when we are converting a "vector union" to a vector with | |
757 | /// insert/extractelement instructions. | |
758 | /// | |
759 | /// Offset is an offset from the original alloca, in bits that need to be | |
760 | /// shifted to the right. | |
761 | Value *ConvertToScalarInfo:: | |
762 | ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, | |
763 | uint64_t Offset, Value* NonConstantIdx, | |
764 | IRBuilder<> &Builder) { | |
765 | // If the load is of the whole new alloca, no conversion is needed. | |
766 | Type *FromType = FromVal->getType(); | |
767 | if (FromType == ToType && Offset == 0) | |
768 | return FromVal; | |
769 | ||
770 | // If the result alloca is a vector type, this is either an element | |
771 | // access or a bitcast to another vector type of the same size. | |
772 | if (VectorType *VTy = dyn_cast<VectorType>(FromType)) { | |
773 | unsigned FromTypeSize = TD.getTypeAllocSize(FromType); | |
774 | unsigned ToTypeSize = TD.getTypeAllocSize(ToType); | |
775 | if (FromTypeSize == ToTypeSize) | |
776 | return Builder.CreateBitCast(FromVal, ToType); | |
777 | ||
778 | // Otherwise it must be an element access. | |
779 | unsigned Elt = 0; | |
780 | if (Offset) { | |
781 | unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); | |
782 | Elt = Offset/EltSize; | |
783 | assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); | |
784 | } | |
785 | // Return the element extracted out of it. | |
786 | Value *Idx; | |
787 | if (NonConstantIdx) { | |
788 | if (Elt) | |
789 | Idx = Builder.CreateAdd(NonConstantIdx, | |
790 | Builder.getInt32(Elt), | |
791 | "dyn.offset"); | |
792 | else | |
793 | Idx = NonConstantIdx; | |
794 | } else | |
795 | Idx = Builder.getInt32(Elt); | |
796 | Value *V = Builder.CreateExtractElement(FromVal, Idx); | |
797 | if (V->getType() != ToType) | |
798 | V = Builder.CreateBitCast(V, ToType); | |
799 | return V; | |
800 | } | |
801 | ||
802 | // If ToType is a first class aggregate, extract out each of the pieces and | |
803 | // use insertvalue's to form the FCA. | |
804 | if (StructType *ST = dyn_cast<StructType>(ToType)) { | |
805 | assert(!NonConstantIdx && | |
806 | "Dynamic indexing into struct types not supported"); | |
807 | const StructLayout &Layout = *TD.getStructLayout(ST); | |
808 | Value *Res = UndefValue::get(ST); | |
809 | for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { | |
810 | Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), | |
811 | Offset+Layout.getElementOffsetInBits(i), | |
812 | 0, Builder); | |
813 | Res = Builder.CreateInsertValue(Res, Elt, i); | |
814 | } | |
815 | return Res; | |
816 | } | |
817 | ||
818 | if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) { | |
819 | assert(!NonConstantIdx && | |
820 | "Dynamic indexing into array types not supported"); | |
821 | uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); | |
822 | Value *Res = UndefValue::get(AT); | |
823 | for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { | |
824 | Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), | |
825 | Offset+i*EltSize, 0, Builder); | |
826 | Res = Builder.CreateInsertValue(Res, Elt, i); | |
827 | } | |
828 | return Res; | |
829 | } | |
830 | ||
831 | // Otherwise, this must be a union that was converted to an integer value. | |
832 | IntegerType *NTy = cast<IntegerType>(FromVal->getType()); | |
833 | ||
834 | // If this is a big-endian system and the load is narrower than the | |
835 | // full alloca type, we need to do a shift to get the right bits. | |
836 | int ShAmt = 0; | |
837 | if (TD.isBigEndian()) { | |
838 | // On big-endian machines, the lowest bit is stored at the bit offset | |
839 | // from the pointer given by getTypeStoreSizeInBits. This matters for | |
840 | // integers with a bitwidth that is not a multiple of 8. | |
841 | ShAmt = TD.getTypeStoreSizeInBits(NTy) - | |
842 | TD.getTypeStoreSizeInBits(ToType) - Offset; | |
843 | } else { | |
844 | ShAmt = Offset; | |
845 | } | |
846 | ||
847 | // Note: we support negative bitwidths (with shl) which are not defined. | |
848 | // We do this to support (f.e.) loads off the end of a structure where | |
849 | // only some bits are used. | |
850 | if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) | |
851 | FromVal = Builder.CreateLShr(FromVal, | |
852 | ConstantInt::get(FromVal->getType(), ShAmt)); | |
853 | else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) | |
854 | FromVal = Builder.CreateShl(FromVal, | |
855 | ConstantInt::get(FromVal->getType(), -ShAmt)); | |
856 | ||
857 | // Finally, unconditionally truncate the integer to the right width. | |
858 | unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); | |
859 | if (LIBitWidth < NTy->getBitWidth()) | |
860 | FromVal = | |
861 | Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), | |
862 | LIBitWidth)); | |
863 | else if (LIBitWidth > NTy->getBitWidth()) | |
864 | FromVal = | |
865 | Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), | |
866 | LIBitWidth)); | |
867 | ||
868 | // If the result is an integer, this is a trunc or bitcast. | |
869 | if (ToType->isIntegerTy()) { | |
870 | // Should be done. | |
871 | } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { | |
872 | // Just do a bitcast, we know the sizes match up. | |
873 | FromVal = Builder.CreateBitCast(FromVal, ToType); | |
874 | } else { | |
875 | // Otherwise must be a pointer. | |
876 | FromVal = Builder.CreateIntToPtr(FromVal, ToType); | |
877 | } | |
878 | assert(FromVal->getType() == ToType && "Didn't convert right?"); | |
879 | return FromVal; | |
880 | } | |
881 | ||
882 | /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer | |
883 | /// or vector value "Old" at the offset specified by Offset. | |
884 | /// | |
885 | /// This happens when we are converting an "integer union" to a | |
886 | /// single integer scalar, or when we are converting a "vector union" to a | |
887 | /// vector with insert/extractelement instructions. | |
888 | /// | |
889 | /// Offset is an offset from the original alloca, in bits that need to be | |
890 | /// shifted to the right. | |
891 | /// | |
892 | /// NonConstantIdx is an index value if there was a GEP with a non-constant | |
893 | /// index value. If this is 0 then all GEPs used to find this insert address | |
894 | /// are constant. | |
895 | Value *ConvertToScalarInfo:: | |
896 | ConvertScalar_InsertValue(Value *SV, Value *Old, | |
897 | uint64_t Offset, Value* NonConstantIdx, | |
898 | IRBuilder<> &Builder) { | |
899 | // Convert the stored type to the actual type, shift it left to insert | |
900 | // then 'or' into place. | |
901 | Type *AllocaType = Old->getType(); | |
902 | LLVMContext &Context = Old->getContext(); | |
903 | ||
904 | if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { | |
905 | uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); | |
906 | uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); | |
907 | ||
908 | // Changing the whole vector with memset or with an access of a different | |
909 | // vector type? | |
910 | if (ValSize == VecSize) | |
911 | return Builder.CreateBitCast(SV, AllocaType); | |
912 | ||
913 | // Must be an element insertion. | |
914 | Type *EltTy = VTy->getElementType(); | |
915 | if (SV->getType() != EltTy) | |
916 | SV = Builder.CreateBitCast(SV, EltTy); | |
917 | uint64_t EltSize = TD.getTypeAllocSizeInBits(EltTy); | |
918 | unsigned Elt = Offset/EltSize; | |
919 | Value *Idx; | |
920 | if (NonConstantIdx) { | |
921 | if (Elt) | |
922 | Idx = Builder.CreateAdd(NonConstantIdx, | |
923 | Builder.getInt32(Elt), | |
924 | "dyn.offset"); | |
925 | else | |
926 | Idx = NonConstantIdx; | |
927 | } else | |
928 | Idx = Builder.getInt32(Elt); | |
929 | return Builder.CreateInsertElement(Old, SV, Idx); | |
930 | } | |
931 | ||
932 | // If SV is a first-class aggregate value, insert each value recursively. | |
933 | if (StructType *ST = dyn_cast<StructType>(SV->getType())) { | |
934 | assert(!NonConstantIdx && | |
935 | "Dynamic indexing into struct types not supported"); | |
936 | const StructLayout &Layout = *TD.getStructLayout(ST); | |
937 | for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { | |
938 | Value *Elt = Builder.CreateExtractValue(SV, i); | |
939 | Old = ConvertScalar_InsertValue(Elt, Old, | |
940 | Offset+Layout.getElementOffsetInBits(i), | |
941 | 0, Builder); | |
942 | } | |
943 | return Old; | |
944 | } | |
945 | ||
946 | if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { | |
947 | assert(!NonConstantIdx && | |
948 | "Dynamic indexing into array types not supported"); | |
949 | uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); | |
950 | for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { | |
951 | Value *Elt = Builder.CreateExtractValue(SV, i); | |
952 | Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, 0, Builder); | |
953 | } | |
954 | return Old; | |
955 | } | |
956 | ||
957 | // If SV is a float, convert it to the appropriate integer type. | |
958 | // If it is a pointer, do the same. | |
959 | unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); | |
960 | unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); | |
961 | unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); | |
962 | unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); | |
963 | if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) | |
964 | SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth)); | |
965 | else if (SV->getType()->isPointerTy()) | |
966 | SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext())); | |
967 | ||
968 | // Zero extend or truncate the value if needed. | |
969 | if (SV->getType() != AllocaType) { | |
970 | if (SV->getType()->getPrimitiveSizeInBits() < | |
971 | AllocaType->getPrimitiveSizeInBits()) | |
972 | SV = Builder.CreateZExt(SV, AllocaType); | |
973 | else { | |
974 | // Truncation may be needed if storing more than the alloca can hold | |
975 | // (undefined behavior). | |
976 | SV = Builder.CreateTrunc(SV, AllocaType); | |
977 | SrcWidth = DestWidth; | |
978 | SrcStoreWidth = DestStoreWidth; | |
979 | } | |
980 | } | |
981 | ||
982 | // If this is a big-endian system and the store is narrower than the | |
983 | // full alloca type, we need to do a shift to get the right bits. | |
984 | int ShAmt = 0; | |
985 | if (TD.isBigEndian()) { | |
