ScalarReplAggregates.cpp revision 1d0be15f89cb5056e20e2d24faa8d6afb1573bca
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
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"
24#include "llvm/Constants.h"
25#include "llvm/DerivedTypes.h"
26#include "llvm/Function.h"
27#include "llvm/GlobalVariable.h"
28#include "llvm/Instructions.h"
29#include "llvm/IntrinsicInst.h"
30#include "llvm/LLVMContext.h"
31#include "llvm/Pass.h"
32#include "llvm/Analysis/Dominators.h"
33#include "llvm/Target/TargetData.h"
34#include "llvm/Transforms/Utils/PromoteMemToReg.h"
35#include "llvm/Transforms/Utils/Local.h"
36#include "llvm/Support/Debug.h"
37#include "llvm/Support/ErrorHandling.h"
38#include "llvm/Support/GetElementPtrTypeIterator.h"
39#include "llvm/Support/IRBuilder.h"
40#include "llvm/Support/MathExtras.h"
41#include "llvm/Support/Compiler.h"
42#include "llvm/ADT/SmallVector.h"
43#include "llvm/ADT/Statistic.h"
44using namespace llvm;
45
46STATISTIC(NumReplaced,  "Number of allocas broken up");
47STATISTIC(NumPromoted,  "Number of allocas promoted");
48STATISTIC(NumConverted, "Number of aggregates converted to scalar");
49STATISTIC(NumGlobals,   "Number of allocas copied from constant global");
50
51namespace {
52  struct VISIBILITY_HIDDEN SROA : public FunctionPass {
53    static char ID; // Pass identification, replacement for typeid
54    explicit SROA(signed T = -1) : FunctionPass(&ID) {
55      if (T == -1)
56        SRThreshold = 128;
57      else
58        SRThreshold = T;
59    }
60
61    bool runOnFunction(Function &F);
62
63    bool performScalarRepl(Function &F);
64    bool performPromotion(Function &F);
65
66    // getAnalysisUsage - This pass does not require any passes, but we know it
67    // will not alter the CFG, so say so.
68    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
69      AU.addRequired<DominatorTree>();
70      AU.addRequired<DominanceFrontier>();
71      AU.addRequired<TargetData>();
72      AU.setPreservesCFG();
73    }
74
75  private:
76    TargetData *TD;
77
78    /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
79    /// information about the uses.  All these fields are initialized to false
80    /// and set to true when something is learned.
81    struct AllocaInfo {
82      /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
83      bool isUnsafe : 1;
84
85      /// needsCleanup - This is set to true if there is some use of the alloca
86      /// that requires cleanup.
87      bool needsCleanup : 1;
88
89      /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
90      bool isMemCpySrc : 1;
91
92      /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
93      bool isMemCpyDst : 1;
94
95      AllocaInfo()
96        : isUnsafe(false), needsCleanup(false),
97          isMemCpySrc(false), isMemCpyDst(false) {}
98    };
99
100    unsigned SRThreshold;
101
102    void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
103
104    int isSafeAllocaToScalarRepl(AllocationInst *AI);
105
106    void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
107                               AllocaInfo &Info);
108    void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
109                         AllocaInfo &Info);
110    void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
111                                        unsigned OpNo, AllocaInfo &Info);
112    void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI,
113                                        AllocaInfo &Info);
114
115    void DoScalarReplacement(AllocationInst *AI,
116                             std::vector<AllocationInst*> &WorkList);
117    void CleanupGEP(GetElementPtrInst *GEP);
118    void CleanupAllocaUsers(AllocationInst *AI);
119    AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
120
121    void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
122                                    SmallVector<AllocaInst*, 32> &NewElts);
123
124    void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
125                                      AllocationInst *AI,
126                                      SmallVector<AllocaInst*, 32> &NewElts);
127    void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI,
128                                       SmallVector<AllocaInst*, 32> &NewElts);
129    void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
130                                      SmallVector<AllocaInst*, 32> &NewElts);
131
132    bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
133                            bool &SawVec, uint64_t Offset, unsigned AllocaSize);
134    void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
135    Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
136                                     uint64_t Offset, IRBuilder<> &Builder);
137    Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
138                                     uint64_t Offset, IRBuilder<> &Builder);
139    static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI);
140  };
141}
142
143char SROA::ID = 0;
144static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
145
146// Public interface to the ScalarReplAggregates pass
147FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) {
148  return new SROA(Threshold);
149}
150
151
152bool SROA::runOnFunction(Function &F) {
153  TD = &getAnalysis<TargetData>();
154
155  bool Changed = performPromotion(F);
156  while (1) {
157    bool LocalChange = performScalarRepl(F);
158    if (!LocalChange) break;   // No need to repromote if no scalarrepl
159    Changed = true;
160    LocalChange = performPromotion(F);
161    if (!LocalChange) break;   // No need to re-scalarrepl if no promotion
162  }
163
164  return Changed;
165}
166
167
168bool SROA::performPromotion(Function &F) {
169  std::vector<AllocaInst*> Allocas;
170  DominatorTree         &DT = getAnalysis<DominatorTree>();
171  DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
172
173  BasicBlock &BB = F.getEntryBlock();  // Get the entry node for the function
174
175  bool Changed = false;
176
177  while (1) {
178    Allocas.clear();
179
180    // Find allocas that are safe to promote, by looking at all instructions in
181    // the entry node
182    for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
183      if (AllocaInst *AI = dyn_cast<AllocaInst>(I))       // Is it an alloca?
184        if (isAllocaPromotable(AI))
185          Allocas.push_back(AI);
186
187    if (Allocas.empty()) break;
188
189    PromoteMemToReg(Allocas, DT, DF, F.getContext());
190    NumPromoted += Allocas.size();
191    Changed = true;
192  }
193
194  return Changed;
195}
196
197/// getNumSAElements - Return the number of elements in the specific struct or
198/// array.
199static uint64_t getNumSAElements(const Type *T) {
200  if (const StructType *ST = dyn_cast<StructType>(T))
201    return ST->getNumElements();
202  return cast<ArrayType>(T)->getNumElements();
203}
204
205// performScalarRepl - This algorithm is a simple worklist driven algorithm,
206// which runs on all of the malloc/alloca instructions in the function, removing
207// them if they are only used by getelementptr instructions.
208//
209bool SROA::performScalarRepl(Function &F) {
210  std::vector<AllocationInst*> WorkList;
211
212  // Scan the entry basic block, adding any alloca's and mallocs to the worklist
213  BasicBlock &BB = F.getEntryBlock();
214  for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
215    if (AllocationInst *A = dyn_cast<AllocationInst>(I))
216      WorkList.push_back(A);
217
218  // Process the worklist
219  bool Changed = false;
220  while (!WorkList.empty()) {
221    AllocationInst *AI = WorkList.back();
222    WorkList.pop_back();
223
224    // Handle dead allocas trivially.  These can be formed by SROA'ing arrays
225    // with unused elements.
226    if (AI->use_empty()) {
227      AI->eraseFromParent();
228      continue;
229    }
230
231    // If this alloca is impossible for us to promote, reject it early.
232    if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
233      continue;
234
235    // Check to see if this allocation is only modified by a memcpy/memmove from
236    // a constant global.  If this is the case, we can change all users to use
237    // the constant global instead.  This is commonly produced by the CFE by
238    // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
239    // is only subsequently read.
240    if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
241      DOUT << "Found alloca equal to global: " << *AI;
242      DOUT << "  memcpy = " << *TheCopy;
243      Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2));
244      AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
245      TheCopy->eraseFromParent();  // Don't mutate the global.
246      AI->eraseFromParent();
247      ++NumGlobals;
248      Changed = true;
249      continue;
250    }
251
252    // Check to see if we can perform the core SROA transformation.  We cannot
253    // transform the allocation instruction if it is an array allocation
254    // (allocations OF arrays are ok though), and an allocation of a scalar
255    // value cannot be decomposed at all.
