//
// 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
+// 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.
//
//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "mem2reg"
#include "llvm/Transforms/Utils/PromoteMemToReg.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.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;
-// Provide DenseMapKeyInfo for all pointers.
+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");
+
+// Provide DenseMapInfo 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);
+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 std::pair<BasicBlock*, unsigned> getTombstoneKey() {
- return std::make_pair((BasicBlock*)-2, 0U);
+ static inline EltTy getTombstoneKey() {
+ return EltTy(reinterpret_cast<BasicBlock*>(-2), 0U);
}
static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
- return DenseMapKeyInfo<void*>::getHashValue(Val.first) + Val.second*2;
+ return DenseMapInfo<void*>::getHashValue(Val.first) + Val.second*2;
+ }
+ static bool isEqual(const EltTy &LHS, const EltTy &RHS) {
+ return LHS == RHS;
}
static bool isPod() { return true; }
};
// 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.
+ if (SI->isVolatile())
+ return false;
} else {
return false; // Not a load or store.
}
}
namespace {
+ struct AllocaInfo;
// Data package used by RenamePass()
class VISIBILITY_HIDDEN RenamePassData {
public:
+ typedef std::vector<Value *> ValVector;
+
+ RenamePassData() {}
RenamePassData(BasicBlock *B, BasicBlock *P,
- const std::vector<Value *> &V) : BB(B), Pred(P), Values(V) {}
+ const ValVector &V) : BB(B), Pred(P), Values(V) {}
BasicBlock *BB;
BasicBlock *Pred;
- std::vector<Value *> Values;
+ ValVector Values;
+
+ void swap(RenamePassData &RHS) {
+ std::swap(BB, RHS.BB);
+ std::swap(Pred, RHS.Pred);
+ Values.swap(RHS.Values);
+ }
};
struct VISIBILITY_HIDDEN PromoteMem2Reg {
///
std::vector<AllocaInst*> Allocas;
SmallVector<AllocaInst*, 16> &RetryList;
- ETForest &ET;
+ DominatorTree &DT;
DominanceFrontier &DF;
- const TargetData &TD;
/// AST - An AliasSetTracker object to update. If null, don't update it.
///
/// non-determinstic behavior.
DenseMap<BasicBlock*, unsigned> BBNumbers;
- /// RenamePassWorkList - Worklist used by RenamePass()
- std::vector<RenamePassData> RenamePassWorkList;
-
+ /// BBNumPreds - Lazily compute the number of predecessors a block has.
+ DenseMap<const BasicBlock*, unsigned> BBNumPreds;
public:
PromoteMem2Reg(const std::vector<AllocaInst*> &A,
- SmallVector<AllocaInst*, 16> &Retry, ETForest &et,
- DominanceFrontier &df, const TargetData &td,
- AliasSetTracker *ast)
- : Allocas(A), RetryList(Retry), ET(et), DF(df), TD(td), AST(ast) {}
+ SmallVector<AllocaInst*, 16> &Retry, DominatorTree &dt,
+ DominanceFrontier &df, AliasSetTracker *ast)
+ : Allocas(A), RetryList(Retry), DT(dt), DF(df), AST(ast) {}
void run();
bool properlyDominates(Instruction *I1, Instruction *I2) const {
if (InvokeInst *II = dyn_cast<InvokeInst>(I1))
I1 = II->getNormalDest()->begin();
- return ET.properlyDominates(I1->getParent(), I2->getParent());
+ return DT.properlyDominates(I1->getParent(), I2->getParent());
}
- /// dominates - Return true if BB1 dominates BB2 using the ETForest.
+ /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
///
bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
- return ET.dominates(BB1, BB2);
+ return DT.dominates(BB1, BB2);
}
private:
- void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
- SmallPtrSet<PHINode*, 16> &DeadPHINodes);
+ 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);
+
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,
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();
if (AST) PointerAllocaValues.resize(Allocas.size());
+ AllocaInfo Info;
+
for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
AllocaInst *AI = Allocas[AllocaNum];
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;
-
- StoreInst *OnlyStore = 0;
- 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.
- Value *AllocaPointerVal = 0;
- 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;
- }
- }
-
- // 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;
- }
+ 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 (DefiningBlocks.size() == 1) {
- // Be aware of loads before the store.