986 | // On big-endian machines, the lowest bit is stored at the bit offset | |
987 | // from the pointer given by getTypeStoreSizeInBits. This matters for | |
988 | // integers with a bitwidth that is not a multiple of 8. | |
989 | ShAmt = DestStoreWidth - SrcStoreWidth - Offset; | |
990 | } else { | |
991 | ShAmt = Offset; | |
992 | } | |
993 | ||
994 | // Note: we support negative bitwidths (with shr) which are not defined. | |
995 | // We do this to support (f.e.) stores off the end of a structure where | |
996 | // only some bits in the structure are set. | |
997 | APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); | |
998 | if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { | |
999 | SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt)); | |
1000 | Mask <<= ShAmt; | |
1001 | } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { | |
1002 | SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt)); | |
1003 | Mask = Mask.lshr(-ShAmt); | |
1004 | } | |
1005 | ||
1006 | // Mask out the bits we are about to insert from the old value, and or | |
1007 | // in the new bits. | |
1008 | if (SrcWidth != DestWidth) { | |
1009 | assert(DestWidth > SrcWidth); | |
1010 | Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); | |
1011 | SV = Builder.CreateOr(Old, SV, "ins"); | |
1012 | } | |
1013 | return SV; | |
1014 | } | |
1015 | ||
1016 | ||
1017 | //===----------------------------------------------------------------------===// | |
1018 | // SRoA Driver | |
1019 | //===----------------------------------------------------------------------===// | |
1020 | ||
1021 | ||
1022 | bool SROA::runOnFunction(Function &F) { | |
970d7e83 | 1023 | TD = getAnalysisIfAvailable<DataLayout>(); |
223e47cc LB |
1024 | |
1025 | bool Changed = performPromotion(F); | |
1026 | ||
970d7e83 | 1027 | // FIXME: ScalarRepl currently depends on DataLayout more than it |
223e47cc LB |
1028 | // theoretically needs to. It should be refactored in order to support |
1029 | // target-independent IR. Until this is done, just skip the actual | |
1030 | // scalar-replacement portion of this pass. | |
1031 | if (!TD) return Changed; | |
1032 | ||
1033 | while (1) { | |
1034 | bool LocalChange = performScalarRepl(F); | |
1035 | if (!LocalChange) break; // No need to repromote if no scalarrepl | |
1036 | Changed = true; | |
1037 | LocalChange = performPromotion(F); | |
1038 | if (!LocalChange) break; // No need to re-scalarrepl if no promotion | |
1039 | } | |
1040 | ||
1041 | return Changed; | |
1042 | } | |
1043 | ||
1044 | namespace { | |
1045 | class AllocaPromoter : public LoadAndStorePromoter { | |
1046 | AllocaInst *AI; | |
1047 | DIBuilder *DIB; | |
1048 | SmallVector<DbgDeclareInst *, 4> DDIs; | |
1049 | SmallVector<DbgValueInst *, 4> DVIs; | |
1050 | public: | |
1051 | AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, | |
1052 | DIBuilder *DB) | |
1053 | : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {} | |
1054 | ||
1055 | void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { | |
1056 | // Remember which alloca we're promoting (for isInstInList). | |
1057 | this->AI = AI; | |
1058 | if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) { | |
1059 | for (Value::use_iterator UI = DebugNode->use_begin(), | |
1060 | E = DebugNode->use_end(); UI != E; ++UI) | |
1061 | if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI)) | |
1062 | DDIs.push_back(DDI); | |
1063 | else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI)) | |
1064 | DVIs.push_back(DVI); | |
1065 | } | |
1066 | ||
1067 | LoadAndStorePromoter::run(Insts); | |
1068 | AI->eraseFromParent(); | |
1069 | for (SmallVector<DbgDeclareInst *, 4>::iterator I = DDIs.begin(), | |
1070 | E = DDIs.end(); I != E; ++I) { | |
1071 | DbgDeclareInst *DDI = *I; | |
1072 | DDI->eraseFromParent(); | |
1073 | } | |
1074 | for (SmallVector<DbgValueInst *, 4>::iterator I = DVIs.begin(), | |
1075 | E = DVIs.end(); I != E; ++I) { | |
1076 | DbgValueInst *DVI = *I; | |
1077 | DVI->eraseFromParent(); | |
1078 | } | |
1079 | } | |
1080 | ||
1081 | virtual bool isInstInList(Instruction *I, | |
1082 | const SmallVectorImpl<Instruction*> &Insts) const { | |
1083 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | |
1084 | return LI->getOperand(0) == AI; | |
1085 | return cast<StoreInst>(I)->getPointerOperand() == AI; | |
1086 | } | |
1087 | ||
1088 | virtual void updateDebugInfo(Instruction *Inst) const { | |
1089 | for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(), | |
1090 | E = DDIs.end(); I != E; ++I) { | |
1091 | DbgDeclareInst *DDI = *I; | |
1092 | if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) | |
1093 | ConvertDebugDeclareToDebugValue(DDI, SI, *DIB); | |
1094 | else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) | |
1095 | ConvertDebugDeclareToDebugValue(DDI, LI, *DIB); | |
1096 | } | |
1097 | for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(), | |
1098 | E = DVIs.end(); I != E; ++I) { | |
1099 | DbgValueInst *DVI = *I; | |
1100 | Value *Arg = NULL; | |
1101 | if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { | |
1102 | // If an argument is zero extended then use argument directly. The ZExt | |
1103 | // may be zapped by an optimization pass in future. | |
1104 | if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) | |
1105 | Arg = dyn_cast<Argument>(ZExt->getOperand(0)); | |
1106 | if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) | |
1107 | Arg = dyn_cast<Argument>(SExt->getOperand(0)); | |
1108 | if (!Arg) | |
1109 | Arg = SI->getOperand(0); | |
1110 | } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { | |
1111 | Arg = LI->getOperand(0); | |
1112 | } else { | |
1113 | continue; | |
1114 | } | |
1115 | Instruction *DbgVal = | |
1116 | DIB->insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()), | |
1117 | Inst); | |
1118 | DbgVal->setDebugLoc(DVI->getDebugLoc()); | |
1119 | } | |
1120 | } | |
1121 | }; | |
1122 | } // end anon namespace | |
1123 | ||
1124 | /// isSafeSelectToSpeculate - Select instructions that use an alloca and are | |
1125 | /// subsequently loaded can be rewritten to load both input pointers and then | |
1126 | /// select between the result, allowing the load of the alloca to be promoted. | |
1127 | /// From this: | |
1128 | /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other | |
1129 | /// %V = load i32* %P2 | |
1130 | /// to: | |
1131 | /// %V1 = load i32* %Alloca -> will be mem2reg'd | |
1132 | /// %V2 = load i32* %Other | |
1133 | /// %V = select i1 %cond, i32 %V1, i32 %V2 | |
1134 | /// | |
1135 | /// We can do this to a select if its only uses are loads and if the operand to | |
1136 | /// the select can be loaded unconditionally. | |
970d7e83 | 1137 | static bool isSafeSelectToSpeculate(SelectInst *SI, const DataLayout *TD) { |
223e47cc LB |
1138 | bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); |
1139 | bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); | |
1140 | ||
1141 | for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); | |
1142 | UI != UE; ++UI) { | |
1143 | LoadInst *LI = dyn_cast<LoadInst>(*UI); | |
1144 | if (LI == 0 || !LI->isSimple()) return false; | |
1145 | ||
1146 | // Both operands to the select need to be dereferencable, either absolutely | |
1147 | // (e.g. allocas) or at this point because we can see other accesses to it. | |
1148 | if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, | |
1149 | LI->getAlignment(), TD)) | |
1150 | return false; | |
1151 | if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, | |
1152 | LI->getAlignment(), TD)) | |
1153 | return false; | |
1154 | } | |
1155 | ||
1156 | return true; | |
1157 | } | |
1158 | ||
1159 | /// isSafePHIToSpeculate - PHI instructions that use an alloca and are | |
1160 | /// subsequently loaded can be rewritten to load both input pointers in the pred | |
1161 | /// blocks and then PHI the results, allowing the load of the alloca to be | |
1162 | /// promoted. | |
1163 | /// From this: | |
1164 | /// %P2 = phi [i32* %Alloca, i32* %Other] | |
1165 | /// %V = load i32* %P2 | |
1166 | /// to: | |
1167 | /// %V1 = load i32* %Alloca -> will be mem2reg'd | |
1168 | /// ... | |
1169 | /// %V2 = load i32* %Other | |
1170 | /// ... | |
1171 | /// %V = phi [i32 %V1, i32 %V2] | |
1172 | /// | |
1173 | /// We can do this to a select if its only uses are loads and if the operand to | |
1174 | /// the select can be loaded unconditionally. | |
970d7e83 | 1175 | static bool isSafePHIToSpeculate(PHINode *PN, const DataLayout *TD) { |
223e47cc LB |
1176 | // For now, we can only do this promotion if the load is in the same block as |
1177 | // the PHI, and if there are no stores between the phi and load. | |
1178 | // TODO: Allow recursive phi users. | |
1179 | // TODO: Allow stores. | |
1180 | BasicBlock *BB = PN->getParent(); | |
1181 | unsigned MaxAlign = 0; | |
1182 | for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); | |
1183 | UI != UE; ++UI) { | |
1184 | LoadInst *LI = dyn_cast<LoadInst>(*UI); | |
1185 | if (LI == 0 || !LI->isSimple()) return false; | |
1186 | ||
1187 | // For now we only allow loads in the same block as the PHI. This is a | |
1188 | // common case that happens when instcombine merges two loads through a PHI. | |
1189 | if (LI->getParent() != BB) return false; | |
1190 | ||
1191 | // Ensure that there are no instructions between the PHI and the load that | |
1192 | // could store. | |
1193 | for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) | |
1194 | if (BBI->mayWriteToMemory()) | |
1195 | return false; | |
1196 | ||
1197 | MaxAlign = std::max(MaxAlign, LI->getAlignment()); | |
1198 | } | |
1199 | ||
1200 | // Okay, we know that we have one or more loads in the same block as the PHI. | |
1201 | // We can transform this if it is safe to push the loads into the predecessor | |
1202 | // blocks. The only thing to watch out for is that we can't put a possibly | |
1203 | // trapping load in the predecessor if it is a critical edge. | |
1204 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { | |
1205 | BasicBlock *Pred = PN->getIncomingBlock(i); | |
1206 | Value *InVal = PN->getIncomingValue(i); | |
1207 | ||
1208 | // If the terminator of the predecessor has side-effects (an invoke), | |
1209 | // there is no safe place to put a load in the predecessor. | |
