ScalarReplAggregates.cpp revision ae6a30509463a82182d090d2b42623e02c118596
1//===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file was developed by the LLVM research group and is distributed under
6// the University of Illinois Open Source 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#include "llvm/Transforms/Scalar.h"
23#include "llvm/Constants.h"
24#include "llvm/DerivedTypes.h"
25#include "llvm/Function.h"
26#include "llvm/Pass.h"
27#include "llvm/Instructions.h"
28#include "llvm/Analysis/Dominators.h"
29#include "llvm/Target/TargetData.h"
30#include "llvm/Transforms/Utils/PromoteMemToReg.h"
31#include "llvm/Support/Debug.h"
32#include "llvm/Support/GetElementPtrTypeIterator.h"
33#include "llvm/Support/MathExtras.h"
34#include "llvm/Support/Compiler.h"
35#include "llvm/ADT/Statistic.h"
36#include "llvm/ADT/StringExtras.h"
37using namespace llvm;
38
39namespace {
40  Statistic NumReplaced("scalarrepl", "Number of allocas broken up");
41  Statistic NumPromoted("scalarrepl", "Number of allocas promoted");
42  Statistic NumConverted("scalarrepl",
43                           "Number of aggregates converted to scalar");
44
45  struct VISIBILITY_HIDDEN SROA : public FunctionPass {
46    bool runOnFunction(Function &F);
47
48    bool performScalarRepl(Function &F);
49    bool performPromotion(Function &F);
50
51    // getAnalysisUsage - This pass does not require any passes, but we know it
52    // will not alter the CFG, so say so.
53    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
54      AU.addRequired<DominatorTree>();
55      AU.addRequired<DominanceFrontier>();
56      AU.addRequired<TargetData>();
57      AU.setPreservesCFG();
58    }
59
60  private:
61    int isSafeElementUse(Value *Ptr);
62    int isSafeUseOfAllocation(Instruction *User);
63    int isSafeAllocaToScalarRepl(AllocationInst *AI);
64    void CanonicalizeAllocaUsers(AllocationInst *AI);
65    AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
66
67    const Type *CanConvertToScalar(Value *V, bool &IsNotTrivial);
68    void ConvertToScalar(AllocationInst *AI, const Type *Ty);
69    void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset);
70  };
71
72  RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
73}
74
75// Public interface to the ScalarReplAggregates pass
76FunctionPass *llvm::createScalarReplAggregatesPass() { return new SROA(); }
77
78
79bool SROA::runOnFunction(Function &F) {
80  bool Changed = performPromotion(F);
81  while (1) {
82    bool LocalChange = performScalarRepl(F);
83    if (!LocalChange) break;   // No need to repromote if no scalarrepl
84    Changed = true;
85    LocalChange = performPromotion(F);
86    if (!LocalChange) break;   // No need to re-scalarrepl if no promotion
87  }
88
89  return Changed;
90}
91
92
93bool SROA::performPromotion(Function &F) {
94  std::vector<AllocaInst*> Allocas;
95  const TargetData &TD = getAnalysis<TargetData>();
96  DominatorTree     &DT = getAnalysis<DominatorTree>();
97  DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
98
99  BasicBlock &BB = F.getEntryBlock();  // Get the entry node for the function
100
101  bool Changed = false;
102
103  while (1) {
104    Allocas.clear();
105
106    // Find allocas that are safe to promote, by looking at all instructions in
107    // the entry node
108    for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
109      if (AllocaInst *AI = dyn_cast<AllocaInst>(I))       // Is it an alloca?
110        if (isAllocaPromotable(AI, TD))
111          Allocas.push_back(AI);
112
113    if (Allocas.empty()) break;
114
115    PromoteMemToReg(Allocas, DT, DF, TD);
116    NumPromoted += Allocas.size();
117    Changed = true;
118  }
119
120  return Changed;
121}
122
123// performScalarRepl - This algorithm is a simple worklist driven algorithm,
124// which runs on all of the malloc/alloca instructions in the function, removing
125// them if they are only used by getelementptr instructions.