256    uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
257
258    // Do not promote any struct whose size is too big.
259    if (AllocaSize > SRThreshold) continue;
260
261    if ((isa<StructType>(AI->getAllocatedType()) ||
262         isa<ArrayType>(AI->getAllocatedType())) &&
263        // Do not promote any struct into more than "32" separate vars.
264        getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) {
265      // Check that all of the users of the allocation are capable of being
266      // transformed.
267      switch (isSafeAllocaToScalarRepl(AI)) {
268      default: llvm_unreachable("Unexpected value!");
269      case 0:  // Not safe to scalar replace.
270        break;
271      case 1:  // Safe, but requires cleanup/canonicalizations first
272        CleanupAllocaUsers(AI);
273        // FALL THROUGH.
274      case 3:  // Safe to scalar replace.
275        DoScalarReplacement(AI, WorkList);
276        Changed = true;
277        continue;
278      }
279    }
280
281    // If we can turn this aggregate value (potentially with casts) into a
282    // simple scalar value that can be mem2reg'd into a register value.
283    // IsNotTrivial tracks whether this is something that mem2reg could have
284    // promoted itself.  If so, we don't want to transform it needlessly.  Note
285    // that we can't just check based on the type: the alloca may be of an i32
286    // but that has pointer arithmetic to set byte 3 of it or something.
287    bool IsNotTrivial = false;
288    const Type *VectorTy = 0;
289    bool HadAVector = false;
290    if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector,
291                           0, unsigned(AllocaSize)) && IsNotTrivial) {
292      AllocaInst *NewAI;
293      // If we were able to find a vector type that can handle this with
294      // insert/extract elements, and if there was at least one use that had
295      // a vector type, promote this to a vector.  We don't want to promote
296      // random stuff that doesn't use vectors (e.g. <9 x double>) because then
297      // we just get a lot of insert/extracts.  If at least one vector is
298      // involved, then we probably really do have a union of vector/array.
299      if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) {
300        DOUT << "CONVERT TO VECTOR: " << *AI << "  TYPE = " << *VectorTy <<"\n";
301
302        // Create and insert the vector alloca.
303        NewAI = new AllocaInst(VectorTy, 0, "",  AI->getParent()->begin());
304        ConvertUsesToScalar(AI, NewAI, 0);
305      } else {
306        DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n";
307
308        // Create and insert the integer alloca.
309        const Type *NewTy = IntegerType::get(AI->getContext(), AllocaSize*8);
310        NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
311        ConvertUsesToScalar(AI, NewAI, 0);
312      }
313      NewAI->takeName(AI);
314      AI->eraseFromParent();
315      ++NumConverted;
316      Changed = true;
317      continue;
318    }
319
320    // Otherwise, couldn't process this alloca.
321  }
322
323  return Changed;
324}
325
326/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
327/// predicate, do SROA now.
328void SROA::DoScalarReplacement(AllocationInst *AI,
329                               std::vector<AllocationInst*> &WorkList) {
330  DOUT << "Found inst to SROA: " << *AI;
331  SmallVector<AllocaInst*, 32> ElementAllocas;
332  if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
333    ElementAllocas.reserve(ST->getNumContainedTypes());
334    for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
335      AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
336                                      AI->getAlignment(),
337                                      AI->getName() + "." + Twine(i), AI);
338      ElementAllocas.push_back(NA);
339      WorkList.push_back(NA);  // Add to worklist for recursive processing
340    }
341  } else {
342    const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
343    ElementAllocas.reserve(AT->getNumElements());
344    const Type *ElTy = AT->getElementType();
345    for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
346      AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
347                                      AI->getName() + "." + Twine(i), AI);
348      ElementAllocas.push_back(NA);
349      WorkList.push_back(NA);  // Add to worklist for recursive processing
350    }
351  }
352
353  // Now that we have created the alloca instructions that we want to use,
354  // expand the getelementptr instructions to use them.
355  //
356  while (!AI->use_empty()) {
357    Instruction *User = cast<Instruction>(AI->use_back());
358    if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) {
359      RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas);
360      BCInst->eraseFromParent();
361      continue;
362    }
363
364    // Replace:
365    //   %res = load { i32, i32 }* %alloc
366    // with:
367    //   %load.0 = load i32* %alloc.0
368    //   %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
369    //   %load.1 = load i32* %alloc.1
370    //   %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
371    // (Also works for arrays instead of structs)
372    if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
373      Value *Insert = UndefValue::get(LI->getType());
374      for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
375        Value *Load = new LoadInst(ElementAllocas[i], "load", LI);
376        Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
377      }
378      LI->replaceAllUsesWith(Insert);
379      LI->eraseFromParent();
380      continue;
381    }
382
383    // Replace:
384    //   store { i32, i32 } %val, { i32, i32 }* %alloc
385    // with:
386    //   %val.0 = extractvalue { i32, i32 } %val, 0
387    //   store i32 %val.0, i32* %alloc.0
388    //   %val.1 = extractvalue { i32, i32 } %val, 1
389    //   store i32 %val.1, i32* %alloc.1
390    // (Also works for arrays instead of structs)
391    if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
392      Value *Val = SI->getOperand(0);
393      for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
394        Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
395        new StoreInst(Extract, ElementAllocas[i], SI);
396      }
397      SI->eraseFromParent();
398      continue;
399    }
400
401    GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
402    // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
403    unsigned Idx =
404       (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
405
406    assert(Idx < ElementAllocas.size() && "Index out of range?");
407    AllocaInst *AllocaToUse = ElementAllocas[Idx];
408
409    Value *RepValue;
410    if (GEPI->getNumOperands() == 3) {
411      // Do not insert a new getelementptr instruction with zero indices, only
412      // to have it optimized out later.
413      RepValue = AllocaToUse;
414    } else {
415      // We are indexing deeply into the structure, so we still need a
416      // getelement ptr instruction to finish the indexing.  This may be
417      // expanded itself once the worklist is rerun.
418      //
419      SmallVector<Value*, 8> NewArgs;
420      NewArgs.push_back(Constant::getNullValue(
421                                           Type::getInt32Ty(AI->getContext())));
422      NewArgs.append(GEPI->op_begin()+3, GEPI->op_end());
423      RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(),
424                                           NewArgs.end(), "", GEPI);
425      RepValue->takeName(GEPI);
426    }
427
428    // If this GEP is to the start of the aggregate, check for memcpys.
429    if (Idx == 0 && GEPI->hasAllZeroIndices())
430      RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas);
431
432    // Move all of the users over to the new GEP.
433    GEPI->replaceAllUsesWith(RepValue);
434    // Delete the old GEP
435    GEPI->eraseFromParent();
436  }
437
438  // Finally, delete the Alloca instruction
439  AI->eraseFromParent();
440  NumReplaced++;
441}
442
443
444/// isSafeElementUse - Check to see if this use is an allowed use for a
445/// getelementptr instruction of an array aggregate allocation.  isFirstElt
446/// indicates whether Ptr is known to the start of the aggregate.
447///
448void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
449                            AllocaInfo &Info) {
450  for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
451       I != E; ++I) {
452    Instruction *User = cast<Instruction>(*I);
453    switch (User->getOpcode()) {
454    case Instruction::Load:  break;
455    case Instruction::Store:
456      // Store is ok if storing INTO the pointer, not storing the pointer
457      if (User->getOperand(0) == Ptr) return MarkUnsafe(Info);
458      break;
459    case Instruction::GetElementPtr: {
460      GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
461      bool AreAllZeroIndices = isFirstElt;
462      if (GEP->getNumOperands() > 1) {
463        if (!isa<ConstantInt>(GEP->getOperand(1)) ||
464            !cast<ConstantInt>(GEP->getOperand(1))->isZero())
465          // Using pointer arithmetic to navigate the array.