- std::set<BasicBlock*> ProcessedBlocks;
- for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i)
- // If the store dominates the block and if we haven't processed it yet,
- // do so now.
- if (dominates(OnlyStore->getParent(), UsingBlocks[i]))
- if (ProcessedBlocks.insert(UsingBlocks[i]).second) {
- BasicBlock *UseBlock = 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 == OnlyStore->getParent()) {
- BasicBlock::iterator I = UseBlock->begin();
- for (; &*I != OnlyStore; ++I) { // scan block for store.
- if (isa<LoadInst>(I) && I->getOperand(0) == AI)
- break;
- }
- if (&*I != 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(OnlyStore->getOperand(0));
- if (AST && isa<PointerType>(LI->getType()))
- AST->deleteValue(LI);
- LI->eraseFromParent();
- }
- }
- }
-
- // Finally, remove this block from the UsingBlock set.
- UsingBlocks[i] = UsingBlocks.back();
- --i; --e;
- }
+ if (Info.DefiningBlocks.size() == 1) {
+ RewriteSingleStoreAlloca(AI, Info);
// Finally, after the scan, check to see if the store is all that is left.
- if (UsingBlocks.empty()) {
+ if (Info.UsingBlocks.empty()) {
+ // Remove the (now dead) store and alloca.
+ Info.OnlyStore->eraseFromParent();
+ if (AST) AST->deleteValue(AI);
+ AI->eraseFromParent();
+
// The alloca has been processed, move on.
- Allocas[AllocaNum] = Allocas.back();
- Allocas.pop_back();
- --AllocaNum;
+ 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 (Info.OnlyUsedInOneBlock) {
+ LocallyUsedAllocas[Info.OnlyBlock].push_back(AI);
+
+ // Remove the alloca from the Allocas list, since it will be processed.
+ RemoveFromAllocasList(AllocaNum);
+ continue;
+ }
- if (AST)
- PointerAllocaValues[AllocaNum] = AllocaPointerVal;
-
// If we haven't computed a numbering for the BB's in the function, do so
// now.
if (BBNumbers.empty()) {
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;
- SmallPtrSet<PHINode*, 16> InsertedPHINodes;
- std::vector<std::pair<unsigned, BasicBlock*> > 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::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 = DFBlocks[i].second;
- if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
- DefiningBlocks.push_back(BB);
- }
- DFBlocks.clear();
- }
- }
-
- // 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 (SmallPtrSet<PHINode*, 16>::iterator I = InsertedPHINodes.begin(),
- E = InsertedPHINodes.end(); I != E; ++I) {
- PHINode *PN = *I;
- 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.
+ // 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;
+
+ // 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);
}
// Process all allocas which are only used in a single basic block.
// 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] = 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
//
- RenamePassWorkList.clear();
+ std::vector<RenamePassData> RenamePassWorkList;
RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
- while(!RenamePassWorkList.empty()) {
- RenamePassData RPD = RenamePassWorkList.back();
+ 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);
+ RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
}
// The renamer uses the Visited set to avoid infinite loops. Clear it now.
if (&BB->front() != SomePHI)
continue;
- // 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 of them.
- if (SomePHI->getNumIncomingValues() == Preds.size())
+ if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
continue;
+
+ // 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
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,
- 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 (BasicBlock* DomBB = BB; DomBB; DomBB = ET.getIDom(DomBB)) {
- 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;
-
- // 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);
- 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;
+ } else 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.back();
+ LiveInBlockWorklist.pop_back();
+
+ // 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);
+ }
+ }
}
+/// 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);
+ }
+ 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) {
+ StoreInst *OnlyStore = Info.OnlyStore;
+ bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
+
+ // Be aware of loads before the store.
+ SmallPtrSet<BasicBlock*, 32> ProcessedBlocks;
+ for (unsigned i = 0, e = Info.UsingBlocks.size(); i != e; ++i) {
+ BasicBlock *UseBlock = Info.UsingBlocks[i];
+
+ // If we already processed this block, don't reprocess it.
+ if (!ProcessedBlocks.insert(UseBlock)) {
+ Info.UsingBlocks[i] = Info.UsingBlocks.back();
+ Info.UsingBlocks.pop_back();
+ --i; --e;
+ continue;
+ }
+
+ // If the store dominates the block and if we haven't processed it yet,
+ // do so now. We can't handle the case where the store doesn't dominate a
+ // block because there may be a path between the store and the use, but we
+ // may need to insert phi nodes to handle dominance properly.