126//
127bool SROA::performScalarRepl(Function &F) {
128  std::vector<AllocationInst*> WorkList;
129
130  // Scan the entry basic block, adding any alloca's and mallocs to the worklist
131  BasicBlock &BB = F.getEntryBlock();
132  for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
133    if (AllocationInst *A = dyn_cast<AllocationInst>(I))
134      WorkList.push_back(A);
135
136  // Process the worklist
137  bool Changed = false;
138  while (!WorkList.empty()) {
139    AllocationInst *AI = WorkList.back();
140    WorkList.pop_back();
141
142    // If we can turn this aggregate value (potentially with casts) into a
143    // simple scalar value that can be mem2reg'd into a register value.
144    bool IsNotTrivial = false;
145    if (const Type *ActualType = CanConvertToScalar(AI, IsNotTrivial))
146      if (IsNotTrivial && ActualType != Type::VoidTy) {
147        ConvertToScalar(AI, ActualType);
148        Changed = true;
149        continue;
150      }
151
152    // We cannot transform the allocation instruction if it is an array
153    // allocation (allocations OF arrays are ok though), and an allocation of a
154    // scalar value cannot be decomposed at all.
155    //
156    if (AI->isArrayAllocation() ||
157        (!isa<StructType>(AI->getAllocatedType()) &&
158         !isa<ArrayType>(AI->getAllocatedType()))) continue;
159
160    // Check that all of the users of the allocation are capable of being
161    // transformed.
162    switch (isSafeAllocaToScalarRepl(AI)) {
163    default: assert(0 && "Unexpected value!");
164    case 0:  // Not safe to scalar replace.
165      continue;
166    case 1:  // Safe, but requires cleanup/canonicalizations first
167      CanonicalizeAllocaUsers(AI);
168    case 3:  // Safe to scalar replace.
169      break;
170    }
171
172    DOUT << "Found inst to xform: " << *AI;
173    Changed = true;
174
175    std::vector<AllocaInst*> ElementAllocas;
176    if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
177      ElementAllocas.reserve(ST->getNumContainedTypes());
178      for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
179        AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
180                                        AI->getAlignment(),
181                                        AI->getName() + "." + utostr(i), AI);
182        ElementAllocas.push_back(NA);
183        WorkList.push_back(NA);  // Add to worklist for recursive processing
184      }
185    } else {
186      const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
187      ElementAllocas.reserve(AT->getNumElements());
188      const Type *ElTy = AT->getElementType();
189      for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
190        AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
191                                        AI->getName() + "." + utostr(i), AI);
192        ElementAllocas.push_back(NA);
193        WorkList.push_back(NA);  // Add to worklist for recursive processing
194      }
195    }
196
197    // Now that we have created the alloca instructions that we want to use,
198    // expand the getelementptr instructions to use them.
199    //
200    while (!AI->use_empty()) {
201      Instruction *User = cast<Instruction>(AI->use_back());
202      GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
203      // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
204      unsigned Idx =
205         (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
206
207      assert(Idx < ElementAllocas.size() && "Index out of range?");
208      AllocaInst *AllocaToUse = ElementAllocas[Idx];
209
210      Value *RepValue;
211      if (GEPI->getNumOperands() == 3) {
212        // Do not insert a new getelementptr instruction with zero indices, only
213        // to have it optimized out later.
214        RepValue = AllocaToUse;
215      } else {
216        // We are indexing deeply into the structure, so we still need a
217        // getelement ptr instruction to finish the indexing.  This may be
218        // expanded itself once the worklist is rerun.
219        //
220        std::string OldName = GEPI->getName();  // Steal the old name.
221        std::vector<Value*> NewArgs;
222        NewArgs.push_back(Constant::getNullValue(Type::IntTy));
223        NewArgs.insert(NewArgs.end(), GEPI->op_begin()+3, GEPI->op_end());
224        GEPI->setName("");
225        RepValue = new GetElementPtrInst(AllocaToUse, NewArgs, OldName, GEPI);
226      }
227
228      // Move all of the users over to the new GEP.
