LLVM API Documentation

Main Page | Namespace List | Class Hierarchy | Alphabetical List | Class List | Directories | File List | Namespace Members | Class Members | File Members | Related Pages

ScalarReplAggregates.cpp

Go to the documentation of this file.
00001 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===//
00002 // 
00003 //                     The LLVM Compiler Infrastructure
00004 //
00005 // This file was developed by the LLVM research group and is distributed under
00006 // the University of Illinois Open Source License. See LICENSE.TXT for details.
00007 // 
00008 //===----------------------------------------------------------------------===//
00009 //
00010 // This transformation implements the well known scalar replacement of
00011 // aggregates transformation.  This xform breaks up alloca instructions of
00012 // aggregate type (structure or array) into individual alloca instructions for
00013 // each member (if possible).  Then, if possible, it transforms the individual
00014 // alloca instructions into nice clean scalar SSA form.
00015 //
00016 // This combines a simple SRoA algorithm with the Mem2Reg algorithm because
00017 // often interact, especially for C++ programs.  As such, iterating between
00018 // SRoA, then Mem2Reg until we run out of things to promote works well.
00019 //
00020 //===----------------------------------------------------------------------===//
00021 
00022 #include "llvm/Transforms/Scalar.h"
00023 #include "llvm/Constants.h"
00024 #include "llvm/DerivedTypes.h"
00025 #include "llvm/Function.h"
00026 #include "llvm/Pass.h"
00027 #include "llvm/Instructions.h"
00028 #include "llvm/Analysis/Dominators.h"
00029 #include "llvm/Support/GetElementPtrTypeIterator.h"
00030 #include "llvm/Target/TargetData.h"
00031 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
00032 #include "llvm/Support/Debug.h"
00033 #include "llvm/ADT/Statistic.h"
00034 #include "llvm/ADT/StringExtras.h"
00035 using namespace llvm;
00036 
00037 namespace {
00038   Statistic<> NumReplaced("scalarrepl", "Number of allocas broken up");
00039   Statistic<> NumPromoted("scalarrepl", "Number of allocas promoted");
00040 
00041   struct SROA : public FunctionPass {
00042     bool runOnFunction(Function &F);
00043 
00044     bool performScalarRepl(Function &F);
00045     bool performPromotion(Function &F);
00046 
00047     // getAnalysisUsage - This pass does not require any passes, but we know it
00048     // will not alter the CFG, so say so.
00049     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
00050       AU.addRequired<DominatorTree>();
00051       AU.addRequired<DominanceFrontier>();
00052       AU.addRequired<TargetData>();
00053       AU.setPreservesCFG();
00054     }
00055 
00056   private:
00057     int isSafeElementUse(Value *Ptr);
00058     int isSafeUseOfAllocation(Instruction *User);
00059     int isSafeAllocaToScalarRepl(AllocationInst *AI);
00060     void CanonicalizeAllocaUsers(AllocationInst *AI);
00061     AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
00062   };
00063 
00064   RegisterOpt<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
00065 }
00066 
00067 // Public interface to the ScalarReplAggregates pass
00068 FunctionPass *llvm::createScalarReplAggregatesPass() { return new SROA(); }
00069 
00070 
00071 bool SROA::runOnFunction(Function &F) {
00072   bool Changed = performPromotion(F);
00073   while (1) {
00074     bool LocalChange = performScalarRepl(F);
00075     if (!LocalChange) break;   // No need to repromote if no scalarrepl
00076     Changed = true;
00077     LocalChange = performPromotion(F);
00078     if (!LocalChange) break;   // No need to re-scalarrepl if no promotion
00079   }
00080 
00081   return Changed;
00082 }
00083 
00084 
00085 bool SROA::performPromotion(Function &F) {
00086   std::vector<AllocaInst*> Allocas;
00087   const TargetData &TD = getAnalysis<TargetData>();
00088   DominatorTree     &DT = getAnalysis<DominatorTree>();
00089   DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
00090 
00091   BasicBlock &BB = F.getEntryBlock();  // Get the entry node for the function
00092 
00093   bool Changed = false;
00094   
00095   while (1) {
00096     Allocas.clear();
00097 
00098     // Find allocas that are safe to promote, by looking at all instructions in
00099     // the entry node
00100     for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
00101       if (AllocaInst *AI = dyn_cast<AllocaInst>(I))       // Is it an alloca?
