1 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===//
3 // The LLVM Compiler Infrastructure
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.
8 //===----------------------------------------------------------------------===//
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.
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.
20 //===----------------------------------------------------------------------===//
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/Support/GetElementPtrTypeIterator.h"
30 #include "llvm/Target/TargetData.h"
31 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/ADT/StringExtras.h"
38 Statistic<> NumReplaced("scalarrepl", "Number of allocas broken up");
39 Statistic<> NumPromoted("scalarrepl", "Number of allocas promoted");
41 struct SROA : public FunctionPass {
42 bool runOnFunction(Function &F);
44 bool performScalarRepl(Function &F);
45 bool performPromotion(Function &F);
47 // getAnalysisUsage - This pass does not require any passes, but we know it
48 // will not alter the CFG, so say so.
49 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
50 AU.addRequired<DominatorTree>();
51 AU.addRequired<DominanceFrontier>();
52 AU.addRequired<TargetData>();
57 int isSafeElementUse(Value *Ptr);
58 int isSafeUseOfAllocation(Instruction *User);
59 int isSafeAllocaToScalarRepl(AllocationInst *AI);
60 void CanonicalizeAllocaUsers(AllocationInst *AI);
61 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
64 RegisterOpt<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
67 // Public interface to the ScalarReplAggregates pass
68 FunctionPass *llvm::createScalarReplAggregatesPass() { return new SROA(); }
71 bool SROA::runOnFunction(Function &F) {
72 bool Changed = performPromotion(F);
74 bool LocalChange = performScalarRepl(F);
75 if (!LocalChange) break; // No need to repromote if no scalarrepl
77 LocalChange = performPromotion(F);
78 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
85 bool SROA::performPromotion(Function &F) {
86 std::vector<AllocaInst*> Allocas;
87 const TargetData &TD = getAnalysis<TargetData>();
88 DominatorTree &DT = getAnalysis<DominatorTree>();
89 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
91 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
98 // Find allocas that are safe to promote, by looking at all instructions in
100 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
101 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
102 if (isAllocaPromotable(AI, TD))
103 Allocas.push_back(AI);
105 if (Allocas.empty()) break;
107 PromoteMemToReg(Allocas, DT, DF, TD);
108 NumPromoted += Allocas.size();
116 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
117 // which runs on all of the malloc/alloca instructions in the function, removing
118 // them if they are only used by getelementptr instructions.
120 bool SROA::performScalarRepl(Function &F) {
121 std::vector<AllocationInst*> WorkList;
123 // Scan the entry basic block, adding any alloca's and mallocs to the worklist
124 BasicBlock &BB = F.getEntryBlock();
125 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
126 if (AllocationInst *A = dyn_cast<AllocationInst>(I))
127 WorkList.push_back(A);
129 // Process the worklist
130 bool Changed = false;
131 while (!WorkList.empty()) {
132 AllocationInst *AI = WorkList.back();
135 // We cannot transform the allocation instruction if it is an array
136 // allocation (allocations OF arrays are ok though), and an allocation of a
137 // scalar value cannot be decomposed at all.
139 if (AI->isArrayAllocation() ||
140 (!isa<StructType>(AI->getAllocatedType()) &&
141 !isa<ArrayType>(AI->getAllocatedType()))) continue;
143 // Check that all of the users of the allocation are capable of being
145 switch (isSafeAllocaToScalarRepl(AI)) {
146 default: assert(0 && "Unexpected value!");
147 case 0: // Not safe to scalar replace.
149 case 1: // Safe, but requires cleanup/canonicalizations first
150 CanonicalizeAllocaUsers(AI);
151 case 3: // Safe to scalar replace.
155 DEBUG(std::cerr << "Found inst to xform: " << *AI);
158 std::vector<AllocaInst*> ElementAllocas;
159 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
160 ElementAllocas.reserve(ST->getNumContainedTypes());
161 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
162 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
164 AI->getName() + "." + utostr(i), AI);
165 ElementAllocas.push_back(NA);
166 WorkList.push_back(NA); // Add to worklist for recursive processing
169 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
170 ElementAllocas.reserve(AT->getNumElements());
171 const Type *ElTy = AT->getElementType();
172 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
173 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
174 AI->getName() + "." + utostr(i), AI);
175 ElementAllocas.push_back(NA);
176 WorkList.push_back(NA); // Add to worklist for recursive processing
180 // Now that we have created the alloca instructions that we want to use,
181 // expand the getelementptr instructions to use them.
183 while (!AI->use_empty()) {
184 Instruction *User = cast<Instruction>(AI->use_back());
185 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
186 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
188 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getRawValue();
190 assert(Idx < ElementAllocas.size() && "Index out of range?");
191 AllocaInst *AllocaToUse = ElementAllocas[Idx];
194 if (GEPI->getNumOperands() == 3) {
195 // Do not insert a new getelementptr instruction with zero indices, only
196 // to have it optimized out later.
197 RepValue = AllocaToUse;
199 // We are indexing deeply into the structure, so we still need a
200 // getelement ptr instruction to finish the indexing. This may be
201 // expanded itself once the worklist is rerun.
203 std::string OldName = GEPI->getName(); // Steal the old name.