229      GEPI->replaceAllUsesWith(RepValue);
230      // Delete the old GEP
231      GEPI->eraseFromParent();
232    }
233
234    // Finally, delete the Alloca instruction
235    AI->getParent()->getInstList().erase(AI);
236    NumReplaced++;
237  }
238
239  return Changed;
240}
241
242
243/// isSafeElementUse - Check to see if this use is an allowed use for a
244/// getelementptr instruction of an array aggregate allocation.
245///
246int SROA::isSafeElementUse(Value *Ptr) {
247  for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
248       I != E; ++I) {
249    Instruction *User = cast<Instruction>(*I);
250    switch (User->getOpcode()) {
251    case Instruction::Load:  break;
252    case Instruction::Store:
253      // Store is ok if storing INTO the pointer, not storing the pointer
254      if (User->getOperand(0) == Ptr) return 0;
255      break;
256    case Instruction::GetElementPtr: {
257      GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
258      if (GEP->getNumOperands() > 1) {
259        if (!isa<Constant>(GEP->getOperand(1)) ||
260            !cast<Constant>(GEP->getOperand(1))->isNullValue())
261          return 0;  // Using pointer arithmetic to navigate the array...
262      }
263      if (!isSafeElementUse(GEP)) return 0;
264      break;
265    }
266    default:
267      DOUT << "  Transformation preventing inst: " << *User;
268      return 0;
269    }
270  }
271  return 3;  // All users look ok :)
272}
273
274/// AllUsersAreLoads - Return true if all users of this value are loads.
275static bool AllUsersAreLoads(Value *Ptr) {
276  for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
277       I != E; ++I)
278    if (cast<Instruction>(*I)->getOpcode() != Instruction::Load)
279      return false;
280  return true;
281}
282
283/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
284/// aggregate allocation.
285///
286int SROA::isSafeUseOfAllocation(Instruction *User) {
287  if (!isa<GetElementPtrInst>(User)) return 0;
288
289  GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
290  gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI);
291
292  // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>".
293  if (I == E ||
294      I.getOperand() != Constant::getNullValue(I.getOperand()->getType()))
295    return 0;
296
297  ++I;
298  if (I == E) return 0;  // ran out of GEP indices??
299
300  // If this is a use of an array allocation, do a bit more checking for sanity.
301  if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
302    uint64_t NumElements = AT->getNumElements();
303
304    if (isa<ConstantInt>(I.getOperand())) {
305      // Check to make sure that index falls within the array.  If not,
306      // something funny is going on, so we won't do the optimization.
307      //
308      if (cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue() >= NumElements)
309        return 0;
310
311      // We cannot scalar repl this level of the array unless any array
312      // sub-indices are in-range constants.  In particular, consider:
313      // A[0][i].  We cannot know that the user isn't doing invalid things like
314      // allowing i to index an out-of-range subscript that accesses A[1].
315      //
316      // Scalar replacing *just* the outer index of the array is probably not
317      // going to be a win anyway, so just give up.
318      for (++I; I != E && (isa<ArrayType>(*I) || isa<PackedType>(*I)); ++I) {
319        uint64_t NumElements;
320        if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*I))
321          NumElements = SubArrayTy->getNumElements();
322        else
323          NumElements = cast<PackedType>(*I)->getNumElements();
324
325        if (!isa<ConstantInt>(I.getOperand())) return 0;
326        if (cast<ConstantInt>(I.getOperand())->getZExtValue() >= NumElements)
327          return 0;
328      }
329
330    } else {
331      // If this is an array index and the index is not constant, we cannot
332      // promote... that is unless the array has exactly one or two elements in
333      // it, in which case we CAN promote it, but we have to canonicalize this
334      // out if this is the only problem.
335      if ((NumElements == 1 || NumElements == 2) &&
336          AllUsersAreLoads(GEPI))
337        return 1;  // Canonicalization required!
338      return 0;
339    }
340  }
341
342  // If there are any non-simple uses of this getelementptr, make sure to reject
343  // them.