00102         if (isAllocaPromotable(AI, TD))
00103           Allocas.push_back(AI);
00104 
00105     if (Allocas.empty()) break;
00106 
00107     PromoteMemToReg(Allocas, DT, DF, TD);
00108     NumPromoted += Allocas.size();
00109     Changed = true;
00110   }
00111 
00112   return Changed;
00113 }
00114 
00115 
00116 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
00117 // which runs on all of the malloc/alloca instructions in the function, removing
00118 // them if they are only used by getelementptr instructions.
00119 //
00120 bool SROA::performScalarRepl(Function &F) {
00121   std::vector<AllocationInst*> WorkList;
00122 
00123   // Scan the entry basic block, adding any alloca's and mallocs to the worklist
00124   BasicBlock &BB = F.getEntryBlock();
00125   for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
00126     if (AllocationInst *A = dyn_cast<AllocationInst>(I))
00127       WorkList.push_back(A);
00128 
00129   // Process the worklist
00130   bool Changed = false;
00131   while (!WorkList.empty()) {
00132     AllocationInst *AI = WorkList.back();
00133     WorkList.pop_back();
00134 
00135     // We cannot transform the allocation instruction if it is an array
00136     // allocation (allocations OF arrays are ok though), and an allocation of a
00137     // scalar value cannot be decomposed at all.
00138     //
00139     if (AI->isArrayAllocation() ||
00140         (!isa<StructType>(AI->getAllocatedType()) &&
00141          !isa<ArrayType>(AI->getAllocatedType()))) continue;
00142 
00143     // Check that all of the users of the allocation are capable of being
00144     // transformed.
00145     switch (isSafeAllocaToScalarRepl(AI)) {
00146     default: assert(0 && "Unexpected value!");
00147     case 0:  // Not safe to scalar replace.
00148       continue;
00149     case 1:  // Safe, but requires cleanup/canonicalizations first
00150       CanonicalizeAllocaUsers(AI);
00151     case 3:  // Safe to scalar replace.
00152       break;
00153     }
00154 
00155     DEBUG(std::cerr << "Found inst to xform: " << *AI);
00156     Changed = true;
00157     
00158     std::vector<AllocaInst*> ElementAllocas;
00159     if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
00160       ElementAllocas.reserve(ST->getNumContainedTypes());
00161       for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
00162         AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
00163                                         AI->getName() + "." + utostr(i), AI);
00164         ElementAllocas.push_back(NA);
00165         WorkList.push_back(NA);  // Add to worklist for recursive processing
00166       }
00167     } else {
00168       const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
00169       ElementAllocas.reserve(AT->getNumElements());
00170       const Type *ElTy = AT->getElementType();
00171       for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
00172         AllocaInst *NA = new AllocaInst(ElTy, 0,
00173                                         AI->getName() + "." + utostr(i), AI);
00174         ElementAllocas.push_back(NA);
00175         WorkList.push_back(NA);  // Add to worklist for recursive processing
00176       }
00177     }
00178     
00179     // Now that we have created the alloca instructions that we want to use,
00180     // expand the getelementptr instructions to use them.
00181     //
00182     while (!AI->use_empty()) {
00183       Instruction *User = cast<Instruction>(AI->use_back());
00184       GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
00185       // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
00186       uint64_t Idx = cast<ConstantInt>(GEPI->getOperand(2))->getRawValue();
00187       
00188       assert(Idx < ElementAllocas.size() && "Index out of range?");
00189       AllocaInst *AllocaToUse = ElementAllocas[Idx];
00190       
00191       Value *RepValue;
00192       if (GEPI->getNumOperands() == 3) {
00193         // Do not insert a new getelementptr instruction with zero indices, only
00194         // to have it optimized out later.
