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/iMemory.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 "Support/Debug.h"
33 #include "Support/Statistic.h"
34 #include "Support/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 bool isSafeElementUse(Value *Ptr);
58 bool isSafeUseOfAllocation(Instruction *User);
59 bool isSafeAllocaToPromote(AllocationInst *AI);
60 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
63 RegisterOpt<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
66 // Public interface to the ScalarReplAggregates pass
67 Pass *llvm::createScalarReplAggregatesPass() { return new SROA(); }
70 bool SROA::runOnFunction(Function &F) {
71 bool Changed = performPromotion(F);
73 bool LocalChange = performScalarRepl(F);
74 if (!LocalChange) break; // No need to repromote if no scalarrepl
76 LocalChange = performPromotion(F);
77 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
84 bool SROA::performPromotion(Function &F) {
85 std::vector<AllocaInst*> Allocas;
86 const TargetData &TD = getAnalysis<TargetData>();
87 DominatorTree &DT = getAnalysis<DominatorTree>();
88 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
90 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
97 // Find allocas that are safe to promote, by looking at all instructions in
99 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
100 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
101 if (isAllocaPromotable(AI, TD))
102 Allocas.push_back(AI);
104 if (Allocas.empty()) break;
106 PromoteMemToReg(Allocas, DT, DF, TD);
107 NumPromoted += Allocas.size();
115 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
116 // which runs on all of the malloc/alloca instructions in the function, removing
117 // them if they are only used by getelementptr instructions.
119 bool SROA::performScalarRepl(Function &F) {
120 std::vector<AllocationInst*> WorkList;
122 // Scan the entry basic block, adding any alloca's and mallocs to the worklist
123 BasicBlock &BB = F.getEntryBlock();
124 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
125 if (AllocationInst *A = dyn_cast<AllocationInst>(I))
126 WorkList.push_back(A);
128 // Process the worklist
129 bool Changed = false;
130 while (!WorkList.empty()) {
131 AllocationInst *AI = WorkList.back();
134 // We cannot transform the allocation instruction if it is an array
135 // allocation (allocations OF arrays are ok though), and an allocation of a
136 // scalar value cannot be decomposed at all.
138 if (AI->isArrayAllocation() ||
139 (!isa<StructType>(AI->getAllocatedType()) &&
140 !isa<ArrayType>(AI->getAllocatedType()))) continue;
142 // Check that all of the users of the allocation are capable of being
144 if (!isSafeAllocaToPromote(AI))
147 DEBUG(std::cerr << "Found inst to xform: " << *AI);
150 std::vector<AllocaInst*> ElementAllocas;
151 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
152 ElementAllocas.reserve(ST->getNumContainedTypes());
153 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
154 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
155 AI->getName() + "." + utostr(i), AI);
156 ElementAllocas.push_back(NA);
157 WorkList.push_back(NA); // Add to worklist for recursive processing
160 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
161 ElementAllocas.reserve(AT->getNumElements());
162 const Type *ElTy = AT->getElementType();
163 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
164 AllocaInst *NA = new AllocaInst(ElTy, 0,
165 AI->getName() + "." + utostr(i), AI);
166 ElementAllocas.push_back(NA);
167 WorkList.push_back(NA); // Add to worklist for recursive processing
171 // Now that we have created the alloca instructions that we want to use,
172 // expand the getelementptr instructions to use them.
174 while (!AI->use_empty()) {
175 Instruction *User = cast<Instruction>(AI->use_back());
176 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
177 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
178 uint64_t Idx = cast<ConstantInt>(GEPI->getOperand(2))->getRawValue();
180 assert(Idx < ElementAllocas.size() && "Index out of range?");
181 AllocaInst *AllocaToUse = ElementAllocas[Idx];
184 if (GEPI->getNumOperands() == 3) {
185 // Do not insert a new getelementptr instruction with zero indices,
186 // only to have it optimized out later.
187 RepValue = AllocaToUse;
189 // We are indexing deeply into the structure, so we still need a
190 // getelement ptr instruction to finish the indexing. This may be
191 // expanded itself once the worklist is rerun.
193 std::string OldName = GEPI->getName(); // Steal the old name...
194 std::vector<Value*> NewArgs;
195 NewArgs.push_back(Constant::getNullValue(Type::IntTy));
196 NewArgs.insert(NewArgs.end(), GEPI->op_begin()+3, GEPI->op_end());
199 new GetElementPtrInst(AllocaToUse, NewArgs, OldName, GEPI);
202 // Move all of the users over to the new GEP.
203 GEPI->replaceAllUsesWith(RepValue);
204 // Delete the old GEP
205 GEPI->getParent()->getInstList().erase(GEPI);
207 assert(0 && "Unexpected instruction type!");
211 // Finally, delete the Alloca instruction
212 AI->getParent()->getInstList().erase(AI);
220 /// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
221 /// aggregate allocation.
223 bool SROA::isSafeUseOfAllocation(Instruction *User) {
224 if (!isa<GetElementPtrInst>(User)) return false;
226 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
227 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI);
229 // The GEP is safe to transform if it is of the form GEP <ptr>, 0, <cst>
231 I.getOperand() != Constant::getNullValue(I.getOperand()->getType()))
235 if (I == E || !isa<ConstantInt>(I.getOperand()))
238 // If this is a use of an array allocation, do a bit more checking for sanity.
239 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
240 uint64_t NumElements = AT->getNumElements();
242 // Check to make sure that index falls within the array. If not,
243 // something funny is going on, so we won't do the optimization.
245 if (cast<ConstantInt>(GEPI->getOperand(2))->getRawValue() >= NumElements)
249 // If there are any non-simple uses of this getelementptr, make sure to reject
251 return isSafeElementUse(GEPI);
254 /// isSafeElementUse - Check to see if this use is an allowed use for a
255 /// getelementptr instruction of an array aggregate allocation.
257 bool SROA::isSafeElementUse(Value *Ptr) {
258 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
260 Instruction *User = cast<Instruction>(*I);
261 switch (User->getOpcode()) {
262 case Instruction::Load: break;
263 case Instruction::Store:
264 // Store is ok if storing INTO the pointer, not storing the pointer
265 if (User->getOperand(0) == Ptr) return false;
267 case Instruction::GetElementPtr: {
268 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
269 if (GEP->getNumOperands() > 1) {
270 if (!isa<Constant>(GEP->getOperand(1)) ||
271 !cast<Constant>(GEP->getOperand(1))->isNullValue())
272 return false; // Using pointer arithmetic to navigate the array...
274 if (!isSafeElementUse(GEP)) return false;
278 DEBUG(std::cerr << " Transformation preventing inst: " << *User);
282 return true; // All users look ok :)
286 /// isSafeStructAllocaToPromote - Check to see if the specified allocation of a
287 /// structure can be broken down into elements.
289 bool SROA::isSafeAllocaToPromote(AllocationInst *AI) {
290 // Loop over the use list of the alloca. We can only transform it if all of
291 // the users are safe to transform.
293 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
295 if (!isSafeUseOfAllocation(cast<Instruction>(*I))) {
296 DEBUG(std::cerr << "Cannot transform: " << *AI << " due to user: "