344  return isSafeElementUse(GEPI);
345}
346
347/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
348/// an aggregate can be broken down into elements.  Return 0 if not, 3 if safe,
349/// or 1 if safe after canonicalization has been performed.
350///
351int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) {
352  // Loop over the use list of the alloca.  We can only transform it if all of
353  // the users are safe to transform.
354  //
355  int isSafe = 3;
356  for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
357       I != E; ++I) {
358    isSafe &= isSafeUseOfAllocation(cast<Instruction>(*I));
359    if (isSafe == 0) {
360      DOUT << "Cannot transform: " << *AI << "  due to user: " << **I;
361      return 0;
362    }
363  }
364  // If we require cleanup, isSafe is now 1, otherwise it is 3.
365  return isSafe;
366}
367
368/// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified
369/// allocation, but only if cleaned up, perform the cleanups required.
370void SROA::CanonicalizeAllocaUsers(AllocationInst *AI) {
371  // At this point, we know that the end result will be SROA'd and promoted, so
372  // we can insert ugly code if required so long as sroa+mem2reg will clean it
373  // up.
374  for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
375       UI != E; ) {
376    GetElementPtrInst *GEPI = cast<GetElementPtrInst>(*UI++);
377    gep_type_iterator I = gep_type_begin(GEPI);
378    ++I;
379
380    if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
381      uint64_t NumElements = AT->getNumElements();
382
383      if (!isa<ConstantInt>(I.getOperand())) {
384        if (NumElements == 1) {
385          GEPI->setOperand(2, Constant::getNullValue(Type::IntTy));
386        } else {
387          assert(NumElements == 2 && "Unhandled case!");
388          // All users of the GEP must be loads.  At each use of the GEP, insert
389          // two loads of the appropriate indexed GEP and select between them.
390          Value *IsOne = BinaryOperator::createSetNE(I.getOperand(),
391                              Constant::getNullValue(I.getOperand()->getType()),
392                                                     "isone", GEPI);
393          // Insert the new GEP instructions, which are properly indexed.
394          std::vector<Value*> Indices(GEPI->op_begin()+1, GEPI->op_end());
395          Indices[1] = Constant::getNullValue(Type::IntTy);
396          Value *ZeroIdx = new GetElementPtrInst(GEPI->getOperand(0), Indices,
397                                                 GEPI->getName()+".0", GEPI);
398          Indices[1] = ConstantInt::get(Type::IntTy, 1);
399          Value *OneIdx = new GetElementPtrInst(GEPI->getOperand(0), Indices,
400                                                GEPI->getName()+".1", GEPI);
401          // Replace all loads of the variable index GEP with loads from both
402          // indexes and a select.
403          while (!GEPI->use_empty()) {
404            LoadInst *LI = cast<LoadInst>(GEPI->use_back());
405            Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
406            Value *One  = new LoadInst(OneIdx , LI->getName()+".1", LI);
407            Value *R = new SelectInst(IsOne, One, Zero, LI->getName(), LI);
408            LI->replaceAllUsesWith(R);
409            LI->eraseFromParent();
410          }
411          GEPI->eraseFromParent();
412        }
413      }
414    }
415  }
416}
417
418/// MergeInType - Add the 'In' type to the accumulated type so far.  If the
419/// types are incompatible, return true, otherwise update Accum and return
420/// false.
421///
422/// There are three cases we handle here:
423///   1) An effectively integer union, where the pieces are stored into as
424///      smaller integers (common with byte swap and other idioms).
425///   2) A union of a vector and its elements.  Here we turn element accesses
426///      into insert/extract element operations.
427///   3) A union of scalar types, such as int/float or int/pointer.  Here we
428///      merge together into integers, allowing the xform to work with #1 as
429///      well.
430static bool MergeInType(const Type *In, const Type *&Accum,
431                        const TargetData &TD) {
432  // If this is our first type, just use it.
433  const PackedType *PTy;
434  if (Accum == Type::VoidTy || In == Accum) {
435    Accum = In;
436  } else if (In->isIntegral() && Accum->isIntegral()) {   // integer union.
437    // Otherwise pick whichever type is larger.