466          return MarkUnsafe(Info);
467
468        if (AreAllZeroIndices)
469          AreAllZeroIndices = GEP->hasAllZeroIndices();
470      }
471      isSafeElementUse(GEP, AreAllZeroIndices, AI, Info);
472      if (Info.isUnsafe) return;
473      break;
474    }
475    case Instruction::BitCast:
476      if (isFirstElt) {
477        isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info);
478        if (Info.isUnsafe) return;
479        break;
480      }
481      DOUT << "  Transformation preventing inst: " << *User;
482      return MarkUnsafe(Info);
483    case Instruction::Call:
484      if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
485        if (isFirstElt) {
486          isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info);
487          if (Info.isUnsafe) return;
488          break;
489        }
490      }
491      DOUT << "  Transformation preventing inst: " << *User;
492      return MarkUnsafe(Info);
493    default:
494      DOUT << "  Transformation preventing inst: " << *User;
495      return MarkUnsafe(Info);
496    }
497  }
498  return;  // All users look ok :)
499}
500
501/// AllUsersAreLoads - Return true if all users of this value are loads.
502static bool AllUsersAreLoads(Value *Ptr) {
503  for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
504       I != E; ++I)
505    if (cast<Instruction>(*I)->getOpcode() != Instruction::Load)
506      return false;
507  return true;
508}
509
510/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
511/// aggregate allocation.
512///
513void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
514                                 AllocaInfo &Info) {
515  if (BitCastInst *C = dyn_cast<BitCastInst>(User))
516    return isSafeUseOfBitCastedAllocation(C, AI, Info);
517
518  if (LoadInst *LI = dyn_cast<LoadInst>(User))
519    if (!LI->isVolatile())
520      return;// Loads (returning a first class aggregrate) are always rewritable
521
522  if (StoreInst *SI = dyn_cast<StoreInst>(User))
523    if (!SI->isVolatile() && SI->getOperand(0) != AI)
524      return;// Store is ok if storing INTO the pointer, not storing the pointer
525
526  GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User);
527  if (GEPI == 0)
528    return MarkUnsafe(Info);
529
530  gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI);
531
532  // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>".
533  if (I == E ||
534      I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) {
535    return MarkUnsafe(Info);
536  }
537
538  ++I;
539  if (I == E) return MarkUnsafe(Info);  // ran out of GEP indices??
540
541  bool IsAllZeroIndices = true;
542
543  // If the first index is a non-constant index into an array, see if we can
544  // handle it as a special case.
545  if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
546    if (!isa<ConstantInt>(I.getOperand())) {
547      IsAllZeroIndices = 0;
548      uint64_t NumElements = AT->getNumElements();
549
550      // If this is an array index and the index is not constant, we cannot
551      // promote... that is unless the array has exactly one or two elements in
552      // it, in which case we CAN promote it, but we have to canonicalize this
553      // out if this is the only problem.
554      if ((NumElements == 1 || NumElements == 2) &&
555          AllUsersAreLoads(GEPI)) {
556        Info.needsCleanup = true;
557        return;  // Canonicalization required!
558      }
559      return MarkUnsafe(Info);
560    }
561  }
562
563  // Walk through the GEP type indices, checking the types that this indexes
564  // into.
565  for (; I != E; ++I) {
566    // Ignore struct elements, no extra checking needed for these.
567    if (isa<StructType>(*I))
568      continue;
569
570    ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand());
571    if (!IdxVal) return MarkUnsafe(Info);
572
573    // Are all indices still zero?
574    IsAllZeroIndices &= IdxVal->isZero();
575
576    if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
577      // This GEP indexes an array.  Verify that this is an in-range constant
578      // integer. Specifically, consider A[0][i]. We cannot know that the user
579      // isn't doing invalid things like allowing i to index an out-of-range
580      // subscript that accesses A[1].  Because of this, we have to reject SROA
581      // of any accesses into structs where any of the components are variables.
582      if (IdxVal->getZExtValue() >= AT->getNumElements())
583        return MarkUnsafe(Info);
584    } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) {
585      if (IdxVal->getZExtValue() >= VT->getNumElements())
586        return MarkUnsafe(Info);
587    }
588  }
589
590  // If there are any non-simple uses of this getelementptr, make sure to reject
591  // them.
592  return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info);
593}
594
595/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory
596/// intrinsic can be promoted by SROA.  At this point, we know that the operand
597/// of the memintrinsic is a pointer to the beginning of the allocation.
598void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
599                                          unsigned OpNo, AllocaInfo &Info) {
600  // If not constant length, give up.
601  ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
602  if (!Length) return MarkUnsafe(Info);
603
604  // If not the whole aggregate, give up.
605  if (Length->getZExtValue() !=
606      TD->getTypeAllocSize(AI->getType()->getElementType()))
607    return MarkUnsafe(Info);
608
609  // We only know about memcpy/memset/memmove.
610  if (!isa<MemIntrinsic>(MI))
611    return MarkUnsafe(Info);
612
613  // Otherwise, we can transform it.  Determine whether this is a memcpy/set
614  // into or out of the aggregate.
615  if (OpNo == 1)
616    Info.isMemCpyDst = true;
617  else {
618    assert(OpNo == 2);
619    Info.isMemCpySrc = true;
620  }
621}
622
623/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast
624/// are
625void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI,
626                                          AllocaInfo &Info) {
627  for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end();
628       UI != E; ++UI) {
629    if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) {
630      isSafeUseOfBitCastedAllocation(BCU, AI, Info);
631    } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) {
632      isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info);
633    } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
634      if (SI->isVolatile())
635        return MarkUnsafe(Info);
636
637      // If storing the entire alloca in one chunk through a bitcasted pointer
638      // to integer, we can transform it.  This happens (for example) when you
639      // cast a {i32,i32}* to i64* and store through it.  This is similar to the
640      // memcpy case and occurs in various "byval" cases and emulated memcpys.
641      if (isa<IntegerType>(SI->getOperand(0)->getType()) &&
642          TD->getTypeAllocSize(SI->getOperand(0)->getType()) ==
643          TD->getTypeAllocSize(AI->getType()->getElementType())) {
644        Info.isMemCpyDst = true;
645        continue;
646      }
647      return MarkUnsafe(Info);
648    } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
649      if (LI->isVolatile())
650        return MarkUnsafe(Info);
651
652      // If loading the entire alloca in one chunk through a bitcasted pointer
653      // to integer, we can transform it.  This happens (for example) when you
654      // cast a {i32,i32}* to i64* and load through it.  This is similar to the
655      // memcpy case and occurs in various "byval" cases and emulated memcpys.
656      if (isa<IntegerType>(LI->getType()) &&
657          TD->getTypeAllocSize(LI->getType()) ==
658          TD->getTypeAllocSize(AI->getType()->getElementType())) {
659        Info.isMemCpySrc = true;
660        continue;
661      }
662      return MarkUnsafe(Info);
663    } else if (isa<DbgInfoIntrinsic>(UI)) {
664      // If one user is DbgInfoIntrinsic then check if all users are
665      // DbgInfoIntrinsics.
666      if (OnlyUsedByDbgInfoIntrinsics(BC)) {
667        Info.needsCleanup = true;
668        return;
669      }
670      else
671        MarkUnsafe(Info);
672    }
673    else {
674      return MarkUnsafe(Info);
675    }
676    if (Info.isUnsafe) return;
677  }
678}
679
680/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes
681/// to its first element.  Transform users of the cast to use the new values
682/// instead.
683void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
684                                      SmallVector<AllocaInst*, 32> &NewElts) {
685  Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end();
686  while (UI != UE) {
687    Instruction *User = cast<Instruction>(*UI++);
688    if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) {
689      RewriteBitCastUserOfAlloca(BCU, AI, NewElts);
690      if (BCU->use_empty()) BCU->eraseFromParent();
691      continue;
692    }
693
694    if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
695      // This must be memcpy/memmove/memset of the entire aggregate.
696      // Split into one per element.
697      RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts);
698      continue;
699    }
700
701    if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
702      // If this is a store of the entire alloca from an integer, rewrite it.