00195         RepValue = AllocaToUse;
00196       } else {
00197         // We are indexing deeply into the structure, so we still need a
00198         // getelement ptr instruction to finish the indexing.  This may be
00199         // expanded itself once the worklist is rerun.
00200         //
00201         std::string OldName = GEPI->getName();  // Steal the old name.
00202         std::vector<Value*> NewArgs;
00203         NewArgs.push_back(Constant::getNullValue(Type::IntTy));
00204         NewArgs.insert(NewArgs.end(), GEPI->op_begin()+3, GEPI->op_end());
00205         GEPI->setName("");
00206         RepValue = new GetElementPtrInst(AllocaToUse, NewArgs, OldName, GEPI);
00207       }
00208       
00209       // Move all of the users over to the new GEP.
00210       GEPI->replaceAllUsesWith(RepValue);
00211       // Delete the old GEP
00212       GEPI->eraseFromParent();
00213     }
00214 
00215     // Finally, delete the Alloca instruction
00216     AI->getParent()->getInstList().erase(AI);
00217     NumReplaced++;
00218   }
00219 
00220   return Changed;
00221 }
00222 
00223 
00224 /// isSafeElementUse - Check to see if this use is an allowed use for a
00225 /// getelementptr instruction of an array aggregate allocation.
00226 ///
00227 int SROA::isSafeElementUse(Value *Ptr) {
00228   for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
00229        I != E; ++I) {
00230     Instruction *User = cast<Instruction>(*I);
00231     switch (User->getOpcode()) {
00232     case Instruction::Load:  break;
00233     case Instruction::Store:
00234       // Store is ok if storing INTO the pointer, not storing the pointer
00235       if (User->getOperand(0) == Ptr) return 0;
00236       break;
00237     case Instruction::GetElementPtr: {
00238       GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
00239       if (GEP->getNumOperands() > 1) {
00240         if (!isa<Constant>(GEP->getOperand(1)) ||
00241             !cast<Constant>(GEP->getOperand(1))->isNullValue())
00242           return 0;  // Using pointer arithmetic to navigate the array...
00243       }
00244       if (!isSafeElementUse(GEP)) return 0;
00245       break;
00246     }
00247     default:
00248       DEBUG(std::cerr << "  Transformation preventing inst: " << *User);
00249       return 0;
00250     }
00251   }
00252   return 3;  // All users look ok :)
00253 }
00254 
00255 /// AllUsersAreLoads - Return true if all users of this value are loads.
00256 static bool AllUsersAreLoads(Value *Ptr) {
00257   for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
00258        I != E; ++I)
00259     if (cast<Instruction>(*I)->getOpcode() != Instruction::Load)
00260       return false;
00261   return true; 
00262 }
00263 
00264 /// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
00265 /// aggregate allocation.
00266 ///
00267 int SROA::isSafeUseOfAllocation(Instruction *User) {
00268   if (!isa<GetElementPtrInst>(User)) return 0;
00269 
00270   GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
00271   gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI);
00272 
00273   // The GEP is safe to transform if it is of the form GEP <ptr>, 0, <cst>
00274   if (I == E ||
00275       I.getOperand() != Constant::getNullValue(I.getOperand()->getType()))
00276     return 0;
00277 
00278   ++I;
00279   if (I == E) return 0;  // ran out of GEP indices??
00280 
00281   // If this is a use of an array allocation, do a bit more checking for sanity.
00282   if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
00283     uint64_t NumElements = AT->getNumElements();
00284 
00285     if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) {
00286       // Check to make sure that index falls within the array.  If not,
00287       // something funny is going on, so we won't do the optimization.