438    if (In->getTypeID() > Accum->getTypeID())
439      Accum = In;
440  } else if (isa<PointerType>(In) && isa<PointerType>(Accum)) {
441    // Pointer unions just stay as one of the pointers.
442  } else if (isa<PackedType>(In) || isa<PackedType>(Accum)) {
443    if ((PTy = dyn_cast<PackedType>(Accum)) &&
444        PTy->getElementType() == In) {
445      // Accum is a vector, and we are accessing an element: ok.
446    } else if ((PTy = dyn_cast<PackedType>(In)) &&
447               PTy->getElementType() == Accum) {
448      // In is a vector, and accum is an element: ok, remember In.
449      Accum = In;
450    } else {
451      // FIXME: Handle packed->packed.
452      return true;
453    }
454  } else {
455    // Pointer/FP/Integer unions merge together as integers.
456    switch (Accum->getTypeID()) {
457    case Type::PointerTyID: Accum = TD.getIntPtrType(); break;
458    case Type::FloatTyID:   Accum = Type::UIntTy; break;
459    case Type::DoubleTyID:  Accum = Type::ULongTy; break;
460    default:
461      assert(Accum->isIntegral() && "Unknown FP type!");
462      break;
463    }
464
465    switch (In->getTypeID()) {
466    case Type::PointerTyID: In = TD.getIntPtrType(); break;
467    case Type::FloatTyID:   In = Type::UIntTy; break;
468    case Type::DoubleTyID:  In = Type::ULongTy; break;
469    default:
470      assert(In->isIntegral() && "Unknown FP type!");
471      break;
472    }
473    return MergeInType(In, Accum, TD);
474  }
475  return false;
476}
477
478/// getUIntAtLeastAsBitAs - Return an unsigned integer type that is at least
479/// as big as the specified type.  If there is no suitable type, this returns
480/// null.
481const Type *getUIntAtLeastAsBitAs(unsigned NumBits) {
482  if (NumBits > 64) return 0;
483  if (NumBits > 32) return Type::ULongTy;
484  if (NumBits > 16) return Type::UIntTy;
485  if (NumBits > 8) return Type::UShortTy;
486  return Type::UByteTy;
487}
488
489/// CanConvertToScalar - V is a pointer.  If we can convert the pointee to a
490/// single scalar integer type, return that type.  Further, if the use is not
491/// a completely trivial use that mem2reg could promote, set IsNotTrivial.  If
492/// there are no uses of this pointer, return Type::VoidTy to differentiate from
493/// failure.
494///
495const Type *SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial) {
496  const Type *UsedType = Type::VoidTy; // No uses, no forced type.
497  const TargetData &TD = getAnalysis<TargetData>();
498  const PointerType *PTy = cast<PointerType>(V->getType());
499
500  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
501    Instruction *User = cast<Instruction>(*UI);
502
503    if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
504      if (MergeInType(LI->getType(), UsedType, TD))
505        return 0;
506
507    } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
508      // Storing the pointer, not the into the value?
509      if (SI->getOperand(0) == V) return 0;
510
511      // NOTE: We could handle storing of FP imms into integers here!
