//
// 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.
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
#include "llvm/Instructions.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/LLVMContext.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/AliasSetTracker.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/StringExtras.h"
+#include "llvm/ADT/STLExtras.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/Compiler.h"
#include <algorithm>
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 DenseMapKeyInfo for all pointers.
+// 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 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;
} else {
- return false; // Not a load or store.
+ return false;
}
return true;
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 VISIBILITY_HIDDEN 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 VISIBILITY_HIDDEN PromoteMem2Reg {
/// Allocas - The alloca instructions being promoted.
///
std::vector<AllocaInst*> Allocas;
- SmallVector<AllocaInst*, 16> &RetryList;
DominatorTree &DT;
DominanceFrontier &DF;
/// AST - An AliasSetTracker object to update. If null, don't update it.
///
AliasSetTracker *AST;
+
+ LLVMContext *Context;
/// AllocaLookup - Reverse mapping of Allocas.
///
/// 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,
- SmallVector<AllocaInst*, 16> &Retry, DominatorTree &dt,
- DominanceFrontier &df, AliasSetTracker *ast)
- : Allocas(A), RetryList(Retry), DT(dt), DF(df), AST(ast) {}
+ PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
+ DominanceFrontier &df, AliasSetTracker *ast,
+ LLVMContext *C)
+ : Allocas(A), DT(dt), DF(df), AST(ast), Context(C) {}
void run();
Allocas.pop_back();
--AllocaIdx;
}
-
- void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info);
- void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
- SmallPtrSet<PHINode*, 16> &DeadPHINodes);
- bool PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI);
- void PromoteLocallyUsedAllocas(BasicBlock *BB,
- const std::vector<AllocaInst*> &AIs);
+ 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,
RenamePassData::ValVector &IncVals,
std::vector<RenamePassData> &Worklist);
/// 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){
+ U != E;) {
Instruction *User = cast<Instruction>(*U);
- if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
+ ++U;
+ 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;
+ }
+ else 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
+ // Otherwise it must be a load instruction, keep track of variable
+ // reads.
UsingBlocks.push_back(LI->getParent());
AllocaPointerVal = LI;
}
}
}
};
-
} // 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];
// analogous to finding the 'uses' and 'definitions' of each variable.
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 (Info.OnlyUsedInOneBlock) {
- LocallyUsedAllocas[Info.OnlyBlock].push_back(AI);
-
- // Remove the alloca from the Allocas list, since it will be processed.
- 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);
+ 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);
}
}
+ // 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) {
+ 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);
+ }
+
+ if (AST) AST->deleteValue(AI);
+ AI->eraseFromParent();
+ LBI.deleteValue(AI);
+
+ // The alloca has been processed, move on.
+ RemoveFromAllocasList(AllocaNum);
+
+ ++NumLocalPromoted;
+ continue;
+ }
+ }
- if (AST)
- PointerAllocaValues[AllocaNum] = Info.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 (!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);
- 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))
- Info.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 = 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 (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;
- }
- // 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*> &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 (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);
- }
+ // 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.
//
RenamePassData::ValVector Values(Allocas.size());
for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
- Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
+ Values[i] = Context->getUndef(Allocas[i]->getAllocatedType());
// Walks all basic blocks in the function performing the SSA rename algorithm
// and inserting the phi nodes we marked as necessary
// Just delete the users now.
//
if (!A->use_empty())
- A->replaceAllUsesWith(UndefValue::get(A->getType()));
+ A->replaceAllUsesWith(Context->getUndef(A->getType()));
if (AST) AST->deleteValue(A);
A->eraseFromParent();
}
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
BasicBlock::iterator BBI = BB->begin();
while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
SomePHI->getNumIncomingValues() == NumBadPreds) {
- Value *UndefVal = UndefValue::get(SomePHI->getType());
+ Value *UndefVal = Context->getUndef(SomePHI->getType());
for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
SomePHI->addIncoming(UndefVal, Preds[pred]);
}
}
+/// 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.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);
+ }
+ }
+}
+
+/// 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) {
+ 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();
- // 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;
+ 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);
- // 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();
+ // 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;
}
}
- // Finally, remove this block from the UsingBlock set.
- Info.UsingBlocks[i] = Info.UsingBlocks.back();
- Info.UsingBlocks.pop_back();
- --i; --e;
+ // Otherwise, we *can* safely rewrite this load.
+ LI->replaceAllUsesWith(OnlyStore->getOperand(0));
+ if (AST && isa<PointerType>(LI->getType()))
+ AST->deleteValue(LI);
+ LI->eraseFromParent();
+ LBI.deleteValue(LI);
}
}
-// 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!
- 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;
-
- // 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);
- }
- }
+/// 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;
}
-}
+};
-/// PromoteLocallyUsedAlloca - Many allocas are only used within a single basic
+/// 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.
///
/// ... 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.
- 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(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 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)) {
- if (SI->getOperand(1) == AI) {
- // Store updates the "current value"...
- CurVal = SI->getOperand(0);
- BB->getInstList().erase(SI);
- }
+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(Context->getUndef(LI->getType()));
+ if (AST && isa<PointerType>(LI->getType()))
+ AST->deleteValue(LI);
+ LBI.deleteValue(LI);
+ LI->eraseFromParent();
}
- }
+ return;
}
-
- // 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
-/// 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()) {
- // 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)) {
- 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);
- SI->eraseFromParent();
- }
- }
+ // 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
//
// Create a PhiNode using the dereferenced type... and add the phi-node to the
// BasicBlock.
- PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
- Allocas[AllocaNo]->getName() + "." +
- utostr(Version++), BB->begin());
+ PN = PHINode::Create(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,
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())) {
- // 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();
+ 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 (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
- if (PredTerm->getSuccessor(i) == BB)
+ 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.
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];
-
- // 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);
- }
- }
+ 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++)
- Worklist.push_back(RenamePassData(TI->getSuccessor(i), BB, IncomingVals));
+ // '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
///
void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
DominatorTree &DT, DominanceFrontier &DF,
- AliasSetTracker *AST) {
+ LLVMContext *Context, AliasSetTracker *AST) {
// If there is nothing to do, bail out...
if (Allocas.empty()) return;
- 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();
- }
+ PromoteMem2Reg(Allocas, DT, DF, AST, Context).run();
}
projects / oota-llvm.git / blobdiff