703      RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
704      continue;
705    }
706
707    if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
708      // If this is a load of the entire alloca to an integer, rewrite it.
709      RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
710      continue;
711    }
712
713    // Otherwise it must be some other user of a gep of the first pointer.  Just
714    // leave these alone.
715    continue;
716  }
717}
718
719/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
720/// Rewrite it to copy or set the elements of the scalarized memory.
721void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
722                                        AllocationInst *AI,
723                                        SmallVector<AllocaInst*, 32> &NewElts) {
724
725  // If this is a memcpy/memmove, construct the other pointer as the
726  // appropriate type.  The "Other" pointer is the pointer that goes to memory
727  // that doesn't have anything to do with the alloca that we are promoting. For
728  // memset, this Value* stays null.
729  Value *OtherPtr = 0;
730  LLVMContext &Context = MI->getContext();
731  unsigned MemAlignment = MI->getAlignment();
732  if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
733    if (BCInst == MTI->getRawDest())
734      OtherPtr = MTI->getRawSource();
735    else {
736      assert(BCInst == MTI->getRawSource());
737      OtherPtr = MTI->getRawDest();
738    }
739  }
740
741  // If there is an other pointer, we want to convert it to the same pointer
742  // type as AI has, so we can GEP through it safely.
743  if (OtherPtr) {
744    // It is likely that OtherPtr is a bitcast, if so, remove it.
745    if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr))
746      OtherPtr = BC->getOperand(0);
747    // All zero GEPs are effectively bitcasts.
748    if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr))
749      if (GEP->hasAllZeroIndices())
750        OtherPtr = GEP->getOperand(0);
751
752    if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr))
753      if (BCE->getOpcode() == Instruction::BitCast)
754        OtherPtr = BCE->getOperand(0);
755
756    // If the pointer is not the right type, insert a bitcast to the right
757    // type.
758    if (OtherPtr->getType() != AI->getType())
759      OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(),
760                                 MI);
761  }
762
763  // Process each element of the aggregate.
764  Value *TheFn = MI->getOperand(0);
765  const Type *BytePtrTy = MI->getRawDest()->getType();
766  bool SROADest = MI->getRawDest() == BCInst;
767
768  Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
769
770  for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
771    // If this is a memcpy/memmove, emit a GEP of the other element address.
772    Value *OtherElt = 0;
773    unsigned OtherEltAlign = MemAlignment;
774
775    if (OtherPtr) {
776      Value *Idx[2] = { Zero,
777                      ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
778      OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2,
779                                           OtherPtr->getNameStr()+"."+Twine(i),
780                                           MI);
781      uint64_t EltOffset;
782      const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
783      if (const StructType *ST =
784            dyn_cast<StructType>(OtherPtrTy->getElementType())) {
785        EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
786      } else {
787        const Type *EltTy =
788          cast<SequentialType>(OtherPtr->getType())->getElementType();
789        EltOffset = TD->getTypeAllocSize(EltTy)*i;
790      }
791
792      // The alignment of the other pointer is the guaranteed alignment of the
793      // element, which is affected by both the known alignment of the whole
794      // mem intrinsic and the alignment of the element.  If the alignment of
795      // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
796      // known alignment is just 4 bytes.
797      OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
798    }
799
800    Value *EltPtr = NewElts[i];
801    const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
802
803    // If we got down to a scalar, insert a load or store as appropriate.
804    if (EltTy->isSingleValueType()) {
805      if (isa<MemTransferInst>(MI)) {
806        if (SROADest) {
807          // From Other to Alloca.
808          Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
809          new StoreInst(Elt, EltPtr, MI);
810        } else {
811          // From Alloca to Other.
812          Value *Elt = new LoadInst(EltPtr, "tmp", MI);
813          new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
814        }
815        continue;
816      }
817      assert(isa<MemSetInst>(MI));
818
819      // If the stored element is zero (common case), just store a null
820      // constant.
821      Constant *StoreVal;
822      if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) {
823        if (CI->isZero()) {
824          StoreVal = Constant::getNullValue(EltTy);  // 0.0, null, 0, <0,0>
825        } else {
826          // If EltTy is a vector type, get the element type.
827          const Type *ValTy = EltTy->getScalarType();
828
829          // Construct an integer with the right value.
830          unsigned EltSize = TD->getTypeSizeInBits(ValTy);
831          APInt OneVal(EltSize, CI->getZExtValue());
832          APInt TotalVal(OneVal);
833          // Set each byte.
834          for (unsigned i = 0; 8*i < EltSize; ++i) {
835            TotalVal = TotalVal.shl(8);
836            TotalVal |= OneVal;
837          }
838
839          // Convert the integer value to the appropriate type.
840          StoreVal = ConstantInt::get(Context, TotalVal);
841          if (isa<PointerType>(ValTy))
842            StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
843          else if (ValTy->isFloatingPoint())
844            StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
845          assert(StoreVal->getType() == ValTy && "Type mismatch!");
846
847          // If the requested value was a vector constant, create it.
848          if (EltTy != ValTy) {
849            unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
850            SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
851            StoreVal = ConstantVector::get(&Elts[0], NumElts);
852          }
853        }
854        new StoreInst(StoreVal, EltPtr, MI);
855        continue;
856      }
857      // Otherwise, if we're storing a byte variable, use a memset call for
858      // this element.
859    }
860
861    // Cast the element pointer to BytePtrTy.
862    if (EltPtr->getType() != BytePtrTy)
863      EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI);
864
865    // Cast the other pointer (if we have one) to BytePtrTy.
866    if (OtherElt && OtherElt->getType() != BytePtrTy)
867      OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(),
868                                 MI);
869
870    unsigned EltSize = TD->getTypeAllocSize(EltTy);
871
872    // Finally, insert the meminst for this element.
873    if (isa<MemTransferInst>(MI)) {
874      Value *Ops[] = {
875        SROADest ? EltPtr : OtherElt,  // Dest ptr
876        SROADest ? OtherElt : EltPtr,  // Src ptr
877        ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
878        // Align
879        ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign)
880      };
881      CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
882    } else {
883      assert(isa<MemSetInst>(MI));
884      Value *Ops[] = {
885        EltPtr, MI->getOperand(2),  // Dest, Value,
886        ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
887        Zero  // Align
888      };
889      CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
890    }
891  }
892  MI->eraseFromParent();
893}
894
895/// RewriteStoreUserOfWholeAlloca - We found an store of an integer that
896/// overwrites the entire allocation.  Extract out the pieces of the stored
897/// integer and store them individually.
898void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI,
899                                         AllocationInst *AI,
900                                         SmallVector<AllocaInst*, 32> &NewElts){
901  // Extract each element out of the integer according to its structure offset
902  // and store the element value to the individual alloca.
903  Value *SrcVal = SI->getOperand(0);
904  const Type *AllocaEltTy = AI->getType()->getElementType();
905  uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
906
907  // If this isn't a store of an integer to the whole alloca, it may be a store
908  // to the first element.  Just ignore the store in this case and normal SROA
909  // will handle it.
910  if (!isa<IntegerType>(SrcVal->getType()) ||
911      TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits)
912    return;
913  // Handle tail padding by extending the operand
914  if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
915    SrcVal = new ZExtInst(SrcVal,
916                          IntegerType::get(SI->getContext(), AllocaSizeBits),
917                          "", SI);
918
919  DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI;
920
921  // There are two forms here: AI could be an array or struct.  Both cases
922  // have different ways to compute the element offset.
923  if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
924    const StructLayout *Layout = TD->getStructLayout(EltSTy);
925
926    for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
927      // Get the number of bits to shift SrcVal to get the value.
928      const Type *FieldTy = EltSTy->getElementType(i);
929      uint64_t Shift = Layout->getElementOffsetInBits(i);
930
931      if (TD->isBigEndian())
932        Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
933
934      Value *EltVal = SrcVal;
935      if (Shift) {
936        Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
937        EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
938                                            "sroa.store.elt", SI);
939      }
940
941      // Truncate down to an integer of the right size.