512
513      if (MergeInType(SI->getOperand(0)->getType(), UsedType, TD))
514        return 0;
515    } else if (CastInst *CI = dyn_cast<CastInst>(User)) {
516      if (!isa<PointerType>(CI->getType())) return 0;
517      IsNotTrivial = true;
518      const Type *SubTy = CanConvertToScalar(CI, IsNotTrivial);
519      if (!SubTy || MergeInType(SubTy, UsedType, TD)) return 0;
520    } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
521      // Check to see if this is stepping over an element: GEP Ptr, int C
522      if (GEP->getNumOperands() == 2 && isa<ConstantInt>(GEP->getOperand(1))) {
523        unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue();
524        unsigned ElSize = TD.getTypeSize(PTy->getElementType());
525        unsigned BitOffset = Idx*ElSize*8;
526        if (BitOffset > 64 || !isPowerOf2_32(ElSize)) return 0;
527
528        IsNotTrivial = true;
529        const Type *SubElt = CanConvertToScalar(GEP, IsNotTrivial);
530        if (SubElt == 0) return 0;
531        if (SubElt != Type::VoidTy && SubElt->isInteger()) {
532          const Type *NewTy =
533            getUIntAtLeastAsBitAs(TD.getTypeSize(SubElt)*8+BitOffset);
534          if (NewTy == 0 || MergeInType(NewTy, UsedType, TD)) return 0;
535          continue;
536        }
537      } else if (GEP->getNumOperands() == 3 &&
538                 isa<ConstantInt>(GEP->getOperand(1)) &&
539                 isa<ConstantInt>(GEP->getOperand(2)) &&
540                 cast<Constant>(GEP->getOperand(1))->isNullValue()) {
541        // We are stepping into an element, e.g. a structure or an array:
542        // GEP Ptr, int 0, uint C
543        const Type *AggTy = PTy->getElementType();
544        unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
545
546        if (const ArrayType *ATy = dyn_cast<ArrayType>(AggTy)) {
547          if (Idx >= ATy->getNumElements()) return 0;  // Out of range.
548        } else if (const PackedType *PackedTy = dyn_cast<PackedType>(AggTy)) {
549          // Getting an element of the packed vector.
550          if (Idx >= PackedTy->getNumElements()) return 0;  // Out of range.
551
552          // Merge in the packed type.
553          if (MergeInType(PackedTy, UsedType, TD)) return 0;
554
555          const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial);
556          if (SubTy == 0) return 0;
557
558          if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, TD))
559            return 0;
560
561          // We'll need to change this to an insert/extract element operation.
562          IsNotTrivial = true;
563          continue;    // Everything looks ok
564
565        } else if (isa<StructType>(AggTy)) {
566          // Structs are always ok.
567        } else {
568          return 0;
569        }
570        const Type *NTy = getUIntAtLeastAsBitAs(TD.getTypeSize(AggTy)*8);
571        if (NTy == 0 || MergeInType(NTy, UsedType, TD)) return 0;
572        const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial);
573        if (SubTy == 0) return 0;
574        if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, TD))
575          return 0;
576        continue;    // Everything looks ok
577      }
578      return 0;
579    } else {
580      // Cannot handle this!
581      return 0;
582    }
583  }
584
585  return UsedType;
586}
587
588/// ConvertToScalar - The specified alloca passes the CanConvertToScalar
589/// predicate and is non-trivial.  Convert it to something that can be trivially
590/// promoted into a register by mem2reg.
591void SROA::ConvertToScalar(AllocationInst *AI, const Type *ActualTy) {
592  DOUT << "CONVERT TO SCALAR: " << *AI << "  TYPE = "
593       << *ActualTy << "\n";
594  ++NumConverted;
595
596  BasicBlock *EntryBlock = AI->getParent();
597  assert(EntryBlock == &EntryBlock->getParent()->front() &&
598         "Not in the entry block!");
599  EntryBlock->getInstList().remove(AI);  // Take the alloca out of the program.
600
601  if (ActualTy->isInteger())
602    ActualTy = ActualTy->getUnsignedVersion();
603
604  // Create and insert the alloca.
605  AllocaInst *NewAI = new AllocaInst(ActualTy, 0, AI->getName(),
606                                     EntryBlock->begin());
607  ConvertUsesToScalar(AI, NewAI, 0);
608  delete AI;
609}
610
611
612/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
613/// directly.  This happens when we are converting an "integer union" to a
614/// single integer scalar, or when we are converting a "vector union" to a
615/// vector with insert/extractelement instructions.
616///
617/// Offset is an offset from the original alloca, in bits that need to be
618/// shifted to the right.  By the end of this, there should be no uses of Ptr.
619void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset) {
620  bool isVectorInsert = isa<PackedType>(NewAI->getType()->getElementType());
621  const TargetData &TD = getAnalysis<TargetData>();
622  while (!Ptr->use_empty()) {
623    Instruction *User = cast<Instruction>(Ptr->use_back());
624
625    if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
626      // The load is a bit extract from NewAI shifted right by Offset bits.