1210 | if (Pred->getTerminator()->mayHaveSideEffects()) | |
1211 | return false; | |
1212 | ||
1213 | // If the value is produced by the terminator of the predecessor | |
1214 | // (an invoke), there is no valid place to put a load in the predecessor. | |
1215 | if (Pred->getTerminator() == InVal) | |
1216 | return false; | |
1217 | ||
1218 | // If the predecessor has a single successor, then the edge isn't critical. | |
1219 | if (Pred->getTerminator()->getNumSuccessors() == 1) | |
1220 | continue; | |
1221 | ||
1222 | // If this pointer is always safe to load, or if we can prove that there is | |
1223 | // already a load in the block, then we can move the load to the pred block. | |
1224 | if (InVal->isDereferenceablePointer() || | |
1225 | isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) | |
1226 | continue; | |
1227 | ||
1228 | return false; | |
1229 | } | |
1230 | ||
1231 | return true; | |
1232 | } | |
1233 | ||
1234 | ||
1235 | /// tryToMakeAllocaBePromotable - This returns true if the alloca only has | |
1236 | /// direct (non-volatile) loads and stores to it. If the alloca is close but | |
1237 | /// not quite there, this will transform the code to allow promotion. As such, | |
1238 | /// it is a non-pure predicate. | |
970d7e83 | 1239 | static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const DataLayout *TD) { |
223e47cc LB |
1240 | SetVector<Instruction*, SmallVector<Instruction*, 4>, |
1241 | SmallPtrSet<Instruction*, 4> > InstsToRewrite; | |
1242 | ||
1243 | for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); | |
1244 | UI != UE; ++UI) { | |
1245 | User *U = *UI; | |
1246 | if (LoadInst *LI = dyn_cast<LoadInst>(U)) { | |
1247 | if (!LI->isSimple()) | |
1248 | return false; | |
1249 | continue; | |
1250 | } | |
1251 | ||
1252 | if (StoreInst *SI = dyn_cast<StoreInst>(U)) { | |
1253 | if (SI->getOperand(0) == AI || !SI->isSimple()) | |
1254 | return false; // Don't allow a store OF the AI, only INTO the AI. | |
1255 | continue; | |
1256 | } | |
1257 | ||
1258 | if (SelectInst *SI = dyn_cast<SelectInst>(U)) { | |
1259 | // If the condition being selected on is a constant, fold the select, yes | |
1260 | // this does (rarely) happen early on. | |
1261 | if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { | |
1262 | Value *Result = SI->getOperand(1+CI->isZero()); | |
1263 | SI->replaceAllUsesWith(Result); | |
1264 | SI->eraseFromParent(); | |
1265 | ||
1266 | // This is very rare and we just scrambled the use list of AI, start | |
1267 | // over completely. | |
1268 | return tryToMakeAllocaBePromotable(AI, TD); | |
1269 | } | |
1270 | ||
1271 | // If it is safe to turn "load (select c, AI, ptr)" into a select of two | |
1272 | // loads, then we can transform this by rewriting the select. | |
1273 | if (!isSafeSelectToSpeculate(SI, TD)) | |
1274 | return false; | |
1275 | ||
1276 | InstsToRewrite.insert(SI); | |
1277 | continue; | |
1278 | } | |
1279 | ||
1280 | if (PHINode *PN = dyn_cast<PHINode>(U)) { | |
1281 | if (PN->use_empty()) { // Dead PHIs can be stripped. | |
1282 | InstsToRewrite.insert(PN); | |
1283 | continue; | |
1284 | } | |
1285 | ||
1286 | // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads | |
1287 | // in the pred blocks, then we can transform this by rewriting the PHI. | |
1288 | if (!isSafePHIToSpeculate(PN, TD)) | |
1289 | return false; | |
1290 | ||
1291 | InstsToRewrite.insert(PN); | |
1292 | continue; | |
1293 | } | |
1294 | ||
1295 | if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { | |
1296 | if (onlyUsedByLifetimeMarkers(BCI)) { | |
1297 | InstsToRewrite.insert(BCI); | |
1298 | continue; | |
1299 | } | |
1300 | } | |
1301 | ||
1302 | return false; | |
1303 | } | |
1304 | ||
1305 | // If there are no instructions to rewrite, then all uses are load/stores and | |
1306 | // we're done! | |
1307 | if (InstsToRewrite.empty()) | |
1308 | return true; | |
1309 | ||
1310 | // If we have instructions that need to be rewritten for this to be promotable | |
1311 | // take care of it now. | |
1312 | for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { | |
1313 | if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) { | |
1314 | // This could only be a bitcast used by nothing but lifetime intrinsics. | |
1315 | for (BitCastInst::use_iterator I = BCI->use_begin(), E = BCI->use_end(); | |
1316 | I != E;) { | |
1317 | Use &U = I.getUse(); | |
1318 | ++I; | |
1319 | cast<Instruction>(U.getUser())->eraseFromParent(); | |
1320 | } | |
1321 | BCI->eraseFromParent(); | |
1322 | continue; | |
1323 | } | |
1324 | ||
1325 | if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { | |
1326 | // Selects in InstsToRewrite only have load uses. Rewrite each as two | |
1327 | // loads with a new select. | |
1328 | while (!SI->use_empty()) { | |
1329 | LoadInst *LI = cast<LoadInst>(SI->use_back()); | |
1330 | ||
1331 | IRBuilder<> Builder(LI); | |
1332 | LoadInst *TrueLoad = | |
1333 | Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); | |
1334 | LoadInst *FalseLoad = | |
1335 | Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f"); | |
1336 | ||
1337 | // Transfer alignment and TBAA info if present. | |
1338 | TrueLoad->setAlignment(LI->getAlignment()); | |
1339 | FalseLoad->setAlignment(LI->getAlignment()); | |
1340 | if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { | |
1341 | TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); | |
1342 | FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); | |
1343 | } | |
1344 | ||
1345 | Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); | |
1346 | V->takeName(LI); | |
1347 | LI->replaceAllUsesWith(V); | |
1348 | LI->eraseFromParent(); | |
1349 | } | |
1350 | ||
1351 | // Now that all the loads are gone, the select is gone too. | |
1352 | SI->eraseFromParent(); | |
1353 | continue; | |
1354 | } | |
1355 | ||
1356 | // Otherwise, we have a PHI node which allows us to push the loads into the | |
1357 | // predecessors. | |
1358 | PHINode *PN = cast<PHINode>(InstsToRewrite[i]); | |
1359 | if (PN->use_empty()) { | |
1360 | PN->eraseFromParent(); | |
1361 | continue; | |
1362 | } | |
1363 | ||
1364 | Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); | |
1365 | PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), | |
1366 | PN->getName()+".ld", PN); | |
1367 | ||
1368 | // Get the TBAA tag and alignment to use from one of the loads. It doesn't | |
1369 | // matter which one we get and if any differ, it doesn't matter. | |
1370 | LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); | |
1371 | MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); | |
1372 | unsigned Align = SomeLoad->getAlignment(); | |
1373 | ||
1374 | // Rewrite all loads of the PN to use the new PHI. | |
1375 | while (!PN->use_empty()) { | |
1376 | LoadInst *LI = cast<LoadInst>(PN->use_back()); | |
1377 | LI->replaceAllUsesWith(NewPN); | |
1378 | LI->eraseFromParent(); | |
1379 | } | |
1380 | ||
1381 | // Inject loads into all of the pred blocks. Keep track of which blocks we | |
1382 | // insert them into in case we have multiple edges from the same block. | |
1383 | DenseMap<BasicBlock*, LoadInst*> InsertedLoads; | |
1384 | ||
1385 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { | |
1386 | BasicBlock *Pred = PN->getIncomingBlock(i); | |
1387 | LoadInst *&Load = InsertedLoads[Pred]; | |
1388 | if (Load == 0) { | |
1389 | Load = new LoadInst(PN->getIncomingValue(i), | |
1390 | PN->getName() + "." + Pred->getName(), | |
1391 | Pred->getTerminator()); | |
1392 | Load->setAlignment(Align); | |
1393 | if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); | |
1394 | } | |
1395 | ||
1396 | NewPN->addIncoming(Load, Pred); | |
1397 | } | |
1398 | ||
1399 | PN->eraseFromParent(); | |
1400 | } | |
1401 | ||
1402 | ++NumAdjusted; | |
1403 | return true; | |
1404 | } | |
1405 | ||
1406 | bool SROA::performPromotion(Function &F) { | |
1407 | std::vector<AllocaInst*> Allocas; | |
1408 | DominatorTree *DT = 0; | |
1409 | if (HasDomTree) | |
1410 | DT = &getAnalysis<DominatorTree>(); | |
1411 | ||
1412 | BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function | |
1413 | DIBuilder DIB(*F.getParent()); | |
1414 | bool Changed = false; | |
1415 | SmallVector<Instruction*, 64> Insts; | |
1416 | while (1) { | |
1417 | Allocas.clear(); | |
1418 | ||
1419 | // Find allocas that are safe to promote, by looking at all instructions in | |
1420 | // the entry node | |
1421 | for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) | |
1422 | if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? | |
1423 | if (tryToMakeAllocaBePromotable(AI, TD)) | |
1424 | Allocas.push_back(AI); | |
1425 | ||
1426 | if (Allocas.empty()) break; | |
1427 | ||
1428 | if (HasDomTree) | |
1429 | PromoteMemToReg(Allocas, *DT); | |
1430 | else { | |
1431 | SSAUpdater SSA; | |
1432 | for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { | |
1433 | AllocaInst *AI = Allocas[i]; | |
1434 | ||
1435 | // Build list of instructions to promote. | |
1436 | for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); | |
1437 | UI != E; ++UI) | |
1438 | Insts.push_back(cast<Instruction>(*UI)); | |
1439 | AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts); | |
1440 | Insts.clear(); | |
1441 | } | |
1442 | } | |
1443 | NumPromoted += Allocas.size(); | |
1444 | Changed = true; | |
1445 | } | |
1446 | ||
1447 | return Changed; | |
1448 | } | |
1449 | ||
1450 | ||
1451 | /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for | |
1452 | /// SROA. It must be a struct or array type with a small number of elements. | |
1453 | bool SROA::ShouldAttemptScalarRepl(AllocaInst *AI) { | |
1454 | Type *T = AI->getAllocatedType(); | |
1455 | // Do not promote any struct that has too many members. | |
1456 | if (StructType *ST = dyn_cast<StructType>(T)) | |
1457 | return ST->getNumElements() <= StructMemberThreshold; | |
1458 | // Do not promote any array that has too many elements. | |
1459 | if (ArrayType *AT = dyn_cast<ArrayType>(T)) | |
1460 | return AT->getNumElements() <= ArrayElementThreshold; | |
1461 | return false; | |
1462 | } | |
1463 | ||
1464 | // performScalarRepl - This algorithm is a simple worklist driven algorithm, | |
1465 | // which runs on all of the alloca instructions in the function, removing them | |
1466 | // if they are only used by getelementptr instructions. | |
