//===- 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 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.
+// 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/Support/CFG.h"
-#include "llvm/Support/StableBasicBlockNumbering.h"
+#include "llvm/Instructions.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/StringExtras.h"
+#include "llvm/Support/CFG.h"
+#include "llvm/Support/Compiler.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");
+
+// Provide DenseMapKeyInfo for all pointers.
+namespace llvm {
+template<>
+struct DenseMapKeyInfo<std::pair<BasicBlock*, unsigned> > {
+ static inline std::pair<BasicBlock*, unsigned> getEmptyKey() {
+ return std::make_pair((BasicBlock*)-1, ~0U);
+ }
+ static inline std::pair<BasicBlock*, unsigned> getTombstoneKey() {
+ return std::make_pair((BasicBlock*)-2, 0U);
+ }
+ static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
+ return DenseMapKeyInfo<void*>::getHashValue(Val.first) + Val.second*2;
+ }
+ static bool isPod() { return true; }
+};
+}
+
/// 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.
} 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)))
- 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; // Not a load or store.
}
-
+
return true;
}
namespace {
- struct PromoteMem2Reg {
- // Allocas - The alloca instructions being promoted
+ struct AllocaInfo;
+
+ // Data package used by RenamePass()
+ class VISIBILITY_HIDDEN RenamePassData {
+ public:
+ typedef std::vector<Value *> ValVector;
+
+ RenamePassData() {}
+ 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);
+ }
+ };
+
+ struct VISIBILITY_HIDDEN PromoteMem2Reg {
+ /// Allocas - The alloca instructions being promoted.
+ ///
std::vector<AllocaInst*> Allocas;
+ SmallVector<AllocaInst*, 16> &RetryList;
DominatorTree &DT;
DominanceFrontier &DF;
- const TargetData &TD;
-
- // AllocaLookup - Reverse mapping of Allocas
- std::map<AllocaInst*, unsigned> AllocaLookup;
- // NewPhiNodes - The PhiNodes we're adding.
- std::map<BasicBlock*, std::vector<PHINode*> > NewPhiNodes;
+ /// AST - An AliasSetTracker object to update. If null, don't update it.
+ ///
+ AliasSetTracker *AST;
- // Visited - The set of basic blocks the renamer has already visited.
- std::set<BasicBlock*> Visited;
+ /// AllocaLookup - Reverse mapping of Allocas.
+ ///
+ std::map<AllocaInst*, unsigned> AllocaLookup;
- // 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;
public:
- PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
- DominanceFrontier &df, const TargetData &td)
- : Allocas(A), DT(dt), DF(df), TD(td) {}
+ PromoteMem2Reg(const std::vector<AllocaInst*> &A,
+ SmallVector<AllocaInst*, 16> &Retry, DominatorTree &dt,
+ DominanceFrontier &df, AliasSetTracker *ast)
+ : Allocas(A), RetryList(Retry), 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 RemoveFromAllocasList(unsigned &AllocaIdx) {
+ Allocas[AllocaIdx] = Allocas.back();
+ Allocas.pop_back();
+ --AllocaIdx;
+ }
+
+ void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info);
+
void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
- std::set<PHINode*> &DeadPHINodes);
- void PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI);
- void PromoteLocallyUsedAllocas(BasicBlock *BB,
+ SmallPtrSet<PHINode*, 16> &DeadPHINodes);
+ bool PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI);
+ void PromoteLocallyUsedAllocas(BasicBlock *BB,
const std::vector<AllocaInst*> &AIs);
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 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());
+ 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();
// 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;
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 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);
+ if (Info.OnlyUsedInOneBlock) {
+ LocallyUsedAllocas[Info.OnlyBlock].push_back(AI);
// Remove the alloca from the Allocas list, since it will be processed.
- Allocas[AllocaNum] = Allocas.back();
- Allocas.pop_back();
- --AllocaNum;
+ RemoveFromAllocasList(AllocaNum);
continue;
}
+ // 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);
+
+ // Finally, after the scan, check to see if the store is all that is left.
+ if (Info.UsingBlocks.empty()) {
+ ++NumSingleStore;
+ // The alloca has been processed, move on.
+ RemoveFromAllocasList(AllocaNum);
+ continue;
+ }
+ }
+
+
+ if (AST)
+ PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
+
// If we haven't computed a numbering for the BB's in the function, do so
// now.
- BBNumbers.compute(F);
+ if (BBNumbers.empty()) {
+ unsigned ID = 0;
+ for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
+ BBNumbers[I] = ID++;
+ }
// 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();
+ 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 PhiNodes
DominanceFrontier::const_iterator it = DF.find(BB);
// (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));
+ for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
+ PE = S.end(); P != PE; ++P)
+ DFBlocks.push_back(std::make_pair(BBNumbers[*P], *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]);
+ BasicBlock *BB = DFBlocks[i].second;
if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
- DefiningBlocks.push_back(BB);
+ Info.DefiningBlocks.push_back(BB);
}
DFBlocks.clear();
}
// 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();
+ for (unsigned i = 0, e = Info.UsingBlocks.size(); i != e; ++i)
+ MarkDominatingPHILive(Info.UsingBlocks[i], AllocaNum, InsertedPHINodes);
+ Info.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(),
+ for (SmallPtrSet<PHINode*, 16>::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);
+ bool Erased=NewPhiNodes.erase(std::make_pair(PN->getParent(), AllocaNum));
+ Erased=Erased;
+ assert(Erased && "PHI already removed?");
+
+ if (AST && isa<PointerType>(PN->getType()))
+ AST->deleteValue(PN);
+ PN->eraseFromParent();
+ PhiToAllocaMap.erase(PN);
}
- // Keep the reverse mapping of the 'Allocas' array.