204 std::vector<Value*> NewArgs;
205 NewArgs.push_back(Constant::getNullValue(Type::IntTy));
206 NewArgs.insert(NewArgs.end(), GEPI->op_begin()+3, GEPI->op_end());
208 RepValue = new GetElementPtrInst(AllocaToUse, NewArgs, OldName, GEPI);
211 // Move all of the users over to the new GEP.
212 GEPI->replaceAllUsesWith(RepValue);
213 // Delete the old GEP
214 GEPI->eraseFromParent();
217 // Finally, delete the Alloca instruction
218 AI->getParent()->getInstList().erase(AI);
226 /// isSafeElementUse - Check to see if this use is an allowed use for a
227 /// getelementptr instruction of an array aggregate allocation.
229 int SROA::isSafeElementUse(Value *Ptr) {
230 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
232 Instruction *User = cast<Instruction>(*I);
233 switch (User->getOpcode()) {
234 case Instruction::Load: break;
235 case Instruction::Store:
236 // Store is ok if storing INTO the pointer, not storing the pointer
237 if (User->getOperand(0) == Ptr) return 0;
239 case Instruction::GetElementPtr: {
240 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
241 if (GEP->getNumOperands() > 1) {
242 if (!isa<Constant>(GEP->getOperand(1)) ||
243 !cast<Constant>(GEP->getOperand(1))->isNullValue())
244 return 0; // Using pointer arithmetic to navigate the array...
246 if (!isSafeElementUse(GEP)) return 0;
250 DEBUG(std::cerr << " Transformation preventing inst: " << *User);
254 return 3; // All users look ok :)
257 /// AllUsersAreLoads - Return true if all users of this value are loads.
258 static bool AllUsersAreLoads(Value *Ptr) {
259 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
261 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load)
266 /// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
267 /// aggregate allocation.
269 int SROA::isSafeUseOfAllocation(Instruction *User) {
270 if (!isa<GetElementPtrInst>(User)) return 0;
272 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
273 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI);
275 // The GEP is safe to transform if it is of the form GEP <ptr>, 0, <cst>
277 I.getOperand() != Constant::getNullValue(I.getOperand()->getType()))
281 if (I == E) return 0; // ran out of GEP indices??
283 // If this is a use of an array allocation, do a bit more checking for sanity.
284 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
285 uint64_t NumElements = AT->getNumElements();
287 if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) {
288 // Check to make sure that index falls within the array. If not,
289 // something funny is going on, so we won't do the optimization.
291 if (cast<ConstantInt>(GEPI->getOperand(2))->getRawValue() >= NumElements)
295 // If this is an array index and the index is not constant, we cannot
296 // promote... that is unless the array has exactly one or two elements in
297 // it, in which case we CAN promote it, but we have to canonicalize this
298 // out if this is the only problem.
299 if (NumElements == 1 || NumElements == 2)
300 return AllUsersAreLoads(GEPI) ? 1 : 0; // Canonicalization required!
305 // If there are any non-simple uses of this getelementptr, make sure to reject
307 return isSafeElementUse(GEPI);
310 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
311 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
312 /// or 1 if safe after canonicalization has been performed.
314 int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) {
315 // Loop over the use list of the alloca. We can only transform it if all of
316 // the users are safe to transform.
319 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
321 isSafe &= isSafeUseOfAllocation(cast<Instruction>(*I));
323 DEBUG(std::cerr << "Cannot transform: " << *AI << " due to user: "
328 // If we require cleanup, isSafe is now 1, otherwise it is 3.
332 /// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified
333 /// allocation, but only if cleaned up, perform the cleanups required.
334 void SROA::CanonicalizeAllocaUsers(AllocationInst *AI) {
335 // At this point, we know that the end result will be SROA'd and promoted, so
336 // we can insert ugly code if required so long as sroa+mem2reg will clean it
338 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
340 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(*UI++);
341 gep_type_iterator I = gep_type_begin(GEPI);
344 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
345 uint64_t NumElements = AT->getNumElements();
347 if (!isa<ConstantInt>(I.getOperand())) {
348 if (NumElements == 1) {
349 GEPI->setOperand(2, Constant::getNullValue(Type::IntTy));
351 assert(NumElements == 2 && "Unhandled case!");
352 // All users of the GEP must be loads. At each use of the GEP, insert
353 // two loads of the appropriate indexed GEP and select between them.
354 Value *IsOne = BinaryOperator::createSetNE(I.getOperand(),
355 Constant::getNullValue(I.getOperand()->getType()),
357 // Insert the new GEP instructions, which are properly indexed.
358 std::vector<Value*> Indices(GEPI->op_begin()+1, GEPI->op_end());
359 Indices[1] = Constant::getNullValue(Type::IntTy);
360 Value *ZeroIdx = new GetElementPtrInst(GEPI->getOperand(0), Indices,
361 GEPI->getName()+".0", GEPI);
362 Indices[1] = ConstantInt::get(Type::IntTy, 1);
363 Value *OneIdx = new GetElementPtrInst(GEPI->getOperand(0), Indices,
364 GEPI->getName()+".1", GEPI);
365 // Replace all loads of the variable index GEP with loads from both
366 // indexes and a select.
367 while (!GEPI->use_empty()) {
368 LoadInst *LI = cast<LoadInst>(GEPI->use_back());
369 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
370 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI);
371 Value *R = new SelectInst(IsOne, One, Zero, LI->getName(), LI);
372 LI->replaceAllUsesWith(R);
373 LI->eraseFromParent();
375 GEPI->eraseFromParent();