1467 | // | |
1468 | bool SROA::performScalarRepl(Function &F) { | |
1469 | std::vector<AllocaInst*> WorkList; | |
1470 | ||
1471 | // Scan the entry basic block, adding allocas to the worklist. | |
1472 | BasicBlock &BB = F.getEntryBlock(); | |
1473 | for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) | |
1474 | if (AllocaInst *A = dyn_cast<AllocaInst>(I)) | |
1475 | WorkList.push_back(A); | |
1476 | ||
1477 | // Process the worklist | |
1478 | bool Changed = false; | |
1479 | while (!WorkList.empty()) { | |
1480 | AllocaInst *AI = WorkList.back(); | |
1481 | WorkList.pop_back(); | |
1482 | ||
1483 | // Handle dead allocas trivially. These can be formed by SROA'ing arrays | |
1484 | // with unused elements. | |
1485 | if (AI->use_empty()) { | |
1486 | AI->eraseFromParent(); | |
1487 | Changed = true; | |
1488 | continue; | |
1489 | } | |
1490 | ||
1491 | // If this alloca is impossible for us to promote, reject it early. | |
1492 | if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) | |
1493 | continue; | |
1494 | ||
1495 | // Check to see if we can perform the core SROA transformation. We cannot | |
1496 | // transform the allocation instruction if it is an array allocation | |
1497 | // (allocations OF arrays are ok though), and an allocation of a scalar | |
1498 | // value cannot be decomposed at all. | |
1499 | uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); | |
1500 | ||
1501 | // Do not promote [0 x %struct]. | |
1502 | if (AllocaSize == 0) continue; | |
1503 | ||
1504 | // Do not promote any struct whose size is too big. | |
1505 | if (AllocaSize > SRThreshold) continue; | |
1506 | ||
1507 | // If the alloca looks like a good candidate for scalar replacement, and if | |
1508 | // all its users can be transformed, then split up the aggregate into its | |
1509 | // separate elements. | |
1510 | if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { | |
1511 | DoScalarReplacement(AI, WorkList); | |
1512 | Changed = true; | |
1513 | continue; | |
1514 | } | |
1515 | ||
1516 | // If we can turn this aggregate value (potentially with casts) into a | |
1517 | // simple scalar value that can be mem2reg'd into a register value. | |
1518 | // IsNotTrivial tracks whether this is something that mem2reg could have | |
1519 | // promoted itself. If so, we don't want to transform it needlessly. Note | |
1520 | // that we can't just check based on the type: the alloca may be of an i32 | |
1521 | // but that has pointer arithmetic to set byte 3 of it or something. | |
1522 | if (AllocaInst *NewAI = ConvertToScalarInfo( | |
1523 | (unsigned)AllocaSize, *TD, ScalarLoadThreshold).TryConvert(AI)) { | |
1524 | NewAI->takeName(AI); | |
1525 | AI->eraseFromParent(); | |
1526 | ++NumConverted; | |
1527 | Changed = true; | |
1528 | continue; | |
1529 | } | |
1530 | ||
1531 | // Otherwise, couldn't process this alloca. | |
1532 | } | |
1533 | ||
1534 | return Changed; | |
1535 | } | |
1536 | ||
1537 | /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl | |
1538 | /// predicate, do SROA now. | |
1539 | void SROA::DoScalarReplacement(AllocaInst *AI, | |
1540 | std::vector<AllocaInst*> &WorkList) { | |
1541 | DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); | |
1542 | SmallVector<AllocaInst*, 32> ElementAllocas; | |
1543 | if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { | |
1544 | ElementAllocas.reserve(ST->getNumContainedTypes()); | |
1545 | for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { | |
1546 | AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, | |
1547 | AI->getAlignment(), | |
1548 | AI->getName() + "." + Twine(i), AI); | |
1549 | ElementAllocas.push_back(NA); | |
1550 | WorkList.push_back(NA); // Add to worklist for recursive processing | |
1551 | } | |
1552 | } else { | |
1553 | ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); | |
1554 | ElementAllocas.reserve(AT->getNumElements()); | |
1555 | Type *ElTy = AT->getElementType(); | |
1556 | for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { | |
1557 | AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), | |
1558 | AI->getName() + "." + Twine(i), AI); | |
1559 | ElementAllocas.push_back(NA); | |
1560 | WorkList.push_back(NA); // Add to worklist for recursive processing | |
1561 | } | |
1562 | } | |
1563 | ||
1564 | // Now that we have created the new alloca instructions, rewrite all the | |
1565 | // uses of the old alloca. | |
1566 | RewriteForScalarRepl(AI, AI, 0, ElementAllocas); | |
1567 | ||
1568 | // Now erase any instructions that were made dead while rewriting the alloca. | |
1569 | DeleteDeadInstructions(); | |
1570 | AI->eraseFromParent(); | |
1571 | ||
1572 | ++NumReplaced; | |
1573 | } | |
1574 | ||
1575 | /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, | |
1576 | /// recursively including all their operands that become trivially dead. | |
1577 | void SROA::DeleteDeadInstructions() { | |
1578 | while (!DeadInsts.empty()) { | |
1579 | Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); | |
1580 | ||
1581 | for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) | |
1582 | if (Instruction *U = dyn_cast<Instruction>(*OI)) { | |
1583 | // Zero out the operand and see if it becomes trivially dead. | |
1584 | // (But, don't add allocas to the dead instruction list -- they are | |
1585 | // already on the worklist and will be deleted separately.) | |
1586 | *OI = 0; | |
1587 | if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) | |
1588 | DeadInsts.push_back(U); | |
1589 | } | |
1590 | ||
1591 | I->eraseFromParent(); | |
1592 | } | |
1593 | } | |
1594 | ||
1595 | /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to | |
1596 | /// performing scalar replacement of alloca AI. The results are flagged in | |
1597 | /// the Info parameter. Offset indicates the position within AI that is | |
1598 | /// referenced by this instruction. | |
1599 | void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, | |
1600 | AllocaInfo &Info) { | |
1601 | for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { | |
1602 | Instruction *User = cast<Instruction>(*UI); | |
1603 | ||
1604 | if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { | |
1605 | isSafeForScalarRepl(BC, Offset, Info); | |
1606 | } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { | |
1607 | uint64_t GEPOffset = Offset; | |
1608 | isSafeGEP(GEPI, GEPOffset, Info); | |
1609 | if (!Info.isUnsafe) | |
1610 | isSafeForScalarRepl(GEPI, GEPOffset, Info); | |
1611 | } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { | |
1612 | ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); | |
1613 | if (Length == 0) | |
1614 | return MarkUnsafe(Info, User); | |
1615 | if (Length->isNegative()) | |
1616 | return MarkUnsafe(Info, User); | |
1617 | ||
1618 | isSafeMemAccess(Offset, Length->getZExtValue(), 0, | |
1619 | UI.getOperandNo() == 0, Info, MI, | |
1620 | true /*AllowWholeAccess*/); | |
1621 | } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { | |
1622 | if (!LI->isSimple()) | |
1623 | return MarkUnsafe(Info, User); | |
1624 | Type *LIType = LI->getType(); | |
1625 | isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), | |
1626 | LIType, false, Info, LI, true /*AllowWholeAccess*/); | |
1627 | Info.hasALoadOrStore = true; | |
1628 | ||
1629 | } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { | |
1630 | // Store is ok if storing INTO the pointer, not storing the pointer | |
1631 | if (!SI->isSimple() || SI->getOperand(0) == I) | |
1632 | return MarkUnsafe(Info, User); | |
1633 | ||
1634 | Type *SIType = SI->getOperand(0)->getType(); | |
1635 | isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), | |
1636 | SIType, true, Info, SI, true /*AllowWholeAccess*/); | |
1637 | Info.hasALoadOrStore = true; | |
1638 | } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { | |
1639 | if (II->getIntrinsicID() != Intrinsic::lifetime_start && | |
1640 | II->getIntrinsicID() != Intrinsic::lifetime_end) | |
1641 | return MarkUnsafe(Info, User); | |
1642 | } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { | |
1643 | isSafePHISelectUseForScalarRepl(User, Offset, Info); | |
1644 | } else { | |
1645 | return MarkUnsafe(Info, User); | |
1646 | } | |
1647 | if (Info.isUnsafe) return; | |
1648 | } | |
1649 | } | |
1650 | ||
1651 | ||
1652 | /// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer | |
1653 | /// derived from the alloca, we can often still split the alloca into elements. | |
1654 | /// This is useful if we have a large alloca where one element is phi'd | |
1655 | /// together somewhere: we can SRoA and promote all the other elements even if | |
1656 | /// we end up not being able to promote this one. | |
1657 | /// | |
1658 | /// All we require is that the uses of the PHI do not index into other parts of | |
1659 | /// the alloca. The most important use case for this is single load and stores | |
1660 | /// that are PHI'd together, which can happen due to code sinking. | |
1661 | void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, | |
1662 | AllocaInfo &Info) { | |
1663 | // If we've already checked this PHI, don't do it again. | |
1664 | if (PHINode *PN = dyn_cast<PHINode>(I)) | |
1665 | if (!Info.CheckedPHIs.insert(PN)) | |
1666 | return; | |
1667 | ||
1668 | for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { | |
1669 | Instruction *User = cast<Instruction>(*UI); | |
1670 | ||
1671 | if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { | |
1672 | isSafePHISelectUseForScalarRepl(BC, Offset, Info); | |
1673 | } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { | |
1674 | // Only allow "bitcast" GEPs for simplicity. We could generalize this, | |
1675 | // but would have to prove that we're staying inside of an element being | |
1676 | // promoted. | |
1677 | if (!GEPI->hasAllZeroIndices()) | |
1678 | return MarkUnsafe(Info, User); | |
1679 | isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); | |
1680 | } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { | |
1681 | if (!LI->isSimple()) | |
1682 | return MarkUnsafe(Info, User); | |
1683 | Type *LIType = LI->getType(); | |
1684 | isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), | |
1685 | LIType, false, Info, LI, false /*AllowWholeAccess*/); | |
1686 | Info.hasALoadOrStore = true; | |
1687 | ||
1688 | } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { | |
1689 | // Store is ok if storing INTO the pointer, not storing the pointer | |
1690 | if (!SI->isSimple() || SI->getOperand(0) == I) | |
1691 | return MarkUnsafe(Info, User); | |
1692 | ||
1693 | Type *SIType = SI->getOperand(0)->getType(); | |
1694 | isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), | |
1695 | SIType, true, Info, SI, false /*AllowWholeAccess*/); | |
1696 | Info.hasALoadOrStore = true; | |
1697 | } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { | |
1698 | isSafePHISelectUseForScalarRepl(User, Offset, Info); | |
1699 | } else { | |
1700 | return MarkUnsafe(Info, User); | |
1701 | } | |
1702 | if (Info.isUnsafe) return; | |
1703 | } | |
1704 | } | |
1705 | ||
1706 | /// isSafeGEP - Check if a GEP instruction can be handled for scalar | |
1707 | /// replacement. It is safe when all the indices are constant, in-bounds | |
1708 | /// references, and when the resulting offset corresponds to an element within | |
1709 | /// the alloca type. The results are flagged in the Info parameter. Upon | |
1710 | /// return, Offset is adjusted as specified by the GEP indices. | |
1711 | void SROA::isSafeGEP(GetElementPtrInst *GEPI, | |
1712 | uint64_t &Offset, AllocaInfo &Info) { | |
1713 | gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); | |
1714 | if (GEPIt == E) | |
1715 | return; | |
1716 | bool NonConstant = false; | |
1717 | unsigned NonConstantIdxSize = 0; | |
1718 | ||
1719 | // Walk through the GEP type indices, checking the types that this indexes | |
1720 | // into. | |
1721 | for (; GEPIt != E; ++GEPIt) { | |
1722 | // Ignore struct elements, no extra checking needed for these. | |
1723 | if ((*GEPIt)->isStructTy()) | |
1724 | continue; | |
1725 | ||
1726 | ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); | |
1727 | if (!IdxVal) { | |
1728 | // Non constant GEPs are only a problem on arrays, structs, and pointers | |
1729 | // Vectors can be dynamically indexed. | |
1730 | // FIXME: Add support for dynamic indexing on arrays. This should be | |
1731 | // ok on any subarrays of the alloca array, eg, a[0][i] is ok, but a[i][0] | |
1732 | // isn't. | |
1733 | if (!(*GEPIt)->isVectorTy()) | |
1734 | return MarkUnsafe(Info, GEPI); | |
1735 | NonConstant = true; | |
1736 | NonConstantIdxSize = TD->getTypeAllocSize(*GEPIt); | |
1737 | } | |
1738 | } | |
1739 | ||
1740 | // Compute the offset due to this GEP and check if the alloca has a | |
1741 | // component element at that offset. | |
1742 | SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); | |
1743 | // If this GEP is non constant then the last operand must have been a | |
1744 | // dynamic index into a vector. Pop this now as it has no impact on the | |
1745 | // constant part of the offset. | |
1746 | if (NonConstant) | |
1747 | Indices.pop_back(); | |
1748 | Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); | |
1749 | if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, | |
1750 | NonConstantIdxSize)) | |
1751 | MarkUnsafe(Info, GEPI); | |
1752 | } | |
1753 | ||
1754 | /// isHomogeneousAggregate - Check if type T is a struct or array containing | |
1755 | /// elements of the same type (which is always true for arrays). If so, | |
1756 | /// return true with NumElts and EltTy set to the number of elements and the | |
1757 | /// element type, respectively. | |
1758 | static bool isHomogeneousAggregate(Type *T, unsigned &NumElts, | |
1759 | Type *&EltTy) { | |
1760 | if (ArrayType *AT = dyn_cast<ArrayType>(T)) { | |
1761 | NumElts = AT->getNumElements(); | |
1762 | EltTy = (NumElts == 0 ? 0 : AT->getElementType()); | |
1763 | return true; | |
1764 | } | |
1765 | if (StructType *ST = dyn_cast<StructType>(T)) { | |
1766 | NumElts = ST->getNumContainedTypes(); | |
1767 | EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); | |
1768 | for (unsigned n = 1; n < NumElts; ++n) { | |
1769 | if (ST->getContainedType(n) != EltTy) | |
1770 | return false; | |
1771 | } | |
1772 | return true; | |
1773 | } | |
1774 | return false; | |
1775 | } | |
1776 | ||
1777 | /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are | |
1778 | /// "homogeneous" aggregates with the same element type and number of elements. | |
1779 | static bool isCompatibleAggregate(Type *T1, Type *T2) { | |
1780 | if (T1 == T2) | |
1781 | return true; | |
1782 | ||
1783 | unsigned NumElts1, NumElts2; | |
1784 | Type *EltTy1, *EltTy2; | |
1785 | if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && | |
1786 | isHomogeneousAggregate(T2, NumElts2, EltTy2) && | |
1787 | NumElts1 == NumElts2 && | |
1788 | EltTy1 == EltTy2) | |
1789 | return true; | |
1790 | ||
1791 | return false; | |
1792 | } | |
1793 | ||
1794 | /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI | |
1795 | /// alloca or has an offset and size that corresponds to a component element | |
1796 | /// within it. The offset checked here may have been formed from a GEP with a | |
1797 | /// pointer bitcasted to a different type. | |
1798 | /// | |
1799 | /// If AllowWholeAccess is true, then this allows uses of the entire alloca as a | |
1800 | /// unit. If false, it only allows accesses known to be in a single element. | |
1801 | void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, | |
1802 | Type *MemOpType, bool isStore, | |
1803 | AllocaInfo &Info, Instruction *TheAccess, | |
1804 | bool AllowWholeAccess) { | |
1805 | // Check if this is a load/store of the entire alloca. | |
1806 | if (Offset == 0 && AllowWholeAccess && | |
1807 | MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { | |
1808 | // This can be safe for MemIntrinsics (where MemOpType is 0) and integer | |
1809 | // loads/stores (which are essentially the same as the MemIntrinsics with | |
1810 | // regard to copying padding between elements). But, if an alloca is | |
1811 | // flagged as both a source and destination of such operations, we'll need | |
1812 | // to check later for padding between elements. | |
1813 | if (!MemOpType || MemOpType->isIntegerTy()) { | |
1814 | if (isStore) | |
1815 | Info.isMemCpyDst = true; | |
1816 | else | |
1817 | Info.isMemCpySrc = true; | |
1818 | return; | |
1819 | } | |
1820 | // This is also safe for references using a type that is compatible with | |
1821 | // the type of the alloca, so that loads/stores can be rewritten using | |
1822 | // insertvalue/extractvalue. | |
1823 | if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { | |
1824 | Info.hasSubelementAccess = true; | |
1825 | return; | |
1826 | } | |
1827 | } | |
1828 | // Check if the offset/size correspond to a component within the alloca type. | |
1829 | Type *T = Info.AI->getAllocatedType(); | |
1830 | if (TypeHasComponent(T, Offset, MemSize)) { | |
1831 | Info.hasSubelementAccess = true; | |
1832 | return; | |
1833 | } | |
1834 | ||
1835 | return MarkUnsafe(Info, TheAccess); | |
1836 | } | |
1837 | ||
1838 | /// TypeHasComponent - Return true if T has a component type with the | |
1839 | /// specified offset and size. If Size is zero, do not check the size. | |
1840 | bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) { | |
1841 | Type *EltTy; | |
1842 | uint64_t EltSize; | |
1843 | if (StructType *ST = dyn_cast<StructType>(T)) { | |
1844 | const StructLayout *Layout = TD->getStructLayout(ST); | |
1845 | unsigned EltIdx = Layout->getElementContainingOffset(Offset); | |
1846 | EltTy = ST->getContainedType(EltIdx); | |
1847 | EltSize = TD->getTypeAllocSize(EltTy); | |
1848 | Offset -= Layout->getElementOffset(EltIdx); | |
1849 | } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { | |
1850 | EltTy = AT->getElementType(); | |
1851 | EltSize = TD->getTypeAllocSize(EltTy); | |
1852 | if (Offset >= AT->getNumElements() * EltSize) | |
1853 | return false; | |
1854 | Offset %= EltSize; | |
1855 | } else if (VectorType *VT = dyn_cast<VectorType>(T)) { | |
1856 | EltTy = VT->getElementType(); | |
1857 | EltSize = TD->getTypeAllocSize(EltTy); | |
1858 | if (Offset >= VT->getNumElements() * EltSize) | |
1859 | return false; | |
1860 | Offset %= EltSize; | |
1861 | } else { | |
1862 | return false; | |
1863 | } | |
1864 | if (Offset == 0 && (Size == 0 || EltSize == Size)) | |
1865 | return true; | |
1866 | // Check if the component spans multiple elements. | |
1867 | if (Offset + Size > EltSize) | |
1868 | return false; | |
1869 | return TypeHasComponent(EltTy, Offset, Size); | |
1870 | } | |
1871 | ||
1872 | /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite | |
1873 | /// the instruction I, which references it, to use the separate elements. | |
1874 | /// Offset indicates the position within AI that is referenced by this | |
1875 | /// instruction. | |
1876 | void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, | |
1877 | SmallVector<AllocaInst*, 32> &NewElts) { | |
1878 | for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { | |
1879 | Use &TheUse = UI.getUse(); | |
1880 | Instruction *User = cast<Instruction>(*UI++); | |
1881 | ||
1882 | if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { | |
1883 | RewriteBitCast(BC, AI, Offset, NewElts); | |
1884 | continue; | |
1885 | } | |
1886 | ||
1887 | if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { | |
1888 | RewriteGEP(GEPI, AI, Offset, NewElts); | |
1889 | continue; | |
1890 | } | |
1891 | ||
1892 | if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { | |
1893 | ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); | |
1894 | uint64_t MemSize = Length->getZExtValue(); | |
1895 | if (Offset == 0 && | |
1896 | MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) | |
1897 | RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); | |
1898 | // Otherwise the intrinsic can only touch a single element and the | |
1899 | // address operand will be updated, so nothing else needs to be done. | |
1900 | continue; | |
1901 | } | |
1902 | ||
1903 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { | |
1904 | if (II->getIntrinsicID() == Intrinsic::lifetime_start || | |
1905 | II->getIntrinsicID() == Intrinsic::lifetime_end) { | |
1906 | RewriteLifetimeIntrinsic(II, AI, Offset, NewElts); | |
1907 | } | |
1908 | continue; | |
1909 | } | |
1910 | ||
1911 | if (LoadInst *LI = dyn_cast<LoadInst>(User)) { | |
1912 | Type *LIType = LI->getType(); | |
1913 | ||
1914 | if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { | |
1915 | // Replace: | |
1916 | // %res = load { i32, i32 }* %alloc | |
1917 | // with: | |
1918 | // %load.0 = load i32* %alloc.0 | |
1919 | // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 | |
1920 | // %load.1 = load i32* %alloc.1 | |
1921 | // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 | |
1922 | // (Also works for arrays instead of structs) | |
1923 | Value *Insert = UndefValue::get(LIType); | |
1924 | IRBuilder<> Builder(LI); | |
1925 | for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { | |
1926 | Value *Load = Builder.CreateLoad(NewElts[i], "load"); | |
1927 | Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); | |
1928 | } | |
1929 | LI->replaceAllUsesWith(Insert); | |
1930 | DeadInsts.push_back(LI); | |
1931 | } else if (LIType->isIntegerTy() && | |
1932 | TD->getTypeAllocSize(LIType) == | |
1933 | TD->getTypeAllocSize(AI->getAllocatedType())) { | |
1934 | // If this is a load of the entire alloca to an integer, rewrite it. | |
1935 | RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); | |
1936 | } | |
1937 | continue; | |
1938 | } | |
1939 | ||
1940 | if (StoreInst *SI = dyn_cast<StoreInst>(User)) { | |
1941 | Value *Val = SI->getOperand(0); | |
1942 | Type *SIType = Val->getType(); | |
1943 | if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { | |
1944 | // Replace: | |
1945 | // store { i32, i32 } %val, { i32, i32 }* %alloc | |
1946 | // with: | |
1947 | // %val.0 = extractvalue { i32, i32 } %val, 0 | |
1948 | // store i32 %val.0, i32* %alloc.0 | |
1949 | // %val.1 = extractvalue { i32, i32 } %val, 1 | |
1950 | // store i32 %val.1, i32* %alloc.1 | |
1951 | // (Also works for arrays instead of structs) | |
1952 | IRBuilder<> Builder(SI); | |
1953 | for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { | |
1954 | Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); | |
1955 | Builder.CreateStore(Extract, NewElts[i]); | |
1956 | } | |
1957 | DeadInsts.push_back(SI); | |
1958 | } else if (SIType->isIntegerTy() && | |
1959 | TD->getTypeAllocSize(SIType) == | |
1960 | TD->getTypeAllocSize(AI->getAllocatedType())) { | |
1961 | // If this is a store of the entire alloca from an integer, rewrite it. | |
1962 | RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); | |
1963 | } | |
1964 | continue; | |
1965 | } | |
1966 | ||
1967 | if (isa<SelectInst>(User) || isa<PHINode>(User)) { | |
1968 | // If we have a PHI user of the alloca itself (as opposed to a GEP or | |
1969 | // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to | |
1970 | // the new pointer. | |
1971 | if (!isa<AllocaInst>(I)) continue; | |
1972 | ||
1973 | assert(Offset == 0 && NewElts[0] && | |
1974 | "Direct alloca use should have a zero offset"); | |
1975 | ||
1976 | // If we have a use of the alloca, we know the derived uses will be | |
1977 | // utilizing just the first element of the scalarized result. Insert a | |
1978 | // bitcast of the first alloca before the user as required. | |
1979 | AllocaInst *NewAI = NewElts[0]; | |
1980 | BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); | |
1981 | NewAI->moveBefore(BCI); | |
1982 | TheUse = BCI; | |
1983 | continue; | |
1984 | } | |
1985 | } | |
1986 | } | |
1987 | ||
1988 | /// RewriteBitCast - Update a bitcast reference to the alloca being replaced | |
1989 | /// and recursively continue updating all of its uses. | |
1990 | void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, | |
1991 | SmallVector<AllocaInst*, 32> &NewElts) { | |
1992 | RewriteForScalarRepl(BC, AI, Offset, NewElts); | |
1993 | if (BC->getOperand(0) != AI) | |
1994 | return; | |
1995 | ||
1996 | // The bitcast references the original alloca. Replace its uses with | |
1997 | // references to the alloca containing offset zero (which is normally at | |
1998 | // index zero, but might not be in cases involving structs with elements | |
1999 | // of size zero). | |
2000 | Type *T = AI->getAllocatedType(); | |
2001 | uint64_t EltOffset = 0; | |
2002 | Type *IdxTy; | |
2003 | uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); | |
2004 | Instruction *Val = NewElts[Idx]; | |
2005 | if (Val->getType() != BC->getDestTy()) { | |
2006 | Val = new BitCastInst(Val, BC->getDestTy(), "", BC); | |
2007 | Val->takeName(BC); | |
2008 | } | |
2009 | BC->replaceAllUsesWith(Val); | |
2010 | DeadInsts.push_back(BC); | |
2011 | } | |
2012 | ||
2013 | /// FindElementAndOffset - Return the index of the element containing Offset | |
2014 | /// within the specified type, which must be either a struct or an array. | |
2015 | /// Sets T to the type of the element and Offset to the offset within that | |
2016 | /// element. IdxTy is set to the type of the index result to be used in a | |
2017 | /// GEP instruction. | |
2018 | uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, | |
2019 | Type *&IdxTy) { | |
2020 | uint64_t Idx = 0; | |
2021 | if (StructType *ST = dyn_cast<StructType>(T)) { | |
2022 | const StructLayout *Layout = TD->getStructLayout(ST); | |
2023 | Idx = Layout->getElementContainingOffset(Offset); | |
2024 | T = ST->getContainedType(Idx); | |
2025 | Offset -= Layout->getElementOffset(Idx); | |
2026 | IdxTy = Type::getInt32Ty(T->getContext()); | |
2027 | return Idx; | |
2028 | } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { | |
2029 | T = AT->getElementType(); | |
2030 | uint64_t EltSize = TD->getTypeAllocSize(T); | |
2031 | Idx = Offset / EltSize; | |
2032 | Offset -= Idx * EltSize; | |
2033 | IdxTy = Type::getInt64Ty(T->getContext()); | |
2034 | return Idx; | |
2035 | } | |
2036 | VectorType *VT = cast<VectorType>(T); | |
2037 | T = VT->getElementType(); | |
2038 | uint64_t EltSize = TD->getTypeAllocSize(T); | |
2039 | Idx = Offset / EltSize; | |
2040 | Offset -= Idx * EltSize; | |
2041 | IdxTy = Type::getInt64Ty(T->getContext()); | |
2042 | return Idx; | |
2043 | } | |
2044 | ||
2045 | /// RewriteGEP - Check if this GEP instruction moves the pointer across | |
2046 | /// elements of the alloca that are being split apart, and if so, rewrite | |
2047 | /// the GEP to be relative to the new element. | |
2048 | void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, | |
2049 | SmallVector<AllocaInst*, 32> &NewElts) { | |
2050 | uint64_t OldOffset = Offset; | |
2051 | SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); | |
2052 | // If the GEP was dynamic then it must have been a dynamic vector lookup. | |
2053 | // In this case, it must be the last GEP operand which is dynamic so keep that | |
2054 | // aside until we've found the constant GEP offset then add it back in at the | |
2055 | // end. | |
2056 | Value* NonConstantIdx = 0; | |
2057 | if (!GEPI->hasAllConstantIndices()) | |
2058 | NonConstantIdx = Indices.pop_back_val(); | |
2059 | Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); | |
2060 | ||
2061 | RewriteForScalarRepl(GEPI, AI, Offset, NewElts); | |
2062 | ||
2063 | Type *T = AI->getAllocatedType(); | |
2064 | Type *IdxTy; | |
2065 | uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); | |
2066 | if (GEPI->getOperand(0) == AI) | |
2067 | OldIdx = ~0ULL; // Force the GEP to be rewritten. | |
2068 | ||
2069 | T = AI->getAllocatedType(); | |
2070 | uint64_t EltOffset = Offset; | |
2071 | uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); | |
2072 | ||
2073 | // If this GEP does not move the pointer across elements of the alloca | |
2074 | // being split, then it does not needs to be rewritten. | |
2075 | if (Idx == OldIdx) | |
2076 | return; | |
2077 | ||
2078 | Type *i32Ty = Type::getInt32Ty(AI->getContext()); | |
2079 | SmallVector<Value*, 8> NewArgs; | |
2080 | NewArgs.push_back(Constant::getNullValue(i32Ty)); | |
2081 | while (EltOffset != 0) { | |
2082 | uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); | |
2083 | NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); | |
2084 | } | |
2085 | if (NonConstantIdx) { | |
2086 | Type* GepTy = T; | |
2087 | // This GEP has a dynamic index. We need to add "i32 0" to index through | |
2088 | // any structs or arrays in the original type until we get to the vector | |
2089 | // to index. | |
2090 | while (!isa<VectorType>(GepTy)) { | |
2091 | NewArgs.push_back(Constant::getNullValue(i32Ty)); | |
2092 | GepTy = cast<CompositeType>(GepTy)->getTypeAtIndex(0U); | |
2093 | } | |
2094 | NewArgs.push_back(NonConstantIdx); | |
2095 | } | |
2096 | Instruction *Val = NewElts[Idx]; | |
2097 | if (NewArgs.size() > 1) { | |
2098 | Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI); | |
2099 | Val->takeName(GEPI); | |
2100 | } | |
2101 | if (Val->getType() != GEPI->getType()) | |
2102 | Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); | |
2103 | GEPI->replaceAllUsesWith(Val); | |
2104 | DeadInsts.push_back(GEPI); | |
2105 | } | |
2106 | ||
2107 | /// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it | |
2108 | /// to mark the lifetime of the scalarized memory. | |
2109 | void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, | |
2110 | uint64_t Offset, | |
2111 | SmallVector<AllocaInst*, 32> &NewElts) { | |
2112 | ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0)); | |
2113 | // Put matching lifetime markers on everything from Offset up to | |
2114 | // Offset+OldSize. | |
2115 | Type *AIType = AI->getAllocatedType(); | |
2116 | uint64_t NewOffset = Offset; | |
2117 | Type *IdxTy; | |
2118 | uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy); | |
2119 | ||
2120 | IRBuilder<> Builder(II); | |
2121 | uint64_t Size = OldSize->getLimitedValue(); | |
2122 | ||
2123 | if (NewOffset) { | |
2124 | // Splice the first element and index 'NewOffset' bytes in. SROA will | |
2125 | // split the alloca again later. | |
2126 | Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy()); | |
2127 | V = Builder.CreateGEP(V, Builder.getInt64(NewOffset)); | |
2128 | ||
2129 | IdxTy = NewElts[Idx]->getAllocatedType(); | |
2130 | uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset; | |
2131 | if (EltSize > Size) { | |
2132 | EltSize = Size; | |
2133 | Size = 0; | |
2134 | } else { | |
2135 | Size -= EltSize; | |
2136 | } | |
2137 | if (II->getIntrinsicID() == Intrinsic::lifetime_start) | |
2138 | Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize)); | |
2139 | else | |
2140 | Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize)); | |
2141 | ++Idx; | |
2142 | } | |
2143 | ||