942      uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
943
944      // Ignore zero sized fields like {}, they obviously contain no data.
945      if (FieldSizeBits == 0) continue;
946
947      if (FieldSizeBits != AllocaSizeBits)
948        EltVal = new TruncInst(EltVal,
949                             IntegerType::get(SI->getContext(), FieldSizeBits),
950                              "", SI);
951      Value *DestField = NewElts[i];
952      if (EltVal->getType() == FieldTy) {
953        // Storing to an integer field of this size, just do it.
954      } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) {
955        // Bitcast to the right element type (for fp/vector values).
956        EltVal = new BitCastInst(EltVal, FieldTy, "", SI);
957      } else {
958        // Otherwise, bitcast the dest pointer (for aggregates).
959        DestField = new BitCastInst(DestField,
960                              PointerType::getUnqual(EltVal->getType()),
961                                    "", SI);
962      }
963      new StoreInst(EltVal, DestField, SI);
964    }
965
966  } else {
967    const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
968    const Type *ArrayEltTy = ATy->getElementType();
969    uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
970    uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
971
972    uint64_t Shift;
973
974    if (TD->isBigEndian())
975      Shift = AllocaSizeBits-ElementOffset;
976    else
977      Shift = 0;
978
979    for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
980      // Ignore zero sized fields like {}, they obviously contain no data.
981      if (ElementSizeBits == 0) continue;
982
983      Value *EltVal = SrcVal;
984      if (Shift) {
985        Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
986        EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
987                                            "sroa.store.elt", SI);
988      }
989
990      // Truncate down to an integer of the right size.
991      if (ElementSizeBits != AllocaSizeBits)
992        EltVal = new TruncInst(EltVal,
993                               IntegerType::get(SI->getContext(),
994                                                ElementSizeBits),"",SI);
995      Value *DestField = NewElts[i];
996      if (EltVal->getType() == ArrayEltTy) {
997        // Storing to an integer field of this size, just do it.
998      } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) {
999        // Bitcast to the right element type (for fp/vector values).
1000        EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI);
1001      } else {
1002        // Otherwise, bitcast the dest pointer (for aggregates).
1003        DestField = new BitCastInst(DestField,
1004                              PointerType::getUnqual(EltVal->getType()),
1005                                    "", SI);
1006      }
1007      new StoreInst(EltVal, DestField, SI);
1008
1009      if (TD->isBigEndian())
1010        Shift -= ElementOffset;
1011      else
1012        Shift += ElementOffset;
1013    }
1014  }
1015
1016  SI->eraseFromParent();
1017}
1018
1019/// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to
1020/// an integer.  Load the individual pieces to form the aggregate value.
1021void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
1022                                        SmallVector<AllocaInst*, 32> &NewElts) {
1023  // Extract each element out of the NewElts according to its structure offset
1024  // and form the result value.
1025  const Type *AllocaEltTy = AI->getType()->getElementType();
1026  uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
1027
1028  // If this isn't a load of the whole alloca to an integer, it may be a load
1029  // of the first element.  Just ignore the load in this case and normal SROA
1030  // will handle it.
1031  if (!isa<IntegerType>(LI->getType()) ||
1032      TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits)
1033    return;
1034
1035  DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI;
1036
1037  // There are two forms here: AI could be an array or struct.  Both cases
1038  // have different ways to compute the element offset.
1039  const StructLayout *Layout = 0;
1040  uint64_t ArrayEltBitOffset = 0;
1041  if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1042    Layout = TD->getStructLayout(EltSTy);
1043  } else {
1044    const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
1045    ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
1046  }
1047
1048  Value *ResultVal =
1049    Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
1050
1051  for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1052    // Load the value from the alloca.  If the NewElt is an aggregate, cast
1053    // the pointer to an integer of the same size before doing the load.
1054    Value *SrcField = NewElts[i];
1055    const Type *FieldTy =
1056      cast<PointerType>(SrcField->getType())->getElementType();
1057    uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1058
1059    // Ignore zero sized fields like {}, they obviously contain no data.
1060    if (FieldSizeBits == 0) continue;
1061
1062    const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
1063                                                     FieldSizeBits);
1064    if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() &&
1065        !isa<VectorType>(FieldTy))
1066      SrcField = new BitCastInst(SrcField,
1067                                 PointerType::getUnqual(FieldIntTy),
1068                                 "", LI);
1069    SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
1070
1071    // If SrcField is a fp or vector of the right size but that isn't an
1072    // integer type, bitcast to an integer so we can shift it.
1073    if (SrcField->getType() != FieldIntTy)
1074      SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
1075
1076    // Zero extend the field to be the same size as the final alloca so that
1077    // we can shift and insert it.
1078    if (SrcField->getType() != ResultVal->getType())
1079      SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
1080
1081    // Determine the number of bits to shift SrcField.
1082    uint64_t Shift;
1083    if (Layout) // Struct case.
1084      Shift = Layout->getElementOffsetInBits(i);
1085    else  // Array case.
1086      Shift = i*ArrayEltBitOffset;
1087
1088    if (TD->isBigEndian())
1089      Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
1090
1091    if (Shift) {
1092      Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
1093      SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
1094    }
1095
1096    ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
1097  }
1098
1099  // Handle tail padding by truncating the result
1100  if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
1101    ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
1102
1103  LI->replaceAllUsesWith(ResultVal);
1104  LI->eraseFromParent();
1105}
1106
1107
1108/// HasPadding - Return true if the specified type has any structure or
1109/// alignment padding, false otherwise.
1110static bool HasPadding(const Type *Ty, const TargetData &TD) {
1111  if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1112    const StructLayout *SL = TD.getStructLayout(STy);
1113    unsigned PrevFieldBitOffset = 0;
1114    for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1115      unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
1116
1117      // Padding in sub-elements?
1118      if (HasPadding(STy->getElementType(i), TD))
1119        return true;
1120
1121      // Check to see if there is any padding between this element and the
1122      // previous one.
1123      if (i) {
1124        unsigned PrevFieldEnd =
1125        PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
1126        if (PrevFieldEnd < FieldBitOffset)
1127          return true;
1128      }
1129
1130      PrevFieldBitOffset = FieldBitOffset;
1131    }
1132
1133    //  Check for tail padding.
1134    if (unsigned EltCount = STy->getNumElements()) {
1135      unsigned PrevFieldEnd = PrevFieldBitOffset +
1136                   TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
1137      if (PrevFieldEnd < SL->getSizeInBits())
1138        return true;
1139    }
1140
1141  } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1142    return HasPadding(ATy->getElementType(), TD);
1143  } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1144    return HasPadding(VTy->getElementType(), TD);
1145  }
1146  return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
1147}
1148
1149/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
1150/// an aggregate can be broken down into elements.  Return 0 if not, 3 if safe,
1151/// or 1 if safe after canonicalization has been performed.
1152///
1153int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) {
1154  // Loop over the use list of the alloca.  We can only transform it if all of
1155  // the users are safe to transform.
1156  AllocaInfo Info;
1157
1158  for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
1159       I != E; ++I) {
1160    isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info);
1161    if (Info.isUnsafe) {
1162      DOUT << "Cannot transform: " << *AI << "  due to user: " << **I;
1163      return 0;
1164    }
1165  }
1166
1167  // Okay, we know all the users are promotable.  If the aggregate is a memcpy
1168  // source and destination, we have to be careful.  In particular, the memcpy
1169  // could be moving around elements that live in structure padding of the LLVM
1170  // types, but may actually be used.  In these cases, we refuse to promote the
1171  // struct.
1172  if (Info.isMemCpySrc && Info.isMemCpyDst &&
1173      HasPadding(AI->getType()->getElementType(), *TD))
1174    return 0;
1175
1176  // If we require cleanup, return 1, otherwise return 3.
