//===- PromoteMemoryToRegister.cpp - Convert allocas to registers ---------===//
-//
+//
// The LLVM Compiler Infrastructure
//
-// This file was developed by the LLVM research group and is distributed under
-// the University of Illinois Open Source License. See LICENSE.TXT for details.
-//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
//===----------------------------------------------------------------------===//
//
-// This file promote memory references to be register references. It promotes
-// alloca instructions which only have loads and stores as uses (or that have
-// PHI nodes which are only loaded from). An alloca is transformed by using
-// dominator frontiers to place PHI nodes, then traversing the function in
-// depth-first order to rewrite loads and stores as appropriate. This is just
-// the standard SSA construction algorithm to construct "pruned" SSA form.
+// This file promotes memory references to be register references. It promotes
+// alloca instructions which only have loads and stores as uses. An alloca is
+// transformed by using dominator frontiers to place PHI nodes, then traversing
+// the function in depth-first order to rewrite loads and stores as appropriate.
+// This is just the standard SSA construction algorithm to construct "pruned"
+// SSA form.
//
//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "mem2reg"
#include "llvm/Transforms/Utils/PromoteMemToReg.h"
-#include "llvm/Analysis/Dominators.h"
-#include "llvm/Instructions.h"
+#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
-#include "llvm/Constant.h"
+#include "llvm/Instructions.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Analysis/AliasSetTracker.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/STLExtras.h"
#include "llvm/Support/CFG.h"
-#include "llvm/Support/StableBasicBlockNumbering.h"
-#include "llvm/ADT/StringExtras.h"
#include <algorithm>
using namespace llvm;
+STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
+STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
+STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
+STATISTIC(NumPHIInsert, "Number of PHI nodes inserted");
+
+namespace llvm {
+template<>
+struct DenseMapInfo<std::pair<BasicBlock*, unsigned> > {
+ typedef std::pair<BasicBlock*, unsigned> EltTy;
+ static inline EltTy getEmptyKey() {
+ return EltTy(reinterpret_cast<BasicBlock*>(-1), ~0U);
+ }
+ static inline EltTy getTombstoneKey() {
+ return EltTy(reinterpret_cast<BasicBlock*>(-2), 0U);
+ }
+ static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
+ return DenseMapInfo<void*>::getHashValue(Val.first) + Val.second*2;
+ }
+ static bool isEqual(const EltTy &LHS, const EltTy &RHS) {
+ return LHS == RHS;
+ }
+};
+}
+
/// isAllocaPromotable - Return true if this alloca is legal for promotion.
-/// This is true if there are only loads and stores to the alloca... of if there
-/// is a PHI node using the address which can be trivially transformed.
+/// This is true if there are only loads and stores to the alloca.
///
-bool llvm::isAllocaPromotable(const AllocaInst *AI, const TargetData &TD) {
+bool llvm::isAllocaPromotable(const AllocaInst *AI) {
// FIXME: If the memory unit is of pointer or integer type, we can permit
// assignments to subsections of the memory unit.
- // Only allow direct loads and stores...
+ // Only allow direct and non-volatile loads and stores...
for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
UI != UE; ++UI) // Loop over all of the uses of the alloca
- if (isa<LoadInst>(*UI)) {
- // noop
+ if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
+ if (LI->isVolatile())
+ return false;
} else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
if (SI->getOperand(0) == AI)
return false; // Don't allow a store OF the AI, only INTO the AI.
- } else if (const PHINode *PN = dyn_cast<PHINode>(*UI)) {
- // We only support PHI nodes in a few simple cases. The PHI node is only
- // allowed to have one use, which must be a load instruction, and can only
- // use alloca instructions (no random pointers). Also, there cannot be
- // any accesses to AI between the PHI node and the use of the PHI.
- if (!PN->hasOneUse()) return false;
-
- // Our transformation causes the unconditional loading of all pointer
- // operands to the PHI node. Because this could cause a fault if there is
- // a critical edge in the CFG and if one of the pointers is illegal, we
- // refuse to promote PHI nodes unless they are obviously safe. For now,
- // obviously safe means that all of the operands are allocas.
- //
- // If we wanted to extend this code to break critical edges, this
- // restriction could be relaxed, and we could even handle uses of the PHI
- // node that are volatile loads or stores.
- //
- for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
- if (!isa<AllocaInst>(PN->getIncomingValue(i)))
- return false;
-
- // Now make sure the one user instruction is in the same basic block as
- // the PHI, and that there are no loads or stores between the PHI node and
- // the access.
- BasicBlock::const_iterator UI = cast<Instruction>(PN->use_back());
- if (!isa<LoadInst>(UI) || cast<LoadInst>(UI)->isVolatile()) return false;
-
- // Scan looking for memory accesses.
- // FIXME: this should REALLY use alias analysis.
- for (--UI; !isa<PHINode>(UI); --UI)
- if (isa<LoadInst>(UI) || isa<StoreInst>(UI) || isa<CallInst>(UI))
- return false;
-
- // If we got this far, we can promote the PHI use.
- } else if (const SelectInst *SI = dyn_cast<SelectInst>(*UI)) {
- // We only support selects in a few simple cases. The select is only
- // allowed to have one use, which must be a load instruction, and can only
- // use alloca instructions (no random pointers). Also, there cannot be
- // any accesses to AI between the PHI node and the use of the PHI.
- if (!SI->hasOneUse()) return false;
-
- // Our transformation causes the unconditional loading of all pointer
- // operands of the select. Because this could cause a fault if there is a
- // critical edge in the CFG and if one of the pointers is illegal, we
- // refuse to promote the select unless it is obviously safe. For now,
- // obviously safe means that all of the operands are allocas.
- //
- if (!isa<AllocaInst>(SI->getOperand(1)) ||
- !isa<AllocaInst>(SI->getOperand(2)))
+ if (SI->isVolatile())
+ return false;
+ } else if (const BitCastInst *BC = dyn_cast<BitCastInst>(*UI)) {
+ // A bitcast that does not feed into debug info inhibits promotion.
+ if (!BC->hasOneUse() || !isa<DbgInfoIntrinsic>(*BC->use_begin()))
+ return false;
+ // If the only use is by debug info, this alloca will not exist in
+ // non-debug code, so don't try to promote; this ensures the same
+ // codegen with debug info. Otherwise, debug info should not
+ // inhibit promotion (but we must examine other uses).
+ if (AI->hasOneUse())
return false;
-
- // Now make sure the one user instruction is in the same basic block as
- // the PHI, and that there are no loads or stores between the PHI node and
- // the access.
- BasicBlock::const_iterator UI = cast<Instruction>(SI->use_back());
- if (!isa<LoadInst>(UI) || cast<LoadInst>(UI)->isVolatile()) return false;
-
- // Scan looking for memory accesses.
- // FIXME: this should REALLY use alias analysis.
- for (--UI; &*UI != SI; --UI)
- if (isa<LoadInst>(UI) || isa<StoreInst>(UI) || isa<CallInst>(UI))
- return false;
-
- // If we got this far, we can promote the select use.
} else {
- return false; // Not a load, store, or promotable PHI?