+ if (!StoringGlobalVal && !dominates(OnlyStore->getParent(), UseBlock))
+ continue;
+
+ // 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 == OnlyStore->getParent()) {
+ BasicBlock::iterator I = UseBlock->begin();
+ for (; &*I != OnlyStore; ++I) { // scan block for store.
+ if (isa<LoadInst>(I) && I->getOperand(0) == AI)
+ break;
+ }
+ if (&*I != OnlyStore)
+ continue; // Do not promote the uses of this in this block.
+ }
+
+ // 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(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();
+ Info.UsingBlocks.pop_back();
+ --i; --e;
+ }
+}
+
+
/// 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
}
// After traversing the basic block, there should be no more uses of the
- // alloca, remove it now.
+ // 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);
+ AI->eraseFromParent();
+
+ ++NumLocalPromoted;
return false;
}
/// 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;
+ DenseMap<AllocaInst*, Value*> CurValues;
for (unsigned i = 0, e = AIs.size(); i != e; ++i)
CurValues[AIs[i]] = 0; // Insert with null value
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);
+ DenseMap<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
if (AIt != CurValues.end()) {
// If loading an uninitialized value, allow the inter-block case to
// handle it. Due to control flow, this might actually be ok.
}
} 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);
+ DenseMap<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);
+ SI->eraseFromParent();
}
}
}
}
+
+ // At the end of the block scan, all allocas in CurValues are dead.
+ for (DenseMap<AllocaInst*, Value*>::iterator I = CurValues.begin(),
+ E = CurValues.end(); I != E; ++I) {
+ AllocaInst *AI = I->first;
+ assert(AI->use_empty() && "Uses of alloca from more than one BB??");
+ if (AST) AST->deleteValue(AI);
+ AI->eraseFromParent();
+ }
+
+ NumLocalPromoted += CurValues.size();
}
PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
Allocas[AllocaNo]->getName() + "." +
utostr(Version++), BB->begin());
+ ++NumPHIInsert;
PhiToAllocaMap[PN] = AllocaNo;
+ PN->reserveOperandSpace(getNumPreds(BB));
InsertedPHINodes.insert(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) {
+ 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 (!HasPredEntries) {
+ // 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();
+
TerminatorInst *PredTerm = Pred->getTerminator();
unsigned NumEdges = 0;
for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
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);
+ // 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);
}
}
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];
+ AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
+ if (!Src) continue;
+
+ std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
+ if (AI == AllocaLookup.end()) continue;
- // 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);
- }
- }
+ 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.
+ // 'Recurse' to our successors.
TerminatorInst *TI = BB->getTerminator();
- for (unsigned i = 0; i != TI->getNumSuccessors(); i++)
- RenamePassWorkList.push_back(RenamePassData(TI->getSuccessor(i), BB, IncomingVals));
+ unsigned NumSuccs = TI->getNumSuccessors();
+ if (NumSuccs == 0) return;
+
+ // Add all-but-one successor to the worklist.
+ for (unsigned i = 0; i != NumSuccs-1; i++)
+ Worklist.push_back(RenamePassData(TI->getSuccessor(i), BB, IncomingVals));
+
+ // Handle the last successor without using the worklist. This allows us to
+ // handle unconditional branches directly, for example.
+ Pred = BB;
+ BB = TI->getSuccessor(NumSuccs-1);
+ goto NextIteration;
}
/// PromoteMemToReg - Promote the specified list of alloca instructions into
/// made to the IR.
///
void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
- ETForest &ET, DominanceFrontier &DF,
- const TargetData &TD, AliasSetTracker *AST) {
+ DominatorTree &DT, DominanceFrontier &DF,
+ AliasSetTracker *AST) {
// If there is nothing to do, bail out...
if (Allocas.empty()) return;
SmallVector<AllocaInst*, 16> RetryList;
- PromoteMem2Reg(Allocas, RetryList, ET, DF, TD, AST).run();
+ 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.
NewAllocas.assign(RetryList.begin(), RetryList.end());
RetryList.clear();
- PromoteMem2Reg(NewAllocas, RetryList, ET, DF, TD, AST).run();
+ PromoteMem2Reg(NewAllocas, RetryList, DT, DF, AST).run();
NewAllocas.clear();
}
}