1177  return Info.needsCleanup ? 1 : 3;
1178}
1179
1180/// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP
1181/// is canonicalized here.
1182void SROA::CleanupGEP(GetElementPtrInst *GEPI) {
1183  gep_type_iterator I = gep_type_begin(GEPI);
1184  ++I;
1185
1186  const ArrayType *AT = dyn_cast<ArrayType>(*I);
1187  if (!AT)
1188    return;
1189
1190  uint64_t NumElements = AT->getNumElements();
1191
1192  if (isa<ConstantInt>(I.getOperand()))
1193    return;
1194
1195  if (NumElements == 1) {
1196    GEPI->setOperand(2,
1197                  Constant::getNullValue(Type::getInt32Ty(GEPI->getContext())));
1198    return;
1199  }
1200
1201  assert(NumElements == 2 && "Unhandled case!");
1202  // All users of the GEP must be loads.  At each use of the GEP, insert
1203  // two loads of the appropriate indexed GEP and select between them.
1204  Value *IsOne = new ICmpInst(GEPI, ICmpInst::ICMP_NE, I.getOperand(),
1205                              Constant::getNullValue(I.getOperand()->getType()),
1206                              "isone");
1207  // Insert the new GEP instructions, which are properly indexed.
1208  SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end());
1209  Indices[1] = Constant::getNullValue(Type::getInt32Ty(GEPI->getContext()));
1210  Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1211                                             Indices.begin(),
1212                                             Indices.end(),
1213                                             GEPI->getName()+".0", GEPI);
1214  Indices[1] = ConstantInt::get(Type::getInt32Ty(GEPI->getContext()), 1);
1215  Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1216                                            Indices.begin(),
1217                                            Indices.end(),
1218                                            GEPI->getName()+".1", GEPI);
1219  // Replace all loads of the variable index GEP with loads from both
1220  // indexes and a select.
1221  while (!GEPI->use_empty()) {
1222    LoadInst *LI = cast<LoadInst>(GEPI->use_back());
1223    Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
1224    Value *One  = new LoadInst(OneIdx , LI->getName()+".1", LI);
1225    Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI);
1226    LI->replaceAllUsesWith(R);
1227    LI->eraseFromParent();
1228  }
1229  GEPI->eraseFromParent();
1230}
1231
1232
1233/// CleanupAllocaUsers - If SROA reported that it can promote the specified
1234/// allocation, but only if cleaned up, perform the cleanups required.
1235void SROA::CleanupAllocaUsers(AllocationInst *AI) {
1236  // At this point, we know that the end result will be SROA'd and promoted, so
1237  // we can insert ugly code if required so long as sroa+mem2reg will clean it
1238  // up.
1239  for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1240       UI != E; ) {
1241    User *U = *UI++;
1242    if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U))
1243      CleanupGEP(GEPI);
1244    else {
1245      Instruction *I = cast<Instruction>(U);
1246      SmallVector<DbgInfoIntrinsic *, 2> DbgInUses;
1247      if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) {
1248        // Safe to remove debug info uses.
1249        while (!DbgInUses.empty()) {
1250          DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back();
1251          DI->eraseFromParent();
1252        }
1253        I->eraseFromParent();
1254      }
1255    }
1256  }
1257}
1258
1259/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at
1260/// the offset specified by Offset (which is specified in bytes).
1261///
1262/// There are two cases we handle here:
1263///   1) A union of vector types of the same size and potentially its elements.
1264///      Here we turn element accesses into insert/extract element operations.
1265///      This promotes a <4 x float> with a store of float to the third element
1266///      into a <4 x float> that uses insert element.
1267///   2) A fully general blob of memory, which we turn into some (potentially
1268///      large) integer type with extract and insert operations where the loads
1269///      and stores would mutate the memory.
1270static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy,
1271                        unsigned AllocaSize, const TargetData &TD,
1272                        LLVMContext &Context) {
1273  // If this could be contributing to a vector, analyze it.
1274  if (VecTy != Type::getVoidTy(Context)) { // either null or a vector type.
1275
1276    // If the In type is a vector that is the same size as the alloca, see if it
1277    // matches the existing VecTy.
1278    if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
1279      if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
1280        // If we're storing/loading a vector of the right size, allow it as a
1281        // vector.  If this the first vector we see, remember the type so that
1282        // we know the element size.
1283        if (VecTy == 0)
1284          VecTy = VInTy;
1285        return;
1286      }
1287    } else if (In == Type::getFloatTy(Context) ||
1288               In == Type::getDoubleTy(Context) ||
1289               (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 &&
1290                isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
1291      // If we're accessing something that could be an element of a vector, see
1292      // if the implied vector agrees with what we already have and if Offset is
1293      // compatible with it.
1294      unsigned EltSize = In->getPrimitiveSizeInBits()/8;
1295      if (Offset % EltSize == 0 &&
1296          AllocaSize % EltSize == 0 &&
1297          (VecTy == 0 ||
1298           cast<VectorType>(VecTy)->getElementType()
1299                 ->getPrimitiveSizeInBits()/8 == EltSize)) {
1300        if (VecTy == 0)
1301          VecTy = VectorType::get(In, AllocaSize/EltSize);
1302        return;
1303      }
1304    }
1305  }
1306
1307  // Otherwise, we have a case that we can't handle with an optimized vector
1308  // form.  We can still turn this into a large integer.
1309  VecTy = Type::getVoidTy(Context);
1310}
1311
1312/// CanConvertToScalar - V is a pointer.  If we can convert the pointee and all
1313/// its accesses to use a to single vector type, return true, and set VecTy to
1314/// the new type.  If we could convert the alloca into a single promotable
1315/// integer, return true but set VecTy to VoidTy.  Further, if the use is not a
1316/// completely trivial use that mem2reg could promote, set IsNotTrivial.  Offset
1317/// is the current offset from the base of the alloca being analyzed.
1318///
1319/// If we see at least one access to the value that is as a vector type, set the
1320/// SawVec flag.
1321///
1322bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
1323                              bool &SawVec, uint64_t Offset,
1324                              unsigned AllocaSize) {
1325  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1326    Instruction *User = cast<Instruction>(*UI);
1327
1328    if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1329      // Don't break volatile loads.
1330      if (LI->isVolatile())
1331        return false;
1332      MergeInType(LI->getType(), Offset, VecTy,
1333                  AllocaSize, *TD, V->getContext());
1334      SawVec |= isa<VectorType>(LI->getType());
1335      continue;
1336    }
1337
1338    if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1339      // Storing the pointer, not into the value?
1340      if (SI->getOperand(0) == V || SI->isVolatile()) return 0;
1341      MergeInType(SI->getOperand(0)->getType(), Offset,
1342                  VecTy, AllocaSize, *TD, V->getContext());
1343      SawVec |= isa<VectorType>(SI->getOperand(0)->getType());
1344      continue;
1345    }
1346
1347    if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
1348      if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset,
1349                              AllocaSize))
1350        return false;
1351      IsNotTrivial = true;
1352      continue;
1353    }
1354
1355    if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1356      // If this is a GEP with a variable indices, we can't handle it.
1357      if (!GEP->hasAllConstantIndices())
1358        return false;
1359
1360      // Compute the offset that this GEP adds to the pointer.
1361      SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1362      uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1363                                                &Indices[0], Indices.size());
1364      // See if all uses can be converted.
1365      if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset,
1366                              AllocaSize))
1367        return false;
1368      IsNotTrivial = true;
1369      continue;
1370    }
1371
1372    // If this is a constant sized memset of a constant value (e.g. 0) we can
1373    // handle it.
1374    if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1375      // Store of constant value and constant size.
1376      if (isa<ConstantInt>(MSI->getValue()) &&
1377          isa<ConstantInt>(MSI->getLength())) {
1378        IsNotTrivial = true;
1379        continue;
1380      }
1381    }
1382
1383    // If this is a memcpy or memmove into or out of the whole allocation, we
1384    // can handle it like a load or store of the scalar type.