+ return false;
}
-
+
return true;
}
namespace {
+ struct AllocaInfo;
+
+ // Data package used by RenamePass()
+ class RenamePassData {
+ public:
+ typedef std::vector<Value *> ValVector;
+
+ RenamePassData() : BB(NULL), Pred(NULL), Values() {}
+ RenamePassData(BasicBlock *B, BasicBlock *P,
+ const ValVector &V) : BB(B), Pred(P), Values(V) {}
+ BasicBlock *BB;
+ BasicBlock *Pred;
+ ValVector Values;
+
+ void swap(RenamePassData &RHS) {
+ std::swap(BB, RHS.BB);
+ std::swap(Pred, RHS.Pred);
+ Values.swap(RHS.Values);
+ }
+ };
+
+ /// LargeBlockInfo - This assigns and keeps a per-bb relative ordering of
+ /// load/store instructions in the block that directly load or store an alloca.
+ ///
+ /// This functionality is important because it avoids scanning large basic
+ /// blocks multiple times when promoting many allocas in the same block.
+ class LargeBlockInfo {
+ /// InstNumbers - For each instruction that we track, keep the index of the
+ /// instruction. The index starts out as the number of the instruction from
+ /// the start of the block.
+ DenseMap<const Instruction *, unsigned> InstNumbers;
+ public:
+
+ /// isInterestingInstruction - This code only looks at accesses to allocas.
+ static bool isInterestingInstruction(const Instruction *I) {
+ return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
+ (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
+ }
+
+ /// getInstructionIndex - Get or calculate the index of the specified
+ /// instruction.
+ unsigned getInstructionIndex(const Instruction *I) {
+ assert(isInterestingInstruction(I) &&
+ "Not a load/store to/from an alloca?");
+
+ // If we already have this instruction number, return it.
+ DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
+ if (It != InstNumbers.end()) return It->second;
+
+ // Scan the whole block to get the instruction. This accumulates
+ // information for every interesting instruction in the block, in order to
+ // avoid gratuitus rescans.
+ const BasicBlock *BB = I->getParent();
+ unsigned InstNo = 0;
+ for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end();
+ BBI != E; ++BBI)
+ if (isInterestingInstruction(BBI))
+ InstNumbers[BBI] = InstNo++;
+ It = InstNumbers.find(I);
+
+ assert(It != InstNumbers.end() && "Didn't insert instruction?");
+ return It->second;
+ }
+
+ void deleteValue(const Instruction *I) {
+ InstNumbers.erase(I);
+ }
+
+ void clear() {
+ InstNumbers.clear();
+ }
+ };
+
struct PromoteMem2Reg {
- // Allocas - The alloca instructions being promoted
+ /// Allocas - The alloca instructions being promoted.
+ ///
std::vector<AllocaInst*> Allocas;
DominatorTree &DT;
DominanceFrontier &DF;
- const TargetData &TD;
- // AllocaLookup - Reverse mapping of Allocas
+ /// AST - An AliasSetTracker object to update. If null, don't update it.
+ ///
+ AliasSetTracker *AST;
+
+ /// AllocaLookup - Reverse mapping of Allocas.
+ ///
std::map<AllocaInst*, unsigned> AllocaLookup;
- // NewPhiNodes - The PhiNodes we're adding.
- std::map<BasicBlock*, std::vector<PHINode*> > NewPhiNodes;
-
- // Visited - The set of basic blocks the renamer has already visited.
- std::set<BasicBlock*> Visited;
-
- // BBNumbers - Contains a stable numbering of basic blocks to avoid
- // non-determinstic behavior.
- StableBasicBlockNumbering BBNumbers;
-
+ /// NewPhiNodes - The PhiNodes we're adding.
+ ///
+ DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
+
+ /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
+ /// it corresponds to.
+ DenseMap<PHINode*, unsigned> PhiToAllocaMap;
+
+ /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
+ /// each alloca that is of pointer type, we keep track of what to copyValue
+ /// to the inserted PHI nodes here.
+ ///
+ std::vector<Value*> PointerAllocaValues;
+
+ /// Visited - The set of basic blocks the renamer has already visited.
+ ///
+ SmallPtrSet<BasicBlock*, 16> Visited;
+
+ /// BBNumbers - Contains a stable numbering of basic blocks to avoid
+ /// non-determinstic behavior.
+ DenseMap<BasicBlock*, unsigned> BBNumbers;
+
+ /// BBNumPreds - Lazily compute the number of predecessors a block has.
+ DenseMap<const BasicBlock*, unsigned> BBNumPreds;
public:
PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
- DominanceFrontier &df, const TargetData &td)
- : Allocas(A), DT(dt), DF(df), TD(td) {}
+ DominanceFrontier &df, AliasSetTracker *ast)
+ : Allocas(A), DT(dt), DF(df), AST(ast) {}
void run();
+ /// properlyDominates - Return true if I1 properly dominates I2.
+ ///
+ bool properlyDominates(Instruction *I1, Instruction *I2) const {
+ if (InvokeInst *II = dyn_cast<InvokeInst>(I1))
+ I1 = II->getNormalDest()->begin();
+ return DT.properlyDominates(I1->getParent(), I2->getParent());
+ }
+
+ /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
+ ///
+ bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
+ return DT.dominates(BB1, BB2);
+ }
+
private:
- void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
- std::set<PHINode*> &DeadPHINodes);
- void PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI);
- void PromoteLocallyUsedAllocas(BasicBlock *BB,
- const std::vector<AllocaInst*> &AIs);
+ void RemoveFromAllocasList(unsigned &AllocaIdx) {
+ Allocas[AllocaIdx] = Allocas.back();
+ Allocas.pop_back();
+ --AllocaIdx;
+ }
+ unsigned getNumPreds(const BasicBlock *BB) {
+ unsigned &NP = BBNumPreds[BB];
+ if (NP == 0)
+ NP = std::distance(pred_begin(BB), pred_end(BB))+1;
+ return NP-1;
+ }
+
+ void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
+ AllocaInfo &Info);
+ void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
+ const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
+ SmallPtrSet<BasicBlock*, 32> &LiveInBlocks);
+
+ void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
+ LargeBlockInfo &LBI);
+ void PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
+ LargeBlockInfo &LBI);
+
+
void RenamePass(BasicBlock *BB, BasicBlock *Pred,
- std::vector<Value*> &IncVals);
+ RenamePassData::ValVector &IncVals,
+ std::vector<RenamePassData> &Worklist);
bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
- std::set<PHINode*> &InsertedPHINodes);
+ SmallPtrSet<PHINode*, 16> &InsertedPHINodes);
+ };
+
+ struct AllocaInfo {
+ std::vector<BasicBlock*> DefiningBlocks;
+ std::vector<BasicBlock*> UsingBlocks;
+
+ StoreInst *OnlyStore;
+ BasicBlock *OnlyBlock;
+ bool OnlyUsedInOneBlock;
+
+ Value *AllocaPointerVal;
+
+ void clear() {
+ DefiningBlocks.clear();
+ UsingBlocks.clear();
+ OnlyStore = 0;
+ OnlyBlock = 0;
+ OnlyUsedInOneBlock = true;
+ AllocaPointerVal = 0;
+ }
+
+ /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our
+ /// ivars.