+ // Keep the reverse mapping of the 'Allocas' array.
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??");
+ const std::vector<AllocaInst*> &LocAllocas = I->second;
+ assert(!LocAllocas.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);
+ if (LocAllocas.size() == 1) {
+ // If we can do the quick promotion pass, do so now.
+ if (PromoteLocallyUsedAlloca(I->first, LocAllocas[0]))
+ RetryList.push_back(LocAllocas[0]); // Failed, retry later.
+ } else {
+ // Locally promote anything possible. Note that if this is unable to
+ // promote a particular alloca, it puts the alloca onto the Allocas vector
+ // for global processing.
+ PromoteLocallyUsedAllocas(I->first, LocAllocas);
+ }
}
if (Allocas.empty())
// 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));
+ while (!RenamePassWorkList.empty()) {
+ RenamePassData RPD;
+ RPD.swap(RenamePassWorkList.back());
+ RenamePassWorkList.pop_back();
+ // RenamePass may add new worklist entries.
+ RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
+ }
+
// 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(true)) {
+ if (!isa<Instruction>(V) ||
+ properlyDominates(cast<Instruction>(V), PN)) {
+ 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) {
+ // 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;
- 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??");
+ // Count the number of preds for BB.
+ SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
// 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() == Preds.size())
+ continue;
+
+ // 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();
+}
+
- // 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]);
+/// 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) {
+ // Be aware of loads before the store.
+ std::set<BasicBlock*> ProcessedBlocks;
+ for (unsigned i = 0, e = Info.UsingBlocks.size(); i != e; ++i) {
+ // If the store dominates the block and if we haven't processed it yet,
+ // do so now.
+ if (!dominates(Info.OnlyStore->getParent(), Info.UsingBlocks[i]))
+ continue;
+
+ if (!ProcessedBlocks.insert(Info.UsingBlocks[i]).second)
+ continue;
+
+ BasicBlock *UseBlock = Info.UsingBlocks[i];
+
+ // If the use and store are in the same block, do a quick scan to
+ // verify that there are no uses before the store.
+ if (UseBlock == Info.OnlyStore->getParent()) {
+ BasicBlock::iterator I = UseBlock->begin();
+ for (; &*I != Info.OnlyStore; ++I) { // scan block for store.
+ if (isa<LoadInst>(I) && I->getOperand(0) == AI)
+ break;
+ }
+ if (&*I != Info.OnlyStore) break; // Do not handle this case.
+ }
+
+ // Otherwise, if this is a different block or if all uses happen
+ // after the store, do a simple linear scan to replace loads with
+ // the stored value.
+ for (BasicBlock::iterator I = UseBlock->begin(),E = UseBlock->end();
+ I != E; ) {
+ if (LoadInst *LI = dyn_cast<LoadInst>(I++)) {
+ if (LI->getOperand(0) == AI) {
+ LI->replaceAllUsesWith(Info.OnlyStore->getOperand(0));
+ if (AST && isa<PointerType>(LI->getType()))
+ AST->deleteValue(LI);
+ LI->eraseFromParent();
}
+ }
}
+
+ // Finally, remove this block from the UsingBlock set.
+ Info.UsingBlocks[i] = Info.UsingBlocks.back();
+ --i; --e;
}
}
+
// 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
// DeadPHINodes set are removed.
//
void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
- std::set<PHINode*> &DeadPHINodes) {
+ SmallPtrSet<PHINode*, 16> &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]) {
+ DomTreeNode *IDomNode = DT.getNode(BB);
+ for (DomTreeNode *IDom = IDomNode; IDom; IDom = IDom->getIDom()) {
+ BasicBlock *DomBB = IDom->getBlock();
+ DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator
+ I = NewPhiNodes.find(std::make_pair(DomBB, AllocaNum));
+ if (I != NewPhiNodes.end()) {
// 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);
+ PHINode *DominatingPHI = I->second;
+ // Find out if we previously thought it was dead. If so, mark it as being
+ // live by removing it from the DeadPHINodes set.
+ if (DeadPHINodes.erase(DominatingPHI)) {
// 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);
/// 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) {
+/// 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.
+///
+bool PromoteMem2Reg::PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI) {
assert(!AI->use_empty() && "There are no uses of the alloca!");
// Handle degenerate cases quickly.
