static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
cl::init(cl::BOU_UNSET), cl::Hidden);
namespace {
+ // Throttle for huge numbers of predecessors (compile speed problems)
+ cl::opt<unsigned>
+ TailMergeThreshold("tail-merge-threshold",
+ cl::desc("Max number of predecessors to consider tail merging"),
+ cl::init(100), cl::Hidden);
+
struct BranchFolder : public MachineFunctionPass {
static char ID;
BranchFolder(bool defaultEnableTailMerge) :
char BranchFolder::ID = 0;
}
-static bool CorrectExtraCFGEdges(MachineBasicBlock &MBB,
- MachineBasicBlock *DestA,
- MachineBasicBlock *DestB,
- bool isCond,
- MachineFunction::iterator FallThru);
-
FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) {
return new BranchFolder(DefaultEnableTailMerge); }
MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
std::vector<MachineOperand> Cond;
if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond))
- EverMadeChange |= CorrectExtraCFGEdges(*MBB, TBB, FBB,
- !Cond.empty(), next(I));
+ EverMadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
}
RegInfo = MF.getTarget().getRegisterInfo();
return true;
else if (p.second->getNumber() > q.second->getNumber())
return false;
- else
+ else {
+ // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
+ // an object with itself.
+#ifndef _GLIBCXX_DEBUG
assert(0 && "Predecessor appears twice");
+#endif
+ return(false);
+ }
}
// See if any of the blocks in MergePotentials (which all have a common single
MergePotentials.push_back(std::make_pair(HashEndOfMBB(I, 2U), I));
}
// See if we can do any tail merging on those.
- MadeChange |= TryMergeBlocks(NULL, NULL);
+ if (MergePotentials.size() < TailMergeThreshold)
+ MadeChange |= TryMergeBlocks(NULL, NULL);
// Look at blocks (IBB) with multiple predecessors (PBB).
// We change each predecessor to a canonical form, by
// transformations.)
for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
- if (!I->succ_empty() && I->pred_size() >= 2) {
+ if (!I->succ_empty() && I->pred_size() >= 2 &&
+ I->pred_size() < TailMergeThreshold) {
MachineBasicBlock *IBB = I;
MachineBasicBlock *PredBB = prior(I);
MergePotentials.clear();
- for (MachineBasicBlock::pred_iterator P = I->pred_begin(), E2 = I->pred_end();
+ for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
+ E2 = I->pred_end();
P != E2; ++P) {
MachineBasicBlock* PBB = *P;
// Skip blocks that loop to themselves, can't tail merge these.
}
-/// CorrectExtraCFGEdges - Various pieces of code can cause excess edges in the
-/// CFG to be inserted. If we have proven that MBB can only branch to DestA and
-/// DestB, remove any other MBB successors from the CFG. DestA and DestB can
-/// be null.
-/// Besides DestA and DestB, retain other edges leading to LandingPads (currently
-/// there can be only one; we don't check or require that here).
-/// Note it is possible that DestA and/or DestB are LandingPads.
-static bool CorrectExtraCFGEdges(MachineBasicBlock &MBB,
- MachineBasicBlock *DestA,
- MachineBasicBlock *DestB,
- bool isCond,
- MachineFunction::iterator FallThru) {
- bool MadeChange = false;
- bool AddedFallThrough = false;
-
- // If this block ends with a conditional branch that falls through to its
- // successor, set DestB as the successor.
- if (isCond) {
- if (DestB == 0 && FallThru != MBB.getParent()->end()) {
- DestB = FallThru;
- AddedFallThrough = true;
- }
- } else {
- // If this is an unconditional branch with no explicit dest, it must just be
- // a fallthrough into DestB.
- if (DestA == 0 && FallThru != MBB.getParent()->end()) {
- DestA = FallThru;
- AddedFallThrough = true;
- }
- }
-
- MachineBasicBlock::succ_iterator SI = MBB.succ_begin();
- MachineBasicBlock *OrigDestA = DestA, *OrigDestB = DestB;
- while (SI != MBB.succ_end()) {
- if (*SI == DestA && DestA == DestB) {
- DestA = DestB = 0;
- ++SI;
- } else if (*SI == DestA) {
- DestA = 0;
- ++SI;
- } else if (*SI == DestB) {
- DestB = 0;
- ++SI;
- } else if ((*SI)->isLandingPad() &&
- *SI!=OrigDestA && *SI!=OrigDestB) {
- ++SI;
- } else {
- // Otherwise, this is a superfluous edge, remove it.
- MBB.removeSuccessor(SI);
- MadeChange = true;
- }
- }
- if (!AddedFallThrough) {
- assert(DestA == 0 && DestB == 0 &&
- "MachineCFG is missing edges!");
- } else if (isCond) {
- assert(DestA == 0 && "MachineCFG is missing edges!");
- }
- return MadeChange;
-}
-
-
/// CanFallThrough - Return true if the specified block (with the specified
/// branch condition) can implicitly transfer control to the block after it by
/// falling off the end of it. This should return false if it can reach the
TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
if (!PriorUnAnalyzable) {
// If the CFG for the prior block has extra edges, remove them.
- MadeChange |= CorrectExtraCFGEdges(PrevBB, PriorTBB, PriorFBB,
- !PriorCond.empty(), MBB);
+ MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB,
+ !PriorCond.empty());
// If the previous branch is conditional and both conditions go to the same
// destination, remove the branch, replacing it with an unconditional one or
bool CurUnAnalyzable = TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond);
if (!CurUnAnalyzable) {
// If the CFG for the prior block has extra edges, remove them.
- MadeChange |= CorrectExtraCFGEdges(*MBB, CurTBB, CurFBB,
- !CurCond.empty(),
- ++MachineFunction::iterator(MBB));
+ MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty());
// If this is a two-way branch, and the FBB branches to this block, reverse
// the condition so the single-basic-block loop is faster. Instead of:
}
// Iterate through all the predecessors, revectoring each in-turn.
- MachineBasicBlock::pred_iterator PI = MBB->pred_begin();
+ size_t PI = 0;
bool DidChange = false;
bool HasBranchToSelf = false;
- while (PI != MBB->pred_end()) {
- if (*PI == MBB) {
+ while(PI != MBB->pred_size()) {
+ MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI);
+ if (PMBB == MBB) {
// If this block has an uncond branch to itself, leave it.
++PI;
HasBranchToSelf = true;
} else {
DidChange = true;
- (*PI)->ReplaceUsesOfBlockWith(MBB, CurTBB);
+ PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
}
}