+ void AnalyzeAlloca(AllocaInst *AI) {
+ clear();
+
+ // As we scan the uses of the alloca instruction, keep track of stores,
+ // and decide whether all of the loads and stores to the alloca are within
+ // the same basic block.
+ for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
+ UI != E;) {
+ Instruction *User = cast<Instruction>(*UI++);
+ if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
+ // Remove any uses of this alloca in DbgInfoInstrinsics.
+ assert(BC->hasOneUse() && "Unexpected alloca uses!");
+ DbgInfoIntrinsic *DI = cast<DbgInfoIntrinsic>(*BC->use_begin());
+ DI->eraseFromParent();
+ BC->eraseFromParent();
+ continue;
+ }
+
+ if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
+ // Remember the basic blocks which define new values for the alloca
+ DefiningBlocks.push_back(SI->getParent());
+ AllocaPointerVal = SI->getOperand(0);
+ OnlyStore = SI;
+ } else {
+ LoadInst *LI = cast<LoadInst>(User);
+ // Otherwise it must be a load instruction, keep track of variable
+ // reads.
+ UsingBlocks.push_back(LI->getParent());
+ AllocaPointerVal = LI;
+ }
+
+ if (OnlyUsedInOneBlock) {
+ if (OnlyBlock == 0)
+ OnlyBlock = User->getParent();
+ else if (OnlyBlock != User->getParent())
+ OnlyUsedInOneBlock = false;
+ }
+ }
+ }
};
} // end of anonymous namespace
+
void PromoteMem2Reg::run() {
Function &F = *DF.getRoot()->getParent();
- // LocallyUsedAllocas - Keep track of all of the alloca instructions which are
- // only used in a single basic block. These instructions can be efficiently
- // promoted by performing a single linear scan over that one block. Since
- // individual basic blocks are sometimes large, we group together all allocas
- // that are live in a single basic block by the basic block they are live in.
- std::map<BasicBlock*, std::vector<AllocaInst*> > LocallyUsedAllocas;
+ if (AST) PointerAllocaValues.resize(Allocas.size());
+ AllocaInfo Info;
+ LargeBlockInfo LBI;
for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
AllocaInst *AI = Allocas[AllocaNum];
- assert(isAllocaPromotable(AI, TD) &&
+ assert(isAllocaPromotable(AI) &&
"Cannot promote non-promotable alloca!");
assert(AI->getParent()->getParent() == &F &&
"All allocas should be in the same function, which is same as DF!");
if (AI->use_empty()) {
// If there are no uses of the alloca, just delete it now.
- AI->getParent()->getInstList().erase(AI);
+ if (AST) AST->deleteValue(AI);
+ AI->eraseFromParent();
// Remove the alloca from the Allocas list, since it has been processed
- Allocas[AllocaNum] = Allocas.back();
- Allocas.pop_back();
- --AllocaNum;
+ RemoveFromAllocasList(AllocaNum);
+ ++NumDeadAlloca;
continue;
}
-
+
// Calculate the set of read and write-locations for each alloca. This is
// analogous to finding the 'uses' and 'definitions' of each variable.
- std::vector<BasicBlock*> DefiningBlocks;
- std::vector<BasicBlock*> UsingBlocks;
-
- BasicBlock *OnlyBlock = 0;
- bool OnlyUsedInOneBlock = true;
-
- // As we scan the uses of the alloca instruction, keep track of stores, and
- // decide whether all of the loads and stores to the alloca are within the
- // same basic block.
- RestartUseScan:
- for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E;++U){
- Instruction *User = cast<Instruction>(*U);
- if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
- // Remember the basic blocks which define new values for the alloca
- DefiningBlocks.push_back(SI->getParent());
- } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
- // Otherwise it must be a load instruction, keep track of variable reads
- UsingBlocks.push_back(LI->getParent());
- } else if (SelectInst *SI = dyn_cast<SelectInst>(User)) {
- // Because of the restrictions we placed on Select instruction uses
- // above things are very simple. Transform the PHI of addresses into a
- // select of loaded values.
- LoadInst *Load = cast<LoadInst>(SI->use_back());
- std::string LoadName = Load->getName(); Load->setName("");
-
- Value *TrueVal = new LoadInst(SI->getOperand(1),
- SI->getOperand(1)->getName()+".val", SI);
- Value *FalseVal = new LoadInst(SI->getOperand(2),
- SI->getOperand(2)->getName()+".val", SI);
-
- Value *NewSI = new SelectInst(SI->getOperand(0), TrueVal,
- FalseVal, Load->getName(), SI);
- Load->replaceAllUsesWith(NewSI);
- Load->getParent()->getInstList().erase(Load);
- SI->getParent()->getInstList().erase(SI);
-
- // Restart our scan of uses...
- DefiningBlocks.clear();
- UsingBlocks.clear();
- goto RestartUseScan;
- } else {
- // Because of the restrictions we placed on PHI node uses above, the PHI
- // node reads the block in any using predecessors. Transform the PHI of
- // addresses into a PHI of loaded values.
- PHINode *PN = cast<PHINode>(User);
- assert(PN->hasOneUse() && "Cannot handle PHI Node with != 1 use!");
- LoadInst *PNUser = cast<LoadInst>(PN->use_back());
- std::string PNUserName = PNUser->getName(); PNUser->setName("");
-
- // Create the new PHI node and insert load instructions as appropriate.
- PHINode *NewPN = new PHINode(AI->getAllocatedType(), PNUserName, PN);
- std::map<BasicBlock*, LoadInst*> NewLoads;
- for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
- BasicBlock *Pred = PN->getIncomingBlock(i);
- LoadInst *&NewLoad = NewLoads[Pred];
- if (NewLoad == 0) // Insert the new load in the predecessor
- NewLoad = new LoadInst(PN->getIncomingValue(i),
- PN->getIncomingValue(i)->getName()+".val",
- Pred->getTerminator());
- NewPN->addIncoming(NewLoad, Pred);
- }
-
- // Remove the old load.
- PNUser->replaceAllUsesWith(NewPN);
- PNUser->getParent()->getInstList().erase(PNUser);
-
- // Remove the old PHI node.
- PN->getParent()->getInstList().erase(PN);
-
- // Restart our scan of uses...
- DefiningBlocks.clear();
- UsingBlocks.clear();
- goto RestartUseScan;
- }
-
- if (OnlyUsedInOneBlock) {
- if (OnlyBlock == 0)
- OnlyBlock = User->getParent();
- else if (OnlyBlock != User->getParent())
- OnlyUsedInOneBlock = false;
+ Info.AnalyzeAlloca(AI);
+
+ // If there is only a single store to this value, replace any loads of
+ // it that are directly dominated by the definition with the value stored.
+ if (Info.DefiningBlocks.size() == 1) {
+ RewriteSingleStoreAlloca(AI, Info, LBI);
+
+ // Finally, after the scan, check to see if the store is all that is left.
+ if (Info.UsingBlocks.empty()) {
+ // Remove the (now dead) store and alloca.