627      Value *NV = new LoadInst(NewAI, LI->getName(), LI);
628      if (NV->getType() != LI->getType()) {
629        if (const PackedType *PTy = dyn_cast<PackedType>(NV->getType())) {
630          // Must be an element access.
631          unsigned Elt = Offset/(TD.getTypeSize(PTy->getElementType())*8);
632          NV = new ExtractElementInst(NV, ConstantInt::get(Type::UIntTy, Elt),
633                                      "tmp", LI);
634        } else if (isa<PointerType>(NV->getType())) {
635          assert(isa<PointerType>(LI->getType()));
636          // Must be ptr->ptr cast.  Anything else would result in NV being
637          // an integer.
638          NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI);
639        } else {
640          assert(NV->getType()->isInteger() && "Unknown promotion!");
641          if (Offset && Offset < TD.getTypeSize(NV->getType())*8) {
642            NV = new ShiftInst(Instruction::LShr, NV,
643                               ConstantInt::get(Type::UByteTy, Offset),
644                               LI->getName(), LI);
645          }
646
647          // If the result is an integer, this is a trunc or bitcast.
648          if (LI->getType()->isIntegral()) {
649            NV = CastInst::createTruncOrBitCast(NV, LI->getType(),
650                                                LI->getName(), LI);
651          } else if (LI->getType()->isFloatingPoint()) {
652            // If needed, truncate the integer to the appropriate size.
653            if (NV->getType()->getPrimitiveSize() >
654                  LI->getType()->getPrimitiveSize())
655              NV = new TruncInst(NV, LI->getType(), LI->getName(), LI);
656
657            // Then do a bitcast.
658            NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI);
659          } else {
660            // Otherwise must be a pointer.
661            NV = new IntToPtrInst(NV, LI->getType(), LI->getName(), LI);
662          }
663        }
664      }
665      LI->replaceAllUsesWith(NV);
666      LI->eraseFromParent();
667    } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
668      assert(SI->getOperand(0) != Ptr && "Consistency error!");
669
670      // Convert the stored type to the actual type, shift it left to insert
671      // then 'or' into place.
672      Value *SV = SI->getOperand(0);
673      const Type *AllocaType = NewAI->getType()->getElementType();
674      if (SV->getType() != AllocaType) {
675        Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI);
676
677        if (const PackedType *PTy = dyn_cast<PackedType>(AllocaType)) {
678          // Must be an element insertion.
679          unsigned Elt = Offset/(TD.getTypeSize(PTy->getElementType())*8);
680          SV = new InsertElementInst(Old, SV,
681                                     ConstantInt::get(Type::UIntTy, Elt),
682                                     "tmp", SI);
683        } else {
684          // If SV is a float, convert it to the appropriate integer type.
685          // If it is a pointer, do the same, and also handle ptr->ptr casts
686          // here.
687          switch (SV->getType()->getTypeID()) {
688          default:
689            assert(!SV->getType()->isFloatingPoint() && "Unknown FP type!");
690            break;
691          case Type::FloatTyID:
692            SV = new BitCastInst(SV, Type::UIntTy, SV->getName(), SI);
693            break;
694          case Type::DoubleTyID:
695            SV = new BitCastInst(SV, Type::ULongTy, SV->getName(), SI);
696            break;
697          case Type::PointerTyID:
698            if (isa<PointerType>(AllocaType))
699              SV = new BitCastInst(SV, AllocaType, SV->getName(), SI);
700            else
701              SV = new PtrToIntInst(SV, TD.getIntPtrType(), SV->getName(), SI);
702            break;
703          }
704
705          unsigned SrcSize = TD.getTypeSize(SV->getType())*8;
706
707          // Always zero extend the value if needed.
708          if (SV->getType() != AllocaType)
709            SV = CastInst::createZExtOrBitCast(SV, AllocaType,
710                                               SV->getName(), SI);
711          if (Offset && Offset < AllocaType->getPrimitiveSizeInBits())
712            SV = new ShiftInst(Instruction::Shl, SV,
713                               ConstantInt::get(Type::UByteTy, Offset),
714                               SV->getName()+".adj", SI);
715          // Mask out the bits we are about to insert from the old value.
