Change the rename pass to be "tail recursive", only adding N-1 successors

to the worklist, and handling the last one with a 'tail call'.  This speeds
up PR1432 from 2.0578s to 2.0012s (2.8%)


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40822 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Chris Lattner 2007-08-04 20:40:27 +00:00
parent e7b653dabe
commit 1e76af3011

View File

@ -801,6 +801,7 @@ bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred, void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
RenamePassData::ValVector &IncomingVals, RenamePassData::ValVector &IncomingVals,
std::vector<RenamePassData> &Worklist) { std::vector<RenamePassData> &Worklist) {
NextIteration:
// If we are inserting any phi nodes into this BB, they will already be in the // If we are inserting any phi nodes into this BB, they will already be in the
// block. // block.
if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) { if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
@ -865,36 +866,49 @@ void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
Instruction *I = II++; // get the instruction, increment iterator Instruction *I = II++; // get the instruction, increment iterator
if (LoadInst *LI = dyn_cast<LoadInst>(I)) { if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) { AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src); if (!Src) continue;
if (AI != AllocaLookup.end()) {
Value *V = IncomingVals[AI->second]; 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 Value *V = IncomingVals[AI->second];
LI->replaceAllUsesWith(V);
if (AST && isa<PointerType>(LI->getType())) // Anything using the load now uses the current value.
AST->deleteValue(LI); LI->replaceAllUsesWith(V);
BB->getInstList().erase(LI); if (AST && isa<PointerType>(LI->getType()))
} AST->deleteValue(LI);
} BB->getInstList().erase(LI);
} else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
// Delete this instruction and mark the name as the current holder of the // Delete this instruction and mark the name as the current holder of the
// value // value
if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) { AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest); if (!Dest) continue;
if (ai != AllocaLookup.end()) {
// what value were we writing? std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
IncomingVals[ai->second] = SI->getOperand(0); if (ai == AllocaLookup.end())
BB->getInstList().erase(SI); 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(); TerminatorInst *TI = BB->getTerminator();
for (unsigned i = 0; i != TI->getNumSuccessors(); i++) 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)); 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 /// PromoteMemToReg - Promote the specified list of alloca instructions into