+ Info.OnlyStore->eraseFromParent();
+ LBI.deleteValue(Info.OnlyStore);
+
+ if (AST) AST->deleteValue(AI);
+ AI->eraseFromParent();
+ LBI.deleteValue(AI);
+
+ // The alloca has been processed, move on.
+ RemoveFromAllocasList(AllocaNum);
+
+ ++NumSingleStore;
+ continue;
}
}
-
+
// If the alloca is only read and written in one basic block, just perform a
// linear sweep over the block to eliminate it.
- if (OnlyUsedInOneBlock) {
- LocallyUsedAllocas[OnlyBlock].push_back(AI);
-
- // Remove the alloca from the Allocas list, since it will be processed.
- Allocas[AllocaNum] = Allocas.back();
- Allocas.pop_back();
- --AllocaNum;
- continue;
- }
-
- // If we haven't computed a numbering for the BB's in the function, do so
- // now.
- BBNumbers.compute(F);
-
- // Compute the locations where PhiNodes need to be inserted. Look at the
- // dominance frontier of EACH basic-block we have a write in.
- //
- unsigned CurrentVersion = 0;
- std::set<PHINode*> InsertedPHINodes;
- std::vector<unsigned> DFBlocks;
- while (!DefiningBlocks.empty()) {
- BasicBlock *BB = DefiningBlocks.back();
- DefiningBlocks.pop_back();
-
- // Look up the DF for this write, add it to PhiNodes
- DominanceFrontier::const_iterator it = DF.find(BB);
- if (it != DF.end()) {
- const DominanceFrontier::DomSetType &S = it->second;
-
- // In theory we don't need the indirection through the DFBlocks vector.
- // In practice, the order of calling QueuePhiNode would depend on the
- // (unspecified) ordering of basic blocks in the dominance frontier,
- // which would give PHI nodes non-determinstic subscripts. Fix this by
- // processing blocks in order of the occurance in the function.
- for (DominanceFrontier::DomSetType::iterator P = S.begin(),PE = S.end();
- P != PE; ++P)
- DFBlocks.push_back(BBNumbers.getNumber(*P));
-
- // Sort by which the block ordering in the function.
- std::sort(DFBlocks.begin(), DFBlocks.end());
-
- for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
- BasicBlock *BB = BBNumbers.getBlock(DFBlocks[i]);
- if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
- DefiningBlocks.push_back(BB);
+ if (Info.OnlyUsedInOneBlock) {
+ PromoteSingleBlockAlloca(AI, Info, LBI);
+
+ // Finally, after the scan, check to see if the stores are all that is
+ // left.
+ if (Info.UsingBlocks.empty()) {
+
+ // Remove the (now dead) stores and alloca.
+ while (!AI->use_empty()) {
+ StoreInst *SI = cast<StoreInst>(AI->use_back());
+ SI->eraseFromParent();
+ LBI.deleteValue(SI);
}
- DFBlocks.clear();
+
+ if (AST) AST->deleteValue(AI);
+ AI->eraseFromParent();
+ LBI.deleteValue(AI);
+
+ // The alloca has been processed, move on.
+ RemoveFromAllocasList(AllocaNum);
+
+ ++NumLocalPromoted;
+ continue;
}
}
-
- // Now that we have inserted PHI nodes along the Iterated Dominance Frontier
- // of the writes to the variable, scan through the reads of the variable,
- // marking PHI nodes which are actually necessary as alive (by removing them
- // from the InsertedPHINodes set). This is not perfect: there may PHI
- // marked alive because of loads which are dominated by stores, but there
- // will be no unmarked PHI nodes which are actually used.
- //
- for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i)
- MarkDominatingPHILive(UsingBlocks[i], AllocaNum, InsertedPHINodes);
- UsingBlocks.clear();
-
- // If there are any PHI nodes which are now known to be dead, remove them!
- for (std::set<PHINode*>::iterator I = InsertedPHINodes.begin(),
- E = InsertedPHINodes.end(); I != E; ++I) {
- PHINode *PN = *I;
- std::vector<PHINode*> &BBPNs = NewPhiNodes[PN->getParent()];
- BBPNs[AllocaNum] = 0;
-
- // Check to see if we just removed the last inserted PHI node from this
- // basic block. If so, remove the entry for the basic block.
- bool HasOtherPHIs = false;
- for (unsigned i = 0, e = BBPNs.size(); i != e; ++i)
- if (BBPNs[i]) {
- HasOtherPHIs = true;
- break;
- }
- if (!HasOtherPHIs)
- NewPhiNodes.erase(PN->getParent());
-
- PN->getParent()->getInstList().erase(PN);
+
+ // If we haven't computed a numbering for the BB's in the function, do so
+ // now.
+ if (BBNumbers.empty()) {
+ unsigned ID = 0;
+ for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
+ BBNumbers[I] = ID++;
}
- // Keep the reverse mapping of the 'Allocas' array.
+ // If we have an AST to keep updated, remember some pointer value that is
+ // stored into the alloca.
+ if (AST)
+ PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
+
+ // Keep the reverse mapping of the 'Allocas' array for the rename pass.
AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
- }
-
- // Process all allocas which are only used in a single basic block.
- for (std::map<BasicBlock*, std::vector<AllocaInst*> >::iterator I =
- LocallyUsedAllocas.begin(), E = LocallyUsedAllocas.end(); I != E; ++I){
- const std::vector<AllocaInst*> &Allocas = I->second;
- assert(!Allocas.empty() && "empty alloca list??");
-
- // It's common for there to only be one alloca in the list. Handle it
- // efficiently.
- if (Allocas.size() == 1)
- PromoteLocallyUsedAlloca(I->first, Allocas[0]);
- else
- PromoteLocallyUsedAllocas(I->first, Allocas);
+
+ // At this point, we're committed to promoting the alloca using IDF's, and
+ // the standard SSA construction algorithm. Determine which blocks need PHI
+ // nodes and see if we can optimize out some work by avoiding insertion of
+ // dead phi nodes.
+ DetermineInsertionPoint(AI, AllocaNum, Info);
}
if (Allocas.empty())
return; // All of the allocas must have been trivial!
+ LBI.clear();
+
+
// Set the incoming values for the basic block to be null values for all of
// the alloca's. We do this in case there is a load of a value that has not
// been stored yet. In this case, it will get this null value.
//
- std::vector<Value *> Values(Allocas.size());
+ RenamePassData::ValVector Values(Allocas.size());
for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
- Values[i] = Constant::getNullValue(Allocas[i]->getAllocatedType());
+ Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
// Walks all basic blocks in the function performing the SSA rename algorithm
// and inserting the phi nodes we marked as necessary
//
- RenamePass(F.begin(), 0, Values);
-
+ std::vector<RenamePassData> RenamePassWorkList;
+ RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
+ do {
+ RenamePassData RPD;
+ RPD.swap(RenamePassWorkList.back());
+ RenamePassWorkList.pop_back();
+ // RenamePass may add new worklist entries.
+ RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
+ } while (!RenamePassWorkList.empty());
+
// The renamer uses the Visited set to avoid infinite loops. Clear it now.
Visited.clear();
- // Remove the allocas themselves from the function...