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()));
+ LI->replaceAllUsesWith(UndefValue::get(AI->getAllocatedType()));
+ if (AST && isa<PointerType>(LI->getType()))
+ AST->deleteValue(LI);
} else {
// Otherwise it must be a store which is never read.
assert(isa<StoreInst>(U));
}
BB->getInstList().erase(U);
} else {
- // Uses of the uninitialized memory location shall get zero...
- Value *CurVal = Constant::getNullValue(AI->getAllocatedType());
-
+ // Uses of the uninitialized memory location shall get undef.
+ Value *CurVal = 0;
+
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) {
+ if (!CurVal) return true; // Could not locally promote!
+
// Loads just returns the "current value"...
LI->replaceAllUsesWith(CurVal);
+ if (AST && isa<PointerType>(LI->getType()))
+ AST->deleteValue(LI);
BB->getInstList().erase(LI);
}
} else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
// 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??");
+ if (AST) AST->deleteValue(AI);
AI->getParent()->getInstList().erase(AI);
+
+ ++NumLocalPromoted;
+ return false;
}
/// PromoteLocallyUsedAllocas - This method is just like
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);
+ // If loading an uninitialized value, allow the inter-block case to
+ // handle it. Due to control flow, this might actually be ok.
+ if (AIt->second == 0) { // Use of locally uninitialized value??
+ RetryList.push_back(AI); // Retry elsewhere.
+ CurValues.erase(AIt); // Stop tracking this here.
+ if (CurValues.empty()) return;
+ } else {
+ // Loads just returns the "current value"...
+ LI->replaceAllUsesWith(AIt->second);
+ if (AST && isa<PointerType>(LI->getType()))
+ AST->deleteValue(LI);
+ BB->getInstList().erase(LI);
+ }
}
}
} else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
//
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 = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
+ Allocas[AllocaNo]->getName() + "." +
+ utostr(Version++), BB->begin());
+ PhiToAllocaMap[PN] = AllocaNo;
+
+ InsertedPHINodes.insert(PN);
+
+ if (AST && isa<PointerType>(PN->getType()))
+ AST->copyValue(PointerAllocaValues[AllocaNo], PN);
+
return true;
}
// 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);
-
- // The currently active variable for this block is now the PHI.
- IncomingVals[k] = PN;
+ RenamePassData::ValVector &IncomingVals,
+ std::vector<RenamePassData> &Worklist) {
+ // 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())) {
+ // Pred may have multiple edges to BB. If so, we want to add N incoming
+ // values to each PHI we are inserting on the first time we see the edge.
+ // Check to see if APN already has incoming values from Pred. This also
+ // prevents us from modifying PHI nodes that are not currently being
+ // inserted.
+ bool HasPredEntries = false;
+ for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) {
+ if (APN->getIncomingBlock(i) == Pred) {
+ HasPredEntries = true;
+ break;
+ }
+ }
+
+ // If we have PHI nodes to update, compute the number of edges from Pred to
+ // BB.
+ if (!HasPredEntries) {
+ TerminatorInst *PredTerm = Pred->getTerminator();
+ unsigned NumEdges = 0;
+ for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
+ if (PredTerm->getSuccessor(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[AllocaNo] = APN;
+
+ // Get the next phi node.
+ ++PNI;
+ APN = dyn_cast<PHINode>(PNI);
+ if (APN == 0) break;
+
+ // Verify it doesn't already have entries for Pred. If it does, it is
+ // not being inserted by this mem2reg invocation.
+ HasPredEntries = false;
+ for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) {
+ if (APN->getIncomingBlock(i) == Pred) {
+ HasPredEntries = true;
+ break;
+ }
+ }
+ } while (!HasPredEntries);
+ }
}
-
- // 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
// walk the use list of this load and replace all uses with r
LI->replaceAllUsesWith(V);
+ if (AST && isa<PointerType>(LI->getType()))
+ AST->deleteValue(LI);
BB->getInstList().erase(LI);
}
}
// 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);
- }
+ for (unsigned i = 0; i != TI->getNumSuccessors(); i++)
+ Worklist.push_back(RenamePassData(TI->getSuccessor(i), BB, IncomingVals));
}
/// 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();
+
+ SmallVector<AllocaInst*, 16> RetryList;
+ PromoteMem2Reg(Allocas, RetryList, DT, DF, AST).run();
+
+ // PromoteMem2Reg may not have been able to promote all of the allocas in one
+ // pass, run it again if needed.
+ std::vector<AllocaInst*> NewAllocas;
+ while (!RetryList.empty()) {
+ // If we need to retry some allocas, this is due to there being no store
+ // before a read in a local block. To counteract this, insert a store of
+ // undef into the alloca right after the alloca itself.
+ for (unsigned i = 0, e = RetryList.size(); i != e; ++i) {
+ BasicBlock::iterator BBI = RetryList[i];
+
+ new StoreInst(UndefValue::get(RetryList[i]->getAllocatedType()),
+ RetryList[i], ++BBI);
+ }
+
+ NewAllocas.assign(RetryList.begin(), RetryList.end());
+ RetryList.clear();
+ PromoteMem2Reg(NewAllocas, RetryList, DT, DF, AST).run();
+ NewAllocas.clear();
+ }
}
projects / oota-llvm.git / blobdiff