2144 | for (; Idx != NewElts.size() && Size; ++Idx) { | |
2145 | IdxTy = NewElts[Idx]->getAllocatedType(); | |
2146 | uint64_t EltSize = TD->getTypeAllocSize(IdxTy); | |
2147 | if (EltSize > Size) { | |
2148 | EltSize = Size; | |
2149 | Size = 0; | |
2150 | } else { | |
2151 | Size -= EltSize; | |
2152 | } | |
2153 | if (II->getIntrinsicID() == Intrinsic::lifetime_start) | |
2154 | Builder.CreateLifetimeStart(NewElts[Idx], | |
2155 | Builder.getInt64(EltSize)); | |
2156 | else | |
2157 | Builder.CreateLifetimeEnd(NewElts[Idx], | |
2158 | Builder.getInt64(EltSize)); | |
2159 | } | |
2160 | DeadInsts.push_back(II); | |
2161 | } | |
2162 | ||
2163 | /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. | |
2164 | /// Rewrite it to copy or set the elements of the scalarized memory. | |
2165 | void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, | |
2166 | AllocaInst *AI, | |
2167 | SmallVector<AllocaInst*, 32> &NewElts) { | |
2168 | // If this is a memcpy/memmove, construct the other pointer as the | |
2169 | // appropriate type. The "Other" pointer is the pointer that goes to memory | |
2170 | // that doesn't have anything to do with the alloca that we are promoting. For | |
2171 | // memset, this Value* stays null. | |
2172 | Value *OtherPtr = 0; | |
2173 | unsigned MemAlignment = MI->getAlignment(); | |
2174 | if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy | |
2175 | if (Inst == MTI->getRawDest()) | |
2176 | OtherPtr = MTI->getRawSource(); | |
2177 | else { | |
2178 | assert(Inst == MTI->getRawSource()); | |
2179 | OtherPtr = MTI->getRawDest(); | |
2180 | } | |
2181 | } | |
2182 | ||
2183 | // If there is an other pointer, we want to convert it to the same pointer | |
2184 | // type as AI has, so we can GEP through it safely. | |
2185 | if (OtherPtr) { | |
2186 | unsigned AddrSpace = | |
2187 | cast<PointerType>(OtherPtr->getType())->getAddressSpace(); | |
2188 | ||
2189 | // Remove bitcasts and all-zero GEPs from OtherPtr. This is an | |
2190 | // optimization, but it's also required to detect the corner case where | |
2191 | // both pointer operands are referencing the same memory, and where | |
2192 | // OtherPtr may be a bitcast or GEP that currently being rewritten. (This | |
2193 | // function is only called for mem intrinsics that access the whole | |
2194 | // aggregate, so non-zero GEPs are not an issue here.) | |
2195 | OtherPtr = OtherPtr->stripPointerCasts(); | |
2196 | ||
2197 | // Copying the alloca to itself is a no-op: just delete it. | |
2198 | if (OtherPtr == AI || OtherPtr == NewElts[0]) { | |
2199 | // This code will run twice for a no-op memcpy -- once for each operand. | |
2200 | // Put only one reference to MI on the DeadInsts list. | |
2201 | for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), | |
2202 | E = DeadInsts.end(); I != E; ++I) | |
2203 | if (*I == MI) return; | |
2204 | DeadInsts.push_back(MI); | |
2205 | return; | |
2206 | } | |
2207 | ||
2208 | // If the pointer is not the right type, insert a bitcast to the right | |
2209 | // type. | |
2210 | Type *NewTy = | |
2211 | PointerType::get(AI->getType()->getElementType(), AddrSpace); | |
2212 | ||
2213 | if (OtherPtr->getType() != NewTy) | |
2214 | OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); | |
2215 | } | |
2216 | ||
2217 | // Process each element of the aggregate. | |
2218 | bool SROADest = MI->getRawDest() == Inst; | |
2219 | ||
2220 | Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); | |
2221 | ||
2222 | for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { | |
2223 | // If this is a memcpy/memmove, emit a GEP of the other element address. | |
2224 | Value *OtherElt = 0; | |
2225 | unsigned OtherEltAlign = MemAlignment; | |
2226 | ||
2227 | if (OtherPtr) { | |
2228 | Value *Idx[2] = { Zero, | |
2229 | ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; | |
2230 | OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, | |
2231 | OtherPtr->getName()+"."+Twine(i), | |
2232 | MI); | |
2233 | uint64_t EltOffset; | |
2234 | PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); | |
2235 | Type *OtherTy = OtherPtrTy->getElementType(); | |
2236 | if (StructType *ST = dyn_cast<StructType>(OtherTy)) { | |
2237 | EltOffset = TD->getStructLayout(ST)->getElementOffset(i); | |
2238 | } else { | |
2239 | Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); | |
2240 | EltOffset = TD->getTypeAllocSize(EltTy)*i; | |
2241 | } | |
2242 | ||
2243 | // The alignment of the other pointer is the guaranteed alignment of the | |
2244 | // element, which is affected by both the known alignment of the whole | |
2245 | // mem intrinsic and the alignment of the element. If the alignment of | |
2246 | // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the | |
2247 | // known alignment is just 4 bytes. | |
2248 | OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); | |
2249 | } | |
2250 | ||
2251 | Value *EltPtr = NewElts[i]; | |
2252 | Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); | |
2253 | ||
2254 | // If we got down to a scalar, insert a load or store as appropriate. | |
2255 | if (EltTy->isSingleValueType()) { | |
2256 | if (isa<MemTransferInst>(MI)) { | |
2257 | if (SROADest) { | |
2258 | // From Other to Alloca. | |
2259 | Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); | |
2260 | new StoreInst(Elt, EltPtr, MI); | |
2261 | } else { | |
2262 | // From Alloca to Other. | |
2263 | Value *Elt = new LoadInst(EltPtr, "tmp", MI); | |
2264 | new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); | |
2265 | } | |
2266 | continue; | |
2267 | } | |
2268 | assert(isa<MemSetInst>(MI)); | |
2269 | ||
2270 | // If the stored element is zero (common case), just store a null | |
2271 | // constant. | |
2272 | Constant *StoreVal; | |
2273 | if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { | |
2274 | if (CI->isZero()) { | |
2275 | StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> | |
2276 | } else { | |
2277 | // If EltTy is a vector type, get the element type. | |
2278 | Type *ValTy = EltTy->getScalarType(); | |
2279 | ||
2280 | // Construct an integer with the right value. | |
2281 | unsigned EltSize = TD->getTypeSizeInBits(ValTy); | |
2282 | APInt OneVal(EltSize, CI->getZExtValue()); | |
2283 | APInt TotalVal(OneVal); | |
2284 | // Set each byte. | |
2285 | for (unsigned i = 0; 8*i < EltSize; ++i) { | |
2286 | TotalVal = TotalVal.shl(8); | |
2287 | TotalVal |= OneVal; | |
2288 | } | |
2289 | ||
2290 | // Convert the integer value to the appropriate type. | |
2291 | StoreVal = ConstantInt::get(CI->getContext(), TotalVal); | |
2292 | if (ValTy->isPointerTy()) | |
2293 | StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); | |
2294 | else if (ValTy->isFloatingPointTy()) | |
2295 | StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); | |
2296 | assert(StoreVal->getType() == ValTy && "Type mismatch!"); | |
2297 | ||
2298 | // If the requested value was a vector constant, create it. | |
2299 | if (EltTy->isVectorTy()) { | |
2300 | unsigned NumElts = cast<VectorType>(EltTy)->getNumElements(); | |
2301 | StoreVal = ConstantVector::getSplat(NumElts, StoreVal); | |
2302 | } | |
2303 | } | |
2304 | new StoreInst(StoreVal, EltPtr, MI); | |
2305 | continue; | |
2306 | } | |
2307 | // Otherwise, if we're storing a byte variable, use a memset call for | |
2308 | // this element. | |
2309 | } | |
2310 | ||
2311 | unsigned EltSize = TD->getTypeAllocSize(EltTy); | |
2312 | if (!EltSize) | |
2313 | continue; | |
2314 | ||
2315 | IRBuilder<> Builder(MI); | |
2316 | ||
2317 | // Finally, insert the meminst for this element. | |
2318 | if (isa<MemSetInst>(MI)) { | |
2319 | Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, | |
2320 | MI->isVolatile()); | |
2321 | } else { | |
2322 | assert(isa<MemTransferInst>(MI)); | |
2323 | Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr | |
2324 | Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr | |
2325 | ||
2326 | if (isa<MemCpyInst>(MI)) | |
2327 | Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); | |
2328 | else | |
2329 | Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); | |
2330 | } | |
2331 | } | |
2332 | DeadInsts.push_back(MI); | |
2333 | } | |
2334 | ||
2335 | /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that | |
2336 | /// overwrites the entire allocation. Extract out the pieces of the stored | |
2337 | /// integer and store them individually. | |
2338 | void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, | |
2339 | SmallVector<AllocaInst*, 32> &NewElts){ | |
2340 | // Extract each element out of the integer according to its structure offset | |
2341 | // and store the element value to the individual alloca. | |
2342 | Value *SrcVal = SI->getOperand(0); | |
2343 | Type *AllocaEltTy = AI->getAllocatedType(); | |
2344 | uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); | |
2345 | ||
2346 | IRBuilder<> Builder(SI); | |
2347 | ||
2348 | // Handle tail padding by extending the operand | |
2349 | if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) | |
2350 | SrcVal = Builder.CreateZExt(SrcVal, | |
2351 | IntegerType::get(SI->getContext(), AllocaSizeBits)); | |
2352 | ||
2353 | DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI | |
2354 | << '\n'); | |
2355 | ||
2356 | // There are two forms here: AI could be an array or struct. Both cases | |
2357 | // have different ways to compute the element offset. | |
2358 | if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { | |
2359 | const StructLayout *Layout = TD->getStructLayout(EltSTy); | |
2360 | ||
2361 | for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { | |
2362 | // Get the number of bits to shift SrcVal to get the value. | |
2363 | Type *FieldTy = EltSTy->getElementType(i); | |
2364 | uint64_t Shift = Layout->getElementOffsetInBits(i); | |
2365 | ||
2366 | if (TD->isBigEndian()) | |
2367 | Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); | |
2368 | ||
2369 | Value *EltVal = SrcVal; | |
2370 | if (Shift) { | |
2371 | Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); | |
2372 | EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); | |
2373 | } | |
2374 | ||
2375 | // Truncate down to an integer of the right size. | |
2376 | uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); | |
2377 | ||