+ // Remove the allocas themselves from the function.
for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
Instruction *A = Allocas[i];
// Just delete the users now.
//
if (!A->use_empty())
- A->replaceAllUsesWith(Constant::getNullValue(A->getType()));
- A->getParent()->getInstList().erase(A);
+ A->replaceAllUsesWith(UndefValue::get(A->getType()));
+ if (AST) AST->deleteValue(A);
+ A->eraseFromParent();
}
+
+ // Loop over all of the PHI nodes and see if there are any that we can get
+ // rid of because they merge all of the same incoming values. This can
+ // happen due to undef values coming into the PHI nodes. This process is
+ // iterative, because eliminating one PHI node can cause others to be removed.
+ bool EliminatedAPHI = true;
+ while (EliminatedAPHI) {
+ EliminatedAPHI = false;
+
+ for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
+ NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
+ PHINode *PN = I->second;
+
+ // If this PHI node merges one value and/or undefs, get the value.
+ if (Value *V = PN->hasConstantValue(&DT)) {
+ if (AST && isa<PointerType>(PN->getType()))
+ AST->deleteValue(PN);
+ PN->replaceAllUsesWith(V);
+ PN->eraseFromParent();
+ NewPhiNodes.erase(I++);
+ EliminatedAPHI = true;
+ continue;
+ }
+ ++I;
+ }
+ }
+
// At this point, the renamer has added entries to PHI nodes for all reachable
- // code. Unfortunately, there may be blocks which are not reachable, which
- // the renamer hasn't traversed. If this is the case, the PHI nodes may not
+ // code. Unfortunately, there may be unreachable blocks which the renamer
+ // hasn't traversed. If this is the case, the PHI nodes may not
// have incoming values for all predecessors. Loop over all PHI nodes we have
- // created, inserting null constants if they are missing any incoming values.
+ // created, inserting undef values if they are missing any incoming values.
//
- for (std::map<BasicBlock*, std::vector<PHINode *> >::iterator I =
+ for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
-
- std::vector<BasicBlock*> Preds(pred_begin(I->first), pred_end(I->first));
- std::vector<PHINode*> &PNs = I->second;
- assert(!PNs.empty() && "Empty PHI node list??");
+ // We want to do this once per basic block. As such, only process a block
+ // when we find the PHI that is the first entry in the block.
+ PHINode *SomePHI = I->second;
+ BasicBlock *BB = SomePHI->getParent();
+ if (&BB->front() != SomePHI)
+ continue;
// Only do work here if there the PHI nodes are missing incoming values. We
// know that all PHI nodes that were inserted in a block will have the same
- // number of incoming values, so we can just check any PHI node.
- PHINode *FirstPHI;
- for (unsigned i = 0; (FirstPHI = PNs[i]) == 0; ++i)
- /*empty*/;
-
- if (Preds.size() != FirstPHI->getNumIncomingValues()) {
- // Ok, now we know that all of the PHI nodes are missing entries for some
- // basic blocks. Start by sorting the incoming predecessors for efficient
- // access.
- std::sort(Preds.begin(), Preds.end());
-
- // Now we loop through all BB's which have entries in FirstPHI and remove
- // them from the Preds list.
- for (unsigned i = 0, e = FirstPHI->getNumIncomingValues(); i != e; ++i) {
- // Do a log(n) search of the Preds list for the entry we want.
- std::vector<BasicBlock*>::iterator EntIt =
- std::lower_bound(Preds.begin(), Preds.end(),
- FirstPHI->getIncomingBlock(i));
- assert(EntIt != Preds.end() && *EntIt == FirstPHI->getIncomingBlock(i)&&
- "PHI node has entry for a block which is not a predecessor!");
-
- // Remove the entry
- Preds.erase(EntIt);
- }
+ // number of incoming values, so we can just check any of them.
+ if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
+ continue;
- // At this point, the blocks left in the preds list must have dummy
- // entries inserted into every PHI nodes for the block.
- for (unsigned i = 0, e = PNs.size(); i != e; ++i)
- if (PHINode *PN = PNs[i]) {
- Value *NullVal = Constant::getNullValue(PN->getType());
- for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
- PN->addIncoming(NullVal, Preds[pred]);
- }
+ // Get the preds for BB.
+ SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
+
+ // Ok, now we know that all of the PHI nodes are missing entries for some
+ // basic blocks. Start by sorting the incoming predecessors for efficient
+ // access.
+ std::sort(Preds.begin(), Preds.end());
+
+ // Now we loop through all BB's which have entries in SomePHI and remove
+ // them from the Preds list.
+ for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
+ // Do a log(n) search of the Preds list for the entry we want.
+ SmallVector<BasicBlock*, 16>::iterator EntIt =
+ std::lower_bound(Preds.begin(), Preds.end(),
+ SomePHI->getIncomingBlock(i));
+ assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
+ "PHI node has entry for a block which is not a predecessor!");
+
+ // Remove the entry
+ Preds.erase(EntIt);
+ }
+
+ // At this point, the blocks left in the preds list must have dummy
+ // entries inserted into every PHI nodes for the block. Update all the phi
+ // nodes in this block that we are inserting (there could be phis before
+ // mem2reg runs).
+ unsigned NumBadPreds = SomePHI->getNumIncomingValues();
+ BasicBlock::iterator BBI = BB->begin();
+ while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
+ SomePHI->getNumIncomingValues() == NumBadPreds) {
+ Value *UndefVal = UndefValue::get(SomePHI->getType());
+ for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
+ SomePHI->addIncoming(UndefVal, Preds[pred]);
}
}
+
+ NewPhiNodes.clear();
}
-// MarkDominatingPHILive - Mem2Reg wants to construct "pruned" SSA form, not
-// "minimal" SSA form. To do this, it inserts all of the PHI nodes on the IDF
-// as usual (inserting the PHI nodes in the DeadPHINodes set), then processes
-// each read of the variable. For each block that reads the variable, this
-// function is called, which removes used PHI nodes from the DeadPHINodes set.
-// After all of the reads have been processed, any PHI nodes left in the
-// DeadPHINodes set are removed.
-//
-void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
- std::set<PHINode*> &DeadPHINodes) {
- // Scan the immediate dominators of this block looking for a block which has a
- // PHI node for Alloca num. If we find it, mark the PHI node as being alive!
- for (DominatorTree::Node *N = DT[BB]; N; N = N->getIDom()) {
- BasicBlock *DomBB = N->getBlock();
- std::map<BasicBlock*, std::vector<PHINode*> >::iterator
- I = NewPhiNodes.find(DomBB);
- if (I != NewPhiNodes.end() && I->second[AllocaNum]) {
- // Ok, we found an inserted PHI node which dominates this value.
- PHINode *DominatingPHI = I->second[AllocaNum];
-
- // Find out if we previously thought it was dead.
- std::set<PHINode*>::iterator DPNI = DeadPHINodes.find(DominatingPHI);
- if (DPNI != DeadPHINodes.end()) {
- // Ok, until now, we thought this PHI node was dead. Mark it as being
- // alive/needed.
- DeadPHINodes.erase(DPNI);
-
- // Now that we have marked the PHI node alive, also mark any PHI nodes
- // which it might use as being alive as well.