1385    if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
1386      if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()))
1387        if (Len->getZExtValue() == AllocaSize && Offset == 0) {
1388          IsNotTrivial = true;
1389          continue;
1390        }
1391    }
1392
1393    // Ignore dbg intrinsic.
1394    if (isa<DbgInfoIntrinsic>(User))
1395      continue;
1396
1397    // Otherwise, we cannot handle this!
1398    return false;
1399  }
1400
1401  return true;
1402}
1403
1404
1405/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
1406/// directly.  This happens when we are converting an "integer union" to a
1407/// single integer scalar, or when we are converting a "vector union" to a
1408/// vector with insert/extractelement instructions.
1409///
1410/// Offset is an offset from the original alloca, in bits that need to be
1411/// shifted to the right.  By the end of this, there should be no uses of Ptr.
1412void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) {
1413  while (!Ptr->use_empty()) {
1414    Instruction *User = cast<Instruction>(Ptr->use_back());
1415
1416    if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
1417      ConvertUsesToScalar(CI, NewAI, Offset);
1418      CI->eraseFromParent();
1419      continue;
1420    }
1421
1422    if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1423      // Compute the offset that this GEP adds to the pointer.
1424      SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1425      uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1426                                                &Indices[0], Indices.size());
1427      ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
1428      GEP->eraseFromParent();
1429      continue;
1430    }
1431
1432    IRBuilder<> Builder(User->getParent(), User);
1433
1434    if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1435      // The load is a bit extract from NewAI shifted right by Offset bits.
1436      Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
1437      Value *NewLoadVal
1438        = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
1439      LI->replaceAllUsesWith(NewLoadVal);
1440      LI->eraseFromParent();
1441      continue;
1442    }
1443
1444    if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1445      assert(SI->getOperand(0) != Ptr && "Consistency error!");
1446      // FIXME: Remove once builder has Twine API.
1447      Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str());
1448      Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
1449                                             Builder);
1450      Builder.CreateStore(New, NewAI);
1451      SI->eraseFromParent();
1452      continue;
1453    }
1454
1455    // If this is a constant sized memset of a constant value (e.g. 0) we can
1456    // transform it into a store of the expanded constant value.
1457    if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1458      assert(MSI->getRawDest() == Ptr && "Consistency error!");
1459      unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
1460      if (NumBytes != 0) {
1461        unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
1462
1463        // Compute the value replicated the right number of times.
1464        APInt APVal(NumBytes*8, Val);
1465
1466        // Splat the value if non-zero.
1467        if (Val)
1468          for (unsigned i = 1; i != NumBytes; ++i)
1469            APVal |= APVal << 8;
1470
1471        // FIXME: Remove once builder has Twine API.
1472        Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str());
1473        Value *New = ConvertScalar_InsertValue(
1474                                    ConstantInt::get(User->getContext(), APVal),
1475                                               Old, Offset, Builder);
1476        Builder.CreateStore(New, NewAI);
1477      }
1478      MSI->eraseFromParent();
1479      continue;
1480    }
1481
1482    // If this is a memcpy or memmove into or out of the whole allocation, we
1483    // can handle it like a load or store of the scalar type.
1484    if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
1485      assert(Offset == 0 && "must be store to start of alloca");
1486
1487      // If the source and destination are both to the same alloca, then this is
1488      // a noop copy-to-self, just delete it.  Otherwise, emit a load and store
1489      // as appropriate.
1490      AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject());
1491
1492      if (MTI->getSource()->getUnderlyingObject() != OrigAI) {
1493        // Dest must be OrigAI, change this to be a load from the original
1494        // pointer (bitcasted), then a store to our new alloca.
1495        assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
1496        Value *SrcPtr = MTI->getSource();
1497        SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType());
1498
1499        LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
1500        SrcVal->setAlignment(MTI->getAlignment());
1501        Builder.CreateStore(SrcVal, NewAI);
1502      } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) {
1503        // Src must be OrigAI, change this to be a load from NewAI then a store
1504        // through the original dest pointer (bitcasted).
1505        assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
1506        LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
1507
1508        Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType());
1509        StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
1510        NewStore->setAlignment(MTI->getAlignment());
1511      } else {
1512        // Noop transfer. Src == Dst
1513      }
1514
1515
1516      MTI->eraseFromParent();
1517      continue;
1518    }
1519
1520    // If user is a dbg info intrinsic then it is safe to remove it.
1521    if (isa<DbgInfoIntrinsic>(User)) {
1522      User->eraseFromParent();
1523      continue;
1524    }
1525
1526    llvm_unreachable("Unsupported operation!");
1527  }
1528}
1529
1530/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
1531/// or vector value FromVal, extracting the bits from the offset specified by
1532/// Offset.  This returns the value, which is of type ToType.
1533///
1534/// This happens when we are converting an "integer union" to a single
1535/// integer scalar, or when we are converting a "vector union" to a vector with
1536/// insert/extractelement instructions.
1537///
1538/// Offset is an offset from the original alloca, in bits that need to be
1539/// shifted to the right.
1540Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
1541                                        uint64_t Offset, IRBuilder<> &Builder) {
1542  // If the load is of the whole new alloca, no conversion is needed.
1543  if (FromVal->getType() == ToType && Offset == 0)
1544    return FromVal;
1545
1546  // If the result alloca is a vector type, this is either an element
1547  // access or a bitcast to another vector type of the same size.
1548  if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
1549    if (isa<VectorType>(ToType))
1550      return Builder.CreateBitCast(FromVal, ToType, "tmp");
1551
1552    // Otherwise it must be an element access.
1553    unsigned Elt = 0;
1554    if (Offset) {
1555      unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
1556      Elt = Offset/EltSize;
1557      assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
1558    }
1559    // Return the element extracted out of it.
1560    Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get(
1561                    Type::getInt32Ty(FromVal->getContext()), Elt), "tmp");
1562    if (V->getType() != ToType)
1563      V = Builder.CreateBitCast(V, ToType, "tmp");
1564    return V;
1565  }
1566
1567  // If ToType is a first class aggregate, extract out each of the pieces and
1568  // use insertvalue's to form the FCA.
1569  if (const StructType *ST = dyn_cast<StructType>(ToType)) {
1570    const StructLayout &Layout = *TD->getStructLayout(ST);
1571    Value *Res = UndefValue::get(ST);
1572    for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1573      Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
1574                                        Offset+Layout.getElementOffsetInBits(i),
1575                                              Builder);
1576      Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1577    }
1578    return Res;
1579  }
1580
1581  if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
1582    uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType());
1583    Value *Res = UndefValue::get(AT);
1584    for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1585      Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
1586                                              Offset+i*EltSize, Builder);
1587      Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1588    }
1589    return Res;
1590  }
1591
1592  // Otherwise, this must be a union that was converted to an integer value.
1593  const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
1594
1595  // If this is a big-endian system and the load is narrower than the
1596  // full alloca type, we need to do a shift to get the right bits.
1597  int ShAmt = 0;
1598  if (TD->isBigEndian()) {
1599    // On big-endian machines, the lowest bit is stored at the bit offset
1600    // from the pointer given by getTypeStoreSizeInBits.  This matters for
1601    // integers with a bitwidth that is not a multiple of 8.
1602    ShAmt = TD->getTypeStoreSizeInBits(NTy) -
1603            TD->getTypeStoreSizeInBits(ToType) - Offset;
1604  } else {
1605    ShAmt = Offset;
1606  }
1607
1608  // Note: we support negative bitwidths (with shl) which are not defined.
1609  // We do this to support (f.e.) loads off the end of a structure where
1610  // only some bits are used.
1611  if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
1612    FromVal = Builder.CreateLShr(FromVal,
1613                                 ConstantInt::get(FromVal->getType(),
1614                                                           ShAmt), "tmp");
1615  else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
1616    FromVal = Builder.CreateShl(FromVal,
1617                                ConstantInt::get(FromVal->getType(),
1618                                                          -ShAmt), "tmp");
1619
1620  // Finally, unconditionally truncate the integer to the right width.