2378 | // Ignore zero sized fields like {}, they obviously contain no data. | |
2379 | if (FieldSizeBits == 0) continue; | |
2380 | ||
2381 | if (FieldSizeBits != AllocaSizeBits) | |
2382 | EltVal = Builder.CreateTrunc(EltVal, | |
2383 | IntegerType::get(SI->getContext(), FieldSizeBits)); | |
2384 | Value *DestField = NewElts[i]; | |
2385 | if (EltVal->getType() == FieldTy) { | |
2386 | // Storing to an integer field of this size, just do it. | |
2387 | } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { | |
2388 | // Bitcast to the right element type (for fp/vector values). | |
2389 | EltVal = Builder.CreateBitCast(EltVal, FieldTy); | |
2390 | } else { | |
2391 | // Otherwise, bitcast the dest pointer (for aggregates). | |
2392 | DestField = Builder.CreateBitCast(DestField, | |
2393 | PointerType::getUnqual(EltVal->getType())); | |
2394 | } | |
2395 | new StoreInst(EltVal, DestField, SI); | |
2396 | } | |
2397 | ||
2398 | } else { | |
2399 | ArrayType *ATy = cast<ArrayType>(AllocaEltTy); | |
2400 | Type *ArrayEltTy = ATy->getElementType(); | |
2401 | uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); | |
2402 | uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); | |
2403 | ||
2404 | uint64_t Shift; | |
2405 | ||
2406 | if (TD->isBigEndian()) | |
2407 | Shift = AllocaSizeBits-ElementOffset; | |
2408 | else | |
2409 | Shift = 0; | |
2410 | ||
2411 | for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { | |
2412 | // Ignore zero sized fields like {}, they obviously contain no data. | |
2413 | if (ElementSizeBits == 0) continue; | |
2414 | ||
2415 | Value *EltVal = SrcVal; | |
2416 | if (Shift) { | |
2417 | Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); | |
2418 | EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); | |
2419 | } | |
2420 | ||
2421 | // Truncate down to an integer of the right size. | |
2422 | if (ElementSizeBits != AllocaSizeBits) | |
2423 | EltVal = Builder.CreateTrunc(EltVal, | |
2424 | IntegerType::get(SI->getContext(), | |
2425 | ElementSizeBits)); | |
2426 | Value *DestField = NewElts[i]; | |
2427 | if (EltVal->getType() == ArrayEltTy) { | |
2428 | // Storing to an integer field of this size, just do it. | |
2429 | } else if (ArrayEltTy->isFloatingPointTy() || | |
2430 | ArrayEltTy->isVectorTy()) { | |
2431 | // Bitcast to the right element type (for fp/vector values). | |
2432 | EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); | |
2433 | } else { | |
2434 | // Otherwise, bitcast the dest pointer (for aggregates). | |
2435 | DestField = Builder.CreateBitCast(DestField, | |
2436 | PointerType::getUnqual(EltVal->getType())); | |
2437 | } | |
2438 | new StoreInst(EltVal, DestField, SI); | |
2439 | ||
2440 | if (TD->isBigEndian()) | |
2441 | Shift -= ElementOffset; | |
2442 | else | |
2443 | Shift += ElementOffset; | |
2444 | } | |
2445 | } | |
2446 | ||
2447 | DeadInsts.push_back(SI); | |
2448 | } | |
2449 | ||
2450 | /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to | |
2451 | /// an integer. Load the individual pieces to form the aggregate value. | |
2452 | void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, | |
2453 | SmallVector<AllocaInst*, 32> &NewElts) { | |
2454 | // Extract each element out of the NewElts according to its structure offset | |
2455 | // and form the result value. | |
2456 | Type *AllocaEltTy = AI->getAllocatedType(); | |
2457 | uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); | |
2458 | ||
2459 | DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI | |
2460 | << '\n'); | |
2461 | ||
2462 | // There are two forms here: AI could be an array or struct. Both cases | |
2463 | // have different ways to compute the element offset. | |
2464 | const StructLayout *Layout = 0; | |
2465 | uint64_t ArrayEltBitOffset = 0; | |
2466 | if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { | |
2467 | Layout = TD->getStructLayout(EltSTy); | |
2468 | } else { | |
2469 | Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); | |
2470 | ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); | |
2471 | } | |
2472 | ||
2473 | Value *ResultVal = | |
2474 | Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); | |
2475 | ||
2476 | for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { | |
2477 | // Load the value from the alloca. If the NewElt is an aggregate, cast | |
2478 | // the pointer to an integer of the same size before doing the load. | |
2479 | Value *SrcField = NewElts[i]; | |
2480 | Type *FieldTy = | |
2481 | cast<PointerType>(SrcField->getType())->getElementType(); | |
2482 | uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); | |
2483 | ||
2484 | // Ignore zero sized fields like {}, they obviously contain no data. | |
2485 | if (FieldSizeBits == 0) continue; | |
2486 | ||
2487 | IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), | |
2488 | FieldSizeBits); | |
2489 | if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && | |
2490 | !FieldTy->isVectorTy()) | |
2491 | SrcField = new BitCastInst(SrcField, | |
2492 | PointerType::getUnqual(FieldIntTy), | |
2493 | "", LI); | |
2494 | SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); | |
2495 | ||
2496 | // If SrcField is a fp or vector of the right size but that isn't an | |
2497 | // integer type, bitcast to an integer so we can shift it. | |
2498 | if (SrcField->getType() != FieldIntTy) | |
2499 | SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); | |
2500 | ||
2501 | // Zero extend the field to be the same size as the final alloca so that | |
2502 | // we can shift and insert it. | |
2503 | if (SrcField->getType() != ResultVal->getType()) | |
2504 | SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); | |
2505 | ||
2506 | // Determine the number of bits to shift SrcField. | |
2507 | uint64_t Shift; | |
2508 | if (Layout) // Struct case. | |
2509 | Shift = Layout->getElementOffsetInBits(i); | |
2510 | else // Array case. | |
2511 | Shift = i*ArrayEltBitOffset; | |
2512 | ||
2513 | if (TD->isBigEndian()) | |
2514 | Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); | |
2515 | ||
2516 | if (Shift) { | |
2517 | Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); | |
2518 | SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); | |
2519 | } | |
2520 | ||
2521 | // Don't create an 'or x, 0' on the first iteration. | |
2522 | if (!isa<Constant>(ResultVal) || | |
2523 | !cast<Constant>(ResultVal)->isNullValue()) | |
2524 | ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); | |
2525 | else | |
2526 | ResultVal = SrcField; | |
2527 | } | |
2528 | ||
2529 | // Handle tail padding by truncating the result | |
2530 | if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) | |
2531 | ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); | |
2532 | ||
2533 | LI->replaceAllUsesWith(ResultVal); | |
2534 | DeadInsts.push_back(LI); | |
2535 | } | |
2536 | ||
2537 | /// HasPadding - Return true if the specified type has any structure or | |
2538 | /// alignment padding in between the elements that would be split apart | |
2539 | /// by SROA; return false otherwise. | |
970d7e83 | 2540 | static bool HasPadding(Type *Ty, const DataLayout &TD) { |
223e47cc LB |
2541 | if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { |
2542 | Ty = ATy->getElementType(); | |
2543 | return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); | |
2544 | } | |
2545 | ||
2546 | // SROA currently handles only Arrays and Structs. | |
2547 | StructType *STy = cast<StructType>(Ty); | |
2548 | const StructLayout *SL = TD.getStructLayout(STy); | |
2549 | unsigned PrevFieldBitOffset = 0; | |
2550 | for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { | |
2551 | unsigned FieldBitOffset = SL->getElementOffsetInBits(i); | |
2552 | ||
2553 | // Check to see if there is any padding between this element and the | |
2554 | // previous one. | |
2555 | if (i) { | |
2556 | unsigned PrevFieldEnd = | |
2557 | PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); | |
2558 | if (PrevFieldEnd < FieldBitOffset) | |
2559 | return true; | |
2560 | } | |
2561 | PrevFieldBitOffset = FieldBitOffset; | |
2562 | } | |
2563 | // Check for tail padding. | |
2564 | if (unsigned EltCount = STy->getNumElements()) { | |
2565 | unsigned PrevFieldEnd = PrevFieldBitOffset + | |
2566 | TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); | |
2567 | if (PrevFieldEnd < SL->getSizeInBits()) | |
2568 | return true; | |
2569 | } | |
2570 | return false; | |
2571 | } | |
2572 | ||
2573 | /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of | |
2574 | /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, | |
2575 | /// or 1 if safe after canonicalization has been performed. | |
2576 | bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { | |
2577 | // Loop over the use list of the alloca. We can only transform it if all of | |
2578 | // the users are safe to transform. | |
2579 | AllocaInfo Info(AI); | |
2580 | ||
2581 | isSafeForScalarRepl(AI, 0, Info); | |
2582 | if (Info.isUnsafe) { | |
2583 | DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); | |
2584 | return false; | |
2585 | } | |
2586 | ||
2587 | // Okay, we know all the users are promotable. If the aggregate is a memcpy | |
2588 | // source and destination, we have to be careful. In particular, the memcpy | |
2589 | // could be moving around elements that live in structure padding of the LLVM | |
2590 | // types, but may actually be used. In these cases, we refuse to promote the | |
2591 | // struct. | |
2592 | if (Info.isMemCpySrc && Info.isMemCpyDst && | |
2593 | HasPadding(AI->getAllocatedType(), *TD)) | |
2594 | return false; | |
2595 | ||
2596 | // If the alloca never has an access to just *part* of it, but is accessed | |
2597 | // via loads and stores, then we should use ConvertToScalarInfo to promote | |
2598 | // the alloca instead of promoting each piece at a time and inserting fission | |
2599 | // and fusion code. | |
2600 | if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { | |
2601 | // If the struct/array just has one element, use basic SRoA. | |
2602 | if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { | |
2603 | if (ST->getNumElements() > 1) return false; | |
2604 | } else { | |
2605 | if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) | |
2606 | return false; | |
2607 | } | |
2608 | } | |
2609 | ||
2610 | return true; | |
2611 | } |