- for (pred_iterator PI = pred_begin(DomBB), PE = pred_end(DomBB);
- PI != PE; ++PI)
- MarkDominatingPHILive(*PI, AllocaNum, DeadPHINodes);
+
+/// ComputeLiveInBlocks - Determine which blocks the value is live in. These
+/// are blocks which lead to uses. Knowing this allows us to avoid inserting
+/// PHI nodes into blocks which don't lead to uses (thus, the inserted phi nodes
+/// would be dead).
+void PromoteMem2Reg::
+ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
+ const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
+ SmallPtrSet<BasicBlock*, 32> &LiveInBlocks) {
+
+ // To determine liveness, we must iterate through the predecessors of blocks
+ // where the def is live. Blocks are added to the worklist if we need to
+ // check their predecessors. Start with all the using blocks.
+ SmallVector<BasicBlock*, 64> LiveInBlockWorklist;
+ LiveInBlockWorklist.insert(LiveInBlockWorklist.end(),
+ Info.UsingBlocks.begin(), Info.UsingBlocks.end());
+
+ // If any of the using blocks is also a definition block, check to see if the
+ // definition occurs before or after the use. If it happens before the use,
+ // the value isn't really live-in.
+ for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
+ BasicBlock *BB = LiveInBlockWorklist[i];
+ if (!DefBlocks.count(BB)) continue;
+
+ // Okay, this is a block that both uses and defines the value. If the first
+ // reference to the alloca is a def (store), then we know it isn't live-in.
+ for (BasicBlock::iterator I = BB->begin(); ; ++I) {
+ if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
+ if (SI->getOperand(1) != AI) continue;
+
+ // We found a store to the alloca before a load. The alloca is not
+ // actually live-in here.
+ LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
+ LiveInBlockWorklist.pop_back();
+ --i, --e;
+ break;
+ }
+
+ if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
+ if (LI->getOperand(0) != AI) continue;
+
+ // Okay, we found a load before a store to the alloca. It is actually
+ // live into this block.
+ break;
}
}
}
+
+ // Now that we have a set of blocks where the phi is live-in, recursively add
+ // their predecessors until we find the full region the value is live.
+ while (!LiveInBlockWorklist.empty()) {
+ BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
+
+ // The block really is live in here, insert it into the set. If already in
+ // the set, then it has already been processed.
+ if (!LiveInBlocks.insert(BB))
+ continue;
+
+ // Since the value is live into BB, it is either defined in a predecessor or
+ // live into it to. Add the preds to the worklist unless they are a
+ // defining block.
+ for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
+ BasicBlock *P = *PI;
+
+ // The value is not live into a predecessor if it defines the value.
+ if (DefBlocks.count(P))
+ continue;
+
+ // Otherwise it is, add to the worklist.
+ LiveInBlockWorklist.push_back(P);
+ }
+ }
}
-/// PromoteLocallyUsedAlloca - Many allocas are only used within a single basic
-/// block. If this is the case, avoid traversing the CFG and inserting a lot of
-/// potentially useless PHI nodes by just performing a single linear pass over
-/// the basic block using the Alloca.
-///
-void PromoteMem2Reg::PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI) {
- assert(!AI->use_empty() && "There are no uses of the alloca!");
-
- // Handle degenerate cases quickly.
- if (AI->hasOneUse()) {
- Instruction *U = cast<Instruction>(AI->use_back());
- if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
- // Must be a load of uninitialized value.
- LI->replaceAllUsesWith(Constant::getNullValue(AI->getAllocatedType()));
- } else {
- // Otherwise it must be a store which is never read.
- assert(isa<StoreInst>(U));
+/// DetermineInsertionPoint - At this point, we're committed to promoting the
+/// alloca using IDF's, and the standard SSA construction algorithm. Determine
+/// which blocks need phi nodes and see if we can optimize out some work by
+/// avoiding insertion of dead phi nodes.
+void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
+ AllocaInfo &Info) {
+
+ // Unique the set of defining blocks for efficient lookup.
+ SmallPtrSet<BasicBlock*, 32> DefBlocks;
+ DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
+
+ // Determine which blocks the value is live in. These are blocks which lead
+ // to uses.
+ SmallPtrSet<BasicBlock*, 32> LiveInBlocks;
+ ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
+
+ // Compute the locations where PhiNodes need to be inserted. Look at the
+ // dominance frontier of EACH basic-block we have a write in.
+ unsigned CurrentVersion = 0;
+ SmallPtrSet<PHINode*, 16> InsertedPHINodes;
+ std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
+ while (!Info.DefiningBlocks.empty()) {
+ BasicBlock *BB = Info.DefiningBlocks.back();
+ Info.DefiningBlocks.pop_back();
+
+ // Look up the DF for this write, add it to defining blocks.
+ DominanceFrontier::const_iterator it = DF.find(BB);
+ if (it == DF.end()) continue;
+
+ const DominanceFrontier::DomSetType &S = it->second;
+
+ // In theory we don't need the indirection through the DFBlocks vector.
+ // In practice, the order of calling QueuePhiNode would depend on the
+ // (unspecified) ordering of basic blocks in the dominance frontier,
+ // which would give PHI nodes non-determinstic subscripts. Fix this by
+ // processing blocks in order of the occurance in the function.
+ for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
+ PE = S.end(); P != PE; ++P) {
+ // If the frontier block is not in the live-in set for the alloca, don't
+ // bother processing it.
+ if (!LiveInBlocks.count(*P))
+ continue;
+
+ DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P));
+ }
+
+ // Sort by which the block ordering in the function.
+ if (DFBlocks.size() > 1)
+ std::sort(DFBlocks.begin(), DFBlocks.end());
+
+ for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
+ BasicBlock *BB = DFBlocks[i].second;
+ if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
+ Info.DefiningBlocks.push_back(BB);
}
- BB->getInstList().erase(U);
- } else {
- // Uses of the uninitialized memory location shall get zero...
- Value *CurVal = Constant::getNullValue(AI->getAllocatedType());
+ DFBlocks.clear();
+ }
+}
+
+/// RewriteSingleStoreAlloca - If there is only a single store to this value,
+/// replace any loads of it that are directly dominated by the definition with
+/// the value stored.
+void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
+ AllocaInfo &Info,
+ LargeBlockInfo &LBI) {
+ StoreInst *OnlyStore = Info.OnlyStore;
+ bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
+ BasicBlock *StoreBB = OnlyStore->getParent();
+ int StoreIndex = -1;
+
+ // Clear out UsingBlocks. We will reconstruct it here if needed.
+ Info.UsingBlocks.clear();
- for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
- Instruction *Inst = I++;
- if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
- if (LI->getOperand(0) == AI) {
- // Loads just returns the "current value"...
- LI->replaceAllUsesWith(CurVal);
- BB->getInstList().erase(LI);
- }
- } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
- if (SI->getOperand(1) == AI) {
- // Store updates the "current value"...