1621  unsigned LIBitWidth = TD->getTypeSizeInBits(ToType);
1622  if (LIBitWidth < NTy->getBitWidth())
1623    FromVal =
1624      Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
1625                                                    LIBitWidth), "tmp");
1626  else if (LIBitWidth > NTy->getBitWidth())
1627    FromVal =
1628       Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
1629                                                    LIBitWidth), "tmp");
1630
1631  // If the result is an integer, this is a trunc or bitcast.
1632  if (isa<IntegerType>(ToType)) {
1633    // Should be done.
1634  } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) {
1635    // Just do a bitcast, we know the sizes match up.
1636    FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
1637  } else {
1638    // Otherwise must be a pointer.
1639    FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
1640  }
1641  assert(FromVal->getType() == ToType && "Didn't convert right?");
1642  return FromVal;
1643}
1644
1645
1646/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
1647/// or vector value "Old" at the offset specified by Offset.
1648///
1649/// This happens when we are converting an "integer union" to a
1650/// single integer scalar, or when we are converting a "vector union" to a
1651/// vector with insert/extractelement instructions.
1652///
1653/// Offset is an offset from the original alloca, in bits that need to be
1654/// shifted to the right.
1655Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old,
1656                                       uint64_t Offset, IRBuilder<> &Builder) {
1657
1658  // Convert the stored type to the actual type, shift it left to insert
1659  // then 'or' into place.
1660  const Type *AllocaType = Old->getType();
1661  LLVMContext &Context = Old->getContext();
1662
1663  if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
1664    uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy);
1665    uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType());
1666
1667    // Changing the whole vector with memset or with an access of a different
1668    // vector type?
1669    if (ValSize == VecSize)
1670      return Builder.CreateBitCast(SV, AllocaType, "tmp");
1671
1672    uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
1673
1674    // Must be an element insertion.
1675    unsigned Elt = Offset/EltSize;
1676
1677    if (SV->getType() != VTy->getElementType())
1678      SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
1679
1680    SV = Builder.CreateInsertElement(Old, SV,
1681                     ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt),
1682                                     "tmp");
1683    return SV;
1684  }
1685
1686  // If SV is a first-class aggregate value, insert each value recursively.
1687  if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
1688    const StructLayout &Layout = *TD->getStructLayout(ST);
1689    for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1690      Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1691      Old = ConvertScalar_InsertValue(Elt, Old,
1692                                      Offset+Layout.getElementOffsetInBits(i),
1693                                      Builder);
1694    }
1695    return Old;
1696  }
1697
1698  if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
1699    uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType());
1700    for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1701      Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1702      Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
1703    }
1704    return Old;
1705  }
1706
1707  // If SV is a float, convert it to the appropriate integer type.
1708  // If it is a pointer, do the same.
1709  unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType());
1710  unsigned DestWidth = TD->getTypeSizeInBits(AllocaType);
1711  unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType());
1712  unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType);
1713  if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType()))
1714    SV = Builder.CreateBitCast(SV,
1715                            IntegerType::get(SV->getContext(),SrcWidth), "tmp");
1716  else if (isa<PointerType>(SV->getType()))
1717    SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(SV->getContext()), "tmp");
1718
1719  // Zero extend or truncate the value if needed.
1720  if (SV->getType() != AllocaType) {
1721    if (SV->getType()->getPrimitiveSizeInBits() <
1722             AllocaType->getPrimitiveSizeInBits())
1723      SV = Builder.CreateZExt(SV, AllocaType, "tmp");
1724    else {
1725      // Truncation may be needed if storing more than the alloca can hold
1726      // (undefined behavior).
1727      SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
1728      SrcWidth = DestWidth;
1729      SrcStoreWidth = DestStoreWidth;
1730    }
1731  }
1732
1733  // If this is a big-endian system and the store is narrower than the
1734  // full alloca type, we need to do a shift to get the right bits.
1735  int ShAmt = 0;
1736  if (TD->isBigEndian()) {
1737    // On big-endian machines, the lowest bit is stored at the bit offset
1738    // from the pointer given by getTypeStoreSizeInBits.  This matters for
1739    // integers with a bitwidth that is not a multiple of 8.
1740    ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1741  } else {
1742    ShAmt = Offset;
1743  }
1744
1745  // Note: we support negative bitwidths (with shr) which are not defined.
1746  // We do this to support (f.e.) stores off the end of a structure where
1747  // only some bits in the structure are set.
1748  APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1749  if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1750    SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(),
1751                           ShAmt), "tmp");
1752    Mask <<= ShAmt;
1753  } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1754    SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(),
1755                            -ShAmt), "tmp");
1756    Mask = Mask.lshr(-ShAmt);
1757  }
1758
1759  // Mask out the bits we are about to insert from the old value, and or
1760  // in the new bits.
1761  if (SrcWidth != DestWidth) {
1762    assert(DestWidth > SrcWidth);
1763    Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
1764    SV = Builder.CreateOr(Old, SV, "ins");
1765  }
1766  return SV;
1767}
1768
1769
1770
1771/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
1772/// some part of a constant global variable.  This intentionally only accepts
1773/// constant expressions because we don't can't rewrite arbitrary instructions.
1774static bool PointsToConstantGlobal(Value *V) {
1775  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1776    return GV->isConstant();
1777  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1778    if (CE->getOpcode() == Instruction::BitCast ||
1779        CE->getOpcode() == Instruction::GetElementPtr)
1780      return PointsToConstantGlobal(CE->getOperand(0));
1781  return false;
1782}
1783
1784/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
1785/// pointer to an alloca.  Ignore any reads of the pointer, return false if we
1786/// see any stores or other unknown uses.  If we see pointer arithmetic, keep
1787/// track of whether it moves the pointer (with isOffset) but otherwise traverse
1788/// the uses.  If we see a memcpy/memmove that targets an unoffseted pointer to
1789/// the alloca, and if the source pointer is a pointer to a constant  global, we
1790/// can optimize this.
1791static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy,
1792                                           bool isOffset) {
1793  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1794    if (LoadInst *LI = dyn_cast<LoadInst>(*UI))
1795      // Ignore non-volatile loads, they are always ok.
1796      if (!LI->isVolatile())
1797        continue;
1798
1799    if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) {
1800      // If uses of the bitcast are ok, we are ok.
1801      if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
1802        return false;
1803      continue;
1804    }
1805    if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
1806      // If the GEP has all zero indices, it doesn't offset the pointer.  If it
1807      // doesn't, it does.
1808      if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
1809                                         isOffset || !GEP->hasAllZeroIndices()))
1810        return false;
1811      continue;
1812    }
1813
1814    // If this is isn't our memcpy/memmove, reject it as something we can't
1815    // handle.
1816    if (!isa<MemTransferInst>(*UI))
1817      return false;
1818
1819    // If we already have seen a copy, reject the second one.
1820    if (TheCopy) return false;
1821
1822    // If the pointer has been offset from the start of the alloca, we can't
1823    // safely handle this.
1824    if (isOffset) return false;
1825
1826    // If the memintrinsic isn't using the alloca as the dest, reject it.
1827    if (UI.getOperandNo() != 1) return false;
1828
1829    MemIntrinsic *MI = cast<MemIntrinsic>(*UI);
1830
1831    // If the source of the memcpy/move is not a constant global, reject it.
1832    if (!PointsToConstantGlobal(MI->getOperand(2)))
1833      return false;
1834
1835    // Otherwise, the transform is safe.  Remember the copy instruction.
1836    TheCopy = MI;
1837  }
1838  return true;
1839}
1840
1841/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
1842/// modified by a copy from a constant global.  If we can prove this, we can
1843/// replace any uses of the alloca with uses of the global directly.
1844Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) {
1845  Instruction *TheCopy = 0;
1846  if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))
1847    return TheCopy;
1848  return 0;
1849}
1850