replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Second attempt at fixing the performance regressions introduced 

by my recent GVN improvement.  Looking through a single layer of
PHI nodes when attempting to sink GEPs, we need to iteratively
look through arbitrary PHI nests.

llvm-svn: 120202
This commit is contained in:
Owen Anderson 2010-11-27 08:15:55 +00:00
parent 762bf86919
commit 8ba5f39f70
2 changed files with 62 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
llvm/include/llvm/Transforms/Utils/AddrModeMatcher.h
View File

@ -39,6 +39,12 @@ struct ExtAddrMode : public TargetLowering::AddrMode {
ExtAddrMode() : BaseReg(0), ScaledReg(0) {}
void print(raw_ostream &OS) const;
void dump() const;
bool operator==(const ExtAddrMode& O) const {
return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
(BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
(HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
}
};
static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {

67
llvm/lib/Transforms/Scalar/CodeGenPrepare.cpp
View File

@ -618,36 +618,67 @@ static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
const Type *AccessTy,
DenseMap<Value*,Value*> &SunkAddrs) {
Value *Repl = Addr;
// Try to collapse single-value PHI nodes. This is necessary to undo
// unprofitable PRE transformations.
Value *Repl = Addr;
if (isa<PHINode>(Addr) && MemoryInst->hasOneUse()) {
PHINode *P = cast<PHINode>(Addr);
Instruction *Consensus = 0;
std::vector<Value*> worklist;
SmallPtrSet<Value*, 4> Visited;
worklist.push_back(Addr);
// Use a worklist to iteratively look through PHI nodes, and ensure that
// the addressing mode obtained from the non-PHI roots of the graph
// are equivalent.
Value *Consensus = 0;
unsigned NumUses = 0;
for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
Instruction *Incoming = dyn_cast<Instruction>(P->getIncomingValue(i));
if (!Incoming || (Consensus && !Incoming->isIdenticalTo(Consensus))) {
SmallVector<Instruction*, 16> AddrModeInsts;
ExtAddrMode AddrMode;
while (!worklist.empty()) {
Value *V = worklist.back();
worklist.pop_back();
// Break use-def graph loops.
if (Visited.count(V)) {
Consensus = 0;
break;
}
if (!Consensus || Incoming->isIdenticalTo(Consensus)) {
if (Incoming->getNumUses() > NumUses) {
Consensus = Incoming;
NumUses = Incoming->getNumUses();
Visited.insert(V);
// For a PHI node, push all of its incoming values.
if (PHINode *P = dyn_cast<PHINode>(V)) {
for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
worklist.push_back(P->getIncomingValue(i));
continue;
}
// For non-PHIs, determine the addressing mode being computed.
SmallVector<Instruction*, 16> NewAddrModeInsts;
ExtAddrMode NewAddrMode =
AddressingModeMatcher::Match(V, AccessTy,MemoryInst,
NewAddrModeInsts, *TLI);
// Ensure that the obtained addressing mode is equivalent to that obtained
// for all other roots of the PHI traversal. Also, when choosing one
// such root as representative, select the one with the most uses in order
// to keep the cost modeling heuristics in AddressingModeMatcher applicable.
if (!Consensus || NewAddrMode == AddrMode) {
if (V->getNumUses() > NumUses) {
Consensus = V;
NumUses = V->getNumUses();
AddrMode = NewAddrMode;
AddrModeInsts = NewAddrModeInsts;
}
continue;
}
Consensus = 0;
break;
}
if (Consensus) Addr = Consensus;
}
// Figure out what addressing mode will be built up for this operation.
SmallVector<Instruction*, 16> AddrModeInsts;
ExtAddrMode AddrMode = AddressingModeMatcher::Match(Addr, AccessTy,MemoryInst,
AddrModeInsts, *TLI);
// If the addressing mode couldn't be determined, or if multiple different
// ones were determined, bail out now.
if (!Consensus) return false;
// Check to see if any of the instructions supersumed by this addr mode are
// non-local to I's BB.