- CurVal = SI->getOperand(0);
- BB->getInstList().erase(SI);
+ for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) {
+ Instruction *UserInst = cast<Instruction>(*UI++);
+ if (!isa<LoadInst>(UserInst)) {
+ assert(UserInst == OnlyStore && "Should only have load/stores");
+ continue;
+ }
+ LoadInst *LI = cast<LoadInst>(UserInst);
+
+ // Okay, if we have a load from the alloca, we want to replace it with the
+ // only value stored to the alloca. We can do this if the value is
+ // dominated by the store. If not, we use the rest of the mem2reg machinery
+ // to insert the phi nodes as needed.
+ if (!StoringGlobalVal) { // Non-instructions are always dominated.
+ if (LI->getParent() == StoreBB) {
+ // If we have a use that is in the same block as the store, compare the
+ // indices of the two instructions to see which one came first. If the
+ // load came before the store, we can't handle it.
+ if (StoreIndex == -1)
+ StoreIndex = LBI.getInstructionIndex(OnlyStore);
+
+ if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
+ // Can't handle this load, bail out.
+ Info.UsingBlocks.push_back(StoreBB);
+ continue;
}
+
+ } else if (LI->getParent() != StoreBB &&
+ !dominates(StoreBB, LI->getParent())) {
+ // If the load and store are in different blocks, use BB dominance to
+ // check their relationships. If the store doesn't dom the use, bail
+ // out.
+ Info.UsingBlocks.push_back(LI->getParent());
+ continue;
}
}
+
+ // Otherwise, we *can* safely rewrite this load.
+ Value *ReplVal = OnlyStore->getOperand(0);
+ // If the replacement value is the load, this must occur in unreachable
+ // code.
+ if (ReplVal == LI)
+ ReplVal = UndefValue::get(LI->getType());
+ LI->replaceAllUsesWith(ReplVal);
+ if (AST && isa<PointerType>(LI->getType()))
+ AST->deleteValue(LI);
+ LI->eraseFromParent();
+ LBI.deleteValue(LI);
}
+}
+
+namespace {
+
+/// StoreIndexSearchPredicate - This is a helper predicate used to search by the
+/// first element of a pair.
+struct StoreIndexSearchPredicate {
+ bool operator()(const std::pair<unsigned, StoreInst*> &LHS,
+ const std::pair<unsigned, StoreInst*> &RHS) {
+ return LHS.first < RHS.first;
+ }
+};
- // After traversing the basic block, there should be no more uses of the
- // alloca, remove it now.
- assert(AI->use_empty() && "Uses of alloca from more than one BB??");
- AI->getParent()->getInstList().erase(AI);
}
-/// PromoteLocallyUsedAllocas - This method is just like
-/// PromoteLocallyUsedAlloca, except that it processes multiple alloca
-/// instructions in parallel. This is important in cases where we have large
-/// basic blocks, as we don't want to rescan the entire basic block for each
-/// alloca which is locally used in it (which might be a lot).
-void PromoteMem2Reg::
-PromoteLocallyUsedAllocas(BasicBlock *BB, const std::vector<AllocaInst*> &AIs) {
- std::map<AllocaInst*, Value*> CurValues;
- for (unsigned i = 0, e = AIs.size(); i != e; ++i)
- CurValues[AIs[i]] = 0; // Insert with null value
-
- for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
- Instruction *Inst = I++;
- if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
- // Is this a load of an alloca we are tracking?
- if (AllocaInst *AI = dyn_cast<AllocaInst>(LI->getOperand(0))) {
- std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
- if (AIt != CurValues.end()) {
- // Loads just returns the "current value"...
- if (AIt->second == 0) // Uninitialized value??
- AIt->second =Constant::getNullValue(AIt->first->getAllocatedType());
- LI->replaceAllUsesWith(AIt->second);
- BB->getInstList().erase(LI);
- }
- }
- } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
- if (AllocaInst *AI = dyn_cast<AllocaInst>(SI->getOperand(1))) {
- std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
- if (AIt != CurValues.end()) {
- // Store updates the "current value"...
- AIt->second = SI->getOperand(0);
- BB->getInstList().erase(SI);
- }
+/// PromoteSingleBlockAlloca - Many allocas are only used within a single basic
+/// block. If this is the case, avoid traversing the CFG and inserting a lot of
+/// potentially useless PHI nodes by just performing a single linear pass over
+/// the basic block using the Alloca.
+///
+/// If we cannot promote this alloca (because it is read before it is written),
+/// return true. This is necessary in cases where, due to control flow, the
+/// alloca is potentially undefined on some control flow paths. e.g. code like
+/// this is potentially correct:
+///
+/// for (...) { if (c) { A = undef; undef = B; } }
+///
+/// ... so long as A is not used before undef is set.
+///
+void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
+ LargeBlockInfo &LBI) {
+ // The trickiest case to handle is when we have large blocks. Because of this,
+ // this code is optimized assuming that large blocks happen. This does not
+ // significantly pessimize the small block case. This uses LargeBlockInfo to
+ // make it efficient to get the index of various operations in the block.
+
+ // Clear out UsingBlocks. We will reconstruct it here if needed.
+ Info.UsingBlocks.clear();
+
+ // Walk the use-def list of the alloca, getting the locations of all stores.
+ typedef SmallVector<std::pair<unsigned, StoreInst*>, 64> StoresByIndexTy;
+ StoresByIndexTy StoresByIndex;
+
+ for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
+ UI != E; ++UI)
+ if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
+ StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
+
+ // If there are no stores to the alloca, just replace any loads with undef.
+ if (StoresByIndex.empty()) {
+ for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;)
+ if (LoadInst *LI = dyn_cast<LoadInst>(*UI++)) {
+ LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
+ if (AST && isa<PointerType>(LI->getType()))
+ AST->deleteValue(LI);
+ LBI.deleteValue(LI);
+ LI->eraseFromParent();
}
+ return;
+ }
+
+ // Sort the stores by their index, making it efficient to do a lookup with a
+ // binary search.
+ std::sort(StoresByIndex.begin(), StoresByIndex.end());
+
+ // Walk all of the loads from this alloca, replacing them with the nearest
+ // store above them, if any.
+ for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
+ LoadInst *LI = dyn_cast<LoadInst>(*UI++);
+ if (!LI) continue;
+
+ unsigned LoadIdx = LBI.getInstructionIndex(LI);
+
+ // Find the nearest store that has a lower than this load.
+ StoresByIndexTy::iterator I =
+ std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
+ std::pair<unsigned, StoreInst*>(LoadIdx, 0),
+ StoreIndexSearchPredicate());
+
+ // If there is no store before this load, then we can't promote this load.
+ if (I == StoresByIndex.begin()) {
+ // Can't handle this load, bail out.
+ Info.UsingBlocks.push_back(LI->getParent());
+ continue;
}
+
+ // Otherwise, there was a store before this load, the load takes its value.
+ --I;
+ LI->replaceAllUsesWith(I->second->getOperand(0));
+ if (AST && isa<PointerType>(LI->getType()))
+ AST->deleteValue(LI);
+ LI->eraseFromParent();
+ LBI.deleteValue(LI);
}
}
-
// QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
// Alloca returns true if there wasn't already a phi-node for that variable
//
bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
unsigned &Version,
- std::set<PHINode*> &InsertedPHINodes) {
- // Look up the basic-block in question
- std::vector<PHINode*> &BBPNs = NewPhiNodes[BB];
- if (BBPNs.empty()) BBPNs.resize(Allocas.size());
+ SmallPtrSet<PHINode*, 16> &InsertedPHINodes) {
+ // Look up the basic-block in question.
+ PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
// If the BB already has a phi node added for the i'th alloca then we're done!
- if (BBPNs[AllocaNo]) return false;
+ if (PN) return false;
// Create a PhiNode using the dereferenced type... and add the phi-node to the
// BasicBlock.
- BBPNs[AllocaNo] = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
- Allocas[AllocaNo]->getName() + "." +
- utostr(Version++), BB->begin());
- InsertedPHINodes.insert(BBPNs[AllocaNo]);
+ PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(),
+ Allocas[AllocaNo]->getName() + "." + Twine(Version++),
+ BB->begin());
+ ++NumPHIInsert;
+ PhiToAllocaMap[PN] = AllocaNo;
+ PN->reserveOperandSpace(getNumPreds(BB));
+
+ InsertedPHINodes.insert(PN);
+
+ if (AST && isa<PointerType>(PN->getType()))
+ AST->copyValue(PointerAllocaValues[AllocaNo], PN);
+
return true;
}
-
// RenamePass - Recursively traverse the CFG of the function, renaming loads and
// stores to the allocas which we are promoting. IncomingVals indicates what
// value each Alloca contains on exit from the predecessor block Pred.
//
void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
- std::vector<Value*> &IncomingVals) {
-
- // If this BB needs a PHI node, update the PHI node for each variable we need
- // PHI nodes for.
- std::map<BasicBlock*, std::vector<PHINode *> >::iterator
- BBPNI = NewPhiNodes.find(BB);
- if (BBPNI != NewPhiNodes.end()) {
- std::vector<PHINode *> &BBPNs = BBPNI->second;
- for (unsigned k = 0; k != BBPNs.size(); ++k)
- if (PHINode *PN = BBPNs[k]) {
- // Add this incoming value to the PHI node.
- PN->addIncoming(IncomingVals[k], Pred);
-
+ RenamePassData::ValVector &IncomingVals,
+ std::vector<RenamePassData> &Worklist) {
+NextIteration:
+ // If we are inserting any phi nodes into this BB, they will already be in the
+ // block.
+ if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
+ // If we have PHI nodes to update, compute the number of edges from Pred to
+ // BB.
+ if (PhiToAllocaMap.count(APN)) {
+ // We want to be able to distinguish between PHI nodes being inserted by
+ // this invocation of mem2reg from those phi nodes that already existed in
+ // the IR before mem2reg was run. We determine that APN is being inserted
+ // because it is missing incoming edges. All other PHI nodes being
+ // inserted by this pass of mem2reg will have the same number of incoming
+ // operands so far. Remember this count.
+ unsigned NewPHINumOperands = APN->getNumOperands();
+
+ unsigned NumEdges = 0;
+ for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
+ if (*I == BB)
+ ++NumEdges;
+ assert(NumEdges && "Must be at least one edge from Pred to BB!");
+
+ // Add entries for all the phis.
+ BasicBlock::iterator PNI = BB->begin();
+ do {
+ unsigned AllocaNo = PhiToAllocaMap[APN];
+
+ // Add N incoming values to the PHI node.
+ for (unsigned i = 0; i != NumEdges; ++i)
+ APN->addIncoming(IncomingVals[AllocaNo], Pred);
+
// The currently active variable for this block is now the PHI.
- IncomingVals[k] = PN;
- }
+ IncomingVals[AllocaNo] = APN;
+
+ // Get the next phi node.
+ ++PNI;
+ APN = dyn_cast<PHINode>(PNI);
+ if (APN == 0) break;
+
+ // Verify that it is missing entries. If not, it is not being inserted
+ // by this mem2reg invocation so we want to ignore it.
+ } while (APN->getNumOperands() == NewPHINumOperands);
+ }
}
-
- // don't revisit nodes
- if (Visited.count(BB)) return;
- // mark as visited
- Visited.insert(BB);
+ // Don't revisit blocks.
+ if (!Visited.insert(BB)) return;
for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
Instruction *I = II++; // get the instruction, increment iterator
if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
- if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) {
- std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
- if (AI != AllocaLookup.end()) {
- Value *V = IncomingVals[AI->second];
-
- // walk the use list of this load and replace all uses with r
- LI->replaceAllUsesWith(V);
- BB->getInstList().erase(LI);
- }
- }
+ AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
+ if (!Src) continue;
+
+ std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
+ if (AI == AllocaLookup.end()) continue;
+
+ Value *V = IncomingVals[AI->second];
+
+ // Anything using the load now uses the current value.
+ LI->replaceAllUsesWith(V);
+ if (AST && isa<PointerType>(LI->getType()))
+ AST->deleteValue(LI);
+ BB->getInstList().erase(LI);
} else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
// Delete this instruction and mark the name as the current holder of the
// value
- if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) {
- std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
- if (ai != AllocaLookup.end()) {
- // what value were we writing?
- IncomingVals[ai->second] = SI->getOperand(0);
- BB->getInstList().erase(SI);
- }
- }
+ AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
+ if (!Dest) continue;
+
+ std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
+ if (ai == AllocaLookup.end())
+ continue;
+
+ // what value were we writing?
+ IncomingVals[ai->second] = SI->getOperand(0);
+ BB->getInstList().erase(SI);
}
}
- // Recurse to our successors.
- TerminatorInst *TI = BB->getTerminator();
- for (unsigned i = 0; i != TI->getNumSuccessors(); i++) {
- std::vector<Value*> OutgoingVals(IncomingVals);
- RenamePass(TI->getSuccessor(i), BB, OutgoingVals);
- }
+ // 'Recurse' to our successors.
+ succ_iterator I = succ_begin(BB), E = succ_end(BB);
+ if (I == E) return;
+
+ // Keep track of the successors so we don't visit the same successor twice
+ SmallPtrSet<BasicBlock*, 8> VisitedSuccs;
+
+ // Handle the first successor without using the worklist.
+ VisitedSuccs.insert(*I);
+ Pred = BB;
+ BB = *I;
+ ++I;
+
+ for (; I != E; ++I)
+ if (VisitedSuccs.insert(*I))
+ Worklist.push_back(RenamePassData(*I, Pred, IncomingVals));
+
+ goto NextIteration;
}
/// PromoteMemToReg - Promote the specified list of alloca instructions into
/// use of DominanceFrontier information. This function does not modify the CFG
/// of the function at all. All allocas must be from the same function.
///
+/// If AST is specified, the specified tracker is updated to reflect changes
+/// made to the IR.
+///
void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
DominatorTree &DT, DominanceFrontier &DF,
- const TargetData &TD) {
+ AliasSetTracker *AST) {
// If there is nothing to do, bail out...
if (Allocas.empty()) return;
- PromoteMem2Reg(Allocas, DT, DF, TD).run();
+
+ PromoteMem2Reg(Allocas, DT, DF, AST).run();
}