00288       //
00289       if (cast<ConstantInt>(GEPI->getOperand(2))->getRawValue() >= NumElements)
00290         return 0;
00291       
00292     } else {
00293       // If this is an array index and the index is not constant, we cannot
00294       // promote... that is unless the array has exactly one or two elements in
00295       // it, in which case we CAN promote it, but we have to canonicalize this
00296       // out if this is the only problem.
00297       if (NumElements == 1 || NumElements == 2)
00298         return AllUsersAreLoads(GEPI) ? 1 : 0;  // Canonicalization required!
00299       return 0;
00300     }
00301   }
00302 
00303   // If there are any non-simple uses of this getelementptr, make sure to reject
00304   // them.
00305   return isSafeElementUse(GEPI);
00306 }
00307 
00308 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
00309 /// an aggregate can be broken down into elements.  Return 0 if not, 3 if safe,
00310 /// or 1 if safe after canonicalization has been performed.
00311 ///
00312 int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) {
00313   // Loop over the use list of the alloca.  We can only transform it if all of
00314   // the users are safe to transform.
00315   //
00316   int isSafe = 3;
00317   for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
00318        I != E; ++I) {
00319     isSafe &= isSafeUseOfAllocation(cast<Instruction>(*I));
00320     if (isSafe == 0) {
00321       DEBUG(std::cerr << "Cannot transform: " << *AI << "  due to user: "
00322             << **I);
00323       return 0;
00324     }
00325   }
00326   // If we require cleanup, isSafe is now 1, otherwise it is 3.
00327   return isSafe;
00328 }
00329 
00330 /// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified
00331 /// allocation, but only if cleaned up, perform the cleanups required.
00332 void SROA::CanonicalizeAllocaUsers(AllocationInst *AI) {
00333   // At this point, we know that the end result will be SROA'd and promoted, so
00334   // we can insert ugly code if required so long as sroa+mem2reg will clean it
00335   // up.
00336   for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
00337        UI != E; ) {
00338     GetElementPtrInst *GEPI = cast<GetElementPtrInst>(*UI++);
00339     gep_type_iterator I = gep_type_begin(GEPI);
00340     ++I;
00341 
00342     if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
00343       uint64_t NumElements = AT->getNumElements();
00344       
00345       if (!isa<ConstantInt>(I.getOperand())) {
00346         if (NumElements == 1) {
00347           GEPI->setOperand(2, Constant::getNullValue(Type::IntTy));
00348         } else {
00349           assert(NumElements == 2 && "Unhandled case!");
00350           // All users of the GEP must be loads.  At each use of the GEP, insert
00351           // two loads of the appropriate indexed GEP and select between them.
00352           Value *IsOne = BinaryOperator::createSetNE(I.getOperand(),
00353                               Constant::getNullValue(I.getOperand()->getType()),
00354                                                      "isone", GEPI);
00355           // Insert the new GEP instructions, which are properly indexed.
00356           std::vector<Value*> Indices(GEPI->op_begin()+1, GEPI->op_end());
00357           Indices[1] = Constant::getNullValue(Type::IntTy);
00358           Value *ZeroIdx = new GetElementPtrInst(GEPI->getOperand(0), Indices,
00359                                                  GEPI->getName()+".0", GEPI);
00360           Indices[1] = ConstantInt::get(Type::IntTy, 1);
00361           Value *OneIdx = new GetElementPtrInst(GEPI->getOperand(0), Indices,
00362                                                 GEPI->getName()+".1", GEPI);
00363           // Replace all loads of the variable index GEP with loads from both
00364           // indexes and a select.
00365           while (!GEPI->use_empty()) {
00366             LoadInst *LI = cast<LoadInst>(GEPI->use_back());
00367             Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
00368             Value *One  = new LoadInst(OneIdx , LI->getName()+".1", LI);
00369             Value *R = new SelectInst(IsOne, One, Zero, LI->getName(), LI);
00370             LI->replaceAllUsesWith(R);
00371             LI->eraseFromParent();
00372           }
00373           GEPI->eraseFromParent();
00374         }
00375       }
00376     }
00377   }
00378 }