716          unsigned TotalBits = TD.getTypeSize(SV->getType())*8;
717          if (TotalBits != SrcSize) {
718            assert(TotalBits > SrcSize);
719            uint64_t Mask = ~(((1ULL << SrcSize)-1) << Offset);
720            Mask = Mask & SV->getType()->getIntegralTypeMask();
721            Old = BinaryOperator::createAnd(Old,
722                                        ConstantInt::get(Old->getType(), Mask),
723                                            Old->getName()+".mask", SI);
724            SV = BinaryOperator::createOr(Old, SV, SV->getName()+".ins", SI);
725          }
726        }
727      }
728      new StoreInst(SV, NewAI, SI);
729      SI->eraseFromParent();
730
731    } else if (CastInst *CI = dyn_cast<CastInst>(User)) {
732      unsigned NewOff = Offset;
733      const TargetData &TD = getAnalysis<TargetData>();
734      if (TD.isBigEndian() && !isVectorInsert) {
735        // Adjust the pointer.  For example, storing 16-bits into a 32-bit
736        // alloca with just a cast makes it modify the top 16-bits.
737        const Type *SrcTy = cast<PointerType>(Ptr->getType())->getElementType();
738        const Type *DstTy = cast<PointerType>(CI->getType())->getElementType();
739        int PtrDiffBits = TD.getTypeSize(SrcTy)*8-TD.getTypeSize(DstTy)*8;
740        NewOff += PtrDiffBits;
741      }
742      ConvertUsesToScalar(CI, NewAI, NewOff);
743      CI->eraseFromParent();
744    } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
745      const PointerType *AggPtrTy =
746        cast<PointerType>(GEP->getOperand(0)->getType());
747      const TargetData &TD = getAnalysis<TargetData>();
748      unsigned AggSizeInBits = TD.getTypeSize(AggPtrTy->getElementType())*8;
749
750      // Check to see if this is stepping over an element: GEP Ptr, int C
751      unsigned NewOffset = Offset;
752      if (GEP->getNumOperands() == 2) {
753        unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue();
754        unsigned BitOffset = Idx*AggSizeInBits;
755
756        if (TD.isLittleEndian() || isVectorInsert)
757          NewOffset += BitOffset;
758        else
759          NewOffset -= BitOffset;
760
761      } else if (GEP->getNumOperands() == 3) {
762        // We know that operand #2 is zero.
763        unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
764        const Type *AggTy = AggPtrTy->getElementType();
765        if (const SequentialType *SeqTy = dyn_cast<SequentialType>(AggTy)) {
766          unsigned ElSizeBits = TD.getTypeSize(SeqTy->getElementType())*8;
767
768          if (TD.isLittleEndian() || isVectorInsert)
769            NewOffset += ElSizeBits*Idx;
770          else
771            NewOffset += AggSizeInBits-ElSizeBits*(Idx+1);
772        } else if (const StructType *STy = dyn_cast<StructType>(AggTy)) {
773          unsigned EltBitOffset = TD.getStructLayout(STy)->MemberOffsets[Idx]*8;
774
775          if (TD.isLittleEndian() || isVectorInsert)
776            NewOffset += EltBitOffset;
777          else {
778            const PointerType *ElPtrTy = cast<PointerType>(GEP->getType());
779            unsigned ElSizeBits = TD.getTypeSize(ElPtrTy->getElementType())*8;
780            NewOffset += AggSizeInBits-(EltBitOffset+ElSizeBits);
781          }
782
783        } else {
784          assert(0 && "Unsupported operation!");
785          abort();
786        }
787      } else {
788        assert(0 && "Unsupported operation!");
789        abort();
790      }
791      ConvertUsesToScalar(GEP, NewAI, NewOffset);
792      GEP->eraseFromParent();
793    } else {
794      assert(0 && "Unsupported operation!");
795      abort();
796    }
797  }
798}
799